ip4: optimize ip4_local

"sh run" says the number of clocks for my tcp based throughput test
dropped from ~43 to ~23

Change-Id: I719439ba7fc079ad36be1432c5d7cf74e3b70d73
Signed-off-by: Florin Coras <fcoras@cisco.com>

2 files changed, 319 insertions, 253 deletions

diff --git a/src/vnet/ip/ip4_forward.c b/src/vnet/ip/ip4_forward.c
index 04cfac0d6be..6d5f07bda2c 100644
--- a/src/vnet/ip/ip4_forward.c
+++ b/src/vnet/ip/ip4_forward.c
@@ -1143,8 +1143,9 @@ VNET_FEATURE_ARC_INIT (ip4_local) =
 /* *INDENT-ON* */
 
 static inline void
-ip4_local_validate_l4 (vlib_main_t * vm, vlib_buffer_t * p, ip4_header_t * ip,
-		       u8 is_udp, u8 * error, u8 * good_tcp_udp)
+ip4_local_l4_csum_validate (vlib_main_t * vm, vlib_buffer_t * p,
+			    ip4_header_t * ip, u8 is_udp, u8 * error,
+			    u8 * good_tcp_udp)
 {
   u32 flags0;
   flags0 = ip4_tcp_udp_validate_checksum (vm, p);
@@ -1165,10 +1166,247 @@ ip4_local_validate_l4 (vlib_main_t * vm, vlib_buffer_t * p, ip4_header_t * ip,
     }
 }
 
