Re: Hard lockup in blk_mq_free_request() / wbt_done() / wake_up_all()

[Chris Boot]

On 12/06/18 17:22, Jens Axboe wrote:
> On 6/12/18 10:19 AM, Chris Boot wrote:
>> On 12/06/18 17:09, Jens Axboe wrote:
>>> On 6/12/18 9:38 AM, Chris Boot wrote:
>>>> Hi folks,
>>>>
>>>> I maintain a large (to me) system with 112 threads (4x Intel E7-4830 v4)
>>>> which has a MegaRAID SAS 9361-24i controller. This system is currently
>>>> running Debian's 4.16.12 kernel (from stretch-backports) with blk_mq
>>>> enabled.
>>>>
>>>> I've run into a lockup which appears to involve blq_mq and writeback
>>>> throttling. It's hard to tell if I've run into this same thing with
>>>> older kernels; I'm trying to track down a deadlock but so far I've been
>>>> fairly certain that involved the OOM killer, but this doesn't seem to.
>> [snip]
>>>
>>> Hmm that's really weird, I don't see how we could be spinning on the
>>> waitqueue lock like that. I haven't seen any wbt bug reports like this
>>> before.
>>>
>>> Are things generally stable if you just turn off wbt? You can do that
>>> for sda, for instance, by doing:
>>>
>>> # echo 0 > /sys/block/sda/queue/wbt_lat_usec
>>>
>>> It'd be interesting to get this data point. Eg leave blk-mq enabled, and
>>> then just disable wbt.
>>
>> Hi Jens,
>>
>> Thanks for the speedy response. I'll see if I can get that tested soon;
>> if the system is stable without blk_mq I can see the users wanting to
>> keep it that way for a while. I'll let you know.
> 
> Understandable. I just get suspicious of the general state of the system,
> if it's locking up there. Could be a hardware issue, or a bug in some
> other area that's messing things up. I have wbt running on literally
> hundreds of thousands of boxes and haven't seen a lockup like this.

Hi Jens,

I got an opportunity yesterday to do some testing. I can't get this
system to crash with blk-mq disabled, or with blk-mq enabled but wbt
disabled. I have a reproducer workload I can launch against the system
and it seems to crash reliably with this, but I doubt I can share it
with you.

I do, however, have a task state dump (SysRq+T) that I managed to get
out of the server once it started locking up. It's pretty large, so I
uploaded it to my Dropbox for now:

https://www.dropbox.com/s/fyo1ab6mmcqk8fq/crash-1.log.gz?dl=0

Hope this helps!

Cheers,
Chris

-- 
Chris Boot
bootc@xxxxxx