xref: /netbsd-src/sys/netinet/tcp_output.c (revision 05a9db8e4f5a54955cacec0e01c45d54b958765c)
1 /*
2  * Copyright (c) 1982, 1986, 1988, 1990 Regents of the University of California.
3  * 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  * 3. All advertising materials mentioning features or use of this software
14  *    must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Berkeley and its contributors.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	from: @(#)tcp_output.c	7.22 (Berkeley) 8/31/90
34  *	$Id: tcp_output.c,v 1.8 1994/04/12 18:09:47 mycroft Exp $
35  */
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/malloc.h>
40 #include <sys/mbuf.h>
41 #include <sys/protosw.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
44 #include <sys/errno.h>
45 
46 #include <net/route.h>
47 
48 #include <netinet/in.h>
49 #include <netinet/in_systm.h>
50 #include <netinet/ip.h>
51 #include <netinet/in_pcb.h>
52 #include <netinet/ip_var.h>
53 #include <netinet/tcp.h>
54 #define	TCPOUTFLAGS
55 #include <netinet/tcp_fsm.h>
56 #include <netinet/tcp_seq.h>
57 #include <netinet/tcp_timer.h>
58 #include <netinet/tcp_var.h>
59 #include <netinet/tcpip.h>
60 #include <netinet/tcp_debug.h>
61 
62 #ifdef notyet
63 extern struct mbuf *m_copypack();
64 #endif
65 
66 /*
67  * Initial options.
68  */
69 u_char	tcp_initopt[4] = { TCPOPT_MAXSEG, 4, 0x0, 0x0, };
70 
71 /*
72  * Tcp output routine: figure out what should be sent and send it.
73  */
74 int
75 tcp_output(tp)
76 	register struct tcpcb *tp;
77 {
78 	register struct socket *so = tp->t_inpcb->inp_socket;
79 	register long len, win;
80 	int off, flags, error;
81 	register struct mbuf *m;
82 	register struct tcpiphdr *ti;
83 	u_char *opt;
84 	unsigned optlen, hdrlen;
85 	int idle, sendalot;
86 
87 	/*
88 	 * Determine length of data that should be transmitted,
89 	 * and flags that will be used.
90 	 * If there is some data or critical controls (SYN, RST)
91 	 * to send, then transmit; otherwise, investigate further.
92 	 */
93 	idle = (tp->snd_max == tp->snd_una);
94 	if (idle && tp->t_idle >= tp->t_rxtcur)
95 		/*
96 		 * We have been idle for "a while" and no acks are
97 		 * expected to clock out any data we send --
98 		 * slow start to get ack "clock" running again.
99 		 */
100 		tp->snd_cwnd = tp->t_maxseg;
101 again:
102 	sendalot = 0;
103 	off = tp->snd_nxt - tp->snd_una;
104 	win = min(tp->snd_wnd, tp->snd_cwnd);
105 
106 	/*
107 	 * If in persist timeout with window of 0, send 1 byte.
108 	 * Otherwise, if window is small but nonzero
109 	 * and timer expired, we will send what we can
110 	 * and go to transmit state.
111 	 */
112 	if (tp->t_force) {
113 		if (win == 0)
114 			win = 1;
115 		else {
116 			tp->t_timer[TCPT_PERSIST] = 0;
117 			tp->t_rxtshift = 0;
118 		}
119 	}
120 
121 	flags = tcp_outflags[tp->t_state];
122 	len = min(so->so_snd.sb_cc, win) - off;
123 
124 	if (len < 0) {
125 		/*
126 		 * If FIN has been sent but not acked,
127 		 * but we haven't been called to retransmit,
128 		 * len will be -1.  Otherwise, window shrank
129 		 * after we sent into it.  If window shrank to 0,
130 		 * cancel pending retransmit and pull snd_nxt
131 		 * back to (closed) window.  We will enter persist
132 		 * state below.  If the window didn't close completely,
133 		 * just wait for an ACK.
134 		 */
135 		len = 0;
136 		if (win == 0) {
137 			tp->t_timer[TCPT_REXMT] = 0;
138 			tp->snd_nxt = tp->snd_una;
139 		}
140 	}
141 	if (len > tp->t_maxseg) {
142 		len = tp->t_maxseg;
143 		sendalot = 1;
144 	}
145 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
146 		flags &= ~TH_FIN;
147 
148 	win = sbspace(&so->so_rcv);
149 
150 	/*
151 	 * Sender silly window avoidance.  If connection is idle
152 	 * and can send all data, a maximum segment,
153 	 * at least a maximum default-size segment do it,
154 	 * or are forced, do it; otherwise don't bother.
