net/mlx5e: XDP, Inline small packets into the TX MPWQE in XDP xmit flow

Upon high packet rate with multiple CPUs TX workloads, much of the HCA's
resources are spent on prefetching TX descriptors, thus affecting
transmission rates.
This patch comes to mitigate this problem by moving some workload to the
CPU and reducing the HW data prefetch overhead for small packets (<= 256B).

When forwarding packets with XDP, a packet that is smaller
than a certain size (set to ~256 bytes) would be sent inline within
its WQE TX descrptor (mem-copied), when the hardware tx queue is congested
beyond a pre-defined water-mark.

This is added to better utilize the HW resources (which now makes
one less packet data prefetch) and allow better scalability, on the
account of CPU usage (which now 'memcpy's the packet into the WQE).

To load balance between HW and CPU and get max packet rate, we use
watermarks to detect how much the HW is congested and move the work
loads back and forth between HW and CPU.

Performance:
Tested packet rate for UDP 64Byte multi-stream
over two dual port ConnectX-5 100Gbps NICs.
CPU: Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

* Tested with hyper-threading disabled

XDP_TX:

|          | before | after   |       |
| 24 rings | 51Mpps | 116Mpps | +126% |
| 1 ring   | 12Mpps | 12Mpps  | same  |

XDP_REDIRECT:

** Below is the transmit rate, not the redirection rate
which might be larger, and is not affected by this patch.

|          | before  | after   |      |
| 32 rings | 64Mpps  | 92Mpps  | +43% |
| 1 ring   | 6.4Mpps | 6.4Mpps | same |

As we can see, feature significantly improves scaling, without
hurting single ring performance.

Signed-off-by: Shay Agroskin <shayag@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>

1 files changed, 13 insertions, 9 deletions

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c
index a9075b526ab9..c3d4efbf60ba 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c
@@ -113,7 +113,9 @@ static void mlx5e_xdp_mpwqe_session_start(struct mlx5e_xdpsq *sq)
 	mlx5e_xdpsq_fetch_wqe(sq, &session->wqe);
 
 	prefetchw(session->wqe->data);
-	session->ds_count = MLX5E_XDP_TX_EMPTY_DS_COUNT;
+	session->ds_count  = MLX5E_XDP_TX_EMPTY_DS_COUNT;
+	session->pkt_count = 0;
+	session->complete  = 0;
 
 	pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
 
@@ -132,6 +134,9 @@ static void mlx5e_xdp_mpwqe_session_start(struct mlx5e_xdpsq *sq)
 		       MLX5E_XDP_MPW_MAX_WQEBBS);
 
 	session->max_ds_count = MLX5_SEND_WQEBB_NUM_DS * wqebbs;
+
+	mlx5e_xdp_update_inline_state(sq);
+
 	stats->mpwqe++;
 }
 
@@ -149,7 +154,7 @@ static void mlx5e_xdp_mpwqe_complete(struct mlx5e_xdpsq *sq)
 	cseg->qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_count);
 
 	wi->num_wqebbs = DIV_ROUND_UP(ds_count, MLX5_SEND_WQEBB_NUM_DS);
-	wi->num_ds     = ds_count - MLX5E_XDP_TX_EMPTY_DS_COUNT;
+	wi->num_pkts   = session->pkt_count;
 
 	sq->pc += wi->num_wqebbs;
 
@@ -164,11 +169,9 @@ static bool mlx5e_xmit_xdp_frame_mpwqe(struct mlx5e_xdpsq *sq,
 	struct mlx5e_xdp_mpwqe *session = &sq->mpwqe;
 	struct mlx5e_xdpsq_stats *stats = sq->stats;
 
-	dma_addr_t dma_addr    = xdpi->dma_addr;
 	struct xdp_frame *xdpf = xdpi->xdpf;
-	unsigned int dma_len   = xdpf->len;
 
-	if (unlikely(sq->hw_mtu < dma_len)) {
+	if (unlikely(sq->hw_mtu < xdpf->len)) {
 		stats->err++;
 		return false;
 	}
@@ -185,9 +188,10 @@ static bool mlx5e_xmit_xdp_frame_mpwqe(struct mlx5e_xdpsq *sq,
 		mlx5e_xdp_mpwqe_session_start(sq);
 	}
 
-	mlx5e_xdp_mpwqe_add_dseg(sq, dma_addr, dma_len);
+	mlx5e_xdp_mpwqe_add_dseg(sq, xdpi, stats);
 
-	if (unlikely(session->ds_count == session->max_ds_count))
+	if (unlikely(session->complete ||
+		     session->ds_count == session->max_ds_count))
 		mlx5e_xdp_mpwqe_complete(sq);
 
 	mlx5e_xdpi_fifo_push(&sq->db.xdpi_fifo, xdpi);
@@ -301,7 +305,7 @@ bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq, struct mlx5e_rq *rq)
 
 			sqcc += wi->num_wqebbs;
 
-			for (j = 0; j < wi->num_ds; j++) {
+			for (j = 0; j < wi->num_pkts; j++) {
 				struct mlx5e_xdp_info xdpi =
 					mlx5e_xdpi_fifo_pop(xdpi_fifo);
 
@@ -342,7 +346,7 @@ void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq, struct mlx5e_rq *rq)
 
 		sq->cc += wi->num_wqebbs;
 
-		for (i = 0; i < wi->num_ds; i++) {
+		for (i = 0; i < wi->num_pkts; i++) {
 			struct mlx5e_xdp_info xdpi =
 				mlx5e_xdpi_fifo_pop(xdpi_fifo);