[DPDK 18.08 v1 12/12] net/mlx5: add RSS flow action

[Nelio Laranjeiro <nelio.laranjeiro@6wind.com>]

Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
---
 drivers/net/mlx5/mlx5_flow.c | 681 +++++++++++++++++++++++++----------
 1 file changed, 484 insertions(+), 197 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_flow.c b/drivers/net/mlx5/mlx5_flow.c
index 4ef0a3fee..27354615f 100644
--- a/drivers/net/mlx5/mlx5_flow.c
+++ b/drivers/net/mlx5/mlx5_flow.c
@@ -54,6 +54,7 @@ extern const struct eth_dev_ops mlx5_dev_ops_isolate;
 /* Action fate on the packet. */
 #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)
@@ -68,15 +69,40 @@ extern const struct eth_dev_ops mlx5_dev_ops_isolate;
 #define MLX5_FLOW_CTRL 0xffffffff
 #define MLX5_FLOW_CTRL_PRIO_OFFSET (MLX5_FLOW_PRIO_L2 + 1)
 
+#define MLX5_RSS_EXP_SUPP(...) \
+	((const enum rte_flow_item_type []) \
+	 { __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END })
+
+/** Supported expansion of items. */
+static const enum rte_flow_item_type *mlx5_supported_expansion[] = {
+	MLX5_RSS_EXP_SUPP(RTE_FLOW_ITEM_TYPE_ETH),
+	MLX5_RSS_EXP_SUPP(RTE_FLOW_ITEM_TYPE_ETH, RTE_FLOW_ITEM_TYPE_IPV4),
+	MLX5_RSS_EXP_SUPP(RTE_FLOW_ITEM_TYPE_ETH,
+			  RTE_FLOW_ITEM_TYPE_IPV4,
+			  RTE_FLOW_ITEM_TYPE_UDP),
+	MLX5_RSS_EXP_SUPP(RTE_FLOW_ITEM_TYPE_ETH,
+			  RTE_FLOW_ITEM_TYPE_IPV4,
+			  RTE_FLOW_ITEM_TYPE_TCP),
+	MLX5_RSS_EXP_SUPP(RTE_FLOW_ITEM_TYPE_ETH, RTE_FLOW_ITEM_TYPE_IPV6),
+	MLX5_RSS_EXP_SUPP(RTE_FLOW_ITEM_TYPE_ETH,
+			  RTE_FLOW_ITEM_TYPE_IPV6,
+			  RTE_FLOW_ITEM_TYPE_UDP),
+	MLX5_RSS_EXP_SUPP(RTE_FLOW_ITEM_TYPE_ETH,
+			  RTE_FLOW_ITEM_TYPE_IPV6,
+			  RTE_FLOW_ITEM_TYPE_TCP),
+};
+
 /** Handles information leading to a drop fate. */
 struct mlx5_flow_verbs {
-	unsigned int size; /**< Size of the attribute. */
+	LIST_ENTRY(mlx5_flow_verbs) next;
+	/**< Pointer to the next Verbs flow structure. */
 	uint32_t layers;
 	/**< Bit-fields of present layers see MLX5_FLOW_ITEMS_*. */
 	uint32_t modifier;
 	/**< Bit-fields of present modifier see MLX5_FLOW_MOD_*. */
 	uint32_t fate;
 	/**< Bit-fields of present fate see MLX5_FLOW_FATE_*. */
+	unsigned int size; /**< Size of the attribute. */
 	struct {
 		struct ibv_flow_attr *attr;
 		/**< Pointer to the Specification buffer. */
@@ -84,14 +110,19 @@ 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. */
 struct rte_flow {
 	TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
 	struct rte_flow_attr attributes; /**< User flow attribute. */
-	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[40]; /**< RSS hash key. */
+	uint16_t (*queue)[]; /**< Destination queues to redirect traffic to. */
 };
 
