Re: [fedora-virt] Fatal PCIe error 928 on KVM GPU Passthrough

[Mario Ceresa]

Thanks Alex, for your kind and detailed response: please keep up with the
good work!

I'm interested in learning more about this and I have access to Quadro 4000
and 5000 cards, so please tell me if I can help with testing. I do not mind
compiling from sources or doing some debug, but I have almost zero
practical knowledge on GPU cards so I'll need some pointers :)

Best,

Mario


On 20 June 2013 23:19, Alex Williamson <alex.williamson(a)redhat.com&gt; wrote:

>  On Thu, 2013-06-20 at 09:30 +0200, Mario Ceresa wrote:
>  > Dear all,
>  > does anybody know if it is currently possible to do GPU passthrough in
>  kvm?
>  > I have access to a dual Quadro 4000 setup and I would like to pass the
>  > second one to kvm for use in a VM.
>  >
>  > I read a lot on internet and it seems feasible [1, 2, 3] and there is
>  even
>  > a reported success case [5].
>  >
>  > However when I start the VM the host reboots with PCIe error 928:
>  > Completion timeout.
>  >
>  > My current setup is:
>  >
>  > * Fedora 18 with 3.9.4-200 x86_64
>  > * dmesg | grep -e DMAR -e IOMMU shows that they are both enabled
>  > * nvidia driver installed from rpmfusion
>  > * I use vfio assignment as per [4]
>
>  GPUs are not well supported yet and you're probably not going to get
>  very far unless you're willing to compile a kernel and qemu and use qemu
>  from the commandline instead of through libvirt.  Some people have luck
>  assigning a GPU as a secondary graphics device to the emulated video,
>  others do not.  Sometimes I suspect this has more to do with addresses
>  assigned to the physical device possibly matching the virtual device...
>  so it really is luck.
>
>  I've posted branches of qemu and kernel that seem to work well for
>  several people (this is what was pointed to in the Arch Linux thread you
>  footnoted).  That code attempts to use the assigned GPU as the primary
>  graphics head for the guest.  I have one Quadro card on hand and it
>  never works as well as the GeForce cards, YMMV.  Graphics cards have all
>  sorts of interesting backdoors to get to PCI config space and I suspect
>  the Quadro cards are not as well supported by the Nouveau project, which
>  we've used for reverse engineered documentation of some of these quirks.
>
>  There's still some work to do before that code can be posted upstream,
>  in particular I've seen cases where a PCI bus won't come back from reset
>  without resetting the host system.  We need to make it more stable
>  before pushing to get it upstream.  Thanks,
>
>  Alex
>
>