ethernet-input optimizations

Change-Id: I4ec7750ef58363bd8966a16a2baeec6db18b7e9e
Signed-off-by: Damjan Marion <damarion@cisco.com>

4 files changed, 552 insertions, 23 deletions

diff --git a/src/vnet/ethernet/ethernet.h b/src/vnet/ethernet/ethernet.h
index ece7aa58217..7435599c347 100644
--- a/src/vnet/ethernet/ethernet.h
+++ b/src/vnet/ethernet/ethernet.h
@@ -45,6 +45,20 @@
 #include <vnet/pg/pg.h>
 #include <vnet/feature/feature.h>
 
+/* ethernet-input frame flags and scalar data */
+
+/* all packets in frame share same sw_if_index */
+#define ETH_INPUT_FRAME_F_SINGLE_SW_IF_IDX (1 << 0)
+
+/* all ip4 packets in frame have correct ip4 checksum */
+#define ETH_INPUT_FRAME_F_IP4_CKSUM_OK (1 << 1)
+
+typedef struct
+{
+  u32 sw_if_index;
+  u32 hw_if_index;
+} ethernet_input_frame_t;
+
 always_inline u64
 ethernet_mac_address_u64 (const u8 * a)
 {
@@ -154,6 +168,7 @@ typedef u32 (ethernet_flag_change_function_t)
 /* Ethernet interface instance. */
 typedef struct ethernet_interface
 {
+  u32 flags;
 
   /* Accept all packets (promiscuous mode). */
 #define ETHERNET_INTERFACE_FLAG_ACCEPT_ALL (1 << 0)
diff --git a/src/vnet/ethernet/interface.c b/src/vnet/ethernet/interface.c
index c66fa8232cf..f1e6785cc6b 100644
--- a/src/vnet/ethernet/interface.c
+++ b/src/vnet/ethernet/interface.c
@@ -379,6 +379,7 @@ ethernet_set_flags (vnet_main_t * vnm, u32 hw_if_index, u32 flags)
   ASSERT (hi->hw_class_index == ethernet_hw_interface_class.index);
 
   ei = pool_elt_at_index (em->interfaces, hi->hw_instance);
+  ei->flags = flags;
   if (ei->flag_change)
     return ei->flag_change (vnm, hi, flags);
   return (u32) ~ 0;
diff --git a/src/vnet/ethernet/node.c b/src/vnet/ethernet/node.c
index 53d5b4eb02d..c39c6d7e0c7 100755
--- a/src/vnet/ethernet/node.c
+++ b/src/vnet/ethernet/node.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015 Cisco and/or its affiliates.
+ * Copyright (c) 2018 Cisco and/or its affiliates.
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
  * You may obtain a copy of the License at:
@@ -49,7 +49,9 @@
 #define foreach_ethernet_input_next		\
   _ (PUNT, "error-punt")			\
   _ (DROP, "error-drop")			\
-  _ (LLC, "llc-input")
+  _ (LLC, "llc-input")				\
+  _ (IP4_INPUT, "ip4-input")			\
+  _ (IP4_INPUT_NCS, "ip4-input-no-checksum")
 
 typedef enum
 {
@@ -62,6 +64,8 @@ typedef enum
 typedef struct
 {
   u8 packet_data[32];
+  u16 frame_flags;
+  ethernet_input_frame_t frame_data;
 } ethernet_input_trace_t;
 
 static u8 *
@@ -70,7 +74,16 @@ format_ethernet_input_trace (u8 * s, va_list * va)
   CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *);
   CLIB_UNUSED (vlib_node_t * node) = va_arg (*va, vlib_node_t *);
   ethernet_input_trace_t *t = va_arg (*va, ethernet_input_trace_t *);
+  u32 indent = format_get_indent (s);
 
