FD.io VPP  v20.01-48-g3e0dafb74
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 tsecr_last_ack; /**< Timestamp echoed to us in last healthy ACK */
398  u32 snd_congestion; /**< snd_una_max when congestion is detected */
399  u32 tx_fifo_size; /**< Tx fifo size. Used to constrain cwnd */
400  tcp_cc_algorithm_t *cc_algo; /**< Congestion control algorithm */
401  u8 cc_data[TCP_CC_DATA_SZ]; /**< Congestion control algo private data */
402 
403  u32 fr_occurences; /**< fast-retransmit occurrences RFC4898
404  tcpEStatsStackFastRetran */
405  u32 tr_occurences; /**< timer-retransmit occurrences */
406  u64 bytes_retrans; /**< RFC4898 tcpEStatsPerfOctetsRetrans */
407  u64 segs_retrans; /**< RFC4898 tcpEStatsPerfSegsRetrans*/
408 
409  /* RTT and RTO */
410  u32 rto; /**< Retransmission timeout */
411  u32 rto_boff; /**< Index for RTO backoff */
412  u32 srtt; /**< Smoothed RTT */
413  u32 rttvar; /**< Smoothed mean RTT difference. Approximates variance */
414  u32 rtt_seq; /**< Sequence number for tracked ACK */
415  f64 rtt_ts; /**< Timestamp for tracked ACK */
416  f64 mrtt_us; /**< High precision mrtt from tracked acks */
417 
418  u32 psh_seq; /**< Add psh header for seg that includes this */
419  u32 next_node_index; /**< Can be used to control next node in output */
420  u32 next_node_opaque; /**< Opaque to pass to next node */
421  u32 limited_transmit; /**< snd_nxt when limited transmit starts */
422  u32 sw_if_index; /**< Interface for the connection */
423 
424  /* Delivery rate estimation */
425  u64 delivered; /**< Total bytes delivered to peer */
426  u64 app_limited; /**< Delivered when app-limited detected */
427  f64 delivered_time; /**< Time last bytes were acked */
428  f64 first_tx_time; /**< Send time for recently delivered/sent */
429  tcp_byte_tracker_t *bt; /**< Tx byte tracker */
430 
431  tcp_errors_t errors; /**< Soft connection errors */
432 
433  f64 start_ts; /**< Timestamp when connection initialized */
434  u32 last_fib_check; /**< Last time we checked fib route for peer */
435  u16 mss; /**< Our max seg size that includes options */
436  u32 timestamp_delta; /**< Offset for timestamp */
437  u32 ipv6_flow_label; /**< flow label for ipv6 header */
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  /** Set if csum offloading is enabled */
562  u8 csum_offload;
564  /** Default congestion control algorithm type */
565  tcp_cc_algorithm_type_e cc_algo;
567  /** Min rwnd, as number of snd_mss segments, for update ack to be sent after
568  * a zero rwnd advertisement */
569  u32 rwnd_min_update_ack;
571  /** Delayed ack time (disabled) */
572  u16 delack_time;
574  /** Timer ticks to wait for close from app */
575  u16 closewait_time;
577  /** Timer ticks to wait in time-wait. Also known as 2MSL */
578  u16 timewait_time;
580  /** Timer ticks to wait in fin-wait1 to send fin and rcv fin-ack */
581  u16 finwait1_time;
583  /** Timer ticks to wait in last ack for ack */
584  u16 lastack_time;
586  /** Timer ticks to wait in fin-wait2 for fin */
587  u16 finwait2_time;
589  /** Timer ticks to wait in closing for fin ack */
590  u16 closing_time;
592  /** Timer ticks to wait before cleaning up the connection */
593  u16 cleanup_time;
595  /** Number of preallocated connections */
596  u32 preallocated_connections;
598  /** Number of preallocated half-open connections */
599  u32 preallocated_half_open_connections;
601  /** Vectors of src addresses. Optional unless one needs > 63K active-opens */
602  ip4_address_t *ip4_src_addrs;
603  ip6_address_t *ip6_src_addrs;
605  /** Fault-injection. Debug only */
606  f64 buffer_fail_fraction;
608 
