[dpdk-dev] [PATCH 1/5] net/mlx4: add simple Tx bypassing ibverbs

[Moti Haimovsky <motih@mellanox.com>]

PMD now sends the single-buffer packets directly to the device
bypassing the ibv Tx post and poll routines.

Signed-off-by: Moti Haimovsky <motih@mellanox.com>
---
 drivers/net/mlx4/mlx4_prm.h  | 253 +++++++++++++++++++++++++++++++++++++++++
 drivers/net/mlx4/mlx4_rxtx.c | 260 +++++++++++++++++++++++++++++++++++--------
 drivers/net/mlx4/mlx4_rxtx.h |  30 ++++-
 drivers/net/mlx4/mlx4_txq.c  |  52 ++++++++-
 mk/rte.app.mk                |   2 +-
 5 files changed, 546 insertions(+), 51 deletions(-)
 create mode 100644 drivers/net/mlx4/mlx4_prm.h

diff --git a/drivers/net/mlx4/mlx4_prm.h b/drivers/net/mlx4/mlx4_prm.h
new file mode 100644
index 0000000..c5ce33b
--- /dev/null
+++ b/drivers/net/mlx4/mlx4_prm.h
@@ -0,0 +1,253 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2017 Mellanox.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef RTE_PMD_MLX4_MLX4_PRM_H_
+#define RTE_PMD_MLX4_MLX4_PRM_H_
+
+/* Verbs headers do not support -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-Wpedantic"
+#endif
+#include <infiniband/verbs.h>
+#include <infiniband/mlx4dv.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-Wpedantic"
+#endif
+
+/* Basic TxQ building block */
+#define TXBB_SHIFT 6
+#define TXBB_SIZE (1 << TXBB_SHIFT)
+
+/* Typical TSO descriptor with 16 gather entries is 352 bytes... */
+#define MAX_WQE_SIZE		512
+#define MAX_WQE_TXBBS		(MAX_WQE_SIZE / TXBB_SIZE)
+
+/* Send Queue Stamping/Invalidating info */
+#define SQ_STAMP_STRIDE		64
+#define SQ_STAMP_DWORDS		(SQ_STAMP_STRIDE / 4)
+#define SQ_STAMP_SHIFT		31
+#define SQ_STAMP_VAL		0x7fffffff
+
+/* WQE flags */
+#define MLX4_OPCODE_SEND	0x0a
+#define MLX4_EN_BIT_WQE_OWN	0x80000000
+
+#define SIZE_TO_TXBBS(size)     (RTE_ALIGN((size), (TXBB_SIZE)) / (TXBB_SIZE))
+
+/**
+ * Update the HW with the new  CQ consumer value.
+ *
+ * @param cq
+ *   Pointer to the cq structure.
+ */
+static inline void
+mlx4_cq_set_ci(struct mlx4_cq *cq)
+{
+	*cq->set_ci_db = rte_cpu_to_be_32(cq->cons_index & 0xffffff);
+}
+
+/**
+ * Returns a pointer to the cqe in position n.
+ *
+ * @param cq
+ *   Pointer to the cq structure.
+ * @param n
+ *   The number of the entry its address we seek.
+ *
+ * @return
+ *   pointer to the cqe.
+ */
+static inline struct mlx4_cqe
+*mlx4_get_cqe(struct mlx4_cq *cq, int n)
+{
+	return (struct mlx4_cqe *)(cq->buf + n * cq->cqe_size);
+}
+
+/**
+ * Returns a pointer to the cqe in position n if it is owned by SW.
+ *
+ * @param cq
+ *   Pointer to the cq structure.
+ * @param n
+ *   The number of the entry its address we seek.
+ *
+ * @return
+ *   pointer to the cqe if owned by SW, otherwise returns NULL.
