net/mlx5: add RSS flow action
authorNelio Laranjeiro <nelio.laranjeiro@6wind.com>
Thu, 12 Jul 2018 09:31:00 +0000 (11:31 +0200)
committerShahaf Shuler <shahafs@mellanox.com>
Thu, 12 Jul 2018 10:10:04 +0000 (12:10 +0200)
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Acked-by: Yongseok Koh <yskoh@mellanox.com>
drivers/net/mlx5/mlx5_flow.c

index 77483bd..758c611 100644 (file)
@@ -51,6 +51,7 @@ extern const struct eth_dev_ops mlx5_dev_ops_isolate;
 /* Actions that modify the fate of matching traffic. */
 #define MLX5_FLOW_FATE_DROP (1u << 0)
 #define MLX5_FLOW_FATE_QUEUE (1u << 1)
+#define MLX5_FLOW_FATE_RSS (1u << 2)
 
 /* Modify a packet. */
 #define MLX5_FLOW_MOD_FLAG (1u << 0)
@@ -60,8 +61,68 @@ extern const struct eth_dev_ops mlx5_dev_ops_isolate;
 #define MLX5_IP_PROTOCOL_TCP 6
 #define MLX5_IP_PROTOCOL_UDP 17
 
+/* Priority reserved for default flows. */
+#define MLX5_FLOW_PRIO_RSVD ((uint32_t)-1)
+
+enum mlx5_expansion {
+       MLX5_EXPANSION_ROOT,
+       MLX5_EXPANSION_ETH,
+       MLX5_EXPANSION_IPV4,
+       MLX5_EXPANSION_IPV4_UDP,
+       MLX5_EXPANSION_IPV4_TCP,
+       MLX5_EXPANSION_IPV6,
+       MLX5_EXPANSION_IPV6_UDP,
+       MLX5_EXPANSION_IPV6_TCP,
+};
+
+/** Supported expansion of items. */
+static const struct rte_flow_expand_node mlx5_support_expansion[] = {
+       [MLX5_EXPANSION_ROOT] = {
+               .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
+                                                MLX5_EXPANSION_IPV4,
+                                                MLX5_EXPANSION_IPV6),
+               .type = RTE_FLOW_ITEM_TYPE_END,
+       },
+       [MLX5_EXPANSION_ETH] = {
+               .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
+                                                MLX5_EXPANSION_IPV6),
+               .type = RTE_FLOW_ITEM_TYPE_ETH,
+       },
+       [MLX5_EXPANSION_IPV4] = {
+               .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4_UDP,
+                                                MLX5_EXPANSION_IPV4_TCP),
+               .type = RTE_FLOW_ITEM_TYPE_IPV4,
+               .rss_types = ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
+                       ETH_RSS_NONFRAG_IPV4_OTHER,
+       },
+       [MLX5_EXPANSION_IPV4_UDP] = {
+               .type = RTE_FLOW_ITEM_TYPE_UDP,
+               .rss_types = ETH_RSS_NONFRAG_IPV4_UDP,
+       },
+       [MLX5_EXPANSION_IPV4_TCP] = {
+               .type = RTE_FLOW_ITEM_TYPE_TCP,
+               .rss_types = ETH_RSS_NONFRAG_IPV4_TCP,
+       },
+       [MLX5_EXPANSION_IPV6] = {
+               .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV6_UDP,
+                                                MLX5_EXPANSION_IPV6_TCP),
+               .type = RTE_FLOW_ITEM_TYPE_IPV6,
+               .rss_types = ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
+                       ETH_RSS_NONFRAG_IPV6_OTHER,
+       },
+       [MLX5_EXPANSION_IPV6_UDP] = {
+               .type = RTE_FLOW_ITEM_TYPE_UDP,
+               .rss_types = ETH_RSS_NONFRAG_IPV6_UDP,
+       },
+       [MLX5_EXPANSION_IPV6_TCP] = {
+               .type = RTE_FLOW_ITEM_TYPE_TCP,
+               .rss_types = ETH_RSS_NONFRAG_IPV6_TCP,
+       },
+};
+
 /** Handles information leading to a drop fate. */
 struct mlx5_flow_verbs {
+       LIST_ENTRY(mlx5_flow_verbs) next;
        unsigned int size; /**< Size of the attribute. */
        struct {
                struct ibv_flow_attr *attr;
@@ -70,6 +131,7 @@ struct mlx5_flow_verbs {
        };
        struct ibv_flow *flow; /**< Verbs flow pointer. */
        struct mlx5_hrxq *hrxq; /**< Hash Rx queue object. */
+       uint64_t hash_fields; /**< Verbs hash Rx queue hash fields. */
 };
 
 /* Flow structure. */
@@ -84,8 +146,12 @@ struct rte_flow {
        uint32_t fate;
        /**< Bit-fields of present fate see MLX5_FLOW_FATE_*. */
        uint8_t l3_protocol; /**< valid when l3_protocol_en is set. */
-       struct mlx5_flow_verbs verbs; /* Verbs flow. */
-       uint16_t queue; /**< Destination queue to redirect traffic to. */
+       LIST_HEAD(verbs, mlx5_flow_verbs) verbs; /**< Verbs flows list. */
+       struct mlx5_flow_verbs *cur_verbs;
+       /**< Current Verbs flow structure being filled. */
+       struct rte_flow_action_rss rss;/**< RSS context. */
+       uint8_t key[MLX5_RSS_HASH_KEY_LEN]; /**< RSS hash key. */
+       uint16_t (*queue)[]; /**< Destination queues to redirect traffic to. */
 };
 
 static const struct rte_flow_ops mlx5_flow_ops = {
@@ -128,16 +194,38 @@ struct ibv_spec_header {
        uint16_t size;
 };
 
- /**
-  * Discover the maximum number of priority available.
-  *
-  * @param[in] dev
-  *   Pointer to Ethernet device.
-  *
-  * @return
-  *   number of supported flow priority on success, a negative errno value
-  *   otherwise and rte_errno is set.
-  */
+/*
+ * Number of sub priorities.
+ * For each kind of pattern matching i.e. L2, L3, L4 to have a correct
+ * matching on the NIC (firmware dependent) L4 most have the higher priority
+ * followed by L3 and ending with L2.
+ */
+#define MLX5_PRIORITY_MAP_L2 2
+#define MLX5_PRIORITY_MAP_L3 1
+#define MLX5_PRIORITY_MAP_L4 0
+#define MLX5_PRIORITY_MAP_MAX 3
+
+/* Map of Verbs to Flow priority with 8 Verbs priorities. */
+static const uint32_t priority_map_3[][MLX5_PRIORITY_MAP_MAX] = {
+       { 0, 1, 2 }, { 2, 3, 4 }, { 5, 6, 7 },
+};
+
+/* Map of Verbs to Flow priority with 16 Verbs priorities. */
+static const uint32_t priority_map_5[][MLX5_PRIORITY_MAP_MAX] = {
+       { 0, 1, 2 }, { 3, 4, 5 }, { 6, 7, 8 },
+       { 9, 10, 11 }, { 12, 13, 14 },
+};
+
+/**
+ * Discover the maximum number of priority available.
+ *
+ * @param[in] dev
+ *   Pointer to Ethernet device.
+ *
+ * @return
+ *   number of supported flow priority on success, a negative errno
+ *   value otherwise and rte_errno is set.
+ */
 int
 mlx5_flow_discover_priorities(struct rte_eth_dev *dev)
 {
@@ -162,6 +250,7 @@ mlx5_flow_discover_priorities(struct rte_eth_dev *dev)
        struct mlx5_hrxq *drop = mlx5_hrxq_drop_new(dev);
        uint16_t vprio[] = { 8, 16 };
        int i;
+       int priority = 0;
 
        if (!drop) {
                rte_errno = ENOTSUP;
@@ -173,11 +262,52 @@ mlx5_flow_discover_priorities(struct rte_eth_dev *dev)
                if (!flow)
                        break;
                claim_zero(mlx5_glue->destroy_flow(flow));
+               priority = vprio[i];
+       }
+       switch (priority) {
+       case 8:
+               priority = RTE_DIM(priority_map_3);
+               break;
+       case 16:
+               priority = RTE_DIM(priority_map_5);
+               break;
+       default:
+               rte_errno = ENOTSUP;
+               DRV_LOG(ERR,
+                       "port %u verbs maximum priority: %d expected 8/16",
+                       dev->data->port_id, vprio[i]);
+               return -rte_errno;
        }
        mlx5_hrxq_drop_release(dev);
        DRV_LOG(INFO, "port %u flow maximum priority: %d",
-               dev->data->port_id, vprio[i - 1]);
-       return vprio[i - 1];
+               dev->data->port_id, priority);
+       return priority;
+}
+
+/**
+ * Adjust flow priority.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param flow
+ *   Pointer to an rte flow.
+ */
+static void
+mlx5_flow_adjust_priority(struct rte_eth_dev *dev, struct rte_flow *flow)
+{
+       struct priv *priv = dev->data->dev_private;
+       uint32_t priority = flow->attributes.priority;
+       uint32_t subpriority = flow->cur_verbs->attr->priority;
+
+       switch (priv->config.flow_prio) {
+       case RTE_DIM(priority_map_3):
+               priority = priority_map_3[priority][subpriority];
+               break;
+       case RTE_DIM(priority_map_5):
+               priority = priority_map_5[priority][subpriority];
+               break;
+       }
+       flow->cur_verbs->attr->priority = priority;
 }
 
 /**
@@ -203,14 +333,15 @@ mlx5_flow_attributes(struct rte_eth_dev *dev,
                     struct rte_flow_error *error)
 {
        uint32_t priority_max =
-               ((struct priv *)dev->data->dev_private)->config.flow_prio;
+               ((struct priv *)dev->data->dev_private)->config.flow_prio - 1;
 
        if (attributes->group)
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
                                          NULL,
                                          "groups is not supported");
-       if (attributes->priority >= priority_max)
+       if (attributes->priority != MLX5_FLOW_PRIO_RSVD &&
+           attributes->priority >= priority_max)
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
                                          NULL,
@@ -231,6 +362,8 @@ mlx5_flow_attributes(struct rte_eth_dev *dev,
                                          NULL,
                                          "ingress attribute is mandatory");
        flow->attributes = *attributes;
+       if (attributes->priority == MLX5_FLOW_PRIO_RSVD)
+               flow->attributes.priority = priority_max;
        return 0;
 }
 
