Difference between revisions of "PCI passthrough via OVMF"

From ArchWiki
Jump to: navigation, search
m (Passthrough all GPUs but the boot GPU: The previous command would fail with certain bus layouts, because they are nested deeper. This approach should catch everything.)
(Isolating the GPU: Readded the instructions regarding loading vfio-pci as a module, since its inclusion is still very new and FS#46505 seems to have reverted those changes)
Line 108: Line 108:
 
{{Warning|Once you reboot after this procedure, whatever GPU you have configured will no longer be usable on the host until you reverse the manipulation. Make sure the GPU you intend to use on the host is properly configured before doing this - your motherboard should be set to display using the host GPU.}}
 
{{Warning|Once you reboot after this procedure, whatever GPU you have configured will no longer be usable on the host until you reverse the manipulation. Make sure the GPU you intend to use on the host is properly configured before doing this - your motherboard should be set to display using the host GPU.}}
  
Starting with Linux 4.1, the kernel includes vfio-pci. This is a VFIO driver, meaning it fulfills the same role as pci-stub did, but it can also control devices to an extent, such as by switching them into their D3 state when they are not in use. Starting with Linux 4.18.16, vfio-pci is compiled-in as opposed to being a module.
+
Starting with Linux 4.1, the kernel includes vfio-pci. This is a VFIO driver, meaning it fulfills the same role as pci-stub did, but it can also control devices to an extent, such as by switching them into their D3 state when they are not in use.
  
 
Vfio-pci normally targets PCI devices by ID, meaning you only need to specify the IDs of the devices you intend to passthrough. For the following IOMMU group, you would want to bind vfio-pci with {{ic|10de:13c2}} and {{ic|10de:0fbb}}, which will be used as example values for the rest of this section.
 
Vfio-pci normally targets PCI devices by ID, meaning you only need to specify the IDs of the devices you intend to passthrough. For the following IOMMU group, you would want to bind vfio-pci with {{ic|10de:13c2}} and {{ic|10de:0fbb}}, which will be used as example values for the rest of this section.
Line 117: Line 117:
 
{{Note|You cannot specify which device to isolate using vendor-device ID pairs if the host GPU and the guest GPU share the same pair (i.e : if both are the same model). If this is your case, read [[#Using identical guest and host GPUs]] instead.}}
 
{{Note|You cannot specify which device to isolate using vendor-device ID pairs if the host GPU and the guest GPU share the same pair (i.e : if both are the same model). If this is your case, read [[#Using identical guest and host GPUs]] instead.}}
  
You can then add those vendor-device ID pairs to the default kernel parameters passed to vfio-pci at boot time.  
+
{{Note|If, as noted in [[#Plugging your guest GPU in an unisolated CPU-based PCIe slot]], your pci root port is part of your IOMMU group, you '''must not''' pass its ID to {{ic|vfio-pci}}, as it needs to remain attached to the host to function properly. Any other device within that group, however, should be left for {{ic|vfio-pci}} to bind with.}}
 +
 
 +
=== With vfio-pci built into the kernel (4.18.16.arch1 onwards) ===
 +
Starting with version 4.18.16, the stock Arch Linux kernel comes with vfio-pci compiled within, as opposed to having it compiled it alongside the kernel but leaving it to be loaded as a module. This means all that should be required to isolate the GPU is to pass the device IDs as [[kernel parameter]]s, like so:
  
 
{{hc|/etc/default/grub|2=
 
{{hc|/etc/default/grub|2=
Line 123: Line 126:
 
}}
 
}}
  
{{Note|If, as noted in [[#Plugging your guest GPU in an unisolated CPU-based PCIe slot]], your pci root port is part of your IOMMU group, you '''must not''' pass its ID to {{ic|vfio-pci}}, as it needs to remain attached to the host to function properly. Any other device within that group, however, should be left for {{ic|vfio-pci}} to bind with.}}
+
In the case of GRUB, since the configuration has been changed, its config file must also be regenerated.
 +
 
 +
  # grub-mkconfig -o /boot/grub/grub.cfg
 +
 
 +
=== With vfio-pci loaded as a module ===
 +
If your kernel image does not include vfio-pci as a built-in module, the configuration differs slightly. First, the vendor-device ID pairs must be specified as default parameters passed to vfio-pci whenever it is inserted into the kernel.
 +
 
 +
{{hc|/etc/modprobe.d/vfio.conf|2=
 +
options vfio-pci ids=10de:13c2,10de:0fbb
 +
}}
 +
 
 +
This, however, does not guarantee that vfio-pci will be loaded before other graphics drivers. To ensure that, we need to statically bind it in the kernel image alongside with its dependencies. That means adding, in this order, {{ic|vfio_pci}}, {{ic|vfio}}, {{ic|vfio_iommu_type1}}, and {{ic|vfio_virqfd}} to [[mkinitcpio]]:
 +
 
 +
{{hc|/etc/mkinitcpio.conf|2=
 +
MODULES=(... vfio_pci vfio vfio_iommu_type1 vfio_virqfd ...)
 +
}}
 +
 
 +
{{Note|If you also have another driver loaded this way for [[Kernel mode setting#Early KMS start|early modesetting]] (such as {{ic|nouveau}}, {{ic|radeon}}, {{ic|amdgpu}}, {{ic|i915}}, etc.), all of the aforementioned VFIO modules must precede it.}}
  
Since the grub configuration has been changed, its config file has to be regenerated.
+
Also, ensure that the modconf hook is included in the HOOKS list of {{ic|mkinitcpio.conf}}:
 +
{{hc|/etc/mkinitcpio.conf|2=
 +
HOOKS=(... modconf ...)
 +
}}
  
  # grub-mkconfig -o /boot/grub/grub.cfg
+
Since new modules have been added to the initramfs configuration, you must [[regenerate the initramfs]]. Should you change the IDs of the devices in {{ic|/etc/modprobe.d/vfio.conf}}, you will also have to regenerate it, as those parameters must be specified in the initramfs to be known during the early boot stages.
  
 +
=== Verifying that the configuration worked ===
 
Reboot and verify that vfio-pci has loaded properly and that it is now bound to the right devices.
 
Reboot and verify that vfio-pci has loaded properly and that it is now bound to the right devices.
  

Revision as of 02:53, 10 November 2018

The Open Virtual Machine Firmware (OVMF) is a project to enable UEFI support for virtual machines. Starting with Linux 3.9 and recent versions of QEMU, it is now possible to passthrough a graphics card, offering the VM native graphics performance which is useful for graphic-intensive tasks.

Provided you have a desktop computer with a spare GPU you can dedicate to the host (be it an integrated GPU or an old OEM card, the brands do not even need to match) and that your hardware supports it (see #Prerequisites), it is possible to have a VM of any OS with its own dedicated GPU and near-native performance. For more information on techniques see the background presentation (pdf).

Contents

Prerequisites

A VGA Passthrough relies on a number of technologies that are not ubiquitous as of today and might not be available on your hardware. You will not be able to do this on your machine unless the following requirements are met :

  • Your CPU must support hardware virtualization (for kvm) and IOMMU (for the passthrough itself)
  • Your motherboard must also support IOMMU
  • Your guest GPU ROM must support UEFI.
    • If you can find any ROM in this list that applies to your specific GPU and is said to support UEFI, you are generally in the clear. All GPUs from 2012 and later should support this, as Microsoft made UEFI a requirement for devices to be marketed as compatible with Windows 8.

You will probably want to have a spare monitor or one with multiple input ports connected to different GPUs (the passthrough GPU will not display anything if there is no screen plugged in and using a VNC or Spice connection will not help your performance), as well as a mouse and a keyboard you can pass to your VM. If anything goes wrong, you will at least have a way to control your host machine this way.

Setting up IOMMU

Note:
  • IOMMU is a generic name for Intel VT-d and AMD-Vi.
  • VT-d stands for Intel Virtualization Technology for Directed I/O and should not be confused with VT-x Intel Virtualization Technology. VT-x allows one hardware platform to function as multiple “virtual” platforms while VT-d improves security and reliability of the systems and also improves performance of I/O devices in virtualized environments.

Using IOMMU opens to features like PCI passthrough and memory protection from faulty or malicious devices, see Wikipedia:Input-output memory management unit#Advantages and Memory Management (computer programming): Could you explain IOMMU in plain English?.

Enabling IOMMU

Ensure that AMD-Vi/Intel VT-d is supported by the CPU and enabled in the BIOS settings. Both normally show up alongside other CPU features (meaning they could be in an overclocking-related menu) either with their actual names ("VT-d" or "AMD-Vi") or in more ambiguous terms such as "Virtualization technology", which may or may not be explained in the manual.

Enable IOMMU support by setting the correct kernel parameter depending on the type of CPU in use:

  • For Intel CPUs (VT-d) set intel_iommu=on
  • For AMD CPUs (AMD-Vi) set amd_iommu=on

You should also append the iommu=pt parameter. This will prevent Linux from touching devices which cannot be passed through.

