The purpose of this page is to describe how to enable SR-IOV
functionality available in OpenStack (using OpenStack Networking). This
functionality was first introduced in the OpenStack Juno release. This
page intends to serve as a guide for how to configure OpenStack
Networking and OpenStack Compute to create SR-IOV ports.

The basics

PCI-SIG Single Root I/O Virtualization and Sharing (SR-IOV)
functionality is available in OpenStack since the Juno release. The
SR-IOV specification defines a standardized mechanism to virtualize PCIe
devices. This mechanism can virtualize a single PCIe Ethernet controller
to appear as multiple PCIe devices. Each device can be directly assigned
to an instance, bypassing the hypervisor and virtual switch layer. As a
result, users are able to achieve low latency and near-line wire
speed.

The following terms are used throughout this document:

SR-IOV agent

The SR-IOV agent allows you to set the admin state of ports,
configure port security (enable and disable spoof checking), and
configure QoS rate limiting and minimum bandwidth. You must include the
SR-IOV agent on each compute node using SR-IOV ports.

Supported Ethernet
controllers

The following manufacturers are known to work:

• Intel
• Mellanox
• QLogic
• Broadcom

For information on Mellanox SR-IOV Ethernet ConnectX
cards, see the Mellanox:
How To Configure SR-IOV VFs on ConnectX-4 or newer.

For information on QLogic SR-IOV Ethernet cards, see
the User’s
Guide OpenStack Deployment with SR-IOV Configuration.

For information on Broadcom NetXtreme Series Ethernet
cards, see the Broadcom
NetXtreme Product Page.

Using SR-IOV interfaces

In order to enable SR-IOV, the following steps are required:

1. Create Virtual Functions (Compute)
2. Configure allow list for PCI devices in nova-compute (Compute)
3. Configure neutron-server (Controller)
4. Configure nova-scheduler (Controller)
5. Enable neutron sriov-agent (Compute)

We recommend using VLAN provider networks for segregation. This way
you can combine instances without SR-IOV ports and instances with SR-IOV
ports on a single network.

Create Virtual Functions
(Compute)

Create the VFs for the network interface that will be used for
SR-IOV. We use eth3 as PF, which is also used as the
interface for the VLAN provider network and has access to the private
networks of all machines.

1. Ensure SR-IOV and VT-d are enabled in BIOS.

2. Enable IOMMU in Linux by adding intel_iommu=on to
   the kernel parameters, for example, using GRUB.

3. On each compute node, create the VFs via the PCI SYS
   interface:

   # echo '8' > /sys/class/net/eth3/device/sriov_numvfs

   Note

   On some PCI devices, observe that when changing the amount of VFs you
   receive the error Device or resource busy. In this case,
   you must first set sriov_numvfs to 0, then set
   it to your new value.

   Note

   A network interface could be used both for PCI passthrough, using the
   PF, and SR-IOV, using the VFs. If the PF is used, the VF number stored
   in the sriov_numvfs file is lost. If the PF is attached
   again to the operating system, the number of VFs assigned to this
   interface will be zero. To keep the number of VFs always assigned to
   this interface, modify the interfaces configuration file adding an
   ifup script command.

   On Ubuntu, modify the /etc/network/interfaces file:

   auto eth3
   iface eth3 inet dhcp
   pre-up echo '4' > /sys/class/net/eth3/device/sriov_numvfs

   On RHEL and derivatives, modify the /sbin/ifup-local
   file:

   #!/bin/sh
   if [[ "$1" == "eth3" ]]
   then
       echo '4' > /sys/class/net/eth3/device/sriov_numvfs
   fi

   Warning

   Alternatively, you can create VFs by passing the max_vfs
   to the kernel module of your network interface. However, the
   max_vfs parameter has been deprecated, so the PCI SYS
   interface is the preferred method.