@@ -296,7 +429,7 @@ mlx5_flow_item_acceptable(const struct rte_flow_item *item,
 }
 
 /**
- * Add a verbs specification into @p flow.
+ * Add a verbs item specification into @p flow.
  *
  * @param[in, out] flow
  *   Pointer to flow structure.
@@ -308,14 +441,16 @@ mlx5_flow_item_acceptable(const struct rte_flow_item *item,
 static void
 mlx5_flow_spec_verbs_add(struct rte_flow *flow, void *src, unsigned int size)
 {
-       if (flow->verbs.specs) {
+       struct mlx5_flow_verbs *verbs = flow->cur_verbs;
+
+       if (verbs->specs) {
                void *dst;
 
-               dst = (void *)(flow->verbs.specs + flow->verbs.size);
+               dst = (void *)(verbs->specs + verbs->size);
                memcpy(dst, src, size);
-               ++flow->verbs.attr->num_of_specs;
+               ++verbs->attr->num_of_specs;
        }
-       flow->verbs.size += size;
+       verbs->size += size;
 }
 
 /**
@@ -390,6 +525,7 @@ mlx5_flow_item_eth(const struct rte_flow_item *item, struct rte_flow *flow,
                }
                eth.val.ether_type &= eth.mask.ether_type;
        }
+       flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
        mlx5_flow_spec_verbs_add(flow, &eth, size);
        return size;
 }
@@ -460,6 +596,7 @@ mlx5_flow_item_vlan(const struct rte_flow_item *item, struct rte_flow *flow,
                .inner_type = RTE_BE16(0xffff),
        };
        unsigned int size = sizeof(struct ibv_flow_spec_eth);
+       struct mlx5_flow_verbs *verbs = flow->cur_verbs;
        struct ibv_flow_spec_eth eth = {
                .type = IBV_FLOW_SPEC_ETH,
                .size = size,
@@ -506,15 +643,16 @@ mlx5_flow_item_vlan(const struct rte_flow_item *item, struct rte_flow *flow,
                                          item->spec,
                                          "VLAN cannot be empty");
        if (!(flow->layers & l2m)) {
-               if (size <= flow_size)
+               if (size <= flow_size) {
+                       flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
                        mlx5_flow_spec_verbs_add(flow, &eth, size);
+               }
        } else {
-               if (flow->verbs.attr)
-                       mlx5_flow_item_vlan_update(flow->verbs.attr, &eth);
+               if (verbs->attr)
+                       mlx5_flow_item_vlan_update(verbs->attr, &eth);
                size = 0; /* Only an update is done in eth specification. */
        }
-       flow->layers |= MLX5_FLOW_LAYER_OUTER_L2 |
-               MLX5_FLOW_LAYER_OUTER_VLAN;
+       flow->layers |= MLX5_FLOW_LAYER_OUTER_L2 | MLX5_FLOW_LAYER_OUTER_VLAN;
        return size;
 }
 
@@ -601,8 +739,18 @@ mlx5_flow_item_ipv4(const struct rte_flow_item *item, struct rte_flow *flow,
        }
        flow->l3_protocol_en = !!ipv4.mask.proto;
        flow->l3_protocol = ipv4.val.proto;
-       if (size <= flow_size)
+       if (size <= flow_size) {
+               uint64_t hash_fields = IBV_RX_HASH_SRC_IPV4 |
+                       IBV_RX_HASH_DST_IPV4;
+
+               if (!(flow->rss.types &
+                     (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
+                      ETH_RSS_NONFRAG_IPV4_OTHER)))
+                       hash_fields = 0;
+               flow->cur_verbs->hash_fields |= hash_fields;
+               flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L3;
                mlx5_flow_spec_verbs_add(flow, &ipv4, size);
+       }
        return size;
 }
 
@@ -714,8 +862,17 @@ mlx5_flow_item_ipv6(const struct rte_flow_item *item, struct rte_flow *flow,
        }
        flow->l3_protocol_en = !!ipv6.mask.next_hdr;
        flow->l3_protocol = ipv6.val.next_hdr;
-       if (size <= flow_size)
+       if (size <= flow_size) {
+               uint64_t hash_fields = IBV_RX_HASH_SRC_IPV6 |
+                       IBV_RX_HASH_DST_IPV6;
+
+               if (!(flow->rss.types &
+                     (ETH_RSS_IPV6 | ETH_RSS_NONFRAG_IPV6_OTHER)))
+                       hash_fields = 0;
+               flow->cur_verbs->hash_fields |= hash_fields;
+               flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L3;
                mlx5_flow_spec_verbs_add(flow, &ipv6, size);
+       }
        return size;
 }
 
@@ -754,22 +911,24 @@ mlx5_flow_item_udp(const struct rte_flow_item *item, struct rte_flow *flow,
        };
        int ret;
 
