xref: /original-bsd/sys/netinet/tcp_output.c (revision b44716a6)
1 /*
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * %sccs.include.redist.c%
6  *
7  *	@(#)tcp_output.c	8.4 (Berkeley) 05/24/95
8  */
9 
10 #include <sys/param.h>
11 #include <sys/systm.h>
12 #include <sys/malloc.h>
13 #include <sys/mbuf.h>
14 #include <sys/protosw.h>
15 #include <sys/socket.h>
16 #include <sys/socketvar.h>
17 #include <sys/errno.h>
18 
19 #include <net/route.h>
20 
21 #include <netinet/in.h>
22 #include <netinet/in_systm.h>
23 #include <netinet/ip.h>
24 #include <netinet/in_pcb.h>
25 #include <netinet/ip_var.h>
26 #include <netinet/tcp.h>
27 #define	TCPOUTFLAGS
28 #include <netinet/tcp_fsm.h>
29 #include <netinet/tcp_seq.h>
30 #include <netinet/tcp_timer.h>
31 #include <netinet/tcp_var.h>
32 #include <netinet/tcpip.h>
33 #include <netinet/tcp_debug.h>
34 
35 #ifdef notyet
36 extern struct mbuf *m_copypack();
37 #endif
38 
39 
40 #define MAX_TCPOPTLEN	32	/* max # bytes that go in options */
41 
42 /*
43  * Tcp output routine: figure out what should be sent and send it.
44  */
45 int
tcp_output(tp)46 tcp_output(tp)
47 	register struct tcpcb *tp;
48 {
49 	register struct socket *so = tp->t_inpcb->inp_socket;
50 	register long len, win;
51 	int off, flags, error;
52 	register struct mbuf *m;
53 	register struct tcpiphdr *ti;
54 	u_char opt[MAX_TCPOPTLEN];
55 	unsigned optlen, hdrlen;
56 	int idle, sendalot;
57 
58 	/*
59 	 * Determine length of data that should be transmitted,
60 	 * and flags that will be used.
61 	 * If there is some data or critical controls (SYN, RST)
62 	 * to send, then transmit; otherwise, investigate further.
63 	 */
64 	idle = (tp->snd_max == tp->snd_una);
65 	if (idle && tp->t_idle >= tp->t_rxtcur)
66 		/*
67 		 * We have been idle for "a while" and no acks are
68 		 * expected to clock out any data we send --
69 		 * slow start to get ack "clock" running again.
70 		 */
71 		tp->snd_cwnd = tp->t_maxseg;
72 again:
73 	sendalot = 0;
74 	off = tp->snd_nxt - tp->snd_una;
75 	win = min(tp->snd_wnd, tp->snd_cwnd);
76 
77 	flags = tcp_outflags[tp->t_state];
78 	/*
79 	 * If in persist timeout with window of 0, send 1 byte.
80 	 * Otherwise, if window is small but nonzero
81 	 * and timer expired, we will send what we can
82 	 * and go to transmit state.
83 	 */
84 	if (tp->t_force) {
85 		if (win == 0) {
86 			/*
87 			 * If we still have some data to send, then
88 			 * clear the FIN bit.  Usually this would
89 			 * happen below when it realizes that we
90 			 * aren't sending all the data.  However,
91 			 * if we have exactly 1 byte of unset data,
92 			 * then it won't clear the FIN bit below,
93 			 * and if we are in persist state, we wind
94 			 * up sending the packet without recording
95 			 * that we sent the FIN bit.
96 			 *
97 			 * We can't just blindly clear the FIN bit,
98 			 * because if we don't have any more data
99 			 * to send then the probe will be the FIN
100 			 * itself.
101 			 */
102 			if (off < so->so_snd.sb_cc)
103 				flags &= ~TH_FIN;
104 			win = 1;
105 		} else {
106 			tp->t_timer[TCPT_PERSIST] = 0;
107 			tp->t_rxtshift = 0;
108 		}
109 	}
110 
111 	len = min(so->so_snd.sb_cc, win) - off;
112 
113 	if (len < 0) {
114 		/*
115 		 * If FIN has been sent but not acked,
116 		 * but we haven't been called to retransmit,
117 		 * len will be -1.  Otherwise, window shrank
118 		 * after we sent into it.  If window shrank to 0,
119 		 * cancel pending retransmit and pull snd_nxt
120 		 * back to (closed) window.  We will enter persist
121 		 * state below.  If the window didn't close completely,
122 		 * just wait for an ACK.
