vlib: add buffer cloning support

Change-Id: I50070611af15b2b4cc29664a8bee4f821ac3c835
Signed-off-by: Damjan Marion <damarion@cisco.com>

8 files changed, 256 insertions, 265 deletions

diff --git a/src/scripts/vnet/mcast/ip4 b/src/scripts/vnet/mcast/ip4
index 69f1ee00fa8..eb6bab27bc6 100644
--- a/src/scripts/vnet/mcast/ip4
+++ b/src/scripts/vnet/mcast/ip4
@@ -2,7 +2,7 @@ packet-generator new {
   name x
   limit 1
   node ip4-input
-  size 64-64
+  size 512-512
   no-recycle
   data {
     ICMP: 1.0.0.2 -> 232.1.1.1
@@ -11,12 +11,15 @@ packet-generator new {
   }
 }
 
-trace add pg-input 100
-loop create
-loop create
-set int state loop0 up
-set int state loop1 up
+create packet-generator interface pg1
+create packet-generator interface pg2
+create packet-generator interface pg3
+
+set int state pg1 up
+set int state pg2 up
+set int state pg3 up
 
 ip mroute add 232.1.1.1 via pg0 Accept
-ip mroute add 232.1.1.1 via loop0 Forward
-ip mroute add 232.1.1.1 via loop1 Forward
+ip mroute add 232.1.1.1 via pg1 Forward
+ip mroute add 232.1.1.1 via pg2 Forward
+ip mroute add 232.1.1.1 via pg3 Forward
diff --git a/src/vlib/buffer.c b/src/vlib/buffer.c
index 95b4344f10b..4f5eb09da91 100644
--- a/src/vlib/buffer.c
+++ b/src/vlib/buffer.c
@@ -68,8 +68,9 @@ format_vlib_buffer (u8 * s, va_list * args)
   vlib_buffer_t *b = va_arg (*args, vlib_buffer_t *);
   uword indent = format_get_indent (s);
 
-  s = format (s, "current data %d, length %d, free-list %d",
-	      b->current_data, b->current_length, b->free_list_index);
+  s = format (s, "current data %d, length %d, free-list %d, clone-count %u",
+	      b->current_data, b->current_length, b->free_list_index,
+	      b->n_add_refs);
 
   if (b->flags & VLIB_BUFFER_TOTAL_LENGTH_VALID)
     s = format (s, ", totlen-nifb %d",
@@ -84,8 +85,10 @@ format_vlib_buffer (u8 * s, va_list * args)
       u32 next_buffer = b->next_buffer;
       b = vlib_get_buffer (vm, next_buffer);
 
-      s = format (s, "\n%Unext-buffer 0x%x, segment length %d",
-		  format_white_space, indent, next_buffer, b->current_length);
+      s =
+	format (s, "\n%Unext-buffer 0x%x, segment length %d, clone-count %u",
+		format_white_space, indent, next_buffer, b->current_length,
+		b->n_add_refs);
     }
 
   return s;
@@ -262,7 +265,7 @@ vlib_main_t **vlib_mains;
 
 /* When dubugging validate that given buffers are either known allocated
    or known free. */
-static void __attribute__ ((unused))
+static void
 vlib_buffer_validate_alloc_free (vlib_main_t * vm,
 				 u32 * buffers,
 				 uword n_buffers,
@@ -362,6 +365,7 @@ vlib_buffer_create_free_list_helper (vlib_main_t * vm,
 