-       if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L3))
+       if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_UDP)
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_ITEM,
                                          item,
-                                         "L3 is mandatory to filter on L4");
-       if (flow->layers & MLX5_FLOW_LAYER_OUTER_L4)
+                                         "protocol filtering not compatible"
+                                         " with UDP layer");
+       if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L3))
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_ITEM,
                                          item,
-                                         "L4 layer is already present");
-       if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_UDP)
+                                         "L3 is mandatory to filter"
+                                         " on L4");
+       if (flow->layers & MLX5_FLOW_LAYER_OUTER_L4)
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_ITEM,
                                          item,
-                                         "protocol filtering not compatible"
-                                         " with UDP layer");
+                                         "L4 layer is already"
+                                         " present");
        if (!mask)
                mask = &rte_flow_item_udp_mask;
        ret = mlx5_flow_item_acceptable
@@ -779,8 +938,6 @@ mlx5_flow_item_udp(const struct rte_flow_item *item, struct rte_flow *flow,
        if (ret < 0)
                return ret;
        flow->layers |= MLX5_FLOW_LAYER_OUTER_L4_UDP;
-       if (size > flow_size)
-               return size;
        if (spec) {
                udp.val.dst_port = spec->hdr.dst_port;
                udp.val.src_port = spec->hdr.src_port;
@@ -790,7 +947,16 @@ mlx5_flow_item_udp(const struct rte_flow_item *item, struct rte_flow *flow,
                udp.val.src_port &= udp.mask.src_port;
                udp.val.dst_port &= udp.mask.dst_port;
        }
-       mlx5_flow_spec_verbs_add(flow, &udp, size);
+       if (size <= flow_size) {
+               uint64_t hash_fields = IBV_RX_HASH_SRC_PORT_UDP |
+                       IBV_RX_HASH_DST_PORT_UDP;
+
+               if (!(flow->rss.types & ETH_RSS_UDP))
+                       hash_fields = 0;
+               flow->cur_verbs->hash_fields |= hash_fields;
+               flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L4;
+               mlx5_flow_spec_verbs_add(flow, &udp, size);
+       }
        return size;
 }
 
@@ -854,8 +1020,6 @@ mlx5_flow_item_tcp(const struct rte_flow_item *item, struct rte_flow *flow,
        if (ret < 0)
                return ret;
        flow->layers |= MLX5_FLOW_LAYER_OUTER_L4_TCP;
-       if (size > flow_size)
-               return size;
        if (spec) {
                tcp.val.dst_port = spec->hdr.dst_port;
                tcp.val.src_port = spec->hdr.src_port;
@@ -865,7 +1029,16 @@ mlx5_flow_item_tcp(const struct rte_flow_item *item, struct rte_flow *flow,
                tcp.val.src_port &= tcp.mask.src_port;
                tcp.val.dst_port &= tcp.mask.dst_port;
        }
-       mlx5_flow_spec_verbs_add(flow, &tcp, size);
+       if (size <= flow_size) {
+               uint64_t hash_fields = IBV_RX_HASH_SRC_PORT_TCP |
+                       IBV_RX_HASH_DST_PORT_TCP;
+
+               if (!(flow->rss.types & ETH_RSS_TCP))
+                       hash_fields = 0;
+               flow->cur_verbs->hash_fields |= hash_fields;
+               flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L4;
+               mlx5_flow_spec_verbs_add(flow, &tcp, size);
+       }
        return size;
 }
 
@@ -1043,11 +1216,95 @@ mlx5_flow_action_queue(struct rte_eth_dev *dev,
                                          RTE_FLOW_ERROR_TYPE_ACTION_CONF,
                                          &queue->index,
                                          "queue is not configured");
-       flow->queue = queue->index;
+       if (flow->queue)
+               (*flow->queue)[0] = queue->index;
+       flow->rss.queue_num = 1;
        flow->fate |= MLX5_FLOW_FATE_QUEUE;
        return 0;
 }
 
