Re: XFS corruption on 2.4.28

[Eric Sandeen <sandeen@xxxxxxx>]

Renaat Dumon wrote:

A couple questions early on - is this stock 2.4.28 from kernel.org?

Are you using extended attributes?  Have you run xfs_fsr on this filesystem?

I doubt that fsr is the culprit here, because your files are only 28 bytes 
long, so fsr would not touch them.

> When I then cd into 0/0/0 and I do a 'du -sk *' :

> 2147483532      000fe1c2b17a7b4b4d2c4eea341cfb08.65536.db

0x7FFFFF8C - hm, a lot of binary 1's in there...


> bacardi 0 # ls -al 000fe1c2b17a7b4b4d2c4eea341cfb08.65536.db 
>
> -rw-------    1 root     root           28 Oct 30 18:53
> 000fe1c2b17a7b4b4d2c4eea341cfb08.65536.db

0x1C

can you try an xfs_bmap -v, and xfs_bmap -a -v of this file?  Just out of 
curiosity.

>  
>
> The correct filesize is indeed 28 bytes! The file mentioned here is just an
> example, but there are quite some files like that actually :(

It might be interesting to gather the reported/correct values for several 
files, so we can possibly identify a pattern.

> Unmounting/remounting the filesystem makes the issue go away temporarily, it
> is back after a couple of hours of operation.

so a file which shows up as bad is ok after a remount?  So at least this 
problem is not on-disk, but...

> I did a xfs_check / xfs_repair before ; but that just dumped (ALMOST)
> EVERYTHING in lost+found , so I'm losing data :(

... something -is- bad on disk.  What is the output of xfs_repair?

>  
>
> The fact that I'm having this on multiple systems is what worries me, the
> filesystems are created with default options, but are mounted with
> su=256,sw=128

what volume manager are you using?

Can you verify whether using the stripe geometry contributes to the problem? 
Without su,sw does the problem go away?


> Does this sound familiar to any of you ? Thanks a bunch!

not quite, the last du reporting problem was only on files with extended 
attributes, and only after an xfs_fsr run - but in your case the files are 
small enough that fsr probably ignores them.

Thanks,

-Eric