Add indirect next hop support

Change-Id: I0626c2650eba7961a15b1e87a664b57bef5503a2
Signed-off-by: Damjan Marion <damarion@cisco.com>

4 files changed, 156 insertions, 40 deletions

diff --git a/vnet/vnet/ip/ip4_forward.c b/vnet/vnet/ip/ip4_forward.c
index bf101bba..63bc0ed8 100644
--- a/vnet/vnet/ip/ip4_forward.c
+++ b/vnet/vnet/ip/ip4_forward.c
@@ -332,10 +332,12 @@ ip4_add_del_route_next_hop (ip4_main_t * im,
 		}
 	      else
 		{
-		  vnm->api_errno = VNET_API_ERROR_NEXT_HOP_NOT_IN_FIB;
-		  error = clib_error_return (0, "next-hop %U/32 not in FIB",
-					     format_ip4_address, next_hop);
-		  goto done;
+		  /* Next hop is not known, so create indirect adj */
+		  ip_adjacency_t add_adj;
+		  add_adj.lookup_next_index = IP_LOOKUP_NEXT_INDIRECT;
+		  add_adj.indirect.next_hop.ip4.as_u32 = next_hop->as_u32;
+		  add_adj.explicit_fib_index = explicit_fib_index;
+		  ip_add_adjacency (lm, &add_adj, 1, &nh_adj_index);
 		}
 	    }
 	  else
@@ -624,7 +626,8 @@ always_inline uword
 ip4_lookup_inline (vlib_main_t * vm,
 		   vlib_node_runtime_t * node,
 		   vlib_frame_t * frame,
-		   int lookup_for_responses_to_locally_received_packets)
+		   int lookup_for_responses_to_locally_received_packets,
+		   int is_indirect)
 {
   ip4_main_t * im = &ip4_main;
   ip_lookup_main_t * lm = &im->lookup_main;
@@ -651,6 +654,7 @@ ip4_lookup_inline (vlib_main_t * vm,
 	  ip_adjacency_t * adj0, * adj1;
 	  ip4_fib_mtrie_t * mtrie0, * mtrie1;
 	  ip4_fib_mtrie_leaf_t leaf0, leaf1;
+	  ip4_address_t * dst_addr0, *dst_addr1;
 	  __attribute__((unused)) u32 pi0, fib_index0, adj_index0, is_tcp_udp0;
 	  __attribute__((unused)) u32 pi1, fib_index1, adj_index1, is_tcp_udp1;
           u32 flow_hash_config0, flow_hash_config1;
@@ -680,6 +684,20 @@ ip4_lookup_inline (vlib_main_t * vm,
 	  ip0 = vlib_buffer_get_current (p0);
 	  ip1 = vlib_buffer_get_current (p1);
 
+	  if (is_indirect)
+	    {
+	      ip_adjacency_t * iadj0, * iadj1;
+	      iadj0 = ip_get_adjacency (lm, vnet_buffer(p0)->ip.adj_index[VLIB_TX]);
+	      iadj1 = ip_get_adjacency (lm, vnet_buffer(p1)->ip.adj_index[VLIB_TX]);
+	      dst_addr0 = &iadj0->indirect.next_hop.ip4;
+	      dst_addr1 = &iadj1->indirect.next_hop.ip4;
+	    }
+	  else
+	    {
+	      dst_addr0 = &ip0->dst_address;
+	      dst_addr1 = &ip1->dst_address;
+	    }
+
 	  fib_index0 = vec_elt (im->fib_index_by_sw_if_index, vnet_buffer (p0)->sw_if_index[VLIB_RX]);
 	  fib_index1 = vec_elt (im->fib_index_by_sw_if_index, vnet_buffer (p1)->sw_if_index[VLIB_RX]);
           fib_index0 = (vnet_buffer(p0)->sw_if_index[VLIB_TX] == (u32)~0) ?
@@ -695,8 +713,8 @@ ip4_lookup_inline (vlib_main_t * vm,
 
 	      leaf0 = leaf1 = IP4_FIB_MTRIE_LEAF_ROOT;
 
