linux/fs/btrfs/extent_io.c

5889 lines
160 KiB
C
Raw Normal View History

License cleanup: add SPDX GPL-2.0 license identifier to files with no license Many source files in the tree are missing licensing information, which makes it harder for compliance tools to determine the correct license. By default all files without license information are under the default license of the kernel, which is GPL version 2. Update the files which contain no license information with the 'GPL-2.0' SPDX license identifier. The SPDX identifier is a legally binding shorthand, which can be used instead of the full boiler plate text. This patch is based on work done by Thomas Gleixner and Kate Stewart and Philippe Ombredanne. How this work was done: Patches were generated and checked against linux-4.14-rc6 for a subset of the use cases: - file had no licensing information it it. - file was a */uapi/* one with no licensing information in it, - file was a */uapi/* one with existing licensing information, Further patches will be generated in subsequent months to fix up cases where non-standard license headers were used, and references to license had to be inferred by heuristics based on keywords. The analysis to determine which SPDX License Identifier to be applied to a file was done in a spreadsheet of side by side results from of the output of two independent scanners (ScanCode & Windriver) producing SPDX tag:value files created by Philippe Ombredanne. Philippe prepared the base worksheet, and did an initial spot review of a few 1000 files. The 4.13 kernel was the starting point of the analysis with 60,537 files assessed. Kate Stewart did a file by file comparison of the scanner results in the spreadsheet to determine which SPDX license identifier(s) to be applied to the file. She confirmed any determination that was not immediately clear with lawyers working with the Linux Foundation. Criteria used to select files for SPDX license identifier tagging was: - Files considered eligible had to be source code files. - Make and config files were included as candidates if they contained >5 lines of source - File already had some variant of a license header in it (even if <5 lines). All documentation files were explicitly excluded. The following heuristics were used to determine which SPDX license identifiers to apply. - when both scanners couldn't find any license traces, file was considered to have no license information in it, and the top level COPYING file license applied. For non */uapi/* files that summary was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 11139 and resulted in the first patch in this series. If that file was a */uapi/* path one, it was "GPL-2.0 WITH Linux-syscall-note" otherwise it was "GPL-2.0". Results of that was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 WITH Linux-syscall-note 930 and resulted in the second patch in this series. - if a file had some form of licensing information in it, and was one of the */uapi/* ones, it was denoted with the Linux-syscall-note if any GPL family license was found in the file or had no licensing in it (per prior point). Results summary: SPDX license identifier # files ---------------------------------------------------|------ GPL-2.0 WITH Linux-syscall-note 270 GPL-2.0+ WITH Linux-syscall-note 169 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) 21 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) 17 LGPL-2.1+ WITH Linux-syscall-note 15 GPL-1.0+ WITH Linux-syscall-note 14 ((GPL-2.0+ WITH Linux-syscall-note) OR BSD-3-Clause) 5 LGPL-2.0+ WITH Linux-syscall-note 4 LGPL-2.1 WITH Linux-syscall-note 3 ((GPL-2.0 WITH Linux-syscall-note) OR MIT) 3 ((GPL-2.0 WITH Linux-syscall-note) AND MIT) 1 and that resulted in the third patch in this series. - when the two scanners agreed on the detected license(s), that became the concluded license(s). - when there was disagreement between the two scanners (one detected a license but the other didn't, or they both detected different licenses) a manual inspection of the file occurred. - In most cases a manual inspection of the information in the file resulted in a clear resolution of the license that should apply (and which scanner probably needed to revisit its heuristics). - When it was not immediately clear, the license identifier was confirmed with lawyers working with the Linux Foundation. - If there was any question as to the appropriate license identifier, the file was flagged for further research and to be revisited later in time. In total, over 70 hours of logged manual review was done on the spreadsheet to determine the SPDX license identifiers to apply to the source files by Kate, Philippe, Thomas and, in some cases, confirmation by lawyers working with the Linux Foundation. Kate also obtained a third independent scan of the 4.13 code base from FOSSology, and compared selected files where the other two scanners disagreed against that SPDX file, to see if there was new insights. The Windriver scanner is based on an older version of FOSSology in part, so they are related. Thomas did random spot checks in about 500 files from the spreadsheets for the uapi headers and agreed with SPDX license identifier in the files he inspected. For the non-uapi files Thomas did random spot checks in about 15000 files. In initial set of patches against 4.14-rc6, 3 files were found to have copy/paste license identifier errors, and have been fixed to reflect the correct identifier. Additionally Philippe spent 10 hours this week doing a detailed manual inspection and review of the 12,461 patched files from the initial patch version early this week with: - a full scancode scan run, collecting the matched texts, detected license ids and scores - reviewing anything where there was a license detected (about 500+ files) to ensure that the applied SPDX license was correct - reviewing anything where there was no detection but the patch license was not GPL-2.0 WITH Linux-syscall-note to ensure that the applied SPDX license was correct This produced a worksheet with 20 files needing minor correction. This worksheet was then exported into 3 different .csv files for the different types of files to be modified. These .csv files were then reviewed by Greg. Thomas wrote a script to parse the csv files and add the proper SPDX tag to the file, in the format that the file expected. This script was further refined by Greg based on the output to detect more types of files automatically and to distinguish between header and source .c files (which need different comment types.) Finally Greg ran the script using the .csv files to generate the patches. Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org> Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2017-11-01 22:07:57 +08:00
// SPDX-License-Identifier: GPL-2.0
#include <linux/bitops.h>
#include <linux/slab.h>
#include <linux/bio.h>
#include <linux/mm.h>
#include <linux/pagemap.h>
#include <linux/page-flags.h>
#include <linux/sched/mm.h>
#include <linux/spinlock.h>
#include <linux/blkdev.h>
#include <linux/swap.h>
#include <linux/writeback.h>
#include <linux/pagevec.h>
#include <linux/prefetch.h>
btrfs: initial fsverity support Add support for fsverity in btrfs. To support the generic interface in fs/verity, we add two new item types in the fs tree for inodes with verity enabled. One stores the per-file verity descriptor and btrfs verity item and the other stores the Merkle tree data itself. Verity checking is done in end_page_read just before a page is marked uptodate. This naturally handles a variety of edge cases like holes, preallocated extents, and inline extents. Some care needs to be taken to not try to verity pages past the end of the file, which are accessed by the generic buffered file reading code under some circumstances like reading to the end of the last page and trying to read again. Direct IO on a verity file falls back to buffered reads. Verity relies on PageChecked for the Merkle tree data itself to avoid re-walking up shared paths in the tree. For this reason, we need to cache the Merkle tree data. Since the file is immutable after verity is turned on, we can cache it at an index past EOF. Use the new inode ro_flags to store verity on the inode item, so that we can enable verity on a file, then rollback to an older kernel and still mount the file system and read the file. Since we can't safely write the file anymore without ruining the invariants of the Merkle tree, we mark a ro_compat flag on the file system when a file has verity enabled. Acked-by: Eric Biggers <ebiggers@google.com> Co-developed-by: Chris Mason <clm@fb.com> Signed-off-by: Chris Mason <clm@fb.com> Signed-off-by: Boris Burkov <boris@bur.io> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-01 04:01:49 +08:00
#include <linux/fsverity.h>
#include "misc.h"
#include "extent_io.h"
#include "extent-io-tree.h"
#include "extent_map.h"
#include "ctree.h"
#include "btrfs_inode.h"
#include "volumes.h"
#include "check-integrity.h"
#include "locking.h"
#include "rcu-string.h"
#include "backref.h"
#include "disk-io.h"
#include "subpage.h"
#include "zoned.h"
#include "block-group.h"
#include "compression.h"
#include "fs.h"
#include "accessors.h"
#include "file-item.h"
#include "file.h"
#include "dev-replace.h"
#include "super.h"
static struct kmem_cache *extent_buffer_cache;
#ifdef CONFIG_BTRFS_DEBUG
static inline void btrfs_leak_debug_add_eb(struct extent_buffer *eb)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
unsigned long flags;
spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
list_add(&eb->leak_list, &fs_info->allocated_ebs);
spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
}
static inline void btrfs_leak_debug_del_eb(struct extent_buffer *eb)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
unsigned long flags;
spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
list_del(&eb->leak_list);
spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
}
void btrfs_extent_buffer_leak_debug_check(struct btrfs_fs_info *fs_info)
{
struct extent_buffer *eb;
unsigned long flags;
/*
* If we didn't get into open_ctree our allocated_ebs will not be
* initialized, so just skip this.
*/
if (!fs_info->allocated_ebs.next)
return;
WARN_ON(!list_empty(&fs_info->allocated_ebs));
spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
while (!list_empty(&fs_info->allocated_ebs)) {
eb = list_first_entry(&fs_info->allocated_ebs,
struct extent_buffer, leak_list);
pr_err(
"BTRFS: buffer leak start %llu len %lu refs %d bflags %lu owner %llu\n",
eb->start, eb->len, atomic_read(&eb->refs), eb->bflags,
btrfs_header_owner(eb));
list_del(&eb->leak_list);
kmem_cache_free(extent_buffer_cache, eb);
}
spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
}
#else
#define btrfs_leak_debug_add_eb(eb) do {} while (0)
#define btrfs_leak_debug_del_eb(eb) do {} while (0)
#endif
/*
* Structure to record info about the bio being assembled, and other info like
* how many bytes are there before stripe/ordered extent boundary.
*/
struct btrfs_bio_ctrl {
struct bio *bio;
int mirror_num;
enum btrfs_compression_type compress_type;
u32 len_to_stripe_boundary;
u32 len_to_oe_boundary;
btrfs_bio_end_io_t end_io_func;
/*
* Tell writepage not to lock the state bits for this range, it still
* does the unlocking.
*/
bool extent_locked;
/* Tell the submit_bio code to use REQ_SYNC */
bool sync_io;
};
static void submit_one_bio(struct btrfs_bio_ctrl *bio_ctrl)
{
struct bio *bio;
struct bio_vec *bv;
struct btrfs_inode *inode;
int mirror_num;
if (!bio_ctrl->bio)
return;
bio = bio_ctrl->bio;
bv = bio_first_bvec_all(bio);
inode = BTRFS_I(bv->bv_page->mapping->host);
mirror_num = bio_ctrl->mirror_num;
/* Caller should ensure the bio has at least some range added */
ASSERT(bio->bi_iter.bi_size);
btrfs: avoid double clean up when submit_one_bio() failed [BUG] When running generic/475 with 64K page size and 4K sector size, it has a very high chance (almost 100%) to hang, with mostly data page locked but no one is going to unlock it. [CAUSE] With commit 1784b7d502a9 ("btrfs: handle csum lookup errors properly on reads"), if we failed to lookup checksum due to metadata IO error, we will return error for btrfs_submit_data_bio(). This will cause the page to be unlocked twice in btrfs_do_readpage(): btrfs_do_readpage() |- submit_extent_page() | |- submit_one_bio() | |- btrfs_submit_data_bio() | |- if (ret) { | |- bio->bi_status = ret; | |- bio_endio(bio); } | In the endio function, we will call end_page_read() | and unlock_extent() to cleanup the subpage range. | |- if (ret) { |- unlock_extent(); end_page_read() } Here we unlock the extent and cleanup the subpage range again. For unlock_extent(), it's mostly double unlock safe. But for end_page_read(), it's not, especially for subpage case, as for subpage case we will call btrfs_subpage_end_reader() to reduce the reader number, and use that to number to determine if we need to unlock the full page. If double accounted, it can underflow the number and leave the page locked without anyone to unlock it. [FIX] The commit 1784b7d502a9 ("btrfs: handle csum lookup errors properly on reads") itself is completely fine, it's our existing code not properly handling the error from bio submission hook properly. This patch will make submit_one_bio() to return void so that the callers will never be able to do cleanup when bio submission hook fails. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-04-12 20:30:13 +08:00
btrfs_bio(bio)->file_offset = page_offset(bv->bv_page) + bv->bv_offset;
if (!is_data_inode(&inode->vfs_inode))
btrfs_submit_metadata_bio(inode, bio, mirror_num);
else if (btrfs_op(bio) == BTRFS_MAP_WRITE)
btrfs_submit_data_write_bio(inode, bio, mirror_num);
else
btrfs_submit_data_read_bio(inode, bio, mirror_num,
bio_ctrl->compress_type);
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
/* The bio is owned by the end_io handler now */
bio_ctrl->bio = NULL;
}
/*
* Submit or fail the current bio in the bio_ctrl structure.
*/
static void submit_write_bio(struct btrfs_bio_ctrl *bio_ctrl, int ret)
{
struct bio *bio = bio_ctrl->bio;
if (!bio)
return;
if (ret) {
ASSERT(ret < 0);
btrfs_bio_end_io(btrfs_bio(bio), errno_to_blk_status(ret));
/* The bio is owned by the end_io handler now */
bio_ctrl->bio = NULL;
} else {
submit_one_bio(bio_ctrl);
}
}
int __init extent_buffer_init_cachep(void)
{
extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
sizeof(struct extent_buffer), 0,
SLAB_MEM_SPREAD, NULL);
if (!extent_buffer_cache)
return -ENOMEM;
btrfs: Fix crash due to not allocating integrity data for a bioset When btrfs creates a bioset, we must also allocate the integrity data pool. Otherwise btrfs will crash when it tries to submit a bio to a checksumming disk: BUG: unable to handle kernel NULL pointer dereference at 0000000000000018 IP: [<ffffffff8111e28a>] mempool_alloc+0x4a/0x150 PGD 2305e4067 PUD 23063d067 PMD 0 Oops: 0000 [#1] PREEMPT SMP Modules linked in: btrfs scsi_debug xfs ext4 jbd2 ext3 jbd mbcache sch_fq_codel eeprom lpc_ich mfd_core nfsd exportfs auth_rpcgss af_packet raid6_pq xor zlib_deflate libcrc32c [last unloaded: scsi_debug] CPU: 1 PID: 4486 Comm: mount Not tainted 3.12.0-rc1-mcsum #2 Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011 task: ffff8802451c9720 ti: ffff880230698000 task.ti: ffff880230698000 RIP: 0010:[<ffffffff8111e28a>] [<ffffffff8111e28a>] mempool_alloc+0x4a/0x150 RSP: 0018:ffff880230699688 EFLAGS: 00010286 RAX: 0000000000000001 RBX: 0000000000000000 RCX: 00000000005f8445 RDX: 0000000000000001 RSI: 0000000000000010 RDI: 0000000000000000 RBP: ffff8802306996f8 R08: 0000000000011200 R09: 0000000000000008 R10: 0000000000000020 R11: ffff88009d6e8000 R12: 0000000000011210 R13: 0000000000000030 R14: ffff8802306996b8 R15: ffff8802451c9720 FS: 00007f25b8a16800(0000) GS:ffff88024fc80000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b CR2: 0000000000000018 CR3: 0000000230576000 CR4: 00000000000007e0 Stack: ffff8802451c9720 0000000000000002 ffffffff81a97100 0000000000281250 ffffffff81a96480 ffff88024fc99150 ffff880228d18200 0000000000000000 0000000000000000 0000000000000040 ffff880230e8c2e8 ffff8802459dc900 Call Trace: [<ffffffff811b2208>] bio_integrity_alloc+0x48/0x1b0 [<ffffffff811b26fc>] bio_integrity_prep+0xac/0x360 [<ffffffff8111e298>] ? mempool_alloc+0x58/0x150 [<ffffffffa03e8041>] ? alloc_extent_state+0x31/0x110 [btrfs] [<ffffffff81241579>] blk_queue_bio+0x1c9/0x460 [<ffffffff8123e58a>] generic_make_request+0xca/0x100 [<ffffffff8123e639>] submit_bio+0x79/0x160 [<ffffffffa03f865e>] btrfs_map_bio+0x48e/0x5b0 [btrfs] [<ffffffffa03c821a>] btree_submit_bio_hook+0xda/0x110 [btrfs] [<ffffffffa03e7eba>] submit_one_bio+0x6a/0xa0 [btrfs] [<ffffffffa03ef450>] read_extent_buffer_pages+0x250/0x310 [btrfs] [<ffffffff8125eef6>] ? __radix_tree_preload+0x66/0xf0 [<ffffffff8125f1c5>] ? radix_tree_insert+0x95/0x260 [<ffffffffa03c66f6>] btree_read_extent_buffer_pages.constprop.128+0xb6/0x120 [btrfs] [<ffffffffa03c8c1a>] read_tree_block+0x3a/0x60 [btrfs] [<ffffffffa03caefd>] open_ctree+0x139d/0x2030 [btrfs] [<ffffffffa03a282a>] btrfs_mount+0x53a/0x7d0 [btrfs] [<ffffffff8113ab0b>] ? pcpu_alloc+0x8eb/0x9f0 [<ffffffff81167305>] ? __kmalloc_track_caller+0x35/0x1e0 [<ffffffff81176ba0>] mount_fs+0x20/0xd0 [<ffffffff81191096>] vfs_kern_mount+0x76/0x120 [<ffffffff81193320>] do_mount+0x200/0xa40 [<ffffffff81135cdb>] ? strndup_user+0x5b/0x80 [<ffffffff81193bf0>] SyS_mount+0x90/0xe0 [<ffffffff8156d31d>] system_call_fastpath+0x1a/0x1f Code: 4c 8d 75 a8 4c 89 6d e8 45 89 e0 4c 8d 6f 30 48 89 5d d8 41 83 e0 af 48 89 fb 49 83 c6 18 4c 89 7d f8 65 4c 8b 3c 25 c0 b8 00 00 <48> 8b 73 18 44 89 c7 44 89 45 98 ff 53 20 48 85 c0 48 89 c2 74 RIP [<ffffffff8111e28a>] mempool_alloc+0x4a/0x150 RSP <ffff880230699688> CR2: 0000000000000018 ---[ end trace 7a96042017ed21e2 ]--- Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Josef Bacik <jbacik@fusionio.com> Signed-off-by: Chris Mason <chris.mason@fusionio.com>
2013-09-20 11:37:07 +08:00
return 0;
}
void __cold extent_buffer_free_cachep(void)
{
/*
* Make sure all delayed rcu free are flushed before we
* destroy caches.
*/
rcu_barrier();
kmem_cache_destroy(extent_buffer_cache);
}
void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
{
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
unsigned long index = start >> PAGE_SHIFT;
unsigned long end_index = end >> PAGE_SHIFT;
struct page *page;
while (index <= end_index) {
page = find_get_page(inode->i_mapping, index);
BUG_ON(!page); /* Pages should be in the extent_io_tree */
clear_page_dirty_for_io(page);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
put_page(page);
index++;
}
}
void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
{
struct address_space *mapping = inode->i_mapping;
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
unsigned long index = start >> PAGE_SHIFT;
unsigned long end_index = end >> PAGE_SHIFT;
struct folio *folio;
while (index <= end_index) {
folio = filemap_get_folio(mapping, index);
filemap_dirty_folio(mapping, folio);
folio_account_redirty(folio);
index += folio_nr_pages(folio);
folio_put(folio);
}
}
/*
* Process one page for __process_pages_contig().
*
* Return >0 if we hit @page == @locked_page.
* Return 0 if we updated the page status.
* Return -EGAIN if the we need to try again.
* (For PAGE_LOCK case but got dirty page or page not belong to mapping)
*/
static int process_one_page(struct btrfs_fs_info *fs_info,
struct address_space *mapping,
struct page *page, struct page *locked_page,
unsigned long page_ops, u64 start, u64 end)
{
u32 len;
ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
len = end + 1 - start;
if (page_ops & PAGE_SET_ORDERED)
btrfs_page_clamp_set_ordered(fs_info, page, start, len);
if (page_ops & PAGE_SET_ERROR)
btrfs_page_clamp_set_error(fs_info, page, start, len);
if (page_ops & PAGE_START_WRITEBACK) {
btrfs_page_clamp_clear_dirty(fs_info, page, start, len);
btrfs_page_clamp_set_writeback(fs_info, page, start, len);
}
if (page_ops & PAGE_END_WRITEBACK)
btrfs_page_clamp_clear_writeback(fs_info, page, start, len);
if (page == locked_page)
return 1;
if (page_ops & PAGE_LOCK) {
int ret;
ret = btrfs_page_start_writer_lock(fs_info, page, start, len);
if (ret)
return ret;
if (!PageDirty(page) || page->mapping != mapping) {
btrfs_page_end_writer_lock(fs_info, page, start, len);
return -EAGAIN;
}
}
if (page_ops & PAGE_UNLOCK)
btrfs_page_end_writer_lock(fs_info, page, start, len);
return 0;
}
static int __process_pages_contig(struct address_space *mapping,
struct page *locked_page,
u64 start, u64 end, unsigned long page_ops,
u64 *processed_end)
{
struct btrfs_fs_info *fs_info = btrfs_sb(mapping->host->i_sb);
pgoff_t start_index = start >> PAGE_SHIFT;
pgoff_t end_index = end >> PAGE_SHIFT;
pgoff_t index = start_index;
unsigned long pages_processed = 0;
struct folio_batch fbatch;
int err = 0;
int i;
if (page_ops & PAGE_LOCK) {
ASSERT(page_ops == PAGE_LOCK);
ASSERT(processed_end && *processed_end == start);
}
if ((page_ops & PAGE_SET_ERROR) && start_index <= end_index)
mapping_set_error(mapping, -EIO);
folio_batch_init(&fbatch);
while (index <= end_index) {
int found_folios;
found_folios = filemap_get_folios_contig(mapping, &index,
end_index, &fbatch);
if (found_folios == 0) {
/*
* Only if we're going to lock these pages, we can find
* nothing at @index.
*/
ASSERT(page_ops & PAGE_LOCK);
err = -EAGAIN;
goto out;
}
for (i = 0; i < found_folios; i++) {
int process_ret;
struct folio *folio = fbatch.folios[i];
process_ret = process_one_page(fs_info, mapping,
&folio->page, locked_page, page_ops,
start, end);
if (process_ret < 0) {
err = -EAGAIN;
folio_batch_release(&fbatch);
goto out;
}
pages_processed += folio_nr_pages(folio);
}
folio_batch_release(&fbatch);
cond_resched();
}
out:
if (err && processed_end) {
/*
* Update @processed_end. I know this is awful since it has
* two different return value patterns (inclusive vs exclusive).
*
* But the exclusive pattern is necessary if @start is 0, or we
* underflow and check against processed_end won't work as
* expected.
*/
if (pages_processed)
*processed_end = min(end,
((u64)(start_index + pages_processed) << PAGE_SHIFT) - 1);
else
*processed_end = start;
}
return err;
}
static noinline void __unlock_for_delalloc(struct inode *inode,
struct page *locked_page,
u64 start, u64 end)
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
{
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
unsigned long index = start >> PAGE_SHIFT;
unsigned long end_index = end >> PAGE_SHIFT;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
ASSERT(locked_page);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
if (index == locked_page->index && end_index == index)
return;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
__process_pages_contig(inode->i_mapping, locked_page, start, end,
PAGE_UNLOCK, NULL);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
}
static noinline int lock_delalloc_pages(struct inode *inode,
struct page *locked_page,
u64 delalloc_start,
u64 delalloc_end)
{
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
unsigned long index = delalloc_start >> PAGE_SHIFT;
unsigned long end_index = delalloc_end >> PAGE_SHIFT;
u64 processed_end = delalloc_start;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
int ret;
ASSERT(locked_page);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
if (index == locked_page->index && index == end_index)
return 0;
ret = __process_pages_contig(inode->i_mapping, locked_page, delalloc_start,
delalloc_end, PAGE_LOCK, &processed_end);
if (ret == -EAGAIN && processed_end > delalloc_start)
__unlock_for_delalloc(inode, locked_page, delalloc_start,
processed_end);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
return ret;
}
/*
* Find and lock a contiguous range of bytes in the file marked as delalloc, no
btrfs: subpage: avoid potential deadlock with compression and delalloc [BUG] With experimental subpage compression enabled, a simple fsstress can lead to self deadlock on page 720896: mkfs.btrfs -f -s 4k $dev > /dev/null mount $dev -o compress $mnt $fsstress -p 1 -n 100 -w -d $mnt -v -s 1625511156 [CAUSE] If we have a file layout looks like below: 0 32K 64K 96K 128K |//| |///////////////| 4K Then we run delalloc range for the inode, it will: - Call find_lock_delalloc_range() with @delalloc_start = 0 Then we got a delalloc range [0, 4K). This range will be COWed. - Call find_lock_delalloc_range() again with @delalloc_start = 4K Since find_lock_delalloc_range() never cares whether the range is still inside page range [0, 64K), it will return range [64K, 128K). This range meets the condition for subpage compression, will go through async COW path. And async COW path will return @page_started. But that @page_started is now for range [64K, 128K), not for range [0, 64K). - writepage_dellloc() returned 1 for page [0, 64K) Thus page [0, 64K) will not be unlocked, nor its page dirty status will be cleared. Next time when we try to lock page [0, 64K) we will deadlock, as there is no one to release page [0, 64K). This problem will never happen for regular page size as one page only contains one sector. After the first find_lock_delalloc_range() call, the @delalloc_end will go beyond @page_end no matter if we found a delalloc range or not Thus this bug only happens for subpage, as now we need multiple runs to exhaust the delalloc range of a page. [FIX] Fix the problem by ensuring the delalloc range we ran at least started inside @locked_page. So that we will never get incorrect @page_started. And to prevent such problem from happening again: - Make find_lock_delalloc_range() return false if the found range is beyond @end value passed in. Since @end will be utilized now, add an ASSERT() to ensure we pass correct @end into find_lock_delalloc_range(). This also means, for selftests we needs to populate @end before calling find_lock_delalloc_range(). - New ASSERT() in find_lock_delalloc_range() Now we will make sure the @start/@end passed in at least covers part of the page. - New ASSERT() in run_delalloc_range() To make sure the range at least starts inside @locked page. - Use @delalloc_start as proper cursor, while @delalloc_end is always reset to @page_end. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-27 15:22:07 +08:00
* more than @max_bytes.
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
*
btrfs: subpage: avoid potential deadlock with compression and delalloc [BUG] With experimental subpage compression enabled, a simple fsstress can lead to self deadlock on page 720896: mkfs.btrfs -f -s 4k $dev > /dev/null mount $dev -o compress $mnt $fsstress -p 1 -n 100 -w -d $mnt -v -s 1625511156 [CAUSE] If we have a file layout looks like below: 0 32K 64K 96K 128K |//| |///////////////| 4K Then we run delalloc range for the inode, it will: - Call find_lock_delalloc_range() with @delalloc_start = 0 Then we got a delalloc range [0, 4K). This range will be COWed. - Call find_lock_delalloc_range() again with @delalloc_start = 4K Since find_lock_delalloc_range() never cares whether the range is still inside page range [0, 64K), it will return range [64K, 128K). This range meets the condition for subpage compression, will go through async COW path. And async COW path will return @page_started. But that @page_started is now for range [64K, 128K), not for range [0, 64K). - writepage_dellloc() returned 1 for page [0, 64K) Thus page [0, 64K) will not be unlocked, nor its page dirty status will be cleared. Next time when we try to lock page [0, 64K) we will deadlock, as there is no one to release page [0, 64K). This problem will never happen for regular page size as one page only contains one sector. After the first find_lock_delalloc_range() call, the @delalloc_end will go beyond @page_end no matter if we found a delalloc range or not Thus this bug only happens for subpage, as now we need multiple runs to exhaust the delalloc range of a page. [FIX] Fix the problem by ensuring the delalloc range we ran at least started inside @locked_page. So that we will never get incorrect @page_started. And to prevent such problem from happening again: - Make find_lock_delalloc_range() return false if the found range is beyond @end value passed in. Since @end will be utilized now, add an ASSERT() to ensure we pass correct @end into find_lock_delalloc_range(). This also means, for selftests we needs to populate @end before calling find_lock_delalloc_range(). - New ASSERT() in find_lock_delalloc_range() Now we will make sure the @start/@end passed in at least covers part of the page. - New ASSERT() in run_delalloc_range() To make sure the range at least starts inside @locked page. - Use @delalloc_start as proper cursor, while @delalloc_end is always reset to @page_end. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-27 15:22:07 +08:00
* @start: The original start bytenr to search.
* Will store the extent range start bytenr.
* @end: The original end bytenr of the search range
* Will store the extent range end bytenr.
*
* Return true if we find a delalloc range which starts inside the original
* range, and @start/@end will store the delalloc range start/end.
*
* Return false if we can't find any delalloc range which starts inside the
* original range, and @start/@end will be the non-delalloc range start/end.
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
*/
EXPORT_FOR_TESTS
noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
struct page *locked_page, u64 *start,
u64 *end)
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
{
btrfs: replace BTRFS_MAX_EXTENT_SIZE with fs_info->max_extent_size On zoned filesystem, data write out is limited by max_zone_append_size, and a large ordered extent is split according the size of a bio. OTOH, the number of extents to be written is calculated using BTRFS_MAX_EXTENT_SIZE, and that estimated number is used to reserve the metadata bytes to update and/or create the metadata items. The metadata reservation is done at e.g, btrfs_buffered_write() and then released according to the estimation changes. Thus, if the number of extent increases massively, the reserved metadata can run out. The increase of the number of extents easily occurs on zoned filesystem if BTRFS_MAX_EXTENT_SIZE > max_zone_append_size. And, it causes the following warning on a small RAM environment with disabling metadata over-commit (in the following patch). [75721.498492] ------------[ cut here ]------------ [75721.505624] BTRFS: block rsv 1 returned -28 [75721.512230] WARNING: CPU: 24 PID: 2327559 at fs/btrfs/block-rsv.c:537 btrfs_use_block_rsv+0x560/0x760 [btrfs] [75721.581854] CPU: 24 PID: 2327559 Comm: kworker/u64:10 Kdump: loaded Tainted: G W 5.18.0-rc2-BTRFS-ZNS+ #109 [75721.597200] Hardware name: Supermicro Super Server/H12SSL-NT, BIOS 2.0 02/22/2021 [75721.607310] Workqueue: btrfs-endio-write btrfs_work_helper [btrfs] [75721.616209] RIP: 0010:btrfs_use_block_rsv+0x560/0x760 [btrfs] [75721.646649] RSP: 0018:ffffc9000fbdf3e0 EFLAGS: 00010286 [75721.654126] RAX: 0000000000000000 RBX: 0000000000004000 RCX: 0000000000000000 [75721.663524] RDX: 0000000000000004 RSI: 0000000000000008 RDI: fffff52001f7be6e [75721.672921] RBP: ffffc9000fbdf420 R08: 0000000000000001 R09: ffff889f8d1fc6c7 [75721.682493] R10: ffffed13f1a3f8d8 R11: 0000000000000001 R12: ffff88980a3c0e28 [75721.692284] R13: ffff889b66590000 R14: ffff88980a3c0e40 R15: ffff88980a3c0e8a [75721.701878] FS: 0000000000000000(0000) GS:ffff889f8d000000(0000) knlGS:0000000000000000 [75721.712601] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [75721.720726] CR2: 000055d12e05c018 CR3: 0000800193594000 CR4: 0000000000350ee0 [75721.730499] Call Trace: [75721.735166] <TASK> [75721.739886] btrfs_alloc_tree_block+0x1e1/0x1100 [btrfs] [75721.747545] ? btrfs_alloc_logged_file_extent+0x550/0x550 [btrfs] [75721.756145] ? btrfs_get_32+0xea/0x2d0 [btrfs] [75721.762852] ? btrfs_get_32+0xea/0x2d0 [btrfs] [75721.769520] ? push_leaf_left+0x420/0x620 [btrfs] [75721.776431] ? memcpy+0x4e/0x60 [75721.781931] split_leaf+0x433/0x12d0 [btrfs] [75721.788392] ? btrfs_get_token_32+0x580/0x580 [btrfs] [75721.795636] ? push_for_double_split.isra.0+0x420/0x420 [btrfs] [75721.803759] ? leaf_space_used+0x15d/0x1a0 [btrfs] [75721.811156] btrfs_search_slot+0x1bc3/0x2790 [btrfs] [75721.818300] ? lock_downgrade+0x7c0/0x7c0 [75721.824411] ? free_extent_buffer.part.0+0x107/0x200 [btrfs] [75721.832456] ? split_leaf+0x12d0/0x12d0 [btrfs] [75721.839149] ? free_extent_buffer.part.0+0x14f/0x200 [btrfs] [75721.846945] ? free_extent_buffer+0x13/0x20 [btrfs] [75721.853960] ? btrfs_release_path+0x4b/0x190 [btrfs] [75721.861429] btrfs_csum_file_blocks+0x85c/0x1500 [btrfs] [75721.869313] ? rcu_read_lock_sched_held+0x16/0x80 [75721.876085] ? lock_release+0x552/0xf80 [75721.881957] ? btrfs_del_csums+0x8c0/0x8c0 [btrfs] [75721.888886] ? __kasan_check_write+0x14/0x20 [75721.895152] ? do_raw_read_unlock+0x44/0x80 [75721.901323] ? _raw_write_lock_irq+0x60/0x80 [75721.907983] ? btrfs_global_root+0xb9/0xe0 [btrfs] [75721.915166] ? btrfs_csum_root+0x12b/0x180 [btrfs] [75721.921918] ? btrfs_get_global_root+0x820/0x820 [btrfs] [75721.929166] ? _raw_write_unlock+0x23/0x40 [75721.935116] ? unpin_extent_cache+0x1e3/0x390 [btrfs] [75721.942041] btrfs_finish_ordered_io.isra.0+0xa0c/0x1dc0 [btrfs] [75721.949906] ? try_to_wake_up+0x30/0x14a0 [75721.955700] ? btrfs_unlink_subvol+0xda0/0xda0 [btrfs] [75721.962661] ? rcu_read_lock_sched_held+0x16/0x80 [75721.969111] ? lock_acquire+0x41b/0x4c0 [75721.974982] finish_ordered_fn+0x15/0x20 [btrfs] [75721.981639] btrfs_work_helper+0x1af/0xa80 [btrfs] [75721.988184] ? _raw_spin_unlock_irq+0x28/0x50 [75721.994643] process_one_work+0x815/0x1460 [75722.000444] ? pwq_dec_nr_in_flight+0x250/0x250 [75722.006643] ? do_raw_spin_trylock+0xbb/0x190 [75722.013086] worker_thread+0x59a/0xeb0 [75722.018511] kthread+0x2ac/0x360 [75722.023428] ? process_one_work+0x1460/0x1460 [75722.029431] ? kthread_complete_and_exit+0x30/0x30 [75722.036044] ret_from_fork+0x22/0x30 [75722.041255] </TASK> [75722.045047] irq event stamp: 0 [75722.049703] hardirqs last enabled at (0): [<0000000000000000>] 0x0 [75722.057610] hardirqs last disabled at (0): [<ffffffff8118a94a>] copy_process+0x1c1a/0x66b0 [75722.067533] softirqs last enabled at (0): [<ffffffff8118a989>] copy_process+0x1c59/0x66b0 [75722.077423] softirqs last disabled at (0): [<0000000000000000>] 0x0 [75722.085335] ---[ end trace 0000000000000000 ]--- To fix the estimation, we need to introduce fs_info->max_extent_size to replace BTRFS_MAX_EXTENT_SIZE, which allow setting the different size for regular vs zoned filesystem. Set fs_info->max_extent_size to BTRFS_MAX_EXTENT_SIZE by default. On zoned filesystem, it is set to fs_info->max_zone_append_size. CC: stable@vger.kernel.org # 5.12+ Fixes: d8e3fb106f39 ("btrfs: zoned: use ZONE_APPEND write for zoned mode") Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-07-09 07:18:40 +08:00
struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
btrfs: subpage: avoid potential deadlock with compression and delalloc [BUG] With experimental subpage compression enabled, a simple fsstress can lead to self deadlock on page 720896: mkfs.btrfs -f -s 4k $dev > /dev/null mount $dev -o compress $mnt $fsstress -p 1 -n 100 -w -d $mnt -v -s 1625511156 [CAUSE] If we have a file layout looks like below: 0 32K 64K 96K 128K |//| |///////////////| 4K Then we run delalloc range for the inode, it will: - Call find_lock_delalloc_range() with @delalloc_start = 0 Then we got a delalloc range [0, 4K). This range will be COWed. - Call find_lock_delalloc_range() again with @delalloc_start = 4K Since find_lock_delalloc_range() never cares whether the range is still inside page range [0, 64K), it will return range [64K, 128K). This range meets the condition for subpage compression, will go through async COW path. And async COW path will return @page_started. But that @page_started is now for range [64K, 128K), not for range [0, 64K). - writepage_dellloc() returned 1 for page [0, 64K) Thus page [0, 64K) will not be unlocked, nor its page dirty status will be cleared. Next time when we try to lock page [0, 64K) we will deadlock, as there is no one to release page [0, 64K). This problem will never happen for regular page size as one page only contains one sector. After the first find_lock_delalloc_range() call, the @delalloc_end will go beyond @page_end no matter if we found a delalloc range or not Thus this bug only happens for subpage, as now we need multiple runs to exhaust the delalloc range of a page. [FIX] Fix the problem by ensuring the delalloc range we ran at least started inside @locked_page. So that we will never get incorrect @page_started. And to prevent such problem from happening again: - Make find_lock_delalloc_range() return false if the found range is beyond @end value passed in. Since @end will be utilized now, add an ASSERT() to ensure we pass correct @end into find_lock_delalloc_range(). This also means, for selftests we needs to populate @end before calling find_lock_delalloc_range(). - New ASSERT() in find_lock_delalloc_range() Now we will make sure the @start/@end passed in at least covers part of the page. - New ASSERT() in run_delalloc_range() To make sure the range at least starts inside @locked page. - Use @delalloc_start as proper cursor, while @delalloc_end is always reset to @page_end. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-27 15:22:07 +08:00
const u64 orig_start = *start;
const u64 orig_end = *end;
btrfs: replace BTRFS_MAX_EXTENT_SIZE with fs_info->max_extent_size On zoned filesystem, data write out is limited by max_zone_append_size, and a large ordered extent is split according the size of a bio. OTOH, the number of extents to be written is calculated using BTRFS_MAX_EXTENT_SIZE, and that estimated number is used to reserve the metadata bytes to update and/or create the metadata items. The metadata reservation is done at e.g, btrfs_buffered_write() and then released according to the estimation changes. Thus, if the number of extent increases massively, the reserved metadata can run out. The increase of the number of extents easily occurs on zoned filesystem if BTRFS_MAX_EXTENT_SIZE > max_zone_append_size. And, it causes the following warning on a small RAM environment with disabling metadata over-commit (in the following patch). [75721.498492] ------------[ cut here ]------------ [75721.505624] BTRFS: block rsv 1 returned -28 [75721.512230] WARNING: CPU: 24 PID: 2327559 at fs/btrfs/block-rsv.c:537 btrfs_use_block_rsv+0x560/0x760 [btrfs] [75721.581854] CPU: 24 PID: 2327559 Comm: kworker/u64:10 Kdump: loaded Tainted: G W 5.18.0-rc2-BTRFS-ZNS+ #109 [75721.597200] Hardware name: Supermicro Super Server/H12SSL-NT, BIOS 2.0 02/22/2021 [75721.607310] Workqueue: btrfs-endio-write btrfs_work_helper [btrfs] [75721.616209] RIP: 0010:btrfs_use_block_rsv+0x560/0x760 [btrfs] [75721.646649] RSP: 0018:ffffc9000fbdf3e0 EFLAGS: 00010286 [75721.654126] RAX: 0000000000000000 RBX: 0000000000004000 RCX: 0000000000000000 [75721.663524] RDX: 0000000000000004 RSI: 0000000000000008 RDI: fffff52001f7be6e [75721.672921] RBP: ffffc9000fbdf420 R08: 0000000000000001 R09: ffff889f8d1fc6c7 [75721.682493] R10: ffffed13f1a3f8d8 R11: 0000000000000001 R12: ffff88980a3c0e28 [75721.692284] R13: ffff889b66590000 R14: ffff88980a3c0e40 R15: ffff88980a3c0e8a [75721.701878] FS: 0000000000000000(0000) GS:ffff889f8d000000(0000) knlGS:0000000000000000 [75721.712601] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [75721.720726] CR2: 000055d12e05c018 CR3: 0000800193594000 CR4: 0000000000350ee0 [75721.730499] Call Trace: [75721.735166] <TASK> [75721.739886] btrfs_alloc_tree_block+0x1e1/0x1100 [btrfs] [75721.747545] ? btrfs_alloc_logged_file_extent+0x550/0x550 [btrfs] [75721.756145] ? btrfs_get_32+0xea/0x2d0 [btrfs] [75721.762852] ? btrfs_get_32+0xea/0x2d0 [btrfs] [75721.769520] ? push_leaf_left+0x420/0x620 [btrfs] [75721.776431] ? memcpy+0x4e/0x60 [75721.781931] split_leaf+0x433/0x12d0 [btrfs] [75721.788392] ? btrfs_get_token_32+0x580/0x580 [btrfs] [75721.795636] ? push_for_double_split.isra.0+0x420/0x420 [btrfs] [75721.803759] ? leaf_space_used+0x15d/0x1a0 [btrfs] [75721.811156] btrfs_search_slot+0x1bc3/0x2790 [btrfs] [75721.818300] ? lock_downgrade+0x7c0/0x7c0 [75721.824411] ? free_extent_buffer.part.0+0x107/0x200 [btrfs] [75721.832456] ? split_leaf+0x12d0/0x12d0 [btrfs] [75721.839149] ? free_extent_buffer.part.0+0x14f/0x200 [btrfs] [75721.846945] ? free_extent_buffer+0x13/0x20 [btrfs] [75721.853960] ? btrfs_release_path+0x4b/0x190 [btrfs] [75721.861429] btrfs_csum_file_blocks+0x85c/0x1500 [btrfs] [75721.869313] ? rcu_read_lock_sched_held+0x16/0x80 [75721.876085] ? lock_release+0x552/0xf80 [75721.881957] ? btrfs_del_csums+0x8c0/0x8c0 [btrfs] [75721.888886] ? __kasan_check_write+0x14/0x20 [75721.895152] ? do_raw_read_unlock+0x44/0x80 [75721.901323] ? _raw_write_lock_irq+0x60/0x80 [75721.907983] ? btrfs_global_root+0xb9/0xe0 [btrfs] [75721.915166] ? btrfs_csum_root+0x12b/0x180 [btrfs] [75721.921918] ? btrfs_get_global_root+0x820/0x820 [btrfs] [75721.929166] ? _raw_write_unlock+0x23/0x40 [75721.935116] ? unpin_extent_cache+0x1e3/0x390 [btrfs] [75721.942041] btrfs_finish_ordered_io.isra.0+0xa0c/0x1dc0 [btrfs] [75721.949906] ? try_to_wake_up+0x30/0x14a0 [75721.955700] ? btrfs_unlink_subvol+0xda0/0xda0 [btrfs] [75721.962661] ? rcu_read_lock_sched_held+0x16/0x80 [75721.969111] ? lock_acquire+0x41b/0x4c0 [75721.974982] finish_ordered_fn+0x15/0x20 [btrfs] [75721.981639] btrfs_work_helper+0x1af/0xa80 [btrfs] [75721.988184] ? _raw_spin_unlock_irq+0x28/0x50 [75721.994643] process_one_work+0x815/0x1460 [75722.000444] ? pwq_dec_nr_in_flight+0x250/0x250 [75722.006643] ? do_raw_spin_trylock+0xbb/0x190 [75722.013086] worker_thread+0x59a/0xeb0 [75722.018511] kthread+0x2ac/0x360 [75722.023428] ? process_one_work+0x1460/0x1460 [75722.029431] ? kthread_complete_and_exit+0x30/0x30 [75722.036044] ret_from_fork+0x22/0x30 [75722.041255] </TASK> [75722.045047] irq event stamp: 0 [75722.049703] hardirqs last enabled at (0): [<0000000000000000>] 0x0 [75722.057610] hardirqs last disabled at (0): [<ffffffff8118a94a>] copy_process+0x1c1a/0x66b0 [75722.067533] softirqs last enabled at (0): [<ffffffff8118a989>] copy_process+0x1c59/0x66b0 [75722.077423] softirqs last disabled at (0): [<0000000000000000>] 0x0 [75722.085335] ---[ end trace 0000000000000000 ]--- To fix the estimation, we need to introduce fs_info->max_extent_size to replace BTRFS_MAX_EXTENT_SIZE, which allow setting the different size for regular vs zoned filesystem. Set fs_info->max_extent_size to BTRFS_MAX_EXTENT_SIZE by default. On zoned filesystem, it is set to fs_info->max_zone_append_size. CC: stable@vger.kernel.org # 5.12+ Fixes: d8e3fb106f39 ("btrfs: zoned: use ZONE_APPEND write for zoned mode") Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-07-09 07:18:40 +08:00
/* The sanity tests may not set a valid fs_info. */
u64 max_bytes = fs_info ? fs_info->max_extent_size : BTRFS_MAX_EXTENT_SIZE;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
u64 delalloc_start;
u64 delalloc_end;
bool found;
struct extent_state *cached_state = NULL;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
int ret;
int loops = 0;
btrfs: subpage: avoid potential deadlock with compression and delalloc [BUG] With experimental subpage compression enabled, a simple fsstress can lead to self deadlock on page 720896: mkfs.btrfs -f -s 4k $dev > /dev/null mount $dev -o compress $mnt $fsstress -p 1 -n 100 -w -d $mnt -v -s 1625511156 [CAUSE] If we have a file layout looks like below: 0 32K 64K 96K 128K |//| |///////////////| 4K Then we run delalloc range for the inode, it will: - Call find_lock_delalloc_range() with @delalloc_start = 0 Then we got a delalloc range [0, 4K). This range will be COWed. - Call find_lock_delalloc_range() again with @delalloc_start = 4K Since find_lock_delalloc_range() never cares whether the range is still inside page range [0, 64K), it will return range [64K, 128K). This range meets the condition for subpage compression, will go through async COW path. And async COW path will return @page_started. But that @page_started is now for range [64K, 128K), not for range [0, 64K). - writepage_dellloc() returned 1 for page [0, 64K) Thus page [0, 64K) will not be unlocked, nor its page dirty status will be cleared. Next time when we try to lock page [0, 64K) we will deadlock, as there is no one to release page [0, 64K). This problem will never happen for regular page size as one page only contains one sector. After the first find_lock_delalloc_range() call, the @delalloc_end will go beyond @page_end no matter if we found a delalloc range or not Thus this bug only happens for subpage, as now we need multiple runs to exhaust the delalloc range of a page. [FIX] Fix the problem by ensuring the delalloc range we ran at least started inside @locked_page. So that we will never get incorrect @page_started. And to prevent such problem from happening again: - Make find_lock_delalloc_range() return false if the found range is beyond @end value passed in. Since @end will be utilized now, add an ASSERT() to ensure we pass correct @end into find_lock_delalloc_range(). This also means, for selftests we needs to populate @end before calling find_lock_delalloc_range(). - New ASSERT() in find_lock_delalloc_range() Now we will make sure the @start/@end passed in at least covers part of the page. - New ASSERT() in run_delalloc_range() To make sure the range at least starts inside @locked page. - Use @delalloc_start as proper cursor, while @delalloc_end is always reset to @page_end. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-27 15:22:07 +08:00
/* Caller should pass a valid @end to indicate the search range end */
ASSERT(orig_end > orig_start);
/* The range should at least cover part of the page */
ASSERT(!(orig_start >= page_offset(locked_page) + PAGE_SIZE ||
orig_end <= page_offset(locked_page)));
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
again:
/* step one, find a bunch of delalloc bytes starting at start */
delalloc_start = *start;
delalloc_end = 0;
found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
max_bytes, &cached_state);
btrfs: subpage: avoid potential deadlock with compression and delalloc [BUG] With experimental subpage compression enabled, a simple fsstress can lead to self deadlock on page 720896: mkfs.btrfs -f -s 4k $dev > /dev/null mount $dev -o compress $mnt $fsstress -p 1 -n 100 -w -d $mnt -v -s 1625511156 [CAUSE] If we have a file layout looks like below: 0 32K 64K 96K 128K |//| |///////////////| 4K Then we run delalloc range for the inode, it will: - Call find_lock_delalloc_range() with @delalloc_start = 0 Then we got a delalloc range [0, 4K). This range will be COWed. - Call find_lock_delalloc_range() again with @delalloc_start = 4K Since find_lock_delalloc_range() never cares whether the range is still inside page range [0, 64K), it will return range [64K, 128K). This range meets the condition for subpage compression, will go through async COW path. And async COW path will return @page_started. But that @page_started is now for range [64K, 128K), not for range [0, 64K). - writepage_dellloc() returned 1 for page [0, 64K) Thus page [0, 64K) will not be unlocked, nor its page dirty status will be cleared. Next time when we try to lock page [0, 64K) we will deadlock, as there is no one to release page [0, 64K). This problem will never happen for regular page size as one page only contains one sector. After the first find_lock_delalloc_range() call, the @delalloc_end will go beyond @page_end no matter if we found a delalloc range or not Thus this bug only happens for subpage, as now we need multiple runs to exhaust the delalloc range of a page. [FIX] Fix the problem by ensuring the delalloc range we ran at least started inside @locked_page. So that we will never get incorrect @page_started. And to prevent such problem from happening again: - Make find_lock_delalloc_range() return false if the found range is beyond @end value passed in. Since @end will be utilized now, add an ASSERT() to ensure we pass correct @end into find_lock_delalloc_range(). This also means, for selftests we needs to populate @end before calling find_lock_delalloc_range(). - New ASSERT() in find_lock_delalloc_range() Now we will make sure the @start/@end passed in at least covers part of the page. - New ASSERT() in run_delalloc_range() To make sure the range at least starts inside @locked page. - Use @delalloc_start as proper cursor, while @delalloc_end is always reset to @page_end. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-27 15:22:07 +08:00
if (!found || delalloc_end <= *start || delalloc_start > orig_end) {
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
*start = delalloc_start;
btrfs: subpage: avoid potential deadlock with compression and delalloc [BUG] With experimental subpage compression enabled, a simple fsstress can lead to self deadlock on page 720896: mkfs.btrfs -f -s 4k $dev > /dev/null mount $dev -o compress $mnt $fsstress -p 1 -n 100 -w -d $mnt -v -s 1625511156 [CAUSE] If we have a file layout looks like below: 0 32K 64K 96K 128K |//| |///////////////| 4K Then we run delalloc range for the inode, it will: - Call find_lock_delalloc_range() with @delalloc_start = 0 Then we got a delalloc range [0, 4K). This range will be COWed. - Call find_lock_delalloc_range() again with @delalloc_start = 4K Since find_lock_delalloc_range() never cares whether the range is still inside page range [0, 64K), it will return range [64K, 128K). This range meets the condition for subpage compression, will go through async COW path. And async COW path will return @page_started. But that @page_started is now for range [64K, 128K), not for range [0, 64K). - writepage_dellloc() returned 1 for page [0, 64K) Thus page [0, 64K) will not be unlocked, nor its page dirty status will be cleared. Next time when we try to lock page [0, 64K) we will deadlock, as there is no one to release page [0, 64K). This problem will never happen for regular page size as one page only contains one sector. After the first find_lock_delalloc_range() call, the @delalloc_end will go beyond @page_end no matter if we found a delalloc range or not Thus this bug only happens for subpage, as now we need multiple runs to exhaust the delalloc range of a page. [FIX] Fix the problem by ensuring the delalloc range we ran at least started inside @locked_page. So that we will never get incorrect @page_started. And to prevent such problem from happening again: - Make find_lock_delalloc_range() return false if the found range is beyond @end value passed in. Since @end will be utilized now, add an ASSERT() to ensure we pass correct @end into find_lock_delalloc_range(). This also means, for selftests we needs to populate @end before calling find_lock_delalloc_range(). - New ASSERT() in find_lock_delalloc_range() Now we will make sure the @start/@end passed in at least covers part of the page. - New ASSERT() in run_delalloc_range() To make sure the range at least starts inside @locked page. - Use @delalloc_start as proper cursor, while @delalloc_end is always reset to @page_end. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-27 15:22:07 +08:00
/* @delalloc_end can be -1, never go beyond @orig_end */
*end = min(delalloc_end, orig_end);
free_extent_state(cached_state);
return false;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
}
/*
* start comes from the offset of locked_page. We have to lock
* pages in order, so we can't process delalloc bytes before
* locked_page
*/
if (delalloc_start < *start)
delalloc_start = *start;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
/*
* make sure to limit the number of pages we try to lock down
*/
if (delalloc_end + 1 - delalloc_start > max_bytes)
delalloc_end = delalloc_start + max_bytes - 1;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
/* step two, lock all the pages after the page that has start */
ret = lock_delalloc_pages(inode, locked_page,
delalloc_start, delalloc_end);
ASSERT(!ret || ret == -EAGAIN);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
if (ret == -EAGAIN) {
/* some of the pages are gone, lets avoid looping by
* shortening the size of the delalloc range we're searching
*/
free_extent_state(cached_state);
cached_state = NULL;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
if (!loops) {
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
max_bytes = PAGE_SIZE;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
loops = 1;
goto again;
} else {
found = false;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
goto out_failed;
}
}
/* step three, lock the state bits for the whole range */
lock_extent(tree, delalloc_start, delalloc_end, &cached_state);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
/* then test to make sure it is all still delalloc */
ret = test_range_bit(tree, delalloc_start, delalloc_end,
EXTENT_DELALLOC, 1, cached_state);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
if (!ret) {
unlock_extent(tree, delalloc_start, delalloc_end,
&cached_state);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
__unlock_for_delalloc(inode, locked_page,
delalloc_start, delalloc_end);
cond_resched();
goto again;
}
free_extent_state(cached_state);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
*start = delalloc_start;
*end = delalloc_end;
out_failed:
return found;
}
void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
struct page *locked_page,
u32 clear_bits, unsigned long page_ops)
{
clear_extent_bit(&inode->io_tree, start, end, clear_bits, NULL);
__process_pages_contig(inode->vfs_inode.i_mapping, locked_page,
start, end, page_ops, NULL);
}
static int insert_failrec(struct btrfs_inode *inode,
struct io_failure_record *failrec)
{
struct rb_node *exist;
spin_lock(&inode->io_failure_lock);
exist = rb_simple_insert(&inode->io_failure_tree, failrec->bytenr,
&failrec->rb_node);
spin_unlock(&inode->io_failure_lock);
return (exist == NULL) ? 0 : -EEXIST;
}
static struct io_failure_record *get_failrec(struct btrfs_inode *inode, u64 start)
{
struct rb_node *node;
struct io_failure_record *failrec = ERR_PTR(-ENOENT);
spin_lock(&inode->io_failure_lock);
node = rb_simple_search(&inode->io_failure_tree, start);
if (node)
failrec = rb_entry(node, struct io_failure_record, rb_node);
spin_unlock(&inode->io_failure_lock);
return failrec;
}
static void free_io_failure(struct btrfs_inode *inode,
struct io_failure_record *rec)
{
spin_lock(&inode->io_failure_lock);
rb_erase(&rec->rb_node, &inode->io_failure_tree);
spin_unlock(&inode->io_failure_lock);
kfree(rec);
}
/*
* this bypasses the standard btrfs submit functions deliberately, as
* the standard behavior is to write all copies in a raid setup. here we only
* want to write the one bad copy. so we do the mapping for ourselves and issue
* submit_bio directly.
* to avoid any synchronization issues, wait for the data after writing, which
* actually prevents the read that triggered the error from finishing.
* currently, there can be no more than two copies of every data bit. thus,
* exactly one rewrite is required.
*/
static int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
u64 length, u64 logical, struct page *page,
unsigned int pg_offset, int mirror_num)
{
struct btrfs_device *dev;
struct bio_vec bvec;
struct bio bio;
u64 map_length = 0;
u64 sector;
struct btrfs_io_context *bioc = NULL;
int ret = 0;
Rename superblock flags (MS_xyz -> SB_xyz) This is a pure automated search-and-replace of the internal kernel superblock flags. The s_flags are now called SB_*, with the names and the values for the moment mirroring the MS_* flags that they're equivalent to. Note how the MS_xyz flags are the ones passed to the mount system call, while the SB_xyz flags are what we then use in sb->s_flags. The script to do this was: # places to look in; re security/*: it generally should *not* be # touched (that stuff parses mount(2) arguments directly), but # there are two places where we really deal with superblock flags. FILES="drivers/mtd drivers/staging/lustre fs ipc mm \ include/linux/fs.h include/uapi/linux/bfs_fs.h \ security/apparmor/apparmorfs.c security/apparmor/include/lib.h" # the list of MS_... constants SYMS="RDONLY NOSUID NODEV NOEXEC SYNCHRONOUS REMOUNT MANDLOCK \ DIRSYNC NOATIME NODIRATIME BIND MOVE REC VERBOSE SILENT \ POSIXACL UNBINDABLE PRIVATE SLAVE SHARED RELATIME KERNMOUNT \ I_VERSION STRICTATIME LAZYTIME SUBMOUNT NOREMOTELOCK NOSEC BORN \ ACTIVE NOUSER" SED_PROG= for i in $SYMS; do SED_PROG="$SED_PROG -e s/MS_$i/SB_$i/g"; done # we want files that contain at least one of MS_..., # with fs/namespace.c and fs/pnode.c excluded. L=$(for i in $SYMS; do git grep -w -l MS_$i $FILES; done| sort|uniq|grep -v '^fs/namespace.c'|grep -v '^fs/pnode.c') for f in $L; do sed -i $f $SED_PROG; done Requested-by: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-28 05:05:09 +08:00
ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
BUG_ON(!mirror_num);
if (btrfs_repair_one_zone(fs_info, logical))
return 0;
btrfs: zoned: relocate block group to repair IO failure in zoned filesystems When a bad checksum is found and if the filesystem has a mirror of the damaged data, we read the correct data from the mirror and writes it to damaged blocks. This however, violates the sequential write constraints of a zoned block device. We can consider three methods to repair an IO failure in zoned filesystems: (1) Reset and rewrite the damaged zone (2) Allocate new device extent and replace the damaged device extent to the new extent (3) Relocate the corresponding block group Method (1) is most similar to a behavior done with regular devices. However, it also wipes non-damaged data in the same device extent, and so it unnecessary degrades non-damaged data. Method (2) is much like device replacing but done in the same device. It is safe because it keeps the device extent until the replacing finish. However, extending device replacing is non-trivial. It assumes "src_dev->physical == dst_dev->physical". Also, the extent mapping replacing function should be extended to support replacing device extent position in one device. Method (3) invokes relocation of the damaged block group and is straightforward to implement. It relocates all the mirrored device extents, so it potentially is a more costly operation than method (1) or (2). But it relocates only used extents which reduce the total IO size. Let's apply method (3) for now. In the future, we can extend device-replace and apply method (2). For protecting a block group gets relocated multiple time with multiple IO errors, this commit introduces "relocating_repair" bit to show it's now relocating to repair IO failures. Also it uses a new kthread "btrfs-relocating-repair", not to block IO path with relocating process. This commit also supports repairing in the scrub process. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-04 18:22:16 +08:00
map_length = length;
/*
* Avoid races with device replace and make sure our bioc has devices
* associated to its stripes that don't go away while we are doing the
* read repair operation.
*/
btrfs_bio_counter_inc_blocked(fs_info);
if (btrfs_is_parity_mirror(fs_info, logical, length)) {
/*
* Note that we don't use BTRFS_MAP_WRITE because it's supposed
* to update all raid stripes, but here we just want to correct
* bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
* stripe's dev and sector.
*/
ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
&map_length, &bioc, 0);
if (ret)
goto out_counter_dec;
ASSERT(bioc->mirror_num == 1);
} else {
ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
&map_length, &bioc, mirror_num);
if (ret)
goto out_counter_dec;
BUG_ON(mirror_num != bioc->mirror_num);
}
sector = bioc->stripes[bioc->mirror_num - 1].physical >> 9;
dev = bioc->stripes[bioc->mirror_num - 1].dev;
btrfs_put_bioc(bioc);
if (!dev || !dev->bdev ||
!test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
ret = -EIO;
goto out_counter_dec;
}
bio_init(&bio, dev->bdev, &bvec, 1, REQ_OP_WRITE | REQ_SYNC);
bio.bi_iter.bi_sector = sector;
__bio_add_page(&bio, page, length, pg_offset);
btrfsic_check_bio(&bio);
ret = submit_bio_wait(&bio);
if (ret) {
/* try to remap that extent elsewhere? */
btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
goto out_bio_uninit;
}
btrfs_info_rl_in_rcu(fs_info,
"read error corrected: ino %llu off %llu (dev %s sector %llu)",
ino, start,
rcu_str_deref(dev->name), sector);
ret = 0;
out_bio_uninit:
bio_uninit(&bio);
out_counter_dec:
btrfs_bio_counter_dec(fs_info);
return ret;
}
int btrfs_repair_eb_io_failure(const struct extent_buffer *eb, int mirror_num)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
u64 start = eb->start;
int i, num_pages = num_extent_pages(eb);
int ret = 0;
if (sb_rdonly(fs_info->sb))
return -EROFS;
for (i = 0; i < num_pages; i++) {
struct page *p = eb->pages[i];
ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
start - page_offset(p), mirror_num);
if (ret)
break;
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
start += PAGE_SIZE;
}
return ret;
}
static int next_mirror(const struct io_failure_record *failrec, int cur_mirror)
{
if (cur_mirror == failrec->num_copies)
return cur_mirror + 1 - failrec->num_copies;
return cur_mirror + 1;
}
static int prev_mirror(const struct io_failure_record *failrec, int cur_mirror)
{
if (cur_mirror == 1)
return failrec->num_copies;
return cur_mirror - 1;
}
/*
* each time an IO finishes, we do a fast check in the IO failure tree
* to see if we need to process or clean up an io_failure_record
*/
int btrfs_clean_io_failure(struct btrfs_inode *inode, u64 start,
struct page *page, unsigned int pg_offset)
{
struct btrfs_fs_info *fs_info = inode->root->fs_info;
struct extent_io_tree *io_tree = &inode->io_tree;
u64 ino = btrfs_ino(inode);
u64 locked_start, locked_end;
struct io_failure_record *failrec;
int mirror;
int ret;
failrec = get_failrec(inode, start);
if (IS_ERR(failrec))
return 0;
BUG_ON(!failrec->this_mirror);
if (sb_rdonly(fs_info->sb))
goto out;
ret = find_first_extent_bit(io_tree, failrec->bytenr, &locked_start,
&locked_end, EXTENT_LOCKED, NULL);
if (ret || locked_start > failrec->bytenr ||
locked_end < failrec->bytenr + failrec->len - 1)
goto out;
mirror = failrec->this_mirror;
do {
mirror = prev_mirror(failrec, mirror);
repair_io_failure(fs_info, ino, start, failrec->len,
failrec->logical, page, pg_offset, mirror);
} while (mirror != failrec->failed_mirror);
out:
free_io_failure(inode, failrec);
return 0;
}
/*
* Can be called when
* - hold extent lock
* - under ordered extent
* - the inode is freeing
*/
void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
{
struct io_failure_record *failrec;
struct rb_node *node, *next;
if (RB_EMPTY_ROOT(&inode->io_failure_tree))
return;
spin_lock(&inode->io_failure_lock);
node = rb_simple_search_first(&inode->io_failure_tree, start);
while (node) {
failrec = rb_entry(node, struct io_failure_record, rb_node);
if (failrec->bytenr > end)
break;
next = rb_next(node);
rb_erase(&failrec->rb_node, &inode->io_failure_tree);
kfree(failrec);
node = next;
}
spin_unlock(&inode->io_failure_lock);
}
static struct io_failure_record *btrfs_get_io_failure_record(struct inode *inode,
struct btrfs_bio *bbio,
unsigned int bio_offset)
{
struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
u64 start = bbio->file_offset + bio_offset;
struct io_failure_record *failrec;
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
const u32 sectorsize = fs_info->sectorsize;
int ret;
failrec = get_failrec(BTRFS_I(inode), start);
if (!IS_ERR(failrec)) {
btrfs_debug(fs_info,
"Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu",
failrec->logical, failrec->bytenr, failrec->len);
/*
* when data can be on disk more than twice, add to failrec here
* (e.g. with a list for failed_mirror) to make
* clean_io_failure() clean all those errors at once.
*/
ASSERT(failrec->this_mirror == bbio->mirror_num);
ASSERT(failrec->len == fs_info->sectorsize);
return failrec;
}
failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
if (!failrec)
return ERR_PTR(-ENOMEM);
RB_CLEAR_NODE(&failrec->rb_node);
failrec->bytenr = start;
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
failrec->len = sectorsize;
failrec->failed_mirror = bbio->mirror_num;
failrec->this_mirror = bbio->mirror_num;
btrfs: fix repair of compressed extents Currently the checksum of compressed extents is verified based on the compressed data and the lower btrfs_bio, but the actual repair process is driven by end_bio_extent_readpage on the upper btrfs_bio for the decompressed data. This has a bunch of issues, including not being able to properly communicate the failed mirror up in case that the I/O submission got preempted, a general loss of if an error was an I/O error or a checksum verification failure, but most importantly that this design causes btrfs_clean_io_failure to eventually write back the uncompressed good data onto the disk sectors that are supposed to contain compressed data. Fix this by moving the repair to the lower btrfs_bio. To do so, a fair amount of code has to be reshuffled: a) the lower btrfs_bio now needs a valid csum pointer. The easiest way to achieve that is to pass NULL btrfs_lookup_bio_sums and just use the btrfs_bio management of csums. For a compressed_bio that is split into multiple btrfs_bios this means additional memory allocations, but the code becomes a lot more regular. b) checksum verification now runs directly on the lower btrfs_bio instead of the compressed_bio. This actually nicely simplifies the end I/O processing. c) btrfs_repair_one_sector can't just look up the logical address for the file offset any more, as there is no corresponding relative offsets that apply to the file offset and the logic address for compressed extents. Instead require that the saved bvec_iter in the btrfs_bio is filled out for all read bios and use that, which again removes a fair amount of code. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: David Sterba <dsterba@suse.com>
2022-07-07 13:33:30 +08:00
failrec->logical = (bbio->iter.bi_sector << SECTOR_SHIFT) + bio_offset;
btrfs_debug(fs_info,
btrfs: fix repair of compressed extents Currently the checksum of compressed extents is verified based on the compressed data and the lower btrfs_bio, but the actual repair process is driven by end_bio_extent_readpage on the upper btrfs_bio for the decompressed data. This has a bunch of issues, including not being able to properly communicate the failed mirror up in case that the I/O submission got preempted, a general loss of if an error was an I/O error or a checksum verification failure, but most importantly that this design causes btrfs_clean_io_failure to eventually write back the uncompressed good data onto the disk sectors that are supposed to contain compressed data. Fix this by moving the repair to the lower btrfs_bio. To do so, a fair amount of code has to be reshuffled: a) the lower btrfs_bio now needs a valid csum pointer. The easiest way to achieve that is to pass NULL btrfs_lookup_bio_sums and just use the btrfs_bio management of csums. For a compressed_bio that is split into multiple btrfs_bios this means additional memory allocations, but the code becomes a lot more regular. b) checksum verification now runs directly on the lower btrfs_bio instead of the compressed_bio. This actually nicely simplifies the end I/O processing. c) btrfs_repair_one_sector can't just look up the logical address for the file offset any more, as there is no corresponding relative offsets that apply to the file offset and the logic address for compressed extents. Instead require that the saved bvec_iter in the btrfs_bio is filled out for all read bios and use that, which again removes a fair amount of code. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: David Sterba <dsterba@suse.com>
2022-07-07 13:33:30 +08:00
"new io failure record logical %llu start %llu",
failrec->logical, start);
btrfs: fix repair of compressed extents Currently the checksum of compressed extents is verified based on the compressed data and the lower btrfs_bio, but the actual repair process is driven by end_bio_extent_readpage on the upper btrfs_bio for the decompressed data. This has a bunch of issues, including not being able to properly communicate the failed mirror up in case that the I/O submission got preempted, a general loss of if an error was an I/O error or a checksum verification failure, but most importantly that this design causes btrfs_clean_io_failure to eventually write back the uncompressed good data onto the disk sectors that are supposed to contain compressed data. Fix this by moving the repair to the lower btrfs_bio. To do so, a fair amount of code has to be reshuffled: a) the lower btrfs_bio now needs a valid csum pointer. The easiest way to achieve that is to pass NULL btrfs_lookup_bio_sums and just use the btrfs_bio management of csums. For a compressed_bio that is split into multiple btrfs_bios this means additional memory allocations, but the code becomes a lot more regular. b) checksum verification now runs directly on the lower btrfs_bio instead of the compressed_bio. This actually nicely simplifies the end I/O processing. c) btrfs_repair_one_sector can't just look up the logical address for the file offset any more, as there is no corresponding relative offsets that apply to the file offset and the logic address for compressed extents. Instead require that the saved bvec_iter in the btrfs_bio is filled out for all read bios and use that, which again removes a fair amount of code. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: David Sterba <dsterba@suse.com>
2022-07-07 13:33:30 +08:00
failrec->num_copies = btrfs_num_copies(fs_info, failrec->logical, sectorsize);
if (failrec->num_copies == 1) {
/*
* We only have a single copy of the data, so don't bother with
* all the retry and error correction code that follows. No
* matter what the error is, it is very likely to persist.
*/
btrfs_debug(fs_info,
"cannot repair logical %llu num_copies %d",
failrec->logical, failrec->num_copies);
kfree(failrec);
return ERR_PTR(-EIO);
}
/* Set the bits in the private failure tree */
ret = insert_failrec(BTRFS_I(inode), failrec);
if (ret) {
kfree(failrec);
return ERR_PTR(ret);
}
return failrec;
}
int btrfs_repair_one_sector(struct btrfs_inode *inode, struct btrfs_bio *failed_bbio,
u32 bio_offset, struct page *page, unsigned int pgoff,
bool submit_buffered)
{
u64 start = failed_bbio->file_offset + bio_offset;
struct io_failure_record *failrec;
struct btrfs_fs_info *fs_info = inode->root->fs_info;
struct bio *failed_bio = &failed_bbio->bio;
const int icsum = bio_offset >> fs_info->sectorsize_bits;
struct bio *repair_bio;
struct btrfs_bio *repair_bbio;
btrfs_debug(fs_info,
"repair read error: read error at %llu", start);
BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
failrec = btrfs_get_io_failure_record(&inode->vfs_inode, failed_bbio, bio_offset);
if (IS_ERR(failrec))
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
return PTR_ERR(failrec);
/*
* There are two premises:
* a) deliver good data to the caller
* b) correct the bad sectors on disk
*
* Since we're only doing repair for one sector, we only need to get
* a good copy of the failed sector and if we succeed, we have setup
* everything for repair_io_failure to do the rest for us.
*/
failrec->this_mirror = next_mirror(failrec, failrec->this_mirror);
if (failrec->this_mirror == failrec->failed_mirror) {
btrfs_debug(fs_info,
"failed to repair num_copies %d this_mirror %d failed_mirror %d",
failrec->num_copies, failrec->this_mirror, failrec->failed_mirror);
free_io_failure(inode, failrec);
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
return -EIO;
}
repair_bio = btrfs_bio_alloc(1, REQ_OP_READ, failed_bbio->end_io,
failed_bbio->private);
repair_bbio = btrfs_bio(repair_bio);
repair_bbio->file_offset = start;
repair_bio->bi_iter.bi_sector = failrec->logical >> 9;
if (failed_bbio->csum) {
const u32 csum_size = fs_info->csum_size;
repair_bbio->csum = repair_bbio->csum_inline;
memcpy(repair_bbio->csum,
failed_bbio->csum + csum_size * icsum, csum_size);
}
bio_add_page(repair_bio, page, failrec->len, pgoff);
repair_bbio->iter = repair_bio->bi_iter;
btrfs_debug(fs_info,
"repair read error: submitting new read to mirror %d",
failrec->this_mirror);
/*
* At this point we have a bio, so any errors from bio submission will
* be handled by the endio on the repair_bio, so we can't return an
* error here.
*/
if (submit_buffered)
btrfs_submit_data_read_bio(inode, repair_bio,
failrec->this_mirror, 0);
else
btrfs_submit_dio_repair_bio(inode, repair_bio, failrec->this_mirror);
return BLK_STS_OK;
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
}
static void end_page_read(struct page *page, bool uptodate, u64 start, u32 len)
{
struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
ASSERT(page_offset(page) <= start &&
start + len <= page_offset(page) + PAGE_SIZE);
if (uptodate) {
btrfs: initial fsverity support Add support for fsverity in btrfs. To support the generic interface in fs/verity, we add two new item types in the fs tree for inodes with verity enabled. One stores the per-file verity descriptor and btrfs verity item and the other stores the Merkle tree data itself. Verity checking is done in end_page_read just before a page is marked uptodate. This naturally handles a variety of edge cases like holes, preallocated extents, and inline extents. Some care needs to be taken to not try to verity pages past the end of the file, which are accessed by the generic buffered file reading code under some circumstances like reading to the end of the last page and trying to read again. Direct IO on a verity file falls back to buffered reads. Verity relies on PageChecked for the Merkle tree data itself to avoid re-walking up shared paths in the tree. For this reason, we need to cache the Merkle tree data. Since the file is immutable after verity is turned on, we can cache it at an index past EOF. Use the new inode ro_flags to store verity on the inode item, so that we can enable verity on a file, then rollback to an older kernel and still mount the file system and read the file. Since we can't safely write the file anymore without ruining the invariants of the Merkle tree, we mark a ro_compat flag on the file system when a file has verity enabled. Acked-by: Eric Biggers <ebiggers@google.com> Co-developed-by: Chris Mason <clm@fb.com> Signed-off-by: Chris Mason <clm@fb.com> Signed-off-by: Boris Burkov <boris@bur.io> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-01 04:01:49 +08:00
if (fsverity_active(page->mapping->host) &&
!PageError(page) &&
!PageUptodate(page) &&
start < i_size_read(page->mapping->host) &&
!fsverity_verify_page(page)) {
btrfs_page_set_error(fs_info, page, start, len);
} else {
btrfs_page_set_uptodate(fs_info, page, start, len);
}
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
} else {
btrfs_page_clear_uptodate(fs_info, page, start, len);
btrfs_page_set_error(fs_info, page, start, len);
}
if (!btrfs_is_subpage(fs_info, page))
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
unlock_page(page);
btrfs: subpage: fix a rare race between metadata endio and eb freeing [BUG] There is a very rare ASSERT() triggering during full fstests run for subpage rw support. No other reproducer so far. The ASSERT() gets triggered for metadata read in btrfs_page_set_uptodate() inside end_page_read(). [CAUSE] There is still a small race window for metadata only, the race could happen like this: T1 | T2 ------------------------------------+----------------------------- end_bio_extent_readpage() | |- btrfs_validate_metadata_buffer() | | |- free_extent_buffer() | | Still have 2 refs | |- end_page_read() | |- if (unlikely(PagePrivate()) | | The page still has Private | | | free_extent_buffer() | | | Only one ref 1, will be | | | released | | |- detach_extent_buffer_page() | | |- btrfs_detach_subpage() |- btrfs_set_page_uptodate() | The page no longer has Private| >>> ASSERT() triggered <<< | This race window is super small, thus pretty hard to hit, even with so many runs of fstests. But the race window is still there, we have to go another way to solve it other than relying on random PagePrivate() check. Data path is not affected, as it will lock the page before reading, while unlocking the page after the last read has finished, thus no race window. [FIX] This patch will fix the bug by repurposing btrfs_subpage::readers. Now btrfs_subpage::readers will be a member shared by both metadata and data. For metadata path, we don't do the page unlock as metadata only relies on extent locking. At the same time, teach page_range_has_eb() to take btrfs_subpage::readers into consideration. So that even if the last eb of a page gets freed, page::private won't be detached as long as there still are pending end_page_read() calls. By this we eliminate the race window, this will slight increase the metadata memory usage, as the page may not be released as frequently as usual. But it should not be a big deal. The code got introduced in ("btrfs: submit read time repair only for each corrupted sector"), but the fix is in a separate patch to keep the problem description and the crash is rare so it should not hurt bisectability. Signed-off-by: Qu Wegruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-06-07 17:02:58 +08:00
else
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
btrfs_subpage_end_reader(fs_info, page, start, len);
}
static void end_sector_io(struct page *page, u64 offset, bool uptodate)
{
struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
const u32 sectorsize = inode->root->fs_info->sectorsize;
struct extent_state *cached = NULL;
end_page_read(page, uptodate, offset, sectorsize);
if (uptodate)
set_extent_uptodate(&inode->io_tree, offset,
offset + sectorsize - 1, &cached, GFP_NOFS);
unlock_extent(&inode->io_tree, offset, offset + sectorsize - 1,
&cached);
}
static void submit_data_read_repair(struct inode *inode,
struct btrfs_bio *failed_bbio,
u32 bio_offset, const struct bio_vec *bvec,
unsigned int error_bitmap)
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
{
const unsigned int pgoff = bvec->bv_offset;
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
struct page *page = bvec->bv_page;
const u64 start = page_offset(bvec->bv_page) + bvec->bv_offset;
const u64 end = start + bvec->bv_len - 1;
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
const u32 sectorsize = fs_info->sectorsize;
const int nr_bits = (end + 1 - start) >> fs_info->sectorsize_bits;
int i;
BUG_ON(bio_op(&failed_bbio->bio) == REQ_OP_WRITE);
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
/* This repair is only for data */
ASSERT(is_data_inode(inode));
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
/* We're here because we had some read errors or csum mismatch */
ASSERT(error_bitmap);
/*
* We only get called on buffered IO, thus page must be mapped and bio
* must not be cloned.
*/
ASSERT(page->mapping && !bio_flagged(&failed_bbio->bio, BIO_CLONED));
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
/* Iterate through all the sectors in the range */
for (i = 0; i < nr_bits; i++) {
const unsigned int offset = i * sectorsize;
bool uptodate = false;
int ret;
if (!(error_bitmap & (1U << i))) {
/*
* This sector has no error, just end the page read
* and unlock the range.
*/
uptodate = true;
goto next;
}
ret = btrfs_repair_one_sector(BTRFS_I(inode), failed_bbio,
bio_offset + offset, page, pgoff + offset,
true);
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
if (!ret) {
/*
* We have submitted the read repair, the page release
* will be handled by the endio function of the
* submitted repair bio.
* Thus we don't need to do any thing here.
*/
continue;
}
/*
* Continue on failed repair, otherwise the remaining sectors
* will not be properly unlocked.
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
*/
next:
end_sector_io(page, start + offset, uptodate);
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
}
}
/* lots and lots of room for performance fixes in the end_bio funcs */
void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
{
struct btrfs_inode *inode;
const bool uptodate = (err == 0);
int ret = 0;
ASSERT(page && page->mapping);
inode = BTRFS_I(page->mapping->host);
btrfs_writepage_endio_finish_ordered(inode, page, start, end, uptodate);
if (!uptodate) {
btrfs: unify regular and subpage error paths in __extent_writepage() [BUG] When running btrfs/160 in a loop for subpage with experimental compression support, it has a high chance to crash (~20%): BTRFS critical (device dm-7): panic in __btrfs_add_ordered_extent:238: inconsistency in ordered tree at offset 0 (errno=-17 Object already exists) ------------[ cut here ]------------ kernel BUG at fs/btrfs/ordered-data.c:238! Internal error: Oops - BUG: 0 [#1] SMP pc : __btrfs_add_ordered_extent+0x550/0x670 [btrfs] lr : __btrfs_add_ordered_extent+0x550/0x670 [btrfs] Call trace: __btrfs_add_ordered_extent+0x550/0x670 [btrfs] btrfs_add_ordered_extent+0x2c/0x50 [btrfs] run_delalloc_nocow+0x81c/0x8fc [btrfs] btrfs_run_delalloc_range+0xa4/0x390 [btrfs] writepage_delalloc+0xc0/0x1ac [btrfs] __extent_writepage+0xf4/0x370 [btrfs] extent_write_cache_pages+0x288/0x4f4 [btrfs] extent_writepages+0x58/0xe0 [btrfs] btrfs_writepages+0x1c/0x30 [btrfs] do_writepages+0x60/0x110 __filemap_fdatawrite_range+0x108/0x170 filemap_fdatawrite_range+0x20/0x30 btrfs_fdatawrite_range+0x34/0x4dc [btrfs] __btrfs_write_out_cache+0x34c/0x480 [btrfs] btrfs_write_out_cache+0x144/0x220 [btrfs] btrfs_start_dirty_block_groups+0x3ac/0x6b0 [btrfs] btrfs_commit_transaction+0xd0/0xbb4 [btrfs] btrfs_sync_fs+0x64/0x1cc [btrfs] sync_fs_one_sb+0x3c/0x50 iterate_supers+0xcc/0x1d4 ksys_sync+0x6c/0xd0 __arm64_sys_sync+0x1c/0x30 invoke_syscall+0x50/0x120 el0_svc_common.constprop.0+0x4c/0xd4 do_el0_svc+0x30/0x9c el0_svc+0x2c/0x54 el0_sync_handler+0x1a8/0x1b0 el0_sync+0x198/0x1c0 ---[ end trace 336f67369ae6e0af ]--- [CAUSE] For subpage case, we can have multiple sectors inside a page, this makes it possible for __extent_writepage() to have part of its page submitted before returning. In btrfs/160, we are using dm-dust to emulate write error, this means for certain pages, we could have everything running fine, but at the end of __extent_writepage(), one of the submitted bios fails due to dm-dust. Then the page is marked Error, and we change @ret from 0 to -EIO. This makes the caller extent_write_cache_pages() to error out, without submitting the remaining pages. Furthermore, since we're erroring out for free space cache, it doesn't really care about the error and will update the inode and retry the writeback. Then we re-run the delalloc range, and will try to insert the same delalloc range while previous delalloc range is still hanging there, triggering the above error. [FIX] The proper fix is to handle errors from __extent_writepage() properly, by ending the remaining ordered extent. But that fix needs the following changes: - Know at exactly which sector the error happened Currently __extent_writepage_io() works for the full page, can't return at which sector we hit the error. - Grab the ordered extent covering the failed sector As a hotfix for subpage case, here we unify the error paths in __extent_writepage(). In fact, the "if (PageError(page))" branch never get executed if @ret is still 0 for non-subpage cases. As for non-subpage case, we never submit current page in __extent_writepage(), but only add current page into bio. The bio can only get submitted in next page. Thus we never get PageError() set due to IO failure, thus when we hit the branch, @ret is never 0. By simply removing that @ret assignment, we let subpage case ignore the IO failure, thus only error out for fatal errors just like regular sectorsize. So that IO error won't be treated as fatal error not trigger the hanging OE problem. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-26 14:35:07 +08:00
const struct btrfs_fs_info *fs_info = inode->root->fs_info;
u32 len;
ASSERT(end + 1 - start <= U32_MAX);
len = end + 1 - start;
btrfs_page_clear_uptodate(fs_info, page, start, len);
btrfs_page_set_error(fs_info, page, start, len);
ret = err < 0 ? err : -EIO;
mapping_set_error(page->mapping, ret);
}
}
/*
* after a writepage IO is done, we need to:
* clear the uptodate bits on error
* clear the writeback bits in the extent tree for this IO
* end_page_writeback if the page has no more pending IO
*
* Scheduling is not allowed, so the extent state tree is expected
* to have one and only one object corresponding to this IO.
*/
static void end_bio_extent_writepage(struct btrfs_bio *bbio)
{
struct bio *bio = &bbio->bio;
int error = blk_status_to_errno(bio->bi_status);
struct bio_vec *bvec;
u64 start;
u64 end;
struct bvec_iter_all iter_all;
bool first_bvec = true;
ASSERT(!bio_flagged(bio, BIO_CLONED));
bio_for_each_segment_all(bvec, bio, iter_all) {
struct page *page = bvec->bv_page;
struct inode *inode = page->mapping->host;
struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
const u32 sectorsize = fs_info->sectorsize;
/* Our read/write should always be sector aligned. */
if (!IS_ALIGNED(bvec->bv_offset, sectorsize))
btrfs_err(fs_info,
"partial page write in btrfs with offset %u and length %u",
bvec->bv_offset, bvec->bv_len);
else if (!IS_ALIGNED(bvec->bv_len, sectorsize))
btrfs_info(fs_info,
"incomplete page write with offset %u and length %u",
bvec->bv_offset, bvec->bv_len);
start = page_offset(page) + bvec->bv_offset;
end = start + bvec->bv_len - 1;
if (first_bvec) {
btrfs_record_physical_zoned(inode, start, bio);
first_bvec = false;
}
end_extent_writepage(page, error, start, end);
btrfs_page_clear_writeback(fs_info, page, start, bvec->bv_len);
}
bio_put(bio);
}
btrfs: add structure to keep track of extent range in end_bio_extent_readpage In end_bio_extent_readpage() we had a strange dance around extent_start/extent_len. Hidden behind the strange dance is, it's just calling endio_readpage_release_extent() on each bvec range. Here is an example to explain the original work flow: Bio is for inode 257, containing 2 pages, for range [1M, 1M+8K) end_bio_extent_extent_readpage() entered |- extent_start = 0; |- extent_end = 0; |- bio_for_each_segment_all() { | |- /* Got the 1st bvec */ | |- start = SZ_1M; | |- end = SZ_1M + SZ_4K - 1; | |- update = 1; | |- if (extent_len == 0) { | | |- extent_start = start; /* SZ_1M */ | | |- extent_len = end + 1 - start; /* SZ_1M */ | | } | | | |- /* Got the 2nd bvec */ | |- start = SZ_1M + 4K; | |- end = SZ_1M + 4K - 1; | |- update = 1; | |- if (extent_start + extent_len == start) { | | |- extent_len += end + 1 - start; /* SZ_8K */ | | } | } /* All bio vec iterated */ | |- if (extent_len) { |- endio_readpage_release_extent(tree, extent_start, extent_len, update); /* extent_start == SZ_1M, extent_len == SZ_8K, uptodate = 1 */ As the above flow shows, the existing code in end_bio_extent_readpage() is accumulates extent_start/extent_len, and when the contiguous range stops, calls endio_readpage_release_extent() for the range. However current behavior has something not really considered: - The inode can change For bio, its pages don't need to have contiguous page_offset. This means, even pages from different inodes can be packed into one bio. - bvec cross page boundary There is a feature called multi-page bvec, where bvec->bv_len can go beyond bvec->bv_page boundary. - Poor readability This patch will address the problem: - Introduce a proper structure, processed_extent, to record processed extent range - Integrate inode/start/end/uptodate check into endio_readpage_release_extent() - Add more comment on each step. This should greatly improve the readability, now in end_bio_extent_readpage() there are only two endio_readpage_release_extent() calls. - Add inode check for contiguity Now we also ensure the inode is the same one before checking if the range is contiguous. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-13 20:51:28 +08:00
/*
* Record previously processed extent range
*
* For endio_readpage_release_extent() to handle a full extent range, reducing
* the extent io operations.
*/
struct processed_extent {
struct btrfs_inode *inode;
/* Start of the range in @inode */
u64 start;
/* End of the range in @inode */
btrfs: add structure to keep track of extent range in end_bio_extent_readpage In end_bio_extent_readpage() we had a strange dance around extent_start/extent_len. Hidden behind the strange dance is, it's just calling endio_readpage_release_extent() on each bvec range. Here is an example to explain the original work flow: Bio is for inode 257, containing 2 pages, for range [1M, 1M+8K) end_bio_extent_extent_readpage() entered |- extent_start = 0; |- extent_end = 0; |- bio_for_each_segment_all() { | |- /* Got the 1st bvec */ | |- start = SZ_1M; | |- end = SZ_1M + SZ_4K - 1; | |- update = 1; | |- if (extent_len == 0) { | | |- extent_start = start; /* SZ_1M */ | | |- extent_len = end + 1 - start; /* SZ_1M */ | | } | | | |- /* Got the 2nd bvec */ | |- start = SZ_1M + 4K; | |- end = SZ_1M + 4K - 1; | |- update = 1; | |- if (extent_start + extent_len == start) { | | |- extent_len += end + 1 - start; /* SZ_8K */ | | } | } /* All bio vec iterated */ | |- if (extent_len) { |- endio_readpage_release_extent(tree, extent_start, extent_len, update); /* extent_start == SZ_1M, extent_len == SZ_8K, uptodate = 1 */ As the above flow shows, the existing code in end_bio_extent_readpage() is accumulates extent_start/extent_len, and when the contiguous range stops, calls endio_readpage_release_extent() for the range. However current behavior has something not really considered: - The inode can change For bio, its pages don't need to have contiguous page_offset. This means, even pages from different inodes can be packed into one bio. - bvec cross page boundary There is a feature called multi-page bvec, where bvec->bv_len can go beyond bvec->bv_page boundary. - Poor readability This patch will address the problem: - Introduce a proper structure, processed_extent, to record processed extent range - Integrate inode/start/end/uptodate check into endio_readpage_release_extent() - Add more comment on each step. This should greatly improve the readability, now in end_bio_extent_readpage() there are only two endio_readpage_release_extent() calls. - Add inode check for contiguity Now we also ensure the inode is the same one before checking if the range is contiguous. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-13 20:51:28 +08:00
u64 end;
bool uptodate;
};
/*
* Try to release processed extent range
*
* May not release the extent range right now if the current range is
* contiguous to processed extent.
*
* Will release processed extent when any of @inode, @uptodate, the range is
* no longer contiguous to the processed range.
*
* Passing @inode == NULL will force processed extent to be released.
*/
static void endio_readpage_release_extent(struct processed_extent *processed,
struct btrfs_inode *inode, u64 start, u64 end,
bool uptodate)
{
struct extent_state *cached = NULL;
btrfs: add structure to keep track of extent range in end_bio_extent_readpage In end_bio_extent_readpage() we had a strange dance around extent_start/extent_len. Hidden behind the strange dance is, it's just calling endio_readpage_release_extent() on each bvec range. Here is an example to explain the original work flow: Bio is for inode 257, containing 2 pages, for range [1M, 1M+8K) end_bio_extent_extent_readpage() entered |- extent_start = 0; |- extent_end = 0; |- bio_for_each_segment_all() { | |- /* Got the 1st bvec */ | |- start = SZ_1M; | |- end = SZ_1M + SZ_4K - 1; | |- update = 1; | |- if (extent_len == 0) { | | |- extent_start = start; /* SZ_1M */ | | |- extent_len = end + 1 - start; /* SZ_1M */ | | } | | | |- /* Got the 2nd bvec */ | |- start = SZ_1M + 4K; | |- end = SZ_1M + 4K - 1; | |- update = 1; | |- if (extent_start + extent_len == start) { | | |- extent_len += end + 1 - start; /* SZ_8K */ | | } | } /* All bio vec iterated */ | |- if (extent_len) { |- endio_readpage_release_extent(tree, extent_start, extent_len, update); /* extent_start == SZ_1M, extent_len == SZ_8K, uptodate = 1 */ As the above flow shows, the existing code in end_bio_extent_readpage() is accumulates extent_start/extent_len, and when the contiguous range stops, calls endio_readpage_release_extent() for the range. However current behavior has something not really considered: - The inode can change For bio, its pages don't need to have contiguous page_offset. This means, even pages from different inodes can be packed into one bio. - bvec cross page boundary There is a feature called multi-page bvec, where bvec->bv_len can go beyond bvec->bv_page boundary. - Poor readability This patch will address the problem: - Introduce a proper structure, processed_extent, to record processed extent range - Integrate inode/start/end/uptodate check into endio_readpage_release_extent() - Add more comment on each step. This should greatly improve the readability, now in end_bio_extent_readpage() there are only two endio_readpage_release_extent() calls. - Add inode check for contiguity Now we also ensure the inode is the same one before checking if the range is contiguous. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-13 20:51:28 +08:00
struct extent_io_tree *tree;
/* The first extent, initialize @processed */
if (!processed->inode)
goto update;
btrfs: add structure to keep track of extent range in end_bio_extent_readpage In end_bio_extent_readpage() we had a strange dance around extent_start/extent_len. Hidden behind the strange dance is, it's just calling endio_readpage_release_extent() on each bvec range. Here is an example to explain the original work flow: Bio is for inode 257, containing 2 pages, for range [1M, 1M+8K) end_bio_extent_extent_readpage() entered |- extent_start = 0; |- extent_end = 0; |- bio_for_each_segment_all() { | |- /* Got the 1st bvec */ | |- start = SZ_1M; | |- end = SZ_1M + SZ_4K - 1; | |- update = 1; | |- if (extent_len == 0) { | | |- extent_start = start; /* SZ_1M */ | | |- extent_len = end + 1 - start; /* SZ_1M */ | | } | | | |- /* Got the 2nd bvec */ | |- start = SZ_1M + 4K; | |- end = SZ_1M + 4K - 1; | |- update = 1; | |- if (extent_start + extent_len == start) { | | |- extent_len += end + 1 - start; /* SZ_8K */ | | } | } /* All bio vec iterated */ | |- if (extent_len) { |- endio_readpage_release_extent(tree, extent_start, extent_len, update); /* extent_start == SZ_1M, extent_len == SZ_8K, uptodate = 1 */ As the above flow shows, the existing code in end_bio_extent_readpage() is accumulates extent_start/extent_len, and when the contiguous range stops, calls endio_readpage_release_extent() for the range. However current behavior has something not really considered: - The inode can change For bio, its pages don't need to have contiguous page_offset. This means, even pages from different inodes can be packed into one bio. - bvec cross page boundary There is a feature called multi-page bvec, where bvec->bv_len can go beyond bvec->bv_page boundary. - Poor readability This patch will address the problem: - Introduce a proper structure, processed_extent, to record processed extent range - Integrate inode/start/end/uptodate check into endio_readpage_release_extent() - Add more comment on each step. This should greatly improve the readability, now in end_bio_extent_readpage() there are only two endio_readpage_release_extent() calls. - Add inode check for contiguity Now we also ensure the inode is the same one before checking if the range is contiguous. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-13 20:51:28 +08:00
/*
* Contiguous to processed extent, just uptodate the end.
*
* Several things to notice:
*
* - bio can be merged as long as on-disk bytenr is contiguous
* This means we can have page belonging to other inodes, thus need to
* check if the inode still matches.
* - bvec can contain range beyond current page for multi-page bvec
* Thus we need to do processed->end + 1 >= start check
*/
if (processed->inode == inode && processed->uptodate == uptodate &&
processed->end + 1 >= start && end >= processed->end) {
processed->end = end;
return;
}
tree = &processed->inode->io_tree;
/*
* Now we don't have range contiguous to the processed range, release
* the processed range now.
*/
unlock_extent(tree, processed->start, processed->end, &cached);
btrfs: add structure to keep track of extent range in end_bio_extent_readpage In end_bio_extent_readpage() we had a strange dance around extent_start/extent_len. Hidden behind the strange dance is, it's just calling endio_readpage_release_extent() on each bvec range. Here is an example to explain the original work flow: Bio is for inode 257, containing 2 pages, for range [1M, 1M+8K) end_bio_extent_extent_readpage() entered |- extent_start = 0; |- extent_end = 0; |- bio_for_each_segment_all() { | |- /* Got the 1st bvec */ | |- start = SZ_1M; | |- end = SZ_1M + SZ_4K - 1; | |- update = 1; | |- if (extent_len == 0) { | | |- extent_start = start; /* SZ_1M */ | | |- extent_len = end + 1 - start; /* SZ_1M */ | | } | | | |- /* Got the 2nd bvec */ | |- start = SZ_1M + 4K; | |- end = SZ_1M + 4K - 1; | |- update = 1; | |- if (extent_start + extent_len == start) { | | |- extent_len += end + 1 - start; /* SZ_8K */ | | } | } /* All bio vec iterated */ | |- if (extent_len) { |- endio_readpage_release_extent(tree, extent_start, extent_len, update); /* extent_start == SZ_1M, extent_len == SZ_8K, uptodate = 1 */ As the above flow shows, the existing code in end_bio_extent_readpage() is accumulates extent_start/extent_len, and when the contiguous range stops, calls endio_readpage_release_extent() for the range. However current behavior has something not really considered: - The inode can change For bio, its pages don't need to have contiguous page_offset. This means, even pages from different inodes can be packed into one bio. - bvec cross page boundary There is a feature called multi-page bvec, where bvec->bv_len can go beyond bvec->bv_page boundary. - Poor readability This patch will address the problem: - Introduce a proper structure, processed_extent, to record processed extent range - Integrate inode/start/end/uptodate check into endio_readpage_release_extent() - Add more comment on each step. This should greatly improve the readability, now in end_bio_extent_readpage() there are only two endio_readpage_release_extent() calls. - Add inode check for contiguity Now we also ensure the inode is the same one before checking if the range is contiguous. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-13 20:51:28 +08:00
update:
/* Update processed to current range */
processed->inode = inode;
processed->start = start;
processed->end = end;
processed->uptodate = uptodate;
}
btrfs: integrate page status update for data read path into begin/end_page_read In btrfs data page read path, the page status update are handled in two different locations: btrfs_do_read_page() { while (cur <= end) { /* No need to read from disk */ if (HOLE/PREALLOC/INLINE){ memset(); set_extent_uptodate(); continue; } /* Read from disk */ ret = submit_extent_page(end_bio_extent_readpage); } end_bio_extent_readpage() { endio_readpage_uptodate_page_status(); } This is fine for sectorsize == PAGE_SIZE case, as for above loop we should only hit one branch and then exit. But for subpage, there is more work to be done in page status update: - Page Unlock condition Unlike regular page size == sectorsize case, we can no longer just unlock a page. Only the last reader of the page can unlock the page. This means, we can unlock the page either in the while() loop, or in the endio function. - Page uptodate condition Since we have multiple sectors to read for a page, we can only mark the full page uptodate if all sectors are uptodate. To handle both subpage and regular cases, introduce a pair of functions to help handling page status update: - begin_page_read() For regular case, it does nothing. For subpage case, it updates the reader counters so that later end_page_read() can know who is the last one to unlock the page. - end_page_read() This is just endio_readpage_uptodate_page_status() renamed. The original name is a little too long and too specific for endio. The new thing added is the condition for page unlock. Now for subpage data, we unlock the page if we're the last reader. This does not only provide the basis for subpage data read, but also hide the special handling of page read from the main read loop. Also, since we're changing how the page lock is handled, there are two existing error paths where we need to manually unlock the page before calling begin_page_read(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-02 10:28:36 +08:00
static void begin_page_read(struct btrfs_fs_info *fs_info, struct page *page)
{
ASSERT(PageLocked(page));
if (!btrfs_is_subpage(fs_info, page))
btrfs: integrate page status update for data read path into begin/end_page_read In btrfs data page read path, the page status update are handled in two different locations: btrfs_do_read_page() { while (cur <= end) { /* No need to read from disk */ if (HOLE/PREALLOC/INLINE){ memset(); set_extent_uptodate(); continue; } /* Read from disk */ ret = submit_extent_page(end_bio_extent_readpage); } end_bio_extent_readpage() { endio_readpage_uptodate_page_status(); } This is fine for sectorsize == PAGE_SIZE case, as for above loop we should only hit one branch and then exit. But for subpage, there is more work to be done in page status update: - Page Unlock condition Unlike regular page size == sectorsize case, we can no longer just unlock a page. Only the last reader of the page can unlock the page. This means, we can unlock the page either in the while() loop, or in the endio function. - Page uptodate condition Since we have multiple sectors to read for a page, we can only mark the full page uptodate if all sectors are uptodate. To handle both subpage and regular cases, introduce a pair of functions to help handling page status update: - begin_page_read() For regular case, it does nothing. For subpage case, it updates the reader counters so that later end_page_read() can know who is the last one to unlock the page. - end_page_read() This is just endio_readpage_uptodate_page_status() renamed. The original name is a little too long and too specific for endio. The new thing added is the condition for page unlock. Now for subpage data, we unlock the page if we're the last reader. This does not only provide the basis for subpage data read, but also hide the special handling of page read from the main read loop. Also, since we're changing how the page lock is handled, there are two existing error paths where we need to manually unlock the page before calling begin_page_read(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-02 10:28:36 +08:00
return;
ASSERT(PagePrivate(page));
btrfs_subpage_start_reader(fs_info, page, page_offset(page), PAGE_SIZE);
}
btrfs: subpage: fix wild pointer access during metadata read failure [BUG] When running fstests for btrfs subpage read-write test, it has a very high chance to crash at generic/475 with the following stack: BTRFS warning (device dm-8): direct IO failed ino 510 rw 1,34817 sector 0xcdf0 len 94208 err no 10 Unable to handle kernel paging request at virtual address ffff80001157e7c0 CPU: 2 PID: 687125 Comm: kworker/u12:4 Tainted: G WC 5.12.0-rc2-custom+ #5 Hardware name: Khadas VIM3 (DT) Workqueue: btrfs-endio-meta btrfs_work_helper [btrfs] pc : queued_spin_lock_slowpath+0x1a0/0x390 lr : do_raw_spin_lock+0xc4/0x11c Call trace: queued_spin_lock_slowpath+0x1a0/0x390 _raw_spin_lock+0x68/0x84 btree_readahead_hook+0x38/0xc0 [btrfs] end_bio_extent_readpage+0x504/0x5f4 [btrfs] bio_endio+0x170/0x1a4 end_workqueue_fn+0x3c/0x60 [btrfs] btrfs_work_helper+0x1b0/0x1b4 [btrfs] process_one_work+0x22c/0x430 worker_thread+0x70/0x3a0 kthread+0x13c/0x140 ret_from_fork+0x10/0x30 Code: 910020e0 8b0200c2 f861d884 aa0203e1 (f8246827) [CAUSE] In end_bio_extent_readpage(), if we hit an error during read, we will handle the error differently for data and metadata. For data we queue a repair, while for metadata, we record the error and let the caller choose what to do. But the code is still using page->private to grab extent buffer, which no longer points to extent buffer for subpage metadata pages. Thus this wild pointer access leads to above crash. [FIX] Introduce a helper, find_extent_buffer_readpage(), to grab extent buffer. The difference against find_extent_buffer_nospinlock() is: - Also handles regular sectorsize == PAGE_SIZE case - No extent buffer refs increase/decrease As extent buffer under IO must have non-zero refs, so this is safe Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-15 13:39:14 +08:00
/*
* Find extent buffer for a givne bytenr.
btrfs: subpage: fix wild pointer access during metadata read failure [BUG] When running fstests for btrfs subpage read-write test, it has a very high chance to crash at generic/475 with the following stack: BTRFS warning (device dm-8): direct IO failed ino 510 rw 1,34817 sector 0xcdf0 len 94208 err no 10 Unable to handle kernel paging request at virtual address ffff80001157e7c0 CPU: 2 PID: 687125 Comm: kworker/u12:4 Tainted: G WC 5.12.0-rc2-custom+ #5 Hardware name: Khadas VIM3 (DT) Workqueue: btrfs-endio-meta btrfs_work_helper [btrfs] pc : queued_spin_lock_slowpath+0x1a0/0x390 lr : do_raw_spin_lock+0xc4/0x11c Call trace: queued_spin_lock_slowpath+0x1a0/0x390 _raw_spin_lock+0x68/0x84 btree_readahead_hook+0x38/0xc0 [btrfs] end_bio_extent_readpage+0x504/0x5f4 [btrfs] bio_endio+0x170/0x1a4 end_workqueue_fn+0x3c/0x60 [btrfs] btrfs_work_helper+0x1b0/0x1b4 [btrfs] process_one_work+0x22c/0x430 worker_thread+0x70/0x3a0 kthread+0x13c/0x140 ret_from_fork+0x10/0x30 Code: 910020e0 8b0200c2 f861d884 aa0203e1 (f8246827) [CAUSE] In end_bio_extent_readpage(), if we hit an error during read, we will handle the error differently for data and metadata. For data we queue a repair, while for metadata, we record the error and let the caller choose what to do. But the code is still using page->private to grab extent buffer, which no longer points to extent buffer for subpage metadata pages. Thus this wild pointer access leads to above crash. [FIX] Introduce a helper, find_extent_buffer_readpage(), to grab extent buffer. The difference against find_extent_buffer_nospinlock() is: - Also handles regular sectorsize == PAGE_SIZE case - No extent buffer refs increase/decrease As extent buffer under IO must have non-zero refs, so this is safe Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-15 13:39:14 +08:00
*
* This is for end_bio_extent_readpage(), thus we can't do any unsafe locking
* in endio context.
*/
static struct extent_buffer *find_extent_buffer_readpage(
struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
{
struct extent_buffer *eb;
/*
* For regular sectorsize, we can use page->private to grab extent
* buffer
*/
if (fs_info->nodesize >= PAGE_SIZE) {
btrfs: subpage: fix wild pointer access during metadata read failure [BUG] When running fstests for btrfs subpage read-write test, it has a very high chance to crash at generic/475 with the following stack: BTRFS warning (device dm-8): direct IO failed ino 510 rw 1,34817 sector 0xcdf0 len 94208 err no 10 Unable to handle kernel paging request at virtual address ffff80001157e7c0 CPU: 2 PID: 687125 Comm: kworker/u12:4 Tainted: G WC 5.12.0-rc2-custom+ #5 Hardware name: Khadas VIM3 (DT) Workqueue: btrfs-endio-meta btrfs_work_helper [btrfs] pc : queued_spin_lock_slowpath+0x1a0/0x390 lr : do_raw_spin_lock+0xc4/0x11c Call trace: queued_spin_lock_slowpath+0x1a0/0x390 _raw_spin_lock+0x68/0x84 btree_readahead_hook+0x38/0xc0 [btrfs] end_bio_extent_readpage+0x504/0x5f4 [btrfs] bio_endio+0x170/0x1a4 end_workqueue_fn+0x3c/0x60 [btrfs] btrfs_work_helper+0x1b0/0x1b4 [btrfs] process_one_work+0x22c/0x430 worker_thread+0x70/0x3a0 kthread+0x13c/0x140 ret_from_fork+0x10/0x30 Code: 910020e0 8b0200c2 f861d884 aa0203e1 (f8246827) [CAUSE] In end_bio_extent_readpage(), if we hit an error during read, we will handle the error differently for data and metadata. For data we queue a repair, while for metadata, we record the error and let the caller choose what to do. But the code is still using page->private to grab extent buffer, which no longer points to extent buffer for subpage metadata pages. Thus this wild pointer access leads to above crash. [FIX] Introduce a helper, find_extent_buffer_readpage(), to grab extent buffer. The difference against find_extent_buffer_nospinlock() is: - Also handles regular sectorsize == PAGE_SIZE case - No extent buffer refs increase/decrease As extent buffer under IO must have non-zero refs, so this is safe Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-15 13:39:14 +08:00
ASSERT(PagePrivate(page) && page->private);
return (struct extent_buffer *)page->private;
}
/* For subpage case, we need to lookup buffer radix tree */
rcu_read_lock();
eb = radix_tree_lookup(&fs_info->buffer_radix,
bytenr >> fs_info->sectorsize_bits);
rcu_read_unlock();
btrfs: subpage: fix wild pointer access during metadata read failure [BUG] When running fstests for btrfs subpage read-write test, it has a very high chance to crash at generic/475 with the following stack: BTRFS warning (device dm-8): direct IO failed ino 510 rw 1,34817 sector 0xcdf0 len 94208 err no 10 Unable to handle kernel paging request at virtual address ffff80001157e7c0 CPU: 2 PID: 687125 Comm: kworker/u12:4 Tainted: G WC 5.12.0-rc2-custom+ #5 Hardware name: Khadas VIM3 (DT) Workqueue: btrfs-endio-meta btrfs_work_helper [btrfs] pc : queued_spin_lock_slowpath+0x1a0/0x390 lr : do_raw_spin_lock+0xc4/0x11c Call trace: queued_spin_lock_slowpath+0x1a0/0x390 _raw_spin_lock+0x68/0x84 btree_readahead_hook+0x38/0xc0 [btrfs] end_bio_extent_readpage+0x504/0x5f4 [btrfs] bio_endio+0x170/0x1a4 end_workqueue_fn+0x3c/0x60 [btrfs] btrfs_work_helper+0x1b0/0x1b4 [btrfs] process_one_work+0x22c/0x430 worker_thread+0x70/0x3a0 kthread+0x13c/0x140 ret_from_fork+0x10/0x30 Code: 910020e0 8b0200c2 f861d884 aa0203e1 (f8246827) [CAUSE] In end_bio_extent_readpage(), if we hit an error during read, we will handle the error differently for data and metadata. For data we queue a repair, while for metadata, we record the error and let the caller choose what to do. But the code is still using page->private to grab extent buffer, which no longer points to extent buffer for subpage metadata pages. Thus this wild pointer access leads to above crash. [FIX] Introduce a helper, find_extent_buffer_readpage(), to grab extent buffer. The difference against find_extent_buffer_nospinlock() is: - Also handles regular sectorsize == PAGE_SIZE case - No extent buffer refs increase/decrease As extent buffer under IO must have non-zero refs, so this is safe Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-15 13:39:14 +08:00
ASSERT(eb);
return eb;
}
/*
* after a readpage IO is done, we need to:
* clear the uptodate bits on error
* set the uptodate bits if things worked
* set the page up to date if all extents in the tree are uptodate
* clear the lock bit in the extent tree
* unlock the page if there are no other extents locked for it
*
* Scheduling is not allowed, so the extent state tree is expected
* to have one and only one object corresponding to this IO.
*/
static void end_bio_extent_readpage(struct btrfs_bio *bbio)
{
struct bio *bio = &bbio->bio;
struct bio_vec *bvec;
btrfs: add structure to keep track of extent range in end_bio_extent_readpage In end_bio_extent_readpage() we had a strange dance around extent_start/extent_len. Hidden behind the strange dance is, it's just calling endio_readpage_release_extent() on each bvec range. Here is an example to explain the original work flow: Bio is for inode 257, containing 2 pages, for range [1M, 1M+8K) end_bio_extent_extent_readpage() entered |- extent_start = 0; |- extent_end = 0; |- bio_for_each_segment_all() { | |- /* Got the 1st bvec */ | |- start = SZ_1M; | |- end = SZ_1M + SZ_4K - 1; | |- update = 1; | |- if (extent_len == 0) { | | |- extent_start = start; /* SZ_1M */ | | |- extent_len = end + 1 - start; /* SZ_1M */ | | } | | | |- /* Got the 2nd bvec */ | |- start = SZ_1M + 4K; | |- end = SZ_1M + 4K - 1; | |- update = 1; | |- if (extent_start + extent_len == start) { | | |- extent_len += end + 1 - start; /* SZ_8K */ | | } | } /* All bio vec iterated */ | |- if (extent_len) { |- endio_readpage_release_extent(tree, extent_start, extent_len, update); /* extent_start == SZ_1M, extent_len == SZ_8K, uptodate = 1 */ As the above flow shows, the existing code in end_bio_extent_readpage() is accumulates extent_start/extent_len, and when the contiguous range stops, calls endio_readpage_release_extent() for the range. However current behavior has something not really considered: - The inode can change For bio, its pages don't need to have contiguous page_offset. This means, even pages from different inodes can be packed into one bio. - bvec cross page boundary There is a feature called multi-page bvec, where bvec->bv_len can go beyond bvec->bv_page boundary. - Poor readability This patch will address the problem: - Introduce a proper structure, processed_extent, to record processed extent range - Integrate inode/start/end/uptodate check into endio_readpage_release_extent() - Add more comment on each step. This should greatly improve the readability, now in end_bio_extent_readpage() there are only two endio_readpage_release_extent() calls. - Add inode check for contiguity Now we also ensure the inode is the same one before checking if the range is contiguous. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-13 20:51:28 +08:00
struct processed_extent processed = { 0 };
/*
* The offset to the beginning of a bio, since one bio can never be
* larger than UINT_MAX, u32 here is enough.
*/
u32 bio_offset = 0;
int mirror;
struct bvec_iter_all iter_all;
ASSERT(!bio_flagged(bio, BIO_CLONED));
bio_for_each_segment_all(bvec, bio, iter_all) {
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
bool uptodate = !bio->bi_status;
struct page *page = bvec->bv_page;
struct inode *inode = page->mapping->host;
struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
const u32 sectorsize = fs_info->sectorsize;
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
unsigned int error_bitmap = (unsigned int)-1;
bool repair = false;
u64 start;
u64 end;
u32 len;
btrfs_debug(fs_info,
"end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
bio->bi_iter.bi_sector, bio->bi_status,
bbio->mirror_num);
/*
* We always issue full-sector reads, but if some block in a
* page fails to read, blk_update_request() will advance
* bv_offset and adjust bv_len to compensate. Print a warning
* for unaligned offsets, and an error if they don't add up to
* a full sector.
*/
if (!IS_ALIGNED(bvec->bv_offset, sectorsize))
btrfs_err(fs_info,
"partial page read in btrfs with offset %u and length %u",
bvec->bv_offset, bvec->bv_len);
else if (!IS_ALIGNED(bvec->bv_offset + bvec->bv_len,
sectorsize))
btrfs_info(fs_info,
"incomplete page read with offset %u and length %u",
bvec->bv_offset, bvec->bv_len);
start = page_offset(page) + bvec->bv_offset;
end = start + bvec->bv_len - 1;
len = bvec->bv_len;
mirror = bbio->mirror_num;
if (likely(uptodate)) {
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
if (is_data_inode(inode)) {
error_bitmap = btrfs_verify_data_csum(bbio,
bio_offset, page, start, end);
if (error_bitmap)
uptodate = false;
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
} else {
if (btrfs_validate_metadata_buffer(bbio,
page, start, end, mirror))
uptodate = false;
btrfs: submit read time repair only for each corrupted sector Currently btrfs_submit_read_repair() has some extra check on whether the failed bio needs extra validation for repair. But we can avoid all these extra mechanisms if we submit the repair for each sector. By this, each read repair can be easily handled without the need to verify which sector is corrupted. This will also benefit subpage, as one subpage bvec can contain several sectors, making the extra verification more complex. So this patch will: - Introduce repair_one_sector() The main code submitting repair, which is more or less the same as old btrfs_submit_read_repair(). But this time, it only repairs one sector. - Make btrfs_submit_read_repair() to handle sectors differently There are 3 different cases: * Good sector We need to release the page and extent, set the range uptodate. * Bad sector and failed to submit repair bio We need to release the page and extent, but not set the range uptodate. * Bad sector but repair bio submitted The page and extent release will be handled by the submitted repair bio. Nothing needs to be done. Since btrfs_submit_read_repair() will handle the page and extent release now, we need to skip to next bvec even we hit some error. - Change the lifespan of @uptodate in end_bio_extent_readpage() Since now btrfs_submit_read_repair() will handle the full bvec which contains any corruption, we don't need to bother updating @uptodate bit anymore. Just let @uptodate to be local variable inside the main loop, so that any error from one bvec won't affect later bvec. - Only export btrfs_repair_one_sector(), unexport btrfs_submit_read_repair() The only outside caller for read repair is DIO, which already submits its repair for just one sector. Only export btrfs_repair_one_sector() for DIO. This patch will focus on the change on the repair path, the extra validation code is still kept as is, and will be cleaned up later. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-03 10:08:55 +08:00
}
}
if (likely(uptodate)) {
loff_t i_size = i_size_read(inode);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
pgoff_t end_index = i_size >> PAGE_SHIFT;
btrfs_clean_io_failure(BTRFS_I(inode), start, page, 0);
btrfs: subpage: fix the false data csum mismatch error [BUG] When running fstresss, we can hit strange data csum mismatch where the on-disk data is in fact correct (passes scrub). With some extra debug info added, we have the following traces: 0482us: btrfs_do_readpage: root=5 ino=284 offset=393216, submit force=0 pgoff=0 iosize=8192 0494us: btrfs_do_readpage: root=5 ino=284 offset=401408, submit force=0 pgoff=8192 iosize=4096 0498us: btrfs_submit_data_bio: root=5 ino=284 bio first bvec=393216 len=8192 0591us: btrfs_do_readpage: root=5 ino=284 offset=405504, submit force=0 pgoff=12288 iosize=36864 0594us: btrfs_submit_data_bio: root=5 ino=284 bio first bvec=401408 len=4096 0863us: btrfs_submit_data_bio: root=5 ino=284 bio first bvec=405504 len=36864 0933us: btrfs_verify_data_csum: root=5 ino=284 offset=393216 len=8192 0967us: btrfs_do_readpage: root=5 ino=284 offset=442368, skip beyond isize pgoff=49152 iosize=16384 1047us: btrfs_verify_data_csum: root=5 ino=284 offset=401408 len=4096 1163us: btrfs_verify_data_csum: root=5 ino=284 offset=405504 len=36864 1290us: check_data_csum: !!! root=5 ino=284 offset=438272 pg_off=45056 !!! 7387us: end_bio_extent_readpage: root=5 ino=284 before pending_read_bios=0 [CAUSE] Normally we expect all submitted bio reads to only touch the range we specified, and under subpage context, it means we should only touch the range specified in each bvec. But in data read path, inside end_bio_extent_readpage(), we have page zeroing which only takes regular page size into consideration. This means for subpage if we have an inode whose content looks like below: 0 16K 32K 48K 64K |///////| |///////| | |//| = data needs to be read from disk | | = hole And i_size is 64K initially. Then the following race can happen: T1 | T2 --------------------------------+-------------------------------- btrfs_do_readpage() | |- isize = 64K; | | At this time, the isize is | | 64K | | | |- submit_extent_page() | | submit previous assembled bio| | assemble bio for [0, 16K) | | | |- submit_extent_page() | submit read bio for [0, 16K) | assemble read bio for | [32K, 48K) | | | btrfs_setsize() | |- i_size_write(, 16K); | Now i_size is only 16K end_io() for [0K, 16K) | |- end_bio_extent_readpage() | |- btrfs_verify_data_csum() | | No csum error | |- i_size = 16K; | |- zero_user_segment(16K, | PAGE_SIZE); | !!! We zeroed range | !!! [32K, 48K) | | end_io for [32K, 48K) | |- end_bio_extent_readpage() | |- btrfs_verify_data_csum() | ! CSUM MISMATCH ! | ! As the range is zeroed now ! [FIX] To fix the problem, make end_bio_extent_readpage() to only zero the range of bvec. The bug only affects subpage read-write support, as for full read-only mount we can't change i_size thus won't hit the race condition. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-01 16:44:22 +08:00
/*
* Zero out the remaining part if this range straddles
* i_size.
*
* Here we should only zero the range inside the bvec,
* not touch anything else.
*
* NOTE: i_size is exclusive while end is inclusive.
*/
if (page->index == end_index && i_size <= end) {
u32 zero_start = max(offset_in_page(i_size),
btrfs: fix wrong offset to zero out range beyond i_size [BUG] The test generic/091 fails , with the following output: fsx -N 10000 -o 128000 -l 500000 -r PSIZE -t BSIZE -w BSIZE -Z -W mapped writes DISABLED Seed set to 1 main: filesystem does not support fallocate mode FALLOC_FL_COLLAPSE_RANGE, disabling! main: filesystem does not support fallocate mode FALLOC_FL_INSERT_RANGE, disabling! skipping zero size read truncating to largest ever: 0xe400 copying to largest ever: 0x1f400 cloning to largest ever: 0x70000 cloning to largest ever: 0x77000 fallocating to largest ever: 0x7a120 Mapped Read: non-zero data past EOF (0x3a7ff) page offset 0x800 is 0xf2e1 <<< ... [CAUSE] In commit c28ea613fafa ("btrfs: subpage: fix the false data csum mismatch error") end_bio_extent_readpage() changes to only zero the range inside the bvec for incoming subpage support. But that commit is using incorrect offset to calculate the start. For subpage, we can have a case that the whole bvec is beyond isize, thus we need to calculate the correct offset. But the offending commit is using @end (bvec end), other than @start (bvec start) to calculate the start offset. This means, we only zero the last byte of the bvec, not from the isize. This stupid bug makes the range beyond isize is not properly zeroed, and failed above test. [FIX] Use correct @start to calculate the range start. Reported-by: kernel test robot <oliver.sang@intel.com> Fixes: c28ea613fafa ("btrfs: subpage: fix the false data csum mismatch error") Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-08 17:20:17 +08:00
offset_in_page(start));
btrfs: subpage: fix the false data csum mismatch error [BUG] When running fstresss, we can hit strange data csum mismatch where the on-disk data is in fact correct (passes scrub). With some extra debug info added, we have the following traces: 0482us: btrfs_do_readpage: root=5 ino=284 offset=393216, submit force=0 pgoff=0 iosize=8192 0494us: btrfs_do_readpage: root=5 ino=284 offset=401408, submit force=0 pgoff=8192 iosize=4096 0498us: btrfs_submit_data_bio: root=5 ino=284 bio first bvec=393216 len=8192 0591us: btrfs_do_readpage: root=5 ino=284 offset=405504, submit force=0 pgoff=12288 iosize=36864 0594us: btrfs_submit_data_bio: root=5 ino=284 bio first bvec=401408 len=4096 0863us: btrfs_submit_data_bio: root=5 ino=284 bio first bvec=405504 len=36864 0933us: btrfs_verify_data_csum: root=5 ino=284 offset=393216 len=8192 0967us: btrfs_do_readpage: root=5 ino=284 offset=442368, skip beyond isize pgoff=49152 iosize=16384 1047us: btrfs_verify_data_csum: root=5 ino=284 offset=401408 len=4096 1163us: btrfs_verify_data_csum: root=5 ino=284 offset=405504 len=36864 1290us: check_data_csum: !!! root=5 ino=284 offset=438272 pg_off=45056 !!! 7387us: end_bio_extent_readpage: root=5 ino=284 before pending_read_bios=0 [CAUSE] Normally we expect all submitted bio reads to only touch the range we specified, and under subpage context, it means we should only touch the range specified in each bvec. But in data read path, inside end_bio_extent_readpage(), we have page zeroing which only takes regular page size into consideration. This means for subpage if we have an inode whose content looks like below: 0 16K 32K 48K 64K |///////| |///////| | |//| = data needs to be read from disk | | = hole And i_size is 64K initially. Then the following race can happen: T1 | T2 --------------------------------+-------------------------------- btrfs_do_readpage() | |- isize = 64K; | | At this time, the isize is | | 64K | | | |- submit_extent_page() | | submit previous assembled bio| | assemble bio for [0, 16K) | | | |- submit_extent_page() | submit read bio for [0, 16K) | assemble read bio for | [32K, 48K) | | | btrfs_setsize() | |- i_size_write(, 16K); | Now i_size is only 16K end_io() for [0K, 16K) | |- end_bio_extent_readpage() | |- btrfs_verify_data_csum() | | No csum error | |- i_size = 16K; | |- zero_user_segment(16K, | PAGE_SIZE); | !!! We zeroed range | !!! [32K, 48K) | | end_io for [32K, 48K) | |- end_bio_extent_readpage() | |- btrfs_verify_data_csum() | ! CSUM MISMATCH ! | ! As the range is zeroed now ! [FIX] To fix the problem, make end_bio_extent_readpage() to only zero the range of bvec. The bug only affects subpage read-write support, as for full read-only mount we can't change i_size thus won't hit the race condition. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-01 16:44:22 +08:00
zero_user_segment(page, zero_start,
offset_in_page(end) + 1);
}
} else if (is_data_inode(inode)) {
/*
* Only try to repair bios that actually made it to a
* device. If the bio failed to be submitted mirror
* is 0 and we need to fail it without retrying.
btrfs: fix repair of compressed extents Currently the checksum of compressed extents is verified based on the compressed data and the lower btrfs_bio, but the actual repair process is driven by end_bio_extent_readpage on the upper btrfs_bio for the decompressed data. This has a bunch of issues, including not being able to properly communicate the failed mirror up in case that the I/O submission got preempted, a general loss of if an error was an I/O error or a checksum verification failure, but most importantly that this design causes btrfs_clean_io_failure to eventually write back the uncompressed good data onto the disk sectors that are supposed to contain compressed data. Fix this by moving the repair to the lower btrfs_bio. To do so, a fair amount of code has to be reshuffled: a) the lower btrfs_bio now needs a valid csum pointer. The easiest way to achieve that is to pass NULL btrfs_lookup_bio_sums and just use the btrfs_bio management of csums. For a compressed_bio that is split into multiple btrfs_bios this means additional memory allocations, but the code becomes a lot more regular. b) checksum verification now runs directly on the lower btrfs_bio instead of the compressed_bio. This actually nicely simplifies the end I/O processing. c) btrfs_repair_one_sector can't just look up the logical address for the file offset any more, as there is no corresponding relative offsets that apply to the file offset and the logic address for compressed extents. Instead require that the saved bvec_iter in the btrfs_bio is filled out for all read bios and use that, which again removes a fair amount of code. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: David Sterba <dsterba@suse.com>
2022-07-07 13:33:30 +08:00
*
* This also includes the high level bios for compressed
* extents - these never make it to a device and repair
* is already handled on the lower compressed bio.
*/
if (mirror > 0)
repair = true;
} else {
struct extent_buffer *eb;
eb = find_extent_buffer_readpage(fs_info, page, start);
set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
eb->read_mirror = mirror;
atomic_dec(&eb->io_pages);
}
if (repair) {
/*
* submit_data_read_repair() will handle all the good
* and bad sectors, we just continue to the next bvec.
*/
submit_data_read_repair(inode, bbio, bio_offset, bvec,
error_bitmap);
} else {
/* Update page status and unlock */
end_page_read(page, uptodate, start, len);
endio_readpage_release_extent(&processed, BTRFS_I(inode),
start, end, PageUptodate(page));
}
ASSERT(bio_offset + len > bio_offset);
bio_offset += len;
}
btrfs: add structure to keep track of extent range in end_bio_extent_readpage In end_bio_extent_readpage() we had a strange dance around extent_start/extent_len. Hidden behind the strange dance is, it's just calling endio_readpage_release_extent() on each bvec range. Here is an example to explain the original work flow: Bio is for inode 257, containing 2 pages, for range [1M, 1M+8K) end_bio_extent_extent_readpage() entered |- extent_start = 0; |- extent_end = 0; |- bio_for_each_segment_all() { | |- /* Got the 1st bvec */ | |- start = SZ_1M; | |- end = SZ_1M + SZ_4K - 1; | |- update = 1; | |- if (extent_len == 0) { | | |- extent_start = start; /* SZ_1M */ | | |- extent_len = end + 1 - start; /* SZ_1M */ | | } | | | |- /* Got the 2nd bvec */ | |- start = SZ_1M + 4K; | |- end = SZ_1M + 4K - 1; | |- update = 1; | |- if (extent_start + extent_len == start) { | | |- extent_len += end + 1 - start; /* SZ_8K */ | | } | } /* All bio vec iterated */ | |- if (extent_len) { |- endio_readpage_release_extent(tree, extent_start, extent_len, update); /* extent_start == SZ_1M, extent_len == SZ_8K, uptodate = 1 */ As the above flow shows, the existing code in end_bio_extent_readpage() is accumulates extent_start/extent_len, and when the contiguous range stops, calls endio_readpage_release_extent() for the range. However current behavior has something not really considered: - The inode can change For bio, its pages don't need to have contiguous page_offset. This means, even pages from different inodes can be packed into one bio. - bvec cross page boundary There is a feature called multi-page bvec, where bvec->bv_len can go beyond bvec->bv_page boundary. - Poor readability This patch will address the problem: - Introduce a proper structure, processed_extent, to record processed extent range - Integrate inode/start/end/uptodate check into endio_readpage_release_extent() - Add more comment on each step. This should greatly improve the readability, now in end_bio_extent_readpage() there are only two endio_readpage_release_extent() calls. - Add inode check for contiguity Now we also ensure the inode is the same one before checking if the range is contiguous. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-13 20:51:28 +08:00
/* Release the last extent */
endio_readpage_release_extent(&processed, NULL, 0, 0, false);
btrfs_bio_free_csum(bbio);
bio_put(bio);
}
/*
* Populate every free slot in a provided array with pages.
*
* @nr_pages: number of pages to allocate
* @page_array: the array to fill with pages; any existing non-null entries in
* the array will be skipped
*
* Return: 0 if all pages were able to be allocated;
* -ENOMEM otherwise, and the caller is responsible for freeing all
* non-null page pointers in the array.
*/
int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array)
{
unsigned int allocated;
for (allocated = 0; allocated < nr_pages;) {
unsigned int last = allocated;
allocated = alloc_pages_bulk_array(GFP_NOFS, nr_pages, page_array);
if (allocated == nr_pages)
return 0;
/*
* During this iteration, no page could be allocated, even
* though alloc_pages_bulk_array() falls back to alloc_page()
* if it could not bulk-allocate. So we must be out of memory.
*/
if (allocated == last)
return -ENOMEM;
memalloc_retry_wait(GFP_NOFS);
}
return 0;
}
/*
* Attempt to add a page to bio.
*
* @bio_ctrl: record both the bio, and its bio_flags
* @page: page to add to the bio
* @disk_bytenr: offset of the new bio or to check whether we are adding
* a contiguous page to the previous one
* @size: portion of page that we want to write
* @pg_offset: starting offset in the page
* @compress_type: compression type of the current bio to see if we can merge them
*
* Attempt to add a page to bio considering stripe alignment etc.
*
* Return >= 0 for the number of bytes added to the bio.
* Can return 0 if the current bio is already at stripe/zone boundary.
* Return <0 for error.
*/
static int btrfs_bio_add_page(struct btrfs_bio_ctrl *bio_ctrl,
struct page *page,
u64 disk_bytenr, unsigned int size,
unsigned int pg_offset,
enum btrfs_compression_type compress_type)
{
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
struct bio *bio = bio_ctrl->bio;
u32 bio_size = bio->bi_iter.bi_size;
u32 real_size;
const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
btrfs: don't merge pages into bio if their page offset is not contiguous [BUG] Zygo reported on latest development branch, he could hit ASSERT()/BUG_ON() caused crash when doing RAID5 recovery (intentionally corrupt one disk, and let btrfs to recover the data during read/scrub). And The following minimal reproducer can cause extent state leakage at rmmod time: mkfs.btrfs -f -d raid5 -m raid5 $dev1 $dev2 $dev3 -b 1G > /dev/null mount $dev1 $mnt fsstress -w -d $mnt -n 25 -s 1660807876 sync fssum -A -f -w /tmp/fssum.saved $mnt umount $mnt # Wipe the dev1 but keeps its super block xfs_io -c "pwrite -S 0x0 1m 1023m" $dev1 mount $dev1 $mnt fssum -r /tmp/fssum.saved $mnt > /dev/null umount $mnt rmmod btrfs This will lead to the following extent states leakage: BTRFS: state leak: start 499712 end 503807 state 5 in tree 1 refs 1 BTRFS: state leak: start 495616 end 499711 state 5 in tree 1 refs 1 BTRFS: state leak: start 491520 end 495615 state 5 in tree 1 refs 1 BTRFS: state leak: start 487424 end 491519 state 5 in tree 1 refs 1 BTRFS: state leak: start 483328 end 487423 state 5 in tree 1 refs 1 BTRFS: state leak: start 479232 end 483327 state 5 in tree 1 refs 1 BTRFS: state leak: start 475136 end 479231 state 5 in tree 1 refs 1 BTRFS: state leak: start 471040 end 475135 state 5 in tree 1 refs 1 [CAUSE] Since commit 7aa51232e204 ("btrfs: pass a btrfs_bio to btrfs_repair_one_sector"), we always use btrfs_bio->file_offset to determine the file offset of a page. But that usage assume that, one bio has all its page having a continuous page offsets. Unfortunately that's not true, btrfs only requires the logical bytenr contiguous when assembling its bios. From above script, we have one bio looks like this: fssum-27671 submit_one_bio: bio logical=217739264 len=36864 fssum-27671 submit_one_bio: r/i=5/261 page_offset=466944 <<< fssum-27671 submit_one_bio: r/i=5/261 page_offset=724992 <<< fssum-27671 submit_one_bio: r/i=5/261 page_offset=729088 fssum-27671 submit_one_bio: r/i=5/261 page_offset=733184 fssum-27671 submit_one_bio: r/i=5/261 page_offset=737280 fssum-27671 submit_one_bio: r/i=5/261 page_offset=741376 fssum-27671 submit_one_bio: r/i=5/261 page_offset=745472 fssum-27671 submit_one_bio: r/i=5/261 page_offset=749568 fssum-27671 submit_one_bio: r/i=5/261 page_offset=753664 Note that the 1st and the 2nd page has non-contiguous page offsets. This means, at repair time, we will have completely wrong file offset passed in: kworker/u32:2-19927 btrfs_repair_one_sector: r/i=5/261 page_off=729088 file_off=475136 bio_offset=8192 Since the file offset is incorrect, we latter incorrectly set the extent states, and no way to really release them. Thus later it causes the leakage. In fact, this can be even worse, since the file offset is incorrect, we can hit cases like the incorrect file offset belongs to a HOLE, and later cause btrfs_num_copies() to trigger error, finally hit BUG_ON()/ASSERT() later. [FIX] Add an extra condition in btrfs_bio_add_page() for uncompressed IO. Now we will have more strict requirement for bio pages: - They should all have the same mapping (the mapping check is already implied by the call chain) - Their logical bytenr should be adjacent This is the same as the old condition. - Their page_offset() (file offset) should be adjacent This is the new check. This would result a slightly increased amount of bios from btrfs (needs holes and inside the same stripe boundary to trigger). But this would greatly reduce the confusion, as it's pretty common to assume a btrfs bio would only contain continuous page cache. Later we may need extra cleanups, as we no longer needs to handle gaps between page offsets in endio functions. Currently this should be the minimal patch to fix commit 7aa51232e204 ("btrfs: pass a btrfs_bio to btrfs_repair_one_sector"). Reported-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org> Fixes: 7aa51232e204 ("btrfs: pass a btrfs_bio to btrfs_repair_one_sector") Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-08-13 16:06:53 +08:00
bool contig = false;
int ret;
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
ASSERT(bio);
/* The limit should be calculated when bio_ctrl->bio is allocated */
ASSERT(bio_ctrl->len_to_oe_boundary && bio_ctrl->len_to_stripe_boundary);
if (bio_ctrl->compress_type != compress_type)
return 0;
btrfs: don't merge pages into bio if their page offset is not contiguous [BUG] Zygo reported on latest development branch, he could hit ASSERT()/BUG_ON() caused crash when doing RAID5 recovery (intentionally corrupt one disk, and let btrfs to recover the data during read/scrub). And The following minimal reproducer can cause extent state leakage at rmmod time: mkfs.btrfs -f -d raid5 -m raid5 $dev1 $dev2 $dev3 -b 1G > /dev/null mount $dev1 $mnt fsstress -w -d $mnt -n 25 -s 1660807876 sync fssum -A -f -w /tmp/fssum.saved $mnt umount $mnt # Wipe the dev1 but keeps its super block xfs_io -c "pwrite -S 0x0 1m 1023m" $dev1 mount $dev1 $mnt fssum -r /tmp/fssum.saved $mnt > /dev/null umount $mnt rmmod btrfs This will lead to the following extent states leakage: BTRFS: state leak: start 499712 end 503807 state 5 in tree 1 refs 1 BTRFS: state leak: start 495616 end 499711 state 5 in tree 1 refs 1 BTRFS: state leak: start 491520 end 495615 state 5 in tree 1 refs 1 BTRFS: state leak: start 487424 end 491519 state 5 in tree 1 refs 1 BTRFS: state leak: start 483328 end 487423 state 5 in tree 1 refs 1 BTRFS: state leak: start 479232 end 483327 state 5 in tree 1 refs 1 BTRFS: state leak: start 475136 end 479231 state 5 in tree 1 refs 1 BTRFS: state leak: start 471040 end 475135 state 5 in tree 1 refs 1 [CAUSE] Since commit 7aa51232e204 ("btrfs: pass a btrfs_bio to btrfs_repair_one_sector"), we always use btrfs_bio->file_offset to determine the file offset of a page. But that usage assume that, one bio has all its page having a continuous page offsets. Unfortunately that's not true, btrfs only requires the logical bytenr contiguous when assembling its bios. From above script, we have one bio looks like this: fssum-27671 submit_one_bio: bio logical=217739264 len=36864 fssum-27671 submit_one_bio: r/i=5/261 page_offset=466944 <<< fssum-27671 submit_one_bio: r/i=5/261 page_offset=724992 <<< fssum-27671 submit_one_bio: r/i=5/261 page_offset=729088 fssum-27671 submit_one_bio: r/i=5/261 page_offset=733184 fssum-27671 submit_one_bio: r/i=5/261 page_offset=737280 fssum-27671 submit_one_bio: r/i=5/261 page_offset=741376 fssum-27671 submit_one_bio: r/i=5/261 page_offset=745472 fssum-27671 submit_one_bio: r/i=5/261 page_offset=749568 fssum-27671 submit_one_bio: r/i=5/261 page_offset=753664 Note that the 1st and the 2nd page has non-contiguous page offsets. This means, at repair time, we will have completely wrong file offset passed in: kworker/u32:2-19927 btrfs_repair_one_sector: r/i=5/261 page_off=729088 file_off=475136 bio_offset=8192 Since the file offset is incorrect, we latter incorrectly set the extent states, and no way to really release them. Thus later it causes the leakage. In fact, this can be even worse, since the file offset is incorrect, we can hit cases like the incorrect file offset belongs to a HOLE, and later cause btrfs_num_copies() to trigger error, finally hit BUG_ON()/ASSERT() later. [FIX] Add an extra condition in btrfs_bio_add_page() for uncompressed IO. Now we will have more strict requirement for bio pages: - They should all have the same mapping (the mapping check is already implied by the call chain) - Their logical bytenr should be adjacent This is the same as the old condition. - Their page_offset() (file offset) should be adjacent This is the new check. This would result a slightly increased amount of bios from btrfs (needs holes and inside the same stripe boundary to trigger). But this would greatly reduce the confusion, as it's pretty common to assume a btrfs bio would only contain continuous page cache. Later we may need extra cleanups, as we no longer needs to handle gaps between page offsets in endio functions. Currently this should be the minimal patch to fix commit 7aa51232e204 ("btrfs: pass a btrfs_bio to btrfs_repair_one_sector"). Reported-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org> Fixes: 7aa51232e204 ("btrfs: pass a btrfs_bio to btrfs_repair_one_sector") Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-08-13 16:06:53 +08:00
if (bio->bi_iter.bi_size == 0) {
/* We can always add a page into an empty bio. */
contig = true;
} else if (bio_ctrl->compress_type == BTRFS_COMPRESS_NONE) {
struct bio_vec *bvec = bio_last_bvec_all(bio);
/*
* The contig check requires the following conditions to be met:
* 1) The pages are belonging to the same inode
* This is implied by the call chain.
*
* 2) The range has adjacent logical bytenr
*
* 3) The range has adjacent file offset
* This is required for the usage of btrfs_bio->file_offset.
*/
if (bio_end_sector(bio) == sector &&
page_offset(bvec->bv_page) + bvec->bv_offset +
bvec->bv_len == page_offset(page) + pg_offset)
contig = true;
} else {
/*
* For compression, all IO should have its logical bytenr
* set to the starting bytenr of the compressed extent.
*/
contig = bio->bi_iter.bi_sector == sector;
btrfs: don't merge pages into bio if their page offset is not contiguous [BUG] Zygo reported on latest development branch, he could hit ASSERT()/BUG_ON() caused crash when doing RAID5 recovery (intentionally corrupt one disk, and let btrfs to recover the data during read/scrub). And The following minimal reproducer can cause extent state leakage at rmmod time: mkfs.btrfs -f -d raid5 -m raid5 $dev1 $dev2 $dev3 -b 1G > /dev/null mount $dev1 $mnt fsstress -w -d $mnt -n 25 -s 1660807876 sync fssum -A -f -w /tmp/fssum.saved $mnt umount $mnt # Wipe the dev1 but keeps its super block xfs_io -c "pwrite -S 0x0 1m 1023m" $dev1 mount $dev1 $mnt fssum -r /tmp/fssum.saved $mnt > /dev/null umount $mnt rmmod btrfs This will lead to the following extent states leakage: BTRFS: state leak: start 499712 end 503807 state 5 in tree 1 refs 1 BTRFS: state leak: start 495616 end 499711 state 5 in tree 1 refs 1 BTRFS: state leak: start 491520 end 495615 state 5 in tree 1 refs 1 BTRFS: state leak: start 487424 end 491519 state 5 in tree 1 refs 1 BTRFS: state leak: start 483328 end 487423 state 5 in tree 1 refs 1 BTRFS: state leak: start 479232 end 483327 state 5 in tree 1 refs 1 BTRFS: state leak: start 475136 end 479231 state 5 in tree 1 refs 1 BTRFS: state leak: start 471040 end 475135 state 5 in tree 1 refs 1 [CAUSE] Since commit 7aa51232e204 ("btrfs: pass a btrfs_bio to btrfs_repair_one_sector"), we always use btrfs_bio->file_offset to determine the file offset of a page. But that usage assume that, one bio has all its page having a continuous page offsets. Unfortunately that's not true, btrfs only requires the logical bytenr contiguous when assembling its bios. From above script, we have one bio looks like this: fssum-27671 submit_one_bio: bio logical=217739264 len=36864 fssum-27671 submit_one_bio: r/i=5/261 page_offset=466944 <<< fssum-27671 submit_one_bio: r/i=5/261 page_offset=724992 <<< fssum-27671 submit_one_bio: r/i=5/261 page_offset=729088 fssum-27671 submit_one_bio: r/i=5/261 page_offset=733184 fssum-27671 submit_one_bio: r/i=5/261 page_offset=737280 fssum-27671 submit_one_bio: r/i=5/261 page_offset=741376 fssum-27671 submit_one_bio: r/i=5/261 page_offset=745472 fssum-27671 submit_one_bio: r/i=5/261 page_offset=749568 fssum-27671 submit_one_bio: r/i=5/261 page_offset=753664 Note that the 1st and the 2nd page has non-contiguous page offsets. This means, at repair time, we will have completely wrong file offset passed in: kworker/u32:2-19927 btrfs_repair_one_sector: r/i=5/261 page_off=729088 file_off=475136 bio_offset=8192 Since the file offset is incorrect, we latter incorrectly set the extent states, and no way to really release them. Thus later it causes the leakage. In fact, this can be even worse, since the file offset is incorrect, we can hit cases like the incorrect file offset belongs to a HOLE, and later cause btrfs_num_copies() to trigger error, finally hit BUG_ON()/ASSERT() later. [FIX] Add an extra condition in btrfs_bio_add_page() for uncompressed IO. Now we will have more strict requirement for bio pages: - They should all have the same mapping (the mapping check is already implied by the call chain) - Their logical bytenr should be adjacent This is the same as the old condition. - Their page_offset() (file offset) should be adjacent This is the new check. This would result a slightly increased amount of bios from btrfs (needs holes and inside the same stripe boundary to trigger). But this would greatly reduce the confusion, as it's pretty common to assume a btrfs bio would only contain continuous page cache. Later we may need extra cleanups, as we no longer needs to handle gaps between page offsets in endio functions. Currently this should be the minimal patch to fix commit 7aa51232e204 ("btrfs: pass a btrfs_bio to btrfs_repair_one_sector"). Reported-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org> Fixes: 7aa51232e204 ("btrfs: pass a btrfs_bio to btrfs_repair_one_sector") Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-08-13 16:06:53 +08:00
}
if (!contig)
return 0;
real_size = min(bio_ctrl->len_to_oe_boundary,
bio_ctrl->len_to_stripe_boundary) - bio_size;
real_size = min(real_size, size);
/*
* If real_size is 0, never call bio_add_*_page(), as even size is 0,
* bio will still execute its endio function on the page!
*/
if (real_size == 0)
return 0;
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
if (bio_op(bio) == REQ_OP_ZONE_APPEND)
ret = bio_add_zone_append_page(bio, page, real_size, pg_offset);
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
else
ret = bio_add_page(bio, page, real_size, pg_offset);
return ret;
}
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
static int calc_bio_boundaries(struct btrfs_bio_ctrl *bio_ctrl,
struct btrfs_inode *inode, u64 file_offset)
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
{
struct btrfs_fs_info *fs_info = inode->root->fs_info;
struct btrfs_io_geometry geom;
struct btrfs_ordered_extent *ordered;
struct extent_map *em;
u64 logical = (bio_ctrl->bio->bi_iter.bi_sector << SECTOR_SHIFT);
int ret;
/*
* Pages for compressed extent are never submitted to disk directly,
* thus it has no real boundary, just set them to U32_MAX.
*
* The split happens for real compressed bio, which happens in
* btrfs_submit_compressed_read/write().
*/
if (bio_ctrl->compress_type != BTRFS_COMPRESS_NONE) {
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
bio_ctrl->len_to_oe_boundary = U32_MAX;
bio_ctrl->len_to_stripe_boundary = U32_MAX;
return 0;
}
em = btrfs_get_chunk_map(fs_info, logical, fs_info->sectorsize);
if (IS_ERR(em))
return PTR_ERR(em);
ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio_ctrl->bio),
logical, &geom);
free_extent_map(em);
if (ret < 0) {
return ret;
}
if (geom.len > U32_MAX)
bio_ctrl->len_to_stripe_boundary = U32_MAX;
else
bio_ctrl->len_to_stripe_boundary = (u32)geom.len;
if (bio_op(bio_ctrl->bio) != REQ_OP_ZONE_APPEND) {
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
bio_ctrl->len_to_oe_boundary = U32_MAX;
return 0;
}
/* Ordered extent not yet created, so we're good */
ordered = btrfs_lookup_ordered_extent(inode, file_offset);
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
if (!ordered) {
bio_ctrl->len_to_oe_boundary = U32_MAX;
return 0;
}
bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
ordered->disk_bytenr + ordered->disk_num_bytes - logical);
btrfs_put_ordered_extent(ordered);
return 0;
}
static int alloc_new_bio(struct btrfs_inode *inode,
struct btrfs_bio_ctrl *bio_ctrl,
struct writeback_control *wbc,
blk_opf_t opf,
u64 disk_bytenr, u32 offset, u64 file_offset,
enum btrfs_compression_type compress_type)
{
struct btrfs_fs_info *fs_info = inode->root->fs_info;
struct bio *bio;
int ret;
ASSERT(bio_ctrl->end_io_func);
bio = btrfs_bio_alloc(BIO_MAX_VECS, opf, bio_ctrl->end_io_func, NULL);
/*
* For compressed page range, its disk_bytenr is always @disk_bytenr
* passed in, no matter if we have added any range into previous bio.
*/
if (compress_type != BTRFS_COMPRESS_NONE)
bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
else
bio->bi_iter.bi_sector = (disk_bytenr + offset) >> SECTOR_SHIFT;
bio_ctrl->bio = bio;
bio_ctrl->compress_type = compress_type;
ret = calc_bio_boundaries(bio_ctrl, inode, file_offset);
if (ret < 0)
goto error;
if (wbc) {
/*
* For Zone append we need the correct block_device that we are
* going to write to set in the bio to be able to respect the
* hardware limitation. Look it up here:
*/
if (bio_op(bio) == REQ_OP_ZONE_APPEND) {
struct btrfs_device *dev;
dev = btrfs_zoned_get_device(fs_info, disk_bytenr,
fs_info->sectorsize);
if (IS_ERR(dev)) {
ret = PTR_ERR(dev);
goto error;
}
bio_set_dev(bio, dev->bdev);
} else {
/*
* Otherwise pick the last added device to support
* cgroup writeback. For multi-device file systems this
* means blk-cgroup policies have to always be set on the
* last added/replaced device. This is a bit odd but has
* been like that for a long time.
*/
bio_set_dev(bio, fs_info->fs_devices->latest_dev->bdev);
}
wbc_init_bio(wbc, bio);
} else {
ASSERT(bio_op(bio) != REQ_OP_ZONE_APPEND);
}
return 0;
error:
bio_ctrl->bio = NULL;
btrfs_bio_end_io(btrfs_bio(bio), errno_to_blk_status(ret));
return ret;
}
/*
* @opf: bio REQ_OP_* and REQ_* flags as one value
* @wbc: optional writeback control for io accounting
* @disk_bytenr: logical bytenr where the write will be
* @page: page to add to the bio
* @size: portion of page that we want to write to
* @pg_offset: offset of the new bio or to check whether we are adding
* a contiguous page to the previous one
* @compress_type: compress type for current bio
*
* The will either add the page into the existing @bio_ctrl->bio, or allocate a
* new one in @bio_ctrl->bio.
* The mirror number for this IO should already be initizlied in
* @bio_ctrl->mirror_num.
*/
static int submit_extent_page(blk_opf_t opf,
struct writeback_control *wbc,
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
struct btrfs_bio_ctrl *bio_ctrl,
u64 disk_bytenr, struct page *page,
size_t size, unsigned long pg_offset,
enum btrfs_compression_type compress_type,
Btrfs: fix read corruption of compressed and shared extents If a file has a range pointing to a compressed extent, followed by another range that points to the same compressed extent and a read operation attempts to read both ranges (either completely or part of them), the pages that correspond to the second range are incorrectly filled with zeroes. Consider the following example: File layout [0 - 8K] [8K - 24K] | | | | points to extent X, points to extent X, offset 4K, length of 8K offset 0, length 16K [extent X, compressed length = 4K uncompressed length = 16K] If a readpages() call spans the 2 ranges, a single bio to read the extent is submitted - extent_io.c:submit_extent_page() would only create a new bio to cover the second range pointing to the extent if the extent it points to had a different logical address than the extent associated with the first range. This has a consequence of the compressed read end io handler (compression.c:end_compressed_bio_read()) finish once the extent is decompressed into the pages covering the first range, leaving the remaining pages (belonging to the second range) filled with zeroes (done by compression.c:btrfs_clear_biovec_end()). So fix this by submitting the current bio whenever we find a range pointing to a compressed extent that was preceded by a range with a different extent map. This is the simplest solution for this corner case. Making the end io callback populate both ranges (or more, if we have multiple pointing to the same extent) is a much more complex solution since each bio is tightly coupled with a single extent map and the extent maps associated to the ranges pointing to the shared extent can have different offsets and lengths. The following test case for fstests triggers the issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_cloner rm -f $seqres.full test_clone_and_read_compressed_extent() { local mount_opts=$1 _scratch_mkfs >>$seqres.full 2>&1 _scratch_mount $mount_opts # Create a test file with a single extent that is compressed (the # data we write into it is highly compressible no matter which # compression algorithm is used, zlib or lzo). $XFS_IO_PROG -f -c "pwrite -S 0xaa 0K 4K" \ -c "pwrite -S 0xbb 4K 8K" \ -c "pwrite -S 0xcc 12K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now clone our extent into an adjacent offset. $CLONER_PROG -s $((4 * 1024)) -d $((16 * 1024)) -l $((8 * 1024)) \ $SCRATCH_MNT/foo $SCRATCH_MNT/foo # Same as before but for this file we clone the extent into a lower # file offset. $XFS_IO_PROG -f -c "pwrite -S 0xaa 8K 4K" \ -c "pwrite -S 0xbb 12K 8K" \ -c "pwrite -S 0xcc 20K 4K" \ $SCRATCH_MNT/bar | _filter_xfs_io $CLONER_PROG -s $((12 * 1024)) -d 0 -l $((8 * 1024)) \ $SCRATCH_MNT/bar $SCRATCH_MNT/bar echo "File digests before unmounting filesystem:" md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch # Evicting the inode or clearing the page cache before reading # again the file would also trigger the bug - reads were returning # all bytes in the range corresponding to the second reference to # the extent with a value of 0, but the correct data was persisted # (it was a bug exclusively in the read path). The issue happened # only if the same readpages() call targeted pages belonging to the # first and second ranges that point to the same compressed extent. _scratch_remount echo "File digests after mounting filesystem again:" # Must match the same digests we got before. md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch } echo -e "\nTesting with zlib compression..." test_clone_and_read_compressed_extent "-o compress=zlib" _scratch_unmount echo -e "\nTesting with lzo compression..." test_clone_and_read_compressed_extent "-o compress=lzo" status=0 exit Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Qu Wenruo<quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
2015-09-14 16:09:31 +08:00
bool force_bio_submit)
{
int ret = 0;
struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
unsigned int cur = pg_offset;
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
ASSERT(bio_ctrl);
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
ASSERT(pg_offset < PAGE_SIZE && size <= PAGE_SIZE &&
pg_offset + size <= PAGE_SIZE);
ASSERT(bio_ctrl->end_io_func);
if (force_bio_submit)
submit_one_bio(bio_ctrl);
while (cur < pg_offset + size) {
u32 offset = cur - pg_offset;
int added;
/* Allocate new bio if needed */
if (!bio_ctrl->bio) {
ret = alloc_new_bio(inode, bio_ctrl, wbc, opf,
disk_bytenr, offset,
page_offset(page) + cur,
compress_type);
if (ret < 0)
return ret;
}
/*
* We must go through btrfs_bio_add_page() to ensure each
* page range won't cross various boundaries.
*/
if (compress_type != BTRFS_COMPRESS_NONE)
added = btrfs_bio_add_page(bio_ctrl, page, disk_bytenr,
size - offset, pg_offset + offset,
compress_type);
else
added = btrfs_bio_add_page(bio_ctrl, page,
disk_bytenr + offset, size - offset,
pg_offset + offset, compress_type);
/* Metadata page range should never be split */
if (!is_data_inode(&inode->vfs_inode))
ASSERT(added == 0 || added == size - offset);
/* At least we added some page, update the account */
if (wbc && added)
wbc_account_cgroup_owner(wbc, page, added);
/* We have reached boundary, submit right now */
if (added < size - offset) {
/* The bio should contain some page(s) */
ASSERT(bio_ctrl->bio->bi_iter.bi_size);
submit_one_bio(bio_ctrl);
}
cur += added;
}
return 0;
}
static int attach_extent_buffer_page(struct extent_buffer *eb,
struct page *page,
struct btrfs_subpage *prealloc)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
int ret = 0;
/*
* If the page is mapped to btree inode, we should hold the private
* lock to prevent race.
* For cloned or dummy extent buffers, their pages are not mapped and
* will not race with any other ebs.