609 typedef struct _tcp_main
610 {
611  /* Per-worker thread tcp connection pools */
613 
614  /* Pool of listeners. */
615  tcp_connection_t *listener_pool;
616 
617  /** Dispatch table by state and flags */
618  tcp_lookup_dispatch_t dispatch_table[TCP_N_STATES][64];
619 
620  u8 log2_tstamp_clocks_per_tick;
621  f64 tstamp_ticks_per_clock;
622 
623  /** per-worker context */
624  tcp_worker_ctx_t *wrk_ctx;
625 
626  /** Pool of half-open connections on which we've sent a SYN */
627  tcp_connection_t *half_open_connections;
628  clib_spinlock_t half_open_lock;
629 
630  /** vlib buffer size */
631  u32 bytes_per_buffer;
632 
633  /** Seed used to generate random iss */
634  tcp_iss_seed_t iss_seed;
635 
636  /** Congestion control algorithms registered */
637  tcp_cc_algorithm_t *cc_algos;
638 
639  /** Hash table of cc algorithms by name */
640  uword *cc_algo_by_name;
641 
642  /** Last cc algo registered */
643  tcp_cc_algorithm_type_e cc_last_type;
644 
645  /** Flag that indicates if stack is on or off */
646  u8 is_enabled;
647 
648  /** Flag that indicates if v4 punting is enabled */
649  u8 punt_unknown4;
650 
651  /** Flag that indicates if v6 punting is enabled */
652  u8 punt_unknown6;
653 
654  /** Rotor for v4 source addresses */
655  u32 last_v4_addr_rotor;
656 
657  /** Rotor for v6 source addresses */
658  u32 last_v6_addr_rotor;
659 
660  /** Protocol configuration */
662 } tcp_main_t;
663 
664 extern tcp_main_t tcp_main;
677 
678 #define tcp_cfg tcp_main.cfg
679 #define tcp_node_index(node_id, is_ip4) \
680  ((is_ip4) ? tcp4_##node_id##_node.index : tcp6_##node_id##_node.index)
681 
684 {
685  return &tcp_main;
686 }
687 
689 tcp_get_worker (u32 thread_index)
690 {
691  return &tcp_main.wrk_ctx[thread_index];
692 }
693 
696 {
697  ASSERT ((signed) b->current_data >= (signed) -VLIB_BUFFER_PRE_DATA_SIZE);
698  return (tcp_header_t *) (b->data + b->current_data
699  + vnet_buffer (b)->tcp.hdr_offset);
700 }
701 
702 #if (VLIB_BUFFER_TRACE_TRAJECTORY)
703 #define tcp_trajectory_add_start(b, start) \
704 { \
705  (*vlib_buffer_trace_trajectory_cb) (b, start); \
706 }
707 #else
708 #define tcp_trajectory_add_start(b, start)
709 #endif
710 
712 
713 void tcp_punt_unknown (vlib_main_t * vm, u8 is_ip4, u8 is_add);
714 
716 tcp_connection_get (u32 conn_index, u32 thread_index)
717 {
718  if (PREDICT_FALSE
719  (pool_is_free_index (tcp_main.connections[thread_index], conn_index)))
720  return 0;
721  return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
722 }
723 
725 tcp_connection_get_if_valid (u32 conn_index, u32 thread_index)
726 {
727  if (tcp_main.connections[thread_index] == 0)
728  return 0;
729  if (pool_is_free_index (tcp_main.connections[thread_index], conn_index))
730  return 0;
731  return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
732 }
733 
736 {
737  return (tcp_connection_t *) tconn;
738 }
739 
740 always_inline void
742 {
743  tc->state = state;
744  TCP_EVT (TCP_EVT_STATE_CHANGE, tc);
745 }
746 
753  tcp_connection_t * base);
757  ip4_address_t * start,
758  ip4_address_t * end, u32 table_id);
760  ip6_address_t * start,
761  ip6_address_t * end, u32 table_id);
762 void tcp_api_reference (void);
763 u8 *format_tcp_connection (u8 * s, va_list * args);
764 u8 *format_tcp_connection_id (u8 * s, va_list * args);
765 
767 tcp_listener_get (u32 tli)
768 {
769  tcp_connection_t *tc = 0;
770  if (!pool_is_free_index (tcp_main.listener_pool, tli))
771  tc = pool_elt_at_index (tcp_main.listener_pool, tli);
772  return tc;
773 }
774 
777 {
778  tcp_connection_t *tc = 0;
779  clib_spinlock_lock_if_init (&tcp_main.half_open_lock);