+ */
+static inline void
+*mlx4_get_sw_cqe(struct mlx4_cq *cq, int n)
+{
+	struct mlx4_cqe *cqe = mlx4_get_cqe(cq, n & (cq->cqe_cnt - 1));
+	struct mlx4_cqe *tcqe = cq->cqe_size == 64 ? cqe + 1 : cqe;
+
+	return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
+		!!(n & cq->cqe_cnt)) ? NULL : cqe;
+}
+
+/**
+ * returns pointer to the wqe at position n.
+ *
+ * @param sq
+ *   Pointer to the sq.
+ * @param n
+ *   The entry number the queue.
+ *
+ * @return
+ *   A pointer to the required entry.
+ */
+static inline void
+*mlx4_get_send_wqe(struct mlx4_sq *sq, unsigned int n)
+{
+	return sq->buf + n * TXBB_SIZE;
+}
+
+/**
+ * returns the size in bytes of this WQE.
+ *
+ * @param wqe
+ *   Pointer to the WQE we want to interrogate.
+ *
+ * @return
+ *   WQE size in bytes.
+ */
+static inline int
+mlx4_wqe_get_real_size(void *wqe)
+{
+	struct mlx4_wqe_ctrl_seg *ctrl = (struct mlx4_wqe_ctrl_seg *)wqe;
+	return ((ctrl->fence_size & 0x3f) << 4);
+}
+
+/**
+ * Fills the ctrl segment of a WQE with info needed for transmitting the packet.
+ *
+ * @param seg
+ *   Pointer to the control structure in the WQE.
+ * @param owner
+ *   The value for the owner field.
+ * @param fence_size
+ *   Fence bit and WQE size in octowords.
+ * @param srcrb_flags
+ *   High 24 bits are SRC remote buffer; low 8 bits are flags.
+ * @param imm
+ *   Immediate data/Invalidation key..
+ */
+static inline void
+mlx4_set_ctrl_seg(struct mlx4_wqe_ctrl_seg *seg, uint32_t owner,
+	     uint8_t fence_size, uint32_t srcrb_flags, uint32_t imm)
+{
+	seg->fence_size = fence_size;
+	seg->srcrb_flags = rte_cpu_to_be_32(srcrb_flags);
+	/*
+	 * The caller should prepare "imm" in advance based on WR opcode.
+	 * For IBV_WR_SEND_WITH_IMM and IBV_WR_RDMA_WRITE_WITH_IMM,
+	 * the "imm" should be assigned as is.
+	 * For the IBV_WR_SEND_WITH_INV, it should be htobe32(imm).
+	 */
+	seg->imm = imm;
+	/*
+	 * Make sure descriptor is fully written before
+	 * setting ownership bit (because HW can start
+	 * executing as soon as we do).
+	 */
+	rte_wmb();
+	seg->owner_opcode = rte_cpu_to_be_32(owner);
+}
+
+/**
+ * Fills a data segment of a WQE with info needed for transmitting
+ * a data fragment.
+ *
+ * @param dseg
+ *   Pointer to a data segment structure in the WQE.
+ *   (WQE may contain several data segments).
+ * @param sq
+ *   fragment info (addr, length, lkey).
+ */
+static inline void
+mlx4_set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ibv_sge *sg)
+{
+	dseg->lkey       = rte_cpu_to_be_32(sg->lkey);
+	dseg->addr       = rte_cpu_to_be_64(sg->addr);
+
+	/*
+	 * Need a barrier here before writing the byte_count field to
+	 * make sure that all the data is visible before the
+	 * byte_count field is set.  Otherwise, if the segment begins
+	 * a new cacheline, the HCA prefetcher could grab the 64-byte
+	 * chunk and get a valid (!= * 0xffffffff) byte count but
+	 * stale data, and end up sending the wrong data.
+	 */
+	rte_io_wmb();
+
+	if (likely(sg->length))
+		dseg->byte_count = rte_cpu_to_be_32(sg->length);
+	else
+		/* Zero len seg is treated as inline segment with zero data */
+		dseg->byte_count = rte_cpu_to_be_32(0x80000000);
+}
+
+/**
+ * Checks if a BBE sized WQE can be inserted to the sq without
+ * overflowing the Q and that theWQE is not over-sized.
+ *
+ * @param sq
+ *   Pointer to the sq we want to put the wqe in.
+ * @param ntxbb
+ *   Number of EBBs the WQE occupies.
+ */
+static inline int
+mlx4_wq_overflow(struct mlx4_sq *sq, int ntxbb)
+{
+	unsigned int cur;
+
+	cur = sq->head - sq->tail;
+	return ((cur + ntxbb + sq->headroom_txbbs >= sq->txbb_cnt) ||
+		(ntxbb > MAX_WQE_TXBBS));
+}
+
+/**
+ * Calc the  WQE size (in bytes) needed for posting this packet.
+ *
+ * @param count
+ *   The number of data-segments the WQE contains.
+ *
+ * @return
+ *   WQE size in bytes.
+ */
+static inline int
+mlx4_wqe_calc_real_size(unsigned int count)
+{
+	return sizeof(struct mlx4_wqe_ctrl_seg) +
+		(count * sizeof(struct mlx4_wqe_data_seg));
+}
+
+#endif /* RTE_PMD_MLX4_MLX4_PRM_H_ */
diff --git a/drivers/net/mlx4/mlx4_rxtx.c b/drivers/net/mlx4/mlx4_rxtx.c
index b5e7777..0720e34 100644
--- a/drivers/net/mlx4/mlx4_rxtx.c
+++ b/drivers/net/mlx4/mlx4_rxtx.c
@@ -55,10 +55,125 @@
 #include <rte_mbuf.h>
 #include <rte_mempool.h>
 #include <rte_prefetch.h>
+#include <rte_io.h>
 
 #include "mlx4.h"
 #include "mlx4_rxtx.h"
 #include "mlx4_utils.h"
+#include "mlx4_prm.h"
+
+
+typedef int bool;
+#define TRUE	1
+#define FALSE	0
+
+/**
+ * stamp a freed wqe so it won't be reused by the HW.
+ *
+ * @param sq
+ *   Pointer to the sq structure.
+ * @param index
+ *   Index of the freed WQE.
+ * @param owner
+ *   The value of the WQE owner bit to use in the stamp.
+ *
+ * @return
+ *   The number of TX EBBs the WQE contained.
+ */
+static int
+mlx4_txq_stamp_freed_wqe(struct mlx4_sq *sq, uint16_t index, uint8_t owner)
+{
+	uint32_t stamp =
+		rte_cpu_to_be_32(SQ_STAMP_VAL | (!!owner << SQ_STAMP_SHIFT));
+	void *end = sq->buf + sq->size;
+	void *wqe = mlx4_get_send_wqe(sq, index & sq->txbb_cnt_mask);
+	uint32_t *ptr = wqe;
+	int num_txbbs = SIZE_TO_TXBBS(mlx4_wqe_get_real_size(wqe));
+	int i;
+
+	/* Optimize the common case when there are no wraparounds */
+	if (likely((char *)wqe + num_txbbs * TXBB_SIZE <= (char *)end)) {
+		/* Stamp the freed descriptor */
+		for (i = 0; i < num_txbbs * TXBB_SIZE; i += SQ_STAMP_STRIDE) {
+			*ptr = stamp;
+			ptr += SQ_STAMP_DWORDS;
+		}
+	} else {
+		/* Stamp the freed descriptor */
+		for (i = 0; i < num_txbbs * TXBB_SIZE; i += SQ_STAMP_STRIDE) {
+			*ptr = stamp;
+			ptr += SQ_STAMP_DWORDS;
+			if ((void *)ptr >= end) {
+				ptr = (uint32_t *)sq->buf;
+				stamp ^= rte_cpu_to_be_32(0x80000000);
+			}
+		}
+	}
+	return num_txbbs;
+}
+
+/**
+ * Poll a CQ for work completions.
+ *
+ * @param txq
+ *   The txq its cq we wish to poll.
+ * @param max_cqes
+ *   Max num of CQEs to handle in this call.
+ *
+ * @return
+ *   The number of pkts that were handled.
+ */
+static int
+mlx4_tx_poll_cq(struct txq *txq, int max_cqes)
+{
+	struct mlx4_cq *cq = &txq->mcq;
+	struct mlx4_sq *sq = &txq->msq;
+	struct mlx4_cqe *cqe;
+	uint32_t cons_index = cq->cons_index;
+	uint16_t new_index;
+	uint16_t nr_txbbs = 0;
+	int pkts = 0;
+
+	while ((max_cqes-- > 0) &&
+		((cqe = mlx4_get_sw_cqe(cq, cons_index)) != NULL)) {
+		cqe += cq->factor;  /* handle CQES with size > 32 */
+		/*
+		 * make sure we read the CQE after we read the
+		 * ownership bit
+		 */
+		rte_rmb();
+		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
+			     MLX4_CQE_OPCODE_ERROR)) {
+			struct mlx4_err_cqe *cqe_err =
+				(struct mlx4_err_cqe *)cqe;
+			ERROR("%p CQE error - vendor syndrome: 0x%x"
+			      " syndrome: 0x%x\n",
+			      txq, cqe_err->vendor_err, cqe_err->syndrome);
+		}
+		/* Get wqe num reported in the cqe */
+		new_index =
+			rte_be_to_cpu_16(cqe->wqe_index) & sq->txbb_cnt_mask;
+		do {
+			/* free next descriptor */
+			nr_txbbs +=
+				mlx4_txq_stamp_freed_wqe(sq,
+				     (sq->tail + nr_txbbs) & sq->txbb_cnt_mask,
+				     !!((sq->tail + nr_txbbs) & sq->txbb_cnt));
+			pkts++;
+		} while (((sq->tail + nr_txbbs) & sq->txbb_cnt_mask) !=
+			 new_index);
+		++cons_index;
+	}
+	/*
+	 * To prevent CQ overflow we first update CQ consumer and only then
+	 * the ring consumer.
+	 */
+	cq->cons_index = cons_index;
+	mlx4_cq_set_ci(cq);
+	rte_wmb();
+	sq->tail = sq->tail + nr_txbbs;
+	return pkts;
+}
 
 /**
  * Manage Tx completions.
@@ -80,16 +195,15 @@
 	unsigned int elts_comp = txq->elts_comp;
 	unsigned int elts_tail = txq->elts_tail;
 	const unsigned int elts_n = txq->elts_n;
-	struct ibv_wc wcs[elts_comp];
 	int wcs_n;
 
 	if (unlikely(elts_comp == 0))
 		return 0;
-	wcs_n = ibv_poll_cq(txq->cq, elts_comp, wcs);
+	wcs_n = mlx4_tx_poll_cq(txq, 1);
 	if (unlikely(wcs_n == 0))
 		return 0;
 	if (unlikely(wcs_n < 0)) {
-		DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
+		DEBUG("%p: mlx4_poll_cq() failed (wcs_n=%d)",
 		      (void *)txq, wcs_n);
 		return -1;
 	}
@@ -99,7 +213,7 @@
 	 * Assume WC status is successful as nothing can be done about it
 	 * anyway.
 	 */
-	elts_tail += wcs_n * txq->elts_comp_cd_init;
+	elts_tail += wcs_n;
 	if (elts_tail >= elts_n)
 		elts_tail -= elts_n;
 	txq->elts_tail = elts_tail;
@@ -183,6 +297,80 @@
 }
 
 /**
+ * Notify mlx4 that work is pending on TXq.
+ *
+ * @param txq
+ *   Pointer to mlx4 Tx queue structure.
+ */
+static inline void
+mlx4_send_flush(struct txq *txq)
+{
+	rte_write32(txq->msq.doorbell_qpn, txq->msq.db);
+}
+
+/**
+ * Posts a single work requests to a send queue.
+ *
+ * @param txq
+ *   The txq to post to.
+ * @param wr
+ *   The work request to handle.
+ * @param bad_wr
+ *   the wr in case that posting had failed.
+ *
+ * @return
+ *   0 - success, -1 error.
+ */
+static int
+mlx4_post_send(struct txq *txq,
+	       struct ibv_send_wr *wr,
+	       struct ibv_send_wr **bad_wr)
+{
+	struct mlx4_wqe_ctrl_seg *ctrl;
+	struct mlx4_wqe_data_seg *dseg;
+	struct mlx4_sq *sq = &txq->msq;
+	uint32_t srcrb_flags;
+	uint8_t fence_size;
+	uint32_t head_idx = sq->head & sq->txbb_cnt_mask;
+	uint32_t owner_opcode;
+	int wqe_real_size, nr_txbbs;
+
+	/* for now we support pkts with one buf only */
+	if (wr->num_sge != 1)
+		goto err;
+	/* Calc the needed wqe size for this packet */
+	wqe_real_size = mlx4_wqe_calc_real_size(wr->num_sge);
+	if (unlikely(!wqe_real_size))
+		goto err;
+	nr_txbbs = SIZE_TO_TXBBS(wqe_real_size);
+	/* Are we too big to handle ? */
+	if (unlikely(mlx4_wq_overflow(sq, nr_txbbs)))
+		goto err;
+	/* Get ctrl and single-data wqe entries */
+	ctrl = mlx4_get_send_wqe(sq, head_idx);
+	dseg = (struct mlx4_wqe_data_seg *)(((char *)ctrl) +
+		sizeof(struct mlx4_wqe_ctrl_seg));
+	mlx4_set_data_seg(dseg, wr->sg_list);
+	/* For raw eth, the SOLICIT flag is used
+	 * to indicate that no icrc should be calculated
+	 */
+	srcrb_flags = MLX4_WQE_CTRL_SOLICIT |
+		      ((wr->send_flags & IBV_SEND_SIGNALED) ?
+				      MLX4_WQE_CTRL_CQ_UPDATE : 0);
+	fence_size = (wr->send_flags & IBV_SEND_FENCE ?
+		MLX4_WQE_CTRL_FENCE : 0) | ((wqe_real_size / 16) & 0x3f);
+	owner_opcode = MLX4_OPCODE_SEND |
+		       ((sq->head & sq->txbb_cnt) ? MLX4_EN_BIT_WQE_OWN : 0);
+	mlx4_set_ctrl_seg(ctrl, owner_opcode, fence_size, srcrb_flags, 0);
+	sq->head += nr_txbbs;
+	rte_wmb();
+	return 0;
+err:
+	*bad_wr = wr;
+	return -1;
+}
+
+/**
  * DPDK callback for Tx.
  *
  * @param dpdk_txq
@@ -199,8 +387,6 @@
 mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 {
 	struct txq *txq = (struct txq *)dpdk_txq;
-	struct ibv_send_wr *wr_head = NULL;
-	struct ibv_send_wr **wr_next = &wr_head;
 	struct ibv_send_wr *wr_bad = NULL;
 	unsigned int elts_head = txq->elts_head;
 	const unsigned int elts_n = txq->elts_n;
@@ -275,6 +461,8 @@
 				elt->buf = NULL;
 				goto stop;
 			}
+			if (buf->pkt_len <= txq->max_inline)
+				send_flags |= IBV_SEND_INLINE;
 			/* Update element. */
 			elt->buf = buf;
 			if (txq->priv->vf)
@@ -285,22 +473,31 @@
 			sge->length = length;
 			sge->lkey = lkey;
 			sent_size += length;
+			/* Set up WR. */
+			wr->sg_list  = sge;
+			wr->num_sge  = segs;
+			wr->opcode   = IBV_WR_SEND;
+			wr->send_flags = send_flags;
+			wr->next     = NULL;
+			/* post the pkt for sending */
+			err = mlx4_post_send(txq, wr, &wr_bad);
+			if (unlikely(err)) {
+				if (unlikely(wr_bad->send_flags &
+					     IBV_SEND_SIGNALED)) {
+					elts_comp_cd = 1;
+					--elts_comp;
+				}
+				elt->buf = NULL;
+				goto stop;
+			}
+			sent_size += length;
 		} else {
 			err = -1;
 			goto stop;
 		}
-		if (sent_size <= txq->max_inline)
-			send_flags |= IBV_SEND_INLINE;
 		elts_head = elts_head_next;
 		/* Increment sent bytes counter. */
 		txq->stats.obytes += sent_size;
-		/* Set up WR. */
-		wr->sg_list = &elt->sge;
-		wr->num_sge = segs;
-		wr->opcode = IBV_WR_SEND;
-		wr->send_flags = send_flags;
-		*wr_next = wr;
-		wr_next = &wr->next;
 	}
 stop:
 	/* Take a shortcut if nothing must be sent. */
@@ -309,38 +506,7 @@
 	/* Increment sent packets counter. */
 	txq->stats.opackets += i;
 	/* Ring QP doorbell. */
-	*wr_next = NULL;
-	assert(wr_head);
-	err = ibv_post_send(txq->qp, wr_head, &wr_bad);
-	if (unlikely(err)) {
-		uint64_t obytes = 0;
-		uint64_t opackets = 0;
-
-		/* Rewind bad WRs. */
-		while (wr_bad != NULL) {
-			int j;
-
-			/* Force completion request if one was lost. */
-			if (wr_bad->send_flags & IBV_SEND_SIGNALED) {
-				elts_comp_cd = 1;
-				--elts_comp;
-			}
-			++opackets;
-			for (j = 0; j < wr_bad->num_sge; ++j)
-				obytes += wr_bad->sg_list[j].length;
-			elts_head = (elts_head ? elts_head : elts_n) - 1;
-			wr_bad = wr_bad->next;
-		}
-		txq->stats.opackets -= opackets;
-		txq->stats.obytes -= obytes;
-		i -= opackets;
-		DEBUG("%p: ibv_post_send() failed, %" PRIu64 " packets"
-		      " (%" PRIu64 " bytes) rejected: %s",
-		      (void *)txq,
-		      opackets,
-		      obytes,
-		      (err <= -1) ? "Internal error" : strerror(err));
-	}
+	mlx4_send_flush(txq);
 	txq->elts_head = elts_head;
 	txq->elts_comp += elts_comp;
 	txq->elts_comp_cd = elts_comp_cd;
diff --git a/drivers/net/mlx4/mlx4_rxtx.h b/drivers/net/mlx4/mlx4_rxtx.h
index fec998a..e442730 100644
--- a/drivers/net/mlx4/mlx4_rxtx.h
+++ b/drivers/net/mlx4/mlx4_rxtx.h
@@ -41,6 +41,7 @@
 #pragma GCC diagnostic ignored "-Wpedantic"
 #endif
 #include <infiniband/verbs.h>
+#include <infiniband/mlx4dv.h>
 #ifdef PEDANTIC
 #pragma GCC diagnostic error "-Wpedantic"
 #endif
@@ -90,7 +91,7 @@ struct txq_elt {
 	struct rte_mbuf *buf; /**< Buffer. */
 };
 
-/** Rx queue counters. */
+/** Tx queue counters. */
 struct mlx4_txq_stats {
 	unsigned int idx; /**< Mapping index. */
 	uint64_t opackets; /**< Total of successfully sent packets. */
