net/mlx5: use flow to enable promiscuous mode
[dpdk.git] / drivers / net / mlx5 / mlx5_flow.c
index 7dd3ebb..8512905 100644 (file)
@@ -52,6 +52,9 @@
 #include "mlx5.h"
 #include "mlx5_prm.h"
 
+/* Define minimal priority for control plane flows. */
+#define MLX5_CTRL_FLOW_PRIORITY 4
+
 static int
 mlx5_flow_create_eth(const struct rte_flow_item *item,
                     const void *default_mask,
@@ -87,19 +90,37 @@ mlx5_flow_create_vxlan(const struct rte_flow_item *item,
                       const void *default_mask,
                       void *data);
 
-struct rte_flow {
-       TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
-       struct ibv_exp_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
-       struct ibv_exp_rwq_ind_table *ind_table; /**< Indirection table. */
+/** Structure for Drop queue. */
+struct mlx5_hrxq_drop {
+       struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
        struct ibv_qp *qp; /**< Verbs queue pair. */
-       struct ibv_exp_flow *ibv_flow; /**< Verbs flow. */
-       struct ibv_exp_wq *wq; /**< Verbs work queue. */
+       struct ibv_wq *wq; /**< Verbs work queue. */
        struct ibv_cq *cq; /**< Verbs completion queue. */
-       uint16_t rxqs_n; /**< Number of queues in this flow, 0 if drop queue. */
+};
+
+/* Flows structures. */
+struct mlx5_flow {
+       uint64_t hash_fields; /**< Fields that participate in the hash. */
+       struct mlx5_hrxq *hrxq; /**< Hash Rx queues. */
+};
+
+/* Drop flows structures. */
+struct mlx5_flow_drop {
+       struct mlx5_hrxq_drop hrxq; /**< Drop hash Rx queue. */
+};
+
+struct rte_flow {
+       TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
        uint32_t mark:1; /**< Set if the flow is marked. */
        uint32_t drop:1; /**< Drop queue. */
-       uint64_t hash_fields; /**< Fields that participate in the hash. */
-       struct rxq *rxqs[]; /**< Pointer to the queues array. */
+       struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
+       struct ibv_flow *ibv_flow; /**< Verbs flow. */
+       uint16_t queues_n; /**< Number of entries in queue[]. */
+       uint16_t (*queues)[]; /**< Queues indexes to use. */
+       union {
+               struct mlx5_flow frxq; /**< Flow with Rx queue. */
+               struct mlx5_flow_drop drxq; /**< Flow with drop Rx queue. */
+       };
 };
 
 /** Static initializer for items. */
@@ -172,7 +193,7 @@ static const struct mlx5_flow_items mlx5_flow_items[] = {
                .default_mask = &rte_flow_item_eth_mask,
                .mask_sz = sizeof(struct rte_flow_item_eth),
                .convert = mlx5_flow_create_eth,
-               .dst_sz = sizeof(struct ibv_exp_flow_spec_eth),
+               .dst_sz = sizeof(struct ibv_flow_spec_eth),
        },
        [RTE_FLOW_ITEM_TYPE_VLAN] = {
                .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
@@ -201,7 +222,7 @@ static const struct mlx5_flow_items mlx5_flow_items[] = {
                .default_mask = &rte_flow_item_ipv4_mask,
                .mask_sz = sizeof(struct rte_flow_item_ipv4),
                .convert = mlx5_flow_create_ipv4,
-               .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv4_ext),
+               .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
        },
        [RTE_FLOW_ITEM_TYPE_IPV6] = {
                .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
@@ -229,7 +250,7 @@ static const struct mlx5_flow_items mlx5_flow_items[] = {
                .default_mask = &rte_flow_item_ipv6_mask,
                .mask_sz = sizeof(struct rte_flow_item_ipv6),
                .convert = mlx5_flow_create_ipv6,
-               .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv6_ext),
+               .dst_sz = sizeof(struct ibv_flow_spec_ipv6),
        },
        [RTE_FLOW_ITEM_TYPE_UDP] = {
                .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
@@ -243,7 +264,7 @@ static const struct mlx5_flow_items mlx5_flow_items[] = {
                .default_mask = &rte_flow_item_udp_mask,
                .mask_sz = sizeof(struct rte_flow_item_udp),
                .convert = mlx5_flow_create_udp,
-               .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
+               .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
        },
        [RTE_FLOW_ITEM_TYPE_TCP] = {
                .actions = valid_actions,
@@ -256,7 +277,7 @@ static const struct mlx5_flow_items mlx5_flow_items[] = {
                .default_mask = &rte_flow_item_tcp_mask,
                .mask_sz = sizeof(struct rte_flow_item_tcp),
                .convert = mlx5_flow_create_tcp,
-               .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
+               .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
        },
        [RTE_FLOW_ITEM_TYPE_VXLAN] = {
                .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
@@ -267,26 +288,11 @@ static const struct mlx5_flow_items mlx5_flow_items[] = {
                .default_mask = &rte_flow_item_vxlan_mask,
                .mask_sz = sizeof(struct rte_flow_item_vxlan),
                .convert = mlx5_flow_create_vxlan,
-               .dst_sz = sizeof(struct ibv_exp_flow_spec_tunnel),
+               .dst_sz = sizeof(struct ibv_flow_spec_tunnel),
        },
 };
 
-/** Structure to pass to the conversion function. */
-struct mlx5_flow {
-       struct ibv_exp_flow_attr *ibv_attr; /**< Verbs attribute. */
-       unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
-       uint32_t inner; /**< Set once VXLAN is encountered. */
-       uint64_t hash_fields; /**< Fields that participate in the hash. */
-};
-
-/** Structure for Drop queue. */
-struct rte_flow_drop {
-       struct ibv_exp_rwq_ind_table *ind_table; /**< Indirection table. */
-       struct ibv_qp *qp; /**< Verbs queue pair. */
-       struct ibv_exp_wq *wq; /**< Verbs work queue. */
-       struct ibv_cq *cq; /**< Verbs completion queue. */
-};
-
+/* Structure to parse actions. */
 struct mlx5_flow_action {
        uint32_t queue:1; /**< Target is a receive queue. */
        uint32_t drop:1; /**< Target is a drop queue. */
@@ -296,6 +302,58 @@ struct mlx5_flow_action {
        uint16_t queues_n; /**< Number of entries in queue[]. */
 };
 
