5 files changed, 3 insertions, 277 deletions

diff --git a/src/plugins/ixge/ixge.c b/src/plugins/ixge/ixge.c
index 0d287250ca4..628d6d71bd7 100644
--- a/src/plugins/ixge/ixge.c
+++ b/src/plugins/ixge/ixge.c
@@ -2454,7 +2454,6 @@ VNET_DEVICE_CLASS (ixge_device_class) = {
   .clear_counters = ixge_clear_hw_interface_counters,
   .admin_up_down_function = ixge_interface_admin_up_down,
   .rx_redirect_to_node = ixge_set_interface_next_node,
-  .flatten_output_chains = 1,
 };
 /* *INDENT-ON* */
 
diff --git a/src/vnet/interface.c b/src/vnet/interface.c
index 1370d048b0b..7c0272b458d 100644
--- a/src/vnet/interface.c
+++ b/src/vnet/interface.c
@@ -783,9 +783,7 @@ vnet_register_interface (vnet_main_t * vnm,
       /* The new class may differ from the old one.
        * Functions have to be updated. */
       node = vlib_get_node (vm, hw->output_node_index);
-      node->function = dev_class->flatten_output_chains ?
-	vnet_interface_output_node_flatten_multiarch_select () :
-	vnet_interface_output_node_multiarch_select ();
+      node->function = vnet_interface_output_node_multiarch_select ();
       node->format_trace = format_vnet_interface_output_trace;
       nrt = vlib_node_get_runtime (vm, hw->output_node_index);
       nrt->function = node->function;
@@ -827,9 +825,7 @@ vnet_register_interface (vnet_main_t * vnm,
 
       r.flags = 0;
       r.name = output_node_name;
-      r.function = dev_class->flatten_output_chains ?
-	vnet_interface_output_node_flatten_multiarch_select () :
-	vnet_interface_output_node_multiarch_select ();
+      r.function = vnet_interface_output_node_multiarch_select ();
       r.format_trace = format_vnet_interface_output_trace;
 
       {
diff --git a/src/vnet/interface.h b/src/vnet/interface.h
index d684e356ca4..ce7700e4ee0 100644
--- a/src/vnet/interface.h
+++ b/src/vnet/interface.h
@@ -198,9 +198,6 @@ typedef struct _vnet_device_class
   /* Link-list of all device classes set up by constructors created below */
   struct _vnet_device_class *next_class_registration;
 
-  /* Splice vnet_interface_output_node into TX path */
-  u8 flatten_output_chains;
-
   /* Function to set mac address. */
   vnet_interface_set_mac_address_function_t *mac_addr_change_function;
 } vnet_device_class_t;
diff --git a/src/vnet/interface_funcs.h b/src/vnet/interface_funcs.h
index 999b72e5a3c..26eef9b9d87 100644
--- a/src/vnet/interface_funcs.h
+++ b/src/vnet/interface_funcs.h
@@ -301,9 +301,8 @@ typedef struct
   u32 is_deleted;
 } vnet_interface_output_runtime_t;
 
-/* Interface output functions. */
+/* Interface output function. */
 void *vnet_interface_output_node_multiarch_select (void);
-void *vnet_interface_output_node_flatten_multiarch_select (void);
 
 word vnet_sw_interface_compare (vnet_main_t * vnm, uword sw_if_index0,
 				uword sw_if_index1);
diff --git a/src/vnet/interface_output.c b/src/vnet/interface_output.c
index 846eb57b95e..cdf18738193 100644
--- a/src/vnet/interface_output.c
+++ b/src/vnet/interface_output.c
@@ -153,271 +153,6 @@ vnet_interface_output_trace (vlib_main_t * vm,
     }
 }
 
-static never_inline u32
-slow_path (vlib_main_t * vm,
-	   u32 bi,
-	   vlib_buffer_t * b,
-	   u32 n_left_to_tx, u32 * to_tx, u32 * n_slow_bytes_result)
-{
-  /* We've already enqueued a single buffer. */
-  u32 n_buffers = 0;
-  u32 n_slow_bytes = 0;
-
-  while (n_left_to_tx > 0)
-    {
-      to_tx[0] = bi;
-      to_tx += 1;
-      n_left_to_tx -= 1;
-      n_buffers += 1;
-      n_slow_bytes += vlib_buffer_length_in_chain (vm, b);
-
-      /* Be grumpy about zero length buffers for benefit of
-         driver tx function. */
-      ASSERT (b->current_length > 0);
-
-      if (!(b->flags & VLIB_BUFFER_NEXT_PRESENT))
-	break;
-
-      bi = b->next_buffer;
-      b = vlib_get_buffer (vm, bi);
-    }
-
-  /* Ran out of space in next frame trying to enqueue buffers? */
-  if (b->flags & VLIB_BUFFER_NEXT_PRESENT)
-    return 0;
-
-  *n_slow_bytes_result = n_slow_bytes;
-  return n_buffers;
-}
-
-/*
- * Increment TX stats. Roll up consecutive increments to the same sw_if_index
- * into one increment.
- */
-static_always_inline void
-incr_output_stats (vnet_main_t * vnm,
-		   u32 thread_index,
-		   u32 length,
-		   u32 sw_if_index,
-		   u32 * last_sw_if_index, u32 * n_packets, u32 * n_bytes)
-{
-  vnet_interface_main_t *im;
-
-  if (PREDICT_TRUE (sw_if_index == *last_sw_if_index))
-    {
-      *n_packets += 1;
-      *n_bytes += length;
-    }
-  else
-    {
-      if (PREDICT_TRUE (*last_sw_if_index != ~0))
-	{
-	  im = &vnm->interface_main;
-
-	  vlib_increment_combined_counter (im->combined_sw_if_counters
-					   + VNET_INTERFACE_COUNTER_TX,
-					   thread_index,
-					   *last_sw_if_index,
-					   *n_packets, *n_bytes);
-	}
-      *last_sw_if_index = sw_if_index;
-      *n_packets = 1;
-      *n_bytes = length;
-    }
-}
-
-
-/* Interface output functions. */
-uword
-vnet_interface_output_node_flatten (vlib_main_t * vm,
-				    vlib_node_runtime_t * node,
-				    vlib_frame_t * frame)
-{
-  vnet_main_t *vnm = vnet_get_main ();
-  vnet_interface_output_runtime_t *rt = (void *) node->runtime_data;
-  vnet_sw_interface_t *si;
-  vnet_hw_interface_t *hi;
-  u32 n_left_to_tx, *from, *from_end, *to_tx;
-  u32 n_bytes, n_buffers, n_packets;
-  u32 last_sw_if_index;
-  u32 thread_index = vm->thread_index;
-
-  n_buffers = frame->n_vectors;
-
-  if (node->flags & VLIB_NODE_FLAG_TRACE)
-    vnet_interface_output_trace (vm, node, frame, n_buffers);
-
-  from = vlib_frame_args (frame);
-
-  if (rt->is_deleted)
-    return vlib_error_drop_buffers (vm, node, from,
-				    /* buffer stride */ 1,
-				    n_buffers,
-				    VNET_INTERFACE_OUTPUT_NEXT_DROP,
-				    node->node_index,
-				    VNET_INTERFACE_OUTPUT_ERROR_INTERFACE_DELETED);
-
-  si = vnet_get_sw_interface (vnm, rt->sw_if_index);
-  hi = vnet_get_sup_hw_interface (vnm, rt->sw_if_index);
-  if (!(si->flags & (VNET_SW_INTERFACE_FLAG_ADMIN_UP |
-		     VNET_SW_INTERFACE_FLAG_BOND_SLAVE)) ||
-      !(hi->flags & VNET_HW_INTERFACE_FLAG_LINK_UP))
-    {
-      vlib_simple_counter_main_t *cm;
-
-      cm = vec_elt_at_index (vnm->interface_main.sw_if_counters,
-			     VNET_INTERFACE_COUNTER_TX_ERROR);
-      vlib_increment_simple_counter (cm, thread_index,
-				     rt->sw_if_index, n_buffers);
-      return vlib_error_drop_buffers (vm, node, from,
-				      /* buffer stride */ 1,
-				      n_buffers,
-				      VNET_INTERFACE_OUTPUT_NEXT_DROP,
-				      node->node_index,
-				      VNET_INTERFACE_OUTPUT_ERROR_INTERFACE_DOWN);
-    }
-
-  from_end = from + n_buffers;
-
-  /* Total byte count of all buffers. */
-  n_bytes = 0;
-  n_packets = 0;
-  last_sw_if_index = ~0;
-
-  while (from < from_end)
-    {
-      /* Get new next frame since previous incomplete frame may have less
-         than VNET_FRAME_SIZE vectors in it. */
-      vlib_get_new_next_frame (vm, node, VNET_INTERFACE_OUTPUT_NEXT_TX,
-			       to_tx, n_left_to_tx);
-
-      while (from + 4 <= from_end && n_left_to_tx >= 2)
-	{
-	  u32 bi0, bi1;
-	  vlib_buffer_t *b0, *b1;
-
-	  /* Prefetch next iteration. */
-	  vlib_prefetch_buffer_with_index (vm, from[2], LOAD);
-	  vlib_prefetch_buffer_with_index (vm, from[3], LOAD);
-
-	  bi0 = from[0];
-	  bi1 = from[1];
-	  to_tx[0] = bi0;
-	  to_tx[1] = bi1;
-	  from += 2;
-	  to_tx += 2;
-	  n_left_to_tx -= 2;
-
-	  b0 = vlib_get_buffer (vm, bi0);
-	  b1 = vlib_get_buffer (vm, bi1);
-
-	  /* Be grumpy about zero length buffers for benefit of
-	     driver tx function. */
-	  ASSERT (b0->current_length > 0);
-	  ASSERT (b1->current_length > 0);
-
-	  if (PREDICT_FALSE
-	      ((b0->flags | b1->flags) & VLIB_BUFFER_NEXT_PRESENT))
-	    {
-	      u32 n_buffers, n_slow_bytes, i;
-
-	      /* Undo. */
-	      from -= 2;
-	      to_tx -= 2;
-	      n_left_to_tx += 2;
-
-	      /* Do slow path two times. */
-	      for (i = 0; i < 2; i++)
-		{
-		  u32 bi = i ? bi1 : bi0;
-		  vlib_buffer_t *b = i ? b1 : b0;
-
-		  n_buffers = slow_path (vm, bi, b,
-					 n_left_to_tx, to_tx, &n_slow_bytes);
-
-		  /* Not enough room for single packet? */
-		  if (n_buffers == 0)
-		    goto put;
-
-		  from += 1;
-		  to_tx += n_buffers;
-		  n_left_to_tx -= n_buffers;
-		  incr_output_stats (vnm, thread_index, n_slow_bytes,
-				     vnet_buffer (b)->sw_if_index[VLIB_TX],
-				     &last_sw_if_index, &n_packets, &n_bytes);
-		}
-	    }
-	  else
-	    {
-	      incr_output_stats (vnm, thread_index,
-				 vlib_buffer_length_in_chain (vm, b0),
-				 vnet_buffer (b0)->sw_if_index[VLIB_TX],
-				 &last_sw_if_index, &n_packets, &n_bytes);
-	      incr_output_stats (vnm, thread_index,
-				 vlib_buffer_length_in_chain (vm, b0),
-				 vnet_buffer (b1)->sw_if_index[VLIB_TX],
-				 &last_sw_if_index, &n_packets, &n_bytes);
-	    }
-	}
-
-      while (from + 1 <= from_end && n_left_to_tx >= 1)
-	{
-	  u32 bi0;
-	  vlib_buffer_t *b0;
-
-	  bi0 = from[0];
-	  to_tx[0] = bi0;
-	  from += 1;
-	  to_tx += 1;
-	  n_left_to_tx -= 1;
-
-	  b0 = vlib_get_buffer (vm, bi0);
-
-	  /* Be grumpy about zero length buffers for benefit of
-	     driver tx function. */
-	  ASSERT (b0->current_length > 0);
-
-	  if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_NEXT_PRESENT))
-	    {
-	      u32 n_buffers, n_slow_bytes;
-
-	      /* Undo. */
-	      from -= 1;
-	      to_tx -= 1;
-	      n_left_to_tx += 1;
-
-	      n_buffers = slow_path (vm, bi0, b0,
-				     n_left_to_tx, to_tx, &n_slow_bytes);
-
-	      /* Not enough room for single packet? */
-	      if (n_buffers == 0)
-		goto put;
-
-	      from += 1;
-	      to_tx += n_buffers;
-	      n_left_to_tx -= n_buffers;
-	    }
-	  incr_output_stats (vnm, thread_index,
-			     vlib_buffer_length_in_chain (vm, b0),
-			     vnet_buffer (b0)->sw_if_index[VLIB_TX],
-			     &last_sw_if_index, &n_packets, &n_bytes);
-	}
-
-    put:
-      vlib_put_next_frame (vm, node, VNET_INTERFACE_OUTPUT_NEXT_TX,
-			   n_left_to_tx);
-    }
-
-  /* Final update of interface stats. */
-  incr_output_stats (vnm, thread_index, 0, ~0,	/* ~0 will flush stats */
-		     &last_sw_if_index, &n_packets, &n_bytes);
-
-  return n_buffers;
-}
-
-VLIB_NODE_FUNCTION_MULTIARCH_CLONE (vnet_interface_output_node_flatten);
-CLIB_MULTIARCH_SELECT_FN (vnet_interface_output_node_flatten);
-
 uword
 vnet_interface_output_node (vlib_main_t * vm,
 			    vlib_node_runtime_t * node, vlib_frame_t * frame)