-	      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->dst_address, 0);
-	      leaf1 = ip4_fib_mtrie_lookup_step (mtrie1, leaf1, &ip1->dst_address, 0);
+	      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, dst_addr0, 0);
+	      leaf1 = ip4_fib_mtrie_lookup_step (mtrie1, leaf1, dst_addr1, 0);
 	    }
 
 	  tcp0 = (void *) (ip0 + 1);
@@ -709,20 +727,20 @@ ip4_lookup_inline (vlib_main_t * vm,
 
 	  if (! lookup_for_responses_to_locally_received_packets)
 	    {
-	      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->dst_address, 1);
-	      leaf1 = ip4_fib_mtrie_lookup_step (mtrie1, leaf1, &ip1->dst_address, 1);
+	      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, dst_addr0, 1);
+	      leaf1 = ip4_fib_mtrie_lookup_step (mtrie1, leaf1, dst_addr1, 1);
 	    }
 
 	  if (! lookup_for_responses_to_locally_received_packets)
 	    {
-	      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->dst_address, 2);
-	      leaf1 = ip4_fib_mtrie_lookup_step (mtrie1, leaf1, &ip1->dst_address, 2);
+	      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, dst_addr0, 2);
+	      leaf1 = ip4_fib_mtrie_lookup_step (mtrie1, leaf1, dst_addr1, 2);
 	    }
 
 	  if (! lookup_for_responses_to_locally_received_packets)
 	    {
-	      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->dst_address, 3);
-	      leaf1 = ip4_fib_mtrie_lookup_step (mtrie1, leaf1, &ip1->dst_address, 3);
+	      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, dst_addr0, 3);
+	      leaf1 = ip4_fib_mtrie_lookup_step (mtrie1, leaf1, dst_addr1, 3);
 	    }
 
 	  if (lookup_for_responses_to_locally_received_packets)
@@ -741,10 +759,10 @@ ip4_lookup_inline (vlib_main_t * vm,
 	    }
 
 	  ASSERT (adj_index0 == ip4_fib_lookup_with_table (im, fib_index0,
-							   &ip0->dst_address,
+							   dst_addr0,
 							   /* no_default_route */ 0));
 	  ASSERT (adj_index1 == ip4_fib_lookup_with_table (im, fib_index1,
-							   &ip1->dst_address,
+							   dst_addr1,
 							   /* no_default_route */ 0));
 	  adj0 = ip_get_adjacency (lm, adj_index0);
 	  adj1 = ip_get_adjacency (lm, adj_index1);
@@ -840,6 +858,7 @@ ip4_lookup_inline (vlib_main_t * vm,
 	  ip_adjacency_t * adj0;
 	  ip4_fib_mtrie_t * mtrie0;
 	  ip4_fib_mtrie_leaf_t leaf0;
+	  ip4_address_t * dst_addr0;
 	  __attribute__((unused)) u32 pi0, fib_index0, adj_index0, is_tcp_udp0;
           u32 flow_hash_config0, hash_c0;
 
@@ -850,6 +869,17 @@ ip4_lookup_inline (vlib_main_t * vm,
 
 	  ip0 = vlib_buffer_get_current (p0);
 
+	  if (is_indirect)
+	    {
+	      ip_adjacency_t * iadj0;
+	      iadj0 = ip_get_adjacency (lm, vnet_buffer(p0)->ip.adj_index[VLIB_TX]);
+	      dst_addr0 = &iadj0->indirect.next_hop.ip4;
+	    }
+	  else
+	    {
+	      dst_addr0 = &ip0->dst_address;
+	    }
+
 	  fib_index0 = vec_elt (im->fib_index_by_sw_if_index, vnet_buffer (p0)->sw_if_index[VLIB_RX]);
           fib_index0 = (vnet_buffer(p0)->sw_if_index[VLIB_TX] == (u32)~0) ?
             fib_index0 : vnet_buffer(p0)->sw_if_index[VLIB_TX];
@@ -860,7 +890,7 @@ ip4_lookup_inline (vlib_main_t * vm,
 
 	      leaf0 = IP4_FIB_MTRIE_LEAF_ROOT;
 