+/**
+ * Ensure the @p action will be understood and used correctly by the  NIC.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param action[in]
+ *   Pointer to flow actions array.
+ * @param flow[in, out]
+ *   Pointer to the rte_flow structure.
+ * @param error[in, out]
+ *   Pointer to error structure.
+ *
+ * @return
+ *   On success @p flow->queue array and @p flow->rss are filled and valid.
+ *   On error, a negative errno value is returned and rte_errno is set.
+ */
+static int
+mlx5_flow_action_rss(struct rte_eth_dev *dev,
+                    const struct rte_flow_action *action,
+                    struct rte_flow *flow,
+                    struct rte_flow_error *error)
+{
+       struct priv *priv = dev->data->dev_private;
+       const struct rte_flow_action_rss *rss = action->conf;
+       unsigned int i;
+
+       if (flow->fate)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION,
+                                         action,
+                                         "multiple fate actions are not"
+                                         " supported");
+       if (rss->func != RTE_ETH_HASH_FUNCTION_DEFAULT &&
+           rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                                         &rss->func,
+                                         "RSS hash function not supported");
+       if (rss->level > 1)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                                         &rss->level,
+                                         "tunnel RSS is not supported");
+       if (rss->key_len < MLX5_RSS_HASH_KEY_LEN)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                                         &rss->key_len,
+                                         "RSS hash key too small");
+       if (rss->key_len > MLX5_RSS_HASH_KEY_LEN)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                                         &rss->key_len,
+                                         "RSS hash key too large");
+       if (rss->queue_num > priv->config.ind_table_max_size)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                                         &rss->queue_num,
+                                         "number of queues too large");
+       if (rss->types & MLX5_RSS_HF_MASK)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                                         &rss->types,
+                                         "some RSS protocols are not"
+                                         " supported");
+       for (i = 0; i != rss->queue_num; ++i) {
+               if (!(*priv->rxqs)[rss->queue[i]])
+                       return rte_flow_error_set
+                               (error, EINVAL,
+                                RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                                &rss->queue[i],
+                                "queue is not configured");
+       }
+       if (flow->queue)
+               memcpy((*flow->queue), rss->queue,
+                      rss->queue_num * sizeof(uint16_t));
+       flow->rss.queue_num = rss->queue_num;
+       memcpy(flow->key, rss->key, MLX5_RSS_HASH_KEY_LEN);
+       flow->rss.types = rss->types;
+       flow->fate |= MLX5_FLOW_FATE_RSS;
+       return 0;
+}
+
 /**
  * Convert the @p action into a Verbs specification after ensuring the NIC
  * will understand and process it correctly.
@@ -1082,6 +1339,7 @@ mlx5_flow_action_flag(const struct rte_flow_action *action,
                .size = size,
                .tag_id = mlx5_flow_mark_set(MLX5_FLOW_MARK_DEFAULT),
        };
+       struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 
        if (flow->modifier & MLX5_FLOW_MOD_FLAG)
                return rte_flow_error_set(error, ENOTSUP,
@@ -1095,32 +1353,34 @@ mlx5_flow_action_flag(const struct rte_flow_action *action,
                                          "flag is not compatible with drop"
                                          " action");
        if (flow->modifier & MLX5_FLOW_MOD_MARK)
-               return 0;
-       flow->modifier |= MLX5_FLOW_MOD_FLAG;
-       if (size <= flow_size)
+               size = 0;
+       else if (size <= flow_size && verbs)
                mlx5_flow_spec_verbs_add(flow, &tag, size);
+       flow->modifier |= MLX5_FLOW_MOD_FLAG;
        return size;
 }
 
 /**
  * Update verbs specification to modify the flag to mark.
  *
- * @param[in, out] flow
- *   Pointer to the rte_flow structure.
+ * @param[in, out] verbs
+ *   Pointer to the mlx5_flow_verbs structure.
  * @param[in] mark_id
  *   Mark identifier to replace the flag.
  */
 static void
-mlx5_flow_verbs_mark_update(struct rte_flow *flow, uint32_t mark_id)
+mlx5_flow_verbs_mark_update(struct mlx5_flow_verbs *verbs, uint32_t mark_id)
 {
        struct ibv_spec_header *hdr;
        int i;
 
+       if (!verbs)
+               return;
        /* Update Verbs specification. */
-       hdr = (struct ibv_spec_header *)flow->verbs.specs;
+       hdr = (struct ibv_spec_header *)verbs->specs;
        if (!hdr)
                return;
-       for (i = 0; i != flow->verbs.attr->num_of_specs; ++i) {
+       for (i = 0; i != verbs->attr->num_of_specs; ++i) {
                if (hdr->type == IBV_FLOW_SPEC_ACTION_TAG) {
                        struct ibv_flow_spec_action_tag *t =
                                (struct ibv_flow_spec_action_tag *)hdr;
@@ -1166,6 +1426,7 @@ mlx5_flow_action_mark(const struct rte_flow_action *action,
                .type = IBV_FLOW_SPEC_ACTION_TAG,
                .size = size,
        };
+       struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 
        if (!mark)
                return rte_flow_error_set(error, EINVAL,
@@ -1190,14 +1451,11 @@ mlx5_flow_action_mark(const struct rte_flow_action *action,
                                          "mark is not compatible with drop"
                                          " action");
        if (flow->modifier & MLX5_FLOW_MOD_FLAG) {
-               mlx5_flow_verbs_mark_update(flow, mark->id);
-               size = 0; /**< Only an update is done in the specification. */
-       } else {
+               mlx5_flow_verbs_mark_update(verbs, mark->id);
+               size = 0;
+       } else if (size <= flow_size) {
                tag.tag_id = mlx5_flow_mark_set(mark->id);
-               if (size <= flow_size) {
-                       tag.tag_id = mlx5_flow_mark_set(mark->id);
-                       mlx5_flow_spec_verbs_add(flow, &tag, size);
-               }
+               mlx5_flow_spec_verbs_add(flow, &tag, size);
        }
        flow->modifier |= MLX5_FLOW_MOD_MARK;
        return size;
@@ -1259,6 +1517,9 @@ mlx5_flow_actions(struct rte_eth_dev *dev,
                case RTE_FLOW_ACTION_TYPE_QUEUE:
                        ret = mlx5_flow_action_queue(dev, actions, flow, error);
                        break;
+               case RTE_FLOW_ACTION_TYPE_RSS:
+                       ret = mlx5_flow_action_rss(dev, actions, flow, error);
+                       break;
                default:
                        return rte_flow_error_set(error, ENOTSUP,
                                                  RTE_FLOW_ERROR_TYPE_ACTION,
@@ -1322,26 +1583,126 @@ mlx5_flow_merge(struct rte_eth_dev *dev, struct rte_flow *flow,
                struct rte_flow_error *error)
 {
        struct rte_flow local_flow = { .layers = 0, };
-       size_t size = sizeof(*flow) + sizeof(struct ibv_flow_attr);
-       int remain = (flow_size > size) ? flow_size - size : 0;
+       size_t size = sizeof(*flow);
+       union {
+               struct rte_flow_expand_rss buf;
+               uint8_t buffer[2048];
+       } expand_buffer;
+       struct rte_flow_expand_rss *buf = &expand_buffer.buf;
+       struct mlx5_flow_verbs *original_verbs = NULL;
+       size_t original_verbs_size = 0;
+       uint32_t original_layers = 0;
+       int expanded_pattern_idx = 0;
        int ret;
+       uint32_t i;
 