@@ -98,6 +99,31 @@ struct mlx4_txq_stats {
 	uint64_t odropped; /**< Total of packets not sent when Tx ring full. */
 };
 
+/** TXQ Info */
+struct mlx4_sq {
+	char	*buf;  /**< SQ buffer. */
+	uint32_t size; /**< SQ size in bytes. */
+	uint32_t head; /**< SQ head counter in units of TXBBS. */
+	uint32_t tail; /**< SQ tail counter in units of TXBBS. */
+	uint32_t txbb_cnt;       /**< Num of WQEBB in the Q (should be ^2). */
+	uint32_t txbb_shift;     /**< The log2 size of the basic block. */
+	uint32_t txbb_cnt_mask;	 /**< txbbs_cnt mask (txbb_cnt is ^2). */
+	uint32_t headroom_txbbs; /**< Num of txbbs that should be kept free .*/
+	uint32_t *db;            /**< Pointer to the doorbell. */
+	uint32_t doorbell_qpn;   /**< qp number to write to the doorbell. */
+};
+
+struct mlx4_cq {
+	char	 *buf;       /**< CQ buffer. */
+	uint32_t size;       /**< CQ size in bytes. */
+	uint32_t cqe_cnt;    /**< Num of entries the CQ has. */
+	int	 cqe_size;   /**< size (in bytes) of a CQE. */
+	uint32_t *set_ci_db; /**< Pionter of the consumer-index doorbell. */
+	uint32_t cons_index; /**< last CQE entry that was handled. */
+	uint32_t factor;     /**< CQ data location in a CQE. */
+	int	 cqn;        /**< CQ number */
+};
+
 /** Tx queue descriptor. */
 struct txq {
 	struct priv *priv; /**< Back pointer to private data. */
@@ -118,6 +144,8 @@ struct txq {
 	unsigned int elts_comp_cd_init; /**< Initial value for countdown. */
 	struct mlx4_txq_stats stats; /**< Tx queue counters. */
 	unsigned int socket; /**< CPU socket ID for allocations. */
+	struct mlx4_sq msq; /**< Info for directly manipulating the SQ. */
+	struct mlx4_cq mcq; /**< Info for directly manipulating the CQ. */
 };
 
 /* mlx4_rxq.c */
diff --git a/drivers/net/mlx4/mlx4_txq.c b/drivers/net/mlx4/mlx4_txq.c
index e0245b0..1273738 100644
--- a/drivers/net/mlx4/mlx4_txq.c
+++ b/drivers/net/mlx4/mlx4_txq.c
@@ -62,6 +62,7 @@
 #include "mlx4_autoconf.h"
 #include "mlx4_rxtx.h"
 #include "mlx4_utils.h"
+#include "mlx4_prm.h"
 
 /**
  * Allocate Tx queue elements.
@@ -109,7 +110,8 @@
 	assert(ret == 0);
 	return 0;
 error:
-	rte_free(elts);
+	if (elts != NULL)
+		rte_free(elts);
 	DEBUG("%p: failed, freed everything", (void *)txq);
 	assert(ret > 0);
 	rte_errno = ret;
@@ -241,6 +243,36 @@ struct txq_mp2mr_mbuf_check_data {
 	mlx4_txq_mp2mr(txq, mp);
 }
 
+static void
+mlx4_txq_fill_dv_obj_info(struct txq *txq, struct mlx4dv_obj *mlxdv)
+{
+	struct mlx4_sq *sq = &txq->msq;
+	struct mlx4_cq *cq = &txq->mcq;
+	struct mlx4dv_qp *dqp = mlxdv->qp.out;
+	struct mlx4dv_cq *dcq = mlxdv->cq.out;
+
+	sq->buf = ((char *)dqp->buf.buf) + dqp->sq.offset;
+	/* Total len, including headroom and spare WQEs*/
+	sq->size = (uint32_t)dqp->rq.offset - (uint32_t)dqp->sq.offset;
+	sq->head = 0;
+	sq->tail = 0;
+	sq->txbb_shift = TXBB_SHIFT;
+	sq->txbb_cnt = (dqp->sq.wqe_cnt << dqp->sq.wqe_shift) >> TXBB_SHIFT;
+	sq->txbb_cnt_mask = sq->txbb_cnt - 1;
+	sq->db = dqp->sdb;
+	sq->doorbell_qpn = dqp->doorbell_qpn;
+	sq->headroom_txbbs = (2048 + (1 << dqp->sq.wqe_shift)) >> TXBB_SHIFT;
+
+	/* Save CQ params */
+	cq->buf = dcq->buf.buf;
+	cq->size = dcq->buf.length;
+	cq->cqe_cnt = dcq->cqe_cnt;
+	cq->cqn = dcq->cqn;
+	cq->set_ci_db = dcq->set_ci_db;
+	cq->cqe_size = dcq->cqe_size;
+	cq->factor = cq->cqe_size > 32 ? 1 : 0;
+}
+
 /**
  * Configure a Tx queue.