-	      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->dst_address, 0);
+	      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, dst_addr0, 0);
 	    }
 
 	  tcp0 = (void *) (ip0 + 1);
@@ -869,13 +899,13 @@ ip4_lookup_inline (vlib_main_t * vm,
 			 || ip0->protocol == IP_PROTOCOL_UDP);
 
 	  if (! lookup_for_responses_to_locally_received_packets)
-	    leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->dst_address, 1);
+	    leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, dst_addr0, 1);
 
 	  if (! lookup_for_responses_to_locally_received_packets)
-	    leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->dst_address, 2);
+	    leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, dst_addr0, 2);
 
 	  if (! lookup_for_responses_to_locally_received_packets)
-	    leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->dst_address, 3);
+	    leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, dst_addr0, 3);
 
 	  if (lookup_for_responses_to_locally_received_packets)
 	    adj_index0 = vnet_buffer (p0)->ip.adj_index[VLIB_RX];
@@ -887,7 +917,7 @@ ip4_lookup_inline (vlib_main_t * vm,
 	    }
 
 	  ASSERT (adj_index0 == ip4_fib_lookup_with_table (im, fib_index0,
-							   &ip0->dst_address,
+							   dst_addr0,
 							   /* no_default_route */ 0));
 
 	  adj0 = ip_get_adjacency (lm, adj_index0);
@@ -945,7 +975,9 @@ ip4_lookup (vlib_main_t * vm,
 	    vlib_node_runtime_t * node,
 	    vlib_frame_t * frame)
 {
-  return ip4_lookup_inline (vm, node, frame, /* lookup_for_responses_to_locally_received_packets */ 0);
+  return ip4_lookup_inline (vm, node, frame,
+			    /* lookup_for_responses_to_locally_received_packets */ 0,
+			    /* is_indirect */ 0);
 
 }
 
@@ -1312,6 +1344,26 @@ VLIB_REGISTER_NODE (ip4_lookup_node) = {
   .next_nodes = IP4_LOOKUP_NEXT_NODES,
 };
 
+static uword
+ip4_indirect (vlib_main_t * vm,
+               vlib_node_runtime_t * node,
+               vlib_frame_t * frame)
+{
+  return ip4_lookup_inline (vm, node, frame,
+			    /* lookup_for_responses_to_locally_received_packets */ 0,
+			    /* is_indirect */ 1);
+}
+
+VLIB_REGISTER_NODE (ip4_indirect_node) = {
+  .function = ip4_indirect,
+  .name = "ip4-indirect",
+  .vector_size = sizeof (u32),
+
+  .n_next_nodes = IP_LOOKUP_N_NEXT,
+  .next_nodes = IP4_LOOKUP_NEXT_NODES,
+};
+
+
 /* Global IP4 main. */
 ip4_main_t ip4_main;
 
diff --git a/vnet/vnet/ip/ip6_forward.c b/vnet/vnet/ip/ip6_forward.c
index db74c75b..e57269a0 100644
--- a/vnet/vnet/ip/ip6_forward.c
+++ b/vnet/vnet/ip/ip6_forward.c
@@ -367,10 +367,12 @@ ip6_add_del_route_next_hop (ip6_main_t * im,
             }
             else
             {
-              vnm->api_errno = VNET_API_ERROR_UNKNOWN_DESTINATION;
-              error = clib_error_return (0, "next-hop %U/128 not in FIB",
-                                         format_ip6_address, next_hop);
-              goto done;
+	      ip_adjacency_t add_adj;
+	      add_adj.lookup_next_index = IP_LOOKUP_NEXT_INDIRECT;
+	      add_adj.indirect.next_hop.ip6.as_u64[0] = next_hop->as_u64[0];
+	      add_adj.indirect.next_hop.ip6.as_u64[1] = next_hop->as_u64[1];
+	      add_adj.explicit_fib_index = explicit_fib_index;
+	      ip_add_adjacency (lm, &add_adj, 1, &nh_adj_index);
             }
           }
           else
@@ -641,10 +643,11 @@ void ip6_delete_matching_routes (ip6_main_t * im,
   ip6_maybe_remap_adjacencies (im, table_index_or_table_id, flags);
 }
 
