Can't unload GVT-g driver after Win 11 VM crashes

Novgorod

New Member
Dec 10, 2025
1
0
1
I'm running Proxmox 9.1.5 on an Intel 10th-gen setup where the iGPU is split via GVT-g (mediated PCI device) between a permanently running TrueNAS VM and an occasionally running Windows 11 VM. Everything works in principle albeit the Win 11 being a bit sluggish which I attribute to the somewhat dated hardware. Recently the GPU has been acting up in the Win 11 VM when accessing it through Anydesk (which does hardware encoding), for example producing graphics glitches in the Chrome browser and sometimes crashing the Windows photo viewer when viewing images.

Once in a while the Win 11 VM would completely crash and kill all remote connections through external tools (like Anydesk or Rustdesk) and also through the Proxmox VNC viewer. Some CPU activity is shown on the summary page but it won't react to anything. When I try to kill the VM, the "stop" task shows this:
Code:
VM quit/powerdown failed - terminating now with SIGTERM
VM still running - terminating now with SIGKILL
actively clean up mediated device with UUID 00000000-0000-0000-0000-000000000102
The mediated device refers to the GVT-g iGPU and it hangs in this state for eternity. The actual VM task is not running anymore at that point but the GVT-g driver can never unload and let the VM boot up again. Manually unlocking it doesn't help because the mediated GPU is unavailable ("could not create mediated device"), and even completely removing the PCI device from the Win 11 hardware config won't let it boot ("TASK ERROR: timeout waiting on systemd").

The other VMs on the system are unaffected, but in this state it's unrecoverable because the GVT-g driver never finishes unloading. Rebooting the host (PVE -> Reboot) will shutdown PVE and just hang forever at some stage (I assume at the GPU shtudown/unload) without remote access. The only solution at that point is to power cycle it which requires physical access or an external KVM.

Is there a way to force-reset the GVT-g device (or the entire iGPU) without a host reboot?