l2: quad loop l2input, l2learn and l2output nodes

Change-Id: I79f76bc9791d228e2ac2867a1b857bb03d9a5279
Signed-off-by: Damjan Marion <damarion@cisco.com>

3 files changed, 265 insertions, 217 deletions

diff --git a/vnet/vnet/l2/l2_input.c b/vnet/vnet/l2/l2_input.c
index dfb7bce2f36..5d4a3761857 100644
--- a/vnet/vnet/l2/l2_input.c
+++ b/vnet/vnet/l2/l2_input.c
@@ -256,9 +256,6 @@ l2input_node_fn (vlib_main_t * vm,
   u32 n_left_from, *from, *to_next;
   l2input_next_t next_index;
   l2input_main_t *msm = &l2input_main;
-  vlib_node_t *n = vlib_get_node (vm, l2input_node.index);
-  u32 node_counter_base_index = n->error_heap_index;
-  vlib_error_main_t *em = &vm->error_main;
   u32 cpu_index = os_get_cpu_number ();
 
   from = vlib_frame_vector_args (frame);
@@ -272,40 +269,32 @@ l2input_node_fn (vlib_main_t * vm,
       /* get space to enqueue frame to graph node "next_index" */
       vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
 
-      while (n_left_from >= 6 && n_left_to_next >= 2)
+      while (n_left_from >= 8 && n_left_to_next >= 4)
 	{
-	  u32 bi0, bi1;
-	  vlib_buffer_t *b0, *b1;
-	  u32 next0, next1;
-	  u32 sw_if_index0, sw_if_index1;
+	  u32 bi0, bi1, bi2, bi3;
+	  vlib_buffer_t *b0, *b1, *b2, *b3;
+	  u32 next0, next1, next2, next3;
+	  u32 sw_if_index0, sw_if_index1, sw_if_index2, sw_if_index3;
 
 	  /* Prefetch next iteration. */
 	  {
-	    vlib_buffer_t *p2, *p3, *p4, *p5;
-	    u32 sw_if_index2, sw_if_index3;
+	    vlib_buffer_t *p4, *p5, *p6, *p7;
 
-	    p2 = vlib_get_buffer (vm, from[2]);
-	    p3 = vlib_get_buffer (vm, from[3]);
 	    p4 = vlib_get_buffer (vm, from[4]);
 	    p5 = vlib_get_buffer (vm, from[5]);
+	    p6 = vlib_get_buffer (vm, from[6]);
+	    p7 = vlib_get_buffer (vm, from[7]);
 
 	    /* Prefetch the buffer header and packet for the N+2 loop iteration */
 	    vlib_prefetch_buffer_header (p4, LOAD);
 	    vlib_prefetch_buffer_header (p5, LOAD);
+	    vlib_prefetch_buffer_header (p6, LOAD);
+	    vlib_prefetch_buffer_header (p7, LOAD);
 
 	    CLIB_PREFETCH (p4->data, CLIB_CACHE_LINE_BYTES, STORE);
 	    CLIB_PREFETCH (p5->data, CLIB_CACHE_LINE_BYTES, STORE);
-
-	    /*
-	     * Prefetch the input config for the N+1 loop iteration
-	     * This depends on the buffer header above
-	     */
-	    sw_if_index2 = vnet_buffer (p2)->sw_if_index[VLIB_RX];
-	    sw_if_index3 = vnet_buffer (p3)->sw_if_index[VLIB_RX];
-	    CLIB_PREFETCH (&msm->configs[sw_if_index2], CLIB_CACHE_LINE_BYTES,
-			   LOAD);
-	    CLIB_PREFETCH (&msm->configs[sw_if_index3], CLIB_CACHE_LINE_BYTES,
-			   LOAD);
+	    CLIB_PREFETCH (p6->data, CLIB_CACHE_LINE_BYTES, STORE);
+	    CLIB_PREFETCH (p7->data, CLIB_CACHE_LINE_BYTES, STORE);
 
