18. Intel Virtual Function Driver

Supported Intel® Ethernet Controllers (see the DPDK Release Notes for details) support the following modes of operation in a virtualized environment:

  • SR-IOV mode: Involves direct assignment of part of the port resources to different guest operating systems using the PCI-SIG Single Root I/O Virtualization (SR IOV) standard, also known as “native mode” or “pass-through” mode. In this chapter, this mode is referred to as IOV mode.
  • VMDq mode: Involves central management of the networking resources by an IO Virtual Machine (IOVM) or a Virtual Machine Monitor (VMM), also known as software switch acceleration mode. In this chapter, this mode is referred to as the Next Generation VMDq mode.

18.1. SR-IOV Mode Utilization in a DPDK Environment

The DPDK uses the SR-IOV feature for hardware-based I/O sharing in IOV mode. Therefore, it is possible to partition SR-IOV capability on Ethernet controller NIC resources logically and expose them to a virtual machine as a separate PCI function called a “Virtual Function”. Refer to Fig. 18.4.

Therefore, a NIC is logically distributed among multiple virtual machines (as shown in Fig. 18.4), while still having global data in common to share with the Physical Function and other Virtual Functions. The DPDK fm10kvf, i40evf, igbvf or ixgbevf as a Poll Mode Driver (PMD) serves for the Intel® 82576 Gigabit Ethernet Controller, Intel® Ethernet Controller I350 family, Intel® 82599 10 Gigabit Ethernet Controller NIC, Intel® Fortville 10/40 Gigabit Ethernet Controller NIC’s virtual PCI function, or PCIe host-interface of the Intel Ethernet Switch FM10000 Series. Meanwhile the DPDK Poll Mode Driver (PMD) also supports “Physical Function” of such NIC’s on the host.

The DPDK PF/VF Poll Mode Driver (PMD) supports the Layer 2 switch on Intel® 82576 Gigabit Ethernet Controller, Intel® Ethernet Controller I350 family, Intel® 82599 10 Gigabit Ethernet Controller, and Intel® Fortville 10/40 Gigabit Ethernet Controller NICs so that guest can choose it for inter virtual machine traffic in SR-IOV mode.

For more detail on SR-IOV, please refer to the following documents:

../_images/single_port_nic.png

Fig. 18.4 Virtualization for a Single Port NIC in SR-IOV Mode

18.1.1. Physical and Virtual Function Infrastructure

The following describes the Physical Function and Virtual Functions infrastructure for the supported Ethernet Controller NICs.

Virtual Functions operate under the respective Physical Function on the same NIC Port and therefore have no access to the global NIC resources that are shared between other functions for the same NIC port.

A Virtual Function has basic access to the queue resources and control structures of the queues assigned to it. For global resource access, a Virtual Function has to send a request to the Physical Function for that port, and the Physical Function operates on the global resources on behalf of the Virtual Function. For this out-of-band communication, an SR-IOV enabled NIC provides a memory buffer for each Virtual Function, which is called a “Mailbox”.

18.1.1.1. Intel® Ethernet Adaptive Virtual Function

Adaptive Virtual Function (AVF) is a SR-IOV Virtual Function with the same device id (8086:1889) on different Intel Ethernet Controller. AVF Driver is VF driver which supports for all future Intel devices without requiring a VM update. And since this happens to be an adaptive VF driver, every new drop of the VF driver would add more and more advanced features that can be turned on in the VM if the underlying HW device supports those advanced features based on a device agnostic way without ever compromising on the base functionality. AVF provides generic hardware interface and interface between AVF driver and a compliant PF driver is specified.

Intel products starting Ethernet Controller 700 Series to support Adaptive Virtual Function.

The way to generate Virtual Function is like normal, and the resource of VF assignment depends on the NIC Infrastructure.

For more detail on SR-IOV, please refer to the following documents:

Note

To use DPDK AVF PMD on Intel® 700 Series Ethernet Controller, the device id (0x1889) need to specified during device assignment in hypervisor. Take qemu for example, the device assignment should carry the AVF device id (0x1889) like -device vfio-pci,x-pci-device-id=0x1889,host=03:0a.0.

