Re:Re: [PATCH-cpuset v10 2/2] cpuset: use Union-Find to optimize the merging of cpumasks

[Xavier <xavier_qy@163.com>]

Hi Michal and Longman,

Please confirm my explanation about cgroup v2 below.


At 2024-07-03 17:40:49, "Michal Koutný" <mkoutny@suse.com> wrote:
>On Wed, Jul 03, 2024 at 02:37:27PM GMT, Xavier <xavier_qy@163.com> wrote:
>> @@ -1102,31 +1101,25 @@ static int generate_sched_domains(cpumask_var_t **domains,
>>  	if (root_load_balance && (csn == 1))
>>  		goto single_root_domain;
>>  
>> -	for (i = 0; i < csn; i++)
>> -		csa[i]->pn = i;
>> -	ndoms = csn;
>> -
>> -restart:
>> -	/* Find the best partition (set of sched domains) */
>> -	for (i = 0; i < csn; i++) {
>> -		struct cpuset *a = csa[i];
>> -		int apn = a->pn;
>> -
>> -		for (j = 0; j < csn; j++) {
>> -			struct cpuset *b = csa[j];
>> -			int bpn = b->pn;
>> -
>> -			if (apn != bpn && cpusets_overlap(a, b)) {
>> -				for (k = 0; k < csn; k++) {
>> -					struct cpuset *c = csa[k];
>> +	if (!cgrpv2) {
>
>I'm surprised that original code wasn't branched on this on you add it
>here. Why is UF used only for v1 code?
>

In the Patch v6, I explained to Longman that based on his new patch, the overlapping check and
merge operations for cpusets are skipped in the case of cgroup v2. Because for cgroup v2,
doms[i] is merely copied from csa[i] rather than merged.
This needs further confirmation from Longman.

	if (cgrpv2) {
		for (i = 0; i < ndoms; i++) {
			cpumask_copy(doms[i], csa[i]->effective_cpus);
			if (dattr)
				dattr[i] = SD_ATTR_INIT;
		}
		goto done;
	}


>> +		for (i = 0; i < csn; i++)
>> +			uf_node_init(&csa[i]->node);
>>  
>> -					if (c->pn == bpn)
>> -						c->pn = apn;
>> -				}
>> -				ndoms--;	/* one less element */
>> -				goto restart;
>> +		/* Merge overlapping cpusets */
>> +		for (i = 0; i < csn; i++) {
>> +			for (j = i + 1; j < csn; j++) {
>> +				if (cpusets_overlap(csa[i], csa[j]))
>> +					uf_union(&csa[i]->node, &csa[j]->node);
>>  			}
>>  		}
>> +
>> +		/* Count the total number of domains */
>> +		for (i = 0; i < csn; i++) {
>> +			if (csa[i]->node.parent == &csa[i]->node)
>> +				ndoms++;
>
>The naked parent access doesn't hide the UF abstraction well.
>I'd consider uf_find(&csa[i]->node) == &csa[i]->node or a specific
>helper like uf_is_representant(&csa[i]->node).

This can be optimized here. I will change it to the first method you mentioned in the next patch.

Best Regards,
Xavier