FD.io VPP  v19.04.4-rc0-5-ge88582fac
Vector Packet Processing
tcp.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2016-2019 Cisco and/or its affiliates.
3  * Licensed under the Apache License, Version 2.0 (the "License");
4  * you may not use this file except in compliance with the License.
5  * You may obtain a copy of the License at:
6  *
7  * http://www.apache.org/licenses/LICENSE-2.0
8  *
9  * Unless required by applicable law or agreed to in writing, software
10  * distributed under the License is distributed on an "AS IS" BASIS,
11  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12  * See the License for the specific language governing permissions and
13  * limitations under the License.
14  */
15 
16 #ifndef _vnet_tcp_h_
17 #define _vnet_tcp_h_
18 
19 #include <vnet/vnet.h>
20 #include <vnet/ip/ip.h>
21 #include <vnet/tcp/tcp_packet.h>
22 #include <vnet/tcp/tcp_timer.h>
23 #include <vnet/session/transport.h>
24 #include <vnet/session/session.h>
25 #include <vnet/tcp/tcp_debug.h>
26 
27 #define TCP_TICK 0.001 /**< TCP tick period (s) */
28 #define THZ (u32) (1/TCP_TICK) /**< TCP tick frequency */
29 #define TCP_TSTAMP_RESOLUTION TCP_TICK /**< Time stamp resolution */
30 #define TCP_PAWS_IDLE 24 * 24 * 60 * 60 * THZ /**< 24 days */
31 #define TCP_FIB_RECHECK_PERIOD 1 * THZ /**< Recheck every 1s */
32 #define TCP_MAX_OPTION_SPACE 40
33 #define TCP_CC_DATA_SZ 24
34 
35 #define TCP_DUPACK_THRESHOLD 3
36 #define TCP_MAX_RX_FIFO_SIZE 32 << 20
37 #define TCP_MIN_RX_FIFO_SIZE 4 << 10
38 #define TCP_IW_N_SEGMENTS 10
39 #define TCP_ALWAYS_ACK 1 /**< On/off delayed acks */
40 #define TCP_USE_SACKS 1 /**< Disable only for testing */
41 
42 /** TCP FSM state definitions as per RFC793. */
43 #define foreach_tcp_fsm_state \
44  _(CLOSED, "CLOSED") \
45  _(LISTEN, "LISTEN") \
46  _(SYN_SENT, "SYN_SENT") \
47  _(SYN_RCVD, "SYN_RCVD") \
48  _(ESTABLISHED, "ESTABLISHED") \
49  _(CLOSE_WAIT, "CLOSE_WAIT") \
50  _(FIN_WAIT_1, "FIN_WAIT_1") \
51  _(LAST_ACK, "LAST_ACK") \
52  _(CLOSING, "CLOSING") \
53  _(FIN_WAIT_2, "FIN_WAIT_2") \
54  _(TIME_WAIT, "TIME_WAIT")
55 
56 typedef enum _tcp_state
57 {
58 #define _(sym, str) TCP_STATE_##sym,
60 #undef _
62 } tcp_state_t;
63 
68 
69 /** TCP timers */
70 #define foreach_tcp_timer \
71  _(RETRANSMIT, "RETRANSMIT") \
72  _(DELACK, "DELAYED ACK") \
73  _(PERSIST, "PERSIST") \
74  _(KEEP, "KEEP") \
75  _(WAITCLOSE, "WAIT CLOSE") \
76  _(RETRANSMIT_SYN, "RETRANSMIT SYN") \
77  _(ESTABLISH, "ESTABLISH") \
78  _(ESTABLISH_AO, "ESTABLISH_AO") \
79 
80 typedef enum _tcp_timers
81 {
82 #define _(sym, str) TCP_TIMER_##sym,
84 #undef _
86 } tcp_timers_e;
87 
88 typedef void (timer_expiration_handler) (u32 index);
89 
94 
95 #define TCP_TIMER_HANDLE_INVALID ((u32) ~0)
96 
97 /* Timer delays as multiples of 100ms */
98 #define TCP_TO_TIMER_TICK TCP_TICK*10 /* Period for converting from TCP
99  * ticks to timer units */
100 #define TCP_DELACK_TIME 1 /* 0.1s */
101 #define TCP_ESTABLISH_TIME 750 /* 75s */
102 #define TCP_SYN_RCVD_TIME 600 /* 60s */
103 #define TCP_2MSL_TIME 300 /* 30s */
104 #define TCP_CLOSEWAIT_TIME 20 /* 2s */
105 #define TCP_TIMEWAIT_TIME 100 /* 10s */
106 #define TCP_FINWAIT1_TIME 600 /* 60s */
107 #define TCP_CLEANUP_TIME 1 /* 0.1s */
108 #define TCP_TIMER_PERSIST_MIN 2 /* 0.2s */
109 
110 #define TCP_RTO_MAX 60 * THZ /* Min max RTO (60s) as per RFC6298 */
111 #define TCP_RTO_MIN 0.2 * THZ /* Min RTO (200ms) - lower than standard */
112 #define TCP_RTT_MAX 30 * THZ /* 30s (probably too much) */
113 #define TCP_RTO_SYN_RETRIES 3 /* SYN retries without doubling RTO */
114 #define TCP_RTO_INIT 1 * THZ /* Initial retransmit timer */
115 
116 /** TCP connection flags */
117 #define foreach_tcp_connection_flag \
118  _(SNDACK, "Send ACK") \
119  _(FINSNT, "FIN sent") \
120  _(RECOVERY, "Recovery") \
121  _(FAST_RECOVERY, "Fast Recovery") \
122  _(DCNT_PENDING, "Disconnect pending") \
123  _(HALF_OPEN_DONE, "Half-open completed") \
124  _(FINPNDG, "FIN pending") \
125  _(FRXT_PENDING, "Fast-retransmit pending") \
126  _(FRXT_FIRST, "Fast-retransmit first again") \
127  _(DEQ_PENDING, "Pending dequeue acked") \
128  _(PSH_PENDING, "PSH pending") \
129  _(FINRCVD, "FIN received") \
130 
131 typedef enum _tcp_connection_flag_bits
132 {
133 #define _(sym, str) TCP_CONN_##sym##_BIT,
135 #undef _
138 
139 typedef enum _tcp_connection_flag
140 {
141 #define _(sym, str) TCP_CONN_##sym = 1 << TCP_CONN_##sym##_BIT,
143 #undef _
146 
147 #define TCP_SCOREBOARD_TRACE (0)
148 #define TCP_MAX_SACK_BLOCKS 256 /**< Max number of SACK blocks stored */
149 #define TCP_INVALID_SACK_HOLE_INDEX ((u32)~0)
150 
151 typedef struct _scoreboard_trace_elt
152 {
153  u32 start;
154  u32 end;
155  u32 ack;
156  u32 snd_una_max;
157  u32 group;
159 
160 typedef struct _sack_scoreboard_hole
161 {
162  u32 next; /**< Index for next entry in linked list */
163  u32 prev; /**< Index for previous entry in linked list */
164  u32 start; /**< Start sequence number */
165  u32 end; /**< End sequence number */
166  u8 is_lost; /**< Mark hole as lost */
168 
169 typedef struct _sack_scoreboard
170 {
171  sack_scoreboard_hole_t *holes; /**< Pool of holes */
172  u32 head; /**< Index of first entry */
173  u32 tail; /**< Index of last entry */
174  u32 sacked_bytes; /**< Number of bytes sacked in sb */
175  u32 last_sacked_bytes; /**< Number of bytes last sacked */
176  u32 last_bytes_delivered; /**< Number of sack bytes delivered */
177  u32 snd_una_adv; /**< Bytes to add to snd_una */
