FD.io VPP  v19.08.3-2-gbabecb413
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 #define TCP_MAX_GSO_SZ 65536
35 #define TCP_RXT_MAX_BURST 10
36 
37 #define TCP_DUPACK_THRESHOLD 3
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  _(WAITCLOSE, "WAIT CLOSE") \
75  _(RETRANSMIT_SYN, "RETRANSMIT SYN") \
76 
77 typedef enum _tcp_timers
78 {
79 #define _(sym, str) TCP_TIMER_##sym,
81 #undef _
83 } tcp_timers_e;
84 
85 typedef void (timer_expiration_handler) (u32 index, u32 thread_index);
86 
91 
92 #define TCP_TIMER_HANDLE_INVALID ((u32) ~0)
93 
94 #define TCP_TIMER_TICK 0.1 /**< Timer tick in seconds */
95 #define TCP_TO_TIMER_TICK TCP_TICK*10 /**< Factor for converting
96  ticks to timer ticks */
97 
98 #define TCP_RTO_MAX 60 * THZ /* Min max RTO (60s) as per RFC6298 */
99 #define TCP_RTO_MIN 0.2 * THZ /* Min RTO (200ms) - lower than standard */
100 #define TCP_RTT_MAX 30 * THZ /* 30s (probably too much) */
101 #define TCP_RTO_SYN_RETRIES 3 /* SYN retries without doubling RTO */
102 #define TCP_RTO_INIT 1 * THZ /* Initial retransmit timer */
103 #define TCP_RTO_BOFF_MAX 8 /* Max number of retries before reset */
104 #define TCP_ESTABLISH_TIME (60 * THZ) /* Connection establish timeout */
106 /** Connection configuration flags */
107 #define foreach_tcp_cfg_flag \
108  _(RATE_SAMPLE, "Rate sampling") \
109  _(NO_CSUM_OFFLOAD, "No csum offload") \
110  _(NO_TSO, "TSO off") \
111  _(TSO, "TSO") \
112 
113 typedef enum tcp_cfg_flag_bits_
114 {
115 #define _(sym, str) TCP_CFG_F_##sym##_BIT,
117 #undef _
120 
121 typedef enum tcp_cfg_flag_
122 {
123 #define _(sym, str) TCP_CFG_F_##sym = 1 << TCP_CFG_F_##sym##_BIT,
125 #undef _
128 
129 /** TCP connection flags */
130 #define foreach_tcp_connection_flag \
131  _(SNDACK, "Send ACK") \
132  _(FINSNT, "FIN sent") \
133  _(RECOVERY, "Recovery") \
134  _(FAST_RECOVERY, "Fast Recovery") \
135  _(DCNT_PENDING, "Disconnect pending") \
136  _(HALF_OPEN_DONE, "Half-open completed") \
137  _(FINPNDG, "FIN pending") \
138  _(RXT_PENDING, "Retransmit pending") \
139  _(FRXT_FIRST, "Retransmit first") \
140  _(DEQ_PENDING, "Dequeue pending ") \
141  _(PSH_PENDING, "PSH pending") \
142  _(FINRCVD, "FIN received") \
143  _(ZERO_RWND_SENT, "Zero RWND sent") \
144 
145 typedef enum tcp_connection_flag_bits_
146 {
147 #define _(sym, str) TCP_CONN_##sym##_BIT,
149 #undef _
152 
153 typedef enum tcp_connection_flag_
154 {
155 #define _(sym, str) TCP_CONN_##sym = 1 << TCP_CONN_##sym##_BIT,
157 #undef _
160 
161 #define TCP_SCOREBOARD_TRACE (0)
162 #define TCP_MAX_SACK_BLOCKS 256 /**< Max number of SACK blocks stored */
163 #define TCP_INVALID_SACK_HOLE_INDEX ((u32)~0)
165 typedef struct _scoreboard_trace_elt
166 {
167  u32 start;
168  u32 end;
169  u32 ack;
170  u32 snd_una_max;
171  u32 group;
173 
174 typedef struct _sack_scoreboard_hole
175 {
176  u32 next; /**< Index for next entry in linked list */
177  u32 prev; /**< Index for previous entry in linked list */
178  u32 start; /**< Start sequence number */
179  u32 end; /**< End sequence number */
180  u8 is_lost; /**< Mark hole as lost */
182 
183 typedef struct _sack_scoreboard
184 {
185  sack_scoreboard_hole_t *holes; /**< Pool of holes */
186  u32 head; /**< Index of first entry */
187  u32 tail; /**< Index of last entry */
188  u32 sacked_bytes; /**< Number of bytes sacked in sb */
189  u32 last_sacked_bytes; /**< Number of bytes last sacked */
190  u32 last_bytes_delivered; /**< Sack bytes delivered to app */
191  u32 rxt_sacked; /**< Rxt bytes last delivered */
192  u32 high_sacked; /**< Highest byte sacked (fack) */
193  u32 high_rxt; /**< Highest retransmitted sequence */
194  u32 rescue_rxt; /**< Rescue sequence number */
195  u32 lost_bytes; /**< Bytes lost as per RFC6675 */
196  u32 last_lost_bytes; /**< Number of bytes last lost */
197  u32 cur_rxt_hole; /**< Retransmitting from this hole */
198  u8 is_reneging;
199 
200 #if TCP_SCOREBOARD_TRACE
202 #endif
203 
205 
206 #if TCP_SCOREBOARD_TRACE
207 #define tcp_scoreboard_trace_add(_tc, _ack) \
208 { \
209  static u64 _group = 0; \
210  sack_scoreboard_t *_sb = &_tc->sack_sb; \
211  sack_block_t *_sack, *_sacks; \
212  scoreboard_trace_elt_t *_elt; \
213  int i; \
214  _group++; \
215  _sacks = _tc->rcv_opts.sacks; \
216  for (i = 0; i < vec_len (_sacks); i++) \
217  { \
218  _sack = &_sacks[i]; \
219  vec_add2 (_sb->trace, _elt, 1); \
220  _elt->start = _sack->start; \
221  _elt->end = _sack->end; \
222  _elt->ack = _elt->end == _ack ? _ack : 0; \
223  _elt->snd_una_max = _elt->end == _ack ? _tc->snd_una_max : 0; \
224  _elt->group = _group; \
225  } \
226 }
227 #else
228 #define tcp_scoreboard_trace_add(_tc, _ack)
229 #endif
230 
233  start, u8 have_sent_1_smss,
234  u8 * can_rescue,
235  u8 * snd_limited);
237  u32 index);
238 
240  sack_scoreboard_hole_t * hole);
242  sack_scoreboard_hole_t * hole);
245 
247 void scoreboard_clear_reneging (sack_scoreboard_t * sb, u32 start, u32 end);
249 void scoreboard_init_rxt (sack_scoreboard_t * sb, u32 snd_una);
250 u8 *format_tcp_scoreboard (u8 * s, va_list * args);
251 
252 #define TCP_BTS_INVALID_INDEX ((u32)~0)
254 typedef enum tcp_bts_flags_
255 {
256  TCP_BTS_IS_RXT = 1,
260 } __clib_packed tcp_bts_flags_t;
261 
262 typedef struct tcp_bt_sample_
263 {
264  u32 next; /**< Next sample index in list */
265  u32 prev; /**< Previous sample index in list */
266  u32 min_seq; /**< Min seq number in sample */
267  u32 max_seq; /**< Max seq number. Set for rxt samples */
268  u64 delivered; /**< Total delivered bytes for sample */
