2
0
mirror of https://github.com/edk2-porting/linux-next.git synced 2024-12-16 09:13:55 +08:00
linux-next/fs/f2fs/node.c

2222 lines
53 KiB
C
Raw Normal View History

/*
* fs/f2fs/node.c
*
* Copyright (c) 2012 Samsung Electronics Co., Ltd.
* http://www.samsung.com/
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 as
* published by the Free Software Foundation.
*/
#include <linux/fs.h>
#include <linux/f2fs_fs.h>
#include <linux/mpage.h>
#include <linux/backing-dev.h>
#include <linux/blkdev.h>
#include <linux/pagevec.h>
#include <linux/swap.h>
#include "f2fs.h"
#include "node.h"
#include "segment.h"
#include "trace.h"
#include <trace/events/f2fs.h>
#define on_build_free_nids(nmi) mutex_is_locked(&nm_i->build_lock)
static struct kmem_cache *nat_entry_slab;
static struct kmem_cache *free_nid_slab;
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
static struct kmem_cache *nat_entry_set_slab;
bool available_free_memory(struct f2fs_sb_info *sbi, int type)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct sysinfo val;
unsigned long avail_ram;
unsigned long mem_size = 0;
bool res = false;
si_meminfo(&val);
/* only uses low memory */
avail_ram = val.totalram - val.totalhigh;
/*
* give 25%, 25%, 50%, 50%, 50% memory for each components respectively
*/
if (type == FREE_NIDS) {
mem_size = (nm_i->fcnt * sizeof(struct free_nid)) >>
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_SHIFT;
res = mem_size < ((avail_ram * nm_i->ram_thresh / 100) >> 2);
} else if (type == NAT_ENTRIES) {
mem_size = (nm_i->nat_cnt * sizeof(struct nat_entry)) >>
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_SHIFT;
res = mem_size < ((avail_ram * nm_i->ram_thresh / 100) >> 2);
} else if (type == DIRTY_DENTS) {
if (sbi->sb->s_bdi->wb.dirty_exceeded)
return false;
mem_size = get_pages(sbi, F2FS_DIRTY_DENTS);
res = mem_size < ((avail_ram * nm_i->ram_thresh / 100) >> 1);
} else if (type == INO_ENTRIES) {
int i;
for (i = 0; i <= UPDATE_INO; i++)
mem_size += (sbi->im[i].ino_num *
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
sizeof(struct ino_entry)) >> PAGE_SHIFT;
res = mem_size < ((avail_ram * nm_i->ram_thresh / 100) >> 1);
} else if (type == EXTENT_CACHE) {
mem_size = (atomic_read(&sbi->total_ext_tree) *
sizeof(struct extent_tree) +
atomic_read(&sbi->total_ext_node) *
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
sizeof(struct extent_node)) >> PAGE_SHIFT;
res = mem_size < ((avail_ram * nm_i->ram_thresh / 100) >> 1);
} else {
if (!sbi->sb->s_bdi->wb.dirty_exceeded)
return true;
}
return res;
}
static void clear_node_page_dirty(struct page *page)
{
struct address_space *mapping = page->mapping;
unsigned int long flags;
if (PageDirty(page)) {
spin_lock_irqsave(&mapping->tree_lock, flags);
radix_tree_tag_clear(&mapping->page_tree,
page_index(page),
PAGECACHE_TAG_DIRTY);
spin_unlock_irqrestore(&mapping->tree_lock, flags);
clear_page_dirty_for_io(page);
dec_page_count(F2FS_M_SB(mapping), F2FS_DIRTY_NODES);
}
ClearPageUptodate(page);
}
static struct page *get_current_nat_page(struct f2fs_sb_info *sbi, nid_t nid)
{
pgoff_t index = current_nat_addr(sbi, nid);
return get_meta_page(sbi, index);
}
static struct page *get_next_nat_page(struct f2fs_sb_info *sbi, nid_t nid)
{
struct page *src_page;
struct page *dst_page;
pgoff_t src_off;
pgoff_t dst_off;
void *src_addr;
void *dst_addr;
struct f2fs_nm_info *nm_i = NM_I(sbi);
src_off = current_nat_addr(sbi, nid);
dst_off = next_nat_addr(sbi, src_off);
/* get current nat block page with lock */
src_page = get_meta_page(sbi, src_off);
dst_page = grab_meta_page(sbi, dst_off);
f2fs_bug_on(sbi, PageDirty(src_page));
src_addr = page_address(src_page);
dst_addr = page_address(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
memcpy(dst_addr, src_addr, PAGE_SIZE);
set_page_dirty(dst_page);
f2fs_put_page(src_page, 1);
set_to_next_nat(nm_i, nid);
return dst_page;
}
static struct nat_entry *__lookup_nat_cache(struct f2fs_nm_info *nm_i, nid_t n)
{
return radix_tree_lookup(&nm_i->nat_root, n);
}
static unsigned int __gang_lookup_nat_cache(struct f2fs_nm_info *nm_i,
nid_t start, unsigned int nr, struct nat_entry **ep)
{
return radix_tree_gang_lookup(&nm_i->nat_root, (void **)ep, start, nr);
}
static void __del_from_nat_cache(struct f2fs_nm_info *nm_i, struct nat_entry *e)
{
list_del(&e->list);
radix_tree_delete(&nm_i->nat_root, nat_get_nid(e));
nm_i->nat_cnt--;
kmem_cache_free(nat_entry_slab, e);
}
static void __set_nat_cache_dirty(struct f2fs_nm_info *nm_i,
struct nat_entry *ne)
{
nid_t set = NAT_BLOCK_OFFSET(ne->ni.nid);
struct nat_entry_set *head;
if (get_nat_flag(ne, IS_DIRTY))
return;
head = radix_tree_lookup(&nm_i->nat_set_root, set);
if (!head) {
head = f2fs_kmem_cache_alloc(nat_entry_set_slab, GFP_NOFS);
INIT_LIST_HEAD(&head->entry_list);
INIT_LIST_HEAD(&head->set_list);
head->set = set;
head->entry_cnt = 0;
f2fs_radix_tree_insert(&nm_i->nat_set_root, set, head);
}
list_move_tail(&ne->list, &head->entry_list);
nm_i->dirty_nat_cnt++;
head->entry_cnt++;
set_nat_flag(ne, IS_DIRTY, true);
}
static void __clear_nat_cache_dirty(struct f2fs_nm_info *nm_i,
struct nat_entry *ne)
{
nid_t set = NAT_BLOCK_OFFSET(ne->ni.nid);
struct nat_entry_set *head;
head = radix_tree_lookup(&nm_i->nat_set_root, set);
if (head) {
list_move_tail(&ne->list, &nm_i->nat_entries);
set_nat_flag(ne, IS_DIRTY, false);
head->entry_cnt--;
nm_i->dirty_nat_cnt--;
}
}
static unsigned int __gang_lookup_nat_set(struct f2fs_nm_info *nm_i,
nid_t start, unsigned int nr, struct nat_entry_set **ep)
{
return radix_tree_gang_lookup(&nm_i->nat_set_root, (void **)ep,
start, nr);
}
int need_dentry_mark(struct f2fs_sb_info *sbi, nid_t nid)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct nat_entry *e;
bool need = false;
down_read(&nm_i->nat_tree_lock);
e = __lookup_nat_cache(nm_i, nid);
if (e) {
if (!get_nat_flag(e, IS_CHECKPOINTED) &&
!get_nat_flag(e, HAS_FSYNCED_INODE))
need = true;
}
up_read(&nm_i->nat_tree_lock);
return need;
}
bool is_checkpointed_node(struct f2fs_sb_info *sbi, nid_t nid)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct nat_entry *e;
bool is_cp = true;
down_read(&nm_i->nat_tree_lock);
e = __lookup_nat_cache(nm_i, nid);
if (e && !get_nat_flag(e, IS_CHECKPOINTED))
is_cp = false;
up_read(&nm_i->nat_tree_lock);
return is_cp;
}
f2fs: fix conditions to remain recovery information in f2fs_sync_file This patch revisited whole the recovery information during the f2fs_sync_file. In this patch, there are three information to make a decision. a) IS_CHECKPOINTED, /* is it checkpointed before? */ b) HAS_FSYNCED_INODE, /* is the inode fsynced before? */ c) HAS_LAST_FSYNC, /* has the latest node fsync mark? */ And, the scenarios for our rule are based on: [Term] F: fsync_mark, D: dentry_mark 1. inode(x) | CP | inode(x) | dnode(F) 2. inode(x) | CP | inode(F) | dnode(F) 3. inode(x) | CP | dnode(F) | inode(x) | inode(F) 4. inode(x) | CP | dnode(F) | inode(F) 5. CP | inode(x) | dnode(F) | inode(DF) 6. CP | inode(DF) | dnode(F) 7. CP | dnode(F) | inode(DF) 8. CP | dnode(F) | inode(x) | inode(DF) For example, #3, the three conditions should be changed as follows. inode(x) | CP | dnode(F) | inode(x) | inode(F) a) x o o o o b) x x x x o c) x o o x o If f2fs_sync_file stops ------^, it should write inode(F) --------------^ So, the need_inode_block_update should return true, since c) get_nat_flag(e, HAS_LAST_FSYNC), is false. For example, #8, CP | alloc | dnode(F) | inode(x) | inode(DF) a) o x x x x b) x x x o c) o o x o If f2fs_sync_file stops -------^, it should write inode(DF) --------------^ Note that, the roll-forward policy should follow this rule, which means, if there are any missing blocks, we doesn't need to recover that inode. Signed-off-by: Huang Ying <ying.huang@intel.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-16 05:50:48 +08:00
bool need_inode_block_update(struct f2fs_sb_info *sbi, nid_t ino)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct nat_entry *e;
f2fs: fix conditions to remain recovery information in f2fs_sync_file This patch revisited whole the recovery information during the f2fs_sync_file. In this patch, there are three information to make a decision. a) IS_CHECKPOINTED, /* is it checkpointed before? */ b) HAS_FSYNCED_INODE, /* is the inode fsynced before? */ c) HAS_LAST_FSYNC, /* has the latest node fsync mark? */ And, the scenarios for our rule are based on: [Term] F: fsync_mark, D: dentry_mark 1. inode(x) | CP | inode(x) | dnode(F) 2. inode(x) | CP | inode(F) | dnode(F) 3. inode(x) | CP | dnode(F) | inode(x) | inode(F) 4. inode(x) | CP | dnode(F) | inode(F) 5. CP | inode(x) | dnode(F) | inode(DF) 6. CP | inode(DF) | dnode(F) 7. CP | dnode(F) | inode(DF) 8. CP | dnode(F) | inode(x) | inode(DF) For example, #3, the three conditions should be changed as follows. inode(x) | CP | dnode(F) | inode(x) | inode(F) a) x o o o o b) x x x x o c) x o o x o If f2fs_sync_file stops ------^, it should write inode(F) --------------^ So, the need_inode_block_update should return true, since c) get_nat_flag(e, HAS_LAST_FSYNC), is false. For example, #8, CP | alloc | dnode(F) | inode(x) | inode(DF) a) o x x x x b) x x x o c) o o x o If f2fs_sync_file stops -------^, it should write inode(DF) --------------^ Note that, the roll-forward policy should follow this rule, which means, if there are any missing blocks, we doesn't need to recover that inode. Signed-off-by: Huang Ying <ying.huang@intel.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-16 05:50:48 +08:00
bool need_update = true;
down_read(&nm_i->nat_tree_lock);
f2fs: fix conditions to remain recovery information in f2fs_sync_file This patch revisited whole the recovery information during the f2fs_sync_file. In this patch, there are three information to make a decision. a) IS_CHECKPOINTED, /* is it checkpointed before? */ b) HAS_FSYNCED_INODE, /* is the inode fsynced before? */ c) HAS_LAST_FSYNC, /* has the latest node fsync mark? */ And, the scenarios for our rule are based on: [Term] F: fsync_mark, D: dentry_mark 1. inode(x) | CP | inode(x) | dnode(F) 2. inode(x) | CP | inode(F) | dnode(F) 3. inode(x) | CP | dnode(F) | inode(x) | inode(F) 4. inode(x) | CP | dnode(F) | inode(F) 5. CP | inode(x) | dnode(F) | inode(DF) 6. CP | inode(DF) | dnode(F) 7. CP | dnode(F) | inode(DF) 8. CP | dnode(F) | inode(x) | inode(DF) For example, #3, the three conditions should be changed as follows. inode(x) | CP | dnode(F) | inode(x) | inode(F) a) x o o o o b) x x x x o c) x o o x o If f2fs_sync_file stops ------^, it should write inode(F) --------------^ So, the need_inode_block_update should return true, since c) get_nat_flag(e, HAS_LAST_FSYNC), is false. For example, #8, CP | alloc | dnode(F) | inode(x) | inode(DF) a) o x x x x b) x x x o c) o o x o If f2fs_sync_file stops -------^, it should write inode(DF) --------------^ Note that, the roll-forward policy should follow this rule, which means, if there are any missing blocks, we doesn't need to recover that inode. Signed-off-by: Huang Ying <ying.huang@intel.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-16 05:50:48 +08:00
e = __lookup_nat_cache(nm_i, ino);
if (e && get_nat_flag(e, HAS_LAST_FSYNC) &&
(get_nat_flag(e, IS_CHECKPOINTED) ||
get_nat_flag(e, HAS_FSYNCED_INODE)))
need_update = false;
up_read(&nm_i->nat_tree_lock);
