[PATCH] net/mlx5/hws: fix completion polling for full queues
Bing Zhao
bingz at nvidia.com
Sun Nov 15 19:28:30 CET 2026
Hi,
LGTM
> -----Original Message-----
> From: Maxime Peim <maxime.peim at gmail.com>
> Sent: Wednesday, November 11, 2026 1:18 AM
> To: Dariusz Sosnowski <dsosnowski at nvidia.com>; Slava Ovsiienko
> <viacheslavo at nvidia.com>; Bing Zhao <bingz at nvidia.com>; Ori Kam
> <orika at nvidia.com>; Suanming Mou <suanmingm at nvidia.com>; Matan Azrad
> <matan at nvidia.com>
> Cc: dev at dpdk.org; stable at dpdk.org
> Subject: [PATCH] net/mlx5/hws: fix completion polling for full queues
>
> External email: Use caution opening links or attachments
>
>
> The software completion ring masks its producer and consumer counters when
> advancing them. When all entries are filled before any completion is
> polled, the counters become equal and the ring is treated as empty.
> Root-table rule operations generate completions directly into this ring,
> so a full batch cannot be dequeued and the queue remains busy.
>
> Keep the counters unmasked and apply the ring mask only when accessing the
> completion array. This distinguishes a full ring from an empty one without
> reducing the configured queue capacity.
>
> The power-of-two ring capacity divides the 16-bit counter modulus, so
> counter wrap preserves masked indices. Capacity also bounds outstanding
> completions below that modulus, keeping full distinct from empty.
>
> Fixes: 3eb748869d2d ("net/mlx5/hws: add send layer")
> Cc: stable at dpdk.org
>
> Signed-off-by: Maxime Peim <maxime.peim at gmail.com>
> ---
> drivers/net/mlx5/hws/mlx5dr_send.c | 8 +++++---
> drivers/net/mlx5/hws/mlx5dr_send.h | 13 ++++++++++---
> 2 files changed, 15 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/net/mlx5/hws/mlx5dr_send.c
> b/drivers/net/mlx5/hws/mlx5dr_send.c
> index 85f613ed39..f94957dd27 100644
> --- a/drivers/net/mlx5/hws/mlx5dr_send.c
> +++ b/drivers/net/mlx5/hws/mlx5dr_send.c
> @@ -646,12 +646,14 @@ static void mlx5dr_send_engine_poll_list(struct
> mlx5dr_send_engine *queue,
>
> while (comp->ci != comp->pi) {
> if (*polled < res_nb) {
> + uint16_t idx = comp->ci & comp->mask;
> +
> res[*polled].status =
> - comp->entries[comp->ci].status;
> + comp->entries[idx].status;
> res[*polled].user_data =
> - comp->entries[comp->ci].user_data;
> + comp->entries[idx].user_data;
> (*polled)++;
> - comp->ci = (comp->ci + 1) & comp->mask;
> + comp->ci++;
> mlx5dr_send_engine_dec_rule(queue);
> } else {
> return;
> diff --git a/drivers/net/mlx5/hws/mlx5dr_send.h
> b/drivers/net/mlx5/hws/mlx5dr_send.h
> index 9209754136..ad2488dca2 100644
> --- a/drivers/net/mlx5/hws/mlx5dr_send.h
> +++ b/drivers/net/mlx5/hws/mlx5dr_send.h
> @@ -138,6 +138,12 @@ struct mlx5dr_completed_poll_entry {
>
> struct mlx5dr_completed_poll {
> struct mlx5dr_completed_poll_entry *entries;
> + /*
> + * Keep counters unmasked to distinguish full from empty, and mask
> + * only when indexing entries. The power-of-two capacity divides
> 2^16,
> + * so uint16_t wrap preserves masked indices. Capacity is less
> than
> + * 2^16 and bounds outstanding completions, so equal counters mean
> empty.
> + */
> uint16_t ci;
> uint16_t pi;
> uint16_t mask;
> @@ -279,11 +285,12 @@ static inline void
> mlx5dr_send_engine_gen_comp(struct mlx5dr_send_engine *queue,
> int comp_status) {
> struct mlx5dr_completed_poll *comp = &queue->completed;
> + uint16_t idx = comp->pi & comp->mask;
>
> - comp->entries[comp->pi].status = comp_status;
> - comp->entries[comp->pi].user_data = user_data;
> + comp->entries[idx].status = comp_status;
> + comp->entries[idx].user_data = user_data;
>
> - comp->pi = (comp->pi + 1) & comp->mask;
> + comp->pi++;
> }
>
> static inline bool mlx5dr_send_engine_err(struct mlx5dr_send_engine
> *queue)
Acked-by: Bing Zhao <bingz at nvidia.com>
Thanks
More information about the dev
mailing list