Changes and New Features

Core

Show SF Device IRQs

[ConnectX-5 and above that support SFs, AFAIR] Added IRQs directory inside SF SYSFS directory. This allows users to know the mapping of SFs and their IRQs.

Configurable SF EQs

[BlueField-2 and above and ConnectX-6 and above] Added support for setting a maximum number of completion EQs for SFs. It allows the user to set the number of queues for an SF.

Netdev

SyncE Userspace Support through Linux Kernel DPLL Subsystem

[All HCAs that support HCA.synce_registers] Added support to use Linux kernel DPLL subsystem as a mechanism for working with clock signals in NVIDIA's proprietary synchronous ethernet protocol daemon. This new mechanism enables the use of VFs and SFs.

Socket-Direct multi-PF Netdev

[ConnectX-6 DX and above with socket direct support] Added support to create single netdev abstraction for multiple socket-direct PFs of the same port that run on multiple NUMA nodes. This single netdev handles traffic locally on each of the NUMA nodes, without upper layer complexities. This improves the applications' locality and performance in a multi-NUMA system.

Support Matching on l4_type for ttc_table

[BlueField3] Added a feature that enables sending CC IFA2.0 prob packets on the same LAG ports as functional packets.

Monitor VFs/SFs

[ConnectX-6 DX and above] Added a feature that exposes the out of buffer counters on the VF/SF to the VF/SF representor, to improve monitoring on the host side.

Support Lightweight EQ

[BlueField-3] Page EQ is not created, and no BOOT/INIT pages are allocated for non-page supplier functions. Optimize RDMA driver initialization by delaying QPs and the corresponding CQs creation. Re-implement write combining test with SQ, and optimize for SF by reading the result from its parent PF.

Support Hardware GRO

[ConnectX-7] Added support for Hardware GRO. Hardware GRO reduces CPU overhead and improves throughput.

Precision Time Measurement Support on ConnectX

[ConnectX-7] Added support for Precision Time Measurement (PTM), that enhances clock synchronization accuracy between the NIC's hardware clock and the system's host clock. PTM resolves inaccuracies caused by asymmetric PCIe request and response times, which affect the PTP. PTM ensures more accurate synchronization, which is important for time sensitive networking.

RDMA

Limit UMAD Receive List

[All HCAs] Added a feature that limits the amount of MAD packets the kernel holds for the UMAD clients. The driver would silently drop newly received packets when the list is full, thus, limiting the effects of a faulty or malicious user overflowing the network with MADs.

Expose req_transport_retries_exceeded Counter

[All HCAs] Added support for exposing req_transport_retries_exceeded counter (per QP counter). It can be seen using the RDMA statistics and it shows the number of times the requester detected transport retries exceeding error.

Support Up to 23 Bits for uar_page_index in create_cq Flow

[ConnectX-7] Added support for allowing users to perform create_cq with a uar_page_index larger than 2^16 as supported by the hardware/firmware.

ODP Memory Scheme

[ConnectX-7 and above] Added support for memory scheme ODP. This allows the usage of ODP on DevX QPs by passing the responsibility of parsing the WQEs and MRs of the page fault to the hardware.