+  if (t->frame_flags)
+    {
+      s = format (s, "frame: flags 0x%x", t->frame_flags);
+      if (t->frame_flags & ETH_INPUT_FRAME_F_SINGLE_SW_IF_IDX)
+	s = format (s, ", hw-if-index %u, sw-if-index %u",
+		    t->frame_data.hw_if_index, t->frame_data.sw_if_index);
+      s = format (s, "\n%U", format_white_space, indent);
+    }
   s = format (s, "%U", format_ethernet_header, t->packet_data);
 
   return s;
@@ -289,16 +302,401 @@ determine_next_node (ethernet_main_t * em,
     }
 }
 
-static_always_inline uword
+typedef enum
+{
+  ETYPE_ID_UNKNOWN = 0,
+  ETYPE_ID_IP4,
+  ETYPE_ID_IP6,
+  ETYPE_ID_MPLS,
+  ETYPE_N_IDS,
+} etype_id_t;
+
+static_always_inline void
+eth_input_advance_and_flags (vlib_main_t * vm, u32 * from, u32 n_left,
+			     i16 advance, u32 and_flags, u32 or_flags)
+{
+  vlib_buffer_t *b[8];
+  while (n_left >= 8)
+    {
+      vlib_get_buffers (vm, from, b, 8);
+      vlib_buffer_advance (b[0], advance);
+      vlib_buffer_advance (b[1], advance);
+      vlib_buffer_advance (b[2], advance);
+      vlib_buffer_advance (b[3], advance);
+      vlib_buffer_advance (b[4], advance);
+      vlib_buffer_advance (b[5], advance);
+      vlib_buffer_advance (b[6], advance);
+      vlib_buffer_advance (b[7], advance);
+      b[0]->flags = (b[0]->flags & and_flags) | or_flags;
+      b[1]->flags = (b[1]->flags & and_flags) | or_flags;
+      b[2]->flags = (b[2]->flags & and_flags) | or_flags;
+      b[3]->flags = (b[3]->flags & and_flags) | or_flags;
+      b[4]->flags = (b[4]->flags & and_flags) | or_flags;
+      b[5]->flags = (b[5]->flags & and_flags) | or_flags;
+      b[6]->flags = (b[6]->flags & and_flags) | or_flags;
+      b[7]->flags = (b[7]->flags & and_flags) | or_flags;
+
+      n_left -= 8;
+      from += 8;
+    }
+  while (n_left)
+    {
+      vlib_get_buffers (vm, from, b, 1);
+      vlib_buffer_advance (b[0], advance);
+      b[0]->flags = (b[0]->flags & and_flags) | or_flags;
+
+      n_left -= 1;
+      from += 1;
+    }
+}
+
+typedef struct
+{
+  u16 etypes[VLIB_FRAME_SIZE];
+  u32 bufs_by_etype[ETYPE_N_IDS][VLIB_FRAME_SIZE];
+  u16 n_bufs_by_etype[ETYPE_N_IDS];
+} eth_input_data_t;
+
+/* following vector code relies on following assumptions */
+STATIC_ASSERT_OFFSET_OF (vlib_buffer_t, current_data, 0);
+STATIC_ASSERT_OFFSET_OF (vlib_buffer_t, current_length, 2);
+STATIC_ASSERT_OFFSET_OF (vlib_buffer_t, flags, 4);
+STATIC_ASSERT (STRUCT_OFFSET_OF (vnet_buffer_opaque_t, l2_hdr_offset) ==
+	       STRUCT_OFFSET_OF (vnet_buffer_opaque_t, l3_hdr_offset) - 2,
+	       "l3_hdr_offset must follow l2_hdr_offset");
+
+static_always_inline void
+eth_input_adv_and_flags_x4 (vlib_buffer_t ** b, i16 adv, u32 flags, int is_l3)
+{
+#ifdef CLIB_HAVE_VEC256
+  /* to reduce number of small loads/stores we are loading first 64 bits
+     of each buffer metadata into 256-bit register so we can advance
+     current_data, current_length and flags.
+     Observed saving of this code is ~2 clocks per packet */
+  u64x4 r, radv;
+
+  /* vector if signed 16 bit integers used in signed vector add operation
+     to advnce current_data and current_length */
+  u32x8 flags4 = { 0, flags, 0, flags, 0, flags, 0, flags };
