You are not logged in.

#1 2026-09-13 00:06:43

qherring
Member
Registered: 2023-05-05
Posts: 88

VFIO passthrough guest resets at Windows bootmgr handoff

Setup: Clevo X580WNT-G (Arrow Lake-HX), RTX 5090 Mobile [10de:2c58] + audio [10de:22e9] on vfio-pci, Windows 11 guest via libvirt/qemu/OVMF.

This configuration has been in stable daily use for several months across multiple kernel updates through the 7.1.x series; the failure appeared immediately and reproducibly on updating to 7.2.4, with no changes to the VM or its configuration. When setting this up there was a Blackwell FLR quirk that was initially present, but has been already handled via a systemd unit that empties reset_method for the GPU (0000:02:00.0) before libvirtd starts, disabling FLR. verified active on both the working and failing kernels.

Symptom (when on 7.2.4): guest resets about 5s after start, 1–2s after OVMF's

BdsDxe: starting ... bootmgfw.efi

. No BSOD even with Windows auto-restart-on-failure disabled (hard fault, not a bugcheck). Host dmesg completely clean, normal vfio-pci reset pairs, no AER/BAR errors. WinRE runs stably for minutes on the same kernel; only the normal boot path dies.

Tried the following downgrades:
edk2-ovmf 202608 → 202605 (kept kernel 7.2.4):    still fails
qemu 11.1.1 → 11.0.3 full set (kept kernel 7.2.4):    still fails
linux 7.2.4 → 7.1.6: works
qemu/ovmf restored to current (kernel 7.1.6):    still works

Regression tracks the kernel alone.

Possibly related to this?:
Proxmox 7.0.x-kernel regression on Blackwell GB203 attributed to nova_core interacting with VFIO at VM start: https://forum.proxmox.com/threads/gpu-p … on.184197/
Note that I have not yet tested blacklisting nova_core while on 7.2.4.

For now my workaround is pinning kernel at 7.1.6

Offline

Board footer

Powered by FluxBB