-#define ip4_local_do_l4_check(is_tcp_udp, flags) 			\
-    (is_tcp_udp && !(flags & VNET_BUFFER_F_L4_CHECKSUM_COMPUTED \
-    || flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM \
-    || flags & VNET_BUFFER_F_OFFLOAD_UDP_CKSUM))
+#define ip4_local_csum_is_offloaded(_b)					\
+    _b->flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM				\
+	|| _b->flags & VNET_BUFFER_F_OFFLOAD_UDP_CKSUM
+
+#define ip4_local_need_csum_check(is_tcp_udp, _b) 			\
+    (is_tcp_udp && !(_b->flags & VNET_BUFFER_F_L4_CHECKSUM_COMPUTED 	\
+	|| ip4_local_csum_is_offloaded (_b)))
+
+#define ip4_local_csum_is_valid(_b)					\
+    (_b->flags & VNET_BUFFER_F_L4_CHECKSUM_CORRECT			\
+	|| (ip4_local_csum_is_offloaded (_b))) != 0
+
+static inline void
+ip4_local_check_l4_csum (vlib_main_t * vm, vlib_buffer_t * b,
+			 ip4_header_t * ih, u8 * error)
+{
+  u8 is_udp, is_tcp_udp, good_tcp_udp;
+
+  is_udp = ih->protocol == IP_PROTOCOL_UDP;
+  is_tcp_udp = is_udp || ih->protocol == IP_PROTOCOL_TCP;
+
+  if (PREDICT_FALSE (ip4_local_need_csum_check (is_tcp_udp, b)))
+    ip4_local_l4_csum_validate (vm, b, ih, is_udp, error, &good_tcp_udp);
+  else
+    good_tcp_udp = ip4_local_csum_is_valid (b);
+
+  ASSERT (IP4_ERROR_TCP_CHECKSUM + 1 == IP4_ERROR_UDP_CHECKSUM);
+  *error = (is_tcp_udp && !good_tcp_udp
+	    ? IP4_ERROR_TCP_CHECKSUM + is_udp : *error);
+}
+
+static inline void
+ip4_local_check_l4_csum_x2 (vlib_main_t * vm, vlib_buffer_t ** b,
+			    ip4_header_t ** ih, u8 * error)
+{
+  u8 is_udp[2], is_tcp_udp[2], good_tcp_udp[2];
+
+  is_udp[0] = ih[0]->protocol == IP_PROTOCOL_UDP;
+  is_udp[1] = ih[1]->protocol == IP_PROTOCOL_UDP;
+
+  is_tcp_udp[0] = is_udp[0] || ih[0]->protocol == IP_PROTOCOL_TCP;
+  is_tcp_udp[1] = is_udp[1] || ih[1]->protocol == IP_PROTOCOL_TCP;
+
+  good_tcp_udp[0] = ip4_local_csum_is_valid (b[0]);
+  good_tcp_udp[1] = ip4_local_csum_is_valid (b[1]);
+
+  if (PREDICT_FALSE (ip4_local_need_csum_check (is_tcp_udp[0], b[0])
+		     || ip4_local_need_csum_check (is_tcp_udp[1], b[1])))
+    {
+      if (is_tcp_udp[0])
+	ip4_local_l4_csum_validate (vm, b[0], ih[0], is_udp[0], &error[0],
+				    &good_tcp_udp[0]);
+      if (is_tcp_udp[1])
+	ip4_local_l4_csum_validate (vm, b[1], ih[1], is_udp[1], &error[1],
+				    &good_tcp_udp[1]);
+    }
+
+  error[0] = (is_tcp_udp[0] && !good_tcp_udp[0] ?
+	      IP4_ERROR_TCP_CHECKSUM + is_udp[0] : error[0]);
+  error[1] = (is_tcp_udp[1] && !good_tcp_udp[1] ?
+	      IP4_ERROR_TCP_CHECKSUM + is_udp[1] : error[1]);
+}
+
+static inline void
+ip4_local_set_next_and_error (vlib_node_runtime_t * error_node,
+			      vlib_buffer_t * b, u16 * next, u8 error,
+			      u8 head_of_feature_arc)
+{
+  u8 arc_index = vnet_feat_arc_ip4_local.feature_arc_index;
+  u32 next_index;
+
+  *next = error != IP4_ERROR_UNKNOWN_PROTOCOL ? IP_LOCAL_NEXT_DROP : *next;
+  b->error = error ? error_node->errors[error] : 0;
+  if (head_of_feature_arc)
+    {
+      next_index = *next;
+      if (PREDICT_TRUE (error == (u8) IP4_ERROR_UNKNOWN_PROTOCOL))
+	{
+	  vnet_feature_arc_start (arc_index,
+				  vnet_buffer (b)->sw_if_index[VLIB_RX],
+				  &next_index, b);
+	  *next = next_index;
+	}
+    }
+}
+
+typedef struct
+{
+  ip4_address_t src;
+  u32 lbi;
+  u8 error;
+} ip4_local_last_check_t;
+
+static inline void
+ip4_local_check_src (vlib_buffer_t * b, ip4_header_t * ip0,
+		     ip4_local_last_check_t * last_check, u8 * error0)
+{
+  ip4_fib_mtrie_leaf_t leaf0;
+  ip4_fib_mtrie_t *mtrie0;
+  const dpo_id_t *dpo0;
+  load_balance_t *lb0;
+  u32 lbi0;
+
+  vnet_buffer (b)->ip.fib_index =
+    vnet_buffer (b)->sw_if_index[VLIB_TX] != ~0 ?
+    vnet_buffer (b)->sw_if_index[VLIB_TX] : vnet_buffer (b)->ip.fib_index;
+
+  if (PREDICT_FALSE (last_check->src.as_u32 != ip0->src_address.as_u32))
+    {
+      mtrie0 = &ip4_fib_get (vnet_buffer (b)->ip.fib_index)->mtrie;
+      leaf0 = ip4_fib_mtrie_lookup_step_one (mtrie0, &ip0->src_address);
+      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->src_address, 2);
+      leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->src_address, 3);
+      lbi0 = ip4_fib_mtrie_leaf_get_adj_index (leaf0);
+
+      vnet_buffer (b)->ip.adj_index[VLIB_TX] = lbi0;
+      vnet_buffer (b)->ip.adj_index[VLIB_RX] = lbi0;
+
+      lb0 = load_balance_get (lbi0);
+      dpo0 = load_balance_get_bucket_i (lb0, 0);
+
+      /*
+       * Must have a route to source otherwise we drop the packet.