After rebooting, check dmesg to confirm that IOMMU has been correctly enabled:

dmesg | grep -e DMAR -e IOMMU
[    0.000000] ACPI: DMAR 0x00000000BDCB1CB0 0000B8 (v01 INTEL  BDW      00000001 INTL 00000001)
[    0.000000] Intel-IOMMU: enabled
[    0.028879] dmar: IOMMU 0: reg_base_addr fed90000 ver 1:0 cap c0000020660462 ecap f0101a
[    0.028883] dmar: IOMMU 1: reg_base_addr fed91000 ver 1:0 cap d2008c20660462 ecap f010da
[    0.028950] IOAPIC id 8 under DRHD base  0xfed91000 IOMMU 1
[    0.536212] DMAR: No ATSR found
[    0.536229] IOMMU 0 0xfed90000: using Queued invalidation
[    0.536230] IOMMU 1 0xfed91000: using Queued invalidation
[    0.536231] IOMMU: Setting RMRR:
[    0.536241] IOMMU: Setting identity map for device 0000:00:02.0 [0xbf000000 - 0xcf1fffff]
[    0.537490] IOMMU: Setting identity map for device 0000:00:14.0 [0xbdea8000 - 0xbdeb6fff]
[    0.537512] IOMMU: Setting identity map for device 0000:00:1a.0 [0xbdea8000 - 0xbdeb6fff]
[    0.537530] IOMMU: Setting identity map for device 0000:00:1d.0 [0xbdea8000 - 0xbdeb6fff]
[    0.537543] IOMMU: Prepare 0-16MiB unity mapping for LPC
[    0.537549] IOMMU: Setting identity map for device 0000:00:1f.0 [0x0 - 0xffffff]
[    2.182790] [drm] DMAR active, disabling use of stolen memory

Ensuring that the groups are valid

The following script should allow you to see how your various PCI devices are mapped to IOMMU groups. If it does not return anything, you either have not enabled IOMMU support properly or your hardware does not support it.

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

Example output:

IOMMU Group 0 00:00.0 Host bridge [0600]: Intel Corporation 2nd Generation Core Processor Family DRAM Controller [8086:0104] (rev 09)
IOMMU Group 1 00:16.0 Communication controller [0780]: Intel Corporation 6 Series/C200 Series Chipset Family MEI Controller #1 [8086:1c3a] (rev 04)
IOMMU Group 2 00:19.0 Ethernet controller [0200]: Intel Corporation 82579LM Gigabit Network Connection [8086:1502] (rev 04)
IOMMU Group 3 00:1a.0 USB controller [0c03]: Intel Corporation 6 Series/C200 Series Chipset Family USB Enhanced Host Controller #2 [8086:1c2d] (rev  
...

An IOMMU group is the smallest set of physical devices that can be passed to a virtual machine. For instance, in the example above, both the GPU in 06:00.0 and its audio controller in 6:00.1 belong to IOMMU group 13 and can only be passed together. The frontal USB controller, however, has its own group (group 2) which is separate from both the USB expansion controller (group 10) and the rear USB controller (group 4), meaning that any of them could be passed to a VM without affecting the others.

Gotchas

Plugging your guest GPU in an unisolated CPU-based PCIe slot

Not all PCI-E slots are the same. Most motherboards have PCIe slots provided by both the CPU and the PCH. Depending on your CPU, it is possible that your processor-based PCIe slot does not support isolation properly, in which case the PCI slot itself will appear to be grouped with the device that is connected to it.

IOMMU Group 1 00:01.0 PCI bridge: Intel Corporation Xeon E3-1200 v2/3rd Gen Core processor PCI Express Root Port (rev 09)
IOMMU Group 1 01:00.0 VGA compatible controller: NVIDIA Corporation GM107 [GeForce GTX 750] (rev a2)
IOMMU Group 1 01:00.1 Audio device: NVIDIA Corporation Device 0fbc (rev a1)

This is fine so long as only your guest GPU is included in here, such as above. Depending on what is plugged in to your other PCIe slots and whether they are allocated to your CPU or your PCH, you may find yourself with additional devices within the same group, which would force you to pass those as well. If you are ok with passing everything that is in there to your VM, you are free to continue. Otherwise, you will either need to try and plug your GPU in your other PCIe slots (if you have any) and see if those provide isolation from the rest or to install the ACS override patch, which comes with its own drawbacks. See #Bypassing the IOMMU groups (ACS override patch) for more information.

Note: If they are grouped with other devices in this manner, pci root ports and bridges should neither be bound to vfio at boot, nor be added to the VM.

Isolating the GPU

In order to assign a device to a virtual machine, this device and all those sharing the same IOMMU group must have their driver replaced by a stub driver or a VFIO driver in order to prevent the host machine from interacting with them. In the case of most devices, this can be done on the fly right before the VM starts.

However, due to their size and complexity, GPU drivers do not tend to support dynamic rebinding very well, so you cannot just have some GPU you use on the host be transparently passed to a VM without having both drivers conflict with each other. Because of this, it is generally advised to bind those placeholder drivers manually before starting the VM, in order to stop other drivers from attempting to claim it.

The following section details how to configure a GPU so those placeholder drivers are bound early during the boot process, which makes said device inactive until a VM claims it or the driver is switched back. This is the preferred method, considering it has less caveats than switching drivers once the system is fully online.

Warning: Once you reboot after this procedure, whatever GPU you have configured will no longer be usable on the host until you reverse the manipulation. Make sure the GPU you intend to use on the host is properly configured before doing this - your motherboard should be set to display using the host GPU.

Starting with Linux 4.1, the kernel includes vfio-pci. This is a VFIO driver, meaning it fulfills the same role as pci-stub did, but it can also control devices to an extent, such as by switching them into their D3 state when they are not in use.

Vfio-pci normally targets PCI devices by ID, meaning you only need to specify the IDs of the devices you intend to passthrough. For the following IOMMU group, you would want to bind vfio-pci with 10de:13c2 and 10de:0fbb, which will be used as example values for the rest of this section.

IOMMU Group 13 06:00.0 VGA compatible controller: NVIDIA Corporation GM204 [GeForce GTX 970] [10de:13c2] (rev a1)
IOMMU Group 13 06:00.1 Audio device: NVIDIA Corporation GM204 High Definition Audio Controller [10de:0fbb] (rev a1)}}
Note: You cannot specify which device to isolate using vendor-device ID pairs if the host GPU and the guest GPU share the same pair (i.e : if both are the same model). If this is your case, read #Using identical guest and host GPUs instead.
Note: If, as noted in #Plugging your guest GPU in an unisolated CPU-based PCIe slot, your pci root port is part of your IOMMU group, you must not pass its ID to vfio-pci, as it needs to remain attached to the host to function properly. Any other device within that group, however, should be left for vfio-pci to bind with.

With vfio-pci built into the kernel (4.18.16.arch1 onwards)

Starting with version 4.18.16, the stock Arch Linux kernel comes with vfio-pci compiled within, as opposed to having it compiled it alongside the kernel but leaving it to be loaded as a module. This means all that should be required to isolate the GPU is to pass the device IDs as kernel parameters, like so:

/etc/default/grub
GRUB_CMDLINE_LINUX_DEFAULT=... vfio-pci.ids=10de:13c2,10de:0fbb

In the case of GRUB, since the configuration has been changed, its config file must also be regenerated.

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

With vfio-pci loaded as a module

If your kernel image does not include vfio-pci as a built-in module, the configuration differs slightly. First, the vendor-device ID pairs must be specified as default parameters passed to vfio-pci whenever it is inserted into the kernel.

/etc/modprobe.d/vfio.conf
options vfio-pci ids=10de:13c2,10de:0fbb

This, however, does not guarantee that vfio-pci will be loaded before other graphics drivers. To ensure that, we need to statically bind it in the kernel image alongside with its dependencies. That means adding, in this order, vfio_pci, vfio, vfio_iommu_type1, and vfio_virqfd to mkinitcpio:

/etc/mkinitcpio.conf
MODULES=(... vfio_pci vfio vfio_iommu_type1 vfio_virqfd ...)
Note: If you also have another driver loaded this way for early modesetting (such as nouveau, radeon, amdgpu, i915, etc.), all of the aforementioned VFIO modules must precede it.

Also, ensure that the modconf hook is included in the HOOKS list of mkinitcpio.conf:

/etc/mkinitcpio.conf
HOOKS=(... modconf ...)

Since new modules have been added to the initramfs configuration, you must regenerate the initramfs. Should you change the IDs of the devices in /etc/modprobe.d/vfio.conf, you will also have to regenerate it, as those parameters must be specified in the initramfs to be known during the early boot stages.

Verifying that the configuration worked

Reboot and verify that vfio-pci has loaded properly and that it is now bound to the right devices.

$ dmesg | grep -i vfio
[    0.329224] VFIO - User Level meta-driver version: 0.3
[    0.341372] vfio_pci: add [10de:13c2[ffff:ffff]] class 0x000000/00000000
[    0.354704] vfio_pci: add [10de:0fbb[ffff:ffff]] class 0x000000/00000000
[    2.061326] vfio-pci 0000:06:00.0: enabling device (0100 -> 0103)

It is not necessary for all devices (or even expected device) from vfio.conf to be in dmesg output. Sometimes a device does not appear in output at boot but actually is able to be visible and operatable in guest VM.

