Re: [PATCH 00/11] mlx5 support for VFIO self test

[David Matlack <dmatlack@google.com>]

On 2026-05-01 01:43 PM, Jason Gunthorpe wrote:
> On Fri, May 01, 2026 at 09:11:11AM -0700, David Matlack wrote:
> > On Thu, Apr 30, 2026 at 5:08 PM Jason Gunthorpe <jgg@nvidia.com> wrote:
> > >
> > > Add an mlx5 driver to VFIO self test. This is largely a remix of the
> > > existing VFIO mlx5 driver in rdma-core. It uses an RDMA loopback QP
> > > to issue RDMA WRITE operations which effectively perform memory
> > > copies using DMA. Since mlx5 has a stable programming ABI this
> > > should work on devices from CX5 to current HW. The device FW must
> > > support the QP loopback configuration.
> > 
> > > This entire series was coded by Claude Code in about 4 days.
> > 
> > Very exciting. Josh Hilke from Google is also working on using AI to
> > create a selftest driver for Intel IGB NICs so VFIO selftests can run
> > in QEMU [1]. So it's encouraging to see you were able to do it with
> > mlx5.
> > 
> > [1] https://www.qemu.org/docs/master/system/devices/igb.html
> 
> Yes! I would feed DPDK in as well in this case? Combined with the
> kernel driver it should be doable. It is much easier if you understand
> how the NIC works, of course. This worked out significantly because I
> guided it through sufficiently small steps and knew where to find all
> the quality reference material..
> 
> > >  - Make it work on a PF too (this is surprisingly hard!).
> > 
> > Can it work on CX VFs? We're interested in continuously performing
> > memory copies across a Live Update using a VF via selftests to
> > demonstrate SR-IOV preservation (when we eventually get there).
> 
> Yes, I started with VF because it is simpler.

Makes sense. I tested it out and was able to get vfio_pci_driver_test
passing with a CX7 VF.

> The PF support flow requires a bunch more complicated stuff.

Do you think it's worth supporting PFs? If anyone with a CX NIC can
enable SR-IOV and run selftests on a VF then we can keep the driver
somewhat simpler.