Re: [PATCH v8 1/3] cpufreq: mediatek-hw: Add support for CPUFREQ HW

[Lukasz Luba]

On 11/19/20 1:40 PM, Hector Yuan wrote:
> On Thu, 2020-11-19 at 12:41 +0000, Lukasz Luba wrote:
> > Hi Hector,
> >
> > On 10/26/20 8:19 AM, Hector Yuan wrote:
> > > From: "Hector.Yuan" <hector.yuan@xxxxxxxxxxxx>
> > >
> > > Add cpufreq HW support.
> > >
> > > Signed-off-by: Hector.Yuan <hector.yuan@xxxxxxxxxxxx>
> >
> > [snip]
> >
> > > +
> > > +static int mtk_cpufreq_hw_cpu_init(struct cpufreq_policy *policy)
> > > +{
> > > +	struct cpufreq_mtk *c;
> > > +	struct device *cpu_dev;
> > > +	struct em_data_callback em_cb = EM_DATA_CB(mtk_cpufreq_get_cpu_power);
> > > +	struct pm_qos_request *qos_request;
> > > +	int sig, pwr_hw = CPUFREQ_HW_STATUS | SVS_HW_STATUS;
> > > +
> > > +	qos_request = kzalloc(sizeof(*qos_request), GFP_KERNEL);
> > > +	if (!qos_request)
> > > +		return -ENOMEM;
> > > +
> > > +	cpu_dev = get_cpu_device(policy->cpu);
> > > +	if (!cpu_dev) {
> > > +		pr_err("failed to get cpu%d device\n", policy->cpu);
> > > +		return -ENODEV;
> > > +	}
> > > +
> > > +	c = mtk_freq_domain_map[policy->cpu];
> > > +	if (!c) {
> > > +		pr_err("No scaling support for CPU%d\n", policy->cpu);
> > > +		return -ENODEV;
> > > +	}
> > > +
> > > +	cpumask_copy(policy->cpus, &c->related_cpus);
> > > +
> > > +	policy->freq_table = c->table;
> > > +	policy->driver_data = c;
> >
> > To control frequency transition rate in schedutil, you might
> > be interested in setting:
> >
> > policy->cpuinfo.transition_latency = <mtk_value_here>;
> >
> > Example, when this latency value comes from FW [1]
> OK, I will add it in v9.
> > > +
> > > +	/* Let CPUs leave idle-off state for SVS CPU initializing */
> > > +	cpu_latency_qos_add_request(qos_request, 0);
> > > +
> > > +	/* HW should be in enabled state to proceed now */
> > > +	writel_relaxed(0x1, c->reg_bases[REG_FREQ_ENABLE]);
> > > +
> > > +	if (readl_poll_timeout(c->reg_bases[REG_FREQ_HW_STATE], sig,
> > > +			       (sig & pwr_hw) == pwr_hw, POLL_USEC,
> > > +			       TIMEOUT_USEC)) {
> > > +		if (!(sig & CPUFREQ_HW_STATUS)) {
> > > +			pr_info("cpufreq hardware of CPU%d is not enabled\n",
> > > +				policy->cpu);
> > > +			return -ENODEV;
> > > +		}
> > > +
> > > +		pr_info("SVS of CPU%d is not enabled\n", policy->cpu);
> > > +	}
> > > +
> > > +	em_dev_register_perf_domain(cpu_dev, c->nr_opp, &em_cb, policy->cpus);
> >
> > Please keep in mind that this is going to be changed soon with a new
> > argument: 'milliwatts'. It's queued in pm/linux-next [2].
> OK, thanks for the remind.
> > Regards,
> > Lukasz
> >
> > [1]
> > https://elixir.bootlin.com/linux/latest/source/drivers/cpufreq/scmi-cpufreq.c#L194
> > [2]
> > https://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm.git/commit/?h=linux-next&id=c250d50fe2ce627ca9805d9c8ac11cbbf922a4a6

Also, based on function mtk_cpufreq_hw_target_index(), which looks
really simple, you might consider to have fast_switch enabled.

It will allow SchedUtil governor to change frequency directly
and not create a dedicated deadline thread for it. It pays off.

You have to experiment with something like:

policy->fast_switch_possible  = true;


static struct cpufreq_driver cpufreq_mtk_hw_driver = {
...
	.fast_switch = mtk_cpufreq_hw_fast_switch
...
}

Again, scmi-cpufreq.c would be a good pattern to follow.