net/mlx5: fix mark enabling for Rx
[dpdk.git] / drivers / net / mlx5 / mlx5_flow.c
index 9904bc5..d7cb1eb 100644 (file)
@@ -164,105 +164,143 @@ mlx5_flow_is_rss_expandable_item(const struct rte_flow_item *item)
        return false;
 }
 
+/**
+ * Network Service Header (NSH) and its next protocol values
+ * are described in RFC-8393.
+ */
+static enum rte_flow_item_type
+mlx5_nsh_proto_to_item_type(uint8_t proto_spec, uint8_t proto_mask)
+{
+       enum rte_flow_item_type type;
+
+       switch (proto_mask & proto_spec) {
+       case RTE_VXLAN_GPE_TYPE_IPV4:
+               type = RTE_FLOW_ITEM_TYPE_IPV4;
+               break;
+       case RTE_VXLAN_GPE_TYPE_IPV6:
+               type = RTE_VXLAN_GPE_TYPE_IPV6;
+               break;
+       case RTE_VXLAN_GPE_TYPE_ETH:
+               type = RTE_FLOW_ITEM_TYPE_ETH;
+               break;
+       default:
+               type = RTE_FLOW_ITEM_TYPE_END;
+       }
+       return type;
+}
+
+static enum rte_flow_item_type
+mlx5_inet_proto_to_item_type(uint8_t proto_spec, uint8_t proto_mask)
+{
+       enum rte_flow_item_type type;
+
+       switch (proto_mask & proto_spec) {
+       case IPPROTO_UDP:
+               type = RTE_FLOW_ITEM_TYPE_UDP;
+               break;
+       case IPPROTO_TCP:
+               type = RTE_FLOW_ITEM_TYPE_TCP;
+               break;
+       case IPPROTO_IP:
+               type = RTE_FLOW_ITEM_TYPE_IPV4;
+               break;
+       case IPPROTO_IPV6:
+               type = RTE_FLOW_ITEM_TYPE_IPV6;
+               break;
+       default:
+               type = RTE_FLOW_ITEM_TYPE_END;
+       }
+       return type;
+}
+
+static enum rte_flow_item_type
+mlx5_ethertype_to_item_type(rte_be16_t type_spec,
+                           rte_be16_t type_mask, bool is_tunnel)
+{
+       enum rte_flow_item_type type;
+
+       switch (rte_be_to_cpu_16(type_spec & type_mask)) {
+       case RTE_ETHER_TYPE_TEB:
+               type = is_tunnel ?
+                      RTE_FLOW_ITEM_TYPE_ETH : RTE_FLOW_ITEM_TYPE_END;
+               break;
+       case RTE_ETHER_TYPE_VLAN:
+               type = !is_tunnel ?
+                      RTE_FLOW_ITEM_TYPE_VLAN : RTE_FLOW_ITEM_TYPE_END;
+               break;
+       case RTE_ETHER_TYPE_IPV4:
+               type = RTE_FLOW_ITEM_TYPE_IPV4;
+               break;
+       case RTE_ETHER_TYPE_IPV6:
+               type = RTE_FLOW_ITEM_TYPE_IPV6;
+               break;
+       default:
+               type = RTE_FLOW_ITEM_TYPE_END;
+       }
+       return type;
+}
+
 static enum rte_flow_item_type
 mlx5_flow_expand_rss_item_complete(const struct rte_flow_item *item)
 {
-       enum rte_flow_item_type ret = RTE_FLOW_ITEM_TYPE_VOID;
-       uint16_t ether_type = 0;
-       uint16_t ether_type_m;
-       uint8_t ip_next_proto = 0;
-       uint8_t ip_next_proto_m;
+#define MLX5_XSET_ITEM_MASK_SPEC(type, fld)                              \
+       do {                                                             \
+               const void *m = item->mask;                              \
+               const void *s = item->spec;                              \
+               mask = m ?                                               \
+                       ((const struct rte_flow_item_##type *)m)->fld :  \
+                       rte_flow_item_##type##_mask.fld;                 \
+               spec = ((const struct rte_flow_item_##type *)s)->fld;    \
+       } while (0)
+
+       enum rte_flow_item_type ret;
+       uint16_t spec, mask;
 
