Benjamin Cheng

Nvidia Compute + VFIO Setup January 15, 2021

I recently got my hands on a Nvidia RTX 2070 Super. Despite being much more powerful than my RX580, at first I was going to use it for VM passthrough only, and not for daily use. However AMD recently discontinued support for ROCm on Polaris chips so I guess I now need to use the 2070 for compute purposes such as machine learning 1. Obviously these two scenarios require different drivers: nvidia for the host compute, and vfio-pci for passthrough. Switching between the two is easy, but requires a sequence of unbinds and binds for each of the (4!!) PCI devices that are exposed.

To help ease this driver switching I wrote a wrote script called pci-bind, which handles the unbinding and binding a new driver for a group of devices with a certain PCI address.

To bind nvidia for compute I can just run:

sudo pci-bind nvidia 0000:0a:00

Replacing nvidia with vfio-pci does the opposite, which is what I put at the top of my qemu script.

On my system lspci gives:

0a:00.0 VGA compatible controller: NVIDIA Corporation TU104 [GeForce RTX 2070 SUPER] (rev a1)
0a:00.1 Audio device: NVIDIA Corporation TU104 HD Audio Controller (rev a1)
0a:00.2 USB controller: NVIDIA Corporation TU104 USB 3.1 Host Controller (rev a1)
0a:00.3 Serial bus controller [0c80]: NVIDIA Corporation TU104 USB Type-C UCSI Controller (rev a1)

and each of these live in /sys/bus/pci/devices/0000:0a:00.{0-3}. Hopefully you can see where the 0000:0a:00 argument comes from (it's just a prefix).

One issue I've experienced with this setup is that XWayland will pick up the Nvidia drivers and hold /dev/nvidia0 open so that the nvidia module will refuse the unbind request. I found that this was due to glvnd loading the Nvidia EGL ICD from /usr/share/glvnd/egl_vendor.d/10_nvidia.json. Unfortunately I couldn't find a way to simply not install the Nvidia GL driver but include the nvidia kernel driver (i.e. a headless driver of some sort) on Arch. Anyways to stop glvnd from loading the nvidia driver, I simply added the following to force EGL to load mesa only:

export __EGL_VENDOR_LIBRARY_FILENAMES=/usr/share/glvnd/egl_vendor.d/50_mesa.json

I also use flatpak, which will end up trying to install the Nvidia GL drivers as a flatpak package. To disable this I just masked out the driver package:

flatpak mask "org.freedesktop.Platform.GL.nvidia*"

screw Nvidia for making CUDA proprietary and not advancing open standards like OpenCL, but also screw AMD for discontinuing ROCm support randomly on a fairly recent card. Hopefully Intel comes up with something better, oh wait they did and SYCL is just another standard.