1 /* $OpenBSD: tcp_output.c,v 1.145 2024/05/14 09:39:02 bluhm Exp $ */
2 /* $NetBSD: tcp_output.c,v 1.16 1997/06/03 16:17:09 kml Exp $ */
3
4 /*
5 * Copyright (c) 1982, 1986, 1988, 1990, 1993
6 * The Regents of the University of California. All rights reserved.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 *
32 * @(#)COPYRIGHT 1.1 (NRL) 17 January 1995
33 *
34 * NRL grants permission for redistribution and use in source and binary
35 * forms, with or without modification, of the software and documentation
36 * created at NRL provided that the following conditions are met:
37 *
38 * 1. Redistributions of source code must retain the above copyright
39 * notice, this list of conditions and the following disclaimer.
40 * 2. Redistributions in binary form must reproduce the above copyright
41 * notice, this list of conditions and the following disclaimer in the
42 * documentation and/or other materials provided with the distribution.
43 * 3. All advertising materials mentioning features or use of this software
44 * must display the following acknowledgements:
45 * This product includes software developed by the University of
46 * California, Berkeley and its contributors.
47 * This product includes software developed at the Information
48 * Technology Division, US Naval Research Laboratory.
49 * 4. Neither the name of the NRL nor the names of its contributors
50 * may be used to endorse or promote products derived from this software
51 * without specific prior written permission.
52 *
53 * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS
54 * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
55 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
56 * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NRL OR
57 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
58 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
59 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
60 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
61 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
62 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
63 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
64 *
65 * The views and conclusions contained in the software and documentation
66 * are those of the authors and should not be interpreted as representing
67 * official policies, either expressed or implied, of the US Naval
68 * Research Laboratory (NRL).
69 */
70
71 #include "pf.h"
72 #include "stoeplitz.h"
73
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/mbuf.h>
77 #include <sys/protosw.h>
78 #include <sys/socket.h>
79 #include <sys/socketvar.h>
80 #include <sys/kernel.h>
81
82 #include <net/if.h>
83 #include <net/if_var.h>
84 #include <net/route.h>
85 #if NPF > 0
86 #include <net/pfvar.h>
87 #endif
88
89 #include <netinet/in.h>
90 #include <netinet/ip.h>
91 #include <netinet/in_pcb.h>
92 #include <netinet/ip_var.h>
93 #include <netinet6/ip6_var.h>
94 #include <netinet/tcp.h>
95 #define TCPOUTFLAGS
96 #include <netinet/tcp_fsm.h>
97 #include <netinet/tcp_seq.h>
98 #include <netinet/tcp_timer.h>
99 #include <netinet/tcp_var.h>
100 #include <netinet/tcp_debug.h>
101
102 #ifdef notyet
103 extern struct mbuf *m_copypack();
104 #endif
105
106 extern int tcprexmtthresh;
107
108 #ifdef TCP_SACK_DEBUG
109 void tcp_print_holes(struct tcpcb *tp);
110
111 void
tcp_print_holes(struct tcpcb * tp)112 tcp_print_holes(struct tcpcb *tp)
113 {
114 struct sackhole *p = tp->snd_holes;
115 if (p == NULL)
116 return;
117 printf("Hole report: start--end dups rxmit\n");
118 while (p) {
119 printf("%x--%x d %d r %x\n", p->start, p->end, p->dups,
120 p->rxmit);
121 p = p->next;
122 }
123 printf("\n");
124 }
125 #endif /* TCP_SACK_DEBUG */
126
127 /*
128 * Returns pointer to a sackhole if there are any pending retransmissions;
129 * NULL otherwise.
130 */
131 struct sackhole *
tcp_sack_output(struct tcpcb * tp)132 tcp_sack_output(struct tcpcb *tp)
133 {
134 struct sackhole *p;
135
136 if (!tp->sack_enable)
137 return (NULL);
138 p = tp->snd_holes;
139 while (p) {
140 if (p->dups >= tcprexmtthresh && SEQ_LT(p->rxmit, p->end)) {
141 if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
142 p = p->next;
143 continue;
144 }
145 #ifdef TCP_SACK_DEBUG
146 if (p)
147 tcp_print_holes(tp);
148 #endif
149 return (p);
150 }
151 p = p->next;
152 }
153 return (NULL);
154 }
155
156 /*
157 * After a timeout, the SACK list may be rebuilt. This SACK information
158 * should be used to avoid retransmitting SACKed data. This function
159 * traverses the SACK list to see if snd_nxt should be moved forward.
160 */
161
162 void
tcp_sack_adjust(struct tcpcb * tp)163 tcp_sack_adjust(struct tcpcb *tp)
164 {
165 struct sackhole *cur = tp->snd_holes;
166 if (cur == NULL)
167 return; /* No holes */
168 if (SEQ_GEQ(tp->snd_nxt, tp->rcv_lastsack))
169 return; /* We're already beyond any SACKed blocks */
170 /*
171 * Two cases for which we want to advance snd_nxt:
172 * i) snd_nxt lies between end of one hole and beginning of another
173 * ii) snd_nxt lies between end of last hole and rcv_lastsack
174 */
175 while (cur->next) {
176 if (SEQ_LT(tp->snd_nxt, cur->end))
177 return;
178 if (SEQ_GEQ(tp->snd_nxt, cur->next->start))
179 cur = cur->next;
180 else {
181 tp->snd_nxt = cur->next->start;
182 return;
183 }
184 }
185 if (SEQ_LT(tp->snd_nxt, cur->end))
186 return;
187 tp->snd_nxt = tp->rcv_lastsack;
188 return;
189 }
190
191 /*
192 * Tcp output routine: figure out what should be sent and send it.