+       * ip4 broadcasts are accepted, e.g. to make dhcp client work
+       *
+       * The checks are:
+       *  - the source is a recieve => it's from us => bogus, do this
+       *    first since it sets a different error code.
+       *  - uRPF check for any route to source - accept if passes.
+       *  - allow packets destined to the broadcast address from unknown sources
+       */
+
+      *error0 = ((*error0 == IP4_ERROR_UNKNOWN_PROTOCOL
+		  && dpo0->dpoi_type == DPO_RECEIVE) ?
+		 IP4_ERROR_SPOOFED_LOCAL_PACKETS : *error0);
+      *error0 = ((*error0 == IP4_ERROR_UNKNOWN_PROTOCOL
+		  && !fib_urpf_check_size (lb0->lb_urpf)
+		  && ip0->dst_address.as_u32 != 0xFFFFFFFF) ?
+		 IP4_ERROR_SRC_LOOKUP_MISS : *error0);
+
+      last_check->src.as_u32 = ip0->src_address.as_u32;
+      last_check->lbi = lbi0;
+      last_check->error = *error0;
+    }
+  else
+    {
+      vnet_buffer (b)->ip.adj_index[VLIB_TX] = last_check->lbi;
+      vnet_buffer (b)->ip.adj_index[VLIB_RX] = last_check->lbi;
+      *error0 = last_check->error;
+    }
+}
+
+static inline void
+ip4_local_check_src_x2 (vlib_buffer_t ** b, ip4_header_t ** ip,
+			ip4_local_last_check_t * last_check, u8 * error)
+{
+  ip4_fib_mtrie_leaf_t leaf[2];
+  ip4_fib_mtrie_t *mtrie[2];
+  const dpo_id_t *dpo[2];
+  load_balance_t *lb[2];
+  u32 not_last_hit = 0;
+  u32 lbi[2];
+
+  not_last_hit |= ip[0]->src_address.as_u32 ^ last_check->src.as_u32;
+  not_last_hit |= ip[1]->src_address.as_u32 ^ last_check->src.as_u32;
+
+  vnet_buffer (b[0])->ip.fib_index =
+    vnet_buffer (b[0])->sw_if_index[VLIB_TX] != ~0 ?
+    vnet_buffer (b[0])->sw_if_index[VLIB_TX] :
+    vnet_buffer (b[0])->ip.fib_index;
+
+  vnet_buffer (b[1])->ip.fib_index =
+    vnet_buffer (b[1])->sw_if_index[VLIB_TX] != ~0 ?
+    vnet_buffer (b[1])->sw_if_index[VLIB_TX] :
+    vnet_buffer (b[1])->ip.fib_index;
+
+  if (PREDICT_FALSE (not_last_hit))
+    {
+      mtrie[0] = &ip4_fib_get (vnet_buffer (b[0])->ip.fib_index)->mtrie;
+      mtrie[1] = &ip4_fib_get (vnet_buffer (b[1])->ip.fib_index)->mtrie;
+
+      leaf[0] = ip4_fib_mtrie_lookup_step_one (mtrie[0], &ip[0]->src_address);
+      leaf[1] = ip4_fib_mtrie_lookup_step_one (mtrie[1], &ip[1]->src_address);
+
+      leaf[0] = ip4_fib_mtrie_lookup_step (mtrie[0], leaf[0],
+					   &ip[0]->src_address, 2);
+      leaf[1] = ip4_fib_mtrie_lookup_step (mtrie[1], leaf[1],
+					   &ip[1]->src_address, 2);
+
+      leaf[0] = ip4_fib_mtrie_lookup_step (mtrie[0], leaf[0],
+					   &ip[0]->src_address, 3);
+      leaf[1] = ip4_fib_mtrie_lookup_step (mtrie[1], leaf[1],
+					   &ip[1]->src_address, 3);
+
+      lbi[0] = ip4_fib_mtrie_leaf_get_adj_index (leaf[0]);
+      lbi[1] = ip4_fib_mtrie_leaf_get_adj_index (leaf[1]);
+
+      vnet_buffer (b[0])->ip.adj_index[VLIB_TX] = lbi[0];
+      vnet_buffer (b[0])->ip.adj_index[VLIB_RX] = lbi[0];
+
+      vnet_buffer (b[1])->ip.adj_index[VLIB_TX] = lbi[1];
+      vnet_buffer (b[1])->ip.adj_index[VLIB_RX] = lbi[1];
+
+      lb[0] = load_balance_get (lbi[0]);
+      lb[1] = load_balance_get (lbi[1]);
+
+      dpo[0] = load_balance_get_bucket_i (lb[0], 0);
+      dpo[1] = load_balance_get_bucket_i (lb[1], 0);
+
+      error[0] = ((error[0] == IP4_ERROR_UNKNOWN_PROTOCOL &&
+		   dpo[0]->dpoi_type == DPO_RECEIVE) ?
+		  IP4_ERROR_SPOOFED_LOCAL_PACKETS : error[0]);
+      error[0] = ((error[0] == IP4_ERROR_UNKNOWN_PROTOCOL &&
+		   !fib_urpf_check_size (lb[0]->lb_urpf) &&
+		   ip[0]->dst_address.as_u32 != 0xFFFFFFFF)
+		  ? IP4_ERROR_SRC_LOOKUP_MISS : error[0]);
+
+      error[1] = ((error[1] == IP4_ERROR_UNKNOWN_PROTOCOL &&
+		   dpo[1]->dpoi_type == DPO_RECEIVE) ?
+		  IP4_ERROR_SPOOFED_LOCAL_PACKETS : error[1]);
+      error[1] = ((error[1] == IP4_ERROR_UNKNOWN_PROTOCOL &&
+		   !fib_urpf_check_size (lb[1]->lb_urpf) &&
+		   ip[1]->dst_address.as_u32 != 0xFFFFFFFF)
+		  ? IP4_ERROR_SRC_LOOKUP_MISS : error[1]);
+
+      last_check->src.as_u32 = ip[1]->src_address.as_u32;
+      last_check->lbi = lbi[1];
+      last_check->error = error[1];
+    }
+  else
+    {
+      vnet_buffer (b[0])->ip.adj_index[VLIB_TX] = last_check->lbi;
+      vnet_buffer (b[0])->ip.adj_index[VLIB_RX] = last_check->lbi;
+
+      vnet_buffer (b[1])->ip.adj_index[VLIB_TX] = last_check->lbi;
+      vnet_buffer (b[1])->ip.adj_index[VLIB_RX] = last_check->lbi;
+
+      error[0] = last_check->error;
+      error[1] = last_check->error;
+    }
+}
 
