GPU passthrough with libvirt qemu kvm

From Gentoo Wiki
Jump to:navigation Jump to:search

GPU passthrough is a technology that allows the Linux kernel to directly present an internal PCI GPU to a virtual machine.

The device acts as if it were directly driven by the VM, and the VM detects the PCI device as if it were physically connected. GPU passthrough is also often known as IOMMU, although this is a bit of a misnomer, since the IOMMU is the hardware technology that provides this feature but also provides other features such as some protection from DMA attacks or ability to address 64-bit memory spaces with 32-bit addresses.

As you can imagine, the most common application for GPU passthrough at least gaming, since GPU passthrough allows a VM direct access to the graphics card with the end result of being able to play games with nearly the same performance as if you were running the game directly on the computer.

QEMU (Quick EMUlator) is a generic, open source hardware emulator and virtualization suite.

Note
This article typically uses KVM as the accelerator of choice due to its GPL licensing and availability. Without KVM nearly all commands described here will still work (unless KVM specific).

Installation

BIOS and UEFI firmware

In order to utilize KVM either VT-x or AMD-V must be supported by the processor. VT-x or AMD-V are Intel and AMD's respective technologies for permitting multiple operating systems to concurrently execute operations on the processors.

To inspect hardware for visualization support issue the following command:

user $grep --color -E "vmx|svm" /proc/cpuinfo

For a period manufacturers were shipping with virtualization turned off by default in the system BIOS

Hardware

  • A CPU that supports Intel VT-d or AMD-Vi. Check List of compatible Intel CPUs (Intel VT-x and Intel VT-d).
  • A motherboard that supports the aforementioned technologies. To find this out, check in your motherboard's BIOS configuration for an option to enable IOMMU or something similar. Chances are that your motherboard will support it if it's from 2013 or newer, but make sure to check since this is a niche technology and some manufacturers may save costs by axing it from their motherboards or delivering a defective implementation (such as Gigabyte's 2015-2016 series) simply because NORPs never use it.
  • At least two GPUs: one for your physical OS, another for your VM. (You can in theory run your computer headless through SSH or a serial console, but it might not work and you risk locking yourself away from your computer if you do so).
  • Optional but recommended: Additional monitor, keyboard and mouse.

EFI configuration

Go into BIOS (EFI) settings and turn on VT-d and IOMMU support.

Note
VT-d and Virtualization configuration params are same
Note
Some EFI doesn't have IOMMU configuration settings

IOMMU

IOMMU – or input–output memory management unit – is a memory management unit (MMU) that connects a direct-memory-access–capable (DMA-capable) I/O bus to the main memory. The IOMMU maps a device-visible virtual address ( I/O virtual address or IOVA) to a physical memory address. In other words, it translates the IOVA into a real physical address.

In an ideal world, every device has its own IOVA address space and no two devices share the same IOVA. But in practice this is often not the case. Moreover, the PCI-Express (PCIe) specifications allow PCIe devices to communicate with each other directly, called peer-to-peer transactions, thereby escaping the IOMMU.

That is where PCI Access Control Services (ACS) are called to the rescue. ACS is able to tell whether or not these peer-to-peer transactions are possible between any two or more devices, and can disable them. ACS features are implemented within the CPU and the chipset.

Unfortunately the implementation of ACS varies greatly between different CPU or chip-set models.

IOMMU kernel configuration

To enable IOMMU support in kernel:

KERNEL
Device Drivers --->
  [*] IOMMU Hardware Support --->
            Generic IOMMU Pagetable Support ----
      [*]   AMD IOMMU support
      <*>     AMD IOMMU Version 2 driver
      [*]   Support for Intel IOMMU using DMA Remapping Devices
      [*]     Support for Shared Virtual Memory with Intel IOMMU
      [*]     Enable Intel DMA Remapping Devices by default
      [*]   Support for Interrupt Remapping

If you have CONFIG_TRIM_UNUSED_KSYMS (Trim unused exported kernel symbols) enabled, you will need to whitelist some symbols. Otherwise, you may get error messages of the form Failed to add group <n> to KVM VFIO device: Invalid argument. See the gentoo forum thread kernel 4.7.0 breaks pci passthrough [SOLVED] and the kvm mailing list thread KVM/VFIO passthrough not working when TRIM_UNUSED_KSYMS is enabled (list of symbols to whitelist in the second post).