   You can determine the maximum number of VFs a PF can support:

   # cat /sys/class/net/eth3/device/sriov_totalvfs
   63

4. Verify that the VFs have been created and are in up
   state. For example:

   # lspci | grep Ethernet
   82:00.0 Ethernet controller: Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection (rev 01)
   82:00.1 Ethernet controller: Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection (rev 01)
   82:10.0 Ethernet controller: Intel Corporation 82599 Ethernet Controller Virtual Function (rev 01)
   82:10.2 Ethernet controller: Intel Corporation 82599 Ethernet Controller Virtual Function (rev 01)
   82:10.4 Ethernet controller: Intel Corporation 82599 Ethernet Controller Virtual Function (rev 01)
   82:10.6 Ethernet controller: Intel Corporation 82599 Ethernet Controller Virtual Function (rev 01)
   82:11.0 Ethernet controller: Intel Corporation 82599 Ethernet Controller Virtual Function (rev 01)
   82:11.2 Ethernet controller: Intel Corporation 82599 Ethernet Controller Virtual Function (rev 01)
   82:11.4 Ethernet controller: Intel Corporation 82599 Ethernet Controller Virtual Function (rev 01)
   82:11.6 Ethernet controller: Intel Corporation 82599 Ethernet Controller Virtual Function (rev 01)

   # ip link show eth3
   8: eth3: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP mode DEFAULT qlen 1000
      link/ether a0:36:9f:8f:3f:b8 brd ff:ff:ff:ff:ff:ff
      vf 0 MAC 00:00:00:00:00:00, spoof checking on, link-state auto
      vf 1 MAC 00:00:00:00:00:00, spoof checking on, link-state auto
      vf 2 MAC 00:00:00:00:00:00, spoof checking on, link-state auto
      vf 3 MAC 00:00:00:00:00:00, spoof checking on, link-state auto
      vf 4 MAC 00:00:00:00:00:00, spoof checking on, link-state auto
      vf 5 MAC 00:00:00:00:00:00, spoof checking on, link-state auto
      vf 6 MAC 00:00:00:00:00:00, spoof checking on, link-state auto
      vf 7 MAC 00:00:00:00:00:00, spoof checking on, link-state auto

   If the interfaces are down, set them to up before
   launching a guest, otherwise the instance will fail to spawn:

   # ip link set eth3 up

5. Persist created VFs on reboot:

   # echo "echo '7' > /sys/class/net/eth3/device/sriov_numvfs" >> /etc/rc.local

   Note

   The suggested way of making PCI SYS settings persistent is through
   the sysfsutils tool. However, this is not available by
   default on many major distributions.

Configuring
allow list for PCI devices nova-compute (Compute)

1. Configure which PCI devices the nova-compute service
   may use. Edit the nova.conf file:

   [pci]
   passthrough_whitelist = { "devname": "eth3", "physical_network": "physnet2"}

   This tells the Compute service that all VFs belonging to
   eth3 are allowed to be passed through to instances and
   belong to the provider network physnet2.

   Alternatively the [pci] passthrough_whitelist parameter
   also supports allowing devices by:

   • PCI address: The address uses the same syntax as in
     lspci and an asterisk (*) can be used to match
     anything.

     [pci]
     passthrough_whitelist = { "address": "[[[[<domain>]:]<bus>]:][<slot>][.[<function>]]", "physical_network": "physnet2" }

     For example, to match any domain, bus 0a, slot
     00, and all functions:

     [pci]
     passthrough_whitelist = { "address": "*:0a:00.*", "physical_network": "physnet2" }

   • PCI vendor_id and product_id as
     displayed by the Linux utility lspci.

     [pci]
     passthrough_whitelist = { "vendor_id": "<id>", "product_id": "<id>", "physical_network": "physnet2" }

   If the device defined by the PCI address or devname
   corresponds to an SR-IOV PF, all VFs under the PF will match the entry.
   Multiple [pci] passthrough_whitelist entries per host are
   supported.

   In order to enable SR-IOV to request “trusted mode”, the
   [pci] passthrough_whitelist parameter also supports a
   trusted tag.

   Note

   This capability is only supported starting with version 18.0.0
   (Rocky) release of the compute service configured to use the libvirt
   driver.

   Important

   There are security implications of enabling trusted ports. The
   trusted VFs can be set into VF promiscuous mode which will enable it to
   receive unmatched and multicast traffic sent to the physical
   function.

   For example, to allow users to request SR-IOV devices with trusted
   capabilities on device eth3:

   [pci]
   passthrough_whitelist = { "devname": "eth3", "physical_network": "physnet2", "trusted":"true" }

   The ports will have to be created with a binding profile to match the
   trusted tag, see Launching instances with
   SR-IOV ports.