$ lspci -nnk -d 10de:13c2
06:00.0 VGA compatible controller: NVIDIA Corporation GM204 [GeForce GTX 970] [10de:13c2] (rev a1)
	Kernel driver in use: vfio-pci
	Kernel modules: nouveau nvidia
$ lspci -nnk -d 10de:0fbb
06:00.1 Audio device: NVIDIA Corporation GM204 High Definition Audio Controller [10de:0fbb] (rev a1)
	Kernel driver in use: vfio-pci
	Kernel modules: snd_hda_intel

Setting up an OVMF-based guest VM

OVMF is an open-source UEFI firmware for QEMU virtual machines. While it is possible to use SeaBIOS to get similar results to an actual PCI passthough, the setup process is different and it is generally preferable to use the EFI method if your hardware supports it.

Configuring libvirt

Libvirt is a wrapper for a number of virtualization utilities that greatly simplifies the configuration and deployment process of virtual machines. In the case of KVM and QEMU, the frontend it provides allows us to avoid dealing with the permissions for QEMU and make it easier to add and remove various devices on a live VM. Its status as a wrapper, however, means that it might not always support all of the latest qemu features, which could end up requiring the use of a wrapper script to provide some extra arguments to QEMU.

After installing qemu, qemu-block-iscsi, libvirt, ovmf, and virt-manager, add the path to your OVMF firmware image and runtime variables template to your libvirt config so virt-install or virt-manager can find those later on.

/etc/libvirt/qemu.conf
nvram = [
	"/usr/share/ovmf/x64/OVMF_CODE.fd:/usr/share/ovmf/x64/OVMF_VARS.fd"
]

You can now enable and start libvirtd.service and its logging component virtlogd.socket.

Setting up the guest OS

The process of setting up a VM using virt-manager is mostly self-explanatory, as most of the process comes with fairly comprehensive on-screen instructions.

If using virt-manager, you have to add your user to the libvirt user group to ensure authentication.

However, you should pay special attention to the following steps :

  • When the VM creation wizard asks you to name your VM (final step before clicking "Finish"), check the "Customize before install" checkbox.
  • In the "Overview" section, set your firmware to "UEFI". If the option is grayed out, make sure that you have correctly specified the location of your firmware in /etc/libvirt/qemu.conf and restart libvirtd.service.
  • In the "CPUs" section, change your CPU model to "host-passthrough". If it is not in the list, you will have to type it by hand. This will ensure that your CPU is detected properly, since it causes libvirt to expose your CPU capabilities exactly as they are instead of only those it recognizes (which is the preferred default behavior to make CPU behavior easier to reproduce). Without it, some applications may complain about your CPU being of an unknown model.
  • If you want to minimize IO overhead, go into "Add Hardware" and add a Controller for SCSI drives of the "VirtIO SCSI" model. You can then change the default IDE disk for a SCSI disk, which will bind to said controller.
    • Windows VMs will not recognize those drives by default, so you need to download the ISO containing the drivers from here and add an IDE (or SATA for Windows 8.1 and newer) CD-ROM storage device linking to said ISO, otherwise you will not be able to get Windows to recognize it during the installation process. When prompted to select a disk to install windows on, load the drivers contained on the CD-ROM under vioscsi.

The rest of the installation process will take place as normal using a standard QXL video adapter running in a window. At this point, there is no need to install additional drivers for the rest of the virtual devices, since most of them will be removed later on. Once the guest OS is done installing, simply turn off the virtual machine. It is possible you will be dropped into the UEFI menu instead of starting the installation upon powering your VM for the first time. Sometimes the correct ISO file was not automatically detected and you will need to manually specify the drive to boot. By typing exit and navigating to "boot manager" you will enter a menu that allows you to choose between devices.

Attaching the PCI devices

With the installation done, it is now possible to edit the hardware details in libvirt and remove virtual integration devices, such as the spice channel and virtual display, the QXL video adapter, the emulated mouse and keyboard and the USB tablet device. Since that leaves you with no input devices, you may want to bind a few USB host devices to your VM as well, but remember to leave at least one mouse and/or keyboard assigned to your host in case something goes wrong with the guest. At this point, it also becomes possible to attach the PCI device that was isolated earlier; simply click on "Add Hardware" and select the PCI Host Devices you want to passthrough. If everything went well, the screen plugged into your GPU should show the OVMF splash screen and your VM should start up normally. From there, you can setup the drivers for the rest of your VM.

Gotchas

Using a non-EFI image on an OVMF-based VM

The OVMF firmware does not support booting off non-EFI mediums. If the installation process drops you in a UEFI shell right after booting, you may have an invalid EFI boot media. Try using an alternate linux/windows image to determine if you have an invalid media.

Performance tuning

Most use cases for PCI passthroughs relate to performance-intensive domains such as video games and GPU-accelerated tasks. While a PCI passthrough on its own is a step towards reaching native performance, there are still a few ajustments on the host and guest to get the most out of your VM.

CPU pinning

The default behavior for KVM guests is to run operations coming from the guest as a number of threads representing virtual processors. Those threads are managed by the Linux scheduler like any other thread and are dispatched to any available CPU cores based on niceness and priority queues. As such, the local CPU cache benefits (L1/L2) are lost each time the host scheduler reschedules the virtual CPU thread on a different physical CPU. This can noticeably harm performance on the guest. CPU pinning aims to resolve this by limiting which physical CPUs the virtual CPUs are allowed to run on. The ideal setup is a one to one mapping such that the virtual CPU cores match physical CPU cores while taking hyperthreading/SMT into account.

Note: For certain users enabling CPU pinning may introduce stuttering and short hangs, especially with the MuQSS scheduler (present in linux-ck and linux-zen kernels). You might want to try disabling pinning first if you experience similar issues, which effectively trades maximum performance for responsiveness at all times.

CPU topology

Most modern CPU's support hardware multitasking, also known as hyper-threading on Intel CPU's or SMT on AMD CPU's. Hyper-threading/SMT is simply a very efficient way of running two threads on one CPU core at any given time. You will want to take into consideration that the CPU pinning you choose will greatly depend on what you do with your host while your VM is running.

To find the topology for your CPU run lscpu -e:

Note: Pay special attention to the 4th column "CORE" as this shows the association of the Physical/Logical CPU cores

lscpu -e on a 6c/12t Ryzen 5 1600:

CPU NODE SOCKET CORE L1d:L1i:L2:L3 ONLINE MAXMHZ    MINMHZ
0   0    0      0    0:0:0:0       yes    3800.0000 1550.0000
1   0    0      0    0:0:0:0       yes    3800.0000 1550.0000
2   0    0      1    1:1:1:0       yes    3800.0000 1550.0000
3   0    0      1    1:1:1:0       yes    3800.0000 1550.0000
4   0    0      2    2:2:2:0       yes    3800.0000 1550.0000
5   0    0      2    2:2:2:0       yes    3800.0000 1550.0000
6   0    0      3    3:3:3:1       yes    3800.0000 1550.0000
7   0    0      3    3:3:3:1       yes    3800.0000 1550.0000
8   0    0      4    4:4:4:1       yes    3800.0000 1550.0000
9   0    0      4    4:4:4:1       yes    3800.0000 1550.0000
10  0    0      5    5:5:5:1       yes    3800.0000 1550.0000
11  0    0      5    5:5:5:1       yes    3800.0000 1550.0000

lscpu -e on a 6c/12t Intel 8700k:

CPU NODE SOCKET CORE L1d:L1i:L2:L3 ONLINE MAXMHZ    MINMHZ
0   0    0      0    0:0:0:0       yes    4600.0000 800.0000
1   0    0      1    1:1:1:0       yes    4600.0000 800.0000
2   0    0      2    2:2:2:0       yes    4600.0000 800.0000
3   0    0      3    3:3:3:0       yes    4600.0000 800.0000
4   0    0      4    4:4:4:0       yes    4600.0000 800.0000
5   0    0      5    5:5:5:0       yes    4600.0000 800.0000
6   0    0      0    0:0:0:0       yes    4600.0000 800.0000
7   0    0      1    1:1:1:0       yes    4600.0000 800.0000
8   0    0      2    2:2:2:0       yes    4600.0000 800.0000
9   0    0      3    3:3:3:0       yes    4600.0000 800.0000
10  0    0      4    4:4:4:0       yes    4600.0000 800.0000
11  0    0      5    5:5:5:0       yes    4600.0000 800.0000

As we see above, with AMD Core 0 is sequential with CPU 0 & 1, whereas Intel places Core 0 on CPU 0 & 6.

If you don't need all cores for the guest, it would then be preferable to leave at the very least one core for the host. Choosing which cores one to use for the host or guest should be based on the specific hardware characteristics of your CPU, however Core 0 is a good choice for the host in most cases. If any cores are reserved for the host, it is recommended to pin the emulator and iothreads, if used, to the host cores rather than the VCPUs. This may improve performance and reduce latency for the guest since those threads will not pollute the cache or contend for scheduling with the guest VCPU threads. If all cores are passed to the guest, there is no need or benefit to pinning the emulator or iothreads.