 	    /*
 	     * Don't bother prefetching the bridge-domain config (which
@@ -319,19 +308,25 @@ l2input_node_fn (vlib_main_t * vm,
 	  /* bi is "buffer index", b is pointer to the buffer */
 	  to_next[0] = bi0 = from[0];
 	  to_next[1] = bi1 = from[1];
-	  from += 2;
-	  to_next += 2;
-	  n_left_from -= 2;
-	  n_left_to_next -= 2;
+	  to_next[2] = bi2 = from[2];
+	  to_next[3] = bi3 = from[3];
+	  from += 4;
+	  to_next += 4;
+	  n_left_from -= 4;
+	  n_left_to_next -= 4;
 
 	  b0 = vlib_get_buffer (vm, bi0);
 	  b1 = vlib_get_buffer (vm, bi1);
+	  b2 = vlib_get_buffer (vm, bi2);
+	  b3 = vlib_get_buffer (vm, bi3);
 
 	  if (PREDICT_FALSE ((node->flags & VLIB_NODE_FLAG_TRACE)))
 	    {
 	      /* RX interface handles */
 	      sw_if_index0 = vnet_buffer (b0)->sw_if_index[VLIB_RX];
 	      sw_if_index1 = vnet_buffer (b1)->sw_if_index[VLIB_RX];
+	      sw_if_index2 = vnet_buffer (b2)->sw_if_index[VLIB_RX];
+	      sw_if_index3 = vnet_buffer (b3)->sw_if_index[VLIB_RX];
 
 	      if (b0->flags & VLIB_BUFFER_IS_TRACED)
 		{
@@ -351,19 +346,40 @@ l2input_node_fn (vlib_main_t * vm,
 		  clib_memcpy (t->src, h1->src_address, 6);
 		  clib_memcpy (t->dst, h1->dst_address, 6);
 		}
+	      if (b2->flags & VLIB_BUFFER_IS_TRACED)
+		{
+		  ethernet_header_t *h2 = vlib_buffer_get_current (b2);
+		  l2input_trace_t *t =
+		    vlib_add_trace (vm, node, b2, sizeof (*t));
+		  t->sw_if_index = sw_if_index2;
+		  clib_memcpy (t->src, h2->src_address, 6);
+		  clib_memcpy (t->dst, h2->dst_address, 6);
+		}
+	      if (b3->flags & VLIB_BUFFER_IS_TRACED)
+		{
+		  ethernet_header_t *h3 = vlib_buffer_get_current (b3);
+		  l2input_trace_t *t =
+		    vlib_add_trace (vm, node, b3, sizeof (*t));
+		  t->sw_if_index = sw_if_index3;
+		  clib_memcpy (t->src, h3->src_address, 6);
+		  clib_memcpy (t->dst, h3->dst_address, 6);
+		}
 	    }
 
-	  em->counters[node_counter_base_index + L2INPUT_ERROR_L2INPUT] += 2;
+	  vlib_node_increment_counter (vm, l2input_node.index,
+				       L2INPUT_ERROR_L2INPUT, 4);
 
 	  classify_and_dispatch (vm, node, cpu_index, msm, b0, &next0);
-
 	  classify_and_dispatch (vm, node, cpu_index, msm, b1, &next1);
+	  classify_and_dispatch (vm, node, cpu_index, msm, b2, &next2);
+	  classify_and_dispatch (vm, node, cpu_index, msm, b3, &next3);
 
 	  /* verify speculative enqueues, maybe switch current next frame */
 	  /* if next0==next1==next_index then nothing special needs to be done */
-	  vlib_validate_buffer_enqueue_x2 (vm, node, next_index,
+	  vlib_validate_buffer_enqueue_x4 (vm, node, next_index,
 					   to_next, n_left_to_next,
-					   bi0, bi1, next0, next1);
+					   bi0, bi1, bi2, bi3,
+					   next0, next1, next2, next3);
 	}
 
       while (n_left_from > 0 && n_left_to_next > 0)
@@ -394,7 +410,8 @@ l2input_node_fn (vlib_main_t * vm,
 	      clib_memcpy (t->dst, h0->dst_address, 6);
 	    }
 
-	  em->counters[node_counter_base_index + L2INPUT_ERROR_L2INPUT] += 1;
+	  vlib_node_increment_counter (vm, l2input_node.index,
+				       L2INPUT_ERROR_L2INPUT, 1);
 
 	  classify_and_dispatch (vm, node, cpu_index, msm, b0, &next0);
 
diff --git a/vnet/vnet/l2/l2_learn.c b/vnet/vnet/l2/l2_learn.c
index e34fe7a83c8..9feb7289466 100644
--- a/vnet/vnet/l2/l2_learn.c
+++ b/vnet/vnet/l2/l2_learn.c
@@ -258,51 +258,65 @@ l2learn_node_fn (vlib_main_t * vm,
       /* get space to enqueue frame to graph node "next_index" */
       vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
 