18.1.1.2. The PCIE host-interface of Intel Ethernet Switch FM10000 Series VF infrastructure

In a virtualized environment, the programmer can enable a maximum of 64 Virtual Functions (VF) globally per PCIE host-interface of the Intel Ethernet Switch FM10000 Series device. Each VF can have a maximum of 16 queue pairs. The Physical Function in host could be only configured by the Linux* fm10k driver (in the case of the Linux Kernel-based Virtual Machine [KVM]), DPDK PMD PF driver doesn’t support it yet.

For example,

  • Using Linux* fm10k driver:

    rmmod fm10k (To remove the fm10k module)
    insmod fm0k.ko max_vfs=2,2 (To enable two Virtual Functions per port)
    

Virtual Function enumeration is performed in the following sequence by the Linux* pci driver for a dual-port NIC. When you enable the four Virtual Functions with the above command, the four enabled functions have a Function# represented by (Bus#, Device#, Function#) in sequence starting from 0 to 3. However:

  • Virtual Functions 0 and 2 belong to Physical Function 0
  • Virtual Functions 1 and 3 belong to Physical Function 1

Note

The above is an important consideration to take into account when targeting specific packets to a selected port.

18.1.1.3. Intel® X710/XL710 Gigabit Ethernet Controller VF Infrastructure

In a virtualized environment, the programmer can enable a maximum of 128 Virtual Functions (VF) globally per Intel® X710/XL710 Gigabit Ethernet Controller NIC device. The number of queue pairs of each VF can be configured by CONFIG_RTE_LIBRTE_I40E_QUEUE_NUM_PER_VF in config file. The Physical Function in host could be either configured by the Linux* i40e driver (in the case of the Linux Kernel-based Virtual Machine [KVM]) or by DPDK PMD PF driver. When using both DPDK PMD PF/VF drivers, the whole NIC will be taken over by DPDK based application.

For example,

  • Using Linux* i40e driver:

    rmmod i40e (To remove the i40e module)
    insmod i40e.ko max_vfs=2,2 (To enable two Virtual Functions per port)
    
  • Using the DPDK PMD PF i40e driver:

    Kernel Params: iommu=pt, intel_iommu=on

    modprobe uio
    insmod igb_uio
    ./dpdk-devbind.py -b igb_uio bb:ss.f
    echo 2 > /sys/bus/pci/devices/0000\:bb\:ss.f/max_vfs (To enable two VFs on a specific PCI device)
    

    Launch the DPDK testpmd/example or your own host daemon application using the DPDK PMD library.

Virtual Function enumeration is performed in the following sequence by the Linux* pci driver for a dual-port NIC. When you enable the four Virtual Functions with the above command, the four enabled functions have a Function# represented by (Bus#, Device#, Function#) in sequence starting from 0 to 3. However:

  • Virtual Functions 0 and 2 belong to Physical Function 0
  • Virtual Functions 1 and 3 belong to Physical Function 1

Note

The above is an important consideration to take into account when targeting specific packets to a selected port.

For Intel® X710/XL710 Gigabit Ethernet Controller, queues are in pairs. One queue pair means one receive queue and one transmit queue. The default number of queue pairs per VF is 4, and can be 16 in maximum.

18.1.1.4. Intel® 82599 10 Gigabit Ethernet Controller VF Infrastructure

The programmer can enable a maximum of 63 Virtual Functions and there must be one Physical Function per Intel® 82599 10 Gigabit Ethernet Controller NIC port. The reason for this is that the device allows for a maximum of 128 queues per port and a virtual/physical function has to have at least one queue pair (RX/TX). The current implementation of the DPDK ixgbevf driver supports a single queue pair (RX/TX) per Virtual Function. The Physical Function in host could be either configured by the Linux* ixgbe driver (in the case of the Linux Kernel-based Virtual Machine [KVM]) or by DPDK PMD PF driver. When using both DPDK PMD PF/VF drivers, the whole NIC will be taken over by DPDK based application.

