tcp_output.c revision cb711a4e
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->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
409    tc->flags |= TCP_CONN_TRACK_BURST;
410
411  if (tc->snd_una == tc->snd_nxt)
412    {
413      tcp_cc_event (tc, TCP_CC_EVT_START_TX);
414      tcp_connection_tx_pacer_reset (tc, tc->cwnd, TRANSPORT_PACER_MIN_MSS);
415    }
416}
417
418#endif /* CLIB_MARCH_VARIANT */
419
420static void *
421tcp_reuse_buffer (vlib_main_t * vm, vlib_buffer_t * b)
422{
423  if (b->flags & VLIB_BUFFER_NEXT_PRESENT)
424    vlib_buffer_free_one (vm, b->next_buffer);
425  /* Zero all flags but free list index and trace flag */
426  b->flags &= VLIB_BUFFER_NEXT_PRESENT - 1;
427  b->current_data = 0;
428  b->current_length = 0;
429  b->total_length_not_including_first_buffer = 0;
430  vnet_buffer (b)->tcp.flags = 0;
431
432  /* Leave enough space for headers */
433  return vlib_buffer_make_headroom (b, TRANSPORT_MAX_HDRS_LEN);
434}
435
436#ifndef CLIB_MARCH_VARIANT
437static void *
438tcp_init_buffer (vlib_main_t * vm, vlib_buffer_t * b)
439{
440  ASSERT ((b->flags & VLIB_BUFFER_NEXT_PRESENT) == 0);
441  b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
442  b->total_length_not_including_first_buffer = 0;
443  b->current_data = 0;
444  vnet_buffer (b)->tcp.flags = 0;
445  VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b);
446  /* Leave enough space for headers */
447  return vlib_buffer_make_headroom (b, TRANSPORT_MAX_HDRS_LEN);
448}
449
450
451/* Compute TCP checksum in software when offloading is disabled for a connection */
452u16
453ip6_tcp_compute_checksum_custom (vlib_main_t * vm, vlib_buffer_t * p0,
454				 ip46_address_t * src, ip46_address_t * dst)
455{
456  ip_csum_t sum0;
457  u16 payload_length_host_byte_order;
458  u32 i;
459
460  /* Initialize checksum with ip header. */
461  sum0 = clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, p0)) +
462    clib_host_to_net_u16 (IP_PROTOCOL_TCP);
463  payload_length_host_byte_order = vlib_buffer_length_in_chain (vm, p0);
464
465  for (i = 0; i < ARRAY_LEN (src->ip6.as_uword); i++)
466    {
467      sum0 = ip_csum_with_carry
468	(sum0, clib_mem_unaligned (&src->ip6.as_uword[i], uword));
469      sum0 = ip_csum_with_carry
470	(sum0, clib_mem_unaligned (&dst->ip6.as_uword[i], uword));
471    }
472
473  return ip_calculate_l4_checksum (vm, p0, sum0,
474				   payload_length_host_byte_order, NULL, 0,
475				   NULL);
476}
477
478u16
479ip4_tcp_compute_checksum_custom (vlib_main_t * vm, vlib_buffer_t * p0,
480				 ip46_address_t * src, ip46_address_t * dst)
481{
482  ip_csum_t sum0;
483  u32 payload_length_host_byte_order;
484
485  payload_length_host_byte_order = vlib_buffer_length_in_chain (vm, p0);
486  sum0 =
487    clib_host_to_net_u32 (payload_length_host_byte_order +
488			  (IP_PROTOCOL_TCP << 16));
489
490  sum0 = ip_csum_with_carry (sum0, clib_mem_unaligned (&src->ip4, u32));
491  sum0 = ip_csum_with_carry (sum0, clib_mem_unaligned (&dst->ip4, u32));
492
493  return ip_calculate_l4_checksum (vm, p0, sum0,
494				   payload_length_host_byte_order, NULL, 0,
495				   NULL);
496}
497
498static inline u16
499tcp_compute_checksum (tcp_connection_t * tc, vlib_buffer_t * b)
500{
501  u16 checksum = 0;
502  if (PREDICT_FALSE (tc->cfg_flags & TCP_CFG_F_NO_CSUM_OFFLOAD))
503    {
504      tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
505      vlib_main_t *vm = wrk->vm;
506
507      if (tc->c_is_ip4)
508	checksum = ip4_tcp_compute_checksum_custom
509	  (vm, b, &tc->c_lcl_ip, &tc->c_rmt_ip);
510      else
511	checksum = ip6_tcp_compute_checksum_custom
512	  (vm, b, &tc->c_lcl_ip, &tc->c_rmt_ip);
513    }
514  else
515    {
516      b->flags |= VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
517    }
518  return checksum;
519}
520
521/**
522 * Prepare ACK
523 */
524static inline void
525tcp_make_ack_i (tcp_connection_t * tc, vlib_buffer_t * b, tcp_state_t state,
526		u8 flags)
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 wnd;
532
533  wnd = tcp_window_to_advertise (tc, state);
534
535  /* Make and write options */
536  tcp_opts_len = tcp_make_established_options (tc, snd_opts);
537  tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
538
539  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->snd_nxt,
540			     tc->rcv_nxt, tcp_hdr_opts_len, flags, wnd);
541
542  tcp_options_write ((u8 *) (th + 1), snd_opts);
543
544  th->checksum = tcp_compute_checksum (tc, b);
545
546  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
547
548  if (wnd == 0)
549    tcp_zero_rwnd_sent_on (tc);
550  else
551    tcp_zero_rwnd_sent_off (tc);
552}
553
554/**
555 * Convert buffer to ACK
556 */
557static inline void
558tcp_make_ack (tcp_connection_t * tc, vlib_buffer_t * b)
559{
560  tcp_make_ack_i (tc, b, TCP_STATE_ESTABLISHED, TCP_FLAG_ACK);
561  TCP_EVT (TCP_EVT_ACK_SENT, tc);
562  tc->rcv_las = tc->rcv_nxt;
563}
564
565/**
566 * Convert buffer to FIN-ACK
567 */
568void
569tcp_make_fin (tcp_connection_t * tc, vlib_buffer_t * b)
570{
571  tcp_make_ack_i (tc, b, TCP_STATE_ESTABLISHED, TCP_FLAG_FIN | TCP_FLAG_ACK);
572}
573
574/**
575 * Convert buffer to SYN
576 */
577void
578tcp_make_syn (tcp_connection_t * tc, vlib_buffer_t * b)
579{
580  u8 tcp_hdr_opts_len, tcp_opts_len;
581  tcp_header_t *th;
582  u16 initial_wnd;
583  tcp_options_t snd_opts;
584
585  initial_wnd = tcp_initial_window_to_advertise (tc);
586
587  /* Make and write options */
588  clib_memset (&snd_opts, 0, sizeof (snd_opts));
589  tcp_opts_len = tcp_make_syn_options (tc, &snd_opts);
590  tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
591
592  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->iss,
593			     tc->rcv_nxt, tcp_hdr_opts_len, TCP_FLAG_SYN,
594			     initial_wnd);
595  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
596  tcp_options_write ((u8 *) (th + 1), &snd_opts);
597  th->checksum = tcp_compute_checksum (tc, b);
598}
599
600/**
601 * Convert buffer to SYN-ACK
602 */
603void
604tcp_make_synack (tcp_connection_t * tc, vlib_buffer_t * b)
605{
606  tcp_options_t _snd_opts, *snd_opts = &_snd_opts;
607  u8 tcp_opts_len, tcp_hdr_opts_len;
608  tcp_header_t *th;
609  u16 initial_wnd;
610
611  clib_memset (snd_opts, 0, sizeof (*snd_opts));
612  initial_wnd = tcp_initial_window_to_advertise (tc);
613  tcp_opts_len = tcp_make_synack_options (tc, snd_opts);
614  tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
615
616  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->iss,
617			     tc->rcv_nxt, tcp_hdr_opts_len,
618			     TCP_FLAG_SYN | TCP_FLAG_ACK, initial_wnd);
619  tcp_options_write ((u8 *) (th + 1), snd_opts);
620
621  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
622  th->checksum = tcp_compute_checksum (tc, b);
623}
624
625always_inline void
626tcp_enqueue_to_ip_lookup_i (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
627			    u8 is_ip4, u32 fib_index, u8 flush)
628{
629  vlib_main_t *vm = wrk->vm;
630  u32 *to_next, next_index;
631  vlib_frame_t *f;
632
633  b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
634  b->error = 0;
635
636  vnet_buffer (b)->sw_if_index[VLIB_TX] = fib_index;
637  vnet_buffer (b)->sw_if_index[VLIB_RX] = 0;
638
639  /* Send to IP lookup */
640  next_index = is_ip4 ? ip4_lookup_node.index : ip6_lookup_node.index;
641  tcp_trajectory_add_start (b, 1);
642
643  f = wrk->ip_lookup_tx_frames[!is_ip4];
644  if (!f)
645    {
646      f = vlib_get_frame_to_node (vm, next_index);
647      ASSERT (f);
648      wrk->ip_lookup_tx_frames[!is_ip4] = f;
649    }
650
651  to_next = vlib_frame_vector_args (f);
652  to_next[f->n_vectors] = bi;
653  f->n_vectors += 1;
654  if (flush || f->n_vectors == VLIB_FRAME_SIZE)
655    {
656      vlib_put_frame_to_node (vm, next_index, f);
657      wrk->ip_lookup_tx_frames[!is_ip4] = 0;
658    }
659}
660
661static void
662tcp_enqueue_to_ip_lookup_now (tcp_worker_ctx_t * wrk, vlib_buffer_t * b,
663			      u32 bi, u8 is_ip4, u32 fib_index)
664{
665  tcp_enqueue_to_ip_lookup_i (wrk, b, bi, is_ip4, fib_index, 1);
666}
667
668static void
669tcp_enqueue_to_ip_lookup (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
670			  u8 is_ip4, u32 fib_index)
671{
672  tcp_enqueue_to_ip_lookup_i (wrk, b, bi, is_ip4, fib_index, 0);
673  if (wrk->vm->thread_index == 0 && vlib_num_workers ())
674    session_flush_frames_main_thread (wrk->vm);
675}
676
677always_inline void
678tcp_enqueue_to_output_i (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
679			 u8 is_ip4, u8 flush)
680{
681  u32 *to_next, next_index;
682  vlib_frame_t *f;
683
684  b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
685  b->error = 0;
686
687  /* Decide where to send the packet */