780  if (!pool_is_free_index (tcp_main.half_open_connections, conn_index))
781  tc = pool_elt_at_index (tcp_main.half_open_connections, conn_index);
782  clib_spinlock_unlock_if_init (&tcp_main.half_open_lock);
783  return tc;
784 }
785 
789  u32 thread_index, u8 is_ip4);
790 void tcp_send_reset (tcp_connection_t * tc);
791 void tcp_send_syn (tcp_connection_t * tc);
793 void tcp_send_fin (tcp_connection_t * tc);
794 void tcp_send_ack (tcp_connection_t * tc);
796 void tcp_update_rto (tcp_connection_t * tc);
799 
803 
804 /*
805  * Rate estimation
806  */
807 
808 /**
809  * Byte tracker initialize
810  *
811  * @param tc connection for which the byte tracker should be allocated and
812  * initialized
813  */
814 void tcp_bt_init (tcp_connection_t * tc);
815 /**
816  * Byte tracker cleanup
817  *
818  * @param tc connection for which the byte tracker should be cleaned up
819  */
820 void tcp_bt_cleanup (tcp_connection_t * tc);
821 /**
822  * Flush byte tracker samples
823  *
824  * @param tc tcp connection for which samples should be flushed
825  */
827 /**
828  * Track a tcp tx burst
829  *
830  * @param tc tcp connection
831  */
833 /**
834  * Track a tcp retransmission
835  *
836  * @param tc tcp connection
837  * @param start start sequence number
838  * @param end end sequence number
839  */
840 void tcp_bt_track_rxt (tcp_connection_t * tc, u32 start, u32 end);
841 /**
842  * Generate a delivery rate sample from recently acked bytes
843  *
844  * @param tc tcp connection
845  * @param rs resulting rate sample
846  */
848  tcp_rate_sample_t * rs);
849 /**
850  * Check if sample to be generated is app limited
851  *
852  * @param tc tcp connection
853  */
855 /**
856  * Check if the byte tracker is in sane state
857  *
858  * Should be used only for testing
859  *
860  * @param bt byte tracker
861  */
863 u8 *format_tcp_bt (u8 * s, va_list * args);
864 
867 {
868  return th->seq_number + tcp_is_syn (th) + tcp_is_fin (th) + len;
869 }
870 
871 /* Modulo arithmetic for TCP sequence numbers */
872 #define seq_lt(_s1, _s2) ((i32)((_s1)-(_s2)) < 0)
873 #define seq_leq(_s1, _s2) ((i32)((_s1)-(_s2)) <= 0)
874 #define seq_gt(_s1, _s2) ((i32)((_s1)-(_s2)) > 0)
875 #define seq_geq(_s1, _s2) ((i32)((_s1)-(_s2)) >= 0)
876 #define seq_max(_s1, _s2) (seq_gt((_s1), (_s2)) ? (_s1) : (_s2))
878 /* Modulo arithmetic for timestamps */
879 #define timestamp_lt(_t1, _t2) ((i32)((_t1)-(_t2)) < 0)
880 #define timestamp_leq(_t1, _t2) ((i32)((_t1)-(_t2)) <= 0)
882 /**
883  * Our estimate of the number of bytes that have left the network
884  */
886 tcp_bytes_out (const tcp_connection_t * tc)
887 {
888  if (tcp_opts_sack_permitted (&tc->rcv_opts))
889  return tc->sack_sb.sacked_bytes + tc->sack_sb.lost_bytes;
890  else
891  return clib_min (tc->rcv_dupacks * tc->snd_mss,
892  tc->snd_nxt - tc->snd_una);
893 }
894 
895 /**
896  * Our estimate of the number of bytes in flight (pipe size)
897  */
900 {
901  int flight_size;
902 
903  flight_size = (int) (tc->snd_nxt - tc->snd_una) - tcp_bytes_out (tc)
904  + tc->snd_rxt_bytes - tc->rxt_delivered;
905 
906  ASSERT (flight_size >= 0);
907 
908  return flight_size;
909 }
910 
911 /**
912  * Initial cwnd as per RFC5681
913  */
916 {
917  if (tcp_cfg.initial_cwnd_multiplier > 0)
918  return tcp_cfg.initial_cwnd_multiplier * tc->snd_mss;
919 
920  if (tc->snd_mss > 2190)
921  return 2 * tc->snd_mss;
922  else if (tc->snd_mss > 1095)
923  return 3 * tc->snd_mss;
924  else
925  return 4 * tc->snd_mss;
926 }
927 
928 /*
929  * Accumulate acked bytes for cwnd increase
930  *
931  * Once threshold bytes are accumulated, snd_mss bytes are added
932  * to the cwnd.
