Re: [RFC PATCH] cpufreq: qcom-cpufreq-hw: allow work to be done on other CPU for PREEMPT_RT

From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
To: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Cc: Andy Gross <agross@kernel.org>,
	Bjorn Andersson <andersson@kernel.org>,
	Konrad Dybcio <konrad.dybcio@linaro.org>,
	"Rafael J. Wysocki" <rafael@kernel.org>,
	Viresh Kumar <viresh.kumar@linaro.org>,
	linux-arm-msm@vger.kernel.org, linux-pm@vger.kernel.org,
	linux-kernel@vger.kernel.org,
	Adrien Thierry <athierry@redhat.com>,
	Brian Masney <bmasney@redhat.com>,
	linux-rt-users@vger.kernel.org
Subject: Re: [RFC PATCH] cpufreq: qcom-cpufreq-hw: allow work to be done on other CPU for PREEMPT_RT
Date: Tue, 21 Mar 2023 11:04:56 +0100	[thread overview]
Message-ID: <20230321100456.0_DhhkZJ@linutronix.de> (raw)
In-Reply-To: <20230315164910.302265-1-krzysztof.kozlowski@linaro.org>

On 2023-03-15 17:49:10 [+0100], Krzysztof Kozlowski wrote:
> Qualcomm cpufreq driver configures interrupts with affinity to each
> cluster, e.g.  dcvsh-irq-0, dcvsh-irq-4 and dcvsh-irq-7 on SM8250.
> Triggered interrupt will schedule delayed work, but, since workqueue
> prefers local CPUs, it might get executed on a CPU dedicated to realtime
> tasks causing unexpected latencies in realtime workload.
> 
> Use unbound workqueue for such case.  This might come with performance
> or energy penalty, e.g. because of cache miss or when other CPU is
> sleeping.

I miss the point where it explains that only PREEMPT_RT is affected by
this.

> Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
> ---
>  drivers/cpufreq/qcom-cpufreq-hw.c | 11 ++++++++++-
>  1 file changed, 10 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/cpufreq/qcom-cpufreq-hw.c b/drivers/cpufreq/qcom-cpufreq-hw.c
> index 2f581d2d617d..c5ff8d25fabb 100644
> --- a/drivers/cpufreq/qcom-cpufreq-hw.c
> +++ b/drivers/cpufreq/qcom-cpufreq-hw.c
> @@ -390,7 +390,16 @@ static irqreturn_t qcom_lmh_dcvs_handle_irq(int irq, void *data)
>  
>  	/* Disable interrupt and enable polling */
>  	disable_irq_nosync(c_data->throttle_irq);
> -	schedule_delayed_work(&c_data->throttle_work, 0);
> +
> +	/*
> +	 * Workqueue prefers local CPUs and since interrupts have set affinity,
> +	 * the work might execute on a CPU dedicated to realtime tasks.
> +	 */
> +	if (IS_ENABLED(CONFIG_PREEMPT_RT))
> +		queue_delayed_work_on(WORK_CPU_UNBOUND, system_unbound_wq,
> +				      &c_data->throttle_work, 0);
> +	else
> +		schedule_delayed_work(&c_data->throttle_work, 0);

You isolated CPUs and use this on PREEMPT_RT. And this special use-case
is your reasoning to make this change and let it depend on PREEMPT_RT?

If you do PREEMPT_RT and you care about latency I would argue that you
either disable cpufreq and set it to PERFORMANCE so that the highest
available frequency is set once and not changed afterwards.

>  	if (qcom_cpufreq.soc_data->reg_intr_clr)
>  		writel_relaxed(GT_IRQ_STATUS,
> -- 
> 2.34.1
> 