devices: af-packet v3 support

Type: feature

CPU usage ~20% less than v2.
Performance improvement 20% more than v2.
High vector rate.

Change-Id: I24bc594200f42664b59d07b44d44578e61068bbc
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>

4 files changed, 274 insertions, 198 deletions

diff --git a/src/vnet/devices/af_packet/af_packet.c b/src/vnet/devices/af_packet/af_packet.c
index bc9caac82f2..60eadf2b0dc 100644
--- a/src/vnet/devices/af_packet/af_packet.c
+++ b/src/vnet/devices/af_packet/af_packet.c
@@ -47,15 +47,13 @@ VNET_HW_INTERFACE_CLASS (af_packet_ip_device_hw_interface_class, static) = {
 #define AF_PACKET_DEFAULT_TX_FRAME_SIZE	      (2048 * 5)
 #define AF_PACKET_TX_BLOCK_NR		1
 
-#define AF_PACKET_DEFAULT_RX_FRAMES_PER_BLOCK 1024
-#define AF_PACKET_DEFAULT_RX_FRAME_SIZE	      (2048 * 5)
-#define AF_PACKET_RX_BLOCK_NR		1
+#define AF_PACKET_DEFAULT_RX_FRAMES_PER_BLOCK 256
+#define AF_PACKET_DEFAULT_RX_FRAME_SIZE	      2048
+#define AF_PACKET_RX_BLOCK_NR		      20
 
 /*defined in net/if.h but clashes with dpdk headers */
 unsigned int if_nametoindex (const char *ifname);
 
-typedef struct tpacket_req tpacket_req_t;
-
 static clib_error_t *
 af_packet_eth_set_max_frame_size (vnet_main_t *vnm, vnet_hw_interface_t *hi,
 				  u32 frame_size)
@@ -130,14 +128,17 @@ is_bridge (const u8 * host_if_name)
 }
 
 static int
-create_packet_v2_sock (int host_if_index, tpacket_req_t * rx_req,
-		       tpacket_req_t * tx_req, int *fd, u8 ** ring)
+create_packet_v3_sock (int host_if_index, tpacket_req3_t *rx_req,
+		       tpacket_req3_t *tx_req, int *fd, u8 **ring,
+		       u32 *hdrlen_ptr)
 {
   af_packet_main_t *apm = &af_packet_main;
-  int ret;
   struct sockaddr_ll sll;
-  int ver = TPACKET_V2;
-  socklen_t req_sz = sizeof (struct tpacket_req);
+  socklen_t req_sz = sizeof (tpacket_req3_t);
+  int ret;
+  int ver = TPACKET_V3;
+  u32 hdrlen = 0;
+  u32 len = sizeof (hdrlen);
   u32 ring_sz = rx_req->tp_block_size * rx_req->tp_block_nr +
     tx_req->tp_block_size * tx_req->tp_block_nr;
 
@@ -166,19 +167,33 @@ create_packet_v2_sock (int host_if_index, tpacket_req_t * rx_req,
 
   if (setsockopt (*fd, SOL_PACKET, PACKET_VERSION, &ver, sizeof (ver)) < 0)
     {
-      vlib_log_debug (apm->log_class,
-		      "Failed to set rx packet interface version: %s (errno %d)",
-		      strerror (errno), errno);
+      vlib_log_debug (
+	apm->log_class,
+	"Failed to set rx packet interface version: %s (errno %d)",
+	strerror (errno), errno);
       ret = VNET_API_ERROR_SYSCALL_ERROR_1;
       goto error;
     }
 
+  if (getsockopt (*fd, SOL_PACKET, PACKET_HDRLEN, &hdrlen, &len) < 0)
+    {
+      vlib_log_debug (
+	apm->log_class,
+	"Failed to get packet hdr len error handling option: %s (errno %d)",
+	strerror (errno), errno);
+      ret = VNET_API_ERROR_SYSCALL_ERROR_1;
+      goto error;
+    }
+  else
+    *hdrlen_ptr = hdrlen;
+
   int opt = 1;
   if (setsockopt (*fd, SOL_PACKET, PACKET_LOSS, &opt, sizeof (opt)) < 0)
     {
-      vlib_log_debug (apm->log_class,
-		      "Failed to set packet tx ring error handling option: %s (errno %d)",
-		      strerror (errno), errno);
+      vlib_log_debug (
+	apm->log_class,
+	"Failed to set packet tx ring error handling option: %s (errno %d)",
+	strerror (errno), errno);
       ret = VNET_API_ERROR_SYSCALL_ERROR_1;
       goto error;
     }
@@ -213,9 +228,8 @@ create_packet_v2_sock (int host_if_index, tpacket_req_t * rx_req,
       goto error;
     }
 
