net/enic: add flow implementation based on Flow Manager API
[dpdk.git] / drivers / net / enic / enic_fm_flow.c
diff --git a/drivers/net/enic/enic_fm_flow.c b/drivers/net/enic/enic_fm_flow.c
new file mode 100644 (file)
index 0000000..a1598b8
--- /dev/null
@@ -0,0 +1,2459 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2008-2019 Cisco Systems, Inc.  All rights reserved.
+ */
+
+#include <errno.h>
+#include <stdint.h>
+#include <rte_log.h>
+#include <rte_ethdev_driver.h>
+#include <rte_flow_driver.h>
+#include <rte_ether.h>
+#include <rte_ip.h>
+#include <rte_udp.h>
+
+#include "enic_compat.h"
+#include "enic.h"
+#include "vnic_dev.h"
+#include "vnic_nic.h"
+
+#define IP_DEFTTL  64   /* from RFC 1340. */
+#define IP_VERSION 0x40
+#define IP_HDRLEN  0x05 /* default IP header length == five 32-bits words. */
+#define IP_VHL_DEF (IP_VERSION | IP_HDRLEN)
+#define IP6_VTC_FLOW 0x60000000
+
+/* Highest Item type supported by Flowman */
+#define FM_MAX_ITEM_TYPE RTE_FLOW_ITEM_TYPE_VXLAN
+
+/* Up to 1024 TCAM entries */
+#define FM_MAX_TCAM_TABLE_SIZE 1024
+
+/* Up to 4096 entries per exact match table */
+#define FM_MAX_EXACT_TABLE_SIZE 4096
+
+/* Number of counters to increase on for each increment */
+#define FM_COUNTERS_EXPAND  100
+
+#define FM_INVALID_HANDLE 0
+
+/*
+ * Flow exact match tables (FET) in the VIC and rte_flow groups.
+ * Use a simple scheme to map groups to tables.
+ * Group 0 uses the single TCAM tables, one for each direction.
+ * Group 1, 2, ... uses its own exact match table.
+ *
+ * The TCAM tables are allocated upfront during init.
+ *
+ * Exact match tables are allocated on demand. 3 paths that lead allocations.
+ *
+ * 1. Add a flow that jumps from group 0 to group N.
+ *
+ * If N does not exist, we allocate an exact match table for it, using
+ * a dummy key. A key is required for the table.
+ *
+ * 2. Add a flow that uses group N.
+ *
+ * If N does not exist, we allocate an exact match table for it, using
+ * the flow's key. Subsequent flows to the same group all should have
+ * the same key.
+ *
+ * Without a jump flow to N, N is not reachable in hardware. No packets
+ * reach N and match.
+ *
+ * 3. Add a flow to an empty group N.
+ *
+ * N has been created via (1) and the dummy key. We free that table, allocate
+ * a new table using the new flow's key. Also re-do the existing jump flow to
+ * point to the new table.
+ */
+#define FM_TCAM_RTE_GROUP 0
+
+struct enic_fm_fet {
+       TAILQ_ENTRY(enic_fm_fet) list;
+       uint32_t group; /* rte_flow group ID */
+       uint64_t handle; /* Exact match table handle from flowman */
+       uint8_t ingress;
+       uint8_t default_key;
+       int ref; /* Reference count via get/put */
+       struct fm_key_template key; /* Key associated with the table */
+};
+
+struct enic_fm_counter {
+       SLIST_ENTRY(enic_fm_counter) next;
+       uint32_t handle;
+};
+
+/* rte_flow.fm */
+struct enic_fm_flow {
+       bool counter_valid;
+       uint64_t entry_handle;
+       uint64_t action_handle;
+       struct enic_fm_counter *counter;
+       struct enic_fm_fet *fet;
+};
+
+struct enic_fm_jump_flow {
+       TAILQ_ENTRY(enic_fm_jump_flow) list;
+       struct rte_flow *flow;
+       uint32_t group;
+       struct fm_tcam_match_entry match;
+       struct fm_action action;
+};
+
+/*
+ * Flowman uses host memory for commands. This structure is allocated
+ * in DMA-able memory.
+ */
+union enic_flowman_cmd_mem {
+       struct fm_tcam_match_table fm_tcam_match_table;
+       struct fm_exact_match_table fm_exact_match_table;
+       struct fm_tcam_match_entry fm_tcam_match_entry;
+       struct fm_exact_match_entry fm_exact_match_entry;
+       struct fm_action fm_action;
+};
+
+struct enic_flowman {
+       struct enic *enic;
+       /* Command buffer */
+       struct {
+               union enic_flowman_cmd_mem *va;
+               dma_addr_t pa;
+       } cmd;
+       /* TCAM tables allocated upfront, used for group 0 */
+       uint64_t ig_tcam_hndl;
+       uint64_t eg_tcam_hndl;
+       /* Counters */
+       SLIST_HEAD(enic_free_counters, enic_fm_counter) counters;
+       void *counter_stack;
+       uint32_t counters_alloced;
+       /* Exact match tables for groups != 0, dynamically allocated */
+       TAILQ_HEAD(fet_list, enic_fm_fet) fet_list;
+       /*
+        * Default exact match tables used for jump actions to
+        * non-existent groups.
+        */
+       struct enic_fm_fet *default_eg_fet;
+       struct enic_fm_fet *default_ig_fet;
+       /* Flows that jump to the default table above */
+       TAILQ_HEAD(jump_flow_list, enic_fm_jump_flow) jump_list;
+       /*
+        * Scratch data used during each invocation of flow_create
+        * and flow_validate.
+        */
+       struct enic_fm_fet *fet;
+       struct fm_tcam_match_entry tcam_entry;
+       struct fm_action action;
+       struct fm_action action_tmp; /* enic_fm_reorder_action_op */
+       int action_op_count;
+};
+
+static int enic_fm_tbl_free(struct enic_flowman *fm, uint64_t handle);
+
+/*
+ * Common arguments passed to copy_item functions. Use this structure
+ * so we can easily add new arguments.
+ * item: Item specification.
+ * fm_tcam_entry: Flowman TCAM match entry.
+ * header_level: 0 for outer header, 1 for inner header.
+ */
+struct copy_item_args {
+       const struct rte_flow_item *item;
+       struct fm_tcam_match_entry *fm_tcam_entry;
+       uint8_t header_level;
+};
+
+/* functions for copying items into flowman match */
+typedef int (enic_copy_item_fn)(struct copy_item_args *arg);
+
+/* Info about how to copy items into flowman match */
+struct enic_fm_items {
+       /* Function for copying and validating an item. */
+       enic_copy_item_fn * const copy_item;
+       /* List of valid previous items. */
+       const enum rte_flow_item_type * const prev_items;
+       /*
+        * True if it's OK for this item to be the first item. For some NIC
+        * versions, it's invalid to start the stack above layer 3.
+        */
+       const uint8_t valid_start_item;
+};
+
+static enic_copy_item_fn enic_fm_copy_item_eth;
+static enic_copy_item_fn enic_fm_copy_item_ipv4;
+static enic_copy_item_fn enic_fm_copy_item_ipv6;
+static enic_copy_item_fn enic_fm_copy_item_raw;
+static enic_copy_item_fn enic_fm_copy_item_sctp;
+static enic_copy_item_fn enic_fm_copy_item_tcp;
+static enic_copy_item_fn enic_fm_copy_item_udp;
+static enic_copy_item_fn enic_fm_copy_item_vlan;
+static enic_copy_item_fn enic_fm_copy_item_vxlan;
+
+/* Ingress actions */
+static const enum rte_flow_action_type enic_fm_supported_ig_actions[] = {
+       RTE_FLOW_ACTION_TYPE_COUNT,
+       RTE_FLOW_ACTION_TYPE_DROP,
+       RTE_FLOW_ACTION_TYPE_FLAG,
+       RTE_FLOW_ACTION_TYPE_JUMP,
+       RTE_FLOW_ACTION_TYPE_MARK,
+       RTE_FLOW_ACTION_TYPE_PORT_ID,
+       RTE_FLOW_ACTION_TYPE_PASSTHRU,
+       RTE_FLOW_ACTION_TYPE_QUEUE,
+       RTE_FLOW_ACTION_TYPE_RSS,
+       RTE_FLOW_ACTION_TYPE_VOID,
+       RTE_FLOW_ACTION_TYPE_VXLAN_ENCAP,
+       RTE_FLOW_ACTION_TYPE_VXLAN_DECAP,
+       RTE_FLOW_ACTION_TYPE_END, /* END must be the last entry */
+};
+
+/* Egress actions */
+static const enum rte_flow_action_type enic_fm_supported_eg_actions[] = {
+       RTE_FLOW_ACTION_TYPE_COUNT,
+       RTE_FLOW_ACTION_TYPE_DROP,
+       RTE_FLOW_ACTION_TYPE_JUMP,
+       RTE_FLOW_ACTION_TYPE_PASSTHRU,
+       RTE_FLOW_ACTION_TYPE_VOID,
+       RTE_FLOW_ACTION_TYPE_VXLAN_ENCAP,
+       RTE_FLOW_ACTION_TYPE_END,
+};
+
+static const struct enic_fm_items enic_fm_items[] = {
+       [RTE_FLOW_ITEM_TYPE_RAW] = {
+               .copy_item = enic_fm_copy_item_raw,
+               .valid_start_item = 0,
+               .prev_items = (const enum rte_flow_item_type[]) {
+                              RTE_FLOW_ITEM_TYPE_UDP,
+                              RTE_FLOW_ITEM_TYPE_END,
+               },
+       },
+       [RTE_FLOW_ITEM_TYPE_ETH] = {
+               .copy_item = enic_fm_copy_item_eth,
+               .valid_start_item = 1,
+               .prev_items = (const enum rte_flow_item_type[]) {
+                              RTE_FLOW_ITEM_TYPE_END,
+               },
+       },
+       [RTE_FLOW_ITEM_TYPE_VLAN] = {
+               .copy_item = enic_fm_copy_item_vlan,
+               .valid_start_item = 1,
+               .prev_items = (const enum rte_flow_item_type[]) {
+                              RTE_FLOW_ITEM_TYPE_ETH,
+                              RTE_FLOW_ITEM_TYPE_END,
+               },
+       },
+       [RTE_FLOW_ITEM_TYPE_IPV4] = {
+               .copy_item = enic_fm_copy_item_ipv4,
+               .valid_start_item = 1,
+               .prev_items = (const enum rte_flow_item_type[]) {
+                              RTE_FLOW_ITEM_TYPE_ETH,
+                              RTE_FLOW_ITEM_TYPE_VLAN,
+                              RTE_FLOW_ITEM_TYPE_END,
+               },
+       },
+       [RTE_FLOW_ITEM_TYPE_IPV6] = {
+               .copy_item = enic_fm_copy_item_ipv6,
+               .valid_start_item = 1,
+               .prev_items = (const enum rte_flow_item_type[]) {
+                              RTE_FLOW_ITEM_TYPE_ETH,
+                              RTE_FLOW_ITEM_TYPE_VLAN,
+                              RTE_FLOW_ITEM_TYPE_END,
+               },
+       },
+       [RTE_FLOW_ITEM_TYPE_UDP] = {
+               .copy_item = enic_fm_copy_item_udp,
+               .valid_start_item = 1,
+               .prev_items = (const enum rte_flow_item_type[]) {
+                              RTE_FLOW_ITEM_TYPE_IPV4,
+                              RTE_FLOW_ITEM_TYPE_IPV6,
+                              RTE_FLOW_ITEM_TYPE_END,
+               },
+       },
+       [RTE_FLOW_ITEM_TYPE_TCP] = {
+               .copy_item = enic_fm_copy_item_tcp,
+               .valid_start_item = 1,
+               .prev_items = (const enum rte_flow_item_type[]) {
+                              RTE_FLOW_ITEM_TYPE_IPV4,
+                              RTE_FLOW_ITEM_TYPE_IPV6,
+                              RTE_FLOW_ITEM_TYPE_END,
+               },
+       },
+       [RTE_FLOW_ITEM_TYPE_SCTP] = {
+               .copy_item = enic_fm_copy_item_sctp,
+               .valid_start_item = 0,
+               .prev_items = (const enum rte_flow_item_type[]) {
+                              RTE_FLOW_ITEM_TYPE_IPV4,
+                              RTE_FLOW_ITEM_TYPE_IPV6,
+                              RTE_FLOW_ITEM_TYPE_END,
+               },
+       },
+       [RTE_FLOW_ITEM_TYPE_VXLAN] = {
+               .copy_item = enic_fm_copy_item_vxlan,
+               .valid_start_item = 1,
+               .prev_items = (const enum rte_flow_item_type[]) {
+                              RTE_FLOW_ITEM_TYPE_UDP,
+                              RTE_FLOW_ITEM_TYPE_END,
+               },
+       },
+};
+
+static int
+enic_fm_copy_item_eth(struct copy_item_args *arg)
+{
+       const struct rte_flow_item *item = arg->item;
+       const struct rte_flow_item_eth *spec = item->spec;
+       const struct rte_flow_item_eth *mask = item->mask;
+       const uint8_t lvl = arg->header_level;
+       struct fm_tcam_match_entry *entry = arg->fm_tcam_entry;
+       struct fm_header_set *fm_data, *fm_mask;
+
+       ENICPMD_FUNC_TRACE();
+       /* Match all if no spec */
+       if (!spec)
+               return 0;
+       if (!mask)
+               mask = &rte_flow_item_eth_mask;
+       fm_data = &entry->ftm_data.fk_hdrset[lvl];
+       fm_mask = &entry->ftm_mask.fk_hdrset[lvl];
+       fm_data->fk_header_select |= FKH_ETHER;
+       fm_mask->fk_header_select |= FKH_ETHER;
+       memcpy(&fm_data->l2.eth, spec, sizeof(*spec));
+       memcpy(&fm_mask->l2.eth, mask, sizeof(*mask));
+       return 0;
+}
+
+static int
+enic_fm_copy_item_vlan(struct copy_item_args *arg)
+{
+       const struct rte_flow_item *item = arg->item;
+       const struct rte_flow_item_vlan *spec = item->spec;
+       const struct rte_flow_item_vlan *mask = item->mask;
+       const uint8_t lvl = arg->header_level;
+       struct fm_tcam_match_entry *entry = arg->fm_tcam_entry;
+       struct fm_header_set *fm_data, *fm_mask;
+       struct rte_ether_hdr *eth_mask;
+       struct rte_ether_hdr *eth_val;
+       uint32_t meta;
+
+       ENICPMD_FUNC_TRACE();
+       fm_data = &entry->ftm_data.fk_hdrset[lvl];
+       fm_mask = &entry->ftm_mask.fk_hdrset[lvl];
+       /* Outer and inner packet vlans need different flags */
+       meta = FKM_VLAN_PRES;
+       if (lvl > 0)
+               meta = FKM_QTAG;
+       fm_data->fk_metadata |= meta;
+       fm_mask->fk_metadata |= meta;
+
+       /* Match all if no spec */
+       if (!spec)
+               return 0;
+       if (!mask)
+               mask = &rte_flow_item_vlan_mask;
+
+       eth_mask = (void *)&fm_mask->l2.eth;
+       eth_val = (void *)&fm_data->l2.eth;
+
+       /* Outer TPID cannot be matched */
+       if (eth_mask->ether_type)
+               return -ENOTSUP;
+
+       /*
+        * When packet matching, the VIC always compares vlan-stripped
+        * L2, regardless of vlan stripping settings. So, the inner type
+        * from vlan becomes the ether type of the eth header.
+        */
+       eth_mask->ether_type = mask->inner_type;
+       eth_val->ether_type = spec->inner_type;
+       fm_data->fk_header_select |= FKH_ETHER | FKH_QTAG;
+       fm_mask->fk_header_select |= FKH_ETHER | FKH_QTAG;
+       fm_data->fk_vlan = rte_be_to_cpu_16(spec->tci);
+       fm_mask->fk_vlan = rte_be_to_cpu_16(mask->tci);
+       return 0;
+}
+
+static int
+enic_fm_copy_item_ipv4(struct copy_item_args *arg)
+{
+       const struct rte_flow_item *item = arg->item;
+       const struct rte_flow_item_ipv4 *spec = item->spec;
+       const struct rte_flow_item_ipv4 *mask = item->mask;
+       const uint8_t lvl = arg->header_level;
+       struct fm_tcam_match_entry *entry = arg->fm_tcam_entry;
+       struct fm_header_set *fm_data, *fm_mask;
+
+       ENICPMD_FUNC_TRACE();
+       fm_data = &entry->ftm_data.fk_hdrset[lvl];
+       fm_mask = &entry->ftm_mask.fk_hdrset[lvl];
+       fm_data->fk_metadata |= FKM_IPV4;
+       fm_mask->fk_metadata |= FKM_IPV4;
+
+       if (!spec)
+               return 0;
+       if (!mask)
+               mask = &rte_flow_item_ipv4_mask;
+
+       fm_data->fk_header_select |= FKH_IPV4;
+       fm_mask->fk_header_select |= FKH_IPV4;
+       memcpy(&fm_data->l3.ip4, spec, sizeof(*spec));
+       memcpy(&fm_mask->l3.ip4, mask, sizeof(*mask));
+       return 0;
+}
+
+static int
+enic_fm_copy_item_ipv6(struct copy_item_args *arg)
+{
+       const struct rte_flow_item *item = arg->item;
+       const struct rte_flow_item_ipv6 *spec = item->spec;
+       const struct rte_flow_item_ipv6 *mask = item->mask;
+       const uint8_t lvl = arg->header_level;
+       struct fm_tcam_match_entry *entry = arg->fm_tcam_entry;
+       struct fm_header_set *fm_data, *fm_mask;
+
+       ENICPMD_FUNC_TRACE();
+       fm_data = &entry->ftm_data.fk_hdrset[lvl];
+       fm_mask = &entry->ftm_mask.fk_hdrset[lvl];
+       fm_data->fk_metadata |= FKM_IPV6;
+       fm_mask->fk_metadata |= FKM_IPV6;
+
+       if (!spec)
+               return 0;
+       if (!mask)
+               mask = &rte_flow_item_ipv6_mask;
+
+       fm_data->fk_header_select |= FKH_IPV6;
+       fm_mask->fk_header_select |= FKH_IPV6;
+       memcpy(&fm_data->l3.ip6, spec, sizeof(*spec));
+       memcpy(&fm_mask->l3.ip6, mask, sizeof(*mask));
+       return 0;
+}
+
+static int
+enic_fm_copy_item_udp(struct copy_item_args *arg)
+{
+       const struct rte_flow_item *item = arg->item;
+       const struct rte_flow_item_udp *spec = item->spec;
+       const struct rte_flow_item_udp *mask = item->mask;
+       const uint8_t lvl = arg->header_level;
+       struct fm_tcam_match_entry *entry = arg->fm_tcam_entry;
+       struct fm_header_set *fm_data, *fm_mask;
+
+       ENICPMD_FUNC_TRACE();
+       fm_data = &entry->ftm_data.fk_hdrset[lvl];
+       fm_mask = &entry->ftm_mask.fk_hdrset[lvl];
+       fm_data->fk_metadata |= FKM_UDP;
+       fm_mask->fk_metadata |= FKM_UDP;
+
+       if (!spec)
+               return 0;
+       if (!mask)
+               mask = &rte_flow_item_udp_mask;
+
+       fm_data->fk_header_select |= FKH_UDP;
+       fm_mask->fk_header_select |= FKH_UDP;
+       memcpy(&fm_data->l4.udp, spec, sizeof(*spec));
+       memcpy(&fm_mask->l4.udp, mask, sizeof(*mask));
+       return 0;
+}
+
+static int
+enic_fm_copy_item_tcp(struct copy_item_args *arg)
+{
+       const struct rte_flow_item *item = arg->item;
+       const struct rte_flow_item_tcp *spec = item->spec;
+       const struct rte_flow_item_tcp *mask = item->mask;
+       const uint8_t lvl = arg->header_level;
+       struct fm_tcam_match_entry *entry = arg->fm_tcam_entry;
+       struct fm_header_set *fm_data, *fm_mask;
+
+       ENICPMD_FUNC_TRACE();
+       fm_data = &entry->ftm_data.fk_hdrset[lvl];
+       fm_mask = &entry->ftm_mask.fk_hdrset[lvl];
+       fm_data->fk_metadata |= FKM_TCP;
+       fm_mask->fk_metadata |= FKM_TCP;
+
+       if (!spec)
+               return 0;
+       if (!mask)
+               mask = &rte_flow_item_tcp_mask;
+
+       fm_data->fk_header_select |= FKH_TCP;
+       fm_mask->fk_header_select |= FKH_TCP;
+       memcpy(&fm_data->l4.tcp, spec, sizeof(*spec));
+       memcpy(&fm_mask->l4.tcp, mask, sizeof(*mask));
+       return 0;
+}
+
+static int
+enic_fm_copy_item_sctp(struct copy_item_args *arg)
+{
+       const struct rte_flow_item *item = arg->item;
+       const struct rte_flow_item_sctp *spec = item->spec;
+       const struct rte_flow_item_sctp *mask = item->mask;
+       const uint8_t lvl = arg->header_level;
+       struct fm_tcam_match_entry *entry = arg->fm_tcam_entry;
+       struct fm_header_set *fm_data, *fm_mask;
+       uint8_t *ip_proto_mask = NULL;
+       uint8_t *ip_proto = NULL;
+       uint32_t l3_fkh;
+
+       ENICPMD_FUNC_TRACE();
+       fm_data = &entry->ftm_data.fk_hdrset[lvl];
+       fm_mask = &entry->ftm_mask.fk_hdrset[lvl];
+       /*
+        * The NIC filter API has no flags for "match sctp", so explicitly
+        * set the protocol number in the IP pattern.
+        */
+       if (fm_data->fk_metadata & FKM_IPV4) {
+               struct rte_ipv4_hdr *ip;
+               ip = (struct rte_ipv4_hdr *)&fm_mask->l3.ip4;
+               ip_proto_mask = &ip->next_proto_id;
+               ip = (struct rte_ipv4_hdr *)&fm_data->l3.ip4;
+               ip_proto = &ip->next_proto_id;
+               l3_fkh = FKH_IPV4;
+       } else if (fm_data->fk_metadata & FKM_IPV6) {
+               struct rte_ipv6_hdr *ip;
+               ip = (struct rte_ipv6_hdr *)&fm_mask->l3.ip6;
+               ip_proto_mask = &ip->proto;
+               ip = (struct rte_ipv6_hdr *)&fm_data->l3.ip6;
+               ip_proto = &ip->proto;
+               l3_fkh = FKH_IPV6;
+       } else {
+               /* Need IPv4/IPv6 pattern first */
+               return -EINVAL;
+       }
+       *ip_proto = IPPROTO_SCTP;
+       *ip_proto_mask = 0xff;
+       fm_data->fk_header_select |= l3_fkh;
+       fm_mask->fk_header_select |= l3_fkh;
+
+       if (!spec)
+               return 0;
+       if (!mask)
+               mask = &rte_flow_item_sctp_mask;
+
+       fm_data->fk_header_select |= FKH_L4RAW;
+       fm_mask->fk_header_select |= FKH_L4RAW;
+       memcpy(fm_data->l4.rawdata, spec, sizeof(*spec));
+       memcpy(fm_mask->l4.rawdata, mask, sizeof(*mask));
+       return 0;
+}
+
+static int
+enic_fm_copy_item_vxlan(struct copy_item_args *arg)
+{
+       const struct rte_flow_item *item = arg->item;
+       const struct rte_flow_item_vxlan *spec = item->spec;
+       const struct rte_flow_item_vxlan *mask = item->mask;
+       struct fm_tcam_match_entry *entry = arg->fm_tcam_entry;
+       struct fm_header_set *fm_data, *fm_mask;
+
+       ENICPMD_FUNC_TRACE();
+       /* Only 2 header levels (outer and inner) allowed */
+       if (arg->header_level > 0)
+               return -EINVAL;
+
+       fm_data = &entry->ftm_data.fk_hdrset[0];
+       fm_mask = &entry->ftm_mask.fk_hdrset[0];
+       fm_data->fk_metadata |= FKM_VXLAN;
+       fm_mask->fk_metadata |= FKM_VXLAN;
+       /* items from here on out are inner header items */
+       arg->header_level = 1;
+
+       /* Match all if no spec */
+       if (!spec)
+               return 0;
+       if (!mask)
+               mask = &rte_flow_item_vxlan_mask;
+
+       fm_data->fk_header_select |= FKH_VXLAN;
+       fm_mask->fk_header_select |= FKH_VXLAN;
+       memcpy(&fm_data->vxlan, spec, sizeof(*spec));
+       memcpy(&fm_mask->vxlan, mask, sizeof(*mask));
+       return 0;
+}
+
+/*
+ * Currently, raw pattern match is very limited. It is intended for matching
+ * UDP tunnel header (e.g. vxlan or geneve).