For example,

  • Using Linux* ixgbe driver:

    rmmod ixgbe (To remove the ixgbe module)
    insmod ixgbe max_vfs=2,2 (To enable two Virtual Functions per port)
    
  • Using the DPDK PMD PF ixgbe driver:

    Kernel Params: iommu=pt, intel_iommu=on

    modprobe uio
    insmod igb_uio
    ./dpdk-devbind.py -b igb_uio bb:ss.f
    echo 2 > /sys/bus/pci/devices/0000\:bb\:ss.f/max_vfs (To enable two VFs on a specific PCI device)
    

    Launch the DPDK testpmd/example or your own host daemon application using the DPDK PMD library.

  • Using the DPDK PMD PF ixgbe driver to enable VF RSS:

    Same steps as above to install the modules of uio, igb_uio, specify max_vfs for PCI device, and launch the DPDK testpmd/example or your own host daemon application using the DPDK PMD library.

    The available queue number (at most 4) per VF depends on the total number of pool, which is determined by the max number of VF at PF initialization stage and the number of queue specified in config:

    • If the max number of VFs (max_vfs) is set in the range of 1 to 32:

      If the number of Rx queues is specified as 4 (--rxq=4 in testpmd), then there are totally 32 pools (ETH_32_POOLS), and each VF could have 4 Rx queues;

      If the number of Rx queues is specified as 2 (--rxq=2 in testpmd), then there are totally 32 pools (ETH_32_POOLS), and each VF could have 2 Rx queues;

    • If the max number of VFs (max_vfs) is in the range of 33 to 64:

      If the number of Rx queues in specified as 4 (--rxq=4 in testpmd), then error message is expected as rxq is not correct at this case;

      If the number of rxq is 2 (--rxq=2 in testpmd), then there is totally 64 pools (ETH_64_POOLS), and each VF have 2 Rx queues;

    On host, to enable VF RSS functionality, rx mq mode should be set as ETH_MQ_RX_VMDQ_RSS or ETH_MQ_RX_RSS mode, and SRIOV mode should be activated (max_vfs >= 1). It also needs config VF RSS information like hash function, RSS key, RSS key length.

Note

The limitation for VF RSS on Intel® 82599 10 Gigabit Ethernet Controller is: The hash and key are shared among PF and all VF, the RETA table with 128 entries is also shared among PF and all VF; So it could not to provide a method to query the hash and reta content per VF on guest, while, if possible, please query them on host for the shared RETA information.

Virtual Function enumeration is performed in the following sequence by the Linux* pci driver for a dual-port NIC. When you enable the four Virtual Functions with the above command, the four enabled functions have a Function# represented by (Bus#, Device#, Function#) in sequence starting from 0 to 3. However:

  • Virtual Functions 0 and 2 belong to Physical Function 0
  • Virtual Functions 1 and 3 belong to Physical Function 1

Note

The above is an important consideration to take into account when targeting specific packets to a selected port.

18.1.1.5. Intel® 82576 Gigabit Ethernet Controller and Intel® Ethernet Controller I350 Family VF Infrastructure

In a virtualized environment, an Intel® 82576 Gigabit Ethernet Controller serves up to eight virtual machines (VMs). The controller has 16 TX and 16 RX queues. They are generally referred to (or thought of) as queue pairs (one TX and one RX queue). This gives the controller 16 queue pairs.

A pool is a group of queue pairs for assignment to the same VF, used for transmit and receive operations. The controller has eight pools, with each pool containing two queue pairs, that is, two TX and two RX queues assigned to each VF.

In a virtualized environment, an Intel® Ethernet Controller I350 family device serves up to eight virtual machines (VMs) per port. The eight queues can be accessed by eight different VMs if configured correctly (the i350 has 4x1GbE ports each with 8T X and 8 RX queues), that means, one Transmit and one Receive queue assigned to each VF.

For example,

  • Using Linux* igb driver:

    rmmod igb (To remove the igb module)
    insmod igb max_vfs=2,2 (To enable two Virtual Functions per port)
    
  • Using DPDK PMD PF igb driver:

    Kernel Params: iommu=pt, intel_iommu=on modprobe uio

    insmod igb_uio
    ./dpdk-devbind.py -b igb_uio bb:ss.f
    echo 2 > /sys/bus/pci/devices/0000\:bb\:ss.f/max_vfs (To enable two VFs on a specific pci device)
    

    Launch DPDK testpmd/example or your own host daemon application using the DPDK PMD library.

