Re: [deadlock] AGI vs AGF ordering deadlocks

[Mark Tinguely <tinguely@sgi.com>]

On 09/12/13 01:28, Jeff Liu wrote:
> On 09/10/2013 03:58 PM, Jeff Liu wrote:
>
>> On 09/10/2013 03:36 PM, Dave Chinner wrote:
>>
>>> FOlks,
>>>
>>> I just got confirmation of a deadlock I suspected has existed for
>>> some time. A concurrent 16-way create and 16-way unlink just locked
>>> up with two threads looking like this:
>>>
>>> fs_mark         D ffff88021bd931c0  3656  7204   7117 0x00000000
>>>   ffff8801e75293a8 0000000000000086 ffff88012c6d0000 ffff8801e7529fd8
>>>   ffff8801e7529fd8 ffff8801e7529fd8 ffff8802d32aae40 ffff88012c6d0000
>>>   ffff8801a2f79d40 7fffffffffffffff ffff8801ee733bb0 0000000000000002
>>> Call Trace:
>>>   [<ffffffff819b0d19>] schedule+0x29/0x70
>>>   [<ffffffff819acd09>] schedule_timeout+0x149/0x1f0
>>>   [<ffffffff819af6bc>] __down_common+0x91/0xe8
>>>   [<ffffffff819af786>] __down+0x1d/0x1f
>>>   [<ffffffff810b5211>] down+0x41/0x50
>>>   [<ffffffff81423dd0>] xfs_buf_lock+0x40/0xf0
>>>   [<ffffffff81424051>] _xfs_buf_find+0x1d1/0x4d0
>>>   [<ffffffff814244f5>] xfs_buf_get_map+0x35/0x180
>>>   [<ffffffff81425517>] xfs_buf_read_map+0x37/0x110
>>>   [<ffffffff8149e299>] xfs_trans_read_buf_map+0x379/0x600
>>>   [<ffffffff81444178>] xfs_read_agf+0xa8/0x100
>>>   [<ffffffff8144423a>] xfs_alloc_read_agf+0x6a/0x250
>>>   [<ffffffff81444950>] xfs_alloc_fix_freelist+0x4f0/0x5a0
>>>   [<ffffffff81444e40>] xfs_alloc_vextent+0x440/0x840
>>>   [<ffffffff8147d0cf>] xfs_ialloc_ag_alloc+0x13f/0x520
>>>   [<ffffffff8147e871>] xfs_dialloc+0x121/0x2d0
>>>   [<ffffffff814803db>] xfs_ialloc+0x5b/0x7c0
>>>   [<ffffffff81480bda>] xfs_dir_ialloc+0x9a/0x2f0
>>>   [<ffffffff8148134d>] xfs_create+0x47d/0x6a0
>>>   [<ffffffff814343ea>] xfs_vn_mknod+0xba/0x1c0
>>>   [<ffffffff81434523>] xfs_vn_create+0x13/0x20
>>>   [<ffffffff811a62a5>] vfs_create+0xb5/0xf0
>>>   [<ffffffff811a6a40>] do_last.isra.56+0x760/0xd10
>>>   [<ffffffff811a70ae>] path_openat+0xbe/0x620
>>>   [<ffffffff811a7bc3>] do_filp_open+0x43/0xa0
>>>   [<ffffffff811969cc>] do_sys_open+0x13c/0x230
>>>   [<ffffffff81196ae2>] SyS_open+0x22/0x30
>>>   [<ffffffff819bae19>] system_call_fastpath+0x16/0x1b
>>>
>>> That a thread holding an AGI and blocking trying to get the AGF to
>>> do an inode chunk allocation.
>>>
>>> rm              D ffff88021bd931c0  3048  7073   7063 0x00000000
>>>   ffff8802bc66d998 0000000000000086 ffff8802d32aae40 ffff8802bc66dfd8
>>>   ffff8802bc66dfd8 ffff8802bc66dfd8 ffff88012c6d5c80 ffff8802d32aae40
>>>   ffff8804091b2b00 7fffffffffffffff ffff8801b943c570 0000000000000002
>>> Call Trace:
>>>   [<ffffffff819b0d19>] schedule+0x29/0x70
>>>   [<ffffffff819acd09>] schedule_timeout+0x149/0x1f0
>>>   [<ffffffff819af6bc>] __down_common+0x91/0xe8
>>>   [<ffffffff819af786>] __down+0x1d/0x1f
>>>   [<ffffffff810b5211>] down+0x41/0x50
>>>   [<ffffffff81423dd0>] xfs_buf_lock+0x40/0xf0
>>>   [<ffffffff81424051>] _xfs_buf_find+0x1d1/0x4d0
>>>   [<ffffffff814244f5>] xfs_buf_get_map+0x35/0x180
>>>   [<ffffffff81425517>] xfs_buf_read_map+0x37/0x110
>>>   [<ffffffff8149e299>] xfs_trans_read_buf_map+0x379/0x600
>>>   [<ffffffff8147d8ca>] xfs_read_agi+0xaa/0x100
>>>   [<ffffffff81481f4e>] xfs_iunlink+0x8e/0x260
>>>   [<ffffffff81482198>] xfs_droplink+0x78/0x80
>>>   [<ffffffff81483671>] xfs_remove+0x331/0x420
>>>   [<ffffffff814340f2>] xfs_vn_unlink+0x52/0xa0
>>>   [<ffffffff811a4f9e>] vfs_unlink+0x9e/0x110
>>>   [<ffffffff811a51b1>] do_unlinkat+0x1a1/0x230
>>>   [<ffffffff811a805b>] SyS_unlinkat+0x1b/0x40
>>>
>>> And that's a thread that has just freed a directory block and so
>>> holds an AGF lock, and is trying to take the AGI lock to add the
>>> inode to the unlinked list.  Everything else is now stuck waiting
>>> for log space because one of the two buffers we've deadlocked on
>>> here pins the tail of the log.
>>>
>>> The solution is to place the inode on the unlinked list before we
>>> remove the directory entry so that we keep the same locking order as
>>> inode allocation.
>>>
>>> I don't have time to look at this for at least a week, so if someone
>>> could work up solution that'd be wonderful...
>>
>> Although I can reproduce it for now, but it looks interesting to me.
>
> Sorry, s/can/can not/.
>
>> I'll take care of this problem.
>
> Still no luck to reproduce it on my poor laptop, so I have to release
> this for someone who can reproduce it and be interesting enough in fix
> it. :)
>
> Thanks,
> -Jeff

Internal testing hit something similar using tar/rm on Linux 3.0-stable.
There are several threads going after each buffer, but if memory is
correct, it was 2 removes that deadlocked.

I set it aside to work on the Linux 3.12 series, I will take a look at
this some more.

--Mark.

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs