net/mlx4: support flow API RSS action
authorVasily Philipov <vasilyf@mellanox.com>
Wed, 5 Jul 2017 08:14:11 +0000 (11:14 +0300)
committerFerruh Yigit <ferruh.yigit@intel.com>
Thu, 6 Jul 2017 13:00:57 +0000 (15:00 +0200)
This commit adds support for the flow API RSS action with the following
limitations:

 - Only supported when isolated mode is enabled.
 - The number of queues specified by the action (rte_flow_action_rss.num)
   must be a power of two.
 - Each queue index can be specified at most once in the configuration
   (rte_flow_action_rss.queue[]).
 - Because a queue can be associated with a single RSS context, it cannot
   be targeted by multiple RSS actions simultaneously.

Signed-off-by: Vasily Philipov <vasilyf@mellanox.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
drivers/net/mlx4/mlx4.c
drivers/net/mlx4/mlx4.h
drivers/net/mlx4/mlx4_flow.c
drivers/net/mlx4/mlx4_flow.h

index e03ce5f..d41552b 100644 (file)
@@ -569,7 +569,7 @@ rxq_cleanup(struct rxq *rxq);
  * @return
  *   Pointer to a parent rxq structure, NULL on failure.
  */
-static struct rxq *
+struct rxq *
 priv_parent_create(struct priv *priv,
                   uint16_t queues[],
                   uint16_t children_n)
@@ -705,10 +705,8 @@ dev_configure(struct rte_eth_dev *dev)
        priv->rss = 1;
        tmp = priv->rxqs_n;
        priv->rxqs_n = rxqs_n;
-       if (priv->isolated) {
-               priv->rss = 0;
+       if (priv->isolated)
                return 0;
-       }
        if (priv_parent_create(priv, NULL, priv->rxqs_n))
                return 0;
        /* Failure, rollback. */
index 4e2a0ae..a2e0ae7 100644 (file)
@@ -372,4 +372,9 @@ rxq_create_qp(struct rxq *rxq,
 void
 rxq_parent_cleanup(struct rxq *parent);
 
+struct rxq *
+priv_parent_create(struct priv *priv,
+                  uint16_t queues[],
+                  uint16_t children_n);
+
 #endif /* RTE_PMD_MLX4_H_ */
index 1344101..8ade106 100644 (file)
@@ -112,6 +112,7 @@ struct rte_flow_drop {
 static const enum rte_flow_action_type valid_actions[] = {
        RTE_FLOW_ACTION_TYPE_DROP,
        RTE_FLOW_ACTION_TYPE_QUEUE,
+       RTE_FLOW_ACTION_TYPE_RSS,
        RTE_FLOW_ACTION_TYPE_END,
 };
 