        if (item == NULL || item->spec == NULL)
-               return ret;
+               return RTE_FLOW_ITEM_TYPE_VOID;
        switch (item->type) {
        case RTE_FLOW_ITEM_TYPE_ETH:
-               if (item->mask)
-                       ether_type_m = ((const struct rte_flow_item_eth *)
-                                               (item->mask))->type;
-               else
-                       ether_type_m = rte_flow_item_eth_mask.type;
-               if (ether_type_m != RTE_BE16(0xFFFF))
-                       break;
-               ether_type = ((const struct rte_flow_item_eth *)
-                               (item->spec))->type;
-               if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV4)
-                       ret = RTE_FLOW_ITEM_TYPE_IPV4;
-               else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV6)
-                       ret = RTE_FLOW_ITEM_TYPE_IPV6;
-               else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_VLAN)
-                       ret = RTE_FLOW_ITEM_TYPE_VLAN;
-               else
-                       ret = RTE_FLOW_ITEM_TYPE_END;
+               MLX5_XSET_ITEM_MASK_SPEC(eth, type);
+               if (!mask)
+                       return RTE_FLOW_ITEM_TYPE_VOID;
+               ret = mlx5_ethertype_to_item_type(spec, mask, false);
                break;
        case RTE_FLOW_ITEM_TYPE_VLAN:
-               if (item->mask)
-                       ether_type_m = ((const struct rte_flow_item_vlan *)
-                                               (item->mask))->inner_type;
-               else
-                       ether_type_m = rte_flow_item_vlan_mask.inner_type;
-               if (ether_type_m != RTE_BE16(0xFFFF))
-                       break;
-               ether_type = ((const struct rte_flow_item_vlan *)
-                               (item->spec))->inner_type;
-               if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV4)
-                       ret = RTE_FLOW_ITEM_TYPE_IPV4;
-               else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV6)
-                       ret = RTE_FLOW_ITEM_TYPE_IPV6;
-               else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_VLAN)
-                       ret = RTE_FLOW_ITEM_TYPE_VLAN;
-               else
-                       ret = RTE_FLOW_ITEM_TYPE_END;
+               MLX5_XSET_ITEM_MASK_SPEC(vlan, inner_type);
+               if (!mask)
+                       return RTE_FLOW_ITEM_TYPE_VOID;
+               ret = mlx5_ethertype_to_item_type(spec, mask, false);
                break;
        case RTE_FLOW_ITEM_TYPE_IPV4:
-               if (item->mask)
-                       ip_next_proto_m = ((const struct rte_flow_item_ipv4 *)
-                                       (item->mask))->hdr.next_proto_id;
-               else
-                       ip_next_proto_m =
-                               rte_flow_item_ipv4_mask.hdr.next_proto_id;
-               if (ip_next_proto_m != 0xFF)
-                       break;
-               ip_next_proto = ((const struct rte_flow_item_ipv4 *)
-                               (item->spec))->hdr.next_proto_id;
-               if (ip_next_proto == IPPROTO_UDP)
-                       ret = RTE_FLOW_ITEM_TYPE_UDP;
-               else if (ip_next_proto == IPPROTO_TCP)
-                       ret = RTE_FLOW_ITEM_TYPE_TCP;
-               else if (ip_next_proto == IPPROTO_IP)
-                       ret = RTE_FLOW_ITEM_TYPE_IPV4;
-               else if (ip_next_proto == IPPROTO_IPV6)
-                       ret = RTE_FLOW_ITEM_TYPE_IPV6;
-               else
-                       ret = RTE_FLOW_ITEM_TYPE_END;
+               MLX5_XSET_ITEM_MASK_SPEC(ipv4, hdr.next_proto_id);
+               if (!mask)
+                       return RTE_FLOW_ITEM_TYPE_VOID;
+               ret = mlx5_inet_proto_to_item_type(spec, mask);
                break;
        case RTE_FLOW_ITEM_TYPE_IPV6:
-               if (item->mask)
-                       ip_next_proto_m = ((const struct rte_flow_item_ipv6 *)
-                                               (item->mask))->hdr.proto;
-               else
-                       ip_next_proto_m =
-                               rte_flow_item_ipv6_mask.hdr.proto;
-               if (ip_next_proto_m != 0xFF)
-                       break;
-               ip_next_proto = ((const struct rte_flow_item_ipv6 *)
-                               (item->spec))->hdr.proto;
-               if (ip_next_proto == IPPROTO_UDP)
-                       ret = RTE_FLOW_ITEM_TYPE_UDP;
-               else if (ip_next_proto == IPPROTO_TCP)
-                       ret = RTE_FLOW_ITEM_TYPE_TCP;
-               else if (ip_next_proto == IPPROTO_IP)
-                       ret = RTE_FLOW_ITEM_TYPE_IPV4;
-               else if (ip_next_proto == IPPROTO_IPV6)
-                       ret = RTE_FLOW_ITEM_TYPE_IPV6;
-               else
-                       ret = RTE_FLOW_ITEM_TYPE_END;
+               MLX5_XSET_ITEM_MASK_SPEC(ipv6, hdr.proto);
+               if (!mask)
+                       return RTE_FLOW_ITEM_TYPE_VOID;
+               ret = mlx5_inet_proto_to_item_type(spec, mask);
+               break;
+       case RTE_FLOW_ITEM_TYPE_GENEVE:
+               MLX5_XSET_ITEM_MASK_SPEC(geneve, protocol);
+               ret = mlx5_ethertype_to_item_type(spec, mask, true);
+               break;
+       case RTE_FLOW_ITEM_TYPE_GRE:
+               MLX5_XSET_ITEM_MASK_SPEC(gre, protocol);
+               ret = mlx5_ethertype_to_item_type(spec, mask, true);
+               break;
+       case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
+               MLX5_XSET_ITEM_MASK_SPEC(vxlan_gpe, protocol);
+               ret = mlx5_nsh_proto_to_item_type(spec, mask);
                break;
        default:
                ret = RTE_FLOW_ITEM_TYPE_VOID;
                break;
        }
        return ret;
+#undef MLX5_XSET_ITEM_MASK_SPEC
 }
 
 static const int *
@@ -530,7 +568,8 @@ enum mlx5_expansion {
        MLX5_EXPANSION_IPV6_UDP,
        MLX5_EXPANSION_IPV6_TCP,
        MLX5_EXPANSION_IPV6_FRAG_EXT,
-       MLX5_EXPANSION_GTP
+       MLX5_EXPANSION_GTP,
+       MLX5_EXPANSION_GENEVE,
 };
 
