btrfs: Don't call btrfs_start_transaction() on frozen fs to avoid deadlock.
authorQu Wenruo <quwenruo@cn.fujitsu.com>
Mon, 19 Jan 2015 07:42:41 +0000 (15:42 +0800)
committerChris Mason <clm@fb.com>
Wed, 21 Jan 2015 01:20:21 +0000 (17:20 -0800)
commita53f4f8e9c8ebe6c9ee3b34c368913aae9876e22
treedfe7b1c873db8cee0c10d3bbdd339b88f941ce88
parent6c9fe14f9d64cc12401a825a60ec5c5723496ca4
btrfs: Don't call btrfs_start_transaction() on frozen fs to avoid deadlock.

Commit 6b5fe46dfa52 (btrfs: do commit in sync_fs if there are pending
changes) will call btrfs_start_transaction() in sync_fs(), to handle
some operations needed to be done in next transaction.

However this can cause deadlock if the filesystem is frozen, with the
following sys_r+w output:
[  143.255932] Call Trace:
[  143.255936]  [<ffffffff816c0e09>] schedule+0x29/0x70
[  143.255939]  [<ffffffff811cb7f3>] __sb_start_write+0xb3/0x100
[  143.255971]  [<ffffffffa040ec06>] start_transaction+0x2e6/0x5a0
[btrfs]
[  143.255992]  [<ffffffffa040f1eb>] btrfs_start_transaction+0x1b/0x20
[btrfs]
[  143.256003]  [<ffffffffa03dc0ba>] btrfs_sync_fs+0xca/0xd0 [btrfs]
[  143.256007]  [<ffffffff811f7be0>] sync_fs_one_sb+0x20/0x30
[  143.256011]  [<ffffffff811cbd01>] iterate_supers+0xe1/0xf0
[  143.256014]  [<ffffffff811f7d75>] sys_sync+0x55/0x90
[  143.256017]  [<ffffffff816c49d2>] system_call_fastpath+0x12/0x17
[  143.256111] Call Trace:
[  143.256114]  [<ffffffff816c0e09>] schedule+0x29/0x70
[  143.256119]  [<ffffffff816c3405>] rwsem_down_write_failed+0x1c5/0x2d0
[  143.256123]  [<ffffffff8133f013>] call_rwsem_down_write_failed+0x13/0x20
[  143.256131]  [<ffffffff811caae8>] thaw_super+0x28/0xc0
[  143.256135]  [<ffffffff811db3e5>] do_vfs_ioctl+0x3f5/0x540
[  143.256187]  [<ffffffff811db5c1>] SyS_ioctl+0x91/0xb0
[  143.256213]  [<ffffffff816c49d2>] system_call_fastpath+0x12/0x17

The reason is like the following:
(Holding s_umount)
VFS sync_fs staff:
|- btrfs_sync_fs()
   |- btrfs_start_transaction()
      |- sb_start_intwrite()
      (Waiting thaw_fs to unfreeze)
VFS thaw_fs staff:
thaw_fs()
(Waiting sync_fs to release
 s_umount)

So deadlock happens.
This can be easily triggered by fstest/generic/068 with inode_cache
mount option.

The fix is to check if the fs is frozen, if the fs is frozen, just
return and waiting for the next transaction.

Cc: David Sterba <dsterba@suse.cz>
Reported-by: Gui Hecheng <guihc.fnst@cn.fujitsu.com>
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
[enhanced comment, changed to SB_FREEZE_WRITE]
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/super.c