Re: [PATCH] cpufreq: schedutil: Don't ignore limit changes when util is unchanged

[Rafael J. Wysocki]

On Thu, Apr 10, 2025 at 9:50 PM Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
>
> On Thu, Apr 10, 2025 at 9:26 PM Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx> wrote:
> >
> > On Thu, Apr 10, 2025 at 08:47:38PM +0200, Rafael J. Wysocki wrote:
> > > On Thu, Apr 10, 2025 at 6:03 PM Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx> wrote:
> > > >
> > > > On Thu, Apr 10, 2025 at 05:34:39PM +0200, Rafael J. Wysocki wrote:
> > > > > On Thu, Apr 10, 2025 at 4:45 AM Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx> wrote:
> > > > > >
> > > > > > From: Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx>
> > > > > >
> > > > > > When utilization is unchanged, a policy limits update is ignored unless
> > > > > > CPUFREQ_NEED_UPDATE_LIMITS is set. This occurs because limits_changed
> > > > > > depends on the old broken behavior of need_freq_update to trigger a call
> > > > > > into cpufreq_driver_resolve_freq() to evaluate the changed policy limits.
> > > > > >
> > > > > > After fixing need_freq_update, limit changes are ignored without
> > > > > > CPUFREQ_NEED_UPDATE_LIMITS, at least until utilization changes enough to
> > > > > > make map_util_freq() return something different.
> > > > > >
> > > > > > Fix the ignored limit changes by preserving the value of limits_changed
> > > > > > until get_next_freq() is called, so limits_changed can trigger a call to
> > > > > > cpufreq_driver_resolve_freq().
> > > > > >
> > > > > > Reported-and-tested-by: Stephan Gerhold <stephan.gerhold@xxxxxxxxxx>
> > > > > > Link: https://lore.kernel.org/lkml/Z_Tlc6Qs-tYpxWYb@xxxxxxxxxx
> > > > > > Fixes: 8e461a1cb43d6 ("cpufreq: schedutil: Fix superfluous updates caused by need_freq_update")
> > > > > > Signed-off-by: Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx>
> > > > > > ---
> > > > > >  kernel/sched/cpufreq_schedutil.c | 5 +++--
> > > > > >  1 file changed, 3 insertions(+), 2 deletions(-)
> > > > > >
> > > > > > 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)
> > > > > >                 return sg_policy->next_freq;
> > > > > >
> > > > > > +       sg_policy->limits_changed = false;
> > > > >
> > > > > AFAICS, after this code modification, a limit change may be missed due
> > > > > to a possible race with sugov_limits() which cannot happen if
> > > > > sg_policy->limits_changed is only cleared when it is set before
> > > > > updating sg_policy->need_freq_update.
> > > >
> > > > I don't think that's the case because sg_policy->limits_changed is cleared
> > > > before the new policy limits are evaluated in cpufreq_driver_resolve_freq().
> > >
> > > sugov_limits() may be triggered by a scaling_max_freq update, for
> > > instance, so it is asynchronous with respect to the usual governor
> > > flow.  It updates sg_policy->limits_changed and assumes that next time
> > > the governor runs, it will call into the driver, for example via
> > > cpufreq_driver_fast_switch(), so the new limits take effect.  This is
> > > not about cpufreq_driver_resolve_freq().
> >
> > OK, though I think there's still a race in cpufreq_driver_resolve_freq().
> >
> > > sugov_limits() runs after the driver's ->verify() callback has
> > > returned and it is conditional on that callback's return value, so the
> > > driver already knows the new limits when sugov_limits() runs, but it
> > > may still need to tell the hardware what the new limits are and that's
> > > why cpufreq_driver_fast_switch() may need to run.
> >
> > Which is why CPUFREQ_NEED_UPDATE_LIMITS exists, right.
> >
> > > Now, if sugov_should_update_freq() sees sg_policy->limits_changed set,
> > > it will set sg_policy->need_freq_update which (for drivers with
> > > CPUFREQ_NEED_UPDATE_LIMITS set) guarantees that the driver will be
> > > invoked and so sg_policy->limits_changed can be cleared.
> > >
> > > If a new instance of sugov_limits() runs at this point, there are two
> > > possibilities.  Either it completes before the
> > > sg_policy->limits_changed update in sugov_should_update_freq(), in
> > > which case the driver already knows the new limits as per the above
> > > and so the subsequent invocation of cpufreq_driver_fast_switch() will
> > > pick them up, or it sets sg_policy->limits_changed again and the
> > > governor will see it set next time it runs.  In both cases the new
> > > limits will be picked up unless they are changed again in the
> > > meantime.
> > >
> > > After the above change, sg_policy->limits_changed may be cleared even
> > > if it has not been set before and that's problematic.  Namely, say it
> > > is unset when sugov_should_update_freq() runs, after being called by
> > > sugov_update_single_freq() via sugov_update_single_common(), and
> > > returns 'true' without setting sg_policy->need_freq_update.  Next,
> > > sugov_update_single_common() returns 'true' and get_next_freq() is
> > > called.  It sees that freq != sg_policy->cached_raw_freq, so it clears
> > > sg_policy->limits_changed.  If sugov_limits() runs on a different CPU
> > > between the check and the sg_policy->limits_changed update in
> > > get_next_freq(), it may be missed and it is still not guaranteed that
> > > cpufreq_driver_fast_switch() will run because
> > > sg_policy->need_freq_update is unset and sugov_hold_freq() may return
> > > 'true'.
> > >
> > > For this to work, sg_policy->limits_changed needs to be cleared only
> > > when it is set and sg_policy->need_freq_update needs to be updated
> > > when sg_policy->limits_changed is cleared.
> >
> > Ah I see, thank you for the detailed explanation. So if I am understanding it
> > correctly: the problem with my patch is that CPUFREQ_NEED_UPDATE_LIMITS might
> > not be honored after a limits change, because CPUFREQ_NEED_UPDATE_LIMITS is only
> > honored when sg_policy->limits_changed is set. So there is the following race:
> >
> >            CPU-A                    CPU-B
> >   sugov_should_update_freq()                    // sg_policy->limits_changed == false, sg_policy->need_freq_update == false
> >                                 sugov_limits()  // sg_policy->limits_changed == true,  sg_policy->need_freq_update == false
> >   get_next_freq()                               // sg_policy->limits_changed == false, sg_policy->need_freq_update == false
> >
> >   // cpufreq driver won't be invoked for the limits change if:
> >   // next_f == sg_policy->next_freq || (sugov_hold_freq() == true && next_f < sg_policy->next_freq)
> >
> > Does that look right?
>
> Yes, it does.
>
> > > It looks like you really want to set sg_policy->need_freq_update to
> > > 'true' in sugov_should_update_freq() when sg_policy->limits_changed is
> > > set, but that would render CPUFREQ_NEED_UPDATE_LIMITS unnecessary.
> > >
> > > > Granted, if we wanted to be really certain of this, we'd need release semantics.
> > >
> > > I don't think so, but feel free to prove me wrong.
> >
> > Well, it appears that there really is synchronization missing between
> > cpufreq_set_policy() and schedutil, since cpufreq_set_policy() changes the live
> > policy->min and policy->max, and schedutil may observe either the old values in
> > there or garbage values due to load/store tearing.
>
> schedutil itself doesn't really read policy->min and policy->max.
> cpufreq_driver_resolve_freq() does this and drivers do, but since
> policy->min and policy->max are integers, no garbage values will be
> observed AFAICS.

Well, I've just realized that this is not exactly the case because of
the way in which __resolve_freq() is used in cpufreq_set_policy().
That needs to be fixed.