Attaching physical PCI devices to guests

The PCI passthrough feature in OpenStack allows full access and direct control of a physical PCI device in guests. This mechanism is generic for any kind of PCI device, and runs with a Network Interface Card (NIC), Graphics Processing Unit (GPU), or any other devices that can be attached to a PCI bus. Correct driver installation is the only requirement for the guest to properly use the devices.

Some PCI devices provide Single Root I/O Virtualization and Sharing (SR-IOV) capabilities. When SR-IOV is used, a physical device is virtualized and appears as multiple PCI devices. Virtual PCI devices are assigned to the same or different guests. In the case of PCI passthrough, the full physical device is assigned to only one guest and cannot be shared.

PCI devices are requested through flavor extra specs, specifically via the pci_passthrough:alias flavor extra spec. This guide demonstrates how to enable PCI passthrough for a type of PCI device with a vendor ID of 8086 and a product ID of 154d - an Intel X520 Network Adapter - by mapping them to the alias a1. You should adjust the instructions for other devices with potentially different capabilities.


For information on creating servers with SR-IOV network interfaces, refer to the Networking Guide.


  • Attaching SR-IOV ports to existing servers was not supported until the 22.0.0 Victoria release. Due to various bugs in libvirt and qemu we recommend to use at least libvirt version 6.0.0 and at least qemu version 4.2.

  • Cold migration (resize) of servers with SR-IOV devices attached was not supported until the 14.0.0 Newton release, see bug 1512800 for details.


Nova only supports PCI addresses where the fields are restricted to the following maximum value:

  • domain - 0xFFFF

  • bus - 0xFF

  • slot - 0x1F

  • function - 0x7

Nova will ignore PCI devices reported by the hypervisor if the address is outside of these ranges.

Enabling PCI passthrough

Configure compute host

To enable PCI passthrough on an x86, Linux-based compute node, the following are required:

  • VT-d enabled in the BIOS

  • IOMMU enabled on the host OS, e.g. by adding the intel_iommu=on or amd_iommu=on parameter to the kernel parameters

  • Assignable PCIe devices

To enable PCI passthrough on a Hyper-V compute node, the following are required:

  • Windows 10 or Windows / Hyper-V Server 2016 or newer

  • VT-d enabled on the host

  • Assignable PCI devices

In order to check the requirements above and if there are any assignable PCI devices, run the following Powershell commands:


If the compute node passes all the requirements, the desired assignable PCI devices to be disabled and unmounted from the host, in order to be assignable by Hyper-V. The following can be read for more details: Hyper-V PCI passthrough.

Configure nova-compute

Once PCI passthrough has been configured for the host, nova-compute must be configured to allow the PCI device to pass through to VMs. This is done using the pci.passthrough_whitelist option. For example, assuming our sample PCI device has a PCI address of 41:00.0 on each host:

passthrough_whitelist = { "address": "0000:41:00.0" }

Refer to pci.passthrough_whitelist for syntax information.

Alternatively, to enable passthrough of all devices with the same product and vendor ID:

passthrough_whitelist = { "vendor_id": "8086", "product_id": "154d" }

If using vendor and product IDs, all PCI devices matching the vendor_id and product_id are added to the pool of PCI devices available for passthrough to VMs.

In addition, it is necessary to configure the pci.alias option, which is a JSON-style configuration option that allows you to map a given device type, identified by the standard PCI vendor_id and (optional) product_id fields, to an arbitrary name or alias. This alias can then be used to request a PCI device using the pci_passthrough:alias flavor extra spec, as discussed previously. For our sample device with a vendor ID of 0x8086 and a product ID of 0x154d, this would be:

alias = { "vendor_id":"8086", "product_id":"154d", "device_type":"type-PF", "name":"a1" }

It’s important to note the addition of the device_type field. This is necessary because this PCI device supports SR-IOV. The nova-compute service categorizes devices into one of three types, depending on the capabilities the devices report:


The device supports SR-IOV and is the parent or root device.


