Problem:

• After successfully passing a GPU to guest OS (I have only tested with Windows 10/11 guests), the system works fine until the guest is shut down.
• As soon as the guest powers down (whether by using shutdown inside the guest or using Virtual Machine Manager’s force stop), the host freezes, leaving a hard reboot as the only option.

Solution:

• Create a file named daemon (with no extension) in /etc/libvirt/hooks/ containing the following code:

#!/bin/bash  
  
if [[ "$2" == "start" ]]  
then  
       echo "Trying to reset gpus"  
       echo > /sys/bus/pci/devices/0000:03:00.0/reset_method  
       echo > /sys/bus/pci/devices/0000:04:00.0/reset_method  
       echo "Finished reset"  
fi

• The location and name of the file ensure it runs when libvirt starts.

Assumption:

• You have already isolated your GPU and passthrough works without hiccups.
• If you are looking for how to isolate a GPU, refer to: Arch Linux PCI Passthrough via OVMF
• This might be overwhelming for someone new to Linux, but it’s worth trying.
• I may write a future tutorial on how I did it, in a simple and concise form.

What does it do?

• This script resets the GPU bindings when libvirt starts.
• I’m not deeply familiar with libvirt and KVM, so I can’t explain why this step is necessary. However, after trying many approaches, this solution has been stable for over six months.

How I found the solution:

• After trying system updates and OS changes (I even switched from Ubuntu to OpenSUSE thinking the latest drivers might help), I spent over 50 hours googling the solution.
• Here are some of the sources that gave me hints:
  • Simple per-VM libvirt hooks with the vfio-tools hook helper
  • Single-GPU Passthrough GitHub
  • Proxmox forum thread on Intel Arc A770M GPU passthrough issues
• While one of these worked temporarily, the final solution I found was by manually running GPU reset commands at different times and noting the correct sequence.
• Once I confirmed it needs to run before I start libvirt (which is started when you launch Virtual Machine Manager), I tested this script.
• I implemented it more than 6 months ago and have not needed to change it since. This should fix the issue for most people facing this problem.

My System:

• OpenSUSE Tumbleweed
• MSI PRO Z690-A WIFI DDR5
• Intel i5-13600K + Arc A770 16 GB
• I have isolated the Arc A770 and use the iGPU for all other host purposes.
• I’m running Windows 10 as a guest OS with the GPU passed through.
• Passing a GPU to Ubuntu guests also works, so other distros should work too.