.. SPDX-License-Identifier: GPL-2.0

==================== mlx5 devlink support

This document describes the devlink features implemented by the mlx5 device driver.

Parameters

.. list-table:: Generic parameters implemented

• • Name
  • Mode
  • Validation
  • Notes
• • enable_roce
  • driverinit
  • Boolean
  • If the device supports RoCE disablement, RoCE enablement state controls device support for RoCE capability. Otherwise, the control occurs in the driver stack. When RoCE is disabled at the driver level, only raw ethernet QPs are supported.
• • io_eq_size
  • driverinit
  • The range is between 64 and 4096.
• • event_eq_size
  • driverinit
  • The range is between 64 and 4096.
• • max_macs
  • driverinit
  • The range is between 1 and 2^31. Only power of 2 values are supported.
• • enable_sriov
  • permanent
  • Boolean
  • Applies to each physical function (PF) independently, if the device supports it. Otherwise, it applies symmetrically to all PFs.
• • total_vfs
  • permanent
  • The range is between 1 and a device-specific max.
  • Applies to each physical function (PF) independently, if the device supports it. Otherwise, it applies symmetrically to all PFs.

Note: permanent parameters such as enable_sriov and total_vfs require FW reset to take effect

.. code-block:: bash

setup parameters

devlink dev param set pci/0000:01:00.0 name enable_sriov value true cmode permanent devlink dev param set pci/0000:01:00.0 name total_vfs value 8 cmode permanent

Fw reset

devlink dev reload pci/0000:01:00.0 action fw_activate

for PCI related config such as sriov PCI reset/rescan is required:

echo 1 >/sys/bus/pci/devices/0000:01:00.0/remove echo 1 >/sys/bus/pci/rescan grep ^ /sys/bus/pci/devices/0000:01:00.0/sriov_*

• • num_doorbells

  • driverinit

  • This controls the number of channel doorbells used by the netdev. In all cases, an additional doorbell is allocated and used for non-channel communication (e.g. for PTP, HWS, etc.). Supported values are:

    • 0: No channel-specific doorbells, use the global one for everything.
    • [1, max_num_channels]: Spread netdev channels equally across these doorbells.

The mlx5 driver also implements the following driver-specific parameters.

.. list-table:: Driver-specific parameters implemented :widths: 5 5 5 85

• • Name
  • Type
  • Mode
  • Description

• • flow_steering_mode

  • string

  • runtime

  • Controls the flow steering mode of the driver

    • dmfs Device managed flow steering. In DMFS mode, the HW steering entities are created and managed through firmware.
    • smfs Software managed flow steering. In SMFS mode, the HW steering entities are created and manage through the driver without firmware intervention.
    • hmfs Hardware managed flow steering. In HMFS mode, the driver is configuring steering rules directly to the HW using Work Queues with a special new type of WQE (Work Queue Element).

    SMFS mode is faster and provides better rule insertion rate compared to default DMFS mode.

• • fdb_large_groups

  • u32

  • driverinit

  • Control the number of large groups (size > 1) in the FDB table.

    • The default value is 15, and the range is between 1 and 1024.

• • esw_multiport

  • Boolean

  • runtime

  • Control MultiPort E-Switch shared fdb mode.

    An experimental mode where a single E-Switch is used and all the vports and physical ports on the NIC are connected to it.

    An example is to send traffic from a VF that is created on PF0 to an uplink that is natively associated with the uplink of PF1

    Note: Future devices, ConnectX-8 and onward, will eventually have this as the default to allow forwarding between all NIC ports in a single E-switch environment and the dual E-switch mode will likely get deprecated.

    Default: disabled

• • esw_port_metadata

  • Boolean

  • runtime

  • When applicable, disabling eswitch metadata can increase packet rate up to 20% depending on the use case and packet sizes.

    Eswitch port metadata state controls whether to internally tag packets with metadata. Metadata tagging must be enabled for multi-port RoCE, failover between representors and stacked devices. By default metadata is enabled on the supported devices in E-switch. Metadata is applicable only for E-switch in switchdev mode and users may disable it when NONE of the below use cases will be in use:

    1. HCA is in Dual/multi-port RoCE mode.
    2. VF/SF representor bonding (Usually used for Live migration)
    3. Stacked devices

    When metadata is disabled, the above use cases will fail to initialize if users try to enable them.