XML examples

Note: Do not use the iothread lines from the XML examples shown below if you have not added an iothread to your disk controller. iothread's only work on virtio-scsi or virtio-blk devices.
4c/1t CPU w/o Hyperthreading Example
$ virsh edit [vmname]
...
<vcpu placement='static'>4</vcpu>
<cputune>
    <vcpupin vcpu='0' cpuset='0'/>
    <vcpupin vcpu='1' cpuset='1'/>
    <vcpupin vcpu='2' cpuset='2'/>
    <vcpupin vcpu='3' cpuset='3'/>
</cputune>
...
6c/2t Intel CPU pinning example
$ virsh edit [vmname]
...
<vcpu placement='static'>8</vcpu>
<iothreads>1</iothreads>
<cputune>
    <vcpupin vcpu='0' cpuset='2'/>
    <vcpupin vcpu='1' cpuset='8'/>
    <vcpupin vcpu='2' cpuset='3'/>
    <vcpupin vcpu='3' cpuset='9'/>
    <vcpupin vcpu='4' cpuset='4'/>
    <vcpupin vcpu='5' cpuset='10'/>
    <vcpupin vcpu='6' cpuset='5'/>
    <vcpupin vcpu='7' cpuset='11'/>
    <emulatorpin cpuset='0,6'/>
    <iothreadpin iothread='1' cpuset='0,6'/>
</cputune>
    ...
    <topology sockets='1' cores='4' threads='2'/>
    ...
4c/2t AMD CPU example
$ virsh edit [vmname]
...
<vcpu placement='static'>8</vcpu>
<iothreads>1</iothreads>
<cputune>
  <vcpupin vcpu='0' cpuset='2'/>
  <vcpupin vcpu='1' cpuset='3'/>
  <vcpupin vcpu='2' cpuset='4'/>
  <vcpupin vcpu='3' cpuset='5'/>
  <vcpupin vcpu='4' cpuset='6'/>
  <vcpupin vcpu='5' cpuset='7'/>
  <vcpupin vcpu='6' cpuset='8'/>
  <vcpupin vcpu='7' cpuset='9'/>
  <emulatorpin cpuset='0-1'/>
  <iothreadpin iothread='1' cpuset='0-1'/>
</cputune>
    ...
    <topology sockets='1' cores='4' threads='2'/>
    ...
Note: If further CPU isolation is needed, consider using the isolcpus kernel command-line parameter on the unused physical/logical cores.

If you do not intend to be doing any computation-heavy work on the host (or even anything at all) at the same time as you would on the VM, you may want to pin your VM threads across all of your cores, so that the VM can fully take advantage of the spare CPU time the host has available. Be aware that pinning all physical and logical cores of your CPU could induce latency in the guest VM.

Huge memory pages

When dealing with applications that require large amounts of memory, memory latency can become a problem since the more memory pages are being used, the more likely it is that this application will attempt to access information across multiple memory "pages", which is the base unit for memory allocation. Resolving the actual address of the memory page takes multiple steps, and so CPUs normally cache information on recently used memory pages to make subsequent uses on the same pages faster. Applications using large amounts of memory run into a problem where, for instance, a virtual machine uses 4 GiB of memory divided into 4 KiB pages (which is the default size for normal pages) for a total of 1.04 million pages, meaning that such cache misses can become extremely frequent and greatly increase memory latency. Huge pages exist to mitigate this issue by giving larger individual pages to those applications, increasing the odds that multiple operations will target the same page in succession.

Transparent huge pages

QEMU will use 2MiB sized transparent huge pages automatically without any explicit configuration in QEMU or Libvirt, subject to some important caveats. When using VFIO the pages are locked in at boot time and transparent huge pages are allocated up front when the VM first boots. If the kernel memory is highly fragmented, or the VM is using a majority of the remaining free memory, it is likely that the kernel will not have enough 2MiB pages to fully satisfy the allocation. In such a case, it silently fails by using a mix of 2MiB and 4KiB pages. Since the pages are locked in VFIO mode, the kernel will not be able to convert those 4KiB pages to huge after the VM starts either. The number of available 2MiB huge pages available to THP is the same as via the #Dynamic huge pages mechanism described in the following sections.

To check how much memory THP is using globally:

$ grep AnonHugePages /proc/meminfo
AnonHugePages:   8091648 kB

To check a specific QEMU instance. QEMU's PID must be substituted in the grep command:

$ grep -P 'AnonHugePages:\s+(?!0)\d+' /proc/[PID]/smaps
AnonHugePages:   8087552 kB

In this example, the VM was allocated 8388608KiB of memory, but only 8087552KiB was available via THP. The remaining 301056KiB are allocated as 4KiB pages. Aside from manually checking, there is no indication when partial allocations occur. As such, THP's effectiveness is very much dependent on the host system's memory fragmentation at the time of VM startup. If this trade off is unacceptable or strict guarantees are required, #Static huge pages is recommended.

Arch kernels have THP compiled in and enabled by default with /sys/kernel/mm/transparent_hugepage/enabled set to madvise mode.

Static huge pages

While transparent huge pages should work in the vast majority of cases, they can also be allocated statically during boot. This should only be needed to make use 1 GiB hugepages on machines that support it, since transparent huge pages normally only go up to 2 MiB.

Warning: Static huge pages lock down the allocated amount of memory, making it unavailable for applications that are not configured to use them. Allocating 4 GiBs worth of huge pages on a machine with 8 GiB of memory will only leave you with 4 GiB of available memory on the host even when the VM is not running.

To allocate huge pages at boot, one must simply specify the desired amount on their kernel command line with hugepages=x. For instance, reserving 1024 pages with hugepages=1024 and the default size of 2048 KiB per huge page creates 2 GiB worth of memory for the virtual machine to use.

If supported by CPU page size could be set manually. 1 GiB huge page support could be verified by grep pdpe1gb /proc/cpuinfo. Setting 1 GiB huge page size via kernel parameters : default_hugepagesz=1G hugepagesz=1G hugepages=X.

Also, since static huge pages can only be used by applications that specifically request it, you must add this section in your libvirt domain configuration to allow kvm to benefit from them :

$ virsh edit [vmname]
...
<memoryBacking>
	<hugepages/>
</memoryBacking>
...

Dynamic huge pages

Tango-inaccurate.pngThe factual accuracy of this article or section is disputed.Tango-inaccurate.png

Reason: Need futher testing if this variant as effective as static one (Discuss in Talk:PCI passthrough via OVMF#)

Hugepages could be allocated manually via vm.nr_overcommit_hugepages sysctl parameter.

/etc/sysctl.d/10-kvm.conf
vm.nr_hugepages = 0
vm.nr_overcommit_hugepages = num

Where num - is the number of huge pages, which default size if 2 MiB. Pages will be automatically allocated, and freed after VM stops.

More manual way:

# echo num > /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages
# echo num > /sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages

For 2 MiB and 1 GiB page size respectively. And they should be manually freed in the same way.

It is hardly recommended to drop caches, compact memory and wait couple of seconds before starting VM, as there could be not enough free contiguous memory for required huge pages blocks. Especially after some uptime of the host system.

# echo 3 > /proc/sys/vm/drop_caches
# echo 1 > /proc/sys/vm/compact_memory

Theoretically, 1 GiB pages works as 2 MiB. But practically - no guaranteed way was found to get contiguous 1 GiB memory blocks. Each consequent request of 1 GiB blocks lead to lesser and lesser dynamically allocated count.

CPU frequency governor

Depending on the way your CPU governor is configured, the VM threads may not hit the CPU load thresholds for the frequency to ramp up. Indeed, KVM cannot actually change the CPU frequency on its own, which can be a problem if it does not scale up with vCPU usage as it would result in underwhelming performance. An easy way to see if it behaves correctly is to check if the frequency reported by watch lscpu goes up when running a CPU-intensive task on the guest. If you are indeed experiencing stutter and the frequency does not go up to reach its reported maximum, it may be due to cpu scaling being controlled by the host OS. In this case, try setting all cores to maximum frequency to see if this improves performance. Note that if you are using a modern intel chip with the default pstate driver, cpupower commands will be ineffective, so monitor /proc/cpuinfo to make sure your cpu is actually at max frequency.

CPU pinning with isolcpus

Alternatively, make sure that you have isolated CPUs properly. In this example, let us assume you are using CPUs 4-7. Use the kernel parameters isolcpus nohz_full rcu_nocbs to completely isolate the CPUs from the kernel.

/etc/defaults/grub
...
GRUB_CMDLINE_LINUX="..your other params.. isolcpus=4-7 nohz_full=4-7 rcu_nocbs=4-7"
...
</nowiki>

Then, run qemu-system-x86_64 with taskset and chrt:

# chrt -r 1 taskset -c 4-7 qemu-system-x86_64 ...

The chrt command will ensure that the task scheduler will round-robin distribute work (otherwise it will all stay on the first cpu). For taskset, the CPU numbers can be comma- and/or dash-separated, like "0,1,2,3" or "0-4" or "1,7-8,10" etc.

See this reddit comment for more info.