193 */
194 int
tcp_output(struct tcpcb * tp)195 tcp_output(struct tcpcb *tp)
196 {
197 struct socket *so = tp->t_inpcb->inp_socket;
198 long len, win, txmaxseg;
199 int off, flags, error;
200 struct mbuf *m;
201 struct tcphdr *th;
202 u_int32_t optbuf[howmany(MAX_TCPOPTLEN, sizeof(u_int32_t))];
203 u_char *opt = (u_char *)optbuf;
204 unsigned int optlen, hdrlen, packetlen;
205 int idle, sendalot = 0;
206 int i, sack_rxmit = 0;
207 struct sackhole *p;
208 uint64_t now;
209 #ifdef TCP_SIGNATURE
210 unsigned int sigoff;
211 #endif /* TCP_SIGNATURE */
212 #ifdef TCP_ECN
213 int needect;
214 #endif
215 int tso;
216
217 if (tp->t_flags & TF_BLOCKOUTPUT) {
218 tp->t_flags |= TF_NEEDOUTPUT;
219 return (0);
220 } else
221 tp->t_flags &= ~TF_NEEDOUTPUT;
222
223 #if defined(TCP_SIGNATURE) && defined(DIAGNOSTIC)
224 if (tp->sack_enable && (tp->t_flags & TF_SIGNATURE))
225 return (EINVAL);
226 #endif /* defined(TCP_SIGNATURE) && defined(DIAGNOSTIC) */
227
228 now = tcp_now();
229
230 /*
231 * Determine length of data that should be transmitted,
232 * and flags that will be used.
233 * If there is some data or critical controls (SYN, RST)
234 * to send, then transmit; otherwise, investigate further.
235 */
236 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
237 if (idle && (now - tp->t_rcvtime) >= tp->t_rxtcur)
238 /*
239 * We have been idle for "a while" and no acks are
240 * expected to clock out any data we send --
241 * slow start to get ack "clock" running again.
242 */
243 tp->snd_cwnd = 2 * tp->t_maxseg;
244
245 /* remember 'idle' for next invocation of tcp_output */
246 if (idle && soissending(so)) {
247 tp->t_flags |= TF_LASTIDLE;
248 idle = 0;
249 } else
250 tp->t_flags &= ~TF_LASTIDLE;
251
252 again:
253 /*
254 * If we've recently taken a timeout, snd_max will be greater than
255 * snd_nxt. There may be SACK information that allows us to avoid
256 * resending already delivered data. Adjust snd_nxt accordingly.
257 */
258 if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max))
259 tcp_sack_adjust(tp);
260 off = tp->snd_nxt - tp->snd_una;
261 win = ulmin(tp->snd_wnd, tp->snd_cwnd);
262
263 flags = tcp_outflags[tp->t_state];
264
265 /*
266 * Send any SACK-generated retransmissions. If we're explicitly trying
267 * to send out new data (when sendalot is 1), bypass this function.
268 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
269 * we're replacing a (future) new transmission with a retransmission
270 * now, and we previously incremented snd_cwnd in tcp_input().
271 */
272 if (tp->sack_enable && !sendalot) {
273 if (tp->t_dupacks >= tcprexmtthresh &&
274 (p = tcp_sack_output(tp))) {
275 off = p->rxmit - tp->snd_una;
276 sack_rxmit = 1;
277 /* Coalesce holes into a single retransmission */
278 len = min(tp->t_maxseg, p->end - p->rxmit);
279 if (SEQ_LT(tp->snd_una, tp->snd_last))
280 tp->snd_cwnd -= tp->t_maxseg;
281 }
282 }
283
284 sendalot = 0;
285 tso = 0;
286 /*
287 * If in persist timeout with window of 0, send 1 byte.
288 * Otherwise, if window is small but nonzero
289 * and timer expired, we will send what we can
290 * and go to transmit state.
291 */
292 if (tp->t_force) {
293 if (win == 0) {
294 /*
295 * If we still have some data to send, then
296 * clear the FIN bit. Usually this would
297 * happen below when it realizes that we
298 * aren't sending all the data. However,
299 * if we have exactly 1 byte of unset data,
300 * then it won't clear the FIN bit below,
301 * and if we are in persist state, we wind
302 * up sending the packet without recording
303 * that we sent the FIN bit.
304 *
305 * We can't just blindly clear the FIN bit,
306 * because if we don't have any more data
307 * to send then the probe will be the FIN
308 * itself.
309 */
310 if (off < so->so_snd.sb_cc)
311 flags &= ~TH_FIN;
312 win = 1;
313 } else {
314 TCP_TIMER_DISARM(tp, TCPT_PERSIST);
315 tp->t_rxtshift = 0;
316 }
317 }
318
319 if (!sack_rxmit) {
320 len = ulmin(so->so_snd.sb_cc, win) - off;
321 }
322
323 if (len < 0) {
324 /*
325 * If FIN has been sent but not acked,
326 * but we haven't been called to retransmit,
327 * len will be -1. Otherwise, window shrank
328 * after we sent into it. If window shrank to 0,
329 * cancel pending retransmit, pull snd_nxt back
330 * to (closed) window, and set the persist timer
331 * if it isn't already going. If the window didn't
332 * close completely, just wait for an ACK.
333 */
334 len = 0;
335 if (win == 0) {
336 TCP_TIMER_DISARM(tp, TCPT_REXMT);
337 tp->t_rxtshift = 0;
338 tp->snd_nxt = tp->snd_una;
339 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
340 tcp_setpersist(tp);
341 }
342 }
343
344 /*
345 * Never send more than half a buffer full. This insures that we can
346 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
347 * therefore acks will never be delayed unless we run out of data to
348 * transmit.
349 */
350 txmaxseg = ulmin(so->so_snd.sb_hiwat / 2, tp->t_maxseg);
351
352 if (len > txmaxseg) {
353 if (tcp_do_tso &&
354 tp->t_inpcb->inp_options == NULL &&
355 tp->t_inpcb->inp_outputopts6 == NULL &&
356 #ifdef TCP_SIGNATURE
357 ((tp->t_flags & TF_SIGNATURE) == 0) &&
358 #endif
359 len >= 2 * tp->t_maxseg &&
360 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
361 !(flags & (TH_SYN|TH_RST|TH_FIN))) {
362 tso = 1;
363 /* avoid small chopped packets */
364 if (len > (len / tp->t_maxseg) * tp->t_maxseg) {
365 len = (len / tp->t_maxseg) * tp->t_maxseg;
366 sendalot = 1;
367 }
368 } else {
369 len = txmaxseg;
370 sendalot = 1;
371 }
372 }
373 if (off + len < so->so_snd.sb_cc)
374 flags &= ~TH_FIN;
375
376 win = sbspace(so, &so->so_rcv);
377
378 /*
379 * Sender silly window avoidance. If connection is idle
380 * and can send all data, a maximum segment,
381 * at least a maximum default-size segment do it,
382 * or are forced, do it; otherwise don't bother.