688  next_index = is_ip4 ? tcp4_output_node.index : tcp6_output_node.index;
689  tcp_trajectory_add_start (b, 2);
690
691  /* Get frame to v4/6 output node */
692  f = wrk->tx_frames[!is_ip4];
693  if (!f)
694    {
695      f = vlib_get_frame_to_node (wrk->vm, next_index);
696      ASSERT (f);
697      wrk->tx_frames[!is_ip4] = f;
698    }
699  to_next = vlib_frame_vector_args (f);
700  to_next[f->n_vectors] = bi;
701  f->n_vectors += 1;
702  if (flush || f->n_vectors == VLIB_FRAME_SIZE)
703    {
704      vlib_put_frame_to_node (wrk->vm, next_index, f);
705      wrk->tx_frames[!is_ip4] = 0;
706    }
707}
708
709static void
710tcp_enqueue_to_output (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
711		       u8 is_ip4)
712{
713  tcp_enqueue_to_output_i (wrk, b, bi, is_ip4, 0);
714}
715
716static void
717tcp_enqueue_to_output_now (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
718			   u8 is_ip4)
719{
720  tcp_enqueue_to_output_i (wrk, b, bi, is_ip4, 1);
721}
722#endif /* CLIB_MARCH_VARIANT */
723
724static int
725tcp_make_reset_in_place (vlib_main_t * vm, vlib_buffer_t * b0,
726			 tcp_state_t state, u8 thread_index, u8 is_ip4)
727{
728  ip4_header_t *ih4;
729  ip6_header_t *ih6;
730  tcp_header_t *th0;
731  ip4_address_t src_ip40, dst_ip40;
732  ip6_address_t src_ip60, dst_ip60;
733  u16 src_port, dst_port;
734  u32 tmp;
735  u32 seq, ack;
736  u8 flags;
737
738  /* Find IP and TCP headers */
739  th0 = tcp_buffer_hdr (b0);
740
741  /* Save src and dst ip */
742  if (is_ip4)
743    {
744      ih4 = vlib_buffer_get_current (b0);
745      ASSERT ((ih4->ip_version_and_header_length & 0xF0) == 0x40);
746      src_ip40.as_u32 = ih4->src_address.as_u32;
747      dst_ip40.as_u32 = ih4->dst_address.as_u32;
748    }
749  else
750    {
751      ih6 = vlib_buffer_get_current (b0);
752      ASSERT ((ih6->ip_version_traffic_class_and_flow_label & 0xF0) == 0x60);
753      clib_memcpy_fast (&src_ip60, &ih6->src_address, sizeof (ip6_address_t));
754      clib_memcpy_fast (&dst_ip60, &ih6->dst_address, sizeof (ip6_address_t));
755    }
756
757  src_port = th0->src_port;
758  dst_port = th0->dst_port;
759
760  /* Try to determine what/why we're actually resetting */
761  if (state == TCP_STATE_CLOSED)
762    {
763      if (!tcp_syn (th0))
764	return -1;
765
766      tmp = clib_net_to_host_u32 (th0->seq_number);
767
768      /* Got a SYN for no listener. */
769      flags = TCP_FLAG_RST | TCP_FLAG_ACK;
770      ack = clib_host_to_net_u32 (tmp + 1);
771      seq = 0;
772    }
773  else
774    {
775      flags = TCP_FLAG_RST;
776      seq = th0->ack_number;
777      ack = 0;
778    }
779
780  tcp_reuse_buffer (vm, b0);
781  tcp_trajectory_add_start (b0, 4);
782  th0 = vlib_buffer_push_tcp_net_order (b0, dst_port, src_port, seq, ack,
783					sizeof (tcp_header_t), flags, 0);
784
785  if (is_ip4)
786    {
787      ih4 = vlib_buffer_push_ip4 (vm, b0, &dst_ip40, &src_ip40,
788				  IP_PROTOCOL_TCP, 1);
789      th0->checksum = ip4_tcp_udp_compute_checksum (vm, b0, ih4);
790    }
791  else
792    {
793      int bogus = ~0;
794      ih6 = vlib_buffer_push_ip6 (vm, b0, &dst_ip60, &src_ip60,
795				  IP_PROTOCOL_TCP);
796      th0->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b0, ih6, &bogus);
797      ASSERT (!bogus);
798    }
799
800  return 0;
801}
802
803#ifndef CLIB_MARCH_VARIANT
804/**
805 *  Send reset without reusing existing buffer
806 *
807 *  It extracts connection info out of original packet
808 */
809void
810tcp_send_reset_w_pkt (tcp_connection_t * tc, vlib_buffer_t * pkt,
811		      u32 thread_index, u8 is_ip4)
812{
813  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
814  vlib_main_t *vm = wrk->vm;
815  vlib_buffer_t *b;
816  u32 bi, sw_if_index, fib_index;
817  u8 tcp_hdr_len, flags = 0;
818  tcp_header_t *th, *pkt_th;
819  u32 seq, ack;
820  ip4_header_t *ih4, *pkt_ih4;
821  ip6_header_t *ih6, *pkt_ih6;
822  fib_protocol_t fib_proto;
823
824  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
825    return;
826
827  b = vlib_get_buffer (vm, bi);
828  sw_if_index = vnet_buffer (pkt)->sw_if_index[VLIB_RX];
829  fib_proto = is_ip4 ? FIB_PROTOCOL_IP4 : FIB_PROTOCOL_IP6;
830  fib_index = fib_table_get_index_for_sw_if_index (fib_proto, sw_if_index);
831  tcp_init_buffer (vm, b);
832
833  /* Make and write options */
834  tcp_hdr_len = sizeof (tcp_header_t);
835
836  if (is_ip4)
837    {
838      pkt_ih4 = vlib_buffer_get_current (pkt);
839      pkt_th = ip4_next_header (pkt_ih4);
840    }
841  else
842    {
843      pkt_ih6 = vlib_buffer_get_current (pkt);
844      pkt_th = ip6_next_header (pkt_ih6);
845    }
846
847  if (tcp_ack (pkt_th))
848    {
849      flags = TCP_FLAG_RST;
850      seq = pkt_th->ack_number;
851      ack = (tc->state >= TCP_STATE_SYN_RCVD) ? tc->rcv_nxt : 0;
852    }
853  else
854    {
855      flags = TCP_FLAG_RST | TCP_FLAG_ACK;
856      seq = 0;
857      ack = clib_host_to_net_u32 (vnet_buffer (pkt)->tcp.seq_end);
858    }
859
860  th = vlib_buffer_push_tcp_net_order (b, pkt_th->dst_port, pkt_th->src_port,
861				       seq, ack, tcp_hdr_len, flags, 0);
862
863  /* Swap src and dst ip */
864  if (is_ip4)
865    {
866      ASSERT ((pkt_ih4->ip_version_and_header_length & 0xF0) == 0x40);
867      ih4 = vlib_buffer_push_ip4 (vm, b, &pkt_ih4->dst_address,
868				  &pkt_ih4->src_address, IP_PROTOCOL_TCP,
869				  tcp_csum_offload (tc));
870      th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih4);
871    }
872  else
873    {
874      int bogus = ~0;
875      ASSERT ((pkt_ih6->ip_version_traffic_class_and_flow_label & 0xF0) ==
876	      0x60);
877      ih6 = vlib_buffer_push_ip6 (vm, b, &pkt_ih6->dst_address,
878				  &pkt_ih6->src_address, IP_PROTOCOL_TCP);
879      th->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b, ih6, &bogus);
880      ASSERT (!bogus);
881    }
882
883  tcp_enqueue_to_ip_lookup_now (wrk, b, bi, is_ip4, fib_index);
884  TCP_EVT (TCP_EVT_RST_SENT, tc);
885  vlib_node_increment_counter (vm, tcp_node_index (output, tc->c_is_ip4),
886			       TCP_ERROR_RST_SENT, 1);
887}
888
889/**
890 * Build and set reset packet for connection
891 */
892void
893tcp_send_reset (tcp_connection_t * tc)
894{
895  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
896  vlib_main_t *vm = wrk->vm;
897  vlib_buffer_t *b;
898  u32 bi;
899  tcp_header_t *th;
900  u16 tcp_hdr_opts_len, advertise_wnd, opts_write_len;
901  u8 flags;
902
903  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
904    return;
905  b = vlib_get_buffer (vm, bi);
906  tcp_init_buffer (vm, b);
907
908  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
909  tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
910  advertise_wnd = tcp_window_to_advertise (tc, TCP_STATE_ESTABLISHED);
911  flags = TCP_FLAG_RST;
912  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->snd_nxt,
913			     tc->rcv_nxt, tcp_hdr_opts_len, flags,
914			     advertise_wnd);
915  opts_write_len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
916  th->checksum = tcp_compute_checksum (tc, b);
917  ASSERT (opts_write_len == tc->snd_opts_len);
918  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
919  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
920  TCP_EVT (TCP_EVT_RST_SENT, tc);
921  vlib_node_increment_counter (vm, tcp_node_index (output, tc->c_is_ip4),
922			       TCP_ERROR_RST_SENT, 1);
923}
924
925static void
926tcp_push_ip_hdr (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
927		 vlib_buffer_t * b)
928{
929  tcp_header_t *th = vlib_buffer_get_current (b);
930  vlib_main_t *vm = wrk->vm;
931  if (tc->c_is_ip4)
932    {
933      ip4_header_t *ih;
934      ih = vlib_buffer_push_ip4 (vm, b, &tc->c_lcl_ip4,
935				 &tc->c_rmt_ip4, IP_PROTOCOL_TCP,
936				 tcp_csum_offload (tc));
937      th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih);
938    }
939  else
940    {
941      ip6_header_t *ih;
942      int bogus = ~0;
943
944      ih = vlib_buffer_push_ip6 (vm, b, &tc->c_lcl_ip6,
945				 &tc->c_rmt_ip6, IP_PROTOCOL_TCP);
946      th->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b, ih, &bogus);
947      ASSERT (!bogus);
948    }
949}
950
951/**
952 *  Send SYN
953 *
954 *  Builds a SYN packet for a half-open connection and sends it to ipx_lookup.
955 *  The packet is not forwarded through tcpx_output to avoid doing lookups
956 *  in the half_open pool.
957 */
958void
959tcp_send_syn (tcp_connection_t * tc)
960{
961  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
962  vlib_main_t *vm = wrk->vm;
963  vlib_buffer_t *b;
964  u32 bi;
965
966  /*
967   * Setup retransmit and establish timers before requesting buffer
968   * such that we can return if we've ran out.
969   */
970  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN,
971		    tc->rto * TCP_TO_TIMER_TICK);
972
973  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
974    {
975      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN, 1);
976      return;
977    }
978
979  b = vlib_get_buffer (vm, bi);
980  tcp_init_buffer (vm, b);
981  tcp_make_syn (tc, b);
982
983  /* Measure RTT with this */
984  tc->rtt_ts = tcp_time_now_us (vlib_num_workers ()? 1 : 0);
985  tc->rtt_seq = tc->snd_nxt;
986  tc->rto_boff = 0;
987
988  tcp_push_ip_hdr (wrk, tc, b);
989  tcp_enqueue_to_ip_lookup (wrk, b, bi, tc->c_is_ip4, tc->c_fib_index);
990  TCP_EVT (TCP_EVT_SYN_SENT, tc);
991}
992
993void
994tcp_send_synack (tcp_connection_t * tc)
995{
996  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
997  vlib_main_t *vm = wrk->vm;
998  vlib_buffer_t *b;
999  u32 bi;
1000
1001  tcp_retransmit_timer_force_update (tc);
1002
1003  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1004    {
1005      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
1006      return;
1007    }
1008
1009  tc->rtt_ts = tcp_time_now_us (tc->c_thread_index);
1010  b = vlib_get_buffer (vm, bi);
1011  tcp_init_buffer (vm, b);
1012  tcp_make_synack (tc, b);
1013  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1014  TCP_EVT (TCP_EVT_SYNACK_SENT, tc);
1015}
1016
1017/**
1018 * Flush tx frame populated by retransmits and timer pops
1019 */
1020void
1021tcp_flush_frame_to_output (tcp_worker_ctx_t * wrk, u8 is_ip4)
1022{
1023  if (wrk->tx_frames[!is_ip4])
1024    {
1025      u32 next_index;
1026      next_index = is_ip4 ? tcp4_output_node.index : tcp6_output_node.index;
1027      vlib_put_frame_to_node (wrk->vm, next_index, wrk->tx_frames[!is_ip4]);
1028      wrk->tx_frames[!is_ip4] = 0;
1029    }
1030}
1031
1032/**
1033 * Flush ip lookup tx frames populated by timer pops
1034 */
1035static void
1036tcp_flush_frame_to_ip_lookup (tcp_worker_ctx_t * wrk, u8 is_ip4)
1037{
1038  if (wrk->ip_lookup_tx_frames[!is_ip4])
1039    {
1040      u32 next_index;
1041      next_index = is_ip4 ? ip4_lookup_node.index : ip6_lookup_node.index;
1042      vlib_put_frame_to_node (wrk->vm, next_index,
1043			      wrk->ip_lookup_tx_frames[!is_ip4]);
1044      wrk->ip_lookup_tx_frames[!is_ip4] = 0;
1045    }
1046}
1047
1048/**
1049 * Flush v4 and v6 tcp and ip-lookup tx frames for thread index
1050 */
1051void
1052tcp_flush_frames_to_output (tcp_worker_ctx_t * wrk)
1053{
1054  tcp_flush_frame_to_output (wrk, 1);
1055  tcp_flush_frame_to_output (wrk, 0);
1056  tcp_flush_frame_to_ip_lookup (wrk, 1);
1057  tcp_flush_frame_to_ip_lookup (wrk, 0);
1058}
1059
1060/**
1061 *  Send FIN
1062 */
1063void
1064tcp_send_fin (tcp_connection_t * tc)
1065{
1066  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
1067  vlib_main_t *vm = wrk->vm;
1068  vlib_buffer_t *b;
1069  u32 bi;
1070  u8 fin_snt = 0;
1071
1072  fin_snt = tc->flags & TCP_CONN_FINSNT;
1073  if (fin_snt)
1074    tc->snd_nxt -= 1;
1075
1076  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1077    {
1078      /* Out of buffers so program fin retransmit ASAP */
1079      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
1080      if (fin_snt)
1081	tc->snd_nxt += 1;
1082      else
1083	/* Make sure retransmit retries a fin not data */
1084	tc->flags |= TCP_CONN_FINSNT;
1085      return;
1086    }
1087
1088  /* If we have non-dupacks programmed, no need to send them */
1089  if ((tc->flags & TCP_CONN_SNDACK) && !tc->pending_dupacks)
1090    tc->flags &= ~TCP_CONN_SNDACK;
1091
1092  tcp_retransmit_timer_force_update (tc);
1093  b = vlib_get_buffer (vm, bi);
1094  tcp_init_buffer (vm, b);
1095  tcp_make_fin (tc, b);
1096  tcp_enqueue_to_output_now (wrk, b, bi, tc->c_is_ip4);
1097  TCP_EVT (TCP_EVT_FIN_SENT, tc);
1098  /* Account for the FIN */
1099  tc->snd_nxt += 1;
1100  if (!fin_snt)
1101    {
1102      tc->flags |= TCP_CONN_FINSNT;
1103      tc->flags &= ~TCP_CONN_FINPNDG;
1104      tc->snd_una_max = seq_max (tc->snd_una_max, tc->snd_nxt);
1105    }
1106}
1107
1108/**
1109 * Push TCP header and update connection variables. Should only be called
1110 * for segments with data, not for 'control' packets.
1111 */
1112always_inline void
1113tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b, u32 snd_nxt,
1114		u8 compute_opts, u8 maybe_burst, u8 update_snd_nxt)
1115{
1116  u8 tcp_hdr_opts_len, flags = TCP_FLAG_ACK;
1117  u32 advertise_wnd, data_len;
1118  tcp_main_t *tm = &tcp_main;
1119  tcp_header_t *th;
1120
1121  data_len = b->current_length;
1122  if (PREDICT_FALSE (b->flags & VLIB_BUFFER_NEXT_PRESENT))
1123    data_len += b->total_length_not_including_first_buffer;
1124
1125  vnet_buffer (b)->tcp.flags = 0;
1126  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
1127
1128  if (compute_opts)
1129    tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
1130
1131  tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
1132
1133  if (maybe_burst)
1134    advertise_wnd = tc->rcv_wnd >> tc->rcv_wscale;
1135  else
1136    advertise_wnd = tcp_window_to_advertise (tc, TCP_STATE_ESTABLISHED);
1137
1138  if (PREDICT_FALSE (tc->flags & TCP_CONN_PSH_PENDING))
1139    {
1140      if (seq_geq (tc->psh_seq, snd_nxt)
1141	  && seq_lt (tc->psh_seq, snd_nxt + data_len))
1142	flags |= TCP_FLAG_PSH;
1143    }
1144  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, snd_nxt,
1145			     tc->rcv_nxt, tcp_hdr_opts_len, flags,
1146			     advertise_wnd);
1147
1148  if (maybe_burst)
1149    {
1150      clib_memcpy_fast ((u8 *) (th + 1),
1151			tm->wrk_ctx[tc->c_thread_index].cached_opts,
1152			tc->snd_opts_len);
1153    }
1154  else
1155    {
1156      u8 len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
1157      ASSERT (len == tc->snd_opts_len);
1158    }
1159
1160  /*
1161   * Update connection variables
1162   */
1163
1164  if (update_snd_nxt)
1165    tc->snd_nxt += data_len;
1166  tc->rcv_las = tc->rcv_nxt;
1167
1168  tc->bytes_out += data_len;
1169  tc->data_segs_out += 1;
1170
1171  th->checksum = tcp_compute_checksum (tc, b);
1172
1173  TCP_EVT (TCP_EVT_PKTIZE, tc);
1174}
1175
1176u32
1177tcp_session_push_header (transport_connection_t * tconn, vlib_buffer_t * b)
1178{
1179  tcp_connection_t *tc = (tcp_connection_t *) tconn;
1180
1181  if (tc->flags & TCP_CONN_TRACK_BURST)
1182    {
1183      tcp_bt_check_app_limited (tc);
1184      tcp_bt_track_tx (tc);
1185      tc->flags &= ~TCP_CONN_TRACK_BURST;
1186    }
1187
1188  tcp_push_hdr_i (tc, b, tc->snd_nxt, /* compute opts */ 0, /* burst */ 1,
1189		  /* update_snd_nxt */ 1);
1190
1191  tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
1192  tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
1193  /* If not tracking an ACK, start tracking */
1194  if (tc->rtt_ts == 0 && !tcp_in_cong_recovery (tc))
1195    {
1196      tc->rtt_ts = tcp_time_now_us (tc->c_thread_index);
1197      tc->rtt_seq = tc->snd_nxt;
1198    }
1199  if (PREDICT_FALSE (!tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT)))
1200    {
1201      tcp_retransmit_timer_set (tc);
1202      tc->rto_boff = 0;
1203    }
1204  tcp_trajectory_add_start (b, 3);
1205  return 0;
1206}
1207
1208void
1209tcp_send_ack (tcp_connection_t * tc)
1210{
1211  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
1212  vlib_main_t *vm = wrk->vm;
1213  vlib_buffer_t *b;
1214  u32 bi;
1215
1216  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1217    {
1218      tcp_update_rcv_wnd (tc);
1219      return;
1220    }
1221  b = vlib_get_buffer (vm, bi);
1222  tcp_init_buffer (vm, b);
1223  tcp_make_ack (tc, b);
1224  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1225}
1226
1227void
1228tcp_program_ack (tcp_connection_t * tc)
1229{
1230  if (!(tc->flags & TCP_CONN_SNDACK))
1231    {
1232      session_add_self_custom_tx_evt (&tc->connection, 1);
1233      tc->flags |= TCP_CONN_SNDACK;
1234    }
1235}
1236
1237void
1238tcp_program_dupack (tcp_connection_t * tc)
1239{
1240  if (!(tc->flags & TCP_CONN_SNDACK))
1241    {
1242      session_add_self_custom_tx_evt (&tc->connection, 1);
1243      tc->flags |= TCP_CONN_SNDACK;
1244    }
1245  if (tc->pending_dupacks < 255)
1246    tc->pending_dupacks += 1;
1247}
1248
1249void
1250tcp_program_retransmit (tcp_connection_t * tc)
1251{
1252  if (!(tc->flags & TCP_CONN_RXT_PENDING))
1253    {
1254      session_add_self_custom_tx_evt (&tc->connection, 0);
1255      tc->flags |= TCP_CONN_RXT_PENDING;
1256    }
1257}
1258
1259/**
1260 * Delayed ack timer handler
1261 *
1262 * Sends delayed ACK when timer expires
1263 */
1264void
1265tcp_timer_delack_handler (u32 index)
1266{
1267  u32 thread_index = vlib_get_thread_index ();
1268  tcp_connection_t *tc;
1269
1270  tc = tcp_connection_get (index, thread_index);
1271  tc->timers[TCP_TIMER_DELACK] = TCP_TIMER_HANDLE_INVALID;
1272  tcp_send_ack (tc);
1273}
1274
1275/**
1276 * Send window update ack
1277 *
1278 * Ensures that it will be sent only once, after a zero rwnd has been
1279 * advertised in a previous ack, and only if rwnd has grown beyond a
1280 * configurable value.
