Re: [PATCH v2 1/2] clocksource/drivers/exynos_mct_v2: introduce Exynos MCT version 2 driver for next Exynos SoC

[Youngmin Nam]

On Wed, Nov 03, 2021 at 10:04:36AM +0100, Krzysztof Kozlowski wrote:
> On 03/11/2021 10:24, Youngmin Nam wrote:
> > On Wed, Nov 03, 2021 at 09:18:07AM +0100, Krzysztof Kozlowski wrote:
> >> On 03/11/2021 01:09, Youngmin Nam wrote:
> >>> On Tue, Nov 02, 2021 at 10:28:10AM +0000, Mark Rutland wrote:
> >>>> On Tue, Nov 02, 2021 at 09:11:21AM +0900, Youngmin Nam wrote:
> >>>>> Exynos MCT version 2 is composed of 1 FRC and 12 comparators.
> >>>>> There are no global timer and local timer anymore.
> >>>>> The 1 of 64bit FRC serves as "up-counter"(not "comparators").
> >>>>> The 12 comaprators(not "counter") can be used as per-cpu event timer
> >>>>> so that it can support upto 12 cores.
> >>>>> And a RTC source can be used as backup clock source.
> >>>>
> >>>> [...]
> >>>>
> >>>>> +static int exynos_mct_starting_cpu(unsigned int cpu)
> >>>>> +{
> >>>>> +	struct mct_clock_event_device *mevt = per_cpu_ptr(&percpu_mct_tick, cpu);
> >>>>> +	struct clock_event_device *evt = &mevt->evt;
> >>>>> +
> >>>>> +	snprintf(mevt->name, sizeof(mevt->name), "mct_comp%d", cpu);
> >>>>> +
> >>>>> +	evt->name = mevt->name;
> >>>>> +	evt->cpumask = cpumask_of(cpu);
> >>>>> +	evt->set_next_event = exynos_comp_set_next_event;
> >>>>> +	evt->set_state_periodic = mct_set_state_periodic;
> >>>>> +	evt->set_state_shutdown = mct_set_state_shutdown;
> >>>>> +	evt->set_state_oneshot = mct_set_state_shutdown;
> >>>>> +	evt->set_state_oneshot_stopped = mct_set_state_shutdown;
> >>>>> +	evt->tick_resume = mct_set_state_shutdown;
> >>>>> +	evt->features = CLOCK_EVT_FEAT_PERIODIC | CLOCK_EVT_FEAT_ONESHOT;
> >>>>> +	evt->rating = 500;	/* use value higher than ARM arch timer */
> >>>>
> >>>> Previously Will asked you to try CLOCK_EVT_FEAT_PERCPU here, and to set
> >>>> the C3STOP flag on the arch timer via the DT when necessary, rather than
> >>>> trying to override the arch timer like this:
> >>>>
> >>>>   https://protect2.fireeye.com/v1/url?k=72526080-2dc9598b-7253ebcf-002590f5b904-ca603717c6462908&q=1&e=be56aa83-dbac-4639-913d-d388620fe3fc&u=https%3A%2F%2Flore.kernel.org%2Fr%2F20211027073458.GA22231%40willie-the-truck
> >>>>
> >>>> There are a bunch of things that depend on the architected timer working
> >>>> as a clocksource (e.g. vdso, kvm), and it *should* work as a lock
> >>>> clockevent_device if configured correctly, and it's much more consistent
> >>>> with *everyone else* to use the arhcitected timer by default.
> >>>>
> >>>> Please try as Will suggested above, so that this works from day one.
> >>>>
> >>>> Thanks,
> >>>> Mark.
> >>>>
> >>>
> >>> Hi Mark.
> >>> It looks like you missed my previous mail.
> >>> https://protect2.fireeye.com/v1/url?k=ab15817a-cbf71c27-ab140a35-000babd9f1ba-123b7f313b1b1ccc&q=1&e=34c8716e-6d2e-4d8e-82fe-04777ebc5eb3&u=https%3A%2F%2Flore.kernel.org%2Fall%2F20211029035422.GA30523%40perf%2F%23t
> >>>
> >>> Yes, I believe Will's suggestion definitely will work.
> >>> But that is for performance not functionality.
> >>> As a driver for new H/W IP I would like to confirm functionality first.
> >>>
> >>> We need more time to test this feature with our exynos core power down feature.
> >>> And we need to do a various regression test whether there is another corner case or not.
> >>> So, how about we apply Will's suggetion later after the current patchset is merged first?
> >>> After doing our regression test with our exynos core power down feature, we can confirm this.
> >>>
> >>
> >> Not really, because once it is merged there is no incentive to fix it or
> >> simply changing it can be forgotten. Also similarly to commit
> >> 6282edb72bed ("clocksource/drivers/exynos_mct: Increase priority over
> >> ARM arch timer"), there should be a valid and serious reason to
> >> prioritize Exynos MCT.
> >>
> > 
> > No, it's not. I also want to decrease MCTv2 timer rating so that we want to use arm arch timer as a default.
> > But this feature has to be confirmed with core power down feature enabled.
> > Without core power down feature, we can't comfirm this.
> > Ater that we need to check whether there is regression or not related power, stability, and so on.
> > I'm not saying I will not apply Will's suggestion but I just want to apply later after some hard test.
> > 
> 
> You repeat the same argument, the same words. Nothing new. Repeating the
> same won't change it, use the lower priority. This is a patch for new
> kernel, so there is a plenty of time to test it and it won't affect your
> production environment.
> 
So, how about we control timer rating value with DT ?
Of course the default rating value should be lower than arm arch timer's.
Do you agree with this?

> Best regards,
> Krzysztof
>