xref: /netbsd-src/sys/netinet/tcp_output.c (revision 0b9f50897e9a9c6709320fafb4c3787fddcc0a45)
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 1.3 1993/05/22 11:42:42 cgd Exp $
35  */
36 
37 #include "param.h"
38 #include "systm.h"
39 #include "malloc.h"
40 #include "mbuf.h"
41 #include "protosw.h"
42 #include "socket.h"
43 #include "socketvar.h"
44 #include "errno.h"
45 
46 #include "../net/route.h"
47 
48 #include "in.h"
49 #include "in_systm.h"
50 #include "ip.h"
51 #include "in_pcb.h"
52 #include "ip_var.h"
53 #include "tcp.h"
54 #define	TCPOUTFLAGS
55 #include "tcp_fsm.h"
56 #include "tcp_seq.h"
57 #include "tcp_timer.h"
58 #include "tcp_var.h"
59 #include "tcpip.h"
60 #include "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 tcp_output(tp)
75 	register struct tcpcb *tp;
76 {
77 	register struct socket *so = tp->t_inpcb->inp_socket;
78 	register long len, win;
79 	int off, flags, error;
80 	register struct mbuf *m;
81 	register struct tcpiphdr *ti;
82 	u_char *opt;
83 	unsigned optlen, hdrlen;
84 	int idle, sendalot;
85 
86 	/*
87 	 * Determine length of data that should be transmitted,
88 	 * and flags that will be used.
89 	 * If there is some data or critical controls (SYN, RST)
90 	 * to send, then transmit; otherwise, investigate further.
91 	 */
92 	idle = (tp->snd_max == tp->snd_una);
93 	if (idle && tp->t_idle >= tp->t_rxtcur)
94 		/*
95 		 * We have been idle for "a while" and no acks are
96 		 * expected to clock out any data we send --
97 		 * slow start to get ack "clock" running again.
98 		 */
99 		tp->snd_cwnd = tp->t_maxseg;
100 again:
101 	sendalot = 0;
102 	off = tp->snd_nxt - tp->snd_una;
103 	win = min(tp->snd_wnd, tp->snd_cwnd);
104 
105 	/*
106 	 * If in persist timeout with window of 0, send 1 byte.
107 	 * Otherwise, if window is small but nonzero
108 	 * and timer expired, we will send what we can
109 	 * and go to transmit state.
110 	 */
111 	if (tp->t_force) {
112 		if (win == 0)
113 			win = 1;
114 		else {
115 			tp->t_timer[TCPT_PERSIST] = 0;
116 			tp->t_rxtshift = 0;
117 		}
118 	}
119 
120 	flags = tcp_outflags[tp->t_state];
121 	len = min(so->so_snd.sb_cc, win) - off;
122 
123 	if (len < 0) {
124 		/*
125 		 * If FIN has been sent but not acked,
126 		 * but we haven't been called to retransmit,
127 		 * len will be -1.  Otherwise, window shrank
128 		 * after we sent into it.  If window shrank to 0,
129 		 * cancel pending retransmit and pull snd_nxt
130 		 * back to (closed) window.  We will enter persist
131 		 * state below.  If the window didn't close completely,
132 		 * just wait for an ACK.
133 		 */
134 		len = 0;
135 		if (win == 0) {
136 			tp->t_timer[TCPT_REXMT] = 0;
137 			tp->snd_nxt = tp->snd_una;
138 		}
139 	}
140 	if (len > tp->t_maxseg) {
141 		len = tp->t_maxseg;
142 		sendalot = 1;
143 	}
144 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
145 		flags &= ~TH_FIN;
146 
147 	win = sbspace(&so->so_rcv);
148 
149 	/*
150 	 * Sender silly window avoidance.  If connection is idle
151 	 * and can send all data, a maximum segment,
152 	 * at least a maximum default-size segment do it,
153 	 * or are forced, do it; otherwise don't bother.
154 	 * If peer's buffer is tiny, then send
155 	 * when window is at least half open.
156 	 * If retransmitting (possibly after persist timer forced us
157 	 * to send into a small window), then must resend.