155 	 * If peer's buffer is tiny, then send
156 	 * when window is at least half open.
157 	 * If retransmitting (possibly after persist timer forced us
158 	 * to send into a small window), then must resend.
159 	 */
160 	if (len) {
161 		if (len == tp->t_maxseg)
162 			goto send;
163 		if ((idle || tp->t_flags & TF_NODELAY) &&
164 		    len + off >= so->so_snd.sb_cc)
165 			goto send;
166 		if (tp->t_force)
167 			goto send;
168 		if (len >= tp->max_sndwnd / 2)
169 			goto send;
170 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
171 			goto send;
172 	}
173 
174 	/*
175 	 * Compare available window to amount of window
176 	 * known to peer (as advertised window less
177 	 * next expected input).  If the difference is at least two
178 	 * max size segments, or at least 50% of the maximum possible
179 	 * window, then want to send a window update to peer.
180 	 */
181 	if (win > 0) {
182 		long adv = win - (tp->rcv_adv - tp->rcv_nxt);
183 
184 		if (adv >= (long) (2 * tp->t_maxseg))
185 			goto send;
186 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
187 			goto send;
188 	}
189 
190 	/*
191 	 * Send if we owe peer an ACK.
192 	 */
193 	if (tp->t_flags & TF_ACKNOW)
194 		goto send;
195 	if (flags & (TH_SYN|TH_RST))
196 		goto send;
197 	if (SEQ_GT(tp->snd_up, tp->snd_una))
198 		goto send;
199 	/*
200 	 * If our state indicates that FIN should be sent
201 	 * and we have not yet done so, or we're retransmitting the FIN,
202 	 * then we need to send.
203 	 */
204 	if (flags & TH_FIN &&
205 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
206 		goto send;
207 
208 	/*
209 	 * TCP window updates are not reliable, rather a polling protocol
210 	 * using ``persist'' packets is used to insure receipt of window
211 	 * updates.  The three ``states'' for the output side are:
212 	 *	idle			not doing retransmits or persists
213 	 *	persisting		to move a small or zero window
214 	 *	(re)transmitting	and thereby not persisting
215 	 *
216 	 * tp->t_timer[TCPT_PERSIST]
217 	 *	is set when we are in persist state.
218 	 * tp->t_force
219 	 *	is set when we are called to send a persist packet.
220 	 * tp->t_timer[TCPT_REXMT]
221 	 *	is set when we are retransmitting
222 	 * The output side is idle when both timers are zero.
223 	 *
224 	 * If send window is too small, there is data to transmit, and no
225 	 * retransmit or persist is pending, then go to persist state.
226 	 * If nothing happens soon, send when timer expires:
227 	 * if window is nonzero, transmit what we can,
228 	 * otherwise force out a byte.
229 	 */
230 	if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
231 	    tp->t_timer[TCPT_PERSIST] == 0) {
232 		tp->t_rxtshift = 0;
233 		tcp_setpersist(tp);
234 	}
235 
236 	/*
237 	 * No reason to send a segment, just return.
238 	 */
239 	return (0);
240 
241 send:
242 	/*
243 	 * Before ESTABLISHED, force sending of initial options
244 	 * unless TCP set not to do any options.
245 	 * NOTE: we assume that the IP/TCP header plus TCP options
246 	 * always fit in a single mbuf, leaving room for a maximum
247 	 * link header, i.e.
248 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MHLEN
249 	 */
250 	optlen = 0;
251 	hdrlen = sizeof (struct tcpiphdr);
252 	if (flags & TH_SYN && (tp->t_flags & TF_NOOPT) == 0) {
253 		opt = tcp_initopt;
254 		optlen = sizeof (tcp_initopt);
255 		hdrlen += sizeof (tcp_initopt);
256 		*(u_short *)(opt + 2) = htons((u_short) tcp_mss(tp, 0));
257 #ifdef DIAGNOSTIC
258 	 	if (max_linkhdr + hdrlen > MHLEN)
259 			panic("tcphdr too big");
260 #endif
261 	}
262 
263 	/*
264 	 * Grab a header mbuf, attaching a copy of data to
265 	 * be transmitted, and initialize the header from
266 	 * the template for sends on this connection.