+/** Structure to pass to the conversion function. */
+struct mlx5_flow_parse {
+       struct ibv_flow_attr *ibv_attr; /**< Verbs attribute. */
+       unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
+       uint32_t inner; /**< Set once VXLAN is encountered. */
+       uint64_t hash_fields; /**< Fields that participate in the hash. */
+       struct mlx5_flow_action actions; /**< Parsed action result. */
+};
+
+static const struct rte_flow_ops mlx5_flow_ops = {
+       .validate = mlx5_flow_validate,
+       .create = mlx5_flow_create,
+       .destroy = mlx5_flow_destroy,
+       .flush = mlx5_flow_flush,
+       .query = NULL,
+       .isolate = mlx5_flow_isolate,
+};
+
+/**
+ * Manage filter operations.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param filter_type
+ *   Filter type.
+ * @param filter_op
+ *   Operation to perform.
+ * @param arg
+ *   Pointer to operation-specific structure.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
+                    enum rte_filter_type filter_type,
+                    enum rte_filter_op filter_op,
+                    void *arg)
+{
+       int ret = EINVAL;
+
+       if (filter_type == RTE_ETH_FILTER_GENERIC) {
+               if (filter_op != RTE_ETH_FILTER_GET)
+                       return -EINVAL;
+               *(const void **)arg = &mlx5_flow_ops;
+               return 0;
+       }
+       ERROR("%p: filter type (%d) not supported",
+             (void *)dev, filter_type);
+       return -ret;
+}
+
 /**
  * Check support for a given item.
  *
@@ -374,8 +432,6 @@ mlx5_flow_item_validate(const struct rte_flow_item *item,
  *   Perform verbose error reporting if not NULL.
  * @param[in, out] flow
  *   Flow structure to update.
- * @param[in, out] action
- *   Action structure to update.
  *
  * @return
  *   0 on success, a negative errno value otherwise and rte_errno is set.
@@ -386,8 +442,7 @@ priv_flow_validate(struct priv *priv,
                   const struct rte_flow_item items[],
                   const struct rte_flow_action actions[],
                   struct rte_flow_error *error,
-                  struct mlx5_flow *flow,
-                  struct mlx5_flow_action *action)
+                  struct mlx5_flow_parse *flow)
 {
        const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 
@@ -399,7 +454,7 @@ priv_flow_validate(struct priv *priv,
                                   "groups are not supported");
                return -rte_errno;
        }
-       if (attr->priority) {
+       if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) {
                rte_flow_error_set(error, ENOTSUP,
                                   RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
                                   NULL,
@@ -469,7 +524,7 @@ priv_flow_validate(struct priv *priv,
                if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
                        continue;
                } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
-                       action->drop = 1;
+                       flow->actions.drop = 1;
                } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
                        const struct rte_flow_action_queue *queue =
                                (const struct rte_flow_action_queue *)
@@ -479,13 +534,13 @@ priv_flow_validate(struct priv *priv,
 
                        if (!queue || (queue->index > (priv->rxqs_n - 1)))
                                goto exit_action_not_supported;
-                       for (n = 0; n < action->queues_n; ++n) {
-                               if (action->queues[n] == queue->index) {
+                       for (n = 0; n < flow->actions.queues_n; ++n) {
+                               if (flow->actions.queues[n] == queue->index) {
                                        found = 1;
                                        break;
                                }
                        }
-                       if (action->queues_n > 1 && !found) {
+                       if (flow->actions.queues_n > 1 && !found) {
                                rte_flow_error_set(error, ENOTSUP,
                                           RTE_FLOW_ERROR_TYPE_ACTION,
                                           actions,
@@ -493,9 +548,9 @@ priv_flow_validate(struct priv *priv,
                                return -rte_errno;
                        }
                        if (!found) {
-                               action->queue = 1;
-                               action->queues_n = 1;
-                               action->queues[0] = queue->index;
+                               flow->actions.queue = 1;
+                               flow->actions.queues_n = 1;
+                               flow->actions.queues[0] = queue->index;
                        }
                } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
                        const struct rte_flow_action_rss *rss =
@@ -510,12 +565,12 @@ priv_flow_validate(struct priv *priv,
                                                   "no valid queues");
                                return -rte_errno;
                        }
-                       if (action->queues_n == 1) {
+                       if (flow->actions.queues_n == 1) {
                                uint16_t found = 0;
 
-                               assert(action->queues_n);
+                               assert(flow->actions.queues_n);
                                for (n = 0; n < rss->num; ++n) {
-                                       if (action->queues[0] ==
+                                       if (flow->actions.queues[0] ==
                                            rss->queue[n]) {
                                                found = 1;
                                                break;
@@ -540,10 +595,10 @@ priv_flow_validate(struct priv *priv,
                                        return -rte_errno;
                                }
                        }
-                       action->queue = 1;
+                       flow->actions.queue = 1;
                        for (n = 0; n < rss->num; ++n)
-                               action->queues[n] = rss->queue[n];
-                       action->queues_n = rss->num;
+                               flow->actions.queues[n] = rss->queue[n];
+                       flow->actions.queues_n = rss->num;
                } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
                        const struct rte_flow_action_mark *mark =
                                (const struct rte_flow_action_mark *)
@@ -563,19 +618,19 @@ priv_flow_validate(struct priv *priv,
                                                   " and 16777199");
                                return -rte_errno;
                        }
-                       action->mark = 1;
-                       action->mark_id = mark->id;
+                       flow->actions.mark = 1;
+                       flow->actions.mark_id = mark->id;
                } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
-                       action->mark = 1;
+                       flow->actions.mark = 1;
                } else {
                        goto exit_action_not_supported;
                }
        }
-       if (action->mark && !flow->ibv_attr && !action->drop)
-               flow->offset += sizeof(struct ibv_exp_flow_spec_action_tag);
-       if (!flow->ibv_attr && action->drop)
-               flow->offset += sizeof(struct ibv_exp_flow_spec_action_drop);
-       if (!action->queue && !action->drop) {
+       if (flow->actions.mark && !flow->ibv_attr && !flow->actions.drop)
+               flow->offset += sizeof(struct ibv_flow_spec_action_tag);
+       if (!flow->ibv_attr && flow->actions.drop)
+               flow->offset += sizeof(struct ibv_flow_spec_action_drop);
+       if (!flow->actions.queue && !flow->actions.drop) {
                rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
                                   NULL, "no valid action");
                return -rte_errno;
@@ -606,18 +661,16 @@ mlx5_flow_validate(struct rte_eth_dev *dev,
 {
        struct priv *priv = dev->data->dev_private;
        int ret;
-       struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr) };
-       struct mlx5_flow_action action = {
-               .queue = 0,
-               .drop = 0,
-               .mark = 0,
-               .mark_id = MLX5_FLOW_MARK_DEFAULT,
-               .queues_n = 0,
+       struct mlx5_flow_parse flow = {
+               .offset = sizeof(struct ibv_flow_attr),
+               .actions = {
+                       .mark_id = MLX5_FLOW_MARK_DEFAULT,
+                       .queues_n = 0,
+               },
        };
 
        priv_lock(priv);
-       ret = priv_flow_validate(priv, attr, items, actions, error, &flow,
-                                &action);
+       ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
        priv_unlock(priv);
        return ret;
 }
@@ -639,17 +692,17 @@ mlx5_flow_create_eth(const struct rte_flow_item *item,
 {
        const struct rte_flow_item_eth *spec = item->spec;
        const struct rte_flow_item_eth *mask = item->mask;
-       struct mlx5_flow *flow = (struct mlx5_flow *)data;
-       struct ibv_exp_flow_spec_eth *eth;
-       const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
+       struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
+       struct ibv_flow_spec_eth *eth;
+       const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
        unsigned int i;
 
