Kconfig option for compile time build coverage (Was: Re: [PATCH] serial/efm32: add new driver)

[Arnd Bergmann]

On Monday 09 January 2012, Uwe Kleine-König wrote:
> On Fri, Dec 23, 2011 at 09:44:28PM +0100, Uwe Kleine-König wrote:
> > On Fri, Dec 23, 2011 at 10:35:22AM +0000, Arnd Bergmann wrote:
> > > On Thursday 22 December 2011, Uwe Kleine-König wrote:
> > > > [...]
> > > > +config SERIAL_EFM32_USART
> > > > + bool "EFM32 USART port."
> > > > + depends on ARCH_EFM32
> > > > + select SERIAL_CORE
> > [...]
> > > I would generally prefer not to make the driver depend on the
> > > platform, so we can get better compile time coverage. I think a better
> > > set of dependencies would be
> > > 
> > >     depends on HAVE_CLK
> > >     depends on OF
> > >     default ARCH_EFM32
> > I'd prefer something like:
> > 
> >       depends on HAVE_CLK
> >       depends on ARCH_EFM32 || I_DO_BUILD_COVERAGE_TESTING
> > 
> > This would make it easier for Joe User to pick the right options for his
> > kernel (assuming he found out to better keep I_DO_BUILD_COVERAGE_TESTING
> > disabled). [...]
>
> What do you think about this I_DO_BUILD_COVERAGE_TESTING option? It
> would allow testers to get all possible options enabled (though there
> will never be an EFM32 USART port on a non-EFM32 machine I guess) and
> still users and distribution packagers would easily keep the option off
> even without cluttering .config and {menu,n,x,whatever}config.

Sorry for the late reply. The same topic has been discussed a lot of
times. We have different ways to handle this in the kernel, but the
way most common way is to use defconfig to decide which drivers to
use on a given machine but give the user the freedom to both enable
and disable any driver whereever possible, even if that would be
a silly thing to do. What we do require normally is that anything
that can be enabled is also able to be built without errors. A lot
of drivers have dependencies on platform specific header files
or other interfaces, so they require that platform to be enabled.

For all others I would not make such a restriction and I would always
recommend to be specific in the dependencies, e.g depending on HAVE_CLK
when you use the clock interfaces, instead of depending on other options
that implicitly enable HAVE_CLK.

	Arnd