Re: [PATCH V3 0/4] cpufreq: cppc: Add support for frequency invariance

[Vincent Guittot]

On Thu, 24 Jun 2021 at 12:48, Ionela Voinescu <ionela.voinescu@xxxxxxx> wrote:
>
> Hi guys,
>
> On Thursday 24 Jun 2021 at 11:49:53 (+0200), Vincent Guittot wrote:
> > On Thu, 24 Jun 2021 at 04:54, Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> > >
> > > On 23-06-21, 08:57, Qian Cai wrote:
> > > > Viresh, I am afraid I don't feel comfortable yet. I have a few new tests in
> > > > development, and will provide an update once ready.
> > >
> > > Oh sure, np.
> > >
> > > > Also, I noticed the delivered perf is even smaller than lowest_perf (100).
> > >
> > > > # cat /sys/devices/system/cpu/cpu8/acpi_cppc/feedback_ctrs
> > > >  ref:103377547901 del:54540736873
> > > > # cat /sys/devices/system/cpu/cpu8/acpi_cppc/feedback_ctrs
> > > >  ref:103379170101 del:54541599117
> > > >
> > > > 100 * (54541599117 - 54540736873) / (103379170101 - 103377547901) = 53
> >
> > I'm not sure that I understand your point. The formula above says that
> > cpu8 run @ 53% of nominal performance
> >
>
> I think this is based on a previous example Qian had where:
>
> /sys/devices/system/cpu/cpu0/acpi_cppc/highest_perf
> 300
> /sys/devices/system/cpu/cpu0/acpi_cppc/lowest_freq
> 1000
> /sys/devices/system/cpu/cpu0/acpi_cppc/lowest_perf
> 100
> /sys/devices/system/cpu/cpu0/acpi_cppc/reference_perf
> 100
>
> ..so the 100 is not from obtaining percentage, is the reference
> performance.
>
> The logic of the formula is to obtain the delivered performance when
> knowing the number of ticks for each counter, so:
>
> So if one gets (103379170101 - 103377547901) ticks for the counter at
> running at 1GHz(perf 100), what is the frequency of the core, if its
> counter ticked (54541599117 - 54540736873) times in the same interval
> of time?
>
> The answer is 530MHz(perf 53), which is lower than the lowest frequency
> at 1GHz(perf 100).

But the nominal_perf is 280 and not 100 if i'm not wrong so the perf
value is 148 > lowest_perf in this case


>
>
> > > >
> > > > My understanding is that the delivered perf should fail into the range between
> > > > lowest_perf and highest_perf. Is that assumption correct? This happens on
> > > > 5.4-based kernel, so I am in process running your series on that system to see
> > > > if there is any differences. In any case, if it is a bug it is pre-existing,
> > > > but I'd like to understand a bit better in that front first.
> > >
> > > Vincent:
> > >
> > > Can that happen because of CPU idle ?
> > >
>
> Not if the counters are implemented properly. The kernel considers that
> both reference and delivered performance counters should stop or reset
> during idle. The kernel would not account for idle itself.
>
> If the reference performance counter does not stop during idle, while
> the core performance counter (delivered) does stop, the behavior above
> should be seen very often.
>
> Qian, do you see these small delivered performance values often or
> seldom?
>
> Thanks,
> Ionela.
>
> > > --
> > > viresh