 static inline uword
 ip4_local_inline (vlib_main_t * vm,
@@ -1177,288 +1415,115 @@ ip4_local_inline (vlib_main_t * vm,
 {
   ip4_main_t *im = &ip4_main;
   ip_lookup_main_t *lm = &im->lookup_main;
-  ip_local_next_t next_index;
-  u32 *from, *to_next, n_left_from, n_left_to_next;
+  u32 *from, n_left_from;
   vlib_node_runtime_t *error_node =
     vlib_node_get_runtime (vm, ip4_input_node.index);
-  u8 arc_index = vnet_feat_arc_ip4_local.feature_arc_index;
+  u16 nexts[VLIB_FRAME_SIZE], *next;
+  vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b;
+  ip4_header_t *ip[2];
+  u8 error[2];
+
+  ip4_local_last_check_t last_check = {
+    .src = {.as_u32 = 0},
+    .lbi = ~0,
+    .error = IP4_ERROR_UNKNOWN_PROTOCOL
+  };
 
   from = vlib_frame_vector_args (frame);
   n_left_from = frame->n_vectors;
-  next_index = node->cached_next_index;
 
   if (node->flags & VLIB_NODE_FLAG_TRACE)
     ip4_forward_next_trace (vm, node, frame, VLIB_TX);
 
-  while (n_left_from > 0)
-    {
-      vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
-
-      while (n_left_from >= 4 && n_left_to_next >= 2)
-	{
-	  vlib_buffer_t *p0, *p1;
-	  ip4_header_t *ip0, *ip1;
-	  ip4_fib_mtrie_t *mtrie0, *mtrie1;
-	  ip4_fib_mtrie_leaf_t leaf0, leaf1;
-	  const dpo_id_t *dpo0, *dpo1;
-	  const load_balance_t *lb0, *lb1;
-	  u32 pi0, next0, lbi0;
-	  u32 pi1, next1, lbi1;
-	  u8 error0, is_udp0, is_tcp_udp0, good_tcp_udp0, proto0;
-	  u8 error1, is_udp1, is_tcp_udp1, good_tcp_udp1, proto1;
-	  u32 sw_if_index0, sw_if_index1;
-
-	  pi0 = to_next[0] = from[0];
-	  pi1 = to_next[1] = from[1];
-	  from += 2;
-	  n_left_from -= 2;
-	  to_next += 2;
-	  n_left_to_next -= 2;
-
-	  next0 = next1 = IP_LOCAL_NEXT_DROP;
-	  error0 = error1 = IP4_ERROR_UNKNOWN_PROTOCOL;
+  vlib_get_buffers (vm, from, bufs, n_left_from);
+  b = bufs;
+  next = nexts;
 