        ++flow->ibv_attr->num_of_specs;
        flow->ibv_attr->priority = 2;
        flow->hash_fields = 0;
        eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
-       *eth = (struct ibv_exp_flow_spec_eth) {
-               .type = flow->inner | IBV_EXP_FLOW_SPEC_ETH,
+       *eth = (struct ibv_flow_spec_eth) {
+               .type = flow->inner | IBV_FLOW_SPEC_ETH,
                .size = eth_size,
        };
        if (!spec)
@@ -688,9 +741,9 @@ mlx5_flow_create_vlan(const struct rte_flow_item *item,
 {
        const struct rte_flow_item_vlan *spec = item->spec;
        const struct rte_flow_item_vlan *mask = item->mask;
-       struct mlx5_flow *flow = (struct mlx5_flow *)data;
-       struct ibv_exp_flow_spec_eth *eth;
-       const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
+       struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
+       struct ibv_flow_spec_eth *eth;
+       const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
 
        eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
        if (!spec)
@@ -720,30 +773,30 @@ mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 {
        const struct rte_flow_item_ipv4 *spec = item->spec;
        const struct rte_flow_item_ipv4 *mask = item->mask;
-       struct mlx5_flow *flow = (struct mlx5_flow *)data;
-       struct ibv_exp_flow_spec_ipv4_ext *ipv4;
-       unsigned int ipv4_size = sizeof(struct ibv_exp_flow_spec_ipv4_ext);
+       struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
+       struct ibv_flow_spec_ipv4_ext *ipv4;
+       unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
 
        ++flow->ibv_attr->num_of_specs;
        flow->ibv_attr->priority = 1;
-       flow->hash_fields = (IBV_EXP_RX_HASH_SRC_IPV4 |
-                            IBV_EXP_RX_HASH_DST_IPV4);
+       flow->hash_fields = (IBV_RX_HASH_SRC_IPV4 |
+                            IBV_RX_HASH_DST_IPV4);
        ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
-       *ipv4 = (struct ibv_exp_flow_spec_ipv4_ext) {
-               .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV4_EXT,
+       *ipv4 = (struct ibv_flow_spec_ipv4_ext) {
+               .type = flow->inner | IBV_FLOW_SPEC_IPV4_EXT,
                .size = ipv4_size,
        };
        if (!spec)
                return 0;
        if (!mask)
                mask = default_mask;
-       ipv4->val = (struct ibv_exp_flow_ipv4_ext_filter){
+       ipv4->val = (struct ibv_flow_ipv4_ext_filter){
                .src_ip = spec->hdr.src_addr,
                .dst_ip = spec->hdr.dst_addr,
                .proto = spec->hdr.next_proto_id,
                .tos = spec->hdr.type_of_service,
        };
-       ipv4->mask = (struct ibv_exp_flow_ipv4_ext_filter){
+       ipv4->mask = (struct ibv_flow_ipv4_ext_filter){
                .src_ip = mask->hdr.src_addr,
                .dst_ip = mask->hdr.dst_addr,
                .proto = mask->hdr.next_proto_id,
@@ -774,18 +827,18 @@ mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 {
        const struct rte_flow_item_ipv6 *spec = item->spec;
        const struct rte_flow_item_ipv6 *mask = item->mask;
-       struct mlx5_flow *flow = (struct mlx5_flow *)data;
-       struct ibv_exp_flow_spec_ipv6_ext *ipv6;
-       unsigned int ipv6_size = sizeof(struct ibv_exp_flow_spec_ipv6_ext);
+       struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
+       struct ibv_flow_spec_ipv6 *ipv6;
+       unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
        unsigned int i;
 
