L2-FWD: use vlib_buffer_enqueue_to_next and cache prefetch fix

before:
  l2-fwd ...  1.72e1  256.00
after:
  l2-fwd ...  1.49e1  256.00

Change-Id: I24b29b799435776abc6e60df0dd0301b74aac99b
Signed-off-by: Neale Ranns <nranns@cisco.com>

4 files changed, 159 insertions, 201 deletions

diff --git a/src/vnet/CMakeLists.txt b/src/vnet/CMakeLists.txt
index 9906a40fcc4..5f89b694afd 100644
--- a/src/vnet/CMakeLists.txt
+++ b/src/vnet/CMakeLists.txt
@@ -158,7 +158,11 @@ list(APPEND VNET_SOURCES
   l2/l2_xcrw.c
 )
 
-list(APPEND VNET_MULTIARCH_SOURCES l2/l2_output.c l2/l2_learn.c)
+list(APPEND VNET_MULTIARCH_SOURCES
+  l2/l2_fwd.c
+  l2/l2_learn.c
+  l2/l2_output.c
+)
 
 list(APPEND VNET_HEADERS
   l2/feat_bitmap.h
@@ -1374,4 +1378,4 @@ if(VPP_BUILD_SESSION_ECHO_APPS)
     )
 endif(VPP_BUILD_SESSION_ECHO_APPS)
 
-##############################################################################
\ No newline at end of file
+##############################################################################
diff --git a/src/vnet/l2/l2_bvi.h b/src/vnet/l2/l2_bvi.h
index cc86072fab7..7abe462054e 100644
--- a/src/vnet/l2/l2_bvi.h
+++ b/src/vnet/l2/l2_bvi.h
@@ -38,7 +38,7 @@ static_always_inline u32
 l2_to_bvi (vlib_main_t * vlib_main,
 	   vnet_main_t * vnet_main,
 	   vlib_buffer_t * b0,
-	   u32 bvi_sw_if_index, next_by_ethertype_t * l3_next, u32 * next0)
+	   u32 bvi_sw_if_index, next_by_ethertype_t * l3_next, u16 * next0)
 {
   /* Perform L3 my-mac filter */
   ethernet_header_t *e0 = vlib_buffer_get_current (b0);
diff --git a/src/vnet/l2/l2_flood.c b/src/vnet/l2/l2_flood.c
index 668301731a9..5d9873e0304 100644
--- a/src/vnet/l2/l2_flood.c
+++ b/src/vnet/l2/l2_flood.c
@@ -155,11 +155,12 @@ l2flood_node_fn (vlib_main_t * vm,
 
       while (n_left_from > 0 && n_left_to_next > 0)
 	{
-	  u32 next0, sw_if_index0, bi0, ci0;
 	  u16 n_clones, n_cloned, clone0;
 	  l2_bridge_domain_t *bd_config;
+	  u32 sw_if_index0, bi0, ci0;
 	  l2_flood_member_t *member;
 	  vlib_buffer_t *b0, *c0;
+	  u16 next0;
 	  u8 in_shg;
 	  i32 mi;
 
diff --git a/src/vnet/l2/l2_fwd.c b/src/vnet/l2/l2_fwd.c
index fd7f54b41ef..7744fcdfa2f 100644
--- a/src/vnet/l2/l2_fwd.c
+++ b/src/vnet/l2/l2_fwd.c
@@ -82,7 +82,11 @@ format_l2fwd_trace (u8 * s, va_list * args)
   return s;
 }
 
+#ifndef CLIB_MARCH_VARIANT
 l2fwd_main_t l2fwd_main;
+#else
+extern l2fwd_main_t l2fwd_main;
+#endif
 
 static vlib_node_registration_t l2fwd_node;
 
@@ -125,7 +129,7 @@ l2fwd_process (vlib_main_t * vm,
 	       l2fwd_main_t * msm,
 	       vlib_error_main_t * em,
 	       vlib_buffer_t * b0,
-	       u32 sw_if_index0, l2fib_entry_result_t * result0, u32 * next0)
+	       u32 sw_if_index0, l2fib_entry_result_t * result0, u16 * next0)
 {
   int try_flood = result0->raw == ~0;
   int flood_error;
@@ -227,229 +231,176 @@ static_always_inline uword
 l2fwd_node_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		   vlib_frame_t * frame, int do_trace)
 {
-  u32 n_left_from, *from, *to_next;
-  l2fwd_next_t next_index;
+  u32 n_left, *from;
   l2fwd_main_t *msm = &l2fwd_main;
   vlib_node_t *n = vlib_get_node (vm, l2fwd_node.index);
   CLIB_UNUSED (u32 node_counter_base_index) = n->error_heap_index;
   vlib_error_main_t *em = &vm->error_main;
   l2fib_entry_key_t cached_key;
   l2fib_entry_result_t cached_result;
+  vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b;
+  u16 nexts[VLIB_FRAME_SIZE], *next;
 
