FD.io VPP  v17.07-30-g839fa73
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 
34 #define TCP_DUPACK_THRESHOLD 3
35 #define TCP_MAX_RX_FIFO_SIZE 4 << 20
36 #define TCP_IW_N_SEGMENTS 10
37 #define TCP_ALWAYS_ACK 1 /**< On/off delayed acks */
38 #define TCP_USE_SACKS 1 /**< Disable only for testing */
39 
40 /** TCP FSM state definitions as per RFC793. */
41 #define foreach_tcp_fsm_state \
42  _(CLOSED, "CLOSED") \
43  _(LISTEN, "LISTEN") \
44  _(SYN_SENT, "SYN_SENT") \
45  _(SYN_RCVD, "SYN_RCVD") \
46  _(ESTABLISHED, "ESTABLISHED") \
47  _(CLOSE_WAIT, "CLOSE_WAIT") \
48  _(FIN_WAIT_1, "FIN_WAIT_1") \
49  _(LAST_ACK, "LAST_ACK") \
50  _(CLOSING, "CLOSING") \
51  _(FIN_WAIT_2, "FIN_WAIT_2") \
52  _(TIME_WAIT, "TIME_WAIT")
53 
54 typedef enum _tcp_state
55 {
56 #define _(sym, str) TCP_STATE_##sym,
58 #undef _
60 } tcp_state_t;
61 
65 
66 /** TCP timers */
67 #define foreach_tcp_timer \
68  _(RETRANSMIT, "RETRANSMIT") \
69  _(DELACK, "DELAYED ACK") \
70  _(PERSIST, "PERSIST") \
71  _(KEEP, "KEEP") \
72  _(WAITCLOSE, "WAIT CLOSE") \
73  _(RETRANSMIT_SYN, "RETRANSMIT SYN") \
74  _(ESTABLISH, "ESTABLISH")
75 
76 typedef enum _tcp_timers
77 {
78 #define _(sym, str) TCP_TIMER_##sym,
80 #undef _
82 } tcp_timers_e;
83 
84 typedef void (timer_expiration_handler) (u32 index);
85 
90 
91 #define TCP_TIMER_HANDLE_INVALID ((u32) ~0)
92 
93 /* Timer delays as multiples of 100ms */
94 #define TCP_TO_TIMER_TICK TCP_TICK*10 /* Period for converting from TCP
95  * ticks to timer units */
96 #define TCP_DELACK_TIME 1 /* 0.1s */
97 #define TCP_ESTABLISH_TIME 750 /* 75s */
98 #define TCP_2MSL_TIME 300 /* 30s */
99 #define TCP_CLOSEWAIT_TIME 20 /* 0.1s */
100 #define TCP_CLEANUP_TIME 5 /* 0.5s Time to wait before cleanup */
101 #define TCP_TIMER_PERSIST_MIN 2 /* 0.2s */
102 
103 #define TCP_RTO_MAX 60 * THZ /* Min max RTO (60s) as per RFC6298 */
104 #define TCP_RTO_MIN 0.2 * THZ /* Min RTO (200ms) - lower than standard */
105 #define TCP_RTT_MAX 30 * THZ /* 30s (probably too much) */
106 #define TCP_RTO_SYN_RETRIES 3 /* SYN retries without doubling RTO */
107 #define TCP_RTO_INIT 1 * THZ /* Initial retransmit timer */
108 
109 /** TCP connection flags */
110 #define foreach_tcp_connection_flag \
111  _(SNDACK, "Send ACK") \
112  _(FINSNT, "FIN sent") \
113  _(SENT_RCV_WND0, "Sent 0 receive window") \
114  _(RECOVERY, "Recovery on") \
115  _(FAST_RECOVERY, "Fast Recovery on") \
116  _(FR_1_SMSS, "Sent 1 SMSS")
117 
118 typedef enum _tcp_connection_flag_bits
119 {
120 #define _(sym, str) TCP_CONN_##sym##_BIT,
122 #undef _
125 
126 typedef enum _tcp_connection_flag
127 {
128 #define _(sym, str) TCP_CONN_##sym = 1 << TCP_CONN_##sym##_BIT,
130 #undef _
133 
134 /** TCP buffer flags */
135 #define foreach_tcp_buf_flag \
136  _ (ACK) /**< Sending ACK. */ \
137  _ (DUPACK) /**< Sending DUPACK. */ \
138 
139 enum
140 {
141 #define _(f) TCP_BUF_BIT_##f,
143 #undef _
145 };
146 
147 enum
148 {
149 #define _(f) TCP_BUF_FLAG_##f = 1 << TCP_BUF_BIT_##f,
151 #undef _
152 };
153 
154 #define TCP_MAX_SACK_BLOCKS 15 /**< Max number of SACK blocks stored */
155 #define TCP_INVALID_SACK_HOLE_INDEX ((u32)~0)
157 typedef struct _sack_scoreboard_hole
158 {
159  u32 next; /**< Index for next entry in linked list */
160  u32 prev; /**< Index for previous entry in linked list */
161  u32 start; /**< Start sequence number */
162  u32 end; /**< End sequence number */
163  u8 is_lost; /**< Mark hole as lost */
165 
166 typedef struct _sack_scoreboard
167 {
168  sack_scoreboard_hole_t *holes; /**< Pool of holes */
169  u32 head; /**< Index of first entry */
170  u32 tail; /**< Index of last entry */
171  u32 sacked_bytes; /**< Number of bytes sacked in sb */
172  u32 last_sacked_bytes; /**< Number of bytes last sacked */
173  u32 last_bytes_delivered; /**< Number of sack bytes delivered */
174  u32 snd_una_adv; /**< Bytes to add to snd_una */
175  u32 high_sacked; /**< Highest byte sacked (fack) */
176  u32 high_rxt; /**< Highest retransmitted sequence */
177  u32 rescue_rxt; /**< Rescue sequence number */
178  u32 lost_bytes; /**< Bytes lost as per RFC6675 */
179  u32 cur_rxt_hole; /**< Retransmitting from this hole */
181 
