net/mlx5: prepare sub-policy for flow with meter
authorLi Zhang <lizh@nvidia.com>
Tue, 27 Apr 2021 10:43:53 +0000 (13:43 +0300)
committerRaslan Darawsheh <rasland@nvidia.com>
Tue, 27 Apr 2021 11:20:35 +0000 (13:20 +0200)
When a flow has a RSS action, the driver splits
each sub flow finally is configured with
a different HW TIR action.

Any RSS action configured in meter policy may cause
a split in the flow configuration.
To save performance, any TIR action will be configured
in different flow table, so policy can be split to
sub-policies per TIR in the flow creation time.

Create a function to prepare the policy and
its sub-policies for a configured flow with meter.

Signed-off-by: Li Zhang <lizh@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
drivers/net/mlx5/mlx5_flow.h
drivers/net/mlx5/mlx5_flow_dv.c

index 98f6132..a80c790 100644 (file)
@@ -1094,6 +1094,11 @@ typedef int (*mlx5_flow_create_mtr_tbls_t)(struct rte_eth_dev *dev,
 typedef void (*mlx5_flow_destroy_mtr_tbls_t)(struct rte_eth_dev *dev,
                                struct mlx5_flow_meter_info *fm);
 typedef void (*mlx5_flow_destroy_mtr_drop_tbls_t)(struct rte_eth_dev *dev);
+typedef struct mlx5_flow_meter_sub_policy *
+       (*mlx5_flow_meter_sub_policy_rss_prepare_t)
+               (struct rte_eth_dev *dev,
+               struct mlx5_flow_meter_policy *mtr_policy,
+               struct mlx5_flow_rss_desc *rss_desc[MLX5_MTR_RTE_COLORS]);
 typedef uint32_t (*mlx5_flow_mtr_alloc_t)
                                            (struct rte_eth_dev *dev);
 typedef void (*mlx5_flow_mtr_free_t)(struct rte_eth_dev *dev,
@@ -1186,6 +1191,7 @@ struct mlx5_flow_driver_ops {
        mlx5_flow_destroy_policy_rules_t destroy_policy_rules;
        mlx5_flow_create_def_policy_t create_def_policy;
        mlx5_flow_destroy_def_policy_t destroy_def_policy;
+       mlx5_flow_meter_sub_policy_rss_prepare_t meter_sub_policy_rss_prepare;
        mlx5_flow_counter_alloc_t counter_alloc;
        mlx5_flow_counter_free_t counter_free;
        mlx5_flow_counter_query_t counter_query;
@@ -1417,6 +1423,10 @@ int mlx5_flow_create_mtr_tbls(struct rte_eth_dev *dev,
 void mlx5_flow_destroy_mtr_tbls(struct rte_eth_dev *dev,
                               struct mlx5_flow_meter_info *fm);
 void mlx5_flow_destroy_mtr_drop_tbls(struct rte_eth_dev *dev);
+struct mlx5_flow_meter_sub_policy *mlx5_flow_meter_sub_policy_rss_prepare
+               (struct rte_eth_dev *dev,
+               struct mlx5_flow_meter_policy *mtr_policy,
+               struct mlx5_flow_rss_desc *rss_desc[MLX5_MTR_RTE_COLORS]);
 int mlx5_flow_dv_discover_counter_offset_support(struct rte_eth_dev *dev);
 int mlx5_action_handle_flush(struct rte_eth_dev *dev);
 void mlx5_release_tunnel_hub(struct mlx5_dev_ctx_shared *sh, uint16_t port_id);
index 6e2a3e8..6bccdf5 100644 (file)
@@ -14874,6 +14874,150 @@ policy_error:
        return -1;
 }
 
