On Fri, Jul 11, 2025 at 11:20:30AM +0530, K Prateek Nayak wrote:

On 7/10/2025 4:27 PM, Li Chen wrote:

quoted

 	/*
 	 * .. and append 'j' levels of NUMA goodness.
 	 */
 	for (j = 1; j < nr_levels; i++, j++) {
-		tl[i] = (struct sched_domain_topology_level){
-			.mask = sd_numa_mask,
-			.sd_flags = cpu_numa_flags,
-			.flags = SDTL_OVERLAP,
-			.numa_level = j,
-			SD_INIT_NAME(NUMA)
-		};
+		tl[i] = SDTL_INIT(sd_numa_mask, cpu_numa_flags, NUMA);
+		tl[i].numa_level = j;
+		tl[i].flags = SDTL_OVERLAP;

Tangential discussion: I was looking at this and was wondering why we
need a "tl->flags" when there is already sd_flags() function and we can
simply add SD_OVERLAP to sd_numa_flags().

I think "tl->flags" was needed when the idea of overlap domains was
added in commit e3589f6c81e4 ("sched: Allow for overlapping sched_domain
spans") when it depended on "FORCE_SD_OVERLAP" sched_feat() which
allowed toggling this off but that was done away with in commit
af85596c74de ("sched/topology: Remove FORCE_SD_OVERLAP") so perhaps we
can get rid of it now?

Relying on SD_NUMA should be enough currently. Peter, Valentin, what do
you think of something like below?

I think you're right. SD_NUMA appears to be the one and only case that
also has SDTL_OVERLAP which then results in setting SD_OVERLAP, making
SD_NUMA and SD_OVERLAP equivalent and SDTL_OVERLAP redundant.

I'll presume you're okay with me adding your SoB to things, and I'll
push out all 5 patches to queue/sched/core to let the robots have a go
at things.