New upstream version 17.11-rc3

Change-Id: I6a5baa40612fe0c20f30b5fa773a6cbbac63a685
Signed-off-by: Luca Boccassi <luca.boccassi@gmail.com>

1 files changed, 1153 insertions, 909 deletions

diff --git a/drivers/net/mlx4/mlx4_flow.c b/drivers/net/mlx4/mlx4_flow.c
index 925c89c5..8b87b298 100644
--- a/drivers/net/mlx4/mlx4_flow.c
+++ b/drivers/net/mlx4/mlx4_flow.c
@@ -2,7 +2,7 @@
  *   BSD LICENSE
  *
  *   Copyright 2017 6WIND S.A.
- *   Copyright 2017 Mellanox.
+ *   Copyright 2017 Mellanox
  *
  *   Redistribution and use in source and binary forms, with or without
  *   modification, are permitted provided that the following conditions
@@ -31,197 +31,328 @@
  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
+/**
+ * @file
+ * Flow API operations for mlx4 driver.
+ */
+
+#include <arpa/inet.h>
 #include <assert.h>
+#include <errno.h>
+#include <stdalign.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+#include <sys/queue.h>
+
+/* Verbs headers do not support -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-Wpedantic"
+#endif
+#include <infiniband/verbs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-Wpedantic"
+#endif
 
+#include <rte_byteorder.h>
+#include <rte_errno.h>
+#include <rte_eth_ctrl.h>
+#include <rte_ethdev.h>
+#include <rte_ether.h>
 #include <rte_flow.h>
 #include <rte_flow_driver.h>
 #include <rte_malloc.h>
 
-/* Generated configuration header. */
-#include "mlx4_autoconf.h"
-
 /* PMD headers. */
 #include "mlx4.h"
 #include "mlx4_flow.h"
+#include "mlx4_rxtx.h"
+#include "mlx4_utils.h"
 
-/** Static initializer for items. */
-#define ITEMS(...) \
+/** Static initializer for a list of subsequent item types. */
+#define NEXT_ITEM(...) \
 	(const enum rte_flow_item_type []){ \
 		__VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 	}
 
-/** Structure to generate a simple graph of layers supported by the NIC. */
-struct mlx4_flow_items {
-	/** List of possible actions for these items. */
-	const enum rte_flow_action_type *const actions;
-	/** Bit-masks corresponding to the possibilities for the item. */
-	const void *mask;
-	/**
-	 * Default bit-masks to use when item->mask is not provided. When
-	 * \default_mask is also NULL, the full supported bit-mask (\mask) is
-	 * used instead.
-	 */
-	const void *default_mask;
-	/** Bit-masks size in bytes. */
+/** Processor structure associated with a flow item. */
+struct mlx4_flow_proc_item {
+	/** Bit-mask for fields supported by this PMD. */
+	const void *mask_support;
+	/** Bit-mask to use when @p item->mask is not provided. */
+	const void *mask_default;
+	/** Size in bytes for @p mask_support and @p mask_default. */
 	const unsigned int mask_sz;
-	/**
-	 * Check support for a given item.
-	 *
-	 * @param item[in]
-	 *   Item specification.
-	 * @param mask[in]
-	 *   Bit-masks covering supported fields to compare with spec,
-	 *   last and mask in
-	 *   \item.
-	 * @param size
-	 *   Bit-Mask size in bytes.
-	 *
-	 * @return
-	 *   0 on success, negative value otherwise.
-	 */
-	int (*validate)(const struct rte_flow_item *item,
-			const uint8_t *mask, unsigned int size);
-	/**
-	 * Conversion function from rte_flow to NIC specific flow.
-	 *
-	 * @param item
-	 *   rte_flow item to convert.
-	 * @param default_mask
-	 *   Default bit-masks to use when item->mask is not provided.
-	 * @param data
-	 *   Internal structure to store the conversion.
-	 *
-	 * @return
-	 *   0 on success, negative value otherwise.
-	 */
-	int (*convert)(const struct rte_flow_item *item,
-		       const void *default_mask,
-		       void *data);
+	/** Merge a pattern item into a flow rule handle. */
+	int (*merge)(struct rte_flow *flow,
+		     const struct rte_flow_item *item,
+		     const struct mlx4_flow_proc_item *proc,
+		     struct rte_flow_error *error);
 	/** Size in bytes of the destination structure. */
 	const unsigned int dst_sz;
-	/** List of possible following items.  */
-	const enum rte_flow_item_type *const items;
+	/** List of possible subsequent items. */
+	const enum rte_flow_item_type *const next_item;
 };
 
-struct rte_flow_drop {
-	struct ibv_qp *qp; /**< Verbs queue pair. */
-	struct ibv_cq *cq; /**< Verbs completion queue. */
+/** Shared resources for drop flow rules. */
+struct mlx4_drop {
+	struct ibv_qp *qp; /**< QP target. */
+	struct ibv_cq *cq; /**< CQ associated with above QP. */
+	struct priv *priv; /**< Back pointer to private data. */
+	uint32_t refcnt; /**< Reference count. */
 };
 
-/** Valid action for this PMD. */
-static const enum rte_flow_action_type valid_actions[] = {
-	RTE_FLOW_ACTION_TYPE_DROP,
-	RTE_FLOW_ACTION_TYPE_QUEUE,
-	RTE_FLOW_ACTION_TYPE_RSS,
-	RTE_FLOW_ACTION_TYPE_END,
-};
+/**
+ * Convert DPDK RSS hash fields to their Verbs equivalent.
+ *
+ * @param rss_hf
+ *   Hash fields in DPDK format (see struct rte_eth_rss_conf).
+ *
+ * @return
+ *   A valid Verbs RSS hash fields mask for mlx4 on success, (uint64_t)-1
+ *   otherwise and rte_errno is set.
+ */
+static uint64_t
+mlx4_conv_rss_hf(uint64_t rss_hf)
+{
+	enum { IPV4, IPV6, TCP, UDP, };
+	const uint64_t in[] = {
+		[IPV4] = (ETH_RSS_IPV4 |
+			  ETH_RSS_FRAG_IPV4 |
+			  ETH_RSS_NONFRAG_IPV4_TCP |
+			  ETH_RSS_NONFRAG_IPV4_UDP |
+			  ETH_RSS_NONFRAG_IPV4_OTHER),
+		[IPV6] = (ETH_RSS_IPV6 |
+			  ETH_RSS_FRAG_IPV6 |
+			  ETH_RSS_NONFRAG_IPV6_TCP |
+			  ETH_RSS_NONFRAG_IPV6_UDP |
+			  ETH_RSS_NONFRAG_IPV6_OTHER |
+			  ETH_RSS_IPV6_EX |
+			  ETH_RSS_IPV6_TCP_EX |
+			  ETH_RSS_IPV6_UDP_EX),
+		[TCP] = (ETH_RSS_NONFRAG_IPV4_TCP |
+			 ETH_RSS_NONFRAG_IPV6_TCP |
+			 ETH_RSS_IPV6_TCP_EX),
+		/*
+		 * UDP support is temporarily disabled due to an
+		 * implementation issue in the kernel.
+		 */
+		[UDP] = 0,
+	};
+	const uint64_t out[RTE_DIM(in)] = {
+		[IPV4] = IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4,
+		[IPV6] = IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6,
+		[TCP] = IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP,
+		[UDP] = IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP,
+	};
+	uint64_t seen = 0;
+	uint64_t conv = 0;
+	unsigned int i;
+
+	for (i = 0; i != RTE_DIM(in); ++i)
+		if (rss_hf & in[i]) {
+			seen |= rss_hf & in[i];
+			conv |= out[i];
+		}
+	if (!(rss_hf & ~seen))
+		return conv;
+	rte_errno = ENOTSUP;
+	return (uint64_t)-1;
+}
 
 /**
- * Convert Ethernet item to Verbs specification.
+ * Merge Ethernet pattern item into flow rule handle.
  *
- * @param item[in]
- *   Item specification.
- * @param default_mask[in]
- *   Default bit-masks to use when item->mask is not provided.
- * @param data[in, out]
- *   User structure.
+ * Additional mlx4-specific constraints on supported fields:
+ *
+ * - No support for partial masks, except in the specific case of matching
+ *   all multicast traffic (@p spec->dst and @p mask->dst equal to
+ *   01:00:00:00:00:00).
+ * - Not providing @p item->spec or providing an empty @p mask->dst is
+ *   *only* supported if the rule doesn't specify additional matching
+ *   criteria (i.e. rule is promiscuous-like).
+ *
+ * @param[in, out] flow
+ *   Flow rule handle to update.
+ * @param[in] item
+ *   Pattern item to merge.
+ * @param[in] proc
+ *   Associated item-processing object.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 static int
-mlx4_flow_create_eth(const struct rte_flow_item *item,
-		     const void *default_mask,
-		     void *data)
+mlx4_flow_merge_eth(struct rte_flow *flow,
+		    const struct rte_flow_item *item,
+		    const struct mlx4_flow_proc_item *proc,
+		    struct rte_flow_error *error)
 {
 	const struct rte_flow_item_eth *spec = item->spec;
-	const struct rte_flow_item_eth *mask = item->mask;
-	struct mlx4_flow *flow = (struct mlx4_flow *)data;
+	const struct rte_flow_item_eth *mask =
+		spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 	struct ibv_flow_spec_eth *eth;
-	const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
+	const char *msg;
 	unsigned int i;
 
