tcp_output.c revision 8a047ed7
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#include <vnet/tcp/tcp.h>
17#include <math.h>
18
19typedef enum _tcp_output_next
20{
21  TCP_OUTPUT_NEXT_DROP,
22  TCP_OUTPUT_NEXT_IP_LOOKUP,
23  TCP_OUTPUT_NEXT_IP_REWRITE,
24  TCP_OUTPUT_NEXT_IP_ARP,
25  TCP_OUTPUT_N_NEXT
26} tcp_output_next_t;
27
28#define foreach_tcp4_output_next              	\
29  _ (DROP, "error-drop")                        \
30  _ (IP_LOOKUP, "ip4-lookup")			\
31  _ (IP_REWRITE, "ip4-rewrite")			\
32  _ (IP_ARP, "ip4-arp")
33
34#define foreach_tcp6_output_next              	\
35  _ (DROP, "error-drop")                        \
36  _ (IP_LOOKUP, "ip6-lookup")			\
37  _ (IP_REWRITE, "ip6-rewrite")			\
38  _ (IP_ARP, "ip6-discover-neighbor")
39
40static char *tcp_error_strings[] = {
41#define tcp_error(n,s) s,
42#include <vnet/tcp/tcp_error.def>
43#undef tcp_error
44};
45
46typedef struct
47{
48  tcp_header_t tcp_header;
49  tcp_connection_t tcp_connection;
50} tcp_tx_trace_t;
51
52static u8 *
53format_tcp_tx_trace (u8 * s, va_list * args)
54{
55  CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *);
56  CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *);
57  tcp_tx_trace_t *t = va_arg (*args, tcp_tx_trace_t *);
58  u32 indent = format_get_indent (s);
59
60  s = format (s, "%U\n%U%U",
61	      format_tcp_header, &t->tcp_header, 128,
62	      format_white_space, indent,
63	      format_tcp_connection, &t->tcp_connection, 1);
64
65  return s;
66}
67
68#ifndef CLIB_MARCH_VARIANT
69static u8
70tcp_window_compute_scale (u32 window)
71{
72  u8 wnd_scale = 0;
73  while (wnd_scale < TCP_MAX_WND_SCALE && (window >> wnd_scale) > TCP_WND_MAX)
74    wnd_scale++;
75  return wnd_scale;
76}
77
78/**
79 * TCP's initial window
80 */
81always_inline u32
82tcp_initial_wnd_unscaled (tcp_connection_t * tc)
83{
84  /* RFC 6928 recommends the value lower. However at the time our connections
85   * are initialized, fifos may not be allocated. Therefore, advertise the
86   * smallest possible unscaled window size and update once fifos are
87   * assigned to the session.
88   */
89  /*
90     tcp_update_rcv_mss (tc);
91     TCP_IW_N_SEGMENTS * tc->mss;
92   */
93  return tcp_cfg.min_rx_fifo;
94}
95
96/**
97 * Compute initial window and scale factor. As per RFC1323, window field in
98 * SYN and SYN-ACK segments is never scaled.
99 */
100u32
101tcp_initial_window_to_advertise (tcp_connection_t * tc)
102{
103  /* Compute rcv wscale only if peer advertised support for it */
104  if (tc->state != TCP_STATE_SYN_RCVD || tcp_opts_wscale (&tc->rcv_opts))
105    tc->rcv_wscale = tcp_window_compute_scale (tcp_cfg.max_rx_fifo);
106
107  tc->rcv_wnd = tcp_initial_wnd_unscaled (tc);
108
109  return clib_min (tc->rcv_wnd, TCP_WND_MAX);
110}
111
112static inline void
113tcp_update_rcv_wnd (tcp_connection_t * tc)
114{
115  u32 available_space, wnd;
116  i32 observed_wnd;
117
118  ASSERT (tc->rcv_opts.mss < transport_rx_fifo_size (&tc->connection));
119
120  /*
121   * Figure out how much space we have available
122   */
123  available_space = transport_max_rx_enqueue (&tc->connection);
124  if (PREDICT_FALSE (available_space < tc->rcv_opts.mss))
125    {
126      tc->rcv_wnd = 0;
127      return;
128    }
129
130  /*
131   * Use the above and what we know about what we've previously advertised
132   * to compute the new window
133   */
134  observed_wnd = (i32) tc->rcv_wnd - (tc->rcv_nxt - tc->rcv_las);
135
136  /* Bad. Thou shalt not shrink */
137  if (PREDICT_FALSE ((i32) available_space < observed_wnd))
138    {
139      wnd = clib_max (observed_wnd, 0);
140      TCP_EVT (TCP_EVT_RCV_WND_SHRUNK, tc, observed_wnd, available_space);
141    }
142  else
143    {
144      wnd = available_space;
145    }
146
147  /* Make sure we have a multiple of rcv_wscale */
148  if (wnd && tc->rcv_wscale)
149    {
150      wnd &= ~((1 << tc->rcv_wscale) - 1);
151      if (wnd == 0)
152	wnd = 1 << tc->rcv_wscale;
153    }
154
155  tc->rcv_wnd = clib_min (wnd, TCP_WND_MAX << tc->rcv_wscale);
156}
157
158/**
159 * Compute and return window to advertise, scaled as per RFC1323
160 */
161static inline u32
162tcp_window_to_advertise (tcp_connection_t * tc, tcp_state_t state)
163{
164  if (state < TCP_STATE_ESTABLISHED)
165    return tcp_initial_window_to_advertise (tc);
166
167  tcp_update_rcv_wnd (tc);
168  return tc->rcv_wnd >> tc->rcv_wscale;
169}
170
171/**
172 * Write TCP options to segment.
173 */
174static u32
175tcp_options_write (u8 * data, tcp_options_t * opts)
176{
177  u32 opts_len = 0;
178  u32 buf, seq_len = 4;
179
180  if (tcp_opts_mss (opts))
181    {
182      *data++ = TCP_OPTION_MSS;
183      *data++ = TCP_OPTION_LEN_MSS;
184      buf = clib_host_to_net_u16 (opts->mss);
185      clib_memcpy_fast (data, &buf, sizeof (opts->mss));
186      data += sizeof (opts->mss);
187      opts_len += TCP_OPTION_LEN_MSS;
188    }
189
190  if (tcp_opts_wscale (opts))
191    {
192      *data++ = TCP_OPTION_WINDOW_SCALE;
193      *data++ = TCP_OPTION_LEN_WINDOW_SCALE;
194      *data++ = opts->wscale;
195      opts_len += TCP_OPTION_LEN_WINDOW_SCALE;
196    }
197
198  if (tcp_opts_sack_permitted (opts))
199    {
200      *data++ = TCP_OPTION_SACK_PERMITTED;
201      *data++ = TCP_OPTION_LEN_SACK_PERMITTED;
202      opts_len += TCP_OPTION_LEN_SACK_PERMITTED;
203    }
204
205  if (tcp_opts_tstamp (opts))
206    {
207      *data++ = TCP_OPTION_TIMESTAMP;
208      *data++ = TCP_OPTION_LEN_TIMESTAMP;
209      buf = clib_host_to_net_u32 (opts->tsval);
210      clib_memcpy_fast (data, &buf, sizeof (opts->tsval));
211      data += sizeof (opts->tsval);
212      buf = clib_host_to_net_u32 (opts->tsecr);
213      clib_memcpy_fast (data, &buf, sizeof (opts->tsecr));
214      data += sizeof (opts->tsecr);
215      opts_len += TCP_OPTION_LEN_TIMESTAMP;
216    }
217
218  if (tcp_opts_sack (opts))
219    {
220      int i;
221
222      if (opts->n_sack_blocks != 0)
223	{
224	  *data++ = TCP_OPTION_SACK_BLOCK;
225	  *data++ = 2 + opts->n_sack_blocks * TCP_OPTION_LEN_SACK_BLOCK;
226	  for (i = 0; i < opts->n_sack_blocks; i++)
227	    {
228	      buf = clib_host_to_net_u32 (opts->sacks[i].start);
229	      clib_memcpy_fast (data, &buf, seq_len);
230	      data += seq_len;
231	      buf = clib_host_to_net_u32 (opts->sacks[i].end);
232	      clib_memcpy_fast (data, &buf, seq_len);
233	      data += seq_len;
234	    }
235	  opts_len += 2 + opts->n_sack_blocks * TCP_OPTION_LEN_SACK_BLOCK;
236	}
237    }
238
239  /* Terminate TCP options */
240  if (opts_len % 4)
241    {
242      *data++ = TCP_OPTION_EOL;
243      opts_len += TCP_OPTION_LEN_EOL;
244    }
245
246  /* Pad with zeroes to a u32 boundary */
247  while (opts_len % 4)
248    {
249      *data++ = TCP_OPTION_NOOP;
250      opts_len += TCP_OPTION_LEN_NOOP;
251    }
252  return opts_len;
253}
254
255static int
256tcp_make_syn_options (tcp_connection_t * tc, tcp_options_t * opts)
257{
258  u8 len = 0;
259
260  opts->flags |= TCP_OPTS_FLAG_MSS;
261  opts->mss = tc->mss;
262  len += TCP_OPTION_LEN_MSS;
263
264  opts->flags |= TCP_OPTS_FLAG_WSCALE;
265  opts->wscale = tc->rcv_wscale;
266  len += TCP_OPTION_LEN_WINDOW_SCALE;
267
268  opts->flags |= TCP_OPTS_FLAG_TSTAMP;
269  opts->tsval = tcp_time_now ();
270  opts->tsecr = 0;
271  len += TCP_OPTION_LEN_TIMESTAMP;
272
273  if (TCP_USE_SACKS)
274    {
275      opts->flags |= TCP_OPTS_FLAG_SACK_PERMITTED;
276      len += TCP_OPTION_LEN_SACK_PERMITTED;
277    }
278
279  /* Align to needed boundary */
280  len += (TCP_OPTS_ALIGN - len % TCP_OPTS_ALIGN) % TCP_OPTS_ALIGN;
281  return len;
282}
283
284static int
285tcp_make_synack_options (tcp_connection_t * tc, tcp_options_t * opts)
286{
287  u8 len = 0;
288
289  opts->flags |= TCP_OPTS_FLAG_MSS;
290  opts->mss = tc->mss;
291  len += TCP_OPTION_LEN_MSS;
292
293  if (tcp_opts_wscale (&tc->rcv_opts))
294    {
295      opts->flags |= TCP_OPTS_FLAG_WSCALE;
296      opts->wscale = tc->rcv_wscale;
297      len += TCP_OPTION_LEN_WINDOW_SCALE;
298    }
299
300  if (tcp_opts_tstamp (&tc->rcv_opts))
301    {
302      opts->flags |= TCP_OPTS_FLAG_TSTAMP;
303      opts->tsval = tcp_time_now ();
304      opts->tsecr = tc->tsval_recent;
305      len += TCP_OPTION_LEN_TIMESTAMP;
306    }
307
308  if (tcp_opts_sack_permitted (&tc->rcv_opts))
309    {
310      opts->flags |= TCP_OPTS_FLAG_SACK_PERMITTED;
311      len += TCP_OPTION_LEN_SACK_PERMITTED;
312    }
313
314  /* Align to needed boundary */
315  len += (TCP_OPTS_ALIGN - len % TCP_OPTS_ALIGN) % TCP_OPTS_ALIGN;
316  return len;
317}
318
319static int
320tcp_make_established_options (tcp_connection_t * tc, tcp_options_t * opts)
321{
322  u8 len = 0;
323
324  opts->flags = 0;
325
326  if (tcp_opts_tstamp (&tc->rcv_opts))
327    {
328      opts->flags |= TCP_OPTS_FLAG_TSTAMP;
329      opts->tsval = tcp_tstamp (tc);
330      opts->tsecr = tc->tsval_recent;
331      len += TCP_OPTION_LEN_TIMESTAMP;
332    }
333  if (tcp_opts_sack_permitted (&tc->rcv_opts))
334    {
335      if (vec_len (tc->snd_sacks))
336	{
337	  opts->flags |= TCP_OPTS_FLAG_SACK;
338	  if (tc->snd_sack_pos >= vec_len (tc->snd_sacks))
339	    tc->snd_sack_pos = 0;
340	  opts->sacks = &tc->snd_sacks[tc->snd_sack_pos];
341	  opts->n_sack_blocks = vec_len (tc->snd_sacks) - tc->snd_sack_pos;
342	  opts->n_sack_blocks = clib_min (opts->n_sack_blocks,
343					  TCP_OPTS_MAX_SACK_BLOCKS);
344	  tc->snd_sack_pos += opts->n_sack_blocks;
345	  len += 2 + TCP_OPTION_LEN_SACK_BLOCK * opts->n_sack_blocks;
346	}
347    }
348
349  /* Align to needed boundary */
350  len += (TCP_OPTS_ALIGN - len % TCP_OPTS_ALIGN) % TCP_OPTS_ALIGN;
351  return len;
352}
353
354always_inline int
355tcp_make_options (tcp_connection_t * tc, tcp_options_t * opts,
356		  tcp_state_t state)
357{
358  switch (state)
359    {
360    case TCP_STATE_ESTABLISHED:
361    case TCP_STATE_CLOSE_WAIT:
362    case TCP_STATE_FIN_WAIT_1:
363    case TCP_STATE_LAST_ACK:
364    case TCP_STATE_CLOSING:
365    case TCP_STATE_FIN_WAIT_2:
366    case TCP_STATE_TIME_WAIT:
367    case TCP_STATE_CLOSED:
368      return tcp_make_established_options (tc, opts);
369    case TCP_STATE_SYN_RCVD:
370      return tcp_make_synack_options (tc, opts);
371    case TCP_STATE_SYN_SENT:
372      return tcp_make_syn_options (tc, opts);
373    default:
374      clib_warning ("State not handled! %d", state);
375      return 0;
376    }
377}
378
379/**
380 * Update burst send vars
381 *
382 * - Updates snd_mss to reflect the effective segment size that we can send
383 * by taking into account all TCP options, including SACKs.
