AMD RX 7900 XTX GPU Passthrough Setup Guide

Host: Fedora 44, AMD Ryzen 9800X3D, Gigabyte Aorus Elite X870m WiFi7

License

Copyright (C) 2026 Caleb Davis Licensed under the GNU General Public License v3 or later. See LICENSE.

Overview

This guide documents the complete host-side setup for AMD GPU passthrough on Fedora with an RX 7900 XTX (reference VBIOS). The setup requires a custom patched kernel to work around a kernel bug where AMD GPU firmware flags (force_iommu/require_direct) prevent VFIO container setup on kernel >= 6.10.

Patch file: vfio-amd-passthrough.patch
Kernel version: 6.19.14-200.vfio_amd.fc44

---

1. BIOS/UEFI Settings

• SVM Mode → Enabled
• IOMMU → Enabled (AMD-Vi)
• Above 4G Decoding → Enabled
• Resizable BAR (ReBAR) → Enabled
• Primary Display → IGD/Integrated (host uses iGPU)

---

2. Kernel Boot Parameters

Add to GRUB_CMDLINE_LINUX in /etc/default/grub:

amd_iommu=on amd_iommu_force_isolation=0 vfio-pci.ids=1002:744c,1002:ab30,1002:7446,1002:7444 amd_vfio_pass_fix=1

• amd_iommu=on — enable AMD IOMMU
• amd_iommu_force_isolation=0 — disable strict IOMMU isolation check
• vfio-pci.ids=... — bind all 4 GPU functions to vfio-pci at boot
• amd_vfio_pass_fix=1 — enable the VFIO passthrough workarounds introduced by this patch (see section 10). Without this parameter the patched kernel behaves identically to stock; the workarounds are completely inactive.

Note: The PCI IDs listed above are specific to the RX 7900 XTX tested here. Yours may differ. Use the following command to find the correct IDs for your card:

lspci -nn | grep -i "VGA\|Audio\|USB\|Serial" | grep AMD

Regenerate GRUB config:

sudo grub2-mkconfig -o /boot/grub2/grub.cfg

---

3. Custom Patched Kernel

The kernel requires code modifications to drivers/iommu/iommu.c and drivers/vfio/vfio_iommu_type1.c to work around the require_direct/force_iommu VBIOS flag issue.

Building the kernel RPM

# Install build dependencies
sudo dnf install -y rpm-build make gcc flex bison openssl-devel \
  elfutils-libelf-devel bc dwarves python3-devel asciidoc xmlto python3-sphinx

# Set up rpmbuild tree
mkdir -p ~/rpmbuild/{SOURCES,SPECS,RPMS,BUILD}

# Download Fedora kernel SRPM
dnf download --source kernel

# Install SRPM
rpm -ivh kernel-*.src.rpm

# Copy patch file
cp vfio-amd-passthrough.patch ~/rpmbuild/SOURCES/

# Edit kernel.spec:
# 1. Add: %define buildid .vfio_amd  (after the commented-out buildid line)
# 2. Add: Patch2: vfio-amd-passthrough.patch  (after Patch1)
# 3. Add: ApplyOptionalPatch vfio-amd-passthrough.patch  (before linux-kernel-test line)

# Build
cd ~/rpmbuild
rpmbuild -bb --target=x86_64 \
  --without debug \
  --without debuginfo \
  --without perf \
  --without tools \
  --without selftests \
  --without cross_headers \
  --without ynl \
  SPECS/kernel.spec

# Install
sudo dnf install \
  ~/rpmbuild/RPMS/x86_64/kernel-*.vfio_amd.*.rpm \
  ~/rpmbuild/RPMS/x86_64/kernel-core-*.vfio_amd.*.rpm \
  ~/rpmbuild/RPMS/x86_64/kernel-modules-*.vfio_amd.*.rpm \
  ~/rpmbuild/RPMS/x86_64/kernel-modules-core-*.vfio_amd.*.rpm \
  ~/rpmbuild/RPMS/x86_64/kernel-modules-extra-*.vfio_amd.*.rpm

---

4. Systemd Service — GPU Function Binding

Functions 03:00.2 (USB) and 03:00.3 (Serial) need to be manually bound to vfio-pci since they're driven by built-in kernel drivers that load before modprobe.

Create /etc/systemd/system/vfio-bind-gpu.service:

[Unit]
Description=Bind AMD GPU functions to vfio-pci and set IOMMU group type
After=systemd-udev-settle.service
Before=libvirtd.service virtqemud.service

[Service]
Type=oneshot
RemainAfterExit=yes
ExecStart=/bin/sh -c '\
  echo "auto" > /sys/bus/pci/devices/0000:03:00.0/iommu_group/type; \
  echo "0000:03:00.2" > /sys/bus/pci/drivers/xhci_hcd/unbind 2>/dev/null || true; \
  echo "0000:03:00.3" > /sys/bus/pci/drivers/i2c-designware-pci/unbind 2>/dev/null || true; \
  echo "vfio-pci" > /sys/bus/pci/devices/0000:03:00.2/driver_override; \
  echo "vfio-pci" > /sys/bus/pci/devices/0000:03:00.3/driver_override; \
  echo "0000:03:00.2" > /sys/bus/pci/drivers/vfio-pci/bind 2>/dev/null || true; \
  echo "0000:03:00.3" > /sys/bus/pci/drivers/vfio-pci/bind 2>/dev/null || true'

[Install]
WantedBy=multi-user.target

Enable:

sudo systemctl enable --now vfio-bind-gpu.service

---

5. libvirt Configuration

Edit /etc/libvirt/qemu.conf and add:

iommufd = 0

This forces QEMU to use the legacy vfio-pci container interface instead of iommufd.