178  u32 high_sacked; /**< Highest byte sacked (fack) */
179  u32 high_rxt; /**< Highest retransmitted sequence */
180  u32 rescue_rxt; /**< Rescue sequence number */
181  u32 lost_bytes; /**< Bytes lost as per RFC6675 */
182  u32 cur_rxt_hole; /**< Retransmitting from this hole */
183 
184 #if TCP_SCOREBOARD_TRACE
186 #endif
187 
189 
190 #if TCP_SCOREBOARD_TRACE
191 #define tcp_scoreboard_trace_add(_tc, _ack) \
192 { \
193  static u64 _group = 0; \
194  sack_scoreboard_t *_sb = &_tc->sack_sb; \
195  sack_block_t *_sack, *_sacks; \
196  scoreboard_trace_elt_t *_elt; \
197  int i; \
198  _group++; \
199  _sacks = _tc->rcv_opts.sacks; \
200  for (i = 0; i < vec_len (_sacks); i++) \
201  { \
202  _sack = &_sacks[i]; \
203  vec_add2 (_sb->trace, _elt, 1); \
204  _elt->start = _sack->start; \
205  _elt->end = _sack->end; \
206  _elt->ack = _elt->end == _ack ? _ack : 0; \
207  _elt->snd_una_max = _elt->end == _ack ? _tc->snd_una_max : 0; \
208  _elt->group = _group; \
209  } \
210 }
211 #else
212 #define tcp_scoreboard_trace_add(_tc, _ack)
213 #endif
214 
217  start, u8 have_sent_1_smss,
218  u8 * can_rescue,
219  u8 * snd_limited);
221  u32 index);
222 
224  sack_scoreboard_hole_t * hole);
226  sack_scoreboard_hole_t * hole);
231 u8 *format_tcp_scoreboard (u8 * s, va_list * args);
232 
233 typedef enum _tcp_cc_algorithm_type
234 {
238 
239 typedef struct _tcp_cc_algorithm tcp_cc_algorithm_t;
240 
241 typedef enum _tcp_cc_ack_t
242 {
246 } tcp_cc_ack_t;
247 
248 typedef struct _tcp_connection
249 {
250  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
251  transport_connection_t connection; /**< Common transport data. First! */
252 
253  u8 state; /**< TCP state as per tcp_state_t */
254  u16 flags; /**< Connection flags (see tcp_conn_flags_e) */
255  u32 timers[TCP_N_TIMERS]; /**< Timer handles into timer wheel */
256 
257  /* TODO RFC4898 */
258 
259  /** Send sequence variables RFC793 */
260  u32 snd_una; /**< oldest unacknowledged sequence number */
261  u32 snd_una_max; /**< newest unacknowledged sequence number + 1*/
262  u32 snd_wnd; /**< send window */
263  u32 snd_wl1; /**< seq number used for last snd.wnd update */
264  u32 snd_wl2; /**< ack number used for last snd.wnd update */
265  u32 snd_nxt; /**< next seq number to be sent */
266  u16 snd_mss; /**< Effective send max seg (data) size */
267 
268  /** Receive sequence variables RFC793 */
269  u32 rcv_nxt; /**< next sequence number expected */
270  u32 rcv_wnd; /**< receive window we expect */
271 
272  u32 rcv_las; /**< rcv_nxt at last ack sent/rcv_wnd update */
273  u32 iss; /**< initial sent sequence */
274  u32 irs; /**< initial remote sequence */
275 
276  /* Options */
277  u8 snd_opts_len; /**< Tx options len */
278  u8 rcv_wscale; /**< Window scale to advertise to peer */
279  u8 snd_wscale; /**< Window scale to use when sending */
280  u32 tsval_recent; /**< Last timestamp received */
281  u32 tsval_recent_age; /**< When last updated tstamp_recent*/
282  tcp_options_t snd_opts; /**< Tx options for connection */
283  tcp_options_t rcv_opts; /**< Rx options for connection */
284 
285  sack_block_t *snd_sacks; /**< Vector of SACKs to send. XXX Fixed size? */
286  u8 snd_sack_pos; /**< Position in vec of first block to send */
287  sack_block_t *snd_sacks_fl; /**< Vector for building new list */
288  sack_scoreboard_t sack_sb; /**< SACK "scoreboard" that tracks holes */
289 
290  u16 rcv_dupacks; /**< Number of DUPACKs received */
291  u8 pending_dupacks; /**< Number of DUPACKs to be sent */
292 
293  /* Congestion control */
294  u32 cwnd; /**< Congestion window */
295  u32 cwnd_acc_bytes; /**< Bytes accumulated for cwnd increment */
296  u32 ssthresh; /**< Slow-start threshold */
297  u32 prev_ssthresh; /**< ssthresh before congestion */
298  u32 prev_cwnd; /**< ssthresh before congestion */
299  u32 bytes_acked; /**< Bytes acknowledged by current segment */
300  u32 burst_acked; /**< Bytes acknowledged in current burst */
301  u32 snd_rxt_bytes; /**< Retransmitted bytes */
302  u32 snd_rxt_ts; /**< Timestamp when first packet is retransmitted */
303  u32 tsecr_last_ack; /**< Timestamp echoed to us in last healthy ACK */
304  u32 snd_congestion; /**< snd_una_max when congestion is detected */
305  tcp_cc_algorithm_t *cc_algo; /**< Congestion control algorithm */
306  u8 cc_data[TCP_CC_DATA_SZ]; /**< Congestion control algo private data */
307 
308  /* RTT and RTO */
309  u32 rto; /**< Retransmission timeout */
310  u32 rto_boff; /**< Index for RTO backoff */
311  u32 srtt; /**< Smoothed RTT */
312  u32 rttvar; /**< Smoothed mean RTT difference. Approximates variance */
313  u32 rtt_seq; /**< Sequence number for tracked ACK */
314  f64 rtt_ts; /**< Timestamp for tracked ACK */
315  f64 mrtt_us; /**< High precision mrtt from tracked acks */
316 
317  u16 mss; /**< Our max seg size that includes options */
318  u32 limited_transmit; /**< snd_nxt when limited transmit starts */
319  u32 last_fib_check; /**< Last time we checked fib route for peer */
320  u32 sw_if_index; /**< Interface for the connection */
321  u32 tx_fifo_size; /**< Tx fifo size. Used to constrain cwnd */
322 
323  u32 psh_seq; /**< Add psh header for seg that includes this */
325 
326 /* *INDENT-OFF* */
327 struct _tcp_cc_algorithm
328 {
329  const char *name;
330  uword (*unformat_cfg) (unformat_input_t * input);
331  void (*rcv_ack) (tcp_connection_t * tc);
332  void (*rcv_cong_ack) (tcp_connection_t * tc, tcp_cc_ack_t ack);