384 * - Cache 'on the wire' options for reuse
385 * - Updates receive window which can be reused for a burst.
386 *
387 * This should *only* be called when doing bursts
388 */
389void
390tcp_update_burst_snd_vars (tcp_connection_t * tc)
391{
392  tcp_main_t *tm = &tcp_main;
393
394  /* Compute options to be used for connection. These may be reused when
395   * sending data or to compute the effective mss (snd_mss) */
396  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts,
397				       TCP_STATE_ESTABLISHED);
398
399  /* XXX check if MTU has been updated */
400  tc->snd_mss = clib_min (tc->mss, tc->rcv_opts.mss) - tc->snd_opts_len;
401  ASSERT (tc->snd_mss > 0);
402
403  tcp_options_write (tm->wrk_ctx[tc->c_thread_index].cached_opts,
404		     &tc->snd_opts);
405
406  tcp_update_rcv_wnd (tc);
407
408  if (tc->flags & TCP_CONN_RATE_SAMPLE)
409    tc->flags |= TCP_CONN_TRACK_BURST;
410
411  if (tc->snd_una == tc->snd_nxt)
412    tcp_cc_event (tc, TCP_CC_EVT_START_TX);
413}
414
415#endif /* CLIB_MARCH_VARIANT */
416
417static void *
418tcp_reuse_buffer (vlib_main_t * vm, vlib_buffer_t * b)
419{
420  if (b->flags & VLIB_BUFFER_NEXT_PRESENT)
421    vlib_buffer_free_one (vm, b->next_buffer);
422  /* Zero all flags but free list index and trace flag */
423  b->flags &= VLIB_BUFFER_NEXT_PRESENT - 1;
424  b->current_data = 0;
425  b->current_length = 0;
426  b->total_length_not_including_first_buffer = 0;
427  vnet_buffer (b)->tcp.flags = 0;
428
429  /* Leave enough space for headers */
430  return vlib_buffer_make_headroom (b, TRANSPORT_MAX_HDRS_LEN);
431}
432
433#ifndef CLIB_MARCH_VARIANT
434static void *
435tcp_init_buffer (vlib_main_t * vm, vlib_buffer_t * b)
436{
437  ASSERT ((b->flags & VLIB_BUFFER_NEXT_PRESENT) == 0);
438  b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
439  b->total_length_not_including_first_buffer = 0;
440  b->current_data = 0;
441  vnet_buffer (b)->tcp.flags = 0;
442  VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b);
443  /* Leave enough space for headers */
444  return vlib_buffer_make_headroom (b, TRANSPORT_MAX_HDRS_LEN);
445}
446
447/**
448 * Prepare ACK
449 */
450static inline void
451tcp_make_ack_i (tcp_connection_t * tc, vlib_buffer_t * b, tcp_state_t state,
452		u8 flags)
453{
454  tcp_options_t _snd_opts, *snd_opts = &_snd_opts;
455  u8 tcp_opts_len, tcp_hdr_opts_len;
456  tcp_header_t *th;
457  u16 wnd;
458
459  wnd = tcp_window_to_advertise (tc, state);
460
461  /* Make and write options */
462  tcp_opts_len = tcp_make_established_options (tc, snd_opts);
463  tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
464
465  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->snd_nxt,
466			     tc->rcv_nxt, tcp_hdr_opts_len, flags, wnd);
467
468  tcp_options_write ((u8 *) (th + 1), snd_opts);
469  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
470
471  if (wnd == 0)
472    tcp_zero_rwnd_sent_on (tc);
473  else
474    tcp_zero_rwnd_sent_off (tc);
475}
476
477/**
478 * Convert buffer to ACK
479 */
480static inline void
481tcp_make_ack (tcp_connection_t * tc, vlib_buffer_t * b)
482{
483  tcp_make_ack_i (tc, b, TCP_STATE_ESTABLISHED, TCP_FLAG_ACK);
484  TCP_EVT (TCP_EVT_ACK_SENT, tc);
485  tc->rcv_las = tc->rcv_nxt;
486}
487
488/**
489 * Convert buffer to FIN-ACK
490 */
491void
492tcp_make_fin (tcp_connection_t * tc, vlib_buffer_t * b)
493{
494  tcp_make_ack_i (tc, b, TCP_STATE_ESTABLISHED, TCP_FLAG_FIN | TCP_FLAG_ACK);
495}
496
497/**
498 * Convert buffer to SYN
499 */
500void
501tcp_make_syn (tcp_connection_t * tc, vlib_buffer_t * b)
502{
503  u8 tcp_hdr_opts_len, tcp_opts_len;
504  tcp_header_t *th;
505  u16 initial_wnd;
506  tcp_options_t snd_opts;
507
508  initial_wnd = tcp_initial_window_to_advertise (tc);
509
510  /* Make and write options */
511  clib_memset (&snd_opts, 0, sizeof (snd_opts));
512  tcp_opts_len = tcp_make_syn_options (tc, &snd_opts);
513  tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
514
515  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->iss,
516			     tc->rcv_nxt, tcp_hdr_opts_len, TCP_FLAG_SYN,
517			     initial_wnd);
518  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
519  tcp_options_write ((u8 *) (th + 1), &snd_opts);
520}
521
522/**
523 * Convert buffer to SYN-ACK
524 */
525void
526tcp_make_synack (tcp_connection_t * tc, vlib_buffer_t * b)
527{
528  tcp_options_t _snd_opts, *snd_opts = &_snd_opts;
529  u8 tcp_opts_len, tcp_hdr_opts_len;
530  tcp_header_t *th;
531  u16 initial_wnd;
532
533  clib_memset (snd_opts, 0, sizeof (*snd_opts));
534  initial_wnd = tcp_initial_window_to_advertise (tc);
535  tcp_opts_len = tcp_make_synack_options (tc, snd_opts);
536  tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
537
538  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->iss,
539			     tc->rcv_nxt, tcp_hdr_opts_len,
540			     TCP_FLAG_SYN | TCP_FLAG_ACK, initial_wnd);
541  tcp_options_write ((u8 *) (th + 1), snd_opts);
542
543  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
544}
545
546always_inline void
547tcp_enqueue_to_ip_lookup_i (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
548			    u8 is_ip4, u32 fib_index, u8 flush)
549{
550  vlib_main_t *vm = wrk->vm;
551  u32 *to_next, next_index;
552  vlib_frame_t *f;
553
554  b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
555  b->error = 0;
556
557  vnet_buffer (b)->sw_if_index[VLIB_TX] = fib_index;
558  vnet_buffer (b)->sw_if_index[VLIB_RX] = 0;
559
560  /* Send to IP lookup */
561  next_index = is_ip4 ? ip4_lookup_node.index : ip6_lookup_node.index;
562  tcp_trajectory_add_start (b, 1);
563
564  f = wrk->ip_lookup_tx_frames[!is_ip4];
565  if (!f)
566    {
567      f = vlib_get_frame_to_node (vm, next_index);
568      ASSERT (f);
569      wrk->ip_lookup_tx_frames[!is_ip4] = f;
570    }
571
572  to_next = vlib_frame_vector_args (f);
573  to_next[f->n_vectors] = bi;
574  f->n_vectors += 1;
575  if (flush || f->n_vectors == VLIB_FRAME_SIZE)
576    {
577      vlib_put_frame_to_node (vm, next_index, f);
578      wrk->ip_lookup_tx_frames[!is_ip4] = 0;
579    }
580}
581
582static void
583tcp_enqueue_to_ip_lookup_now (tcp_worker_ctx_t * wrk, vlib_buffer_t * b,
584			      u32 bi, u8 is_ip4, u32 fib_index)
585{
586  tcp_enqueue_to_ip_lookup_i (wrk, b, bi, is_ip4, fib_index, 1);
587}
588
589static void
590tcp_enqueue_to_ip_lookup (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
591			  u8 is_ip4, u32 fib_index)
592{
593  tcp_enqueue_to_ip_lookup_i (wrk, b, bi, is_ip4, fib_index, 0);
594  if (wrk->vm->thread_index == 0 && vlib_num_workers ())
595    session_flush_frames_main_thread (wrk->vm);
596}
597
598always_inline void
599tcp_enqueue_to_output_i (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
600			 u8 is_ip4, u8 flush)
601{
602  u32 *to_next, next_index;
603  vlib_frame_t *f;
604
605  b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
606  b->error = 0;
607
608  /* Decide where to send the packet */
609  next_index = is_ip4 ? tcp4_output_node.index : tcp6_output_node.index;
610  tcp_trajectory_add_start (b, 2);
611
612  /* Get frame to v4/6 output node */
613  f = wrk->tx_frames[!is_ip4];
614  if (!f)
615    {
616      f = vlib_get_frame_to_node (wrk->vm, next_index);
617      ASSERT (f);
618      wrk->tx_frames[!is_ip4] = f;
619    }
620  to_next = vlib_frame_vector_args (f);
621  to_next[f->n_vectors] = bi;
622  f->n_vectors += 1;
623  if (flush || f->n_vectors == VLIB_FRAME_SIZE)
624    {
625      vlib_put_frame_to_node (wrk->vm, next_index, f);
626      wrk->tx_frames[!is_ip4] = 0;
627    }
628}
629
630static void
631tcp_enqueue_to_output (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
632		       u8 is_ip4)
633{
634  tcp_enqueue_to_output_i (wrk, b, bi, is_ip4, 0);
635}
636
637static void
638tcp_enqueue_to_output_now (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
639			   u8 is_ip4)
640{
641  tcp_enqueue_to_output_i (wrk, b, bi, is_ip4, 1);
642}
643#endif /* CLIB_MARCH_VARIANT */
644
645static int
646tcp_make_reset_in_place (vlib_main_t * vm, vlib_buffer_t * b0,
647			 tcp_state_t state, u8 thread_index, u8 is_ip4)
648{
649  ip4_header_t *ih4;
650  ip6_header_t *ih6;
651  tcp_header_t *th0;
652  ip4_address_t src_ip40, dst_ip40;
653  ip6_address_t src_ip60, dst_ip60;
654  u16 src_port, dst_port;
655  u32 tmp;
656  u32 seq, ack;
657  u8 flags;
658
659  /* Find IP and TCP headers */
660  th0 = tcp_buffer_hdr (b0);
661
662  /* Save src and dst ip */
663  if (is_ip4)
664    {
665      ih4 = vlib_buffer_get_current (b0);
666      ASSERT ((ih4->ip_version_and_header_length & 0xF0) == 0x40);
667      src_ip40.as_u32 = ih4->src_address.as_u32;
668      dst_ip40.as_u32 = ih4->dst_address.as_u32;
669    }
670  else
671    {
672      ih6 = vlib_buffer_get_current (b0);
673      ASSERT ((ih6->ip_version_traffic_class_and_flow_label & 0xF0) == 0x60);
674      clib_memcpy_fast (&src_ip60, &ih6->src_address, sizeof (ip6_address_t));
675      clib_memcpy_fast (&dst_ip60, &ih6->dst_address, sizeof (ip6_address_t));
676    }
677
678  src_port = th0->src_port;
679  dst_port = th0->dst_port;
680
681  /* Try to determine what/why we're actually resetting */
682  if (state == TCP_STATE_CLOSED)
683    {
684      if (!tcp_syn (th0))
685	return -1;
686
687      tmp = clib_net_to_host_u32 (th0->seq_number);
688
689      /* Got a SYN for no listener. */
690      flags = TCP_FLAG_RST | TCP_FLAG_ACK;
691      ack = clib_host_to_net_u32 (tmp + 1);
692      seq = 0;
693    }
694  else
695    {
696      flags = TCP_FLAG_RST;
697      seq = th0->ack_number;
698      ack = 0;
699    }
700
701  tcp_reuse_buffer (vm, b0);
702  tcp_trajectory_add_start (b0, 4);
703  th0 = vlib_buffer_push_tcp_net_order (b0, dst_port, src_port, seq, ack,