Improving performance on AMD CPUs

Previously, Nested Page Tables (NPT) had to be disabled on AMD systems running KVM to improve GPU performance because of a very old bug, but the trade off was decreased CPU performance, including stuttering.

There is a kernel patch that resolves this issue, which was accepted into kernel 4.14-stable and 4.9-stable. If you are running the official linux or linux-lts kernel the patch has already been applied (make sure you are on the latest). If you are running another kernel you might need to manually patch yourself.

Note: Several Ryzen users (see this Reddit thread) have tested the patch, and can confirm that it works, bringing GPU passthrough performance up to near native quality.

Starting with QEMU 3.1 the TOPOEXT cpuid flag is disabled by default. In order to use hyperthreading(SMT) on AMD CPU's you need to manually enable it:

 <cpu mode='host-passthrough' check='none'>
 <topology sockets='1' cores='4' threads='2'/>
 <feature policy='require' name='topoext'/>
 </cpu>

commit: https://git.qemu.org/?p=qemu.git;a=commit;h=7210a02c58572b2686a3a8d610c6628f87864aed

Further tuning

More specialized VM tuning tips are available at Red Hat's Virtualization Tuning and Optimization Guide. This guide may be especially helpful if you are experiencing:

  • Moderate CPU load on the host during downloads/uploads from within the guest: See Bridge Zero Copy Transmit for a potential fix.
  • Guests capping out at certain network speeds during downloads/uploads despite virtio-net being used: See Multi-queue virtio-net for a potential fix.
  • Guests "stuttering" under high I/O, despite the same workload not affecting hosts to the same degree: See Multi-queue virtio-scsi for a potential fix.

Special procedures

Certain setups require specific configuration tweaks in order to work properly. If you are having problems getting your host or your VM to work properly, see if your system matches one of the cases below and try adjusting your configuration accordingly.

Using identical guest and host GPUs

Tango-view-fullscreen.pngThis article or section needs expansion.Tango-view-fullscreen.png

Reason: A number of users have been having issues with this, it should probably be adressed by the article. (Discuss in Talk:PCI passthrough via OVMF#Additionnal sections)

Due to how vfio-pci uses your vendor and device id pair to identify which device they need to bind to at boot, if you have two GPUs sharing such an ID pair you will not be able to get your passthough driver to bind with just one of them. This sort of setup makes it necessary to use a script, so that whichever driver you are using is instead assigned by pci bus address using the driver_override mechanism.

Script variants

Passthrough all GPUs but the boot GPU

Here, we will make a script to bind vfio-pci to all GPUs but the boot gpu. Create the script /usr/bin/vfio-pci-override.sh:

#!/bin/sh

for i in /sys/bus/pci/devices/*/boot_vga; do
	if [ $(cat "$i") -eq 0 ]; then
		GPU="${i%/boot_vga}"
		AUDIO="$(echo "$GPU" | sed -e "s/0$/1/")"
		echo "vfio-pci" > "$GPU/driver_override"
		if [ -d "$AUDIO" ]; then
			echo "vfio-pci" > "$AUDIO/driver_override"
		fi
	fi
done

modprobe -i vfio-pci
Passthrough selected GPU

In this case we manually specify the GPU to bind.

#!/bin/sh

GROUP="0000:00:03.0"
DEVS="0000:03:00.0 0000:03:00.1 ."

if [ ! -z "$(ls -A /sys/class/iommu)" ]; then
	for DEV in $DEVS; do
		echo "vfio-pci" > /sys/bus/pci/devices/$GROUP/$DEV/driver_override
	done
fi

modprobe -i vfio-pci

Script installation

Note: This will not work as of Linux 4.18.16, since vfio-pci is compiled-in. Has to be done through an initcpio hook.

Create /etc/modprobe.d/vfio.conf with the following:

install vfio-pci /usr/bin/vfio-pci-override.sh

Edit /etc/mkinitcpio.conf

Remove any video drivers from MODULES, and add vfio-pci, and vfio_iommu_type1

MODULES=(ext4 vfat vfio-pci vfio_iommu_type1)

Add /etc/modprobe.d/vfio.conf and /usr/bin/vfio-pci-override.sh to FILES:

FILES=(/etc/modprobe.d/vfio.conf /usr/bin/vfio-pci-override.sh)

Regenerate the initramfs and reboot:

Passing the boot GPU to the guest

Tango-view-fullscreen.pngThis article or section needs expansion.Tango-view-fullscreen.png

Reason: This is related to VBIOS issues and should be moved into a separate section regarding VBIOS compatibility. (Discuss in Talk:PCI passthrough via OVMF#UEFI (OVMF) Compatibility in VBIOS)

The GPU marked as boot_vga is a special case when it comes to doing PCI passthroughs, since the BIOS needs to use it in order to display things like boot messages or the BIOS configuration menu. To do that, it makes a copy of the VGA boot ROM which can then be freely modified. This modified copy is the version the system gets to see, which the passthrough driver may reject as invalid. As such, it is generally recommended to change the boot GPU in the BIOS configuration so the host GPU is used instead or, if that is not possible, to swap the host and guest cards in the machine itself.

Using Looking Glass to stream guest screen to the host

It is possible to make VM share the monitor, and optionally a keyboard and a mouse with a help of Looking Glass.

Adding IVSHMEM Device to VM

Looking glass works by creating a shared memory buffer between a host and a guest. This is a lot faster than streaming frames via localhost, but requires additional setup.

With your VM turned off open the machine configuration

$ virsh edit [vmname]
...
<devices>
    ...
  <shmem name='looking-glass'>
    <model type='ivshmem-plain'/>
    <size unit='M'>32</size>
  </shmem>
</devices>
...

You should replace 32 with your own calculated value based on what resolution you are going to pass through. It can be calculated like this:

width x height x 4 x 2 = total bytes
total bytes / 1024 / 1024 = total mebibytes + 2

For example, in case of 1920x1080

1920 x 1080 x 4 x 2 = 16,588,800 bytes
16,588,800 / 1024 / 1024 = 15.82 MiB + 2 = 17.82

The result must be rounded up to the nearest power of two, and since 17.82 is bigger than 16 we should choose 32.

Next create a script to create a shared memory file.

/usr/local/bin/looking-glass-init.sh
#!/bin/sh

touch /dev/shm/looking-glass
chown user:kvm /dev/shm/looking-glass
chmod 660 /dev/shm/looking-glass

Replace user with your username.

Remember to make the script executable.

Create a systemd unit to execute this script during boot

/etc/systemd/system/looking-glass-init.service
[Unit]
Description=Create shared memory for looking glass

[Service]
Type=oneshot
ExecStart=/usr/local/bin/looking-glass-init.sh

[Install]
WantedBy=multi-user.target

Start and enable looking-glass-init.service

Installing the IVSHMEM Host to Windows guest

Currently Windows would not notify users about a new IVSHMEM device, it would silently install a dummy driver. To actually enable the device you have to go into device manager and update the driver for the device under the "System Devices" node for "PCI standard RAM Controller". Download a signed driver from issue 217, as it is not yet available elsewhere.

Once the driver is installed you must download the latest looking-glass-host from the looking glass website and start it on your guest. In order to run it you would also need to install Microsoft Visual C++ Redistributable from Microsoft It is also possible to make it start automatically on VM boot by editing the HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows\CurrentVersion\Run registry and adding a path to the downloaded executable.

Getting a client

Looking glass client can be installed from AUR using looking-glassAUR or looking-glass-gitAUR packages.

You can start it once the VM is set up and running

$ looking-glass-client

If you do not want to use Spice to control the guest mouse and keyboard you can disable the Spice server.

$ looking-glass-client -s

Additionally you may want to start Looking Glass Client as full screen, otherwise the image may be scaled down resulting in poor image fidelity.

$ looking-glass-client -F

Launch with the --help option for further information.

Swap peripherals to and from the Host

Looking Glass includes a Spice client in order to control mouse movement on the Windows guest. However this may have too much latency for certain applications, such as gaming. An alternative method is passing through specific USB devices for minimal latency. This allows for switching the devices between host and guest.

First create a .xml file for the device(s) you wish to pass-through, which libvirt will use to identify the device.

/home/$USER/.VFIOinput/input_1.xml
<hostdev mode='subsystem' type='usb' managed='no'>
<source>
<vendor id='0x[Before Colon]'/>
<product id='0x[After Colon]'/>
</source>
</hostdev>

Replace [Before/After Colon] with the contents of the 'lsusb' command, specific to the device you want to pass-through.

For instance my mouse is Bus 005 Device 002: ID 1532:0037 Razer USA, Ltd so I would replace vendor id with 1532, and product id with 1037.

Repeat this process for any additional USB devices you want to pass-through. If your mouse / keyboard has multiple entries in lsusb, perhaps if it is wireless, then create additional xml files for each.

Note: Do not forget to change the path & name of the script(s) above and below to match your user and specific system.

Next a bash script file is needed to tell libvirt what to attach/detach the USB devices to the guest.

