Having file extent items with ranges that overlap each other is a
serious issue that leads to all sorts of corruptions and crashes (like a
BUG_ON() during the course of __btrfs_drop_extents() when it traims file
extent items). Therefore teach the tree checker to detect such cases.
This is motivated by a recently fixed bug (race between ranged full
fsync and writeback or adjacent ranges).
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When we do a full fsync (the bit BTRFS_INODE_NEEDS_FULL_SYNC is set in the
inode) that happens to be ranged, which happens during a msync() or writes
for files opened with O_SYNC for example, we can end up with a corrupt log,
due to different file extent items representing ranges that overlap with
each other, or hit some assertion failures.
When doing a ranged fsync we only flush delalloc and wait for ordered
exents within that range. If while we are logging items from our inode
ordered extents for adjacent ranges complete, we end up in a race that can
make us insert the file extent items that overlap with others we logged
previously and the assertion failures.
For example, if tree-log.c:copy_items() receives a leaf that has the
following file extents items, all with a length of 4K and therefore there
is an implicit hole in the range 68K to 72K - 1:
(257 EXTENT_ITEM 64K), (257 EXTENT_ITEM 72K), (257 EXTENT_ITEM 76K), ...
It copies them to the log tree. However due to the need to detect implicit
holes, it may release the path, in order to look at the previous leaf to
detect an implicit hole, and then later it will search again in the tree
for the first file extent item key, with the goal of locking again the
leaf (which might have changed due to concurrent changes to other inodes).
However when it locks again the leaf containing the first key, the key
corresponding to the extent at offset 72K may not be there anymore since
there is an ordered extent for that range that is finishing (that is,
somewhere in the middle of btrfs_finish_ordered_io()), and it just
removed the file extent item but has not yet replaced it with a new file
extent item, so the part of copy_items() that does hole detection will
decide that there is a hole in the range starting from 68K to 76K - 1,
and therefore insert a file extent item to represent that hole, having
a key offset of 68K. After that we now have a log tree with 2 different
extent items that have overlapping ranges:
1) The file extent item copied before copy_items() released the path,
which has a key offset of 72K and a length of 4K, representing the
file range 72K to 76K - 1.
2) And a file extent item representing a hole that has a key offset of
68K and a length of 8K, representing the range 68K to 76K - 1. This
item was inserted after releasing the path, and overlaps with the
extent item inserted before.
The overlapping extent items can cause all sorts of unpredictable and
incorrect behaviour, either when replayed or if a fast (non full) fsync
happens later, which can trigger a BUG_ON() when calling
btrfs_set_item_key_safe() through __btrfs_drop_extents(), producing a
trace like the following:
[61666.783269] ------------[ cut here ]------------
[61666.783943] kernel BUG at fs/btrfs/ctree.c:3182!
[61666.784644] invalid opcode: 0000 [#1] PREEMPT SMP
(...)
[61666.786253] task: ffff880117b88c40 task.stack: ffffc90008168000
[61666.786253] RIP: 0010:btrfs_set_item_key_safe+0x7c/0xd2 [btrfs]
[61666.786253] RSP: 0018:ffffc9000816b958 EFLAGS: 00010246
[61666.786253] RAX: 0000000000000000 RBX: 000000000000000f RCX: 0000000000030000
[61666.786253] RDX: 0000000000000000 RSI: ffffc9000816ba4f RDI: ffffc9000816b937
[61666.786253] RBP: ffffc9000816b998 R08: ffff88011dae2428 R09: 0000000000001000
[61666.786253] R10: 0000160000000000 R11: 6db6db6db6db6db7 R12: ffff88011dae2418
[61666.786253] R13: ffffc9000816ba4f R14: ffff8801e10c4118 R15: ffff8801e715c000
[61666.786253] FS: 00007f6060a18700(0000) GS:ffff88023f5c0000(0000) knlGS:0000000000000000
[61666.786253] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[61666.786253] CR2: 00007f6060a28000 CR3: 0000000213e69000 CR4: 00000000000006e0
[61666.786253] Call Trace:
[61666.786253] __btrfs_drop_extents+0x5e3/0xaad [btrfs]
[61666.786253] ? time_hardirqs_on+0x9/0x14
[61666.786253] btrfs_log_changed_extents+0x294/0x4e0 [btrfs]
[61666.786253] ? release_extent_buffer+0x38/0xb4 [btrfs]
[61666.786253] btrfs_log_inode+0xb6e/0xcdc [btrfs]
[61666.786253] ? lock_acquire+0x131/0x1c5
[61666.786253] ? btrfs_log_inode_parent+0xee/0x659 [btrfs]
[61666.786253] ? arch_local_irq_save+0x9/0xc
[61666.786253] ? btrfs_log_inode_parent+0x1f5/0x659 [btrfs]
[61666.786253] btrfs_log_inode_parent+0x223/0x659 [btrfs]
[61666.786253] ? arch_local_irq_save+0x9/0xc
[61666.786253] ? lockref_get_not_zero+0x2c/0x34
[61666.786253] ? rcu_read_unlock+0x3e/0x5d
[61666.786253] btrfs_log_dentry_safe+0x60/0x7b [btrfs]
[61666.786253] btrfs_sync_file+0x317/0x42c [btrfs]
[61666.786253] vfs_fsync_range+0x8c/0x9e
[61666.786253] SyS_msync+0x13c/0x1c9
[61666.786253] entry_SYSCALL_64_fastpath+0x18/0xad
A sample of a corrupt log tree leaf with overlapping extents I got from
running btrfs/072:
item 14 key (295 108 200704) itemoff 2599 itemsize 53
extent data disk bytenr 0 nr 0
extent data offset 0 nr 458752 ram 458752
item 15 key (295 108 659456) itemoff 2546 itemsize 53
extent data disk bytenr 4343541760 nr 770048
extent data offset 606208 nr 163840 ram 770048
item 16 key (295 108 663552) itemoff 2493 itemsize 53
extent data disk bytenr 4343541760 nr 770048
extent data offset 610304 nr 155648 ram 770048
item 17 key (295 108 819200) itemoff 2440 itemsize 53
extent data disk bytenr 4334788608 nr 4096
extent data offset 0 nr 4096 ram 4096
The file extent item at offset 659456 (item 15) ends at offset 823296
(659456 + 163840) while the next file extent item (item 16) starts at
offset 663552.
