FD.io VPP  v19.01.2-3-gf61a1a8
Vector Packet Processing
tcp.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2016 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  _(SENT_RCV_WND0, "Sent 0 rcv_wnd") \
121  _(RECOVERY, "Recovery") \
122  _(FAST_RECOVERY, "Fast Recovery") \
123  _(DCNT_PENDING, "Disconnect pending") \
124  _(HALF_OPEN_DONE, "Half-open completed") \
125  _(FINPNDG, "FIN pending") \
126  _(FRXT_PENDING, "Fast-retransmit pending") \
127  _(FRXT_FIRST, "Fast-retransmit first again") \
128  _(DEQ_PENDING, "Pending dequeue acked") \
129  _(PSH_PENDING, "PSH pending") \
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 /** TCP buffer flags */
148 #define foreach_tcp_buf_flag \
149  _ (ACK) /**< Sending ACK. */ \
150  _ (DUPACK) /**< Sending DUPACK. */ \
151 
152 enum
153 {
154 #define _(f) TCP_BUF_BIT_##f,
156 #undef _
158 };
159 
160 enum
161 {
162 #define _(f) TCP_BUF_FLAG_##f = 1 << TCP_BUF_BIT_##f,
164 #undef _
165 };
166 
167 #define TCP_SCOREBOARD_TRACE (0)
168 #define TCP_MAX_SACK_BLOCKS 256 /**< Max number of SACK blocks stored */
169 #define TCP_INVALID_SACK_HOLE_INDEX ((u32)~0)
171 typedef struct _scoreboard_trace_elt
172 {
173  u32 start;
174  u32 end;
175  u32 ack;
176  u32 snd_una_max;
177  u32 group;
179 
180 typedef struct _sack_scoreboard_hole
181 {
182  u32 next; /**< Index for next entry in linked list */
183  u32 prev; /**< Index for previous entry in linked list */
184  u32 start; /**< Start sequence number */
185  u32 end; /**< End sequence number */
186  u8 is_lost; /**< Mark hole as lost */
188 
189 typedef struct _sack_scoreboard
190 {
191  sack_scoreboard_hole_t *holes; /**< Pool of holes */
192  u32 head; /**< Index of first entry */
193  u32 tail; /**< Index of last entry */
194  u32 sacked_bytes; /**< Number of bytes sacked in sb */
195  u32 last_sacked_bytes; /**< Number of bytes last sacked */
196  u32 last_bytes_delivered; /**< Number of sack bytes delivered */
197  u32 snd_una_adv; /**< Bytes to add to snd_una */
198  u32 high_sacked; /**< Highest byte sacked (fack) */
199  u32 high_rxt; /**< Highest retransmitted sequence */
200  u32 rescue_rxt; /**< Rescue sequence number */
201  u32 lost_bytes; /**< Bytes lost as per RFC6675 */
202  u32 cur_rxt_hole; /**< Retransmitting from this hole */
203 
204 #if TCP_SCOREBOARD_TRACE
206 #endif
207 
209 
210 #if TCP_SCOREBOARD_TRACE
211 #define tcp_scoreboard_trace_add(_tc, _ack) \
212 { \
213  static u64 _group = 0; \
214  sack_scoreboard_t *_sb = &_tc->sack_sb; \
215  sack_block_t *_sack, *_sacks; \
216  scoreboard_trace_elt_t *_elt; \
217  int i; \
218  _group++; \
219  _sacks = _tc->rcv_opts.sacks; \
220  for (i = 0; i < vec_len (_sacks); i++) \
221  { \
222  _sack = &_sacks[i]; \
223  vec_add2 (_sb->trace, _elt, 1); \
224  _elt->start = _sack->start; \
225  _elt->end = _sack->end; \
226  _elt->ack = _elt->end == _ack ? _ack : 0; \
227  _elt->snd_una_max = _elt->end == _ack ? _tc->snd_una_max : 0; \
228  _elt->group = _group; \
229  } \
230 }
231 #else
232 #define tcp_scoreboard_trace_add(_tc, _ack)
233 #endif
234 
237  start, u8 have_sent_1_smss,
238  u8 * can_rescue,
239  u8 * snd_limited);
241  u32 index);
242 
244  sack_scoreboard_hole_t * hole);
246  sack_scoreboard_hole_t * hole);
251 u8 *format_tcp_scoreboard (u8 * s, va_list * args);
252 
253 typedef enum _tcp_cc_algorithm_type
254 {
258 
259 typedef struct _tcp_cc_algorithm tcp_cc_algorithm_t;
261 typedef enum _tcp_cc_ack_t
262 {
263  TCP_CC_ACK,
267 
268 typedef struct _tcp_connection
269 {
270  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
271  transport_connection_t connection; /**< Common transport data. First! */
272 
273  u8 state; /**< TCP state as per tcp_state_t */
274  u16 flags; /**< Connection flags (see tcp_conn_flags_e) */
275  u32 timers[TCP_N_TIMERS]; /**< Timer handles into timer wheel */
276 
277  /* TODO RFC4898 */
278 
279  /** Send sequence variables RFC793 */
280  u32 snd_una; /**< oldest unacknowledged sequence number */
281  u32 snd_una_max; /**< newest unacknowledged sequence number + 1*/