 static const struct rte_flow_ops mlx5_flow_ops = {
@@ -144,6 +175,8 @@ void
 mlx5_flow_print(struct rte_flow *flow __rte_unused)
 {
 #ifndef NDEBUG
+	struct mlx5_flow_verbs *verbs = LIST_FIRST(&flow->verbs);
+
 	fprintf(stdout, "---------8<------------\n");
 	fprintf(stdout, "%s: flow information\n", MLX5_DRIVER_NAME);
 	fprintf(stdout, " attributes: group %u priority %u ingress %d egress %d"
@@ -152,27 +185,36 @@ mlx5_flow_print(struct rte_flow *flow __rte_unused)
 		flow->attributes.ingress,
 		flow->attributes.egress,
 		flow->attributes.transfer);
-	fprintf(stdout, " layers: %s/%s/%s\n",
-		flow->verbs.layers & MLX5_FLOW_LAYER_L2 ? "l2" : "-",
-		flow->verbs.layers & MLX5_FLOW_LAYER_L3 ? "l3" : "-",
-		flow->verbs.layers & MLX5_FLOW_LAYER_L4 ? "l4" : "-");
-	if (flow->verbs.fate & MLX5_FLOW_FATE_DROP)
+	if (verbs->fate & MLX5_FLOW_FATE_DROP) {
 		fprintf(stdout, " fate: drop queue\n");
-	else if (flow->verbs.fate & MLX5_FLOW_FATE_QUEUE)
-		fprintf(stdout, " fate: target queue %u\n", flow->queue);
-	if (flow->verbs.attr) {
-		struct ibv_spec_header *hdr =
-			(struct ibv_spec_header *)flow->verbs.specs;
-		const int n = flow->verbs.attr->num_of_specs;
-		int i;
-
-		fprintf(stdout, " Verbs attributes: priority %u specs_n %u\n",
-			flow->verbs.attr->priority,
-			flow->verbs.attr->num_of_specs);
-		for (i = 0; i != n; ++i) {
-			rte_hexdump(stdout, " ", hdr, hdr->size);
-			hdr = (struct ibv_spec_header *)
-				((uint8_t *)hdr + hdr->size);
+	} else {
+		uint16_t i;
+
+		fprintf(stdout, " fate: target queues");
+		for (i = 0; i != flow->rss.queue_num; ++i)
+			fprintf(stdout, " %u", (*flow->queue)[i]);
+		fprintf(stdout, "\n");
+	}
+	LIST_FOREACH(verbs, &flow->verbs, next) {
+		fprintf(stdout, " layers: %s/%s/%s\n",
+			verbs->layers & MLX5_FLOW_LAYER_L2 ? "l2" : "-",
+			verbs->layers & MLX5_FLOW_LAYER_L3 ? "l3" : "-",
+			verbs->layers & MLX5_FLOW_LAYER_L4 ? "l4" : "-");
+		if (verbs->attr) {
+			struct ibv_spec_header *hdr =
+				(struct ibv_spec_header *)verbs->specs;
+			const int n = verbs->attr->num_of_specs;
+			int i;
+
+			fprintf(stdout, " Verbs attributes: priority %u"
+				" specs_n %u\n",
+				verbs->attr->priority,
+				verbs->attr->num_of_specs);
+			for (i = 0; i != n; ++i) {
+				rte_hexdump(stdout, " ", hdr, hdr->size);
+				hdr = (struct ibv_spec_header *)
+					((uint8_t *)hdr + hdr->size);
+			}
 		}
 	}
 	fprintf(stdout, "--------->8------------\n");
@@ -289,24 +331,25 @@ mlx5_flow_item_validate(const struct rte_flow_item *item,
 /**
  * Add a verbs specification.
  *
- * @param flow
- *   Pointer to flow structure.
+ * @param verbs
+ *   Pointer to mlx5_flow_verbs structure.
  * @param src
  *   Create specification.
  * @param size
  *   Size in bytes of the specification to copy.
  */
 static void
-mlx5_flow_spec_verbs_add(struct rte_flow *flow, void *src, unsigned int size)
+mlx5_flow_spec_verbs_add(struct mlx5_flow_verbs *verbs, void *src,
+			 unsigned int size)
 {
-	if (flow->verbs.specs) {
+	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;
 }
 
 /**
@@ -334,13 +377,14 @@ mlx5_flow_item_eth(const struct rte_flow_item *item, struct rte_flow *flow,
 		.type = RTE_BE16(0xffff),
 	};
 	const 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,
 	};
 	int ret;
 
-	if (flow->verbs.layers & MLX5_FLOW_LAYER_L2)
+	if (verbs->layers & MLX5_FLOW_LAYER_L2)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ITEM,
 					  item,
@@ -369,8 +413,8 @@ mlx5_flow_item_eth(const struct rte_flow_item *item, struct rte_flow *flow,
 		}
 		eth.val.ether_type &= eth.mask.ether_type;
 	}
-	mlx5_flow_spec_verbs_add(flow, &eth, size);
-	flow->verbs.layers |= MLX5_FLOW_LAYER_L2;
+	mlx5_flow_spec_verbs_add(verbs, &eth, size);
+	verbs->layers |= MLX5_FLOW_LAYER_L2;
 	return 0;
 }
 
@@ -426,6 +470,7 @@ mlx5_flow_item_vlan(const struct rte_flow_item *item, struct rte_flow *flow,
 	const struct rte_flow_item_vlan nic_mask = {
 		.tci = RTE_BE16(0x0fff),
 	};
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 	const unsigned int size = sizeof(struct ibv_flow_spec_eth);
 	struct ibv_flow_spec_eth eth = {
 		.type = IBV_FLOW_SPEC_ETH,
@@ -436,12 +481,12 @@ mlx5_flow_item_vlan(const struct rte_flow_item *item, struct rte_flow *flow,
 	const uint32_t vlanm = MLX5_FLOW_LAYER_VLAN;
 	const uint32_t l2m = MLX5_FLOW_LAYER_L2;
 