+ */
+static int
+enic_fm_copy_item_raw(struct copy_item_args *arg)
+{
+       const struct rte_flow_item *item = arg->item;
+       const struct rte_flow_item_raw *spec = item->spec;
+       const struct rte_flow_item_raw *mask = item->mask;
+       const uint8_t lvl = arg->header_level;
+       struct fm_tcam_match_entry *entry = arg->fm_tcam_entry;
+       struct fm_header_set *fm_data, *fm_mask;
+
+       ENICPMD_FUNC_TRACE();
+       /* Cannot be used for inner packet */
+       if (lvl > 0)
+               return -EINVAL;
+       /* Need both spec and mask */
+       if (!spec || !mask)
+               return -EINVAL;
+       /* Only supports relative with offset 0 */
+       if (!spec->relative || spec->offset != 0 || spec->search ||
+           spec->limit)
+               return -EINVAL;
+       /* Need non-null pattern that fits within the NIC's filter pattern */
+       if (spec->length == 0 ||
+           spec->length + sizeof(struct rte_udp_hdr) > FM_LAYER_SIZE ||
+           !spec->pattern || !mask->pattern)
+               return -EINVAL;
+       /*
+        * Mask fields, including length, are often set to zero. Assume that
+        * means "same as spec" to avoid breaking existing apps. If length
+        * is not zero, then it should be >= spec length.
+        *
+        * No more pattern follows this, so append to the L4 layer instead of
+        * L5 to work with both recent and older VICs.
+        */
+       if (mask->length != 0 && mask->length < spec->length)
+               return -EINVAL;
+
+       fm_data = &entry->ftm_data.fk_hdrset[lvl];
+       fm_mask = &entry->ftm_mask.fk_hdrset[lvl];
+       fm_data->fk_header_select |= FKH_L4RAW;
+       fm_mask->fk_header_select |= FKH_L4RAW;
+       fm_data->fk_header_select &= ~FKH_UDP;
+       fm_mask->fk_header_select &= ~FKH_UDP;
+       memcpy(fm_data->l4.rawdata + sizeof(struct rte_udp_hdr),
+              spec->pattern, spec->length);
+       memcpy(fm_mask->l4.rawdata + sizeof(struct rte_udp_hdr),
+              mask->pattern, spec->length);
+       return 0;
+}
+
+static int
+enic_fet_alloc(struct enic_flowman *fm, uint8_t ingress,
+              struct fm_key_template *key, int entries,
+              struct enic_fm_fet **fet_out)
+{
+       struct fm_exact_match_table *cmd;
+       struct fm_header_set *hdr;
+       struct enic_fm_fet *fet;
+       u64 args[3];
+       int ret;
+
+       ENICPMD_FUNC_TRACE();
+       fet = calloc(1, sizeof(struct enic_fm_fet));
+       if (fet == NULL)
+               return -ENOMEM;
+       cmd = &fm->cmd.va->fm_exact_match_table;
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->fet_direction = ingress ? FM_INGRESS : FM_EGRESS;
+       cmd->fet_stage = FM_STAGE_LAST;
+       cmd->fet_max_entries = entries ? entries : FM_MAX_EXACT_TABLE_SIZE;
+       if (key == NULL) {
+               hdr = &cmd->fet_key.fk_hdrset[0];
+               memset(hdr, 0, sizeof(*hdr));
+               hdr->fk_header_select = FKH_IPV4 | FKH_UDP;
+               hdr->l3.ip4.fk_saddr = 0xFFFFFFFF;
+               hdr->l3.ip4.fk_daddr = 0xFFFFFFFF;
+               hdr->l4.udp.fk_source = 0xFFFF;
+               hdr->l4.udp.fk_dest = 0xFFFF;
+               fet->default_key = 1;
+       } else {
+               memcpy(&cmd->fet_key, key, sizeof(*key));
+               memcpy(&fet->key, key, sizeof(*key));
+               fet->default_key = 0;
+       }
+       cmd->fet_key.fk_packet_tag = 1;
+
+       args[0] = FM_EXACT_TABLE_ALLOC;
+       args[1] = fm->cmd.pa;
+       ret = vnic_dev_flowman_cmd(fm->enic->vdev, args, 2);
+       if (ret) {
+               ENICPMD_LOG(ERR, "cannot alloc exact match table: rc=%d", ret);
+               free(fet);
+               return ret;
+       }
+       fet->handle = args[0];
+       fet->ingress = ingress;
+       ENICPMD_LOG(DEBUG, "allocated exact match table: handle=0x%" PRIx64,
+                   fet->handle);
+       *fet_out = fet;
+       return 0;
+}
+
+static void
+enic_fet_free(struct enic_flowman *fm, struct enic_fm_fet *fet)
+{
+       ENICPMD_FUNC_TRACE();
+       enic_fm_tbl_free(fm, fet->handle);
+       if (!fet->default_key)
+               TAILQ_REMOVE(&fm->fet_list, fet, list);
+       free(fet);
+}
+
+/*
+ * Get the exact match table for the given combination of
+ * <group, ingress, key>. Allocate one on the fly as necessary.
+ */
+static int
+enic_fet_get(struct enic_flowman *fm,
+            uint32_t group,
+            uint8_t ingress,
+            struct fm_key_template *key,
+            struct enic_fm_fet **fet_out,
+            struct rte_flow_error *error)
+{
+       struct enic_fm_fet *fet;
+
+       ENICPMD_FUNC_TRACE();
+       /* See if we already have this table open */
+       TAILQ_FOREACH(fet, &fm->fet_list, list) {
+               if (fet->group == group && fet->ingress == ingress)
+                       break;
+       }
+       if (fet == NULL) {
+               /* Jumping to a non-existing group? Use the default table */
+               if (key == NULL) {
+                       fet = ingress ? fm->default_ig_fet : fm->default_eg_fet;
+               } else if (enic_fet_alloc(fm, ingress, key, 0, &fet)) {
+                       return rte_flow_error_set(error, EINVAL,
+                               RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                               NULL, "enic: cannot get exact match table");
+               }
+               fet->group = group;
+               /* Default table is never on the open table list */
+               if (!fet->default_key)
+                       TAILQ_INSERT_HEAD(&fm->fet_list, fet, list);
+       }
+       fet->ref++;
+       *fet_out = fet;
+       ENICPMD_LOG(DEBUG, "fet_get: %s %s group=%u ref=%u",
+                   fet->default_key ? "default" : "",
+                   fet->ingress ? "ingress" : "egress",
+                   fet->group, fet->ref);
+       return 0;
+}
+
+static void
+enic_fet_put(struct enic_flowman *fm, struct enic_fm_fet *fet)
+{
+       ENICPMD_FUNC_TRACE();
+       RTE_ASSERT(fet->ref > 0);
+       fet->ref--;
+       ENICPMD_LOG(DEBUG, "fet_put: %s %s group=%u ref=%u",
+                   fet->default_key ? "default" : "",
+                   fet->ingress ? "ingress" : "egress",
+                   fet->group, fet->ref);
+       if (fet->ref == 0)
+               enic_fet_free(fm, fet);
+}
+
+/* Return 1 if current item is valid on top of the previous one. */
+static int
+fm_item_stacking_valid(enum rte_flow_item_type prev_item,
+                      const struct enic_fm_items *item_info,
+                      uint8_t is_first_item)
+{
+       enum rte_flow_item_type const *allowed_items = item_info->prev_items;
+
+       ENICPMD_FUNC_TRACE();
+       for (; *allowed_items != RTE_FLOW_ITEM_TYPE_END; allowed_items++) {
+               if (prev_item == *allowed_items)
+                       return 1;
+       }
+
+       /* This is the first item in the stack. Check if that's cool */
+       if (is_first_item && item_info->valid_start_item)
+               return 1;
+       return 0;
+}
+
+/*
+ * Build the flow manager match entry structure from the provided pattern.
+ * The pattern is validated as the items are copied.
+ */
+static int
+enic_fm_copy_entry(struct enic_flowman *fm,
+                  const struct rte_flow_item pattern[],
+                  struct rte_flow_error *error)
+{
+       const struct enic_fm_items *item_info;
+       enum rte_flow_item_type prev_item;
+       const struct rte_flow_item *item;
+       struct copy_item_args args;
+       uint8_t prev_header_level;
+       uint8_t is_first_item;
+       int ret;
+
+       ENICPMD_FUNC_TRACE();
+       item = pattern;
+       is_first_item = 1;
+       prev_item = RTE_FLOW_ITEM_TYPE_END;
+
+       args.fm_tcam_entry = &fm->tcam_entry;
+       args.header_level = 0;
+       prev_header_level = 0;
+       for (; item->type != RTE_FLOW_ITEM_TYPE_END; item++) {
+               /*
+                * Get info about how to validate and copy the item. If NULL
+                * is returned the nic does not support the item.
