Re: [PATCH] sched/cache: Ignore non-fair current in alb_break_llc()

[Tim Chen]

On Tue, 2026-10-20 at 11:14 -0700, Tim Chen wrote:
> On Tue, 2026-10-20 at 19:17 +0300, Kayra Cizmeci wrote:
> > Commit d6b9afab44e23 ("sched/cache: Fix
> > checking active load balance by only considering the CFS task")
> > 
> > Stopped using task_util() for non-fair tasks. Since the values
> > of non-fair tasks can be stale.
> > 
> > __active_load_balance_cpu_stop() calls detach_one_task(), that only looks
> > to cfs_tasks list. So if the rq->curr is not fair, doing a decision
> > according to it is wrong, since we only migrate tasks that is fair.
> > 
> > Commit f0d243a96f26 ("sched/fair: Avoid creating misfits during
> > cache-aware balancing") uses cur without checking if the cur is in fair
> > class or not. Fix this.

I suggest rephrasing the log a bit:

  Commit d6b9afab44e23 ("sched/cache: Fix checking active load balance
  by only considering the CFS task") stopped using task_util() of a
  non-fair curr in alb_break_llc(), as the util of non-fair tasks can
  be stale.

  However, commit f0d243a96f26 ("sched/fair: Avoid creating misfits
  during cache-aware balancing") still passes curr to
  task_misfits_asym_cpu() regardless of its class, which then calls
  task_fits_cpu() on it. A non-fair curr is never migrated by active
  balance, since detach_one_task() only picks tasks from cfs_tasks.
  Ignore a non-fair curr so it does not affect the decision.

Tim


> > 
> > Fixes: f0d243a96f26 ("sched/fair: Avoid creating misfits during cache-aware balancing")
> > Signed-off-by: Kayra Cizmeci <kayracizmeci@xxxxxxxxx>
> > ---
> >  kernel/sched/fair.c | 4 +++-
> >  1 file changed, 3 insertions(+), 1 deletion(-)
> > 
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 0feec172fd37..6347ae4cb419 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -11016,7 +11016,9 @@ alb_break_llc(struct lb_env *env)
> >  			return true;
> >  
> >  		cur = rcu_dereference_all(env->src_rq->curr);
> > -		if (cur && cur->sched_class == &fair_sched_class)
> > +		if (cur && cur->sched_class != &fair_sched_class)
> 
> Thanks. util was already 0 for a non-fair curr, but cur was still
> passed to task_misfits_asym_cpu(), which would evaluate
> task_fits_cpu() on a non-CFS task. Since nr_pref_llc_running and
> h_nr_runnable only count CFS tasks, curr can be e.g. an RT task here
> with a CFS task queued behind it, and active balance would only ever
> detach a CFS task anyway. So the fix makes sense.
> 
> Suggest that you put a comment here:
> 
>                   /*
>                    * Active balance only detaches CFS tasks, so a non-fair
>                    * curr is never migrated. Ignore it and avoid using its
>                    * stale util in the misfit and LLC checks.
>                    */
>                   if (cur && cur->sched_class != &fair_sched_class)
>                           cur = NULL;
>                   if (cur)
>                           util = task_util(cur);
> 
> Could you update the changelog to say explicitly that the issue is the
> non-fair curr being fed to task_misfits_asym_cpu(), not util?
> 
> Tim
> 
> > +			cur = NULL;
> > +		if (cur)
> >  			util = task_util(cur);
> >  
> >  		if (task_misfits_asym_cpu(env, cur) ||