[PATCH] sched/pelt: Fix truncation of the PELT period count

[PATCH] sched/pelt: Fix truncation of the PELT period count

[Hui Su]

accumulate_sum() computes the number of elapsed PELT periods in a
u64 but returns u32. When the count is a multiple of 2^32, it is
truncated to zero even though the PELT sums have been updated.

The caller interprets zero as no period boundary crossed and skips
recomputing the averages, leaving *_avg stale relative to *_sum.

Return only whether any period elapsed, which is all the caller
needs.

Signed-off-by: Hui Su <sh_def@163.com>
---
 kernel/sched/pelt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/kernel/sched/pelt.c b/kernel/sched/pelt.c
index 779eb58a4261..519bc9c40043 100644
--- a/kernel/sched/pelt.c
+++ b/kernel/sched/pelt.c
@@ -146,7 +146,7 @@ accumulate_sum(u64 delta, struct sched_avg *sa,
 	if (running)
 		sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
 
-	return periods;
+	return !!periods;
 }
 
 /*
-- 
2.55.0

Re: [PATCH] sched/pelt: Fix truncation of the PELT period count

[K Prateek Nayak]

Hello Hui,

On 10/22/2026 9:19 AM, Hui Su wrote:
> accumulate_sum() computes the number of elapsed PELT periods in a
> u64 but returns u32. When the count is a multiple of 2^32, it is

So at least an hour between PELT updates! I suppose it is possible
with nohz so not too far fetched.

> truncated to zero even though the PELT sums have been updated.
> 
> The caller interprets zero as no period boundary crossed and skips
> recomputing the averages, leaving *_avg stale relative to *_sum.
> 
> Return only whether any period elapsed, which is all the caller
> needs.
> 

Probably needs a:

Fixes: a481db34b9be ("sched/fair: Optimize ___update_sched_avg()") # c079629862b2

> Signed-off-by: Hui Su <sh_def@163.com>
> ---
>  kernel/sched/pelt.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/kernel/sched/pelt.c b/kernel/sched/pelt.c
> index 779eb58a4261..519bc9c40043 100644
> --- a/kernel/sched/pelt.c
> +++ b/kernel/sched/pelt.c
> @@ -146,7 +146,7 @@ accumulate_sum(u64 delta, struct sched_avg *sa,
>  	if (running)
>  		sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
>  
> -	return periods;
> +	return !!periods;

I would actually favor changing return type to u64. Since
accumulate_sum() is only called in one place, and that too within an
'if' condition, I'm sure compiler can optimize it similarly.

>  }
>  
>  /*

-- 
Thanks and Regards,
Prateek

Re: [PATCH] sched/pelt: Fix truncation of the PELT period count

[Vincent Guittot]

On Thu, 22 Oct 2026 at 10:35, K Prateek Nayak <kprateek.nayak@amd.com> wrote:
>
> Hello Hui,
>
> On 10/22/2026 9:19 AM, Hui Su wrote:
> > accumulate_sum() computes the number of elapsed PELT periods in a
> > u64 but returns u32. When the count is a multiple of 2^32, it is
>
> So at least an hour between PELT updates! I suppose it is possible
> with nohz so not too far fetched.

Period is 1ms so it's more in the range of 50 days

>
> > truncated to zero even though the PELT sums have been updated.
> >
> > The caller interprets zero as no period boundary crossed and skips
> > recomputing the averages, leaving *_avg stale relative to *_sum.
> >
> > Return only whether any period elapsed, which is all the caller
> > needs.
> >
>
> Probably needs a:
>
> Fixes: a481db34b9be ("sched/fair: Optimize ___update_sched_avg()") # c079629862b2
>
> > Signed-off-by: Hui Su <sh_def@163.com>
> > ---
> >  kernel/sched/pelt.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/kernel/sched/pelt.c b/kernel/sched/pelt.c
> > index 779eb58a4261..519bc9c40043 100644
> > --- a/kernel/sched/pelt.c
> > +++ b/kernel/sched/pelt.c
> > @@ -146,7 +146,7 @@ accumulate_sum(u64 delta, struct sched_avg *sa,
> >       if (running)
> >               sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
> >
> > -     return periods;
> > +     return !!periods;
>
> I would actually favor changing return type to u64. Since
> accumulate_sum() is only called in one place, and that too within an
> 'if' condition, I'm sure compiler can optimize it similarly.
>
> >  }
> >
> >  /*
>
> --
> Thanks and Regards,
> Prateek
>

Re: [PATCH] sched/pelt: Fix truncation of the PELT period count

[Hui Su]

> Probably needs a:
>
> Fixes: a481db34b9be ("sched/fair: Optimize ___update_sched_avg()") # c079629862b2
>
> I would actually favor changing return type to u64. Since
> accumulate_sum() is only called in one place, and that too within an
> 'if' condition, I'm sure compiler can optimize it similarly.

Thanks for the review.

I will use a481db34b9be in v2's Fixes tag and change accumulate_sum() to return u64.

Thanks,
Hui

Re: [PATCH] sched/pelt: Fix truncation of the PELT period count

[Vincent Guittot]

On Thu, 22 Oct 2026 at 05:50, Hui Su <sh_def@163.com> wrote:
>
> accumulate_sum() computes the number of elapsed PELT periods in a
> u64 but returns u32. When the count is a multiple of 2^32, it is
> truncated to zero even though the PELT sums have been updated.

Did you face this problem, or did you find it only during a code review?

This would happen every multiple of exactly 2^32 ms (~50days) without
updating pelt whereas the latter is supposed to be updated far more
often (within second) until it reaches 0. Then such a large period
would keep it at 0 anyway

>
> The caller interprets zero as no period boundary crossed and skips
> recomputing the averages, leaving *_avg stale relative to *_sum.
>
> Return only whether any period elapsed, which is all the caller
> needs.
>
> Signed-off-by: Hui Su <sh_def@163.com>

Even if I doubt it can happen, it's better to fix it anyway

Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>

> ---
>  kernel/sched/pelt.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/sched/pelt.c b/kernel/sched/pelt.c
> index 779eb58a4261..519bc9c40043 100644
> --- a/kernel/sched/pelt.c
> +++ b/kernel/sched/pelt.c
> @@ -146,7 +146,7 @@ accumulate_sum(u64 delta, struct sched_avg *sa,
>         if (running)
>                 sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
>
> -       return periods;
> +       return !!periods;
>  }
>
>  /*
> --
> 2.55.0
>

Re: [PATCH] sched/pelt: Fix truncation of the PELT period count

[Hui Su]

On Thu, Oct 22, 2026, Vincent Guittot wrote:

> Did you face this problem, or did you find it only during a code review?

I found it during code review, not from a production failure.

I reproduced it with a KUnit test that advances PELT by exactly 2^32 periods.
Before the fix, util_avg remained 43 instead of 1024; after the fix, the test
passed. A standalone harness confirmed the same result.

Thanks for the review.

Thanks,
Hui