+	if (!mask) {
+		flow->promisc = 1;
+	} else {
+		uint32_t sum_dst = 0;
+		uint32_t sum_src = 0;
+
+		for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
+			sum_dst += mask->dst.addr_bytes[i];
+			sum_src += mask->src.addr_bytes[i];
+		}
+		if (sum_src) {
+			msg = "mlx4 does not support source MAC matching";
+			goto error;
+		} else if (!sum_dst) {
+			flow->promisc = 1;
+		} else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
+			if (!(spec->dst.addr_bytes[0] & 1)) {
+				msg = "mlx4 does not support the explicit"
+					" exclusion of all multicast traffic";
+				goto error;
+			}
+			flow->allmulti = 1;
+		} else if (sum_dst != (UINT8_C(0xff) * ETHER_ADDR_LEN)) {
+			msg = "mlx4 does not support matching partial"
+				" Ethernet fields";
+			goto error;
+		}
+	}
+	if (!flow->ibv_attr)
+		return 0;
+	if (flow->promisc) {
+		flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
+		return 0;
+	}
+	if (flow->allmulti) {
+		flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
+		return 0;
+	}
 	++flow->ibv_attr->num_of_specs;
-	flow->ibv_attr->priority = 2;
-	eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
+	eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 	*eth = (struct ibv_flow_spec_eth) {
 		.type = IBV_FLOW_SPEC_ETH,
-		.size = eth_size,
+		.size = sizeof(*eth),
 	};
-	if (!spec) {
-		flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
-		return 0;
-	}
-	if (!mask)
-		mask = default_mask;
 	memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
-	memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 	memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
-	memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 	/* Remove unwanted bits from values. */
 	for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 		eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
-		eth->val.src_mac[i] &= eth->mask.src_mac[i];
 	}
 	return 0;
+error:
+	return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
+				  item, msg);
 }
 
 /**
- * Convert VLAN item to Verbs specification.
+ * Merge VLAN pattern item into flow rule handle.
  *
- * @param item[in]
- *   Item specification.
- * @param default_mask[in]
- *   Default bit-masks to use when item->mask is not provided.
- * @param data[in, out]
- *   User structure.
+ * Additional mlx4-specific constraints on supported fields:
+ *
+ * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
+ *   empty @p item->mask would also include non-VLAN traffic. Doing so is
+ *   therefore unsupported.
+ * - No support for partial masks.
+ *
+ * @param[in, out] flow
+ *   Flow rule handle to update.
+ * @param[in] item
+ *   Pattern item to merge.
+ * @param[in] proc
+ *   Associated item-processing object.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 static int
-mlx4_flow_create_vlan(const struct rte_flow_item *item,
-		      const void *default_mask,
-		      void *data)
+mlx4_flow_merge_vlan(struct rte_flow *flow,
+		     const struct rte_flow_item *item,
+		     const struct mlx4_flow_proc_item *proc,
+		     struct rte_flow_error *error)
 {
 	const struct rte_flow_item_vlan *spec = item->spec;
-	const struct rte_flow_item_vlan *mask = item->mask;
-	struct mlx4_flow *flow = (struct mlx4_flow *)data;
+	const struct rte_flow_item_vlan *mask =
+		spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 	struct ibv_flow_spec_eth *eth;
-	const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
+	const char *msg;
 
-	eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
-	if (!spec)
+	if (!mask || !mask->tci) {
+		msg = "mlx4 cannot match all VLAN traffic while excluding"
+			" non-VLAN traffic, TCI VID must be specified";
+		goto error;
+	}
+	if (mask->tci != RTE_BE16(0x0fff)) {
+		msg = "mlx4 does not support partial TCI VID matching";
+		goto error;
+	}
+	if (!flow->ibv_attr)
 		return 0;
-	if (!mask)
-		mask = default_mask;
+	eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
+		       sizeof(*eth));
 	eth->val.vlan_tag = spec->tci;
 	eth->mask.vlan_tag = mask->tci;
 	eth->val.vlan_tag &= eth->mask.vlan_tag;
 	return 0;
+error:
+	return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
+				  item, msg);
 }
 
 /**
- * Convert IPv4 item to Verbs specification.
+ * Merge IPv4 pattern item into flow rule handle.
  *
- * @param item[in]
- *   Item specification.
- * @param default_mask[in]
- *   Default bit-masks to use when item->mask is not provided.
- * @param data[in, out]
- *   User structure.
+ * Additional mlx4-specific constraints on supported fields:
+ *
+ * - No support for partial masks.
+ *
+ * @param[in, out] flow
+ *   Flow rule handle to update.
+ * @param[in] item
+ *   Pattern item to merge.
+ * @param[in] proc
+ *   Associated item-processing object.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 static int
-mlx4_flow_create_ipv4(const struct rte_flow_item *item,
-		      const void *default_mask,
-		      void *data)
+mlx4_flow_merge_ipv4(struct rte_flow *flow,
+		     const struct rte_flow_item *item,
+		     const struct mlx4_flow_proc_item *proc,
+		     struct rte_flow_error *error)
 {
 	const struct rte_flow_item_ipv4 *spec = item->spec;
-	const struct rte_flow_item_ipv4 *mask = item->mask;
-	struct mlx4_flow *flow = (struct mlx4_flow *)data;
+	const struct rte_flow_item_ipv4 *mask =
+		spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 	struct ibv_flow_spec_ipv4 *ipv4;
-	unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4);
+	const char *msg;
 
+	if (mask &&
+	    ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
+	     (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
+		msg = "mlx4 does not support matching partial IPv4 fields";
+		goto error;
+	}
+	if (!flow->ibv_attr)
+		return 0;
 	++flow->ibv_attr->num_of_specs;
-	flow->ibv_attr->priority = 1;
-	ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
+	ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 	*ipv4 = (struct ibv_flow_spec_ipv4) {
 		.type = IBV_FLOW_SPEC_IPV4,
-		.size = ipv4_size,
+		.size = sizeof(*ipv4),
 	};
 	if (!spec)
 		return 0;
@@ -229,8 +360,6 @@ mlx4_flow_create_ipv4(const struct rte_flow_item *item,
 		.src_ip = spec->hdr.src_addr,
 		.dst_ip = spec->hdr.dst_addr,
 	};
-	if (!mask)
-		mask = default_mask;
 	ipv4->mask = (struct ibv_flow_ipv4_filter) {
 		.src_ip = mask->hdr.src_addr,
 		.dst_ip = mask->hdr.dst_addr,
@@ -239,528 +368,504 @@ mlx4_flow_create_ipv4(const struct rte_flow_item *item,
 	ipv4->val.src_ip &= ipv4->mask.src_ip;
 	ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 	return 0;
+error:
+	return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
+				  item, msg);
 }
 
 /**
- * Convert UDP item to Verbs specification.
+ * Merge UDP pattern item into flow rule handle.
  *
- * @param item[in]
- *   Item specification.
- * @param default_mask[in]
- *   Default bit-masks to use when item->mask is not provided.
- * @param data[in, out]
- *   User structure.
+ * Additional mlx4-specific constraints on supported fields:
+ *
+ * - No support for partial masks.
+ *
+ * @param[in, out] flow
+ *   Flow rule handle to update.
+ * @param[in] item
+ *   Pattern item to merge.
+ * @param[in] proc
+ *   Associated item-processing object.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 static int
-mlx4_flow_create_udp(const struct rte_flow_item *item,
-		     const void *default_mask,
-		     void *data)
+mlx4_flow_merge_udp(struct rte_flow *flow,
+		    const struct rte_flow_item *item,
+		    const struct mlx4_flow_proc_item *proc,
+		    struct rte_flow_error *error)
 {
 	const struct rte_flow_item_udp *spec = item->spec;
-	const struct rte_flow_item_udp *mask = item->mask;
-	struct mlx4_flow *flow = (struct mlx4_flow *)data;
+	const struct rte_flow_item_udp *mask =
+		spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 	struct ibv_flow_spec_tcp_udp *udp;
-	unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
+	const char *msg;
 
+	if (mask &&
+	    ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
+	     (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
+		msg = "mlx4 does not support matching partial UDP fields";
+		goto error;
+	}
+	if (!flow->ibv_attr)
+		return 0;
 	++flow->ibv_attr->num_of_specs;
-	flow->ibv_attr->priority = 0;
-	udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
+	udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 	*udp = (struct ibv_flow_spec_tcp_udp) {
 		.type = IBV_FLOW_SPEC_UDP,
-		.size = udp_size,
+		.size = sizeof(*udp),
 	};
 	if (!spec)
 		return 0;
 	udp->val.dst_port = spec->hdr.dst_port;
 	udp->val.src_port = spec->hdr.src_port;
-	if (!mask)
-		mask = default_mask;
 	udp->mask.dst_port = mask->hdr.dst_port;
 	udp->mask.src_port = mask->hdr.src_port;
 	/* Remove unwanted bits from values. */
 	udp->val.src_port &= udp->mask.src_port;
 	udp->val.dst_port &= udp->mask.dst_port;
 	return 0;
+error:
+	return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
+				  item, msg);
 }
 
 /**
- * Convert TCP item to Verbs specification.
+ * Merge TCP pattern item into flow rule handle.
  *
- * @param item[in]
- *   Item specification.
- * @param default_mask[in]
- *   Default bit-masks to use when item->mask is not provided.
- * @param data[in, out]
- *   User structure.
+ * Additional mlx4-specific constraints on supported fields:
+ *
+ * - No support for partial masks.
+ *
+ * @param[in, out] flow
+ *   Flow rule handle to update.
+ * @param[in] item
+ *   Pattern item to merge.
+ * @param[in] proc
+ *   Associated item-processing object.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 static int
-mlx4_flow_create_tcp(const struct rte_flow_item *item,
-		     const void *default_mask,
-		     void *data)
+mlx4_flow_merge_tcp(struct rte_flow *flow,
+		    const struct rte_flow_item *item,
+		    const struct mlx4_flow_proc_item *proc,
+		    struct rte_flow_error *error)
 {
 	const struct rte_flow_item_tcp *spec = item->spec;
-	const struct rte_flow_item_tcp *mask = item->mask;
-	struct mlx4_flow *flow = (struct mlx4_flow *)data;
+	const struct rte_flow_item_tcp *mask =
+		spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 	struct ibv_flow_spec_tcp_udp *tcp;
-	unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
+	const char *msg;
 