f2fs: fix conditions to remain recovery information in f2fs_sync_file This patch revisited whole the recovery information during the f2fs_sync_file. In this patch, there are three information to make a decision. a) IS_CHECKPOINTED, /* is it checkpointed before? */ b) HAS_FSYNCED_INODE, /* is the inode fsynced before? */ c) HAS_LAST_FSYNC, /* has the latest node fsync mark? */ And, the scenarios for our rule are based on: [Term] F: fsync_mark, D: dentry_mark 1. inode(x) | CP | inode(x) | dnode(F) 2. inode(x) | CP | inode(F) | dnode(F) 3. inode(x) | CP | dnode(F) | inode(x) | inode(F) 4. inode(x) | CP | dnode(F) | inode(F) 5. CP | inode(x) | dnode(F) | inode(DF) 6. CP | inode(DF) | dnode(F) 7. CP | dnode(F) | inode(DF) 8. CP | dnode(F) | inode(x) | inode(DF) For example, #3, the three conditions should be changed as follows. inode(x) | CP | dnode(F) | inode(x) | inode(F) a) x o o o o b) x x x x o c) x o o x o If f2fs_sync_file stops ------^, it should write inode(F) --------------^ So, the need_inode_block_update should return true, since c) get_nat_flag(e, HAS_LAST_FSYNC), is false. For example, #8, CP | alloc | dnode(F) | inode(x) | inode(DF) a) o x x x x b) x x x o c) o o x o If f2fs_sync_file stops -------^, it should write inode(DF) --------------^ Note that, the roll-forward policy should follow this rule, which means, if there are any missing blocks, we doesn't need to recover that inode. Signed-off-by: Huang Ying <ying.huang@intel.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-16 05:50:48 +08:00
return need_update;
}
static struct nat_entry *grab_nat_entry(struct f2fs_nm_info *nm_i, nid_t nid)
{
struct nat_entry *new;
new = f2fs_kmem_cache_alloc(nat_entry_slab, GFP_NOFS);
f2fs_radix_tree_insert(&nm_i->nat_root, nid, new);
memset(new, 0, sizeof(struct nat_entry));
nat_set_nid(new, nid);
f2fs: fix conditions to remain recovery information in f2fs_sync_file This patch revisited whole the recovery information during the f2fs_sync_file. In this patch, there are three information to make a decision. a) IS_CHECKPOINTED, /* is it checkpointed before? */ b) HAS_FSYNCED_INODE, /* is the inode fsynced before? */ c) HAS_LAST_FSYNC, /* has the latest node fsync mark? */ And, the scenarios for our rule are based on: [Term] F: fsync_mark, D: dentry_mark 1. inode(x) | CP | inode(x) | dnode(F) 2. inode(x) | CP | inode(F) | dnode(F) 3. inode(x) | CP | dnode(F) | inode(x) | inode(F) 4. inode(x) | CP | dnode(F) | inode(F) 5. CP | inode(x) | dnode(F) | inode(DF) 6. CP | inode(DF) | dnode(F) 7. CP | dnode(F) | inode(DF) 8. CP | dnode(F) | inode(x) | inode(DF) For example, #3, the three conditions should be changed as follows. inode(x) | CP | dnode(F) | inode(x) | inode(F) a) x o o o o b) x x x x o c) x o o x o If f2fs_sync_file stops ------^, it should write inode(F) --------------^ So, the need_inode_block_update should return true, since c) get_nat_flag(e, HAS_LAST_FSYNC), is false. For example, #8, CP | alloc | dnode(F) | inode(x) | inode(DF) a) o x x x x b) x x x o c) o o x o If f2fs_sync_file stops -------^, it should write inode(DF) --------------^ Note that, the roll-forward policy should follow this rule, which means, if there are any missing blocks, we doesn't need to recover that inode. Signed-off-by: Huang Ying <ying.huang@intel.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-16 05:50:48 +08:00
nat_reset_flag(new);
list_add_tail(&new->list, &nm_i->nat_entries);
nm_i->nat_cnt++;
return new;
}
static void cache_nat_entry(struct f2fs_sb_info *sbi, nid_t nid,
struct f2fs_nat_entry *ne)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct nat_entry *e;
e = __lookup_nat_cache(nm_i, nid);
if (!e) {
e = grab_nat_entry(nm_i, nid);
node_info_from_raw_nat(&e->ni, ne);
} else {
f2fs_bug_on(sbi, nat_get_ino(e) != ne->ino ||
nat_get_blkaddr(e) != ne->block_addr ||
nat_get_version(e) != ne->version);
}
}
static void set_node_addr(struct f2fs_sb_info *sbi, struct node_info *ni,
block_t new_blkaddr, bool fsync_done)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct nat_entry *e;
down_write(&nm_i->nat_tree_lock);
e = __lookup_nat_cache(nm_i, ni->nid);
if (!e) {
e = grab_nat_entry(nm_i, ni->nid);
copy_node_info(&e->ni, ni);
f2fs_bug_on(sbi, ni->blk_addr == NEW_ADDR);
} else if (new_blkaddr == NEW_ADDR) {
/*
* when nid is reallocated,
* previous nat entry can be remained in nat cache.
* So, reinitialize it with new information.
*/
copy_node_info(&e->ni, ni);
f2fs_bug_on(sbi, ni->blk_addr != NULL_ADDR);
}
/* sanity check */
f2fs_bug_on(sbi, nat_get_blkaddr(e) != ni->blk_addr);
f2fs_bug_on(sbi, nat_get_blkaddr(e) == NULL_ADDR &&
new_blkaddr == NULL_ADDR);
f2fs_bug_on(sbi, nat_get_blkaddr(e) == NEW_ADDR &&
new_blkaddr == NEW_ADDR);
f2fs_bug_on(sbi, nat_get_blkaddr(e) != NEW_ADDR &&
nat_get_blkaddr(e) != NULL_ADDR &&
new_blkaddr == NEW_ADDR);
/* increment version no as node is removed */
if (nat_get_blkaddr(e) != NEW_ADDR && new_blkaddr == NULL_ADDR) {
unsigned char version = nat_get_version(e);
nat_set_version(e, inc_node_version(version));
/* in order to reuse the nid */
if (nm_i->next_scan_nid > ni->nid)
nm_i->next_scan_nid = ni->nid;
}
/* change address */
nat_set_blkaddr(e, new_blkaddr);
f2fs: fix conditions to remain recovery information in f2fs_sync_file This patch revisited whole the recovery information during the f2fs_sync_file. In this patch, there are three information to make a decision. a) IS_CHECKPOINTED, /* is it checkpointed before? */ b) HAS_FSYNCED_INODE, /* is the inode fsynced before? */ c) HAS_LAST_FSYNC, /* has the latest node fsync mark? */ And, the scenarios for our rule are based on: [Term] F: fsync_mark, D: dentry_mark 1. inode(x) | CP | inode(x) | dnode(F) 2. inode(x) | CP | inode(F) | dnode(F) 3. inode(x) | CP | dnode(F) | inode(x) | inode(F) 4. inode(x) | CP | dnode(F) | inode(F) 5. CP | inode(x) | dnode(F) | inode(DF) 6. CP | inode(DF) | dnode(F) 7. CP | dnode(F) | inode(DF) 8. CP | dnode(F) | inode(x) | inode(DF) For example, #3, the three conditions should be changed as follows. inode(x) | CP | dnode(F) | inode(x) | inode(F) a) x o o o o b) x x x x o c) x o o x o If f2fs_sync_file stops ------^, it should write inode(F) --------------^ So, the need_inode_block_update should return true, since c) get_nat_flag(e, HAS_LAST_FSYNC), is false. For example, #8, CP | alloc | dnode(F) | inode(x) | inode(DF) a) o x x x x b) x x x o c) o o x o If f2fs_sync_file stops -------^, it should write inode(DF) --------------^ Note that, the roll-forward policy should follow this rule, which means, if there are any missing blocks, we doesn't need to recover that inode. Signed-off-by: Huang Ying <ying.huang@intel.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-16 05:50:48 +08:00
if (new_blkaddr == NEW_ADDR || new_blkaddr == NULL_ADDR)
set_nat_flag(e, IS_CHECKPOINTED, false);
__set_nat_cache_dirty(nm_i, e);
/* update fsync_mark if its inode nat entry is still alive */
if (ni->nid != ni->ino)
e = __lookup_nat_cache(nm_i, ni->ino);
f2fs: fix conditions to remain recovery information in f2fs_sync_file This patch revisited whole the recovery information during the f2fs_sync_file. In this patch, there are three information to make a decision. a) IS_CHECKPOINTED, /* is it checkpointed before? */ b) HAS_FSYNCED_INODE, /* is the inode fsynced before? */ c) HAS_LAST_FSYNC, /* has the latest node fsync mark? */ And, the scenarios for our rule are based on: [Term] F: fsync_mark, D: dentry_mark 1. inode(x) | CP | inode(x) | dnode(F) 2. inode(x) | CP | inode(F) | dnode(F) 3. inode(x) | CP | dnode(F) | inode(x) | inode(F) 4. inode(x) | CP | dnode(F) | inode(F) 5. CP | inode(x) | dnode(F) | inode(DF) 6. CP | inode(DF) | dnode(F) 7. CP | dnode(F) | inode(DF) 8. CP | dnode(F) | inode(x) | inode(DF) For example, #3, the three conditions should be changed as follows. inode(x) | CP | dnode(F) | inode(x) | inode(F) a) x o o o o b) x x x x o c) x o o x o If f2fs_sync_file stops ------^, it should write inode(F) --------------^ So, the need_inode_block_update should return true, since c) get_nat_flag(e, HAS_LAST_FSYNC), is false. For example, #8, CP | alloc | dnode(F) | inode(x) | inode(DF) a) o x x x x b) x x x o c) o o x o If f2fs_sync_file stops -------^, it should write inode(DF) --------------^ Note that, the roll-forward policy should follow this rule, which means, if there are any missing blocks, we doesn't need to recover that inode. Signed-off-by: Huang Ying <ying.huang@intel.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-16 05:50:48 +08:00
if (e) {
if (fsync_done && ni->nid == ni->ino)
set_nat_flag(e, HAS_FSYNCED_INODE, true);
set_nat_flag(e, HAS_LAST_FSYNC, fsync_done);
}
up_write(&nm_i->nat_tree_lock);
}
int try_to_free_nats(struct f2fs_sb_info *sbi, int nr_shrink)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
int nr = nr_shrink;
if (!down_write_trylock(&nm_i->nat_tree_lock))
return 0;
while (nr_shrink && !list_empty(&nm_i->nat_entries)) {
struct nat_entry *ne;
ne = list_first_entry(&nm_i->nat_entries,
struct nat_entry, list);
__del_from_nat_cache(nm_i, ne);
nr_shrink--;
}
up_write(&nm_i->nat_tree_lock);
return nr - nr_shrink;
}
/*
* This function always returns success
*/
void get_node_info(struct f2fs_sb_info *sbi, nid_t nid, struct node_info *ni)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_HOT_DATA);
struct f2fs_journal *journal = curseg->journal;
nid_t start_nid = START_NID(nid);
struct f2fs_nat_block *nat_blk;
struct page *page = NULL;
struct f2fs_nat_entry ne;
struct nat_entry *e;
int i;
ni->nid = nid;
/* Check nat cache */
down_read(&nm_i->nat_tree_lock);
e = __lookup_nat_cache(nm_i, nid);
if (e) {
ni->ino = nat_get_ino(e);
ni->blk_addr = nat_get_blkaddr(e);
ni->version = nat_get_version(e);
up_read(&nm_i->nat_tree_lock);
return;
}
memset(&ne, 0, sizeof(struct f2fs_nat_entry));
/* Check current segment summary */
down_read(&curseg->journal_rwsem);
i = lookup_journal_in_cursum(journal, NAT_JOURNAL, nid, 0);
if (i >= 0) {
ne = nat_in_journal(journal, i);
node_info_from_raw_nat(ni, &ne);
}
up_read(&curseg->journal_rwsem);
if (i >= 0)
goto cache;
/* Fill node_info from nat page */
page = get_current_nat_page(sbi, start_nid);
nat_blk = (struct f2fs_nat_block *)page_address(page);
ne = nat_blk->entries[nid - start_nid];
node_info_from_raw_nat(ni, &ne);
f2fs_put_page(page, 1);
cache:
up_read(&nm_i->nat_tree_lock);
/* cache nat entry */
down_write(&nm_i->nat_tree_lock);
cache_nat_entry(sbi, nid, &ne);
up_write(&nm_i->nat_tree_lock);
}
pgoff_t get_next_page_offset(struct dnode_of_data *dn, pgoff_t pgofs)
{
const long direct_index = ADDRS_PER_INODE(dn->inode);
const long direct_blks = ADDRS_PER_BLOCK;
const long indirect_blks = ADDRS_PER_BLOCK * NIDS_PER_BLOCK;
unsigned int skipped_unit = ADDRS_PER_BLOCK;
int cur_level = dn->cur_level;
int max_level = dn->max_level;
pgoff_t base = 0;
if (!dn->max_level)
return pgofs + 1;
while (max_level-- > cur_level)
skipped_unit *= NIDS_PER_BLOCK;
switch (dn->max_level) {
case 3:
base += 2 * indirect_blks;
case 2:
base += 2 * direct_blks;
case 1:
base += direct_index;
break;
default:
f2fs_bug_on(F2FS_I_SB(dn->inode), 1);
}
return ((pgofs - base) / skipped_unit + 1) * skipped_unit + base;
}
/*
* The maximum depth is four.
* Offset[0] will have raw inode offset.
