2024-04-04 10:42:23

by Vincent Guittot

[permalink] [raw]
Subject: [PATCH] PM:EM: fix wrong utilization estimation in em_cpu_energy()

Commit 1b600da51073 ("PM: EM: Optimize em_cpu_energy() and remove division")
has added back map_util_perf() in em_cpu_energy() computation which has
been removed with the rework of scheduler/cpufreq interface.
This is wrong because sugov_effective_cpu_perf() already takes care of
mapping the utilization to a performance level.

Fixes: 1b600da51073 ("PM: EM: Optimize em_cpu_energy() and remove division")
Signed-off-by: Vincent Guittot <[email protected]>
---
include/linux/energy_model.h | 1 -
1 file changed, 1 deletion(-)

diff --git a/include/linux/energy_model.h b/include/linux/energy_model.h
index 770755df852f..70cd7258cd29 100644
--- a/include/linux/energy_model.h
+++ b/include/linux/energy_model.h
@@ -245,7 +245,6 @@ static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
* max utilization to the allowed CPU capacity before calculating
* effective performance.
*/
- max_util = map_util_perf(max_util);
max_util = min(max_util, allowed_cpu_cap);

/*
--
2.34.1



2024-04-04 11:05:52

by Lukasz Luba

[permalink] [raw]
Subject: Re: [PATCH] PM:EM: fix wrong utilization estimation in em_cpu_energy()

Hi Vincent,

On 4/4/24 11:42, Vincent Guittot wrote:
> Commit 1b600da51073 ("PM: EM: Optimize em_cpu_energy() and remove division")
> has added back map_util_perf() in em_cpu_energy() computation which has
> been removed with the rework of scheduler/cpufreq interface.
> This is wrong because sugov_effective_cpu_perf() already takes care of
> mapping the utilization to a performance level.
>
> Fixes: 1b600da51073 ("PM: EM: Optimize em_cpu_energy() and remove division")
> Signed-off-by: Vincent Guittot <[email protected]>
> ---
> include/linux/energy_model.h | 1 -
> 1 file changed, 1 deletion(-)
>
> diff --git a/include/linux/energy_model.h b/include/linux/energy_model.h
> index 770755df852f..70cd7258cd29 100644
> --- a/include/linux/energy_model.h
> +++ b/include/linux/energy_model.h
> @@ -245,7 +245,6 @@ static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
> * max utilization to the allowed CPU capacity before calculating
> * effective performance.
> */
> - max_util = map_util_perf(max_util);
> max_util = min(max_util, allowed_cpu_cap);
>
> /*

LGTM. It was developed in parallel IIRC and that change which removes
the extra margin to the util was lost from my radar. I can see it
landed first.

Reviewed-by: Lukasz Luba <[email protected]>

Regards,
Lukasz

2024-04-04 19:30:42

by Rafael J. Wysocki

[permalink] [raw]
Subject: Re: [PATCH] PM:EM: fix wrong utilization estimation in em_cpu_energy()

On Thu, Apr 4, 2024 at 1:05 PM Lukasz Luba <[email protected]> wrote:
>
> Hi Vincent,
>
> On 4/4/24 11:42, Vincent Guittot wrote:
> > Commit 1b600da51073 ("PM: EM: Optimize em_cpu_energy() and remove division")
> > has added back map_util_perf() in em_cpu_energy() computation which has
> > been removed with the rework of scheduler/cpufreq interface.
> > This is wrong because sugov_effective_cpu_perf() already takes care of
> > mapping the utilization to a performance level.
> >
> > Fixes: 1b600da51073 ("PM: EM: Optimize em_cpu_energy() and remove division")
> > Signed-off-by: Vincent Guittot <[email protected]>
> > ---
> > include/linux/energy_model.h | 1 -
> > 1 file changed, 1 deletion(-)
> >
> > diff --git a/include/linux/energy_model.h b/include/linux/energy_model.h
> > index 770755df852f..70cd7258cd29 100644
> > --- a/include/linux/energy_model.h
> > +++ b/include/linux/energy_model.h
> > @@ -245,7 +245,6 @@ static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
> > * max utilization to the allowed CPU capacity before calculating
> > * effective performance.
> > */
> > - max_util = map_util_perf(max_util);
> > max_util = min(max_util, allowed_cpu_cap);
> >
> > /*
>
> LGTM. It was developed in parallel IIRC and that change which removes
> the extra margin to the util was lost from my radar. I can see it
> landed first.
>
> Reviewed-by: Lukasz Luba <[email protected]>

Applied as 6.9-rc material, thanks!