net/mlx5: use an RSS drop queue
[dpdk.git] / drivers / net / mlx5 / mlx5_flow.c
index a885f6e..e789c4c 100644 (file)
@@ -52,6 +52,9 @@
 #include "mlx5.h"
 #include "mlx5_prm.h"
 
+/* Number of Work Queue necessary for the DROP queue. */
+#define MLX5_DROP_WQ_N 4
+
 static int
 mlx5_flow_create_eth(const struct rte_flow_item *item,
                     const void *default_mask,
@@ -95,8 +98,11 @@ struct rte_flow {
        struct ibv_exp_flow *ibv_flow; /**< Verbs flow. */
        struct ibv_exp_wq *wq; /**< Verbs work queue. */
        struct ibv_cq *cq; /**< Verbs completion queue. */
-       struct rxq *rxq; /**< Pointer to the queue, NULL if drop queue. */
+       struct rxq *(*rxqs)[]; /**< Pointer to the queues array. */
+       uint16_t rxqs_n; /**< Number of queues in this flow, 0 if drop queue. */
        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. */
 };
 
 /** Static initializer for items. */
@@ -146,6 +152,7 @@ static const enum rte_flow_action_type valid_actions[] = {
        RTE_FLOW_ACTION_TYPE_DROP,
        RTE_FLOW_ACTION_TYPE_QUEUE,
        RTE_FLOW_ACTION_TYPE_MARK,
+       RTE_FLOW_ACTION_TYPE_FLAG,
        RTE_FLOW_ACTION_TYPE_END,
 };
 
@@ -272,14 +279,24 @@ 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 *wqs[MLX5_DROP_WQ_N]; /**< Verbs work queue. */
+       struct ibv_cq *cq; /**< Verbs completion queue. */
 };
 
 struct mlx5_flow_action {
        uint32_t queue:1; /**< Target is a receive queue. */
        uint32_t drop:1; /**< Target is a drop queue. */
        uint32_t mark:1; /**< Mark is present in the flow. */
-       uint32_t queue_id; /**< Identifier of the queue. */
        uint32_t mark_id; /**< Mark identifier. */
+       uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
+       uint16_t queues_n; /**< Number of entries in queue[]. */
 };
 
 /**
@@ -452,10 +469,56 @@ priv_flow_validate(struct priv *priv,
                        const struct rte_flow_action_queue *queue =
                                (const struct rte_flow_action_queue *)
                                actions->conf;
+                       uint16_t n;
+                       uint16_t found = 0;
 
                        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) {
+                                       found = 1;
+                                       break;
+                               }
+                       }
+                       if (action.queues_n && !found) {
+                               rte_flow_error_set(error, ENOTSUP,
+                                          RTE_FLOW_ERROR_TYPE_ACTION,
+                                          actions,
+                                          "queue action not in RSS queues");
+                               return -rte_errno;
+                       }
                        action.queue = 1;
+                       action.queues_n = 1;
+                       action.queues[0] = queue->index;
+               } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
+                       const struct rte_flow_action_rss *rss =
+                               (const struct rte_flow_action_rss *)
+                               actions->conf;
+                       uint16_t n;
+
+                       if (action.queues_n == 1) {
+                               uint16_t found = 0;
+
+                               assert(action.queues_n);
+                               for (n = 0; n < rss->num; ++n) {
+                                       if (action.queues[0] == rss->queue[n]) {
+                                               found = 1;
+                                               break;
+                                       }
+                               }
+                               if (!found) {
+                                       rte_flow_error_set(error, ENOTSUP,
+                                                  RTE_FLOW_ERROR_TYPE_ACTION,
+                                                  actions,
+                                                  "queue action not in RSS"
+                                                  " queues");
+                                       return -rte_errno;
+                               }
+                       }
+                       action.queue = 1;
+                       for (n = 0; n < rss->num; ++n)
+                               action.queues[n] = rss->queue[n];
+                       action.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 *)
@@ -476,6 +539,8 @@ priv_flow_validate(struct priv *priv,
                                return -rte_errno;
                        }
                        action.mark = 1;
+               } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
+                       action.mark = 1;
                } else {
                        goto exit_action_not_supported;
                }
@@ -545,6 +610,7 @@ mlx5_flow_create_eth(const struct rte_flow_item *item,
 