Virtual Function enumeration is performed in the following sequence by the Linux* pci driver for a four-port NIC. When you enable the four Virtual Functions with the above command, the four enabled functions have a Function# represented by (Bus#, Device#, Function#) in sequence, starting from 0 to 7. However:

  • Virtual Functions 0 and 4 belong to Physical Function 0
  • Virtual Functions 1 and 5 belong to Physical Function 1
  • Virtual Functions 2 and 6 belong to Physical Function 2
  • Virtual Functions 3 and 7 belong to Physical Function 3

Note

The above is an important consideration to take into account when targeting specific packets to a selected port.

18.1.2. Validated Hypervisors

The validated hypervisor is:

  • KVM (Kernel Virtual Machine) with Qemu, version 0.14.0

However, the hypervisor is bypassed to configure the Virtual Function devices using the Mailbox interface, the solution is hypervisor-agnostic. Xen* and VMware* (when SR- IOV is supported) will also be able to support the DPDK with Virtual Function driver support.

18.1.3. Expected Guest Operating System in Virtual Machine

The expected guest operating systems in a virtualized environment are:

  • Fedora* 14 (64-bit)
  • Ubuntu* 10.04 (64-bit)

For supported kernel versions, refer to the DPDK Release Notes.

18.2. Setting Up a KVM Virtual Machine Monitor

The following describes a target environment:

  • Host Operating System: Fedora 14
  • Hypervisor: KVM (Kernel Virtual Machine) with Qemu version 0.14.0
  • Guest Operating System: Fedora 14
  • Linux Kernel Version: Refer to the DPDK Getting Started Guide
  • Target Applications: l2fwd, l3fwd-vf

The setup procedure is as follows:

  1. Before booting the Host OS, open BIOS setup and enable Intel® VT features.

  2. While booting the Host OS kernel, pass the intel_iommu=on kernel command line argument using GRUB. When using DPDK PF driver on host, pass the iommu=pt kernel command line argument in GRUB.

  3. Download qemu-kvm-0.14.0 from http://sourceforge.net/projects/kvm/files/qemu-kvm/ and install it in the Host OS using the following steps:

    When using a recent kernel (2.6.25+) with kvm modules included:

    tar xzf qemu-kvm-release.tar.gz
    cd qemu-kvm-release
    ./configure --prefix=/usr/local/kvm
    make
    sudo make install
    sudo /sbin/modprobe kvm-intel
    

    When using an older kernel, or a kernel from a distribution without the kvm modules, you must download (from the same link), compile and install the modules yourself:

    tar xjf kvm-kmod-release.tar.bz2
    cd kvm-kmod-release
    ./configure
    make
    sudo make install
    sudo /sbin/modprobe kvm-intel
    

    qemu-kvm installs in the /usr/local/bin directory.

    For more details about KVM configuration and usage, please refer to:

    http://www.linux-kvm.org/page/HOWTO1.

  4. Create a Virtual Machine and install Fedora 14 on the Virtual Machine. This is referred to as the Guest Operating System (Guest OS).

  5. Download and install the latest ixgbe driver from:

    http://downloadcenter.intel.com/Detail_Desc.aspx?agr=Y&DwnldID=14687

  6. In the Host OS

    When using Linux kernel ixgbe driver, unload the Linux ixgbe driver and reload it with the max_vfs=2,2 argument:

    rmmod ixgbe
    modprobe ixgbe max_vfs=2,2
    

    When using DPDK PMD PF driver, insert DPDK kernel module igb_uio and set the number of VF by sysfs max_vfs:

    modprobe uio
    insmod igb_uio
    ./dpdk-devbind.py -b igb_uio 02:00.0 02:00.1 0e:00.0 0e:00.1
    echo 2 > /sys/bus/pci/devices/0000\:02\:00.0/max_vfs
    echo 2 > /sys/bus/pci/devices/0000\:02\:00.1/max_vfs
    echo 2 > /sys/bus/pci/devices/0000\:0e\:00.0/max_vfs
    echo 2 > /sys/bus/pci/devices/0000\:0e\:00.1/max_vfs
    

    Note

    You need to explicitly specify number of vfs for each port, for example, in the command above, it creates two vfs for the first two ixgbe ports.

