devices: af-packet gso mtu

Type: fix

Set the GSO flag when buffer length exceeds the
linux mtu. Don't listen for mtu changes on linux
side for now.

This also fixes a TX issue, as we only search for
valid frames on tx to the extent of n_left, we might
stay stuck.

Change-Id: Idf0bdd88990254a614962c2f7bc3e0292ccfd61a
Signed-off-by: Nathan Skrzypczak <nathan.skrzypczak@gmail.com>

4 files changed, 126 insertions, 38 deletions

diff --git a/src/vnet/devices/af_packet/af_packet.c b/src/vnet/devices/af_packet/af_packet.c
index ba6bf3d5a46..2cc0cc70bca 100644
--- a/src/vnet/devices/af_packet/af_packet.c
+++ b/src/vnet/devices/af_packet/af_packet.c
@@ -83,11 +83,32 @@ af_packet_eth_flag_change (vnet_main_t * vnm, vnet_hw_interface_t * hi,
 	  clib_error_free (error);
 	  return VNET_API_ERROR_SYSCALL_ERROR_1;
 	}
+      else
+	apif->host_mtu = hi->max_packet_bytes;
     }
 
   return 0;
 }
 
+static int
+af_packet_read_mtu (af_packet_if_t *apif)
+{
+  af_packet_main_t *apm = &af_packet_main;
+  clib_error_t *error;
+  u8 *s;
+  s = format (0, "/sys/class/net/%s/mtu%c", apif->host_if_name, 0);
+  error = clib_sysfs_read ((char *) s, "%d", &apif->host_mtu);
+  vec_free (s);
+  if (error)
+    {
+      vlib_log_err (apm->log_class, "sysfs read failed to get MTU: %U",
+		    format_clib_error, error);
+      clib_error_free (error);
+      return VNET_API_ERROR_SYSCALL_ERROR_1;
+    }
+  return 0;
+}
+
 static clib_error_t *
 af_packet_fd_read_ready (clib_file_t * uf)
 {
@@ -338,6 +359,10 @@ af_packet_create_if (vlib_main_t * vm, u8 * host_if_name, u8 * hw_addr_set,
   apif->next_tx_frame = 0;
   apif->next_rx_frame = 0;
 
+  ret = af_packet_read_mtu (apif);
+  if (ret != 0)
+    goto error;
+
   if (tm->n_vlib_mains > 1)
     clib_spinlock_init (&apif->lockp);
 
diff --git a/src/vnet/devices/af_packet/af_packet.h b/src/vnet/devices/af_packet/af_packet.h
index 395f2d93fdf..fc35b48b85e 100644
--- a/src/vnet/devices/af_packet/af_packet.h
+++ b/src/vnet/devices/af_packet/af_packet.h
@@ -48,6 +48,7 @@ typedef struct
   u32 per_interface_next_index;
   u8 is_admin_up;
   u32 queue_index;
+  u32 host_mtu;
 } af_packet_if_t;
 
 typedef struct
diff --git a/src/vnet/devices/af_packet/device.c b/src/vnet/devices/af_packet/device.c
index 2886fb5b6d5..b6b99a0465c 100644
--- a/src/vnet/devices/af_packet/device.c
+++ b/src/vnet/devices/af_packet/device.c
@@ -67,7 +67,48 @@ format_af_packet_device_name (u8 * s, va_list * args)
 static u8 *
 format_af_packet_device (u8 * s, va_list * args)
 {
-  s = format (s, "Linux PACKET socket interface");
+  u32 dev_instance = va_arg (*args, u32);
+  u32 indent = format_get_indent (s);
+  int __clib_unused verbose = va_arg (*args, int);
+
+  af_packet_main_t *apm = &af_packet_main;
+  af_packet_if_t *apif = pool_elt_at_index (apm->interfaces, dev_instance);
+  clib_spinlock_lock_if_init (&apif->lockp);
+  u32 block_size = apif->tx_req->tp_block_size;
+  u32 frame_size = apif->tx_req->tp_frame_size;
+  u32 frame_num = apif->tx_req->tp_frame_nr;
+  int block = 0;
+  u8 *block_start = apif->tx_ring + block * block_size;
+  u32 tx_frame = apif->next_tx_frame;
+  struct tpacket2_hdr *tph;
+
+  s = format (s, "Linux PACKET socket interface\n");
+  s = format (s, "%Ublock:%d frame:%d\n", format_white_space, indent,
+	      block_size, frame_size);
+  s = format (s, "%Unext frame:%d\n", format_white_space, indent,
+	      apif->next_tx_frame);
+
+  int n_send_req = 0, n_avail = 0, n_sending = 0, n_tot = 0, n_wrong = 0;
+  do
+    {
+      tph = (struct tpacket2_hdr *) (block_start + tx_frame * frame_size);
+      tx_frame = (tx_frame + 1) % frame_num;
+      if (tph->tp_status == 0)
+	n_avail++;
+      else if (tph->tp_status & TP_STATUS_SEND_REQUEST)
+	n_send_req++;
+      else if (tph->tp_status & TP_STATUS_SENDING)
+	n_sending++;
+      else
+	n_wrong++;
+      n_tot++;
+    }
+  while (tx_frame != apif->next_tx_frame);
+  s = format (s, "%Uavailable:%d request:%d sending:%d wrong:%d total:%d\n",
+	      format_white_space, indent, n_avail, n_send_req, n_sending,
+	      n_wrong, n_tot);
+
+  clib_spinlock_unlock_if_init (&apif->lockp);
   return s;
 }
 
@@ -99,7 +140,7 @@ VNET_DEVICE_CLASS_TX_FN (af_packet_device_class) (vlib_main_t * vm,
   struct tpacket2_hdr *tph;
   u32 frame_not_ready = 0;
 
-  while (n_left > 0)
+  while (n_left)
     {
       u32 len;
       u32 offset = 0;
@@ -108,13 +149,17 @@ VNET_DEVICE_CLASS_TX_FN (af_packet_device_class) (vlib_main_t * vm,
       u32 bi = buffers[0];
       buffers++;
 
+    nextframe:
       tph = (struct tpacket2_hdr *) (block_start + tx_frame * frame_size);
-
-      if (PREDICT_FALSE
-	  (tph->tp_status & (TP_STATUS_SEND_REQUEST | TP_STATUS_SENDING)))
+      if (PREDICT_FALSE (tph->tp_status &
+			 (TP_STATUS_SEND_REQUEST | TP_STATUS_SENDING)))
 	{
+	  tx_frame = (tx_frame + 1) % frame_num;
 	  frame_not_ready++;
-	  goto next;
+	  /* check if we've exhausted the ring */
+	  if (PREDICT_FALSE (frame_not_ready + n_sent == frame_num))
+	    break;
+	  goto nextframe;
 	}
 
       do
@@ -132,7 +177,7 @@ VNET_DEVICE_CLASS_TX_FN (af_packet_device_class) (vlib_main_t * vm,
       tph->tp_len = tph->tp_snaplen = offset;
       tph->tp_status = TP_STATUS_SEND_REQUEST;
       n_sent++;
-    next:
+
       tx_frame = (tx_frame + 1) % frame_num;
 
       /* check if we've exhausted the ring */
@@ -142,23 +187,22 @@ VNET_DEVICE_CLASS_TX_FN (af_packet_device_class) (vlib_main_t * vm,
 
   CLIB_MEMORY_BARRIER ();
 
-  if (PREDICT_TRUE (n_sent))
-    {
-      apif->next_tx_frame = tx_frame;
+  apif->next_tx_frame = tx_frame;
 
-      if (PREDICT_FALSE (sendto (apif->fd, NULL, 0,
-				 MSG_DONTWAIT, NULL, 0) == -1))
-	{
-	  /* Uh-oh, drop & move on, but count whether it was fatal or not.
-	   * Note that we have no reliable way to properly determine the
-	   * disposition of the packets we just enqueued for delivery.
-	   */
-	  vlib_error_count (vm, node->node_index,
-			    unix_error_is_fatal (errno) ?
+  if (PREDICT_TRUE (n_sent))
+    if (PREDICT_FALSE (sendto (apif->fd, NULL, 0, MSG_DONTWAIT, NULL, 0) ==
+		       -1))
+      {
+	/* Uh-oh, drop & move on, but count whether it was fatal or not.
+	 * Note that we have no reliable way to properly determine the
+	 * disposition of the packets we just enqueued for delivery.
+	 */
+	vlib_error_count (vm, node->node_index,
+			  unix_error_is_fatal (errno) ?
 			    AF_PACKET_TX_ERROR_TXRING_FATAL :
-			    AF_PACKET_TX_ERROR_TXRING_EAGAIN, n_sent);
-	}
-    }
+			    AF_PACKET_TX_ERROR_TXRING_EAGAIN,
+			  n_sent);
+      }
 
   clib_spinlock_unlock_if_init (&apif->lockp);
 
diff --git a/src/vnet/devices/af_packet/node.c b/src/vnet/devices/af_packet/node.c
index d444b3b6eea..f4db6399a57 100644
--- a/src/vnet/devices/af_packet/node.c
+++ b/src/vnet/devices/af_packet/node.c
@@ -109,7 +109,15 @@ buffer_add_to_chain (vlib_main_t * vm, u32 bi, u32 first_bi, u32 prev_bi)
 }
 
 static_always_inline void
-mark_tcp_udp_cksum_calc (vlib_buffer_t * b)
+fill_gso_buffer_flags (vlib_buffer_t *b, u32 gso_size, u8 l4_hdr_sz)
+{
+  b->flags |= VNET_BUFFER_F_GSO;
+  vnet_buffer2 (b)->gso_size = gso_size;
+  vnet_buffer2 (b)->gso_l4_hdr_sz = l4_hdr_sz;
+}
+
+static_always_inline void
+mark_tcp_udp_cksum_calc (vlib_buffer_t *b, u8 *l4_hdr_sz)
 {
   ethernet_header_t *eth = vlib_buffer_get_current (b);
   if (clib_net_to_host_u16 (eth->type) == ETHERNET_TYPE_IP4)
@@ -120,18 +128,20 @@ mark_tcp_udp_cksum_calc (vlib_buffer_t * b)
       if (ip4->protocol == IP_PROTOCOL_TCP)
 	{
 	  b->flags |= VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
-	  ((tcp_header_t
-	    *) (vlib_buffer_get_current (b) +
-		sizeof (ethernet_header_t) +
-		ip4_header_bytes (ip4)))->checksum = 0;
+	  tcp_header_t *tcp = (tcp_header_t *) (vlib_buffer_get_current (b) +
+						sizeof (ethernet_header_t) +
+						ip4_header_bytes (ip4));
+	  tcp->checksum = 0;
+	  *l4_hdr_sz = tcp_header_bytes (tcp);
 	}
       else if (ip4->protocol == IP_PROTOCOL_UDP)
 	{
 	  b->flags |= VNET_BUFFER_F_OFFLOAD_UDP_CKSUM;
-	  ((udp_header_t
-	    *) (vlib_buffer_get_current (b) +
-		sizeof (ethernet_header_t) +
-		ip4_header_bytes (ip4)))->checksum = 0;
+	  udp_header_t *udp = (udp_header_t *) (vlib_buffer_get_current (b) +
+						sizeof (ethernet_header_t) +
+						ip4_header_bytes (ip4));
+	  udp->checksum = 0;
+	  *l4_hdr_sz = sizeof (*udp);
 	}
       vnet_buffer (b)->l3_hdr_offset = sizeof (ethernet_header_t);
       vnet_buffer (b)->l4_hdr_offset =
@@ -156,16 +166,20 @@ mark_tcp_udp_cksum_calc (vlib_buffer_t * b)
       if (ip6->protocol == IP_PROTOCOL_TCP)
 	{
 	  b->flags |= VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
-	  ((tcp_header_t
-	    *) (vlib_buffer_get_current (b) +
-		sizeof (ethernet_header_t) + ip6_hdr_len))->checksum = 0;
+	  tcp_header_t *tcp =
+	    (tcp_header_t *) (vlib_buffer_get_current (b) +
+			      sizeof (ethernet_header_t) + ip6_hdr_len);
+	  tcp->checksum = 0;
+	  *l4_hdr_sz = tcp_header_bytes (tcp);
 	}
       else if (ip6->protocol == IP_PROTOCOL_UDP)
 	{
 	  b->flags |= VNET_BUFFER_F_OFFLOAD_UDP_CKSUM;
-	  ((udp_header_t
-	    *) (vlib_buffer_get_current (b) +
-		sizeof (ethernet_header_t) + ip6_hdr_len))->checksum = 0;
+	  udp_header_t *udp =
+	    (udp_header_t *) (vlib_buffer_get_current (b) +
+			      sizeof (ethernet_header_t) + ip6_hdr_len);
+	  udp->checksum = 0;
+	  *l4_hdr_sz = sizeof (*udp);
 	}
       vnet_buffer (b)->l3_hdr_offset = sizeof (ethernet_header_t);
       vnet_buffer (b)->l4_hdr_offset =
@@ -221,6 +235,7 @@ af_packet_device_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  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)
 	    {
@@ -275,7 +290,10 @@ af_packet_device_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 		  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);
+		    mark_tcp_udp_cksum_calc (first_b0, &l4_hdr_sz);
+		  if (tph->tp_snaplen > apif->host_mtu)
+		    fill_gso_buffer_flags (first_b0, apif->host_mtu,
+					   l4_hdr_sz);
 		}
 	      else
 		buffer_add_to_chain (vm, bi0, first_bi0, prev_bi0);