282  u32 snd_wnd; /**< send window */
283  u32 snd_wl1; /**< seq number used for last snd.wnd update */
284  u32 snd_wl2; /**< ack number used for last snd.wnd update */
285  u32 snd_nxt; /**< next seq number to be sent */
286  u16 snd_mss; /**< Effective send max seg (data) size */
287 
288  /** Receive sequence variables RFC793 */
289  u32 rcv_nxt; /**< next sequence number expected */
290  u32 rcv_wnd; /**< receive window we expect */
291 
292  u32 rcv_las; /**< rcv_nxt at last ack sent/rcv_wnd update */
293  u32 iss; /**< initial sent sequence */
294  u32 irs; /**< initial remote sequence */
295 
296  /* Options */
297  u8 snd_opts_len; /**< Tx options len */
298  u8 rcv_wscale; /**< Window scale to advertise to peer */
299  u8 snd_wscale; /**< Window scale to use when sending */
300  u32 tsval_recent; /**< Last timestamp received */
301  u32 tsval_recent_age; /**< When last updated tstamp_recent*/
302  tcp_options_t snd_opts; /**< Tx options for connection */
303  tcp_options_t rcv_opts; /**< Rx options for connection */
304 
305  sack_block_t *snd_sacks; /**< Vector of SACKs to send. XXX Fixed size? */
306  sack_scoreboard_t sack_sb; /**< SACK "scoreboard" that tracks holes */
307 
308  u16 rcv_dupacks; /**< Number of DUPACKs received */
309  u8 pending_dupacks; /**< Number of DUPACKs to be sent */
310 
311  /* Congestion control */
312  u32 cwnd; /**< Congestion window */
313  u32 cwnd_acc_bytes; /**< Bytes accumulated for cwnd increment */
314  u32 ssthresh; /**< Slow-start threshold */
315  u32 prev_ssthresh; /**< ssthresh before congestion */
316  u32 prev_cwnd; /**< ssthresh before congestion */
317  u32 bytes_acked; /**< Bytes acknowledged by current segment */
318  u32 burst_acked; /**< Bytes acknowledged in current burst */
319  u32 snd_rxt_bytes; /**< Retransmitted bytes */
320  u32 snd_rxt_ts; /**< Timestamp when first packet is retransmitted */
321  u32 tsecr_last_ack; /**< Timestamp echoed to us in last healthy ACK */
322  u32 snd_congestion; /**< snd_una_max when congestion is detected */
323  tcp_cc_algorithm_t *cc_algo; /**< Congestion control algorithm */
324  u8 cc_data[TCP_CC_DATA_SZ]; /**< Congestion control algo private data */
325 
326  /* RTT and RTO */
327  u32 rto; /**< Retransmission timeout */
328  u32 rto_boff; /**< Index for RTO backoff */
329  u32 srtt; /**< Smoothed RTT */
330  u32 rttvar; /**< Smoothed mean RTT difference. Approximates variance */
331  u32 rtt_seq; /**< Sequence number for tracked ACK */
332  f64 rtt_ts; /**< Timestamp for tracked ACK */
333  f64 mrtt_us; /**< High precision mrtt from tracked acks */
334 
335  u16 mss; /**< Our max seg size that includes options */
336  u32 limited_transmit; /**< snd_nxt when limited transmit starts */
337  u32 last_fib_check; /**< Last time we checked fib route for peer */
338  u32 sw_if_index; /**< Interface for the connection */
339  u32 tx_fifo_size; /**< Tx fifo size. Used to constrain cwnd */
340 
341  u32 psh_seq; /**< Add psh header for seg that includes this */
343 
344 /* *INDENT-OFF* */
345 struct _tcp_cc_algorithm
346 {
347  const char *name;
348  uword (*unformat_cfg) (unformat_input_t * input);
349  void (*rcv_ack) (tcp_connection_t * tc);
350  void (*rcv_cong_ack) (tcp_connection_t * tc, tcp_cc_ack_t ack);
351  void (*congestion) (tcp_connection_t * tc);
352  void (*recovered) (tcp_connection_t * tc);
353  void (*init) (tcp_connection_t * tc);
354 };
355 /* *INDENT-ON* */
356 
357 #define tcp_fastrecovery_on(tc) (tc)->flags |= TCP_CONN_FAST_RECOVERY
358 #define tcp_fastrecovery_off(tc) (tc)->flags &= ~TCP_CONN_FAST_RECOVERY
359 #define tcp_recovery_on(tc) (tc)->flags |= TCP_CONN_RECOVERY
360 #define tcp_recovery_off(tc) (tc)->flags &= ~TCP_CONN_RECOVERY
361 #define tcp_in_fastrecovery(tc) ((tc)->flags & TCP_CONN_FAST_RECOVERY)
362 #define tcp_in_recovery(tc) ((tc)->flags & (TCP_CONN_RECOVERY))
363 #define tcp_in_slowstart(tc) (tc->cwnd < tc->ssthresh)
364 #define tcp_disconnect_pending(tc) ((tc)->flags & TCP_CONN_DCNT_PENDING)
365 #define tcp_disconnect_pending_on(tc) ((tc)->flags |= TCP_CONN_DCNT_PENDING)
366 #define tcp_disconnect_pending_off(tc) ((tc)->flags &= ~TCP_CONN_DCNT_PENDING)
367 #define tcp_fastrecovery_first(tc) ((tc)->flags & TCP_CONN_FRXT_FIRST)
368 #define tcp_fastrecovery_first_on(tc) ((tc)->flags |= TCP_CONN_FRXT_FIRST)
369 #define tcp_fastrecovery_first_off(tc) ((tc)->flags &= ~TCP_CONN_FRXT_FIRST)
371 #define tcp_in_cong_recovery(tc) ((tc)->flags & \
372  (TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY))
373 
374 always_inline void
376 {
377  tc->flags &= ~(TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY);
379 }
380 
381 typedef enum _tcp_error
382 {
383 #define tcp_error(n,s) TCP_ERROR_##n,
385 #undef tcp_error
386  TCP_N_ERROR,
388 
389 typedef struct _tcp_lookup_dispatch
390 {
391  u8 next, error;
393 
394 typedef struct tcp_worker_ctx_
395 {
396  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
397  /** worker time */
398  u32 time_now;
400  /** worker timer wheel */
401  tw_timer_wheel_16t_2w_512sl_t timer_wheel;
403  /** tx buffer free list */
404  u32 *tx_buffers;
406  /** tx frames for tcp 4/6 output nodes */
407  vlib_frame_t *tx_frames[2];
409  /** tx frames for ip 4/6 lookup nodes */
410  vlib_frame_t *ip_lookup_tx_frames[2];
412  /** vector of connections needing fast rxt */
413  u32 *pending_fast_rxt;
415  /** vector of connections now doing fast rxt */
416  u32 *ongoing_fast_rxt;
418  /** vector of connections that will do fast rxt */
419  u32 *postponed_fast_rxt;
421  /** vector of pending ack dequeues */
422  u32 *pending_deq_acked;
424  /** vector of pending acks */
425  u32 *pending_acks;
427  /** vector of pending disconnect notifications */
428  u32 *pending_disconnects;
430  /** convenience pointer to this thread's vlib main */
431  vlib_main_t *vm;
433  CLIB_CACHE_LINE_ALIGN_MARK (cacheline1);
434 
435  /** cached 'on the wire' options for bursts */
436  u8 cached_opts[40];
439 
440 typedef struct tcp_iss_seed_
441 {
442  u64 first;
443  u64 second;
445 
446 typedef struct _tcp_main
447 {
448  /* Per-worker thread tcp connection pools */
449  tcp_connection_t **connections;
450 
451  /* Pool of listeners. */
452  tcp_connection_t *listener_pool;
453 
454  /** Dispatch table by state and flags */
455  tcp_lookup_dispatch_t dispatch_table[TCP_N_STATES][64];
456 
457  u8 log2_tstamp_clocks_per_tick;
458  f64 tstamp_ticks_per_clock;
459 
460  /** per-worker context */
461  tcp_worker_ctx_t *wrk_ctx;
462 
463  /* Pool of half-open connections on which we've sent a SYN */
464  tcp_connection_t *half_open_connections;
465  clib_spinlock_t half_open_lock;
466 
467  /* Congestion control algorithms registered */
468  tcp_cc_algorithm_t *cc_algos;
469 
470  /** vlib buffer size */
471  u32 bytes_per_buffer;
472 
473  /* Seed used to generate random iss */
474  tcp_iss_seed_t iss_seed;
475 
476  /*
477  * Configuration
478  */
479 
480  /* Flag that indicates if stack is on or off */
481  u8 is_enabled;
482 
483  /** Max rx fifo size for a session. It is used in to compute the
484  * rfc 7323 window scaling factor */
485  u32 max_rx_fifo;
486 
487  /** Number of preallocated connections */
488  u32 preallocated_connections;
489  u32 preallocated_half_open_connections;
490 
491  /** Vectors of src addresses. Optional unless one needs > 63K active-opens */
492  ip4_address_t *ip4_src_addresses;
493  u32 last_v4_address_rotor;
494  u32 last_v6_address_rotor;
495  ip6_address_t *ip6_src_addresses;
496 
497  /** Enable tx pacing for new connections */
498  u8 tx_pacing;
499 
500  u8 punt_unknown4;
501  u8 punt_unknown6;
502 
503  /** fault-injection */
504  f64 buffer_fail_fraction;
505 
506  /** Default congestion control algorithm type */
507  tcp_cc_algorithm_type_e cc_algo;
508 
509 } tcp_main_t;
510 
511 extern tcp_main_t tcp_main;
516 
519 {
520  return &tcp_main;
521 }
522 
524 tcp_get_worker (u32 thread_index)
525 {
526  return &tcp_main.wrk_ctx[thread_index];
527 }
528 
531 {
532  ASSERT ((signed) b->current_data >= (signed) -VLIB_BUFFER_PRE_DATA_SIZE);
533  return (tcp_header_t *) (b->data + b->current_data