+                */
+               if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
+                       continue;
+
+               item_info = &enic_fm_items[item->type];
+
+               if (item->type > FM_MAX_ITEM_TYPE ||
+                   item_info->copy_item == NULL) {
+                       return rte_flow_error_set(error, ENOTSUP,
+                               RTE_FLOW_ERROR_TYPE_ITEM,
+                               NULL, "enic: unsupported item");
+               }
+
+               /* check to see if item stacking is valid */
+               if (!fm_item_stacking_valid(prev_item, item_info,
+                                           is_first_item))
+                       goto stacking_error;
+
+               args.item = item;
+               ret = item_info->copy_item(&args);
+               if (ret)
+                       goto item_not_supported;
+               /* Going from outer to inner? Treat it as a new packet start */
+               if (prev_header_level != args.header_level) {
+                       prev_item = RTE_FLOW_ITEM_TYPE_END;
+                       is_first_item = 1;
+               } else {
+                       prev_item = item->type;
+                       is_first_item = 0;
+               }
+               prev_header_level = args.header_level;
+       }
+       return 0;
+
+item_not_supported:
+       return rte_flow_error_set(error, -ret, RTE_FLOW_ERROR_TYPE_ITEM,
+                                 NULL, "enic: unsupported item type");
+
+stacking_error:
+       return rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM,
+                                 item, "enic: unsupported item stack");
+}
+
+static void
+flow_item_skip_void(const struct rte_flow_item **item)
+{
+       for ( ; ; (*item)++)
+               if ((*item)->type != RTE_FLOW_ITEM_TYPE_VOID)
+                       return;
+}
+
+static void
+append_template(void **template, uint8_t *off, const void *data, int len)
+{
+       memcpy(*template, data, len);
+       *template = (char *)*template + len;
+       *off = *off + len;
+}
+
+static int
+enic_fm_append_action_op(struct enic_flowman *fm,
+                        struct fm_action_op *fm_op,
+                        struct rte_flow_error *error)
+{
+       int count;
+
+       count = fm->action_op_count;
+       ENICPMD_LOG(DEBUG, "append action op: idx=%d op=%u",
+                   count, fm_op->fa_op);
+       if (count == FM_ACTION_OP_MAX) {
+               return rte_flow_error_set(error, EINVAL,
+                       RTE_FLOW_ERROR_TYPE_ACTION, NULL,
+                       "too many action operations");
+       }
+       fm->action.fma_action_ops[count] = *fm_op;
+       fm->action_op_count = count + 1;
+       return 0;
+}
+
+/* Steer operations need to appear before other ops */
+static void
+enic_fm_reorder_action_op(struct enic_flowman *fm)
+{
+       struct fm_action_op *dst, *dst_head, *src, *src_head;
+
+       ENICPMD_FUNC_TRACE();
+       /* Move steer ops to the front. */
+       src = fm->action.fma_action_ops;
+       src_head = src;
+       dst = fm->action_tmp.fma_action_ops;
+       dst_head = dst;
+       /* Copy steer ops to tmp */
+       while (src->fa_op != FMOP_END) {
+               if (src->fa_op == FMOP_RQ_STEER) {
+                       ENICPMD_LOG(DEBUG, "move op: %ld -> dst %ld",
+                                   (long)(src - src_head),
+                                   (long)(dst - dst_head));
+                       *dst = *src;
+                       dst++;
+               }
+               src++;
+       }
+       /* Then append non-steer ops */
+       src = src_head;
+       while (src->fa_op != FMOP_END) {
+               if (src->fa_op != FMOP_RQ_STEER) {
+                       ENICPMD_LOG(DEBUG, "move op: %ld -> dst %ld",
+                                   (long)(src - src_head),
+                                   (long)(dst - dst_head));
+                       *dst = *src;
+                       dst++;
+               }
+               src++;
+       }
+       /* Copy END */
+       *dst = *src;
+       /* Finally replace the original action with the reordered one */
+       memcpy(fm->action.fma_action_ops, fm->action_tmp.fma_action_ops,
+              sizeof(fm->action.fma_action_ops));
+}
+
+/* VXLAN decap is done via flowman compound action */
+static int
+enic_fm_copy_vxlan_decap(struct enic_flowman *fm,
+                        struct fm_tcam_match_entry *fmt,
+                        const struct rte_flow_action *action,
+                        struct rte_flow_error *error)
+{
+       struct fm_header_set *fm_data;
+       struct fm_action_op fm_op;
+
+       ENICPMD_FUNC_TRACE();
+       fm_data = &fmt->ftm_data.fk_hdrset[0];
+       if (!(fm_data->fk_metadata & FKM_VXLAN)) {
+               return rte_flow_error_set(error, EINVAL,
+                       RTE_FLOW_ERROR_TYPE_ACTION, action,
+                       "vxlan-decap: vxlan must be in pattern");
+       }
+
+       memset(&fm_op, 0, sizeof(fm_op));
+       fm_op.fa_op = FMOP_DECAP_NOSTRIP;
+       return enic_fm_append_action_op(fm, &fm_op, error);
+}
+
+/* VXLAN encap is done via flowman compound action */
+static int
+enic_fm_copy_vxlan_encap(struct enic_flowman *fm,
+                        const struct rte_flow_item *item,
+                        struct rte_flow_error *error)
+{
+       struct fm_action_op fm_op;
+       struct rte_ether_hdr *eth;
+       uint16_t *ethertype;
+       void *template;
+       uint8_t off;
+
+       ENICPMD_FUNC_TRACE();
+       memset(&fm_op, 0, sizeof(fm_op));
+       fm_op.fa_op = FMOP_ENCAP;
+       template = fm->action.fma_data;
+       off = 0;
+       /*
+        * Copy flow items to the flowman template starting L2.
+        * L2 must be ethernet.
+        */
+       flow_item_skip_void(&item);
+       if (item->type != RTE_FLOW_ITEM_TYPE_ETH)
+               return rte_flow_error_set(error, EINVAL,
+                       RTE_FLOW_ERROR_TYPE_ITEM, item,
+                       "vxlan-encap: first item should be ethernet");
+       eth = (struct rte_ether_hdr *)template;
+       ethertype = &eth->ether_type;
+       append_template(&template, &off, item->spec,
+                       sizeof(struct rte_flow_item_eth));
+       item++;
+       flow_item_skip_void(&item);
+       /* Optional VLAN */
+       if (item->type == RTE_FLOW_ITEM_TYPE_VLAN) {
+               const struct rte_flow_item_vlan *spec;
+
+               ENICPMD_LOG(DEBUG, "vxlan-encap: vlan");
+               spec = item->spec;
+               fm_op.encap.outer_vlan = rte_be_to_cpu_16(spec->tci);
+               item++;
+               flow_item_skip_void(&item);
+       }
+       /* L3 must be IPv4, IPv6 */
+       switch (item->type) {
+       case RTE_FLOW_ITEM_TYPE_IPV4:
+       {
+               struct rte_ipv4_hdr *ip4;
+
+               ENICPMD_LOG(DEBUG, "vxlan-encap: ipv4");
+               *ethertype = rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV4);
+               ip4 = (struct rte_ipv4_hdr *)template;
+               /*
+                * Offset of IPv4 length field and its initial value
+                * (IP + UDP + VXLAN) are specified in the action. The NIC
+                * will add inner packet length.
+                */
+               fm_op.encap.len1_offset = off +
+                       offsetof(struct rte_ipv4_hdr, total_length);
+               fm_op.encap.len1_delta = sizeof(struct rte_ipv4_hdr) +
+                       sizeof(struct rte_udp_hdr) +
+                       sizeof(struct rte_vxlan_hdr);
+               append_template(&template, &off, item->spec,
+                               sizeof(struct rte_ipv4_hdr));
+               ip4->version_ihl = IP_VHL_DEF;
+               if (ip4->time_to_live == 0)
+                       ip4->time_to_live = IP_DEFTTL;
+               ip4->next_proto_id = IPPROTO_UDP;
+               break;
+       }
+       case RTE_FLOW_ITEM_TYPE_IPV6:
+       {
+               struct rte_ipv6_hdr *ip6;
+
+               ENICPMD_LOG(DEBUG, "vxlan-encap: ipv6");
+               *ethertype = rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV6);
+               ip6 = (struct rte_ipv6_hdr *)template;
+               fm_op.encap.len1_offset = off +
+                       offsetof(struct rte_ipv6_hdr, payload_len);
+               fm_op.encap.len1_delta = sizeof(struct rte_udp_hdr) +
+                       sizeof(struct rte_vxlan_hdr);
+               append_template(&template, &off, item->spec,
+                               sizeof(struct rte_ipv6_hdr));
+               ip6->vtc_flow |= rte_cpu_to_be_32(IP6_VTC_FLOW);
+               if (ip6->hop_limits == 0)
+                       ip6->hop_limits = IP_DEFTTL;
+               ip6->proto = IPPROTO_UDP;
+               break;
+       }
+       default:
+               return rte_flow_error_set(error,
+                       EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
+                       "vxlan-encap: L3 must be IPv4/IPv6");
+       }
+       item++;
+       flow_item_skip_void(&item);
+
+       /* L4 is UDP */
+       if (item->type != RTE_FLOW_ITEM_TYPE_UDP)
+               return rte_flow_error_set(error, EINVAL,
+                       RTE_FLOW_ERROR_TYPE_ITEM, item,
+                       "vxlan-encap: UDP must follow IPv4/IPv6");
+       /* UDP length = UDP + VXLAN. NIC will add inner packet length. */
+       fm_op.encap.len2_offset =
+               off + offsetof(struct rte_udp_hdr, dgram_len);
+       fm_op.encap.len2_delta =
+               sizeof(struct rte_udp_hdr) + sizeof(struct rte_vxlan_hdr);
+       append_template(&template, &off, item->spec,
+                       sizeof(struct rte_udp_hdr));
+       item++;
+       flow_item_skip_void(&item);
+
+       /* Finally VXLAN */
+       if (item->type != RTE_FLOW_ITEM_TYPE_VXLAN)
+               return rte_flow_error_set(error,
+                       EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
+                       "vxlan-encap: VXLAN must follow UDP");
+       append_template(&template, &off, item->spec,
+                       sizeof(struct rte_flow_item_vxlan));
+
+       /*
+        * Fill in the rest of the action structure.
+        * Indicate that we want to encap with vxlan at packet start.
+        */
+       fm_op.encap.template_offset = 0;
+       fm_op.encap.template_len = off;
+       return enic_fm_append_action_op(fm, &fm_op, error);
+}
+
+static int
+enic_fm_find_vnic(struct enic *enic, const struct rte_pci_addr *addr,
+                 uint64_t *handle)
+{
+       uint32_t bdf;
+       u64 args[2];
+       int rc;
+
+       ENICPMD_FUNC_TRACE();
+       ENICPMD_LOG(DEBUG, "bdf=%x:%x:%x", addr->bus, addr->devid,
+                   addr->function);
+       bdf = addr->bus << 8 | addr->devid << 3 | addr->function;
+       args[0] = FM_VNIC_FIND;
+       args[1] = bdf;
+       rc = vnic_dev_flowman_cmd(enic->vdev, args, 2);
+       if (rc != 0) {
+               ENICPMD_LOG(ERR, "allocating counters rc=%d", rc);
+               return rc;
+       }
+       *handle = args[0];
+       ENICPMD_LOG(DEBUG, "found vnic: handle=0x%" PRIx64, *handle);
+       return 0;
+}
+
+/* Translate flow actions to flowman TCAM entry actions */
+static int
+enic_fm_copy_action(struct enic_flowman *fm,
+                   const struct rte_flow_action actions[],
+                   uint8_t ingress,
+                   struct rte_flow_error *error)
+{
+       enum {
+               FATE = 1 << 0,
+               MARK = 1 << 1,
+               PASSTHRU = 1 << 2,
+               COUNT = 1 << 3,
+               ENCAP = 1 << 4,
+       };
+       struct fm_tcam_match_entry *fmt;
+       struct fm_action_op fm_op;
+       struct enic *enic;
+       uint32_t overlap;
+       uint64_t vnic_h;
+       bool first_rq;
+       int ret;
+
+       ENICPMD_FUNC_TRACE();
+       fmt = &fm->tcam_entry;
+       first_rq = true;
+       enic = fm->enic;
+       overlap = 0;
+       vnic_h = 0; /* 0 = current vNIC */
+       for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
+               switch (actions->type) {
+               case RTE_FLOW_ACTION_TYPE_VOID:
+                       continue;
+               case RTE_FLOW_ACTION_TYPE_PASSTHRU: {
+                       if (overlap & PASSTHRU)
+                               goto unsupported;
+                       overlap |= PASSTHRU;
+                       break;
+               }
+               case RTE_FLOW_ACTION_TYPE_JUMP: {
+                       const struct rte_flow_action_jump *jump =
+                               actions->conf;
+                       struct enic_fm_fet *fet;
+
+                       if (overlap & FATE)
+                               goto unsupported;
+                       ret = enic_fet_get(fm, jump->group, ingress, NULL,
+                                          &fet, error);
+                       if (ret)
+                               return ret;
+                       overlap |= FATE;
+                       memset(&fm_op, 0, sizeof(fm_op));
+                       fm_op.fa_op = FMOP_EXACT_MATCH;
+                       fm_op.exact.handle = fet->handle;
+                       fm->fet = fet;
+                       ret = enic_fm_append_action_op(fm, &fm_op, error);
+                       if (ret)
+                               return ret;
+                       break;
+               }
+               case RTE_FLOW_ACTION_TYPE_MARK: {
+                       const struct rte_flow_action_mark *mark =
+                               actions->conf;
+
+                       if (overlap & MARK)
+                               goto unsupported;
+                       overlap |= MARK;
+                       if (mark->id >= ENIC_MAGIC_FILTER_ID - 1)
+                               return rte_flow_error_set(error, EINVAL,
+                                       RTE_FLOW_ERROR_TYPE_ACTION,
+                                       NULL, "invalid mark id");
+                       memset(&fm_op, 0, sizeof(fm_op));
+                       fm_op.fa_op = FMOP_MARK;
+                       fm_op.mark.mark = mark->id + 1;
+                       ret = enic_fm_append_action_op(fm, &fm_op, error);
+                       if (ret)
+                               return ret;
+                       break;
+               }
+               case RTE_FLOW_ACTION_TYPE_FLAG: {
+                       if (overlap & MARK)
+                               goto unsupported;
+                       overlap |= MARK;
+                       /* ENIC_MAGIC_FILTER_ID is reserved for flagging */
+                       memset(&fm_op, 0, sizeof(fm_op));
+                       fm_op.fa_op = FMOP_MARK;
+                       fm_op.mark.mark = ENIC_MAGIC_FILTER_ID;
+                       ret = enic_fm_append_action_op(fm, &fm_op, error);
+                       if (ret)
+                               return ret;
+                       break;
+               }
+               case RTE_FLOW_ACTION_TYPE_QUEUE: {
+                       const struct rte_flow_action_queue *queue =
+                               actions->conf;
+
+                       /*
+                        * If other fate kind is set, fail.  Multiple
+                        * queue actions are ok.