/home/$USER/.VFIOinput/input_attach.sh
#!/bin/bash

virsh attach-device [VM-Name] [USBdevice]

Replace [VM-Name] with the name of your virtual machine, which can be seen under virt-manager. Additionally replace [USBdevice] with the full path to the .xml file for the device you wish to pass-through. Add additional lines for more than 1 device. For example here is my script:

/home/$USER/.VFIOinput/input_attach.sh
#!/bin/bash

virsh attach-device win10 /home/$USER/.VFIOinput/input_mouse.xml
virsh attach-device win10 /home/$USER/.VFIOinput/input_keyboard.xml

Next duplicate the script file and replace attach-device with detach-device. Ensure both scripts are executable with chmod +x $script.sh

This 2 script files can now be executed to attach or detach your USB devices from the host to the guest VM. It is important to note that they may need to be executed as root. To run the script from the Windows VM, one possibility is using PuTTY to SSH into the host, and execute the script. On Windows PuTTY comes with plink.exe which can execute singular commands over SSH before then logging out, instead of opening a SSH terminal, all in the background.

detach_devices.bat
"C:\Program Files\PuTTY\plink.exe" root@$HOST_IP -pw $ROOTPASSWORD /home/$USER/.VFIOinput/input_detach.sh

Replace $HOST_IP with the Host IP Address and $ROOTPASSWORD with the root password.

Warning: This method is insecure if somebody has access to your VM, since they could open the file and read your password. It is advisable to use SSH keys instead!

You may also want to execute the script files using key binds. On Windows one option is Autohotkey, and on the Host Xbindkeys. Because of the need to run the scripts as root, you may also need to use Polkit which can be used to authenticate specific executables as able to run as root without needing a password.

Bypassing the IOMMU groups (ACS override patch)

If you find your PCI devices grouped among others that you do not wish to pass through, you may be able to seperate them using Alex Williamson's ACS override patch. Make sure you understand the potential risk of doing so.

You will need a kernel with the patch applied. The easiest method to acquiring this is through the linux-vfioAUR package.

In addition, the ACS override patch needs to be enabled with kernel command line options. The patch file adds the following documentation:

pcie_acs_override =
        [PCIE] Override missing PCIe ACS support for:
    downstream
        All downstream ports - full ACS capabilties
    multifunction
        All multifunction devices - multifunction ACS subset
    id:nnnn:nnnn
        Specfic device - full ACS capabilities
        Specified as vid:did (vendor/device ID) in hex

The option pcie_acs_override=downstream is typically sufficient.

After installation and configuration, reconfigure your bootloader kernel parameters to load the new kernel with the pcie_acs_override= option enabled.

Plain QEMU without libvirt

Instead of setting up a virtual machine with the help of libvirt, plain QEMU commands with custom parameters can be used for running the VM intended to be used with PCI passthrough. This is desirable for some use cases like scripted setups, where the flexibility for usage with other scripts is needed.

To achieve this after #Setting up IOMMU and #Isolating the GPU, follow the QEMU article to setup the virtualized environment, enable KVM on it and use the flag -device vfio-pci,host=07:00.0 replacing the identifier (07:00.0) with your actual device's ID that you used for the GPU isolation earlier.

For utilizing the OVMF firmware, make sure the ovmf package is installed, copy the UEFI variables from /usr/share/ovmf/x64/OVMF_VARS.fd to temporary location like /tmp/MY_VARS.fd and finally specify the OVMF paths by appending the following parameters to the QEMU command (order matters):

  • -drive if=pflash,format=raw,readonly,file=/usr/share/ovmf/x64/OVMF_CODE.fd for the actual OVMF firmware binary, note the readonly option
  • -drive if=pflash,format=raw,file=/tmp/MY_VARS.fd for the variables
Note: QEMU's default SeaBIOS can be used instead of OVMF, but it is not recommended as it can cause issues with passthrough setups.

It is recommended to study the QEMU article for ways to enhance the performance by using the virtio drivers and other further configurations for the setup.

You also might have to use the -cpu host,kvm=off parameter to forward the host's CPU model info to the VM and fool the virtualization detection used by Nvidia's and possibly other manufacturers' device drivers trying to block the full hardware usage inside a virtualized system.

Passing though other devices

USB controller

If your motherboard has multiple USB controllers mapped to multiple groups, it is possible to pass those instead of USB devices. Passing an actual controller over an individual USB device provides the following advantages :

  • If a device disconnects or changes ID over the course of an given operation (such as a phone undergoing an update), the VM will not suddenly stop seeing it.
  • Any USB port managed by this controller is directly handled by the VM and can have its devices unplugged, replugged and changed without having to notify the hypervisor.
  • Libvirt will not complain if one of the USB devices you usually pass to the guest is missing when starting the VM.

Unlike with GPUs, drivers for most USB controllers do not require any specific configuration to work on a VM and control can normally be passed back and forth between the host and guest systems with no side effects.

Warning: Make sure your USB controller supports resetting: #Passing through a device that does not support resetting

You can find out which PCI devices correspond to which controller and how various ports and devices are assigned to each one of them using this command :

$ for usb_ctrl in $(find /sys/bus/usb/devices/usb* -maxdepth 0 -type l); do pci_path="$(dirname "$(realpath "${usb_ctrl}")")"; echo "Bus $(cat "${usb_ctrl}/busnum") --> $(basename $pci_path) (IOMMU group $(basename $(realpath $pci_path/iommu_group)))"; lsusb -s "$(cat "${usb_ctrl}/busnum"):"; echo; done
Bus 1 --> 0000:00:1a.0 (IOMMU group 4)
Bus 001 Device 004: ID 04f2:b217 Chicony Electronics Co., Ltd Lenovo Integrated Camera (0.3MP)
Bus 001 Device 007: ID 0a5c:21e6 Broadcom Corp. BCM20702 Bluetooth 4.0 [ThinkPad]
Bus 001 Device 008: ID 0781:5530 SanDisk Corp. Cruzer
Bus 001 Device 002: ID 8087:0024 Intel Corp. Integrated Rate Matching Hub
Bus 001 Device 001: ID 1d6b:0002 Linux Foundation 2.0 root hub

Bus 2 --> 0000:00:1d.0 (IOMMU group 9)
Bus 002 Device 006: ID 0451:e012 Texas Instruments, Inc. TI-Nspire Calculator
Bus 002 Device 002: ID 8087:0024 Intel Corp. Integrated Rate Matching Hub
Bus 002 Device 001: ID 1d6b:0002 Linux Foundation 2.0 root hub

This laptop has 3 USB ports managed by 2 USB controllers, each with their own IOMMU group. In this example, Bus 001 manages a single USB port (with a SanDisk USB pendrive plugged into it so it appears on the list), but also a number of internal devices, such as the internal webcam and the bluetooth card. Bus 002, on the other hand, does not apprear to manage anything except for the calculator that is plugged into it. The third port is empty, which is why it does not show up on the list, but is actually managed by Bus 002.

Once you have identified which controller manages which ports by plugging various devices into them and decided which one you want to passthrough, simply add it to the list of PCI host devices controlled by the VM in your guest configuration. No other configuration should be needed.

Note: If your USB controller does not support resetting, is not in an isolated group, or is otherwise unable to be passed through then it may still be possible to accomplish similar results through udev rules. See [1] which allows any device connected to specified USB ports to be automatically attached to a virtual machine.

Passing VM audio to host via PulseAudio

It is possible to route the virtual machine's audio to the host as an application using libvirt. This has the advantage of multiple audio streams being routable to one host output, and working with audio output devices that do not support passthrough. PulseAudio is required for this to work.

First, remove the comment from the #user = "" line. Then add your username in the quotations. This tells QEMU which user's pulseaudio stream to route through.

/etc/libvirt/qemu.conf
user = "example"

Next, modify the libvirt configuration

$ virsh edit [vmname]
<domain type='kvm'>

to

$ virsh edit [vmname]
<domain type='kvm' xmlns:qemu='http://libvirt.org/schemas/domain/qemu/1.0'>

Then set the QEMU PulseAudio environment variables at the bottom of the libvirt xml file

$ virsh edit [vmname]
    </devices>
   </domain>

to

$ virsh edit [vmname]
    </devices>
      <qemu:commandline>
        <qemu:env name='QEMU_AUDIO_DRV' value='pa'/>
        <qemu:env name='QEMU_PA_SERVER' value='/run/user/1000/pulse/native'/>
      </qemu:commandline>
 </domain>

Change 1000 under the user directory to your user uid (which can be found by running the id command. Remember to save the file and exit it without ending the process before continuing, otherwise the changes will not register. If you get the message Domain [vmname] XML configuration edited. after exiting, it means that your changes have been applied.

Restart libvirtd.service and user service pulseaudio.service.

Virtual Machine audio will now be routed through the host as an application. The application pavucontrol can be used to control the output device. Be aware that on Windows guests, this can cause audio crackling without using Message-Signaled Interrupts.

QEMU 3.0 audio changes

As of QEMU 3.0 part of the audio patches have been merged (reddit link). The qemu-patchedAUR package currently includes some additional audio patches, as some of the patches have not been officially up-streamed yet.

