net/mlx5: fix unnecessary init in mark conversion
[dpdk.git] / drivers / net / mlx5 / mlx5_flow_dv.c
index 8b5b683..5339980 100644 (file)
@@ -8,16 +8,6 @@
 #include <string.h>
 #include <unistd.h>
 
-/* Verbs header. */
-/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
-#ifdef PEDANTIC
-#pragma GCC diagnostic ignored "-Wpedantic"
-#endif
-#include <infiniband/verbs.h>
-#ifdef PEDANTIC
-#pragma GCC diagnostic error "-Wpedantic"
-#endif
-
 #include <rte_common.h>
 #include <rte_ether.h>
 #include <rte_ethdev_driver.h>
 #include <rte_gre.h>
 #include <rte_vxlan.h>
 #include <rte_gtp.h>
+#include <rte_eal_paging.h>
+#include <rte_mpls.h>
 
+#include <mlx5_glue.h>
 #include <mlx5_devx_cmds.h>
 #include <mlx5_prm.h>
+#include <mlx5_malloc.h>
 
 #include "mlx5_defs.h"
 #include "mlx5.h"
@@ -1158,8 +1152,7 @@ flow_dv_convert_action_mark(struct rte_eth_dev *dev,
                .mask = &mask,
        };
        struct field_modify_info reg_c_x[] = {
-               {4, 0, 0}, /* dynamic instead of MLX5_MODI_META_REG_C_1. */
-               {0, 0, 0},
+               [1] = {0, 0, 0},
        };
        int reg;
 
@@ -1179,7 +1172,7 @@ flow_dv_convert_action_mark(struct rte_eth_dev *dev,
                mask = rte_cpu_to_be_32(mask) & msk_c0;
                mask = rte_cpu_to_be_32(mask << shl_c0);
        }
-       reg_c_x[0].id = reg_to_field[reg];
+       reg_c_x[0] = (struct field_modify_info){4, 0, reg_to_field[reg]};
        return flow_dv_convert_modify_action(&item, reg_c_x, NULL, resource,
                                             MLX5_MODIFICATION_TYPE_SET, error);
 }
@@ -1839,7 +1832,17 @@ flow_dv_validate_action_pop_vlan(struct rte_eth_dev *dev,
                                          RTE_FLOW_ERROR_TYPE_ACTION, action,
                                          "no support for multiple VLAN "
                                          "actions");
-       if (!(item_flags & MLX5_FLOW_LAYER_OUTER_VLAN))
+       /* Pop VLAN with preceding Decap requires inner header with VLAN. */
+       if ((action_flags & MLX5_FLOW_ACTION_DECAP) &&
+           !(item_flags & MLX5_FLOW_LAYER_INNER_VLAN))
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                         NULL,
+                                         "cannot pop vlan after decap without "
+                                         "match on inner vlan in the flow");
+       /* Pop VLAN without preceding Decap requires outer header with VLAN. */
+       if (!(action_flags & MLX5_FLOW_ACTION_DECAP) &&
+           !(item_flags & MLX5_FLOW_LAYER_OUTER_VLAN))
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL,
@@ -1948,22 +1951,11 @@ flow_dv_validate_action_push_vlan(struct rte_eth_dev *dev,
        const struct rte_flow_action_of_push_vlan *push_vlan = action->conf;
        const struct mlx5_priv *priv = dev->data->dev_private;
 
-       if (!attr->transfer && attr->ingress)
-               return rte_flow_error_set(error, ENOTSUP,
-                                         RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
-                                         NULL,
-                                         "push VLAN action not supported for "
-                                         "ingress");
        if (push_vlan->ethertype != RTE_BE16(RTE_ETHER_TYPE_VLAN) &&
            push_vlan->ethertype != RTE_BE16(RTE_ETHER_TYPE_QINQ))
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ACTION, action,
                                          "invalid vlan ethertype");
-       if (action_flags & MLX5_FLOW_VLAN_ACTIONS)
-               return rte_flow_error_set(error, ENOTSUP,
-                                         RTE_FLOW_ERROR_TYPE_ACTION, action,
-                                         "no support for multiple VLAN "
-                                         "actions");
        if (action_flags & MLX5_FLOW_ACTION_PORT_ID)
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ACTION, action,
@@ -2428,6 +2420,11 @@ flow_dv_validate_action_decap(struct rte_eth_dev *dev,
 {
        const struct mlx5_priv *priv = dev->data->dev_private;
 
+       if (priv->config.hca_attr.scatter_fcs_w_decap_disable &&
+           !priv->config.decap_en)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION, NULL,
+                                         "decap is not enabled");
        if (action_flags & MLX5_FLOW_XCAP_ACTIONS)
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_ACTION, NULL,
@@ -2615,7 +2612,7 @@ flow_dv_encap_decap_resource_register
                                        (sh->ctx, domain, cache_resource,
                                         &cache_resource->action);
        if (ret) {
-               rte_free(cache_resource);
+               mlx5_free(cache_resource);
                return rte_flow_error_set(error, ENOMEM,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot create action");
@@ -2772,7 +2769,7 @@ flow_dv_port_id_action_resource_register
                                (priv->sh->fdb_domain, resource->port_id,
                                 &cache_resource->action);
        if (ret) {
-               rte_free(cache_resource);
+               mlx5_free(cache_resource);
                return rte_flow_error_set(error, ENOMEM,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot create action");
@@ -2851,7 +2848,7 @@ flow_dv_push_vlan_action_resource_register
                                        (domain, resource->vlan_tag,
                                         &cache_resource->action);
        if (ret) {
-               rte_free(cache_resource);
+               mlx5_free(cache_resource);
                return rte_flow_error_set(error, ENOMEM,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot create action");
@@ -2869,7 +2866,7 @@ flow_dv_push_vlan_action_resource_register
        return 0;
 }
 /**
- * Get the size of specific rte_flow_item_type
+ * Get the size of specific rte_flow_item_type hdr size
  *
  * @param[in] item_type
  *   Tested rte_flow_item_type.
@@ -2878,43 +2875,39 @@ flow_dv_push_vlan_action_resource_register
  *   sizeof struct item_type, 0 if void or irrelevant.
  */
 static size_t
-flow_dv_get_item_len(const enum rte_flow_item_type item_type)
+flow_dv_get_item_hdr_len(const enum rte_flow_item_type item_type)
 {
        size_t retval;
 
        switch (item_type) {
        case RTE_FLOW_ITEM_TYPE_ETH:
-               retval = sizeof(struct rte_flow_item_eth);
+               retval = sizeof(struct rte_ether_hdr);
                break;
        case RTE_FLOW_ITEM_TYPE_VLAN:
-               retval = sizeof(struct rte_flow_item_vlan);
+               retval = sizeof(struct rte_vlan_hdr);
                break;
        case RTE_FLOW_ITEM_TYPE_IPV4:
-               retval = sizeof(struct rte_flow_item_ipv4);
+               retval = sizeof(struct rte_ipv4_hdr);
                break;
        case RTE_FLOW_ITEM_TYPE_IPV6:
-               retval = sizeof(struct rte_flow_item_ipv6);
+               retval = sizeof(struct rte_ipv6_hdr);
                break;
        case RTE_FLOW_ITEM_TYPE_UDP:
-               retval = sizeof(struct rte_flow_item_udp);
+               retval = sizeof(struct rte_udp_hdr);
                break;
        case RTE_FLOW_ITEM_TYPE_TCP:
-               retval = sizeof(struct rte_flow_item_tcp);
+               retval = sizeof(struct rte_tcp_hdr);
                break;
        case RTE_FLOW_ITEM_TYPE_VXLAN:
-               retval = sizeof(struct rte_flow_item_vxlan);
+       case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
+               retval = sizeof(struct rte_vxlan_hdr);
                break;
        case RTE_FLOW_ITEM_TYPE_GRE:
-               retval = sizeof(struct rte_flow_item_gre);
-               break;
        case RTE_FLOW_ITEM_TYPE_NVGRE:
-               retval = sizeof(struct rte_flow_item_nvgre);
-               break;
-       case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
-               retval = sizeof(struct rte_flow_item_vxlan_gpe);
+               retval = sizeof(struct rte_gre_hdr);
                break;
        case RTE_FLOW_ITEM_TYPE_MPLS:
-               retval = sizeof(struct rte_flow_item_mpls);
+               retval = sizeof(struct rte_mpls_hdr);
                break;
        case RTE_FLOW_ITEM_TYPE_VOID: /* Fall through. */
        default:
@@ -2967,7 +2960,7 @@ flow_dv_convert_encap_data(const struct rte_flow_item *items, uint8_t *buf,
                                          RTE_FLOW_ERROR_TYPE_ACTION,
                                          NULL, "invalid empty data");
        for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
-               len = flow_dv_get_item_len(items->type);
+               len = flow_dv_get_item_hdr_len(items->type);
                if (len + temp_size > MLX5_ENCAP_MAX_LEN)
                        return rte_flow_error_set(error, EINVAL,
                                                  RTE_FLOW_ERROR_TYPE_ACTION,
@@ -4024,8 +4017,9 @@ flow_dv_modify_hdr_resource_register
                }
        }
        /* Register new modify-header resource. */
-       cache_resource = rte_calloc(__func__, 1,
-                                   sizeof(*cache_resource) + actions_len, 0);
+       cache_resource = mlx5_malloc(MLX5_MEM_ZERO,
+                                   sizeof(*cache_resource) + actions_len, 0,
+                                   SOCKET_ID_ANY);
        if (!cache_resource)
                return rte_flow_error_set(error, ENOMEM,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
@@ -4036,7 +4030,7 @@ flow_dv_modify_hdr_resource_register
                                        (sh->ctx, ns, cache_resource,
                                         actions_len, &cache_resource->action);
        if (ret) {
-               rte_free(cache_resource);
+               mlx5_free(cache_resource);
                return rte_flow_error_set(error, ENOMEM,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot create action");
@@ -4175,7 +4169,14 @@ flow_dv_create_counter_stat_mem_mng(struct rte_eth_dev *dev, int raws_n)
                        MLX5_COUNTERS_PER_POOL +
                        sizeof(struct mlx5_counter_stats_raw)) * raws_n +
                        sizeof(struct mlx5_counter_stats_mem_mng);
-       uint8_t *mem = rte_calloc(__func__, 1, size, sysconf(_SC_PAGESIZE));
+       size_t pgsize = rte_mem_page_size();
+       if (pgsize == (size_t)-1) {
+               DRV_LOG(ERR, "Failed to get mem page size");
+               rte_errno = ENOMEM;
+               return NULL;
+       }
+       uint8_t *mem = mlx5_malloc(MLX5_MEM_ZERO, size, pgsize,
+                                 SOCKET_ID_ANY);
        int i;
 
        if (!mem) {
@@ -4188,7 +4189,7 @@ flow_dv_create_counter_stat_mem_mng(struct rte_eth_dev *dev, int raws_n)
                                                 IBV_ACCESS_LOCAL_WRITE);
        if (!mem_mng->umem) {
                rte_errno = errno;
-               rte_free(mem);
+               mlx5_free(mem);
                return NULL;
        }
        mkey_attr.addr = (uintptr_t)mem;
@@ -4207,7 +4208,7 @@ flow_dv_create_counter_stat_mem_mng(struct rte_eth_dev *dev, int raws_n)
        if (!mem_mng->dm) {
                mlx5_glue->devx_umem_dereg(mem_mng->umem);
                rte_errno = errno;
-               rte_free(mem);
+               mlx5_free(mem);
                return NULL;
        }
        mem_mng->raws = (struct mlx5_counter_stats_raw *)(mem + size);
@@ -4244,7 +4245,7 @@ flow_dv_container_resize(struct rte_eth_dev *dev,
        void *old_pools = cont->pools;
        uint32_t resize = cont->n + MLX5_CNT_CONTAINER_RESIZE;
        uint32_t mem_size = sizeof(struct mlx5_flow_counter_pool *) * resize;
-       void *pools = rte_calloc(__func__, 1, mem_size, 0);
+       void *pools = mlx5_malloc(MLX5_MEM_ZERO, mem_size, 0, SOCKET_ID_ANY);
 
        if (!pools) {
                rte_errno = ENOMEM;
@@ -4263,7 +4264,7 @@ flow_dv_container_resize(struct rte_eth_dev *dev,
                mem_mng = flow_dv_create_counter_stat_mem_mng(dev,
                          MLX5_CNT_CONTAINER_RESIZE + MLX5_MAX_PENDING_QUERIES);
                if (!mem_mng) {
-                       rte_free(pools);
+                       mlx5_free(pools);
                        return -ENOMEM;
                }
                for (i = 0; i < MLX5_MAX_PENDING_QUERIES; ++i)
@@ -4278,7 +4279,7 @@ flow_dv_container_resize(struct rte_eth_dev *dev,
        cont->pools = pools;
        rte_spinlock_unlock(&cont->resize_sl);
        if (old_pools)
-               rte_free(old_pools);
+               mlx5_free(old_pools);
        return 0;
 }
 
@@ -4367,7 +4368,7 @@ flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
        size += MLX5_COUNTERS_PER_POOL * CNT_SIZE;
        size += (batch ? 0 : MLX5_COUNTERS_PER_POOL * CNTEXT_SIZE);
        size += (!age ? 0 : MLX5_COUNTERS_PER_POOL * AGE_SIZE);
-       pool = rte_calloc(__func__, 1, size, 0);
+       pool = mlx5_malloc(MLX5_MEM_ZERO, size, 0, SOCKET_ID_ANY);
        if (!pool) {
                rte_errno = ENOMEM;
                return NULL;
@@ -4403,38 +4404,65 @@ flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
 }
 
 /**
- * Update the minimum dcs-id for aged or no-aged counter pool.
+ * Restore skipped counters in the pool.
+ *
+ * As counter pool query requires the first counter dcs
+ * ID start with 4 alinged, if the pool counters with
+ * min_dcs ID are not aligned with 4, the counters will
+ * be skipped.
+ * Once other min_dcs ID less than these skipped counter
+ * dcs ID appears, the skipped counters will be safe to
+ * use.
+ * Should be called when min_dcs is updated.
  *
- * @param[in] dev
- *   Pointer to the Ethernet device structure.
  * @param[in] pool
  *   Current counter pool.
- * @param[in] batch
- *   Whether the pool is for counter that was allocated by batch command.
- * @param[in] age
- *   Whether the counter is for aging.
+ * @param[in] last_min_dcs
+ *   Last min_dcs.
  */
 static void
-flow_dv_counter_update_min_dcs(struct rte_eth_dev *dev,
-                       struct mlx5_flow_counter_pool *pool,
-                       uint32_t batch, uint32_t age)
+flow_dv_counter_restore(struct mlx5_flow_counter_pool *pool,
+                       struct mlx5_devx_obj *last_min_dcs)
 {
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_flow_counter_pool *other;
-       struct mlx5_pools_container *cont;
+       struct mlx5_flow_counter_ext *cnt_ext;
+       uint32_t offset, new_offset;
+       uint32_t skip_cnt = 0;
+       uint32_t i;
 
-       cont = MLX5_CNT_CONTAINER(priv->sh, batch, (age ^ 0x1));
-       other = flow_dv_find_pool_by_id(cont, pool->min_dcs->id);
-       if (!other)
+       if (!pool->skip_cnt)
                return;
-       if (pool->min_dcs->id < other->min_dcs->id) {
-               rte_atomic64_set(&other->a64_dcs,
-                       rte_atomic64_read(&pool->a64_dcs));
-       } else {
-               rte_atomic64_set(&pool->a64_dcs,
-                       rte_atomic64_read(&other->a64_dcs));
+       /*
+        * If last min_dcs is not valid. The skipped counter may even after
+        * last min_dcs, set the offset to the whole pool.
+        */
+       if (last_min_dcs->id & (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1))
+               offset = MLX5_COUNTERS_PER_POOL;
+       else
+               offset = last_min_dcs->id % MLX5_COUNTERS_PER_POOL;
+       new_offset = pool->min_dcs->id % MLX5_COUNTERS_PER_POOL;
+       /*
+        * Check the counters from 1 to the last_min_dcs range. Counters
+        * before new min_dcs indicates pool still has skipped counters.
+        * Counters be skipped after new min_dcs will be ready to use.
+        * Offset 0 counter must be empty or min_dcs, start from 1.
+        */
+       for (i = 1; i < offset; i++) {
+               cnt_ext = MLX5_GET_POOL_CNT_EXT(pool, i);
+               if (cnt_ext->skipped) {
+                       if (i > new_offset) {
+                               cnt_ext->skipped = 0;
+                               TAILQ_INSERT_TAIL
+                                       (&pool->counters[pool->query_gen],
+                                        MLX5_POOL_GET_CNT(pool, i), next);
+                       } else {
+                               skip_cnt++;
+                       }
+               }
        }
+       if (!skip_cnt)
+               pool->skip_cnt = 0;
 }
+
 /**
  * Prepare a new counter and/or a new counter pool.
  *
@@ -4460,33 +4488,84 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
        struct mlx5_pools_container *cont;
        struct mlx5_flow_counter_pool *pool;
        struct mlx5_counters tmp_tq;
+       struct mlx5_devx_obj *last_min_dcs;
        struct mlx5_devx_obj *dcs = NULL;
        struct mlx5_flow_counter *cnt;
+       uint32_t add2other;
        uint32_t i;
 
        cont = MLX5_CNT_CONTAINER(priv->sh, batch, age);
        if (!batch) {
+retry:
+               add2other = 0;
                /* bulk_bitmap must be 0 for single counter allocation. */
                dcs = mlx5_devx_cmd_flow_counter_alloc(priv->sh->ctx, 0);
                if (!dcs)
                        return NULL;
                pool = flow_dv_find_pool_by_id(cont, dcs->id);
+               /* Check if counter belongs to exist pool ID range. */
                if (!pool) {
-                       pool = flow_dv_pool_create(dev, dcs, batch, age);
-                       if (!pool) {
-                               mlx5_devx_cmd_destroy(dcs);
-                               return NULL;
+                       pool = flow_dv_find_pool_by_id
+                              (MLX5_CNT_CONTAINER
+                              (priv->sh, batch, (age ^ 0x1)), dcs->id);
+                       /*
+                        * Pool eixsts, counter will be added to the other
+                        * container, need to reallocate it later.
+                        */
+                       if (pool) {
+                               add2other = 1;
+                       } else {
+                               pool = flow_dv_pool_create(dev, dcs, batch,
+                                                          age);
+                               if (!pool) {
+                                       mlx5_devx_cmd_destroy(dcs);
+                                       return NULL;
+                               }
                        }
-               } else if (dcs->id < pool->min_dcs->id) {
+               }
+               if ((dcs->id < pool->min_dcs->id ||
+                   pool->min_dcs->id &
+                   (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1)) &&
+                   !(dcs->id & (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1))) {
+                       /*
+                        * Update the pool min_dcs only if current dcs is
+                        * valid and exist min_dcs is not valid or greater
+                        * than new dcs.
+                        */
+                       last_min_dcs = pool->min_dcs;
                        rte_atomic64_set(&pool->a64_dcs,
                                         (int64_t)(uintptr_t)dcs);
+                       /*
+                        * Restore any skipped counters if the new min_dcs
+                        * ID is smaller or min_dcs is not valid.
+                        */
+                       if (dcs->id < last_min_dcs->id ||
+                           last_min_dcs->id &
+                           (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1))
+                               flow_dv_counter_restore(pool, last_min_dcs);
                }
-               flow_dv_counter_update_min_dcs(dev,
-                                               pool, batch, age);
                i = dcs->id % MLX5_COUNTERS_PER_POOL;
                cnt = MLX5_POOL_GET_CNT(pool, i);
                cnt->pool = pool;
                MLX5_GET_POOL_CNT_EXT(pool, i)->dcs = dcs;
+               /*
+                * If min_dcs is not valid, it means the new allocated dcs
+                * also fail to become the valid min_dcs, just skip it.
+                * Or if min_dcs is valid, and new dcs ID is smaller than
+                * min_dcs, but not become the min_dcs, also skip it.
+                */
+               if (pool->min_dcs->id &
+                   (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1) ||
+                   dcs->id < pool->min_dcs->id) {
+                       MLX5_GET_POOL_CNT_EXT(pool, i)->skipped = 1;
+                       pool->skip_cnt = 1;
+                       goto retry;
+               }
+               if (add2other) {
+                       TAILQ_INSERT_TAIL(&pool->counters[pool->query_gen],
+                                         cnt, next);
+                       goto retry;
+               }
                *cnt_free = cnt;
                return pool;
        }
@@ -4923,6 +5002,17 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                        .hop_limits = 0xff,
                },
        };
+       const struct rte_flow_item_ecpri nic_ecpri_mask = {
+               .hdr = {
+                       .common = {
+                               .u32 =
+                               RTE_BE32(((const struct rte_ecpri_common_hdr) {
+                                       .type = 0xFF,
+                                       }).u32),
+                       },
+                       .dummy[0] = 0xffffffff,
+               },
+       };
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_dev_config *dev_conf = &priv->config;
        uint16_t queue_index = 0xFFFF;
@@ -5173,6 +5263,17 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                return ret;
                        last_item = MLX5_FLOW_LAYER_GTP;
                        break;
+               case RTE_FLOW_ITEM_TYPE_ECPRI:
+                       /* Capacity will be checked in the translate stage. */
+                       ret = mlx5_flow_validate_item_ecpri(items, item_flags,
+                                                           last_item,
+                                                           ether_type,
+                                                           &nic_ecpri_mask,
+                                                           error);
+                       if (ret < 0)
+                               return ret;
+                       last_item = MLX5_FLOW_LAYER_ECPRI;
+                       break;
                default:
                        return rte_flow_error_set(error, ENOTSUP,