+	if (mask &&
+	    ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
+	     (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
+		msg = "mlx4 does not support matching partial TCP fields";
+		goto error;
+	}
+	if (!flow->ibv_attr)
+		return 0;
 	++flow->ibv_attr->num_of_specs;
-	flow->ibv_attr->priority = 0;
-	tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
+	tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 	*tcp = (struct ibv_flow_spec_tcp_udp) {
 		.type = IBV_FLOW_SPEC_TCP,
-		.size = tcp_size,
+		.size = sizeof(*tcp),
 	};
 	if (!spec)
 		return 0;
 	tcp->val.dst_port = spec->hdr.dst_port;
 	tcp->val.src_port = spec->hdr.src_port;
-	if (!mask)
-		mask = default_mask;
 	tcp->mask.dst_port = mask->hdr.dst_port;
 	tcp->mask.src_port = mask->hdr.src_port;
 	/* Remove unwanted bits from values. */
 	tcp->val.src_port &= tcp->mask.src_port;
 	tcp->val.dst_port &= tcp->mask.dst_port;
 	return 0;
+error:
+	return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
+				  item, msg);
 }
 
 /**
- * Check support for a given item.
+ * Perform basic sanity checks on a pattern item.
  *
- * @param item[in]
+ * @param[in] item
  *   Item specification.
- * @param mask[in]
- *   Bit-masks covering supported fields to compare with spec, last and mask in
- *   \item.
- * @param size
- *   Bit-Mask size in bytes.
+ * @param[in] proc
+ *   Associated item-processing object.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
  *
  * @return
- *   0 on success, negative value otherwise.
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 static int
-mlx4_flow_item_validate(const struct rte_flow_item *item,
-			const uint8_t *mask, unsigned int size)
+mlx4_flow_item_check(const struct rte_flow_item *item,
+		     const struct mlx4_flow_proc_item *proc,
+		     struct rte_flow_error *error)
 {
-	int ret = 0;
+	const uint8_t *mask;
+	unsigned int i;
 
+	/* item->last and item->mask cannot exist without item->spec. */
 	if (!item->spec && (item->mask || item->last))
-		return -1;
-	if (item->spec && !item->mask) {
-		unsigned int i;
-		const uint8_t *spec = item->spec;
-
-		for (i = 0; i < size; ++i)
-			if ((spec[i] | mask[i]) != mask[i])
-				return -1;
-	}
-	if (item->last && !item->mask) {
-		unsigned int i;
-		const uint8_t *spec = item->last;
-
-		for (i = 0; i < size; ++i)
-			if ((spec[i] | mask[i]) != mask[i])
-				return -1;
-	}
-	if (item->spec && item->last) {
-		uint8_t spec[size];
-		uint8_t last[size];
-		const uint8_t *apply = mask;
-		unsigned int i;
-
-		if (item->mask)
-			apply = item->mask;
-		for (i = 0; i < size; ++i) {
-			spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
-			last[i] = ((const uint8_t *)item->last)[i] & apply[i];
-		}
-		ret = memcmp(spec, last, size);
-	}
-	return ret;
-}
-
-static int
-mlx4_flow_validate_eth(const struct rte_flow_item *item,
-		       const uint8_t *mask, unsigned int size)
-{
-	if (item->mask) {
-		const struct rte_flow_item_eth *mask = item->mask;
-
-		if (mask->dst.addr_bytes[0] != 0xff ||
-				mask->dst.addr_bytes[1] != 0xff ||
-				mask->dst.addr_bytes[2] != 0xff ||
-				mask->dst.addr_bytes[3] != 0xff ||
-				mask->dst.addr_bytes[4] != 0xff ||
-				mask->dst.addr_bytes[5] != 0xff)
-			return -1;
-	}
-	return mlx4_flow_item_validate(item, mask, size);
-}
-
-static int
-mlx4_flow_validate_vlan(const struct rte_flow_item *item,
-			const uint8_t *mask, unsigned int size)
-{
-	if (item->mask) {
-		const struct rte_flow_item_vlan *mask = item->mask;
-
-		if (mask->tci != 0 &&
-		    ntohs(mask->tci) != 0x0fff)
-			return -1;
-	}
-	return mlx4_flow_item_validate(item, mask, size);
-}
-
-static int
-mlx4_flow_validate_ipv4(const struct rte_flow_item *item,
-			const uint8_t *mask, unsigned int size)
-{
-	if (item->mask) {
-		const struct rte_flow_item_ipv4 *mask = item->mask;
-
-		if (mask->hdr.src_addr != 0 &&
-		    mask->hdr.src_addr != 0xffffffff)
-			return -1;
-		if (mask->hdr.dst_addr != 0 &&
-		    mask->hdr.dst_addr != 0xffffffff)
-			return -1;
-	}
-	return mlx4_flow_item_validate(item, mask, size);
-}
-
-static int
-mlx4_flow_validate_udp(const struct rte_flow_item *item,
-		       const uint8_t *mask, unsigned int size)
-{
-	if (item->mask) {
-		const struct rte_flow_item_udp *mask = item->mask;
-
-		if (mask->hdr.src_port != 0 &&
-		    mask->hdr.src_port != 0xffff)
-			return -1;
-		if (mask->hdr.dst_port != 0 &&
-		    mask->hdr.dst_port != 0xffff)
-			return -1;
-	}
-	return mlx4_flow_item_validate(item, mask, size);
-}
-
-static int
-mlx4_flow_validate_tcp(const struct rte_flow_item *item,
-		       const uint8_t *mask, unsigned int size)
-{
-	if (item->mask) {
-		const struct rte_flow_item_tcp *mask = item->mask;
-
-		if (mask->hdr.src_port != 0 &&
-		    mask->hdr.src_port != 0xffff)
-			return -1;
-		if (mask->hdr.dst_port != 0 &&
-		    mask->hdr.dst_port != 0xffff)
-			return -1;
+		return rte_flow_error_set
+			(error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
+			 "\"mask\" or \"last\" field provided without a"
+			 " corresponding \"spec\"");
+	/* No spec, no mask, no problem. */
+	if (!item->spec)
+		return 0;
+	mask = item->mask ?
+		(const uint8_t *)item->mask :
+		(const uint8_t *)proc->mask_default;
+	assert(mask);
+	/*
+	 * Single-pass check to make sure that:
+	 * - Mask is supported, no bits are set outside proc->mask_support.
+	 * - Both item->spec and item->last are included in mask.
+	 */
+	for (i = 0; i != proc->mask_sz; ++i) {
+		if (!mask[i])
+			continue;
+		if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
+		    ((const uint8_t *)proc->mask_support)[i])
+			return rte_flow_error_set
+				(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
+				 item, "unsupported field found in \"mask\"");
+		if (item->last &&
+		    (((const uint8_t *)item->spec)[i] & mask[i]) !=
+		    (((const uint8_t *)item->last)[i] & mask[i]))
+			return rte_flow_error_set
+				(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
+				 item,
+				 "range between \"spec\" and \"last\""
+				 " is larger than \"mask\"");
 	}
-	return mlx4_flow_item_validate(item, mask, size);
+	return 0;
 }
 
 /** Graph of supported items and associated actions. */
-static const struct mlx4_flow_items mlx4_flow_items[] = {
+static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
 	[RTE_FLOW_ITEM_TYPE_END] = {
-		.items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
+		.next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
 	},
 	[RTE_FLOW_ITEM_TYPE_ETH] = {
-		.items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
-			       RTE_FLOW_ITEM_TYPE_IPV4),
-		.actions = valid_actions,
-		.mask = &(const struct rte_flow_item_eth){
+		.next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
+				       RTE_FLOW_ITEM_TYPE_IPV4),
+		.mask_support = &(const struct rte_flow_item_eth){
+			/* Only destination MAC can be matched. */
 			.dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
-			.src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 		},
-		.default_mask = &rte_flow_item_eth_mask,
+		.mask_default = &rte_flow_item_eth_mask,
 		.mask_sz = sizeof(struct rte_flow_item_eth),
-		.validate = mlx4_flow_validate_eth,
-		.convert = mlx4_flow_create_eth,
+		.merge = mlx4_flow_merge_eth,
 		.dst_sz = sizeof(struct ibv_flow_spec_eth),
 	},
 	[RTE_FLOW_ITEM_TYPE_VLAN] = {
-		.items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4),
-		.actions = valid_actions,
-		.mask = &(const struct rte_flow_item_vlan){
-		/* rte_flow_item_vlan_mask is invalid for mlx4. */
-#if RTE_BYTE_ORDER == RTE_BIG_ENDIAN
-			.tci = 0x0fff,
-#else
-			.tci = 0xff0f,
-#endif
+		.next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
+		.mask_support = &(const struct rte_flow_item_vlan){
+			/* Only TCI VID matching is supported. */
+			.tci = RTE_BE16(0x0fff),
 		},
+		.mask_default = &rte_flow_item_vlan_mask,
 		.mask_sz = sizeof(struct rte_flow_item_vlan),
-		.validate = mlx4_flow_validate_vlan,
-		.convert = mlx4_flow_create_vlan,
+		.merge = mlx4_flow_merge_vlan,
 		.dst_sz = 0,
 	},
 	[RTE_FLOW_ITEM_TYPE_IPV4] = {
-		.items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
-			       RTE_FLOW_ITEM_TYPE_TCP),
-		.actions = valid_actions,
-		.mask = &(const struct rte_flow_item_ipv4){
+		.next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
+				       RTE_FLOW_ITEM_TYPE_TCP),
+		.mask_support = &(const struct rte_flow_item_ipv4){
 			.hdr = {
-				.src_addr = -1,
-				.dst_addr = -1,
+				.src_addr = RTE_BE32(0xffffffff),
+				.dst_addr = RTE_BE32(0xffffffff),
 			},
 		},
-		.default_mask = &rte_flow_item_ipv4_mask,
+		.mask_default = &rte_flow_item_ipv4_mask,
 		.mask_sz = sizeof(struct rte_flow_item_ipv4),
-		.validate = mlx4_flow_validate_ipv4,
-		.convert = mlx4_flow_create_ipv4,
+		.merge = mlx4_flow_merge_ipv4,
 		.dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 	},
 	[RTE_FLOW_ITEM_TYPE_UDP] = {
-		.actions = valid_actions,
-		.mask = &(const struct rte_flow_item_udp){
+		.mask_support = &(const struct rte_flow_item_udp){
 			.hdr = {
-				.src_port = -1,
-				.dst_port = -1,
+				.src_port = RTE_BE16(0xffff),
+				.dst_port = RTE_BE16(0xffff),
 			},
 		},
-		.default_mask = &rte_flow_item_udp_mask,
+		.mask_default = &rte_flow_item_udp_mask,
 		.mask_sz = sizeof(struct rte_flow_item_udp),
-		.validate = mlx4_flow_validate_udp,
-		.convert = mlx4_flow_create_udp,
+		.merge = mlx4_flow_merge_udp,
 		.dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 	},
 	[RTE_FLOW_ITEM_TYPE_TCP] = {
-		.actions = valid_actions,
-		.mask = &(const struct rte_flow_item_tcp){
+		.mask_support = &(const struct rte_flow_item_tcp){
 			.hdr = {
-				.src_port = -1,
-				.dst_port = -1,
+				.src_port = RTE_BE16(0xffff),
+				.dst_port = RTE_BE16(0xffff),
 			},
 		},
-		.default_mask = &rte_flow_item_tcp_mask,
+		.mask_default = &rte_flow_item_tcp_mask,
 		.mask_sz = sizeof(struct rte_flow_item_tcp),
-		.validate = mlx4_flow_validate_tcp,
-		.convert = mlx4_flow_create_tcp,
+		.merge = mlx4_flow_merge_tcp,
 		.dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 	},
 };
 
