Changes and New Features

Core

ConnectX-7 Multihost+LAG Support for Tencent New Server Platform

[ConnectX-7] Added a feature that enables each host to get partial PFs and to own its unique LAG under multi-host deployment. All the lags are independent.

Trust Lockdown Health Syndrome

[ConnectX-8] Added a new health syndrome for trust lockdown and exposed crr bit in the health buffer. When crr bit is set, the syndrome requires cold reset to recover.

User-Space Firmware Access via FWctl

[ConnectX-5 and onward] Added support for FWctl, which is a new subsystem designed to establish a common framework for securely exposing firmware interfaces to user space. It aims to provide consistent access for debugging, configuration, and provisioning of devices. This also adds support for mlx5, enabling user-space access to debug and configuration registers of the ConnectX hardware family, including NICs, PCI switches, and SmartNIC SoCs.

Netdev

Group of Groups

[ConnectX-6 Dx and above] Added support for a group of groups, which enables multi-tier resource control, where parent groups enforce aggregate limits while child groups dynamically share bandwidth.

This structure is crucial for QoS enforcement, burst management, and fair bandwidth distribution, ensuring that workloads receive adequate resources.

Support PTP Device

[ConnectX-6 Dx] Added support for one PTP device per hardware clock. The PTP device represents the shared PHC when it is running in real time mode.

Lanes Number

[All HCAs] Added support for specifying the number of lanes (protocol width) when auto-negotiation is off. As the combination of speed and number of lanes corresponds to one link mode, this enables the user to select exactly one link node when auto negotiation is off. This feature follows a firmware update that expects the driver to specify a single link mode when auto-negotiation is off, rather than a bitmap corresponding to the user-defined speed.

RDMA

Devx Qps Parallel Cleanup

[All HCAs] Added a feature that considerably reduces cleanup time of userspace applications that use DevX QPs without changing anything on the user end. When using a Kernel that includes this patch automatically, DevX QPs destruction upon cleanup will be sent to the firmware in parallel instead of in a serial method, which results in a faster cleanup time. The more DevX QPs the application has the more noticeable it would be.

Log IB State Transitions

[All HCAs] Added a feature that enhances the visibility for IB device state transitions by adding log messages to the Kernel log (dmesg). Every time an IB device changes state, a relevant print will be printed.

Example: "mlx5_0: Port: 1 Link DOWN"

Optional Counters Binding Support

[All HCAs] Added support for a new optional filter named optional-counter to the iproute2 command: rdma stat qp set link [link_name] auto. It indicates that when binding counters to a QP we also want the currently enabled optional-counters on the link to be bound as well. In addition manual QP binding in iproute2 now also binds the currently configured optional-counters for the said counter.

This feature allows tracking optional counters on a per (QP)/(QP group) granularity.