Re: [patch 2/3] cpufreq: Define dbs_mutex purpose and cleanup its usage

[Mathieu Desnoyers <mathieu.desnoyers-scC8bbJcJLCw5LPnMra/2Q@public.gmane.org>]

* Pallipadi, Venkatesh (venkatesh.pallipadi-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org) wrote:
> On Thu, 2009-06-25 at 12:46 -0700, Mathieu Desnoyers wrote:
> > * venkatesh.pallipadi-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org (venkatesh.pallipadi-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org) wrote:
> > > Commit b14893a62c73af0eca414cfed505b8c09efc613c although it was very
> > > much needed to cleanup ondemand timer cleanly, openup a can of worms
> > > related to locking dependencies in cpufreq.
> > > 
> > > Patch here defines the need for dbs_mutex and cleans up its usage in
> > > ondemand governor. This also resolves the lockdep warnings reported here
> > > 
> > > http://lkml.indiana.edu/hypermail/linux/kernel/0906.1/01925.html
> > > 
> 
> > > @@ -598,14 +593,16 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
> > >  				max(min_sampling_rate,
> > >  				    latency * LATENCY_MULTIPLIER);
> > >  		}
> > > +		mutex_unlock(&dbs_mutex);
> > > +
> > >  		dbs_timer_init(this_dbs_info);
> > >  
> > > -		mutex_unlock(&dbs_mutex);
> > >  		break;
> > >  
> > >  	case CPUFREQ_GOV_STOP:
> > > -		mutex_lock(&dbs_mutex);
> > >  		dbs_timer_exit(this_dbs_info);
> > 
> > Hrm, so.. how do we protect against concurrent :
> > 
> > CPUFREQ_GOV_START/CPUFREQ_GOV_STOP now ?
> 
> concurrent _START _STOP across CPUs does not matter for timer_init and
> timer_exit.

Given those are per-cpu anyway I guess. Hopefully it works OK with CPU
hotplug.

> On same CPU, there cannot be two concurrent _START as upper
> level cpufreq will have policy_rwsem in write mode.

Agreed.

> I cannot think of a
> flow where _START and _STOP on same CPU is possible.
> 

_STOP is not protected by any mutex now. So it could be preempted, and
then a _START executed, and there is your race.

> However two concurrent _STOP for same CPU is still possible, as we are
> releasing the rwsem lock before STOP callback. "Back to drawing board"
> time to figure this all out..

I fear that it is indeed the case. If you can come up with a document
explaining the expected interactions between :

- cpu hotplug
- policy lock
- cpufreq driver lock
- timer lock

that would be awesome. :)

Mathieu

> 
> Thanks,
> Venki
> 

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68