704					sizeof (tcp_header_t), flags, 0);
705
706  if (is_ip4)
707    {
708      ih4 = vlib_buffer_push_ip4 (vm, b0, &dst_ip40, &src_ip40,
709				  IP_PROTOCOL_TCP, 1);
710      th0->checksum = ip4_tcp_udp_compute_checksum (vm, b0, ih4);
711    }
712  else
713    {
714      int bogus = ~0;
715      ih6 = vlib_buffer_push_ip6 (vm, b0, &dst_ip60, &src_ip60,
716				  IP_PROTOCOL_TCP);
717      th0->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b0, ih6, &bogus);
718      ASSERT (!bogus);
719    }
720
721  return 0;
722}
723
724#ifndef CLIB_MARCH_VARIANT
725/**
726 *  Send reset without reusing existing buffer
727 *
728 *  It extracts connection info out of original packet
729 */
730void
731tcp_send_reset_w_pkt (tcp_connection_t * tc, vlib_buffer_t * pkt,
732		      u32 thread_index, u8 is_ip4)
733{
734  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
735  vlib_main_t *vm = wrk->vm;
736  vlib_buffer_t *b;
737  u32 bi, sw_if_index, fib_index;
738  u8 tcp_hdr_len, flags = 0;
739  tcp_header_t *th, *pkt_th;
740  u32 seq, ack;
741  ip4_header_t *ih4, *pkt_ih4;
742  ip6_header_t *ih6, *pkt_ih6;
743  fib_protocol_t fib_proto;
744
745  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
746    return;
747
748  b = vlib_get_buffer (vm, bi);
749  sw_if_index = vnet_buffer (pkt)->sw_if_index[VLIB_RX];
750  fib_proto = is_ip4 ? FIB_PROTOCOL_IP4 : FIB_PROTOCOL_IP6;
751  fib_index = fib_table_get_index_for_sw_if_index (fib_proto, sw_if_index);
752  tcp_init_buffer (vm, b);
753
754  /* Make and write options */
755  tcp_hdr_len = sizeof (tcp_header_t);
756
757  if (is_ip4)
758    {
759      pkt_ih4 = vlib_buffer_get_current (pkt);
760      pkt_th = ip4_next_header (pkt_ih4);
761    }
762  else
763    {
764      pkt_ih6 = vlib_buffer_get_current (pkt);
765      pkt_th = ip6_next_header (pkt_ih6);
766    }
767
768  if (tcp_ack (pkt_th))
769    {
770      flags = TCP_FLAG_RST;
771      seq = pkt_th->ack_number;
772      ack = (tc->state >= TCP_STATE_SYN_RCVD) ? tc->rcv_nxt : 0;
773    }
774  else
775    {
776      flags = TCP_FLAG_RST | TCP_FLAG_ACK;
777      seq = 0;
778      ack = clib_host_to_net_u32 (vnet_buffer (pkt)->tcp.seq_end);
779    }
780
781  th = vlib_buffer_push_tcp_net_order (b, pkt_th->dst_port, pkt_th->src_port,
782				       seq, ack, tcp_hdr_len, flags, 0);
783
784  /* Swap src and dst ip */
785  if (is_ip4)
786    {
787      ASSERT ((pkt_ih4->ip_version_and_header_length & 0xF0) == 0x40);
788      ih4 = vlib_buffer_push_ip4 (vm, b, &pkt_ih4->dst_address,
789				  &pkt_ih4->src_address, IP_PROTOCOL_TCP, 1);
790      th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih4);
791    }
792  else
793    {
794      int bogus = ~0;
795      ASSERT ((pkt_ih6->ip_version_traffic_class_and_flow_label & 0xF0) ==
796	      0x60);
797      ih6 = vlib_buffer_push_ip6 (vm, b, &pkt_ih6->dst_address,
798				  &pkt_ih6->src_address, IP_PROTOCOL_TCP);
799      th->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b, ih6, &bogus);
800      ASSERT (!bogus);
801    }
802
803  tcp_enqueue_to_ip_lookup_now (wrk, b, bi, is_ip4, fib_index);
804  TCP_EVT (TCP_EVT_RST_SENT, tc);
805  vlib_node_increment_counter (vm, tcp_node_index (output, tc->c_is_ip4),
806			       TCP_ERROR_RST_SENT, 1);
807}
808
809/**
810 * Build and set reset packet for connection
811 */
812void
813tcp_send_reset (tcp_connection_t * tc)
814{
815  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
816  vlib_main_t *vm = wrk->vm;
817  vlib_buffer_t *b;
818  u32 bi;
819  tcp_header_t *th;
820  u16 tcp_hdr_opts_len, advertise_wnd, opts_write_len;
821  u8 flags;
822
823  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
824    return;
825  b = vlib_get_buffer (vm, bi);
826  tcp_init_buffer (vm, b);
827
828  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
829  tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
830  advertise_wnd = tcp_window_to_advertise (tc, TCP_STATE_ESTABLISHED);
831  flags = TCP_FLAG_RST;
832  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->snd_nxt,
833			     tc->rcv_nxt, tcp_hdr_opts_len, flags,
834			     advertise_wnd);
835  opts_write_len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
836  ASSERT (opts_write_len == tc->snd_opts_len);
837  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
838  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
839  TCP_EVT (TCP_EVT_RST_SENT, tc);
840  vlib_node_increment_counter (vm, tcp_node_index (output, tc->c_is_ip4),
841			       TCP_ERROR_RST_SENT, 1);
842}
843
844static void
845tcp_push_ip_hdr (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
846		 vlib_buffer_t * b)
847{
848  tcp_header_t *th = vlib_buffer_get_current (b);
849  vlib_main_t *vm = wrk->vm;
850  if (tc->c_is_ip4)
851    {
852      ip4_header_t *ih;
853      ih = vlib_buffer_push_ip4 (vm, b, &tc->c_lcl_ip4,
854				 &tc->c_rmt_ip4, IP_PROTOCOL_TCP, 1);
855      th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih);
856    }
857  else
858    {
859      ip6_header_t *ih;
860      int bogus = ~0;
861
862      ih = vlib_buffer_push_ip6 (vm, b, &tc->c_lcl_ip6,
863				 &tc->c_rmt_ip6, IP_PROTOCOL_TCP);
864      th->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b, ih, &bogus);
865      ASSERT (!bogus);
866    }
867}
868
869/**
870 *  Send SYN
871 *
872 *  Builds a SYN packet for a half-open connection and sends it to ipx_lookup.
873 *  The packet is not forwarded through tcpx_output to avoid doing lookups
874 *  in the half_open pool.
875 */
876void
877tcp_send_syn (tcp_connection_t * tc)
878{
879  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
880  vlib_main_t *vm = wrk->vm;
881  vlib_buffer_t *b;
882  u32 bi;
883
884  /*
885   * Setup retransmit and establish timers before requesting buffer
886   * such that we can return if we've ran out.
887   */
888  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN,
889		    tc->rto * TCP_TO_TIMER_TICK);
890
891  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
892    {
893      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN, 1);
894      return;
895    }
896
897  b = vlib_get_buffer (vm, bi);
898  tcp_init_buffer (vm, b);
899  tcp_make_syn (tc, b);
900
901  /* Measure RTT with this */
902  tc->rtt_ts = tcp_time_now_us (vlib_num_workers ()? 1 : 0);
903  tc->rtt_seq = tc->snd_nxt;
904  tc->rto_boff = 0;
905
906  tcp_push_ip_hdr (wrk, tc, b);
907  tcp_enqueue_to_ip_lookup (wrk, b, bi, tc->c_is_ip4, tc->c_fib_index);
908  TCP_EVT (TCP_EVT_SYN_SENT, tc);
909}
910
911void
912tcp_send_synack (tcp_connection_t * tc)
913{
914  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
915  vlib_main_t *vm = wrk->vm;
916  vlib_buffer_t *b;
917  u32 bi;
918
919  tcp_retransmit_timer_force_update (tc);
920
921  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
922    {
923      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
924      return;
925    }
926
927  tc->rtt_ts = tcp_time_now_us (tc->c_thread_index);
928  b = vlib_get_buffer (vm, bi);
929  tcp_init_buffer (vm, b);
930  tcp_make_synack (tc, b);
931  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
932  TCP_EVT (TCP_EVT_SYNACK_SENT, tc);
933}
934
935/**
936 * Flush tx frame populated by retransmits and timer pops
937 */
938void
939tcp_flush_frame_to_output (tcp_worker_ctx_t * wrk, u8 is_ip4)
940{
941  if (wrk->tx_frames[!is_ip4])
942    {
943      u32 next_index;
944      next_index = is_ip4 ? tcp4_output_node.index : tcp6_output_node.index;
945      vlib_put_frame_to_node (wrk->vm, next_index, wrk->tx_frames[!is_ip4]);
946      wrk->tx_frames[!is_ip4] = 0;
947    }
948}
949
950/**
951 * Flush ip lookup tx frames populated by timer pops
952 */
953static void
954tcp_flush_frame_to_ip_lookup (tcp_worker_ctx_t * wrk, u8 is_ip4)
955{
956  if (wrk->ip_lookup_tx_frames[!is_ip4])
957    {
958      u32 next_index;
959      next_index = is_ip4 ? ip4_lookup_node.index : ip6_lookup_node.index;
960      vlib_put_frame_to_node (wrk->vm, next_index,
961			      wrk->ip_lookup_tx_frames[!is_ip4]);
962      wrk->ip_lookup_tx_frames[!is_ip4] = 0;
963    }
964}
965
966/**
967 * Flush v4 and v6 tcp and ip-lookup tx frames for thread index
968 */
969void
970tcp_flush_frames_to_output (tcp_worker_ctx_t * wrk)
971{
972  tcp_flush_frame_to_output (wrk, 1);
973  tcp_flush_frame_to_output (wrk, 0);
974  tcp_flush_frame_to_ip_lookup (wrk, 1);
975  tcp_flush_frame_to_ip_lookup (wrk, 0);
976}
977
978/**
979 *  Send FIN
980 */
981void
982tcp_send_fin (tcp_connection_t * tc)
983{
984  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
985  vlib_main_t *vm = wrk->vm;
986  vlib_buffer_t *b;
987  u32 bi;
988  u8 fin_snt = 0;
989
990  fin_snt = tc->flags & TCP_CONN_FINSNT;
991  if (fin_snt)
992    tc->snd_nxt -= 1;
993
994  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
995    {
996      /* Out of buffers so program fin retransmit ASAP */
997      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
998      if (fin_snt)
999	tc->snd_nxt += 1;
1000      else
1001	/* Make sure retransmit retries a fin not data */
1002	tc->flags |= TCP_CONN_FINSNT;
1003      return;
1004    }
1005
1006  tcp_retransmit_timer_force_update (tc);
1007  b = vlib_get_buffer (vm, bi);
1008  tcp_init_buffer (vm, b);
1009  tcp_make_fin (tc, b);
1010  tcp_enqueue_to_output_now (wrk, b, bi, tc->c_is_ip4);
1011  TCP_EVT (TCP_EVT_FIN_SENT, tc);
1012  /* Account for the FIN */
1013  tc->snd_nxt += 1;
1014  if (!fin_snt)
1015    {
1016      tc->flags |= TCP_CONN_FINSNT;
1017      tc->flags &= ~TCP_CONN_FINPNDG;
1018      tc->snd_una_max = seq_max (tc->snd_una_max, tc->snd_nxt);
1019    }
1020}
1021
1022/**
1023 * Push TCP header and update connection variables. Should only be called
1024 * for segments with data, not for 'control' packets.