933  */
934 always_inline void
935 tcp_cwnd_accumulate (tcp_connection_t * tc, u32 thresh, u32 bytes)
936 {
937  tc->cwnd_acc_bytes += bytes;
938  if (tc->cwnd_acc_bytes >= thresh)
939  {
940  u32 inc = tc->cwnd_acc_bytes / thresh;
941  tc->cwnd_acc_bytes -= inc * thresh;
942  tc->cwnd += inc * tc->snd_mss;
943  tc->cwnd = clib_min (tc->cwnd, tc->tx_fifo_size);
944  }
945 }
946 
948 tcp_loss_wnd (const tcp_connection_t * tc)
949 {
950  /* Whatever we have in flight + the packet we're about to send */
951  return tcp_flight_size (tc) + tc->snd_mss;
952 }
953 
956 {
957  return clib_min (tc->cwnd, tc->snd_wnd);
958 }
959 
962 {
963  u32 available_wnd = tcp_available_snd_wnd (tc);
964  int flight_size = (int) (tc->snd_nxt - tc->snd_una);
965 
966  if (available_wnd <= flight_size)
967  return 0;
968 
969  return available_wnd - flight_size;
970 }
971 
972 /**
973  * Estimate of how many bytes we can still push into the network
974  */
977 {
978  u32 available_wnd = tcp_available_snd_wnd (tc);
979  u32 flight_size = tcp_flight_size (tc);
980 
981  if (available_wnd <= flight_size)
982  return 0;
983 
984  return available_wnd - flight_size;
985 }
986 
989 {
990  if ((tc->flags & TCP_CONN_FINSNT) && (tc->snd_una_max - tc->snd_una == 1))
991  return 1;
992  return 0;
993 }
994 
997 
999 
1000 /* Made public for unit testing only */
1001 void tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end);
1003 
1005 tcp_time_now (void)
1007  return tcp_main.wrk_ctx[vlib_get_thread_index ()].time_now;
1008 }
1009 
1011 tcp_time_now_w_thread (u32 thread_index)
1013  return tcp_main.wrk_ctx[thread_index].time_now;
1014 }
1015 
1016 /**
1017  * Generate timestamp for tcp connection
1018  */
1022  return (tcp_main.wrk_ctx[tc->c_thread_index].time_now -
1023  tc->timestamp_delta);
1024 }
1025 
1027 tcp_time_now_us (u32 thread_index)
1029  return transport_time_now (thread_index);
1030 }
1031 
1035  wrk->time_now = clib_cpu_time_now () * tcp_main.tstamp_ticks_per_clock;
1036  return wrk->time_now;
1037 }
1038 
1040  vlib_buffer_t * b);
1041 int tcp_session_custom_tx (void *conn, u32 max_burst_size);
1042 
1049  u32 start_bucket);
1050 
1051 always_inline void
1054  tc->cc_algo->rcv_ack (tc, rs);
1055  tc->tsecr_last_ack = tc->rcv_opts.tsecr;
1056 }
1057 
1058 static inline void
1061 {
1062  tc->cc_algo->rcv_cong_ack (tc, ack_type, rs);
1063 }
1064 
1065 static inline void
1068  tc->cc_algo->congestion (tc);
1069 }
1070 
1071 static inline void
1074  tc->cc_algo->loss (tc);
1075 }
1076 
1077 static inline void
1080  tc->cc_algo->recovered (tc);
1081 }
1082 
1083 static inline void
1086  if (tc->cc_algo->undo_recovery)
1087  tc->cc_algo->undo_recovery (tc);
1088 }
1089 
1090 static inline void
1093  if (tc->cc_algo->event)
1094  tc->cc_algo->event (tc, evt);
1095 }
1096 
1097 static inline u64
1100  if (tc->cc_algo->get_pacing_rate)
1101  return tc->cc_algo->get_pacing_rate (tc);
1102 
1103  f64 srtt = clib_min ((f64) tc->srtt * TCP_TICK, tc->mrtt_us);
1104 
1105  /* TODO should constrain to interface's max throughput but
1106  * we don't have link speeds for sw ifs ..*/
1107  return ((f64) tc->cwnd / srtt);
1108 }
1109 
1110 always_inline void
1111 tcp_timer_set (tcp_connection_t * tc, u8 timer_id, u32 interval)
1113  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
1114  ASSERT (tc->timers[timer_id] == TCP_TIMER_HANDLE_INVALID);
1115  tc->timers[timer_id] =
1116  tw_timer_start_16t_2w_512sl (&tcp_main.
1117  wrk_ctx[tc->c_thread_index].timer_wheel,
1118  tc->c_c_index, timer_id, interval);
1119 }
1120 
1121 always_inline void
1122 tcp_timer_reset (tcp_connection_t * tc, u8 timer_id)
1124  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
1125  if (tc->timers[timer_id] == TCP_TIMER_HANDLE_INVALID)
1126  return;
1127 
1128  tw_timer_stop_16t_2w_512sl (&tcp_main.
