Re: [PATCH v5 00/14] iommu/amd: Introduce Nested Translation support

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Nov 13, 2025 at 01:52:02PM -0800, Nicolin Chen wrote:
> On Wed, Nov 12, 2025 at 06:24:52PM +0000, Suravee Suthikulpanit wrote:
> > Note: This series is rebased on top of:
> >  * Git repo: git://git.kernel.org/pub/scm/linux/kernel/git/iommu/linux.git 
> >    Branch: next
> >    Commit: 91920a9d87f5 ("Merge branches 'arm/smmu/updates', 'arm/smmu/bindings',
> >                           'mediatek', 'nvidia/tegra', 'amd/amd-vi' and 'core' 
> >                           into next")
> >  * [PATCH v5] iommu/amd: Add support for hw_info for iommu capability query
> >    https://lore.kernel.org/linux-iommu/20250926141901.511313-1-suravee.suthikulpanit@amd.com/T/#u 
> 
> Nit: this patch doesn't apply cleanly on 91920a9d87f5 :-/
> 
> >  drivers/iommu/amd/Makefile          |   2 +-
> >  drivers/iommu/amd/amd_iommu.h       |  36 ++++
> >  drivers/iommu/amd/amd_iommu_types.h |  48 +++++-
> >  drivers/iommu/amd/init.c            |   8 +
> >  drivers/iommu/amd/iommu.c           | 221 +++++++++++++++---------
> >  drivers/iommu/amd/iommufd.c         |  50 ++++++
> >  drivers/iommu/amd/iommufd.h         |   5 +
> >  drivers/iommu/amd/nested.c          | 259 ++++++++++++++++++++++++++++
> >  include/uapi/linux/iommufd.h        |  11 ++
> 
> So, this seems to be a preparatory series for AMD vIOMMU, yet it
> doesn't properly work since it's missing IOMMUFD_VIOMMU_TYPE_AMD
> and the invalidation component (HW_QUEUE).
> 
> However, the series does declare IOMMU_HWPT_DATA_AMD_GUEST in the
> uAPI header. I am afraid that might confuse user who might think
> AMD now supports virtualization using the HWPT-based mode, like
> Intel VT-d.
> 
> So, maybe we should either:
>   - leave a note at IOMMU_HWPT_DATA_AMD_GUEST to declare it is
>     incomplete yet, and remove later
>   - keep IOMMU_HWPT_DATA_AMD_GUEST in an AMD driver header, and
>     move to the uAPI header later
> 
> Jason?

Yeah, I like to see this incremental work, but Alex recently raised
that we should be a bit more careful about how userspace perceives
these partially complete things.

I don't think tricks with head files work well, I think what you'd
want to do is leave some critical system call disabled until all the
work is finished so the VMM never has to see a half working
implementation?

The patch to get the info would have been a nice choice for this purpose..

Jason