 /** Supported expansion of items. */
@@ -574,6 +613,7 @@ static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
                .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VXLAN,
                                                  MLX5_EXPANSION_VXLAN_GPE,
                                                  MLX5_EXPANSION_MPLS,
+                                                 MLX5_EXPANSION_GENEVE,
                                                  MLX5_EXPANSION_GTP),
                .type = RTE_FLOW_ITEM_TYPE_UDP,
                .rss_types = RTE_ETH_RSS_NONFRAG_IPV4_UDP,
@@ -598,6 +638,7 @@ static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
                .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VXLAN,
                                                  MLX5_EXPANSION_VXLAN_GPE,
                                                  MLX5_EXPANSION_MPLS,
+                                                 MLX5_EXPANSION_GENEVE,
                                                  MLX5_EXPANSION_GTP),
                .type = RTE_FLOW_ITEM_TYPE_UDP,
                .rss_types = RTE_ETH_RSS_NONFRAG_IPV6_UDP,
@@ -628,7 +669,8 @@ static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
                .type = RTE_FLOW_ITEM_TYPE_VXLAN_GPE,
        },
        [MLX5_EXPANSION_GRE] = {
-               .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
+               .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
+                                                 MLX5_EXPANSION_IPV4,
                                                  MLX5_EXPANSION_IPV6,
                                                  MLX5_EXPANSION_GRE_KEY,
                                                  MLX5_EXPANSION_MPLS),
@@ -701,6 +743,12 @@ static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
                                                  MLX5_EXPANSION_IPV6),
                .type = RTE_FLOW_ITEM_TYPE_GTP,
        },
+       [MLX5_EXPANSION_GENEVE] = {
+               .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
+                                                 MLX5_EXPANSION_IPV4,
+                                                 MLX5_EXPANSION_IPV6),
+               .type = RTE_FLOW_ITEM_TYPE_GENEVE,
+       },
 };
 
 static struct rte_flow_action_handle *
@@ -748,6 +796,14 @@ mlx5_flow_tunnel_get_restore_info(struct rte_eth_dev *dev,
                                  struct rte_mbuf *m,
                                  struct rte_flow_restore_info *info,
                                  struct rte_flow_error *err);
+static struct rte_flow_item_flex_handle *
+mlx5_flow_flex_item_create(struct rte_eth_dev *dev,
+                          const struct rte_flow_item_flex_conf *conf,
+                          struct rte_flow_error *error);
+static int
+mlx5_flow_flex_item_release(struct rte_eth_dev *dev,
+                           const struct rte_flow_item_flex_handle *handle,
+                           struct rte_flow_error *error);
 
 static const struct rte_flow_ops mlx5_flow_ops = {
        .validate = mlx5_flow_validate,
@@ -767,6 +823,8 @@ static const struct rte_flow_ops mlx5_flow_ops = {
        .tunnel_action_decap_release = mlx5_flow_tunnel_action_release,
        .tunnel_item_release = mlx5_flow_tunnel_item_release,
        .get_restore_info = mlx5_flow_tunnel_get_restore_info,
+       .flex_item_create = mlx5_flow_flex_item_create,
+       .flex_item_release = mlx5_flow_flex_item_release,
 };
 
 /* Tunnel information. */
@@ -897,6 +955,7 @@ mlx5_flow_get_reg_id(struct rte_eth_dev *dev,
        case MLX5_MTR_COLOR:
        case MLX5_ASO_FLOW_HIT:
        case MLX5_ASO_CONNTRACK:
+       case MLX5_SAMPLE_ID:
                /* All features use the same REG_C. */
                MLX5_ASSERT(priv->mtr_color_reg != REG_NON);
                return priv->mtr_color_reg;
@@ -1162,7 +1221,7 @@ flow_rxq_tunnel_ptype_update(struct mlx5_rxq_ctrl *rxq_ctrl)
 }
 
 /**
- * Set the Rx queue flags (Mark/Flag and Tunnel Ptypes) according to the devive
+ * Set the Rx queue flags (Mark/Flag and Tunnel Ptypes) according to the device
  * flow.
  *
  * @param[in] dev
@@ -1175,7 +1234,6 @@ flow_drv_rxq_flags_set(struct rte_eth_dev *dev,
                       struct mlx5_flow_handle *dev_handle)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       const int mark = dev_handle->mark;
        const int tunnel = !!(dev_handle->layers & MLX5_FLOW_LAYER_TUNNEL);
        struct mlx5_ind_table_obj *ind_tbl = NULL;
        unsigned int i;
@@ -1200,24 +1258,16 @@ flow_drv_rxq_flags_set(struct rte_eth_dev *dev,
                return;
        for (i = 0; i != ind_tbl->queues_n; ++i) {
                int idx = ind_tbl->queues[i];
-               struct mlx5_rxq_ctrl *rxq_ctrl =
-                       container_of((*priv->rxqs)[idx],
-                                    struct mlx5_rxq_ctrl, rxq);
+               struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev, idx);
 