1025 */
1026always_inline void
1027tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b, u32 snd_nxt,
1028		u8 compute_opts, u8 maybe_burst, u8 update_snd_nxt)
1029{
1030  u8 tcp_hdr_opts_len, flags = TCP_FLAG_ACK;
1031  u32 advertise_wnd, data_len;
1032  tcp_main_t *tm = &tcp_main;
1033  tcp_header_t *th;
1034
1035  data_len = b->current_length;
1036  if (PREDICT_FALSE (b->flags & VLIB_BUFFER_NEXT_PRESENT))
1037    data_len += b->total_length_not_including_first_buffer;
1038
1039  vnet_buffer (b)->tcp.flags = 0;
1040  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
1041
1042  if (compute_opts)
1043    tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
1044
1045  tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
1046
1047  if (maybe_burst)
1048    advertise_wnd = tc->rcv_wnd >> tc->rcv_wscale;
1049  else
1050    advertise_wnd = tcp_window_to_advertise (tc, TCP_STATE_ESTABLISHED);
1051
1052  if (PREDICT_FALSE (tc->flags & TCP_CONN_PSH_PENDING))
1053    {
1054      if (seq_geq (tc->psh_seq, snd_nxt)
1055	  && seq_lt (tc->psh_seq, snd_nxt + data_len))
1056	flags |= TCP_FLAG_PSH;
1057    }
1058  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, snd_nxt,
1059			     tc->rcv_nxt, tcp_hdr_opts_len, flags,
1060			     advertise_wnd);
1061
1062  if (maybe_burst)
1063    {
1064      clib_memcpy_fast ((u8 *) (th + 1),
1065			tm->wrk_ctx[tc->c_thread_index].cached_opts,
1066			tc->snd_opts_len);
1067    }
1068  else
1069    {
1070      u8 len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
1071      ASSERT (len == tc->snd_opts_len);
1072    }
1073
1074  /*
1075   * Update connection variables
1076   */
1077
1078  if (update_snd_nxt)
1079    tc->snd_nxt += data_len;
1080  tc->rcv_las = tc->rcv_nxt;
1081
1082  tc->bytes_out += data_len;
1083  tc->data_segs_out += 1;
1084
1085  TCP_EVT (TCP_EVT_PKTIZE, tc);
1086}
1087
1088u32
1089tcp_session_push_header (transport_connection_t * tconn, vlib_buffer_t * b)
1090{
1091  tcp_connection_t *tc = (tcp_connection_t *) tconn;
1092
1093  if (tc->flags & TCP_CONN_TRACK_BURST)
1094    {
1095      tcp_bt_check_app_limited (tc);
1096      tcp_bt_track_tx (tc);
1097      tc->flags &= ~TCP_CONN_TRACK_BURST;
1098    }
1099
1100  tcp_push_hdr_i (tc, b, tc->snd_nxt, /* compute opts */ 0, /* burst */ 1,
1101		  /* update_snd_nxt */ 1);
1102
1103  tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
1104  tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
1105  /* If not tracking an ACK, start tracking */
1106  if (tc->rtt_ts == 0 && !tcp_in_cong_recovery (tc))
1107    {
1108      tc->rtt_ts = tcp_time_now_us (tc->c_thread_index);
1109      tc->rtt_seq = tc->snd_nxt;
1110    }
1111  if (PREDICT_FALSE (!tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT)))
1112    {
1113      tcp_retransmit_timer_set (tc);
1114      tc->rto_boff = 0;
1115    }
1116  tcp_trajectory_add_start (b, 3);
1117  return 0;
1118}
1119
1120void
1121tcp_send_ack (tcp_connection_t * tc)
1122{
1123  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
1124  vlib_main_t *vm = wrk->vm;
1125  vlib_buffer_t *b;
1126  u32 bi;
1127
1128  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1129    {
1130      tcp_update_rcv_wnd (tc);
1131      return;
1132    }
1133  b = vlib_get_buffer (vm, bi);
1134  tcp_init_buffer (vm, b);
1135  tcp_make_ack (tc, b);
1136  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1137}
1138
1139void
1140tcp_program_ack (tcp_connection_t * tc)
1141{
1142  if (!(tc->flags & TCP_CONN_SNDACK))
1143    {
1144      session_add_self_custom_tx_evt (&tc->connection, 1);
1145      tc->flags |= TCP_CONN_SNDACK;
1146    }
1147}
1148
1149void
1150tcp_program_dupack (tcp_connection_t * tc)
1151{
1152  if (!(tc->flags & TCP_CONN_SNDACK))
1153    {
1154      session_add_self_custom_tx_evt (&tc->connection, 1);
1155      tc->flags |= TCP_CONN_SNDACK;
1156    }
1157  if (tc->pending_dupacks < 255)
1158    tc->pending_dupacks += 1;
1159}
1160
1161void
1162tcp_program_retransmit (tcp_connection_t * tc)
1163{
1164  if (!(tc->flags & TCP_CONN_RXT_PENDING))
1165    {
1166      session_add_self_custom_tx_evt (&tc->connection, 0);
1167      tc->flags |= TCP_CONN_RXT_PENDING;
1168    }
1169}
1170
1171/**
1172 * Delayed ack timer handler
1173 *
1174 * Sends delayed ACK when timer expires
1175 */
1176void
1177tcp_timer_delack_handler (u32 index)
1178{
1179  u32 thread_index = vlib_get_thread_index ();
1180  tcp_connection_t *tc;
1181
1182  tc = tcp_connection_get (index, thread_index);
1183  tc->timers[TCP_TIMER_DELACK] = TCP_TIMER_HANDLE_INVALID;
1184  tcp_send_ack (tc);
1185}
1186
1187/**
1188 * Send window update ack
1189 *
1190 * Ensures that it will be sent only once, after a zero rwnd has been
1191 * advertised in a previous ack, and only if rwnd has grown beyond a
1192 * configurable value.
1193 */
1194void
1195tcp_send_window_update_ack (tcp_connection_t * tc)
1196{
1197  if (tcp_zero_rwnd_sent (tc))
1198    {
1199      tcp_update_rcv_wnd (tc);
1200      if (tc->rcv_wnd >= tcp_cfg.rwnd_min_update_ack * tc->snd_mss)
1201	{
1202	  tcp_zero_rwnd_sent_off (tc);
1203	  tcp_program_ack (tc);
1204	}
1205    }
1206}
1207
1208/**
1209 * Allocate a new buffer and build a new tcp segment
1210 *
1211 * @param wrk		tcp worker
1212 * @param tc		connection for which the segment will be allocated
1213 * @param offset	offset of the first byte in the tx fifo
1214 * @param max_deq_byte	segment size
1215 * @param[out] b	pointer to buffer allocated
1216 *
1217 * @return 	the number of bytes in the segment or 0 if buffer cannot be
1218 * 		allocated or no data available
1219 */
1220static int
1221tcp_prepare_segment (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1222		     u32 offset, u32 max_deq_bytes, vlib_buffer_t ** b)
1223{
1224  u32 bytes_per_buffer = vnet_get_tcp_main ()->bytes_per_buffer;
1225  vlib_main_t *vm = wrk->vm;
1226  u32 bi, seg_size;
1227  int n_bytes = 0;
1228  u8 *data;
1229
1230  seg_size = max_deq_bytes + TRANSPORT_MAX_HDRS_LEN;
1231
1232  /*
1233   * Prepare options
1234   */
1235  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
1236
1237  /*
1238   * Allocate and fill in buffer(s)
1239   */
1240
1241  /* Easy case, buffer size greater than mss */
1242  if (PREDICT_TRUE (seg_size <= bytes_per_buffer))
1243    {
1244      if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1245	return 0;
1246      *b = vlib_get_buffer (vm, bi);
1247      data = tcp_init_buffer (vm, *b);
1248      n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1249					    max_deq_bytes);
1250      ASSERT (n_bytes == max_deq_bytes);
1251      b[0]->current_length = n_bytes;
1252      tcp_push_hdr_i (tc, *b, tc->snd_una + offset, /* compute opts */ 0,
1253		      /* burst */ 0, /* update_snd_nxt */ 0);
1254    }
1255  /* Split mss into multiple buffers */
1256  else
1257    {
1258      u32 chain_bi = ~0, n_bufs_per_seg, n_bufs;
1259      u16 n_peeked, len_to_deq;
1260      vlib_buffer_t *chain_b, *prev_b;
1261      int i;
1262
1263      /* Make sure we have enough buffers */
1264      n_bufs_per_seg = ceil ((double) seg_size / bytes_per_buffer);
1265      vec_validate_aligned (wrk->tx_buffers, n_bufs_per_seg - 1,
1266			    CLIB_CACHE_LINE_BYTES);
1267      n_bufs = vlib_buffer_alloc (vm, wrk->tx_buffers, n_bufs_per_seg);
1268      if (PREDICT_FALSE (n_bufs != n_bufs_per_seg))
1269	{
1270	  if (n_bufs)
1271	    vlib_buffer_free (vm, wrk->tx_buffers, n_bufs);
1272	  return 0;
1273	}
1274
1275      *b = vlib_get_buffer (vm, wrk->tx_buffers[--n_bufs]);
1276      data = tcp_init_buffer (vm, *b);
1277      n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1278					    bytes_per_buffer -
1279					    TRANSPORT_MAX_HDRS_LEN);
1280      b[0]->current_length = n_bytes;
1281      b[0]->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
1282      b[0]->total_length_not_including_first_buffer = 0;
1283      max_deq_bytes -= n_bytes;
1284
1285      chain_b = *b;
1286      for (i = 1; i < n_bufs_per_seg; i++)
1287	{
1288	  prev_b = chain_b;
1289	  len_to_deq = clib_min (max_deq_bytes, bytes_per_buffer);
1290	  chain_bi = wrk->tx_buffers[--n_bufs];
1291	  chain_b = vlib_get_buffer (vm, chain_bi);
1292	  chain_b->current_data = 0;
1293	  data = vlib_buffer_get_current (chain_b);
1294	  n_peeked = session_tx_fifo_peek_bytes (&tc->connection, data,
1295						 offset + n_bytes,
1296						 len_to_deq);
1297	  ASSERT (n_peeked == len_to_deq);
1298	  n_bytes += n_peeked;
1299	  chain_b->current_length = n_peeked;
1300	  chain_b->next_buffer = 0;
1301
1302	  /* update previous buffer */
1303	  prev_b->next_buffer = chain_bi;
1304	  prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT;
1305
1306	  max_deq_bytes -= n_peeked;
1307	  b[0]->total_length_not_including_first_buffer += n_peeked;
1308	}
1309
1310      tcp_push_hdr_i (tc, *b, tc->snd_una + offset, /* compute opts */ 0,
1311		      /* burst */ 0, /* update_snd_nxt */ 0);
1312
1313      if (PREDICT_FALSE (n_bufs))
1314	{
1315	  clib_warning ("not all buffers consumed");
1316	  vlib_buffer_free (vm, wrk->tx_buffers, n_bufs);
1317	}
1318    }
1319
1320  ASSERT (n_bytes > 0);
1321  ASSERT (((*b)->current_data + (*b)->current_length) <= bytes_per_buffer);
1322
1323  return n_bytes;
1324}
1325
1326/**
1327 * Build a retransmit segment
1328 *
1329 * @return the number of bytes in the segment or 0 if there's nothing to
1330 *         retransmit
1331 */
1332static u32
1333tcp_prepare_retransmit_segment (tcp_worker_ctx_t * wrk,
1334				tcp_connection_t * tc, u32 offset,