-      while (n_left_from >= 4 && n_left_to_next >= 2)
+      while (n_left_from >= 8 && n_left_to_next >= 4)
 	{
-	  u32 bi0, bi1;
-	  vlib_buffer_t *b0, *b1;
-	  u32 next0, next1;
-	  u32 sw_if_index0, sw_if_index1;
-	  ethernet_header_t *h0, *h1;
-	  l2fib_entry_key_t key0, key1;
-	  l2fib_entry_result_t result0, result1;
-	  u32 bucket0, bucket1;
+	  u32 bi0, bi1, bi2, bi3;
+	  vlib_buffer_t *b0, *b1, *b2, *b3;
+	  u32 next0, next1, next2, next3;
+	  u32 sw_if_index0, sw_if_index1, sw_if_index2, sw_if_index3;
+	  ethernet_header_t *h0, *h1, *h2, *h3;
+	  l2fib_entry_key_t key0, key1, key2, key3;
+	  l2fib_entry_result_t result0, result1, result2, result3;
+	  u32 bucket0, bucket1, bucket2, bucket3;
 
 	  /* Prefetch next iteration. */
 	  {
-	    vlib_buffer_t *p2, *p3;
-
-	    p2 = vlib_get_buffer (vm, from[2]);
-	    p3 = vlib_get_buffer (vm, from[3]);
-
-	    vlib_prefetch_buffer_header (p2, LOAD);
-	    vlib_prefetch_buffer_header (p3, LOAD);
-
-	    CLIB_PREFETCH (p2->data, CLIB_CACHE_LINE_BYTES, STORE);
-	    CLIB_PREFETCH (p3->data, CLIB_CACHE_LINE_BYTES, STORE);
+	    vlib_buffer_t *p4, *p5, *p6, *p7;;
+
+	    p4 = vlib_get_buffer (vm, from[4]);
+	    p5 = vlib_get_buffer (vm, from[5]);
+	    p6 = vlib_get_buffer (vm, from[6]);
+	    p7 = vlib_get_buffer (vm, from[7]);
+
+	    vlib_prefetch_buffer_header (p4, LOAD);
+	    vlib_prefetch_buffer_header (p5, LOAD);
+	    vlib_prefetch_buffer_header (p6, LOAD);
+	    vlib_prefetch_buffer_header (p7, LOAD);
+
+	    CLIB_PREFETCH (p4->data, CLIB_CACHE_LINE_BYTES, STORE);
+	    CLIB_PREFETCH (p5->data, CLIB_CACHE_LINE_BYTES, STORE);
+	    CLIB_PREFETCH (p6->data, CLIB_CACHE_LINE_BYTES, STORE);
+	    CLIB_PREFETCH (p7->data, CLIB_CACHE_LINE_BYTES, STORE);
 	  }
 
 	  /* speculatively enqueue b0 and b1 to the current next frame */
 	  /* bi is "buffer index", b is pointer to the buffer */
 	  to_next[0] = bi0 = from[0];
 	  to_next[1] = bi1 = from[1];
-	  from += 2;
-	  to_next += 2;
-	  n_left_from -= 2;
-	  n_left_to_next -= 2;
+	  to_next[2] = bi2 = from[2];
+	  to_next[3] = bi3 = from[3];
+	  from += 4;
+	  to_next += 4;
+	  n_left_from -= 4;
+	  n_left_to_next -= 4;
 
 	  b0 = vlib_get_buffer (vm, bi0);
 	  b1 = vlib_get_buffer (vm, bi1);
+	  b2 = vlib_get_buffer (vm, bi2);
+	  b3 = vlib_get_buffer (vm, bi3);
 
 	  /* RX interface handles */
 	  sw_if_index0 = vnet_buffer (b0)->sw_if_index[VLIB_RX];
 	  sw_if_index1 = vnet_buffer (b1)->sw_if_index[VLIB_RX];
+	  sw_if_index2 = vnet_buffer (b2)->sw_if_index[VLIB_RX];
+	  sw_if_index3 = vnet_buffer (b3)->sw_if_index[VLIB_RX];
 
