1 /*-
2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3 * The Regents of the University of California. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 4. Neither the name of the University nor the names of its contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 *
29 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
30 * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.100.2.7 2005/05/04 13:59:26 andre Exp $
31 */
32
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ipsec.h"
36 #include "opt_mac.h"
37 #include "opt_tcpdebug.h"
38 #include "opt_tcp_sack.h"
39
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #include <sys/kernel.h>
44 #include <sys/lock.h>
45 #include <sys/mac.h>
46 #include <sys/mbuf.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52
53 #include <net/route.h>
54
55 #include <netinet/in.h>
56 #include <netinet/in_systm.h>
57 #include <netinet/ip.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/ip_var.h>
60 #ifdef INET6
61 #include <netinet6/in6_pcb.h>
62 #include <netinet/ip6.h>
63 #include <netinet6/ip6_var.h>
64 #endif
65 #include <netinet/tcp.h>
66 #define TCPOUTFLAGS
67 #include <netinet/tcp_fsm.h>
68 #include <netinet/tcp_seq.h>
69 #include <netinet/tcp_timer.h>
70 #include <netinet/tcp_var.h>
71 #include <netinet/tcpip.h>
72 #ifdef TCPDEBUG
73 #include <netinet/tcp_debug.h>
74 #endif
75
76 #ifdef IPSEC
77 #include <netinet6/ipsec.h>
78 #endif /*IPSEC*/
79
80 #ifdef FAST_IPSEC
81 #include <netipsec/ipsec.h>
82 #define IPSEC
83 #endif /*FAST_IPSEC*/
84
85 #include <machine/in_cksum.h>
86
87 #ifdef notyet
88 extern struct mbuf *m_copypack();
89 #endif
90
91 int path_mtu_discovery = 1;
92 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
93 &path_mtu_discovery, 1, "Enable Path MTU Discovery");
94
95 int ss_fltsz = 1;
96 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
97 &ss_fltsz, 1, "Slow start flight size");
98
99 int ss_fltsz_local = 4;
100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
101 &ss_fltsz_local, 1, "Slow start flight size for local networks");
102
103 int tcp_do_newreno = 1;
104 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno,
105 0, "Enable NewReno Algorithms");
106
107 /*
108 * Tcp output routine: figure out what should be sent and send it.
109 */
110 int
111 tcp_output(struct tcpcb *tp)
112 {
113 struct socket *so = tp->t_inpcb->inp_socket;
114 long len, recwin, sendwin;
115 int off, flags, error;
116 #ifdef TCP_SIGNATURE
117 int sigoff = 0;
118 #endif
119 struct mbuf *m;
120 struct ip *ip = NULL;
121 struct ipovly *ipov = NULL;
122 struct tcphdr *th;
123 u_char opt[TCP_MAXOLEN];
124 unsigned ipoptlen, optlen, hdrlen;
125 int idle, sendalot;
126 int i, sack_rxmit;
127 int sack_bytes_rxmt;
128 struct sackhole *p;
129 #if 0
130 int maxburst = TCP_MAXBURST;
131 #endif
132 struct rmxp_tao tao;
133 #ifdef INET6
134 struct ip6_hdr *ip6 = NULL;
135 int isipv6;
136
137 bzero(&tao, sizeof(tao));
138 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
139 #endif
140
141 INP_LOCK_ASSERT(tp->t_inpcb);
142
143 /*
144 * Determine length of data that should be transmitted,
145 * and flags that will be used.
146 * If there is some data or critical controls (SYN, RST)
147 * to send, then transmit; otherwise, investigate further.
148 */
149 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
150 if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
151 /*
152 * We have been idle for "a while" and no acks are
153 * expected to clock out any data we send --
154 * slow start to get ack "clock" running again.
155 *
156 * Set the slow-start flight size depending on whether
157 * this is a local network or not.
158 */
159 int ss = ss_fltsz;
160 #ifdef INET6
161 if (isipv6) {
162 if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
163 ss = ss_fltsz_local;
164 } else
165 #endif /* INET6 */
166 if (in_localaddr(tp->t_inpcb->inp_faddr))
167 ss = ss_fltsz_local;
168 tp->snd_cwnd = tp->t_maxseg * ss;
169 }
170 tp->t_flags &= ~TF_LASTIDLE;
171 if (idle) {
172 if (tp->t_flags & TF_MORETOCOME) {
173 tp->t_flags |= TF_LASTIDLE;
174 idle = 0;
175 }
176 }
177 again:
178 /*
179 * If we've recently taken a timeout, snd_max will be greater than
180 * snd_nxt. There may be SACK information that allows us to avoid
181 * resending already delivered data. Adjust snd_nxt accordingly.
182 */
183 if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max))
184 tcp_sack_adjust(tp);
185 sendalot = 0;
186 off = tp->snd_nxt - tp->snd_una;
187 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
188 sendwin = min(sendwin, tp->snd_bwnd);
189
190 flags = tcp_outflags[tp->t_state];
191 /*
192 * Send any SACK-generated retransmissions. If we're explicitly trying
193 * to send out new data (when sendalot is 1), bypass this function.