KERNEL
[*] Enable loadable module support --->
    [*]   Trim unused exported kernel symbols
    (/path/to/whitelist) Whitelist of symbols to keep in ksymtab
FILE /path/to/whitelist
vfio_group_get_external_user
vfio_external_group_match_file
vfio_group_put_external_user
vfio_group_set_kvm
vfio_external_check_extension
vfio_external_user_iommu_id
mdev_get_iommu_device
mdev_bus_type

Re-build the kernel.

GRUB bootloader

When using GRUB as the secondary bootloader, IOMMU will need to be enabled by modifying kernel's commandline parameters. Edit the /etc/default/grub file and add the following values to the GRUB_CMDLINE_LINUX variable:

FILE /etc/default/grub
GRUB_CMDLINE_LINUX="... iommu=pt intel_iommu=on pcie_acs_override=downstream,multifunction ..."
Note
If the system hangs after rebooting, check the BIOS and IOMMU settings.

Apply changes:

root #grub-mkconfig -o /boot/grub/grub.cfg

Verify IOMMU has been enabled and is operational:

user $dmesg | grep 'IOMMU enabled'
[    0.000000] DMAR: IOMMU enabled
Note
For CPU on XEN architecture, run:
user $lspci -vv | grep -i 'Access Control Services'

IOMMU groups

Passing through PCI or VGA devices requires you to pass through all devices within an IOMMU group. The exception to this rule are PCI root devices that reside in the same IOMMU group with the device(s) we want to pass through. These root devices cannot be passed through as they often perform important tasks for the host. A number of (Intel) CPUs, usually consumer-grade CPUs with integrated graphics (IGD), share a root device in the same IOMMU group as the first PCIe 16x slot.

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

IOMMU Group 13 01:00.0 VGA compatible controller [0300]: NVIDIA Corporation GP104 [GeForce GTX 1080] [10de:1b80] (rev a1)

IOMMU Group 15 02:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI] Ellesmere [Radeon Pro WX 7100] [1002:67c4]

IOMMU Group 16 02:00.1 Audio device [0403]: Advanced Micro Devices, Inc. [AMD/ATI] Ellesmere [Radeon RX 580] [1002:aaf0]

...

Nvidia in IOMMU Group 13 and AMD Video Card in IOMMU group 15 and 16. Everything looks fine. But if you have buggy IOMMU support and all devices within one IOMMU group, hardware can't guarantee good device isolation. Unfortunately, it is not possible to fix that. The only workaround - use ACS override patch witch ignore IOMMU hardware check. See ACS override patch.

Other devices in my IOMMU group

ACS override patch

root #git clone https://github.com/feniksa/gentoo_ACS_override_patch.git /etc/portage/patches

Next re-emerge the kernel

root #emerge gentoo-sources
 * Applying 4400_alpha-sysctl-uac.patch (-p1) ...                                       [ ok ]
 * Applying 4567_distro-Gentoo-Kconfig.patch (-p1) ...                                  [ ok ]
>>> Source unpacked in /var/tmp/portage/sys-kernel/gentoo-sources-4.14.52/work
>>> Preparing source in /var/tmp/portage/sys-kernel/gentoo-sources-4.14.52/work/linux-4.14.52-gentoo ...
 * Applying override_for_missing_acs_capabilities.patch ...                             [ ok ]
 * User patches applied.

VFIO

Kernel drivers:

KERNEL
Device Drivers --->
  <M> VFIO Non-Privileged userpsace driver framework --->
      [*]   VFIO No-IOMMU support ----
      <M>   VFIO support for PCI devices
      [*]     VFIO PCI support for VGA devices
      < >   Mediated device driver framework

Search for VGA card IDs. Run:

root #lspci -nn
...
04:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI] Vega 10 XL/XT [Radeon RX Vega 56/64] [1002:687f] (rev c1)
04:00.1 Audio device [0403]: Advanced Micro Devices, Inc. [AMD/ATI] Device [1002:aaf8]
..


