Re: workqueue lockup debug

[John Garry <john.g.garry@oracle.com>]

On 24/10/2024 19:56, Tejun Heo wrote:

Hi Tejun,

> 
> On Thu, Oct 24, 2024 at 04:49:40PM +0100, John Garry wrote:
>> Hi workqueue and scheduler maintainers,
>>
>> As reported in https://urldefense.com/v3/__https://lore.kernel.org/linux-fsdevel/df9db1ce-17d9-49f1-ab6d-7ed9a4f1f9c0@oracle.com/T/*m506b9edb1340cdddd87c6d14d20222ca8d7e8796__;Iw!!ACWV5N9M2RV99hQ!Jn5gFi_KIzLb-7uCOvHRIcf9Zd4RyHCTnm15Lo8u0KsgaMmwEPZHH2m4l-w4Fr-b6aRICoRWb5DU$ ,
>> I am experiencing a workqueue lockup for v6.12-rcX.
>>
>> At the point it occurs, the system becomes unresponsive and I cannot bring
>> it back to life.
>>
>> Enabling /proc/sys/kernel/softlockup_all_cpu_backtrace does not give
>> anything extra in the way of debug. All I get is something like this:
>>
>> Message from syslogd@jgarry-atomic-write-exp-e4-8-instance-20231214-1221 at
>> Oct 24 15:34:02 ...
>>   kernel:watchdog: BUG: soft lockup - CPU#29 stuck for 22s! [mysqld:14352]
>>
>> Message from syslogd@jgarry-atomic-write-exp-e4-8-instance-20231214-1221 at
>> Oct 24 15:34:02 ...
>>   kernel:BUG: workqueue lockup - pool cpus=4 node=0 flags=0x0 nice=0 stuck
>> for 30s!
>>
>> Message from syslogd@jgarry-atomic-write-exp-e4-8-instance-20231214-1221 at
>> Oct 24 15:34:02 ...
>>   kernel:BUG: workqueue lockup - pool cpus=31 node=0 flags=0x0 nice=0 stuck
>> for 49s!
>> ^C
>>
>> Can you advise on a robust method to get some debug from this system?
>>
>> Maybe this is a scheduler issue, as Dave mentioned in that same thread.
> 
> So, while the symptom *could* be a workqueue bug, wq lockups are usually
> more a side-effect of some other problem.

Sure

> All that wq is saying is that some
> work item hasn't made forward progress for too long. Given that there are
> soft lockup messages, which can't be caused by wq lockups but can definitely
> trigger wq lockups, I'd start with debugging what's happening with the CPU
> that triggered soft lockup. 

ok

> Soft lockups are nice in that they're
> synchronous failures, so probably start with the backtrace?
> 

I'm not getting any backtrace, which is not helping the issue.

But I'll continue to look.

Thanks,
John