*/
static int get_node_path(struct inode *inode, long block,
int offset[4], unsigned int noffset[4])
{
const long direct_index = ADDRS_PER_INODE(inode);
const long direct_blks = ADDRS_PER_BLOCK;
const long dptrs_per_blk = NIDS_PER_BLOCK;
const long indirect_blks = ADDRS_PER_BLOCK * NIDS_PER_BLOCK;
const long dindirect_blks = indirect_blks * NIDS_PER_BLOCK;
int n = 0;
int level = 0;
noffset[0] = 0;
if (block < direct_index) {
offset[n] = block;
goto got;
}
block -= direct_index;
if (block < direct_blks) {
offset[n++] = NODE_DIR1_BLOCK;
noffset[n] = 1;
offset[n] = block;
level = 1;
goto got;
}
block -= direct_blks;
if (block < direct_blks) {
offset[n++] = NODE_DIR2_BLOCK;
noffset[n] = 2;
offset[n] = block;
level = 1;
goto got;
}
block -= direct_blks;
if (block < indirect_blks) {
offset[n++] = NODE_IND1_BLOCK;
noffset[n] = 3;
offset[n++] = block / direct_blks;
noffset[n] = 4 + offset[n - 1];
offset[n] = block % direct_blks;
level = 2;
goto got;
}
block -= indirect_blks;
if (block < indirect_blks) {
offset[n++] = NODE_IND2_BLOCK;
noffset[n] = 4 + dptrs_per_blk;
offset[n++] = block / direct_blks;
noffset[n] = 5 + dptrs_per_blk + offset[n - 1];
offset[n] = block % direct_blks;
level = 2;
goto got;
}
block -= indirect_blks;
if (block < dindirect_blks) {
offset[n++] = NODE_DIND_BLOCK;
noffset[n] = 5 + (dptrs_per_blk * 2);
offset[n++] = block / indirect_blks;
noffset[n] = 6 + (dptrs_per_blk * 2) +
offset[n - 1] * (dptrs_per_blk + 1);
offset[n++] = (block / direct_blks) % dptrs_per_blk;
noffset[n] = 7 + (dptrs_per_blk * 2) +
offset[n - 2] * (dptrs_per_blk + 1) +
offset[n - 1];
offset[n] = block % direct_blks;
level = 3;
goto got;
} else {
BUG();
}
got:
return level;
}
/*
* Caller should call f2fs_put_dnode(dn).
* Also, it should grab and release a rwsem by calling f2fs_lock_op() and
* f2fs_unlock_op() only if ro is not set RDONLY_NODE.
f2fs: introduce a new global lock scheme In the previous version, f2fs uses global locks according to the usage types, such as directory operations, block allocation, block write, and so on. Reference the following lock types in f2fs.h. enum lock_type { RENAME, /* for renaming operations */ DENTRY_OPS, /* for directory operations */ DATA_WRITE, /* for data write */ DATA_NEW, /* for data allocation */ DATA_TRUNC, /* for data truncate */ NODE_NEW, /* for node allocation */ NODE_TRUNC, /* for node truncate */ NODE_WRITE, /* for node write */ NR_LOCK_TYPE, }; In that case, we lose the performance under the multi-threading environment, since every types of operations must be conducted one at a time. In order to address the problem, let's share the locks globally with a mutex array regardless of any types. So, let users grab a mutex and perform their jobs in parallel as much as possbile. For this, I propose a new global lock scheme as follows. 0. Data structure - f2fs_sb_info -> mutex_lock[NR_GLOBAL_LOCKS] - f2fs_sb_info -> node_write 1. mutex_lock_op(sbi) - try to get an avaiable lock from the array. - returns the index of the gottern lock variable. 2. mutex_unlock_op(sbi, index of the lock) - unlock the given index of the lock. 3. mutex_lock_all(sbi) - grab all the locks in the array before the checkpoint. 4. mutex_unlock_all(sbi) - release all the locks in the array after checkpoint. 5. block_operations() - call mutex_lock_all() - sync_dirty_dir_inodes() - grab node_write - sync_node_pages() Note that, the pairs of mutex_lock_op()/mutex_unlock_op() and mutex_lock_all()/mutex_unlock_all() should be used together. Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2012-11-22 15:21:29 +08:00
* In the case of RDONLY_NODE, we don't need to care about mutex.
*/
int get_dnode_of_data(struct dnode_of_data *dn, pgoff_t index, int mode)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(dn->inode);
struct page *npage[4];
struct page *parent = NULL;
int offset[4];
unsigned int noffset[4];
nid_t nids[4];
int level, i = 0;
int err = 0;
level = get_node_path(dn->inode, index, offset, noffset);
nids[0] = dn->inode->i_ino;
npage[0] = dn->inode_page;
if (!npage[0]) {
npage[0] = get_node_page(sbi, nids[0]);
if (IS_ERR(npage[0]))
return PTR_ERR(npage[0]);
}
/* if inline_data is set, should not report any block indices */
if (f2fs_has_inline_data(dn->inode) && index) {
err = -ENOENT;
f2fs_put_page(npage[0], 1);
goto release_out;
}
parent = npage[0];
if (level != 0)
nids[1] = get_nid(parent, offset[0], true);
dn->inode_page = npage[0];
dn->inode_page_locked = true;
/* get indirect or direct nodes */
for (i = 1; i <= level; i++) {
bool done = false;
if (!nids[i] && mode == ALLOC_NODE) {
/* alloc new node */
if (!alloc_nid(sbi, &(nids[i]))) {
err = -ENOSPC;
goto release_pages;
}
dn->nid = nids[i];
npage[i] = new_node_page(dn, noffset[i], NULL);
if (IS_ERR(npage[i])) {
alloc_nid_failed(sbi, nids[i]);
err = PTR_ERR(npage[i]);
goto release_pages;
}
set_nid(parent, offset[i - 1], nids[i], i == 1);
alloc_nid_done(sbi, nids[i]);
done = true;
} else if (mode == LOOKUP_NODE_RA && i == level && level > 1) {
npage[i] = get_node_page_ra(parent, offset[i - 1]);
if (IS_ERR(npage[i])) {
err = PTR_ERR(npage[i]);
goto release_pages;
}
done = true;
}
if (i == 1) {
dn->inode_page_locked = false;
unlock_page(parent);
} else {
f2fs_put_page(parent, 1);
}
if (!done) {
npage[i] = get_node_page(sbi, nids[i]);
if (IS_ERR(npage[i])) {
err = PTR_ERR(npage[i]);
f2fs_put_page(npage[0], 0);
goto release_out;
}
}
if (i < level) {
parent = npage[i];
nids[i + 1] = get_nid(parent, offset[i], false);
}
}
dn->nid = nids[level];
dn->ofs_in_node = offset[level];
dn->node_page = npage[level];
dn->data_blkaddr = datablock_addr(dn->node_page, dn->ofs_in_node);
return 0;
release_pages:
f2fs_put_page(parent, 1);
if (i > 1)
f2fs_put_page(npage[0], 0);
release_out:
dn->inode_page = NULL;
dn->node_page = NULL;
if (err == -ENOENT) {
dn->cur_level = i;
dn->max_level = level;
}
return err;
}
static void truncate_node(struct dnode_of_data *dn)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(dn->inode);
struct node_info ni;
get_node_info(sbi, dn->nid, &ni);
if (dn->inode->i_blocks == 0) {
f2fs_bug_on(sbi, ni.blk_addr != NULL_ADDR);
goto invalidate;
}
f2fs_bug_on(sbi, ni.blk_addr == NULL_ADDR);
/* Deallocate node address */
invalidate_blocks(sbi, ni.blk_addr);
dec_valid_node_count(sbi, dn->inode);
set_node_addr(sbi, &ni, NULL_ADDR, false);
if (dn->nid == dn->inode->i_ino) {
remove_orphan_inode(sbi, dn->nid);
dec_valid_inode_count(sbi);
} else {
sync_inode_page(dn);
}
invalidate:
clear_node_page_dirty(dn->node_page);
set_sbi_flag(sbi, SBI_IS_DIRTY);
f2fs_put_page(dn->node_page, 1);
invalidate_mapping_pages(NODE_MAPPING(sbi),
dn->node_page->index, dn->node_page->index);
dn->node_page = NULL;
trace_f2fs_truncate_node(dn->inode, dn->nid, ni.blk_addr);
}
static int truncate_dnode(struct dnode_of_data *dn)
{
struct page *page;
if (dn->nid == 0)
return 1;
/* get direct node */
page = get_node_page(F2FS_I_SB(dn->inode), dn->nid);
if (IS_ERR(page) && PTR_ERR(page) == -ENOENT)
return 1;
else if (IS_ERR(page))
return PTR_ERR(page);
/* Make dnode_of_data for parameter */
dn->node_page = page;
dn->ofs_in_node = 0;
truncate_data_blocks(dn);
truncate_node(dn);
return 1;
}
static int truncate_nodes(struct dnode_of_data *dn, unsigned int nofs,
int ofs, int depth)
{
struct dnode_of_data rdn = *dn;
struct page *page;
struct f2fs_node *rn;
nid_t child_nid;
unsigned int child_nofs;
int freed = 0;
int i, ret;
if (dn->nid == 0)
return NIDS_PER_BLOCK + 1;
trace_f2fs_truncate_nodes_enter(dn->inode, dn->nid, dn->data_blkaddr);
page = get_node_page(F2FS_I_SB(dn->inode), dn->nid);
if (IS_ERR(page)) {
trace_f2fs_truncate_nodes_exit(dn->inode, PTR_ERR(page));
return PTR_ERR(page);
}
rn = F2FS_NODE(page);
if (depth < 3) {
for (i = ofs; i < NIDS_PER_BLOCK; i++, freed++) {
child_nid = le32_to_cpu(rn->in.nid[i]);
if (child_nid == 0)
continue;
rdn.nid = child_nid;
ret = truncate_dnode(&rdn);
if (ret < 0)
goto out_err;
if (set_nid(page, i, 0, false))
dn->node_changed = true;
}
} else {
child_nofs = nofs + ofs * (NIDS_PER_BLOCK + 1) + 1;
for (i = ofs; i < NIDS_PER_BLOCK; i++) {
child_nid = le32_to_cpu(rn->in.nid[i]);
if (child_nid == 0) {
child_nofs += NIDS_PER_BLOCK + 1;
continue;
}
rdn.nid = child_nid;
ret = truncate_nodes(&rdn, child_nofs, 0, depth - 1);
if (ret == (NIDS_PER_BLOCK + 1)) {
if (set_nid(page, i, 0, false))
dn->node_changed = true;
child_nofs += ret;
} else if (ret < 0 && ret != -ENOENT) {
goto out_err;
}
}
freed = child_nofs;
}
if (!ofs) {
/* remove current indirect node */
dn->node_page = page;
truncate_node(dn);
freed++;
} else {
f2fs_put_page(page, 1);
}
trace_f2fs_truncate_nodes_exit(dn->inode, freed);
return freed;
out_err:
f2fs_put_page(page, 1);
trace_f2fs_truncate_nodes_exit(dn->inode, ret);
return ret;
}
static int truncate_partial_nodes(struct dnode_of_data *dn,
struct f2fs_inode *ri, int *offset, int depth)
{
struct page *pages[2];
nid_t nid[3];
nid_t child_nid;
int err = 0;
int i;
int idx = depth - 2;
nid[0] = le32_to_cpu(ri->i_nid[offset[0] - NODE_DIR1_BLOCK]);
if (!nid[0])
return 0;
/* get indirect nodes in the path */
f2fs: fix truncate_partial_nodes bug The truncate_partial_nodes puts pages incorrectly in the following two cases. Note that the value for argc 'depth' can only be 2 or 3. Please see truncate_inode_blocks() and truncate_partial_nodes(). 1) An err is occurred in the first 'for' loop When err is occurred with depth = 2, pages[0] is invalid, so this page doesn't need to be put. There is no problem, however, when depth is 3, it doesn't put the pages correctly where pages[0] is valid and pages[1] is invalid. In this case, depth is set to 2 (ref to statemnt depth = i + 1), and then 'goto fail'. In label 'fail', for (i = depth - 3; i >= 0; i--) cannot meet the condition because i = -1, so pages[0] cann't be put. 2) An err happened in the second 'for' loop Now we've got pages[0] with depth = 2, or we've got pages[0] and pages[1] with depth = 3. When an err is detected, we need 'goto fail' to put such the pages. When depth is 2, in label 'fail', for (i = depth - 3; i >= 0; i--) cann't meet the condition because i = -1, so pages[0] cann't be put. When depth is 3, in label 'fail', for (i = depth - 3; i >= 0; i--) can only put pages[0], pages[1] also cann't be put. Note that 'depth' has been changed before first 'goto fail' (ref to statemnt depth = i + 1), so passing this modified 'depth' to the tracepoint, trace_f2fs_truncate_partial_nodes, is also incorrect. Signed-off-by: Shifei Ge <shifei10.ge@samsung.com> [Jaegeuk Kim: modify the description and fix one bug] Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2013-10-29 15:32:34 +08:00
for (i = 0; i < idx + 1; i++) {
/* reference count'll be increased */
pages[i] = get_node_page(F2FS_I_SB(dn->inode), nid[i]);
if (IS_ERR(pages[i])) {
err = PTR_ERR(pages[i]);
f2fs: fix truncate_partial_nodes bug The truncate_partial_nodes puts pages incorrectly in the following two cases. Note that the value for argc 'depth' can only be 2 or 3. Please see truncate_inode_blocks() and truncate_partial_nodes(). 1) An err is occurred in the first 'for' loop When err is occurred with depth = 2, pages[0] is invalid, so this page doesn't need to be put. There is no problem, however, when depth is 3, it doesn't put the pages correctly where pages[0] is valid and pages[1] is invalid. In this case, depth is set to 2 (ref to statemnt depth = i + 1), and then 'goto fail'. In label 'fail', for (i = depth - 3; i >= 0; i--) cannot meet the condition because i = -1, so pages[0] cann't be put. 2) An err happened in the second 'for' loop Now we've got pages[0] with depth = 2, or we've got pages[0] and pages[1] with depth = 3. When an err is detected, we need 'goto fail' to put such the pages. When depth is 2, in label 'fail', for (i = depth - 3; i >= 0; i--) cann't meet the condition because i = -1, so pages[0] cann't be put. When depth is 3, in label 'fail', for (i = depth - 3; i >= 0; i--) can only put pages[0], pages[1] also cann't be put. Note that 'depth' has been changed before first 'goto fail' (ref to statemnt depth = i + 1), so passing this modified 'depth' to the tracepoint, trace_f2fs_truncate_partial_nodes, is also incorrect. Signed-off-by: Shifei Ge <shifei10.ge@samsung.com> [Jaegeuk Kim: modify the description and fix one bug] Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2013-10-29 15:32:34 +08:00