+  i16x16 adv4 = {
+    adv, -adv, 0, 0, adv, -adv, 0, 0,
+    adv, -adv, 0, 0, adv, -adv, 0, 0
+  };
+
+  /* load 4 x 64 bits */
+  r = u64x4_gather (b[0], b[1], b[2], b[3]);
+
+  /* set flags */
+  r |= (u64x4) flags4;
+
+  /* advance buffer */
+  radv = (u64x4) ((i16x16) r + adv4);
+
+  /* write 4 x 64 bits */
+  u64x4_scatter (is_l3 ? radv : r, b[0], b[1], b[2], b[3]);
+
+  /* use old current_data as l2_hdr_offset and new current_data as
+     l3_hdr_offset */
+  r = (u64x4) u16x16_blend (r, radv << 16, 0xaa);
+
+  /* store both l2_hdr_offset and l3_hdr_offset in single store operation */
+  u32x8_scatter_one ((u32x8) r, 0, &vnet_buffer (b[0])->l2_hdr_offset);
+  u32x8_scatter_one ((u32x8) r, 2, &vnet_buffer (b[1])->l2_hdr_offset);
+  u32x8_scatter_one ((u32x8) r, 4, &vnet_buffer (b[2])->l2_hdr_offset);
+  u32x8_scatter_one ((u32x8) r, 6, &vnet_buffer (b[3])->l2_hdr_offset);
+
+  ASSERT (b[0]->current_data == vnet_buffer (b[0])->l3_hdr_offset);
+  ASSERT (b[1]->current_data == vnet_buffer (b[1])->l3_hdr_offset);
+  ASSERT (b[2]->current_data == vnet_buffer (b[2])->l3_hdr_offset);
+  ASSERT (b[3]->current_data == vnet_buffer (b[3])->l3_hdr_offset);
+
+  ASSERT (b[0]->current_data - vnet_buffer (b[0])->l2_hdr_offset == adv);
+  ASSERT (b[1]->current_data - vnet_buffer (b[1])->l2_hdr_offset == adv);
+  ASSERT (b[2]->current_data - vnet_buffer (b[2])->l2_hdr_offset == adv);
+  ASSERT (b[3]->current_data - vnet_buffer (b[3])->l2_hdr_offset == adv);
+
+#else
+  vnet_buffer (b[0])->l2_hdr_offset = b[0]->current_data;
+  vnet_buffer (b[1])->l2_hdr_offset = b[1]->current_data;
+  vnet_buffer (b[2])->l2_hdr_offset = b[2]->current_data;
+  vnet_buffer (b[3])->l2_hdr_offset = b[3]->current_data;
+  vnet_buffer (b[0])->l3_hdr_offset = b[0]->current_data + adv;
+  vnet_buffer (b[1])->l3_hdr_offset = b[1]->current_data + adv;
+  vnet_buffer (b[2])->l3_hdr_offset = b[2]->current_data + adv;
+  vnet_buffer (b[3])->l3_hdr_offset = b[3]->current_data + adv;
+
+  if (is_l3)
+    {
+      vlib_buffer_advance (b[0], adv);
+      vlib_buffer_advance (b[1], adv);
+      vlib_buffer_advance (b[2], adv);
+      vlib_buffer_advance (b[3], adv);
+    }
+
+  b[0]->flags |= flags;
+  b[1]->flags |= flags;
+  b[2]->flags |= flags;
+  b[3]->flags |= flags;
+#endif
+
+  if (!is_l3)
+    {
+      vnet_buffer (b[0])->l2.l2_len = adv;
+      vnet_buffer (b[1])->l2.l2_len = adv;
+      vnet_buffer (b[2])->l2.l2_len = adv;
+      vnet_buffer (b[3])->l2.l2_len = adv;
+    }
+}
+
+static_always_inline void
+eth_input_adv_and_flags_x1 (vlib_buffer_t ** b, i16 adv, u32 flags, int is_l3)
+{
+  vnet_buffer (b[0])->l2_hdr_offset = b[0]->current_data;
+  vnet_buffer (b[0])->l3_hdr_offset = b[0]->current_data + adv;
+
+  if (is_l3)
+    vlib_buffer_advance (b[0], adv);
+  b[0]->flags |= flags;
+  if (!is_l3)
+    vnet_buffer (b[0])->l2.l2_len = adv;
+}
+
+static_always_inline void
+eth_input_process_frame (vlib_main_t * vm, u32 * from, u16 * etype,
+			 u32 n_left, int is_l3)
+{
+  vlib_buffer_t *b[16];
+  ethernet_header_t *e;
+  int adv = sizeof (ethernet_header_t);
+
+  u32 flags = VNET_BUFFER_F_L2_HDR_OFFSET_VALID |
+    VNET_BUFFER_F_L3_HDR_OFFSET_VALID;