182 typedef enum _tcp_cc_algorithm_type
183 {
186 
187 typedef struct _tcp_cc_algorithm tcp_cc_algorithm_t;
189 typedef enum _tcp_cc_ack_t
190 {
191  TCP_CC_ACK,
195 
196 typedef struct _tcp_connection
197 {
198  transport_connection_t connection; /**< Common transport data. First! */
199 
200  u8 state; /**< TCP state as per tcp_state_t */
201  u16 flags; /**< Connection flags (see tcp_conn_flags_e) */
202  u32 timers[TCP_N_TIMERS]; /**< Timer handles into timer wheel */
203 
204  /* TODO RFC4898 */
205 
206  /** Send sequence variables RFC793 */
207  u32 snd_una; /**< oldest unacknowledged sequence number */
208  u32 snd_una_max; /**< newest unacknowledged sequence number + 1*/
209  u32 snd_wnd; /**< send window */
210  u32 snd_wl1; /**< seq number used for last snd.wnd update */
211  u32 snd_wl2; /**< ack number used for last snd.wnd update */
212  u32 snd_nxt; /**< next seq number to be sent */
213  u16 snd_mss; /**< Effective send max seg (data) size */
214 
215  /** Receive sequence variables RFC793 */
216  u32 rcv_nxt; /**< next sequence number expected */
217  u32 rcv_wnd; /**< receive window we expect */
218 
219  u32 rcv_las; /**< rcv_nxt at last ack sent/rcv_wnd update */
220  u32 iss; /**< initial sent sequence */
221  u32 irs; /**< initial remote sequence */
222 
223  /* Options */
224  tcp_options_t rcv_opts; /**< Rx options for connection */
225  tcp_options_t snd_opts; /**< Tx options for connection */
226  u8 snd_opts_len; /**< Tx options len */
227  u8 rcv_wscale; /**< Window scale to advertise to peer */
228  u8 snd_wscale; /**< Window scale to use when sending */
229  u32 tsval_recent; /**< Last timestamp received */
230  u32 tsval_recent_age; /**< When last updated tstamp_recent*/
231 
232  sack_block_t *snd_sacks; /**< Vector of SACKs to send. XXX Fixed size? */
233  sack_scoreboard_t sack_sb; /**< SACK "scoreboard" that tracks holes */
234 
235  u16 rcv_dupacks; /**< Number of DUPACKs received */
236  u8 snt_dupacks; /**< Number of DUPACKs sent in a burst */
237 
238  /* Congestion control */
239  u32 cwnd; /**< Congestion window */
240  u32 ssthresh; /**< Slow-start threshold */
241  u32 prev_ssthresh; /**< ssthresh before congestion */
242  u32 prev_cwnd; /**< ssthresh before congestion */
243  u32 bytes_acked; /**< Bytes acknowledged by current segment */
244  u32 snd_rxt_bytes; /**< Retransmitted bytes */
245  u32 snd_rxt_ts; /**< Timestamp when first packet is retransmitted */
246  u32 tsecr_last_ack; /**< Timestamp echoed to us in last healthy ACK */
247  u32 snd_congestion; /**< snd_una_max when congestion is detected */
248  tcp_cc_algorithm_t *cc_algo; /**< Congestion control algorithm */
249 
250  /* RTT and RTO */
251  u32 rto; /**< Retransmission timeout */
252  u32 rto_boff; /**< Index for RTO backoff */
253  u32 srtt; /**< Smoothed RTT */
254  u32 rttvar; /**< Smoothed mean RTT difference. Approximates variance */
255  u32 rtt_ts; /**< Timestamp for tracked ACK */
256  u32 rtt_seq; /**< Sequence number for tracked ACK */
257 
258  u16 mss; /**< Our max seg size that includes options */
259  u32 limited_transmit; /**< snd_nxt when limited transmit starts */
260  u32 last_fib_check; /**< Last time we checked fib route for peer */
262 
263 struct _tcp_cc_algorithm
264 {
265  void (*rcv_ack) (tcp_connection_t * tc);
266  void (*rcv_cong_ack) (tcp_connection_t * tc, tcp_cc_ack_t ack);
267  void (*congestion) (tcp_connection_t * tc);
268  void (*recovered) (tcp_connection_t * tc);
269  void (*init) (tcp_connection_t * tc);
270 };
271 
272 #define tcp_fastrecovery_on(tc) (tc)->flags |= TCP_CONN_FAST_RECOVERY
273 #define tcp_fastrecovery_off(tc) (tc)->flags &= ~TCP_CONN_FAST_RECOVERY
274 #define tcp_recovery_on(tc) (tc)->flags |= TCP_CONN_RECOVERY
275 #define tcp_recovery_off(tc) (tc)->flags &= ~TCP_CONN_RECOVERY
276 #define tcp_in_fastrecovery(tc) ((tc)->flags & TCP_CONN_FAST_RECOVERY)
277 #define tcp_in_recovery(tc) ((tc)->flags & (TCP_CONN_RECOVERY))
278 #define tcp_in_slowstart(tc) (tc->cwnd < tc->ssthresh)
279 #define tcp_fastrecovery_sent_1_smss(tc) ((tc)->flags & TCP_CONN_FR_1_SMSS)
280 #define tcp_fastrecovery_1_smss_on(tc) ((tc)->flags |= TCP_CONN_FR_1_SMSS)
