Re: [PATCH 17/22] sched/core: fix opencoded cpumask_any_but()

[Valentin Schneider <vschneid@redhat.com>]

On 10/05/22 18:37, Peter Zijlstra wrote:
> On Tue, May 10, 2022 at 08:47:45AM -0700, Yury Norov wrote:
>> sched_core_cpu_starting() and sched_core_cpu_deactivate() implement
>> opencoded cpumask_any_but(). Fix it.
>>
>> CC: Ben Segall <bsegall@google.com>
>> CC: Daniel Bristot de Oliveira <bristot@redhat.com>
>> CC: Dietmar Eggemann <dietmar.eggemann@arm.com>
>> CC: Ingo Molnar <mingo@redhat.com>
>> CC: Juri Lelli <juri.lelli@redhat.com>
>> CC: Mel Gorman <mgorman@suse.de>
>> CC: Peter Zijlstra <peterz@infradead.org>
>> CC: Steven Rostedt <rostedt@goodmis.org>
>> CC: Valentin Schneider <vschneid@redhat.com>
>> CC: Vincent Guittot <vincent.guittot@linaro.org>
>> CC: linux-kernel@vger.kernel.org
>> Signed-off-by: Yury Norov <yury.norov@gmail.com>
>> ---
>>  kernel/sched/core.c | 33 +++++++++++++--------------------
>>  1 file changed, 13 insertions(+), 20 deletions(-)
>>
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index f5ebc392493d..9700001948d0 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -6125,7 +6125,7 @@ static void queue_core_balance(struct rq *rq)
>>  static void sched_core_cpu_starting(unsigned int cpu)
>>  {
>>      const struct cpumask *smt_mask = cpu_smt_mask(cpu);
>> -	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
>> +	struct rq *rq = cpu_rq(cpu), *core_rq;
>>      unsigned long flags;
>>      int t;
>>
>> @@ -6138,19 +6138,16 @@ static void sched_core_cpu_starting(unsigned int cpu)
>>              goto unlock;
>>
>>      /* find the leader */
>> -	for_each_cpu(t, smt_mask) {
>> -		if (t == cpu)
>> -			continue;
>> -		rq = cpu_rq(t);
>> -		if (rq->core == rq) {
>> -			core_rq = rq;
>> -			break;
>> -		}
>> -	}
>> +	t = cpumask_any_but(smt_mask, cpu);
>> +	if (t >= nr_cpu_ids)
>> +		goto unlock;
>>
>> -	if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
>> +	rq = cpu_rq(t);
>> +	if (WARN_ON_ONCE(rq->core != rq)) /* whoopsie */
>>              goto unlock;
>>
>> +	core_rq = rq;
>> +
>>      /* install and validate core_rq */
>>      for_each_cpu(t, smt_mask) {
>>              rq = cpu_rq(t);
>
> I don't think this is equivalent. Imagine SMT4, with:
>
>   rqN->core_rq = rq0
>
> Now, further suppose smt0-2 are online and we're about to online smt3.
> Then t above is free to be smt2, which then results in insta triggering:
>
> +	if (WARN_ON_ONCE(rq->core != rq)) /* whoopsie */
>
> You seem to have lost how the first loop searches for rq->core.
>

cpumask_any() is actually cpumask_first(), so t should be smt0 in that
case. However, if for some reason rq->core isn't the first online CPU in
smt_mask \ {cpu} (which I think can happen if you offline smt0-1 then
re-online smt0), then yes that splats.

> Please, be more careful. Also, all of this is super cold path don't
> bother with optimizations. Much of the patches you have in this series
> fall under that.

I tend to agree, I do like the cpumask_weight_eq() stuff because it's a low
hanging fruit and can even be autopatched with coccinelle, but the
open-coded stuff in cold paths isn't as relevant (nor as obvious as it may
look :)).