*/
if (page->mapping)
lockdep_assert_held(&page->mapping->private_lock);
if (fs_info->nodesize >= PAGE_SIZE) {
if (!PagePrivate(page))
attach_page_private(page, eb);
else
WARN_ON(page->private != (unsigned long)eb);
return 0;
}
/* Already mapped, just free prealloc */
if (PagePrivate(page)) {
btrfs_free_subpage(prealloc);
return 0;
}
if (prealloc)
/* Has preallocated memory for subpage */
attach_page_private(page, prealloc);
else
/* Do new allocation to attach subpage */
ret = btrfs_attach_subpage(fs_info, page,
BTRFS_SUBPAGE_METADATA);
return ret;
}
int set_page_extent_mapped(struct page *page)
{
struct btrfs_fs_info *fs_info;
ASSERT(page->mapping);
if (PagePrivate(page))
return 0;
fs_info = btrfs_sb(page->mapping->host->i_sb);
if (btrfs_is_subpage(fs_info, page))
return btrfs_attach_subpage(fs_info, page, BTRFS_SUBPAGE_DATA);
attach_page_private(page, (void *)EXTENT_PAGE_PRIVATE);
return 0;
}
void clear_page_extent_mapped(struct page *page)
{
struct btrfs_fs_info *fs_info;
ASSERT(page->mapping);
if (!PagePrivate(page))
return;
fs_info = btrfs_sb(page->mapping->host->i_sb);
if (btrfs_is_subpage(fs_info, page))
return btrfs_detach_subpage(fs_info, page);
detach_page_private(page);
}
static struct extent_map *
__get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
u64 start, u64 len, struct extent_map **em_cached)
{
struct extent_map *em;
if (em_cached && *em_cached) {
em = *em_cached;
if (extent_map_in_tree(em) && start >= em->start &&
start < extent_map_end(em)) {
refcount_inc(&em->refs);
return em;
}
free_extent_map(em);
*em_cached = NULL;
}
em = btrfs_get_extent(BTRFS_I(inode), page, pg_offset, start, len);
if (em_cached && !IS_ERR(em)) {
BUG_ON(*em_cached);
refcount_inc(&em->refs);
*em_cached = em;
}
return em;
}
/*
* basic readpage implementation. Locked extent state structs are inserted
* into the tree that are removed when the IO is done (by the end_io
* handlers)
* XXX JDM: This needs looking at to ensure proper page locking
* return 0 on success, otherwise return error
*/
static int btrfs_do_readpage(struct page *page, struct extent_map **em_cached,
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
struct btrfs_bio_ctrl *bio_ctrl,
blk_opf_t read_flags, u64 *prev_em_start)
{
struct inode *inode = page->mapping->host;
btrfs: integrate page status update for data read path into begin/end_page_read In btrfs data page read path, the page status update are handled in two different locations: btrfs_do_read_page() { while (cur <= end) { /* No need to read from disk */ if (HOLE/PREALLOC/INLINE){ memset(); set_extent_uptodate(); continue; } /* Read from disk */ ret = submit_extent_page(end_bio_extent_readpage); } end_bio_extent_readpage() { endio_readpage_uptodate_page_status(); } This is fine for sectorsize == PAGE_SIZE case, as for above loop we should only hit one branch and then exit. But for subpage, there is more work to be done in page status update: - Page Unlock condition Unlike regular page size == sectorsize case, we can no longer just unlock a page. Only the last reader of the page can unlock the page. This means, we can unlock the page either in the while() loop, or in the endio function. - Page uptodate condition Since we have multiple sectors to read for a page, we can only mark the full page uptodate if all sectors are uptodate. To handle both subpage and regular cases, introduce a pair of functions to help handling page status update: - begin_page_read() For regular case, it does nothing. For subpage case, it updates the reader counters so that later end_page_read() can know who is the last one to unlock the page. - end_page_read() This is just endio_readpage_uptodate_page_status() renamed. The original name is a little too long and too specific for endio. The new thing added is the condition for page unlock. Now for subpage data, we unlock the page if we're the last reader. This does not only provide the basis for subpage data read, but also hide the special handling of page read from the main read loop. Also, since we're changing how the page lock is handled, there are two existing error paths where we need to manually unlock the page before calling begin_page_read(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-02 10:28:36 +08:00
struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
u64 start = page_offset(page);
const u64 end = start + PAGE_SIZE - 1;
u64 cur = start;
u64 extent_offset;
u64 last_byte = i_size_read(inode);
u64 block_start;
struct extent_map *em;
int ret = 0;
size_t pg_offset = 0;
size_t iosize;
size_t blocksize = inode->i_sb->s_blocksize;
struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
ret = set_page_extent_mapped(page);
if (ret < 0) {
unlock_extent(tree, start, end, NULL);
btrfs: integrate page status update for data read path into begin/end_page_read In btrfs data page read path, the page status update are handled in two different locations: btrfs_do_read_page() { while (cur <= end) { /* No need to read from disk */ if (HOLE/PREALLOC/INLINE){ memset(); set_extent_uptodate(); continue; } /* Read from disk */ ret = submit_extent_page(end_bio_extent_readpage); } end_bio_extent_readpage() { endio_readpage_uptodate_page_status(); } This is fine for sectorsize == PAGE_SIZE case, as for above loop we should only hit one branch and then exit. But for subpage, there is more work to be done in page status update: - Page Unlock condition Unlike regular page size == sectorsize case, we can no longer just unlock a page. Only the last reader of the page can unlock the page. This means, we can unlock the page either in the while() loop, or in the endio function. - Page uptodate condition Since we have multiple sectors to read for a page, we can only mark the full page uptodate if all sectors are uptodate. To handle both subpage and regular cases, introduce a pair of functions to help handling page status update: - begin_page_read() For regular case, it does nothing. For subpage case, it updates the reader counters so that later end_page_read() can know who is the last one to unlock the page. - end_page_read() This is just endio_readpage_uptodate_page_status() renamed. The original name is a little too long and too specific for endio. The new thing added is the condition for page unlock. Now for subpage data, we unlock the page if we're the last reader. This does not only provide the basis for subpage data read, but also hide the special handling of page read from the main read loop. Also, since we're changing how the page lock is handled, there are two existing error paths where we need to manually unlock the page before calling begin_page_read(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-02 10:28:36 +08:00
btrfs_page_set_error(fs_info, page, start, PAGE_SIZE);
unlock_page(page);
goto out;
}
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
if (page->index == last_byte >> PAGE_SHIFT) {
size_t zero_offset = offset_in_page(last_byte);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
if (zero_offset) {
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
iosize = PAGE_SIZE - zero_offset;
btrfs: use memzero_page() instead of open coded kmap pattern There are many places where kmap/memset/kunmap patterns occur. Use the newly lifted memzero_page() to eliminate direct uses of kmap and leverage the new core functions use of kmap_local_page(). The development of this patch was aided by the following coccinelle script: // <smpl> // SPDX-License-Identifier: GPL-2.0-only // Find kmap/memset/kunmap pattern and replace with memset*page calls // // NOTE: Offsets and other expressions may be more complex than what the script // will automatically generate. Therefore a catchall rule is provided to find // the pattern which then must be evaluated by hand. // // Confidence: Low // Copyright: (C) 2021 Intel Corporation // URL: http://coccinelle.lip6.fr/ // Comments: // Options: // // Then the memset pattern // @ memset_rule1 @ expression page, V, L, Off; identifier ptr; type VP; @@ ( -VP ptr = kmap(page); | -ptr = kmap(page); | -VP ptr = kmap_atomic(page); | -ptr = kmap_atomic(page); ) <+... ( -memset(ptr, 0, L); +memzero_page(page, 0, L); | -memset(ptr + Off, 0, L); +memzero_page(page, Off, L); | -memset(ptr, V, L); +memset_page(page, V, 0, L); | -memset(ptr + Off, V, L); +memset_page(page, V, Off, L); ) ...+> ( -kunmap(page); | -kunmap_atomic(ptr); ) // Remove any pointers left unused @ depends on memset_rule1 @ identifier memset_rule1.ptr; type VP, VP1; @@ -VP ptr; ... when != ptr; ? VP1 ptr; // // Catch all // @ memset_rule2 @ expression page; identifier ptr; expression GenTo, GenSize, GenValue; type VP; @@ ( -VP ptr = kmap(page); | -ptr = kmap(page); | -VP ptr = kmap_atomic(page); | -ptr = kmap_atomic(page); ) <+... ( // // Some call sites have complex expressions within the memset/memcpy // The follow are catch alls which need to be evaluated by hand. // -memset(GenTo, 0, GenSize); +memzero_pageExtra(page, GenTo, GenSize); | -memset(GenTo, GenValue, GenSize); +memset_pageExtra(page, GenValue, GenTo, GenSize); ) ...+> ( -kunmap(page); | -kunmap_atomic(ptr); ) // Remove any pointers left unused @ depends on memset_rule2 @ identifier memset_rule2.ptr; type VP, VP1; @@ -VP ptr; ... when != ptr; ? VP1 ptr; // </smpl> Link: https://lkml.kernel.org/r/20210309212137.2610186-4-ira.weiny@intel.com Signed-off-by: Ira Weiny <ira.weiny@intel.com> Reviewed-by: David Sterba <dsterba@suse.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com> Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 09:40:07 +08:00
memzero_page(page, zero_offset, iosize);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
}
}
bio_ctrl->end_io_func = end_bio_extent_readpage;
btrfs: integrate page status update for data read path into begin/end_page_read In btrfs data page read path, the page status update are handled in two different locations: btrfs_do_read_page() { while (cur <= end) { /* No need to read from disk */ if (HOLE/PREALLOC/INLINE){ memset(); set_extent_uptodate(); continue; } /* Read from disk */ ret = submit_extent_page(end_bio_extent_readpage); } end_bio_extent_readpage() { endio_readpage_uptodate_page_status(); } This is fine for sectorsize == PAGE_SIZE case, as for above loop we should only hit one branch and then exit. But for subpage, there is more work to be done in page status update: - Page Unlock condition Unlike regular page size == sectorsize case, we can no longer just unlock a page. Only the last reader of the page can unlock the page. This means, we can unlock the page either in the while() loop, or in the endio function. - Page uptodate condition Since we have multiple sectors to read for a page, we can only mark the full page uptodate if all sectors are uptodate. To handle both subpage and regular cases, introduce a pair of functions to help handling page status update: - begin_page_read() For regular case, it does nothing. For subpage case, it updates the reader counters so that later end_page_read() can know who is the last one to unlock the page. - end_page_read() This is just endio_readpage_uptodate_page_status() renamed. The original name is a little too long and too specific for endio. The new thing added is the condition for page unlock. Now for subpage data, we unlock the page if we're the last reader. This does not only provide the basis for subpage data read, but also hide the special handling of page read from the main read loop. Also, since we're changing how the page lock is handled, there are two existing error paths where we need to manually unlock the page before calling begin_page_read(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-02 10:28:36 +08:00
begin_page_read(fs_info, page);
while (cur <= end) {
unsigned long this_bio_flag = 0;
Btrfs: fix read corruption of compressed and shared extents If a file has a range pointing to a compressed extent, followed by another range that points to the same compressed extent and a read operation attempts to read both ranges (either completely or part of them), the pages that correspond to the second range are incorrectly filled with zeroes. Consider the following example: File layout [0 - 8K] [8K - 24K] | | | | points to extent X, points to extent X, offset 4K, length of 8K offset 0, length 16K [extent X, compressed length = 4K uncompressed length = 16K] If a readpages() call spans the 2 ranges, a single bio to read the extent is submitted - extent_io.c:submit_extent_page() would only create a new bio to cover the second range pointing to the extent if the extent it points to had a different logical address than the extent associated with the first range. This has a consequence of the compressed read end io handler (compression.c:end_compressed_bio_read()) finish once the extent is decompressed into the pages covering the first range, leaving the remaining pages (belonging to the second range) filled with zeroes (done by compression.c:btrfs_clear_biovec_end()). So fix this by submitting the current bio whenever we find a range pointing to a compressed extent that was preceded by a range with a different extent map. This is the simplest solution for this corner case. Making the end io callback populate both ranges (or more, if we have multiple pointing to the same extent) is a much more complex solution since each bio is tightly coupled with a single extent map and the extent maps associated to the ranges pointing to the shared extent can have different offsets and lengths. The following test case for fstests triggers the issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_cloner rm -f $seqres.full test_clone_and_read_compressed_extent() { local mount_opts=$1 _scratch_mkfs >>$seqres.full 2>&1 _scratch_mount $mount_opts # Create a test file with a single extent that is compressed (the # data we write into it is highly compressible no matter which # compression algorithm is used, zlib or lzo). $XFS_IO_PROG -f -c "pwrite -S 0xaa 0K 4K" \ -c "pwrite -S 0xbb 4K 8K" \ -c "pwrite -S 0xcc 12K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now clone our extent into an adjacent offset. $CLONER_PROG -s $((4 * 1024)) -d $((16 * 1024)) -l $((8 * 1024)) \ $SCRATCH_MNT/foo $SCRATCH_MNT/foo # Same as before but for this file we clone the extent into a lower # file offset. $XFS_IO_PROG -f -c "pwrite -S 0xaa 8K 4K" \ -c "pwrite -S 0xbb 12K 8K" \ -c "pwrite -S 0xcc 20K 4K" \ $SCRATCH_MNT/bar | _filter_xfs_io $CLONER_PROG -s $((12 * 1024)) -d 0 -l $((8 * 1024)) \ $SCRATCH_MNT/bar $SCRATCH_MNT/bar echo "File digests before unmounting filesystem:" md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch # Evicting the inode or clearing the page cache before reading # again the file would also trigger the bug - reads were returning # all bytes in the range corresponding to the second reference to # the extent with a value of 0, but the correct data was persisted # (it was a bug exclusively in the read path). The issue happened # only if the same readpages() call targeted pages belonging to the # first and second ranges that point to the same compressed extent. _scratch_remount echo "File digests after mounting filesystem again:" # Must match the same digests we got before. md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch } echo -e "\nTesting with zlib compression..." test_clone_and_read_compressed_extent "-o compress=zlib" _scratch_unmount echo -e "\nTesting with lzo compression..." test_clone_and_read_compressed_extent "-o compress=lzo" status=0 exit Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Qu Wenruo<quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
2015-09-14 16:09:31 +08:00
bool force_bio_submit = false;
u64 disk_bytenr;
btrfs: subpage: make add_ra_bio_pages() compatible [BUG] If we remove the subpage limitation in add_ra_bio_pages(), then read a compressed extent which has part of its range in next page, like the following inode layout: 0 32K 64K 96K 128K |<--------------|-------------->| Btrfs will trigger ASSERT() in endio function: assertion failed: atomic_read(&subpage->readers) >= nbits ------------[ cut here ]------------ kernel BUG at fs/btrfs/ctree.h:3431! Internal error: Oops - BUG: 0 [#1] SMP Workqueue: btrfs-endio btrfs_work_helper [btrfs] Call trace: assertfail.constprop.0+0x28/0x2c [btrfs] btrfs_subpage_end_reader+0x148/0x14c [btrfs] end_page_read+0x8c/0x100 [btrfs] end_bio_extent_readpage+0x320/0x6b0 [btrfs] bio_endio+0x15c/0x1dc end_workqueue_fn+0x44/0x64 [btrfs] btrfs_work_helper+0x74/0x250 [btrfs] process_one_work+0x1d4/0x47c worker_thread+0x180/0x400 kthread+0x11c/0x120 ret_from_fork+0x10/0x30 ---[ end trace c8b7b552d3bb408c ]--- [CAUSE] When we read the page range [0, 64K), we find it's a compressed extent, and we will try to add extra pages in add_ra_bio_pages() to avoid reading the same compressed extent. But when we add such page into the read bio, it doesn't follow the behavior of btrfs_do_readpage() to properly set subpage::readers. This means, for page [64K, 128K), its subpage::readers is still 0. And when endio is executed on both pages, since page [64K, 128K) has 0 subpage::readers, it triggers above ASSERT() [FIX] Function add_ra_bio_pages() is far from subpage compatible, it always assume PAGE_SIZE == sectorsize, thus when it skip to next range it always just skip PAGE_SIZE. Make it subpage compatible by: - Skip to next page properly when needed If we find there is already a page cache, we need to skip to next page. For that case, we shouldn't just skip PAGE_SIZE bytes, but use @pg_index to calculate the next bytenr and continue. - Only add the page range covered by current extent map We need to calculate which range is covered by current extent map and only add that part into the read bio. - Update subpage::readers before submitting the bio - Use proper cursor other than confusing @last_offset - Calculate the missed threshold based on sector size It's no longer using missed pages, as for 64K page size, we have at most 3 pages to skip. (If aligned only 2 pages) - Add ASSERT() to make sure our bytenr is always aligned - Add comment for the function Add a special note for subpage case, as the function won't really work well for subpage cases. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-27 15:21:47 +08:00
ASSERT(IS_ALIGNED(cur, fs_info->sectorsize));
if (cur >= last_byte) {
struct extent_state *cached = NULL;
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
iosize = PAGE_SIZE - pg_offset;
btrfs: use memzero_page() instead of open coded kmap pattern There are many places where kmap/memset/kunmap patterns occur. Use the newly lifted memzero_page() to eliminate direct uses of kmap and leverage the new core functions use of kmap_local_page(). The development of this patch was aided by the following coccinelle script: // <smpl> // SPDX-License-Identifier: GPL-2.0-only // Find kmap/memset/kunmap pattern and replace with memset*page calls // // NOTE: Offsets and other expressions may be more complex than what the script // will automatically generate. Therefore a catchall rule is provided to find // the pattern which then must be evaluated by hand. // // Confidence: Low // Copyright: (C) 2021 Intel Corporation // URL: http://coccinelle.lip6.fr/ // Comments: // Options: // // Then the memset pattern // @ memset_rule1 @ expression page, V, L, Off; identifier ptr; type VP; @@ ( -VP ptr = kmap(page); | -ptr = kmap(page); | -VP ptr = kmap_atomic(page); | -ptr = kmap_atomic(page); ) <+... ( -memset(ptr, 0, L); +memzero_page(page, 0, L); | -memset(ptr + Off, 0, L); +memzero_page(page, Off, L); | -memset(ptr, V, L); +memset_page(page, V, 0, L); | -memset(ptr + Off, V, L); +memset_page(page, V, Off, L); ) ...+> ( -kunmap(page); | -kunmap_atomic(ptr); ) // Remove any pointers left unused @ depends on memset_rule1 @ identifier memset_rule1.ptr; type VP, VP1; @@ -VP ptr; ... when != ptr; ? VP1 ptr; // // Catch all // @ memset_rule2 @ expression page; identifier ptr; expression GenTo, GenSize, GenValue; type VP; @@ ( -VP ptr = kmap(page); | -ptr = kmap(page); | -VP ptr = kmap_atomic(page); | -ptr = kmap_atomic(page); ) <+... ( // // Some call sites have complex expressions within the memset/memcpy // The follow are catch alls which need to be evaluated by hand. // -memset(GenTo, 0, GenSize); +memzero_pageExtra(page, GenTo, GenSize); | -memset(GenTo, GenValue, GenSize); +memset_pageExtra(page, GenValue, GenTo, GenSize); ) ...+> ( -kunmap(page); | -kunmap_atomic(ptr); ) // Remove any pointers left unused @ depends on memset_rule2 @ identifier memset_rule2.ptr; type VP, VP1; @@ -VP ptr; ... when != ptr; ? VP1 ptr; // </smpl> Link: https://lkml.kernel.org/r/20210309212137.2610186-4-ira.weiny@intel.com Signed-off-by: Ira Weiny <ira.weiny@intel.com> Reviewed-by: David Sterba <dsterba@suse.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com> Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 09:40:07 +08:00
memzero_page(page, pg_offset, iosize);
set_extent_uptodate(tree, cur, cur + iosize - 1,
&cached, GFP_NOFS);
unlock_extent(tree, cur, cur + iosize - 1, &cached);
btrfs: integrate page status update for data read path into begin/end_page_read In btrfs data page read path, the page status update are handled in two different locations: btrfs_do_read_page() { while (cur <= end) { /* No need to read from disk */ if (HOLE/PREALLOC/INLINE){ memset(); set_extent_uptodate(); continue; } /* Read from disk */ ret = submit_extent_page(end_bio_extent_readpage); } end_bio_extent_readpage() { endio_readpage_uptodate_page_status(); } This is fine for sectorsize == PAGE_SIZE case, as for above loop we should only hit one branch and then exit. But for subpage, there is more work to be done in page status update: - Page Unlock condition Unlike regular page size == sectorsize case, we can no longer just unlock a page. Only the last reader of the page can unlock the page. This means, we can unlock the page either in the while() loop, or in the endio function. - Page uptodate condition Since we have multiple sectors to read for a page, we can only mark the full page uptodate if all sectors are uptodate. To handle both subpage and regular cases, introduce a pair of functions to help handling page status update: - begin_page_read() For regular case, it does nothing. For subpage case, it updates the reader counters so that later end_page_read() can know who is the last one to unlock the page. - end_page_read() This is just endio_readpage_uptodate_page_status() renamed. The original name is a little too long and too specific for endio. The new thing added is the condition for page unlock. Now for subpage data, we unlock the page if we're the last reader. This does not only provide the basis for subpage data read, but also hide the special handling of page read from the main read loop. Also, since we're changing how the page lock is handled, there are two existing error paths where we need to manually unlock the page before calling begin_page_read(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-02 10:28:36 +08:00
end_page_read(page, true, cur, iosize);
break;
}
em = __get_extent_map(inode, page, pg_offset, cur,
end - cur + 1, em_cached);
if (IS_ERR(em)) {
unlock_extent(tree, cur, end, NULL);
btrfs: integrate page status update for data read path into begin/end_page_read In btrfs data page read path, the page status update are handled in two different locations: btrfs_do_read_page() { while (cur <= end) { /* No need to read from disk */ if (HOLE/PREALLOC/INLINE){ memset(); set_extent_uptodate(); continue; } /* Read from disk */ ret = submit_extent_page(end_bio_extent_readpage); } end_bio_extent_readpage() { endio_readpage_uptodate_page_status(); } This is fine for sectorsize == PAGE_SIZE case, as for above loop we should only hit one branch and then exit. But for subpage, there is more work to be done in page status update: - Page Unlock condition Unlike regular page size == sectorsize case, we can no longer just unlock a page. Only the last reader of the page can unlock the page. This means, we can unlock the page either in the while() loop, or in the endio function. - Page uptodate condition Since we have multiple sectors to read for a page, we can only mark the full page uptodate if all sectors are uptodate. To handle both subpage and regular cases, introduce a pair of functions to help handling page status update: - begin_page_read() For regular case, it does nothing. For subpage case, it updates the reader counters so that later end_page_read() can know who is the last one to unlock the page. - end_page_read() This is just endio_readpage_uptodate_page_status() renamed. The original name is a little too long and too specific for endio. The new thing added is the condition for page unlock. Now for subpage data, we unlock the page if we're the last reader. This does not only provide the basis for subpage data read, but also hide the special handling of page read from the main read loop. Also, since we're changing how the page lock is handled, there are two existing error paths where we need to manually unlock the page before calling begin_page_read(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-02 10:28:36 +08:00
end_page_read(page, false, cur, end + 1 - cur);
ret = PTR_ERR(em);
break;
}
extent_offset = cur - em->start;
BUG_ON(extent_map_end(em) <= cur);
BUG_ON(end < cur);
if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
this_bio_flag = em->compress_type;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
iosize = min(extent_map_end(em) - cur, end - cur + 1);
iosize = ALIGN(iosize, blocksize);
if (this_bio_flag != BTRFS_COMPRESS_NONE)
disk_bytenr = em->block_start;
else
disk_bytenr = em->block_start + extent_offset;
block_start = em->block_start;
if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
block_start = EXTENT_MAP_HOLE;
Btrfs: fix read corruption of compressed and shared extents If a file has a range pointing to a compressed extent, followed by another range that points to the same compressed extent and a read operation attempts to read both ranges (either completely or part of them), the pages that correspond to the second range are incorrectly filled with zeroes. Consider the following example: File layout [0 - 8K] [8K - 24K] | | | | points to extent X, points to extent X, offset 4K, length of 8K offset 0, length 16K [extent X, compressed length = 4K uncompressed length = 16K] If a readpages() call spans the 2 ranges, a single bio to read the extent is submitted - extent_io.c:submit_extent_page() would only create a new bio to cover the second range pointing to the extent if the extent it points to had a different logical address than the extent associated with the first range. This has a consequence of the compressed read end io handler (compression.c:end_compressed_bio_read()) finish once the extent is decompressed into the pages covering the first range, leaving the remaining pages (belonging to the second range) filled with zeroes (done by compression.c:btrfs_clear_biovec_end()). So fix this by submitting the current bio whenever we find a range pointing to a compressed extent that was preceded by a range with a different extent map. This is the simplest solution for this corner case. Making the end io callback populate both ranges (or more, if we have multiple pointing to the same extent) is a much more complex solution since each bio is tightly coupled with a single extent map and the extent maps associated to the ranges pointing to the shared extent can have different offsets and lengths. The following test case for fstests triggers the issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_cloner rm -f $seqres.full test_clone_and_read_compressed_extent() { local mount_opts=$1 _scratch_mkfs >>$seqres.full 2>&1 _scratch_mount $mount_opts # Create a test file with a single extent that is compressed (the # data we write into it is highly compressible no matter which # compression algorithm is used, zlib or lzo). $XFS_IO_PROG -f -c "pwrite -S 0xaa 0K 4K" \ -c "pwrite -S 0xbb 4K 8K" \ -c "pwrite -S 0xcc 12K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now clone our extent into an adjacent offset. $CLONER_PROG -s $((4 * 1024)) -d $((16 * 1024)) -l $((8 * 1024)) \ $SCRATCH_MNT/foo $SCRATCH_MNT/foo # Same as before but for this file we clone the extent into a lower # file offset. $XFS_IO_PROG -f -c "pwrite -S 0xaa 8K 4K" \ -c "pwrite -S 0xbb 12K 8K" \ -c "pwrite -S 0xcc 20K 4K" \ $SCRATCH_MNT/bar | _filter_xfs_io $CLONER_PROG -s $((12 * 1024)) -d 0 -l $((8 * 1024)) \ $SCRATCH_MNT/bar $SCRATCH_MNT/bar echo "File digests before unmounting filesystem:" md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch # Evicting the inode or clearing the page cache before reading # again the file would also trigger the bug - reads were returning # all bytes in the range corresponding to the second reference to # the extent with a value of 0, but the correct data was persisted # (it was a bug exclusively in the read path). The issue happened # only if the same readpages() call targeted pages belonging to the # first and second ranges that point to the same compressed extent. _scratch_remount echo "File digests after mounting filesystem again:" # Must match the same digests we got before. md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch } echo -e "\nTesting with zlib compression..." test_clone_and_read_compressed_extent "-o compress=zlib" _scratch_unmount echo -e "\nTesting with lzo compression..." test_clone_and_read_compressed_extent "-o compress=lzo" status=0 exit Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Qu Wenruo<quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
2015-09-14 16:09:31 +08:00
/*
* If we have a file range that points to a compressed extent
* and it's followed by a consecutive file range that points
Btrfs: fix read corruption of compressed and shared extents If a file has a range pointing to a compressed extent, followed by another range that points to the same compressed extent and a read operation attempts to read both ranges (either completely or part of them), the pages that correspond to the second range are incorrectly filled with zeroes. Consider the following example: File layout [0 - 8K] [8K - 24K] | | | | points to extent X, points to extent X, offset 4K, length of 8K offset 0, length 16K [extent X, compressed length = 4K uncompressed length = 16K] If a readpages() call spans the 2 ranges, a single bio to read the extent is submitted - extent_io.c:submit_extent_page() would only create a new bio to cover the second range pointing to the extent if the extent it points to had a different logical address than the extent associated with the first range. This has a consequence of the compressed read end io handler (compression.c:end_compressed_bio_read()) finish once the extent is decompressed into the pages covering the first range, leaving the remaining pages (belonging to the second range) filled with zeroes (done by compression.c:btrfs_clear_biovec_end()). So fix this by submitting the current bio whenever we find a range pointing to a compressed extent that was preceded by a range with a different extent map. This is the simplest solution for this corner case. Making the end io callback populate both ranges (or more, if we have multiple pointing to the same extent) is a much more complex solution since each bio is tightly coupled with a single extent map and the extent maps associated to the ranges pointing to the shared extent can have different offsets and lengths. The following test case for fstests triggers the issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_cloner rm -f $seqres.full test_clone_and_read_compressed_extent() { local mount_opts=$1 _scratch_mkfs >>$seqres.full 2>&1 _scratch_mount $mount_opts # Create a test file with a single extent that is compressed (the # data we write into it is highly compressible no matter which # compression algorithm is used, zlib or lzo). $XFS_IO_PROG -f -c "pwrite -S 0xaa 0K 4K" \ -c "pwrite -S 0xbb 4K 8K" \ -c "pwrite -S 0xcc 12K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now clone our extent into an adjacent offset. $CLONER_PROG -s $((4 * 1024)) -d $((16 * 1024)) -l $((8 * 1024)) \ $SCRATCH_MNT/foo $SCRATCH_MNT/foo # Same as before but for this file we clone the extent into a lower # file offset. $XFS_IO_PROG -f -c "pwrite -S 0xaa 8K 4K" \ -c "pwrite -S 0xbb 12K 8K" \ -c "pwrite -S 0xcc 20K 4K" \ $SCRATCH_MNT/bar | _filter_xfs_io $CLONER_PROG -s $((12 * 1024)) -d 0 -l $((8 * 1024)) \ $SCRATCH_MNT/bar $SCRATCH_MNT/bar echo "File digests before unmounting filesystem:" md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch # Evicting the inode or clearing the page cache before reading # again the file would also trigger the bug - reads were returning # all bytes in the range corresponding to the second reference to # the extent with a value of 0, but the correct data was persisted # (it was a bug exclusively in the read path). The issue happened # only if the same readpages() call targeted pages belonging to the # first and second ranges that point to the same compressed extent. _scratch_remount echo "File digests after mounting filesystem again:" # Must match the same digests we got before. md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch } echo -e "\nTesting with zlib compression..." test_clone_and_read_compressed_extent "-o compress=zlib" _scratch_unmount echo -e "\nTesting with lzo compression..." test_clone_and_read_compressed_extent "-o compress=lzo" status=0 exit Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Qu Wenruo<quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
2015-09-14 16:09:31 +08:00
* to the same compressed extent (possibly with a different
* offset and/or length, so it either points to the whole extent
* or only part of it), we must make sure we do not submit a
* single bio to populate the pages for the 2 ranges because
* this makes the compressed extent read zero out the pages
* belonging to the 2nd range. Imagine the following scenario:
*
* File layout
* [0 - 8K] [8K - 24K]
* | |
* | |
* points to extent X, points to extent X,
* offset 4K, length of 8K offset 0, length 16K
*
* [extent X, compressed length = 4K uncompressed length = 16K]
*
* If the bio to read the compressed extent covers both ranges,
* it will decompress extent X into the pages belonging to the
* first range and then it will stop, zeroing out the remaining
* pages that belong to the other range that points to extent X.
* So here we make sure we submit 2 bios, one for the first
* range and another one for the third range. Both will target
* the same physical extent from disk, but we can't currently
* make the compressed bio endio callback populate the pages
* for both ranges because each compressed bio is tightly
* coupled with a single extent map, and each range can have
* an extent map with a different offset value relative to the
* uncompressed data of our extent and different lengths. This
* is a corner case so we prioritize correctness over
* non-optimal behavior (submitting 2 bios for the same extent).
*/
if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
prev_em_start && *prev_em_start != (u64)-1 &&
Btrfs: fix corruption reading shared and compressed extents after hole punching In the past we had data corruption when reading compressed extents that are shared within the same file and they are consecutive, this got fixed by commit 005efedf2c7d0 ("Btrfs: fix read corruption of compressed and shared extents") and by commit 808f80b46790f ("Btrfs: update fix for read corruption of compressed and shared extents"). However there was a case that was missing in those fixes, which is when the shared and compressed extents are referenced with a non-zero offset. The following shell script creates a reproducer for this issue: #!/bin/bash mkfs.btrfs -f /dev/sdc &> /dev/null mount -o compress /dev/sdc /mnt/sdc # Create a file with 3 consecutive compressed extents, each has an # uncompressed size of 128Kb and a compressed size of 4Kb. for ((i = 1; i <= 3; i++)); do head -c 4096 /dev/zero for ((j = 1; j <= 31; j++)); do head -c 4096 /dev/zero | tr '\0' "\377" done done > /mnt/sdc/foobar sync echo "Digest after file creation: $(md5sum /mnt/sdc/foobar)" # Clone the first extent into offsets 128K and 256K. xfs_io -c "reflink /mnt/sdc/foobar 0 128K 128K" /mnt/sdc/foobar xfs_io -c "reflink /mnt/sdc/foobar 0 256K 128K" /mnt/sdc/foobar sync echo "Digest after cloning: $(md5sum /mnt/sdc/foobar)" # Punch holes into the regions that are already full of zeroes. xfs_io -c "fpunch 0 4K" /mnt/sdc/foobar xfs_io -c "fpunch 128K 4K" /mnt/sdc/foobar xfs_io -c "fpunch 256K 4K" /mnt/sdc/foobar sync echo "Digest after hole punching: $(md5sum /mnt/sdc/foobar)" echo "Dropping page cache..." sysctl -q vm.drop_caches=1 echo "Digest after hole punching: $(md5sum /mnt/sdc/foobar)" umount /dev/sdc When running the script we get the following output: Digest after file creation: 5a0888d80d7ab1fd31c229f83a3bbcc8 /mnt/sdc/foobar linked 131072/131072 bytes at offset 131072 128 KiB, 1 ops; 0.0033 sec (36.960 MiB/sec and 295.6830 ops/sec) linked 131072/131072 bytes at offset 262144 128 KiB, 1 ops; 0.0015 sec (78.567 MiB/sec and 628.5355 ops/sec) Digest after cloning: 5a0888d80d7ab1fd31c229f83a3bbcc8 /mnt/sdc/foobar Digest after hole punching: 5a0888d80d7ab1fd31c229f83a3bbcc8 /mnt/sdc/foobar Dropping page cache... Digest after hole punching: fba694ae8664ed0c2e9ff8937e7f1484 /mnt/sdc/foobar This happens because after reading all the pages of the extent in the range from 128K to 256K for example, we read the hole at offset 256K and then when reading the page at offset 260K we don't submit the existing bio, which is responsible for filling all the page in the range 128K to 256K only, therefore adding the pages from range 260K to 384K to the existing bio and submitting it after iterating over the entire range. Once the bio completes, the uncompressed data fills only the pages in the range 128K to 256K because there's no more data read from disk, leaving the pages in the range 260K to 384K unfilled. It is just a slightly different variant of what was solved by commit 005efedf2c7d0 ("Btrfs: fix read corruption of compressed and shared extents"). Fix this by forcing a bio submit, during readpages(), whenever we find a compressed extent map for a page that is different from the extent map for the previous page or has a different starting offset (in case it's the same compressed extent), instead of the extent map's original start offset. A test case for fstests follows soon. Reported-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org> Fixes: 808f80b46790f ("Btrfs: update fix for read corruption of compressed and shared extents") Fixes: 005efedf2c7d0 ("Btrfs: fix read corruption of compressed and shared extents") Cc: stable@vger.kernel.org # 4.3+ Tested-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-14 23:17:20 +08:00
*prev_em_start != em->start)
Btrfs: fix read corruption of compressed and shared extents If a file has a range pointing to a compressed extent, followed by another range that points to the same compressed extent and a read operation attempts to read both ranges (either completely or part of them), the pages that correspond to the second range are incorrectly filled with zeroes. Consider the following example: File layout [0 - 8K] [8K - 24K] | | | | points to extent X, points to extent X, offset 4K, length of 8K offset 0, length 16K [extent X, compressed length = 4K uncompressed length = 16K] If a readpages() call spans the 2 ranges, a single bio to read the extent is submitted - extent_io.c:submit_extent_page() would only create a new bio to cover the second range pointing to the extent if the extent it points to had a different logical address than the extent associated with the first range. This has a consequence of the compressed read end io handler (compression.c:end_compressed_bio_read()) finish once the extent is decompressed into the pages covering the first range, leaving the remaining pages (belonging to the second range) filled with zeroes (done by compression.c:btrfs_clear_biovec_end()). So fix this by submitting the current bio whenever we find a range pointing to a compressed extent that was preceded by a range with a different extent map. This is the simplest solution for this corner case. Making the end io callback populate both ranges (or more, if we have multiple pointing to the same extent) is a much more complex solution since each bio is tightly coupled with a single extent map and the extent maps associated to the ranges pointing to the shared extent can have different offsets and lengths. The following test case for fstests triggers the issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_cloner rm -f $seqres.full test_clone_and_read_compressed_extent() { local mount_opts=$1 _scratch_mkfs >>$seqres.full 2>&1 _scratch_mount $mount_opts # Create a test file with a single extent that is compressed (the # data we write into it is highly compressible no matter which # compression algorithm is used, zlib or lzo). $XFS_IO_PROG -f -c "pwrite -S 0xaa 0K 4K" \ -c "pwrite -S 0xbb 4K 8K" \ -c "pwrite -S 0xcc 12K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now clone our extent into an adjacent offset. $CLONER_PROG -s $((4 * 1024)) -d $((16 * 1024)) -l $((8 * 1024)) \ $SCRATCH_MNT/foo $SCRATCH_MNT/foo # Same as before but for this file we clone the extent into a lower # file offset. $XFS_IO_PROG -f -c "pwrite -S 0xaa 8K 4K" \ -c "pwrite -S 0xbb 12K 8K" \ -c "pwrite -S 0xcc 20K 4K" \ $SCRATCH_MNT/bar | _filter_xfs_io $CLONER_PROG -s $((12 * 1024)) -d 0 -l $((8 * 1024)) \ $SCRATCH_MNT/bar $SCRATCH_MNT/bar echo "File digests before unmounting filesystem:" md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch # Evicting the inode or clearing the page cache before reading # again the file would also trigger the bug - reads were returning # all bytes in the range corresponding to the second reference to # the extent with a value of 0, but the correct data was persisted # (it was a bug exclusively in the read path). The issue happened # only if the same readpages() call targeted pages belonging to the # first and second ranges that point to the same compressed extent. _scratch_remount echo "File digests after mounting filesystem again:" # Must match the same digests we got before. md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch } echo -e "\nTesting with zlib compression..." test_clone_and_read_compressed_extent "-o compress=zlib" _scratch_unmount echo -e "\nTesting with lzo compression..." test_clone_and_read_compressed_extent "-o compress=lzo" status=0 exit Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Qu Wenruo<quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
2015-09-14 16:09:31 +08:00
force_bio_submit = true;
if (prev_em_start)
Btrfs: fix corruption reading shared and compressed extents after hole punching In the past we had data corruption when reading compressed extents that are shared within the same file and they are consecutive, this got fixed by commit 005efedf2c7d0 ("Btrfs: fix read corruption of compressed and shared extents") and by commit 808f80b46790f ("Btrfs: update fix for read corruption of compressed and shared extents"). However there was a case that was missing in those fixes, which is when the shared and compressed extents are referenced with a non-zero offset. The following shell script creates a reproducer for this issue: #!/bin/bash mkfs.btrfs -f /dev/sdc &> /dev/null mount -o compress /dev/sdc /mnt/sdc # Create a file with 3 consecutive compressed extents, each has an # uncompressed size of 128Kb and a compressed size of 4Kb. for ((i = 1; i <= 3; i++)); do head -c 4096 /dev/zero for ((j = 1; j <= 31; j++)); do head -c 4096 /dev/zero | tr '\0' "\377" done done > /mnt/sdc/foobar sync echo "Digest after file creation: $(md5sum /mnt/sdc/foobar)" # Clone the first extent into offsets 128K and 256K. xfs_io -c "reflink /mnt/sdc/foobar 0 128K 128K" /mnt/sdc/foobar xfs_io -c "reflink /mnt/sdc/foobar 0 256K 128K" /mnt/sdc/foobar sync echo "Digest after cloning: $(md5sum /mnt/sdc/foobar)" # Punch holes into the regions that are already full of zeroes. xfs_io -c "fpunch 0 4K" /mnt/sdc/foobar xfs_io -c "fpunch 128K 4K" /mnt/sdc/foobar xfs_io -c "fpunch 256K 4K" /mnt/sdc/foobar sync echo "Digest after hole punching: $(md5sum /mnt/sdc/foobar)" echo "Dropping page cache..." sysctl -q vm.drop_caches=1 echo "Digest after hole punching: $(md5sum /mnt/sdc/foobar)" umount /dev/sdc When running the script we get the following output: Digest after file creation: 5a0888d80d7ab1fd31c229f83a3bbcc8 /mnt/sdc/foobar linked 131072/131072 bytes at offset 131072 128 KiB, 1 ops; 0.0033 sec (36.960 MiB/sec and 295.6830 ops/sec) linked 131072/131072 bytes at offset 262144 128 KiB, 1 ops; 0.0015 sec (78.567 MiB/sec and 628.5355 ops/sec) Digest after cloning: 5a0888d80d7ab1fd31c229f83a3bbcc8 /mnt/sdc/foobar Digest after hole punching: 5a0888d80d7ab1fd31c229f83a3bbcc8 /mnt/sdc/foobar Dropping page cache... Digest after hole punching: fba694ae8664ed0c2e9ff8937e7f1484 /mnt/sdc/foobar This happens because after reading all the pages of the extent in the range from 128K to 256K for example, we read the hole at offset 256K and then when reading the page at offset 260K we don't submit the existing bio, which is responsible for filling all the page in the range 128K to 256K only, therefore adding the pages from range 260K to 384K to the existing bio and submitting it after iterating over the entire range. Once the bio completes, the uncompressed data fills only the pages in the range 128K to 256K because there's no more data read from disk, leaving the pages in the range 260K to 384K unfilled. It is just a slightly different variant of what was solved by commit 005efedf2c7d0 ("Btrfs: fix read corruption of compressed and shared extents"). Fix this by forcing a bio submit, during readpages(), whenever we find a compressed extent map for a page that is different from the extent map for the previous page or has a different starting offset (in case it's the same compressed extent), instead of the extent map's original start offset. A test case for fstests follows soon. Reported-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org> Fixes: 808f80b46790f ("Btrfs: update fix for read corruption of compressed and shared extents") Fixes: 005efedf2c7d0 ("Btrfs: fix read corruption of compressed and shared extents") Cc: stable@vger.kernel.org # 4.3+ Tested-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-14 23:17:20 +08:00
*prev_em_start = em->start;
Btrfs: fix read corruption of compressed and shared extents If a file has a range pointing to a compressed extent, followed by another range that points to the same compressed extent and a read operation attempts to read both ranges (either completely or part of them), the pages that correspond to the second range are incorrectly filled with zeroes. Consider the following example: File layout [0 - 8K] [8K - 24K] | | | | points to extent X, points to extent X, offset 4K, length of 8K offset 0, length 16K [extent X, compressed length = 4K uncompressed length = 16K] If a readpages() call spans the 2 ranges, a single bio to read the extent is submitted - extent_io.c:submit_extent_page() would only create a new bio to cover the second range pointing to the extent if the extent it points to had a different logical address than the extent associated with the first range. This has a consequence of the compressed read end io handler (compression.c:end_compressed_bio_read()) finish once the extent is decompressed into the pages covering the first range, leaving the remaining pages (belonging to the second range) filled with zeroes (done by compression.c:btrfs_clear_biovec_end()). So fix this by submitting the current bio whenever we find a range pointing to a compressed extent that was preceded by a range with a different extent map. This is the simplest solution for this corner case. Making the end io callback populate both ranges (or more, if we have multiple pointing to the same extent) is a much more complex solution since each bio is tightly coupled with a single extent map and the extent maps associated to the ranges pointing to the shared extent can have different offsets and lengths. The following test case for fstests triggers the issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_cloner rm -f $seqres.full test_clone_and_read_compressed_extent() { local mount_opts=$1 _scratch_mkfs >>$seqres.full 2>&1 _scratch_mount $mount_opts # Create a test file with a single extent that is compressed (the # data we write into it is highly compressible no matter which # compression algorithm is used, zlib or lzo). $XFS_IO_PROG -f -c "pwrite -S 0xaa 0K 4K" \ -c "pwrite -S 0xbb 4K 8K" \ -c "pwrite -S 0xcc 12K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now clone our extent into an adjacent offset. $CLONER_PROG -s $((4 * 1024)) -d $((16 * 1024)) -l $((8 * 1024)) \ $SCRATCH_MNT/foo $SCRATCH_MNT/foo # Same as before but for this file we clone the extent into a lower # file offset. $XFS_IO_PROG -f -c "pwrite -S 0xaa 8K 4K" \ -c "pwrite -S 0xbb 12K 8K" \ -c "pwrite -S 0xcc 20K 4K" \ $SCRATCH_MNT/bar | _filter_xfs_io $CLONER_PROG -s $((12 * 1024)) -d 0 -l $((8 * 1024)) \ $SCRATCH_MNT/bar $SCRATCH_MNT/bar echo "File digests before unmounting filesystem:" md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch # Evicting the inode or clearing the page cache before reading # again the file would also trigger the bug - reads were returning # all bytes in the range corresponding to the second reference to # the extent with a value of 0, but the correct data was persisted # (it was a bug exclusively in the read path). The issue happened # only if the same readpages() call targeted pages belonging to the # first and second ranges that point to the same compressed extent. _scratch_remount echo "File digests after mounting filesystem again:" # Must match the same digests we got before. md5sum $SCRATCH_MNT/foo | _filter_scratch md5sum $SCRATCH_MNT/bar | _filter_scratch } echo -e "\nTesting with zlib compression..." test_clone_and_read_compressed_extent "-o compress=zlib" _scratch_unmount echo -e "\nTesting with lzo compression..." test_clone_and_read_compressed_extent "-o compress=lzo" status=0 exit Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Qu Wenruo<quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
2015-09-14 16:09:31 +08:00
free_extent_map(em);
em = NULL;
/* we've found a hole, just zero and go on */
if (block_start == EXTENT_MAP_HOLE) {
struct extent_state *cached = NULL;
btrfs: use memzero_page() instead of open coded kmap pattern There are many places where kmap/memset/kunmap patterns occur. Use the newly lifted memzero_page() to eliminate direct uses of kmap and leverage the new core functions use of kmap_local_page(). The development of this patch was aided by the following coccinelle script: // <smpl> // SPDX-License-Identifier: GPL-2.0-only // Find kmap/memset/kunmap pattern and replace with memset*page calls // // NOTE: Offsets and other expressions may be more complex than what the script // will automatically generate. Therefore a catchall rule is provided to find // the pattern which then must be evaluated by hand. // // Confidence: Low // Copyright: (C) 2021 Intel Corporation // URL: http://coccinelle.lip6.fr/ // Comments: // Options: // // Then the memset pattern // @ memset_rule1 @ expression page, V, L, Off; identifier ptr; type VP; @@ ( -VP ptr = kmap(page); | -ptr = kmap(page); | -VP ptr = kmap_atomic(page); | -ptr = kmap_atomic(page); ) <+... ( -memset(ptr, 0, L); +memzero_page(page, 0, L); | -memset(ptr + Off, 0, L); +memzero_page(page, Off, L); | -memset(ptr, V, L); +memset_page(page, V, 0, L); | -memset(ptr + Off, V, L); +memset_page(page, V, Off, L); ) ...+> ( -kunmap(page); | -kunmap_atomic(ptr); ) // Remove any pointers left unused @ depends on memset_rule1 @ identifier memset_rule1.ptr; type VP, VP1; @@ -VP ptr; ... when != ptr; ? VP1 ptr; // // Catch all // @ memset_rule2 @ expression page; identifier ptr; expression GenTo, GenSize, GenValue; type VP; @@ ( -VP ptr = kmap(page); | -ptr = kmap(page); | -VP ptr = kmap_atomic(page); | -ptr = kmap_atomic(page); ) <+... ( // // Some call sites have complex expressions within the memset/memcpy // The follow are catch alls which need to be evaluated by hand. // -memset(GenTo, 0, GenSize); +memzero_pageExtra(page, GenTo, GenSize); | -memset(GenTo, GenValue, GenSize); +memset_pageExtra(page, GenValue, GenTo, GenSize); ) ...+> ( -kunmap(page); | -kunmap_atomic(ptr); ) // Remove any pointers left unused @ depends on memset_rule2 @ identifier memset_rule2.ptr; type VP, VP1; @@ -VP ptr; ... when != ptr; ? VP1 ptr; // </smpl> Link: https://lkml.kernel.org/r/20210309212137.2610186-4-ira.weiny@intel.com Signed-off-by: Ira Weiny <ira.weiny@intel.com> Reviewed-by: David Sterba <dsterba@suse.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com> Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 09:40:07 +08:00
memzero_page(page, pg_offset, iosize);
set_extent_uptodate(tree, cur, cur + iosize - 1,
&cached, GFP_NOFS);
unlock_extent(tree, cur, cur + iosize - 1, &cached);
btrfs: integrate page status update for data read path into begin/end_page_read In btrfs data page read path, the page status update are handled in two different locations: btrfs_do_read_page() { while (cur <= end) { /* No need to read from disk */ if (HOLE/PREALLOC/INLINE){ memset(); set_extent_uptodate(); continue; } /* Read from disk */ ret = submit_extent_page(end_bio_extent_readpage); } end_bio_extent_readpage() { endio_readpage_uptodate_page_status(); } This is fine for sectorsize == PAGE_SIZE case, as for above loop we should only hit one branch and then exit. But for subpage, there is more work to be done in page status update: - Page Unlock condition Unlike regular page size == sectorsize case, we can no longer just unlock a page. Only the last reader of the page can unlock the page. This means, we can unlock the page either in the while() loop, or in the endio function. - Page uptodate condition Since we have multiple sectors to read for a page, we can only mark the full page uptodate if all sectors are uptodate. To handle both subpage and regular cases, introduce a pair of functions to help handling page status update: - begin_page_read() For regular case, it does nothing. For subpage case, it updates the reader counters so that later end_page_read() can know who is the last one to unlock the page. - end_page_read() This is just endio_readpage_uptodate_page_status() renamed. The original name is a little too long and too specific for endio. The new thing added is the condition for page unlock. Now for subpage data, we unlock the page if we're the last reader. This does not only provide the basis for subpage data read, but also hide the special handling of page read from the main read loop. Also, since we're changing how the page lock is handled, there are two existing error paths where we need to manually unlock the page before calling begin_page_read(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-02 10:28:36 +08:00
end_page_read(page, true, cur, iosize);
cur = cur + iosize;
pg_offset += iosize;
continue;
}
/* the get_extent function already copied into the page */
if (block_start == EXTENT_MAP_INLINE) {
unlock_extent(tree, cur, cur + iosize - 1, NULL);
2022-08-19 10:44:08 +08:00
end_page_read(page, true, cur, iosize);
cur = cur + iosize;
pg_offset += iosize;
continue;
}
ret = submit_extent_page(REQ_OP_READ | read_flags, NULL,
bio_ctrl, disk_bytenr, page, iosize,
pg_offset, this_bio_flag,
force_bio_submit);
if (ret) {
/*
* We have to unlock the remaining range, or the page
* will never be unlocked.
*/
unlock_extent(tree, cur, end, NULL);
end_page_read(page, false, cur, end + 1 - cur);
goto out;
}
cur = cur + iosize;
pg_offset += iosize;
}
out:
return ret;
}
Page cache changes for 5.19 - Appoint myself page cache maintainer - Fix how scsicam uses the page cache - Use the memalloc_nofs_save() API to replace AOP_FLAG_NOFS - Remove the AOP flags entirely - Remove pagecache_write_begin() and pagecache_write_end() - Documentation updates - Convert several address_space operations to use folios: - is_dirty_writeback - readpage becomes read_folio - releasepage becomes release_folio - freepage becomes free_folio - Change filler_t to require a struct file pointer be the first argument like ->read_folio -----BEGIN PGP SIGNATURE----- iQEzBAABCgAdFiEEejHryeLBw/spnjHrDpNsjXcpgj4FAmKNMDUACgkQDpNsjXcp gj4/mwf/bpHhXH4ZoNIvtUpTF6rZbqeffmc0VrbxCZDZ6igRnRPglxZ9H9v6L53O 7B0FBQIfxgNKHZpdqGdOkv8cjg/GMe/HJUbEy5wOakYPo4L9fZpHbDZ9HM2Eankj xBqLIBgBJ7doKr+Y62DAN19TVD8jfRfVtli5mqXJoNKf65J7BkxljoTH1L3EXD9d nhLAgyQjR67JQrT/39KMW+17GqLhGefLQ4YnAMONtB6TVwX/lZmigKpzVaCi4r26 bnk5vaR/3PdjtNxIoYvxdc71y2Eg05n2jEq9Wcy1AaDv/5vbyZUlZ2aBSaIVbtKX WfrhN9O3L0bU5qS7p9PoyfLc9wpq8A== =djLv -----END PGP SIGNATURE----- Merge tag 'folio-5.19' of git://git.infradead.org/users/willy/pagecache Pull page cache updates from Matthew Wilcox: - Appoint myself page cache maintainer - Fix how scsicam uses the page cache - Use the memalloc_nofs_save() API to replace AOP_FLAG_NOFS - Remove the AOP flags entirely - Remove pagecache_write_begin() and pagecache_write_end() - Documentation updates - Convert several address_space operations to use folios: - is_dirty_writeback - readpage becomes read_folio - releasepage becomes release_folio - freepage becomes free_folio - Change filler_t to require a struct file pointer be the first argument like ->read_folio * tag 'folio-5.19' of git://git.infradead.org/users/willy/pagecache: (107 commits) nilfs2: Fix some kernel-doc comments Appoint myself page cache maintainer fs: Remove aops->freepage secretmem: Convert to free_folio nfs: Convert to free_folio orangefs: Convert to free_folio fs: Add free_folio address space operation fs: Convert drop_buffers() to use a folio fs: Change try_to_free_buffers() to take a folio jbd2: Convert release_buffer_page() to use a folio jbd2: Convert jbd2_journal_try_to_free_buffers to take a folio reiserfs: Convert release_buffer_page() to use a folio fs: Remove last vestiges of releasepage ubifs: Convert to release_folio reiserfs: Convert to release_folio orangefs: Convert to release_folio ocfs2: Convert to release_folio nilfs2: Remove comment about releasepage nfs: Convert to release_folio jfs: Convert to release_folio ...
2022-05-25 10:55:07 +08:00
int btrfs_read_folio(struct file *file, struct folio *folio)
{
Page cache changes for 5.19 - Appoint myself page cache maintainer - Fix how scsicam uses the page cache - Use the memalloc_nofs_save() API to replace AOP_FLAG_NOFS - Remove the AOP flags entirely - Remove pagecache_write_begin() and pagecache_write_end() - Documentation updates - Convert several address_space operations to use folios: - is_dirty_writeback - readpage becomes read_folio - releasepage becomes release_folio - freepage becomes free_folio - Change filler_t to require a struct file pointer be the first argument like ->read_folio -----BEGIN PGP SIGNATURE----- iQEzBAABCgAdFiEEejHryeLBw/spnjHrDpNsjXcpgj4FAmKNMDUACgkQDpNsjXcp gj4/mwf/bpHhXH4ZoNIvtUpTF6rZbqeffmc0VrbxCZDZ6igRnRPglxZ9H9v6L53O 7B0FBQIfxgNKHZpdqGdOkv8cjg/GMe/HJUbEy5wOakYPo4L9fZpHbDZ9HM2Eankj xBqLIBgBJ7doKr+Y62DAN19TVD8jfRfVtli5mqXJoNKf65J7BkxljoTH1L3EXD9d nhLAgyQjR67JQrT/39KMW+17GqLhGefLQ4YnAMONtB6TVwX/lZmigKpzVaCi4r26 bnk5vaR/3PdjtNxIoYvxdc71y2Eg05n2jEq9Wcy1AaDv/5vbyZUlZ2aBSaIVbtKX WfrhN9O3L0bU5qS7p9PoyfLc9wpq8A== =djLv -----END PGP SIGNATURE----- Merge tag 'folio-5.19' of git://git.infradead.org/users/willy/pagecache Pull page cache updates from Matthew Wilcox: - Appoint myself page cache maintainer - Fix how scsicam uses the page cache - Use the memalloc_nofs_save() API to replace AOP_FLAG_NOFS - Remove the AOP flags entirely - Remove pagecache_write_begin() and pagecache_write_end() - Documentation updates - Convert several address_space operations to use folios: - is_dirty_writeback - readpage becomes read_folio - releasepage becomes release_folio - freepage becomes free_folio - Change filler_t to require a struct file pointer be the first argument like ->read_folio * tag 'folio-5.19' of git://git.infradead.org/users/willy/pagecache: (107 commits) nilfs2: Fix some kernel-doc comments Appoint myself page cache maintainer fs: Remove aops->freepage secretmem: Convert to free_folio nfs: Convert to free_folio orangefs: Convert to free_folio fs: Add free_folio address space operation fs: Convert drop_buffers() to use a folio fs: Change try_to_free_buffers() to take a folio jbd2: Convert release_buffer_page() to use a folio jbd2: Convert jbd2_journal_try_to_free_buffers to take a folio reiserfs: Convert release_buffer_page() to use a folio fs: Remove last vestiges of releasepage ubifs: Convert to release_folio reiserfs: Convert to release_folio orangefs: Convert to release_folio ocfs2: Convert to release_folio nilfs2: Remove comment about releasepage nfs: Convert to release_folio jfs: Convert to release_folio ...
2022-05-25 10:55:07 +08:00
struct page *page = &folio->page;
struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
u64 start = page_offset(page);
u64 end = start + PAGE_SIZE - 1;
struct btrfs_bio_ctrl bio_ctrl = { 0 };
int ret;
btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
ret = btrfs_do_readpage(page, NULL, &bio_ctrl, 0, NULL);
/*
* If btrfs_do_readpage() failed we will want to submit the assembled
* bio to do the cleanup.
*/
submit_one_bio(&bio_ctrl);
return ret;
}
static inline void contiguous_readpages(struct page *pages[], int nr_pages,
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
u64 start, u64 end,
struct extent_map **em_cached,
struct btrfs_bio_ctrl *bio_ctrl,
u64 *prev_em_start)
{
struct btrfs_inode *inode = BTRFS_I(pages[0]->mapping->host);
int index;
btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
for (index = 0; index < nr_pages; index++) {
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
btrfs_do_readpage(pages[index], em_cached, bio_ctrl,
REQ_RAHEAD, prev_em_start);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
put_page(pages[index]);
}
}
/*
* helper for __extent_writepage, doing all of the delayed allocation setup.
*
* This returns 1 if btrfs_run_delalloc_range function did all the work required
* to write the page (copy into inline extent). In this case the IO has
* been started and the page is already unlocked.
*
* This returns 0 if all went well (page still locked)
* This returns < 0 if there were errors (page still locked)
*/
static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
struct page *page, struct writeback_control *wbc)
{
btrfs: subpage: avoid potential deadlock with compression and delalloc [BUG] With experimental subpage compression enabled, a simple fsstress can lead to self deadlock on page 720896: mkfs.btrfs -f -s 4k $dev > /dev/null mount $dev -o compress $mnt $fsstress -p 1 -n 100 -w -d $mnt -v -s 1625511156 [CAUSE] If we have a file layout looks like below: 0 32K 64K 96K 128K |//| |///////////////| 4K Then we run delalloc range for the inode, it will: - Call find_lock_delalloc_range() with @delalloc_start = 0 Then we got a delalloc range [0, 4K). This range will be COWed. - Call find_lock_delalloc_range() again with @delalloc_start = 4K Since find_lock_delalloc_range() never cares whether the range is still inside page range [0, 64K), it will return range [64K, 128K). This range meets the condition for subpage compression, will go through async COW path. And async COW path will return @page_started. But that @page_started is now for range [64K, 128K), not for range [0, 64K). - writepage_dellloc() returned 1 for page [0, 64K) Thus page [0, 64K) will not be unlocked, nor its page dirty status will be cleared. Next time when we try to lock page [0, 64K) we will deadlock, as there is no one to release page [0, 64K). This problem will never happen for regular page size as one page only contains one sector. After the first find_lock_delalloc_range() call, the @delalloc_end will go beyond @page_end no matter if we found a delalloc range or not Thus this bug only happens for subpage, as now we need multiple runs to exhaust the delalloc range of a page. [FIX] Fix the problem by ensuring the delalloc range we ran at least started inside @locked_page. So that we will never get incorrect @page_started. And to prevent such problem from happening again: - Make find_lock_delalloc_range() return false if the found range is beyond @end value passed in. Since @end will be utilized now, add an ASSERT() to ensure we pass correct @end into find_lock_delalloc_range(). This also means, for selftests we needs to populate @end before calling find_lock_delalloc_range(). - New ASSERT() in find_lock_delalloc_range() Now we will make sure the @start/@end passed in at least covers part of the page. - New ASSERT() in run_delalloc_range() To make sure the range at least starts inside @locked page. - Use @delalloc_start as proper cursor, while @delalloc_end is always reset to @page_end. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-27 15:22:07 +08:00
const u64 page_end = page_offset(page) + PAGE_SIZE - 1;
u64 delalloc_start = page_offset(page);
u64 delalloc_to_write = 0;
/* How many pages are started by btrfs_run_delalloc_range() */
unsigned long nr_written = 0;
int ret;
int page_started = 0;
btrfs: subpage: avoid potential deadlock with compression and delalloc [BUG] With experimental subpage compression enabled, a simple fsstress can lead to self deadlock on page 720896: mkfs.btrfs -f -s 4k $dev > /dev/null mount $dev -o compress $mnt $fsstress -p 1 -n 100 -w -d $mnt -v -s 1625511156 [CAUSE] If we have a file layout looks like below: 0 32K 64K 96K 128K |//| |///////////////| 4K Then we run delalloc range for the inode, it will: - Call find_lock_delalloc_range() with @delalloc_start = 0 Then we got a delalloc range [0, 4K). This range will be COWed. - Call find_lock_delalloc_range() again with @delalloc_start = 4K Since find_lock_delalloc_range() never cares whether the range is still inside page range [0, 64K), it will return range [64K, 128K). This range meets the condition for subpage compression, will go through async COW path. And async COW path will return @page_started. But that @page_started is now for range [64K, 128K), not for range [0, 64K). - writepage_dellloc() returned 1 for page [0, 64K) Thus page [0, 64K) will not be unlocked, nor its page dirty status will be cleared. Next time when we try to lock page [0, 64K) we will deadlock, as there is no one to release page [0, 64K). This problem will never happen for regular page size as one page only contains one sector. After the first find_lock_delalloc_range() call, the @delalloc_end will go beyond @page_end no matter if we found a delalloc range or not Thus this bug only happens for subpage, as now we need multiple runs to exhaust the delalloc range of a page. [FIX] Fix the problem by ensuring the delalloc range we ran at least started inside @locked_page. So that we will never get incorrect @page_started. And to prevent such problem from happening again: - Make find_lock_delalloc_range() return false if the found range is beyond @end value passed in. Since @end will be utilized now, add an ASSERT() to ensure we pass correct @end into find_lock_delalloc_range(). This also means, for selftests we needs to populate @end before calling find_lock_delalloc_range(). - New ASSERT() in find_lock_delalloc_range() Now we will make sure the @start/@end passed in at least covers part of the page. - New ASSERT() in run_delalloc_range() To make sure the range at least starts inside @locked page. - Use @delalloc_start as proper cursor, while @delalloc_end is always reset to @page_end. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-27 15:22:07 +08:00
while (delalloc_start < page_end) {
u64 delalloc_end = page_end;
bool found;
found = find_lock_delalloc_range(&inode->vfs_inode, page,
&delalloc_start,
&delalloc_end);
if (!found) {
delalloc_start = delalloc_end + 1;
continue;
}
ret = btrfs_run_delalloc_range(inode, page, delalloc_start,
delalloc_end, &page_started, &nr_written, wbc);
if (ret) {
btrfs: unify regular and subpage error paths in __extent_writepage() [BUG] When running btrfs/160 in a loop for subpage with experimental compression support, it has a high chance to crash (~20%): BTRFS critical (device dm-7): panic in __btrfs_add_ordered_extent:238: inconsistency in ordered tree at offset 0 (errno=-17 Object already exists) ------------[ cut here ]------------ kernel BUG at fs/btrfs/ordered-data.c:238! Internal error: Oops - BUG: 0 [#1] SMP pc : __btrfs_add_ordered_extent+0x550/0x670 [btrfs] lr : __btrfs_add_ordered_extent+0x550/0x670 [btrfs] Call trace: __btrfs_add_ordered_extent+0x550/0x670 [btrfs] btrfs_add_ordered_extent+0x2c/0x50 [btrfs] run_delalloc_nocow+0x81c/0x8fc [btrfs] btrfs_run_delalloc_range+0xa4/0x390 [btrfs] writepage_delalloc+0xc0/0x1ac [btrfs] __extent_writepage+0xf4/0x370 [btrfs] extent_write_cache_pages+0x288/0x4f4 [btrfs] extent_writepages+0x58/0xe0 [btrfs] btrfs_writepages+0x1c/0x30 [btrfs] do_writepages+0x60/0x110 __filemap_fdatawrite_range+0x108/0x170 filemap_fdatawrite_range+0x20/0x30 btrfs_fdatawrite_range+0x34/0x4dc [btrfs] __btrfs_write_out_cache+0x34c/0x480 [btrfs] btrfs_write_out_cache+0x144/0x220 [btrfs] btrfs_start_dirty_block_groups+0x3ac/0x6b0 [btrfs] btrfs_commit_transaction+0xd0/0xbb4 [btrfs] btrfs_sync_fs+0x64/0x1cc [btrfs] sync_fs_one_sb+0x3c/0x50 iterate_supers+0xcc/0x1d4 ksys_sync+0x6c/0xd0 __arm64_sys_sync+0x1c/0x30 invoke_syscall+0x50/0x120 el0_svc_common.constprop.0+0x4c/0xd4 do_el0_svc+0x30/0x9c el0_svc+0x2c/0x54 el0_sync_handler+0x1a8/0x1b0 el0_sync+0x198/0x1c0 ---[ end trace 336f67369ae6e0af ]--- [CAUSE] For subpage case, we can have multiple sectors inside a page, this makes it possible for __extent_writepage() to have part of its page submitted before returning. In btrfs/160, we are using dm-dust to emulate write error, this means for certain pages, we could have everything running fine, but at the end of __extent_writepage(), one of the submitted bios fails due to dm-dust. Then the page is marked Error, and we change @ret from 0 to -EIO. This makes the caller extent_write_cache_pages() to error out, without submitting the remaining pages. Furthermore, since we're erroring out for free space cache, it doesn't really care about the error and will update the inode and retry the writeback. Then we re-run the delalloc range, and will try to insert the same delalloc range while previous delalloc range is still hanging there, triggering the above error. [FIX] The proper fix is to handle errors from __extent_writepage() properly, by ending the remaining ordered extent. But that fix needs the following changes: - Know at exactly which sector the error happened Currently __extent_writepage_io() works for the full page, can't return at which sector we hit the error. - Grab the ordered extent covering the failed sector As a hotfix for subpage case, here we unify the error paths in __extent_writepage(). In fact, the "if (PageError(page))" branch never get executed if @ret is still 0 for non-subpage cases. As for non-subpage case, we never submit current page in __extent_writepage(), but only add current page into bio. The bio can only get submitted in next page. Thus we never get PageError() set due to IO failure, thus when we hit the branch, @ret is never 0. By simply removing that @ret assignment, we let subpage case ignore the IO failure, thus only error out for fatal errors just like regular sectorsize. So that IO error won't be treated as fatal error not trigger the hanging OE problem. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-26 14:35:07 +08:00
btrfs_page_set_error(inode->root->fs_info, page,
page_offset(page), PAGE_SIZE);
return ret;
}
/*
* delalloc_end is already one less than the total length, so
* we don't subtract one from PAGE_SIZE
*/
delalloc_to_write += (delalloc_end - delalloc_start +
PAGE_SIZE) >> PAGE_SHIFT;
delalloc_start = delalloc_end + 1;
}
if (wbc->nr_to_write < delalloc_to_write) {
int thresh = 8192;
if (delalloc_to_write < thresh * 2)
thresh = delalloc_to_write;
wbc->nr_to_write = min_t(u64, delalloc_to_write,
thresh);
}
/* Did btrfs_run_dealloc_range() already unlock and start the IO? */
if (page_started) {
/*
* We've unlocked the page, so we can't update the mapping's
* writeback index, just update nr_to_write.
*/
wbc->nr_to_write -= nr_written;
return 1;
}
return 0;
}
/*
* Find the first byte we need to write.
*
* For subpage, one page can contain several sectors, and
* __extent_writepage_io() will just grab all extent maps in the page
* range and try to submit all non-inline/non-compressed extents.
*
* This is a big problem for subpage, we shouldn't re-submit already written
* data at all.
* This function will lookup subpage dirty bit to find which range we really
* need to submit.
*
* Return the next dirty range in [@start, @end).
* If no dirty range is found, @start will be page_offset(page) + PAGE_SIZE.
*/
static void find_next_dirty_byte(struct btrfs_fs_info *fs_info,
struct page *page, u64 *start, u64 *end)
{
struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
struct btrfs_subpage_info *spi = fs_info->subpage_info;
u64 orig_start = *start;
/* Declare as unsigned long so we can use bitmap ops */
unsigned long flags;
int range_start_bit;
int range_end_bit;
/*
* For regular sector size == page size case, since one page only
* contains one sector, we return the page offset directly.
*/
if (!btrfs_is_subpage(fs_info, page)) {
*start = page_offset(page);
*end = page_offset(page) + PAGE_SIZE;
return;
}
range_start_bit = spi->dirty_offset +
(offset_in_page(orig_start) >> fs_info->sectorsize_bits);
/* We should have the page locked, but just in case */
spin_lock_irqsave(&subpage->lock, flags);
bitmap_next_set_region(subpage->bitmaps, &range_start_bit, &range_end_bit,
spi->dirty_offset + spi->bitmap_nr_bits);
spin_unlock_irqrestore(&subpage->lock, flags);
range_start_bit -= spi->dirty_offset;
range_end_bit -= spi->dirty_offset;
*start = page_offset(page) + range_start_bit * fs_info->sectorsize;
*end = page_offset(page) + range_end_bit * fs_info->sectorsize;
}
/*
* helper for __extent_writepage. This calls the writepage start hooks,
* and does the loop to map the page into extents and bios.
*
* We return 1 if the IO is started and the page is unlocked,
* 0 if all went well (page still locked)
* < 0 if there were errors (page still locked)
*/
static noinline_for_stack int __extent_writepage_io(struct btrfs_inode *inode,
struct page *page,
struct writeback_control *wbc,
struct btrfs_bio_ctrl *bio_ctrl,
loff_t i_size,
int *nr_ret)
{
struct btrfs_fs_info *fs_info = inode->root->fs_info;
u64 cur = page_offset(page);
u64 end = cur + PAGE_SIZE - 1;
u64 extent_offset;
u64 block_start;
struct extent_map *em;
int saved_ret = 0;
int ret = 0;
int nr = 0;
enum req_op op = REQ_OP_WRITE;
const blk_opf_t write_flags = wbc_to_write_flags(wbc);
bool has_error = false;
bool compressed;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
ret = btrfs_writepage_cow_fixup(page);
if (ret) {
/* Fixup worker will requeue */
redirty_page_for_writepage(wbc, page);
unlock_page(page);
return 1;
}
/*
* we don't want to touch the inode after unlocking the page,
* so we update the mapping writeback index now
*/
wbc->nr_to_write--;
bio_ctrl->end_io_func = end_bio_extent_writepage;
while (cur <= end) {
u64 disk_bytenr;
u64 em_end;
u64 dirty_range_start = cur;
u64 dirty_range_end;
u32 iosize;
if (cur >= i_size) {
btrfs_writepage_endio_finish_ordered(inode, page, cur,
end, true);
btrfs: subpage: fix writeback which does not have ordered extent [BUG] When running fsstress with subpage RW support, there are random BUG_ON()s triggered with the following trace: kernel BUG at fs/btrfs/file-item.c:667! Internal error: Oops - BUG: 0 [#1] SMP CPU: 1 PID: 3486 Comm: kworker/u13:2 5.11.0-rc4-custom+ #43 Hardware name: Radxa ROCK Pi 4B (DT) Workqueue: btrfs-worker-high btrfs_work_helper [btrfs] pstate: 60000005 (nZCv daif -PAN -UAO -TCO BTYPE=--) pc : btrfs_csum_one_bio+0x420/0x4e0 [btrfs] lr : btrfs_csum_one_bio+0x400/0x4e0 [btrfs] Call trace: btrfs_csum_one_bio+0x420/0x4e0 [btrfs] btrfs_submit_bio_start+0x20/0x30 [btrfs] run_one_async_start+0x28/0x44 [btrfs] btrfs_work_helper+0x128/0x1b4 [btrfs] process_one_work+0x22c/0x430 worker_thread+0x70/0x3a0 kthread+0x13c/0x140 ret_from_fork+0x10/0x30 [CAUSE] Above BUG_ON() means there is some bio range which doesn't have ordered extent, which indeed is worth a BUG_ON(). Unlike regular sectorsize == PAGE_SIZE case, in subpage we have extra subpage dirty bitmap to record which range is dirty and should be written back. This means, if we submit bio for a subpage range, we do not only need to clear page dirty, but also need to clear subpage dirty bits. In __extent_writepage_io(), we will call btrfs_page_clear_dirty() for any range we submit a bio. But there is loophole, if we hit a range which is beyond i_size, we just call btrfs_writepage_endio_finish_ordered() to finish the ordered io, then break out, without clearing the subpage dirty. This means, if we hit above branch, the subpage dirty bits are still there, if other range of the page get dirtied and we need to writeback that page again, we will submit bio for the old range, leaving a wild bio range which doesn't have ordered extent. [FIX] Fix it by always calling btrfs_page_clear_dirty() in __extent_writepage_io(). Also to avoid such problem from happening again, add a new assert, btrfs_page_assert_not_dirty(), to make sure both page dirty and subpage dirty bits are cleared before exiting __extent_writepage_io(). Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-26 14:34:58 +08:00
/*
* This range is beyond i_size, thus we don't need to
* bother writing back.
* But we still need to clear the dirty subpage bit, or
* the next time the page gets dirtied, we will try to
* writeback the sectors with subpage dirty bits,
* causing writeback without ordered extent.
*/
btrfs_page_clear_dirty(fs_info, page, cur, end + 1 - cur);
break;
}
find_next_dirty_byte(fs_info, page, &dirty_range_start,
&dirty_range_end);
if (cur < dirty_range_start) {
cur = dirty_range_start;
continue;
}
em = btrfs_get_extent(inode, NULL, 0, cur, end - cur + 1);
if (IS_ERR(em)) {
btrfs_page_set_error(fs_info, page, cur, end - cur + 1);
Btrfs: fix hang on error (such as ENOSPC) when writing extent pages When running low on available disk space and having several processes doing buffered file IO, I got the following trace in dmesg: [ 4202.720152] INFO: task kworker/u8:1:5450 blocked for more than 120 seconds. [ 4202.720401] Not tainted 3.13.0-fdm-btrfs-next-26+ #1 [ 4202.720596] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 4202.720874] kworker/u8:1 D 0000000000000001 0 5450 2 0x00000000 [ 4202.720904] Workqueue: btrfs-flush_delalloc normal_work_helper [btrfs] [ 4202.720908] ffff8801f62ddc38 0000000000000082 ffff880203ac2490 00000000001d3f40 [ 4202.720913] ffff8801f62ddfd8 00000000001d3f40 ffff8800c4f0c920 ffff880203ac2490 [ 4202.720918] 00000000001d4a40 ffff88020fe85a40 ffff88020fe85ab8 0000000000000001 [ 4202.720922] Call Trace: [ 4202.720931] [<ffffffff816a3cb9>] schedule+0x29/0x70 [ 4202.720950] [<ffffffffa01ec48d>] btrfs_start_ordered_extent+0x6d/0x110 [btrfs] [ 4202.720956] [<ffffffff8108e620>] ? bit_waitqueue+0xc0/0xc0 [ 4202.720972] [<ffffffffa01ec559>] btrfs_run_ordered_extent_work+0x29/0x40 [btrfs] [ 4202.720988] [<ffffffffa0201987>] normal_work_helper+0x137/0x2c0 [btrfs] [ 4202.720994] [<ffffffff810680e5>] process_one_work+0x1f5/0x530 (...) [ 4202.721027] 2 locks held by kworker/u8:1/5450: [ 4202.721028] #0: (%s-%s){++++..}, at: [<ffffffff81068083>] process_one_work+0x193/0x530 [ 4202.721037] #1: ((&work->normal_work)){+.+...}, at: [<ffffffff81068083>] process_one_work+0x193/0x530 [ 4202.721054] INFO: task btrfs:7891 blocked for more than 120 seconds. [ 4202.721258] Not tainted 3.13.0-fdm-btrfs-next-26+ #1 [ 4202.721444] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 4202.721699] btrfs D 0000000000000001 0 7891 7890 0x00000001 [ 4202.721704] ffff88018c2119e8 0000000000000086 ffff8800a33d2490 00000000001d3f40 [ 4202.721710] ffff88018c211fd8 00000000001d3f40 ffff8802144b0000 ffff8800a33d2490 [ 4202.721714] ffff8800d8576640 ffff88020fe85bc0 ffff88020fe85bc8 7fffffffffffffff [ 4202.721718] Call Trace: [ 4202.721723] [<ffffffff816a3cb9>] schedule+0x29/0x70 [ 4202.721727] [<ffffffff816a2ebc>] schedule_timeout+0x1dc/0x270 [ 4202.721732] [<ffffffff8109bd79>] ? mark_held_locks+0xb9/0x140 [ 4202.721736] [<ffffffff816a90c0>] ? _raw_spin_unlock_irq+0x30/0x40 [ 4202.721740] [<ffffffff8109bf0d>] ? trace_hardirqs_on_caller+0x10d/0x1d0 [ 4202.721744] [<ffffffff816a488f>] wait_for_completion+0xdf/0x120 [ 4202.721749] [<ffffffff8107fa90>] ? try_to_wake_up+0x310/0x310 [ 4202.721765] [<ffffffffa01ebee4>] btrfs_wait_ordered_extents+0x1f4/0x280 [btrfs] [ 4202.721781] [<ffffffffa020526e>] btrfs_mksubvol.isra.62+0x30e/0x5a0 [btrfs] [ 4202.721786] [<ffffffff8108e620>] ? bit_waitqueue+0xc0/0xc0 [ 4202.721799] [<ffffffffa02056a9>] btrfs_ioctl_snap_create_transid+0x1a9/0x1b0 [btrfs] [ 4202.721813] [<ffffffffa020583a>] btrfs_ioctl_snap_create_v2+0x10a/0x170 [btrfs] (...) It turns out that extent_io.c:__extent_writepage(), which ends up being called through filemap_fdatawrite_range() in btrfs_start_ordered_extent(), was getting -ENOSPC when calling the fill_delalloc callback. In this situation, it returned without the writepage_end_io_hook callback (inode.c:btrfs_writepage_end_io_hook) ever being called for the respective page, which prevents the ordered extent's bytes_left count from ever reaching 0, and therefore a finish_ordered_fn work is never queued into the endio_write_workers queue. This makes the task that called btrfs_start_ordered_extent() hang forever on the wait queue of the ordered extent. This is fairly easy to reproduce using a small filesystem and fsstress on a quad core vm: mkfs.btrfs -f -b `expr 2100 \* 1024 \* 1024` /dev/sdd mount /dev/sdd /mnt fsstress -p 6 -d /mnt -n 100000 -x \ "btrfs subvolume snapshot -r /mnt /mnt/mysnap" \ -f allocsp=0 \ -f bulkstat=0 \ -f bulkstat1=0 \ -f chown=0 \ -f creat=1 \ -f dread=0 \ -f dwrite=0 \ -f fallocate=1 \ -f fdatasync=0 \ -f fiemap=0 \ -f freesp=0 \ -f fsync=0 \ -f getattr=0 \ -f getdents=0 \ -f link=0 \ -f mkdir=0 \ -f mknod=0 \ -f punch=1 \ -f read=0 \ -f readlink=0 \ -f rename=0 \ -f resvsp=0 \ -f rmdir=0 \ -f setxattr=0 \ -f stat=0 \ -f symlink=0 \ -f sync=0 \ -f truncate=1 \ -f unlink=0 \ -f unresvsp=0 \ -f write=4 So just ensure that if an error happens while writing the extent page we call the writepage_end_io_hook callback. Also make it return the error code and ensure the caller (extent_write_cache_pages) processes all pages in the page vector even if an error happens only for some of them, so that ordered extents end up released. Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-05-10 00:17:40 +08:00
ret = PTR_ERR_OR_ZERO(em);
has_error = true;
if (!saved_ret)
saved_ret = ret;
break;
}
extent_offset = cur - em->start;
em_end = extent_map_end(em);
ASSERT(cur <= em_end);
ASSERT(cur < end);
ASSERT(IS_ALIGNED(em->start, fs_info->sectorsize));
ASSERT(IS_ALIGNED(em->len, fs_info->sectorsize));
block_start = em->block_start;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
disk_bytenr = em->block_start + extent_offset;
/*
* Note that em_end from extent_map_end() and dirty_range_end from
* find_next_dirty_byte() are all exclusive
*/
iosize = min(min(em_end, end + 1), dirty_range_end) - cur;
if (btrfs_use_zone_append(inode, em->block_start))
op = REQ_OP_ZONE_APPEND;
free_extent_map(em);
em = NULL;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
/*
* compressed and inline extents are written through other
* paths in the FS
*/
if (compressed || block_start == EXTENT_MAP_HOLE ||
block_start == EXTENT_MAP_INLINE) {
if (compressed)
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
nr++;
else
btrfs_writepage_endio_finish_ordered(inode,
page, cur, cur + iosize - 1, true);
btrfs: subpage: fix writeback which does not have ordered extent [BUG] When running fsstress with subpage RW support, there are random BUG_ON()s triggered with the following trace: kernel BUG at fs/btrfs/file-item.c:667! Internal error: Oops - BUG: 0 [#1] SMP CPU: 1 PID: 3486 Comm: kworker/u13:2 5.11.0-rc4-custom+ #43 Hardware name: Radxa ROCK Pi 4B (DT) Workqueue: btrfs-worker-high btrfs_work_helper [btrfs] pstate: 60000005 (nZCv daif -PAN -UAO -TCO BTYPE=--) pc : btrfs_csum_one_bio+0x420/0x4e0 [btrfs] lr : btrfs_csum_one_bio+0x400/0x4e0 [btrfs] Call trace: btrfs_csum_one_bio+0x420/0x4e0 [btrfs] btrfs_submit_bio_start+0x20/0x30 [btrfs] run_one_async_start+0x28/0x44 [btrfs] btrfs_work_helper+0x128/0x1b4 [btrfs] process_one_work+0x22c/0x430 worker_thread+0x70/0x3a0 kthread+0x13c/0x140 ret_from_fork+0x10/0x30 [CAUSE] Above BUG_ON() means there is some bio range which doesn't have ordered extent, which indeed is worth a BUG_ON(). Unlike regular sectorsize == PAGE_SIZE case, in subpage we have extra subpage dirty bitmap to record which range is dirty and should be written back. This means, if we submit bio for a subpage range, we do not only need to clear page dirty, but also need to clear subpage dirty bits. In __extent_writepage_io(), we will call btrfs_page_clear_dirty() for any range we submit a bio. But there is loophole, if we hit a range which is beyond i_size, we just call btrfs_writepage_endio_finish_ordered() to finish the ordered io, then break out, without clearing the subpage dirty. This means, if we hit above branch, the subpage dirty bits are still there, if other range of the page get dirtied and we need to writeback that page again, we will submit bio for the old range, leaving a wild bio range which doesn't have ordered extent. [FIX] Fix it by always calling btrfs_page_clear_dirty() in __extent_writepage_io(). Also to avoid such problem from happening again, add a new assert, btrfs_page_assert_not_dirty(), to make sure both page dirty and subpage dirty bits are cleared before exiting __extent_writepage_io(). Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-26 14:34:58 +08:00
btrfs_page_clear_dirty(fs_info, page, cur, iosize);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
cur += iosize;
continue;
}
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
btrfs_set_range_writeback(inode, cur, cur + iosize - 1);
if (!PageWriteback(page)) {
btrfs_err(inode->root->fs_info,
"page %lu not writeback, cur %llu end %llu",
page->index, cur, end);
}
/*
* Although the PageDirty bit is cleared before entering this
* function, subpage dirty bit is not cleared.
* So clear subpage dirty bit here so next time we won't submit
* page for range already written to disk.
*/
btrfs_page_clear_dirty(fs_info, page, cur, iosize);
ret = submit_extent_page(op | write_flags, wbc,
bio_ctrl, disk_bytenr,
page, iosize,
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
cur - page_offset(page),
0, false);
if (ret) {
has_error = true;
if (!saved_ret)
saved_ret = ret;
btrfs_page_set_error(fs_info, page, cur, iosize);
if (PageWriteback(page))
btrfs_page_clear_writeback(fs_info, page, cur,
iosize);
}
cur += iosize;
nr++;
}
btrfs: subpage: fix writeback which does not have ordered extent [BUG] When running fsstress with subpage RW support, there are random BUG_ON()s triggered with the following trace: kernel BUG at fs/btrfs/file-item.c:667! Internal error: Oops - BUG: 0 [#1] SMP CPU: 1 PID: 3486 Comm: kworker/u13:2 5.11.0-rc4-custom+ #43 Hardware name: Radxa ROCK Pi 4B (DT) Workqueue: btrfs-worker-high btrfs_work_helper [btrfs] pstate: 60000005 (nZCv daif -PAN -UAO -TCO BTYPE=--) pc : btrfs_csum_one_bio+0x420/0x4e0 [btrfs] lr : btrfs_csum_one_bio+0x400/0x4e0 [btrfs] Call trace: btrfs_csum_one_bio+0x420/0x4e0 [btrfs] btrfs_submit_bio_start+0x20/0x30 [btrfs] run_one_async_start+0x28/0x44 [btrfs] btrfs_work_helper+0x128/0x1b4 [btrfs] process_one_work+0x22c/0x430 worker_thread+0x70/0x3a0 kthread+0x13c/0x140 ret_from_fork+0x10/0x30 [CAUSE] Above BUG_ON() means there is some bio range which doesn't have ordered extent, which indeed is worth a BUG_ON(). Unlike regular sectorsize == PAGE_SIZE case, in subpage we have extra subpage dirty bitmap to record which range is dirty and should be written back. This means, if we submit bio for a subpage range, we do not only need to clear page dirty, but also need to clear subpage dirty bits. In __extent_writepage_io(), we will call btrfs_page_clear_dirty() for any range we submit a bio. But there is loophole, if we hit a range which is beyond i_size, we just call btrfs_writepage_endio_finish_ordered() to finish the ordered io, then break out, without clearing the subpage dirty. This means, if we hit above branch, the subpage dirty bits are still there, if other range of the page get dirtied and we need to writeback that page again, we will submit bio for the old range, leaving a wild bio range which doesn't have ordered extent. [FIX] Fix it by always calling btrfs_page_clear_dirty() in __extent_writepage_io(). Also to avoid such problem from happening again, add a new assert, btrfs_page_assert_not_dirty(), to make sure both page dirty and subpage dirty bits are cleared before exiting __extent_writepage_io(). Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-26 14:34:58 +08:00
/*
* If we finish without problem, we should not only clear page dirty,
* but also empty subpage dirty bits
*/
if (!has_error)
btrfs: subpage: fix writeback which does not have ordered extent [BUG] When running fsstress with subpage RW support, there are random BUG_ON()s triggered with the following trace: kernel BUG at fs/btrfs/file-item.c:667! Internal error: Oops - BUG: 0 [#1] SMP CPU: 1 PID: 3486 Comm: kworker/u13:2 5.11.0-rc4-custom+ #43 Hardware name: Radxa ROCK Pi 4B (DT) Workqueue: btrfs-worker-high btrfs_work_helper [btrfs] pstate: 60000005 (nZCv daif -PAN -UAO -TCO BTYPE=--) pc : btrfs_csum_one_bio+0x420/0x4e0 [btrfs] lr : btrfs_csum_one_bio+0x400/0x4e0 [btrfs] Call trace: btrfs_csum_one_bio+0x420/0x4e0 [btrfs] btrfs_submit_bio_start+0x20/0x30 [btrfs] run_one_async_start+0x28/0x44 [btrfs] btrfs_work_helper+0x128/0x1b4 [btrfs] process_one_work+0x22c/0x430 worker_thread+0x70/0x3a0 kthread+0x13c/0x140 ret_from_fork+0x10/0x30 [CAUSE] Above BUG_ON() means there is some bio range which doesn't have ordered extent, which indeed is worth a BUG_ON(). Unlike regular sectorsize == PAGE_SIZE case, in subpage we have extra subpage dirty bitmap to record which range is dirty and should be written back. This means, if we submit bio for a subpage range, we do not only need to clear page dirty, but also need to clear subpage dirty bits. In __extent_writepage_io(), we will call btrfs_page_clear_dirty() for any range we submit a bio. But there is loophole, if we hit a range which is beyond i_size, we just call btrfs_writepage_endio_finish_ordered() to finish the ordered io, then break out, without clearing the subpage dirty. This means, if we hit above branch, the subpage dirty bits are still there, if other range of the page get dirtied and we need to writeback that page again, we will submit bio for the old range, leaving a wild bio range which doesn't have ordered extent. [FIX] Fix it by always calling btrfs_page_clear_dirty() in __extent_writepage_io(). Also to avoid such problem from happening again, add a new assert, btrfs_page_assert_not_dirty(), to make sure both page dirty and subpage dirty bits are cleared before exiting __extent_writepage_io(). Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-26 14:34:58 +08:00
btrfs_page_assert_not_dirty(fs_info, page);
else
ret = saved_ret;
*nr_ret = nr;
return ret;
}
/*
* the writepage semantics are similar to regular writepage. extent
* records are inserted to lock ranges in the tree, and as dirty areas
* are found, they are marked writeback. Then the lock bits are removed
* and the end_io handler clears the writeback ranges
*
* Return 0 if everything goes well.
* Return <0 for error.