        ++flow->ibv_attr->num_of_specs;
        flow->ibv_attr->priority = 1;
-       flow->hash_fields = (IBV_EXP_RX_HASH_SRC_IPV6 |
-                            IBV_EXP_RX_HASH_DST_IPV6);
+       flow->hash_fields = (IBV_RX_HASH_SRC_IPV6 |
+                            IBV_RX_HASH_DST_IPV6);
        ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
-       *ipv6 = (struct ibv_exp_flow_spec_ipv6_ext) {
-               .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV6_EXT,
+       *ipv6 = (struct ibv_flow_spec_ipv6) {
+               .type = flow->inner | IBV_FLOW_SPEC_IPV6,
                .size = ipv6_size,
        };
        if (!spec)
@@ -831,17 +884,17 @@ mlx5_flow_create_udp(const struct rte_flow_item *item,
 {
        const struct rte_flow_item_udp *spec = item->spec;
        const struct rte_flow_item_udp *mask = item->mask;
-       struct mlx5_flow *flow = (struct mlx5_flow *)data;
-       struct ibv_exp_flow_spec_tcp_udp *udp;
-       unsigned int udp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
+       struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
+       struct ibv_flow_spec_tcp_udp *udp;
+       unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 
        ++flow->ibv_attr->num_of_specs;
        flow->ibv_attr->priority = 0;
-       flow->hash_fields |= (IBV_EXP_RX_HASH_SRC_PORT_UDP |
-                             IBV_EXP_RX_HASH_DST_PORT_UDP);
+       flow->hash_fields |= (IBV_RX_HASH_SRC_PORT_UDP |
+                             IBV_RX_HASH_DST_PORT_UDP);
        udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
-       *udp = (struct ibv_exp_flow_spec_tcp_udp) {
-               .type = flow->inner | IBV_EXP_FLOW_SPEC_UDP,
+       *udp = (struct ibv_flow_spec_tcp_udp) {
+               .type = flow->inner | IBV_FLOW_SPEC_UDP,
                .size = udp_size,
        };
        if (!spec)
@@ -875,17 +928,17 @@ mlx5_flow_create_tcp(const struct rte_flow_item *item,
 {
        const struct rte_flow_item_tcp *spec = item->spec;
        const struct rte_flow_item_tcp *mask = item->mask;
-       struct mlx5_flow *flow = (struct mlx5_flow *)data;
-       struct ibv_exp_flow_spec_tcp_udp *tcp;
-       unsigned int tcp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
+       struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
+       struct ibv_flow_spec_tcp_udp *tcp;
+       unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 
        ++flow->ibv_attr->num_of_specs;
        flow->ibv_attr->priority = 0;
-       flow->hash_fields |= (IBV_EXP_RX_HASH_SRC_PORT_TCP |
-                             IBV_EXP_RX_HASH_DST_PORT_TCP);
+       flow->hash_fields |= (IBV_RX_HASH_SRC_PORT_TCP |
+                             IBV_RX_HASH_DST_PORT_TCP);
        tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
-       *tcp = (struct ibv_exp_flow_spec_tcp_udp) {
-               .type = flow->inner | IBV_EXP_FLOW_SPEC_TCP,
+       *tcp = (struct ibv_flow_spec_tcp_udp) {
+               .type = flow->inner | IBV_FLOW_SPEC_TCP,
                .size = tcp_size,
        };
        if (!spec)
@@ -919,9 +972,9 @@ mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 {
        const struct rte_flow_item_vxlan *spec = item->spec;
        const struct rte_flow_item_vxlan *mask = item->mask;
-       struct mlx5_flow *flow = (struct mlx5_flow *)data;
-       struct ibv_exp_flow_spec_tunnel *vxlan;
-       unsigned int size = sizeof(struct ibv_exp_flow_spec_tunnel);
+       struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
+       struct ibv_flow_spec_tunnel *vxlan;
+       unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
        union vni {
                uint32_t vlan_id;
                uint8_t vni[4];
@@ -931,11 +984,11 @@ mlx5_flow_create_vxlan(const struct rte_flow_item *item,
        flow->ibv_attr->priority = 0;
        id.vni[0] = 0;
        vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
-       *vxlan = (struct ibv_exp_flow_spec_tunnel) {
-               .type = flow->inner | IBV_EXP_FLOW_SPEC_VXLAN_TUNNEL,
+       *vxlan = (struct ibv_flow_spec_tunnel) {
+               .type = flow->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
                .size = size,
        };
-       flow->inner = IBV_EXP_FLOW_SPEC_INNER;
+       flow->inner = IBV_FLOW_SPEC_INNER;
        if (!spec)
                return 0;
        if (!mask)
@@ -958,14 +1011,14 @@ mlx5_flow_create_vxlan(const struct rte_flow_item *item,
  *   Mark identifier.
  */
 static int
-mlx5_flow_create_flag_mark(struct mlx5_flow *flow, uint32_t mark_id)
+mlx5_flow_create_flag_mark(struct mlx5_flow_parse *flow, uint32_t mark_id)
 {
-       struct ibv_exp_flow_spec_action_tag *tag;
-       unsigned int size = sizeof(struct ibv_exp_flow_spec_action_tag);
+       struct ibv_flow_spec_action_tag *tag;
+       unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
 
        tag = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
-       *tag = (struct ibv_exp_flow_spec_action_tag){
-               .type = IBV_EXP_FLOW_SPEC_ACTION_TAG,
+       *tag = (struct ibv_flow_spec_action_tag){
+               .type = IBV_FLOW_SPEC_ACTION_TAG,
                .size = size,
                .tag_id = mlx5_flow_mark_set(mark_id),
        };
@@ -988,12 +1041,12 @@ mlx5_flow_create_flag_mark(struct mlx5_flow *flow, uint32_t mark_id)
  */
 static struct rte_flow *
 priv_flow_create_action_queue_drop(struct priv *priv,
-                                  struct mlx5_flow *flow,
+                                  struct mlx5_flow_parse *flow,
                                   struct rte_flow_error *error)
 {
        struct rte_flow *rte_flow;
-       struct ibv_exp_flow_spec_action_drop *drop;
-       unsigned int size = sizeof(struct ibv_exp_flow_spec_action_drop);
+       struct ibv_flow_spec_action_drop *drop;
+       unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
 
        assert(priv->pd);
        assert(priv->ctx);
@@ -1005,18 +1058,18 @@ priv_flow_create_action_queue_drop(struct priv *priv,
        }
        rte_flow->drop = 1;
        drop = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
-       *drop = (struct ibv_exp_flow_spec_action_drop){
-                       .type = IBV_EXP_FLOW_SPEC_ACTION_DROP,
+       *drop = (struct ibv_flow_spec_action_drop){
+                       .type = IBV_FLOW_SPEC_ACTION_DROP,
                        .size = size,
        };
        ++flow->ibv_attr->num_of_specs;
-       flow->offset += sizeof(struct ibv_exp_flow_spec_action_drop);
+       flow->offset += sizeof(struct ibv_flow_spec_action_drop);
        rte_flow->ibv_attr = flow->ibv_attr;
-       if (!priv->started)
+       if (!priv->dev->data->dev_started)
                return rte_flow;
-       rte_flow->qp = priv->flow_drop_queue->qp;
-       rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
-                                                rte_flow->ibv_attr);
+       rte_flow->drxq.hrxq.qp = priv->flow_drop_queue->qp;
+       rte_flow->ibv_flow = ibv_create_flow(rte_flow->drxq.hrxq.qp,
+                                            rte_flow->ibv_attr);
        if (!rte_flow->ibv_flow) {
                rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
                                   NULL, "flow rule creation failure");
@@ -1036,8 +1089,6 @@ error:
  *   Pointer to private structure.
  * @param flow
  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
- * @param action
- *   Target action structure.
  * @param[out] error
  *   Perform verbose error reporting if not NULL.
  *
@@ -1046,86 +1097,62 @@ error:
  */
 static struct rte_flow *
 priv_flow_create_action_queue(struct priv *priv,
-                             struct mlx5_flow *flow,
-                             struct mlx5_flow_action *action,
+                             struct mlx5_flow_parse *flow,
                              struct rte_flow_error *error)
 {
        struct rte_flow *rte_flow;
        unsigned int i;
-       unsigned int j;
-       const unsigned int wqs_n = 1 << log2above(action->queues_n);
-       struct ibv_exp_wq *wqs[wqs_n];
 
