net/mlx5e: Ring the TX doorbell on DMA errors

Bugzilla: https://bugzilla.redhat.com/2112940
Upstream-status: v5.19-rc7

commit 5b759bf
Author: Maxim Mikityanskiy <maximmi@nvidia.com>
Date:   Mon May 30 14:01:37 2022 +0300

    net/mlx5e: Ring the TX doorbell on DMA errors

    TX doorbells may be postponed, because sometimes the driver knows that
    another packet follows (for example, when xmit_more is true, or when a
    MPWQE session is closed before transmitting a packet).

    However, the DMA mapping may fail for the next packet, in which case a
    new WQE is not posted, the doorbell isn't updated either, and the
    transmission of the previous packet will be delayed indefinitely.

    This commit fixes the described rare error flow by posting a NOP and
    ringing the doorbell on errors to flush all the previous packets. The
    MPWQE session is closed before that. DMA mapping in the MPWQE flow is
    moved to the beginning of mlx5e_sq_xmit_mpwqe, because empty sessions
    are not allowed. Stop room always has enough space for a NOP, because
    the actual TX WQE is not posted.

    Fixes: e586b3b ("net/mlx5: Ethernet Datapath files")
    Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
    Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
    Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>

Signed-off-by: Mohammad Kabat <mkabat@redhat.com>

Author: Mohammad Kabat

drivers/net/ethernet/mellanox/mlx5/core/en_tx.c

@@ -318,6 +318,26 @@ static void mlx5e_tx_check_stop(struct mlx5e_txqsq *sq)
 	}
 }
 
+static void mlx5e_tx_flush(struct mlx5e_txqsq *sq)
+{
+	struct mlx5e_tx_wqe_info *wi;
+	struct mlx5e_tx_wqe *wqe;
+	u16 pi;
+
+	/* Must not be called when a MPWQE session is active but empty. */
+	mlx5e_tx_mpwqe_ensure_complete(sq);
+
+	pi = mlx5_wq_cyc_ctr2ix(&sq->wq, sq->pc);
+	wi = &sq->db.wqe_info[pi];
+
+	*wi = (struct mlx5e_tx_wqe_info) {
+		.num_wqebbs = 1,
+	};
+
+	wqe = mlx5e_post_nop(&sq->wq, sq->sqn, &sq->pc);
+	mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, &wqe->ctrl);
+}
+
 static inline void
 mlx5e_txwqe_complete(struct mlx5e_txqsq *sq, struct sk_buff *skb,
 		     const struct mlx5e_tx_attr *attr,
@@ -410,6 +430,7 @@ mlx5e_sq_xmit_wqe(struct mlx5e_txqsq *sq, struct sk_buff *skb,
 err_drop:
 	stats->dropped++;
 	dev_kfree_skb_any(skb);
+	mlx5e_tx_flush(sq);
 }
 
 static bool mlx5e_tx_skb_supports_mpwqe(struct sk_buff *skb, struct mlx5e_tx_attr *attr)
@@ -511,6 +532,13 @@ mlx5e_sq_xmit_mpwqe(struct mlx5e_txqsq *sq, struct sk_buff *skb,
 	struct mlx5_wqe_ctrl_seg *cseg;
 	struct mlx5e_xmit_data txd;
 
+	txd.data = skb->data;
+	txd.len = skb->len;
+
+	txd.dma_addr = dma_map_single(sq->pdev, txd.data, txd.len, DMA_TO_DEVICE);
+	if (unlikely(dma_mapping_error(sq->pdev, txd.dma_addr)))
+		goto err_unmap;
+
 	if (!mlx5e_tx_mpwqe_session_is_active(sq)) {
 		mlx5e_tx_mpwqe_session_start(sq, eseg);
 	} else if (!mlx5e_tx_mpwqe_same_eseg(sq, eseg)) {
@@ -520,18 +548,9 @@ mlx5e_sq_xmit_mpwqe(struct mlx5e_txqsq *sq, struct sk_buff *skb,
 
 	sq->stats->xmit_more += xmit_more;
 
-	txd.data = skb->data;
-	txd.len = skb->len;
-
-	txd.dma_addr = dma_map_single(sq->pdev, txd.data, txd.len, DMA_TO_DEVICE);
-	if (unlikely(dma_mapping_error(sq->pdev, txd.dma_addr)))
-		goto err_unmap;
 	mlx5e_dma_push(sq, txd.dma_addr, txd.len, MLX5E_DMA_MAP_SINGLE);
-
 	mlx5e_skb_fifo_push(&sq->db.skb_fifo, skb);
-
 	mlx5e_tx_mpwqe_add_dseg(sq, &txd);
-
 	mlx5e_tx_skb_update_hwts_flags(skb);
 
 	if (unlikely(mlx5e_tx_mpwqe_is_full(&sq->mpwqe, sq->max_sq_mpw_wqebbs))) {
@@ -553,6 +572,7 @@ mlx5e_sq_xmit_mpwqe(struct mlx5e_txqsq *sq, struct sk_buff *skb,
 	mlx5e_dma_unmap_wqe_err(sq, 1);
 	sq->stats->dropped++;
 	dev_kfree_skb_any(skb);
+	mlx5e_tx_flush(sq);
 }
 
 void mlx5e_tx_mpwqe_ensure_complete(struct mlx5e_txqsq *sq)
@@ -935,5 +955,6 @@ void mlx5i_sq_xmit(struct mlx5e_txqsq *sq, struct sk_buff *skb,
 err_drop:
 	stats->dropped++;
 	dev_kfree_skb_any(skb);
+	mlx5e_tx_flush(sq);
 }
 #endif