-	if (flow->verbs.layers & vlanm)
+	if (verbs->layers & vlanm)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ITEM,
 					  item,
 					  "L2 layers already configured");
-	else if ((flow->verbs.layers & lm) != 0)
+	else if ((verbs->layers & lm) != 0)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ITEM,
 					  item,
@@ -477,11 +522,11 @@ mlx5_flow_item_vlan(const struct rte_flow_item *item, struct rte_flow *flow,
 					  item->spec,
 					  "VLAN TPID matching is not"
 					  " supported");
-	if (!(flow->verbs.layers & l2m))
-		mlx5_flow_spec_verbs_add(flow, &eth, size);
+	if (!(verbs->layers & l2m))
+		mlx5_flow_spec_verbs_add(verbs, &eth, size);
 	else
-		mlx5_flow_item_vlan_update(flow->verbs.attr, &eth);
-	flow->verbs.layers |= MLX5_FLOW_LAYER_L2 | MLX5_FLOW_LAYER_VLAN;
+		mlx5_flow_item_vlan_update(verbs->attr, &eth);
+	verbs->layers |= MLX5_FLOW_LAYER_L2 | MLX5_FLOW_LAYER_VLAN;
 	return 0;
 }
 
@@ -512,6 +557,7 @@ mlx5_flow_item_ipv4(const struct rte_flow_item *item, struct rte_flow *flow,
 			.next_proto_id = 0xff,
 		},
 	};
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 	unsigned int size = sizeof(struct ibv_flow_spec_ipv4_ext);
 	struct ibv_flow_spec_ipv4_ext ipv4 = {
 		.type = IBV_FLOW_SPEC_IPV4_EXT,
@@ -519,12 +565,12 @@ mlx5_flow_item_ipv4(const struct rte_flow_item *item, struct rte_flow *flow,
 	};
 	int ret;
 
-	if (flow->verbs.layers & MLX5_FLOW_LAYER_L3)
+	if (verbs->layers & MLX5_FLOW_LAYER_L3)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ITEM,
 					  item,
 					  "multiple L3 layers not supported");
-	else if (flow->verbs.layers & MLX5_FLOW_LAYER_L4)
+	else if (verbs->layers & MLX5_FLOW_LAYER_L4)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ITEM,
 					  item,
@@ -555,8 +601,8 @@ mlx5_flow_item_ipv4(const struct rte_flow_item *item, struct rte_flow *flow,
 		ipv4.val.proto &= ipv4.mask.proto;
 		ipv4.val.tos &= ipv4.mask.tos;
 	}
-	mlx5_flow_spec_verbs_add(flow, &ipv4, size);
-	flow->verbs.layers |= MLX5_FLOW_LAYER_L3_IPV4;
+	mlx5_flow_spec_verbs_add(verbs, &ipv4, size);
+	verbs->layers |= MLX5_FLOW_LAYER_L3_IPV4;
 	return 0;
 }
 
@@ -592,6 +638,7 @@ mlx5_flow_item_ipv6(const struct rte_flow_item *item, struct rte_flow *flow,
 			.hop_limits = 0xff,
 		},
 	};
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 	unsigned int size = sizeof(struct ibv_flow_spec_ipv6);
 	struct ibv_flow_spec_ipv6 ipv6 = {
 		.type = IBV_FLOW_SPEC_IPV6,
@@ -599,12 +646,12 @@ mlx5_flow_item_ipv6(const struct rte_flow_item *item, struct rte_flow *flow,
 	};
 	int ret;
 
-	if (flow->verbs.layers & MLX5_FLOW_LAYER_L3)
+	if (verbs->layers & MLX5_FLOW_LAYER_L3)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ITEM,
 					  item,
 					  "multiple L3 layers not supported");
-	else if (flow->verbs.layers & MLX5_FLOW_LAYER_L4)
+	else if (verbs->layers & MLX5_FLOW_LAYER_L4)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ITEM,
 					  item,
@@ -655,8 +702,8 @@ mlx5_flow_item_ipv6(const struct rte_flow_item *item, struct rte_flow *flow,
 		ipv6.val.next_hdr &= ipv6.mask.next_hdr;
 		ipv6.val.hop_limit &= ipv6.mask.hop_limit;
 	}
-	mlx5_flow_spec_verbs_add(flow, &ipv6, size);
-	flow->verbs.layers |= MLX5_FLOW_LAYER_L3_IPV6;
+	mlx5_flow_spec_verbs_add(verbs, &ipv6, size);
+	verbs->layers |= MLX5_FLOW_LAYER_L3_IPV6;
 	return 0;
 }
 