-  *ring =
-    mmap (NULL, ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_LOCKED, *fd,
-	  0);
+  *ring = mmap (NULL, ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_LOCKED,
+		*fd, 0);
   if (*ring == MAP_FAILED)
     {
       vlib_log_debug (apm->log_class, "mmap failure: %s (errno %d)",
@@ -240,8 +254,8 @@ af_packet_create_if (af_packet_create_if_arg_t *arg)
   af_packet_main_t *apm = &af_packet_main;
   vlib_main_t *vm = vlib_get_main ();
   int ret, fd = -1, fd2 = -1;
-  struct tpacket_req *rx_req = 0;
-  struct tpacket_req *tx_req = 0;
+  tpacket_req3_t *rx_req = 0;
+  tpacket_req3_t *tx_req = 0;
   struct ifreq ifr;
   u8 *ring = 0;
   af_packet_if_t *apif = 0;
@@ -255,6 +269,8 @@ af_packet_create_if (af_packet_create_if_arg_t *arg)
   int host_if_index = -1;
   u32 rx_frames_per_block, tx_frames_per_block;
   u32 rx_frame_size, tx_frame_size;
+  u32 hdrlen = 0;
+  u32 i = 0;
 
   p = mhash_get (&apm->if_index_by_host_if_name, arg->host_if_name);
   if (p)
@@ -282,12 +298,18 @@ af_packet_create_if (af_packet_create_if_arg_t *arg)
   rx_req->tp_frame_size = rx_frame_size;
   rx_req->tp_block_nr = AF_PACKET_RX_BLOCK_NR;
   rx_req->tp_frame_nr = AF_PACKET_RX_BLOCK_NR * rx_frames_per_block;
+  rx_req->tp_retire_blk_tov = 0;
+  rx_req->tp_feature_req_word = 0;
+  rx_req->tp_sizeof_priv = 0;
 
   vec_validate (tx_req, 0);
   tx_req->tp_block_size = tx_frame_size * tx_frames_per_block;
   tx_req->tp_frame_size = tx_frame_size;
   tx_req->tp_block_nr = AF_PACKET_TX_BLOCK_NR;
   tx_req->tp_frame_nr = AF_PACKET_TX_BLOCK_NR * tx_frames_per_block;
+  tx_req->tp_retire_blk_tov = 0;
+  tx_req->tp_sizeof_priv = 0;
+  tx_req->tp_feature_req_word = 0;
 
   /*
    * make sure host side of interface is 'UP' before binding AF_PACKET
@@ -343,7 +365,8 @@ af_packet_create_if (af_packet_create_if_arg_t *arg)
       fd2 = -1;
     }
 
-  ret = create_packet_v2_sock (host_if_index, rx_req, tx_req, &fd, &ring);
+  ret =
+    create_packet_v3_sock (host_if_index, rx_req, tx_req, &fd, &ring, &hdrlen);
 
   if (ret != 0)
     goto error;
@@ -359,15 +382,29 @@ af_packet_create_if (af_packet_create_if_arg_t *arg)
 
   apif->host_if_index = host_if_index;
   apif->fd = fd;
-  apif->rx_ring = ring;
-  apif->tx_ring = ring + rx_req->tp_block_size * rx_req->tp_block_nr;
+
+  vec_validate (apif->rx_ring, rx_req->tp_block_nr - 1);
+  vec_foreach_index (i, apif->rx_ring)
+    {
+      apif->rx_ring[i] = ring + i * rx_req->tp_block_size;
+    }
+
+  ring = ring + rx_req->tp_block_size * rx_req->tp_block_nr;
+
+  vec_validate (apif->tx_ring, tx_req->tp_block_nr - 1);
+  vec_foreach_index (i, apif->tx_ring)
+    {
+      apif->tx_ring[i] = ring + i * tx_req->tp_block_size;
+    }
+
   apif->rx_req = rx_req;
   apif->tx_req = tx_req;
   apif->host_if_name = host_if_name_dup;
   apif->per_interface_next_index = ~0;
   apif->next_tx_frame = 0;
-  apif->next_rx_frame = 0;
+  apif->next_rx_block = 0;
   apif->mode = arg->mode;
+  apif->hdrlen = hdrlen;
 
