net/mlx5: make sample and mirror action thread safe
[dpdk.git] / drivers / net / mlx5 / linux / mlx5_verbs.c
index 56570b7..494ddba 100644 (file)
@@ -4,7 +4,6 @@
 
 #include <stddef.h>
 #include <errno.h>
-#include <stdbool.h>
 #include <string.h>
 #include <stdint.h>
 #include <unistd.h>
@@ -97,21 +96,87 @@ mlx5_rxq_obj_modify_wq_vlan_strip(struct mlx5_rxq_obj *rxq_obj, int on)
  *
  * @param rxq_obj
  *   Verbs Rx queue object.
+ * @param type
+ *   Type of change queue state.
  *
  * @return
  *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 static int
-mlx5_ibv_modify_wq(struct mlx5_rxq_obj *rxq_obj, bool is_start)
+mlx5_ibv_modify_wq(struct mlx5_rxq_obj *rxq_obj, uint8_t type)
 {
        struct ibv_wq_attr mod = {
                .attr_mask = IBV_WQ_ATTR_STATE,
-               .wq_state = is_start ? IBV_WQS_RDY : IBV_WQS_RESET,
+               .wq_state = (enum ibv_wq_state)type,
        };
 
        return mlx5_glue->modify_wq(rxq_obj->wq, &mod);
 }
 
+/**
+ * Modify QP using Verbs API.
+ *
+ * @param txq_obj
+ *   Verbs Tx queue object.
+ * @param type
+ *   Type of change queue state.
+ * @param dev_port
+ *   IB device port number.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static int
+mlx5_ibv_modify_qp(struct mlx5_txq_obj *obj, enum mlx5_txq_modify_type type,
+                  uint8_t dev_port)
+{
+       struct ibv_qp_attr mod = {
+               .qp_state = IBV_QPS_RESET,
+               .port_num = dev_port,
+       };
+       int attr_mask = (IBV_QP_STATE | IBV_QP_PORT);
+       int ret;
+
+       if (type != MLX5_TXQ_MOD_RST2RDY) {
+               ret = mlx5_glue->modify_qp(obj->qp, &mod, IBV_QP_STATE);
+               if (ret) {
+                       DRV_LOG(ERR, "Cannot change Tx QP state to RESET %s",
+                               strerror(errno));
+                       rte_errno = errno;
+                       return ret;
+               }
+               if (type == MLX5_TXQ_MOD_RDY2RST)
+                       return 0;
+       }
+       if (type == MLX5_TXQ_MOD_ERR2RDY)
+               attr_mask = IBV_QP_STATE;
+       mod.qp_state = IBV_QPS_INIT;
+       ret = mlx5_glue->modify_qp(obj->qp, &mod, attr_mask);
+       if (ret) {
+               DRV_LOG(ERR, "Cannot change Tx QP state to INIT %s",
+                       strerror(errno));
+               rte_errno = errno;
+               return ret;
+       }
+       mod.qp_state = IBV_QPS_RTR;
+       ret = mlx5_glue->modify_qp(obj->qp, &mod, IBV_QP_STATE);
+       if (ret) {
+               DRV_LOG(ERR, "Cannot change Tx QP state to RTR %s",
+                       strerror(errno));
+               rte_errno = errno;
+               return ret;
+       }
+       mod.qp_state = IBV_QPS_RTS;
+       ret = mlx5_glue->modify_qp(obj->qp, &mod, IBV_QP_STATE);
+       if (ret) {
+               DRV_LOG(ERR, "Cannot change Tx QP state to RTS %s",
+                       strerror(errno));
+               rte_errno = errno;
+               return ret;
+       }
+       return 0;
+}
+
 /**
  * Create a CQ Verbs object.
  *
@@ -303,7 +368,6 @@ mlx5_rxq_ibv_obj_new(struct rte_eth_dev *dev, uint16_t idx)
        MLX5_ASSERT(tmpl);
        priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_RX_QUEUE;
        priv->verbs_alloc_ctx.obj = rxq_ctrl;
-       tmpl->type = MLX5_RXQ_OBJ_TYPE_IBV;
        tmpl->rxq_ctrl = rxq_ctrl;
        if (rxq_ctrl->irq) {
                tmpl->ibv_channel =
@@ -354,7 +418,7 @@ mlx5_rxq_ibv_obj_new(struct rte_eth_dev *dev, uint16_t idx)
                goto error;
        }
        /* Change queue state to ready. */
-       ret = mlx5_ibv_modify_wq(tmpl, true);
+       ret = mlx5_ibv_modify_wq(tmpl, IBV_WQS_RDY);
        if (ret) {
                DRV_LOG(ERR,
                        "Port %u Rx queue %u WQ state to IBV_WQS_RDY failed.",
@@ -440,67 +504,49 @@ exit:
 }
 
 /**
- * Create an indirection table.