@@ -679,6 +726,7 @@ mlx5_flow_item_udp(const struct rte_flow_item *item, struct rte_flow *flow,
 {
 	const struct rte_flow_item_udp *spec = item->spec;
 	const struct rte_flow_item_udp *mask = item->mask;
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 	unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
 	struct ibv_flow_spec_tcp_udp udp = {
 		.type = IBV_FLOW_SPEC_UDP,
@@ -686,7 +734,7 @@ mlx5_flow_item_udp(const struct rte_flow_item *item, struct rte_flow *flow,
 	};
 	int ret;
 
-	if (flow->verbs.layers & MLX5_FLOW_LAYER_L4)
+	if (verbs->layers & MLX5_FLOW_LAYER_L4)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ITEM,
 					  item,
@@ -707,8 +755,8 @@ 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);
-	flow->verbs.layers |= MLX5_FLOW_LAYER_L4_UDP;
+	mlx5_flow_spec_verbs_add(verbs, &udp, size);
+	verbs->layers |= MLX5_FLOW_LAYER_L4_UDP;
 	return 0;
 }
 
@@ -731,6 +779,7 @@ mlx5_flow_item_tcp(const struct rte_flow_item *item, struct rte_flow *flow,
 {
 	const struct rte_flow_item_tcp *spec = item->spec;
 	const struct rte_flow_item_tcp *mask = item->mask;
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 	unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
 	struct ibv_flow_spec_tcp_udp tcp = {
 		.type = IBV_FLOW_SPEC_TCP,
@@ -738,7 +787,7 @@ mlx5_flow_item_tcp(const struct rte_flow_item *item, struct rte_flow *flow,
 	};
 	int ret;
 
-	if (flow->verbs.layers & MLX5_FLOW_LAYER_L4)
+	if (verbs->layers & MLX5_FLOW_LAYER_L4)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ITEM,
 					  item,
@@ -759,8 +808,8 @@ 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);
-	flow->verbs.layers |= MLX5_FLOW_LAYER_L4_TCP;
+	mlx5_flow_spec_verbs_add(verbs, &tcp, size);
+	verbs->layers |= MLX5_FLOW_LAYER_L4_TCP;
 	return 0;
 }
 
@@ -782,6 +831,8 @@ mlx5_flow_items(const struct rte_flow_item items[],
 		struct rte_flow *flow __rte_unused,
 		struct rte_flow_error *error)
 {
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
+
 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
 		int ret = 0;
 
@@ -815,7 +866,7 @@ mlx5_flow_items(const struct rte_flow_item items[],
 		if (ret < 0)
 			return ret;
 	}
-	if (!flow->verbs.layers) {
+	if (!verbs->layers) {
 		const struct rte_flow_item item = {
 			.type = RTE_FLOW_ITEM_TYPE_ETH,
 		};
@@ -845,15 +896,16 @@ mlx5_flow_action_drop(const struct rte_flow_action *actions,
 			.type = IBV_FLOW_SPEC_ACTION_DROP,
 			.size = size,
 	};
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 
-	if (flow->verbs.fate)
+	if (verbs->fate)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ACTION,
 					  actions,
 					  "multiple fate actions are not"
 					  " supported");
-	mlx5_flow_spec_verbs_add(flow, &drop, size);
-	flow->verbs.fate |= MLX5_FLOW_FATE_DROP;
+	mlx5_flow_spec_verbs_add(verbs, &drop, size);
+	verbs->fate |= MLX5_FLOW_FATE_DROP;
 	return 0;
 }
 