Another different problem that the race can trigger is a failure in the
assertions at tree-log.c:copy_items(), which expect that the first file
extent item key we found before releasing the path exists after we have
released path and that the last key we found before releasing the path
also exists after releasing the path:
$ cat -n fs/btrfs/tree-log.c
4080 if (need_find_last_extent) {
4081 /* btrfs_prev_leaf could return 1 without releasing the path */
4082 btrfs_release_path(src_path);
4083 ret = btrfs_search_slot(NULL, inode->root, &first_key,
4084 src_path, 0, 0);
4085 if (ret < 0)
4086 return ret;
4087 ASSERT(ret == 0);
(...)
4103 if (i >= btrfs_header_nritems(src_path->nodes[0])) {
4104 ret = btrfs_next_leaf(inode->root, src_path);
4105 if (ret < 0)
4106 return ret;
4107 ASSERT(ret == 0);
4108 src = src_path->nodes[0];
4109 i = 0;
4110 need_find_last_extent = true;
4111 }
(...)
The second assertion implicitly expects that the last key before the path
release still exists, because the surrounding while loop only stops after
we have found that key. When this assertion fails it produces a stack like
this:
[139590.037075] assertion failed: ret == 0, file: fs/btrfs/tree-log.c, line: 4107
[139590.037406] ------------[ cut here ]------------
[139590.037707] kernel BUG at fs/btrfs/ctree.h:3546!
[139590.038034] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC PTI
[139590.038340] CPU: 1 PID: 31841 Comm: fsstress Tainted: G W 5.0.0-btrfs-next-46 #1
(...)
[139590.039354] RIP: 0010:assfail.constprop.24+0x18/0x1a [btrfs]
(...)
[139590.040397] RSP: 0018:ffffa27f48f2b9b0 EFLAGS: 00010282
[139590.040730] RAX: 0000000000000041 RBX: ffff897c635d92c8 RCX: 0000000000000000
[139590.041105] RDX: 0000000000000000 RSI: ffff897d36a96868 RDI: ffff897d36a96868
[139590.041470] RBP: ffff897d1b9a0708 R08: 0000000000000000 R09: 0000000000000000
[139590.041815] R10: 0000000000000008 R11: 0000000000000000 R12: 0000000000000013
[139590.042159] R13: 0000000000000227 R14: ffff897cffcbba88 R15: 0000000000000001
[139590.042501] FS: 00007f2efc8dee80(0000) GS:ffff897d36a80000(0000) knlGS:0000000000000000
[139590.042847] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[139590.043199] CR2: 00007f8c064935e0 CR3: 0000000232252002 CR4: 00000000003606e0
[139590.043547] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[139590.043899] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[139590.044250] Call Trace:
[139590.044631] copy_items+0xa3f/0x1000 [btrfs]
[139590.045009] ? generic_bin_search.constprop.32+0x61/0x200 [btrfs]
[139590.045396] btrfs_log_inode+0x7b3/0xd70 [btrfs]
[139590.045773] btrfs_log_inode_parent+0x2b3/0xce0 [btrfs]
[139590.046143] ? do_raw_spin_unlock+0x49/0xc0
[139590.046510] btrfs_log_dentry_safe+0x4a/0x70 [btrfs]
[139590.046872] btrfs_sync_file+0x3b6/0x440 [btrfs]
[139590.047243] btrfs_file_write_iter+0x45b/0x5c0 [btrfs]
[139590.047592] __vfs_write+0x129/0x1c0
[139590.047932] vfs_write+0xc2/0x1b0
[139590.048270] ksys_write+0x55/0xc0
[139590.048608] do_syscall_64+0x60/0x1b0
[139590.048946] entry_SYSCALL_64_after_hwframe+0x49/0xbe
[139590.049287] RIP: 0033:0x7f2efc4be190
(...)