-	  p0 = vlib_get_buffer (vm, pi0);
-	  p1 = vlib_get_buffer (vm, pi1);
+  while (n_left_from >= 6)
+    {
+      u32 is_nat, not_batch = 0;
 
-	  ip0 = vlib_buffer_get_current (p0);
-	  ip1 = vlib_buffer_get_current (p1);
+      /* Prefetch next iteration. */
+      {
+	vlib_prefetch_buffer_header (b[4], LOAD);
+	vlib_prefetch_buffer_header (b[5], LOAD);
 
-	  vnet_buffer (p0)->l3_hdr_offset = p0->current_data;
-	  vnet_buffer (p1)->l3_hdr_offset = p1->current_data;
+	CLIB_PREFETCH (b[4]->data, CLIB_CACHE_LINE_BYTES, LOAD);
+	CLIB_PREFETCH (b[5]->data, CLIB_CACHE_LINE_BYTES, LOAD);
+      }
 
-	  sw_if_index0 = vnet_buffer (p0)->sw_if_index[VLIB_RX];
-	  sw_if_index1 = vnet_buffer (p1)->sw_if_index[VLIB_RX];
+      error[0] = error[1] = IP4_ERROR_UNKNOWN_PROTOCOL;
 
-	  proto0 = ip0->protocol;
-	  proto1 = ip1->protocol;
+      ip[0] = vlib_buffer_get_current (b[0]);
+      ip[1] = vlib_buffer_get_current (b[1]);
 
-	  if (head_of_feature_arc == 0)
-	    goto skip_checks;
+      vnet_buffer (b[0])->l3_hdr_offset = b[0]->current_data;
+      vnet_buffer (b[1])->l3_hdr_offset = b[1]->current_data;
 
-	  is_udp0 = proto0 == IP_PROTOCOL_UDP;
-	  is_udp1 = proto1 == IP_PROTOCOL_UDP;
-	  is_tcp_udp0 = is_udp0 || proto0 == IP_PROTOCOL_TCP;
-	  is_tcp_udp1 = is_udp1 || proto1 == IP_PROTOCOL_TCP;
+      is_nat = b[0]->flags & VNET_BUFFER_F_IS_NATED;
+      not_batch |= is_nat ^ (b[1]->flags & VNET_BUFFER_F_IS_NATED);
 
-	  good_tcp_udp0 =
-	    (p0->flags & VNET_BUFFER_F_L4_CHECKSUM_CORRECT
-	     || (p0->flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM
-		 || p0->flags & VNET_BUFFER_F_OFFLOAD_UDP_CKSUM)) != 0;
-	  good_tcp_udp1 = (p1->flags & VNET_BUFFER_F_L4_CHECKSUM_CORRECT
-			   || (p1->flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM
-			       || p1->flags &
-			       VNET_BUFFER_F_OFFLOAD_UDP_CKSUM)) != 0;
+      if (head_of_feature_arc == 0 || (is_nat && not_batch == 0))
+	goto skip_checks;
 