1281 */
1282void
1283tcp_send_window_update_ack (tcp_connection_t * tc)
1284{
1285  if (tcp_zero_rwnd_sent (tc))
1286    {
1287      tcp_update_rcv_wnd (tc);
1288      if (tc->rcv_wnd >= tcp_cfg.rwnd_min_update_ack * tc->snd_mss)
1289	{
1290	  tcp_zero_rwnd_sent_off (tc);
1291	  tcp_program_ack (tc);
1292	}
1293    }
1294}
1295
1296/**
1297 * Allocate a new buffer and build a new tcp segment
1298 *
1299 * @param wrk		tcp worker
1300 * @param tc		connection for which the segment will be allocated
1301 * @param offset	offset of the first byte in the tx fifo
1302 * @param max_deq_byte	segment size
1303 * @param[out] b	pointer to buffer allocated
1304 *
1305 * @return 	the number of bytes in the segment or 0 if buffer cannot be
1306 * 		allocated or no data available
1307 */
1308static int
1309tcp_prepare_segment (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1310		     u32 offset, u32 max_deq_bytes, vlib_buffer_t ** b)
1311{
1312  u32 bytes_per_buffer = vnet_get_tcp_main ()->bytes_per_buffer;
1313  vlib_main_t *vm = wrk->vm;
1314  u32 bi, seg_size;
1315  int n_bytes = 0;
1316  u8 *data;
1317
1318  seg_size = max_deq_bytes + TRANSPORT_MAX_HDRS_LEN;
1319
1320  /*
1321   * Prepare options
1322   */
1323  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
1324
1325  /*
1326   * Allocate and fill in buffer(s)
1327   */
1328
1329  /* Easy case, buffer size greater than mss */
1330  if (PREDICT_TRUE (seg_size <= bytes_per_buffer))
1331    {
1332      if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1333	return 0;
1334      *b = vlib_get_buffer (vm, bi);
1335      data = tcp_init_buffer (vm, *b);
1336      n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1337					    max_deq_bytes);
1338      ASSERT (n_bytes == max_deq_bytes);
1339      b[0]->current_length = n_bytes;
1340      tcp_push_hdr_i (tc, *b, tc->snd_una + offset, /* compute opts */ 0,
1341		      /* burst */ 0, /* update_snd_nxt */ 0);
1342    }
1343  /* Split mss into multiple buffers */
1344  else
1345    {
1346      u32 chain_bi = ~0, n_bufs_per_seg, n_bufs;
1347      u16 n_peeked, len_to_deq;
1348      vlib_buffer_t *chain_b, *prev_b;
1349      int i;
1350
1351      /* Make sure we have enough buffers */
1352      n_bufs_per_seg = ceil ((double) seg_size / bytes_per_buffer);
1353      vec_validate_aligned (wrk->tx_buffers, n_bufs_per_seg - 1,
1354			    CLIB_CACHE_LINE_BYTES);
1355      n_bufs = vlib_buffer_alloc (vm, wrk->tx_buffers, n_bufs_per_seg);
1356      if (PREDICT_FALSE (n_bufs != n_bufs_per_seg))
1357	{
1358	  if (n_bufs)
1359	    vlib_buffer_free (vm, wrk->tx_buffers, n_bufs);
1360	  return 0;
1361	}
1362
1363      *b = vlib_get_buffer (vm, wrk->tx_buffers[--n_bufs]);
1364      data = tcp_init_buffer (vm, *b);
1365      n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1366					    bytes_per_buffer -
1367					    TRANSPORT_MAX_HDRS_LEN);
1368      b[0]->current_length = n_bytes;
1369      b[0]->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
1370      b[0]->total_length_not_including_first_buffer = 0;
1371      max_deq_bytes -= n_bytes;
1372
1373      chain_b = *b;
1374      for (i = 1; i < n_bufs_per_seg; i++)
1375	{
1376	  prev_b = chain_b;
1377	  len_to_deq = clib_min (max_deq_bytes, bytes_per_buffer);
1378	  chain_bi = wrk->tx_buffers[--n_bufs];
1379	  chain_b = vlib_get_buffer (vm, chain_bi);
1380	  chain_b->current_data = 0;
1381	  data = vlib_buffer_get_current (chain_b);
1382	  n_peeked = session_tx_fifo_peek_bytes (&tc->connection, data,
1383						 offset + n_bytes,
1384						 len_to_deq);
1385	  ASSERT (n_peeked == len_to_deq);
1386	  n_bytes += n_peeked;
1387	  chain_b->current_length = n_peeked;
1388	  chain_b->next_buffer = 0;
1389
1390	  /* update previous buffer */
1391	  prev_b->next_buffer = chain_bi;
1392	  prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT;
1393
1394	  max_deq_bytes -= n_peeked;
1395	  b[0]->total_length_not_including_first_buffer += n_peeked;
1396	}
1397
1398      tcp_push_hdr_i (tc, *b, tc->snd_una + offset, /* compute opts */ 0,
1399		      /* burst */ 0, /* update_snd_nxt */ 0);
1400
1401      if (PREDICT_FALSE (n_bufs))
1402	{
1403	  clib_warning ("not all buffers consumed");
1404	  vlib_buffer_free (vm, wrk->tx_buffers, n_bufs);
1405	}
1406    }
1407
1408  ASSERT (n_bytes > 0);
1409  ASSERT (((*b)->current_data + (*b)->current_length) <= bytes_per_buffer);
1410
1411  return n_bytes;
1412}
1413
1414/**
1415 * Build a retransmit segment
1416 *
1417 * @return the number of bytes in the segment or 0 if there's nothing to
1418 *         retransmit
1419 */
1420static u32
1421tcp_prepare_retransmit_segment (tcp_worker_ctx_t * wrk,
1422				tcp_connection_t * tc, u32 offset,
1423				u32 max_deq_bytes, vlib_buffer_t ** b)
1424{
1425  u32 start, available_bytes;
1426  int n_bytes = 0;
1427
1428  ASSERT (tc->state >= TCP_STATE_ESTABLISHED);
1429  ASSERT (max_deq_bytes != 0);
1430
1431  /*
1432   * Make sure we can retransmit something
1433   */
1434  available_bytes = transport_max_tx_dequeue (&tc->connection);
1435  ASSERT (available_bytes >= offset);
1436  available_bytes -= offset;
1437  if (!available_bytes)
1438    return 0;
1439
1440  max_deq_bytes = clib_min (tc->snd_mss, max_deq_bytes);
1441  max_deq_bytes = clib_min (available_bytes, max_deq_bytes);
1442
1443  start = tc->snd_una + offset;
1444  n_bytes = tcp_prepare_segment (wrk, tc, offset, max_deq_bytes, b);
1445  if (!n_bytes)
1446    return 0;
1447
1448  tc->snd_rxt_bytes += n_bytes;
1449
1450  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
1451    tcp_bt_track_rxt (tc, start, start + n_bytes);
1452
1453  tc->bytes_retrans += n_bytes;
1454  tc->segs_retrans += 1;
1455  TCP_EVT (TCP_EVT_CC_RTX, tc, offset, n_bytes);
1456
1457  return n_bytes;
1458}
1459
1460static void
1461tcp_check_sack_reneging (tcp_connection_t * tc)
1462{
1463  sack_scoreboard_t *sb = &tc->sack_sb;
1464  sack_scoreboard_hole_t *hole;
1465
1466  hole = scoreboard_first_hole (sb);
1467  if (!sb->is_reneging && (!hole || hole->start == tc->snd_una))
1468    return;
1469
1470  scoreboard_clear_reneging (sb, tc->snd_una, tc->snd_nxt);
1471}
1472
1473/**
1474 * Reset congestion control, switch cwnd to loss window and try again.
1475 */
1476static void
1477tcp_cc_init_rxt_timeout (tcp_connection_t * tc)
1478{
1479  TCP_EVT (TCP_EVT_CC_EVT, tc, 6);
1480
1481  tc->prev_ssthresh = tc->ssthresh;
1482  tc->prev_cwnd = tc->cwnd;
1483
1484  /* If we entrered loss without fast recovery, notify cc algo of the
1485   * congestion event such that it can update ssthresh and its state */
1486  if (!tcp_in_fastrecovery (tc))
1487    tcp_cc_congestion (tc);
1488
1489  /* Let cc algo decide loss cwnd and ssthresh post unrecovered loss */
1490  tcp_cc_loss (tc);
1491
1492  tc->rtt_ts = 0;
1493  tc->cwnd_acc_bytes = 0;
1494  tc->tr_occurences += 1;
1495  tcp_recovery_on (tc);
1496}
1497
1498void
1499tcp_timer_retransmit_handler (u32 tc_index)
1500{
1501  u32 thread_index = vlib_get_thread_index ();
1502  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
1503  vlib_main_t *vm = wrk->vm;
1504  tcp_connection_t *tc;
1505  vlib_buffer_t *b = 0;
1506  u32 bi, n_bytes;
1507
1508  tc = tcp_connection_get (tc_index, thread_index);
1509
1510  /* Note: the connection may have been closed and pool_put */
1511  if (PREDICT_FALSE (tc == 0 || tc->state == TCP_STATE_SYN_SENT))
1512    return;
1513
1514  tc->timers[TCP_TIMER_RETRANSMIT] = TCP_TIMER_HANDLE_INVALID;
1515
1516  /* Wait-close and retransmit could pop at the same time */
1517  if (tc->state == TCP_STATE_CLOSED)
1518    return;
1519
1520  if (tc->state >= TCP_STATE_ESTABLISHED)
1521    {
1522      TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
1523
1524      /* Lost FIN, retransmit and return */
1525      if (tc->flags & TCP_CONN_FINSNT)
1526	{
1527	  tcp_send_fin (tc);
1528	  tc->rto_boff += 1;
1529	  tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1530	  return;
1531	}
1532
1533      /* Shouldn't be here. This condition is tricky because it has to take
1534       * into account boff > 0 due to persist timeout. */
1535      if ((tc->rto_boff == 0 && tc->snd_una == tc->snd_nxt)
1536	  || (tc->rto_boff > 0 && seq_geq (tc->snd_una, tc->snd_congestion)
1537	      && !tcp_flight_size (tc)))
1538	{
1539	  ASSERT (!tcp_in_recovery (tc));
1540	  tc->rto_boff = 0;
1541	  return;
1542	}
1543
1544      /* We're not in recovery so make sure rto_boff is 0. Can be non 0 due
1545       * to persist timer timeout */
1546      if (!tcp_in_recovery (tc) && tc->rto_boff > 0)
1547	{
1548	  tc->rto_boff = 0;
1549	  tcp_update_rto (tc);
1550	}
1551
1552      /* Peer is dead or network connectivity is lost. Close connection.