383 * If peer's buffer is tiny, then send
384 * when window is at least half open.
385 * If retransmitting (possibly after persist timer forced us
386 * to send into a small window), then must resend.
387 */
388 if (len) {
389 if (len >= txmaxseg)
390 goto send;
391 if ((idle || (tp->t_flags & TF_NODELAY)) &&
392 len + off >= so->so_snd.sb_cc && !soissending(so) &&
393 (tp->t_flags & TF_NOPUSH) == 0)
394 goto send;
395 if (tp->t_force)
396 goto send;
397 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
398 goto send;
399 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
400 goto send;
401 if (sack_rxmit)
402 goto send;
403 }
404
405 /*
406 * Compare available window to amount of window
407 * known to peer (as advertised window less
408 * next expected input). If the difference is at least two
409 * max size segments, or at least 50% of the maximum possible
410 * window, then want to send a window update to peer.
411 */
412 if (win > 0) {
413 /*
414 * "adv" is the amount we can increase the window,
415 * taking into account that we are limited by
416 * TCP_MAXWIN << tp->rcv_scale.
417 */
418 long adv = lmin(win, (long)TCP_MAXWIN << tp->rcv_scale) -
419 (tp->rcv_adv - tp->rcv_nxt);
420
421 if (adv >= (long) (2 * tp->t_maxseg))
422 goto send;
423 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
424 goto send;
425 }
426
427 /*
428 * Send if we owe peer an ACK.
429 */
430 if (tp->t_flags & TF_ACKNOW)
431 goto send;
432 if (flags & (TH_SYN|TH_RST))
433 goto send;
434 if (SEQ_GT(tp->snd_up, tp->snd_una))
435 goto send;
436 /*
437 * If our state indicates that FIN should be sent
438 * and we have not yet done so, or we're retransmitting the FIN,
439 * then we need to send.
440 */
441 if (flags & TH_FIN &&
442 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
443 goto send;
444 /*
445 * In SACK, it is possible for tcp_output to fail to send a segment
446 * after the retransmission timer has been turned off. Make sure
447 * that the retransmission timer is set.
448 */
449 if (SEQ_GT(tp->snd_max, tp->snd_una) &&
450 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
451 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
452 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
453 return (0);
454 }
455
456 /*
457 * TCP window updates are not reliable, rather a polling protocol
458 * using ``persist'' packets is used to insure receipt of window
459 * updates. The three ``states'' for the output side are:
460 * idle not doing retransmits or persists
461 * persisting to move a small or zero window
462 * (re)transmitting and thereby not persisting
463 *
464 * tp->t_timer[TCPT_PERSIST]
465 * is set when we are in persist state.
466 * tp->t_force
467 * is set when we are called to send a persist packet.
468 * tp->t_timer[TCPT_REXMT]
469 * is set when we are retransmitting
470 * The output side is idle when both timers are zero.
471 *
472 * If send window is too small, there is data to transmit, and no
473 * retransmit or persist is pending, then go to persist state.
474 * If nothing happens soon, send when timer expires:
475 * if window is nonzero, transmit what we can,
476 * otherwise force out a byte.
477 */
478 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
479 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
480 tp->t_rxtshift = 0;
481 tcp_setpersist(tp);
482 }
483
484 /*
485 * No reason to send a segment, just return.
486 */
487 return (0);
488
489 send:
490 /*
491 * Before ESTABLISHED, force sending of initial options
492 * unless TCP set not to do any options.
493 * NOTE: we assume that the IP/TCP header plus TCP options
494 * always fit in a single mbuf, leaving room for a maximum
495 * link header, i.e.
496 * max_linkhdr + sizeof(network header) + sizeof(struct tcphdr +
497 * optlen <= MHLEN
498 */
499 optlen = 0;
500
501 switch (tp->pf) {
502 case 0: /*default to PF_INET*/
503 case PF_INET:
504 hdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
505 break;
506 #ifdef INET6
507 case PF_INET6:
508 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
509 break;
510 #endif /* INET6 */
511 default:
512 return (EPFNOSUPPORT);
513 }
514
515 if (flags & TH_SYN) {
516 tp->snd_nxt = tp->iss;
517 if ((tp->t_flags & TF_NOOPT) == 0) {
518 u_int16_t mss;
519
520 opt[0] = TCPOPT_MAXSEG;
521 opt[1] = 4;
522 mss = htons((u_int16_t) tcp_mss(tp, 0));
523 memcpy(opt + 2, &mss, sizeof(mss));
524 optlen = 4;
525
526 if (flags & TH_ACK)
527 tcp_mss_update(tp);
528 /*
529 * If this is the first SYN of connection (not a SYN
530 * ACK), include SACK_PERMIT_HDR option. If this is a
531 * SYN ACK, include SACK_PERMIT_HDR option if peer has
532 * already done so.
533 */
534 if (tp->sack_enable && ((flags & TH_ACK) == 0 ||
535 (tp->t_flags & TF_SACK_PERMIT))) {
536 *((u_int32_t *) (opt + optlen)) =
537 htonl(TCPOPT_SACK_PERMIT_HDR);
538 optlen += 4;
539 }
540 if ((tp->t_flags & TF_REQ_SCALE) &&
541 ((flags & TH_ACK) == 0 ||
542 (tp->t_flags & TF_RCVD_SCALE))) {
543 *((u_int32_t *) (opt + optlen)) = htonl(
544 TCPOPT_NOP << 24 |
545 TCPOPT_WINDOW << 16 |
546 TCPOLEN_WINDOW << 8 |
547 tp->request_r_scale);
548 optlen += 4;
549 }
550 }
551 }
552
553 /*
554 * Send a timestamp and echo-reply if this is a SYN and our side
555 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
556 * and our peer have sent timestamps in our SYN's.
