From: Adam Li Date: Sat, 11 Oct 2025 06:43:22 +0000 (+0000) Subject: sched/fair: Only update stats for allowed CPUs when looking for dst group X-Git-Url: http://git.ipfire.org/gitweb.cgi?a=commitdiff_plain;h=82d6e01a0699800efd8b048eb584c907ccb47b7a;p=thirdparty%2Flinux.git sched/fair: Only update stats for allowed CPUs when looking for dst group Load imbalance is observed when the workload frequently forks new threads. Due to CPU affinity, the workload can run on CPU 0-7 in the first group, and only on CPU 8-11 in the second group. CPU 12-15 are always idle. { 0 1 2 3 4 5 6 7 } {8 9 10 11 12 13 14 15} * * * * * * * * * * * * When looking for dst group for newly forked threads, in many times update_sg_wakeup_stats() reports the second group has more idle CPUs than the first group. The scheduler thinks the second group is less busy. Then it selects least busy CPUs among CPU 8-11. Therefore CPU 8-11 can be crowded with newly forked threads, at the same time CPU 0-7 can be idle. A task may not use all the CPUs in a schedule group due to CPU affinity. Only update schedule group statistics for allowed CPUs. Signed-off-by: Adam Li Signed-off-by: Peter Zijlstra (Intel) --- diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 00f9d6c05d4cf..ac881df0eebe8 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -10683,7 +10683,7 @@ static inline void update_sg_wakeup_stats(struct sched_domain *sd, if (sd->flags & SD_ASYM_CPUCAPACITY) sgs->group_misfit_task_load = 1; - for_each_cpu(i, sched_group_span(group)) { + for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { struct rq *rq = cpu_rq(i); unsigned int local;