   ret = af_packet_read_mtu (apif);
   if (ret != 0)
diff --git a/src/vnet/devices/af_packet/af_packet.h b/src/vnet/devices/af_packet/af_packet.h
index d7d16b91c8c..ed3f10bb092 100644
--- a/src/vnet/devices/af_packet/af_packet.h
+++ b/src/vnet/devices/af_packet/af_packet.h
@@ -17,9 +17,15 @@
  *------------------------------------------------------------------
  */
 
+#include <linux/if_packet.h>
+
 #include <vppinfra/lock.h>
 #include <vlib/log.h>
 
+typedef struct tpacket_block_desc block_desc_t;
+typedef struct tpacket_req3 tpacket_req3_t;
+typedef struct tpacket3_hdr tpacket3_hdr_t;
+
 typedef enum
 {
   AF_PACKET_IF_MODE_ETHERNET = 1,
@@ -39,15 +45,16 @@ typedef struct
   u8 *host_if_name;
   int host_if_index;
   int fd;
-  struct tpacket_req *rx_req;
-  struct tpacket_req *tx_req;
-  u8 *rx_ring;
-  u8 *tx_ring;
+  tpacket_req3_t *rx_req;
+  tpacket_req3_t *tx_req;
+  u8 **rx_ring;
+  u8 **tx_ring;
+  u32 hdrlen;
   u32 hw_if_index;
   u32 sw_if_index;
   u32 clib_file_index;
 
-  u32 next_rx_frame;
+  u32 next_rx_block;
   u32 next_tx_frame;
 
   u32 per_interface_next_index;
diff --git a/src/vnet/devices/af_packet/device.c b/src/vnet/devices/af_packet/device.c
index c8e59c3566c..23b18832cc8 100644
--- a/src/vnet/devices/af_packet/device.c
+++ b/src/vnet/devices/af_packet/device.c
@@ -83,9 +83,9 @@ format_af_packet_device (u8 * s, va_list * args)
   u32 rx_frame_nr = apif->rx_req->tp_frame_nr;
   u32 rx_block_nr = apif->rx_req->tp_block_nr;
   int block = 0;
-  u8 *tx_block_start = apif->tx_ring + block * tx_block_sz;
+  u8 *tx_block_start = apif->tx_ring[block];
   u32 tx_frame = apif->next_tx_frame;
-  struct tpacket2_hdr *tph;
+  tpacket3_hdr_t *tph;
 
   s = format (s, "Linux PACKET socket interface\n");
   s = format (s, "%UTX block size:%d nr:%d  TX frame size:%d nr:%d\n",
@@ -100,7 +100,7 @@ format_af_packet_device (u8 * s, va_list * args)
   int n_send_req = 0, n_avail = 0, n_sending = 0, n_tot = 0, n_wrong = 0;
   do
     {
-      tph = (struct tpacket2_hdr *) (tx_block_start + tx_frame * tx_frame_sz);
+      tph = (struct tpacket3_hdr *) (tx_block_start + tx_frame * tx_frame_sz);
       tx_frame = (tx_frame + 1) % tx_frame_nr;
       if (tph->tp_status == 0)
 	n_avail++;
@@ -140,13 +140,12 @@ VNET_DEVICE_CLASS_TX_FN (af_packet_device_class) (vlib_main_t * vm,
   af_packet_if_t *apif =
     pool_elt_at_index (apm->interfaces, rd->dev_instance);
   clib_spinlock_lock_if_init (&apif->lockp);
-  int block = 0;
-  u32 block_size = apif->tx_req->tp_block_size;
+  u32 block = 0;
   u32 frame_size = apif->tx_req->tp_frame_size;
   u32 frame_num = apif->tx_req->tp_frame_nr;
-  u8 *block_start = apif->tx_ring + block * block_size;
+  u8 *block_start = apif->tx_ring[block];
   u32 tx_frame = apif->next_tx_frame;
-  struct tpacket2_hdr *tph;
+  tpacket3_hdr_t *tph;
   u32 frame_not_ready = 0;
 
   while (n_left)
@@ -158,7 +157,7 @@ VNET_DEVICE_CLASS_TX_FN (af_packet_device_class) (vlib_main_t * vm,
       u32 bi = buffers[0];
       buffers++;
 
-      tph = (struct tpacket2_hdr *) (block_start + tx_frame * frame_size);
+      tph = (struct tpacket3_hdr *) (block_start + tx_frame * frame_size);
       if (PREDICT_FALSE (tph->tp_status &
 			 (TP_STATUS_SEND_REQUEST | TP_STATUS_SENDING)))
 	{
@@ -170,9 +169,9 @@ VNET_DEVICE_CLASS_TX_FN (af_packet_device_class) (vlib_main_t * vm,
 	{
 	  b0 = vlib_get_buffer (vm, bi);
 	  len = b0->current_length;
-	  clib_memcpy_fast ((u8 *) tph +
-			    TPACKET_ALIGN (sizeof (struct tpacket2_hdr)) +
-			    offset, vlib_buffer_get_current (b0), len);
+	  clib_memcpy_fast (
+	    (u8 *) tph + TPACKET_ALIGN (sizeof (struct tpacket3_hdr)) + offset,
+	    vlib_buffer_get_current (b0), len);
 	  offset += len;
 	}
       while ((bi =
diff --git a/src/vnet/devices/af_packet/node.c b/src/vnet/devices/af_packet/node.c
index efe701642bb..229e0501af7 100644
--- a/src/vnet/devices/af_packet/node.c
+++ b/src/vnet/devices/af_packet/node.c
@@ -51,7 +51,10 @@ typedef struct
   u32 next_index;
   u32 hw_if_index;
   int block;
-  struct tpacket2_hdr tph;
+  u32 num_pkts;
+  void *block_start;
+  block_desc_t bd;
+  tpacket3_hdr_t tph;
 } af_packet_input_trace_t;
 
 static u8 *
@@ -65,26 +68,27 @@ format_af_packet_input_trace (u8 * s, va_list * args)
   s = format (s, "af_packet: hw_if_index %d next-index %d",
 	      t->hw_if_index, t->next_index);
 
+  s = format (s,
+	      "\n%Ublock %u:\n%Uaddress %p version %u seq_num %lu"
+	      " num_pkts %u",
+	      format_white_space, indent + 2, t->block, format_white_space,
+	      indent + 4, t->block_start, t->bd.version, t->bd.hdr.bh1.seq_num,
+	      t->num_pkts);
   s =
     format (s,
-	    "\n%Utpacket2_hdr:\n%Ustatus 0x%x len %u snaplen %u mac %u net %u"
+	    "\n%Utpacket3_hdr:\n%Ustatus 0x%x len %u snaplen %u mac %u net %u"
 	    "\n%Usec 0x%x nsec 0x%x vlan %U"
 #ifdef TP_STATUS_VLAN_TPID_VALID
 	    " vlan_tpid %u"
 #endif
 	    ,
-	    format_white_space, indent + 2,
-	    format_white_space, indent + 4,
-	    t->tph.tp_status,
-	    t->tph.tp_len,
-	    t->tph.tp_snaplen,
-	    t->tph.tp_mac,
-	    t->tph.tp_net,
-	    format_white_space, indent + 4,
-	    t->tph.tp_sec,
-	    t->tph.tp_nsec, format_ethernet_vlan_tci, t->tph.tp_vlan_tci
+	    format_white_space, indent + 2, format_white_space, indent + 4,
+	    t->tph.tp_status, t->tph.tp_len, t->tph.tp_snaplen, t->tph.tp_mac,
+	    t->tph.tp_net, format_white_space, indent + 4, t->tph.tp_sec,
+	    t->tph.tp_nsec, format_ethernet_vlan_tci, t->tph.hv1.tp_vlan_tci
 #ifdef TP_STATUS_VLAN_TPID_VALID
-	    , t->tph.tp_vlan_tpid
+	    ,
+	    t->tph.hv1.tp_vlan_tpid
 #endif
     );
   return s;
@@ -193,23 +197,23 @@ af_packet_device_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 			   vlib_frame_t * frame, af_packet_if_t * apif)
 {
   af_packet_main_t *apm = &af_packet_main;
-  struct tpacket2_hdr *tph;
+  tpacket3_hdr_t *tph;
   u32 next_index;
-  u32 block = 0;
-  u32 rx_frame;
   u32 n_free_bufs;
   u32 n_rx_packets = 0;
   u32 n_rx_bytes = 0;
   u32 *to_next = 0;
-  u32 block_size = apif->rx_req->tp_block_size;
-  u32 frame_size = apif->rx_req->tp_frame_size;
-  u32 frame_num = apif->rx_req->tp_frame_nr;
-  u8 *block_start = apif->rx_ring + block * block_size;
+  u32 block = apif->next_rx_block;
+  u32 block_nr = apif->rx_req->tp_block_nr;
+  u8 *block_start = 0;
   uword n_trace = vlib_get_trace_count (vm, node);
   u32 thread_index = vm->thread_index;
   u32 n_buffer_bytes = vlib_buffer_get_default_data_size (vm);
   u32 min_bufs = apif->rx_req->tp_frame_size / n_buffer_bytes;
   u32 eth_header_size = 0;
+  u32 num_pkts = 0;
+  u32 rx_frame_offset = 0;
+  block_desc_t *bd = 0;
   vlib_buffer_t bt;
 
