6.6-stable review patch. If anyone has any objections, please let me know.
------------------
From: Rahul Rameshbabu rrameshbabu@nvidia.com
[ Upstream commit 92214be5979c0961a471b7eaaaeacab41bdf456c ]
Previously, mlx5e_ptp_poll_ts_cq would update the device doorbell with the incremented consumer index after the relevant software counters in the kernel were updated. In the mlx5e_sq_xmit_wqe context, this would lead to either overrunning the device CQ or exceeding the expected software buffer size in the device CQ if the device CQ size was greater than the software buffer size. Update the relevant software counter only after updating the device CQ consumer index in the port timestamping napi_poll context.
Log: mlx5_core 0000:08:00.0: cq_err_event_notifier:517:(pid 0): CQ error on CQN 0x487, syndrome 0x1 mlx5_core 0000:08:00.0 eth2: mlx5e_cq_error_event: cqn=0x000487 event=0x04
Fixes: 1880bc4e4a96 ("net/mlx5e: Add TX port timestamp support") Signed-off-by: Rahul Rameshbabu rrameshbabu@nvidia.com Signed-off-by: Saeed Mahameed saeedm@nvidia.com Link: https://lore.kernel.org/r/20231114215846.5902-12-saeed@kernel.org Signed-off-by: Jakub Kicinski kuba@kernel.org Signed-off-by: Sasha Levin sashal@kernel.org --- .../net/ethernet/mellanox/mlx5/core/en/ptp.c | 20 +++++++++++++++---- 1 file changed, 16 insertions(+), 4 deletions(-)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/ptp.c b/drivers/net/ethernet/mellanox/mlx5/core/en/ptp.c index bb11e644d24f7..af3928eddafd1 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/en/ptp.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/en/ptp.c @@ -177,6 +177,8 @@ static void mlx5e_ptpsq_mark_ts_cqes_undelivered(struct mlx5e_ptpsq *ptpsq,
static void mlx5e_ptp_handle_ts_cqe(struct mlx5e_ptpsq *ptpsq, struct mlx5_cqe64 *cqe, + u8 *md_buff, + u8 *md_buff_sz, int budget) { struct mlx5e_ptp_port_ts_cqe_list *pending_cqe_list = ptpsq->ts_cqe_pending_list; @@ -211,19 +213,24 @@ static void mlx5e_ptp_handle_ts_cqe(struct mlx5e_ptpsq *ptpsq, mlx5e_ptpsq_mark_ts_cqes_undelivered(ptpsq, hwtstamp); out: napi_consume_skb(skb, budget); - mlx5e_ptp_metadata_fifo_push(&ptpsq->metadata_freelist, metadata_id); + md_buff[*md_buff_sz++] = metadata_id; if (unlikely(mlx5e_ptp_metadata_map_unhealthy(&ptpsq->metadata_map)) && !test_and_set_bit(MLX5E_SQ_STATE_RECOVERING, &sq->state)) queue_work(ptpsq->txqsq.priv->wq, &ptpsq->report_unhealthy_work); }
-static bool mlx5e_ptp_poll_ts_cq(struct mlx5e_cq *cq, int budget) +static bool mlx5e_ptp_poll_ts_cq(struct mlx5e_cq *cq, int napi_budget) { struct mlx5e_ptpsq *ptpsq = container_of(cq, struct mlx5e_ptpsq, ts_cq); - struct mlx5_cqwq *cqwq = &cq->wq; + int budget = min(napi_budget, MLX5E_TX_CQ_POLL_BUDGET); + u8 metadata_buff[MLX5E_TX_CQ_POLL_BUDGET]; + u8 metadata_buff_sz = 0; + struct mlx5_cqwq *cqwq; struct mlx5_cqe64 *cqe; int work_done = 0;
+ cqwq = &cq->wq; + if (unlikely(!test_bit(MLX5E_SQ_STATE_ENABLED, &ptpsq->txqsq.state))) return false;
@@ -234,7 +241,8 @@ static bool mlx5e_ptp_poll_ts_cq(struct mlx5e_cq *cq, int budget) do { mlx5_cqwq_pop(cqwq);
- mlx5e_ptp_handle_ts_cqe(ptpsq, cqe, budget); + mlx5e_ptp_handle_ts_cqe(ptpsq, cqe, + metadata_buff, &metadata_buff_sz, napi_budget); } while ((++work_done < budget) && (cqe = mlx5_cqwq_get_cqe(cqwq)));
mlx5_cqwq_update_db_record(cqwq); @@ -242,6 +250,10 @@ static bool mlx5e_ptp_poll_ts_cq(struct mlx5e_cq *cq, int budget) /* ensure cq space is freed before enabling more cqes */ wmb();
+ while (metadata_buff_sz > 0) + mlx5e_ptp_metadata_fifo_push(&ptpsq->metadata_freelist, + metadata_buff[--metadata_buff_sz]); + mlx5e_txqsq_wake(&ptpsq->txqsq);
return work_done == budget;