281 #define tcp_fastrecovery_1_smss_off(tc) ((tc)->flags &= ~TCP_CONN_FR_1_SMSS)
283 #define tcp_in_cong_recovery(tc) ((tc)->flags & \
284  (TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY))
285 
286 always_inline void
288 {
289  tc->flags &= ~(TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY);
291 }
292 
293 typedef enum
294 {
295  TCP_IP4,
299 
300 typedef enum _tcp_error
301 {
302 #define tcp_error(n,s) TCP_ERROR_##n,
304 #undef tcp_error
305  TCP_N_ERROR,
307 
308 typedef struct _tcp_lookup_dispatch
309 {
310  u8 next, error;
312 
313 typedef struct _tcp_main
314 {
315  /* Per-worker thread tcp connection pools */
316  tcp_connection_t **connections;
317 
318  /* Pool of listeners. */
319  tcp_connection_t *listener_pool;
320 
321  /** Dispatch table by state and flags */
322  tcp_lookup_dispatch_t dispatch_table[TCP_N_STATES][64];
323 
324  u8 log2_tstamp_clocks_per_tick;
325  f64 tstamp_ticks_per_clock;
326 
327  /** per-worker tx buffer free lists */
328  u32 **tx_buffers;
329 
330  /* Per worker-thread timer wheel for connections timers */
331  tw_timer_wheel_16t_2w_512sl_t *timer_wheels;
332 
333 // /* Convenience per worker-thread vector of connections to DELACK */
334 // u32 **delack_connections;
335 
336  /* Pool of half-open connections on which we've sent a SYN */
337  tcp_connection_t *half_open_connections;
338 
339  /* Pool of local TCP endpoints */
340  transport_endpoint_t *local_endpoints;
341 
342  /* Local endpoints lookup table */
343  transport_endpoint_table_t local_endpoints_table;
344 
345  /* Congestion control algorithms registered */
346  tcp_cc_algorithm_t *cc_algos;
347 
348  /* Flag that indicates if stack is on or off */
349  u8 is_enabled;
350 
351  /* convenience */
356 } tcp_main_t;
357 
358 extern tcp_main_t tcp_main;
363 
366 {
367  return &tcp_main;
368 }
369 
372 {
373  ASSERT ((signed) b->current_data >= (signed) -VLIB_BUFFER_PRE_DATA_SIZE);
374  return (tcp_header_t *) (b->data + b->current_data
375  + vnet_buffer (b)->tcp.hdr_offset);
376 }
377 
379 
381 tcp_connection_get (u32 conn_index, u32 thread_index)
382 {
383  if (pool_is_free_index (tcp_main.connections[thread_index], conn_index))
384  return 0;
385  return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
386 }
387 
389 tcp_connection_get_if_valid (u32 conn_index, u32 thread_index)
390 {
391  if (tcp_main.connections[thread_index] == 0)
392  return 0;
393  if (pool_is_free_index (tcp_main.connections[thread_index], conn_index))
394  return 0;
395  return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
396 }
397 
402 
403 u8 *format_tcp_connection_id (u8 * s, va_list * args);
404 u8 *format_tcp_connection (u8 * s, va_list * args);
405 u8 *format_tcp_scoreboard (u8 * s, va_list * args);
406 
408 tcp_listener_get (u32 tli)
409 {
410  return pool_elt_at_index (tcp_main.listener_pool, tli);
411 }
412 
415 {
416  return pool_elt_at_index (tcp_main.half_open_connections, conn_index);
417 }
418 
422 void tcp_send_reset (vlib_buffer_t * pkt, u8 is_ip4);
423 void tcp_send_syn (tcp_connection_t * tc);
424 void tcp_send_fin (tcp_connection_t * tc);
425 void tcp_init_mss (tcp_connection_t * tc);
427 void tcp_update_rto (tcp_connection_t * tc);
428 
430 tcp_end_seq (tcp_header_t * th, u32 len)
431 {
432  return th->seq_number + tcp_is_syn (th) + tcp_is_fin (th) + len;
433 }
434 
435 /* Modulo arithmetic for TCP sequence numbers */
436 #define seq_lt(_s1, _s2) ((i32)((_s1)-(_s2)) < 0)
437 #define seq_leq(_s1, _s2) ((i32)((_s1)-(_s2)) <= 0)
438 #define seq_gt(_s1, _s2) ((i32)((_s1)-(_s2)) > 0)
439 #define seq_geq(_s1, _s2) ((i32)((_s1)-(_s2)) >= 0)
440 #define seq_max(_s1, _s2) (seq_gt((_s1), (_s2)) ? (_s1) : (_s2))
442 /* Modulo arithmetic for timestamps */
443 #define timestamp_lt(_t1, _t2) ((i32)((_t1)-(_t2)) < 0)
444 #define timestamp_leq(_t1, _t2) ((i32)((_t1)-(_t2)) <= 0)
446 /**
447  * Our estimate of the number of bytes that have left the network
448  */
450 tcp_bytes_out (const tcp_connection_t * tc)
451 {
452  if (tcp_opts_sack_permitted (&tc->rcv_opts))
453  return tc->sack_sb.sacked_bytes + tc->sack_sb.lost_bytes;
454  else