Add VGA PCI IDs to VFIO

FILE /etc/modprobe.d/vfio.conf
options vfio-pci ids=1002:687f,1002:aaf8

Libvirt

Windows

Create Windows 10 as usual via libvirt manager. Edit virtual image, click Add Hardware, select AMD Ati Vega 64 and AMD Ati device. Click Apply.

Now start the Windows 10 guest OS.

AMD cards have 2 devices on PCIe bus -> one video output and another is HDMI output. Windows drivers works only if KVM will bypass to Windows both AMD devices.

Fixed Vega 56/64 reset bug

AMD Vega 56/64 is unable to initialize itself after Guest host shutdown/reboot, because drivers left card in "garbage" state. As workaround of this bug, VFIO should load AMD card ROM at guest startup. To do that:

  1. Install clear Windows 10 somewhere (not in libvirt. A BARE METAL Windows 10 installation.)
  2. Install all latest Windows 10 updates.
  3. Install AMD vga drivers.
  4. Reboot.
  5. Go again to the bare metal Windows 10 installation.
  6. Install GPU-Z.
  7. In GPU-Z in main tab, near BIOS version will be small button "Save ROM". Click it and save the ROM somewhere. This ROM will be needed for Gentoo and libvirt. For example, for a Vega64 the ROM can be saved as Vega64.rom
  8. Reboot into Gentoo.
  9. root #mkdir /etc/firmware
  10. Copy to /etc/firmware the ROM file (for this example it is Vega64.rom)
  11. Go to /etc/libvirt/qemu
  12. Edit the xml file with description of the Windows 10 guest.
  13. Find section with AMD Video Card device (not AMD HDMI. You can always re-check with lspci)

In my case:

FILE /etc/libvirt/qemu/win10.xml
...
    <hostdev mode='subsystem' type='pci' managed='yes'>
      <source>
        <address domain='0x0000' bus='0x04' slot='0x00' function='0x0'/>
      </source>
      <address type='pci' domain='0x0000' bus='0x00' slot='0x08' function='0x0'/>
    </hostdev>
    ...

14. Add path to vga rom

  <rom bar='on' file='/etc/firmware/Vega64.rom'/>

So, it should be:

FILE /etc/libvirt/qemu/win10.xml
...
    <hostdev mode='subsystem' type='pci' managed='yes'>
      <source>
        <address domain='0x0000' bus='0x04' slot='0x00' function='0x0'/>
      </source>
      <address type='pci' domain='0x0000' bus='0x00' slot='0x08' function='0x0'/>
      <rom bar='on' file='/etc/firmware/Vega64.rom'/>
    </hostdev>
...

Fixed Navi reset bug

AMD Navi 10 series GPUs require a vendor specific reset procedure. According to AMD a PSP mode 2 reset should be enough however at this time the details of how to perform this are not available.

Instead kernel can signal the SMU to enter and exit BACO which has the same desired effect.

To apply workaround (for kernel 4.19.72. For newer kernel replace number 4.19.72 with newer kernel):

  1. Download patchset https://github.com/feniksa/gentoo_ACS_override_patch/blob/master/sys-kernel/gentoo-sources-4.19.72/navi_reset.patch
  2. Put patchset into: /etc/portage/patches/sys-kernel/gentoo-sources-4.19.72
  3. Re-emerge gentoo-sources package:
    root #emerge gentoo-sources
  4. Re-compile the kernel

Applied patchset contain custom logic for reset GPU.

Sound

root #mkdir /home/qemu
root #cp /home/<user>/.config/pulse /home/qemu
root #chown qemu:qemu -R /home/qemu

Change the home directory for the qemu user:

root #usermod -d /home/qemu qemu

QEMU

In case you want to use QEMU directly, here are some configurations to get you started. In general, as a typical QEMU call will usually require many command-line flags, it is typically advised to place the QEMU call in a bash script and to run it that way. Don't forget to make the script file executable!

Minimal

This minimal configuration will simply boot into the bios - there aren't any drives connected so there is nothing else for QEMU to do. However, this allows us to verify that the GPU passthrough is actually working.

