FIB2.0: Adjacency complete pull model (VPP-487)

Change the adjacency completion model to pull not push.

A complete adjacency has a rewirte string, an incomplete one does not. the re-write string for a peer comes either from a discovery protocol (i.e. ARP/ND) or can be directly derived from the link type (i.e. GRE tunnels). Which method it is, is interface type specific.
For each packet type sent on a link to a peer there is a corresponding adjacency. For example, if there is a peer 10.0.0.1 on Eth0 and we need to send to it IPv4 and MPLS packets, there will be two adjacencies; one for the IPv4 and one for the MPLS packets. The adjacencies are thus distinguished by the packets the carry, this is known as the adjacency's 'link-type'. It is not an L3 packet type, since the adjacency can have a link type of Ethernet (for L2 over GRE).
The discovery protocols are not aware of all the link types required - only the FIB is. the FIB will create adjacencies as and when they are required, and it is thus then desirable to 'pull' from the discovery protocol the re-write required. The alternative (that we have now) is that the discovery protocol pushes (i.e. creates) adjacencies for each link type - this creates more adjacencies than we need.
To pull, FIB now requests from the interface-type to 'complete' the adjacency. The interface can then delegate to the discovery protocol (on ethernet links) or directly build the re-write (i.e on GRE).

Change-Id: I61451789ae03f26b1012d8d6524007b769b6c6ee
Signed-off-by: Neale Ranns <nranns@cisco.com>

1 files changed, 15 insertions, 29 deletions

diff --git a/vnet/vnet/fib/fib_entry.c b/vnet/vnet/fib/fib_entry.c
index 5429da2983d..404f0f40da7 100644
--- a/vnet/vnet/fib/fib_entry.c
+++ b/vnet/vnet/fib/fib_entry.c
@@ -402,35 +402,21 @@ fib_entry_back_walk_notify (fib_node_t *node,
                                             fib_entry_get_index(fib_entry)));
     }
 
-    if (FIB_NODE_BW_REASON_FLAG_ADJ_UPDATE & ctx->fnbw_reason)
-    {
-        /*
-         * ADJ updates (complete<->incomplete) do not need to propagate to
-         * recursive entries.
-         * The only reason its needed as far back as here, is that the adj
-         * and the incomplete adj are a different DPO type, so the LBs need
-         * to re-stack.
-         */
-        return (FIB_NODE_BACK_WALK_CONTINUE);
-    }
-    else
-    {
-        /*
-         * all other walk types can be reclassifed to a re-evaluate to
-         * all recursive dependents.
-         * By reclassifying we ensure that should any of these walk types meet
-         * they can be merged.
-         */
-        ctx->fnbw_reason = FIB_NODE_BW_REASON_FLAG_EVALUATE;
-
-        /*
-         * propagate the backwalk further if we haven't already reached the
-         * maximum depth.
-         */
-        fib_walk_sync(FIB_NODE_TYPE_ENTRY,
-                      fib_entry_get_index(fib_entry),
-                      ctx);
-    }
+    /*
+     * all other walk types can be reclassifed to a re-evaluate to
+     * all recursive dependents.
+     * By reclassifying we ensure that should any of these walk types meet
+     * they can be merged.
+     */
+    ctx->fnbw_reason = FIB_NODE_BW_REASON_FLAG_EVALUATE;
+
+    /*
+     * propagate the backwalk further if we haven't already reached the
+     * maximum depth.
+     */
+    fib_walk_sync(FIB_NODE_TYPE_ENTRY,
+		  fib_entry_get_index(fib_entry),
+		  ctx);
 
     return (FIB_NODE_BACK_WALK_CONTINUE);
 }