+ * Creates a receive work queue as a filed of indirection table.
  *
  * @param dev
  *   Pointer to Ethernet device.
- * @param queues
- *   Queues entering in the indirection table.
- * @param queues_n
- *   Number of queues in the array.
+ * @param log_n
+ *   Log of number of queues in the array.
+ * @param ind_tbl
+ *   Verbs indirection table object.
  *
  * @return
- *   The Verbs object initialized, NULL otherwise and rte_errno is set.
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
-static struct mlx5_ind_table_obj *
-mlx5_ibv_ind_table_obj_new(struct rte_eth_dev *dev, const uint16_t *queues,
-                          uint32_t queues_n)
+static int
+mlx5_ibv_ind_table_new(struct rte_eth_dev *dev, const unsigned int log_n,
+                      struct mlx5_ind_table_obj *ind_tbl)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_ind_table_obj *ind_tbl;
-       const unsigned int wq_n = rte_is_power_of_2(queues_n) ?
-                                 log2above(queues_n) :
-                                 log2above(priv->config.ind_table_max_size);
-       struct ibv_wq *wq[1 << wq_n];
-       unsigned int i = 0, j = 0, k = 0;
-
-       ind_tbl = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*ind_tbl) +
-                             queues_n * sizeof(uint16_t), 0, SOCKET_ID_ANY);
-       if (!ind_tbl) {
-               rte_errno = ENOMEM;
-               return NULL;
-       }
-       for (i = 0; i != queues_n; ++i) {
-               struct mlx5_rxq_ctrl *rxq = mlx5_rxq_get(dev, queues[i]);
-               if (!rxq)
-                       goto error;
-               wq[i] = rxq->obj->wq;
-               ind_tbl->queues[i] = queues[i];
+       struct ibv_wq *wq[1 << log_n];
+       unsigned int i, j;
+
+       MLX5_ASSERT(ind_tbl);
+       for (i = 0; i != ind_tbl->queues_n; ++i) {
+               struct mlx5_rxq_data *rxq = (*priv->rxqs)[ind_tbl->queues[i]];
+               struct mlx5_rxq_ctrl *rxq_ctrl =
+                               container_of(rxq, struct mlx5_rxq_ctrl, rxq);
+
+               wq[i] = rxq_ctrl->obj->wq;
        }
-       ind_tbl->queues_n = queues_n;
+       MLX5_ASSERT(i > 0);
        /* Finalise indirection table. */
-       k = i; /* Retain value of i for use in error case. */
-       for (j = 0; k != (unsigned int)(1 << wq_n); ++k, ++j)
-               wq[k] = wq[j];
+       for (j = 0; i != (unsigned int)(1 << log_n); ++j, ++i)
+               wq[i] = wq[j];
        ind_tbl->ind_table = mlx5_glue->create_rwq_ind_table(priv->sh->ctx,
                                        &(struct ibv_rwq_ind_table_init_attr){
-                                               .log_ind_tbl_size = wq_n,
+                                               .log_ind_tbl_size = log_n,
                                                .ind_tbl = wq,
                                                .comp_mask = 0,
                                        });
        if (!ind_tbl->ind_table) {
                rte_errno = errno;
-               goto error;
+               return -rte_errno;
        }
-       rte_atomic32_inc(&ind_tbl->refcnt);
-       LIST_INSERT_HEAD(&priv->ind_tbls, ind_tbl, next);
-       return ind_tbl;
-error:
-       for (j = 0; j < i; j++)
-               mlx5_rxq_release(dev, ind_tbl->queues[j]);
-       mlx5_free(ind_tbl);
-       DEBUG("Port %u cannot create indirection table.", dev->data->port_id);
-       return NULL;
+       return 0;
 }
 
 /**
@@ -510,7 +556,7 @@ error:
  *   Indirection table to release.
