Re: [PATCH v2] sched/fair: Prefer waker CPU for reciprocal sync wakeups

[Shubhang]

Hi Vineeth,

On Fri, 24 Jul 2026, Madadi Vineeth Reddy wrote:

> On 23/07/26 04:20, Shubhang Kaushik (Ampere) wrote:
> > +			    READ_ONCE(p->last_wakee) == current &&
> > +			    prefer_sync_pair_cpu(p, cpu))
> > +				return cpu;
>
>
> The difference is SMT. POWER10/11 are SMT8: stacking there puts the pair on one thread while up to
> seven siblings on the same core sit idle, sharing the same LLC.
>
> So instead of returning the waker's CPU, I'm looking at letting the waker's *core* count as idle
> when the waker's runqueue has a single runnable task, so the wakee lands on a sibling thread.

Thanks for the review, the distinction makes sense.

In the case of non-SMT, the benefit I am seeing comes from stopping 
SIS from moving the reciprocal WF_SYNC wakee away from the wake-affine 
target.

For v3, my plan is to keep this scoped to that non-SMT case under 
the existing SD_WAKE_AFFINE path and fold in Prateek's suggested 
cleanups. Would still wait for Christian/Prateek's SMT numbers to make 
sure this does not need additional SMT handling in this series.

Regards,
Shubhang Kaushik