 /**
- * Validate a flow supported by the NIC.
+ * Make sure a flow rule is supported and initialize associated structure.
  *
  * @param priv
  *   Pointer to private structure.
  * @param[in] attr
  *   Flow rule attributes.
- * @param[in] items
+ * @param[in] pattern
  *   Pattern specification (list terminated by the END pattern item).
  * @param[in] actions
  *   Associated actions (list terminated by the END action).
  * @param[out] error
  *   Perform verbose error reporting if not NULL.
- * @param[in, out] flow
- *   Flow structure to update.
+ * @param[in, out] addr
+ *   Buffer where the resulting flow rule handle pointer must be stored.
+ *   If NULL, stop processing after validation stage.
  *
  * @return
  *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 static int
-priv_flow_validate(struct priv *priv,
-		   const struct rte_flow_attr *attr,
-		   const struct rte_flow_item items[],
-		   const struct rte_flow_action actions[],
-		   struct rte_flow_error *error,
-		   struct mlx4_flow *flow)
+mlx4_flow_prepare(struct priv *priv,
+		  const struct rte_flow_attr *attr,
+		  const struct rte_flow_item pattern[],
+		  const struct rte_flow_action actions[],
+		  struct rte_flow_error *error,
+		  struct rte_flow **addr)
 {
-	const struct mlx4_flow_items *cur_item = mlx4_flow_items;
-	struct mlx4_flow_action action = {
-		.queue = 0,
-		.drop = 0,
-	};
-
-	(void)priv;
-	if (attr->group) {
-		rte_flow_error_set(error, ENOTSUP,
-				   RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
-				   NULL,
-				   "groups are not supported");
-		return -rte_errno;
-	}
-	if (attr->priority) {
-		rte_flow_error_set(error, ENOTSUP,
-				   RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
-				   NULL,
-				   "priorities are not supported");
-		return -rte_errno;
-	}
-	if (attr->egress) {
-		rte_flow_error_set(error, ENOTSUP,
-				   RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
-				   NULL,
-				   "egress is not supported");
-		return -rte_errno;
-	}
-	if (!attr->ingress) {
-		rte_flow_error_set(error, ENOTSUP,
-				   RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
-				   NULL,
-				   "only ingress is supported");
-		return -rte_errno;
-	}
-	/* Go over items list. */
-	for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
-		const struct mlx4_flow_items *token = NULL;
+	const struct rte_flow_item *item;
+	const struct rte_flow_action *action;
+	const struct mlx4_flow_proc_item *proc;
+	struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
+	struct rte_flow *flow = &temp;
+	const char *msg = NULL;
+
+	if (attr->group)
+		return rte_flow_error_set
+			(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
+			 NULL, "groups are not supported");
+	if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
+		return rte_flow_error_set
+			(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
+			 NULL, "maximum priority level is "
+			 MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
+	if (attr->egress)
+		return rte_flow_error_set
+			(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
+			 NULL, "egress is not supported");
+	if (!attr->ingress)
+		return rte_flow_error_set
+			(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
+			 NULL, "only ingress is supported");
+fill:
+	proc = mlx4_flow_proc_item_list;
+	/* Go over pattern. */
+	for (item = pattern; item->type; ++item) {
+		const struct mlx4_flow_proc_item *next = NULL;
 		unsigned int i;
 		int err;
 
-		if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
+		if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
+			continue;
+		if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
+			flow->internal = 1;
 			continue;
-		/*
-		 * The nic can support patterns with NULL eth spec only
-		 * if eth is a single item in a rule.
-		 */
-		if (!items->spec &&
-			items->type == RTE_FLOW_ITEM_TYPE_ETH) {
-			const struct rte_flow_item *next = items + 1;
-
-			if (next->type != RTE_FLOW_ITEM_TYPE_END) {
-				rte_flow_error_set(error, ENOTSUP,
-						   RTE_FLOW_ERROR_TYPE_ITEM,
-						   items,
-						   "the rule requires"
-						   " an Ethernet spec");
-				return -rte_errno;
-			}
 		}
-		for (i = 0;
-		     cur_item->items &&
-		     cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
-		     ++i) {
-			if (cur_item->items[i] == items->type) {
-				token = &mlx4_flow_items[items->type];
+		if (flow->promisc || flow->allmulti) {
+			msg = "mlx4 does not support additional matching"
+				" criteria combined with indiscriminate"
+				" matching on Ethernet headers";
+			goto exit_item_not_supported;
+		}
+		for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
+			if (proc->next_item[i] == item->type) {
+				next = &mlx4_flow_proc_item_list[item->type];
 				break;
 			}
 		}
-		if (!token)
-			goto exit_item_not_supported;
-		cur_item = token;
-		err = cur_item->validate(items,
-					(const uint8_t *)cur_item->mask,
-					 cur_item->mask_sz);
-		if (err)
+		if (!next)
 			goto exit_item_not_supported;
-		if (flow->ibv_attr && cur_item->convert) {
-			err = cur_item->convert(items,
-						(cur_item->default_mask ?
-						 cur_item->default_mask :
-						 cur_item->mask),
-						 flow);
+		proc = next;
+		/*
+		 * Perform basic sanity checks only once, while handle is
+		 * not allocated.
+		 */
+		if (flow == &temp) {
+			err = mlx4_flow_item_check(item, proc, error);
 			if (err)
-				goto exit_item_not_supported;
+				return err;
 		}
-		flow->offset += cur_item->dst_sz;
+		if (proc->merge) {
+			err = proc->merge(flow, item, proc, error);
+			if (err)
+				return err;
+		}
+		flow->ibv_attr_size += proc->dst_sz;
 	}
-	/* Go over actions list */
-	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
-		if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
-			continue;
-		} else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
-			action.drop = 1;
-		} else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
-			const struct rte_flow_action_queue *queue =
-				(const struct rte_flow_action_queue *)
-				actions->conf;
+	/* Go over actions list. */
+	for (action = actions; action->type; ++action) {
+		switch (action->type) {
+			const struct rte_flow_action_queue *queue;
+			const struct rte_flow_action_rss *rss;
+			const struct rte_eth_rss_conf *rss_conf;
+			unsigned int i;
 
