SD_OVERLAP exists to allow overlapping groups, overlapping groups appear in NUMA topologies that aren't fully connected. The typical result of not fully connected NUMA is that each cpu (or rather node) will have different spans for a particular distance. However due to how sched domains are traversed -- only the first cpu in the mask goes one level up -- the next level only cares about the spans of the cpus that went up. Due to this two things were observed to be broken: - build_overlap_sched_groups() -- since its possible the cpu we're building the groups for exists in multiple (or all) groups, the selection criteria of the first group didn't ensure there was a cpu for which is was true that cpumask_first(span) == cpu. Thus load- balancing would terminate. - update_group_power() -- assumed that the cpu span of the first group of the domain was covered by all groups of the child domain. The above explains why this isn't true, so deal with it.
Showing the changes to kernel (2 of 2 files). Show all
Unified · Split
| @@ -6030,11 +6030,14 @@ build_overlap_sched_groups(struct sched_domain *sd, int cpu) · 6029 unmodified lines | ||
| 6030 | 6030 | |
| 6031 | 6031 | cpumask_or(covered, covered, sg_span); |
| 6032 | 6032 | |
| 6033 | − sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span)); | |
| 6033 | + sg->sgp = *per_cpu_ptr(sdd->sgp, i); | |
| 6034 | 6034 | atomic_inc(&sg->sgp->ref); |
| 6035 | 6035 | |
| 6036 | − if (cpumask_test_cpu(cpu, sg_span)) | |
| 6036 | + if ((!groups && cpumask_test_cpu(cpu, sg_span)) || | |
| 6037 | + cpumask_first(sg_span) == cpu) { | |
| 6038 | + WARN_ON_ONCE(!cpumask_test_cpu(cpu, sg_span)); | |
| 6037 | 6039 | groups = sg; |
| 6040 | + } | |
| 6038 | 6041 | |
| 6039 | 6042 | if (!first) |
| 6040 | 6043 | first = sg; |
| @@ -3574,11 +3574,26 @@ void update_group_power(struct sched_domain *sd, int cpu) · 3573 unmodified lines | ||
| 3574 | 3574 | |
| 3575 | 3575 | power = 0; |
| 3576 | 3576 | |
| 3577 | − group = child->groups; | |
| 3578 | − do { | |
| 3579 | − power += group->sgp->power; | |
| 3580 | − group = group->next; | |
| 3581 | − } while (group != child->groups); | |
| 3577 | + if (child->flags & SD_OVERLAP) { | |
| 3578 | + /* | |
| 3579 | + * SD_OVERLAP domains cannot assume that child groups | |
| 3580 | + * span the current group. | |
| 3581 | + */ | |
| 3582 | + | |
| 3583 | + for_each_cpu(cpu, sched_group_cpus(sdg)) | |
| 3584 | + power += power_of(cpu); | |
| 3585 | + } else { | |
| 3586 | + /* | |
| 3587 | + * !SD_OVERLAP domains can assume that child groups | |
| 3588 | + * span the current group. | |
| 3589 | + */ | |
| 3590 | + | |
| 3591 | + group = child->groups; | |
| 3592 | + do { | |
| 3593 | + power += group->sgp->power; | |
| 3594 | + group = group->next; | |
| 3595 | + } while (group != child->groups); | |
| 3596 | + } | |
| 3582 | 3597 | |
| 3583 | 3598 | sdg->sgp->power = power; |
| 3584 | 3599 | } |