@@ -672,6 +673,76 @@ priv_flow_validate(struct priv *priv,
                        if (!queue || (queue->index > (priv->rxqs_n - 1)))
                                goto exit_action_not_supported;
                        action.queue = 1;
+                       action.queues_n = 1;
+                       action.queues[0] = queue->index;
+               } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
+                       int i;
+                       int ierr;
+                       const struct rte_flow_action_rss *rss =
+                               (const struct rte_flow_action_rss *)
+                               actions->conf;
+
+                       if (!priv->hw_rss) {
+                               rte_flow_error_set(error, ENOTSUP,
+                                          RTE_FLOW_ERROR_TYPE_ACTION,
+                                          actions,
+                                          "RSS cannot be used with "
+                                          "the current configuration");
+                               return -rte_errno;
+                       }
+                       if (!priv->isolated) {
+                               rte_flow_error_set(error, ENOTSUP,
+                                          RTE_FLOW_ERROR_TYPE_ACTION,
+                                          actions,
+                                          "RSS cannot be used without "
+                                          "isolated mode");
+                               return -rte_errno;
+                       }
+                       if (!rte_is_power_of_2(rss->num)) {
+                               rte_flow_error_set(error, ENOTSUP,
+                                          RTE_FLOW_ERROR_TYPE_ACTION,
+                                          actions,
+                                          "the number of queues "
+                                          "should be power of two");
+                               return -rte_errno;
+                       }
+                       if (priv->max_rss_tbl_sz < rss->num) {
+                               rte_flow_error_set(error, ENOTSUP,
+                                          RTE_FLOW_ERROR_TYPE_ACTION,
+                                          actions,
+                                          "the number of queues "
+                                          "is too large");
+                               return -rte_errno;
+                       }
+                       /* checking indexes array */
+                       ierr = 0;
+                       for (i = 0; i < rss->num; ++i) {
+                               int j;
+                               if (rss->queue[i] >= priv->rxqs_n)
+                                       ierr = 1;
+                               /*
+                                * Prevent the user from specifying
+                                * the same queue twice in the RSS array.
+                                */
+                               for (j = i + 1; j < rss->num && !ierr; ++j)
+                                       if (rss->queue[j] == rss->queue[i])
+                                               ierr = 1;
+                               if (ierr) {
+                                       rte_flow_error_set(
+                                               error,
+                                               ENOTSUP,
+                                               RTE_FLOW_ERROR_TYPE_HANDLE,
+                                               NULL,
+                                               "RSS action only supports "
+                                               "unique queue indices "
+                                               "in a list");
+                                       return -rte_errno;
+                               }
+                       }
+                       action.queue = 1;
+                       action.queues_n = rss->num;
+                       for (i = 0; i < rss->num; ++i)
+                               action.queues[i] = rss->queue[i];
                } else {
                        goto exit_action_not_supported;
                }
@@ -796,6 +867,82 @@ err:
        return -1;
 }
 
+/**
+ * Get RSS parent rxq structure for given queues.
+ *
+ * Creates a new or returns an existed one.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param queues
+ *   queues indices array, NULL in default RSS case.
+ * @param children_n
+ *   the size of queues array.
+ *
+ * @return
+ *   Pointer to a parent rxq structure, NULL on failure.
+ */
+static struct rxq *
+priv_parent_get(struct priv *priv,
+               uint16_t queues[],
+               uint16_t children_n,
+               struct rte_flow_error *error)
+{
+       unsigned int i;
+       struct rxq *parent;
+
+       for (parent = LIST_FIRST(&priv->parents);
+            parent;
+            parent = LIST_NEXT(parent, next)) {
+               unsigned int same = 0;
+               unsigned int overlap = 0;
+
+               /*
+                * Find out whether an appropriate parent queue already exists
+                * and can be reused, otherwise make sure there are no overlaps.
+                */
+               for (i = 0; i < children_n; ++i) {
+                       unsigned int j;
+
+                       for (j = 0; j < parent->rss.queues_n; ++j) {
+                               if (parent->rss.queues[j] != queues[i])
+                                       continue;
+                               ++overlap;
+                               if (i == j)
+                                       ++same;
+                       }
+               }
+               if (same == children_n &&
+                       children_n == parent->rss.queues_n)
+                       return parent;
+               else if (overlap)
+                       goto error;
+       }
+       /* Exclude the cases when some QPs were created without RSS */
+       for (i = 0; i < children_n; ++i) {
+               struct rxq *rxq = (*priv->rxqs)[queues[i]];
+               if (rxq->qp)
+                       goto error;
+       }
+       parent = priv_parent_create(priv, queues, children_n);
+       if (!parent) {
+               rte_flow_error_set(error,
+                                  ENOMEM, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                  NULL, "flow rule creation failure");
+               return NULL;
+       }
+       return parent;
+
+error:
+       rte_flow_error_set(error,
+                          EEXIST,
+                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                          NULL,
+                          "sharing a queue between several"
+                          " RSS groups is not supported");
+       return NULL;
+}
+
 /**
  * Complete flow rule creation.
  *
@@ -819,6 +966,7 @@ priv_flow_create_action_queue(struct priv *priv,
 {
        struct ibv_qp *qp;
        struct rte_flow *rte_flow;
+       struct rxq *rxq_parent = NULL;
 