Restart libvirt:

sudo systemctl restart virtqemud virtnetworkd

---

6. virtqemud Memory Lock Limit

Create /etc/systemd/system/virtqemud.service.d/override.conf:

[Service]
LimitMEMLOCK=infinity

sudo systemctl daemon-reload
sudo systemctl restart virtqemud

---

7. VBIOS ROM File

The GPU VBIOS ROM is required for OVMF/UEFI to initialize the GPU display output (GOP) during VM boot. Download the matching VBIOS for your card from TechPowerUp (search for RX 7900 XTX, VBIOS version 022.001.002.009.000001).

Place in a QEMU-accessible location:

sudo cp AMD.RX7900XTX.24576.221115.rom /usr/share/qemu/
sudo chmod 644 /usr/share/qemu/AMD.RX7900XTX.24576.221115.rom

---

8. VM XML Configuration

Key elements of the libvirt VM definition:

<!-- UEFI firmware with secure boot -->
<os firmware="efi">
  <type arch="x86_64" machine="q35">hvm</type>
  <firmware>
    <feature enabled="yes" name="enrolled-keys"/>
    <feature enabled="yes" name="secure-boot"/>
  </firmware>
</os>

<!-- CPU passthrough with AMD SMT fix -->
<cpu mode="host-passthrough" check="none" migratable="off">
  <topology sockets="1" dies="1" clusters="1" cores="6" threads="2"/>
  <feature policy="require" name="topoext"/>
</cpu>

<!-- Hypervisor hiding -->
<features>
  <hyperv mode="custom">
    <vendor_id state="on" value="AuthenticAMD"/>
  </hyperv>
  <kvm><hidden state="on"/></kvm>
  <ioapic driver="qemu"/>
  <smm state="on"/>
</features>

<!-- GPU passthrough with VBIOS ROM -->
<hostdev mode="subsystem" type="pci" managed="yes">
  <source>
    <address domain="0x0000" bus="0x03" slot="0x00" function="0x0"/>
  </source>
  <rom bar="on" file="/usr/share/qemu/AMD.RX7900XTX.24576.221115.rom"/>
</hostdev>

<!-- GPU audio, USB controller, serial bus -->
<hostdev mode="subsystem" type="pci" managed="yes">
  <source><address domain="0x0000" bus="0x03" slot="0x00" function="0x1"/></source>
</hostdev>
<hostdev mode="subsystem" type="pci" managed="yes">
  <source><address domain="0x0000" bus="0x03" slot="0x00" function="0x2"/></source>
</hostdev>
<hostdev mode="subsystem" type="pci" managed="yes">
  <source><address domain="0x0000" bus="0x03" slot="0x00" function="0x3"/></source>
</hostdev>

---

9. Verification

After setup, verify all 4 GPU functions are bound to vfio-pci:

lspci -nnk | grep -A 2 "03:00"
# All should show: Kernel driver in use: vfio-pci

Check IOMMU groups:

for d in /sys/kernel/iommu_groups/*/devices/*; do
  n=${d#*/iommu_groups/*}; n=${n%%/*}
  printf 'IOMMU Group %s ' "$n"
  lspci -nns "${d##*/}"
done | grep "03:00"

Start VM:

sudo virsh start win11

---

10. The Kernel Patch — Technical Background

The RX 7900 XTX reference VBIOS sets a force_iommu flag that causes the kernel to assign an identity IOMMU domain to the GPU and set require_direct=1. On kernel >= 6.10, this prevents VFIO from setting up the container needed for passthrough.

Boot parameter gate

All workarounds introduced by this patch are disabled by default. They are activated exclusively when the kernel is booted with:

amd_vfio_pass_fix=1

This is implemented via a standard __setup() early-param handler defined in drivers/iommu/iommu.c. The resulting bool amd_vfio_pass_fix variable is exported via EXPORT_SYMBOL_GPL and read as an extern in drivers/vfio/vfio_iommu_type1.c. Without the parameter, every patched code path falls back to identical stock kernel behaviour — no safety checks are relaxed, no IOMMU logic is altered.

Patches applied

All eight hunks below are individually conditioned on amd_vfio_pass_fix:

1. iommu.c: Don't re-set require_direct=1 from RESV_DIRECT regions
2. iommu.c: Skip identity mapping of address 0 (conflicts with VFIO DMA)
3. iommu.c: Allow overriding identity domains with DMA_FQ domain type
4. iommu.c: Disable firmware 1:1 mapping rejection for VFIO/iommufd
5. vfio_iommu_type1.c: Skip RLIMIT_MEMLOCK check for multi-group passthrough
6. vfio_iommu_type1.c: Always allow memory locking for VFIO containers
7. vfio_iommu_type1.c: Don't exclude address 0 from IOVA range
8. vfio_iommu_type1.c: Clear require_direct before VFIO domain allocation