tcp: refactor flags into cfg and conn flags

Type: refactor

Change-Id: Ib5e2ecf55732599e1322736e3662178c9da49c7d
Signed-off-by: Florin Coras <fcoras@cisco.com>

4 files changed, 89 insertions, 31 deletions

diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 6ef03dc093d..86e5d949bd1 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -271,7 +271,7 @@ tcp_connection_cleanup (tcp_connection_t * tc)
       vec_free (tc->rcv_opts.sacks);
       pool_free (tc->sack_sb.holes);
 
-      if (tc->flags & TCP_CONN_RATE_SAMPLE)
+      if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
 	tcp_bt_cleanup (tc);
 
       /* Poison the entry */
@@ -737,9 +737,12 @@ tcp_connection_init_vars (tcp_connection_t * tc)
       || tcp_cfg.enable_tx_pacing)
     tcp_enable_pacing (tc);
 
-  if (tc->flags & TCP_CONN_RATE_SAMPLE)
+  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
     tcp_bt_init (tc);
 
+  if (!tcp_cfg.allow_tso)
+    tc->cfg_flags |= TCP_CFG_F_NO_TSO;
+
   tc->start_ts = tcp_time_now_us (tc->c_thread_index);
 }
 
@@ -839,6 +842,31 @@ format_tcp_state (u8 * s, va_list * args)
   return s;
 }
 
+const char *tcp_cfg_flags_str[] = {
+#define _(sym, str) str,
+  foreach_tcp_cfg_flag
+#undef _
+};
+
+static u8 *
+format_tcp_cfg_flags (u8 * s, va_list * args)
+{
+  tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
+  int i, last = -1;
+
+  for (i = 0; i < TCP_CFG_N_FLAG_BITS; i++)
+    if (tc->cfg_flags & (1 << i))
+      last = i;
+  for (i = 0; i < last; i++)
+    {
+      if (tc->cfg_flags & (1 << i))
+	s = format (s, "%s, ", tcp_cfg_flags_str[i]);
+    }
+  if (last >= 0)
+    s = format (s, "%s", tcp_cfg_flags_str[last]);
+  return s;
+}
+
 const char *tcp_connection_flags_str[] = {
 #define _(sym, str) str,
   foreach_tcp_connection_flag
@@ -963,8 +991,9 @@ static u8 *
 format_tcp_vars (u8 * s, va_list * args)
 {
   tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
-  s = format (s, " index: %u flags: %U timers: %U\n", tc->c_c_index,
-	      format_tcp_connection_flags, tc, format_tcp_timers, tc);
+  s = format (s, " index: %u cfg: %U flags: %U timers: %U\n", tc->c_c_index,
+	      format_tcp_cfg_flags, tc, format_tcp_connection_flags, tc,
+	      format_tcp_timers, tc);
   s = format (s, " snd_una %u snd_nxt %u snd_una_max %u",
 	      tc->snd_una - tc->iss, tc->snd_nxt - tc->iss,
 	      tc->snd_una_max - tc->iss);
@@ -1219,10 +1248,8 @@ tcp_session_send_mss (transport_connection_t * trans_conn)
    * the current state of the connection. */
   tcp_update_burst_snd_vars (tc);
 
-  if (PREDICT_FALSE (tc->is_tso))
-    {
-      return tcp_session_cal_goal_size (tc);
-    }
+  if (PREDICT_FALSE (tc->cfg_flags & TCP_CFG_F_TSO))
+    return tcp_session_cal_goal_size (tc);
 
   return tc->snd_mss;
 }
@@ -1621,6 +1648,7 @@ tcp_configuration_init (void)
   tcp_cfg.default_mtu = 1500;
   tcp_cfg.initial_cwnd_multiplier = 0;
   tcp_cfg.enable_tx_pacing = 1;
+  tcp_cfg.allow_tso = 0;
   tcp_cfg.cc_algo = TCP_CC_NEWRENO;
   tcp_cfg.rwnd_min_update_ack = 1;
 
