SR-IOV

SR-IOV

The purpose of this page is to describe how to enable SR-IOV functionality available in OpenStack (using OpenStack Networking) as of the Juno release. It is intended to server 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:

Term Definition
PF Physical Function. The physical Ethernet controller that supports SR-IOV.
VF Virtual Function. The virtual PCIe device created from a physical Ethernet controller.

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. You must include the SR-IOV agent on each compute node using SR-IOV ports.

Note

The SR-IOV agent was optional before Mitaka, and was not enabled by default before Liberty.

Note

The ability to control port security and QoS rate limit settings was added in Liberty.

Supported Ethernet controllers

The following manufacturers are known to work:

  • Intel
  • Mellanox
  • QLogic

For information on Mellanox SR-IOV Ethernet ConnectX-3/ConnectX-3 Pro cards, see Mellanox: How To Configure SR-IOV VFs.

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

Using SR-IOV interfaces

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

  1. Create Virtual Functions (Compute)
  2. Whitelist 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.

Note

Throughout this guide, eth3 is used as the PF and physnet2 is used as the provider network configured as a VLAN range. These ports may vary in different environments.

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.

Note

The steps detail how to create VFs using Mellanox ConnectX-4 and newer/Intel SR-IOV Ethernet cards on an Intel system. Steps may differ for different hardware configurations.

  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.

    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:

    # 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.

Whitelist PCI devices nova-compute (Compute)

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

    [default]
    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 whitelisting 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.

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

Configure neutron-server (Controller)

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

    mechanism_drivers = openvswitch,sriovnicswitch
    
  2. Add the ml2_conf_sriov.ini file as parameter to the neutron-server service. Edit the appropriate initialization script to configure the neutron-server service to load the SR-IOV configuration file:

    --config-file /etc/neutron/neutron.conf
    --config-file /etc/neutron/plugin.ini
    --config-file /etc/neutron/plugins/ml2/ml2_conf_sriov.ini
    
  3. Restart the neutron-server service.

Configure nova-scheduler (Controller)

  1. On every controller node running the nova-scheduler service, add PciPassthroughFilter to scheduler_default_filters to enable PciPassthroughFilter by default. Also ensure scheduler_available_filters parameter under the [DEFAULT] section in nova.conf is set to all_filters to enable all filters provided by the Compute service.

    [DEFAULT]
    scheduler_default_filters = RetryFilter, AvailabilityZoneFilter, RamFilter, ComputeFilter, ComputeCapabilitiesFilter, ImagePropertiesFilter, ServerGroupAntiAffinityFilter, ServerGroupAffinityFilter, PciPassthroughFilter
    scheduler_available_filters = nova.scheduler.filters.all_filters
    
  2. Restart the nova-scheduler service.

Enable neutron sriov-agent (Compute)

  1. Install the SR-IOV agent.

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

    $ net_id=`neutron net-show net04 | grep "\ id\ " | awk '{ print $4 }'`
    
  2. Create the SR-IOV port. vnic_type=direct is used here, but other options include normal, direct-physical, and macvtap:

    $ port_id=`neutron port-create $net_id --name sriov_port --binding:vnic_type direct | grep "\ id\ " | awk '{ print $4 }'`
    
  3. Create the instance. Specify the SR-IOV port created in step two for the NIC:

    $ nova boot --flavor m1.large --image ubuntu_14.04 --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.

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 master, and any other will act as a slave that will take over when the master 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 is not supported for instances with SR-IOV ports.

Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.