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

[Doug Smythies <dsmythies@telus.net>]

Hi Rafael,

On Wed, Jul 26, 2023 at 7:43 AM Rafael J. Wysocki <rafael@kernel.org> wrote:
> On Tue, Jul 25, 2023 at 9:12 PM Doug Smythies <dsmythies@telus.net> wrote:
> > On Tue, Jul 25, 2023 at 11:31 AM Rafael J. Wysocki <rafael@kernel.org> wrote:
> > > On Tue, Jul 25, 2023 at 2:14 AM Doug Smythies <dsmythies@telus.net> 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@linux.intel.com>
> > > > Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=217597
> > > > Signed-off-by: Doug Smythies <dsmythies@telus.net>
> > > > ---
> > > >  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.

Okay, but I would need some help with it. I already tried to
figure out a fix before starting this thread, and have tried
again since your comment. I haven't been able to figure
it out.

An example of the issue:
Use the ondemand governor and set some
minimum and also put a load on CPU 5 such that the
governor asks for a non-min and non-max pstate.
Then switch to the schedutil governor, and terminate
the load on CPU 5, and look at CPU frequencies:

$ grep . /sys/devices/system/cpu/cpufreq/policy*/scaling_cur_freq
/sys/devices/system/cpu/cpufreq/policy0/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy10/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy11/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy1/scaling_cur_freq:4799871
/sys/devices/system/cpu/cpufreq/policy2/scaling_cur_freq:4800027
/sys/devices/system/cpu/cpufreq/policy3/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy4/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy5/scaling_cur_freq:1300000
/sys/devices/system/cpu/cpufreq/policy6/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy7/scaling_cur_freq:4800736
/sys/devices/system/cpu/cpufreq/policy8/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy9/scaling_cur_freq:1000000

non stale frequencies are identified by non round numbers.
observe that CPU 5 still indicates pstate 13.
observe the other stale frequencies are the pstate 10 min
that I set when the governor was ondemand.
Now change the minimum to 1.1 GHz and check it:

$ grep . /sys/devices/system/cpu/cpufreq/policy*/scaling_min_freq
/sys/devices/system/cpu/cpufreq/policy0/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy10/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy11/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy1/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy2/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy3/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy4/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy5/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy6/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy7/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy8/scaling_min_freq:1100000
/sys/devices/system/cpu/cpufreq/policy9/scaling_min_freq:1100000

and look at current again:

$ grep . /sys/devices/system/cpu/cpufreq/policy*/scaling_cur_freq
/sys/devices/system/cpu/cpufreq/policy0/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy10/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy11/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy1/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy2/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy3/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy4/scaling_cur_freq:1000000
/sys/devices/system/cpu/cpufreq/policy5/scaling_cur_freq:1300000
/sys/devices/system/cpu/cpufreq/policy6/scaling_cur_freq:4800585
/sys/devices/system/cpu/cpufreq/policy7/scaling_cur_freq:4800177
/sys/devices/system/cpu/cpufreq/policy8/scaling_cur_freq:4799992
/sys/devices/system/cpu/cpufreq/policy9/scaling_cur_freq:4800015

Observe the stale frequencies are unchanged and outside of the
range limits.

> > 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.

Okay, I'll submit a new patch shortly, with this:

diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
index 50bbc969ffe5..4e91169a83f5 100644
--- a/drivers/cpufreq/cpufreq.c
+++ b/drivers/cpufreq/cpufreq.c
@@ -1768,7 +1768,7 @@ unsigned int cpufreq_quick_get(unsigned int cpu)

        policy = cpufreq_cpu_get(cpu);
        if (policy) {
-               ret_freq = policy->cur;
+               ret_freq = policy->cur ? policy->cur : policy->min;
                cpufreq_cpu_put(policy);
        }

The testing results are in the bugzilla report here:
https://bugzilla.kernel.org/attachment.cgi?id=304734

... Doug