-			if (!queue || (queue->index > (priv->rxqs_n - 1)))
+		case RTE_FLOW_ACTION_TYPE_VOID:
+			continue;
+		case RTE_FLOW_ACTION_TYPE_DROP:
+			flow->drop = 1;
+			break;
+		case RTE_FLOW_ACTION_TYPE_QUEUE:
+			if (flow->rss)
+				break;
+			queue = action->conf;
+			if (queue->index >= priv->dev->data->nb_rx_queues) {
+				msg = "queue target index beyond number of"
+					" configured Rx queues";
 				goto exit_action_not_supported;
-			action.queue = 1;
-			action.queues_n = 1;
-			action.queues[0] = queue->index;
-		} else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
-			int i;
-			int ierr;
-			const struct rte_flow_action_rss *rss =
-				(const struct rte_flow_action_rss *)
-				actions->conf;
-
-			if (!priv->hw_rss) {
-				rte_flow_error_set(error, ENOTSUP,
-					   RTE_FLOW_ERROR_TYPE_ACTION,
-					   actions,
-					   "RSS cannot be used with "
-					   "the current configuration");
-				return -rte_errno;
 			}
-			if (!priv->isolated) {
-				rte_flow_error_set(error, ENOTSUP,
-					   RTE_FLOW_ERROR_TYPE_ACTION,
-					   actions,
-					   "RSS cannot be used without "
-					   "isolated mode");
-				return -rte_errno;
+			flow->rss = mlx4_rss_get
+				(priv, 0, mlx4_rss_hash_key_default, 1,
+				 &queue->index);
+			if (!flow->rss) {
+				msg = "not enough resources for additional"
+					" single-queue RSS context";
+				goto exit_action_not_supported;
+			}
+			break;
+		case RTE_FLOW_ACTION_TYPE_RSS:
+			if (flow->rss)
+				break;
+			rss = action->conf;
+			/* Default RSS configuration if none is provided. */
+			rss_conf =
+				rss->rss_conf ?
+				rss->rss_conf :
+				&(struct rte_eth_rss_conf){
+					.rss_key = mlx4_rss_hash_key_default,
+					.rss_key_len = MLX4_RSS_HASH_KEY_SIZE,
+					.rss_hf = (ETH_RSS_IPV4 |
+						   ETH_RSS_NONFRAG_IPV4_TCP |
+						   ETH_RSS_IPV6 |
+						   ETH_RSS_NONFRAG_IPV6_TCP),
+				};
+			/* Sanity checks. */
+			for (i = 0; i < rss->num; ++i)
+				if (rss->queue[i] >=
+				    priv->dev->data->nb_rx_queues)
+					break;
+			if (i != rss->num) {
+				msg = "queue index target beyond number of"
+					" configured Rx queues";
+				goto exit_action_not_supported;
 			}
 			if (!rte_is_power_of_2(rss->num)) {
-				rte_flow_error_set(error, ENOTSUP,
-					   RTE_FLOW_ERROR_TYPE_ACTION,
-					   actions,
-					   "the number of queues "
-					   "should be power of two");
-				return -rte_errno;
+				msg = "for RSS, mlx4 requires the number of"
+					" queues to be a power of two";
+				goto exit_action_not_supported;
 			}
-			if (priv->max_rss_tbl_sz < rss->num) {
-				rte_flow_error_set(error, ENOTSUP,
-					   RTE_FLOW_ERROR_TYPE_ACTION,
-					   actions,
-					   "the number of queues "
-					   "is too large");
-				return -rte_errno;
+			if (rss_conf->rss_key_len !=
+			    sizeof(flow->rss->key)) {
+				msg = "mlx4 supports exactly one RSS hash key"
+					" length: "
+					MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
+				goto exit_action_not_supported;
 			}
-			/* checking indexes array */
-			ierr = 0;
-			for (i = 0; i < rss->num; ++i) {
-				int j;
-				if (rss->queue[i] >= priv->rxqs_n)
-					ierr = 1;
-				/*
-				 * Prevent the user from specifying
-				 * the same queue twice in the RSS array.
-				 */
-				for (j = i + 1; j < rss->num && !ierr; ++j)
-					if (rss->queue[j] == rss->queue[i])
-						ierr = 1;
-				if (ierr) {
-					rte_flow_error_set(
-						error,
-						ENOTSUP,
-						RTE_FLOW_ERROR_TYPE_HANDLE,
-						NULL,
-						"RSS action only supports "
-						"unique queue indices "
-						"in a list");
-					return -rte_errno;
-				}
+			for (i = 1; i < rss->num; ++i)
+				if (rss->queue[i] - rss->queue[i - 1] != 1)
+					break;
+			if (i != rss->num) {
+				msg = "mlx4 requires RSS contexts to use"
+					" consecutive queue indices only";
+				goto exit_action_not_supported;
 			}
-			action.queue = 1;
-			action.queues_n = rss->num;
-			for (i = 0; i < rss->num; ++i)
-				action.queues[i] = rss->queue[i];
-		} else {
+			if (rss->queue[0] % rss->num) {
+				msg = "mlx4 requires the first queue of a RSS"
+					" context to be aligned on a multiple"
+					" of the context size";
+				goto exit_action_not_supported;
+			}
+			flow->rss = mlx4_rss_get
+				(priv, mlx4_conv_rss_hf(rss_conf->rss_hf),
+				 rss_conf->rss_key, rss->num, rss->queue);
+			if (!flow->rss) {
+				msg = "either invalid parameters or not enough"
+					" resources for additional multi-queue"
+					" RSS context";
+				goto exit_action_not_supported;
+			}
+			break;
+		default:
 			goto exit_action_not_supported;
 		}
 	}
-	if (!action.queue && !action.drop) {
-		rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
-				   NULL, "no valid action");
-		return -rte_errno;
+	if (!flow->rss && !flow->drop)
+		return rte_flow_error_set
+			(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+			 NULL, "no valid action");
+	/* Validation ends here. */
+	if (!addr) {
+		if (flow->rss)
+			mlx4_rss_put(flow->rss);
+		return 0;
 	}
+	if (flow == &temp) {
+		/* Allocate proper handle based on collected data. */
+		const struct mlx4_malloc_vec vec[] = {
+			{
+				.align = alignof(struct rte_flow),
+				.size = sizeof(*flow),
+				.addr = (void **)&flow,
+			},
+			{
+				.align = alignof(struct ibv_flow_attr),
+				.size = temp.ibv_attr_size,
+				.addr = (void **)&temp.ibv_attr,
+			},
+		};
+
+		if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec)))
+			return rte_flow_error_set
+				(error, -rte_errno,
+				 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
+				 "flow rule handle allocation failure");
+		/* Most fields will be updated by second pass. */
+		*flow = (struct rte_flow){
+			.ibv_attr = temp.ibv_attr,
+			.ibv_attr_size = sizeof(*flow->ibv_attr),
+			.rss = temp.rss,
+		};
+		*flow->ibv_attr = (struct ibv_flow_attr){
+			.type = IBV_FLOW_ATTR_NORMAL,
+			.size = sizeof(*flow->ibv_attr),
+			.priority = attr->priority,
+			.port = priv->port,
+		};
+		goto fill;
+	}
+	*addr = flow;
 	return 0;
 exit_item_not_supported:
-	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
-			   items, "item not supported");
-	return -rte_errno;
+	return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
+				  item, msg ? msg : "item not supported");
 exit_action_not_supported:
-	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
-			   actions, "action not supported");
-	return -rte_errno;
+	return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
+				  action, msg ? msg : "action not supported");
 }
 
 /**
@@ -769,552 +874,691 @@ exit_action_not_supported:
  * @see rte_flow_validate()
  * @see rte_flow_ops
  */
-int
+static int
 mlx4_flow_validate(struct rte_eth_dev *dev,
 		   const struct rte_flow_attr *attr,
-		   const struct rte_flow_item items[],
+		   const struct rte_flow_item pattern[],
 		   const struct rte_flow_action actions[],
 		   struct rte_flow_error *error)
 {
 	struct priv *priv = dev->data->dev_private;
-	int ret;
-	struct mlx4_flow flow = { .offset = sizeof(struct ibv_flow_attr) };
-
-	priv_lock(priv);
-	ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
-	priv_unlock(priv);
-	return ret;
-}
-
-/**
- * Destroy a drop queue.
- *
- * @param priv
- *   Pointer to private structure.
- */
-static void
-mlx4_flow_destroy_drop_queue(struct priv *priv)
-{
-	if (priv->flow_drop_queue) {
-		struct rte_flow_drop *fdq = priv->flow_drop_queue;
 
-		priv->flow_drop_queue = NULL;
-		claim_zero(ibv_destroy_qp(fdq->qp));
-		claim_zero(ibv_destroy_cq(fdq->cq));
-		rte_free(fdq);
-	}
+	return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 }
 
 /**
- * Create a single drop queue for all drop flows.
+ * Get a drop flow rule resources instance.
  *
  * @param priv
  *   Pointer to private structure.
  *
  * @return
- *   0 on success, negative value otherwise.
+ *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
+ *   is set.
  */
-static int
-mlx4_flow_create_drop_queue(struct priv *priv)
+static struct mlx4_drop *
+mlx4_drop_get(struct priv *priv)
 {
-	struct ibv_qp *qp;
-	struct ibv_cq *cq;
-	struct rte_flow_drop *fdq;
+	struct mlx4_drop *drop = priv->drop;
 
-	fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
-	if (!fdq) {
-		ERROR("Cannot allocate memory for drop struct");
-		goto err;
-	}
-	cq = ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
-			      &(struct ibv_exp_cq_init_attr){
-					.comp_mask = 0,
-			      });
-	if (!cq) {
-		ERROR("Cannot create drop CQ");
-		goto err_create_cq;
-	}
-	qp = ibv_exp_create_qp(priv->ctx,
-			      &(struct ibv_exp_qp_init_attr){
-					.send_cq = cq,
-					.recv_cq = cq,
-					.cap = {
-						.max_recv_wr = 1,
-						.max_recv_sge = 1,
-					},
-					.qp_type = IBV_QPT_RAW_PACKET,
-					.comp_mask =
-						IBV_EXP_QP_INIT_ATTR_PD |
-						IBV_EXP_QP_INIT_ATTR_PORT,
-					.pd = priv->pd,
-					.port_num = priv->port,
-			      });
-	if (!qp) {
-		ERROR("Cannot create drop QP");
-		goto err_create_qp;
+	if (drop) {
+		assert(drop->refcnt);
+		assert(drop->priv == priv);
+		++drop->refcnt;
+		return drop;
 	}
-	*fdq = (struct rte_flow_drop){
-		.qp = qp,
-		.cq = cq,
+	drop = rte_malloc(__func__, sizeof(*drop), 0);
+	if (!drop)
+		goto error;
+	*drop = (struct mlx4_drop){
+		.priv = priv,
+		.refcnt = 1,
 	};
-	priv->flow_drop_queue = fdq;
-	return 0;
-err_create_qp:
-	claim_zero(ibv_destroy_cq(cq));
-err_create_cq:
-	rte_free(fdq);
-err:
-	return -1;
+	drop->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
+	if (!drop->cq)
+		goto error;
+	drop->qp = ibv_create_qp(priv->pd,
+				 &(struct ibv_qp_init_attr){
+					.send_cq = drop->cq,
+					.recv_cq = drop->cq,
+					.qp_type = IBV_QPT_RAW_PACKET,
+				 });
+	if (!drop->qp)
+		goto error;
+	priv->drop = drop;
+	return drop;
+error:
+	if (drop->qp)
+		claim_zero(ibv_destroy_qp(drop->qp));
+	if (drop->cq)
+		claim_zero(ibv_destroy_cq(drop->cq));
+	if (drop)
+		rte_free(drop);
+	rte_errno = ENOMEM;
+	return NULL;
 }
 