455  return tc->rcv_dupacks * tc->snd_mss;
456 }
457 
458 /**
459  * Our estimate of the number of bytes in flight (pipe size)
460  */
463 {
464  int flight_size;
465 
466  flight_size = (int) (tc->snd_una_max - tc->snd_una) - tcp_bytes_out (tc)
467  + tc->snd_rxt_bytes;
468 
469  if (flight_size < 0)
470  {
471  if (0)
473  ("Negative: %u %u %u dupacks %u sacked bytes %u flags %d",
474  tc->snd_una_max - tc->snd_una, tcp_bytes_out (tc),
475  tc->snd_rxt_bytes, tc->rcv_dupacks, tc->sack_sb.sacked_bytes,
476  tc->rcv_opts.flags);
477  return 0;
478  }
479 
480  return flight_size;
481 }
482 
483 /**
484  * Initial cwnd as per RFC5681
485  */
488 {
489  if (tc->snd_mss > 2190)
490  return 2 * tc->snd_mss;
491  else if (tc->snd_mss > 1095)
492  return 3 * tc->snd_mss;
493  else
494  return 4 * tc->snd_mss;
495 }
496 
498 tcp_loss_wnd (const tcp_connection_t * tc)
499 {
500  return tc->snd_mss;
501 }
502 
505 {
506  return clib_min (tc->cwnd, tc->snd_wnd);
507 }
508 
511 {
512  u32 available_wnd = tcp_available_wnd (tc);
513  u32 flight_size = tcp_flight_size (tc);
514 
515  if (available_wnd <= flight_size)
516  return 0;
517 
518  return available_wnd - flight_size;
519 }
520 
524 
532 
534 
535 /* Made public for unit testing only */
536 void tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end);
537 
539 tcp_time_now (void)
540 {
541  return clib_cpu_time_now () * tcp_main.tstamp_ticks_per_clock;
542 }
543 
544 always_inline void
545 tcp_update_time (f64 now, u32 thread_index)
546 {
547  tw_timer_expire_timers_16t_2w_512sl (&tcp_main.timer_wheels[thread_index],
548  now);
549 }
550 
552 
553 u32
555  u32 offset, u32 max_bytes);
556 
560 
561 always_inline void
563 {
564  /* Reset flags, make sure ack is sent */
565  tc->flags = TCP_CONN_SNDACK;
566  vnet_buffer (b)->tcp.flags &= ~TCP_BUF_FLAG_DUPACK;
567 }
568 
569 always_inline void
570 tcp_timer_set (tcp_connection_t * tc, u8 timer_id, u32 interval)
571 {
572  tc->timers[timer_id]
573  = tw_timer_start_16t_2w_512sl (&tcp_main.timer_wheels[tc->c_thread_index],
574  tc->c_c_index, timer_id, interval);
575 }
576 
577 always_inline void
578 tcp_timer_reset (tcp_connection_t * tc, u8 timer_id)
579 {
580  if (tc->timers[timer_id] == TCP_TIMER_HANDLE_INVALID)
581  return;
582 
583  tw_timer_stop_16t_2w_512sl (&tcp_main.timer_wheels[tc->c_thread_index],
584  tc->timers[timer_id]);
585  tc->timers[timer_id] = TCP_TIMER_HANDLE_INVALID;
586 }
587 
588 always_inline void
589 tcp_timer_update (tcp_connection_t * tc, u8 timer_id, u32 interval)
590 {
591  if (tc->timers[timer_id] != TCP_TIMER_HANDLE_INVALID)
592  tw_timer_stop_16t_2w_512sl (&tcp_main.timer_wheels[tc->c_thread_index],
593  tc->timers[timer_id]);
594  tc->timers[timer_id] =
595  tw_timer_start_16t_2w_512sl (&tcp_main.timer_wheels[tc->c_thread_index],
596  tc->c_c_index, timer_id, interval);
597 }
598 
599 /* XXX Switch retransmit to faster TW */
600 always_inline void
602 {
603  tcp_timer_set (tc, TCP_TIMER_RETRANSMIT,
604  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
605 }
606 
607 always_inline void
609 {
610  tcp_timer_reset (tc, TCP_TIMER_RETRANSMIT);
611 }
612 
613 always_inline void
615 {
616  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
617  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
618 }
619 
620 always_inline void
622 {
623  /* Reuse RTO. It's backed off in handler */
624  tcp_timer_set (tc, TCP_TIMER_PERSIST,
625  clib_max (tc->rto * TCP_TO_TIMER_TICK,
627 }
628 
629 always_inline void
631 {
632  tcp_timer_update (tc, TCP_TIMER_PERSIST,
633  clib_max (tc->rto * TCP_TO_TIMER_TICK,
635 }
636 
637 always_inline void
639 {
640  tcp_timer_reset (tc, TCP_TIMER_PERSIST);
641 }
642 
643 always_inline void
645 {
646  if (tc->snd_una == tc->snd_una_max)
647  {
649  if (tc->snd_wnd < tc->snd_mss)
651  }
652  else
653  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
654  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
655 }
656 
659 {
660  return tc->timers[timer] != TCP_TIMER_HANDLE_INVALID;
661 }
662 
663 #define tcp_validate_txf_size(_tc, _a) \
664  ASSERT(_tc->state != TCP_STATE_ESTABLISHED \
665  || stream_session_tx_fifo_max_dequeue (&_tc->connection) >= _a)