[139590.050342] RSP: 002b:00007ffe743243a8 EFLAGS: 00000246 ORIG_RAX: 0000000000000001
[139590.050701] RAX: ffffffffffffffda RBX: 0000000000008d58 RCX: 00007f2efc4be190
[139590.051067] RDX: 0000000000008d58 RSI: 00005567eca0f370 RDI: 0000000000000003
[139590.051459] RBP: 0000000000000024 R08: 0000000000000003 R09: 0000000000008d60
[139590.051863] R10: 0000000000000078 R11: 0000000000000246 R12: 0000000000000003
[139590.052252] R13: 00000000003d3507 R14: 00005567eca0f370 R15: 0000000000000000
(...)
[139590.055128] ---[ end trace 193f35d0215cdeeb ]---
So fix this race between a full ranged fsync and writeback of adjacent
ranges by flushing all delalloc and waiting for all ordered extents to
complete before logging the inode. This is the simplest way to solve the
problem because currently the full fsync path does not deal with ranges
at all (it assumes a full range from 0 to LLONG_MAX) and it always needs
to look at adjacent ranges for hole detection. For use cases of ranged
fsyncs this can make a few fsyncs slower but on the other hand it can
make some following fsyncs to other ranges do less work or no need to do
anything at all. A full fsync is rare anyway and happens only once after
loading/creating an inode and once after less common operations such as a
shrinking truncate.
This is an issue that exists for a long time, and was often triggered by
generic/127, because it does mmap'ed writes and msync (which triggers a
ranged fsync). Adding support for the tree checker to detect overlapping
extents (next patch in the series) and trigger a WARN() when such cases
are found, and then calling btrfs_check_leaf_full() at the end of
btrfs_insert_file_extent() made the issue much easier to detect. Running
btrfs/072 with that change to the tree checker and making fsstress open
files always with O_SYNC made it much easier to trigger the issue (as
triggering it with generic/127 is very rare).
CC: stable@vger.kernel.org # 3.16+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When we are doing a full fsync (bit BTRFS_INODE_NEEDS_FULL_SYNC set) of a
file that has holes and has file extent items spanning two or more leafs,
we can end up falling to back to a full transaction commit due to a logic
bug that leads to failure to insert a duplicate file extent item that is
meant to represent a hole between the last file extent item of a leaf and
the first file extent item in the next leaf. The failure (EEXIST error)
leads to a transaction commit (as most errors when logging an inode do).
For example, we have the two following leafs:
Leaf N:
-----------------------------------------------
| ..., ..., ..., (257, FILE_EXTENT_ITEM, 64K) |
-----------------------------------------------
The file extent item at the end of leaf N has a length of 4Kb,
representing the file range from 64K to 68K - 1.
Leaf N + 1:
-----------------------------------------------
| (257, FILE_EXTENT_ITEM, 72K), ..., ..., ... |
-----------------------------------------------
The file extent item at the first slot of leaf N + 1 has a length of
4Kb too, representing the file range from 72K to 76K - 1.
During the full fsync path, when we are at tree-log.c:copy_items() with
leaf N as a parameter, after processing the last file extent item, that
represents the extent at offset 64K, we take a look at the first file
extent item at the next leaf (leaf N + 1), and notice there's a 4K hole
between the two extents, and therefore we insert a file extent item
representing that hole, starting at file offset 68K and ending at offset
72K - 1. However we don't update the value of *last_extent, which is used
to represent the end offset (plus 1, non-inclusive end) of the last file
extent item inserted in the log, so it stays with a value of 68K and not
with a value of 72K.
Then, when copy_items() is called for leaf N + 1, because the value of
*last_extent is smaller then the offset of the first extent item in the
leaf (68K < 72K), we look at the last file extent item in the previous
leaf (leaf N) and see it there's a 4K gap between it and our first file
extent item (again, 68K < 72K), so we decide to insert a file extent item
representing the hole, starting at file offset 68K and ending at offset
72K - 1, this insertion will fail with -EEXIST being returned from
btrfs_insert_file_extent() because we already inserted a file extent item
representing a hole for this offset (68K) in the previous call to
copy_items(), when processing leaf N.
The -EEXIST error gets propagated to the fsync callback, btrfs_sync_file(),
which falls back to a full transaction commit.
Fix this by adjusting *last_extent after inserting a hole when we had to
look at the next leaf.
Fixes: 4ee3fad34a ("Btrfs: fix fsync after hole punching when using no-holes feature")
Cc: stable@vger.kernel.org # 4.14+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Commit ddf30cf03f ("btrfs: extent-tree: Use btrfs_ref to refactor
add_pinned_bytes()") refactored add_pinned_bytes(), but during that
refactor, there are two callers which add the pinned bytes instead
of subtracting.
That refactor misses those two caller, causing incorrect pinned bytes
calculation and resulting unexpected ENOSPC error.
Fix it by adding a new parameter @sign to restore the original behavior.
Reported-by: kernel test robot <rong.a.chen@intel.com>
Fixes: ddf30cf03f ("btrfs: extent-tree: Use btrfs_ref to refactor add_pinned_bytes()")
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
A failed call to kobject_init_and_add() must be followed by a call to
kobject_put(). Currently in the error path when adding fs_devices we
are missing this call. This could be fixed by calling
btrfs_sysfs_remove_fsid() if btrfs_sysfs_add_fsid() returns an error or
by adding a call to kobject_put() directly in btrfs_sysfs_add_fsid().
Here we choose the second option because it prevents the slightly
unusual error path handling requirements of kobject from leaking out
into btrfs functions.
