net/mlx5: fix RSS expansion for inner tunnel VLAN
authorLior Margalit <lmargalit@nvidia.com>
Tue, 3 Aug 2021 18:13:22 +0000 (21:13 +0300)
committerRaslan Darawsheh <rasland@nvidia.com>
Thu, 19 Aug 2021 08:13:35 +0000 (10:13 +0200)
The RSS expansion algorithm is using a graph to find the possible
expansion paths. The VLAN item in the flow pattern requires special
treatment, because it should not be added implicitly by the expansion
algorithm.  If the flow pattern ends with ETH item, the pattern will be
expanded with IPv4 and IPv6.
For example:
testpmd> flow create ... eth / end actions rss / end
ETH END
ETH IPV4 END
ETH IPV6 END
If a VLAN item follows the ETH item in the flow pattern, the pattern
will be expanded with IPv4 and IPv6 following the VLAN item.
For example:
testpmd> flow create ... eth / vlan / end actions rss level 1 / end
ETH VLAN END
ETH VLAN IPV4 END
ETH VLAN IPV6 END

The case of inner tunnel VLAN item was not taken care of so the flow
pattern did not expand with IPv6 and IPv4 as expected.
Example with inner VLAN:
testpmd> flow create ... / vxlan / eth / vlan / end actions rss level 2
/ end
The current result of the expansion alg:
ETH IPV6 UDP VXLAN ETH VLAN END
The expected result of the expansion alg:
ETH IPV6 UDP VXLAN ETH VLAN END
ETH IPV6 UDP VXLAN ETH VLAN IPV4 END
ETH IPV6 UDP VXLAN ETH VLAN IPV6 END

The fix is to introduce a new flag to set on a graph expansion node
to apply the 'explicit' behavior, meaning the node is not added to
the expanded pattern, if it is not found in the flow pattern, but the
expansion alg can go deeper to its next nodes.

Fixes: c7870bfe09dc ("ethdev: move RSS expansion code to mlx5 driver")
Cc: stable@dpdk.org
Signed-off-by: Lior Margalit <lmargalit@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
drivers/net/mlx5/mlx5_flow.c

index e63a297..4762fa0 100644 (file)
@@ -100,10 +100,25 @@ struct mlx5_flow_expand_node {
         * RSS types bit-field associated with this node
         * (see ETH_RSS_* definitions).
         */
-       uint8_t optional;
-       /**< optional expand field. Default 0 to expand, 1 not go deeper. */
+       uint64_t node_flags;
+       /**<
+        *  Bit-fields that define how the node is used in the expansion.
+        * (see MLX5_EXPANSION_NODE_* definitions).
+        */
 };
 