idx = i - 1;
goto fail;
}
nid[i + 1] = get_nid(pages[i], offset[i + 1], false);
}
/* free direct nodes linked to a partial indirect node */
f2fs: fix truncate_partial_nodes bug The truncate_partial_nodes puts pages incorrectly in the following two cases. Note that the value for argc 'depth' can only be 2 or 3. Please see truncate_inode_blocks() and truncate_partial_nodes(). 1) An err is occurred in the first 'for' loop When err is occurred with depth = 2, pages[0] is invalid, so this page doesn't need to be put. There is no problem, however, when depth is 3, it doesn't put the pages correctly where pages[0] is valid and pages[1] is invalid. In this case, depth is set to 2 (ref to statemnt depth = i + 1), and then 'goto fail'. In label 'fail', for (i = depth - 3; i >= 0; i--) cannot meet the condition because i = -1, so pages[0] cann't be put. 2) An err happened in the second 'for' loop Now we've got pages[0] with depth = 2, or we've got pages[0] and pages[1] with depth = 3. When an err is detected, we need 'goto fail' to put such the pages. When depth is 2, in label 'fail', for (i = depth - 3; i >= 0; i--) cann't meet the condition because i = -1, so pages[0] cann't be put. When depth is 3, in label 'fail', for (i = depth - 3; i >= 0; i--) can only put pages[0], pages[1] also cann't be put. Note that 'depth' has been changed before first 'goto fail' (ref to statemnt depth = i + 1), so passing this modified 'depth' to the tracepoint, trace_f2fs_truncate_partial_nodes, is also incorrect. Signed-off-by: Shifei Ge <shifei10.ge@samsung.com> [Jaegeuk Kim: modify the description and fix one bug] Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2013-10-29 15:32:34 +08:00
for (i = offset[idx + 1]; i < NIDS_PER_BLOCK; i++) {
child_nid = get_nid(pages[idx], i, false);
if (!child_nid)
continue;
dn->nid = child_nid;
err = truncate_dnode(dn);
if (err < 0)
goto fail;
if (set_nid(pages[idx], i, 0, false))
dn->node_changed = true;
}
f2fs: fix truncate_partial_nodes bug The truncate_partial_nodes puts pages incorrectly in the following two cases. Note that the value for argc 'depth' can only be 2 or 3. Please see truncate_inode_blocks() and truncate_partial_nodes(). 1) An err is occurred in the first 'for' loop When err is occurred with depth = 2, pages[0] is invalid, so this page doesn't need to be put. There is no problem, however, when depth is 3, it doesn't put the pages correctly where pages[0] is valid and pages[1] is invalid. In this case, depth is set to 2 (ref to statemnt depth = i + 1), and then 'goto fail'. In label 'fail', for (i = depth - 3; i >= 0; i--) cannot meet the condition because i = -1, so pages[0] cann't be put. 2) An err happened in the second 'for' loop Now we've got pages[0] with depth = 2, or we've got pages[0] and pages[1] with depth = 3. When an err is detected, we need 'goto fail' to put such the pages. When depth is 2, in label 'fail', for (i = depth - 3; i >= 0; i--) cann't meet the condition because i = -1, so pages[0] cann't be put. When depth is 3, in label 'fail', for (i = depth - 3; i >= 0; i--) can only put pages[0], pages[1] also cann't be put. Note that 'depth' has been changed before first 'goto fail' (ref to statemnt depth = i + 1), so passing this modified 'depth' to the tracepoint, trace_f2fs_truncate_partial_nodes, is also incorrect. Signed-off-by: Shifei Ge <shifei10.ge@samsung.com> [Jaegeuk Kim: modify the description and fix one bug] Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2013-10-29 15:32:34 +08:00
if (offset[idx + 1] == 0) {
dn->node_page = pages[idx];
dn->nid = nid[idx];
truncate_node(dn);
} else {
f2fs_put_page(pages[idx], 1);
}
offset[idx]++;
f2fs: fix truncate_partial_nodes bug The truncate_partial_nodes puts pages incorrectly in the following two cases. Note that the value for argc 'depth' can only be 2 or 3. Please see truncate_inode_blocks() and truncate_partial_nodes(). 1) An err is occurred in the first 'for' loop When err is occurred with depth = 2, pages[0] is invalid, so this page doesn't need to be put. There is no problem, however, when depth is 3, it doesn't put the pages correctly where pages[0] is valid and pages[1] is invalid. In this case, depth is set to 2 (ref to statemnt depth = i + 1), and then 'goto fail'. In label 'fail', for (i = depth - 3; i >= 0; i--) cannot meet the condition because i = -1, so pages[0] cann't be put. 2) An err happened in the second 'for' loop Now we've got pages[0] with depth = 2, or we've got pages[0] and pages[1] with depth = 3. When an err is detected, we need 'goto fail' to put such the pages. When depth is 2, in label 'fail', for (i = depth - 3; i >= 0; i--) cann't meet the condition because i = -1, so pages[0] cann't be put. When depth is 3, in label 'fail', for (i = depth - 3; i >= 0; i--) can only put pages[0], pages[1] also cann't be put. Note that 'depth' has been changed before first 'goto fail' (ref to statemnt depth = i + 1), so passing this modified 'depth' to the tracepoint, trace_f2fs_truncate_partial_nodes, is also incorrect. Signed-off-by: Shifei Ge <shifei10.ge@samsung.com> [Jaegeuk Kim: modify the description and fix one bug] Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2013-10-29 15:32:34 +08:00
offset[idx + 1] = 0;
idx--;
fail:
f2fs: fix truncate_partial_nodes bug The truncate_partial_nodes puts pages incorrectly in the following two cases. Note that the value for argc 'depth' can only be 2 or 3. Please see truncate_inode_blocks() and truncate_partial_nodes(). 1) An err is occurred in the first 'for' loop When err is occurred with depth = 2, pages[0] is invalid, so this page doesn't need to be put. There is no problem, however, when depth is 3, it doesn't put the pages correctly where pages[0] is valid and pages[1] is invalid. In this case, depth is set to 2 (ref to statemnt depth = i + 1), and then 'goto fail'. In label 'fail', for (i = depth - 3; i >= 0; i--) cannot meet the condition because i = -1, so pages[0] cann't be put. 2) An err happened in the second 'for' loop Now we've got pages[0] with depth = 2, or we've got pages[0] and pages[1] with depth = 3. When an err is detected, we need 'goto fail' to put such the pages. When depth is 2, in label 'fail', for (i = depth - 3; i >= 0; i--) cann't meet the condition because i = -1, so pages[0] cann't be put. When depth is 3, in label 'fail', for (i = depth - 3; i >= 0; i--) can only put pages[0], pages[1] also cann't be put. Note that 'depth' has been changed before first 'goto fail' (ref to statemnt depth = i + 1), so passing this modified 'depth' to the tracepoint, trace_f2fs_truncate_partial_nodes, is also incorrect. Signed-off-by: Shifei Ge <shifei10.ge@samsung.com> [Jaegeuk Kim: modify the description and fix one bug] Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2013-10-29 15:32:34 +08:00
for (i = idx; i >= 0; i--)
f2fs_put_page(pages[i], 1);
trace_f2fs_truncate_partial_nodes(dn->inode, nid, depth, err);
return err;
}
/*
* All the block addresses of data and nodes should be nullified.
*/
int truncate_inode_blocks(struct inode *inode, pgoff_t from)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
int err = 0, cont = 1;
int level, offset[4], noffset[4];
unsigned int nofs = 0;
struct f2fs_inode *ri;
struct dnode_of_data dn;
struct page *page;
trace_f2fs_truncate_inode_blocks_enter(inode, from);
level = get_node_path(inode, from, offset, noffset);
restart:
page = get_node_page(sbi, inode->i_ino);
if (IS_ERR(page)) {
trace_f2fs_truncate_inode_blocks_exit(inode, PTR_ERR(page));
return PTR_ERR(page);
}
set_new_dnode(&dn, inode, page, NULL, 0);
unlock_page(page);
ri = F2FS_INODE(page);
switch (level) {
case 0:
case 1:
nofs = noffset[1];
break;
case 2:
nofs = noffset[1];
if (!offset[level - 1])
goto skip_partial;
err = truncate_partial_nodes(&dn, ri, offset, level);
if (err < 0 && err != -ENOENT)
goto fail;
nofs += 1 + NIDS_PER_BLOCK;
break;
case 3:
nofs = 5 + 2 * NIDS_PER_BLOCK;
if (!offset[level - 1])
goto skip_partial;
err = truncate_partial_nodes(&dn, ri, offset, level);
if (err < 0 && err != -ENOENT)
goto fail;
break;
default:
BUG();
}
skip_partial:
while (cont) {
dn.nid = le32_to_cpu(ri->i_nid[offset[0] - NODE_DIR1_BLOCK]);
switch (offset[0]) {
case NODE_DIR1_BLOCK:
case NODE_DIR2_BLOCK:
err = truncate_dnode(&dn);
break;
case NODE_IND1_BLOCK:
case NODE_IND2_BLOCK:
err = truncate_nodes(&dn, nofs, offset[1], 2);
break;
case NODE_DIND_BLOCK:
err = truncate_nodes(&dn, nofs, offset[1], 3);
cont = 0;
break;
default:
BUG();
}
if (err < 0 && err != -ENOENT)
goto fail;
if (offset[1] == 0 &&
ri->i_nid[offset[0] - NODE_DIR1_BLOCK]) {
lock_page(page);
if (unlikely(page->mapping != NODE_MAPPING(sbi))) {
f2fs_put_page(page, 1);
goto restart;
}
f2fs_wait_on_page_writeback(page, NODE, true);
ri->i_nid[offset[0] - NODE_DIR1_BLOCK] = 0;
set_page_dirty(page);
unlock_page(page);
}
offset[1] = 0;
offset[0]++;
nofs += err;
}
fail:
f2fs_put_page(page, 0);
trace_f2fs_truncate_inode_blocks_exit(inode, err);
return err > 0 ? 0 : err;
}
int truncate_xattr_node(struct inode *inode, struct page *page)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
nid_t nid = F2FS_I(inode)->i_xattr_nid;
struct dnode_of_data dn;
struct page *npage;
if (!nid)
return 0;
npage = get_node_page(sbi, nid);
if (IS_ERR(npage))
return PTR_ERR(npage);
F2FS_I(inode)->i_xattr_nid = 0;
/* need to do checkpoint during fsync */
F2FS_I(inode)->xattr_ver = cur_cp_version(F2FS_CKPT(sbi));
set_new_dnode(&dn, inode, page, npage, nid);
if (page)
dn.inode_page_locked = true;
truncate_node(&dn);
return 0;
}
f2fs: introduce a new global lock scheme In the previous version, f2fs uses global locks according to the usage types, such as directory operations, block allocation, block write, and so on. Reference the following lock types in f2fs.h. enum lock_type { RENAME, /* for renaming operations */ DENTRY_OPS, /* for directory operations */ DATA_WRITE, /* for data write */ DATA_NEW, /* for data allocation */ DATA_TRUNC, /* for data truncate */ NODE_NEW, /* for node allocation */ NODE_TRUNC, /* for node truncate */ NODE_WRITE, /* for node write */ NR_LOCK_TYPE, }; In that case, we lose the performance under the multi-threading environment, since every types of operations must be conducted one at a time. In order to address the problem, let's share the locks globally with a mutex array regardless of any types. So, let users grab a mutex and perform their jobs in parallel as much as possbile. For this, I propose a new global lock scheme as follows. 0. Data structure - f2fs_sb_info -> mutex_lock[NR_GLOBAL_LOCKS] - f2fs_sb_info -> node_write 1. mutex_lock_op(sbi) - try to get an avaiable lock from the array. - returns the index of the gottern lock variable. 2. mutex_unlock_op(sbi, index of the lock) - unlock the given index of the lock. 3. mutex_lock_all(sbi) - grab all the locks in the array before the checkpoint. 4. mutex_unlock_all(sbi) - release all the locks in the array after checkpoint. 5. block_operations() - call mutex_lock_all() - sync_dirty_dir_inodes() - grab node_write - sync_node_pages() Note that, the pairs of mutex_lock_op()/mutex_unlock_op() and mutex_lock_all()/mutex_unlock_all() should be used together. Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2012-11-22 15:21:29 +08:00
/*
* Caller should grab and release a rwsem by calling f2fs_lock_op() and
* f2fs_unlock_op().