194 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
195 * we're replacing a (future) new transmission with a retransmission
196 * now, and we previously incremented snd_cwnd in tcp_input().
197 */
198 /*
199 * Still in sack recovery , reset rxmit flag to zero.
200 */
201 sack_rxmit = 0;
202 sack_bytes_rxmt = 0;
203 len = 0;
204 p = NULL;
205 if (tp->sack_enable && IN_FASTRECOVERY(tp) &&
206 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
207 long cwin;
208
209 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
210 if (cwin < 0)
211 cwin = 0;
212 /* Do not retransmit SACK segments beyond snd_recover */
213 if (SEQ_GT(p->end, tp->snd_recover)) {
214 /*
215 * (At least) part of sack hole extends beyond
216 * snd_recover. Check to see if we can rexmit data
217 * for this hole.
218 */
219 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
220 /*
221 * Can't rexmit any more data for this hole.
222 * That data will be rexmitted in the next
223 * sack recovery episode, when snd_recover
224 * moves past p->rxmit.
225 */
226 p = NULL;
227 goto after_sack_rexmit;
228 } else
229 /* Can rexmit part of the current hole */
230 len = ((long)ulmin(cwin,
231 tp->snd_recover - p->rxmit));
232 } else
233 len = ((long)ulmin(cwin, p->end - p->rxmit));
234 off = p->rxmit - tp->snd_una;
235 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
236 __func__, off));
237 if (len > 0) {
238 sack_rxmit = 1;
239 sendalot = 1;
240 tcpstat.tcps_sack_rexmits++;
241 tcpstat.tcps_sack_rexmit_bytes +=
242 min(len, tp->t_maxseg);
243 }
244 }
245 after_sack_rexmit:
246 /*
247 * Get standard flags, and add SYN or FIN if requested by 'hidden'
248 * state flags.
249 */
250 if (tp->t_flags & TF_NEEDFIN)
251 flags |= TH_FIN;
252 if (tp->t_flags & TF_NEEDSYN)
253 flags |= TH_SYN;
254
255 SOCKBUF_LOCK(&so->so_snd);
256 /*
257 * If in persist timeout with window of 0, send 1 byte.
258 * Otherwise, if window is small but nonzero
259 * and timer expired, we will send what we can
260 * and go to transmit state.
261 */
262 if (tp->t_force) {
263 if (sendwin == 0) {
264 /*
265 * If we still have some data to send, then
266 * clear the FIN bit. Usually this would
267 * happen below when it realizes that we
268 * aren't sending all the data. However,
269 * if we have exactly 1 byte of unsent data,
270 * then it won't clear the FIN bit below,
271 * and if we are in persist state, we wind
272 * up sending the packet without recording
273 * that we sent the FIN bit.
274 *
275 * We can't just blindly clear the FIN bit,
276 * because if we don't have any more data
277 * to send then the probe will be the FIN
278 * itself.
279 */
280 if (off < so->so_snd.sb_cc)
281 flags &= ~TH_FIN;
282 sendwin = 1;
283 } else {
284 callout_stop(tp->tt_persist);
285 tp->t_rxtshift = 0;
286 }
287 }
288
289 /*
290 * If snd_nxt == snd_max and we have transmitted a FIN, the
291 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
292 * a negative length. This can also occur when TCP opens up
293 * its congestion window while receiving additional duplicate
294 * acks after fast-retransmit because TCP will reset snd_nxt
295 * to snd_max after the fast-retransmit.
296 *
297 * In the normal retransmit-FIN-only case, however, snd_nxt will
298 * be set to snd_una, the offset will be 0, and the length may
299 * wind up 0.
300 *
301 * If sack_rxmit is true we are retransmitting from the scoreboard
302 * in which case len is already set.
303 */
304 if (sack_rxmit == 0) {
305 if (sack_bytes_rxmt == 0)
306 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
307 else {
308 long cwin;
309
310 /*
311 * We are inside of a SACK recovery episode and are
312 * sending new data, having retransmitted all the
313 * data possible in the scoreboard.
314 */
315 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
316 - off);
317 /*
318 * Don't remove this (len > 0) check !
319 * We explicitly check for len > 0 here (although it
320 * isn't really necessary), to work around a gcc
321 * optimization issue - to force gcc to compute
322 * len above. Without this check, the computation
323 * of len is bungled by the optimizer.
324 */
325 if (len > 0) {
326 cwin = tp->snd_cwnd -
327 (tp->snd_nxt - tp->sack_newdata) -
328 sack_bytes_rxmt;
329 if (cwin < 0)
330 cwin = 0;
331 len = lmin(len, cwin);
332 }
333 }
334 }
335
336 /*
337 * Lop off SYN bit if it has already been sent. However, if this
338 * is SYN-SENT state and if segment contains data and if we don't
339 * know that foreign host supports TAO, suppress sending segment.
340 */
341 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
342 flags &= ~TH_SYN;
343 off--, len++;
344 if (tcp_do_rfc1644)
345 tcp_hc_gettao(&tp->t_inpcb->inp_inc, &tao);
346 if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
347 tao.tao_ccsent == 0)
348 goto just_return;
349 }
350
351 /*
352 * Be careful not to send data and/or FIN on SYN segments
353 * in cases when no CC option will be sent.
354 * This measure is needed to prevent interoperability problems
355 * with not fully conformant TCP implementations.
356 */
357 if ((flags & TH_SYN) &&
358 ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
359 ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
360 len = 0;
361 flags &= ~TH_FIN;
362 }
363
364 if (len < 0) {
365 /*
366 * If FIN has been sent but not acked,
367 * but we haven't been called to retransmit,
368 * len will be < 0. Otherwise, window shrank
369 * after we sent into it. If window shrank to 0,
370 * cancel pending retransmit, pull snd_nxt back
371 * to (closed) window, and set the persist timer
372 * if it isn't already going. If the window didn't
373 * close completely, just wait for an ACK.
374 */
375 len = 0;
376 if (sendwin == 0) {
377 callout_stop(tp->tt_rexmt);
378 tp->t_rxtshift = 0;
379 tp->snd_nxt = tp->snd_una;
380 if (!callout_active(tp->tt_persist))
381 tcp_setpersist(tp);
382 }
383 }
384
385 /*
386 * len will be >= 0 after this point. Truncate to the maximum
387 * segment length and ensure that FIN is removed if the length
388 * no longer contains the last data byte.
389 */
390 if (len > tp->t_maxseg) {
391 len = tp->t_maxseg;
392 sendalot = 1;
393 }
394 if (sack_rxmit) {
395 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
396 flags &= ~TH_FIN;
397 } else {
398 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
399 flags &= ~TH_FIN;
400 }
401
402 recwin = sbspace(&so->so_rcv);
403
404 /*
405 * Sender silly window avoidance. We transmit under the following
406 * conditions when len is non-zero:
407 *
408 * - We have a full segment
409 * - This is the last buffer in a write()/send() and we are
410 * either idle or running NODELAY
411 * - we've timed out (e.g. persist timer)
412 * - we have more then 1/2 the maximum send window's worth of
413 * data (receiver may be limited the window size)
414 * - we need to retransmit
415 */
416 if (len) {
417 if (len == tp->t_maxseg)
418 goto send;
419 /*
420 * NOTE! on localhost connections an 'ack' from the remote
421 * end may occur synchronously with the output and cause
422 * us to flush a buffer queued with moretocome. XXX
423 *
424 * note: the len + off check is almost certainly unnecessary.
425 */
426 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
427 (idle || (tp->t_flags & TF_NODELAY)) &&
428 len + off >= so->so_snd.sb_cc &&
429 (tp->t_flags & TF_NOPUSH) == 0) {
430 goto send;
431 }
432 if (tp->t_force) /* typ. timeout case */
433 goto send;
434 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
435 goto send;
436 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
437 goto send;
438 if (sack_rxmit)
439 goto send;
440 }
441
442 /*
443 * Compare available window to amount of window
444 * known to peer (as advertised window less
445 * next expected input). If the difference is at least two
446 * max size segments, or at least 50% of the maximum possible
447 * window, then want to send a window update to peer.
448 * Skip this if the connection is in T/TCP half-open state.
449 */
450 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) {
451 /*
452 * "adv" is the amount we can increase the window,
453 * taking into account that we are limited by
454 * TCP_MAXWIN << tp->rcv_scale.
455 */
456 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
457 (tp->rcv_adv - tp->rcv_nxt);
458
459 if (adv >= (long) (2 * tp->t_maxseg))
460 goto send;
461 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
462 goto send;
463 }
464
465 /*
466 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
467 * is also a catch-all for the retransmit timer timeout case.
468 */
469 if (tp->t_flags & TF_ACKNOW)
470 goto send;
471 if ((flags & TH_RST) ||
472 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
473 goto send;
474 if (SEQ_GT(tp->snd_up, tp->snd_una))
475 goto send;
476 /*
477 * If our state indicates that FIN should be sent
478 * and we have not yet done so, then we need to send.
479 */
480 if (flags & TH_FIN &&
481 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
482 goto send;
483 /*
484 * In SACK, it is possible for tcp_output to fail to send a segment
485 * after the retransmission timer has been turned off. Make sure
486 * that the retransmission timer is set.
487 */
488 if (tp->sack_enable && SEQ_GT(tp->snd_max, tp->snd_una) &&
489 !callout_active(tp->tt_rexmt) &&
490 !callout_active(tp->tt_persist)) {
491 callout_reset(tp->tt_rexmt, tp->t_rxtcur,
492 tcp_timer_rexmt, tp);
493 goto just_return;
494 }
495 /*
496 * TCP window updates are not reliable, rather a polling protocol
497 * using ``persist'' packets is used to insure receipt of window
498 * updates. The three ``states'' for the output side are:
499 * idle not doing retransmits or persists
500 * persisting to move a small or zero window
501 * (re)transmitting and thereby not persisting
502 *
503 * callout_active(tp->tt_persist)
504 * is true when we are in persist state.