  */
 static void
-mlx5_ibv_ind_table_obj_destroy(struct mlx5_ind_table_obj *ind_tbl)
+mlx5_ibv_ind_table_destroy(struct mlx5_ind_table_obj *ind_tbl)
 {
        claim_zero(mlx5_glue->destroy_rwq_ind_table(ind_tbl->ind_table));
 }
@@ -520,46 +566,24 @@ mlx5_ibv_ind_table_obj_destroy(struct mlx5_ind_table_obj *ind_tbl)
  *
  * @param dev
  *   Pointer to Ethernet device.
- * @param rss_key
- *   RSS key for the Rx hash queue.
- * @param rss_key_len
- *   RSS key length.
- * @param hash_fields
- *   Verbs protocol hash field to make the RSS on.
- * @param queues
- *   Queues entering in hash queue. In case of empty hash_fields only the
- *   first queue index will be taken for the indirection table.
- * @param queues_n
- *   Number of queues.
+ * @param hrxq
+ *   Pointer to Rx Hash queue.
  * @param tunnel
  *   Tunnel type.
  *
  * @return
- *   The Verbs object initialized index, 0 otherwise and rte_errno is set.
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
-static uint32_t
-mlx5_ibv_hrxq_new(struct rte_eth_dev *dev,
-                 const uint8_t *rss_key, uint32_t rss_key_len,
-                 uint64_t hash_fields,
-                 const uint16_t *queues, uint32_t queues_n,
+static int
+mlx5_ibv_hrxq_new(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq,
                  int tunnel __rte_unused)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_hrxq *hrxq = NULL;
-       uint32_t hrxq_idx = 0;
        struct ibv_qp *qp = NULL;
-       struct mlx5_ind_table_obj *ind_tbl;
+       struct mlx5_ind_table_obj *ind_tbl = hrxq->ind_table;
+       const uint8_t *rss_key = hrxq->rss_key;
+       uint64_t hash_fields = hrxq->hash_fields;
        int err;
-
-       queues_n = hash_fields ? queues_n : 1;
-       ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
-       if (!ind_tbl)
-               ind_tbl = priv->obj_ops->ind_table_obj_new(dev, queues,
-                                                          queues_n);
-       if (!ind_tbl) {
-               rte_errno = ENOMEM;
-               return 0;
-       }
 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
        struct mlx5dv_qp_init_attr qp_init_attr;
 
@@ -589,7 +613,7 @@ mlx5_ibv_hrxq_new(struct rte_eth_dev *dev,
                                .rx_hash_conf = (struct ibv_rx_hash_conf){
                                        .rx_hash_function =
                                                IBV_RX_HASH_FUNC_TOEPLITZ,
-                                       .rx_hash_key_len = rss_key_len,
+                                       .rx_hash_key_len = hrxq->rss_key_len,
                                        .rx_hash_key =
                                                (void *)(uintptr_t)rss_key,
                                        .rx_hash_fields_mask = hash_fields,
@@ -610,7 +634,7 @@ mlx5_ibv_hrxq_new(struct rte_eth_dev *dev,
                                .rx_hash_conf = (struct ibv_rx_hash_conf){
                                        .rx_hash_function =
                                                IBV_RX_HASH_FUNC_TOEPLITZ,
-                                       .rx_hash_key_len = rss_key_len,
+                                       .rx_hash_key_len = hrxq->rss_key_len,
                                        .rx_hash_key =
                                                (void *)(uintptr_t)rss_key,
                                        .rx_hash_fields_mask = hash_fields,
@@ -623,10 +647,6 @@ mlx5_ibv_hrxq_new(struct rte_eth_dev *dev,
                rte_errno = errno;
                goto error;
        }
-       hrxq = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_HRXQ], &hrxq_idx);
-       if (!hrxq)
-               goto error;
-       hrxq->ind_table = ind_tbl;
        hrxq->qp = qp;
 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
        hrxq->action = mlx5_glue->dv_create_flow_action_dest_ibv_qp(hrxq->qp);
@@ -635,22 +655,13 @@ mlx5_ibv_hrxq_new(struct rte_eth_dev *dev,