+               MLX5_ASSERT(rxq_ctrl != NULL);
+               if (rxq_ctrl == NULL)
+                       continue;
                /*
                 * To support metadata register copy on Tx loopback,
                 * this must be always enabled (metadata may arive
                 * from other port - not from local flows only.
                 */
-               if (priv->config.dv_flow_en &&
-                   priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
-                   mlx5_flow_ext_mreg_supported(dev)) {
-                       rxq_ctrl->rxq.mark = 1;
-                       rxq_ctrl->flow_mark_n = 1;
-               } else if (mark) {
-                       rxq_ctrl->rxq.mark = 1;
-                       rxq_ctrl->flow_mark_n++;
-               }
                if (tunnel) {
                        unsigned int j;
 
@@ -1235,6 +1285,20 @@ flow_drv_rxq_flags_set(struct rte_eth_dev *dev,
        }
 }
 
+static void
+flow_rxq_mark_flag_set(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_rxq_ctrl *rxq_ctrl;
+
+       if (priv->mark_enabled)
+               return;
+       LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
+               rxq_ctrl->rxq.mark = 1;
+       }
+       priv->mark_enabled = 1;
+}
+
 /**
  * Set the Rx queue flags (Mark/Flag and Tunnel Ptypes) for a flow
  *
@@ -1249,7 +1313,11 @@ flow_rxq_flags_set(struct rte_eth_dev *dev, struct rte_flow *flow)
        struct mlx5_priv *priv = dev->data->dev_private;
        uint32_t handle_idx;
        struct mlx5_flow_handle *dev_handle;
+       struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
 
+       MLX5_ASSERT(wks);
+       if (wks->mark)
+               flow_rxq_mark_flag_set(dev);
        SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
                       handle_idx, dev_handle, next)
                flow_drv_rxq_flags_set(dev, dev_handle);
@@ -1269,7 +1337,6 @@ flow_drv_rxq_flags_trim(struct rte_eth_dev *dev,
                        struct mlx5_flow_handle *dev_handle)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       const int mark = dev_handle->mark;
        const int tunnel = !!(dev_handle->layers & MLX5_FLOW_LAYER_TUNNEL);
        struct mlx5_ind_table_obj *ind_tbl = NULL;
        unsigned int i;
@@ -1295,19 +1362,11 @@ flow_drv_rxq_flags_trim(struct rte_eth_dev *dev,
        MLX5_ASSERT(dev->data->dev_started);
        for (i = 0; i != ind_tbl->queues_n; ++i) {
                int idx = ind_tbl->queues[i];
-               struct mlx5_rxq_ctrl *rxq_ctrl =
-                       container_of((*priv->rxqs)[idx],
-                                    struct mlx5_rxq_ctrl, rxq);
-
-               if (priv->config.dv_flow_en &&
-                   priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
-                   mlx5_flow_ext_mreg_supported(dev)) {
-                       rxq_ctrl->rxq.mark = 1;
-                       rxq_ctrl->flow_mark_n = 1;
-               } else if (mark) {
-                       rxq_ctrl->flow_mark_n--;
-                       rxq_ctrl->rxq.mark = !!rxq_ctrl->flow_mark_n;
-               }
+               struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev, idx);
+
+               MLX5_ASSERT(rxq_ctrl != NULL);
+               if (rxq_ctrl == NULL)
+                       continue;
                if (tunnel) {
                        unsigned int j;
 
@@ -1359,19 +1418,17 @@ flow_rxq_flags_clear(struct rte_eth_dev *dev)
        unsigned int i;
 
        for (i = 0; i != priv->rxqs_n; ++i) {
-               struct mlx5_rxq_ctrl *rxq_ctrl;
+               struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, i);
                unsigned int j;
 
-               if (!(*priv->rxqs)[i])
+               if (rxq == NULL || rxq->ctrl == NULL)
                        continue;
-               rxq_ctrl = container_of((*priv->rxqs)[i],
-                                       struct mlx5_rxq_ctrl, rxq);
-               rxq_ctrl->flow_mark_n = 0;
-               rxq_ctrl->rxq.mark = 0;
+               rxq->ctrl->rxq.mark = 0;
                for (j = 0; j != MLX5_FLOW_TUNNEL; ++j)
-                       rxq_ctrl->flow_tunnels_n[j] = 0;
-               rxq_ctrl->rxq.tunnel = 0;
+                       rxq->ctrl->flow_tunnels_n[j] = 0;
+               rxq->ctrl->rxq.tunnel = 0;
        }
+       priv->mark_enabled = 0;
 }
 
