Re: [PATCH v1 4/6] arm64: dts: exynos: gs101: Add 'local-timer-stop' to cpuidle nodes

[William McVicker <willmcvicker@google.com>]

On 04/03/2025, Youngmin Nam wrote:
> On Wed, Apr 02, 2025 at 02:59:31PM -0700, William McVicker wrote:
> > Hi Youngmin,
> > 
> > On 04/02/2025, Youngmin Nam wrote:
> > > On Mon, Mar 31, 2025 at 04:00:26PM -0700, Will McVicker wrote:
> > > > From: Will Deacon <willdeacon@google.com>
> > > > 
> > > > In preparation for switching to the architected timer as the primary
> > > > clockevents device, mark the cpuidle nodes with the 'local-timer-stop'
> > > > property to indicate that an alternative clockevents device must be
> > > > used for waking up from the "c2" idle state.
> > > > 
> > > > Signed-off-by: Will Deacon <willdeacon@google.com>
> > > > [Original commit from https://android.googlesource.com/kernel/gs/+/a896fd98638047989513d05556faebd28a62b27c]
> > > > Signed-off-by: Will McVicker <willmcvicker@google.com>
> > > > ---
> > > >  arch/arm64/boot/dts/exynos/google/gs101.dtsi | 3 +++
> > > >  1 file changed, 3 insertions(+)
> > > > 
> > > > diff --git a/arch/arm64/boot/dts/exynos/google/gs101.dtsi b/arch/arm64/boot/dts/exynos/google/gs101.dtsi
> > > > index 3de3a758f113..fd0badf24e6f 100644
> > > > --- a/arch/arm64/boot/dts/exynos/google/gs101.dtsi
> > > > +++ b/arch/arm64/boot/dts/exynos/google/gs101.dtsi
> > > > @@ -155,6 +155,7 @@ ananke_cpu_sleep: cpu-ananke-sleep {
> > > >  				idle-state-name = "c2";
> > > >  				compatible = "arm,idle-state";
> > > >  				arm,psci-suspend-param = <0x0010000>;
> > > > +				local-timer-stop;
> > > >  				entry-latency-us = <70>;
> > > >  				exit-latency-us = <160>;
> > > >  				min-residency-us = <2000>;
> > > > @@ -164,6 +165,7 @@ enyo_cpu_sleep: cpu-enyo-sleep {
> > > >  				idle-state-name = "c2";
> > > >  				compatible = "arm,idle-state";
> > > >  				arm,psci-suspend-param = <0x0010000>;
> > > > +				local-timer-stop;
> > > >  				entry-latency-us = <150>;
> > > >  				exit-latency-us = <190>;
> > > >  				min-residency-us = <2500>;
> > > > @@ -173,6 +175,7 @@ hera_cpu_sleep: cpu-hera-sleep {
> > > >  				idle-state-name = "c2";
> > > >  				compatible = "arm,idle-state";
> > > >  				arm,psci-suspend-param = <0x0010000>;
> > > > +				local-timer-stop;
> > > >  				entry-latency-us = <235>;
> > > >  				exit-latency-us = <220>;
> > > >  				min-residency-us = <3500>;
> > > > -- 
> > > > 2.49.0.472.ge94155a9ec-goog
> > > > 
> > > Hi Will.
> > > 
> > > Are you using this property in production?
> > > If so, have you noticed any performance improvements?
> > 
> > On Pixel 6, I have only recently switched to using the arch_timer as the
> > default clocksource. I haven't noticed any major perf improvements to the main
> > benchmarks, but also haven't seen any regressions. Based on the ChromeOS perf
> > analysis in [1,2], there was a significant perf difference found.
> > 
> > [1] https://lore.kernel.org/linux-samsung-soc/CAJFHJrrgWGc4XGQB0ysLufAg3Wouz-aYXu97Sy2Kp=HzK+akVQ@mail.gmail.com/
> > [2] https://lore.kernel.org/linux-samsung-soc/CAASgrz2Nr69tpfC8ka9gbs2OvjLEGsvgAj4vBCFxhsamuFum7w@mail.gmail.com/
> > 
> > If it helps, I found that Pixel 8 and 9 devices (didn't check Pixel 7)
> > are already using the arch_timer with this 'local-timer-stop' as the default
> > clocksource in the production kernel.
> > 
> > Thanks,
> > Will
> > 
> > [...]
> > 
> 
> Hi Will,
> 
> Thanks for sharing the status of Pixel devices.
> 
> I agree that using the arch_timer as a clock source device brings significant benefits.
> The links you shared are definitely related to that.
> 
> However, I would also like to know whether arch_timer is used as a clock event device in Pixel production.

For Pixel 8 and 9, the arch_timer is used as both the clocksource and
clockevent device which is what my series is proposing for Pixel 6 upstream.
The MCT device is solely being used as the alternative clockevents device for
waking up from the "c2" state. The reason for using the arch_timer as the
clockevents device is because we were seeing hrtimer stability issues where
a 10ms interval timer would delay about 300ms-1s before starting the callback.
This resulted in several media-related latency issues.

Thanks,
Will

[...]