267 	 */
268 	if (len) {
269 		if (tp->t_force && len == 1)
270 			tcpstat.tcps_sndprobe++;
271 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
272 			tcpstat.tcps_sndrexmitpack++;
273 			tcpstat.tcps_sndrexmitbyte += len;
274 		} else {
275 			tcpstat.tcps_sndpack++;
276 			tcpstat.tcps_sndbyte += len;
277 		}
278 #ifdef notyet
279 		if ((m = m_copypack(so->so_snd.sb_mb, off,
280 		    (int)len, max_linkhdr + hdrlen)) == 0) {
281 			error = ENOBUFS;
282 			goto out;
283 		}
284 		/*
285 		 * m_copypack left space for our hdr; use it.
286 		 */
287 		m->m_len += hdrlen;
288 		m->m_data -= hdrlen;
289 #else
290 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
291 		if (m == NULL) {
292 			error = ENOBUFS;
293 			goto out;
294 		}
295 		m->m_data += max_linkhdr;
296 		m->m_len = hdrlen;
297 		if (len <= MHLEN - hdrlen - max_linkhdr) {
298 			m_copydata(so->so_snd.sb_mb, off, (int) len,
299 			    mtod(m, caddr_t) + hdrlen);
300 			m->m_len += len;
301 		} else {
302 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
303 			if (m->m_next == 0)
304 				len = 0;
305 		}
306 #endif
307 		/*
308 		 * If we're sending everything we've got, set PUSH.
309 		 * (This will keep happy those implementations which only
310 		 * give data to the user when a buffer fills or
311 		 * a PUSH comes in.)
312 		 */
313 		if (off + len == so->so_snd.sb_cc)
314 			flags |= TH_PUSH;
315 	} else {
316 		if (tp->t_flags & TF_ACKNOW)
317 			tcpstat.tcps_sndacks++;
318 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
319 			tcpstat.tcps_sndctrl++;
320 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
321 			tcpstat.tcps_sndurg++;
322 		else
323 			tcpstat.tcps_sndwinup++;
324 
325 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
326 		if (m == NULL) {
327 			error = ENOBUFS;
328 			goto out;
329 		}
330 		m->m_data += max_linkhdr;
331 		m->m_len = hdrlen;
332 	}
333 	m->m_pkthdr.rcvif = (struct ifnet *)0;
334 	ti = mtod(m, struct tcpiphdr *);
335 	if (tp->t_template == 0)
336 		panic("tcp_output");
337 	bcopy((caddr_t)tp->t_template, (caddr_t)ti, sizeof (struct tcpiphdr));
338 
339 	/*
340 	 * Fill in fields, remembering maximum advertised
341 	 * window for use in delaying messages about window sizes.
342 	 * If resending a FIN, be sure not to use a new sequence number.
343 	 */
344 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
345 	    tp->snd_nxt == tp->snd_max)
346 		tp->snd_nxt--;
347 	if (len || (flags & (TH_SYN|TH_FIN)) || tp->t_timer[TCPT_PERSIST])
348 		ti->ti_seq = htonl(tp->snd_nxt);
349 	else
350 		ti->ti_seq = htonl(tp->snd_max);
351 	ti->ti_ack = htonl(tp->rcv_nxt);
352 	if (optlen) {
353 		bcopy((caddr_t)opt, (caddr_t)(ti + 1), optlen);
354 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
355 	}
356 	ti->ti_flags = flags;
357 	/*
358 	 * Calculate receive window.  Don't shrink window,
359 	 * but avoid silly window syndrome.
360 	 */
361 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
362 		win = 0;
363 	if (win > TCP_MAXWIN)
364 		win = TCP_MAXWIN;
365 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
366 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
367 	ti->ti_win = htons((u_short)win);
368 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
369 		ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
370 		ti->ti_flags |= TH_URG;
371 	} else
372 		/*
373 		 * If no urgent pointer to send, then we pull
374 		 * the urgent pointer to the left edge of the send window
375 		 * so that it doesn't drift into the send window on sequence
376 		 * number wraparound.
377 		 */
378 		tp->snd_up = tp->snd_una;		/* drag it along */
379 
380 	/*
381 	 * Put TCP length in extended header, and then
382 	 * checksum extended header and data.
