[PATCH] drm: Hack around CONFIG_AGP=m build failures

[Michal Marek]

On 2015-10-01 12:17, Daniel Vetter wrote:
> On Thu, Oct 01, 2015 at 11:58:32AM +0200, Michal Marek wrote:
>> On 2015-10-01 10:19, Jani Nikula wrote:
>>> On Thu, 01 Oct 2015, Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
>>>> Surprisingly kbuild can't cope with tristates in the
>>>> <module>-$(CONFIG_FOO) pattern. This patch hacks up a solution.
>>>
>>> Given that it's two distinct Makefile variables (foo-y and foo-m) being
>>> assigned to, I don't really find this surprising. Maybe this could be
>>> made to work as a convenience, but there might be other, more surpising
>>> consequences.
>>
>> I actually think that kbuild should be able to handle this. The likely
>> reason why it is not doing it right now is that in an ideal world,
>> modules are modules can be built out of tree against just the kernel and
>> their static dependencies. In real world, we sometimes have features in
>> modules that are enabled if other modules are enabled. I'll post a patch
>> later. We also have lots of tests fo CONFIG_FOO || CONFIG_FOO_MODULE in
>> built-in code, which is a similar case.
> 
> Cool. I'll keep this hack in drm-misc then. Please cc me on the proper
> solution so I know when I can revert it again.

It's not as trivial as it seemed, because there are at least three
Makefiles that rely on the current behavior:

init/Makefile
drivers/misc/ibmasm/Makefile
fs/logfs/Makefile

While ibmasm and logfs can and probably should be fixed to work with
modular 8250 and mtd, respectively, init/Makefile would need a
workaround to only pick up do_mounts_rd.o and do_mounts_md.o if the
respective block drivers are built-in. So we would be trading one hack
for another.

Michal