                goto error;
        }
 #endif
-       hrxq->rss_key_len = rss_key_len;
-       hrxq->hash_fields = hash_fields;
-       memcpy(hrxq->rss_key, rss_key, rss_key_len);
-       rte_atomic32_inc(&hrxq->refcnt);
-       ILIST_INSERT(priv->sh->ipool[MLX5_IPOOL_HRXQ], &priv->hrxqs, hrxq_idx,
-                    hrxq, next);
-       return hrxq_idx;
+       return 0;
 error:
        err = rte_errno; /* Save rte_errno before cleanup. */
-       mlx5_ind_table_obj_release(dev, ind_tbl);
        if (qp)
                claim_zero(mlx5_glue->destroy_qp(qp));
-       if (hrxq)
-               mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
        rte_errno = err; /* Restore rte_errno. */
-       return 0;
+       return -rte_errno;
 }
 
 /**
@@ -665,14 +676,408 @@ mlx5_ibv_qp_destroy(struct mlx5_hrxq *hrxq)
        claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
 }
 
+/**
+ * Release a drop Rx queue Verbs object.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ */
+static void
+mlx5_rxq_ibv_obj_drop_release(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_rxq_obj *rxq = priv->drop_queue.rxq;
+
+       if (rxq->wq)
+               claim_zero(mlx5_glue->destroy_wq(rxq->wq));
+       if (rxq->ibv_cq)
+               claim_zero(mlx5_glue->destroy_cq(rxq->ibv_cq));
+       mlx5_free(rxq);
+       priv->drop_queue.rxq = NULL;
+}
+
+/**
+ * Create a drop Rx queue Verbs object.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static int
+mlx5_rxq_ibv_obj_drop_create(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct ibv_context *ctx = priv->sh->ctx;
+       struct mlx5_rxq_obj *rxq = priv->drop_queue.rxq;
+
+       if (rxq)
+               return 0;
+       rxq = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*rxq), 0, SOCKET_ID_ANY);
+       if (!rxq) {
+               DEBUG("Port %u cannot allocate drop Rx queue memory.",
+                     dev->data->port_id);
+               rte_errno = ENOMEM;
+               return -rte_errno;
+       }
+       priv->drop_queue.rxq = rxq;
+       rxq->ibv_cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
+       if (!rxq->ibv_cq) {
+               DEBUG("Port %u cannot allocate CQ for drop queue.",
+                     dev->data->port_id);
+               rte_errno = errno;
+               goto error;
+       }
+       rxq->wq = mlx5_glue->create_wq(ctx, &(struct ibv_wq_init_attr){
+                                                   .wq_type = IBV_WQT_RQ,
+                                                   .max_wr = 1,
+                                                   .max_sge = 1,
+                                                   .pd = priv->sh->pd,
+                                                   .cq = rxq->ibv_cq,
+                                             });
+       if (!rxq->wq) {
+               DEBUG("Port %u cannot allocate WQ for drop queue.",
+                     dev->data->port_id);
+               rte_errno = errno;
+               goto error;
+       }
+       priv->drop_queue.rxq = rxq;
+       return 0;
+error:
+       mlx5_rxq_ibv_obj_drop_release(dev);
+       return -rte_errno;
+}
+
+/**
+ * Create a Verbs drop action for Rx Hash queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static int
+mlx5_ibv_drop_action_create(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
+       struct ibv_rwq_ind_table *ind_tbl = NULL;
+       struct mlx5_rxq_obj *rxq;
+       int ret;
+
+       MLX5_ASSERT(hrxq && hrxq->ind_table);
+       ret = mlx5_rxq_ibv_obj_drop_create(dev);
+       if (ret < 0)
+               goto error;
+       rxq = priv->drop_queue.rxq;
+       ind_tbl = mlx5_glue->create_rwq_ind_table
+                               (priv->sh->ctx,
+                                &(struct ibv_rwq_ind_table_init_attr){
+                                       .log_ind_tbl_size = 0,
+                                       .ind_tbl = (struct ibv_wq **)&rxq->wq,