534  + vnet_buffer (b)->tcp.hdr_offset);
535 }
536 
537 #if (VLIB_BUFFER_TRACE_TRAJECTORY)
538 #define tcp_trajectory_add_start(b, start) \
539 { \
540  (*vlib_buffer_trace_trajectory_cb) (b, start); \
541 }
542 #else
543 #define tcp_trajectory_add_start(b, start)
544 #endif
545 
547 
548 void tcp_punt_unknown (vlib_main_t * vm, u8 is_ip4, u8 is_add);
549 
551 tcp_connection_get (u32 conn_index, u32 thread_index)
552 {
553  if (PREDICT_FALSE
554  (pool_is_free_index (tcp_main.connections[thread_index], conn_index)))
555  return 0;
556  return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
557 }
558 
560 tcp_connection_get_if_valid (u32 conn_index, u32 thread_index)
561 {
562  if (tcp_main.connections[thread_index] == 0)
563  return 0;
564  if (pool_is_free_index (tcp_main.connections[thread_index], conn_index))
565  return 0;
566  return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
567 }
568 
571 {
572  return (tcp_connection_t *) tconn;
573 }
574 
575 always_inline void
577 {
578  tc->state = state;
579  TCP_EVT_DBG (TCP_EVT_STATE_CHANGE, tc);
580 }
581 
590  ip4_address_t * start,
591  ip4_address_t * end, u32 table_id);
593  ip6_address_t * start,
594  ip6_address_t * end, u32 table_id);
595 void tcp_api_reference (void);
596 u8 *format_tcp_connection (u8 * s, va_list * args);
597 
599 tcp_listener_get (u32 tli)
600 {
601  return pool_elt_at_index (tcp_main.listener_pool, tli);
602 }
603 
606 {
607  tcp_connection_t *tc = 0;
608  clib_spinlock_lock_if_init (&tcp_main.half_open_lock);
609  if (!pool_is_free_index (tcp_main.half_open_connections, conn_index))
610  tc = pool_elt_at_index (tcp_main.half_open_connections, conn_index);
611  clib_spinlock_unlock_if_init (&tcp_main.half_open_lock);
612  return tc;
613 }
614 
619  u32 thread_index, u8 is_ip4);
620 void tcp_send_reset (tcp_connection_t * tc);
621 void tcp_send_syn (tcp_connection_t * tc);
623 void tcp_send_fin (tcp_connection_t * tc);
624 void tcp_init_mss (tcp_connection_t * tc);
626 void tcp_update_rto (tcp_connection_t * tc);
627 void tcp_flush_frame_to_output (tcp_worker_ctx_t * wrk, u8 is_ip4);
630  tcp_connection_t * tc);
632 
635 void tcp_send_acks (tcp_worker_ctx_t * wrk);
636 
639 {
640  return th->seq_number + tcp_is_syn (th) + tcp_is_fin (th) + len;
641 }
642 
643 /* Modulo arithmetic for TCP sequence numbers */
644 #define seq_lt(_s1, _s2) ((i32)((_s1)-(_s2)) < 0)
645 #define seq_leq(_s1, _s2) ((i32)((_s1)-(_s2)) <= 0)
646 #define seq_gt(_s1, _s2) ((i32)((_s1)-(_s2)) > 0)
647 #define seq_geq(_s1, _s2) ((i32)((_s1)-(_s2)) >= 0)
648 #define seq_max(_s1, _s2) (seq_gt((_s1), (_s2)) ? (_s1) : (_s2))
650 /* Modulo arithmetic for timestamps */
651 #define timestamp_lt(_t1, _t2) ((i32)((_t1)-(_t2)) < 0)
652 #define timestamp_leq(_t1, _t2) ((i32)((_t1)-(_t2)) <= 0)
654 /**
655  * Our estimate of the number of bytes that have left the network
656  */
658 tcp_bytes_out (const tcp_connection_t * tc)
659 {
660  if (tcp_opts_sack_permitted (&tc->rcv_opts))
661  return tc->sack_sb.sacked_bytes + tc->sack_sb.lost_bytes;
662  else
663  return tc->rcv_dupacks * tc->snd_mss;
664 }
665 
666 /**
667  * Our estimate of the number of bytes in flight (pipe size)
668  */
671 {
672  int flight_size;
673 
674  flight_size = (int) (tc->snd_una_max - tc->snd_una) - tcp_bytes_out (tc)
675  + tc->snd_rxt_bytes;
676 
677  if (flight_size < 0)
678  {
679  if (0)
681  ("Negative: %u %u %u dupacks %u sacked bytes %u flags %d",
682  tc->snd_una_max - tc->snd_una, tcp_bytes_out (tc),
683  tc->snd_rxt_bytes, tc->rcv_dupacks, tc->sack_sb.sacked_bytes,
684  tc->rcv_opts.flags);
685  return 0;
686  }
687 
688  return flight_size;
689 }
690 
691 /**
692  * Initial cwnd as per RFC5681
693  */
696 {
697  if (tc->snd_mss > 2190)
698  return 2 * tc->snd_mss;
699  else if (tc->snd_mss > 1095)
700  return 3 * tc->snd_mss;
701  else
702  return 4 * tc->snd_mss;
703 }
704 
705 /*
706  * Accumulate acked bytes for cwnd increase
707  *
708  * Once threshold bytes are accumulated, snd_mss bytes are added
709  * to the cwnd.