Add a call to kobject_put() in the error path of kobject_add_and_init().
This causes the release method to be called if kobject_init_and_add()
fails. open_tree() is the function that calls btrfs_sysfs_add_fsid()
and the error code in this function is already written with the
assumption that the release method is called during the error path of
open_tree() (as seen by the call to btrfs_sysfs_remove_fsid() under the
fail_fsdev_sysfs label).
Cc: stable@vger.kernel.org # v4.4+
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Tobin C. Harding <tobin@kernel.org>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
If a call to kobject_init_and_add() fails we must call kobject_put()
otherwise we leak memory.
Calling kobject_put() when kobject_init_and_add() fails drops the
refcount back to 0 and calls the ktype release method (which in turn
calls the percpu destroy and kfree).
Add call to kobject_put() in the error path of call to
kobject_init_and_add().
Cc: stable@vger.kernel.org # v4.4+
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Tobin C. Harding <tobin@kernel.org>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently when we fail to COW a path at btrfs_update_root() we end up
always aborting the transaction. However all the current callers of
btrfs_update_root() are able to deal with errors returned from it, many do
end up aborting the transaction themselves (directly or not, such as the
transaction commit path), other BUG_ON() or just gracefully cancel whatever
they were doing.
When syncing the fsync log, we call btrfs_update_root() through
tree-log.c:update_log_root(), and if it returns an -ENOSPC error, the log
sync code does not abort the transaction, instead it gracefully handles
the error and returns -EAGAIN to the fsync handler, so that it falls back
to a transaction commit. Any other error different from -ENOSPC, makes the
log sync code abort the transaction.
So remove the transaction abort from btrfs_update_log() when we fail to
COW a path to update the root item, so that if an -ENOSPC failure happens
we avoid aborting the current transaction and have a chance of the fsync
succeeding after falling back to a transaction commit.
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=203413
Fixes: 79787eaab4 ("btrfs: replace many BUG_ONs with proper error handling")
Cc: stable@vger.kernel.org # 4.4+
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We're now reserving an extra items worth of space for property
inheritance. We only have one property at the moment so this covers us,
but if we add more in the future this will allow us to not get bitten by
the extra space reservation. If we do add more properties in the future
we should re-visit how we calculate the space reservation needs by the
callers.
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
[ refreshed on top of prop/xattr cleanups ]
Signed-off-by: David Sterba <dsterba@suse.com>
If we have an error writing out a delalloc range in
btrfs_punch_hole_lock_range we'll unlock the inode and then goto
out_only_mutex, where we will again unlock the inode. This is bad,
don't do this.
Fixes: f27451f229 ("Btrfs: add support for fallocate's zero range operation")
CC: stable@vger.kernel.org # 4.19+
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When a file's compression property is set as zlib or zstd but leave
the compression mount option not be set, that means btrfs will try
to compress the file with default compression level. But in
btrfs_compress_pages(), it calls get_workspace() with level = 0.
This will return a workspace with a wrong compression level.
For zlib, the compression level in the workspace will be 0
(that means "store only"). And for zstd, the compression in the
workspace will be 1, not the default level 3.
How to reproduce:
mkfs -t btrfs /dev/sdb
mount /dev/sdb /mnt/
mkdir /mnt/zlib
btrfs property set /mnt/zlib/ compression zlib
dd if=/dev/zero of=/mnt/zlib/compression-friendly-file-10M bs=1M count=10
sync
btrfs-debugfs -f /mnt/zlib/compression-friendly-file-10M
btrfs-debugfs output:
* before:
...
(258 9961472): ram 524288 disk 1106247680 disk_size 524288
file: ... extents 20 disk size 10485760 logical size 10485760 ratio 1.00
* after:
...
(258 10354688): ram 131072 disk 14217216 disk_size 4096
file: ... extents 80 disk size 327680 logical size 10485760 ratio 32.00
The steps for zstd are similar, but need to put a debugging message to
show the level of the return workspace in zstd_get_workspace().
This commit adds a check of the compression level before getting a
workspace by set_level().
CC: stable@vger.kernel.org # 5.1+
Signed-off-by: Johnny Chang <johnnyc@synology.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Recent refactoring of cow_file_range_async means it's now possible to
request a rather large physically contiguous memory via kmalloc. The
size is dependent on the number of 512k chunks that the compressed range
consists of. David reported multiple OOM messages on such large
allocations. Fix it by switching to using kvmalloc.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Irrespective of whether the compress code fell back to uncompressed or
a compressed extent has to be submitted, the extent range is always
locked. So factor out the common lock_extent call at the beginning of
the loop. No functional changes just removes one duplicate lock_extent
call.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The inode never changes so it's sufficient to dereference it and get
the iotree only once, before the execution of the main loop. No
functional changes, only the size of the function is decreased:
add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-44 (-44)
Function old new delta
submit_compressed_extents 1240 1196 -44
Total: Before=88476, After=88432, chg -0.05%
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
All context this function needs is held within struct async_chunk.
Currently we not only pass the struct but also every individual member.
This is redundant, simplify it by only passing struct async_chunk and
leaving it to compress_file_range to extract the values it requires.