-static uword
-ip6_lookup (vlib_main_t * vm,
-	    vlib_node_runtime_t * node,
-	    vlib_frame_t * frame)
+always_inline uword
+ip6_lookup_inline (vlib_main_t * vm,
+		   vlib_node_runtime_t * node,
+		   vlib_frame_t * frame,
+		   int is_indirect)
 {
   ip6_main_t * im = &ip6_main;
   ip_lookup_main_t * lm = &im->lookup_main;
@@ -669,6 +672,7 @@ ip6_lookup (vlib_main_t * vm,
 	  ip_lookup_next_t next0, next1;
 	  ip6_header_t * ip0, * ip1;
 	  ip_adjacency_t * adj0, * adj1;
+	  ip6_address_t * dst_addr0, * dst_addr1;
           u32 fib_index0, fib_index1;
           u32 flow_hash_config0, flow_hash_config1;
 
@@ -694,6 +698,20 @@ ip6_lookup (vlib_main_t * vm,
 	  ip0 = vlib_buffer_get_current (p0);
 	  ip1 = vlib_buffer_get_current (p1);
 
+	  if (is_indirect)
+	    {
+	      ip_adjacency_t * iadj0, * iadj1;
+	      iadj0 = ip_get_adjacency (lm, vnet_buffer(p0)->ip.adj_index[VLIB_TX]);
+	      iadj1 = ip_get_adjacency (lm, vnet_buffer(p1)->ip.adj_index[VLIB_TX]);
+	      dst_addr0 = &iadj0->indirect.next_hop.ip6;
+	      dst_addr1 = &iadj1->indirect.next_hop.ip6;
+	    }
+	  else
+	    {
+	      dst_addr0 = &ip0->dst_address;
+	      dst_addr1 = &ip1->dst_address;
+	    }
+
 	  fib_index0 = vec_elt (im->fib_index_by_sw_if_index, vnet_buffer (p0)->sw_if_index[VLIB_RX]);
 	  fib_index1 = vec_elt (im->fib_index_by_sw_if_index, vnet_buffer (p1)->sw_if_index[VLIB_RX]);
 
@@ -702,10 +720,8 @@ ip6_lookup (vlib_main_t * vm,
           fib_index1 = (vnet_buffer(p1)->sw_if_index[VLIB_TX] == (u32)~0) ?
             fib_index1 : vnet_buffer(p1)->sw_if_index[VLIB_TX];
 
-	  adj_index0 = ip6_fib_lookup_with_table (im, fib_index0, 
-                                                  &ip0->dst_address);
-	  adj_index1 = ip6_fib_lookup_with_table (im, fib_index1, 
-                                                  &ip1->dst_address);
+	  adj_index0 = ip6_fib_lookup_with_table (im, fib_index0, dst_addr0);
+	  adj_index1 = ip6_fib_lookup_with_table (im, fib_index1, dst_addr1);
 
 	  adj0 = ip_get_adjacency (lm, adj_index0);
 	  adj1 = ip_get_adjacency (lm, adj_index1);
@@ -713,13 +729,13 @@ ip6_lookup (vlib_main_t * vm,
           if (PREDICT_FALSE (adj0->explicit_fib_index != ~0))
             {
               adj_index0 = ip6_fib_lookup_with_table 
-                (im, adj0->explicit_fib_index, &ip0->dst_address);
+                (im, adj0->explicit_fib_index, dst_addr0);
               adj0 = ip_get_adjacency (lm, adj_index0);
             }
           if (PREDICT_FALSE (adj1->explicit_fib_index != ~0))
             {
               adj_index1 = ip6_fib_lookup_with_table 
-                (im, adj1->explicit_fib_index, &ip1->dst_address);
+                (im, adj1->explicit_fib_index, dst_addr1);
               adj1 = ip_get_adjacency (lm, adj_index1);
             }
 