You will need to change the chipset accordingly to how your VM is set up, i.e. pc-q35-3.0 or pc-i440fx-3.0 (after installing qemu 3.0) to use the new code paths:

$ virsh edit [vmname]
<domain type='kvm'>
  ...
  <os>
    <type arch='x86_64' machine='pc-q35-3.0'>hvm</type>
    ...
  </os>
$ virsh edit [vmname]
<domain type='kvm'>
  ...
  <os>
    <type arch='x86_64' machine='pc-i440fx-3.0'>hvm</type>
    ...
  </os>
Note:
  • To speed up compilation time with qemu-patchedAUR use --target-list=x86_64-softmmu to compile qemu with only x86_64 guest support.
  • Since Qemu 3.0 the XML arguments qemu:env above are not needed if you run PulseAudio as your user and you have nographics_allow_host_audio = 1 enabled in /etc/libvirt/qemu.conf. If you use a different user with QEMU/Libvirt, you will need to keep the QEMU_PA_SERVER variable otherwise permission errors will occur.

Physical disk/partition

A whole disk or a partition may be used as a whole for improved I/O performance by adding an entry to the XML

Virtio-BLK Example:

$ virsh edit [vmname]
 
<devices>
...
  <disk type='block' device='disk'>
    <driver name='qemu' type='raw' cache='none' io='native'/>
    <source dev='/dev/disk/by-id/xxxxxxxx'/>
    <target dev='vda' bus='virtio'/>
  </disk>
...
</devices>

Virtio-SCSI Example:

$ virsh edit [vmname]
 
<devices>
...
  <disk type='block' device='disk'>
    <driver name='qemu' type='raw' cache='none' io='native'/>
    <source dev='/dev/disk/by-id/xxxxxxxx'/>
    <target dev='sda' bus='scsi'/>
  </disk>
...
</devices>

To find out which disk/partition is associated with the one you would like to pass:

$ ls -l /dev/disk/by-id
 ata-ST1000LM002-9VQ14L_Z0501SZ9 -> ../../sdd
 ata-ST1000LM002-9VQ14L_Z0501SZ9-part1 -> ../../sdd1

You can also add the disk with Virt-Manager's Add Hardware menu and then type the disk you want in the Select or create custom storage box, e.g. /dev/disk/by-id/ata-ST1000LM002-9VQ14L_Z0501SZ9

Depending on which bus you use, the above step will require either the VIOSTOR(bus=virtio) or VIOSCSI(bus=scsi) driver on Windows guests, refer to #Setting up the guest OS for the driver ISO.

Gotchas

Passing through a device that does not support resetting

When the VM shuts down, all devices used by the guest are deinitialized by its OS in preparation for shutdown. In this state, those devices are no longer functional and must then be power-cycled before they can resume normal operation. Linux can handle this power-cycling on its own, but when a device has no known reset methods, it remains in this disabled state and becomes unavailable. Since Libvirt and Qemu both expect all host PCI devices to be ready to reattach to the host before completely stopping the VM, when encountering a device that will not reset, they will hang in a "Shutting down" state where they will not be able to be restarted until the host system has been rebooted. It is therefore reccomanded to only pass through PCI devices which the kernel is able to reset, as evidenced by the presence of a reset file in the PCI device sysfs node, such as /sys/bus/pci/devices/0000:00:1a.0/reset.

The following bash command shows which devices can and cannot be reset.

for iommu_group in $(find /sys/kernel/iommu_groups/ -maxdepth 1 -mindepth 1 -type d);do echo "IOMMU group $(basename "$iommu_group")"; for device in $(\ls -1 "$iommu_group"/devices/); do if [[ -e "$iommu_group"/devices/"$device"/reset ]]; then echo -n "[RESET]"; fi; echo -n $'\t';lspci -nns "$device"; done; done
IOMMU group 0
	00:00.0 Host bridge [0600]: Intel Corporation Xeon E3-1200 v2/Ivy Bridge DRAM Controller [8086:0158] (rev 09)
IOMMU group 1
	00:01.0 PCI bridge [0604]: Intel Corporation Xeon E3-1200 v2/3rd Gen Core processor PCI Express Root Port [8086:0151] (rev 09)
	01:00.0 VGA compatible controller [0300]: NVIDIA Corporation GK208 [GeForce GT 720] [10de:1288] (rev a1)
	01:00.1 Audio device [0403]: NVIDIA Corporation GK208 HDMI/DP Audio Controller [10de:0e0f] (rev a1)
IOMMU group 2
	00:14.0 USB controller [0c03]: Intel Corporation 7 Series/C210 Series Chipset Family USB xHCI Host Controller [8086:1e31] (rev 04)
IOMMU group 4
[RESET]	00:1a.0 USB controller [0c03]: Intel Corporation 7 Series/C210 Series Chipset Family USB Enhanced Host Controller #2 [8086:1e2d] (rev 04)
IOMMU group 5
[RESET]	00:1b.0 Audio device [0403]: Intel Corporation 7 Series/C210 Series Chipset Family High Definition Audio Controller [8086:1e20] (rev 04)
IOMMU group 10
[RESET]	00:1d.0 USB controller [0c03]: Intel Corporation 7 Series/C210 Series Chipset Family USB Enhanced Host Controller #1 [8086:1e26] (rev 04)
IOMMU group 13
	06:00.0 VGA compatible controller [0300]: NVIDIA Corporation GM204 [GeForce GTX 970] [10de:13c2] (rev a1)
	06:00.1 Audio device [0403]: NVIDIA Corporation GM204 High Definition Audio Controller [10de:0fbb] (rev a1)

This signals that the xHCI USB controller in 00:14.0 cannot be reset and will therefore stop the VM from shutting down properly, while the integrated sound card in 00:1b.0 and the other two controllers in 00:1a.0 and 00:1d.0 do not share this problem and can be passed without issue.

Complete setups and examples

For many reasons users may seek to see complete passthrough setup examples.

These examples offer a supplement to existing hardware compatibility lists. Additionally, if you have trouble configuring a certain mechanism in your setup, you might find these examples very valuable. Users there have described their setups in detail, and some have provided examples of their configuration files as well.

We encourage those who successfully build their system from this resource to help improve it by contributing their builds. Due to the many different hardware manufacturers involved, the seemingly significant lack of sufficient documentation, as well as other issues due to the nature of this process, community contributions are necessary.

Troubleshooting

If your issue is not mentioned below, you may want to browse QEMU#Troubleshooting.

"Error 43: Driver failed to load" on Nvidia GPUs passed to Windows VMs

Note:
  • This may also fix SYSTEM_THREAD_EXCEPTION_NOT_HANDLED boot crashes related to Nvidia drivers.
  • This may also fix problems under linux guests.

Since version 337.88, Nvidia drivers on Windows check if an hypervisor is running and fail if it detects one, which results in an Error 43 in the Windows device manager. Starting with QEMU 2.5.0 and libvirt 1.3.3, the vendor_id for the hypervisor can be spoofed, which is enough to fool the Nvidia drivers into loading anyway. All one must do is add hv_vendor_id=whatever to the cpu parameters in their QEMU command line, or by adding the following line to their libvirt domain configuration. It may help for the ID to be set to a 12-character alphanumeric (e.g. '1234567890ab') as opposed to longer or shorter strings.

$ virsh edit [vmname]
...
<features>
	<hyperv>
		...
		<vendor_id state='on' value='whatever'/>
		...
	</hyperv>
	...
	<kvm>
	<hidden state='on'/>
	</kvm>
</features>
...

Users with older versions of QEMU and/or libvirt will instead have to disable a few hypervisor extensions, which can degrade performance substantially. If this is what you want to do, do the following replacement in your libvirt domain config file.

$ virsh edit [vmname]
...
<features>
	<hyperv>
		<relaxed state='on'/>
		<vapic state='on'/>
		<spinlocks state='on' retries='8191'/>
	</hyperv>
	...
</features>
...
<clock offset='localtime'>
	<timer name='hypervclock' present='yes'/>
</clock>
...
...
<clock offset='localtime'>
	<timer name='hypervclock' present='no'/>
</clock>
...
<features>
	<kvm>
	<hidden state='on'/>
	</kvm>
	...
	<hyperv>
		<relaxed state='off'/>
		<vapic state='off'/>
		<spinlocks state='off'/>
	</hyperv>
	...
</features>
...

"BAR 3: cannot reserve [mem]" error in dmesg after starting VM

Tango-view-fullscreen.pngThis article or section needs expansion.Tango-view-fullscreen.png

Reason: This error is actually related to the boot_vgs issue and should be merged together with everything else concerning GPU ROMs. (Discuss in Talk:PCI passthrough via OVMF#UEFI (OVMF) Compatibility in VBIOS)

With respect to this article:

If you still have code 43 check dmesg for memory reservation errors after starting up VM, if you have similar it could be the case:

vfio-pci 0000:09:00.0: BAR 3: cannot reserve [mem 0xf0000000-0xf1ffffff 64bit pref]

Find out a PCI Bridge your graphic card is connected to. This will give actual hierarchy of devices:

$ lspci -t

Before starting VM run following lines replacing IDs with actual from previous output.