@@ -877,8 +929,9 @@ mlx5_flow_action_queue(struct rte_eth_dev *dev,
 {
 	struct priv *priv = dev->data->dev_private;
 	const struct rte_flow_action_queue *queue = actions->conf;
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 
-	if (flow->verbs.fate)
+	if (verbs->fate)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_ACTION,
 					  actions,
@@ -894,8 +947,144 @@ 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;
-	flow->verbs.fate |= MLX5_FLOW_FATE_QUEUE;
+	if (flow->queue)
+		(*flow->queue)[0] = queue->index;
+	flow->rss.queue_num = 1;
+	verbs->fate |= MLX5_FLOW_FATE_QUEUE;
+	return 0;
+}
+
+/**
+ * Store the Verbs hash field according to the layer and types.
+ *
+ * @param verbs
+ *   Pointer to a struct mlx5_flow_verb.
+ * @param types
+ *   RSS types for this flow (see ETH_RSS_*).
+ */
+static void
+mlx5_flow_action_rss_hash_filed(struct mlx5_flow_verbs *verbs, uint32_t types)
+{
+	if ((types & ETH_RSS_NONFRAG_IPV4_TCP) &&
+	    (verbs->layers & MLX5_FLOW_LAYER_L4_TCP))
+		verbs->hash_fields = IBV_RX_HASH_SRC_IPV4 |
+			IBV_RX_HASH_DST_IPV4 |
+			IBV_RX_HASH_SRC_PORT_TCP |
+			IBV_RX_HASH_DST_PORT_TCP;
+	else if ((types & ETH_RSS_NONFRAG_IPV4_UDP) &&
+		 (verbs->layers & MLX5_FLOW_LAYER_L4_UDP))
+		verbs->hash_fields = IBV_RX_HASH_SRC_IPV4 |
+			IBV_RX_HASH_DST_IPV4 |
+			IBV_RX_HASH_SRC_PORT_UDP |
+			IBV_RX_HASH_DST_PORT_UDP;
+	else if ((types & (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4)) &&
+		 (verbs->layers & MLX5_FLOW_LAYER_L3_IPV4))
+		verbs->hash_fields = IBV_RX_HASH_SRC_IPV4 |
+			IBV_RX_HASH_DST_IPV4;
+	else if ((types & ETH_RSS_NONFRAG_IPV6_TCP) &&
+		 (verbs->layers & MLX5_FLOW_LAYER_L4_TCP))
+		verbs->hash_fields = IBV_RX_HASH_SRC_IPV6 |
+			IBV_RX_HASH_DST_IPV6 |
+			IBV_RX_HASH_SRC_PORT_TCP |
+			IBV_RX_HASH_DST_PORT_TCP;
+	else if ((types & ETH_RSS_NONFRAG_IPV6_UDP) &&
+		 (verbs->layers & MLX5_FLOW_LAYER_L3_IPV6))
+		verbs->hash_fields = IBV_RX_HASH_SRC_IPV6 |
+			IBV_RX_HASH_DST_IPV6 |
+			IBV_RX_HASH_SRC_PORT_UDP |
+			IBV_RX_HASH_DST_PORT_UDP;
+	else if ((types & (ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6)) &&
+		 (verbs->layers & MLX5_FLOW_LAYER_L3_IPV6))
+		verbs->hash_fields = IBV_RX_HASH_SRC_IPV6 |
+			IBV_RX_HASH_DST_IPV6;
+	else
+		verbs->hash_fields = 0;
+}
+
+/**
+ * Validate action queue provided by the user.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param actions
+ *   Pointer to flow actions array.
+ * @param flow
+ *   Pointer to the rte_flow structure.
+ * @param error
+ *   Pointer to error structure.
+ */
+static int
+mlx5_flow_action_rss(struct rte_eth_dev *dev,
+		     const struct rte_flow_action *actions,
+		     struct rte_flow *flow,
+		     struct rte_flow_error *error)
+{
+	struct priv *priv = dev->data->dev_private;
+	const struct rte_flow_action_rss *rss = actions->conf;
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
+	unsigned int i;
+
+	/*
+	 * If verbs list is not empty, this action has already been parsed
+	 * with another items list.
+	 */
+	if (!LIST_EMPTY(&flow->verbs)) {
+		verbs->fate |= MLX5_FLOW_FATE_RSS;
+		return 0;
+	}
+	if (verbs->fate)
+		return rte_flow_error_set(error, ENOTSUP,
+					  RTE_FLOW_ERROR_TYPE_ACTION,
+					  actions,
+					  "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 < rss_hash_default_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 > rss_hash_default_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, rss_hash_default_key_len);
+	flow->rss.types = rss->types;
+	verbs->fate |= MLX5_FLOW_FATE_RSS;
 	return 0;
 }
 
@@ -919,31 +1108,32 @@ mlx5_flow_action_flag(struct rte_flow *flow)
 		.size = size,
 		.tag_id = mlx5_flow_mark_set(MLX5_FLOW_MARK_DEFAULT),
 	};
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 
-	if (flow->verbs.modifier & MLX5_FLOW_MOD_MARK)
+	if (verbs->modifier & MLX5_FLOW_MOD_MARK)
 		return 0;
-	mlx5_flow_spec_verbs_add(flow, &tag, size);
-	flow->verbs.modifier |= MLX5_FLOW_MOD_FLAG;
+	mlx5_flow_spec_verbs_add(verbs, &tag, size);
+	verbs->modifier |= MLX5_FLOW_MOD_FLAG;
 	return 0;
 }
 
 /**
  * Update verbs specification to modify the flag to mark.
  *
- * @param flow
- *   Pointer to the rte_flow structure.
+ * @param verbs
+ *   Pointer to the mlx5_flow_verbs structure.
  * @param mark_id
  *   Mark identifier to replace the flag.
  */
 static void
-mlx5_flow_action_mark_fate_queue(struct rte_flow *flow, uint32_t mark_id)
+mlx5_flow_verbs_mark_update(struct mlx5_flow_verbs *verbs, uint32_t mark_id)
 {
 	int i;
 
