tcp: use sacks for timer based recovery

Type: feature

If available, reuse sack scoreboard in timer triggered retransmit to
minimize spurious retransmits.

Additional changes/refactoring:
- limited transmit updates
- add sacked rxt count to scoreboard
- prr pacing of fast retransmits
- startup pacing updates
- changed loss window to flight + mss

Change-Id: I057de6a9d6401698bd1031d5cf5cfbb62f2bdf61
Signed-off-by: Florin Coras <fcoras@cisco.com>
(cherry picked from commit 36ebcfffbc7ab0e83b4bb8dfaec16bf16cafb954)

src/vnet/session/transport.c

@@ -578,7 +578,8 @@ spacer_max_burst (spacer_t * pacer, u64 norm_time_now)
   u64 n_periods = norm_time_now - pacer->last_update;
   u64 inc;
 
-  if (n_periods > 0 && (inc = n_periods * pacer->tokens_per_period) > 10)
+  if (n_periods > 0
+      && (inc = (f32) n_periods * pacer->tokens_per_period) > 10)
     {
       pacer->last_update = norm_time_now;
       pacer->bucket = clib_min (pacer->bucket + inc, pacer->bytes_per_sec);
@@ -608,6 +609,13 @@ spacer_pace_rate (spacer_t * pacer)
   return pacer->bytes_per_sec;
 }
 
+static inline void
+spacer_reset_bucket (spacer_t * pacer, u64 norm_time_now)
+{
+  pacer->last_update = norm_time_now;
+  pacer->bucket = 0;
+}
+
 void
 transport_connection_tx_pacer_reset (transport_connection_t * tc,
 				     u32 rate_bytes_per_sec,
@@ -646,6 +654,14 @@ transport_connection_tx_pacer_burst (transport_connection_t * tc,
   return spacer_max_burst (&tc->pacer, time_now);
 }
 
+void
+transport_connection_tx_pacer_reset_bucket (transport_connection_t * tc,
+					    u64 time_now)
+{
+  time_now >>= SPACER_CPU_TICKS_PER_PERIOD_SHIFT;
+  spacer_reset_bucket (&tc->pacer, time_now);
+}
+
 u32
 transport_connection_snd_space (transport_connection_t * tc, u64 time_now,
 				u16 mss)

src/vnet/session/transport.h

@@ -219,6 +219,15 @@ u32 transport_connection_tx_pacer_burst (transport_connection_t * tc,
  */
 u64 transport_connection_tx_pacer_rate (transport_connection_t * tc);
 
+/**
+ * Reset tx pacer bucket
+ *
+ * @param tc		transport connection
+ * @param time_now	current cpu time
+ */
+void transport_connection_tx_pacer_reset_bucket (transport_connection_t * tc,
+						 u64 time_now);
+
 /**
  * Initialize period for tx pacers
  *

src/vnet/tcp/tcp.c

@@ -699,17 +699,15 @@ tcp_init_snd_vars (tcp_connection_t * tc)
   tc->snd_una = tc->iss;
   tc->snd_nxt = tc->iss + 1;
   tc->snd_una_max = tc->snd_nxt;
-  tc->srtt = 0;
+  tc->srtt = 100;		/* 100 ms */
 }
 
 void
 tcp_enable_pacing (tcp_connection_t * tc)
 {
-  u32 initial_bucket, byte_rate;
-  initial_bucket = 16 * tc->snd_mss;
-  byte_rate = 2 << 16;
-  transport_connection_tx_pacer_init (&tc->connection, byte_rate,
-				      initial_bucket);
+  u32 byte_rate;
+  byte_rate = tc->cwnd / (tc->srtt * TCP_TICK);
+  transport_connection_tx_pacer_init (&tc->connection, byte_rate, tc->cwnd);
   tc->mrtt_us = (u32) ~ 0;
 }
 
