net/mlx4: merge flow creation and validation code
[dpdk.git] / drivers / net / mlx4 / mlx4_flow.c
index ab37e7d..ac66444 100644 (file)
  *   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_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;
+/** Processor structure associated with a flow item. */
+struct mlx4_flow_proc_item {
        /** Bit-masks corresponding to the possibilities for the item. */
        const void *mask;
        /**
@@ -88,19 +110,19 @@ struct mlx4_flow_items {
         *   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.
+        * @param flow
+        *   Flow rule handle to update.
         *
         * @return
         *   0 on success, negative value otherwise.
         */
        int (*convert)(const struct rte_flow_item *item,
                       const void *default_mask,
-                      void *data);
+                      struct rte_flow *flow);
        /** 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 {
@@ -108,14 +130,6 @@ struct rte_flow_drop {
        struct ibv_cq *cq; /**< Verbs completion queue. */
 };
 
-/** 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 Ethernet item to Verbs specification.
  *
@@ -123,24 +137,23 @@ static const enum rte_flow_action_type valid_actions[] = {
  *   Item specification.
  * @param default_mask[in]
  *   Default bit-masks to use when item->mask is not provided.
- * @param data[in, out]
- *   User structure.
+ * @param flow[in, out]
+ *   Flow rule handle to update.
  */
 static int
 mlx4_flow_create_eth(const struct rte_flow_item *item,
                     const void *default_mask,
-                    void *data)
+                    struct rte_flow *flow)
 {
        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;
        struct ibv_flow_spec_eth *eth;
        const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
        unsigned int i;
 
        ++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,
@@ -170,21 +183,21 @@ mlx4_flow_create_eth(const struct rte_flow_item *item,
  *   Item specification.
  * @param default_mask[in]
  *   Default bit-masks to use when item->mask is not provided.
- * @param data[in, out]
- *   User structure.
+ * @param flow[in, out]
+ *   Flow rule handle to update.
  */
 static int
 mlx4_flow_create_vlan(const struct rte_flow_item *item,
                      const void *default_mask,
-                     void *data)
+                     struct rte_flow *flow)
 {
        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;
        struct ibv_flow_spec_eth *eth;
        const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
 
-       eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
+       eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
+                      eth_size);
        if (!spec)
                return 0;
        if (!mask)
@@ -202,23 +215,22 @@ mlx4_flow_create_vlan(const struct rte_flow_item *item,
  *   Item specification.
  * @param default_mask[in]
  *   Default bit-masks to use when item->mask is not provided.
- * @param data[in, out]
- *   User structure.
+ * @param flow[in, out]
+ *   Flow rule handle to update.
  */
 static int
 mlx4_flow_create_ipv4(const struct rte_flow_item *item,
                      const void *default_mask,
-                     void *data)
+                     struct rte_flow *flow)
 {
        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;
        struct ibv_flow_spec_ipv4 *ipv4;
        unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4);
 
        ++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,
@@ -248,23 +260,22 @@ mlx4_flow_create_ipv4(const struct rte_flow_item *item,
  *   Item specification.
  * @param default_mask[in]
  *   Default bit-masks to use when item->mask is not provided.
- * @param data[in, out]
- *   User structure.
+ * @param flow[in, out]
+ *   Flow rule handle to update.
  */
 static int
 mlx4_flow_create_udp(const struct rte_flow_item *item,
                     const void *default_mask,
-                    void *data)
+                    struct rte_flow *flow)
 {
        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;
        struct ibv_flow_spec_tcp_udp *udp;
        unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 
        ++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,
@@ -290,23 +301,22 @@ mlx4_flow_create_udp(const struct rte_flow_item *item,
  *   Item specification.
  * @param default_mask[in]
  *   Default bit-masks to use when item->mask is not provided.
- * @param data[in, out]
- *   User structure.
+ * @param flow[in, out]
+ *   Flow rule handle to update.
  */
 static int
 mlx4_flow_create_tcp(const struct rte_flow_item *item,
                     const void *default_mask,
-                    void *data)
+                    struct rte_flow *flow)
 {
        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;
        struct ibv_flow_spec_tcp_udp *tcp;
        unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 
        ++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,
@@ -464,14 +474,13 @@ mlx4_flow_validate_tcp(const struct rte_flow_item *item,
 }
 