710  */
711 always_inline void
712 tcp_cwnd_accumulate (tcp_connection_t * tc, u32 thresh, u32 bytes)
713 {
714  tc->cwnd_acc_bytes += bytes;
715  if (tc->cwnd_acc_bytes >= thresh)
716  {
717  u32 inc = tc->cwnd_acc_bytes / thresh;
718  tc->cwnd_acc_bytes -= inc * thresh;
719  tc->cwnd += inc * tc->snd_mss;
720  tc->cwnd = clib_min (tc->cwnd, tc->tx_fifo_size);
721  }
722 }
723 
725 tcp_loss_wnd (const tcp_connection_t * tc)
726 {
727  return tc->snd_mss;
728 }
729 
732 {
733  return clib_min (tc->cwnd, tc->snd_wnd);
734 }
735 
738 {
739  u32 available_wnd = tcp_available_snd_wnd (tc);
740  int flight_size = (int) (tc->snd_nxt - tc->snd_una);
741 
742  if (available_wnd <= flight_size)
743  return 0;
744 
745  return available_wnd - flight_size;
746 }
747 
748 /**
749  * Estimate of how many bytes we can still push into the network
750  */
753 {
754  u32 available_wnd = tcp_available_snd_wnd (tc);
755  u32 flight_size = tcp_flight_size (tc);
756 
757  if (available_wnd <= flight_size)
758  return 0;
759 
760  return available_wnd - flight_size;
761 }
762 
765 {
766  if ((tc->flags & TCP_CONN_FINSNT) && tc->snd_una_max - tc->snd_una == 1)
767  return 1;
768  return 0;
769 }
770 
773  tcp_connection_t * tc);
775  tcp_connection_t * tc, u32 burst_size);
777  u32 burst_size);
779  u32 burst_size);
782 
784 
785 /* Made public for unit testing only */
786 void tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end);
788 
790 tcp_time_now (void)
791 {
792  return tcp_main.wrk_ctx[vlib_get_thread_index ()].time_now;
793 }
794 
796 tcp_time_now_w_thread (u32 thread_index)
797 {
798  return tcp_main.wrk_ctx[thread_index].time_now;
799 }
800 
802 tcp_time_now_us (u32 thread_index)
803 {
804  return transport_time_now (thread_index);
805 }
806 
809 {
810  wrk->time_now = clib_cpu_time_now () * tcp_main.tstamp_ticks_per_clock;
811  return wrk->time_now;
812 }
813 
815 
822  u32 start_bucket);
823 
824 always_inline void
826 {
827  tc->cc_algo->rcv_ack (tc);
828  tc->tsecr_last_ack = tc->rcv_opts.tsecr;
829 }
830 
831 always_inline void
832 tcp_timer_set (tcp_connection_t * tc, u8 timer_id, u32 interval)
833 {
834  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
835  ASSERT (tc->timers[timer_id] == TCP_TIMER_HANDLE_INVALID);
836  tc->timers[timer_id] =
837  tw_timer_start_16t_2w_512sl (&tcp_main.
838  wrk_ctx[tc->c_thread_index].timer_wheel,
839  tc->c_c_index, timer_id, interval);
840 }
841 
842 always_inline void
843 tcp_timer_reset (tcp_connection_t * tc, u8 timer_id)
844 {
845  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
846  if (tc->timers[timer_id] == TCP_TIMER_HANDLE_INVALID)
847  return;
848 
849  tw_timer_stop_16t_2w_512sl (&tcp_main.
850  wrk_ctx[tc->c_thread_index].timer_wheel,
851  tc->timers[timer_id]);
852  tc->timers[timer_id] = TCP_TIMER_HANDLE_INVALID;
853 }
854 
855 always_inline void
856 tcp_timer_update (tcp_connection_t * tc, u8 timer_id, u32 interval)
857 {
858  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
859  if (tc->timers[timer_id] != TCP_TIMER_HANDLE_INVALID)
860  tw_timer_update_16t_2w_512sl (&tcp_main.
861  wrk_ctx[tc->c_thread_index].timer_wheel,
862  tc->timers[timer_id], interval);
863  else
864  tc->timers[timer_id] =
865  tw_timer_start_16t_2w_512sl (&tcp_main.