+                        */
+                       if ((overlap & FATE) && first_rq)
+                               goto unsupported;
+                       first_rq = false;
+                       overlap |= FATE;
+                       memset(&fm_op, 0, sizeof(fm_op));
+                       fm_op.fa_op = FMOP_RQ_STEER;
+                       fm_op.rq_steer.rq_index =
+                               enic_rte_rq_idx_to_sop_idx(queue->index);
+                       fm_op.rq_steer.vnic_handle = vnic_h;
+                       ret = enic_fm_append_action_op(fm, &fm_op, error);
+                       if (ret)
+                               return ret;
+                       ENICPMD_LOG(DEBUG, "create QUEUE action rq: %u",
+                                   fm_op.rq_steer.rq_index);
+                       break;
+               }
+               case RTE_FLOW_ACTION_TYPE_DROP: {
+                       if (overlap & FATE)
+                               goto unsupported;
+                       overlap |= FATE;
+                       memset(&fm_op, 0, sizeof(fm_op));
+                       fm_op.fa_op = FMOP_DROP;
+                       ret = enic_fm_append_action_op(fm, &fm_op, error);
+                       if (ret)
+                               return ret;
+                       ENICPMD_LOG(DEBUG, "create DROP action");
+                       break;
+               }
+               case RTE_FLOW_ACTION_TYPE_COUNT: {
+                       if (overlap & COUNT)
+                               goto unsupported;
+                       overlap |= COUNT;
+                       /* Count is associated with entry not action on VIC. */
+                       fmt->ftm_flags |= FMEF_COUNTER;
+                       break;
+               }
+               case RTE_FLOW_ACTION_TYPE_RSS: {
+                       const struct rte_flow_action_rss *rss = actions->conf;
+                       bool allow;
+                       uint16_t i;
+
+                       /*
+                        * Hardware does not support general RSS actions, but
+                        * we can still support the dummy one that is used to
+                        * "receive normally".
+                        */
+                       allow = rss->func == RTE_ETH_HASH_FUNCTION_DEFAULT &&
+                               rss->level == 0 &&
+                               (rss->types == 0 ||
+                                rss->types == enic->rss_hf) &&
+                               rss->queue_num == enic->rq_count &&
+                               rss->key_len == 0;
+                       /* Identity queue map is ok */
+                       for (i = 0; i < rss->queue_num; i++)
+                               allow = allow && (i == rss->queue[i]);
+                       if (!allow)
+                               goto unsupported;
+                       if (overlap & FATE)
+                               goto unsupported;
+                       /* Need MARK or FLAG */
+                       if (!(overlap & MARK))
+                               goto unsupported;
+                       overlap |= FATE;
+                       break;
+               }
+               case RTE_FLOW_ACTION_TYPE_PORT_ID: {
+                       const struct rte_flow_action_port_id *port;
+                       struct rte_pci_device *pdev;
+                       struct rte_eth_dev *dev;
+
+                       port = actions->conf;
+                       if (port->original) {
+                               vnic_h = 0; /* This port */
+                               break;
+                       }
+                       ENICPMD_LOG(DEBUG, "port id %u", port->id);
+                       if (!rte_eth_dev_is_valid_port(port->id)) {
+                               return rte_flow_error_set(error, EINVAL,
+                                       RTE_FLOW_ERROR_TYPE_ACTION,
+                                       NULL, "invalid port_id");
+                       }
+                       dev = &rte_eth_devices[port->id];
+                       if (!dev_is_enic(dev)) {
+                               return rte_flow_error_set(error, EINVAL,
+                                       RTE_FLOW_ERROR_TYPE_ACTION,
+                                       NULL, "port_id is not enic");
+                       }
+                       pdev = RTE_ETH_DEV_TO_PCI(dev);
+                       if (enic_fm_find_vnic(enic, &pdev->addr, &vnic_h)) {
+                               return rte_flow_error_set(error, EINVAL,
+                                       RTE_FLOW_ERROR_TYPE_ACTION,
+                                       NULL, "port_id is not vnic");
+                       }
+                       break;
+               }
+               case RTE_FLOW_ACTION_TYPE_VXLAN_DECAP: {
+                       ret = enic_fm_copy_vxlan_decap(fm, fmt, actions,
+                               error);
+                       if (ret != 0)
+                               return ret;
+                       break;
+               }
+               case RTE_FLOW_ACTION_TYPE_VXLAN_ENCAP: {
+                       const struct rte_flow_action_vxlan_encap *encap;
+
+                       encap = actions->conf;
+                       if (overlap & ENCAP)
+                               goto unsupported;
+                       overlap |= ENCAP;
+                       ret = enic_fm_copy_vxlan_encap(fm, encap->definition,
+                               error);
+                       if (ret != 0)
+                               return ret;
+                       break;
+               }
+               default:
+                       goto unsupported;
+               }
+       }
+
+       if (!(overlap & (FATE | PASSTHRU | COUNT)))
+               goto unsupported;
+       memset(&fm_op, 0, sizeof(fm_op));
+       fm_op.fa_op = FMOP_END;
+       ret = enic_fm_append_action_op(fm, &fm_op, error);
+       if (ret)
+               return ret;
+       enic_fm_reorder_action_op(fm);
+       return 0;
+
+unsupported:
+       return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
+                                 NULL, "enic: unsupported action");
+}
+
+/** Check if the action is supported */
+static int
+enic_fm_match_action(const struct rte_flow_action *action,
+                    const enum rte_flow_action_type *supported_actions)
+{
+       for (; *supported_actions != RTE_FLOW_ACTION_TYPE_END;
+            supported_actions++) {
+               if (action->type == *supported_actions)
+                       return 1;
+       }
+       return 0;
+}
+
+/* Debug function to dump internal NIC action structure. */
+static void
+enic_fm_dump_tcam_actions(const struct fm_action *fm_action)
+{
+       /* Manually keep in sync with FMOP commands */
+       const char *fmop_str[FMOP_OP_MAX] = {
+               [FMOP_END] = "end",
+               [FMOP_DROP] = "drop",
+               [FMOP_RQ_STEER] = "steer",
+               [FMOP_EXACT_MATCH] = "exmatch",
+               [FMOP_MARK] = "mark",
+               [FMOP_EXT_MARK] = "ext_mark",
+               [FMOP_TAG] = "tag",
+               [FMOP_EG_HAIRPIN] = "eg_hairpin",
+               [FMOP_IG_HAIRPIN] = "ig_hairpin",
+               [FMOP_ENCAP_IVLAN] = "encap_ivlan",
+               [FMOP_ENCAP_NOIVLAN] = "encap_noivlan",
+               [FMOP_ENCAP] = "encap",
+               [FMOP_SET_OVLAN] = "set_ovlan",
+               [FMOP_DECAP_NOSTRIP] = "decap_nostrip",
+       };
+       const struct fm_action_op *op = &fm_action->fma_action_ops[0];
+       char buf[128], *bp = buf;
+       const char *op_str;
+       int i, n, buf_len;
+
+       buf[0] = '\0';
+       buf_len = sizeof(buf);
+       for (i = 0; i < FM_ACTION_OP_MAX; i++) {
+               if (op->fa_op == FMOP_END)
+                       break;
+               if (op->fa_op >= FMOP_OP_MAX)
+                       op_str = "unknown";
+               else
+                       op_str = fmop_str[op->fa_op];
+               n = snprintf(bp, buf_len, "%s,", op_str);
+               if (n > 0 && n < buf_len) {
+                       bp += n;
+                       buf_len -= n;
+               }
+               op++;
+       }
+       /* Remove trailing comma */
+       if (buf[0])
+               *(bp - 1) = '\0';
+       ENICPMD_LOG(DEBUG, "       Acions: %s", buf);
+}
+
+static int
+bits_to_str(uint32_t bits, const char *strings[], int max,
+           char *buf, int buf_len)
+{
+       int i, n = 0, len = 0;
+
+       for (i = 0; i < max; i++) {
+               if (bits & (1 << i)) {
+                       n = snprintf(buf, buf_len, "%s,", strings[i]);
+                       if (n > 0 && n < buf_len) {
+                               buf += n;
+                               buf_len -= n;
+                               len += n;
+                       }
+               }
+       }
+       /* Remove trailing comma */
+       if (len) {
+               *(buf - 1) = '\0';
+               len--;
+       }
+       return len;
+}
+
+/* Debug function to dump internal NIC filter structure. */
+static void
+__enic_fm_dump_tcam_match(const struct fm_header_set *fk_hdrset, char *buf,
+                         int buf_len)
+{
+       /* Manually keep in sync with FKM_BITS */
+       const char *fm_fkm_str[FKM_BIT_COUNT] = {
+               [FKM_QTAG_BIT] = "qtag",
+               [FKM_CMD_BIT] = "cmd",
+               [FKM_IPV4_BIT] = "ip4",
+               [FKM_IPV6_BIT] = "ip6",
+               [FKM_ROCE_BIT] = "roce",
+               [FKM_UDP_BIT] = "udp",
+               [FKM_TCP_BIT] = "tcp",
+               [FKM_TCPORUDP_BIT] = "tcpportudp",
+               [FKM_IPFRAG_BIT] = "ipfrag",
+               [FKM_NVGRE_BIT] = "nvgre",
+               [FKM_VXLAN_BIT] = "vxlan",
+               [FKM_GENEVE_BIT] = "geneve",
+               [FKM_NSH_BIT] = "nsh",
+               [FKM_ROCEV2_BIT] = "rocev2",
+               [FKM_VLAN_PRES_BIT] = "vlan_pres",
+               [FKM_IPOK_BIT] = "ipok",
+               [FKM_L4OK_BIT] = "l4ok",
+               [FKM_ROCEOK_BIT] = "roceok",
+               [FKM_FCSOK_BIT] = "fcsok",
+               [FKM_EG_SPAN_BIT] = "eg_span",
+               [FKM_IG_SPAN_BIT] = "ig_span",
+               [FKM_EG_HAIRPINNED_BIT] = "eg_hairpinned",
+       };
+       /* Manually keep in sync with FKH_BITS */
+       const char *fm_fkh_str[FKH_BIT_COUNT] = {
+               [FKH_ETHER_BIT] = "eth",
+               [FKH_QTAG_BIT] = "qtag",
+               [FKH_L2RAW_BIT] = "l2raw",
+               [FKH_IPV4_BIT] = "ip4",
+               [FKH_IPV6_BIT] = "ip6",
+               [FKH_L3RAW_BIT] = "l3raw",
+               [FKH_UDP_BIT] = "udp",
+               [FKH_TCP_BIT] = "tcp",
+               [FKH_ICMP_BIT] = "icmp",
+               [FKH_VXLAN_BIT] = "vxlan",
+               [FKH_L4RAW_BIT] = "l4raw",
+       };
+       uint32_t fkh_bits = fk_hdrset->fk_header_select;
+       uint32_t fkm_bits = fk_hdrset->fk_metadata;
+       int n;
+
+       if (!fkm_bits && !fkh_bits)
+               return;
+       n = snprintf(buf, buf_len, "metadata(");
+       if (n > 0 && n < buf_len) {
+               buf += n;
+               buf_len -= n;
+       }
+       n = bits_to_str(fkm_bits, fm_fkm_str, FKM_BIT_COUNT, buf, buf_len);
+       if (n > 0 && n < buf_len) {
+               buf += n;
+               buf_len -= n;
+       }
+       n = snprintf(buf, buf_len, ") valid hdr fields(");
+       if (n > 0 && n < buf_len) {
+               buf += n;
+               buf_len -= n;
+       }
+       n = bits_to_str(fkh_bits, fm_fkh_str, FKH_BIT_COUNT, buf, buf_len);
+       if (n > 0 && n < buf_len) {
+               buf += n;
+               buf_len -= n;
+       }
+       snprintf(buf, buf_len, ")");
+}
+
+static void
+enic_fm_dump_tcam_match(const struct fm_tcam_match_entry *match,
+                       uint8_t ingress)
+{
+       char buf[256];
+
+       memset(buf, 0, sizeof(buf));
+       __enic_fm_dump_tcam_match(&match->ftm_mask.fk_hdrset[0],
+                                 buf, sizeof(buf));
+       ENICPMD_LOG(DEBUG, " TCAM %s Outer: %s %scounter",
+                   (ingress) ? "IG" : "EG", buf,
+                   (match->ftm_flags & FMEF_COUNTER) ? "" : "no ");
+       memset(buf, 0, sizeof(buf));
