Re: [PATCH 00/14] Modularize schedutil

[Rafael J. Wysocki]

On Tue, May 12, 2020 at 5:11 PM Quentin Perret <qperret@xxxxxxxxxx> wrote:
>
> On Tuesday 12 May 2020 at 16:08:56 (+0200), Rafael J. Wysocki wrote:
> > If some piece of kernel code is modular, it still needs to be build.
> > The difference is when and how it gets loaded, so can you possibly
> > elaborate here?
>
> Sure thing, sorry if that wasn't clear.

No worries.

> The end goal with GKI is the following: Google will release a single
> binary kernel image (signed, etc etc) that all devices using a given
> Android version will be required to use. That image is however going to
> be only for the core of the kernel (no drivers or anything of the sort).
> Vendors and OEMs, on their end, will be responsible to build and ship
> GKI-compatible modules for their respective devices. So, Android devices
> will eventually ship with a Google-issued GKI, plus a bunch of
> vendor-provided modules loaded during boot.

If that is the case, then I absolutely think that schedutil should be
part of the GKI.

Moreover, that would have been my opinion even if it had been modular
in the first place.

> This is a significant shift from the current model where vendors
> completely own the kernel, and are largely free to use the kernel config
> they want. Today, those who don't use schedutil are free to turn the
> config off, for example.

So why is this regarded as a good thing?

> But GKI changes that. The 'core' GKI config is effectively imposed to
> the entire ecosystem. As of now, because it is 'bool' we have no choice
> but to compile schedutil in the core GKI as some (most) partners use it.
> But as you can imagine, that is not the preferred option of those who
> _don't_ use schedutil.

OTOH, it may as well be an incentive for them to switch over and
report problems with it that they see.

I absolutely would like to make schedutil the clearly preferred option
and IMO avoiding to use it, especially for non-technical reasons,
should be clearly less attractive.

> Modularizing avoids any potential friction since
> the vendors who want to use it will be able load the module, and the
> others will simply not. That really is the reason for that series.

If the long-term target is for everyone to use schedutil, then I don't
quite see why making it easy to not include it in one's system is
going to help.

> Then there is an important question: why should upstream care about all
> that stuff? That's obviously debatable, but my biased opinion is that
> GKI is a good thing(TM). It's our opportunity to put some order in the
> android ecosystem and to reduce the delta with mainline. That'll
> definitely take time, and there will be Android-specific churn in GKI in
> the beginning, but we'd like to keep that as small as possible, and to
> converge to 0 looking forwards.

That's a good goal, but I'm not sure if the least resistance path to
it is the right one. :-)

Cheers!