+                                       .comp_mask = 0,
+                                });
+       if (!ind_tbl) {
+               DEBUG("Port %u cannot allocate indirection table for drop"
+                     " queue.", dev->data->port_id);
+               rte_errno = errno;
+               goto error;
+       }
+       hrxq->qp = mlx5_glue->create_qp_ex(priv->sh->ctx,
+                &(struct ibv_qp_init_attr_ex){
+                       .qp_type = IBV_QPT_RAW_PACKET,
+                       .comp_mask = IBV_QP_INIT_ATTR_PD |
+                                    IBV_QP_INIT_ATTR_IND_TABLE |
+                                    IBV_QP_INIT_ATTR_RX_HASH,
+                       .rx_hash_conf = (struct ibv_rx_hash_conf){
+                               .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
+                               .rx_hash_key_len = MLX5_RSS_HASH_KEY_LEN,
+                               .rx_hash_key = rss_hash_default_key,
+                               .rx_hash_fields_mask = 0,
+                               },
+                       .rwq_ind_tbl = ind_tbl,
+                       .pd = priv->sh->pd
+                });
+       if (!hrxq->qp) {
+               DEBUG("Port %u cannot allocate QP for drop queue.",
+                     dev->data->port_id);
+               rte_errno = errno;
+               goto error;
+       }
+#ifdef HAVE_IBV_FLOW_DV_SUPPORT
+       hrxq->action = mlx5_glue->dv_create_flow_action_dest_ibv_qp(hrxq->qp);
+       if (!hrxq->action) {
+               rte_errno = errno;
+               goto error;
+       }
+#endif
+       hrxq->ind_table->ind_table = ind_tbl;
+       return 0;
+error:
+       if (hrxq->qp)
+               claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
+       if (ind_tbl)
+               claim_zero(mlx5_glue->destroy_rwq_ind_table(ind_tbl));
+       if (priv->drop_queue.rxq)
+               mlx5_rxq_ibv_obj_drop_release(dev);
+       return -rte_errno;
+}
+
+/**
+ * Release a drop hash Rx queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ */
+static void
+mlx5_ibv_drop_action_destroy(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
+       struct ibv_rwq_ind_table *ind_tbl = hrxq->ind_table->ind_table;
+
+#ifdef HAVE_IBV_FLOW_DV_SUPPORT
+       claim_zero(mlx5_glue->destroy_flow_action(hrxq->action));
+#endif
+       claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
+       claim_zero(mlx5_glue->destroy_rwq_ind_table(ind_tbl));
+       mlx5_rxq_ibv_obj_drop_release(dev);
+}
+
+/**
+ * Create a QP Verbs object.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param idx
+ *   Queue index in DPDK Tx queue array.
+ *
+ * @return
+ *   The QP Verbs object, NULL otherwise and rte_errno is set.
+ */
+static struct ibv_qp *
+mlx5_txq_ibv_qp_create(struct rte_eth_dev *dev, uint16_t idx)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
+       struct mlx5_txq_ctrl *txq_ctrl =
+                       container_of(txq_data, struct mlx5_txq_ctrl, txq);
+       struct ibv_qp *qp_obj = NULL;
+       struct ibv_qp_init_attr_ex qp_attr = { 0 };
+       const int desc = 1 << txq_data->elts_n;
+
+       MLX5_ASSERT(txq_ctrl->obj->cq);
+       /* CQ to be associated with the send queue. */
+       qp_attr.send_cq = txq_ctrl->obj->cq;
+       /* CQ to be associated with the receive queue. */
+       qp_attr.recv_cq = txq_ctrl->obj->cq;
+       /* Max number of outstanding WRs. */
+       qp_attr.cap.max_send_wr = ((priv->sh->device_attr.max_qp_wr < desc) ?
+                                  priv->sh->device_attr.max_qp_wr : desc);
+       /*
+        * Max number of scatter/gather elements in a WR, must be 1 to prevent
+        * libmlx5 from trying to affect must be 1 to prevent libmlx5 from
+        * trying to affect too much memory. TX gather is not impacted by the
+        * device_attr.max_sge limit and will still work properly.
