Re: CPUFREQ-CPUHOTPLUG: Possible circular locking dependency

[Gautham R Shenoy <ego@in.ibm.com>]

Hi Venki,
On Wed, Dec 06, 2006 at 10:27:01AM -0800, Pallipadi, Venkatesh wrote:

> But, if we make cpufreq more affected_cpus aware and have a per_cpu
> target()
> call by moving set_cpus_allowed() from driver into cpufreq core and
> define
> the target function to be atomic/non-sleeping type, then we really don't
> need a hotplug lock for the driver any more. Driver can have
> get_cpu/put_cpu
> pair to disable preemption and then change the frequency.

Well, we would still need to keep the affected_cpus map in sync with the
cpu_online map. That would still require hotplug protection, right?

Besides, I would love to see a way of implementing target function to be
atomic/non-sleeping type. But as of now, the target functions call
cpufreq_notify_transition which might sleep.

That's not the last of my worries. The ondemand-workqueue interaction
in the cpu_hotplug callback path can cause a deadlock if we go for
per-subsystem hotcpu mutexes. Can you think of a way by which we can 
avoid destroying the kondemand workqueue from the cpu-hotplug callback
path ? Please see http://lkml.org/lkml/2006/11/30/9 for the
culprit-callpath.

Thanks and Regards
gautham.
-- 
Gautham R Shenoy
Linux Technology Center
IBM India.
"Freedom comes with a price tag of responsibility, which is still a bargain,
because Freedom is priceless!"