        assert(priv->pd);
        assert(priv->ctx);
-       assert(!action->drop);
-       rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow) +
-                             sizeof(*rte_flow->rxqs) * action->queues_n, 0);
+       assert(!flow->actions.drop);
+       rte_flow =
+               rte_calloc(__func__, 1,
+                          sizeof(*flow) +
+                          flow->actions.queues_n * sizeof(uint16_t),
+                          0);
        if (!rte_flow) {
                rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
                                   NULL, "cannot allocate flow memory");
                return NULL;
        }
-       for (i = 0; i < action->queues_n; ++i) {
-               struct rxq_ctrl *rxq;
-
-               rxq = container_of((*priv->rxqs)[action->queues[i]],
-                                  struct rxq_ctrl, rxq);
-               wqs[i] = rxq->wq;
-               rte_flow->rxqs[i] = &rxq->rxq;
-               ++rte_flow->rxqs_n;
-               rxq->rxq.mark |= action->mark;
-       }
-       /* finalise indirection table. */
-       for (j = 0; i < wqs_n; ++i, ++j) {
-               wqs[i] = wqs[j];
-               if (j == action->queues_n)
-                       j = 0;
-       }
-       rte_flow->mark = action->mark;
+       rte_flow->mark = flow->actions.mark;
        rte_flow->ibv_attr = flow->ibv_attr;
-       rte_flow->hash_fields = flow->hash_fields;
-       rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
-               priv->ctx,
-               &(struct ibv_exp_rwq_ind_table_init_attr){
-                       .pd = priv->pd,
-                       .log_ind_tbl_size = log2above(action->queues_n),
-                       .ind_tbl = wqs,
-                       .comp_mask = 0,
-               });
-       if (!rte_flow->ind_table) {
+       rte_flow->queues = (uint16_t (*)[])(rte_flow + 1);
+       memcpy(rte_flow->queues, flow->actions.queues,
+              flow->actions.queues_n * sizeof(uint16_t));
+       rte_flow->queues_n = flow->actions.queues_n;
+       rte_flow->frxq.hash_fields = flow->hash_fields;
+       rte_flow->frxq.hrxq = mlx5_priv_hrxq_get(priv, rss_hash_default_key,
+                                                rss_hash_default_key_len,
+                                                flow->hash_fields,
+                                                (*rte_flow->queues),
+                                                rte_flow->queues_n);
+       if (rte_flow->frxq.hrxq) {
                rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
-                                  NULL, "cannot allocate indirection table");
+                                  NULL, "duplicated flow");
                goto error;
        }
-       rte_flow->qp = ibv_exp_create_qp(
-               priv->ctx,
-               &(struct ibv_exp_qp_init_attr){
-                       .qp_type = IBV_QPT_RAW_PACKET,
-                       .comp_mask =
-                               IBV_EXP_QP_INIT_ATTR_PD |
-                               IBV_EXP_QP_INIT_ATTR_PORT |
-                               IBV_EXP_QP_INIT_ATTR_RX_HASH,
-                       .pd = priv->pd,
-                       .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
-                               .rx_hash_function =
-                                       IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
-                               .rx_hash_key_len = rss_hash_default_key_len,
-                               .rx_hash_key = rss_hash_default_key,
-                               .rx_hash_fields_mask = rte_flow->hash_fields,
-                               .rwq_ind_tbl = rte_flow->ind_table,
-                       },
-                       .port_num = priv->port,
-               });
-       if (!rte_flow->qp) {
+       rte_flow->frxq.hrxq = mlx5_priv_hrxq_new(priv, rss_hash_default_key,
+                                                rss_hash_default_key_len,
+                                                flow->hash_fields,
+                                                (*rte_flow->queues),
+                                                rte_flow->queues_n);
+       if (!rte_flow->frxq.hrxq) {
                rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
-                                  NULL, "cannot allocate QP");
+                                  NULL, "cannot create hash rxq");
                goto error;
        }
-       if (!priv->started)
+       for (i = 0; i != flow->actions.queues_n; ++i) {
+               struct mlx5_rxq_data *q =
+                       (*priv->rxqs)[flow->actions.queues[i]];
+
+               q->mark |= flow->actions.mark;
+       }
+       if (!priv->dev->data->dev_started)
                return rte_flow;
-       rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
-                                                rte_flow->ibv_attr);
+       rte_flow->ibv_flow = ibv_create_flow(rte_flow->frxq.hrxq->qp,
+                                            rte_flow->ibv_attr);
        if (!rte_flow->ibv_flow) {
                rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
                                   NULL, "flow rule creation failure");
@@ -1134,10 +1161,8 @@ priv_flow_create_action_queue(struct priv *priv,
        return rte_flow;
 error:
        assert(rte_flow);
-       if (rte_flow->qp)
-               ibv_destroy_qp(rte_flow->qp);
-       if (rte_flow->ind_table)
-               ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
+       if (rte_flow->frxq.hrxq)
+               mlx5_priv_hrxq_release(priv, rte_flow->frxq.hrxq);
        rte_free(rte_flow);
        return NULL;
 }
@@ -1147,6 +1172,8 @@ error:
  *
  * @param priv
  *   Pointer to private structure.
+ * @param list
+ *   Pointer to a TAILQ flow list.
  * @param[in] attr
  *   Flow rule attributes.
  * @param[in] pattern
@@ -1161,58 +1188,60 @@ error:
  */
 static struct rte_flow *
 priv_flow_create(struct priv *priv,
+                struct mlx5_flows *list,
                 const struct rte_flow_attr *attr,
                 const struct rte_flow_item items[],
                 const struct rte_flow_action actions[],
                 struct rte_flow_error *error)
 {
        struct rte_flow *rte_flow;
-       struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr), };
-       struct mlx5_flow_action action = {
-               .queue = 0,
-               .drop = 0,
-               .mark = 0,
-               .mark_id = MLX5_FLOW_MARK_DEFAULT,
-               .queues_n = 0,
+       struct mlx5_flow_parse flow = {
+               .offset = sizeof(struct ibv_flow_attr),
+               .actions = {
+                       .mark_id = MLX5_FLOW_MARK_DEFAULT,
+                       .queues = { 0 },
+                       .queues_n = 0,
+               },
        };
        int err;
 