505 * tp->t_force
506 * is set when we are called to send a persist packet.
507 * callout_active(tp->tt_rexmt)
508 * is set when we are retransmitting
509 * The output side is idle when both timers are zero.
510 *
511 * If send window is too small, there is data to transmit, and no
512 * retransmit or persist is pending, then go to persist state.
513 * If nothing happens soon, send when timer expires:
514 * if window is nonzero, transmit what we can,
515 * otherwise force out a byte.
516 */
517 if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
518 !callout_active(tp->tt_persist)) {
519 tp->t_rxtshift = 0;
520 tcp_setpersist(tp);
521 }
522
523 /*
524 * No reason to send a segment, just return.
525 */
526 just_return:
527 SOCKBUF_UNLOCK(&so->so_snd);
528 return (0);
529
530 send:
531 SOCKBUF_LOCK_ASSERT(&so->so_snd);
532 /*
533 * Before ESTABLISHED, force sending of initial options
534 * unless TCP set not to do any options.
535 * NOTE: we assume that the IP/TCP header plus TCP options
536 * always fit in a single mbuf, leaving room for a maximum
537 * link header, i.e.
538 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
539 */
540 optlen = 0;
541 #ifdef INET6
542 if (isipv6)
543 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
544 else
545 #endif
546 hdrlen = sizeof (struct tcpiphdr);
547 if (flags & TH_SYN) {
548 tp->snd_nxt = tp->iss;
549 if ((tp->t_flags & TF_NOOPT) == 0) {
550 u_short mss;
551
552 opt[0] = TCPOPT_MAXSEG;
553 opt[1] = TCPOLEN_MAXSEG;
554 mss = htons((u_short) tcp_mssopt(&tp->t_inpcb->inp_inc));
555 (void)memcpy(opt + 2, &mss, sizeof(mss));
556 optlen = TCPOLEN_MAXSEG;
557
558 /*
559 * If this is the first SYN of connection (not a SYN
560 * ACK), include SACK_PERMIT_HDR option. If this is a
561 * SYN ACK, include SACK_PERMIT_HDR option if peer has
562 * already done so. This is only for active connect,
563 * since the syncache takes care of the passive connect.
564 */
565 if (tp->sack_enable && ((flags & TH_ACK) == 0 ||
566 (tp->t_flags & TF_SACK_PERMIT))) {
567 *((u_int32_t *) (opt + optlen)) =
568 htonl(TCPOPT_SACK_PERMIT_HDR);
569 optlen += 4;
570 }
571 if ((tp->t_flags & TF_REQ_SCALE) &&
572 ((flags & TH_ACK) == 0 ||
573 (tp->t_flags & TF_RCVD_SCALE))) {
574 *((u_int32_t *)(opt + optlen)) = htonl(
575 TCPOPT_NOP << 24 |
576 TCPOPT_WINDOW << 16 |
577 TCPOLEN_WINDOW << 8 |
578 tp->request_r_scale);
579 optlen += 4;
580 }
581 }
582 }
583
584 /*
585 * Send a timestamp and echo-reply if this is a SYN and our side
586 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
587 * and our peer have sent timestamps in our SYN's.
588 */
589 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
590 (flags & TH_RST) == 0 &&
591 ((flags & TH_ACK) == 0 ||
592 (tp->t_flags & TF_RCVD_TSTMP))) {
593 u_int32_t *lp = (u_int32_t *)(opt + optlen);
594
595 /* Form timestamp option as shown in appendix A of RFC 1323. */
596 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
597 *lp++ = htonl(ticks);
598 *lp = htonl(tp->ts_recent);
599 optlen += TCPOLEN_TSTAMP_APPA;
600 }
601
602 /*
603 * Send SACKs if necessary. This should be the last option processed.
604 * Only as many SACKs are sent as are permitted by the maximum options
605 * size. No more than three SACKs are sent.
606 */
607 if (tp->sack_enable && tp->t_state == TCPS_ESTABLISHED &&
608 (tp->t_flags & (TF_SACK_PERMIT|TF_NOOPT)) == TF_SACK_PERMIT &&
609 tp->rcv_numsacks) {
610 u_int32_t *lp = (u_int32_t *)(opt + optlen);
611 u_int32_t *olp = lp++;
612 int count = 0; /* actual number of SACKs inserted */
613 int maxsack = (MAX_TCPOPTLEN - (optlen + 4))/TCPOLEN_SACK;
614
615 tcpstat.tcps_sack_send_blocks++;
616 maxsack = min(maxsack, TCP_MAX_SACK);
617 for (i = 0; (i < tp->rcv_numsacks && count < maxsack); i++) {
618 struct sackblk sack = tp->sackblks[i];
619 if (sack.start == 0 && sack.end == 0)
620 continue;
621 *lp++ = htonl(sack.start);
622 *lp++ = htonl(sack.end);
623 count++;
624 }
625 *olp = htonl(TCPOPT_SACK_HDR|(TCPOLEN_SACK*count+2));
626 optlen += TCPOLEN_SACK*count + 4; /* including leading NOPs */
627 }
628 /*
629 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
630 * options are allowed (!TF_NOOPT) and it's not a RST.
