On 10/11/2026 0:15, Rishikesh Jethwani wrote:
This write relies on an earlier may_move check in mlx5e_select_queue(),
without locking the old queue or rechecking wmem.
Right, I assumed the socket lock serializes selection, which doesn't
hold with an upper device's qdisc in between.
Selection needs the same initial-claim CAS and locked handoff: recheck
the owner and current wmem under the old queue's lock, and leave
ownership unchanged on failure. Locking only the write leaves the wmem
check stale. Preserve zero self-charge for non-decrypted, possibly
partially orphaned skbs.
A failed transfer can fall back to the owner's queue, updating the
cache, if the index is in [0, real_num_tx_queues), still maps to a live
SQ and is not a PTP index. Keep the xmit ownership guard.
Agreed, v2 does that. Selection and xmit move the connection through
one helper, as you describe, and a failed move falls back to the
owner's TX queue under those conditions. The xmit guard stays.
Thanks,
Tariq