Re: [PATCH] sched/fair: Skip sched-domain traversal when wake affine is disabled

[Kayra Cizmeci <kayracizmeci@gmail.com>]

On Mon, Nov 09, 2026 at 10:01:20PM +0530, Srikar Dronamraju wrote:
> * Kayra Cizmeci <kayracizmeci@gmail.com> [2026-11-09 18:09:56]:
> 
> Hi Kayra,
> 
> > 
> > On Mon, Nov 09, 2026 at 05:35:22PM +0530, Srikar Dronamraju wrote:
> > > In select_task_rq_fair(), TTWU wakeups for which wake_wide() returns
> > > true, or where the waking CPU is not in the task's allowed CPU mask,
> > > want_affine is set to false.
> > > 
> > > For these wakeups, the scheduler does not perform wake-affine placement.
> > > The existing code already treats this as a fast-path by exiting the
> > > sched-domain walk before locating an SD_BALANCE_WAKE domain, ultimately
> > > falling through to select_idle_sibling().
> > 
> > Not before locating. If it doesn't locates SD_BALANCE_WAKE in the first
> > check it just leaves.
> > 
> > AFAICT SD_BALANCE_WAKE is not set anywhere tho.
> 
> from kernel/sched/sched.h
> #define WF_TTWU                 0x08 /* Wakeup;            maps to SD_BALANCE_WAKE *
> The code searches for WF_TTWU in lieu of SD_BALANCE_WAKE.

Yeah I know. But I couldn't find a place where SD_BALANCE_WAKE gets set
in the domain.

> 
> > 
> > So if you wanna do this, maybe you can do something like:
> > 
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index d4cc0caee626..b40b2eb5e577 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -10016,6 +10016,9 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
> >  		}
> >  
> >  		want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
> > +
> > +		if (!want_affine)
> > +			return select_idle_sibling(p, prev_cpu, new_cpu);
> >  	}
> 
> Yes, this also has the same effect.
> 
> >  
> >  	for_each_domain(cpu, tmp) {
> > 
> > > 
> > > However, select_task_rq_fair() still enters the sched-domain traversal
> > > only to break out almost immediately. Avoid this unnecessary work by
> > > jumping directly to the existing fast path when want_affine is false.
> > > 
> > > As part of this cleanup:
> > > 
> > >   - Move want_affine into the WF_TTWU block where it is used.
> > >   - Skip the sched-domain traversal entirely when want_affine is false.
> > >   - Remove the now redundant want_affine test from the
> > >     SD_WAKE_AFFINE check.
> > >   - Remove the corresponding !want_affine loop termination logic.
> > > 
> > > This preserves existing wakeup placement behavior while avoiding
> > > unnecessary sched-domain iteration for TTWU wakeups that cannot make
> > > use of wake-affine placement.
> > > 
> > > Signed-off-by: Srikar Dronamraju <srikar@linux.ibm.com>
> > > ---
> > >  kernel/sched/fair.c | 9 +++++----
> > >  1 file changed, 5 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index 3ebec186f982..58ff55e093cb 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -8832,7 +8832,6 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
> > >  	struct sched_domain *tmp, *sd = NULL;
> > >  	int cpu = smp_processor_id();
> > >  	int new_cpu = prev_cpu;
> > > -	int want_affine = 0;
> > >  	/* SD_flags and WF_flags share the first nibble */
> > >  	int sd_flag = wake_flags & 0xF;
> > >  
> > > @@ -8841,6 +8840,8 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
> > >  	 */
> > >  	lockdep_assert_held(&p->pi_lock);
> > >  	if (wake_flags & WF_TTWU) {
> > > +		int want_affine;
> > > +
> > >  		record_wakee(p);
> > >  
> > >  		if ((wake_flags & WF_CURRENT_CPU) &&
> > > @@ -8855,6 +8856,8 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
> > >  		}
> > >  
> > >  		want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
> > > +		if (!want_affine)
> > > +			goto fast_path;
> > >  	}
> > >  
> > >  	for_each_domain(cpu, tmp) {
> > > @@ -8862,7 +8865,7 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
> > >  		 * If both 'cpu' and 'prev_cpu' are part of this domain,
> > >  		 * cpu is a valid SD_WAKE_AFFINE target.
> > >  		 */
> > > -		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
> > > +		if ((tmp->flags & SD_WAKE_AFFINE) &&
> > >  		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
> > >  			if (cpu != prev_cpu)
> > >  				new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
> > > @@ -8878,8 +8881,6 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
> > >  		 */
> > >  		if (tmp->flags & sd_flag)
> > >  			sd = tmp;
> > > -		else if (!want_affine)
> > > -			break;
> > >  	}
> > 
> > 
> > This just changes the behavior.
> > 
> > Old Code WF_FORK 1, WF_TTWU 0:
> > 
> > want_affine = 0, so we skip the first if block in the loop then leave
> > once we find a domain without our flag. And go with slow path.
> > 
> > 
> > New Code:
> > 
> > want_affine is moved away and early exit is removed. So we enter the if
> > block and try to find a domain with prev_cpu. Because WF_TTWU 
> > is 0, we just return new_cpu. Or if we can't fimd the domain with
> > prev_cpu, slow path.
> > 
> 
> When WF_TTWU is set, we are always suppose take the fast path.
> slowpath is for WF_FORK and WF_EXEC.

I know. And that is the exact problem with the above diff :-).

Like let's say prev_cpu is in the first domain we are looking, the
domain has SD_WAKE_AFFINE and SD_BALANCE_FORK set and next domain has
only SD_WAKE_AFFINE set:

Old Code:

sd = first domain and we enter the slow path with the first domain.

With your patch:

sd = NULL and we return new_cpu


Or am I missing something?

Cheers,

Kayra