Re: [PATCH] iommufd: fix selftest link dependency on amdv1 driver

[Jason Gunthorpe]

On Wed, Dec 24, 2025 at 11:17:13AM +0100, Arnd Bergmann wrote:
> On Wed, Dec 24, 2025, at 02:52, Jason Gunthorpe wrote:
> > On Tue, Dec 23, 2025 at 10:51:22PM +0100, Arnd Bergmann wrote:
> >> From: Arnd Bergmann <arnd@xxxxxxxx>
> >> 
> >> When CONFIG_AMD_IOMMU is disabled, the selftest runs into a link failure:
> >> 
> >> aarch64-linux-ld: drivers/iommu/iommufd/selftest.o: in function `mock_domain_alloc_pgtable.isra.0':
> >> selftest.c:(.text+0x1590): undefined reference to `pt_iommu_amdv1_init'
> >> aarch64-linux-ld: drivers/iommu/iommufd/selftest.o:(.rodata+0x280): undefined reference to `pt_iommu_amdv1_map_pages'
> >> aarch64-linux-ld: drivers/iommu/iommufd/selftest.o:(.rodata+0x288): undefined reference to `pt_iommu_amdv1_unmap_pages'
> >> aarch64-linux-ld: drivers/iommu/iommufd/selftest.o:(.rodata+0x2b0): undefined reference to `pt_iommu_amdv1_iova_to_phys'
> >> aarch64-linux-ld: drivers/iommu/iommufd/selftest.o:(.rodata+0x2d8): undefined reference to `pt_iommu_amdv1_read_and_clear_dirty'
> >> 
> >> Change the top-level function to skip all the amdv1 tests in this
> >> configuration.
> >
> > kconfig shouldn't let this build happen:
> >
> > config IOMMUFD_TEST
> >         bool "IOMMU Userspace API Test support"
> >         depends on DEBUG_KERNEL
> >         depends on FAULT_INJECTION
> >         depends on RUNTIME_TESTING_MENU
> >         depends on IOMMU_PT_AMDV1
> >         select IOMMUFD_DRIVER
> >
> > I think this is more fall out from Geert's change to make GENERIC_PT 
> > unselectable, the 'depends on' is no longer enough??
> >
> > I guess the right fix is to add a "select GENERIC_PT" above, does that
> > work for you?
> 
> No, it's already set.
> 
> The configuration that had the build failure contains
> 
> CONFIG_ARM64=y
> # CONFIG_AMD_IOMMU is not set
> CONFIG_GENERIC_PT=y
> CONFIG_IOMMU_PT=m
> CONFIG_IOMMU_PT_AMDV1=m
> CONFIG_IOMMUFD=y
> CONFIG_IOMMUFD_TEST=y
> 
> I suspect this is because IOMMUFD_TEST is 'bool' and can be enabled
> even when its dependencies are =m, but if the test code is built-in,
> it can't actually link.

Okay, that makes more sense..

> This should be fixable using
> 
>        depends on IOMMU_PT_AMDV1=y || IOMMUFD=IOMMU_PT_AMDV1

This is probably the best option, can you check it with the
problematic .config?

> or by forcing the iommufd code to be a loadable module when
> IOMMU_PT_AMDV=m, e.g. through
> 
> config IOMMUFD
>       depends on IOMMU_PT_AMDV1 || !IOMMU_PT_AMDV1 || !RUNTIME_TESTING_MENU

Because we don't want this restriction for production builds..

> A cleaner approach might be to split out IOMMUFD_TEST into a separate
> module rather than including it inside of the IOMMUFD module,
> but that is likely much more work.

I think there are cross calls that would make that difficult :\

Jason