    Note: Setting this parameter does not take effect immediately. Setting must happen in legacy mode and eswitch port metadata takes effect after enabling switchdev mode.

• • hairpin_num_queues

  • u32

  • driverinit

  • We refer to a TC NIC rule that involves forwarding as "hairpin". Hairpin queues are mlx5 hardware specific implementation for hardware forwarding of such packets.

    Control the number of hairpin queues.

• • hairpin_queue_size
  • u32
  • driverinit
  • Control the size (in packets) of the hairpin queues.

• • pcie_cong_inbound_high

  • u16

  • driverinit

  • High threshold configuration for PCIe congestion events. The firmware will send an event once device side inbound PCIe traffic went above the configured high threshold for a long enough period (at least 200ms).

    See pci_bw_inbound_high ethtool stat.

    Units are 0.01 %. Accepted values are in range [0, 10000]. pcie_cong_inbound_low < pcie_cong_inbound_high. Default value: 9000 (Corresponds to 90%).

• • pcie_cong_inbound_low

  • u16

  • driverinit

  • Low threshold configuration for PCIe congestion events. The firmware will send an event once device side inbound PCIe traffic went below the configured low threshold, only after having been previously in a congested state.

    See pci_bw_inbound_low ethtool stat.

    Units are 0.01 %. Accepted values are in range [0, 10000]. pcie_cong_inbound_low < pcie_cong_inbound_high. Default value: 7500.

• • pcie_cong_outbound_high

  • u16

  • driverinit

  • High threshold configuration for PCIe congestion events. The firmware will send an event once device side outbound PCIe traffic went above the configured high threshold for a long enough period (at least 200ms).

    See pci_bw_outbound_high ethtool stat.

    Units are 0.01 %. Accepted values are in range [0, 10000]. pcie_cong_outbound_low < pcie_cong_outbound_high. Default value: 9000 (Corresponds to 90%).

• • pcie_cong_outbound_low

  • u16

  • driverinit

  • Low threshold configuration for PCIe congestion events. The firmware will send an event once device side outbound PCIe traffic went below the configured low threshold, only after having been previously in a congested state.

    See pci_bw_outbound_low ethtool stat.

    Units are 0.01 %. Accepted values are in range [0, 10000]. pcie_cong_outbound_low < pcie_cong_outbound_high. Default value: 7500.

• • cqe_compress_type
  • string
  • permanent
  • Configure which mechanism/algorithm should be used by the NIC that will affect the rate (aggressiveness) of compressed CQEs depending on PCIe bus conditions and other internal NIC factors. This mode affects all queues that enable compression.
    • balanced : Merges fewer CQEs, resulting in a moderate compression ratio but maintaining a balance between bandwidth savings and performance
    • aggressive : Merges more CQEs into a single entry, achieving a higher compression rate and maximizing performance, particularly under high traffic loads

• • swp_l4_csum_mode

  • string

  • permanent

  • Configure how the L4 checksum is calculated by the device when using Software Parser (SWP) hints for header locations.

    • default : Use the device's default checksum calculation mode. The driver will discover during init whether or full_csum or l4_only is in use. Setting this value explicitly from userspace is not allowed, but some firmware versions may return this value on param read.
    • full_csum : Calculate full checksum including the pseudo-header
    • l4_only : Calculate L4-only checksum, excluding the pseudo-header

The mlx5 driver supports reloading via DEVLINK_CMD_RELOAD

Info versions

The mlx5 driver reports the following versions

.. list-table:: devlink info versions implemented :widths: 5 5 90

• • Name
  • Type
  • Description
• • fw.psid
  • fixed
  • Used to represent the board id of the device.
• • fw.version
  • stored, running
  • Three digit major.minor.subminor firmware version number.

Health reporters

tx reporter

The tx reporter is responsible for reporting and recovering of the following three error scenarios:

• tx timeout Report on kernel tx timeout detection. Recover by searching lost interrupts.
• tx error completion Report on error tx completion. Recover by flushing the tx queue and reset it.
• tx PTP port timestamping CQ unhealthy Report too many CQEs never delivered on port ts CQ. Recover by flushing and re-creating all PTP channels.

tx reporter also support on demand diagnose callback, on which it provides real time information of its send queues status.

User commands examples:

• Diagnose send queues status::

  $ devlink health diagnose pci/0000:82:00.0 reporter tx

.. note:: This command has valid output only when interface is up, otherwise the command has empty output.

