btrfs: fix fsfreeze hang caused by delayed iputs deal
authorWang Xiaoguang <wangxg.fnst@cn.fujitsu.com>
Mon, 1 Aug 2016 05:28:08 +0000 (13:28 +0800)
committerChris Mason <clm@fb.com>
Thu, 25 Aug 2016 10:58:26 +0000 (03:58 -0700)
commit9e7cc91a6d18a4973c6d2cc104871439c9e94f3d
tree7ad5c06fa1c4439040665948da4fbf7f7c94d49e
parent18513091af9483ba84328d42092bd4d42a3c958f
btrfs: fix fsfreeze hang caused by delayed iputs deal

When running fstests generic/068, sometimes we got below deadlock:
  xfs_io          D ffff8800331dbb20     0  6697   6693 0x00000080
  ffff8800331dbb20 ffff88007acfc140 ffff880034d895c0 ffff8800331dc000
  ffff880032d243e8 fffffffeffffffff ffff880032d24400 0000000000000001
  ffff8800331dbb38 ffffffff816a9045 ffff880034d895c0 ffff8800331dbba8
  Call Trace:
  [<ffffffff816a9045>] schedule+0x35/0x80
  [<ffffffff816abab2>] rwsem_down_read_failed+0xf2/0x140
  [<ffffffff8118f5e1>] ? __filemap_fdatawrite_range+0xd1/0x100
  [<ffffffff8134f978>] call_rwsem_down_read_failed+0x18/0x30
  [<ffffffffa06631fc>] ? btrfs_alloc_block_rsv+0x2c/0xb0 [btrfs]
  [<ffffffff810d32b5>] percpu_down_read+0x35/0x50
  [<ffffffff81217dfc>] __sb_start_write+0x2c/0x40
  [<ffffffffa067f5d5>] start_transaction+0x2a5/0x4d0 [btrfs]
  [<ffffffffa067f857>] btrfs_join_transaction+0x17/0x20 [btrfs]
  [<ffffffffa068ba34>] btrfs_evict_inode+0x3c4/0x5d0 [btrfs]
  [<ffffffff81230a1a>] evict+0xba/0x1a0
  [<ffffffff812316b6>] iput+0x196/0x200
  [<ffffffffa06851d0>] btrfs_run_delayed_iputs+0x70/0xc0 [btrfs]
  [<ffffffffa067f1d8>] btrfs_commit_transaction+0x928/0xa80 [btrfs]
  [<ffffffffa0646df0>] btrfs_freeze+0x30/0x40 [btrfs]
  [<ffffffff81218040>] freeze_super+0xf0/0x190
  [<ffffffff81229275>] do_vfs_ioctl+0x4a5/0x5c0
  [<ffffffff81003176>] ? do_audit_syscall_entry+0x66/0x70
  [<ffffffff810038cf>] ? syscall_trace_enter_phase1+0x11f/0x140
  [<ffffffff81229409>] SyS_ioctl+0x79/0x90
  [<ffffffff81003c12>] do_syscall_64+0x62/0x110
  [<ffffffff816acbe1>] entry_SYSCALL64_slow_path+0x25/0x25

>From this warning, freeze_super() already holds SB_FREEZE_FS, but
btrfs_freeze() will call btrfs_commit_transaction() again, if
btrfs_commit_transaction() finds that it has delayed iputs to handle,
it'll start_transaction(), which will try to get SB_FREEZE_FS lock
again, then deadlock occurs.

The root cause is that in btrfs, sync_filesystem(sb) does not make
sure all metadata is updated. There still maybe some codes adding
delayed iputs, see below sample race window:

         CPU1                                  |         CPU2
|-> freeze_super()                             |
    |-> sync_filesystem(sb);                   |
    |                                          |-> cleaner_kthread()
    |                                          |   |-> btrfs_delete_unused_bgs()
    |                                          |       |-> btrfs_remove_chunk()
    |                                          |           |-> btrfs_remove_block_group()
    |                                          |               |-> btrfs_add_delayed_iput()
    |                                          |
    |-> sb->s_writers.frozen = SB_FREEZE_FS;   |
    |-> sb_wait_write(sb, SB_FREEZE_FS);       |
    |   acquire SB_FREEZE_FS lock.             |
    |                                          |
    |-> btrfs_freeze()                         |
        |-> btrfs_commit_transaction()         |
            |-> btrfs_run_delayed_iputs()      |
            |   will handle delayed iputs,     |
            |   that means start_transaction() |
            |   will be called, which will try |
            |   to get SB_FREEZE_FS lock.      |

To fix this issue, introduce a "int fs_frozen" to record internally whether
fs has been frozen. If fs has been frozen, we can not handle delayed iputs.

Signed-off-by: Wang Xiaoguang <wangxg.fnst@cn.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ add comment to btrfs_freeze ]
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/ctree.h
fs/btrfs/disk-io.c
fs/btrfs/super.c
fs/btrfs/transaction.c