                                                  RTE_FLOW_ERROR_TYPE_ITEM,
@@ -5653,21 +5754,38 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                                  actions,
                                                  "no fate action is found");
        }
-       /* Continue validation for Xcap actions.*/
-       if ((action_flags & MLX5_FLOW_XCAP_ACTIONS) && (queue_index == 0xFFFF ||
-           mlx5_rxq_get_type(dev, queue_index) != MLX5_RXQ_TYPE_HAIRPIN)) {
+       /* Continue validation for Xcap and VLAN actions.*/
+       if ((action_flags & (MLX5_FLOW_XCAP_ACTIONS |
+                            MLX5_FLOW_VLAN_ACTIONS)) &&
+           (queue_index == 0xFFFF ||
+            mlx5_rxq_get_type(dev, queue_index) != MLX5_RXQ_TYPE_HAIRPIN)) {
                if ((action_flags & MLX5_FLOW_XCAP_ACTIONS) ==
                    MLX5_FLOW_XCAP_ACTIONS)
                        return rte_flow_error_set(error, ENOTSUP,
                                                  RTE_FLOW_ERROR_TYPE_ACTION,
                                                  NULL, "encap and decap "
                                                  "combination aren't supported");
-               if (!attr->transfer && attr->ingress && (action_flags &
-                                                       MLX5_FLOW_ACTION_ENCAP))
-                       return rte_flow_error_set(error, ENOTSUP,
-                                                 RTE_FLOW_ERROR_TYPE_ACTION,
-                                                 NULL, "encap is not supported"
-                                                 " for ingress traffic");
+               if (!attr->transfer && attr->ingress) {
+                       if (action_flags & MLX5_FLOW_ACTION_ENCAP)
+                               return rte_flow_error_set
+                                               (error, ENOTSUP,
+                                                RTE_FLOW_ERROR_TYPE_ACTION,
+                                                NULL, "encap is not supported"
+                                                " for ingress traffic");
+                       else if (action_flags & MLX5_FLOW_ACTION_OF_PUSH_VLAN)
+                               return rte_flow_error_set
+                                               (error, ENOTSUP,
+                                                RTE_FLOW_ERROR_TYPE_ACTION,
+                                                NULL, "push VLAN action not "
+                                                "supported for ingress");
+                       else if ((action_flags & MLX5_FLOW_VLAN_ACTIONS) ==
+                                       MLX5_FLOW_VLAN_ACTIONS)
+                               return rte_flow_error_set
+                                               (error, ENOTSUP,
+                                                RTE_FLOW_ERROR_TYPE_ACTION,
+                                                NULL, "no support for "
+                                                "multiple VLAN actions");
+               }
        }
        /* Hairpin flow will add one more TAG action. */
        if (hairpin > 0)
@@ -5737,7 +5855,15 @@ flow_dv_prepare(struct rte_eth_dev *dev,
        dev_flow = &((struct mlx5_flow *)priv->inter_flows)[priv->flow_idx++];
        dev_flow->handle = dev_handle;
        dev_flow->handle_idx = handle_idx;
-       dev_flow->dv.value.size = MLX5_ST_SZ_BYTES(fte_match_param);
+       /*
+        * In some old rdma-core releases, before continuing, a check of the
+        * length of matching parameter will be done at first. It needs to use
+        * the length without misc4 param. If the flow has misc4 support, then
+        * the length needs to be adjusted accordingly. Each param member is
+        * aligned with a 64B boundary naturally.
+        */
+       dev_flow->dv.value.size = MLX5_ST_SZ_BYTES(fte_match_param) -
+                                 MLX5_ST_SZ_BYTES(fte_match_set_misc4);
        /*
         * The matching value needs to be cleared to 0 before using. In the
         * past, it will be automatically cleared when using rte_*alloc
@@ -5882,6 +6008,7 @@ flow_dv_translate_item_eth(void *matcher, void *key,
         * Set match on ethertype only if ETH header is not followed by VLAN.
         * HW is optimized for IPv4/IPv6. In such cases, avoid setting
         * ethertype, and use ip_version field instead.
+        * eCPRI over Ether layer will use type value 0xAEFE.
         */
        if (eth_v->type == RTE_BE16(RTE_ETHER_TYPE_IPV4) &&
            eth_m->type == 0xFFFF) {
@@ -6416,8 +6543,8 @@ flow_dv_translate_item_nvgre(void *matcher, void *key,
        const struct rte_flow_item_nvgre *nvgre_v = item->spec;
        void *misc_m = MLX5_ADDR_OF(fte_match_param, matcher, misc_parameters);
        void *misc_v = MLX5_ADDR_OF(fte_match_param, key, misc_parameters);
-       const char *tni_flow_id_m = (const char *)nvgre_m->tni;
-       const char *tni_flow_id_v = (const char *)nvgre_v->tni;
+       const char *tni_flow_id_m;
+       const char *tni_flow_id_v;
        char *gre_key_m;
        char *gre_key_v;
        int size;
@@ -6442,6 +6569,8 @@ flow_dv_translate_item_nvgre(void *matcher, void *key,
                return;
        if (!nvgre_m)
                nvgre_m = &rte_flow_item_nvgre_mask;
+       tni_flow_id_m = (const char *)nvgre_m->tni;
+       tni_flow_id_v = (const char *)nvgre_v->tni;
        size = sizeof(nvgre_m->tni) + sizeof(nvgre_m->flow_id);
        gre_key_m = MLX5_ADDR_OF(fte_match_set_misc, misc_m, gre_key_h);
        gre_key_v = MLX5_ADDR_OF(fte_match_set_misc, misc_v, gre_key_h);
@@ -7236,6 +7365,90 @@ flow_dv_translate_item_gtp(void *matcher, void *key,
                 rte_be_to_cpu_32(gtp_v->teid & gtp_m->teid));
 }
 
+/**
+ * Add eCPRI item to matcher and to the value.
+ *
+ * @param[in] dev
+ *   The devich to configure through.
+ * @param[in, out] matcher
+ *   Flow matcher.
+ * @param[in, out] key
+ *   Flow matcher value.
+ * @param[in] item
+ *   Flow pattern to translate.
+ * @param[in] samples
+ *   Sample IDs to be used in the matching.
+ */
+static void
+flow_dv_translate_item_ecpri(struct rte_eth_dev *dev, void *matcher,
+                            void *key, const struct rte_flow_item *item)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       const struct rte_flow_item_ecpri *ecpri_m = item->mask;
+       const struct rte_flow_item_ecpri *ecpri_v = item->spec;
+       void *misc4_m = MLX5_ADDR_OF(fte_match_param, matcher,
+                                    misc_parameters_4);
+       void *misc4_v = MLX5_ADDR_OF(fte_match_param, key, misc_parameters_4);
+       uint32_t *samples;
+       void *dw_m;
+       void *dw_v;
+
+       if (!ecpri_v)
+               return;
+       if (!ecpri_m)
+               ecpri_m = &rte_flow_item_ecpri_mask;
+       /*
+        * Maximal four DW samples are supported in a single matching now.
+        * Two are used now for a eCPRI matching:
+        * 1. Type: one byte, mask should be 0x00ff0000 in network order
+        * 2. ID of a message: one or two bytes, mask 0xffff0000 or 0xff000000
+        *    if any.
+        */
+       if (!ecpri_m->hdr.common.u32)
+               return;
+       samples = priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0].ids;
+       /* Need to take the whole DW as the mask to fill the entry. */
+       dw_m = MLX5_ADDR_OF(fte_match_set_misc4, misc4_m,
+                           prog_sample_field_value_0);
+       dw_v = MLX5_ADDR_OF(fte_match_set_misc4, misc4_v,
+                           prog_sample_field_value_0);
+       /* Already big endian (network order) in the header. */
+       *(uint32_t *)dw_m = ecpri_m->hdr.common.u32;
+       *(uint32_t *)dw_v = ecpri_v->hdr.common.u32;
+       /* Sample#0, used for matching type, offset 0. */
+       MLX5_SET(fte_match_set_misc4, misc4_m,
+                prog_sample_field_id_0, samples[0]);
+       /* It makes no sense to set the sample ID in the mask field. */
+       MLX5_SET(fte_match_set_misc4, misc4_v,
+                prog_sample_field_id_0, samples[0]);
+       /*
+        * Checking if message body part needs to be matched.
+        * Some wildcard rules only matching type field should be supported.
+        */
+       if (ecpri_m->hdr.dummy[0]) {
+               switch (ecpri_v->hdr.common.type) {
+               case RTE_ECPRI_MSG_TYPE_IQ_DATA:
+               case RTE_ECPRI_MSG_TYPE_RTC_CTRL:
+               case RTE_ECPRI_MSG_TYPE_DLY_MSR:
+                       dw_m = MLX5_ADDR_OF(fte_match_set_misc4, misc4_m,
+                                           prog_sample_field_value_1);
+                       dw_v = MLX5_ADDR_OF(fte_match_set_misc4, misc4_v,
+                                           prog_sample_field_value_1);
+                       *(uint32_t *)dw_m = ecpri_m->hdr.dummy[0];
+                       *(uint32_t *)dw_v = ecpri_v->hdr.dummy[0];
+                       /* Sample#1, to match message body, offset 4. */
+                       MLX5_SET(fte_match_set_misc4, misc4_m,
+                                prog_sample_field_id_1, samples[1]);
+                       MLX5_SET(fte_match_set_misc4, misc4_v,
+                                prog_sample_field_id_1, samples[1]);
+                       break;
+               default:
+                       /* Others, do not match any sample ID. */
+                       break;
+               }
+       }
+}
+
 static uint32_t matcher_zero[MLX5_ST_SZ_DW(fte_match_param)] = { 0 };
 
