Re: deadlocked xfs

[Mark Goodwin <markgw@sgi.com>]

Thanks for the report Eric. This looks very similar to a
deadlock Lachlan recently hit in the patch for
"Use atomics for iclog reference counting"
http://oss.sgi.com/archives/xfs/2008-02/msg00130.html

It seems this patch can cause deadlocks under heavy log traffic.
I don't think anyone has a fix yet ... Lachlan is out this week,
but Tim can follow-up here ...

Cheers
-- Mark

Eric Sandeen wrote:
> while running fs_mark:
> 
> fs_mark -d /mnt/test -D 256 -n 100000 -t 4 -s 20480 -F -S 0
> 
> which runs 4 threads making 20k files,things got stuffed up in the log,
> see sysrq-w output at:
> 
> http://sandeen.fedorapeople.org/xfs/deadxfs
> 
> 2 threads down this path:
> 
>  [<ffffffffa03c89f5>] :xfs:xlog_state_get_iclog_space+0x83/0x187
>  [<ffffffffa03c8c1a>] :xfs:xlog_write+0x121/0x486
>  [<ffffffffa03c9043>] :xfs:xfs_log_write+0x2a/0x50
> 
> this is on a 4 cpu AMD.
> 
> I made the fs with:
> 
> mkfs.xfs -dsize=1t,agcount=32 -l lazy-count=1 -f /dev/sdb
> 
> and mounted with:
> 
> mount -o nobarrier,logbsize=256k,logbufs=8,inode64 /dev/sdb /mnt/test
> 
> has this been seen before?
> 
> Thanks,
> 
> -Eric
> 
> 

-- 

  Mark Goodwin                                  markgw@sgi.com
  Engineering Manager for XFS and PCP    Phone: +61-3-99631937
  SGI Australian Software Group           Cell: +61-4-18969583
-------------------------------------------------------------