+        */
+       qp_attr.cap.max_send_sge = 1;
+       qp_attr.qp_type = IBV_QPT_RAW_PACKET,
+       /* Do *NOT* enable this, completions events are managed per Tx burst. */
+       qp_attr.sq_sig_all = 0;
+       qp_attr.pd = priv->sh->pd;
+       qp_attr.comp_mask = IBV_QP_INIT_ATTR_PD;
+       if (txq_data->inlen_send)
+               qp_attr.cap.max_inline_data = txq_ctrl->max_inline_data;
+       if (txq_data->tso_en) {
+               qp_attr.max_tso_header = txq_ctrl->max_tso_header;
+               qp_attr.comp_mask |= IBV_QP_INIT_ATTR_MAX_TSO_HEADER;
+       }
+       qp_obj = mlx5_glue->create_qp_ex(priv->sh->ctx, &qp_attr);
+       if (qp_obj == NULL) {
+               DRV_LOG(ERR, "Port %u Tx queue %u QP creation failure.",
+                       dev->data->port_id, idx);
+               rte_errno = errno;
+       }
+       return qp_obj;
+}
+
+/**
+ * Create the Tx queue Verbs object.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param idx
+ *   Queue index in DPDK Tx queue array.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_txq_ibv_obj_new(struct rte_eth_dev *dev, uint16_t idx)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
+       struct mlx5_txq_ctrl *txq_ctrl =
+               container_of(txq_data, struct mlx5_txq_ctrl, txq);
+       struct mlx5_txq_obj *txq_obj = txq_ctrl->obj;
+       unsigned int cqe_n;
+       struct mlx5dv_qp qp;
+       struct mlx5dv_cq cq_info;
+       struct mlx5dv_obj obj;
+       const int desc = 1 << txq_data->elts_n;
+       int ret = 0;
+
+       MLX5_ASSERT(txq_data);
+       MLX5_ASSERT(txq_obj);
+       txq_obj->txq_ctrl = txq_ctrl;
+       priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_TX_QUEUE;
+       priv->verbs_alloc_ctx.obj = txq_ctrl;
+       if (mlx5_getenv_int("MLX5_ENABLE_CQE_COMPRESSION")) {
+               DRV_LOG(ERR, "Port %u MLX5_ENABLE_CQE_COMPRESSION "
+                       "must never be set.", dev->data->port_id);
+               rte_errno = EINVAL;
+               return -rte_errno;
+       }
+       cqe_n = desc / MLX5_TX_COMP_THRESH +
+               1 + MLX5_TX_COMP_THRESH_INLINE_DIV;
+       txq_obj->cq = mlx5_glue->create_cq(priv->sh->ctx, cqe_n, NULL, NULL, 0);
+       if (txq_obj->cq == NULL) {
+               DRV_LOG(ERR, "Port %u Tx queue %u CQ creation failure.",
+                       dev->data->port_id, idx);
+               rte_errno = errno;
+               goto error;
+       }
+       txq_obj->qp = mlx5_txq_ibv_qp_create(dev, idx);
+       if (txq_obj->qp == NULL) {
+               rte_errno = errno;
+               goto error;
+       }
+       ret = mlx5_ibv_modify_qp(txq_obj, MLX5_TXQ_MOD_RST2RDY,
+                                (uint8_t)priv->dev_port);
+       if (ret) {
+               DRV_LOG(ERR, "Port %u Tx queue %u QP state modifying failed.",
+                       dev->data->port_id, idx);
+               rte_errno = errno;
+               goto error;
+       }
+       qp.comp_mask = MLX5DV_QP_MASK_UAR_MMAP_OFFSET;
+#ifdef HAVE_IBV_FLOW_DV_SUPPORT
+       /* If using DevX, need additional mask to read tisn value. */
+       if (priv->sh->devx && !priv->sh->tdn)
+               qp.comp_mask |= MLX5DV_QP_MASK_RAW_QP_HANDLES;
+#endif
+       obj.cq.in = txq_obj->cq;
+       obj.cq.out = &cq_info;
+       obj.qp.in = txq_obj->qp;
+       obj.qp.out = &qp;
+       ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_CQ | MLX5DV_OBJ_QP);
+       if (ret != 0) {
+               rte_errno = errno;
+               goto error;
+       }
+       if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) {
+               DRV_LOG(ERR,
+                       "Port %u wrong MLX5_CQE_SIZE environment variable"
+                       " value: it should be set to %u.",
+                       dev->data->port_id, RTE_CACHE_LINE_SIZE);
+               rte_errno = EINVAL;
+               goto error;
+       }
+       txq_data->cqe_n = log2above(cq_info.cqe_cnt);
+       txq_data->cqe_s = 1 << txq_data->cqe_n;
+       txq_data->cqe_m = txq_data->cqe_s - 1;
+       txq_data->qp_num_8s = ((struct ibv_qp *)txq_obj->qp)->qp_num << 8;
+       txq_data->wqes = qp.sq.buf;
+       txq_data->wqe_n = log2above(qp.sq.wqe_cnt);
+       txq_data->wqe_s = 1 << txq_data->wqe_n;
+       txq_data->wqe_m = txq_data->wqe_s - 1;
+       txq_data->wqes_end = txq_data->wqes + txq_data->wqe_s;
+       txq_data->qp_db = &qp.dbrec[MLX5_SND_DBR];
+       txq_data->cq_db = cq_info.dbrec;
+       txq_data->cqes = (volatile struct mlx5_cqe *)cq_info.buf;
+       txq_data->cq_ci = 0;
+       txq_data->cq_pi = 0;
+       txq_data->wqe_ci = 0;
+       txq_data->wqe_pi = 0;
+       txq_data->wqe_comp = 0;
+       txq_data->wqe_thres = txq_data->wqe_s / MLX5_TX_COMP_THRESH_INLINE_DIV;
+#ifdef HAVE_IBV_FLOW_DV_SUPPORT
+       /*
+        * If using DevX need to query and store TIS transport domain value.