557 */
558 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
559 (flags & TH_RST) == 0 &&
560 ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
561 (tp->t_flags & TF_RCVD_TSTMP))) {
562 u_int32_t *lp = (u_int32_t *)(opt + optlen);
563
564 /* Form timestamp option as shown in appendix A of RFC 1323. */
565 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
566 *lp++ = htonl(now + tp->ts_modulate);
567 *lp = htonl(tp->ts_recent);
568 optlen += TCPOLEN_TSTAMP_APPA;
569 }
570 /* Set receive buffer autosizing timestamp. */
571 if (tp->rfbuf_ts == 0) {
572 tp->rfbuf_ts = now;
573 tp->rfbuf_cnt = 0;
574 }
575
576 #ifdef TCP_SIGNATURE
577 if (tp->t_flags & TF_SIGNATURE) {
578 u_int8_t *bp = (u_int8_t *)(opt + optlen);
579
580 /* Send signature option */
581 *(bp++) = TCPOPT_SIGNATURE;
582 *(bp++) = TCPOLEN_SIGNATURE;
583 sigoff = optlen + 2;
584
585 {
586 unsigned int i;
587
588 for (i = 0; i < 16; i++)
589 *(bp++) = 0;
590 }
591
592
593 /* Pad options list to the next 32 bit boundary and
594 * terminate it.
595 */
596 *bp++ = TCPOPT_NOP;
597 *bp++ = TCPOPT_NOP;
598
599 optlen += TCPOLEN_SIGLEN;
600 }
601 #endif /* TCP_SIGNATURE */
602
603 /*
604 * Send SACKs if necessary. This should be the last option processed.
605 * Only as many SACKs are sent as are permitted by the maximum options
606 * size. No more than three SACKs are sent.
607 */
608 if (tp->sack_enable && tp->t_state == TCPS_ESTABLISHED &&
609 (tp->t_flags & (TF_SACK_PERMIT|TF_NOOPT)) == TF_SACK_PERMIT &&
610 tp->rcv_numsacks) {
611 u_int32_t *lp = (u_int32_t *)(opt + optlen);
612 u_int32_t *olp = lp++;
613 int count = 0; /* actual number of SACKs inserted */
614 int maxsack = (MAX_TCPOPTLEN - (optlen + 4))/TCPOLEN_SACK;
615
616 tcpstat_inc(tcps_sack_snd_opts);
617 maxsack = min(maxsack, TCP_MAX_SACK);
618 for (i = 0; (i < tp->rcv_numsacks && count < maxsack); i++) {
619 struct sackblk sack = tp->sackblks[i];
620 if (sack.start == 0 && sack.end == 0)
621 continue;
622 *lp++ = htonl(sack.start);
623 *lp++ = htonl(sack.end);
624 count++;
625 }
626 *olp = htonl(TCPOPT_SACK_HDR|(TCPOLEN_SACK*count+2));
627 optlen += TCPOLEN_SACK*count + 4; /* including leading NOPs */
628 }
629
630 #ifdef DIAGNOSTIC
631 if (optlen > MAX_TCPOPTLEN)
632 panic("tcp_output: options too long");
633 #endif /* DIAGNOSTIC */
634
635 hdrlen += optlen;
636
637 /*
638 * Adjust data length if insertion of options will
639 * bump the packet length beyond the t_maxopd length.
640 * Clear the FIN bit because we cut off the tail of
641 * the segment.
642 */
643 if (len > tp->t_maxopd - optlen) {
644 if (tso) {
645 if (len + hdrlen + max_linkhdr > MAXMCLBYTES) {
646 len = MAXMCLBYTES - hdrlen - max_linkhdr;
647 sendalot = 1;
648 }
649 } else {
650 len = tp->t_maxopd - optlen;
651 sendalot = 1;
652 }
653 flags &= ~TH_FIN;
654 }
655
656 #ifdef DIAGNOSTIC
657 if (max_linkhdr + hdrlen > MCLBYTES)
658 panic("tcphdr too big");
659 #endif
660
661 /*
662 * Grab a header mbuf, attaching a copy of data to
663 * be transmitted, and initialize the header from
664 * the template for sends on this connection.
665 */
666 if (len) {
667 if (tp->t_force && len == 1)
668 tcpstat_inc(tcps_sndprobe);
669 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
670 tcpstat_pkt(tcps_sndrexmitpack, tcps_sndrexmitbyte,
671 len);
672 tp->t_sndrexmitpack++;
673 } else {
674 tcpstat_pkt(tcps_sndpack, tcps_sndbyte, len);
675 }
676 #ifdef notyet
677 if ((m = m_copypack(so->so_snd.sb_mb, off,
678 (int)len, max_linkhdr + hdrlen)) == 0) {
679 error = ENOBUFS;
680 goto out;
681 }
682 /*
683 * m_copypack left space for our hdr; use it.
684 */
685 m->m_len += hdrlen;
686 m->m_data -= hdrlen;
687 #else
688 MGETHDR(m, M_DONTWAIT, MT_HEADER);
689 if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
690 MCLGET(m, M_DONTWAIT);
691 if ((m->m_flags & M_EXT) == 0) {
692 m_freem(m);
693 m = NULL;
694 }
695 }
696 if (m == NULL) {
697 error = ENOBUFS;
698 goto out;
699 }
700 m->m_data += max_linkhdr;
701 m->m_len = hdrlen;
702 if (len <= m_trailingspace(m)) {
703 m_copydata(so->so_snd.sb_mb, off, (int) len,
704 mtod(m, caddr_t) + hdrlen);
705 m->m_len += len;
706 } else {
707 m->m_next = m_copym(so->so_snd.sb_mb, off, (int) len,
708 M_NOWAIT);
709 if (m->m_next == 0) {
710 (void) m_free(m);
711 error = ENOBUFS;
712 goto out;
713 }
714 }
715 if (so->so_snd.sb_mb->m_flags & M_PKTHDR)
716 m->m_pkthdr.ph_loopcnt =
717 so->so_snd.sb_mb->m_pkthdr.ph_loopcnt;
718 #endif
719 /*
720 * If we're sending everything we've got, set PUSH.
721 * (This will keep happy those implementations which only
722 * give data to the user when a buffer fills or
723 * a PUSH comes in.)