1335				u32 max_deq_bytes, vlib_buffer_t ** b)
1336{
1337  u32 start, available_bytes;
1338  int n_bytes = 0;
1339
1340  ASSERT (tc->state >= TCP_STATE_ESTABLISHED);
1341  ASSERT (max_deq_bytes != 0);
1342
1343  /*
1344   * Make sure we can retransmit something
1345   */
1346  available_bytes = transport_max_tx_dequeue (&tc->connection);
1347  ASSERT (available_bytes >= offset);
1348  available_bytes -= offset;
1349  if (!available_bytes)
1350    return 0;
1351
1352  max_deq_bytes = clib_min (tc->snd_mss, max_deq_bytes);
1353  max_deq_bytes = clib_min (available_bytes, max_deq_bytes);
1354
1355  /* Start is beyond snd_congestion */
1356  start = tc->snd_una + offset;
1357  if (seq_geq (start, tc->snd_congestion))
1358    return 0;
1359
1360  /* Don't overshoot snd_congestion */
1361  if (seq_gt (start + max_deq_bytes, tc->snd_congestion))
1362    {
1363      max_deq_bytes = tc->snd_congestion - start;
1364      if (max_deq_bytes == 0)
1365	return 0;
1366    }
1367
1368  n_bytes = tcp_prepare_segment (wrk, tc, offset, max_deq_bytes, b);
1369  if (!n_bytes)
1370    return 0;
1371
1372  tc->snd_rxt_bytes += n_bytes;
1373
1374  if (tc->flags & TCP_CONN_RATE_SAMPLE)
1375    tcp_bt_track_rxt (tc, start, start + n_bytes);
1376
1377  tc->bytes_retrans += n_bytes;
1378  tc->segs_retrans += 1;
1379  TCP_EVT (TCP_EVT_CC_RTX, tc, offset, n_bytes);
1380
1381  return n_bytes;
1382}
1383
1384static void
1385tcp_check_sack_reneging (tcp_connection_t * tc)
1386{
1387  sack_scoreboard_t *sb = &tc->sack_sb;
1388  sack_scoreboard_hole_t *hole;
1389
1390  hole = scoreboard_first_hole (sb);
1391  if (!sb->is_reneging && (!hole || hole->start == tc->snd_una))
1392    return;
1393
1394  scoreboard_clear_reneging (sb, tc->snd_una, tc->snd_nxt);
1395}
1396
1397/**
1398 * Reset congestion control, switch cwnd to loss window and try again.
1399 */
1400static void
1401tcp_cc_init_rxt_timeout (tcp_connection_t * tc)
1402{
1403  TCP_EVT (TCP_EVT_CC_EVT, tc, 6);
1404
1405  tc->prev_ssthresh = tc->ssthresh;
1406  tc->prev_cwnd = tc->cwnd;
1407
1408  /* If we entrered loss without fast recovery, notify cc algo of the
1409   * congestion event such that it can update ssthresh and its state */
1410  if (!tcp_in_fastrecovery (tc))
1411    tcp_cc_congestion (tc);
1412
1413  /* Let cc algo decide loss cwnd and ssthresh post unrecovered loss */
1414  tcp_cc_loss (tc);
1415
1416  tc->rtt_ts = 0;
1417  tc->cwnd_acc_bytes = 0;
1418  tc->tr_occurences += 1;
1419  tcp_recovery_on (tc);
1420}
1421
1422void
1423tcp_timer_retransmit_handler (u32 tc_index)
1424{
1425  u32 thread_index = vlib_get_thread_index ();
1426  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
1427  vlib_main_t *vm = wrk->vm;
1428  tcp_connection_t *tc;
1429  vlib_buffer_t *b = 0;
1430  u32 bi, n_bytes;
1431
1432  tc = tcp_connection_get (tc_index, thread_index);
1433
1434  /* Note: the connection may have been closed and pool_put */
1435  if (PREDICT_FALSE (tc == 0 || tc->state == TCP_STATE_SYN_SENT))
1436    return;
1437
1438  tc->timers[TCP_TIMER_RETRANSMIT] = TCP_TIMER_HANDLE_INVALID;
1439
1440  /* Wait-close and retransmit could pop at the same time */
1441  if (tc->state == TCP_STATE_CLOSED)
1442    return;
1443
1444  if (tc->state >= TCP_STATE_ESTABLISHED)
1445    {
1446      TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
1447
1448      /* Lost FIN, retransmit and return */
1449      if (tc->flags & TCP_CONN_FINSNT)
1450	{
1451	  tcp_send_fin (tc);
1452	  tc->rto_boff += 1;
1453	  tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1454	  return;
1455	}
1456
1457      /* Shouldn't be here. This condition is tricky because it has to take
1458       * into account boff > 0 due to persist timeout. */
1459      if ((tc->rto_boff == 0 && tc->snd_una == tc->snd_nxt)
1460	  || (tc->rto_boff > 0 && seq_geq (tc->snd_una, tc->snd_congestion)
1461	      && !tcp_flight_size (tc)))
1462	{
1463	  ASSERT (!tcp_in_recovery (tc));
1464	  tc->rto_boff = 0;
1465	  return;
1466	}
1467
1468      /* We're not in recovery so make sure rto_boff is 0. Can be non 0 due
1469       * to persist timer timeout */
1470      if (!tcp_in_recovery (tc) && tc->rto_boff > 0)
1471	{
1472	  tc->rto_boff = 0;
1473	  tcp_update_rto (tc);
1474	}
1475
1476      /* Peer is dead or network connectivity is lost. Close connection.
1477       * RFC 1122 section 4.2.3.5 recommends a value of at least 100s. For
1478       * a min rto of 0.2s we need to retry about 8 times. */
1479      if (tc->rto_boff >= TCP_RTO_BOFF_MAX)
1480	{
1481	  tcp_send_reset (tc);
1482	  tcp_connection_set_state (tc, TCP_STATE_CLOSED);
1483	  session_transport_closing_notify (&tc->connection);
1484	  tcp_connection_timers_reset (tc);
1485	  tcp_timer_update (tc, TCP_TIMER_WAITCLOSE, tcp_cfg.closewait_time);
1486	  return;
1487	}
1488
1489      if (tcp_opts_sack_permitted (&tc->rcv_opts))
1490	tcp_check_sack_reneging (tc);
1491
1492      /* Update send congestion to make sure that rxt has data to send */
1493      tc->snd_congestion = tc->snd_nxt;
1494
1495      /* Send the first unacked segment. If we're short on buffers, return
1496       * as soon as possible */
1497      n_bytes = tcp_prepare_retransmit_segment (wrk, tc, 0, tc->snd_mss, &b);
1498      if (!n_bytes)
1499	{
1500	  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
1501	  return;
1502	}
1503
1504      bi = vlib_get_buffer_index (vm, b);
1505      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1506
1507      tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1508      tcp_retransmit_timer_force_update (tc);
1509
1510      tc->rto_boff += 1;
1511      if (tc->rto_boff == 1)
1512	{
1513	  tcp_cc_init_rxt_timeout (tc);
1514	  /* Record timestamp. Eifel detection algorithm RFC3522 */
1515	  tc->snd_rxt_ts = tcp_tstamp (tc);
1516	}
1517
1518      if (tcp_opts_sack_permitted (&tc->rcv_opts))
1519	scoreboard_init_high_rxt (&tc->sack_sb, tc->snd_una + tc->snd_mss);
1520
1521      tcp_program_retransmit (tc);
1522    }
1523  /* Retransmit SYN-ACK */
1524  else if (tc->state == TCP_STATE_SYN_RCVD)
1525    {
1526      TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
1527
1528      tc->rtt_ts = 0;
1529
1530      /* Passive open establish timeout */
1531      if (tc->rto > TCP_ESTABLISH_TIME >> 1)
1532	{
1533	  tcp_connection_set_state (tc, TCP_STATE_CLOSED);
1534	  tcp_connection_timers_reset (tc);
1535	  tcp_timer_update (tc, TCP_TIMER_WAITCLOSE, tcp_cfg.cleanup_time);
1536	  return;
1537	}
1538
1539      if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1540	{
1541	  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
1542	  return;
1543	}
1544
1545      tc->rto_boff += 1;
1546      if (tc->rto_boff > TCP_RTO_SYN_RETRIES)
1547	tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1548
1549      tcp_retransmit_timer_force_update (tc);
1550
1551      b = vlib_get_buffer (vm, bi);
1552      tcp_init_buffer (vm, b);
1553      tcp_make_synack (tc, b);
1554      TCP_EVT (TCP_EVT_SYN_RXT, tc, 1);
1555
1556      /* Retransmit timer already updated, just enqueue to output */
1557      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1558    }
1559  else
1560    {
1561      ASSERT (tc->state == TCP_STATE_CLOSED);
1562      return;
1563    }
1564}
1565
1566/**
1567 * SYN retransmit timer handler. Active open only.
1568 */
1569void
1570tcp_timer_retransmit_syn_handler (u32 tc_index)
1571{
1572  u32 thread_index = vlib_get_thread_index ();
1573  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
1574  vlib_main_t *vm = wrk->vm;
1575  tcp_connection_t *tc;
1576  vlib_buffer_t *b = 0;
1577  u32 bi;
1578
1579  tc = tcp_half_open_connection_get (tc_index);
1580
1581  /* Note: the connection may have transitioned to ESTABLISHED... */
1582  if (PREDICT_FALSE (tc == 0 || tc->state != TCP_STATE_SYN_SENT))
1583    return;
1584
1585  tc->timers[TCP_TIMER_RETRANSMIT_SYN] = TCP_TIMER_HANDLE_INVALID;
1586
1587  /* Half-open connection actually moved to established but we were
1588   * waiting for syn retransmit to pop to call cleanup from the right
1589   * thread. */
1590  if (tc->flags & TCP_CONN_HALF_OPEN_DONE)
1591    {
1592      if (tcp_half_open_connection_cleanup (tc))
1593	TCP_DBG ("could not remove half-open connection");
1594      return;
1595    }
1596
1597  TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
1598  tc->rtt_ts = 0;
1599
1600  /* Active open establish timeout */
1601  if (tc->rto >= TCP_ESTABLISH_TIME >> 1)
1602    {
1603      session_stream_connect_notify (&tc->connection, 1 /* fail */ );
1604      tcp_connection_cleanup (tc);
1605      return;
1606    }
1607
1608  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1609    {
1610      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN, 1);
1611      return;
1612    }
1613
1614  /* Try without increasing RTO a number of times. If this fails,
1615   * start growing RTO exponentially */
1616  tc->rto_boff += 1;
1617  if (tc->rto_boff > TCP_RTO_SYN_RETRIES)
1618    tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1619
1620  b = vlib_get_buffer (vm, bi);
1621  tcp_init_buffer (vm, b);
1622  tcp_make_syn (tc, b);
1623
1624  TCP_EVT (TCP_EVT_SYN_RXT, tc, 0);
1625
1626  /* This goes straight to ipx_lookup */
1627  tcp_push_ip_hdr (wrk, tc, b);
1628  tcp_enqueue_to_ip_lookup (wrk, b, bi, tc->c_is_ip4, tc->c_fib_index);
1629
1630  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN,
1631		    tc->rto * TCP_TO_TIMER_TICK);
1632}
1633
1634/**
1635 * Got 0 snd_wnd from peer, try to do something about it.