No functional changes.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The associated btrfs_work already contains a reference to the fs_info so
use that instead of passing it via async_chunk. No functional changes.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Now that we have an explicit async_chunk struct rename references to
variables of this type to async_chunk. No functional changes.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
This commit changes the implementation of cow_file_range_async in order
to get rid of the BUG_ON in the middle of the loop. Additionally it
reworks the inner loop in the hopes of making it more understandable.
The idea is to make async_cow be a top-level structured, shared amongst
all chunks being sent for compression. This allows to perform one memory
allocation at the beginning and gracefully fail the IO if there isn't
enough memory. Now, each chunk is going to be described by an
async_chunk struct. It's the responsibility of the final chunk
to actually free the memory.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
With the per-inode block reserves we started refilling the reserve based
on the calculated size of the outstanding csum bytes and extents for the
inode, including the amount we were adding with the new operation.
However, generic/224 exposed a problem with this approach. With 1000
files all writing at the same time we ended up with a bunch of bytes
being reserved but unusable.
When you write to a file we reserve space for the csum leaves for those
bytes, the number of extent items required to cover those bytes, and a
single transaction item for updating the inode at ordered extent finish
for that range of bytes. This is held until the ordered extent finishes
and we release all of the reserved space.
If a second write comes in at this point we would add a single
reservation for the new outstanding extent and however many reservations
for the csum leaves. At this point we find the delta of how much we
have reserved and how much outstanding size this is and attempt to
reserve this delta. If the first write finishes it will not release any
space, because the space it had reserved for the initial write is still
needed for the second write. However some space would have been used,
as we have added csums, extent items, and dirtied the inode. Our
reserved space would be > 0 but less than the total needed reserved
space.
This is just for a single inode, now consider generic/224. This has
1000 inodes writing in parallel to a very small file system, 1GiB. In
my testing this usually means we get about a 120MiB metadata area to
work with, more than enough to allow the writes to continue, but not
enough if all of the inodes are stuck trying to reserve the slack space
while continuing to hold their leftovers from their initial writes.
Fix this by pre-reserved _only_ for the space we are currently trying to
add. Then once that is successful modify our inodes csum count and
outstanding extents, and then add the newly reserved space to the inodes
block_rsv. This allows us to actually pass generic/224 without running
out of metadata space.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When diagnosing a slowdown of generic/224 I noticed we were not doing
anything when calling into shrink_delalloc(). This is because all
writes in 224 are O_DIRECT, not delalloc, and thus our delalloc_bytes
counter is 0, which short circuits most of the work inside of
shrink_delalloc(). However O_DIRECT writes still consume metadata
resources and generate ordered extents, which we can still wait on.
Fix this by tracking outstanding DIO write bytes, and use this as well
as the delalloc bytes counter to decide if we need to lookup and wait on
any ordered extents. If we have more DIO writes than delalloc bytes
we'll go ahead and wait on any ordered extents regardless of our flush
state as flushing delalloc is likely to not gain us anything.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
[ use dio instead of odirect in identifiers ]
Signed-off-by: David Sterba <dsterba@suse.com>
Since now the trans argument is never NULL in btrfs_set_prop we don't
have to check. So delete it and use btrfs_setxattr that makes use of
that.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The last consumer of btrfs_set_prop_trans() was taken away by the patch
("btrfs: start transaction in xattr_handler_set_prop") so now this
function can be deleted.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs specific extended attributes on the inode are set using
btrfs_xattr_handler_set_prop(), and the required transaction for this
update is started by btrfs_setxattr(). For better visibility of the
transaction start and end, do this in btrfs_xattr_handler_set_prop().
For which this patch copied code of btrfs_setxattr() as it is in the
original, which needs proper error handling.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
There isn't real use of making struct inode::i_mode a local copy, it
saves a dereference one time, not much. Just use it directly.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs_inode_flags_to_fsflags() is copied into @old_fsflags and used only
once. Instead used it directly.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Instead of updating the binode::flags directly, update a local copy, and
then at the point of no error, store copy it to the binode::flags.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The patch ("btrfs: start transaction in btrfs_ioctl_setflags()") used
btrfs_set_prop() instead of btrfs_set_prop_trans() by which now the
inode::i_flags update functions such as
btrfs_sync_inode_flags_to_i_flags() and btrfs_update_inode() is called
in btrfs_ioctl_setflags() instead of
btrfs_set_prop_trans()->btrfs_setxattr() as earlier. So the
inode::i_flags remains unmodified until the thread has checked all the
conditions. So drop the saved inode::i_flags in out_i_flags.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Inode attribute can be set through the FS_IOC_SETFLAGS ioctl. This
flags also includes compression attribute for which we would set/reset
the compression extended attribute. While doing this there is a bit of
duplicate code, the following things happens twice:
- start/end_transaction
- inode_inc_iversion()
- current_time update to inode->i_ctime
- and btrfs_update_inode()
These are updated both at btrfs_ioctl_setflags() and btrfs_set_props()
as well. This patch merges these two duplicate codes at
btrfs_ioctl_setflags().
