40. NVIDIA MLX4 Ethernet Driver

Note

NVIDIA acquired Mellanox Technologies in 2020. The DPDK documentation and code might still include instances of or references to Mellanox trademarks (like BlueField and ConnectX) that are now NVIDIA trademarks.

The MLX4 poll mode driver library (librte_net_mlx4) implements support for NVIDIA ConnectX-3 and NVIDIA ConnectX-3 Pro 10/40 Gbps adapters as well as their virtual functions (VF) in SR-IOV context.

There is also a section dedicated to this poll mode driver.

40.1. Implementation details

Most NVIDIA ConnectX-3 devices provide two ports but expose a single PCI bus address, thus unlike most drivers, librte_net_mlx4 registers itself as a PCI driver that allocates one Ethernet device per detected port.

For this reason, one cannot block (or allow) a single port without also blocking (or allowing) the others on the same device.

Besides its dependency on libibverbs (that implies libmlx4 and associated kernel support), librte_net_mlx4 relies heavily on system calls for control operations such as querying/updating the MTU and flow control parameters.

For security reasons and robustness, this driver only deals with virtual memory addresses. The way resources allocations are handled by the kernel combined with hardware specifications that allow it to handle virtual memory addresses directly ensure that DPDK applications cannot access random physical memory (or memory that does not belong to the current process).

This capability allows the PMD to coexist with kernel network interfaces which remain functional, although they stop receiving unicast packets as long as they share the same MAC address.

The Flow isolated mode is supported.

Compiling librte_net_mlx4 causes DPDK to be linked against libibverbs.

40.2. Configuration

40.2.1. Compilation options

The ibverbs libraries can be linked with this PMD in a number of ways, configured by the ibverbs_link build option:

  • shared (default): the PMD depends on some .so files.
  • dlopen: Split the dependencies glue in a separate library loaded when needed by dlopen. It make dependencies on libibverbs and libmlx4 optional, and has no performance impact.
  • static: Embed static flavor of the dependencies libibverbs and libmlx4 in the PMD shared library or the executable static binary.

40.2.2. Environment variables

  • MLX4_GLUE_PATH

    A list of directories in which to search for the rdma-core “glue” plug-in, separated by colons or semi-colons.

40.2.3. Run-time configuration

  • librte_net_mlx4 brings kernel network interfaces up during initialization because it is affected by their state. Forcing them down prevents packets reception.

  • ethtool operations on related kernel interfaces also affect the PMD.

  • port parameter [int]

    This parameter provides a physical port to probe and can be specified multiple times for additional ports. All ports are probed by default if left unspecified.

  • mr_ext_memseg_en parameter [int]

    A nonzero value enables extending memseg when registering DMA memory. If enabled, the number of entries in MR (Memory Region) lookup table on datapath is minimized and it benefits performance. On the other hand, it worsens memory utilization because registered memory is pinned by kernel driver. Even if a page in the extended chunk is freed, that doesn’t become reusable until the entire memory is freed.

    Enabled by default.

40.2.4. Kernel module parameters

The mlx4_core kernel module has several parameters that affect the behavior and/or the performance of librte_net_mlx4. Some of them are described below.

  • num_vfs (integer or triplet, optionally prefixed by device address strings)

    Create the given number of VFs on the specified devices.

  • log_num_mgm_entry_size (integer)

    Device-managed flow steering (DMFS) is required by DPDK applications. It is enabled by using a negative value, the last four bits of which have a special meaning.

    • -1: force device-managed flow steering (DMFS).
    • -7: configure optimized steering mode to improve performance with the following limitation: VLAN filtering is not supported with this mode. This is the recommended mode in case VLAN filter is not needed.

40.3. Limitations

  • For secondary process:
    • Forked secondary process not supported.
    • External memory unregistered in EAL memseg list cannot be used for DMA unless such memory has been registered by mlx4_mr_update_ext_mp() in primary process and remapped to the same virtual address in secondary process. If the external memory is registered by primary process but has different virtual address in secondary process, unexpected error may happen.
  • CRC stripping is supported by default and always reported as “true”. The ability to enable/disable CRC stripping requires OFED version 4.3-1.5.0.0 and above or rdma-core version v18 and above.
  • TSO (Transmit Segmentation Offload) is supported in OFED version 4.4 and above.

40.4. Prerequisites

This driver relies on external libraries and kernel drivers for resources allocations and initialization. The following dependencies are not part of DPDK and must be installed separately:

  • libibverbs (provided by rdma-core package)

    User space verbs framework used by librte_net_mlx4. This library provides a generic interface between the kernel and low-level user space drivers such as libmlx4.

    It allows slow and privileged operations (context initialization, hardware resources allocations) to be managed by the kernel and fast operations to never leave user space.

  • libmlx4 (provided by rdma-core package)

    Low-level user space driver library for NVIDIA ConnectX-3 devices, it is automatically loaded by libibverbs.

    This library basically implements send/receive calls to the hardware queues.

  • Kernel modules

    They provide the kernel-side verbs API and low level device drivers that manage actual hardware initialization and resources sharing with user space processes.

    Unlike most other PMDs, these modules must remain loaded and bound to their devices:

    • mlx4_core: hardware driver managing NVIDIA ConnectX-3 devices.
    • mlx4_en: Ethernet device driver that provides kernel network interfaces.
    • mlx4_ib: InfiniBand device driver.
    • ib_uverbs: user space driver for verbs (entry point for libibverbs).
  • Firmware update

    NVIDIA MLNX_OFED releases include firmware updates for ConnectX-3 adapters.

    Because each release provides new features, these updates must be applied to match the kernel modules and libraries they come with.

Note

Both libraries are BSD and GPL licensed. Linux kernel modules are GPL licensed.

Depending on system constraints and user preferences either RDMA core library with a recent enough Linux kernel release (recommended) or NVIDIA MLNX_OFED, which provides compatibility with older releases.

40.4.2. NVIDIA MLNX_OFED as a fallback

  • NVIDIA MLNX_OFED version: 4.4, 4.5, 4.6.
  • firmware version: 2.42.5000 and above.

Note

Several versions of NVIDIA MLNX_OFED are available. Installing the version this DPDK release was developed and tested against is strongly recommended. Please check the prerequisites.

40.4.2.1. Installing NVIDIA MLNX_OFED

  1. Download latest NVIDIA MLNX_OFED.

  2. Install the required libraries and kernel modules either by installing only the required set, or by installing the entire NVIDIA MLNX_OFED:

    For bare metal use:

    ./mlnxofedinstall --dpdk --upstream-libs
    

    For SR-IOV hypervisors use:

    ./mlnxofedinstall --dpdk --upstream-libs --enable-sriov --hypervisor
    

    For SR-IOV virtual machine use:

    ./mlnxofedinstall --dpdk --upstream-libs --guest
    
  3. Verify the firmware is the correct one:

    ibv_devinfo
    
  4. Set all ports links to Ethernet, follow instructions on the screen:

    connectx_port_config
    
  5. Continue with section 2 of the Quick Start Guide.

40.5. Quick Start Guide

  1. Set all ports links to Ethernet:

    PCI=<NIC PCI address>
    echo eth > "/sys/bus/pci/devices/$PCI/mlx4_port0"
    echo eth > "/sys/bus/pci/devices/$PCI/mlx4_port1"
    

    Note

    If using NVIDIA MLNX_OFED one can permanently set the port link to Ethernet using connectx_port_config tool provided by it. NVIDIA MLNX_OFED as a fallback:

  1. In case of bare metal or hypervisor, configure optimized steering mode by adding the following line to /etc/modprobe.d/mlx4_core.conf:

    options mlx4_core log_num_mgm_entry_size=-7
    

    Note

    If VLAN filtering is used, set log_num_mgm_entry_size=-1. Performance degradation can occur on this case.

  2. Restart the driver:

    /etc/init.d/openibd restart
    

    or:

    service openibd restart
    
  3. Install DPDK and you are ready to go. See compilation instructions.

40.6. Performance tuning

  1. Verify the optimized steering mode is configured:

    cat /sys/module/mlx4_core/parameters/log_num_mgm_entry_size
    
  2. Use the CPU near local NUMA node to which the PCIe adapter is connected, for better performance. For VMs, verify that the right CPU and NUMA node are pinned according to the above. Run:

    lstopo-no-graphics --merge
    

    to identify the NUMA node to which the PCIe adapter is connected.

  3. If more than one adapter is used, and root complex capabilities allow to put both adapters on the same NUMA node without PCI bandwidth degradation, it is recommended to locate both adapters on the same NUMA node. This in order to forward packets from one to the other without NUMA performance penalty.

  4. Disable pause frames:

    ethtool -A <netdev> rx off tx off
    
  5. Verify IO non-posted prefetch is disabled by default. This can be checked via the BIOS configuration. Please contact you server provider for more information about the settings.

Note

On some machines, depends on the machine integrator, it is beneficial to set the PCI max read request parameter to 1K. This can be done in the following way:

To query the read request size use:

setpci -s <NIC PCI address> 68.w

If the output is different than 3XXX, set it by:

setpci -s <NIC PCI address> 68.w=3XXX

The XXX can be different on different systems. Make sure to configure according to the setpci output.