1129  wrk_ctx[tc->c_thread_index].timer_wheel,
1130  tc->timers[timer_id]);
1131  tc->timers[timer_id] = TCP_TIMER_HANDLE_INVALID;
1132 }
1133 
1134 always_inline void
1135 tcp_timer_update (tcp_connection_t * tc, u8 timer_id, u32 interval)
1137  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
1138  if (tc->timers[timer_id] != TCP_TIMER_HANDLE_INVALID)
1139  tw_timer_update_16t_2w_512sl (&tcp_main.
1140  wrk_ctx[tc->c_thread_index].timer_wheel,
1141  tc->timers[timer_id], interval);
1142  else
1143  tc->timers[timer_id] =
1144  tw_timer_start_16t_2w_512sl (&tcp_main.
1145  wrk_ctx[tc->c_thread_index].timer_wheel,
1146  tc->c_c_index, timer_id, interval);
1147 }
1148 
1149 always_inline void
1152  ASSERT (tc->snd_una != tc->snd_una_max);
1153  tcp_timer_set (tc, TCP_TIMER_RETRANSMIT,
1154  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
1155 }
1156 
1157 always_inline void
1160  tcp_timer_reset (tc, TCP_TIMER_RETRANSMIT);
1161 }
1162 
1163 always_inline void
1166  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
1167  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
1168 }
1169 
1170 always_inline void
1173  /* Reuse RTO. It's backed off in handler */
1174  tcp_timer_set (tc, TCP_TIMER_PERSIST,
1175  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
1176 }
1177 
1178 always_inline void
1181  u32 interval;
1182 
1183  if (seq_leq (tc->snd_una, tc->snd_congestion + tc->burst_acked))
1184  interval = 1;
1185  else
1186  interval = clib_max (tc->rto * TCP_TO_TIMER_TICK, 1);
1187 
1188  tcp_timer_update (tc, TCP_TIMER_PERSIST, interval);
1189 }
1190 
1191 always_inline void
1194  tcp_timer_reset (tc, TCP_TIMER_PERSIST);
1195 }
1196 
1197 always_inline void
1200  if (tc->snd_una == tc->snd_nxt)
1201  {
1203  if (tc->snd_wnd < tc->snd_mss)
1205  }
1206  else
1207  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
1208  clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
1209 }
1210 
1214  return tc->timers[timer] != TCP_TIMER_HANDLE_INVALID;
1215 }
1216 
1217 #define tcp_validate_txf_size(_tc, _a) \
1218  ASSERT(_tc->state != TCP_STATE_ESTABLISHED \
1219  || transport_max_tx_dequeue (&_tc->connection) >= _a)
1220 
1221 void tcp_rcv_sacks (tcp_connection_t * tc, u32 ack);
1222 u8 *tcp_scoreboard_replay (u8 * s, tcp_connection_t * tc, u8 verbose);
1223 
1224 /**
1225  * Register exiting cc algo type
1226  */
1228  const tcp_cc_algorithm_t * vft);
1229 
1230 /**
1231  * Register new cc algo type
1232  */
1235 
1236 static inline void *
1239  return (void *) tc->cc_data;
1240 }
1241 
1243  tcp_rate_sample_t * rs);
1244 
1245 /**
1246  * Push TCP header to buffer
1247  *
1248  * @param vm - vlib_main
1249  * @param b - buffer to write the header to
1250  * @param sp_net - source port net order
1251  * @param dp_net - destination port net order
1252  * @param seq - sequence number net order
1253  * @param ack - ack number net order
1254  * @param tcp_hdr_opts_len - header and options length in bytes
1255  * @param flags - header flags
1256  * @param wnd - window size
1257  *
1258  * @return - pointer to start of TCP header
1259  */
1260 always_inline void *
1262  u32 ack, u8 tcp_hdr_opts_len, u8 flags,
1263  u16 wnd)
1264 {
1265  tcp_header_t *th;
1266 
1267  th = vlib_buffer_push_uninit (b, tcp_hdr_opts_len);
1268 
1269  th->src_port = sp;
1270  th->dst_port = dp;
1271  th->seq_number = seq;
1272  th->ack_number = ack;
1273  th->data_offset_and_reserved = (tcp_hdr_opts_len >> 2) << 4;
1274  th->flags = flags;
1275  th->window = wnd;
1276  th->checksum = 0;
1277  th->urgent_pointer = 0;
1278  vnet_buffer (b)->l4_hdr_offset = (u8 *) th - b->data;
1279  b->flags |= VNET_BUFFER_F_L4_HDR_OFFSET_VALID;
1280  return th;
1281 }
1282 
1283 /**
1284  * Push TCP header to buffer
1285  *
1286  * @param b - buffer to write the header to
1287  * @param sp_net - source port net order
1288  * @param dp_net - destination port net order
1289  * @param seq - sequence number host order
1290  * @param ack - ack number host order
1291  * @param tcp_hdr_opts_len - header and options length in bytes
1292  * @param flags - header flags
1293  * @param wnd - window size
1294  *
1295  * @return - pointer to start of TCP header
1296  */
1297 always_inline void *
1298 vlib_buffer_push_tcp (vlib_buffer_t * b, u16 sp_net, u16 dp_net, u32 seq,