123 		 */
124 		len = 0;
125 		if (win == 0) {
126 			tp->t_timer[TCPT_REXMT] = 0;
127 			tp->snd_nxt = tp->snd_una;
128 		}
129 	}
130 	if (len > tp->t_maxseg) {
131 		len = tp->t_maxseg;
132 		sendalot = 1;
133 	}
134 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
135 		flags &= ~TH_FIN;
136 
137 	win = sbspace(&so->so_rcv);
138 
139 	/*
140 	 * Sender silly window avoidance.  If connection is idle
141 	 * and can send all data, a maximum segment,
142 	 * at least a maximum default-size segment do it,
143 	 * or are forced, do it; otherwise don't bother.
144 	 * If peer's buffer is tiny, then send
145 	 * when window is at least half open.
146 	 * If retransmitting (possibly after persist timer forced us
147 	 * to send into a small window), then must resend.
148 	 */
149 	if (len) {
150 		if (len == tp->t_maxseg)
151 			goto send;
152 		if ((idle || tp->t_flags & TF_NODELAY) &&
153 		    len + off >= so->so_snd.sb_cc)
154 			goto send;
155 		if (tp->t_force)
156 			goto send;
157 		if (len >= tp->max_sndwnd / 2)
158 			goto send;
159 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
160 			goto send;
161 	}
162 
163 	/*
164 	 * Compare available window to amount of window
165 	 * known to peer (as advertised window less
166 	 * next expected input).  If the difference is at least two
167 	 * max size segments, or at least 50% of the maximum possible
168 	 * window, then want to send a window update to peer.
169 	 */
170 	if (win > 0) {
171 		/*
172 		 * "adv" is the amount we can increase the window,
173 		 * taking into account that we are limited by
174 		 * TCP_MAXWIN << tp->rcv_scale.
175 		 */
176 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
177 			(tp->rcv_adv - tp->rcv_nxt);
178 
179 		if (adv >= (long) (2 * tp->t_maxseg))
180 			goto send;
181 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
182 			goto send;
183 	}
184 
185 	/*
186 	 * Send if we owe peer an ACK.
187 	 */
188 	if (tp->t_flags & TF_ACKNOW)
189 		goto send;
190 	if (flags & (TH_SYN|TH_RST))
191 		goto send;
192 	if (SEQ_GT(tp->snd_up, tp->snd_una))
193 		goto send;
194 	/*
195 	 * If our state indicates that FIN should be sent
196 	 * and we have not yet done so, or we're retransmitting the FIN,
197 	 * then we need to send.
198 	 */
199 	if (flags & TH_FIN &&
200 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
201 		goto send;
202 
203 	/*
204 	 * TCP window updates are not reliable, rather a polling protocol
205 	 * using ``persist'' packets is used to insure receipt of window
206 	 * updates.  The three ``states'' for the output side are:
207 	 *	idle			not doing retransmits or persists
208 	 *	persisting		to move a small or zero window
209 	 *	(re)transmitting	and thereby not persisting
210 	 *
211 	 * tp->t_timer[TCPT_PERSIST]
212 	 *	is set when we are in persist state.
213 	 * tp->t_force
214 	 *	is set when we are called to send a persist packet.
215 	 * tp->t_timer[TCPT_REXMT]
216 	 *	is set when we are retransmitting
217 	 * The output side is idle when both timers are zero.
218 	 *
219 	 * If send window is too small, there is data to transmit, and no
220 	 * retransmit or persist is pending, then go to persist state.
221 	 * If nothing happens soon, send when timer expires:
222 	 * if window is nonzero, transmit what we can,
223 	 * otherwise force out a byte.
224 	 */
225 	if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
226 	    tp->t_timer[TCPT_PERSIST] == 0) {
227 		tp->t_rxtshift = 0;
228 		tcp_setpersist(tp);
229 	}
230 
231 	/*
232 	 * No reason to send a segment, just return.