666 
667 void
669  sack_scoreboard_hole_t * hole);
672  u32 prev_index, u32 start,
673  u32 end);
676  start, u8 have_sent_1_smss,
677  u8 * can_rescue,
678  u8 * snd_limited);
680 
683 {
684  if (index != TCP_INVALID_SACK_HOLE_INDEX)
685  return pool_elt_at_index (sb->holes, index);
686  return 0;
687 }
688 
691 {
692  if (hole->next != TCP_INVALID_SACK_HOLE_INDEX)
693  return pool_elt_at_index (sb->holes, hole->next);
694  return 0;
695 }
696 
699 {
700  if (hole->prev != TCP_INVALID_SACK_HOLE_INDEX)
701  return pool_elt_at_index (sb->holes, hole->prev);
702  return 0;
703 }
704 
707 {
708  if (sb->head != TCP_INVALID_SACK_HOLE_INDEX)
709  return pool_elt_at_index (sb->holes, sb->head);
710  return 0;
711 }
712 
715 {
716  if (sb->tail != TCP_INVALID_SACK_HOLE_INDEX)
717  return pool_elt_at_index (sb->holes, sb->tail);
718  return 0;
719 }
720 
721 always_inline void
723 {
725  while ((hole = scoreboard_first_hole (sb)))
726  {
727  scoreboard_remove_hole (sb, hole);
728  }
729  ASSERT (sb->head == sb->tail && sb->head == TCP_INVALID_SACK_HOLE_INDEX);
730  sb->sacked_bytes = 0;
731  sb->last_sacked_bytes = 0;
732  sb->last_bytes_delivered = 0;
733  sb->snd_una_adv = 0;
734  sb->high_sacked = 0;
735  sb->high_rxt = 0;
736  sb->lost_bytes = 0;
737  sb->cur_rxt_hole = TCP_INVALID_SACK_HOLE_INDEX;
738 }
739 
742 {
743  return hole->end - hole->start;
744 }
745 
748 {
749  return hole - sb->holes;
750 }
751 
752 always_inline void
754 {
755  sb->head = TCP_INVALID_SACK_HOLE_INDEX;
756  sb->tail = TCP_INVALID_SACK_HOLE_INDEX;
757  sb->cur_rxt_hole = TCP_INVALID_SACK_HOLE_INDEX;
758 }
759 
760 void tcp_rcv_sacks (tcp_connection_t * tc, u32 ack);
761 
762 always_inline void
764  const tcp_cc_algorithm_t * vft)
765 {
766  tcp_main_t *tm = vnet_get_tcp_main ();
767  vec_validate (tm->cc_algos, type);
768 
769  tm->cc_algos[type] = *vft;
770 }
771 
774 {
775  tcp_main_t *tm = vnet_get_tcp_main ();
776  return &tm->cc_algos[type];
777 }
778 
779 void tcp_cc_init (tcp_connection_t * tc);
780 
781 /**
782  * Push TCP header to buffer
783  *
784  * @param vm - vlib_main
785  * @param b - buffer to write the header to
786  * @param sp_net - source port net order
787  * @param dp_net - destination port net order
788  * @param seq - sequence number net order
789  * @param ack - ack number net order
790  * @param tcp_hdr_opts_len - header and options length in bytes
791  * @param flags - header flags
792  * @param wnd - window size
793  *
794  * @return - pointer to start of TCP header
795  */
796 always_inline void *
798  u32 ack, u8 tcp_hdr_opts_len, u8 flags,
799  u16 wnd)
800 {
801  tcp_header_t *th;
802 
803  th = vlib_buffer_push_uninit (b, tcp_hdr_opts_len);
804 
805  th->src_port = sp;
806  th->dst_port = dp;
807  th->seq_number = seq;
808  th->ack_number = ack;
809  th->data_offset_and_reserved = (tcp_hdr_opts_len >> 2) << 4;
810  th->flags = flags;
811  th->window = wnd;
812  th->checksum = 0;
813  th->urgent_pointer = 0;
814  return th;
815 }
816 
817 /**
818  * Push TCP header to buffer
819  *
820  * @param b - buffer to write the header to
821  * @param sp_net - source port net order
822  * @param dp_net - destination port net order
823  * @param seq - sequence number host order
824  * @param ack - ack number host order
825  * @param tcp_hdr_opts_len - header and options length in bytes
826  * @param flags - header flags
827  * @param wnd - window size
828  *
829  * @return - pointer to start of TCP header
830  */
831 always_inline void *
832 vlib_buffer_push_tcp (vlib_buffer_t * b, u16 sp_net, u16 dp_net, u32 seq,
833  u32 ack, u8 tcp_hdr_opts_len, u8 flags, u16 wnd)
834 {
835  return vlib_buffer_push_tcp_net_order (b, sp_net, dp_net,
836  clib_host_to_net_u32 (seq),
837  clib_host_to_net_u32 (ack),
838  tcp_hdr_opts_len, flags,
839  clib_host_to_net_u16 (wnd));
840 }
841 
842 #endif /* _vnet_tcp_h_ */
843 
844 /*
845  * fd.io coding-style-patch-verification: ON
846  *
847  * Local Variables:
848  * eval: (c-set-style "gnu")
849  * End:
850  */
#define vec_validate(V, I)
Make sure vector is long enough for given index (no header, unspecified alignment) ...