   if (apif->mode == AF_PACKET_IF_MODE_IP)
@@ -227,170 +231,199 @@ af_packet_device_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
       vnet_feature_start_device_input_x1 (apif->sw_if_index, &next_index, &bt);
     }
 
-  n_free_bufs = vec_len (apm->rx_buffers[thread_index]);
-  if (PREDICT_FALSE (n_free_bufs < VLIB_FRAME_SIZE))
+  while ((((block_desc_t *) (block_start = apif->rx_ring[block]))
+	    ->hdr.bh1.block_status &
+	  TP_STATUS_USER) != 0)
     {
-      vec_validate (apm->rx_buffers[thread_index],
-		    VLIB_FRAME_SIZE + n_free_bufs - 1);
-      n_free_bufs +=
-	vlib_buffer_alloc (vm, &apm->rx_buffers[thread_index][n_free_bufs],
-			   VLIB_FRAME_SIZE);
-      _vec_len (apm->rx_buffers[thread_index]) = n_free_bufs;
-    }
+      u32 n_required = 0;
 
-  rx_frame = apif->next_rx_frame;
-  tph = (struct tpacket2_hdr *) (block_start + rx_frame * frame_size);
-  while ((tph->tp_status & TP_STATUS_USER) && (n_free_bufs > min_bufs))
-    {
-      vlib_buffer_t *b0 = 0, *first_b0 = 0;
-      u32 next0 = next_index;
+      if (PREDICT_FALSE (num_pkts == 0))
+	{
+	  bd = (block_desc_t *) block_start;
+	  num_pkts = bd->hdr.bh1.num_pkts;
+	  rx_frame_offset = sizeof (block_desc_t);
+	}
 
-      u32 n_left_to_next;
-      vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
-      while ((tph->tp_status & TP_STATUS_USER) && (n_free_bufs > min_bufs) &&
-	     n_left_to_next)
+      n_required = clib_max (num_pkts, VLIB_FRAME_SIZE);
+      n_free_bufs = vec_len (apm->rx_buffers[thread_index]);
+      if (PREDICT_FALSE (n_free_bufs < n_required))
 	{
-	  u32 data_len = tph->tp_snaplen;
-	  u32 offset = 0;
-	  u32 bi0 = 0, first_bi0 = 0, prev_bi0;
-	  u8 l4_hdr_sz = 0;
+	  vec_validate (apm->rx_buffers[thread_index],
+			n_required + n_free_bufs - 1);
+	  n_free_bufs += vlib_buffer_alloc (
+	    vm, &apm->rx_buffers[thread_index][n_free_bufs], n_required);
+	  _vec_len (apm->rx_buffers[thread_index]) = n_free_bufs;
+	}
 