2. Restart the nova-compute service for the changes to
   go into effect.

Configure neutron-server
(Controller)

Note

This section does not apply to remote-managed ports of SmartNIC DPU
devices which also use SR-IOV at the host side but do not rely on the
sriovnicswitch mechanism driver.

1. Add sriovnicswitch as mechanism driver. Edit the
   ml2_conf.ini file on each controller:

   [ml2]
   mechanism_drivers = openvswitch,sriovnicswitch

2. Ensure your physnet is configured for the chosen network type.
   Edit the ml2_conf.ini file on each controller:

   [ml2_type_vlan]
   network_vlan_ranges = physnet2

3. Add the plugin.ini file as a parameter to the
   neutron-server service. Edit the appropriate initialization
   script to configure the neutron-server service to load the
   plugin configuration file:

   --config-file /etc/neutron/neutron.conf
   --config-file /etc/neutron/plugin.ini

4. Restart the neutron-server service.

Configure nova-scheduler
(Controller)

1. On every controller node running the nova-scheduler
   service, add PciPassthroughFilter to
   [filter_scheduler] enabled_filters to enable this filter.
   Ensure [filter_scheduler] available_filters is set to the
   default of nova.scheduler.filters.all_filters:

   [filter_scheduler]
   enabled_filters = AvailabilityZoneFilter, ComputeFilter, ComputeCapabilitiesFilter, ImagePropertiesFilter, ServerGroupAntiAffinityFilter, ServerGroupAffinityFilter, PciPassthroughFilter
   available_filters = nova.scheduler.filters.all_filters

2. Restart the nova-scheduler service.

Enable
neutron-sriov-nic-agent (Compute)

1. Install the SR-IOV agent, if necessary.

2. Edit the sriov_agent.ini file on each compute node.
   For example:

   [securitygroup]
   firewall_driver = neutron.agent.firewall.NoopFirewallDriver
   
   [sriov_nic]
   physical_device_mappings = physnet2:eth3
   exclude_devices =

   Note

   The physical_device_mappings parameter is not limited to
   be a 1-1 mapping between physical networks and NICs. This enables you to
   map the same physical network to more than one NIC. For example, if
   physnet2 is connected to eth3 and
   eth4, then physnet2:eth3,physnet2:eth4 is a
   valid option.

   The exclude_devices parameter is empty, therefore, all
   the VFs associated with eth3 may be configured by the agent. To exclude
   specific VFs, add them to the exclude_devices parameter as
   follows:

   exclude_devices = eth1:0000:07:00.2;0000:07:00.3,eth2:0000:05:00.1;0000:05:00.2

3. Ensure the SR-IOV agent runs successfully:

   # neutron-sriov-nic-agent \
     --config-file /etc/neutron/neutron.conf \
     --config-file /etc/neutron/plugins/ml2/sriov_agent.ini

4. Enable the neutron SR-IOV agent service.

   If installing from source, you must configure a daemon file for the
   init system manually.

(Optional) FDB L2 agent
extension

Forwarding DataBase (FDB) population is an L2 agent extension to OVS
agent or Linux bridge. Its objective is to update the FDB table for
existing instance using normal port. This enables communication between
SR-IOV instances and normal instances. The use cases of the FDB
population extension are:

• Direct port and normal port instances reside on the same compute
  node.
• Direct port instance that uses floating IP address and network node
  are located on the same host.

For additional information describing the problem, refer to: Virtual
switching technologies and Linux bridge.

1. Edit the ovs_agent.ini or
   linuxbridge_agent.ini file on each compute node. For
   example:

   [agent]
   extensions = fdb

2. Add the FDB section and the
   shared_physical_device_mappings parameter. This parameter
   maps each physical port to its physical network name. Each physical
   network can be mapped to several ports:

   [FDB]
   shared_physical_device_mappings = physnet1:p1p1, physnet1:p1p2

Launching instances with
SR-IOV ports

Once configuration is complete, you can launch instances with SR-IOV
ports.