        ++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,
@@ -624,6 +690,8 @@ mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 
        ++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);
        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,
@@ -676,6 +744,8 @@ mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 
        ++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);
        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,
@@ -725,6 +795,8 @@ mlx5_flow_create_udp(const struct rte_flow_item *item,
 
        ++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);
        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,
@@ -767,6 +839,8 @@ mlx5_flow_create_tcp(const struct rte_flow_item *item,
 
        ++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);
        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,
@@ -857,13 +931,60 @@ mlx5_flow_create_flag_mark(struct mlx5_flow *flow, uint32_t mark_id)
        return 0;
 }
 
+/**
+ * Complete flow rule creation with a drop queue.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param flow
+ *   MLX5 flow attributes (filled by mlx5_flow_validate()).
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ *
+ * @return
+ *   A flow if the rule could be created.
+ */
+static struct rte_flow *
+priv_flow_create_action_queue_drop(struct priv *priv,
+                                  struct mlx5_flow *flow,
+                                  struct rte_flow_error *error)
+{
+       struct rte_flow *rte_flow;
+
+       assert(priv->pd);
+       assert(priv->ctx);
+       rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
+       if (!rte_flow) {
+               rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
+                                  NULL, "cannot allocate flow memory");
+               return NULL;
+       }
+       rte_flow->drop = 1;
+       rte_flow->ibv_attr = flow->ibv_attr;
+       rte_flow->qp = priv->flow_drop_queue->qp;
+       if (!priv->started)
+               return rte_flow;
+       rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->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");
+               goto error;
+       }
+       return rte_flow;
+error:
+       assert(rte_flow);
+       rte_free(rte_flow);
+       return NULL;
+}
+
 /**
  * Complete flow rule creation.
  *
  * @param priv
  *   Pointer to private structure.
- * @param ibv_attr
- *   Verbs flow attributes.
+ * @param flow
+ *   MLX5 flow attributes (filled by mlx5_flow_validate()).
  * @param action
  *   Target action structure.
  * @param[out] error
@@ -874,62 +995,55 @@ mlx5_flow_create_flag_mark(struct mlx5_flow *flow, uint32_t mark_id)
  */
 static struct rte_flow *
 priv_flow_create_action_queue(struct priv *priv,
-                             struct ibv_exp_flow_attr *ibv_attr,
+                             struct mlx5_flow *flow,
                              struct mlx5_flow_action *action,
                              struct rte_flow_error *error)
 {
-       struct rxq_ctrl *rxq;
        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);
-       rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
+       assert(!action->drop);
+       rte_flow = rte_calloc(__func__, 1,
+                             sizeof(*rte_flow) + sizeof(struct rxq *) *
+                             action->queues_n, 0);
        if (!rte_flow) {
                rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
                                   NULL, "cannot allocate flow memory");
                return NULL;
        }
-       if (action->drop) {
-               rte_flow->cq =
-                       ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
-                                         &(struct ibv_exp_cq_init_attr){
-                                                 .comp_mask = 0,
-                                         });
-               if (!rte_flow->cq) {
-                       rte_flow_error_set(error, ENOMEM,
-                                          RTE_FLOW_ERROR_TYPE_HANDLE,
-                                          NULL, "cannot allocate CQ");
-                       goto error;
-               }
-               rte_flow->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,
-                                                .pd = priv->pd,
-                                                .cq = rte_flow->cq,
-                                                });
-               if (!rte_flow->wq) {
-                       rte_flow_error_set(error, ENOMEM,
-                                          RTE_FLOW_ERROR_TYPE_HANDLE,
-                                          NULL, "cannot allocate WQ");
-                       goto error;
-               }
-       } else {
-               rxq = container_of((*priv->rxqs)[action->queue_id],
+       rte_flow->rxqs = (struct rxq *(*)[])((uintptr_t)rte_flow +
+                                            sizeof(struct rxq *) *
+                                            action->queues_n);
+       for (i = 0; i < action->queues_n; ++i) {
+               struct rxq_ctrl *rxq;
+
+               rxq = container_of((*priv->rxqs)[action->queues[i]],
                                   struct rxq_ctrl, rxq);
-               rte_flow->rxq = &rxq->rxq;
+               wqs[i] = rxq->wq;
+               (*rte_flow->rxqs)[i] = &rxq->rxq;
