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

[Shubhang]

Hi Prateek,

On Thu, 23 Jul 2026, K Prateek Nayak wrote:
> > +	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
> > +		return false;
>
> want_affine already covers the affinity test now that you've moved it
> into the SD_WAKE_AFFINE branch.

Agreed, I'll drop this from the helper.

> > +	if (sched_asym_cpucap_active()) {
> > +		sync_entity_load_avg(&p->se);
> > +		if (!task_fits_cpu(p, cpu))
> > +			return false;
> > +	}
> > +
> > +	return true;
>
> Reads a lot like WA_IDLE. Couldn't this simply be an early returns
> in select_idle_sibling() if we passed wake_flags there?

This does show up as select_idle_sibling() overriding the wake-affine
target with an idle CPU.

But if we move the check there, wake_flags alone is not enough to 
preserve the SD_WAKE_AFFINE/isolcpus fix from v2. SIS would also 
need to know that the existing SD_WAKE_AFFINE check was reached.

Also, putting the early return only under sched_smt_active() &&
!has_idle_core handles the SMT case, but would drop the non-SMT case 
where the Ampere Altra improvement comes from.

So I can either keep the smaller check in the SD_WAKE_AFFINE block with
the cleanups, or move it into SIS with an explicit affine_sd argument 
and separate non-SMT/SMT handling.

> >  		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
> >  		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
> > +			if (sync &&
> > +			    READ_ONCE(p->last_wakee) == current &&
>
> No need for READ_ONCE(), Last wakee is recorded when "p" has last
> woken up a task and given we are now doing a wakeup for "p", it
> is completely off the rq and p->last_wakee is stable.

Makes sense, will drop that. On the non-SMT Ampere Altra system I tested 
on, keeping the smaller SD_WAKE_AFFINE version with the cleanups gives 
the beter result in comparison.

Regards,
Shubhang Kaushik