724 */
725 if (off + len == so->so_snd.sb_cc && !soissending(so))
726 flags |= TH_PUSH;
727 tp->t_sndtime = now;
728 } else {
729 if (tp->t_flags & TF_ACKNOW)
730 tcpstat_inc(tcps_sndacks);
731 else if (flags & (TH_SYN|TH_FIN|TH_RST))
732 tcpstat_inc(tcps_sndctrl);
733 else if (SEQ_GT(tp->snd_up, tp->snd_una))
734 tcpstat_inc(tcps_sndurg);
735 else
736 tcpstat_inc(tcps_sndwinup);
737
738 MGETHDR(m, M_DONTWAIT, MT_HEADER);
739 if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
740 MCLGET(m, M_DONTWAIT);
741 if ((m->m_flags & M_EXT) == 0) {
742 m_freem(m);
743 m = NULL;
744 }
745 }
746 if (m == NULL) {
747 error = ENOBUFS;
748 goto out;
749 }
750 m->m_data += max_linkhdr;
751 m->m_len = hdrlen;
752 }
753 m->m_pkthdr.ph_ifidx = 0;
754 m->m_pkthdr.len = hdrlen + len;
755
756 /* Enable TSO and specify the size of the resulting segments. */
757 if (tso) {
758 SET(m->m_pkthdr.csum_flags, M_TCP_TSO);
759 m->m_pkthdr.ph_mss = tp->t_maxseg;
760 }
761
762 if (!tp->t_template)
763 panic("tcp_output");
764 #ifdef DIAGNOSTIC
765 if (tp->t_template->m_len != hdrlen - optlen)
766 panic("tcp_output: template len != hdrlen - optlen");
767 #endif /* DIAGNOSTIC */
768 memcpy(mtod(m, caddr_t), mtod(tp->t_template, caddr_t),
769 tp->t_template->m_len);
770 th = (struct tcphdr *)(mtod(m, caddr_t) + tp->t_template->m_len -
771 sizeof(struct tcphdr));
772
773 /*
774 * Fill in fields, remembering maximum advertised
775 * window for use in delaying messages about window sizes.
776 * If resending a FIN, be sure not to use a new sequence number.
777 */
778 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
779 (tp->snd_nxt == tp->snd_max))
780 tp->snd_nxt--;
781 /*
782 * If we are doing retransmissions, then snd_nxt will
783 * not reflect the first unsent octet. For ACK only
784 * packets, we do not want the sequence number of the
785 * retransmitted packet, we want the sequence number
786 * of the next unsent octet. So, if there is no data
787 * (and no SYN or FIN), use snd_max instead of snd_nxt
788 * when filling in ti_seq. But if we are in persist
789 * state, snd_max might reflect one byte beyond the
790 * right edge of the window, so use snd_nxt in that
791 * case, since we know we aren't doing a retransmission.
792 * (retransmit and persist are mutually exclusive...)
793 */
794 if (len || (flags & (TH_SYN|TH_FIN)) ||
795 TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
796 th->th_seq = htonl(tp->snd_nxt);
797 else
798 th->th_seq = htonl(tp->snd_max);
799
800 if (sack_rxmit) {
801 /*
802 * If sendalot was turned on (due to option stuffing), turn it
803 * off. Properly set th_seq field. Advance the ret'x pointer
804 * by len.
805 */
806 if (sendalot)
807 sendalot = 0;
808 th->th_seq = htonl(p->rxmit);
809 p->rxmit += len;
810 tcpstat_pkt(tcps_sack_rexmits, tcps_sack_rexmit_bytes, len);
811 }
812
813 th->th_ack = htonl(tp->rcv_nxt);
814 if (optlen) {
815 memcpy(th + 1, opt, optlen);
816 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
817 }
818 #ifdef TCP_ECN
819 if (tcp_do_ecn) {
820 /*
821 * if we have received congestion experienced segs,
822 * set ECE bit.
823 */
824 if (tp->t_flags & TF_RCVD_CE) {
825 flags |= TH_ECE;
826 tcpstat_inc(tcps_ecn_sndece);
827 }
828 if (!(tp->t_flags & TF_DISABLE_ECN)) {
829 /*
830 * if this is a SYN seg, set ECE and CWR.
831 * set only ECE for SYN-ACK if peer supports ECN.
832 */
833 if ((flags & (TH_SYN|TH_ACK)) == TH_SYN)
834 flags |= (TH_ECE|TH_CWR);
835 else if ((tp->t_flags & TF_ECN_PERMIT) &&
836 (flags & (TH_SYN|TH_ACK)) == (TH_SYN|TH_ACK))
837 flags |= TH_ECE;
838 }
839 /*
840 * if we have reduced the congestion window, notify
841 * the peer by setting CWR bit.
842 */
843 if ((tp->t_flags & TF_ECN_PERMIT) &&
844 (tp->t_flags & TF_SEND_CWR)) {
845 flags |= TH_CWR;
846 tp->t_flags &= ~TF_SEND_CWR;
847 tcpstat_inc(tcps_ecn_sndcwr);
848 }
849 }
850 #endif
851 th->th_flags = flags;
852
853 /*
854 * Calculate receive window. Don't shrink window,
855 * but avoid silly window syndrome.
856 */
857 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
858 win = 0;
859 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
860 win = (long)TCP_MAXWIN << tp->rcv_scale;
861 if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
862 win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
863 if (flags & TH_RST)
864 win = 0;
865 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
866 if (th->th_win == 0)
867 tp->t_sndzerowin++;
868 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
869 u_int32_t urp = tp->snd_up - tp->snd_nxt;
870 if (urp > IP_MAXPACKET)
871 urp = IP_MAXPACKET;
872 th->th_urp = htons((u_int16_t)urp);
873 th->th_flags |= TH_URG;
874 } else
875 /*
876 * If no urgent pointer to send, then we pull
877 * the urgent pointer to the left edge of the send window
878 * so that it doesn't drift into the send window on sequence
879 * number wraparound.
