Followup to last test with 2.4.23pre6aa3. same config as before (ext3 
/ql2300 / bonnie++) I forgot to mention that I have quotas enabled on 
the partition under test.

There have been a few 2.4 / quota / ext3 deadlock-type  mails on the 
list recently so I'm sot sure if this is related. I can see an atime 
update in the call trace of one of the bonnies but no dquot_ calls 
anywhere.  I haven't applied Jan's patch as it came out before I could 
reliably reproduce this problem.

In the meantime I'll restart testing with noatime - which is how I'd 
usually mount such a fs anyhow.

The box seems to deadlock after 3-6 days of filesystem exercise. I have 
repeated this 3 times. The last run was 6 days and and then I see all 
disk I/O has ceased. The system remains pingable and sshd connects at 
the TCP layer but nothing happens after that. The root filesystem is 
ext3 but is a completely separate device (motherboard scsi) to the 
filesystems running bonnie (external FC RAID). It seems that all 
filesystem activity ceases.

I have run the alt-sysrq-t ouput through ksymoops and attached.  Hope 
this is useful to someone who knows.

-jamie

Jamie Clark wrote:

> I made the quick fix (disabling rq_mergeable) and started the load test.
> Will let it run for a week or so.
>
> Thanks for the help!
> -Jamie
>
> Andrea Arcangeli wrote:
>
>> On Tue, Nov 04, 2003 at 10:49:29AM +0800, Jamie Clark wrote:
>>
>>> Hi,
>>>
>>> Consistent oops with 2.4.23pre6aa3 after 3-4 hours running bonnie on 
>>> ext3 fs through qla2300 HBA. (w SMP, HIGHMEM) The test machine was 
>>> completely wedged so I ended up transcribing the oops from the vga 
>>> console. No typos I think.
>>>
>> I need to release an update that has a chance to fix it. Jens identified
>> problem in his last_merge scsi patch so I will back it out.
>>
>> you can try to backout it by yourself in the meantime, it's called
>> \*elevator-merge-fast-path\* . or you can disable it with this patch:
>