158 	 */
159 	if (len) {
160 		if (len == tp->t_maxseg)
161 			goto send;
162 		if ((idle || tp->t_flags & TF_NODELAY) &&
163 		    len + off >= so->so_snd.sb_cc)
164 			goto send;
165 		if (tp->t_force)
166 			goto send;
167 		if (len >= tp->max_sndwnd / 2)
168 			goto send;
169 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
170 			goto send;
171 	}
172 
173 	/*
174 	 * Compare available window to amount of window
175 	 * known to peer (as advertised window less
176 	 * next expected input).  If the difference is at least two
177 	 * max size segments, or at least 50% of the maximum possible
178 	 * window, then want to send a window update to peer.
179 	 */
180 	if (win > 0) {
181 		long adv = win - (tp->rcv_adv - tp->rcv_nxt);
182 
183 		if (adv >= (long) (2 * tp->t_maxseg))
184 			goto send;
185 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
186 			goto send;
187 	}
188 
189 	/*
190 	 * Send if we owe peer an ACK.
191 	 */
192 	if (tp->t_flags & TF_ACKNOW)
193 		goto send;
194 	if (flags & (TH_SYN|TH_RST))
195 		goto send;
196 	if (SEQ_GT(tp->snd_up, tp->snd_una))
197 		goto send;
198 	/*
199 	 * If our state indicates that FIN should be sent
200 	 * and we have not yet done so, or we're retransmitting the FIN,
201 	 * then we need to send.
202 	 */
203 	if (flags & TH_FIN &&
204 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
205 		goto send;
206 
207 	/*
208 	 * TCP window updates are not reliable, rather a polling protocol
209 	 * using ``persist'' packets is used to insure receipt of window
210 	 * updates.  The three ``states'' for the output side are:
211 	 *	idle			not doing retransmits or persists
212 	 *	persisting		to move a small or zero window
213 	 *	(re)transmitting	and thereby not persisting
214 	 *
215 	 * tp->t_timer[TCPT_PERSIST]
216 	 *	is set when we are in persist state.
217 	 * tp->t_force
218 	 *	is set when we are called to send a persist packet.
219 	 * tp->t_timer[TCPT_REXMT]
220 	 *	is set when we are retransmitting
221 	 * The output side is idle when both timers are zero.
222 	 *
223 	 * If send window is too small, there is data to transmit, and no
224 	 * retransmit or persist is pending, then go to persist state.
225 	 * If nothing happens soon, send when timer expires:
226 	 * if window is nonzero, transmit what we can,
227 	 * otherwise force out a byte.
228 	 */
229 	if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
230 	    tp->t_timer[TCPT_PERSIST] == 0) {
231 		tp->t_rxtshift = 0;
232 		tcp_setpersist(tp);
233 	}
234 
235 	/*
236 	 * No reason to send a segment, just return.
237 	 */
238 	return (0);
239 
240 send:
241 	/*
242 	 * Before ESTABLISHED, force sending of initial options
243 	 * unless TCP set not to do any options.
244 	 * NOTE: we assume that the IP/TCP header plus TCP options
245 	 * always fit in a single mbuf, leaving room for a maximum
246 	 * link header, i.e.
247 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MHLEN
248 	 */
249 	optlen = 0;
250 	hdrlen = sizeof (struct tcpiphdr);
251 	if (flags & TH_SYN && (tp->t_flags & TF_NOOPT) == 0) {
252 		opt = tcp_initopt;
253 		optlen = sizeof (tcp_initopt);
254 		hdrlen += sizeof (tcp_initopt);
255 		*(u_short *)(opt + 2) = htons((u_short) tcp_mss(tp, 0));
256 #ifdef DIAGNOSTIC
257 	 	if (max_linkhdr + hdrlen > MHLEN)
258 			panic("tcphdr too big");
259 #endif
260 	}
261 
262 	/*
263 	 * Grab a header mbuf, attaching a copy of data to
264 	 * be transmitted, and initialize the header from
265 	 * the template for sends on this connection.
266 	 */
267 	if (len) {
268 		if (tp->t_force && len == 1)
269 			tcpstat.tcps_sndprobe++;
270 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
271 			tcpstat.tcps_sndrexmitpack++;
272 			tcpstat.tcps_sndrexmitbyte += len;
273 		} else {
274 			tcpstat.tcps_sndpack++;
275 			tcpstat.tcps_sndbyte += len;
276 		}
277 #ifdef notyet
278 		if ((m = m_copypack(so->so_snd.sb_mb, off,
279 		    (int)len, max_linkhdr + hdrlen)) == 0) {
280 			error = ENOBUFS;
281 			goto out;
282 		}
283 		/*
284 		 * m_copypack left space for our hdr; use it.