@@ -1744,6 +1772,8 @@ tcp_config_fn (vlib_main_t * vm, unformat_input_t * input)
 	tcp_cfg.initial_cwnd_multiplier = cwnd_multiplier;
       else if (unformat (input, "no-tx-pacing"))
 	tcp_cfg.enable_tx_pacing = 0;
+      else if (unformat (input, "tso"))
+	tcp_cfg.allow_tso = 1;
       else if (unformat (input, "cc-algo %U", unformat_tcp_cc_algo,
 			 &tcp_cfg.cc_algo))
 	;
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 5e683f72dbd..097a1475185 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -103,6 +103,29 @@ extern timer_expiration_handler tcp_timer_retransmit_syn_handler;
 #define TCP_RTO_BOFF_MAX 8	/* Max number of retries before reset */
 #define TCP_ESTABLISH_TIME (60 * THZ)	/* Connection establish timeout */
 
+/** Connection configuration flags */
+#define foreach_tcp_cfg_flag 			\
+  _(RATE_SAMPLE, "Rate sampling")		\
+  _(NO_CSUM_OFFLOAD, "No csum offload")    	\
+  _(NO_TSO, "TSO off")				\
+  _(TSO, "TSO")					\
+
+typedef enum tcp_cfg_flag_bits_
+{
+#define _(sym, str) TCP_CFG_F_##sym##_BIT,
+  foreach_tcp_cfg_flag
+#undef _
+  TCP_CFG_N_FLAG_BITS
+} tcp_cfg_flag_bits_e;
+
+typedef enum tcp_cfg_flag_
+{
+#define _(sym, str) TCP_CFG_F_##sym = 1 << TCP_CFG_F_##sym##_BIT,
+  foreach_tcp_cfg_flag
+#undef _
+  TCP_CFG_N_FLAGS
+} tcp_cfg_flags_e;
+
 /** TCP connection flags */
 #define foreach_tcp_connection_flag             \
   _(SNDACK, "Send ACK")                         \
@@ -113,16 +136,14 @@ extern timer_expiration_handler tcp_timer_retransmit_syn_handler;
   _(HALF_OPEN_DONE, "Half-open completed")	\
   _(FINPNDG, "FIN pending")			\
   _(RXT_PENDING, "Retransmit pending")		\
-  _(FRXT_FIRST, "Fast-retransmit first again")	\
-  _(DEQ_PENDING, "Pending dequeue acked")	\
+  _(FRXT_FIRST, "Retransmit first")		\
+  _(DEQ_PENDING, "Dequeue pending ")		\
   _(PSH_PENDING, "PSH pending")			\
   _(FINRCVD, "FIN received")			\
-  _(RATE_SAMPLE, "Conn does rate sampling")	\
   _(TRACK_BURST, "Track burst")			\
   _(ZERO_RWND_SENT, "Zero RWND sent")		\
-  _(NO_CSUM_OFFLOAD, "No Checksum Offload")     \
 
-typedef enum _tcp_connection_flag_bits
+typedef enum tcp_connection_flag_bits_
 {
 #define _(sym, str) TCP_CONN_##sym##_BIT,
   foreach_tcp_connection_flag
@@ -130,7 +151,7 @@ typedef enum _tcp_connection_flag_bits
   TCP_CONN_N_FLAG_BITS
 } tcp_connection_flag_bits_e;
 
-typedef enum _tcp_connection_flag
+typedef enum tcp_connection_flag_
 {
 #define _(sym, str) TCP_CONN_##sym = 1 << TCP_CONN_##sym##_BIT,
   foreach_tcp_connection_flag
@@ -310,7 +331,7 @@ typedef struct _tcp_connection
   transport_connection_t connection;  /**< Common transport data. First! */
 
   u8 state;			/**< TCP state as per tcp_state_t */
-  u8 is_tso;	  /** is connection could use tso */
+  u8 cfg_flags;			/**< Connection configuration flags */
   u16 flags;			/**< Connection flags (see tcp_conn_flags_e) */
   u32 timers[TCP_N_TIMERS];	/**< Timer handles into timer wheel */
 