 /** 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,
+               .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
+                                      RTE_FLOW_ITEM_TYPE_IPV4),
                .mask = &(const struct rte_flow_item_eth){
                        .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
                        .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
@@ -483,15 +492,10 @@ static const struct mlx4_flow_items mlx4_flow_items[] = {
                .dst_sz = sizeof(struct ibv_flow_spec_eth),
        },
        [RTE_FLOW_ITEM_TYPE_VLAN] = {
-               .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4),
-               .actions = valid_actions,
+               .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
                .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
+                       /* Only TCI VID matching is supported. */
+                       .tci = RTE_BE16(0x0fff),
                },
                .mask_sz = sizeof(struct rte_flow_item_vlan),
                .validate = mlx4_flow_validate_vlan,
@@ -499,13 +503,12 @@ static const struct mlx4_flow_items mlx4_flow_items[] = {
                .dst_sz = 0,
        },
        [RTE_FLOW_ITEM_TYPE_IPV4] = {
-               .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
-                              RTE_FLOW_ITEM_TYPE_TCP),
-               .actions = valid_actions,
+               .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
+                                      RTE_FLOW_ITEM_TYPE_TCP),
                .mask = &(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,
@@ -515,11 +518,10 @@ static const struct mlx4_flow_items mlx4_flow_items[] = {
                .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
        },
        [RTE_FLOW_ITEM_TYPE_UDP] = {
-               .actions = valid_actions,
                .mask = &(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,
@@ -529,11 +531,10 @@ static const struct mlx4_flow_items mlx4_flow_items[] = {
                .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
        },
        [RTE_FLOW_ITEM_TYPE_TCP] = {
-               .actions = valid_actions,
                .mask = &(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,
@@ -545,222 +546,183 @@ static const struct mlx4_flow_items mlx4_flow_items[] = {
 };
 
 /**
- * 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;
+       uint32_t priority_override = 0;
+
+       if (attr->group)
+               return rte_flow_error_set
+                       (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
+                        NULL, "groups are not supported");
+       if (priv->isolated)
+               priority_override = attr->priority;
+       else if (attr->priority)
+               return rte_flow_error_set
+                       (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
+                        NULL,
+                        "priorities are not supported outside isolated mode");
+       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;
                /*
                 * 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;
-                       }
+               if (!item->spec && item->type == RTE_FLOW_ITEM_TYPE_ETH) {
+                       const struct rte_flow_item *next = item + 1;
+
+                       if (next->type)
+                               return rte_flow_error_set
+                                       (error, ENOTSUP,
+                                        RTE_FLOW_ERROR_TYPE_ITEM, item,
+                                        "the rule requires an Ethernet spec");
                }
-               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];
+               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)
+               if (!next)
                        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)
-                       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 validation once, while handle is not allocated. */
+               if (flow == &temp) {
+                       err = proc->validate(item, proc->mask, proc->mask_sz);
+                       if (err)
+                               goto exit_item_not_supported;
+               } else if (proc->convert) {
+                       err = proc->convert(item,
+                                           (proc->default_mask ?
+                                            proc->default_mask :
+                                            proc->mask),
+                                           flow);
                        if (err)
                                goto exit_item_not_supported;
                }
-               flow->offset += cur_item->dst_sz;
+               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) {
+       /* Use specified priority level when in isolated mode. */
+       if (priv->isolated && flow != &temp)
+               flow->ibv_attr->priority = priority_override;
+       /* Go over actions list. */
+       for (action = actions; action->type; ++action) {
+               switch (action->type) {
+                       const struct rte_flow_action_queue *queue;
+
+               case 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;
-
-                       if (!queue || (queue->index > (priv->rxqs_n - 1)))
+               case RTE_FLOW_ACTION_TYPE_DROP:
+                       flow->drop = 1;
+                       break;
+               case RTE_FLOW_ACTION_TYPE_QUEUE:
+                       queue = action->conf;
+                       if (queue->index >= priv->dev->data->nb_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;
-                       }
-                       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;
-                       }
-                       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;
-                       }
-                       /* 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;
-                               }
-                       }
-                       action.queue = 1;
-                       action.queues_n = rss->num;
-                       for (i = 0; i < rss->num; ++i)
-                               action.queues[i] = rss->queue[i];
-               } else {
+                       flow->queue = 1;
+                       flow->queue_id = queue->index;
+                       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->queue && !flow->drop)
+               return rte_flow_error_set
+                       (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                        NULL, "no valid action");
+       /* Validation ends here. */
+       if (!addr)
+               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),
+               };
+               *flow->ibv_attr = (struct ibv_flow_attr){
+                       .type = IBV_FLOW_ATTR_NORMAL,
+                       .size = sizeof(*flow->ibv_attr),
+                       .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, "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, "action not supported");
 }
 
 /**
@@ -769,21 +731,16 @@ 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;
+       return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 }
 
 /**
@@ -826,29 +783,21 @@ mlx4_flow_create_drop_queue(struct priv *priv)
                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,
-                             });
+       cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 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,
-                             });
+       qp = ibv_create_qp(priv->pd,
+                          &(struct ibv_qp_init_attr){
+                               .send_cq = cq,
+                               .recv_cq = cq,
+                               .cap = {
+                                       .max_recv_wr = 1,
+                                       .max_recv_sge = 1,
+                               },
+                               .qp_type = IBV_QPT_RAW_PACKET,
+                          });
        if (!qp) {
                ERROR("Cannot create drop QP");
                goto err_create_qp;
@@ -868,259 +817,66 @@ err:
 }
 
