Re: CPU excessively long times between frequency scaling driver calls - bisected

[Thomas Gleixner <tglx@linutronix.de>]

Rafael,

On Tue, Feb 22 2022 at 19:04, Rafael J. Wysocki wrote:
> On Tue, Feb 22, 2022 at 8:41 AM Feng Tang <feng.tang@intel.com> wrote:
>> There is periodic activity in between, related to active load balancing
>> in scheduler (since last frequency was higher these small work will
>> also run at higher frequency). But those threads are not CFS class, so
>> scheduler callback will not be called for them.
>>
>> So removing the patch removed a trigger which would have caused a
>> sched_switch to a CFS task and call a cpufreq/intel_pstate callback.
>
> And so this behavior needs to be restored for the time being which
> means reverting the problematic commit for 5.17 if possible.

No. This is just papering over the problem. Just because the clocksource
watchdog has the side effect of making cpufreq "work", does not make it
a prerequisite for cpufreq. The commit unearthed a problem in the
cpufreq code, so it needs to be fixed there.

Even if we'd revert it then, you can produce the same effect by adding
'tsc=reliable' to the kernel command line which disables the clocksource
watchdog too. The commit is there to deal with modern hardware without
requiring people to add 'tsc=reliable' to the command line.

Thanks,

        tglx