Re: [PATCH 7/7] iommu/vtd: Split paging_domain_compatible()

[Baolu Lu <baolu.lu@linux.intel.com>]

On 6/12/25 21:47, Jason Gunthorpe wrote:
> On Wed, Jun 11, 2025 at 12:50:11PM +0800, Baolu Lu wrote:
> 
>>> When this was first done that was harder, but now I think we have
>>> enough infrastructure to do it easily..
>> Yes. I can follow up with a patch series for broad discussion if you
>> don't want to include it in this series.
> Please, I'd rather keep this as is.
> 
> There is also a problem here with AMD, it turns out only its V1 page
> table can support disabling no-snoop, the v2 page table cannot and
> there is nothing like the PGSNP there.
> 
> The iommufd side assumes that if a device supports coherencey then all
> paging iommu_domains support it.. Maybe a creation flag can fix that
> also?

I guess so. For the AMD case, a FORCE_COHERENCY allocation flag
indicates that the IOMMU driver should select a right page table type to
meet this requirement. This requirement will also be enforced in the
domain attaching paths, meaning it will be an incompatible failure if a
FORCE_COHERENCY domain is attached to any device that doesn't support
this feature.

Anyway, we can discuss the details when we have that code.

Thanks,
baolu