  *
@@ -263,7 +295,10 @@ struct txq_mp2mr_mbuf_check_data {
 	       unsigned int socket, const struct rte_eth_txconf *conf)
 {
 	struct priv *priv = dev->data->dev_private;
-	struct txq tmpl = {
+	struct mlx4dv_obj mlxdv;
+	struct mlx4dv_qp dv_qp;
+	struct mlx4dv_cq dv_cq;
+		struct txq tmpl = {
 		.priv = priv,
 		.socket = socket
 	};
@@ -308,6 +343,8 @@ struct txq_mp2mr_mbuf_check_data {
 			/* Max number of scatter/gather elements in a WR. */
 			.max_send_sge = 1,
 			.max_inline_data = MLX4_PMD_MAX_INLINE,
+			.max_recv_wr = 0,
+			.max_recv_sge = 0,
 		},
 		.qp_type = IBV_QPT_RAW_PACKET,
 		/*
@@ -370,6 +407,17 @@ struct txq_mp2mr_mbuf_check_data {
 	DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
 	/* Pre-register known mempools. */
 	rte_mempool_walk(mlx4_txq_mp2mr_iter, txq);
+	/* Retrieve device Q info */
+	mlxdv.cq.in = txq->cq;
+	mlxdv.cq.out = &dv_cq;
+	mlxdv.qp.in = txq->qp;
+	mlxdv.qp.out = &dv_qp;
+	ret = mlx4dv_init_obj(&mlxdv, MLX4DV_OBJ_QP | MLX4DV_OBJ_CQ);
+	if (ret) {
+		ERROR("%p: Failed to retrieve device obj info", (void *)dev);
+		goto error;
+	}
+	mlx4_txq_fill_dv_obj_info(txq, &mlxdv);
 	return 0;
 error:
 	ret = rte_errno;
diff --git a/mk/rte.app.mk b/mk/rte.app.mk
index c25fdd9..2f1286e 100644
--- a/mk/rte.app.mk
+++ b/mk/rte.app.mk
@@ -128,7 +128,7 @@ ifeq ($(CONFIG_RTE_LIBRTE_KNI),y)
 _LDLIBS-$(CONFIG_RTE_LIBRTE_PMD_KNI)        += -lrte_pmd_kni
 endif
 _LDLIBS-$(CONFIG_RTE_LIBRTE_LIO_PMD)        += -lrte_pmd_lio
-_LDLIBS-$(CONFIG_RTE_LIBRTE_MLX4_PMD)       += -lrte_pmd_mlx4 -libverbs
+_LDLIBS-$(CONFIG_RTE_LIBRTE_MLX4_PMD)       += -lrte_pmd_mlx4 -libverbs -lmlx4
 _LDLIBS-$(CONFIG_RTE_LIBRTE_MLX5_PMD)       += -lrte_pmd_mlx5 -libverbs
 _LDLIBS-$(CONFIG_RTE_LIBRTE_NFP_PMD)        += -lrte_pmd_nfp
 _LDLIBS-$(CONFIG_RTE_LIBRTE_PMD_NULL)       += -lrte_pmd_null
-- 
1.8.3.1