 	/* Update Verbs specification. */
-	for (i = 0; i != flow->verbs.attr->num_of_specs; ++i) {
+	for (i = 0; i != verbs->attr->num_of_specs; ++i) {
 		struct ibv_spec_header *hdr =
-			(struct ibv_spec_header *)flow->verbs.attr;
+			(struct ibv_spec_header *)verbs->attr;
 
 		if (hdr->type == IBV_FLOW_SPEC_ACTION_TAG) {
 			struct ibv_flow_spec_action_tag *t =
@@ -978,6 +1168,7 @@ mlx5_flow_action_mark(const struct rte_flow_action *actions,
 		.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,
@@ -990,13 +1181,13 @@ mlx5_flow_action_mark(const struct rte_flow_action *actions,
 					  &mark->id,
 					  "mark must be between 0 and"
 					  " 16777199");
-	if (flow->verbs.modifier & MLX5_FLOW_MOD_FLAG) {
-		mlx5_flow_action_mark_fate_queue(flow, mark->id);
+	if (verbs->modifier & MLX5_FLOW_MOD_FLAG) {
+		mlx5_flow_verbs_mark_update(verbs, mark->id);
 	} else {
 		tag.tag_id = mlx5_flow_mark_set(mark->id);
-		mlx5_flow_spec_verbs_add(flow, &tag, size);
+		mlx5_flow_spec_verbs_add(verbs, &tag, size);
 	}
-	flow->verbs.modifier |= MLX5_FLOW_MOD_MARK;
+	verbs->modifier |= MLX5_FLOW_MOD_MARK;
 	return 0;
 }
 
@@ -1019,13 +1210,14 @@ mlx5_flow_actions(struct rte_eth_dev *dev,
 		  struct rte_flow_error *error)
 {
 	int ret;
+	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 
 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
 		switch (actions->type) {
 		case RTE_FLOW_ACTION_TYPE_VOID:
 			break;
 		case RTE_FLOW_ACTION_TYPE_FLAG:
-			if (flow->verbs.modifier & MLX5_FLOW_MOD_MARK)
+			if (verbs->modifier & MLX5_FLOW_MOD_MARK)
 				return rte_flow_error_set
 					(error, ENOTSUP,
 					 RTE_FLOW_ERROR_TYPE_ACTION,
@@ -1042,6 +1234,9 @@ 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,
@@ -1051,7 +1246,7 @@ mlx5_flow_actions(struct rte_eth_dev *dev,
 		if (ret < 0)
 			return ret;
 	}
-	if (!flow->verbs.fate)
+	if (!verbs->fate)
 		return rte_flow_error_set(error, ENOTSUP,
 					  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 					  NULL,
@@ -1059,6 +1254,30 @@ mlx5_flow_actions(struct rte_eth_dev *dev,
 	return 0;
 }
 
+/**
+ * Update the verbs specification according to the pattern it matches.
+ *
+ * @param layers
+ *   Bit-fields of present layers see MLX5_FLOW_ITEMS_*.
+ * @param attr[in, out]
+ *   Pointer to Verbs attribute to update.
+ * @param control[in]
+ *   The specification is used for default PMD flows.
+ */
+static void
+mlx5_flow_verbs_priority(uint32_t layers, struct ibv_flow_attr *attr,
+			 const uint32_t control)
+{
+	if (layers & MLX5_FLOW_LAYER_L4)
+		attr->priority = MLX5_FLOW_PRIO_L4;
+	else if (layers & MLX5_FLOW_LAYER_L3)
+		attr->priority = MLX5_FLOW_PRIO_L3;
+	else
+		attr->priority = MLX5_FLOW_PRIO_L2;
+	if (control)
+		attr->priority += MLX5_FLOW_CTRL_PRIO_OFFSET;
+}
+
 /**
  * Validate the rule and return a flow structure filled accordingly.
  *
@@ -1089,31 +1308,89 @@ mlx5_flow_merge(struct rte_eth_dev *dev, struct rte_flow *flow,
 		const struct rte_flow_action actions[],
 		struct rte_flow_error *error)
 {
-	struct rte_flow vflow = { .verbs.layers = 0, };
-	size_t size;
+	struct mlx5_flow_verbs vverbs = { .layers = 0, };
+	struct rte_flow vflow = { .cur_verbs = &vverbs, };
+	struct rte_flow_expand_rss *buf;
 	int ret;
+	size_t size;
+	uint32_t i;
 
