Re: [PATCH v4] sched/fair: Skip sched_balance_running cmpxchg when balance is not due

[Srikar Dronamraju]

* Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx> [2025-11-10 10:47:35]:

> The NUMA sched domain sets the SD_SERIALIZE flag by default, allowing
> only one NUMA load balancing operation to run system-wide at a time.
> 
> @@ -11757,17 +11772,26 @@ static int sched_balance_rq(int this_cpu, struct rq *this_rq,
>  		.fbq_type	= all,
>  		.tasks		= LIST_HEAD_INIT(env.tasks),
>  	};
> +	bool need_unlock;
>  
>  	cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
>  
>  	schedstat_inc(sd->lb_count[idle]);
>  
>  redo:
> +	need_unlock = false;
>  	if (!should_we_balance(&env)) {
>  		*continue_balancing = 0;
>  		goto out_balanced;
>  	}
>  
> +	if (sd->flags & SD_SERIALIZE) {
> +		if (atomic_cmpxchg_acquire(&sched_balance_running, 0, 1)) {
> +			goto out_balanced;
> +		}
> +		need_unlock = true;
> +	}
> +

Moving the serialize check to sched_balance_rq is better since we only take
when its really needed. Previously we could have skipped the balancing for
the appropriate CPU.

>  	group = sched_balance_find_src_group(&env);
>  	if (!group) {
>  		schedstat_inc(sd->lb_nobusyg[idle]);
> @@ -11892,6 +11916,9 @@ static int sched_balance_rq(int this_cpu, struct rq *this_rq,
>  			if (!cpumask_subset(cpus, env.dst_grpmask)) {
>  				env.loop = 0;
>  				env.loop_break = SCHED_NR_MIGRATE_BREAK;
> +				if (need_unlock)
> +					atomic_set_release(&sched_balance_running, 0);
> +

One nit:
While the current code is good, would conditionally resetting the
need_unlock just after resetting the atomic variable better than
unconditional reset that we do now?

>  				goto redo;

Otherwise looks good to me.

Reviewed-by: Srikar Dronamraju <srikar@xxxxxxxxxxxxx>

-- 
Thanks and Regards
Srikar Dronamraju