-       err = priv_flow_validate(priv, attr, items, actions, error, &flow,
-                                &action);
+       err = priv_flow_validate(priv, attr, items, actions, error, &flow);
        if (err)
                goto exit;
        flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
-       flow.offset = sizeof(struct ibv_exp_flow_attr);
+       flow.offset = sizeof(struct ibv_flow_attr);
        if (!flow.ibv_attr) {
                rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
                                   NULL, "cannot allocate ibv_attr memory");
                goto exit;
        }
-       *flow.ibv_attr = (struct ibv_exp_flow_attr){
-               .type = IBV_EXP_FLOW_ATTR_NORMAL,
-               .size = sizeof(struct ibv_exp_flow_attr),
+       *flow.ibv_attr = (struct ibv_flow_attr){
+               .type = IBV_FLOW_ATTR_NORMAL,
+               .size = sizeof(struct ibv_flow_attr),
                .priority = attr->priority,
                .num_of_specs = 0,
                .port = 0,
                .flags = 0,
-               .reserved = 0,
        };
        flow.inner = 0;
        flow.hash_fields = 0;
        claim_zero(priv_flow_validate(priv, attr, items, actions,
-                                     error, &flow, &action));
-       if (action.mark && !action.drop) {
-               mlx5_flow_create_flag_mark(&flow, action.mark_id);
-               flow.offset += sizeof(struct ibv_exp_flow_spec_action_tag);
+                                     error, &flow));
+       if (flow.actions.mark && !flow.actions.drop) {
+               mlx5_flow_create_flag_mark(&flow, flow.actions.mark_id);
+               flow.offset += sizeof(struct ibv_flow_spec_action_tag);
        }
-       if (action.drop)
+       if (flow.actions.drop)
                rte_flow =
                        priv_flow_create_action_queue_drop(priv, &flow, error);
        else
-               rte_flow = priv_flow_create_action_queue(priv, &flow, &action,
-                                                        error);
+               rte_flow = priv_flow_create_action_queue(priv, &flow, error);
        if (!rte_flow)
                goto exit;
+       if (rte_flow) {
+               TAILQ_INSERT_TAIL(list, rte_flow, next);
+               DEBUG("Flow created %p", (void *)rte_flow);
+       }
        return rte_flow;
 exit:
        rte_free(flow.ibv_attr);
@@ -1236,11 +1265,8 @@ mlx5_flow_create(struct rte_eth_dev *dev,
        struct rte_flow *flow;
 
        priv_lock(priv);
-       flow = priv_flow_create(priv, attr, items, actions, error);
-       if (flow) {
-               TAILQ_INSERT_TAIL(&priv->flows, flow, next);
-               DEBUG("Flow created %p", (void *)flow);
-       }
+       flow = priv_flow_create(priv, &priv->flows, attr, items, actions,
+                               error);
        priv_unlock(priv);
        return flow;
 }
@@ -1250,55 +1276,54 @@ mlx5_flow_create(struct rte_eth_dev *dev,
  *
  * @param priv
  *   Pointer to private structure.
+ * @param list
+ *   Pointer to a TAILQ flow list.
  * @param[in] flow
  *   Flow to destroy.
  */
 static void
 priv_flow_destroy(struct priv *priv,
+                 struct mlx5_flows *list,
                  struct rte_flow *flow)
 {
-       TAILQ_REMOVE(&priv->flows, flow, next);
-       if (flow->ibv_flow)
-               claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
-       if (flow->drop)
+       unsigned int i;
+       uint16_t *queues;
+       uint16_t queues_n;
+
+       if (flow->drop || !flow->mark)
                goto free;
-       if (flow->qp)
-               claim_zero(ibv_destroy_qp(flow->qp));
-       if (flow->ind_table)
-               claim_zero(ibv_exp_destroy_rwq_ind_table(flow->ind_table));
-       if (flow->mark) {
+       queues = flow->frxq.hrxq->ind_table->queues;
+       queues_n = flow->frxq.hrxq->ind_table->queues_n;
+       for (i = 0; i != queues_n; ++i) {
                struct rte_flow *tmp;
-               struct rxq *rxq;
-               uint32_t mark_n = 0;
-               uint32_t queue_n;
+               struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[queues[i]];
+               int mark = 0;
 
                /*
                 * To remove the mark from the queue, the queue must not be
                 * present in any other marked flow (RSS or not).
                 */
-               for (queue_n = 0; queue_n < flow->rxqs_n; ++queue_n) {
-                       rxq = flow->rxqs[queue_n];
-                       for (tmp = TAILQ_FIRST(&priv->flows);
-                            tmp;
-                            tmp = TAILQ_NEXT(tmp, next)) {
-                               uint32_t tqueue_n;
-
-                               if (tmp->drop)
-                                       continue;
-                               for (tqueue_n = 0;
-                                    tqueue_n < tmp->rxqs_n;
-                                    ++tqueue_n) {
-                                       struct rxq *trxq;
-
-                                       trxq = tmp->rxqs[tqueue_n];
-                                       if (rxq == trxq)
-                                               ++mark_n;
-                               }
-                       }
-                       rxq->mark = !!mark_n;
+               TAILQ_FOREACH(tmp, list, next) {
+                       unsigned int j;
+
+                       if (!tmp->mark)
+                               continue;
+                       for (j = 0;
+                            (j != tmp->frxq.hrxq->ind_table->queues_n) &&
+                            !mark;
+                            j++)
+                               if (tmp->frxq.hrxq->ind_table->queues[j] ==
+                                   queues[i])
+                                       mark = 1;
                }
+               rxq_data->mark = mark;
        }
 free:
+       if (flow->ibv_flow)
+               claim_zero(ibv_destroy_flow(flow->ibv_flow));
+       if (!flow->drop)
+               mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
+       TAILQ_REMOVE(list, flow, next);
        rte_free(flow->ibv_attr);
        DEBUG("Flow destroyed %p", (void *)flow);
        rte_free(flow);
@@ -1319,7 +1344,7 @@ mlx5_flow_destroy(struct rte_eth_dev *dev,
 
        (void)error;
        priv_lock(priv);
-       priv_flow_destroy(priv, flow);
+       priv_flow_destroy(priv, &priv->flows, flow);
        priv_unlock(priv);
        return 0;
 }
@@ -1329,15 +1354,17 @@ mlx5_flow_destroy(struct rte_eth_dev *dev,
  *
  * @param priv
  *   Pointer to private structure.
+ * @param list
+ *   Pointer to a TAILQ flow list.
  */
-static void
-priv_flow_flush(struct priv *priv)
+void
+priv_flow_flush(struct priv *priv, struct mlx5_flows *list)
 {
-       while (!TAILQ_EMPTY(&priv->flows)) {
+       while (!TAILQ_EMPTY(list)) {
                struct rte_flow *flow;
 
-               flow = TAILQ_FIRST(&priv->flows);
-               priv_flow_destroy(priv, flow);
+               flow = TAILQ_FIRST(list);
+               priv_flow_destroy(priv, list, flow);
        }
 }
 