@@ -451,6 +472,8 @@ struct _tcp_cc_algorithm
 #define tcp_in_cong_recovery(tc) ((tc)->flags & 		\
 	  (TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY))
 
+#define tcp_csum_offload(tc) (!((tc)->cfg_flags & TCP_CFG_F_NO_CSUM_OFFLOAD))
+
 always_inline void
 tcp_cong_recovery_off (tcp_connection_t * tc)
 {
@@ -534,6 +557,9 @@ typedef struct tcp_configuration_
   /** Enable tx pacing for new connections */
   u8 enable_tx_pacing;
 
+  /** Allow use of TSO whenever available */
+  u8 allow_tso;
+
   /** Default congestion control algorithm type */
   tcp_cc_algorithm_type_e cc_algo;
 
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index ea64e4ab754..61e5aa8c662 100755
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -503,7 +503,8 @@ tcp_update_rtt (tcp_connection_t * tc, tcp_rate_sample_t * rs, u32 ack)
   if (tcp_in_cong_recovery (tc))
     {
       /* Accept rtt estimates for samples that have not been retransmitted */
-      if ((tc->flags & TCP_CONN_RATE_SAMPLE) && !(rs->flags & TCP_BTS_IS_RXT))
+      if ((tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
+	  && !(rs->flags & TCP_BTS_IS_RXT))
 	{
 	  mrtt = rs->rtt_time * THZ;
 	  goto estimate_rtt;
@@ -1604,7 +1605,7 @@ process_ack:
   tc->snd_una = vnet_buffer (b)->tcp.ack_number;
   tcp_validate_txf_size (tc, tc->bytes_acked);
 
-  if (tc->flags & TCP_CONN_RATE_SAMPLE)
+  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
     tcp_bt_sample_delivery_rate (tc, &rs);
 
   tcp_program_dequeue (wrk, tc);
@@ -2361,11 +2362,10 @@ tcp_check_tx_offload (tcp_connection_t * tc, int is_ipv4)
   sw_if_idx = dpo->dpoi_index;
   hw_if = vnet_get_sup_hw_interface (vnm, sw_if_idx);
 
-  tc->is_tso =
-    ((hw_if->flags & VNET_HW_INTERFACE_FLAG_SUPPORTS_GSO) == 0) ? 0 : 1;
+  if (hw_if->flags & VNET_HW_INTERFACE_FLAG_SUPPORTS_GSO)
+    tc->cfg_flags |= TCP_CFG_F_TSO;
 }
 
-
 always_inline uword
 tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		       vlib_frame_t * from_frame, int is_ip4)
@@ -2582,7 +2582,8 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  goto drop;
 	}
 
-      tcp_check_tx_offload (new_tc0, is_ip4);
+      if (!(new_tc0->cfg_flags & TCP_CFG_F_NO_TSO))
+	tcp_check_tx_offload (new_tc0, is_ip4);
 
       /* Read data, if any */
       if (PREDICT_FALSE (vnet_buffer (b0)->tcp.data_len))
@@ -2771,7 +2772,8 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  tc0->state = TCP_STATE_ESTABLISHED;
 	  TCP_EVT (TCP_EVT_STATE_CHANGE, tc0);
 
-	  tcp_check_tx_offload (tc0, is_ip4);
+	  if (!(tc0->cfg_flags & TCP_CFG_F_NO_TSO))
+	    tcp_check_tx_offload (tc0, is_ip4);
 
 	  /* Initialize session variables */
 	  tc0->snd_una = vnet_buffer (b0)->tcp.ack_number;
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 047247e9ebe..fb806446dcc 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -405,7 +405,7 @@ tcp_update_burst_snd_vars (tcp_connection_t * tc)
 
   tcp_update_rcv_wnd (tc);
 
