Received-SPF: pass (google.com: domain of linux-kernel+bounces-169638-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Date: Mon, 6 May 2024 12:05:09 +0200
From: Peter Zijlstra <peterz@infradead.org>
To: Qais Yousef <qyousef@layalina.io>
Cc: "Rafael J. Wysocki" <rafael@kernel.org>,
	Viresh Kumar <viresh.kumar@linaro.org>,
	Ingo Molnar <mingo@kernel.org>,
	Vincent Guittot <vincent.guittot@linaro.org>,
	Juri Lelli <juri.lelli@redhat.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Ben Segall <bsegall@google.com>, Mel Gorman <mgorman@suse.de>,
	Daniel Bristot de Oliveira <bristot@redhat.com>,
	Valentin Schneider <vschneid@redhat.com>,
	Christian Loehle <christian.loehle@arm.com>,
	linux-pm@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH v2] sched: Consolidate cpufreq updates
Message-ID: <20240506100509.GL40213@noisy.programming.kicks-ass.net>
References: <20240505233103.168766-1-qyousef@layalina.io>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20240505233103.168766-1-qyousef@layalina.io>

On Mon, May 06, 2024 at 12:31:03AM +0100, Qais Yousef wrote:

> +static inline void update_cpufreq_ctx_switch(struct rq *rq, struct task_struct *prev)
> +{
> +#ifdef CONFIG_CPU_FREQ
> +	unsigned int flags = 0;
> +
> +#ifdef CONFIG_SMP
> +	if (unlikely(current->sched_class == &stop_sched_class))
> +		return;
> +#endif

why do we care about the stop class? It shouldn't, in general, consume a
lot of cycles.

> +
> +	if (unlikely(current->sched_class == &idle_sched_class))
> +		return;

And why do we care about idle? Specifically this test doesn't capture
force-idle threads. Notably see is_idle_task().

> +
> +	if (unlikely(task_has_idle_policy(current)))
> +		return;
> +
> +	if (likely(fair_policy(current->policy))) {
> +
> +		if (unlikely(current->in_iowait)) {
> +			flags |= SCHED_CPUFREQ_IOWAIT | SCHED_CPUFREQ_FORCE_UPDATE;
> +			goto force_update;
> +		}
> +
> +#ifdef CONFIG_SMP
> +		/*
> +		 * Allow cpufreq updates once for every update_load_avg() decay.
> +		 */
> +		if (unlikely(rq->cfs.decayed)) {
> +			rq->cfs.decayed = false;
> +			goto force_update;
> +		}
> +#endif
> +		return;
> +	}
> +
> +	/*
> +	 * RT and DL should always send a freq update. But we can do some
> +	 * simple checks to avoid it when we know it's not necessary.
> +	 */
> +	if (rt_task(current) && rt_task(prev)) {

IIRC dl tasks also match rt_task, so your else clause might not work the
way you've intended.

> +#ifdef CONFIG_UCLAMP_TASK
> +		unsigned long curr_uclamp_min = uclamp_eff_value(current, UCLAMP_MIN);
> +		unsigned long prev_uclamp_min = uclamp_eff_value(prev, UCLAMP_MIN);
> +
> +		if (curr_uclamp_min == prev_uclamp_min)
> +#endif
> +			return;
> +	} else if (dl_task(current) && current->dl.flags & SCHED_FLAG_SUGOV) {

Notably DL tasks also match rt_task(), so I don't think this clause
exactly does as you expect. Also, isn't the flags check sufficient on
it's own?

> +		/* Ignore sugov kthreads, they're responding to our requests */
> +		return;
> +	}
> +
> +	flags |= SCHED_CPUFREQ_FORCE_UPDATE;
> +
> +force_update:
> +	cpufreq_update_util(rq, flags);
> +#endif
> +}

But over-all the thing seems very messy, mixing sched_class, policy and
prio based selection methods.

Can't this be cleaned up somewhat?


Notably, if you structure it something like so:

	if (fair_policy(current)) {
		...
		return;
	}

	if (rt_policy(current)) {
		if (dl_task(current) && current->dl.flags & SCHED_FLAG_SUGOV)
			return;
		if (rt_policy(prev) && uclamps_match(current, prev))
			return;
		...
		return;
	}

	/* everybody else gets nothing */
	return;

You get a lot less branches in the common paths, no?