tcp: improve mss computation

Type: fix

Use different mss for v4 and v6 packets.

Change-Id: Ifcd322c1fd74f749a42bae671dd6cfed8c32b150
Signed-off-by: Florin Coras <fcoras@cisco.com>

3 files changed, 48 insertions, 43 deletions

diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index bfab1d6f0c5..cd7b9e80c7a 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -638,6 +638,47 @@ tcp_generate_random_iss (tcp_connection_t * tc)
 }
 
 /**
+ * Initialize max segment size we're able to process.
+ *
+ * The value is constrained by the output interface's MTU and by the size
+ * of the IP and TCP headers (see RFC6691). It is also what we advertise
+ * to our peer.
+ */
+static void
+tcp_init_rcv_mss (tcp_connection_t * tc)
+{
+  u8 ip_hdr_len;
+
+  ip_hdr_len = tc->c_is_ip4 ? sizeof (ip4_header_t) : sizeof (ip6_header_t);
+  tc->mss = tcp_cfg.default_mtu - sizeof (tcp_header_t) - ip_hdr_len;
+}
+
+static void
+tcp_init_mss (tcp_connection_t * tc)
+{
+  u16 default_min_mss = 536;
+
+  tcp_init_rcv_mss (tc);
+
+  /* TODO consider PMTU discovery */
+  tc->snd_mss = clib_min (tc->rcv_opts.mss, tc->mss);
+
+  if (tc->snd_mss < 45)
+    {
+      /* Assume that at least the min default mss works */
+      tc->snd_mss = default_min_mss;
+      tc->rcv_opts.mss = default_min_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->rcv_opts))
+    tc->snd_mss -= TCP_OPTION_LEN_TIMESTAMP;
+}
+
+/**
  * Initialize connection send variables.
  */
 void
@@ -651,6 +692,7 @@ tcp_init_snd_vars (tcp_connection_t * tc)
    */
   tcp_set_time_now (tcp_get_worker (vlib_get_thread_index ()));
 
+  tcp_init_rcv_mss (tc);
   tc->iss = tcp_generate_random_iss (tc);
   tc->snd_una = tc->iss;
   tc->snd_nxt = tc->iss + 1;
@@ -1541,7 +1583,7 @@ tcp_configuration_init (void)
   tcp_cfg.max_rx_fifo = 32 << 20;
   tcp_cfg.min_rx_fifo = 4 << 10;
 
-  tcp_cfg.default_mtu = 1460;
+  tcp_cfg.default_mtu = 1500;
   tcp_cfg.initial_cwnd_multiplier = 0;
   tcp_cfg.enable_tx_pacing = 1;
   tcp_cfg.cc_algo = TCP_CC_NEWRENO;
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index cad105b0c23..d55ff801c00 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -743,7 +743,6 @@ void tcp_send_reset (tcp_connection_t * tc);
 void tcp_send_syn (tcp_connection_t * tc);
 void tcp_send_synack (tcp_connection_t * tc);
 void tcp_send_fin (tcp_connection_t * tc);
-void tcp_init_mss (tcp_connection_t * tc);
 void tcp_update_burst_snd_vars (tcp_connection_t * tc);
 void tcp_update_rto (tcp_connection_t * tc);
 void tcp_flush_frame_to_output (tcp_worker_ctx_t * wrk, u8 is_ip4);
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 862f4a2f307..239f6035d5f 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -76,19 +76,6 @@ tcp_window_compute_scale (u32 window)
 }
 
 /**
- * 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 = tcp_cfg.default_mtu - sizeof (tcp_header_t);
-}
-
-/**
  * TCP's initial window
  */
 always_inline u32
@@ -263,16 +250,16 @@ tcp_options_write (u8 * data, tcp_options_t * opts)
 }
 
 static int
-tcp_make_syn_options (tcp_options_t * opts, u8 wnd_scale)
+tcp_make_syn_options (tcp_connection_t * tc, tcp_options_t * opts)
 {
   u8 len = 0;
 
   opts->flags |= TCP_OPTS_FLAG_MSS;
-  opts->mss = tcp_cfg.default_mtu;	/*XXX discover that */
+  opts->mss = tc->mss;
   len += TCP_OPTION_LEN_MSS;
 
   opts->flags |= TCP_OPTS_FLAG_WSCALE;
-  opts->wscale = wnd_scale;
+  opts->wscale = tc->rcv_wscale;
   len += TCP_OPTION_LEN_WINDOW_SCALE;
 
   opts->flags |= TCP_OPTS_FLAG_TSTAMP;
@@ -379,7 +366,7 @@ tcp_make_options (tcp_connection_t * tc, tcp_options_t * opts,
     case TCP_STATE_SYN_RCVD:
       return tcp_make_synack_options (tc, opts);
     case TCP_STATE_SYN_SENT:
-      return tcp_make_syn_options (opts, tc->rcv_wscale);
+      return tcp_make_syn_options (tc, opts);
     default:
       clib_warning ("State not handled! %d", state);
       return 0;
@@ -422,29 +409,6 @@ tcp_update_burst_snd_vars (tcp_connection_t * tc)
     tcp_cc_event (tc, TCP_CC_EVT_START_TX);
 }
 
-void
-tcp_init_mss (tcp_connection_t * tc)
-{
-  u16 default_min_mss = 536;
-  tcp_update_rcv_mss (tc);
-
-  /* TODO cache mss and consider PMTU discovery */
-  tc->snd_mss = clib_min (tc->rcv_opts.mss, tc->mss);
-
-  if (tc->snd_mss < 45)
-    {
-      /* Assume that at least the min default mss works */
-      tc->snd_mss = default_min_mss;
-      tc->rcv_opts.mss = default_min_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->rcv_opts))
-    tc->snd_mss -= TCP_OPTION_LEN_TIMESTAMP;
-}
 #endif /* CLIB_MARCH_VARIANT */
 
 static void *
@@ -542,7 +506,7 @@ tcp_make_syn (tcp_connection_t * tc, vlib_buffer_t * b)
 
   /* Make and write options */
   clib_memset (&snd_opts, 0, sizeof (snd_opts));
-  tcp_opts_len = tcp_make_syn_options (&snd_opts, tc->rcv_wscale);
+  tcp_opts_len = tcp_make_syn_options (tc, &snd_opts);
   tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
 
   th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->iss,