*/
static int __extent_writepage(struct page *page, struct writeback_control *wbc,
struct btrfs_bio_ctrl *bio_ctrl)
{
struct folio *folio = page_folio(page);
struct inode *inode = page->mapping->host;
struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
const u64 page_start = page_offset(page);
const u64 page_end = page_start + PAGE_SIZE - 1;
int ret;
int nr = 0;
size_t pg_offset;
loff_t i_size = i_size_read(inode);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
unsigned long end_index = i_size >> PAGE_SHIFT;
trace___extent_writepage(page, inode, wbc);
WARN_ON(!PageLocked(page));
btrfs: unify regular and subpage error paths in __extent_writepage() [BUG] When running btrfs/160 in a loop for subpage with experimental compression support, it has a high chance to crash (~20%): BTRFS critical (device dm-7): panic in __btrfs_add_ordered_extent:238: inconsistency in ordered tree at offset 0 (errno=-17 Object already exists) ------------[ cut here ]------------ kernel BUG at fs/btrfs/ordered-data.c:238! Internal error: Oops - BUG: 0 [#1] SMP pc : __btrfs_add_ordered_extent+0x550/0x670 [btrfs] lr : __btrfs_add_ordered_extent+0x550/0x670 [btrfs] Call trace: __btrfs_add_ordered_extent+0x550/0x670 [btrfs] btrfs_add_ordered_extent+0x2c/0x50 [btrfs] run_delalloc_nocow+0x81c/0x8fc [btrfs] btrfs_run_delalloc_range+0xa4/0x390 [btrfs] writepage_delalloc+0xc0/0x1ac [btrfs] __extent_writepage+0xf4/0x370 [btrfs] extent_write_cache_pages+0x288/0x4f4 [btrfs] extent_writepages+0x58/0xe0 [btrfs] btrfs_writepages+0x1c/0x30 [btrfs] do_writepages+0x60/0x110 __filemap_fdatawrite_range+0x108/0x170 filemap_fdatawrite_range+0x20/0x30 btrfs_fdatawrite_range+0x34/0x4dc [btrfs] __btrfs_write_out_cache+0x34c/0x480 [btrfs] btrfs_write_out_cache+0x144/0x220 [btrfs] btrfs_start_dirty_block_groups+0x3ac/0x6b0 [btrfs] btrfs_commit_transaction+0xd0/0xbb4 [btrfs] btrfs_sync_fs+0x64/0x1cc [btrfs] sync_fs_one_sb+0x3c/0x50 iterate_supers+0xcc/0x1d4 ksys_sync+0x6c/0xd0 __arm64_sys_sync+0x1c/0x30 invoke_syscall+0x50/0x120 el0_svc_common.constprop.0+0x4c/0xd4 do_el0_svc+0x30/0x9c el0_svc+0x2c/0x54 el0_sync_handler+0x1a8/0x1b0 el0_sync+0x198/0x1c0 ---[ end trace 336f67369ae6e0af ]--- [CAUSE] For subpage case, we can have multiple sectors inside a page, this makes it possible for __extent_writepage() to have part of its page submitted before returning. In btrfs/160, we are using dm-dust to emulate write error, this means for certain pages, we could have everything running fine, but at the end of __extent_writepage(), one of the submitted bios fails due to dm-dust. Then the page is marked Error, and we change @ret from 0 to -EIO. This makes the caller extent_write_cache_pages() to error out, without submitting the remaining pages. Furthermore, since we're erroring out for free space cache, it doesn't really care about the error and will update the inode and retry the writeback. Then we re-run the delalloc range, and will try to insert the same delalloc range while previous delalloc range is still hanging there, triggering the above error. [FIX] The proper fix is to handle errors from __extent_writepage() properly, by ending the remaining ordered extent. But that fix needs the following changes: - Know at exactly which sector the error happened Currently __extent_writepage_io() works for the full page, can't return at which sector we hit the error. - Grab the ordered extent covering the failed sector As a hotfix for subpage case, here we unify the error paths in __extent_writepage(). In fact, the "if (PageError(page))" branch never get executed if @ret is still 0 for non-subpage cases. As for non-subpage case, we never submit current page in __extent_writepage(), but only add current page into bio. The bio can only get submitted in next page. Thus we never get PageError() set due to IO failure, thus when we hit the branch, @ret is never 0. By simply removing that @ret assignment, we let subpage case ignore the IO failure, thus only error out for fatal errors just like regular sectorsize. So that IO error won't be treated as fatal error not trigger the hanging OE problem. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-26 14:35:07 +08:00
btrfs_page_clear_error(btrfs_sb(inode->i_sb), page,
page_offset(page), PAGE_SIZE);
pg_offset = offset_in_page(i_size);
if (page->index > end_index ||
(page->index == end_index && !pg_offset)) {
folio_invalidate(folio, 0, folio_size(folio));
folio_unlock(folio);
return 0;
}
if (page->index == end_index)
btrfs: use memzero_page() instead of open coded kmap pattern There are many places where kmap/memset/kunmap patterns occur. Use the newly lifted memzero_page() to eliminate direct uses of kmap and leverage the new core functions use of kmap_local_page(). The development of this patch was aided by the following coccinelle script: // <smpl> // SPDX-License-Identifier: GPL-2.0-only // Find kmap/memset/kunmap pattern and replace with memset*page calls // // NOTE: Offsets and other expressions may be more complex than what the script // will automatically generate. Therefore a catchall rule is provided to find // the pattern which then must be evaluated by hand. // // Confidence: Low // Copyright: (C) 2021 Intel Corporation // URL: http://coccinelle.lip6.fr/ // Comments: // Options: // // Then the memset pattern // @ memset_rule1 @ expression page, V, L, Off; identifier ptr; type VP; @@ ( -VP ptr = kmap(page); | -ptr = kmap(page); | -VP ptr = kmap_atomic(page); | -ptr = kmap_atomic(page); ) <+... ( -memset(ptr, 0, L); +memzero_page(page, 0, L); | -memset(ptr + Off, 0, L); +memzero_page(page, Off, L); | -memset(ptr, V, L); +memset_page(page, V, 0, L); | -memset(ptr + Off, V, L); +memset_page(page, V, Off, L); ) ...+> ( -kunmap(page); | -kunmap_atomic(ptr); ) // Remove any pointers left unused @ depends on memset_rule1 @ identifier memset_rule1.ptr; type VP, VP1; @@ -VP ptr; ... when != ptr; ? VP1 ptr; // // Catch all // @ memset_rule2 @ expression page; identifier ptr; expression GenTo, GenSize, GenValue; type VP; @@ ( -VP ptr = kmap(page); | -ptr = kmap(page); | -VP ptr = kmap_atomic(page); | -ptr = kmap_atomic(page); ) <+... ( // // Some call sites have complex expressions within the memset/memcpy // The follow are catch alls which need to be evaluated by hand. // -memset(GenTo, 0, GenSize); +memzero_pageExtra(page, GenTo, GenSize); | -memset(GenTo, GenValue, GenSize); +memset_pageExtra(page, GenValue, GenTo, GenSize); ) ...+> ( -kunmap(page); | -kunmap_atomic(ptr); ) // Remove any pointers left unused @ depends on memset_rule2 @ identifier memset_rule2.ptr; type VP, VP1; @@ -VP ptr; ... when != ptr; ? VP1 ptr; // </smpl> Link: https://lkml.kernel.org/r/20210309212137.2610186-4-ira.weiny@intel.com Signed-off-by: Ira Weiny <ira.weiny@intel.com> Reviewed-by: David Sterba <dsterba@suse.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com> Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 09:40:07 +08:00
memzero_page(page, pg_offset, PAGE_SIZE - pg_offset);
ret = set_page_extent_mapped(page);
if (ret < 0) {
SetPageError(page);
goto done;
}
if (!bio_ctrl->extent_locked) {
ret = writepage_delalloc(BTRFS_I(inode), page, wbc);
if (ret == 1)
return 0;
if (ret)
goto done;
}
ret = __extent_writepage_io(BTRFS_I(inode), page, wbc, bio_ctrl, i_size,
&nr);
if (ret == 1)
return 0;
done:
if (nr == 0) {
/* make sure the mapping tag for page dirty gets cleared */
set_page_writeback(page);
end_page_writeback(page);
}
btrfs: unify regular and subpage error paths in __extent_writepage() [BUG] When running btrfs/160 in a loop for subpage with experimental compression support, it has a high chance to crash (~20%): BTRFS critical (device dm-7): panic in __btrfs_add_ordered_extent:238: inconsistency in ordered tree at offset 0 (errno=-17 Object already exists) ------------[ cut here ]------------ kernel BUG at fs/btrfs/ordered-data.c:238! Internal error: Oops - BUG: 0 [#1] SMP pc : __btrfs_add_ordered_extent+0x550/0x670 [btrfs] lr : __btrfs_add_ordered_extent+0x550/0x670 [btrfs] Call trace: __btrfs_add_ordered_extent+0x550/0x670 [btrfs] btrfs_add_ordered_extent+0x2c/0x50 [btrfs] run_delalloc_nocow+0x81c/0x8fc [btrfs] btrfs_run_delalloc_range+0xa4/0x390 [btrfs] writepage_delalloc+0xc0/0x1ac [btrfs] __extent_writepage+0xf4/0x370 [btrfs] extent_write_cache_pages+0x288/0x4f4 [btrfs] extent_writepages+0x58/0xe0 [btrfs] btrfs_writepages+0x1c/0x30 [btrfs] do_writepages+0x60/0x110 __filemap_fdatawrite_range+0x108/0x170 filemap_fdatawrite_range+0x20/0x30 btrfs_fdatawrite_range+0x34/0x4dc [btrfs] __btrfs_write_out_cache+0x34c/0x480 [btrfs] btrfs_write_out_cache+0x144/0x220 [btrfs] btrfs_start_dirty_block_groups+0x3ac/0x6b0 [btrfs] btrfs_commit_transaction+0xd0/0xbb4 [btrfs] btrfs_sync_fs+0x64/0x1cc [btrfs] sync_fs_one_sb+0x3c/0x50 iterate_supers+0xcc/0x1d4 ksys_sync+0x6c/0xd0 __arm64_sys_sync+0x1c/0x30 invoke_syscall+0x50/0x120 el0_svc_common.constprop.0+0x4c/0xd4 do_el0_svc+0x30/0x9c el0_svc+0x2c/0x54 el0_sync_handler+0x1a8/0x1b0 el0_sync+0x198/0x1c0 ---[ end trace 336f67369ae6e0af ]--- [CAUSE] For subpage case, we can have multiple sectors inside a page, this makes it possible for __extent_writepage() to have part of its page submitted before returning. In btrfs/160, we are using dm-dust to emulate write error, this means for certain pages, we could have everything running fine, but at the end of __extent_writepage(), one of the submitted bios fails due to dm-dust. Then the page is marked Error, and we change @ret from 0 to -EIO. This makes the caller extent_write_cache_pages() to error out, without submitting the remaining pages. Furthermore, since we're erroring out for free space cache, it doesn't really care about the error and will update the inode and retry the writeback. Then we re-run the delalloc range, and will try to insert the same delalloc range while previous delalloc range is still hanging there, triggering the above error. [FIX] The proper fix is to handle errors from __extent_writepage() properly, by ending the remaining ordered extent. But that fix needs the following changes: - Know at exactly which sector the error happened Currently __extent_writepage_io() works for the full page, can't return at which sector we hit the error. - Grab the ordered extent covering the failed sector As a hotfix for subpage case, here we unify the error paths in __extent_writepage(). In fact, the "if (PageError(page))" branch never get executed if @ret is still 0 for non-subpage cases. As for non-subpage case, we never submit current page in __extent_writepage(), but only add current page into bio. The bio can only get submitted in next page. Thus we never get PageError() set due to IO failure, thus when we hit the branch, @ret is never 0. By simply removing that @ret assignment, we let subpage case ignore the IO failure, thus only error out for fatal errors just like regular sectorsize. So that IO error won't be treated as fatal error not trigger the hanging OE problem. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-26 14:35:07 +08:00
/*
* Here we used to have a check for PageError() and then set @ret and
* call end_extent_writepage().
*
* But in fact setting @ret here will cause different error paths
* between subpage and regular sectorsize.
*
* For regular page size, we never submit current page, but only add
* current page to current bio.
* The bio submission can only happen in next page.
* Thus if we hit the PageError() branch, @ret is already set to
* non-zero value and will not get updated for regular sectorsize.
*
* But for subpage case, it's possible we submit part of current page,
* thus can get PageError() set by submitted bio of the same page,
* while our @ret is still 0.
*
* So here we unify the behavior and don't set @ret.
* Error can still be properly passed to higher layer as page will
* be set error, here we just don't handle the IO failure.
*
* NOTE: This is just a hotfix for subpage.
* The root fix will be properly ending ordered extent when we hit
* an error during writeback.
*
* But that needs a bigger refactoring, as we not only need to grab the
* submitted OE, but also need to know exactly at which bytenr we hit
* the error.
* Currently the full page based __extent_writepage_io() is not
* capable of that.
*/
if (PageError(page))
end_extent_writepage(page, ret, page_start, page_end);
if (bio_ctrl->extent_locked) {
/*
* If bio_ctrl->extent_locked, it's from extent_write_locked_range(),
* the page can either be locked by lock_page() or
* process_one_page().
* Let btrfs_page_unlock_writer() handle both cases.
*/
ASSERT(wbc);
btrfs_page_unlock_writer(fs_info, page, wbc->range_start,
wbc->range_end + 1 - wbc->range_start);
} else {
unlock_page(page);
}
ASSERT(ret <= 0);
return ret;
}
void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
{
sched: Remove proliferation of wait_on_bit() action functions The current "wait_on_bit" interface requires an 'action' function to be provided which does the actual waiting. There are over 20 such functions, many of them identical. Most cases can be satisfied by one of just two functions, one which uses io_schedule() and one which just uses schedule(). So: Rename wait_on_bit and wait_on_bit_lock to wait_on_bit_action and wait_on_bit_lock_action to make it explicit that they need an action function. Introduce new wait_on_bit{,_lock} and wait_on_bit{,_lock}_io which are *not* given an action function but implicitly use a standard one. The decision to error-out if a signal is pending is now made based on the 'mode' argument rather than being encoded in the action function. All instances of the old wait_on_bit and wait_on_bit_lock which can use the new version have been changed accordingly and their action functions have been discarded. wait_on_bit{_lock} does not return any specific error code in the event of a signal so the caller must check for non-zero and interpolate their own error code as appropriate. The wait_on_bit() call in __fscache_wait_on_invalidate() was ambiguous as it specified TASK_UNINTERRUPTIBLE but used fscache_wait_bit_interruptible as an action function. David Howells confirms this should be uniformly "uninterruptible" The main remaining user of wait_on_bit{,_lock}_action is NFS which needs to use a freezer-aware schedule() call. A comment in fs/gfs2/glock.c notes that having multiple 'action' functions is useful as they display differently in the 'wchan' field of 'ps'. (and /proc/$PID/wchan). As the new bit_wait{,_io} functions are tagged "__sched", they will not show up at all, but something higher in the stack. So the distinction will still be visible, only with different function names (gds2_glock_wait versus gfs2_glock_dq_wait in the gfs2/glock.c case). Since first version of this patch (against 3.15) two new action functions appeared, on in NFS and one in CIFS. CIFS also now uses an action function that makes the same freezer aware schedule call as NFS. Signed-off-by: NeilBrown <neilb@suse.de> Acked-by: David Howells <dhowells@redhat.com> (fscache, keys) Acked-by: Steven Whitehouse <swhiteho@redhat.com> (gfs2) Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Steve French <sfrench@samba.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20140707051603.28027.72349.stgit@notabene.brown Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-07 13:16:04 +08:00
wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
TASK_UNINTERRUPTIBLE);
}
Btrfs: fix unwritten extent buffers and hangs on future writeback attempts The lock_extent_buffer_io() returns 1 to the caller to tell it everything went fine and the callers needs to start writeback for the extent buffer (submit a bio, etc), 0 to tell the caller everything went fine but it does not need to start writeback for the extent buffer, and a negative value if some error happened. When it's about to return 1 it tries to lock all pages, and if a try lock on a page fails, and we didn't flush any existing bio in our "epd", it calls flush_write_bio(epd) and overwrites the return value of 1 to 0 or an error. The page might have been locked elsewhere, not with the goal of starting writeback of the extent buffer, and even by some code other than btrfs, like page migration for example, so it does not mean the writeback of the extent buffer was already started by some other task, so returning a 0 tells the caller (btree_write_cache_pages()) to not start writeback for the extent buffer. Note that epd might currently have either no bio, so flush_write_bio() returns 0 (success) or it might have a bio for another extent buffer with a lower index (logical address). Since we return 0 with the EXTENT_BUFFER_WRITEBACK bit set on the extent buffer and writeback is never started for the extent buffer, future attempts to writeback the extent buffer will hang forever waiting on that bit to be cleared, since it can only be cleared after writeback completes. Such hang is reported with a trace like the following: [49887.347053] INFO: task btrfs-transacti:1752 blocked for more than 122 seconds. [49887.347059] Not tainted 5.2.13-gentoo #2 [49887.347060] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [49887.347062] btrfs-transacti D 0 1752 2 0x80004000 [49887.347064] Call Trace: [49887.347069] ? __schedule+0x265/0x830 [49887.347071] ? bit_wait+0x50/0x50 [49887.347072] ? bit_wait+0x50/0x50 [49887.347074] schedule+0x24/0x90 [49887.347075] io_schedule+0x3c/0x60 [49887.347077] bit_wait_io+0x8/0x50 [49887.347079] __wait_on_bit+0x6c/0x80 [49887.347081] ? __lock_release.isra.29+0x155/0x2d0 [49887.347083] out_of_line_wait_on_bit+0x7b/0x80 [49887.347084] ? var_wake_function+0x20/0x20 [49887.347087] lock_extent_buffer_for_io+0x28c/0x390 [49887.347089] btree_write_cache_pages+0x18e/0x340 [49887.347091] do_writepages+0x29/0xb0 [49887.347093] ? kmem_cache_free+0x132/0x160 [49887.347095] ? convert_extent_bit+0x544/0x680 [49887.347097] filemap_fdatawrite_range+0x70/0x90 [49887.347099] btrfs_write_marked_extents+0x53/0x120 [49887.347100] btrfs_write_and_wait_transaction.isra.4+0x38/0xa0 [49887.347102] btrfs_commit_transaction+0x6bb/0x990 [49887.347103] ? start_transaction+0x33e/0x500 [49887.347105] transaction_kthread+0x139/0x15c So fix this by not overwriting the return value (ret) with the result from flush_write_bio(). We also need to clear the EXTENT_BUFFER_WRITEBACK bit in case flush_write_bio() returns an error, otherwise it will hang any future attempts to writeback the extent buffer, and undo all work done before (set back EXTENT_BUFFER_DIRTY, etc). This is a regression introduced in the 5.2 kernel. Fixes: 2e3c25136adfb ("btrfs: extent_io: add proper error handling to lock_extent_buffer_for_io()") Fixes: f4340622e0226 ("btrfs: extent_io: Move the BUG_ON() in flush_write_bio() one level up") Reported-by: Zdenek Sojka <zsojka@seznam.cz> Link: https://lore.kernel.org/linux-btrfs/GpO.2yos.3WGDOLpx6t%7D.1TUDYM@seznam.cz/T/#u Reported-by: Stefan Priebe - Profihost AG <s.priebe@profihost.ag> Link: https://lore.kernel.org/linux-btrfs/5c4688ac-10a7-fb07-70e8-c5d31a3fbb38@profihost.ag/T/#t Reported-by: Drazen Kacar <drazen.kacar@oradian.com> Link: https://lore.kernel.org/linux-btrfs/DB8PR03MB562876ECE2319B3E579590F799C80@DB8PR03MB5628.eurprd03.prod.outlook.com/ Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=204377 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-12 00:42:00 +08:00
static void end_extent_buffer_writeback(struct extent_buffer *eb)
{
clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
smp_mb__after_atomic();
wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
}
/*
* Lock extent buffer status and pages for writeback.
*
* May try to flush write bio if we can't get the lock.
*
* Return 0 if the extent buffer doesn't need to be submitted.
* (E.g. the extent buffer is not dirty)
* Return >0 is the extent buffer is submitted to bio.
* Return <0 if something went wrong, no page is locked.
*/
static noinline_for_stack int lock_extent_buffer_for_io(struct extent_buffer *eb,
struct btrfs_bio_ctrl *bio_ctrl)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
btrfs: avoid double clean up when submit_one_bio() failed [BUG] When running generic/475 with 64K page size and 4K sector size, it has a very high chance (almost 100%) to hang, with mostly data page locked but no one is going to unlock it. [CAUSE] With commit 1784b7d502a9 ("btrfs: handle csum lookup errors properly on reads"), if we failed to lookup checksum due to metadata IO error, we will return error for btrfs_submit_data_bio(). This will cause the page to be unlocked twice in btrfs_do_readpage(): btrfs_do_readpage() |- submit_extent_page() | |- submit_one_bio() | |- btrfs_submit_data_bio() | |- if (ret) { | |- bio->bi_status = ret; | |- bio_endio(bio); } | In the endio function, we will call end_page_read() | and unlock_extent() to cleanup the subpage range. | |- if (ret) { |- unlock_extent(); end_page_read() } Here we unlock the extent and cleanup the subpage range again. For unlock_extent(), it's mostly double unlock safe. But for end_page_read(), it's not, especially for subpage case, as for subpage case we will call btrfs_subpage_end_reader() to reduce the reader number, and use that to number to determine if we need to unlock the full page. If double accounted, it can underflow the number and leave the page locked without anyone to unlock it. [FIX] The commit 1784b7d502a9 ("btrfs: handle csum lookup errors properly on reads") itself is completely fine, it's our existing code not properly handling the error from bio submission hook properly. This patch will make submit_one_bio() to return void so that the callers will never be able to do cleanup when bio submission hook fails. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-04-12 20:30:13 +08:00
int i, num_pages;
int flush = 0;
int ret = 0;
if (!btrfs_try_tree_write_lock(eb)) {
submit_write_bio(bio_ctrl, 0);
flush = 1;
btrfs_tree_lock(eb);
}
if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
btrfs_tree_unlock(eb);
if (!bio_ctrl->sync_io)
return 0;
if (!flush) {
submit_write_bio(bio_ctrl, 0);
flush = 1;
}
while (1) {
wait_on_extent_buffer_writeback(eb);
btrfs_tree_lock(eb);
if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
break;
btrfs_tree_unlock(eb);
}
}
/*
* We need to do this to prevent races in people who check if the eb is
* under IO since we can end up having no IO bits set for a short period
* of time.
*/
spin_lock(&eb->refs_lock);
if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
spin_unlock(&eb->refs_lock);
btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
-eb->len,
fs_info->dirty_metadata_batch);
ret = 1;
} else {
spin_unlock(&eb->refs_lock);
}
btrfs_tree_unlock(eb);
/*
* Either we don't need to submit any tree block, or we're submitting
* subpage eb.
* Subpage metadata doesn't use page locking at all, so we can skip
* the page locking.
*/
if (!ret || fs_info->nodesize < PAGE_SIZE)
return ret;
num_pages = num_extent_pages(eb);
for (i = 0; i < num_pages; i++) {
struct page *p = eb->pages[i];
if (!trylock_page(p)) {
if (!flush) {
submit_write_bio(bio_ctrl, 0);
flush = 1;
}
lock_page(p);
}
}
return ret;
}
static void set_btree_ioerr(struct page *page, struct extent_buffer *eb)
Btrfs: be aware of btree inode write errors to avoid fs corruption While we have a transaction ongoing, the VM might decide at any time to call btree_inode->i_mapping->a_ops->writepages(), which will start writeback of dirty pages belonging to btree nodes/leafs. This call might return an error or the writeback might finish with an error before we attempt to commit the running transaction. If this happens, we might have no way of knowing that such error happened when we are committing the transaction - because the pages might no longer be marked dirty nor tagged for writeback (if a subsequent modification to the extent buffer didn't happen before the transaction commit) which makes filemap_fdata[write|wait]_range unable to find such pages (even if they're marked with SetPageError). So if this happens we must abort the transaction, otherwise we commit a super block with btree roots that point to btree nodes/leafs whose content on disk is invalid - either garbage or the content of some node/leaf from a past generation that got cowed or deleted and is no longer valid (for this later case we end up getting error messages like "parent transid verify failed on 10826481664 wanted 25748 found 29562" when reading btree nodes/leafs from disk). Note that setting and checking AS_EIO/AS_ENOSPC in the btree inode's i_mapping would not be enough because we need to distinguish between log tree extents (not fatal) vs non-log tree extents (fatal) and because the next call to filemap_fdatawait_range() will catch and clear such errors in the mapping - and that call might be from a log sync and not from a transaction commit, which means we would not know about the error at transaction commit time. Also, checking for the eb flag EXTENT_BUFFER_IOERR at transaction commit time isn't done and would not be completely reliable, as the eb might be removed from memory and read back when trying to get it, which clears that flag right before reading the eb's pages from disk, making us not know about the previous write error. Using the new 3 flags for the btree inode also makes us achieve the goal of AS_EIO/AS_ENOSPC when writepages() returns success, started writeback for all dirty pages and before filemap_fdatawait_range() is called, the writeback for all dirty pages had already finished with errors - because we were not using AS_EIO/AS_ENOSPC, filemap_fdatawait_range() would return success, as it could not know that writeback errors happened (the pages were no longer tagged for writeback). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-26 19:25:56 +08:00
{
struct btrfs_fs_info *fs_info = eb->fs_info;
Btrfs: be aware of btree inode write errors to avoid fs corruption While we have a transaction ongoing, the VM might decide at any time to call btree_inode->i_mapping->a_ops->writepages(), which will start writeback of dirty pages belonging to btree nodes/leafs. This call might return an error or the writeback might finish with an error before we attempt to commit the running transaction. If this happens, we might have no way of knowing that such error happened when we are committing the transaction - because the pages might no longer be marked dirty nor tagged for writeback (if a subsequent modification to the extent buffer didn't happen before the transaction commit) which makes filemap_fdata[write|wait]_range unable to find such pages (even if they're marked with SetPageError). So if this happens we must abort the transaction, otherwise we commit a super block with btree roots that point to btree nodes/leafs whose content on disk is invalid - either garbage or the content of some node/leaf from a past generation that got cowed or deleted and is no longer valid (for this later case we end up getting error messages like "parent transid verify failed on 10826481664 wanted 25748 found 29562" when reading btree nodes/leafs from disk). Note that setting and checking AS_EIO/AS_ENOSPC in the btree inode's i_mapping would not be enough because we need to distinguish between log tree extents (not fatal) vs non-log tree extents (fatal) and because the next call to filemap_fdatawait_range() will catch and clear such errors in the mapping - and that call might be from a log sync and not from a transaction commit, which means we would not know about the error at transaction commit time. Also, checking for the eb flag EXTENT_BUFFER_IOERR at transaction commit time isn't done and would not be completely reliable, as the eb might be removed from memory and read back when trying to get it, which clears that flag right before reading the eb's pages from disk, making us not know about the previous write error. Using the new 3 flags for the btree inode also makes us achieve the goal of AS_EIO/AS_ENOSPC when writepages() returns success, started writeback for all dirty pages and before filemap_fdatawait_range() is called, the writeback for all dirty pages had already finished with errors - because we were not using AS_EIO/AS_ENOSPC, filemap_fdatawait_range() would return success, as it could not know that writeback errors happened (the pages were no longer tagged for writeback). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-26 19:25:56 +08:00
btrfs_page_set_error(fs_info, page, eb->start, eb->len);
Btrfs: be aware of btree inode write errors to avoid fs corruption While we have a transaction ongoing, the VM might decide at any time to call btree_inode->i_mapping->a_ops->writepages(), which will start writeback of dirty pages belonging to btree nodes/leafs. This call might return an error or the writeback might finish with an error before we attempt to commit the running transaction. If this happens, we might have no way of knowing that such error happened when we are committing the transaction - because the pages might no longer be marked dirty nor tagged for writeback (if a subsequent modification to the extent buffer didn't happen before the transaction commit) which makes filemap_fdata[write|wait]_range unable to find such pages (even if they're marked with SetPageError). So if this happens we must abort the transaction, otherwise we commit a super block with btree roots that point to btree nodes/leafs whose content on disk is invalid - either garbage or the content of some node/leaf from a past generation that got cowed or deleted and is no longer valid (for this later case we end up getting error messages like "parent transid verify failed on 10826481664 wanted 25748 found 29562" when reading btree nodes/leafs from disk). Note that setting and checking AS_EIO/AS_ENOSPC in the btree inode's i_mapping would not be enough because we need to distinguish between log tree extents (not fatal) vs non-log tree extents (fatal) and because the next call to filemap_fdatawait_range() will catch and clear such errors in the mapping - and that call might be from a log sync and not from a transaction commit, which means we would not know about the error at transaction commit time. Also, checking for the eb flag EXTENT_BUFFER_IOERR at transaction commit time isn't done and would not be completely reliable, as the eb might be removed from memory and read back when trying to get it, which clears that flag right before reading the eb's pages from disk, making us not know about the previous write error. Using the new 3 flags for the btree inode also makes us achieve the goal of AS_EIO/AS_ENOSPC when writepages() returns success, started writeback for all dirty pages and before filemap_fdatawait_range() is called, the writeback for all dirty pages had already finished with errors - because we were not using AS_EIO/AS_ENOSPC, filemap_fdatawait_range() would return success, as it could not know that writeback errors happened (the pages were no longer tagged for writeback). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-26 19:25:56 +08:00
if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
return;
btrfs: clear extent buffer uptodate when we fail to write it I got dmesg errors on generic/281 on our overnight fstests. Looking at the history this happens occasionally, with errors like this WARNING: CPU: 0 PID: 673217 at fs/btrfs/extent_io.c:6848 assert_eb_page_uptodate+0x3f/0x50 CPU: 0 PID: 673217 Comm: kworker/u4:13 Tainted: G W 5.16.0-rc2+ #469 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014 Workqueue: btrfs-cache btrfs_work_helper RIP: 0010:assert_eb_page_uptodate+0x3f/0x50 RSP: 0018:ffffae598230bc60 EFLAGS: 00010246 RAX: 0017ffffc0002112 RBX: ffffebaec4100900 RCX: 0000000000001000 RDX: ffffebaec45733c7 RSI: ffffebaec4100900 RDI: ffff9fd98919f340 RBP: 0000000000000d56 R08: ffff9fd98e300000 R09: 0000000000000000 R10: 0001207370a91c50 R11: 0000000000000000 R12: 00000000000007b0 R13: ffff9fd98919f340 R14: 0000000001500000 R15: 0000000001cb0000 FS: 0000000000000000(0000) GS:ffff9fd9fbc00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007f549fcf8940 CR3: 0000000114908004 CR4: 0000000000370ef0 Call Trace: extent_buffer_test_bit+0x3f/0x70 free_space_test_bit+0xa6/0xc0 load_free_space_tree+0x1d6/0x430 caching_thread+0x454/0x630 ? rcu_read_lock_sched_held+0x12/0x60 ? rcu_read_lock_sched_held+0x12/0x60 ? rcu_read_lock_sched_held+0x12/0x60 ? lock_release+0x1f0/0x2d0 btrfs_work_helper+0xf2/0x3e0 ? lock_release+0x1f0/0x2d0 ? finish_task_switch.isra.0+0xf9/0x3a0 process_one_work+0x270/0x5a0 worker_thread+0x55/0x3c0 ? process_one_work+0x5a0/0x5a0 kthread+0x174/0x1a0 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 This happens because we're trying to read from a extent buffer page that is !PageUptodate. This happens because we will clear the page uptodate when we have an IO error, but we don't clear the extent buffer uptodate. If we do a read later and find this extent buffer we'll think its valid and not return an error, and then trip over this warning. Fix this by also clearing uptodate on the extent buffer when this happens, so that we get an error when we do a btrfs_search_slot() and find this block later. CC: stable@vger.kernel.org # 5.4+ Signed-off-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-11-25 03:14:23 +08:00
/*
* A read may stumble upon this buffer later, make sure that it gets an
* error and knows there was an error.
*/
clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
/*
* We need to set the mapping with the io error as well because a write
* error will flip the file system readonly, and then syncfs() will
* return a 0 because we are readonly if we don't modify the err seq for
* the superblock.
*/
mapping_set_error(page->mapping, -EIO);
/*
* If we error out, we should add back the dirty_metadata_bytes
* to make it consistent.
*/
percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
eb->len, fs_info->dirty_metadata_batch);
Btrfs: be aware of btree inode write errors to avoid fs corruption While we have a transaction ongoing, the VM might decide at any time to call btree_inode->i_mapping->a_ops->writepages(), which will start writeback of dirty pages belonging to btree nodes/leafs. This call might return an error or the writeback might finish with an error before we attempt to commit the running transaction. If this happens, we might have no way of knowing that such error happened when we are committing the transaction - because the pages might no longer be marked dirty nor tagged for writeback (if a subsequent modification to the extent buffer didn't happen before the transaction commit) which makes filemap_fdata[write|wait]_range unable to find such pages (even if they're marked with SetPageError). So if this happens we must abort the transaction, otherwise we commit a super block with btree roots that point to btree nodes/leafs whose content on disk is invalid - either garbage or the content of some node/leaf from a past generation that got cowed or deleted and is no longer valid (for this later case we end up getting error messages like "parent transid verify failed on 10826481664 wanted 25748 found 29562" when reading btree nodes/leafs from disk). Note that setting and checking AS_EIO/AS_ENOSPC in the btree inode's i_mapping would not be enough because we need to distinguish between log tree extents (not fatal) vs non-log tree extents (fatal) and because the next call to filemap_fdatawait_range() will catch and clear such errors in the mapping - and that call might be from a log sync and not from a transaction commit, which means we would not know about the error at transaction commit time. Also, checking for the eb flag EXTENT_BUFFER_IOERR at transaction commit time isn't done and would not be completely reliable, as the eb might be removed from memory and read back when trying to get it, which clears that flag right before reading the eb's pages from disk, making us not know about the previous write error. Using the new 3 flags for the btree inode also makes us achieve the goal of AS_EIO/AS_ENOSPC when writepages() returns success, started writeback for all dirty pages and before filemap_fdatawait_range() is called, the writeback for all dirty pages had already finished with errors - because we were not using AS_EIO/AS_ENOSPC, filemap_fdatawait_range() would return success, as it could not know that writeback errors happened (the pages were no longer tagged for writeback). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-26 19:25:56 +08:00
/*
* If writeback for a btree extent that doesn't belong to a log tree
* failed, increment the counter transaction->eb_write_errors.
* We do this because while the transaction is running and before it's
* committing (when we call filemap_fdata[write|wait]_range against
* the btree inode), we might have
* btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
* returns an error or an error happens during writeback, when we're
* committing the transaction we wouldn't know about it, since the pages
* can be no longer dirty nor marked anymore for writeback (if a
* subsequent modification to the extent buffer didn't happen before the
* transaction commit), which makes filemap_fdata[write|wait]_range not
* able to find the pages tagged with SetPageError at transaction
* commit time. So if this happens we must abort the transaction,
* otherwise we commit a super block with btree roots that point to
* btree nodes/leafs whose content on disk is invalid - either garbage
* or the content of some node/leaf from a past generation that got
* cowed or deleted and is no longer valid.
*
* Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
* not be enough - we need to distinguish between log tree extents vs
* non-log tree extents, and the next filemap_fdatawait_range() call
* will catch and clear such errors in the mapping - and that call might
* be from a log sync and not from a transaction commit. Also, checking
* for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
* not done and would not be reliable - the eb might have been released
* from memory and reading it back again means that flag would not be
* set (since it's a runtime flag, not persisted on disk).
*
* Using the flags below in the btree inode also makes us achieve the
* goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
* writeback for all dirty pages and before filemap_fdatawait_range()
* is called, the writeback for all dirty pages had already finished
* with errors - because we were not using AS_EIO/AS_ENOSPC,
* filemap_fdatawait_range() would return success, as it could not know
* that writeback errors happened (the pages were no longer tagged for
* writeback).
*/
switch (eb->log_index) {
case -1:
set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
Btrfs: be aware of btree inode write errors to avoid fs corruption While we have a transaction ongoing, the VM might decide at any time to call btree_inode->i_mapping->a_ops->writepages(), which will start writeback of dirty pages belonging to btree nodes/leafs. This call might return an error or the writeback might finish with an error before we attempt to commit the running transaction. If this happens, we might have no way of knowing that such error happened when we are committing the transaction - because the pages might no longer be marked dirty nor tagged for writeback (if a subsequent modification to the extent buffer didn't happen before the transaction commit) which makes filemap_fdata[write|wait]_range unable to find such pages (even if they're marked with SetPageError). So if this happens we must abort the transaction, otherwise we commit a super block with btree roots that point to btree nodes/leafs whose content on disk is invalid - either garbage or the content of some node/leaf from a past generation that got cowed or deleted and is no longer valid (for this later case we end up getting error messages like "parent transid verify failed on 10826481664 wanted 25748 found 29562" when reading btree nodes/leafs from disk). Note that setting and checking AS_EIO/AS_ENOSPC in the btree inode's i_mapping would not be enough because we need to distinguish between log tree extents (not fatal) vs non-log tree extents (fatal) and because the next call to filemap_fdatawait_range() will catch and clear such errors in the mapping - and that call might be from a log sync and not from a transaction commit, which means we would not know about the error at transaction commit time. Also, checking for the eb flag EXTENT_BUFFER_IOERR at transaction commit time isn't done and would not be completely reliable, as the eb might be removed from memory and read back when trying to get it, which clears that flag right before reading the eb's pages from disk, making us not know about the previous write error. Using the new 3 flags for the btree inode also makes us achieve the goal of AS_EIO/AS_ENOSPC when writepages() returns success, started writeback for all dirty pages and before filemap_fdatawait_range() is called, the writeback for all dirty pages had already finished with errors - because we were not using AS_EIO/AS_ENOSPC, filemap_fdatawait_range() would return success, as it could not know that writeback errors happened (the pages were no longer tagged for writeback). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-26 19:25:56 +08:00
break;
case 0:
set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
Btrfs: be aware of btree inode write errors to avoid fs corruption While we have a transaction ongoing, the VM might decide at any time to call btree_inode->i_mapping->a_ops->writepages(), which will start writeback of dirty pages belonging to btree nodes/leafs. This call might return an error or the writeback might finish with an error before we attempt to commit the running transaction. If this happens, we might have no way of knowing that such error happened when we are committing the transaction - because the pages might no longer be marked dirty nor tagged for writeback (if a subsequent modification to the extent buffer didn't happen before the transaction commit) which makes filemap_fdata[write|wait]_range unable to find such pages (even if they're marked with SetPageError). So if this happens we must abort the transaction, otherwise we commit a super block with btree roots that point to btree nodes/leafs whose content on disk is invalid - either garbage or the content of some node/leaf from a past generation that got cowed or deleted and is no longer valid (for this later case we end up getting error messages like "parent transid verify failed on 10826481664 wanted 25748 found 29562" when reading btree nodes/leafs from disk). Note that setting and checking AS_EIO/AS_ENOSPC in the btree inode's i_mapping would not be enough because we need to distinguish between log tree extents (not fatal) vs non-log tree extents (fatal) and because the next call to filemap_fdatawait_range() will catch and clear such errors in the mapping - and that call might be from a log sync and not from a transaction commit, which means we would not know about the error at transaction commit time. Also, checking for the eb flag EXTENT_BUFFER_IOERR at transaction commit time isn't done and would not be completely reliable, as the eb might be removed from memory and read back when trying to get it, which clears that flag right before reading the eb's pages from disk, making us not know about the previous write error. Using the new 3 flags for the btree inode also makes us achieve the goal of AS_EIO/AS_ENOSPC when writepages() returns success, started writeback for all dirty pages and before filemap_fdatawait_range() is called, the writeback for all dirty pages had already finished with errors - because we were not using AS_EIO/AS_ENOSPC, filemap_fdatawait_range() would return success, as it could not know that writeback errors happened (the pages were no longer tagged for writeback). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-26 19:25:56 +08:00
break;
case 1:
set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
Btrfs: be aware of btree inode write errors to avoid fs corruption While we have a transaction ongoing, the VM might decide at any time to call btree_inode->i_mapping->a_ops->writepages(), which will start writeback of dirty pages belonging to btree nodes/leafs. This call might return an error or the writeback might finish with an error before we attempt to commit the running transaction. If this happens, we might have no way of knowing that such error happened when we are committing the transaction - because the pages might no longer be marked dirty nor tagged for writeback (if a subsequent modification to the extent buffer didn't happen before the transaction commit) which makes filemap_fdata[write|wait]_range unable to find such pages (even if they're marked with SetPageError). So if this happens we must abort the transaction, otherwise we commit a super block with btree roots that point to btree nodes/leafs whose content on disk is invalid - either garbage or the content of some node/leaf from a past generation that got cowed or deleted and is no longer valid (for this later case we end up getting error messages like "parent transid verify failed on 10826481664 wanted 25748 found 29562" when reading btree nodes/leafs from disk). Note that setting and checking AS_EIO/AS_ENOSPC in the btree inode's i_mapping would not be enough because we need to distinguish between log tree extents (not fatal) vs non-log tree extents (fatal) and because the next call to filemap_fdatawait_range() will catch and clear such errors in the mapping - and that call might be from a log sync and not from a transaction commit, which means we would not know about the error at transaction commit time. Also, checking for the eb flag EXTENT_BUFFER_IOERR at transaction commit time isn't done and would not be completely reliable, as the eb might be removed from memory and read back when trying to get it, which clears that flag right before reading the eb's pages from disk, making us not know about the previous write error. Using the new 3 flags for the btree inode also makes us achieve the goal of AS_EIO/AS_ENOSPC when writepages() returns success, started writeback for all dirty pages and before filemap_fdatawait_range() is called, the writeback for all dirty pages had already finished with errors - because we were not using AS_EIO/AS_ENOSPC, filemap_fdatawait_range() would return success, as it could not know that writeback errors happened (the pages were no longer tagged for writeback). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-26 19:25:56 +08:00
break;
default:
BUG(); /* unexpected, logic error */
}
}
/*
* The endio specific version which won't touch any unsafe spinlock in endio
* context.
*/
static struct extent_buffer *find_extent_buffer_nolock(
struct btrfs_fs_info *fs_info, u64 start)
{
struct extent_buffer *eb;
rcu_read_lock();
eb = radix_tree_lookup(&fs_info->buffer_radix,
start >> fs_info->sectorsize_bits);
if (eb && atomic_inc_not_zero(&eb->refs)) {
rcu_read_unlock();
return eb;
}
rcu_read_unlock();
return NULL;
}
/*
* The endio function for subpage extent buffer write.
*
* Unlike end_bio_extent_buffer_writepage(), we only call end_page_writeback()
* after all extent buffers in the page has finished their writeback.
*/
static void end_bio_subpage_eb_writepage(struct btrfs_bio *bbio)
{
struct bio *bio = &bbio->bio;
struct btrfs_fs_info *fs_info;
struct bio_vec *bvec;
struct bvec_iter_all iter_all;
fs_info = btrfs_sb(bio_first_page_all(bio)->mapping->host->i_sb);
ASSERT(fs_info->nodesize < PAGE_SIZE);
ASSERT(!bio_flagged(bio, BIO_CLONED));
bio_for_each_segment_all(bvec, bio, iter_all) {
struct page *page = bvec->bv_page;
u64 bvec_start = page_offset(page) + bvec->bv_offset;
u64 bvec_end = bvec_start + bvec->bv_len - 1;
u64 cur_bytenr = bvec_start;
ASSERT(IS_ALIGNED(bvec->bv_len, fs_info->nodesize));
/* Iterate through all extent buffers in the range */
while (cur_bytenr <= bvec_end) {
struct extent_buffer *eb;
int done;
/*
* Here we can't use find_extent_buffer(), as it may
* try to lock eb->refs_lock, which is not safe in endio
* context.
*/
eb = find_extent_buffer_nolock(fs_info, cur_bytenr);
ASSERT(eb);
cur_bytenr = eb->start + eb->len;
ASSERT(test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags));
done = atomic_dec_and_test(&eb->io_pages);
ASSERT(done);
if (bio->bi_status ||
test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
ClearPageUptodate(page);
set_btree_ioerr(page, eb);
}
btrfs_subpage_clear_writeback(fs_info, page, eb->start,
eb->len);
end_extent_buffer_writeback(eb);
/*
* free_extent_buffer() will grab spinlock which is not
* safe in endio context. Thus here we manually dec
* the ref.
*/
atomic_dec(&eb->refs);
}
}
bio_put(bio);
}
static void end_bio_extent_buffer_writepage(struct btrfs_bio *bbio)
{
struct bio *bio = &bbio->bio;
struct bio_vec *bvec;
struct extent_buffer *eb;
int done;
struct bvec_iter_all iter_all;
ASSERT(!bio_flagged(bio, BIO_CLONED));
bio_for_each_segment_all(bvec, bio, iter_all) {
struct page *page = bvec->bv_page;
eb = (struct extent_buffer *)page->private;
BUG_ON(!eb);
done = atomic_dec_and_test(&eb->io_pages);
if (bio->bi_status ||
test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
ClearPageUptodate(page);
set_btree_ioerr(page, eb);
}
end_page_writeback(page);
if (!done)
continue;
end_extent_buffer_writeback(eb);
}
bio_put(bio);
}
static void prepare_eb_write(struct extent_buffer *eb)
{
u32 nritems;
unsigned long start;
unsigned long end;
clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
atomic_set(&eb->io_pages, num_extent_pages(eb));
/* Set btree blocks beyond nritems with 0 to avoid stale content */
nritems = btrfs_header_nritems(eb);
if (btrfs_header_level(eb) > 0) {
end = btrfs_node_key_ptr_offset(nritems);
memzero_extent_buffer(eb, end, eb->len - end);
} else {
/*
* Leaf:
* header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
*/
start = btrfs_item_nr_offset(nritems);
end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(eb);
memzero_extent_buffer(eb, start, end - start);
}
}
/*
* Unlike the work in write_one_eb(), we rely completely on extent locking.
* Page locking is only utilized at minimum to keep the VMM code happy.
*/
static int write_one_subpage_eb(struct extent_buffer *eb,
struct writeback_control *wbc,
struct btrfs_bio_ctrl *bio_ctrl)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
struct page *page = eb->pages[0];
for-5.20-tag -----BEGIN PGP SIGNATURE----- iQIzBAABCgAdFiEE8rQSAMVO+zA4DBdWxWXV+ddtWDsFAmLnyNUACgkQxWXV+ddt WDt9vA/9HcF+v5EkknyW07tatTap/Hm/ZB86Z5OZi6ikwIEcHsWhp3rUICejm88e GecDPIluDtCtyD6x4stuqkwOm22aDP5q2T9H6+gyw92ozyb436OV1Z8IrmftzXKY EpZO70PHZT+E6E/WYvyoTmmoCrjib7YlqCWZZhSLUFpsqqlOInmHEH49PW6KvM4r acUZ/RxHurKdmI3kNY6ECbAQl6CASvtTdYcVCx8fT2zN0azoLIQxpYa7n/9ca1R6 8WnYilCbLbNGtcUXvO2M3tMZ4/5kvxrwQsUn93ccCJYuiN0ASiDXbLZ2g4LZ+n56 JGu+y5v5oBwjpVf+46cuvnENP5BQ61594WPseiVjrqODWnPjN28XkcVC0XmPsiiZ lszeHO2cuIrIFoCah8ELMl8usu8+qxfXmPxIXtPu9rEyKsDtOjxVYc8SMXqLp0qQ qYtBoFm0JcZHqtZRpB+dhQ37/xXtH4ljUi/mI6x8iALVujeR273URs7yO9zgIdeW uZoFtbwpHFLUk+TL7Ku82/zOXp3fCwtDpNmlYbxeMbea/be3ShjncM4+mYzvHYri dYON2LFrq+mnRDqtIXTCaAYwX7zU8Y18Ev9QwlNll8dKlKwS89+jpqLoa+eVYy3c /HitHFza70KxmOj4dvDVZlzDpPvl7kW1UBkmskg4u3jnNWzedkM= =sS1q -----END PGP SIGNATURE----- Merge tag 'for-5.20-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux Pull btrfs updates from David Sterba: "This brings some long awaited changes, the send protocol bump, otherwise lots of small improvements and fixes. The main core part is reworking bio handling, cleaning up the submission and endio and improving error handling. There are some changes outside of btrfs adding helpers or updating API, listed at the end of the changelog. Features: - sysfs: - export chunk size, in debug mode add tunable for setting its size - show zoned among features (was only in debug mode) - show commit stats (number, last/max/total duration) - send protocol updated to 2 - new commands: - ability write larger data chunks than 64K - send raw compressed extents (uses the encoded data ioctls), ie. no decompression on send side, no compression needed on receive side if supported - send 'otime' (inode creation time) among other timestamps - send file attributes (a.k.a file flags and xflags) - this is first version bump, backward compatibility on send and receive side is provided - there are still some known and wanted commands that will be implemented in the near future, another version bump will be needed, however we want to minimize that to avoid causing usability issues - print checksum type and implementation at mount time - don't print some messages at mount (mentioned as people asked about it), we want to print messages namely for new features so let's make some space for that - big metadata - this has been supported for a long time and is not a feature that's worth mentioning - skinny metadata - same reason, set by default by mkfs Performance improvements: - reduced amount of reserved metadata for delayed items - when inserted items can be batched into one leaf - when deleting batched directory index items - when deleting delayed items used for deletion - overall improved count of files/sec, decreased subvolume lock contention - metadata item access bounds checker micro-optimized, with a few percent of improved runtime for metadata-heavy operations - increase direct io limit for read to 256 sectors, improved throughput by 3x on sample workload Notable fixes: - raid56 - reduce parity writes, skip sectors of stripe when there are no data updates - restore reading from on-disk data instead of using stripe cache, this reduces chances to damage correct data due to RMW cycle - refuse to replay log with unknown incompat read-only feature bit set - zoned - fix page locking when COW fails in the middle of allocation - improved tracking of active zones, ZNS drives may limit the number and there are ENOSPC errors due to that limit and not actual lack of space - adjust maximum extent size for zone append so it does not cause late ENOSPC due to underreservation - mirror reading error messages show the mirror number - don't fallback to buffered IO for NOWAIT direct IO writes, we don't have the NOWAIT semantics for buffered io yet - send, fix sending link commands for existing file paths when there are deleted and created hardlinks for same files - repair all mirrors for profiles with more than 1 copy (raid1c34) - fix repair of compressed extents, unify where error detection and repair happen Core changes: - bio completion cleanups - don't double defer compression bios - simplify endio workqueues - add more data to btrfs_bio to avoid allocation for read requests - rework bio error handling so it's same what block layer does, the submission works and errors are consumed in endio - when asynchronous bio offload fails fall back to synchronous checksum calculation to avoid errors under writeback or memory pressure - new trace points - raid56 events - ordered extent operations - super block log_root_transid deprecated (never used) - mixed_backref and big_metadata sysfs feature files removed, they've been default for sufficiently long time, there are no known users and mixed_backref could be confused with mixed_groups Non-btrfs changes, API updates: - minor highmem API update to cover const arguments - switch all kmap/kmap_atomic to kmap_local - remove redundant flush_dcache_page() - address_space_operations::writepage callback removed - add bdev_max_segments() helper" * tag 'for-5.20-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux: (163 commits) btrfs: don't call btrfs_page_set_checked in finish_compressed_bio_read btrfs: fix repair of compressed extents btrfs: remove the start argument to check_data_csum and export btrfs: pass a btrfs_bio to btrfs_repair_one_sector btrfs: simplify the pending I/O counting in struct compressed_bio btrfs: repair all known bad mirrors btrfs: merge btrfs_dev_stat_print_on_error with its only caller btrfs: join running log transaction when logging new name btrfs: simplify error handling in btrfs_lookup_dentry btrfs: send: always use the rbtree based inode ref management infrastructure btrfs: send: fix sending link commands for existing file paths btrfs: send: introduce recorded_ref_alloc and recorded_ref_free btrfs: zoned: wait until zone is finished when allocation didn't progress btrfs: zoned: write out partially allocated region btrfs: zoned: activate necessary block group btrfs: zoned: activate metadata block group on flush_space btrfs: zoned: disable metadata overcommit for zoned btrfs: zoned: introduce space_info->active_total_bytes btrfs: zoned: finish least available block group on data bg allocation btrfs: let can_allocate_chunk return error ...
2022-08-04 05:54:52 +08:00
blk_opf_t write_flags = wbc_to_write_flags(wbc);
bool no_dirty_ebs = false;
int ret;
prepare_eb_write(eb);
/* clear_page_dirty_for_io() in subpage helper needs page locked */
lock_page(page);
btrfs_subpage_set_writeback(fs_info, page, eb->start, eb->len);
/* Check if this is the last dirty bit to update nr_written */
no_dirty_ebs = btrfs_subpage_clear_and_test_dirty(fs_info, page,
eb->start, eb->len);
if (no_dirty_ebs)
clear_page_dirty_for_io(page);
bio_ctrl->end_io_func = end_bio_subpage_eb_writepage;
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
ret = submit_extent_page(REQ_OP_WRITE | write_flags, wbc,
bio_ctrl, eb->start, page, eb->len,
eb->start - page_offset(page), 0, false);
if (ret) {
btrfs_subpage_clear_writeback(fs_info, page, eb->start, eb->len);
set_btree_ioerr(page, eb);
unlock_page(page);
if (atomic_dec_and_test(&eb->io_pages))
end_extent_buffer_writeback(eb);
return -EIO;
}
unlock_page(page);
/*
* Submission finished without problem, if no range of the page is
* dirty anymore, we have submitted a page. Update nr_written in wbc.
*/
if (no_dirty_ebs)
wbc->nr_to_write--;
return ret;
}
static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
struct writeback_control *wbc,
struct btrfs_bio_ctrl *bio_ctrl)
{
u64 disk_bytenr = eb->start;
int i, num_pages;
for-5.20-tag -----BEGIN PGP SIGNATURE----- iQIzBAABCgAdFiEE8rQSAMVO+zA4DBdWxWXV+ddtWDsFAmLnyNUACgkQxWXV+ddt WDt9vA/9HcF+v5EkknyW07tatTap/Hm/ZB86Z5OZi6ikwIEcHsWhp3rUICejm88e GecDPIluDtCtyD6x4stuqkwOm22aDP5q2T9H6+gyw92ozyb436OV1Z8IrmftzXKY EpZO70PHZT+E6E/WYvyoTmmoCrjib7YlqCWZZhSLUFpsqqlOInmHEH49PW6KvM4r acUZ/RxHurKdmI3kNY6ECbAQl6CASvtTdYcVCx8fT2zN0azoLIQxpYa7n/9ca1R6 8WnYilCbLbNGtcUXvO2M3tMZ4/5kvxrwQsUn93ccCJYuiN0ASiDXbLZ2g4LZ+n56 JGu+y5v5oBwjpVf+46cuvnENP5BQ61594WPseiVjrqODWnPjN28XkcVC0XmPsiiZ lszeHO2cuIrIFoCah8ELMl8usu8+qxfXmPxIXtPu9rEyKsDtOjxVYc8SMXqLp0qQ qYtBoFm0JcZHqtZRpB+dhQ37/xXtH4ljUi/mI6x8iALVujeR273URs7yO9zgIdeW uZoFtbwpHFLUk+TL7Ku82/zOXp3fCwtDpNmlYbxeMbea/be3ShjncM4+mYzvHYri dYON2LFrq+mnRDqtIXTCaAYwX7zU8Y18Ev9QwlNll8dKlKwS89+jpqLoa+eVYy3c /HitHFza70KxmOj4dvDVZlzDpPvl7kW1UBkmskg4u3jnNWzedkM= =sS1q -----END PGP SIGNATURE----- Merge tag 'for-5.20-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux Pull btrfs updates from David Sterba: "This brings some long awaited changes, the send protocol bump, otherwise lots of small improvements and fixes. The main core part is reworking bio handling, cleaning up the submission and endio and improving error handling. There are some changes outside of btrfs adding helpers or updating API, listed at the end of the changelog. Features: - sysfs: - export chunk size, in debug mode add tunable for setting its size - show zoned among features (was only in debug mode) - show commit stats (number, last/max/total duration) - send protocol updated to 2 - new commands: - ability write larger data chunks than 64K - send raw compressed extents (uses the encoded data ioctls), ie. no decompression on send side, no compression needed on receive side if supported - send 'otime' (inode creation time) among other timestamps - send file attributes (a.k.a file flags and xflags) - this is first version bump, backward compatibility on send and receive side is provided - there are still some known and wanted commands that will be implemented in the near future, another version bump will be needed, however we want to minimize that to avoid causing usability issues - print checksum type and implementation at mount time - don't print some messages at mount (mentioned as people asked about it), we want to print messages namely for new features so let's make some space for that - big metadata - this has been supported for a long time and is not a feature that's worth mentioning - skinny metadata - same reason, set by default by mkfs Performance improvements: - reduced amount of reserved metadata for delayed items - when inserted items can be batched into one leaf - when deleting batched directory index items - when deleting delayed items used for deletion - overall improved count of files/sec, decreased subvolume lock contention - metadata item access bounds checker micro-optimized, with a few percent of improved runtime for metadata-heavy operations - increase direct io limit for read to 256 sectors, improved throughput by 3x on sample workload Notable fixes: - raid56 - reduce parity writes, skip sectors of stripe when there are no data updates - restore reading from on-disk data instead of using stripe cache, this reduces chances to damage correct data due to RMW cycle - refuse to replay log with unknown incompat read-only feature bit set - zoned - fix page locking when COW fails in the middle of allocation - improved tracking of active zones, ZNS drives may limit the number and there are ENOSPC errors due to that limit and not actual lack of space - adjust maximum extent size for zone append so it does not cause late ENOSPC due to underreservation - mirror reading error messages show the mirror number - don't fallback to buffered IO for NOWAIT direct IO writes, we don't have the NOWAIT semantics for buffered io yet - send, fix sending link commands for existing file paths when there are deleted and created hardlinks for same files - repair all mirrors for profiles with more than 1 copy (raid1c34) - fix repair of compressed extents, unify where error detection and repair happen Core changes: - bio completion cleanups - don't double defer compression bios - simplify endio workqueues - add more data to btrfs_bio to avoid allocation for read requests - rework bio error handling so it's same what block layer does, the submission works and errors are consumed in endio - when asynchronous bio offload fails fall back to synchronous checksum calculation to avoid errors under writeback or memory pressure - new trace points - raid56 events - ordered extent operations - super block log_root_transid deprecated (never used) - mixed_backref and big_metadata sysfs feature files removed, they've been default for sufficiently long time, there are no known users and mixed_backref could be confused with mixed_groups Non-btrfs changes, API updates: - minor highmem API update to cover const arguments - switch all kmap/kmap_atomic to kmap_local - remove redundant flush_dcache_page() - address_space_operations::writepage callback removed - add bdev_max_segments() helper" * tag 'for-5.20-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux: (163 commits) btrfs: don't call btrfs_page_set_checked in finish_compressed_bio_read btrfs: fix repair of compressed extents btrfs: remove the start argument to check_data_csum and export btrfs: pass a btrfs_bio to btrfs_repair_one_sector btrfs: simplify the pending I/O counting in struct compressed_bio btrfs: repair all known bad mirrors btrfs: merge btrfs_dev_stat_print_on_error with its only caller btrfs: join running log transaction when logging new name btrfs: simplify error handling in btrfs_lookup_dentry btrfs: send: always use the rbtree based inode ref management infrastructure btrfs: send: fix sending link commands for existing file paths btrfs: send: introduce recorded_ref_alloc and recorded_ref_free btrfs: zoned: wait until zone is finished when allocation didn't progress btrfs: zoned: write out partially allocated region btrfs: zoned: activate necessary block group btrfs: zoned: activate metadata block group on flush_space btrfs: zoned: disable metadata overcommit for zoned btrfs: zoned: introduce space_info->active_total_bytes btrfs: zoned: finish least available block group on data bg allocation btrfs: let can_allocate_chunk return error ...
2022-08-04 05:54:52 +08:00
blk_opf_t write_flags = wbc_to_write_flags(wbc);
int ret = 0;
prepare_eb_write(eb);
bio_ctrl->end_io_func = end_bio_extent_buffer_writepage;
num_pages = num_extent_pages(eb);
for (i = 0; i < num_pages; i++) {
struct page *p = eb->pages[i];
clear_page_dirty_for_io(p);
set_page_writeback(p);
ret = submit_extent_page(REQ_OP_WRITE | write_flags, wbc,
bio_ctrl, disk_bytenr, p,
PAGE_SIZE, 0, 0, false);
if (ret) {
set_btree_ioerr(p, eb);
if (PageWriteback(p))
end_page_writeback(p);
if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
end_extent_buffer_writeback(eb);
ret = -EIO;
break;
}
disk_bytenr += PAGE_SIZE;
wbc->nr_to_write--;
unlock_page(p);
}
if (unlikely(ret)) {
for (; i < num_pages; i++) {
struct page *p = eb->pages[i];
clear_page_dirty_for_io(p);
unlock_page(p);
}
}
return ret;
}
/*
* Submit one subpage btree page.
*
* The main difference to submit_eb_page() is:
* - Page locking
* For subpage, we don't rely on page locking at all.
*
* - Flush write bio
* We only flush bio if we may be unable to fit current extent buffers into
* current bio.
*
* Return >=0 for the number of submitted extent buffers.
* Return <0 for fatal error.
*/
static int submit_eb_subpage(struct page *page,
struct writeback_control *wbc,
struct btrfs_bio_ctrl *bio_ctrl)
{
struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
int submitted = 0;
u64 page_start = page_offset(page);
int bit_start = 0;
int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
int ret;
/* Lock and write each dirty extent buffers in the range */
while (bit_start < fs_info->subpage_info->bitmap_nr_bits) {
struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
struct extent_buffer *eb;
unsigned long flags;
u64 start;
/*
* Take private lock to ensure the subpage won't be detached
* in the meantime.
*/
spin_lock(&page->mapping->private_lock);
if (!PagePrivate(page)) {
spin_unlock(&page->mapping->private_lock);
break;
}
spin_lock_irqsave(&subpage->lock, flags);
if (!test_bit(bit_start + fs_info->subpage_info->dirty_offset,
subpage->bitmaps)) {
spin_unlock_irqrestore(&subpage->lock, flags);
spin_unlock(&page->mapping->private_lock);
bit_start++;
continue;
}
start = page_start + bit_start * fs_info->sectorsize;
bit_start += sectors_per_node;
/*
* Here we just want to grab the eb without touching extra
* spin locks, so call find_extent_buffer_nolock().
*/
eb = find_extent_buffer_nolock(fs_info, start);
spin_unlock_irqrestore(&subpage->lock, flags);
spin_unlock(&page->mapping->private_lock);
/*
* The eb has already reached 0 refs thus find_extent_buffer()
* doesn't return it. We don't need to write back such eb
* anyway.
*/
if (!eb)
continue;
ret = lock_extent_buffer_for_io(eb, bio_ctrl);
if (ret == 0) {
free_extent_buffer(eb);
continue;
}
if (ret < 0) {
free_extent_buffer(eb);
goto cleanup;
}
ret = write_one_subpage_eb(eb, wbc, bio_ctrl);
free_extent_buffer(eb);
if (ret < 0)
goto cleanup;
submitted++;
}
return submitted;
cleanup:
/* We hit error, end bio for the submitted extent buffers */
submit_write_bio(bio_ctrl, ret);
return ret;
}
/*
* Submit all page(s) of one extent buffer.
*
* @page: the page of one extent buffer
* @eb_context: to determine if we need to submit this page, if current page
* belongs to this eb, we don't need to submit
*
* The caller should pass each page in their bytenr order, and here we use
* @eb_context to determine if we have submitted pages of one extent buffer.
*
* If we have, we just skip until we hit a new page that doesn't belong to
* current @eb_context.
*
* If not, we submit all the page(s) of the extent buffer.
*
* Return >0 if we have submitted the extent buffer successfully.
* Return 0 if we don't need to submit the page, as it's already submitted by
* previous call.
* Return <0 for fatal error.
*/
static int submit_eb_page(struct page *page, struct writeback_control *wbc,
struct btrfs_bio_ctrl *bio_ctrl,
struct extent_buffer **eb_context)
{
struct address_space *mapping = page->mapping;
struct btrfs_block_group *cache = NULL;
struct extent_buffer *eb;
int ret;
if (!PagePrivate(page))
return 0;
if (btrfs_sb(page->mapping->host->i_sb)->nodesize < PAGE_SIZE)
return submit_eb_subpage(page, wbc, bio_ctrl);
spin_lock(&mapping->private_lock);
if (!PagePrivate(page)) {
spin_unlock(&mapping->private_lock);
return 0;
}
eb = (struct extent_buffer *)page->private;
/*
* Shouldn't happen and normally this would be a BUG_ON but no point
* crashing the machine for something we can survive anyway.
*/
if (WARN_ON(!eb)) {
spin_unlock(&mapping->private_lock);
return 0;
}
if (eb == *eb_context) {
spin_unlock(&mapping->private_lock);
return 0;
}
ret = atomic_inc_not_zero(&eb->refs);
spin_unlock(&mapping->private_lock);
if (!ret)
return 0;
if (!btrfs_check_meta_write_pointer(eb->fs_info, eb, &cache)) {
/*
* If for_sync, this hole will be filled with
* trasnsaction commit.
*/
if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
ret = -EAGAIN;
else
ret = 0;
free_extent_buffer(eb);
return ret;
}
*eb_context = eb;
ret = lock_extent_buffer_for_io(eb, bio_ctrl);
if (ret <= 0) {
btrfs_revert_meta_write_pointer(cache, eb);
if (cache)
btrfs_put_block_group(cache);
free_extent_buffer(eb);
return ret;
}
if (cache) {
/*
* Implies write in zoned mode. Mark the last eb in a block group.
*/
btrfs_schedule_zone_finish_bg(cache, eb);
btrfs_put_block_group(cache);
}
ret = write_one_eb(eb, wbc, bio_ctrl);
free_extent_buffer(eb);
if (ret < 0)
return ret;
return 1;
}
int btree_write_cache_pages(struct address_space *mapping,
struct writeback_control *wbc)
{
struct extent_buffer *eb_context = NULL;
struct btrfs_bio_ctrl bio_ctrl = {
.extent_locked = 0,
.sync_io = (wbc->sync_mode == WB_SYNC_ALL),
};
btrfs: Don't submit any btree write bio if the fs has errors [BUG] There is a fuzzed image which could cause KASAN report at unmount time. BUG: KASAN: use-after-free in btrfs_queue_work+0x2c1/0x390 Read of size 8 at addr ffff888067cf6848 by task umount/1922 CPU: 0 PID: 1922 Comm: umount Tainted: G W 5.0.21 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1ubuntu1 04/01/2014 Call Trace: dump_stack+0x5b/0x8b print_address_description+0x70/0x280 kasan_report+0x13a/0x19b btrfs_queue_work+0x2c1/0x390 btrfs_wq_submit_bio+0x1cd/0x240 btree_submit_bio_hook+0x18c/0x2a0 submit_one_bio+0x1be/0x320 flush_write_bio.isra.41+0x2c/0x70 btree_write_cache_pages+0x3bb/0x7f0 do_writepages+0x5c/0x130 __writeback_single_inode+0xa3/0x9a0 writeback_single_inode+0x23d/0x390 write_inode_now+0x1b5/0x280 iput+0x2ef/0x600 close_ctree+0x341/0x750 generic_shutdown_super+0x126/0x370 kill_anon_super+0x31/0x50 btrfs_kill_super+0x36/0x2b0 deactivate_locked_super+0x80/0xc0 deactivate_super+0x13c/0x150 cleanup_mnt+0x9a/0x130 task_work_run+0x11a/0x1b0 exit_to_usermode_loop+0x107/0x130 do_syscall_64+0x1e5/0x280 entry_SYSCALL_64_after_hwframe+0x44/0xa9 [CAUSE] The fuzzed image has a completely screwd up extent tree: leaf 29421568 gen 8 total ptrs 6 free space 3587 owner EXTENT_TREE refs 2 lock (w:0 r:0 bw:0 br:0 sw:0 sr:0) lock_owner 0 current 5938 item 0 key (12587008 168 4096) itemoff 3942 itemsize 53 extent refs 1 gen 9 flags 1 ref#0: extent data backref root 5 objectid 259 offset 0 count 1 item 1 key (12591104 168 8192) itemoff 3889 itemsize 53 extent refs 1 gen 9 flags 1 ref#0: extent data backref root 5 objectid 271 offset 0 count 1 item 2 key (12599296 168 4096) itemoff 3836 itemsize 53 extent refs 1 gen 9 flags 1 ref#0: extent data backref root 5 objectid 259 offset 4096 count 1 item 3 key (29360128 169 0) itemoff 3803 itemsize 33 extent refs 1 gen 9 flags 2 ref#0: tree block backref root 5 item 4 key (29368320 169 1) itemoff 3770 itemsize 33 extent refs 1 gen 9 flags 2 ref#0: tree block backref root 5 item 5 key (29372416 169 0) itemoff 3737 itemsize 33 extent refs 1 gen 9 flags 2 ref#0: tree block backref root 5 Note that leaf 29421568 doesn't have its backref in the extent tree. Thus extent allocator can re-allocate leaf 29421568 for other trees. In short, the bug is caused by: - Existing tree block gets allocated to log tree This got its generation bumped. - Log tree balance cleaned dirty bit of offending tree block It will not be written back to disk, thus no WRITTEN flag. - Original owner of the tree block gets COWed Since the tree block has higher transid, no WRITTEN flag, it's reused, and not traced by transaction::dirty_pages. - Transaction aborted Tree blocks get cleaned according to transaction::dirty_pages. But the offending tree block is not recorded at all. - Filesystem unmount All pages are assumed to be are clean, destroying all workqueue, then call iput(btree_inode). But offending tree block is still dirty, which triggers writeback, and causes use-after-free bug. The detailed sequence looks like this: - Initial status eb: 29421568, header=WRITTEN bflags_dirty=0, page_dirty=0, gen=8, not traced by any dirty extent_iot_tree. - New tree block is allocated Since there is no backref for 29421568, it's re-allocated as new tree block. Keep in mind that tree block 29421568 is still referred by extent tree. - Tree block 29421568 is filled for log tree eb: 29421568, header=0 bflags_dirty=1, page_dirty=1, gen=9 << (gen bumped) traced by btrfs_root::dirty_log_pages - Some log tree operations Since the fs is using node size 4096, the log tree can easily go a level higher. - Log tree needs balance Tree block 29421568 gets all its content pushed to right, thus now it is empty, and we don't need it. btrfs_clean_tree_block() from __push_leaf_right() get called. eb: 29421568, header=0 bflags_dirty=0, page_dirty=0, gen=9 traced by btrfs_root::dirty_log_pages - Log tree write back btree_write_cache_pages() goes through dirty pages ranges, but since page of tree block 29421568 gets cleaned already, it's not written back to disk. Thus it doesn't have WRITTEN bit set. But ranges in dirty_log_pages are cleared. eb: 29421568, header=0 bflags_dirty=0, page_dirty=0, gen=9 not traced by any dirty extent_iot_tree. - Extent tree update when committing transaction Since tree block 29421568 has transid equal to running trans, and has no WRITTEN bit, should_cow_block() will use it directly without adding it to btrfs_transaction::dirty_pages. eb: 29421568, header=0 bflags_dirty=1, page_dirty=1, gen=9 not traced by any dirty extent_iot_tree. At this stage, we're doomed. We have a dirty eb not tracked by any extent io tree. - Transaction gets aborted due to corrupted extent tree Btrfs cleans up dirty pages according to transaction::dirty_pages and btrfs_root::dirty_log_pages. But since tree block 29421568 is not tracked by neither of them, it's still dirty. eb: 29421568, header=0 bflags_dirty=1, page_dirty=1, gen=9 not traced by any dirty extent_iot_tree. - Filesystem unmount Since all cleanup is assumed to be done, all workqueus are destroyed. Then iput(btree_inode) is called, expecting no dirty pages. But tree 29421568 is still dirty, thus triggering writeback. Since all workqueues are already freed, we cause use-after-free. This shows us that, log tree blocks + bad extent tree can cause wild dirty pages. [FIX] To fix the problem, don't submit any btree write bio if the filesytem has any error. This is the last safe net, just in case other cleanup haven't caught catch it. Link: https://github.com/bobfuzzer/CVE/tree/master/CVE-2019-19377 CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-02-12 14:12:44 +08:00
struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
int ret = 0;
int done = 0;
int nr_to_write_done = 0;
struct pagevec pvec;
int nr_pages;
pgoff_t index;
pgoff_t end; /* Inclusive */
int scanned = 0;
xa_mark_t tag;
pagevec_init(&pvec);
if (wbc->range_cyclic) {
index = mapping->writeback_index; /* Start from prev offset */
end = -1;
btrfs: fix improper setting of scanned for range cyclic write cache pages We noticed that we were having regular CG OOM kills in cases where there was still enough dirty pages to avoid OOM'ing. It turned out there's this corner case in btrfs's handling of range_cyclic where files that were being redirtied were not getting fully written out because of how we do range_cyclic writeback. We unconditionally were setting scanned = 1; the first time we found any pages in the inode. This isn't actually what we want, we want it to be set if we've scanned the entire file. For range_cyclic we could be starting in the middle or towards the end of the file, so we could write one page and then not write any of the other dirty pages in the file because we set scanned = 1. Fix this by not setting scanned = 1 if we find pages. The rules for setting scanned should be 1) !range_cyclic. In this case we have a specified range to write out. 2) range_cyclic && index == 0. In this case we've started at the beginning and there is no need to loop around a second time. 3) range_cyclic && we started at index > 0 and we've reached the end of the file without satisfying our nr_to_write. This patch fixes both of our writepages implementations to make sure these rules hold true. This fixed our over zealous CG OOMs in production. Fixes: d1310b2e0cd9 ("Btrfs: Split the extent_map code into two parts") Signed-off-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: David Sterba <dsterba@suse.com> [ add comment ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-01-03 23:38:44 +08:00
/*
* Start from the beginning does not need to cycle over the
* range, mark it as scanned.
*/
scanned = (index == 0);
} else {
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
index = wbc->range_start >> PAGE_SHIFT;
end = wbc->range_end >> PAGE_SHIFT;
scanned = 1;
}
if (wbc->sync_mode == WB_SYNC_ALL)
tag = PAGECACHE_TAG_TOWRITE;
else
tag = PAGECACHE_TAG_DIRTY;
btrfs_zoned_meta_io_lock(fs_info);
retry:
if (wbc->sync_mode == WB_SYNC_ALL)
tag_pages_for_writeback(mapping, index, end);
while (!done && !nr_to_write_done && (index <= end) &&
(nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
tag))) {
unsigned i;
for (i = 0; i < nr_pages; i++) {
struct page *page = pvec.pages[i];
ret = submit_eb_page(page, wbc, &bio_ctrl, &eb_context);
if (ret == 0)
continue;
if (ret < 0) {
done = 1;
break;
}
/*
* the filesystem may choose to bump up nr_to_write.
* We have to make sure to honor the new nr_to_write
* at any time
*/
nr_to_write_done = wbc->nr_to_write <= 0;
}
pagevec_release(&pvec);
cond_resched();
}
if (!scanned && !done) {
/*
* We hit the last page and there is more work to be done: wrap
* back to the start of the file
*/
scanned = 1;
index = 0;
goto retry;
}
btrfs: Don't submit any btree write bio if the fs has errors [BUG] There is a fuzzed image which could cause KASAN report at unmount time. BUG: KASAN: use-after-free in btrfs_queue_work+0x2c1/0x390 Read of size 8 at addr ffff888067cf6848 by task umount/1922 CPU: 0 PID: 1922 Comm: umount Tainted: G W 5.0.21 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1ubuntu1 04/01/2014 Call Trace: dump_stack+0x5b/0x8b print_address_description+0x70/0x280 kasan_report+0x13a/0x19b btrfs_queue_work+0x2c1/0x390 btrfs_wq_submit_bio+0x1cd/0x240 btree_submit_bio_hook+0x18c/0x2a0 submit_one_bio+0x1be/0x320 flush_write_bio.isra.41+0x2c/0x70 btree_write_cache_pages+0x3bb/0x7f0 do_writepages+0x5c/0x130 __writeback_single_inode+0xa3/0x9a0 writeback_single_inode+0x23d/0x390 write_inode_now+0x1b5/0x280 iput+0x2ef/0x600 close_ctree+0x341/0x750 generic_shutdown_super+0x126/0x370 kill_anon_super+0x31/0x50 btrfs_kill_super+0x36/0x2b0 deactivate_locked_super+0x80/0xc0 deactivate_super+0x13c/0x150 cleanup_mnt+0x9a/0x130 task_work_run+0x11a/0x1b0 exit_to_usermode_loop+0x107/0x130 do_syscall_64+0x1e5/0x280 entry_SYSCALL_64_after_hwframe+0x44/0xa9 [CAUSE] The fuzzed image has a completely screwd up extent tree: leaf 29421568 gen 8 total ptrs 6 free space 3587 owner EXTENT_TREE refs 2 lock (w:0 r:0 bw:0 br:0 sw:0 sr:0) lock_owner 0 current 5938 item 0 key (12587008 168 4096) itemoff 3942 itemsize 53 extent refs 1 gen 9 flags 1 ref#0: extent data backref root 5 objectid 259 offset 0 count 1 item 1 key (12591104 168 8192) itemoff 3889 itemsize 53 extent refs 1 gen 9 flags 1 ref#0: extent data backref root 5 objectid 271 offset 0 count 1 item 2 key (12599296 168 4096) itemoff 3836 itemsize 53 extent refs 1 gen 9 flags 1 ref#0: extent data backref root 5 objectid 259 offset 4096 count 1 item 3 key (29360128 169 0) itemoff 3803 itemsize 33 extent refs 1 gen 9 flags 2 ref#0: tree block backref root 5 item 4 key (29368320 169 1) itemoff 3770 itemsize 33 extent refs 1 gen 9 flags 2 ref#0: tree block backref root 5 item 5 key (29372416 169 0) itemoff 3737 itemsize 33 extent refs 1 gen 9 flags 2 ref#0: tree block backref root 5 Note that leaf 29421568 doesn't have its backref in the extent tree. Thus extent allocator can re-allocate leaf 29421568 for other trees. In short, the bug is caused by: - Existing tree block gets allocated to log tree This got its generation bumped. - Log tree balance cleaned dirty bit of offending tree block It will not be written back to disk, thus no WRITTEN flag. - Original owner of the tree block gets COWed Since the tree block has higher transid, no WRITTEN flag, it's reused, and not traced by transaction::dirty_pages. - Transaction aborted Tree blocks get cleaned according to transaction::dirty_pages. But the offending tree block is not recorded at all. - Filesystem unmount All pages are assumed to be are clean, destroying all workqueue, then call iput(btree_inode). But offending tree block is still dirty, which triggers writeback, and causes use-after-free bug. The detailed sequence looks like this: - Initial status eb: 29421568, header=WRITTEN bflags_dirty=0, page_dirty=0, gen=8, not traced by any dirty extent_iot_tree. - New tree block is allocated Since there is no backref for 29421568, it's re-allocated as new tree block. Keep in mind that tree block 29421568 is still referred by extent tree. - Tree block 29421568 is filled for log tree eb: 29421568, header=0 bflags_dirty=1, page_dirty=1, gen=9 << (gen bumped) traced by btrfs_root::dirty_log_pages - Some log tree operations Since the fs is using node size 4096, the log tree can easily go a level higher. - Log tree needs balance Tree block 29421568 gets all its content pushed to right, thus now it is empty, and we don't need it. btrfs_clean_tree_block() from __push_leaf_right() get called. eb: 29421568, header=0 bflags_dirty=0, page_dirty=0, gen=9 traced by btrfs_root::dirty_log_pages - Log tree write back btree_write_cache_pages() goes through dirty pages ranges, but since page of tree block 29421568 gets cleaned already, it's not written back to disk. Thus it doesn't have WRITTEN bit set. But ranges in dirty_log_pages are cleared. eb: 29421568, header=0 bflags_dirty=0, page_dirty=0, gen=9 not traced by any dirty extent_iot_tree. - Extent tree update when committing transaction Since tree block 29421568 has transid equal to running trans, and has no WRITTEN bit, should_cow_block() will use it directly without adding it to btrfs_transaction::dirty_pages. eb: 29421568, header=0 bflags_dirty=1, page_dirty=1, gen=9 not traced by any dirty extent_iot_tree. At this stage, we're doomed. We have a dirty eb not tracked by any extent io tree. - Transaction gets aborted due to corrupted extent tree Btrfs cleans up dirty pages according to transaction::dirty_pages and btrfs_root::dirty_log_pages. But since tree block 29421568 is not tracked by neither of them, it's still dirty. eb: 29421568, header=0 bflags_dirty=1, page_dirty=1, gen=9 not traced by any dirty extent_iot_tree. - Filesystem unmount Since all cleanup is assumed to be done, all workqueus are destroyed. Then iput(btree_inode) is called, expecting no dirty pages. But tree 29421568 is still dirty, thus triggering writeback. Since all workqueues are already freed, we cause use-after-free. This shows us that, log tree blocks + bad extent tree can cause wild dirty pages. [FIX] To fix the problem, don't submit any btree write bio if the filesytem has any error. This is the last safe net, just in case other cleanup haven't caught catch it. Link: https://github.com/bobfuzzer/CVE/tree/master/CVE-2019-19377 CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-02-12 14:12:44 +08:00
/*
* If something went wrong, don't allow any metadata write bio to be
* submitted.
*
* This would prevent use-after-free if we had dirty pages not
* cleaned up, which can still happen by fuzzed images.
*
* - Bad extent tree
* Allowing existing tree block to be allocated for other trees.
*
* - Log tree operations
* Exiting tree blocks get allocated to log tree, bumps its
* generation, then get cleaned in tree re-balance.
* Such tree block will not be written back, since it's clean,
* thus no WRITTEN flag set.