631 */
632 if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
633 (flags & TH_RST) == 0) {
634 switch (flags & (TH_SYN|TH_ACK)) {
635 /*
636 * This is a normal ACK, send CC if we received CC before
637 * from our peer.
638 */
639 case TH_ACK:
640 if (!(tp->t_flags & TF_RCVD_CC))
641 break;
642 /*FALLTHROUGH*/
643
644 /*
645 * We can only get here in T/TCP's SYN_SENT* state, when
646 * we're a sending a non-SYN segment without waiting for
647 * the ACK of our SYN. A check above assures that we only
648 * do this if our peer understands T/TCP.
649 */
650 case 0:
651 opt[optlen++] = TCPOPT_NOP;
652 opt[optlen++] = TCPOPT_NOP;
653 opt[optlen++] = TCPOPT_CC;
654 opt[optlen++] = TCPOLEN_CC;
655 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
656
657 optlen += 4;
658 break;
659
660 /*
661 * This is our initial SYN, check whether we have to use
662 * CC or CC.new.
663 */
664 case TH_SYN:
665 opt[optlen++] = TCPOPT_NOP;
666 opt[optlen++] = TCPOPT_NOP;
667 opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
668 TCPOPT_CCNEW : TCPOPT_CC;
669 opt[optlen++] = TCPOLEN_CC;
670 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
671 optlen += 4;
672 break;
673
674 /*
675 * This is a SYN,ACK; send CC and CC.echo if we received
676 * CC from our peer.
677 */
678 case (TH_SYN|TH_ACK):
679 if (tp->t_flags & TF_RCVD_CC) {
680 opt[optlen++] = TCPOPT_NOP;
681 opt[optlen++] = TCPOPT_NOP;
682 opt[optlen++] = TCPOPT_CC;
683 opt[optlen++] = TCPOLEN_CC;
684 *(u_int32_t *)&opt[optlen] =
685 htonl(tp->cc_send);
686 optlen += 4;
687 opt[optlen++] = TCPOPT_NOP;
688 opt[optlen++] = TCPOPT_NOP;
689 opt[optlen++] = TCPOPT_CCECHO;
690 opt[optlen++] = TCPOLEN_CC;
691 *(u_int32_t *)&opt[optlen] =
692 htonl(tp->cc_recv);
693 optlen += 4;
694 }
695 break;
696 }
697 }
698
699 #ifdef TCP_SIGNATURE
700 #ifdef INET6
701 if (!isipv6)
702 #endif
703 if (tp->t_flags & TF_SIGNATURE) {
704 int i;
705 u_char *bp;
706
707 /* Initialize TCP-MD5 option (RFC2385) */
708 bp = (u_char *)opt + optlen;
709 *bp++ = TCPOPT_SIGNATURE;
710 *bp++ = TCPOLEN_SIGNATURE;
711 sigoff = optlen + 2;
712 for (i = 0; i < TCP_SIGLEN; i++)
713 *bp++ = 0;
714 optlen += TCPOLEN_SIGNATURE;
715
716 /* Terminate options list and maintain 32-bit alignment. */
717 *bp++ = TCPOPT_NOP;
718 *bp++ = TCPOPT_EOL;
719 optlen += 2;
720 }
721 #endif /* TCP_SIGNATURE */
722
723 hdrlen += optlen;
724
725 #ifdef INET6
726 if (isipv6)
727 ipoptlen = ip6_optlen(tp->t_inpcb);
728 else
729 #endif
730 if (tp->t_inpcb->inp_options)
731 ipoptlen = tp->t_inpcb->inp_options->m_len -
732 offsetof(struct ipoption, ipopt_list);
733 else
734 ipoptlen = 0;
735 #ifdef IPSEC
736 ipoptlen += ipsec_hdrsiz_tcp(tp);
737 #endif
738
739 /*
740 * Adjust data length if insertion of options will
741 * bump the packet length beyond the t_maxopd length.
742 * Clear the FIN bit because we cut off the tail of
743 * the segment.
744 */
745 if (len + optlen + ipoptlen > tp->t_maxopd) {
746 /*
747 * If there is still more to send, don't close the connection.
748 */
749 flags &= ~TH_FIN;
750 len = tp->t_maxopd - optlen - ipoptlen;
751 sendalot = 1;
752 }
753
754 /*#ifdef DIAGNOSTIC*/
755 #ifdef INET6
756 if (max_linkhdr + hdrlen > MCLBYTES)
757 #else
758 if (max_linkhdr + hdrlen > MHLEN)
759 #endif
760 panic("tcphdr too big");
761 /*#endif*/
762
763 /*
764 * Grab a header mbuf, attaching a copy of data to
765 * be transmitted, and initialize the header from
766 * the template for sends on this connection.