383 	 */
384 	if (len + optlen)
385 		ti->ti_len = htons((u_short)(sizeof (struct tcphdr) +
386 		    optlen + len));
387 	ti->ti_sum = in_cksum(m, (int)(hdrlen + len));
388 
389 	/*
390 	 * In transmit state, time the transmission and arrange for
391 	 * the retransmit.  In persist state, just set snd_max.
392 	 */
393 	if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) {
394 		tcp_seq startseq = tp->snd_nxt;
395 
396 		/*
397 		 * Advance snd_nxt over sequence space of this segment.
398 		 */
399 		if (flags & (TH_SYN|TH_FIN)) {
400 			if (flags & TH_SYN)
401 				tp->snd_nxt++;
402 			if (flags & TH_FIN) {
403 				tp->snd_nxt++;
404 				tp->t_flags |= TF_SENTFIN;
405 			}
406 		}
407 		tp->snd_nxt += len;
408 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
409 			tp->snd_max = tp->snd_nxt;
410 			/*
411 			 * Time this transmission if not a retransmission and
412 			 * not currently timing anything.
413 			 */
414 			if (tp->t_rtt == 0) {
415 				tp->t_rtt = 1;
416 				tp->t_rtseq = startseq;
417 				tcpstat.tcps_segstimed++;
418 			}
419 		}
420 
421 		/*
422 		 * Set retransmit timer if not currently set,
423 		 * and not doing an ack or a keep-alive probe.
424 		 * Initial value for retransmit timer is smoothed
425 		 * round-trip time + 2 * round-trip time variance.
426 		 * Initialize shift counter which is used for backoff
427 		 * of retransmit time.
428 		 */
429 		if (tp->t_timer[TCPT_REXMT] == 0 &&
430 		    tp->snd_nxt != tp->snd_una) {
431 			tp->t_timer[TCPT_REXMT] = tp->t_rxtcur;
432 			if (tp->t_timer[TCPT_PERSIST]) {
433 				tp->t_timer[TCPT_PERSIST] = 0;
434 				tp->t_rxtshift = 0;
435 			}
436 		}
437 	} else
438 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
439 			tp->snd_max = tp->snd_nxt + len;
440 
441 	/*
442 	 * Trace.
443 	 */
444 	if (so->so_options & SO_DEBUG)
445 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
446 
447 	/*
448 	 * Fill in IP length and desired time to live and
449 	 * send to IP level.  There should be a better way
450 	 * to handle ttl and tos; we could keep them in
451 	 * the template, but need a way to checksum without them.
452 	 */
453 	m->m_pkthdr.len = hdrlen + len;
454 	((struct ip *)ti)->ip_len = m->m_pkthdr.len;
455 	((struct ip *)ti)->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;	/* XXX */
456 	((struct ip *)ti)->ip_tos = tp->t_inpcb->inp_ip.ip_tos;	/* XXX */
457 #if BSD >= 43
458 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
459 	    so->so_options & SO_DONTROUTE, NULL);
460 #else
461 	error = ip_output(m, (struct mbuf *)0, &tp->t_inpcb->inp_route,
462 	    so->so_options & SO_DONTROUTE);
463 #endif
464 	if (error) {
465 out:
466 		if (error == ENOBUFS) {
467 			tcp_quench(tp->t_inpcb, 0);
468 			return (0);
469 		}
470 		if ((error == EHOSTUNREACH || error == ENETDOWN)
471 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
472 			tp->t_softerror = error;
473 			return (0);
474 		}
475 		return (error);
476 	}
477 	tcpstat.tcps_sndtotal++;
478 
479 	/*
480 	 * Data sent (as far as we can tell).
481 	 * If this advertises a larger window than any other segment,
482 	 * then remember the size of the advertised window.
483 	 * Any pending ACK has now been sent.
484 	 */
485 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
486 		tp->rcv_adv = tp->rcv_nxt + win;
487 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
488 	if (sendalot)
489 		goto again;
490 	return (0);
491 }
492 
493 void
494 tcp_setpersist(tp)
495 	register struct tcpcb *tp;
496 {
497 	register t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
498 
499 	if (tp->t_timer[TCPT_REXMT])
500 		panic("tcp_output REXMT");
501 	/*
502 	 * Start/restart persistance timer.
503 	 */
504 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
505 	    t * tcp_backoff[tp->t_rxtshift],
506 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
507 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
508 		tp->t_rxtshift++;
509 }
510