* And after log writes back, this tree block is not traced by
* any dirty extent_io_tree.
*
* - Offending tree block gets re-dirtied from its original owner
* Since it has bumped generation, no WRITTEN flag, it can be
* reused without COWing. This tree block will not be traced
* by btrfs_transaction::dirty_pages.
*
* Now such dirty tree block will not be cleaned by any dirty
* extent io tree. Thus we don't want to submit such wild eb
* if the fs already has error.
*
btrfs: avoid double clean up when submit_one_bio() failed [BUG] When running generic/475 with 64K page size and 4K sector size, it has a very high chance (almost 100%) to hang, with mostly data page locked but no one is going to unlock it. [CAUSE] With commit 1784b7d502a9 ("btrfs: handle csum lookup errors properly on reads"), if we failed to lookup checksum due to metadata IO error, we will return error for btrfs_submit_data_bio(). This will cause the page to be unlocked twice in btrfs_do_readpage(): btrfs_do_readpage() |- submit_extent_page() | |- submit_one_bio() | |- btrfs_submit_data_bio() | |- if (ret) { | |- bio->bi_status = ret; | |- bio_endio(bio); } | In the endio function, we will call end_page_read() | and unlock_extent() to cleanup the subpage range. | |- if (ret) { |- unlock_extent(); end_page_read() } Here we unlock the extent and cleanup the subpage range again. For unlock_extent(), it's mostly double unlock safe. But for end_page_read(), it's not, especially for subpage case, as for subpage case we will call btrfs_subpage_end_reader() to reduce the reader number, and use that to number to determine if we need to unlock the full page. If double accounted, it can underflow the number and leave the page locked without anyone to unlock it. [FIX] The commit 1784b7d502a9 ("btrfs: handle csum lookup errors properly on reads") itself is completely fine, it's our existing code not properly handling the error from bio submission hook properly. This patch will make submit_one_bio() to return void so that the callers will never be able to do cleanup when bio submission hook fails. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-04-12 20:30:13 +08:00
* We can get ret > 0 from submit_extent_page() indicating how many ebs
* were submitted. Reset it to 0 to avoid false alerts for the caller.
*/
if (ret > 0)
ret = 0;
if (!ret && BTRFS_FS_ERROR(fs_info))
ret = -EROFS;
submit_write_bio(&bio_ctrl, ret);
btrfs_zoned_meta_io_unlock(fs_info);
return ret;
}
/*
btrfs: fix parameter description for functions in extent_io.c This makes the file W=1 clean and fixes the following warnings: fs/btrfs/extent_io.c:414: warning: Function parameter or member 'tree' not described in '__etree_search' fs/btrfs/extent_io.c:414: warning: Function parameter or member 'offset' not described in '__etree_search' fs/btrfs/extent_io.c:414: warning: Function parameter or member 'next_ret' not described in '__etree_search' fs/btrfs/extent_io.c:414: warning: Function parameter or member 'prev_ret' not described in '__etree_search' fs/btrfs/extent_io.c:414: warning: Function parameter or member 'p_ret' not described in '__etree_search' fs/btrfs/extent_io.c:414: warning: Function parameter or member 'parent_ret' not described in '__etree_search' fs/btrfs/extent_io.c:1607: warning: Function parameter or member 'tree' not described in 'find_contiguous_extent_bit' fs/btrfs/extent_io.c:1607: warning: Function parameter or member 'start' not described in 'find_contiguous_extent_bit' fs/btrfs/extent_io.c:1607: warning: Function parameter or member 'start_ret' not described in 'find_contiguous_extent_bit' fs/btrfs/extent_io.c:1607: warning: Function parameter or member 'end_ret' not described in 'find_contiguous_extent_bit' fs/btrfs/extent_io.c:1607: warning: Function parameter or member 'bits' not described in 'find_contiguous_extent_bit' fs/btrfs/extent_io.c:1644: warning: Function parameter or member 'tree' not described in 'find_first_clear_extent_bit' fs/btrfs/extent_io.c:1644: warning: Function parameter or member 'start' not described in 'find_first_clear_extent_bit' fs/btrfs/extent_io.c:1644: warning: Function parameter or member 'start_ret' not described in 'find_first_clear_extent_bit' fs/btrfs/extent_io.c:1644: warning: Function parameter or member 'end_ret' not described in 'find_first_clear_extent_bit' fs/btrfs/extent_io.c:1644: warning: Function parameter or member 'bits' not described in 'find_first_clear_extent_bit' fs/btrfs/extent_io.c:4187: warning: Function parameter or member 'epd' not described in 'extent_write_cache_pages' fs/btrfs/extent_io.c:4187: warning: Excess function parameter 'data' description in 'extent_write_cache_pages' Signed-off-by: Nikolay Borisov <nborisov@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-22 17:58:03 +08:00
* Walk the list of dirty pages of the given address space and write all of them.
*
* @mapping: address space structure to write
* @wbc: subtract the number of written pages from *@wbc->nr_to_write
* @bio_ctrl: holds context for the write, namely the bio
*
* If a page is already under I/O, write_cache_pages() skips it, even
* if it's dirty. This is desirable behaviour for memory-cleaning writeback,
* but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
* and msync() need to guarantee that all the data which was dirty at the time
* the call was made get new I/O started against them. If wbc->sync_mode is
* WB_SYNC_ALL then we were called for data integrity and we must wait for
* existing IO to complete.
*/
static int extent_write_cache_pages(struct address_space *mapping,
struct writeback_control *wbc,
struct btrfs_bio_ctrl *bio_ctrl)
{
struct inode *inode = mapping->host;
int ret = 0;
int done = 0;
int nr_to_write_done = 0;
struct pagevec pvec;
int nr_pages;
pgoff_t index;
pgoff_t end; /* Inclusive */
pgoff_t done_index;
int range_whole = 0;
int scanned = 0;
xa_mark_t tag;
/*
* We have to hold onto the inode so that ordered extents can do their
* work when the IO finishes. The alternative to this is failing to add
* an ordered extent if the igrab() fails there and that is a huge pain
* to deal with, so instead just hold onto the inode throughout the
* writepages operation. If it fails here we are freeing up the inode
* anyway and we'd rather not waste our time writing out stuff that is
* going to be truncated anyway.
*/
if (!igrab(inode))
return 0;
pagevec_init(&pvec);
if (wbc->range_cyclic) {
index = mapping->writeback_index; /* Start from prev offset */
end = -1;
btrfs: fix improper setting of scanned for range cyclic write cache pages We noticed that we were having regular CG OOM kills in cases where there was still enough dirty pages to avoid OOM'ing. It turned out there's this corner case in btrfs's handling of range_cyclic where files that were being redirtied were not getting fully written out because of how we do range_cyclic writeback. We unconditionally were setting scanned = 1; the first time we found any pages in the inode. This isn't actually what we want, we want it to be set if we've scanned the entire file. For range_cyclic we could be starting in the middle or towards the end of the file, so we could write one page and then not write any of the other dirty pages in the file because we set scanned = 1. Fix this by not setting scanned = 1 if we find pages. The rules for setting scanned should be 1) !range_cyclic. In this case we have a specified range to write out. 2) range_cyclic && index == 0. In this case we've started at the beginning and there is no need to loop around a second time. 3) range_cyclic && we started at index > 0 and we've reached the end of the file without satisfying our nr_to_write. This patch fixes both of our writepages implementations to make sure these rules hold true. This fixed our over zealous CG OOMs in production. Fixes: d1310b2e0cd9 ("Btrfs: Split the extent_map code into two parts") Signed-off-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: David Sterba <dsterba@suse.com> [ add comment ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-01-03 23:38:44 +08:00
/*
* Start from the beginning does not need to cycle over the
* range, mark it as scanned.
*/
scanned = (index == 0);
} else {
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
index = wbc->range_start >> PAGE_SHIFT;
end = wbc->range_end >> PAGE_SHIFT;
if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
range_whole = 1;
scanned = 1;
}
btrfs: use tagged writepage to mitigate livelock of snapshot Snapshot is expected to be fast. But if there are writers steadily creating dirty pages in our subvolume, the snapshot may take a very long time to complete. To fix the problem, we use tagged writepage for snapshot flusher as we do in the generic write_cache_pages(), so we can omit pages dirtied after the snapshot command. This does not change the semantics regarding which data get to the snapshot, if there are pages being dirtied during the snapshotting operation. There's a sync called before snapshot is taken in old/new case, any IO in flight just after that may be in the snapshot but this depends on other system effects that might still sync the IO. We do a simple snapshot speed test on a Intel D-1531 box: fio --ioengine=libaio --iodepth=32 --bs=4k --rw=write --size=64G --direct=0 --thread=1 --numjobs=1 --time_based --runtime=120 --filename=/mnt/sub/testfile --name=job1 --group_reporting & sleep 5; time btrfs sub snap -r /mnt/sub /mnt/snap; killall fio original: 1m58sec patched: 6.54sec This is the best case for this patch since for a sequential write case, we omit nearly all pages dirtied after the snapshot command. For a multi writers, random write test: fio --ioengine=libaio --iodepth=32 --bs=4k --rw=randwrite --size=64G --direct=0 --thread=1 --numjobs=4 --time_based --runtime=120 --filename=/mnt/sub/testfile --name=job1 --group_reporting & sleep 5; time btrfs sub snap -r /mnt/sub /mnt/snap; killall fio original: 15.83sec patched: 10.35sec The improvement is smaller compared to the sequential write case, since we omit only half of the pages dirtied after snapshot command. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Ethan Lien <ethanlien@synology.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-11-01 14:49:03 +08:00
/*
* We do the tagged writepage as long as the snapshot flush bit is set
* and we are the first one who do the filemap_flush() on this inode.
*
* The nr_to_write == LONG_MAX is needed to make sure other flushers do
* not race in and drop the bit.
*/
if (range_whole && wbc->nr_to_write == LONG_MAX &&
test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
&BTRFS_I(inode)->runtime_flags))
wbc->tagged_writepages = 1;
if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
tag = PAGECACHE_TAG_TOWRITE;
else
tag = PAGECACHE_TAG_DIRTY;
retry:
btrfs: use tagged writepage to mitigate livelock of snapshot Snapshot is expected to be fast. But if there are writers steadily creating dirty pages in our subvolume, the snapshot may take a very long time to complete. To fix the problem, we use tagged writepage for snapshot flusher as we do in the generic write_cache_pages(), so we can omit pages dirtied after the snapshot command. This does not change the semantics regarding which data get to the snapshot, if there are pages being dirtied during the snapshotting operation. There's a sync called before snapshot is taken in old/new case, any IO in flight just after that may be in the snapshot but this depends on other system effects that might still sync the IO. We do a simple snapshot speed test on a Intel D-1531 box: fio --ioengine=libaio --iodepth=32 --bs=4k --rw=write --size=64G --direct=0 --thread=1 --numjobs=1 --time_based --runtime=120 --filename=/mnt/sub/testfile --name=job1 --group_reporting & sleep 5; time btrfs sub snap -r /mnt/sub /mnt/snap; killall fio original: 1m58sec patched: 6.54sec This is the best case for this patch since for a sequential write case, we omit nearly all pages dirtied after the snapshot command. For a multi writers, random write test: fio --ioengine=libaio --iodepth=32 --bs=4k --rw=randwrite --size=64G --direct=0 --thread=1 --numjobs=4 --time_based --runtime=120 --filename=/mnt/sub/testfile --name=job1 --group_reporting & sleep 5; time btrfs sub snap -r /mnt/sub /mnt/snap; killall fio original: 15.83sec patched: 10.35sec The improvement is smaller compared to the sequential write case, since we omit only half of the pages dirtied after snapshot command. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Ethan Lien <ethanlien@synology.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-11-01 14:49:03 +08:00
if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
tag_pages_for_writeback(mapping, index, end);
done_index = index;
while (!done && !nr_to_write_done && (index <= end) &&
(nr_pages = pagevec_lookup_range_tag(&pvec, mapping,
&index, end, tag))) {
unsigned i;
for (i = 0; i < nr_pages; i++) {
struct page *page = pvec.pages[i];
btrfs: Avoid getting stuck during cyclic writebacks During a cyclic writeback, extent_write_cache_pages() uses done_index to update the writeback_index after the current run is over. However, instead of current index + 1, it gets to to the current index itself. Unfortunately, this, combined with returning on EOF instead of looping back, can lead to the following pathlogical behavior. 1. There is a single file which has accumulated enough dirty pages to trigger balance_dirty_pages() and the writer appending to the file with a series of short writes. 2. balance_dirty_pages kicks in, wakes up background writeback and sleeps. 3. Writeback kicks in and the cursor is on the last page of the dirty file. Writeback is started or skipped if already in progress. As it's EOF, extent_write_cache_pages() returns and the cursor is set to done_index which is pointing to the last page. 4. Writeback is done. Nothing happens till balance_dirty_pages finishes, at which point we go back to #1. This can almost completely stall out writing back of the file and keep the system over dirty threshold for a long time which can mess up the whole system. We encountered this issue in production with a package handling application which can reliably reproduce the issue when running under tight memory limits. Reading the comment in the error handling section, this seems to be to avoid accidentally skipping a page in case the write attempt on the page doesn't succeed. However, this concern seems bogus. On each page, the code either: * Skips and moves onto the next page. * Fails issue and sets done_index to index + 1. * Successfully issues and continue to the next page if budget allows and not EOF. IOW, as long as it's not EOF and there's budget, the code never retries writing back the same page. Only when a page happens to be the last page of a particular run, we end up retrying the page, which can't possibly guarantee anything data integrity related. Besides, cyclic writes are only used for non-syncing writebacks meaning that there's no data integrity implication to begin with. Fix it by always setting done_index past the current page being processed. Note that this problem exists in other writepages too. CC: stable@vger.kernel.org # 4.19+ Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-10-03 22:27:13 +08:00
done_index = page->index + 1;
/*
* At this point we hold neither the i_pages lock nor
* the page lock: the page may be truncated or
* invalidated (changing page->mapping to NULL),
* or even swizzled back from swapper_space to
* tmpfs file mapping
*/
if (!trylock_page(page)) {
submit_write_bio(bio_ctrl, 0);
lock_page(page);
}
if (unlikely(page->mapping != mapping)) {
unlock_page(page);
continue;
}
if (wbc->sync_mode != WB_SYNC_NONE) {
btrfs: avoid double clean up when submit_one_bio() failed [BUG] When running generic/475 with 64K page size and 4K sector size, it has a very high chance (almost 100%) to hang, with mostly data page locked but no one is going to unlock it. [CAUSE] With commit 1784b7d502a9 ("btrfs: handle csum lookup errors properly on reads"), if we failed to lookup checksum due to metadata IO error, we will return error for btrfs_submit_data_bio(). This will cause the page to be unlocked twice in btrfs_do_readpage(): btrfs_do_readpage() |- submit_extent_page() | |- submit_one_bio() | |- btrfs_submit_data_bio() | |- if (ret) { | |- bio->bi_status = ret; | |- bio_endio(bio); } | In the endio function, we will call end_page_read() | and unlock_extent() to cleanup the subpage range. | |- if (ret) { |- unlock_extent(); end_page_read() } Here we unlock the extent and cleanup the subpage range again. For unlock_extent(), it's mostly double unlock safe. But for end_page_read(), it's not, especially for subpage case, as for subpage case we will call btrfs_subpage_end_reader() to reduce the reader number, and use that to number to determine if we need to unlock the full page. If double accounted, it can underflow the number and leave the page locked without anyone to unlock it. [FIX] The commit 1784b7d502a9 ("btrfs: handle csum lookup errors properly on reads") itself is completely fine, it's our existing code not properly handling the error from bio submission hook properly. This patch will make submit_one_bio() to return void so that the callers will never be able to do cleanup when bio submission hook fails. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-04-12 20:30:13 +08:00
if (PageWriteback(page))
submit_write_bio(bio_ctrl, 0);
wait_on_page_writeback(page);
}
if (PageWriteback(page) ||
!clear_page_dirty_for_io(page)) {
unlock_page(page);
continue;
}
ret = __extent_writepage(page, wbc, bio_ctrl);
if (ret < 0) {
done = 1;
break;
}
/*
* the filesystem may choose to bump up nr_to_write.
* We have to make sure to honor the new nr_to_write
* at any time
*/
nr_to_write_done = wbc->nr_to_write <= 0;
}
pagevec_release(&pvec);
cond_resched();
}
if (!scanned && !done) {
/*
* We hit the last page and there is more work to be done: wrap
* back to the start of the file
*/
scanned = 1;
index = 0;
btrfs: flush write bio if we loop in extent_write_cache_pages There exists a deadlock with range_cyclic that has existed forever. If we loop around with a bio already built we could deadlock with a writer who has the page locked that we're attempting to write but is waiting on a page in our bio to be written out. The task traces are as follows PID: 1329874 TASK: ffff889ebcdf3800 CPU: 33 COMMAND: "kworker/u113:5" #0 [ffffc900297bb658] __schedule at ffffffff81a4c33f #1 [ffffc900297bb6e0] schedule at ffffffff81a4c6e3 #2 [ffffc900297bb6f8] io_schedule at ffffffff81a4ca42 #3 [ffffc900297bb708] __lock_page at ffffffff811f145b #4 [ffffc900297bb798] __process_pages_contig at ffffffff814bc502 #5 [ffffc900297bb8c8] lock_delalloc_pages at ffffffff814bc684 #6 [ffffc900297bb900] find_lock_delalloc_range at ffffffff814be9ff #7 [ffffc900297bb9a0] writepage_delalloc at ffffffff814bebd0 #8 [ffffc900297bba18] __extent_writepage at ffffffff814bfbf2 #9 [ffffc900297bba98] extent_write_cache_pages at ffffffff814bffbd PID: 2167901 TASK: ffff889dc6a59c00 CPU: 14 COMMAND: "aio-dio-invalid" #0 [ffffc9003b50bb18] __schedule at ffffffff81a4c33f #1 [ffffc9003b50bba0] schedule at ffffffff81a4c6e3 #2 [ffffc9003b50bbb8] io_schedule at ffffffff81a4ca42 #3 [ffffc9003b50bbc8] wait_on_page_bit at ffffffff811f24d6 #4 [ffffc9003b50bc60] prepare_pages at ffffffff814b05a7 #5 [ffffc9003b50bcd8] btrfs_buffered_write at ffffffff814b1359 #6 [ffffc9003b50bdb0] btrfs_file_write_iter at ffffffff814b5933 #7 [ffffc9003b50be38] new_sync_write at ffffffff8128f6a8 #8 [ffffc9003b50bec8] vfs_write at ffffffff81292b9d #9 [ffffc9003b50bf00] ksys_pwrite64 at ffffffff81293032 I used drgn to find the respective pages we were stuck on page_entry.page 0xffffea00fbfc7500 index 8148 bit 15 pid 2167901 page_entry.page 0xffffea00f9bb7400 index 7680 bit 0 pid 1329874 As you can see the kworker is waiting for bit 0 (PG_locked) on index 7680, and aio-dio-invalid is waiting for bit 15 (PG_writeback) on index 8148. aio-dio-invalid has 7680, and the kworker epd looks like the following crash> struct extent_page_data ffffc900297bbbb0 struct extent_page_data { bio = 0xffff889f747ed830, tree = 0xffff889eed6ba448, extent_locked = 0, sync_io = 0 } Probably worth mentioning as well that it waits for writeback of the page to complete while holding a lock on it (at prepare_pages()). Using drgn I walked the bio pages looking for page 0xffffea00fbfc7500 which is the one we're waiting for writeback on bio = Object(prog, 'struct bio', address=0xffff889f747ed830) for i in range(0, bio.bi_vcnt.value_()): bv = bio.bi_io_vec[i] if bv.bv_page.value_() == 0xffffea00fbfc7500: print("FOUND IT") which validated what I suspected. The fix for this is simple, flush the epd before we loop back around to the beginning of the file during writeout. Fixes: b293f02e1423 ("Btrfs: Add writepages support") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-01-24 04:33:02 +08:00
/*
* If we're looping we could run into a page that is locked by a
* writer and that writer could be waiting on writeback for a
* page in our current bio, and thus deadlock, so flush the
* write bio here.
*/
submit_write_bio(bio_ctrl, 0);
btrfs: avoid double clean up when submit_one_bio() failed [BUG] When running generic/475 with 64K page size and 4K sector size, it has a very high chance (almost 100%) to hang, with mostly data page locked but no one is going to unlock it. [CAUSE] With commit 1784b7d502a9 ("btrfs: handle csum lookup errors properly on reads"), if we failed to lookup checksum due to metadata IO error, we will return error for btrfs_submit_data_bio(). This will cause the page to be unlocked twice in btrfs_do_readpage(): btrfs_do_readpage() |- submit_extent_page() | |- submit_one_bio() | |- btrfs_submit_data_bio() | |- if (ret) { | |- bio->bi_status = ret; | |- bio_endio(bio); } | In the endio function, we will call end_page_read() | and unlock_extent() to cleanup the subpage range. | |- if (ret) { |- unlock_extent(); end_page_read() } Here we unlock the extent and cleanup the subpage range again. For unlock_extent(), it's mostly double unlock safe. But for end_page_read(), it's not, especially for subpage case, as for subpage case we will call btrfs_subpage_end_reader() to reduce the reader number, and use that to number to determine if we need to unlock the full page. If double accounted, it can underflow the number and leave the page locked without anyone to unlock it. [FIX] The commit 1784b7d502a9 ("btrfs: handle csum lookup errors properly on reads") itself is completely fine, it's our existing code not properly handling the error from bio submission hook properly. This patch will make submit_one_bio() to return void so that the callers will never be able to do cleanup when bio submission hook fails. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-04-12 20:30:13 +08:00
goto retry;
}
if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
mapping->writeback_index = done_index;
btrfs_add_delayed_iput(BTRFS_I(inode));
return ret;
}
/*
* Submit the pages in the range to bio for call sites which delalloc range has
* already been ran (aka, ordered extent inserted) and all pages are still
* locked.
*/
int extent_write_locked_range(struct inode *inode, u64 start, u64 end)
{
bool found_error = false;
int first_error = 0;
int ret = 0;
struct address_space *mapping = inode->i_mapping;
struct page *page;
u64 cur = start;
unsigned long nr_pages;
const u32 sectorsize = btrfs_sb(inode->i_sb)->sectorsize;
struct btrfs_bio_ctrl bio_ctrl = {
.extent_locked = 1,
.sync_io = 1,
};
struct writeback_control wbc_writepages = {
.sync_mode = WB_SYNC_ALL,
.range_start = start,
.range_end = end + 1,
/* We're called from an async helper function */
.punt_to_cgroup = 1,
.no_cgroup_owner = 1,
};
ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(end + 1, sectorsize));
nr_pages = (round_up(end, PAGE_SIZE) - round_down(start, PAGE_SIZE)) >>
PAGE_SHIFT;
wbc_writepages.nr_to_write = nr_pages * 2;
wbc_attach_fdatawrite_inode(&wbc_writepages, inode);
while (cur <= end) {
u64 cur_end = min(round_down(cur, PAGE_SIZE) + PAGE_SIZE - 1, end);
page = find_get_page(mapping, cur >> PAGE_SHIFT);
/*
* All pages in the range are locked since
* btrfs_run_delalloc_range(), thus there is no way to clear
* the page dirty flag.
*/
ASSERT(PageLocked(page));
ASSERT(PageDirty(page));
clear_page_dirty_for_io(page);
ret = __extent_writepage(page, &wbc_writepages, &bio_ctrl);
ASSERT(ret <= 0);
if (ret < 0) {
found_error = true;
first_error = ret;
}
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
put_page(page);
cur = cur_end + 1;
}
submit_write_bio(&bio_ctrl, found_error ? ret : 0);
wbc_detach_inode(&wbc_writepages);
if (found_error)
return first_error;
return ret;
}
int extent_writepages(struct address_space *mapping,
struct writeback_control *wbc)
{
struct inode *inode = mapping->host;
int ret = 0;
struct btrfs_bio_ctrl bio_ctrl = {
.extent_locked = 0,
.sync_io = (wbc->sync_mode == WB_SYNC_ALL),
};
/*
* Allow only a single thread to do the reloc work in zoned mode to
* protect the write pointer updates.
*/
btrfs_zoned_data_reloc_lock(BTRFS_I(inode));
ret = extent_write_cache_pages(mapping, wbc, &bio_ctrl);
submit_write_bio(&bio_ctrl, ret);
btrfs_zoned_data_reloc_unlock(BTRFS_I(inode));
return ret;
}
void extent_readahead(struct readahead_control *rac)
{
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
struct btrfs_bio_ctrl bio_ctrl = { 0 };
struct page *pagepool[16];
struct extent_map *em_cached = NULL;
Btrfs: update fix for read corruption of compressed and shared extents My previous fix in commit 005efedf2c7d ("Btrfs: fix read corruption of compressed and shared extents") was effective only if the compressed extents cover a file range with a length that is not a multiple of 16 pages. That's because the detection of when we reached a different range of the file that shares the same compressed extent as the previously processed range was done at extent_io.c:__do_contiguous_readpages(), which covers subranges with a length up to 16 pages, because extent_readpages() groups the pages in clusters no larger than 16 pages. So fix this by tracking the start of the previously processed file range's extent map at extent_readpages(). The following test case for fstests reproduces the issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_cloner rm -f $seqres.full test_clone_and_read_compressed_extent() { local mount_opts=$1 _scratch_mkfs >>$seqres.full 2>&1 _scratch_mount $mount_opts # Create our test file with a single extent of 64Kb that is going to # be compressed no matter which compression algo is used (zlib/lzo). $XFS_IO_PROG -f -c "pwrite -S 0xaa 0K 64K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now clone the compressed extent into an adjacent file offset. $CLONER_PROG -s 0 -d $((64 * 1024)) -l $((64 * 1024)) \ $SCRATCH_MNT/foo $SCRATCH_MNT/foo echo "File digest before unmount:" md5sum $SCRATCH_MNT/foo | _filter_scratch # Remount the fs or clear the page cache to trigger the bug in # btrfs. Because the extent has an uncompressed length that is a # multiple of 16 pages, all the pages belonging to the second range # of the file (64K to 128K), which points to the same extent as the # first range (0K to 64K), had their contents full of zeroes instead # of the byte 0xaa. This was a bug exclusively in the read path of # compressed extents, the correct data was stored on disk, btrfs # just failed to fill in the pages correctly. _scratch_remount echo "File digest after remount:" # Must match the digest we got before. md5sum $SCRATCH_MNT/foo | _filter_scratch } echo -e "\nTesting with zlib compression..." test_clone_and_read_compressed_extent "-o compress=zlib" _scratch_unmount echo -e "\nTesting with lzo compression..." test_clone_and_read_compressed_extent "-o compress=lzo" status=0 exit Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Tested-by: Timofey Titovets <nefelim4ag@gmail.com>
2015-09-28 16:56:26 +08:00
u64 prev_em_start = (u64)-1;
int nr;
while ((nr = readahead_page_batch(rac, pagepool))) {
u64 contig_start = readahead_pos(rac);
u64 contig_end = contig_start + readahead_batch_length(rac) - 1;
contiguous_readpages(pagepool, nr, contig_start, contig_end,
btrfs: refactor submit_extent_page() to make bio and its flag tracing easier There is a lot of code inside extent_io.c needs both "struct bio **bio_ret" and "unsigned long prev_bio_flags", along with some parameters like "unsigned long bio_flags". Such strange parameters are here for bio assembly. For example, we have such inode page layout: 0 4K 8K 12K |<-- Extent A-->|<- EB->| Then what we do is: - Page [0, 4K) *bio_ret = NULL So we allocate a new bio to bio_ret, Add page [0, 4K) to *bio_ret. - Page [4K, 8K) *bio_ret != NULL We found this page is continuous to *bio_ret, and if we're not at stripe boundary, we add page [4K, 8K) to *bio_ret. - Page [8K, 12K) *bio_ret != NULL But we found this page is not continuous, so we submit *bio_ret, then allocate a new bio, and add page [8K, 12K) to the new bio. This means we need to record both the bio and its bio_flag, but we record them manually using those strange parameter list, other than encapsulating them into their own structure. So this patch will introduce a new structure, btrfs_bio_ctrl, to record both the bio, and its bio_flags. Also, in above case, for all pages added to the bio, we need to check if the new page crosses stripe boundary. This check itself can be time consuming, and we don't really need to do that for each page. This patch also integrates the stripe boundary check into btrfs_bio_ctrl. When a new bio is allocated, the stripe and ordered extent boundary is also calculated, so no matter how large the bio will be, we only calculate the boundaries once, to save some CPU time. The following functions/structures are affected: - struct extent_page_data Replace its bio pointer with structure btrfs_bio_ctrl (embedded structure, not pointer) - end_write_bio() - flush_write_bio() Just change how bio is fetched - btrfs_bio_add_page() Use pre-calculated boundaries instead of re-calculating them. And use @bio_ctrl to replace @bio and @prev_bio_flags. - calc_bio_boundaries() New function - submit_extent_page() callers - btrfs_do_readpage() callers - contiguous_readpages() callers To Use @bio_ctrl to replace @bio and @prev_bio_flags, and how to grab bio. - btrfs_bio_fits_in_ordered_extent() Removed, as now the ordered extent size limit is done at bio allocation time, no need to check for each page range. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-14 16:42:15 +08:00
&em_cached, &bio_ctrl, &prev_em_start);
}
if (em_cached)
free_extent_map(em_cached);
submit_one_bio(&bio_ctrl);
}
/*
* basic invalidate_folio code, this waits on any locked or writeback
* ranges corresponding to the folio, and then deletes any extent state
* records from the tree
*/
int extent_invalidate_folio(struct extent_io_tree *tree,
struct folio *folio, size_t offset)
{
struct extent_state *cached_state = NULL;
u64 start = folio_pos(folio);
u64 end = start + folio_size(folio) - 1;
size_t blocksize = folio->mapping->host->i_sb->s_blocksize;
/* This function is only called for the btree inode */
ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
start += ALIGN(offset, blocksize);
if (start > end)
return 0;
lock_extent(tree, start, end, &cached_state);
folio_wait_writeback(folio);
/*
* Currently for btree io tree, only EXTENT_LOCKED is utilized,
* so here we only need to unlock the extent range to free any
* existing extent state.
*/
unlock_extent(tree, start, end, &cached_state);
return 0;
}
/*
* a helper for release_folio, this tests for areas of the page that
* are locked or under IO and drops the related state bits if it is safe
* to drop the page.
*/
static int try_release_extent_state(struct extent_io_tree *tree,
struct page *page, gfp_t mask)
{
u64 start = page_offset(page);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
u64 end = start + PAGE_SIZE - 1;
int ret = 1;
if (test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL)) {
ret = 0;
} else {
u32 clear_bits = ~(EXTENT_LOCKED | EXTENT_NODATASUM |
EXTENT_DELALLOC_NEW | EXTENT_CTLBITS);
/*
btrfs: update the number of bytes used by an inode atomically There are several occasions where we do not update the inode's number of used bytes atomically, resulting in a concurrent stat(2) syscall to report a value of used blocks that does not correspond to a valid value, that is, a value that does not match neither what we had before the operation nor what we get after the operation completes. In extreme cases it can result in stat(2) reporting zero used blocks, which can cause problems for some userspace tools where they can consider a file with a non-zero size and zero used blocks as completely sparse and skip reading data, as reported/discussed a long time ago in some threads like the following: https://lists.gnu.org/archive/html/bug-tar/2016-07/msg00001.html The cases where this can happen are the following: -> Case 1 If we do a write (buffered or direct IO) against a file region for which there is already an allocated extent (or multiple extents), then we have a short time window where we can report a number of used blocks to stat(2) that does not take into account the file region being overwritten. This short time window happens when completing the ordered extent(s). This happens because when we drop the extents in the write range we decrement the inode's number of bytes and later on when we insert the new extent(s) we increment the number of bytes in the inode, resulting in a short time window where a stat(2) syscall can get an incorrect number of used blocks. If we do writes that overwrite an entire file, then we have a short time window where we report 0 used blocks to stat(2). Example reproducer: $ cat reproducer-1.sh #!/bin/bash MNT=/mnt/sdi DEV=/dev/sdi stat_loop() { trap "wait; exit" SIGTERM local filepath=$1 local expected=$2 local got while :; do got=$(stat -c %b $filepath) if [ $got -ne $expected ]; then echo -n "ERROR: unexpected used blocks" echo " (got: $got expected: $expected)" fi done } mkfs.btrfs -f $DEV > /dev/null # mkfs.xfs -f $DEV > /dev/null # mkfs.ext4 -F $DEV > /dev/null # mkfs.f2fs -f $DEV > /dev/null # mkfs.reiserfs -f $DEV > /dev/null mount $DEV $MNT xfs_io -f -s -c "pwrite -b 64K 0 64K" $MNT/foobar >/dev/null expected=$(stat -c %b $MNT/foobar) # Create a process to keep calling stat(2) on the file and see if the # reported number of blocks used (disk space used) changes, it should # not because we are not increasing the file size nor punching holes. stat_loop $MNT/foobar $expected & loop_pid=$! for ((i = 0; i < 50000; i++)); do xfs_io -s -c "pwrite -b 64K 0 64K" $MNT/foobar >/dev/null done kill $loop_pid &> /dev/null wait umount $DEV $ ./reproducer-1.sh ERROR: unexpected used blocks (got: 0 expected: 128) ERROR: unexpected used blocks (got: 0 expected: 128) (...) Note that since this is a short time window where the race can happen, the reproducer may not be able to always trigger the bug in one run, or it may trigger it multiple times. -> Case 2 If we do a buffered write against a file region that does not have any allocated extents, like a hole or beyond EOF, then during ordered extent completion we have a short time window where a concurrent stat(2) syscall can report a number of used blocks that does not correspond to the value before or after the write operation, a value that is actually larger than the value after the write completes. This happens because once we start a buffered write into an unallocated file range we increment the inode's 'new_delalloc_bytes', to make sure any stat(2) call gets a correct used blocks value before delalloc is flushed and completes. However at ordered extent completion, after we inserted the new extent, we increment the inode's number of bytes used with the size of the new extent, and only later, when clearing the range in the inode's iotree, we decrement the inode's 'new_delalloc_bytes' counter with the size of the extent. So this results in a short time window where a concurrent stat(2) syscall can report a number of used blocks that accounts for the new extent twice. Example reproducer: $ cat reproducer-2.sh #!/bin/bash MNT=/mnt/sdi DEV=/dev/sdi stat_loop() { trap "wait; exit" SIGTERM local filepath=$1 local expected=$2 local got while :; do got=$(stat -c %b $filepath) if [ $got -ne $expected ]; then echo -n "ERROR: unexpected used blocks" echo " (got: $got expected: $expected)" fi done } mkfs.btrfs -f $DEV > /dev/null # mkfs.xfs -f $DEV > /dev/null # mkfs.ext4 -F $DEV > /dev/null # mkfs.f2fs -f $DEV > /dev/null # mkfs.reiserfs -f $DEV > /dev/null mount $DEV $MNT touch $MNT/foobar write_size=$((64 * 1024)) for ((i = 0; i < 16384; i++)); do offset=$(($i * $write_size)) xfs_io -c "pwrite -S 0xab $offset $write_size" $MNT/foobar >/dev/null blocks_used=$(stat -c %b $MNT/foobar) # Fsync the file to trigger writeback and keep calling stat(2) on it # to see if the number of blocks used changes. stat_loop $MNT/foobar $blocks_used & loop_pid=$! xfs_io -c "fsync" $MNT/foobar kill $loop_pid &> /dev/null wait $loop_pid done umount $DEV $ ./reproducer-2.sh ERROR: unexpected used blocks (got: 265472 expected: 265344) ERROR: unexpected used blocks (got: 284032 expected: 283904) (...) Note that since this is a short time window where the race can happen, the reproducer may not be able to always trigger the bug in one run, or it may trigger it multiple times. -> Case 3 Another case where such problems happen is during other operations that replace extents in a file range with other extents. Those operations are extent cloning, deduplication and fallocate's zero range operation. The cause of the problem is similar to the first case. When we drop the extents from a range, we decrement the inode's number of bytes, and later on, after inserting the new extents we increment it. Since this is not done atomically, a concurrent stat(2) call can see and return a number of used blocks that is smaller than it should be, does not match the number of used blocks before or after the clone/deduplication/zero operation. Like for the first case, when doing a clone, deduplication or zero range operation against an entire file, we end up having a time window where we can report 0 used blocks to a stat(2) call. Example reproducer: $ cat reproducer-3.sh #!/bin/bash MNT=/mnt/sdi DEV=/dev/sdi mkfs.btrfs -f $DEV > /dev/null # mkfs.xfs -f -m reflink=1 $DEV > /dev/null mount $DEV $MNT extent_size=$((64 * 1024)) num_extents=16384 file_size=$(($extent_size * $num_extents)) # File foo has many small extents. xfs_io -f -s -c "pwrite -S 0xab -b $extent_size 0 $file_size" $MNT/foo \ > /dev/null # File bar has much less extents and has exactly the same data as foo. xfs_io -f -c "pwrite -S 0xab 0 $file_size" $MNT/bar > /dev/null expected=$(stat -c %b $MNT/foo) # Now deduplicate bar into foo. While the deduplication is in progres, # the number of used blocks/file size reported by stat should not change xfs_io -c "dedupe $MNT/bar 0 0 $file_size" $MNT/foo > /dev/null & dedupe_pid=$! while [ -n "$(ps -p $dedupe_pid -o pid=)" ]; do used=$(stat -c %b $MNT/foo) if [ $used -ne $expected ]; then echo "Unexpected blocks used: $used (expected: $expected)" fi done umount $DEV $ ./reproducer-3.sh Unexpected blocks used: 2076800 (expected: 2097152) Unexpected blocks used: 2097024 (expected: 2097152) Unexpected blocks used: 2079872 (expected: 2097152) (...) Note that since this is a short time window where the race can happen, the reproducer may not be able to always trigger the bug in one run, or it may trigger it multiple times. So fix this by: 1) Making btrfs_drop_extents() not decrement the VFS inode's number of bytes, and instead return the number of bytes; 2) Making any code that drops extents and adds new extents update the inode's number of bytes atomically, while holding the btrfs inode's spinlock, which is also used by the stat(2) callback to get the inode's number of bytes; 3) For ranges in the inode's iotree that are marked as 'delalloc new', corresponding to previously unallocated ranges, increment the inode's number of bytes when clearing the 'delalloc new' bit from the range, in the same critical section that decrements the inode's 'new_delalloc_bytes' counter, delimited by the btrfs inode's spinlock. An alternative would be to have btrfs_getattr() wait for any IO (ordered extents in progress) and locking the whole range (0 to (u64)-1) while it it computes the number of blocks used. But that would mean blocking stat(2), which is a very used syscall and expected to be fast, waiting for writes, clone/dedupe, fallocate, page reads, fiemap, etc. CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-04 19:07:34 +08:00
* At this point we can safely clear everything except the
* locked bit, the nodatasum bit and the delalloc new bit.
* The delalloc new bit will be cleared by ordered extent
* completion.
*/
ret = __clear_extent_bit(tree, start, end, clear_bits, NULL,
mask, NULL);
/* if clear_extent_bit failed for enomem reasons,
* we can't allow the release to continue.
*/
if (ret < 0)
ret = 0;
else
ret = 1;
}
return ret;
}
/*
* a helper for release_folio. As long as there are no locked extents
* in the range corresponding to the page, both state records and extent
* map records are removed
*/
int try_release_extent_mapping(struct page *page, gfp_t mask)
{
struct extent_map *em;
u64 start = page_offset(page);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
u64 end = start + PAGE_SIZE - 1;
Btrfs: fix file data corruption after cloning a range and fsync When we clone a range into a file we can end up dropping existing extent maps (or trimming them) and replacing them with new ones if the range to be cloned overlaps with a range in the destination inode. When that happens we add the new extent maps to the list of modified extents in the inode's extent map tree, so that a "fast" fsync (the flag BTRFS_INODE_NEEDS_FULL_SYNC not set in the inode) will see the extent maps and log corresponding extent items. However, at the end of range cloning operation we do truncate all the pages in the affected range (in order to ensure future reads will not get stale data). Sometimes this truncation will release the corresponding extent maps besides the pages from the page cache. If this happens, then a "fast" fsync operation will miss logging some extent items, because it relies exclusively on the extent maps being present in the inode's extent tree, leading to data loss/corruption if the fsync ends up using the same transaction used by the clone operation (that transaction was not committed in the meanwhile). An extent map is released through the callback btrfs_invalidatepage(), which gets called by truncate_inode_pages_range(), and it calls __btrfs_releasepage(). The later ends up calling try_release_extent_mapping() which will release the extent map if some conditions are met, like the file size being greater than 16Mb, gfp flags allow blocking and the range not being locked (which is the case during the clone operation) nor being the extent map flagged as pinned (also the case for cloning). The following example, turned into a test for fstests, reproduces the issue: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ xfs_io -f -c "pwrite -S 0x18 9000K 6908K" /mnt/foo $ xfs_io -f -c "pwrite -S 0x20 2572K 156K" /mnt/bar $ xfs_io -c "fsync" /mnt/bar # reflink destination offset corresponds to the size of file bar, # 2728Kb minus 4Kb. $ xfs_io -c ""reflink ${SCRATCH_MNT}/foo 0 2724K 15908K" /mnt/bar $ xfs_io -c "fsync" /mnt/bar $ md5sum /mnt/bar 95a95813a8c2abc9aa75a6c2914a077e /mnt/bar <power fail> $ mount /dev/sdb /mnt $ md5sum /mnt/bar 207fd8d0b161be8a84b945f0df8d5f8d /mnt/bar # digest should be 95a95813a8c2abc9aa75a6c2914a077e like before the # power failure In the above example, the destination offset of the clone operation corresponds to the size of the "bar" file minus 4Kb. So during the clone operation, the extent map covering the range from 2572Kb to 2728Kb gets trimmed so that it ends at offset 2724Kb, and a new extent map covering the range from 2724Kb to 11724Kb is created. So at the end of the clone operation when we ask to truncate the pages in the range from 2724Kb to 2724Kb + 15908Kb, the page invalidation callback ends up removing the new extent map (through try_release_extent_mapping()) when the page at offset 2724Kb is passed to that callback. Fix this by setting the bit BTRFS_INODE_NEEDS_FULL_SYNC whenever an extent map is removed at try_release_extent_mapping(), forcing the next fsync to search for modified extents in the fs/subvolume tree instead of relying on the presence of extent maps in memory. This way we can continue doing a "fast" fsync if the destination range of a clone operation does not overlap with an existing range or if any of the criteria necessary to remove an extent map at try_release_extent_mapping() is not met (file size not bigger then 16Mb or gfp flags do not allow blocking). CC: stable@vger.kernel.org # 3.16+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-07-12 08:36:43 +08:00
struct btrfs_inode *btrfs_inode = BTRFS_I(page->mapping->host);
struct extent_io_tree *tree = &btrfs_inode->io_tree;
struct extent_map_tree *map = &btrfs_inode->extent_tree;
mm, page_alloc: distinguish between being unable to sleep, unwilling to sleep and avoiding waking kswapd __GFP_WAIT has been used to identify atomic context in callers that hold spinlocks or are in interrupts. They are expected to be high priority and have access one of two watermarks lower than "min" which can be referred to as the "atomic reserve". __GFP_HIGH users get access to the first lower watermark and can be called the "high priority reserve". Over time, callers had a requirement to not block when fallback options were available. Some have abused __GFP_WAIT leading to a situation where an optimisitic allocation with a fallback option can access atomic reserves. This patch uses __GFP_ATOMIC to identify callers that are truely atomic, cannot sleep and have no alternative. High priority users continue to use __GFP_HIGH. __GFP_DIRECT_RECLAIM identifies callers that can sleep and are willing to enter direct reclaim. __GFP_KSWAPD_RECLAIM to identify callers that want to wake kswapd for background reclaim. __GFP_WAIT is redefined as a caller that is willing to enter direct reclaim and wake kswapd for background reclaim. This patch then converts a number of sites o __GFP_ATOMIC is used by callers that are high priority and have memory pools for those requests. GFP_ATOMIC uses this flag. o Callers that have a limited mempool to guarantee forward progress clear __GFP_DIRECT_RECLAIM but keep __GFP_KSWAPD_RECLAIM. bio allocations fall into this category where kswapd will still be woken but atomic reserves are not used as there is a one-entry mempool to guarantee progress. o Callers that are checking if they are non-blocking should use the helper gfpflags_allow_blocking() where possible. This is because checking for __GFP_WAIT as was done historically now can trigger false positives. Some exceptions like dm-crypt.c exist where the code intent is clearer if __GFP_DIRECT_RECLAIM is used instead of the helper due to flag manipulations. o Callers that built their own GFP flags instead of starting with GFP_KERNEL and friends now also need to specify __GFP_KSWAPD_RECLAIM. The first key hazard to watch out for is callers that removed __GFP_WAIT and was depending on access to atomic reserves for inconspicuous reasons. In some cases it may be appropriate for them to use __GFP_HIGH. The second key hazard is callers that assembled their own combination of GFP flags instead of starting with something like GFP_KERNEL. They may now wish to specify __GFP_KSWAPD_RECLAIM. It's almost certainly harmless if it's missed in most cases as other activity will wake kswapd. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Michal Hocko <mhocko@suse.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Vitaly Wool <vitalywool@gmail.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-07 08:28:21 +08:00
if (gfpflags_allow_blocking(mask) &&
page->mapping->host->i_size > SZ_16M) {
u64 len;
while (start <= end) {
struct btrfs_fs_info *fs_info;
u64 cur_gen;
len = end - start + 1;
write_lock(&map->lock);
em = lookup_extent_mapping(map, start, len);
if (!em) {
write_unlock(&map->lock);
break;
}
if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
em->start != start) {
write_unlock(&map->lock);
free_extent_map(em);
break;
}
btrfs: fix race between page release and a fast fsync When releasing an extent map, done through the page release callback, we can race with an ongoing fast fsync and cause the fsync to miss a new extent and not log it. The steps for this to happen are the following: 1) A page is dirtied for some inode I; 2) Writeback for that page is triggered by a path other than fsync, for example by the system due to memory pressure; 3) When the ordered extent for the extent (a single 4K page) finishes, we unpin the corresponding extent map and set its generation to N, the current transaction's generation; 4) The btrfs_releasepage() callback is invoked by the system due to memory pressure for that no longer dirty page of inode I; 5) At the same time, some task calls fsync on inode I, joins transaction N, and at btrfs_log_inode() it sees that the inode does not have the full sync flag set, so we proceed with a fast fsync. But before we get into btrfs_log_changed_extents() and lock the inode's extent map tree: 6) Through btrfs_releasepage() we end up at try_release_extent_mapping() and we remove the extent map for the new 4Kb extent, because it is neither pinned anymore nor locked. By calling remove_extent_mapping(), we remove the extent map from the list of modified extents, since the extent map does not have the logging flag set. We unlock the inode's extent map tree; 7) The task doing the fast fsync now enters btrfs_log_changed_extents(), locks the inode's extent map tree and iterates its list of modified extents, which no longer has the 4Kb extent in it, so it does not log the extent; 8) The fsync finishes; 9) Before transaction N is committed, a power failure happens. After replaying the log, the 4K extent of inode I will be missing, since it was not logged due to the race with try_release_extent_mapping(). So fix this by teaching try_release_extent_mapping() to not remove an extent map if it's still in the list of modified extents. Fixes: ff44c6e36dc9dc ("Btrfs: do not hold the write_lock on the extent tree while logging") CC: stable@vger.kernel.org # 5.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-07-22 19:28:37 +08:00
if (test_range_bit(tree, em->start,
extent_map_end(em) - 1,
EXTENT_LOCKED, 0, NULL))
goto next;
/*
* If it's not in the list of modified extents, used
* by a fast fsync, we can remove it. If it's being
* logged we can safely remove it since fsync took an
* extra reference on the em.
*/
if (list_empty(&em->list) ||
test_bit(EXTENT_FLAG_LOGGING, &em->flags))
goto remove_em;
/*
* If it's in the list of modified extents, remove it
* only if its generation is older then the current one,
* in which case we don't need it for a fast fsync.
* Otherwise don't remove it, we could be racing with an
* ongoing fast fsync that could miss the new extent.
*/
fs_info = btrfs_inode->root->fs_info;
spin_lock(&fs_info->trans_lock);
cur_gen = fs_info->generation;
spin_unlock(&fs_info->trans_lock);
if (em->generation >= cur_gen)
goto next;
remove_em:
btrfs: do not set the full sync flag on the inode during page release When removing an extent map at try_release_extent_mapping(), called through the page release callback (btrfs_releasepage()), we always set the full sync flag on the inode, which forces the next fsync to use a slower code path. This hurts performance for workloads that dirty an amount of data that exceeds or is very close to the system's RAM memory and do frequent fsync operations (like database servers can for example). In particular if there are concurrent fsyncs against different files, by falling back to a full fsync we do a lot more checksum lookups in the checksums btree, as we do it for all the extents created in the current transaction, instead of only the new ones since the last fsync. These checksums lookups not only take some time but, more importantly, they also cause contention on the checksums btree locks due to the concurrency with checksum insertions in the btree by ordered extents from other inodes. We actually don't need to set the full sync flag on the inode, because we only remove extent maps that are in the list of modified extents if they were created in a past transaction, in which case an fsync skips them as it's pointless to log them. So stop setting the full fsync flag on the inode whenever we remove an extent map. This patch is part of a patchset that consists of 3 patches, which have the following subjects: 1/3 btrfs: fix race between page release and a fast fsync 2/3 btrfs: release old extent maps during page release 3/3 btrfs: do not set the full sync flag on the inode during page release Performance tests were ran against a branch (misc-next) containing the whole patchset. The test exercises a workload where there are multiple processes writing to files and fsyncing them (each writing and fsyncing its own file), and in total the amount of data dirtied ranges from 2x to 4x the system's RAM memory (16GiB), so that the page release callback is invoked frequently. The following script, using fio, was used to perform the tests: $ cat test-fsync.sh #!/bin/bash DEV=/dev/sdk MNT=/mnt/sdk MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" if [ $# -ne 3 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 cat <<EOF > /tmp/fio-job.ini [writers] rw=write fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=64k ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS thread EOF echo "Using config:" echo cat /tmp/fio-job.ini echo mkfs.btrfs -f $MKFS_OPTIONS $DEV &> /dev/null mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The tests were performed for different numbers of jobs, file sizes and fsync frequency. A qemu VM using kvm was used, with 8 cores (the host has 12 cores, with cpu governance set to performance mode on all cores), 16GiB of ram (the host has 64GiB) and using a NVMe device directly (without an intermediary filesystem in the host). While running the tests, the host was not used for anything else, to avoid disturbing the tests. The obtained results were the following, and the last line printed by fio is pasted (includes aggregated throughput and test run time). ***************************************************** **** 1 job, 32GiB file, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=29.1MiB/s (30.5MB/s), 29.1MiB/s-29.1MiB/s (30.5MB/s-30.5MB/s), io=32.0GiB (34.4GB), run=1127557-1127557msec After patchset: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=32.0GiB (34.4GB), run=1119042-1119042msec (+0.7% throughput, -0.8% run time) ***************************************************** **** 2 jobs, 16GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=33.5MiB/s (35.1MB/s), 33.5MiB/s-33.5MiB/s (35.1MB/s-35.1MB/s), io=32.0GiB (34.4GB), run=979000-979000msec After patchset: WRITE: bw=39.9MiB/s (41.8MB/s), 39.9MiB/s-39.9MiB/s (41.8MB/s-41.8MB/s), io=32.0GiB (34.4GB), run=821283-821283msec (+19.1% throughput, -16.1% runtime) ***************************************************** **** 4 jobs, 8GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=52.1MiB/s (54.6MB/s), 52.1MiB/s-52.1MiB/s (54.6MB/s-54.6MB/s), io=32.0GiB (34.4GB), run=629130-629130msec After patchset: WRITE: bw=71.8MiB/s (75.3MB/s), 71.8MiB/s-71.8MiB/s (75.3MB/s-75.3MB/s), io=32.0GiB (34.4GB), run=456357-456357msec (+37.8% throughput, -27.5% runtime) ***************************************************** **** 8 jobs, 4GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=76.1MiB/s (79.8MB/s), 76.1MiB/s-76.1MiB/s (79.8MB/s-79.8MB/s), io=32.0GiB (34.4GB), run=430708-430708msec After patchset: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=245458-245458msec (+74.7% throughput, -43.0% run time) ***************************************************** **** 16 jobs, 2GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=74.7MiB/s (78.3MB/s), 74.7MiB/s-74.7MiB/s (78.3MB/s-78.3MB/s), io=32.0GiB (34.4GB), run=438625-438625msec After patchset: WRITE: bw=184MiB/s (193MB/s), 184MiB/s-184MiB/s (193MB/s-193MB/s), io=32.0GiB (34.4GB), run=177864-177864msec (+146.3% throughput, -59.5% run time) ***************************************************** **** 32 jobs, 2GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=72.6MiB/s (76.1MB/s), 72.6MiB/s-72.6MiB/s (76.1MB/s-76.1MB/s), io=64.0GiB (68.7GB), run=902615-902615msec After patchset: WRITE: bw=227MiB/s (238MB/s), 227MiB/s-227MiB/s (238MB/s-238MB/s), io=64.0GiB (68.7GB), run=288936-288936msec (+212.7% throughput, -68.0% run time) ***************************************************** **** 64 jobs, 1GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=98.8MiB/s (104MB/s), 98.8MiB/s-98.8MiB/s (104MB/s-104MB/s), io=64.0GiB (68.7GB), run=663126-663126msec After patchset: WRITE: bw=294MiB/s (308MB/s), 294MiB/s-294MiB/s (308MB/s-308MB/s), io=64.0GiB (68.7GB), run=222940-222940msec (+197.6% throughput, -66.4% run time) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-07-22 19:29:01 +08:00
/*
* We only remove extent maps that are not in the list of
* modified extents or that are in the list but with a
* generation lower then the current generation, so there
* is no need to set the full fsync flag on the inode (it
* hurts the fsync performance for workloads with a data
* size that exceeds or is close to the system's memory).
*/
remove_extent_mapping(map, em);
/* once for the rb tree */
free_extent_map(em);
btrfs: fix race between page release and a fast fsync When releasing an extent map, done through the page release callback, we can race with an ongoing fast fsync and cause the fsync to miss a new extent and not log it. The steps for this to happen are the following: 1) A page is dirtied for some inode I; 2) Writeback for that page is triggered by a path other than fsync, for example by the system due to memory pressure; 3) When the ordered extent for the extent (a single 4K page) finishes, we unpin the corresponding extent map and set its generation to N, the current transaction's generation; 4) The btrfs_releasepage() callback is invoked by the system due to memory pressure for that no longer dirty page of inode I; 5) At the same time, some task calls fsync on inode I, joins transaction N, and at btrfs_log_inode() it sees that the inode does not have the full sync flag set, so we proceed with a fast fsync. But before we get into btrfs_log_changed_extents() and lock the inode's extent map tree: 6) Through btrfs_releasepage() we end up at try_release_extent_mapping() and we remove the extent map for the new 4Kb extent, because it is neither pinned anymore nor locked. By calling remove_extent_mapping(), we remove the extent map from the list of modified extents, since the extent map does not have the logging flag set. We unlock the inode's extent map tree; 7) The task doing the fast fsync now enters btrfs_log_changed_extents(), locks the inode's extent map tree and iterates its list of modified extents, which no longer has the 4Kb extent in it, so it does not log the extent; 8) The fsync finishes; 9) Before transaction N is committed, a power failure happens. After replaying the log, the 4K extent of inode I will be missing, since it was not logged due to the race with try_release_extent_mapping(). So fix this by teaching try_release_extent_mapping() to not remove an extent map if it's still in the list of modified extents. Fixes: ff44c6e36dc9dc ("Btrfs: do not hold the write_lock on the extent tree while logging") CC: stable@vger.kernel.org # 5.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-07-22 19:28:37 +08:00
next:
start = extent_map_end(em);
write_unlock(&map->lock);
/* once for us */
free_extent_map(em);
fs/btrfs: Add cond_resched() for try_release_extent_mapping() stalls Very large I/Os can cause the following RCU CPU stall warning: RIP: 0010:rb_prev+0x8/0x50 Code: 49 89 c0 49 89 d1 48 89 c2 48 89 f8 e9 e5 fd ff ff 4c 89 48 10 c3 4c = 89 06 c3 4c 89 40 10 c3 0f 1f 00 48 8b 0f 48 39 cf 74 38 <48> 8b 47 10 48 85 c0 74 22 48 8b 50 08 48 85 d2 74 0c 48 89 d0 48 RSP: 0018:ffffc9002212bab0 EFLAGS: 00000287 ORIG_RAX: ffffffffffffff13 RAX: ffff888821f93630 RBX: ffff888821f93630 RCX: ffff888821f937e0 RDX: 0000000000000000 RSI: 0000000000102000 RDI: ffff888821f93630 RBP: 0000000000103000 R08: 000000000006c000 R09: 0000000000000238 R10: 0000000000102fff R11: ffffc9002212bac8 R12: 0000000000000001 R13: ffffffffffffffff R14: 0000000000102000 R15: ffff888821f937e0 __lookup_extent_mapping+0xa0/0x110 try_release_extent_mapping+0xdc/0x220 btrfs_releasepage+0x45/0x70 shrink_page_list+0xa39/0xb30 shrink_inactive_list+0x18f/0x3b0 shrink_lruvec+0x38e/0x6b0 shrink_node+0x14d/0x690 do_try_to_free_pages+0xc6/0x3e0 try_to_free_mem_cgroup_pages+0xe6/0x1e0 reclaim_high.constprop.73+0x87/0xc0 mem_cgroup_handle_over_high+0x66/0x150 exit_to_usermode_loop+0x82/0xd0 do_syscall_64+0xd4/0x100 entry_SYSCALL_64_after_hwframe+0x44/0xa9 On a PREEMPT=n kernel, the try_release_extent_mapping() function's "while" loop might run for a very long time on a large I/O. This commit therefore adds a cond_resched() to this loop, providing RCU any needed quiescent states. Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-05-09 05:15:37 +08:00
cond_resched(); /* Allow large-extent preemption. */
}
}
return try_release_extent_state(tree, page, mask);
}
btrfs: fiemap: Cache and merge fiemap extent before submit it to user [BUG] Cycle mount btrfs can cause fiemap to return different result. Like: # mount /dev/vdb5 /mnt/btrfs # dd if=/dev/zero bs=16K count=4 oflag=dsync of=/mnt/btrfs/file # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 # umount /mnt/btrfs # mount /dev/vdb5 /mnt/btrfs # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..31]: 25088..25119 32 0x0 1: [32..63]: 25120..25151 32 0x0 2: [64..95]: 25152..25183 32 0x0 3: [96..127]: 25184..25215 32 0x1 But after above fiemap, we get correct merged result if we call fiemap again. # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 [REASON] Btrfs will try to merge extent map when inserting new extent map. btrfs_fiemap(start=0 len=(u64)-1) |- extent_fiemap(start=0 len=(u64)-1) |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=0 len=64k) | | Found on-disk (ino, EXTENT_DATA, 0) | |- add_extent_mapping() | |- Return (em->start=0, len=16k) | |- fiemap_fill_next_extent(logic=0 phys=X len=16k) | |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=16k len=48k) | | Found on-disk (ino, EXTENT_DATA, 16k) | |- add_extent_mapping() | | |- try_merge_map() | | Merge with previous em start=0 len=16k | | resulting em start=0 len=32k | |- Return (em->start=0, len=32K) << Merged result |- Stripe off the unrelated range (0~16K) of return em |- fiemap_fill_next_extent(logic=16K phys=X+16K len=16K) ^^^ Causing split fiemap extent. And since in add_extent_mapping(), em is already merged, in next fiemap() call, we will get merged result. [FIX] Here we introduce a new structure, fiemap_cache, which records previous fiemap extent. And will always try to merge current fiemap_cache result before calling fiemap_fill_next_extent(). Only when we failed to merge current fiemap extent with cached one, we will call fiemap_fill_next_extent() to submit cached one. So by this method, we can merge all fiemap extents. It can also be done in fs/ioctl.c, however the problem is if fieinfo->fi_extents_max == 0, we have no space to cache previous fiemap extent. So I choose to merge it in btrfs. Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2017-04-07 10:43:15 +08:00
/*
* To cache previous fiemap extent
*
* Will be used for merging fiemap extent
*/
struct fiemap_cache {
u64 offset;
u64 phys;
u64 len;
u32 flags;
bool cached;
};
/*
* Helper to submit fiemap extent.
*
* Will try to merge current fiemap extent specified by @offset, @phys,
* @len and @flags with cached one.
* And only when we fails to merge, cached one will be submitted as
* fiemap extent.
*
* Return value is the same as fiemap_fill_next_extent().
*/
static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
struct fiemap_cache *cache,
u64 offset, u64 phys, u64 len, u32 flags)
{
int ret = 0;
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
/* Set at the end of extent_fiemap(). */
ASSERT((flags & FIEMAP_EXTENT_LAST) == 0);
btrfs: fiemap: Cache and merge fiemap extent before submit it to user [BUG] Cycle mount btrfs can cause fiemap to return different result. Like: # mount /dev/vdb5 /mnt/btrfs # dd if=/dev/zero bs=16K count=4 oflag=dsync of=/mnt/btrfs/file # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 # umount /mnt/btrfs # mount /dev/vdb5 /mnt/btrfs # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..31]: 25088..25119 32 0x0 1: [32..63]: 25120..25151 32 0x0 2: [64..95]: 25152..25183 32 0x0 3: [96..127]: 25184..25215 32 0x1 But after above fiemap, we get correct merged result if we call fiemap again. # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 [REASON] Btrfs will try to merge extent map when inserting new extent map. btrfs_fiemap(start=0 len=(u64)-1) |- extent_fiemap(start=0 len=(u64)-1) |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=0 len=64k) | | Found on-disk (ino, EXTENT_DATA, 0) | |- add_extent_mapping() | |- Return (em->start=0, len=16k) | |- fiemap_fill_next_extent(logic=0 phys=X len=16k) | |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=16k len=48k) | | Found on-disk (ino, EXTENT_DATA, 16k) | |- add_extent_mapping() | | |- try_merge_map() | | Merge with previous em start=0 len=16k | | resulting em start=0 len=32k | |- Return (em->start=0, len=32K) << Merged result |- Stripe off the unrelated range (0~16K) of return em |- fiemap_fill_next_extent(logic=16K phys=X+16K len=16K) ^^^ Causing split fiemap extent. And since in add_extent_mapping(), em is already merged, in next fiemap() call, we will get merged result. [FIX] Here we introduce a new structure, fiemap_cache, which records previous fiemap extent. And will always try to merge current fiemap_cache result before calling fiemap_fill_next_extent(). Only when we failed to merge current fiemap extent with cached one, we will call fiemap_fill_next_extent() to submit cached one. So by this method, we can merge all fiemap extents. It can also be done in fs/ioctl.c, however the problem is if fieinfo->fi_extents_max == 0, we have no space to cache previous fiemap extent. So I choose to merge it in btrfs. Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2017-04-07 10:43:15 +08:00
if (!cache->cached)
goto assign;
/*
* Sanity check, extent_fiemap() should have ensured that new
* fiemap extent won't overlap with cached one.
btrfs: fiemap: Cache and merge fiemap extent before submit it to user [BUG] Cycle mount btrfs can cause fiemap to return different result. Like: # mount /dev/vdb5 /mnt/btrfs # dd if=/dev/zero bs=16K count=4 oflag=dsync of=/mnt/btrfs/file # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 # umount /mnt/btrfs # mount /dev/vdb5 /mnt/btrfs # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..31]: 25088..25119 32 0x0 1: [32..63]: 25120..25151 32 0x0 2: [64..95]: 25152..25183 32 0x0 3: [96..127]: 25184..25215 32 0x1 But after above fiemap, we get correct merged result if we call fiemap again. # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 [REASON] Btrfs will try to merge extent map when inserting new extent map. btrfs_fiemap(start=0 len=(u64)-1) |- extent_fiemap(start=0 len=(u64)-1) |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=0 len=64k) | | Found on-disk (ino, EXTENT_DATA, 0) | |- add_extent_mapping() | |- Return (em->start=0, len=16k) | |- fiemap_fill_next_extent(logic=0 phys=X len=16k) | |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=16k len=48k) | | Found on-disk (ino, EXTENT_DATA, 16k) | |- add_extent_mapping() | | |- try_merge_map() | | Merge with previous em start=0 len=16k | | resulting em start=0 len=32k | |- Return (em->start=0, len=32K) << Merged result |- Stripe off the unrelated range (0~16K) of return em |- fiemap_fill_next_extent(logic=16K phys=X+16K len=16K) ^^^ Causing split fiemap extent. And since in add_extent_mapping(), em is already merged, in next fiemap() call, we will get merged result. [FIX] Here we introduce a new structure, fiemap_cache, which records previous fiemap extent. And will always try to merge current fiemap_cache result before calling fiemap_fill_next_extent(). Only when we failed to merge current fiemap extent with cached one, we will call fiemap_fill_next_extent() to submit cached one. So by this method, we can merge all fiemap extents. It can also be done in fs/ioctl.c, however the problem is if fieinfo->fi_extents_max == 0, we have no space to cache previous fiemap extent. So I choose to merge it in btrfs. Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2017-04-07 10:43:15 +08:00
* Not recoverable.
*
* NOTE: Physical address can overlap, due to compression
*/
if (cache->offset + cache->len > offset) {
WARN_ON(1);
return -EINVAL;
}
/*
* Only merges fiemap extents if
* 1) Their logical addresses are continuous
*
* 2) Their physical addresses are continuous
* So truly compressed (physical size smaller than logical size)
* extents won't get merged with each other
*
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
* 3) Share same flags
btrfs: fiemap: Cache and merge fiemap extent before submit it to user [BUG] Cycle mount btrfs can cause fiemap to return different result. Like: # mount /dev/vdb5 /mnt/btrfs # dd if=/dev/zero bs=16K count=4 oflag=dsync of=/mnt/btrfs/file # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 # umount /mnt/btrfs # mount /dev/vdb5 /mnt/btrfs # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..31]: 25088..25119 32 0x0 1: [32..63]: 25120..25151 32 0x0 2: [64..95]: 25152..25183 32 0x0 3: [96..127]: 25184..25215 32 0x1 But after above fiemap, we get correct merged result if we call fiemap again. # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 [REASON] Btrfs will try to merge extent map when inserting new extent map. btrfs_fiemap(start=0 len=(u64)-1) |- extent_fiemap(start=0 len=(u64)-1) |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=0 len=64k) | | Found on-disk (ino, EXTENT_DATA, 0) | |- add_extent_mapping() | |- Return (em->start=0, len=16k) | |- fiemap_fill_next_extent(logic=0 phys=X len=16k) | |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=16k len=48k) | | Found on-disk (ino, EXTENT_DATA, 16k) | |- add_extent_mapping() | | |- try_merge_map() | | Merge with previous em start=0 len=16k | | resulting em start=0 len=32k | |- Return (em->start=0, len=32K) << Merged result |- Stripe off the unrelated range (0~16K) of return em |- fiemap_fill_next_extent(logic=16K phys=X+16K len=16K) ^^^ Causing split fiemap extent. And since in add_extent_mapping(), em is already merged, in next fiemap() call, we will get merged result. [FIX] Here we introduce a new structure, fiemap_cache, which records previous fiemap extent. And will always try to merge current fiemap_cache result before calling fiemap_fill_next_extent(). Only when we failed to merge current fiemap extent with cached one, we will call fiemap_fill_next_extent() to submit cached one. So by this method, we can merge all fiemap extents. It can also be done in fs/ioctl.c, however the problem is if fieinfo->fi_extents_max == 0, we have no space to cache previous fiemap extent. So I choose to merge it in btrfs. Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2017-04-07 10:43:15 +08:00
*/
if (cache->offset + cache->len == offset &&
cache->phys + cache->len == phys &&
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
cache->flags == flags) {
btrfs: fiemap: Cache and merge fiemap extent before submit it to user [BUG] Cycle mount btrfs can cause fiemap to return different result. Like: # mount /dev/vdb5 /mnt/btrfs # dd if=/dev/zero bs=16K count=4 oflag=dsync of=/mnt/btrfs/file # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 # umount /mnt/btrfs # mount /dev/vdb5 /mnt/btrfs # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..31]: 25088..25119 32 0x0 1: [32..63]: 25120..25151 32 0x0 2: [64..95]: 25152..25183 32 0x0 3: [96..127]: 25184..25215 32 0x1 But after above fiemap, we get correct merged result if we call fiemap again. # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 [REASON] Btrfs will try to merge extent map when inserting new extent map. btrfs_fiemap(start=0 len=(u64)-1) |- extent_fiemap(start=0 len=(u64)-1) |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=0 len=64k) | | Found on-disk (ino, EXTENT_DATA, 0) | |- add_extent_mapping() | |- Return (em->start=0, len=16k) | |- fiemap_fill_next_extent(logic=0 phys=X len=16k) | |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=16k len=48k) | | Found on-disk (ino, EXTENT_DATA, 16k) | |- add_extent_mapping() | | |- try_merge_map() | | Merge with previous em start=0 len=16k | | resulting em start=0 len=32k | |- Return (em->start=0, len=32K) << Merged result |- Stripe off the unrelated range (0~16K) of return em |- fiemap_fill_next_extent(logic=16K phys=X+16K len=16K) ^^^ Causing split fiemap extent. And since in add_extent_mapping(), em is already merged, in next fiemap() call, we will get merged result. [FIX] Here we introduce a new structure, fiemap_cache, which records previous fiemap extent. And will always try to merge current fiemap_cache result before calling fiemap_fill_next_extent(). Only when we failed to merge current fiemap extent with cached one, we will call fiemap_fill_next_extent() to submit cached one. So by this method, we can merge all fiemap extents. It can also be done in fs/ioctl.c, however the problem is if fieinfo->fi_extents_max == 0, we have no space to cache previous fiemap extent. So I choose to merge it in btrfs. Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2017-04-07 10:43:15 +08:00
cache->len += len;
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
return 0;
btrfs: fiemap: Cache and merge fiemap extent before submit it to user [BUG] Cycle mount btrfs can cause fiemap to return different result. Like: # mount /dev/vdb5 /mnt/btrfs # dd if=/dev/zero bs=16K count=4 oflag=dsync of=/mnt/btrfs/file # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 # umount /mnt/btrfs # mount /dev/vdb5 /mnt/btrfs # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..31]: 25088..25119 32 0x0 1: [32..63]: 25120..25151 32 0x0 2: [64..95]: 25152..25183 32 0x0 3: [96..127]: 25184..25215 32 0x1 But after above fiemap, we get correct merged result if we call fiemap again. # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 [REASON] Btrfs will try to merge extent map when inserting new extent map. btrfs_fiemap(start=0 len=(u64)-1) |- extent_fiemap(start=0 len=(u64)-1) |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=0 len=64k) | | Found on-disk (ino, EXTENT_DATA, 0) | |- add_extent_mapping() | |- Return (em->start=0, len=16k) | |- fiemap_fill_next_extent(logic=0 phys=X len=16k) | |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=16k len=48k) | | Found on-disk (ino, EXTENT_DATA, 16k) | |- add_extent_mapping() | | |- try_merge_map() | | Merge with previous em start=0 len=16k | | resulting em start=0 len=32k | |- Return (em->start=0, len=32K) << Merged result |- Stripe off the unrelated range (0~16K) of return em |- fiemap_fill_next_extent(logic=16K phys=X+16K len=16K) ^^^ Causing split fiemap extent. And since in add_extent_mapping(), em is already merged, in next fiemap() call, we will get merged result. [FIX] Here we introduce a new structure, fiemap_cache, which records previous fiemap extent. And will always try to merge current fiemap_cache result before calling fiemap_fill_next_extent(). Only when we failed to merge current fiemap extent with cached one, we will call fiemap_fill_next_extent() to submit cached one. So by this method, we can merge all fiemap extents. It can also be done in fs/ioctl.c, however the problem is if fieinfo->fi_extents_max == 0, we have no space to cache previous fiemap extent. So I choose to merge it in btrfs. Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2017-04-07 10:43:15 +08:00
}
/* Not mergeable, need to submit cached one */
ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
cache->len, cache->flags);
cache->cached = false;
if (ret)
return ret;
assign:
cache->cached = true;
cache->offset = offset;
cache->phys = phys;
cache->len = len;
cache->flags = flags;
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
return 0;
btrfs: fiemap: Cache and merge fiemap extent before submit it to user [BUG] Cycle mount btrfs can cause fiemap to return different result. Like: # mount /dev/vdb5 /mnt/btrfs # dd if=/dev/zero bs=16K count=4 oflag=dsync of=/mnt/btrfs/file # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 # umount /mnt/btrfs # mount /dev/vdb5 /mnt/btrfs # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..31]: 25088..25119 32 0x0 1: [32..63]: 25120..25151 32 0x0 2: [64..95]: 25152..25183 32 0x0 3: [96..127]: 25184..25215 32 0x1 But after above fiemap, we get correct merged result if we call fiemap again. # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 [REASON] Btrfs will try to merge extent map when inserting new extent map. btrfs_fiemap(start=0 len=(u64)-1) |- extent_fiemap(start=0 len=(u64)-1) |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=0 len=64k) | | Found on-disk (ino, EXTENT_DATA, 0) | |- add_extent_mapping() | |- Return (em->start=0, len=16k) | |- fiemap_fill_next_extent(logic=0 phys=X len=16k) | |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=16k len=48k) | | Found on-disk (ino, EXTENT_DATA, 16k) | |- add_extent_mapping() | | |- try_merge_map() | | Merge with previous em start=0 len=16k | | resulting em start=0 len=32k | |- Return (em->start=0, len=32K) << Merged result |- Stripe off the unrelated range (0~16K) of return em |- fiemap_fill_next_extent(logic=16K phys=X+16K len=16K) ^^^ Causing split fiemap extent. And since in add_extent_mapping(), em is already merged, in next fiemap() call, we will get merged result. [FIX] Here we introduce a new structure, fiemap_cache, which records previous fiemap extent. And will always try to merge current fiemap_cache result before calling fiemap_fill_next_extent(). Only when we failed to merge current fiemap extent with cached one, we will call fiemap_fill_next_extent() to submit cached one. So by this method, we can merge all fiemap extents. It can also be done in fs/ioctl.c, however the problem is if fieinfo->fi_extents_max == 0, we have no space to cache previous fiemap extent. So I choose to merge it in btrfs. Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2017-04-07 10:43:15 +08:00
}
/*
* Emit last fiemap cache
btrfs: fiemap: Cache and merge fiemap extent before submit it to user [BUG] Cycle mount btrfs can cause fiemap to return different result. Like: # mount /dev/vdb5 /mnt/btrfs # dd if=/dev/zero bs=16K count=4 oflag=dsync of=/mnt/btrfs/file # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 # umount /mnt/btrfs # mount /dev/vdb5 /mnt/btrfs # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..31]: 25088..25119 32 0x0 1: [32..63]: 25120..25151 32 0x0 2: [64..95]: 25152..25183 32 0x0 3: [96..127]: 25184..25215 32 0x1 But after above fiemap, we get correct merged result if we call fiemap again. # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 [REASON] Btrfs will try to merge extent map when inserting new extent map. btrfs_fiemap(start=0 len=(u64)-1) |- extent_fiemap(start=0 len=(u64)-1) |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=0 len=64k) | | Found on-disk (ino, EXTENT_DATA, 0) | |- add_extent_mapping() | |- Return (em->start=0, len=16k) | |- fiemap_fill_next_extent(logic=0 phys=X len=16k) | |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=16k len=48k) | | Found on-disk (ino, EXTENT_DATA, 16k) | |- add_extent_mapping() | | |- try_merge_map() | | Merge with previous em start=0 len=16k | | resulting em start=0 len=32k | |- Return (em->start=0, len=32K) << Merged result |- Stripe off the unrelated range (0~16K) of return em |- fiemap_fill_next_extent(logic=16K phys=X+16K len=16K) ^^^ Causing split fiemap extent. And since in add_extent_mapping(), em is already merged, in next fiemap() call, we will get merged result. [FIX] Here we introduce a new structure, fiemap_cache, which records previous fiemap extent. And will always try to merge current fiemap_cache result before calling fiemap_fill_next_extent(). Only when we failed to merge current fiemap extent with cached one, we will call fiemap_fill_next_extent() to submit cached one. So by this method, we can merge all fiemap extents. It can also be done in fs/ioctl.c, however the problem is if fieinfo->fi_extents_max == 0, we have no space to cache previous fiemap extent. So I choose to merge it in btrfs. Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2017-04-07 10:43:15 +08:00
*
* The last fiemap cache may still be cached in the following case:
* 0 4k 8k
* |<- Fiemap range ->|
* |<------------ First extent ----------->|
*
* In this case, the first extent range will be cached but not emitted.
* So we must emit it before ending extent_fiemap().
btrfs: fiemap: Cache and merge fiemap extent before submit it to user [BUG] Cycle mount btrfs can cause fiemap to return different result. Like: # mount /dev/vdb5 /mnt/btrfs # dd if=/dev/zero bs=16K count=4 oflag=dsync of=/mnt/btrfs/file # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 # umount /mnt/btrfs # mount /dev/vdb5 /mnt/btrfs # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..31]: 25088..25119 32 0x0 1: [32..63]: 25120..25151 32 0x0 2: [64..95]: 25152..25183 32 0x0 3: [96..127]: 25184..25215 32 0x1 But after above fiemap, we get correct merged result if we call fiemap again. # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 [REASON] Btrfs will try to merge extent map when inserting new extent map. btrfs_fiemap(start=0 len=(u64)-1) |- extent_fiemap(start=0 len=(u64)-1) |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=0 len=64k) | | Found on-disk (ino, EXTENT_DATA, 0) | |- add_extent_mapping() | |- Return (em->start=0, len=16k) | |- fiemap_fill_next_extent(logic=0 phys=X len=16k) | |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=16k len=48k) | | Found on-disk (ino, EXTENT_DATA, 16k) | |- add_extent_mapping() | | |- try_merge_map() | | Merge with previous em start=0 len=16k | | resulting em start=0 len=32k | |- Return (em->start=0, len=32K) << Merged result |- Stripe off the unrelated range (0~16K) of return em |- fiemap_fill_next_extent(logic=16K phys=X+16K len=16K) ^^^ Causing split fiemap extent. And since in add_extent_mapping(), em is already merged, in next fiemap() call, we will get merged result. [FIX] Here we introduce a new structure, fiemap_cache, which records previous fiemap extent. And will always try to merge current fiemap_cache result before calling fiemap_fill_next_extent(). Only when we failed to merge current fiemap extent with cached one, we will call fiemap_fill_next_extent() to submit cached one. So by this method, we can merge all fiemap extents. It can also be done in fs/ioctl.c, however the problem is if fieinfo->fi_extents_max == 0, we have no space to cache previous fiemap extent. So I choose to merge it in btrfs. Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2017-04-07 10:43:15 +08:00
*/
static int emit_last_fiemap_cache(struct fiemap_extent_info *fieinfo,
struct fiemap_cache *cache)
btrfs: fiemap: Cache and merge fiemap extent before submit it to user [BUG] Cycle mount btrfs can cause fiemap to return different result. Like: # mount /dev/vdb5 /mnt/btrfs # dd if=/dev/zero bs=16K count=4 oflag=dsync of=/mnt/btrfs/file # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 # umount /mnt/btrfs # mount /dev/vdb5 /mnt/btrfs # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..31]: 25088..25119 32 0x0 1: [32..63]: 25120..25151 32 0x0 2: [64..95]: 25152..25183 32 0x0 3: [96..127]: 25184..25215 32 0x1 But after above fiemap, we get correct merged result if we call fiemap again. # xfs_io -c "fiemap -v" /mnt/btrfs/file /mnt/test/file: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..127]: 25088..25215 128 0x1 [REASON] Btrfs will try to merge extent map when inserting new extent map. btrfs_fiemap(start=0 len=(u64)-1) |- extent_fiemap(start=0 len=(u64)-1) |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=0 len=64k) | | Found on-disk (ino, EXTENT_DATA, 0) | |- add_extent_mapping() | |- Return (em->start=0, len=16k) | |- fiemap_fill_next_extent(logic=0 phys=X len=16k) | |- get_extent_skip_holes(start=0 len=64k) | |- btrfs_get_extent_fiemap(start=0 len=64k) | |- btrfs_get_extent(start=16k len=48k) | | Found on-disk (ino, EXTENT_DATA, 16k) | |- add_extent_mapping() | | |- try_merge_map() | | Merge with previous em start=0 len=16k | | resulting em start=0 len=32k | |- Return (em->start=0, len=32K) << Merged result |- Stripe off the unrelated range (0~16K) of return em |- fiemap_fill_next_extent(logic=16K phys=X+16K len=16K) ^^^ Causing split fiemap extent. And since in add_extent_mapping(), em is already merged, in next fiemap() call, we will get merged result. [FIX] Here we introduce a new structure, fiemap_cache, which records previous fiemap extent. And will always try to merge current fiemap_cache result before calling fiemap_fill_next_extent(). Only when we failed to merge current fiemap extent with cached one, we will call fiemap_fill_next_extent() to submit cached one. So by this method, we can merge all fiemap extents. It can also be done in fs/ioctl.c, however the problem is if fieinfo->fi_extents_max == 0, we have no space to cache previous fiemap extent. So I choose to merge it in btrfs. Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2017-04-07 10:43:15 +08:00
{
int ret;
if (!cache->cached)
return 0;
ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
cache->len, cache->flags);
cache->cached = false;
if (ret > 0)
ret = 0;
return ret;
}
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
static int fiemap_next_leaf_item(struct btrfs_inode *inode, struct btrfs_path *path)
{
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
struct extent_buffer *clone;
struct btrfs_key key;
int slot;
int ret;
path->slots[0]++;
if (path->slots[0] < btrfs_header_nritems(path->nodes[0]))
return 0;
ret = btrfs_next_leaf(inode->root, path);
if (ret != 0)
return ret;
/*
* Don't bother with cloning if there are no more file extent items for
* our inode.