The device is a child device of a device that supports SR-IOV.


The device does not support SR-IOV.

By default, it is only possible to attach type-PCI devices using PCI passthrough. If you wish to attach type-PF or type-VF devices, you must specify the device_type field in the config option. If the device was a device that did not support SR-IOV, the device_type field could be omitted.

Refer to pci.alias for syntax information.


This option must also be configured on controller nodes. This is discussed later in this document.

Once configured, restart the nova-compute service.

Configure nova-scheduler

The nova-scheduler service must be configured to enable the PciPassthroughFilter. To do this, add this filter to the list of filters specified in filter_scheduler.enabled_filters and set filter_scheduler.available_filters to the default of nova.scheduler.filters.all_filters. For example:

enabled_filters = ...,PciPassthroughFilter
available_filters = nova.scheduler.filters.all_filters

Once done, restart the nova-scheduler service.

Configure nova-api

It is necessary to also configure the pci.alias config option on the controller. This configuration should match the configuration found on the compute nodes. For example:

alias = { "vendor_id":"8086", "product_id":"154d", "device_type":"type-PF", "name":"a1", "numa_policy":"preferred" }

Refer to pci.alias for syntax information. Refer to Affinity for numa_policy information.

Once configured, restart the nova-api service.

Configuring a flavor or image

Once the alias has been configured, it can be used for an flavor extra spec. For example, to request two of the PCI devices referenced by alias a1, run:

$ openstack flavor set m1.large --property "pci_passthrough:alias"="a1:2"

For more information about the syntax for pci_passthrough:alias, refer to the documentation.

PCI-NUMA affinity policies

By default, the libvirt driver enforces strict NUMA affinity for PCI devices, be they PCI passthrough devices or neutron SR-IOV interfaces. This means that by default a PCI device must be allocated from the same host NUMA node as at least one of the instance’s CPUs. This isn’t always necessary, however, and you can configure this policy using the hw:pci_numa_affinity_policy flavor extra spec or equivalent image metadata property. There are three possible values allowed:


This policy means that nova will boot instances with PCI devices only if at least one of the NUMA nodes of the instance is associated with these PCI devices. It means that if NUMA node info for some PCI devices could not be determined, those PCI devices wouldn’t be consumable by the instance. This provides maximum performance.


This policy means that the PCI device must be affined to the same host socket as at least one of the guest NUMA nodes. For example, consider a system with two sockets, each with two NUMA nodes, numbered node 0 and node 1 on socket 0, and node 2 and node 3 on socket 1. There is a PCI device affined to node 0. An PCI instance with two guest NUMA nodes and the socket policy can be affined to either:

  • node 0 and node 1

  • node 0 and node 2

  • node 0 and node 3

  • node 1 and node 2

  • node 1 and node 3

The instance cannot be affined to node 2 and node 3, as neither of those are on the same socket as the PCI device. If the other nodes are consumed by other instances and only nodes 2 and 3 are available, the instance will not boot.


This policy means that nova-scheduler will choose a compute host with minimal consideration for the NUMA affinity of PCI devices. nova-compute will attempt a best effort selection of PCI devices based on NUMA affinity, however, if this is not possible then nova-compute will fall back to scheduling on a NUMA node that is not associated with the PCI device.


This is the default policy and it describes the current nova behavior. Usually we have information about association of PCI devices with NUMA nodes. However, some PCI devices do not provide such information. The legacy value will mean that nova will boot instances with PCI device if either:

  • The PCI device is associated with at least one NUMA nodes on which the instance will be booted

  • There is no information about PCI-NUMA affinity available

For example, to configure a flavor to use the preferred PCI NUMA affinity policy for any neutron SR-IOV interfaces attached by the user:

$ openstack flavor set $FLAVOR \
    --property hw:pci_numa_affinity_policy=preferred

You can also configure this for PCI passthrough devices by specifying the policy in the alias configuration via pci.alias. For more information, refer to the documentation.