-       if (!remain)
+       if (size > flow_size)
                flow = &local_flow;
        ret = mlx5_flow_attributes(dev, attributes, flow, error);
        if (ret < 0)
                return ret;
-       ret = mlx5_flow_items(pattern, flow, remain, error);
-       if (ret < 0)
-               return ret;
-       size += ret;
-       remain = (flow_size > size) ? flow_size - size : 0;
-       ret = mlx5_flow_actions(dev, actions, flow, remain, error);
+       ret = mlx5_flow_actions(dev, actions, &local_flow, 0, error);
        if (ret < 0)
                return ret;
-       size += ret;
+       if (local_flow.rss.types) {
+               ret = rte_flow_expand_rss(buf, sizeof(expand_buffer.buffer),
+                                         pattern, local_flow.rss.types,
+                                         mlx5_support_expansion,
+                                         MLX5_EXPANSION_ROOT);
+               assert(ret > 0 &&
+                      (unsigned int)ret < sizeof(expand_buffer.buffer));
+       } else {
+               buf->entries = 1;
+               buf->entry[0].pattern = (void *)(uintptr_t)pattern;
+       }
+       size += RTE_ALIGN_CEIL(local_flow.rss.queue_num * sizeof(uint16_t),
+                              sizeof(void *));
        if (size <= flow_size)
-               flow->verbs.attr->priority = flow->attributes.priority;
+               flow->queue = (void *)(flow + 1);
+       LIST_INIT(&flow->verbs);
+       flow->layers = 0;
+       flow->modifier = 0;
+       flow->fate = 0;
+       for (i = 0; i != buf->entries; ++i) {
+               size_t off = size;
+               size_t off2;
+
+               flow->layers = original_layers;
+               size += sizeof(struct ibv_flow_attr) +
+                       sizeof(struct mlx5_flow_verbs);
+               off2 = size;
+               if (size < flow_size) {
+                       flow->cur_verbs = (void *)((uintptr_t)flow + off);
+                       flow->cur_verbs->attr = (void *)(flow->cur_verbs + 1);
+                       flow->cur_verbs->specs =
+                               (void *)(flow->cur_verbs->attr + 1);
+               }
+               /* First iteration convert the pattern into Verbs. */
+               if (i == 0) {
+                       /* Actions don't need to be converted several time. */
+                       ret = mlx5_flow_actions(dev, actions, flow,
+                                               (size < flow_size) ?
+                                               flow_size - size : 0,
+                                               error);
+                       if (ret < 0)
+                               return ret;
+                       size += ret;
+               } else {
+                       /*
+                        * Next iteration means the pattern has already been
+                        * converted and an expansion is necessary to match
+                        * the user RSS request.  For that only the expanded
+                        * items will be converted, the common part with the
+                        * user pattern are just copied into the next buffer
+                        * zone.
+                        */
+                       size += original_verbs_size;
+                       if (size < flow_size) {
+                               rte_memcpy(flow->cur_verbs->attr,
+                                          original_verbs->attr,
+                                          original_verbs_size +
+                                          sizeof(struct ibv_flow_attr));
+                               flow->cur_verbs->size = original_verbs_size;
+                       }
+               }
+               ret = mlx5_flow_items
+                       ((const struct rte_flow_item *)
+                        &buf->entry[i].pattern[expanded_pattern_idx],
+                        flow,
+                        (size < flow_size) ? flow_size - size : 0, error);
+               if (ret < 0)
+                       return ret;
+               size += ret;
+               if (size <= flow_size) {
+                       mlx5_flow_adjust_priority(dev, flow);
+                       LIST_INSERT_HEAD(&flow->verbs, flow->cur_verbs, next);
+               }
+               /*
+                * Keep a pointer of the first verbs conversion and the layers
+                * it has encountered.
+                */
+               if (i == 0) {
+                       original_verbs = flow->cur_verbs;
+                       original_verbs_size = size - off2;
+                       original_layers = flow->layers;
+                       /*
+                        * move the index of the expanded pattern to the
+                        * first item not addressed yet.
+                        */
+                       if (pattern->type == RTE_FLOW_ITEM_TYPE_END) {
+                               expanded_pattern_idx++;
+                       } else {
+                               const struct rte_flow_item *item = pattern;
+
+                               for (item = pattern;
+                                    item->type != RTE_FLOW_ITEM_TYPE_END;
+                                    ++item)
+                                       expanded_pattern_idx++;
+                       }
+               }
+       }
+       /* Restore the origin layers in the flow. */
+       flow->layers = original_layers;
        return size;
 }
 
