Re: [PATCH v4 2/3] sched/topology: Rework CPU capacity asymmetry detection

[Beata Michalska]

On Mon, May 17, 2021 at 04:06:05PM +0100, Valentin Schneider wrote:
> On 17/05/21 14:18, Beata Michalska wrote:
> > On Mon, May 17, 2021 at 01:04:25PM +0100, Valentin Schneider wrote:
> >> On 17/05/21 09:23, Beata Michalska wrote:
> >> > +static void asym_cpu_capacity_scan(const struct cpumask *cpu_map)
> >> > +{
> >> > +	struct asym_cap_data *entry, *next;
> >> > +	int cpu;
> >> >
> >> > -		for_each_sd_topology(tl) {
> >> > -			if (tl_id < asym_level)
> >> > -				goto next_level;
> >> > +	if (!list_empty(&asym_cap_list))
> >> > +		list_for_each_entry(entry, &asym_cap_list, link)
> >> > +			cpumask_clear(entry->cpu_mask);
> >> >
> >>
> >> The topology isn't going to change between domain rebuilds, so why
> >> recompute the masks? The sched_domain spans are already masked by cpu_map,
> >> so no need to do this masking twice. I'm thinking this scan should be done
> >> once against the cpu_possible_mask - kinda like sched_init_numa() done once
> >> against the possible nodes.
> >>
> > This is currently done, as what you have mentioned earlier, the tl->mask
> > may contain CPUs that are not 'available'. So it makes sure that the masks
> > kept on  the list are representing only those CPUs that are online.
> > And it is also needed case all CPUs of given capacity go offline - not to to
> > lose the full asymmetry that might change because of that ( empty masks are
> > being removed from the list).
> >
> > I could change that and use the CPU mask that represents the online CPUs as
> > a checkpoint but then it also means additional tracking which items on the
> > list are actually available at a given point of time.
> > So if the CPUs masks on the list are to be set once (as you are suggesting)
> > than it needs additional logic to count the number of available capacities
> > to decide whether there is a full asymmetry or not.
> >
> 
> That should be doable by counting non-empty intersections between each
> entry->cpumask and the cpu_online_mask in _classify().
> 
> That said I'm afraid cpufreq module loading forces us to dynamically update
> those masks, as you've done. The first domain build could see asymmetry
> without cpufreq loaded, and a later one with cpufreq loaded would need an
> update. Conversely, as much of a fringe case as it is, we'd have to cope
> with the cpufreq module being unloaded later on...
> 
> :(
So it got me thinking that maybe we could actually make it more
'update-on-demand' and use the cpufreq policy notifier to trigger the update.
I could try to draft smth generic enough to make it ... relatively easy to adapt
to different archs case needed.
Any thoughts ?

---
BR
B.