@@ -723,10 +721,11 @@ tcp_connection_init_vars (tcp_connection_t * tc)
   tcp_connection_timers_init (tc);
   tcp_init_mss (tc);
   scoreboard_init (&tc->sack_sb);
-  tcp_cc_init (tc);
   if (tc->state == TCP_STATE_SYN_RCVD)
     tcp_init_snd_vars (tc);
 
+  tcp_cc_init (tc);
+
   if (!tc->c_is_ip4 && ip6_address_is_link_local_unicast (&tc->c_rmt_ip6))
     tcp_add_del_adjacency (tc, 1);
 
@@ -919,7 +918,7 @@ format_tcp_congestion (u8 * s, va_list * args)
   s = format (s, "%U ", format_tcp_congestion_status, tc);
   s = format (s, "algo %s cwnd %u ssthresh %u bytes_acked %u\n",
 	      tc->cc_algo->name, tc->cwnd, tc->ssthresh, tc->bytes_acked);
-  s = format (s, "%Ucc space %u prev_cwnd %u prev_ssthresh %u rtx_bytes %u\n",
+  s = format (s, "%Ucc space %u prev_cwnd %u prev_ssthresh %u rxt_bytes %u\n",
 	      format_white_space, indent, tcp_available_cc_snd_space (tc),
 	      tc->prev_cwnd, tc->prev_ssthresh, tc->snd_rxt_bytes);
   s = format (s, "%Usnd_congestion %u dupack %u limited_transmit %u\n",
@@ -1139,8 +1138,9 @@ format_tcp_scoreboard (u8 * s, va_list * args)
   sack_scoreboard_hole_t *hole;
   u32 indent = format_get_indent (s);
 
-  s = format (s, "sacked_bytes %u last_sacked_bytes %u lost_bytes %u\n",
-	      sb->sacked_bytes, sb->last_sacked_bytes, sb->lost_bytes);
+  s = format (s, "sacked %u last_sacked %u lost %u last_lost %u\n",
+	      sb->sacked_bytes, sb->last_sacked_bytes, sb->lost_bytes,
+	      sb->last_lost_bytes);
   s = format (s, "%Ulast_bytes_delivered %u high_sacked %u is_reneging %u\n",
 	      format_white_space, indent, sb->last_bytes_delivered,
 	      sb->high_sacked - tc->iss, sb->is_reneging);
@@ -1245,7 +1245,7 @@ tcp_round_snd_space (tcp_connection_t * tc, u32 snd_space)
 static inline u32
 tcp_snd_space_inline (tcp_connection_t * tc)
 {
-  int snd_space, snt_limited;
+  int snd_space;
 
   if (PREDICT_FALSE (tcp_in_fastrecovery (tc)
 		     || tc->state == TCP_STATE_CLOSED))
@@ -1253,18 +1253,21 @@ tcp_snd_space_inline (tcp_connection_t * tc)
 
   snd_space = tcp_available_output_snd_space (tc);
 
-  /* If we haven't gotten dupacks or if we did and have gotten sacked
-   * bytes then we can still send as per Limited Transmit (RFC3042) */
-  if (PREDICT_FALSE (tc->rcv_dupacks != 0
-		     && (tcp_opts_sack_permitted (tc)
-			 && tc->sack_sb.last_sacked_bytes == 0)))
+  /* If we got dupacks or sacked bytes but we're not yet in recovery, try
+   * to force the peer to send enough dupacks to start retransmitting as
+   * per Limited Transmit (RFC3042)
+   */
+  if (PREDICT_FALSE (tc->rcv_dupacks != 0 || tc->sack_sb.sacked_bytes))
     {
-      if (tc->rcv_dupacks == 1 && tc->limited_transmit != tc->snd_nxt)
+      if (tc->limited_transmit != tc->snd_nxt
+	  && (seq_lt (tc->limited_transmit, tc->snd_nxt - 2 * tc->snd_mss)
+	      || seq_gt (tc->limited_transmit, tc->snd_nxt)))
 	tc->limited_transmit = tc->snd_nxt;
+
       ASSERT (seq_leq (tc->limited_transmit, tc->snd_nxt));
 
