Re: [PATCH] cpufreq: conservative: Ignore idle periods when a policy CPU is busy

[hu.shengming]

Zhongqiu wrote:

> On 9/7/2026 6:55 PM, hu.shengming@xxxxxxxxxx wrote:
> > Zhongqiu wrote:
> >> Hi Shengming,
> >> Thanks for the patch.
> > 
> > Hi Zhongqiu,
> > Thanks for the review!
> > 
> >> On 9/2/2026 3:47 PM, hu.shengming@xxxxxxxxxx wrote:
> >>> From: Shengming Hu <hu.shengming@xxxxxxxxxx>
> >>>
> >>> For a shared cpufreq policy, dbs_update() derives the load from the
> >>> highest utilization among its CPUs, but it also records deferred idle
> >>> periods from any CPU whose idle time exceeds two sampling intervals.
> >>>
> >>> This lets a single update report both a high load (from a busy CPU)
> >>> and several deferred idle periods (from an idle sibling). Since
> >>> conservative applies the deferred down steps before the up step
> >>> triggered by the high load, the down steps can outweigh the single
> >>> up step.
> >>>
> >>> The issue reproduces on a policy shared by CPUs 2 and 3: a CPU-bound
> >>> SCHED_EXT task keeps CPU 2 at 100% utilization while CPU 3 stays
> >>> idle. On this system SCHED_EXT generates update-util callbacks less
> >>> frequently than CFS, so DBS updates are sparse, tracing shows:
> >>>
> >>>    load=100 idle_periods=7 interval=59 ms
> >>>    load=100 idle_periods=4 interval=39 ms
> >>>    load=100 idle_periods=2 interval=19 ms
> >>>    load=100 idle_periods=7 interval=59 ms
> >>>
> >>> With the default 5% step and a 2.6 GHz ceiling, conservative first
> >>> removes seven 130 MHz steps and then adds only one. Repeating this
> >>> sequence keeps the policy near 530 MHz despite CPU 2 being fully busy.
> >>>
> >>> Only retain deferred idle periods when every CPU in the policy meets
> >>> the long-idle condition. This keeps the existing behavior for
> >>> single-CPU and fully idle shared policies, while preventing an idle
> >>> sibling from downscaling a policy that contains a busy CPU.
> >>>
> >>> Cc: stable@xxxxxxxxxxxxxxx
> >>> Fixes: 00bfe05889e9 ("cpufreq: conservative: Decrease frequency faster for deferred updates")
> >>> Reviewed-by: Luo Haiyang <luo.haiyang@xxxxxxxxxx>
> >>> Reviewed-by: Run Zhang <zhang.run@xxxxxxxxxx>
> >>> Signed-off-by: Shengming Hu <hu.shengming@xxxxxxxxxx>
> >>> ---
> >>>    drivers/cpufreq/cpufreq_governor.c | 5 ++++-
> >>>    1 file changed, 4 insertions(+), 1 deletion(-)
> >>>
> >>> diff --git a/drivers/cpufreq/cpufreq_governor.c b/drivers/cpufreq/cpufreq_governor.c
> >>> index 710d93ec89b5..64eb6b5f08a4 100644
> >>> --- a/drivers/cpufreq/cpufreq_governor.c
> >>> +++ b/drivers/cpufreq/cpufreq_governor.c
> >>> @@ -126,6 +126,7 @@ unsigned int dbs_update(struct cpufreq_policy *policy)
> >>>        unsigned int ignore_nice = dbs_data->ignore_nice_load;
> >>>        unsigned int max_load = 0, idle_periods = UINT_MAX;
> >>>        unsigned int sampling_rate, io_busy, j;
> >>> +    bool all_cpus_idle = true;
> >>>        u64 cur_nice;
> >>>
> >>>        /*
> >>> @@ -233,13 +234,15 @@ unsigned int dbs_update(struct cpufreq_policy *policy)
> >>>
> >>>                if (periods < idle_periods)
> >>>                    idle_periods = periods;
> >>> +        } else {
> >>> +            all_cpus_idle = false;
> >>
> >> The problem is real, but I don't think this condition is the right one.
> >> idle_time > 2 * sampling_rate tells us how many sampling periods were
> >> deferred for that CPU, so its negation means "this CPU was sampled on
> >> time", not "this CPU is busy".
> >>
> >> Since all_cpus_idle is per-policy, one such CPU is enough to discard the
> >> deferred periods for the whole policy, and in a shared policy it is
> >> possible. That effectively disables the optimization from 00bfe05889e9
> >> for shared policies, which is the opposite of what we want for power.
> > 
> > Agreed that not meeting the long-idle condition does not necessarily
> > mean that the CPU was busy. The condition is based on accumulated idle
> > time, so it is not a reliable indication of whether that CPU should
> > prevent deferred downscaling.
> > 
> >> What matters is whether the CPU was busy over the sample, that is,
> >> whether the skipped sampling periods would have led to a frequency
> >> reduction at all. It seems more appropriate to key that off the load
> >> measured over the sample (kept separate from the possibly inherited one)
> >> against up_threshold, so an idle-but-punctually-sampled sibling does not
> > 
> > Thanks for the suggestion. I agree that the load actually measured over
> > the current sample should be kept separate from the load that may inherit
> > prev_load. However, I don't think up_threshold is the appropriate
> > boundary for deciding whether deferred down steps should be applied.
> > 
> > For example, suppose CPU A has been idle for several sampling periods
> > while CPU B has a sustained load of 75%, with up_threshold at 80 and
> > down_threshold at 20. The policy is then in conservative's hold region,
> > so the load itself would trigger neither an increase nor a decrease.
> > If deferred downscaling is gated only by up_threshold, CPU B would
> > not block it, so CPU A's deferred idle periods could still reduce
> > the policy frequency.
> 
> It seems not, in func cs_dbs_update(), idle_periods only affects the
> local variable requested_freq, and that variable is never actually
> applied to change the CPU frequency while the policy remains in the hold
> region.