233 	 */
234 	return (0);
235 
236 send:
237 	/*
238 	 * Before ESTABLISHED, force sending of initial options
239 	 * unless TCP set not to do any options.
240 	 * NOTE: we assume that the IP/TCP header plus TCP options
241 	 * always fit in a single mbuf, leaving room for a maximum
242 	 * link header, i.e.
243 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MHLEN
244 	 */
245 	optlen = 0;
246 	hdrlen = sizeof (struct tcpiphdr);
247 	if (flags & TH_SYN) {
248 		tp->snd_nxt = tp->iss;
249 		if ((tp->t_flags & TF_NOOPT) == 0) {
250 			u_short mss;
251 
252 			opt[0] = TCPOPT_MAXSEG;
253 			opt[1] = 4;
254 			mss = htons((u_short) tcp_mss(tp, 0));
255 			bcopy((caddr_t)&mss, (caddr_t)(opt + 2), sizeof(mss));
256 			optlen = 4;
257 
258 			if ((tp->t_flags & TF_REQ_SCALE) &&
259 			    ((flags & TH_ACK) == 0 ||
260 			    (tp->t_flags & TF_RCVD_SCALE))) {
261 				*((u_long *) (opt + optlen)) = htonl(
262 					TCPOPT_NOP << 24 |
263 					TCPOPT_WINDOW << 16 |
264 					TCPOLEN_WINDOW << 8 |
265 					tp->request_r_scale);
266 				optlen += 4;
267 			}
268 		}
269  	}
270 
271  	/*
272 	 * Send a timestamp and echo-reply if this is a SYN and our side
273 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
274 	 * and our peer have sent timestamps in our SYN's.
275  	 */
276  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
277  	     (flags & TH_RST) == 0 &&
278  	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
279 	     (tp->t_flags & TF_RCVD_TSTMP))) {
280 		u_long *lp = (u_long *)(opt + optlen);
281 
282  		/* Form timestamp option as shown in appendix A of RFC 1323. */
283  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
284  		*lp++ = htonl(tcp_now);
285  		*lp   = htonl(tp->ts_recent);
286  		optlen += TCPOLEN_TSTAMP_APPA;
287  	}
288 
289  	hdrlen += optlen;
290 
291 	/*
292 	 * Adjust data length if insertion of options will
293 	 * bump the packet length beyond the t_maxseg length.
294 	 */
295 	if (len > tp->t_maxseg - optlen) {
296 		len = tp->t_maxseg - optlen;
297 		sendalot = 1;
298 		flags &= ~TH_FIN;
299 	 }
300 
301 
302 #ifdef DIAGNOSTIC
303  	if (max_linkhdr + hdrlen > MHLEN)
304 		panic("tcphdr too big");
305 #endif
306 
307 	/*
308 	 * Grab a header mbuf, attaching a copy of data to
309 	 * be transmitted, and initialize the header from
310 	 * the template for sends on this connection.
311 	 */
312 	if (len) {
313 		if (tp->t_force && len == 1)
314 			tcpstat.tcps_sndprobe++;
315 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
316 			tcpstat.tcps_sndrexmitpack++;
317 			tcpstat.tcps_sndrexmitbyte += len;
318 		} else {
319 			tcpstat.tcps_sndpack++;
320 			tcpstat.tcps_sndbyte += len;
321 		}
322 #ifdef notyet
323 		if ((m = m_copypack(so->so_snd.sb_mb, off,
324 		    (int)len, max_linkhdr + hdrlen)) == 0) {
325 			error = ENOBUFS;
326 			goto out;
327 		}
328 		/*
329 		 * m_copypack left space for our hdr; use it.
330 		 */
331 		m->m_len += hdrlen;
332 		m->m_data -= hdrlen;
333 #else
334 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
335 		if (m == NULL) {
336 			error = ENOBUFS;
337 			goto out;
338 		}
339 		m->m_data += max_linkhdr;
340 		m->m_len = hdrlen;
341 		if (len <= MHLEN - hdrlen - max_linkhdr) {
342 			m_copydata(so->so_snd.sb_mb, off, (int) len,
343 			    mtod(m, caddr_t) + hdrlen);
344 			m->m_len += len;
345 		} else {
346 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
347 			if (m->m_next == 0) {
348 				(void) m_free(m);
349 				error = ENOBUFS;
350 				goto out;
351 			}
352 		}
353 #endif
354 		/*
355 		 * If we're sending everything we've got, set PUSH.
356 		 * (This will keep happy those implementations which only
357 		 * give data to the user when a buffer fills or
358 		 * a PUSH comes in.)
359 		 */
360 		if (off + len == so->so_snd.sb_cc)
361 			flags |= TH_PUSH;
362 	} else {
363 		if (tp->t_flags & TF_ACKNOW)
364 			tcpstat.tcps_sndacks++;
365 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
366 			tcpstat.tcps_sndctrl++;
367 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
368 			tcpstat.tcps_sndurg++;
369 		else
370 			tcpstat.tcps_sndwinup++;
371 
372 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
373 		if (m == NULL) {
374 			error = ENOBUFS;
375 			goto out;
376 		}
377 		m->m_data += max_linkhdr;
378 		m->m_len = hdrlen;
379 	}
380 	m->m_pkthdr.rcvif = (struct ifnet *)0;
381 	ti = mtod(m, struct tcpiphdr *);
382 	if (tp->t_template == 0)
383 		panic("tcp_output");
384 	bcopy((caddr_t)tp->t_template, (caddr_t)ti, sizeof (struct tcpiphdr));
385 
386 	/*
387 	 * Fill in fields, remembering maximum advertised
388 	 * window for use in delaying messages about window sizes.
389 	 * If resending a FIN, be sure not to use a new sequence number.
390 	 */
391 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
392 	    tp->snd_nxt == tp->snd_max)
393 		tp->snd_nxt--;
394 	/*
395 	 * If we are doing retransmissions, then snd_nxt will
396 	 * not reflect the first unsent octet.  For ACK only
397 	 * packets, we do not want the sequence number of the
398 	 * retransmitted packet, we want the sequence number
399 	 * of the next unsent octet.  So, if there is no data
400 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
401 	 * when filling in ti_seq.  But if we are in persist
402 	 * state, snd_max might reflect one byte beyond the
403 	 * right edge of the window, so use snd_nxt in that
404 	 * case, since we know we aren't doing a retransmission.
405 	 * (retransmit and persist are mutually exclusive...)
406 	 */
407 	if (len || (flags & (TH_SYN|TH_FIN)) || tp->t_timer[TCPT_PERSIST])
408 		ti->ti_seq = htonl(tp->snd_nxt);
409 	else
410 		ti->ti_seq = htonl(tp->snd_max);
411 	ti->ti_ack = htonl(tp->rcv_nxt);
412 	if (optlen) {
413 		bcopy((caddr_t)opt, (caddr_t)(ti + 1), optlen);
414 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
415 	}
416 	ti->ti_flags = flags;
417 	/*
418 	 * Calculate receive window.  Don't shrink window,
419 	 * but avoid silly window syndrome.
420 	 */
421 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
422 		win = 0;
423 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
424 		win = (long)TCP_MAXWIN << tp->rcv_scale;
425 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
426 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
427 	ti->ti_win = htons((u_short) (win>>tp->rcv_scale));
428 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
429 		ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
430 		ti->ti_flags |= TH_URG;
431 	} else
432 		/*
433 		 * If no urgent pointer to send, then we pull
434 		 * the urgent pointer to the left edge of the send window
435 		 * so that it doesn't drift into the send window on sequence
436 		 * number wraparound.
437 		 */
438 		tp->snd_up = tp->snd_una;		/* drag it along */
439 
440 	/*
441 	 * Put TCP length in extended header, and then
442 	 * checksum extended header and data.
443 	 */
444 	if (len + optlen)
445 		ti->ti_len = htons((u_short)(sizeof (struct tcphdr) +
446 		    optlen + len));
447 	ti->ti_sum = in_cksum(m, (int)(hdrlen + len));
448 
449 	/*
450 	 * In transmit state, time the transmission and arrange for
451 	 * the retransmit.  In persist state, just set snd_max.
452 	 */
453 	if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) {
454 		tcp_seq startseq = tp->snd_nxt;
455 
456 		/*
457 		 * Advance snd_nxt over sequence space of this segment.