1299  u32 ack, u8 tcp_hdr_opts_len, u8 flags, u16 wnd)
1300 {
1301  return vlib_buffer_push_tcp_net_order (b, sp_net, dp_net,
1302  clib_host_to_net_u32 (seq),
1303  clib_host_to_net_u32 (ack),
1304  tcp_hdr_opts_len, flags,
1305  clib_host_to_net_u16 (wnd));
1306 }
1307 
1308 #endif /* _vnet_tcp_h_ */
1309 
1310 /*
1311  * fd.io coding-style-patch-verification: ON
1312  *
1313  * Local Variables:
1314  * eval: (c-set-style "gnu")
1315  * End:
1316  */
tcp_main_t tcp_main
Definition: tcp.c:29
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:3392
void tcp_program_retransmit(tcp_connection_t *tc)
Definition: tcp_output.c:1198
static vlib_cli_command_t trace
(constructor) VLIB_CLI_COMMAND (trace)
Definition: vlib_api_cli.c:898
#define CLIB_CACHE_LINE_ALIGN_MARK(mark)
Definition: cache.h:60
void tcp_send_window_update_ack(tcp_connection_t *tc)
Send window update ack.
Definition: tcp_output.c:1229
#define clib_min(x, y)
Definition: clib.h:295
void scoreboard_clear(sack_scoreboard_t *sb)
Definition: tcp_input.c:927
static f64 tcp_time_now_us(u32 thread_index)
Definition: tcp.h:1028
static void tcp_retransmit_timer_set(tcp_connection_t *tc)
Definition: tcp.h:1151
static clib_time_type_t transport_time_now(u32 thread_index)
Definition: session.h:513
#define seq_leq(_s1, _s2)
Definition: tcp.h:874
void tcp_make_fin(tcp_connection_t *tc, vlib_buffer_t *b)
Convert buffer to FIN-ACK.
Definition: tcp_output.c:569
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:519
#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:2456
void scoreboard_init(sack_scoreboard_t *sb)
Definition: tcp_input.c:919
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:887
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:726
void tcp_connection_del(tcp_connection_t *tc)
Connection removal.
Definition: tcp.c:291
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:1743
enum tcp_bts_flags_ tcp_bts_flags_t
u32 tcp_snd_space(tcp_connection_t *tc)
Definition: tcp.c:1322
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:777
sack_scoreboard_hole_t * scoreboard_last_hole(sack_scoreboard_t *sb)
Definition: tcp_input.c:681
sack_scoreboard_hole_t * scoreboard_next_hole(sack_scoreboard_t *sb, sack_scoreboard_hole_t *hole)
Definition: tcp_input.c:657
#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:2476
u8 data[0]
Packet data.
Definition: buffer.h:181
static u64 clib_cpu_time_now(void)
Definition: time.h:75
timer_expiration_handler tcp_timer_retransmit_handler
u8 * format_tcp_scoreboard(u8 *s, va_list *args)
Definition: tcp.c:1181
u32 tcp_session_push_header(transport_connection_t *tconn, vlib_buffer_t *b)
Definition: tcp_output.c:1129
struct _tcp_lookup_dispatch tcp_lookup_dispatch_t
void tcp_connection_tx_pacer_update(tcp_connection_t *tc)
Definition: tcp.c:1396
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:390
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:977
static u32 tcp_available_snd_wnd(const tcp_connection_t *tc)
Definition: tcp.h:956
static tcp_connection_t * tcp_get_connection_from_transport(transport_connection_t *tconn)
Definition: tcp.h:736
void tcp_connection_cleanup(tcp_connection_t *tc)
Cleans up connection state.
Definition: tcp.c:238
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:342
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:2723
#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:2172
#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:742
#define tcp_cfg
Definition: tcp.h:679
vl_api_interface_index_t sw_if_index
Definition: gre.api:59
sack_scoreboard_hole_t * scoreboard_get_hole(sack_scoreboard_t *sb, u32 index)
Definition: tcp_input.c:649
struct tcp_bt_sample_ tcp_bt_sample_t
static u32 tcp_available_output_snd_space(const tcp_connection_t *tc)
Definition: tcp.h:962
static void tcp_cc_loss(tcp_connection_t *tc)
Definition: tcp.h:1073
static tcp_header_t * tcp_buffer_hdr(vlib_buffer_t *b)
Definition: tcp.h:696
enum _tcp_state tcp_state_t
timer_expiration_handler tcp_timer_retransmit_syn_handler
static u32 tcp_time_now(void)
Definition: tcp.h:1006
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:845
void tcp_api_reference(void)
Definition: tcp_api.c:119
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:1112
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:519
void tcp_bt_init(tcp_connection_t *tc)
Byte tracker initialize.
