GPU passthrough hot-swapping driver development?

[nucleare2 <nucleare2@protonmail.com>]

Hello, I'd like to discuss the possibility of developing a method for hot-swapping video cards between a KVM host and the guest machine when using passthrough.

It seems that currently passthrough is pretty restrictive as far as assignment to the host vs guest goes.  This means for applications like accelerating macOS fully a video card needs to be dedicated to passthrough for macOS so that native drivers grab a hold of it at macOS guest startup.

I've been thinking about the possibility of creating some kind of a dummy passthrough driver that may, at some minimal level, virtualize the GPU access so as to allow the KVM host and guest (macOS, Windows, Linux, whatever would need the full GPU access) to "hot-swap" access to the GPU.

As I said above, right now it appears that it's necessary to have multiple GPUs for the host vs guests that need/want passthrough, but if an appropriate driver is developed that somehow captures some core functionality, couldn't a kernel level key combination capture be implemented that would flip passthrough between host and guest(s)?

Basically if a macOS or Windows (or Linux) guest BELIEVES it still has control of the GPU, then the guest kernel should not panic and should still remain operational while the host get's access.  I'm not talking about any kind of capturing here, just simply making the host or guest think it's still getting GPU access and happily spinning away while the user is flipping between host/guest instances.

Has anyone considered this or put any work into this so far?

I've seen someone else mention this before somewhere, but it seemed to not get any attention. This is a pretty critical function that would benefit KVM users enormously.

-nuc