FILE MinimalPassthrough.sh
#!/bin/bash

virsh nodedev-detach pci_0000_09_00_0
virsh nodedev-detach pci_0000_09_00_1
qemu-system-x86_64 \
    -nodefaults \
    -enable-kvm \
    -cpu host,kvm=off \
    -m 8G \
    -name "BlankVM" \
    -smp cores=4 \
    -device pcie-root-port,id=pcie.1,bus=pcie.0,addr=1c.0,slot=1,chassis=1,multifunction=on
    -device vfio-pci,host=09:00.0,bus=pcie.1,addr=00.0,x-vga=on,multifunction=on,romfile=GP107_patched.rom \
    -device vfio-pci,host=09:00.1,bus=pcie.1,addr=00.1 \
    -monitor stdio \
    -nographic \
    -vga none \
    $@

virsh nodedev-reattach pci_0000_09_00_0
virsh nodedev-reattach pci_0000_09_00_1

Here's an explanation of each line:

  1. -nodefaults stops qemu from creating some default devices. Specifically, it creates a VGA device by default, which interferes with our attempt to pass through the video card (if you have a multi-video card host this may not be an issue for you)
  2. -enable-kvm enables acceleration
  3. -cpu host, kvm=off \ this makes the virtual machine match the CPU architecture of the host. Not really sure what `kvm=off` does...
  4. -m 8G give the guest 8 gigabytes of RAM
  5. -name "BlankVM" I guess it just gives the virtual machine a name
  6. -smp cores=4 how many cores the guest should have. I'm matching the host.
  7. -device pcie-root-port,id=pcie.1... a dedicate root port other than pcie.0 is required by amd gpu for windows driver
  8. -device vfio-pci,host=09:00.0... add a device using vfio-pci kernel module, from the host's address "09:00.0"
  9. ...addr=.. video must on .0 and audio on .1 while both video and audio must be on the same pci-root-port other than pcie.0
  10. ...x-vga=on this is an option for the vfio-pci module (I think)
  11. ...multifunction=on since our card is doing both audio and video, it needs multifunction (I think...)
  12. ...romfile=GP107_patched.rom due to known issues on NVIDIA cards, it may be necessary to use a modified vbios. This is how you make qemu use that modified vbios.
  13. -device vfio-pci,host=09:00.1 just like above - this is the audio device that is in the same IOMMU group as the video device.
  14. -monitor stdio this will drop you into a qemu "command line" (they call it a monitor) once you launch the VM, allowing you to do things.
  15. -vga none this is probably redundant since we did "nodefaults"

As noted above, there are certain known issues with NVIDIA drivers. I used this tool to patch my vbios, after first downloading my vbios in windows 10 using this gpuz tool.

Linux Guest

Here is a slightly more complicated qemu call, that actually loads a Gentoo VM.

FILE GentooPassthrough.sh
#!/bin/bash

exec qemu-system-x86_64 \
    -nodefaults \
    -enable-kvm \
    -cpu host,kvm=off,hv_vendor_id=1234567890ab \
    -m 8G \
    -name "Gentoo VM" \
    -smp cores=4 \
    -boot order=d \
    -drive file=Gentoo_VM.img,if=virtio \
    -monitor stdio \
    -serial none \
    -net nic \
    -net user,hostfwd=tcp::50000-:22,hostfwd=tcp::50001-:5900,hostname=gentoo_qemu \
    -nographic \
    -vga none \
    -device vfio-pci,host=09:00.0,x-vga=on,multifunction=on,romfile=GP107_patched.rom \
    -device vfio-pci,host=09:00.1 \
    -usb \
    -device usb-host,vendorid=0x1532,productid=0x0101,id=mouse \
    -device usb-host,vendorid=0x04f2,productid=0x0833,id=keyboard \
    $@

Here is an explanation of the new configuration options:

  1. ...hv_vendor_id=... despite the patched vbios, the NVIDIA driver still recognized that it is being run in a virtual machine and refuses to load. This "spoofs" the vendor id (somewhere) and tricks the driver
  2. -boot order=d boot the hard drive first
  3. -drive file=Gentoo_VM.img,if=virtio this is a drive that is emulated in the VM. The "Gentoo_VM.img" file is a qcow QEMU-style virtual drive file.
  4. -serial none actually, I can't remember why I put this in there....
  5. -net nic create a Ethernet in the guest vm
  6. -net user,hostfwd... forwards the ports from host 50000 and 50001 to the guest ports 22 and 5900. Now, from the host, you can ssh into the guest using `ssh -p 50000 myuser@127.0.0.1`, and if you have a vnc server running in the guest on port 5900, you can access it using port 50001 in the host
  7. -nographic this may not be needed if you have a dedicated graphics card for the guest
  8. -usb emulate a USB device on the guest
  9. -device usb-host,... these two lines forward the keyboard and mouse from the host to the guest. The vendorid and productid can be found using lsusb in the host.

Please note that without the `hv_vendor_id` portion, you can boot in and use the console in the guest with the forwarded graphics card. But whenever you launch X, which initialized the proprietary NVIDIA driver, it will fail.


Here is a little variation of the above qemu script for Gentoo host and Gentoo guest. It uses separate CPUs for the guest. Works on a notebook with Ryzen CPU, where the 2nd NVIDIA GPU is passed through to the guest. The guest runs the NVIDIA driver. Installation is performed according to the Gentoo installation guide using UEFI and a GPT partition table. It uses no custom ROMs.

FILE gentooPassthrough.sh
#!/bin/bash

name=genpass
pid="${$}"
cpus="8-15"
ncpus=8
cgrouprootfs="/sys/fs/cgroup"
cgroupfs="${cgrouprootfs}/${name}"

echo "PID: ${pid}"

# using separate CPUs for VM
# cgroup usage see https://www.kernel.org/doc/html/latest/admin-guide/cgroup-v2.html
# 'lscpu -e' to see which cpus to use
echo "+cpuset" > ${cgrouprootfs}/cgroup.subtree_control
mkdir -p ${cgroupfs}
echo ${cpus} > ${cgroupfs}/cpuset.cpus
echo "root" > ${cgroupfs}/cpuset.cpus.partition
echo "${pid}" > ${cgroupfs}/cgroup.procs

# setting performance governor for QEMU CPUs
for i in `seq 8 15` ; do
  echo performance >/sys/devices/system/cpu/cpu${i}/cpufreq/scaling_governor
done

qemu-system-x86_64 \
    -M q35 \
    -monitor stdio \
    -bios /usr/share/edk2-ovmf/OVMF_CODE.fd \
    -accel kvm,kernel-irqchip=on \
    -cpu host,kvm=off \
    -smp ${ncpus} \
    -m 4G \
    -name "${name}" \
    -device vfio-pci,host=01:00.0,multifunction=on \
    -device vfio-pci,host=01:00.1 \
    -nographic \
    -vga none \
    -serial none \
    -parallel none \
    -hda hda.qcow2 \
    -usb \
    -device usb-host,vendorid=0x046D,productid=0xC52B \
    $@ 

# removing cgroup cpuset
echo "${pid}" > ${cgrouprootfs}/cgroup.procs
rmdir ${cgroupfs}

# setting schedutil governor for qemu cpus
for i in `seq 8 15` ; do
  echo schedutil >/sys/devices/system/cpu/cpu${i}/cpufreq/scaling_governor
done

The kernel of the Gentoo host has been build with genkernel --virtio all. The NVIDIA GPU has been bound to vfio-pci with /etc/modprobe.d/local.conf on the host:

FILE /etc/modprobe.d/local.conf
alias pci:v000010DEd00001F95sv0000103Csd000087B2bc03sc00i00 vfio-pci
alias pci:v000010DEd000010FAsv0000103Csd000087B2bc04sc03i00 vfio-pci
options vfio-pci ids=10de:1f95,10de:10fa

This way the internal graphic of the Ryzen processor shows the host on the laptop display, Gentoo guest is displayed on the monitor connected to the HDMI of the NVIDIA graphic. To get sound in the VM, i have to replug the HDMI cable after the VM has booted. Maybe this issue is related to the HDMI cable or the external monitor.

See also

  • QEMU — a generic, open source hardware emulator and virtualization suite.

External resources