  1. To minimize overhead of searching Memory Regions:
    • ‘–socket-mem’ is recommended to pin memory by predictable amount.
    • Configure per-lcore cache when creating Mempools for packet buffer.
    • Refrain from dynamically allocating/freeing memory in run-time.

40.7. Usage example

This section demonstrates how to launch testpmd with NVIDIA ConnectX-3 devices managed by librte_net_mlx4.

  1. Load the kernel modules:

    modprobe -a ib_uverbs mlx4_en mlx4_core mlx4_ib
    

    Alternatively if MLNX_OFED is fully installed, the following script can be run:

    /etc/init.d/openibd restart
    

    Note

    User space I/O kernel modules (uio and igb_uio) are not used and do not have to be loaded.

  2. Make sure Ethernet interfaces are in working order and linked to kernel verbs. Related sysfs entries should be present:

    ls -d /sys/class/net/*/device/infiniband_verbs/uverbs* | cut -d / -f 5
    

    Example output:

    eth2
    eth3
    eth4
    eth5
    
  3. Optionally, retrieve their PCI bus addresses to be used with the allow argument:

    {
        for intf in eth2 eth3 eth4 eth5;
        do
            (cd "/sys/class/net/${intf}/device/" && pwd -P);
        done;
    } |
    sed -n 's,.*/\(.*\),-a \1,p'
    

    Example output:

    -a 0000:83:00.0
    -a 0000:83:00.0
    -a 0000:84:00.0
    -a 0000:84:00.0
    

    Note

    There are only two distinct PCI bus addresses because the NVIDIA ConnectX-3 adapters installed on this system are dual port.

  4. Request huge pages:

    dpdk-hugepages.py --setup 2G
    
  5. Start testpmd with basic parameters:

    dpdk-testpmd -l 8-15 -n 4 -a 0000:83:00.0 -a 0000:84:00.0 -- --rxq=2 --txq=2 -i
    

    Example output:

    [...]
    EAL: PCI device 0000:83:00.0 on NUMA socket 1
    EAL:   probe driver: 15b3:1007 librte_net_mlx4
    PMD: librte_net_mlx4: PCI information matches, using device "mlx4_0" (VF: false)
    PMD: librte_net_mlx4: 2 port(s) detected
    PMD: librte_net_mlx4: port 1 MAC address is 00:02:c9:b5:b7:50
    PMD: librte_net_mlx4: port 2 MAC address is 00:02:c9:b5:b7:51
    EAL: PCI device 0000:84:00.0 on NUMA socket 1
    EAL:   probe driver: 15b3:1007 librte_net_mlx4
    PMD: librte_net_mlx4: PCI information matches, using device "mlx4_1" (VF: false)
    PMD: librte_net_mlx4: 2 port(s) detected
    PMD: librte_net_mlx4: port 1 MAC address is 00:02:c9:b5:ba:b0
    PMD: librte_net_mlx4: port 2 MAC address is 00:02:c9:b5:ba:b1
    Interactive-mode selected
    Configuring Port 0 (socket 0)
    PMD: librte_net_mlx4: 0x867d60: TX queues number update: 0 -> 2
    PMD: librte_net_mlx4: 0x867d60: RX queues number update: 0 -> 2
    Port 0: 00:02:C9:B5:B7:50
    Configuring Port 1 (socket 0)
    PMD: librte_net_mlx4: 0x867da0: TX queues number update: 0 -> 2
    PMD: librte_net_mlx4: 0x867da0: RX queues number update: 0 -> 2
    Port 1: 00:02:C9:B5:B7:51
    Configuring Port 2 (socket 0)
    PMD: librte_net_mlx4: 0x867de0: TX queues number update: 0 -> 2
    PMD: librte_net_mlx4: 0x867de0: RX queues number update: 0 -> 2
    Port 2: 00:02:C9:B5:BA:B0
    Configuring Port 3 (socket 0)
    PMD: librte_net_mlx4: 0x867e20: TX queues number update: 0 -> 2
    PMD: librte_net_mlx4: 0x867e20: RX queues number update: 0 -> 2
    Port 3: 00:02:C9:B5:BA:B1
    Checking link statuses...
    Port 0 Link Up - speed 10000 Mbps - full-duplex
    Port 1 Link Up - speed 40000 Mbps - full-duplex
    Port 2 Link Up - speed 10000 Mbps - full-duplex
    Port 3 Link Up - speed 40000 Mbps - full-duplex
    Done
    testpmd>