 /**
@@ -1384,13 +1441,15 @@ void
 mlx5_flow_rxq_dynf_metadata_set(struct rte_eth_dev *dev)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_data *data;
        unsigned int i;
 
        for (i = 0; i != priv->rxqs_n; ++i) {
-               if (!(*priv->rxqs)[i])
+               struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, i);
+               struct mlx5_rxq_data *data;
+
+               if (rxq == NULL || rxq->ctrl == NULL)
                        continue;
-               data = (*priv->rxqs)[i];
+               data = &rxq->ctrl->rxq;
                if (!rte_flow_dynf_metadata_avail()) {
                        data->dynf_meta = 0;
                        data->flow_meta_mask = 0;
@@ -1581,7 +1640,7 @@ mlx5_flow_validate_action_queue(const struct rte_flow_action *action,
                                          RTE_FLOW_ERROR_TYPE_ACTION_CONF,
                                          &queue->index,
                                          "queue index out of range");
-       if (!(*priv->rxqs)[queue->index])
+       if (mlx5_rxq_get(dev, queue->index) == NULL)
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ACTION_CONF,
                                          &queue->index,
@@ -1612,7 +1671,7 @@ mlx5_flow_validate_action_queue(const struct rte_flow_action *action,
  *   0 on success, a negative errno code on error.
  */
 static int
-mlx5_validate_rss_queues(const struct rte_eth_dev *dev,
+mlx5_validate_rss_queues(struct rte_eth_dev *dev,
                         const uint16_t *queues, uint32_t queues_n,
                         const char **error, uint32_t *queue_idx)
 {
@@ -1621,20 +1680,19 @@ mlx5_validate_rss_queues(const struct rte_eth_dev *dev,
        uint32_t i;
 
        for (i = 0; i != queues_n; ++i) {
-               struct mlx5_rxq_ctrl *rxq_ctrl;
+               struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev,
+                                                                  queues[i]);
 
                if (queues[i] >= priv->rxqs_n) {
                        *error = "queue index out of range";
                        *queue_idx = i;
                        return -EINVAL;
                }
-               if (!(*priv->rxqs)[queues[i]]) {
+               if (rxq_ctrl == NULL) {
                        *error =  "queue is not configured";
                        *queue_idx = i;
                        return -EINVAL;
                }
-               rxq_ctrl = container_of((*priv->rxqs)[queues[i]],
-                                       struct mlx5_rxq_ctrl, rxq);
                if (i == 0)
                        rxq_type = rxq_ctrl->type;
                if (rxq_type != rxq_ctrl->type) {
@@ -2963,7 +3021,7 @@ mlx5_flow_validate_item_geneve_opt(const struct rte_flow_item *item,
        if ((uint32_t)spec->option_len > MLX5_GENEVE_OPTLEN_MASK)
                return rte_flow_error_set
                        (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM, item,
-                       "Geneve TLV opt length exceeeds the limit (31)");
+                       "Geneve TLV opt length exceeds the limit (31)");
        /* Check if class type and length masks are full. */
        if (full_mask.option_class != mask->option_class ||
            full_mask.option_type != mask->option_type ||
@@ -3912,7 +3970,7 @@ find_graph_root(uint32_t rss_level)
  *  subflow.
  *
  * @param[in] dev_flow
- *   Pointer the created preifx subflow.
+ *   Pointer the created prefix subflow.
  *
  * @return
  *   The layers get from prefix subflow.
@@ -4239,7 +4297,7 @@ flow_dv_mreg_create_cb(void *tool_ctx, void *cb_ctx)
                [3] = { .type = RTE_FLOW_ACTION_TYPE_END, },
        };
 
-       /* Fill the register fileds in the flow. */
+       /* Fill the register fields in the flow. */
        ret = mlx5_flow_get_reg_id(dev, MLX5_FLOW_MARK, 0, error);
        if (ret < 0)
                return NULL;
@@ -4308,7 +4366,7 @@ flow_dv_mreg_create_cb(void *tool_ctx, void *cb_ctx)
        /*
         * The copy Flows are not included in any list. There
         * ones are referenced from other Flows and can not
-        * be applied, removed, deleted in ardbitrary order
+        * be applied, removed, deleted in arbitrary order
         * by list traversing.
         */
        mcp_res->rix_flow = flow_list_create(dev, MLX5_FLOW_TYPE_MCP,
@@ -4751,6 +4809,7 @@ flow_create_split_inner(struct rte_eth_dev *dev,
                        struct rte_flow_error *error)
 {
        struct mlx5_flow *dev_flow;
+       struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
 
