Re: [PATCH 1/2] clocksource: tegra: Use rating when registering clock source

[Thierry Reding]

On Fri, Jun 14, 2019 at 08:00:09PM +0300, Dmitry Osipenko wrote:
> 14.06.2019 18:37, Thierry Reding ÐÐÑÐÑ:
> > On Fri, Jun 14, 2019 at 05:06:48PM +0300, Dmitry Osipenko wrote:
> >> 14.06.2019 17:02, Dmitry Osipenko ÐÐÑÐÑ:
> >>> 14.06.2019 16:53, Thierry Reding ÐÐÑÐÑ:
> >>>> On Fri, Jun 14, 2019 at 04:29:17PM +0300, Dmitry Osipenko wrote:
> >>>>> 14.06.2019 16:22, Thierry Reding ÐÐÑÐÑ:
> >>>>>> On Fri, Jun 14, 2019 at 03:24:07PM +0300, Dmitry Osipenko wrote:
> >>>>>>> 14.06.2019 13:47, Thierry Reding ÐÐÑÐÑ:
> >>>>>>>> From: Thierry Reding <treding@xxxxxxxxxx>
> >>>>>>>>
> >>>>>>>> The rating is parameterized depending on SoC generation to make sure it
> >>>>>>>> takes precedence on implementations where the architected timer can't be
> >>>>>>>> used. This rating is already used for the clock event device. Use the
> >>>>>>>> same rating for the clock source to be consistent.
> >>>>>>>>
> >>>>>>>> Signed-off-by: Thierry Reding <treding@xxxxxxxxxx>
> >>>>>>>> ---
> >>>>>>>>  drivers/clocksource/timer-tegra.c | 2 +-
> >>>>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
> >>>>>>>>
> >>>>>>>> diff --git a/drivers/clocksource/timer-tegra.c b/drivers/clocksource/timer-tegra.c
> >>>>>>>> index f6a8eb0d7322..e6608141cccb 100644
> >>>>>>>> --- a/drivers/clocksource/timer-tegra.c
> >>>>>>>> +++ b/drivers/clocksource/timer-tegra.c
> >>>>>>>> @@ -318,7 +318,7 @@ static int __init tegra_init_timer(struct device_node *np, bool tegra20,
> >>>>>>>>  	sched_clock_register(tegra_read_sched_clock, 32, TIMER_1MHz);
> >>>>>>>>  
> >>>>>>>>  	ret = clocksource_mmio_init(timer_reg_base + TIMERUS_CNTR_1US,
> >>>>>>>> -				    "timer_us", TIMER_1MHz, 300, 32,
> >>>>>>>> +				    "timer_us", TIMER_1MHz, rating, 32,
> >>>>>>>>  				    clocksource_mmio_readl_up);
> >>>>>>>>  	if (ret)
> >>>>>>>>  		pr_err("failed to register clocksource: %d\n", ret);
> >>>>>>>>
> >>>>>>>
> >>>>>>> Looks good. Although, could you please clarify whether arch-timer stops on T210 when CPU
> >>>>>>> enters deepest (powerdown) idle state? I'm starting to lose track a bit already. Because
> >>>>>>> if arch-timer stops in the deepest idle state, then it's a bit odd that Joseph didn't add
> >>>>>>> the clocksource for T210 in the first place and v5.1 probably shouldn't work well because
> >>>>>>> of that already.
> >>>>>>
> >>>>>> Yes, the architected timer doesn't work across an SC7 (which is what the
> >>>>>> deepest idle state is called on Tegra210) transition, hence why we can't
> >>>>>> use it as a suspend clocksource. I actually sent out a patch to do that,
> >>>>>> earlier.
> >>>>>>
> >>>>>> And yes, it's entirely possible that v5.1 doesn't work in this regard,
> >>>>>> but we're not noticing that because we don't have suspend/resume support
> >>>>>> for Tegra210 anyway. There are a couple of missing pieces that we need
> >>>>>> in order to make it work.
> >>>>>>
> >>>>>> This change in particular is only going to affect the CPU idle state
> >>>>>> (CC7). Since the architected timer doesn't survive that either, we need
> >>>>>> the Tegra timer to be preferred over the architected timer for normal
> >>>>>> operation.
> >>>>>>
> >>>>>> All of these issues go away on Tegra186 and later, where the architected
> >>>>>> timer is in an always-on partition and has a PLL that remains on during
> >>>>>> SC7 (and CC7).
> >>>>>
> >>>>> Thank you very much for the clarification. But then what about the
> >>>>> sched_clock? I suppose sched_clock will suffer on T210 as well and it's
> >>>>> a bit trickier case because apparently arch-timer always wins since it
> >>>>> has a higher precision. I guess the best solution will be to just bail
> >>>>> out from arch-timer's driver probe in a case of T210.
> >>>>>
> >>>>> if (of_machine_is_compatible("nvidia,tegra210"))
> >>>>> 	return 0.
> >>>>
> >>>> I don't think there's any issue with the scheduler clock on Tegra210.
> >>>> Before the CPU can be turned off, all tasks scheduled on that CPU have
> >>>> to be migrated to another CPU, right? Conversely, before any tasks can
> >>>> be scheduled on a CPU that CPU needs to be brought online, at which
> >>>> point the architected timer should work fine again.
> >>>
> >>> Is SC7 a CPU-idle state that cpuidle driver may enter or it's a
> >>> system-wide suspend state? It's still not clear to me.
> >>>
> >>
> >> Ah, looks like I see now. So CC7 (CPU idle state) also affects the
> >> arch-timer (like SC7) and hence scheduler clock will be stopped while it
> >> shouldn't, which doesn't sound good.
> > 
> > We enable CC7 on Jetson TX1 and I've just verified on Jetson Nano that
> > there are no issues if CC7 is enabled. From the boot log I can see that
> > the architected timer is still used as scheduler clock.
> > 
> > So that either means that the scheduler doesn't mind if the clock is
> > disabled when a CPU is asleep or it means that CC7 does not impact the
> > architected timer. I thought we had already confirmed that the latter
> > isn't true, i.e. that the architected timer is disabled during CC7, so
> > that would mean that indeed the scheduler continues to work fine if the
> > clock is off during sleep. I also don't understand why it would break,
> > given that it's only put to sleep when there are no longer any tasks
> > running on it.
> 
> CPU may enter idling state while task is sleeping, i.e. waiting for some event. To be
> honest, I don't know much about how scheduling actually works in the kernel and what
> are the actual purposes of scheduler clock. Maybe Daniel could clarify it all for us?

My understanding is that a task that goes to sleep can be migrated to
another CPU, so there's got to be code in place already to account for
that. Migrating from one CPU to another is not very different from
resuming a task on the same CPU.

But yeah, I'm not very firm on the fundamentals here either, so a bit of
clarification from Daniel or Thomas would help.

Thierry

Attachment: signature.asc
Description: PGP signature