@@ -1355,7 +1382,7 @@ mlx5_flow_flush(struct rte_eth_dev *dev,
 
        (void)error;
        priv_lock(priv);
-       priv_flow_flush(priv);
+       priv_flow_flush(priv, &priv->flows);
        priv_unlock(priv);
        return 0;
 }
@@ -1372,7 +1399,7 @@ mlx5_flow_flush(struct rte_eth_dev *dev,
 static int
 priv_flow_create_drop_queue(struct priv *priv)
 {
-       struct rte_flow_drop *fdq = NULL;
+       struct mlx5_hrxq_drop *fdq = NULL;
 
        assert(priv->pd);
        assert(priv->ctx);
@@ -1381,19 +1408,16 @@ priv_flow_create_drop_queue(struct priv *priv)
                WARN("cannot allocate memory for drop queue");
                goto error;
        }
-       fdq->cq = ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
-                       &(struct ibv_exp_cq_init_attr){
-                       .comp_mask = 0,
-                       });
+       fdq->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
        if (!fdq->cq) {
                WARN("cannot allocate CQ for drop queue");
                goto error;
        }
-       fdq->wq = ibv_exp_create_wq(priv->ctx,
-                       &(struct ibv_exp_wq_init_attr){
-                       .wq_type = IBV_EXP_WQT_RQ,
-                       .max_recv_wr = 1,
-                       .max_recv_sge = 1,
+       fdq->wq = ibv_create_wq(priv->ctx,
+                       &(struct ibv_wq_init_attr){
+                       .wq_type = IBV_WQT_RQ,
+                       .max_wr = 1,
+                       .max_sge = 1,
                        .pd = priv->pd,
                        .cq = fdq->cq,
                        });
@@ -1401,9 +1425,8 @@ priv_flow_create_drop_queue(struct priv *priv)
                WARN("cannot allocate WQ for drop queue");
                goto error;
        }
-       fdq->ind_table = ibv_exp_create_rwq_ind_table(priv->ctx,
-                       &(struct ibv_exp_rwq_ind_table_init_attr){
-                       .pd = priv->pd,
+       fdq->ind_table = ibv_create_rwq_ind_table(priv->ctx,
+                       &(struct ibv_rwq_ind_table_init_attr){
                        .log_ind_tbl_size = 0,
                        .ind_tbl = &fdq->wq,
                        .comp_mask = 0,
@@ -1412,24 +1435,23 @@ priv_flow_create_drop_queue(struct priv *priv)
                WARN("cannot allocate indirection table for drop queue");
                goto error;
        }
-       fdq->qp = ibv_exp_create_qp(priv->ctx,
-               &(struct ibv_exp_qp_init_attr){
+       fdq->qp = ibv_create_qp_ex(priv->ctx,
+               &(struct ibv_qp_init_attr_ex){
                        .qp_type = IBV_QPT_RAW_PACKET,
                        .comp_mask =
-                               IBV_EXP_QP_INIT_ATTR_PD |
-                               IBV_EXP_QP_INIT_ATTR_PORT |
-                               IBV_EXP_QP_INIT_ATTR_RX_HASH,
-                       .pd = priv->pd,
-                       .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
+                               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_EXP_RX_HASH_FUNC_TOEPLITZ,
+                                       IBV_RX_HASH_FUNC_TOEPLITZ,
                                .rx_hash_key_len = rss_hash_default_key_len,
                                .rx_hash_key = rss_hash_default_key,
                                .rx_hash_fields_mask = 0,
-                               .rwq_ind_tbl = fdq->ind_table,
                                },
-                       .port_num = priv->port,
-                       });
+                       .rwq_ind_tbl = fdq->ind_table,
+                       .pd = priv->pd
+               });
        if (!fdq->qp) {
                WARN("cannot allocate QP for drop queue");
                goto error;
@@ -1440,9 +1462,9 @@ error:
        if (fdq->qp)
                claim_zero(ibv_destroy_qp(fdq->qp));
        if (fdq->ind_table)
-               claim_zero(ibv_exp_destroy_rwq_ind_table(fdq->ind_table));
+               claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
        if (fdq->wq)
-               claim_zero(ibv_exp_destroy_wq(fdq->wq));
+               claim_zero(ibv_destroy_wq(fdq->wq));
        if (fdq->cq)
                claim_zero(ibv_destroy_cq(fdq->cq));
        if (fdq)
@@ -1460,16 +1482,16 @@ error:
 static void
 priv_flow_delete_drop_queue(struct priv *priv)
 {
-       struct rte_flow_drop *fdq = priv->flow_drop_queue;
+       struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
 
        if (!fdq)
                return;
        if (fdq->qp)
                claim_zero(ibv_destroy_qp(fdq->qp));
        if (fdq->ind_table)
-               claim_zero(ibv_exp_destroy_rwq_ind_table(fdq->ind_table));
+               claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
        if (fdq->wq)
-               claim_zero(ibv_exp_destroy_wq(fdq->wq));
+               claim_zero(ibv_destroy_wq(fdq->wq));
        if (fdq->cq)
                claim_zero(ibv_destroy_cq(fdq->cq));
        rte_free(fdq);
@@ -1483,20 +1505,26 @@ priv_flow_delete_drop_queue(struct priv *priv)
  *
  * @param priv
  *   Pointer to private structure.
+ * @param list
+ *   Pointer to a TAILQ flow list.
  */
 void
-priv_flow_stop(struct priv *priv)
+priv_flow_stop(struct priv *priv, struct mlx5_flows *list)
 {
        struct rte_flow *flow;
 
-       TAILQ_FOREACH_REVERSE(flow, &priv->flows, mlx5_flows, next) {
-               claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
+       TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
+               claim_zero(ibv_destroy_flow(flow->ibv_flow));
                flow->ibv_flow = NULL;
+               mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
+               flow->frxq.hrxq = NULL;
                if (flow->mark) {
                        unsigned int n;
+                       struct mlx5_ind_table_ibv *ind_tbl =
+                               flow->frxq.hrxq->ind_table;
 
