Linux Kernel Upstream Release Notes v6.17

Changes and New Features

Upstream Kernel 6.17

Internal Ref.

Feature

Description

Support Added in User Space Version

Support Added in Firmware Version

Net Dev

4370178

ETS Rate Limit for VFs Group

[ConnectX-6 Dx and above] Added support for specifying bandwidth proportions between traffic classes (TC) in the devlink-rate API. This option allows users to allocate bandwidth across multiple traffic classes in a single command. This feature provides a more granular control over traffic management, especially for scenarios requiring Enhanced Transmission Selection. Users can now define a specific relative bandwidth share for each traffic class, such as allocating 20% for TC0 (TCP/UDP) and 80% for TC5 (RoCE).

iproute2

12.43.0272

3358561

TCP Zero Copy

[ConnectX-7 and above] Exposed Header-Data Split to the TCP layer to support TCP Zero Copy. This allows headers up to and including TCP to be received by the software, while packet data to be received directly by the GPU via DMA from the NIC, without the CPU's intervention, achieving true zero-copy.

N/A

N/A

RDMA

4372311

Network Namespace

[ConnectX-6 and above] Added support for setting a device's network namespace upon allocation. The name space is inherited from its devlink instance

following a devlink reload command. The IPoIB driver is now namespace-aware, creating its interfaces in the same namespace as the parent RDMA device.

N/A

N/A

4220319

TPH and Steering Tags Support

[ConnectX-8] Added support for TPH and steering tags, which allows the device to provide hints about where an MR TLPs should be processed in system memory.

rdma-core v59

xx.46.1006

4063646

DMABUF Optimization

[ConnectX-4 and above] Added support for o ptimizing the page size used for DMABUF memory keys (mkeys) in the mlx5 driver. The goal is to enhance performance by selecting the most appropriate page size based on the underlying system configuration and workload requirements, which can lead to improved memory management and BW.

N/A

N/A

4032404

Flow Control for Solicited MADs

[All HCAs] Added support for flow control mechanism for non-user solicited MADs to prevent congestion and excessive retries in the MAD QP layer. This limits the number of outstanding solicited MAD requests an agent can issue, holding excess requests in a backlog until the MAD QP is ready to process more.

N/A

N/A

© Copyright 2025, NVIDIA. Last updated on Nov 10, 2025