Re: [PATCH] cpufreq, store_scaling_governor requires policy->rwsem to be held for duration of changing governors [v2]

[Prarit Bhargava <prarit@redhat.com>]

On 08/05/2014 03:46 AM, Viresh Kumar wrote:
> On 2 August 2014 01:06, Stephen Boyd <sboyd@codeaurora.org> wrote:
>> I have the same options. The difference is that my driver has a governor
>> per policy. That's set with the CPUFREQ_HAVE_GOVERNOR_PER_POLICY flag.
> 
> You may call me stupid but I got a bit confused after looking into the code
> again. Why does the crash dump depends on this flag?

Nope, not a stupid question.  After reproducing (finally!) yesterday I've been
wondering the same thing.

> 
> We *always* remove the governor specific directory while switching governors
> (Ofcourse only if its updated for All CPUs). And so on a dual core platform,
> where both CPU 0 & 1 share a clock line, switching of governors should result
> in this crash dump?

I've been looking into *exactly* this.  On any platform where
cpu_weight(affected_cpus) == 1 for a particular cpu this lockdep trace should
happen.


> 
> I may know the answer to the stupid question I had, but not sure why that is a
> problem. The only (and quite significant) difference that this flag makes
> is the location of governor-specific directory:
> - w/o this flag: /sys/devices/system/cpu/cpufreq/<here>
> - w/ this flag: /sys/devices/system/cpu/cpu*/cpufreq/<here>
> 

cpufreq_global_kobject vs the policy's kobject.

> So, is there some issue with the sysfs lock for <cpu*/cpufreq/> node as while
> switching governor we change  <cpu*/cpufreq/scaling_governor> at the same
> location?

That's what I'm wondering too.  I'm going to instrument the code to find out
this morning.  I'm wondering if this comes down to a lockdep class issue
(perhaps lockdep puts globally defined locks like cpufreq_global_kobject in a
different class?).

P.