 /**
- * Get RSS parent rxq structure for given queues.
+ * Give back a drop flow rule resources instance.
  *
- * Creates a new or returns an existed one.
- *
- * @param priv
- *   Pointer to private structure.
- * @param queues
- *   queues indices array, NULL in default RSS case.
- * @param children_n
- *   the size of queues array.
- *
- * @return
- *   Pointer to a parent rxq structure, NULL on failure.
+ * @param drop
+ *   Pointer to drop flow rule resources.
  */
-static struct rxq *
-priv_parent_get(struct priv *priv,
-		uint16_t queues[],
-		uint16_t children_n,
-		struct rte_flow_error *error)
+static void
+mlx4_drop_put(struct mlx4_drop *drop)
 {
-	unsigned int i;
-	struct rxq *parent;
-
-	for (parent = LIST_FIRST(&priv->parents);
-	     parent;
-	     parent = LIST_NEXT(parent, next)) {
-		unsigned int same = 0;
-		unsigned int overlap = 0;
-
-		/*
-		 * Find out whether an appropriate parent queue already exists
-		 * and can be reused, otherwise make sure there are no overlaps.
-		 */
-		for (i = 0; i < children_n; ++i) {
-			unsigned int j;
-
-			for (j = 0; j < parent->rss.queues_n; ++j) {
-				if (parent->rss.queues[j] != queues[i])
-					continue;
-				++overlap;
-				if (i == j)
-					++same;
-			}
-		}
-		if (same == children_n &&
-			children_n == parent->rss.queues_n)
-			return parent;
-		else if (overlap)
-			goto error;
-	}
-	/* Exclude the cases when some QPs were created without RSS */
-	for (i = 0; i < children_n; ++i) {
-		struct rxq *rxq = (*priv->rxqs)[queues[i]];
-		if (rxq->qp)
-			goto error;
-	}
-	parent = priv_parent_create(priv, queues, children_n);
-	if (!parent) {
-		rte_flow_error_set(error,
-				   ENOMEM, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-				   NULL, "flow rule creation failure");
-		return NULL;
-	}
-	return parent;
-
-error:
-	rte_flow_error_set(error,
-			   EEXIST,
-			   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-			   NULL,
-			   "sharing a queue between several"
-			   " RSS groups is not supported");
-	return NULL;
+	assert(drop->refcnt);
+	if (--drop->refcnt)
+		return;
+	drop->priv->drop = NULL;
+	claim_zero(ibv_destroy_qp(drop->qp));
+	claim_zero(ibv_destroy_cq(drop->cq));
+	rte_free(drop);
 }
 
 /**
- * Complete flow rule creation.
+ * Toggle a configured flow rule.
  *
  * @param priv
  *   Pointer to private structure.
- * @param ibv_attr
- *   Verbs flow attributes.
- * @param action
- *   Target action structure.
+ * @param flow
+ *   Flow rule handle to toggle.
+ * @param enable
+ *   Whether associated Verbs flow must be created or removed.
  * @param[out] error
  *   Perform verbose error reporting if not NULL.
  *
  * @return
- *   A flow if the rule could be created.
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
-static struct rte_flow *
-priv_flow_create_action_queue(struct priv *priv,
-			      struct ibv_flow_attr *ibv_attr,
-			      struct mlx4_flow_action *action,
-			      struct rte_flow_error *error)
+static int
+mlx4_flow_toggle(struct priv *priv,
+		 struct rte_flow *flow,
+		 int enable,
+		 struct rte_flow_error *error)
 {
-	struct ibv_qp *qp;
-	struct rte_flow *rte_flow;
-	struct rxq *rxq_parent = NULL;
+	struct ibv_qp *qp = NULL;
+	const char *msg;
+	int err;
 
-	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;
+	if (!enable) {
+		if (!flow->ibv_flow)
+			return 0;
+		claim_zero(ibv_destroy_flow(flow->ibv_flow));
+		flow->ibv_flow = NULL;
+		if (flow->drop)
+			mlx4_drop_put(priv->drop);
+		else if (flow->rss)
+			mlx4_rss_detach(flow->rss);
+		return 0;
 	}
-	if (action->drop) {
-		qp = priv->flow_drop_queue ? priv->flow_drop_queue->qp : NULL;
-	} else {
-		int ret;
+	assert(flow->ibv_attr);
+	if (!flow->internal &&
+	    !priv->isolated &&
+	    flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
+		if (flow->ibv_flow) {
+			claim_zero(ibv_destroy_flow(flow->ibv_flow));
+			flow->ibv_flow = NULL;
+			if (flow->drop)
+				mlx4_drop_put(priv->drop);
+			else if (flow->rss)
+				mlx4_rss_detach(flow->rss);
+		}
+		err = EACCES;
+		msg = ("priority level "
+		       MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
+		       " is reserved when not in isolated mode");
+		goto error;
+	}
+	if (flow->rss) {
+		struct mlx4_rss *rss = flow->rss;
+		int missing = 0;
 		unsigned int i;
-		struct rxq *rxq = NULL;
 
-		if (action->queues_n > 1) {
-			rxq_parent = priv_parent_get(priv, action->queues,
-						     action->queues_n, error);
-			if (!rxq_parent)
-				goto error;
+		/* Stop at the first nonexistent target queue. */
+		for (i = 0; i != rss->queues; ++i)
+			if (rss->queue_id[i] >=
+			    priv->dev->data->nb_rx_queues ||
+			    !priv->dev->data->rx_queues[rss->queue_id[i]]) {
+				missing = 1;
+				break;
+			}
+		if (flow->ibv_flow) {
+			if (missing ^ !flow->drop)
+				return 0;
+			/* Verbs flow needs updating. */
+			claim_zero(ibv_destroy_flow(flow->ibv_flow));
+			flow->ibv_flow = NULL;
+			if (flow->drop)
+				mlx4_drop_put(priv->drop);
+			else
+				mlx4_rss_detach(rss);
 		}
-		for (i = 0; i < action->queues_n; ++i) {
-			rxq = (*priv->rxqs)[action->queues[i]];
-			/*
-			 * In case of isolated mode we postpone
-			 * ibv receive queue creation till the first
-			 * rte_flow rule will be applied on that queue.
-			 */
-			if (!rxq->qp) {
-				assert(priv->isolated);
-				ret = rxq_create_qp(rxq, rxq->elts_n,
-						    0, 0, rxq_parent);
-				if (ret) {
-					rte_flow_error_set(
-						error,
-						ENOMEM,
-						RTE_FLOW_ERROR_TYPE_HANDLE,
-						NULL,
-						"flow rule creation failure");
-					goto error;
-				}
+		if (!missing) {
+			err = mlx4_rss_attach(rss);
+			if (err) {
+				err = -err;
+				msg = "cannot create indirection table or hash"
+					" QP to associate flow rule with";
+				goto error;
 			}
+			qp = rss->qp;
 		}
-		qp = action->queues_n > 1 ? rxq_parent->qp : rxq->qp;
-		rte_flow->qp = qp;
+		/* A missing target queue drops traffic implicitly. */
+		flow->drop = missing;
 	}
-	rte_flow->ibv_attr = ibv_attr;
-	if (!priv->started)
-		return rte_flow;
-	rte_flow->ibv_flow = ibv_create_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;
+	if (flow->drop) {
+		mlx4_drop_get(priv);
+		if (!priv->drop) {
+			err = rte_errno;
+			msg = "resources for drop flow rule cannot be created";
+			goto error;
+		}
+		qp = priv->drop->qp;
 	}
-	return rte_flow;
-
+	assert(qp);
+	if (flow->ibv_flow)
+		return 0;
+	flow->ibv_flow = ibv_create_flow(qp, flow->ibv_attr);
+	if (flow->ibv_flow)
+		return 0;
+	if (flow->drop)
+		mlx4_drop_put(priv->drop);
+	else if (flow->rss)
+		mlx4_rss_detach(flow->rss);
+	err = errno;
+	msg = "flow rule rejected by device";
 error:
-	if (rxq_parent)
-		rxq_parent_cleanup(rxq_parent);
-	rte_free(rte_flow);
-	return NULL;
+	return rte_flow_error_set
+		(error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
 }
 
 /**
- * Convert a flow.
- *
- * @param priv
- *   Pointer to private structure.
- * @param[in] attr
- *   Flow rule attributes.
- * @param[in] items
- *   Pattern specification (list terminated by the END pattern item).
- * @param[in] actions
- *   Associated actions (list terminated by the END action).
- * @param[out] error
- *   Perform verbose error reporting if not NULL.
+ * Create a flow.
  *
- * @return
- *   A flow on success, NULL otherwise.
+ * @see rte_flow_create()
+ * @see rte_flow_ops
  */
 static struct rte_flow *
-priv_flow_create(struct priv *priv,
+mlx4_flow_create(struct rte_eth_dev *dev,
 		 const struct rte_flow_attr *attr,
-		 const struct rte_flow_item items[],
+		 const struct rte_flow_item pattern[],
 		 const struct rte_flow_action actions[],
 		 struct rte_flow_error *error)
 {
-	struct rte_flow *rte_flow;
-	struct mlx4_flow_action action;
-	struct mlx4_flow flow = { .offset = sizeof(struct ibv_flow_attr), };
+	struct priv *priv = dev->data->dev_private;
+	struct rte_flow *flow;
 	int err;
 
-	err = priv_flow_validate(priv, attr, items, actions, error, &flow);
+	err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
 	if (err)
 		return NULL;
-	flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
-	if (!flow.ibv_attr) {
-		rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
-				   NULL, "cannot allocate ibv_attr memory");
-		return NULL;
-	}
-	flow.offset = sizeof(struct ibv_flow_attr);
-	*flow.ibv_attr = (struct ibv_flow_attr){
-		.comp_mask = 0,
-		.type = IBV_FLOW_ATTR_NORMAL,
-		.size = sizeof(struct ibv_flow_attr),
-		.priority = attr->priority,
-		.num_of_specs = 0,
-		.port = priv->port,
-		.flags = 0,
-	};
-	claim_zero(priv_flow_validate(priv, attr, items, actions,
-				      error, &flow));
-	action = (struct mlx4_flow_action){
-		.queue = 0,
-		.drop = 0,
-	};
-	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
-		if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
-			continue;
-		} else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
-			action.queue = 1;
-			action.queues_n = 1;
-			action.queues[0] =
-				((const struct rte_flow_action_queue *)
-				 actions->conf)->index;
-		} else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
-			action.drop = 1;
-		} else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
-			unsigned int i;
-			const struct rte_flow_action_rss *rss =
-				(const struct rte_flow_action_rss *)
-				 actions->conf;
+	err = mlx4_flow_toggle(priv, flow, priv->started, error);
+	if (!err) {
+		struct rte_flow *curr = LIST_FIRST(&priv->flows);
 