333  void (*congestion) (tcp_connection_t * tc);
334  void (*recovered) (tcp_connection_t * tc);
335  void (*init) (tcp_connection_t * tc);
336 };
337 /* *INDENT-ON* */
338 
339 #define tcp_fastrecovery_on(tc) (tc)->flags |= TCP_CONN_FAST_RECOVERY
340 #define tcp_fastrecovery_off(tc) (tc)->flags &= ~TCP_CONN_FAST_RECOVERY
341 #define tcp_recovery_on(tc) (tc)->flags |= TCP_CONN_RECOVERY
342 #define tcp_recovery_off(tc) (tc)->flags &= ~TCP_CONN_RECOVERY
343 #define tcp_in_fastrecovery(tc) ((tc)->flags & TCP_CONN_FAST_RECOVERY)
344 #define tcp_in_recovery(tc) ((tc)->flags & (TCP_CONN_RECOVERY))
345 #define tcp_in_slowstart(tc) (tc->cwnd < tc->ssthresh)
346 #define tcp_disconnect_pending(tc) ((tc)->flags & TCP_CONN_DCNT_PENDING)
347 #define tcp_disconnect_pending_on(tc) ((tc)->flags |= TCP_CONN_DCNT_PENDING)
348 #define tcp_disconnect_pending_off(tc) ((tc)->flags &= ~TCP_CONN_DCNT_PENDING)
349 #define tcp_fastrecovery_first(tc) ((tc)->flags & TCP_CONN_FRXT_FIRST)
350 #define tcp_fastrecovery_first_on(tc) ((tc)->flags |= TCP_CONN_FRXT_FIRST)
351 #define tcp_fastrecovery_first_off(tc) ((tc)->flags &= ~TCP_CONN_FRXT_FIRST)
352 
353 #define tcp_in_cong_recovery(tc) ((tc)->flags & \
354  (TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY))
355 
356 always_inline void
358 {
359  tc->flags &= ~(TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY);
361 }
362 
363 typedef enum _tcp_error
364 {
365 #define tcp_error(n,s) TCP_ERROR_##n,
366 #include <vnet/tcp/tcp_error.def>
367 #undef tcp_error
369 } tcp_error_t;
370 
371 typedef struct _tcp_lookup_dispatch
372 {
373  u8 next, error;
375 
376 typedef struct tcp_worker_ctx_
377 {
378  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
379  /** worker time */
381 
382  /** worker timer wheel */
383  tw_timer_wheel_16t_2w_512sl_t timer_wheel;
384 
385  /** tx buffer free list */
387 
388  /** tx frames for tcp 4/6 output nodes */
389  vlib_frame_t *tx_frames[2];
390 
391  /** tx frames for ip 4/6 lookup nodes */
392  vlib_frame_t *ip_lookup_tx_frames[2];
393 
394  /** vector of connections needing fast rxt */
396 
397  /** vector of connections now doing fast rxt */
399 
400  /** vector of connections that will do fast rxt */
402 
403  /** vector of pending ack dequeues */
405 
406  /** vector of pending acks */
408 
409  /** vector of pending disconnect notifications */
411 
412  /** convenience pointer to this thread's vlib main */
414 
415  CLIB_CACHE_LINE_ALIGN_MARK (cacheline1);
416 
417  /** cached 'on the wire' options for bursts */
418  u8 cached_opts[40];
419 
421 
422 typedef struct tcp_iss_seed_
423 {
427 
428 typedef struct _tcp_main
429 {
430  /* Per-worker thread tcp connection pools */
432 
433  /* Pool of listeners. */
434  tcp_connection_t *listener_pool;
435 
436  /** Dispatch table by state and flags */
437  tcp_lookup_dispatch_t dispatch_table[TCP_N_STATES][64];
438 
439  u8 log2_tstamp_clocks_per_tick;
440  f64 tstamp_ticks_per_clock;
441 
442  /** per-worker context */
443  tcp_worker_ctx_t *wrk_ctx;
444 
445  /* Pool of half-open connections on which we've sent a SYN */
446  tcp_connection_t *half_open_connections;
447  clib_spinlock_t half_open_lock;
448 
449  /* Congestion control algorithms registered */
450  tcp_cc_algorithm_t *cc_algos;
451 
452  /** vlib buffer size */
453  u32 bytes_per_buffer;
454 
455  /* Seed used to generate random iss */
456  tcp_iss_seed_t iss_seed;
457 
458  /** Hash table of cc algorithms by name */
459  uword *cc_algo_by_name;
460 
461  /*
462  * Configuration
463  */
464 
465  /* Flag that indicates if stack is on or off */
466  u8 is_enabled;
467 
468  /** Max rx fifo size for a session. It is used in to compute the
469  * rfc 7323 window scaling factor */
470  u32 max_rx_fifo;
471 
472  /** Default MTU to be used when establishing connections */
473  u16 default_mtu;
474 
475  /** Number of preallocated connections */
476  u32 preallocated_connections;
477  u32 preallocated_half_open_connections;
478 
479  /** Vectors of src addresses. Optional unless one needs > 63K active-opens */
480  ip4_address_t *ip4_src_addresses;
481  u32 last_v4_address_rotor;
482  u32 last_v6_address_rotor;
483  ip6_address_t *ip6_src_addresses;
484 
485  /** Enable tx pacing for new connections */
486  u8 tx_pacing;
487 
488  u8 punt_unknown4;
489  u8 punt_unknown6;
490 
491  /** fault-injection */
492  f64 buffer_fail_fraction;
493 
494  /** Default congestion control algorithm type */
495  tcp_cc_algorithm_type_e cc_algo;
496 
497 } tcp_main_t;
498 
499 extern tcp_main_t tcp_main;
512 
515 {
516  return &tcp_main;
517 }
518 
520 tcp_get_worker (u32 thread_index)
521 {
522  return &tcp_main.wrk_ctx[thread_index];
523 }
524 
527 {
528  ASSERT ((signed) b->current_data >= (signed) -VLIB_BUFFER_PRE_DATA_SIZE);
529  return (tcp_header_t *) (b->data + b->current_data
530  + vnet_buffer (b)->tcp.hdr_offset);
531 }
532 
533 #if (VLIB_BUFFER_TRACE_TRAJECTORY)
534 #define tcp_trajectory_add_start(b, start) \
535 { \
536  (*vlib_buffer_trace_trajectory_cb) (b, start); \
537 }
538 #else
539 #define tcp_trajectory_add_start(b, start)
540 #endif
541 
543 
544 void tcp_punt_unknown (vlib_main_t * vm, u8 is_ip4, u8 is_add);
545 
547 tcp_connection_get (u32 conn_index, u32 thread_index)
548 {
549  if (PREDICT_FALSE
550  (pool_is_free_index (tcp_main.connections[thread_index], conn_index)))
551  return 0;
552  return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
553 }
554 
556 tcp_connection_get_if_valid (u32 conn_index, u32 thread_index)