 /**
- * Get RSS parent rxq structure for given queues.
- *
- * Creates a new or returns an existed one.
+ * Toggle a configured flow rule.
  *
  * @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.
- */
-static struct rxq *
-priv_parent_get(struct priv *priv,
-               uint16_t queues[],
-               uint16_t children_n,
-               struct rte_flow_error *error)
-{
-       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;
-}
-
-/**
- * Complete flow rule creation.
- *
- * @param priv
- *   Pointer to private structure.
- * @param ibv_attr
- *   Verbs flow attributes.
- * @param action
- *   Target action structure.
- * @param[out] error
- *   Perform verbose error reporting if not NULL.
- *
- * @return
- *   A flow if the rule could be created.
- */
-static struct rte_flow *
-priv_flow_create_action_queue(struct priv *priv,
-                             struct ibv_flow_attr *ibv_attr,
-                             struct mlx4_flow_action *action,
-                             struct rte_flow_error *error)
-{
-       struct ibv_qp *qp;
-       struct rte_flow *rte_flow;
-       struct rxq *rxq_parent = NULL;
-
-       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 (action->drop) {
-               qp = priv->flow_drop_queue ? priv->flow_drop_queue->qp : NULL;
-       } else {
-               int ret;
-               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;
-               }
-               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;
-                               }
-                       }
-               }
-               qp = action->queues_n > 1 ? rxq_parent->qp : rxq->qp;
-               rte_flow->qp = qp;
-       }
-       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;
-       }
-       return rte_flow;
-
-error:
-       if (rxq_parent)
-               rxq_parent_cleanup(rxq_parent);
-       rte_free(rte_flow);
-       return NULL;
-}
-
-/**
- * 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 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 on success, NULL otherwise.
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
-static struct rte_flow *
-priv_flow_create(struct priv *priv,
-                const struct rte_flow_attr *attr,
-                const struct rte_flow_item items[],
-                const struct rte_flow_action actions[],
+static int
+mlx4_flow_toggle(struct priv *priv,
+                struct rte_flow *flow,
+                int enable,
                 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 ibv_qp *qp = NULL;
+       const char *msg;
        int err;
 
-       err = priv_flow_validate(priv, attr, items, 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;
+       if (!enable) {
+               if (!flow->ibv_flow)
+                       return 0;
+               claim_zero(ibv_destroy_flow(flow->ibv_flow));
+               flow->ibv_flow = NULL;
+               return 0;
        }
-       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;
-
-                       action.queue = 1;
-                       action.queues_n = rss->num;
-                       for (i = 0; i < rss->num; ++i)
-                               action.queues[i] = rss->queue[i];
-               } else {
-                       rte_flow_error_set(error, ENOTSUP,
-                                          RTE_FLOW_ERROR_TYPE_ACTION,
-                                          actions, "unsupported action");
-                       goto exit;
+       if (flow->ibv_flow)
+               return 0;
+       assert(flow->queue ^ flow->drop);
+       if (flow->queue) {
+               struct rxq *rxq;
+
+               assert(flow->queue_id < priv->dev->data->nb_rx_queues);
+               rxq = priv->dev->data->rx_queues[flow->queue_id];
+               if (!rxq) {
+                       err = EINVAL;
+                       msg = "target queue must be configured first";
+                       goto error;
                }
+               qp = rxq->qp;
        }
-       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);
-       return NULL;
+       if (flow->drop) {
+               assert(priv->flow_drop_queue);
+               qp = priv->flow_drop_queue->qp;
+       }
+       assert(qp);
+       assert(flow->ibv_attr);
+       flow->ibv_flow = ibv_create_flow(qp, flow->ibv_attr);
+       if (flow->ibv_flow)
+               return 0;
+       err = errno;
+       msg = "flow rule rejected by device";
+error:
+       return rte_flow_error_set
+               (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
 }
 
 /**
@@ -1129,176 +885,132 @@ exit:
  * @see rte_flow_create()
  * @see rte_flow_ops
  */
-struct rte_flow *
+static 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_item pattern[],
                 const struct rte_flow_action actions[],
                 struct rte_flow_error *error)
 {
        struct priv *priv = dev->data->dev_private;
        struct rte_flow *flow;
+       int err;
 
-       priv_lock(priv);
-       flow = priv_flow_create(priv, attr, items, actions, error);
-       if (flow) {
+       err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
+       if (err)
+               return NULL;
+       err = mlx4_flow_toggle(priv, flow, priv->started, error);
+       if (!err) {
                LIST_INSERT_HEAD(&priv->flows, flow, next);
-               DEBUG("Flow created %p", (void *)flow);
+               return flow;
        }
-       priv_unlock(priv);
-       return flow;
+       rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
+                          error->message);
+       rte_free(flow);
+       return NULL;
 }
 