-	  /* Process 2 x pkts */
+	  /* Process 4 x pkts */
 
 	  h0 = vlib_buffer_get_current (b0);
 	  h1 = vlib_buffer_get_current (b1);
+	  h2 = vlib_buffer_get_current (b2);
+	  h3 = vlib_buffer_get_current (b3);
 
 	  if (PREDICT_FALSE ((node->flags & VLIB_NODE_FLAG_TRACE)))
 	    {
@@ -324,18 +338,42 @@ l2learn_node_fn (vlib_main_t * vm,
 		  clib_memcpy (t->src, h1->src_address, 6);
 		  clib_memcpy (t->dst, h1->dst_address, 6);
 		}
+	      if (b2->flags & VLIB_BUFFER_IS_TRACED)
+		{
+		  l2learn_trace_t *t =
+		    vlib_add_trace (vm, node, b2, sizeof (*t));
+		  t->sw_if_index = sw_if_index2;
+		  t->bd_index = vnet_buffer (b2)->l2.bd_index;
+		  clib_memcpy (t->src, h2->src_address, 6);
+		  clib_memcpy (t->dst, h2->dst_address, 6);
+		}
+	      if (b3->flags & VLIB_BUFFER_IS_TRACED)
+		{
+		  l2learn_trace_t *t =
+		    vlib_add_trace (vm, node, b3, sizeof (*t));
+		  t->sw_if_index = sw_if_index3;
+		  t->bd_index = vnet_buffer (b3)->l2.bd_index;
+		  clib_memcpy (t->src, h3->src_address, 6);
+		  clib_memcpy (t->dst, h3->dst_address, 6);
+		}
 	    }
 
-	  /* process 2 pkts */
-	  em->counters[node_counter_base_index + L2LEARN_ERROR_L2LEARN] += 2;
+	  /* process 4 pkts */
+	  vlib_node_increment_counter (vm, l2learn_node.index,
+				       L2LEARN_ERROR_L2LEARN, 4);
 
-	  l2fib_lookup_2 (msm->mac_table, &cached_key, &cached_result,
+	  l2fib_lookup_4 (msm->mac_table, &cached_key, &cached_result,
 			  h0->src_address,
 			  h1->src_address,
+			  h2->src_address,
+			  h3->src_address,
 			  vnet_buffer (b0)->l2.bd_index,
 			  vnet_buffer (b1)->l2.bd_index,
-			  &key0,
-			  &key1, &bucket0, &bucket1, &result0, &result1);
+			  vnet_buffer (b2)->l2.bd_index,
+			  vnet_buffer (b3)->l2.bd_index,
+			  &key0, &key1, &key2, &key3,
+			  &bucket0, &bucket1, &bucket2, &bucket3,
+			  &result0, &result1, &result2, &result3);
 
 	  l2learn_process (node, msm, &em->counters[node_counter_base_index],
 			   b0, sw_if_index0, &key0, &cached_key,
@@ -345,11 +383,20 @@ l2learn_node_fn (vlib_main_t * vm,
 			   b1, sw_if_index1, &key1, &cached_key,
 			   &bucket1, &result1, &next1);
 
+	  l2learn_process (node, msm, &em->counters[node_counter_base_index],
+			   b2, sw_if_index2, &key2, &cached_key,
+			   &bucket2, &result2, &next2);
+
+	  l2learn_process (node, msm, &em->counters[node_counter_base_index],
+			   b3, sw_if_index3, &key3, &cached_key,
+			   &bucket3, &result3, &next3);
+
 	  /* verify speculative enqueues, maybe switch current next frame */
 	  /* if next0==next1==next_index then nothing special needs to be done */
-	  vlib_validate_buffer_enqueue_x2 (vm, node, next_index,
+	  vlib_validate_buffer_enqueue_x4 (vm, node, next_index,
 					   to_next, n_left_to_next,
-					   bi0, bi1, next0, next1);
+					   bi0, bi1, bi2, bi3,
+					   next0, next1, next2, next3);
 	}
 
       while (n_left_from > 0 && n_left_to_next > 0)
@@ -388,7 +435,9 @@ l2learn_node_fn (vlib_main_t * vm,
 	    }
 
 	  /* process 1 pkt */
-	  em->counters[node_counter_base_index + L2LEARN_ERROR_L2LEARN] += 1;
+	  vlib_node_increment_counter (vm, l2learn_node.index,
+				       L2LEARN_ERROR_L2LEARN, 1);
+
 
 	  l2fib_lookup_1 (msm->mac_table, &cached_key, &cached_result,
 			  h0->src_address, vnet_buffer (b0)->l2.bd_index,
diff --git a/vnet/vnet/l2/l2_output.c b/vnet/vnet/l2/l2_output.c
index 22ce663b899..953fcb0222a 100644
--- a/vnet/vnet/l2/l2_output.c
+++ b/vnet/vnet/l2/l2_output.c
@@ -92,6 +92,51 @@ split_horizon_violation (u8 shg1, u8 shg2)
     }
 }
 