557 {
558  if (tcp_main.connections[thread_index] == 0)
559  return 0;
560  if (pool_is_free_index (tcp_main.connections[thread_index], conn_index))
561  return 0;
562  return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
563 }
564 
567 {
568  return (tcp_connection_t *) tconn;
569 }
570 
571 always_inline void
573 {
574  tc->state = state;
575  TCP_EVT_DBG (TCP_EVT_STATE_CHANGE, tc);
576 }
577 
586  ip4_address_t * start,
587  ip4_address_t * end, u32 table_id);
589  ip6_address_t * start,
590  ip6_address_t * end, u32 table_id);
591 void tcp_api_reference (void);
592 u8 *format_tcp_connection (u8 * s, va_list * args);
593 
596 {
597  return pool_elt_at_index (tcp_main.listener_pool, tli);
598 }
599 
602 {
603  tcp_connection_t *tc = 0;
604  clib_spinlock_lock_if_init (&tcp_main.half_open_lock);
605  if (!pool_is_free_index (tcp_main.half_open_connections, conn_index))
606  tc = pool_elt_at_index (tcp_main.half_open_connections, conn_index);
607  clib_spinlock_unlock_if_init (&tcp_main.half_open_lock);
608  return tc;
609 }
610 
614  u32 thread_index, u8 is_ip4);
615 void tcp_send_reset (tcp_connection_t * tc);
616 void tcp_send_syn (tcp_connection_t * tc);
618 void tcp_send_fin (tcp_connection_t * tc);
619 void tcp_init_mss (tcp_connection_t * tc);
621 void tcp_update_rto (tcp_connection_t * tc);
622 void tcp_flush_frame_to_output (tcp_worker_ctx_t * wrk, u8 is_ip4);
625  tcp_connection_t * tc);
627 
630 void tcp_send_acks (tcp_worker_ctx_t * wrk);
631 
634 {
635  return th->seq_number + tcp_is_syn (th) + tcp_is_fin (th) + len;
636 }
637 
638 /* Modulo arithmetic for TCP sequence numbers */
639 #define seq_lt(_s1, _s2) ((i32)((_s1)-(_s2)) < 0)
640 #define seq_leq(_s1, _s2) ((i32)((_s1)-(_s2)) <= 0)
641 #define seq_gt(_s1, _s2) ((i32)((_s1)-(_s2)) > 0)
642 #define seq_geq(_s1, _s2) ((i32)((_s1)-(_s2)) >= 0)
643 #define seq_max(_s1, _s2) (seq_gt((_s1), (_s2)) ? (_s1) : (_s2))
644 
645 /* Modulo arithmetic for timestamps */
646 #define timestamp_lt(_t1, _t2) ((i32)((_t1)-(_t2)) < 0)
647 #define timestamp_leq(_t1, _t2) ((i32)((_t1)-(_t2)) <= 0)
648 
649 /**
650  * Our estimate of the number of bytes that have left the network
651  */
654 {
655  if (tcp_opts_sack_permitted (&tc->rcv_opts))
656  return tc->sack_sb.sacked_bytes + tc->sack_sb.lost_bytes;
657  else
658  return tc->rcv_dupacks * tc->snd_mss;
659 }
660 
661 /**
662  * Our estimate of the number of bytes in flight (pipe size)
663  */
666 {
667  int flight_size;
668 
669  flight_size = (int) (tc->snd_nxt - tc->snd_una) - tcp_bytes_out (tc)
670  + tc->snd_rxt_bytes;
671 
672  if (flight_size < 0)
673  {
674  if (0)
676  ("Negative: %u %u %u dupacks %u sacked bytes %u flags %d",
677  tc->snd_una_max - tc->snd_una, tcp_bytes_out (tc),
678  tc->snd_rxt_bytes, tc->rcv_dupacks, tc->sack_sb.sacked_bytes,
679  tc->rcv_opts.flags);
680  return 0;
681  }
682 
683  return flight_size;
684 }
685 
686 /**
687  * Initial cwnd as per RFC5681
688  */
691 {
692  if (tc->snd_mss > 2190)
693  return 2 * tc->snd_mss;
694  else if (tc->snd_mss > 1095)
695  return 3 * tc->snd_mss;
696  else
697  return 4 * tc->snd_mss;
698 }
699 
700 /*
701  * Accumulate acked bytes for cwnd increase
702  *
703  * Once threshold bytes are accumulated, snd_mss bytes are added
704  * to the cwnd.
705  */
706 always_inline void
708 {
709  tc->cwnd_acc_bytes += bytes;
710  if (tc->cwnd_acc_bytes >= thresh)
711  {
712  u32 inc = tc->cwnd_acc_bytes / thresh;
713  tc->cwnd_acc_bytes -= inc * thresh;
714  tc->cwnd += inc * tc->snd_mss;
715  tc->cwnd = clib_min (tc->cwnd, tc->tx_fifo_size);
716  }
717 }
718 
721 {
722  return tc->snd_mss;
723 }
724 
727 {
728  return clib_min (tc->cwnd, tc->snd_wnd);
729 }
730 
733 {
734  u32 available_wnd = tcp_available_snd_wnd (tc);
735  int flight_size = (int) (tc->snd_nxt - tc->snd_una);
736 
737  if (available_wnd <= flight_size)
738  return 0;
739 
740  return available_wnd - flight_size;
741 }
742 
743 /**
744  * Estimate of how many bytes we can still push into the network
745  */
748 {
749  u32 available_wnd = tcp_available_snd_wnd (tc);
750  u32 flight_size = tcp_flight_size (tc);
751 
752  if (available_wnd <= flight_size)
753  return 0;
754 
755  return available_wnd - flight_size;
756 }
757 
760 {
761  if ((tc->flags & TCP_CONN_FINSNT) && tc->snd_una_max - tc->snd_una == 1)
762  return 1;
763  return 0;
764 }
765 
768  tcp_connection_t * tc);
770  tcp_connection_t * tc, u32 burst_size);
772  u32 burst_size);
774  u32 burst_size);
777 
779 
780 /* Made public for unit testing only */
781 void tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end);
783 
786 {
787  return tcp_main.wrk_ctx[vlib_get_thread_index ()].time_now;
788 }
789 
792 {
793  return tcp_main.wrk_ctx[thread_index].time_now;
794 }
795 
797 tcp_time_now_us (u32 thread_index)
798 {
799  return transport_time_now (thread_index);
800 }
801 
804 {
805  wrk->time_now = clib_cpu_time_now () * tcp_main.tstamp_ticks_per_clock;
806  return wrk->time_now;
807 }
808 
810  vlib_buffer_t * b);
811 
818  u32 start_bucket);
819 
820 always_inline void
822 {
823  tc->cc_algo->rcv_ack (tc);
824  tc->tsecr_last_ack = tc->rcv_opts.tsecr;
825 }
826 
827 always_inline void
828 tcp_timer_set (tcp_connection_t * tc, u8 timer_id, u32 interval)
829 {
830  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
831  ASSERT (tc->timers[timer_id] == TCP_TIMER_HANDLE_INVALID);
832  tc->timers[timer_id] =
833  tw_timer_start_16t_2w_512sl (&tcp_main.
