Re: [PATCH] arm64: dts: rockchip: enable built-in thermal monitoring on rk3588

[Alexey Charkov]

On Thu, Jan 18, 2024 at 10:48 PM Dragan Simic <dsimic@xxxxxxxxxxx> wrote:
>
> On 2024-01-08 14:41, Alexey Charkov wrote:
> > Hello Dragan,
>
> Hello Alexey! :)
>
> I apologize for my delayed response.  It took me almost a month to
> nearly fully recover from some really nasty flu that eventually went
> into my lungs.  It was awful and I'm still not back to my 100%. :(

Ouch, I hope you get well soon!

> > Thanks a lot for your review and comments! Some reflections below.
>
> Thank you for your work and for your detailed response.  Please see my
> comments below, which apply to your v2 submission as a well, to which
> I'll respond separately a bit later.
>
> > On Sun, Jan 7, 2024 at 2:54 AM Dragan Simic <dsimic@xxxxxxxxxxx> wrote:
> >> On 2024-01-06 23:23, Alexey Charkov wrote:
> >> > Include thermal zones information in device tree for rk3588 variants
> >> > and enable the built-in thermal sensing ADC on RADXA Rock 5B
> >> >
> >> > Signed-off-by: Alexey Charkov <alchark@xxxxxxxxx>
> >> > ---
> >> > diff --git a/arch/arm64/boot/dts/rockchip/rk3588s.dtsi
> >> > b/arch/arm64/boot/dts/rockchip/rk3588s.dtsi
> >> > index 8aa0499f9b03..8235991e3112 100644
> >> > --- a/arch/arm64/boot/dts/rockchip/rk3588s.dtsi
> >> > +++ b/arch/arm64/boot/dts/rockchip/rk3588s.dtsi
> >> > @@ -10,6 +10,7 @@
> >> >  #include <dt-bindings/reset/rockchip,rk3588-cru.h>
> >> >  #include <dt-bindings/phy/phy.h>
> >> >  #include <dt-bindings/ata/ahci.h>
> >> > +#include <dt-bindings/thermal/thermal.h>
> >> >
> >> >  / {
> >> >       compatible = "rockchip,rk3588";
> >> > @@ -2112,6 +2113,148 @@ tsadc: tsadc@fec00000 {
> >> >               status = "disabled";
> >> >       };
> >> >
> >> > +     thermal_zones: thermal-zones {
> >> > +             soc_thermal: soc-thermal {
> >>
> >> It should be better to name it cpu_thermal instead.  In the end,
> >> that's
> >> what it is.
> >
> > The TRM document says the first TSADC channel (to which this section
> > applies) measures the temperature near the center of the SoC die,
> > which implies not only the CPU but also the GPU at least. RADXA's
> > kernel for Rock 5B also has GPU passive cooling as one of the cooling
> > maps under this node (not included here, as we don't have the GPU node
> > in .dtsi just yet). So perhaps naming this one cpu_thermal could be
> > misleading?
>
> Ah, I see now, thanks for reminding;  it's all described on page 1,372
> of the first part of the RK3588 TRM v1.0.
>
> Having that in mind, I'd suggest that we end up naming it
> package_thermal.
> The temperature near the center of the chip is usually considered to be
> the overall package temperature;  for example, that's how the
> user-facing
> CPU temperatures are measured in the x86_64 world.

Sounds good, will rename in v3!

> >> > +                     trips {
> >> > +                             threshold: trip-point-0 {
> >>
> >> It should be better to name it cpu_alert0 instead, because that's what
> >> other newer dtsi files already use.
> >
> > Reflecting on your comments here and below, I'm thinking that maybe it
> > would be better to define only the critical trip point for the SoC
> > overall, and then have alerts along with the respective cooling maps
> > separately for A76-0,1, A76-2,3, A55-0,1,2,3? After all, given that we
> > have more granular temperature measurement here than in previous RK
> > chipsets it might be better to only throttle the "offending" cores,
> > not the full package.
> >
> > What do you think?
> >
> > Downstream DT doesn't follow this approach though, so maybe there's
> > something I'm missing here.
>
> I agree, it's better to fully utilize the higher measurement granularity
> made possible by having multiple temperature sensors available.
>
> I also agree that we should have only the critical trip defined for the
> package-level temperature sensor.  Let's have the separate temperature
> measurements for the CPU (sub)clusters do the thermal throttling, and
> let's keep the package-level measurement for the critical shutdowns
> only.
> IIRC, some MediaTek SoC dtsi already does exactly that.
>
> Of course, there are no reasons not to have the critical trips defined
> for the CPU (sub)clusters as well.