        dev_flow = flow_drv_prepare(dev, flow, attr, items, actions,
                                    flow_split_info->flow_idx, error);
@@ -4765,12 +4824,14 @@ flow_create_split_inner(struct rte_eth_dev *dev,
        /*
         * If dev_flow is as one of the suffix flow, some actions in suffix
         * flow may need some user defined item layer flags, and pass the
-        * Metadate rxq mark flag to suffix flow as well.
+        * Metadata rxq mark flag to suffix flow as well.
         */
        if (flow_split_info->prefix_layers)
                dev_flow->handle->layers = flow_split_info->prefix_layers;
-       if (flow_split_info->prefix_mark)
-               dev_flow->handle->mark = 1;
+       if (flow_split_info->prefix_mark) {
+               MLX5_ASSERT(wks);
+               wks->mark = 1;
+       }
        if (sub_flow)
                *sub_flow = dev_flow;
 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
@@ -5206,6 +5267,8 @@ exit:
  *   Pointer to the Q/RSS action.
  * @param[in] actions_n
  *   Number of original actions.
+ * @param[in] mtr_sfx
+ *   Check if it is in meter suffix table.
  * @param[out] error
  *   Perform verbose error reporting if not NULL.
  *
@@ -5218,7 +5281,8 @@ flow_mreg_split_qrss_prep(struct rte_eth_dev *dev,
                          struct rte_flow_action *split_actions,
                          const struct rte_flow_action *actions,
                          const struct rte_flow_action *qrss,
-                         int actions_n, struct rte_flow_error *error)
+                         int actions_n, int mtr_sfx,
+                         struct rte_flow_error *error)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_rte_flow_action_set_tag *set_tag;
@@ -5233,15 +5297,15 @@ flow_mreg_split_qrss_prep(struct rte_eth_dev *dev,
         * - Add jump to mreg CP_TBL.
         * As a result, there will be one more action.
         */
-       ++actions_n;
        memcpy(split_actions, actions, sizeof(*split_actions) * actions_n);
+       /* Count MLX5_RTE_FLOW_ACTION_TYPE_TAG. */
+       ++actions_n;
        set_tag = (void *)(split_actions + actions_n);
        /*
-        * If tag action is not set to void(it means we are not the meter
-        * suffix flow), add the tag action. Since meter suffix flow already
-        * has the tag added.
+        * If we are not the meter suffix flow, add the tag action.
+        * Since meter suffix flow already has the tag added.
         */
-       if (split_actions[qrss_idx].type != RTE_FLOW_ACTION_TYPE_VOID) {
+       if (!mtr_sfx) {
                /*
                 * Allocate the new subflow ID. This one is unique within
                 * device and not shared with representors. Otherwise,
@@ -5274,6 +5338,12 @@ flow_mreg_split_qrss_prep(struct rte_eth_dev *dev,
                                MLX5_RTE_FLOW_ACTION_TYPE_TAG,
                        .conf = set_tag,
                };
+       } else {
+               /*
+                * If we are the suffix flow of meter, tag already exist.
+                * Set the QUEUE/RSS action to void.
+                */
+               split_actions[qrss_idx].type = RTE_FLOW_ACTION_TYPE_VOID;
        }
        /* JUMP action to jump to mreg copy table (CP_TBL). */
        jump = (void *)(set_tag + 1);
@@ -5305,7 +5375,7 @@ flow_mreg_split_qrss_prep(struct rte_eth_dev *dev,
  * @param[out] error
  *   Perform verbose error reporting if not NULL.
  * @param[in] encap_idx
- *   The encap action inndex.
+ *   The encap action index.
  *
  * @return
  *   0 on success, negative value otherwise
@@ -5570,7 +5640,7 @@ flow_sample_split_prep(struct rte_eth_dev *dev,
                /* Prepare the prefix tag action. */
                append_index++;
                set_tag = (void *)(actions_pre + actions_n + append_index);
-               ret = mlx5_flow_get_reg_id(dev, MLX5_APP_TAG, 0, error);
+               ret = mlx5_flow_get_reg_id(dev, MLX5_SAMPLE_ID, 0, error);
                if (ret < 0)
                        return ret;
                mlx5_ipool_malloc(priv->sh->ipool
@@ -5728,17 +5798,6 @@ flow_create_split_metadata(struct rte_eth_dev *dev,
                                                  RTE_FLOW_ERROR_TYPE_ACTION,
                                                  NULL, "no memory to split "
                                                  "metadata flow");
-               /*
-                * If we are the suffix flow of meter, tag already exist.
-                * Set the tag action to void.
-                */
-               if (mtr_sfx)
-                       ext_actions[qrss - actions].type =
-                                               RTE_FLOW_ACTION_TYPE_VOID;
-               else
-                       ext_actions[qrss - actions].type =
-                                               (enum rte_flow_action_type)
-                                               MLX5_RTE_FLOW_ACTION_TYPE_TAG;
                /*
                 * Create the new actions list with removed Q/RSS action
                 * and appended set tag and jump to register copy table
@@ -5746,7 +5805,8 @@ flow_create_split_metadata(struct rte_eth_dev *dev,
                 * in advance, because it is needed for set tag action.
                 */
                qrss_id = flow_mreg_split_qrss_prep(dev, ext_actions, actions,
-                                                   qrss, actions_n, error);
+                                                   qrss, actions_n,
+                                                   mtr_sfx, error);
                if (!mtr_sfx && !qrss_id) {
                        ret = -rte_errno;
                        goto exit;
@@ -5837,6 +5897,7 @@ flow_create_split_metadata(struct rte_eth_dev *dev,
                /* Add suffix subflow to execute Q/RSS. */
                flow_split_info->prefix_layers = layers;
                flow_split_info->prefix_mark = 0;
+               flow_split_info->table_id = 0;
                ret = flow_create_split_inner(dev, flow, &dev_flow,
                                              &q_attr, mtr_sfx ? items :
                                              q_items, q_actions,
@@ -6051,7 +6112,6 @@ flow_create_split_meter(struct rte_eth_dev *dev,
                        goto exit;
                }
                /* Add the prefix subflow. */
-               flow_split_info->prefix_mark = 0;
                skip_scale_restore = flow_split_info->skip_scale;
                flow_split_info->skip_scale |=
                        1 << MLX5_SCALE_JUMP_FLOW_GROUP_BIT;
@@ -6084,7 +6144,7 @@ flow_create_split_meter(struct rte_eth_dev *dev,
                                 MLX5_FLOW_TABLE_LEVEL_METER;
                flow_split_info->prefix_layers =
                                flow_get_prefix_layer_flags(dev_flow);
-               flow_split_info->prefix_mark = dev_flow->handle->mark;
+               flow_split_info->prefix_mark |= wks->mark;
                flow_split_info->table_id = MLX5_MTR_TABLE_ID_SUFFIX;
        }
        /* Add the prefix subflow. */
@@ -6150,6 +6210,7 @@ flow_create_split_sample(struct rte_eth_dev *dev,
        struct mlx5_flow_dv_sample_resource *sample_res;
        struct mlx5_flow_tbl_data_entry *sfx_tbl_data;
        struct mlx5_flow_tbl_resource *sfx_tbl;
+       struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
 #endif
        size_t act_size;
        size_t item_size;
@@ -6236,7 +6297,8 @@ flow_create_split_sample(struct rte_eth_dev *dev,
                }
                flow_split_info->prefix_layers =
                                flow_get_prefix_layer_flags(dev_flow);
-               flow_split_info->prefix_mark = dev_flow->handle->mark;
+               MLX5_ASSERT(wks);
+               flow_split_info->prefix_mark |= wks->mark;
                /* Suffix group level already be scaled with factor, set
                 * MLX5_SCALE_FLOW_GROUP_BIT of skip_scale to 1 to avoid scale
                 * again in translation.
@@ -6840,7 +6902,7 @@ flow_list_destroy(struct rte_eth_dev *dev, enum mlx5_flow_type type,
  * @param type
  *   Flow type to be flushed.
  * @param active
- *   If flushing is called avtively.
+ *   If flushing is called actively.
  */
 void
 mlx5_flow_list_flush(struct rte_eth_dev *dev, enum mlx5_flow_type type,
@@ -7764,7 +7826,6 @@ mlx5_counter_query(struct rte_eth_dev *dev, uint32_t cnt,
 static int
 mlx5_flow_create_counter_stat_mem_mng(struct mlx5_dev_ctx_shared *sh)
 {
-       struct mlx5_devx_mkey_attr mkey_attr;
        struct mlx5_counter_stats_mem_mng *mem_mng;
        volatile struct flow_counter_stats *raw_data;
        int raws_n = MLX5_CNT_CONTAINER_RESIZE + MLX5_MAX_PENDING_QUERIES;
@@ -7774,6 +7835,7 @@ mlx5_flow_create_counter_stat_mem_mng(struct mlx5_dev_ctx_shared *sh)
                        sizeof(struct mlx5_counter_stats_mem_mng);
        size_t pgsize = rte_mem_page_size();
        uint8_t *mem;
+       int ret;
        int i;
 
