Re: net/bluetooth: workqueue destruction WARNING in hci_unregister_dev

[Jiri Slaby]

Hello,

On 03/18/2016, 09:52 PM, Tejun Heo wrote:
> On Thu, Mar 17, 2016 at 01:00:13PM +0100, Jiri Slaby wrote:
>>>> I have not done that yet, but today, I see:
>>>> destroy_workqueue: name='req_hci0' pwq=ffff88002f590300
>>>> wq->dfl_pwq=ffff88002f591e00 pwq->refcnt=2 pwq->nr_active=0 delayed_works:
>>>>    pwq 12: cpus=0-1 node=0 flags=0x4 nice=-20 active=0/1
>>>>      in-flight: 18568:wq_barrier_func
>>>
>>> So, this means that there's flush_work() racing against workqueue
>>> destruction, which can't be safe. :(
>>
>> But I cannot trigger the WARN_ONs in the attached patch, so I am
>> confused how this can happen :(. (While I am still seeing the destroy
>> WARNINGs.)
> 
> So, no operations should be in progress when destroy_workqueue() is
> called.  If somebody was flushing a work item, the flush call must
> have returned before destroy_workqueue() was invoked, which doesn't
> seem to be the case here.  Can you trigger BUG_ON() or sysrq-t when
> the above triggers?  There must be a task which is flushing a work
> item there and it shouldn't be difficult to pinpoint what's going on
> from it.

The output of sysrq-t is here (> 200k), but I cannot see anything
suspicious in it:
http://www.fi.muni.cz/~xslaby/sklad/panics/jctl.txt

This is what the code does now:
+               if ((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) {
+                       pr_info("%s: name='%s' pwq=%p wq->dfl_pwq=%p
pwq->refcnt=%d pwq->nr_active=%d delayed_works:",
+                                       __func__, wq->name, pwq,
wq->dfl_pwq,
+                                       pwq->refcnt, pwq->nr_active);
+
+                       show_pwq(pwq);
+
+                       mutex_unlock(&wq->mutex);
+                       show_state();
+                       show_workqueue_state();
+                       WARN_ON(1);
+                       return;
+               }

thanks,
-- 
js
suse labs