Include TCP options in segment size computation

Ensure that TCP data plus options does not exceed peer's advertised MSS.

Change-Id: I0de824cb3619346f0394dd694942fc1cf33a82b7
Signed-off-by: Florin Coras <fcoras@cisco.com>

5 files changed, 97 insertions, 42 deletions

diff --git a/src/vnet/session/node.c b/src/vnet/session/node.c
index 210754fa723..2d12ee2bac9 100644
--- a/src/vnet/session/node.c
+++ b/src/vnet/session/node.c
@@ -95,8 +95,8 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
   tc0 = transport_vft->get_connection (s0->connection_index, thread_index);
 
   /* Make sure we have space to send and there's something to dequeue */
-  snd_space0 = transport_vft->send_space (tc0);
   snd_mss0 = transport_vft->send_mss (tc0);
+  snd_space0 = transport_vft->send_space (tc0);
 
   /* Can't make any progress */
   if (snd_space0 == 0 || snd_mss0 == 0)
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index e80e2ec9b7c..224ee0dd160 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -331,7 +331,7 @@ void
 tcp_connection_init_vars (tcp_connection_t * tc)
 {
   tcp_connection_timers_init (tc);
-  tcp_set_snd_mss (tc);
+  tcp_init_mss (tc);
   scoreboard_init (&tc->sack_sb);
   tcp_cc_init (tc);
 }
@@ -560,10 +560,23 @@ tcp_half_open_session_get_transport (u32 conn_index)
   return &tc->connection;
 }
 
+/**
+ * Compute maximum segment size for session layer.
+ *
+ * Since the result needs to be the actual data length, it first computes
+ * the tcp options to be used in the next burst and subtracts their
+ * length from the connection's snd_mss.
+ */
 u16
 tcp_session_send_mss (transport_connection_t * trans_conn)
 {
   tcp_connection_t *tc = (tcp_connection_t *) trans_conn;
+
+  /* Ensure snd_mss does accurately reflect the amount of data we can push
+   * in a segment. This also makes sure that options are updated according to
+   * the current state of the connection. */
+  tcp_update_snd_mss (tc);
+
   return tc->snd_mss;
 }
 
@@ -607,7 +620,7 @@ tcp_session_send_space (transport_connection_t * trans_conn)
       tc->snd_nxt = tc->snd_una_max;
       snd_space = tcp_available_wnd (tc) - tc->rtx_bytes
 	- (tc->snd_una_max - tc->snd_congestion);
-      if (snd_space <= 0)
+      if (snd_space <= 0 || (tc->snd_una_max - tc->snd_una) >= tc->snd_wnd)
 	return 0;
       return tcp_round_snd_space (tc, snd_space);
     }
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index c75479dce5c..8212ada785e 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -211,7 +211,9 @@ typedef struct _tcp_connection
   u32 irs;		/**< initial remote sequence */
 
   /* Options */
-  tcp_options_t opt;	/**< TCP connection options parsed */
+  tcp_options_t opt;		/**< TCP connection options parsed */
+  tcp_options_t snd_opts;	/**< Tx options for connection */
+  u8 snd_opts_len;		/**< Tx options len */
   u8 rcv_wscale;	/**< Window scale to advertise to peer */
   u8 snd_wscale;	/**< Window scale to use when sending */
   u32 tsval_recent;	/**< Last timestamp received */
@@ -241,7 +243,8 @@ typedef struct _tcp_connection
   u32 rtt_ts;		/**< Timestamp for tracked ACK */
   u32 rtt_seq;		/**< Sequence number for tracked ACK */
 
-  u16 snd_mss;		/**< Send MSS */
+  u16 snd_mss;		/**< Effective send max seg (data) size */
+  u16 mss;		/**< Our max seg size that includes options */
 } tcp_connection_t;
 
 struct _tcp_cc_algorithm
@@ -405,7 +408,8 @@ void tcp_make_synack (tcp_connection_t * ts, vlib_buffer_t * b);
 void tcp_send_reset (vlib_buffer_t * pkt, u8 is_ip4);
 void tcp_send_syn (tcp_connection_t * tc);
 void tcp_send_fin (tcp_connection_t * tc);
-void tcp_set_snd_mss (tcp_connection_t * tc);
+void tcp_init_mss (tcp_connection_t * tc);
+void tcp_update_snd_mss (tcp_connection_t * tc);
 
 always_inline u32
 tcp_end_seq (tcp_header_t * th, u32 len)
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index e9c52c5e27f..d268251cb38 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -389,10 +389,14 @@ static int
 tcp_update_rtt (tcp_connection_t * tc, u32 ack)
 {
   u32 mrtt = 0;
+  u8 rtx_acked;
+
+  /* Determine if only rtx bytes are acked. TODO fast retransmit */
+  rtx_acked = tc->rto_boff && (tc->bytes_acked <= tc->snd_mss);
 
   /* Karn's rule, part 1. Don't use retransmitted segments to estimate
    * RTT because they're ambiguous. */
-  if (tc->rtt_ts && seq_geq (ack, tc->rtt_seq) && !tc->rto_boff)
+  if (tc->rtt_ts && seq_geq (ack, tc->rtt_seq) && !rtx_acked)
     {
       mrtt = tcp_time_now () - tc->rtt_ts;
     }
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 7ee930c61af..2a1b140750c 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -64,26 +64,6 @@ format_tcp_tx_trace (u8 * s, va_list * args)
   return s;
 }
 
-void
-tcp_set_snd_mss (tcp_connection_t * tc)
-{
-  u16 snd_mss;
-
-  /* TODO find our iface MTU */
-  snd_mss = dummy_mtu;
-
-  /* TODO cache mss and consider PMTU discovery */
-  snd_mss = clib_min (tc->opt.mss, snd_mss);
-
-  tc->snd_mss = snd_mss;
-
-  if (tc->snd_mss == 0)
-    {
-      clib_warning ("snd mss is 0");
-      tc->snd_mss = dummy_mtu;
-    }
-}
-
 static u8
 tcp_window_compute_scale (u32 available_space)
 {
@@ -100,7 +80,7 @@ tcp_window_compute_scale (u32 available_space)
 always_inline u32
 tcp_initial_wnd_unscaled (tcp_connection_t * tc)
 {
-  return TCP_IW_N_SEGMENTS * dummy_mtu;
+  return TCP_IW_N_SEGMENTS * tc->mss;
 }
 
 /**
@@ -310,7 +290,7 @@ tcp_make_synack_options (tcp_connection_t * tc, tcp_options_t * opts)
   u8 len = 0;
 
   opts->flags |= TCP_OPTS_FLAG_MSS;
-  opts->mss = dummy_mtu;	/*XXX discover that */
+  opts->mss = tc->mss;
   len += TCP_OPTION_LEN_MSS;
 
   if (tcp_opts_wscale (&tc->opt))
@@ -389,6 +369,57 @@ tcp_make_options (tcp_connection_t * tc, tcp_options_t * opts,
     }
 }
 
+/**
+ * Update max segment size we're able to process.
+ *
+ * The value is constrained by our interface's MTU and IP options. It is
+ * also what we advertise to our peer.
+ */
+void
+tcp_update_rcv_mss (tcp_connection_t * tc)
+{
+  /* TODO find our iface MTU */
+  tc->mss = dummy_mtu;
+}
+
+/**
+ * Update snd_mss to reflect the effective segment size that we can send
+ * by taking into account all TCP options, including SACKs
+ */
+void
+tcp_update_snd_mss (tcp_connection_t * tc)
+{
+  /* Compute options to be used for connection. These may be reused when
+   * sending data or to compute the effective mss (snd_mss) */
+  tc->snd_opts_len =
+    tcp_make_options (tc, &tc->snd_opts, TCP_STATE_ESTABLISHED);
+
+  /* XXX check if MTU has been updated */
+  tc->snd_mss = clib_min (tc->mss, tc->opt.mss) - tc->snd_opts_len;
+}
+
+void
+tcp_init_mss (tcp_connection_t * tc)
+{
+  tcp_update_rcv_mss (tc);
+
+  /* TODO cache mss and consider PMTU discovery */
+  tc->snd_mss = clib_min (tc->opt.mss, tc->mss);
+
+  if (tc->snd_mss == 0)
+    {
+      clib_warning ("snd mss is 0");
+      tc->snd_mss = tc->mss;
+    }
+
+  /* We should have enough space for 40 bytes of options */
+  ASSERT (tc->snd_mss > 45);
+
+  /* If we use timestamp option, account for it */
+  if (tcp_opts_tstamp (&tc->opt))
+    tc->snd_mss -= TCP_OPTION_LEN_TIMESTAMP;
+}
+
 #define tcp_get_free_buffer_index(tm, bidx)                             \
 do {                                                                    \
   u32 *my_tx_buffers, n_free_buffers;                                   \
@@ -886,20 +917,20 @@ tcp_make_state_flags (tcp_state_t next_state)
  */
 static void
 tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
-		tcp_state_t next_state)
+		tcp_state_t next_state, u8 compute_opts)
 {
   u32 advertise_wnd, data_len;
-  u8 tcp_opts_len, tcp_hdr_opts_len, opts_write_len, flags;
-  tcp_options_t _snd_opts, *snd_opts = &_snd_opts;
+  u8 tcp_hdr_opts_len, opts_write_len, flags;
   tcp_header_t *th;
 
   data_len = b->current_length;
   vnet_buffer (b)->tcp.flags = 0;
 
-  /* Make and write options */
-  memset (snd_opts, 0, sizeof (*snd_opts));
-  tcp_opts_len = tcp_make_options (tc, snd_opts, next_state);
-  tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
+  if (compute_opts)
+    tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
+
+  /* Write pre-computed options */
+  tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
 
   /* Get rcv window to advertise */
   advertise_wnd = tcp_window_to_advertise (tc, next_state);
@@ -910,9 +941,9 @@ tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
 			     tc->rcv_nxt, tcp_hdr_opts_len, flags,
 			     advertise_wnd);
 
-  opts_write_len = tcp_options_write ((u8 *) (th + 1), snd_opts);
+  opts_write_len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
 
-  ASSERT (opts_write_len == tcp_opts_len);
+  ASSERT (opts_write_len == tc->snd_opts_len);
 
   /* Tag the buffer with the connection index  */
   vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
@@ -993,6 +1024,8 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, vlib_buffer_t * b,
 	goto done;
     }
 
+  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
+
   ASSERT (max_bytes <= tc->snd_mss);
 
   n_bytes = stream_session_peek_bytes (&tc->connection,
@@ -1000,7 +1033,7 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, vlib_buffer_t * b,
 				       max_bytes);
   ASSERT (n_bytes != 0);
   b->current_length = n_bytes;
-  tcp_push_hdr_i (tc, b, tc->state);
+  tcp_push_hdr_i (tc, b, tc->state, 0);
   tc->rtx_bytes += n_bytes;
 
 done:
@@ -1097,7 +1130,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 
       vlib_buffer_make_headroom (b, MAX_HDRS_LEN);
 
-      tcp_push_hdr_i (tc, b, tc->state);
+      tcp_push_hdr_i (tc, b, tc->state, 1);
 
       /* Account for the SYN */
       tc->snd_nxt += 1;
@@ -1168,6 +1201,7 @@ tcp_timer_persist_handler (u32 index)
   /* Try to force the first unsent segment  */
   tcp_get_free_buffer_index (tm, &bi);
   b = vlib_get_buffer (vm, bi);
+  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
   n_bytes = stream_session_peek_bytes (&tc->connection,
 				       vlib_buffer_get_current (b),
 				       tc->snd_una_max - tc->snd_una,
@@ -1180,7 +1214,7 @@ tcp_timer_persist_handler (u32 index)
     }
 
   b->current_length = n_bytes;
-  tcp_push_hdr_i (tc, b, tc->state);
+  tcp_push_hdr_i (tc, b, tc->state, 0);
   tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
 
   /* Re-enable persist timer */
@@ -1507,7 +1541,7 @@ tcp_push_header (transport_connection_t * tconn, vlib_buffer_t * b)
   tcp_connection_t *tc;
 
   tc = (tcp_connection_t *) tconn;
-  tcp_push_hdr_i (tc, b, TCP_STATE_ESTABLISHED);
+  tcp_push_hdr_i (tc, b, TCP_STATE_ESTABLISHED, 0);
   return 0;
 }