xfs errors while unlinking filenames with hash collisions

[Mark Tinguely]

On 03/27/14 16:15, Hannes Frederic Sowa wrote:
> On Thu, Mar 27, 2014 at 03:57:45PM -0500, Mark Tinguely wrote:
>> On 03/27/14 10:24, Hannes Frederic Sowa wrote:
>>> On Thu, Mar 27, 2014 at 10:15:01AM -0500, Mark Tinguely wrote:
>>>> On 03/27/14 09:05, Hannes Frederic Sowa wrote:
>>>>> On Thu, Mar 27, 2014 at 08:34:14AM -0500, Mark Tinguely wrote:
>>>>>> On 03/27/14 08:23, Hannes Frederic Sowa wrote:
>>>>>>> On Thu, Mar 27, 2014 at 08:14:06AM -0500, Mark Tinguely wrote:
>>>>>>>> Have you tried to run a xfs_repair on the filesystem after the reboot?
>>>>>>>
>>>>>>> Yes, I did. I still use the filesystem and it works. As soon as I try
>>>>>>> to
>>>>>>> remove the directory again the same splash from above happens again.
>>>>>>
>>>>>> Is it the latest xfsprogs' repair?
>>>>>>
>>>>>> Do you have the output from the repair still?
>>>>>
>>>>> I can easily test this here, so you can throw any commands and tests at
>>>>> me. ;)
>>>>>
>>>>> This is the output:
>>>>>
>>>>> (I replayed the journal before that)
>>>>>
>>>>> pre-mount:/# xfs_repair -V
>>>>> xfs_repair version 3.1.11
>>>>> pre-mount:/# xfs_repair -v /dev/vda1
>>>>> Phase 1 - find and verify superblock...
>>>>>          - block cache size set to 372848 entries
>>>>> Phase 2 - using internal log
>>>>>          - zero log...
>>>>> zero_log: head block 5071 tail block 5071
>>>>>          - scan filesystem freespace and inode maps...
>>>>>          - found root inode chunk
>>>>> Phase 3 - for each AG...
>>>>>          - scan and clear agi unlinked lists...
>>>>>          - process known inodes and perform inode discovery...
>>>>>          - agno = 0
>>>>> bad hash ordering in block 8388739 of directory inode 3543184
>>>>
>>>> interesting. I will see if I can recreate it.
>>>>
>>>> Are you open to making it an xfstest?
>>>
>>> Sure, I'll put it on my todo list for the weekend.
>>>
>> I will bisect to find the introduction. It appears be somewhere between
>> Linux 3.9 and 3.10.
>
> Thanks!
>
> Maybe it would be best to add a seed to the hashing function (and the super
> block)?
>

Good idea - replicate a test like fsstress.

I bet you could narrow down the iterations required to cause the hang.
I have been using wall clock and disk blocks used as a guide so far.

--Mark.