Re: [PATCH] cpufreq: schedutil: Don't skip freq update when limits change

["Rafael J. Wysocki" <rafael@kernel.org>]

On Mon, Jul 29, 2019 at 10:32 AM Viresh Kumar <viresh.kumar@linaro.org> wrote:
>
> On 29-07-19, 00:55, Doug Smythies wrote:
> > On 2019.07.25 23:58 Viresh Kumar wrote:
> > > Hmm, so I tried to reproduce your setup on my ARM board.
> > > - booted only with CPU0 so I hit the sugov_update_single() routine
> > > - And applied below diff to make CPU look permanently busy:
> > >
> > > -------------------------8<-------------------------
> > >diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> > > index 2f382b0959e5..afb47490e5dc 100644
> > > --- a/kernel/sched/cpufreq_schedutil.c
> > > +++ b/kernel/sched/cpufreq_schedutil.c
> > > @@ -121,6 +121,7 @@ static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
> > >         if (!sugov_update_next_freq(sg_policy, time, next_freq))
> > >                return;
> > >
> > > +       pr_info("%s: %d: %u\n", __func__, __LINE__, freq);
> >
> > ?? there is no "freq" variable here, and so this doesn't compile. However this works:
> >
> > +       pr_info("%s: %d: %u\n", __func__, __LINE__, next_freq);
>
> There are two paths we can take to change the frequency, normal
> sleep-able path (sugov_work) or fast path. Only one of them is taken
> by any driver ever. In your case it is the fast path always and in
> mine it was the slow path.
>
> I only tested the diff with slow-path and copy pasted to fast path
> while giving out to you and so the build issue. Sorry about that.
>
> Also make sure that the print is added after sugov_update_next_freq()
> is called, not before it.
>
> > >         next_freq = cpufreq_driver_fast_switch(policy, next_freq);
> > >        if (!next_freq)
> > >                return;
> > > @@ -424,14 +425,10 @@ static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
> > > #ifdef CONFIG_NO_HZ_COMMON
> > > static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
> > > {
> > > -       unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
> > > -       bool ret = idle_calls == sg_cpu->saved_idle_calls;
> > > -
> > > -       sg_cpu->saved_idle_calls = idle_calls;
> > > -       return ret;
> > > +       return true;
> > >  }
> > >  #else
> > > -static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
> > > +static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return true; }
> > > #endif /* CONFIG_NO_HZ_COMMON */
> > >
> > >  /*
> > > @@ -565,6 +562,7 @@ static void sugov_work(struct kthread_work *work)
> > >         sg_policy->work_in_progress = false;
> > >         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
> > >
> > > +       pr_info("%s: %d: %u\n", __func__, __LINE__, freq);
> > >         mutex_lock(&sg_policy->work_lock);
> > >         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
> > >         mutex_unlock(&sg_policy->work_lock);
> > >
> > > -------------------------8<-------------------------
> > >
> > > Now, the frequency never gets down and so gets set to the maximum
> > > possible after a bit.
> > >
> > > - Then I did:
> > >
> > > echo <any-low-freq-value> > /sys/devices/system/cpu/cpufreq/policy0/scaling_max_freq
> > >
> > > Without my patch applied:
> > >        The print never gets printed and so frequency doesn't go down.
> > >
> > > With my patch applied:
> > >        The print gets printed immediately from sugov_work() and so
> > >        the frequency reduces.
> > >
> > > Can you try with this diff along with my Patch2 ? I suspect there may
> > > be something wrong with the intel_cpufreq driver as the patch fixes
> > > the only path we have in the schedutil governor which takes busyness
> > > of a CPU into account.
> >
> > With this diff along with your patch2 There is never a print message
> > from sugov_work. There are from sugov_fast_switch.
>
> Which is okay. sugov_work won't get hit in your case as I explained
> above.
>
> > Note that for the intel_cpufreq CPU scaling driver and the schedutil
> > governor I adjust the maximum clock frequency this way:
> >
> > echo <any-low-percent> > /sys/devices/system/cpu/intel_pstate/max_perf_pct
>
> This should eventually call sugov_limits() in schedutil governor, this
> can be easily checked with another print message.
>
> > I also applied the pr_info messages to the reverted kernel, and re-did
> > my tests (where everything works as expected). There is never a print
> > message from sugov_work. There are from sugov_fast_switch.
>
> that's fine.
>
> > Notes:
> >
> > I do not know if:
> > /sys/devices/system/cpu/cpufreq/policy*/scaling_max_freq
> > /sys/devices/system/cpu/cpufreq/policy*/scaling_min_freq
> > Need to be accurate when using the intel_pstate driver in passive mode.
> > They are not.
> > The commit comment for 9083e4986124389e2a7c0ffca95630a4983887f0
> > suggests that they might need to be representative.
> > I wonder if something similar to that commit is needed
> > for other global changes, such as max_perf_pct and min_perf_pct?
>
> We are already calling intel_pstate_update_policies() in that case, so
> it should be fine I believe.
>
> > intel_cpufreq/ondemand doesn't work properly on the reverted kernel.
>
> reverted kernel ? The patch you reverted was only for schedutil and it
> shouldn't have anything to do with ondemand.
>
> > (just discovered, not investigated)
> > I don't know about other governors.
>
> When you do:
>
> echo <any-low-percent> > /sys/devices/system/cpu/intel_pstate/max_perf_pct
>
> How soon does the print from sugov_fast_switch() gets printed ?
> Immediately ? Check with both the kernels, with my patch and with the
> reverted patch.
>
> Also see if there is any difference in the next_freq value in both the
> kernels when you change max_perf_pct.
>
> FWIW, we now know the difference between intel-pstate and
> acpi-cpufreq/my testcase and why we see differences here. In the cases
> where my patch fixed the issue (acpi/ARM), we were really changing the
> limits, i.e. policy->min/max. This happened because we touched
> scaling_max_freq directly.
>
> For the case of intel-pstate, you are changing max_perf_pct which
> doesn't change policy->max directly. I am not very sure how all of it
> work really, but at least schedutil will not see policy->max changing.
>
> @Rafael: Do you understand why things don't work properly with
> intel_cpufreq driver ?

I haven't tried to understand this yet, so no.

My somewhat educated guess is that using max_perf_pct has to do with
it, so I would try to retest to see if there's any difference when
scaling_max_freq is used instead of that.