Re: [PATCH net] net/mlx5e: Fix stalled channels with threaded NAPI

[Josef Bacik <josef@toxicpanda.com>]

On Thu, Oct 29, 2026 at 02:52:06PM +0200, Tariq Toukan wrote:
> We confirmed it on the device, on ConnectX-6 Dx (the generation in your
> report) and on ConnectX-7: an arm raises an immediate event only when
> the CQ holds a CQE newer than the last reported event. A backlog the
> last event already reported raises nothing.

Thanks for checking that.

> Please hold your v2 until these are out, and rebase on top: your patch
> then stands as what it is, saving a complete-and-rearm cycle per budget
> with threaded NAPI.

Will do.  Please Cc me on them.  On top of your fix it is no longer a
fix, so v2 will drop the Fixes tag and go to net-next.

What changes in v2 besides that: the test moves into
mlx5e_channel_no_affinity_change(), as Jakub and Bjorn asked, and it
tests NAPI_STATE_SCHED_THREADED instead of NAPI_STATE_THREADED, so that
it follows who is polling and not the mode that was asked for.

> One request for the code comment when you reword it: today it reads as
> if arming never raises an event for pending CQEs, "only for a new
> completion".

The comment in v2 only says why the CPU of the NAPI thread tells
nothing about the interrupt's affinity, it no longer says anything
about arming.  The changelog will use your wording.

pw-bot: cr

Thanks,
Josef