-      snt_limited = tc->snd_nxt - tc->limited_transmit;
-      snd_space = clib_max (2 * tc->snd_mss - snt_limited, 0);
+      int snt_limited = tc->snd_nxt - tc->limited_transmit;
+      snd_space = clib_max ((int) 2 * tc->snd_mss - snt_limited, 0);
     }
   return tcp_round_snd_space (tc, snd_space);
 }
@@ -1358,9 +1361,9 @@ tcp_connection_tx_pacer_reset (tcp_connection_t * tc, u32 window,
 			       u32 start_bucket)
 {
   tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
-  u32 byte_rate = window / ((f64) TCP_TICK * tc->srtt);
+  f64 srtt = clib_min ((f64) tc->srtt * TCP_TICK, tc->mrtt_us);
   u64 last_time = wrk->vm->clib_time.last_cpu_time;
-  transport_connection_tx_pacer_reset (&tc->connection, byte_rate,
+  transport_connection_tx_pacer_reset (&tc->connection, window / srtt,
 				       start_bucket, last_time);
 }
 

src/vnet/tcp/tcp.h

@@ -32,6 +32,7 @@
 #define TCP_MAX_OPTION_SPACE 40
 #define TCP_CC_DATA_SZ 24
 #define TCP_MAX_GSO_SZ 65536
+#define TCP_RXT_MAX_BURST 10
 
 #define TCP_DUPACK_THRESHOLD 	3
 #define TCP_IW_N_SEGMENTS 	10
@@ -111,7 +112,7 @@ extern timer_expiration_handler tcp_timer_retransmit_syn_handler;
   _(DCNT_PENDING, "Disconnect pending")		\
   _(HALF_OPEN_DONE, "Half-open completed")	\
   _(FINPNDG, "FIN pending")			\
-  _(FRXT_PENDING, "Fast-retransmit pending")	\
+  _(RXT_PENDING, "Retransmit pending")		\
   _(FRXT_FIRST, "Fast-retransmit first again")	\
   _(DEQ_PENDING, "Pending dequeue acked")	\
   _(PSH_PENDING, "PSH pending")			\
@@ -166,6 +167,7 @@ typedef struct _sack_scoreboard
   u32 sacked_bytes;			/**< Number of bytes sacked in sb */
   u32 last_sacked_bytes;		/**< Number of bytes last sacked */
   u32 last_bytes_delivered;		/**< Sack bytes delivered to app */
+  u32 rxt_sacked;			/**< Rxt last delivered */
   u32 high_sacked;			/**< Highest byte sacked (fack) */
   u32 high_rxt;				/**< Highest retransmitted sequence */
   u32 rescue_rxt;			/**< Rescue sequence number */
@@ -219,8 +221,11 @@ sack_scoreboard_hole_t *scoreboard_prev_hole (sack_scoreboard_t * sb,
 					      sack_scoreboard_hole_t * hole);
 sack_scoreboard_hole_t *scoreboard_first_hole (sack_scoreboard_t * sb);
 sack_scoreboard_hole_t *scoreboard_last_hole (sack_scoreboard_t * sb);
+
 void scoreboard_clear (sack_scoreboard_t * sb);
+void scoreboard_clear_reneging (sack_scoreboard_t * sb, u32 start, u32 end);
 void scoreboard_init (sack_scoreboard_t * sb);
+void scoreboard_init_high_rxt (sack_scoreboard_t * sb, u32 snd_una);
 u8 *format_tcp_scoreboard (u8 * s, va_list * args);
 