1553       * RFC 1122 section 4.2.3.5 recommends a value of at least 100s. For
1554       * a min rto of 0.2s we need to retry about 8 times. */
1555      if (tc->rto_boff >= TCP_RTO_BOFF_MAX)
1556	{
1557	  tcp_send_reset (tc);
1558	  tcp_connection_set_state (tc, TCP_STATE_CLOSED);
1559	  session_transport_closing_notify (&tc->connection);
1560	  tcp_connection_timers_reset (tc);
1561	  tcp_timer_update (tc, TCP_TIMER_WAITCLOSE, tcp_cfg.closewait_time);
1562	  return;
1563	}
1564
1565      if (tcp_opts_sack_permitted (&tc->rcv_opts))
1566	tcp_check_sack_reneging (tc);
1567
1568      /* Update send congestion to make sure that rxt has data to send */
1569      tc->snd_congestion = tc->snd_nxt;
1570
1571      /* Send the first unacked segment. If we're short on buffers, return
1572       * as soon as possible */
1573      n_bytes = tcp_prepare_retransmit_segment (wrk, tc, 0, tc->snd_mss, &b);
1574      if (!n_bytes)
1575	{
1576	  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
1577	  return;
1578	}
1579
1580      bi = vlib_get_buffer_index (vm, b);
1581      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1582
1583      tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1584      tcp_retransmit_timer_force_update (tc);
1585
1586      tc->rto_boff += 1;
1587      if (tc->rto_boff == 1)
1588	{
1589	  tcp_cc_init_rxt_timeout (tc);
1590	  /* Record timestamp. Eifel detection algorithm RFC3522 */
1591	  tc->snd_rxt_ts = tcp_tstamp (tc);
1592	}
1593
1594      if (tcp_opts_sack_permitted (&tc->rcv_opts))
1595	scoreboard_init_rxt (&tc->sack_sb, tc->snd_una + n_bytes);
1596
1597      tcp_program_retransmit (tc);
1598    }
1599  /* Retransmit SYN-ACK */
1600  else if (tc->state == TCP_STATE_SYN_RCVD)
1601    {
1602      TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
1603
1604      tc->rtt_ts = 0;
1605
1606      /* Passive open establish timeout */
1607      if (tc->rto > TCP_ESTABLISH_TIME >> 1)
1608	{
1609	  tcp_connection_set_state (tc, TCP_STATE_CLOSED);
1610	  tcp_connection_timers_reset (tc);
1611	  tcp_timer_update (tc, TCP_TIMER_WAITCLOSE, tcp_cfg.cleanup_time);
1612	  return;
1613	}
1614
1615      if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1616	{
1617	  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
1618	  return;
1619	}
1620
1621      tc->rto_boff += 1;
1622      if (tc->rto_boff > TCP_RTO_SYN_RETRIES)
1623	tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1624
1625      tcp_retransmit_timer_force_update (tc);
1626
1627      b = vlib_get_buffer (vm, bi);
1628      tcp_init_buffer (vm, b);
1629      tcp_make_synack (tc, b);
1630      TCP_EVT (TCP_EVT_SYN_RXT, tc, 1);
1631
1632      /* Retransmit timer already updated, just enqueue to output */
1633      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1634    }
1635  else
1636    {
1637      ASSERT (tc->state == TCP_STATE_CLOSED);
1638      return;
1639    }
1640}
1641
1642/**
1643 * SYN retransmit timer handler. Active open only.
1644 */
1645void
1646tcp_timer_retransmit_syn_handler (u32 tc_index)
1647{
1648  u32 thread_index = vlib_get_thread_index ();
1649  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
1650  vlib_main_t *vm = wrk->vm;
1651  tcp_connection_t *tc;
1652  vlib_buffer_t *b = 0;
1653  u32 bi;
1654
1655  tc = tcp_half_open_connection_get (tc_index);
1656
1657  /* Note: the connection may have transitioned to ESTABLISHED... */
1658  if (PREDICT_FALSE (tc == 0 || tc->state != TCP_STATE_SYN_SENT))
1659    return;
1660
1661  tc->timers[TCP_TIMER_RETRANSMIT_SYN] = TCP_TIMER_HANDLE_INVALID;
1662
1663  /* Half-open connection actually moved to established but we were
1664   * waiting for syn retransmit to pop to call cleanup from the right
1665   * thread. */
1666  if (tc->flags & TCP_CONN_HALF_OPEN_DONE)
1667    {
1668      if (tcp_half_open_connection_cleanup (tc))
1669	TCP_DBG ("could not remove half-open connection");
1670      return;
1671    }
1672
1673  TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
1674  tc->rtt_ts = 0;
1675
1676  /* Active open establish timeout */
1677  if (tc->rto >= TCP_ESTABLISH_TIME >> 1)
1678    {
1679      session_stream_connect_notify (&tc->connection, 1 /* fail */ );
1680      tcp_connection_cleanup (tc);
1681      return;
1682    }
1683
1684  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1685    {
1686      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN, 1);
1687      return;
1688    }
1689
1690  /* Try without increasing RTO a number of times. If this fails,
1691   * start growing RTO exponentially */
1692  tc->rto_boff += 1;
1693  if (tc->rto_boff > TCP_RTO_SYN_RETRIES)
1694    tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1695
1696  b = vlib_get_buffer (vm, bi);
1697  tcp_init_buffer (vm, b);
1698  tcp_make_syn (tc, b);
1699
1700  TCP_EVT (TCP_EVT_SYN_RXT, tc, 0);
1701
1702  /* This goes straight to ipx_lookup */
1703  tcp_push_ip_hdr (wrk, tc, b);
1704  tcp_enqueue_to_ip_lookup (wrk, b, bi, tc->c_is_ip4, tc->c_fib_index);
1705
1706  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN,
1707		    tc->rto * TCP_TO_TIMER_TICK);
1708}
1709
1710/**
1711 * Got 0 snd_wnd from peer, try to do something about it.
1712 *
1713 */
1714void
1715tcp_timer_persist_handler (u32 index)
1716{
1717  u32 thread_index = vlib_get_thread_index ();
1718  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
1719  u32 bi, max_snd_bytes, available_bytes, offset;
1720  tcp_main_t *tm = vnet_get_tcp_main ();
1721  vlib_main_t *vm = wrk->vm;
1722  tcp_connection_t *tc;
1723  vlib_buffer_t *b;
1724  int n_bytes = 0;
1725  u8 *data;
1726
1727  tc = tcp_connection_get_if_valid (index, thread_index);
1728  if (!tc)
1729    return;
1730
1731  /* Make sure timer handle is set to invalid */
1732  tc->timers[TCP_TIMER_PERSIST] = TCP_TIMER_HANDLE_INVALID;
1733
1734  /* Problem already solved or worse */
1735  if (tc->state == TCP_STATE_CLOSED || tc->snd_wnd > tc->snd_mss
1736      || (tc->flags & TCP_CONN_FINSNT))
1737    return;
1738
1739  available_bytes = transport_max_tx_dequeue (&tc->connection);
1740  offset = tc->snd_nxt - tc->snd_una;
1741
1742  /* Reprogram persist if no new bytes available to send. We may have data
1743   * next time */
1744  if (!available_bytes)
1745    {
1746      tcp_persist_timer_set (tc);
1747      return;
1748    }
1749
1750  if (available_bytes <= offset)
1751    {
1752      ASSERT (tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT));
1753      return;
1754    }
1755
1756  /* Increment RTO backoff */
1757  tc->rto_boff += 1;
1758  tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1759
1760  /*
1761   * Try to force the first unsent segment (or buffer)
1762   */
1763  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1764    {
1765      tcp_persist_timer_set (tc);
1766      return;
1767    }
1768  b = vlib_get_buffer (vm, bi);
1769  data = tcp_init_buffer (vm, b);
1770
1771  tcp_validate_txf_size (tc, offset);
1772  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
1773  max_snd_bytes = clib_min (tc->snd_mss,
1774			    tm->bytes_per_buffer - TRANSPORT_MAX_HDRS_LEN);
1775  n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1776					max_snd_bytes);
1777  b->current_length = n_bytes;
1778  ASSERT (n_bytes != 0 && (tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT)
1779			   || tc->snd_nxt == tc->snd_una_max
1780			   || tc->rto_boff > 1));
1781
1782  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
1783    {
1784      tcp_bt_check_app_limited (tc);
1785      tcp_bt_track_tx (tc);
1786    }
1787
1788  tcp_push_hdr_i (tc, b, tc->snd_nxt, /* compute opts */ 0,
1789		  /* burst */ 0, /* update_snd_nxt */ 1);
1790  tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
1791  tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
1792  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1793
1794  /* Just sent new data, enable retransmit */
1795  tcp_retransmit_timer_update (tc);
1796}
1797
1798/**
1799 * Retransmit first unacked segment
1800 */
1801int
1802tcp_retransmit_first_unacked (tcp_worker_ctx_t * wrk, tcp_connection_t * tc)
1803{
1804  vlib_main_t *vm = wrk->vm;
1805  vlib_buffer_t *b;
1806  u32 bi, n_bytes;
1807
1808  TCP_EVT (TCP_EVT_CC_EVT, tc, 1);
1809
1810  n_bytes = tcp_prepare_retransmit_segment (wrk, tc, 0, tc->snd_mss, &b);
1811  if (!n_bytes)
1812    return -1;
1813
1814  bi = vlib_get_buffer_index (vm, b);
1815  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1816
1817  return 0;
1818}
1819
1820static int
1821tcp_transmit_unsent (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1822		     u32 burst_size)
1823{
1824  u32 offset, n_segs = 0, n_written, bi, available_wnd;
1825  vlib_main_t *vm = wrk->vm;
1826  vlib_buffer_t *b = 0;
1827
1828  offset = tc->snd_nxt - tc->snd_una;
1829  available_wnd = tc->snd_wnd - offset;