269  f64 delivered_time; /**< Delivered time when sample taken */
270  f64 tx_time; /**< Transmit time for the burst */
271  f64 first_tx_time; /**< Connection first tx time at tx */
272  tcp_bts_flags_t flags; /**< Sample flag */
274 
275 typedef struct tcp_rate_sample_
276 {
277  u64 prior_delivered; /**< Delivered of sample used for rate, i.e.,
278  total bytes delivered at prior_time */
279  f64 prior_time; /**< Delivered time of sample used for rate */
280  f64 interval_time; /**< Time to ack the bytes delivered */
281  f64 rtt_time; /**< RTT for sample */
282  u32 delivered; /**< Bytes delivered in interval_time */
283  u32 acked_and_sacked; /**< Bytes acked + sacked now */
284  u32 lost; /**< Bytes lost now */
285  tcp_bts_flags_t flags; /**< Rate sample flags from bt sample */
287 
288 typedef struct tcp_byte_tracker_
289 {
290  tcp_bt_sample_t *samples; /**< Pool of samples */
291  rb_tree_t sample_lookup; /**< Rbtree for sample lookup by min_seq */
292  u32 head; /**< Head of samples linked list */
293  u32 tail; /**< Tail of samples linked list */
294  u32 last_ooo; /**< Cached last ooo sample */
296 
297 typedef enum _tcp_cc_algorithm_type
298 {
303 
304 typedef struct _tcp_cc_algorithm tcp_cc_algorithm_t;
306 typedef enum _tcp_cc_ack_t
307 {
308  TCP_CC_ACK,
312 
313 typedef enum tcp_cc_event_
314 {
317 
318 /*
319  * As per RFC4898 tcpEStatsStackSoftErrors
320  */
321 typedef struct tcp_errors_
322 {
323  u32 below_data_wnd; /**< All data in seg is below snd_una */
324  u32 above_data_wnd; /**< Some data in segment is above snd_wnd */
325  u32 below_ack_wnd; /**< Acks for data below snd_una */
326  u32 above_ack_wnd; /**< Acks for data not sent */
328 
329 typedef struct _tcp_connection
330 {
331  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
332  transport_connection_t connection; /**< Common transport data. First! */
333 
334  u8 state; /**< TCP state as per tcp_state_t */
335  u8 cfg_flags; /**< Connection configuration flags */
336  u16 flags; /**< Connection flags (see tcp_conn_flags_e) */
337  u32 timers[TCP_N_TIMERS]; /**< Timer handles into timer wheel */
338 
339  u64 segs_in; /** RFC4022/4898 tcpHCInSegs/tcpEStatsPerfSegsIn */
340  u64 bytes_in; /** RFC4898 tcpEStatsPerfHCDataOctetsIn */
341  u64 segs_out; /** RFC4898 tcpEStatsPerfSegsOut */
342  u64 bytes_out; /** RFC4898 tcpEStatsPerfHCDataOctetsOut */
343 
344  /** Send sequence variables RFC793 */
345  u32 snd_una; /**< oldest unacknowledged sequence number */
346  u32 snd_una_max; /**< newest unacknowledged sequence number + 1*/
347  u32 snd_wnd; /**< send window */
348  u32 snd_wl1; /**< seq number used for last snd.wnd update */
349  u32 snd_wl2; /**< ack number used for last snd.wnd update */
350  u32 snd_nxt; /**< next seq number to be sent */
351  u16 snd_mss; /**< Effective send max seg (data) size */
352 
353  u64 data_segs_in; /** RFC4898 tcpEStatsPerfDataSegsIn */
354  u64 data_segs_out; /** RFC4898 tcpEStatsPerfDataSegsOut */
355 
356  /** Receive sequence variables RFC793 */
357  u32 rcv_nxt; /**< next sequence number expected */
358  u32 rcv_wnd; /**< receive window we expect */
359 
360  u32 rcv_las; /**< rcv_nxt at last ack sent/rcv_wnd update */
361  u32 iss; /**< initial sent sequence */
362  u32 irs; /**< initial remote sequence */
363 
364  /* Options */
365  u8 snd_opts_len; /**< Tx options len */
366  u8 rcv_wscale; /**< Window scale to advertise to peer */
367  u8 snd_wscale; /**< Window scale to use when sending */
368  u32 tsval_recent; /**< Last timestamp received */
369  u32 tsval_recent_age; /**< When last updated tstamp_recent*/
370  tcp_options_t snd_opts; /**< Tx options for connection */
371  tcp_options_t rcv_opts; /**< Rx options for connection */
372 
373  sack_block_t *snd_sacks; /**< Vector of SACKs to send. XXX Fixed size? */
374  u8 snd_sack_pos; /**< Position in vec of first block to send */
375  sack_block_t *snd_sacks_fl; /**< Vector for building new list */
376  sack_scoreboard_t sack_sb; /**< SACK "scoreboard" that tracks holes */
377 
378  u16 rcv_dupacks; /**< Number of recent DUPACKs received */
379  u32 dupacks_in; /**< RFC4898 tcpEStatsStackDupAcksIn*/
380  u8 pending_dupacks; /**< Number of DUPACKs to be sent */
381  u32 dupacks_out; /**< RFC4898 tcpEStatsPathDupAcksOut */
382 
383  /* Congestion control */
384  u32 cwnd; /**< Congestion window */
385  u32 cwnd_acc_bytes; /**< Bytes accumulated for cwnd increment */
386  u32 ssthresh; /**< Slow-start threshold */
387  u32 prev_ssthresh; /**< ssthresh before congestion */
388  u32 prev_cwnd; /**< ssthresh before congestion */
389  u32 bytes_acked; /**< Bytes acknowledged by current segment */
390  u32 burst_acked; /**< Bytes acknowledged in current burst */
391  u32 snd_rxt_bytes; /**< Retransmitted bytes during current cc event */
392  u32 snd_rxt_ts; /**< Timestamp when first packet is retransmitted */
393  u32 prr_delivered; /**< RFC6937 bytes delivered during current event */
394  u32 prr_start; /**< snd_una when prr starts */
395  u32 rxt_delivered; /**< Rxt bytes delivered during current cc event */
396  u32 rxt_head; /**< snd_una last time we re rxted the head */
397  u32 prev_dsegs_out; /**< Number of dsegs after last ack */
398  u32 tsecr_last_ack; /**< Timestamp echoed to us in last healthy ACK */
399  u32 snd_congestion; /**< snd_una_max when congestion is detected */
400  u32 tx_fifo_size; /**< Tx fifo size. Used to constrain cwnd */
401  tcp_cc_algorithm_t *cc_algo; /**< Congestion control algorithm */
