Re: [PATCH 2/6] ASoC: codecs: AK4641 depends on GPIOLIB

[Arnd Bergmann <arnd@arndb.de>]

On Sunday 02 October 2011 22:27:11 Mark Brown wrote:
> > Is there any other symbol that I can test then?
> 
> You shouldn't be testing anything - the client side GPIO API (which is
> what this driver is using) is supposed to stub itself out when not in
> use so drivers should just be able to use it without worrying about
> dependencies.  You didn't report the problem but I'd expect that
> whatever you saw will be an issue in whatever platform you were trying
> to build for (I'm guessing it hasn't provided gpio_request_one()),
> though it could be a problem in the gpiolib stubs if that's being used.

I don't remember where I first saw it. If the problem comes back,
I'll do a full bug report. I've verified now that it works on
various platforms with and without GPIOLIB.

I didn't know how the GPIO bits fit together, so I ended up doing
something that made the problem go away, whatever it was. This
is of course a problem with the randconfig fixing: One really needs
to understand every possible corner of the kernel to get it right,
and if you don't you end up with a patch that avoids the symptom
without fixing the underlying bug and then you make it harder to
find.

I really appreciate you doing the thorough review of the patches
to make sure we find the actual bugs, which is one of the main
things I want to achieve here anyway.

> > I noticed that a lot of places use 'depends on GPIOLIB' or
> > '#ifdef CONFIG_GPIOLIB', are those usually wrong, too?
> 
> Checks for gpiolib in drivers providing GPIOs are sensible, if a
> platform hasn't used gpiolib then it's generally not even got an
> interface for drivers to provide GPIOs.
> 
> On the user side these are usually due to people making the sort of
> changes you're making here due to a random build coverage issue - it
> seems unfortunately common for people to just shove a dependency in
> Kconfig when they run into a build coverage issue without looking at
> what's going on.  For a lot of the stuff you see on PCs it's going to
> make sense but for some of the "service" APIs like GPIOs that are more
> commonly used only in embedded contexts the use of the API is usually
> completely optional (eg, in this case the driver is controlling power
> and reset lines which could easily just be strapped in the hardware with
> no soft control and are supplied as optional platform data) so for many
> systems the driver is going to work completely happily without doing
> anything with GPIOs.
> 
> Adding dependencies to all the users needlessly restricts which systems
> can use the drivers.  Adding ifdefs to the drivers is repetitive and
> isn't great for legiblity, having the header stub itself out is simpler
> and easier to use on the driver side.

Ok, makes sense. Thanks for the background information!

	Arnd