Re: [PATCH RFC/RFT] vfio/pci: Create feature to disable MSI virtualization

[Jason Gunthorpe <jgg@ziepe.ca>]

On Fri, Jan 03, 2025 at 02:47:11PM +0000, David Woodhouse wrote:

> Probably best for the hypervisor to have some way to *advertise* that
> it's handling this though, as guests also might want to run on
> hypervisors which don't.

If the hypervisor doesn't properly virtualize the device it shouldn't
assign it to a VM to start with :\

Intel looked at the question of advertising clean interrupt remapping
when trying to virtualize IMS and it didn't seem so great.

Bare metal machines need to work, so any test they could think of
adding would either fail on bare metal or fail on existing VMs.

VMM's have taken the approach of not telling the guest they are in VMs
and then also not implementing the bare metal HW behaviors with full
fidelity. So we have no way to discover that the VMM is, in fact,
emulating broken "hw".

Thus we get this push that all kernels need to accomodate the worst
VMM behaviors :(

Jason