Re: [PATCH v3] sched/fair: bring back select_idle_smt, but differently

[Mel Gorman <mgorman@techsingularity.net>]

On Wed, Apr 07, 2021 at 12:15:13PM +0200, Peter Zijlstra wrote:
> On Wed, Apr 07, 2021 at 10:41:06AM +0100, Mel Gorman wrote:
> 
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -6112,6 +6112,27 @@ static int select_idle_core(struct task_
> > >  	return -1;
> > >  }
> > >  
> > > +/*
> > > + * Scan the local SMT mask for idle CPUs.
> > > + */
> > > +static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
> > > +{
> > > +	int cpu;
> > > +
> > > +	if (!static_branch_likely(&sched_smt_present))
> > > +		return -1;
> > > +
> > > +	for_each_cpu(cpu, cpu_smt_mask(target)) {
> > > +		if (!cpumask_test_cpu(cpu, p->cpus_ptr) ||
> > > +		    !cpumask_test_cpu(cpu, sched_domain_span(sd)))
> > > +			continue;
> > 
> > While I know that !cpumask_test_cpu(cpu, sched_domain_span(sd)) was
> > done previously, I found it hard to believe that the test matters. If
> > target/prev share a the LLC domain, why would the SMT siblings *not*
> > share a LLC?
> 
> I think the reason for it is that a cpuset might have split the siblings
> apart and disabled load-balancing across them or something.
> 
> Then the affinity mask can still cross the partition, but we shouldn't
> ever move into it through balancing.

Ok, cpusets do split domains. I can't imagine the logic of splitting SMT
siblings across cpusets but if it's possible, it has to be checked and
protecting that with cpusets_enabled() would be a little overkill and
possibly miss some other corner case :(

Thanks.

-- 
Mel Gorman
SUSE Labs