I think I'll also add a board-specific active cooling mechanism on the
package level in the next iteration, given that Rock 5B has a PWM fan
defined as a cooling device. That will go in the separate patch that
updates rk3588-rock-5b.dts (your feedback to v2 of this patch is also
duly noted, thank you!)

> >> > +                                     temperature = <75000>;
> >> > +                                     hysteresis = <2000>;
> >> > +                                     type = "passive";
> >> > +                             };
> >> > +                             target: trip-point-1 {
> >>
> >> It should be better to name it cpu_alert1 instead, because that's what
> >> other newer dtsi files already use.
> >>
> >> > +                                     temperature = <85000>;
> >> > +                                     hysteresis = <2000>;
> >> > +                                     type = "passive";
> >> > +                             };
> >> > +                             soc_crit: soc-crit {
> >>
> >> It should be better to name it cpu_crit instead, because that's what
> >> other newer dtsi files already use.
> >
> > Seems to me that if I define separate trips for the three groups of
> > CPU cores as mentioned above this would better stay as soc_crit, as it
> > applies to the whole die rather than the CPU cluster alone. Then
> > 'threshold' and 'target' will go altogether, and I'll have separate
> > *_alert0 and *_alert1 per CPU group.
>
> It should perhaps be the best to have "passive", "hot" and "critical"
> trips defined for all three CPU groups/(sub)clusters, separately of
> course, to have even higher granularity when it comes to the resulting
> thermal throttling.

I looked through drivers/thermal/rockchip_thermal.c, and it doesn't
seem to provide any callback for the "hot" trip as part of its struct
thermal_zone_device_ops, so I guess it would be redundant in our case
here? I couldn't find any generic mechanism to react to "hot" trips,
and they seem to be purely driver-specific, thus no-op in case of
Rockchips - or am I missing something?

> >> > +                                     hysteresis = <2000>;
> >> > +                                     type = "critical";
> >> > +                             };
> >> > +                     };
> >> > +                     cooling-maps {
> >> > +                             map0 {
> >> > +                                     trip = <&target>;
> >>
> >> Shouldn't &threshold (i.e. &cpu_alert0) be referenced here instead?
> >>
> >> > +                                     cooling-device = <&cpu_l0 THERMAL_NO_LIMIT THERMAL_NO_LIMIT>;
> >>
> >> Shouldn't all big CPU cores be listed here instead?
> >
> > I guess if a separate trip point is defined for cpu_l0,1,2,3 then it
> > would need to throttle at 75C, and then cpu_b0,1 and cpu_b2,3 at 85C
> > each. Logic being that if a sensor stacked in the middle of a group of
> > four cores shows 75C then one of the four might well be in abnormal
> > temperature range already (85+), while sensors next to only two big
> > cores each will likely show temperatures similar to the actual core
> > temperature.
>
> I think we shouldn't make any assumptions of how the CPU cores heat up
> and affect each other, because we don't really know the required
> details.
> Instead, we should simply define the reasonable values for the
> "passive",
> "hot" and "critical" trips, and leave the rest to the standard thermal
> throttling logic.  I hope you agree.
>
> In the end, that's why we have separate thermal sensors available.

Indeed! I'll add extra "passive" alerts though (at 75C) to enable the
power allocation governor to initialize its PID parameters calculation
before the control temperature setpoint gets hit (per Daniel's
feedback under separate cover).

Thanks again for your review and comments on this one!

Best regards,
Alexey