866  wrk_ctx[tc->c_thread_index].timer_wheel,
867  tc->c_c_index, timer_id, interval);
868 }
869 
870 always_inline void
872 {
873  ASSERT (tc->snd_una != tc->snd_una_max);
874  tcp_timer_set (tc, TCP_TIMER_RETRANSMIT,
875  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
876 }
877 
878 always_inline void
880 {
881  tcp_timer_reset (tc, TCP_TIMER_RETRANSMIT);
882 }
883 
884 always_inline void
886 {
887  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
888  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
889 }
890 
891 always_inline void
893 {
894  /* Reuse RTO. It's backed off in handler */
895  tcp_timer_set (tc, TCP_TIMER_PERSIST,
896  clib_max (tc->rto * TCP_TO_TIMER_TICK,
898 }
899 
900 always_inline void
902 {
903  tcp_timer_update (tc, TCP_TIMER_PERSIST,
904  clib_max (tc->rto * TCP_TO_TIMER_TICK,
906 }
907 
908 always_inline void
910 {
911  tcp_timer_reset (tc, TCP_TIMER_PERSIST);
912 }
913 
914 always_inline void
916 {
917  if (tc->snd_una == tc->snd_una_max)
918  {
920  if (tc->snd_wnd < tc->snd_mss)
922  }
923  else
924  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
925  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
926 }
927 
930 {
931  return tc->timers[timer] != TCP_TIMER_HANDLE_INVALID;
932 }
933 
934 #define tcp_validate_txf_size(_tc, _a) \
935  ASSERT(_tc->state != TCP_STATE_ESTABLISHED \
936  || session_tx_fifo_max_dequeue (&_tc->connection) >= _a)
937 
938 void tcp_rcv_sacks (tcp_connection_t * tc, u32 ack);
939 u8 *tcp_scoreboard_replay (u8 * s, tcp_connection_t * tc, u8 verbose);
940 
942  const tcp_cc_algorithm_t * vft);
943 
945 
946 static inline void *
948 {
949  return (void *) tc->cc_data;
950 }
951 
953 
954 /**
955  * Push TCP header to buffer
956  *
957  * @param vm - vlib_main
958  * @param b - buffer to write the header to
959  * @param sp_net - source port net order
960  * @param dp_net - destination port net order
961  * @param seq - sequence number net order
962  * @param ack - ack number net order
963  * @param tcp_hdr_opts_len - header and options length in bytes
964  * @param flags - header flags
965  * @param wnd - window size
966  *
967  * @return - pointer to start of TCP header
968  */
969 always_inline void *
971  u32 ack, u8 tcp_hdr_opts_len, u8 flags,
972  u16 wnd)
973 {
974  tcp_header_t *th;
975 
976  th = vlib_buffer_push_uninit (b, tcp_hdr_opts_len);
977 
978  th->src_port = sp;
979  th->dst_port = dp;
980  th->seq_number = seq;
981  th->ack_number = ack;
982  th->data_offset_and_reserved = (tcp_hdr_opts_len >> 2) << 4;
983  th->flags = flags;
984  th->window = wnd;
985  th->checksum = 0;
986  th->urgent_pointer = 0;
987  return th;
988 }
989 
990 /**
991  * Push TCP header to buffer
992  *
993  * @param b - buffer to write the header to
994  * @param sp_net - source port net order
995  * @param dp_net - destination port net order
996  * @param seq - sequence number host order
997  * @param ack - ack number host order
998  * @param tcp_hdr_opts_len - header and options length in bytes
999  * @param flags - header flags
1000  * @param wnd - window size
1001  *
1002  * @return - pointer to start of TCP header
1003  */
1004 always_inline void *
1005 vlib_buffer_push_tcp (vlib_buffer_t * b, u16 sp_net, u16 dp_net, u32 seq,
1006  u32 ack, u8 tcp_hdr_opts_len, u8 flags, u16 wnd)
1007 {
1008  return vlib_buffer_push_tcp_net_order (b, sp_net, dp_net,
1009  clib_host_to_net_u32 (seq),
1010  clib_host_to_net_u32 (ack),
1011  tcp_hdr_opts_len, flags,
1012  clib_host_to_net_u16 (wnd));
1013 }
1014 
1015 #endif /* _vnet_tcp_h_ */
1016 
1017 /*
1018  * fd.io coding-style-patch-verification: ON
1019  *
1020  * Local Variables:
1021  * eval: (c-set-style "gnu")
1022  * End:
1023  */
tcp_main_t tcp_main
Definition: tcp.c:29
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
void scoreboard_clear(sack_scoreboard_t *sb)
Definition: tcp_input.c:853
static f64 tcp_time_now_us(u32 thread_index)
Definition: tcp.h:803
static void tcp_retransmit_timer_set(tcp_connection_t *tc)
Definition: tcp.h:872
void tcp_make_fin(tcp_connection_t *tc, vlib_buffer_t *b)
Convert buffer to FIN-ACK.
Definition: tcp_output.c:575
struct _sack_block sack_block_t
void tcp_cc_init_congestion(tcp_connection_t *tc)
Init loss recovery/fast recovery.
Definition: tcp_input.c:1111
struct _scoreboard_trace_elt scoreboard_trace_elt_t
void tcp_connection_timers_reset(tcp_connection_t *tc)
Stop all connection timers.
Definition: tcp.c:426
static f64 transport_time_now(u32 thread_index)
Definition: session.h:557
struct _transport_connection transport_connection_t
#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:2226
void scoreboard_init(sack_scoreboard_t *sb)
Definition: tcp_input.c:845
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:659
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:561
void tcp_connection_del(tcp_connection_t *tc)
Connection removal.
Definition: tcp.c:243
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:1695
u32 tcp_snd_space(tcp_connection_t *tc)
Definition: tcp.c:1118
static tcp_connection_t * tcp_half_open_connection_get(u32 conn_index)
Definition: tcp.h:606
sack_scoreboard_hole_t * scoreboard_last_hole(sack_scoreboard_t *sb)
Definition: tcp_input.c:647
sack_scoreboard_hole_t * scoreboard_next_hole(sack_scoreboard_t *sb, sack_scoreboard_hole_t *hole)
Definition: tcp_input.c:623
#define VLIB_BUFFER_PRE_DATA_SIZE
Definition: buffer.h:52
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:2249
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:998
struct _tcp_lookup_dispatch tcp_lookup_dispatch_t
void tcp_connection_tx_pacer_update(tcp_connection_t *tc)
Definition: tcp.c:1197
void tcp_update_burst_snd_vars(tcp_connection_t *tc)
Update burst send vars.
