Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Tue, 25 Oct 2022 10:32:26 +0100
From:   Mel Gorman <mgorman@suse.de>
To:     Hao Jia <jiahao.os@bytedance.com>
Cc:     mingo@redhat.com, peterz@infradead.org, mingo@kernel.org,
        juri.lelli@redhat.com, vincent.guittot@linaro.org,
        dietmar.eggemann@arm.com, rostedt@goodmis.org, bsegall@google.com,
        bristot@redhat.com, vschneid@redhat.com,
        mgorman@techsingularity.net, linux-kernel@vger.kernel.org
Subject: Re: [External] Re: [PATCH 1/2] sched/numa: Stop an exhastive search
 if an idle core is found
Message-ID: <20221025093226.dm4sjvdq2tofkwvc@suse.de>
References: <20221021061558.34767-1-jiahao.os@bytedance.com>
 <20221021061558.34767-2-jiahao.os@bytedance.com>
 <20221024133435.e2kajx5k7jzznp25@suse.de>
 <ced7c05a-121b-a77d-0c57-3e60abaecacd@bytedance.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-15
Content-Disposition: inline
In-Reply-To: <ced7c05a-121b-a77d-0c57-3e60abaecacd@bytedance.com>
Precedence: bulk

On Tue, Oct 25, 2022 at 11:16:29AM +0800, Hao Jia wrote:
> > Remove the change in the first hunk and call break in the second hunk
> > after updating ns->idle_cpu.
> > 
> 
> Yes, thanks for your review.
> If I understand correctly, some things might look like this.
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index e4a0b8bd941c..dfcb620bfe50 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -1792,7 +1792,7 @@ static void update_numa_stats(struct task_numa_env
> *env,
>                 ns->nr_running += rq->cfs.h_nr_running;
>                 ns->compute_capacity += capacity_of(cpu);
> 
> -               if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
> +               if (find_idle && idle_core < 0 && !rq->nr_running &&
> idle_cpu(cpu)) {
>                         if (READ_ONCE(rq->numa_migrate_on) ||
>                             !cpumask_test_cpu(cpu, env->p->cpus_ptr))
>                                 continue;
> 

I meant more like the below but today I wondered why did I not do this in
the first place? The answer is because it's wrong and broken in concept.

The full loop is needed to calculate approximate NUMA stats at a
point in time. For example, the src and dst nr_running is needed by
task_numa_find_cpu. The search for an idle CPU or core in update_numa_stats
is simply taking advantage of the fact we are scanning anyway to keep
track of an idle CPU or core to avoid a second search as per ff7db0bf24db
("sched/numa: Prefer using an idle CPU as a migration target instead of
comparing tasks")

The patch I had in mind is below but that said, for both your version and
my initial suggestion

Naked-by: Mel Gorman <mgorman@suse.de>

For the record, this is what I was suggesting initially because it's more
efficient but it's wrong, don't do it.

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index e4a0b8bd941c..7f1f6a1736a5 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -1800,7 +1800,12 @@ static void update_numa_stats(struct task_numa_env *env,
 			if (ns->idle_cpu == -1)
 				ns->idle_cpu = cpu;
 
+			/* If we find an idle core, stop searching. */
 			idle_core = numa_idle_core(idle_core, cpu);
+			if (idle_core >= 0) {
+				ns->idle_cpu = idle_core;
+				break;
+			}
 		}
 	}
 	rcu_read_unlock();
@@ -1808,9 +1813,6 @@ static void update_numa_stats(struct task_numa_env *env,
 	ns->weight = cpumask_weight(cpumask_of_node(nid));
 
 	ns->node_type = numa_classify(env->imbalance_pct, ns);
-
-	if (idle_core >= 0)
-		ns->idle_cpu = idle_core;
 }
 
 static void task_numa_assign(struct task_numa_env *env,