402  u8 cc_data[TCP_CC_DATA_SZ]; /**< Congestion control algo private data */
403 
404  u32 fr_occurences; /**< fast-retransmit occurrences RFC4898
405  tcpEStatsStackFastRetran */
406  u32 tr_occurences; /**< timer-retransmit occurrences */
407  u64 bytes_retrans; /**< RFC4898 tcpEStatsPerfOctetsRetrans */
408  u64 segs_retrans; /**< RFC4898 tcpEStatsPerfSegsRetrans*/
409 
410  /* RTT and RTO */
411  u32 rto; /**< Retransmission timeout */
412  u32 rto_boff; /**< Index for RTO backoff */
413  u32 srtt; /**< Smoothed RTT */
414  u32 rttvar; /**< Smoothed mean RTT difference. Approximates variance */
415  u32 rtt_seq; /**< Sequence number for tracked ACK */
416  f64 rtt_ts; /**< Timestamp for tracked ACK */
417  f64 mrtt_us; /**< High precision mrtt from tracked acks */
418 
419  u32 psh_seq; /**< Add psh header for seg that includes this */
420  u32 next_node_index; /**< Can be used to control next node in output */
421  u32 next_node_opaque; /**< Opaque to pass to next node */
422  u32 limited_transmit; /**< snd_nxt when limited transmit starts */
423  u32 sw_if_index; /**< Interface for the connection */
424 
425  /* Delivery rate estimation */
426  u64 delivered; /**< Total bytes delivered to peer */
427  u64 app_limited; /**< Delivered when app-limited detected */
428  f64 delivered_time; /**< Time last bytes were acked */
429  f64 first_tx_time; /**< Send time for recently delivered/sent */
430  tcp_byte_tracker_t *bt; /**< Tx byte tracker */
431 
432  tcp_errors_t errors; /**< Soft connection errors */
433 
434  f64 start_ts; /**< Timestamp when connection initialized */
435  u32 last_fib_check; /**< Last time we checked fib route for peer */
436  u16 mss; /**< Our max seg size that includes options */
437  u32 timestamp_delta; /**< Offset for timestamp */
439 
440 /* *INDENT-OFF* */
441 struct _tcp_cc_algorithm
442 {
443  const char *name;
444  uword (*unformat_cfg) (unformat_input_t * input);
445  void (*init) (tcp_connection_t * tc);
446  void (*cleanup) (tcp_connection_t * tc);
447  void (*rcv_ack) (tcp_connection_t * tc, tcp_rate_sample_t *rs);
448  void (*rcv_cong_ack) (tcp_connection_t * tc, tcp_cc_ack_t ack,
449  tcp_rate_sample_t *rs);
450  void (*congestion) (tcp_connection_t * tc);
451  void (*loss) (tcp_connection_t * tc);
452  void (*recovered) (tcp_connection_t * tc);
453  void (*undo_recovery) (tcp_connection_t * tc);
454  void (*event) (tcp_connection_t *tc, tcp_cc_event_t evt);
455  u64 (*get_pacing_rate) (tcp_connection_t *tc);
456 };
457 /* *INDENT-ON* */
458 
459 #define tcp_fastrecovery_on(tc) (tc)->flags |= TCP_CONN_FAST_RECOVERY
460 #define tcp_fastrecovery_off(tc) (tc)->flags &= ~TCP_CONN_FAST_RECOVERY
461 #define tcp_recovery_on(tc) (tc)->flags |= TCP_CONN_RECOVERY
462 #define tcp_recovery_off(tc) (tc)->flags &= ~TCP_CONN_RECOVERY
463 #define tcp_in_fastrecovery(tc) ((tc)->flags & TCP_CONN_FAST_RECOVERY)
464 #define tcp_in_recovery(tc) ((tc)->flags & (TCP_CONN_RECOVERY))
465 #define tcp_in_slowstart(tc) (tc->cwnd < tc->ssthresh)
466 #define tcp_disconnect_pending(tc) ((tc)->flags & TCP_CONN_DCNT_PENDING)
467 #define tcp_disconnect_pending_on(tc) ((tc)->flags |= TCP_CONN_DCNT_PENDING)
468 #define tcp_disconnect_pending_off(tc) ((tc)->flags &= ~TCP_CONN_DCNT_PENDING)
469 #define tcp_fastrecovery_first(tc) ((tc)->flags & TCP_CONN_FRXT_FIRST)
470 #define tcp_fastrecovery_first_on(tc) ((tc)->flags |= TCP_CONN_FRXT_FIRST)
471 #define tcp_fastrecovery_first_off(tc) ((tc)->flags &= ~TCP_CONN_FRXT_FIRST)
473 #define tcp_in_cong_recovery(tc) ((tc)->flags & \
474  (TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY))
475 
476 #define tcp_csum_offload(tc) (!((tc)->cfg_flags & TCP_CFG_F_NO_CSUM_OFFLOAD))
478 always_inline void
480 {
481  tc->flags &= ~(TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY);
483 }
484 
485 #define tcp_zero_rwnd_sent(tc) ((tc)->flags & TCP_CONN_ZERO_RWND_SENT)
486 #define tcp_zero_rwnd_sent_on(tc) (tc)->flags |= TCP_CONN_ZERO_RWND_SENT
487 #define tcp_zero_rwnd_sent_off(tc) (tc)->flags &= ~TCP_CONN_ZERO_RWND_SENT
489 typedef enum _tcp_error
490 {
491 #define tcp_error(n,s) TCP_ERROR_##n,
493 #undef tcp_error
494  TCP_N_ERROR,
496 
497 typedef struct _tcp_lookup_dispatch
498 {
499  u8 next, error;
501 
502 typedef struct tcp_worker_ctx_
503 {
504  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
505  /** worker time */
506  u32 time_now;
508  /** worker timer wheel */
509  tw_timer_wheel_16t_2w_512sl_t timer_wheel;
511  /** tx buffer free list */
512  u32 *tx_buffers;
514  /** tx frames for ip 4/6 lookup nodes */
515  vlib_frame_t *ip_lookup_tx_frames[2];
517  /** vector of pending ack dequeues */
518  u32 *pending_deq_acked;
520  /** vector of pending disconnect notifications */
521  u32 *pending_disconnects;
523  /** convenience pointer to this thread's vlib main */
524  vlib_main_t *vm;
526  CLIB_CACHE_LINE_ALIGN_MARK (cacheline1);
527 
528  /** cached 'on the wire' options for bursts */
529  u8 cached_opts[40];
532 
533 typedef struct tcp_iss_seed_
534 {
535  u64 first;
536  u64 second;
538 
539 typedef struct tcp_configuration_
540 {
541  /** Max rx fifo size for a session (in bytes). It is used in to compute the
542  * rfc 7323 window scaling factor */
543  u32 max_rx_fifo;
545  /** Min rx fifo for a session (in bytes) */
546  u32 min_rx_fifo;
548  /** Default MTU to be used when establishing connections */
549  u16 default_mtu;
551  /** Initial CWND multiplier, which multiplies MSS to determine initial CWND.
552  * Set 0 to determine the initial CWND by another way */
553  u16 initial_cwnd_multiplier;
555  /** Enable tx pacing for new connections */
556  u8 enable_tx_pacing;
558  /** Allow use of TSO whenever available */
559  u8 allow_tso;
561  /** Default congestion control algorithm type */
562  tcp_cc_algorithm_type_e cc_algo;
564  /** Min rwnd, as number of snd_mss segments, for update ack to be sent after
565  * a zero rwnd advertisement */
566  u32 rwnd_min_update_ack;
568  /** Delayed ack time (disabled) */
569  u16 delack_time;
571  /** Timer ticks to wait for close from app */
572  u16 closewait_time;
574  /** Timer ticks to wait in time-wait. Also known as 2MSL */
575  u16 timewait_time;
577  /** Timer ticks to wait in fin-wait1 to send fin and rcv fin-ack */
578  u16 finwait1_time;
580  /** Timer ticks to wait in last ack for ack */
581  u16 lastack_time;
583  /** Timer ticks to wait in fin-wait2 for fin */
584  u16 finwait2_time;
586  /** Timer ticks to wait in closing for fin ack */
587  u16 closing_time;
589  /** Timer ticks to wait before cleaning up the connection */
590  u16 cleanup_time;
592  /** Number of preallocated connections */
593  u32 preallocated_connections;
595  /** Number of preallocated half-open connections */
596  u32 preallocated_half_open_connections;
598  /** Vectors of src addresses. Optional unless one needs > 63K active-opens */
599  ip4_address_t *ip4_src_addrs;
600  ip6_address_t *ip6_src_addrs;
602  /** Fault-injection. Debug only */
603  f64 buffer_fail_fraction;
605 
606 typedef struct _tcp_main
607 {
608  /* Per-worker thread tcp connection pools */
610 
611  /* Pool of listeners. */
612  tcp_connection_t *listener_pool;
613 
614  /** Dispatch table by state and flags */
615  tcp_lookup_dispatch_t dispatch_table[TCP_N_STATES][64];
616 
617  u8 log2_tstamp_clocks_per_tick;
618  f64 tstamp_ticks_per_clock;
619 
620  /** per-worker context */
621  tcp_worker_ctx_t *wrk_ctx;
622 
623  /** Pool of half-open connections on which we've sent a SYN */
624  tcp_connection_t *half_open_connections;
625  clib_spinlock_t half_open_lock;
626 
627  /** vlib buffer size */
628  u32 bytes_per_buffer;
629 
630  /** Seed used to generate random iss */
631  tcp_iss_seed_t iss_seed;
632 
633  /** Congestion control algorithms registered */
634  tcp_cc_algorithm_t *cc_algos;
635 
636  /** Hash table of cc algorithms by name */
637  uword *cc_algo_by_name;
638 
639  /** Last cc algo registered */
640  tcp_cc_algorithm_type_e cc_last_type;
641 
642  /** Flag that indicates if stack is on or off */
643  u8 is_enabled;
644 
645  /** Flag that indicates if v4 punting is enabled */
646  u8 punt_unknown4;
647 
648  /** Flag that indicates if v6 punting is enabled */
649  u8 punt_unknown6;
650 
651  /** Rotor for v4 source addresses */
652  u32 last_v4_addr_rotor;
653 
654  /** Rotor for v6 source addresses */
655  u32 last_v6_addr_rotor;
656 
657  /** Protocol configuration */
659 } tcp_main_t;
660 
661 extern tcp_main_t tcp_main;
674 
675 #define tcp_cfg tcp_main.cfg
676 #define tcp_node_index(node_id, is_ip4) \
677  ((is_ip4) ? tcp4_##node_id##_node.index : tcp6_##node_id##_node.index)
678 
681 {
682  return &tcp_main;
683 }
684 
686 tcp_get_worker (u32 thread_index)
687 {
688  return &tcp_main.wrk_ctx[thread_index];
689 }
690 
693 {
694  ASSERT ((signed) b->current_data >= (signed) -VLIB_BUFFER_PRE_DATA_SIZE);
695  return (tcp_header_t *) (b->data + b->current_data
696  + vnet_buffer (b)->tcp.hdr_offset);
697 }
698 
699 #if (VLIB_BUFFER_TRACE_TRAJECTORY)
700 #define tcp_trajectory_add_start(b, start) \
701 { \
702  (*vlib_buffer_trace_trajectory_cb) (b, start); \
703 }
704 #else
705 #define tcp_trajectory_add_start(b, start)
706 #endif
707 
709 
710 void tcp_punt_unknown (vlib_main_t * vm, u8 is_ip4, u8 is_add);
711 
713 tcp_connection_get (u32 conn_index, u32 thread_index)
714 {
715  if (PREDICT_FALSE
716  (pool_is_free_index (tcp_main.connections[thread_index], conn_index)))
717  return 0;
718  return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
719 }
720 
722 tcp_connection_get_if_valid (u32 conn_index, u32 thread_index)
723 {
724  if (tcp_main.connections[thread_index] == 0)
725  return 0;
726  if (pool_is_free_index (tcp_main.connections[thread_index], conn_index))
727  return 0;
728  return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
729 }
730 
733 {
734  return (tcp_connection_t *) tconn;
735 }
736 
737 always_inline void
739 {
740  tc->state = state;
741  TCP_EVT (TCP_EVT_STATE_CHANGE, tc);
742 }
743 
750  tcp_connection_t * base);
754  ip4_address_t * start,
755  ip4_address_t * end, u32 table_id);
757  ip6_address_t * start,
758  ip6_address_t * end, u32 table_id);
759 void tcp_api_reference (void);
760 u8 *format_tcp_connection (u8 * s, va_list * args);
761 u8 *format_tcp_connection_id (u8 * s, va_list * args);
762 
764 tcp_listener_get (u32 tli)
765 {
766  return pool_elt_at_index (tcp_main.listener_pool, tli);
767 }
768 
771 {
772  tcp_connection_t *tc = 0;
773  clib_spinlock_lock_if_init (&tcp_main.half_open_lock);
774  if (!pool_is_free_index (tcp_main.half_open_connections, conn_index))
775  tc = pool_elt_at_index (tcp_main.half_open_connections, conn_index);
776  clib_spinlock_unlock_if_init (&tcp_main.half_open_lock);
777  return tc;
778 }
779 
783  u32 thread_index, u8 is_ip4);
784 void tcp_send_reset (tcp_connection_t * tc);
785 void tcp_send_syn (tcp_connection_t * tc);
787 void tcp_send_fin (tcp_connection_t * tc);
788 void tcp_send_ack (tcp_connection_t * tc);
790 void tcp_update_rto (tcp_connection_t * tc);
793 
797 
798 /*
799  * Rate estimation
800  */
801 
802 /**
803  * Byte tracker initialize
804  *
805  * @param tc connection for which the byte tracker should be allocated and
806  * initialized
807  */
808 void tcp_bt_init (tcp_connection_t * tc);
809 /**
810  * Byte tracker cleanup
811  *
812  * @param tc connection for which the byte tracker should be cleaned up
813  */
814 void tcp_bt_cleanup (tcp_connection_t * tc);
815 /**
816  * Flush byte tracker samples
817  *
818  * @param tc tcp connection for which samples should be flushed
819  */
821 /**
822  * Track a tcp tx burst
823  *
824  * @param tc tcp connection
825  */
827 /**
828  * Track a tcp retransmission
829  *
830  * @param tc tcp connection
831  * @param start start sequence number
832  * @param end end sequence number
833  */
834 void tcp_bt_track_rxt (tcp_connection_t * tc, u32 start, u32 end);
835 /**
836  * Generate a delivery rate sample from recently acked bytes
837  *
838  * @param tc tcp connection
839  * @param rs resulting rate sample
840  */
842  tcp_rate_sample_t * rs);
843 /**
844  * Check if sample to be generated is app limited
845  *
846  * @param tc tcp connection
847  */
849 /**
850  * Check if the byte tracker is in sane state
851  *
852  * Should be used only for testing
853  *
854  * @param bt byte tracker
855  */
857 u8 *format_tcp_bt (u8 * s, va_list * args);
858 
861 {
862  return th->seq_number + tcp_is_syn (th) + tcp_is_fin (th) + len;
863 }
864 
865 /* Modulo arithmetic for TCP sequence numbers */
866 #define seq_lt(_s1, _s2) ((i32)((_s1)-(_s2)) < 0)
867 #define seq_leq(_s1, _s2) ((i32)((_s1)-(_s2)) <= 0)
868 #define seq_gt(_s1, _s2) ((i32)((_s1)-(_s2)) > 0)
869 #define seq_geq(_s1, _s2) ((i32)((_s1)-(_s2)) >= 0)
870 #define seq_max(_s1, _s2) (seq_gt((_s1), (_s2)) ? (_s1) : (_s2))
872 /* Modulo arithmetic for timestamps */
873 #define timestamp_lt(_t1, _t2) ((i32)((_t1)-(_t2)) < 0)
874 #define timestamp_leq(_t1, _t2) ((i32)((_t1)-(_t2)) <= 0)
876 /**
877  * Our estimate of the number of bytes that have left the network
878  */
880 tcp_bytes_out (const tcp_connection_t * tc)
881 {
882  if (tcp_opts_sack_permitted (&tc->rcv_opts))
883  return tc->sack_sb.sacked_bytes + tc->sack_sb.lost_bytes;
884  else
885  return clib_min (tc->rcv_dupacks * tc->snd_mss,
886  tc->snd_nxt - tc->snd_una);
887 }
888 
889 /**
890  * Our estimate of the number of bytes in flight (pipe size)
891  */
894 {
895  int flight_size;
896 
897  flight_size = (int) (tc->snd_nxt - tc->snd_una) - tcp_bytes_out (tc)
898  + tc->snd_rxt_bytes - tc->rxt_delivered;
899 
900  ASSERT (flight_size >= 0);
901 
902  return flight_size;
903 }
904 
905 /**
906  * Initial cwnd as per RFC5681
907  */
910 {
911  if (tcp_cfg.initial_cwnd_multiplier > 0)
912  return tcp_cfg.initial_cwnd_multiplier * tc->snd_mss;
913 
914  if (tc->snd_mss > 2190)
915  return 2 * tc->snd_mss;
916  else if (tc->snd_mss > 1095)
917  return 3 * tc->snd_mss;
918  else
919  return 4 * tc->snd_mss;
920 }
921 
922 /*
923  * Accumulate acked bytes for cwnd increase
924  *
925  * Once threshold bytes are accumulated, snd_mss bytes are added
926  * to the cwnd.
