Re: [PATCH 1/2] cpufreq: schedutil: Fix superfluous updates caused by need_freq_update

[Xuewen Yan]

Hi Sultan,

On Wed, Apr 9, 2025 at 12:50 AM Stephan Gerhold
<stephan.gerhold@xxxxxxxxxx> wrote:
>
> Hi Sultan,
>
> On Tue, Apr 08, 2025 at 08:22:20AM -0700, Sultan Alsawaf wrote:
> > On Tue, Apr 08, 2025 at 10:59:31AM +0200, Stephan Gerhold wrote:
> > > On Wed, Dec 11, 2024 at 05:57:32PM -0800, Sultan Alsawaf wrote:
> > > > From: "Sultan Alsawaf (unemployed)" <sultan@xxxxxxxxxxxxxxx>
> > > >
> > > > A redundant frequency update is only truly needed when there is a policy
> > > > limits change with a driver that specifies CPUFREQ_NEED_UPDATE_LIMITS.
> > > >
> > > > In spite of that, drivers specifying CPUFREQ_NEED_UPDATE_LIMITS receive a
> > > > frequency update _all the time_, not just for a policy limits change,
> > > > because need_freq_update is never cleared.
> > > >
> > > > Furthermore, ignore_dl_rate_limit()'s usage of need_freq_update also leads
> > > > to a redundant frequency update, regardless of whether or not the driver
> > > > specifies CPUFREQ_NEED_UPDATE_LIMITS, when the next chosen frequency is the
> > > > same as the current one.
> > > >
> > > > Fix the superfluous updates by only honoring CPUFREQ_NEED_UPDATE_LIMITS
> > > > when there's a policy limits change, and clearing need_freq_update when a
> > > > requisite redundant update occurs.
> > > >
> > > > This is neatly achieved by moving up the CPUFREQ_NEED_UPDATE_LIMITS test
> > > > and instead setting need_freq_update to false in sugov_update_next_freq().
> > > >
> > > > Signed-off-by: Sultan Alsawaf (unemployed) <sultan@xxxxxxxxxxxxxxx>
> > > > ---
> > > >  kernel/sched/cpufreq_schedutil.c | 4 ++--
> > > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> > > > index 28c77904ea74..e51d5ce730be 100644
> > > > --- a/kernel/sched/cpufreq_schedutil.c
> > > > +++ b/kernel/sched/cpufreq_schedutil.c
> > > > @@ -83,7 +83,7 @@ static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
> > > >
> > > >   if (unlikely(sg_policy->limits_changed)) {
> > > >           sg_policy->limits_changed = false;
> > > > -         sg_policy->need_freq_update = true;
> > > > +         sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
> > > >           return true;
> > > >   }
> > > >
> > > > @@ -96,7 +96,7 @@ static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
> > > >                              unsigned int next_freq)
> > > >  {
> > > >   if (sg_policy->need_freq_update)
> > > > -         sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
> > > > +         sg_policy->need_freq_update = false;
> > > >   else if (sg_policy->next_freq == next_freq)
> > > >           return false;
> > > >
> > >
> > > This patch breaks cpufreq throttling (e.g. for thermal cooling) for
> > > cpufreq drivers that:
> > >
> > >  - Have policy->fast_switch_enabled/fast_switch_possible set, but
> > >  - Do not have CPUFREQ_NEED_UPDATE_LIMITS flag set
> > >
> > > There are several examples for this in the tree (search for
> > > "fast_switch_possible"). Of all those drivers, only intel-pstate and
> > > amd-pstate (sometimes) set CPUFREQ_NEED_UPDATE_LIMITS.
> > >
> > > I can reliably reproduce this with scmi-cpufreq on a Qualcomm X1E
> > > laptop:
> > >
> > >  1. I added some low temperature trip points in the device tree,
> > >     together with passive cpufreq cooling.
> > >  2. I run a CPU stress test on all CPUs and monitor the temperatures
> > >     and CPU frequencies.
> > >
> > > When using "performance" governor instead of "schedutil", the CPU
> > > frequencies are being throttled as expected, as soon as the temperature
> > > trip points are reached.
> > >
> > > When using "schedutil", the CPU frequencies stay at maximum as long as
> > > the stress test is running. No throttling happens, so the device heats
> > > up far beyond the defined temperature trip points. Throttling is applied
> > > only after stopping the stress test, since this forces schedutil to
> > > re-evaluate the CPU frequency.
> > >
> > > Reverting this commit fixes the problem.
> > >
> > > Looking at the code, I think the problem is that:
> > >  - sg_policy->limits_changed does not result in
> > >    sg->policy->need_freq_update without CPUFREQ_NEED_UPDATE_LIMITS
> > >    anymore, and
> > >  - Without sg->policy->need_freq_update, get_next_freq() skips calling
> > >    cpufreq_driver_resolve_freq(), which would normally apply the policy
> > >    min/max constraints.
> > >
> > > Do we need to set CPUFREQ_NEED_UPDATE_LIMITS for all cpufreq drivers
> > > that set policy->fast_switch_possible? If I'm reading the documentation
> > > comment correctly, that flag is just supposed to enable notifications if
> > > the policy min/max changes, but the resolved target frequency is still
> > > the same. This is not the case here, the target frequency needs to be
> > > throttled, but schedutil isn't applying the new limits.
> > >
> > > Any suggestions how to fix this? I'm happy to test patches with my
> > > setup.
> >
> > Thank you for reporting this. As I see it, sg_policy->need_freq_update is
> > working correctly now; however, sg_policy->limits_changed relied on the broken
> > behavior of sg_policy->need_freq_update and therefore sg_policy->limits_changed
> > needs to be fixed.
>
> Thanks for the quick reply and the patch!
>
> >
> > Can you try this patch:
> >
> > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> > index 1a19d69b91ed3..f37b999854d52 100644
> > --- a/kernel/sched/cpufreq_schedutil.c
> > +++ b/kernel/sched/cpufreq_schedutil.c
> > @@ -82,7 +82,6 @@ static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
> >               return false;
> >
> >       if (unlikely(sg_policy->limits_changed)) {
> > -             sg_policy->limits_changed = false;
> >               sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
> >               return true;
> >       }
> > @@ -171,9 +170,11 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy,
> >       freq = get_capacity_ref_freq(policy);
> >       freq = map_util_freq(util, freq, max);
> >
> > -     if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
> > +     if (freq == sg_policy->cached_raw_freq && !sg_policy->limits_changed &&
> > +         !sg_policy->need_freq_update)

We also should add the limits_changed in the sugov_update_single_freq().

> >               return sg_policy->next_freq;
> >
> > +     sg_policy->limits_changed = false;

Maybe this should be add in the sugov_update_next_freq()，because, both
sugov_update_single_freq() and sugov_update_shared(),
the sugov_update_next_freq() is after the freq check.

> >       sg_policy->cached_raw_freq = freq;
> >       return cpufreq_driver_resolve_freq(policy, freq);
> >  }
> >
>
> This is working correctly for me, CPU frequency is being throttled again
> when the temperature trip points are reached. If you send this, feel
> free to add:
>
> Tested-by: Stephan Gerhold <stephan.gerhold@xxxxxxxxxx>
>
> Thanks!
> Stephan
>