You are right if both the measured load and the load returned by
dbs_update() remain at 75%. In that case neither frequency branch is
entered, so the idle-adjusted local requested_freq is not passed to
__cpufreq_driver_target().

My example was incomplete. I was referring to a case where the measured
policy load is 75%, but the load returned by dbs_update() is above
up_threshold because another CPU reuses a high prev_load.

For example, consider a policy shared by CPUs A and B, with
up_threshold=80 and down_threshold=20:

CPU A:
        current sample load = 0
        prev_load           = 100
        idle_periods        = 7

CPU B:
        current sample load = 75
        no prev_load reuse

For CPU A, the long-idle branch uses:

load = j_cdbs->prev_load;
j_cdbs->prev_load = 0;

If the load calculated from the current counters is retained separately
before that substitution, the policy-level values are:

max_sample_load = max(0, 75)   = 75
returned load   = max(100, 75) = 100
idle_periods                   = 7

With an up_threshold gate, max_sample_load=75 does not suppress the
deferred adjustment. cs_dbs_update() therefore first subtracts seven
steps and then, because the returned load is 100, enters the increase
branch, adds one step, and passes the result to
__cpufreq_driver_target().

With a down_threshold gate, max_sample_load=75 is not in the decrease
region, so the seven deferred steps are skipped. The returned load of
100 then executes only the normal increase step, subject to policy->max,
so it cannot lower the previous request.

> > I think deferred down steps should instead be applied only when the
> > maximum load actually measured across the policy is below
> > down_threshold. To keep this independent of the load returned by
> 
> The two gates only differ when the measured load lands between
> down_threshold and up_threshold and the load used for the decision (the
> inherited prev_load in that case) triggers one of the branches - if no
> CPU took the reuse path the two values are equal and the outcome is the
> same.

Agreed that the gates can differ when the measured and returned loads diverge.
The example above illustrates the high-prev_load reuse case I am concerned about.

> If we use up_threshold --> the deferred downscale is only given up when
> the CPU is genuinely busy enough to warrant a frequency increase; in all
> other cases it still scales down as much as possible. This stays closer
> to the design of 00bfe05889e9 ("cpufreq: conservative: Decrease
> frequency faster for deferred updates"). When the measured load is above
> up_threshold, we skip the deferred downscaling; when it falls between
> down_threshold and up_threshold and the load used for the decision is in
> that band as well, the frequency is left unchanged either way. This
> fixes the bug you described while avoiding any significant power
> regression.

In the scenario above, the difference between the two choices is how the
deferred down steps are handled when the measured policy load is in the
hold region but the load returned by dbs_update() exceeds up_threshold
due to prev_load reuse. With an up_threshold gate, the deferred down
steps are still applied before the increase step, which may result in
a net frequency reduction. With a down_threshold gate, those deferred
down steps are skipped because the measured load is not in the governor's
downscaling region.

Although an up_threshold gate preserves more of the existing behavior,
00bfe05889e9 addressed the case where the workload had finished and the
CPU was idle. In that case max_sample_load is below down_threshold, so
both gates preserve the original deferred-downscale optimization.

> If we use down_threshold --> the deferred downscale is skipped whenever
> the load is not in the lowest (downscale) region. it can cause power
> regression.

I agree that using down_threshold can result in a higher requested frequency
in this case, so I cannot rule out a power regression without measurements.
My concern is whether deferred idle periods from one CPU should reduce the
shared policy frequency when another CPU's measured load is in the hold region.
Do you consider that reduction intentional?

> > dbs_update(), which may inherit prev_load, we could record the maximum
> > measured load separately in struct policy_dbs_info, for example as
> > max_sample_load.
> > 
> > The conservative governor could then gate the deferred reductions with
> > something like:
> > 
> >      if (policy_dbs->max_sample_load < cs_tuners->down_threshold &&
> >          policy_dbs->idle_periods < UINT_MAX) {
> >          ...
> >      }
> > 
> > This preserves deferred downscaling when the measured policy load is
> > below down_threshold, while avoiding deferred reductions when any CPU
> > is in either the hold or upscale region.
> > 
> >> May I know could you comment and try this patch on your scenario?  Once
> >> everyone agrees I can send this formally:
> > 
> > I'll rework the patch along these lines, keeping the measured load
> > separate from the inherited load and using down_threshold for the
> > deferred-downscale condition.
> > 
> > I'll send a v2, with a Suggested-by tag for your suggestion.

--
With Best Regards,
Shengming