Windows Guest Stalling on 5.4.26 and later

Upgrading my host kernel to 5.7.9 to fix a network stack bug with my system I found that the windows VFIO performance became extremely unstable to the point of being unusable.

The symptoms of this bug as I experienced it:

  • Seems to only affect windows guests
  • NVidia GPU stalls in the guest, windows performs driver recovery repeatedly.
  • Dragging a selection rectangle on the desktop can freeze
  • Physical output of the GPU goes black during the GPU recovery.
  • BSOD due to failed GPU recovery in the guest.

I have spent the better part of a day tracking this one down, If others have experienced this issue the culprit is this commit that was introduced in 5.4.26:

Reverting this commit fixes this behavior.