Re: 4.14: WARNING: CPU: 4 PID: 2895 at block/blk-mq.c:1144 with virtio-blk

[Christian Borntraeger]

On 11/20/2017 08:42 PM, Jens Axboe wrote:
> On 11/20/2017 12:29 PM, Christian Borntraeger wrote:
>>
>>
>> On 11/20/2017 08:20 PM, Bart Van Assche wrote:
>>> On Fri, 2017-11-17 at 15:42 +0100, Christian Borntraeger wrote:
>>>> This is 
>>>>
>>>> b7a71e66d (Jens Axboe                2017-08-01 09:28:24 -0600 1141)     * are mapped to it.
>>>> b7a71e66d (Jens Axboe                2017-08-01 09:28:24 -0600 1142)     */
>>>> 6a83e74d2 (Bart Van Assche           2016-11-02 10:09:51 -0600 1143)    WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
>>>> 6a83e74d2 (Bart Van Assche           2016-11-02 10:09:51 -0600 1144)            cpu_online(hctx->next_cpu));
>>>> 6a83e74d2 (Bart Van Assche           2016-11-02 10:09:51 -0600 1145) 
>>>> b7a71e66d (Jens Axboe                2017-08-01 09:28:24 -0600 1146)    /*
>>>
>>> Did you really try to figure out when the code that reported the warning
>>> was introduced? I think that warning was introduced through the following
>>> commit:
>>
>> This was more a cut'n'paste to show which warning triggered since line numbers are somewhat volatile.
>>
>>>
>>> commit fd1270d5df6a005e1248e87042159a799cc4b2c9
>>> Date:   Wed Apr 16 09:23:48 2014 -0600
>>>
>>>     blk-mq: don't use preempt_count() to check for right CPU
>>>      
>>>     UP or CONFIG_PREEMPT_NONE will return 0, and what we really
>>>     want to check is whether or not we are on the right CPU.
>>>     So don't make PREEMPT part of this, just test the CPU in
>>>     the mask directly.
>>>
>>> Anyway, I think that warning is appropriate and useful. So the next step
>>> is to figure out what work item was involved and why that work item got
>>> executed on the wrong CPU.
>>
>> It seems to be related to virtio-blk (is triggered by fio on such disks). Your comment basically
>> says: "no this is not a known issue" then :-)
>> I will try to take a dump to find out the work item
> 
> blk-mq does not attempt to freeze/sync existing work if a CPU goes away,
> and we reconfigure the mappings. So I don't think the above is unexpected,
> if you are doing CPU hot unplug while running a fio job.

I did a cpu hot plug (adding a CPU) and I started fio AFTER that.

 
> While it's a bit annoying that we trigger the WARN_ON() for a condition
> that can happen, we're basically interested in it if it triggers for
> normal operations.

I think we should never trigger a WARN_ON on conditions that can happen. I know some
folks enabling panic_on_warn to detect/avoid data integrity issues. FWIW, this also seems
to happen wit 4.13 and 4.12