1. If it does not already exist, create a network and subnet for the
   chosen physnet. This is the network to which SR-IOV ports will be
   attached. For example:

   $ openstack network create --provider-physical-network physnet2 \
       --provider-network-type vlan --provider-segment 1000 \
       sriov-net
   
   $ openstack subnet create --network sriov-net \
       --subnet-pool shared-default-subnetpool-v4 \
       sriov-subnet

2. Get the id of the network where you want the SR-IOV
   port to be created:

   $ net_id=$(openstack network show sriov-net -c id -f value)

3. Create the SR-IOV port. vnic-type=direct is used
   here, but other options include normal,
   direct-physical, and macvtap:

   $ openstack port create --network $net_id --vnic-type direct \
       sriov-port

   Alternatively, to request that the SR-IOV port accept trusted
   capabilities, the binding profile should be enhanced with the
   trusted tag.

   $ openstack port create --network $net_id --vnic-type direct \
       --binding-profile trusted=true \
       sriov-port

4. Get the id of the created port:

   $ port_id=$(openstack port show sriov-port -c id -f value)

5. Create the instance. Specify the SR-IOV port created in step two
   for the NIC:

   $ openstack server create --flavor m1.large --image ubuntu_18.04 \
       --nic port-id=$port_id \
       test-sriov

   Note

   There are two ways to attach VFs to an instance. You can create an
   SR-IOV port or use the pci_alias in the Compute service.
   For more information about using pci_alias, refer to nova-api configuration__.

   __ https://docs.openstack.org/nova/latest/admin/pci-passthrough.html#configure-nova-api-controller

SR-IOV
with ConnectX-3/ConnectX-3 Pro Dual Port Ethernet

In contrast to Mellanox newer generation NICs, ConnectX-3 family
network adapters expose a single PCI device (PF) in the system
regardless of the number of physical ports. When the device is
dual port and SR-IOV is enabled and configured we can
observe some inconsistencies in linux networking subsystem.

Example: A system with ConnectX-3 dual port device
and a total of four VFs configured, two VFs assigned to port one and two
VFs assigned to port two.

$ lspci | grep Mellanox
04:00.0 Network controller: Mellanox Technologies MT27520 Family [ConnectX-3 Pro]
04:00.1 Network controller: Mellanox Technologies MT27500/MT27520 Family [ConnectX-3/ConnectX-3 Pro Virtual Function]
04:00.2 Network controller: Mellanox Technologies MT27500/MT27520 Family [ConnectX-3/ConnectX-3 Pro Virtual Function]
04:00.3 Network controller: Mellanox Technologies MT27500/MT27520 Family [ConnectX-3/ConnectX-3 Pro Virtual Function]
04:00.4 Network controller: Mellanox Technologies MT27500/MT27520 Family [ConnectX-3/ConnectX-3 Pro Virtual Function]

Four VFs are available in the system, however,

$ ip link show
31: enp4s0: <BROADCAST,MULTICAST> mtu 1500 qdisc noop master ovs-system state DOWN mode DEFAULT group default qlen 1000
    link/ether f4:52:14:01:d9:e1 brd ff:ff:ff:ff:ff:ff
    vf 0 MAC 00:00:00:00:00:00, vlan 4095, spoof checking off, link-state auto
    vf 1 MAC 00:00:00:00:00:00, vlan 4095, spoof checking off, link-state auto
    vf 2 MAC 00:00:00:00:00:00, vlan 4095, spoof checking off, link-state auto
    vf 3 MAC 00:00:00:00:00:00, vlan 4095, spoof checking off, link-state auto
32: enp4s0d1: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN mode DEFAULT group default qlen 1000
    link/ether f4:52:14:01:d9:e2 brd ff:ff:ff:ff:ff:ff
    vf 0 MAC 00:00:00:00:00:00, vlan 4095, spoof checking off, link-state auto
    vf 1 MAC 00:00:00:00:00:00, vlan 4095, spoof checking off, link-state auto
    vf 2 MAC 00:00:00:00:00:00, vlan 4095, spoof checking off, link-state auto
    vf 3 MAC 00:00:00:00:00:00, vlan 4095, spoof checking off, link-state auto

ip command identifies each PF associated net device
as having four VFs each.

