[dpdk-dev] [PATCH v1 25/29] net/mlx4: convert Rx path to work queues

Adrien Mazarguil adrien.mazarguil at 6wind.com
Wed Oct 11 16:35:27 CEST 2017


Work queues (WQs) are lower-level than standard queue pairs (QPs). They are
dedicated to one traffic direction and have to be used in conjunction with
indirection tables and special "hash" QPs to get the same level of
functionality.

These extra objects however are the building blocks for RSS support brought
by subsequent commits, as a single "hash" QP can manage several WQs through
an indirection table according to a hash algorithm and other parameters.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil at 6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro at 6wind.com>
---
 drivers/net/mlx4/mlx4.h      |  3 ++
 drivers/net/mlx4/mlx4_rxq.c  | 74 ++++++++++++++++++++++++++++++++-------
 drivers/net/mlx4/mlx4_rxtx.c |  2 +-
 drivers/net/mlx4/mlx4_rxtx.h |  2 ++
 4 files changed, 68 insertions(+), 13 deletions(-)

diff --git a/drivers/net/mlx4/mlx4.h b/drivers/net/mlx4/mlx4.h
index a27399a..b04a104 100644
--- a/drivers/net/mlx4/mlx4.h
+++ b/drivers/net/mlx4/mlx4.h
@@ -61,6 +61,9 @@
 /** Maximum size for inline data. */
 #define MLX4_PMD_MAX_INLINE 0
 