*/
btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
if (key.objectid != btrfs_ino(inode) || key.type != BTRFS_EXTENT_DATA_KEY)
return 1;
/* See the comment at fiemap_search_slot() about why we clone. */
clone = btrfs_clone_extent_buffer(path->nodes[0]);
if (!clone)
return -ENOMEM;
slot = path->slots[0];
btrfs_release_path(path);
path->nodes[0] = clone;
path->slots[0] = slot;
return 0;
}
/*
* Search for the first file extent item that starts at a given file offset or
* the one that starts immediately before that offset.
* Returns: 0 on success, < 0 on error, 1 if not found.
*/
static int fiemap_search_slot(struct btrfs_inode *inode, struct btrfs_path *path,
u64 file_offset)
{
const u64 ino = btrfs_ino(inode);
struct btrfs_root *root = inode->root;
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
struct extent_buffer *clone;
struct btrfs_key key;
int slot;
int ret;
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
key.objectid = ino;
key.type = BTRFS_EXTENT_DATA_KEY;
key.offset = file_offset;
ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
if (ret < 0)
return ret;
if (ret > 0 && path->slots[0] > 0) {
btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
path->slots[0]--;
}
if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
ret = btrfs_next_leaf(root, path);
if (ret != 0)
return ret;
btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
return 1;
}
btrfs: return whole extents in fiemap `xfs_io -c 'fiemap <off> <len>' <file>` can give surprising results on btrfs that differ from xfs. btrfs prints out extents trimmed to fit the user input. If the user's fiemap request has an offset, then rather than returning each whole extent which intersects that range, we also trim the start extent to not have start < off. Documentation in filesystems/fiemap.txt and the xfs_io man page suggests that returning the whole extent is expected. Some cases which all yield the same fiemap in xfs, but not btrfs: dd if=/dev/zero of=$f bs=4k count=1 sudo xfs_io -c 'fiemap 0 1024' $f 0: [0..7]: 26624..26631 sudo xfs_io -c 'fiemap 2048 1024' $f 0: [4..7]: 26628..26631 sudo xfs_io -c 'fiemap 2048 4096' $f 0: [4..7]: 26628..26631 sudo xfs_io -c 'fiemap 3584 512' $f 0: [7..7]: 26631..26631 sudo xfs_io -c 'fiemap 4091 5' $f 0: [7..6]: 26631..26630 I believe this is a consequence of the logic for merging contiguous extents represented by separate extent items. That logic needs to track the last offset as it loops through the extent items, which happens to pick up the start offset on the first iteration, and trim off the beginning of the full extent. To fix it, start `off` at 0 rather than `start` so that we keep the iteration/merging intact without cutting off the start of the extent. after the fix, all the above commands give: 0: [0..7]: 26624..26631 The merging logic is exercised by fstest generic/483, and I have written a new fstest for checking we don't have backwards or zero-length fiemaps for cases like those above. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-07 06:31:18 +08:00
/*
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
* We clone the leaf and use it during fiemap. This is because while
* using the leaf we do expensive things like checking if an extent is
* shared, which can take a long time. In order to prevent blocking
* other tasks for too long, we use a clone of the leaf. We have locked
* the file range in the inode's io tree, so we know none of our file
* extent items can change. This way we avoid blocking other tasks that
* want to insert items for other inodes in the same leaf or b+tree
* rebalance operations (triggered for example when someone is trying
* to push items into this leaf when trying to insert an item in a
* neighbour leaf).
* We also need the private clone because holding a read lock on an
* extent buffer of the subvolume's b+tree will make lockdep unhappy
* when we call fiemap_fill_next_extent(), because that may cause a page
* fault when filling the user space buffer with fiemap data.
btrfs: return whole extents in fiemap `xfs_io -c 'fiemap <off> <len>' <file>` can give surprising results on btrfs that differ from xfs. btrfs prints out extents trimmed to fit the user input. If the user's fiemap request has an offset, then rather than returning each whole extent which intersects that range, we also trim the start extent to not have start < off. Documentation in filesystems/fiemap.txt and the xfs_io man page suggests that returning the whole extent is expected. Some cases which all yield the same fiemap in xfs, but not btrfs: dd if=/dev/zero of=$f bs=4k count=1 sudo xfs_io -c 'fiemap 0 1024' $f 0: [0..7]: 26624..26631 sudo xfs_io -c 'fiemap 2048 1024' $f 0: [4..7]: 26628..26631 sudo xfs_io -c 'fiemap 2048 4096' $f 0: [4..7]: 26628..26631 sudo xfs_io -c 'fiemap 3584 512' $f 0: [7..7]: 26631..26631 sudo xfs_io -c 'fiemap 4091 5' $f 0: [7..6]: 26631..26630 I believe this is a consequence of the logic for merging contiguous extents represented by separate extent items. That logic needs to track the last offset as it loops through the extent items, which happens to pick up the start offset on the first iteration, and trim off the beginning of the full extent. To fix it, start `off` at 0 rather than `start` so that we keep the iteration/merging intact without cutting off the start of the extent. after the fix, all the above commands give: 0: [0..7]: 26624..26631 The merging logic is exercised by fstest generic/483, and I have written a new fstest for checking we don't have backwards or zero-length fiemaps for cases like those above. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-07 06:31:18 +08:00
*/
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
clone = btrfs_clone_extent_buffer(path->nodes[0]);
if (!clone)
return -ENOMEM;
slot = path->slots[0];
btrfs_release_path(path);
path->nodes[0] = clone;
path->slots[0] = slot;
return 0;
}
/*
* Process a range which is a hole or a prealloc extent in the inode's subvolume
* btree. If @disk_bytenr is 0, we are dealing with a hole, otherwise a prealloc
* extent. The end offset (@end) is inclusive.
*/
static int fiemap_process_hole(struct btrfs_inode *inode,
struct fiemap_extent_info *fieinfo,
struct fiemap_cache *cache,
struct btrfs_backref_share_check_ctx *backref_ctx,
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
u64 disk_bytenr, u64 extent_offset,
u64 extent_gen,
u64 start, u64 end)
{
const u64 i_size = i_size_read(&inode->vfs_inode);
u64 cur_offset = start;
u64 last_delalloc_end = 0;
u32 prealloc_flags = FIEMAP_EXTENT_UNWRITTEN;
bool checked_extent_shared = false;
int ret;
/*
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
* There can be no delalloc past i_size, so don't waste time looking for
* it beyond i_size.
*/
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
while (cur_offset < end && cur_offset < i_size) {
u64 delalloc_start;
u64 delalloc_end;
u64 prealloc_start;
u64 prealloc_len = 0;
bool delalloc;
delalloc = btrfs_find_delalloc_in_range(inode, cur_offset, end,
&delalloc_start,
&delalloc_end);
if (!delalloc)
break;
/*
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
* If this is a prealloc extent we have to report every section
* of it that has no delalloc.
*/
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
if (disk_bytenr != 0) {
if (last_delalloc_end == 0) {
prealloc_start = start;
prealloc_len = delalloc_start - start;
} else {
prealloc_start = last_delalloc_end + 1;
prealloc_len = delalloc_start - prealloc_start;
}
}
if (prealloc_len > 0) {
if (!checked_extent_shared && fieinfo->fi_extents_max) {
ret = btrfs_is_data_extent_shared(inode,
disk_bytenr,
extent_gen,
backref_ctx);
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
if (ret < 0)
return ret;
else if (ret > 0)
prealloc_flags |= FIEMAP_EXTENT_SHARED;
checked_extent_shared = true;
}
ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
disk_bytenr + extent_offset,
prealloc_len, prealloc_flags);
if (ret)
return ret;
extent_offset += prealloc_len;
}
ret = emit_fiemap_extent(fieinfo, cache, delalloc_start, 0,
delalloc_end + 1 - delalloc_start,
FIEMAP_EXTENT_DELALLOC |
FIEMAP_EXTENT_UNKNOWN);
if (ret)
return ret;
last_delalloc_end = delalloc_end;
cur_offset = delalloc_end + 1;
extent_offset += cur_offset - delalloc_start;
cond_resched();
}
/*
* Either we found no delalloc for the whole prealloc extent or we have
* a prealloc extent that spans i_size or starts at or after i_size.
*/
if (disk_bytenr != 0 && last_delalloc_end < end) {
u64 prealloc_start;
u64 prealloc_len;
if (last_delalloc_end == 0) {
prealloc_start = start;
prealloc_len = end + 1 - start;
} else {
prealloc_start = last_delalloc_end + 1;
prealloc_len = end + 1 - prealloc_start;
}
if (!checked_extent_shared && fieinfo->fi_extents_max) {
ret = btrfs_is_data_extent_shared(inode,
disk_bytenr,
extent_gen,
backref_ctx);
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
if (ret < 0)
return ret;
else if (ret > 0)
prealloc_flags |= FIEMAP_EXTENT_SHARED;
}
ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
disk_bytenr + extent_offset,
prealloc_len, prealloc_flags);
if (ret)
return ret;
}
return 0;
}
static int fiemap_find_last_extent_offset(struct btrfs_inode *inode,
struct btrfs_path *path,
u64 *last_extent_end_ret)
{
const u64 ino = btrfs_ino(inode);
struct btrfs_root *root = inode->root;
struct extent_buffer *leaf;
struct btrfs_file_extent_item *ei;
struct btrfs_key key;
u64 disk_bytenr;
int ret;
/*
* Lookup the last file extent. We're not using i_size here because
* there might be preallocation past i_size.
*/
ret = btrfs_lookup_file_extent(NULL, root, path, ino, (u64)-1, 0);
/* There can't be a file extent item at offset (u64)-1 */
ASSERT(ret != 0);
if (ret < 0)
return ret;
/*
* For a non-existing key, btrfs_search_slot() always leaves us at a
* slot > 0, except if the btree is empty, which is impossible because
* at least it has the inode item for this inode and all the items for
* the root inode 256.
*/
ASSERT(path->slots[0] > 0);
path->slots[0]--;
leaf = path->nodes[0];
btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY) {
/* No file extent items in the subvolume tree. */
*last_extent_end_ret = 0;
return 0;
}
/*
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
* For an inline extent, the disk_bytenr is where inline data starts at,
* so first check if we have an inline extent item before checking if we
* have an implicit hole (disk_bytenr == 0).
*/
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_INLINE) {
*last_extent_end_ret = btrfs_file_extent_end(path);
return 0;
}
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
/*
* Find the last file extent item that is not a hole (when NO_HOLES is
* not enabled). This should take at most 2 iterations in the worst
* case: we have one hole file extent item at slot 0 of a leaf and
* another hole file extent item as the last item in the previous leaf.
* This is because we merge file extent items that represent holes.
*/
disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
while (disk_bytenr == 0) {
ret = btrfs_previous_item(root, path, ino, BTRFS_EXTENT_DATA_KEY);
if (ret < 0) {
return ret;
} else if (ret > 0) {
/* No file extent items that are not holes. */
*last_extent_end_ret = 0;
return 0;
}
leaf = path->nodes[0];
ei = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_file_extent_item);
disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
}
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
*last_extent_end_ret = btrfs_file_extent_end(path);
return 0;
}
int extent_fiemap(struct btrfs_inode *inode, struct fiemap_extent_info *fieinfo,
u64 start, u64 len)
{
const u64 ino = btrfs_ino(inode);
struct extent_state *cached_state = NULL;
struct btrfs_path *path;
struct fiemap_cache cache = { 0 };
struct btrfs_backref_share_check_ctx *backref_ctx;
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
u64 last_extent_end;
u64 prev_extent_end;
u64 lockstart;
u64 lockend;
bool stopped = false;
int ret;
backref_ctx = btrfs_alloc_backref_share_check_ctx();
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
path = btrfs_alloc_path();
if (!backref_ctx || !path) {
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
ret = -ENOMEM;
goto out;
}
lockstart = round_down(start, inode->root->fs_info->sectorsize);
lockend = round_up(start + len, inode->root->fs_info->sectorsize);
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
prev_extent_end = lockstart;
lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
ret = fiemap_find_last_extent_offset(inode, path, &last_extent_end);
if (ret < 0)
goto out_unlock;
btrfs_release_path(path);
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
path->reada = READA_FORWARD;
ret = fiemap_search_slot(inode, path, lockstart);
if (ret < 0) {
goto out_unlock;
} else if (ret > 0) {
/*
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
* No file extent item found, but we may have delalloc between
* the current offset and i_size. So check for that.
*/
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
ret = 0;
goto check_eof_delalloc;
}
while (prev_extent_end < lockend) {
struct extent_buffer *leaf = path->nodes[0];
struct btrfs_file_extent_item *ei;
struct btrfs_key key;
u64 extent_end;
u64 extent_len;
u64 extent_offset = 0;
u64 extent_gen;
u64 disk_bytenr = 0;
u64 flags = 0;
int extent_type;
u8 compression;
btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
break;
extent_end = btrfs_file_extent_end(path);
/*
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
* The first iteration can leave us at an extent item that ends
* before our range's start. Move to the next item.
*/
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
if (extent_end <= lockstart)
goto next_item;
btrfs: avoid duplicated resolution of indirect backrefs during fiemap During fiemap, when determining if a data extent is shared or not, if we don't find the extent is directly shared, then we need to determine if it's shared through subtrees. For that we need to resolve the indirect reference we found in order to figure out the path in the inode's fs tree, which is a path starting at the fs tree's root node and going down to the leaf that contains the file extent item that points to the data extent. We then proceed to determine if any extent buffer in that path is shared with other trees or not. Currently whenever we find the data extent that a file extent item points to is not directly shared, we always resolve the path in the fs tree, and then check if any extent buffer in the path is shared. This is a lot of work and when we have file extent items that belong to the same leaf, we have the same path, so we only need to calculate it once. This change does that, it keeps track of the current and previous leaf, and when we find that a data extent is not directly shared, we try to compute the fs tree path only once and then use it for every other file extent item in the same leaf, using the existing cached path result for the leaf as long as the cache results are valid. This saves us from doing expensive b+tree searches in the fs tree of our target inode, as well as other minor work. The following test was run on a non-debug kernel (Debian's default kernel config): $ cat test-with-snapshots.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi umount $DEV &> /dev/null mkfs.btrfs -f $DEV # Use compression to quickly create files with a lot of extents # (each with a size of 128K). mount -o compress=lzo $DEV $MNT # 40G gives 327680 extents, each with a size of 128K. xfs_io -f -c "pwrite -S 0xab -b 1M 0 40G" $MNT/foobar # Add some more files to increase the size of the fs and extent # trees (in the real world there's a lot of files and extents # from other files). xfs_io -f -c "pwrite -S 0xcd -b 1M 0 20G" $MNT/file1 xfs_io -f -c "pwrite -S 0xef -b 1M 0 20G" $MNT/file2 xfs_io -f -c "pwrite -S 0x73 -b 1M 0 20G" $MNT/file3 # Create a snapshot so all the extents become indirectly shared # through subtrees, with a generation less than or equals to the # generation used to create the snapshot. btrfs subvolume snapshot -r $MNT $MNT/snap1 umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" echo start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Result before applying this patch: (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1204 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 729 milliseconds (metadata cached) Result after applying this patch: (...) /mnt/sdi/foobar: 327680 extents found fiemap took 732 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 421 milliseconds (metadata cached) That's a -46.1% total reduction for the metadata not cached case, and a -42.2% reduction for the cached metadata case. The test is somewhat limited in the sense the gains may be higher in practice, because in the test the filesystem is small, so we have small fs and extent trees, plus there's no concurrent access to the trees as well, therefore no lock contention there. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-10-11 20:17:08 +08:00
backref_ctx->curr_leaf_bytenr = leaf->start;
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
/* We have in implicit hole (NO_HOLES feature enabled). */
if (prev_extent_end < key.offset) {
const u64 range_end = min(key.offset, lockend) - 1;
btrfs: skip unnecessary extent buffer sharedness checks during fiemap During fiemap, for each file extent we find, we must check if it's shared or not. The sharedness check starts by verifying if the extent is directly shared (its refcount in the extent tree is > 1), and if it is not directly shared, then we will check if every node in the subvolume b+tree leading from the root to the leaf that has the file extent item (in reverse order), is shared (through snapshots). However this second step is not needed if our extent was created in a transaction more recent than the last transaction where a snapshot of the inode's root happened, because it can't be shared indirectly (through shared subtrees) without a snapshot created in a more recent transaction. So grab the generation of the extent from the extent map and pass it to btrfs_is_data_extent_shared(), which will skip this second phase when the generation is more recent than the root's last snapshot value. Note that we skip this optimization if the extent map is the result of merging 2 or more extent maps, because in this case its generation is the maximum of the generations of all merged extent maps. The fact the we use extent maps and they can be merged despite the underlying extents being distinct (different file extent items in the subvolume b+tree and different extent items in the extent b+tree), can result in some bugs when reporting shared extents. But this is a problem of the current implementation of fiemap relying on extent maps. One example where we get incorrect results is: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This is z problem that existed before this change, and remains after this change, as it can't be easily fixed. The next patch in the series reworks fiemap to primarily use file extent items instead of extent maps (except for checking for delalloc ranges), with the goal of improving its scalability and performance, but it also ends up fixing this particular bug caused by extent map merging. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: Qu Wenruo <wqu@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:29 +08:00
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
ret = fiemap_process_hole(inode, fieinfo, &cache,
backref_ctx, 0, 0, 0,
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
prev_extent_end, range_end);
if (ret < 0) {
goto out_unlock;
} else if (ret > 0) {
/* fiemap_fill_next_extent() told us to stop. */
stopped = true;
break;
}
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
/* We've reached the end of the fiemap range, stop. */
if (key.offset >= lockend) {
stopped = true;
break;
}
}
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
extent_len = extent_end - key.offset;
ei = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_file_extent_item);
compression = btrfs_file_extent_compression(leaf, ei);
extent_type = btrfs_file_extent_type(leaf, ei);
extent_gen = btrfs_file_extent_generation(leaf, ei);
if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
if (compression == BTRFS_COMPRESS_NONE)
extent_offset = btrfs_file_extent_offset(leaf, ei);
}
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
if (compression != BTRFS_COMPRESS_NONE)
flags |= FIEMAP_EXTENT_ENCODED;
if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
flags |= FIEMAP_EXTENT_DATA_INLINE;
flags |= FIEMAP_EXTENT_NOT_ALIGNED;
ret = emit_fiemap_extent(fieinfo, &cache, key.offset, 0,
extent_len, flags);
} else if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
ret = fiemap_process_hole(inode, fieinfo, &cache,
backref_ctx,
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
disk_bytenr, extent_offset,
extent_gen, key.offset,
extent_end - 1);
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
} else if (disk_bytenr == 0) {
/* We have an explicit hole. */
ret = fiemap_process_hole(inode, fieinfo, &cache,
backref_ctx, 0, 0, 0,
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
key.offset, extent_end - 1);
} else {
/* We have a regular extent. */
if (fieinfo->fi_extents_max) {
ret = btrfs_is_data_extent_shared(inode,
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
disk_bytenr,
extent_gen,
backref_ctx);
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
if (ret < 0)
goto out_unlock;
else if (ret > 0)
flags |= FIEMAP_EXTENT_SHARED;
}
ret = emit_fiemap_extent(fieinfo, &cache, key.offset,
disk_bytenr + extent_offset,
extent_len, flags);
}
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
if (ret < 0) {
goto out_unlock;
} else if (ret > 0) {
/* fiemap_fill_next_extent() told us to stop. */
stopped = true;
break;
}
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
prev_extent_end = extent_end;
next_item:
if (fatal_signal_pending(current)) {
ret = -EINTR;
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
goto out_unlock;
}
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
ret = fiemap_next_leaf_item(inode, path);
if (ret < 0) {
goto out_unlock;
} else if (ret > 0) {
/* No more file extent items for this inode. */
break;
}
cond_resched();
}
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
check_eof_delalloc:
/*
* Release (and free) the path before emitting any final entries to
* fiemap_fill_next_extent() to keep lockdep happy. This is because
* once we find no more file extent items exist, we may have a
* non-cloned leaf, and fiemap_fill_next_extent() can trigger page
* faults when copying data to the user space buffer.
*/
btrfs_free_path(path);
path = NULL;
if (!stopped && prev_extent_end < lockend) {
ret = fiemap_process_hole(inode, fieinfo, &cache, backref_ctx,
0, 0, 0, prev_extent_end, lockend - 1);
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
if (ret < 0)
goto out_unlock;
prev_extent_end = lockend;
}
if (cache.cached && cache.offset + cache.len >= last_extent_end) {
const u64 i_size = i_size_read(&inode->vfs_inode);
if (prev_extent_end < i_size) {
u64 delalloc_start;
u64 delalloc_end;
bool delalloc;
delalloc = btrfs_find_delalloc_in_range(inode,
prev_extent_end,
i_size - 1,
&delalloc_start,
&delalloc_end);
if (!delalloc)
cache.flags |= FIEMAP_EXTENT_LAST;
} else {
cache.flags |= FIEMAP_EXTENT_LAST;
}
}
ret = emit_last_fiemap_cache(fieinfo, &cache);
out_unlock:
unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
btrfs: make fiemap more efficient and accurate reporting extent sharedness The current fiemap implementation does not scale very well with the number of extents a file has. This is both because the main algorithm to find out the extents has a high algorithmic complexity and because for each extent we have to check if it's shared. This second part, checking if an extent is shared, is significantly improved by the two previous patches in this patchset, while the first part is improved by this specific patch. Every now and then we get reports from users mentioning fiemap is too slow or even unusable for files with a very large number of extents, such as the two recent reports referred to by the Link tags at the bottom of this change log. To understand why the part of finding which extents a file has is very inefficient, consider the example of doing a full ranged fiemap against a file that has over 100K extents (normal for example for a file with more than 10G of data and using compression, which limits the extent size to 128K). When we enter fiemap at extent_fiemap(), the following happens: 1) Before entering the main loop, we call get_extent_skip_holes() to get the first extent map. This leads us to btrfs_get_extent_fiemap(), which in turn calls btrfs_get_extent(), to find the first extent map that covers the file range [0, LLONG_MAX). btrfs_get_extent() will first search the inode's extent map tree, to see if we have an extent map there that covers the range. If it does not find one, then it will search the inode's subvolume b+tree for a fitting file extent item. After finding the file extent item, it will allocate an extent map, fill it in with information extracted from the file extent item, and add it to the inode's extent map tree (which requires a search for insertion in the tree). 2) Then we enter the main loop at extent_fiemap(), emit the details of the extent, and call again get_extent_skip_holes(), with a start offset matching the end of the extent map we previously processed. We end up at btrfs_get_extent() again, will search the extent map tree and then search the subvolume b+tree for a file extent item if we could not find an extent map in the extent tree. We allocate an extent map, fill it in with the details in the file extent item, and then insert it into the extent map tree (yet another search in this tree). 3) The second step is repeated over and over, until we have processed the whole file range. Each iteration ends at btrfs_get_extent(), which does a red black tree search on the extent map tree, then searches the subvolume b+tree, allocates an extent map and then does another search in the extent map tree in order to insert the extent map. In the best scenario we have all the extent maps already in the extent tree, and so for each extent we do a single search on a red black tree, so we have a complexity of O(n log n). In the worst scenario we don't have any extent map already loaded in the extent map tree, or have very few already there. In this case the complexity is much higher since we do: - A red black tree search on the extent map tree, which has O(log n) complexity, initially very fast since the tree is empty or very small, but as we end up allocating extent maps and adding them to the tree when we don't find them there, each subsequent search on the tree gets slower, since it's getting bigger and bigger after each iteration. - A search on the subvolume b+tree, also O(log n) complexity, but it has items for all inodes in the subvolume, not just items for our inode. Plus on a filesystem with concurrent operations on other inodes, we can block doing the search due to lock contention on b+tree nodes/leaves. - Allocate an extent map - this can block, and can also fail if we are under serious memory pressure. - Do another search on the extent maps red black tree, with the goal of inserting the extent map we just allocated. Again, after every iteration this tree is getting bigger by 1 element, so after many iterations the searches are slower and slower. - We will not need the allocated extent map anymore, so it's pointless to add it to the extent map tree. It's just wasting time and memory. In short we end up searching the extent map tree multiple times, on a tree that is growing bigger and bigger after each iteration. And besides that we visit the same leaf of the subvolume b+tree many times, since a leaf with the default size of 16K can easily have more than 200 file extent items. This is very inefficient overall. This patch changes the algorithm to instead iterate over the subvolume b+tree, visiting each leaf only once, and only searching in the extent map tree for file ranges that have holes or prealloc extents, in order to figure out if we have delalloc there. It will never allocate an extent map and add it to the extent map tree. This is very similar to what was previously done for the lseek's hole and data seeking features. Also, the current implementation relying on extent maps for figuring out which extents we have is not correct. This is because extent maps can be merged even if they represent different extents - we do this to minimize memory utilization and keep extent map trees smaller. For example if we have two extents that are contiguous on disk, once we load the two extent maps, they get merged into a single one - however if only one of the extents is shared, we end up reporting both as shared or both as not shared, which is incorrect. This reproducer triggers that bug: $ cat fiemap-bug.sh #!/bin/bash DEV=/dev/sdj MNT=/mnt/sdj mkfs.btrfs -f $DEV mount $DEV $MNT # Create a file with two 256K extents. # Since there is no other write activity, they will be contiguous, # and their extent maps merged, despite having two distinct extents. xfs_io -f -c "pwrite -S 0xab 0 256K" \ -c "fsync" \ -c "pwrite -S 0xcd 256K 256K" \ -c "fsync" \ $MNT/foo # Now clone only the second extent into another file. xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar # Filefrag will report a single 512K extent, and say it's not shared. echo filefrag -v $MNT/foo umount $MNT Running the reproducer: $ ./fiemap-bug.sh wrote 262144/262144 bytes at offset 0 256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec) wrote 262144/262144 bytes at offset 262144 256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec) linked 262144/262144 bytes at offset 0 256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec) Filesystem type is: 9123683e File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes) ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 127: 3328.. 3455: 128: last,eof /mnt/sdj/foo: 1 extent found We end up reporting that we have a single 512K that is not shared, however we have two 256K extents, and the second one is shared. Changing the reproducer to clone instead the first extent into file 'bar', makes us report a single 512K extent that is shared, which is algo incorrect since we have two 256K extents and only the first one is shared. This patch is part of a larger patchset that is comprised of the following patches: btrfs: allow hole and data seeking to be interruptible btrfs: make hole and data seeking a lot more efficient btrfs: remove check for impossible block start for an extent map at fiemap btrfs: remove zero length check when entering fiemap btrfs: properly flush delalloc when entering fiemap btrfs: allow fiemap to be interruptible btrfs: rename btrfs_check_shared() to a more descriptive name btrfs: speedup checking for extent sharedness during fiemap btrfs: skip unnecessary extent buffer sharedness checks during fiemap btrfs: make fiemap more efficient and accurate reporting extent sharedness The patchset was tested on a machine running a non-debug kernel (Debian's default config) and compared the tests below on a branch without the patchset versus the same branch with the whole patchset applied. The following test for a large compressed file without holes: $ cat fiemap-perf-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f $DEV mount -o compress=lzo $DEV $MNT # 40G gives 327680 128K file extents (due to compression). xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata not cached)" start=$(date +%s%N) filefrag $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "fiemap took $dur milliseconds (metadata cached)" umount $MNT Before patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 3597 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 2107 milliseconds (metadata cached) After patchset: $ ./fiemap-perf-test.sh (...) /mnt/sdi/foobar: 327680 extents found fiemap took 1214 milliseconds (metadata not cached) /mnt/sdi/foobar: 327680 extents found fiemap took 684 milliseconds (metadata cached) That's a speedup of about 3x for both cases (no metadata cached and all metadata cached). The test provided by Pavel (first Link tag at the bottom), which uses files with a large number of holes, was also used to measure the gains, and it consists on a small C program and a shell script to invoke it. The C program is the following: $ cat pavels-test.c #include <stdio.h> #include <unistd.h> #include <stdlib.h> #include <fcntl.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/ioctl.h> #include <linux/fs.h> #include <linux/fiemap.h> #define FILE_INTERVAL (1<<13) /* 8Kb */ long long interval(struct timeval t1, struct timeval t2) { long long val = 0; val += (t2.tv_usec - t1.tv_usec); val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000; return val; } int main(int argc, char **argv) { struct fiemap fiemap = {}; struct timeval t1, t2; char data = 'a'; struct stat st; int fd, off, file_size = FILE_INTERVAL; if (argc != 3 && argc != 2) { printf("usage: %s <path> [size]\n", argv[0]); return 1; } if (argc == 3) file_size = atoi(argv[2]); if (file_size < FILE_INTERVAL) file_size = FILE_INTERVAL; file_size -= file_size % FILE_INTERVAL; fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644); if (fd < 0) { perror("open"); return 1; } for (off = 0; off < file_size; off += FILE_INTERVAL) { if (pwrite(fd, &data, 1, off) != 1) { perror("pwrite"); close(fd); return 1; } } if (ftruncate(fd, file_size)) { perror("ftruncate"); close(fd); return 1; } if (fstat(fd, &st) < 0) { perror("fstat"); close(fd); return 1; } printf("size: %ld\n", st.st_size); printf("actual size: %ld\n", st.st_blocks * 512); fiemap.fm_length = FIEMAP_MAX_OFFSET; gettimeofday(&t1, NULL); if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) { perror("fiemap"); close(fd); return 1; } gettimeofday(&t2, NULL); printf("fiemap: fm_mapped_extents = %d\n", fiemap.fm_mapped_extents); printf("time = %lld us\n", interval(t1, t2)); close(fd); return 0; } $ gcc -o pavels_test pavels_test.c And the wrapper shell script: $ cat fiemap-pavels-test.sh #!/bin/bash DEV=/dev/sdi MNT=/mnt/sdi mkfs.btrfs -f -O no-holes $DEV mount $DEV $MNT echo echo "*********** 256M ***********" echo ./pavels-test $MNT/testfile $((1 << 28)) echo ./pavels-test $MNT/testfile $((1 << 28)) echo echo "*********** 512M ***********" echo ./pavels-test $MNT/testfile $((1 << 29)) echo ./pavels-test $MNT/testfile $((1 << 29)) echo echo "*********** 1G ***********" echo ./pavels-test $MNT/testfile $((1 << 30)) echo ./pavels-test $MNT/testfile $((1 << 30)) umount $MNT Running his reproducer before applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4003133 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 4895330 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 30123675 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 33450934 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 224924074 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 217239242 us Running it after applying the patchset: *********** 256M *********** size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29475 us size: 268435456 actual size: 134217728 fiemap: fm_mapped_extents = 32768 time = 29307 us *********** 512M *********** size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 58996 us size: 536870912 actual size: 268435456 fiemap: fm_mapped_extents = 65536 time = 59115 us *********** 1G *********** size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 116251 time = 124141 us size: 1073741824 actual size: 536870912 fiemap: fm_mapped_extents = 131072 time = 119387 us The speedup is massive, both on the first fiemap call and on the second one as well, as his test creates files with many holes and small extents (every extent follows a hole and precedes another hole). For the 256M file we go from 4 seconds down to 29 milliseconds in the first run, and then from 4.9 seconds down to 29 milliseconds again in the second run, a speedup of 138x and 169x, respectively. For the 512M file we go from 30.1 seconds down to 59 milliseconds in the first run, and then from 33.5 seconds down to 59 milliseconds again in the second run, a speedup of 510x and 568x, respectively. For the 1G file, we go from 225 seconds down to 124 milliseconds in the first run, and then from 217 seconds down to 119 milliseconds in the second run, a speedup of 1815x and 1824x, respectively. Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com> Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/ Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com> Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-01 21:18:30 +08:00
out:
btrfs_free_backref_share_ctx(backref_ctx);
btrfs_free_path(path);
return ret;
}
static void __free_extent_buffer(struct extent_buffer *eb)
{
kmem_cache_free(extent_buffer_cache, eb);
}
int extent_buffer_under_io(const struct extent_buffer *eb)
{
return (atomic_read(&eb->io_pages) ||
test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
}
btrfs: support subpage for extent buffer page release In btrfs_release_extent_buffer_pages(), we need to add extra handling for subpage. Introduce a helper, detach_extent_buffer_page(), to do different handling for regular and subpage cases. For subpage case, handle detaching page private. For unmapped (dummy or cloned) ebs, we can detach the page private immediately as the page can only be attached to one unmapped eb. For mapped ebs, we have to ensure there are no eb in the page range before we delete it, as page->private is shared between all ebs in the same page. But there is a subpage specific race, where we can race with extent buffer allocation, and clear the page private while new eb is still being utilized, like this: Extent buffer A is the new extent buffer which will be allocated, while extent buffer B is the last existing extent buffer of the page. T1 (eb A) | T2 (eb B) -------------------------------+------------------------------ alloc_extent_buffer() | btrfs_release_extent_buffer_pages() |- p = find_or_create_page() | | |- attach_extent_buffer_page() | | | | |- detach_extent_buffer_page() | | |- if (!page_range_has_eb()) | | | No new eb in the page range yet | | | As new eb A hasn't yet been | | | inserted into radix tree. | | |- btrfs_detach_subpage() | | |- detach_page_private(); |- radix_tree_insert() | Then we have a metadata eb whose page has no private bit. To avoid such race, we introduce a subpage metadata-specific member, btrfs_subpage::eb_refs. In alloc_extent_buffer() we increase eb_refs in the critical section of private_lock. Then page_range_has_eb() will return true for detach_extent_buffer_page(), and will not detach page private. The section is marked by: - btrfs_page_inc_eb_refs() - btrfs_page_dec_eb_refs() Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-26 16:33:50 +08:00
static bool page_range_has_eb(struct btrfs_fs_info *fs_info, struct page *page)
{
btrfs: support subpage for extent buffer page release In btrfs_release_extent_buffer_pages(), we need to add extra handling for subpage. Introduce a helper, detach_extent_buffer_page(), to do different handling for regular and subpage cases. For subpage case, handle detaching page private. For unmapped (dummy or cloned) ebs, we can detach the page private immediately as the page can only be attached to one unmapped eb. For mapped ebs, we have to ensure there are no eb in the page range before we delete it, as page->private is shared between all ebs in the same page. But there is a subpage specific race, where we can race with extent buffer allocation, and clear the page private while new eb is still being utilized, like this: Extent buffer A is the new extent buffer which will be allocated, while extent buffer B is the last existing extent buffer of the page. T1 (eb A) | T2 (eb B) -------------------------------+------------------------------ alloc_extent_buffer() | btrfs_release_extent_buffer_pages() |- p = find_or_create_page() | | |- attach_extent_buffer_page() | | | | |- detach_extent_buffer_page() | | |- if (!page_range_has_eb()) | | | No new eb in the page range yet | | | As new eb A hasn't yet been | | | inserted into radix tree. | | |- btrfs_detach_subpage() | | |- detach_page_private(); |- radix_tree_insert() | Then we have a metadata eb whose page has no private bit. To avoid such race, we introduce a subpage metadata-specific member, btrfs_subpage::eb_refs. In alloc_extent_buffer() we increase eb_refs in the critical section of private_lock. Then page_range_has_eb() will return true for detach_extent_buffer_page(), and will not detach page private. The section is marked by: - btrfs_page_inc_eb_refs() - btrfs_page_dec_eb_refs() Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-26 16:33:50 +08:00
struct btrfs_subpage *subpage;
btrfs: support subpage for extent buffer page release In btrfs_release_extent_buffer_pages(), we need to add extra handling for subpage. Introduce a helper, detach_extent_buffer_page(), to do different handling for regular and subpage cases. For subpage case, handle detaching page private. For unmapped (dummy or cloned) ebs, we can detach the page private immediately as the page can only be attached to one unmapped eb. For mapped ebs, we have to ensure there are no eb in the page range before we delete it, as page->private is shared between all ebs in the same page. But there is a subpage specific race, where we can race with extent buffer allocation, and clear the page private while new eb is still being utilized, like this: Extent buffer A is the new extent buffer which will be allocated, while extent buffer B is the last existing extent buffer of the page. T1 (eb A) | T2 (eb B) -------------------------------+------------------------------ alloc_extent_buffer() | btrfs_release_extent_buffer_pages() |- p = find_or_create_page() | | |- attach_extent_buffer_page() | | | | |- detach_extent_buffer_page() | | |- if (!page_range_has_eb()) | | | No new eb in the page range yet | | | As new eb A hasn't yet been | | | inserted into radix tree. | | |- btrfs_detach_subpage() | | |- detach_page_private(); |- radix_tree_insert() | Then we have a metadata eb whose page has no private bit. To avoid such race, we introduce a subpage metadata-specific member, btrfs_subpage::eb_refs. In alloc_extent_buffer() we increase eb_refs in the critical section of private_lock. Then page_range_has_eb() will return true for detach_extent_buffer_page(), and will not detach page private. The section is marked by: - btrfs_page_inc_eb_refs() - btrfs_page_dec_eb_refs() Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-26 16:33:50 +08:00
lockdep_assert_held(&page->mapping->private_lock);
btrfs: support subpage for extent buffer page release In btrfs_release_extent_buffer_pages(), we need to add extra handling for subpage. Introduce a helper, detach_extent_buffer_page(), to do different handling for regular and subpage cases. For subpage case, handle detaching page private. For unmapped (dummy or cloned) ebs, we can detach the page private immediately as the page can only be attached to one unmapped eb. For mapped ebs, we have to ensure there are no eb in the page range before we delete it, as page->private is shared between all ebs in the same page. But there is a subpage specific race, where we can race with extent buffer allocation, and clear the page private while new eb is still being utilized, like this: Extent buffer A is the new extent buffer which will be allocated, while extent buffer B is the last existing extent buffer of the page. T1 (eb A) | T2 (eb B) -------------------------------+------------------------------ alloc_extent_buffer() | btrfs_release_extent_buffer_pages() |- p = find_or_create_page() | | |- attach_extent_buffer_page() | | | | |- detach_extent_buffer_page() | | |- if (!page_range_has_eb()) | | | No new eb in the page range yet | | | As new eb A hasn't yet been | | | inserted into radix tree. | | |- btrfs_detach_subpage() | | |- detach_page_private(); |- radix_tree_insert() | Then we have a metadata eb whose page has no private bit. To avoid such race, we introduce a subpage metadata-specific member, btrfs_subpage::eb_refs. In alloc_extent_buffer() we increase eb_refs in the critical section of private_lock. Then page_range_has_eb() will return true for detach_extent_buffer_page(), and will not detach page private. The section is marked by: - btrfs_page_inc_eb_refs() - btrfs_page_dec_eb_refs() Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-26 16:33:50 +08:00
if (PagePrivate(page)) {
subpage = (struct btrfs_subpage *)page->private;
if (atomic_read(&subpage->eb_refs))
return true;
btrfs: subpage: fix a rare race between metadata endio and eb freeing [BUG] There is a very rare ASSERT() triggering during full fstests run for subpage rw support. No other reproducer so far. The ASSERT() gets triggered for metadata read in btrfs_page_set_uptodate() inside end_page_read(). [CAUSE] There is still a small race window for metadata only, the race could happen like this: T1 | T2 ------------------------------------+----------------------------- end_bio_extent_readpage() | |- btrfs_validate_metadata_buffer() | | |- free_extent_buffer() | | Still have 2 refs | |- end_page_read() | |- if (unlikely(PagePrivate()) | | The page still has Private | | | free_extent_buffer() | | | Only one ref 1, will be | | | released | | |- detach_extent_buffer_page() | | |- btrfs_detach_subpage() |- btrfs_set_page_uptodate() | The page no longer has Private| >>> ASSERT() triggered <<< | This race window is super small, thus pretty hard to hit, even with so many runs of fstests. But the race window is still there, we have to go another way to solve it other than relying on random PagePrivate() check. Data path is not affected, as it will lock the page before reading, while unlocking the page after the last read has finished, thus no race window. [FIX] This patch will fix the bug by repurposing btrfs_subpage::readers. Now btrfs_subpage::readers will be a member shared by both metadata and data. For metadata path, we don't do the page unlock as metadata only relies on extent locking. At the same time, teach page_range_has_eb() to take btrfs_subpage::readers into consideration. So that even if the last eb of a page gets freed, page::private won't be detached as long as there still are pending end_page_read() calls. By this we eliminate the race window, this will slight increase the metadata memory usage, as the page may not be released as frequently as usual. But it should not be a big deal. The code got introduced in ("btrfs: submit read time repair only for each corrupted sector"), but the fix is in a separate patch to keep the problem description and the crash is rare so it should not hurt bisectability. Signed-off-by: Qu Wegruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-06-07 17:02:58 +08:00
/*
* Even there is no eb refs here, we may still have
* end_page_read() call relying on page::private.
*/
if (atomic_read(&subpage->readers))
return true;
btrfs: support subpage for extent buffer page release In btrfs_release_extent_buffer_pages(), we need to add extra handling for subpage. Introduce a helper, detach_extent_buffer_page(), to do different handling for regular and subpage cases. For subpage case, handle detaching page private. For unmapped (dummy or cloned) ebs, we can detach the page private immediately as the page can only be attached to one unmapped eb. For mapped ebs, we have to ensure there are no eb in the page range before we delete it, as page->private is shared between all ebs in the same page. But there is a subpage specific race, where we can race with extent buffer allocation, and clear the page private while new eb is still being utilized, like this: Extent buffer A is the new extent buffer which will be allocated, while extent buffer B is the last existing extent buffer of the page. T1 (eb A) | T2 (eb B) -------------------------------+------------------------------ alloc_extent_buffer() | btrfs_release_extent_buffer_pages() |- p = find_or_create_page() | | |- attach_extent_buffer_page() | | | | |- detach_extent_buffer_page() | | |- if (!page_range_has_eb()) | | | No new eb in the page range yet | | | As new eb A hasn't yet been | | | inserted into radix tree. | | |- btrfs_detach_subpage() | | |- detach_page_private(); |- radix_tree_insert() | Then we have a metadata eb whose page has no private bit. To avoid such race, we introduce a subpage metadata-specific member, btrfs_subpage::eb_refs. In alloc_extent_buffer() we increase eb_refs in the critical section of private_lock. Then page_range_has_eb() will return true for detach_extent_buffer_page(), and will not detach page private. The section is marked by: - btrfs_page_inc_eb_refs() - btrfs_page_dec_eb_refs() Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-26 16:33:50 +08:00
}
return false;
}
btrfs: support subpage for extent buffer page release In btrfs_release_extent_buffer_pages(), we need to add extra handling for subpage. Introduce a helper, detach_extent_buffer_page(), to do different handling for regular and subpage cases. For subpage case, handle detaching page private. For unmapped (dummy or cloned) ebs, we can detach the page private immediately as the page can only be attached to one unmapped eb. For mapped ebs, we have to ensure there are no eb in the page range before we delete it, as page->private is shared between all ebs in the same page. But there is a subpage specific race, where we can race with extent buffer allocation, and clear the page private while new eb is still being utilized, like this: Extent buffer A is the new extent buffer which will be allocated, while extent buffer B is the last existing extent buffer of the page. T1 (eb A) | T2 (eb B) -------------------------------+------------------------------ alloc_extent_buffer() | btrfs_release_extent_buffer_pages() |- p = find_or_create_page() | | |- attach_extent_buffer_page() | | | | |- detach_extent_buffer_page() | | |- if (!page_range_has_eb()) | | | No new eb in the page range yet | | | As new eb A hasn't yet been | | | inserted into radix tree. | | |- btrfs_detach_subpage() | | |- detach_page_private(); |- radix_tree_insert() | Then we have a metadata eb whose page has no private bit. To avoid such race, we introduce a subpage metadata-specific member, btrfs_subpage::eb_refs. In alloc_extent_buffer() we increase eb_refs in the critical section of private_lock. Then page_range_has_eb() will return true for detach_extent_buffer_page(), and will not detach page private. The section is marked by: - btrfs_page_inc_eb_refs() - btrfs_page_dec_eb_refs() Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-26 16:33:50 +08:00
static void detach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
/*
* For mapped eb, we're going to change the page private, which should
* be done under the private_lock.
*/
if (mapped)
spin_lock(&page->mapping->private_lock);
if (!PagePrivate(page)) {
if (mapped)
btrfs: support subpage for extent buffer page release In btrfs_release_extent_buffer_pages(), we need to add extra handling for subpage. Introduce a helper, detach_extent_buffer_page(), to do different handling for regular and subpage cases. For subpage case, handle detaching page private. For unmapped (dummy or cloned) ebs, we can detach the page private immediately as the page can only be attached to one unmapped eb. For mapped ebs, we have to ensure there are no eb in the page range before we delete it, as page->private is shared between all ebs in the same page. But there is a subpage specific race, where we can race with extent buffer allocation, and clear the page private while new eb is still being utilized, like this: Extent buffer A is the new extent buffer which will be allocated, while extent buffer B is the last existing extent buffer of the page. T1 (eb A) | T2 (eb B) -------------------------------+------------------------------ alloc_extent_buffer() | btrfs_release_extent_buffer_pages() |- p = find_or_create_page() | | |- attach_extent_buffer_page() | | | | |- detach_extent_buffer_page() | | |- if (!page_range_has_eb()) | | | No new eb in the page range yet | | | As new eb A hasn't yet been | | | inserted into radix tree. | | |- btrfs_detach_subpage() | | |- detach_page_private(); |- radix_tree_insert() | Then we have a metadata eb whose page has no private bit. To avoid such race, we introduce a subpage metadata-specific member, btrfs_subpage::eb_refs. In alloc_extent_buffer() we increase eb_refs in the critical section of private_lock. Then page_range_has_eb() will return true for detach_extent_buffer_page(), and will not detach page private. The section is marked by: - btrfs_page_inc_eb_refs() - btrfs_page_dec_eb_refs() Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-26 16:33:50 +08:00
spin_unlock(&page->mapping->private_lock);
return;
}
if (fs_info->nodesize >= PAGE_SIZE) {
/*
* We do this since we'll remove the pages after we've
* removed the eb from the radix tree, so we could race
* and have this page now attached to the new eb. So
* only clear page_private if it's still connected to
* this eb.
*/
if (PagePrivate(page) &&
page->private == (unsigned long)eb) {
BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
BUG_ON(PageDirty(page));
BUG_ON(PageWriteback(page));
/*
* We need to make sure we haven't be attached
* to a new eb.
*/
detach_page_private(page);
}
if (mapped)
spin_unlock(&page->mapping->private_lock);
btrfs: support subpage for extent buffer page release In btrfs_release_extent_buffer_pages(), we need to add extra handling for subpage. Introduce a helper, detach_extent_buffer_page(), to do different handling for regular and subpage cases. For subpage case, handle detaching page private. For unmapped (dummy or cloned) ebs, we can detach the page private immediately as the page can only be attached to one unmapped eb. For mapped ebs, we have to ensure there are no eb in the page range before we delete it, as page->private is shared between all ebs in the same page. But there is a subpage specific race, where we can race with extent buffer allocation, and clear the page private while new eb is still being utilized, like this: Extent buffer A is the new extent buffer which will be allocated, while extent buffer B is the last existing extent buffer of the page. T1 (eb A) | T2 (eb B) -------------------------------+------------------------------ alloc_extent_buffer() | btrfs_release_extent_buffer_pages() |- p = find_or_create_page() | | |- attach_extent_buffer_page() | | | | |- detach_extent_buffer_page() | | |- if (!page_range_has_eb()) | | | No new eb in the page range yet | | | As new eb A hasn't yet been | | | inserted into radix tree. | | |- btrfs_detach_subpage() | | |- detach_page_private(); |- radix_tree_insert() | Then we have a metadata eb whose page has no private bit. To avoid such race, we introduce a subpage metadata-specific member, btrfs_subpage::eb_refs. In alloc_extent_buffer() we increase eb_refs in the critical section of private_lock. Then page_range_has_eb() will return true for detach_extent_buffer_page(), and will not detach page private. The section is marked by: - btrfs_page_inc_eb_refs() - btrfs_page_dec_eb_refs() Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-26 16:33:50 +08:00
return;
}
/*
* For subpage, we can have dummy eb with page private. In this case,
* we can directly detach the private as such page is only attached to
* one dummy eb, no sharing.
*/
if (!mapped) {
btrfs_detach_subpage(fs_info, page);
return;
}
btrfs_page_dec_eb_refs(fs_info, page);
/*
* We can only detach the page private if there are no other ebs in the
btrfs: subpage: fix a rare race between metadata endio and eb freeing [BUG] There is a very rare ASSERT() triggering during full fstests run for subpage rw support. No other reproducer so far. The ASSERT() gets triggered for metadata read in btrfs_page_set_uptodate() inside end_page_read(). [CAUSE] There is still a small race window for metadata only, the race could happen like this: T1 | T2 ------------------------------------+----------------------------- end_bio_extent_readpage() | |- btrfs_validate_metadata_buffer() | | |- free_extent_buffer() | | Still have 2 refs | |- end_page_read() | |- if (unlikely(PagePrivate()) | | The page still has Private | | | free_extent_buffer() | | | Only one ref 1, will be | | | released | | |- detach_extent_buffer_page() | | |- btrfs_detach_subpage() |- btrfs_set_page_uptodate() | The page no longer has Private| >>> ASSERT() triggered <<< | This race window is super small, thus pretty hard to hit, even with so many runs of fstests. But the race window is still there, we have to go another way to solve it other than relying on random PagePrivate() check. Data path is not affected, as it will lock the page before reading, while unlocking the page after the last read has finished, thus no race window. [FIX] This patch will fix the bug by repurposing btrfs_subpage::readers. Now btrfs_subpage::readers will be a member shared by both metadata and data. For metadata path, we don't do the page unlock as metadata only relies on extent locking. At the same time, teach page_range_has_eb() to take btrfs_subpage::readers into consideration. So that even if the last eb of a page gets freed, page::private won't be detached as long as there still are pending end_page_read() calls. By this we eliminate the race window, this will slight increase the metadata memory usage, as the page may not be released as frequently as usual. But it should not be a big deal. The code got introduced in ("btrfs: submit read time repair only for each corrupted sector"), but the fix is in a separate patch to keep the problem description and the crash is rare so it should not hurt bisectability. Signed-off-by: Qu Wegruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-06-07 17:02:58 +08:00
* page range and no unfinished IO.
btrfs: support subpage for extent buffer page release In btrfs_release_extent_buffer_pages(), we need to add extra handling for subpage. Introduce a helper, detach_extent_buffer_page(), to do different handling for regular and subpage cases. For subpage case, handle detaching page private. For unmapped (dummy or cloned) ebs, we can detach the page private immediately as the page can only be attached to one unmapped eb. For mapped ebs, we have to ensure there are no eb in the page range before we delete it, as page->private is shared between all ebs in the same page. But there is a subpage specific race, where we can race with extent buffer allocation, and clear the page private while new eb is still being utilized, like this: Extent buffer A is the new extent buffer which will be allocated, while extent buffer B is the last existing extent buffer of the page. T1 (eb A) | T2 (eb B) -------------------------------+------------------------------ alloc_extent_buffer() | btrfs_release_extent_buffer_pages() |- p = find_or_create_page() | | |- attach_extent_buffer_page() | | | | |- detach_extent_buffer_page() | | |- if (!page_range_has_eb()) | | | No new eb in the page range yet | | | As new eb A hasn't yet been | | | inserted into radix tree. | | |- btrfs_detach_subpage() | | |- detach_page_private(); |- radix_tree_insert() | Then we have a metadata eb whose page has no private bit. To avoid such race, we introduce a subpage metadata-specific member, btrfs_subpage::eb_refs. In alloc_extent_buffer() we increase eb_refs in the critical section of private_lock. Then page_range_has_eb() will return true for detach_extent_buffer_page(), and will not detach page private. The section is marked by: - btrfs_page_inc_eb_refs() - btrfs_page_dec_eb_refs() Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-26 16:33:50 +08:00
*/
if (!page_range_has_eb(fs_info, page))
btrfs_detach_subpage(fs_info, page);
spin_unlock(&page->mapping->private_lock);
}
/* Release all pages attached to the extent buffer */
static void btrfs_release_extent_buffer_pages(struct extent_buffer *eb)
{
int i;
int num_pages;
ASSERT(!extent_buffer_under_io(eb));
num_pages = num_extent_pages(eb);
for (i = 0; i < num_pages; i++) {
struct page *page = eb->pages[i];
if (!page)
continue;
detach_extent_buffer_page(eb, page);
/* One for when we allocated the page */
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
put_page(page);
}
}
/*
* Helper for releasing the extent buffer.
*/
static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
{
btrfs_release_extent_buffer_pages(eb);
btrfs_leak_debug_del_eb(eb);
__free_extent_buffer(eb);
}
static struct extent_buffer *
__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
unsigned long len)
{
struct extent_buffer *eb = NULL;
btrfs: Prevent from early transaction abort Btrfs relies on GFP_NOFS allocation when committing the transaction but this allocation context is rather weak wrt. reclaim capabilities. The page allocator currently tries hard to not fail these allocations if they are small (<=PAGE_ALLOC_COSTLY_ORDER) so this is not a problem currently but there is an attempt to move away from the default no-fail behavior and allow these allocation to fail more eagerly. And this would lead to a pre-mature transaction abort as follows: [ 55.328093] Call Trace: [ 55.328890] [<ffffffff8154e6f0>] dump_stack+0x4f/0x7b [ 55.330518] [<ffffffff8108fa28>] ? console_unlock+0x334/0x363 [ 55.332738] [<ffffffff8110873e>] __alloc_pages_nodemask+0x81d/0x8d4 [ 55.334910] [<ffffffff81100752>] pagecache_get_page+0x10e/0x20c [ 55.336844] [<ffffffffa007d916>] alloc_extent_buffer+0xd0/0x350 [btrfs] [ 55.338973] [<ffffffffa0059d8c>] btrfs_find_create_tree_block+0x15/0x17 [btrfs] [ 55.341329] [<ffffffffa004f728>] btrfs_alloc_tree_block+0x18c/0x405 [btrfs] [ 55.343566] [<ffffffffa003fa34>] split_leaf+0x1e4/0x6a6 [btrfs] [ 55.345577] [<ffffffffa0040567>] btrfs_search_slot+0x671/0x831 [btrfs] [ 55.347679] [<ffffffff810682d7>] ? get_parent_ip+0xe/0x3e [ 55.349434] [<ffffffffa0041cb2>] btrfs_insert_empty_items+0x5d/0xa8 [btrfs] [ 55.351681] [<ffffffffa004ecfb>] __btrfs_run_delayed_refs+0x7a6/0xf35 [btrfs] [ 55.353979] [<ffffffffa00512ea>] btrfs_run_delayed_refs+0x6e/0x226 [btrfs] [ 55.356212] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.358378] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.360626] [<ffffffffa0060221>] btrfs_commit_transaction+0x4c/0xaba [btrfs] [ 55.362894] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.365221] [<ffffffffa0073428>] btrfs_sync_file+0x29c/0x310 [btrfs] [ 55.367273] [<ffffffff81186808>] vfs_fsync_range+0x8f/0x9e [ 55.369047] [<ffffffff81186833>] vfs_fsync+0x1c/0x1e [ 55.370654] [<ffffffff81186869>] do_fsync+0x34/0x4e [ 55.372246] [<ffffffff81186ab3>] SyS_fsync+0x10/0x14 [ 55.373851] [<ffffffff81554f97>] system_call_fastpath+0x12/0x6f [ 55.381070] BTRFS: error (device hdb1) in btrfs_run_delayed_refs:2821: errno=-12 Out of memory [ 55.382431] BTRFS warning (device hdb1): Skipping commit of aborted transaction. [ 55.382433] BTRFS warning (device hdb1): cleanup_transaction:1692: Aborting unused transaction(IO failure). [ 55.384280] ------------[ cut here ]------------ [ 55.384312] WARNING: CPU: 0 PID: 3010 at fs/btrfs/delayed-ref.c:438 btrfs_select_ref_head+0xd9/0xfe [btrfs]() [...] [ 55.384337] Call Trace: [ 55.384353] [<ffffffff8154e6f0>] dump_stack+0x4f/0x7b [ 55.384357] [<ffffffff8107f717>] ? down_trylock+0x2d/0x37 [ 55.384359] [<ffffffff81046977>] warn_slowpath_common+0xa1/0xbb [ 55.384398] [<ffffffffa00a1d6b>] ? btrfs_select_ref_head+0xd9/0xfe [btrfs] [ 55.384400] [<ffffffff81046a34>] warn_slowpath_null+0x1a/0x1c [ 55.384423] [<ffffffffa00a1d6b>] btrfs_select_ref_head+0xd9/0xfe [btrfs] [ 55.384446] [<ffffffffa004e5f7>] ? __btrfs_run_delayed_refs+0xa2/0xf35 [btrfs] [ 55.384455] [<ffffffffa004e600>] __btrfs_run_delayed_refs+0xab/0xf35 [btrfs] [ 55.384476] [<ffffffffa00512ea>] btrfs_run_delayed_refs+0x6e/0x226 [btrfs] [ 55.384499] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.384521] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.384543] [<ffffffffa0060221>] btrfs_commit_transaction+0x4c/0xaba [btrfs] [ 55.384565] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.384588] [<ffffffffa0073428>] btrfs_sync_file+0x29c/0x310 [btrfs] [ 55.384591] [<ffffffff81186808>] vfs_fsync_range+0x8f/0x9e [ 55.384592] [<ffffffff81186833>] vfs_fsync+0x1c/0x1e [ 55.384593] [<ffffffff81186869>] do_fsync+0x34/0x4e [ 55.384594] [<ffffffff81186ab3>] SyS_fsync+0x10/0x14 [ 55.384595] [<ffffffff81554f97>] system_call_fastpath+0x12/0x6f [...] [ 55.384608] ---[ end trace c29799da1d4dd621 ]--- [ 55.437323] BTRFS info (device hdb1): forced readonly [ 55.438815] BTRFS info (device hdb1): delayed_refs has NO entry Fix this by being explicit about the no-fail behavior of this allocation path and use __GFP_NOFAIL. Signed-off-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-19 20:17:40 +08:00
eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
eb->start = start;
eb->len = len;
eb->fs_info = fs_info;
btrfs: switch extent buffer tree lock to rw_semaphore Historically we've implemented our own locking because we wanted to be able to selectively spin or sleep based on what we were doing in the tree. For instance, if all of our nodes were in cache then there's rarely a reason to need to sleep waiting for node locks, as they'll likely become available soon. At the time this code was written the rw_semaphore didn't do adaptive spinning, and thus was orders of magnitude slower than our home grown locking. However now the opposite is the case. There are a few problems with how we implement blocking locks, namely that we use a normal waitqueue and simply wake everybody up in reverse sleep order. This leads to some suboptimal performance behavior, and a lot of context switches in highly contended cases. The rw_semaphores actually do this properly, and also have adaptive spinning that works relatively well. The locking code is also a bit of a bear to understand, and we lose the benefit of lockdep for the most part because the blocking states of the lock are simply ad-hoc and not mapped into lockdep. So rework the locking code to drop all of this custom locking stuff, and simply use a rw_semaphore for everything. This makes the locking much simpler for everything, as we can now drop a lot of cruft and blocking transitions. The performance numbers vary depending on the workload, because generally speaking there doesn't tend to be a lot of contention on the btree. However, on my test system which is an 80 core single socket system with 256GiB of RAM and a 2TiB NVMe drive I get the following results (with all debug options off): dbench 200 baseline Throughput 216.056 MB/sec 200 clients 200 procs max_latency=1471.197 ms dbench 200 with patch Throughput 737.188 MB/sec 200 clients 200 procs max_latency=714.346 ms Previously we also used fs_mark to test this sort of contention, and those results are far less impressive, mostly because there's not enough tasks to really stress the locking fs_mark -d /d[0-15] -S 0 -L 20 -n 100000 -s 0 -t 16 baseline Average Files/sec: 160166.7 p50 Files/sec: 165832 p90 Files/sec: 123886 p99 Files/sec: 123495 real 3m26.527s user 2m19.223s sys 48m21.856s patched Average Files/sec: 164135.7 p50 Files/sec: 171095 p90 Files/sec: 122889 p99 Files/sec: 113819 real 3m29.660s user 2m19.990s sys 44m12.259s Signed-off-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-20 23:46:09 +08:00
init_rwsem(&eb->lock);
Btrfs: Change btree locking to use explicit blocking points Most of the btrfs metadata operations can be protected by a spinlock, but some operations still need to schedule. So far, btrfs has been using a mutex along with a trylock loop, most of the time it is able to avoid going for the full mutex, so the trylock loop is a big performance gain. This commit is step one for getting rid of the blocking locks entirely. btrfs_tree_lock takes a spinlock, and the code explicitly switches to a blocking lock when it starts an operation that can schedule. We'll be able get rid of the blocking locks in smaller pieces over time. Tracing allows us to find the most common cause of blocking, so we can start with the hot spots first. The basic idea is: btrfs_tree_lock() returns with the spin lock held btrfs_set_lock_blocking() sets the EXTENT_BUFFER_BLOCKING bit in the extent buffer flags, and then drops the spin lock. The buffer is still considered locked by all of the btrfs code. If btrfs_tree_lock gets the spinlock but finds the blocking bit set, it drops the spin lock and waits on a wait queue for the blocking bit to go away. Much of the code that needs to set the blocking bit finishes without actually blocking a good percentage of the time. So, an adaptive spin is still used against the blocking bit to avoid very high context switch rates. btrfs_clear_lock_blocking() clears the blocking bit and returns with the spinlock held again. btrfs_tree_unlock() can be called on either blocking or spinning locks, it does the right thing based on the blocking bit. ctree.c has a helper function to set/clear all the locked buffers in a path as blocking. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:25:08 +08:00
btrfs_leak_debug_add_eb(eb);
INIT_LIST_HEAD(&eb->release_list);
spin_lock_init(&eb->refs_lock);
atomic_set(&eb->refs, 1);
atomic_set(&eb->io_pages, 0);
ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
return eb;
}
struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
{
int i;
struct extent_buffer *new;
int num_pages = num_extent_pages(src);
int ret;
new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
if (new == NULL)
return NULL;
/*
* Set UNMAPPED before calling btrfs_release_extent_buffer(), as
* btrfs_release_extent_buffer() have different behavior for
* UNMAPPED subpage extent buffer.
*/
set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
ret = btrfs_alloc_page_array(num_pages, new->pages);
if (ret) {
btrfs_release_extent_buffer(new);
return NULL;
}
for (i = 0; i < num_pages; i++) {
int ret;
struct page *p = new->pages[i];
ret = attach_extent_buffer_page(new, p, NULL);
if (ret < 0) {
btrfs_release_extent_buffer(new);
return NULL;
}
WARN_ON(PageDirty(p));
copy_page(page_address(p), page_address(src->pages[i]));
}
set_extent_buffer_uptodate(new);
return new;
}
struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
u64 start, unsigned long len)
{
struct extent_buffer *eb;
int num_pages;
int i;
int ret;
eb = __alloc_extent_buffer(fs_info, start, len);
if (!eb)
return NULL;
num_pages = num_extent_pages(eb);
ret = btrfs_alloc_page_array(num_pages, eb->pages);
if (ret)
goto err;
for (i = 0; i < num_pages; i++) {
struct page *p = eb->pages[i];
ret = attach_extent_buffer_page(eb, p, NULL);
if (ret < 0)
goto err;
}
set_extent_buffer_uptodate(eb);
btrfs_set_header_nritems(eb, 0);
set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
return eb;
err:
for (i = 0; i < num_pages; i++) {
if (eb->pages[i]) {
detach_extent_buffer_page(eb, eb->pages[i]);
__free_page(eb->pages[i]);
}
}
__free_extent_buffer(eb);
return NULL;
}
struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
u64 start)
{
return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
}
static void check_buffer_tree_ref(struct extent_buffer *eb)
{
int refs;
btrfs: fix fatal extent_buffer readahead vs releasepage race Under somewhat convoluted conditions, it is possible to attempt to release an extent_buffer that is under io, which triggers a BUG_ON in btrfs_release_extent_buffer_pages. This relies on a few different factors. First, extent_buffer reads done as readahead for searching use WAIT_NONE, so they free the local extent buffer reference while the io is outstanding. However, they should still be protected by TREE_REF. However, if the system is doing signficant reclaim, and simultaneously heavily accessing the extent_buffers, it is possible for releasepage to race with two concurrent readahead attempts in a way that leaves TREE_REF unset when the readahead extent buffer is released. Essentially, if two tasks race to allocate a new extent_buffer, but the winner who attempts the first io is rebuffed by a page being locked (likely by the reclaim itself) then the loser will still go ahead with issuing the readahead. The loser's call to find_extent_buffer must also race with the reclaim task reading the extent_buffer's refcount as 1 in a way that allows the reclaim to re-clear the TREE_REF checked by find_extent_buffer. The following represents an example execution demonstrating the race: CPU0 CPU1 CPU2 reada_for_search reada_for_search readahead_tree_block readahead_tree_block find_create_tree_block find_create_tree_block alloc_extent_buffer alloc_extent_buffer find_extent_buffer // not found allocates eb lock pages associate pages to eb insert eb into radix tree set TREE_REF, refs == 2 unlock pages read_extent_buffer_pages // WAIT_NONE not uptodate (brand new eb) lock_page if !trylock_page goto unlock_exit // not an error free_extent_buffer release_extent_buffer atomic_dec_and_test refs to 1 find_extent_buffer // found try_release_extent_buffer take refs_lock reads refs == 1; no io atomic_inc_not_zero refs to 2 mark_buffer_accessed check_buffer_tree_ref // not STALE, won't take refs_lock refs == 2; TREE_REF set // no action read_extent_buffer_pages // WAIT_NONE clear TREE_REF release_extent_buffer atomic_dec_and_test refs to 1 unlock_page still not uptodate (CPU1 read failed on trylock_page) locks pages set io_pages > 0 submit io return free_extent_buffer release_extent_buffer dec refs to 0 delete from radix tree btrfs_release_extent_buffer_pages BUG_ON(io_pages > 0)!!! We observe this at a very low rate in production and were also able to reproduce it in a test environment by introducing some spurious delays and by introducing probabilistic trylock_page failures. To fix it, we apply check_tree_ref at a point where it could not possibly be unset by a competing task: after io_pages has been incremented. All the codepaths that clear TREE_REF check for io, so they would not be able to clear it after this point until the io is done. Stack trace, for reference: [1417839.424739] ------------[ cut here ]------------ [1417839.435328] kernel BUG at fs/btrfs/extent_io.c:4841! [1417839.447024] invalid opcode: 0000 [#1] SMP [1417839.502972] RIP: 0010:btrfs_release_extent_buffer_pages+0x20/0x1f0 [1417839.517008] Code: ed e9 ... [1417839.558895] RSP: 0018:ffffc90020bcf798 EFLAGS: 00010202 [1417839.570816] RAX: 0000000000000002 RBX: ffff888102d6def0 RCX: 0000000000000028 [1417839.586962] RDX: 0000000000000002 RSI: ffff8887f0296482 RDI: ffff888102d6def0 [1417839.603108] RBP: ffff88885664a000 R08: 0000000000000046 R09: 0000000000000238 [1417839.619255] R10: 0000000000000028 R11: ffff88885664af68 R12: 0000000000000000 [1417839.635402] R13: 0000000000000000 R14: ffff88875f573ad0 R15: ffff888797aafd90 [1417839.651549] FS: 00007f5a844fa700(0000) GS:ffff88885f680000(0000) knlGS:0000000000000000 [1417839.669810] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [1417839.682887] CR2: 00007f7884541fe0 CR3: 000000049f609002 CR4: 00000000003606e0 [1417839.699037] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [1417839.715187] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [1417839.731320] Call Trace: [1417839.737103] release_extent_buffer+0x39/0x90 [1417839.746913] read_block_for_search.isra.38+0x2a3/0x370 [1417839.758645] btrfs_search_slot+0x260/0x9b0 [1417839.768054] btrfs_lookup_file_extent+0x4a/0x70 [1417839.778427] btrfs_get_extent+0x15f/0x830 [1417839.787665] ? submit_extent_page+0xc4/0x1c0 [1417839.797474] ? __do_readpage+0x299/0x7a0 [1417839.806515] __do_readpage+0x33b/0x7a0 [1417839.815171] ? btrfs_releasepage+0x70/0x70 [1417839.824597] extent_readpages+0x28f/0x400 [1417839.833836] read_pages+0x6a/0x1c0 [1417839.841729] ? startup_64+0x2/0x30 [1417839.849624] __do_page_cache_readahead+0x13c/0x1a0 [1417839.860590] filemap_fault+0x6c7/0x990 [1417839.869252] ? xas_load+0x8/0x80 [1417839.876756] ? xas_find+0x150/0x190 [1417839.884839] ? filemap_map_pages+0x295/0x3b0 [1417839.894652] __do_fault+0x32/0x110 [1417839.902540] __handle_mm_fault+0xacd/0x1000 [1417839.912156] handle_mm_fault+0xaa/0x1c0 [1417839.921004] __do_page_fault+0x242/0x4b0 [1417839.930044] ? page_fault+0x8/0x30 [1417839.937933] page_fault+0x1e/0x30 [1417839.945631] RIP: 0033:0x33c4bae [1417839.952927] Code: Bad RIP value. [1417839.960411] RSP: 002b:00007f5a844f7350 EFLAGS: 00010206 [1417839.972331] RAX: 000000000000006e RBX: 1614b3ff6a50398a RCX: 0000000000000000 [1417839.988477] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000002 [1417840.004626] RBP: 00007f5a844f7420 R08: 000000000000006e R09: 00007f5a94aeccb8 [1417840.020784] R10: 00007f5a844f7350 R11: 0000000000000000 R12: 00007f5a94aecc79 [1417840.036932] R13: 00007f5a94aecc78 R14: 00007f5a94aecc90 R15: 00007f5a94aecc40 CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2020-06-18 02:35:19 +08:00
/*
* The TREE_REF bit is first set when the extent_buffer is added
* to the radix tree. It is also reset, if unset, when a new reference
* is created by find_extent_buffer.
*
btrfs: fix fatal extent_buffer readahead vs releasepage race Under somewhat convoluted conditions, it is possible to attempt to release an extent_buffer that is under io, which triggers a BUG_ON in btrfs_release_extent_buffer_pages. This relies on a few different factors. First, extent_buffer reads done as readahead for searching use WAIT_NONE, so they free the local extent buffer reference while the io is outstanding. However, they should still be protected by TREE_REF. However, if the system is doing signficant reclaim, and simultaneously heavily accessing the extent_buffers, it is possible for releasepage to race with two concurrent readahead attempts in a way that leaves TREE_REF unset when the readahead extent buffer is released. Essentially, if two tasks race to allocate a new extent_buffer, but the winner who attempts the first io is rebuffed by a page being locked (likely by the reclaim itself) then the loser will still go ahead with issuing the readahead. The loser's call to find_extent_buffer must also race with the reclaim task reading the extent_buffer's refcount as 1 in a way that allows the reclaim to re-clear the TREE_REF checked by find_extent_buffer. The following represents an example execution demonstrating the race: CPU0 CPU1 CPU2 reada_for_search reada_for_search readahead_tree_block readahead_tree_block find_create_tree_block find_create_tree_block alloc_extent_buffer alloc_extent_buffer find_extent_buffer // not found allocates eb lock pages associate pages to eb insert eb into radix tree set TREE_REF, refs == 2 unlock pages read_extent_buffer_pages // WAIT_NONE not uptodate (brand new eb) lock_page if !trylock_page goto unlock_exit // not an error free_extent_buffer release_extent_buffer atomic_dec_and_test refs to 1 find_extent_buffer // found try_release_extent_buffer take refs_lock reads refs == 1; no io atomic_inc_not_zero refs to 2 mark_buffer_accessed check_buffer_tree_ref // not STALE, won't take refs_lock refs == 2; TREE_REF set // no action read_extent_buffer_pages // WAIT_NONE clear TREE_REF release_extent_buffer atomic_dec_and_test refs to 1 unlock_page still not uptodate (CPU1 read failed on trylock_page) locks pages set io_pages > 0 submit io return free_extent_buffer release_extent_buffer dec refs to 0 delete from radix tree btrfs_release_extent_buffer_pages BUG_ON(io_pages > 0)!!! We observe this at a very low rate in production and were also able to reproduce it in a test environment by introducing some spurious delays and by introducing probabilistic trylock_page failures. To fix it, we apply check_tree_ref at a point where it could not possibly be unset by a competing task: after io_pages has been incremented. All the codepaths that clear TREE_REF check for io, so they would not be able to clear it after this point until the io is done. Stack trace, for reference: [1417839.424739] ------------[ cut here ]------------ [1417839.435328] kernel BUG at fs/btrfs/extent_io.c:4841! [1417839.447024] invalid opcode: 0000 [#1] SMP [1417839.502972] RIP: 0010:btrfs_release_extent_buffer_pages+0x20/0x1f0 [1417839.517008] Code: ed e9 ... [1417839.558895] RSP: 0018:ffffc90020bcf798 EFLAGS: 00010202 [1417839.570816] RAX: 0000000000000002 RBX: ffff888102d6def0 RCX: 0000000000000028 [1417839.586962] RDX: 0000000000000002 RSI: ffff8887f0296482 RDI: ffff888102d6def0 [1417839.603108] RBP: ffff88885664a000 R08: 0000000000000046 R09: 0000000000000238 [1417839.619255] R10: 0000000000000028 R11: ffff88885664af68 R12: 0000000000000000 [1417839.635402] R13: 0000000000000000 R14: ffff88875f573ad0 R15: ffff888797aafd90 [1417839.651549] FS: 00007f5a844fa700(0000) GS:ffff88885f680000(0000) knlGS:0000000000000000 [1417839.669810] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [1417839.682887] CR2: 00007f7884541fe0 CR3: 000000049f609002 CR4: 00000000003606e0 [1417839.699037] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [1417839.715187] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [1417839.731320] Call Trace: [1417839.737103] release_extent_buffer+0x39/0x90 [1417839.746913] read_block_for_search.isra.38+0x2a3/0x370 [1417839.758645] btrfs_search_slot+0x260/0x9b0 [1417839.768054] btrfs_lookup_file_extent+0x4a/0x70 [1417839.778427] btrfs_get_extent+0x15f/0x830 [1417839.787665] ? submit_extent_page+0xc4/0x1c0 [1417839.797474] ? __do_readpage+0x299/0x7a0 [1417839.806515] __do_readpage+0x33b/0x7a0 [1417839.815171] ? btrfs_releasepage+0x70/0x70 [1417839.824597] extent_readpages+0x28f/0x400 [1417839.833836] read_pages+0x6a/0x1c0 [1417839.841729] ? startup_64+0x2/0x30 [1417839.849624] __do_page_cache_readahead+0x13c/0x1a0 [1417839.860590] filemap_fault+0x6c7/0x990 [1417839.869252] ? xas_load+0x8/0x80 [1417839.876756] ? xas_find+0x150/0x190 [1417839.884839] ? filemap_map_pages+0x295/0x3b0 [1417839.894652] __do_fault+0x32/0x110 [1417839.902540] __handle_mm_fault+0xacd/0x1000 [1417839.912156] handle_mm_fault+0xaa/0x1c0 [1417839.921004] __do_page_fault+0x242/0x4b0 [1417839.930044] ? page_fault+0x8/0x30 [1417839.937933] page_fault+0x1e/0x30 [1417839.945631] RIP: 0033:0x33c4bae [1417839.952927] Code: Bad RIP value. [1417839.960411] RSP: 002b:00007f5a844f7350 EFLAGS: 00010206 [1417839.972331] RAX: 000000000000006e RBX: 1614b3ff6a50398a RCX: 0000000000000000 [1417839.988477] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000002 [1417840.004626] RBP: 00007f5a844f7420 R08: 000000000000006e R09: 00007f5a94aeccb8 [1417840.020784] R10: 00007f5a844f7350 R11: 0000000000000000 R12: 00007f5a94aecc79 [1417840.036932] R13: 00007f5a94aecc78 R14: 00007f5a94aecc90 R15: 00007f5a94aecc40 CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2020-06-18 02:35:19 +08:00
* It is only cleared in two cases: freeing the last non-tree
* reference to the extent_buffer when its STALE bit is set or
* calling release_folio when the tree reference is the only reference.