+        * This is done once per port.
+        * Will use this value on Rx, when creating matching TIR.
+        */
+       if (priv->sh->devx && !priv->sh->tdn) {
+               ret = mlx5_devx_cmd_qp_query_tis_td(txq_obj->qp, qp.tisn,
+                                                   &priv->sh->tdn);
+               if (ret) {
+                       DRV_LOG(ERR, "Fail to query port %u Tx queue %u QP TIS "
+                               "transport domain.", dev->data->port_id, idx);
+                       rte_errno = EINVAL;
+                       goto error;
+               } else {
+                       DRV_LOG(DEBUG, "Port %u Tx queue %u TIS number %d "
+                               "transport domain %d.", dev->data->port_id,
+                               idx, qp.tisn, priv->sh->tdn);
+               }
+       }
+#endif
+       txq_ctrl->bf_reg = qp.bf.reg;
+       if (qp.comp_mask & MLX5DV_QP_MASK_UAR_MMAP_OFFSET) {
+               txq_ctrl->uar_mmap_offset = qp.uar_mmap_offset;
+               DRV_LOG(DEBUG, "Port %u: uar_mmap_offset 0x%" PRIx64 ".",
+                       dev->data->port_id, txq_ctrl->uar_mmap_offset);
+       } else {
+               DRV_LOG(ERR,
+                       "Port %u failed to retrieve UAR info, invalid"
+                       " libmlx5.so",
+                       dev->data->port_id);
+               rte_errno = EINVAL;
+               goto error;
+       }
+       txq_uar_init(txq_ctrl);
+       priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
+       return 0;
+error:
+       ret = rte_errno; /* Save rte_errno before cleanup. */
+       if (txq_obj->cq)
+               claim_zero(mlx5_glue->destroy_cq(txq_obj->cq));
+       if (txq_obj->qp)
+               claim_zero(mlx5_glue->destroy_qp(txq_obj->qp));
+       priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
+       rte_errno = ret; /* Restore rte_errno. */
+       return -rte_errno;
+}
+
+/**
+ * Release an Tx verbs queue object.
+ *
+ * @param txq_obj
+ *   Verbs Tx queue object..
+ */
+void
+mlx5_txq_ibv_obj_release(struct mlx5_txq_obj *txq_obj)
+{
+       MLX5_ASSERT(txq_obj);
+       claim_zero(mlx5_glue->destroy_qp(txq_obj->qp));
+       claim_zero(mlx5_glue->destroy_cq(txq_obj->cq));
+}
+
 struct mlx5_obj_ops ibv_obj_ops = {
        .rxq_obj_modify_vlan_strip = mlx5_rxq_obj_modify_wq_vlan_strip,
        .rxq_obj_new = mlx5_rxq_ibv_obj_new,
        .rxq_event_get = mlx5_rx_ibv_get_event,
        .rxq_obj_modify = mlx5_ibv_modify_wq,
        .rxq_obj_release = mlx5_rxq_ibv_obj_release,
-       .ind_table_obj_new = mlx5_ibv_ind_table_obj_new,
-       .ind_table_obj_destroy = mlx5_ibv_ind_table_obj_destroy,
+       .ind_table_new = mlx5_ibv_ind_table_new,
+       .ind_table_destroy = mlx5_ibv_ind_table_destroy,
        .hrxq_new = mlx5_ibv_hrxq_new,
        .hrxq_destroy = mlx5_ibv_qp_destroy,
+       .drop_action_create = mlx5_ibv_drop_action_create,
+       .drop_action_destroy = mlx5_ibv_drop_action_destroy,
+       .txq_obj_new = mlx5_txq_ibv_obj_new,
+       .txq_obj_modify = mlx5_ibv_modify_qp,
+       .txq_obj_release = mlx5_txq_ibv_obj_release,
 };