834  wrk_ctx[tc->c_thread_index].timer_wheel,
835  tc->c_c_index, timer_id, interval);
836 }
837 
838 always_inline void
840 {
841  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
842  if (tc->timers[timer_id] == TCP_TIMER_HANDLE_INVALID)
843  return;
844 
845  tw_timer_stop_16t_2w_512sl (&tcp_main.
846  wrk_ctx[tc->c_thread_index].timer_wheel,
847  tc->timers[timer_id]);
848  tc->timers[timer_id] = TCP_TIMER_HANDLE_INVALID;
849 }
850 
851 always_inline void
852 tcp_timer_update (tcp_connection_t * tc, u8 timer_id, u32 interval)
853 {
854  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
855  if (tc->timers[timer_id] != TCP_TIMER_HANDLE_INVALID)
856  tw_timer_update_16t_2w_512sl (&tcp_main.
857  wrk_ctx[tc->c_thread_index].timer_wheel,
858  tc->timers[timer_id], interval);
859  else
860  tc->timers[timer_id] =
861  tw_timer_start_16t_2w_512sl (&tcp_main.
862  wrk_ctx[tc->c_thread_index].timer_wheel,
863  tc->c_c_index, timer_id, interval);
864 }
865 
866 always_inline void
868 {
869  ASSERT (tc->snd_una != tc->snd_una_max);
870  tcp_timer_set (tc, TCP_TIMER_RETRANSMIT,
871  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
872 }
873 
874 always_inline void
876 {
877  tcp_timer_reset (tc, TCP_TIMER_RETRANSMIT);
878 }
879 
880 always_inline void
882 {
883  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
884  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
885 }
886 
887 always_inline void
889 {
890  /* Reuse RTO. It's backed off in handler */
891  tcp_timer_set (tc, TCP_TIMER_PERSIST,
892  clib_max (tc->rto * TCP_TO_TIMER_TICK,
894 }
895 
896 always_inline void
898 {
899  tcp_timer_update (tc, TCP_TIMER_PERSIST,
900  clib_max (tc->rto * TCP_TO_TIMER_TICK,
902 }
903 
904 always_inline void
906 {
907  tcp_timer_reset (tc, TCP_TIMER_PERSIST);
908 }
909 
910 always_inline void
912 {
913  if (tc->snd_una == tc->snd_nxt)
914  {
916  if (tc->snd_wnd < tc->snd_mss)
918  }
919  else
920  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
921  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
922 }
923 
926 {
927  return tc->timers[timer] != TCP_TIMER_HANDLE_INVALID;
928 }
929 
930 #define tcp_validate_txf_size(_tc, _a) \
931  ASSERT(_tc->state != TCP_STATE_ESTABLISHED \
932  || transport_max_tx_dequeue (&_tc->connection) >= _a)
933 
934 void tcp_rcv_sacks (tcp_connection_t * tc, u32 ack);
935 u8 *tcp_scoreboard_replay (u8 * s, tcp_connection_t * tc, u8 verbose);
936 
938  const tcp_cc_algorithm_t * vft);
939 
941 
942 static inline void *
944 {
945  return (void *) tc->cc_data;
946 }
947 
949 
950 /**
951  * Push TCP header to buffer
952  *
953  * @param vm - vlib_main
954  * @param b - buffer to write the header to
955  * @param sp_net - source port net order
956  * @param dp_net - destination port net order
957  * @param seq - sequence number net order
958  * @param ack - ack number net order
959  * @param tcp_hdr_opts_len - header and options length in bytes
960  * @param flags - header flags
961  * @param wnd - window size
962  *
963  * @return - pointer to start of TCP header
964  */
965 always_inline void *
967  u32 ack, u8 tcp_hdr_opts_len, u8 flags,
968  u16 wnd)
969 {
970  tcp_header_t *th;
971 
972  th = vlib_buffer_push_uninit (b, tcp_hdr_opts_len);
973 
974  th->src_port = sp;
975  th->dst_port = dp;
976  th->seq_number = seq;
977  th->ack_number = ack;
978  th->data_offset_and_reserved = (tcp_hdr_opts_len >> 2) << 4;
979  th->flags = flags;
980  th->window = wnd;
981  th->checksum = 0;
982  th->urgent_pointer = 0;
983  return th;
984 }
985 
986 /**
987  * Push TCP header to buffer
988  *
989  * @param b - buffer to write the header to
990  * @param sp_net - source port net order
991  * @param dp_net - destination port net order
992  * @param seq - sequence number host order
993  * @param ack - ack number host order
994  * @param tcp_hdr_opts_len - header and options length in bytes
995  * @param flags - header flags
996  * @param wnd - window size
997  *
998  * @return - pointer to start of TCP header
999  */
1000 always_inline void *
1001 vlib_buffer_push_tcp (vlib_buffer_t * b, u16 sp_net, u16 dp_net, u32 seq,
1002  u32 ack, u8 tcp_hdr_opts_len, u8 flags, u16 wnd)
1003 {
1004  return vlib_buffer_push_tcp_net_order (b, sp_net, dp_net,
1005  clib_host_to_net_u32 (seq),
1006  clib_host_to_net_u32 (ack),
1007  tcp_hdr_opts_len, flags,
1008  clib_host_to_net_u16 (wnd));
1009 }
1010 
1011 #endif /* _vnet_tcp_h_ */
1012 
1013 /*
1014  * fd.io coding-style-patch-verification: ON
1015  *
1016  * Local Variables:
1017  * eval: (c-set-style "gnu")
1018  * End:
1019  */
u32 sw_if_index
Definition: ipsec_gre.api:37
tcp_main_t tcp_main
Definition: tcp.c:29
vlib_node_registration_t tcp6_listen_node
(constructor) VLIB_REGISTER_NODE (tcp6_listen_node)
Definition: tcp_input.c:3259
static vlib_cli_command_t trace
(constructor) VLIB_CLI_COMMAND (trace)
Definition: vlib_api_cli.c:862
#define CLIB_CACHE_LINE_ALIGN_MARK(mark)
Definition: cache.h:60
u32 flags
Definition: vhost_user.h:115
#define clib_min(x, y)
Definition: clib.h:295
u32 * pending_acks
vector of pending acks
Definition: tcp.h:407
u32 * pending_disconnects
vector of pending disconnect notifications
Definition: tcp.h:410
void scoreboard_clear(sack_scoreboard_t *sb)
Definition: tcp_input.c:900
static f64 tcp_time_now_us(u32 thread_index)
Definition: tcp.h:797
static void tcp_retransmit_timer_set(tcp_connection_t *tc)
Definition: tcp.h:867
void tcp_make_fin(tcp_connection_t *tc, vlib_buffer_t *b)
Convert buffer to FIN-ACK.
Definition: tcp_output.c:522
struct _sack_block sack_block_t
void tcp_cc_init_congestion(tcp_connection_t *tc)
Init loss recovery/fast recovery.
Definition: tcp_input.c:1161
struct _scoreboard_trace_elt scoreboard_trace_elt_t
void tcp_connection_timers_reset(tcp_connection_t *tc)
Stop all connection timers.
Definition: tcp.c:442
static f64 transport_time_now(u32 thread_index)
Definition: session.h:426
#define TCP_TO_TIMER_TICK
Definition: tcp.h:98
vlib_node_registration_t tcp4_output_node
(constructor) VLIB_REGISTER_NODE (tcp4_output_node)
Definition: tcp_output.c:2154
void scoreboard_init(sack_scoreboard_t *sb)
Definition: tcp_input.c:892
static u32 tcp_bytes_out(const tcp_connection_t *tc)
Our estimate of the number of bytes that have left the network.
Definition: tcp.h:653
i16 current_data
signed offset in data[], pre_data[] that we are currently processing.
Definition: buffer.h:110
unsigned long u64
Definition: types.h:89
static tcp_connection_t * tcp_connection_get_if_valid(u32 conn_index, u32 thread_index)
Definition: tcp.h:556
void tcp_connection_del(tcp_connection_t *tc)
Connection removal.
Definition: tcp.c:246
struct _sack_scoreboard sack_scoreboard_t
void tcp_update_sack_list(tcp_connection_t *tc, u32 start, u32 end)
Build SACK list as per RFC2018.
Definition: tcp_input.c:1736
u32 tcp_snd_space(tcp_connection_t *tc)
Definition: tcp.c:1141
static tcp_connection_t * tcp_half_open_connection_get(u32 conn_index)
Definition: tcp.h:601
sack_scoreboard_hole_t * scoreboard_last_hole(sack_scoreboard_t *sb)
Definition: tcp_input.c:690
sack_scoreboard_hole_t * scoreboard_next_hole(sack_scoreboard_t *sb, sack_scoreboard_hole_t *hole)
Definition: tcp_input.c:666
#define VLIB_BUFFER_PRE_DATA_SIZE
Definition: buffer.h:51
static_always_inline void clib_spinlock_unlock_if_init(clib_spinlock_t *p)
Definition: lock.h:98
struct _tcp_main tcp_main_t
vlib_node_registration_t tcp6_output_node
(constructor) VLIB_REGISTER_NODE (tcp6_output_node)
Definition: tcp_output.c:2174
u8 data[0]
Packet data.
Definition: buffer.h:181
static u64 clib_cpu_time_now(void)
Definition: time.h:75
timer_expiration_handler tcp_timer_retransmit_handler
u8 * format_tcp_scoreboard(u8 *s, va_list *args)
Definition: tcp.c:1021
u32 tcp_session_push_header(transport_connection_t *tconn, vlib_buffer_t *b)
Definition: tcp_output.c:1129
struct _tcp_lookup_dispatch tcp_lookup_dispatch_t
void tcp_connection_tx_pacer_update(tcp_connection_t *tc)
Definition: tcp.c:1213
void tcp_update_burst_snd_vars(tcp_connection_t *tc)
Update burst send vars.
Definition: tcp_output.c:408
struct _tcp_connection tcp_connection_t
static u32 tcp_available_cc_snd_space(const tcp_connection_t *tc)
Estimate of how many bytes we can still push into the network.
Definition: tcp.h:747
static u32 tcp_available_snd_wnd(const tcp_connection_t *tc)
Definition: tcp.h:726
static tcp_connection_t * tcp_get_connection_from_transport(transport_connection_t *tconn)
Definition: tcp.h:566
void tcp_connection_cleanup(tcp_connection_t *tc)
Cleans up connection state.
Definition: tcp.c:199
static void tcp_cc_rcv_ack(tcp_connection_t *tc)
Definition: tcp.h:821
format_function_t format_tcp_flags
Definition: tcp.h:65
u32 * ongoing_fast_rxt
vector of connections now doing fast rxt
Definition: tcp.h:398
struct _tcp_header tcp_header_t
void tcp_connection_reset(tcp_connection_t *tc)
Notify session that connection has been reset.
Definition: tcp.c:284
u32 * pending_deq_acked
vector of pending ack dequeues
Definition: tcp.h:404
unsigned char u8
Definition: types.h:56
struct _sack_scoreboard_hole sack_scoreboard_hole_t
double f64
Definition: types.h:142
u8 *() format_function_t(u8 *s, va_list *args)
Definition: format.h:48
vlib_node_registration_t tcp4_syn_sent_node
(constructor) VLIB_REGISTER_NODE (tcp4_syn_sent_node)
Definition: tcp_input.c:2617
u8 * tcp_scoreboard_replay(u8 *s, tcp_connection_t *tc, u8 verbose)
Definition: tcp.c:1965
#define tcp_is_fin(_th)
Definition: tcp_packet.h:90
static void tcp_connection_set_state(tcp_connection_t *tc, tcp_state_t state)
Definition: tcp.h:572
sack_scoreboard_hole_t * scoreboard_get_hole(sack_scoreboard_t *sb, u32 index)
Definition: tcp_input.c:658
#define always_inline
Definition: clib.h:98
static u32 tcp_available_output_snd_space(const tcp_connection_t *tc)
Definition: tcp.h:732
int tcp_retransmit_first_unacked(tcp_worker_ctx_t *wrk, tcp_connection_t *tc)
Retransmit first unacked segment.
Definition: tcp_output.c:1726
static tcp_header_t * tcp_buffer_hdr(vlib_buffer_t *b)
Definition: tcp.h:526
static timer_callback_t * timers
Definition: timer.c:61
tw_timer_wheel_16t_2w_512sl_t timer_wheel
worker timer wheel
Definition: tcp.h:383
enum _tcp_state tcp_state_t
vhost_vring_state_t state
Definition: vhost_user.h:120
timer_expiration_handler tcp_timer_retransmit_syn_handler
static u32 tcp_time_now(void)
Definition: tcp.h:785
unsigned int u32
Definition: types.h:88
static ct_connection_t * connections
sack_scoreboard_hole_t * scoreboard_next_rxt_hole(sack_scoreboard_t *sb, sack_scoreboard_hole_t *start, u8 have_sent_1_smss, u8 *can_rescue, u8 *snd_limited)
Figure out the next hole to retransmit.