+/* Optional expand field. The expansion alg will not go deeper. */
+#define MLX5_EXPANSION_NODE_OPTIONAL (UINT64_C(1) << 0)
+
+/* The node is not added implicitly as expansion to the flow pattern.
+ * If the node type does not match the flow pattern item type, the
+ * expansion alg will go deeper to its next items.
+ * In the current implementation, the list of next nodes indexes can
+ * have up to one node with this flag set and it has to be the last
+ * node index (before the list terminator).
+ */
+#define MLX5_EXPANSION_NODE_EXPLICIT (UINT64_C(1) << 1)
+
 /** Object returned by mlx5_flow_expand_rss(). */
 struct mlx5_flow_expand_rss {
        uint32_t entries;
@@ -308,10 +323,17 @@ mlx5_flow_expand_rss(struct mlx5_flow_expand_rss *buf, size_t size,
                        continue;
                }
                last_item = item;
-               for (i = 0; node->next && node->next[i]; ++i) {
+               i = 0;
+               while (node->next && node->next[i]) {
                        next = &graph[node->next[i]];
                        if (next->type == item->type)
                                break;
+                       if (next->node_flags & MLX5_EXPANSION_NODE_EXPLICIT) {
+                               node = next;
+                               i = 0;
+                       } else {
+                               ++i;
+                       }
                }
                if (next)
                        node = next;
@@ -370,6 +392,16 @@ mlx5_flow_expand_rss(struct mlx5_flow_expand_rss *buf, size_t size,
        }
        memset(flow_items, 0, sizeof(flow_items));
        next_node = node->next;
+       while (next_node) {
+               /*
+                * Skip the nodes with the MLX5_EXPANSION_NODE_EXPLICIT
+                * flag set, because they were not found in the flow pattern.
+                */
+               node = &graph[*next_node];
+               if (!(node->node_flags & MLX5_EXPANSION_NODE_EXPLICIT))
+                       break;
+               next_node = node->next;
+       }
        stack[stack_pos] = next_node;
        node = next_node ? &graph[*next_node] : NULL;
        while (node) {
@@ -404,7 +436,8 @@ mlx5_flow_expand_rss(struct mlx5_flow_expand_rss *buf, size_t size,
                        addr = (void *)(((uintptr_t)addr) + n);
                }
                /* Go deeper. */
-               if (!node->optional && node->next) {
+               if (!(node->node_flags & MLX5_EXPANSION_NODE_OPTIONAL) &&
+                               node->next) {
                        next_node = node->next;
                        if (stack_pos++ == MLX5_RSS_EXP_ELT_N) {
                                rte_errno = E2BIG;
@@ -429,10 +462,7 @@ mlx5_flow_expand_rss(struct mlx5_flow_expand_rss *buf, size_t size,
 enum mlx5_expansion {
        MLX5_EXPANSION_ROOT,
        MLX5_EXPANSION_ROOT_OUTER,
-       MLX5_EXPANSION_ROOT_ETH_VLAN,
-       MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN,
        MLX5_EXPANSION_OUTER_ETH,
-       MLX5_EXPANSION_OUTER_ETH_VLAN,
        MLX5_EXPANSION_OUTER_VLAN,
        MLX5_EXPANSION_OUTER_IPV4,
        MLX5_EXPANSION_OUTER_IPV4_UDP,
@@ -447,7 +477,6 @@ enum mlx5_expansion {
        MLX5_EXPANSION_GRE_KEY,
        MLX5_EXPANSION_MPLS,
        MLX5_EXPANSION_ETH,
-       MLX5_EXPANSION_ETH_VLAN,
        MLX5_EXPANSION_VLAN,
        MLX5_EXPANSION_IPV4,
        MLX5_EXPANSION_IPV4_UDP,
@@ -473,22 +502,7 @@ static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
                                                  MLX5_EXPANSION_OUTER_IPV6),
                .type = RTE_FLOW_ITEM_TYPE_END,
        },
-       [MLX5_EXPANSION_ROOT_ETH_VLAN] = {
-               .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH_VLAN),
-               .type = RTE_FLOW_ITEM_TYPE_END,
-       },
-       [MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN] = {
-               .next = MLX5_FLOW_EXPAND_RSS_NEXT
-                                               (MLX5_EXPANSION_OUTER_ETH_VLAN),
-               .type = RTE_FLOW_ITEM_TYPE_END,
-       },
        [MLX5_EXPANSION_OUTER_ETH] = {
-               .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_IPV4,
-                                                 MLX5_EXPANSION_OUTER_IPV6),
-               .type = RTE_FLOW_ITEM_TYPE_ETH,
-               .rss_types = 0,
-       },
-       [MLX5_EXPANSION_OUTER_ETH_VLAN] = {
                .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_VLAN),
                .type = RTE_FLOW_ITEM_TYPE_ETH,
                .rss_types = 0,
@@ -497,6 +511,7 @@ static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
                .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_IPV4,
                                                  MLX5_EXPANSION_OUTER_IPV6),
                .type = RTE_FLOW_ITEM_TYPE_VLAN,
+               .node_flags = MLX5_EXPANSION_NODE_EXPLICIT,
        },
        [MLX5_EXPANSION_OUTER_IPV4] = {
                .next = MLX5_FLOW_EXPAND_RSS_NEXT
@@ -570,7 +585,7 @@ static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
                                                  MLX5_EXPANSION_IPV6,
                                                  MLX5_EXPANSION_MPLS),
                .type = RTE_FLOW_ITEM_TYPE_GRE_KEY,
-               .optional = 1,
+               .node_flags = MLX5_EXPANSION_NODE_OPTIONAL,
        },
        [MLX5_EXPANSION_NVGRE] = {
                .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH),
@@ -581,14 +596,9 @@ static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
                                                  MLX5_EXPANSION_IPV6,
                                                  MLX5_EXPANSION_ETH),
                .type = RTE_FLOW_ITEM_TYPE_MPLS,
-               .optional = 1,
+               .node_flags = MLX5_EXPANSION_NODE_OPTIONAL,
        },
        [MLX5_EXPANSION_ETH] = {
-               .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
-                                                 MLX5_EXPANSION_IPV6),
-               .type = RTE_FLOW_ITEM_TYPE_ETH,
-       },
-       [MLX5_EXPANSION_ETH_VLAN] = {
                .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VLAN),
                .type = RTE_FLOW_ITEM_TYPE_ETH,
        },
@@ -596,6 +606,7 @@ static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
                .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
                                                  MLX5_EXPANSION_IPV6),
                .type = RTE_FLOW_ITEM_TYPE_VLAN,
+               .node_flags = MLX5_EXPANSION_NODE_EXPLICIT,
        },
        [MLX5_EXPANSION_IPV4] = {
                .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4_UDP,
@@ -632,10 +643,10 @@ static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
                .type = RTE_FLOW_ITEM_TYPE_IPV6_FRAG_EXT,
        },
        [MLX5_EXPANSION_GTP] = {
-                       .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
-                                                         MLX5_EXPANSION_IPV6),
-                       .type = RTE_FLOW_ITEM_TYPE_GTP
-       }
+               .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
+                                                 MLX5_EXPANSION_IPV6),
+               .type = RTE_FLOW_ITEM_TYPE_GTP,
+       },
 };
 