-                       for (n = 0; n < flow->rxqs_n; ++n)
-                               flow->rxqs[n]->mark = 0;
+                       for (n = 0; n < ind_tbl->queues_n; ++n)
+                               (*priv->rxqs)[ind_tbl->queues[n]]->mark = 0;
                }
                DEBUG("Flow %p removed", (void *)flow);
        }
@@ -1508,12 +1536,14 @@ priv_flow_stop(struct priv *priv)
  *
  * @param priv
  *   Pointer to private structure.
+ * @param list
+ *   Pointer to a TAILQ flow list.
  *
  * @return
  *   0 on success, a errno value otherwise and rte_errno is set.
  */
 int
-priv_flow_start(struct priv *priv)
+priv_flow_start(struct priv *priv, struct mlx5_flows *list)
 {
        int ret;
        struct rte_flow *flow;
@@ -1521,14 +1551,32 @@ priv_flow_start(struct priv *priv)
        ret = priv_flow_create_drop_queue(priv);
        if (ret)
                return -1;
-       TAILQ_FOREACH(flow, &priv->flows, next) {
-               struct ibv_qp *qp;
-
-               if (flow->drop)
-                       qp = priv->flow_drop_queue->qp;
-               else
-                       qp = flow->qp;
-               flow->ibv_flow = ibv_exp_create_flow(qp, flow->ibv_attr);
+       TAILQ_FOREACH(flow, list, next) {
+               if (flow->frxq.hrxq)
+                       goto flow_create;
+               flow->frxq.hrxq =
+                       mlx5_priv_hrxq_get(priv, rss_hash_default_key,
+                                          rss_hash_default_key_len,
+                                          flow->frxq.hash_fields,
+                                          (*flow->queues),
+                                          flow->queues_n);
+               if (flow->frxq.hrxq)
+                       goto flow_create;
+               flow->frxq.hrxq =
+                       mlx5_priv_hrxq_new(priv, rss_hash_default_key,
+                                          rss_hash_default_key_len,
+                                          flow->frxq.hash_fields,
+                                          (*flow->queues),
+                                          flow->queues_n);
+               if (!flow->frxq.hrxq) {
+                       DEBUG("Flow %p cannot be applied",
+                             (void *)flow);
+                       rte_errno = EINVAL;
+                       return rte_errno;
+               }
+flow_create:
+               flow->ibv_flow = ibv_create_flow(flow->frxq.hrxq->qp,
+                                                flow->ibv_attr);
                if (!flow->ibv_flow) {
                        DEBUG("Flow %p cannot be applied", (void *)flow);
                        rte_errno = EINVAL;
@@ -1538,39 +1586,13 @@ priv_flow_start(struct priv *priv)
                if (flow->mark) {
                        unsigned int n;
 
-                       for (n = 0; n < flow->rxqs_n; ++n)
-                               flow->rxqs[n]->mark = 1;
-               }
-       }
-       return 0;
-}
-
-/**
- * Verify if the Rx queue is used in a flow.
- *
- * @param priv
- *   Pointer to private structure.
- * @param rxq
- *   Pointer to the queue to search.
- *
- * @return
- *   Nonzero if the queue is used by a flow.
- */
-int
-priv_flow_rxq_in_use(struct priv *priv, struct rxq *rxq)
-{
-       struct rte_flow *flow;
-
-       for (flow = TAILQ_FIRST(&priv->flows);
-            flow;
-            flow = TAILQ_NEXT(flow, next)) {
-               unsigned int n;
-
-               if (flow->drop)
-                       continue;
-               for (n = 0; n < flow->rxqs_n; ++n) {
-                       if (flow->rxqs[n] == rxq)
-                               return 1;
+                       for (n = 0;
+                            n < flow->frxq.hrxq->ind_table->queues_n;
+                            ++n) {
+                               uint16_t idx =
+                                       flow->frxq.hrxq->ind_table->queues[n];
+                               (*priv->rxqs)[idx]->mark = 1;
+                       }
                }
        }
        return 0;
@@ -1590,7 +1612,7 @@ mlx5_flow_isolate(struct rte_eth_dev *dev,
        struct priv *priv = dev->data->dev_private;
 
        priv_lock(priv);
-       if (priv->started) {
+       if (dev->data->dev_started) {
                rte_flow_error_set(error, EBUSY,
                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                   NULL,
@@ -1602,3 +1624,112 @@ mlx5_flow_isolate(struct rte_eth_dev *dev,
        priv_unlock(priv);
        return 0;
 }
+
+/**
+ * Verify the flow list is empty
+ *
+ * @param priv
+ *  Pointer to private structure.
+ *
+ * @return the number of flows not released.
+ */
+int
+priv_flow_verify(struct priv *priv)
+{
+       struct rte_flow *flow;
+       int ret = 0;
+
+       TAILQ_FOREACH(flow, &priv->flows, next) {
+               DEBUG("%p: flow %p still referenced", (void *)priv,
+                     (void *)flow);
+               ++ret;
+       }
+       return ret;
+}
+
+/**
+ * Enable/disable a control flow configured from the control plane.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param spec
+ *   An Ethernet flow spec to apply.
+ * @param mask
+ *   An Ethernet flow mask to apply.
+ * @param enable
+ *   Enable/disable the flow.
+ *
+ * @return
+ *   0 on success.
+ */
+int
+mlx5_ctrl_flow(struct rte_eth_dev *dev,
+              struct rte_flow_item_eth *spec,
+              struct rte_flow_item_eth *mask,
+              unsigned int enable)
+{
+       struct priv *priv = dev->data->dev_private;
+       const struct rte_flow_attr attr = {
+               .ingress = 1,
+               .priority = MLX5_CTRL_FLOW_PRIORITY,
+       };
+       struct rte_flow_item items[] = {
+               {
+                       .type = RTE_FLOW_ITEM_TYPE_ETH,
+                       .spec = spec,
+                       .last = NULL,
+                       .mask = mask,
+               },
+               {
+                       .type = RTE_FLOW_ITEM_TYPE_END,
+               },
+       };
+       struct rte_flow_action actions[] = {
+               {
+                       .type = RTE_FLOW_ACTION_TYPE_QUEUE,
+                       .conf = &(struct rte_flow_action_queue){
+                               .index = 0,
+                       },
+               },
+               {
+                       .type = RTE_FLOW_ACTION_TYPE_END,
+               },
+       };
+       struct rte_flow *flow;
+       struct rte_flow_error error;
+
+       if (enable) {
+               flow = priv_flow_create(priv, &priv->ctrl_flows, &attr, items,
+                                       actions, &error);
+               if (!flow)
+                       return 1;
+       } else {
+               struct spec {
+                       struct ibv_flow_attr ibv_attr;
+                       struct ibv_flow_spec_eth eth;
+               } spec;
+               struct mlx5_flow_parse parser = {
+                       .ibv_attr = &spec.ibv_attr,
+                       .offset = sizeof(struct ibv_flow_attr),
+               };
+               struct ibv_flow_spec_eth *eth;
+               const unsigned int attr_size = sizeof(struct ibv_flow_attr);
+
+               claim_zero(mlx5_flow_create_eth(&items[0], NULL, &parser));
+               TAILQ_FOREACH(flow, &priv->ctrl_flows, next) {
+                       eth = (void *)((uintptr_t)flow->ibv_attr + attr_size);
+                       assert(eth->type == IBV_FLOW_SPEC_ETH);
+                       if (!memcmp(eth, &spec.eth, sizeof(*eth)))
+                               break;
+               }
+               if (flow) {
+                       claim_zero(ibv_destroy_flow(flow->ibv_flow));
+                       mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
+                       rte_free(flow->ibv_attr);
+                       DEBUG("Control flow destroyed %p", (void *)flow);
+                       TAILQ_REMOVE(&priv->ctrl_flows, flow, next);
+                       rte_free(flow);
+               }
+       }
+       return 0;
+}