AMD Navi ROM BAR Quirk

AMD RDNA 2 and RDNA 3 GPUs (RX 6000 and RX 7000 series) have a known issue when used with vfio-pci: the ROM BAR returns an invalid signature (0xffff instead of 0xaa55). This causes a 65-second PCIe bus reset hang and leaves the device stuck in the D3cold power state.

Symptoms

vfio-pci 0000:08:00.0: Invalid PCI ROM header signature: expecting 0xaa55, got 0xffff
error getting device from group 22: No such device

QEMU exits immediately after this message. The GPU cannot be used until the host is cold-rebooted.

Fix: supply a VBIOS dump

Do not use rombar=1 without also supplying a VBIOS file. vee defaults to rombar=0 to avoid this hang. To re-enable ROM BAR (some games need it), you must supply the correct VBIOS dump via rom_file.

Get the VBIOS

If vfio-pci owns the GPU at boot (the typical setup), amdgpu never loads and the sysfs ROM interface is unavailable. Use one of these sources:

TechPowerUp VGABIOS database — download the ROM for your exact board revision from techpowerup.com/vgabios. This is the easiest option.

Sysfs dump — only possible if a second GPU drives the host display (so amdgpu never touches the passthrough card):

echo 1 | sudo tee /sys/bus/pci/devices/0000:08:00.0/rom
sudo cat /sys/bus/pci/devices/0000:08:00.0/rom > ~/.vee/gpu.rom
echo 0 | sudo tee /sys/bus/pci/devices/0000:08:00.0/rom

Configure vm.yaml

gpu:
  mode: passthrough
  pci_addr: "0000:08:00.0"
  rom_file: "/home/youruser/.vee/gpu.rom"

With rom_file set, QEMU serves the VBIOS from the file instead of probing the BAR, and the 65-second hang is avoided.

D3cold recovery

If the GPU is already stuck in D3cold from a previous failed start, vee will attempt a PCI function-level reset before the next vee start. If that fails, a cold reboot of the host is required to recover the device.