f2fs: introduce a new global lock scheme In the previous version, f2fs uses global locks according to the usage types, such as directory operations, block allocation, block write, and so on. Reference the following lock types in f2fs.h. enum lock_type { RENAME, /* for renaming operations */ DENTRY_OPS, /* for directory operations */ DATA_WRITE, /* for data write */ DATA_NEW, /* for data allocation */ DATA_TRUNC, /* for data truncate */ NODE_NEW, /* for node allocation */ NODE_TRUNC, /* for node truncate */ NODE_WRITE, /* for node write */ NR_LOCK_TYPE, }; In that case, we lose the performance under the multi-threading environment, since every types of operations must be conducted one at a time. In order to address the problem, let's share the locks globally with a mutex array regardless of any types. So, let users grab a mutex and perform their jobs in parallel as much as possbile. For this, I propose a new global lock scheme as follows. 0. Data structure - f2fs_sb_info -> mutex_lock[NR_GLOBAL_LOCKS] - f2fs_sb_info -> node_write 1. mutex_lock_op(sbi) - try to get an avaiable lock from the array. - returns the index of the gottern lock variable. 2. mutex_unlock_op(sbi, index of the lock) - unlock the given index of the lock. 3. mutex_lock_all(sbi) - grab all the locks in the array before the checkpoint. 4. mutex_unlock_all(sbi) - release all the locks in the array after checkpoint. 5. block_operations() - call mutex_lock_all() - sync_dirty_dir_inodes() - grab node_write - sync_node_pages() Note that, the pairs of mutex_lock_op()/mutex_unlock_op() and mutex_lock_all()/mutex_unlock_all() should be used together. Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2012-11-22 15:21:29 +08:00
*/
int remove_inode_page(struct inode *inode)
{
struct dnode_of_data dn;
int err;
set_new_dnode(&dn, inode, NULL, NULL, inode->i_ino);
err = get_dnode_of_data(&dn, 0, LOOKUP_NODE);
if (err)
return err;
err = truncate_xattr_node(inode, dn.inode_page);
if (err) {
f2fs_put_dnode(&dn);
return err;
}
/* remove potential inline_data blocks */
if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
S_ISLNK(inode->i_mode))
truncate_data_blocks_range(&dn, 1);
/* 0 is possible, after f2fs_new_inode() has failed */
f2fs_bug_on(F2FS_I_SB(inode),
inode->i_blocks != 0 && inode->i_blocks != 1);
/* will put inode & node pages */
truncate_node(&dn);
return 0;
}
struct page *new_inode_page(struct inode *inode)
{
struct dnode_of_data dn;
/* allocate inode page for new inode */
set_new_dnode(&dn, inode, NULL, NULL, inode->i_ino);
/* caller should f2fs_put_page(page, 1); */
return new_node_page(&dn, 0, NULL);
}
struct page *new_node_page(struct dnode_of_data *dn,
unsigned int ofs, struct page *ipage)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(dn->inode);
struct node_info old_ni, new_ni;
struct page *page;
int err;
if (unlikely(is_inode_flag_set(F2FS_I(dn->inode), FI_NO_ALLOC)))
return ERR_PTR(-EPERM);
page = grab_cache_page(NODE_MAPPING(sbi), dn->nid);
if (!page)
return ERR_PTR(-ENOMEM);
if (unlikely(!inc_valid_node_count(sbi, dn->inode))) {
err = -ENOSPC;
goto fail;
}
get_node_info(sbi, dn->nid, &old_ni);
/* Reinitialize old_ni with new node page */
f2fs_bug_on(sbi, old_ni.blk_addr != NULL_ADDR);
new_ni = old_ni;
new_ni.ino = dn->inode->i_ino;
set_node_addr(sbi, &new_ni, NEW_ADDR, false);
f2fs_wait_on_page_writeback(page, NODE, true);
fill_node_footer(page, dn->nid, dn->inode->i_ino, ofs, true);
set_cold_node(dn->inode, page);
SetPageUptodate(page);
if (set_page_dirty(page))
dn->node_changed = true;
if (f2fs_has_xattr_block(ofs))
F2FS_I(dn->inode)->i_xattr_nid = dn->nid;
dn->node_page = page;
if (ipage)
update_inode(dn->inode, ipage);
else
sync_inode_page(dn);
if (ofs == 0)
inc_valid_inode_count(sbi);
return page;
fail:
clear_node_page_dirty(page);
f2fs_put_page(page, 1);
return ERR_PTR(err);
}
/*
* Caller should do after getting the following values.
* 0: f2fs_put_page(page, 0)
* LOCKED_PAGE or error: f2fs_put_page(page, 1)
*/
static int read_node_page(struct page *page, int rw)
{
struct f2fs_sb_info *sbi = F2FS_P_SB(page);
struct node_info ni;
struct f2fs_io_info fio = {
.sbi = sbi,
.type = NODE,
.rw = rw,
.page = page,
.encrypted_page = NULL,
};
get_node_info(sbi, page->index, &ni);
if (unlikely(ni.blk_addr == NULL_ADDR)) {
ClearPageUptodate(page);
return -ENOENT;
}
if (PageUptodate(page))
return LOCKED_PAGE;
fio.new_blkaddr = fio.old_blkaddr = ni.blk_addr;
return f2fs_submit_page_bio(&fio);
}
/*
* Readahead a node page
*/
void ra_node_page(struct f2fs_sb_info *sbi, nid_t nid)
{
struct page *apage;
int err;
if (!nid)
return;
f2fs_bug_on(sbi, check_nid_range(sbi, nid));
rcu_read_lock();
apage = radix_tree_lookup(&NODE_MAPPING(sbi)->page_tree, nid);
rcu_read_unlock();
if (apage)
return;
apage = grab_cache_page(NODE_MAPPING(sbi), nid);
if (!apage)
return;
err = read_node_page(apage, READA);
f2fs_put_page(apage, err ? 1 : 0);
}
/*
* readahead MAX_RA_NODE number of node pages.
*/
static void ra_node_pages(struct page *parent, int start)
{
struct f2fs_sb_info *sbi = F2FS_P_SB(parent);
struct blk_plug plug;
int i, end;
nid_t nid;
blk_start_plug(&plug);
/* Then, try readahead for siblings of the desired node */
end = start + MAX_RA_NODE;
end = min(end, NIDS_PER_BLOCK);
for (i = start; i < end; i++) {
nid = get_nid(parent, i, false);
ra_node_page(sbi, nid);
}
blk_finish_plug(&plug);
}
static struct page *__get_node_page(struct f2fs_sb_info *sbi, pgoff_t nid,
struct page *parent, int start)
{
struct page *page;
int err;
if (!nid)
return ERR_PTR(-ENOENT);
f2fs_bug_on(sbi, check_nid_range(sbi, nid));
repeat:
page = grab_cache_page(NODE_MAPPING(sbi), nid);
if (!page)
return ERR_PTR(-ENOMEM);
err = read_node_page(page, READ_SYNC);
if (err < 0) {
f2fs_put_page(page, 1);
return ERR_PTR(err);
} else if (err == LOCKED_PAGE) {
goto page_hit;
}
if (parent)
ra_node_pages(parent, start + 1);
f2fs: give a chance to merge IOs by IO scheduler Previously, background GC submits many 4KB read requests to load victim blocks and/or its (i)node blocks. ... f2fs_gc : f2fs_readpage: ino = 1, page_index = 0xb61, blkaddr = 0x3b964ed f2fs_gc : block_rq_complete: 8,16 R () 499854968 + 8 [0] f2fs_gc : f2fs_readpage: ino = 1, page_index = 0xb6f, blkaddr = 0x3b964ee f2fs_gc : block_rq_complete: 8,16 R () 499854976 + 8 [0] f2fs_gc : f2fs_readpage: ino = 1, page_index = 0xb79, blkaddr = 0x3b964ef f2fs_gc : block_rq_complete: 8,16 R () 499854984 + 8 [0] ... However, by the fact that many IOs are sequential, we can give a chance to merge the IOs by IO scheduler. In order to do that, let's use blk_plug. ... f2fs_gc : f2fs_iget: ino = 143 f2fs_gc : f2fs_readpage: ino = 143, page_index = 0x1c6, blkaddr = 0x2e6ee f2fs_gc : f2fs_iget: ino = 143 f2fs_gc : f2fs_readpage: ino = 143, page_index = 0x1c7, blkaddr = 0x2e6ef <idle> : block_rq_complete: 8,16 R () 1519616 + 8 [0] <idle> : block_rq_complete: 8,16 R () 1519848 + 8 [0] <idle> : block_rq_complete: 8,16 R () 1520432 + 96 [0] <idle> : block_rq_complete: 8,16 R () 1520536 + 104 [0] <idle> : block_rq_complete: 8,16 R () 1521008 + 112 [0] <idle> : block_rq_complete: 8,16 R () 1521440 + 152 [0] <idle> : block_rq_complete: 8,16 R () 1521688 + 144 [0] <idle> : block_rq_complete: 8,16 R () 1522128 + 192 [0] <idle> : block_rq_complete: 8,16 R () 1523256 + 328 [0] ... Note that this issue should be addressed in checkpoint, and some readahead flows too. Reviewed-by: Namjae Jeon <namjae.jeon@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2013-04-24 12:19:56 +08:00
lock_page(page);
if (unlikely(!PageUptodate(page))) {
f2fs_put_page(page, 1);
return ERR_PTR(-EIO);
}
if (unlikely(page->mapping != NODE_MAPPING(sbi))) {
f2fs_put_page(page, 1);
goto repeat;
}
page_hit:
f2fs_bug_on(sbi, nid != nid_of_node(page));
return page;
}
struct page *get_node_page(struct f2fs_sb_info *sbi, pgoff_t nid)
{
return __get_node_page(sbi, nid, NULL, 0);
}
struct page *get_node_page_ra(struct page *parent, int start)
{
struct f2fs_sb_info *sbi = F2FS_P_SB(parent);
nid_t nid = get_nid(parent, start, false);
return __get_node_page(sbi, nid, parent, start);
}
void sync_inode_page(struct dnode_of_data *dn)
{
int ret = 0;
if (IS_INODE(dn->node_page) || dn->inode_page == dn->node_page) {
ret = update_inode(dn->inode, dn->node_page);
} else if (dn->inode_page) {
if (!dn->inode_page_locked)
lock_page(dn->inode_page);
ret = update_inode(dn->inode, dn->inode_page);
if (!dn->inode_page_locked)
unlock_page(dn->inode_page);
} else {
ret = update_inode_page(dn->inode);
}
dn->node_changed = ret ? true: false;
}
static void flush_inline_data(struct f2fs_sb_info *sbi, nid_t ino)
{
struct inode *inode;
struct page *page;
/* should flush inline_data before evict_inode */
inode = ilookup(sbi->sb, ino);
if (!inode)
return;
f2fs: fix to avoid deadlock when merging inline data When testing with fsstress, kworker and user threads were both blocked: INFO: task kworker/u16:1:16580 blocked for more than 120 seconds. "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. kworker/u16:1 D ffff8803f2595390 0 16580 2 0x00000000 Workqueue: writeback bdi_writeback_workfn (flush-251:0) ffff8802730e5760 0000000000000046 ffff880274729fc0 0000000000012440 ffff8802730e5fd8 ffff8802730e4010 0000000000012440 0000000000012440 ffff8802730e5fd8 0000000000012440 ffff880274729fc0 ffff88026eb50000 Call Trace: [<ffffffff816fe9d9>] schedule+0x29/0x70 [<ffffffff816ff895>] rwsem_down_read_failed+0xa5/0xf9 [<ffffffff81378584>] call_rwsem_down_read_failed+0x14/0x30 [<ffffffffa0694feb>] f2fs_write_data_page+0x31b/0x420 [f2fs] [<ffffffffa0690f1a>] __f2fs_writepage+0x1a/0x50 [f2fs] [<ffffffffa06922a0>] f2fs_write_data_pages+0xe0/0x290 [f2fs] [<ffffffff811473b3>] do_writepages+0x23/0x40 [<ffffffff811cc3ee>] __writeback_single_inode+0x4e/0x250 [<ffffffff811cd4f1>] writeback_sb_inodes+0x2c1/0x470 [<ffffffff811cd73e>] __writeback_inodes_wb+0x9e/0xd0 [<ffffffff811cda0b>] wb_writeback+0x1fb/0x2d0 [<ffffffff811cdb7c>] wb_do_writeback+0x9c/0x220 [<ffffffff811ce232>] bdi_writeback_workfn+0x72/0x1c0 [<ffffffff8106b74e>] process_one_work+0x1de/0x5b0 [<ffffffff8106e78f>] worker_thread+0x11f/0x3e0 [<ffffffff810750ce>] kthread+0xde/0xf0 [<ffffffff817093f8>] ret_from_fork+0x58/0x90 fsstress thread stack: [<ffffffff81139f0e>] sleep_on_page+0xe/0x20 [<ffffffff81139ef7>] __lock_page+0x67/0x70 [<ffffffff8113b100>] find_lock_page+0x50/0x80 [<ffffffff8113b24f>] find_or_create_page+0x3f/0xb0 [<ffffffffa06983a9>] sync_node_pages+0x259/0x810 [f2fs] [<ffffffffa068d874>] write_checkpoint+0x1a4/0xce0 [f2fs] [<ffffffffa0686b0c>] f2fs_sync_fs+0x7c/0xd0 [f2fs] [<ffffffffa067c813>] f2fs_sync_file+0x143/0x5f0 [f2fs] [<ffffffff811d301b>] vfs_fsync_range+0x2b/0x40 [<ffffffff811d304c>] vfs_fsync+0x1c/0x20 [<ffffffff811d3291>] do_fsync+0x41/0x70 [<ffffffff811d32d3>] SyS_fdatasync+0x13/0x20 [<ffffffff817094a2>] system_call_fastpath+0x16/0x1b [<ffffffffffffffff>] 0xffffffffffffffff The reason of this issue is: CPU0: CPU1: - f2fs_write_data_pages - f2fs_sync_fs - write_checkpoint - block_operations - f2fs_lock_all - down_write(sbi->cp_rwsem) - lock_page(page) - f2fs_write_data_page - sync_node_pages - flush_inline_data - pagecache_get_page(page, GFP_LOCK) - f2fs_lock_op - down_read(sbi->cp_rwsem) This patch alters to use trylock_page in flush_inline_data to fix this ABBA deadlock issue. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2016-02-26 09:33:04 +08:00
page = pagecache_get_page(inode->i_mapping, 0, FGP_NOWAIT, 0);
if (!page)
goto iput_out;