Definition: tcp_output.c:424
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:753
static u32 tcp_available_snd_wnd(const tcp_connection_t *tc)
Definition: tcp.h:732
static tcp_connection_t * tcp_get_connection_from_transport(transport_connection_t *tconn)
Definition: tcp.h:571
void tcp_connection_cleanup(tcp_connection_t *tc)
Cleans up connection state.
Definition: tcp.c:199
u8 *( format_function_t)(u8 *s, va_list *args)
Definition: format.h:48
static void tcp_cc_rcv_ack(tcp_connection_t *tc)
Definition: tcp.h:826
format_function_t format_tcp_flags
Definition: tcp.h:65
struct _tcp_header tcp_header_t
void tcp_connection_reset(tcp_connection_t *tc)
Notify session that connection has been reset.
Definition: tcp.c:277
unsigned char u8
Definition: types.h:56
struct _sack_scoreboard_hole sack_scoreboard_hole_t
double f64
Definition: types.h:142
u8 * tcp_scoreboard_replay(u8 *s, tcp_connection_t *tc, u8 verbose)
Definition: tcp.c:1941
i16 current_data
signed offset in data[], pre_data[] that we are currently processing.
Definition: buffer.h:110
#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:577
sack_scoreboard_hole_t * scoreboard_get_hole(sack_scoreboard_t *sb, u32 index)
Definition: tcp_input.c:615
u32 sw_if_index
Definition: vxlan_gbp.api:37
#define always_inline
Definition: clib.h:98
static u32 tcp_available_output_snd_space(const tcp_connection_t *tc)
Definition: tcp.h:738
int tcp_retransmit_first_unacked(tcp_worker_ctx_t *wrk, tcp_connection_t *tc)
Retransmit first unacked segment.
Definition: tcp_output.c:1784
static tcp_header_t * tcp_buffer_hdr(vlib_buffer_t *b)
Definition: tcp.h:531
static timer_callback_t * timers
Definition: timer.c:61
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:791
unsigned int u32
Definition: types.h:88
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:777
void tcp_api_reference(void)
Definition: tcp_api.c:109
#define TCP_EVT_DBG(_evt, _args...)
Definition: tcp_debug.h:238
static void tcp_timer_set(tcp_connection_t *tc, u8 timer_id, u32 interval)
Definition: tcp.h:833
void tcp_do_fastretransmits(tcp_worker_ctx_t *wrk)
Definition: tcp_input.c:1255
static heap_elt_t * first(heap_header_t *h)
Definition: heap.c:59
struct tcp_worker_ctx_ tcp_worker_ctx_t
#define pool_elt_at_index(p, i)
Returns pointer to element at given index.
Definition: pool.h:511
timer_expiration_handler tcp_timer_persist_handler
u32 tcp_push_header(tcp_connection_t *tconn, vlib_buffer_t *b)
Definition: tcp_output.c:1216
u32 tcp_sack_list_bytes(tcp_connection_t *tc)
Definition: tcp_input.c:1749
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:971
void tcp_rcv_sacks(tcp_connection_t *tc, u32 ack)
Definition: tcp_input.c:888
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:1840
clib_error_t * vnet_tcp_enable_disable(vlib_main_t *vm, u8 is_en)
Definition: tcp.c:1483
void tcp_send_syn(tcp_connection_t *tc)
Send SYN.
Definition: tcp_output.c:987
static void * tcp_cc_data(tcp_connection_t *tc)
Definition: tcp.h:948
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:631
#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:671
#define PREDICT_FALSE(x)
Definition: clib.h:111
tcp_connection_t * tcp_connection_alloc(u8 thread_index)
Definition: tcp.c:251
void tcp_connection_tx_pacer_reset(tcp_connection_t *tc, u32 window, u32 start_bucket)
Definition: tcp.c:1213
void tcp_cc_algo_register(tcp_cc_algorithm_type_e type, const tcp_cc_algorithm_t *vft)
Definition: tcp.c:540
void( timer_expiration_handler)(u32 index)
Definition: tcp.h:88
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:713
u8 len
Definition: ip_types.api:49
static void tcp_timer_reset(tcp_connection_t *tc, u8 timer_id)
Definition: tcp.h:844
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:1006
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:1625
static_always_inline uword vlib_get_thread_index(void)
Definition: threads.h:212
vlib_main_t * vm
Definition: buffer.c:301
void tcp_send_reset(tcp_connection_t *tc)
Build and set reset packet for connection.
Definition: tcp_output.c:909
void tcp_punt_unknown(vlib_main_t *vm, u8 is_ip4, u8 is_add)
Definition: tcp.c:1501
format_function_t format_tcp_state
Definition: tcp.h:64
void tcp_program_ack(tcp_worker_ctx_t *wrk, tcp_connection_t *tc)
Definition: tcp_output.c:1258
#define clib_warning(format, args...)