1830  burst_size = clib_min (burst_size, available_wnd / tc->snd_mss);
1831
1832  while (n_segs < burst_size)
1833    {
1834      n_written = tcp_prepare_segment (wrk, tc, offset, tc->snd_mss, &b);
1835      if (!n_written)
1836	goto done;
1837
1838      bi = vlib_get_buffer_index (vm, b);
1839      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1840      offset += n_written;
1841      n_segs += 1;
1842
1843      tc->snd_nxt += n_written;
1844      tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
1845    }
1846
1847done:
1848  return n_segs;
1849}
1850
1851/**
1852 * Estimate send space using proportional rate reduction (RFC6937)
1853 */
1854int
1855tcp_fastrecovery_prr_snd_space (tcp_connection_t * tc)
1856{
1857  u32 pipe, prr_out;
1858  int space;
1859
1860  pipe = tcp_flight_size (tc);
1861  prr_out = tc->snd_rxt_bytes + (tc->snd_nxt - tc->snd_congestion);
1862
1863  if (pipe > tc->ssthresh)
1864    {
1865      space = ((int) tc->prr_delivered * ((f64) tc->ssthresh / tc->prev_cwnd))
1866	- prr_out;
1867    }
1868  else
1869    {
1870      int limit;
1871      limit = clib_max ((int) (tc->prr_delivered - prr_out), 0) + tc->snd_mss;
1872      space = clib_min (tc->ssthresh - pipe, limit);
1873    }
1874  space = clib_max (space, prr_out ? 0 : tc->snd_mss);
1875  return space;
1876}
1877
1878static inline u8
1879tcp_retransmit_should_retry_head (tcp_connection_t * tc,
1880				  sack_scoreboard_t * sb)
1881{
1882  u32 tx_adv_sack = sb->high_sacked - tc->snd_congestion;
1883  f64 rr = (f64) tc->ssthresh / tc->prev_cwnd;
1884
1885  return (tx_adv_sack > (tc->snd_una - tc->prr_start) * rr);
1886}
1887
1888static inline u8
1889tcp_max_tx_deq (tcp_connection_t * tc)
1890{
1891  return (transport_max_tx_dequeue (&tc->connection)
1892	  - (tc->snd_nxt - tc->snd_una));
1893}
1894
1895#define scoreboard_rescue_rxt_valid(_sb, _tc)			\
1896    (seq_geq (_sb->rescue_rxt, _tc->snd_una) 			\
1897	&& seq_leq (_sb->rescue_rxt, _tc->snd_congestion))
1898
1899/**
1900 * Do retransmit with SACKs
1901 */
1902static int
1903tcp_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1904		     u32 burst_size)
1905{
1906  u8 snd_limited = 0, can_rescue = 0, reset_pacer = 0;
1907  u32 n_written = 0, offset, max_bytes, n_segs = 0;
1908  u32 bi, max_deq, burst_bytes, sent_bytes;
1909  sack_scoreboard_hole_t *hole;
1910  vlib_main_t *vm = wrk->vm;
1911  vlib_buffer_t *b = 0;
1912  sack_scoreboard_t *sb;
1913  int snd_space;
1914  u64 time_now;
1915
1916  ASSERT (tcp_in_cong_recovery (tc));
1917
1918  time_now = wrk->vm->clib_time.last_cpu_time;
1919  burst_bytes = transport_connection_tx_pacer_burst (&tc->connection,
1920						     time_now);
1921  burst_size = clib_min (burst_size, burst_bytes / tc->snd_mss);
1922  if (!burst_size)
1923    {
1924      tcp_program_retransmit (tc);
1925      return 0;
1926    }
1927
1928  if (tcp_in_recovery (tc))
1929    snd_space = tcp_available_cc_snd_space (tc);
1930  else
1931    snd_space = tcp_fastrecovery_prr_snd_space (tc);
1932
1933  if (snd_space < tc->snd_mss)
1934    {
1935      reset_pacer = burst_bytes > tc->snd_mss;
1936      goto done;
1937    }
1938
1939  reset_pacer = snd_space < burst_bytes;
1940
1941  sb = &tc->sack_sb;
1942
1943  /* Check if snd_una is a lost retransmit */
1944  if (pool_elts (sb->holes)
1945      && seq_gt (sb->high_sacked, tc->snd_congestion)
1946      && tc->rxt_head != tc->snd_una
1947      && tcp_retransmit_should_retry_head (tc, sb))
1948    {
1949      n_written = tcp_prepare_retransmit_segment (wrk, tc, 0, tc->snd_mss,
1950						  &b);
1951      if (!n_written)
1952	{
1953	  tcp_program_retransmit (tc);
1954	  goto done;
1955	}
1956      bi = vlib_get_buffer_index (vm, b);
1957      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1958      n_segs = 1;
1959
1960      tc->rxt_head = tc->snd_una;
1961      tc->rxt_delivered += n_written;
1962      tc->prr_delivered += n_written;
1963      ASSERT (tc->rxt_delivered <= tc->snd_rxt_bytes);
1964    }
1965
1966  TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
1967  hole = scoreboard_get_hole (sb, sb->cur_rxt_hole);
1968
1969  max_deq = transport_max_tx_dequeue (&tc->connection);
1970  max_deq -= tc->snd_nxt - tc->snd_una;
1971
1972  while (snd_space > 0 && n_segs < burst_size)
1973    {
1974      hole = scoreboard_next_rxt_hole (sb, hole, max_deq, &can_rescue,
1975				       &snd_limited);
1976      if (!hole)
1977	{
1978	  /* We are out of lost holes to retransmit so send some new data. */
1979	  if (max_deq > tc->snd_mss)
1980	    {
1981	      u32 n_segs_new;
1982	      int av_wnd;
1983
1984	      av_wnd = (int) tc->snd_wnd - (tc->snd_nxt - tc->snd_una);
1985	      av_wnd = clib_max (av_wnd, 0);
1986	      snd_space = clib_min (snd_space, av_wnd);
1987	      snd_space = clib_min (max_deq, snd_space);
1988	      burst_size = clib_min (burst_size - n_segs,
1989				     snd_space / tc->snd_mss);
1990	      burst_size = clib_min (burst_size, TCP_RXT_MAX_BURST);
1991	      n_segs_new = tcp_transmit_unsent (wrk, tc, burst_size);
1992	      if (max_deq > n_segs_new * tc->snd_mss)
1993		tcp_program_retransmit (tc);
1994
1995	      n_segs += n_segs_new;
1996	      goto done;
1997	    }
1998
1999	  if (tcp_in_recovery (tc) || !can_rescue
2000	      || scoreboard_rescue_rxt_valid (sb, tc))
2001	    break;
2002
2003	  /* If rescue rxt undefined or less than snd_una then one segment of
2004	   * up to SMSS octets that MUST include the highest outstanding
2005	   * unSACKed sequence number SHOULD be returned, and RescueRxt set to
2006	   * RecoveryPoint. HighRxt MUST NOT be updated.
2007	   */
2008	  max_bytes = clib_min (tc->snd_mss,
2009				tc->snd_congestion - tc->snd_una);
2010	  max_bytes = clib_min (max_bytes, snd_space);
2011	  offset = tc->snd_congestion - tc->snd_una - max_bytes;
2012	  sb->rescue_rxt = tc->snd_congestion;
2013	  n_written = tcp_prepare_retransmit_segment (wrk, tc, offset,
2014						      max_bytes, &b);
2015	  if (!n_written)
2016	    goto done;
2017
2018	  bi = vlib_get_buffer_index (vm, b);
2019	  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
2020	  n_segs += 1;
2021	  break;
2022	}
2023
2024      max_bytes = clib_min (hole->end - sb->high_rxt, snd_space);
2025      max_bytes = snd_limited ? clib_min (max_bytes, tc->snd_mss) : max_bytes;
2026      if (max_bytes == 0)
2027	break;
2028
2029      offset = sb->high_rxt - tc->snd_una;
2030      n_written = tcp_prepare_retransmit_segment (wrk, tc, offset, max_bytes,
2031						  &b);
2032      ASSERT (n_written <= snd_space);
2033
2034      /* Nothing left to retransmit */
2035      if (n_written == 0)
2036	break;
2037
2038      bi = vlib_get_buffer_index (vm, b);
2039      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
2040
2041      sb->high_rxt += n_written;
2042      snd_space -= n_written;
2043      n_segs += 1;
2044    }
2045
2046  if (hole)
2047    tcp_program_retransmit (tc);
2048
2049done:
2050
2051  if (reset_pacer)
2052    {
2053      transport_connection_tx_pacer_reset_bucket (&tc->connection,
2054						  vm->clib_time.
2055						  last_cpu_time);
2056    }
2057  else
2058    {
2059      sent_bytes = clib_min (n_segs * tc->snd_mss, burst_bytes);
2060      transport_connection_tx_pacer_update_bytes (&tc->connection,
2061						  sent_bytes);
2062    }
2063
2064  return n_segs;
2065}
2066
2067/**
2068 * Fast retransmit without SACK info
2069 */
2070static int
2071tcp_retransmit_no_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
2072			u32 burst_size)
2073{
2074  u32 n_written = 0, offset = 0, bi, max_deq, n_segs_now;
2075  u32 burst_bytes, sent_bytes;
2076  vlib_main_t *vm = wrk->vm;
2077  int snd_space, n_segs = 0;
2078  u8 cc_limited = 0;
2079  vlib_buffer_t *b;
2080  u64 time_now;
2081
2082  ASSERT (tcp_in_fastrecovery (tc));
2083  TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
2084
2085  time_now = wrk->vm->clib_time.last_cpu_time;
2086  burst_bytes = transport_connection_tx_pacer_burst (&tc->connection,
2087						     time_now);
2088  burst_size = clib_min (burst_size, burst_bytes / tc->snd_mss);
2089  if (!burst_size)
2090    {
2091      tcp_program_retransmit (tc);
2092      return 0;
2093    }
2094
2095  snd_space = tcp_available_cc_snd_space (tc);
2096  cc_limited = snd_space < burst_bytes;
2097
2098  if (!tcp_fastrecovery_first (tc))
2099    goto send_unsent;
2100
2101  /* RFC 6582: [If a partial ack], retransmit the first unacknowledged
2102   * segment. */
2103  while (snd_space > 0 && n_segs < burst_size)
2104    {
2105      n_written = tcp_prepare_retransmit_segment (wrk, tc, offset,
2106						  tc->snd_mss, &b);
2107
2108      /* Nothing left to retransmit */
2109      if (n_written == 0)
2110	break;
2111
2112      bi = vlib_get_buffer_index (vm, b);
2113      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
2114      snd_space -= n_written;
2115      offset += n_written;
2116      n_segs += 1;
2117    }
2118
2119  if (n_segs == burst_size)
2120    goto done;
2121
2122send_unsent:
2123
2124  /* RFC 6582: Send a new segment if permitted by the new value of cwnd. */
2125  if (snd_space < tc->snd_mss || tc->snd_mss == 0)
2126    goto done;
2127
2128  max_deq = transport_max_tx_dequeue (&tc->connection);
2129  max_deq -= tc->snd_nxt - tc->snd_una;
2130  if (max_deq)
2131    {
2132      snd_space = clib_min (max_deq, snd_space);
2133      burst_size = clib_min (burst_size - n_segs, snd_space / tc->snd_mss);
2134      n_segs_now = tcp_transmit_unsent (wrk, tc, burst_size);
2135      if (max_deq > n_segs_now * tc->snd_mss)
2136	tcp_program_retransmit (tc);
2137      n_segs += n_segs_now;
2138    }
2139
2140done:
2141  tcp_fastrecovery_first_off (tc);
2142
2143  sent_bytes = clib_min (n_segs * tc->snd_mss, burst_bytes);
2144  sent_bytes = cc_limited ? burst_bytes : sent_bytes;
2145  transport_connection_tx_pacer_update_bytes (&tc->connection, sent_bytes);
2146
2147  return n_segs;
2148}
2149
2150static int
2151tcp_send_acks (tcp_connection_t * tc, u32 max_burst_size)
2152{
2153  int j, n_acks;
2154
2155  if (!tc->pending_dupacks)
2156    {
2157      if (tcp_in_cong_recovery (tc) || !tcp_max_tx_deq (tc)
2158	  || tc->state != TCP_STATE_ESTABLISHED)
2159	{
2160	  tcp_send_ack (tc);
2161	  return 1;
2162	}
2163      return 0;
2164    }
2165
2166  /* If we're supposed to send dupacks but have no ooo data
2167   * send only one ack */
2168  if (!vec_len (tc->snd_sacks))
2169    {
2170      tcp_send_ack (tc);
2171      tc->pending_dupacks = 0;
2172      return 1;
2173    }
2174
2175  /* Start with first sack block */
2176  tc->snd_sack_pos = 0;
2177
2178  /* Generate enough dupacks to cover all sack blocks. Do not generate
2179   * more sacks than the number of packets received. But do generate at
2180   * least 3, i.e., the number needed to signal congestion, if needed. */
2181  n_acks = vec_len (tc->snd_sacks) / TCP_OPTS_MAX_SACK_BLOCKS;
2182  n_acks = clib_min (n_acks, tc->pending_dupacks);
2183  n_acks = clib_max (n_acks, clib_min (tc->pending_dupacks, 3));
2184  for (j = 0; j < clib_min (n_acks, max_burst_size); j++)
2185    tcp_send_ack (tc);
2186
2187  if (n_acks < max_burst_size)
2188    {
2189      tc->pending_dupacks = 0;
2190      tc->snd_sack_pos = 0;
2191      tc->dupacks_out += n_acks;
2192      return n_acks;
2193    }
2194  else
2195    {
2196      TCP_DBG ("constrained by burst size");
2197      tc->pending_dupacks = n_acks - max_burst_size;
2198      tc->dupacks_out += max_burst_size;
2199      tcp_program_dupack (tc);
2200      return max_burst_size;
2201    }
2202}
2203
2204static int
2205tcp_do_retransmit (tcp_connection_t * tc, u32 max_burst_size)
2206{
2207  tcp_worker_ctx_t *wrk;
2208  u32 n_segs;
2209
2210  wrk = tcp_get_worker (tc->c_thread_index);
2211
2212  if (tcp_opts_sack_permitted (&tc->rcv_opts))
2213    n_segs = tcp_retransmit_sack (wrk, tc, max_burst_size);
2214  else
2215    n_segs = tcp_retransmit_no_sack (wrk, tc, max_burst_size);
2216
2217  return n_segs;
2218}
2219
2220int
2221tcp_session_custom_tx (void *conn, u32 max_burst_size)
2222{
2223  tcp_connection_t *tc = (tcp_connection_t *) conn;
2224  u32 n_segs = 0;
2225
2226  if (tcp_in_cong_recovery (tc) && (tc->flags & TCP_CONN_RXT_PENDING))
2227    {
2228      tc->flags &= ~TCP_CONN_RXT_PENDING;
2229      n_segs = tcp_do_retransmit (tc, max_burst_size);
2230      max_burst_size -= n_segs;
2231    }
2232
2233  if (!(tc->flags & TCP_CONN_SNDACK))
2234    return n_segs;
2235
2236  tc->flags &= ~TCP_CONN_SNDACK;
2237
2238  /* We have retransmitted packets and no dupack */
2239  if (n_segs && !tc->pending_dupacks)
2240    return n_segs;
2241
2242  if (!max_burst_size)
2243    {
2244      tcp_program_ack (tc);
2245      return max_burst_size;
2246    }
2247
2248  n_segs += tcp_send_acks (tc, max_burst_size);
2249
2250  return n_segs;
2251}
2252#endif /* CLIB_MARCH_VARIANT */
2253
2254static void
2255tcp_output_handle_link_local (tcp_connection_t * tc0, vlib_buffer_t * b0,
2256			      u16 * next0, u32 * error0)
2257{
2258  ip_adjacency_t *adj;
2259  adj_index_t ai;
2260
2261  /* Not thread safe but as long as the connection exists the adj should
2262   * not be removed */
2263  ai = adj_nbr_find (FIB_PROTOCOL_IP6, VNET_LINK_IP6, &tc0->c_rmt_ip,
2264		     tc0->sw_if_index);
2265  if (ai == ADJ_INDEX_INVALID)
2266    {
2267      vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
2268      *next0 = TCP_OUTPUT_NEXT_DROP;
2269      *error0 = TCP_ERROR_LINK_LOCAL_RW;
2270      return;
2271    }
2272
2273  adj = adj_get (ai);
2274  if (PREDICT_TRUE (adj->lookup_next_index == IP_LOOKUP_NEXT_REWRITE))
2275    *next0 = TCP_OUTPUT_NEXT_IP_REWRITE;
2276  else if (adj->lookup_next_index == IP_LOOKUP_NEXT_ARP)
2277    *next0 = TCP_OUTPUT_NEXT_IP_ARP;
2278  else
2279    {
2280      *next0 = TCP_OUTPUT_NEXT_DROP;
2281      *error0 = TCP_ERROR_LINK_LOCAL_RW;
2282    }
2283  vnet_buffer (b0)->ip.adj_index[VLIB_TX] = ai;
2284}
2285
2286static void
2287tcp46_output_trace_frame (vlib_main_t * vm, vlib_node_runtime_t * node,
2288			  u32 * to_next, u32 n_bufs)
2289{
2290  u32 n_trace = vlib_get_trace_count (vm, node);
2291  tcp_connection_t *tc;
2292  tcp_tx_trace_t *t;
2293  vlib_buffer_t *b;
2294  tcp_header_t *th;
2295  int i;
2296
2297  for (i = 0; i < clib_min (n_trace, n_bufs); i++)
2298    {
2299      b = vlib_get_buffer (vm, to_next[i]);
2300      th = vlib_buffer_get_current (b);
2301      tc = tcp_connection_get (vnet_buffer (b)->tcp.connection_index,
2302			       vm->thread_index);
2303      t = vlib_add_trace (vm, node, b, sizeof (*t));
2304      clib_memcpy_fast (&t->tcp_header, th, sizeof (t->tcp_header));
2305      clib_memcpy_fast (&t->tcp_connection, tc, sizeof (t->tcp_connection));
2306    }
2307}
2308
2309always_inline void
2310tcp_output_push_ip (vlib_main_t * vm, vlib_buffer_t * b0,
2311		    tcp_connection_t * tc0, u8 is_ip4)
2312{
2313  u8 __clib_unused *ih0;
2314  tcp_header_t __clib_unused *th0 = vlib_buffer_get_current (b0);
2315
2316  TCP_EVT (TCP_EVT_OUTPUT, tc0, th0->flags, b0->current_length);
2317
2318  if (is_ip4)
2319    ih0 = vlib_buffer_push_ip4 (vm, b0, &tc0->c_lcl_ip4, &tc0->c_rmt_ip4,
2320				IP_PROTOCOL_TCP, tcp_csum_offload (tc0));
2321  else
2322    ih0 = vlib_buffer_push_ip6 (vm, b0, &tc0->c_lcl_ip6, &tc0->c_rmt_ip6,
2323				IP_PROTOCOL_TCP);
2324
2325}
2326
2327always_inline void
2328tcp_check_if_gso (tcp_connection_t * tc, vlib_buffer_t * b)
2329{
2330  if (PREDICT_TRUE (!(tc->cfg_flags & TCP_CFG_F_TSO)))
2331    return;
2332
2333  u16 data_len = b->current_length - sizeof (tcp_header_t) - tc->snd_opts_len;
2334
2335  if (PREDICT_FALSE (b->flags & VLIB_BUFFER_TOTAL_LENGTH_VALID))
2336    data_len += b->total_length_not_including_first_buffer;
2337
2338  if (PREDICT_TRUE (data_len <= tc->snd_mss))
2339    return;
2340  else
2341    {
2342      ASSERT ((b->flags & VNET_BUFFER_F_L3_HDR_OFFSET_VALID) != 0);
2343      ASSERT ((b->flags & VNET_BUFFER_F_L4_HDR_OFFSET_VALID) != 0);
2344      b->flags |= VNET_BUFFER_F_GSO;
2345      vnet_buffer2 (b)->gso_l4_hdr_sz =
2346	sizeof (tcp_header_t) + tc->snd_opts_len;
2347      vnet_buffer2 (b)->gso_size = tc->snd_mss;
2348    }
2349}
2350
2351always_inline void
2352tcp_output_handle_packet (tcp_connection_t * tc0, vlib_buffer_t * b0,
2353			  vlib_node_runtime_t * error_node, u16 * next0,
2354			  u8 is_ip4)
2355{
2356  /* If next_index is not drop use it */
2357  if (tc0->next_node_index)
2358    {
2359      *next0 = tc0->next_node_index;
2360      vnet_buffer (b0)->tcp.next_node_opaque = tc0->next_node_opaque;
2361    }
2362  else
2363    {
2364      *next0 = TCP_OUTPUT_NEXT_IP_LOOKUP;
2365    }
2366
2367  vnet_buffer (b0)->sw_if_index[VLIB_TX] = tc0->c_fib_index;
2368  vnet_buffer (b0)->sw_if_index[VLIB_RX] = 0;
2369
2370  if (!is_ip4)
2371    {
2372      u32 error0 = 0;
2373
2374      if (PREDICT_FALSE (ip6_address_is_link_local_unicast (&tc0->c_rmt_ip6)))
2375	tcp_output_handle_link_local (tc0, b0, next0, &error0);
2376
2377      if (PREDICT_FALSE (error0))
2378	{
2379	  b0->error = error_node->errors[error0];
2380	  return;
2381	}
2382    }
2383
2384  if (!TCP_ALWAYS_ACK)
2385    tcp_timer_reset (tc0, TCP_TIMER_DELACK);
2386
2387  tc0->segs_out += 1;
2388}
2389
2390always_inline uword
2391tcp46_output_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
2392		     vlib_frame_t * frame, int is_ip4)
2393{
2394  u32 n_left_from, *from, thread_index = vm->thread_index;
2395  vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b;
2396  u16 nexts[VLIB_FRAME_SIZE], *next;
2397  vlib_node_runtime_t *error_node;
2398
2399  error_node = vlib_node_get_runtime (vm, tcp_node_index (output, is_ip4));
