Re: [PATCH] cpufreq: stats: Make the stats code non-modular

[Rafael J. Wysocki]

On Mon, May 30, 2016 at 6:53 AM, Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> On 28-05-16, 15:15, Rafael J. Wysocki wrote:
>> It does, but there's a problem.
>>
>> If fast frequency switching is in use, the stats attributes just sit there
>> full of zeros (because the stats are not updated then) which is confusing.
>>
>> Of course, the ultimate solution here would be to make the stats actually
>> work with fast frequency switching, but that requires some major surgery of
>> the stats code, so for now I'd like to simply make those attributes return
>> nothing if fast frequency switching is enabled for the policy (returning
>> -EBUSY from there, which would have been cleaner, unfortunately breaks
>> powertop).
>
> What's stopping us from doing that? Sorry I don't remember that well :(
>
> I mean, why can't we call cpufreq_stats_record_transition() somehow via that
> code ?

Because it does things that aren't allowed to be done from the
scheduler paths (using the non-raw spinlock in the first place).

It also isn't efficient enough.  For example, the fast switch path in
schedutil for non-shared policies is lockless and it would be a shame
to add a lock to it just for the stats.