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.

Note

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

Limitations

• 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.

Note

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.

Changed in version 26.0.0: (Zed): PCI passthrough device inventories now can be tracked in Placement. For more information, refer to PCI tracking in Placement.

Changed in version 26.0.0: (Zed): The nova-compute service will refuse to start if both the parent PF and its children VFs are configured in pci.device_spec. For more information, refer to PCI tracking in Placement.

Changed in version 26.0.0: (Zed): The nova-compute service will refuse to start with pci.device_spec configuration that uses the devname field.

Changed in version 27.0.0: (2023.1 Antelope): Nova provides Placement based scheduling support for servers with flavor based PCI requests. This support is disable by default.

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

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.device_spec option. For example, assuming our sample PCI device has a PCI address of 41:00.0 on each host:

[pci]
device_spec = { "address": "0000:41:00.0" }

Refer to pci.device_spec for syntax information.

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

[pci]
device_spec = { "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:

[pci]
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:

type-PF

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

type-VF

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

type-PCI

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.

Important

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

Once configured, restart the nova-compute service.

Special Tags

When specified in pci.device_spec some tags have special meaning:

physical_network

Associates a device with a physical network label which corresponds to the physical_network attribute of a network segment object in Neutron. For virtual networks such as overlays a value of null should be specified as follows: "physical_network": null. In the case of physical networks, this tag is used to supply the metadata necessary for identifying a switched fabric to which a PCI device belongs and associate the port with the correct network segment in the networking backend. Besides typical SR-IOV scenarios, this tag can be used for remote-managed devices in conjunction with the remote_managed tag.

Note

The use of "physical_network": null is only supported in single segment networks. This is due to Nova not supporting multisegment networks for SR-IOV ports. See bug 1983570 for details.

remote_managed

Used to specify whether a PCI device is managed remotely or not. By default, devices are implicitly tagged as "remote_managed": "false" but and they must be tagged as "remote_managed": "true" if ports with VNIC_TYPE_REMOTE_MANAGED are intended to be used. Once that is done, those PCI devices will not be available for allocation for regular PCI passthrough use. Specifying "remote_managed": "true" is only valid for SR-IOV VFs and specifying it for PFs is prohibited.

Important

It is recommended that PCI VFs that are meant to be remote-managed (e.g. the ones provided by SmartNIC DPUs) are tagged as remote-managed in order to prevent them from being allocated for regular PCI passthrough since they have to be programmed accordingly at the host that has access to the NIC switch control plane. If this is not done, instances requesting regular SR-IOV ports may get a device that will not be configured correctly and will not be usable for sending network traffic.

Important

For the Libvirt virt driver, clearing a VLAN by programming VLAN 0 must not result in errors in the VF kernel driver at the compute host. Before v8.1.0 Libvirt clears a VLAN before passing a VF through to the guest which may result in an error depending on your driver and kernel version (see, for example, this bug which discusses a case relevant to one driver). As of Libvirt v8.1.0, EPERM errors encountered while programming a VLAN are ignored if VLAN clearing is not explicitly requested in the device XML.

trusted

If a port is requested to be trusted by specifying an extra option during port creation via --binding-profile trusted=true, only devices tagged as trusted: "true" will be allocated to instances. Nova will then configure those devices as trusted by the network controller through its PF device driver. The specific set of features allowed by the trusted mode of a VF will differ depending on the network controller itself, its firmware version and what a PF device driver version allows to pass to the NIC. Common features to be affected by this tag are changing the VF MAC address, enabling promiscuous mode or multicast promiscuous mode.

Important

While the trusted tag does not directly conflict with the remote_managed tag, network controllers in SmartNIC DPUs may prohibit setting the trusted mode on a VF via a PF device driver in the first place. It is recommended to test specific devices, drivers and firmware versions before assuming this feature can be used.

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:

[filter_scheduler]
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:

[pci]
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:

required

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.

socket

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.

preferred

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.

legacy

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.

PCI tracking in Placement

Note

The feature described below are optional and disabled by default in nova 26.0.0. (Zed). The legacy PCI tracker code path is still supported and enabled. The Placement PCI tracking can be enabled via the pci.report_in_placement configuration. But please note that once it is enabled on a given compute host it cannot be disabled there any more.

Since nova 26.0.0 (Zed) PCI passthrough device inventories are tracked in Placement. If a PCI device exists on the hypervisor and matches one of the device specifications configured via pci.device_spec then Placement will have a representation of the device. Each PCI device of type type-PCI and type-PF will be modeled as a Placement resource provider (RP) with the name <hypervisor_hostname>_<pci_address>. A devices with type type-VF is represented by its parent PCI device, the PF, as resource provider.

By default nova will use CUSTOM_PCI_<vendor_id>_<product_id> as the resource class in PCI inventories in Placement. However the name of the resource class can be customized via the resource_class tag in the pci.device_spec option. There is also a new traits tag in that configuration that allows specifying a list of placement traits to be added to the resource provider representing the matching PCI devices.