1636 *
1637 */
1638void
1639tcp_timer_persist_handler (u32 index)
1640{
1641  u32 thread_index = vlib_get_thread_index ();
1642  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
1643  u32 bi, max_snd_bytes, available_bytes, offset;
1644  tcp_main_t *tm = vnet_get_tcp_main ();
1645  vlib_main_t *vm = wrk->vm;
1646  tcp_connection_t *tc;
1647  vlib_buffer_t *b;
1648  int n_bytes = 0;
1649  u8 *data;
1650
1651  tc = tcp_connection_get_if_valid (index, thread_index);
1652  if (!tc)
1653    return;
1654
1655  /* Make sure timer handle is set to invalid */
1656  tc->timers[TCP_TIMER_PERSIST] = TCP_TIMER_HANDLE_INVALID;
1657
1658  /* Problem already solved or worse */
1659  if (tc->state == TCP_STATE_CLOSED || tc->snd_wnd > tc->snd_mss
1660      || (tc->flags & TCP_CONN_FINSNT))
1661    return;
1662
1663  available_bytes = transport_max_tx_dequeue (&tc->connection);
1664  offset = tc->snd_nxt - tc->snd_una;
1665
1666  /* Reprogram persist if no new bytes available to send. We may have data
1667   * next time */
1668  if (!available_bytes)
1669    {
1670      tcp_persist_timer_set (tc);
1671      return;
1672    }
1673
1674  if (available_bytes <= offset)
1675    {
1676      ASSERT (tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT));
1677      return;
1678    }
1679
1680  /* Increment RTO backoff */
1681  tc->rto_boff += 1;
1682  tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1683
1684  /*
1685   * Try to force the first unsent segment (or buffer)
1686   */
1687  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1688    {
1689      tcp_persist_timer_set (tc);
1690      return;
1691    }
1692  b = vlib_get_buffer (vm, bi);
1693  data = tcp_init_buffer (vm, b);
1694
1695  tcp_validate_txf_size (tc, offset);
1696  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
1697  max_snd_bytes = clib_min (tc->snd_mss,
1698			    tm->bytes_per_buffer - TRANSPORT_MAX_HDRS_LEN);
1699  n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1700					max_snd_bytes);
1701  b->current_length = n_bytes;
1702  ASSERT (n_bytes != 0 && (tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT)
1703			   || tc->snd_nxt == tc->snd_una_max
1704			   || tc->rto_boff > 1));
1705
1706  if (tc->flags & TCP_CONN_RATE_SAMPLE)
1707    {
1708      tcp_bt_check_app_limited (tc);
1709      tcp_bt_track_tx (tc);
1710    }
1711
1712  tcp_push_hdr_i (tc, b, tc->snd_nxt, /* compute opts */ 0,
1713		  /* burst */ 0, /* update_snd_nxt */ 1);
1714  tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
1715  tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
1716  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1717
1718  /* Just sent new data, enable retransmit */
1719  tcp_retransmit_timer_update (tc);
1720}
1721
1722/**
1723 * Retransmit first unacked segment
1724 */
1725int
1726tcp_retransmit_first_unacked (tcp_worker_ctx_t * wrk, tcp_connection_t * tc)
1727{
1728  vlib_main_t *vm = wrk->vm;
1729  vlib_buffer_t *b;
1730  u32 bi, n_bytes;
1731
1732  TCP_EVT (TCP_EVT_CC_EVT, tc, 1);
1733
1734  n_bytes = tcp_prepare_retransmit_segment (wrk, tc, 0, tc->snd_mss, &b);
1735  if (!n_bytes)
1736    return -1;
1737
1738  bi = vlib_get_buffer_index (vm, b);
1739  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1740
1741  return 0;
1742}
1743
1744static int
1745tcp_transmit_unsent (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1746		     u32 burst_size)
1747{
1748  u32 offset, n_segs = 0, n_written, bi, available_wnd;
1749  vlib_main_t *vm = wrk->vm;
1750  vlib_buffer_t *b = 0;
1751
1752  offset = tc->snd_nxt - tc->snd_una;
1753  available_wnd = tc->snd_wnd - offset;
1754  burst_size = clib_min (burst_size, available_wnd / tc->snd_mss);
1755
1756  while (n_segs < burst_size)
1757    {
1758      n_written = tcp_prepare_segment (wrk, tc, offset, tc->snd_mss, &b);
1759      if (!n_written)
1760	goto done;
1761
1762      bi = vlib_get_buffer_index (vm, b);
1763      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1764      offset += n_written;
1765      n_segs += 1;
1766
1767      tc->snd_nxt += n_written;
1768      tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
1769    }
1770
1771done:
1772  return n_segs;
1773}
1774
1775/**
1776 * Estimate send space using proportional rate reduction (RFC6937)
1777 */
1778static int
1779tcp_fastrecovery_prr_snd_space (tcp_connection_t * tc)
1780{
1781  u32 pipe, prr_out;
1782  int space;
1783
1784  pipe = tcp_flight_size (tc);
1785  prr_out = tc->snd_rxt_bytes + (tc->snd_nxt - tc->snd_congestion);
1786
1787  if (pipe > tc->ssthresh)
1788    {
1789      space = ((int) tc->prr_delivered * ((f64) tc->ssthresh / tc->prev_cwnd))
1790	- prr_out;
1791    }
1792  else
1793    {
1794      int limit = tc->prr_delivered - prr_out + tc->snd_mss;
1795      space = clib_min (tc->ssthresh - pipe, limit);
1796    }
1797  space = clib_max (space, prr_out ? 0 : tc->snd_mss);
1798  return space;
1799}
1800
1801#define scoreboard_rescue_rxt_valid(_sb, _tc)			\
1802    (seq_geq (_sb->rescue_rxt, _tc->snd_una) 			\
1803	&& seq_leq (_sb->rescue_rxt, _tc->snd_congestion))
1804
1805/**
1806 * Do retransmit with SACKs
1807 */
1808static int
1809tcp_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1810		     u32 burst_size)
1811{
1812  u32 n_written = 0, offset, max_bytes, n_segs = 0;
1813  sack_scoreboard_hole_t *hole;
1814  vlib_main_t *vm = wrk->vm;
1815  vlib_buffer_t *b = 0;
1816  sack_scoreboard_t *sb;
1817  u32 bi, max_deq;
1818  int snd_space;
1819  u8 snd_limited = 0, can_rescue = 0;
1820
1821  ASSERT (tcp_in_cong_recovery (tc));
1822
1823  if (tcp_in_recovery (tc))
1824    snd_space = tcp_available_cc_snd_space (tc);
1825  else
1826    snd_space = tcp_fastrecovery_prr_snd_space (tc);
1827
1828  if (snd_space < tc->snd_mss)
1829    {
1830      /* We're cc constrained so don't accumulate tokens */
1831      transport_connection_tx_pacer_reset_bucket (&tc->connection,
1832						  vm->
1833						  clib_time.last_cpu_time);
1834      return 0;
1835    }
1836
1837  TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
1838  sb = &tc->sack_sb;
1839  hole = scoreboard_get_hole (sb, sb->cur_rxt_hole);
1840
1841  max_deq = transport_max_tx_dequeue (&tc->connection);
1842  max_deq -= tc->snd_nxt - tc->snd_una;
1843
1844  while (snd_space > 0 && n_segs < burst_size)
1845    {
1846      hole = scoreboard_next_rxt_hole (sb, hole, max_deq, &can_rescue,
1847				       &snd_limited);
1848      if (!hole)
1849	{
1850	  /* We are out of lost holes to retransmit so send some new data. */
1851	  if (max_deq)
1852	    {
1853	      u32 n_segs_new, av_window;
1854	      av_window = tc->snd_wnd - (tc->snd_nxt - tc->snd_una);
1855	      snd_space = clib_min (snd_space, av_window);
1856	      snd_space = clib_min (max_deq, snd_space);
1857	      burst_size = clib_min (burst_size - n_segs,
1858				     snd_space / tc->snd_mss);
1859	      burst_size = clib_min (burst_size, TCP_RXT_MAX_BURST);
1860	      n_segs_new = tcp_transmit_unsent (wrk, tc, burst_size);
1861	      if (max_deq > n_segs_new * tc->snd_mss)
1862		tcp_program_retransmit (tc);
1863
1864	      n_segs += n_segs_new;
1865	      goto done;
1866	    }
1867
1868	  if (tcp_in_recovery (tc) || !can_rescue
1869	      || scoreboard_rescue_rxt_valid (sb, tc))
1870	    break;
1871
1872	  /* If rescue rxt undefined or less than snd_una then one segment of
1873	   * up to SMSS octets that MUST include the highest outstanding
1874	   * unSACKed sequence number SHOULD be returned, and RescueRxt set to
1875	   * RecoveryPoint. HighRxt MUST NOT be updated.