@@ -1359,12 +1720,17 @@ mlx5_flow_rxq_mark_set(struct rte_eth_dev *dev, struct rte_flow *flow)
        struct priv *priv = dev->data->dev_private;
 
        if (flow->modifier & (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK)) {
-               struct mlx5_rxq_ctrl *rxq_ctrl =
-                       container_of((*priv->rxqs)[flow->queue],
-                                    struct mlx5_rxq_ctrl, rxq);
+               unsigned int i;
+
+               for (i = 0; i != flow->rss.queue_num; ++i) {
+                       int idx = (*flow->queue)[i];
+                       struct mlx5_rxq_ctrl *rxq_ctrl =
+                               container_of((*priv->rxqs)[idx],
+                                            struct mlx5_rxq_ctrl, rxq);
 
-               rxq_ctrl->rxq.mark = 1;
-               rxq_ctrl->flow_mark_n++;
+                       rxq_ctrl->rxq.mark = 1;
+                       rxq_ctrl->flow_mark_n++;
+               }
        }
 }
 
@@ -1383,12 +1749,17 @@ mlx5_flow_rxq_mark_trim(struct rte_eth_dev *dev, struct rte_flow *flow)
        struct priv *priv = dev->data->dev_private;
 
        if (flow->modifier & (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK)) {
-               struct mlx5_rxq_ctrl *rxq_ctrl =
-                       container_of((*priv->rxqs)[flow->queue],
-                                    struct mlx5_rxq_ctrl, rxq);
+               unsigned int i;
+
+               for (i = 0; i != flow->rss.queue_num; ++i) {
+                       int idx = (*flow->queue)[i];
+                       struct mlx5_rxq_ctrl *rxq_ctrl =
+                               container_of((*priv->rxqs)[idx],
+                                            struct mlx5_rxq_ctrl, rxq);
 
-               rxq_ctrl->flow_mark_n--;
-               rxq_ctrl->rxq.mark = !!rxq_ctrl->flow_mark_n;
+                       rxq_ctrl->flow_mark_n--;
+                       rxq_ctrl->rxq.mark = !!rxq_ctrl->flow_mark_n;
+               }
        }
 }
 
@@ -1449,18 +1820,20 @@ mlx5_flow_validate(struct rte_eth_dev *dev,
 static void
 mlx5_flow_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
 {
-       if (flow->fate & MLX5_FLOW_FATE_DROP) {
-               if (flow->verbs.flow) {
-                       claim_zero(mlx5_glue->destroy_flow(flow->verbs.flow));
-                       flow->verbs.flow = NULL;
+       struct mlx5_flow_verbs *verbs;
+
+       LIST_FOREACH(verbs, &flow->verbs, next) {
+               if (verbs->flow) {
+                       claim_zero(mlx5_glue->destroy_flow(verbs->flow));
+                       verbs->flow = NULL;
+               }
+               if (verbs->hrxq) {
+                       if (flow->fate & MLX5_FLOW_FATE_DROP)
+                               mlx5_hrxq_drop_release(dev);
+                       else
+                               mlx5_hrxq_release(dev, verbs->hrxq);
+                       verbs->hrxq = NULL;
                }
-       }
-       if (flow->verbs.hrxq) {
-               if (flow->fate & MLX5_FLOW_FATE_DROP)
-                       mlx5_hrxq_drop_release(dev);
-               else if (flow->fate & MLX5_FLOW_FATE_QUEUE)
-                       mlx5_hrxq_release(dev, flow->verbs.hrxq);
-               flow->verbs.hrxq = NULL;
        }
 }
 
