XFS stat/statvfs problems with 2.6.38-rc2-00175-g6fb1b30

Since updating past 2.6.37, I'm seeing du(1) sometimes report files taking up 8GB even though ls(1) reports the correct size (hundreds of bytes). At the same time, df reports that the filesystem is 95% or 100% full (although I can still write to the filesystem). Repeated du calls show persistent results (once st_blocks is wrong, it stays wrong). Rebooting clears the problem for a few hours, then it seems to come back, although the exact files affected seems to change. xfs_check doesn't find any errors, and rebooting into 2.6.37 clears up the problem entirely.

There's nothing in dmesg indicating a problem:

Jan 27 08:04:37 pyron kernel: [    0.000000] Linux version 2.6.38-rc2-00175-g6fb1b30 (adi@pyron) (gcc version 4.4.5 (Debian 4.4.5-10) ) #75 SMP Thu Jan 27 00:39:11 PST 2011
Jan 27 08:04:37 pyron kernel: [    8.420170] SGI XFS with ACLs, security attributes, large block/inode numbers, no debug enabled
Jan 27 08:04:37 pyron kernel: [    8.434527] XFS mounting filesystem dm-8
Jan 27 08:04:37 pyron kernel: [    8.676644] Starting XFS recovery on filesystem: dm-8 (logdev: internal)
Jan 27 08:04:37 pyron kernel: [    8.734279] Ending XFS recovery on filesystem: dm-8 (logdev: internal)
Jan 27 08:04:37 pyron kernel: [    8.762724] XFS mounting filesystem dm-9
Jan 27 08:04:37 pyron kernel: [    8.957986] Ending clean XFS mount for filesystem: dm-9
Jan 27 08:04:37 pyron kernel: [    8.985909] XFS mounting filesystem dm-10
Jan 27 08:04:37 pyron kernel: [    9.098404] Starting XFS recovery on filesystem: dm-10 (logdev: internal)
Jan 27 08:04:37 pyron kernel: [    9.205018] Ending XFS recovery on filesystem: dm-10 (logdev: internal)

/d1 is a 200GB XFS on DM on AHCI SATA on a Core i7 desktop board. The erroneous statvfs reports are visible in this munin graph:
http://web.hexapodia.org/~adi/tmp/20110201-df-pyron.png
(the filesystem is actually 72% full currently, and should be linearly filling as is visible around Jan 26.)

% sudo xfs_info /d1

meta-data=/dev/mapper/vg0-d1     isize=256    agcount=4, agsize=13107200 blks
         =                       sectsz=512   attr=2
data     =                       bsize=4096   blocks=52428800, imaxpct=25
         =                       sunit=0      swidth=0 blks
naming   =version 2              bsize=4096   ascii-ci=0
log      =internal               bsize=4096   blocks=25600, version=2
         =                       sectsz=512   sunit=0 blks, lazy-count=1
realtime =none                   extsz=4096   blocks=0, rtextents=0

I'm back on 2.6.37 right now, I'll try a new git tip tonight.

-andy

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

It'll be the excessive specualtive preallocation bug introduced in .38-rc1 which is fixed in .38-rc3.

Cheers,

Dave.

Dave Chinner
[email protected]

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

Indeed, I've been running 2.6.38-rc3-00019-gafe8a88 for 18 hours now with the same load and not seen any issues.

Thanks,
-andy

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs