tcp_output.c revision 2089c69e
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    tcp_bt_check_app_limited (tc);
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
677static void
678tcp_enqueue_to_output (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
679		       u8 is_ip4)
680{
681  session_type_t st;
682
683  b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
684  b->error = 0;
685
686  st = session_type_from_proto_and_ip (TRANSPORT_PROTO_TCP, is_ip4);
687  session_add_pending_tx_buffer (st, wrk->vm->thread_index, bi);
688}
689
690#endif /* CLIB_MARCH_VARIANT */
691
692static int
693tcp_make_reset_in_place (vlib_main_t * vm, vlib_buffer_t * b0,
694			 tcp_state_t state, u8 thread_index, u8 is_ip4)
695{
696  ip4_header_t *ih4;
697  ip6_header_t *ih6;
698  tcp_header_t *th0;
699  ip4_address_t src_ip40, dst_ip40;
700  ip6_address_t src_ip60, dst_ip60;
701  u16 src_port, dst_port;
702  u32 tmp;
703  u32 seq, ack;
704  u8 flags;
705
706  /* Find IP and TCP headers */
707  th0 = tcp_buffer_hdr (b0);
708
709  /* Save src and dst ip */
710  if (is_ip4)
711    {
712      ih4 = vlib_buffer_get_current (b0);
713      ASSERT ((ih4->ip_version_and_header_length & 0xF0) == 0x40);
714      src_ip40.as_u32 = ih4->src_address.as_u32;
715      dst_ip40.as_u32 = ih4->dst_address.as_u32;
716    }
717  else
718    {
719      ih6 = vlib_buffer_get_current (b0);
720      ASSERT ((ih6->ip_version_traffic_class_and_flow_label & 0xF0) == 0x60);
721      clib_memcpy_fast (&src_ip60, &ih6->src_address, sizeof (ip6_address_t));
722      clib_memcpy_fast (&dst_ip60, &ih6->dst_address, sizeof (ip6_address_t));
723    }
724
725  src_port = th0->src_port;
726  dst_port = th0->dst_port;
727
728  /* Try to determine what/why we're actually resetting */
729  if (state == TCP_STATE_CLOSED)
730    {
731      if (!tcp_syn (th0))
732	return -1;
733
734      tmp = clib_net_to_host_u32 (th0->seq_number);
735
736      /* Got a SYN for no listener. */
737      flags = TCP_FLAG_RST | TCP_FLAG_ACK;
738      ack = clib_host_to_net_u32 (tmp + 1);
739      seq = 0;
740    }
741  else
742    {
743      flags = TCP_FLAG_RST;
744      seq = th0->ack_number;
745      ack = 0;
746    }
747
748  tcp_reuse_buffer (vm, b0);
749  tcp_trajectory_add_start (b0, 4);
750  th0 = vlib_buffer_push_tcp_net_order (b0, dst_port, src_port, seq, ack,
751					sizeof (tcp_header_t), flags, 0);
752
753  if (is_ip4)
754    {
755      ih4 = vlib_buffer_push_ip4 (vm, b0, &dst_ip40, &src_ip40,
756				  IP_PROTOCOL_TCP, 1);
757      th0->checksum = ip4_tcp_udp_compute_checksum (vm, b0, ih4);
758    }
759  else
760    {
761      int bogus = ~0;
762      ih6 = vlib_buffer_push_ip6 (vm, b0, &dst_ip60, &src_ip60,
763				  IP_PROTOCOL_TCP);
764      th0->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b0, ih6, &bogus);
765      ASSERT (!bogus);
766    }
767
768  return 0;
769}
770
771#ifndef CLIB_MARCH_VARIANT
772/**
773 *  Send reset without reusing existing buffer
774 *
775 *  It extracts connection info out of original packet
776 */
777void
778tcp_send_reset_w_pkt (tcp_connection_t * tc, vlib_buffer_t * pkt,
779		      u32 thread_index, u8 is_ip4)
780{
781  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
782  vlib_main_t *vm = wrk->vm;
783  vlib_buffer_t *b;
784  u32 bi, sw_if_index, fib_index;
785  u8 tcp_hdr_len, flags = 0;
786  tcp_header_t *th, *pkt_th;
787  u32 seq, ack;
788  ip4_header_t *ih4, *pkt_ih4;
789  ip6_header_t *ih6, *pkt_ih6;
790  fib_protocol_t fib_proto;
791
792  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
793    return;
794
795  b = vlib_get_buffer (vm, bi);
796  sw_if_index = vnet_buffer (pkt)->sw_if_index[VLIB_RX];
797  fib_proto = is_ip4 ? FIB_PROTOCOL_IP4 : FIB_PROTOCOL_IP6;
798  fib_index = fib_table_get_index_for_sw_if_index (fib_proto, sw_if_index);
799  tcp_init_buffer (vm, b);
800
801  /* Make and write options */
802  tcp_hdr_len = sizeof (tcp_header_t);
803
804  if (is_ip4)
805    {
806      pkt_ih4 = vlib_buffer_get_current (pkt);
807      pkt_th = ip4_next_header (pkt_ih4);
808    }
809  else
810    {
811      pkt_ih6 = vlib_buffer_get_current (pkt);
812      pkt_th = ip6_next_header (pkt_ih6);
813    }
814
815  if (tcp_ack (pkt_th))
816    {
817      flags = TCP_FLAG_RST;
818      seq = pkt_th->ack_number;
819      ack = (tc->state >= TCP_STATE_SYN_RCVD) ? tc->rcv_nxt : 0;
820    }
821  else
822    {
823      flags = TCP_FLAG_RST | TCP_FLAG_ACK;
824      seq = 0;
825      ack = clib_host_to_net_u32 (vnet_buffer (pkt)->tcp.seq_end);
826    }
827
828  th = vlib_buffer_push_tcp_net_order (b, pkt_th->dst_port, pkt_th->src_port,
829				       seq, ack, tcp_hdr_len, flags, 0);
830
831  /* Swap src and dst ip */
832  if (is_ip4)
833    {
834      ASSERT ((pkt_ih4->ip_version_and_header_length & 0xF0) == 0x40);
835      ih4 = vlib_buffer_push_ip4 (vm, b, &pkt_ih4->dst_address,
836				  &pkt_ih4->src_address, IP_PROTOCOL_TCP,
837				  tcp_csum_offload (tc));
838      th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih4);
839    }
840  else
841    {
842      int bogus = ~0;
843      ASSERT ((pkt_ih6->ip_version_traffic_class_and_flow_label & 0xF0) ==
844	      0x60);
845      ih6 = vlib_buffer_push_ip6_custom (vm, b, &pkt_ih6->dst_address,
846					 &pkt_ih6->src_address,
847					 IP_PROTOCOL_TCP,
848					 tc->ipv6_flow_label);
849      th->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b, ih6, &bogus);
850      ASSERT (!bogus);
851    }
852
853  tcp_enqueue_to_ip_lookup_now (wrk, b, bi, is_ip4, fib_index);
854  TCP_EVT (TCP_EVT_RST_SENT, tc);
855  vlib_node_increment_counter (vm, tcp_node_index (output, tc->c_is_ip4),
856			       TCP_ERROR_RST_SENT, 1);
857}
858
859/**
860 * Build and set reset packet for connection
861 */
862void
863tcp_send_reset (tcp_connection_t * tc)
864{
865  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
866  vlib_main_t *vm = wrk->vm;
867  vlib_buffer_t *b;
868  u32 bi;
869  tcp_header_t *th;
870  u16 tcp_hdr_opts_len, advertise_wnd, opts_write_len;
871  u8 flags;
872
873  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
874    return;
875  b = vlib_get_buffer (vm, bi);
876  tcp_init_buffer (vm, b);
877
878  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
879  tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
880  advertise_wnd = tcp_window_to_advertise (tc, TCP_STATE_ESTABLISHED);
881  flags = TCP_FLAG_RST;
882  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->snd_nxt,
883			     tc->rcv_nxt, tcp_hdr_opts_len, flags,
884			     advertise_wnd);
885  opts_write_len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
886  th->checksum = tcp_compute_checksum (tc, b);
887  ASSERT (opts_write_len == tc->snd_opts_len);
888  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
889  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
890  TCP_EVT (TCP_EVT_RST_SENT, tc);
891  vlib_node_increment_counter (vm, tcp_node_index (output, tc->c_is_ip4),
892			       TCP_ERROR_RST_SENT, 1);
893}
894
895static void
896tcp_push_ip_hdr (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
897		 vlib_buffer_t * b)
898{
899  tcp_header_t *th = vlib_buffer_get_current (b);
900  vlib_main_t *vm = wrk->vm;
901  if (tc->c_is_ip4)
902    {
903      ip4_header_t *ih;
904      ih = vlib_buffer_push_ip4 (vm, b, &tc->c_lcl_ip4,
905				 &tc->c_rmt_ip4, IP_PROTOCOL_TCP,
906				 tcp_csum_offload (tc));
907      th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih);
908    }
909  else
910    {
911      ip6_header_t *ih;
912      int bogus = ~0;
913
914      ih = vlib_buffer_push_ip6_custom (vm, b, &tc->c_lcl_ip6,
915					&tc->c_rmt_ip6, IP_PROTOCOL_TCP,
916					tc->ipv6_flow_label);
917      th->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b, ih, &bogus);
918      ASSERT (!bogus);
919    }
920}
921
922/**
923 *  Send SYN
924 *
925 *  Builds a SYN packet for a half-open connection and sends it to ipx_lookup.
926 *  The packet is not forwarded through tcpx_output to avoid doing lookups
927 *  in the half_open pool.
928 */
929void
930tcp_send_syn (tcp_connection_t * tc)
931{
932  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
933  vlib_main_t *vm = wrk->vm;
934  vlib_buffer_t *b;
935  u32 bi;
936
937  /*
938   * Setup retransmit and establish timers before requesting buffer
939   * such that we can return if we've ran out.
940   */
941  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN,
942		    tc->rto * TCP_TO_TIMER_TICK);
943
944  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
945    {
946      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN, 1);
947      return;
948    }
949
950  b = vlib_get_buffer (vm, bi);
951  tcp_init_buffer (vm, b);
952  tcp_make_syn (tc, b);
953
954  /* Measure RTT with this */
955  tc->rtt_ts = tcp_time_now_us (vlib_num_workers ()? 1 : 0);
956  tc->rtt_seq = tc->snd_nxt;
957  tc->rto_boff = 0;
958
959  tcp_push_ip_hdr (wrk, tc, b);
960  tcp_enqueue_to_ip_lookup (wrk, b, bi, tc->c_is_ip4, tc->c_fib_index);
961  TCP_EVT (TCP_EVT_SYN_SENT, tc);
962}
963
964void
965tcp_send_synack (tcp_connection_t * tc)
966{
967  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
968  vlib_main_t *vm = wrk->vm;
969  vlib_buffer_t *b;
970  u32 bi;
971
972  tcp_retransmit_timer_force_update (tc);
973
974  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
975    {
976      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
977      return;
978    }
979
980  tc->rtt_ts = tcp_time_now_us (tc->c_thread_index);
981  b = vlib_get_buffer (vm, bi);
982  tcp_init_buffer (vm, b);
983  tcp_make_synack (tc, b);
984  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
985  TCP_EVT (TCP_EVT_SYNACK_SENT, tc);
986}
987
988/**
989 * Flush ip lookup tx frames populated by timer pops
990 */
991static void
992tcp_flush_frame_to_ip_lookup (tcp_worker_ctx_t * wrk, u8 is_ip4)
993{
994  if (wrk->ip_lookup_tx_frames[!is_ip4])
995    {
996      u32 next_index;
997      next_index = is_ip4 ? ip4_lookup_node.index : ip6_lookup_node.index;
998      vlib_put_frame_to_node (wrk->vm, next_index,
999			      wrk->ip_lookup_tx_frames[!is_ip4]);
1000      wrk->ip_lookup_tx_frames[!is_ip4] = 0;
1001    }
1002}
1003
1004/**
1005 * Flush v4 and v6 tcp and ip-lookup tx frames for thread index
1006 */
1007void
1008tcp_flush_frames_to_output (tcp_worker_ctx_t * wrk)
1009{
1010  tcp_flush_frame_to_ip_lookup (wrk, 1);
1011  tcp_flush_frame_to_ip_lookup (wrk, 0);
1012}
1013
1014/**
1015 *  Send FIN
1016 */
1017void
1018tcp_send_fin (tcp_connection_t * tc)
1019{
1020  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
1021  vlib_main_t *vm = wrk->vm;
1022  vlib_buffer_t *b;
1023  u32 bi;
1024  u8 fin_snt = 0;
1025
1026  fin_snt = tc->flags & TCP_CONN_FINSNT;
1027  if (fin_snt)
1028    tc->snd_nxt -= 1;
1029
1030  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1031    {
1032      /* Out of buffers so program fin retransmit ASAP */
1033      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
1034      if (fin_snt)
1035	tc->snd_nxt += 1;
1036      else
1037	/* Make sure retransmit retries a fin not data */
1038	tc->flags |= TCP_CONN_FINSNT;
1039      return;
1040    }
1041
1042  /* If we have non-dupacks programmed, no need to send them */
1043  if ((tc->flags & TCP_CONN_SNDACK) && !tc->pending_dupacks)
1044    tc->flags &= ~TCP_CONN_SNDACK;
1045
1046  tcp_retransmit_timer_force_update (tc);
1047  b = vlib_get_buffer (vm, bi);
1048  tcp_init_buffer (vm, b);
1049  tcp_make_fin (tc, b);
1050  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1051  TCP_EVT (TCP_EVT_FIN_SENT, tc);
1052  /* Account for the FIN */
1053  tc->snd_nxt += 1;
1054  if (!fin_snt)
1055    {
1056      tc->flags |= TCP_CONN_FINSNT;
1057      tc->flags &= ~TCP_CONN_FINPNDG;
1058      tc->snd_una_max = seq_max (tc->snd_una_max, tc->snd_nxt);
1059    }
1060}
1061
1062/**
1063 * Push TCP header and update connection variables. Should only be called
1064 * for segments with data, not for 'control' packets.
1065 */
1066always_inline void
1067tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b, u32 snd_nxt,
1068		u8 compute_opts, u8 maybe_burst, u8 update_snd_nxt)
1069{
1070  u8 tcp_hdr_opts_len, flags = TCP_FLAG_ACK;
1071  u32 advertise_wnd, data_len;
1072  tcp_main_t *tm = &tcp_main;
1073  tcp_header_t *th;
1074
1075  data_len = b->current_length;
1076  if (PREDICT_FALSE (b->flags & VLIB_BUFFER_NEXT_PRESENT))
1077    data_len += b->total_length_not_including_first_buffer;
1078
1079  vnet_buffer (b)->tcp.flags = 0;
1080  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
1081
1082  if (compute_opts)
1083    tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
1084
1085  tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
1086
1087  if (maybe_burst)
1088    advertise_wnd = tc->rcv_wnd >> tc->rcv_wscale;
1089  else
1090    advertise_wnd = tcp_window_to_advertise (tc, TCP_STATE_ESTABLISHED);
1091
1092  if (PREDICT_FALSE (tc->flags & TCP_CONN_PSH_PENDING))
1093    {
1094      if (seq_geq (tc->psh_seq, snd_nxt)
1095	  && seq_lt (tc->psh_seq, snd_nxt + data_len))
1096	flags |= TCP_FLAG_PSH;
1097    }
1098  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, snd_nxt,
1099			     tc->rcv_nxt, tcp_hdr_opts_len, flags,
1100			     advertise_wnd);
1101
1102  if (maybe_burst)
1103    {
1104      clib_memcpy_fast ((u8 *) (th + 1),
1105			tm->wrk_ctx[tc->c_thread_index].cached_opts,
1106			tc->snd_opts_len);
1107    }
1108  else
1109    {
1110      u8 len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
1111      ASSERT (len == tc->snd_opts_len);
1112    }
1113
1114  /*
1115   * Update connection variables
1116   */
1117
1118  if (update_snd_nxt)
1119    tc->snd_nxt += data_len;
1120  tc->rcv_las = tc->rcv_nxt;
1121
1122  tc->bytes_out += data_len;
1123  tc->data_segs_out += 1;
1124
1125  th->checksum = tcp_compute_checksum (tc, b);
1126
1127  TCP_EVT (TCP_EVT_PKTIZE, tc);
1128}
1129
1130always_inline u32
1131tcp_buffer_len (vlib_buffer_t * b)
1132{
1133  u32 data_len = b->current_length;
1134  if (PREDICT_FALSE (b->flags & VLIB_BUFFER_NEXT_PRESENT))
1135    data_len += b->total_length_not_including_first_buffer;
1136  return data_len;
1137}
1138
1139u32
1140tcp_session_push_header (transport_connection_t * tconn, vlib_buffer_t * b)
1141{
1142  tcp_connection_t *tc = (tcp_connection_t *) tconn;
1143
1144  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
1145    tcp_bt_track_tx (tc, tcp_buffer_len (b));
1146
1147  tcp_push_hdr_i (tc, b, tc->snd_nxt, /* compute opts */ 0, /* burst */ 1,
1148		  /* update_snd_nxt */ 1);
1149
1150  tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
1151  tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
1152  /* If not tracking an ACK, start tracking */
1153  if (tc->rtt_ts == 0 && !tcp_in_cong_recovery (tc))
1154    {
1155      tc->rtt_ts = tcp_time_now_us (tc->c_thread_index);
1156      tc->rtt_seq = tc->snd_nxt;
1157    }
1158  if (PREDICT_FALSE (!tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT)))
1159    {
1160      tcp_retransmit_timer_set (tc);
1161      tc->rto_boff = 0;
1162    }
1163  tcp_trajectory_add_start (b, 3);
1164  return 0;
1165}
1166
1167void
1168tcp_send_ack (tcp_connection_t * tc)
1169{
1170  tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
1171  vlib_main_t *vm = wrk->vm;
1172  vlib_buffer_t *b;
1173  u32 bi;
1174
1175  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1176    {
1177      tcp_update_rcv_wnd (tc);
1178      return;
1179    }
1180  b = vlib_get_buffer (vm, bi);
1181  tcp_init_buffer (vm, b);
1182  tcp_make_ack (tc, b);
1183  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1184}
1185
1186void
1187tcp_program_ack (tcp_connection_t * tc)
1188{
1189  if (!(tc->flags & TCP_CONN_SNDACK))
1190    {
1191      session_add_self_custom_tx_evt (&tc->connection, 1);
1192      tc->flags |= TCP_CONN_SNDACK;
1193    }
1194}
1195
1196void
1197tcp_program_dupack (tcp_connection_t * tc)
1198{
1199  if (!(tc->flags & TCP_CONN_SNDACK))
1200    {
1201      session_add_self_custom_tx_evt (&tc->connection, 1);
1202      tc->flags |= TCP_CONN_SNDACK;
1203    }
1204  if (tc->pending_dupacks < 255)
1205    tc->pending_dupacks += 1;
1206}
1207
1208void
1209tcp_program_retransmit (tcp_connection_t * tc)
1210{
1211  if (!(tc->flags & TCP_CONN_RXT_PENDING))
1212    {
1213      session_add_self_custom_tx_evt (&tc->connection, 0);
1214      tc->flags |= TCP_CONN_RXT_PENDING;
1215    }
1216}
1217
1218/**
1219 * Delayed ack timer handler
1220 *
1221 * Sends delayed ACK when timer expires
1222 */
1223void
1224tcp_timer_delack_handler (u32 index)
1225{
1226  u32 thread_index = vlib_get_thread_index ();
1227  tcp_connection_t *tc;
1228
1229  tc = tcp_connection_get (index, thread_index);
1230  tc->timers[TCP_TIMER_DELACK] = TCP_TIMER_HANDLE_INVALID;
1231  tcp_send_ack (tc);
1232}
1233
1234/**
1235 * Send window update ack
1236 *
1237 * Ensures that it will be sent only once, after a zero rwnd has been
1238 * advertised in a previous ack, and only if rwnd has grown beyond a
1239 * configurable value.
1240 */
1241void
1242tcp_send_window_update_ack (tcp_connection_t * tc)
1243{
1244  if (tcp_zero_rwnd_sent (tc))
1245    {
1246      tcp_update_rcv_wnd (tc);
1247      if (tc->rcv_wnd >= tcp_cfg.rwnd_min_update_ack * tc->snd_mss)
1248	{
1249	  tcp_zero_rwnd_sent_off (tc);
1250	  tcp_program_ack (tc);
1251	}
1252    }
1253}
1254
1255/**
1256 * Allocate a new buffer and build a new tcp segment
1257 *
1258 * @param wrk		tcp worker
1259 * @param tc		connection for which the segment will be allocated
1260 * @param offset	offset of the first byte in the tx fifo
1261 * @param max_deq_byte	segment size
1262 * @param[out] b	pointer to buffer allocated
1263 *
1264 * @return 	the number of bytes in the segment or 0 if buffer cannot be
1265 * 		allocated or no data available
1266 */
1267static int
1268tcp_prepare_segment (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1269		     u32 offset, u32 max_deq_bytes, vlib_buffer_t ** b)
1270{
1271  u32 bytes_per_buffer = vnet_get_tcp_main ()->bytes_per_buffer;
1272  vlib_main_t *vm = wrk->vm;
1273  u32 bi, seg_size;
1274  int n_bytes = 0;
1275  u8 *data;
1276
1277  seg_size = max_deq_bytes + TRANSPORT_MAX_HDRS_LEN;
1278
1279  /*
1280   * Prepare options
1281   */
1282  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
1283
1284  /*
1285   * Allocate and fill in buffer(s)
1286   */
1287
1288  /* Easy case, buffer size greater than mss */
1289  if (PREDICT_TRUE (seg_size <= bytes_per_buffer))
1290    {
1291      if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1292	return 0;
1293      *b = vlib_get_buffer (vm, bi);
1294      data = tcp_init_buffer (vm, *b);
1295      n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1296					    max_deq_bytes);
1297      ASSERT (n_bytes == max_deq_bytes);
1298      b[0]->current_length = n_bytes;
1299      tcp_push_hdr_i (tc, *b, tc->snd_una + offset, /* compute opts */ 0,
1300		      /* burst */ 0, /* update_snd_nxt */ 0);
1301    }
1302  /* Split mss into multiple buffers */
1303  else
1304    {
1305      u32 chain_bi = ~0, n_bufs_per_seg, n_bufs;
1306      u16 n_peeked, len_to_deq;
1307      vlib_buffer_t *chain_b, *prev_b;
1308      int i;
1309
1310      /* Make sure we have enough buffers */
1311      n_bufs_per_seg = ceil ((double) seg_size / bytes_per_buffer);
1312      vec_validate_aligned (wrk->tx_buffers, n_bufs_per_seg - 1,
1313			    CLIB_CACHE_LINE_BYTES);
1314      n_bufs = vlib_buffer_alloc (vm, wrk->tx_buffers, n_bufs_per_seg);
1315      if (PREDICT_FALSE (n_bufs != n_bufs_per_seg))
1316	{
1317	  if (n_bufs)
1318	    vlib_buffer_free (vm, wrk->tx_buffers, n_bufs);
1319	  return 0;
1320	}
1321
1322      *b = vlib_get_buffer (vm, wrk->tx_buffers[--n_bufs]);
1323      data = tcp_init_buffer (vm, *b);
1324      n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1325					    bytes_per_buffer -
1326					    TRANSPORT_MAX_HDRS_LEN);
1327      b[0]->current_length = n_bytes;
1328      b[0]->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
1329      b[0]->total_length_not_including_first_buffer = 0;
1330      max_deq_bytes -= n_bytes;
1331
1332      chain_b = *b;
1333      for (i = 1; i < n_bufs_per_seg; i++)
1334	{
1335	  prev_b = chain_b;
1336	  len_to_deq = clib_min (max_deq_bytes, bytes_per_buffer);
1337	  chain_bi = wrk->tx_buffers[--n_bufs];
1338	  chain_b = vlib_get_buffer (vm, chain_bi);
1339	  chain_b->current_data = 0;
1340	  data = vlib_buffer_get_current (chain_b);
1341	  n_peeked = session_tx_fifo_peek_bytes (&tc->connection, data,
1342						 offset + n_bytes,
1343						 len_to_deq);
1344	  ASSERT (n_peeked == len_to_deq);
1345	  n_bytes += n_peeked;
1346	  chain_b->current_length = n_peeked;
1347	  chain_b->next_buffer = 0;
1348
1349	  /* update previous buffer */
1350	  prev_b->next_buffer = chain_bi;
1351	  prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT;
1352
1353	  max_deq_bytes -= n_peeked;
1354	  b[0]->total_length_not_including_first_buffer += n_peeked;
1355	}
1356
1357      tcp_push_hdr_i (tc, *b, tc->snd_una + offset, /* compute opts */ 0,
1358		      /* burst */ 0, /* update_snd_nxt */ 0);
1359
1360      if (PREDICT_FALSE (n_bufs))
1361	{
1362	  clib_warning ("not all buffers consumed");
1363	  vlib_buffer_free (vm, wrk->tx_buffers, n_bufs);
1364	}
1365    }
1366
1367  ASSERT (n_bytes > 0);
1368  ASSERT (((*b)->current_data + (*b)->current_length) <= bytes_per_buffer);
1369
1370  return n_bytes;
1371}
1372
1373/**
1374 * Build a retransmit segment
1375 *
1376 * @return the number of bytes in the segment or 0 if there's nothing to
1377 *         retransmit
1378 */
1379static u32
1380tcp_prepare_retransmit_segment (tcp_worker_ctx_t * wrk,
1381				tcp_connection_t * tc, u32 offset,
1382				u32 max_deq_bytes, vlib_buffer_t ** b)
1383{
1384  u32 start, available_bytes;
1385  int n_bytes = 0;
1386
1387  ASSERT (tc->state >= TCP_STATE_ESTABLISHED);
1388  ASSERT (max_deq_bytes != 0);
1389
1390  /*
1391   * Make sure we can retransmit something
1392   */
1393  available_bytes = transport_max_tx_dequeue (&tc->connection);
1394  ASSERT (available_bytes >= offset);
1395  available_bytes -= offset;
1396  if (!available_bytes)
1397    return 0;
1398
1399  max_deq_bytes = clib_min (tc->snd_mss, max_deq_bytes);
1400  max_deq_bytes = clib_min (available_bytes, max_deq_bytes);
1401
1402  start = tc->snd_una + offset;
1403  ASSERT (seq_leq (start + max_deq_bytes, tc->snd_nxt));
1404
1405  n_bytes = tcp_prepare_segment (wrk, tc, offset, max_deq_bytes, b);
1406  if (!n_bytes)
1407    return 0;
1408
1409  tc->snd_rxt_bytes += n_bytes;
1410
1411  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
1412    tcp_bt_track_rxt (tc, start, start + n_bytes);
1413
1414  tc->bytes_retrans += n_bytes;
1415  tc->segs_retrans += 1;
1416  TCP_EVT (TCP_EVT_CC_RTX, tc, offset, n_bytes);
1417
1418  return n_bytes;
1419}
1420
1421static void
1422tcp_check_sack_reneging (tcp_connection_t * tc)
1423{
1424  sack_scoreboard_t *sb = &tc->sack_sb;
1425  sack_scoreboard_hole_t *hole;
1426
1427  hole = scoreboard_first_hole (sb);
1428  if (!sb->is_reneging && (!hole || hole->start == tc->snd_una))
1429    return;
1430
1431  scoreboard_clear_reneging (sb, tc->snd_una, tc->snd_nxt);
1432}
1433
1434/**
1435 * Reset congestion control, switch cwnd to loss window and try again.
1436 */
1437static void
1438tcp_cc_init_rxt_timeout (tcp_connection_t * tc)
1439{
1440  TCP_EVT (TCP_EVT_CC_EVT, tc, 6);
1441
1442  tc->prev_ssthresh = tc->ssthresh;
1443  tc->prev_cwnd = tc->cwnd;
1444
1445  /* If we entrered loss without fast recovery, notify cc algo of the
1446   * congestion event such that it can update ssthresh and its state */
1447  if (!tcp_in_fastrecovery (tc))
1448    tcp_cc_congestion (tc);
1449
1450  /* Let cc algo decide loss cwnd and ssthresh post unrecovered loss */
1451  tcp_cc_loss (tc);
1452
1453  tc->rtt_ts = 0;
1454  tc->cwnd_acc_bytes = 0;
1455  tc->tr_occurences += 1;
1456  tcp_recovery_on (tc);
1457}
1458
1459void
1460tcp_timer_retransmit_handler (u32 tc_index)
1461{
1462  u32 thread_index = vlib_get_thread_index ();
1463  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
1464  vlib_main_t *vm = wrk->vm;
1465  tcp_connection_t *tc;
1466  vlib_buffer_t *b = 0;
1467  u32 bi, n_bytes;
1468
1469  tc = tcp_connection_get (tc_index, thread_index);
1470
1471  /* Note: the connection may have been closed and pool_put */
1472  if (PREDICT_FALSE (tc == 0 || tc->state == TCP_STATE_SYN_SENT))
1473    return;
1474
1475  tc->timers[TCP_TIMER_RETRANSMIT] = TCP_TIMER_HANDLE_INVALID;
1476
1477  /* Wait-close and retransmit could pop at the same time */
1478  if (tc->state == TCP_STATE_CLOSED)
1479    return;
1480
1481  if (tc->state >= TCP_STATE_ESTABLISHED)
1482    {
1483      TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
1484
1485      /* Lost FIN, retransmit and return */
1486      if (tc->flags & TCP_CONN_FINSNT)
1487	{
1488	  tcp_send_fin (tc);
1489	  tc->rto_boff += 1;
1490	  tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1491	  return;
1492	}
1493
1494      /* Shouldn't be here. This condition is tricky because it has to take
1495       * into account boff > 0 due to persist timeout. */
1496      if ((tc->rto_boff == 0 && tc->snd_una == tc->snd_nxt)
1497	  || (tc->rto_boff > 0 && seq_geq (tc->snd_una, tc->snd_congestion)
1498	      && !tcp_flight_size (tc)))
1499	{
1500	  ASSERT (!tcp_in_recovery (tc));
1501	  tc->rto_boff = 0;
1502	  return;
1503	}
1504
1505      /* We're not in recovery so make sure rto_boff is 0. Can be non 0 due
1506       * to persist timer timeout */
1507      if (!tcp_in_recovery (tc) && tc->rto_boff > 0)
1508	{
1509	  tc->rto_boff = 0;
1510	  tcp_update_rto (tc);
1511	}
1512
1513      /* Peer is dead or network connectivity is lost. Close connection.
1514       * RFC 1122 section 4.2.3.5 recommends a value of at least 100s. For
1515       * a min rto of 0.2s we need to retry about 8 times. */
1516      if (tc->rto_boff >= TCP_RTO_BOFF_MAX)
1517	{
1518	  tcp_send_reset (tc);
1519	  tcp_connection_set_state (tc, TCP_STATE_CLOSED);
1520	  session_transport_closing_notify (&tc->connection);
1521	  tcp_connection_timers_reset (tc);
1522	  tcp_timer_update (tc, TCP_TIMER_WAITCLOSE, tcp_cfg.closewait_time);
1523	  return;
1524	}
1525
1526      if (tcp_opts_sack_permitted (&tc->rcv_opts))
1527	tcp_check_sack_reneging (tc);
1528
1529      /* Update send congestion to make sure that rxt has data to send */
1530      tc->snd_congestion = tc->snd_nxt;
1531
1532      /* Send the first unacked segment. If we're short on buffers, return
1533       * as soon as possible */
1534      n_bytes = clib_min (tc->snd_mss, tc->snd_nxt - tc->snd_una);
1535      n_bytes = tcp_prepare_retransmit_segment (wrk, tc, 0, n_bytes, &b);
1536      if (!n_bytes)
1537	{
1538	  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
1539	  return;
1540	}
1541
1542      bi = vlib_get_buffer_index (vm, b);
1543      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1544
1545      tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1546      tcp_retransmit_timer_force_update (tc);
1547
1548      tc->rto_boff += 1;
1549      if (tc->rto_boff == 1)
1550	{
1551	  tcp_cc_init_rxt_timeout (tc);
1552	  /* Record timestamp. Eifel detection algorithm RFC3522 */
1553	  tc->snd_rxt_ts = tcp_tstamp (tc);
1554	}
1555
1556      if (tcp_opts_sack_permitted (&tc->rcv_opts))
1557	scoreboard_init_rxt (&tc->sack_sb, tc->snd_una + n_bytes);
1558
1559      tcp_program_retransmit (tc);
1560    }
1561  /* Retransmit SYN-ACK */
1562  else if (tc->state == TCP_STATE_SYN_RCVD)
1563    {
1564      TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
1565
1566      tc->rtt_ts = 0;
1567
1568      /* Passive open establish timeout */
1569      if (tc->rto > TCP_ESTABLISH_TIME >> 1)
1570	{
1571	  tcp_connection_set_state (tc, TCP_STATE_CLOSED);
1572	  tcp_connection_timers_reset (tc);
1573	  tcp_timer_update (tc, TCP_TIMER_WAITCLOSE, tcp_cfg.cleanup_time);
1574	  return;
1575	}
1576
1577      if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1578	{
1579	  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
1580	  return;
1581	}
1582
1583      tc->rto_boff += 1;
1584      if (tc->rto_boff > TCP_RTO_SYN_RETRIES)
1585	tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1586
1587      tcp_retransmit_timer_force_update (tc);
1588
1589      b = vlib_get_buffer (vm, bi);
1590      tcp_init_buffer (vm, b);
1591      tcp_make_synack (tc, b);
1592      TCP_EVT (TCP_EVT_SYN_RXT, tc, 1);
1593
1594      /* Retransmit timer already updated, just enqueue to output */
1595      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1596    }
1597  else
1598    {
1599      ASSERT (tc->state == TCP_STATE_CLOSED);
1600      return;
1601    }
1602}
1603
1604/**
1605 * SYN retransmit timer handler. Active open only.
1606 */
1607void
1608tcp_timer_retransmit_syn_handler (u32 tc_index)
1609{
1610  u32 thread_index = vlib_get_thread_index ();
1611  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
1612  vlib_main_t *vm = wrk->vm;
1613  tcp_connection_t *tc;
1614  vlib_buffer_t *b = 0;
1615  u32 bi;
1616
1617  tc = tcp_half_open_connection_get (tc_index);
1618
1619  /* Note: the connection may have transitioned to ESTABLISHED... */
1620  if (PREDICT_FALSE (tc == 0 || tc->state != TCP_STATE_SYN_SENT))
1621    return;
1622
1623  tc->timers[TCP_TIMER_RETRANSMIT_SYN] = TCP_TIMER_HANDLE_INVALID;
1624
1625  /* Half-open connection actually moved to established but we were
1626   * waiting for syn retransmit to pop to call cleanup from the right
1627   * thread. */
1628  if (tc->flags & TCP_CONN_HALF_OPEN_DONE)
1629    {
1630      if (tcp_half_open_connection_cleanup (tc))
1631	TCP_DBG ("could not remove half-open connection");
1632      return;
1633    }
1634
1635  TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
1636  tc->rtt_ts = 0;
1637
1638  /* Active open establish timeout */
1639  if (tc->rto >= TCP_ESTABLISH_TIME >> 1)
1640    {
1641      session_stream_connect_notify (&tc->connection, 1 /* fail */ );
1642      tcp_connection_cleanup (tc);
1643      return;
1644    }
1645
1646  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1647    {
1648      tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN, 1);
1649      return;
1650    }
1651
1652  /* Try without increasing RTO a number of times. If this fails,
1653   * start growing RTO exponentially */
1654  tc->rto_boff += 1;
1655  if (tc->rto_boff > TCP_RTO_SYN_RETRIES)
1656    tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1657
1658  b = vlib_get_buffer (vm, bi);
1659  tcp_init_buffer (vm, b);
1660  tcp_make_syn (tc, b);
1661
1662  TCP_EVT (TCP_EVT_SYN_RXT, tc, 0);
1663
1664  /* This goes straight to ipx_lookup */
1665  tcp_push_ip_hdr (wrk, tc, b);
1666  tcp_enqueue_to_ip_lookup (wrk, b, bi, tc->c_is_ip4, tc->c_fib_index);
1667
1668  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT_SYN,
1669		    tc->rto * TCP_TO_TIMER_TICK);
1670}
1671
1672/**
1673 * Got 0 snd_wnd from peer, try to do something about it.
1674 *
1675 */
1676void
1677tcp_timer_persist_handler (u32 index)
1678{
1679  u32 thread_index = vlib_get_thread_index ();
1680  tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
1681  u32 bi, max_snd_bytes, available_bytes, offset;
1682  tcp_main_t *tm = vnet_get_tcp_main ();
1683  vlib_main_t *vm = wrk->vm;
1684  tcp_connection_t *tc;
1685  vlib_buffer_t *b;
1686  int n_bytes = 0;
1687  u8 *data;
1688
1689  tc = tcp_connection_get_if_valid (index, thread_index);
1690  if (!tc)
1691    return;
1692
1693  /* Make sure timer handle is set to invalid */
1694  tc->timers[TCP_TIMER_PERSIST] = TCP_TIMER_HANDLE_INVALID;
1695
1696  /* Problem already solved or worse */
1697  if (tc->state == TCP_STATE_CLOSED || tc->snd_wnd > tc->snd_mss
1698      || (tc->flags & TCP_CONN_FINSNT))
1699    return;
1700
1701  available_bytes = transport_max_tx_dequeue (&tc->connection);
1702  offset = tc->snd_nxt - tc->snd_una;
1703
1704  /* Reprogram persist if no new bytes available to send. We may have data
1705   * next time */
1706  if (!available_bytes)
1707    {
1708      tcp_persist_timer_set (tc);
1709      return;
1710    }
1711
1712  if (available_bytes <= offset)
1713    {
1714      ASSERT (tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT));
1715      return;
1716    }
1717
1718  /* Increment RTO backoff */
1719  tc->rto_boff += 1;
1720  tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1721
1722  /*
1723   * Try to force the first unsent segment (or buffer)
1724   */
1725  if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1726    {
1727      tcp_persist_timer_set (tc);
1728      return;
1729    }
1730  b = vlib_get_buffer (vm, bi);
1731  data = tcp_init_buffer (vm, b);
1732
1733  tcp_validate_txf_size (tc, offset);
1734  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
1735  max_snd_bytes = clib_min (tc->snd_mss,
1736			    tm->bytes_per_buffer - TRANSPORT_MAX_HDRS_LEN);
1737  n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1738					max_snd_bytes);
1739  b->current_length = n_bytes;
1740  ASSERT (n_bytes != 0 && (tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT)
1741			   || tc->snd_nxt == tc->snd_una_max
1742			   || tc->rto_boff > 1));
1743
1744  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
1745    {
1746      tcp_bt_check_app_limited (tc);
1747      tcp_bt_track_tx (tc, n_bytes);
1748    }
1749
1750  tcp_push_hdr_i (tc, b, tc->snd_nxt, /* compute opts */ 0,
1751		  /* burst */ 0, /* update_snd_nxt */ 1);
1752  tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
1753  tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
1754  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1755
1756  /* Just sent new data, enable retransmit */
1757  tcp_retransmit_timer_update (tc);
1758}
1759
1760/**
1761 * Retransmit first unacked segment
1762 */
1763int
1764tcp_retransmit_first_unacked (tcp_worker_ctx_t * wrk, tcp_connection_t * tc)
1765{
1766  vlib_main_t *vm = wrk->vm;
1767  vlib_buffer_t *b;
1768  u32 bi, n_bytes;
1769
1770  TCP_EVT (TCP_EVT_CC_EVT, tc, 1);
1771
1772  n_bytes = tcp_prepare_retransmit_segment (wrk, tc, 0, tc->snd_mss, &b);
1773  if (!n_bytes)
1774    return -1;
1775
1776  bi = vlib_get_buffer_index (vm, b);
1777  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1778
1779  return 0;
1780}
1781
1782static int
1783tcp_transmit_unsent (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1784		     u32 burst_size)
1785{
1786  u32 offset, n_segs = 0, n_written, bi, available_wnd;
1787  vlib_main_t *vm = wrk->vm;
1788  vlib_buffer_t *b = 0;
1789
1790  offset = tc->snd_nxt - tc->snd_una;
1791  available_wnd = tc->snd_wnd - offset;
1792  burst_size = clib_min (burst_size, available_wnd / tc->snd_mss);
1793
1794  if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
1795    tcp_bt_check_app_limited (tc);
1796
1797  while (n_segs < burst_size)
1798    {
1799      n_written = tcp_prepare_segment (wrk, tc, offset, tc->snd_mss, &b);
1800      if (!n_written)
1801	goto done;
1802
1803      bi = vlib_get_buffer_index (vm, b);
1804      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1805      offset += n_written;
1806      n_segs += 1;
1807
1808      if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
1809	tcp_bt_track_tx (tc, n_written);
1810
1811      tc->snd_nxt += n_written;
1812      tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
1813    }
1814
1815done:
1816  return n_segs;
1817}
1818
1819/**
1820 * Estimate send space using proportional rate reduction (RFC6937)
1821 */
1822int
1823tcp_fastrecovery_prr_snd_space (tcp_connection_t * tc)
1824{
1825  u32 pipe, prr_out;
1826  int space;
1827
1828  pipe = tcp_flight_size (tc);
1829  prr_out = tc->snd_rxt_bytes + (tc->snd_nxt - tc->snd_congestion);
1830
1831  if (pipe > tc->ssthresh)
1832    {
1833      space = ((int) tc->prr_delivered * ((f64) tc->ssthresh / tc->prev_cwnd))
1834	- prr_out;
1835    }
1836  else
1837    {
1838      int limit;
1839      limit = clib_max ((int) (tc->prr_delivered - prr_out), 0) + tc->snd_mss;
1840      space = clib_min (tc->ssthresh - pipe, limit);
1841    }
1842  space = clib_max (space, prr_out ? 0 : tc->snd_mss);
1843  return space;
1844}
1845
1846static inline u8
1847tcp_retransmit_should_retry_head (tcp_connection_t * tc,
1848				  sack_scoreboard_t * sb)
1849{
1850  u32 tx_adv_sack = sb->high_sacked - tc->snd_congestion;
1851  f64 rr = (f64) tc->ssthresh / tc->prev_cwnd;
1852
1853  return (tx_adv_sack > (tc->snd_una - tc->prr_start) * rr);
1854}
1855
1856static inline u8
1857tcp_max_tx_deq (tcp_connection_t * tc)
1858{
1859  return (transport_max_tx_dequeue (&tc->connection)
1860	  - (tc->snd_nxt - tc->snd_una));
1861}
1862
1863#define scoreboard_rescue_rxt_valid(_sb, _tc)			\
1864    (seq_geq (_sb->rescue_rxt, _tc->snd_una) 			\
1865	&& seq_leq (_sb->rescue_rxt, _tc->snd_congestion))
1866
1867/**
1868 * Do retransmit with SACKs
1869 */
1870static int
1871tcp_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1872		     u32 burst_size)
1873{
1874  u8 snd_limited = 0, can_rescue = 0, reset_pacer = 0;
1875  u32 n_written = 0, offset, max_bytes, n_segs = 0;
1876  u32 bi, max_deq, burst_bytes, sent_bytes;
1877  sack_scoreboard_hole_t *hole;
1878  vlib_main_t *vm = wrk->vm;
1879  vlib_buffer_t *b = 0;
1880  sack_scoreboard_t *sb;
1881  int snd_space;
1882
1883  ASSERT (tcp_in_cong_recovery (tc));
1884
1885  burst_bytes = transport_connection_tx_pacer_burst (&tc->connection);
1886  burst_size = clib_min (burst_size, burst_bytes / tc->snd_mss);
1887  if (!burst_size)
1888    {
1889      tcp_program_retransmit (tc);
1890      return 0;
1891    }
1892
1893  if (tcp_in_recovery (tc))
1894    snd_space = tcp_available_cc_snd_space (tc);
1895  else
1896    snd_space = tcp_fastrecovery_prr_snd_space (tc);
1897
1898  if (snd_space < tc->snd_mss)
1899    {
1900      reset_pacer = burst_bytes > tc->snd_mss;
1901      goto done;
1902    }
1903
1904  reset_pacer = snd_space < burst_bytes;
1905
1906  sb = &tc->sack_sb;
1907
1908  /* Check if snd_una is a lost retransmit */
1909  if (pool_elts (sb->holes)
1910      && seq_gt (sb->high_sacked, tc->snd_congestion)
1911      && tc->rxt_head != tc->snd_una
1912      && tcp_retransmit_should_retry_head (tc, sb))
1913    {
1914      n_written = tcp_prepare_retransmit_segment (wrk, tc, 0, tc->snd_mss,
1915						  &b);
1916      if (!n_written)
1917	{
1918	  tcp_program_retransmit (tc);
1919	  goto done;
1920	}
1921      bi = vlib_get_buffer_index (vm, b);
1922      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1923      n_segs = 1;
1924
1925      tc->rxt_head = tc->snd_una;
1926      tc->rxt_delivered += n_written;
1927      tc->prr_delivered += n_written;
1928      ASSERT (tc->rxt_delivered <= tc->snd_rxt_bytes);
1929    }
1930
1931  TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
1932  hole = scoreboard_get_hole (sb, sb->cur_rxt_hole);
1933
1934  max_deq = transport_max_tx_dequeue (&tc->connection);
1935  max_deq -= tc->snd_nxt - tc->snd_una;
1936
1937  while (snd_space > 0 && n_segs < burst_size)
1938    {
1939      hole = scoreboard_next_rxt_hole (sb, hole, max_deq != 0, &can_rescue,
1940				       &snd_limited);
1941      if (!hole)
1942	{
1943	  /* We are out of lost holes to retransmit so send some new data. */
1944	  if (max_deq > tc->snd_mss)
1945	    {
1946	      u32 n_segs_new;
1947	      int av_wnd;
1948
1949	      av_wnd = (int) tc->snd_wnd - (tc->snd_nxt - tc->snd_una);
1950	      av_wnd = clib_max (av_wnd, 0);
1951	      snd_space = clib_min (snd_space, av_wnd);
1952	      snd_space = clib_min (max_deq, snd_space);
1953	      burst_size = clib_min (burst_size - n_segs,
1954				     snd_space / tc->snd_mss);
1955	      burst_size = clib_min (burst_size, TCP_RXT_MAX_BURST);
1956	      n_segs_new = tcp_transmit_unsent (wrk, tc, burst_size);
1957	      if (max_deq > n_segs_new * tc->snd_mss)
1958		tcp_program_retransmit (tc);
1959
1960	      n_segs += n_segs_new;
1961	      goto done;
1962	    }
1963
1964	  if (tcp_in_recovery (tc) || !can_rescue
1965	      || scoreboard_rescue_rxt_valid (sb, tc))
1966	    break;
1967
1968	  /* If rescue rxt undefined or less than snd_una then one segment of
1969	   * up to SMSS octets that MUST include the highest outstanding
1970	   * unSACKed sequence number SHOULD be returned, and RescueRxt set to
1971	   * RecoveryPoint. HighRxt MUST NOT be updated.
1972	   */
1973	  max_bytes = clib_min (tc->snd_mss,
1974				tc->snd_congestion - tc->snd_una);
1975	  max_bytes = clib_min (max_bytes, snd_space);
1976	  offset = tc->snd_congestion - tc->snd_una - max_bytes;
1977	  sb->rescue_rxt = tc->snd_congestion;
1978	  n_written = tcp_prepare_retransmit_segment (wrk, tc, offset,
1979						      max_bytes, &b);
1980	  if (!n_written)
1981	    goto done;
1982
1983	  bi = vlib_get_buffer_index (vm, b);
1984	  tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1985	  n_segs += 1;
1986	  break;
1987	}
1988
1989      max_bytes = clib_min (hole->end - sb->high_rxt, snd_space);
1990      max_bytes = snd_limited ? clib_min (max_bytes, tc->snd_mss) : max_bytes;
1991      if (max_bytes == 0)
1992	break;
1993
1994      offset = sb->high_rxt - tc->snd_una;
1995      n_written = tcp_prepare_retransmit_segment (wrk, tc, offset, max_bytes,
1996						  &b);
1997      ASSERT (n_written <= snd_space);
1998
1999      /* Nothing left to retransmit */
2000      if (n_written == 0)
2001	break;
2002
2003      bi = vlib_get_buffer_index (vm, b);
2004      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
2005
2006      sb->high_rxt += n_written;
2007      ASSERT (seq_leq (sb->high_rxt, tc->snd_nxt));
2008
2009      snd_space -= n_written;
2010      n_segs += 1;
2011    }
2012
2013  if (hole)
2014    tcp_program_retransmit (tc);
2015
2016done:
2017
2018  if (reset_pacer)
2019    {
2020      transport_connection_tx_pacer_reset_bucket (&tc->connection);
2021    }
2022  else
2023    {
2024      sent_bytes = clib_min (n_segs * tc->snd_mss, burst_bytes);
2025      transport_connection_tx_pacer_update_bytes (&tc->connection,
2026						  sent_bytes);
2027    }
2028
2029  return n_segs;
2030}
2031
2032/**
2033 * Fast retransmit without SACK info
2034 */
2035static int
2036tcp_retransmit_no_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
2037			u32 burst_size)
2038{
2039  u32 n_written = 0, offset = 0, bi, max_deq, n_segs_now;
2040  u32 burst_bytes, sent_bytes;
2041  vlib_main_t *vm = wrk->vm;
2042  int snd_space, n_segs = 0;
2043  u8 cc_limited = 0;
2044  vlib_buffer_t *b;
2045
2046  ASSERT (tcp_in_fastrecovery (tc));
2047  TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
2048
2049  burst_bytes = transport_connection_tx_pacer_burst (&tc->connection);
2050  burst_size = clib_min (burst_size, burst_bytes / tc->snd_mss);
2051  if (!burst_size)
2052    {
2053      tcp_program_retransmit (tc);
2054      return 0;
2055    }
2056
2057  snd_space = tcp_available_cc_snd_space (tc);
2058  cc_limited = snd_space < burst_bytes;
2059
2060  if (!tcp_fastrecovery_first (tc))
2061    goto send_unsent;
2062
2063  /* RFC 6582: [If a partial ack], retransmit the first unacknowledged
2064   * segment. */
2065  while (snd_space > 0 && n_segs < burst_size)
2066    {
2067      n_written = tcp_prepare_retransmit_segment (wrk, tc, offset,
2068						  tc->snd_mss, &b);
2069
2070      /* Nothing left to retransmit */
2071      if (n_written == 0)
2072	break;
2073
2074      bi = vlib_get_buffer_index (vm, b);
2075      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
2076      snd_space -= n_written;
2077      offset += n_written;
2078      n_segs += 1;
2079    }
2080
2081  if (n_segs == burst_size)
2082    goto done;
2083
2084send_unsent:
2085
2086  /* RFC 6582: Send a new segment if permitted by the new value of cwnd. */
2087  if (snd_space < tc->snd_mss || tc->snd_mss == 0)
2088    goto done;
2089
2090  max_deq = transport_max_tx_dequeue (&tc->connection);
2091  max_deq -= tc->snd_nxt - tc->snd_una;
2092  if (max_deq)
2093    {
2094      snd_space = clib_min (max_deq, snd_space);
2095      burst_size = clib_min (burst_size - n_segs, snd_space / tc->snd_mss);
2096      n_segs_now = tcp_transmit_unsent (wrk, tc, burst_size);
2097      if (max_deq > n_segs_now * tc->snd_mss)
2098	tcp_program_retransmit (tc);
2099      n_segs += n_segs_now;
2100    }
2101
2102done:
2103  tcp_fastrecovery_first_off (tc);
2104
2105  sent_bytes = clib_min (n_segs * tc->snd_mss, burst_bytes);
2106  sent_bytes = cc_limited ? burst_bytes : sent_bytes;
2107  transport_connection_tx_pacer_update_bytes (&tc->connection, sent_bytes);
2108
2109  return n_segs;
2110}
2111
2112static int
2113tcp_send_acks (tcp_connection_t * tc, u32 max_burst_size)
2114{
2115  int j, n_acks;
2116
2117  if (!tc->pending_dupacks)
2118    {
2119      if (tcp_in_cong_recovery (tc) || !tcp_max_tx_deq (tc)
2120	  || tc->state != TCP_STATE_ESTABLISHED)
2121	{
2122	  tcp_send_ack (tc);
2123	  return 1;
2124	}
2125      return 0;
2126    }
2127
2128  /* If we're supposed to send dupacks but have no ooo data
2129   * send only one ack */
2130  if (!vec_len (tc->snd_sacks))
2131    {
2132      tcp_send_ack (tc);
2133      tc->pending_dupacks = 0;
2134      return 1;
2135    }
2136
2137  /* Start with first sack block */
2138  tc->snd_sack_pos = 0;
2139
2140  /* Generate enough dupacks to cover all sack blocks. Do not generate
2141   * more sacks than the number of packets received. But do generate at
2142   * least 3, i.e., the number needed to signal congestion, if needed. */
2143  n_acks = vec_len (tc->snd_sacks) / TCP_OPTS_MAX_SACK_BLOCKS;
2144  n_acks = clib_min (n_acks, tc->pending_dupacks);
2145  n_acks = clib_max (n_acks, clib_min (tc->pending_dupacks, 3));
2146  for (j = 0; j < clib_min (n_acks, max_burst_size); j++)
2147    tcp_send_ack (tc);
2148
2149  if (n_acks < max_burst_size)
2150    {
2151      tc->pending_dupacks = 0;
2152      tc->snd_sack_pos = 0;
2153      tc->dupacks_out += n_acks;
2154      return n_acks;
2155    }
2156  else
2157    {
2158      TCP_DBG ("constrained by burst size");
2159      tc->pending_dupacks = n_acks - max_burst_size;
2160      tc->dupacks_out += max_burst_size;
2161      tcp_program_dupack (tc);
2162      return max_burst_size;
2163    }
2164}
2165
2166static int
2167tcp_do_retransmit (tcp_connection_t * tc, u32 max_burst_size)
2168{
2169  tcp_worker_ctx_t *wrk;
2170  u32 n_segs;
2171
2172  wrk = tcp_get_worker (tc->c_thread_index);
2173
2174  if (tcp_opts_sack_permitted (&tc->rcv_opts))
2175    n_segs = tcp_retransmit_sack (wrk, tc, max_burst_size);
2176  else
2177    n_segs = tcp_retransmit_no_sack (wrk, tc, max_burst_size);
2178
2179  return n_segs;
2180}
2181
2182int
2183tcp_session_custom_tx (void *conn, u32 max_burst_size)
2184{
2185  tcp_connection_t *tc = (tcp_connection_t *) conn;
2186  u32 n_segs = 0;
2187
2188  if (tcp_in_cong_recovery (tc) && (tc->flags & TCP_CONN_RXT_PENDING))
2189    {
2190      tc->flags &= ~TCP_CONN_RXT_PENDING;
2191      n_segs = tcp_do_retransmit (tc, max_burst_size);
2192      max_burst_size -= n_segs;
2193    }
2194
2195  if (!(tc->flags & TCP_CONN_SNDACK))
2196    return n_segs;
2197
2198  tc->flags &= ~TCP_CONN_SNDACK;
2199
2200  /* We have retransmitted packets and no dupack */
2201  if (n_segs && !tc->pending_dupacks)
2202    return n_segs;
2203
2204  if (!max_burst_size)
2205    {
2206      tcp_program_ack (tc);
2207      return max_burst_size;
2208    }
2209
2210  n_segs += tcp_send_acks (tc, max_burst_size);
2211
2212  return n_segs;
2213}
2214#endif /* CLIB_MARCH_VARIANT */
2215
2216static void
2217tcp_output_handle_link_local (tcp_connection_t * tc0, vlib_buffer_t * b0,
2218			      u16 * next0, u32 * error0)
2219{
2220  ip_adjacency_t *adj;
2221  adj_index_t ai;
2222
2223  /* Not thread safe but as long as the connection exists the adj should
2224   * not be removed */
2225  ai = adj_nbr_find (FIB_PROTOCOL_IP6, VNET_LINK_IP6, &tc0->c_rmt_ip,
2226		     tc0->sw_if_index);
2227  if (ai == ADJ_INDEX_INVALID)
2228    {
2229      vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
2230      *next0 = TCP_OUTPUT_NEXT_DROP;
2231      *error0 = TCP_ERROR_LINK_LOCAL_RW;
2232      return;
2233    }
2234
2235  adj = adj_get (ai);
2236  if (PREDICT_TRUE (adj->lookup_next_index == IP_LOOKUP_NEXT_REWRITE))
2237    *next0 = TCP_OUTPUT_NEXT_IP_REWRITE;
2238  else if (adj->lookup_next_index == IP_LOOKUP_NEXT_ARP)
2239    *next0 = TCP_OUTPUT_NEXT_IP_ARP;
2240  else
2241    {
2242      *next0 = TCP_OUTPUT_NEXT_DROP;
2243      *error0 = TCP_ERROR_LINK_LOCAL_RW;
2244    }
2245  vnet_buffer (b0)->ip.adj_index[VLIB_TX] = ai;
2246}
2247
2248static void
2249tcp46_output_trace_frame (vlib_main_t * vm, vlib_node_runtime_t * node,
2250			  u32 * to_next, u32 n_bufs)
2251{
2252  u32 n_trace = vlib_get_trace_count (vm, node);
2253  tcp_connection_t *tc;
2254  tcp_tx_trace_t *t;
2255  vlib_buffer_t *b;
2256  tcp_header_t *th;
2257  int i;
2258
2259  for (i = 0; i < clib_min (n_trace, n_bufs); i++)
2260    {
2261      b = vlib_get_buffer (vm, to_next[i]);
2262      th = vlib_buffer_get_current (b);
2263      tc = tcp_connection_get (vnet_buffer (b)->tcp.connection_index,
2264			       vm->thread_index);
2265      t = vlib_add_trace (vm, node, b, sizeof (*t));
2266      clib_memcpy_fast (&t->tcp_header, th, sizeof (t->tcp_header));
2267      clib_memcpy_fast (&t->tcp_connection, tc, sizeof (t->tcp_connection));
2268    }
2269}
2270
2271always_inline void
2272tcp_output_push_ip (vlib_main_t * vm, vlib_buffer_t * b0,
2273		    tcp_connection_t * tc0, u8 is_ip4)
2274{
2275  u8 __clib_unused *ih0;
2276  tcp_header_t __clib_unused *th0 = vlib_buffer_get_current (b0);
2277
2278  TCP_EVT (TCP_EVT_OUTPUT, tc0, th0->flags, b0->current_length);
2279
2280  if (is_ip4)
2281    ih0 = vlib_buffer_push_ip4 (vm, b0, &tc0->c_lcl_ip4, &tc0->c_rmt_ip4,
2282				IP_PROTOCOL_TCP, tcp_csum_offload (tc0));
2283  else
2284    ih0 =
2285      vlib_buffer_push_ip6_custom (vm, b0, &tc0->c_lcl_ip6, &tc0->c_rmt_ip6,
2286				   IP_PROTOCOL_TCP, tc0->ipv6_flow_label);
2287
2288}
2289
2290always_inline void
2291tcp_check_if_gso (tcp_connection_t * tc, vlib_buffer_t * b)
2292{
2293  if (PREDICT_TRUE (!(tc->cfg_flags & TCP_CFG_F_TSO)))
2294    return;
2295
2296  u16 data_len = b->current_length - sizeof (tcp_header_t) - tc->snd_opts_len;
2297
2298  if (PREDICT_FALSE (b->flags & VLIB_BUFFER_TOTAL_LENGTH_VALID))
2299    data_len += b->total_length_not_including_first_buffer;
2300
2301  if (PREDICT_TRUE (data_len <= tc->snd_mss))
2302    return;
2303  else
2304    {
2305      ASSERT ((b->flags & VNET_BUFFER_F_L3_HDR_OFFSET_VALID) != 0);
2306      ASSERT ((b->flags & VNET_BUFFER_F_L4_HDR_OFFSET_VALID) != 0);
2307      b->flags |= VNET_BUFFER_F_GSO;
2308      vnet_buffer2 (b)->gso_l4_hdr_sz =
2309	sizeof (tcp_header_t) + tc->snd_opts_len;
2310      vnet_buffer2 (b)->gso_size = tc->snd_mss;
2311    }
2312}
2313
2314always_inline void
2315tcp_output_handle_packet (tcp_connection_t * tc0, vlib_buffer_t * b0,
2316			  vlib_node_runtime_t * error_node, u16 * next0,
2317			  u8 is_ip4)
2318{
2319  /* If next_index is not drop use it */
2320  if (tc0->next_node_index)
2321    {
2322      *next0 = tc0->next_node_index;
2323      vnet_buffer (b0)->tcp.next_node_opaque = tc0->next_node_opaque;
2324    }
2325  else
2326    {
2327      *next0 = TCP_OUTPUT_NEXT_IP_LOOKUP;
2328    }
2329
2330  vnet_buffer (b0)->sw_if_index[VLIB_TX] = tc0->c_fib_index;
2331  vnet_buffer (b0)->sw_if_index[VLIB_RX] = 0;
2332
2333  if (!is_ip4)
2334    {
2335      u32 error0 = 0;
2336
2337      if (PREDICT_FALSE (ip6_address_is_link_local_unicast (&tc0->c_rmt_ip6)))
2338	tcp_output_handle_link_local (tc0, b0, next0, &error0);
2339
2340      if (PREDICT_FALSE (error0))
2341	{
2342	  b0->error = error_node->errors[error0];
2343	  return;
2344	}
2345    }
2346
2347  if (!TCP_ALWAYS_ACK)
2348    tcp_timer_reset (tc0, TCP_TIMER_DELACK);
2349
2350  tc0->segs_out += 1;
2351}
2352
2353always_inline uword
2354tcp46_output_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
2355		     vlib_frame_t * frame, int is_ip4)
2356{
2357  u32 n_left_from, *from, thread_index = vm->thread_index;
2358  vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b;
2359  u16 nexts[VLIB_FRAME_SIZE], *next;
2360  vlib_node_runtime_t *error_node;
2361
2362  error_node = vlib_node_get_runtime (vm, tcp_node_index (output, is_ip4));
2363
2364  from = vlib_frame_vector_args (frame);
2365  n_left_from = frame->n_vectors;
2366  tcp_set_time_now (tcp_get_worker (thread_index));
2367
2368  if (PREDICT_FALSE (node->flags & VLIB_NODE_FLAG_TRACE))
2369    tcp46_output_trace_frame (vm, node, from, n_left_from);
2370
2371  vlib_get_buffers (vm, from, bufs, n_left_from);
2372  b = bufs;
2373  next = nexts;
2374
2375  while (n_left_from >= 4)
2376    {
2377      tcp_connection_t *tc0, *tc1;
2378
2379      {
2380	vlib_prefetch_buffer_header (b[2], STORE);
2381	CLIB_PREFETCH (b[2]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2382
2383	vlib_prefetch_buffer_header (b[3], STORE);
2384	CLIB_PREFETCH (b[3]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2385      }
2386
2387      tc0 = tcp_connection_get (vnet_buffer (b[0])->tcp.connection_index,
2388				thread_index);
2389      tc1 = tcp_connection_get (vnet_buffer (b[1])->tcp.connection_index,
2390				thread_index);
2391
2392      if (PREDICT_TRUE (!tc0 + !tc1 == 0))
2393	{
2394	  tcp_output_push_ip (vm, b[0], tc0, is_ip4);
2395	  tcp_output_push_ip (vm, b[1], tc1, is_ip4);
2396
2397	  tcp_check_if_gso (tc0, b[0]);
2398	  tcp_check_if_gso (tc1, b[1]);
2399
2400	  tcp_output_handle_packet (tc0, b[0], error_node, &next[0], is_ip4);
2401	  tcp_output_handle_packet (tc1, b[1], error_node, &next[1], is_ip4);
2402	}
2403      else
2404	{
2405	  if (tc0 != 0)
2406	    {
2407	      tcp_output_push_ip (vm, b[0], tc0, is_ip4);
2408	      tcp_check_if_gso (tc0, b[0]);
2409	      tcp_output_handle_packet (tc0, b[0], error_node, &next[0],
2410					is_ip4);
2411	    }
2412	  else
2413	    {
2414	      b[0]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2415	      next[0] = TCP_OUTPUT_NEXT_DROP;
2416	    }
2417	  if (tc1 != 0)
2418	    {
2419	      tcp_output_push_ip (vm, b[1], tc1, is_ip4);
2420	      tcp_check_if_gso (tc1, b[1]);
