Re: [PATCH] drm/amdkfd: fix build error with missing AMD_IOMMU_V2

[Christian König]

Am 08.03.21 um 21:02 schrieb Felix Kuehling:
> Am 2021-03-08 um 2:33 p.m. schrieb Arnd Bergmann:
> > On Mon, Mar 8, 2021 at 8:11 PM Felix Kuehling <felix.kuehling@xxxxxxx> wrote:
> > > Am 2021-03-08 um 2:05 p.m. schrieb Arnd Bergmann:
> > > > On Mon, Mar 8, 2021 at 5:24 PM Felix Kuehling <felix.kuehling@xxxxxxx> wrote:
> > > > > The driver build should work without IOMMUv2. In amdkfd/Makefile, we
> > > > > have this condition:
> > > > >
> > > > > ifneq ($(CONFIG_AMD_IOMMU_V2),)
> > > > > AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o
> > > > > endif
> > > > >
> > > > > In amdkfd/kfd_iommu.h we define inline stubs of the functions that are
> > > > > causing your link-failures if IOMMU_V2 is not enabled:
> > > > >
> > > > > #if defined(CONFIG_AMD_IOMMU_V2_MODULE) || defined(CONFIG_AMD_IOMMU_V2)
> > > > > ... function declarations ...
> > > > > #else
> > > > > ... stubs ...
> > > > > #endif
> > > > Right, that is the problem I tried to explain in my patch description.
> > > >
> > > > Should we just drop the 'imply' then and add a proper dependency like this?
> > > >
> > > >        depends on DRM_AMDGPU && (X86_64 || ARM64 || PPC64)
> > > >        depends on AMD_IOMMU_V2=y || DRM_AMDGPU=m
> > > >
> > > > I can send a v2 after some testing if you prefer this version.
> > > No. My point is, there should not be a hard dependency. The build should
> > > work without CONFIG_AMD_IOMMU_V2. I don't understand why it's not
> > > working for you. It looks like you're building kfd_iommu.o, which should
> > > not be happening when AMD_IOMMU_V2 is not enabled. The condition in
> > > amdkfd/Makefile should make sure that kfd_iommu.o doesn't get built with
> > > your kernel config.
> > Again, as I explained in the changelog text, AMD_IOMMU_V2 configured as
> > a loadable module, while AMDGPU is configured as built-in.
> I'm sorry, I didn't read it carefully. And I thought "imply" was meant
> to fix exactly this kind of issue.
>
> I don't want to create a hard dependency on AMD_IOMMU_V2 if I can avoid
> it, because it is only really needed for a small number of AMD APUs, and
> even there it is now optional for more recent ones.
>
> Is there a better way to avoid build failures without creating a hard
> dependency?

What you need is the same trick we used for AGP on radeon/nouveau:

depends on AMD_IOMMU_V2 || !AMD_IOMMU_V2

This way when AMD_IOMMU_V2 is build as a module DRM_AMDGPU will be build 
as a module as well. When it is disabled completely we don't care.

Regards,
Christian.

>    The documentation in
> Documentation/kbuild/kconfig-language.rst suggests using if
> (IS_REACHABLE(CONFIG_AMD_IOMMU_V2)) to guard those problematic function
> calls. I think more generally, we could guard all of kfd_iommu.c with
>
>      #if IS_REACHABLE(CONFIG_AMD_IOMMU_V2)
>
> And use the same condition to define the stubs in kfd_iommu.h.
>
> Regards,
>    Felix
>
>
> > The causes a link failure for the vmlinux file, because the linker cannot
> > resolve addresses of loadable modules at compile time -- they have
> > not been loaded yet.
> >
> >        Arnd
> > _______________________________________________
> > dri-devel mailing list
> > dri-devel@xxxxxxxxxxxxxxxxxxxxx
> > https://lists.freedesktop.org/mailman/listinfo/dri-devel
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@xxxxxxxxxxxxxxxxxxxxx
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx