net/mlx5: skip table zero to improve insertion rate
authorDekel Peled <dekelp@mellanox.com>
Wed, 11 Sep 2019 11:03:36 +0000 (14:03 +0300)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 20 Sep 2019 08:19:41 +0000 (10:19 +0200)
E-switch tables one and above provide higher insertion rate
than table zero, as well as enhanced functionality.

This patch adds a mechanism to utilize these advantages, by creating
a default rule on port start, which directs all packets from e-switch
table zero to table one.
Other flow rules, requested for group n, will be created in
e-switch table n+1.
Jump action to e-switch group n will be created to group n+1.

Utility function mlx5_flow_group_to_table() is added to translate the
rte_flow group value to HW table value, and is called by PMD flow
engine on flow rule validation and creation.

Signed-off-by: Dekel Peled <dekelp@mellanox.com>
Acked-by: Matan Azrad <matan@mellanox.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
drivers/net/mlx5/mlx5.h
drivers/net/mlx5/mlx5_flow.c
drivers/net/mlx5/mlx5_flow.h
drivers/net/mlx5/mlx5_flow_dv.c
drivers/net/mlx5/mlx5_flow_verbs.c
drivers/net/mlx5/mlx5_trigger.c

index 239b56c..7982490 100644 (file)
@@ -513,8 +513,8 @@ struct mlx5_flow_tbl_resource {
        rte_atomic32_t refcnt; /**< Reference counter. */
 };
 
-#define MLX5_MAX_TABLES 0xffff
-#define MLX5_MAX_TABLES_FDB 0xffff
+#define MLX5_MAX_TABLES UINT16_MAX
+#define MLX5_MAX_TABLES_FDB UINT16_MAX
 
 #define MLX5_DBR_PAGE_SIZE 4096 /* Must be >= 512. */
 #define MLX5_DBR_SIZE 8
@@ -829,6 +829,7 @@ int mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
 int mlx5_ctrl_flow(struct rte_eth_dev *dev,
                   struct rte_flow_item_eth *eth_spec,
                   struct rte_flow_item_eth *eth_mask);
+struct rte_flow *mlx5_flow_create_esw_table_zero_flow(struct rte_eth_dev *dev);
 int mlx5_flow_create_drop_queue(struct rte_eth_dev *dev);
 void mlx5_flow_delete_drop_queue(struct rte_eth_dev *dev);
 void mlx5_flow_async_pool_query_handle(struct mlx5_ibv_shared *sh,
index eb36052..482f65b 100644 (file)
@@ -2013,6 +2013,7 @@ flow_null_validate(struct rte_eth_dev *dev __rte_unused,
                   const struct rte_flow_attr *attr __rte_unused,
                   const struct rte_flow_item items[] __rte_unused,
                   const struct rte_flow_action actions[] __rte_unused,
+                  bool external __rte_unused,
                   struct rte_flow_error *error)
 {
        return rte_flow_error_set(error, ENOTSUP,
@@ -2125,6 +2126,8 @@ flow_get_drv_type(struct rte_eth_dev *dev, const struct rte_flow_attr *attr)
  *   Pointer to the list of items.
  * @param[in] actions
  *   Pointer to the list of actions.
+ * @param[in] external
+ *   This flow rule is created by request external to PMD.
  * @param[out] error
  *   Pointer to the error structure.
  *
@@ -2136,13 +2139,13 @@ flow_drv_validate(struct rte_eth_dev *dev,
                  const struct rte_flow_attr *attr,
                  const struct rte_flow_item items[],
                  const struct rte_flow_action actions[],
-                 struct rte_flow_error *error)
+                 bool external, struct rte_flow_error *error)
 {
        const struct mlx5_flow_driver_ops *fops;
        enum mlx5_flow_drv_type type = flow_get_drv_type(dev, attr);
 
        fops = flow_get_drv_ops(type);
-       return fops->validate(dev, attr, items, actions, error);
+       return fops->validate(dev, attr, items, actions, external, error);
 }
 
 /**
@@ -2314,7 +2317,7 @@ mlx5_flow_validate(struct rte_eth_dev *dev,
 {
        int ret;
 
