[vfio-users] GPU instance ID not persistent on guest restart

[Alex Williamson]

On Mon, 2015-09-14 at 15:54 +0000, Abdulla Bubshait wrote:
> I can get my GPU to work well once per host reboot cycle. It would appear
> on subsequent guest reboots the GPU instance ID changes slightly.
> 
> This is what I get on my windows system:
> Working
> PCI\VEN_1002&DEV_7300&SUBSYS_0B361002&REV_C8\
> On Reboot
> PCI\VEN_1002&DEV_7300&SUBSYS_0B361002&REV_00\
> 
> This results in the GPU not being detected as compatible by the driver and
> defaults back to the microsoft display driver. That means the GPU is
> detected properly only once per host reboot (or suspend). On subsequent
> qemu restarts the GPU appears as an unsupported A-series APU, instead of an
> r9 Fury.
> 
> Is there something that could be done to keep the GPU persistent across
> guest reboots?

It's surprising that the driver wouldn't claim it if only the revision
number is getting reset.  I thought most drivers don't really care about
the revision ID.  Perhaps test whether anything else in config space is
not quite right.  Try running:

$ sudo lspci -vvvxxxx -s <bus:dev.fn>

in the host and compare the output after the first VM boot to that after
the second VM boot.  If it's only the chip revision we could probably
quirk something, but that register is read-only according to the PCI
spec, so it shouldn't be affected by a bus reset.  Thanks,

Alex