f2fs: fix to avoid deadlock when merging inline data When testing with fsstress, kworker and user threads were both blocked: INFO: task kworker/u16:1:16580 blocked for more than 120 seconds. "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. kworker/u16:1 D ffff8803f2595390 0 16580 2 0x00000000 Workqueue: writeback bdi_writeback_workfn (flush-251:0) ffff8802730e5760 0000000000000046 ffff880274729fc0 0000000000012440 ffff8802730e5fd8 ffff8802730e4010 0000000000012440 0000000000012440 ffff8802730e5fd8 0000000000012440 ffff880274729fc0 ffff88026eb50000 Call Trace: [<ffffffff816fe9d9>] schedule+0x29/0x70 [<ffffffff816ff895>] rwsem_down_read_failed+0xa5/0xf9 [<ffffffff81378584>] call_rwsem_down_read_failed+0x14/0x30 [<ffffffffa0694feb>] f2fs_write_data_page+0x31b/0x420 [f2fs] [<ffffffffa0690f1a>] __f2fs_writepage+0x1a/0x50 [f2fs] [<ffffffffa06922a0>] f2fs_write_data_pages+0xe0/0x290 [f2fs] [<ffffffff811473b3>] do_writepages+0x23/0x40 [<ffffffff811cc3ee>] __writeback_single_inode+0x4e/0x250 [<ffffffff811cd4f1>] writeback_sb_inodes+0x2c1/0x470 [<ffffffff811cd73e>] __writeback_inodes_wb+0x9e/0xd0 [<ffffffff811cda0b>] wb_writeback+0x1fb/0x2d0 [<ffffffff811cdb7c>] wb_do_writeback+0x9c/0x220 [<ffffffff811ce232>] bdi_writeback_workfn+0x72/0x1c0 [<ffffffff8106b74e>] process_one_work+0x1de/0x5b0 [<ffffffff8106e78f>] worker_thread+0x11f/0x3e0 [<ffffffff810750ce>] kthread+0xde/0xf0 [<ffffffff817093f8>] ret_from_fork+0x58/0x90 fsstress thread stack: [<ffffffff81139f0e>] sleep_on_page+0xe/0x20 [<ffffffff81139ef7>] __lock_page+0x67/0x70 [<ffffffff8113b100>] find_lock_page+0x50/0x80 [<ffffffff8113b24f>] find_or_create_page+0x3f/0xb0 [<ffffffffa06983a9>] sync_node_pages+0x259/0x810 [f2fs] [<ffffffffa068d874>] write_checkpoint+0x1a4/0xce0 [f2fs] [<ffffffffa0686b0c>] f2fs_sync_fs+0x7c/0xd0 [f2fs] [<ffffffffa067c813>] f2fs_sync_file+0x143/0x5f0 [f2fs] [<ffffffff811d301b>] vfs_fsync_range+0x2b/0x40 [<ffffffff811d304c>] vfs_fsync+0x1c/0x20 [<ffffffff811d3291>] do_fsync+0x41/0x70 [<ffffffff811d32d3>] SyS_fdatasync+0x13/0x20 [<ffffffff817094a2>] system_call_fastpath+0x16/0x1b [<ffffffffffffffff>] 0xffffffffffffffff The reason of this issue is: CPU0: CPU1: - f2fs_write_data_pages - f2fs_sync_fs - write_checkpoint - block_operations - f2fs_lock_all - down_write(sbi->cp_rwsem) - lock_page(page) - f2fs_write_data_page - sync_node_pages - flush_inline_data - pagecache_get_page(page, GFP_LOCK) - f2fs_lock_op - down_read(sbi->cp_rwsem) This patch alters to use trylock_page in flush_inline_data to fix this ABBA deadlock issue. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2016-02-26 09:33:04 +08:00
if (!trylock_page(page))
goto release_out;
if (!PageUptodate(page))
goto page_out;
if (!PageDirty(page))
goto page_out;
if (!clear_page_dirty_for_io(page))
goto page_out;
if (!f2fs_write_inline_data(inode, page))
inode_dec_dirty_pages(inode);
else
set_page_dirty(page);
page_out:
f2fs: fix to avoid deadlock when merging inline data When testing with fsstress, kworker and user threads were both blocked: INFO: task kworker/u16:1:16580 blocked for more than 120 seconds. "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. kworker/u16:1 D ffff8803f2595390 0 16580 2 0x00000000 Workqueue: writeback bdi_writeback_workfn (flush-251:0) ffff8802730e5760 0000000000000046 ffff880274729fc0 0000000000012440 ffff8802730e5fd8 ffff8802730e4010 0000000000012440 0000000000012440 ffff8802730e5fd8 0000000000012440 ffff880274729fc0 ffff88026eb50000 Call Trace: [<ffffffff816fe9d9>] schedule+0x29/0x70 [<ffffffff816ff895>] rwsem_down_read_failed+0xa5/0xf9 [<ffffffff81378584>] call_rwsem_down_read_failed+0x14/0x30 [<ffffffffa0694feb>] f2fs_write_data_page+0x31b/0x420 [f2fs] [<ffffffffa0690f1a>] __f2fs_writepage+0x1a/0x50 [f2fs] [<ffffffffa06922a0>] f2fs_write_data_pages+0xe0/0x290 [f2fs] [<ffffffff811473b3>] do_writepages+0x23/0x40 [<ffffffff811cc3ee>] __writeback_single_inode+0x4e/0x250 [<ffffffff811cd4f1>] writeback_sb_inodes+0x2c1/0x470 [<ffffffff811cd73e>] __writeback_inodes_wb+0x9e/0xd0 [<ffffffff811cda0b>] wb_writeback+0x1fb/0x2d0 [<ffffffff811cdb7c>] wb_do_writeback+0x9c/0x220 [<ffffffff811ce232>] bdi_writeback_workfn+0x72/0x1c0 [<ffffffff8106b74e>] process_one_work+0x1de/0x5b0 [<ffffffff8106e78f>] worker_thread+0x11f/0x3e0 [<ffffffff810750ce>] kthread+0xde/0xf0 [<ffffffff817093f8>] ret_from_fork+0x58/0x90 fsstress thread stack: [<ffffffff81139f0e>] sleep_on_page+0xe/0x20 [<ffffffff81139ef7>] __lock_page+0x67/0x70 [<ffffffff8113b100>] find_lock_page+0x50/0x80 [<ffffffff8113b24f>] find_or_create_page+0x3f/0xb0 [<ffffffffa06983a9>] sync_node_pages+0x259/0x810 [f2fs] [<ffffffffa068d874>] write_checkpoint+0x1a4/0xce0 [f2fs] [<ffffffffa0686b0c>] f2fs_sync_fs+0x7c/0xd0 [f2fs] [<ffffffffa067c813>] f2fs_sync_file+0x143/0x5f0 [f2fs] [<ffffffff811d301b>] vfs_fsync_range+0x2b/0x40 [<ffffffff811d304c>] vfs_fsync+0x1c/0x20 [<ffffffff811d3291>] do_fsync+0x41/0x70 [<ffffffff811d32d3>] SyS_fdatasync+0x13/0x20 [<ffffffff817094a2>] system_call_fastpath+0x16/0x1b [<ffffffffffffffff>] 0xffffffffffffffff The reason of this issue is: CPU0: CPU1: - f2fs_write_data_pages - f2fs_sync_fs - write_checkpoint - block_operations - f2fs_lock_all - down_write(sbi->cp_rwsem) - lock_page(page) - f2fs_write_data_page - sync_node_pages - flush_inline_data - pagecache_get_page(page, GFP_LOCK) - f2fs_lock_op - down_read(sbi->cp_rwsem) This patch alters to use trylock_page in flush_inline_data to fix this ABBA deadlock issue. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2016-02-26 09:33:04 +08:00
unlock_page(page);
release_out:
f2fs_put_page(page, 0);
iput_out:
iput(inode);
}
int sync_node_pages(struct f2fs_sb_info *sbi, nid_t ino,
struct writeback_control *wbc)
{
pgoff_t index, end;
struct pagevec pvec;
int step = ino ? 2 : 0;
int nwritten = 0;
pagevec_init(&pvec, 0);
next_step:
index = 0;
end = ULONG_MAX;
while (index <= end) {
int i, nr_pages;
nr_pages = pagevec_lookup_tag(&pvec, NODE_MAPPING(sbi), &index,
PAGECACHE_TAG_DIRTY,
min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
if (nr_pages == 0)
break;
for (i = 0; i < nr_pages; i++) {
struct page *page = pvec.pages[i];
if (unlikely(f2fs_cp_error(sbi))) {
pagevec_release(&pvec);
return -EIO;
}
/*
* flushing sequence with step:
* 0. indirect nodes
* 1. dentry dnodes
* 2. file dnodes
*/
if (step == 0 && IS_DNODE(page))
continue;
if (step == 1 && (!IS_DNODE(page) ||
is_cold_node(page)))
continue;
if (step == 2 && (!IS_DNODE(page) ||
!is_cold_node(page)))
continue;
/*
* If an fsync mode,
* we should not skip writing node pages.
*/
lock_node:
if (ino && ino_of_node(page) == ino)
lock_page(page);
else if (!trylock_page(page))
continue;
if (unlikely(page->mapping != NODE_MAPPING(sbi))) {
continue_unlock:
unlock_page(page);
continue;
}
if (ino && ino_of_node(page) != ino)
goto continue_unlock;
if (!PageDirty(page)) {
/* someone wrote it for us */
goto continue_unlock;
}
/* flush inline_data */
if (!ino && is_inline_node(page)) {
clear_inline_node(page);
unlock_page(page);
flush_inline_data(sbi, ino_of_node(page));
goto lock_node;
}
f2fs_wait_on_page_writeback(page, NODE, true);
BUG_ON(PageWriteback(page));
if (!clear_page_dirty_for_io(page))
goto continue_unlock;
/* called by fsync() */
if (ino && IS_DNODE(page)) {
set_fsync_mark(page, 1);
if (IS_INODE(page))
set_dentry_mark(page,
need_dentry_mark(sbi, ino));
nwritten++;
} else {
set_fsync_mark(page, 0);
set_dentry_mark(page, 0);
}
if (NODE_MAPPING(sbi)->a_ops->writepage(page, wbc))
unlock_page(page);
if (--wbc->nr_to_write == 0)
break;
}
pagevec_release(&pvec);
cond_resched();
if (wbc->nr_to_write == 0) {
step = 2;
break;
}
}
if (step < 2) {
step++;
goto next_step;
}
return nwritten;
}
int wait_on_node_pages_writeback(struct f2fs_sb_info *sbi, nid_t ino)
{
pgoff_t index = 0, end = ULONG_MAX;
struct pagevec pvec;
int ret2 = 0, ret = 0;
pagevec_init(&pvec, 0);
while (index <= end) {
int i, nr_pages;
nr_pages = pagevec_lookup_tag(&pvec, NODE_MAPPING(sbi), &index,
PAGECACHE_TAG_WRITEBACK,
min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
if (nr_pages == 0)
break;
for (i = 0; i < nr_pages; i++) {
struct page *page = pvec.pages[i];
/* until radix tree lookup accepts end_index */
if (unlikely(page->index > end))
continue;
if (ino && ino_of_node(page) == ino) {
f2fs_wait_on_page_writeback(page, NODE, true);
if (TestClearPageError(page))
ret = -EIO;
}
}
pagevec_release(&pvec);
cond_resched();
}
if (unlikely(test_and_clear_bit(AS_ENOSPC, &NODE_MAPPING(sbi)->flags)))
ret2 = -ENOSPC;
if (unlikely(test_and_clear_bit(AS_EIO, &NODE_MAPPING(sbi)->flags)))
ret2 = -EIO;
if (!ret)
ret = ret2;
return ret;
}
static int f2fs_write_node_page(struct page *page,
struct writeback_control *wbc)
{
struct f2fs_sb_info *sbi = F2FS_P_SB(page);
nid_t nid;
struct node_info ni;
struct f2fs_io_info fio = {
.sbi = sbi,
.type = NODE,
.rw = (wbc->sync_mode == WB_SYNC_ALL) ? WRITE_SYNC : WRITE,
.page = page,
.encrypted_page = NULL,
};
trace_f2fs_writepage(page, NODE);
if (unlikely(is_sbi_flag_set(sbi, SBI_POR_DOING)))
goto redirty_out;
if (unlikely(f2fs_cp_error(sbi)))
goto redirty_out;
/* get old block addr of this node page */
nid = nid_of_node(page);
f2fs_bug_on(sbi, page->index != nid);
if (wbc->for_reclaim) {
if (!down_read_trylock(&sbi->node_write))
goto redirty_out;
} else {
down_read(&sbi->node_write);
}
get_node_info(sbi, nid, &ni);
/* This page is already truncated */
if (unlikely(ni.blk_addr == NULL_ADDR)) {
ClearPageUptodate(page);
f2fs: introduce a new global lock scheme In the previous version, f2fs uses global locks according to the usage types, such as directory operations, block allocation, block write, and so on. Reference the following lock types in f2fs.h. enum lock_type { RENAME, /* for renaming operations */ DENTRY_OPS, /* for directory operations */ DATA_WRITE, /* for data write */ DATA_NEW, /* for data allocation */ DATA_TRUNC, /* for data truncate */ NODE_NEW, /* for node allocation */ NODE_TRUNC, /* for node truncate */ NODE_WRITE, /* for node write */ NR_LOCK_TYPE, }; In that case, we lose the performance under the multi-threading environment, since every types of operations must be conducted one at a time. In order to address the problem, let's share the locks globally with a mutex array regardless of any types. So, let users grab a mutex and perform their jobs in parallel as much as possbile. For this, I propose a new global lock scheme as follows. 0. Data structure - f2fs_sb_info -> mutex_lock[NR_GLOBAL_LOCKS] - f2fs_sb_info -> node_write 1. mutex_lock_op(sbi) - try to get an avaiable lock from the array. - returns the index of the gottern lock variable. 2. mutex_unlock_op(sbi, index of the lock) - unlock the given index of the lock. 3. mutex_lock_all(sbi) - grab all the locks in the array before the checkpoint. 4. mutex_unlock_all(sbi) - release all the locks in the array after checkpoint. 5. block_operations() - call mutex_lock_all() - sync_dirty_dir_inodes() - grab node_write - sync_node_pages() Note that, the pairs of mutex_lock_op()/mutex_unlock_op() and mutex_lock_all()/mutex_unlock_all() should be used together. Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2012-11-22 15:21:29 +08:00
dec_page_count(sbi, F2FS_DIRTY_NODES);
up_read(&sbi->node_write);
f2fs: introduce a new global lock scheme In the previous version, f2fs uses global locks according to the usage types, such as directory operations, block allocation, block write, and so on. Reference the following lock types in f2fs.h. enum lock_type { RENAME, /* for renaming operations */ DENTRY_OPS, /* for directory operations */ DATA_WRITE, /* for data write */ DATA_NEW, /* for data allocation */ DATA_TRUNC, /* for data truncate */ NODE_NEW, /* for node allocation */ NODE_TRUNC, /* for node truncate */ NODE_WRITE, /* for node write */ NR_LOCK_TYPE, }; In that case, we lose the performance under the multi-threading environment, since every types of operations must be conducted one at a time. In order to address the problem, let's share the locks globally with a mutex array regardless of any types. So, let users grab a mutex and perform their jobs in parallel as much as possbile. For this, I propose a new global lock scheme as follows. 0. Data structure - f2fs_sb_info -> mutex_lock[NR_GLOBAL_LOCKS] - f2fs_sb_info -> node_write 1. mutex_lock_op(sbi) - try to get an avaiable lock from the array. - returns the index of the gottern lock variable. 2. mutex_unlock_op(sbi, index of the lock) - unlock the given index of the lock. 3. mutex_lock_all(sbi) - grab all the locks in the array before the checkpoint. 4. mutex_unlock_all(sbi) - release all the locks in the array after checkpoint. 5. block_operations() - call mutex_lock_all() - sync_dirty_dir_inodes() - grab node_write - sync_node_pages() Note that, the pairs of mutex_lock_op()/mutex_unlock_op() and mutex_lock_all()/mutex_unlock_all() should be used together. Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2012-11-22 15:21:29 +08:00
unlock_page(page);
return 0;
}
set_page_writeback(page);
fio.old_blkaddr = ni.blk_addr;
write_node_page(nid, &fio);
set_node_addr(sbi, &ni, fio.new_blkaddr, is_fsync_dnode(page));
dec_page_count(sbi, F2FS_DIRTY_NODES);
up_read(&sbi->node_write);
if (wbc->for_reclaim)
f2fs_submit_merged_bio_cond(sbi, NULL, page, 0, NODE, WRITE);
unlock_page(page);
if (unlikely(f2fs_cp_error(sbi)))
f2fs_submit_merged_bio(sbi, NODE, WRITE);
return 0;
redirty_out:
redirty_page_for_writepage(wbc, page);
return AOP_WRITEPAGE_ACTIVATE;
}
static int f2fs_write_node_pages(struct address_space *mapping,
struct writeback_control *wbc)
{
struct f2fs_sb_info *sbi = F2FS_M_SB(mapping);
long diff;