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Make btrfs_set_prop() a non-static function, so that it can be called
from btrfs_ioctl_setflags(). We need btrfs_set_prop() instead of
btrfs_set_prop_trans() so that we can use the transaction which is
already started in the current thread.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
In preparation to merge multiple transactions when setting the
compression flags, split btrfs_set_props() validation part outside of
it.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Allowing error injection for btrfs_check_leaf_full() and
btrfs_check_node() is useful to test the failure path of btrfs write
time tree check.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Commit 41bd606769 ("Btrfs: fix fsync of files with multiple hard links
in new directories") introduced a path that makes fsync fallback to a full
transaction commit in order to avoid losing hard links and new ancestors
of the fsynced inode. That path is triggered only when the inode has more
than one hard link and either has a new hard link created in the current
transaction or the inode was evicted and reloaded in the current
transaction.
That path ends up getting triggered very often (hundreds of times) during
the course of pgbench benchmarks, resulting in performance drops of about
20%.
This change restores the performance by not triggering the full transaction
commit in those cases, and instead iterate the fs/subvolume tree in search
of all possible new ancestors, for all hard links, to log them.
Reported-by: Zhao Yuhu <zyuhu@suse.com>
Tested-by: James Wang <jnwang@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Send operates on read only trees and expects them to never change while it
is using them. This is part of its initial design, and this expection is
due to two different reasons:
1) When it was introduced, no operations were allowed to modifiy read-only
subvolumes/snapshots (including defrag for example).
2) It keeps send from having an impact on other filesystem operations.
Namely send does not need to keep locks on the trees nor needs to hold on
to transaction handles and delay transaction commits. This ends up being
a consequence of the former reason.
However the deduplication feature was introduced later (on September 2013,
while send was introduced in July 2012) and it allowed for deduplication
with destination files that belong to read-only trees (subvolumes and
snapshots).
That means that having a send operation (either full or incremental) running
in parallel with a deduplication that has the destination inode in one of
the trees used by the send operation, can result in tree nodes and leaves
getting freed and reused while send is using them. This problem is similar
to the problem solved for the root nodes getting freed and reused when a
snapshot is made against one tree that is currenly being used by a send
operation, fixed in commits [1] and [2]. These commits explain in detail
how the problem happens and the explanation is valid for any node or leaf
that is not the root of a tree as well. This problem was also discussed
and explained recently in a thread [3].
The problem is very easy to reproduce when using send with large trees
(snapshots) and just a few concurrent deduplication operations that target
files in the trees used by send. A stress test case is being sent for
fstests that triggers the issue easily. The most common error to hit is
the send ioctl return -EIO with the following messages in dmesg/syslog:
[1631617.204075] BTRFS error (device sdc): did not find backref in send_root. inode=63292, offset=0, disk_byte=5228134400 found extent=5228134400
[1631633.251754] BTRFS error (device sdc): parent transid verify failed on 32243712 wanted 24 found 27
The first one is very easy to hit while the second one happens much less
frequently, except for very large trees (in that test case, snapshots
with 100000 files having large xattrs to get deep and wide trees).
Less frequently, at least one BUG_ON can be hit:
[1631742.130080] ------------[ cut here ]------------
[1631742.130625] kernel BUG at fs/btrfs/ctree.c:1806!
[1631742.131188] invalid opcode: 0000 [#6] SMP DEBUG_PAGEALLOC PTI
[1631742.131726] CPU: 1 PID: 13394 Comm: btrfs Tainted: G B D W 5.0.0-rc8-btrfs-next-45 #1
[1631742.132265] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014
[1631742.133399] RIP: 0010:read_node_slot+0x122/0x130 [btrfs]
(...)
[1631742.135061] RSP: 0018:ffffb530021ebaa0 EFLAGS: 00010246
[1631742.135615] RAX: ffff93ac8912e000 RBX: 000000000000009d RCX: 0000000000000002
[1631742.136173] RDX: 000000000000009d RSI: ffff93ac564b0d08 RDI: ffff93ad5b48c000
[1631742.136759] RBP: ffffb530021ebb7d R08: 0000000000000001 R09: ffffb530021ebb7d
[1631742.137324] R10: ffffb530021eba70 R11: 0000000000000000 R12: ffff93ac87d0a708
[1631742.137900] R13: 0000000000000000 R14: 0000000000000000 R15: 0000000000000001
[1631742.138455] FS: 00007f4cdb1528c0(0000) GS:ffff93ad76a80000(0000) knlGS:0000000000000000
[1631742.139010] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[1631742.139568] CR2: 00007f5acb3d0420 CR3: 000000012be3e006 CR4: 00000000003606e0
[1631742.140131] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[1631742.140719] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[1631742.141272] Call Trace:
[1631742.141826] ? do_raw_spin_unlock+0x49/0xc0
[1631742.142390] tree_advance+0x173/0x1d0 [btrfs]
[1631742.142948] btrfs_compare_trees+0x268/0x690 [btrfs]
[1631742.143533] ? process_extent+0x1070/0x1070 [btrfs]
[1631742.144088] btrfs_ioctl_send+0x1037/0x1270 [btrfs]
[1631742.144645] _btrfs_ioctl_send+0x80/0x110 [btrfs]
[1631742.145161] ? trace_sched_stick_numa+0xe0/0xe0
[1631742.145685] btrfs_ioctl+0x13fe/0x3120 [btrfs]
[1631742.146179] ? account_entity_enqueue+0xd3/0x100
[1631742.146662] ? reweight_entity+0x154/0x1a0
[1631742.147135] ? update_curr+0x20/0x2a0
[1631742.147593] ? check_preempt_wakeup+0x103/0x250
[1631742.148053] ? do_vfs_ioctl+0xa2/0x6f0
[1631742.148510] ? btrfs_ioctl_get_supported_features+0x30/0x30 [btrfs]
[1631742.148942] do_vfs_ioctl+0xa2/0x6f0
[1631742.149361] ? __fget+0x113/0x200
[1631742.149767] ksys_ioctl+0x70/0x80
[1631742.150159] __x64_sys_ioctl+0x16/0x20
[1631742.150543] do_syscall_64+0x60/0x1b0
[1631742.150931] entry_SYSCALL_64_after_hwframe+0x49/0xbe
[1631742.151326] RIP: 0033:0x7f4cd9f5add7
(...)
