[PATCH v2] ARM: initial multiplatform support

[Arnd Bergmann]

On Tuesday 04 September 2012, Rob Herring wrote:
> On 09/04/2012 01:56 AM, Arnd Bergmann wrote:
> > On Monday 03 September 2012, Rob Herring wrote:
> >> On 09/03/2012 11:34 AM, Arnd Bergmann wrote:
> >>>> @@ -254,27 +254,9 @@ config MMU
> >>>>  #
> >>>>  choice
> >>>>         prompt "ARM system type"
> >>>> +       depends on !ARCH_MULTIPLATFORM
> >>>>         default ARCH_VERSATILE
> >>>
> >>> Why did you move ARCH_MULTIPLATFORM out of the "choice" statement?
> >>> If we leave it in there, and make it the default, then we don't
> >>> even have to change the defconfigs any more (except the versatile
> >>> one, which is no longer the default), which I think is quite clever
> >>> and helps git-bisecting across this commit.
> >>
> >> We don't have to change and of the defconfigs or .config in this patch.
> > 
> > How does ARCH_MULTIPLATFORM get set then when building the existing
> > defconfig files for the converted platforms?
> 
> Ahhh, right. Those need to be updated. It doesn't really matter if
> ARCH_MULTIPLATFORM is in the choice or not. It matters if we move
> platforms out of the choice or not. More on that below.

But if ARCH_MULTIPLATFORM is one of the choice statements rather than
a separate option, you don't need to change any of the defconfigs,
and you can keep using 'make oldconfig' when bisecting over this
commit.
> > 
> > I was thinking of other cases actually:
> > 
> > * device drivers that we don't really care about much but that rely
> >   on a mach/* header file. In some cases it can be hard to clean up
> >   those headers, so as an intermediate step, we can just disable the
> >   drivers when building multiplatform.
> 
> depends on !ARCH_MULTIPLATFORM ?

Yes, of course, that is what you'd do in in the drivers, but it only works
if the same platform can be used both with and without ARCH_MULTIPLATFORM.

> > * At least the Samsung platforms have files that are built for separate
> >   platforms in a way that does not work across platforms: 
> >   arch/arm/plat-samsung/devs.c depends heavily on constants defined
> >   in mach/*.h header files. When building for DT-only, we can decide to
> >   just not build this file.
> 
> I still think that should be possible on a per platform basis, but why
> mandate that all platforms do that?

I wasn't suggesting we do that. The platforms that you have already
converted are all ok to stay multiplatform-only.

> We have ARCH_FOO in the choice today. Assuming we want to keep the
> choice name as is to not break configs (as Russell wanted), we have to
> leave this alone. So then we create ARCH_FOO_MULTI for the mult-platform
> case and also ARCH_FOO_OPTIONS to avoid duplicating all the per mach
> selects for ARCH_FOO and ARCH_FOO_MULTI. However, then we have ARCH_FOO
> as a depends all over the tree. I'd like to avoid having to fix-up all
> those to be depends on ARCH_FOO_OPTIONS or (ARCH_FOO || ARCH_FOO_MULTI).

I've been going back and forth a few times between the different methods
when I created the earlier version of this patch, trying to find the one
that is least invasive. As you say, creating a new ARCH_FOO_MULTI in parallel
with the existing (choice) ARCH_FOO means we have to change all the Kconfig
statements that refer to ARCH_FOO. The good news is that I think we now only
have to do it inside of the mach-foo directories, which we might want to
rework anyway so you can more easily build e.g. a THUMB2 kernel that enables
all the ARMv7 platforms but not the ARMv6 ones.

All device drivers that have a dependency on a specific platform should
fall into one of two categories:

a) Those using platform specific header files. We just decided to not allow
this in multiplatform kernels, so having a dependency on a symbol that
only exists for single-platform builds is appropriate.

b) Drivers that only make sense for one platform but can be built anywhere,
e.g. after fixing the include path from a). I'd argue that we should allow
building them anyway, to improve build coverage, so we should better remove
the dependency, or turn it into a more specific one.

	Arnd