-       ret = flow_drv_validate(dev, attr, items, actions, error);
+       ret = flow_drv_validate(dev, attr, items, actions, true, error);
        if (ret < 0)
                return ret;
        return 0;
@@ -2376,6 +2379,8 @@ find_graph_root(const struct rte_flow_item pattern[], uint32_t rss_level)
  *   Pattern specification (list terminated by the END pattern item).
  * @param[in] actions
  *   Associated actions (list terminated by the END action).
+ * @param[in] external
+ *   This flow rule is created by request external to PMD.
  * @param[out] error
  *   Perform verbose error reporting if not NULL.
  *
@@ -2387,7 +2392,7 @@ flow_list_create(struct rte_eth_dev *dev, struct mlx5_flows *list,
                 const struct rte_flow_attr *attr,
                 const struct rte_flow_item items[],
                 const struct rte_flow_action actions[],
-                struct rte_flow_error *error)
+                bool external, struct rte_flow_error *error)
 {
        struct rte_flow *flow = NULL;
        struct mlx5_flow *dev_flow;
@@ -2401,7 +2406,7 @@ flow_list_create(struct rte_eth_dev *dev, struct mlx5_flows *list,
        uint32_t i;
        uint32_t flow_size;
 
-       ret = flow_drv_validate(dev, attr, items, actions, error);
+       ret = flow_drv_validate(dev, attr, items, actions, external, error);
        if (ret < 0)
                return NULL;
        flow_size = sizeof(struct rte_flow);
@@ -2443,6 +2448,7 @@ flow_list_create(struct rte_eth_dev *dev, struct mlx5_flows *list,
                if (!dev_flow)
                        goto error;
                dev_flow->flow = flow;
+               dev_flow->external = external;
                LIST_INSERT_HEAD(&flow->dev_flows, dev_flow, next);
                ret = flow_drv_translate(dev, dev_flow, attr,
                                         buf->entry[i].pattern,
@@ -2467,6 +2473,55 @@ error:
        return NULL;
 }
 
+/**
+ * Create a dedicated flow rule on e-switch table 0 (root table), to direct all
+ * incoming packets to table 1.
+ *
+ * Other flow rules, requested for group n, will be created in
+ * e-switch table n+1.
+ * Jump action to e-switch group n will be created to group n+1.
+ *
+ * Used when working in switchdev mode, to utilise advantages of table 1
+ * and above.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ *
+ * @return
+ *   Pointer to flow on success, NULL otherwise and rte_errno is set.
+ */
+struct rte_flow *
+mlx5_flow_create_esw_table_zero_flow(struct rte_eth_dev *dev)
+{
+       const struct rte_flow_attr attr = {
+               .group = 0,
+               .priority = 0,
+               .ingress = 1,
+               .egress = 0,
+               .transfer = 1,
+       };
+       const struct rte_flow_item pattern = {
+               .type = RTE_FLOW_ITEM_TYPE_END,
+       };
+       struct rte_flow_action_jump jump = {
+               .group = 1,
+       };
+       const struct rte_flow_action actions[] = {
+               {
+                       .type = RTE_FLOW_ACTION_TYPE_JUMP,
+                       .conf = &jump,
+               },
+               {
+                       .type = RTE_FLOW_ACTION_TYPE_END,
+               },
+       };
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct rte_flow_error error;
+
+       return flow_list_create(dev, &priv->ctrl_flows, &attr, &pattern,
+                               actions, false, &error);
+}
+
 /**
  * Create a flow.
  *
@@ -2483,7 +2538,7 @@ mlx5_flow_create(struct rte_eth_dev *dev,
        struct mlx5_priv *priv = dev->data->dev_private;
 
        return flow_list_create(dev, &priv->flows,
-                               attr, items, actions, error);
+                               attr, items, actions, true, error);
 }
 