880 */
881 tp->snd_up = tp->snd_una; /* drag it along */
882
883 #ifdef TCP_SIGNATURE
884 if (tp->t_flags & TF_SIGNATURE) {
885 int iphlen;
886 union sockaddr_union src, dst;
887 struct tdb *tdb;
888
889 bzero(&src, sizeof(union sockaddr_union));
890 bzero(&dst, sizeof(union sockaddr_union));
891
892 switch (tp->pf) {
893 case 0: /*default to PF_INET*/
894 case AF_INET:
895 iphlen = sizeof(struct ip);
896 src.sa.sa_len = sizeof(struct sockaddr_in);
897 src.sa.sa_family = AF_INET;
898 src.sin.sin_addr = mtod(m, struct ip *)->ip_src;
899 dst.sa.sa_len = sizeof(struct sockaddr_in);
900 dst.sa.sa_family = AF_INET;
901 dst.sin.sin_addr = mtod(m, struct ip *)->ip_dst;
902 break;
903 #ifdef INET6
904 case AF_INET6:
905 iphlen = sizeof(struct ip6_hdr);
906 src.sa.sa_len = sizeof(struct sockaddr_in6);
907 src.sa.sa_family = AF_INET6;
908 src.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_src;
909 dst.sa.sa_len = sizeof(struct sockaddr_in6);
910 dst.sa.sa_family = AF_INET6;
911 dst.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_dst;
912 break;
913 #endif /* INET6 */
914 }
915
916 tdb = gettdbbysrcdst(rtable_l2(tp->t_inpcb->inp_rtableid),
917 0, &src, &dst, IPPROTO_TCP);
918 if (tdb == NULL) {
919 m_freem(m);
920 return (EPERM);
921 }
922
923 if (tcp_signature(tdb, tp->pf, m, th, iphlen, 0,
924 mtod(m, caddr_t) + hdrlen - optlen + sigoff) < 0) {
925 m_freem(m);
926 tdb_unref(tdb);
927 return (EINVAL);
928 }
929 tdb_unref(tdb);
930 }
931 #endif /* TCP_SIGNATURE */
932
933 /* Defer checksumming until later (ip_output() or hardware) */
934 m->m_pkthdr.csum_flags |= M_TCP_CSUM_OUT;
935
936 /*
937 * In transmit state, time the transmission and arrange for
938 * the retransmit. In persist state, just set snd_max.
939 */
940 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
941 tcp_seq startseq = tp->snd_nxt;
942
943 /*
944 * Advance snd_nxt over sequence space of this segment.
945 */
946 if (flags & (TH_SYN|TH_FIN)) {
947 if (flags & TH_SYN)
948 tp->snd_nxt++;
949 if (flags & TH_FIN) {
950 tp->snd_nxt++;
951 tp->t_flags |= TF_SENTFIN;
952 }
953 }
954 if (tp->sack_enable) {
955 if (sack_rxmit && (p->rxmit != tp->snd_nxt)) {
956 goto timer;
957 }
958 }
959 tp->snd_nxt += len;
960 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
961 tp->snd_max = tp->snd_nxt;
962 /*
963 * Time this transmission if not a retransmission and
964 * not currently timing anything.
965 */
966 if (tp->t_rtttime == 0) {
967 tp->t_rtttime = now;
968 tp->t_rtseq = startseq;
969 tcpstat_inc(tcps_segstimed);
970 }
971 }
972
973 /*
974 * Set retransmit timer if not currently set,
975 * and not doing an ack or a keep-alive probe.
976 * Initial value for retransmit timer is smoothed
977 * round-trip time + 2 * round-trip time variance.
978 * Initialize shift counter which is used for backoff
979 * of retransmit time.
980 */
981 timer:
982 if (tp->sack_enable && sack_rxmit &&
983 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
984 tp->snd_nxt != tp->snd_max) {
985 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
986 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
987 TCP_TIMER_DISARM(tp, TCPT_PERSIST);
988 tp->t_rxtshift = 0;
989 }
990 }
991
992 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
993 tp->snd_nxt != tp->snd_una) {
994 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
995 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
996 TCP_TIMER_DISARM(tp, TCPT_PERSIST);
997 tp->t_rxtshift = 0;
998 }
999 }
1000
1001 if (len == 0 && so->so_snd.sb_cc &&
1002 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1003 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1004 /*
1005 * Avoid a situation where we do not set persist timer
1006 * after a zero window condition. For example:
1007 * 1) A -> B: packet with enough data to fill the window
1008 * 2) B -> A: ACK for #1 + new data (0 window
1009 * advertisement)
1010 * 3) A -> B: ACK for #2, 0 len packet
1011 *
1012 * In this case, A will not activate the persist timer,
1013 * because it chose to send a packet. Unless tcp_output
1014 * is called for some other reason (delayed ack timer,
1015 * another input packet from B, socket syscall), A will
1016 * not send zero window probes.
1017 *
1018 * So, if you send a 0-length packet, but there is data
1019 * in the socket buffer, and neither the rexmt or
1020 * persist timer is already set, then activate the
1021 * persist timer.
1022 */
1023 tp->t_rxtshift = 0;
1024 tcp_setpersist(tp);
1025 }
1026 } else
1027 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1028 tp->snd_max = tp->snd_nxt + len;
1029
1030 tcp_update_sndspace(tp);
1031
1032 /*
1033 * Trace.
1034 */
1035 if (so->so_options & SO_DEBUG)
1036 tcp_trace(TA_OUTPUT, tp->t_state, tp, tp, mtod(m, caddr_t), 0,
1037 len);
1038
1039 /*
1040 * Fill in IP length and desired time to live and
1041 * send to IP level. There should be a better way
1042 * to handle ttl and tos; we could keep them in
1043 * the template, but need a way to checksum without them.
1044 */
1045
1046 #ifdef TCP_ECN
1047 /*
1048 * if peer is ECN capable, set the ECT bit in the IP header.
1049 * but don't set ECT for a pure ack, a retransmit or a window probe.