   /* Setup free buffer template. */
   f->buffer_init_template.free_list_index = f->index;
+  f->buffer_init_template.n_add_refs = 0;
 
   if (is_public)
     {
@@ -620,19 +624,11 @@ vlib_buffer_free_inline (vlib_main_t * vm,
 {
   vlib_buffer_main_t *bm = vm->buffer_main;
   vlib_buffer_free_list_t *fl;
-  static u32 *next_to_free[2];	/* smp bad */
-  u32 i_next_to_free, *b, *n, *f, fi;
-  uword n_left;
+  u32 fi;
   int i;
-  static vlib_buffer_free_list_t **announce_list;
-  vlib_buffer_free_list_t *fl0 = 0, *fl1 = 0;
-  u32 bi0 = (u32) ~ 0, bi1 = (u32) ~ 0, fi0, fi1 = (u32) ~ 0;
-  u8 free0, free1 = 0, free_next0, free_next1;
   u32 (*cb) (vlib_main_t * vm, u32 * buffers, u32 n_buffers,
 	     u32 follow_buffer_next);
 
-  ASSERT (os_get_cpu_number () == 0);
-
   cb = bm->buffer_free_callback;
 
   if (PREDICT_FALSE (cb != 0))
@@ -641,203 +637,68 @@ vlib_buffer_free_inline (vlib_main_t * vm,
   if (!n_buffers)
     return;
 
-  /* Use first buffer to get default free list. */
-  {
-    u32 bi0 = buffers[0];
-    vlib_buffer_t *b0;
-
-    b0 = vlib_get_buffer (vm, bi0);
-    fl = vlib_buffer_get_buffer_free_list (vm, b0, &fi);
-    if (fl->buffers_added_to_freelist_function)
-      vec_add1 (announce_list, fl);
-  }
-
-  vec_validate (next_to_free[0], n_buffers - 1);
-  vec_validate (next_to_free[1], n_buffers - 1);
-
-  i_next_to_free = 0;
-  n_left = n_buffers;
-  b = buffers;
-
-again:
-  /* Verify that buffers are known allocated. */
-  vlib_buffer_validate_alloc_free (vm, b,
-				   n_left, VLIB_BUFFER_KNOWN_ALLOCATED);
-
-  vec_add2_aligned (fl->buffers, f, n_left, CLIB_CACHE_LINE_BYTES);
-
-  n = next_to_free[i_next_to_free];
-  while (n_left >= 4)
-    {
-      vlib_buffer_t *b0, *b1, *binit0, *binit1, dummy_buffers[2];
-
-      bi0 = b[0];
-      bi1 = b[1];
-
-      f[0] = bi0;
-      f[1] = bi1;
-      f += 2;
-      b += 2;
-      n_left -= 2;
-
-      /* Prefetch buffers for next iteration. */
-      vlib_prefetch_buffer_with_index (vm, b[0], WRITE);
-      vlib_prefetch_buffer_with_index (vm, b[1], WRITE);
-
-      b0 = vlib_get_buffer (vm, bi0);
-      b1 = vlib_get_buffer (vm, bi1);
-
-      free0 = (b0->flags & VLIB_BUFFER_RECYCLE) == 0;
-      free1 = (b1->flags & VLIB_BUFFER_RECYCLE) == 0;
-
-      /* Must be before init which will over-write buffer flags. */
-      if (follow_buffer_next)
-	{
-	  n[0] = b0->next_buffer;
-	  free_next0 = free0 && (b0->flags & VLIB_BUFFER_NEXT_PRESENT) != 0;
-	  n += free_next0;
-
-	  n[0] = b1->next_buffer;
-	  free_next1 = free1 && (b1->flags & VLIB_BUFFER_NEXT_PRESENT) != 0;
-	  n += free_next1;
-	}
-      else
-	free_next0 = free_next1 = 0;
-
-      /* Must be before init which will over-write buffer free list. */
-      fi0 = b0->free_list_index;
-      fi1 = b1->free_list_index;
-
-      if (PREDICT_FALSE (fi0 != fi || fi1 != fi))
-	goto slow_path_x2;
-
-      binit0 = free0 ? b0 : &dummy_buffers[0];
-      binit1 = free1 ? b1 : &dummy_buffers[1];
-
-      vlib_buffer_init_two_for_free_list (binit0, binit1, fl);
-      continue;
-
-    slow_path_x2:
-      /* Backup speculation. */
-      f -= 2;
-      n -= free_next0 + free_next1;
-
-      _vec_len (fl->buffers) = f - fl->buffers;
-
-      fl0 = pool_elt_at_index (bm->buffer_free_list_pool, fi0);
-      fl1 = pool_elt_at_index (bm->buffer_free_list_pool, fi1);
-
-      vlib_buffer_add_to_free_list (vm, fl0, bi0, free0);
-      if (PREDICT_FALSE (fl0->buffers_added_to_freelist_function != 0))
-	{
-	  int i;
-	  for (i = 0; i < vec_len (announce_list); i++)
-	    if (fl0 == announce_list[i])
-	      goto no_fl0;
-	  vec_add1 (announce_list, fl0);
-	}
-    no_fl0:
-      if (PREDICT_FALSE (fl1->buffers_added_to_freelist_function != 0))
-	{
-	  int i;
-	  for (i = 0; i < vec_len (announce_list); i++)
-	    if (fl1 == announce_list[i])
-	      goto no_fl1;
-	  vec_add1 (announce_list, fl1);
-	}
-
-    no_fl1:
-      vlib_buffer_add_to_free_list (vm, fl1, bi1, free1);
-
-      /* Possibly change current free list. */
-      if (fi0 != fi && fi1 != fi)
-	{
-	  fi = fi1;
-	  fl = pool_elt_at_index (bm->buffer_free_list_pool, fi);
-	}
-
-      vec_add2_aligned (fl->buffers, f, n_left, CLIB_CACHE_LINE_BYTES);
-    }
-
-  while (n_left >= 1)
+  for (i = 0; i < n_buffers; i++)
     {
-      vlib_buffer_t *b0, *binit0, dummy_buffers[1];
+      vlib_buffer_t *b;
+      u32 bi = buffers[i];
 
-      bi0 = b[0];
-      f[0] = bi0;
-      f += 1;
-      b += 1;
-      n_left -= 1;
-
-      b0 = vlib_get_buffer (vm, bi0);
+      b = vlib_get_buffer (vm, bi);
 
-      free0 = (b0->flags & VLIB_BUFFER_RECYCLE) == 0;
+      fl = vlib_buffer_get_buffer_free_list (vm, b, &fi);
 
-      /* Must be before init which will over-write buffer flags. */
-      if (follow_buffer_next)
+      /* The only current use of this callback: multicast recycle */
+      if (PREDICT_FALSE (fl->buffers_added_to_freelist_function != 0))
 	{
-	  n[0] = b0->next_buffer;
-	  free_next0 = free0 && (b0->flags & VLIB_BUFFER_NEXT_PRESENT) != 0;
-	  n += free_next0;
+	  int j;
+
+	  vlib_buffer_add_to_free_list
+	    (vm, fl, buffers[i], (b->flags & VLIB_BUFFER_RECYCLE) == 0);
+
+	  for (j = 0; j < vec_len (bm->announce_list); j++)
+	    {
+	      if (fl == bm->announce_list[j])
+		goto already_announced;
+	    }
+	  vec_add1 (bm->announce_list, fl);
+	already_announced:
+	  ;
 	}
       else
-	free_next0 = 0;
-
-      /* Must be before init which will over-write buffer free list. */
-      fi0 = b0->free_list_index;
-
-      if (PREDICT_FALSE (fi0 != fi))
-	goto slow_path_x1;
-
-      binit0 = free0 ? b0 : &dummy_buffers[0];
-
-      vlib_buffer_init_for_free_list (binit0, fl);
-      continue;
-
-    slow_path_x1:
-      /* Backup speculation. */
-      f -= 1;
-      n -= free_next0;
-
-      _vec_len (fl->buffers) = f - fl->buffers;
-
-      fl0 = pool_elt_at_index (bm->buffer_free_list_pool, fi0);
-
-      vlib_buffer_add_to_free_list (vm, fl0, bi0, free0);
-      if (PREDICT_FALSE (fl0->buffers_added_to_freelist_function != 0))
 	{
-	  int i;
-	  for (i = 0; i < vec_len (announce_list); i++)
-	    if (fl0 == announce_list[i])
-	      goto no_fl00;
-	  vec_add1 (announce_list, fl0);
+	  if (PREDICT_TRUE ((b->flags & VLIB_BUFFER_RECYCLE) == 0))
+	    {
+	      u32 flags, next;
+
+	      do
+		{
+		  vlib_buffer_t *nb = vlib_get_buffer (vm, bi);
+		  flags = nb->flags;
+		  next = nb->next_buffer;
+		  if (nb->n_add_refs)
+		    nb->n_add_refs--;
+		  else
+		    {
+		      vlib_buffer_validate_alloc_free (vm, &bi, 1,
+						       VLIB_BUFFER_KNOWN_ALLOCATED);
+		      vlib_buffer_add_to_free_list (vm, fl, bi, 1);
+		    }
+		  bi = next;
+		}
+	      while (follow_buffer_next
+		     && (flags & VLIB_BUFFER_NEXT_PRESENT));
+
+	    }
 	}
-
-    no_fl00:
-      fi = fi0;
-      fl = pool_elt_at_index (bm->buffer_free_list_pool, fi);
-
-      vec_add2_aligned (fl->buffers, f, n_left, CLIB_CACHE_LINE_BYTES);
     }
-
-  if (follow_buffer_next && ((n_left = n - next_to_free[i_next_to_free]) > 0))
-    {
-      b = next_to_free[i_next_to_free];
-      i_next_to_free ^= 1;
-      goto again;
-    }
-
-  _vec_len (fl->buffers) = f - fl->buffers;
-
-  if (vec_len (announce_list))
+  if (vec_len (bm->announce_list))
     {
       vlib_buffer_free_list_t *fl;
-      for (i = 0; i < vec_len (announce_list); i++)
+      for (i = 0; i < vec_len (bm->announce_list); i++)
 	{
-	  fl = announce_list[i];
+	  fl = bm->announce_list[i];
 	  fl->buffers_added_to_freelist_function (vm, fl);
 	}
-      _vec_len (announce_list) = 0;
+      _vec_len (bm->announce_list) = 0;
     }
 }
 
@@ -922,6 +783,7 @@ vlib_packet_template_init (vlib_main_t * vm,
   fl->buffer_init_template.current_data = 0;
   fl->buffer_init_template.current_length = n_packet_data_bytes;
   fl->buffer_init_template.flags = 0;
+  fl->buffer_init_template.n_add_refs = 0;
   vlib_worker_thread_barrier_release (vm);
 }
 
diff --git a/src/vlib/buffer.h b/src/vlib/buffer.h
index 8ea79502455..b4015b302c3 100644
--- a/src/vlib/buffer.h
+++ b/src/vlib/buffer.h
@@ -119,7 +119,9 @@ typedef struct
                            feature node
                         */
 
-  u8 dont_waste_me[3]; /**< Available space in the (precious)
+  u8 n_add_refs; /**< Number of additional references to this buffer. */
+
+  u8 dont_waste_me[2]; /**< Available space in the (precious)
                           first 32 octets of buffer metadata
                           Before allocating any of it, discussion required!
                        */
diff --git a/src/vlib/buffer_funcs.h b/src/vlib/buffer_funcs.h
index 0b583a61994..e0fde5f2194 100644
--- a/src/vlib/buffer_funcs.h
+++ b/src/vlib/buffer_funcs.h
@@ -530,6 +530,110 @@ vlib_buffer_copy (vlib_main_t * vm, vlib_buffer_t * b)
   return fd;
 }
 
+/** \brief Create multiple clones of buffer and store them in the supplied array
+
+    @param vm - (vlib_main_t *) vlib main data structure pointer
+    @param src_buffer - (u32) source buffer index
+    @param buffers - (u32 * ) buffer index array
+    @param n_buffers - (u8) number of buffer clones requested
+    @param head_end_offset - (u16) offset relative to current position
+           where packet head ends
+    @return - (u8) number of buffers actually cloned, may be
+    less than the number requested or zero
+*/
+
+always_inline u8
+vlib_buffer_clone (vlib_main_t * vm, u32 src_buffer, u32 * buffers,
+		   u8 n_buffers, u16 head_end_offset)
+{
+  u8 i;
+  vlib_buffer_t *s = vlib_get_buffer (vm, src_buffer);
+
+  ASSERT (s->n_add_refs == 0);
+  ASSERT (n_buffers);
+
+  if (s->current_length <= head_end_offset + CLIB_CACHE_LINE_BYTES * 2)
+    {
+      buffers[0] = src_buffer;
+      for (i = 1; i < n_buffers; i++)
+	{
+	  vlib_buffer_t *d;
+	  d = vlib_buffer_copy (vm, s);
+	  if (d == 0)
+	    return i;
+	  buffers[i] = vlib_get_buffer_index (vm, d);
+
+	}
+      return n_buffers;
+    }
+
+  n_buffers = vlib_buffer_alloc_from_free_list (vm, buffers, n_buffers,
+						s->free_list_index);
+  if (PREDICT_FALSE (n_buffers == 0))
+    {
+      buffers[0] = src_buffer;
+      return 1;
+    }
+
+  for (i = 0; i < n_buffers; i++)
+    {
+      vlib_buffer_t *d = vlib_get_buffer (vm, buffers[i]);
+      d->current_data = s->current_data;
+      d->current_length = head_end_offset;
+      d->free_list_index = s->free_list_index;
+      d->total_length_not_including_first_buffer =
+	s->total_length_not_including_first_buffer + s->current_length -
+	head_end_offset;
+      d->flags = s->flags | VLIB_BUFFER_NEXT_PRESENT;
+      d->flags &= ~VLIB_BUFFER_EXT_HDR_VALID;
+      clib_memcpy (d->opaque, s->opaque, sizeof (s->opaque));
+      clib_memcpy (vlib_buffer_get_current (d), vlib_buffer_get_current (s),
+		   head_end_offset);
+      d->next_buffer = src_buffer;
+    }
+  vlib_buffer_advance (s, head_end_offset);
+  s->n_add_refs = n_buffers - 1;
+  while (s->flags & VLIB_BUFFER_NEXT_PRESENT)
+    {
+      s = vlib_get_buffer (vm, s->next_buffer);
+      s->n_add_refs = n_buffers - 1;
+    }
+
+  return n_buffers;
+}
+
+/** \brief Attach cloned tail to the buffer
+
+    @param vm - (vlib_main_t *) vlib main data structure pointer
+    @param head - (vlib_buffer_t *) head buffer
+    @param tail - (Vlib buffer_t *) tail buffer to clone and attach to head
+*/
+
+always_inline void
+vlib_buffer_attach_clone (vlib_main_t * vm, vlib_buffer_t * head,
+			  vlib_buffer_t * tail)
+{
+  ASSERT ((head->flags & VLIB_BUFFER_NEXT_PRESENT) == 0);
+  ASSERT (head->free_list_index == tail->free_list_index);
+
+  head->flags |= VLIB_BUFFER_NEXT_PRESENT;
+  head->flags &= ~VLIB_BUFFER_TOTAL_LENGTH_VALID;
+  head->flags &= ~VLIB_BUFFER_EXT_HDR_VALID;
+  head->flags |= (tail->flags & VLIB_BUFFER_TOTAL_LENGTH_VALID);
+  head->next_buffer = vlib_get_buffer_index (vm, tail);
+  head->total_length_not_including_first_buffer = tail->current_length +
+    tail->total_length_not_including_first_buffer;
+
+next_segment:
+  __sync_add_and_fetch (&tail->n_add_refs, 1);
+
+  if (tail->flags & VLIB_BUFFER_NEXT_PRESENT)
+    {
+      tail = vlib_get_buffer (vm, tail->next_buffer);
+      goto next_segment;
+    }
+}
+
 /* Initializes the buffer as an empty packet with no chained buffers. */
 always_inline void
 vlib_buffer_chain_init (vlib_buffer_t * first)
@@ -695,7 +799,8 @@ vlib_buffer_init_for_free_list (vlib_buffer_t * dst,
   _(flags);
   _(free_list_index);
 #undef _
-  ASSERT (dst->total_length_not_including_first_buffer == 0);
+  dst->total_length_not_including_first_buffer = 0;
+  ASSERT (dst->n_add_refs == 0);
 }
 
 always_inline void
@@ -727,8 +832,10 @@ vlib_buffer_init_two_for_free_list (vlib_buffer_t * dst0,
   _(flags);
   _(free_list_index);
 #undef _
-  ASSERT (dst0->total_length_not_including_first_buffer == 0);
-  ASSERT (dst1->total_length_not_including_first_buffer == 0);
+  dst0->total_length_not_including_first_buffer = 0;
+  dst1->total_length_not_including_first_buffer = 0;
+  ASSERT (dst0->n_add_refs == 0);
+  ASSERT (dst1->n_add_refs == 0);
 }
 
 #if CLIB_DEBUG > 0
diff --git a/src/vnet/devices/dpdk/buffer.c b/src/vnet/devices/dpdk/buffer.c
index 007093e493a..f95d4cb5c38 100644
--- a/src/vnet/devices/dpdk/buffer.c
+++ b/src/vnet/devices/dpdk/buffer.c
@@ -79,20 +79,46 @@
 STATIC_ASSERT (VLIB_BUFFER_PRE_DATA_SIZE == RTE_PKTMBUF_HEADROOM,
 	       "VLIB_BUFFER_PRE_DATA_SIZE must be equal to RTE_PKTMBUF_HEADROOM");
 
+static_always_inline void
+dpdk_rte_pktmbuf_free (vlib_main_t * vm, vlib_buffer_t * b)
+{
+  vlib_buffer_t *hb = b;
+  struct rte_mbuf *mb;
+  u32 next, flags;
+  mb = rte_mbuf_from_vlib_buffer (hb);
+
+next:
+  flags = b->flags;
+  next = b->next_buffer;
+  mb = rte_mbuf_from_vlib_buffer (b);
+
+  if (PREDICT_FALSE (b->n_add_refs))
+    {
+      rte_mbuf_refcnt_update (mb, b->n_add_refs);
+      b->n_add_refs = 0;
+    }
+
+  rte_pktmbuf_free_seg (mb);
+
+  if (flags & VLIB_BUFFER_NEXT_PRESENT)
+    {
+      b = vlib_get_buffer (vm, next);
+      goto next;
+    }
+}
+
 static void
 del_free_list (vlib_main_t * vm, vlib_buffer_free_list_t * f)
 {
   u32 i;
-  struct rte_mbuf *mb;
   vlib_buffer_t *b;
 
   for (i = 0; i < vec_len (f->buffers); i++)
     {
       b = vlib_get_buffer (vm, f->buffers[i]);
-      mb = rte_mbuf_from_vlib_buffer (b);
-      ASSERT (rte_mbuf_refcnt_read (mb) == 1);
-      rte_pktmbuf_free (mb);
+      dpdk_rte_pktmbuf_free (vm, b);
     }
+
   vec_free (f->name);
   vec_free (f->buffers);
 }
@@ -325,7 +351,6 @@ vlib_buffer_free_inline (vlib_main_t * vm,
   for (i = 0; i < n_buffers; i++)
     {
       vlib_buffer_t *b;
-      struct rte_mbuf *mb;
 
       b = vlib_get_buffer (vm, buffers[i]);
 
@@ -351,11 +376,7 @@ vlib_buffer_free_inline (vlib_main_t * vm,
       else
 	{
 	  if (PREDICT_TRUE ((b->flags & VLIB_BUFFER_RECYCLE) == 0))
-	    {
-	      mb = rte_mbuf_from_vlib_buffer (b);
-	      ASSERT (rte_mbuf_refcnt_read (mb) == 1);
-	      rte_pktmbuf_free (mb);
-	    }
+	    dpdk_rte_pktmbuf_free (vm, b);
 	}
     }
   if (vec_len (bm->announce_list))
diff --git a/src/vnet/devices/dpdk/device.c b/src/vnet/devices/dpdk/device.c
index c9d9a567b7a..17397900c59 100644
--- a/src/vnet/devices/dpdk/device.c
+++ b/src/vnet/devices/dpdk/device.c
@@ -168,13 +168,11 @@ dpdk_validate_rte_mbuf (vlib_main_t * vm, vlib_buffer_t * b,
 	{
 	  b2 = vlib_get_buffer (vm, b2->next_buffer);
 	  mb = rte_mbuf_from_vlib_buffer (b2);
-	  last_mb->next = mb;
-	  last_mb = mb;
 	  rte_pktmbuf_reset (mb);
 	}
     }
 
-  first_mb = mb = rte_mbuf_from_vlib_buffer (b);
+  last_mb = first_mb = mb = rte_mbuf_from_vlib_buffer (b);
   first_mb->nb_segs = 1;
   mb->data_len = b->current_length;
   mb->pkt_len = maybe_multiseg ? vlib_buffer_length_in_chain (vm, b) :
@@ -185,10 +183,17 @@ dpdk_validate_rte_mbuf (vlib_main_t * vm, vlib_buffer_t * b,
     {
       b = vlib_get_buffer (vm, b->next_buffer);
       mb = rte_mbuf_from_vlib_buffer (b);
+      last_mb->next = mb;
+      last_mb = mb;
       mb->data_len = b->current_length;
       mb->pkt_len = b->current_length;
       mb->data_off = VLIB_BUFFER_PRE_DATA_SIZE + b->current_data;
       first_mb->nb_segs++;
+      if (PREDICT_FALSE (b->n_add_refs))
+	{
+	  rte_mbuf_refcnt_update (mb, b->n_add_refs);
+	  b->n_add_refs = 0;
+	}
     }
 }
 
diff --git a/src/vnet/dpo/replicate_dpo.c b/src/vnet/dpo/replicate_dpo.c
index a67b19c893f..a9f334be7cf 100644
--- a/src/vnet/dpo/replicate_dpo.c
+++ b/src/vnet/dpo/replicate_dpo.c
@@ -625,6 +625,7 @@ replicate_inline (vlib_main_t * vm,
                   vlib_frame_t * frame)
 {
     vlib_combined_counter_main_t * cm = &replicate_main.repm_counters;
+    replicate_main_t * rm = &replicate_main;
     u32 n_left_from, * from, * to_next, next_index;
     u32 cpu_index = os_get_cpu_number();
 
@@ -645,13 +646,11 @@ replicate_inline (vlib_main_t * vm,
             const replicate_t *rep0;
             vlib_buffer_t * b0, *c0;
             const dpo_id_t *dpo0;
+	    u8 num_cloned;
 
             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);
             repi0 = vnet_buffer (b0)->ip.adj_index[VLIB_TX];
@@ -661,50 +660,21 @@ replicate_inline (vlib_main_t * vm,
                 cm, cpu_index, repi0, 1,
                 vlib_buffer_length_in_chain(vm, b0));
 
-            /* ship the original to the first bucket */
-            dpo0 = replicate_get_bucket_i(rep0, 0);
-            next0 = dpo0->dpoi_next_node;
-            vnet_buffer (b0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
+	    vec_validate (rm->clones[cpu_index], rep0->rep_n_buckets - 1);
 
-            if (PREDICT_FALSE(b0->flags & VLIB_BUFFER_IS_TRACED))
-            {
-                replicate_trace_t *t = vlib_add_trace (vm, node, b0, sizeof (*t));
-                t->rep_index = repi0;
-                t->dpo = *dpo0;
-            }
-            vlib_validate_buffer_enqueue_x1 (vm, node, next_index,
-                                             to_next, n_left_to_next,
-                                             bi0, next0);
+	    num_cloned = vlib_buffer_clone (vm, bi0, rm->clones[cpu_index], rep0->rep_n_buckets, 128);
 
-            /* ship copies to the rest of the buckets */
-            for (bucket = 1; bucket < rep0->rep_n_buckets; bucket++)
-            {
-                /*
-                 * After the enqueue of the first buffer, and of all subsequent
-                 * buffers in this loop, it is possible that we over-flow the
-                 * frame of the to-next node. When this happens we need to 'put'
-                 * that full frame to the node and get a fresh empty one.
-                 * Note that these are macros with side effects that change
-                 * to_next & n_left_to_next
-                 */
-                if (PREDICT_FALSE(0 == n_left_to_next))
-                {
-                    vlib_put_next_frame (vm, node, next_index, n_left_to_next);
-                    vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
-                }
+	    if (num_cloned != rep0->rep_n_buckets)
+	      {
+		vlib_node_increment_counter
+		  (vm, node->node_index,
+		   REPLICATE_DPO_ERROR_BUFFER_ALLOCATION_FAILURE, 1);
+	      }
 
-                /* Make a copy. This can fail, so deal with it. */
-                c0 = vlib_buffer_copy(vm, b0);
-                if (PREDICT_FALSE (c0 == 0))
-                  {
-                    vlib_node_increment_counter 
-                      (vm, node->node_index, 
-                       REPLICATE_DPO_ERROR_BUFFER_ALLOCATION_FAILURE,
-                       1);
-                    continue;
-                  }
-                
-                ci0 = vlib_get_buffer_index(vm, c0);
+            for (bucket = 0; bucket < num_cloned; bucket++)
+            {
+                ci0 = rm->clones[cpu_index][bucket];
+                c0 = vlib_get_buffer(vm, ci0);
 
                 to_next[0] = ci0;
                 to_next += 1;
@@ -724,7 +694,13 @@ replicate_inline (vlib_main_t * vm,
                 vlib_validate_buffer_enqueue_x1 (vm, node, next_index,
                                                  to_next, n_left_to_next,
                                                  ci0, next0);
+		if (PREDICT_FALSE (n_left_to_next == 0))
+		  {
+		    vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+		    vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
+		  }
             }
+	    vec_reset_length (rm->clones[cpu_index]);
         }
 
         vlib_put_next_frame (vm, node, next_index, n_left_to_next);
@@ -797,3 +773,15 @@ VLIB_REGISTER_NODE (ip6_replicate_node) = {
       [0] = "error-drop",
   },
 };
+
+clib_error_t *
+replicate_dpo_init (vlib_main_t * vm)
+{
+  replicate_main_t * rm = &replicate_main;
+
+  vec_validate (rm->clones, vlib_num_workers());
+
+  return 0;
+}
+
+VLIB_INIT_FUNCTION (replicate_dpo_init);
diff --git a/src/vnet/dpo/replicate_dpo.h b/src/vnet/dpo/replicate_dpo.h
index a564739c9f2..77273015c9e 100644
--- a/src/vnet/dpo/replicate_dpo.h
+++ b/src/vnet/dpo/replicate_dpo.h
@@ -32,6 +32,9 @@
 typedef struct replicate_main_t_
 {
     vlib_combined_counter_main_t repm_counters;
+
+    /* per-cpu vector of cloned packets */
+    u32 **clones;
 } replicate_main_t;
 
 extern replicate_main_t replicate_main;