 /**
@@ -2680,7 +2735,7 @@ mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
        for (i = 0; i != priv->reta_idx_n; ++i)
                queue[i] = (*priv->reta_idx)[i];
        flow = flow_list_create(dev, &priv->ctrl_flows,
-                               &attr, items, actions, &error);
+                               &attr, items, actions, false, &error);
        if (!flow)
                return -rte_errno;
        return 0;
@@ -3094,7 +3149,8 @@ flow_fdir_filter_add(struct rte_eth_dev *dev,
                goto error;
        }
        flow = flow_list_create(dev, &priv->flows, &fdir_flow->attr,
-                               fdir_flow->items, fdir_flow->actions, NULL);
+                               fdir_flow->items, fdir_flow->actions, true,
+                               NULL);
        if (!flow)
                goto error;
        assert(!flow->fdir);
@@ -3442,3 +3498,39 @@ mlx5_flow_async_pool_query_handle(struct mlx5_ibv_shared *sh,
        pool->raw_hw = NULL;
        sh->cmng.pending_queries--;
 }
+
+/**
+ * Translate the rte_flow group index to HW table value.
+ *
+ * @param[in] attributes
+ *   Pointer to flow attributes
+ * @param[in] external
+ *   Value is part of flow rule created by request external to PMD.
+ * @param[in] group
+ *   rte_flow group index value.
+ * @param[out] table
+ *   HW table value.
+ * @param[out] error
+ *   Pointer to error structure.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_flow_group_to_table(const struct rte_flow_attr *attributes, bool external,
+                        uint32_t group, uint32_t *table,
+                        struct rte_flow_error *error)
+{
+       if (attributes->transfer && external) {
+               if (group == UINT32_MAX)
+                       return rte_flow_error_set
+                                               (error, EINVAL,
+                                                RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
+                                                NULL,
+                                                "group index not supported");
+               *table = group + 1;
+       } else {
+               *table = group;
+       }
+       return 0;
+}
index 95abbb8..235bccd 100644 (file)
@@ -391,6 +391,7 @@ struct mlx5_flow {
 #endif
                struct mlx5_flow_verbs verbs;
        };
+       bool external; /**< true if the flow is created external to PMD. */
 };
 
 /* Flow structure. */
@@ -417,6 +418,7 @@ typedef int (*mlx5_flow_validate_t)(struct rte_eth_dev *dev,
                                    const struct rte_flow_attr *attr,
                                    const struct rte_flow_item items[],
                                    const struct rte_flow_action actions[],
+                                   bool external,
                                    struct rte_flow_error *error);
 typedef struct mlx5_flow *(*mlx5_flow_prepare_t)
        (const struct rte_flow_attr *attr, const struct rte_flow_item items[],
@@ -455,6 +457,9 @@ struct mlx5_flow_driver_ops {
 