927  */
928 always_inline void
929 tcp_cwnd_accumulate (tcp_connection_t * tc, u32 thresh, u32 bytes)
930 {
931  tc->cwnd_acc_bytes += bytes;
932  if (tc->cwnd_acc_bytes >= thresh)
933  {
934  u32 inc = tc->cwnd_acc_bytes / thresh;
935  tc->cwnd_acc_bytes -= inc * thresh;
936  tc->cwnd += inc * tc->snd_mss;
937  tc->cwnd = clib_min (tc->cwnd, tc->tx_fifo_size);
938  }
939 }
940 
942 tcp_loss_wnd (const tcp_connection_t * tc)
943 {
944  /* Whatever we have in flight + the packet we're about to send */
945  return tcp_flight_size (tc) + tc->snd_mss;
946 }
947 
950 {
951  return clib_min (tc->cwnd, tc->snd_wnd);
952 }
953 
956 {
957  u32 available_wnd = tcp_available_snd_wnd (tc);
958  int flight_size = (int) (tc->snd_nxt - tc->snd_una);
959 
960  if (available_wnd <= flight_size)
961  return 0;
962 
963  return available_wnd - flight_size;
964 }
965 
966 /**
967  * Estimate of how many bytes we can still push into the network
968  */
971 {
972  u32 available_wnd = tcp_available_snd_wnd (tc);
973  u32 flight_size = tcp_flight_size (tc);
974 
975  if (available_wnd <= flight_size)
976  return 0;
977 
978  return available_wnd - flight_size;
979 }
980 
983 {
984  if ((tc->flags & TCP_CONN_FINSNT) && (tc->snd_una_max - tc->snd_una == 1))
985  return 1;
986  return 0;
987 }
988 
991 
993 
994 /* Made public for unit testing only */
995 void tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end);
997 
999 tcp_time_now (void)
1001  return tcp_main.wrk_ctx[vlib_get_thread_index ()].time_now;
1002 }
1003 
1005 tcp_time_now_w_thread (u32 thread_index)
1007  return tcp_main.wrk_ctx[thread_index].time_now;
1008 }
1009 
1010 /**
1011  * Generate timestamp for tcp connection
1012  */
1016  return (tcp_main.wrk_ctx[tc->c_thread_index].time_now -
1017  tc->timestamp_delta);
1018 }
1019 
1021 tcp_time_now_us (u32 thread_index)
1023  return transport_time_now (thread_index);
1024 }
1025 
1029  wrk->time_now = clib_cpu_time_now () * tcp_main.tstamp_ticks_per_clock;
1030  return wrk->time_now;
1031 }
1032 
1034  vlib_buffer_t * b);
1035 int tcp_session_custom_tx (void *conn, u32 max_burst_size);
1036 
1043  u32 start_bucket);
1044 
1045 always_inline void
1048  tc->cc_algo->rcv_ack (tc, rs);
1049  tc->tsecr_last_ack = tc->rcv_opts.tsecr;
1050 }
1051 
1052 static inline void
1055 {
1056  tc->cc_algo->rcv_cong_ack (tc, ack_type, rs);
1057 }
1058 
1059 static inline void
1062  tc->cc_algo->congestion (tc);
1063 }
1064 
1065 static inline void
1068  tc->cc_algo->loss (tc);
1069 }
1070 
1071 static inline void
1074  tc->cc_algo->recovered (tc);
1075 }
1076 
1077 static inline void
1080  if (tc->cc_algo->undo_recovery)
1081  tc->cc_algo->undo_recovery (tc);
1082 }
1083 
1084 static inline void
1087  if (tc->cc_algo->event)
1088  tc->cc_algo->event (tc, evt);
1089 }
1090 
1091 static inline u64
1094  if (tc->cc_algo->get_pacing_rate)
1095  return tc->cc_algo->get_pacing_rate (tc);
1096 
1097  f64 srtt = clib_min ((f64) tc->srtt * TCP_TICK, tc->mrtt_us);
1098 
1099  /* TODO should constrain to interface's max throughput but
1100  * we don't have link speeds for sw ifs ..*/
1101  return ((f64) tc->cwnd / srtt);
1102 }
1103 
1104 always_inline void
1105 tcp_timer_set (tcp_connection_t * tc, u8 timer_id, u32 interval)
1107  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
1108  ASSERT (tc->timers[timer_id] == TCP_TIMER_HANDLE_INVALID);
1109  tc->timers[timer_id] =
1110  tw_timer_start_16t_2w_512sl (&tcp_main.
1111  wrk_ctx[tc->c_thread_index].timer_wheel,
1112  tc->c_c_index, timer_id, interval);
1113 }
1114 
1115 always_inline void
1116 tcp_timer_reset (tcp_connection_t * tc, u8 timer_id)
1118  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
1119  if (tc->timers[timer_id] == TCP_TIMER_HANDLE_INVALID)
1120  return;
1121 
1122  tw_timer_stop_16t_2w_512sl (&tcp_main.
1123  wrk_ctx[tc->c_thread_index].timer_wheel,
1124  tc->timers[timer_id]);
1125  tc->timers[timer_id] = TCP_TIMER_HANDLE_INVALID;
1126 }
1127 
1128 always_inline void
1129 tcp_timer_update (tcp_connection_t * tc, u8 timer_id, u32 interval)
1131  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
1132  if (tc->timers[timer_id] != TCP_TIMER_HANDLE_INVALID)
1133  tw_timer_update_16t_2w_512sl (&tcp_main.
1134  wrk_ctx[tc->c_thread_index].timer_wheel,
1135  tc->timers[timer_id], interval);
1136  else
1137  tc->timers[timer_id] =
1138  tw_timer_start_16t_2w_512sl (&tcp_main.