@@ -818,6 +834,7 @@ ip6_lookup (vlib_main_t * vm,
 	  u32 pi0, adj_index0;
 	  ip_lookup_next_t next0;
 	  ip_adjacency_t * adj0;
+	  ip6_address_t * dst_addr0;
           u32 fib_index0, flow_hash_config0;
 
 	  pi0 = from[0];
@@ -827,6 +844,17 @@ ip6_lookup (vlib_main_t * vm,
 
 	  ip0 = vlib_buffer_get_current (p0);
 
+	  if (is_indirect)
+	    {
+	      ip_adjacency_t * iadj0;
+	      iadj0 = ip_get_adjacency (lm, vnet_buffer(p0)->ip.adj_index[VLIB_TX]);
+	      dst_addr0 = &iadj0->indirect.next_hop.ip6;
+	    }
+	  else
+	    {
+	      dst_addr0 = &ip0->dst_address;
+	    }
+
 	  fib_index0 = vec_elt (im->fib_index_by_sw_if_index, vnet_buffer (p0)->sw_if_index[VLIB_RX]);
           fib_index0 = (vnet_buffer(p0)->sw_if_index[VLIB_TX] == (u32)~0) ?
             fib_index0 : vnet_buffer(p0)->sw_if_index[VLIB_TX];
@@ -834,15 +862,14 @@ ip6_lookup (vlib_main_t * vm,
           flow_hash_config0 = 
               vec_elt_at_index (im->fibs,fib_index0)->flow_hash_config;
 
-	  adj_index0 = ip6_fib_lookup_with_table (im, fib_index0, 
-                                                  &ip0->dst_address);
+	  adj_index0 = ip6_fib_lookup_with_table (im, fib_index0, dst_addr0);
 
 	  adj0 = ip_get_adjacency (lm, adj_index0);
 
           if (PREDICT_FALSE (adj0->explicit_fib_index != ~0))
             {
-              adj_index0 = ip6_fib_lookup_with_table 
-                (im, adj0->explicit_fib_index, &ip0->dst_address);
+              adj_index0 = ip6_fib_lookup_with_table
+                (im, adj0->explicit_fib_index, dst_addr0);
               adj0 = ip_get_adjacency (lm, adj_index0);
             }
 
@@ -1206,6 +1233,14 @@ ip6_sw_interface_add_del (vnet_main_t * vnm,
 
 VNET_SW_INTERFACE_ADD_DEL_FUNCTION (ip6_sw_interface_add_del);
 
+static uword
+ip6_lookup (vlib_main_t * vm,
+	    vlib_node_runtime_t * node,
+	    vlib_frame_t * frame)
+{
+  return ip6_lookup_inline (vm, node, frame, /* is_indirect */ 0);
+}
+
 VLIB_REGISTER_NODE (ip6_lookup_node) = {
   .function = ip6_lookup,
   .name = "ip6-lookup",
@@ -1215,6 +1250,24 @@ VLIB_REGISTER_NODE (ip6_lookup_node) = {
   .next_nodes = IP6_LOOKUP_NEXT_NODES,
 };
 
