Re: [PATCH] sched: Reduce rq lock contention in load_balance()

[Abel Wu <wuyun.abel@bytedance.com>]

On 12/13/22 11:13 AM, chenying wrote:
> [nit]
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index e4a0b8bd941c..aeb4fa9ac93a 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -10295,6 +10295,7 @@ static int load_balance(int this_cpu, struct rq 
> *this_rq,
>                  goto out_balanced;
>          }
> 
> +refind:
>          busiest = find_busiest_queue(&env, group);
>          if (!busiest) {
>                  schedstat_inc(sd->lb_nobusyq[idle]);
> @@ -10303,6 +10304,14 @@ static int load_balance(int this_cpu, struct rq 
> *this_rq,
> 
>          WARN_ON_ONCE(busiest == env.dst_rq);
> 
> +       if (READ_ONCE(busiest->balancing)) {
> +               __cpumask_clear_cpu(cpu_of(busiest), cpus);
> +               if (cpumask_intersects(sched_group_span(group), cpus))
> +                       goto refind;
> +
> +               goto out_balanced;
> +       }
> +

Here removing the cpu from @cpus will prevent it being selected once
a redo is triggered due to all tasks on the busiest cpu pinned by cpu
affinity. If that is the case, the removed cpu can still be the busiest
but not in balancing at that moment.

IMHO it'd be better skip the in-balancing cpus in find_busiest_queue()
without modifying @cpus to keep consistence among the redos.

Thanks & Best,
	Abel