-	  if (PREDICT_FALSE (ip4_local_do_l4_check (is_tcp_udp0, p0->flags)
-			     || ip4_local_do_l4_check (is_tcp_udp1,
-						       p1->flags)))
+      if (PREDICT_TRUE (not_batch == 0))
+	{
+	  ip4_local_check_l4_csum_x2 (vm, b, ip, error);
+	  ip4_local_check_src_x2 (b, ip, &last_check, error);
+	}
+      else
+	{
+	  if (!(b[0]->flags & VNET_BUFFER_F_IS_NATED))
 	    {
-	      if (is_tcp_udp0)
-		ip4_local_validate_l4 (vm, p0, ip0, is_udp0, &error0,
-				       &good_tcp_udp0);
-	      if (is_tcp_udp1)
-		ip4_local_validate_l4 (vm, p1, ip1, is_udp1, &error1,
-				       &good_tcp_udp1);
+	      ip4_local_check_l4_csum (vm, b[0], ip[0], &error[0]);
+	      ip4_local_check_src (b[0], ip[0], &last_check, &error[0]);
 	    }
-
-	  ASSERT (IP4_ERROR_TCP_CHECKSUM + 1 == IP4_ERROR_UDP_CHECKSUM);
-	  error0 = (is_tcp_udp0 && !good_tcp_udp0
-		    ? IP4_ERROR_TCP_CHECKSUM + is_udp0 : error0);
-	  error1 = (is_tcp_udp1 && !good_tcp_udp1
-		    ? IP4_ERROR_TCP_CHECKSUM + is_udp1 : error1);
-
-	  vnet_buffer (p0)->ip.fib_index =
-	    vnet_buffer (p0)->sw_if_index[VLIB_TX] != ~0 ?
-	    vnet_buffer (p0)->sw_if_index[VLIB_TX] :
-	    vnet_buffer (p0)->ip.fib_index;
-
-	  vnet_buffer (p1)->ip.fib_index =
-	    vnet_buffer (p1)->sw_if_index[VLIB_TX] != ~0 ?
-	    vnet_buffer (p1)->sw_if_index[VLIB_TX] :
-	    vnet_buffer (p1)->ip.fib_index;
-
-
-	  mtrie0 = &ip4_fib_get (vnet_buffer (p0)->ip.fib_index)->mtrie;
-	  mtrie1 = &ip4_fib_get (vnet_buffer (p1)->ip.fib_index)->mtrie;
-
-	  leaf0 = ip4_fib_mtrie_lookup_step_one (mtrie0, &ip0->src_address);
-	  leaf1 = ip4_fib_mtrie_lookup_step_one (mtrie1, &ip1->src_address);
-	  leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->src_address,
-					     2);
-	  leaf1 = ip4_fib_mtrie_lookup_step (mtrie1, leaf1, &ip1->src_address,
-					     2);
-	  leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->src_address,
-					     3);
-	  leaf1 = ip4_fib_mtrie_lookup_step (mtrie1, leaf1, &ip1->src_address,
-					     3);
-
-	  vnet_buffer (p0)->ip.adj_index[VLIB_RX] = lbi0 =
-	    ip4_fib_mtrie_leaf_get_adj_index (leaf0);
-	  vnet_buffer (p0)->ip.adj_index[VLIB_TX] = lbi0;
-
-	  vnet_buffer (p1)->ip.adj_index[VLIB_RX] = lbi1 =
-	    ip4_fib_mtrie_leaf_get_adj_index (leaf1);
-	  vnet_buffer (p1)->ip.adj_index[VLIB_TX] = lbi1;
-
-	  lb0 = load_balance_get (lbi0);
-	  lb1 = load_balance_get (lbi1);
-	  dpo0 = load_balance_get_bucket_i (lb0, 0);
-	  dpo1 = load_balance_get_bucket_i (lb1, 0);
-
-	  /*
-	   * Must have a route to source otherwise we drop the packet.
-	   * ip4 broadcasts are accepted, e.g. to make dhcp client work
-	   *
-	   * The checks are:
-	   *  - the source is a recieve => it's from us => bogus, do this
-	   *    first since it sets a different error code.
-	   *  - uRPF check for any route to source - accept if passes.
-	   *  - allow packets destined to the broadcast address from unknown sources
-	   */
-	  if (p0->flags & VNET_BUFFER_F_IS_NATED)
-	    goto skip_check0;
-
-	  error0 = ((error0 == IP4_ERROR_UNKNOWN_PROTOCOL &&
-		     dpo0->dpoi_type == DPO_RECEIVE) ?
-		    IP4_ERROR_SPOOFED_LOCAL_PACKETS : error0);
-	  error0 = ((error0 == IP4_ERROR_UNKNOWN_PROTOCOL &&
-		     !fib_urpf_check_size (lb0->lb_urpf) &&
-		     ip0->dst_address.as_u32 != 0xFFFFFFFF)
-		    ? IP4_ERROR_SRC_LOOKUP_MISS : error0);
-
-	skip_check0:
-	  if (p1->flags & VNET_BUFFER_F_IS_NATED)
-	    goto skip_checks;
-
-	  error1 = ((error1 == IP4_ERROR_UNKNOWN_PROTOCOL &&
-		     dpo1->dpoi_type == DPO_RECEIVE) ?
-		    IP4_ERROR_SPOOFED_LOCAL_PACKETS : error1);
-	  error1 = ((error1 == IP4_ERROR_UNKNOWN_PROTOCOL &&
-		     !fib_urpf_check_size (lb1->lb_urpf) &&
-		     ip1->dst_address.as_u32 != 0xFFFFFFFF)
-		    ? IP4_ERROR_SRC_LOOKUP_MISS : error1);
-
-	skip_checks:
-
-	  next0 = lm->local_next_by_ip_protocol[proto0];
-	  next1 = lm->local_next_by_ip_protocol[proto1];
-
-	  next0 =
-	    error0 != IP4_ERROR_UNKNOWN_PROTOCOL ? IP_LOCAL_NEXT_DROP : next0;
-	  next1 =
-	    error1 != IP4_ERROR_UNKNOWN_PROTOCOL ? IP_LOCAL_NEXT_DROP : next1;
-
-	  p0->error = error0 ? error_node->errors[error0] : 0;
-	  p1->error = error1 ? error_node->errors[error1] : 0;
-
-	  if (head_of_feature_arc)
+	  if (!(b[1]->flags & VNET_BUFFER_F_IS_NATED))
 	    {
-	      if (PREDICT_TRUE (error0 == (u8) IP4_ERROR_UNKNOWN_PROTOCOL))
-		vnet_feature_arc_start (arc_index, sw_if_index0, &next0, p0);
-	      if (PREDICT_TRUE (error1 == (u8) IP4_ERROR_UNKNOWN_PROTOCOL))
-		vnet_feature_arc_start (arc_index, sw_if_index1, &next1, p1);
+	      ip4_local_check_l4_csum (vm, b[1], ip[1], &error[1]);
+	      ip4_local_check_src (b[1], ip[1], &last_check, &error[1]);
 	    }
-
-	  vlib_validate_buffer_enqueue_x2 (vm, node, next_index, to_next,
-					   n_left_to_next, pi0, pi1,
-					   next0, next1);
 	}
 