+       __enic_fm_dump_tcam_match(&match->ftm_mask.fk_hdrset[1],
+                                 buf, sizeof(buf));
+       if (buf[0])
+               ENICPMD_LOG(DEBUG, "         Inner: %s", buf);
+}
+
+/* Debug function to dump internal NIC flow structures. */
+static void
+enic_fm_dump_tcam_entry(const struct fm_tcam_match_entry *fm_match,
+                       const struct fm_action *fm_action,
+                       uint8_t ingress)
+{
+       if (rte_log_get_level(enic_pmd_logtype) < (int)RTE_LOG_DEBUG)
+               return;
+       enic_fm_dump_tcam_match(fm_match, ingress);
+       enic_fm_dump_tcam_actions(fm_action);
+}
+
+static int
+enic_fm_flow_parse(struct enic_flowman *fm,
+                  const struct rte_flow_attr *attrs,
+                  const struct rte_flow_item pattern[],
+                  const struct rte_flow_action actions[],
+                  struct rte_flow_error *error)
+{
+       const struct rte_flow_action *action;
+       unsigned int ret;
+       static const enum rte_flow_action_type *sa;
+
+       ENICPMD_FUNC_TRACE();
+       ret = 0;
+       if (!pattern) {
+               rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM_NUM,
+                                  NULL, "no pattern specified");
+               return -rte_errno;
+       }
+
+       if (!actions) {
+               rte_flow_error_set(error, EINVAL,
+                                  RTE_FLOW_ERROR_TYPE_ACTION_NUM,
+                                  NULL, "no action specified");
+               return -rte_errno;
+       }
+
+       if (attrs) {
+               if (attrs->priority) {
+                       rte_flow_error_set(error, ENOTSUP,
+                                          RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
+                                          NULL,
+                                          "priorities are not supported");
+                       return -rte_errno;
+               } else if (attrs->transfer) {
+                       rte_flow_error_set(error, ENOTSUP,
+                                          RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
+                                          NULL,
+                                          "transfer is not supported");
+                       return -rte_errno;
+               } else if (attrs->ingress && attrs->egress) {
+                       rte_flow_error_set(error, ENOTSUP,
+                                          RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
+                                          NULL,
+                                          "bidirectional rules not supported");
+                       return -rte_errno;
+               }
+
+       } else {
+               rte_flow_error_set(error, EINVAL,
+                                  RTE_FLOW_ERROR_TYPE_ATTR,
+                                  NULL, "no attribute specified");
+               return -rte_errno;
+       }
+
+       /* Verify Actions. */
+       sa = (attrs->ingress) ? enic_fm_supported_ig_actions :
+            enic_fm_supported_eg_actions;
+       for (action = &actions[0]; action->type != RTE_FLOW_ACTION_TYPE_END;
+            action++) {
+               if (action->type == RTE_FLOW_ACTION_TYPE_VOID)
+                       continue;
+               else if (!enic_fm_match_action(action, sa))
+                       break;
+       }
+       if (action->type != RTE_FLOW_ACTION_TYPE_END) {
+               rte_flow_error_set(error, EPERM, RTE_FLOW_ERROR_TYPE_ACTION,
+                                  action, "invalid action");
+               return -rte_errno;
+       }
+       ret = enic_fm_copy_entry(fm, pattern, error);
+       if (ret)
+               return ret;
+       ret = enic_fm_copy_action(fm, actions, attrs->ingress, error);
+       return ret;
+}
+
+static void
+enic_fm_counter_free(struct enic_flowman *fm, struct enic_fm_flow *fm_flow)
+{
+       if (!fm_flow->counter_valid)
+               return;
+       SLIST_INSERT_HEAD(&fm->counters, fm_flow->counter, next);
+       fm_flow->counter_valid = false;
+}
+
+static int
+enic_fm_more_counters(struct enic_flowman *fm)
+{
+       struct enic_fm_counter *new_stack;
+       struct enic_fm_counter *ctrs;
+       struct enic *enic;
+       int i, rc;
+       u64 args[2];
+
+       ENICPMD_FUNC_TRACE();
+       enic = fm->enic;
+       new_stack = rte_realloc(fm->counter_stack, (fm->counters_alloced +
+                               FM_COUNTERS_EXPAND) *
+                               sizeof(struct enic_fm_counter), 0);
+       if (new_stack == NULL) {
+               ENICPMD_LOG(ERR, "cannot alloc counter memory");
+               return -ENOMEM;
+       }
+       fm->counter_stack = new_stack;
+
+       args[0] = FM_COUNTER_BRK;
+       args[1] = fm->counters_alloced + FM_COUNTERS_EXPAND;
+       rc = vnic_dev_flowman_cmd(enic->vdev, args, 2);
+       if (rc != 0) {
+               ENICPMD_LOG(ERR, "cannot alloc counters rc=%d", rc);
+               return rc;
+       }
+       ctrs = (struct enic_fm_counter *)fm->counter_stack +
+               fm->counters_alloced;
+       for (i = 0; i < FM_COUNTERS_EXPAND; i++, ctrs++) {
+               ctrs->handle = fm->counters_alloced + i;
+               SLIST_INSERT_HEAD(&fm->counters, ctrs, next);
+       }
+       fm->counters_alloced += FM_COUNTERS_EXPAND;
+       ENICPMD_LOG(DEBUG, "%u counters allocated, total: %u",
+                   FM_COUNTERS_EXPAND, fm->counters_alloced);
+       return 0;
+}
+
+static int
+enic_fm_counter_zero(struct enic_flowman *fm, struct enic_fm_counter *c)
+{
+       struct enic *enic;
+       u64 args[3];
+       int ret;
+
+       ENICPMD_FUNC_TRACE();
+       enic = fm->enic;
+       args[0] = FM_COUNTER_QUERY;
+       args[1] = c->handle;
+       args[2] = 1; /* clear */
+       ret = vnic_dev_flowman_cmd(enic->vdev, args, 3);
+       if (ret) {
+               ENICPMD_LOG(ERR, "counter init: rc=%d handle=0x%x",
+                           ret, c->handle);
+               return ret;
+       }
+       return 0;
+}
+
+static int
+enic_fm_counter_alloc(struct enic_flowman *fm, struct rte_flow_error *error,
+                     struct enic_fm_counter **ctr)
+{
+       struct enic_fm_counter *c;
+       int ret;
+
+       ENICPMD_FUNC_TRACE();
+       *ctr = NULL;
+       if (SLIST_EMPTY(&fm->counters)) {
+               ret = enic_fm_more_counters(fm);
+               if (ret)
+                       return rte_flow_error_set(error, -ret,
+                               RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                               NULL, "enic: out of counters");
+       }
+       c = SLIST_FIRST(&fm->counters);
+       SLIST_REMOVE_HEAD(&fm->counters, next);
+       *ctr = c;
+       return 0;
+}
+
+static int
+enic_fm_action_free(struct enic_flowman *fm, uint64_t handle)
+{
+       u64 args[2];
+       int rc;
+
+       ENICPMD_FUNC_TRACE();
+       args[0] = FM_ACTION_FREE;
+       args[1] = handle;
+       rc = vnic_dev_flowman_cmd(fm->enic->vdev, args, 2);
+       if (rc)
+               ENICPMD_LOG(ERR, "cannot free action: rc=%d handle=0x%" PRIx64,
+                           rc, handle);
+       return rc;
+}
+
+static int
+enic_fm_entry_free(struct enic_flowman *fm, uint64_t handle)
+{
+       u64 args[2];
+       int rc;
+
+       ENICPMD_FUNC_TRACE();
+       args[0] = FM_MATCH_ENTRY_REMOVE;
+       args[1] = handle;
+       rc = vnic_dev_flowman_cmd(fm->enic->vdev, args, 2);
+       if (rc)
+               ENICPMD_LOG(ERR, "cannot free match entry: rc=%d"
+                           " handle=0x%" PRIx64, rc, handle);
+       return rc;
+}
+
+static struct enic_fm_jump_flow *
+find_jump_flow(struct enic_flowman *fm, uint32_t group)
+{
+       struct enic_fm_jump_flow *j;
+
+       ENICPMD_FUNC_TRACE();
+       TAILQ_FOREACH(j, &fm->jump_list, list) {
+               if (j->group == group)
+                       return j;
+       }
+       return NULL;
+}
+
+static void
+remove_jump_flow(struct enic_flowman *fm, struct rte_flow *flow)
+{
+       struct enic_fm_jump_flow *j;
+
+       ENICPMD_FUNC_TRACE();
+       TAILQ_FOREACH(j, &fm->jump_list, list) {
+               if (j->flow == flow) {
+                       TAILQ_REMOVE(&fm->jump_list, j, list);
+                       free(j);
+                       return;
+               }
+       }
+}
+
+static int
+save_jump_flow(struct enic_flowman *fm,
+              struct rte_flow *flow,
+              uint32_t group,
+              struct fm_tcam_match_entry *match,
+              struct fm_action *action)
+{
+       struct enic_fm_jump_flow *j;
+
+       ENICPMD_FUNC_TRACE();
+       j = calloc(1, sizeof(struct enic_fm_jump_flow));
+       if (j == NULL)
+               return -ENOMEM;
+       j->flow = flow;
+       j->group = group;
+       j->match = *match;
+       j->action = *action;
+       TAILQ_INSERT_HEAD(&fm->jump_list, j, list);
+       ENICPMD_LOG(DEBUG, "saved jump flow: flow=%p group=%u", flow, group);
+       return 0;
+}
+
+static void
+__enic_fm_flow_free(struct enic_flowman *fm, struct enic_fm_flow *fm_flow)
+{
+       if (fm_flow->entry_handle != FM_INVALID_HANDLE) {
+               enic_fm_entry_free(fm, fm_flow->entry_handle);
+               fm_flow->entry_handle = FM_INVALID_HANDLE;
+       }
+       if (fm_flow->action_handle != FM_INVALID_HANDLE) {
+               enic_fm_action_free(fm, fm_flow->action_handle);
+               fm_flow->action_handle = FM_INVALID_HANDLE;
+       }
+       enic_fm_counter_free(fm, fm_flow);
+       if (fm_flow->fet) {
+               enic_fet_put(fm, fm_flow->fet);
+               fm_flow->fet = NULL;
+       }
+}
+
+static void
+enic_fm_flow_free(struct enic_flowman *fm, struct rte_flow *flow)
+{
+       if (flow->fm->fet && flow->fm->fet->default_key)
+               remove_jump_flow(fm, flow);
+       __enic_fm_flow_free(fm, flow->fm);
+       free(flow->fm);
+       free(flow);
+}
+
+static int
+enic_fm_add_tcam_entry(struct enic_flowman *fm,
+                      struct fm_tcam_match_entry *match_in,
+                      uint64_t *entry_handle,
+                      uint8_t ingress,
+                      struct rte_flow_error *error)
+{
+       struct fm_tcam_match_entry *ftm;
+       u64 args[3];
+       int ret;
+
+       ENICPMD_FUNC_TRACE();
+       /* Copy entry to the command buffer */
+       ftm = &fm->cmd.va->fm_tcam_match_entry;
+       memcpy(ftm, match_in, sizeof(*ftm));
+       /* Add TCAM entry */
+       args[0] = FM_TCAM_ENTRY_INSTALL;
+       args[1] = ingress ? fm->ig_tcam_hndl : fm->eg_tcam_hndl;
+       args[2] = fm->cmd.pa;
+       ret = vnic_dev_flowman_cmd(fm->enic->vdev, args, 3);
+       if (ret != 0) {
+               ENICPMD_LOG(ERR, "cannot add %s TCAM entry: rc=%d",
+                           ingress ? "ingress" : "egress", ret);
+               rte_flow_error_set(error, ret, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                       NULL, "enic: devcmd(tcam-entry-install)");
+               return ret;
+       }
+       ENICPMD_LOG(DEBUG, "installed %s TCAM entry: handle=0x%" PRIx64,
+                   ingress ? "ingress" : "egress", (uint64_t)args[0]);
+       *entry_handle = args[0];
+       return 0;
+}
+
+static int
+enic_fm_add_exact_entry(struct enic_flowman *fm,
+                       struct fm_tcam_match_entry *match_in,
+                       uint64_t *entry_handle,
+                       struct enic_fm_fet *fet,
+                       struct rte_flow_error *error)
+{
+       struct fm_exact_match_entry *fem;
+       u64 args[3];
+       int ret;
+
+       ENICPMD_FUNC_TRACE();
+       /* The new entry must have the table's key */
+       if (memcmp(fet->key.fk_hdrset, match_in->ftm_mask.fk_hdrset,
+                  sizeof(struct fm_header_set) * FM_HDRSET_MAX)) {
+               return rte_flow_error_set(error, EINVAL,
+                       RTE_FLOW_ERROR_TYPE_ITEM, NULL,
+                       "enic: key does not match group's key");
+       }
+
+       /* Copy entry to the command buffer */
+       fem = &fm->cmd.va->fm_exact_match_entry;
+       /*
+        * Translate TCAM entry to exact entry. As is only need to drop
+        * position and mask. The mask is part of the exact match table.