+static_always_inline void
+l2output_vtr (vlib_node_runtime_t * node, l2_output_config_t * config,
+	      u32 feature_bitmap, vlib_buffer_t * b, u32 * next)
+{
+  if (PREDICT_FALSE (config->out_vtr_flag))
+    {
+      /* Perform pre-vtr EFP filter check if configured */
+      if (config->output_vtr.push_and_pop_bytes)
+	{
+	  /*
+	   * Perform output vlan tag rewrite and the pre-vtr EFP filter check.
+	   * The EFP Filter only needs to be run if there is an output VTR
+	   * configured. The flag for the post-vtr EFP Filter node is used
+	   * to trigger the pre-vtr check as well.
+	   */
+	  u32 failed1 = (feature_bitmap & L2OUTPUT_FEAT_EFP_FILTER)
+	    && (l2_efp_filter_process (b, &(config->input_vtr)));
+	  u32 failed2 = l2_vtr_process (b, &(config->output_vtr));
+
+	  if (PREDICT_FALSE (failed1 | failed2))
+	    {
+	      *next = L2OUTPUT_NEXT_DROP;
+	      if (failed2)
+		{
+		  b->error = node->errors[L2OUTPUT_ERROR_VTR_DROP];
+		}
+	      if (failed1)
+		{
+		  b->error = node->errors[L2OUTPUT_ERROR_EFP_DROP];
+		}
+	    }
+	}
+      // perform the PBB rewrite
+      else if (config->output_pbb_vtr.push_and_pop_bytes)
+	{
+	  u32 failed = l2_pbb_process (b, &(config->output_pbb_vtr));
+	  if (PREDICT_FALSE (failed))
+	    {
+	      *next = L2OUTPUT_NEXT_DROP;
+	      b->error = node->errors[L2OUTPUT_ERROR_VTR_DROP];
+	    }
+	}
+    }
+}
+
 
 static vlib_node_registration_t l2output_node;
 
@@ -102,9 +147,6 @@ l2output_node_fn (vlib_main_t * vm,
   u32 n_left_from, *from, *to_next;
   l2output_next_t next_index;
   l2output_main_t *msm = &l2output_main;
-  vlib_node_t *n = vlib_get_node (vm, l2output_node.index);
-  u32 node_counter_base_index = n->error_heap_index;
-  vlib_error_main_t *em = &vm->error_main;
   u32 cached_sw_if_index;
   u32 cached_next_index;
 
@@ -123,64 +165,61 @@ l2output_node_fn (vlib_main_t * vm,
       /* get space to enqueue frame to graph node "next_index" */
       vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
 
-      while (n_left_from >= 6 && n_left_to_next >= 2)
+      while (n_left_from >= 8 && n_left_to_next >= 4)
 	{
-	  u32 bi0, bi1;
-	  vlib_buffer_t *b0, *b1;
-	  u32 next0, next1;
-	  u32 sw_if_index0, sw_if_index1;
-	  ethernet_header_t *h0, *h1;
-	  l2_output_config_t *config0, *config1;
+	  u32 bi0, bi1, bi2, bi3;
+	  vlib_buffer_t *b0, *b1, *b2, *b3;
+	  u32 next0, next1, next2, next3;
+	  u32 sw_if_index0, sw_if_index1, sw_if_index2, sw_if_index3;
+	  ethernet_header_t *h0, *h1, *h2, *h3;
+	  l2_output_config_t *config0, *config1, *config2, *config3;
 	  u32 feature_bitmap0, feature_bitmap1;
+	  u32 feature_bitmap2, feature_bitmap3;
 
 	  /* Prefetch next iteration. */
 	  {
-	    vlib_buffer_t *p2, *p3, *p4, *p5;
-	    u32 sw_if_index2, sw_if_index3;
+	    vlib_buffer_t *p4, *p5, *p6, *p7;
 
-	    p2 = vlib_get_buffer (vm, from[2]);
-	    p3 = vlib_get_buffer (vm, from[3]);
 	    p4 = vlib_get_buffer (vm, from[4]);
 	    p5 = vlib_get_buffer (vm, from[5]);
+	    p6 = vlib_get_buffer (vm, from[6]);
+	    p7 = vlib_get_buffer (vm, from[7]);
 
