On 05/29/2018 02:27 AM, Juri Lelli wrote: > On 28/05/18 21:24, Waiman Long wrote: >> On 05/28/2018 09:12 PM, Waiman Long wrote: >>> On 05/24/2018 06:28 AM, Juri Lelli wrote: >>>> On 17/05/18 16:55, Waiman Long wrote: >>>> >>>> [...] >>>> >>>>> @@ -849,7 +860,12 @@ static void rebuild_sched_domains_locked(void) >>>>> * passing doms with offlined cpu to partition_sched_domains(). >>>>> * Anyways, hotplug work item will rebuild sched domains. >>>>> */ >>>>> - if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask)) >>>>> + if (!top_cpuset.isolation_count && >>>>> + !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask)) >>>>> + goto out; >>>>> + >>>>> + if (top_cpuset.isolation_count && >>>>> + !cpumask_subset(top_cpuset.effective_cpus, cpu_active_mask)) >>>>> goto out; >>>> Do we cover the case in which hotplug removed one of the isolated cpus >>>> from cpu_active_mask? >>> Yes, you are right. That is the remnant of my original patch that allows >>> only one isolated_cpus at root. Thanks for spotting that. >> I am sorry. I would like to take it back my previous comment. The code >> above looks for inconsistency in the state of the effective_cpus mask to >> find out if it is racing with a hotplug event. If it is, we can skip the >> domain generation as the hotplug event will do that too. The checks are >> still valid with the current patchset. So I don't think we need to make >> any change here. > Yes, these checks are valid, but don't we also need to check for hotplug > races w.r.t. isolated CPUs (of some other sub domain)? It is not actually a race. Both the hotplug event and any changes to cpu lists or flags are serialized by the cpuset_mutex. It is just that we may be doing the same work twice that we are wasting cpu cycles. So we are doing a quick check to avoid this. The check isn't exhaustive and we can certainly miss some cases. Doing a more throughout check may need as much time as doing the sched domain generation itself and so you are actually wasting more CPU cycles on average as the chance of a hotplug event is very low. Cheers, Longman -- To unsubscribe from this list: send the line "unsubscribe cgroups" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