1876	   */
1877	  max_bytes = clib_min (tc->snd_mss,
1878				tc->snd_congestion - tc->snd_una);
1879	  max_bytes = clib_min (max_bytes, snd_space);
1880	  offset = tc->snd_congestion - tc->snd_una - max_bytes;
1881	  sb->rescue_rxt = tc->snd_congestion;
1882	  n_written = tcp_prepare_retransmit_segment (wrk, tc, offset,
1883						      max_bytes, &b);
1884	  if (!n_written)
1885	    goto done;
1886
1887	  bi = vlib_get_buffer_index (vm, b);
1888	  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1889	  n_segs += 1;
1890	  break;
1891	}
1892
1893      max_bytes = clib_min (hole->end - sb->high_rxt, snd_space);
1894      max_bytes = snd_limited ? clib_min (max_bytes, tc->snd_mss) : max_bytes;
1895      if (max_bytes == 0)
1896	break;
1897
1898      offset = sb->high_rxt - tc->snd_una;
1899      n_written = tcp_prepare_retransmit_segment (wrk, tc, offset, max_bytes,
1900						  &b);
1901      ASSERT (n_written <= snd_space);
1902
1903      /* Nothing left to retransmit */
1904      if (n_written == 0)
1905	break;
1906
1907      bi = vlib_get_buffer_index (vm, b);
1908      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1909
1910      sb->high_rxt += n_written;
1911      snd_space -= n_written;
1912      n_segs += 1;
1913    }
1914
1915  if (hole)
1916    tcp_program_retransmit (tc);
1917
1918done:
1919
1920  return n_segs;
1921}
1922
1923/**
1924 * Fast retransmit without SACK info
1925 */
1926static int
1927tcp_retransmit_no_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1928			u32 burst_size)
1929{
1930  u32 n_written = 0, offset = 0, bi, max_deq, n_segs_now;
1931  vlib_main_t *vm = wrk->vm;
1932  int snd_space, n_segs = 0;
1933  vlib_buffer_t *b;
1934
1935  ASSERT (tcp_in_fastrecovery (tc));
1936  TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
1937
1938  snd_space = tcp_available_cc_snd_space (tc);
1939
1940  if (!tcp_fastrecovery_first (tc))
1941    goto send_unsent;
1942
1943  /* RFC 6582: [If a partial ack], retransmit the first unacknowledged
1944   * segment. */
1945  while (snd_space > 0 && n_segs < burst_size)
1946    {
1947      n_written = tcp_prepare_retransmit_segment (wrk, tc, offset,
1948						  tc->snd_mss, &b);
1949
1950      /* Nothing left to retransmit */
1951      if (n_written == 0)
1952	break;
1953
1954      bi = vlib_get_buffer_index (vm, b);
1955      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1956      snd_space -= n_written;
1957      offset += n_written;
1958      n_segs += 1;
1959    }
1960
1961  if (n_segs == burst_size)
1962    goto done;
1963
1964send_unsent:
1965
1966  /* RFC 6582: Send a new segment if permitted by the new value of cwnd. */
1967  if (snd_space < tc->snd_mss || tc->snd_mss == 0)
1968    goto done;
1969
1970  max_deq = transport_max_tx_dequeue (&tc->connection);
1971  max_deq -= tc->snd_nxt - tc->snd_una;
1972  if (max_deq)
1973    {
1974      snd_space = clib_min (max_deq, snd_space);
1975      burst_size = clib_min (burst_size - n_segs, snd_space / tc->snd_mss);
1976      n_segs_now = tcp_transmit_unsent (wrk, tc, burst_size);
1977      if (max_deq > n_segs_now * tc->snd_mss)
1978	tcp_program_retransmit (tc);
1979      n_segs += n_segs_now;
1980    }
1981
1982done:
1983  tcp_fastrecovery_first_off (tc);
1984  return n_segs;
1985}
1986
1987/**
1988 * Do fast retransmit
1989 */
1990static int
1991tcp_retransmit (tcp_worker_ctx_t * wrk, tcp_connection_t * tc, u32 burst_size)
1992{
1993  if (tcp_opts_sack_permitted (&tc->rcv_opts))
1994    return tcp_retransmit_sack (wrk, tc, burst_size);
1995  else
1996    return tcp_retransmit_no_sack (wrk, tc, burst_size);
1997}
1998
1999static int
2000tcp_send_acks (tcp_connection_t * tc, u32 max_burst_size)
2001{
2002  int j, n_acks;
2003
2004  if (!tc->pending_dupacks)
2005    {
2006      tcp_send_ack (tc);
2007      return 1;
2008    }
2009
2010  /* If we're supposed to send dupacks but have no ooo data
2011   * send only one ack */
2012  if (!vec_len (tc->snd_sacks))
2013    {
2014      tcp_send_ack (tc);
2015      return 1;
2016    }
2017
2018  /* Start with first sack block */
2019  tc->snd_sack_pos = 0;
2020
2021  /* Generate enough dupacks to cover all sack blocks. Do not generate
2022   * more sacks than the number of packets received. But do generate at
2023   * least 3, i.e., the number needed to signal congestion, if needed. */
2024  n_acks = vec_len (tc->snd_sacks) / TCP_OPTS_MAX_SACK_BLOCKS;
2025  n_acks = clib_min (n_acks, tc->pending_dupacks);
2026  n_acks = clib_max (n_acks, clib_min (tc->pending_dupacks, 3));
2027  for (j = 0; j < clib_min (n_acks, max_burst_size); j++)
2028    tcp_send_ack (tc);
2029
2030  if (n_acks < max_burst_size)
2031    {
2032      tc->pending_dupacks = 0;
2033      tc->snd_sack_pos = 0;
2034      tc->dupacks_out += n_acks;
2035      return n_acks;
2036    }
2037  else
2038    {
2039      TCP_DBG ("constrained by burst size");
2040      tc->pending_dupacks = n_acks - max_burst_size;
2041      tc->dupacks_out += max_burst_size;
2042      tcp_program_dupack (tc);
2043      return max_burst_size;
2044    }
2045}
2046
2047static int
2048tcp_do_retransmit (tcp_connection_t * tc, u32 max_burst_size)
2049{
2050  u32 n_segs = 0, burst_size, sent_bytes, burst_bytes;
2051  tcp_worker_ctx_t *wrk;
2052
2053  wrk = tcp_get_worker (tc->c_thread_index);
2054  burst_bytes = transport_connection_tx_pacer_burst (&tc->connection,
2055						     wrk->vm->
2056						     clib_time.last_cpu_time);
2057  burst_size = clib_min (max_burst_size, burst_bytes / tc->snd_mss);
2058  if (!burst_size)
2059    {
2060      tcp_program_retransmit (tc);
2061      return 0;
2062    }
2063
2064  n_segs = tcp_retransmit (wrk, tc, burst_size);
2065  sent_bytes = clib_min (n_segs * tc->snd_mss, burst_bytes);
2066  transport_connection_tx_pacer_update_bytes (&tc->connection, sent_bytes);
2067  return n_segs;
2068}
2069
2070int
2071tcp_session_custom_tx (void *conn, u32 max_burst_size)
2072{
2073  tcp_connection_t *tc = (tcp_connection_t *) conn;
2074  u32 n_segs = 0;
2075
2076  if (tcp_in_cong_recovery (tc) && (tc->flags & TCP_CONN_RXT_PENDING))
2077    {
2078      tc->flags &= ~TCP_CONN_RXT_PENDING;
2079      n_segs = tcp_do_retransmit (tc, max_burst_size);
2080      max_burst_size -= n_segs;
2081    }
2082
2083  if (!(tc->flags & TCP_CONN_SNDACK))
2084    return n_segs;
2085
2086  tc->flags &= ~TCP_CONN_SNDACK;
2087
2088  /* We have retransmitted packets and no dupack */
2089  if (n_segs && !tc->pending_dupacks)
2090    return n_segs;
2091
2092  if (!max_burst_size)
2093    {
2094      tcp_program_ack (tc);
2095      return max_burst_size;
2096    }
2097
2098  n_segs += tcp_send_acks (tc, max_burst_size);
2099
2100  return n_segs;
2101}
2102#endif /* CLIB_MARCH_VARIANT */
2103
2104static void
2105tcp_output_handle_link_local (tcp_connection_t * tc0, vlib_buffer_t * b0,
2106			      u16 * next0, u32 * error0)
2107{
2108  ip_adjacency_t *adj;
2109  adj_index_t ai;
2110
2111  /* Not thread safe but as long as the connection exists the adj should
2112   * not be removed */
2113  ai = adj_nbr_find (FIB_PROTOCOL_IP6, VNET_LINK_IP6, &tc0->c_rmt_ip,
2114		     tc0->sw_if_index);
2115  if (ai == ADJ_INDEX_INVALID)
2116    {
2117      vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
2118      *next0 = TCP_OUTPUT_NEXT_DROP;
2119      *error0 = TCP_ERROR_LINK_LOCAL_RW;
2120      return;
2121    }
2122
2123  adj = adj_get (ai);
2124  if (PREDICT_TRUE (adj->lookup_next_index == IP_LOOKUP_NEXT_REWRITE))
2125    *next0 = TCP_OUTPUT_NEXT_IP_REWRITE;
2126  else if (adj->lookup_next_index == IP_LOOKUP_NEXT_ARP)
2127    *next0 = TCP_OUTPUT_NEXT_IP_ARP;
2128  else
2129    {
2130      *next0 = TCP_OUTPUT_NEXT_DROP;
2131      *error0 = TCP_ERROR_LINK_LOCAL_RW;
2132    }
2133  vnet_buffer (b0)->ip.adj_index[VLIB_TX] = ai;
2134}
2135
2136static void
2137tcp46_output_trace_frame (vlib_main_t * vm, vlib_node_runtime_t * node,
2138			  u32 * to_next, u32 n_bufs)
2139{
2140  u32 n_trace = vlib_get_trace_count (vm, node);
2141  tcp_connection_t *tc;
2142  tcp_tx_trace_t *t;
2143  vlib_buffer_t *b;
2144  tcp_header_t *th;
2145  int i;
2146
2147  for (i = 0; i < clib_min (n_trace, n_bufs); i++)
2148    {
2149      b = vlib_get_buffer (vm, to_next[i]);
2150      th = vlib_buffer_get_current (b);
2151      tc = tcp_connection_get (vnet_buffer (b)->tcp.connection_index,
2152			       vm->thread_index);
2153      t = vlib_add_trace (vm, node, b, sizeof (*t));
2154      clib_memcpy_fast (&t->tcp_header, th, sizeof (t->tcp_header));
2155      clib_memcpy_fast (&t->tcp_connection, tc, sizeof (t->tcp_connection));
2156    }
2157}
2158
2159always_inline void
2160tcp_output_push_ip (vlib_main_t * vm, vlib_buffer_t * b0,
2161		    tcp_connection_t * tc0, u8 is_ip4)
2162{
2163  tcp_header_t *th0 = 0;
2164
2165  th0 = vlib_buffer_get_current (b0);
2166  TCP_EVT (TCP_EVT_OUTPUT, tc0, th0->flags, b0->current_length);
2167  if (is_ip4)
2168    {
2169      vlib_buffer_push_ip4 (vm, b0, &tc0->c_lcl_ip4, &tc0->c_rmt_ip4,
2170			    IP_PROTOCOL_TCP, 1);
2171      b0->flags |= VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
2172      vnet_buffer (b0)->l4_hdr_offset = (u8 *) th0 - b0->data;
2173      th0->checksum = 0;
2174    }
2175  else
2176    {
2177      ip6_header_t *ih0;
2178      ih0 = vlib_buffer_push_ip6 (vm, b0, &tc0->c_lcl_ip6,
2179				  &tc0->c_rmt_ip6, IP_PROTOCOL_TCP);
2180      b0->flags |= VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
2181      vnet_buffer (b0)->l3_hdr_offset = (u8 *) ih0 - b0->data;
2182      vnet_buffer (b0)->l4_hdr_offset = (u8 *) th0 - b0->data;
2183      b0->flags |=
2184	VNET_BUFFER_F_L3_HDR_OFFSET_VALID | VNET_BUFFER_F_L4_HDR_OFFSET_VALID;
2185      th0->checksum = 0;
2186    }
2187}
2188
2189always_inline void
2190tcp_check_if_gso (tcp_connection_t * tc, vlib_buffer_t * b)
2191{
2192  if (!tc->is_tso)
2193    return;
2194  u16 data_len = b->current_length - sizeof (tcp_header_t) - tc->snd_opts_len;
2195
2196  if (PREDICT_FALSE (b->flags & VLIB_BUFFER_TOTAL_LENGTH_VALID))
2197    data_len += b->total_length_not_including_first_buffer;
2198
2199  if (PREDICT_TRUE (data_len <= tc->snd_mss))
2200    return;
2201  else
2202    {
2203      ASSERT ((b->flags & VNET_BUFFER_F_L3_HDR_OFFSET_VALID) != 0);
2204      ASSERT ((b->flags & VNET_BUFFER_F_L4_HDR_OFFSET_VALID) != 0);
2205      b->flags |= VNET_BUFFER_F_GSO;
2206      vnet_buffer2 (b)->gso_l4_hdr_sz =
