xfs_repair issue with ACLs on v5 XFS when beyond v4 limits

[Michael L. Semon]

On 06/18/2014 11:35 PM, Dave Chinner wrote:
> On Thu, Jun 12, 2014 at 10:28:02PM -0400, Michael L. Semon wrote:
>> On 06/10/2014 01:52 AM, Dave Chinner wrote:
>>> On Mon, Jun 09, 2014 at 11:21:03PM -0400, Michael L. Semon wrote:
>>>> Hi!  I've been running around in circles trying to work with too many 
>>>> ACLs, even losing my ability to count for a while.  Along the way, 
>>>> xfs_repair from git xfsprogs (last commit around May 27) is showing 
>>>> the following symptoms:
>>>>
>>>> On v5-superblock XFS...
>>>>
>>>> 1) When the ACL count is just above the limit from v4-superblock XFS--
>>>> 96 is a good test figure--`xfs_repair -n` and `xfs_repair` will both 
>>>> end in a segmentation fault.
>>>
>>> I couldn't reproduce this - I suspect that this is a problem with
>>> the ACL struct having a hardcoded array size or userspace not
>>> having the correct padding in the on-disk structure definition and
>>> you are on a 32bit system. I think I've fixed that in the patch
>>> below.
>>
>> Maybe.  Pentium III has a narrower cacheline than the Pentium 4, so 
>> I was not surprised to see holes in the XFS kernel code, even in the 
>> non-XFS kernel structs.  Do I need to upgrade something (ACL, system 
>> kernel headers, etc.) or would a pahole trip through libxfs be more 
>> revealing?
>>
>> What I'm getting is that if xfs_repair is counting between 200 and 
>> 256 ACLs, it will mention that there are too many ACLs, and it will 
>> segfault.  With your patch, the areas below and above this range are 
>> OK.
>>
>> A sample session like the one I overwrote last time looks like this:
>>
>> Phase 1 - find and verify superblock...
>> Phase 2 - using internal log
>>         - zero log...
>>         - 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
>> Too many ACL entries, count 250
>> entry contains illegal value in attribute named SGI_ACL_FILE or SGI_ACL_DEFAULT
>> (segfault, either Error 4 or Error 5, forgot to bring dmesg)
> 
> Ok, your test found a bug in the patch that was causing segv's - at
> about 20 ACLs, not 250. It's not the same as what you have reported,
> but it was a stack corruption bug and so may just be triggering
> differently on your machines.
> 
> Can you try the patch below?

This patch works!  The range from 4 ACL entries to the ACL limit 
seems to be fine to xfs_repair.  No segfaults, and the ACL limit is 
OK for this case.
 
>> Maybe not...your E-mail patch doesn't have the git version at the 
>> bottom, so I wondered whether I installed the entire patch.  What 
>> I did get went through `git am` just fine, with one whitespace error.
> 
> That's because I didn't use git directly to generate it. As you
> found out, it's still a valid patch...

Indeed.  All is well here, and hopefully, xfs_repair is a patch or two 
more ready for the masses.

Good work!

Michael