2421	      tcp_output_handle_packet (tc1, b[1], error_node, &next[1],
2422					is_ip4);
2423	    }
2424	  else
2425	    {
2426	      b[1]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2427	      next[1] = TCP_OUTPUT_NEXT_DROP;
2428	    }
2429	}
2430
2431      b += 2;
2432      next += 2;
2433      n_left_from -= 2;
2434    }
2435  while (n_left_from > 0)
2436    {
2437      tcp_connection_t *tc0;
2438
2439      if (n_left_from > 1)
2440	{
2441	  vlib_prefetch_buffer_header (b[1], STORE);
2442	  CLIB_PREFETCH (b[1]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2443	}
2444
2445      tc0 = tcp_connection_get (vnet_buffer (b[0])->tcp.connection_index,
2446				thread_index);
2447
2448      if (PREDICT_TRUE (tc0 != 0))
2449	{
2450	  tcp_output_push_ip (vm, b[0], tc0, is_ip4);
2451	  tcp_check_if_gso (tc0, b[0]);
2452	  tcp_output_handle_packet (tc0, b[0], error_node, &next[0], is_ip4);
2453	}
2454      else
2455	{
2456	  b[0]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2457	  next[0] = TCP_OUTPUT_NEXT_DROP;
2458	}
2459
2460      b += 1;
2461      next += 1;
2462      n_left_from -= 1;
2463    }
2464
2465  vlib_buffer_enqueue_to_next (vm, node, from, nexts, frame->n_vectors);
2466  vlib_node_increment_counter (vm, tcp_node_index (output, is_ip4),
2467			       TCP_ERROR_PKTS_SENT, frame->n_vectors);
2468  return frame->n_vectors;
2469}
2470
2471VLIB_NODE_FN (tcp4_output_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2472				 vlib_frame_t * from_frame)
2473{
2474  return tcp46_output_inline (vm, node, from_frame, 1 /* is_ip4 */ );
2475}
2476
2477VLIB_NODE_FN (tcp6_output_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2478				 vlib_frame_t * from_frame)
2479{
2480  return tcp46_output_inline (vm, node, from_frame, 0 /* is_ip4 */ );
2481}
2482
2483/* *INDENT-OFF* */
2484VLIB_REGISTER_NODE (tcp4_output_node) =
2485{
2486  .name = "tcp4-output",
2487  /* Takes a vector of packets. */
2488  .vector_size = sizeof (u32),
2489  .n_errors = TCP_N_ERROR,
2490  .protocol_hint = VLIB_NODE_PROTO_HINT_TCP,
2491  .error_strings = tcp_error_strings,
2492  .n_next_nodes = TCP_OUTPUT_N_NEXT,
2493  .next_nodes = {
2494#define _(s,n) [TCP_OUTPUT_NEXT_##s] = n,
2495    foreach_tcp4_output_next
2496#undef _
2497  },
2498  .format_buffer = format_tcp_header,
2499  .format_trace = format_tcp_tx_trace,
2500};
2501/* *INDENT-ON* */
2502
2503/* *INDENT-OFF* */
2504VLIB_REGISTER_NODE (tcp6_output_node) =
2505{
2506  .name = "tcp6-output",
2507    /* Takes a vector of packets. */
2508  .vector_size = sizeof (u32),
2509  .n_errors = TCP_N_ERROR,
2510  .protocol_hint = VLIB_NODE_PROTO_HINT_TCP,
2511  .error_strings = tcp_error_strings,
2512  .n_next_nodes = TCP_OUTPUT_N_NEXT,
2513  .next_nodes = {
2514#define _(s,n) [TCP_OUTPUT_NEXT_##s] = n,
2515    foreach_tcp6_output_next
2516#undef _
2517  },
2518  .format_buffer = format_tcp_header,
2519  .format_trace = format_tcp_tx_trace,
2520};
2521/* *INDENT-ON* */
2522
2523typedef enum _tcp_reset_next
2524{
2525  TCP_RESET_NEXT_DROP,
2526  TCP_RESET_NEXT_IP_LOOKUP,
2527  TCP_RESET_N_NEXT
2528} tcp_reset_next_t;
2529
2530#define foreach_tcp4_reset_next        	\
2531  _(DROP, "error-drop")                 \
2532  _(IP_LOOKUP, "ip4-lookup")
2533
2534#define foreach_tcp6_reset_next        	\
2535  _(DROP, "error-drop")                 \
2536  _(IP_LOOKUP, "ip6-lookup")
2537
2538static uword
2539tcp46_send_reset_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
2540			 vlib_frame_t * from_frame, u8 is_ip4)
2541{
2542  u32 n_left_from, next_index, *from, *to_next;
2543  u32 my_thread_index = vm->thread_index;
2544
2545  from = vlib_frame_vector_args (from_frame);
2546  n_left_from = from_frame->n_vectors;
2547
2548  next_index = node->cached_next_index;
2549
2550  while (n_left_from > 0)
2551    {
2552      u32 n_left_to_next;
2553
2554      vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
2555
2556      while (n_left_from > 0 && n_left_to_next > 0)
2557	{
2558	  u32 bi0;
2559	  vlib_buffer_t *b0;
2560	  tcp_tx_trace_t *t0;
2561	  tcp_header_t *th0;
2562	  u32 error0 = TCP_ERROR_RST_SENT, next0 = TCP_RESET_NEXT_IP_LOOKUP;
2563
2564	  bi0 = from[0];
2565	  to_next[0] = bi0;
2566	  from += 1;
2567	  to_next += 1;
2568	  n_left_from -= 1;
2569	  n_left_to_next -= 1;
2570
2571	  b0 = vlib_get_buffer (vm, bi0);
2572
2573	  if (tcp_make_reset_in_place (vm, b0, vnet_buffer (b0)->tcp.flags,
2574				       my_thread_index, is_ip4))
2575	    {
2576	      error0 = TCP_ERROR_LOOKUP_DROPS;
2577	      next0 = TCP_RESET_NEXT_DROP;
2578	      goto done;
2579	    }
2580
2581	  /* Prepare to send to IP lookup */
2582	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
2583	  next0 = TCP_RESET_NEXT_IP_LOOKUP;
2584
2585	done:
2586	  b0->error = node->errors[error0];
2587	  b0->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
2588	  if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
2589	    {
2590	      th0 = vlib_buffer_get_current (b0);
2591	      if (is_ip4)
2592		th0 = ip4_next_header ((ip4_header_t *) th0);
2593	      else
2594		th0 = ip6_next_header ((ip6_header_t *) th0);
2595	      t0 = vlib_add_trace (vm, node, b0, sizeof (*t0));
2596	      clib_memcpy_fast (&t0->tcp_header, th0,
2597				sizeof (t0->tcp_header));
2598	    }
2599
2600	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
2601					   n_left_to_next, bi0, next0);
2602	}
2603      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
2604    }
2605  return from_frame->n_vectors;
2606}
2607
2608VLIB_NODE_FN (tcp4_reset_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2609				vlib_frame_t * from_frame)
2610{
2611  return tcp46_send_reset_inline (vm, node, from_frame, 1);
2612}
2613
2614VLIB_NODE_FN (tcp6_reset_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2615				vlib_frame_t * from_frame)
2616{
2617  return tcp46_send_reset_inline (vm, node, from_frame, 0);
2618}
2619
2620/* *INDENT-OFF* */
2621VLIB_REGISTER_NODE (tcp4_reset_node) = {
2622  .name = "tcp4-reset",
2623  .vector_size = sizeof (u32),
2624  .n_errors = TCP_N_ERROR,
2625  .error_strings = tcp_error_strings,
2626  .n_next_nodes = TCP_RESET_N_NEXT,
2627  .next_nodes = {
2628#define _(s,n) [TCP_RESET_NEXT_##s] = n,
2629    foreach_tcp4_reset_next
2630#undef _
2631  },
2632  .format_trace = format_tcp_tx_trace,
2633};
2634/* *INDENT-ON* */
2635
2636/* *INDENT-OFF* */
2637VLIB_REGISTER_NODE (tcp6_reset_node) = {
2638  .name = "tcp6-reset",
2639  .vector_size = sizeof (u32),
2640  .n_errors = TCP_N_ERROR,
2641  .error_strings = tcp_error_strings,
2642  .n_next_nodes = TCP_RESET_N_NEXT,
2643  .next_nodes = {
2644#define _(s,n) [TCP_RESET_NEXT_##s] = n,
2645    foreach_tcp6_reset_next
2646#undef _
2647  },
2648  .format_trace = format_tcp_tx_trace,
2649};
2650/* *INDENT-ON* */
2651
2652/*
2653 * fd.io coding-style-patch-verification: ON
2654 *
2655 * Local Variables:
2656 * eval: (c-set-style "gnu")
2657 * End:
2658 */
2659