 	    /* Prefetch the buffer header for the N+2 loop iteration */
 	    vlib_prefetch_buffer_header (p4, LOAD);
 	    vlib_prefetch_buffer_header (p5, LOAD);
-	    /*
-	     * Note: no need to prefetch packet data.
-	     * This node doesn't reference it.
-	     *
-	     * Prefetch the input config for the N+1 loop iteration
-	     * This depends on the buffer header above
-	     */
-	    sw_if_index2 = vnet_buffer (p2)->sw_if_index[VLIB_TX];
-	    sw_if_index3 = vnet_buffer (p3)->sw_if_index[VLIB_TX];
-	    CLIB_PREFETCH (&msm->configs[sw_if_index2], CLIB_CACHE_LINE_BYTES,
-			   LOAD);
-	    CLIB_PREFETCH (&msm->configs[sw_if_index3], CLIB_CACHE_LINE_BYTES,
-			   LOAD);
+	    vlib_prefetch_buffer_header (p6, LOAD);
+	    vlib_prefetch_buffer_header (p7, LOAD);
 	  }
 
 	  /* speculatively enqueue b0 and b1 to the current next frame */
 	  /* bi is "buffer index", b is pointer to the buffer */
 	  to_next[0] = bi0 = from[0];
 	  to_next[1] = bi1 = from[1];
-	  from += 2;
-	  to_next += 2;
-	  n_left_from -= 2;
-	  n_left_to_next -= 2;
+	  to_next[2] = bi2 = from[2];
+	  to_next[3] = bi3 = from[3];
+	  from += 4;
+	  to_next += 4;
+	  n_left_from -= 4;
+	  n_left_to_next -= 4;
 
 	  b0 = vlib_get_buffer (vm, bi0);
 	  b1 = vlib_get_buffer (vm, bi1);
+	  b2 = vlib_get_buffer (vm, bi2);
+	  b3 = vlib_get_buffer (vm, bi3);
 
 	  /* TX interface handles */
 	  sw_if_index0 = vnet_buffer (b0)->sw_if_index[VLIB_TX];
 	  sw_if_index1 = vnet_buffer (b1)->sw_if_index[VLIB_TX];
+	  sw_if_index2 = vnet_buffer (b2)->sw_if_index[VLIB_TX];
+	  sw_if_index3 = vnet_buffer (b3)->sw_if_index[VLIB_TX];
 
 	  if (PREDICT_FALSE ((node->flags & VLIB_NODE_FLAG_TRACE)))
 	    {
 	      h0 = vlib_buffer_get_current (b0);
 	      h1 = vlib_buffer_get_current (b1);
+	      h2 = vlib_buffer_get_current (b2);
+	      h3 = vlib_buffer_get_current (b3);
 	      if (b0->flags & VLIB_BUFFER_IS_TRACED)
 		{
 		  l2output_trace_t *t =
@@ -197,14 +236,32 @@ l2output_node_fn (vlib_main_t * vm,
 		  clib_memcpy (t->src, h1->src_address, 6);
 		  clib_memcpy (t->dst, h1->dst_address, 6);
 		}
+	      if (b2->flags & VLIB_BUFFER_IS_TRACED)
+		{
+		  l2output_trace_t *t =
+		    vlib_add_trace (vm, node, b2, sizeof (*t));
+		  t->sw_if_index = sw_if_index2;
+		  clib_memcpy (t->src, h2->src_address, 6);
+		  clib_memcpy (t->dst, h2->dst_address, 6);
+		}
+	      if (b3->flags & VLIB_BUFFER_IS_TRACED)
+		{
+		  l2output_trace_t *t =
+		    vlib_add_trace (vm, node, b3, sizeof (*t));
+		  t->sw_if_index = sw_if_index3;
+		  clib_memcpy (t->src, h3->src_address, 6);
+		  clib_memcpy (t->dst, h3->dst_address, 6);
+		}
 	    }
 
-	  em->counters[node_counter_base_index + L2OUTPUT_ERROR_L2OUTPUT] +=
-	    2;
+	  vlib_node_increment_counter (vm, l2output_node.index,
+				       L2OUTPUT_ERROR_L2OUTPUT, 4);
 
 	  /* Get config for the output interface */
 	  config0 = vec_elt_at_index (msm->configs, sw_if_index0);
 	  config1 = vec_elt_at_index (msm->configs, sw_if_index1);
+	  config2 = vec_elt_at_index (msm->configs, sw_if_index2);
+	  config3 = vec_elt_at_index (msm->configs, sw_if_index3);
 
 	  /*
 	   * Get features from the config
@@ -212,6 +269,8 @@ l2output_node_fn (vlib_main_t * vm,
 	   */
 	  feature_bitmap0 = config0->feature_bitmap;
 	  feature_bitmap1 = config1->feature_bitmap;
+	  feature_bitmap2 = config2->feature_bitmap;
+	  feature_bitmap3 = config3->feature_bitmap;
 