        assert(priv->pd);
        assert(priv->ctx);
@@ -832,23 +980,38 @@ priv_flow_create_action_queue(struct priv *priv,
                qp = priv->flow_drop_queue->qp;
        } else {
                int ret;
-               struct rxq *rxq = (*priv->rxqs)[action->queue_id];
-
-               if (!rxq->qp) {
-                       assert(priv->isolated);
-                       ret = rxq_create_qp(rxq, rxq->elts_n,
-                                           0, 0, NULL);
-                       if (ret) {
-                               rte_flow_error_set(
-                                       error,
-                                       ENOMEM,
-                                       RTE_FLOW_ERROR_TYPE_HANDLE,
-                                       NULL,
-                                       "flow rule creation failure");
+               unsigned int i;
+               struct rxq *rxq = NULL;
+
+               if (action->queues_n > 1) {
+                       rxq_parent = priv_parent_get(priv, action->queues,
+                                                    action->queues_n, error);
+                       if (!rxq_parent)
                                goto error;
+               }
+               for (i = 0; i < action->queues_n; ++i) {
+                       rxq = (*priv->rxqs)[action->queues[i]];
+                       /*
+                        * In case of isolated mode we postpone
+                        * ibv receive queue creation till the first
+                        * rte_flow rule will be applied on that queue.
+                        */
+                       if (!rxq->qp) {
+                               assert(priv->isolated);
+                               ret = rxq_create_qp(rxq, rxq->elts_n,
+                                                   0, 0, rxq_parent);
+                               if (ret) {
+                                       rte_flow_error_set(
+                                               error,
+                                               ENOMEM,
+                                               RTE_FLOW_ERROR_TYPE_HANDLE,
+                                               NULL,
+                                               "flow rule creation failure");
+                                       goto error;
+                               }
                        }
                }
-               qp = rxq->qp;
+               qp = action->queues_n > 1 ? rxq_parent->qp : rxq->qp;
                rte_flow->qp = qp;
        }
        rte_flow->ibv_attr = ibv_attr;
@@ -861,6 +1024,8 @@ priv_flow_create_action_queue(struct priv *priv,
        return rte_flow;
 
 error:
+       if (rxq_parent)
+               rxq_parent_cleanup(rxq_parent);
        rte_free(rte_flow);
        return NULL;
 }
@@ -924,11 +1089,22 @@ priv_flow_create(struct priv *priv,
                        continue;
                } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
                        action.queue = 1;
-                       action.queue_id =
+                       action.queues_n = 1;
+                       action.queues[0] =
                                ((const struct rte_flow_action_queue *)
                                 actions->conf)->index;
                } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
                        action.drop = 1;
+               } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
+                       unsigned int i;
+                       const struct rte_flow_action_rss *rss =
+                               (const struct rte_flow_action_rss *)
+                                actions->conf;
+
+                       action.queue = 1;
+                       action.queues_n = rss->num;
+                       for (i = 0; i < rss->num; ++i)
+                               action.queues[i] = rss->queue[i];
                } else {
                        rte_flow_error_set(error, ENOTSUP,
                                           RTE_FLOW_ERROR_TYPE_ACTION,
index 4d007da..beabcf2 100644 (file)
@@ -98,7 +98,8 @@ mlx4_flow_isolate(struct rte_eth_dev *dev,
 struct mlx4_flow_action {
        uint32_t drop:1; /**< Target is a drop queue. */
        uint32_t queue:1; /**< Target is a receive queue. */
-       uint32_t queue_id; /**< Identifier of the queue. */
+       uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queue indices to use. */
+       uint16_t queues_n; /**< Number of entries in queue[] */
 };
 
 int mlx4_priv_flow_start(struct priv *priv);