Definition: error.h:59
enum _tcp_timers tcp_timers_e
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:409
void tcp_make_synack(tcp_connection_t *ts, vlib_buffer_t *b)
Convert buffer to SYN-ACK.
Definition: tcp_output.c:618
#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:3569
fib_node_index_t tcp_lookup_rmt_in_fib(tcp_connection_t *tc)
void tcp_make_ack(tcp_connection_t *ts, vlib_buffer_t *b)
Convert buffer to ACK.
Definition: tcp_output.c:561
void tcp_program_fastretransmit(tcp_worker_ctx_t *wrk, tcp_connection_t *tc)
Definition: tcp_input.c:1245
void tcp_send_synack(tcp_connection_t *tc)
Definition: tcp_output.c:1020
static void tcp_timer_update(tcp_connection_t *tc, u8 timer_id, u32 interval)
Definition: tcp.h:857
#define ASSERT(truth)
enum _tcp_cc_algorithm_type tcp_cc_algorithm_type_e
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:1072
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:614
u8 * format_tcp_connection(u8 *s, va_list *args)
Definition: tcp.c:887
static u32 tcp_end_seq(tcp_header_t *th, u32 len)
Definition: tcp.h:639
static void init(void)
Definition: client.c:115
#define tcp_fastrecovery_first_off(tc)
Definition: tcp.h:370
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:1041
struct _tcp_cc_algorithm tcp_cc_algorithm_t
Definition: tcp.h:260
#define clib_max(x, y)
Definition: clib.h:288
static u32 tcp_time_now_w_thread(u32 thread_index)
Definition: tcp.h:797
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
static void tcp_persist_timer_update(tcp_connection_t *tc)
Definition: tcp.h:902
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:829
static void * vlib_buffer_push_uninit(vlib_buffer_t *b, u8 size)
Prepend uninitialized data to buffer.
Definition: buffer.h:320
vlib_node_registration_t tcp4_input_node
(constructor) VLIB_REGISTER_NODE (tcp4_input_node)
Definition: tcp_input.c:3546
tcp_cc_algorithm_t * tcp_cc_algo_get(tcp_cc_algorithm_type_e type)
Definition: tcp.c:550
static u32 tcp_initial_cwnd(const tcp_connection_t *tc)
Initial cwnd as per RFC5681.
Definition: tcp.h:696
#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:1084
#define foreach_tcp_connection_flag
TCP connection flags.
Definition: tcp.h:117
#define foreach_tcp_timer
TCP timers.
Definition: tcp.h:70
static u8 tcp_is_lost_fin(tcp_connection_t *tc)
Definition: tcp.h:765
sack_scoreboard_hole_t * scoreboard_first_hole(sack_scoreboard_t *sb)
Definition: tcp_input.c:639
static tcp_worker_ctx_t * tcp_get_worker(u32 thread_index)
Definition: tcp.h:525
static void tcp_retransmit_timer_update(tcp_connection_t *tc)
Definition: tcp.h:916
u64 uword
Definition: types.h:112
u32 time_now
worker time
Definition: tcp.h:399
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:1949
int tcp_fast_retransmit(tcp_worker_ctx_t *wrk, tcp_connection_t *tc, u32 burst_size)
Do fast retransmit.
Definition: tcp_output.c:2018
void tcp_init_snd_vars(tcp_connection_t *tc)
Initialize connection send variables.
Definition: tcp.c:581
void tcp_connection_close(tcp_connection_t *tc)
Begin connection closing procedure.
Definition: tcp.c:333
static void tcp_retransmit_timer_force_update(tcp_connection_t *tc)
Definition: tcp.h:886
#define vnet_buffer(b)
Definition: buffer.h:368
static tcp_connection_t * tcp_connection_get(u32 conn_index, u32 thread_index)
Definition: tcp.h:552
void tcp_update_rto(tcp_connection_t *tc)
Definition: tcp_input.c:428
void tcp_init_mss(tcp_connection_t *tc)
Definition: tcp_output.c:444
u8 data[0]
Packet data.
Definition: buffer.h:176
#define foreach_tcp_buf_flag
TCP buffer flags.
Definition: tcp.h:148
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:1280
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:726
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:1717
static void tcp_persist_timer_set(tcp_connection_t *tc)
Definition: tcp.h:893
static tcp_main_t * vnet_get_tcp_main()
Definition: tcp.h:519
static void tcp_cong_recovery_off(tcp_connection_t *tc)
Definition: tcp.h:376
timer_expiration_handler tcp_timer_delack_handler
void tcp_connection_free(tcp_connection_t *tc)
Definition: tcp.c:264
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:880
void tcp_program_dupack(tcp_worker_ctx_t *wrk, tcp_connection_t *tc)
Definition: tcp_output.c:1268
void tcp_cc_fastrecovery_exit(tcp_connection_t *tc)
Definition: tcp_input.c:1136
static u32 tcp_set_time_now(tcp_worker_ctx_t *wrk)
Definition: tcp.h:809
static u8 tcp_timer_is_active(tcp_connection_t *tc, tcp_timers_e timer)
Definition: tcp.h:930
static tcp_connection_t * tcp_listener_get(u32 tli)
Definition: tcp.h:600
static void tcp_persist_timer_reset(tcp_connection_t *tc)
Definition: tcp.h:910