+
+  while (n_left >= 16)
+    {
+      vlib_buffer_t **ph = b + 12, **pd = b + 8;
+      vlib_get_buffers (vm, from, b, 4);
+      vlib_get_buffers (vm, from + 8, b + 8, 8);
+
+      vlib_prefetch_buffer_header (ph[0], LOAD);
+      vlib_prefetch_buffer_data (pd[0], LOAD);
+      e = vlib_buffer_get_current (b[0]);
+      etype[0] = e->type;
+
+      vlib_prefetch_buffer_header (ph[1], LOAD);
+      vlib_prefetch_buffer_data (pd[1], LOAD);
+      e = vlib_buffer_get_current (b[1]);
+      etype[1] = e->type;
+
+      vlib_prefetch_buffer_header (ph[2], LOAD);
+      vlib_prefetch_buffer_data (pd[2], LOAD);
+      e = vlib_buffer_get_current (b[2]);
+      etype[2] = e->type;
+
+      vlib_prefetch_buffer_header (ph[3], LOAD);
+      vlib_prefetch_buffer_data (pd[3], LOAD);
+      e = vlib_buffer_get_current (b[3]);
+      etype[3] = e->type;
+
+      eth_input_adv_and_flags_x4 (b, adv, flags, is_l3);
+
+      /* next */
+      n_left -= 4;
+      etype += 4;
+      from += 4;
+    }
+  while (n_left >= 4)
+    {
+      vlib_get_buffers (vm, from, b, 4);
+
+      e = vlib_buffer_get_current (b[0]);
+      etype[0] = e->type;
+
+      e = vlib_buffer_get_current (b[1]);
+      etype[1] = e->type;
+
+      e = vlib_buffer_get_current (b[2]);
+      etype[2] = e->type;
+
+      e = vlib_buffer_get_current (b[3]);
+      etype[3] = e->type;
+
+      eth_input_adv_and_flags_x4 (b, adv, flags, is_l3);
+
+      /* next */
+      n_left -= 4;
+      etype += 4;
+      from += 4;
+    }
+  while (n_left)
+    {
+      vlib_get_buffers (vm, from, b, 1);
+
+      e = vlib_buffer_get_current (b[0]);
+      etype[0] = e->type;
+
+      eth_input_adv_and_flags_x1 (b, adv, flags, is_l3);
+
+      /* next */
+      n_left -= 1;
+      etype += 1;
+      from += 1;
+    }
+}
+
+static_always_inline void
+eth_input_sort (vlib_main_t * vm, u32 * from, u32 n_packets,
+		eth_input_data_t * d)
+{
+  u16 *etype = d->etypes;
+  i32 n_left = n_packets;
+
+#if defined (CLIB_HAVE_VEC256)
+  u16x16 e16;
+  u16x16 et16_ip4 = u16x16_splat (clib_host_to_net_u16 (ETHERNET_TYPE_IP4));
+  u16x16 et16_ip6 = u16x16_splat (clib_host_to_net_u16 (ETHERNET_TYPE_IP6));
+  u16x16 et16_mpls = u16x16_splat (clib_host_to_net_u16 (ETHERNET_TYPE_MPLS));
+  u16x16 id16_ip4 = u16x16_splat (ETYPE_ID_IP4);
+  u16x16 id16_ip6 = u16x16_splat (ETYPE_ID_IP6);
+  u16x16 id16_mpls = u16x16_splat (ETYPE_ID_MPLS);
+
+  while (n_left > 0)
+    {
+      u16x16 r = { 0 };
+      e16 = u16x16_load_unaligned (etype);
+      r += (e16 == et16_ip4) & id16_ip4;
+      r += (e16 == et16_ip6) & id16_ip6;
+      r += (e16 == et16_mpls) & id16_mpls;
+      u16x16_store_unaligned (r, etype);
+      etype += 16;
+      n_left -= 16;
+    }
+#elif defined (CLIB_HAVE_VEC128)
+  u16x8 e8;
+  u16x8 et8_ip4 = u16x8_splat (clib_host_to_net_u16 (ETHERNET_TYPE_IP4));
+  u16x8 et8_ip6 = u16x8_splat (clib_host_to_net_u16 (ETHERNET_TYPE_IP6));
+  u16x8 et8_mpls = u16x8_splat (clib_host_to_net_u16 (ETHERNET_TYPE_MPLS));
+  u16x8 id8_ip4 = u16x8_splat (ETYPE_ID_IP4);
+  u16x8 id8_ip6 = u16x8_splat (ETYPE_ID_IP6);
+  u16x8 id8_mpls = u16x8_splat (ETYPE_ID_MPLS);
+
+  while (n_left > 0)
+    {