# echo 1 > /sys/bus/pci/devices/0000\:00\:03.1/remove
# echo 1 > /sys/bus/pci/rescan
Note: Probably setting kernel parameter video=efifb:off is required as well. Source

UEFI (OVMF) compatibility in VBIOS

Tango-edit-cut.pngThis section is being considered for removal.Tango-edit-cut.png

Reason: Flashing you guest GPU for the purpose of a GPU passthrough is never good advice. A full section should be dedicated to VBIOS compatibility. (Discuss in Talk:PCI passthrough via OVMF#UEFI (OVMF) Compatibility in VBIOS)

With respect to this article:

Error 43 can be caused by the GPU's VBIOS without UEFI support. To check whenever your VBIOS supports it, you will have to use rom-parser:

$ git clone https://github.com/awilliam/rom-parser
$ cd rom-parser && make

Dump the GPU VBIOS:

# echo 1 > /sys/bus/pci/devices/0000:01:00.0/rom
# cat /sys/bus/pci/devices/0000:01:00.0/rom > /tmp/image.rom
# echo 0 > /sys/bus/pci/devices/0000:01:00.0/rom

And test it for compatibility:

$ ./rom-parser /tmp/image.rom
Valid ROM signature found @600h, PCIR offset 190h
	PCIR: type 0 (x86 PC-AT), vendor: 10de, device: 1184, class: 030000
	PCIR: revision 0, vendor revision: 1
Valid ROM signature found @fa00h, PCIR offset 1ch
	PCIR: type 3 (EFI), vendor: 10de, device: 1184, class: 030000
	PCIR: revision 3, vendor revision: 0
		EFI: Signature Valid, Subsystem: Boot, Machine: X64
	Last image

To be UEFI compatible, you need a "type 3 (EFI)" in the result. If it is not there, try updating your GPU VBIOS. GPU manufacturers often share VBIOS upgrades on their support pages. A large database of known compatible and working VBIOSes (along with their UEFI compatibility status!) is available on TechPowerUp.

Updated VBIOS can be used in the VM without flashing. To load it in QEMU:

-device vfio-pci,host=07:00.0,......,romfile=/path/to/your/gpu/bios.bin \

And in libvirt:

<hostdev>
     ...
     <rom file='/path/to/your/gpu/bios.bin'/>
     ...
   </hostdev>

One should compare VBIOS versions between host and guest systems using nvflash (Linux versions under Show more versions) or GPU-Z (in Windows guest). To check the currently loaded VBIOS:

$ ./nvflash --version
...
Version               : 80.04.XX.00.97
...
UEFI Support          : No
UEFI Version          : N/A
UEFI Variant Id       : N/A ( Unknown )
UEFI Signer(s)        : Unsigned
...

And to check a given VBIOS file:

$ ./nvflash --version NV299MH.rom
...
Version               : 80.04.XX.00.95
...
UEFI Support          : Yes
UEFI Version          : 0x10022 (Jul  2 2013 @ 16377903 )
UEFI Variant Id       : 0x0000000000000004 ( GK1xx )
UEFI Signer(s)        : Microsoft Corporation UEFI CA 2011
...

If the external ROM did not work as it should in the guest, you will have to flash the newer VBIOS image to the GPU. In some cases it is possible to create your own VBIOS image with UEFI support using GOPUpd tool, however this is risky and may result in GPU brick.

Warning: Failure during flashing may "brick" your GPU - recovery may be possible, but rarely easy and often requires additional hardware. DO NOT flash VBIOS images for other GPU models (different boards may use different VBIOSes, clocks, fan configuration). If it breaks, you get to keep all the pieces.

In order to avoid the irreparable damage to your graphics adapter it is necessary to unload the NVIDIA kernel driver first:

# modprobe -r nvidia_modeset nvidia 

Flashing the VBIOS can be done with:

# ./nvflash romfile.bin
Warning: DO NOT interrupt the flashing process, even if it looks like it is stuck. Flashing should take about a minute on most GPUs, but may take longer.

Slowed down audio pumped through HDMI on the video card

For some users VM's audio slows down/starts stuttering/becomes demonic after a while when it is pumped through HDMI on the video card. This usually also slows down graphics. A possible solution consists of enabling MSI (Message Signaled-Based Interrupts) instead of the default (Line-Based Interrupts).

In order to check whether MSI is supported or enabled, run the following command as root:

# lspci -vs $device | grep 'MSI:'

where `$device` is the card's address (e.g. `01:00.0`).

The output should be similar to:

Capabilities: [60] MSI: Enable- Count=1/1 Maskable- 64bit+

A - after Enable means MSI is supported, but not used by the VM, while a + says that the VM is using it.

The procedure to enable it is quite complex, instructions and an overview of the setting can be found here.

Other hints can be found on the lime-technology's wiki, or on this article on VFIO tips and tricks.

A UI tool called MSI Utility (FOSS Version 2) works with Windows 10 64-bit and simplifies the process.

In order to fix the issues enabling MSI on the 0 function of a nVidia card (01:00.0 VGA compatible controller: NVIDIA Corporation GM206 [GeForce GTX 960] (rev a1) (prog-if 00 [VGA controller])) was not enough; it will also be required to enable it on the other function (01:00.1 Audio device: NVIDIA Corporation Device 0fba (rev a1)) to fix the issue.

No HDMI audio output on host when intel_iommu is enabled

If after enabling intel_iommu the HDMI output device of Intel GPU becomes unusable on the host then setting the option igfx_off (i.e. intel_iommu=on,igfx_off) might bring the audio back, please read Graphics Problems? in Intel-IOMMU.txt for details about setting igfx_off.

X does not start after enabling vfio_pci

This is related to the host GPU being detected as a secondary GPU, which causes X to fail/crash when it tries to load a driver for the guest GPU. To circumvent this, a Xorg configuration file specifying the BusID for the host GPU is required. The correct BusID can be acquired from lspci or the Xorg log. Source →

/etc/X11/xorg.conf.d/10-intel.conf
Section "Device"
        Identifier "Intel GPU"
        Driver "modesetting"
        BusID  "PCI:0:2:0"
EndSection

Chromium ignores integrated graphics for acceleration

Chromium and friends will try to detect as many GPUs as they can in the system and pick which one is preferred (usually discrete NVIDIA/AMD graphics). It tries to pick a GPU by looking at PCI devices, not OpenGL renderers available in the system - the result is that Chromium may ignore the integrated GPU available for rendering and try to use the dedicated GPU bound to the vfio-pci driver, and unusable on the host system, regardless of whenever a guest VM is running or not. This results in software rendering being used (leading to higher CPU load, which may also result in choppy video playback, scrolling and general un-smoothness).

This can be fixed by explicitly telling Chromium which GPU you want to use.

VM only uses one core

For some users, even if IOMMU is enabled and the core count is set to more than 1, the VM still only uses one CPU core and thread. To solve this enable "Manually set CPU topology" in virt-manager and set it to the desirable amount of CPU sockets, cores and threads. Keep in mind that "Threads" refers to the thread count per CPU, not the total count.

Passthrough seems to work but no output is displayed

Make sure if you are using virt-manager that UEFI firmware is selected for your virtual machine. Also, make sure you have passed the correct device to the VM.

virt-manager has permission issues

If you are getting a permission error with virt-manager add the following to your /etc/libvirt/qemu.conf:

group="kvm"
user="user"

If that does not work make sure your user is added to the kvm and libvirt user groups.

Host lockup after VM shutdown

This issue seems to primarily affect users running a Windows 10 guest and usually after the VM has been run for a prolonged period of time: the host will experience multiple CPU core lockups (see [2]). To fix this try enabling Message Signal Interrupts on the GPU passed through to the guest. A good guide for how to do this can be found in [3].

Host lockup if guest is left running during sleep

VFIO-enabled virtual machines tend to become unstable if left running through a sleep/wakeup cycle and have been known to cause the host machine to lockup when an attempt is then made to shut them down. In order to avoid this, one can simply prevent the host from going into sleep while the guest is running using the following libvirt hook script and systemd unit. The hook file needs executable permissions to work.

/etc/libvirt/hooks/qemu
#!/bin/bash

OBJECT="$1"
OPERATION="$2"
SUBOPERATION="$3"
EXTRA_ARG="$4"

case "$OPERATION" in
        "prepare")
                systemctl start libvirt-nosleep@"$OBJECT"
                ;;
        "release")
                systemctl stop libvirt-nosleep@"$OBJECT"
                ;;
esac
/etc/systemd/system/libvirt-nosleep@.service
[Unit]
Description=Preventing sleep while libvirt domain "%i" is running

[Service]
Type=simple
ExecStart=/usr/bin/systemd-inhibit --what=sleep --why="Libvirt domain \"%i\" is running" --who=%U --mode=block sleep infinity

Cannot boot after upgrading ovmf

If you cannot boot after upgrading from ovmf-1:r23112.018432f0ce-1 then you need to remove the old *VARS.fd file in /var/lib/libvirt/qemu/nvram:

mv /var/lib/libvirt/qemu/nvram/vmname_VARS.fd /var/lib/libvirt/qemu/nvram/vmname_VARS.fd.old

See FS#57825 for further details.

See also