Re: [PATCH v3 00/10] Convert Intel VT-d to use the generic iommu page table

[David Matlack <dmatlack@google.com>]

On 2025-10-23 03:22 PM, Jason Gunthorpe wrote:
> Replace the VT-d iommu_domain implementation of the VT-d second stage and
> first stage page tables with the iommupt VTDSS and x86_64
> pagetables. x86_64 is shared with the AMD driver.
> 
> VT-d has HW that requires an incoherent page table walker, the majority of
> the patches are adding generic support for the required cache flushing to
> iommupt. This is modeled after the existing ARM64 version and is intended
> to be re-used there.
> 
> Applies on top of the AMD conversion:
>   https://patch.msgid.link/r/0-v2-5c26bde5c22d+58b-iommu_pt_jgg@nvidia.com
> 
> This is on github: https://github.com/jgunthorpe/linux/commits/iommu_pt_vtd

I ran VFIO selftests against this series and did not detect any
regressions.

Notably, vfio_dma_mapping_test confirmed that this series did not
regress VFIO's or IOMMUFD's ability to map 2MB and 1GB HugeTLB pages
as 2MB and 1GB entries in the I/O page tables as expected.  And
vfio_pci_driver_test validates that a device (Intel DSA in this case)
can do DMA, and exercises I/O faulting handling.

Tested-by: David Matlack <dmatlack@google.com>