memif: zero copy slave

Change-Id: I65306fb1f8e39221dd1d8c00737a7fb1c0129ba8
Signed-off-by: Damjan Marion <damarion@cisco.com>

6 files changed, 653 insertions, 145 deletions

diff --git a/src/plugins/memif/cli.c b/src/plugins/memif/cli.c
index ed2b1b7e6d0..f2c8829aa6a 100644
--- a/src/plugins/memif/cli.c
+++ b/src/plugins/memif/cli.c
@@ -183,6 +183,8 @@ memif_create_command_fn (vlib_main_t * vm, unformat_input_t * input,
   if (!unformat_user (input, unformat_line_input, line_input))
     return 0;
 
+  args.is_zero_copy = 1;
+
   while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT)
     {
       if (unformat (line_input, "id %u", &args.id))
@@ -203,6 +205,8 @@ memif_create_command_fn (vlib_main_t * vm, unformat_input_t * input,
 	args.is_master = 1;
       else if (unformat (line_input, "slave"))
 	args.is_master = 0;
+      else if (unformat (line_input, "no-zero-copy"))
+	args.is_zero_copy = 0;
       else if (unformat (line_input, "mode ip"))
 	args.mode = MEMIF_INTERFACE_MODE_IP;
       else if (unformat (line_input, "hw-addr %U",
@@ -421,6 +425,7 @@ memif_show_command_fn (vlib_main_t * vm, unformat_input_t * input,
   memif_main_t *mm = &memif_main;
   memif_if_t *mif;
   vnet_main_t *vnm = vnet_get_main ();
+  memif_region_t *mr;
   memif_queue_t *mq;
   uword i;
   int show_descr = 0;
@@ -498,10 +503,10 @@ memif_show_command_fn (vlib_main_t * vm, unformat_input_t * input,
       vlib_cli_output (vm, "  listener-fd %d conn-fd %d",
 		       msf->sock ? msf->sock->fd : 0,
 		       mif->sock ? mif->sock->fd : 0);
-      vlib_cli_output (vm,
-		       "  num-s2m-rings %u num-m2s-rings %u buffer-size %u",
+      vlib_cli_output (vm, "  num-s2m-rings %u num-m2s-rings %u "
+		       "buffer-size %u num-regions %u",
 		       mif->run.num_s2m_rings, mif->run.num_m2s_rings,
-		       mif->run.buffer_size);
+		       mif->run.buffer_size, vec_len (mif->regions));
 
       if (mif->local_disc_string)
 	vlib_cli_output (vm, "  local-disc-reason \"%s\"",
@@ -510,20 +515,28 @@ memif_show_command_fn (vlib_main_t * vm, unformat_input_t * input,
 	vlib_cli_output (vm, "  remote-disc-reason \"%s\"",
 			 mif->remote_disc_string);
 
+      /* *INDENT-OFF* */
+      vec_foreach_index (i, mif->regions)
+	{
+	  mr = vec_elt_at_index (mif->regions, i);
+	  vlib_cli_output (vm, "  region %u size %u fd %d", i,
+			   mr->region_size, mr->fd);
+	}
       vec_foreach_index (i, mif->tx_queues)
-      {
-	mq = vec_elt_at_index (mif->tx_queues, i);
-	vlib_cli_output (vm, "  %U", format_memif_queue, mq, i);
-	if (show_descr)
-	  vlib_cli_output (vm, "  %U", format_memif_descriptor, mif, mq);
-      }
+	{
+	  mq = vec_elt_at_index (mif->tx_queues, i);
+	  vlib_cli_output (vm, "  %U", format_memif_queue, mq, i);
+	  if (show_descr)
+	    vlib_cli_output (vm, "  %U", format_memif_descriptor, mif, mq);
+	}
       vec_foreach_index (i, mif->rx_queues)
-      {
-	mq = vec_elt_at_index (mif->rx_queues, i);
-	vlib_cli_output (vm, "  %U", format_memif_queue, mq, i);
-	if (show_descr)
-	  vlib_cli_output (vm, "  %U", format_memif_descriptor, mif, mq);
-      }
+	{
+	  mq = vec_elt_at_index (mif->rx_queues, i);
+	  vlib_cli_output (vm, "  %U", format_memif_queue, mq, i);
+	  if (show_descr)
+	    vlib_cli_output (vm, "  %U", format_memif_descriptor, mif, mq);
+	}
+      /* *INDENT-ON* */
     }
 done:
   vec_free (hw_if_indices);
diff --git a/src/plugins/memif/device.c b/src/plugins/memif/device.c
index 112db57b4b4..6accad6e673 100644
--- a/src/plugins/memif/device.c
+++ b/src/plugins/memif/device.c
@@ -99,35 +99,20 @@ memif_add_copy_op (memif_per_thread_data_t * ptd, void *data, u32 len,
 static_always_inline uword
 memif_interface_tx_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 			   vlib_frame_t * frame, memif_if_t * mif,
-			   memif_ring_type_t type)
+			   memif_ring_type_t type, memif_queue_t * mq,
+			   memif_per_thread_data_t * ptd)
 {
-  u8 qid;
   memif_ring_t *ring;
   u32 *buffers = vlib_frame_args (frame);
   u32 n_left = frame->n_vectors;
   u32 n_copy_op;
   u16 ring_size, mask, slot, free_slots;
-  u32 thread_index = vlib_get_thread_index ();
-  memif_per_thread_data_t *ptd = vec_elt_at_index (memif_main.per_thread_data,
-						   thread_index);
-  u8 tx_queues = vec_len (mif->tx_queues);
-  memif_queue_t *mq;
   int n_retries = 5;
   vlib_buffer_t *b0, *b1, *b2, *b3;
   memif_copy_op_t *co;
   memif_region_index_t last_region = ~0;
   void *last_region_shm = 0;
 
-  if (tx_queues < vec_len (vlib_mains))
-    {
-      ASSERT (tx_queues > 0);
-      qid = thread_index % tx_queues;
-      clib_spinlock_lock_if_init (&mif->lockp);
-    }
-  else
-    qid = thread_index;
-
-  mq = vec_elt_at_index (mif->tx_queues, qid);
   ring = mq->ring;
   ring_size = 1 << mq->log2_ring_size;
   mask = ring_size - 1;
@@ -307,6 +292,113 @@ no_free_slots:
   return frame->n_vectors;
 }
 
+static_always_inline uword
+memif_interface_tx_zc_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
+			      vlib_frame_t * frame, memif_if_t * mif,
+			      memif_queue_t * mq,
+			      memif_per_thread_data_t * ptd)
+{
+  memif_ring_t *ring = mq->ring;
+  u32 *buffers = vlib_frame_args (frame);
+  u32 n_left = frame->n_vectors;
+  u16 slot, free_slots, n_free;
+  u16 ring_size = 1 << mq->log2_ring_size;
+  u16 mask = ring_size - 1;
+  int n_retries = 5;
+  vlib_buffer_t *b0;
+
+retry:
+  n_free = ring->tail - mq->last_tail;
+  if (n_free >= 16)
+    {
+      vlib_buffer_free_from_ring_no_next (vm, mq->buffers, mq->last_tail,
+					  ring_size, n_free);
+      mq->last_tail += n_free;
+    }
+
+  slot = ring->head;
+  free_slots = ring_size - ring->head + mq->last_tail;
+
+  while (n_left && free_slots)
+    {
+      u16 s0;
+      u16 slots_in_packet = 1;
+      memif_desc_t *d0;
+      u32 bi0;
+
+      CLIB_PREFETCH (&ring->desc[(slot + 8) & mask], CLIB_CACHE_LINE_BYTES,
+		     STORE);
+
+      if (PREDICT_TRUE (n_left >= 4))
+	vlib_prefetch_buffer_header (vlib_get_buffer (vm, buffers[3]), LOAD);
+
+      bi0 = buffers[0];
+
+    next_in_chain:
+      s0 = slot & mask;
+      d0 = &ring->desc[s0];
+      mq->buffers[s0] = bi0;
+      b0 = vlib_get_buffer (vm, bi0);
+
+      d0->region = b0->buffer_pool_index + 1;
+      d0->offset = (void *) b0->data + b0->current_data -
+	mif->regions[d0->region].shm;
+      d0->length = b0->current_length;
+
+      free_slots--;
+      slot++;
+
+      if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_NEXT_PRESENT))
+	{
+	  if (PREDICT_FALSE (free_slots == 0))
+	    {
+	      /* revert to last fully processed packet */
+	      free_slots += slots_in_packet;
+	      slot -= slots_in_packet;
+	      goto no_free_slots;
+	    }
+
+	  d0->flags = MEMIF_DESC_FLAG_NEXT;
+	  bi0 = b0->next_buffer;
+
+	  /* next */
+	  slots_in_packet++;
+	  goto next_in_chain;
+	}
+
+      d0->flags = 0;
+
+      /* next from */
+      buffers++;
+      n_left--;
+    }
+no_free_slots:
+
+  CLIB_MEMORY_STORE_BARRIER ();
+  ring->head = slot;
+
+  if (n_left && n_retries--)
+    goto retry;
+
+  clib_spinlock_unlock_if_init (&mif->lockp);
+
+  if (n_left)
+    {
+      vlib_error_count (vm, node->node_index, MEMIF_TX_ERROR_NO_FREE_SLOTS,
+			n_left);
+      vlib_buffer_free (vm, buffers, n_left);
+    }
+
+  if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0 && mq->int_fd > -1)
+    {
+      u64 b = 1;
+      CLIB_UNUSED (int r) = write (mq->int_fd, &b, sizeof (b));
+      mq->int_count++;
+    }
+
+  return frame->n_vectors;
+}
+
 uword
 CLIB_MULTIARCH_FN (memif_interface_tx) (vlib_main_t * vm,
 					vlib_node_runtime_t * node,
@@ -315,11 +407,29 @@ CLIB_MULTIARCH_FN (memif_interface_tx) (vlib_main_t * vm,
   memif_main_t *nm = &memif_main;
   vnet_interface_output_runtime_t *rund = (void *) node->runtime_data;
   memif_if_t *mif = pool_elt_at_index (nm->interfaces, rund->dev_instance);
+  memif_queue_t *mq;
+  u32 thread_index = vlib_get_thread_index ();
+  memif_per_thread_data_t *ptd = vec_elt_at_index (memif_main.per_thread_data,
+						   thread_index);
+  u8 tx_queues = vec_len (mif->tx_queues);
+
+  if (tx_queues < vec_len (vlib_mains))
+    {
+      ASSERT (tx_queues > 0);
+      mq = vec_elt_at_index (mif->tx_queues, thread_index % tx_queues);
+      clib_spinlock_lock_if_init (&mif->lockp);
+    }
+  else
+    mq = vec_elt_at_index (mif->tx_queues, thread_index);
 
-  if (mif->flags & MEMIF_IF_FLAG_IS_SLAVE)
-    return memif_interface_tx_inline (vm, node, frame, mif, MEMIF_RING_S2M);
+  if (mif->flags & MEMIF_IF_FLAG_ZERO_COPY)
+    return memif_interface_tx_zc_inline (vm, node, frame, mif, mq, ptd);
+  else if (mif->flags & MEMIF_IF_FLAG_IS_SLAVE)
+    return memif_interface_tx_inline (vm, node, frame, mif, MEMIF_RING_S2M,
+				      mq, ptd);
   else
-    return memif_interface_tx_inline (vm, node, frame, mif, MEMIF_RING_M2S);
+    return memif_interface_tx_inline (vm, node, frame, mif, MEMIF_RING_M2S,
+				      mq, ptd);
 }
 
 static __clib_unused void
diff --git a/src/plugins/memif/memif.c b/src/plugins/memif/memif.c
index c83a955090c..f533ab97e80 100644
--- a/src/plugins/memif/memif.c
+++ b/src/plugins/memif/memif.c
@@ -110,36 +110,42 @@ memif_disconnect (memif_if_t * mif, clib_error_t * err)
       clib_mem_free (mif->sock);
     }
 
+  /* *INDENT-OFF* */
   vec_foreach_index (i, mif->rx_queues)
-  {
-    mq = vec_elt_at_index (mif->rx_queues, i);
-    if (mq->ring)
-      {
-	int rv;
-	rv = vnet_hw_interface_unassign_rx_thread (vnm, mif->hw_if_index, i);
-	if (rv)
-	  DBG ("Warning: unable to unassign interface %d, "
-	       "queue %d: rc=%d", mif->hw_if_index, i, rv);
-	mq->ring = 0;
-      }
-  }
+    {
+      mq = vec_elt_at_index (mif->rx_queues, i);
+      if (mq->ring)
+	{
+	  int rv;
+	  rv = vnet_hw_interface_unassign_rx_thread (vnm, mif->hw_if_index, i);
+	  if (rv)
+	    DBG ("Warning: unable to unassign interface %d, "
+		 "queue %d: rc=%d", mif->hw_if_index, i, rv);
+	  mq->ring = 0;
+	}
+    }
 
   /* free tx and rx queues */
-  vec_foreach (mq, mif->rx_queues) memif_queue_intfd_close (mq);
+  vec_foreach (mq, mif->rx_queues)
+    memif_queue_intfd_close (mq);
   vec_free (mif->rx_queues);
 
-  vec_foreach (mq, mif->tx_queues) memif_queue_intfd_close (mq);
+  vec_foreach (mq, mif->tx_queues)
+    memif_queue_intfd_close (mq);
   vec_free (mif->tx_queues);
 
   /* free memory regions */
   vec_foreach (mr, mif->regions)
-  {
-    int rv;
-    if ((rv = munmap (mr->shm, mr->region_size)))
-      clib_warning ("munmap failed, rv = %d", rv);
-    if (mr->fd > -1)
-      close (mr->fd);
-  }
+    {
+      int rv;
+      if (mr->is_external)
+	continue;
+      if ((rv = munmap (mr->shm, mr->region_size)))
+	clib_warning ("munmap failed, rv = %d", rv);
+      if (mr->fd > -1)
+	close (mr->fd);
+    }
+  /* *INDENT-ON* */
   vec_free (mif->regions);
   vec_free (mif->remote_name);
   vec_free (mif->remote_if_name);
@@ -184,66 +190,70 @@ memif_connect (memif_if_t * mif)
   vec_free (mif->local_disc_string);
   vec_free (mif->remote_disc_string);
 
+  /* *INDENT-OFF* */
   vec_foreach (mr, mif->regions)
-  {
-    if (mr->shm)
-      continue;
+    {
+      if (mr->shm)
+	continue;
 
-    if (mr->fd < 0)
-      clib_error_return (0, "no memory region fd");
+      if (mr->fd < 0)
+	clib_error_return (0, "no memory region fd");
 
-    if ((mr->shm = mmap (NULL, mr->region_size, PROT_READ | PROT_WRITE,
-			 MAP_SHARED, mr->fd, 0)) == MAP_FAILED)
-      return clib_error_return_unix (0, "mmap");
-  }
+      if ((mr->shm = mmap (NULL, mr->region_size, PROT_READ | PROT_WRITE,
+			   MAP_SHARED, mr->fd, 0)) == MAP_FAILED)
+	return clib_error_return_unix (0, "mmap");
+    }
+  /* *INDENT-ON* */
 
   template.read_function = memif_int_fd_read_ready;
 
+  /* *INDENT-OFF* */
   vec_foreach_index (i, mif->tx_queues)
-  {
-    memif_queue_t *mq = vec_elt_at_index (mif->tx_queues, i);
+    {
+      memif_queue_t *mq = vec_elt_at_index (mif->tx_queues, i);
 
-    mq->ring = mif->regions[mq->region].shm + mq->offset;
-    if (mq->ring->cookie != MEMIF_COOKIE)
-      return clib_error_return (0, "wrong cookie on tx ring %u", i);
-  }
+      mq->ring = mif->regions[mq->region].shm + mq->offset;
+      if (mq->ring->cookie != MEMIF_COOKIE)
+	return clib_error_return (0, "wrong cookie on tx ring %u", i);
+    }
 
   vec_foreach_index (i, mif->rx_queues)
-  {
-    memif_queue_t *mq = vec_elt_at_index (mif->rx_queues, i);
-    int rv;
-
-    mq->ring = mif->regions[mq->region].shm + mq->offset;
-    if (mq->ring->cookie != MEMIF_COOKIE)
-      return clib_error_return (0, "wrong cookie on tx ring %u", i);
-
-    if (mq->int_fd > -1)
-      {
-	template.file_descriptor = mq->int_fd;
-	template.private_data = (mif->dev_instance << 16) | (i & 0xFFFF);
-	template.description = format (0, "%U rx %u int",
-				       format_memif_device_name,
-				       mif->dev_instance, i);
-	memif_file_add (&mq->int_clib_file_index, &template);
-      }
-    vnet_hw_interface_assign_rx_thread (vnm, mif->hw_if_index, i, ~0);
-    rv = vnet_hw_interface_set_rx_mode (vnm, mif->hw_if_index, i,
-					VNET_HW_INTERFACE_RX_MODE_DEFAULT);
-    if (rv)
-      clib_warning
-	("Warning: unable to set rx mode for interface %d queue %d: "
-	 "rc=%d", mif->hw_if_index, i, rv);
-    else
-      {
-	vnet_hw_interface_rx_mode rxmode;
-	vnet_hw_interface_get_rx_mode (vnm, mif->hw_if_index, i, &rxmode);
-
-	if (rxmode == VNET_HW_INTERFACE_RX_MODE_POLLING)
-	  mq->ring->flags |= MEMIF_RING_FLAG_MASK_INT;
-	else
-	  vnet_device_input_set_interrupt_pending (vnm, mif->hw_if_index, i);
-      }
-  }
+    {
+      memif_queue_t *mq = vec_elt_at_index (mif->rx_queues, i);
+      int rv;
+
+      mq->ring = mif->regions[mq->region].shm + mq->offset;
+      if (mq->ring->cookie != MEMIF_COOKIE)
+	return clib_error_return (0, "wrong cookie on tx ring %u", i);
+
+      if (mq->int_fd > -1)
+	{
+	  template.file_descriptor = mq->int_fd;
+	  template.private_data = (mif->dev_instance << 16) | (i & 0xFFFF);
+	  template.description = format (0, "%U rx %u int",
+					 format_memif_device_name,
+					 mif->dev_instance, i);
+	  memif_file_add (&mq->int_clib_file_index, &template);
+	}
+      vnet_hw_interface_assign_rx_thread (vnm, mif->hw_if_index, i, ~0);
+      rv = vnet_hw_interface_set_rx_mode (vnm, mif->hw_if_index, i,
+					  VNET_HW_INTERFACE_RX_MODE_DEFAULT);
+      if (rv)
+	clib_warning
+	  ("Warning: unable to set rx mode for interface %d queue %d: "
+	   "rc=%d", mif->hw_if_index, i, rv);
+      else
+	{
+	  vnet_hw_interface_rx_mode rxmode;
+	  vnet_hw_interface_get_rx_mode (vnm, mif->hw_if_index, i, &rxmode);
+
+	  if (rxmode == VNET_HW_INTERFACE_RX_MODE_POLLING)
+	    mq->ring->flags |= MEMIF_RING_FLAG_MASK_INT;
+	  else
+	    vnet_device_input_set_interrupt_pending (vnm, mif->hw_if_index, i);
+	}
+    }
+  /* *INDENT-ON* */
 
   mif->flags &= ~MEMIF_IF_FLAG_CONNECTING;
   mif->flags |= MEMIF_IF_FLAG_CONNECTED;
@@ -270,6 +280,7 @@ memif_get_ring (memif_if_t * mif, memif_ring_type_t type, u16 ring_num)
 clib_error_t *
 memif_init_regions_and_queues (memif_if_t * mif)
 {
+  vlib_main_t *vm = vlib_get_main ();
   memif_ring_t *ring = NULL;
   int i, j;
   u64 buffer_offset;
@@ -277,16 +288,18 @@ memif_init_regions_and_queues (memif_if_t * mif)
   clib_mem_vm_alloc_t alloc = { 0 };
   clib_error_t *err;
 
-  vec_validate_aligned (mif->regions, 0, CLIB_CACHE_LINE_BYTES);
-  r = vec_elt_at_index (mif->regions, 0);
+  ASSERT (vec_len (mif->regions) == 0);
+  vec_add2_aligned (mif->regions, r, 1, CLIB_CACHE_LINE_BYTES);
 
   buffer_offset = (mif->run.num_s2m_rings + mif->run.num_m2s_rings) *
     (sizeof (memif_ring_t) +
      sizeof (memif_desc_t) * (1 << mif->run.log2_ring_size));
 
-  r->region_size = buffer_offset +
-    mif->run.buffer_size * (1 << mif->run.log2_ring_size) *
-    (mif->run.num_s2m_rings + mif->run.num_m2s_rings);
+  r->region_size = buffer_offset;
+
+  if ((mif->flags & MEMIF_IF_FLAG_ZERO_COPY) == 0)
+    r->region_size += mif->run.buffer_size * (1 << mif->run.log2_ring_size) *
+      (mif->run.num_s2m_rings + mif->run.num_m2s_rings);
 
   alloc.name = "memif region";
   alloc.size = r->region_size;
@@ -299,11 +312,32 @@ memif_init_regions_and_queues (memif_if_t * mif)
   r->fd = alloc.fd;
   r->shm = alloc.addr;
 
+  if (mif->flags & MEMIF_IF_FLAG_ZERO_COPY)
+    {
+      vlib_buffer_pool_t *bp;
+      /* *INDENT-OFF* */
+      vec_foreach (bp, buffer_main.buffer_pools)
+	{
+	  vlib_physmem_region_t *pr;
+	  pr = vlib_physmem_get_region (vm, bp->physmem_region);
+	  vec_add2_aligned (mif->regions, r, 1, CLIB_CACHE_LINE_BYTES);
+	  r->fd = pr->fd;
+	  r->region_size = pr->size;
+	  r->shm = pr->mem;
+	  r->is_external = 1;
+	}
+      /* *INDENT-ON* */
+    }
+
   for (i = 0; i < mif->run.num_s2m_rings; i++)
     {
       ring = memif_get_ring (mif, MEMIF_RING_S2M, i);
       ring->head = ring->tail = 0;
       ring->cookie = MEMIF_COOKIE;
+
+      if (mif->flags & MEMIF_IF_FLAG_ZERO_COPY)
+	continue;
+
       for (j = 0; j < (1 << mif->run.log2_ring_size); j++)
 	{
 	  u16 slot = i * (1 << mif->run.log2_ring_size) + j;
@@ -318,6 +352,10 @@ memif_init_regions_and_queues (memif_if_t * mif)
       ring = memif_get_ring (mif, MEMIF_RING_M2S, i);
       ring->head = ring->tail = 0;
       ring->cookie = MEMIF_COOKIE;
+
+      if (mif->flags & MEMIF_IF_FLAG_ZERO_COPY)
+	continue;
+
       for (j = 0; j < (1 << mif->run.log2_ring_size); j++)
 	{
 	  u16 slot =
@@ -332,36 +370,48 @@ memif_init_regions_and_queues (memif_if_t * mif)
   ASSERT (mif->tx_queues == 0);
   vec_validate_aligned (mif->tx_queues, mif->run.num_s2m_rings - 1,
 			CLIB_CACHE_LINE_BYTES);
+
+  /* *INDENT-OFF* */
   vec_foreach_index (i, mif->tx_queues)
-  {
-    memif_queue_t *mq = vec_elt_at_index (mif->tx_queues, i);
-    if ((mq->int_fd = eventfd (0, EFD_NONBLOCK)) < 0)
-      return clib_error_return_unix (0, "eventfd[tx queue %u]", i);
-    mq->int_clib_file_index = ~0;
-    mq->ring = memif_get_ring (mif, MEMIF_RING_S2M, i);
-    mq->log2_ring_size = mif->cfg.log2_ring_size;
-    mq->region = 0;
-    mq->offset = (void *) mq->ring - (void *) mif->regions[mq->region].shm;
-    mq->last_head = 0;
-    mq->type = MEMIF_RING_S2M;
-  }
+    {
+      memif_queue_t *mq = vec_elt_at_index (mif->tx_queues, i);
+      if ((mq->int_fd = eventfd (0, EFD_NONBLOCK)) < 0)
+	return clib_error_return_unix (0, "eventfd[tx queue %u]", i);
+      mq->int_clib_file_index = ~0;
+      mq->ring = memif_get_ring (mif, MEMIF_RING_S2M, i);
+      mq->log2_ring_size = mif->cfg.log2_ring_size;
+      mq->region = 0;
+      mq->offset = (void *) mq->ring - (void *) mif->regions[mq->region].shm;
+      mq->last_head = 0;
+      mq->type = MEMIF_RING_S2M;
+      if (mif->flags & MEMIF_IF_FLAG_ZERO_COPY)
+	vec_validate_aligned (mq->buffers, 1 << mq->log2_ring_size,
+			      CLIB_CACHE_LINE_BYTES);
+    }
+  /* *INDENT-ON* */
 
   ASSERT (mif->rx_queues == 0);
   vec_validate_aligned (mif->rx_queues, mif->run.num_m2s_rings - 1,
 			CLIB_CACHE_LINE_BYTES);
+
+  /* *INDENT-OFF* */
   vec_foreach_index (i, mif->rx_queues)
-  {
-    memif_queue_t *mq = vec_elt_at_index (mif->rx_queues, i);
-    if ((mq->int_fd = eventfd (0, EFD_NONBLOCK)) < 0)
-      return clib_error_return_unix (0, "eventfd[rx queue %u]", i);
-    mq->int_clib_file_index = ~0;
-    mq->ring = memif_get_ring (mif, MEMIF_RING_M2S, i);
-    mq->log2_ring_size = mif->cfg.log2_ring_size;
-    mq->region = 0;
-    mq->offset = (void *) mq->ring - (void *) mif->regions[mq->region].shm;
-    mq->last_head = 0;
-    mq->type = MEMIF_RING_M2S;
-  }
+    {
+      memif_queue_t *mq = vec_elt_at_index (mif->rx_queues, i);
+      if ((mq->int_fd = eventfd (0, EFD_NONBLOCK)) < 0)
+	return clib_error_return_unix (0, "eventfd[rx queue %u]", i);
+      mq->int_clib_file_index = ~0;
+      mq->ring = memif_get_ring (mif, MEMIF_RING_M2S, i);
+      mq->log2_ring_size = mif->cfg.log2_ring_size;
+      mq->region = 0;
+      mq->offset = (void *) mq->ring - (void *) mif->regions[mq->region].shm;
+      mq->last_head = 0;
+      mq->type = MEMIF_RING_M2S;
+      if (mif->flags & MEMIF_IF_FLAG_ZERO_COPY)
+	vec_validate_aligned (mq->buffers, 1 << mq->log2_ring_size,
+			      CLIB_CACHE_LINE_BYTES);
+    }
+  /* *INDENT-ON* */
 
   return 0;
 }
@@ -616,10 +666,10 @@ memif_delete_if (vlib_main_t * vm, memif_if_t * mif)
       if (msf->is_listener)
 	{
 	  int i;
+	  /* *INDENT-OFF* */
 	  vec_foreach_index (i, msf->pending_clients)
-	  {
 	    memif_socket_close (msf->pending_clients + i);
-	  }
+	  /* *INDENT-ON* */
 	  memif_socket_close (&msf->sock);
 	  vec_free (msf->pending_clients);
 	}
@@ -854,7 +904,11 @@ memif_create_if (vlib_main_t * vm, memif_create_if_args_t * args)
   msf->ref_cnt++;
 
   if (args->is_master == 0)
-    mif->flags |= MEMIF_IF_FLAG_IS_SLAVE;
+    {
+      mif->flags |= MEMIF_IF_FLAG_IS_SLAVE;
+      if (args->is_zero_copy)
+	mif->flags |= MEMIF_IF_FLAG_ZERO_COPY;
+    }
 
   hw = vnet_get_hw_interface (vnm, mif->hw_if_index);
   hw->flags |= VNET_HW_INTERFACE_FLAG_SUPPORTS_INT_MODE;
diff --git a/src/plugins/memif/node.c b/src/plugins/memif/node.c
index e8223d12435..e1ee63b47bb 100644
--- a/src/plugins/memif/node.c
+++ b/src/plugins/memif/node.c
@@ -141,6 +141,35 @@ memif_add_to_chain (vlib_main_t * vm, vlib_buffer_t * b, u32 * buffers,
     }
 }
 
+static_always_inline u32
+sat_sub (u32 x, u32 y)
+{
+  u32 res = x - y;
+  res &= -(res <= x);
+  return res;
+}
+
+/* branchless validation of the descriptor - uses saturated subtraction */
+static_always_inline u32
+memif_desc_is_invalid (memif_if_t * mif, memif_desc_t * d, u32 buffer_length)
+{
+  u32 rv;
+  u16 valid_flags = MEMIF_DESC_FLAG_NEXT;
+
+  rv = d->flags & (~valid_flags);
+  rv |= sat_sub (d->region + 1, vec_len (mif->regions));
+  rv |= sat_sub (d->length, buffer_length);
+  rv |= sat_sub (d->offset + d->length, mif->regions[d->region].region_size);
+
+  if (PREDICT_FALSE (rv))
+    {
+      mif->flags |= MEMIF_IF_FLAG_ERROR;
+      return 1;
+    }
+
+  return 0;
+}
+
 static_always_inline uword
 memif_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 			   vlib_frame_t * frame, memif_if_t * mif,
@@ -182,7 +211,6 @@ memif_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
      so start with IP header at offset 14 */
   start_offset = (mode == MEMIF_INTERFACE_MODE_IP) ? 14 : 0;
 
-
   /* for S2M rings, we are consumers of packet buffers, and for M2S rings we
      are producers of empty buffers */
   cur_slot = (type == MEMIF_RING_S2M) ? mq->last_head : mq->last_tail;
@@ -492,6 +520,289 @@ refill:
   return n_rx_packets;
 }
 
+static_always_inline uword
+memif_device_input_zc_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
+			      vlib_frame_t * frame, memif_if_t * mif,
+			      u16 qid, memif_interface_mode_t mode)
+{
+  vnet_main_t *vnm = vnet_get_main ();
+  memif_main_t *mm = &memif_main;
+  memif_ring_t *ring;
+  memif_queue_t *mq;
+  u32 next_index;
+  uword n_trace = vlib_get_trace_count (vm, node);
+  u32 n_rx_packets = 0, n_rx_bytes = 0;
+  u32 *to_next = 0, *buffers;
+  u32 bi0, bi1, bi2, bi3;
+  vlib_buffer_t *b0, *b1, *b2, *b3;
+  u32 thread_index = vlib_get_thread_index ();
+  memif_per_thread_data_t *ptd = vec_elt_at_index (mm->per_thread_data,
+						   thread_index);
+  vlib_buffer_t *bt = &ptd->buffer_template;
+  u16 cur_slot, last_slot, ring_size, n_slots, mask, head;
+  i16 start_offset;
+  u32 buffer_length;
+  u16 n_alloc;
+
+  mq = vec_elt_at_index (mif->rx_queues, qid);
+  ring = mq->ring;
+  ring_size = 1 << mq->log2_ring_size;
+  mask = ring_size - 1;
+
+  next_index = (mode == MEMIF_INTERFACE_MODE_IP) ?
+    VNET_DEVICE_INPUT_NEXT_IP6_INPUT : VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT;
+
+  /* asume that somebody will want to add ethernet header on the packet
+     so start with IP header at offset 14 */
+  start_offset = (mode == MEMIF_INTERFACE_MODE_IP) ? 14 : 0;
+  buffer_length = VLIB_BUFFER_DEFAULT_FREE_LIST_BYTES - start_offset;
+
+  cur_slot = mq->last_tail;
+  last_slot = ring->tail;
+  if (cur_slot == last_slot)
+    goto refill;
+  n_slots = last_slot - cur_slot;
+
+  /* process ring slots */
+  vec_validate_aligned (ptd->buffers, MEMIF_RX_VECTOR_SZ,
+			CLIB_CACHE_LINE_BYTES);
+  while (n_slots && n_rx_packets < MEMIF_RX_VECTOR_SZ)
+    {
+      u16 s0;
+      memif_desc_t *d0;
+      vlib_buffer_t *hb;
+
+      s0 = cur_slot & mask;
+      bi0 = mq->buffers[s0];
+      ptd->buffers[n_rx_packets++] = bi0;
+
+      CLIB_PREFETCH (&ring->desc[(cur_slot + 8) & mask],
+		     CLIB_CACHE_LINE_BYTES, LOAD);
+      d0 = &ring->desc[s0];
+      hb = b0 = vlib_get_buffer (vm, bi0);
+      b0->current_data = start_offset;
+      b0->current_length = start_offset + d0->length;
+
+
+      if (0 && memif_desc_is_invalid (mif, d0, buffer_length))
+	return 0;
+
+      cur_slot++;
+      n_slots--;
+      if (PREDICT_FALSE ((d0->flags & MEMIF_DESC_FLAG_NEXT) && n_slots))
+	{
+	  hb->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
+	next_slot:
+	  s0 = cur_slot & mask;
+	  d0 = &ring->desc[s0];
+	  bi0 = mq->buffers[s0];
+
+	  /*previous buffer */
+	  b0->next_buffer = bi0;
+	  b0->flags |= VLIB_BUFFER_NEXT_PRESENT;
+
+	  /* current buffer */
+	  b0 = vlib_get_buffer (vm, bi0);
+	  b0->current_data = start_offset;
+	  b0->current_length = start_offset + d0->length;
+	  hb->total_length_not_including_first_buffer += d0->length;
+
+	  cur_slot++;
+	  n_slots--;
+	  if ((d0->flags & MEMIF_DESC_FLAG_NEXT) && n_slots)
+	    goto next_slot;
+	}
+    }
+
+  /* release slots from the ring */
+  mq->last_tail = cur_slot;
+
+  u32 n_from = n_rx_packets;
+
+  vnet_buffer (bt)->sw_if_index[VLIB_RX] = mif->sw_if_index;
+
+  buffers = ptd->buffers;
+
+  while (n_from)
+    {
+      u32 n_left_to_next;
+      u32 next0, next1, next2, next3;
+
+      vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
+      while (n_from >= 8 && n_left_to_next >= 4)
+	{
+	  b0 = vlib_get_buffer (vm, buffers[4]);
+	  b1 = vlib_get_buffer (vm, buffers[5]);
+	  b2 = vlib_get_buffer (vm, buffers[6]);
+	  b3 = vlib_get_buffer (vm, buffers[7]);
+	  vlib_prefetch_buffer_header (b0, STORE);
+	  vlib_prefetch_buffer_header (b1, STORE);
+	  vlib_prefetch_buffer_header (b2, STORE);
+	  vlib_prefetch_buffer_header (b3, STORE);
+
+	  /* enqueue buffer */
+	  to_next[0] = bi0 = buffers[0];
+	  to_next[1] = bi1 = buffers[1];
+	  to_next[2] = bi2 = buffers[2];
+	  to_next[3] = bi3 = buffers[3];
+	  to_next += 4;
+	  n_left_to_next -= 4;
+	  buffers += 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);
+
+	  vnet_buffer (b0)->sw_if_index[VLIB_RX] = mif->sw_if_index;
+	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
+	  vnet_buffer (b1)->sw_if_index[VLIB_RX] = mif->sw_if_index;
+	  vnet_buffer (b1)->sw_if_index[VLIB_TX] = ~0;
+	  vnet_buffer (b2)->sw_if_index[VLIB_RX] = mif->sw_if_index;
+	  vnet_buffer (b2)->sw_if_index[VLIB_TX] = ~0;
+	  vnet_buffer (b3)->sw_if_index[VLIB_RX] = mif->sw_if_index;
+	  vnet_buffer (b3)->sw_if_index[VLIB_TX] = ~0;
+
+	  if (mode == MEMIF_INTERFACE_MODE_IP)
+	    {
+	      next0 = memif_next_from_ip_hdr (node, b0);
+	      next1 = memif_next_from_ip_hdr (node, b1);
+	      next2 = memif_next_from_ip_hdr (node, b2);
+	      next3 = memif_next_from_ip_hdr (node, b3);
+	    }
+	  else if (mode == MEMIF_INTERFACE_MODE_ETHERNET)
+	    {
+	      if (PREDICT_FALSE (mif->per_interface_next_index != ~0))
+		{
+		  next0 = mif->per_interface_next_index;
+		  next1 = mif->per_interface_next_index;
+		  next2 = mif->per_interface_next_index;
+		  next3 = mif->per_interface_next_index;
+		}
+	      else
+		{
+		  next0 = next1 = next2 = next3 = next_index;
+		  /* redirect if feature path enabled */
+		  vnet_feature_start_device_input_x1 (mif->sw_if_index,
+						      &next0, b0);
+		  vnet_feature_start_device_input_x1 (mif->sw_if_index,
+						      &next1, b1);
+		  vnet_feature_start_device_input_x1 (mif->sw_if_index,
+						      &next2, b2);
+		  vnet_feature_start_device_input_x1 (mif->sw_if_index,
+						      &next3, b3);
+		}
+	    }
+
+	  /* trace */
+	  if (PREDICT_FALSE (n_trace > 0))
+	    {
+	      memif_trace_buffer (vm, node, mif, b0, next0, qid, &n_trace);
+	      if (PREDICT_FALSE (n_trace > 0))
+		memif_trace_buffer (vm, node, mif, b1, next1, qid, &n_trace);
+	      if (PREDICT_FALSE (n_trace > 0))
+		memif_trace_buffer (vm, node, mif, b2, next2, qid, &n_trace);
+	      if (PREDICT_FALSE (n_trace > 0))
+		memif_trace_buffer (vm, node, mif, b3, next3, qid, &n_trace);
+	    }
+
+	  /* enqueue */
+	  vlib_validate_buffer_enqueue_x4 (vm, node, next_index, to_next,
+					   n_left_to_next, bi0, bi1, bi2, bi3,
+					   next0, next1, next2, next3);
+
+	  /* next */
+	  n_from -= 4;
+	}
+      while (n_from && n_left_to_next)
+	{
+	  /* enqueue buffer */
+	  to_next[0] = bi0 = buffers[0];
+	  to_next += 1;
+	  n_left_to_next--;
+	  buffers += 1;
+
+	  b0 = vlib_get_buffer (vm, bi0);
+	  vnet_buffer (b0)->sw_if_index[VLIB_RX] = mif->sw_if_index;
+	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
+
+	  if (mode == MEMIF_INTERFACE_MODE_IP)
+	    {
+	      next0 = memif_next_from_ip_hdr (node, b0);
+	    }
+	  else if (mode == MEMIF_INTERFACE_MODE_ETHERNET)
+	    {
+	      if (PREDICT_FALSE (mif->per_interface_next_index != ~0))
+		next0 = mif->per_interface_next_index;
+	      else
+		{
+		  next0 = next_index;
+		  /* redirect if feature path enabled */
+		  vnet_feature_start_device_input_x1 (mif->sw_if_index,
+						      &next0, b0);
+		}
+	    }
+
+	  /* trace */
+	  if (PREDICT_FALSE (n_trace > 0))
+	    memif_trace_buffer (vm, node, mif, b0, next0, qid, &n_trace);
+
+	  /* enqueue */
+	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
+					   n_left_to_next, bi0, next0);
+
+	  /* next */
+	  n_from--;
+	}
+      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+    }
+
+  vlib_increment_combined_counter (vnm->interface_main.combined_sw_if_counters
+				   + VNET_INTERFACE_COUNTER_RX, thread_index,
+				   mif->hw_if_index, n_rx_packets,
+				   n_rx_bytes);
+
+  /* refill ring with empty buffers */
+refill:
+  vec_reset_length (ptd->buffers);
+
+  head = ring->head;
+  n_slots = ring_size - head + mq->last_tail;
+
+  if (n_slots < 8)
+    goto done;
+
+  memif_desc_t *dt = &ptd->desc_template;
+  memset (dt, 0, sizeof (memif_desc_t));
+  dt->length = VLIB_BUFFER_DEFAULT_FREE_LIST_BYTES - start_offset;
+
+  n_alloc = vlib_buffer_alloc_to_ring (vm, mq->buffers, head & mask,
+				       ring_size, n_slots);
+
+  if (PREDICT_FALSE (n_alloc != n_slots))
+    {
+      vlib_error_count (vm, node->node_index,
+			MEMIF_INPUT_ERROR_BUFFER_ALLOC_FAIL, 1);
+    }
+
+  while (n_alloc--)
+    {
+      u16 s = head++ & mask;
+      memif_desc_t *d = &ring->desc[s];
+      clib_memcpy (d, dt, sizeof (memif_desc_t));
+      b0 = vlib_get_buffer (vm, mq->buffers[s]);
+      d->region = b0->buffer_pool_index + 1;
+      d->offset =
+	(void *) b0->data - mif->regions[d->region].shm + start_offset;
+    }
+
+  CLIB_MEMORY_STORE_BARRIER ();
+  ring->head = head;
+
+done:
+  return n_rx_packets;
+}
+
 uword
 CLIB_MULTIARCH_FN (memif_input_fn) (vlib_main_t * vm,
 				    vlib_node_runtime_t * node,
@@ -501,6 +812,8 @@ CLIB_MULTIARCH_FN (memif_input_fn) (vlib_main_t * vm,
   memif_main_t *mm = &memif_main;
   vnet_device_input_runtime_t *rt = (void *) node->runtime_data;
   vnet_device_and_queue_t *dq;
+  memif_interface_mode_t mode_ip = MEMIF_INTERFACE_MODE_IP;
+  memif_interface_mode_t mode_eth = MEMIF_INTERFACE_MODE_ETHERNET;
 
   foreach_device_and_queue (dq, rt->devices_and_queues)
   {
@@ -509,27 +822,36 @@ CLIB_MULTIARCH_FN (memif_input_fn) (vlib_main_t * vm,
     if ((mif->flags & MEMIF_IF_FLAG_ADMIN_UP) &&
 	(mif->flags & MEMIF_IF_FLAG_CONNECTED))
       {
+	if (mif->flags & MEMIF_IF_FLAG_ZERO_COPY)
+	  {
+	    if (mif->mode == MEMIF_INTERFACE_MODE_IP)
+	      n_rx += memif_device_input_zc_inline (vm, node, frame, mif,
+						    dq->queue_id, mode_ip);
+	    else
+	      n_rx += memif_device_input_zc_inline (vm, node, frame, mif,
+						    dq->queue_id, mode_eth);
+	  }
 	if (mif->flags & MEMIF_IF_FLAG_IS_SLAVE)
 	  {
 	    if (mif->mode == MEMIF_INTERFACE_MODE_IP)
 	      n_rx += memif_device_input_inline (vm, node, frame, mif,
 						 MEMIF_RING_M2S, dq->queue_id,
-						 MEMIF_INTERFACE_MODE_IP);
+						 mode_ip);
 	    else
 	      n_rx += memif_device_input_inline (vm, node, frame, mif,
 						 MEMIF_RING_M2S, dq->queue_id,
-						 MEMIF_INTERFACE_MODE_ETHERNET);
+						 mode_eth);
 	  }
 	else
 	  {
 	    if (mif->mode == MEMIF_INTERFACE_MODE_IP)
 	      n_rx += memif_device_input_inline (vm, node, frame, mif,
 						 MEMIF_RING_S2M, dq->queue_id,
-						 MEMIF_INTERFACE_MODE_IP);
+						 mode_ip);
 	    else
 	      n_rx += memif_device_input_inline (vm, node, frame, mif,
 						 MEMIF_RING_S2M, dq->queue_id,
-						 MEMIF_INTERFACE_MODE_ETHERNET);
+						 mode_eth);
 	  }
       }
   }
diff --git a/src/plugins/memif/private.h b/src/plugins/memif/private.h
index f4ace4cee5d..ad6295e4bb8 100644
--- a/src/plugins/memif/private.h
+++ b/src/plugins/memif/private.h
@@ -83,6 +83,7 @@ typedef struct
   void *shm;
   memif_region_size_t region_size;
   int fd;
+  u8 is_external;
 } memif_region_t;
 
 typedef struct
@@ -101,6 +102,7 @@ typedef struct
 
   u16 last_head;
   u16 last_tail;
+  u32 *buffers;
 
   /* interrupts */
   int int_fd;
@@ -116,7 +118,9 @@ typedef struct
   _(1, IS_SLAVE, "slave")		\
   _(2, CONNECTING, "connecting")	\
   _(3, CONNECTED, "connected")		\
-  _(4, DELETING, "deleting")
+  _(4, DELETING, "deleting")		\
+  _(5, ZERO_COPY, "zero-copy")		\
+  _(6, ERROR, "error")
 
 typedef enum
 {
@@ -201,6 +205,7 @@ typedef struct
 
   /* buffer template */
   vlib_buffer_t buffer_template;
+  memif_desc_t desc_template;
 } memif_per_thread_data_t;
 
 typedef struct
@@ -238,6 +243,7 @@ typedef struct
   u32 socket_id;
   u8 *secret;
   u8 is_master;
+  u8 is_zero_copy;
   memif_interface_mode_t mode:8;
   memif_log2_ring_size_t log2_ring_size;
   u16 buffer_size;
diff --git a/src/plugins/memif/socket.c b/src/plugins/memif/socket.c
index 34bff166d79..39ff99e3d8d 100644
--- a/src/plugins/memif/socket.c
+++ b/src/plugins/memif/socket.c
@@ -437,11 +437,14 @@ memif_msg_receive (memif_if_t ** mifp, clib_socket_t * sock, clib_file_t * uf)
       if ((err = memif_init_regions_and_queues (mif)))
 	return err;
       memif_msg_enq_init (mif);
-      memif_msg_enq_add_region (mif, 0);
+      /* *INDENT-OFF* */
+      vec_foreach_index (i, mif->regions)
+	memif_msg_enq_add_region (mif, i);
       vec_foreach_index (i, mif->tx_queues)
 	memif_msg_enq_add_ring (mif, i, MEMIF_RING_S2M);
       vec_foreach_index (i, mif->rx_queues)
 	memif_msg_enq_add_ring (mif, i, MEMIF_RING_M2S);
+      /* *INDENT-ON* */
       memif_msg_enq_connect (mif);
       break;