(Updated) Linux TC HW Offload not working on ConnectX-5/6

Goal : Mirror Incoming/Outgoing Packets Between Two Ethernet Ports in Hardware

Env : Fresh, unmodified install of Unbuntu 22.04 on bare metal Intel x86_64
Dual port ConnectX-5 in PCI slots 0000:01.00/0000:01.01
Two ethernet links ‘enp1s0f0’, ‘enp1s0f1’

Basic Solution: Run these Linux Commands:

0 # enable HW offload
1 ethtool -K enp1s0f0 hw-tc-offload on
2 ethtool -K enp1s0f1 hw-tc-offload on
3 # ingress mirror in HW
4 tc qdisc add dev enp1s0f1 ingress
5 tc filter add dev enp1s0f1 ingress flower verbose skip_sw action mirred egress mirror dev enp1s0f0
7 # egress mirror in HW
8 tc qdisc add dev enp1s0f1 root prio
9 tc filter add dev enp1s0f1 root flower verbose skip_sw action mirred egress mirror dev enp1s0f0

The filter commands give error:
Error: mlx5_core: Not implemented offload action. There is no additional detail in dmesg on Ubuntu.

These filter variations DO WORK:

  • In line 5, 9 replace keyword “mirror” with “redirect” i.e. HW redirect works
  • In line 5, 9 remove keyword “skip_sw” i.e. mirror works in kernel/software

This is a very basic problem:

  • A successful install of MLNX_OFED_LINUX-23.10-
    with or without updating the ConnectX-5 firmware give the same error
  • The same errors occurs on stock RHEL8. In the RHEL dmesg cryptically mentions
    class(ifier) error.
  • I’ve seen this issue on Connect6-X NICs too
  • I’ve seen the same errors on Ubuntu 24.xx

This question is a repeat of,

which was never answered right. Variations can be found across the WWW too.

This error comes from omitting act->id=5 (MIRRER) in the MLX5_CORE module but supporting REDIRECT (v5.15.136 is the kernel version for Ubuntu 22.04LTS):


This means the MLX5_CORE driver does not support the TC fragment “action mirred (ingress|egress) mirror”.

How can packets be mirrored then between ethernet links?

Hello @shane.miller,

Thank you for posting your query on our community. in order to debug this issue further, we would request you to submit a support ticket along with a sysinfo snapshot captured from the node for further troubleshooting. The support ticket can be opened by emailing " Networking-support@nvidia.com ".


NVIDIA support has been no help.

However, I believe I was able to figure out how to do this. When testing is done I will post a link here to the solution, which is how NVIDIA should provide technical documentation.