To allow neutron SR-IOV agent to properly identify the VFs that
belong to the correct PF network device (thus to the correct network
port) Admin is required to provide the exclude_devices
configuration option in sriov_agent.ini

Step 1: derive the VF to Port mapping from mlx4
driver configuration file: /etc/modprobe.d/mlnx.conf or
/etc/modprobe.d/mlx4.conf

$ cat /etc/modprobe.d/mlnx.conf | grep "options mlx4_core"
options mlx4_core port_type_array=2,2 num_vfs=2,2,0 probe_vf=2,2,0 log_num_mgm_entry_size=-1

Where:

num_vfs=n1,n2,n3 – The driver will enable
n1 VFs on physical port 1, n2 VFs on physical
port 2 and n3 dual port VFs (applies only to dual port HCA
when all ports are Ethernet ports).

probe_vfs=m1,m2,m3 – the driver probes m1
single port VFs on physical port 1, m2 single port VFs on
physical port 2 (applies only if such a port exist) m3 dual
port VFs. Those VFs are attached to the hypervisor. (applies only if all
ports are configured as Ethernet).

The VFs will be enumerated in the following order:

1. port 1 VFs
2. port 2 VFs
3. dual port VFs

In our example:

Step 2: Update exclude_devices
configuration option in sriov_agent.ini with the correct
mapping

Each PF associated net device shall exclude the
other port’s VFs

[sriov_nic]
physical_device_mappings = physnet1:enp4s0,physnet2:enp4s0d1
exclude_devices = enp4s0:0000:04:00.3;0000:04:00.4,enp4s0d1:0000:04:00.1;0000:04:00.2

SR-IOV with InfiniBand

The support for SR-IOV with InfiniBand allows a Virtual PCI device
(VF) to be directly mapped to the guest, allowing higher performance and
advanced features such as RDMA (remote direct memory access). To use
this feature, you must:

1. Use InfiniBand enabled network adapters.

2. Run InfiniBand subnet managers to enable InfiniBand fabric.

   All InfiniBand networks must have a subnet manager running for the
   network to function. This is true even when doing a simple network of
   two machines with no switch and the cards are plugged in back-to-back. A
   subnet manager is required for the link on the cards to come up. It is
   possible to have more than one subnet manager. In this case, one of them
   will act as the primary, and any other will act as a backup that will
   take over when the primary subnet manager fails.

3. Install the ebrctl utility on the compute nodes.

   Check that ebrctl is listed somewhere in
   /etc/nova/rootwrap.d/*:

   $ grep 'ebrctl' /etc/nova/rootwrap.d/*

   If ebrctl does not appear in any of the rootwrap files,
   add this to the /etc/nova/rootwrap.d/compute.filters file
   in the [Filters] section.

   [Filters]
   ebrctl: CommandFilter, ebrctl, root

Known limitations

• When using Quality of Service (QoS), max_burst_kbps
  (burst over max_kbps) is not supported. In addition,
  max_kbps is rounded to Mbps.

• Security groups are not supported when using SR-IOV, thus, the
  firewall driver must be disabled. This can be done in the
  neutron.conf file.

  [securitygroup]
  firewall_driver = neutron.agent.firewall.NoopFirewallDriver

• SR-IOV is not integrated into the OpenStack Dashboard (horizon).
  Users must use the CLI or API to configure SR-IOV interfaces.

• Live migration support has been added to the Libvirt Nova
  virt-driver in the Train release for instances with neutron SR-IOV
  ports. Indirect mode SR-IOV interfaces (vnic-type: macvtap or
  virtio-forwarder) can now be migrated transparently to the guest. Direct
  mode SR-IOV interfaces (vnic-type: direct or direct-physical) are
  detached before the migration and reattached after the migration so this
  is not transparent to the guest. To avoid loss of network connectivy
  when live migrating with direct mode sriov the user should create a
  failover bond in the guest with a transparently live migration port type
  e.g. vnic-type normal or indirect mode SR-IOV.

  Note

  SR-IOV features may require a specific NIC driver version, depending
  on the vendor. Intel NICs, for example, require ixgbe version 4.4.6 or
  greater, and ixgbevf version 3.2.2 or greater.

• Attaching SR-IOV ports to existing servers is supported starting
  with the Victoria release.