Definition: vec.h:436
tcp_main_t tcp_main
Definition: tcp.c:22
#define tcp_fastrecovery_1_smss_off(tc)
Definition: tcp.h:282
#define clib_min(x, y)
Definition: clib.h:332
void scoreboard_init_high_rxt(sack_scoreboard_t *sb)
Definition: tcp_input.c:677
static void tcp_retransmit_timer_set(tcp_connection_t *tc)
Definition: tcp.h:602
static u32 tcp_available_wnd(const tcp_connection_t *tc)
Definition: tcp.h:505
void tcp_make_fin(tcp_connection_t *tc, vlib_buffer_t *b)
Convert buffer to FIN-ACK.
Definition: tcp_output.c:518
void tcp_send_reset(vlib_buffer_t *pkt, u8 is_ip4)
Send reset without reusing existing buffer.
Definition: tcp_output.c:684
struct _sack_block sack_block_t
void tcp_cc_init_congestion(tcp_connection_t *tc)
Definition: tcp_input.c:875
void tcp_connection_timers_reset(tcp_connection_t *tc)
Stop all connection timers.
Definition: tcp.c:337
struct _transport_connection transport_connection_t
#define TCP_TO_TIMER_TICK
Definition: tcp.h:94
vlib_node_registration_t tcp4_output_node
(constructor) VLIB_REGISTER_NODE (tcp4_output_node)
Definition: tcp_output.c:1594
void tcp_fast_retransmit(tcp_connection_t *tc)
Do fast retransmit.
Definition: tcp_output.c:1408
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:451
sack_scoreboard_hole_t * scoreboard_insert_hole(sack_scoreboard_t *sb, u32 prev_index, u32 start, u32 end)
Definition: tcp_input.c:541
static tcp_connection_t * tcp_connection_get_if_valid(u32 conn_index, u32 thread_index)
Definition: tcp.h:390
void tcp_connection_del(tcp_connection_t *tc)
Connection removal.
Definition: tcp.c:140
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:1247
u32 tcp_snd_space(tcp_connection_t *tc)
Compute tx window session is allowed to fill.
Definition: tcp.c:823
static tcp_connection_t * tcp_half_open_connection_get(u32 conn_index)
Definition: tcp.h:415
#define VLIB_BUFFER_PRE_DATA_SIZE
Definition: buffer.h:51
void scoreboard_update_lost(tcp_connection_t *tc, sack_scoreboard_t *sb)
struct _tcp_main tcp_main_t
vlib_node_registration_t tcp6_output_node
(constructor) VLIB_REGISTER_NODE (tcp6_output_node)
Definition: tcp_output.c:1615
static u64 clib_cpu_time_now(void)
Definition: time.h:73
timer_expiration_handler tcp_timer_retransmit_handler
static void scoreboard_init(sack_scoreboard_t *sb)
Definition: tcp.h:754
u8 * format_tcp_scoreboard(u8 *s, va_list *args)
Definition: tcp.c:739
struct _vlib_node_registration vlib_node_registration_t
static sack_scoreboard_hole_t * scoreboard_prev_hole(sack_scoreboard_t *sb, sack_scoreboard_hole_t *hole)
Definition: tcp.h:699
struct _tcp_lookup_dispatch tcp_lookup_dispatch_t
u32 tcp_prepare_retransmit_segment(tcp_connection_t *tc, vlib_buffer_t *b, u32 offset, u32 max_bytes)
Build a retransmit segment.
Definition: tcp_output.c:989
struct _tcp_connection tcp_connection_t
void tcp_fast_retransmit_sack(tcp_connection_t *tc)
Do fast retransmit with SACKs.
Definition: tcp_output.c:1291
void tcp_connection_cleanup(tcp_connection_t *tc)
Cleans up connection state.
Definition: tcp.c:104
u8 *( format_function_t)(u8 *s, va_list *args)
Definition: format.h:48
static void scoreboard_clear(sack_scoreboard_t *sb)
Definition: tcp.h:723
format_function_t format_tcp_flags
Definition: tcp.h:63
struct _tcp_header tcp_header_t
static u32 tcp_available_snd_space(const tcp_connection_t *tc)
Definition: tcp.h:511
void tcp_connection_reset(tcp_connection_t *tc)
Notify session that connection has been reset.