-  if (tc->flags & TCP_CONN_RATE_SAMPLE)
+  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
     tc->flags |= TCP_CONN_TRACK_BURST;
 
   if (tc->snd_una == tc->snd_nxt)
@@ -499,7 +499,7 @@ static inline u16
 tcp_compute_checksum (tcp_connection_t * tc, vlib_buffer_t * b)
 {
   u16 checksum = 0;
-  if (PREDICT_FALSE (tc->flags & TCP_CONN_NO_CSUM_OFFLOAD))
+  if (PREDICT_FALSE (tc->cfg_flags & TCP_CFG_F_NO_CSUM_OFFLOAD))
     {
       tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
       vlib_main_t *vm = wrk->vm;
@@ -867,7 +867,7 @@ tcp_send_reset_w_pkt (tcp_connection_t * tc, vlib_buffer_t * pkt,
       ASSERT ((pkt_ih4->ip_version_and_header_length & 0xF0) == 0x40);
       ih4 = vlib_buffer_push_ip4 (vm, b, &pkt_ih4->dst_address,
 				  &pkt_ih4->src_address, IP_PROTOCOL_TCP,
-				  (!(tc->flags & TCP_CONN_NO_CSUM_OFFLOAD)));
+				  tcp_csum_offload (tc));
       th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih4);
     }
   else
@@ -934,7 +934,7 @@ tcp_push_ip_hdr (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
       ip4_header_t *ih;
       ih = vlib_buffer_push_ip4 (vm, b, &tc->c_lcl_ip4,
 				 &tc->c_rmt_ip4, IP_PROTOCOL_TCP,
-				 (!(tc->flags & TCP_CONN_NO_CSUM_OFFLOAD)));
+				 tcp_csum_offload (tc));
       th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih);
     }
   else
@@ -1445,7 +1445,7 @@ tcp_prepare_retransmit_segment (tcp_worker_ctx_t * wrk,
 
   tc->snd_rxt_bytes += n_bytes;
 
-  if (tc->flags & TCP_CONN_RATE_SAMPLE)
+  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
     tcp_bt_track_rxt (tc, start, start + n_bytes);
 
   tc->bytes_retrans += n_bytes;
@@ -1777,7 +1777,7 @@ tcp_timer_persist_handler (u32 index)
 			   || tc->snd_nxt == tc->snd_una_max
 			   || tc->rto_boff > 1));
 
-  if (tc->flags & TCP_CONN_RATE_SAMPLE)
+  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
     {
       tcp_bt_check_app_limited (tc);
       tcp_bt_track_tx (tc);
@@ -2301,8 +2301,7 @@ tcp_output_push_ip (vlib_main_t * vm, vlib_buffer_t * b0,
 
   if (is_ip4)
     ih0 = vlib_buffer_push_ip4 (vm, b0, &tc0->c_lcl_ip4, &tc0->c_rmt_ip4,
-				IP_PROTOCOL_TCP,
-				(!(tc0->flags & TCP_CONN_NO_CSUM_OFFLOAD)));
+				IP_PROTOCOL_TCP, tcp_csum_offload (tc0));
   else
     ih0 = vlib_buffer_push_ip6 (vm, b0, &tc0->c_lcl_ip6, &tc0->c_rmt_ip6,
 				IP_PROTOCOL_TCP);
@@ -2312,8 +2311,9 @@ tcp_output_push_ip (vlib_main_t * vm, vlib_buffer_t * b0,
 always_inline void
 tcp_check_if_gso (tcp_connection_t * tc, vlib_buffer_t * b)
 {
-  if (!tc->is_tso)
+  if (PREDICT_TRUE (!(tc->cfg_flags & TCP_CFG_F_TSO)))
     return;
+
   u16 data_len = b->current_length - sizeof (tcp_header_t) - tc->snd_opts_len;
 
   if (PREDICT_FALSE (b->flags & VLIB_BUFFER_TOTAL_LENGTH_VALID))