 /* mlx5_flow.c */
 
+int mlx5_flow_group_to_table(const struct rte_flow_attr *attributes,
+                            bool external, uint32_t group, uint32_t *table,
+                            struct rte_flow_error *error);
 uint64_t mlx5_flow_hashfields_adjust(struct mlx5_flow *dev_flow, int tunnel,
                                     uint64_t layer_types,
                                     uint64_t hash_fields);
index 330a868..c234d13 100644 (file)
@@ -2366,8 +2366,10 @@ flow_dv_validate_action_modify_ttl(const uint64_t action_flags,
  *   Pointer to the jump action.
  * @param[in] action_flags
  *   Holds the actions detected until now.
- * @param[in] group
- *   The group of the current flow.
+ * @param[in] attributes
+ *   Pointer to flow attributes
+ * @param[in] external
+ *   Action belongs to flow rule created by request external to PMD.
  * @param[out] error
  *   Pointer to error structure.
  *
@@ -2377,9 +2379,14 @@ flow_dv_validate_action_modify_ttl(const uint64_t action_flags,
 static int
 flow_dv_validate_action_jump(const struct rte_flow_action *action,
                             uint64_t action_flags,
-                            uint32_t group,
-                            struct rte_flow_error *error)
+                            const struct rte_flow_attr *attributes,
+                            bool external, struct rte_flow_error *error)
 {
+       uint32_t max_group = attributes->transfer ? MLX5_MAX_TABLES_FDB :
+                                                   MLX5_MAX_TABLES;
+       uint32_t target_group, table;
+       int ret = 0;
+
        if (action_flags & (MLX5_FLOW_FATE_ACTIONS |
                            MLX5_FLOW_FATE_ESWITCH_ACTIONS))
                return rte_flow_error_set(error, EINVAL,
@@ -2390,10 +2397,20 @@ flow_dv_validate_action_jump(const struct rte_flow_action *action,
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ACTION_CONF,
                                          NULL, "action configuration not set");
-       if (group >= ((const struct rte_flow_action_jump *)action->conf)->group)
+       target_group =
+               ((const struct rte_flow_action_jump *)action->conf)->group;
+       ret = mlx5_flow_group_to_table(attributes, external, target_group,
+                                      &table, error);
+       if (ret)
+               return ret;
+       if (table >= max_group)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ATTR_GROUP, NULL,
+                                         "target group index out of range");
+       if (attributes->group >= target_group)
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ACTION, NULL,
-                                         "target group must be higher then"
+                                         "target group must be higher than"
                                          " the current flow group");
        return 0;
 }
@@ -3174,6 +3191,8 @@ flow_dv_counter_release(struct rte_eth_dev *dev,
  *   Pointer to dev struct.
  * @param[in] attributes
  *   Pointer to flow attributes
+ * @param[in] external
+ *   This flow rule is created by request external to PMD.
  * @param[out] error
  *   Pointer to error structure.
  *
@@ -3183,6 +3202,7 @@ flow_dv_counter_release(struct rte_eth_dev *dev,
 static int
 flow_dv_validate_attributes(struct rte_eth_dev *dev,
                            const struct rte_flow_attr *attributes,
+                           bool external __rte_unused,
                            struct rte_flow_error *error)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
@@ -3193,7 +3213,22 @@ flow_dv_validate_attributes(struct rte_eth_dev *dev,
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
                                          NULL,
-                                         "groups is not supported");
+                                         "groups are not supported");
+#else
+       uint32_t max_group = attributes->transfer ? MLX5_MAX_TABLES_FDB :
+                                                   MLX5_MAX_TABLES;
+       uint32_t table;
+       int ret;
+
+       ret = mlx5_flow_group_to_table(attributes, external,
+                                      attributes->group,
+                                      &table, error);
+       if (ret)
+               return ret;
+       if (table >= max_group)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ATTR_GROUP, NULL,
+                                         "group index out of range");
 #endif
        if (attributes->priority != MLX5_FLOW_PRIO_RSVD &&
            attributes->priority >= priority_max)
@@ -3217,12 +3252,6 @@ flow_dv_validate_attributes(struct rte_eth_dev *dev,
                                (error, ENOTSUP,
                                 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, attributes,
                                 "egress is not supported");
-               if (attributes->group >= MLX5_MAX_TABLES_FDB)
-                       return rte_flow_error_set
-                               (error, EINVAL,
-                                RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
-                                NULL, "group must be smaller than "
-                                RTE_STR(MLX5_MAX_TABLES_FDB));
        }
        if (!(attributes->egress ^ attributes->ingress))
                return rte_flow_error_set(error, ENOTSUP,
@@ -3243,6 +3272,8 @@ flow_dv_validate_attributes(struct rte_eth_dev *dev,
  *   Pointer to the list of items.
  * @param[in] actions
  *   Pointer to the list of actions.
+ * @param[in] external
+ *   This flow rule is created by request external to PMD.
  * @param[out] error
  *   Pointer to the error structure.
  *
@@ -3253,7 +3284,7 @@ static int
 flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                 const struct rte_flow_item items[],
                 const struct rte_flow_action actions[],
-                struct rte_flow_error *error)
+                bool external, struct rte_flow_error *error)
 {
        int ret;
        uint64_t action_flags = 0;
@@ -3272,7 +3303,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
 
        if (items == NULL)
                return -1;
-       ret = flow_dv_validate_attributes(dev, attr, error);
+       ret = flow_dv_validate_attributes(dev, attr, external, error);
        if (ret < 0)
                return ret;
        for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
@@ -3675,7 +3706,8 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                case RTE_FLOW_ACTION_TYPE_JUMP:
                        ret = flow_dv_validate_action_jump(actions,
                                                           action_flags,
-                                                          attr->group, error);
+                                                          attr, external,
+                                                          error);
                        if (ret)
                                return ret;
                        ++actions_n;
@@ -5142,8 +5174,14 @@ flow_dv_translate(struct rte_eth_dev *dev,
        struct rte_vlan_hdr vlan = { 0 };
        bool vlan_inherited = false;
        uint16_t vlan_tci;
+       uint32_t table;
+       int ret = 0;
 