+      u16x8 r = { 0 };
+      e8 = u16x8_load_unaligned (etype);
+      r += (e8 == et8_ip4) & id8_ip4;
+      r += (e8 == et8_ip6) & id8_ip6;
+      r += (e8 == et8_mpls) & id8_mpls;
+      u16x8_store_unaligned (r, etype);
+      etype += 8;
+      n_left -= 8;
+    }
+#else
+  while (n_left)
+    {
+      if (etype[0] == ETHERNET_TYPE_IP4)
+	etype[0] = ETYPE_ID_IP4;
+      else if (etype[0] == ETHERNET_TYPE_IP6)
+	etype[0] = ETYPE_ID_IP6;
+      else if (etype[0] == ETHERNET_TYPE_MPLS)
+	etype[0] = ETYPE_ID_MPLS;
+      else
+	etype[0] = ETYPE_ID_UNKNOWN;
+
+      etype += 1;
+      n_left -= 1;
+    }
+#endif
+
+  etype = d->etypes;
+  n_left = n_packets;
+
+  clib_memset_u16 (d->n_bufs_by_etype, 0, ETYPE_N_IDS);
+  while (n_left)
+    {
+      u16 x, y;
+      x = etype[0];
+      y = d->n_bufs_by_etype[x];
+
+#ifdef CLIB_HAVE_VEC256
+      if (n_left >= 16 && u16x16_is_all_equal (u16x16_load_unaligned (etype),
+					       etype[0]))
+	{
+	  clib_memcpy_fast (&d->bufs_by_etype[x][y], from, 16 * sizeof (u32));
+	  d->n_bufs_by_etype[x] += 16;
+
+	  /* next */
+	  n_left -= 16;
+	  etype += 16;
+	  from += 16;
+	  continue;
+	}
+#endif
+#ifdef CLIB_HAVE_VEC128
+      if (n_left >= 8 && u16x8_is_all_equal (u16x8_load_unaligned (etype),
+					     etype[0]))
+	{
+	  clib_memcpy_fast (&d->bufs_by_etype[x][y], from, 8 * sizeof (u32));
+	  d->n_bufs_by_etype[x] += 8;
+
+	  /* next */
+	  n_left -= 8;
+	  etype += 8;
+	  from += 8;
+	  continue;
+	}
+#endif
+      d->bufs_by_etype[x][y] = from[0];
+      d->n_bufs_by_etype[x]++;
+
+      /* next */
+      n_left -= 1;
+      etype += 1;
+      from += 1;
+    }
+}
+
+static_always_inline void
+ethernet_input_trace (vlib_main_t * vm, vlib_node_runtime_t * node,
+		      vlib_frame_t * from_frame)
+{
+  u32 *from, n_left;
+  if ((node->flags & VLIB_NODE_FLAG_TRACE) == 0)
+    return;
+
+  from = vlib_frame_vector_args (from_frame);
+  n_left = from_frame->n_vectors;
+
+  while (n_left)
+    {
+      ethernet_input_trace_t *t0;
+      vlib_buffer_t *b0 = vlib_get_buffer (vm, from[0]);
+
+      if (b0->flags & VLIB_BUFFER_IS_TRACED)
+	{
+	  t0 = vlib_add_trace (vm, node, b0, sizeof (ethernet_input_trace_t));
+	  clib_memcpy_fast (t0->packet_data, b0->data + b0->current_data,
+			    sizeof (t0->packet_data));
+	  t0->frame_flags = from_frame->flags;
+	  clib_memcpy_fast (&t0->frame_data,
+			    vlib_frame_scalar_args (from_frame),
+			    sizeof (ethernet_input_frame_t));
+	}
+      from += 1;
+      n_left -= 1;
+    }
+}
+
+static_always_inline void
 ethernet_input_inline (vlib_main_t * vm,
 		       vlib_node_runtime_t * node,
-		       vlib_frame_t * from_frame,
+		       u32 * from, u32 n_packets,
 		       ethernet_input_variant_t variant)
 {
   vnet_main_t *vnm = vnet_get_main ();
   ethernet_main_t *em = &ethernet_main;
   vlib_node_runtime_t *error_node;
-  u32 n_left_from, next_index, *from, *to_next;
+  u32 n_left_from, next_index, *to_next;
   u32 stats_sw_if_index, stats_n_packets, stats_n_bytes;
   u32 thread_index = vm->thread_index;
   u32 cached_sw_if_index = ~0;
@@ -310,15 +708,7 @@ ethernet_input_inline (vlib_main_t * vm,
   else
     error_node = node;
 
-  from = vlib_frame_vector_args (from_frame);
-  n_left_from = from_frame->n_vectors;
-
-  if (node->flags & VLIB_NODE_FLAG_TRACE)
-    vlib_trace_frame_buffers_only (vm, node,
-				   from,
-				   n_left_from,
-				   sizeof (from[0]),
-				   sizeof (ethernet_input_trace_t));
+  n_left_from = n_packets;
 
   next_index = node->cached_next_index;
   stats_sw_if_index = node->runtime_data[0];
@@ -764,32 +1154,136 @@ ethernet_input_inline (vlib_main_t * vm,
 	 thread_index, stats_sw_if_index, stats_n_packets, stats_n_bytes);
       node->runtime_data[0] = stats_sw_if_index;
     }
+}
 
-  return from_frame->n_vectors;
+static_always_inline void
+eth_input_enqueue_untagged (vlib_main_t * vm, vlib_node_runtime_t * node,
+			    eth_input_data_t * d, int ip4_cksum_ok, int is_l3)
+{
+  ethernet_main_t *em = &ethernet_main;
+  etype_id_t id;
+  u32 next_index;
+
+  id = ETYPE_ID_IP4;
+  if (d->n_bufs_by_etype[id])
+    {
+      if (is_l3)
+	{
+	  next_index = em->l3_next.input_next_ip4;
+	  if (next_index == ETHERNET_INPUT_NEXT_IP4_INPUT && ip4_cksum_ok)
+	    next_index = ETHERNET_INPUT_NEXT_IP4_INPUT_NCS;
+	}
+      else
+	next_index = em->l2_next;
+
+      vlib_buffer_enqueue_to_single_next (vm, node, d->bufs_by_etype[id],
+					  next_index, d->n_bufs_by_etype[id]);
+    }
+
+  id = ETYPE_ID_IP6;
+  if (d->n_bufs_by_etype[id])
+    {
+      next_index = is_l3 ? em->l3_next.input_next_ip6 : em->l2_next;
+      vlib_buffer_enqueue_to_single_next (vm, node, d->bufs_by_etype[id],
+					  next_index, d->n_bufs_by_etype[id]);
+    }
+
+  id = ETYPE_ID_MPLS;
+  if (d->n_bufs_by_etype[id])
+    {
+      next_index = is_l3 ? em->l3_next.input_next_mpls : em->l2_next;
+      vlib_buffer_enqueue_to_single_next (vm, node, d->bufs_by_etype[id],
+					  next_index, d->n_bufs_by_etype[id]);
+    }
+
+  id = ETYPE_ID_UNKNOWN;
+  if (d->n_bufs_by_etype[id])
+    {
+      /* in case of l3 interfaces, we already advanced  buffer so we need to
+         roll back */
+      if (is_l3)
+	eth_input_advance_and_flags (vm, d->bufs_by_etype[id],
+				     d->n_bufs_by_etype[id],
+				     -(i16) sizeof (ethernet_header_t),
+				     ~VNET_BUFFER_F_L3_HDR_OFFSET_VALID, 0);
+      ethernet_input_inline (vm, node, d->bufs_by_etype[id],
+			     d->n_bufs_by_etype[id],
+			     ETHERNET_INPUT_VARIANT_ETHERNET);
+    }
 }
 
 VLIB_NODE_FN (ethernet_input_node) (vlib_main_t * vm,
 				    vlib_node_runtime_t * node,
-				    vlib_frame_t * from_frame)
+				    vlib_frame_t * frame)
 {
-  return ethernet_input_inline (vm, node, from_frame,
-				ETHERNET_INPUT_VARIANT_ETHERNET);
+  vnet_main_t *vnm = vnet_get_main ();
+  ethernet_main_t *em = &ethernet_main;
+  u32 *from = vlib_frame_vector_args (frame);
+  u32 n_packets = frame->n_vectors;
+
+  ethernet_input_trace (vm, node, frame);
+
+  if (frame->flags & ETH_INPUT_FRAME_F_SINGLE_SW_IF_IDX)
+    {
+      eth_input_data_t data, *d = &data;
+      ethernet_input_frame_t *ef = vlib_frame_scalar_args (frame);
+      vnet_hw_interface_t *hi = vnet_get_hw_interface (vnm, ef->hw_if_index);
+      main_intf_t *intf0 = vec_elt_at_index (em->main_intfs, hi->hw_if_index);
+      subint_config_t *subint0 = &intf0->untagged_subint;
+      int ip4_cksum_ok = (frame->flags & ETH_INPUT_FRAME_F_IP4_CKSUM_OK) != 0;
+
+      if (subint0->flags & SUBINT_CONFIG_L2)
+	{
+	  /* untagged packets are treated as L2 */
+	  eth_input_process_frame (vm, from, d->etypes, n_packets, 0);
+	  eth_input_sort (vm, from, n_packets, d);
+	  eth_input_enqueue_untagged (vm, node, d, ip4_cksum_ok, 0);
+	}
+      else
+	{
+	  ethernet_interface_t *ei;
+	  ei = pool_elt_at_index (em->interfaces, hi->hw_instance);
+
+	  /* currently only slowpath deals with dmac check */
+	  if (ei->flags & ETHERNET_INTERFACE_FLAG_ACCEPT_ALL)
+	    goto slowpath;
+
+	  /* untagged packets are treated as L3 */
+	  eth_input_process_frame (vm, from, d->etypes, n_packets, 1);
+	  eth_input_sort (vm, from, n_packets, d);
+	  eth_input_enqueue_untagged (vm, node, d, ip4_cksum_ok, 1);
+	}
+      return n_packets;
+    }
+
+slowpath:
+  ethernet_input_inline (vm, node, from, n_packets,
+			 ETHERNET_INPUT_VARIANT_ETHERNET);
+  return n_packets;
 }
 
 VLIB_NODE_FN (ethernet_input_type_node) (vlib_main_t * vm,
 					 vlib_node_runtime_t * node,
 					 vlib_frame_t * from_frame)
 {
-  return ethernet_input_inline (vm, node, from_frame,
-				ETHERNET_INPUT_VARIANT_ETHERNET_TYPE);
+  u32 *from = vlib_frame_vector_args (from_frame);
+  u32 n_packets = from_frame->n_vectors;
+  ethernet_input_trace (vm, node, from_frame);
+  ethernet_input_inline (vm, node, from, n_packets,
+			 ETHERNET_INPUT_VARIANT_ETHERNET_TYPE);
+  return n_packets;
 }
 
 VLIB_NODE_FN (ethernet_input_not_l2_node) (vlib_main_t * vm,
 					   vlib_node_runtime_t * node,
 					   vlib_frame_t * from_frame)
 {
-  return ethernet_input_inline (vm, node, from_frame,
-				ETHERNET_INPUT_VARIANT_NOT_L2);
+  u32 *from = vlib_frame_vector_args (from_frame);
+  u32 n_packets = from_frame->n_vectors;
+  ethernet_input_trace (vm, node, from_frame);
+  ethernet_input_inline (vm, node, from, n_packets,
+			 ETHERNET_INPUT_VARIANT_NOT_L2);
+  return n_packets;
 }
 
 
@@ -1159,6 +1653,7 @@ VLIB_REGISTER_NODE (ethernet_input_node) = {
   .name = "ethernet-input",
   /* Takes a vector of packets. */
   .vector_size = sizeof (u32),
+  .scalar_size = sizeof (ethernet_input_frame_t),
   .n_errors = ETHERNET_N_ERROR,
   .error_strings = ethernet_error_strings,
   .n_next_nodes = ETHERNET_INPUT_N_NEXT,
diff --git a/src/vnet/pg/input.c b/src/vnet/pg/input.c
index 8b46688929f..ee6aad4995e 100644
--- a/src/vnet/pg/input.c
+++ b/src/vnet/pg/input.c
@@ -40,6 +40,7 @@
 #include <vlib/vlib.h>
 #include <vnet/pg/pg.h>
 #include <vnet/vnet.h>
+#include <vnet/ethernet/ethernet.h>
 #include <vnet/feature/feature.h>
 #include <vnet/devices/devices.h>
 
@@ -1496,7 +1497,24 @@ pg_generate_packets (vlib_node_runtime_t * node,
     {
       u32 *head, *start, *end;
 
-      vlib_get_next_frame (vm, node, next_index, to_next, n_left);
+      if (PREDICT_TRUE (next_index == VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT))
+	{
+	  vlib_next_frame_t *nf;
+	  vlib_frame_t *f;
+	  ethernet_input_frame_t *ef;
+	  pg_interface_t *pi;
+	  vlib_get_new_next_frame (vm, node, next_index, to_next, n_left);
+	  nf = vlib_node_runtime_get_next_frame (vm, node, next_index);
+	  f = vlib_get_frame (vm, nf->frame_index);
+	  f->flags = ETH_INPUT_FRAME_F_SINGLE_SW_IF_IDX;
+
+	  ef = vlib_frame_scalar_args (f);
+	  pi = pool_elt_at_index (pg->interfaces, s->pg_if_index);
+	  ef->sw_if_index = pi->sw_if_index;
+	  ef->hw_if_index = pi->hw_if_index;
+	}
+      else
+	vlib_get_next_frame (vm, node, next_index, to_next, n_left);
 
       n_this_frame = n_packets_to_generate;
       if (n_this_frame > n_left)