-			action.queue = 1;
-			action.queues_n = rss->num;
-			for (i = 0; i < rss->num; ++i)
-				action.queues[i] = rss->queue[i];
+		/* New rules are inserted after internal ones. */
+		if (!curr || !curr->internal) {
+			LIST_INSERT_HEAD(&priv->flows, flow, next);
 		} else {
-			rte_flow_error_set(error, ENOTSUP,
-					   RTE_FLOW_ERROR_TYPE_ACTION,
-					   actions, "unsupported action");
-			goto exit;
+			while (LIST_NEXT(curr, next) &&
+			       LIST_NEXT(curr, next)->internal)
+				curr = LIST_NEXT(curr, next);
+			LIST_INSERT_AFTER(curr, flow, next);
 		}
+		return flow;
 	}
-	rte_flow = priv_flow_create_action_queue(priv, flow.ibv_attr,
-						 &action, error);
-	if (rte_flow)
-		return rte_flow;
-exit:
-	rte_free(flow.ibv_attr);
+	if (flow->rss)
+		mlx4_rss_put(flow->rss);
+	rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
+			   error->message);
+	rte_free(flow);
 	return NULL;
 }
 
 /**
- * Create a flow.
+ * Configure isolated mode.
  *
- * @see rte_flow_create()
- * @see rte_flow_ops
- */
-struct rte_flow *
-mlx4_flow_create(struct rte_eth_dev *dev,
-		 const struct rte_flow_attr *attr,
-		 const struct rte_flow_item items[],
-		 const struct rte_flow_action actions[],
-		 struct rte_flow_error *error)
-{
-	struct priv *priv = dev->data->dev_private;
-	struct rte_flow *flow;
-
-	priv_lock(priv);
-	flow = priv_flow_create(priv, attr, items, actions, error);
-	if (flow) {
-		LIST_INSERT_HEAD(&priv->flows, flow, next);
-		DEBUG("Flow created %p", (void *)flow);
-	}
-	priv_unlock(priv);
-	return flow;
-}
-
-/**
  * @see rte_flow_isolate()
- *
- * Must be done before calling dev_configure().
- *
- * @param dev
- *   Pointer to the ethernet device structure.
- * @param enable
- *   Nonzero to enter isolated mode, attempt to leave it otherwise.
- * @param[out] error
- *   Perform verbose error reporting if not NULL. PMDs initialize this
- *   structure in case of error only.
- *
- * @return
- *   0 on success, a negative value on error.
+ * @see rte_flow_ops
  */
-int
+static int
 mlx4_flow_isolate(struct rte_eth_dev *dev,
 		  int enable,
 		  struct rte_flow_error *error)
 {
 	struct priv *priv = dev->data->dev_private;
 
-	priv_lock(priv);
-	if (priv->rxqs) {
-		rte_flow_error_set(error, ENOTSUP,
-				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-				   NULL, "isolated mode must be set"
-				   " before configuring the device");
-		priv_unlock(priv);
+	if (!!enable == !!priv->isolated)
+		return 0;
+	priv->isolated = !!enable;
+	if (mlx4_flow_sync(priv, error)) {
+		priv->isolated = !enable;
 		return -rte_errno;
 	}
-	priv->isolated = !!enable;
-	priv_unlock(priv);
 	return 0;
 }
 
 /**
- * Destroy a flow.
+ * Destroy a flow rule.
  *
- * @param priv
- *   Pointer to private structure.
- * @param[in] flow
- *   Flow to destroy.
+ * @see rte_flow_destroy()
+ * @see rte_flow_ops
  */
-static void
-priv_flow_destroy(struct priv *priv, struct rte_flow *flow)
+static int
+mlx4_flow_destroy(struct rte_eth_dev *dev,
+		  struct rte_flow *flow,
+		  struct rte_flow_error *error)
 {
-	(void)priv;
+	struct priv *priv = dev->data->dev_private;
+	int err = mlx4_flow_toggle(priv, flow, 0, error);
+
+	if (err)
+		return err;
 	LIST_REMOVE(flow, next);
-	if (flow->ibv_flow)
-		claim_zero(ibv_destroy_flow(flow->ibv_flow));
-	rte_free(flow->ibv_attr);
-	DEBUG("Flow destroyed %p", (void *)flow);
+	if (flow->rss)
+		mlx4_rss_put(flow->rss);
 	rte_free(flow);
+	return 0;
 }
 
 /**
- * Destroy a flow.
+ * Destroy user-configured flow rules.
  *
- * @see rte_flow_destroy()
+ * This function skips internal flows rules.
+ *
+ * @see rte_flow_flush()
  * @see rte_flow_ops
  */
-int
-mlx4_flow_destroy(struct rte_eth_dev *dev,
-		  struct rte_flow *flow,
-		  struct rte_flow_error *error)
+static int
+mlx4_flow_flush(struct rte_eth_dev *dev,
+		struct rte_flow_error *error)
 {
 	struct priv *priv = dev->data->dev_private;
+	struct rte_flow *flow = LIST_FIRST(&priv->flows);
+
+	while (flow) {
+		struct rte_flow *next = LIST_NEXT(flow, next);
 
-	(void)error;
-	priv_lock(priv);
-	priv_flow_destroy(priv, flow);
-	priv_unlock(priv);
+		if (!flow->internal)
+			mlx4_flow_destroy(dev, flow, error);
+		flow = next;
+	}
 	return 0;
 }
 
 /**
- * Destroy all flows.
+ * Helper function to determine the next configured VLAN filter.
  *
  * @param priv
  *   Pointer to private structure.
+ * @param vlan
+ *   VLAN ID to use as a starting point.
+ *
+ * @return
+ *   Next configured VLAN ID or a high value (>= 4096) if there is none.
  */
-static void
-priv_flow_flush(struct priv *priv)
+static uint16_t
+mlx4_flow_internal_next_vlan(struct priv *priv, uint16_t vlan)
+{
+	while (vlan < 4096) {
+		if (priv->dev->data->vlan_filter_conf.ids[vlan / 64] &
+		    (UINT64_C(1) << (vlan % 64)))
+			return vlan;
+		++vlan;
+	}
+	return vlan;
+}
+
+/**
+ * Generate internal flow rules.
+ *
+ * Various flow rules are created depending on the mode the device is in:
+ *
+ * 1. Promiscuous: port MAC + catch-all (VLAN filtering is ignored).
+ * 2. All multicast: port MAC/VLAN + catch-all multicast.
+ * 3. Otherwise: port MAC/VLAN + broadcast MAC/VLAN.
+ *
+ * About MAC flow rules:
+ *
+ * - MAC flow rules are generated from @p dev->data->mac_addrs
+ *   (@p priv->mac array).
+ * - An additional flow rule for Ethernet broadcasts is also generated.
+ * - All these are per-VLAN if @p dev->data->dev_conf.rxmode.hw_vlan_filter
+ *   is enabled and VLAN filters are configured.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static int
+mlx4_flow_internal(struct priv *priv, struct rte_flow_error *error)
 {
-	while (!LIST_EMPTY(&priv->flows)) {
-		struct rte_flow *flow;
+	struct rte_flow_attr attr = {
+		.priority = MLX4_FLOW_PRIORITY_LAST,
+		.ingress = 1,
+	};
+	struct rte_flow_item_eth eth_spec;
+	const struct rte_flow_item_eth eth_mask = {
+		.dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
+	};
+	const struct rte_flow_item_eth eth_allmulti = {
+		.dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
+	};
+	struct rte_flow_item_vlan vlan_spec;
+	const struct rte_flow_item_vlan vlan_mask = {
+		.tci = RTE_BE16(0x0fff),
+	};
+	struct rte_flow_item pattern[] = {
+		{
+			.type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
+		},
+		{
+			.type = RTE_FLOW_ITEM_TYPE_ETH,
+			.spec = &eth_spec,
+			.mask = &eth_mask,
+		},
+		{
+			/* Replaced with VLAN if filtering is enabled. */
+			.type = RTE_FLOW_ITEM_TYPE_END,
+		},
+		{
+			.type = RTE_FLOW_ITEM_TYPE_END,
+		},
+	};
+	/*
+	 * Round number of queues down to their previous power of 2 to
+	 * comply with RSS context limitations. Extra queues silently do not
+	 * get RSS by default.
+	 */
+	uint32_t queues =
+		rte_align32pow2(priv->dev->data->nb_rx_queues + 1) >> 1;
+	alignas(struct rte_flow_action_rss) uint8_t rss_conf_data
+		[offsetof(struct rte_flow_action_rss, queue) +
+		 sizeof(((struct rte_flow_action_rss *)0)->queue[0]) * queues];
+	struct rte_flow_action_rss *rss_conf = (void *)rss_conf_data;
+	struct rte_flow_action actions[] = {
+		{
+			.type = RTE_FLOW_ACTION_TYPE_RSS,
+			.conf = rss_conf,
+		},
+		{
+			.type = RTE_FLOW_ACTION_TYPE_END,
+		},
+	};
+	struct ether_addr *rule_mac = &eth_spec.dst;
+	rte_be16_t *rule_vlan =
+		priv->dev->data->dev_conf.rxmode.hw_vlan_filter &&
+		!priv->dev->data->promiscuous ?
+		&vlan_spec.tci :
+		NULL;
+	int broadcast =
+		!priv->dev->data->promiscuous &&
+		!priv->dev->data->all_multicast;
+	uint16_t vlan = 0;
+	struct rte_flow *flow;
+	unsigned int i;
+	int err = 0;
 