    Let say we have a machine with four physical ixgbe ports:

    0000:02:00.0

    0000:02:00.1

    0000:0e:00.0

    0000:0e:00.1

    The command above creates two vfs for device 0000:02:00.0:

    ls -alrt /sys/bus/pci/devices/0000\:02\:00.0/virt*
    lrwxrwxrwx. 1 root root 0 Apr 13 05:40 /sys/bus/pci/devices/0000:02:00.0/virtfn1 -> ../0000:02:10.2
    lrwxrwxrwx. 1 root root 0 Apr 13 05:40 /sys/bus/pci/devices/0000:02:00.0/virtfn0 -> ../0000:02:10.0
    

    It also creates two vfs for device 0000:02:00.1:

    ls -alrt /sys/bus/pci/devices/0000\:02\:00.1/virt*
    lrwxrwxrwx. 1 root root 0 Apr 13 05:51 /sys/bus/pci/devices/0000:02:00.1/virtfn1 -> ../0000:02:10.3
    lrwxrwxrwx. 1 root root 0 Apr 13 05:51 /sys/bus/pci/devices/0000:02:00.1/virtfn0 -> ../0000:02:10.1
    
  7. List the PCI devices connected and notice that the Host OS shows two Physical Functions (traditional ports) and four Virtual Functions (two for each port). This is the result of the previous step.

  8. Insert the pci_stub module to hold the PCI devices that are freed from the default driver using the following command (see http://www.linux-kvm.org/page/How_to_assign_devices_with_VT-d_in_KVM Section 4 for more information):

    sudo /sbin/modprobe pci-stub
    

    Unbind the default driver from the PCI devices representing the Virtual Functions. A script to perform this action is as follows:

    echo "8086 10ed" > /sys/bus/pci/drivers/pci-stub/new_id
    echo 0000:08:10.0 > /sys/bus/pci/devices/0000:08:10.0/driver/unbind
    echo 0000:08:10.0 > /sys/bus/pci/drivers/pci-stub/bind
    

    where, 0000:08:10.0 belongs to the Virtual Function visible in the Host OS.

  9. Now, start the Virtual Machine by running the following command:

    /usr/local/kvm/bin/qemu-system-x86_64 -m 4096 -smp 4 -boot c -hda lucid.qcow2 -device pci-assign,host=08:10.0
    

    where:

    — -m = memory to assign

    —-smp = number of smp cores

    — -boot = boot option

    —-hda = virtual disk image

    — -device = device to attach

    Note

    — The pci-assign,host=08:10.0 value indicates that you want to attach a PCI device to a Virtual Machine and the respective (Bus:Device.Function) numbers should be passed for the Virtual Function to be attached.

    — qemu-kvm-0.14.0 allows a maximum of four PCI devices assigned to a VM, but this is qemu-kvm version dependent since qemu-kvm-0.14.1 allows a maximum of five PCI devices.

    — qemu-system-x86_64 also has a -cpu command line option that is used to select the cpu_model to emulate in a Virtual Machine. Therefore, it can be used as:

    /usr/local/kvm/bin/qemu-system-x86_64 -cpu ?
    
    (to list all available cpu_models)
    
    /usr/local/kvm/bin/qemu-system-x86_64 -m 4096 -cpu host -smp 4 -boot c -hda lucid.qcow2 -device pci-assign,host=08:10.0
    
    (to use the same cpu_model equivalent to the host cpu)
    

    For more information, please refer to: http://wiki.qemu.org/Features/CPUModels.