458 		 */
459 		if (flags & (TH_SYN|TH_FIN)) {
460 			if (flags & TH_SYN)
461 				tp->snd_nxt++;
462 			if (flags & TH_FIN) {
463 				tp->snd_nxt++;
464 				tp->t_flags |= TF_SENTFIN;
465 			}
466 		}
467 		tp->snd_nxt += len;
468 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
469 			tp->snd_max = tp->snd_nxt;
470 			/*
471 			 * Time this transmission if not a retransmission and
472 			 * not currently timing anything.
473 			 */
474 			if (tp->t_rtt == 0) {
475 				tp->t_rtt = 1;
476 				tp->t_rtseq = startseq;
477 				tcpstat.tcps_segstimed++;
478 			}
479 		}
480 
481 		/*
482 		 * Set retransmit timer if not currently set,
483 		 * and not doing an ack or a keep-alive probe.
484 		 * Initial value for retransmit timer is smoothed
485 		 * round-trip time + 2 * round-trip time variance.
486 		 * Initialize shift counter which is used for backoff
487 		 * of retransmit time.
488 		 */
489 		if (tp->t_timer[TCPT_REXMT] == 0 &&
490 		    tp->snd_nxt != tp->snd_una) {
491 			tp->t_timer[TCPT_REXMT] = tp->t_rxtcur;
492 			if (tp->t_timer[TCPT_PERSIST]) {
493 				tp->t_timer[TCPT_PERSIST] = 0;
494 				tp->t_rxtshift = 0;
495 			}
496 		}
497 	} else
498 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
499 			tp->snd_max = tp->snd_nxt + len;
500 
501 	/*
502 	 * Trace.
503 	 */
504 	if (so->so_options & SO_DEBUG)
505 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
506 
507 	/*
508 	 * Fill in IP length and desired time to live and
509 	 * send to IP level.  There should be a better way
510 	 * to handle ttl and tos; we could keep them in
511 	 * the template, but need a way to checksum without them.
512 	 */
513 	m->m_pkthdr.len = hdrlen + len;
514 #ifdef TUBA
515 	if (tp->t_tuba_pcb)
516 		error = tuba_output(m, tp);
517 	else
518 #endif
519     {
520 	((struct ip *)ti)->ip_len = m->m_pkthdr.len;
521 	((struct ip *)ti)->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;	/* XXX */
522 	((struct ip *)ti)->ip_tos = tp->t_inpcb->inp_ip.ip_tos;	/* XXX */
523 #if BSD >= 43
524 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
525 	    so->so_options & SO_DONTROUTE, 0);
526 #else
527 	error = ip_output(m, (struct mbuf *)0, &tp->t_inpcb->inp_route,
528 	    so->so_options & SO_DONTROUTE);
529 #endif
530     }
531 	if (error) {
532 out:
533 		if (error == ENOBUFS) {
534 			tcp_quench(tp->t_inpcb, 0);
535 			return (0);
536 		}
537 		if ((error == EHOSTUNREACH || error == ENETDOWN)
538 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
539 			tp->t_softerror = error;
540 			return (0);
541 		}
542 		return (error);
543 	}
544 	tcpstat.tcps_sndtotal++;
545 
546 	/*
547 	 * Data sent (as far as we can tell).
548 	 * If this advertises a larger window than any other segment,
549 	 * then remember the size of the advertised window.
550 	 * Any pending ACK has now been sent.
551 	 */
552 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
553 		tp->rcv_adv = tp->rcv_nxt + win;
554 	tp->last_ack_sent = tp->rcv_nxt;
555 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
556 	if (sendalot)
557 		goto again;
558 	return (0);
559 }
560 
561 void
tcp_setpersist(tp)562 tcp_setpersist(tp)
563 	register struct tcpcb *tp;
564 {
565 	register t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
566 
567 	if (tp->t_timer[TCPT_REXMT])
568 		panic("tcp_output REXMT");
569 	/*
570 	 * Start/restart persistance timer.
571 	 */
572 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
573 	    t * tcp_backoff[tp->t_rxtshift],
574 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
575 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
576 		tp->t_rxtshift++;
577 }
578