-      while (n_left_from > 0 && n_left_to_next > 0)
-	{
-	  vlib_buffer_t *p0;
-	  ip4_header_t *ip0;
-	  ip4_fib_mtrie_t *mtrie0;
-	  ip4_fib_mtrie_leaf_t leaf0;
-	  u32 pi0, next0, lbi0;
-	  u8 error0, is_udp0, is_tcp_udp0, good_tcp_udp0, proto0;
-	  load_balance_t *lb0;
-	  const dpo_id_t *dpo0;
-	  u32 sw_if_index0;
-
-	  pi0 = to_next[0] = from[0];
-	  from += 1;
-	  n_left_from -= 1;
-	  to_next += 1;
-	  n_left_to_next -= 1;
+    skip_checks:
 
-	  next0 = IP_LOCAL_NEXT_DROP;
-	  error0 = IP4_ERROR_UNKNOWN_PROTOCOL;
+      next[0] = lm->local_next_by_ip_protocol[ip[0]->protocol];
+      next[1] = lm->local_next_by_ip_protocol[ip[1]->protocol];
+      ip4_local_set_next_and_error (error_node, b[0], &next[0], error[0],
+				    head_of_feature_arc);
+      ip4_local_set_next_and_error (error_node, b[1], &next[1], error[1],
+				    head_of_feature_arc);
 
-	  p0 = vlib_get_buffer (vm, pi0);
-	  ip0 = vlib_buffer_get_current (p0);
-	  vnet_buffer (p0)->l3_hdr_offset = p0->current_data;
-	  sw_if_index0 = vnet_buffer (p0)->sw_if_index[VLIB_RX];
-
-	  proto0 = ip0->protocol;
-
-	  if (head_of_feature_arc == 0 || p0->flags & VNET_BUFFER_F_IS_NATED)
-	    goto skip_check;
-
-	  is_udp0 = proto0 == IP_PROTOCOL_UDP;
-	  is_tcp_udp0 = is_udp0 || proto0 == IP_PROTOCOL_TCP;
-
-	  good_tcp_udp0 =
-	    (p0->flags & VNET_BUFFER_F_L4_CHECKSUM_CORRECT
-	     || (p0->flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM
-		 || p0->flags & VNET_BUFFER_F_OFFLOAD_UDP_CKSUM)) != 0;
-
-	  if (PREDICT_FALSE (ip4_local_do_l4_check (is_tcp_udp0, p0->flags)))
-	    {
-	      ip4_local_validate_l4 (vm, p0, ip0, is_udp0, &error0,
-				     &good_tcp_udp0);
-	    }
+      b += 2;
+      next += 2;
+      n_left_from -= 2;
+    }
 