   /* Clear the one-entry cache in case mac table was updated */
   cached_key.raw = ~0;
   cached_result.raw = ~0;
 
   from = vlib_frame_vector_args (frame);
-  n_left_from = frame->n_vectors;	/* number of packets to process */
-  next_index = node->cached_next_index;
+  n_left = frame->n_vectors;	/* number of packets to process */
+  vlib_get_buffers (vm, from, bufs, n_left);
+  next = nexts;
+  b = bufs;
 
-  while (n_left_from > 0)
+  while (n_left >= 8)
     {
-      u32 n_left_to_next;
-
-      /* 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 >= 8 && n_left_to_next >= 4)
-	{
-	  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 *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];
-	  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];
-
-	  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 (do_trace)
-	    {
-	      if (b0->flags & VLIB_BUFFER_IS_TRACED)
-		{
-		  l2fwd_trace_t *t =
-		    vlib_add_trace (vm, node, b0, sizeof (*t));
-		  t->sw_if_index = sw_if_index0;
-		  t->bd_index = vnet_buffer (b0)->l2.bd_index;
-		  clib_memcpy (t->src, h0->src_address, 6);
-		  clib_memcpy (t->dst, h0->dst_address, 6);
-		}
-	      if (b1->flags & VLIB_BUFFER_IS_TRACED)
-		{
-		  l2fwd_trace_t *t =
-		    vlib_add_trace (vm, node, b1, sizeof (*t));
-		  t->sw_if_index = sw_if_index1;
-		  t->bd_index = vnet_buffer (b1)->l2.bd_index;
-		  clib_memcpy (t->src, h1->src_address, 6);
-		  clib_memcpy (t->dst, h1->dst_address, 6);
-		}
-	      if (b2->flags & VLIB_BUFFER_IS_TRACED)
-		{
-		  l2fwd_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)
-		{
-		  l2fwd_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 */
-#ifdef COUNTERS
-	  em->counters[node_counter_base_index + L2FWD_ERROR_L2FWD] += 4;
-#endif
-	  /* *INDENT-OFF* */
-	  l2fib_lookup_4 (msm->mac_table, &cached_key, &cached_result,
-			  h0->dst_address, h1->dst_address,
-			  h2->dst_address, h3->dst_address,
-			  vnet_buffer (b0)->l2.bd_index,
-			  vnet_buffer (b1)->l2.bd_index,
-			  vnet_buffer (b2)->l2.bd_index,
-			  vnet_buffer (b3)->l2.bd_index,
-			  &key0,	/* not used */
-			  &key1,	/* not used */
-			  &key2,	/* not used */
-			  &key3,	/* not used */
-			  &bucket0,	/* not used */
-			  &bucket1,	/* not used */
-			  &bucket2,	/* not used */
-			  &bucket3,	/* not used */
-			  &result0,
-			  &result1,
-			  &result2,
-			  &result3);
-	  /* *INDENT-ON* */
-	  l2fwd_process (vm, node, msm, em, b0, sw_if_index0, &result0,
-			 &next0);
-	  l2fwd_process (vm, node, msm, em, b1, sw_if_index1, &result1,
-			 &next1);
-	  l2fwd_process (vm, node, msm, em, b2, sw_if_index2, &result2,
-			 &next2);
-	  l2fwd_process (vm, node, msm, em, b3, sw_if_index3, &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_x4 (vm, node, next_index,
-					   to_next, n_left_to_next,
-					   bi0, bi1, bi2, bi3,
-					   next0, next1, next2, next3);
-	}
-
-      while (n_left_from > 0 && n_left_to_next > 0)
+      u32 sw_if_index0, sw_if_index1, sw_if_index2, sw_if_index3;
+      const 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_prefetch_buffer_header (b[4], LOAD);
+	vlib_prefetch_buffer_header (b[5], LOAD);
+	vlib_prefetch_buffer_header (b[6], LOAD);
+	vlib_prefetch_buffer_header (b[7], LOAD);
+
+	CLIB_PREFETCH (b[4]->data, CLIB_CACHE_LINE_BYTES, LOAD);
+	CLIB_PREFETCH (b[5]->data, CLIB_CACHE_LINE_BYTES, LOAD);
+	CLIB_PREFETCH (b[6]->data, CLIB_CACHE_LINE_BYTES, LOAD);
+	CLIB_PREFETCH (b[7]->data, CLIB_CACHE_LINE_BYTES, LOAD);
+      }
+
+      /* RX interface handles */
+      sw_if_index0 = vnet_buffer (b[0])->sw_if_index[VLIB_RX];
+      sw_if_index1 = vnet_buffer (b[1])->sw_if_index[VLIB_RX];
+      sw_if_index2 = vnet_buffer (b[2])->sw_if_index[VLIB_RX];
+      sw_if_index3 = vnet_buffer (b[3])->sw_if_index[VLIB_RX];
+
+      h0 = vlib_buffer_get_current (b[0]);
+      h1 = vlib_buffer_get_current (b[1]);
+      h2 = vlib_buffer_get_current (b[2]);
+      h3 = vlib_buffer_get_current (b[3]);
+
+      if (do_trace)
 	{
-	  u32 bi0;
-	  vlib_buffer_t *b0;
-	  u32 next0;
-	  u32 sw_if_index0;
-	  ethernet_header_t *h0;
-	  l2fib_entry_key_t key0;
-	  l2fib_entry_result_t result0;
-	  u32 bucket0;
-
-	  /* speculatively enqueue b0 to the current next frame */
-	  bi0 = from[0];
-	  to_next[0] = bi0;
-	  from += 1;
-	  to_next += 1;
-	  n_left_from -= 1;
-	  n_left_to_next -= 1;
-
-	  b0 = vlib_get_buffer (vm, bi0);
-
-	  sw_if_index0 = vnet_buffer (b0)->sw_if_index[VLIB_RX];
-
-	  h0 = vlib_buffer_get_current (b0);
-
-	  if (do_trace && PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
+	  if (b[0]->flags & VLIB_BUFFER_IS_TRACED)
 	    {
-	      l2fwd_trace_t *t = vlib_add_trace (vm, node, b0, sizeof (*t));
+	      l2fwd_trace_t *t = vlib_add_trace (vm, node, b[0], sizeof (*t));
 	      t->sw_if_index = sw_if_index0;
-	      t->bd_index = vnet_buffer (b0)->l2.bd_index;
+	      t->bd_index = vnet_buffer (b[0])->l2.bd_index;
 	      clib_memcpy (t->src, h0->src_address, 6);
 	      clib_memcpy (t->dst, h0->dst_address, 6);
 	    }
+	  if (b[1]->flags & VLIB_BUFFER_IS_TRACED)
+	    {
+	      l2fwd_trace_t *t = vlib_add_trace (vm, node, b[1], sizeof (*t));
+	      t->sw_if_index = sw_if_index1;
+	      t->bd_index = vnet_buffer (b[1])->l2.bd_index;
+	      clib_memcpy (t->src, h1->src_address, 6);
+	      clib_memcpy (t->dst, h1->dst_address, 6);
+	    }
+	  if (b[2]->flags & VLIB_BUFFER_IS_TRACED)
+	    {
+	      l2fwd_trace_t *t = vlib_add_trace (vm, node, b[2], sizeof (*t));
+	      t->sw_if_index = sw_if_index2;
+	      t->bd_index = vnet_buffer (b[2])->l2.bd_index;
+	      clib_memcpy (t->src, h2->src_address, 6);
+	      clib_memcpy (t->dst, h2->dst_address, 6);
+	    }
+	  if (b[3]->flags & VLIB_BUFFER_IS_TRACED)
+	    {
+	      l2fwd_trace_t *t = vlib_add_trace (vm, node, b[3], sizeof (*t));
+	      t->sw_if_index = sw_if_index3;
+	      t->bd_index = vnet_buffer (b[3])->l2.bd_index;
+	      clib_memcpy (t->src, h3->src_address, 6);
+	      clib_memcpy (t->dst, h3->dst_address, 6);
+	    }
+	}
 
-	  /* process 1 pkt */
 #ifdef COUNTERS
-	  em->counters[node_counter_base_index + L2FWD_ERROR_L2FWD] += 1;
+      em->counters[node_counter_base_index + L2FWD_ERROR_L2FWD] += 4;
 #endif
-	  l2fib_lookup_1 (msm->mac_table, &cached_key, &cached_result, h0->dst_address, vnet_buffer (b0)->l2.bd_index, &key0,	/* not used */
-			  &bucket0,	/* not used */
-			  &result0);
-	  l2fwd_process (vm, node, msm, em, b0, sw_if_index0, &result0,
-			 &next0);
-
-	  /* verify speculative enqueue, maybe switch current next frame */
-	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index,
-					   to_next, n_left_to_next,
-					   bi0, next0);
+      /* *INDENT-OFF* */
+      l2fib_lookup_4 (msm->mac_table, &cached_key, &cached_result,
+                      h0->dst_address, h1->dst_address,
+                      h2->dst_address, h3->dst_address,
+                      vnet_buffer (b[0])->l2.bd_index,
+                      vnet_buffer (b[1])->l2.bd_index,
+                      vnet_buffer (b[2])->l2.bd_index,
+                      vnet_buffer (b[3])->l2.bd_index,
+                      &key0,	/* not used */
+                      &key1,	/* not used */
+                      &key2,	/* not used */
+                      &key3,	/* not used */
+                      &bucket0,	/* not used */
+                      &bucket1,	/* not used */
+                      &bucket2,	/* not used */
+                      &bucket3,	/* not used */
+                      &result0,
+                      &result1,
+                      &result2,
+                      &result3);
+      /* *INDENT-ON* */
+      l2fwd_process (vm, node, msm, em, b[0], sw_if_index0, &result0, next);
+      l2fwd_process (vm, node, msm, em, b[1], sw_if_index1, &result1,
+		     next + 1);
+      l2fwd_process (vm, node, msm, em, b[2], sw_if_index2, &result2,
+		     next + 2);
+      l2fwd_process (vm, node, msm, em, b[3], sw_if_index3, &result3,
+		     next + 3);
+
+      /* verify speculative enqueues, maybe switch current next frame */
+      /* if next0==next1==next_index then nothing special needs to be done */
+      next += 4;
+      b += 4;
+      n_left -= 4;
+    }
+
+  while (n_left > 0)
+    {
+      u32 sw_if_index0;
+      ethernet_header_t *h0;
+      l2fib_entry_key_t key0;
+      l2fib_entry_result_t result0;
+      u32 bucket0;
+
+      sw_if_index0 = vnet_buffer (b[0])->sw_if_index[VLIB_RX];
+
+      h0 = vlib_buffer_get_current (b[0]);
+
+      if (do_trace && PREDICT_FALSE (b[0]->flags & VLIB_BUFFER_IS_TRACED))
+	{
+	  l2fwd_trace_t *t = vlib_add_trace (vm, node, b[0], sizeof (*t));
+	  t->sw_if_index = sw_if_index0;
+	  t->bd_index = vnet_buffer (b[0])->l2.bd_index;
+	  clib_memcpy (t->src, h0->src_address, 6);
+	  clib_memcpy (t->dst, h0->dst_address, 6);
 	}
 
-      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+      /* process 1 pkt */
+#ifdef COUNTERS
+      em->counters[node_counter_base_index + L2FWD_ERROR_L2FWD] += 1;
+#endif
+      l2fib_lookup_1 (msm->mac_table, &cached_key, &cached_result, h0->dst_address, vnet_buffer (b[0])->l2.bd_index, &key0,	/* not used */
+		      &bucket0,	/* not used */
+		      &result0);
+      l2fwd_process (vm, node, msm, em, b[0], sw_if_index0, &result0, next);
+
+      /* verify speculative enqueue, maybe switch current next frame */
+      next += 1;
+      b += 1;
+      n_left -= 1;
     }
 
+  vlib_buffer_enqueue_to_next (vm, node, from, nexts, frame->n_vectors);
+
   return frame->n_vectors;
 }
 
-static uword
-l2fwd_node_fn (vlib_main_t * vm,
-	       vlib_node_runtime_t * node, vlib_frame_t * frame)
+VLIB_NODE_FN (l2fwd_node) (vlib_main_t * vm,
+			   vlib_node_runtime_t * node, vlib_frame_t * frame)
 {
   if (PREDICT_FALSE ((node->flags & VLIB_NODE_FLAG_TRACE)))
     return l2fwd_node_inline (vm, node, frame, 1 /* do_trace */ );
@@ -458,7 +409,6 @@ l2fwd_node_fn (vlib_main_t * vm,
 
 /* *INDENT-OFF* */
 VLIB_REGISTER_NODE (l2fwd_node,static) = {
-  .function = l2fwd_node_fn,
   .name = "l2-fwd",
   .vector_size = sizeof (u32),
   .format_trace = format_l2fwd_trace,
@@ -477,8 +427,9 @@ VLIB_REGISTER_NODE (l2fwd_node,static) = {
 };
 /* *INDENT-ON* */
 
-VLIB_NODE_FUNCTION_MULTIARCH (l2fwd_node, l2fwd_node_fn)
-     clib_error_t *l2fwd_init (vlib_main_t * vm)
+#ifndef CLIB_MARCH_VARIANT
+clib_error_t *
+l2fwd_init (vlib_main_t * vm)
 {
   l2fwd_main_t *mp = &l2fwd_main;
 
@@ -577,6 +528,8 @@ VLIB_CLI_COMMAND (int_fwd_cli, static) = {
 };
 /* *INDENT-ON* */
 
+#endif
+
 /*
  * fd.io coding-style-patch-verification: ON
  *