Definition: tcp_input.c:822
void tcp_api_reference(void)
Definition: tcp_api.c:109
#define TCP_EVT_DBG(_evt, _args...)
Definition: tcp_debug.h:241
static void tcp_timer_set(tcp_connection_t *tc, u8 timer_id, u32 interval)
Definition: tcp.h:828
void tcp_do_fastretransmits(tcp_worker_ctx_t *wrk)
Definition: tcp_input.c:1311
struct tcp_worker_ctx_ tcp_worker_ctx_t
vlib_main_t * vm
convenience pointer to this thread&#39;s vlib main
Definition: tcp.h:413
#define pool_elt_at_index(p, i)
Returns pointer to element at given index.
Definition: pool.h:514
timer_expiration_handler tcp_timer_persist_handler
u32 * postponed_fast_rxt
vector of connections that will do fast rxt
Definition: tcp.h:401
vlib_node_registration_t tcp6_syn_sent_node
(constructor) VLIB_REGISTER_NODE (tcp6_syn_sent_node)
Definition: tcp_input.c:2636
u32 tcp_sack_list_bytes(tcp_connection_t *tc)
Definition: tcp_input.c:1785
static void * vlib_buffer_push_tcp_net_order(vlib_buffer_t *b, u16 sp, u16 dp, u32 seq, u32 ack, u8 tcp_hdr_opts_len, u8 flags, u16 wnd)
Push TCP header to buffer.
Definition: tcp.h:966
void tcp_rcv_sacks(tcp_connection_t *tc, u32 ack)
Definition: tcp_input.c:937
int tcp_fast_retransmit_sack(tcp_worker_ctx_t *wrk, tcp_connection_t *tc, u32 burst_size)
Do fast retransmit with SACKs.
Definition: tcp_output.c:1780
clib_error_t * vnet_tcp_enable_disable(vlib_main_t *vm, u8 is_en)
Definition: tcp.c:1498
void tcp_send_syn(tcp_connection_t *tc)
Send SYN.
Definition: tcp_output.c:919
static void * tcp_cc_data(tcp_connection_t *tc)
Definition: tcp.h:943
struct _unformat_input_t unformat_input_t
unsigned short u16
Definition: types.h:57
sack_scoreboard_hole_t * scoreboard_prev_hole(sack_scoreboard_t *sb, sack_scoreboard_hole_t *hole)
Definition: tcp_input.c:674
#define TCP_TIMER_HANDLE_INVALID
Definition: tcp.h:95
static u32 tcp_flight_size(const tcp_connection_t *tc)
Our estimate of the number of bytes in flight (pipe size)
Definition: tcp.h:665
#define PREDICT_FALSE(x)
Definition: clib.h:111
tcp_connection_t * tcp_connection_alloc(u8 thread_index)
Definition: tcp.c:253
void tcp_connection_tx_pacer_reset(tcp_connection_t *tc, u32 window, u32 start_bucket)
Definition: tcp.c:1229
void tcp_cc_algo_register(tcp_cc_algorithm_type_e type, const tcp_cc_algorithm_t *vft)
Definition: tcp.c:556
u8 name[64]
Definition: memclnt.api:152
enum _tcp_cc_ack_t tcp_cc_ack_t
static void tcp_cwnd_accumulate(tcp_connection_t *tc, u32 thresh, u32 bytes)
Definition: tcp.h:707
u8 len
Definition: ip_types.api:49
static void tcp_timer_reset(tcp_connection_t *tc, u8 timer_id)
Definition: tcp.h:839
vlib_node_registration_t tcp4_listen_node
(constructor) VLIB_REGISTER_NODE (tcp4_listen_node)
Definition: tcp_input.c:3240
enum _tcp_error tcp_error_t
static void * vlib_buffer_push_tcp(vlib_buffer_t *b, u16 sp_net, u16 dp_net, u32 seq, u32 ack, u8 tcp_hdr_opts_len, u8 flags, u16 wnd)
Push TCP header to buffer.
Definition: tcp.h:1001
int tcp_configure_v4_source_address_range(vlib_main_t *vm, ip4_address_t *start, ip4_address_t *end, u32 table_id)
Configure an ipv4 source address range.
Definition: tcp.c:1649
static_always_inline uword vlib_get_thread_index(void)
Definition: threads.h:212
void tcp_send_reset(tcp_connection_t *tc)
Build and set reset packet for connection.
Definition: tcp_output.c:841
void tcp_punt_unknown(vlib_main_t *vm, u8 is_ip4, u8 is_add)
Definition: tcp.c:1516
format_function_t format_tcp_state
Definition: tcp.h:64
u64 second
Definition: tcp.h:425
void tcp_program_ack(tcp_worker_ctx_t *wrk, tcp_connection_t *tc)
Definition: tcp_output.c:1171
#define clib_warning(format, args...)
Definition: error.h:59
enum _tcp_timers tcp_timers_e
struct _transport_connection transport_connection_t
int tcp_half_open_connection_cleanup(tcp_connection_t *tc)
Try to cleanup half-open connection.
Definition: tcp.c:170
u32 fib_node_index_t
A typedef of a node index.
Definition: fib_types.h:30
void tcp_connection_timers_init(tcp_connection_t *tc)
Initialize all connection timers as invalid.
Definition: tcp.c:425
void tcp_make_synack(tcp_connection_t *ts, vlib_buffer_t *b)
Convert buffer to SYN-ACK.
Definition: tcp_output.c:556
#define pool_is_free_index(P, I)
Use free bitmap to query whether given index is free.
Definition: pool.h:283
format_function_t format_tcp_rcv_sacks
Definition: tcp.h:67
vlib_node_registration_t tcp6_input_node
(constructor) VLIB_REGISTER_NODE (tcp6_input_node)
Definition: tcp_input.c:3598
fib_node_index_t tcp_lookup_rmt_in_fib(tcp_connection_t *tc)
void tcp_program_fastretransmit(tcp_worker_ctx_t *wrk, tcp_connection_t *tc)
Definition: tcp_input.c:1301
void tcp_send_synack(tcp_connection_t *tc)
Definition: tcp_output.c:955
static void tcp_timer_update(tcp_connection_t *tc, u8 timer_id, u32 interval)
Definition: tcp.h:852
#define ASSERT(truth)
enum _tcp_cc_algorithm_type tcp_cc_algorithm_type_e
u8 is_add
Definition: ipsec_gre.api:36
void tcp_flush_frames_to_output(tcp_worker_ctx_t *wrk)
Flush v4 and v6 tcp and ip-lookup tx frames for thread index.
Definition: tcp_output.c:1013
enum _tcp_connection_flag_bits tcp_connection_flag_bits_e
void tcp_connection_init_vars(tcp_connection_t *tc)
Initialize tcp connection variables.