 static struct rte_flow_action_handle *
@@ -3774,20 +3785,8 @@ flow_get_shared_rss_action(struct rte_eth_dev *dev,
 }
 
 static unsigned int
-find_graph_root(const struct rte_flow_item pattern[], uint32_t rss_level)
+find_graph_root(uint32_t rss_level)
 {
-       const struct rte_flow_item *item;
-       unsigned int has_vlan = 0;
-
-       for (item = pattern; item->type != RTE_FLOW_ITEM_TYPE_END; item++) {
-               if (item->type == RTE_FLOW_ITEM_TYPE_VLAN) {
-                       has_vlan = 1;
-                       break;
-               }
-       }
-       if (has_vlan)
-               return rss_level < 2 ? MLX5_EXPANSION_ROOT_ETH_VLAN :
-                                      MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN;
        return rss_level < 2 ? MLX5_EXPANSION_ROOT :
                               MLX5_EXPANSION_ROOT_OUTER;
 }
@@ -6281,7 +6280,7 @@ flow_list_create(struct rte_eth_dev *dev, enum mlx5_flow_type type,
        int indir_actions_n = MLX5_MAX_INDIRECT_ACTIONS;
        union {
                struct mlx5_flow_expand_rss buf;
-               uint8_t buffer[2048];
+               uint8_t buffer[4096];
        } expand_buffer;
        union {
                struct rte_flow_action actions[MLX5_MAX_SPLIT_ACTIONS];
@@ -6372,7 +6371,7 @@ flow_list_create(struct rte_eth_dev *dev, enum mlx5_flow_type type,
        if (rss && rss->types) {
                unsigned int graph_root;
 
-               graph_root = find_graph_root(items, rss->level);
+               graph_root = find_graph_root(rss->level);
                ret = mlx5_flow_expand_rss(buf, sizeof(expand_buffer.buffer),
                                           items, rss->types,
                                           mlx5_support_expansion, graph_root);