Re: [BUG] While changing the cpufreq governor, kernel hits a bug in workqueue.c

[Johannes Weiner <hannes@saeurebad.de>]

Johannes Weiner <hannes@saeurebad.de> writes:

> Hi Nageswara,
>
> Nageswara R Sastry <rnsastry@linux.vnet.ibm.com> writes:
>
>> Any updates on this bug.
>
> Sorry.  I was just looking into it again and still see no resolution.
> Added Venkatesh and Alexey to CC.

Sorry, wrong address for Alexey.  Hope this one works.

> To summarize again:
>
> The problem we have is that cpufreq_ondemand has a self-rearming worker
> function which we should but can not cancel synchroneously with the
> current code.  The locking order is like this:
>
> store()
>   policy_rwsem write
>     cpufreq_governor_dbs()
>     dbs_mutex
>
> work lock
>   do_dbs_timer()
>     policy_rwsem write
>
> which gives a locking hierarchy of:
>
> work lock
>   policy_rwsem
>     dbs_mutex
>
> Now, we _should_ cancel the worker synchroneously, but that won't fly
> since when cpufreq_governor_dbs() is called, we already hold
> policy_rwsem and grabbing the work lock is illegal.
>
> I will continue to read code and see if I can come up with a solution,
> but any help is really appreciated.
>
> I have not yet looked at cpufreq_conservative...
>
>> Thanks and Regards
>> R.Nageswara Sastry
>
> Thanks for your patience and insistence,
>
> 	Hannes