-       flow->group = attr->group;
+       ret = mlx5_flow_group_to_table(attr, dev_flow->external, attr->group,
+                                      &table, error);
+       if (ret)
+               return ret;
+       flow->group = table;
        if (attr->transfer)
                res.ft_type = MLX5DV_FLOW_TABLE_TYPE_FDB;
        if (priority == MLX5_FLOW_PRIO_RSVD)
@@ -5229,7 +5267,7 @@ flow_dv_translate(struct rte_eth_dev *dev,
                        flow->counter = flow_dv_counter_alloc(dev,
                                                              count->shared,
                                                              count->id,
-                                                             attr->group);
+                                                             flow->group);
                        if (flow->counter == NULL)
                                goto cnt_err;
                        dev_flow->dv.actions[actions_n++] =
@@ -5371,7 +5409,12 @@ cnt_err:
                        break;
                case RTE_FLOW_ACTION_TYPE_JUMP:
                        jump_data = action->conf;
-                       tbl = flow_dv_tbl_resource_get(dev, jump_data->group,
+                       ret = mlx5_flow_group_to_table(attr, dev_flow->external,
+                                                      jump_data->group, &table,
+                                                      error);
+                       if (ret)
+                               return ret;
+                       tbl = flow_dv_tbl_resource_get(dev, table,
                                                       attr->egress,
                                                       attr->transfer, error);
                        if (!tbl)
@@ -5522,7 +5565,7 @@ cnt_err:
                        mlx5_flow_tunnel_ip_check(items, next_protocol,
                                                  &item_flags, &tunnel);
                        flow_dv_translate_item_ipv4(match_mask, match_value,
-                                                   items, tunnel, attr->group);
+                                                   items, tunnel, flow->group);
                        matcher.priority = MLX5_PRIORITY_MAP_L3;
                        dev_flow->dv.hash_fields |=
                                mlx5_flow_hashfields_adjust
@@ -5549,7 +5592,7 @@ cnt_err:
                        mlx5_flow_tunnel_ip_check(items, next_protocol,
                                                  &item_flags, &tunnel);
                        flow_dv_translate_item_ipv6(match_mask, match_value,
-                                                   items, tunnel, attr->group);
+                                                   items, tunnel, flow->group);
                        matcher.priority = MLX5_PRIORITY_MAP_L3;
                        dev_flow->dv.hash_fields |=
                                mlx5_flow_hashfields_adjust
@@ -5668,7 +5711,7 @@ cnt_err:
        matcher.priority = mlx5_flow_adjust_priority(dev, priority,
                                                     matcher.priority);
        matcher.egress = attr->egress;
-       matcher.group = attr->group;
+       matcher.group = flow->group;
        matcher.transfer = attr->transfer;
        if (flow_dv_matcher_register(dev, &matcher, dev_flow, error))
                return -rte_errno;
index c5b28e3..23110f2 100644 (file)
@@ -1016,6 +1016,8 @@ flow_verbs_translate_action_count(struct mlx5_flow *dev_flow,
  *   Pointer to the list of items.
  * @param[in] actions
  *   Pointer to the list of actions.
+ * @param[in] external
+ *   This flow rule is created by request external to PMD.
  * @param[out] error
  *   Pointer to the error structure.
  *
@@ -1027,6 +1029,7 @@ flow_verbs_validate(struct rte_eth_dev *dev,
                    const struct rte_flow_attr *attr,
                    const struct rte_flow_item items[],
                    const struct rte_flow_action actions[],
+                   bool external __rte_unused,
                    struct rte_flow_error *error)
 {
        int ret;
index aa323ad..122f31c 100644 (file)
@@ -286,6 +286,9 @@ mlx5_traffic_enable(struct rte_eth_dev *dev)
        unsigned int j;
        int ret;
 
+       if (priv->config.dv_esw_en && !priv->config.vf)
+               if (!mlx5_flow_create_esw_table_zero_flow(dev))
+                       goto error;
        if (priv->isolated)
                return 0;
        if (dev->data->promiscuous) {