Re: [PATCH] arm64: rk3399: Add capacity-dmips-mhz attributes

[Marc Zyngier <marc.zyngier@arm.com>]

On Sun, 24 Feb 2019 20:07:49 +0000
Robin Murphy <robin.murphy@arm.com> wrote:

Hi Robin,

> Hi Marc,
> 
> On 2019-02-24 2:35 pm, Marc Zyngier wrote:
> > The RK3399 has the interesting property to be a so called "big-little"
> > system, where not all the CPUs are equal (the A53s are much weaker
> > than the A72s).
> > 
> > So far, we're not telling the OS that there is such a difference in
> > processing capacity, and Linux assumes that they are equal. Too bad.
> > 
> > Let's tell the OS about this by using the capacity-dmips-mhz
> > property. The values used here are those used on the Juno platform,
> > which is quite similar. This leads to the scheduler knowing that
> > it can pack more tasks on the A72s, and leads to a better interactive
> > experience.
> > 
> > Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
> > ---
> >   arch/arm64/boot/dts/rockchip/rk3399.dtsi | 7 ++++++-
> >   1 file changed, 6 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/arm64/boot/dts/rockchip/rk3399.dtsi b/arch/arm64/boot/dts/rockchip/rk3399.dtsi
> > index 6cc1c9fa4ea6..7eab31184aee 100644
> > --- a/arch/arm64/boot/dts/rockchip/rk3399.dtsi
> > +++ b/arch/arm64/boot/dts/rockchip/rk3399.dtsi
> > @@ -75,6 +75,7 @@
> >   			#cooling-cells = <2>; /* min followed by max */
> >   			dynamic-power-coefficient = <100>;
> >   			cpu-idle-states = <&CPU_SLEEP &CLUSTER_SLEEP>;
> > +			capacity-dmips-mhz = <578>;  
> 
> Probably better to use the value from Juno r2 (485) which should more
> appropriately represent A53 vs. A72, rather than A53 vs. A57.

Indeed. Although I doubt that there is such a difference between A57
and A72, but I guess it won't really show.

> >   		};  
> >   >   		cpu_l1: cpu@1 {  
> > @@ -86,6 +87,7 @@
> >   			#cooling-cells = <2>; /* min followed by
> > max */ dynamic-power-coefficient = <100>;
> >   			cpu-idle-states = <&CPU_SLEEP
> > &CLUSTER_SLEEP>;
> > +			capacity-dmips-mhz = <578>;
> >   		};  
> >   >   		cpu_l2: cpu@2 {  
> > @@ -97,6 +99,7 @@
> >   			#cooling-cells = <2>; /* min followed by
> > max */ dynamic-power-coefficient = <100>;
> >   			cpu-idle-states = <&CPU_SLEEP
> > &CLUSTER_SLEEP>;
> > +			capacity-dmips-mhz = <578>;
> >   		};  
> >   >   		cpu_l3: cpu@3 {  
> > @@ -108,6 +111,7 @@
> >   			#cooling-cells = <2>; /* min followed by
> > max */ dynamic-power-coefficient = <100>;
> >   			cpu-idle-states = <&CPU_SLEEP
> > &CLUSTER_SLEEP>;
> > +			capacity-dmips-mhz = <578>;
> >   		};  
> >   >   		cpu_b0: cpu@100 {  
> > @@ -118,7 +122,7 @@
> >   			clocks = <&cru ARMCLKB>;
> >   			#cooling-cells = <2>; /* min followed by
> > max */ dynamic-power-coefficient = <436>;
> > -			cpu-idle-states = <&CPU_SLEEP
> > &CLUSTER_SLEEP>;  
> 
> Oops.

Gahhhh! Rebase fail. What an idiot. Thanks for spotting this!

> 
> Otherwise, though, I was on the verge of tackling this myself now
> that my 3399 board is building its own kernels - annoyance at the
> linking bottleneck steadfastly refusing to move off CPU0 (except
> occasionally to another little core) has been growing steadily :)

Yeah, tell me about it. I still regularly build the kernel on my kevin,
and while the compilation time is reasonable for this kind of device
(and the storage media being an sdcard), the linking is just a pig.
Even on the A72.

> With the nits fixed,
> 
> Tested-by: Robin Murphy <robin.murphy@arm.com>

Cheers Robin. I'll post a v2 after -rc1.

	M.
-- 
Without deviation from the norm, progress is not possible.