-	  while (data_len)
+      while (num_pkts && (n_free_bufs > min_bufs))
+	{
+	  vlib_buffer_t *b0 = 0, *first_b0 = 0;
+	  u32 next0 = next_index;
+	  u32 n_left_to_next;
+	  vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
+	  while (num_pkts && n_left_to_next && (n_free_bufs > min_bufs))
 	    {
-	      /* grab free buffer */
-	      u32 last_empty_buffer =
-		vec_len (apm->rx_buffers[thread_index]) - 1;
-	      prev_bi0 = bi0;
-	      bi0 = apm->rx_buffers[thread_index][last_empty_buffer];
-	      b0 = vlib_get_buffer (vm, bi0);
-	      _vec_len (apm->rx_buffers[thread_index]) = last_empty_buffer;
-	      n_free_bufs--;
-
-	      /* copy data */
-	      u32 bytes_to_copy =
-		data_len > n_buffer_bytes ? n_buffer_bytes : data_len;
-	      u32 vlan_len = 0;
-	      u32 bytes_copied = 0;
-	      b0->current_data = 0;
-	      /* Kernel removes VLAN headers, so reconstruct VLAN */
-	      if (PREDICT_FALSE (tph->tp_status & TP_STATUS_VLAN_VALID))
+	      tph = (tpacket3_hdr_t *) (block_start + rx_frame_offset);
+
+	      if (num_pkts > 1)
+		CLIB_PREFETCH (block_start + rx_frame_offset +
+				 tph->tp_next_offset,
+			       2 * CLIB_CACHE_LINE_BYTES, LOAD);
+	      u32 data_len = tph->tp_snaplen;
+	      u32 offset = 0;
+	      u32 bi0 = 0, first_bi0 = 0, prev_bi0;
+	      u8 l4_hdr_sz = 0;
+
+	      while (data_len)
 		{
-		  if (PREDICT_TRUE (offset == 0))
+		  /* grab free buffer */
+		  u32 last_empty_buffer =
+		    vec_len (apm->rx_buffers[thread_index]) - 1;
+		  prev_bi0 = bi0;
+		  bi0 = apm->rx_buffers[thread_index][last_empty_buffer];
+		  b0 = vlib_get_buffer (vm, bi0);
+		  _vec_len (apm->rx_buffers[thread_index]) = last_empty_buffer;
+		  n_free_bufs--;
+
+		  /* copy data */
+		  u32 bytes_to_copy =
+		    data_len > n_buffer_bytes ? n_buffer_bytes : data_len;
+		  u32 vlan_len = 0;
+		  u32 bytes_copied = 0;
+		  b0->current_data = 0;
+		  /* Kernel removes VLAN headers, so reconstruct VLAN */
+		  if (PREDICT_FALSE (tph->tp_status & TP_STATUS_VLAN_VALID))
 		    {
-		      clib_memcpy_fast (vlib_buffer_get_current (b0),
-					(u8 *) tph + tph->tp_mac,
-					sizeof (ethernet_header_t));
-		      ethernet_header_t *eth = vlib_buffer_get_current (b0);
-		      ethernet_vlan_header_t *vlan =
-			(ethernet_vlan_header_t *) (eth + 1);
-		      vlan->priority_cfi_and_id =
-			clib_host_to_net_u16 (tph->tp_vlan_tci);
-		      vlan->type = eth->type;
-		      eth->type = clib_host_to_net_u16 (ETHERNET_TYPE_VLAN);
-		      vlan_len = sizeof (ethernet_vlan_header_t);
-		      bytes_copied = sizeof (ethernet_header_t);
+		      if (PREDICT_TRUE (offset == 0))
+			{
+			  clib_memcpy_fast (vlib_buffer_get_current (b0),
+					    (u8 *) tph + tph->tp_mac,
+					    sizeof (ethernet_header_t));
+			  ethernet_header_t *eth =
+			    vlib_buffer_get_current (b0);
+			  ethernet_vlan_header_t *vlan =
+			    (ethernet_vlan_header_t *) (eth + 1);
+			  vlan->priority_cfi_and_id =
+			    clib_host_to_net_u16 (tph->hv1.tp_vlan_tci);
+			  vlan->type = eth->type;
+			  eth->type =
+			    clib_host_to_net_u16 (ETHERNET_TYPE_VLAN);
+			  vlan_len = sizeof (ethernet_vlan_header_t);
+			  bytes_copied = sizeof (ethernet_header_t);
+			}
 		    }
-		}
-	      clib_memcpy_fast (((u8 *) vlib_buffer_get_current (b0)) +
-				bytes_copied + vlan_len,
-				(u8 *) tph + tph->tp_mac + offset +
-				bytes_copied, (bytes_to_copy - bytes_copied));
+		  clib_memcpy_fast (((u8 *) vlib_buffer_get_current (b0)) +
+				      bytes_copied + vlan_len,
+				    (u8 *) tph + tph->tp_mac + offset +
+				      bytes_copied,
+				    (bytes_to_copy - bytes_copied));
+
+		  /* fill buffer header */
+		  b0->current_length = bytes_to_copy + vlan_len;
 