+               ++rte_flow->rxqs_n;
                rxq->rxq.mark |= action->mark;
-               rte_flow->wq = rxq->wq;
+       }
+       /* 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->ibv_attr = ibv_attr;
+       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 = 0,
-                       .ind_tbl = &rte_flow->wq,
+                       .log_ind_tbl_size = log2above(action->queues_n),
+                       .ind_tbl = wqs,
                        .comp_mask = 0,
                });
        if (!rte_flow->ind_table) {
@@ -951,7 +1065,7 @@ priv_flow_create_action_queue(struct priv *priv,
                                        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 = 0,
+                               .rx_hash_fields_mask = rte_flow->hash_fields,
                                .rwq_ind_tbl = rte_flow->ind_table,
                        },
                        .port_num = priv->port,
@@ -977,10 +1091,6 @@ error:
                ibv_destroy_qp(rte_flow->qp);
        if (rte_flow->ind_table)
                ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
-       if (!rte_flow->rxq && rte_flow->wq)
-               ibv_exp_destroy_wq(rte_flow->wq);
-       if (!rte_flow->rxq && rte_flow->cq)
-               ibv_destroy_cq(rte_flow->cq);
        rte_free(rte_flow);
        return NULL;
 }
@@ -1034,6 +1144,7 @@ priv_flow_create(struct priv *priv,
                .reserved = 0,
        };
        flow.inner = 0;
+       flow.hash_fields = 0;
        claim_zero(priv_flow_validate(priv, attr, items, actions,
                                      error, &flow));
        action = (struct mlx5_flow_action){
@@ -1047,9 +1158,19 @@ priv_flow_create(struct priv *priv,
                        continue;
                } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
                        action.queue = 1;
-                       action.queue_id =
+                       action.queues[action.queues_n++] =
                                ((const struct rte_flow_action_queue *)
                                 actions->conf)->index;
+               } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
+                       const struct rte_flow_action_rss *rss =
+                               (const struct rte_flow_action_rss *)
+                                actions->conf;
+                       uint16_t n;
+
+                       action.queue = 1;
+                       action.queues_n = rss->num;
+                       for (n = 0; n < rss->num; ++n)
+                               action.queues[n] = rss->queue[n];
                } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
                        action.drop = 1;
                        action.mark = 0;
@@ -1061,6 +1182,8 @@ priv_flow_create(struct priv *priv,
                        if (mark)
                                action.mark_id = mark->id;
                        action.mark = !action.drop;
+               } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
+                       action.mark = 1;
                } else {
                        rte_flow_error_set(error, ENOTSUP,
                                           RTE_FLOW_ERROR_TYPE_ACTION,
@@ -1072,8 +1195,12 @@ priv_flow_create(struct priv *priv,
                mlx5_flow_create_flag_mark(&flow, action.mark_id);
                flow.offset += sizeof(struct ibv_exp_flow_spec_action_tag);
        }
-       rte_flow = priv_flow_create_action_queue(priv, flow.ibv_attr,
-                                                &action, error);
+       if (action.drop)
+               rte_flow =
+                       priv_flow_create_action_queue_drop(priv, &flow, error);
+       else
+               rte_flow = priv_flow_create_action_queue(priv, &flow, &action,
+                                                        error);
        if (!rte_flow)
                goto exit;
        return rte_flow;
@@ -1124,26 +1251,49 @@ priv_flow_destroy(struct priv *priv,
        LIST_REMOVE(flow, next);
        if (flow->ibv_flow)
                claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
+       if (flow->drop)
+               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->rxq && flow->wq)
+       if (flow->drop && flow->wq)
                claim_zero(ibv_exp_destroy_wq(flow->wq));
-       if (!flow->rxq && flow->cq)
+       if (flow->drop && flow->cq)
                claim_zero(ibv_destroy_cq(flow->cq));
        if (flow->mark) {
                struct rte_flow *tmp;
+               struct rxq *rxq;
                uint32_t mark_n = 0;
-
-               for (tmp = LIST_FIRST(&priv->flows);
-                    tmp;
-                    tmp = LIST_NEXT(tmp, next)) {
-                       if ((flow->rxq == tmp->rxq) && tmp->mark)
-                               ++mark_n;
+               uint32_t queue_n;
+
+               /*
+                * 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 = LIST_FIRST(&priv->flows);
+                            tmp;
+                            tmp = LIST_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;
                }
-               flow->rxq->mark = !!mark_n;
        }
+free:
        rte_free(flow->ibv_attr);
        DEBUG("Flow destroyed %p", (void *)flow);
        rte_free(flow);
@@ -1205,6 +1355,125 @@ mlx5_flow_flush(struct rte_eth_dev *dev,
        return 0;
 }
 
