Skip to content

rdma/mlx5_dv: force RoCE IPv4 TOS 96 on RTR regardless of SL - #6

Merged
jleeh merged 1 commit into
rebase/v26.05from
fix/mlx5-dv-force-roce-tos96
Aug 28, 2026
Merged

jleeh merged 1 commit into
rebase/v26.05from
fix/mlx5-dv-force-roce-tos96

Conversation

@jleeh

@jleeh jleeh commented Aug 28, 2026

Copy link
Copy Markdown

Summary

  • Reset.13 inferred traffic_class from ah_attr.sl. mlx5_dv RTR AVs keep sl=0 while rdma_set_option(TOS) only programs the SQ, so the helper was a no-op and switch Q3 stayed frozen.
  • On RTR, if IBV_QP_AV is present, force TOS 96 / SL 3 (CS3 / DSCP 24), matching dataEngineRdmaPriorityClass v2=3. Do not add IBV_QP_AV if the kernel did not already supply a full AV.
  • NOTICELOG the mask / sl / tclass / is_global so a canary IM roll can prove the helper fired.

Test plan

  • Merge, tag v26.05.0+linkpool.5, pin dep-versions, build IM v1.13.0-reset.14
  • Canary-roll ma5-11 (CR-delete v2 IM only). IM logs contain RoCE RTR TOS 96
  • ma5-leaf-1/2 xe-1/1/5 Q3 increments with ma5-11 writes; if not, check uplinks / xe-1/1/7 (EVPN VXLAN)

Made with Cursor

Inferring traffic_class from ah_attr.sl was a no-op: mlx5_dv RTR AVs
keep sl=0 while rdma_set_option only programs the SQ. Stamp TOS 96 /
SL 3 (CS3) whenever IBV_QP_AV is present, and log the AV so a canary
IM roll can prove the helper fired.

Co-authored-by: Cursor <cursoragent@cursor.com>
@jleeh
jleeh merged commit c4a19ec into rebase/v26.05 Aug 28, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant