[dpdk-dev] [PATCH v1 23/29] net/mlx4: drop live queue reconfiguration support

[Adrien Mazarguil <adrien.mazarguil@6wind.com>]

DPDK ensures that setup functions are never called on configured queues,
or only if they have previously been released.

PMDs therefore do not need to deal with the unexpected reconfiguration of
live queues which may fail with no easy way to recover. Dropping support
for this scenario greatly simplifies the code as allocation and setup steps
and checks can be merged.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
---
 drivers/net/mlx4/mlx4_rxq.c  | 281 ++++++++++++++------------------------
 drivers/net/mlx4/mlx4_rxtx.h |   2 -
 drivers/net/mlx4/mlx4_txq.c  | 239 +++++++++++---------------------
 3 files changed, 184 insertions(+), 338 deletions(-)

diff --git a/drivers/net/mlx4/mlx4_rxq.c b/drivers/net/mlx4/mlx4_rxq.c
index 693db4f..30b0654 100644
--- a/drivers/net/mlx4/mlx4_rxq.c
+++ b/drivers/net/mlx4/mlx4_rxq.c
@@ -162,36 +162,12 @@ mlx4_rxq_free_elts(struct rxq *rxq)
 }
 
 /**
- * Clean up a Rx queue.
- *
- * Destroy objects, free allocated memory and reset the structure for reuse.
- *
- * @param rxq
- *   Pointer to Rx queue structure.
- */
-void
-mlx4_rxq_cleanup(struct rxq *rxq)
-{
-	DEBUG("cleaning up %p", (void *)rxq);
-	mlx4_rxq_free_elts(rxq);
-	if (rxq->qp != NULL)
-		claim_zero(ibv_destroy_qp(rxq->qp));
-	if (rxq->cq != NULL)
-		claim_zero(ibv_destroy_cq(rxq->cq));
-	if (rxq->channel != NULL)
-		claim_zero(ibv_destroy_comp_channel(rxq->channel));
-	if (rxq->mr != NULL)
-		claim_zero(ibv_dereg_mr(rxq->mr));
-	memset(rxq, 0, sizeof(*rxq));
-}
-
-/**
- * Configure a Rx queue.
+ * DPDK callback to configure a Rx queue.
  *
  * @param dev
  *   Pointer to Ethernet device structure.
- * @param rxq
- *   Pointer to Rx queue structure.
+ * @param idx
+ *   Rx queue index.
  * @param desc
  *   Number of descriptors to configure in queue.
  * @param socket
@@ -204,30 +180,53 @@ mlx4_rxq_cleanup(struct rxq *rxq)
  * @return
  *   0 on success, negative errno value otherwise and rte_errno is set.
  */
-static int
-mlx4_rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
-	       unsigned int socket, const struct rte_eth_rxconf *conf,
-	       struct rte_mempool *mp)
+int
+mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
+		    unsigned int socket, const struct rte_eth_rxconf *conf,
+		    struct rte_mempool *mp)
 {
 	struct priv *priv = dev->data->dev_private;
-	struct rxq tmpl = {
-		.priv = priv,
-		.mp = mp,
-		.socket = socket
-	};
-	struct ibv_qp_attr mod;
-	struct ibv_qp_init_attr qp_init;
-	struct ibv_recv_wr *bad_wr;
-	unsigned int mb_len;
+	uint32_t mb_len = rte_pktmbuf_data_room_size(mp);
+	struct rte_flow_error error;
+	struct rxq *rxq;
 	int ret;
 
 	(void)conf; /* Thresholds configuration (ignored). */
-	mb_len = rte_pktmbuf_data_room_size(mp);
-	if (desc == 0) {
+	DEBUG("%p: configuring queue %u for %u descriptors",
+	      (void *)dev, idx, desc);
+	if (idx >= dev->data->nb_rx_queues) {
+		rte_errno = EOVERFLOW;
+		ERROR("%p: queue index out of range (%u >= %u)",
+		      (void *)dev, idx, dev->data->nb_rx_queues);
+		return -rte_errno;
+	}
+	rxq = dev->data->rx_queues[idx];
+	if (rxq) {
+		rte_errno = EEXIST;
+		ERROR("%p: Rx queue %u already configured, release it first",
+		      (void *)dev, idx);
+		return -rte_errno;
+	}
+	if (!desc) {
 		rte_errno = EINVAL;
 		ERROR("%p: invalid number of Rx descriptors", (void *)dev);
-		goto error;
+		return -rte_errno;
+	}
+	/* Allocate and initialize Rx queue. */
+	rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
+	if (!rxq) {
+		rte_errno = ENOMEM;
+		ERROR("%p: unable to allocate queue index %u",
+		      (void *)dev, idx);
+		return -rte_errno;
 	}
+	*rxq = (struct rxq){
+		.priv = priv,
+		.mp = mp,
+		.port_id = dev->data->port_id,
+		.stats.idx = idx,
+		.socket = socket,
+	};
 	/* Enable scattered packets support for this queue if necessary. */
 	assert(mb_len >= RTE_PKTMBUF_HEADROOM);
 	if (dev->data->dev_conf.rxmode.max_rx_pkt_len <=
@@ -246,201 +245,115 @@ mlx4_rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
 		     mb_len - RTE_PKTMBUF_HEADROOM);
 	}
 	/* Use the entire Rx mempool as the memory region. */
-	tmpl.mr = mlx4_mp2mr(priv->pd, mp);
-	if (tmpl.mr == NULL) {
+	rxq->mr = mlx4_mp2mr(priv->pd, mp);
+	if (!rxq->mr) {
 		rte_errno = EINVAL;
 		ERROR("%p: MR creation failure: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
 	if (dev->data->dev_conf.intr_conf.rxq) {
-		tmpl.channel = ibv_create_comp_channel(priv->ctx);
-		if (tmpl.channel == NULL) {
+		rxq->channel = ibv_create_comp_channel(priv->ctx);
+		if (rxq->channel == NULL) {
 			rte_errno = ENOMEM;
 			ERROR("%p: Rx interrupt completion channel creation"
 			      " failure: %s",
 			      (void *)dev, strerror(rte_errno));
 			goto error;
 		}
-		if (mlx4_fd_set_non_blocking(tmpl.channel->fd) < 0) {
+		if (mlx4_fd_set_non_blocking(rxq->channel->fd) < 0) {
 			ERROR("%p: unable to make Rx interrupt completion"
 			      " channel non-blocking: %s",
 			      (void *)dev, strerror(rte_errno));
 			goto error;
 		}
 	}
-	tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, tmpl.channel, 0);
-	if (tmpl.cq == NULL) {
+	rxq->cq = ibv_create_cq(priv->ctx, desc, NULL, rxq->channel, 0);
+	if (!rxq->cq) {
 		rte_errno = ENOMEM;
 		ERROR("%p: CQ creation failure: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	DEBUG("priv->device_attr.max_qp_wr is %d",
-	      priv->device_attr.max_qp_wr);
-	DEBUG("priv->device_attr.max_sge is %d",
-	      priv->device_attr.max_sge);
-	qp_init = (struct ibv_qp_init_attr){
-		/* CQ to be associated with the send queue. */
-		.send_cq = tmpl.cq,
-		/* CQ to be associated with the receive queue. */
-		.recv_cq = tmpl.cq,
-		.cap = {
-			/* Max number of outstanding WRs. */
-			.max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
-					priv->device_attr.max_qp_wr :
-					desc),
-			/* Max number of scatter/gather elements in a WR. */
-			.max_recv_sge = 1,
-		},
-		.qp_type = IBV_QPT_RAW_PACKET,
-	};
-	tmpl.qp = ibv_create_qp(priv->pd, &qp_init);
-	if (tmpl.qp == NULL) {
+	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,
+			},
+			.qp_type = IBV_QPT_RAW_PACKET,
+		 });
+	if (!rxq->qp) {
 		rte_errno = errno ? errno : EINVAL;
 		ERROR("%p: QP creation failure: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	mod = (struct ibv_qp_attr){
-		/* Move the QP to this state. */
-		.qp_state = IBV_QPS_INIT,
-		/* Primary port number. */
-		.port_num = priv->port
-	};
-	ret = ibv_modify_qp(tmpl.qp, &mod, IBV_QP_STATE | IBV_QP_PORT);
+	ret = ibv_modify_qp
+		(rxq->qp,
+		 &(struct ibv_qp_attr){
+			.qp_state = IBV_QPS_INIT,
+			.port_num = priv->port,
+		 },
+		 IBV_QP_STATE | IBV_QP_PORT);
 	if (ret) {
 		rte_errno = ret;
 		ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	ret = mlx4_rxq_alloc_elts(&tmpl, desc);
+	ret = mlx4_rxq_alloc_elts(rxq, desc);
 	if (ret) {
 		ERROR("%p: RXQ allocation failed: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	ret = ibv_post_recv(tmpl.qp, &(*tmpl.elts)[0].wr, &bad_wr);
+	ret = ibv_post_recv(rxq->qp, &(*rxq->elts)[0].wr,
+			    &(struct ibv_recv_wr *){ NULL });
 	if (ret) {
 		rte_errno = ret;
-		ERROR("%p: ibv_post_recv() failed for WR %p: %s",
+		ERROR("%p: ibv_post_recv() failed: %s",
 		      (void *)dev,
-		      (void *)bad_wr,
 		      strerror(rte_errno));
 		goto error;
 	}
-	mod = (struct ibv_qp_attr){
-		.qp_state = IBV_QPS_RTR
-	};
-	ret = ibv_modify_qp(tmpl.qp, &mod, IBV_QP_STATE);
+	ret = ibv_modify_qp
+		(rxq->qp,
+		 &(struct ibv_qp_attr){
+			.qp_state = IBV_QPS_RTR,
+		 },
+		 IBV_QP_STATE);
 	if (ret) {
 		rte_errno = ret;
 		ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	/* Save port ID. */
-	tmpl.port_id = dev->data->port_id;
-	DEBUG("%p: RTE port ID: %u", (void *)rxq, tmpl.port_id);
-	/* Clean up rxq in case we're reinitializing it. */
-	DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq);
-	mlx4_rxq_cleanup(rxq);
-	*rxq = tmpl;
-	DEBUG("%p: rxq updated with %p", (void *)rxq, (void *)&tmpl);
-	return 0;
+	DEBUG("%p: adding Rx queue %p to list", (void *)dev, (void *)rxq);
+	dev->data->rx_queues[idx] = rxq;
+	/* Enable associated flows. */
+	ret = mlx4_flow_sync(priv, &error);
+	if (!ret)
+		return 0;
+	ERROR("cannot re-attach flow rules to queue %u"
+	      " (code %d, \"%s\"), flow error type %d, cause %p, message: %s",
+	      idx, -ret, strerror(-ret), error.type, error.cause,
+	      error.message ? error.message : "(unspecified)");
 error:
+	dev->data->rx_queues[idx] = NULL;
 	ret = rte_errno;
-	mlx4_rxq_cleanup(&tmpl);
+	mlx4_rx_queue_release(rxq);
 	rte_errno = ret;
 	assert(rte_errno > 0);
 	return -rte_errno;
 }
 
 /**
- * DPDK callback to configure a Rx queue.
- *
- * @param dev
- *   Pointer to Ethernet device structure.
- * @param idx
- *   Rx queue index.
- * @param desc
- *   Number of descriptors to configure in queue.
- * @param socket
- *   NUMA socket on which memory must be allocated.
- * @param[in] conf
- *   Thresholds parameters.
- * @param mp
- *   Memory pool for buffer allocations.
- *
- * @return
- *   0 on success, negative errno value otherwise and rte_errno is set.
- */
-int
-mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
-		    unsigned int socket, const struct rte_eth_rxconf *conf,
-		    struct rte_mempool *mp)
-{
-	struct priv *priv = dev->data->dev_private;
-	struct rxq *rxq = dev->data->rx_queues[idx];
-	int ret;
-
-	DEBUG("%p: configuring queue %u for %u descriptors",
-	      (void *)dev, idx, desc);
-	if (idx >= dev->data->nb_rx_queues) {
-		rte_errno = EOVERFLOW;
-		ERROR("%p: queue index out of range (%u >= %u)",
-		      (void *)dev, idx, dev->data->nb_rx_queues);
-		return -rte_errno;
-	}
-	if (rxq != NULL) {
-		DEBUG("%p: reusing already allocated queue index %u (%p)",
-		      (void *)dev, idx, (void *)rxq);
-		if (priv->started) {
-			rte_errno = EEXIST;
-			return -rte_errno;
-		}
-		dev->data->rx_queues[idx] = NULL;
-		/* Disable associated flows. */
-		mlx4_flow_sync(priv, NULL);
-		mlx4_rxq_cleanup(rxq);
-	} else {
-		rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
-		if (rxq == NULL) {
-			rte_errno = ENOMEM;
-			ERROR("%p: unable to allocate queue index %u",
-			      (void *)dev, idx);
-			return -rte_errno;
-		}
-	}
-	ret = mlx4_rxq_setup(dev, rxq, desc, socket, conf, mp);
-	if (ret) {
-		rte_free(rxq);
-	} else {
-		struct rte_flow_error error;
-
-		rxq->stats.idx = idx;
-		DEBUG("%p: adding Rx queue %p to list",
-		      (void *)dev, (void *)rxq);
-		dev->data->rx_queues[idx] = rxq;
-		/* Re-enable associated flows. */
-		ret = mlx4_flow_sync(priv, &error);
-		if (ret) {
-			ERROR("cannot re-attach flow rules to queue %u"
-			      " (code %d, \"%s\"), flow error type %d,"
-			      " cause %p, message: %s", idx,
-			      -ret, strerror(-ret), error.type, error.cause,
-			      error.message ? error.message : "(unspecified)");
-			dev->data->rx_queues[idx] = NULL;
-			mlx4_rxq_cleanup(rxq);
-			rte_free(rxq);
-			return ret;
-		}
-	}
-	return ret;
-}
-
-/**
  * DPDK callback to release a Rx queue.
  *
  * @param dpdk_rxq
@@ -464,6 +377,14 @@ mlx4_rx_queue_release(void *dpdk_rxq)
 			break;
 		}
 	mlx4_flow_sync(priv, NULL);
-	mlx4_rxq_cleanup(rxq);
+	mlx4_rxq_free_elts(rxq);
+	if (rxq->qp)
+		claim_zero(ibv_destroy_qp(rxq->qp));
+	if (rxq->cq)
+		claim_zero(ibv_destroy_cq(rxq->cq));
+	if (rxq->channel)
+		claim_zero(ibv_destroy_comp_channel(rxq->channel));
+	if (rxq->mr)
+		claim_zero(ibv_dereg_mr(rxq->mr));
 	rte_free(rxq);
 }
diff --git a/drivers/net/mlx4/mlx4_rxtx.h b/drivers/net/mlx4/mlx4_rxtx.h
index 7a2c982..d62120e 100644
--- a/drivers/net/mlx4/mlx4_rxtx.h
+++ b/drivers/net/mlx4/mlx4_rxtx.h
@@ -122,7 +122,6 @@ struct txq {
 
 /* mlx4_rxq.c */
 
-void mlx4_rxq_cleanup(struct rxq *rxq);
 int mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
 			uint16_t desc, unsigned int socket,
 			const struct rte_eth_rxconf *conf,
@@ -143,7 +142,6 @@ uint16_t mlx4_rx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
 
 /* mlx4_txq.c */
 
-void mlx4_txq_cleanup(struct txq *txq);
 int mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
 			uint16_t desc, unsigned int socket,
 			const struct rte_eth_txconf *conf);
diff --git a/drivers/net/mlx4/mlx4_txq.c b/drivers/net/mlx4/mlx4_txq.c
index 3cece3e..f102c68 100644
--- a/drivers/net/mlx4/mlx4_txq.c
+++ b/drivers/net/mlx4/mlx4_txq.c
@@ -155,34 +155,6 @@ mlx4_txq_free_elts(struct txq *txq)
 	rte_free(elts);
 }
 
-/**
- * Clean up a Tx queue.
- *
- * Destroy objects, free allocated memory and reset the structure for reuse.
- *
- * @param txq
- *   Pointer to Tx queue structure.
- */
-void
-mlx4_txq_cleanup(struct txq *txq)
-{
-	size_t i;
-
-	DEBUG("cleaning up %p", (void *)txq);
-	mlx4_txq_free_elts(txq);
-	if (txq->qp != NULL)
-		claim_zero(ibv_destroy_qp(txq->qp));
-	if (txq->cq != NULL)
-		claim_zero(ibv_destroy_cq(txq->cq));
-	for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
-		if (txq->mp2mr[i].mp == NULL)
-			break;
-		assert(txq->mp2mr[i].mr != NULL);
-		claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
-	}
-	memset(txq, 0, sizeof(*txq));
-}
-
 struct txq_mp2mr_mbuf_check_data {
 	int ret;
 };
@@ -242,12 +214,12 @@ mlx4_txq_mp2mr_iter(struct rte_mempool *mp, void *arg)
 }
 
 /**
- * Configure a Tx queue.
+ * DPDK callback to configure a Tx queue.
  *
  * @param dev
  *   Pointer to Ethernet device structure.
- * @param txq
- *   Pointer to Tx queue structure.
+ * @param idx
+ *   Tx queue index.
  * @param desc
  *   Number of descriptors to configure in queue.
  * @param socket
@@ -258,190 +230,135 @@ mlx4_txq_mp2mr_iter(struct rte_mempool *mp, void *arg)
  * @return
  *   0 on success, negative errno value otherwise and rte_errno is set.
  */
-static int
-mlx4_txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
-	       unsigned int socket, const struct rte_eth_txconf *conf)
+int
+mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
+		    unsigned int socket, const struct rte_eth_txconf *conf)
 {
 	struct priv *priv = dev->data->dev_private;
-	struct txq tmpl = {
-		.priv = priv,
-		.socket = socket
-	};
-	union {
-		struct ibv_qp_init_attr init;
-		struct ibv_qp_attr mod;
-	} attr;
+	struct ibv_qp_init_attr qp_init_attr;
+	struct txq *txq;
 	int ret;
 
 	(void)conf; /* Thresholds configuration (ignored). */
-	if (priv == NULL) {
-		rte_errno = EINVAL;
-		goto error;
+	DEBUG("%p: configuring queue %u for %u descriptors",
+	      (void *)dev, idx, desc);
+	if (idx >= dev->data->nb_tx_queues) {
+		rte_errno = EOVERFLOW;
+		ERROR("%p: queue index out of range (%u >= %u)",
+		      (void *)dev, idx, dev->data->nb_tx_queues);
+		return -rte_errno;
+	}
+	txq = dev->data->tx_queues[idx];
+	if (txq) {
+		rte_errno = EEXIST;
+		DEBUG("%p: Tx queue %u already configured, release it first",
+		      (void *)dev, idx);
+		return -rte_errno;
 	}
-	if (desc == 0) {
+	if (!desc) {
 		rte_errno = EINVAL;
 		ERROR("%p: invalid number of Tx descriptors", (void *)dev);
-		goto error;
+		return -rte_errno;
 	}
-	/* MRs will be registered in mp2mr[] later. */
-	tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
-	if (tmpl.cq == NULL) {
+	/* Allocate and initialize Tx queue. */
+	txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
+	if (!txq) {
+		rte_errno = ENOMEM;
+		ERROR("%p: unable to allocate queue index %u",
+		      (void *)dev, idx);
+		return -rte_errno;
+	}
+	*txq = (struct txq){
+		.priv = priv,
+		.stats.idx = idx,
+		.socket = socket,
+	};
+	txq->cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
+	if (!txq->cq) {
 		rte_errno = ENOMEM;
 		ERROR("%p: CQ creation failure: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	DEBUG("priv->device_attr.max_qp_wr is %d",
-	      priv->device_attr.max_qp_wr);
-	DEBUG("priv->device_attr.max_sge is %d",
-	      priv->device_attr.max_sge);
-	attr.init = (struct ibv_qp_init_attr){
-		/* CQ to be associated with the send queue. */
-		.send_cq = tmpl.cq,
-		/* CQ to be associated with the receive queue. */
-		.recv_cq = tmpl.cq,
+	qp_init_attr = (struct ibv_qp_init_attr){
+		.send_cq = txq->cq,
+		.recv_cq = txq->cq,
 		.cap = {
-			/* Max number of outstanding WRs. */
-			.max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
-					priv->device_attr.max_qp_wr :
-					desc),
-			/* Max number of scatter/gather elements in a WR. */
+			.max_send_wr =
+				RTE_MIN(priv->device_attr.max_qp_wr, desc),
 			.max_send_sge = 1,
 			.max_inline_data = MLX4_PMD_MAX_INLINE,
 		},
 		.qp_type = IBV_QPT_RAW_PACKET,
-		/*
-		 * Do *NOT* enable this, completions events are managed per
-		 * Tx burst.
-		 */
+		/* No completion events must occur by default. */
 		.sq_sig_all = 0,
 	};
-	tmpl.qp = ibv_create_qp(priv->pd, &attr.init);
-	if (tmpl.qp == NULL) {
+	txq->qp = ibv_create_qp(priv->pd, &qp_init_attr);
+	if (!txq->qp) {
 		rte_errno = errno ? errno : EINVAL;
 		ERROR("%p: QP creation failure: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	/* ibv_create_qp() updates this value. */
-	tmpl.max_inline = attr.init.cap.max_inline_data;
-	attr.mod = (struct ibv_qp_attr){
-		/* Move the QP to this state. */
-		.qp_state = IBV_QPS_INIT,
-		/* Primary port number. */
-		.port_num = priv->port
-	};
-	ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE | IBV_QP_PORT);
+	txq->max_inline = qp_init_attr.cap.max_inline_data;
+	ret = ibv_modify_qp
+		(txq->qp,
+		 &(struct ibv_qp_attr){
+			.qp_state = IBV_QPS_INIT,
+			.port_num = priv->port,
+		 },
+		 IBV_QP_STATE | IBV_QP_PORT);
 	if (ret) {
 		rte_errno = ret;
 		ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	ret = mlx4_txq_alloc_elts(&tmpl, desc);
+	ret = mlx4_txq_alloc_elts(txq, desc);
 	if (ret) {
 		ERROR("%p: TXQ allocation failed: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	attr.mod = (struct ibv_qp_attr){
-		.qp_state = IBV_QPS_RTR
-	};
-	ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
+	ret = ibv_modify_qp
+		(txq->qp,
+		 &(struct ibv_qp_attr){
+			.qp_state = IBV_QPS_RTR,
+		 },
+		 IBV_QP_STATE);
 	if (ret) {
 		rte_errno = ret;
 		ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	attr.mod.qp_state = IBV_QPS_RTS;
-	ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
+	ret = ibv_modify_qp
+		(txq->qp,
+		 &(struct ibv_qp_attr){
+			.qp_state = IBV_QPS_RTS,
+		 },
+		 IBV_QP_STATE);
 	if (ret) {
 		rte_errno = ret;
 		ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
 		      (void *)dev, strerror(rte_errno));
 		goto error;
 	}
-	/* Clean up txq in case we're reinitializing it. */
-	DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
-	mlx4_txq_cleanup(txq);
-	*txq = tmpl;
-	DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
 	/* Pre-register known mempools. */
 	rte_mempool_walk(mlx4_txq_mp2mr_iter, txq);
+	DEBUG("%p: adding Tx queue %p to list", (void *)dev, (void *)txq);
+	dev->data->tx_queues[idx] = txq;
 	return 0;
 error:
+	dev->data->tx_queues[idx] = NULL;
 	ret = rte_errno;
-	mlx4_txq_cleanup(&tmpl);
+	mlx4_tx_queue_release(txq);
 	rte_errno = ret;
 	assert(rte_errno > 0);
 	return -rte_errno;
 }
 
 /**
- * DPDK callback to configure a Tx queue.
- *
- * @param dev
- *   Pointer to Ethernet device structure.
- * @param idx
- *   Tx queue index.
- * @param desc
- *   Number of descriptors to configure in queue.
- * @param socket
- *   NUMA socket on which memory must be allocated.
- * @param[in] conf
- *   Thresholds parameters.
- *
- * @return
- *   0 on success, negative errno value otherwise and rte_errno is set.
- */
-int
-mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
-		    unsigned int socket, const struct rte_eth_txconf *conf)
-{
-	struct priv *priv = dev->data->dev_private;
-	struct txq *txq = dev->data->tx_queues[idx];
-	int ret;
-
-	DEBUG("%p: configuring queue %u for %u descriptors",
-	      (void *)dev, idx, desc);
-	if (idx >= dev->data->nb_tx_queues) {
-		rte_errno = EOVERFLOW;
-		ERROR("%p: queue index out of range (%u >= %u)",
-		      (void *)dev, idx, dev->data->nb_tx_queues);
-		return -rte_errno;
-	}
-	if (txq != NULL) {
-		DEBUG("%p: reusing already allocated queue index %u (%p)",
-		      (void *)dev, idx, (void *)txq);
-		if (priv->started) {
-			rte_errno = EEXIST;
-			return -rte_errno;
-		}
-		dev->data->tx_queues[idx] = NULL;
-		mlx4_txq_cleanup(txq);
-	} else {
-		txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
-		if (txq == NULL) {
-			rte_errno = ENOMEM;
-			ERROR("%p: unable to allocate queue index %u",
-			      (void *)dev, idx);
-			return -rte_errno;
-		}
-	}
-	ret = mlx4_txq_setup(dev, txq, desc, socket, conf);
-	if (ret) {
-		rte_free(txq);
-	} else {
-		txq->stats.idx = idx;
-		DEBUG("%p: adding Tx queue %p to list",
-		      (void *)dev, (void *)txq);
-		dev->data->tx_queues[idx] = txq;
-	}
-	return ret;
-}
-
-/**
  * DPDK callback to release a Tx queue.
  *
  * @param dpdk_txq
@@ -464,6 +381,16 @@ mlx4_tx_queue_release(void *dpdk_txq)
 			priv->dev->data->tx_queues[i] = NULL;
 			break;
 		}
-	mlx4_txq_cleanup(txq);
+	mlx4_txq_free_elts(txq);
+	if (txq->qp)
+		claim_zero(ibv_destroy_qp(txq->qp));
+	if (txq->cq)
+		claim_zero(ibv_destroy_cq(txq->cq));
+	for (i = 0; i != RTE_DIM(txq->mp2mr); ++i) {
+		if (!txq->mp2mr[i].mp)
+			break;
+		assert(txq->mp2mr[i].mr);
+		claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
+	}
 	rte_free(txq);
 }
-- 
2.1.4