+/**
+ * Create drop queue.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   0 on success.
+ */
+static int
+priv_flow_create_drop_queue(struct priv *priv)
+{
+       struct rte_flow_drop *fdq = NULL;
+       unsigned int i;
+
+       assert(priv->pd);
+       assert(priv->ctx);
+       fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
+       if (!fdq) {
+               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,
+                       });
+       if (!fdq->cq) {
+               WARN("cannot allocate CQ for drop queue");
+               goto error;
+       }
+       for (i = 0; i != MLX5_DROP_WQ_N; ++i) {
+               fdq->wqs[i] = 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,
+                               .pd = priv->pd,
+                               .cq = fdq->cq,
+                               });
+               if (!fdq->wqs[i]) {
+                       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,
+                       .log_ind_tbl_size = 0,
+                       .ind_tbl = fdq->wqs,
+                       .comp_mask = 0,
+                       });
+       if (!fdq->ind_table) {
+               WARN("cannot allocate indirection table for drop queue");
+               goto error;
+       }
+       fdq->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 = 0,
+                               .rwq_ind_tbl = fdq->ind_table,
+                               },
+                       .port_num = priv->port,
+                       });
+       if (!fdq->qp) {
+               WARN("cannot allocate QP for drop queue");
+               goto error;
+       }
+       priv->flow_drop_queue = fdq;
+       return 0;
+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));
+       for (i = 0; i != MLX5_DROP_WQ_N; ++i) {
+               if (fdq->wqs[i])
+                       claim_zero(ibv_exp_destroy_wq(fdq->wqs[i]));
+       }
+       if (fdq->cq)
+               claim_zero(ibv_destroy_cq(fdq->cq));
+       if (fdq)
+               rte_free(fdq);
+       priv->flow_drop_queue = NULL;
+       return -1;
+}
+
+/**
+ * Delete drop queue.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+static void
+priv_flow_delete_drop_queue(struct priv *priv)
+{
+       struct rte_flow_drop *fdq = priv->flow_drop_queue;
+       unsigned int i;
+
+       claim_zero(ibv_destroy_qp(fdq->qp));
+       claim_zero(ibv_exp_destroy_rwq_ind_table(fdq->ind_table));
+       for (i = 0; i != MLX5_DROP_WQ_N; ++i) {
+               assert(fdq->wqs[i]);
+               claim_zero(ibv_exp_destroy_wq(fdq->wqs[i]));
+       }
+       claim_zero(ibv_destroy_cq(fdq->cq));
+       rte_free(fdq);
+       priv->flow_drop_queue = NULL;
+}
+
 /**
  * Remove all flows.
  *
@@ -1223,10 +1492,15 @@ priv_flow_stop(struct priv *priv)
             flow = LIST_NEXT(flow, next)) {
                claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
                flow->ibv_flow = NULL;
-               if (flow->mark)
-                       flow->rxq->mark = 0;
+               if (flow->mark) {
+                       unsigned int n;
+
+                       for (n = 0; n < flow->rxqs_n; ++n)
+                               (*flow->rxqs)[n]->mark = 0;
+               }
                DEBUG("Flow %p removed", (void *)flow);
        }
+       priv_flow_delete_drop_queue(priv);
 }
 
 /**
@@ -1241,21 +1515,34 @@ priv_flow_stop(struct priv *priv)
 int
 priv_flow_start(struct priv *priv)
 {
+       int ret;
        struct rte_flow *flow;
 
+       ret = priv_flow_create_drop_queue(priv);
+       if (ret)
+               return -1;
        for (flow = LIST_FIRST(&priv->flows);
             flow;
             flow = LIST_NEXT(flow, next)) {
-               flow->ibv_flow = ibv_exp_create_flow(flow->qp,
-                                                    flow->ibv_attr);
+               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);
                if (!flow->ibv_flow) {
                        DEBUG("Flow %p cannot be applied", (void *)flow);
                        rte_errno = EINVAL;
                        return rte_errno;
                }
                DEBUG("Flow %p applied", (void *)flow);
-               if (flow->rxq)
-                       flow->rxq->mark |= flow->mark;
+               if (flow->mark) {
+                       unsigned int n;
+
+                       for (n = 0; n < flow->rxqs_n; ++n)
+                               (*flow->rxqs)[n]->mark = 1;
+               }
        }
        return 0;
 }