 	  /* Determine next node */
 	  l2_output_dispatch (msm->vlib_main,
@@ -232,86 +291,35 @@ l2output_node_fn (vlib_main_t * vm,
 			      &msm->next_nodes,
 			      b1, sw_if_index1, feature_bitmap1, &next1);
 
-	  if (PREDICT_FALSE (config0->out_vtr_flag))
-	    {
-	      /* Perform pre-vtr EFP filter check if configured */
-	      if (config0->output_vtr.push_and_pop_bytes)
-		{
-		  /*
-		   * Perform output vlan tag rewrite and the pre-vtr EFP filter check.
-		   * The EFP Filter only needs to be run if there is an output VTR
-		   * configured. The flag for the post-vtr EFP Filter node is used
-		   * to trigger the pre-vtr check as well.
-		   */
-		  u32 failed1 = (feature_bitmap0 & L2OUTPUT_FEAT_EFP_FILTER)
-		    && (l2_efp_filter_process (b0, &(config0->input_vtr)));
-		  u32 failed2 = l2_vtr_process (b0, &(config0->output_vtr));
-
-		  if (PREDICT_FALSE (failed1 | failed2))
-		    {
-		      next0 = L2OUTPUT_NEXT_DROP;
-		      if (failed2)
-			{
-			  b0->error = node->errors[L2OUTPUT_ERROR_VTR_DROP];
-			}
-		      if (failed1)
-			{
-			  b0->error = node->errors[L2OUTPUT_ERROR_EFP_DROP];
-			}
-		    }
-		}
-	      // perform the PBB rewrite
-	      else if (config0->output_pbb_vtr.push_and_pop_bytes)
-		{
-		  u32 failed =
-		    l2_pbb_process (b0, &(config0->output_pbb_vtr));
-		  if (PREDICT_FALSE (failed))
-		    {
-		      next0 = L2OUTPUT_NEXT_DROP;
-		      b0->error = node->errors[L2OUTPUT_ERROR_VTR_DROP];
-		    }
-		}
-	    }
-	  if (PREDICT_FALSE (config1->out_vtr_flag))
-	    {
-	      /* Perform pre-vtr EFP filter check if configured */
-	      if (config1->output_vtr.push_and_pop_bytes)
-		{
-		  u32 failed1 = (feature_bitmap1 & L2OUTPUT_FEAT_EFP_FILTER)
-		    && (l2_efp_filter_process (b1, &(config1->input_vtr)));
-		  u32 failed2 = l2_vtr_process (b1, &(config1->output_vtr));
-
-		  if (PREDICT_FALSE (failed1 | failed2))
-		    {
-		      next1 = L2OUTPUT_NEXT_DROP;
-		      if (failed2)
-			{
-			  b1->error = node->errors[L2OUTPUT_ERROR_VTR_DROP];
-			}
-		      if (failed1)
-			{
-			  b1->error = node->errors[L2OUTPUT_ERROR_EFP_DROP];
-			}
-		    }
-		}
-	      // perform the PBB rewrite
-	      else if (config1->output_pbb_vtr.push_and_pop_bytes)
-		{
-		  u32 failed =
-		    l2_pbb_process (b0, &(config1->output_pbb_vtr));
-		  if (PREDICT_FALSE (failed))
-		    {
-		      next1 = L2OUTPUT_NEXT_DROP;
-		      b1->error = node->errors[L2OUTPUT_ERROR_VTR_DROP];
-		    }
-		}
-	    }
+	  l2_output_dispatch (msm->vlib_main,
+			      msm->vnet_main,
+			      node,
+			      l2output_node.index,
+			      &cached_sw_if_index,
+			      &cached_next_index,
+			      &msm->next_nodes,
+			      b2, sw_if_index2, feature_bitmap2, &next2);
+
+	  l2_output_dispatch (msm->vlib_main,
+			      msm->vnet_main,
+			      node,
+			      l2output_node.index,
+			      &cached_sw_if_index,
+			      &cached_next_index,
+			      &msm->next_nodes,
+			      b3, sw_if_index3, feature_bitmap3, &next3);
+
+	  l2output_vtr (node, config0, feature_bitmap0, b0, &next0);
+	  l2output_vtr (node, config1, feature_bitmap1, b1, &next1);
+	  l2output_vtr (node, config2, feature_bitmap2, b2, &next2);
+	  l2output_vtr (node, config3, feature_bitmap3, b3, &next3);
 
 	  /*
 	   * Perform the split horizon check
 	   * The check can only fail for non-zero shg's
 	   */
-	  if (PREDICT_FALSE (config0->shg + config1->shg))
+	  if (PREDICT_FALSE (config0->shg + config1->shg +
+			     config2->shg + config3->shg))
 	    {
 	      /* one of the checks might fail, check both */
 	      if (split_horizon_violation
@@ -326,13 +334,26 @@ l2output_node_fn (vlib_main_t * vm,
 		  next1 = L2OUTPUT_NEXT_DROP;
 		  b1->error = node->errors[L2OUTPUT_ERROR_SHG_DROP];
 		}
+	      if (split_horizon_violation
+		  (config2->shg, vnet_buffer (b2)->l2.shg))
+		{
+		  next2 = L2OUTPUT_NEXT_DROP;
+		  b2->error = node->errors[L2OUTPUT_ERROR_SHG_DROP];
+		}
+	      if (split_horizon_violation
+		  (config3->shg, vnet_buffer (b3)->l2.shg))
+		{
+		  next3 = L2OUTPUT_NEXT_DROP;
+		  b3->error = node->errors[L2OUTPUT_ERROR_SHG_DROP];
+		}
 	    }
 
 	  /* verify speculative enqueues, maybe switch current next frame */
 	  /* if next0==next1==next_index then nothing special needs to be done */
-	  vlib_validate_buffer_enqueue_x2 (vm, node, next_index,
+	  vlib_validate_buffer_enqueue_x4 (vm, node, next_index,
 					   to_next, n_left_to_next,
-					   bi0, bi1, next0, next1);
+					   bi0, bi1, bi2, bi3,
+					   next0, next1, next2, next3);
 	}
 
       while (n_left_from > 0 && n_left_to_next > 0)
@@ -368,8 +389,8 @@ l2output_node_fn (vlib_main_t * vm,
 	      clib_memcpy (t->dst, h0->dst_address, 6);
 	    }
 
-	  em->counters[node_counter_base_index +
-		       L2OUTPUT_ERROR_L2OUTPUT] += 1;
+	  vlib_node_increment_counter (vm, l2output_node.index,
+				       L2OUTPUT_ERROR_L2OUTPUT, 1);
 
 	  /* Get config for the output interface */
 	  config0 = vec_elt_at_index (msm->configs, sw_if_index0);
@@ -390,47 +411,8 @@ l2output_node_fn (vlib_main_t * vm,
 			      &msm->next_nodes,
 			      b0, sw_if_index0, feature_bitmap0, &next0);
 
-	  if (PREDICT_FALSE (config0->out_vtr_flag))
-	    {
-	      /*
-	       * Perform output vlan tag rewrite and the pre-vtr EFP filter check.
-	       * The EFP Filter only needs to be run if there is an output VTR
-	       * configured. The flag for the post-vtr EFP Filter node is used
-	       * to trigger the pre-vtr check as well.
-	       */
-
-	      if (config0->output_vtr.push_and_pop_bytes)
-		{
-		  /* Perform pre-vtr EFP filter check if configured */
-		  u32 failed1 = (feature_bitmap0 & L2OUTPUT_FEAT_EFP_FILTER)
-		    && (l2_efp_filter_process (b0, &(config0->input_vtr)));
-		  u32 failed2 = l2_vtr_process (b0, &(config0->output_vtr));
-
-		  if (PREDICT_FALSE (failed1 | failed2))
-		    {
-		      next0 = L2OUTPUT_NEXT_DROP;
-		      if (failed2)
-			{
-			  b0->error = node->errors[L2OUTPUT_ERROR_VTR_DROP];
-			}
-		      if (failed1)
-			{
-			  b0->error = node->errors[L2OUTPUT_ERROR_EFP_DROP];
-			}
-		    }
-		}
-	      // perform the PBB rewrite
-	      else if (config0->output_pbb_vtr.push_and_pop_bytes)
-		{
-		  u32 failed =
-		    l2_pbb_process (b0, &(config0->output_pbb_vtr));
-		  if (PREDICT_FALSE (failed))
-		    {
-		      next0 = L2OUTPUT_NEXT_DROP;
-		      b0->error = node->errors[L2OUTPUT_ERROR_VTR_DROP];
-		    }
-		}
-	    }
+	  l2output_vtr (node, config0, feature_bitmap0, b0, &next0);
+
 	  /* Perform the split horizon check */
 	  if (PREDICT_FALSE
 	      (split_horizon_violation