 	/* Make a first virtual parse. */
 	ret = mlx5_flow_actions(dev, actions, &vflow, error);
 	if (ret)
 		return ret;
-	ret = mlx5_flow_items(items, &vflow, error);
+	ret = rte_flow_expand_rss(NULL, 0, items, vflow.rss.types,
+				  mlx5_supported_expansion);
+	assert(ret > 0);
+	buf = rte_calloc(__func__, 1, ret, 0);
+	if (!buf) {
+		rte_flow_error_set(error, ENOMEM,
+				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+				   NULL,
+				   "not enough memory to expand the RSS flow");
+		goto error;
+	}
+	ret = rte_flow_expand_rss(buf, ret, items, vflow.rss.types,
+				  mlx5_supported_expansion);
+	assert(ret > 0);
+	/* Create the flow. */
+	size = sizeof(vflow) + vflow.rss.queue_num * sizeof(uint16_t);
+	size = RTE_ALIGN_CEIL(size, sizeof(void *));
+	if (size >= flow_size)
+		flow = &vflow;
+	else
+		flow->queue = (void *)(flow + 1);
+	LIST_INIT(&flow->verbs);
+	ret = mlx5_flow_attributes(attr, flow, error);
 	if (ret)
 		return ret;
-	/* Size of the verbs specification is now known. */
-	size = sizeof(vflow) + sizeof(struct ibv_flow_attr) + vflow.verbs.size;
-	if (size <= flow_size) {
-		ret = mlx5_flow_attributes(attr, flow, error);
-		if (ret)
-			return ret;
-		ret = mlx5_flow_items(items, flow, error);
+	for (i = 0; i != buf->entries; ++i) {
+		size_t off = size;
+
+		memset(&vverbs, 0, sizeof(vverbs));
+		flow->cur_verbs = &vverbs;
+		ret = mlx5_flow_items
+			((const struct rte_flow_item *)buf->patterns[i],
+			 flow, error);
 		if (ret)
-			return ret;
+			goto error;
 		ret = mlx5_flow_actions(dev, actions, flow, error);
 		if (ret)
-			return ret;
+			goto error;
+		/* Size of the verbs specification is now known. */
+		size += sizeof(struct ibv_flow_attr) + sizeof(vverbs) +
+			vverbs.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);
+			ret = mlx5_flow_items
+				((const struct rte_flow_item *)buf->patterns[i],
+				 flow, error);
+			if (ret)
+				goto error;
+			ret = mlx5_flow_actions(dev, actions, flow, error);
+			if (ret)
+				goto error;
+			/*
+			 * Note: This verbs priority adjustment could be done
+			 * using the priority reported by the expansion.
+			 */
+			mlx5_flow_verbs_priority(flow->cur_verbs->layers,
+						 flow->cur_verbs->attr,
+						 flow->attributes.priority ==
+						 MLX5_FLOW_CTRL);
+			mlx5_flow_action_rss_hash_filed(flow->cur_verbs,
+							flow->rss.types);
+			LIST_INSERT_HEAD(&flow->verbs, flow->cur_verbs, next);
+		}
 	}
+	rte_free(buf);
 	return size;
+error:
+	rte_free(buf);
+	return ret;
 }
 
 /**
@@ -1131,25 +1408,43 @@ mlx5_flow_rxq_mark(struct rte_eth_dev *dev, struct rte_flow *flow, int enable)
 {
 	struct priv *priv = dev->data->dev_private;
 	struct mlx5_rxq_data *rxq_data;
+	uint32_t i;
 	const uint32_t mask = MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK;
-	struct rte_flow *tmp;
-	uint32_t mark = !!enable;
+	struct mlx5_flow_verbs *verbs;
 
-	if (!(flow->verbs.modifier & mask))
+	verbs = LIST_FIRST(&flow->verbs);
+	if (!(verbs->modifier & mask))
 		return;
-	rxq_data = (*priv->rxqs)[flow->queue];
-	/**
-	 * Mark/Flag bit can only be disabled when there is no other
-	 * flow applied using the same queue has a MARK/FLOW action
-	 * configured.
-	 */
-	TAILQ_FOREACH(tmp, &priv->flows, next) {
-		if (tmp == flow)
-			continue;
-		if (tmp->queue == flow->queue)
-			mark |= !!(tmp->verbs.modifier & mask);
+	for (i = 0; i != flow->rss.queue_num; ++i) {
+		int mark = 0;
+		unsigned int qidx = (*flow->queue)[i];
+		struct rte_flow *tmp;
+		rxq_data = (*priv->rxqs)[qidx];
+
+		if (enable) {
+			rxq_data->mark = !!enable;
+			break;
+		}
+		/**
+		 * Mark/Flag bit can only be disabled when there is no other
+		 * flow applied using the same queue has a MARK/FLOW action
+		 * configured.
+		 */
+		TAILQ_FOREACH(tmp, &priv->flows, next) {
+			uint32_t j;
+			struct mlx5_flow_verbs *vtmp;
+
+			if (tmp == flow)
+				continue;
+			vtmp = LIST_FIRST(&tmp->verbs);
+			for (j = 0; j != tmp->rss.queue_num; ++j)
+				if ((*tmp->queue)[i] == qidx) {
+					mark |= !!(vtmp->modifier & mask);
+					break;
+				}
+		}
+		rxq_data->mark = mark;
 	}
-	rxq_data->mark = mark;
 }
 