+        * Position (aka priority) is not supported in the exact match table.
+        */
+       fem->fem_data = match_in->ftm_data;
+       fem->fem_flags = match_in->ftm_flags;
+       fem->fem_action = match_in->ftm_action;
+       fem->fem_counter = match_in->ftm_counter;
+
+       /* Add exact entry */
+       args[0] = FM_EXACT_ENTRY_INSTALL;
+       args[1] = fet->handle;
+       args[2] = fm->cmd.pa;
+       ret = vnic_dev_flowman_cmd(fm->enic->vdev, args, 3);
+       if (ret != 0) {
+               ENICPMD_LOG(ERR, "cannot add %s exact entry: group=%u",
+                           fet->ingress ? "ingress" : "egress", fet->group);
+               rte_flow_error_set(error, ret, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                       NULL, "enic: devcmd(exact-entry-install)");
+               return ret;
+       }
+       ENICPMD_LOG(DEBUG, "installed %s exact entry: group=%u"
+                   " handle=0x%" PRIx64,
+                   fet->ingress ? "ingress" : "egress", fet->group,
+                   (uint64_t)args[0]);
+       *entry_handle = args[0];
+       return 0;
+}
+
+/* Push match-action to the NIC. */
+static int
+__enic_fm_flow_add_entry(struct enic_flowman *fm,
+                        struct enic_fm_flow *fm_flow,
+                        struct fm_tcam_match_entry *match_in,
+                        struct fm_action *action_in,
+                        uint32_t group,
+                        uint8_t ingress,
+                        struct rte_flow_error *error)
+{
+       struct enic_fm_counter *ctr;
+       struct fm_action *fma;
+       uint64_t action_h;
+       uint64_t entry_h;
+       u64 args[3];
+       int ret;
+
+       ENICPMD_FUNC_TRACE();
+       /* Allocate action. */
+       fma = &fm->cmd.va->fm_action;
+       memcpy(fma, action_in, sizeof(*fma));
+       args[0] = FM_ACTION_ALLOC;
+       args[1] = fm->cmd.pa;
+       ret = vnic_dev_flowman_cmd(fm->enic->vdev, args, 2);
+       if (ret != 0) {
+               ENICPMD_LOG(ERR, "allocating TCAM table action rc=%d", ret);
+               rte_flow_error_set(error, ret, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                       NULL, "enic: devcmd(action-alloc)");
+               return ret;
+       }
+       action_h = args[0];
+       fm_flow->action_handle = action_h;
+       match_in->ftm_action = action_h;
+       ENICPMD_LOG(DEBUG, "action allocated: handle=0x%" PRIx64, action_h);
+
+       /* Allocate counter if requested. */
+       if (match_in->ftm_flags & FMEF_COUNTER) {
+               ret = enic_fm_counter_alloc(fm, error, &ctr);
+               if (ret) /* error has been filled in */
+                       return ret;
+               fm_flow->counter_valid = true;
+               fm_flow->counter = ctr;
+               match_in->ftm_counter = ctr->handle;
+       }
+
+       /*
+        * Get the group's table (either TCAM or exact match table) and
+        * add entry to it. If we use the exact match table, the handler
+        * will translate the TCAM entry (match_in) to the appropriate
+        * exact match entry and use that instead.
+        */
+       entry_h = FM_INVALID_HANDLE;
+       if (group == FM_TCAM_RTE_GROUP) {
+               ret = enic_fm_add_tcam_entry(fm, match_in, &entry_h, ingress,
+                                            error);
+               if (ret)
+                       return ret;
+               /* Jump action might have a ref to fet */
+               fm_flow->fet = fm->fet;
+               fm->fet = NULL;
+       } else {
+               struct enic_fm_fet *fet = NULL;
+
+               ret = enic_fet_get(fm, group, ingress,
+                                  &match_in->ftm_mask, &fet, error);
+               if (ret)
+                       return ret;
+               fm_flow->fet = fet;
+               ret = enic_fm_add_exact_entry(fm, match_in, &entry_h, fet,
+                                             error);
+               if (ret)
+                       return ret;
+       }
+       /* Clear counter after adding entry, as it requires in-use counter */
+       if (fm_flow->counter_valid) {
+               ret = enic_fm_counter_zero(fm, fm_flow->counter);
+               if (ret)
+                       return ret;
+       }
+       fm_flow->entry_handle = entry_h;
+       return 0;
+}
+
+/* Push match-action to the NIC. */
+static struct rte_flow *
+enic_fm_flow_add_entry(struct enic_flowman *fm,
+                      struct fm_tcam_match_entry *match_in,
+                      struct fm_action *action_in,
+                      const struct rte_flow_attr *attrs,
+                      struct rte_flow_error *error)
+{
+       struct enic_fm_flow *fm_flow;
+       struct rte_flow *flow;
+
+       ENICPMD_FUNC_TRACE();
+       enic_fm_dump_tcam_entry(match_in, action_in, attrs->ingress);
+       flow = calloc(1, sizeof(*flow));
+       fm_flow = calloc(1, sizeof(*fm_flow));
+       if (flow == NULL || fm_flow == NULL) {
+               rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
+                       NULL, "enic: cannot allocate rte_flow");
+               free(flow);
+               free(fm_flow);
+               return NULL;
+       }
+       flow->fm = fm_flow;
+       fm_flow->action_handle = FM_INVALID_HANDLE;
+       fm_flow->entry_handle = FM_INVALID_HANDLE;
+       if (__enic_fm_flow_add_entry(fm, fm_flow, match_in, action_in,
+                                    attrs->group, attrs->ingress, error)) {
+               enic_fm_flow_free(fm, flow);
+               return NULL;
+       }
+       return flow;
+}
+
+static void
+convert_jump_flows(struct enic_flowman *fm, struct enic_fm_fet *fet,
+                  struct rte_flow_error *error)
+{
+       struct enic_fm_flow *fm_flow;
+       struct enic_fm_jump_flow *j;
+       struct fm_action *fma;
+       uint32_t group;
+
+       ENICPMD_FUNC_TRACE();
+       /*
+        * Find the saved flows that should jump to the new table (fet).
+        * Then delete the old TCAM entry that jumps to the default table,
+        * and add a new one that jumps to the new table.
+        */
+       group = fet->group;
+       j = find_jump_flow(fm, group);
+       while (j) {
+               ENICPMD_LOG(DEBUG, "convert jump flow: flow=%p group=%u",
+                           j->flow, group);
+               /* Delete old entry */
+               fm_flow = j->flow->fm;
+               __enic_fm_flow_free(fm, fm_flow);
+
+               /* Add new entry */
+               fma = &j->action;
+               fma->fma_action_ops[0].exact.handle = fet->handle;
+               if (__enic_fm_flow_add_entry(fm, fm_flow, &j->match, fma,
+                       FM_TCAM_RTE_GROUP, fet->ingress, error)) {
+                       /* Cannot roll back changes at the moment */
+                       ENICPMD_LOG(ERR, "cannot convert jump flow: flow=%p",
+                                   j->flow);
+               } else {
+                       fm_flow->fet = fet;
+                       fet->ref++;
+                       ENICPMD_LOG(DEBUG, "convert ok: group=%u ref=%u",
+                                   fet->group, fet->ref);
+               }
+
+               TAILQ_REMOVE(&fm->jump_list, j, list);
+               free(j);
+               j = find_jump_flow(fm, group);
+       }
+}
+
+static void
+enic_fm_open_scratch(struct enic_flowman *fm)
+{
+       fm->action_op_count = 0;
+       fm->fet = NULL;
+       memset(&fm->tcam_entry, 0, sizeof(fm->tcam_entry));
+       memset(&fm->action, 0, sizeof(fm->action));
+}
+
+static void
+enic_fm_close_scratch(struct enic_flowman *fm)
+{
+       if (fm->fet) {
+               enic_fet_put(fm, fm->fet);
+               fm->fet = NULL;
+       }
+       fm->action_op_count = 0;
+}
+
+static int
+enic_fm_flow_validate(struct rte_eth_dev *dev,
+                     const struct rte_flow_attr *attrs,
+                     const struct rte_flow_item pattern[],
+                     const struct rte_flow_action actions[],
+                     struct rte_flow_error *error)
+{
+       struct fm_tcam_match_entry *fm_tcam_entry;
+       struct fm_action *fm_action;
+       struct enic_flowman *fm;
+       int ret;
+
+       ENICPMD_FUNC_TRACE();
+       fm = pmd_priv(dev)->fm;
+       if (fm == NULL)
+               return -ENOTSUP;
+       enic_fm_open_scratch(fm);
+       ret = enic_fm_flow_parse(fm, attrs, pattern, actions, error);
+       if (!ret) {
+               fm_tcam_entry = &fm->tcam_entry;
+               fm_action = &fm->action;
+               enic_fm_dump_tcam_entry(fm_tcam_entry, fm_action,
+                                       attrs->ingress);
+       }
+       enic_fm_close_scratch(fm);
+       return ret;
+}
+
+static int
+enic_fm_flow_query_count(struct rte_eth_dev *dev,
+                        struct rte_flow *flow, void *data,
+                        struct rte_flow_error *error)
+{
+       struct rte_flow_query_count *query;
+       struct enic_fm_flow *fm_flow;
+       struct enic *enic;
+       u64 args[3];
+       int rc;
+
+       ENICPMD_FUNC_TRACE();
+       enic = pmd_priv(dev);
+       query = data;
+       fm_flow = flow->fm;
+       if (!fm_flow->counter_valid)
+               return rte_flow_error_set(error, ENOTSUP,
+                       RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
+                       "enic: flow does not have counter");
+
+       args[0] = FM_COUNTER_QUERY;
+       args[1] = fm_flow->counter->handle;
+       args[2] = query->reset;
+       rc = vnic_dev_flowman_cmd(enic->vdev, args, 3);
+       if (rc) {
+               ENICPMD_LOG(ERR, "cannot query counter: rc=%d handle=0x%x",
+                           rc, fm_flow->counter->handle);
+               return rc;
+       }
+       query->hits_set = 1;
+       query->hits = args[0];
+       query->bytes_set = 1;
+       query->bytes = args[1];
+       return 0;
+}
+
+static int
+enic_fm_flow_query(struct rte_eth_dev *dev,
+                  struct rte_flow *flow,
+                  const struct rte_flow_action *actions,
+                  void *data,
+                  struct rte_flow_error *error)
+{
+       int ret = 0;
+
+       ENICPMD_FUNC_TRACE();
+       for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
+               switch (actions->type) {
+               case RTE_FLOW_ACTION_TYPE_VOID:
+                       break;
+               case RTE_FLOW_ACTION_TYPE_COUNT:
+                       ret = enic_fm_flow_query_count(dev, flow, data, error);
+                       break;
+               default:
+                       return rte_flow_error_set(error, ENOTSUP,
+                                                 RTE_FLOW_ERROR_TYPE_ACTION,
+                                                 actions,
+                                                 "action not supported");
+               }
+               if (ret < 0)
+                       return ret;
+       }
+       return 0;
+}
+
+static struct rte_flow *
+enic_fm_flow_create(struct rte_eth_dev *dev,
+                   const struct rte_flow_attr *attrs,
+                   const struct rte_flow_item pattern[],
+                   const struct rte_flow_action actions[],
+                   struct rte_flow_error *error)
+{
+       struct fm_tcam_match_entry *fm_tcam_entry;
+       struct fm_action *fm_action;
+       struct enic_flowman *fm;
+       struct enic_fm_fet *fet;
+       struct rte_flow *flow;
+       struct enic *enic;
+       int ret;
+
+       ENICPMD_FUNC_TRACE();
+       enic = pmd_priv(dev);
+       fm = enic->fm;
+       if (fm == NULL) {
+               rte_flow_error_set(error, ENOTSUP,
+                       RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
+                       "flowman is not initialized");
+               return NULL;
+       }
+       enic_fm_open_scratch(fm);
+       flow = NULL;
+       ret = enic_fm_flow_parse(fm, attrs, pattern, actions, error);
+       if (ret < 0)
+               goto error_with_scratch;
+       fm_tcam_entry = &fm->tcam_entry;
+       fm_action = &fm->action;
+       flow = enic_fm_flow_add_entry(fm, fm_tcam_entry, fm_action,
+                                     attrs, error);
+       if (flow) {
+               LIST_INSERT_HEAD(&enic->flows, flow, next);
+               fet = flow->fm->fet;
+               if (fet && fet->default_key) {
+                       /*
+                        * Jump to non-existent group? Save the relevant info
+                        * so we can convert this flow when that group
+                        * materializes.