Definition: tcp.c:152
struct _sack_scoreboard_hole sack_scoreboard_hole_t
static u32 scoreboard_hole_index(sack_scoreboard_t *sb, sack_scoreboard_hole_t *hole)
Definition: tcp.h:748
i16 current_data
signed offset in data[], pre_data[] that we are currently processing.
Definition: buffer.h:67
#define tcp_is_fin(_th)
Definition: tcp_packet.h:90
i32 tcp_rcv_wnd_available(tcp_connection_t *tc)
Definition: tcp.c:876
#define always_inline
Definition: clib.h:84
static tcp_header_t * tcp_buffer_hdr(vlib_buffer_t *b)
Definition: tcp.h:372
int i32
Definition: types.h:81
static timer_callback_t * timers
Definition: timer.c:56
enum _tcp_state tcp_state_t
timer_expiration_handler tcp_timer_retransmit_syn_handler
static u32 tcp_time_now(void)
Definition: tcp.h:540
static tcp_cc_algorithm_t * tcp_cc_algo_get(tcp_cc_algorithm_type_e type)
Definition: tcp.h:774
static u32 scoreboard_hole_bytes(sack_scoreboard_hole_t *hole)
Definition: tcp.h:742
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:622
static void tcp_timer_set(tcp_connection_t *tc, u8 timer_id, u32 interval)
Definition: tcp.h:571
Definition: tcp.h:298
#define TCP_INVALID_SACK_HOLE_INDEX
Definition: tcp.h:156
#define pool_elt_at_index(p, i)
Returns pointer to element at given index.
Definition: pool.h:397
timer_expiration_handler tcp_timer_persist_handler
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:798
void tcp_retransmit_first_unacked(tcp_connection_t *tc)
Retransmit first unacked segment.
Definition: tcp_output.c:1258
void tcp_rcv_sacks(tcp_connection_t *tc, u32 ack)
Definition: tcp_input.c:686
clib_error_t * vnet_tcp_enable_disable(vlib_main_t *vm, u8 is_en)
Definition: tcp.c:1097
void tcp_send_syn(tcp_connection_t *tc)
Send SYN.
Definition: tcp_output.c:789
static sack_scoreboard_hole_t * scoreboard_next_hole(sack_scoreboard_t *sb, sack_scoreboard_hole_t *hole)
Definition: tcp.h:691
#define TCP_TIMER_HANDLE_INVALID
Definition: tcp.h:91
void tcp_fast_retransmit_no_sack(tcp_connection_t *tc)
Fast retransmit without SACK info.
Definition: tcp_output.c:1365
static u32 tcp_flight_size(const tcp_connection_t *tc)
Our estimate of the number of bytes in flight (pipe size)
Definition: tcp.h:463
vnet_main_t vnet_main
Definition: misc.c:43
u32 tcp_push_header(transport_connection_t *tconn, vlib_buffer_t *b)
Definition: tcp_output.c:1637
void( timer_expiration_handler)(u32 index)
Definition: tcp.h:84
static sack_scoreboard_hole_t * scoreboard_last_hole(sack_scoreboard_t *sb)
Definition: tcp.h:715
enum _tcp_cc_ack_t tcp_cc_ack_t
static void tcp_timer_reset(tcp_connection_t *tc, u8 timer_id)
Definition: tcp.h:579
static sack_scoreboard_hole_t * scoreboard_first_hole(sack_scoreboard_t *sb)
Definition: tcp.h:707
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:833
void tcp_cc_init(tcp_connection_t *tc)
Definition: tcp_input.c:1138
clib_bihash_24_8_t transport_endpoint_table_t
Definition: transport.h:236
format_function_t format_tcp_state
Definition: tcp.h:62
#define clib_warning(format, args...)
Definition: error.h:59
enum _tcp_timers tcp_timers_e
static void tcp_connection_force_ack(tcp_connection_t *tc, vlib_buffer_t *b)
Definition: tcp.h:563
u32 fib_node_index_t
A typedef of a node index.
Definition: fib_types.h:28
void tcp_connection_timers_init(tcp_connection_t *tc)
Initialize all connection timers as invalid.
Definition: tcp.c:320
void tcp_make_synack(tcp_connection_t *ts, vlib_buffer_t *b)
Convert buffer to SYN-ACK.
Definition: tcp_output.c:538
#define pool_is_free_index(P, I)
Use free bitmap to query whether given index is free.
Definition: pool.h:238
u8 * format_tcp_connection_id(u8 *s, va_list *args)
Definition: tcp.c:645
vlib_node_registration_t tcp6_input_node
(constructor) VLIB_REGISTER_NODE (tcp6_input_node)
Definition: tcp_input.c:2846
fib_node_index_t tcp_lookup_rmt_in_fib(tcp_connection_t *tc)
Definition: tcp.c:396
void tcp_make_ack(tcp_connection_t *ts, vlib_buffer_t *b)
Convert buffer to ACK.
Definition: tcp_output.c:503
static void tcp_timer_update(tcp_connection_t *tc, u8 timer_id, u32 interval)
Definition: tcp.h:590
int vlib_main(vlib_main_t *volatile vm, unformat_input_t *input)
Definition: main.c:1692
#define ASSERT(truth)
unsigned int u32
Definition: types.h:88
Definition: tcp.h:296
enum _tcp_cc_algorithm_type tcp_cc_algorithm_type_e
ip6_main_t ip6_main
Definition: ip6_forward.c:2926
IPv4 main type.
