Re: [PATCH] x86/aperfmperf: Make stale CPU frequency response within limits.

[Rafael J. Wysocki]

Hi Doug,

On Tue, Jul 25, 2023 at 9:12 PM Doug Smythies <dsmythies@xxxxxxxxx> wrote:
>
> Hi Rafael,
>
> On Tue, Jul 25, 2023 at 11:31 AM Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
> > On Tue, Jul 25, 2023 at 2:14 AM Doug Smythies <dsmythies@xxxxxxxxx> wrote:
> > >
> > > Currently, when the CPU frequency is stale the nominal clock frequency
> > > is returned by calls to arch_freq_get_on_cpu(). Some users are
> > > confused by the high reported frequency when their system is idle
> > > and/or it is above a reduced maximum they set.
> > >
> > > This patch will return the policy minimum as the stale frequency reply
> > > from arch_freq_get_on_cpu().
> > >
> > > Reported-by: Yang Jie <yang.jie@xxxxxxxxxxxxxxx>
> > > Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=217597
> > > Signed-off-by: Doug Smythies <dsmythies@xxxxxxxxx>
> > > ---
> > >  arch/x86/kernel/cpu/aperfmperf.c | 13 +++++--------
> > >  drivers/cpufreq/cpufreq.c        | 18 ++++++++++++++++++
> > >  include/linux/cpufreq.h          |  5 +++++
> > >  3 files changed, 28 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/arch/x86/kernel/cpu/aperfmperf.c b/arch/x86/kernel/cpu/aperfmperf.c
> > > index fdbb5f07448f..44cc96864d94 100644
> > > --- a/arch/x86/kernel/cpu/aperfmperf.c
> > > +++ b/arch/x86/kernel/cpu/aperfmperf.c
> > > @@ -418,9 +418,10 @@ unsigned int arch_freq_get_on_cpu(int cpu)
> > >         unsigned long last;
> > >         u64 acnt, mcnt;
> > >
> > > -       if (!cpu_feature_enabled(X86_FEATURE_APERFMPERF))
> > > -               goto fallback;
> > > -
> > > +       if (!cpu_feature_enabled(X86_FEATURE_APERFMPERF)){
> > > +               freq = cpufreq_quick_get(cpu);
> > > +               return freq ? freq : cpufreq_quick_get_min(cpu);
> > > +       }
> > >         do {
> > >                 seq = raw_read_seqcount_begin(&s->seq);
> > >                 last = s->last_update;
> > > @@ -433,13 +434,9 @@ unsigned int arch_freq_get_on_cpu(int cpu)
> > >          * which covers idle and NOHZ full CPUs.
> > >          */
> > >         if (!mcnt || (jiffies - last) > MAX_SAMPLE_AGE)
> > > -               goto fallback;
> > > +               return cpufreq_quick_get_min(cpu);
> > >
> > >         return div64_u64((cpu_khz * acnt), mcnt);
> > > -
> > > -fallback:
> > > -       freq = cpufreq_quick_get(cpu);
> > > -       return freq ? freq : cpu_khz;
> >
> > It looks to me like modifying cpufreq_quick_get) to return policy->min
> > if policy->cur is 0 would work in a similar way, wouldn't it?
>
> For the configuration of intel_cpufreq driver (intel_pstate in
> passive mode), schedutil governor, HWP enabled, for
> a stale frequency  policy->cur is not 0 and will always
> be whatever the min value was when the driver was initialized,
> regardless of what has been set since.

So I would prefer to address this in the intel_pstate driver than to
work around it in the core.

> The patch I submitted deals with that situation also.
>
> A complete list of driver/governor/HWP stale frequency
> replies can be found on the bugzilla report at:
>
> https://bugzilla.kernel.org/attachment.cgi?id=304694
>
> There might be push back on some of the performance
> governor stale frequency replies. I could not figure out
> a performance governor dependant reply.
>
> Also there are other callers to cpufreq_quick_get
> and I was not sure I could mess with the function
> response for them. For example
> drivers/devfreq/tegra30-devfreq.c
> and drivers/thermal/cpufreq_cooling.c
> and drivers/powercap/dtpm_cpu.c

IIUC, all of the above rely on policy->cur being nonzero.

There are other users doing questionable things when
cpufreq_quick_get() returns 0 that I think would be better off if the
min is returned instead.