*
btrfs: fix fatal extent_buffer readahead vs releasepage race Under somewhat convoluted conditions, it is possible to attempt to release an extent_buffer that is under io, which triggers a BUG_ON in btrfs_release_extent_buffer_pages. This relies on a few different factors. First, extent_buffer reads done as readahead for searching use WAIT_NONE, so they free the local extent buffer reference while the io is outstanding. However, they should still be protected by TREE_REF. However, if the system is doing signficant reclaim, and simultaneously heavily accessing the extent_buffers, it is possible for releasepage to race with two concurrent readahead attempts in a way that leaves TREE_REF unset when the readahead extent buffer is released. Essentially, if two tasks race to allocate a new extent_buffer, but the winner who attempts the first io is rebuffed by a page being locked (likely by the reclaim itself) then the loser will still go ahead with issuing the readahead. The loser's call to find_extent_buffer must also race with the reclaim task reading the extent_buffer's refcount as 1 in a way that allows the reclaim to re-clear the TREE_REF checked by find_extent_buffer. The following represents an example execution demonstrating the race: CPU0 CPU1 CPU2 reada_for_search reada_for_search readahead_tree_block readahead_tree_block find_create_tree_block find_create_tree_block alloc_extent_buffer alloc_extent_buffer find_extent_buffer // not found allocates eb lock pages associate pages to eb insert eb into radix tree set TREE_REF, refs == 2 unlock pages read_extent_buffer_pages // WAIT_NONE not uptodate (brand new eb) lock_page if !trylock_page goto unlock_exit // not an error free_extent_buffer release_extent_buffer atomic_dec_and_test refs to 1 find_extent_buffer // found try_release_extent_buffer take refs_lock reads refs == 1; no io atomic_inc_not_zero refs to 2 mark_buffer_accessed check_buffer_tree_ref // not STALE, won't take refs_lock refs == 2; TREE_REF set // no action read_extent_buffer_pages // WAIT_NONE clear TREE_REF release_extent_buffer atomic_dec_and_test refs to 1 unlock_page still not uptodate (CPU1 read failed on trylock_page) locks pages set io_pages > 0 submit io return free_extent_buffer release_extent_buffer dec refs to 0 delete from radix tree btrfs_release_extent_buffer_pages BUG_ON(io_pages > 0)!!! We observe this at a very low rate in production and were also able to reproduce it in a test environment by introducing some spurious delays and by introducing probabilistic trylock_page failures. To fix it, we apply check_tree_ref at a point where it could not possibly be unset by a competing task: after io_pages has been incremented. All the codepaths that clear TREE_REF check for io, so they would not be able to clear it after this point until the io is done. Stack trace, for reference: [1417839.424739] ------------[ cut here ]------------ [1417839.435328] kernel BUG at fs/btrfs/extent_io.c:4841! [1417839.447024] invalid opcode: 0000 [#1] SMP [1417839.502972] RIP: 0010:btrfs_release_extent_buffer_pages+0x20/0x1f0 [1417839.517008] Code: ed e9 ... [1417839.558895] RSP: 0018:ffffc90020bcf798 EFLAGS: 00010202 [1417839.570816] RAX: 0000000000000002 RBX: ffff888102d6def0 RCX: 0000000000000028 [1417839.586962] RDX: 0000000000000002 RSI: ffff8887f0296482 RDI: ffff888102d6def0 [1417839.603108] RBP: ffff88885664a000 R08: 0000000000000046 R09: 0000000000000238 [1417839.619255] R10: 0000000000000028 R11: ffff88885664af68 R12: 0000000000000000 [1417839.635402] R13: 0000000000000000 R14: ffff88875f573ad0 R15: ffff888797aafd90 [1417839.651549] FS: 00007f5a844fa700(0000) GS:ffff88885f680000(0000) knlGS:0000000000000000 [1417839.669810] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [1417839.682887] CR2: 00007f7884541fe0 CR3: 000000049f609002 CR4: 00000000003606e0 [1417839.699037] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [1417839.715187] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [1417839.731320] Call Trace: [1417839.737103] release_extent_buffer+0x39/0x90 [1417839.746913] read_block_for_search.isra.38+0x2a3/0x370 [1417839.758645] btrfs_search_slot+0x260/0x9b0 [1417839.768054] btrfs_lookup_file_extent+0x4a/0x70 [1417839.778427] btrfs_get_extent+0x15f/0x830 [1417839.787665] ? submit_extent_page+0xc4/0x1c0 [1417839.797474] ? __do_readpage+0x299/0x7a0 [1417839.806515] __do_readpage+0x33b/0x7a0 [1417839.815171] ? btrfs_releasepage+0x70/0x70 [1417839.824597] extent_readpages+0x28f/0x400 [1417839.833836] read_pages+0x6a/0x1c0 [1417839.841729] ? startup_64+0x2/0x30 [1417839.849624] __do_page_cache_readahead+0x13c/0x1a0 [1417839.860590] filemap_fault+0x6c7/0x990 [1417839.869252] ? xas_load+0x8/0x80 [1417839.876756] ? xas_find+0x150/0x190 [1417839.884839] ? filemap_map_pages+0x295/0x3b0 [1417839.894652] __do_fault+0x32/0x110 [1417839.902540] __handle_mm_fault+0xacd/0x1000 [1417839.912156] handle_mm_fault+0xaa/0x1c0 [1417839.921004] __do_page_fault+0x242/0x4b0 [1417839.930044] ? page_fault+0x8/0x30 [1417839.937933] page_fault+0x1e/0x30 [1417839.945631] RIP: 0033:0x33c4bae [1417839.952927] Code: Bad RIP value. [1417839.960411] RSP: 002b:00007f5a844f7350 EFLAGS: 00010206 [1417839.972331] RAX: 000000000000006e RBX: 1614b3ff6a50398a RCX: 0000000000000000 [1417839.988477] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000002 [1417840.004626] RBP: 00007f5a844f7420 R08: 000000000000006e R09: 00007f5a94aeccb8 [1417840.020784] R10: 00007f5a844f7350 R11: 0000000000000000 R12: 00007f5a94aecc79 [1417840.036932] R13: 00007f5a94aecc78 R14: 00007f5a94aecc90 R15: 00007f5a94aecc40 CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2020-06-18 02:35:19 +08:00
* In both cases, care is taken to ensure that the extent_buffer's
* pages are not under io. However, release_folio can be concurrently
btrfs: fix fatal extent_buffer readahead vs releasepage race Under somewhat convoluted conditions, it is possible to attempt to release an extent_buffer that is under io, which triggers a BUG_ON in btrfs_release_extent_buffer_pages. This relies on a few different factors. First, extent_buffer reads done as readahead for searching use WAIT_NONE, so they free the local extent buffer reference while the io is outstanding. However, they should still be protected by TREE_REF. However, if the system is doing signficant reclaim, and simultaneously heavily accessing the extent_buffers, it is possible for releasepage to race with two concurrent readahead attempts in a way that leaves TREE_REF unset when the readahead extent buffer is released. Essentially, if two tasks race to allocate a new extent_buffer, but the winner who attempts the first io is rebuffed by a page being locked (likely by the reclaim itself) then the loser will still go ahead with issuing the readahead. The loser's call to find_extent_buffer must also race with the reclaim task reading the extent_buffer's refcount as 1 in a way that allows the reclaim to re-clear the TREE_REF checked by find_extent_buffer. The following represents an example execution demonstrating the race: CPU0 CPU1 CPU2 reada_for_search reada_for_search readahead_tree_block readahead_tree_block find_create_tree_block find_create_tree_block alloc_extent_buffer alloc_extent_buffer find_extent_buffer // not found allocates eb lock pages associate pages to eb insert eb into radix tree set TREE_REF, refs == 2 unlock pages read_extent_buffer_pages // WAIT_NONE not uptodate (brand new eb) lock_page if !trylock_page goto unlock_exit // not an error free_extent_buffer release_extent_buffer atomic_dec_and_test refs to 1 find_extent_buffer // found try_release_extent_buffer take refs_lock reads refs == 1; no io atomic_inc_not_zero refs to 2 mark_buffer_accessed check_buffer_tree_ref // not STALE, won't take refs_lock refs == 2; TREE_REF set // no action read_extent_buffer_pages // WAIT_NONE clear TREE_REF release_extent_buffer atomic_dec_and_test refs to 1 unlock_page still not uptodate (CPU1 read failed on trylock_page) locks pages set io_pages > 0 submit io return free_extent_buffer release_extent_buffer dec refs to 0 delete from radix tree btrfs_release_extent_buffer_pages BUG_ON(io_pages > 0)!!! We observe this at a very low rate in production and were also able to reproduce it in a test environment by introducing some spurious delays and by introducing probabilistic trylock_page failures. To fix it, we apply check_tree_ref at a point where it could not possibly be unset by a competing task: after io_pages has been incremented. All the codepaths that clear TREE_REF check for io, so they would not be able to clear it after this point until the io is done. Stack trace, for reference: [1417839.424739] ------------[ cut here ]------------ [1417839.435328] kernel BUG at fs/btrfs/extent_io.c:4841! [1417839.447024] invalid opcode: 0000 [#1] SMP [1417839.502972] RIP: 0010:btrfs_release_extent_buffer_pages+0x20/0x1f0 [1417839.517008] Code: ed e9 ... [1417839.558895] RSP: 0018:ffffc90020bcf798 EFLAGS: 00010202 [1417839.570816] RAX: 0000000000000002 RBX: ffff888102d6def0 RCX: 0000000000000028 [1417839.586962] RDX: 0000000000000002 RSI: ffff8887f0296482 RDI: ffff888102d6def0 [1417839.603108] RBP: ffff88885664a000 R08: 0000000000000046 R09: 0000000000000238 [1417839.619255] R10: 0000000000000028 R11: ffff88885664af68 R12: 0000000000000000 [1417839.635402] R13: 0000000000000000 R14: ffff88875f573ad0 R15: ffff888797aafd90 [1417839.651549] FS: 00007f5a844fa700(0000) GS:ffff88885f680000(0000) knlGS:0000000000000000 [1417839.669810] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [1417839.682887] CR2: 00007f7884541fe0 CR3: 000000049f609002 CR4: 00000000003606e0 [1417839.699037] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [1417839.715187] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [1417839.731320] Call Trace: [1417839.737103] release_extent_buffer+0x39/0x90 [1417839.746913] read_block_for_search.isra.38+0x2a3/0x370 [1417839.758645] btrfs_search_slot+0x260/0x9b0 [1417839.768054] btrfs_lookup_file_extent+0x4a/0x70 [1417839.778427] btrfs_get_extent+0x15f/0x830 [1417839.787665] ? submit_extent_page+0xc4/0x1c0 [1417839.797474] ? __do_readpage+0x299/0x7a0 [1417839.806515] __do_readpage+0x33b/0x7a0 [1417839.815171] ? btrfs_releasepage+0x70/0x70 [1417839.824597] extent_readpages+0x28f/0x400 [1417839.833836] read_pages+0x6a/0x1c0 [1417839.841729] ? startup_64+0x2/0x30 [1417839.849624] __do_page_cache_readahead+0x13c/0x1a0 [1417839.860590] filemap_fault+0x6c7/0x990 [1417839.869252] ? xas_load+0x8/0x80 [1417839.876756] ? xas_find+0x150/0x190 [1417839.884839] ? filemap_map_pages+0x295/0x3b0 [1417839.894652] __do_fault+0x32/0x110 [1417839.902540] __handle_mm_fault+0xacd/0x1000 [1417839.912156] handle_mm_fault+0xaa/0x1c0 [1417839.921004] __do_page_fault+0x242/0x4b0 [1417839.930044] ? page_fault+0x8/0x30 [1417839.937933] page_fault+0x1e/0x30 [1417839.945631] RIP: 0033:0x33c4bae [1417839.952927] Code: Bad RIP value. [1417839.960411] RSP: 002b:00007f5a844f7350 EFLAGS: 00010206 [1417839.972331] RAX: 000000000000006e RBX: 1614b3ff6a50398a RCX: 0000000000000000 [1417839.988477] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000002 [1417840.004626] RBP: 00007f5a844f7420 R08: 000000000000006e R09: 00007f5a94aeccb8 [1417840.020784] R10: 00007f5a844f7350 R11: 0000000000000000 R12: 00007f5a94aecc79 [1417840.036932] R13: 00007f5a94aecc78 R14: 00007f5a94aecc90 R15: 00007f5a94aecc40 CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2020-06-18 02:35:19 +08:00
* called with creating new references, which is prone to race
* conditions between the calls to check_buffer_tree_ref in those
* codepaths and clearing TREE_REF in try_release_extent_buffer.
*
btrfs: fix fatal extent_buffer readahead vs releasepage race Under somewhat convoluted conditions, it is possible to attempt to release an extent_buffer that is under io, which triggers a BUG_ON in btrfs_release_extent_buffer_pages. This relies on a few different factors. First, extent_buffer reads done as readahead for searching use WAIT_NONE, so they free the local extent buffer reference while the io is outstanding. However, they should still be protected by TREE_REF. However, if the system is doing signficant reclaim, and simultaneously heavily accessing the extent_buffers, it is possible for releasepage to race with two concurrent readahead attempts in a way that leaves TREE_REF unset when the readahead extent buffer is released. Essentially, if two tasks race to allocate a new extent_buffer, but the winner who attempts the first io is rebuffed by a page being locked (likely by the reclaim itself) then the loser will still go ahead with issuing the readahead. The loser's call to find_extent_buffer must also race with the reclaim task reading the extent_buffer's refcount as 1 in a way that allows the reclaim to re-clear the TREE_REF checked by find_extent_buffer. The following represents an example execution demonstrating the race: CPU0 CPU1 CPU2 reada_for_search reada_for_search readahead_tree_block readahead_tree_block find_create_tree_block find_create_tree_block alloc_extent_buffer alloc_extent_buffer find_extent_buffer // not found allocates eb lock pages associate pages to eb insert eb into radix tree set TREE_REF, refs == 2 unlock pages read_extent_buffer_pages // WAIT_NONE not uptodate (brand new eb) lock_page if !trylock_page goto unlock_exit // not an error free_extent_buffer release_extent_buffer atomic_dec_and_test refs to 1 find_extent_buffer // found try_release_extent_buffer take refs_lock reads refs == 1; no io atomic_inc_not_zero refs to 2 mark_buffer_accessed check_buffer_tree_ref // not STALE, won't take refs_lock refs == 2; TREE_REF set // no action read_extent_buffer_pages // WAIT_NONE clear TREE_REF release_extent_buffer atomic_dec_and_test refs to 1 unlock_page still not uptodate (CPU1 read failed on trylock_page) locks pages set io_pages > 0 submit io return free_extent_buffer release_extent_buffer dec refs to 0 delete from radix tree btrfs_release_extent_buffer_pages BUG_ON(io_pages > 0)!!! We observe this at a very low rate in production and were also able to reproduce it in a test environment by introducing some spurious delays and by introducing probabilistic trylock_page failures. To fix it, we apply check_tree_ref at a point where it could not possibly be unset by a competing task: after io_pages has been incremented. All the codepaths that clear TREE_REF check for io, so they would not be able to clear it after this point until the io is done. Stack trace, for reference: [1417839.424739] ------------[ cut here ]------------ [1417839.435328] kernel BUG at fs/btrfs/extent_io.c:4841! [1417839.447024] invalid opcode: 0000 [#1] SMP [1417839.502972] RIP: 0010:btrfs_release_extent_buffer_pages+0x20/0x1f0 [1417839.517008] Code: ed e9 ... [1417839.558895] RSP: 0018:ffffc90020bcf798 EFLAGS: 00010202 [1417839.570816] RAX: 0000000000000002 RBX: ffff888102d6def0 RCX: 0000000000000028 [1417839.586962] RDX: 0000000000000002 RSI: ffff8887f0296482 RDI: ffff888102d6def0 [1417839.603108] RBP: ffff88885664a000 R08: 0000000000000046 R09: 0000000000000238 [1417839.619255] R10: 0000000000000028 R11: ffff88885664af68 R12: 0000000000000000 [1417839.635402] R13: 0000000000000000 R14: ffff88875f573ad0 R15: ffff888797aafd90 [1417839.651549] FS: 00007f5a844fa700(0000) GS:ffff88885f680000(0000) knlGS:0000000000000000 [1417839.669810] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [1417839.682887] CR2: 00007f7884541fe0 CR3: 000000049f609002 CR4: 00000000003606e0 [1417839.699037] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [1417839.715187] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [1417839.731320] Call Trace: [1417839.737103] release_extent_buffer+0x39/0x90 [1417839.746913] read_block_for_search.isra.38+0x2a3/0x370 [1417839.758645] btrfs_search_slot+0x260/0x9b0 [1417839.768054] btrfs_lookup_file_extent+0x4a/0x70 [1417839.778427] btrfs_get_extent+0x15f/0x830 [1417839.787665] ? submit_extent_page+0xc4/0x1c0 [1417839.797474] ? __do_readpage+0x299/0x7a0 [1417839.806515] __do_readpage+0x33b/0x7a0 [1417839.815171] ? btrfs_releasepage+0x70/0x70 [1417839.824597] extent_readpages+0x28f/0x400 [1417839.833836] read_pages+0x6a/0x1c0 [1417839.841729] ? startup_64+0x2/0x30 [1417839.849624] __do_page_cache_readahead+0x13c/0x1a0 [1417839.860590] filemap_fault+0x6c7/0x990 [1417839.869252] ? xas_load+0x8/0x80 [1417839.876756] ? xas_find+0x150/0x190 [1417839.884839] ? filemap_map_pages+0x295/0x3b0 [1417839.894652] __do_fault+0x32/0x110 [1417839.902540] __handle_mm_fault+0xacd/0x1000 [1417839.912156] handle_mm_fault+0xaa/0x1c0 [1417839.921004] __do_page_fault+0x242/0x4b0 [1417839.930044] ? page_fault+0x8/0x30 [1417839.937933] page_fault+0x1e/0x30 [1417839.945631] RIP: 0033:0x33c4bae [1417839.952927] Code: Bad RIP value. [1417839.960411] RSP: 002b:00007f5a844f7350 EFLAGS: 00010206 [1417839.972331] RAX: 000000000000006e RBX: 1614b3ff6a50398a RCX: 0000000000000000 [1417839.988477] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000002 [1417840.004626] RBP: 00007f5a844f7420 R08: 000000000000006e R09: 00007f5a94aeccb8 [1417840.020784] R10: 00007f5a844f7350 R11: 0000000000000000 R12: 00007f5a94aecc79 [1417840.036932] R13: 00007f5a94aecc78 R14: 00007f5a94aecc90 R15: 00007f5a94aecc40 CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2020-06-18 02:35:19 +08:00
* The actual lifetime of the extent_buffer in the radix tree is
* adequately protected by the refcount, but the TREE_REF bit and
* its corresponding reference are not. To protect against this
* class of races, we call check_buffer_tree_ref from the codepaths
* which trigger io after they set eb->io_pages. Note that once io is
* initiated, TREE_REF can no longer be cleared, so that is the
* moment at which any such race is best fixed.
*/
refs = atomic_read(&eb->refs);
if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
return;
spin_lock(&eb->refs_lock);
if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
atomic_inc(&eb->refs);
spin_unlock(&eb->refs_lock);
}
mm: non-atomically mark page accessed during page cache allocation where possible aops->write_begin may allocate a new page and make it visible only to have mark_page_accessed called almost immediately after. Once the page is visible the atomic operations are necessary which is noticable overhead when writing to an in-memory filesystem like tmpfs but should also be noticable with fast storage. The objective of the patch is to initialse the accessed information with non-atomic operations before the page is visible. The bulk of filesystems directly or indirectly use grab_cache_page_write_begin or find_or_create_page for the initial allocation of a page cache page. This patch adds an init_page_accessed() helper which behaves like the first call to mark_page_accessed() but may called before the page is visible and can be done non-atomically. The primary APIs of concern in this care are the following and are used by most filesystems. find_get_page find_lock_page find_or_create_page grab_cache_page_nowait grab_cache_page_write_begin All of them are very similar in detail to the patch creates a core helper pagecache_get_page() which takes a flags parameter that affects its behavior such as whether the page should be marked accessed or not. Then old API is preserved but is basically a thin wrapper around this core function. Each of the filesystems are then updated to avoid calling mark_page_accessed when it is known that the VM interfaces have already done the job. There is a slight snag in that the timing of the mark_page_accessed() has now changed so in rare cases it's possible a page gets to the end of the LRU as PageReferenced where as previously it might have been repromoted. This is expected to be rare but it's worth the filesystem people thinking about it in case they see a problem with the timing change. It is also the case that some filesystems may be marking pages accessed that previously did not but it makes sense that filesystems have consistent behaviour in this regard. The test case used to evaulate this is a simple dd of a large file done multiple times with the file deleted on each iterations. The size of the file is 1/10th physical memory to avoid dirty page balancing. In the async case it will be possible that the workload completes without even hitting the disk and will have variable results but highlight the impact of mark_page_accessed for async IO. The sync results are expected to be more stable. The exception is tmpfs where the normal case is for the "IO" to not hit the disk. The test machine was single socket and UMA to avoid any scheduling or NUMA artifacts. Throughput and wall times are presented for sync IO, only wall times are shown for async as the granularity reported by dd and the variability is unsuitable for comparison. As async results were variable do to writback timings, I'm only reporting the maximum figures. The sync results were stable enough to make the mean and stddev uninteresting. The performance results are reported based on a run with no profiling. Profile data is based on a separate run with oprofile running. async dd 3.15.0-rc3 3.15.0-rc3 vanilla accessed-v2 ext3 Max elapsed 13.9900 ( 0.00%) 11.5900 ( 17.16%) tmpfs Max elapsed 0.5100 ( 0.00%) 0.4900 ( 3.92%) btrfs Max elapsed 12.8100 ( 0.00%) 12.7800 ( 0.23%) ext4 Max elapsed 18.6000 ( 0.00%) 13.3400 ( 28.28%) xfs Max elapsed 12.5600 ( 0.00%) 2.0900 ( 83.36%) The XFS figure is a bit strange as it managed to avoid a worst case by sheer luck but the average figures looked reasonable. samples percentage ext3 86107 0.9783 vmlinux-3.15.0-rc4-vanilla mark_page_accessed ext3 23833 0.2710 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed ext3 5036 0.0573 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed ext4 64566 0.8961 vmlinux-3.15.0-rc4-vanilla mark_page_accessed ext4 5322 0.0713 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed ext4 2869 0.0384 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed xfs 62126 1.7675 vmlinux-3.15.0-rc4-vanilla mark_page_accessed xfs 1904 0.0554 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed xfs 103 0.0030 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed btrfs 10655 0.1338 vmlinux-3.15.0-rc4-vanilla mark_page_accessed btrfs 2020 0.0273 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed btrfs 587 0.0079 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed tmpfs 59562 3.2628 vmlinux-3.15.0-rc4-vanilla mark_page_accessed tmpfs 1210 0.0696 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed tmpfs 94 0.0054 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed [akpm@linux-foundation.org: don't run init_page_accessed() against an uninitialised pointer] Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Jan Kara <jack@suse.cz> Cc: Michal Hocko <mhocko@suse.cz> Cc: Hugh Dickins <hughd@google.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Theodore Ts'o <tytso@mit.edu> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Rik van Riel <riel@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Tested-by: Prabhakar Lad <prabhakar.csengg@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-05 07:10:31 +08:00
static void mark_extent_buffer_accessed(struct extent_buffer *eb,
struct page *accessed)
{
int num_pages, i;
check_buffer_tree_ref(eb);
num_pages = num_extent_pages(eb);
for (i = 0; i < num_pages; i++) {
struct page *p = eb->pages[i];
mm: non-atomically mark page accessed during page cache allocation where possible aops->write_begin may allocate a new page and make it visible only to have mark_page_accessed called almost immediately after. Once the page is visible the atomic operations are necessary which is noticable overhead when writing to an in-memory filesystem like tmpfs but should also be noticable with fast storage. The objective of the patch is to initialse the accessed information with non-atomic operations before the page is visible. The bulk of filesystems directly or indirectly use grab_cache_page_write_begin or find_or_create_page for the initial allocation of a page cache page. This patch adds an init_page_accessed() helper which behaves like the first call to mark_page_accessed() but may called before the page is visible and can be done non-atomically. The primary APIs of concern in this care are the following and are used by most filesystems. find_get_page find_lock_page find_or_create_page grab_cache_page_nowait grab_cache_page_write_begin All of them are very similar in detail to the patch creates a core helper pagecache_get_page() which takes a flags parameter that affects its behavior such as whether the page should be marked accessed or not. Then old API is preserved but is basically a thin wrapper around this core function. Each of the filesystems are then updated to avoid calling mark_page_accessed when it is known that the VM interfaces have already done the job. There is a slight snag in that the timing of the mark_page_accessed() has now changed so in rare cases it's possible a page gets to the end of the LRU as PageReferenced where as previously it might have been repromoted. This is expected to be rare but it's worth the filesystem people thinking about it in case they see a problem with the timing change. It is also the case that some filesystems may be marking pages accessed that previously did not but it makes sense that filesystems have consistent behaviour in this regard. The test case used to evaulate this is a simple dd of a large file done multiple times with the file deleted on each iterations. The size of the file is 1/10th physical memory to avoid dirty page balancing. In the async case it will be possible that the workload completes without even hitting the disk and will have variable results but highlight the impact of mark_page_accessed for async IO. The sync results are expected to be more stable. The exception is tmpfs where the normal case is for the "IO" to not hit the disk. The test machine was single socket and UMA to avoid any scheduling or NUMA artifacts. Throughput and wall times are presented for sync IO, only wall times are shown for async as the granularity reported by dd and the variability is unsuitable for comparison. As async results were variable do to writback timings, I'm only reporting the maximum figures. The sync results were stable enough to make the mean and stddev uninteresting. The performance results are reported based on a run with no profiling. Profile data is based on a separate run with oprofile running. async dd 3.15.0-rc3 3.15.0-rc3 vanilla accessed-v2 ext3 Max elapsed 13.9900 ( 0.00%) 11.5900 ( 17.16%) tmpfs Max elapsed 0.5100 ( 0.00%) 0.4900 ( 3.92%) btrfs Max elapsed 12.8100 ( 0.00%) 12.7800 ( 0.23%) ext4 Max elapsed 18.6000 ( 0.00%) 13.3400 ( 28.28%) xfs Max elapsed 12.5600 ( 0.00%) 2.0900 ( 83.36%) The XFS figure is a bit strange as it managed to avoid a worst case by sheer luck but the average figures looked reasonable. samples percentage ext3 86107 0.9783 vmlinux-3.15.0-rc4-vanilla mark_page_accessed ext3 23833 0.2710 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed ext3 5036 0.0573 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed ext4 64566 0.8961 vmlinux-3.15.0-rc4-vanilla mark_page_accessed ext4 5322 0.0713 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed ext4 2869 0.0384 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed xfs 62126 1.7675 vmlinux-3.15.0-rc4-vanilla mark_page_accessed xfs 1904 0.0554 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed xfs 103 0.0030 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed btrfs 10655 0.1338 vmlinux-3.15.0-rc4-vanilla mark_page_accessed btrfs 2020 0.0273 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed btrfs 587 0.0079 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed tmpfs 59562 3.2628 vmlinux-3.15.0-rc4-vanilla mark_page_accessed tmpfs 1210 0.0696 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed tmpfs 94 0.0054 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed [akpm@linux-foundation.org: don't run init_page_accessed() against an uninitialised pointer] Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Jan Kara <jack@suse.cz> Cc: Michal Hocko <mhocko@suse.cz> Cc: Hugh Dickins <hughd@google.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Theodore Ts'o <tytso@mit.edu> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Rik van Riel <riel@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Tested-by: Prabhakar Lad <prabhakar.csengg@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-05 07:10:31 +08:00
if (p != accessed)
mark_page_accessed(p);
}
}
struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
u64 start)
{
struct extent_buffer *eb;
eb = find_extent_buffer_nolock(fs_info, start);
if (!eb)
return NULL;
/*
* Lock our eb's refs_lock to avoid races with free_extent_buffer().
* When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
* another task running free_extent_buffer() might have seen that flag
* set, eb->refs == 2, that the buffer isn't under IO (dirty and
* writeback flags not set) and it's still in the tree (flag
* EXTENT_BUFFER_TREE_REF set), therefore being in the process of
* decrementing the extent buffer's reference count twice. So here we
* could race and increment the eb's reference count, clear its stale
* flag, mark it as dirty and drop our reference before the other task
* finishes executing free_extent_buffer, which would later result in
* an attempt to free an extent buffer that is dirty.
*/
if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
spin_lock(&eb->refs_lock);
spin_unlock(&eb->refs_lock);
}
mark_extent_buffer_accessed(eb, NULL);
return eb;
}
#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
u64 start)
{
struct extent_buffer *eb, *exists = NULL;
int ret;
eb = find_extent_buffer(fs_info, start);
if (eb)
return eb;
eb = alloc_dummy_extent_buffer(fs_info, start);
if (!eb)
return ERR_PTR(-ENOMEM);
eb->fs_info = fs_info;
again:
ret = radix_tree_preload(GFP_NOFS);
if (ret) {
exists = ERR_PTR(ret);
goto free_eb;
}
spin_lock(&fs_info->buffer_lock);
ret = radix_tree_insert(&fs_info->buffer_radix,
start >> fs_info->sectorsize_bits, eb);
spin_unlock(&fs_info->buffer_lock);
radix_tree_preload_end();
if (ret == -EEXIST) {
exists = find_extent_buffer(fs_info, start);
if (exists)
goto free_eb;
else
goto again;
}
check_buffer_tree_ref(eb);
set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
return eb;
free_eb:
btrfs_release_extent_buffer(eb);
return exists;
}
#endif
static struct extent_buffer *grab_extent_buffer(
struct btrfs_fs_info *fs_info, struct page *page)
{
struct extent_buffer *exists;
/*
* For subpage case, we completely rely on radix tree to ensure we
* don't try to insert two ebs for the same bytenr. So here we always
* return NULL and just continue.
*/
if (fs_info->nodesize < PAGE_SIZE)
return NULL;
/* Page not yet attached to an extent buffer */
if (!PagePrivate(page))
return NULL;
/*
* We could have already allocated an eb for this page and attached one
* so lets see if we can get a ref on the existing eb, and if we can we
* know it's good and we can just return that one, else we know we can
* just overwrite page->private.
*/
exists = (struct extent_buffer *)page->private;
if (atomic_inc_not_zero(&exists->refs))
return exists;
WARN_ON(PageDirty(page));
detach_page_private(page);
return NULL;
}
static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
{
if (!IS_ALIGNED(start, fs_info->sectorsize)) {
btrfs_err(fs_info, "bad tree block start %llu", start);
return -EINVAL;
}
if (fs_info->nodesize < PAGE_SIZE &&
offset_in_page(start) + fs_info->nodesize > PAGE_SIZE) {
btrfs_err(fs_info,
"tree block crosses page boundary, start %llu nodesize %u",
start, fs_info->nodesize);
return -EINVAL;
}
if (fs_info->nodesize >= PAGE_SIZE &&
!PAGE_ALIGNED(start)) {
btrfs_err(fs_info,
"tree block is not page aligned, start %llu nodesize %u",
start, fs_info->nodesize);
return -EINVAL;
}
return 0;
}
struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
u64 start, u64 owner_root, int level)
{
unsigned long len = fs_info->nodesize;
int num_pages;
int i;
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
unsigned long index = start >> PAGE_SHIFT;
struct extent_buffer *eb;
struct extent_buffer *exists = NULL;
struct page *p;
struct address_space *mapping = fs_info->btree_inode->i_mapping;
btrfs: fix lockdep splat with reloc root extent buffers We have been hitting the following lockdep splat with btrfs/187 recently WARNING: possible circular locking dependency detected 5.19.0-rc8+ #775 Not tainted ------------------------------------------------------ btrfs/752500 is trying to acquire lock: ffff97e1875a97b8 (btrfs-treloc-02#2){+.+.}-{3:3}, at: __btrfs_tree_lock+0x24/0x110 but task is already holding lock: ffff97e1875a9278 (btrfs-tree-01/1){+.+.}-{3:3}, at: __btrfs_tree_lock+0x24/0x110 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (btrfs-tree-01/1){+.+.}-{3:3}: down_write_nested+0x41/0x80 __btrfs_tree_lock+0x24/0x110 btrfs_init_new_buffer+0x7d/0x2c0 btrfs_alloc_tree_block+0x120/0x3b0 __btrfs_cow_block+0x136/0x600 btrfs_cow_block+0x10b/0x230 btrfs_search_slot+0x53b/0xb70 btrfs_lookup_inode+0x2a/0xa0 __btrfs_update_delayed_inode+0x5f/0x280 btrfs_async_run_delayed_root+0x24c/0x290 btrfs_work_helper+0xf2/0x3e0 process_one_work+0x271/0x590 worker_thread+0x52/0x3b0 kthread+0xf0/0x120 ret_from_fork+0x1f/0x30 -> #1 (btrfs-tree-01){++++}-{3:3}: down_write_nested+0x41/0x80 __btrfs_tree_lock+0x24/0x110 btrfs_search_slot+0x3c3/0xb70 do_relocation+0x10c/0x6b0 relocate_tree_blocks+0x317/0x6d0 relocate_block_group+0x1f1/0x560 btrfs_relocate_block_group+0x23e/0x400 btrfs_relocate_chunk+0x4c/0x140 btrfs_balance+0x755/0xe40 btrfs_ioctl+0x1ea2/0x2c90 __x64_sys_ioctl+0x88/0xc0 do_syscall_64+0x38/0x90 entry_SYSCALL_64_after_hwframe+0x63/0xcd -> #0 (btrfs-treloc-02#2){+.+.}-{3:3}: __lock_acquire+0x1122/0x1e10 lock_acquire+0xc2/0x2d0 down_write_nested+0x41/0x80 __btrfs_tree_lock+0x24/0x110 btrfs_lock_root_node+0x31/0x50 btrfs_search_slot+0x1cb/0xb70 replace_path+0x541/0x9f0 merge_reloc_root+0x1d6/0x610 merge_reloc_roots+0xe2/0x260 relocate_block_group+0x2c8/0x560 btrfs_relocate_block_group+0x23e/0x400 btrfs_relocate_chunk+0x4c/0x140 btrfs_balance+0x755/0xe40 btrfs_ioctl+0x1ea2/0x2c90 __x64_sys_ioctl+0x88/0xc0 do_syscall_64+0x38/0x90 entry_SYSCALL_64_after_hwframe+0x63/0xcd other info that might help us debug this: Chain exists of: btrfs-treloc-02#2 --> btrfs-tree-01 --> btrfs-tree-01/1 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(btrfs-tree-01/1); lock(btrfs-tree-01); lock(btrfs-tree-01/1); lock(btrfs-treloc-02#2); *** DEADLOCK *** 7 locks held by btrfs/752500: #0: ffff97e292fdf460 (sb_writers#12){.+.+}-{0:0}, at: btrfs_ioctl+0x208/0x2c90 #1: ffff97e284c02050 (&fs_info->reclaim_bgs_lock){+.+.}-{3:3}, at: btrfs_balance+0x55f/0xe40 #2: ffff97e284c00878 (&fs_info->cleaner_mutex){+.+.}-{3:3}, at: btrfs_relocate_block_group+0x236/0x400 #3: ffff97e292fdf650 (sb_internal#2){.+.+}-{0:0}, at: merge_reloc_root+0xef/0x610 #4: ffff97e284c02378 (btrfs_trans_num_writers){++++}-{0:0}, at: join_transaction+0x1a8/0x5a0 #5: ffff97e284c023a0 (btrfs_trans_num_extwriters){++++}-{0:0}, at: join_transaction+0x1a8/0x5a0 #6: ffff97e1875a9278 (btrfs-tree-01/1){+.+.}-{3:3}, at: __btrfs_tree_lock+0x24/0x110 stack backtrace: CPU: 1 PID: 752500 Comm: btrfs Not tainted 5.19.0-rc8+ #775 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014 Call Trace: dump_stack_lvl+0x56/0x73 check_noncircular+0xd6/0x100 ? lock_is_held_type+0xe2/0x140 __lock_acquire+0x1122/0x1e10 lock_acquire+0xc2/0x2d0 ? __btrfs_tree_lock+0x24/0x110 down_write_nested+0x41/0x80 ? __btrfs_tree_lock+0x24/0x110 __btrfs_tree_lock+0x24/0x110 btrfs_lock_root_node+0x31/0x50 btrfs_search_slot+0x1cb/0xb70 ? lock_release+0x137/0x2d0 ? _raw_spin_unlock+0x29/0x50 ? release_extent_buffer+0x128/0x180 replace_path+0x541/0x9f0 merge_reloc_root+0x1d6/0x610 merge_reloc_roots+0xe2/0x260 relocate_block_group+0x2c8/0x560 btrfs_relocate_block_group+0x23e/0x400 btrfs_relocate_chunk+0x4c/0x140 btrfs_balance+0x755/0xe40 btrfs_ioctl+0x1ea2/0x2c90 ? lock_is_held_type+0xe2/0x140 ? lock_is_held_type+0xe2/0x140 ? __x64_sys_ioctl+0x88/0xc0 __x64_sys_ioctl+0x88/0xc0 do_syscall_64+0x38/0x90 entry_SYSCALL_64_after_hwframe+0x63/0xcd This isn't necessarily new, it's just tricky to hit in practice. There are two competing things going on here. With relocation we create a snapshot of every fs tree with a reloc tree. Any extent buffers that get initialized here are initialized with the reloc root lockdep key. However since it is a snapshot, any blocks that are currently in cache that originally belonged to the fs tree will have the normal tree lockdep key set. This creates the lock dependency of reloc tree -> normal tree for the extent buffer locking during the first phase of the relocation as we walk down the reloc root to relocate blocks. However this is problematic because the final phase of the relocation is merging the reloc root into the original fs root. This involves searching down to any keys that exist in the original fs root and then swapping the relocated block and the original fs root block. We have to search down to the fs root first, and then go search the reloc root for the block we need to replace. This creates the dependency of normal tree -> reloc tree which is why lockdep complains. Additionally even if we were to fix this particular mismatch with a different nesting for the merge case, we're still slotting in a block that has a owner of the reloc root objectid into a normal tree, so that block will have its lockdep key set to the tree reloc root, and create a lockdep splat later on when we wander into that block from the fs root. Unfortunately the only solution here is to make sure we do not set the lockdep key to the reloc tree lockdep key normally, and then reset any blocks we wander into from the reloc root when we're doing the merged. This solves the problem of having mixed tree reloc keys intermixed with normal tree keys, and then allows us to make sure in the merge case we maintain the lock order of normal tree -> reloc tree We handle this by setting a bit on the reloc root when we do the search for the block we want to relocate, and any block we search into or COW at that point gets set to the reloc tree key. This works correctly because we only ever COW down to the parent node, so we aren't resetting the key for the block we're linking into the fs root. With this patch we no longer have the lockdep splat in btrfs/187. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-07-27 04:24:04 +08:00
u64 lockdep_owner = owner_root;
int uptodate = 1;
int ret;
if (check_eb_alignment(fs_info, start))
return ERR_PTR(-EINVAL);
btrfs: more graceful errors/warnings on 32bit systems when reaching limits Btrfs uses internally mapped u64 address space for all its metadata. Due to the page cache limit on 32bit systems, btrfs can't access metadata at or beyond (ULONG_MAX + 1) << PAGE_SHIFT. See how MAX_LFS_FILESIZE and page::index are defined. This is 16T for 4K page size while 256T for 64K page size. Users can have a filesystem which doesn't have metadata beyond the boundary at mount time, but later balance can cause it to create metadata beyond the boundary. And modification to MM layer is unrealistic just for such minor use case. We can't do more than to prevent mounting such filesystem or warn early when the numbers are still within the limits. To address such problem, this patch will introduce the following checks: - Mount time rejection This will reject any fs which has metadata chunk at or beyond the boundary. - Mount time early warning If there is any metadata chunk beyond 5/8th of the boundary, we do an early warning and hope the end user will see it. - Runtime extent buffer rejection If we're going to allocate an extent buffer at or beyond the boundary, reject such request with EOVERFLOW. This is definitely going to cause problems like transaction abort, but we have no better ways. - Runtime extent buffer early warning If an extent buffer beyond 5/8th of the max file size is allocated, do an early warning. Above error/warning message will only be printed once for each fs to reduce dmesg flood. If the mount is rejected, the filesystem will be mountable only on a 64bit host. Link: https://lore.kernel.org/linux-btrfs/1783f16d-7a28-80e6-4c32-fdf19b705ed0@gmx.com/ Reported-by: Erik Jensen <erikjensen@rkjnsn.net> Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-25 09:18:14 +08:00
#if BITS_PER_LONG == 32
if (start >= MAX_LFS_FILESIZE) {
btrfs_err_rl(fs_info,
"extent buffer %llu is beyond 32bit page cache limit", start);
btrfs_err_32bit_limit(fs_info);
return ERR_PTR(-EOVERFLOW);
}
if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
btrfs_warn_32bit_limit(fs_info);
#endif
eb = find_extent_buffer(fs_info, start);
if (eb)
return eb;
eb = __alloc_extent_buffer(fs_info, start, len);
if (!eb)
return ERR_PTR(-ENOMEM);
btrfs: fix lockdep splat with reloc root extent buffers We have been hitting the following lockdep splat with btrfs/187 recently WARNING: possible circular locking dependency detected 5.19.0-rc8+ #775 Not tainted ------------------------------------------------------ btrfs/752500 is trying to acquire lock: ffff97e1875a97b8 (btrfs-treloc-02#2){+.+.}-{3:3}, at: __btrfs_tree_lock+0x24/0x110 but task is already holding lock: ffff97e1875a9278 (btrfs-tree-01/1){+.+.}-{3:3}, at: __btrfs_tree_lock+0x24/0x110 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (btrfs-tree-01/1){+.+.}-{3:3}: down_write_nested+0x41/0x80 __btrfs_tree_lock+0x24/0x110 btrfs_init_new_buffer+0x7d/0x2c0 btrfs_alloc_tree_block+0x120/0x3b0 __btrfs_cow_block+0x136/0x600 btrfs_cow_block+0x10b/0x230 btrfs_search_slot+0x53b/0xb70 btrfs_lookup_inode+0x2a/0xa0 __btrfs_update_delayed_inode+0x5f/0x280 btrfs_async_run_delayed_root+0x24c/0x290 btrfs_work_helper+0xf2/0x3e0 process_one_work+0x271/0x590 worker_thread+0x52/0x3b0 kthread+0xf0/0x120 ret_from_fork+0x1f/0x30 -> #1 (btrfs-tree-01){++++}-{3:3}: down_write_nested+0x41/0x80 __btrfs_tree_lock+0x24/0x110 btrfs_search_slot+0x3c3/0xb70 do_relocation+0x10c/0x6b0 relocate_tree_blocks+0x317/0x6d0 relocate_block_group+0x1f1/0x560 btrfs_relocate_block_group+0x23e/0x400 btrfs_relocate_chunk+0x4c/0x140 btrfs_balance+0x755/0xe40 btrfs_ioctl+0x1ea2/0x2c90 __x64_sys_ioctl+0x88/0xc0 do_syscall_64+0x38/0x90 entry_SYSCALL_64_after_hwframe+0x63/0xcd -> #0 (btrfs-treloc-02#2){+.+.}-{3:3}: __lock_acquire+0x1122/0x1e10 lock_acquire+0xc2/0x2d0 down_write_nested+0x41/0x80 __btrfs_tree_lock+0x24/0x110 btrfs_lock_root_node+0x31/0x50 btrfs_search_slot+0x1cb/0xb70 replace_path+0x541/0x9f0 merge_reloc_root+0x1d6/0x610 merge_reloc_roots+0xe2/0x260 relocate_block_group+0x2c8/0x560 btrfs_relocate_block_group+0x23e/0x400 btrfs_relocate_chunk+0x4c/0x140 btrfs_balance+0x755/0xe40 btrfs_ioctl+0x1ea2/0x2c90 __x64_sys_ioctl+0x88/0xc0 do_syscall_64+0x38/0x90 entry_SYSCALL_64_after_hwframe+0x63/0xcd other info that might help us debug this: Chain exists of: btrfs-treloc-02#2 --> btrfs-tree-01 --> btrfs-tree-01/1 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(btrfs-tree-01/1); lock(btrfs-tree-01); lock(btrfs-tree-01/1); lock(btrfs-treloc-02#2); *** DEADLOCK *** 7 locks held by btrfs/752500: #0: ffff97e292fdf460 (sb_writers#12){.+.+}-{0:0}, at: btrfs_ioctl+0x208/0x2c90 #1: ffff97e284c02050 (&fs_info->reclaim_bgs_lock){+.+.}-{3:3}, at: btrfs_balance+0x55f/0xe40 #2: ffff97e284c00878 (&fs_info->cleaner_mutex){+.+.}-{3:3}, at: btrfs_relocate_block_group+0x236/0x400 #3: ffff97e292fdf650 (sb_internal#2){.+.+}-{0:0}, at: merge_reloc_root+0xef/0x610 #4: ffff97e284c02378 (btrfs_trans_num_writers){++++}-{0:0}, at: join_transaction+0x1a8/0x5a0 #5: ffff97e284c023a0 (btrfs_trans_num_extwriters){++++}-{0:0}, at: join_transaction+0x1a8/0x5a0 #6: ffff97e1875a9278 (btrfs-tree-01/1){+.+.}-{3:3}, at: __btrfs_tree_lock+0x24/0x110 stack backtrace: CPU: 1 PID: 752500 Comm: btrfs Not tainted 5.19.0-rc8+ #775 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014 Call Trace: dump_stack_lvl+0x56/0x73 check_noncircular+0xd6/0x100 ? lock_is_held_type+0xe2/0x140 __lock_acquire+0x1122/0x1e10 lock_acquire+0xc2/0x2d0 ? __btrfs_tree_lock+0x24/0x110 down_write_nested+0x41/0x80 ? __btrfs_tree_lock+0x24/0x110 __btrfs_tree_lock+0x24/0x110 btrfs_lock_root_node+0x31/0x50 btrfs_search_slot+0x1cb/0xb70 ? lock_release+0x137/0x2d0 ? _raw_spin_unlock+0x29/0x50 ? release_extent_buffer+0x128/0x180 replace_path+0x541/0x9f0 merge_reloc_root+0x1d6/0x610 merge_reloc_roots+0xe2/0x260 relocate_block_group+0x2c8/0x560 btrfs_relocate_block_group+0x23e/0x400 btrfs_relocate_chunk+0x4c/0x140 btrfs_balance+0x755/0xe40 btrfs_ioctl+0x1ea2/0x2c90 ? lock_is_held_type+0xe2/0x140 ? lock_is_held_type+0xe2/0x140 ? __x64_sys_ioctl+0x88/0xc0 __x64_sys_ioctl+0x88/0xc0 do_syscall_64+0x38/0x90 entry_SYSCALL_64_after_hwframe+0x63/0xcd This isn't necessarily new, it's just tricky to hit in practice. There are two competing things going on here. With relocation we create a snapshot of every fs tree with a reloc tree. Any extent buffers that get initialized here are initialized with the reloc root lockdep key. However since it is a snapshot, any blocks that are currently in cache that originally belonged to the fs tree will have the normal tree lockdep key set. This creates the lock dependency of reloc tree -> normal tree for the extent buffer locking during the first phase of the relocation as we walk down the reloc root to relocate blocks. However this is problematic because the final phase of the relocation is merging the reloc root into the original fs root. This involves searching down to any keys that exist in the original fs root and then swapping the relocated block and the original fs root block. We have to search down to the fs root first, and then go search the reloc root for the block we need to replace. This creates the dependency of normal tree -> reloc tree which is why lockdep complains. Additionally even if we were to fix this particular mismatch with a different nesting for the merge case, we're still slotting in a block that has a owner of the reloc root objectid into a normal tree, so that block will have its lockdep key set to the tree reloc root, and create a lockdep splat later on when we wander into that block from the fs root. Unfortunately the only solution here is to make sure we do not set the lockdep key to the reloc tree lockdep key normally, and then reset any blocks we wander into from the reloc root when we're doing the merged. This solves the problem of having mixed tree reloc keys intermixed with normal tree keys, and then allows us to make sure in the merge case we maintain the lock order of normal tree -> reloc tree We handle this by setting a bit on the reloc root when we do the search for the block we want to relocate, and any block we search into or COW at that point gets set to the reloc tree key. This works correctly because we only ever COW down to the parent node, so we aren't resetting the key for the block we're linking into the fs root. With this patch we no longer have the lockdep splat in btrfs/187. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-07-27 04:24:04 +08:00
/*
* The reloc trees are just snapshots, so we need them to appear to be
* just like any other fs tree WRT lockdep.
*/
if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID)
lockdep_owner = BTRFS_FS_TREE_OBJECTID;
btrfs_set_buffer_lockdep_class(lockdep_owner, eb, level);
num_pages = num_extent_pages(eb);
for (i = 0; i < num_pages; i++, index++) {
struct btrfs_subpage *prealloc = NULL;
btrfs: Prevent from early transaction abort Btrfs relies on GFP_NOFS allocation when committing the transaction but this allocation context is rather weak wrt. reclaim capabilities. The page allocator currently tries hard to not fail these allocations if they are small (<=PAGE_ALLOC_COSTLY_ORDER) so this is not a problem currently but there is an attempt to move away from the default no-fail behavior and allow these allocation to fail more eagerly. And this would lead to a pre-mature transaction abort as follows: [ 55.328093] Call Trace: [ 55.328890] [<ffffffff8154e6f0>] dump_stack+0x4f/0x7b [ 55.330518] [<ffffffff8108fa28>] ? console_unlock+0x334/0x363 [ 55.332738] [<ffffffff8110873e>] __alloc_pages_nodemask+0x81d/0x8d4 [ 55.334910] [<ffffffff81100752>] pagecache_get_page+0x10e/0x20c [ 55.336844] [<ffffffffa007d916>] alloc_extent_buffer+0xd0/0x350 [btrfs] [ 55.338973] [<ffffffffa0059d8c>] btrfs_find_create_tree_block+0x15/0x17 [btrfs] [ 55.341329] [<ffffffffa004f728>] btrfs_alloc_tree_block+0x18c/0x405 [btrfs] [ 55.343566] [<ffffffffa003fa34>] split_leaf+0x1e4/0x6a6 [btrfs] [ 55.345577] [<ffffffffa0040567>] btrfs_search_slot+0x671/0x831 [btrfs] [ 55.347679] [<ffffffff810682d7>] ? get_parent_ip+0xe/0x3e [ 55.349434] [<ffffffffa0041cb2>] btrfs_insert_empty_items+0x5d/0xa8 [btrfs] [ 55.351681] [<ffffffffa004ecfb>] __btrfs_run_delayed_refs+0x7a6/0xf35 [btrfs] [ 55.353979] [<ffffffffa00512ea>] btrfs_run_delayed_refs+0x6e/0x226 [btrfs] [ 55.356212] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.358378] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.360626] [<ffffffffa0060221>] btrfs_commit_transaction+0x4c/0xaba [btrfs] [ 55.362894] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.365221] [<ffffffffa0073428>] btrfs_sync_file+0x29c/0x310 [btrfs] [ 55.367273] [<ffffffff81186808>] vfs_fsync_range+0x8f/0x9e [ 55.369047] [<ffffffff81186833>] vfs_fsync+0x1c/0x1e [ 55.370654] [<ffffffff81186869>] do_fsync+0x34/0x4e [ 55.372246] [<ffffffff81186ab3>] SyS_fsync+0x10/0x14 [ 55.373851] [<ffffffff81554f97>] system_call_fastpath+0x12/0x6f [ 55.381070] BTRFS: error (device hdb1) in btrfs_run_delayed_refs:2821: errno=-12 Out of memory [ 55.382431] BTRFS warning (device hdb1): Skipping commit of aborted transaction. [ 55.382433] BTRFS warning (device hdb1): cleanup_transaction:1692: Aborting unused transaction(IO failure). [ 55.384280] ------------[ cut here ]------------ [ 55.384312] WARNING: CPU: 0 PID: 3010 at fs/btrfs/delayed-ref.c:438 btrfs_select_ref_head+0xd9/0xfe [btrfs]() [...] [ 55.384337] Call Trace: [ 55.384353] [<ffffffff8154e6f0>] dump_stack+0x4f/0x7b [ 55.384357] [<ffffffff8107f717>] ? down_trylock+0x2d/0x37 [ 55.384359] [<ffffffff81046977>] warn_slowpath_common+0xa1/0xbb [ 55.384398] [<ffffffffa00a1d6b>] ? btrfs_select_ref_head+0xd9/0xfe [btrfs] [ 55.384400] [<ffffffff81046a34>] warn_slowpath_null+0x1a/0x1c [ 55.384423] [<ffffffffa00a1d6b>] btrfs_select_ref_head+0xd9/0xfe [btrfs] [ 55.384446] [<ffffffffa004e5f7>] ? __btrfs_run_delayed_refs+0xa2/0xf35 [btrfs] [ 55.384455] [<ffffffffa004e600>] __btrfs_run_delayed_refs+0xab/0xf35 [btrfs] [ 55.384476] [<ffffffffa00512ea>] btrfs_run_delayed_refs+0x6e/0x226 [btrfs] [ 55.384499] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.384521] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.384543] [<ffffffffa0060221>] btrfs_commit_transaction+0x4c/0xaba [btrfs] [ 55.384565] [<ffffffffa0060e21>] ? start_transaction+0x192/0x534 [btrfs] [ 55.384588] [<ffffffffa0073428>] btrfs_sync_file+0x29c/0x310 [btrfs] [ 55.384591] [<ffffffff81186808>] vfs_fsync_range+0x8f/0x9e [ 55.384592] [<ffffffff81186833>] vfs_fsync+0x1c/0x1e [ 55.384593] [<ffffffff81186869>] do_fsync+0x34/0x4e [ 55.384594] [<ffffffff81186ab3>] SyS_fsync+0x10/0x14 [ 55.384595] [<ffffffff81554f97>] system_call_fastpath+0x12/0x6f [...] [ 55.384608] ---[ end trace c29799da1d4dd621 ]--- [ 55.437323] BTRFS info (device hdb1): forced readonly [ 55.438815] BTRFS info (device hdb1): delayed_refs has NO entry Fix this by being explicit about the no-fail behavior of this allocation path and use __GFP_NOFAIL. Signed-off-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-19 20:17:40 +08:00
p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
if (!p) {
exists = ERR_PTR(-ENOMEM);
goto free_eb;
}
/*
* Preallocate page->private for subpage case, so that we won't
* allocate memory with private_lock hold. The memory will be
* freed by attach_extent_buffer_page() or freed manually if
* we exit earlier.
*
* Although we have ensured one subpage eb can only have one
* page, but it may change in the future for 16K page size
* support, so we still preallocate the memory in the loop.
*/
if (fs_info->nodesize < PAGE_SIZE) {
prealloc = btrfs_alloc_subpage(fs_info, BTRFS_SUBPAGE_METADATA);
if (IS_ERR(prealloc)) {
ret = PTR_ERR(prealloc);
unlock_page(p);
put_page(p);
exists = ERR_PTR(ret);
goto free_eb;
}
}
spin_lock(&mapping->private_lock);
exists = grab_extent_buffer(fs_info, p);
if (exists) {
spin_unlock(&mapping->private_lock);
unlock_page(p);
put_page(p);
mark_extent_buffer_accessed(exists, p);
btrfs_free_subpage(prealloc);
goto free_eb;
}
/* Should not fail, as we have preallocated the memory */
ret = attach_extent_buffer_page(eb, p, prealloc);
ASSERT(!ret);
btrfs: support subpage for extent buffer page release In btrfs_release_extent_buffer_pages(), we need to add extra handling for subpage. Introduce a helper, detach_extent_buffer_page(), to do different handling for regular and subpage cases. For subpage case, handle detaching page private. For unmapped (dummy or cloned) ebs, we can detach the page private immediately as the page can only be attached to one unmapped eb. For mapped ebs, we have to ensure there are no eb in the page range before we delete it, as page->private is shared between all ebs in the same page. But there is a subpage specific race, where we can race with extent buffer allocation, and clear the page private while new eb is still being utilized, like this: Extent buffer A is the new extent buffer which will be allocated, while extent buffer B is the last existing extent buffer of the page. T1 (eb A) | T2 (eb B) -------------------------------+------------------------------ alloc_extent_buffer() | btrfs_release_extent_buffer_pages() |- p = find_or_create_page() | | |- attach_extent_buffer_page() | | | | |- detach_extent_buffer_page() | | |- if (!page_range_has_eb()) | | | No new eb in the page range yet | | | As new eb A hasn't yet been | | | inserted into radix tree. | | |- btrfs_detach_subpage() | | |- detach_page_private(); |- radix_tree_insert() | Then we have a metadata eb whose page has no private bit. To avoid such race, we introduce a subpage metadata-specific member, btrfs_subpage::eb_refs. In alloc_extent_buffer() we increase eb_refs in the critical section of private_lock. Then page_range_has_eb() will return true for detach_extent_buffer_page(), and will not detach page private. The section is marked by: - btrfs_page_inc_eb_refs() - btrfs_page_dec_eb_refs() Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-26 16:33:50 +08:00
/*
* To inform we have extra eb under allocation, so that
* detach_extent_buffer_page() won't release the page private
* when the eb hasn't yet been inserted into radix tree.
*
* The ref will be decreased when the eb released the page, in
* detach_extent_buffer_page().
* Thus needs no special handling in error path.
*/
btrfs_page_inc_eb_refs(fs_info, p);
spin_unlock(&mapping->private_lock);
WARN_ON(btrfs_page_test_dirty(fs_info, p, eb->start, eb->len));
eb->pages[i] = p;
if (!PageUptodate(p))
uptodate = 0;
/*
* We can't unlock the pages just yet since the extent buffer
* hasn't been properly inserted in the radix tree, this
* opens a race with btree_release_folio which can free a page
* while we are still filling in all pages for the buffer and
* we could crash.
*/
}
if (uptodate)
Btrfs: Change btree locking to use explicit blocking points Most of the btrfs metadata operations can be protected by a spinlock, but some operations still need to schedule. So far, btrfs has been using a mutex along with a trylock loop, most of the time it is able to avoid going for the full mutex, so the trylock loop is a big performance gain. This commit is step one for getting rid of the blocking locks entirely. btrfs_tree_lock takes a spinlock, and the code explicitly switches to a blocking lock when it starts an operation that can schedule. We'll be able get rid of the blocking locks in smaller pieces over time. Tracing allows us to find the most common cause of blocking, so we can start with the hot spots first. The basic idea is: btrfs_tree_lock() returns with the spin lock held btrfs_set_lock_blocking() sets the EXTENT_BUFFER_BLOCKING bit in the extent buffer flags, and then drops the spin lock. The buffer is still considered locked by all of the btrfs code. If btrfs_tree_lock gets the spinlock but finds the blocking bit set, it drops the spin lock and waits on a wait queue for the blocking bit to go away. Much of the code that needs to set the blocking bit finishes without actually blocking a good percentage of the time. So, an adaptive spin is still used against the blocking bit to avoid very high context switch rates. btrfs_clear_lock_blocking() clears the blocking bit and returns with the spinlock held again. btrfs_tree_unlock() can be called on either blocking or spinning locks, it does the right thing based on the blocking bit. ctree.c has a helper function to set/clear all the locked buffers in a path as blocking. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:25:08 +08:00
set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
again:
ret = radix_tree_preload(GFP_NOFS);
if (ret) {
exists = ERR_PTR(ret);
goto free_eb;
}
spin_lock(&fs_info->buffer_lock);
ret = radix_tree_insert(&fs_info->buffer_radix,
start >> fs_info->sectorsize_bits, eb);
spin_unlock(&fs_info->buffer_lock);
radix_tree_preload_end();
if (ret == -EEXIST) {
exists = find_extent_buffer(fs_info, start);
if (exists)
goto free_eb;
else
goto again;
}
/* add one reference for the tree */
check_buffer_tree_ref(eb);
set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
/*
* Now it's safe to unlock the pages because any calls to
* btree_release_folio will correctly detect that a page belongs to a
* live buffer and won't free them prematurely.
*/
for (i = 0; i < num_pages; i++)
unlock_page(eb->pages[i]);
return eb;
free_eb:
WARN_ON(!atomic_dec_and_test(&eb->refs));
for (i = 0; i < num_pages; i++) {
if (eb->pages[i])
unlock_page(eb->pages[i]);
}
btrfs_release_extent_buffer(eb);
return exists;
}
static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
{
struct extent_buffer *eb =
container_of(head, struct extent_buffer, rcu_head);
__free_extent_buffer(eb);
}
static int release_extent_buffer(struct extent_buffer *eb)
__releases(&eb->refs_lock)
{
lockdep_assert_held(&eb->refs_lock);
WARN_ON(atomic_read(&eb->refs) == 0);
if (atomic_dec_and_test(&eb->refs)) {
if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
struct btrfs_fs_info *fs_info = eb->fs_info;
spin_unlock(&eb->refs_lock);
spin_lock(&fs_info->buffer_lock);
radix_tree_delete(&fs_info->buffer_radix,
eb->start >> fs_info->sectorsize_bits);
spin_unlock(&fs_info->buffer_lock);
} else {
spin_unlock(&eb->refs_lock);
}
btrfs_leak_debug_del_eb(eb);
/* Should be safe to release our pages at this point */
btrfs_release_extent_buffer_pages(eb);
#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
__free_extent_buffer(eb);
return 1;
}
#endif
call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
return 1;
}
spin_unlock(&eb->refs_lock);
return 0;
}
void free_extent_buffer(struct extent_buffer *eb)
{
int refs;
if (!eb)
return;
refs = atomic_read(&eb->refs);
while (1) {
if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
|| (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
refs == 1))
break;
if (atomic_try_cmpxchg(&eb->refs, &refs, refs - 1))
return;
}
spin_lock(&eb->refs_lock);
if (atomic_read(&eb->refs) == 2 &&
test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
!extent_buffer_under_io(eb) &&
test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
atomic_dec(&eb->refs);
/*
* I know this is terrible, but it's temporary until we stop tracking
* the uptodate bits and such for the extent buffers.
*/
release_extent_buffer(eb);
}
void free_extent_buffer_stale(struct extent_buffer *eb)
{
if (!eb)
return;
spin_lock(&eb->refs_lock);
set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
atomic_dec(&eb->refs);
release_extent_buffer(eb);
}
btrfs: make set/clear_extent_buffer_dirty() subpage compatible For set_extent_buffer_dirty() to support subpage sized metadata, just call btrfs_page_set_dirty() to handle both cases. For clear_extent_buffer_dirty(), it needs to clear the page dirty if and only if all extent buffers in the page range are no longer dirty. Also do the same for page error. This is pretty different from the existing clear_extent_buffer_dirty() routine, so add a new helper function, clear_subpage_extent_buffer_dirty() to do this for subpage metadata. Also since the main part of clearing page dirty code is still the same, extract that into btree_clear_page_dirty() so that it can be utilized for both cases. But there is a special race between set_extent_buffer_dirty() and clear_extent_buffer_dirty(), where we can clear the page dirty. [POSSIBLE RACE WINDOW] For the race window between clear_subpage_extent_buffer_dirty() and set_extent_buffer_dirty(), due to the fact that we can't call clear_page_dirty_for_io() under subpage spin lock, we can race like below: T1 (eb1 in the same page) | T2 (eb2 in the same page) -------------------------------+------------------------------ set_extent_buffer_dirty() | clear_extent_buffer_dirty() |- was_dirty = false; | |- clear_subpagE_extent_buffer_dirty() | | |- btrfs_clear_and_test_dirty() | | | Since eb2 is the last dirty page | | | we got: | | | last == true; | | | |- btrfs_page_set_dirty() | | | We set the page dirty and | | | subpage dirty bitmap | | | | |- if (last) | | | Since we don't have subpage lock | | | held, now @last is no longer | | | correct | | |- btree_clear_page_dirty() | | Now PageDirty == false, even if | | we have dirty_bitmap not zero. |- ASSERT(PageDirty()); | ^^^^ CRASH The solution here is to also lock the eb->pages[0] for subpage case of set_extent_buffer_dirty(), to prevent racing with clear_extent_buffer_dirty(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-25 15:14:43 +08:00
static void btree_clear_page_dirty(struct page *page)
{
ASSERT(PageDirty(page));
ASSERT(PageLocked(page));
clear_page_dirty_for_io(page);
xa_lock_irq(&page->mapping->i_pages);
if (!PageDirty(page))
__xa_clear_mark(&page->mapping->i_pages,
page_index(page), PAGECACHE_TAG_DIRTY);
xa_unlock_irq(&page->mapping->i_pages);
}
static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
struct page *page = eb->pages[0];
bool last;
/* btree_clear_page_dirty() needs page locked */
lock_page(page);
last = btrfs_subpage_clear_and_test_dirty(fs_info, page, eb->start,
eb->len);
if (last)
btree_clear_page_dirty(page);
unlock_page(page);
WARN_ON(atomic_read(&eb->refs) == 0);
}
void clear_extent_buffer_dirty(const struct extent_buffer *eb)
{
int i;
int num_pages;
struct page *page;
if (eb->fs_info->nodesize < PAGE_SIZE)
btrfs: make set/clear_extent_buffer_dirty() subpage compatible For set_extent_buffer_dirty() to support subpage sized metadata, just call btrfs_page_set_dirty() to handle both cases. For clear_extent_buffer_dirty(), it needs to clear the page dirty if and only if all extent buffers in the page range are no longer dirty. Also do the same for page error. This is pretty different from the existing clear_extent_buffer_dirty() routine, so add a new helper function, clear_subpage_extent_buffer_dirty() to do this for subpage metadata. Also since the main part of clearing page dirty code is still the same, extract that into btree_clear_page_dirty() so that it can be utilized for both cases. But there is a special race between set_extent_buffer_dirty() and clear_extent_buffer_dirty(), where we can clear the page dirty. [POSSIBLE RACE WINDOW] For the race window between clear_subpage_extent_buffer_dirty() and set_extent_buffer_dirty(), due to the fact that we can't call clear_page_dirty_for_io() under subpage spin lock, we can race like below: T1 (eb1 in the same page) | T2 (eb2 in the same page) -------------------------------+------------------------------ set_extent_buffer_dirty() | clear_extent_buffer_dirty() |- was_dirty = false; | |- clear_subpagE_extent_buffer_dirty() | | |- btrfs_clear_and_test_dirty() | | | Since eb2 is the last dirty page | | | we got: | | | last == true; | | | |- btrfs_page_set_dirty() | | | We set the page dirty and | | | subpage dirty bitmap | | | | |- if (last) | | | Since we don't have subpage lock | | | held, now @last is no longer | | | correct | | |- btree_clear_page_dirty() | | Now PageDirty == false, even if | | we have dirty_bitmap not zero. |- ASSERT(PageDirty()); | ^^^^ CRASH The solution here is to also lock the eb->pages[0] for subpage case of set_extent_buffer_dirty(), to prevent racing with clear_extent_buffer_dirty(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-25 15:14:43 +08:00
return clear_subpage_extent_buffer_dirty(eb);
num_pages = num_extent_pages(eb);
for (i = 0; i < num_pages; i++) {
page = eb->pages[i];
if (!PageDirty(page))
continue;
lock_page(page);
btrfs: make set/clear_extent_buffer_dirty() subpage compatible For set_extent_buffer_dirty() to support subpage sized metadata, just call btrfs_page_set_dirty() to handle both cases. For clear_extent_buffer_dirty(), it needs to clear the page dirty if and only if all extent buffers in the page range are no longer dirty. Also do the same for page error. This is pretty different from the existing clear_extent_buffer_dirty() routine, so add a new helper function, clear_subpage_extent_buffer_dirty() to do this for subpage metadata. Also since the main part of clearing page dirty code is still the same, extract that into btree_clear_page_dirty() so that it can be utilized for both cases. But there is a special race between set_extent_buffer_dirty() and clear_extent_buffer_dirty(), where we can clear the page dirty. [POSSIBLE RACE WINDOW] For the race window between clear_subpage_extent_buffer_dirty() and set_extent_buffer_dirty(), due to the fact that we can't call clear_page_dirty_for_io() under subpage spin lock, we can race like below: T1 (eb1 in the same page) | T2 (eb2 in the same page) -------------------------------+------------------------------ set_extent_buffer_dirty() | clear_extent_buffer_dirty() |- was_dirty = false; | |- clear_subpagE_extent_buffer_dirty() | | |- btrfs_clear_and_test_dirty() | | | Since eb2 is the last dirty page | | | we got: | | | last == true; | | | |- btrfs_page_set_dirty() | | | We set the page dirty and | | | subpage dirty bitmap | | | | |- if (last) | | | Since we don't have subpage lock | | | held, now @last is no longer | | | correct | | |- btree_clear_page_dirty() | | Now PageDirty == false, even if | | we have dirty_bitmap not zero. |- ASSERT(PageDirty()); | ^^^^ CRASH The solution here is to also lock the eb->pages[0] for subpage case of set_extent_buffer_dirty(), to prevent racing with clear_extent_buffer_dirty(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-25 15:14:43 +08:00
btree_clear_page_dirty(page);
ClearPageError(page);
unlock_page(page);
}
WARN_ON(atomic_read(&eb->refs) == 0);
}
bool set_extent_buffer_dirty(struct extent_buffer *eb)
{
int i;
int num_pages;
bool was_dirty;
check_buffer_tree_ref(eb);
was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
num_pages = num_extent_pages(eb);
WARN_ON(atomic_read(&eb->refs) == 0);
WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
btrfs: make set/clear_extent_buffer_dirty() subpage compatible For set_extent_buffer_dirty() to support subpage sized metadata, just call btrfs_page_set_dirty() to handle both cases. For clear_extent_buffer_dirty(), it needs to clear the page dirty if and only if all extent buffers in the page range are no longer dirty. Also do the same for page error. This is pretty different from the existing clear_extent_buffer_dirty() routine, so add a new helper function, clear_subpage_extent_buffer_dirty() to do this for subpage metadata. Also since the main part of clearing page dirty code is still the same, extract that into btree_clear_page_dirty() so that it can be utilized for both cases. But there is a special race between set_extent_buffer_dirty() and clear_extent_buffer_dirty(), where we can clear the page dirty. [POSSIBLE RACE WINDOW] For the race window between clear_subpage_extent_buffer_dirty() and set_extent_buffer_dirty(), due to the fact that we can't call clear_page_dirty_for_io() under subpage spin lock, we can race like below: T1 (eb1 in the same page) | T2 (eb2 in the same page) -------------------------------+------------------------------ set_extent_buffer_dirty() | clear_extent_buffer_dirty() |- was_dirty = false; | |- clear_subpagE_extent_buffer_dirty() | | |- btrfs_clear_and_test_dirty() | | | Since eb2 is the last dirty page | | | we got: | | | last == true; | | | |- btrfs_page_set_dirty() | | | We set the page dirty and | | | subpage dirty bitmap | | | | |- if (last) | | | Since we don't have subpage lock | | | held, now @last is no longer | | | correct | | |- btree_clear_page_dirty() | | Now PageDirty == false, even if | | we have dirty_bitmap not zero. |- ASSERT(PageDirty()); | ^^^^ CRASH The solution here is to also lock the eb->pages[0] for subpage case of set_extent_buffer_dirty(), to prevent racing with clear_extent_buffer_dirty(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-25 15:14:43 +08:00
if (!was_dirty) {
bool subpage = eb->fs_info->nodesize < PAGE_SIZE;
btrfs: make set/clear_extent_buffer_dirty() subpage compatible For set_extent_buffer_dirty() to support subpage sized metadata, just call btrfs_page_set_dirty() to handle both cases. For clear_extent_buffer_dirty(), it needs to clear the page dirty if and only if all extent buffers in the page range are no longer dirty. Also do the same for page error. This is pretty different from the existing clear_extent_buffer_dirty() routine, so add a new helper function, clear_subpage_extent_buffer_dirty() to do this for subpage metadata. Also since the main part of clearing page dirty code is still the same, extract that into btree_clear_page_dirty() so that it can be utilized for both cases. But there is a special race between set_extent_buffer_dirty() and clear_extent_buffer_dirty(), where we can clear the page dirty. [POSSIBLE RACE WINDOW] For the race window between clear_subpage_extent_buffer_dirty() and set_extent_buffer_dirty(), due to the fact that we can't call clear_page_dirty_for_io() under subpage spin lock, we can race like below: T1 (eb1 in the same page) | T2 (eb2 in the same page) -------------------------------+------------------------------ set_extent_buffer_dirty() | clear_extent_buffer_dirty() |- was_dirty = false; | |- clear_subpagE_extent_buffer_dirty() | | |- btrfs_clear_and_test_dirty() | | | Since eb2 is the last dirty page | | | we got: | | | last == true; | | | |- btrfs_page_set_dirty() | | | We set the page dirty and | | | subpage dirty bitmap | | | | |- if (last) | | | Since we don't have subpage lock | | | held, now @last is no longer | | | correct | | |- btree_clear_page_dirty() | | Now PageDirty == false, even if | | we have dirty_bitmap not zero. |- ASSERT(PageDirty()); | ^^^^ CRASH The solution here is to also lock the eb->pages[0] for subpage case of set_extent_buffer_dirty(), to prevent racing with clear_extent_buffer_dirty(). Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-03-25 15:14:43 +08:00
/*
* For subpage case, we can have other extent buffers in the
* same page, and in clear_subpage_extent_buffer_dirty() we
* have to clear page dirty without subpage lock held.
* This can cause race where our page gets dirty cleared after
* we just set it.
*
* Thankfully, clear_subpage_extent_buffer_dirty() has locked
* its page for other reasons, we can use page lock to prevent
* the above race.
*/
if (subpage)
lock_page(eb->pages[0]);
for (i = 0; i < num_pages; i++)
btrfs_page_set_dirty(eb->fs_info, eb->pages[i],
eb->start, eb->len);
if (subpage)
unlock_page(eb->pages[0]);
}
#ifdef CONFIG_BTRFS_DEBUG
for (i = 0; i < num_pages; i++)
ASSERT(PageDirty(eb->pages[i]));
#endif
return was_dirty;
}
void clear_extent_buffer_uptodate(struct extent_buffer *eb)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
struct page *page;
int num_pages;
int i;
Btrfs: Change btree locking to use explicit blocking points Most of the btrfs metadata operations can be protected by a spinlock, but some operations still need to schedule. So far, btrfs has been using a mutex along with a trylock loop, most of the time it is able to avoid going for the full mutex, so the trylock loop is a big performance gain. This commit is step one for getting rid of the blocking locks entirely. btrfs_tree_lock takes a spinlock, and the code explicitly switches to a blocking lock when it starts an operation that can schedule. We'll be able get rid of the blocking locks in smaller pieces over time. Tracing allows us to find the most common cause of blocking, so we can start with the hot spots first. The basic idea is: btrfs_tree_lock() returns with the spin lock held btrfs_set_lock_blocking() sets the EXTENT_BUFFER_BLOCKING bit in the extent buffer flags, and then drops the spin lock. The buffer is still considered locked by all of the btrfs code. If btrfs_tree_lock gets the spinlock but finds the blocking bit set, it drops the spin lock and waits on a wait queue for the blocking bit to go away. Much of the code that needs to set the blocking bit finishes without actually blocking a good percentage of the time. So, an adaptive spin is still used against the blocking bit to avoid very high context switch rates. btrfs_clear_lock_blocking() clears the blocking bit and returns with the spinlock held again. btrfs_tree_unlock() can be called on either blocking or spinning locks, it does the right thing based on the blocking bit. ctree.c has a helper function to set/clear all the locked buffers in a path as blocking. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:25:08 +08:00
clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
num_pages = num_extent_pages(eb);
for (i = 0; i < num_pages; i++) {
page = eb->pages[i];
if (!page)
continue;
/*
* This is special handling for metadata subpage, as regular
* btrfs_is_subpage() can not handle cloned/dummy metadata.
*/
if (fs_info->nodesize >= PAGE_SIZE)
ClearPageUptodate(page);
else
btrfs_subpage_clear_uptodate(fs_info, page, eb->start,
eb->len);
}
}
void set_extent_buffer_uptodate(struct extent_buffer *eb)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
struct page *page;
int num_pages;
int i;
set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
num_pages = num_extent_pages(eb);
for (i = 0; i < num_pages; i++) {
page = eb->pages[i];
/*
* This is special handling for metadata subpage, as regular
* btrfs_is_subpage() can not handle cloned/dummy metadata.
*/
if (fs_info->nodesize >= PAGE_SIZE)
SetPageUptodate(page);
else
btrfs_subpage_set_uptodate(fs_info, page, eb->start,
eb->len);
}
}
static int read_extent_buffer_subpage(struct extent_buffer *eb, int wait,
int mirror_num,
struct btrfs_tree_parent_check *check)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
struct extent_io_tree *io_tree;
struct page *page = eb->pages[0];
struct extent_state *cached_state = NULL;
struct btrfs_bio_ctrl bio_ctrl = {
.mirror_num = mirror_num,
};
int ret = 0;
ASSERT(!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags));
ASSERT(PagePrivate(page));
ASSERT(check);
io_tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
if (wait == WAIT_NONE) {
if (!try_lock_extent(io_tree, eb->start, eb->start + eb->len - 1,
&cached_state))
return -EAGAIN;
} else {
ret = lock_extent(io_tree, eb->start, eb->start + eb->len - 1,
&cached_state);
if (ret < 0)
return ret;
}
ret = 0;
if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags) ||
PageUptodate(page) ||
btrfs_subpage_test_uptodate(fs_info, page, eb->start, eb->len)) {
set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
unlock_extent(io_tree, eb->start, eb->start + eb->len - 1,
&cached_state);
return ret;
}
clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
eb->read_mirror = 0;
atomic_set(&eb->io_pages, 1);
check_buffer_tree_ref(eb);
bio_ctrl.end_io_func = end_bio_extent_readpage;
btrfs_subpage_clear_error(fs_info, page, eb->start, eb->len);
btrfs: subpage: fix a rare race between metadata endio and eb freeing [BUG] There is a very rare ASSERT() triggering during full fstests run for subpage rw support. No other reproducer so far. The ASSERT() gets triggered for metadata read in btrfs_page_set_uptodate() inside end_page_read(). [CAUSE] There is still a small race window for metadata only, the race could happen like this: T1 | T2 ------------------------------------+----------------------------- end_bio_extent_readpage() | |- btrfs_validate_metadata_buffer() | | |- free_extent_buffer() | | Still have 2 refs | |- end_page_read() | |- if (unlikely(PagePrivate()) | | The page still has Private | | | free_extent_buffer() | | | Only one ref 1, will be | | | released | | |- detach_extent_buffer_page() | | |- btrfs_detach_subpage() |- btrfs_set_page_uptodate() | The page no longer has Private| >>> ASSERT() triggered <<< | This race window is super small, thus pretty hard to hit, even with so many runs of fstests. But the race window is still there, we have to go another way to solve it other than relying on random PagePrivate() check. Data path is not affected, as it will lock the page before reading, while unlocking the page after the last read has finished, thus no race window. [FIX] This patch will fix the bug by repurposing btrfs_subpage::readers. Now btrfs_subpage::readers will be a member shared by both metadata and data. For metadata path, we don't do the page unlock as metadata only relies on extent locking. At the same time, teach page_range_has_eb() to take btrfs_subpage::readers into consideration. So that even if the last eb of a page gets freed, page::private won't be detached as long as there still are pending end_page_read() calls. By this we eliminate the race window, this will slight increase the metadata memory usage, as the page may not be released as frequently as usual. But it should not be a big deal. The code got introduced in ("btrfs: submit read time repair only for each corrupted sector"), but the fix is in a separate patch to keep the problem description and the crash is rare so it should not hurt bisectability. Signed-off-by: Qu Wegruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-06-07 17:02:58 +08:00
btrfs_subpage_start_reader(fs_info, page, eb->start, eb->len);
ret = submit_extent_page(REQ_OP_READ, NULL, &bio_ctrl,
eb->start, page, eb->len,
eb->start - page_offset(page), 0, true);
if (ret) {
/*
* In the endio function, if we hit something wrong we will
* increase the io_pages, so here we need to decrease it for
* error path.
*/
atomic_dec(&eb->io_pages);
}
memcpy(&btrfs_bio(bio_ctrl.bio)->parent_check, check, sizeof(*check));
submit_one_bio(&bio_ctrl);
if (ret || wait != WAIT_COMPLETE) {
free_extent_state(cached_state);
return ret;
}
wait_extent_bit(io_tree, eb->start, eb->start + eb->len - 1,
EXTENT_LOCKED, &cached_state);
if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
ret = -EIO;
return ret;
}
int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
struct btrfs_tree_parent_check *check)
{
int i;
struct page *page;
int err;
int ret = 0;
int locked_pages = 0;
int all_uptodate = 1;
int num_pages;
unsigned long num_reads = 0;
struct btrfs_bio_ctrl bio_ctrl = {
.mirror_num = mirror_num,
};
Btrfs: Change btree locking to use explicit blocking points Most of the btrfs metadata operations can be protected by a spinlock, but some operations still need to schedule. So far, btrfs has been using a mutex along with a trylock loop, most of the time it is able to avoid going for the full mutex, so the trylock loop is a big performance gain. This commit is step one for getting rid of the blocking locks entirely. btrfs_tree_lock takes a spinlock, and the code explicitly switches to a blocking lock when it starts an operation that can schedule. We'll be able get rid of the blocking locks in smaller pieces over time. Tracing allows us to find the most common cause of blocking, so we can start with the hot spots first. The basic idea is: btrfs_tree_lock() returns with the spin lock held btrfs_set_lock_blocking() sets the EXTENT_BUFFER_BLOCKING bit in the extent buffer flags, and then drops the spin lock. The buffer is still considered locked by all of the btrfs code. If btrfs_tree_lock gets the spinlock but finds the blocking bit set, it drops the spin lock and waits on a wait queue for the blocking bit to go away. Much of the code that needs to set the blocking bit finishes without actually blocking a good percentage of the time. So, an adaptive spin is still used against the blocking bit to avoid very high context switch rates. btrfs_clear_lock_blocking() clears the blocking bit and returns with the spinlock held again. btrfs_tree_unlock() can be called on either blocking or spinning locks, it does the right thing based on the blocking bit. ctree.c has a helper function to set/clear all the locked buffers in a path as blocking. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:25:08 +08:00
if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
return 0;
btrfs: check WRITE_ERR when trying to read an extent buffer Filipe reported a hang when we have errors on btrfs. This turned out to be a side-effect of my fix c2e39305299f01 ("btrfs: clear extent buffer uptodate when we fail to write it") which made it so we clear EXTENT_BUFFER_UPTODATE on an eb when we fail to write it out. Below is a paste of Filipe's analysis he got from using drgn to debug the hang """ btree readahead code calls read_extent_buffer_pages(), sets ->io_pages to a value while writeback of all pages has not yet completed: --> writeback for the first 3 pages finishes, we clear EXTENT_BUFFER_UPTODATE from eb on the first page when we get an error. --> at this point eb->io_pages is 1 and we cleared Uptodate bit from the first 3 pages --> read_extent_buffer_pages() does not see EXTENT_BUFFER_UPTODATE() so it continues, it's able to lock the pages since we obviously don't hold the pages locked during writeback --> read_extent_buffer_pages() then computes 'num_reads' as 3, and sets eb->io_pages to 3, since only the first page does not have Uptodate bit set at this point --> writeback for the remaining page completes, we ended decrementing eb->io_pages by 1, resulting in eb->io_pages == 2, and therefore never calling end_extent_buffer_writeback(), so EXTENT_BUFFER_WRITEBACK remains in the eb's flags --> of course, when the read bio completes, it doesn't and shouldn't call end_extent_buffer_writeback() --> we should clear EXTENT_BUFFER_UPTODATE only after all pages of the eb finished writeback? or maybe make the read pages code wait for writeback of all pages of the eb to complete before checking which pages need to be read, touch ->io_pages, submit read bio, etc writeback bit never cleared means we can hang when aborting a transaction, at: btrfs_cleanup_one_transaction() btrfs_destroy_marked_extents() wait_on_extent_buffer_writeback() """ This is a problem because our writes are not synchronized with reads in any way. We clear the UPTODATE flag and then we can easily come in and try to read the EB while we're still waiting on other bio's to complete. We have two options here, we could lock all the pages, and then check to see if eb->io_pages != 0 to know if we've already got an outstanding write on the eb. Or we can simply check to see if we have WRITE_ERR set on this extent buffer. We set this bit _before_ we clear UPTODATE, so if the read gets triggered because we aren't UPTODATE because of a write error we're guaranteed to have WRITE_ERR set, and in this case we can simply return -EIO. This will fix the reported hang. Reported-by: Filipe Manana <fdmanana@suse.com> Fixes: c2e39305299f01 ("btrfs: clear extent buffer uptodate when we fail to write it") CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-14 03:22:33 +08:00
/*
* We could have had EXTENT_BUFFER_UPTODATE cleared by the write
* operation, which could potentially still be in flight. In this case
* we simply want to return an error.
*/
if (unlikely(test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)))
return -EIO;
if (eb->fs_info->nodesize < PAGE_SIZE)
return read_extent_buffer_subpage(eb, wait, mirror_num, check);
num_pages = num_extent_pages(eb);
for (i = 0; i < num_pages; i++) {
page = eb->pages[i];
if (wait == WAIT_NONE) {
/*
* WAIT_NONE is only utilized by readahead. If we can't
* acquire the lock atomically it means either the eb
* is being read out or under modification.
* Either way the eb will be or has been cached,
* readahead can exit safely.
*/
if (!trylock_page(page))
goto unlock_exit;
} else {
lock_page(page);
}
locked_pages++;
Btrfs: fix memory leak in reading btree blocks So we can read a btree block via readahead or intentional read, and we can end up with a memory leak when something happens as follows, 1) readahead starts to read block A but does not wait for read completion, 2) btree_readpage_end_io_hook finds that block A is corrupted, and it needs to clear all block A's pages' uptodate bit. 3) meanwhile an intentional read kicks in and checks block A's pages' uptodate to decide which page needs to be read. 4) when some pages have the uptodate bit during 3)'s check so 3) doesn't count them for eb->io_pages, but they are later cleared by 2) so we has to readpage on the page, we get the wrong eb->io_pages which results in a memory leak of this block. This fixes the problem by firstly getting all pages's locking and then checking pages' uptodate bit. t1(readahead) t2(readahead endio) t3(the following read) read_extent_buffer_pages end_bio_extent_readpage for pg in eb: for page 0,1,2 in eb: if pg is uptodate: btree_readpage_end_io_hook(pg) num_reads++ if uptodate: eb->io_pages = num_reads SetPageUptodate(pg) _______________ for pg in eb: for page 3 in eb: read_extent_buffer_pages if pg is NOT uptodate: btree_readpage_end_io_hook(pg) for pg in eb: __extent_read_full_page(pg) sanity check reports something wrong if pg is uptodate: clear_extent_buffer_uptodate(eb) num_reads++ for pg in eb: eb->io_pages = num_reads ClearPageUptodate(page) _______________ for pg in eb: if pg is NOT uptodate: __extent_read_full_page(pg) So t3's eb->io_pages is not consistent with the number of pages it's reading, and during endio(), atomic_dec_and_test(&eb->io_pages) will get a negative number so that we're not able to free the eb. Signed-off-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2016-08-04 03:33:01 +08:00
}
/*
* We need to firstly lock all pages to make sure that
* the uptodate bit of our pages won't be affected by
* clear_extent_buffer_uptodate().