• Show number of tx errors indicated, number of recover flows ended successfully, is autorecover enabled and graceful period from last recover::

  $ devlink health show pci/0000:82:00.0 reporter tx

rx reporter

The rx reporter is responsible for reporting and recovering of the following two error scenarios:

• rx queues' initialization (population) timeout Population of rx queues' descriptors on ring initialization is done in napi context via triggering an irq. In case of a failure to get the minimum amount of descriptors, a timeout would occur, and descriptors could be recovered by polling the EQ (Event Queue).
• rx completions with errors (reported by HW on interrupt context) Report on rx completion error. Recover (if needed) by flushing the related queue and reset it.

rx reporter also supports on demand diagnose callback, on which it provides real time information of its receive queues' status.

• Diagnose rx queues' status and corresponding completion queue::

  $ devlink health diagnose pci/0000:82:00.0 reporter rx

.. note:: This command has valid output only when interface is up. Otherwise, the command has empty output.

• Show number of rx errors indicated, number of recover flows ended successfully, is autorecover enabled, and graceful period from last recover::

  $ devlink health show pci/0000:82:00.0 reporter rx

fw reporter

The fw reporter implements diagnose and dump callbacks. It follows symptoms of fw error such as fw syndrome by triggering fw core dump and storing it into the dump buffer. The fw reporter diagnose command can be triggered any time by the user to check current fw status.

User commands examples:

• Check fw heath status::

  $ devlink health diagnose pci/0000:82:00.0 reporter fw

• Read FW core dump if already stored or trigger new one::

  $ devlink health dump show pci/0000:82:00.0 reporter fw

.. note:: This command can run only on the PF which has fw tracer ownership, running it on other PF or any VF will return "Operation not permitted".

fw fatal reporter

The fw fatal reporter implements dump and recover callbacks. It follows fatal errors indications by CR-space dump and recover flow. The CR-space dump uses vsc interface which is valid even if the FW command interface is not functional, which is the case in most FW fatal errors. The recover function runs recover flow which reloads the driver and triggers fw reset if needed. On firmware error, the health buffer is dumped into the dmesg. The log level is derived from the error's severity (given in health buffer).

User commands examples:

• Run fw recover flow manually::

  $ devlink health recover pci/0000:82:00.0 reporter fw_fatal

• Read FW CR-space dump if already stored or trigger new one::

  $ devlink health dump show pci/0000:82:00.1 reporter fw_fatal

.. note:: This command can run only on PF.

vnic reporter

The vnic reporter implements only the diagnose callback. It is responsible for querying the vnic diagnostic counters from fw and displaying them in realtime.

Description of the vnic counters:

• total_error_queues number of queues in an error state due to an async error or errored command.
• send_queue_priority_update_flow number of QP/SQ priority/SL update events.
• cq_overrun number of times CQ entered an error state due to an overflow.
• async_eq_overrun number of times an EQ mapped to async events was overrun.
• comp_eq_overrun number of times an EQ mapped to completion events was overrun.
• quota_exceeded_command number of commands issued and failed due to quota exceeded.
• invalid_command number of commands issued and failed dues to any reason other than quota exceeded.
• nic_receive_steering_discard number of packets that completed RX flow steering but were discarded due to a mismatch in flow table.
• generated_pkt_steering_fail number of packets generated by the VNIC experiencing unexpected steering failure (at any point in steering flow).
• handled_pkt_steering_fail number of packets handled by the VNIC experiencing unexpected steering failure (at any point in steering flow owned by the VNIC, including the FDB for the eswitch owner).
• icm_consumption amount of Interconnect Host Memory (ICM) consumed by the vnic in granularity of 4KB. ICM is host memory allocated by SW upon HCA request and is used for storing data structures that control HCA operation.
• bar_uar_access number of WRITE or READ access operations to the UAR on the PCIe BAR.
• odp_local_triggered_page_fault number of locally-triggered page-faults due to ODP.
• odp_remote_triggered_page_fault number of remotly-triggered page-faults due to ODP.

User commands examples:

• Diagnose PF/VF vnic counters::

  Copy

    $ devlink health diagnose pci/0000:82:00.1 reporter vnic
  

• Diagnose representor vnic counters (performed by supplying devlink port of the representor, which can be obtained via devlink port command)::

  Copy

    $ devlink health diagnose pci/0000:82:00.1/65537 reporter vnic
  

.. note:: This command can run over all interfaces such as PF/VF and representor ports.