Re: [Patch v3 2/6] sched/topology: Record number of cores in sched group

[Tim Chen]

On Mon, 2023-07-10 at 21:33 +0100, Valentin Schneider wrote:
> On 07/07/23 15:57, Tim Chen wrote:
> > From: Tim C Chen <tim.c.chen@xxxxxxxxxxxxxxx>
> > 
> > When balancing sibling domains that have different number of cores,
> > tasks in respective sibling domain should be proportional to the number
> > of cores in each domain. In preparation of implementing such a policy,
> > record the number of tasks in a scheduling group.
> > 
> > Signed-off-by: Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx>
> > ---
> >  kernel/sched/sched.h    |  1 +
> >  kernel/sched/topology.c | 10 +++++++++-
> >  2 files changed, 10 insertions(+), 1 deletion(-)
> > 
> > diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
> > index 3d0eb36350d2..5f7f36e45b87 100644
> > --- a/kernel/sched/sched.h
> > +++ b/kernel/sched/sched.h
> > @@ -1860,6 +1860,7 @@ struct sched_group {
> >       atomic_t		ref;
> > 
> >       unsigned int		group_weight;
> > +	unsigned int		cores;
> >       struct sched_group_capacity *sgc;
> >       int			asym_prefer_cpu;	/* CPU of highest priority in group */
> >       int			flags;
> > diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
> > index 6d5628fcebcf..6b099dbdfb39 100644
> > --- a/kernel/sched/topology.c
> > +++ b/kernel/sched/topology.c
> > @@ -1275,14 +1275,22 @@ build_sched_groups(struct sched_domain *sd, int cpu)
> >  static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
> >  {
> >       struct sched_group *sg = sd->groups;
> > +	struct cpumask *mask = sched_domains_tmpmask2;
> > 
> >       WARN_ON(!sg);
> > 
> >       do {
> > -		int cpu, max_cpu = -1;
> > +		int cpu, cores = 0, max_cpu = -1;
> > 
> >               sg->group_weight = cpumask_weight(sched_group_span(sg));
> > 
> > +		cpumask_copy(mask, sched_group_span(sg));
> > +		for_each_cpu(cpu, mask) {
> > +			cores++;
> > +			cpumask_andnot(mask, mask, cpu_smt_mask(cpu));
> > +		}
> 
> 
> This rekindled my desire for an SMT core cpumask/iterator. I played around
> with a global mask but that's a headache: what if we end up with a core
> whose SMT threads are split across two exclusive cpusets?

Peter and I pondered that for a while.  But it seems like partitioning
threads in a core between two different sched domains is not a very
reasonable thing to do. 

https://lore.kernel.org/all/20230612112945.GK4253@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/

Tim

> 
> I ended up necro'ing a patch from Peter [1], but didn't get anywhere nice
> (the LLC shared storage caused me issues).
> 
> All that to say, I couldn't think of a nicer way :(
> 
> [1]: https://lore.kernel.org/all/20180530143106.082002139@xxxxxxxxxxxxx/#t
>