 /**
- * @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.
+ * Configure isolated mode.
  *
- * @return
- *   0 on success, a negative value on error.
+ * @see rte_flow_isolate()
+ * @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);
-               return -rte_errno;
-       }
+       if (!!enable == !!priv->isolated)
+               return 0;
        priv->isolated = !!enable;
-       priv_unlock(priv);
+       if (enable) {
+               mlx4_mac_addr_del(priv);
+       } else if (mlx4_mac_addr_add(priv) < 0) {
+               priv->isolated = 1;
+               return rte_flow_error_set(error, rte_errno,
+                                         RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                         NULL, "cannot leave isolated mode");
+       }
        return 0;
 }
 
 /**
- * Destroy a flow.
- *
- * @param priv
- *   Pointer to private structure.
- * @param[in] flow
- *   Flow to destroy.
- */
-static void
-priv_flow_destroy(struct priv *priv, struct rte_flow *flow)
-{
-       (void)priv;
-       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);
-       rte_free(flow);
-}
-
-/**
- * Destroy a flow.
+ * Destroy a flow rule.
  *
  * @see rte_flow_destroy()
  * @see rte_flow_ops
  */
-int
+static int
 mlx4_flow_destroy(struct rte_eth_dev *dev,
                  struct rte_flow *flow,
                  struct rte_flow_error *error)
 {
        struct priv *priv = dev->data->dev_private;
+       int err = mlx4_flow_toggle(priv, flow, 0, error);
 
-       (void)error;
-       priv_lock(priv);
-       priv_flow_destroy(priv, flow);
-       priv_unlock(priv);
+       if (err)
+               return err;
+       LIST_REMOVE(flow, next);
+       rte_free(flow);
        return 0;
 }
 
 /**
- * Destroy all flows.
- *
- * @param priv
- *   Pointer to private structure.
- */
-static void
-priv_flow_flush(struct priv *priv)
-{
-       while (!LIST_EMPTY(&priv->flows)) {
-               struct rte_flow *flow;
-
-               flow = LIST_FIRST(&priv->flows);
-               priv_flow_destroy(priv, flow);
-       }
-}
-
-/**
- * Destroy all flows.
+ * Destroy all flow rules.
  *
  * @see rte_flow_flush()
  * @see rte_flow_ops
  */
-int
+static int
 mlx4_flow_flush(struct rte_eth_dev *dev,
                struct rte_flow_error *error)
 {
        struct priv *priv = dev->data->dev_private;
 
-       (void)error;
-       priv_lock(priv);
-       priv_flow_flush(priv);
-       priv_unlock(priv);
+       while (!LIST_EMPTY(&priv->flows)) {
+               struct rte_flow *flow;
+
+               flow = LIST_FIRST(&priv->flows);
+               mlx4_flow_destroy(dev, flow, error);
+       }
        return 0;
 }
 
 /**
- * Remove all flows.
- *
- * Called by dev_stop() to remove all flows.
+ * Disable flow rules.
  *
  * @param priv
  *   Pointer to private structure.
  */
 void
-mlx4_priv_flow_stop(struct priv *priv)
+mlx4_flow_stop(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);
+               claim_zero(mlx4_flow_toggle(priv, flow, 0, NULL));
        }
        mlx4_flow_destroy_drop_queue(priv);
 }
 
 /**
- * Add all flows.
+ * Enable flow rules.
  *
  * @param priv
  *   Pointer to private structure.
  *
  * @return
- *   0 on success, a errno value otherwise and rte_errno is set.
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 int
-mlx4_priv_flow_start(struct priv *priv)
+mlx4_flow_start(struct priv *priv)
 {
        int ret;
-       struct ibv_qp *qp;
        struct rte_flow *flow;
 
        ret = mlx4_flow_create_drop_queue(priv);
@@ -1307,14 +1019,55 @@ mlx4_priv_flow_start(struct priv *priv)
        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;
+               ret = mlx4_flow_toggle(priv, flow, 1, NULL);
+               if (unlikely(ret)) {
+                       mlx4_flow_stop(priv);
+                       return ret;
                }
-               DEBUG("Flow %p applied", (void *)flow);
        }
        return 0;
 }
+
+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,
+};
+
+/**
+ * Manage filter operations.
+ *
+ * @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, negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx4_filter_ctrl(struct rte_eth_dev *dev,
+                enum rte_filter_type filter_type,
+                enum rte_filter_op filter_op,
+                void *arg)
+{
+       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;
+       }
+       rte_errno = ENOTSUP;
+       return -rte_errno;
+}