 #define TCP_BTS_INVALID_INDEX	((u32)~0)
@@ -360,8 +365,10 @@ typedef struct _tcp_connection
   u32 prev_cwnd;	/**< ssthresh before congestion */
   u32 bytes_acked;	/**< Bytes acknowledged by current segment */
   u32 burst_acked;	/**< Bytes acknowledged in current burst */
-  u32 snd_rxt_bytes;	/**< Retransmitted bytes */
+  u32 snd_rxt_bytes;	/**< Retransmitted bytes during current cc event */
   u32 snd_rxt_ts;	/**< Timestamp when first packet is retransmitted */
+  u32 prr_delivered;	/**< RFC6937 bytes delivered during current event */
+  u32 rxt_delivered;	/**< Rxt bytes delivered during current cc event */
   u32 tsecr_last_ack;	/**< Timestamp echoed to us in last healthy ACK */
   u32 snd_congestion;	/**< snd_una_max when congestion is detected */
   u32 tx_fifo_size;	/**< Tx fifo size. Used to constrain cwnd */
@@ -757,7 +764,7 @@ void tcp_send_window_update_ack (tcp_connection_t * tc);
 
 void tcp_program_ack (tcp_connection_t * tc);
 void tcp_program_dupack (tcp_connection_t * tc);
-void tcp_program_fastretransmit (tcp_connection_t * tc);
+void tcp_program_retransmit (tcp_connection_t * tc);
 
 /*
  * Rate estimation
@@ -857,18 +864,9 @@ tcp_flight_size (const tcp_connection_t * tc)
   int flight_size;
 
   flight_size = (int) (tc->snd_nxt - tc->snd_una) - tcp_bytes_out (tc)
-    + tc->snd_rxt_bytes;
+    + tc->snd_rxt_bytes - tc->rxt_delivered;
 
-  if (flight_size < 0)
-    {
-      if (0)
-	clib_warning
-	  ("Negative: %u %u %u dupacks %u sacked bytes %u flags %d",
-	   tc->snd_una_max - tc->snd_una, tcp_bytes_out (tc),
-	   tc->snd_rxt_bytes, tc->rcv_dupacks, tc->sack_sb.sacked_bytes,
-	   tc->rcv_opts.flags);
-      return 0;
-    }
+  ASSERT (flight_size >= 0);
 
   return flight_size;
 }
@@ -912,7 +910,8 @@ tcp_cwnd_accumulate (tcp_connection_t * tc, u32 thresh, u32 bytes)
 always_inline u32
 tcp_loss_wnd (const tcp_connection_t * tc)
 {
-  return tc->snd_mss;
+  /* Whatever we have in flight + the packet we're about to send */
+  return tcp_flight_size (tc) + tc->snd_mss;
 }
 
 always_inline u32
@@ -951,22 +950,14 @@ tcp_available_cc_snd_space (const tcp_connection_t * tc)
 always_inline u8
 tcp_is_lost_fin (tcp_connection_t * tc)
 {
-  if ((tc->flags & TCP_CONN_FINSNT) && tc->snd_una_max - tc->snd_una == 1)
+  if ((tc->flags & TCP_CONN_FINSNT) && (tc->snd_una_max - tc->snd_una == 1))
     return 1;
   return 0;
 }
 
 u32 tcp_snd_space (tcp_connection_t * tc);
-int tcp_retransmit_first_unacked (tcp_worker_ctx_t * wrk,
-				  tcp_connection_t * tc);
-int tcp_fast_retransmit_no_sack (tcp_worker_ctx_t * wrk,
-				 tcp_connection_t * tc, u32 burst_size);
-int tcp_fast_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
-			      u32 burst_size);
-int tcp_fast_retransmit (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
-			 u32 burst_size);
-void tcp_cc_init_congestion (tcp_connection_t * tc);
-void tcp_cc_fastrecovery_clear (tcp_connection_t * tc);
+//void tcp_cc_init_congestion (tcp_connection_t * tc);
+//void tcp_cc_fastrecovery_clear (tcp_connection_t * tc);
 
 fib_node_index_t tcp_lookup_rmt_in_fib (tcp_connection_t * tc);
 