1050 */
1051 needect = 0;
1052 if (tcp_do_ecn && (tp->t_flags & TF_ECN_PERMIT)) {
1053 if (len == 0 || SEQ_LT(tp->snd_nxt, tp->snd_max) ||
1054 (tp->t_force && len == 1)) {
1055 /* don't set ECT */
1056 } else {
1057 needect = 1;
1058 tcpstat_inc(tcps_ecn_sndect);
1059 }
1060 }
1061 #endif
1062
1063 /* force routing table */
1064 m->m_pkthdr.ph_rtableid = tp->t_inpcb->inp_rtableid;
1065
1066 #if NPF > 0
1067 pf_mbuf_link_inpcb(m, tp->t_inpcb);
1068 #endif
1069
1070 switch (tp->pf) {
1071 case 0: /*default to PF_INET*/
1072 case AF_INET:
1073 {
1074 struct ip *ip;
1075
1076 ip = mtod(m, struct ip *);
1077 ip->ip_len = htons(m->m_pkthdr.len);
1078 packetlen = m->m_pkthdr.len;
1079 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1080 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1081 #ifdef TCP_ECN
1082 if (needect)
1083 ip->ip_tos |= IPTOS_ECN_ECT0;
1084 #endif
1085 }
1086 #if NSTOEPLITZ > 0
1087 m->m_pkthdr.ph_flowid = tp->t_inpcb->inp_flowid;
1088 SET(m->m_pkthdr.csum_flags, M_FLOWID);
1089 #endif
1090 error = ip_output(m, tp->t_inpcb->inp_options,
1091 &tp->t_inpcb->inp_route,
1092 (ip_mtudisc ? IP_MTUDISC : 0), NULL,
1093 &tp->t_inpcb->inp_seclevel, 0);
1094 break;
1095 #ifdef INET6
1096 case AF_INET6:
1097 {
1098 struct ip6_hdr *ip6;
1099
1100 ip6 = mtod(m, struct ip6_hdr *);
1101 ip6->ip6_plen = m->m_pkthdr.len -
1102 sizeof(struct ip6_hdr);
1103 packetlen = m->m_pkthdr.len;
1104 ip6->ip6_nxt = IPPROTO_TCP;
1105 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb);
1106 #ifdef TCP_ECN
1107 if (needect)
1108 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1109 #endif
1110 }
1111 error = ip6_output(m, tp->t_inpcb->inp_outputopts6,
1112 &tp->t_inpcb->inp_route, 0, NULL,
1113 &tp->t_inpcb->inp_seclevel);
1114 break;
1115 #endif /* INET6 */
1116 }
1117
1118 if (error) {
1119 out:
1120 if (error == ENOBUFS) {
1121 /*
1122 * If the interface queue is full, or IP cannot
1123 * get an mbuf, trigger TCP slow start.
1124 */
1125 tp->snd_cwnd = tp->t_maxseg;
1126 return (0);
1127 }
1128 if (error == EMSGSIZE) {
1129 /*
1130 * ip_output() will have already fixed the route
1131 * for us. tcp_mtudisc() will, as its last action,
1132 * initiate retransmission, so it is important to
1133 * not do so here.
1134 */
1135 tcp_mtudisc(tp->t_inpcb, -1);
1136 return (0);
1137 }
1138 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1139 TCPS_HAVERCVDSYN(tp->t_state)) {
1140 tp->t_softerror = error;
1141 return (0);
1142 }
1143
1144 /* Restart the delayed ACK timer, if necessary. */
1145 if (TCP_TIMER_ISARMED(tp, TCPT_DELACK))
1146 TCP_TIMER_ARM(tp, TCPT_DELACK, tcp_delack_msecs);
1147
1148 return (error);
1149 }
1150
1151 if (packetlen > tp->t_pmtud_mtu_sent)
1152 tp->t_pmtud_mtu_sent = packetlen;
1153
1154 tcpstat_inc(tcps_sndtotal);
1155 if (TCP_TIMER_ISARMED(tp, TCPT_DELACK))
1156 tcpstat_inc(tcps_delack);
1157
1158 /*
1159 * Data sent (as far as we can tell).
1160 * If this advertises a larger window than any other segment,
1161 * then remember the size of the advertised window.
1162 * Any pending ACK has now been sent.
1163 */
1164 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1165 tp->rcv_adv = tp->rcv_nxt + win;
1166 tp->last_ack_sent = tp->rcv_nxt;
1167 tp->t_sndacktime = now;
1168 tp->t_flags &= ~TF_ACKNOW;
1169 TCP_TIMER_DISARM(tp, TCPT_DELACK);
1170 if (sendalot)
1171 goto again;
1172 return (0);
1173 }
1174
1175 void
tcp_setpersist(struct tcpcb * tp)1176 tcp_setpersist(struct tcpcb *tp)
1177 {
1178 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + TCP_RTT_BASE_SHIFT);
1179 int msec;
1180
1181 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1182 panic("tcp_output REXMT");
1183 /*
1184 * Start/restart persistence timer.
1185 */
1186 if (t < tp->t_rttmin)
1187 t = tp->t_rttmin;
1188 TCPT_RANGESET(msec, t * tcp_backoff[tp->t_rxtshift],
1189 TCPTV_PERSMIN, TCPTV_PERSMAX);
1190 TCP_TIMER_ARM(tp, TCPT_PERSIST, msec);
1191 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1192 tp->t_rxtshift++;
1193 }
1194
1195 int
tcp_chopper(struct mbuf * m0,struct mbuf_list * ml,struct ifnet * ifp,u_int mss)1196 tcp_chopper(struct mbuf *m0, struct mbuf_list *ml, struct ifnet *ifp,
1197 u_int mss)
1198 {
1199 struct ip *ip = NULL;
1200 #ifdef INET6
1201 struct ip6_hdr *ip6 = NULL;
1202 #endif
1203 struct tcphdr *th;
1204 int firstlen, iphlen, hlen, tlen, off;
1205 int error;
1206
1207 ml_init(ml);
1208 ml_enqueue(ml, m0);
1209
1210 if (mss == 0) {
1211 error = EINVAL;
1212 goto bad;
1213 }
1214
1215 ip = mtod(m0, struct ip *);
1216 switch (ip->ip_v) {
1217 case 4:
1218 iphlen = ip->ip_hl << 2;
1219 if (ISSET(ip->ip_off, htons(IP_OFFMASK | IP_MF)) ||
1220 iphlen != sizeof(struct ip) || ip->ip_p != IPPROTO_TCP) {
1221 /* only TCP without fragment or IP option supported */
1222 error = EPROTOTYPE;
1223 goto bad;
1224 }
1225 break;
1226 #ifdef INET6
1227 case 6:
1228 ip = NULL;
1229 ip6 = mtod(m0, struct ip6_hdr *);
1230 iphlen = sizeof(struct ip6_hdr);
1231 if (ip6->ip6_nxt != IPPROTO_TCP) {
1232 /* only TCP without IPv6 header chain supported */
1233 error = EPROTOTYPE;
1234 goto bad;
1235 }
1236 break;
1237 #endif
1238 default:
1239 panic("%s: unknown ip version %d", __func__, ip->ip_v);
1240 }
1241
1242 tlen = m0->m_pkthdr.len;
1243 if (tlen < iphlen + sizeof(struct tcphdr)) {
1244 error = ENOPROTOOPT;
1245 goto bad;
1246 }
1247 /* IP and TCP header should be contiguous, this check is paranoia */
1248 if (m0->m_len < iphlen + sizeof(*th)) {
1249 ml_dequeue(ml);
1250 if ((m0 = m_pullup(m0, iphlen + sizeof(*th))) == NULL) {
1251 error = ENOBUFS;
1252 goto bad;
1253 }
1254 ml_enqueue(ml, m0);
1255 }
1256 th = (struct tcphdr *)(mtod(m0, caddr_t) + iphlen);
1257 hlen = iphlen + (th->th_off << 2);
1258 if (tlen < hlen) {
1259 error = ENOPROTOOPT;
1260 goto bad;
1261 }
1262 firstlen = MIN(tlen - hlen, mss);
1263
1264 CLR(m0->m_pkthdr.csum_flags, M_TCP_TSO);
1265
1266 /*
1267 * Loop through length of payload after first segment,
1268 * make new header and copy data of each part and link onto chain.