/* balancing f2fs's metadata in background */
f2fs_balance_fs_bg(sbi);
/* collect a number of dirty node pages and write together */
if (get_pages(sbi, F2FS_DIRTY_NODES) < nr_pages_to_skip(sbi, NODE))
goto skip_write;
trace_f2fs_writepages(mapping->host, wbc, NODE);
diff = nr_pages_to_write(sbi, NODE, wbc);
wbc->sync_mode = WB_SYNC_NONE;
sync_node_pages(sbi, 0, wbc);
wbc->nr_to_write = max((long)0, wbc->nr_to_write - diff);
return 0;
skip_write:
wbc->pages_skipped += get_pages(sbi, F2FS_DIRTY_NODES);
trace_f2fs_writepages(mapping->host, wbc, NODE);
return 0;
}
static int f2fs_set_node_page_dirty(struct page *page)
{
trace_f2fs_set_page_dirty(page, NODE);
SetPageUptodate(page);
if (!PageDirty(page)) {
__set_page_dirty_nobuffers(page);
inc_page_count(F2FS_P_SB(page), F2FS_DIRTY_NODES);
SetPagePrivate(page);
f2fs_trace_pid(page);
return 1;
}
return 0;
}
/*
* Structure of the f2fs node operations
*/
const struct address_space_operations f2fs_node_aops = {
.writepage = f2fs_write_node_page,
.writepages = f2fs_write_node_pages,
.set_page_dirty = f2fs_set_node_page_dirty,
.invalidatepage = f2fs_invalidate_page,
.releasepage = f2fs_release_page,
};
static struct free_nid *__lookup_free_nid_list(struct f2fs_nm_info *nm_i,
nid_t n)
{
return radix_tree_lookup(&nm_i->free_nid_root, n);
}
static void __del_from_free_nid_list(struct f2fs_nm_info *nm_i,
struct free_nid *i)
{
list_del(&i->list);
radix_tree_delete(&nm_i->free_nid_root, i->nid);
}
static int add_free_nid(struct f2fs_sb_info *sbi, nid_t nid, bool build)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct free_nid *i;
struct nat_entry *ne;
bool allocated = false;
if (!available_free_memory(sbi, FREE_NIDS))
return -1;
/* 0 nid should not be used */
if (unlikely(nid == 0))
return 0;
if (build) {
/* do not add allocated nids */
ne = __lookup_nat_cache(nm_i, nid);
if (ne && (!get_nat_flag(ne, IS_CHECKPOINTED) ||
nat_get_blkaddr(ne) != NULL_ADDR))
allocated = true;
if (allocated)
return 0;
}
i = f2fs_kmem_cache_alloc(free_nid_slab, GFP_NOFS);
i->nid = nid;
i->state = NID_NEW;
if (radix_tree_preload(GFP_NOFS)) {
kmem_cache_free(free_nid_slab, i);
return 0;
}
spin_lock(&nm_i->free_nid_list_lock);
if (radix_tree_insert(&nm_i->free_nid_root, i->nid, i)) {
spin_unlock(&nm_i->free_nid_list_lock);
radix_tree_preload_end();
kmem_cache_free(free_nid_slab, i);
return 0;
}
list_add_tail(&i->list, &nm_i->free_nid_list);
nm_i->fcnt++;
spin_unlock(&nm_i->free_nid_list_lock);
radix_tree_preload_end();
return 1;
}
static void remove_free_nid(struct f2fs_nm_info *nm_i, nid_t nid)
{
struct free_nid *i;
bool need_free = false;
spin_lock(&nm_i->free_nid_list_lock);
i = __lookup_free_nid_list(nm_i, nid);
if (i && i->state == NID_NEW) {
__del_from_free_nid_list(nm_i, i);
nm_i->fcnt--;
need_free = true;
}
spin_unlock(&nm_i->free_nid_list_lock);
if (need_free)
kmem_cache_free(free_nid_slab, i);
}
static void scan_nat_page(struct f2fs_sb_info *sbi,
struct page *nat_page, nid_t start_nid)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct f2fs_nat_block *nat_blk = page_address(nat_page);
block_t blk_addr;
int i;
i = start_nid % NAT_ENTRY_PER_BLOCK;
for (; i < NAT_ENTRY_PER_BLOCK; i++, start_nid++) {
if (unlikely(start_nid >= nm_i->max_nid))
break;
blk_addr = le32_to_cpu(nat_blk->entries[i].block_addr);
f2fs_bug_on(sbi, blk_addr == NEW_ADDR);
if (blk_addr == NULL_ADDR) {
if (add_free_nid(sbi, start_nid, true) < 0)
break;
}
}
}
static void build_free_nids(struct f2fs_sb_info *sbi)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_HOT_DATA);
struct f2fs_journal *journal = curseg->journal;
int i = 0;
nid_t nid = nm_i->next_scan_nid;
/* Enough entries */
if (nm_i->fcnt > NAT_ENTRY_PER_BLOCK)
return;
/* readahead nat pages to be scanned */
ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), FREE_NID_PAGES,
META_NAT, true);
down_read(&nm_i->nat_tree_lock);
while (1) {
struct page *page = get_current_nat_page(sbi, nid);
scan_nat_page(sbi, page, nid);
f2fs_put_page(page, 1);
nid += (NAT_ENTRY_PER_BLOCK - (nid % NAT_ENTRY_PER_BLOCK));
if (unlikely(nid >= nm_i->max_nid))
nid = 0;
if (++i >= FREE_NID_PAGES)
break;
}
/* go to the next free nat pages to find free nids abundantly */
nm_i->next_scan_nid = nid;
/* find free nids from current sum_pages */
down_read(&curseg->journal_rwsem);
for (i = 0; i < nats_in_cursum(journal); i++) {
block_t addr;
addr = le32_to_cpu(nat_in_journal(journal, i).block_addr);
nid = le32_to_cpu(nid_in_journal(journal, i));
if (addr == NULL_ADDR)
add_free_nid(sbi, nid, true);
else
remove_free_nid(nm_i, nid);
}
up_read(&curseg->journal_rwsem);
up_read(&nm_i->nat_tree_lock);
ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nm_i->next_scan_nid),
nm_i->ra_nid_pages, META_NAT, false);
}
/*
* If this function returns success, caller can obtain a new nid
* from second parameter of this function.
* The returned nid could be used ino as well as nid when inode is created.
*/
bool alloc_nid(struct f2fs_sb_info *sbi, nid_t *nid)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct free_nid *i = NULL;
retry:
if (unlikely(sbi->total_valid_node_count + 1 > nm_i->available_nids))
return false;
spin_lock(&nm_i->free_nid_list_lock);
/* We should not use stale free nids created by build_free_nids */
if (nm_i->fcnt && !on_build_free_nids(nm_i)) {
f2fs_bug_on(sbi, list_empty(&nm_i->free_nid_list));
list_for_each_entry(i, &nm_i->free_nid_list, list)
if (i->state == NID_NEW)
break;
f2fs_bug_on(sbi, i->state != NID_NEW);
*nid = i->nid;
i->state = NID_ALLOC;
nm_i->fcnt--;
spin_unlock(&nm_i->free_nid_list_lock);
return true;
}
spin_unlock(&nm_i->free_nid_list_lock);
/* Let's scan nat pages and its caches to get free nids */
mutex_lock(&nm_i->build_lock);
build_free_nids(sbi);
mutex_unlock(&nm_i->build_lock);
goto retry;
}
/*
* alloc_nid() should be called prior to this function.
*/
void alloc_nid_done(struct f2fs_sb_info *sbi, nid_t nid)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct free_nid *i;
spin_lock(&nm_i->free_nid_list_lock);
i = __lookup_free_nid_list(nm_i, nid);
f2fs_bug_on(sbi, !i || i->state != NID_ALLOC);
__del_from_free_nid_list(nm_i, i);
spin_unlock(&nm_i->free_nid_list_lock);
kmem_cache_free(free_nid_slab, i);
}
/*
* alloc_nid() should be called prior to this function.
*/
void alloc_nid_failed(struct f2fs_sb_info *sbi, nid_t nid)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct free_nid *i;
bool need_free = false;
if (!nid)
return;
spin_lock(&nm_i->free_nid_list_lock);
i = __lookup_free_nid_list(nm_i, nid);
f2fs_bug_on(sbi, !i || i->state != NID_ALLOC);
if (!available_free_memory(sbi, FREE_NIDS)) {
__del_from_free_nid_list(nm_i, i);
need_free = true;
} else {
i->state = NID_NEW;
nm_i->fcnt++;
}
spin_unlock(&nm_i->free_nid_list_lock);
if (need_free)
kmem_cache_free(free_nid_slab, i);
}
int try_to_free_nids(struct f2fs_sb_info *sbi, int nr_shrink)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct free_nid *i, *next;
int nr = nr_shrink;
if (!mutex_trylock(&nm_i->build_lock))
return 0;
spin_lock(&nm_i->free_nid_list_lock);
list_for_each_entry_safe(i, next, &nm_i->free_nid_list, list) {
if (nr_shrink <= 0 || nm_i->fcnt <= NAT_ENTRY_PER_BLOCK)
break;
if (i->state == NID_ALLOC)
continue;
__del_from_free_nid_list(nm_i, i);
kmem_cache_free(free_nid_slab, i);
nm_i->fcnt--;
nr_shrink--;
}
spin_unlock(&nm_i->free_nid_list_lock);
mutex_unlock(&nm_i->build_lock);
return nr - nr_shrink;
}
void recover_inline_xattr(struct inode *inode, struct page *page)
{
void *src_addr, *dst_addr;
size_t inline_size;
struct page *ipage;
struct f2fs_inode *ri;
ipage = get_node_page(F2FS_I_SB(inode), inode->i_ino);
f2fs_bug_on(F2FS_I_SB(inode), IS_ERR(ipage));
ri = F2FS_INODE(page);
if (!(ri->i_inline & F2FS_INLINE_XATTR)) {
clear_inode_flag(F2FS_I(inode), FI_INLINE_XATTR);
goto update_inode;
}
dst_addr = inline_xattr_addr(ipage);
src_addr = inline_xattr_addr(page);
inline_size = inline_xattr_size(inode);
f2fs_wait_on_page_writeback(ipage, NODE, true);
memcpy(dst_addr, src_addr, inline_size);
update_inode:
update_inode(inode, ipage);
f2fs_put_page(ipage, 1);
}
void recover_xattr_data(struct inode *inode, struct page *page, block_t blkaddr)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
nid_t prev_xnid = F2FS_I(inode)->i_xattr_nid;
nid_t new_xnid = nid_of_node(page);
struct node_info ni;
/* 1: invalidate the previous xattr nid */
if (!prev_xnid)
goto recover_xnid;
/* Deallocate node address */
get_node_info(sbi, prev_xnid, &ni);
f2fs_bug_on(sbi, ni.blk_addr == NULL_ADDR);
invalidate_blocks(sbi, ni.blk_addr);
dec_valid_node_count(sbi, inode);
set_node_addr(sbi, &ni, NULL_ADDR, false);
recover_xnid:
/* 2: allocate new xattr nid */
if (unlikely(!inc_valid_node_count(sbi, inode)))
f2fs_bug_on(sbi, 1);
remove_free_nid(NM_I(sbi), new_xnid);
get_node_info(sbi, new_xnid, &ni);
ni.ino = inode->i_ino;
set_node_addr(sbi, &ni, NEW_ADDR, false);
F2FS_I(inode)->i_xattr_nid = new_xnid;
/* 3: update xattr blkaddr */
refresh_sit_entry(sbi, NEW_ADDR, blkaddr);
set_node_addr(sbi, &ni, blkaddr, false);
update_inode_page(inode);
}
int recover_inode_page(struct f2fs_sb_info *sbi, struct page *page)
{
struct f2fs_inode *src, *dst;
nid_t ino = ino_of_node(page);
struct node_info old_ni, new_ni;
struct page *ipage;
get_node_info(sbi, ino, &old_ni);
if (unlikely(old_ni.blk_addr != NULL_ADDR))
return -EINVAL;
ipage = grab_cache_page(NODE_MAPPING(sbi), ino);
if (!ipage)
return -ENOMEM;
/* Should not use this inode from free nid list */
remove_free_nid(NM_I(sbi), ino);
SetPageUptodate(ipage);
fill_node_footer(ipage, ino, ino, 0, true);
src = F2FS_INODE(page);
dst = F2FS_INODE(ipage);
memcpy(dst, src, (unsigned long)&src->i_ext - (unsigned long)src);
dst->i_size = 0;
dst->i_blocks = cpu_to_le64(1);
dst->i_links = cpu_to_le32(1);
dst->i_xattr_nid = 0;
dst->i_inline = src->i_inline & F2FS_INLINE_XATTR;
new_ni = old_ni;
new_ni.ino = ino;
if (unlikely(!inc_valid_node_count(sbi, NULL)))
WARN_ON(1);
set_node_addr(sbi, &new_ni, NEW_ADDR, false);
inc_valid_inode_count(sbi);
set_page_dirty(ipage);
f2fs_put_page(ipage, 1);
return 0;
}
int restore_node_summary(struct f2fs_sb_info *sbi,
unsigned int segno, struct f2fs_summary_block *sum)
{
struct f2fs_node *rn;
struct f2fs_summary *sum_entry;
block_t addr;
int bio_blocks = MAX_BIO_BLOCKS(sbi);
int i, idx, last_offset, nrpages;
/* scan the node segment */
last_offset = sbi->blocks_per_seg;
addr = START_BLOCK(sbi, segno);
sum_entry = &sum->entries[0];
for (i = 0; i < last_offset; i += nrpages, addr += nrpages) {
nrpages = min(last_offset - i, bio_blocks);
/* readahead node pages */
ra_meta_pages(sbi, addr, nrpages, META_POR, true);
for (idx = addr; idx < addr + nrpages; idx++) {
struct page *page = get_tmp_page(sbi, idx);
rn = F2FS_NODE(page);
sum_entry->nid = rn->footer.nid;
sum_entry->version = 0;
sum_entry->ofs_in_node = 0;
sum_entry++;
f2fs_put_page(page, 1);
}
invalidate_mapping_pages(META_MAPPING(sbi), addr,
addr + nrpages);
}
return 0;
}
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
static void remove_nats_in_journal(struct f2fs_sb_info *sbi)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_HOT_DATA);
struct f2fs_journal *journal = curseg->journal;
int i;
down_write(&curseg->journal_rwsem);
for (i = 0; i < nats_in_cursum(journal); i++) {
struct nat_entry *ne;
struct f2fs_nat_entry raw_ne;
nid_t nid = le32_to_cpu(nid_in_journal(journal, i));
raw_ne = nat_in_journal(journal, i);
ne = __lookup_nat_cache(nm_i, nid);
if (!ne) {
ne = grab_nat_entry(nm_i, nid);
node_info_from_raw_nat(&ne->ni, &raw_ne);
}
__set_nat_cache_dirty(nm_i, ne);
}
update_nats_in_cursum(journal, -i);
up_write(&curseg->journal_rwsem);
}
static void __adjust_nat_entry_set(struct nat_entry_set *nes,
struct list_head *head, int max)
{
struct nat_entry_set *cur;
if (nes->entry_cnt >= max)
goto add_out;
list_for_each_entry(cur, head, set_list) {
if (cur->entry_cnt >= nes->entry_cnt) {
list_add(&nes->set_list, cur->set_list.prev);
return;
}
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
}
add_out:
list_add_tail(&nes->set_list, head);
}
static void __flush_nat_entry_set(struct f2fs_sb_info *sbi,
struct nat_entry_set *set)
{
struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_HOT_DATA);
struct f2fs_journal *journal = curseg->journal;
nid_t start_nid = set->set * NAT_ENTRY_PER_BLOCK;
bool to_journal = true;
struct f2fs_nat_block *nat_blk;
struct nat_entry *ne, *cur;
struct page *page = NULL;
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
/*
* there are two steps to flush nat entries:
* #1, flush nat entries to journal in current hot data summary block.
