[PATCH] net/mlx5: fix storage handling of shared Rx queues

[PATCH] net/mlx5: fix storage handling of shared Rx queues

[Gregory Etelson]

The MLX5 PMD maintains 2 lists for Rx queues:
- mlx5_priv::rxqsctrl - for non-shared and shared Rx queues
- mlx5_dev_ctx_shared::shared_rxqs - for shared Rx queues only

The PMD used the `rxqsctrl` as the primary list for Rx queues
maintenance.

The PMD wipes out port mlx5_priv object after an application closed
the port.

If PMD shared Rx queues between the transfer proxy port and
representor ports and closed the transfer proxy port before
representor, the representor port cannot iterate its shared Rx queues
because Rx queues list head was wiped out.

The patch separates Rx queue storage list according to the list type:
- shared Rx queues are stored in the `shared_rxqs` only
- non-shared Rx queues are stored in the `rxqsctrl` list only.

Fixes: 6886b5f39d66 ("net/mlx5: fix hairpin queue release")

Signed-off-by: Gregory Etelson <getelson@nvidia.com>
Acked-by: Dariusz Sosnowski <dsosnowski@nvidia.com>
---
 drivers/net/mlx5/linux/mlx5_os.c | 13 ++++++++++---
 drivers/net/mlx5/mlx5_flow.c     |  6 ++++++
 drivers/net/mlx5/mlx5_rxq.c      |  6 ++++--
 3 files changed, 20 insertions(+), 5 deletions(-)

diff --git a/drivers/net/mlx5/linux/mlx5_os.c b/drivers/net/mlx5/linux/mlx5_os.c
index 696a3e12c7..0266f71bb5 100644
--- a/drivers/net/mlx5/linux/mlx5_os.c
+++ b/drivers/net/mlx5/linux/mlx5_os.c
@@ -747,13 +747,20 @@ void
 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 {
 	struct mlx5_dev_ctx_shared *sh = priv->sh;
-#ifdef HAVE_MLX5DV_DR
-	int i;
-#endif
+	struct mlx5_rxq_ctrl *rxq_ctrl;
+	int i = 0;
 
 	MLX5_ASSERT(sh && sh->refcnt);
 	if (sh->refcnt > 1)
 		return;
+	LIST_FOREACH(rxq_ctrl, &sh->shared_rxqs, next) {
+		DRV_LOG(DEBUG, "port %u Rx Queue %u still referenced",
+			priv->dev_data->port_id, rxq_ctrl->rxq.idx);
+		++i;
+	}
+	if (i > 0)
+		DRV_LOG(WARNING, "port %u some Rx queues still remain %d",
+			priv->dev_data->port_id, i);
 	MLX5_ASSERT(LIST_EMPTY(&sh->shared_rxqs));
 #ifdef HAVE_MLX5DV_DR
 	if (sh->rx_domain) {
diff --git a/drivers/net/mlx5/mlx5_flow.c b/drivers/net/mlx5/mlx5_flow.c
index 8db372123c..fa8b95df16 100644
--- a/drivers/net/mlx5/mlx5_flow.c
+++ b/drivers/net/mlx5/mlx5_flow.c
@@ -1652,12 +1652,18 @@ flow_rxq_mark_flag_set(struct rte_eth_dev *dev)
 			LIST_FOREACH(rxq_ctrl, &opriv->rxqsctrl, next) {
 				rxq_ctrl->rxq.mark = 1;
 			}
+			LIST_FOREACH(rxq_ctrl, &opriv->sh->shared_rxqs, next) {
+				rxq_ctrl->rxq.mark = 1;
+			}
 			opriv->mark_enabled = 1;
 		}
 	} else {
 		LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
 			rxq_ctrl->rxq.mark = 1;
 		}
+		LIST_FOREACH(rxq_ctrl, &priv->sh->shared_rxqs, next) {
+			rxq_ctrl->rxq.mark = 1;
+		}
 		priv->mark_enabled = 1;
 	}
 	priv->sh->shared_mark_enabled = 1;
diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index 77c5848c37..1425886a22 100644
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -2033,8 +2033,9 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 		tmpl->share_group = conf->share_group;
 		tmpl->share_qid = conf->share_qid;
 		LIST_INSERT_HEAD(&priv->sh->shared_rxqs, tmpl, share_entry);
+	} else {
+		LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
 	}
-	LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
 	rte_atomic_store_explicit(&tmpl->ctrl_ref, 1, rte_memory_order_relaxed);
 	return tmpl;
 error:
@@ -2365,7 +2366,8 @@ mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
 					(&rxq_ctrl->rxq.mr_ctrl.cache_bh);
 			if (rxq_ctrl->rxq.shared)
 				LIST_REMOVE(rxq_ctrl, share_entry);
-			LIST_REMOVE(rxq_ctrl, next);
+			else
+				LIST_REMOVE(rxq_ctrl, next);
 			mlx5_free(rxq_ctrl->rxq.rq_win_data);
 			mlx5_free(rxq_ctrl);
 		}
-- 
2.48.1

Re: [PATCH] net/mlx5: fix storage handling of shared Rx queues

[Raslan Darawsheh]

Hi,


On 31/07/2025 1:41 PM, Gregory Etelson wrote:
> The MLX5 PMD maintains 2 lists for Rx queues:
> - mlx5_priv::rxqsctrl - for non-shared and shared Rx queues
> - mlx5_dev_ctx_shared::shared_rxqs - for shared Rx queues only
> 
> The PMD used the `rxqsctrl` as the primary list for Rx queues
> maintenance.
> 
> The PMD wipes out port mlx5_priv object after an application closed
> the port.
> 
> If PMD shared Rx queues between the transfer proxy port and
> representor ports and closed the transfer proxy port before
> representor, the representor port cannot iterate its shared Rx queues
> because Rx queues list head was wiped out.
> 
> The patch separates Rx queue storage list according to the list type:
> - shared Rx queues are stored in the `shared_rxqs` only
> - non-shared Rx queues are stored in the `rxqsctrl` list only.
> 
> Fixes: 6886b5f39d66 ("net/mlx5: fix hairpin queue release")
> 
> Signed-off-by: Gregory Etelson <getelson@nvidia.com>
> Acked-by: Dariusz Sosnowski <dsosnowski@nvidia.com>

Patch applied to next-net-mlx,

Kindest regards
Raslan Darawsheh