767 */
768 if (len) {
769 if (tp->t_force && len == 1)
770 tcpstat.tcps_sndprobe++;
771 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
772 tcpstat.tcps_sndrexmitpack++;
773 tcpstat.tcps_sndrexmitbyte += len;
774 } else {
775 tcpstat.tcps_sndpack++;
776 tcpstat.tcps_sndbyte += len;
777 }
778 #ifdef notyet
779 if ((m = m_copypack(so->so_snd.sb_mb, off,
780 (int)len, max_linkhdr + hdrlen)) == 0) {
781 SOCKBUF_UNLOCK(&so->so_snd);
782 error = ENOBUFS;
783 goto out;
784 }
785 /*
786 * m_copypack left space for our hdr; use it.
787 */
788 m->m_len += hdrlen;
789 m->m_data -= hdrlen;
790 #else
791 MGETHDR(m, M_DONTWAIT, MT_HEADER);
792 if (m == NULL) {
793 SOCKBUF_UNLOCK(&so->so_snd);
794 error = ENOBUFS;
795 goto out;
796 }
797 #ifdef INET6
798 if (MHLEN < hdrlen + max_linkhdr) {
799 MCLGET(m, M_DONTWAIT);
800 if ((m->m_flags & M_EXT) == 0) {
801 SOCKBUF_UNLOCK(&so->so_snd);
802 m_freem(m);
803 error = ENOBUFS;
804 goto out;
805 }
806 }
807 #endif
808 m->m_data += max_linkhdr;
809 m->m_len = hdrlen;
810 if (len <= MHLEN - hdrlen - max_linkhdr) {
811 m_copydata(so->so_snd.sb_mb, off, (int) len,
812 mtod(m, caddr_t) + hdrlen);
813 m->m_len += len;
814 } else {
815 m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
816 if (m->m_next == 0) {
817 SOCKBUF_UNLOCK(&so->so_snd);
818 (void) m_free(m);
819 error = ENOBUFS;
820 goto out;
821 }
822 }
823 #endif
824 /*
825 * If we're sending everything we've got, set PUSH.
826 * (This will keep happy those implementations which only
827 * give data to the user when a buffer fills or
828 * a PUSH comes in.)
829 */
830 if (off + len == so->so_snd.sb_cc)
831 flags |= TH_PUSH;
832 SOCKBUF_UNLOCK(&so->so_snd);
833 } else {
834 SOCKBUF_UNLOCK(&so->so_snd);
835 if (tp->t_flags & TF_ACKNOW)
836 tcpstat.tcps_sndacks++;
837 else if (flags & (TH_SYN|TH_FIN|TH_RST))
838 tcpstat.tcps_sndctrl++;
839 else if (SEQ_GT(tp->snd_up, tp->snd_una))
840 tcpstat.tcps_sndurg++;
841 else
842 tcpstat.tcps_sndwinup++;
843
844 MGETHDR(m, M_DONTWAIT, MT_HEADER);
845 if (m == NULL) {
846 error = ENOBUFS;
847 goto out;
848 }
849 #ifdef INET6
850 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
851 MHLEN >= hdrlen) {
852 MH_ALIGN(m, hdrlen);
853 } else
854 #endif
855 m->m_data += max_linkhdr;
856 m->m_len = hdrlen;
857 }
858 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
859 m->m_pkthdr.rcvif = (struct ifnet *)0;
860 #ifdef MAC
861 mac_create_mbuf_from_inpcb(tp->t_inpcb, m);
862 #endif
863 #ifdef INET6
864 if (isipv6) {
865 ip6 = mtod(m, struct ip6_hdr *);
866 th = (struct tcphdr *)(ip6 + 1);
867 tcpip_fillheaders(tp->t_inpcb, ip6, th);
868 } else
869 #endif /* INET6 */
870 {
871 ip = mtod(m, struct ip *);
872 ipov = (struct ipovly *)ip;
873 th = (struct tcphdr *)(ip + 1);
874 tcpip_fillheaders(tp->t_inpcb, ip, th);
875 }
876
877 /*
878 * Fill in fields, remembering maximum advertised
879 * window for use in delaying messages about window sizes.
880 * If resending a FIN, be sure not to use a new sequence number.
881 */
882 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
883 tp->snd_nxt == tp->snd_max)
884 tp->snd_nxt--;
885 /*
886 * If we are doing retransmissions, then snd_nxt will
887 * not reflect the first unsent octet. For ACK only
888 * packets, we do not want the sequence number of the
889 * retransmitted packet, we want the sequence number
890 * of the next unsent octet. So, if there is no data
891 * (and no SYN or FIN), use snd_max instead of snd_nxt
892 * when filling in ti_seq. But if we are in persist
893 * state, snd_max might reflect one byte beyond the
894 * right edge of the window, so use snd_nxt in that
895 * case, since we know we aren't doing a retransmission.
896 * (retransmit and persist are mutually exclusive...)