-	      /* fill buffer header */
-	      b0->current_length = bytes_to_copy + vlan_len;
+		  if (offset == 0)
+		    {
+		      b0->total_length_not_including_first_buffer = 0;
+		      b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID;
+		      vnet_buffer (b0)->sw_if_index[VLIB_RX] =
+			apif->sw_if_index;
+		      vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~0;
+		      first_bi0 = bi0;
+		      first_b0 = vlib_get_buffer (vm, first_bi0);
+		      if (tph->tp_status & TP_STATUS_CSUMNOTREADY)
+			mark_tcp_udp_cksum_calc (first_b0, &l4_hdr_sz);
+		      /* This is a trade-off for GSO. As kernel isn't passing
+		       * us the GSO state or size, we guess it by comparing it
+		       * to the host MTU of the interface */
+		      if (tph->tp_snaplen > (apif->host_mtu + eth_header_size))
+			fill_gso_buffer_flags (first_b0, apif->host_mtu,
+					       l4_hdr_sz);
+		    }
+		  else
+		    buffer_add_to_chain (vm, bi0, first_bi0, prev_bi0);
 
-	      if (offset == 0)
+		  offset += bytes_to_copy;
+		  data_len -= bytes_to_copy;
+		}
+	      n_rx_packets++;
+	      n_rx_bytes += tph->tp_snaplen;
+	      to_next[0] = first_bi0;
+	      to_next += 1;
+	      n_left_to_next--;
+
+	      /* drop partial packets */
+	      if (PREDICT_FALSE (tph->tp_len != tph->tp_snaplen))
 		{
-		  b0->total_length_not_including_first_buffer = 0;
-		  b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID;
-		  vnet_buffer (b0)->sw_if_index[VLIB_RX] = apif->sw_if_index;
-		  vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0;
-		  first_bi0 = bi0;
-		  first_b0 = vlib_get_buffer (vm, first_bi0);
-		  if (tph->tp_status & TP_STATUS_CSUMNOTREADY)
-		    mark_tcp_udp_cksum_calc (first_b0, &l4_hdr_sz);
-		  /* This is a trade-off for GSO. As kernel isn't passing
-		   * us the GSO state or size, we guess it by comparing it
-		   * to the host MTU of the interface */
-		  if (tph->tp_snaplen > (apif->host_mtu + eth_header_size))
-		    fill_gso_buffer_flags (first_b0, apif->host_mtu,
-					   l4_hdr_sz);
+		  next0 = VNET_DEVICE_INPUT_NEXT_DROP;
+		  first_b0->error =
+		    node->errors[AF_PACKET_INPUT_ERROR_PARTIAL_PKT];
 		}
 	      else
-		buffer_add_to_chain (vm, bi0, first_bi0, prev_bi0);
-
-	      offset += bytes_to_copy;
-	      data_len -= bytes_to_copy;
-	    }
-	  n_rx_packets++;
-	  n_rx_bytes += tph->tp_snaplen;
-	  to_next[0] = first_bi0;
-	  to_next += 1;
-	  n_left_to_next--;
-
-	  /* drop partial packets */
-	  if (PREDICT_FALSE (tph->tp_len != tph->tp_snaplen))
-	    {
-	      next0 = VNET_DEVICE_INPUT_NEXT_DROP;
-	      first_b0->error =
-		node->errors[AF_PACKET_INPUT_ERROR_PARTIAL_PKT];
-	    }
-	  else
-	    {
-	      if (PREDICT_FALSE (apif->mode == AF_PACKET_IF_MODE_IP))
 		{
-		  switch (first_b0->data[0] & 0xf0)
+		  if (PREDICT_FALSE (apif->mode == AF_PACKET_IF_MODE_IP))
 		    {
-		    case 0x40:
-		      next0 = VNET_DEVICE_INPUT_NEXT_IP4_INPUT;
-		      break;
-		    case 0x60:
-		      next0 = VNET_DEVICE_INPUT_NEXT_IP6_INPUT;
-		      break;
-		    default:
-		      next0 = VNET_DEVICE_INPUT_NEXT_DROP;
-		      break;
+		      switch (first_b0->data[0] & 0xf0)
+			{
+			case 0x40:
+			  next0 = VNET_DEVICE_INPUT_NEXT_IP4_INPUT;
+			  break;
+			case 0x60:
+			  next0 = VNET_DEVICE_INPUT_NEXT_IP6_INPUT;
+			  break;
+			default:
+			  next0 = VNET_DEVICE_INPUT_NEXT_DROP;
+			  break;
+			}
+		      if (PREDICT_FALSE (apif->per_interface_next_index != ~0))
+			next0 = apif->per_interface_next_index;
+		    }
+		  else
+		    {
+		      /* copy feature arc data from template */
+		      first_b0->current_config_index = bt.current_config_index;
+		      vnet_buffer (first_b0)->feature_arc_index =
+			vnet_buffer (&bt)->feature_arc_index;
 		    }
-		  if (PREDICT_FALSE (apif->per_interface_next_index != ~0))
-		    next0 = apif->per_interface_next_index;
 		}
-	      else
+
+	      /* trace */
+	      if (PREDICT_FALSE (n_trace > 0 &&
+				 vlib_trace_buffer (vm, node, next0, first_b0,
+						    /* follow_chain */ 0)))
 		{
-		  /* copy feature arc data from template */
-		  first_b0->current_config_index = bt.current_config_index;
-		  vnet_buffer (first_b0)->feature_arc_index =
-		    vnet_buffer (&bt)->feature_arc_index;
+		  af_packet_input_trace_t *tr;
+		  vlib_set_trace_count (vm, node, --n_trace);
+		  tr = vlib_add_trace (vm, node, first_b0, sizeof (*tr));
+		  tr->next_index = next0;
+		  tr->hw_if_index = apif->hw_if_index;
+		  tr->block = block;
+		  tr->block_start = bd;
+		  tr->num_pkts = num_pkts;
+		  clib_memcpy_fast (&tr->bd, bd, sizeof (block_desc_t));
+		  clib_memcpy_fast (&tr->tph, tph, sizeof (tpacket3_hdr_t));
 		}
-	    }
 
