Btrfs: fix race between block group creation and their cache writeout
authorFilipe Manana <fdmanana@suse.com>
Wed, 6 May 2015 15:15:09 +0000 (16:15 +0100)
committerChris Mason <clm@fb.com>
Mon, 11 May 2015 14:59:10 +0000 (07:59 -0700)
commitff1f8250a9e47c1032eb5c86ababff461a11f3a0
tree6181e90101a94706cf1b2496c9a8cff615e711ed
parent28aeeac1dd3080db5108b7b446be69f05c470a90
Btrfs: fix race between block group creation and their cache writeout

So creating a block group has 2 distinct phases:

Phase 1 - creates the btrfs_block_group_cache item and adds it to the
rbtree fs_info->block_group_cache_tree and to the corresponding list
space_info->block_groups[];

Phase 2 - adds the block group item to the extent tree and corresponding
items to the chunk tree.

The first phase adds the block_group_cache_item to a list of pending block
groups in the transaction handle, and phase 2 happens when
btrfs_end_transaction() is called against the transaction handle.

It happens that once phase 1 completes, other concurrent tasks that use
their own transaction handle, but points to the same running transaction
(struct btrfs_trans_handle->transaction), can use this block group for
space allocations and therefore mark it dirty. Dirty block groups are
tracked in a list belonging to the currently running transaction (struct
btrfs_transaction) and not in the transaction handle (btrfs_trans_handle).

This is a problem because once a task calls btrfs_commit_transaction(),
it calls btrfs_start_dirty_block_groups() which will see all dirty block
groups and attempt to start their writeout, including those that are
still attached to the transaction handle of some concurrent task that
hasn't called btrfs_end_transaction() yet - which means those block
groups haven't gone through phase 2 yet and therefore when
write_one_cache_group() is called, it won't find the block group items
in the extent tree and abort the current transaction with -ENOENT,
turning the fs into readonly mode and require a remount.

Fix this by ignoring -ENOENT when looking for block group items in the
extent tree when we attempt to start the writeout of the block group
caches outside the critical section of the transaction commit. We will
try again later during the critical section and if there we still don't
find the block group item in the extent tree, we then abort the current
transaction.

This issue happened twice, once while running fstests btrfs/067 and once
for btrfs/078, which produced the following trace:

[ 3278.703014] WARNING: CPU: 7 PID: 18499 at fs/btrfs/super.c:260 __btrfs_abort_transaction+0x52/0x114 [btrfs]()
[ 3278.707329] BTRFS: Transaction aborted (error -2)
(...)
[ 3278.731555] Call Trace:
[ 3278.732396]  [<ffffffff8142fa46>] dump_stack+0x4f/0x7b
[ 3278.733860]  [<ffffffff8108b6a2>] ? console_unlock+0x361/0x3ad
[ 3278.735312]  [<ffffffff81045ea5>] warn_slowpath_common+0xa1/0xbb
[ 3278.736874]  [<ffffffffa03ada6d>] ? __btrfs_abort_transaction+0x52/0x114 [btrfs]
[ 3278.738302]  [<ffffffff81045f05>] warn_slowpath_fmt+0x46/0x48
[ 3278.739520]  [<ffffffffa03ada6d>] __btrfs_abort_transaction+0x52/0x114 [btrfs]
[ 3278.741222]  [<ffffffffa03b9e56>] write_one_cache_group+0xae/0xbf [btrfs]
[ 3278.742797]  [<ffffffffa03c487b>] btrfs_start_dirty_block_groups+0x170/0x2b2 [btrfs]
[ 3278.744492]  [<ffffffffa03d309c>] btrfs_commit_transaction+0x130/0x9c9 [btrfs]
[ 3278.746084]  [<ffffffff8107d33d>] ? trace_hardirqs_on+0xd/0xf
[ 3278.747249]  [<ffffffffa03e5660>] btrfs_sync_file+0x313/0x387 [btrfs]
[ 3278.748744]  [<ffffffff8117acad>] vfs_fsync_range+0x95/0xa4
[ 3278.749958]  [<ffffffff81435b54>] ? ret_from_sys_call+0x1d/0x58
[ 3278.751218]  [<ffffffff8117acd8>] vfs_fsync+0x1c/0x1e
[ 3278.754197]  [<ffffffff8117ae54>] do_fsync+0x34/0x4e
[ 3278.755192]  [<ffffffff8117b07c>] SyS_fsync+0x10/0x14
[ 3278.756236]  [<ffffffff81435b32>] system_call_fastpath+0x12/0x17
[ 3278.757366] ---[ end trace 9a4d4df4969709aa ]---

Fixes: 1bbc621ef284 ("Btrfs: allow block group cache writeout
                      outside critical section in commit")

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/extent-tree.c