897 */
898 if (sack_rxmit == 0) {
899 if (len || (flags & (TH_SYN|TH_FIN))
900 || callout_active(tp->tt_persist))
901 th->th_seq = htonl(tp->snd_nxt);
902 else
903 th->th_seq = htonl(tp->snd_max);
904 } else {
905 th->th_seq = htonl(p->rxmit);
906 p->rxmit += len;
907 }
908 th->th_ack = htonl(tp->rcv_nxt);
909 if (optlen) {
910 bcopy(opt, th + 1, optlen);
911 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
912 }
913 th->th_flags = flags;
914 /*
915 * Calculate receive window. Don't shrink window,
916 * but avoid silly window syndrome.
917 */
918 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
919 recwin < (long)tp->t_maxseg)
920 recwin = 0;
921 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
922 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
923 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
924 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
925 th->th_win = htons((u_short) (recwin >> tp->rcv_scale));
926
927
928 /*
929 * Adjust the RXWIN0SENT flag - indicate that we have advertised
930 * a 0 window. This may cause the remote transmitter to stall. This
931 * flag tells soreceive() to disable delayed acknowledgements when
932 * draining the buffer. This can occur if the receiver is attempting
933 * to read more data then can be buffered prior to transmitting on
934 * the connection.
935 */
936 if (recwin == 0)
937 tp->t_flags |= TF_RXWIN0SENT;
938 else
939 tp->t_flags &= ~TF_RXWIN0SENT;
940 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
941 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
942 th->th_flags |= TH_URG;
943 } else
944 /*
945 * If no urgent pointer to send, then we pull
946 * the urgent pointer to the left edge of the send window
947 * so that it doesn't drift into the send window on sequence
948 * number wraparound.
949 */
950 tp->snd_up = tp->snd_una; /* drag it along */
951
952 #ifdef TCP_SIGNATURE
953 #ifdef INET6
954 if (!isipv6)
955 #endif
956 if (tp->t_flags & TF_SIGNATURE)
957 tcp_signature_compute(m, sizeof(struct ip), len, optlen,
958 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
959 #endif
960
961 /*
962 * Put TCP length in extended header, and then
963 * checksum extended header and data.
964 */
965 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
966 #ifdef INET6
967 if (isipv6)
968 /*
969 * ip6_plen is not need to be filled now, and will be filled
970 * in ip6_output.
971 */
972 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
973 sizeof(struct tcphdr) + optlen + len);
974 else
975 #endif /* INET6 */
976 {
977 m->m_pkthdr.csum_flags = CSUM_TCP;
978 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
979 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
980 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
981
982 /* IP version must be set here for ipv4/ipv6 checking later */
983 KASSERT(ip->ip_v == IPVERSION,
984 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
985 }
986
987 /*
988 * In transmit state, time the transmission and arrange for
989 * the retransmit. In persist state, just set snd_max.
990 */
991 if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
992 tcp_seq startseq = tp->snd_nxt;
993
994 /*
995 * Advance snd_nxt over sequence space of this segment.
996 */
997 if (flags & (TH_SYN|TH_FIN)) {
998 if (flags & TH_SYN)
999 tp->snd_nxt++;
1000 if (flags & TH_FIN) {
1001 tp->snd_nxt++;
1002 tp->t_flags |= TF_SENTFIN;
1003 }
1004 }
1005 if (sack_rxmit)
1006 goto timer;
1007 tp->snd_nxt += len;
1008 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1009 tp->snd_max = tp->snd_nxt;
1010 /*
1011 * Time this transmission if not a retransmission and
1012 * not currently timing anything.
1013 */
1014 if (tp->t_rtttime == 0) {
1015 tp->t_rtttime = ticks;
1016 tp->t_rtseq = startseq;
1017 tcpstat.tcps_segstimed++;
1018 }
1019 }
1020
1021 /*
1022 * Set retransmit timer if not currently set,
1023 * and not doing a pure ack or a keep-alive probe.
1024 * Initial value for retransmit timer is smoothed
1025 * round-trip time + 2 * round-trip time variance.
1026 * Initialize shift counter which is used for backoff
1027 * of retransmit time.
1028 */
1029 timer:
1030 if (!callout_active(tp->tt_rexmt) &&
1031 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1032 (tp->snd_nxt != tp->snd_una))) {
1033 if (callout_active(tp->tt_persist)) {
1034 callout_stop(tp->tt_persist);
1035 tp->t_rxtshift = 0;
1036 }
1037 callout_reset(tp->tt_rexmt, tp->t_rxtcur,
1038 tcp_timer_rexmt, tp);
1039 }
1040 } else {
1041 /*
1042 * Persist case, update snd_max but since we are in
1043 * persist mode (no window) we do not update snd_nxt.
1044 */
1045 int xlen = len;
1046 if (flags & TH_SYN)
1047 ++xlen;
1048 if (flags & TH_FIN) {
1049 ++xlen;
1050 tp->t_flags |= TF_SENTFIN;
1051 }
1052 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1053 tp->snd_max = tp->snd_nxt + len;
1054 }
1055
1056 #ifdef TCPDEBUG
1057 /*
1058 * Trace.