-	  /* trace */
-	  if (PREDICT_FALSE
-	      (n_trace > 0 && vlib_trace_buffer (vm, node, next0, first_b0,
-						 /* follow_chain */ 0)))
-	    {
-	      af_packet_input_trace_t *tr;
-	      vlib_set_trace_count (vm, node, --n_trace);
-	      tr = vlib_add_trace (vm, node, first_b0, sizeof (*tr));
-	      tr->next_index = next0;
-	      tr->hw_if_index = apif->hw_if_index;
-	      clib_memcpy_fast (&tr->tph, tph, sizeof (struct tpacket2_hdr));
-	    }
+	      /* enque and take next packet */
+	      vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
+					       n_left_to_next, first_bi0,
+					       next0);
 
-	  /* enque and take next packet */
-	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
-					   n_left_to_next, first_bi0, next0);
+	      /* next packet */
+	      num_pkts--;
+	      rx_frame_offset += tph->tp_next_offset;
+	    }
 
-	  /* next packet */
-	  tph->tp_status = TP_STATUS_KERNEL;
-	  rx_frame = (rx_frame + 1) % frame_num;
-	  tph = (struct tpacket2_hdr *) (block_start + rx_frame * frame_size);
+	  vlib_put_next_frame (vm, node, next_index, n_left_to_next);
 	}
 
-      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+      if (PREDICT_TRUE (num_pkts == 0))
+	{
+	  bd->hdr.bh1.block_status = TP_STATUS_KERNEL;
+	  block = (block + 1) % block_nr;
+	}
     }
 
-  apif->next_rx_frame = rx_frame;
-
+  apif->next_rx_block = block;
   vlib_increment_combined_counter
     (vnet_get_main ()->interface_main.combined_sw_if_counters
      + VNET_INTERFACE_COUNTER_RX,