1139  wrk_ctx[tc->c_thread_index].timer_wheel,
1140  tc->c_c_index, timer_id, interval);
1141 }
1142 
1143 always_inline void
1146  ASSERT (tc->snd_una != tc->snd_una_max);
1147  tcp_timer_set (tc, TCP_TIMER_RETRANSMIT,
1148  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
1149 }
1150 
1151 always_inline void
1154  tcp_timer_reset (tc, TCP_TIMER_RETRANSMIT);
1155 }
1156 
1157 always_inline void
1160  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
1161  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
1162 }
1163 
1164 always_inline void
1167  /* Reuse RTO. It's backed off in handler */
1168  tcp_timer_set (tc, TCP_TIMER_PERSIST,
1169  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
1170 }
1171 
1172 always_inline void
1175  u32 interval;
1176 
1177  if (seq_leq (tc->snd_una, tc->snd_congestion + tc->burst_acked))
1178  interval = 1;
1179  else
1180  interval = clib_max (tc->rto * TCP_TO_TIMER_TICK, 1);
1181 
1182  tcp_timer_update (tc, TCP_TIMER_PERSIST, interval);
1183 }
1184 
1185 always_inline void
1188  tcp_timer_reset (tc, TCP_TIMER_PERSIST);
1189 }
1190 
1191 always_inline void
1194  if (tc->snd_una == tc->snd_nxt)
1195  {
1197  if (tc->snd_wnd < tc->snd_mss)
1199  }
1200  else
1201  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
1202  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
1203 }
1204 
1208  return tc->timers[timer] != TCP_TIMER_HANDLE_INVALID;
1209 }
1210 
1211 #define tcp_validate_txf_size(_tc, _a) \
1212  ASSERT(_tc->state != TCP_STATE_ESTABLISHED \
1213  || transport_max_tx_dequeue (&_tc->connection) >= _a)
1214 
1215 void tcp_rcv_sacks (tcp_connection_t * tc, u32 ack);
1216 u8 *tcp_scoreboard_replay (u8 * s, tcp_connection_t * tc, u8 verbose);
1217 
1218 /**
1219  * Register exiting cc algo type
1220  */
1222  const tcp_cc_algorithm_t * vft);
1223 
1224 /**
1225  * Register new cc algo type
1226  */
1229 
1230 static inline void *
1233  return (void *) tc->cc_data;
1234 }
1235 
1237  tcp_rate_sample_t * rs);
1238 
1239 /**
1240  * Push TCP header to buffer
1241  *
1242  * @param vm - vlib_main
1243  * @param b - buffer to write the header to
1244  * @param sp_net - source port net order
1245  * @param dp_net - destination port net order
1246  * @param seq - sequence number net order
1247  * @param ack - ack number net order
1248  * @param tcp_hdr_opts_len - header and options length in bytes
1249  * @param flags - header flags
1250  * @param wnd - window size
1251  *
1252  * @return - pointer to start of TCP header
1253  */
1254 always_inline void *
1256  u32 ack, u8 tcp_hdr_opts_len, u8 flags,
1257  u16 wnd)
1258 {
1259  tcp_header_t *th;
1260 
1261  th = vlib_buffer_push_uninit (b, tcp_hdr_opts_len);
1262 
1263  th->src_port = sp;
1264  th->dst_port = dp;
1265  th->seq_number = seq;
1266  th->ack_number = ack;
1267  th->data_offset_and_reserved = (tcp_hdr_opts_len >> 2) << 4;
1268  th->flags = flags;
1269  th->window = wnd;
1270  th->checksum = 0;
1271  th->urgent_pointer = 0;
1272  vnet_buffer (b)->l4_hdr_offset = (u8 *) th - b->data;
1273  b->flags |= VNET_BUFFER_F_L4_HDR_OFFSET_VALID;
1274  return th;
1275 }
1276 
1277 /**
1278  * Push TCP header to buffer
1279  *
1280  * @param b - buffer to write the header to
1281  * @param sp_net - source port net order
1282  * @param dp_net - destination port net order
1283  * @param seq - sequence number host order
1284  * @param ack - ack number host order
1285  * @param tcp_hdr_opts_len - header and options length in bytes
1286  * @param flags - header flags
1287  * @param wnd - window size
1288  *
1289  * @return - pointer to start of TCP header
1290  */
1291 always_inline void *
1292 vlib_buffer_push_tcp (vlib_buffer_t * b, u16 sp_net, u16 dp_net, u32 seq,
1293  u32 ack, u8 tcp_hdr_opts_len, u8 flags, u16 wnd)
1294 {
1295  return vlib_buffer_push_tcp_net_order (b, sp_net, dp_net,
1296  clib_host_to_net_u32 (seq),
1297  clib_host_to_net_u32 (ack),
1298  tcp_hdr_opts_len, flags,
1299  clib_host_to_net_u16 (wnd));
1300 }
1301 
1302 #endif /* _vnet_tcp_h_ */
1303 
1304 /*
1305  * fd.io coding-style-patch-verification: ON
1306  *
1307  * Local Variables:
1308  * eval: (c-set-style "gnu")
1309  * End:
1310  */
tcp_main_t tcp_main
Definition: tcp.c:30
u32 flags
buffer flags: VLIB_BUFFER_FREE_LIST_INDEX_MASK: bits used to store free list index, VLIB_BUFFER_IS_TRACED: trace this buffer.
Definition: buffer.h:124
enum tcp_connection_flag_ tcp_connection_flags_e
#define foreach_tcp_cfg_flag
Connection configuration flags.
Definition: tcp.h:108
void() timer_expiration_handler(u32 index, u32 thread_index)
Definition: tcp.h:85
vlib_node_registration_t tcp6_listen_node
(constructor) VLIB_REGISTER_NODE (tcp6_listen_node)
Definition: tcp_input.c:3368
void tcp_program_retransmit(tcp_connection_t *tc)
Definition: tcp_output.c:1203
static vlib_cli_command_t trace
(constructor) VLIB_CLI_COMMAND (trace)
Definition: vlib_api_cli.c:889
#define CLIB_CACHE_LINE_ALIGN_MARK(mark)
Definition: cache.h:60
u32 flags
Definition: vhost_user.h:141
void tcp_send_window_update_ack(tcp_connection_t *tc)
Send window update ack.
Definition: tcp_output.c:1234
#define clib_min(x, y)
Definition: clib.h:302
void scoreboard_clear(sack_scoreboard_t *sb)
Definition: tcp_input.c:949
static f64 tcp_time_now_us(u32 thread_index)
Definition: tcp.h:1022
static void tcp_retransmit_timer_set(tcp_connection_t *tc)
Definition: tcp.h:1145
static clib_time_type_t transport_time_now(u32 thread_index)
Definition: session.h:513
#define seq_leq(_s1, _s2)
Definition: tcp.h:868
void tcp_make_fin(tcp_connection_t *tc, vlib_buffer_t *b)
Convert buffer to FIN-ACK.
Definition: tcp_output.c:566
struct _sack_block sack_block_t
struct _scoreboard_trace_elt scoreboard_trace_elt_t
void tcp_connection_timers_reset(tcp_connection_t *tc)
Stop all connection timers.
Definition: tcp.c:520
#define TCP_TO_TIMER_TICK
Factor for converting ticks to timer ticks.
Definition: tcp.h:95
vlib_node_registration_t tcp4_output_node
(constructor) VLIB_REGISTER_NODE (tcp4_output_node)
Definition: tcp_output.c:2479
void scoreboard_init(sack_scoreboard_t *sb)
Definition: tcp_input.c:941
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:881
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:723
void tcp_connection_del(tcp_connection_t *tc)
Connection removal.
Definition: tcp.c:292
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:1762
enum tcp_bts_flags_ tcp_bts_flags_t
u32 tcp_snd_space(tcp_connection_t *tc)
Definition: tcp.c:1318
int tcp_bt_is_sane(tcp_byte_tracker_t *bt)
Check if the byte tracker is in sane state.
Definition: tcp_bt.c:213
static tcp_connection_t * tcp_half_open_connection_get(u32 conn_index)
Definition: tcp.h:771
sack_scoreboard_hole_t * scoreboard_last_hole(sack_scoreboard_t *sb)
Definition: tcp_input.c:703
sack_scoreboard_hole_t * scoreboard_next_hole(sack_scoreboard_t *sb, sack_scoreboard_hole_t *hole)
Definition: tcp_input.c:679
#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:110
struct _tcp_main tcp_main_t
vlib_node_registration_t tcp6_output_node
(constructor) VLIB_REGISTER_NODE (tcp6_output_node)
Definition: tcp_output.c:2499
static u64 clib_cpu_time_now(void)
Definition: time.h:81
timer_expiration_handler tcp_timer_retransmit_handler
u8 * format_tcp_scoreboard(u8 *s, va_list *args)
Definition: tcp.c:1177
u32 tcp_session_push_header(transport_connection_t *tconn, vlib_buffer_t *b)
Definition: tcp_output.c:1134
struct _tcp_lookup_dispatch tcp_lookup_dispatch_t
void tcp_connection_tx_pacer_update(tcp_connection_t *tc)
Definition: tcp.c:1392
enum tcp_cc_event_ tcp_cc_event_t
void tcp_update_burst_snd_vars(tcp_connection_t *tc)
Update burst send vars.
Definition: tcp_output.c:387
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:971
static u32 tcp_available_snd_wnd(const tcp_connection_t *tc)
Definition: tcp.h:950
static tcp_connection_t * tcp_get_connection_from_transport(transport_connection_t *tconn)
Definition: tcp.h:733
void tcp_connection_cleanup(tcp_connection_t *tc)
Cleans up connection state.
Definition: tcp.c:239
format_function_t format_tcp_flags
Definition: tcp.h:65
struct _tcp_header tcp_header_t
tcp_cfg_flag_bits_
Definition: tcp.h:114
void tcp_connection_reset(tcp_connection_t *tc)
Notify session that connection has been reset.
Definition: tcp.c:343
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:2713
#define TCP_TICK
TCP tick period (s)
Definition: tcp.h:27
u8 * tcp_scoreboard_replay(u8 *s, tcp_connection_t *tc, u8 verbose)
Definition: tcp.c:2161
#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:739
#define tcp_cfg
Definition: tcp.h:676
vl_api_interface_index_t sw_if_index
Definition: gre.api:50
sack_scoreboard_hole_t * scoreboard_get_hole(sack_scoreboard_t *sb, u32 index)
Definition: tcp_input.c:671
#define always_inline
Definition: clib.h:99
struct tcp_bt_sample_ tcp_bt_sample_t
static u32 tcp_available_output_snd_space(const tcp_connection_t *tc)
Definition: tcp.h:956
static void tcp_cc_loss(tcp_connection_t *tc)
Definition: tcp.h:1067
static tcp_header_t * tcp_buffer_hdr(vlib_buffer_t *b)
Definition: tcp.h:693
enum _tcp_state tcp_state_t
vhost_vring_state_t state
Definition: vhost_user.h:146
timer_expiration_handler tcp_timer_retransmit_syn_handler
static u32 tcp_time_now(void)
Definition: tcp.h:1000
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:867
void tcp_api_reference(void)
Definition: tcp_api.c:109
struct tcp_byte_tracker_ tcp_byte_tracker_t
static void tcp_timer_set(tcp_connection_t *tc, u8 timer_id, u32 interval)
Definition: tcp.h:1106
static heap_elt_t * first(heap_header_t *h)
Definition: heap.c:59
vl_api_fib_path_type_t type
Definition: fib_types.api:123
struct tcp_worker_ctx_ tcp_worker_ctx_t
tcp_cfg_flag_
Definition: tcp.h:122
enum tcp_cfg_flag_bits_ tcp_cfg_flag_bits_e
#define pool_elt_at_index(p, i)
Returns pointer to element at given index.
Definition: pool.h:514
void tcp_bt_init(tcp_connection_t *tc)
Byte tracker initialize.
Definition: tcp_bt.c:655
timer_expiration_handler tcp_timer_persist_handler
void tcp_send_ack(tcp_connection_t *tc)
Definition: tcp_output.c:1162
vlib_node_registration_t tcp6_syn_sent_node
(constructor) VLIB_REGISTER_NODE (tcp6_syn_sent_node)
Definition: tcp_input.c:2732
u32 tcp_sack_list_bytes(tcp_connection_t *tc)
Definition: tcp_input.c:1811
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:1256
int tcp_fastrecovery_prr_snd_space(tcp_connection_t *tc)
Estimate send space using proportional rate reduction (RFC6937)
Definition: tcp_output.c:1803
void tcp_rcv_sacks(tcp_connection_t *tc, u32 ack)
Definition: tcp_input.c:1003
clib_error_t * vnet_tcp_enable_disable(vlib_main_t *vm, u8 is_en)
Definition: tcp.c:1627
void tcp_send_syn(tcp_connection_t *tc)
Send SYN.
Definition: tcp_output.c:924
static void * tcp_cc_data(tcp_connection_t *tc)
Definition: tcp.h:1232
struct _unformat_input_t unformat_input_t
unsigned short u16
Definition: types.h:57
enum tcp_connection_flag_bits_ tcp_connection_flag_bits_e
sack_scoreboard_hole_t * scoreboard_prev_hole(sack_scoreboard_t *sb, sack_scoreboard_hole_t *hole)
Definition: tcp_input.c:687
static void cleanup(void)
Definition: client.c:131
#define TCP_TIMER_HANDLE_INVALID
Definition: tcp.h:92
static void tcp_cc_rcv_cong_ack(tcp_connection_t *tc, tcp_cc_ack_t ack_type, tcp_rate_sample_t *rs)
Definition: tcp.h:1054
static u32 tcp_flight_size(const tcp_connection_t *tc)
Our estimate of the number of bytes in flight (pipe size)
Definition: tcp.h:894
#define PREDICT_FALSE(x)
Definition: clib.h:112
tcp_connection_t * tcp_connection_alloc(u8 thread_index)
Definition: tcp.c:299
void tcp_connection_tx_pacer_reset(tcp_connection_t *tc, u32 window, u32 start_bucket)
Definition: tcp.c:1402
void tcp_cc_algo_register(tcp_cc_algorithm_type_e type, const tcp_cc_algorithm_t *vft)
Register exiting cc algo type.
Definition: tcp.c:87
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:930
void newreno_rcv_cong_ack(tcp_connection_t *tc, tcp_cc_ack_t ack_type, tcp_rate_sample_t *rs)
Definition: tcp_newreno.c:61
u8 len
Definition: ip_types.api:90
tcp_bts_flags_
Definition: tcp.h:255
static void tcp_timer_reset(tcp_connection_t *tc, u8 timer_id)
Definition: tcp.h:1117
vlib_node_registration_t tcp4_listen_node
(constructor) VLIB_REGISTER_NODE (tcp4_listen_node)
Definition: tcp_input.c:3349
void scoreboard_init_rxt(sack_scoreboard_t *sb, u32 snd_una)
Definition: tcp_input.c:927
tcp_connection_flag_bits_
Definition: tcp.h:146
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:1293
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:1850
void scoreboard_clear_reneging(sack_scoreboard_t *sb, u32 start, u32 end)
Definition: tcp_input.c:968
static_always_inline uword vlib_get_thread_index(void)
Definition: threads.h:218
vlib_main_t * vm
Definition: buffer.c:323
void tcp_program_dupack(tcp_connection_t *tc)
Definition: tcp_output.c:1191
void tcp_send_reset(tcp_connection_t *tc)
Build and set reset packet for connection.
Definition: tcp_output.c:858
static u32 tcp_tstamp(tcp_connection_t *tc)
Generate timestamp for tcp connection.
Definition: tcp.h:1015
void tcp_punt_unknown(vlib_main_t *vm, u8 is_ip4, u8 is_add)
Definition: tcp.c:1645
format_function_t format_tcp_state
Definition: tcp.h:64
tcp_connection_flag_
Definition: tcp.h:154
static void tcp_cc_undo_recovery(tcp_connection_t *tc)
Definition: tcp.h:1079
enum _tcp_timers tcp_timers_e
u8 data[]
Packet data.
Definition: buffer.h:181
void tcp_bt_track_tx(tcp_connection_t *tc, u32 len)
Track a tcp tx burst.
Definition: tcp_bt.c:297
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:211
u8 * format_tcp_bt(u8 *s, va_list *args)
Definition: tcp_bt.c:681
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:503
void tcp_make_synack(tcp_connection_t *ts, vlib_buffer_t *b)
Convert buffer to SYN-ACK.
Definition: tcp_output.c:601
#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
u8 * format_tcp_connection_id(u8 *s, va_list *args)
Definition: tcp.c:1036
vlib_node_registration_t tcp6_input_node
(constructor) VLIB_REGISTER_NODE (tcp6_input_node)
Definition: tcp_input.c:3795
fib_node_index_t tcp_lookup_rmt_in_fib(tcp_connection_t *tc)
void tcp_send_synack(tcp_connection_t *tc)
Definition: tcp_output.c:959
static void tcp_timer_update(tcp_connection_t *tc, u8 timer_id, u32 interval)
Definition: tcp.h:1130
#define ASSERT(truth)
enum _tcp_cc_algorithm_type tcp_cc_algorithm_type_e
struct tcp_rate_sample_ tcp_rate_sample_t
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:1002
void tcp_connection_init_vars(tcp_connection_t *tc)
Initialize tcp connection variables.
Definition: tcp.c:724
void tcp_bt_sample_delivery_rate(tcp_connection_t *tc, tcp_rate_sample_t *rs)
Generate a delivery rate sample from recently acked bytes.
Definition: tcp_bt.c:586
vlib_node_registration_t tcp4_rcv_process_node
(constructor) VLIB_REGISTER_NODE (tcp4_rcv_process_node)
Definition: tcp_input.c:3146
u8 * format_tcp_connection(u8 *s, va_list *args)
Definition: tcp.c:1060
static u32 tcp_end_seq(tcp_header_t *th, u32 len)
Definition: tcp.h:861
static void init(void)
Definition: client.c:116
vlib_node_registration_t tcp6_established_node
(constructor) VLIB_REGISTER_NODE (tcp6_established_node)
Definition: tcp_input.c:2283
#define tcp_fastrecovery_first_off(tc)
Definition: tcp.h:472
struct _tcp_cc_algorithm tcp_cc_algorithm_t
Definition: tcp.h:305
#define clib_max(x, y)
Definition: clib.h:295
static u32 tcp_time_now_w_thread(u32 thread_index)
Definition: tcp.h:1006
tcp_cc_event_
Definition: tcp.h:314
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:1174
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:775
static void * vlib_buffer_push_uninit(vlib_buffer_t *b, u8 size)
Prepend uninitialized data to buffer.
Definition: buffer.h:335
void tcp_bt_check_app_limited(tcp_connection_t *tc)
Check if sample to be generated is app limited.
Definition: tcp_bt.c:282
vlib_node_registration_t tcp4_input_node
(constructor) VLIB_REGISTER_NODE (tcp4_input_node)
Definition: tcp_input.c:3775
tcp_cc_algorithm_t * tcp_cc_algo_get(tcp_cc_algorithm_type_e type)
Definition: tcp.c:98
static u32 tcp_initial_cwnd(const tcp_connection_t *tc)
Initial cwnd as per RFC5681.
Definition: tcp.h:910
vlib_node_registration_t tcp6_rcv_process_node
(constructor) VLIB_REGISTER_NODE (tcp6_rcv_process_node)
Definition: tcp_input.c:3165
#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:1012
#define foreach_tcp_connection_flag
TCP connection flags.
Definition: tcp.h:131
#define foreach_tcp_timer
TCP timers.
Definition: tcp.h:70
static u8 tcp_is_lost_fin(tcp_connection_t *tc)
Definition: tcp.h:983
static void tcp_cc_rcv_ack(tcp_connection_t *tc, tcp_rate_sample_t *rs)
Definition: tcp.h:1047
sack_scoreboard_hole_t * scoreboard_first_hole(sack_scoreboard_t *sb)
Definition: tcp_input.c:695
static tcp_worker_ctx_t * tcp_get_worker(u32 thread_index)
Definition: tcp.h:687
static void tcp_retransmit_timer_update(tcp_connection_t *tc)
Definition: tcp.h:1193
VLIB buffer representation.
Definition: buffer.h:102
u64 uword
Definition: types.h:112
u32 time_now
worker time
Definition: tcp.h:507
static void tcp_cc_recovered(tcp_connection_t *tc)
Definition: tcp.h:1073
void tcp_init_snd_vars(tcp_connection_t *tc)
Initialize connection send variables.
Definition: tcp.c:692
struct tcp_configuration_ tcp_configuration_t
void tcp_connection_close(tcp_connection_t *tc)
Begin connection closing procedure.
Definition: tcp.c:402
static void tcp_retransmit_timer_force_update(tcp_connection_t *tc)
Definition: tcp.h:1159
#define vnet_buffer(b)
Definition: buffer.h:365
static tcp_connection_t * tcp_connection_get(u32 conn_index, u32 thread_index)
Definition: tcp.h:714
static void tcp_cc_event(tcp_connection_t *tc, tcp_cc_event_t evt)
Definition: tcp.h:1086
void tcp_update_rto(tcp_connection_t *tc)
Definition: tcp_input.c:478
format_function_t format_tcp_sacks
Definition: tcp.h:66
tcp_cc_algorithm_type_e tcp_cc_algo_new_type(const tcp_cc_algorithm_t *vft)
Register new cc algo type.
Definition: tcp.c:105
void tcp_program_ack(tcp_connection_t *tc)
Definition: tcp_output.c:1181
#define tcp_opts_sack_permitted(_to)
Definition: tcp_packet.h:159
static u32 tcp_loss_wnd(const tcp_connection_t *tc)
Definition: tcp.h:943
struct tcp_errors_ tcp_errors_t
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:1938
u32 table_id
Definition: fib_types.api:118
static u64 tcp_cc_get_pacing_rate(tcp_connection_t *tc)
Definition: tcp.h:1093
static void tcp_persist_timer_set(tcp_connection_t *tc)
Definition: tcp.h:1166
static tcp_main_t * vnet_get_tcp_main()
Definition: tcp.h:681
static void tcp_cong_recovery_off(tcp_connection_t *tc)
Definition: tcp.h:480
timer_expiration_handler tcp_timer_delack_handler
void tcp_connection_free(tcp_connection_t *tc)
Definition: tcp.c:325
static_always_inline void clib_spinlock_lock_if_init(clib_spinlock_t *p)
Definition: lock.h:95
static void tcp_retransmit_timer_reset(tcp_connection_t *tc)
Definition: tcp.h:1153
tcp_connection_t * tcp_connection_alloc_w_base(u8 thread_index, tcp_connection_t *base)
Definition: tcp.c:312
static u32 tcp_set_time_now(tcp_worker_ctx_t *wrk)
Definition: tcp.h:1028
void tcp_bt_track_rxt(tcp_connection_t *tc, u32 start, u32 end)
Track a tcp retransmission.
Definition: tcp_bt.c:333
enum tcp_cfg_flag_ tcp_cfg_flags_e
void tcp_bt_flush_samples(tcp_connection_t *tc)
Flush byte tracker samples.
Definition: tcp_bt.c:619
static u8 tcp_timer_is_active(tcp_connection_t *tc, tcp_timers_e timer)
Definition: tcp.h:1207
void tcp_bt_cleanup(tcp_connection_t *tc)
Byte tracker cleanup.
Definition: tcp_bt.c:644
static void tcp_cc_congestion(tcp_connection_t *tc)
Definition: tcp.h:1061
static tcp_connection_t * tcp_listener_get(u32 tli)
Definition: tcp.h:765
vlib_node_registration_t tcp4_established_node
(constructor) VLIB_REGISTER_NODE (tcp4_established_node)
Definition: tcp_input.c:2264
static void tcp_persist_timer_reset(tcp_connection_t *tc)
Definition: tcp.h:1187
#define TCP_EVT(_evt, _args...)
Definition: tcp_debug.h:145
int tcp_session_custom_tx(void *conn, u32 max_burst_size)
Definition: tcp_output.c:2178