2400
2401  from = vlib_frame_vector_args (frame);
2402  n_left_from = frame->n_vectors;
2403  tcp_set_time_now (tcp_get_worker (thread_index));
2404
2405  if (PREDICT_FALSE (node->flags & VLIB_NODE_FLAG_TRACE))
2406    tcp46_output_trace_frame (vm, node, from, n_left_from);
2407
2408  vlib_get_buffers (vm, from, bufs, n_left_from);
2409  b = bufs;
2410  next = nexts;
2411
2412  while (n_left_from >= 4)
2413    {
2414      tcp_connection_t *tc0, *tc1;
2415
2416      {
2417	vlib_prefetch_buffer_header (b[2], STORE);
2418	CLIB_PREFETCH (b[2]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2419
2420	vlib_prefetch_buffer_header (b[3], STORE);
2421	CLIB_PREFETCH (b[3]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2422      }
2423
2424      tc0 = tcp_connection_get (vnet_buffer (b[0])->tcp.connection_index,
2425				thread_index);
2426      tc1 = tcp_connection_get (vnet_buffer (b[1])->tcp.connection_index,
2427				thread_index);
2428
2429      if (PREDICT_TRUE (!tc0 + !tc1 == 0))
2430	{
2431	  tcp_output_push_ip (vm, b[0], tc0, is_ip4);
2432	  tcp_output_push_ip (vm, b[1], tc1, is_ip4);
2433
2434	  tcp_check_if_gso (tc0, b[0]);
2435	  tcp_check_if_gso (tc1, b[1]);
2436
2437	  tcp_output_handle_packet (tc0, b[0], error_node, &next[0], is_ip4);
2438	  tcp_output_handle_packet (tc1, b[1], error_node, &next[1], is_ip4);
2439	}
2440      else
2441	{
2442	  if (tc0 != 0)
2443	    {
2444	      tcp_output_push_ip (vm, b[0], tc0, is_ip4);
2445	      tcp_check_if_gso (tc0, b[0]);
2446	      tcp_output_handle_packet (tc0, b[0], error_node, &next[0],
2447					is_ip4);
2448	    }
2449	  else
2450	    {
2451	      b[0]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2452	      next[0] = TCP_OUTPUT_NEXT_DROP;
2453	    }
2454	  if (tc1 != 0)
2455	    {
2456	      tcp_output_push_ip (vm, b[1], tc1, is_ip4);
2457	      tcp_check_if_gso (tc1, b[1]);
2458	      tcp_output_handle_packet (tc1, b[1], error_node, &next[1],
2459					is_ip4);
2460	    }
2461	  else
2462	    {
2463	      b[1]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2464	      next[1] = TCP_OUTPUT_NEXT_DROP;
2465	    }
2466	}
2467
2468      b += 2;
2469      next += 2;
2470      n_left_from -= 2;
2471    }
2472  while (n_left_from > 0)
2473    {
2474      tcp_connection_t *tc0;
2475
2476      if (n_left_from > 1)
2477	{
2478	  vlib_prefetch_buffer_header (b[1], STORE);
2479	  CLIB_PREFETCH (b[1]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2480	}
2481
2482      tc0 = tcp_connection_get (vnet_buffer (b[0])->tcp.connection_index,
2483				thread_index);
2484
2485      if (PREDICT_TRUE (tc0 != 0))
2486	{
2487	  tcp_output_push_ip (vm, b[0], tc0, is_ip4);
2488	  tcp_check_if_gso (tc0, b[0]);
2489	  tcp_output_handle_packet (tc0, b[0], error_node, &next[0], is_ip4);
2490	}
2491      else
2492	{
2493	  b[0]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2494	  next[0] = TCP_OUTPUT_NEXT_DROP;
2495	}
2496
2497      b += 1;
2498      next += 1;
2499      n_left_from -= 1;
2500    }
2501
2502  vlib_buffer_enqueue_to_next (vm, node, from, nexts, frame->n_vectors);
2503  vlib_node_increment_counter (vm, tcp_node_index (output, is_ip4),
2504			       TCP_ERROR_PKTS_SENT, frame->n_vectors);
2505  return frame->n_vectors;
2506}
2507
2508VLIB_NODE_FN (tcp4_output_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2509				 vlib_frame_t * from_frame)
2510{
2511  return tcp46_output_inline (vm, node, from_frame, 1 /* is_ip4 */ );
2512}
2513
2514VLIB_NODE_FN (tcp6_output_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2515				 vlib_frame_t * from_frame)
2516{
2517  return tcp46_output_inline (vm, node, from_frame, 0 /* is_ip4 */ );
2518}
2519
2520/* *INDENT-OFF* */
2521VLIB_REGISTER_NODE (tcp4_output_node) =
2522{
2523  .name = "tcp4-output",
2524  /* Takes a vector of packets. */
2525  .vector_size = sizeof (u32),
2526  .n_errors = TCP_N_ERROR,
2527  .protocol_hint = VLIB_NODE_PROTO_HINT_TCP,
2528  .error_strings = tcp_error_strings,
2529  .n_next_nodes = TCP_OUTPUT_N_NEXT,
2530  .next_nodes = {
2531#define _(s,n) [TCP_OUTPUT_NEXT_##s] = n,
2532    foreach_tcp4_output_next
2533#undef _
2534  },
2535  .format_buffer = format_tcp_header,
2536  .format_trace = format_tcp_tx_trace,
2537};
2538/* *INDENT-ON* */
2539
2540/* *INDENT-OFF* */
2541VLIB_REGISTER_NODE (tcp6_output_node) =
2542{
2543  .name = "tcp6-output",
2544    /* Takes a vector of packets. */
2545  .vector_size = sizeof (u32),
2546  .n_errors = TCP_N_ERROR,
2547  .protocol_hint = VLIB_NODE_PROTO_HINT_TCP,
2548  .error_strings = tcp_error_strings,
2549  .n_next_nodes = TCP_OUTPUT_N_NEXT,
2550  .next_nodes = {
2551#define _(s,n) [TCP_OUTPUT_NEXT_##s] = n,
2552    foreach_tcp6_output_next
2553#undef _
2554  },
2555  .format_buffer = format_tcp_header,
2556  .format_trace = format_tcp_tx_trace,
2557};
2558/* *INDENT-ON* */
2559
2560typedef enum _tcp_reset_next
2561{
2562  TCP_RESET_NEXT_DROP,
2563  TCP_RESET_NEXT_IP_LOOKUP,
2564  TCP_RESET_N_NEXT
2565} tcp_reset_next_t;
2566
2567#define foreach_tcp4_reset_next        	\
2568  _(DROP, "error-drop")                 \
2569  _(IP_LOOKUP, "ip4-lookup")
2570
2571#define foreach_tcp6_reset_next        	\
2572  _(DROP, "error-drop")                 \
2573  _(IP_LOOKUP, "ip6-lookup")
2574
2575static uword
2576tcp46_send_reset_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
2577			 vlib_frame_t * from_frame, u8 is_ip4)
2578{
2579  u32 n_left_from, next_index, *from, *to_next;
2580  u32 my_thread_index = vm->thread_index;
2581
2582  from = vlib_frame_vector_args (from_frame);
2583  n_left_from = from_frame->n_vectors;
2584
2585  next_index = node->cached_next_index;
2586
2587  while (n_left_from > 0)
2588    {
2589      u32 n_left_to_next;
2590
2591      vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
2592
2593      while (n_left_from > 0 && n_left_to_next > 0)
2594	{
2595	  u32 bi0;
2596	  vlib_buffer_t *b0;
2597	  tcp_tx_trace_t *t0;
2598	  tcp_header_t *th0;
2599	  u32 error0 = TCP_ERROR_RST_SENT, next0 = TCP_RESET_NEXT_IP_LOOKUP;
2600
2601	  bi0 = from[0];
2602	  to_next[0] = bi0;
2603	  from += 1;
2604	  to_next += 1;
2605	  n_left_from -= 1;
2606	  n_left_to_next -= 1;
2607
2608	  b0 = vlib_get_buffer (vm, bi0);
2609
2610	  if (tcp_make_reset_in_place (vm, b0, vnet_buffer (b0)->tcp.flags,
2611				       my_thread_index, is_ip4))
2612	    {
2613	      error0 = TCP_ERROR_LOOKUP_DROPS;
2614	      next0 = TCP_RESET_NEXT_DROP;
2615	      goto done;
2616	    }
2617
2618	  /* Prepare to send to IP lookup */
2619	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
2620	  next0 = TCP_RESET_NEXT_IP_LOOKUP;
2621
2622	done:
2623	  b0->error = node->errors[error0];
2624	  b0->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
2625	  if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
2626	    {
2627	      th0 = vlib_buffer_get_current (b0);
2628	      if (is_ip4)
2629		th0 = ip4_next_header ((ip4_header_t *) th0);
2630	      else
2631		th0 = ip6_next_header ((ip6_header_t *) th0);
2632	      t0 = vlib_add_trace (vm, node, b0, sizeof (*t0));
2633	      clib_memcpy_fast (&t0->tcp_header, th0,
2634				sizeof (t0->tcp_header));
2635	    }
2636
2637	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
2638					   n_left_to_next, bi0, next0);
2639	}
2640      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
2641    }
2642  return from_frame->n_vectors;
2643}
2644
2645VLIB_NODE_FN (tcp4_reset_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2646				vlib_frame_t * from_frame)
2647{
2648  return tcp46_send_reset_inline (vm, node, from_frame, 1);
2649}
2650
2651VLIB_NODE_FN (tcp6_reset_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2652				vlib_frame_t * from_frame)
2653{
2654  return tcp46_send_reset_inline (vm, node, from_frame, 0);
2655}
2656
2657/* *INDENT-OFF* */
2658VLIB_REGISTER_NODE (tcp4_reset_node) = {
2659  .name = "tcp4-reset",
2660  .vector_size = sizeof (u32),
2661  .n_errors = TCP_N_ERROR,
2662  .error_strings = tcp_error_strings,
2663  .n_next_nodes = TCP_RESET_N_NEXT,
2664  .next_nodes = {
2665#define _(s,n) [TCP_RESET_NEXT_##s] = n,
2666    foreach_tcp4_reset_next
2667#undef _
2668  },
2669  .format_trace = format_tcp_tx_trace,
2670};
2671/* *INDENT-ON* */
2672
2673/* *INDENT-OFF* */
2674VLIB_REGISTER_NODE (tcp6_reset_node) = {
2675  .name = "tcp6-reset",
2676  .vector_size = sizeof (u32),
2677  .n_errors = TCP_N_ERROR,
2678  .error_strings = tcp_error_strings,
2679  .n_next_nodes = TCP_RESET_N_NEXT,
2680  .next_nodes = {
2681#define _(s,n) [TCP_RESET_NEXT_##s] = n,
2682    foreach_tcp6_reset_next
2683#undef _
2684  },
2685  .format_trace = format_tcp_tx_trace,
2686};
2687/* *INDENT-ON* */
2688
2689/*
2690 * fd.io coding-style-patch-verification: ON
2691 *
2692 * Local Variables:
2693 * eval: (c-set-style "gnu")
2694 * End:
2695 */
2696