*/
for (i = 0; i < num_pages; i++) {
Btrfs: fix memory leak in reading btree blocks So we can read a btree block via readahead or intentional read, and we can end up with a memory leak when something happens as follows, 1) readahead starts to read block A but does not wait for read completion, 2) btree_readpage_end_io_hook finds that block A is corrupted, and it needs to clear all block A's pages' uptodate bit. 3) meanwhile an intentional read kicks in and checks block A's pages' uptodate to decide which page needs to be read. 4) when some pages have the uptodate bit during 3)'s check so 3) doesn't count them for eb->io_pages, but they are later cleared by 2) so we has to readpage on the page, we get the wrong eb->io_pages which results in a memory leak of this block. This fixes the problem by firstly getting all pages's locking and then checking pages' uptodate bit. t1(readahead) t2(readahead endio) t3(the following read) read_extent_buffer_pages end_bio_extent_readpage for pg in eb: for page 0,1,2 in eb: if pg is uptodate: btree_readpage_end_io_hook(pg) num_reads++ if uptodate: eb->io_pages = num_reads SetPageUptodate(pg) _______________ for pg in eb: for page 3 in eb: read_extent_buffer_pages if pg is NOT uptodate: btree_readpage_end_io_hook(pg) for pg in eb: __extent_read_full_page(pg) sanity check reports something wrong if pg is uptodate: clear_extent_buffer_uptodate(eb) num_reads++ for pg in eb: eb->io_pages = num_reads ClearPageUptodate(page) _______________ for pg in eb: if pg is NOT uptodate: __extent_read_full_page(pg) So t3's eb->io_pages is not consistent with the number of pages it's reading, and during endio(), atomic_dec_and_test(&eb->io_pages) will get a negative number so that we're not able to free the eb. Signed-off-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2016-08-04 03:33:01 +08:00
page = eb->pages[i];
if (!PageUptodate(page)) {
num_reads++;
all_uptodate = 0;
}
}
Btrfs: fix memory leak in reading btree blocks So we can read a btree block via readahead or intentional read, and we can end up with a memory leak when something happens as follows, 1) readahead starts to read block A but does not wait for read completion, 2) btree_readpage_end_io_hook finds that block A is corrupted, and it needs to clear all block A's pages' uptodate bit. 3) meanwhile an intentional read kicks in and checks block A's pages' uptodate to decide which page needs to be read. 4) when some pages have the uptodate bit during 3)'s check so 3) doesn't count them for eb->io_pages, but they are later cleared by 2) so we has to readpage on the page, we get the wrong eb->io_pages which results in a memory leak of this block. This fixes the problem by firstly getting all pages's locking and then checking pages' uptodate bit. t1(readahead) t2(readahead endio) t3(the following read) read_extent_buffer_pages end_bio_extent_readpage for pg in eb: for page 0,1,2 in eb: if pg is uptodate: btree_readpage_end_io_hook(pg) num_reads++ if uptodate: eb->io_pages = num_reads SetPageUptodate(pg) _______________ for pg in eb: for page 3 in eb: read_extent_buffer_pages if pg is NOT uptodate: btree_readpage_end_io_hook(pg) for pg in eb: __extent_read_full_page(pg) sanity check reports something wrong if pg is uptodate: clear_extent_buffer_uptodate(eb) num_reads++ for pg in eb: eb->io_pages = num_reads ClearPageUptodate(page) _______________ for pg in eb: if pg is NOT uptodate: __extent_read_full_page(pg) So t3's eb->io_pages is not consistent with the number of pages it's reading, and during endio(), atomic_dec_and_test(&eb->io_pages) will get a negative number so that we're not able to free the eb. Signed-off-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2016-08-04 03:33:01 +08:00
if (all_uptodate) {
set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
goto unlock_exit;
}
Btrfs: be aware of btree inode write errors to avoid fs corruption While we have a transaction ongoing, the VM might decide at any time to call btree_inode->i_mapping->a_ops->writepages(), which will start writeback of dirty pages belonging to btree nodes/leafs. This call might return an error or the writeback might finish with an error before we attempt to commit the running transaction. If this happens, we might have no way of knowing that such error happened when we are committing the transaction - because the pages might no longer be marked dirty nor tagged for writeback (if a subsequent modification to the extent buffer didn't happen before the transaction commit) which makes filemap_fdata[write|wait]_range unable to find such pages (even if they're marked with SetPageError). So if this happens we must abort the transaction, otherwise we commit a super block with btree roots that point to btree nodes/leafs whose content on disk is invalid - either garbage or the content of some node/leaf from a past generation that got cowed or deleted and is no longer valid (for this later case we end up getting error messages like "parent transid verify failed on 10826481664 wanted 25748 found 29562" when reading btree nodes/leafs from disk). Note that setting and checking AS_EIO/AS_ENOSPC in the btree inode's i_mapping would not be enough because we need to distinguish between log tree extents (not fatal) vs non-log tree extents (fatal) and because the next call to filemap_fdatawait_range() will catch and clear such errors in the mapping - and that call might be from a log sync and not from a transaction commit, which means we would not know about the error at transaction commit time. Also, checking for the eb flag EXTENT_BUFFER_IOERR at transaction commit time isn't done and would not be completely reliable, as the eb might be removed from memory and read back when trying to get it, which clears that flag right before reading the eb's pages from disk, making us not know about the previous write error. Using the new 3 flags for the btree inode also makes us achieve the goal of AS_EIO/AS_ENOSPC when writepages() returns success, started writeback for all dirty pages and before filemap_fdatawait_range() is called, the writeback for all dirty pages had already finished with errors - because we were not using AS_EIO/AS_ENOSPC, filemap_fdatawait_range() would return success, as it could not know that writeback errors happened (the pages were no longer tagged for writeback). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-26 19:25:56 +08:00
clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
eb->read_mirror = 0;
atomic_set(&eb->io_pages, num_reads);
btrfs: fix fatal extent_buffer readahead vs releasepage race Under somewhat convoluted conditions, it is possible to attempt to release an extent_buffer that is under io, which triggers a BUG_ON in btrfs_release_extent_buffer_pages. This relies on a few different factors. First, extent_buffer reads done as readahead for searching use WAIT_NONE, so they free the local extent buffer reference while the io is outstanding. However, they should still be protected by TREE_REF. However, if the system is doing signficant reclaim, and simultaneously heavily accessing the extent_buffers, it is possible for releasepage to race with two concurrent readahead attempts in a way that leaves TREE_REF unset when the readahead extent buffer is released. Essentially, if two tasks race to allocate a new extent_buffer, but the winner who attempts the first io is rebuffed by a page being locked (likely by the reclaim itself) then the loser will still go ahead with issuing the readahead. The loser's call to find_extent_buffer must also race with the reclaim task reading the extent_buffer's refcount as 1 in a way that allows the reclaim to re-clear the TREE_REF checked by find_extent_buffer. The following represents an example execution demonstrating the race: CPU0 CPU1 CPU2 reada_for_search reada_for_search readahead_tree_block readahead_tree_block find_create_tree_block find_create_tree_block alloc_extent_buffer alloc_extent_buffer find_extent_buffer // not found allocates eb lock pages associate pages to eb insert eb into radix tree set TREE_REF, refs == 2 unlock pages read_extent_buffer_pages // WAIT_NONE not uptodate (brand new eb) lock_page if !trylock_page goto unlock_exit // not an error free_extent_buffer release_extent_buffer atomic_dec_and_test refs to 1 find_extent_buffer // found try_release_extent_buffer take refs_lock reads refs == 1; no io atomic_inc_not_zero refs to 2 mark_buffer_accessed check_buffer_tree_ref // not STALE, won't take refs_lock refs == 2; TREE_REF set // no action read_extent_buffer_pages // WAIT_NONE clear TREE_REF release_extent_buffer atomic_dec_and_test refs to 1 unlock_page still not uptodate (CPU1 read failed on trylock_page) locks pages set io_pages > 0 submit io return free_extent_buffer release_extent_buffer dec refs to 0 delete from radix tree btrfs_release_extent_buffer_pages BUG_ON(io_pages > 0)!!! We observe this at a very low rate in production and were also able to reproduce it in a test environment by introducing some spurious delays and by introducing probabilistic trylock_page failures. To fix it, we apply check_tree_ref at a point where it could not possibly be unset by a competing task: after io_pages has been incremented. All the codepaths that clear TREE_REF check for io, so they would not be able to clear it after this point until the io is done. Stack trace, for reference: [1417839.424739] ------------[ cut here ]------------ [1417839.435328] kernel BUG at fs/btrfs/extent_io.c:4841! [1417839.447024] invalid opcode: 0000 [#1] SMP [1417839.502972] RIP: 0010:btrfs_release_extent_buffer_pages+0x20/0x1f0 [1417839.517008] Code: ed e9 ... [1417839.558895] RSP: 0018:ffffc90020bcf798 EFLAGS: 00010202 [1417839.570816] RAX: 0000000000000002 RBX: ffff888102d6def0 RCX: 0000000000000028 [1417839.586962] RDX: 0000000000000002 RSI: ffff8887f0296482 RDI: ffff888102d6def0 [1417839.603108] RBP: ffff88885664a000 R08: 0000000000000046 R09: 0000000000000238 [1417839.619255] R10: 0000000000000028 R11: ffff88885664af68 R12: 0000000000000000 [1417839.635402] R13: 0000000000000000 R14: ffff88875f573ad0 R15: ffff888797aafd90 [1417839.651549] FS: 00007f5a844fa700(0000) GS:ffff88885f680000(0000) knlGS:0000000000000000 [1417839.669810] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [1417839.682887] CR2: 00007f7884541fe0 CR3: 000000049f609002 CR4: 00000000003606e0 [1417839.699037] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [1417839.715187] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [1417839.731320] Call Trace: [1417839.737103] release_extent_buffer+0x39/0x90 [1417839.746913] read_block_for_search.isra.38+0x2a3/0x370 [1417839.758645] btrfs_search_slot+0x260/0x9b0 [1417839.768054] btrfs_lookup_file_extent+0x4a/0x70 [1417839.778427] btrfs_get_extent+0x15f/0x830 [1417839.787665] ? submit_extent_page+0xc4/0x1c0 [1417839.797474] ? __do_readpage+0x299/0x7a0 [1417839.806515] __do_readpage+0x33b/0x7a0 [1417839.815171] ? btrfs_releasepage+0x70/0x70 [1417839.824597] extent_readpages+0x28f/0x400 [1417839.833836] read_pages+0x6a/0x1c0 [1417839.841729] ? startup_64+0x2/0x30 [1417839.849624] __do_page_cache_readahead+0x13c/0x1a0 [1417839.860590] filemap_fault+0x6c7/0x990 [1417839.869252] ? xas_load+0x8/0x80 [1417839.876756] ? xas_find+0x150/0x190 [1417839.884839] ? filemap_map_pages+0x295/0x3b0 [1417839.894652] __do_fault+0x32/0x110 [1417839.902540] __handle_mm_fault+0xacd/0x1000 [1417839.912156] handle_mm_fault+0xaa/0x1c0 [1417839.921004] __do_page_fault+0x242/0x4b0 [1417839.930044] ? page_fault+0x8/0x30 [1417839.937933] page_fault+0x1e/0x30 [1417839.945631] RIP: 0033:0x33c4bae [1417839.952927] Code: Bad RIP value. [1417839.960411] RSP: 002b:00007f5a844f7350 EFLAGS: 00010206 [1417839.972331] RAX: 000000000000006e RBX: 1614b3ff6a50398a RCX: 0000000000000000 [1417839.988477] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000002 [1417840.004626] RBP: 00007f5a844f7420 R08: 000000000000006e R09: 00007f5a94aeccb8 [1417840.020784] R10: 00007f5a844f7350 R11: 0000000000000000 R12: 00007f5a94aecc79 [1417840.036932] R13: 00007f5a94aecc78 R14: 00007f5a94aecc90 R15: 00007f5a94aecc40 CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2020-06-18 02:35:19 +08:00
/*
* It is possible for release_folio to clear the TREE_REF bit before we
btrfs: fix fatal extent_buffer readahead vs releasepage race Under somewhat convoluted conditions, it is possible to attempt to release an extent_buffer that is under io, which triggers a BUG_ON in btrfs_release_extent_buffer_pages. This relies on a few different factors. First, extent_buffer reads done as readahead for searching use WAIT_NONE, so they free the local extent buffer reference while the io is outstanding. However, they should still be protected by TREE_REF. However, if the system is doing signficant reclaim, and simultaneously heavily accessing the extent_buffers, it is possible for releasepage to race with two concurrent readahead attempts in a way that leaves TREE_REF unset when the readahead extent buffer is released. Essentially, if two tasks race to allocate a new extent_buffer, but the winner who attempts the first io is rebuffed by a page being locked (likely by the reclaim itself) then the loser will still go ahead with issuing the readahead. The loser's call to find_extent_buffer must also race with the reclaim task reading the extent_buffer's refcount as 1 in a way that allows the reclaim to re-clear the TREE_REF checked by find_extent_buffer. The following represents an example execution demonstrating the race: CPU0 CPU1 CPU2 reada_for_search reada_for_search readahead_tree_block readahead_tree_block find_create_tree_block find_create_tree_block alloc_extent_buffer alloc_extent_buffer find_extent_buffer // not found allocates eb lock pages associate pages to eb insert eb into radix tree set TREE_REF, refs == 2 unlock pages read_extent_buffer_pages // WAIT_NONE not uptodate (brand new eb) lock_page if !trylock_page goto unlock_exit // not an error free_extent_buffer release_extent_buffer atomic_dec_and_test refs to 1 find_extent_buffer // found try_release_extent_buffer take refs_lock reads refs == 1; no io atomic_inc_not_zero refs to 2 mark_buffer_accessed check_buffer_tree_ref // not STALE, won't take refs_lock refs == 2; TREE_REF set // no action read_extent_buffer_pages // WAIT_NONE clear TREE_REF release_extent_buffer atomic_dec_and_test refs to 1 unlock_page still not uptodate (CPU1 read failed on trylock_page) locks pages set io_pages > 0 submit io return free_extent_buffer release_extent_buffer dec refs to 0 delete from radix tree btrfs_release_extent_buffer_pages BUG_ON(io_pages > 0)!!! We observe this at a very low rate in production and were also able to reproduce it in a test environment by introducing some spurious delays and by introducing probabilistic trylock_page failures. To fix it, we apply check_tree_ref at a point where it could not possibly be unset by a competing task: after io_pages has been incremented. All the codepaths that clear TREE_REF check for io, so they would not be able to clear it after this point until the io is done. Stack trace, for reference: [1417839.424739] ------------[ cut here ]------------ [1417839.435328] kernel BUG at fs/btrfs/extent_io.c:4841! [1417839.447024] invalid opcode: 0000 [#1] SMP [1417839.502972] RIP: 0010:btrfs_release_extent_buffer_pages+0x20/0x1f0 [1417839.517008] Code: ed e9 ... [1417839.558895] RSP: 0018:ffffc90020bcf798 EFLAGS: 00010202 [1417839.570816] RAX: 0000000000000002 RBX: ffff888102d6def0 RCX: 0000000000000028 [1417839.586962] RDX: 0000000000000002 RSI: ffff8887f0296482 RDI: ffff888102d6def0 [1417839.603108] RBP: ffff88885664a000 R08: 0000000000000046 R09: 0000000000000238 [1417839.619255] R10: 0000000000000028 R11: ffff88885664af68 R12: 0000000000000000 [1417839.635402] R13: 0000000000000000 R14: ffff88875f573ad0 R15: ffff888797aafd90 [1417839.651549] FS: 00007f5a844fa700(0000) GS:ffff88885f680000(0000) knlGS:0000000000000000 [1417839.669810] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [1417839.682887] CR2: 00007f7884541fe0 CR3: 000000049f609002 CR4: 00000000003606e0 [1417839.699037] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [1417839.715187] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [1417839.731320] Call Trace: [1417839.737103] release_extent_buffer+0x39/0x90 [1417839.746913] read_block_for_search.isra.38+0x2a3/0x370 [1417839.758645] btrfs_search_slot+0x260/0x9b0 [1417839.768054] btrfs_lookup_file_extent+0x4a/0x70 [1417839.778427] btrfs_get_extent+0x15f/0x830 [1417839.787665] ? submit_extent_page+0xc4/0x1c0 [1417839.797474] ? __do_readpage+0x299/0x7a0 [1417839.806515] __do_readpage+0x33b/0x7a0 [1417839.815171] ? btrfs_releasepage+0x70/0x70 [1417839.824597] extent_readpages+0x28f/0x400 [1417839.833836] read_pages+0x6a/0x1c0 [1417839.841729] ? startup_64+0x2/0x30 [1417839.849624] __do_page_cache_readahead+0x13c/0x1a0 [1417839.860590] filemap_fault+0x6c7/0x990 [1417839.869252] ? xas_load+0x8/0x80 [1417839.876756] ? xas_find+0x150/0x190 [1417839.884839] ? filemap_map_pages+0x295/0x3b0 [1417839.894652] __do_fault+0x32/0x110 [1417839.902540] __handle_mm_fault+0xacd/0x1000 [1417839.912156] handle_mm_fault+0xaa/0x1c0 [1417839.921004] __do_page_fault+0x242/0x4b0 [1417839.930044] ? page_fault+0x8/0x30 [1417839.937933] page_fault+0x1e/0x30 [1417839.945631] RIP: 0033:0x33c4bae [1417839.952927] Code: Bad RIP value. [1417839.960411] RSP: 002b:00007f5a844f7350 EFLAGS: 00010206 [1417839.972331] RAX: 000000000000006e RBX: 1614b3ff6a50398a RCX: 0000000000000000 [1417839.988477] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000002 [1417840.004626] RBP: 00007f5a844f7420 R08: 000000000000006e R09: 00007f5a94aeccb8 [1417840.020784] R10: 00007f5a844f7350 R11: 0000000000000000 R12: 00007f5a94aecc79 [1417840.036932] R13: 00007f5a94aecc78 R14: 00007f5a94aecc90 R15: 00007f5a94aecc40 CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2020-06-18 02:35:19 +08:00
* set io_pages. See check_buffer_tree_ref for a more detailed comment.
*/
check_buffer_tree_ref(eb);
bio_ctrl.end_io_func = end_bio_extent_readpage;
for (i = 0; i < num_pages; i++) {
page = eb->pages[i];
if (!PageUptodate(page)) {
if (ret) {
atomic_dec(&eb->io_pages);
unlock_page(page);
continue;
}
ClearPageError(page);
err = submit_extent_page(REQ_OP_READ, NULL,
&bio_ctrl, page_offset(page), page,
PAGE_SIZE, 0, 0, false);
if (err) {
/*
* We failed to submit the bio so it's the
* caller's responsibility to perform cleanup
* i.e unlock page/set error bit.
*/
ret = err;
SetPageError(page);
unlock_page(page);
atomic_dec(&eb->io_pages);
}
} else {
unlock_page(page);
}
}
memcpy(&btrfs_bio(bio_ctrl.bio)->parent_check, check, sizeof(*check));
submit_one_bio(&bio_ctrl);
if (ret || wait != WAIT_COMPLETE)
return ret;
for (i = 0; i < num_pages; i++) {
page = eb->pages[i];
wait_on_page_locked(page);
if (!PageUptodate(page))
ret = -EIO;
}
return ret;
unlock_exit:
while (locked_pages > 0) {
locked_pages--;
page = eb->pages[locked_pages];
unlock_page(page);
}
return ret;
}
btrfs: extent_io: do extra check for extent buffer read write functions Although we have start, len check for extent buffer reader/write (e.g. read_extent_buffer()), these checks have limitations: - No overflow check Values like start = 1024 len = -1024 can still pass the basic (start + len) > eb->len check. - Checks are not consistent For read_extent_buffer() we only check (start + len) against eb->len. While for memcmp_extent_buffer() we also check start against eb->len. - Different error reporting mechanism We use WARN() in read_extent_buffer() but BUG() in memcpy_extent_buffer(). - Still modify memory if the request is obviously wrong In read_extent_buffer() even we find (start + len) > eb->len, we still call memset(dst, 0, len), which can easily cause memory access error if start + len overflows. To address above problems, this patch creates a new common function to check such access, check_eb_range(). - Add overflow check This function checks start, start + len against eb->len and overflow check. - Unified checks - Unified error reports Will call WARN() if CONFIG_BTRFS_DEBUG is configured. And also do btrfs_warn() message for non-debug build. - Exit ASAP if check fails No more possible memory corruption. - Add extra comment for @start @len used in those functions as it's sometimes confused with the logical addressing instead of a range inside the eb space Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202817 [ Inspired by above report, the report itself is already addressed ] Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> [ use check_add_overflow ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-19 14:35:47 +08:00
static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
unsigned long len)
{
btrfs_warn(eb->fs_info,
"access to eb bytenr %llu len %lu out of range start %lu len %lu",
eb->start, eb->len, start, len);
WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
return true;
}
/*
* Check if the [start, start + len) range is valid before reading/writing
* the eb.
* NOTE: @start and @len are offset inside the eb, not logical address.
*
* Caller should not touch the dst/src memory if this function returns error.
*/
static inline int check_eb_range(const struct extent_buffer *eb,
unsigned long start, unsigned long len)
{
unsigned long offset;
/* start, start + len should not go beyond eb->len nor overflow */
if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
return report_eb_range(eb, start, len);
return false;
}
void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
unsigned long start, unsigned long len)
{
size_t cur;
size_t offset;
struct page *page;
char *kaddr;
char *dst = (char *)dstv;
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
unsigned long i = get_eb_page_index(start);
btrfs: extent_io: do extra check for extent buffer read write functions Although we have start, len check for extent buffer reader/write (e.g. read_extent_buffer()), these checks have limitations: - No overflow check Values like start = 1024 len = -1024 can still pass the basic (start + len) > eb->len check. - Checks are not consistent For read_extent_buffer() we only check (start + len) against eb->len. While for memcmp_extent_buffer() we also check start against eb->len. - Different error reporting mechanism We use WARN() in read_extent_buffer() but BUG() in memcpy_extent_buffer(). - Still modify memory if the request is obviously wrong In read_extent_buffer() even we find (start + len) > eb->len, we still call memset(dst, 0, len), which can easily cause memory access error if start + len overflows. To address above problems, this patch creates a new common function to check such access, check_eb_range(). - Add overflow check This function checks start, start + len against eb->len and overflow check. - Unified checks - Unified error reports Will call WARN() if CONFIG_BTRFS_DEBUG is configured. And also do btrfs_warn() message for non-debug build. - Exit ASAP if check fails No more possible memory corruption. - Add extra comment for @start @len used in those functions as it's sometimes confused with the logical addressing instead of a range inside the eb space Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202817 [ Inspired by above report, the report itself is already addressed ] Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> [ use check_add_overflow ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-19 14:35:47 +08:00
if (check_eb_range(eb, start, len))
return;
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
offset = get_eb_offset_in_page(eb, start);
while (len > 0) {
page = eb->pages[i];
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
cur = min(len, (PAGE_SIZE - offset));
kaddr = page_address(page);
memcpy(dst, kaddr + offset, cur);
dst += cur;
len -= cur;
offset = 0;
i++;
}
}
btrfs: fix potential deadlock in the search ioctl With the conversion of the tree locks to rwsem I got the following lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 5.8.0-rc7-00165-g04ec4da5f45f-dirty #922 Not tainted ------------------------------------------------------ compsize/11122 is trying to acquire lock: ffff889fabca8768 (&mm->mmap_lock#2){++++}-{3:3}, at: __might_fault+0x3e/0x90 but task is already holding lock: ffff889fe720fe40 (btrfs-fs-00){++++}-{3:3}, at: __btrfs_tree_read_lock+0x39/0x180 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (btrfs-fs-00){++++}-{3:3}: down_write_nested+0x3b/0x70 __btrfs_tree_lock+0x24/0x120 btrfs_search_slot+0x756/0x990 btrfs_lookup_inode+0x3a/0xb4 __btrfs_update_delayed_inode+0x93/0x270 btrfs_async_run_delayed_root+0x168/0x230 btrfs_work_helper+0xd4/0x570 process_one_work+0x2ad/0x5f0 worker_thread+0x3a/0x3d0 kthread+0x133/0x150 ret_from_fork+0x1f/0x30 -> #1 (&delayed_node->mutex){+.+.}-{3:3}: __mutex_lock+0x9f/0x930 btrfs_delayed_update_inode+0x50/0x440 btrfs_update_inode+0x8a/0xf0 btrfs_dirty_inode+0x5b/0xd0 touch_atime+0xa1/0xd0 btrfs_file_mmap+0x3f/0x60 mmap_region+0x3a4/0x640 do_mmap+0x376/0x580 vm_mmap_pgoff+0xd5/0x120 ksys_mmap_pgoff+0x193/0x230 do_syscall_64+0x50/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xa9 -> #0 (&mm->mmap_lock#2){++++}-{3:3}: __lock_acquire+0x1272/0x2310 lock_acquire+0x9e/0x360 __might_fault+0x68/0x90 _copy_to_user+0x1e/0x80 copy_to_sk.isra.32+0x121/0x300 search_ioctl+0x106/0x200 btrfs_ioctl_tree_search_v2+0x7b/0xf0 btrfs_ioctl+0x106f/0x30a0 ksys_ioctl+0x83/0xc0 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x50/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xa9 other info that might help us debug this: Chain exists of: &mm->mmap_lock#2 --> &delayed_node->mutex --> btrfs-fs-00 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(btrfs-fs-00); lock(&delayed_node->mutex); lock(btrfs-fs-00); lock(&mm->mmap_lock#2); *** DEADLOCK *** 1 lock held by compsize/11122: #0: ffff889fe720fe40 (btrfs-fs-00){++++}-{3:3}, at: __btrfs_tree_read_lock+0x39/0x180 stack backtrace: CPU: 17 PID: 11122 Comm: compsize Kdump: loaded Not tainted 5.8.0-rc7-00165-g04ec4da5f45f-dirty #922 Hardware name: Quanta Tioga Pass Single Side 01-0030993006/Tioga Pass Single Side, BIOS F08_3A18 12/20/2018 Call Trace: dump_stack+0x78/0xa0 check_noncircular+0x165/0x180 __lock_acquire+0x1272/0x2310 lock_acquire+0x9e/0x360 ? __might_fault+0x3e/0x90 ? find_held_lock+0x72/0x90 __might_fault+0x68/0x90 ? __might_fault+0x3e/0x90 _copy_to_user+0x1e/0x80 copy_to_sk.isra.32+0x121/0x300 ? btrfs_search_forward+0x2a6/0x360 search_ioctl+0x106/0x200 btrfs_ioctl_tree_search_v2+0x7b/0xf0 btrfs_ioctl+0x106f/0x30a0 ? __do_sys_newfstat+0x5a/0x70 ? ksys_ioctl+0x83/0xc0 ksys_ioctl+0x83/0xc0 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x50/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xa9 The problem is we're doing a copy_to_user() while holding tree locks, which can deadlock if we have to do a page fault for the copy_to_user(). This exists even without my locking changes, so it needs to be fixed. Rework the search ioctl to do the pre-fault and then copy_to_user_nofault for the copying. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-10 23:42:27 +08:00
int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
void __user *dstv,
unsigned long start, unsigned long len)
{
size_t cur;
size_t offset;
struct page *page;
char *kaddr;
char __user *dst = (char __user *)dstv;
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
unsigned long i = get_eb_page_index(start);
int ret = 0;
WARN_ON(start > eb->len);
WARN_ON(start + len > eb->start + eb->len);
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
offset = get_eb_offset_in_page(eb, start);
while (len > 0) {
page = eb->pages[i];
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
cur = min(len, (PAGE_SIZE - offset));
kaddr = page_address(page);
btrfs: fix potential deadlock in the search ioctl With the conversion of the tree locks to rwsem I got the following lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 5.8.0-rc7-00165-g04ec4da5f45f-dirty #922 Not tainted ------------------------------------------------------ compsize/11122 is trying to acquire lock: ffff889fabca8768 (&mm->mmap_lock#2){++++}-{3:3}, at: __might_fault+0x3e/0x90 but task is already holding lock: ffff889fe720fe40 (btrfs-fs-00){++++}-{3:3}, at: __btrfs_tree_read_lock+0x39/0x180 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (btrfs-fs-00){++++}-{3:3}: down_write_nested+0x3b/0x70 __btrfs_tree_lock+0x24/0x120 btrfs_search_slot+0x756/0x990 btrfs_lookup_inode+0x3a/0xb4 __btrfs_update_delayed_inode+0x93/0x270 btrfs_async_run_delayed_root+0x168/0x230 btrfs_work_helper+0xd4/0x570 process_one_work+0x2ad/0x5f0 worker_thread+0x3a/0x3d0 kthread+0x133/0x150 ret_from_fork+0x1f/0x30 -> #1 (&delayed_node->mutex){+.+.}-{3:3}: __mutex_lock+0x9f/0x930 btrfs_delayed_update_inode+0x50/0x440 btrfs_update_inode+0x8a/0xf0 btrfs_dirty_inode+0x5b/0xd0 touch_atime+0xa1/0xd0 btrfs_file_mmap+0x3f/0x60 mmap_region+0x3a4/0x640 do_mmap+0x376/0x580 vm_mmap_pgoff+0xd5/0x120 ksys_mmap_pgoff+0x193/0x230 do_syscall_64+0x50/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xa9 -> #0 (&mm->mmap_lock#2){++++}-{3:3}: __lock_acquire+0x1272/0x2310 lock_acquire+0x9e/0x360 __might_fault+0x68/0x90 _copy_to_user+0x1e/0x80 copy_to_sk.isra.32+0x121/0x300 search_ioctl+0x106/0x200 btrfs_ioctl_tree_search_v2+0x7b/0xf0 btrfs_ioctl+0x106f/0x30a0 ksys_ioctl+0x83/0xc0 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x50/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xa9 other info that might help us debug this: Chain exists of: &mm->mmap_lock#2 --> &delayed_node->mutex --> btrfs-fs-00 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(btrfs-fs-00); lock(&delayed_node->mutex); lock(btrfs-fs-00); lock(&mm->mmap_lock#2); *** DEADLOCK *** 1 lock held by compsize/11122: #0: ffff889fe720fe40 (btrfs-fs-00){++++}-{3:3}, at: __btrfs_tree_read_lock+0x39/0x180 stack backtrace: CPU: 17 PID: 11122 Comm: compsize Kdump: loaded Not tainted 5.8.0-rc7-00165-g04ec4da5f45f-dirty #922 Hardware name: Quanta Tioga Pass Single Side 01-0030993006/Tioga Pass Single Side, BIOS F08_3A18 12/20/2018 Call Trace: dump_stack+0x78/0xa0 check_noncircular+0x165/0x180 __lock_acquire+0x1272/0x2310 lock_acquire+0x9e/0x360 ? __might_fault+0x3e/0x90 ? find_held_lock+0x72/0x90 __might_fault+0x68/0x90 ? __might_fault+0x3e/0x90 _copy_to_user+0x1e/0x80 copy_to_sk.isra.32+0x121/0x300 ? btrfs_search_forward+0x2a6/0x360 search_ioctl+0x106/0x200 btrfs_ioctl_tree_search_v2+0x7b/0xf0 btrfs_ioctl+0x106f/0x30a0 ? __do_sys_newfstat+0x5a/0x70 ? ksys_ioctl+0x83/0xc0 ksys_ioctl+0x83/0xc0 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x50/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xa9 The problem is we're doing a copy_to_user() while holding tree locks, which can deadlock if we have to do a page fault for the copy_to_user(). This exists even without my locking changes, so it needs to be fixed. Rework the search ioctl to do the pre-fault and then copy_to_user_nofault for the copying. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-10 23:42:27 +08:00
if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
ret = -EFAULT;
break;
}
dst += cur;
len -= cur;
offset = 0;
i++;
}
return ret;
}
int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
unsigned long start, unsigned long len)
{
size_t cur;
size_t offset;
struct page *page;
char *kaddr;
char *ptr = (char *)ptrv;
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
unsigned long i = get_eb_page_index(start);
int ret = 0;
btrfs: extent_io: do extra check for extent buffer read write functions Although we have start, len check for extent buffer reader/write (e.g. read_extent_buffer()), these checks have limitations: - No overflow check Values like start = 1024 len = -1024 can still pass the basic (start + len) > eb->len check. - Checks are not consistent For read_extent_buffer() we only check (start + len) against eb->len. While for memcmp_extent_buffer() we also check start against eb->len. - Different error reporting mechanism We use WARN() in read_extent_buffer() but BUG() in memcpy_extent_buffer(). - Still modify memory if the request is obviously wrong In read_extent_buffer() even we find (start + len) > eb->len, we still call memset(dst, 0, len), which can easily cause memory access error if start + len overflows. To address above problems, this patch creates a new common function to check such access, check_eb_range(). - Add overflow check This function checks start, start + len against eb->len and overflow check. - Unified checks - Unified error reports Will call WARN() if CONFIG_BTRFS_DEBUG is configured. And also do btrfs_warn() message for non-debug build. - Exit ASAP if check fails No more possible memory corruption. - Add extra comment for @start @len used in those functions as it's sometimes confused with the logical addressing instead of a range inside the eb space Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202817 [ Inspired by above report, the report itself is already addressed ] Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> [ use check_add_overflow ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-19 14:35:47 +08:00
if (check_eb_range(eb, start, len))
return -EINVAL;
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
offset = get_eb_offset_in_page(eb, start);
while (len > 0) {
page = eb->pages[i];
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
cur = min(len, (PAGE_SIZE - offset));
kaddr = page_address(page);
ret = memcmp(ptr, kaddr + offset, cur);
if (ret)
break;
ptr += cur;
len -= cur;
offset = 0;
i++;
}
return ret;
}
/*
* Check that the extent buffer is uptodate.
*
* For regular sector size == PAGE_SIZE case, check if @page is uptodate.
* For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
*/
static void assert_eb_page_uptodate(const struct extent_buffer *eb,
struct page *page)
{
struct btrfs_fs_info *fs_info = eb->fs_info;
btrfs: do not WARN_ON() if we have PageError set Whenever we do any extent buffer operations we call assert_eb_page_uptodate() to complain loudly if we're operating on an non-uptodate page. Our overnight tests caught this warning earlier this week WARNING: CPU: 1 PID: 553508 at fs/btrfs/extent_io.c:6849 assert_eb_page_uptodate+0x3f/0x50 CPU: 1 PID: 553508 Comm: kworker/u4:13 Tainted: G W 5.17.0-rc3+ #564 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014 Workqueue: btrfs-cache btrfs_work_helper RIP: 0010:assert_eb_page_uptodate+0x3f/0x50 RSP: 0018:ffffa961440a7c68 EFLAGS: 00010246 RAX: 0017ffffc0002112 RBX: ffffe6e74453f9c0 RCX: 0000000000001000 RDX: ffffe6e74467c887 RSI: ffffe6e74453f9c0 RDI: ffff8d4c5efc2fc0 RBP: 0000000000000d56 R08: ffff8d4d4a224000 R09: 0000000000000000 R10: 00015817fa9d1ef0 R11: 000000000000000c R12: 00000000000007b1 R13: ffff8d4c5efc2fc0 R14: 0000000001500000 R15: 0000000001cb1000 FS: 0000000000000000(0000) GS:ffff8d4dbbd00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007ff31d3448d8 CR3: 0000000118be8004 CR4: 0000000000370ee0 Call Trace: extent_buffer_test_bit+0x3f/0x70 free_space_test_bit+0xa6/0xc0 load_free_space_tree+0x1f6/0x470 caching_thread+0x454/0x630 ? rcu_read_lock_sched_held+0x12/0x60 ? rcu_read_lock_sched_held+0x12/0x60 ? rcu_read_lock_sched_held+0x12/0x60 ? lock_release+0x1f0/0x2d0 btrfs_work_helper+0xf2/0x3e0 ? lock_release+0x1f0/0x2d0 ? finish_task_switch.isra.0+0xf9/0x3a0 process_one_work+0x26d/0x580 ? process_one_work+0x580/0x580 worker_thread+0x55/0x3b0 ? process_one_work+0x580/0x580 kthread+0xf0/0x120 ? kthread_complete_and_exit+0x20/0x20 ret_from_fork+0x1f/0x30 This was partially fixed by c2e39305299f01 ("btrfs: clear extent buffer uptodate when we fail to write it"), however all that fix did was keep us from finding extent buffers after a failed writeout. It didn't keep us from continuing to use a buffer that we already had found. In this case we're searching the commit root to cache the block group, so we can start committing the transaction and switch the commit root and then start writing. After the switch we can look up an extent buffer that hasn't been written yet and start processing that block group. Then we fail to write that block out and clear Uptodate on the page, and then we start spewing these errors. Normally we're protected by the tree lock to a certain degree here. If we read a block we have that block read locked, and we block the writer from locking the block before we submit it for the write. However this isn't necessarily fool proof because the read could happen before we do the submit_bio and after we locked and unlocked the extent buffer. Also in this particular case we have path->skip_locking set, so that won't save us here. We'll simply get a block that was valid when we read it, but became invalid while we were using it. What we really want is to catch the case where we've "read" a block but it's not marked Uptodate. On read we ClearPageError(), so if we're !Uptodate and !Error we know we didn't do the right thing for reading the page. Fix this by checking !Uptodate && !Error, this way we will not complain if our buffer gets invalidated while we're using it, and we'll maintain the spirit of the check which is to make sure we have a fully in-cache block while we're messing with it. CC: stable@vger.kernel.org # 5.4+ Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-18 23:17:39 +08:00
/*
* If we are using the commit root we could potentially clear a page
* Uptodate while we're using the extent buffer that we've previously
* looked up. We don't want to complain in this case, as the page was
* valid before, we just didn't write it out. Instead we want to catch
* the case where we didn't actually read the block properly, which
* would have !PageUptodate && !PageError, as we clear PageError before
* reading.
*/
if (fs_info->nodesize < PAGE_SIZE) {
btrfs: do not WARN_ON() if we have PageError set Whenever we do any extent buffer operations we call assert_eb_page_uptodate() to complain loudly if we're operating on an non-uptodate page. Our overnight tests caught this warning earlier this week WARNING: CPU: 1 PID: 553508 at fs/btrfs/extent_io.c:6849 assert_eb_page_uptodate+0x3f/0x50 CPU: 1 PID: 553508 Comm: kworker/u4:13 Tainted: G W 5.17.0-rc3+ #564 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014 Workqueue: btrfs-cache btrfs_work_helper RIP: 0010:assert_eb_page_uptodate+0x3f/0x50 RSP: 0018:ffffa961440a7c68 EFLAGS: 00010246 RAX: 0017ffffc0002112 RBX: ffffe6e74453f9c0 RCX: 0000000000001000 RDX: ffffe6e74467c887 RSI: ffffe6e74453f9c0 RDI: ffff8d4c5efc2fc0 RBP: 0000000000000d56 R08: ffff8d4d4a224000 R09: 0000000000000000 R10: 00015817fa9d1ef0 R11: 000000000000000c R12: 00000000000007b1 R13: ffff8d4c5efc2fc0 R14: 0000000001500000 R15: 0000000001cb1000 FS: 0000000000000000(0000) GS:ffff8d4dbbd00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007ff31d3448d8 CR3: 0000000118be8004 CR4: 0000000000370ee0 Call Trace: extent_buffer_test_bit+0x3f/0x70 free_space_test_bit+0xa6/0xc0 load_free_space_tree+0x1f6/0x470 caching_thread+0x454/0x630 ? rcu_read_lock_sched_held+0x12/0x60 ? rcu_read_lock_sched_held+0x12/0x60 ? rcu_read_lock_sched_held+0x12/0x60 ? lock_release+0x1f0/0x2d0 btrfs_work_helper+0xf2/0x3e0 ? lock_release+0x1f0/0x2d0 ? finish_task_switch.isra.0+0xf9/0x3a0 process_one_work+0x26d/0x580 ? process_one_work+0x580/0x580 worker_thread+0x55/0x3b0 ? process_one_work+0x580/0x580 kthread+0xf0/0x120 ? kthread_complete_and_exit+0x20/0x20 ret_from_fork+0x1f/0x30 This was partially fixed by c2e39305299f01 ("btrfs: clear extent buffer uptodate when we fail to write it"), however all that fix did was keep us from finding extent buffers after a failed writeout. It didn't keep us from continuing to use a buffer that we already had found. In this case we're searching the commit root to cache the block group, so we can start committing the transaction and switch the commit root and then start writing. After the switch we can look up an extent buffer that hasn't been written yet and start processing that block group. Then we fail to write that block out and clear Uptodate on the page, and then we start spewing these errors. Normally we're protected by the tree lock to a certain degree here. If we read a block we have that block read locked, and we block the writer from locking the block before we submit it for the write. However this isn't necessarily fool proof because the read could happen before we do the submit_bio and after we locked and unlocked the extent buffer. Also in this particular case we have path->skip_locking set, so that won't save us here. We'll simply get a block that was valid when we read it, but became invalid while we were using it. What we really want is to catch the case where we've "read" a block but it's not marked Uptodate. On read we ClearPageError(), so if we're !Uptodate and !Error we know we didn't do the right thing for reading the page. Fix this by checking !Uptodate && !Error, this way we will not complain if our buffer gets invalidated while we're using it, and we'll maintain the spirit of the check which is to make sure we have a fully in-cache block while we're messing with it. CC: stable@vger.kernel.org # 5.4+ Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-18 23:17:39 +08:00
bool uptodate, error;
uptodate = btrfs_subpage_test_uptodate(fs_info, page,
eb->start, eb->len);
btrfs: do not WARN_ON() if we have PageError set Whenever we do any extent buffer operations we call assert_eb_page_uptodate() to complain loudly if we're operating on an non-uptodate page. Our overnight tests caught this warning earlier this week WARNING: CPU: 1 PID: 553508 at fs/btrfs/extent_io.c:6849 assert_eb_page_uptodate+0x3f/0x50 CPU: 1 PID: 553508 Comm: kworker/u4:13 Tainted: G W 5.17.0-rc3+ #564 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014 Workqueue: btrfs-cache btrfs_work_helper RIP: 0010:assert_eb_page_uptodate+0x3f/0x50 RSP: 0018:ffffa961440a7c68 EFLAGS: 00010246 RAX: 0017ffffc0002112 RBX: ffffe6e74453f9c0 RCX: 0000000000001000 RDX: ffffe6e74467c887 RSI: ffffe6e74453f9c0 RDI: ffff8d4c5efc2fc0 RBP: 0000000000000d56 R08: ffff8d4d4a224000 R09: 0000000000000000 R10: 00015817fa9d1ef0 R11: 000000000000000c R12: 00000000000007b1 R13: ffff8d4c5efc2fc0 R14: 0000000001500000 R15: 0000000001cb1000 FS: 0000000000000000(0000) GS:ffff8d4dbbd00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007ff31d3448d8 CR3: 0000000118be8004 CR4: 0000000000370ee0 Call Trace: extent_buffer_test_bit+0x3f/0x70 free_space_test_bit+0xa6/0xc0 load_free_space_tree+0x1f6/0x470 caching_thread+0x454/0x630 ? rcu_read_lock_sched_held+0x12/0x60 ? rcu_read_lock_sched_held+0x12/0x60 ? rcu_read_lock_sched_held+0x12/0x60 ? lock_release+0x1f0/0x2d0 btrfs_work_helper+0xf2/0x3e0 ? lock_release+0x1f0/0x2d0 ? finish_task_switch.isra.0+0xf9/0x3a0 process_one_work+0x26d/0x580 ? process_one_work+0x580/0x580 worker_thread+0x55/0x3b0 ? process_one_work+0x580/0x580 kthread+0xf0/0x120 ? kthread_complete_and_exit+0x20/0x20 ret_from_fork+0x1f/0x30 This was partially fixed by c2e39305299f01 ("btrfs: clear extent buffer uptodate when we fail to write it"), however all that fix did was keep us from finding extent buffers after a failed writeout. It didn't keep us from continuing to use a buffer that we already had found. In this case we're searching the commit root to cache the block group, so we can start committing the transaction and switch the commit root and then start writing. After the switch we can look up an extent buffer that hasn't been written yet and start processing that block group. Then we fail to write that block out and clear Uptodate on the page, and then we start spewing these errors. Normally we're protected by the tree lock to a certain degree here. If we read a block we have that block read locked, and we block the writer from locking the block before we submit it for the write. However this isn't necessarily fool proof because the read could happen before we do the submit_bio and after we locked and unlocked the extent buffer. Also in this particular case we have path->skip_locking set, so that won't save us here. We'll simply get a block that was valid when we read it, but became invalid while we were using it. What we really want is to catch the case where we've "read" a block but it's not marked Uptodate. On read we ClearPageError(), so if we're !Uptodate and !Error we know we didn't do the right thing for reading the page. Fix this by checking !Uptodate && !Error, this way we will not complain if our buffer gets invalidated while we're using it, and we'll maintain the spirit of the check which is to make sure we have a fully in-cache block while we're messing with it. CC: stable@vger.kernel.org # 5.4+ Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-18 23:17:39 +08:00
error = btrfs_subpage_test_error(fs_info, page, eb->start, eb->len);
WARN_ON(!uptodate && !error);
} else {
btrfs: do not WARN_ON() if we have PageError set Whenever we do any extent buffer operations we call assert_eb_page_uptodate() to complain loudly if we're operating on an non-uptodate page. Our overnight tests caught this warning earlier this week WARNING: CPU: 1 PID: 553508 at fs/btrfs/extent_io.c:6849 assert_eb_page_uptodate+0x3f/0x50 CPU: 1 PID: 553508 Comm: kworker/u4:13 Tainted: G W 5.17.0-rc3+ #564 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014 Workqueue: btrfs-cache btrfs_work_helper RIP: 0010:assert_eb_page_uptodate+0x3f/0x50 RSP: 0018:ffffa961440a7c68 EFLAGS: 00010246 RAX: 0017ffffc0002112 RBX: ffffe6e74453f9c0 RCX: 0000000000001000 RDX: ffffe6e74467c887 RSI: ffffe6e74453f9c0 RDI: ffff8d4c5efc2fc0 RBP: 0000000000000d56 R08: ffff8d4d4a224000 R09: 0000000000000000 R10: 00015817fa9d1ef0 R11: 000000000000000c R12: 00000000000007b1 R13: ffff8d4c5efc2fc0 R14: 0000000001500000 R15: 0000000001cb1000 FS: 0000000000000000(0000) GS:ffff8d4dbbd00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007ff31d3448d8 CR3: 0000000118be8004 CR4: 0000000000370ee0 Call Trace: extent_buffer_test_bit+0x3f/0x70 free_space_test_bit+0xa6/0xc0 load_free_space_tree+0x1f6/0x470 caching_thread+0x454/0x630 ? rcu_read_lock_sched_held+0x12/0x60 ? rcu_read_lock_sched_held+0x12/0x60 ? rcu_read_lock_sched_held+0x12/0x60 ? lock_release+0x1f0/0x2d0 btrfs_work_helper+0xf2/0x3e0 ? lock_release+0x1f0/0x2d0 ? finish_task_switch.isra.0+0xf9/0x3a0 process_one_work+0x26d/0x580 ? process_one_work+0x580/0x580 worker_thread+0x55/0x3b0 ? process_one_work+0x580/0x580 kthread+0xf0/0x120 ? kthread_complete_and_exit+0x20/0x20 ret_from_fork+0x1f/0x30 This was partially fixed by c2e39305299f01 ("btrfs: clear extent buffer uptodate when we fail to write it"), however all that fix did was keep us from finding extent buffers after a failed writeout. It didn't keep us from continuing to use a buffer that we already had found. In this case we're searching the commit root to cache the block group, so we can start committing the transaction and switch the commit root and then start writing. After the switch we can look up an extent buffer that hasn't been written yet and start processing that block group. Then we fail to write that block out and clear Uptodate on the page, and then we start spewing these errors. Normally we're protected by the tree lock to a certain degree here. If we read a block we have that block read locked, and we block the writer from locking the block before we submit it for the write. However this isn't necessarily fool proof because the read could happen before we do the submit_bio and after we locked and unlocked the extent buffer. Also in this particular case we have path->skip_locking set, so that won't save us here. We'll simply get a block that was valid when we read it, but became invalid while we were using it. What we really want is to catch the case where we've "read" a block but it's not marked Uptodate. On read we ClearPageError(), so if we're !Uptodate and !Error we know we didn't do the right thing for reading the page. Fix this by checking !Uptodate && !Error, this way we will not complain if our buffer gets invalidated while we're using it, and we'll maintain the spirit of the check which is to make sure we have a fully in-cache block while we're messing with it. CC: stable@vger.kernel.org # 5.4+ Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-18 23:17:39 +08:00
WARN_ON(!PageUptodate(page) && !PageError(page));
}
}
void write_extent_buffer_chunk_tree_uuid(const struct extent_buffer *eb,
const void *srcv)
{
char *kaddr;
assert_eb_page_uptodate(eb, eb->pages[0]);
kaddr = page_address(eb->pages[0]) +
get_eb_offset_in_page(eb, offsetof(struct btrfs_header,
chunk_tree_uuid));
memcpy(kaddr, srcv, BTRFS_FSID_SIZE);
}
void write_extent_buffer_fsid(const struct extent_buffer *eb, const void *srcv)
{
char *kaddr;
assert_eb_page_uptodate(eb, eb->pages[0]);
kaddr = page_address(eb->pages[0]) +
get_eb_offset_in_page(eb, offsetof(struct btrfs_header, fsid));
memcpy(kaddr, srcv, BTRFS_FSID_SIZE);
}
void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
unsigned long start, unsigned long len)
{
size_t cur;
size_t offset;
struct page *page;
char *kaddr;
char *src = (char *)srcv;
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
unsigned long i = get_eb_page_index(start);
WARN_ON(test_bit(EXTENT_BUFFER_NO_CHECK, &eb->bflags));
btrfs: extent_io: do extra check for extent buffer read write functions Although we have start, len check for extent buffer reader/write (e.g. read_extent_buffer()), these checks have limitations: - No overflow check Values like start = 1024 len = -1024 can still pass the basic (start + len) > eb->len check. - Checks are not consistent For read_extent_buffer() we only check (start + len) against eb->len. While for memcmp_extent_buffer() we also check start against eb->len. - Different error reporting mechanism We use WARN() in read_extent_buffer() but BUG() in memcpy_extent_buffer(). - Still modify memory if the request is obviously wrong In read_extent_buffer() even we find (start + len) > eb->len, we still call memset(dst, 0, len), which can easily cause memory access error if start + len overflows. To address above problems, this patch creates a new common function to check such access, check_eb_range(). - Add overflow check This function checks start, start + len against eb->len and overflow check. - Unified checks - Unified error reports Will call WARN() if CONFIG_BTRFS_DEBUG is configured. And also do btrfs_warn() message for non-debug build. - Exit ASAP if check fails No more possible memory corruption. - Add extra comment for @start @len used in those functions as it's sometimes confused with the logical addressing instead of a range inside the eb space Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202817 [ Inspired by above report, the report itself is already addressed ] Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> [ use check_add_overflow ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-19 14:35:47 +08:00
if (check_eb_range(eb, start, len))
return;
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
offset = get_eb_offset_in_page(eb, start);
while (len > 0) {
page = eb->pages[i];
assert_eb_page_uptodate(eb, page);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
cur = min(len, PAGE_SIZE - offset);
kaddr = page_address(page);
memcpy(kaddr + offset, src, cur);
src += cur;
len -= cur;
offset = 0;
i++;
}
}
void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
unsigned long len)
{
size_t cur;
size_t offset;
struct page *page;
char *kaddr;
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
unsigned long i = get_eb_page_index(start);
btrfs: extent_io: do extra check for extent buffer read write functions Although we have start, len check for extent buffer reader/write (e.g. read_extent_buffer()), these checks have limitations: - No overflow check Values like start = 1024 len = -1024 can still pass the basic (start + len) > eb->len check. - Checks are not consistent For read_extent_buffer() we only check (start + len) against eb->len. While for memcmp_extent_buffer() we also check start against eb->len. - Different error reporting mechanism We use WARN() in read_extent_buffer() but BUG() in memcpy_extent_buffer(). - Still modify memory if the request is obviously wrong In read_extent_buffer() even we find (start + len) > eb->len, we still call memset(dst, 0, len), which can easily cause memory access error if start + len overflows. To address above problems, this patch creates a new common function to check such access, check_eb_range(). - Add overflow check This function checks start, start + len against eb->len and overflow check. - Unified checks - Unified error reports Will call WARN() if CONFIG_BTRFS_DEBUG is configured. And also do btrfs_warn() message for non-debug build. - Exit ASAP if check fails No more possible memory corruption. - Add extra comment for @start @len used in those functions as it's sometimes confused with the logical addressing instead of a range inside the eb space Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202817 [ Inspired by above report, the report itself is already addressed ] Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> [ use check_add_overflow ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-19 14:35:47 +08:00
if (check_eb_range(eb, start, len))
return;
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
offset = get_eb_offset_in_page(eb, start);
while (len > 0) {
page = eb->pages[i];
assert_eb_page_uptodate(eb, page);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
cur = min(len, PAGE_SIZE - offset);
kaddr = page_address(page);
memset(kaddr + offset, 0, cur);
len -= cur;
offset = 0;
i++;
}
}
void copy_extent_buffer_full(const struct extent_buffer *dst,
const struct extent_buffer *src)
{
int i;
int num_pages;
ASSERT(dst->len == src->len);
if (dst->fs_info->nodesize >= PAGE_SIZE) {
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
num_pages = num_extent_pages(dst);
for (i = 0; i < num_pages; i++)
copy_page(page_address(dst->pages[i]),
page_address(src->pages[i]));
} else {
size_t src_offset = get_eb_offset_in_page(src, 0);
size_t dst_offset = get_eb_offset_in_page(dst, 0);
ASSERT(src->fs_info->nodesize < PAGE_SIZE);
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
memcpy(page_address(dst->pages[0]) + dst_offset,
page_address(src->pages[0]) + src_offset,
src->len);
}
}
void copy_extent_buffer(const struct extent_buffer *dst,
const struct extent_buffer *src,
unsigned long dst_offset, unsigned long src_offset,
unsigned long len)
{
u64 dst_len = dst->len;
size_t cur;
size_t offset;
struct page *page;
char *kaddr;
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
unsigned long i = get_eb_page_index(dst_offset);
btrfs: extent_io: do extra check for extent buffer read write functions Although we have start, len check for extent buffer reader/write (e.g. read_extent_buffer()), these checks have limitations: - No overflow check Values like start = 1024 len = -1024 can still pass the basic (start + len) > eb->len check. - Checks are not consistent For read_extent_buffer() we only check (start + len) against eb->len. While for memcmp_extent_buffer() we also check start against eb->len. - Different error reporting mechanism We use WARN() in read_extent_buffer() but BUG() in memcpy_extent_buffer(). - Still modify memory if the request is obviously wrong In read_extent_buffer() even we find (start + len) > eb->len, we still call memset(dst, 0, len), which can easily cause memory access error if start + len overflows. To address above problems, this patch creates a new common function to check such access, check_eb_range(). - Add overflow check This function checks start, start + len against eb->len and overflow check. - Unified checks - Unified error reports Will call WARN() if CONFIG_BTRFS_DEBUG is configured. And also do btrfs_warn() message for non-debug build. - Exit ASAP if check fails No more possible memory corruption. - Add extra comment for @start @len used in those functions as it's sometimes confused with the logical addressing instead of a range inside the eb space Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202817 [ Inspired by above report, the report itself is already addressed ] Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> [ use check_add_overflow ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-19 14:35:47 +08:00
if (check_eb_range(dst, dst_offset, len) ||
check_eb_range(src, src_offset, len))
return;
WARN_ON(src->len != dst_len);
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
offset = get_eb_offset_in_page(dst, dst_offset);
while (len > 0) {
page = dst->pages[i];
assert_eb_page_uptodate(dst, page);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
cur = min(len, (unsigned long)(PAGE_SIZE - offset));
kaddr = page_address(page);
read_extent_buffer(src, kaddr + offset, src_offset, cur);
src_offset += cur;
len -= cur;
offset = 0;
i++;
}
}
/*
* eb_bitmap_offset() - calculate the page and offset of the byte containing the
* given bit number
* @eb: the extent buffer
* @start: offset of the bitmap item in the extent buffer
* @nr: bit number
* @page_index: return index of the page in the extent buffer that contains the
* given bit number
* @page_offset: return offset into the page given by page_index
*
* This helper hides the ugliness of finding the byte in an extent buffer which
* contains a given bit.
*/
static inline void eb_bitmap_offset(const struct extent_buffer *eb,
unsigned long start, unsigned long nr,
unsigned long *page_index,
size_t *page_offset)
{
size_t byte_offset = BIT_BYTE(nr);
size_t offset;
/*
* The byte we want is the offset of the extent buffer + the offset of
* the bitmap item in the extent buffer + the offset of the byte in the
* bitmap item.
*/
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
offset = start + offset_in_page(eb->start) + byte_offset;
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
*page_index = offset >> PAGE_SHIFT;
*page_offset = offset_in_page(offset);
}
/*
* Determine whether a bit in a bitmap item is set.
*
* @eb: the extent buffer
* @start: offset of the bitmap item in the extent buffer
* @nr: bit number to test
*/
int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
unsigned long nr)
{
u8 *kaddr;
struct page *page;
unsigned long i;
size_t offset;
eb_bitmap_offset(eb, start, nr, &i, &offset);
page = eb->pages[i];
assert_eb_page_uptodate(eb, page);
kaddr = page_address(page);
return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
}
/*
* Set an area of a bitmap to 1.
*
* @eb: the extent buffer
* @start: offset of the bitmap item in the extent buffer
* @pos: bit number of the first bit
* @len: number of bits to set
*/
void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
unsigned long pos, unsigned long len)
{
u8 *kaddr;
struct page *page;
unsigned long i;
size_t offset;
const unsigned int size = pos + len;
int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
eb_bitmap_offset(eb, start, pos, &i, &offset);
page = eb->pages[i];
assert_eb_page_uptodate(eb, page);
kaddr = page_address(page);
while (len >= bits_to_set) {
kaddr[offset] |= mask_to_set;
len -= bits_to_set;
bits_to_set = BITS_PER_BYTE;
mask_to_set = ~0;
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
if (++offset >= PAGE_SIZE && len > 0) {
offset = 0;
page = eb->pages[++i];
assert_eb_page_uptodate(eb, page);
kaddr = page_address(page);
}
}
if (len) {
mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
kaddr[offset] |= mask_to_set;
}
}
/*
* Clear an area of a bitmap.
*
* @eb: the extent buffer
* @start: offset of the bitmap item in the extent buffer
* @pos: bit number of the first bit
* @len: number of bits to clear
*/
void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
unsigned long start, unsigned long pos,
unsigned long len)
{
u8 *kaddr;
struct page *page;
unsigned long i;
size_t offset;
const unsigned int size = pos + len;
int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
eb_bitmap_offset(eb, start, pos, &i, &offset);
page = eb->pages[i];
assert_eb_page_uptodate(eb, page);
kaddr = page_address(page);
while (len >= bits_to_clear) {
kaddr[offset] &= ~mask_to_clear;
len -= bits_to_clear;
bits_to_clear = BITS_PER_BYTE;
mask_to_clear = ~0;
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
if (++offset >= PAGE_SIZE && len > 0) {
offset = 0;
page = eb->pages[++i];
assert_eb_page_uptodate(eb, page);
kaddr = page_address(page);
}
}
if (len) {
mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
kaddr[offset] &= ~mask_to_clear;
}
}
btrfs: properly handle overlapping areas in memmove_extent_buffer Fix data corruption caused by memcpy() usage on overlapping data. I've observed it first when found out usermode linux crash on btrfs. ?all chain is the following: ------------[ cut here ]------------ WARNING: at /home/slyfox/linux-2.6/fs/btrfs/extent_io.c:3900 memcpy_extent_buffer+0x1a5/0x219() Call Trace: 6fa39a58: [<601b495e>] _raw_spin_unlock_irqrestore+0x18/0x1c 6fa39a68: [<60029ad9>] warn_slowpath_common+0x59/0x70 6fa39aa8: [<60029b05>] warn_slowpath_null+0x15/0x17 6fa39ab8: [<600efc97>] memcpy_extent_buffer+0x1a5/0x219 6fa39b48: [<600efd9f>] memmove_extent_buffer+0x94/0x208 6fa39bc8: [<600becbf>] btrfs_del_items+0x214/0x473 6fa39c78: [<600ce1b0>] btrfs_delete_one_dir_name+0x7c/0xda 6fa39cc8: [<600dad6b>] __btrfs_unlink_inode+0xad/0x25d 6fa39d08: [<600d7864>] btrfs_start_transaction+0xe/0x10 6fa39d48: [<600dc9ff>] btrfs_unlink_inode+0x1b/0x3b 6fa39d78: [<600e04bc>] btrfs_unlink+0x70/0xef 6fa39dc8: [<6007f0d0>] vfs_unlink+0x58/0xa3 6fa39df8: [<60080278>] do_unlinkat+0xd4/0x162 6fa39e48: [<600517db>] call_rcu_sched+0xe/0x10 6fa39e58: [<600452a8>] __put_cred+0x58/0x5a 6fa39e78: [<6007446c>] sys_faccessat+0x154/0x166 6fa39ed8: [<60080317>] sys_unlink+0x11/0x13 6fa39ee8: [<60016b80>] handle_syscall+0x58/0x70 6fa39f08: [<60021377>] userspace+0x2d4/0x381 6fa39fc8: [<60014507>] fork_handler+0x62/0x69 ---[ end trace 70b0ca2ef0266b93 ]--- http://www.mail-archive.com/linux-btrfs@vger.kernel.org/msg09302.html Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org> Reviewed-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2011-04-12 05:52:52 +08:00
static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
{
unsigned long distance = (src > dst) ? src - dst : dst - src;
return distance < len;
}
static void copy_pages(struct page *dst_page, struct page *src_page,
unsigned long dst_off, unsigned long src_off,
unsigned long len)
{
char *dst_kaddr = page_address(dst_page);
char *src_kaddr;
int must_memmove = 0;
btrfs: properly handle overlapping areas in memmove_extent_buffer Fix data corruption caused by memcpy() usage on overlapping data. I've observed it first when found out usermode linux crash on btrfs. ?all chain is the following: ------------[ cut here ]------------ WARNING: at /home/slyfox/linux-2.6/fs/btrfs/extent_io.c:3900 memcpy_extent_buffer+0x1a5/0x219() Call Trace: 6fa39a58: [<601b495e>] _raw_spin_unlock_irqrestore+0x18/0x1c 6fa39a68: [<60029ad9>] warn_slowpath_common+0x59/0x70 6fa39aa8: [<60029b05>] warn_slowpath_null+0x15/0x17 6fa39ab8: [<600efc97>] memcpy_extent_buffer+0x1a5/0x219 6fa39b48: [<600efd9f>] memmove_extent_buffer+0x94/0x208 6fa39bc8: [<600becbf>] btrfs_del_items+0x214/0x473 6fa39c78: [<600ce1b0>] btrfs_delete_one_dir_name+0x7c/0xda 6fa39cc8: [<600dad6b>] __btrfs_unlink_inode+0xad/0x25d 6fa39d08: [<600d7864>] btrfs_start_transaction+0xe/0x10 6fa39d48: [<600dc9ff>] btrfs_unlink_inode+0x1b/0x3b 6fa39d78: [<600e04bc>] btrfs_unlink+0x70/0xef 6fa39dc8: [<6007f0d0>] vfs_unlink+0x58/0xa3 6fa39df8: [<60080278>] do_unlinkat+0xd4/0x162 6fa39e48: [<600517db>] call_rcu_sched+0xe/0x10 6fa39e58: [<600452a8>] __put_cred+0x58/0x5a 6fa39e78: [<6007446c>] sys_faccessat+0x154/0x166 6fa39ed8: [<60080317>] sys_unlink+0x11/0x13 6fa39ee8: [<60016b80>] handle_syscall+0x58/0x70 6fa39f08: [<60021377>] userspace+0x2d4/0x381 6fa39fc8: [<60014507>] fork_handler+0x62/0x69 ---[ end trace 70b0ca2ef0266b93 ]--- http://www.mail-archive.com/linux-btrfs@vger.kernel.org/msg09302.html Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org> Reviewed-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2011-04-12 05:52:52 +08:00
if (dst_page != src_page) {
src_kaddr = page_address(src_page);
btrfs: properly handle overlapping areas in memmove_extent_buffer Fix data corruption caused by memcpy() usage on overlapping data. I've observed it first when found out usermode linux crash on btrfs. ?all chain is the following: ------------[ cut here ]------------ WARNING: at /home/slyfox/linux-2.6/fs/btrfs/extent_io.c:3900 memcpy_extent_buffer+0x1a5/0x219() Call Trace: 6fa39a58: [<601b495e>] _raw_spin_unlock_irqrestore+0x18/0x1c 6fa39a68: [<60029ad9>] warn_slowpath_common+0x59/0x70 6fa39aa8: [<60029b05>] warn_slowpath_null+0x15/0x17 6fa39ab8: [<600efc97>] memcpy_extent_buffer+0x1a5/0x219 6fa39b48: [<600efd9f>] memmove_extent_buffer+0x94/0x208 6fa39bc8: [<600becbf>] btrfs_del_items+0x214/0x473 6fa39c78: [<600ce1b0>] btrfs_delete_one_dir_name+0x7c/0xda 6fa39cc8: [<600dad6b>] __btrfs_unlink_inode+0xad/0x25d 6fa39d08: [<600d7864>] btrfs_start_transaction+0xe/0x10 6fa39d48: [<600dc9ff>] btrfs_unlink_inode+0x1b/0x3b 6fa39d78: [<600e04bc>] btrfs_unlink+0x70/0xef 6fa39dc8: [<6007f0d0>] vfs_unlink+0x58/0xa3 6fa39df8: [<60080278>] do_unlinkat+0xd4/0x162 6fa39e48: [<600517db>] call_rcu_sched+0xe/0x10 6fa39e58: [<600452a8>] __put_cred+0x58/0x5a 6fa39e78: [<6007446c>] sys_faccessat+0x154/0x166 6fa39ed8: [<60080317>] sys_unlink+0x11/0x13 6fa39ee8: [<60016b80>] handle_syscall+0x58/0x70 6fa39f08: [<60021377>] userspace+0x2d4/0x381 6fa39fc8: [<60014507>] fork_handler+0x62/0x69 ---[ end trace 70b0ca2ef0266b93 ]--- http://www.mail-archive.com/linux-btrfs@vger.kernel.org/msg09302.html Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org> Reviewed-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2011-04-12 05:52:52 +08:00
} else {
src_kaddr = dst_kaddr;
if (areas_overlap(src_off, dst_off, len))
must_memmove = 1;
btrfs: properly handle overlapping areas in memmove_extent_buffer Fix data corruption caused by memcpy() usage on overlapping data. I've observed it first when found out usermode linux crash on btrfs. ?all chain is the following: ------------[ cut here ]------------ WARNING: at /home/slyfox/linux-2.6/fs/btrfs/extent_io.c:3900 memcpy_extent_buffer+0x1a5/0x219() Call Trace: 6fa39a58: [<601b495e>] _raw_spin_unlock_irqrestore+0x18/0x1c 6fa39a68: [<60029ad9>] warn_slowpath_common+0x59/0x70 6fa39aa8: [<60029b05>] warn_slowpath_null+0x15/0x17 6fa39ab8: [<600efc97>] memcpy_extent_buffer+0x1a5/0x219 6fa39b48: [<600efd9f>] memmove_extent_buffer+0x94/0x208 6fa39bc8: [<600becbf>] btrfs_del_items+0x214/0x473 6fa39c78: [<600ce1b0>] btrfs_delete_one_dir_name+0x7c/0xda 6fa39cc8: [<600dad6b>] __btrfs_unlink_inode+0xad/0x25d 6fa39d08: [<600d7864>] btrfs_start_transaction+0xe/0x10 6fa39d48: [<600dc9ff>] btrfs_unlink_inode+0x1b/0x3b 6fa39d78: [<600e04bc>] btrfs_unlink+0x70/0xef 6fa39dc8: [<6007f0d0>] vfs_unlink+0x58/0xa3 6fa39df8: [<60080278>] do_unlinkat+0xd4/0x162 6fa39e48: [<600517db>] call_rcu_sched+0xe/0x10 6fa39e58: [<600452a8>] __put_cred+0x58/0x5a 6fa39e78: [<6007446c>] sys_faccessat+0x154/0x166 6fa39ed8: [<60080317>] sys_unlink+0x11/0x13 6fa39ee8: [<60016b80>] handle_syscall+0x58/0x70 6fa39f08: [<60021377>] userspace+0x2d4/0x381 6fa39fc8: [<60014507>] fork_handler+0x62/0x69 ---[ end trace 70b0ca2ef0266b93 ]--- http://www.mail-archive.com/linux-btrfs@vger.kernel.org/msg09302.html Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org> Reviewed-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2011-04-12 05:52:52 +08:00
}
if (must_memmove)
memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
else
memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
}
void memcpy_extent_buffer(const struct extent_buffer *dst,
unsigned long dst_offset, unsigned long src_offset,
unsigned long len)
{
size_t cur;
size_t dst_off_in_page;
size_t src_off_in_page;
unsigned long dst_i;
unsigned long src_i;
btrfs: extent_io: do extra check for extent buffer read write functions Although we have start, len check for extent buffer reader/write (e.g. read_extent_buffer()), these checks have limitations: - No overflow check Values like start = 1024 len = -1024 can still pass the basic (start + len) > eb->len check. - Checks are not consistent For read_extent_buffer() we only check (start + len) against eb->len. While for memcmp_extent_buffer() we also check start against eb->len. - Different error reporting mechanism We use WARN() in read_extent_buffer() but BUG() in memcpy_extent_buffer(). - Still modify memory if the request is obviously wrong In read_extent_buffer() even we find (start + len) > eb->len, we still call memset(dst, 0, len), which can easily cause memory access error if start + len overflows. To address above problems, this patch creates a new common function to check such access, check_eb_range(). - Add overflow check This function checks start, start + len against eb->len and overflow check. - Unified checks - Unified error reports Will call WARN() if CONFIG_BTRFS_DEBUG is configured. And also do btrfs_warn() message for non-debug build. - Exit ASAP if check fails No more possible memory corruption. - Add extra comment for @start @len used in those functions as it's sometimes confused with the logical addressing instead of a range inside the eb space Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202817 [ Inspired by above report, the report itself is already addressed ] Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> [ use check_add_overflow ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-19 14:35:47 +08:00
if (check_eb_range(dst, dst_offset, len) ||
check_eb_range(dst, src_offset, len))
return;
while (len > 0) {
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
dst_off_in_page = get_eb_offset_in_page(dst, dst_offset);
src_off_in_page = get_eb_offset_in_page(dst, src_offset);
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
dst_i = get_eb_page_index(dst_offset);
src_i = get_eb_page_index(src_offset);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
cur = min(len, (unsigned long)(PAGE_SIZE -
src_off_in_page));
cur = min_t(unsigned long, cur,
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 20:29:47 +08:00
(unsigned long)(PAGE_SIZE - dst_off_in_page));
copy_pages(dst->pages[dst_i], dst->pages[src_i],
dst_off_in_page, src_off_in_page, cur);
src_offset += cur;
dst_offset += cur;
len -= cur;
}
}
void memmove_extent_buffer(const struct extent_buffer *dst,
unsigned long dst_offset, unsigned long src_offset,
unsigned long len)
{
size_t cur;
size_t dst_off_in_page;
size_t src_off_in_page;
unsigned long dst_end = dst_offset + len - 1;
unsigned long src_end = src_offset + len - 1;
unsigned long dst_i;
unsigned long src_i;
btrfs: extent_io: do extra check for extent buffer read write functions Although we have start, len check for extent buffer reader/write (e.g. read_extent_buffer()), these checks have limitations: - No overflow check Values like start = 1024 len = -1024 can still pass the basic (start + len) > eb->len check. - Checks are not consistent For read_extent_buffer() we only check (start + len) against eb->len. While for memcmp_extent_buffer() we also check start against eb->len. - Different error reporting mechanism We use WARN() in read_extent_buffer() but BUG() in memcpy_extent_buffer(). - Still modify memory if the request is obviously wrong In read_extent_buffer() even we find (start + len) > eb->len, we still call memset(dst, 0, len), which can easily cause memory access error if start + len overflows. To address above problems, this patch creates a new common function to check such access, check_eb_range(). - Add overflow check This function checks start, start + len against eb->len and overflow check. - Unified checks - Unified error reports Will call WARN() if CONFIG_BTRFS_DEBUG is configured. And also do btrfs_warn() message for non-debug build. - Exit ASAP if check fails No more possible memory corruption. - Add extra comment for @start @len used in those functions as it's sometimes confused with the logical addressing instead of a range inside the eb space Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202817 [ Inspired by above report, the report itself is already addressed ] Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> [ use check_add_overflow ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-19 14:35:47 +08:00
if (check_eb_range(dst, dst_offset, len) ||
check_eb_range(dst, src_offset, len))
return;
if (dst_offset < src_offset) {
memcpy_extent_buffer(dst, dst_offset, src_offset, len);
return;
}
while (len > 0) {
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
dst_i = get_eb_page_index(dst_end);
src_i = get_eb_page_index(src_end);
btrfs: handle sectorsize < PAGE_SIZE case for extent buffer accessors To support sectorsize < PAGE_SIZE case, we need to take extra care of extent buffer accessors. Since sectorsize is smaller than PAGE_SIZE, one page can contain multiple tree blocks, we must use eb->start to determine the real offset to read/write for extent buffer accessors. This patch introduces two helpers to do this: - get_eb_page_index() This is to calculate the index to access extent_buffer::pages. It's just a simple wrapper around "start >> PAGE_SHIFT". For sectorsize == PAGE_SIZE case, nothing is changed. For sectorsize < PAGE_SIZE case, we always get index as 0, and the existing page shift also works. - get_eb_offset_in_page() This is to calculate the offset to access extent_buffer::pages. This needs to take extent_buffer::start into consideration. For sectorsize == PAGE_SIZE case, extent_buffer::start is always aligned to PAGE_SIZE, thus adding extent_buffer::start to offset_in_page() won't change the result. For sectorsize < PAGE_SIZE case, adding extent_buffer::start gives us the correct offset to access. This patch will touch the following parts to cover all extent buffer accessors: - BTRFS_SETGET_HEADER_FUNCS() - read_extent_buffer() - read_extent_buffer_to_user() - memcmp_extent_buffer() - write_extent_buffer_chunk_tree_uuid() - write_extent_buffer_fsid() - write_extent_buffer() - memzero_extent_buffer() - copy_extent_buffer_full() - copy_extent_buffer() - memcpy_extent_buffer() - memmove_extent_buffer() - btrfs_get_token_##bits() - btrfs_get_##bits() - btrfs_set_token_##bits() - btrfs_set_##bits() - generic_bin_search() Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-02 14:48:04 +08:00
dst_off_in_page = get_eb_offset_in_page(dst, dst_end);
src_off_in_page = get_eb_offset_in_page(dst, src_end);
cur = min_t(unsigned long, len, src_off_in_page + 1);
cur = min(cur, dst_off_in_page + 1);
copy_pages(dst->pages[dst_i], dst->pages[src_i],
dst_off_in_page - cur + 1,
src_off_in_page - cur + 1, cur);
dst_end -= cur;
src_end -= cur;
len -= cur;
}
}
#define GANG_LOOKUP_SIZE 16
static struct extent_buffer *get_next_extent_buffer(
struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
{
struct extent_buffer *gang[GANG_LOOKUP_SIZE];
struct extent_buffer *found = NULL;
u64 page_start = page_offset(page);
u64 cur = page_start;
ASSERT(in_range(bytenr, page_start, PAGE_SIZE));
lockdep_assert_held(&fs_info->buffer_lock);
while (cur < page_start + PAGE_SIZE) {
int ret;
int i;
ret = radix_tree_gang_lookup(&fs_info->buffer_radix,
(void **)gang, cur >> fs_info->sectorsize_bits,
min_t(unsigned int, GANG_LOOKUP_SIZE,
PAGE_SIZE / fs_info->nodesize));
if (ret == 0)
goto out;
for (i = 0; i < ret; i++) {
/* Already beyond page end */
if (gang[i]->start >= page_start + PAGE_SIZE)
goto out;
/* Found one */
if (gang[i]->start >= bytenr) {
found = gang[i];
goto out;
}
}
cur = gang[ret - 1]->start + gang[ret - 1]->len;
}
out:
return found;
}
static int try_release_subpage_extent_buffer(struct page *page)
{
struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
u64 cur = page_offset(page);
const u64 end = page_offset(page) + PAGE_SIZE;
int ret;
while (cur < end) {
struct extent_buffer *eb = NULL;
/*
* Unlike try_release_extent_buffer() which uses page->private
* to grab buffer, for subpage case we rely on radix tree, thus
* we need to ensure radix tree consistency.
*
* We also want an atomic snapshot of the radix tree, thus go
* with spinlock rather than RCU.
*/
spin_lock(&fs_info->buffer_lock);
eb = get_next_extent_buffer(fs_info, page, cur);
if (!eb) {
/* No more eb in the page range after or at cur */
spin_unlock(&fs_info->buffer_lock);
break;
}
cur = eb->start + eb->len;
/*
* The same as try_release_extent_buffer(), to ensure the eb
* won't disappear out from under us.
*/
spin_lock(&eb->refs_lock);
if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
spin_unlock(&eb->refs_lock);
spin_unlock(&fs_info->buffer_lock);
break;
}
spin_unlock(&fs_info->buffer_lock);
/*
* If tree ref isn't set then we know the ref on this eb is a
* real ref, so just return, this eb will likely be freed soon
* anyway.
*/
if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
spin_unlock(&eb->refs_lock);
break;
}
/*
* Here we don't care about the return value, we will always
* check the page private at the end. And
* release_extent_buffer() will release the refs_lock.
*/
release_extent_buffer(eb);
}
/*
* Finally to check if we have cleared page private, as if we have
* released all ebs in the page, the page private should be cleared now.
*/
spin_lock(&page->mapping->private_lock);
if (!PagePrivate(page))
ret = 1;
else
ret = 0;
spin_unlock(&page->mapping->private_lock);
return ret;
}
int try_release_extent_buffer(struct page *page)
{
struct extent_buffer *eb;
if (btrfs_sb(page->mapping->host->i_sb)->nodesize < PAGE_SIZE)
return try_release_subpage_extent_buffer(page);
/*
* We need to make sure nobody is changing page->private, as we rely on
* page->private as the pointer to extent buffer.
*/
spin_lock(&page->mapping->private_lock);
if (!PagePrivate(page)) {
spin_unlock(&page->mapping->private_lock);
return 1;
}
eb = (struct extent_buffer *)page->private;
BUG_ON(!eb);
/*
* This is a little awful but should be ok, we need to make sure that
* the eb doesn't disappear out from under us while we're looking at
* this page.
*/
spin_lock(&eb->refs_lock);
if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
spin_unlock(&eb->refs_lock);
spin_unlock(&page->mapping->private_lock);
return 0;
}
spin_unlock(&page->mapping->private_lock);
/*
* If tree ref isn't set then we know the ref on this eb is a real ref,
* so just return, this page will likely be freed soon anyway.
*/
if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
spin_unlock(&eb->refs_lock);
return 0;
}
return release_extent_buffer(eb);
}
/*
* btrfs_readahead_tree_block - attempt to readahead a child block
* @fs_info: the fs_info
* @bytenr: bytenr to read
* @owner_root: objectid of the root that owns this eb
* @gen: generation for the uptodate check, can be 0
* @level: level for the eb
*
* Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
* normal uptodate check of the eb, without checking the generation. If we have
* to read the block we will not block on anything.
*/
void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
u64 bytenr, u64 owner_root, u64 gen, int level)
{
struct btrfs_tree_parent_check check = {
.has_first_key = 0,
.level = level,
.transid = gen
};
struct extent_buffer *eb;
int ret;
eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
if (IS_ERR(eb))
return;
if (btrfs_buffer_uptodate(eb, gen, 1)) {
free_extent_buffer(eb);
return;
}
ret = read_extent_buffer_pages(eb, WAIT_NONE, 0, &check);
if (ret < 0)
free_extent_buffer_stale(eb);
else
free_extent_buffer(eb);
}
/*
* btrfs_readahead_node_child - readahead a node's child block
* @node: parent node we're reading from
* @slot: slot in the parent node for the child we want to read
*
* A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
* the slot in the node provided.
*/
void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
{
btrfs_readahead_tree_block(node->fs_info,
btrfs_node_blockptr(node, slot),
btrfs_header_owner(node),
btrfs_node_ptr_generation(node, slot),
btrfs_header_level(node) - 1);
}