-		flow = LIST_FIRST(&priv->flows);
-		priv_flow_destroy(priv, flow);
+	/* Nothing to be done if there are no Rx queues. */
+	if (!queues)
+		goto error;
+	/* Prepare default RSS configuration. */
+	*rss_conf = (struct rte_flow_action_rss){
+		.rss_conf = NULL, /* Rely on default fallback settings. */
+		.num = queues,
+	};
+	for (i = 0; i != queues; ++i)
+		rss_conf->queue[i] = i;
+	/*
+	 * Set up VLAN item if filtering is enabled and at least one VLAN
+	 * filter is configured.
+	 */
+	if (rule_vlan) {
+		vlan = mlx4_flow_internal_next_vlan(priv, 0);
+		if (vlan < 4096) {
+			pattern[2] = (struct rte_flow_item){
+				.type = RTE_FLOW_ITEM_TYPE_VLAN,
+				.spec = &vlan_spec,
+				.mask = &vlan_mask,
+			};
+next_vlan:
+			*rule_vlan = rte_cpu_to_be_16(vlan);
+		} else {
+			rule_vlan = NULL;
+		}
 	}
+	for (i = 0; i != RTE_DIM(priv->mac) + broadcast; ++i) {
+		const struct ether_addr *mac;
+
+		/* Broadcasts are handled by an extra iteration. */
+		if (i < RTE_DIM(priv->mac))
+			mac = &priv->mac[i];
+		else
+			mac = &eth_mask.dst;
+		if (is_zero_ether_addr(mac))
+			continue;
+		/* Check if MAC flow rule is already present. */
+		for (flow = LIST_FIRST(&priv->flows);
+		     flow && flow->internal;
+		     flow = LIST_NEXT(flow, next)) {
+			const struct ibv_flow_spec_eth *eth =
+				(const void *)((uintptr_t)flow->ibv_attr +
+					       sizeof(*flow->ibv_attr));
+			unsigned int j;
+
+			if (!flow->mac)
+				continue;
+			assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
+			assert(flow->ibv_attr->num_of_specs == 1);
+			assert(eth->type == IBV_FLOW_SPEC_ETH);
+			assert(flow->rss);
+			if (rule_vlan &&
+			    (eth->val.vlan_tag != *rule_vlan ||
+			     eth->mask.vlan_tag != RTE_BE16(0x0fff)))
+				continue;
+			if (!rule_vlan && eth->mask.vlan_tag)
+				continue;
+			for (j = 0; j != sizeof(mac->addr_bytes); ++j)
+				if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
+				    eth->mask.dst_mac[j] != UINT8_C(0xff) ||
+				    eth->val.src_mac[j] != UINT8_C(0x00) ||
+				    eth->mask.src_mac[j] != UINT8_C(0x00))
+					break;
+			if (j != sizeof(mac->addr_bytes))
+				continue;
+			if (flow->rss->queues != queues ||
+			    memcmp(flow->rss->queue_id, rss_conf->queue,
+				   queues * sizeof(flow->rss->queue_id[0])))
+				continue;
+			break;
+		}
+		if (!flow || !flow->internal) {
+			/* Not found, create a new flow rule. */
+			memcpy(rule_mac, mac, sizeof(*mac));
+			flow = mlx4_flow_create(priv->dev, &attr, pattern,
+						actions, error);
+			if (!flow) {
+				err = -rte_errno;
+				goto error;
+			}
+		}
+		flow->select = 1;
+		flow->mac = 1;
+	}
+	if (rule_vlan) {
+		vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
+		if (vlan < 4096)
+			goto next_vlan;
+	}
+	/* Take care of promiscuous and all multicast flow rules. */
+	if (!broadcast) {
+		for (flow = LIST_FIRST(&priv->flows);
+		     flow && flow->internal;
+		     flow = LIST_NEXT(flow, next)) {
+			if (priv->dev->data->promiscuous) {
+				if (flow->promisc)
+					break;
+			} else {
+				assert(priv->dev->data->all_multicast);
+				if (flow->allmulti)
+					break;
+			}
+		}
+		if (flow && flow->internal) {
+			assert(flow->rss);
+			if (flow->rss->queues != queues ||
+			    memcmp(flow->rss->queue_id, rss_conf->queue,
+				   queues * sizeof(flow->rss->queue_id[0])))
+				flow = NULL;
+		}
+		if (!flow || !flow->internal) {
+			/* Not found, create a new flow rule. */
+			if (priv->dev->data->promiscuous) {
+				pattern[1].spec = NULL;
+				pattern[1].mask = NULL;
+			} else {
+				assert(priv->dev->data->all_multicast);
+				pattern[1].spec = &eth_allmulti;
+				pattern[1].mask = &eth_allmulti;
+			}
+			pattern[2] = pattern[3];
+			flow = mlx4_flow_create(priv->dev, &attr, pattern,
+						actions, error);
+			if (!flow) {
+				err = -rte_errno;
+				goto error;
+			}
+		}
+		assert(flow->promisc || flow->allmulti);
+		flow->select = 1;
+	}
+error:
+	/* Clear selection and clean up stale internal flow rules. */
+	flow = LIST_FIRST(&priv->flows);
+	while (flow && flow->internal) {
+		struct rte_flow *next = LIST_NEXT(flow, next);
+
+		if (!flow->select)
+			claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
+		else
+			flow->select = 0;
+		flow = next;
+	}
+	return err;
 }
 
 /**
- * Destroy all flows.
+ * Synchronize flow rules.
  *
- * @see rte_flow_flush()
- * @see rte_flow_ops
+ * This function synchronizes flow rules with the state of the device by
+ * taking into account isolated mode and whether target queues are
+ * configured.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 int
-mlx4_flow_flush(struct rte_eth_dev *dev,
-		struct rte_flow_error *error)
+mlx4_flow_sync(struct priv *priv, struct rte_flow_error *error)
 {
-	struct priv *priv = dev->data->dev_private;
+	struct rte_flow *flow;
+	int ret;
 
-	(void)error;
-	priv_lock(priv);
-	priv_flow_flush(priv);
-	priv_unlock(priv);
+	/* Internal flow rules are guaranteed to come first in the list. */
+	if (priv->isolated) {
+		/*
+		 * Get rid of them in isolated mode, stop at the first
+		 * non-internal rule found.
+		 */
+		for (flow = LIST_FIRST(&priv->flows);
+		     flow && flow->internal;
+		     flow = LIST_FIRST(&priv->flows))
+			claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
+	} else {
+		/* Refresh internal rules. */
+		ret = mlx4_flow_internal(priv, error);
+		if (ret)
+			return ret;
+	}
+	/* Toggle the remaining flow rules . */
+	LIST_FOREACH(flow, &priv->flows, next) {
+		ret = mlx4_flow_toggle(priv, flow, priv->started, error);
+		if (ret)
+			return ret;
+	}
+	if (!priv->started)
+		assert(!priv->drop);
 	return 0;
 }
 
 /**
- * Remove all flows.
+ * Clean up all flow rules.
  *
- * Called by dev_stop() to remove all flows.
+ * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
+ * rules regardless of whether they are internal or user-configured.
  *
  * @param priv
  *   Pointer to private structure.
  */
 void
-mlx4_priv_flow_stop(struct priv *priv)
+mlx4_flow_clean(struct priv *priv)
 {
 	struct rte_flow *flow;
 
-	for (flow = LIST_FIRST(&priv->flows);
-	     flow;
-	     flow = LIST_NEXT(flow, next)) {
-		claim_zero(ibv_destroy_flow(flow->ibv_flow));
-		flow->ibv_flow = NULL;
-		DEBUG("Flow %p removed", (void *)flow);
-	}
-	mlx4_flow_destroy_drop_queue(priv);
+	while ((flow = LIST_FIRST(&priv->flows)))
+		mlx4_flow_destroy(priv->dev, flow, NULL);
+	assert(LIST_EMPTY(&priv->rss));
 }
 
+static const struct rte_flow_ops mlx4_flow_ops = {
+	.validate = mlx4_flow_validate,
+	.create = mlx4_flow_create,
+	.destroy = mlx4_flow_destroy,
+	.flush = mlx4_flow_flush,
+	.isolate = mlx4_flow_isolate,
+};
+
 /**
- * Add all flows.
+ * Manage filter operations.
  *
- * @param priv
- *   Pointer to private structure.
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param filter_type
+ *   Filter type.
+ * @param filter_op
+ *   Operation to perform.
+ * @param arg
+ *   Pointer to operation-specific structure.
  *
  * @return
- *   0 on success, a errno value otherwise and rte_errno is set.
+ *   0 on success, negative errno value otherwise and rte_errno is set.
  */
 int
-mlx4_priv_flow_start(struct priv *priv)
+mlx4_filter_ctrl(struct rte_eth_dev *dev,
+		 enum rte_filter_type filter_type,
+		 enum rte_filter_op filter_op,
+		 void *arg)
 {
-	int ret;
-	struct ibv_qp *qp;
-	struct rte_flow *flow;
-
-	ret = mlx4_flow_create_drop_queue(priv);
-	if (ret)
-		return -1;
-	for (flow = LIST_FIRST(&priv->flows);
-	     flow;
-	     flow = LIST_NEXT(flow, next)) {
-		qp = flow->qp ? flow->qp : priv->flow_drop_queue->qp;
-		flow->ibv_flow = ibv_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);
+	switch (filter_type) {
+	case RTE_ETH_FILTER_GENERIC:
+		if (filter_op != RTE_ETH_FILTER_GET)
+			break;
+		*(const void **)arg = &mlx4_flow_ops;
+		return 0;
+	default:
+		ERROR("%p: filter type (%d) not supported",
+		      (void *)dev, filter_type);
+		break;
 	}
-	return 0;
+	rte_errno = ENOTSUP;
+	return -rte_errno;
 }