Definition: tcp_bt.c:650
timer_expiration_handler tcp_timer_persist_handler
void tcp_send_ack(tcp_connection_t *tc)
Definition: tcp_output.c:1157
vlib_node_registration_t tcp6_syn_sent_node
(constructor) VLIB_REGISTER_NODE (tcp6_syn_sent_node)
Definition: tcp_input.c:2742
u32 tcp_sack_list_bytes(tcp_connection_t *tc)
Definition: tcp_input.c:1792
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:1262
int tcp_fastrecovery_prr_snd_space(tcp_connection_t *tc)
Estimate send space using proportional rate reduction (RFC6937)
Definition: tcp_output.c:1798
void tcp_rcv_sacks(tcp_connection_t *tc, u32 ack)
Definition: tcp_input.c:981
clib_error_t * vnet_tcp_enable_disable(vlib_main_t *vm, u8 is_en)
Definition: tcp.c:1636
void tcp_send_syn(tcp_connection_t *tc)
Send SYN.
Definition: tcp_output.c:919
static void * tcp_cc_data(tcp_connection_t *tc)
Definition: tcp.h:1238
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:665
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:1060
static u32 tcp_flight_size(const tcp_connection_t *tc)
Our estimate of the number of bytes in flight (pipe size)
Definition: tcp.h:900
#define PREDICT_FALSE(x)
Definition: clib.h:111
#define always_inline
Definition: ipsec.h:28
tcp_connection_t * tcp_connection_alloc(u8 thread_index)
Definition: tcp.c:298
void tcp_connection_tx_pacer_reset(tcp_connection_t *tc, u32 window, u32 start_bucket)
Definition: tcp.c:1409
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:86
vlib_main_t * vm
Definition: in2out_ed.c:1810
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:936
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:91
tcp_bts_flags_
Definition: tcp.h:255
static void tcp_timer_reset(tcp_connection_t *tc, u8 timer_id)
Definition: tcp.h:1123
vlib_node_registration_t tcp4_listen_node
(constructor) VLIB_REGISTER_NODE (tcp4_listen_node)
Definition: tcp_input.c:3373
void scoreboard_init_rxt(sack_scoreboard_t *sb, u32 snd_una)
Definition: tcp_input.c:905
tcp_connection_flag_bits_
Definition: tcp.h:146
u8 is_ip4
Definition: lisp_gpe.api:232
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:1299
u32 flags
Definition: vhost_user.h:141
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:1862
void scoreboard_clear_reneging(sack_scoreboard_t *sb, u32 start, u32 end)
Definition: tcp_input.c:946
static_always_inline uword vlib_get_thread_index(void)
Definition: threads.h:218
void tcp_program_dupack(tcp_connection_t *tc)
Definition: tcp_output.c:1186
void tcp_send_reset(tcp_connection_t *tc)
Build and set reset packet for connection.
Definition: tcp_output.c:863
static u32 tcp_tstamp(tcp_connection_t *tc)
Generate timestamp for tcp connection.
Definition: tcp.h:1021
void tcp_punt_unknown(vlib_main_t *vm, u8 is_ip4, u8 is_add)
Definition: tcp.c:1654
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:1085
enum _tcp_timers tcp_timers_e
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:210
u8 * format_tcp_bt(u8 *s, va_list *args)
Definition: tcp_bt.c:676
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:502
void tcp_make_synack(tcp_connection_t *ts, vlib_buffer_t *b)
Convert buffer to SYN-ACK.
Definition: tcp_output.c:604
#define pool_is_free_index(P, I)
Use free bitmap to query whether given index is free.
Definition: pool.h:284
format_function_t format_tcp_rcv_sacks
Definition: tcp.h:67
u8 * format_tcp_connection_id(u8 *s, va_list *args)
Definition: tcp.c:1040
vlib_node_registration_t tcp6_input_node
(constructor) VLIB_REGISTER_NODE (tcp6_input_node)
Definition: tcp_input.c:3806
fib_node_index_t tcp_lookup_rmt_in_fib(tcp_connection_t *tc)
string name[64]
Definition: ip.api:44
void tcp_send_synack(tcp_connection_t *tc)
Definition: tcp_output.c:954
static void tcp_timer_update(tcp_connection_t *tc, u8 timer_id, u32 interval)
Definition: tcp.h:1136
#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:997
void tcp_connection_init_vars(tcp_connection_t *tc)
Initialize tcp connection variables.