+static uword
+ip6_indirect (vlib_main_t * vm,
+	      vlib_node_runtime_t * node,
+	      vlib_frame_t * frame)
+{
+  return ip6_lookup_inline (vm, node, frame, /* is_indirect */ 1);
+}
+
+
+VLIB_REGISTER_NODE (ip6_indirect_node) = {
+  .function = ip6_indirect,
+  .name = "ip6-indirect",
+  .vector_size = sizeof (u32),
+
+  .n_next_nodes = IP_LOOKUP_N_NEXT,
+  .next_nodes = IP6_LOOKUP_NEXT_NODES,
+};
+
 typedef struct {
   /* Adjacency taken. */
   u32 adj_index;
diff --git a/vnet/vnet/ip/lookup.c b/vnet/vnet/ip/lookup.c
index df14a5fe..2f638f77 100644
--- a/vnet/vnet/ip/lookup.c
+++ b/vnet/vnet/ip/lookup.c
@@ -948,6 +948,7 @@ u8 * format_ip_lookup_next (u8 * s, va_list * args)
     case IP_LOOKUP_NEXT_MAP: t = "map"; break;
     case IP_LOOKUP_NEXT_MAP_T: t = "map-t"; break;
     case IP_LOOKUP_NEXT_SIXRD: t = "sixrd"; break;
+    case IP_LOOKUP_NEXT_INDIRECT: t="indirect"; break;
     case IP_LOOKUP_NEXT_REWRITE:
       break;
     }
@@ -1009,6 +1010,8 @@ u8 * format_ip_adjacency (u8 * s, va_list * args)
         case IP_LOOKUP_NEXT_CLASSIFY:
             s = format (s, " table %d", adj->classify.table_index);
 
+        case IP_LOOKUP_NEXT_INDIRECT:
+	    s = format (s, " via %U", format_ip46_address, &adj->indirect.next_hop);
 	default:
 	  break;
 	}
diff --git a/vnet/vnet/ip/lookup.h b/vnet/vnet/ip/lookup.h
index 62de2105..e9a49124 100644
--- a/vnet/vnet/ip/lookup.h
+++ b/vnet/vnet/ip/lookup.h
@@ -79,6 +79,9 @@ typedef enum {
   /* This packets needs to go to 6RD (RFC5969) */
   IP_LOOKUP_NEXT_SIXRD,
 
+  /* This packets needs to go to indirect next hop */
+  IP_LOOKUP_NEXT_INDIRECT,
+
   /* Hop-by-hop header handling */
   IP_LOOKUP_NEXT_HOP_BY_HOP,
   IP_LOOKUP_NEXT_ADD_HOP_BY_HOP,
@@ -101,6 +104,7 @@ typedef enum {
     [IP_LOOKUP_NEXT_HOP_BY_HOP] = "ip4-hop-by-hop",		\
     [IP_LOOKUP_NEXT_ADD_HOP_BY_HOP] = "ip4-add-hop-by-hop",	\
     [IP_LOOKUP_NEXT_POP_HOP_BY_HOP] = "ip4-pop-hop-by-hop",	\
+    [IP_LOOKUP_NEXT_INDIRECT] = "ip4-indirect",			\
 }
 
 #define IP6_LOOKUP_NEXT_NODES {					\
@@ -117,6 +121,7 @@ typedef enum {
     [IP_LOOKUP_NEXT_HOP_BY_HOP] = "ip6-hop-by-hop",		\
     [IP_LOOKUP_NEXT_ADD_HOP_BY_HOP] = "ip6-add-hop-by-hop",	\
     [IP_LOOKUP_NEXT_POP_HOP_BY_HOP] = "ip6-pop-hop-by-hop",	\
+    [IP_LOOKUP_NEXT_INDIRECT] = "ip6-indirect",			\
 }
 
 /* Flow hash configuration */
@@ -170,12 +175,15 @@ typedef struct {
     /* IP_LOOKUP_NEXT_ARP only */
     struct {
       ip46_address_t next_hop;
-      u32 next_adj_index_with_same_next_hop;
     } arp;
     /* IP_LOOKUP_NEXT_CLASSIFY only */
     struct {
       u16 table_index;
     } classify;
+    /* IP_LOOKUP_NEXT_INDIRECT only */
+    struct {
+        ip46_address_t next_hop;
+    } indirect;
   };
 
   STRUCT_MARK(signature_end);