Re: [PATCH] ASoC: SOF: sort out Kconfig, again

From: Arnd Bergmann
Date: Wed Apr 29 2020 - 04:25:06 EST


On Wed, Apr 29, 2020 at 1:00 AM Pierre-Louis Bossart
<pierre-louis.bossart@xxxxxxxxxxxxxxx> wrote:
> >>> Thanks Arnd, do you mind sharing your config?
> >>
> >> https://pastebin.com/HRX5xi3R
> >
> > will give it a try, thanks!
> >
> >>> We noticed last week that
> >>> there's a depend/select confusion might be simpler to fix, see
> >>> https://github.com/thesofproject/linux/pull/2047/commits
> >>>
> >>> If I look at the first line I see a IMX_DSP=n which looks exactly like
> >>> what we wanted to fix.
> >>
> >> Yes, I think that fix addresses the build warning as well, but looking
> >> more closely I don't think it's what you want: If you do this on
> >> a config that has the IMX_DSP disabled, it would appear to the
> >> user that you have enabled the drivers, but the actual code is still
> >> disabled.
> >
> > Are you sure? we added a select IMX_DSP, so not sure how it can be
> > disabled?
>
> I just tested Arnd's config with the patch we came up with for SOF
> (attached) and it makes the unmet dependency go away and builds fine.
> the problem is really using select IMX_DSP if it can be disabled by
> something else. My proposal looks simpler but I will agree it's not
> necessarily super elegant to move the dependency on IMX_BOX into SOF, so
> no sustained objection from me on Arnd's proposal.

Ok, thanks for testing!

I looked at the bigger picture again and found that the more fundamental
problem is the dependency reversal in sound/soc/sof/sof-of-dev.c, where
you have common code that knows about and links against a hardware
specific driver. This is something we try hard do avoid in general in the
kernel, as it causes all kinds of problems:

- Expressing the Kconfig dependencies is rather unnatural and error-prone,
as you found

- Adding multiple new drivers at the same time leads to merge conflicts

- A kernel that supports multiple SoC families, like all general-purpose
distros do, and Android is going to do in the future means that you have
to load every hardware specific module in order to just use one of them.

- In Android's case, it also breaks the model of having one vendor provide
support for a new SoC by enabling additional modules they ship in
their vendor specific partition

I think this is all solved by moving the "module_platform_driver()"
and of_device_id array for each driver into the module that defines
the corresponding sof_dev_desc structure, and have those drivers
call the exported sof_of_probe() and sof_of_remove() functions.

Arnd