-	  ASSERT (IP4_ERROR_TCP_CHECKSUM + 1 == IP4_ERROR_UDP_CHECKSUM);
-	  error0 = (is_tcp_udp0 && !good_tcp_udp0
-		    ? IP4_ERROR_TCP_CHECKSUM + is_udp0 : error0);
-
-	  vnet_buffer (p0)->ip.fib_index =
-	    vnet_buffer (p0)->sw_if_index[VLIB_TX] != ~0 ?
-	    vnet_buffer (p0)->sw_if_index[VLIB_TX] :
-	    vnet_buffer (p0)->ip.fib_index;
-
-	  mtrie0 = &ip4_fib_get (vnet_buffer (p0)->ip.fib_index)->mtrie;
-	  leaf0 = ip4_fib_mtrie_lookup_step_one (mtrie0, &ip0->src_address);
-	  leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->src_address,
-					     2);
-	  leaf0 = ip4_fib_mtrie_lookup_step (mtrie0, leaf0, &ip0->src_address,
-					     3);
-	  lbi0 = ip4_fib_mtrie_leaf_get_adj_index (leaf0);
-	  vnet_buffer (p0)->ip.adj_index[VLIB_TX] = lbi0;
-	  vnet_buffer (p0)->ip.adj_index[VLIB_RX] = lbi0;
+  while (n_left_from > 0)
+    {
+      error[0] = IP4_ERROR_UNKNOWN_PROTOCOL;
 
-	  lb0 = load_balance_get (lbi0);
-	  dpo0 = load_balance_get_bucket_i (lb0, 0);
+      ip[0] = vlib_buffer_get_current (b[0]);
+      vnet_buffer (b[0])->l3_hdr_offset = b[0]->current_data;
 
-	  error0 = ((error0 == IP4_ERROR_UNKNOWN_PROTOCOL &&
-		     dpo0->dpoi_type == DPO_RECEIVE) ?
-		    IP4_ERROR_SPOOFED_LOCAL_PACKETS : error0);
-	  error0 = ((error0 == IP4_ERROR_UNKNOWN_PROTOCOL &&
-		     !fib_urpf_check_size (lb0->lb_urpf) &&
-		     ip0->dst_address.as_u32 != 0xFFFFFFFF)
-		    ? IP4_ERROR_SRC_LOOKUP_MISS : error0);
+      if (head_of_feature_arc == 0 || (b[0]->flags & VNET_BUFFER_F_IS_NATED))
+	goto skip_check;
 
-	skip_check:
-	  next0 = lm->local_next_by_ip_protocol[proto0];
-	  next0 =
-	    error0 != IP4_ERROR_UNKNOWN_PROTOCOL ? IP_LOCAL_NEXT_DROP : next0;
+      ip4_local_check_l4_csum (vm, b[0], ip[0], &error[0]);
+      ip4_local_check_src (b[0], ip[0], &last_check, &error[0]);
 
-	  p0->error = error0 ? error_node->errors[error0] : 0;
+    skip_check:
 
-	  if (head_of_feature_arc)
-	    {
-	      if (PREDICT_TRUE (error0 == (u8) IP4_ERROR_UNKNOWN_PROTOCOL))
-		vnet_feature_arc_start (arc_index, sw_if_index0, &next0, p0);
-	    }
+      next[0] = lm->local_next_by_ip_protocol[ip[0]->protocol];
+      ip4_local_set_next_and_error (error_node, b[0], &next[0], error[0],
+				    head_of_feature_arc);
 
-	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
-					   n_left_to_next, pi0, next0);
-	}
-      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+      b += 1;
+      next += 1;
+      n_left_from -= 1;
     }
 
+  vlib_buffer_enqueue_to_next (vm, node, from, nexts, frame->n_vectors);
   return frame->n_vectors;
 }
 
diff --git a/src/vnet/ip/ip6_forward.c b/src/vnet/ip/ip6_forward.c
index 275bcc67b69..a66070302c0 100644
--- a/src/vnet/ip/ip6_forward.c
+++ b/src/vnet/ip/ip6_forward.c
@@ -1335,6 +1335,7 @@ ip6_local_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      error0 = (!ip6_urpf_loose_check (im, p0, ip0)
 			? IP6_ERROR_SRC_LOOKUP_MISS : error0);
 	    }
+
 	  vnet_buffer (p0)->ip.fib_index =
 	    vnet_buffer (p0)->sw_if_index[VLIB_TX] != ~0 ?
 	    vnet_buffer (p0)->sw_if_index[VLIB_TX] :