 /**
@@ -1183,43 +1478,21 @@ mlx5_flow_validate(struct rte_eth_dev *dev,
 static void
 mlx5_flow_fate_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
 {
-	if (flow->verbs.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 (verbs->fate & MLX5_FLOW_FATE_DROP)
+				mlx5_hrxq_drop_release(dev, verbs->hrxq);
+			else
+				mlx5_hrxq_release(dev, verbs->hrxq);
+			verbs->hrxq = NULL;
 		}
 	}
-	if (flow->verbs.hrxq) {
-		if (flow->verbs.fate & MLX5_FLOW_FATE_DROP)
-			mlx5_hrxq_drop_release(dev, flow->verbs.hrxq);
-		else if (flow->verbs.fate & MLX5_FLOW_FATE_QUEUE)
-			mlx5_hrxq_release(dev, flow->verbs.hrxq);
-		flow->verbs.hrxq = NULL;
-	}
-}
-
-/**
- * Update the verbs specification according to the pattern it matches.
- *
- * @param layers
- *   Bit-fields of present layers see MLX5_FLOW_ITEMS_*.
- * @param attr[in, out]
- *   Pointer to Verbs attribute to update.
- * @param control[in]
- *   The specification is used for default PMD flows.
- */
-static void
-mlx5_flow_verbs_priority(uint32_t layers, struct ibv_flow_attr *attr,
-			 const uint32_t control)
-{
-	if (layers & MLX5_FLOW_LAYER_L4)
-		attr->priority = MLX5_FLOW_PRIO_L4;
-	else if (layers & MLX5_FLOW_LAYER_L3)
-		attr->priority = MLX5_FLOW_PRIO_L3;
-	else
-		attr->priority = MLX5_FLOW_PRIO_L2;
-	if (control)
-		attr->priority += MLX5_FLOW_CTRL_PRIO_OFFSET;
 }
 
 /**
@@ -1239,53 +1512,66 @@ static int
 mlx5_flow_fate_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
 	       struct rte_flow_error *error)
 {
-	if (flow->verbs.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->verbs.fate & MLX5_FLOW_FATE_QUEUE) {
-		struct mlx5_hrxq *hrxq;
-
-		hrxq = mlx5_hrxq_get(dev, rss_hash_default_key,
-				     rss_hash_default_key_len, 0,
-				     &flow->queue, 1, 0, 0);
-		if (!hrxq)
-			hrxq = mlx5_hrxq_new(dev, rss_hash_default_key,
-					     rss_hash_default_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;
-	} else {
-		return rte_flow_error_set(error, ENOTSUP,
-					  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-					  NULL,
-					  "no fate found");
+	struct mlx5_flow_verbs *verbs;
+	int err;
+
+	LIST_FOREACH(verbs, &flow->verbs, next) {
+		if (verbs->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,
+					     rss_hash_default_key_len,
+					     verbs->hash_fields,
+					     (*flow->queue),
+					     flow->rss.queue_num, 0, 0);
+			if (!hrxq)
+				hrxq = mlx5_hrxq_new(dev, flow->key,
+						     rss_hash_default_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;
+		}
 	}
-	mlx5_flow_verbs_priority(flow->verbs.layers, flow->verbs.attr,
-				 flow->attributes.priority == MLX5_FLOW_CTRL);
-	flow->verbs.flow =
-		mlx5_glue->create_flow(flow->verbs.hrxq->qp, flow->verbs.attr);
-	if (!flow->verbs.flow) {
-		if (flow->verbs.fate & MLX5_FLOW_FATE_DROP)
-			mlx5_hrxq_drop_release(dev, flow->verbs.hrxq);
+	return 0;
+error:
+	err = rte_errno; /* Save rte_errno before cleanup. */
+	LIST_FOREACH(verbs, &flow->verbs, next) {
+		if (verbs->fate & MLX5_FLOW_FATE_DROP)
+			mlx5_hrxq_drop_release(dev, verbs->hrxq);
 		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");
+			mlx5_hrxq_release(dev, verbs->hrxq);
+		verbs->hrxq = NULL;
 	}
-	return 0;
+	rte_errno = err; /* Restore rte_errno. */
+	return -rte_errno;
 }
 
 /**
@@ -1315,42 +1601,43 @@ 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)
 		return flow;
 	ret = mlx5_flow_fate_apply(dev, flow, error);
-	if (ret < 0)
-		goto error;
+	if (ret < 0) {
+		ret = rte_errno; /* Save rte_errno before cleanup. */
+		if (flow) {
+			mlx5_flow_fate_remove(dev, flow);
+			rte_free(flow);
+		}
+		rte_errno = ret; /* Restore rte_errno. */
+		return NULL;
+	}
 	mlx5_flow_rxq_mark(dev, flow, 1);
 	TAILQ_INSERT_TAIL(list, flow, next);
 	return flow;
-error:
-	ret = rte_errno; /* Save rte_errno before cleanup. */
-	mlx5_flow_fate_remove(dev, flow);
-	rte_free(flow);
-	rte_errno = ret; /* Restore rte_errno. */
-	return NULL;
 }
 
 /**
-- 
2.17.0