        if (pgsize == (size_t)-1) {
@@ -7788,23 +7850,10 @@ mlx5_flow_create_counter_stat_mem_mng(struct mlx5_dev_ctx_shared *sh)
        }
        mem_mng = (struct mlx5_counter_stats_mem_mng *)(mem + size) - 1;
        size = sizeof(*raw_data) * MLX5_COUNTERS_PER_POOL * raws_n;
-       mem_mng->umem = mlx5_os_umem_reg(sh->cdev->ctx, mem, size,
-                                                IBV_ACCESS_LOCAL_WRITE);
-       if (!mem_mng->umem) {
-               rte_errno = errno;
-               mlx5_free(mem);
-               return -rte_errno;
-       }
-       memset(&mkey_attr, 0, sizeof(mkey_attr));
-       mkey_attr.addr = (uintptr_t)mem;
-       mkey_attr.size = size;
-       mkey_attr.umem_id = mlx5_os_get_umem_id(mem_mng->umem);
-       mkey_attr.pd = sh->cdev->pdn;
-       mkey_attr.relaxed_ordering_write = sh->cmng.relaxed_ordering_write;
-       mkey_attr.relaxed_ordering_read = sh->cmng.relaxed_ordering_read;
-       mem_mng->dm = mlx5_devx_cmd_mkey_create(sh->cdev->ctx, &mkey_attr);
-       if (!mem_mng->dm) {
-               mlx5_os_umem_dereg(mem_mng->umem);
+       ret = mlx5_os_wrapped_mkey_create(sh->cdev->ctx, sh->cdev->pd,
+                                         sh->cdev->pdn, mem, size,
+                                         &mem_mng->wm);
+       if (ret) {
                rte_errno = errno;
                mlx5_free(mem);
                return -rte_errno;
@@ -7923,7 +7972,7 @@ mlx5_flow_query_alarm(void *arg)
        ret = mlx5_devx_cmd_flow_counter_query(pool->min_dcs, 0,
                                               MLX5_COUNTERS_PER_POOL,
                                               NULL, NULL,
-                                              pool->raw_hw->mem_mng->dm->id,
+                                              pool->raw_hw->mem_mng->wm.lkey,
                                               (void *)(uintptr_t)
                                               pool->raw_hw->data,
                                               sh->devx_comp,
@@ -8500,7 +8549,7 @@ mlx5_flow_dev_dump_sh_all(struct rte_eth_dev *dev,
  *   Perform verbose error reporting if not NULL. PMDs initialize this
  *   structure in case of error only.
  * @return
- *   0 on success, a nagative value otherwise.
+ *   0 on success, a negative value otherwise.
  */
 int
 mlx5_flow_dev_dump(struct rte_eth_dev *dev, struct rte_flow *flow_idx,
@@ -8978,7 +9027,7 @@ mlx5_get_tof(const struct rte_flow_item *item,
 }
 