Definition: ip4.h:83
enum _tcp_connection_flag_bits tcp_connection_flag_bits_e
vhost_vring_state_t state
Definition: vhost-user.h:81
void tcp_connection_init_vars(tcp_connection_t *tc)
Initialize tcp connection variables.
Definition: tcp.c:444
u8 * format_tcp_connection(u8 *s, va_list *args)
Definition: tcp.c:669
static u32 tcp_end_seq(tcp_header_t *th, u32 len)
Definition: tcp.h:431
static void init(void)
Definition: client.c:76
struct _tcp_cc_algorithm tcp_cc_algorithm_t
Definition: tcp.h:188
void scoreboard_remove_hole(sack_scoreboard_t *sb, sack_scoreboard_hole_t *hole)
Definition: tcp_input.c:510
#define clib_max(x, y)
Definition: clib.h:325
template key/value backing page structure
Definition: bihash_doc.h:44
#define tcp_is_syn(_th)
Definition: tcp_packet.h:89
void tcp_update_snd_mss(tcp_connection_t *tc)
Update snd_mss to reflect the effective segment size that we can send by taking into account all TCP ...
Definition: tcp_output.c:390
static void tcp_persist_timer_update(tcp_connection_t *tc)
Definition: tcp.h:631
static sack_scoreboard_hole_t * scoreboard_get_hole(sack_scoreboard_t *sb, u32 index)
Definition: tcp.h:683
Definition: tcp.h:297
static void * vlib_buffer_push_uninit(vlib_buffer_t *b, u8 size)
Prepend uninitialized data to buffer.
Definition: buffer.h:285
unsigned short u16
Definition: types.h:57
vlib_node_registration_t tcp4_input_node
(constructor) VLIB_REGISTER_NODE (tcp4_input_node)
Definition: tcp_input.c:2823
static u32 tcp_initial_cwnd(const tcp_connection_t *tc)
Initial cwnd as per RFC5681.
Definition: tcp.h:488
tcp_af_t
Definition: tcp.h:294
#define foreach_tcp_fsm_state
TCP FSM state definitions as per RFC793.
Definition: tcp.h:41
void tcp_send_fin(tcp_connection_t *tc)
Send FIN.
Definition: tcp_output.c:866
#define foreach_tcp_connection_flag
TCP connection flags.
Definition: tcp.h:110
#define foreach_tcp_timer
TCP timers.
Definition: tcp.h:67
static void tcp_update_time(f64 now, u32 thread_index)
Definition: tcp.h:546
double f64
Definition: types.h:142
unsigned char u8
Definition: types.h:56
static void tcp_retransmit_timer_update(tcp_connection_t *tc)
Definition: tcp.h:645
struct _transport_endpoint transport_endpoint_t
void tcp_connection_close(tcp_connection_t *tc)
Begin connection closing procedure.
Definition: tcp.c:194
static void tcp_retransmit_timer_force_update(tcp_connection_t *tc)
Definition: tcp.h:615
#define vnet_buffer(b)
Definition: buffer.h:303
static tcp_connection_t * tcp_connection_get(u32 conn_index, u32 thread_index)
Definition: tcp.h:382
ip4_main_t ip4_main
Global ip4 main structure.
Definition: ip4_forward.c:1168
void tcp_update_rto(tcp_connection_t *tc)
Definition: tcp_input.c:405
void tcp_init_mss(tcp_connection_t *tc)
Definition: tcp_output.c:403
u8 data[0]
Packet data.
Definition: buffer.h:152
static void tcp_cc_algo_register(tcp_cc_algorithm_type_e type, const tcp_cc_algorithm_t *vft)
Definition: tcp.h:764
#define foreach_tcp_buf_flag
TCP buffer flags.
Definition: tcp.h:135
enum _tcp_connection_flag tcp_connection_flags_e
format_function_t format_tcp_sacks
Definition: tcp.h:64
#define TCP_TIMER_PERSIST_MIN
Definition: tcp.h:101
#define tcp_opts_sack_permitted(_to)
Definition: tcp_packet.h:160
static u32 tcp_loss_wnd(const tcp_connection_t *tc)
Definition: tcp.h:499
u32 flags
Definition: vhost-user.h:76
static void tcp_persist_timer_set(tcp_connection_t *tc)
Definition: tcp.h:622
static tcp_main_t * vnet_get_tcp_main()
Definition: tcp.h:366
static void tcp_cong_recovery_off(tcp_connection_t *tc)
Definition: tcp.h:288
timer_expiration_handler tcp_timer_delack_handler
static void tcp_retransmit_timer_reset(tcp_connection_t *tc)
Definition: tcp.h:609
void tcp_cc_fastrecovery_exit(tcp_connection_t *tc)
Definition: tcp_input.c:894
void tcp_update_rcv_wnd(tcp_connection_t *tc)
Definition: tcp_output.c:127
static u8 tcp_timer_is_active(tcp_connection_t *tc, tcp_timers_e timer)
Definition: tcp.h:659
int tcp_cc_recover(tcp_connection_t *tc)
Definition: tcp_input.c:925
static tcp_connection_t * tcp_listener_get(u32 tli)
Definition: tcp.h:409
static void tcp_persist_timer_reset(tcp_connection_t *tc)
Definition: tcp.h:639