Re: [PATCH RESEND] sched/fair: Only update stats for allowed CPUs when looking for dst group

[Chen Yu]

On 10/14/25 19:37, Peter Zijlstra wrote:
> On Sat, Oct 11, 2025 at 06:43:22AM +0000, Adam Li wrote:
> > Load imbalance is observed when the workload frequently forks new threads.
> > Due to CPU affinity, the workload can run on CPU 0-7 in the first
> > group, and only on CPU 8-11 in the second group. CPU 12-15 are always idle.
> >
> > { 0 1 2 3 4 5 6 7 } {8 9 10 11 12 13 14 15}
> >    * * * * * * * *    * * *  *
> >
> > When looking for dst group for newly forked threads, in many times
> > update_sg_wakeup_stats() reports the second group has more idle CPUs
> > than the first group. The scheduler thinks the second group is less
> > busy. Then it selects least busy CPUs among CPU 8-11. Therefore CPU 8-11
> > can be crowded with newly forked threads, at the same time CPU 0-7
> > can be idle.
> >
> > A task may not use all the CPUs in a schedule group due to CPU affinity.
> > Only update schedule group statistics for allowed CPUs.
> >
> > Signed-off-by: Adam Li <adamli@xxxxxxxxxxxxxxxxxxxxxx>
> > ---
> > Resending this patch from the patchset:
> > https://lore.kernel.org/lkml/20250717062036.432243-2-adamli@xxxxxxxxxxxxxxxxxxxxxx/
>
> Right, lets start with this then ;-)
>
> No need to do the cpumask_and() thing, that's just more changes vs
> update_sg_lb_stats().


I just saw Peter's comments on this.  I'm OK with the current
version and Adam please feel free to keep my Reviewed-by tag
for this non-cpumask_and version.

thanks,
Chenyu