1059 */
1060 if (so->so_options & SO_DEBUG) {
1061 u_short save = 0;
1062 #ifdef INET6
1063 if (!isipv6)
1064 #endif
1065 {
1066 save = ipov->ih_len;
1067 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1068 }
1069 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1070 #ifdef INET6
1071 if (!isipv6)
1072 #endif
1073 ipov->ih_len = save;
1074 }
1075 #endif
1076
1077 /*
1078 * Fill in IP length and desired time to live and
1079 * send to IP level. There should be a better way
1080 * to handle ttl and tos; we could keep them in
1081 * the template, but need a way to checksum without them.
1082 */
1083 /*
1084 * m->m_pkthdr.len should have been set before cksum calcuration,
1085 * because in6_cksum() need it.
1086 */
1087 #ifdef INET6
1088 if (isipv6) {
1089 /*
1090 * we separately set hoplimit for every segment, since the
1091 * user might want to change the value via setsockopt.
1092 * Also, desired default hop limit might be changed via
1093 * Neighbor Discovery.
1094 */
1095 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1096
1097 /* TODO: IPv6 IP6TOS_ECT bit on */
1098 error = ip6_output(m,
1099 tp->t_inpcb->in6p_outputopts, NULL,
1100 (so->so_options & SO_DONTROUTE), NULL, NULL,
1101 tp->t_inpcb);
1102 } else
1103 #endif /* INET6 */
1104 {
1105 ip->ip_len = m->m_pkthdr.len;
1106 #ifdef INET6
1107 if (INP_CHECK_SOCKAF(so, AF_INET6))
1108 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1109 #endif /* INET6 */
1110 /*
1111 * If we do path MTU discovery, then we set DF on every packet.
1112 * This might not be the best thing to do according to RFC3390
1113 * Section 2. However the tcp hostcache migitates the problem
1114 * so it affects only the first tcp connection with a host.
1115 */
1116 if (path_mtu_discovery)
1117 ip->ip_off |= IP_DF;
1118
1119 error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1120 (so->so_options & SO_DONTROUTE), 0, tp->t_inpcb);
1121 }
1122 if (error) {
1123
1124 /*
1125 * We know that the packet was lost, so back out the
1126 * sequence number advance, if any.
1127 */
1128 if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
1129 /*
1130 * No need to check for TH_FIN here because
1131 * the TF_SENTFIN flag handles that case.
1132 */
1133 if ((flags & TH_SYN) == 0) {
1134 if (sack_rxmit)
1135 p->rxmit -= len;
1136 else
1137 tp->snd_nxt -= len;
1138 }
1139 }
1140
1141 out:
1142 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1143 if (error == ENOBUFS) {
1144 if (!callout_active(tp->tt_rexmt) &&
1145 !callout_active(tp->tt_persist))
1146 callout_reset(tp->tt_rexmt, tp->t_rxtcur,
1147 tcp_timer_rexmt, tp);
1148 tp->snd_cwnd = tp->t_maxseg;
1149 return (0);
1150 }
1151 if (error == EMSGSIZE) {
1152 /*
1153 * ip_output() will have already fixed the route
1154 * for us. tcp_mtudisc() will, as its last action,
1155 * initiate retransmission, so it is important to
1156 * not do so here.
1157 */
1158 tcp_mtudisc(tp->t_inpcb, 0);
1159 return 0;
1160 }
1161 if ((error == EHOSTUNREACH || error == ENETDOWN)
1162 && TCPS_HAVERCVDSYN(tp->t_state)) {
1163 tp->t_softerror = error;
1164 return (0);
1165 }
1166 return (error);
1167 }
1168 tcpstat.tcps_sndtotal++;
1169
1170 /*
1171 * Data sent (as far as we can tell).
1172 * If this advertises a larger window than any other segment,
1173 * then remember the size of the advertised window.
1174 * Any pending ACK has now been sent.
1175 */
1176 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1177 tp->rcv_adv = tp->rcv_nxt + recwin;
1178 tp->last_ack_sent = tp->rcv_nxt;
1179 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1180 if (callout_active(tp->tt_delack))
1181 callout_stop(tp->tt_delack);
1182 #if 0
1183 /*
1184 * This completely breaks TCP if newreno is turned on. What happens
1185 * is that if delayed-acks are turned on on the receiver, this code
1186 * on the transmitter effectively destroys the TCP window, forcing
1187 * it to four packets (1.5Kx4 = 6K window).
1188 */
1189 if (sendalot && (!tcp_do_newreno || --maxburst))
1190 goto again;
1191 #endif
1192 if (sendalot)
1193 goto again;
1194 return (0);
1195 }
1196
1197 void
1198 tcp_setpersist(tp)
1199 register struct tcpcb *tp;
1200 {
1201 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1202 int tt;
1203
1204 if (callout_active(tp->tt_rexmt))
1205 panic("tcp_setpersist: retransmit pending");
1206 /*
1207 * Start/restart persistance timer.
1208 */
1209 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1210 TCPTV_PERSMIN, TCPTV_PERSMAX);
1211 callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1212 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1213 tp->t_rxtshift++;
1214 }
Cache object: 0090cc79db10a4b251eadc63868c7473
|