@@ -1035,6 +1026,12 @@ tcp_cc_rcv_cong_ack (tcp_connection_t * tc, tcp_cc_ack_t ack_type,
   tc->cc_algo->rcv_cong_ack (tc, ack_type, rs);
 }
 
+static inline void
+tcp_cc_congestion (tcp_connection_t * tc)
+{
+  tc->cc_algo->congestion (tc);
+}
+
 static inline void
 tcp_cc_loss (tcp_connection_t * tc)
 {
@@ -1068,9 +1065,10 @@ tcp_cc_get_pacing_rate (tcp_connection_t * tc)
     return tc->cc_algo->get_pacing_rate (tc);
 
   f64 srtt = clib_min ((f64) tc->srtt * TCP_TICK, tc->mrtt_us);
+
   /* TODO should constrain to interface's max throughput but
    * we don't have link speeds for sw ifs ..*/
-  return (tc->cwnd / srtt);
+  return ((f64) tc->cwnd / srtt);
 }
 
 always_inline void

src/vnet/tcp/tcp_cubic.c

@@ -103,10 +103,7 @@ cubic_congestion (tcp_connection_t * tc)
 
   cd->w_max = w_max;
   tc->ssthresh = clib_max (tc->cwnd * beta_cubic, 2 * tc->snd_mss);
-
   tc->cwnd = tc->ssthresh;
-  if (!tcp_opts_sack_permitted (&tc->rcv_opts))
-    tc->cwnd += 3 * tc->snd_mss;
 }
 
 static void
@@ -114,11 +111,10 @@ cubic_loss (tcp_connection_t * tc)
 {
   cubic_data_t *cd = (cubic_data_t *) tcp_cc_data (tc);
 
-  tc->ssthresh = clib_max (tc->cwnd * beta_cubic, 2 * tc->snd_mss);
   tc->cwnd = tcp_loss_wnd (tc);
   cd->t_start = cubic_time (tc->c_thread_index);
   cd->K = 0;
-  cd->w_max = 0;
+  cd->w_max = tc->cwnd / tc->snd_mss;
 }
 
 static void
@@ -159,7 +155,7 @@ cubic_rcv_ack (tcp_connection_t * tc, tcp_rate_sample_t * rs)
 
   if (tcp_in_slowstart (tc))
     {
-      tc->cwnd += clib_min (tc->snd_mss, tc->bytes_acked);
+      tc->cwnd += tc->bytes_acked;
       return;
     }
 

src/vnet/tcp/tcp_newreno.c

@@ -20,17 +20,11 @@ newreno_congestion (tcp_connection_t * tc)
 {
   tc->ssthresh = clib_max (tcp_flight_size (tc) / 2, 2 * tc->snd_mss);
   tc->cwnd = tc->ssthresh;
-  /* Post retransmit update cwnd to ssthresh and account for the
-   * three segments that have left the network and should've been
-   * buffered at the receiver XXX */
-  if (!tcp_opts_sack_permitted (&tc->rcv_opts))
-    tc->cwnd += 3 * tc->snd_mss;
 }
 
 static void
 newreno_loss (tcp_connection_t * tc)
 {
-  tc->ssthresh = clib_max (tcp_flight_size (tc) / 2, 2 * tc->snd_mss);
   tc->cwnd = tcp_loss_wnd (tc);
 }
 

src/vnet/tcp/tcp_packet.h

@@ -143,9 +143,8 @@ typedef struct _sack_block
 typedef struct
 {
   u8 flags;		/** Option flags, see above */
-
-  u16 mss;		/**< Maximum segment size advertised */
   u8 wscale;		/**< Window scale advertised */
+  u16 mss;		/**< Maximum segment size advertised */
   u32 tsval;		/**< Timestamp value */
   u32 tsecr;		/**< Echoed/reflected time stamp */
   sack_block_t *sacks;	/**< SACK blocks */