* #2, flush nat entries to nat page.
*/
if (!__has_cursum_space(journal, set->entry_cnt, NAT_JOURNAL))
to_journal = false;
if (to_journal) {
down_write(&curseg->journal_rwsem);
} else {
page = get_next_nat_page(sbi, start_nid);
nat_blk = page_address(page);
f2fs_bug_on(sbi, !nat_blk);
}
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
/* flush dirty nats in nat entry set */
list_for_each_entry_safe(ne, cur, &set->entry_list, list) {
struct f2fs_nat_entry *raw_ne;
nid_t nid = nat_get_nid(ne);
int offset;
if (nat_get_blkaddr(ne) == NEW_ADDR)
continue;
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
if (to_journal) {
offset = lookup_journal_in_cursum(journal,
NAT_JOURNAL, nid, 1);
f2fs_bug_on(sbi, offset < 0);
raw_ne = &nat_in_journal(journal, offset);
nid_in_journal(journal, offset) = cpu_to_le32(nid);
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
} else {
raw_ne = &nat_blk->entries[nid - start_nid];
}
raw_nat_from_node_info(raw_ne, &ne->ni);
nat_reset_flag(ne);
__clear_nat_cache_dirty(NM_I(sbi), ne);
if (nat_get_blkaddr(ne) == NULL_ADDR)
add_free_nid(sbi, nid, false);
}
if (to_journal)
up_write(&curseg->journal_rwsem);
else
f2fs_put_page(page, 1);
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
f2fs_bug_on(sbi, set->entry_cnt);
radix_tree_delete(&NM_I(sbi)->nat_set_root, set->set);
kmem_cache_free(nat_entry_set_slab, set);
}
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
/*
* This function is called during the checkpointing process.
*/
void flush_nat_entries(struct f2fs_sb_info *sbi)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_HOT_DATA);
struct f2fs_journal *journal = curseg->journal;
struct nat_entry_set *setvec[SETVEC_SIZE];
struct nat_entry_set *set, *tmp;
unsigned int found;
nid_t set_idx = 0;
LIST_HEAD(sets);
if (!nm_i->dirty_nat_cnt)
return;
down_write(&nm_i->nat_tree_lock);
/*
* if there are no enough space in journal to store dirty nat
* entries, remove all entries from journal and merge them
* into nat entry set.
*/
if (!__has_cursum_space(journal, nm_i->dirty_nat_cnt, NAT_JOURNAL))
remove_nats_in_journal(sbi);
while ((found = __gang_lookup_nat_set(nm_i,
set_idx, SETVEC_SIZE, setvec))) {
unsigned idx;
set_idx = setvec[found - 1]->set + 1;
for (idx = 0; idx < found; idx++)
__adjust_nat_entry_set(setvec[idx], &sets,
MAX_NAT_JENTRIES(journal));
}
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
/* flush dirty nats in nat entry set */
list_for_each_entry_safe(set, tmp, &sets, set_list)
__flush_nat_entry_set(sbi, set);
up_write(&nm_i->nat_tree_lock);
f2fs_bug_on(sbi, nm_i->dirty_nat_cnt);
}
static int init_node_manager(struct f2fs_sb_info *sbi)
{
struct f2fs_super_block *sb_raw = F2FS_RAW_SUPER(sbi);
struct f2fs_nm_info *nm_i = NM_I(sbi);
unsigned char *version_bitmap;
unsigned int nat_segs, nat_blocks;
nm_i->nat_blkaddr = le32_to_cpu(sb_raw->nat_blkaddr);
/* segment_count_nat includes pair segment so divide to 2. */
nat_segs = le32_to_cpu(sb_raw->segment_count_nat) >> 1;
nat_blocks = nat_segs << le32_to_cpu(sb_raw->log_blocks_per_seg);
nm_i->max_nid = NAT_ENTRY_PER_BLOCK * nat_blocks;
/* not used nids: 0, node, meta, (and root counted as valid node) */
nm_i->available_nids = nm_i->max_nid - F2FS_RESERVED_NODE_NUM;
nm_i->fcnt = 0;
nm_i->nat_cnt = 0;
nm_i->ram_thresh = DEF_RAM_THRESHOLD;
nm_i->ra_nid_pages = DEF_RA_NID_PAGES;
nm_i->dirty_nats_ratio = DEF_DIRTY_NAT_RATIO_THRESHOLD;
INIT_RADIX_TREE(&nm_i->free_nid_root, GFP_ATOMIC);
INIT_LIST_HEAD(&nm_i->free_nid_list);
INIT_RADIX_TREE(&nm_i->nat_root, GFP_NOIO);
INIT_RADIX_TREE(&nm_i->nat_set_root, GFP_NOIO);
INIT_LIST_HEAD(&nm_i->nat_entries);
mutex_init(&nm_i->build_lock);
spin_lock_init(&nm_i->free_nid_list_lock);
init_rwsem(&nm_i->nat_tree_lock);
nm_i->next_scan_nid = le32_to_cpu(sbi->ckpt->next_free_nid);
nm_i->bitmap_size = __bitmap_size(sbi, NAT_BITMAP);
version_bitmap = __bitmap_ptr(sbi, NAT_BITMAP);
if (!version_bitmap)
return -EFAULT;
nm_i->nat_bitmap = kmemdup(version_bitmap, nm_i->bitmap_size,
GFP_KERNEL);
if (!nm_i->nat_bitmap)
return -ENOMEM;
return 0;
}
int build_node_manager(struct f2fs_sb_info *sbi)
{
int err;
sbi->nm_info = kzalloc(sizeof(struct f2fs_nm_info), GFP_KERNEL);
if (!sbi->nm_info)
return -ENOMEM;
err = init_node_manager(sbi);
if (err)
return err;
build_free_nids(sbi);
return 0;
}
void destroy_node_manager(struct f2fs_sb_info *sbi)
{
struct f2fs_nm_info *nm_i = NM_I(sbi);
struct free_nid *i, *next_i;
struct nat_entry *natvec[NATVEC_SIZE];
struct nat_entry_set *setvec[SETVEC_SIZE];
nid_t nid = 0;
unsigned int found;
if (!nm_i)
return;
/* destroy free nid list */
spin_lock(&nm_i->free_nid_list_lock);
list_for_each_entry_safe(i, next_i, &nm_i->free_nid_list, list) {
f2fs_bug_on(sbi, i->state == NID_ALLOC);
__del_from_free_nid_list(nm_i, i);
nm_i->fcnt--;
spin_unlock(&nm_i->free_nid_list_lock);
kmem_cache_free(free_nid_slab, i);
spin_lock(&nm_i->free_nid_list_lock);
}
f2fs_bug_on(sbi, nm_i->fcnt);
spin_unlock(&nm_i->free_nid_list_lock);
/* destroy nat cache */
down_write(&nm_i->nat_tree_lock);
while ((found = __gang_lookup_nat_cache(nm_i,
nid, NATVEC_SIZE, natvec))) {
unsigned idx;
nid = nat_get_nid(natvec[found - 1]) + 1;
for (idx = 0; idx < found; idx++)
__del_from_nat_cache(nm_i, natvec[idx]);
}
f2fs_bug_on(sbi, nm_i->nat_cnt);
/* destroy nat set cache */
nid = 0;
while ((found = __gang_lookup_nat_set(nm_i,
nid, SETVEC_SIZE, setvec))) {
unsigned idx;
nid = setvec[found - 1]->set + 1;
for (idx = 0; idx < found; idx++) {
/* entry_cnt is not zero, when cp_error was occurred */
f2fs_bug_on(sbi, !list_empty(&setvec[idx]->entry_list));
radix_tree_delete(&nm_i->nat_set_root, setvec[idx]->set);
kmem_cache_free(nat_entry_set_slab, setvec[idx]);
}
}
up_write(&nm_i->nat_tree_lock);
kfree(nm_i->nat_bitmap);
sbi->nm_info = NULL;
kfree(nm_i);
}
int __init create_node_manager_caches(void)
{
nat_entry_slab = f2fs_kmem_cache_create("nat_entry",
sizeof(struct nat_entry));
if (!nat_entry_slab)
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
goto fail;
free_nid_slab = f2fs_kmem_cache_create("free_nid",
sizeof(struct free_nid));
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
if (!free_nid_slab)
goto destroy_nat_entry;
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
nat_entry_set_slab = f2fs_kmem_cache_create("nat_entry_set",
sizeof(struct nat_entry_set));
if (!nat_entry_set_slab)
goto destroy_free_nid;
return 0;
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
destroy_free_nid:
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
kmem_cache_destroy(free_nid_slab);
destroy_nat_entry:
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
kmem_cache_destroy(nat_entry_slab);
fail:
return -ENOMEM;
}
void destroy_node_manager_caches(void)
{
f2fs: refactor flush_nat_entries codes for reducing NAT writes Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time. In this patch we merge dirty entries located in same NAT block to nat entry set, and linked all set to list, sorted ascending order by entries' count of set. Later we flush entries in sparse set into journal as many as we can, and then flush merged entries to disk. In this way we can not only gain in performance, but also save lifetime of flash device. In my testing environment, it shows this patch can help to reduce NAT block writes obviously. In hard disk test case: cost time of fsstress is stablely reduced by about 5%. 1. virtual machine + hard disk: fsstress -p 20 -n 200 -l 5 node num cp count nodes/cp based 4599.6 1803.0 2.551 patched 2714.6 1829.6 1.483 2. virtual machine + 32g micro SD card: fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0 -f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5 -f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S node num cp count nodes/cp based 84.5 43.7 1.933 patched 49.2 40.0 1.23 Our latency of merging op shows not bad when handling extreme case like: merging a great number of dirty nats: latency(ns) dirty nat count 3089219 24922 5129423 27422 4000250 24523 change log from v1: o fix wrong logic in add_nat_entry when grab a new nat entry set. o swith to create slab cache in create_node_manager_caches. o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency. change log from v2: o make comment position more appropriate suggested by Jaegeuk Kim. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-06-24 09:18:20 +08:00
kmem_cache_destroy(nat_entry_set_slab);
kmem_cache_destroy(free_nid_slab);
kmem_cache_destroy(nat_entry_slab);
}