Re: [PATCH v2] sched/topology: improve topology_span_sane speed

[Valentin Schneider]

Hi,

Sorry it took me this long to get to this.

On 31/10/24 15:04, Steve Wahl wrote:
> Use a different approach to topology_span_sane(), that checks for the
> same constraint of no partial overlaps for any two CPU sets for
> non-NUMA topology levels, but does so in a way that is O(N) rather
> than O(N^2).
>
> Instead of comparing with all other masks to detect collisions, keep
> one mask that includes all CPUs seen so far and detect collisions with
> a single cpumask_intersects test.
>
> If the current mask has no collisions with previously seen masks, it
> should be a new mask, which can be uniquely identified ("id") by the
> lowest bit set in this mask.  Mark that we've seen a mask with this
> id, and add the CPUs in this mask to the list of those seen.
>
> If the current mask does collide with previously seen masks, it should
> be exactly equal to a mask seen before, identified once again by the
> lowest bit the current mask has set.  It's an error if we haven't seen
> a mask with that id, or if the current mask doesn't match the one we
> get by looking up that id.
>
> Move the topology_span_sane() check out of the existing topology level
> loop, let it do its own looping to match the needs of this algorithm.
>
> On a system with 1920 processors (16 sockets, 60 cores, 2 threads),
> the average time to take one processor offline is reduced from 2.18
> seconds to 1.01 seconds.  (Off-lining 959 of 1920 processors took
> 34m49.765s without this change, 16m10.038s with this change in place.)
>
> Signed-off-by: Steve Wahl <steve.wahl@xxxxxxx>
> ---
>
> Version 2: Adopted suggestion by K Prateek Nayak that removes an array and
> simplifies the code, and eliminates the erroneous use of
> num_possible_cpus() that Peter Zijlstra noted.
>
> Version 1 discussion:
>     https://lore.kernel.org/all/20241010155111.230674-1-steve.wahl@xxxxxxx/
>
>  kernel/sched/topology.c | 73 +++++++++++++++++++++++++++--------------
>  1 file changed, 48 insertions(+), 25 deletions(-)
>
> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
> index 9748a4c8d668..6a2a3e91d59e 100644
> --- a/kernel/sched/topology.c
> +++ b/kernel/sched/topology.c
> @@ -2356,35 +2356,58 @@ static struct sched_domain *build_sched_domain(struct sched_domain_topology_leve
>  
>  /*
>   * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
> - * any two given CPUs at this (non-NUMA) topology level.
> + * any two given CPUs on non-NUMA topology levels.
>   */
> -static bool topology_span_sane(struct sched_domain_topology_level *tl,
> -			      const struct cpumask *cpu_map, int cpu)
> +static bool topology_span_sane(const struct cpumask *cpu_map)
>  {
> -	int i = cpu + 1;
> +	struct sched_domain_topology_level *tl;
> +	struct cpumask *covered, *id_seen;
> +	int cpu;
>  
> -	/* NUMA levels are allowed to overlap */
> -	if (tl->flags & SDTL_OVERLAP)
> -		return true;
> +	lockdep_assert_held(&sched_domains_mutex);
> +	covered = sched_domains_tmpmask;
> +	id_seen = sched_domains_tmpmask2;
> +
> +	for_each_sd_topology(tl) {
> +
> +		/* NUMA levels are allowed to overlap */
> +		if (tl->flags & SDTL_OVERLAP)
> +			continue;
> +
> +		cpumask_clear(covered);
> +		cpumask_clear(id_seen);
>  
> -	/*
> -	 * Non-NUMA levels cannot partially overlap - they must be either
> -	 * completely equal or completely disjoint. Otherwise we can end up
> -	 * breaking the sched_group lists - i.e. a later get_group() pass
> -	 * breaks the linking done for an earlier span.
> -	 */
> -	for_each_cpu_from(i, cpu_map) {
>  		/*
> -		 * We should 'and' all those masks with 'cpu_map' to exactly
> -		 * match the topology we're about to build, but that can only
> -		 * remove CPUs, which only lessens our ability to detect
> -		 * overlaps
> +		 * Non-NUMA levels cannot partially overlap - they must be either
> +		 * completely equal or completely disjoint. Otherwise we can end up
> +		 * breaking the sched_group lists - i.e. a later get_group() pass
> +		 * breaks the linking done for an earlier span.
>  		 */
> -		if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
> -		    cpumask_intersects(tl->mask(cpu), tl->mask(i)))
> -			return false;
> +		for_each_cpu(cpu, cpu_map) {
> +			const struct cpumask *tl_cpu_mask = tl->mask(cpu);
> +			int id;
> +
> +			/* lowest bit set in this mask is used as a unique id */
> +			id = cpumask_first(tl_cpu_mask);
> +

Using the first CPU like we do in get_group() is clever; if we've seen the
first CPU then the spans must exactly match; else it must be a completely
new span.


> +			/* if this mask doesn't collide with what we've already seen */
> +			if (!cpumask_intersects(tl_cpu_mask, covered)) {
> +				/* Really odd case when cpu != id, likely not sane */
> +				if ((cpu != id) && !cpumask_equal(tl_cpu_mask, tl->mask(id)))
> +					return false;
> +				if (cpumask_test_and_set_cpu(id, id_seen))
> +					return false;
> +				cpumask_or(covered, tl_cpu_mask, covered);
> +			} else if ((!cpumask_test_cpu(id, id_seen)) ||
> +				    !cpumask_equal(tl->mask(id), tl_cpu_mask)) {
> +				/*
> +				 * a collision with covered should have exactly matched
> +				 * a previously seen mask with the same id
> +				 */
> +				return false;
> +			}


Per the previous comment, if you hinge the condition on @id itself,
couldn't you "simplify" it as below?

	if (cpumask_test_cpu(id, id_seen)) {
		/* First CPU has already been seen, ensure identical spans */
		if (!cpumask_equal(tl->mask(id), tl_cpu_mask))
			return false;
	} else {
		/* First CPU hasn't been seen before, ensure it's a completely new span */
		if (cpumask_intersects(tl_cpu_mask, covered))
			return false;

		cpumask_or(covered, covered, tl_cpu_mask);
		cpumask_set_cpu(id, id_seen);
	}