Note

In nova 26.0.0 (Zed) the Placement resource tracking of PCI devices does not support SR-IOV devices intended to be consumed via Neutron ports and therefore having physical_network tag in pci.device_spec. Such devices are supported via the legacy PCI tracker code path in Nova.

Note

Having different resource class or traits configuration for VFs under the same parent PF is not supported and the nova-compute service will refuse to start with such configuration.

Important

While nova supported configuring both the PF and its children VFs for PCI passthrough in the past, it only allowed consuming either the parent PF or its children VFs. Since 26.0.0. (Zed) the nova-compute service will enforce the same rule for the configuration as well and will refuse to start if both the parent PF and its VFs are configured.

Important

While nova supported configuring PCI devices by device name via the devname parameter in pci.device_spec in the past, this proved to be problematic as the netdev name of a PCI device could change for multiple reasons during hypervisor reboot. So since nova 26.0.0 (Zed) the nova-compute service will refuse to start with such configuration. It is suggested to use the PCI address of the device instead.

The nova-compute service makes sure that existing instances with PCI allocations in the nova DB will have a corresponding PCI allocation in placement. This allocation healing also acts on any new instances regardless of the status of the scheduling part of this feature to make sure that the nova DB and placement are in sync. There is one limitation of the healing logic. It assumes that there is no in-progress migration when the nova-compute service is upgraded. If there is an in-progress migration then the PCI allocation on the source host of the migration will not be healed. The placement view will be consistent after such migration is completed or reverted.

Reconfiguring the PCI devices on the hypervisor or changing the pci.device_spec configuration option and restarting the nova-compute service is supported in the following cases:

• new devices are added

• devices without allocation are removed

Removing a device that has allocations is not supported. If a device having any allocation is removed then the nova-compute service will keep the device and the allocation exists in the nova DB and in placement and logs a warning. If a device with any allocation is reconfigured in a way that an allocated PF is removed and VFs from the same PF is configured (or vice versa) then nova-compute will refuse to start as it would create a situation where both the PF and its VFs are made available for consumption.

Since nova 27.0.0 (2023.1 Antelope) scheduling and allocation of PCI devices in Placement can also be enabled via filter_scheduler.pci_in_placement. Please note that this should only be enabled after all the computes in the system is configured to report PCI inventory in Placement via enabling pci.report_in_placement. In Antelope flavor based PCI requests are support but Neutron port base PCI requests are not handled in Placement.

If you are upgrading from an earlier version with already existing servers with PCI usage then you must enable pci.report_in_placement first on all your computes having PCI allocations and then restart the nova-compute service, before you enable filter_scheduler.pci_in_placement. The compute service will heal the missing PCI allocation in placement during startup and will continue healing missing allocations for future servers until the scheduling support is enabled.

If a flavor requests multiple type-VF devices via pci_passthrough:alias then it is important to consider the value of group_policy as well. The value none allows nova to select VFs from the same parent PF to fulfill the request. The value isolate restricts nova to select each VF from a different parent PF to fulfill the request. If group_policy is not provided in such flavor then it will defaulted to none.

Symmetrically with the resource_class and traits fields of pci.device_spec the pci.alias configuration option supports requesting devices by Placement resource class name via the resource_class field and also support requesting traits to be present on the selected devices via the traits field in the alias. If the resource_class field is not specified in the alias then it is defaulted by nova to CUSTOM_PCI_<vendor_id>_<product_id>.

For deeper technical details please read the nova specification.

Virtual IOMMU support

With provided hw:viommu_model flavor extra spec or equivalent image metadata property hw_viommu_model and with the guest CPU architecture and OS allows, we can enable vIOMMU in libvirt driver.

Note

Enable vIOMMU might introduce significant performance overhead. You can see performance comparison table from AMD vIOMMU session on KVM Forum 2021. For the above reason, vIOMMU should only be enabled for workflow that require it.

Here are four possible values allowed for hw:viommu_model (and hw_viommu_model):

virtio

Supported on Libvirt since 8.3.0, for Q35 and ARM virt guests.

smmuv3

Supported on Libvirt since 5.5.0, for ARM virt guests.

intel

Supported for for Q35 guests.

auto

This option will translate to virtio if Libvirt supported, else intel on X86 (Q35) and smmuv3 on AArch64.

For the viommu attributes:

• intremap, caching_mode, and iotlb options for viommu (These attributes are driver attributes defined in Libvirt IOMMU Domain) will directly enabled.

• eim will directly enabled if machine type is Q35. eim is driver attribute defined in Libvirt IOMMU Domain.

Note

eim(Extended Interrupt Mode) attribute (with possible values on and off) can be used to configure Extended Interrupt Mode. A q35 domain with split I/O APIC (as described in hypervisor features), and both interrupt remapping and EIM turned on for the IOMMU, will be able to use more than 255 vCPUs. Since 3.4.0 (QEMU/KVM only).

• aw_bits attribute can used to set the address width to allow mapping larger iova addresses in the guest. Since Qemu current supported values are 39 and 48, we directly set this to larger width (48) if Libvirt supported. aw_bits is driver attribute defined in Libvirt IOMMU Domain.