+/**
+ * Find the policy table for prefix table with RSS.
+ *
+ * @param[in] dev
+ *   Pointer to Ethernet device.
+ * @param[in] mtr_policy
+ *   Pointer to meter policy table.
+ * @param[in] rss_desc
+ *   Pointer to rss_desc
+ * @return
+ *   Pointer to table set on success, NULL otherwise and rte_errno is set.
+ */
+static struct mlx5_flow_meter_sub_policy *
+flow_dv_meter_sub_policy_rss_prepare(struct rte_eth_dev *dev,
+               struct mlx5_flow_meter_policy *mtr_policy,
+               struct mlx5_flow_rss_desc *rss_desc[MLX5_MTR_RTE_COLORS])
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_flow_meter_sub_policy *sub_policy = NULL;
+       uint32_t sub_policy_idx = 0;
+       uint32_t hrxq_idx[MLX5_MTR_RTE_COLORS] = {0};
+       uint32_t i, j;
+       struct mlx5_hrxq *hrxq;
+       struct mlx5_flow_handle dh;
+       struct mlx5_meter_policy_action_container *act_cnt;
+       uint32_t domain = MLX5_MTR_DOMAIN_INGRESS;
+       uint16_t sub_policy_num;
+
+       rte_spinlock_lock(&mtr_policy->sl);
+       for (i = 0; i < MLX5_MTR_RTE_COLORS; i++) {
+               if (!rss_desc[i])
+                       continue;
+               hrxq_idx[i] = mlx5_hrxq_get(dev, rss_desc[i]);
+               if (!hrxq_idx[i]) {
+                       rte_spinlock_unlock(&mtr_policy->sl);
+                       return NULL;
+               }
+       }
+       sub_policy_num = (mtr_policy->sub_policy_num >>
+                       (MLX5_MTR_SUB_POLICY_NUM_SHIFT * domain)) &
+                       MLX5_MTR_SUB_POLICY_NUM_MASK;
+       for (i = 0; i < sub_policy_num;
+               i++) {
+               for (j = 0; j < MLX5_MTR_RTE_COLORS; j++) {
+                       if (rss_desc[j] &&
+                               hrxq_idx[j] !=
+                       mtr_policy->sub_policys[domain][i]->rix_hrxq[j])
+                               break;
+               }
+               if (j >= MLX5_MTR_RTE_COLORS) {
+                       /*
+                        * Found the sub policy table with
+                        * the same queue per color
+                        */
+                       rte_spinlock_unlock(&mtr_policy->sl);
+                       for (j = 0; j < MLX5_MTR_RTE_COLORS; j++)
+                               mlx5_hrxq_release(dev, hrxq_idx[j]);
+                       return mtr_policy->sub_policys[domain][i];
+               }
+       }
+       /* Create sub policy. */
+       if (!mtr_policy->sub_policys[domain][0]->rix_hrxq[0]) {
+               /* Reuse the first dummy sub_policy*/
+               sub_policy = mtr_policy->sub_policys[domain][0];
+               sub_policy_idx = sub_policy->idx;
+       } else {
+               sub_policy = mlx5_ipool_zmalloc
+                               (priv->sh->ipool[MLX5_IPOOL_MTR_POLICY],
+                               &sub_policy_idx);
+               if (!sub_policy ||
+                       sub_policy_idx > MLX5_MAX_SUB_POLICY_TBL_NUM) {
+                       for (i = 0; i < MLX5_MTR_RTE_COLORS; i++)
+                               mlx5_hrxq_release(dev, hrxq_idx[i]);
+                       goto rss_sub_policy_error;
+               }
+               sub_policy->idx = sub_policy_idx;
+               sub_policy->main_policy = mtr_policy;
+       }
+       for (i = 0; i < MLX5_MTR_RTE_COLORS; i++) {
+               if (!rss_desc[i])
+                       continue;
+               sub_policy->rix_hrxq[i] = hrxq_idx[i];
+               /*
+                * Overwrite the last action from
+                * RSS action to Queue action.
+                */
+               hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
+                             hrxq_idx[i]);
+               if (!hrxq) {
+                       DRV_LOG(ERR, "Failed to create policy hrxq");
+                       goto rss_sub_policy_error;
+               }
+               act_cnt = &mtr_policy->act_cnt[i];
+               if (act_cnt->rix_mark || act_cnt->modify_hdr) {
+                       memset(&dh, 0, sizeof(struct mlx5_flow_handle));
+                       if (act_cnt->rix_mark)
+                               dh.mark = 1;
+                       dh.fate_action = MLX5_FLOW_FATE_QUEUE;
+                       dh.rix_hrxq = hrxq_idx[i];
+                       flow_drv_rxq_flags_set(dev, &dh);
+               }
+       }
+       if (__flow_dv_create_policy_acts_rules(dev, mtr_policy,
+               sub_policy, domain)) {
+               DRV_LOG(ERR, "Failed to create policy "
+                       "rules per domain.");
+               goto rss_sub_policy_error;
+       }
+       if (sub_policy != mtr_policy->sub_policys[domain][0]) {
+               i = (mtr_policy->sub_policy_num >>
+                       (MLX5_MTR_SUB_POLICY_NUM_SHIFT * domain)) &
+                       MLX5_MTR_SUB_POLICY_NUM_MASK;
+               mtr_policy->sub_policys[domain][i] = sub_policy;
+               i++;
+               if (i > MLX5_MTR_RSS_MAX_SUB_POLICY)
+                       goto rss_sub_policy_error;
+               mtr_policy->sub_policy_num &= ~(MLX5_MTR_SUB_POLICY_NUM_MASK <<
+                       (MLX5_MTR_SUB_POLICY_NUM_SHIFT * domain));
+               mtr_policy->sub_policy_num |=
+                       (i & MLX5_MTR_SUB_POLICY_NUM_MASK) <<
+                       (MLX5_MTR_SUB_POLICY_NUM_SHIFT * domain);
+       }
+       rte_spinlock_unlock(&mtr_policy->sl);
+       return sub_policy;
+rss_sub_policy_error:
+       if (sub_policy) {
+               __flow_dv_destroy_sub_policy_rules(dev, sub_policy);
+               if (sub_policy != mtr_policy->sub_policys[domain][0]) {
+                       i = (mtr_policy->sub_policy_num >>
+                       (MLX5_MTR_SUB_POLICY_NUM_SHIFT * domain)) &
+                       MLX5_MTR_SUB_POLICY_NUM_MASK;
+                       mtr_policy->sub_policys[domain][i] = NULL;
+                       mlx5_ipool_free
+                       (priv->sh->ipool[MLX5_IPOOL_MTR_POLICY],
+                                       sub_policy->idx);
+               }
+       }
+       if (sub_policy_idx)
+               mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MTR_POLICY],
+                       sub_policy_idx);
+       rte_spinlock_unlock(&mtr_policy->sl);
+       return NULL;
+}
+
 /**
  * Validate the batch counter support in root table.
  *
@@ -15464,6 +15608,7 @@ const struct mlx5_flow_driver_ops mlx5_flow_dv_drv_ops = {
        .destroy_policy_rules = flow_dv_destroy_policy_rules,
        .create_def_policy = flow_dv_create_def_policy,
        .destroy_def_policy = flow_dv_destroy_def_policy,
+       .meter_sub_policy_rss_prepare = flow_dv_meter_sub_policy_rss_prepare,
        .counter_alloc = flow_dv_counter_allocate,
        .counter_free = flow_dv_counter_free,
        .counter_query = flow_dv_counter_query,