 #define HEADER_IS_ZERO(match_criteria, headers)                                     \
@@ -7271,6 +7484,9 @@ flow_dv_matcher_enable(uint32_t *match_criteria)
        match_criteria_enable |=
                (!HEADER_IS_ZERO(match_criteria, misc_parameters_3)) <<
                MLX5_MATCH_CRITERIA_ENABLE_MISC3_BIT;
+       match_criteria_enable |=
+               (!HEADER_IS_ZERO(match_criteria, misc_parameters_4)) <<
+               MLX5_MATCH_CRITERIA_ENABLE_MISC4_BIT;
        return match_criteria_enable;
 }
 
@@ -7467,7 +7683,8 @@ flow_dv_matcher_register(struct rte_eth_dev *dev,
                }
        }
        /* Register new matcher. */
-       cache_matcher = rte_calloc(__func__, 1, sizeof(*cache_matcher), 0);
+       cache_matcher = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*cache_matcher), 0,
+                                   SOCKET_ID_ANY);
        if (!cache_matcher) {
                flow_dv_tbl_resource_release(dev, tbl);
                return rte_flow_error_set(error, ENOMEM,
@@ -7483,7 +7700,7 @@ flow_dv_matcher_register(struct rte_eth_dev *dev,
        ret = mlx5_flow_os_create_flow_matcher(sh->ctx, &dv_attr, tbl->obj,
                                               &cache_matcher->matcher_object);
        if (ret) {
-               rte_free(cache_matcher);
+               mlx5_free(cache_matcher);
 #ifdef HAVE_MLX5DV_DR
                flow_dv_tbl_resource_release(dev, tbl);
 #endif
@@ -7558,7 +7775,7 @@ flow_dv_tag_resource_register
        ret = mlx5_flow_os_create_flow_action_tag(tag_be24,
                                                  &cache_resource->action);
        if (ret) {
-               rte_free(cache_resource);
+               mlx5_free(cache_resource);
                return rte_flow_error_set(error, ENOMEM,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot create action");
@@ -7567,7 +7784,7 @@ flow_dv_tag_resource_register
        rte_atomic32_inc(&cache_resource->refcnt);
        if (mlx5_hlist_insert(sh->tag_table, &cache_resource->entry)) {
                mlx5_flow_os_destroy_flow_action(cache_resource->action);
-               rte_free(cache_resource);
+               mlx5_free(cache_resource);
                return rte_flow_error_set(error, EEXIST,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot insert tag");
@@ -7869,7 +8086,8 @@ __flow_dv_translate(struct rte_eth_dev *dev,
        uint64_t priority = attr->priority;
        struct mlx5_flow_dv_matcher matcher = {
                .mask = {
-                       .size = sizeof(matcher.mask.buf),
+                       .size = sizeof(matcher.mask.buf) -
+                               MLX5_ST_SZ_BYTES(fte_match_set_misc4),
                },
        };
        int actions_n = 0;
@@ -8462,8 +8680,7 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                case RTE_FLOW_ITEM_TYPE_GRE:
                        flow_dv_translate_item_gre(match_mask, match_value,
                                                   items, tunnel);
-                       matcher.priority = rss_desc->level >= 2 ?
-                                   MLX5_PRIORITY_MAP_L2 : MLX5_PRIORITY_MAP_L4;
+                       matcher.priority = MLX5_TUNNEL_PRIO_GET(rss_desc);
                        last_item = MLX5_FLOW_LAYER_GRE;
                        break;
                case RTE_FLOW_ITEM_TYPE_GRE_KEY:
@@ -8474,37 +8691,32 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                case RTE_FLOW_ITEM_TYPE_NVGRE:
                        flow_dv_translate_item_nvgre(match_mask, match_value,
                                                     items, tunnel);
-                       matcher.priority = rss_desc->level >= 2 ?
-                                   MLX5_PRIORITY_MAP_L2 : MLX5_PRIORITY_MAP_L4;
+                       matcher.priority = MLX5_TUNNEL_PRIO_GET(rss_desc);
                        last_item = MLX5_FLOW_LAYER_GRE;
                        break;
                case RTE_FLOW_ITEM_TYPE_VXLAN:
                        flow_dv_translate_item_vxlan(match_mask, match_value,
                                                     items, tunnel);
-                       matcher.priority = rss_desc->level >= 2 ?
-                                   MLX5_PRIORITY_MAP_L2 : MLX5_PRIORITY_MAP_L4;
+                       matcher.priority = MLX5_TUNNEL_PRIO_GET(rss_desc);
                        last_item = MLX5_FLOW_LAYER_VXLAN;
                        break;
                case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
                        flow_dv_translate_item_vxlan_gpe(match_mask,
                                                         match_value, items,
                                                         tunnel);
-                       matcher.priority = rss_desc->level >= 2 ?
-                                   MLX5_PRIORITY_MAP_L2 : MLX5_PRIORITY_MAP_L4;
+                       matcher.priority = MLX5_TUNNEL_PRIO_GET(rss_desc);
                        last_item = MLX5_FLOW_LAYER_VXLAN_GPE;
                        break;
                case RTE_FLOW_ITEM_TYPE_GENEVE:
                        flow_dv_translate_item_geneve(match_mask, match_value,
                                                      items, tunnel);
-                       matcher.priority = rss_desc->level >= 2 ?
-                                   MLX5_PRIORITY_MAP_L2 : MLX5_PRIORITY_MAP_L4;
+                       matcher.priority = MLX5_TUNNEL_PRIO_GET(rss_desc);
                        last_item = MLX5_FLOW_LAYER_GENEVE;
                        break;
                case RTE_FLOW_ITEM_TYPE_MPLS:
                        flow_dv_translate_item_mpls(match_mask, match_value,
                                                    items, last_item, tunnel);
-                       matcher.priority = rss_desc->level >= 2 ?
-                                   MLX5_PRIORITY_MAP_L2 : MLX5_PRIORITY_MAP_L4;
+                       matcher.priority = MLX5_TUNNEL_PRIO_GET(rss_desc);
                        last_item = MLX5_FLOW_LAYER_MPLS;
                        break;
                case RTE_FLOW_ITEM_TYPE_MARK:
@@ -8546,10 +8758,29 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                case RTE_FLOW_ITEM_TYPE_GTP:
                        flow_dv_translate_item_gtp(match_mask, match_value,
                                                   items, tunnel);
-                       matcher.priority = rss_desc->level >= 2 ?
-                                   MLX5_PRIORITY_MAP_L2 : MLX5_PRIORITY_MAP_L4;
+                       matcher.priority = MLX5_TUNNEL_PRIO_GET(rss_desc);
                        last_item = MLX5_FLOW_LAYER_GTP;
                        break;
+               case RTE_FLOW_ITEM_TYPE_ECPRI:
+                       if (!mlx5_flex_parser_ecpri_exist(dev)) {
+                               /* Create it only the first time to be used. */
+                               ret = mlx5_flex_parser_ecpri_alloc(dev);
+                               if (ret)
+                                       return rte_flow_error_set
+                                               (error, -ret,
+                                               RTE_FLOW_ERROR_TYPE_ITEM,
+                                               NULL,
+                                               "cannot create eCPRI parser");
+                       }
+                       /* Adjust the length matcher and device flow value. */
+                       matcher.mask.size = MLX5_ST_SZ_BYTES(fte_match_param);
+                       dev_flow->dv.value.size =
+                                       MLX5_ST_SZ_BYTES(fte_match_param);
+                       flow_dv_translate_item_ecpri(dev, match_mask,
+                                                    match_value, items);
+                       /* No other protocol should follow eCPRI layer. */
+                       last_item = MLX5_FLOW_LAYER_ECPRI;
+                       break;
                default:
                        break;
                }
@@ -8769,7 +9000,7 @@ flow_dv_matcher_release(struct rte_eth_dev *dev,
                LIST_REMOVE(matcher, next);
                /* table ref-- in release interface. */
                flow_dv_tbl_resource_release(dev, matcher->tbl);
-               rte_free(matcher);
+               mlx5_free(matcher);
                DRV_LOG(DEBUG, "port %u matcher %p: removed",
                        dev->data->port_id, (void *)matcher);
                return 0;
@@ -8911,7 +9142,7 @@ flow_dv_modify_hdr_resource_release(struct mlx5_flow_handle *handle)
                claim_zero(mlx5_flow_os_destroy_flow_action
                                                (cache_resource->action));
                LIST_REMOVE(cache_resource, next);
-               rte_free(cache_resource);
+               mlx5_free(cache_resource);
                DRV_LOG(DEBUG, "modify-header resource %p: removed",
                        (void *)cache_resource);
                return 0;
@@ -9284,7 +9515,7 @@ flow_dv_destroy_mtr_tbl(struct rte_eth_dev *dev,
                flow_dv_tbl_resource_release(dev, mtd->transfer.sfx_tbl);
        if (mtd->drop_actn)
                claim_zero(mlx5_flow_os_destroy_flow_action(mtd->drop_actn));
-       rte_free(mtd);
+       mlx5_free(mtd);
        return 0;
 }
 
@@ -9417,7 +9648,7 @@ flow_dv_create_mtr_tbl(struct rte_eth_dev *dev,
                rte_errno = ENOTSUP;
                return NULL;
        }
-       mtb = rte_calloc(__func__, 1, sizeof(*mtb), 0);
+       mtb = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*mtb), 0, SOCKET_ID_ANY);
        if (!mtb) {
                DRV_LOG(ERR, "Failed to allocate memory for meter.");
                return NULL;
@@ -9830,3 +10061,4 @@ const struct mlx5_flow_driver_ops mlx5_flow_dv_drv_ops = {
 };
 
 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
+