NVIDIA GPU passthrough on XCP-ng 8.3 fails after reboot — UUID/PCI ID changes
-
I’m trying to use passthrough for an NVIDIA GPU on XCP-ng 8.3.
The host detects the GPU (lspci shows VGA + Audio), and IOMMU is enabled.However, whenever I apply xen-pciback.hide and reboot the host, XCP-ng generates a new internal UUID and new PCI ID for the GPU.
As a result, I cannot even assign the GPU to a VM, because the device “changes” its reference after each reboot.Additional important context:
I previously had a different GPU doing passthrough in a VM.
That GPU was removed.
The new GPU was installed in a different slot.
I suspect the issues may be related to leftover metadata from the previous GPU, which prevents the new GPU from being recognized consistently for passthrough.
Question:
Has anyone encountered this problem? Is there a way to completely clear the old passthrough state on the host or ensure that the new GPU is recognized consistently for passthrough? -
Hi,
Clear entirely the boot parameter with
/opt/xensource/libexec/xen-cmdline --delete-dom0 xen-pciback.hide, reboot, and re-assign. -
Thank you for your reply.
I followed your instructions; however, when I passthrough the GPU NVIDIA on the XenOrchestra and he same error occurs when I reboot.
Change the PCI and UUID of the graphics card to the one before rebooting.
-
This post is deleted! -
Can you tell us on what platform you are doing this? It looks like a buggy PCI reset.
-
I have a the XCP8.3 installed.
Xen Orchestra, commit bcee5 .When I add it manually, the audio, for example, it asks for a reboot, but then, when I restart the server, it has another pci ID, and another ID.
-
Oo I'm not sure to follow. But let me add @Team-Hypervisor-Kernel in case that rings a bell.
-
Hi,
@samuelolavo can you run the following command on the dom0 and paste its output please:
grep -A5 "'XCP-ng'" /etc/grub.cfgThen also run
lspci -vvv xe pci-list xl pci-assignable-list