Definition: tcp.c:726
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:582
vlib_node_registration_t tcp4_rcv_process_node
(constructor) VLIB_REGISTER_NODE (tcp4_rcv_process_node)
Definition: tcp_input.c:3156
u8 * format_tcp_connection(u8 *s, va_list *args)
Definition: tcp.c:1064
static u32 tcp_end_seq(tcp_header_t *th, u32 len)
Definition: tcp.h:867
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:2263
#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:288
static u32 tcp_time_now_w_thread(u32 thread_index)
Definition: tcp.h:1012
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:1180
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:778
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:3786
tcp_cc_algorithm_t * tcp_cc_algo_get(tcp_cc_algorithm_type_e type)
Definition: tcp.c:97
static u32 tcp_initial_cwnd(const tcp_connection_t *tc)
Initial cwnd as per RFC5681.
Definition: tcp.h:916
vlib_node_registration_t tcp6_rcv_process_node
(constructor) VLIB_REGISTER_NODE (tcp6_rcv_process_node)
Definition: tcp_input.c:3175
#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:1007
#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:989
static void tcp_cc_rcv_ack(tcp_connection_t *tc, tcp_rate_sample_t *rs)
Definition: tcp.h:1053
sack_scoreboard_hole_t * scoreboard_first_hole(sack_scoreboard_t *sb)
Definition: tcp_input.c:673
static tcp_worker_ctx_t * tcp_get_worker(u32 thread_index)
Definition: tcp.h:690
static void tcp_retransmit_timer_update(tcp_connection_t *tc)
Definition: tcp.h:1199
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:1079
void tcp_init_snd_vars(tcp_connection_t *tc)
Initialize connection send variables.
Definition: tcp.c:691
struct tcp_configuration_ tcp_configuration_t
void tcp_connection_close(tcp_connection_t *tc)
Begin connection closing procedure.
Definition: tcp.c:401
static void tcp_retransmit_timer_force_update(tcp_connection_t *tc)
Definition: tcp.h:1165
#define vnet_buffer(b)
Definition: buffer.h:408
static tcp_connection_t * tcp_connection_get(u32 conn_index, u32 thread_index)
Definition: tcp.h:717
static void tcp_cc_event(tcp_connection_t *tc, tcp_cc_event_t evt)
Definition: tcp.h:1092
void tcp_update_rto(tcp_connection_t *tc)
Definition: tcp_input.c:478
vl_api_dhcp_client_state_t state
Definition: dhcp.api:201
format_function_t format_tcp_sacks
Definition: tcp.h:66
f64 end
end of the time range
Definition: mactime.api:44
tcp_cc_algorithm_type_e tcp_cc_algo_new_type(const tcp_cc_algorithm_t *vft)
Register new cc algo type.
Definition: tcp.c:104
void tcp_program_ack(tcp_connection_t *tc)
Definition: tcp_output.c:1176
#define tcp_opts_sack_permitted(_to)
Definition: tcp_packet.h:159
static u32 tcp_loss_wnd(const tcp_connection_t *tc)
Definition: tcp.h:949
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:1949
u32 table_id
Definition: fib_types.api:118
static u64 tcp_cc_get_pacing_rate(tcp_connection_t *tc)
Definition: tcp.h:1099
static void tcp_persist_timer_set(tcp_connection_t *tc)
Definition: tcp.h:1172
static tcp_main_t * vnet_get_tcp_main()
Definition: tcp.h:684
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:324
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:1159
tcp_connection_t * tcp_connection_alloc_w_base(u8 thread_index, tcp_connection_t *base)
Definition: tcp.c:311
static u32 tcp_set_time_now(tcp_worker_ctx_t *wrk)
Definition: tcp.h:1034
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:614
static u8 tcp_timer_is_active(tcp_connection_t *tc, tcp_timers_e timer)
Definition: tcp.h:1213
void tcp_bt_cleanup(tcp_connection_t *tc)
Byte tracker cleanup.
Definition: tcp_bt.c:639
static void tcp_cc_congestion(tcp_connection_t *tc)
Definition: tcp.h:1067
static tcp_connection_t * tcp_listener_get(u32 tli)
Definition: tcp.h:768
vlib_node_registration_t tcp4_established_node
(constructor) VLIB_REGISTER_NODE (tcp4_established_node)
Definition: tcp_input.c:2244
static void tcp_persist_timer_reset(tcp_connection_t *tc)
Definition: tcp.h:1193
#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:2157