2207	sizeof (tcp_header_t) + tc->snd_opts_len;
2208      vnet_buffer2 (b)->gso_size = tc->snd_mss;
2209    }
2210}
2211
2212always_inline void
2213tcp_output_handle_packet (tcp_connection_t * tc0, vlib_buffer_t * b0,
2214			  vlib_node_runtime_t * error_node, u16 * next0,
2215			  u8 is_ip4)
2216{
2217  /* If next_index is not drop use it */
2218  if (tc0->next_node_index)
2219    {
2220      *next0 = tc0->next_node_index;
2221      vnet_buffer (b0)->tcp.next_node_opaque = tc0->next_node_opaque;
2222    }
2223  else
2224    {
2225      *next0 = TCP_OUTPUT_NEXT_IP_LOOKUP;
2226    }
2227
2228  vnet_buffer (b0)->sw_if_index[VLIB_TX] = tc0->c_fib_index;
2229  vnet_buffer (b0)->sw_if_index[VLIB_RX] = 0;
2230
2231  if (!is_ip4)
2232    {
2233      u32 error0 = 0;
2234
2235      if (PREDICT_FALSE (ip6_address_is_link_local_unicast (&tc0->c_rmt_ip6)))
2236	tcp_output_handle_link_local (tc0, b0, next0, &error0);
2237
2238      if (PREDICT_FALSE (error0))
2239	{
2240	  b0->error = error_node->errors[error0];
2241	  return;
2242	}
2243    }
2244
2245  if (!TCP_ALWAYS_ACK)
2246    tcp_timer_reset (tc0, TCP_TIMER_DELACK);
2247
2248  tc0->segs_out += 1;
2249}
2250
2251always_inline uword
2252tcp46_output_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
2253		     vlib_frame_t * frame, int is_ip4)
2254{
2255  u32 n_left_from, *from, thread_index = vm->thread_index;
2256  vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b;
2257  u16 nexts[VLIB_FRAME_SIZE], *next;
2258  vlib_node_runtime_t *error_node;
2259
2260  error_node = vlib_node_get_runtime (vm, tcp_node_index (output, is_ip4));
2261
2262  from = vlib_frame_vector_args (frame);
2263  n_left_from = frame->n_vectors;
2264  tcp_set_time_now (tcp_get_worker (thread_index));
2265
2266  if (PREDICT_FALSE (node->flags & VLIB_NODE_FLAG_TRACE))
2267    tcp46_output_trace_frame (vm, node, from, n_left_from);
2268
2269  vlib_get_buffers (vm, from, bufs, n_left_from);
2270  b = bufs;
2271  next = nexts;
2272
2273  while (n_left_from >= 4)
2274    {
2275      tcp_connection_t *tc0, *tc1;
2276
2277      {
2278	vlib_prefetch_buffer_header (b[2], STORE);
2279	CLIB_PREFETCH (b[2]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2280
2281	vlib_prefetch_buffer_header (b[3], STORE);
2282	CLIB_PREFETCH (b[3]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2283      }
2284
2285      tc0 = tcp_connection_get (vnet_buffer (b[0])->tcp.connection_index,
2286				thread_index);
2287      tc1 = tcp_connection_get (vnet_buffer (b[1])->tcp.connection_index,
2288				thread_index);
2289
2290      if (PREDICT_TRUE (!tc0 + !tc1 == 0))
2291	{
2292	  tcp_output_push_ip (vm, b[0], tc0, is_ip4);
2293	  tcp_output_push_ip (vm, b[1], tc1, is_ip4);
2294
2295	  tcp_check_if_gso (tc0, b[0]);
2296	  tcp_check_if_gso (tc1, b[1]);
2297
2298	  tcp_output_handle_packet (tc0, b[0], error_node, &next[0], is_ip4);
2299	  tcp_output_handle_packet (tc1, b[1], error_node, &next[1], is_ip4);
2300	}
2301      else
2302	{
2303	  if (tc0 != 0)
2304	    {
2305	      tcp_output_push_ip (vm, b[0], tc0, is_ip4);
2306	      tcp_check_if_gso (tc0, b[0]);
2307	      tcp_output_handle_packet (tc0, b[0], error_node, &next[0],
2308					is_ip4);
2309	    }
2310	  else
2311	    {
2312	      b[0]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2313	      next[0] = TCP_OUTPUT_NEXT_DROP;
2314	    }
2315	  if (tc1 != 0)
2316	    {
2317	      tcp_output_push_ip (vm, b[1], tc1, is_ip4);
2318	      tcp_check_if_gso (tc1, b[1]);
2319	      tcp_output_handle_packet (tc1, b[1], error_node, &next[1],
2320					is_ip4);
2321	    }
2322	  else
2323	    {
2324	      b[1]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2325	      next[1] = TCP_OUTPUT_NEXT_DROP;
2326	    }
2327	}
2328
2329      b += 2;
2330      next += 2;
2331      n_left_from -= 2;
2332    }
2333  while (n_left_from > 0)
2334    {
2335      tcp_connection_t *tc0;
2336
2337      if (n_left_from > 1)
2338	{
2339	  vlib_prefetch_buffer_header (b[1], STORE);
2340	  CLIB_PREFETCH (b[1]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2341	}
2342
2343      tc0 = tcp_connection_get (vnet_buffer (b[0])->tcp.connection_index,
2344				thread_index);
2345
2346      if (PREDICT_TRUE (tc0 != 0))
2347	{
2348	  tcp_output_push_ip (vm, b[0], tc0, is_ip4);
2349	  tcp_check_if_gso (tc0, b[0]);
2350	  tcp_output_handle_packet (tc0, b[0], error_node, &next[0], is_ip4);
2351	}
2352      else
2353	{
2354	  b[0]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2355	  next[0] = TCP_OUTPUT_NEXT_DROP;
2356	}
2357
2358      b += 1;
2359      next += 1;
2360      n_left_from -= 1;
2361    }
2362
2363  vlib_buffer_enqueue_to_next (vm, node, from, nexts, frame->n_vectors);
2364  vlib_node_increment_counter (vm, tcp_node_index (output, is_ip4),
2365			       TCP_ERROR_PKTS_SENT, frame->n_vectors);
2366  return frame->n_vectors;
2367}
2368
2369VLIB_NODE_FN (tcp4_output_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2370				 vlib_frame_t * from_frame)
2371{
2372  return tcp46_output_inline (vm, node, from_frame, 1 /* is_ip4 */ );
2373}
2374
2375VLIB_NODE_FN (tcp6_output_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2376				 vlib_frame_t * from_frame)
2377{
2378  return tcp46_output_inline (vm, node, from_frame, 0 /* is_ip4 */ );
2379}
2380
2381/* *INDENT-OFF* */
2382VLIB_REGISTER_NODE (tcp4_output_node) =
2383{
2384  .name = "tcp4-output",
2385  /* Takes a vector of packets. */
2386  .vector_size = sizeof (u32),
2387  .n_errors = TCP_N_ERROR,
2388  .protocol_hint = VLIB_NODE_PROTO_HINT_TCP,
2389  .error_strings = tcp_error_strings,
2390  .n_next_nodes = TCP_OUTPUT_N_NEXT,
2391  .next_nodes = {
2392#define _(s,n) [TCP_OUTPUT_NEXT_##s] = n,
2393    foreach_tcp4_output_next
2394#undef _
2395  },
2396  .format_buffer = format_tcp_header,
2397  .format_trace = format_tcp_tx_trace,
2398};
2399/* *INDENT-ON* */
2400
2401/* *INDENT-OFF* */
2402VLIB_REGISTER_NODE (tcp6_output_node) =
2403{
2404  .name = "tcp6-output",
2405    /* Takes a vector of packets. */
2406  .vector_size = sizeof (u32),
2407  .n_errors = TCP_N_ERROR,
2408  .protocol_hint = VLIB_NODE_PROTO_HINT_TCP,
2409  .error_strings = tcp_error_strings,
2410  .n_next_nodes = TCP_OUTPUT_N_NEXT,
2411  .next_nodes = {
2412#define _(s,n) [TCP_OUTPUT_NEXT_##s] = n,
2413    foreach_tcp6_output_next
2414#undef _
2415  },
2416  .format_buffer = format_tcp_header,
2417  .format_trace = format_tcp_tx_trace,
2418};
2419/* *INDENT-ON* */
2420
2421typedef enum _tcp_reset_next
2422{
2423  TCP_RESET_NEXT_DROP,
2424  TCP_RESET_NEXT_IP_LOOKUP,
2425  TCP_RESET_N_NEXT
2426} tcp_reset_next_t;
2427
2428#define foreach_tcp4_reset_next        	\
2429  _(DROP, "error-drop")                 \
2430  _(IP_LOOKUP, "ip4-lookup")
2431
2432#define foreach_tcp6_reset_next        	\
2433  _(DROP, "error-drop")                 \
2434  _(IP_LOOKUP, "ip6-lookup")
2435
2436static uword
2437tcp46_send_reset_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
2438			 vlib_frame_t * from_frame, u8 is_ip4)
2439{
2440  u32 n_left_from, next_index, *from, *to_next;
2441  u32 my_thread_index = vm->thread_index;
2442
2443  from = vlib_frame_vector_args (from_frame);
2444  n_left_from = from_frame->n_vectors;
2445
2446  next_index = node->cached_next_index;
2447
2448  while (n_left_from > 0)
2449    {
2450      u32 n_left_to_next;
2451
2452      vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
2453
2454      while (n_left_from > 0 && n_left_to_next > 0)
2455	{
2456	  u32 bi0;
2457	  vlib_buffer_t *b0;
2458	  tcp_tx_trace_t *t0;
2459	  tcp_header_t *th0;
2460	  u32 error0 = TCP_ERROR_RST_SENT, next0 = TCP_RESET_NEXT_IP_LOOKUP;
2461
2462	  bi0 = from[0];
2463	  to_next[0] = bi0;
2464	  from += 1;
2465	  to_next += 1;
2466	  n_left_from -= 1;
2467	  n_left_to_next -= 1;
2468
2469	  b0 = vlib_get_buffer (vm, bi0);
2470
2471	  if (tcp_make_reset_in_place (vm, b0, vnet_buffer (b0)->tcp.flags,
2472				       my_thread_index, is_ip4))
2473	    {
2474	      error0 = TCP_ERROR_LOOKUP_DROPS;
2475	      next0 = TCP_RESET_NEXT_DROP;
2476	      goto done;
2477	    }
2478
2479	  /* Prepare to send to IP lookup */
2480	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
2481	  next0 = TCP_RESET_NEXT_IP_LOOKUP;
2482
2483	done:
2484	  b0->error = node->errors[error0];
2485	  b0->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
2486	  if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
2487	    {
2488	      th0 = vlib_buffer_get_current (b0);
2489	      if (is_ip4)
2490		th0 = ip4_next_header ((ip4_header_t *) th0);
2491	      else
2492		th0 = ip6_next_header ((ip6_header_t *) th0);
2493	      t0 = vlib_add_trace (vm, node, b0, sizeof (*t0));
2494	      clib_memcpy_fast (&t0->tcp_header, th0,
2495				sizeof (t0->tcp_header));
2496	    }
2497
2498	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
2499					   n_left_to_next, bi0, next0);
2500	}
2501      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
2502    }
2503  return from_frame->n_vectors;
2504}
2505
2506VLIB_NODE_FN (tcp4_reset_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2507				vlib_frame_t * from_frame)
2508{
2509  return tcp46_send_reset_inline (vm, node, from_frame, 1);
2510}
2511
2512VLIB_NODE_FN (tcp6_reset_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2513				vlib_frame_t * from_frame)
2514{
2515  return tcp46_send_reset_inline (vm, node, from_frame, 0);
2516}
2517
2518/* *INDENT-OFF* */
2519VLIB_REGISTER_NODE (tcp4_reset_node) = {
2520  .name = "tcp4-reset",
2521  .vector_size = sizeof (u32),
2522  .n_errors = TCP_N_ERROR,
2523  .error_strings = tcp_error_strings,
2524  .n_next_nodes = TCP_RESET_N_NEXT,
2525  .next_nodes = {
2526#define _(s,n) [TCP_RESET_NEXT_##s] = n,
2527    foreach_tcp4_reset_next
2528#undef _
2529  },
2530  .format_trace = format_tcp_tx_trace,
2531};
2532/* *INDENT-ON* */
2533
2534/* *INDENT-OFF* */
2535VLIB_REGISTER_NODE (tcp6_reset_node) = {
2536  .name = "tcp6-reset",
2537  .vector_size = sizeof (u32),
2538  .n_errors = TCP_N_ERROR,
2539  .error_strings = tcp_error_strings,
2540  .n_next_nodes = TCP_RESET_N_NEXT,
2541  .next_nodes = {
2542#define _(s,n) [TCP_RESET_NEXT_##s] = n,
2543    foreach_tcp6_reset_next
2544#undef _
2545  },
2546  .format_trace = format_tcp_tx_trace,
2547};
2548/* *INDENT-ON* */
2549
2550/*
2551 * fd.io coding-style-patch-verification: ON
2552 *
2553 * Local Variables:
2554 * eval: (c-set-style "gnu")
2555 * End:
2556 */
2557