Re: [PATCH v3 xfstests] add tests for unlinking directories with hash collisions

From: Mark Tinguely <tinguely@sgi.com>
To: Hannes Frederic Sowa <hannes@stressinduktion.org>
Cc: xfs@oss.sgi.com
Subject: Re: [PATCH v3 xfstests] add tests for unlinking directories with hash collisions
Date: Wed, 02 Apr 2014 08:07:41 -0500	[thread overview]
Message-ID: <533C0B9D.4070307@sgi.com> (raw)
In-Reply-To: <20140402123442.GF16157@order.stressinduktion.org>

On 04/02/14 07:34, Hannes Frederic Sowa wrote:
> This tests creates several directories that have the same small (8)
> group of hashes to ensure the hash ordering of file and directories
> are preserved.
>
> Sample backtrace this test tries to prevent in future:
>
> [ 3856.245843] XFS (vda1): Internal error xfs_trans_cancel at line 966 of file fs/xfs/xfs_trans.c.  Caller 0xffffffffa01186bc
> [ 3856.249049] CPU: 1 PID: 866 Comm: rm Not tainted 3.13.6-200.fc20.x86_64 #1
> [ 3856.250966] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
> [ 3856.252615]  000000000000000c ffff8800d23a7d68 ffffffff8168730c ffff8800cf5462b8
> [ 3856.254823]  ffff8800d23a7d80 ffffffffa00d00cb ffffffffa01186bc ffff8800d23a7da8
> [ 3856.257241]  ffffffffa00e5459 ffff8800d9ac3400 ffff8800d23a7e30 ffff8800371b6800
> [ 3856.259420] Call Trace:
> [ 3856.260172]  [<ffffffff8168730c>] dump_stack+0x45/0x56
> [ 3856.261717]  [<ffffffffa00d00cb>] xfs_error_report+0x3b/0x40 [xfs]
> [ 3856.263472]  [<ffffffffa01186bc>] ? xfs_remove+0x1ac/0x370 [xfs]
> [ 3856.270838]  [<ffffffffa00e5459>] xfs_trans_cancel+0xd9/0x100 [xfs]
> [ 3856.272783]  [<ffffffffa01186bc>] xfs_remove+0x1ac/0x370 [xfs]
> [ 3856.274531]  [<ffffffffa00db40b>] xfs_vn_unlink+0x4b/0x90 [xfs]
> [ 3856.276286]  [<ffffffff811c61b8>] vfs_rmdir+0xa8/0x100
> [ 3856.277821]  [<ffffffff811c638d>] do_rmdir+0x17d/0x1d0
> [ 3856.281021]  [<ffffffff811ba7fe>] ? ____fput+0xe/0x10
> [ 3856.285261]  [<ffffffff8108c11c>] ? task_work_run+0xac/0xe0
> [ 3856.286952]  [<ffffffff81013a31>] ? do_notify_resume+0x61/0xa0
> [ 3856.288693]  [<ffffffff811c9a65>] SyS_unlinkat+0x25/0x40
> [ 3856.290407]  [<ffffffff816962e9>] system_call_fastpath+0x16/0x1b
> [ 3856.292685] XFS (vda1): xfs_do_force_shutdown(0x8) called from line 967 of file fs/xfs/xfs_trans.c.  Return address = 0xffffffffa00e5472
> [ 3856.627330] XFS (vda1): Corruption of in-memory data detected.  Shutting down filesystem
> [ 3856.627332] XFS (vda1): Please umount the filesystem and rectify the problem(s)
>
> With help from Mark Tinguely, thanks very much!
>
> Cc: Dave Chinner<david@fromorbit.com>
> Cc: Mark Tinguely<tinguely@sgi.com>
> Signed-off-by: Hannes Frederic Sowa<hannes@stressinduktion.org>
> ---
> Changelog:
>
> v2)
> * first serious proposal
>
> v3)
> * reduced number of possible generated hashes to 8 and thus lowered the number of
>    generated files to 10_000 which still generate the corruption in all of my
>    10 tests. This speeds up the test considerable. Maybe we can add
>    quick to the group description now?
> * updated changelog
>
> Also:
> When testing this program with reduced number of generated hashes and
> huge amount of test files, xfs_repair needs a considerable amount of
> time to check the directory (I aborted it). I guess this is because the
> hash tables get flattened to linked lists.
>
> I don't know if there are other runtime explosions in other parts of
> the code (maybe in the kernel).  I suggest to add a random perturbation
> to the hash function, which unluckily seems to be included into the
> superblock then, too.
>
> Please have a look!
>
> Thanks,
>
>    Hannes

Tested-by: Mark Tinguely <tinguely@sgi.com>

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