wireguard: add handoff node

All timer and control plane functions happen from main thread

Type: fix

Change-Id: I4fc333c644485cd17e6f426493feef91688d9b24
Signed-off-by: Artem Glazychev <artem.glazychev@xored.com>

1 files changed, 140 insertions, 115 deletions

diff --git a/src/plugins/wireguard/wireguard_input.c b/src/plugins/wireguard/wireguard_input.c
index cdd65f87b51..b15c265cdac 100755
--- a/src/plugins/wireguard/wireguard_input.c
+++ b/src/plugins/wireguard/wireguard_input.c
@@ -30,6 +30,7 @@
   _(DECRYPTION, "Failed during decryption")             \
   _(KEEPALIVE_SEND, "Failed while sending Keepalive")   \
   _(HANDSHAKE_SEND, "Failed while sending Handshake")   \
+  _(TOO_BIG, "Packet too big")                          \
   _(UNDEFINED, "Undefined error")
 
 typedef enum
@@ -51,7 +52,7 @@ typedef struct
   message_type_t type;
   u16 current_length;
   bool is_keepalive;
-
+  index_t peer;
 } wg_input_trace_t;
 
 u8 *
@@ -79,6 +80,7 @@ format_wg_input_trace (u8 * s, va_list * args)
 
   s = format (s, "WG input: \n");
   s = format (s, "  Type: %U\n", format_wg_message_type, t->type);
+  s = format (s, "  peer: %d\n", t->peer);
   s = format (s, "  Length: %d\n", t->current_length);
   s = format (s, "  Keepalive: %s", t->is_keepalive ? "true" : "false");
 
@@ -87,6 +89,8 @@ format_wg_input_trace (u8 * s, va_list * args)
 
 typedef enum
 {
+  WG_INPUT_NEXT_HANDOFF_HANDSHAKE,
+  WG_INPUT_NEXT_HANDOFF_DATA,
   WG_INPUT_NEXT_IP4_INPUT,
   WG_INPUT_NEXT_PUNT,
   WG_INPUT_NEXT_ERROR,
@@ -106,6 +110,8 @@ typedef enum
 static wg_input_error_t
 wg_handshake_process (vlib_main_t * vm, wg_main_t * wmp, vlib_buffer_t * b)
 {
+  ASSERT (vm->thread_index == 0);
+
   enum cookie_mac_state mac_state;
   bool packet_needs_cookie;
   bool under_load;
@@ -129,17 +135,15 @@ wg_handshake_process (vlib_main_t * vm, wg_main_t * wmp, vlib_buffer_t * b)
   if (NULL == wg_if)
     return WG_INPUT_ERROR_INTERFACE;
 
-  if (header->type == MESSAGE_HANDSHAKE_COOKIE)
+  if (PREDICT_FALSE (header->type == MESSAGE_HANDSHAKE_COOKIE))
     {
       message_handshake_cookie_t *packet =
 	(message_handshake_cookie_t *) current_b_data;
       u32 *entry =
 	wg_index_table_lookup (&wmp->index_table, packet->receiver_index);
       if (entry)
-	{
-	  peer = pool_elt_at_index (wmp->peers, *entry);
-	}
-      if (!peer)
+	peer = wg_peer_get (*entry);
+      else
 	return WG_INPUT_ERROR_PEER;
 
       // TODO: Implement cookie_maker_consume_payload
@@ -178,17 +182,17 @@ wg_handshake_process (vlib_main_t * vm, wg_main_t * wmp, vlib_buffer_t * b)
 	    // TODO: Add processing
 	  }
 	noise_remote_t *rp;
-
 	if (noise_consume_initiation
-	    (wmp->vlib_main, &wg_if->local, &rp, message->sender_index,
-	     message->unencrypted_ephemeral, message->encrypted_static,
-	     message->encrypted_timestamp))
+	    (vm, noise_local_get (wg_if->local_idx), &rp,
+	     message->sender_index, message->unencrypted_ephemeral,
+	     message->encrypted_static, message->encrypted_timestamp))
 	  {
-	    peer = pool_elt_at_index (wmp->peers, rp->r_peer_idx);
+	    peer = wg_peer_get (rp->r_peer_idx);
+	  }
+	else
+	  {
+	    return WG_INPUT_ERROR_PEER;
 	  }
-
-	if (!peer)
-	  return WG_INPUT_ERROR_PEER;
 
 	// set_peer_address (peer, ip4_src, udp_src_port);
 	if (PREDICT_FALSE (!wg_send_handshake_response (vm, peer)))
@@ -203,15 +207,18 @@ wg_handshake_process (vlib_main_t * vm, wg_main_t * wmp, vlib_buffer_t * b)
 	message_handshake_response_t *resp = current_b_data;
 	u32 *entry =
 	  wg_index_table_lookup (&wmp->index_table, resp->receiver_index);
-	if (entry)
+
+	if (PREDICT_TRUE (entry != NULL))
 	  {
-	    peer = pool_elt_at_index (wmp->peers, *entry);
-	    if (!peer || peer->is_dead)
+	    peer = wg_peer_get (*entry);
+	    if (peer->is_dead)
 	      return WG_INPUT_ERROR_PEER;
 	  }
+	else
+	  return WG_INPUT_ERROR_PEER;
 
 	if (!noise_consume_response
-	    (wmp->vlib_main, &peer->remote, resp->sender_index,
+	    (vm, &peer->remote, resp->sender_index,
 	     resp->receiver_index, resp->unencrypted_ephemeral,
 	     resp->encrypted_nothing))
 	  {
@@ -223,8 +230,9 @@ wg_handshake_process (vlib_main_t * vm, wg_main_t * wmp, vlib_buffer_t * b)
 	  }
 
 	// set_peer_address (peer, ip4_src, udp_src_port);
-	if (noise_remote_begin_session (wmp->vlib_main, &peer->remote))
+	if (noise_remote_begin_session (vm, &peer->remote))
 	  {
+
 	    wg_timers_session_derived (peer);
 	    wg_timers_handshake_complete (peer);
 	    if (PREDICT_FALSE (!wg_send_keepalive (vm, peer)))
@@ -272,6 +280,7 @@ VLIB_NODE_FN (wg_input_node) (vlib_main_t * vm,
   u32 *from;
   vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b;
   u16 nexts[VLIB_FRAME_SIZE], *next;
+  u32 thread_index = vm->thread_index;
 
   from = vlib_frame_vector_args (frame);
   n_left_from = frame->n_vectors;
@@ -289,120 +298,132 @@ VLIB_NODE_FN (wg_input_node) (vlib_main_t * vm,
       next[0] = WG_INPUT_NEXT_PUNT;
       header_type =
 	((message_header_t *) vlib_buffer_get_current (b[0]))->type;
+      u32 *peer_idx;
 
-      switch (header_type)
+      if (PREDICT_TRUE (header_type == MESSAGE_DATA))
 	{
-	case MESSAGE_HANDSHAKE_INITIATION:
-	case MESSAGE_HANDSHAKE_RESPONSE:
-	case MESSAGE_HANDSHAKE_COOKIE:
-	  {
-	    wg_input_error_t ret = wg_handshake_process (vm, wmp, b[0]);
-	    if (ret != WG_INPUT_ERROR_NONE)
-	      {
-		next[0] = WG_INPUT_NEXT_ERROR;
-		b[0]->error = node->errors[ret];
-	      }
-	    break;
-	  }
-	case MESSAGE_DATA:
-	  {
-	    message_data_t *data = vlib_buffer_get_current (b[0]);
-	    u32 *entry =
-	      wg_index_table_lookup (&wmp->index_table, data->receiver_index);
-
-	    if (entry)
-	      {
-		peer = pool_elt_at_index (wmp->peers, *entry);
-	      }
-	    else
-	      {
-		next[0] = WG_INPUT_NEXT_ERROR;
-		b[0]->error = node->errors[WG_INPUT_ERROR_PEER];
-		goto out;
-	      }
+	  message_data_t *data = vlib_buffer_get_current (b[0]);
 
-	    u16 encr_len = b[0]->current_length - sizeof (message_data_t);
-	    u16 decr_len = encr_len - NOISE_AUTHTAG_LEN;
-	    u8 *decr_data = clib_mem_alloc (decr_len);
+	  peer_idx = wg_index_table_lookup (&wmp->index_table,
+					    data->receiver_index);
 
-	    enum noise_state_crypt state_cr =
-	      noise_remote_decrypt (wmp->vlib_main,
-				    &peer->remote,
-				    data->receiver_index,
-				    data->counter,
-				    data->encrypted_data,
-				    encr_len,
-				    decr_data);
+	  if (peer_idx)
+	    {
+	      peer = wg_peer_get (*peer_idx);
+	    }
+	  else
+	    {
+	      next[0] = WG_INPUT_NEXT_ERROR;
+	      b[0]->error = node->errors[WG_INPUT_ERROR_PEER];
+	      goto out;
+	    }
 
-	    switch (state_cr)
-	      {
-	      case SC_OK:
-		break;
-	      case SC_CONN_RESET:
-		wg_timers_handshake_complete (peer);
-		break;
-	      case SC_KEEP_KEY_FRESH:
-		if (PREDICT_FALSE (!wg_send_handshake (vm, peer, false)))
-		  {
-		    vlib_node_increment_counter (vm, wg_input_node.index,
-						 WG_INPUT_ERROR_HANDSHAKE_SEND,
-						 1);
-		  }
-		break;
-	      case SC_FAILED:
-		next[0] = WG_INPUT_NEXT_ERROR;
-		b[0]->error = node->errors[WG_INPUT_ERROR_DECRYPTION];
-		goto out;
-	      default:
-		break;
-	      }
+	  if (PREDICT_FALSE (~0 == peer->input_thread_index))
+	    {
+	      /* this is the first packet to use this peer, claim the peer
+	       * for this thread.
+	       */
+	      clib_atomic_cmp_and_swap (&peer->input_thread_index, ~0,
+					wg_peer_assign_thread (thread_index));
+	    }
 
-	    clib_memcpy (vlib_buffer_get_current (b[0]), decr_data, decr_len);
-	    b[0]->current_length = decr_len;
-	    b[0]->flags &= ~VNET_BUFFER_F_OFFLOAD_UDP_CKSUM;
+	  if (PREDICT_TRUE (thread_index != peer->input_thread_index))
+	    {
+	      next[0] = WG_INPUT_NEXT_HANDOFF_DATA;
+	      goto next;
+	    }
 
-	    clib_mem_free (decr_data);
+	  u16 encr_len = b[0]->current_length - sizeof (message_data_t);
+	  u16 decr_len = encr_len - NOISE_AUTHTAG_LEN;
+	  if (PREDICT_FALSE (decr_len >= WG_DEFAULT_DATA_SIZE))
+	    {
+	      b[0]->error = node->errors[WG_INPUT_ERROR_TOO_BIG];
+	      goto out;
+	    }
 
-	    wg_timers_any_authenticated_packet_received (peer);
-	    wg_timers_any_authenticated_packet_traversal (peer);
+	  u8 *decr_data = wmp->per_thread_data[thread_index].data;
 
-	    if (decr_len == 0)
-	      {
-		is_keepalive = true;
-		goto out;
-	      }
+	  enum noise_state_crypt state_cr = noise_remote_decrypt (vm,
+								  &peer->remote,
+								  data->receiver_index,
+								  data->counter,
+								  data->encrypted_data,
+								  encr_len,
+								  decr_data);
 
-	    wg_timers_data_received (peer);
+	  if (PREDICT_FALSE (state_cr == SC_CONN_RESET))
+	    {
+	      wg_timers_handshake_complete (peer);
+	    }
+	  else if (PREDICT_FALSE (state_cr == SC_KEEP_KEY_FRESH))
+	    {
+	      wg_send_handshake_from_mt (*peer_idx, false);
+	    }
+	  else if (PREDICT_FALSE (state_cr == SC_FAILED))
+	    {
+	      next[0] = WG_INPUT_NEXT_ERROR;
+	      b[0]->error = node->errors[WG_INPUT_ERROR_DECRYPTION];
+	      goto out;
+	    }
 
-	    ip4_header_t *iph = vlib_buffer_get_current (b[0]);
+	  clib_memcpy (vlib_buffer_get_current (b[0]), decr_data, decr_len);
+	  b[0]->current_length = decr_len;
+	  b[0]->flags &= ~VNET_BUFFER_F_OFFLOAD_UDP_CKSUM;
 
-	    const wg_peer_allowed_ip_t *allowed_ip;
-	    bool allowed = false;
+	  wg_timers_any_authenticated_packet_received (peer);
+	  wg_timers_any_authenticated_packet_traversal (peer);
 
-	    /*
-	     * we could make this into an ACL, but the expectation
-	     * is that there aren't many allowed IPs and thus a linear
-	     * walk is fater than an ACL
-	     */
-	    vec_foreach (allowed_ip, peer->allowed_ips)
+	  /* Keepalive packet has zero length */
+	  if (decr_len == 0)
 	    {
-	      if (fib_prefix_is_cover_addr_4 (&allowed_ip->prefix,
-					      &iph->src_address))
-		{
-		  allowed = true;
-		  break;
-		}
+	      is_keepalive = true;
+	      goto out;
 	    }
-	    if (allowed)
+
+	  wg_timers_data_received (peer);
+
+	  ip4_header_t *iph = vlib_buffer_get_current (b[0]);
+
+	  const wg_peer_allowed_ip_t *allowed_ip;
+	  bool allowed = false;
+
+	  /*
+	   * we could make this into an ACL, but the expectation
+	   * is that there aren't many allowed IPs and thus a linear
+	   * walk is fater than an ACL
+	   */
+	  vec_foreach (allowed_ip, peer->allowed_ips)
+	  {
+	    if (fib_prefix_is_cover_addr_4 (&allowed_ip->prefix,
+					    &iph->src_address))
 	      {
-		vnet_buffer (b[0])->sw_if_index[VLIB_RX] =
-		  peer->wg_sw_if_index;
-		next[0] = WG_INPUT_NEXT_IP4_INPUT;
+		allowed = true;
+		break;
 	      }
-	    break;
 	  }
-	default:
-	  break;
+	  if (allowed)
+	    {
+	      vnet_buffer (b[0])->sw_if_index[VLIB_RX] = peer->wg_sw_if_index;
+	      next[0] = WG_INPUT_NEXT_IP4_INPUT;
+	    }
+	}
+      else
+	{
+	  peer_idx = NULL;
+
+	  /* Handshake packets should be processed in main thread */
+	  if (thread_index != 0)
+	    {
+	      next[0] = WG_INPUT_NEXT_HANDOFF_HANDSHAKE;
+	      goto next;
+	    }
+
+	  wg_input_error_t ret = wg_handshake_process (vm, wmp, b[0]);
+	  if (ret != WG_INPUT_ERROR_NONE)
+	    {
+	      next[0] = WG_INPUT_NEXT_ERROR;
+	      b[0]->error = node->errors[ret];
+	    }
 	}
 
     out:
@@ -413,7 +434,9 @@ VLIB_NODE_FN (wg_input_node) (vlib_main_t * vm,
 	  t->type = header_type;
 	  t->current_length = b[0]->current_length;
 	  t->is_keepalive = is_keepalive;
+	  t->peer = peer_idx ? *peer_idx : INDEX_INVALID;
 	}
+    next:
       n_left_from -= 1;
       next += 1;
       b += 1;
@@ -435,6 +458,8 @@ VLIB_REGISTER_NODE (wg_input_node) =
   .n_next_nodes = WG_INPUT_N_NEXT,
   /* edit / add dispositions here */
   .next_nodes = {
+        [WG_INPUT_NEXT_HANDOFF_HANDSHAKE] = "wg-handshake-handoff",
+        [WG_INPUT_NEXT_HANDOFF_DATA] = "wg-input-data-handoff",
         [WG_INPUT_NEXT_IP4_INPUT] = "ip4-input-no-checksum",
         [WG_INPUT_NEXT_PUNT] = "error-punt",
         [WG_INPUT_NEXT_ERROR] = "error-drop",