2.6.31+2.6.31.4: XFS - All I/O locks up to D-state after 24-48 hours (sysrq-t+w available)

[Justin Piszcz]

On Tue, 20 Oct 2009, Justin Piszcz wrote:


>
>
> On Tue, 20 Oct 2009, Dave Chinner wrote:
>
>> On Mon, Oct 19, 2009 at 06:18:58AM -0400, Justin Piszcz wrote:
>>> On Mon, 19 Oct 2009, Dave Chinner wrote:
>>>> On Sun, Oct 18, 2009 at 04:17:42PM -0400, Justin Piszcz wrote:
>>>>> It has happened again, all sysrq-X output was saved this time.
>>>> .....
>>>> 
>>>> All pointing to log IO not completing.
>>>> 
>> ....
>>> So far I do not have a reproducible test case,
>> 
>> Ok. What sort of load is being placed on the machine?
> Hello, generally the load is low, it mainly serves out some samba shares.
>
>> 
>> It appears that both the xfslogd and the xfsdatad on CPU 0 are in
>> the running state but don't appear to be consuming any significant
>> CPU time. If they remain like this then I think that means they are
>> stuck waiting on the run queue.  Do these XFS threads always appear
>> like this when the hang occurs? If so, is there something else that
>> is hogging CPU 0 preventing these threads from getting the CPU?
> Yes, the XFS threads show up like this on each time the kernel crashed.  So 
> far
> with 2.6.30.9 after ~48hrs+ it has not crashed.  So it appears to be some 
> issue
> between 2.6.30.9 and 2.6.31.x when this began happening.  Any recommendations
> on how to catch this bug w/certain options enabled/etc?
>
>
>> 
>> Cheers,
>> 
>> Dave.
>> -- 
>> Dave Chinner
>> david at fromorbit.com
>> 
>

Uptime with 2.6.30.9:

  06:18:41 up 2 days, 14:10, 14 users,  load average: 0.41, 0.21, 0.07

No issues yet, so it first started happening in 2.6.(31).(x).

Any further recommendations on how to debug this issue?  BTW: Do you view this
as an XFS bug or MD/VFS layer issue based on the logs/output thus far?

Justin.