+/** Fixed RSS hash key size in bytes. Cannot be modified. */
+#define MLX4_RSS_HASH_KEY_SIZE 40
+
 /**
  * Maximum number of cached Memory Pools (MPs) per TX queue. Each RTE MP
  * from which buffers are to be transmitted will have to be mapped by this
diff --git a/drivers/net/mlx4/mlx4_rxq.c b/drivers/net/mlx4/mlx4_rxq.c
index 03e6af5..b56f1ff 100644
--- a/drivers/net/mlx4/mlx4_rxq.c
+++ b/drivers/net/mlx4/mlx4_rxq.c
@@ -268,18 +268,64 @@ mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	rxq->qp = ibv_create_qp
-		(priv->pd,
-		 &(struct ibv_qp_init_attr){
-			.send_cq = rxq->cq,
-			.recv_cq = rxq->cq,
-			.cap = {
-				.max_recv_wr =
-					RTE_MIN(priv->device_attr.max_qp_wr,
-						desc),
-				.max_recv_sge = 1,
+	rxq->wq = ibv_create_wq
+		(priv->ctx,
+		 &(struct ibv_wq_init_attr){
+			.wq_type = IBV_WQT_RQ,
+			.max_wr = RTE_MIN(priv->device_attr.max_qp_wr, desc),
+			.max_sge = 1,
+			.pd = priv->pd,
+			.cq = rxq->cq,
+		 });
+	if (!rxq->wq) {
+		rte_errno = errno ? errno : EINVAL;
+		ERROR("%p: WQ creation failure: %s",
+		      (void *)dev, strerror(rte_errno));
+		goto error;
+	}
+	ret = ibv_modify_wq
+		(rxq->wq,
+		 &(struct ibv_wq_attr){
+			.attr_mask = IBV_WQ_ATTR_STATE,
+			.wq_state = IBV_WQS_RDY,
+		 });
+	if (ret) {
+		rte_errno = ret;
+		ERROR("%p: WQ state to IBV_WPS_RDY failed: %s",
+		      (void *)dev, strerror(rte_errno));
+		goto error;
+	}
+	rxq->ind = ibv_create_rwq_ind_table
+		(priv->ctx,
+		 &(struct ibv_rwq_ind_table_init_attr){
+			.log_ind_tbl_size = 0,
+			.ind_tbl = (struct ibv_wq *[]){
+				rxq->wq,
 			},
+			.comp_mask = 0,
+		 });
+	if (!rxq->ind) {
+		rte_errno = errno ? errno : EINVAL;
+		ERROR("%p: indirection table creation failure: %s",
+		      (void *)dev, strerror(errno));
+		goto error;
+	}
+	rxq->qp = ibv_create_qp_ex
+		(priv->ctx,
+		 &(struct ibv_qp_init_attr_ex){
+			.comp_mask = (IBV_QP_INIT_ATTR_PD |
+				      IBV_QP_INIT_ATTR_RX_HASH |
+				      IBV_QP_INIT_ATTR_IND_TABLE),
 			.qp_type = IBV_QPT_RAW_PACKET,
+			.pd = priv->pd,
+			.rwq_ind_tbl = rxq->ind,
+			.rx_hash_conf = {
+				.rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
+				.rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
+				.rx_hash_key =
+					(uint8_t [MLX4_RSS_HASH_KEY_SIZE]){ 0 },
+				.rx_hash_fields_mask = 0,
+			},
 		 });
 	if (!rxq->qp) {
 		rte_errno = errno ? errno : EINVAL;
@@ -306,8 +352,8 @@ mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	ret = ibv_post_recv(rxq->qp, &(*rxq->elts)[0].wr,
-			    &(struct ibv_recv_wr *){ NULL });
+	ret = ibv_post_wq_recv(rxq->wq, &(*rxq->elts)[0].wr,
+			       &(struct ibv_recv_wr *){ NULL });
 	if (ret) {
 		rte_errno = ret;
 		ERROR("%p: ibv_post_recv() failed: %s",
@@ -373,6 +419,10 @@ mlx4_rx_queue_release(void *dpdk_rxq)
 	mlx4_rxq_free_elts(rxq);
 	if (rxq->qp)
 		claim_zero(ibv_destroy_qp(rxq->qp));
+	if (rxq->ind)
+		claim_zero(ibv_destroy_rwq_ind_table(rxq->ind));
+	if (rxq->wq)
+		claim_zero(ibv_destroy_wq(rxq->wq));
 	if (rxq->cq)
 		claim_zero(ibv_destroy_cq(rxq->cq));
 	if (rxq->channel)
diff --git a/drivers/net/mlx4/mlx4_rxtx.c b/drivers/net/mlx4/mlx4_rxtx.c
index b5e7777..859f1bd 100644
--- a/drivers/net/mlx4/mlx4_rxtx.c
+++ b/drivers/net/mlx4/mlx4_rxtx.c
@@ -459,7 +459,7 @@ mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	/* Repost WRs. */
 	*wr_next = NULL;
 	assert(wr_head);
-	ret = ibv_post_recv(rxq->qp, wr_head, &wr_bad);
+	ret = ibv_post_wq_recv(rxq->wq, wr_head, &wr_bad);
 	if (unlikely(ret)) {
 		/* Inability to repost WRs is fatal. */
 		DEBUG("%p: recv_burst(): failed (ret=%d)",
diff --git a/drivers/net/mlx4/mlx4_rxtx.h b/drivers/net/mlx4/mlx4_rxtx.h
index d90f2f9..897fd2a 100644
--- a/drivers/net/mlx4/mlx4_rxtx.h
+++ b/drivers/net/mlx4/mlx4_rxtx.h
@@ -73,6 +73,8 @@ struct rxq {
 	struct rte_mempool *mp; /**< Memory pool for allocations. */
 	struct ibv_mr *mr; /**< Memory region (for mp). */
 	struct ibv_cq *cq; /**< Completion queue. */
+	struct ibv_wq *wq; /**< Work queue. */
+	struct ibv_rwq_ind_table *ind; /**< Indirection table. */
 	struct ibv_qp *qp; /**< Queue pair. */
 	struct ibv_comp_channel *channel; /**< Rx completion channel. */
 	unsigned int port_id; /**< Port ID for incoming packets. */
-- 
2.1.4



More information about the dev mailing list