 /**
- * tunnel offload functionalilty is defined for DV environment only
+ * tunnel offload functionality is defined for DV environment only
  */
 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 __extension__
@@ -9491,33 +9540,37 @@ err:
        return err;
 }
 
-static inline bool
+static inline int
 mlx5_flow_tunnel_validate(struct rte_eth_dev *dev,
                          struct rte_flow_tunnel *tunnel,
-                         const char *err_msg)
+                         struct rte_flow_error *error)
 {
-       err_msg = NULL;
-       if (!is_tunnel_offload_active(dev)) {
-               err_msg = "tunnel offload was not activated";
-               goto out;
-       } else if (!tunnel) {
-               err_msg = "no application tunnel";
-               goto out;
-       }
+       struct mlx5_priv *priv = dev->data->dev_private;
 
+       if (!priv->config.dv_flow_en)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
+                                         "flow DV interface is off");
+       if (!is_tunnel_offload_active(dev))
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
+                                         "tunnel offload was not activated");
+       if (!tunnel)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
+                                         "no application tunnel");
        switch (tunnel->type) {
        default:
-               err_msg = "unsupported tunnel type";
-               goto out;
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
+                                         "unsupported tunnel type");
        case RTE_FLOW_ITEM_TYPE_VXLAN:
        case RTE_FLOW_ITEM_TYPE_GRE:
        case RTE_FLOW_ITEM_TYPE_NVGRE:
        case RTE_FLOW_ITEM_TYPE_GENEVE:
                break;
        }
-
-out:
-       return !err_msg;
+       return 0;
 }
 
 static int
@@ -9527,15 +9580,11 @@ mlx5_flow_tunnel_decap_set(struct rte_eth_dev *dev,
                    uint32_t *num_of_actions,
                    struct rte_flow_error *error)
 {
-       int ret;
        struct mlx5_flow_tunnel *tunnel;
-       const char *err_msg = NULL;
-       bool verdict = mlx5_flow_tunnel_validate(dev, app_tunnel, err_msg);
+       int ret = mlx5_flow_tunnel_validate(dev, app_tunnel, error);
 
-       if (!verdict)
-               return rte_flow_error_set(error, EINVAL,
-                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
-                                         err_msg);
+       if (ret)
+               return ret;
        ret = mlx5_get_flow_tunnel(dev, app_tunnel, &tunnel);
        if (ret < 0) {
                return rte_flow_error_set(error, ret,
@@ -9554,15 +9603,11 @@ mlx5_flow_tunnel_match(struct rte_eth_dev *dev,
                       uint32_t *num_of_items,
                       struct rte_flow_error *error)
 {
-       int ret;
        struct mlx5_flow_tunnel *tunnel;
-       const char *err_msg = NULL;
-       bool verdict = mlx5_flow_tunnel_validate(dev, app_tunnel, err_msg);
+       int ret = mlx5_flow_tunnel_validate(dev, app_tunnel, error);
 
-       if (!verdict)
-               return rte_flow_error_set(error, EINVAL,
-                                         RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
-                                         err_msg);
+       if (ret)
+               return ret;
        ret = mlx5_get_flow_tunnel(dev, app_tunnel, &tunnel);
        if (ret < 0) {
                return rte_flow_error_set(error, ret,
@@ -9788,6 +9833,45 @@ mlx5_release_tunnel_hub(__rte_unused struct mlx5_dev_ctx_shared *sh,
 }
 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 
+/* Flex flow item API */
+static struct rte_flow_item_flex_handle *
+mlx5_flow_flex_item_create(struct rte_eth_dev *dev,
+                          const struct rte_flow_item_flex_conf *conf,
+                          struct rte_flow_error *error)
+{
+       static const char err_msg[] = "flex item creation unsupported";
+       struct rte_flow_attr attr = { .transfer = 0 };
+       const struct mlx5_flow_driver_ops *fops =
+                       flow_get_drv_ops(flow_get_drv_type(dev, &attr));
+
+       if (!fops->item_create) {
+               DRV_LOG(ERR, "port %u %s.", dev->data->port_id, err_msg);
+               rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
+                                  NULL, err_msg);
+               return NULL;
+       }
+       return fops->item_create(dev, conf, error);
+}
+
+static int
+mlx5_flow_flex_item_release(struct rte_eth_dev *dev,
+                           const struct rte_flow_item_flex_handle *handle,
+                           struct rte_flow_error *error)
+{
+       static const char err_msg[] = "flex item release unsupported";
+       struct rte_flow_attr attr = { .transfer = 0 };
+       const struct mlx5_flow_driver_ops *fops =
+                       flow_get_drv_ops(flow_get_drv_type(dev, &attr));
+
+       if (!fops->item_release) {
+               DRV_LOG(ERR, "port %u %s.", dev->data->port_id, err_msg);
+               rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
+                                  NULL, err_msg);
+               return -rte_errno;
+       }
+       return fops->item_release(dev, handle, error);
+}
+
 static void
 mlx5_dbg__print_pattern(const struct rte_flow_item *item)
 {