]> Pileus Git - ~andy/linux/commit
Btrfs: fix tree mod logging
authorFilipe David Borba Manana <fdmanana@gmail.com>
Fri, 20 Dec 2013 15:17:46 +0000 (15:17 +0000)
committerChris Mason <clm@fb.com>
Tue, 28 Jan 2014 21:20:04 +0000 (13:20 -0800)
commit5de865eebb8330eee19c37b31fb6f315a09d4273
treece94fb50edc52a2a185bc73ebbe1a36876953eab
parent66ef7d65c3fc6e5300b9359f1c6537efb23781bb
Btrfs: fix tree mod logging

While running the test btrfs/004 from xfstests in a loop, it failed
about 1 time out of 20 runs in my desktop. The failure happened in
the backref walking part of the test, and the test's error message was
like this:

  btrfs/004 93s ... [failed, exit status 1] - output mismatch (see /home/fdmanana/git/hub/xfstests_2/results//btrfs/004.out.bad)
      --- tests/btrfs/004.out 2013-11-26 18:25:29.263333714 +0000
      +++ /home/fdmanana/git/hub/xfstests_2/results//btrfs/004.out.bad 2013-12-10 15:25:10.327518516 +0000
      @@ -1,3 +1,8 @@
       QA output created by 004
       *** test backref walking
      -*** done
      +unexpected output from
      + /home/fdmanana/git/hub/btrfs-progs/btrfs inspect-internal logical-resolve -P 141512704 /home/fdmanana/btrfs-tests/scratch_1
      +expected inum: 405, expected address: 454656, file: /home/fdmanana/btrfs-tests/scratch_1/snap1/p0/d6/d3d/d156/fce, got:
      +
       ...
       (Run 'diff -u tests/btrfs/004.out /home/fdmanana/git/hub/xfstests_2/results//btrfs/004.out.bad' to see the entire diff)
  Ran: btrfs/004
  Failures: btrfs/004
  Failed 1 of 1 tests

But immediately after the test finished, the btrfs inspect-internal command
returned the expected output:

  $ btrfs inspect-internal logical-resolve -P 141512704 /home/fdmanana/btrfs-tests/scratch_1
  inode 405 offset 454656 root 258
  inode 405 offset 454656 root 5

It turned out this was because the btrfs_search_old_slot() calls performed
during backref walking (backref.c:__resolve_indirect_ref) were not finding
anything. The reason for this turned out to be that the tree mod logging
code was not logging some node multi-step operations atomically, therefore
btrfs_search_old_slot() callers iterated often over an incomplete tree that
wasn't fully consistent with any tree state from the past. Besides missing
items, this often (but not always) resulted in -EIO errors during old slot
searches, reported in dmesg like this:

[ 4299.933936] ------------[ cut here ]------------
[ 4299.933949] WARNING: CPU: 0 PID: 23190 at fs/btrfs/ctree.c:1343 btrfs_search_old_slot+0x57b/0xab0 [btrfs]()
[ 4299.933950] Modules linked in: btrfs raid6_pq xor pci_stub vboxpci(O) vboxnetadp(O) vboxnetflt(O) vboxdrv(O) bnep rfcomm bluetooth parport_pc ppdev binfmt_misc joydev snd_hda_codec_h
[ 4299.933977] CPU: 0 PID: 23190 Comm: btrfs Tainted: G        W  O 3.12.0-fdm-btrfs-next-16+ #70
[ 4299.933978] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./Z77 Pro4, BIOS P1.50 09/04/2012
[ 4299.933979]  000000000000053f ffff8806f3fd98f8 ffffffff8176d284 0000000000000007
[ 4299.933982]  0000000000000000 ffff8806f3fd9938 ffffffff8104a81c ffff880659c64b70
[ 4299.933984]  ffff880659c643d0 ffff8806599233d8 ffff880701e2e938 0000160000000000
[ 4299.933987] Call Trace:
[ 4299.933991]  [<ffffffff8176d284>] dump_stack+0x55/0x76
[ 4299.933994]  [<ffffffff8104a81c>] warn_slowpath_common+0x8c/0xc0
[ 4299.933997]  [<ffffffff8104a86a>] warn_slowpath_null+0x1a/0x20
[ 4299.934003]  [<ffffffffa065d3bb>] btrfs_search_old_slot+0x57b/0xab0 [btrfs]
[ 4299.934005]  [<ffffffff81775f3b>] ? _raw_read_unlock+0x2b/0x50
[ 4299.934010]  [<ffffffffa0655001>] ? __tree_mod_log_search+0x81/0xc0 [btrfs]
[ 4299.934019]  [<ffffffffa06dd9b0>] __resolve_indirect_refs+0x130/0x5f0 [btrfs]
[ 4299.934027]  [<ffffffffa06a21f1>] ? free_extent_buffer+0x61/0xc0 [btrfs]
[ 4299.934034]  [<ffffffffa06de39c>] find_parent_nodes+0x1fc/0xe40 [btrfs]
[ 4299.934042]  [<ffffffffa06b13e0>] ? defrag_lookup_extent+0xe0/0xe0 [btrfs]
[ 4299.934048]  [<ffffffffa06b13e0>] ? defrag_lookup_extent+0xe0/0xe0 [btrfs]
[ 4299.934056]  [<ffffffffa06df980>] iterate_extent_inodes+0xe0/0x250 [btrfs]
[ 4299.934058]  [<ffffffff817762db>] ? _raw_spin_unlock+0x2b/0x50
[ 4299.934065]  [<ffffffffa06dfb82>] iterate_inodes_from_logical+0x92/0xb0 [btrfs]
[ 4299.934071]  [<ffffffffa06b13e0>] ? defrag_lookup_extent+0xe0/0xe0 [btrfs]
[ 4299.934078]  [<ffffffffa06b7015>] btrfs_ioctl+0xf65/0x1f60 [btrfs]
[ 4299.934080]  [<ffffffff811658b8>] ? handle_mm_fault+0x278/0xb00
[ 4299.934083]  [<ffffffff81075563>] ? up_read+0x23/0x40
[ 4299.934085]  [<ffffffff8177a41c>] ? __do_page_fault+0x20c/0x5a0
[ 4299.934088]  [<ffffffff811b2946>] do_vfs_ioctl+0x96/0x570
[ 4299.934090]  [<ffffffff81776e23>] ? error_sti+0x5/0x6
[ 4299.934093]  [<ffffffff810b71e8>] ? trace_hardirqs_off_caller+0x28/0xd0
[ 4299.934096]  [<ffffffff81776a09>] ? retint_swapgs+0xe/0x13
[ 4299.934098]  [<ffffffff811b2eb1>] SyS_ioctl+0x91/0xb0
[ 4299.934100]  [<ffffffff813eecde>] ? trace_hardirqs_on_thunk+0x3a/0x3f
[ 4299.934102]  [<ffffffff8177ef12>] system_call_fastpath+0x16/0x1b
[ 4299.934102]  [<ffffffff8177ef12>] system_call_fastpath+0x16/0x1b
[ 4299.934104] ---[ end trace 48f0cfc902491414 ]---
[ 4299.934378] btrfs bad fsid on block 0

These tree mod log operations that must be performed atomically, tree_mod_log_free_eb,
tree_mod_log_eb_copy, tree_mod_log_insert_root and tree_mod_log_insert_move, used to
be performed atomically before the following commit:

  c8cc6341653721b54760480b0d0d9b5f09b46741
  (Btrfs: stop using GFP_ATOMIC for the tree mod log allocations)

That change removed the atomicity of such operations. This patch restores the
atomicity while still not doing the GFP_ATOMIC allocations of tree_mod_elem
structures, so it has to do the allocations using GFP_NOFS before acquiring
the mod log lock.

This issue has been experienced by several users recently, such as for example:

  http://www.spinics.net/lists/linux-btrfs/msg28574.html

After running the btrfs/004 test for 679 consecutive iterations with this
patch applied, I didn't ran into the issue anymore.

Cc: stable@vger.kernel.org
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/ctree.c