Re: [PATCH 00/14] Modularize schedutil

[Valentin Schneider]

On 08/05/20 14:15, Quentin Perret wrote:
> Hey Valentin,
>
> On Thursday 07 May 2020 at 22:34:17 (+0100), Valentin Schneider wrote:
>> I'm curious; why would some Android device not want to roll with schedutil?
>>
>> When it comes to dynamic policies (i.e. forget performance / powersave, and
>> put userspace in a corner), I'd be willing to take a stand and say you
>> should only really be using schedutil nowadays - alignment with the
>> scheduler, uclamp, yadda yadda.
>>
>> AFAIA the only schedutil-related quirk we oughta fix for arm/arm64 is that
>> arch_scale_freq_invariant() thingie, and FWIW I'm hoping to get something
>> regarding this out sometime soonish. After that, I'd actually want to make
>> schedutil the default governor for arm/arm64.
>
> As in setting CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL=y in the arm64
> defconfig? If so, you have my Acked-by already :)
>

I'm actually thinking of making it the unconditional default for arm/arm64
in cpufreq's Kconfig, following what has been recently done for
intel_pstate.

>> I'm not opiniated on the modularization, but if you can, could you please
>> share some more details as to why schedutil cannot fulfill its role of holy
>> messiah of governors for GKI?
>
> I guess I answered some of that in the other thread with Peter, but all
> in all I'm definitely not trying to make an argument that schedutil
> isn't good enough here. I'm trying to say that mandating it in *GKI* is
> just likely to cause unnecessary friction, and trust me there is already
> enough of that with other topics.

Right, I appreciate it must be an "interesting" tug of war. My own opinion
has also already been expanded in the rest of the thread; i.e. we should
strive to make schedutil good enough that folks don't feel like they still
need to use ondemand/whatever frankengov. That said, even without GKI, I
get that making some vendors change their already tested-and-tuned setup is
an obstacle course in and of itself.

> Giving the option of having sugov as a
> module doesn't prevent us from making it a default for a few arches, so
> I think there is ground for an agreement!
>
> Cheers,
> Quentin