Changes and New Features

Upstream Kernel 6.5

Internal Ref.

Feature

Description

Support Added in User Space Version

Support Added in Firmware Version

ASAP

3376168

Bridge Debuggability Extensions: Expose FDB Data via debugfs

[ConnectX-6 Dx and above] This new capability improves the bridge's offload debuggability.

N/A

N/A

Core

3427627

Full Chip Reset on BlueField-2 and BlueField-3 DPUs

[BlueField-2 and BlueField-3] Added support for full chip reset in DPU mode on BlueField-2 and BlueField-3 DPUs, using mlxfwreset the "--sync 1" option.

During this flow, Arm is going through reboot and firmware is reloaded.

N/A

xx.38.1002

3405789

Embedded CPU Virtual Functions

[BlueField-2 and BlueField-3] Enabled the creation of Virtual Functions within the Arm.

Note: This capability requires setting the following parameters:

  • In mlxconfig: PF_NUM_OF_VF_VALID=True

  • For each Arm and Host PF: PF_NUM_OF_VF

N/A

xx.36.1010

3307360

QEMU VFIO Migration pre-copy

[ConnectX-6 Dx and above] VFIO migration pre-copy support extends basic VFIO migration by allowing the device state's migration while the VM and the device are running.

This new capability reduces migration downtime, especially for devices that use a lot of resources.

QEMU 8.1

xx.37.1014

30011500

Light Weight Local SFs

[ConnectX-5 and above] Probing local SFs with devlink instance only. The local SFs-SFs spawns over the device (PF/ECPF) which is the eSwitch manager.

This new capability decreases the amount of time needed to probe and configure SFs by saving the time required for the devlink to reload the SF.

N/A

N/A

2831943

4 Ports VF LAG

[ConnectX-7] Enabled VF LAG over 4 ports HCAs.

Note: This capability is supported ONLY in LAGs that included all HCA ports, e.g.: with 4 port HCA, only 4 port LAG is supported. 2 ports or 3 ports LAG is not supported.

N/A

xx.36.1010

NetDev

3320236

Indication for Packet Drops due to Severe Steering Errors

[All HCAs] This new capability exposes the generated_pkt_steering_fail and handled_pkt_steering_fail counters through the devlink health reporter to provide the user an indication on any severe Steering errors

N/A

N/A

RDMA

3376634

RDMA, Static Rate

This new capability reduces the time consumption of “resolve route” during a large number of rdma_cm connections establishment by setting the rdma_cm RoCE static rate to 0.

N/A

N/A

3355352

QKEY Mitigation in Kernel

[All HCAs] Non-privileged users are now blocked by default from setting controlled/privileged QKEYs (QKEY with MSB set).

N/A

xx.38.1002

3113659

Expose DC MRA Caps

[All HCAs] The current system behavior reports only the value for the RC QPs, thus the users running DC need to query this value and negotiate server-client with this.

Up until now, the value for both RC and DC was set to 16 which might have resulted in performance issues. To avoid performance issues, we changed the max outstanding DC atomic reads to 32, and have different RC and DC values.

N/A

N/A

© Copyright 2023, NVIDIA. Last updated on Nov 20, 2023.