[vfio-users] NVIDIA GPU Passthrough to Win10 - Driver Disabled (Code 43)

[Alex Williamson]

On Mon, 25 Jul 2016 21:39:56 -0400
Steven Bell <stv.bell07 at gmail.com> wrote:

> Hey Alex,
> 
> I really appreciate all your help and your quick responses (and everyone
> else too, thanks!).
> 
> I have found a Utility from MSI that will detect if there is an updated
> version of the my card's firmware, but it won't run on the VM (I assume
> it's looking for some version bit for Win10_64bit that just isn't set).

I would not suggest ever trying to update device firmware in a VM.
 
> So I'm left with having to physically remove the GPU from the host,
> re-install it in my Windows desktop, then run the utility in the hopes that
> it's got a firmware update (likely, since I don't think I've ever flashed
> it before), and that that update enables it to function with UEFI.

You haven't said what the specific card vendor/model is, but perhaps
there's a UEFI capable BIOS on techpowerup:

https://www.techpowerup.com/vgabios/?model=GTX+670

> If this doesn't work, I'll look into patching the kernel. I'm already using
> kernel 4.7 from the rawhide repo so as to enable quirks for ACS for the
> Skylake processor I'm using. I don't know how to go about patching the
> kernel so I'll put that off as my last resort if I can't get the OVMF
> working with this last try.
> 
> I do have two questions for you though, Alex. Would completely disabling
> the IGD help in anyway? It's not really necessary in my setup since the
> host is headless and the IGD just provides a physical terminal if ever SSH
> becomes unreachable.

Yes it would help, but then your GTX would become the primary console
and you'll need to go through some extra steps to prevent the non-pci
drivers from attaching to the video head too.  Minimally
video=efifb:off,vesafb:off but people always seem to find other drivers
that latch onto vga it seems.  I'd certainly want a serial console for
that, I know from assigning IGD on my laptop what a pain it is to run
completely headless.
 
> Also, for the OVMF, I see two UEFI x86_64 modes available when I create a
> virtual machine:
> 
> UEFI x86_64: /usr/share/edk2.git/ovmf-x64/OVMF_CODE-pure-efi.fd
> UEFI x86_64 /usr/share/edk2/ovmf/OVMF_CODE.fd
> 
> I don't understand why there's two or what the difference is between them.
> All my tests have been using the first one (pure-efi). Am I likely to see
> different results if I try with the OVMF_CODE.fd one?

I use /usr/share/edk2.git/ovmf-x64/OVMF-pure-efi.fd, which is the one
from Gerd's firmware repo (newer).  The other is delivered by Fedora
(older).  Thanks,

Alex