Re: [RFC PATCH 0/2] vfio: Improve DMA mapping performance for huge pages

[Jason Gunthorpe <jgg@nvidia.com>]

On Tue, Dec 23, 2025 at 11:00:42PM +0000, Aaron Lewis wrote:
> This RFC explores the current state of DMA mapping performance across
> vfio, and proposes an implementation to improve the performance
> for "vfio_type1_iommu" for huge pages.

We are trying to sunset type1, please just use the memfd path on
iommufd which is supposed to be the fastest available option. You can
try to improve iommufd as well.

But we definately should not be making type 1 the best option and then
encouraging people to use it - that is the opposite of what we are
trying to do!

> This is being sent as an RFC because while there is a proposed solution
> for "vfio_type1_iommu", there are no solutions for the other two iommu
> modes.  Attached is a callstack in patch 1/2 showing where the latency
> issues are for iommufd, however, I haven't posted one
> for "iommufd_compat_type1". I'm also not clear on what the intention is
> for "iommufd_compat_type1" w.r.t. this issue.  Especially given it is so
> much slower than the others.

Probably your test is using VFIO_TYPE1_IOMMU for iommufd and comparing
it to VFIO_TYPE1v2_IOMMU for vfio. v0 forces 4k pages at the iommu
domain which is much slower.

It is also really weird because I have seen other people reporting a
30x speedup using iommufd and here you are reporting it is the slowest
option?

Jason