  10. If use vfio-pci to pass through device instead of pci-assign, steps 8 and 9 need to be updated to bind device to vfio-pci and replace pci-assign with vfio-pci when start virtual machine.

    sudo /sbin/modprobe vfio-pci
    
    echo "8086 10ed" > /sys/bus/pci/drivers/vfio-pci/new_id
    echo 0000:08:10.0 > /sys/bus/pci/devices/0000:08:10.0/driver/unbind
    echo 0000:08:10.0 > /sys/bus/pci/drivers/vfio-pci/bind
    
    /usr/local/kvm/bin/qemu-system-x86_64 -m 4096 -smp 4 -boot c -hda lucid.qcow2 -device vfio-pci,host=08:10.0
    
  11. Install and run DPDK host app to take over the Physical Function. Eg.

    make install T=x86_64-native-linuxapp-gcc
    ./x86_64-native-linuxapp-gcc/app/testpmd -l 0-3 -n 4 -- -i
    
  12. Finally, access the Guest OS using vncviewer with the localhost:5900 port and check the lspci command output in the Guest OS. The virtual functions will be listed as available for use.

  13. Configure and install the DPDK with an x86_64-native-linuxapp-gcc configuration on the Guest OS as normal, that is, there is no change to the normal installation procedure.

    make config T=x86_64-native-linuxapp-gcc O=x86_64-native-linuxapp-gcc
    cd x86_64-native-linuxapp-gcc
    make
    

Note

If you are unable to compile the DPDK and you are getting “error: CPU you selected does not support x86-64 instruction set”, power off the Guest OS and start the virtual machine with the correct -cpu option in the qemu- system-x86_64 command as shown in step 9. You must select the best x86_64 cpu_model to emulate or you can select host option if available.

Note

Run the DPDK l2fwd sample application in the Guest OS with Hugepages enabled. For the expected benchmark performance, you must pin the cores from the Guest OS to the Host OS (taskset can be used to do this) and you must also look at the PCI Bus layout on the board to ensure you are not running the traffic over the QPI Interface.

Note

  • The Virtual Machine Manager (the Fedora package name is virt-manager) is a utility for virtual machine management that can also be used to create, start, stop and delete virtual machines. If this option is used, step 2 and 6 in the instructions provided will be different.
  • virsh, a command line utility for virtual machine management, can also be used to bind and unbind devices to a virtual machine in Ubuntu. If this option is used, step 6 in the instructions provided will be different.
  • The Virtual Machine Monitor (see Fig. 18.5) is equivalent to a Host OS with KVM installed as described in the instructions.
../_images/perf_benchmark.png

Fig. 18.5 Performance Benchmark Setup

18.3. DPDK SR-IOV PMD PF/VF Driver Usage Model

18.3.1. Fast Host-based Packet Processing

Software Defined Network (SDN) trends are demanding fast host-based packet handling. In a virtualization environment, the DPDK VF PMD driver performs the same throughput result as a non-VT native environment.

With such host instance fast packet processing, lots of services such as filtering, QoS, DPI can be offloaded on the host fast path.

Fig. 18.6 shows the scenario where some VMs directly communicate externally via a VFs, while others connect to a virtual switch and share the same uplink bandwidth.

../_images/fast_pkt_proc.png

Fig. 18.6 Fast Host-based Packet Processing

18.4. SR-IOV (PF/VF) Approach for Inter-VM Communication

Inter-VM data communication is one of the traffic bottle necks in virtualization platforms. SR-IOV device assignment helps a VM to attach the real device, taking advantage of the bridge in the NIC. So VF-to-VF traffic within the same physical port (VM0<->VM1) have hardware acceleration. However, when VF crosses physical ports (VM0<->VM2), there is no such hardware bridge. In this case, the DPDK PMD PF driver provides host forwarding between such VMs.

Fig. 18.7 shows an example. In this case an update of the MAC address lookup tables in both the NIC and host DPDK application is required.

In the NIC, writing the destination of a MAC address belongs to another cross device VM to the PF specific pool. So when a packet comes in, its destination MAC address will match and forward to the host DPDK PMD application.

In the host DPDK application, the behavior is similar to L2 forwarding, that is, the packet is forwarded to the correct PF pool. The SR-IOV NIC switch forwards the packet to a specific VM according to the MAC destination address which belongs to the destination VF on the VM.

../_images/inter_vm_comms.png

Fig. 18.7 Inter-VM Communication