[1631742.152509] RSP: 002b:00007ffe91017708 EFLAGS: 00000202 ORIG_RAX: 0000000000000010
[1631742.152892] RAX: ffffffffffffffda RBX: 0000000000000105 RCX: 00007f4cd9f5add7
[1631742.153268] RDX: 00007ffe91017790 RSI: 0000000040489426 RDI: 0000000000000007
[1631742.153633] RBP: 0000000000000007 R08: 00007f4cd9e79700 R09: 00007f4cd9e79700
[1631742.153999] R10: 00007f4cd9e799d0 R11: 0000000000000202 R12: 0000000000000003
[1631742.154365] R13: 0000555dfae53020 R14: 0000000000000000 R15: 0000000000000001
(...)
[1631742.156696] ---[ end trace 5dac9f96dcc3fd6b ]---
That BUG_ON happens because while send is using a node, that node is COWed
by a concurrent deduplication, gets freed and gets reused as a leaf (because
a transaction commit happened in between), so when it attempts to read a
slot from the extent buffer, at ctree.c:read_node_slot(), the extent buffer
contents were wiped out and it now matches a leaf (which can even belong to
some other tree now), hitting the BUG_ON(level == 0).
Fix this concurrency issue by not allowing send and deduplication to run
in parallel if both operate on the same readonly trees, returning EAGAIN
to user space and logging an exlicit warning in dmesg/syslog.
[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=be6821f82c3cc36e026f5afd10249988852b35ea
[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=6f2f0b394b54e2b159ef969a0b5274e9bbf82ff2
[3] https://lore.kernel.org/linux-btrfs/CAL3q7H7iqSEEyFaEtpRZw3cp613y+4k2Q8b4W7mweR3tZA05bQ@mail.gmail.com/
CC: stable@vger.kernel.org # 4.4+
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When we set a subvolume to read-only mode we do not flush dellaloc for any
of its inodes (except if the filesystem is mounted with -o flushoncommit),
since it does not affect correctness for any subsequent operations - except
for a future send operation. The send operation will not be able to see the
delalloc data since the respective file extent items, inode item updates,
backreferences, etc, have not hit yet the subvolume and extent trees.
Effectively this means data loss, since the send stream will not contain
any data from existing delalloc. Another problem from this is that if the
writeback starts and finishes while the send operation is in progress, we
have the subvolume tree being being modified concurrently which can result
in send failing unexpectedly with EIO or hitting runtime errors, assertion
failures or hitting BUG_ONs, etc.
Simple reproducer:
$ mkfs.btrfs -f /dev/sdb
$ mount /dev/sdb /mnt
$ btrfs subvolume create /mnt/sv
$ xfs_io -f -c "pwrite -S 0xea 0 108K" /mnt/sv/foo
$ btrfs property set /mnt/sv ro true
$ btrfs send -f /tmp/send.stream /mnt/sv
$ od -t x1 -A d /mnt/sv/foo
0000000 ea ea ea ea ea ea ea ea ea ea ea ea ea ea ea ea
*
0110592
$ umount /mnt
$ mkfs.btrfs -f /dev/sdc
$ mount /dev/sdc /mnt
$ btrfs receive -f /tmp/send.stream /mnt
$ echo $?
0
$ od -t x1 -A d /mnt/sv/foo
0000000
# ---> empty file
Since this a problem that affects send only, fix it in send by flushing
dellaloc for all the roots used by the send operation before send starts
to process the commit roots.
This is a problem that affects send since it was introduced (commit
31db9f7c23 ("Btrfs: introduce BTRFS_IOC_SEND for btrfs send/receive"))
but backporting it to older kernels has some dependencies:
- For kernels between 3.19 and 4.20, it depends on commit 3cd24c6980
("btrfs: use tagged writepage to mitigate livelock of snapshot") because
the function btrfs_start_delalloc_snapshot() does not exist before that
commit. So one has to either pick that commit or replace the calls to
btrfs_start_delalloc_snapshot() in this patch with calls to
btrfs_start_delalloc_inodes().
- For kernels older than 3.19 it also requires commit e5fa8f865b
("Btrfs: ensure send always works on roots without orphans") because
it depends on the function ensure_commit_roots_uptodate() which that
commits introduced.
- No dependencies for 5.0+ kernels.
A test case for fstests follows soon.
CC: stable@vger.kernel.org # 3.19+
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
During fiemap, for regular extents (non inline) we need to check if they
are shared and if they are, set the shared bit. Checking if an extent is
shared requires checking the delayed references of the currently running
transaction, since some reference might have not yet hit the extent tree
and be only in the in-memory delayed references.