+                        */
+                       save_jump_flow(fm, flow, fet->group,
+                                      fm_tcam_entry, fm_action);
+               } else if (fet && fet->ref == 1) {
+                       /*
+                        * A new table is created. Convert the saved flows
+                        * that should jump to this group.
+                        */
+                       convert_jump_flows(fm, fet, error);
+               }
+       }
+
+error_with_scratch:
+       enic_fm_close_scratch(fm);
+       return flow;
+}
+
+static int
+enic_fm_flow_destroy(struct rte_eth_dev *dev, struct rte_flow *flow,
+                    __rte_unused struct rte_flow_error *error)
+{
+       struct enic *enic = pmd_priv(dev);
+
+       ENICPMD_FUNC_TRACE();
+       if (enic->fm == NULL)
+               return 0;
+       LIST_REMOVE(flow, next);
+       enic_fm_flow_free(enic->fm, flow);
+       return 0;
+}
+
+static int
+enic_fm_flow_flush(struct rte_eth_dev *dev,
+                  __rte_unused struct rte_flow_error *error)
+{
+       struct enic_fm_flow *fm_flow;
+       struct enic_flowman *fm;
+       struct rte_flow *flow;
+       struct enic *enic = pmd_priv(dev);
+
+       ENICPMD_FUNC_TRACE();
+       if (enic->fm == NULL)
+               return 0;
+       fm = enic->fm;
+       while (!LIST_EMPTY(&enic->flows)) {
+               flow = LIST_FIRST(&enic->flows);
+               fm_flow = flow->fm;
+               LIST_REMOVE(flow, next);
+               /*
+                * If tables are null, then vNIC is closing, and the firmware
+                * has already cleaned up flowman state. So do not try to free
+                * resources, as it only causes errors.
+                */
+               if (fm->ig_tcam_hndl == FM_INVALID_HANDLE) {
+                       fm_flow->entry_handle = FM_INVALID_HANDLE;
+                       fm_flow->action_handle = FM_INVALID_HANDLE;
+                       fm_flow->fet = NULL;
+               }
+               enic_fm_flow_free(fm, flow);
+       }
+       return 0;
+}
+
+static int
+enic_fm_tbl_free(struct enic_flowman *fm, uint64_t handle)
+{
+       u64 args[2];
+       int rc;
+
+       args[0] = FM_MATCH_TABLE_FREE;
+       args[1] = handle;
+       rc = vnic_dev_flowman_cmd(fm->enic->vdev, args, 2);
+       if (rc)
+               ENICPMD_LOG(ERR, "cannot free table: rc=%d handle=0x%" PRIx64,
+                           rc, handle);
+       return rc;
+}
+
+static int
+enic_fm_tcam_tbl_alloc(struct enic_flowman *fm, uint32_t direction,
+                       uint32_t max_entries, uint64_t *handle)
+{
+       struct fm_tcam_match_table *tcam_tbl;
+       struct enic *enic;
+       u64 args[2];
+       int rc;
+
+       ENICPMD_FUNC_TRACE();
+       enic = fm->enic;
+       tcam_tbl = &fm->cmd.va->fm_tcam_match_table;
+       tcam_tbl->ftt_direction = direction;
+       tcam_tbl->ftt_stage = FM_STAGE_LAST;
+       tcam_tbl->ftt_max_entries = max_entries;
+       args[0] = FM_TCAM_TABLE_ALLOC;
+       args[1] = fm->cmd.pa;
+       rc = vnic_dev_flowman_cmd(enic->vdev, args, 2);
+       if (rc) {
+               ENICPMD_LOG(ERR, "cannot alloc %s TCAM table: rc=%d",
+                           (direction == FM_INGRESS) ? "IG" : "EG", rc);
+               return rc;
+       }
+       *handle = args[0];
+       ENICPMD_LOG(DEBUG, "%s TCAM table allocated, handle=0x%" PRIx64,
+                   (direction == FM_INGRESS) ? "IG" : "EG", *handle);
+       return 0;
+}
+
+static int
+enic_fm_init_counters(struct enic_flowman *fm)
+{
+       ENICPMD_FUNC_TRACE();
+       SLIST_INIT(&fm->counters);
+       return enic_fm_more_counters(fm);
+}
+
+static void
+enic_fm_free_all_counters(struct enic_flowman *fm)
+{
+       struct enic *enic;
+       u64 args[2];
+       int rc;
+
+       enic = fm->enic;
+       args[0] = FM_COUNTER_BRK;
+       args[1] = 0;
+       rc = vnic_dev_flowman_cmd(enic->vdev, args, 2);
+       if (rc != 0)
+               ENICPMD_LOG(ERR, "cannot free counters: rc=%d", rc);
+       rte_free(fm->counter_stack);
+}
+
+static int
+enic_fm_alloc_tcam_tables(struct enic_flowman *fm)
+{
+       int rc;
+
+       ENICPMD_FUNC_TRACE();
+       rc = enic_fm_tcam_tbl_alloc(fm, FM_INGRESS, FM_MAX_TCAM_TABLE_SIZE,
+                                   &fm->ig_tcam_hndl);
+       if (rc)
+               return rc;
+       rc = enic_fm_tcam_tbl_alloc(fm, FM_EGRESS, FM_MAX_TCAM_TABLE_SIZE,
+                                   &fm->eg_tcam_hndl);
+       return rc;
+}
+
+static void
+enic_fm_free_tcam_tables(struct enic_flowman *fm)
+{
+       ENICPMD_FUNC_TRACE();
+       if (fm->ig_tcam_hndl) {
+               ENICPMD_LOG(DEBUG, "free IG TCAM table handle=0x%" PRIx64,
+                           fm->ig_tcam_hndl);
+               enic_fm_tbl_free(fm, fm->ig_tcam_hndl);
+               fm->ig_tcam_hndl = FM_INVALID_HANDLE;
+       }
+       if (fm->eg_tcam_hndl) {
+               ENICPMD_LOG(DEBUG, "free EG TCAM table handle=0x%" PRIx64,
+                           fm->eg_tcam_hndl);
+               enic_fm_tbl_free(fm, fm->eg_tcam_hndl);
+               fm->eg_tcam_hndl = FM_INVALID_HANDLE;
+       }
+}
+
+int
+enic_fm_init(struct enic *enic)
+{
+       struct enic_flowman *fm;
+       u8 name[NAME_MAX];
+       int rc;
+
+       if (enic->flow_filter_mode != FILTER_FLOWMAN)
+               return 0;
+       ENICPMD_FUNC_TRACE();
+       fm = calloc(1, sizeof(*fm));
+       if (fm == NULL) {
+               ENICPMD_LOG(ERR, "cannot alloc flowman struct");
+               return -ENOMEM;
+       }
+       fm->enic = enic;
+       TAILQ_INIT(&fm->fet_list);
+       TAILQ_INIT(&fm->jump_list);
+       /* Allocate host memory for flowman commands */
+       snprintf((char *)name, NAME_MAX, "fm-cmd-%s", enic->bdf_name);
+       fm->cmd.va = enic_alloc_consistent(enic,
+               sizeof(union enic_flowman_cmd_mem), &fm->cmd.pa, name);
+       if (!fm->cmd.va) {
+               ENICPMD_LOG(ERR, "cannot allocate flowman command memory");
+               rc = -ENOMEM;
+               goto error_fm;
+       }
+       /* Allocate TCAM tables upfront as they are the main tables */
+       rc = enic_fm_alloc_tcam_tables(fm);
+       if (rc) {
+               ENICPMD_LOG(ERR, "cannot alloc TCAM tables");
+               goto error_cmd;
+       }
+       /* Then a number of counters */
+       rc = enic_fm_init_counters(fm);
+       if (rc) {
+               ENICPMD_LOG(ERR, "cannot alloc counters");
+               goto error_tables;
+       }
+       /*
+        * One default exact match table for each direction. We hold onto
+        * it until close.
+        */
+       rc = enic_fet_alloc(fm, 1, NULL, 128, &fm->default_ig_fet);
+       if (rc) {
+               ENICPMD_LOG(ERR, "cannot alloc default IG exact match table");
+               goto error_counters;
+       }
+       fm->default_ig_fet->ref = 1;
+       rc = enic_fet_alloc(fm, 0, NULL, 128, &fm->default_eg_fet);
+       if (rc) {
+               ENICPMD_LOG(ERR, "cannot alloc default EG exact match table");
+               goto error_ig_fet;
+       }
+       fm->default_eg_fet->ref = 1;
+       enic->fm = fm;
+       return 0;
+
+error_ig_fet:
+       enic_fet_free(fm, fm->default_ig_fet);
+error_counters:
+       enic_fm_free_all_counters(fm);
+error_tables:
+       enic_fm_free_tcam_tables(fm);
+error_cmd:
+       enic_free_consistent(enic, sizeof(union enic_flowman_cmd_mem),
+               fm->cmd.va, fm->cmd.pa);
+error_fm:
+       free(fm);
+       return rc;
+}
+
+void
+enic_fm_destroy(struct enic *enic)
+{
+       struct enic_flowman *fm;
+       struct enic_fm_fet *fet;
+
+       if (enic->fm == NULL)
+               return;
+       ENICPMD_FUNC_TRACE();
+       fm = enic->fm;
+       enic_fet_free(fm, fm->default_eg_fet);
+       enic_fet_free(fm, fm->default_ig_fet);
+       /* Free all exact match tables still open */
+       while (!TAILQ_EMPTY(&fm->fet_list)) {
+               fet = TAILQ_FIRST(&fm->fet_list);
+               enic_fet_free(fm, fet);
+       }
+       enic_fm_free_tcam_tables(fm);
+       enic_fm_free_all_counters(fm);
+       enic_free_consistent(enic, sizeof(union enic_flowman_cmd_mem),
+               fm->cmd.va, fm->cmd.pa);
+       fm->cmd.va = NULL;
+       free(fm);
+       enic->fm = NULL;
+}
+
+const struct rte_flow_ops enic_fm_flow_ops = {
+       .validate = enic_fm_flow_validate,
+       .create = enic_fm_flow_create,
+       .destroy = enic_fm_flow_destroy,
+       .flush = enic_fm_flow_flush,
+       .query = enic_fm_flow_query,
+};