Definition: tcp.c:631
vlib_node_registration_t tcp4_rcv_process_node
(constructor) VLIB_REGISTER_NODE (tcp4_rcv_process_node)
Definition: tcp_input.c:3037
u8 * format_tcp_connection(u8 *s, va_list *args)
Definition: tcp.c:906
static u32 tcp_end_seq(tcp_header_t *th, u32 len)
Definition: tcp.h:633
static void init(void)
Definition: client.c:115
vlib_node_registration_t tcp6_established_node
(constructor) VLIB_REGISTER_NODE (tcp6_established_node)
Definition: tcp_input.c:2251
#define tcp_fastrecovery_first_off(tc)
Definition: tcp.h:351
void tcp_flush_frame_to_output(tcp_worker_ctx_t *wrk, u8 is_ip4)
Flush tx frame populated by retransmits and timer pops.
Definition: tcp_output.c:982
struct _tcp_cc_algorithm tcp_cc_algorithm_t
Definition: tcp.h:239
#define clib_max(x, y)
Definition: clib.h:288
static u32 tcp_time_now_w_thread(u32 thread_index)
Definition: tcp.h:791
struct _vlib_node_registration vlib_node_registration_t
#define tcp_is_syn(_th)
Definition: tcp_packet.h:89
struct tcp_iss_seed_ tcp_iss_seed_t
u64 first
Definition: tcp.h:424
static void tcp_persist_timer_update(tcp_connection_t *tc)
Definition: tcp.h:897
void tcp_send_reset_w_pkt(tcp_connection_t *tc, vlib_buffer_t *pkt, u32 thread_index, u8 is_ip4)
Send reset without reusing existing buffer.
Definition: tcp_output.c:761
static void * vlib_buffer_push_uninit(vlib_buffer_t *b, u8 size)
Prepend uninitialized data to buffer.
Definition: buffer.h:335
vlib_node_registration_t tcp4_input_node
(constructor) VLIB_REGISTER_NODE (tcp4_input_node)
Definition: tcp_input.c:3578
tcp_cc_algorithm_t * tcp_cc_algo_get(tcp_cc_algorithm_type_e type)
Definition: tcp.c:567
static u32 tcp_initial_cwnd(const tcp_connection_t *tc)
Initial cwnd as per RFC5681.
Definition: tcp.h:690
vlib_node_registration_t tcp6_rcv_process_node
(constructor) VLIB_REGISTER_NODE (tcp6_rcv_process_node)
Definition: tcp_input.c:3056
#define TCP_CC_DATA_SZ
Definition: tcp.h:33
#define foreach_tcp_fsm_state
TCP FSM state definitions as per RFC793.
Definition: tcp.h:43
void tcp_send_fin(tcp_connection_t *tc)
Send FIN.
Definition: tcp_output.c:1025
#define foreach_tcp_connection_flag
TCP connection flags.
Definition: tcp.h:117
#define foreach_tcp_timer
TCP timers.
Definition: tcp.h:70
void() timer_expiration_handler(u32 index)
Definition: tcp.h:88
static u8 tcp_is_lost_fin(tcp_connection_t *tc)
Definition: tcp.h:759
sack_scoreboard_hole_t * scoreboard_first_hole(sack_scoreboard_t *sb)
Definition: tcp_input.c:682
static tcp_worker_ctx_t * tcp_get_worker(u32 thread_index)
Definition: tcp.h:520
static void tcp_retransmit_timer_update(tcp_connection_t *tc)
Definition: tcp.h:911
VLIB buffer representation.
Definition: buffer.h:102
u64 uword
Definition: types.h:112
u32 time_now
worker time
Definition: tcp.h:380
int tcp_fast_retransmit_no_sack(tcp_worker_ctx_t *wrk, tcp_connection_t *tc, u32 burst_size)
Fast retransmit without SACK info.
Definition: tcp_output.c:1883
int tcp_fast_retransmit(tcp_worker_ctx_t *wrk, tcp_connection_t *tc, u32 burst_size)
Do fast retransmit.
Definition: tcp_output.c:1947
void tcp_init_snd_vars(tcp_connection_t *tc)
Initialize connection send variables.
Definition: tcp.c:598
void tcp_connection_close(tcp_connection_t *tc)
Begin connection closing procedure.
Definition: tcp.c:340
static void tcp_retransmit_timer_force_update(tcp_connection_t *tc)
Definition: tcp.h:881
#define vnet_buffer(b)
Definition: buffer.h:369
static tcp_connection_t * tcp_connection_get(u32 conn_index, u32 thread_index)
Definition: tcp.h:547
void tcp_update_rto(tcp_connection_t *tc)
Definition: tcp_input.c:465
u32 * pending_fast_rxt
vector of connections needing fast rxt
Definition: tcp.h:395
void tcp_init_mss(tcp_connection_t *tc)
Definition: tcp_output.c:428
enum _tcp_connection_flag tcp_connection_flags_e
format_function_t format_tcp_sacks
Definition: tcp.h:66
#define TCP_TIMER_PERSIST_MIN
Definition: tcp.h:108
#define tcp_opts_sack_permitted(_to)
Definition: tcp_packet.h:160
void tcp_send_acks(tcp_worker_ctx_t *wrk)
Definition: tcp_output.c:1193
void newreno_rcv_cong_ack(tcp_connection_t *tc, tcp_cc_ack_t ack_type)
Definition: tcp_newreno.c:45
static u32 tcp_loss_wnd(const tcp_connection_t *tc)
Definition: tcp.h:720
int tcp_configure_v6_source_address_range(vlib_main_t *vm, ip6_address_t *start, ip6_address_t *end, u32 table_id)
Configure an ipv6 source address range.
Definition: tcp.c:1741
static void tcp_persist_timer_set(tcp_connection_t *tc)
Definition: tcp.h:888
static tcp_main_t * vnet_get_tcp_main()
Definition: tcp.h:514
static void tcp_cong_recovery_off(tcp_connection_t *tc)
Definition: tcp.h:357
timer_expiration_handler tcp_timer_delack_handler
void tcp_connection_free(tcp_connection_t *tc)
Definition: tcp.c:266
static_always_inline void clib_spinlock_lock_if_init(clib_spinlock_t *p)
Definition: lock.h:82
static void tcp_retransmit_timer_reset(tcp_connection_t *tc)
Definition: tcp.h:875
void tcp_program_dupack(tcp_worker_ctx_t *wrk, tcp_connection_t *tc)
Definition: tcp_output.c:1181
void tcp_cc_fastrecovery_exit(tcp_connection_t *tc)
Definition: tcp_input.c:1187
static u32 tcp_set_time_now(tcp_worker_ctx_t *wrk)
Definition: tcp.h:803
static u8 tcp_timer_is_active(tcp_connection_t *tc, tcp_timers_e timer)
Definition: tcp.h:925
static tcp_connection_t * tcp_listener_get(u32 tli)
Definition: tcp.h:595
u32 * tx_buffers
tx buffer free list
Definition: tcp.h:386
vlib_node_registration_t tcp4_established_node
(constructor) VLIB_REGISTER_NODE (tcp4_established_node)
Definition: tcp_input.c:2232
static void tcp_persist_timer_reset(tcp_connection_t *tc)
Definition: tcp.h:905