Re: [PATCH 4/8] thermal/drivers/Kconfig: Convert the CPU cooling device to a choice

[Daniel Thompson]

On Thu, Jan 25, 2018 at 02:36:12PM +0100, Daniel Lezcano wrote:
> On 25/01/2018 11:57, Daniel Thompson wrote:
> > On Wed, Jan 24, 2018 at 05:59:09PM +0100, Daniel Lezcano wrote:
> >> On 24/01/2018 17:34, Daniel Thompson wrote:
> >> However, we are talking about distros here but the CPU cooling mechanism
> >> is for mobile and in this case the kernel (usually Android based) comes
> >> with a specific configuration file and this is where the SoC vendor has
> >> to choose the right strategy.
> > 
> > I agree its hard to conceive of a modern Android device that doesn't implement
> > both the needed features but the performance figures in the covering 
> > letter come from Hikey (and they look pretty good) and Hikey is 
> > supported by a good number of regular Linux distros now.
> > 
> > Using Kconfig to force what should be runtime decisions to happen at 
> > compile time means that Hikey becomes an example of a platform that
> > is unable to run at max performance on the distros that have added 
> > support for it.
> 
> I disagree. The ARM64's defconfig is not distro ready. We have always to
> change the default option and fix things in the Kconfig to make the
> hikey to work (eg. the missing hi655x clock config), without speaking
> about the hikey960 which is not yet ready for full support.
> 
> Hence, tweaking the Kconfig to choose the better strategy is not
> necessarily a problem for this first iteration of code.

The problem is not about tweaking config for a distro. No distro I know 
defconfig kernels. Tweaking is normal and expected.

The problem is that a distro cannot generate a config that performs
optimally on all devices that it supports because enabling one form of
CPU cooling prevents other forms from being selected. There is no
correct value for us to select if we don't know in advance what board
the kernel image will be loaded on.

Given the massive work that has gone on (especially in ARM ecosystem) 
to allow one kernel binary to support many boards at once it surprises 
me to see new features proposed that cannot be exploited without
recompiling the kernel from platform to platform.


> Note I'm not against changing the code to make it runtime selectable but
> that will need a major rework of the current CPU cooling code especially
> handling the change while the thermal framework is doing the mitigation
> (and probably also changes of the thermal framework).

Perhaps I should have distinguished more between runtime-meaning-boot-time
and runtime-meaning-full-system-operation. To be clear none of my
comments are about being able to enable/disable idle injection on a
fully running system. It is the impact on multi-platform binaries that
attracted my attention.


> I prefer to keep simple self-encapsulated feature code and make it
> evolve to something better instead of sending a blowing patch series
> taking into account all possible combinations. Choosing the strategy at
> compile time may be look restrictive but we can live with that without
> problem and iteratively do the change until the choice becomes the
> default strategy selection option.

You won't find me arguing against iterative improvement. However I
did observe that the idle injection code consists almost exclusively 
of new lines of code. Thus I don't understand why this new code 
interferes with cpufreq cooling to the point that the existing cooling
options must compiled out of the kernel before we can exploit it.

I can see the combo code does have tentacles in more places but even so
I have the same question. What prevents the existing cooling strategy
from being compiled at the same time.

You appear to be saying that there's not yet enough infrastructure to
decide which type of cooler to instantiate[1] so its OK to hack around
the decision by forcing it to be made it at compile time. Even if we
want to force a decision by some means, is compile time really the best
point to do that?


Daniel.


[1] Is that because the DT isn't rich enough for us to figure out 
    the trade off between using real OPPs and virtual idle-injected
    ones nor whether cpuidle entry/exit is fast enough for cooling
    to be effective?

> 
> 
> 
> 
> 
> 
>  <http://www.linaro.org/> Linaro.org â Open source software for ARM SoCs
> 
> Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
> <http://twitter.com/#!/linaroorg> Twitter |
> <http://www.linaro.org/linaro-blog/> Blog
>