1269 */
1270 for (off = hlen + firstlen; off < tlen; off += mss) {
1271 struct mbuf *m;
1272 struct tcphdr *mhth;
1273 int len;
1274
1275 len = MIN(tlen - off, mss);
1276
1277 MGETHDR(m, M_DONTWAIT, MT_HEADER);
1278 if (m == NULL) {
1279 error = ENOBUFS;
1280 goto bad;
1281 }
1282 ml_enqueue(ml, m);
1283 if ((error = m_dup_pkthdr(m, m0, M_DONTWAIT)) != 0)
1284 goto bad;
1285
1286 /* IP and TCP header to the end, space for link layer header */
1287 m->m_len = hlen;
1288 m_align(m, hlen);
1289
1290 /* copy and adjust TCP header */
1291 mhth = (struct tcphdr *)(mtod(m, caddr_t) + iphlen);
1292 memcpy(mhth, th, hlen - iphlen);
1293 mhth->th_seq = htonl(ntohl(th->th_seq) + (off - hlen));
1294 if (off + len < tlen)
1295 CLR(mhth->th_flags, TH_PUSH|TH_FIN);
1296
1297 /* add mbuf chain with payload */
1298 m->m_pkthdr.len = hlen + len;
1299 if ((m->m_next = m_copym(m0, off, len, M_DONTWAIT)) == NULL) {
1300 error = ENOBUFS;
1301 goto bad;
1302 }
1303
1304 /* copy and adjust IP header, calculate checksum */
1305 SET(m->m_pkthdr.csum_flags, M_TCP_CSUM_OUT);
1306 if (ip) {
1307 struct ip *mhip;
1308
1309 mhip = mtod(m, struct ip *);
1310 *mhip = *ip;
1311 mhip->ip_len = htons(hlen + len);
1312 mhip->ip_id = htons(ip_randomid());
1313 in_hdr_cksum_out(m, ifp);
1314 in_proto_cksum_out(m, ifp);
1315 }
1316 #ifdef INET6
1317 if (ip6) {
1318 struct ip6_hdr *mhip6;
1319
1320 mhip6 = mtod(m, struct ip6_hdr *);
1321 *mhip6 = *ip6;
1322 mhip6->ip6_plen = htons(hlen - iphlen + len);
1323 in6_proto_cksum_out(m, ifp);
1324 }
1325 #endif
1326 }
1327
1328 /*
1329 * Update first segment by trimming what's been copied out
1330 * and updating header, then send each segment (in order).
1331 */
1332 if (hlen + firstlen < tlen) {
1333 m_adj(m0, hlen + firstlen - tlen);
1334 CLR(th->th_flags, TH_PUSH|TH_FIN);
1335 }
1336 /* adjust IP header, calculate checksum */
1337 SET(m0->m_pkthdr.csum_flags, M_TCP_CSUM_OUT);
1338 if (ip) {
1339 ip->ip_len = htons(m0->m_pkthdr.len);
1340 in_hdr_cksum_out(m0, ifp);
1341 in_proto_cksum_out(m0, ifp);
1342 }
1343 #ifdef INET6
1344 if (ip6) {
1345 ip6->ip6_plen = htons(m0->m_pkthdr.len - iphlen);
1346 in6_proto_cksum_out(m0, ifp);
1347 }
1348 #endif
1349
1350 tcpstat_add(tcps_outpkttso, ml_len(ml));
1351 return 0;
1352
1353 bad:
1354 tcpstat_inc(tcps_outbadtso);
1355 ml_purge(ml);
1356 return error;
1357 }
1358
1359 int
tcp_if_output_tso(struct ifnet * ifp,struct mbuf ** mp,struct sockaddr * dst,struct rtentry * rt,uint32_t ifcap,u_int mtu)1360 tcp_if_output_tso(struct ifnet *ifp, struct mbuf **mp, struct sockaddr *dst,
1361 struct rtentry *rt, uint32_t ifcap, u_int mtu)
1362 {
1363 struct mbuf_list ml;
1364 int error;
1365
1366 /* caller must fail later or fragment */
1367 if (!ISSET((*mp)->m_pkthdr.csum_flags, M_TCP_TSO))
1368 return 0;
1369 if ((*mp)->m_pkthdr.ph_mss > mtu) {
1370 CLR((*mp)->m_pkthdr.csum_flags, M_TCP_TSO);
1371 return 0;
1372 }
1373
1374 /* network interface hardware will do TSO */
1375 if (in_ifcap_cksum(*mp, ifp, ifcap)) {
1376 if (ISSET(ifcap, IFCAP_TSOv4)) {
1377 in_hdr_cksum_out(*mp, ifp);
1378 in_proto_cksum_out(*mp, ifp);
1379 }
1380 #ifdef INET6
1381 if (ISSET(ifcap, IFCAP_TSOv6))
1382 in6_proto_cksum_out(*mp, ifp);
1383 #endif
1384 error = ifp->if_output(ifp, *mp, dst, rt);
1385 if (!error)
1386 tcpstat_inc(tcps_outhwtso);
1387 goto done;
1388 }
1389
1390 /* as fallback do TSO in software */
1391 if ((error = tcp_chopper(*mp, &ml, ifp, (*mp)->m_pkthdr.ph_mss)) ||
1392 (error = if_output_ml(ifp, &ml, dst, rt)))
1393 goto done;
1394 tcpstat_inc(tcps_outswtso);
1395
1396 done:
1397 *mp = NULL;
1398 return error;
1399 }
1400