@@ -1481,46 +1854,68 @@ static int
 mlx5_flow_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
                struct rte_flow_error *error)
 {
-       if (flow->fate & MLX5_FLOW_FATE_DROP) {
-               flow->verbs.hrxq = mlx5_hrxq_drop_new(dev);
-               if (!flow->verbs.hrxq)
-                       return rte_flow_error_set
-                               (error, errno,
-                                RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-                                NULL,
-                                "cannot allocate Drop queue");
-       } else if (flow->fate & MLX5_FLOW_FATE_QUEUE) {
-               struct mlx5_hrxq *hrxq;
-
-               hrxq = mlx5_hrxq_get(dev, rss_hash_default_key,
-                                    MLX5_RSS_HASH_KEY_LEN, 0,
-                                    &flow->queue, 1, 0, 0);
-               if (!hrxq)
-                       hrxq = mlx5_hrxq_new(dev, rss_hash_default_key,
-                                            MLX5_RSS_HASH_KEY_LEN, 0,
-                                            &flow->queue, 1, 0, 0);
-               if (!hrxq)
-                       return rte_flow_error_set(error, rte_errno,
-                                       RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-                                       NULL,
-                                       "cannot create flow");
-               flow->verbs.hrxq = hrxq;
-       }
-       flow->verbs.flow =
-               mlx5_glue->create_flow(flow->verbs.hrxq->qp, flow->verbs.attr);
-       if (!flow->verbs.flow) {
-               if (flow->fate & MLX5_FLOW_FATE_DROP)
-                       mlx5_hrxq_drop_release(dev);
-               else
-                       mlx5_hrxq_release(dev, flow->verbs.hrxq);
-               flow->verbs.hrxq = NULL;
-               return rte_flow_error_set(error, errno,
-                                         RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-                                         NULL,
-                                         "kernel module refuses to create"
-                                         " flow");
+       struct mlx5_flow_verbs *verbs;
+       int err;
+
+       LIST_FOREACH(verbs, &flow->verbs, next) {
+               if (flow->fate & MLX5_FLOW_FATE_DROP) {
+                       verbs->hrxq = mlx5_hrxq_drop_new(dev);
+                       if (!verbs->hrxq) {
+                               rte_flow_error_set
+                                       (error, errno,
+                                        RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                        NULL,
+                                        "cannot get drop hash queue");
+                               goto error;
+                       }
+               } else {
+                       struct mlx5_hrxq *hrxq;
+
+                       hrxq = mlx5_hrxq_get(dev, flow->key,
+                                            MLX5_RSS_HASH_KEY_LEN,
+                                            verbs->hash_fields,
+                                            (*flow->queue),
+                                            flow->rss.queue_num, 0, 0);
+                       if (!hrxq)
+                               hrxq = mlx5_hrxq_new(dev, flow->key,
+                                                    MLX5_RSS_HASH_KEY_LEN,
+                                                    verbs->hash_fields,
+                                                    (*flow->queue),
+                                                    flow->rss.queue_num, 0, 0);
+                       if (!hrxq) {
+                               rte_flow_error_set
+                                       (error, rte_errno,
+                                        RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                        NULL,
+                                        "cannot get hash queue");
+                               goto error;
+                       }
+                       verbs->hrxq = hrxq;
+               }
+               verbs->flow =
+                       mlx5_glue->create_flow(verbs->hrxq->qp, verbs->attr);
+               if (!verbs->flow) {
+                       rte_flow_error_set(error, errno,
+                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                          NULL,
+                                          "hardware refuses to create flow");
+                       goto error;
+               }
        }
        return 0;
+error:
+       err = rte_errno; /* Save rte_errno before cleanup. */
+       LIST_FOREACH(verbs, &flow->verbs, next) {
+               if (verbs->hrxq) {
+                       if (flow->fate & MLX5_FLOW_FATE_DROP)
+                               mlx5_hrxq_drop_release(dev);
+                       else
+                               mlx5_hrxq_release(dev, verbs->hrxq);
+                       verbs->hrxq = NULL;
+               }
+       }
+       rte_errno = err; /* Restore rte_errno. */
+       return -rte_errno;
 }
 
 /**
@@ -1550,42 +1945,43 @@ mlx5_flow_list_create(struct rte_eth_dev *dev,
                      const struct rte_flow_action actions[],
                      struct rte_flow_error *error)
 {
-       struct rte_flow *flow;
-       size_t size;
+       struct rte_flow *flow = NULL;
+       size_t size = 0;
        int ret;
 
-       ret = mlx5_flow_merge(dev, NULL, 0, attr, items, actions, error);
+       ret = mlx5_flow_merge(dev, flow, size, attr, items, actions, error);
        if (ret < 0)
                return NULL;
        size = ret;
-       flow = rte_zmalloc(__func__, size, 0);
+       flow = rte_calloc(__func__, 1, size, 0);
        if (!flow) {
                rte_flow_error_set(error, ENOMEM,
                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                   NULL,
-                                  "cannot allocate memory");
+                                  "not enough memory to create flow");
                return NULL;
        }
-       flow->verbs.attr = (struct ibv_flow_attr *)(flow + 1);
-       flow->verbs.specs = (uint8_t *)(flow->verbs.attr + 1);
        ret = mlx5_flow_merge(dev, flow, size, attr, items, actions, error);
-       if (ret < 0)
-               goto error;
+       if (ret < 0) {
+               rte_free(flow);
+               return NULL;
+       }
        assert((size_t)ret == size);
        if (dev->data->dev_started) {
                ret = mlx5_flow_apply(dev, flow, error);
-               if (ret < 0)
-                       goto error;
+               if (ret < 0) {
+                       ret = rte_errno; /* Save rte_errno before cleanup. */
+                       if (flow) {
+                               mlx5_flow_remove(dev, flow);
+                               rte_free(flow);
+                       }
+                       rte_errno = ret; /* Restore rte_errno. */
+                       return NULL;
+               }
        }
        mlx5_flow_rxq_mark_set(dev, flow);
        TAILQ_INSERT_TAIL(list, flow, next);
        return flow;
-error:
-       ret = rte_errno; /* Save rte_errno before cleanup. */
-       mlx5_flow_remove(dev, flow);
-       rte_free(flow);
-       rte_errno = ret; /* Restore rte_errno. */
-       return NULL;
 }
 
 /**
@@ -1745,7 +2141,7 @@ mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
        struct priv *priv = dev->data->dev_private;
        const struct rte_flow_attr attr = {
                .ingress = 1,
-               .priority = priv->config.flow_prio - 1,
+               .priority = MLX5_FLOW_PRIO_RSVD,
        };
        struct rte_flow_item items[] = {
                {