Re: [PATCH v2 1/2] sched/topology: Skip updating masks for non-online nodes

[Valentin Schneider]

On 01/07/21 09:45, Srikar Dronamraju wrote:
> @@ -1891,12 +1894,30 @@ void sched_init_numa(void)
>  void sched_domains_numa_masks_set(unsigned int cpu)
>  {
>       int node = cpu_to_node(cpu);
> -	int i, j;
> +	int i, j, empty;
>
> +	empty = cpumask_empty(sched_domains_numa_masks[0][node]);
>       for (i = 0; i < sched_domains_numa_levels; i++) {
>               for (j = 0; j < nr_node_ids; j++) {
> -			if (node_distance(j, node) <= sched_domains_numa_distance[i])
> +			if (!node_online(j))
> +				continue;
> +
> +			if (node_distance(j, node) <= sched_domains_numa_distance[i]) {
>                               cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
> +
> +				/*
> +				 * We skip updating numa_masks for offline
> +				 * nodes. However now that the node is
> +				 * finally online, CPUs that were added
> +				 * earlier, should now be accommodated into
> +				 * newly oneline node's numa mask.
> +				 */
> +				if (node != j && empty) {
> +					cpumask_or(sched_domains_numa_masks[i][node],
> +							sched_domains_numa_masks[i][node],
> +							sched_domains_numa_masks[0][j]);
> +				}
> +			}

Hmph, so we're playing games with masks of offline nodes - is that really
necessary? Your modification of sched_init_numa() still scans all of the
nodes (regardless of their online status) to build the distance map, and
that is never updated (sched_init_numa() is pretty much an __init
function).

So AFAICT this is all to cope with topology_span_sane() not applying
'cpu_map' to its masks. That seemed fine to me back when I wrote it, but in
light of having bogus distance values for offline nodes, not so much...

What about the below instead?

---
diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
index b77ad49dc14f..c2d9caad4aa6 100644
--- a/kernel/sched/topology.c
+++ b/kernel/sched/topology.c
@@ -2075,6 +2075,7 @@ static struct sched_domain *build_sched_domain(struct sched_domain_topology_leve
 static bool topology_span_sane(struct sched_domain_topology_level *tl,
 			      const struct cpumask *cpu_map, int cpu)
 {
+	struct cpumask *intersect = sched_domains_tmpmask;
 	int i;
 
 	/* NUMA levels are allowed to overlap */
@@ -2090,14 +2091,17 @@ static bool topology_span_sane(struct sched_domain_topology_level *tl,
 	for_each_cpu(i, cpu_map) {
 		if (i == cpu)
 			continue;
+
 		/*
-		 * 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
+		 * We shouldn't have to bother with cpu_map here, unfortunately
+		 * some architectures (powerpc says hello) have to deal with
+		 * offline NUMA nodes reporting bogus distance values. This can
+		 * lead to funky NODE domain spans, but since those are offline
+		 * we can mask them out.
 		 */
+		cpumask_and(intersect, tl->mask(cpu), tl->mask(i));
 		if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
-		    cpumask_intersects(tl->mask(cpu), tl->mask(i)))
+		    cpumask_intersects(intersect, cpu_map))
 			return false;
 	}