Re: [PATCH 3/4] sched: add sched_numa_find_nth_cpu()

[Yury Norov]

On Fri, Nov 11, 2022 at 01:42:29PM +0200, Andy Shevchenko wrote:
> On Thu, Nov 10, 2022 at 08:00:26PM -0800, Yury Norov wrote:
> > The function finds Nth set CPU in a given cpumask starting from a given
> > node.
> > 
> > Leveraging the fact that each hop in sched_domains_numa_masks includes the
> > same or greater number of CPUs than the previous one, we can use binary
> > search on hops instead of linear walk, which makes the overall complexity
> > of O(log n) in terms of number of cpumask_weight() calls.
> 
> ...
> 
> > +int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
> > +{
> > +	unsigned int first = 0, mid, last = sched_domains_numa_levels;
> > +	struct cpumask ***masks;
> 
> *** ?
> Hmm... Do we really need such deep indirection?

It's 2d array of pointers, so - yes.
 
> > +	int w, ret = nr_cpu_ids;
> > +
> > +	rcu_read_lock();
> > +	masks = rcu_dereference(sched_domains_numa_masks);
> > +	if (!masks)
> > +		goto out;
> > +
> > +	while (last >= first) {
> > +		mid = (last + first) / 2;
> > +
> > +		if (cpumask_weight_and(cpus, masks[mid][node]) <= cpu) {
> > +			first = mid + 1;
> > +			continue;
> > +		}
> > +
> > +		w = (mid == 0) ? 0 : cpumask_weight_and(cpus, masks[mid - 1][node]);
> 
> See below.
> 
> > +		if (w <= cpu)
> > +			break;
> > +
> > +		last = mid - 1;
> > +	}
> 
> We have lib/bsearch.h. I haven't really looked deeply into the above, but my
> gut feelings that that might be useful here. Can you check that?

Yes we do. I tried it, and it didn't work because nodes arrays are
allocated dynamically, and distance between different pairs of hops
for a given node is not a constant, which is a requirement for
bsearch.

However, distance between hops pointers in 1st level array should be
constant, and we can try feeding bsearch with it. I'll experiment with
bsearch for more.

> > +	ret = (mid == 0) ?
> > +		cpumask_nth_and(cpu - w, cpus, masks[mid][node]) :
> > +		cpumask_nth_and_andnot(cpu - w, cpus, masks[mid][node], masks[mid - 1][node]);
> 
> You can also shorten this by inversing the conditional:
> 
> 	ret = mid ? ...not 0... : ...for 0...;

Yep, why not.

> > +out:
> 
> out_unlock: ?

Do you think it's better?

> > +	rcu_read_unlock();
> > +	return ret;
> > +}
> 
> -- 
> With Best Regards,
> Andy Shevchenko
>