On Thu, Jul 29, 2021 at 09:20:22AM +0200, Christoph Hellwig wrote:

On Wed, Jul 28, 2021 at 02:59:25PM -0300, Jason Gunthorpe wrote:

...

On Wed, Jul 28, 2021 at 01:38:58PM +0000, Wang, Zhi A wrote:

...

I guess those APIs you were talking about are KVM-only. For other hypervisors, e.g. Xen, ARCN cannot use the APIs you mentioned. Not sure if you have already noticed that VFIO is KVM-only right now.

There is very little hard connection between VFIO and KVM, so no, I don't think that is completely true.

The only connection is the SET_KVM notifier as far as I can tell. Which is used by a total of two drivers, including i915/gvt. That being said gvt does not only use vfio, but also does quite a few direct cals to KVM.

...

In an event, an in-tree version of other hypervisor support for GVT needs to go through enabling VFIO support so that the existing API multiplexers we have can be used properly, not adding a shim layer trying to recreate VFIO inside a GPU driver.

Yes. And if we go back to actually looking at the series a lot of it just removes entirely pointless indirect calls that go to generic code and not even the kvm code, or questionable data structure designs. If we were to support another upstream hypervisor we'd just need to union a few fields in struct intel_gpu and maybe introduce a few methods. Preferably in a way that avoids expensive indirect calls in the fast path.

fwiw I concur with the direction of this series. gvt landed 5 years ago, that should have been plenty of time to merge at least one of the other backends that float around. If it didn't happen in 5 years it aint suddenly happening in the next few, and the abstraction layer should be sunset.

Also yes structuring it more as a helper layer with some unions/subclassing than full blown backend abstractor layer would be a good idea too I guess (it usually is the right thing to do). -Daniel