However we were using a transaction join for this, which creates a new
transaction when there is no transaction currently running. That means
that two more potential failures can happen: creating the transaction and
committing it. Further, if no write activity is currently happening in the
system, and fiemap calls keep being done, we end up creating and
committing transactions that do nothing.
In some extreme cases this can result in the commit of the transaction
created by fiemap to fail with ENOSPC when updating the root item of a
subvolume tree because a join does not reserve any space, leading to a
trace like the following:
heisenberg kernel: ------------[ cut here ]------------
heisenberg kernel: BTRFS: Transaction aborted (error -28)
heisenberg kernel: WARNING: CPU: 0 PID: 7137 at fs/btrfs/root-tree.c:136 btrfs_update_root+0x22b/0x320 [btrfs]
(...)
heisenberg kernel: CPU: 0 PID: 7137 Comm: btrfs-transacti Not tainted 4.19.0-4-amd64 #1 Debian 4.19.28-2
heisenberg kernel: Hardware name: FUJITSU LIFEBOOK U757/FJNB2A5, BIOS Version 1.21 03/19/2018
heisenberg kernel: RIP: 0010:btrfs_update_root+0x22b/0x320 [btrfs]
(...)
heisenberg kernel: RSP: 0018:ffffb5448828bd40 EFLAGS: 00010286
heisenberg kernel: RAX: 0000000000000000 RBX: ffff8ed56bccef50 RCX: 0000000000000006
heisenberg kernel: RDX: 0000000000000007 RSI: 0000000000000092 RDI: ffff8ed6bda166a0
heisenberg kernel: RBP: 00000000ffffffe4 R08: 00000000000003df R09: 0000000000000007
heisenberg kernel: R10: 0000000000000000 R11: 0000000000000001 R12: ffff8ed63396a078
heisenberg kernel: R13: ffff8ed092d7c800 R14: ffff8ed64f5db028 R15: ffff8ed6bd03d068
heisenberg kernel: FS: 0000000000000000(0000) GS:ffff8ed6bda00000(0000) knlGS:0000000000000000
heisenberg kernel: CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
heisenberg kernel: CR2: 00007f46f75f8000 CR3: 0000000310a0a002 CR4: 00000000003606f0
heisenberg kernel: DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
heisenberg kernel: DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
heisenberg kernel: Call Trace:
heisenberg kernel: commit_fs_roots+0x166/0x1d0 [btrfs]
heisenberg kernel: ? _cond_resched+0x15/0x30
heisenberg kernel: ? btrfs_run_delayed_refs+0xac/0x180 [btrfs]
heisenberg kernel: btrfs_commit_transaction+0x2bd/0x870 [btrfs]
heisenberg kernel: ? start_transaction+0x9d/0x3f0 [btrfs]
heisenberg kernel: transaction_kthread+0x147/0x180 [btrfs]
heisenberg kernel: ? btrfs_cleanup_transaction+0x530/0x530 [btrfs]
heisenberg kernel: kthread+0x112/0x130
heisenberg kernel: ? kthread_bind+0x30/0x30
heisenberg kernel: ret_from_fork+0x35/0x40
heisenberg kernel: ---[ end trace 05de912e30e012d9 ]---
Since fiemap (and btrfs_check_shared()) is a read-only operation, do not do
a transaction join to avoid the overhead of creating a new transaction (if
there is currently no running transaction) and introducing a potential
point of failure when the new transaction gets committed, instead use a
transaction attach to grab a handle for the currently running transaction
if any.
Reported-by: Christoph Anton Mitterer <calestyo@scientia.net>
Link: https://lore.kernel.org/linux-btrfs/b2a668d7124f1d3e410367f587926f622b3f03a4.camel@scientia.net/
Fixes: afce772e87 ("btrfs: fix check_shared for fiemap ioctl")
CC: stable@vger.kernel.org # 4.14+
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since reloc tree doesn't contribute to qgroup numbers, just skip them.
This should catch the final cause of unnecessary data ref processing
when running balance of metadata with qgroups on.
The 4G data 16 snapshots test (*) should explain it pretty well:
| delayed subtree | refactor delayed ref | this patch
---------------------------------------------------------------------
relocated | 22653 | 22673 | 22744
qgroup dirty | 122792 | 48360 | 70
time | 24.494 | 11.606 | 3.944
Finally, we're at the stage where qgroup + metadata balance cost no
obvious overhead.
Test environment:
Test VM:
- vRAM 8G
- vCPU 8
- block dev vitrio-blk, 'unsafe' cache mode
- host block 850evo
Test workload:
- Copy 4G data from /usr/ to one subvolume
- Create 16 snapshots of that subvolume, and modify 3 files in each
snapshot
- Enable quota, rescan
- Time "btrfs balance start -m"
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Similar to btrfs_inc_extent_ref(), use btrfs_ref to replace the long
parameter list and the confusing @owner parameter.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Use the new btrfs_ref structure and replace parameter list to clean up
the usage of owner and level to distinguish the extent types.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since add_pinned_bytes() only needs to know if the extent is metadata
and if it's a chunk tree extent, btrfs_ref is a perfect match for it, as
we don't need various owner/level trick to determine extent type.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
