Re: md data-check causes soft lockup

[Gabriele Trombetti <gabriele.trombetti@itb.cnr.it>]

Robin Hill wrote:
> On Tue Sep 22, 2009 at 07:59:45AM -0700, Lee Howard wrote:
>
>   
>> Majed B. wrote:
>>     
>>> I must have missed that part. It may not work for your case, but worth trying.
>>>
>>> Perhaps Neil Brown, or someone involved could shed some light on this.
>>>
>>> If I remember correctly, those soft lockups were harmless anyway.
>>>   
>>>       
>> Not harmless for production use.  Yes, data is not harmed, and yes, the 
>> problem state does recover when the data-check finishes, but during the 
>> data-check the system is virtually unresponsive and all other use of the 
>> system is stalled.
>>
>>     
> Are you sure this is caused by these soft lockups, and that you're not
> just running with too high a /sys/block/mdX/md/sync_speed_max setting?
> I've had issues with this on some servers, where the I/O demand for the
> sync/check is causing the system to become totally unresponsive.
>   

That's correct for me in the sense that lowering sync_speed_max solves
the problem, see my post, however I'd call it a bug if a value of
sync_speed_max too high starves the system forever. The resync is
supposed to be less prioritarian than normal I/O disk operations, but it
doesn't happen this way. Also note that lowering the value of
stripe_cache_size also solves the problem: how would this fit into your
reasoning?

(BTW I have not checked the mentioned patch yet, I'm not sure I can do
that in a short time because our servers are into production now)