285 		 */
286 		m->m_len += hdrlen;
287 		m->m_data -= hdrlen;
288 #else
289 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
290 		if (m == NULL) {
291 			error = ENOBUFS;
292 			goto out;
293 		}
294 		m->m_data += max_linkhdr;
295 		m->m_len = hdrlen;
296 		if (len <= MHLEN - hdrlen - max_linkhdr) {
297 			m_copydata(so->so_snd.sb_mb, off, (int) len,
298 			    mtod(m, caddr_t) + hdrlen);
299 			m->m_len += len;
300 		} else {
301 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
302 			if (m->m_next == 0)
303 				len = 0;
304 		}
305 #endif
306 		/*
307 		 * If we're sending everything we've got, set PUSH.
308 		 * (This will keep happy those implementations which only
309 		 * give data to the user when a buffer fills or
310 		 * a PUSH comes in.)
311 		 */
312 		if (off + len == so->so_snd.sb_cc)
313 			flags |= TH_PUSH;
314 	} else {
315 		if (tp->t_flags & TF_ACKNOW)
316 			tcpstat.tcps_sndacks++;
317 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
318 			tcpstat.tcps_sndctrl++;
319 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
320 			tcpstat.tcps_sndurg++;
321 		else
322 			tcpstat.tcps_sndwinup++;
323 
324 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
325 		if (m == NULL) {
326 			error = ENOBUFS;
327 			goto out;
328 		}
329 		m->m_data += max_linkhdr;
330 		m->m_len = hdrlen;
331 	}
332 	m->m_pkthdr.rcvif = (struct ifnet *)0;
333 	ti = mtod(m, struct tcpiphdr *);
334 	if (tp->t_template == 0)
335 		panic("tcp_output");
336 	bcopy((caddr_t)tp->t_template, (caddr_t)ti, sizeof (struct tcpiphdr));
337 
338 	/*
339 	 * Fill in fields, remembering maximum advertised
340 	 * window for use in delaying messages about window sizes.
341 	 * If resending a FIN, be sure not to use a new sequence number.
342 	 */
343 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
344 	    tp->snd_nxt == tp->snd_max)
345 		tp->snd_nxt--;
346 	ti->ti_seq = htonl(tp->snd_nxt);
347 	ti->ti_ack = htonl(tp->rcv_nxt);
348 	if (optlen) {
349 		bcopy((caddr_t)opt, (caddr_t)(ti + 1), optlen);
350 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
351 	}
352 	ti->ti_flags = flags;
353 	/*
354 	 * Calculate receive window.  Don't shrink window,
355 	 * but avoid silly window syndrome.
356 	 */
357 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
358 		win = 0;
359 	if (win > TCP_MAXWIN)
360 		win = TCP_MAXWIN;
361 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
362 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
363 	ti->ti_win = htons((u_short)win);
364 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
365 		ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
366 		ti->ti_flags |= TH_URG;
367 	} else
368 		/*
369 		 * If no urgent pointer to send, then we pull
370 		 * the urgent pointer to the left edge of the send window
371 		 * so that it doesn't drift into the send window on sequence
372 		 * number wraparound.
373 		 */
374 		tp->snd_up = tp->snd_una;		/* drag it along */
375 
376 	/*
377 	 * Put TCP length in extended header, and then
378 	 * checksum extended header and data.
379 	 */
380 	if (len + optlen)
381 		ti->ti_len = htons((u_short)(sizeof (struct tcphdr) +
382 		    optlen + len));
383 	ti->ti_sum = in_cksum(m, (int)(hdrlen + len));
384 
385 	/*
386 	 * In transmit state, time the transmission and arrange for
387 	 * the retransmit.  In persist state, just set snd_max.
388 	 */
389 	if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) {
390 		tcp_seq startseq = tp->snd_nxt;
391 
392 		/*
393 		 * Advance snd_nxt over sequence space of this segment.
394 		 */
395 		if (flags & (TH_SYN|TH_FIN)) {
396 			if (flags & TH_SYN)
397 				tp->snd_nxt++;
398 			if (flags & TH_FIN) {
399 				tp->snd_nxt++;
400 				tp->t_flags |= TF_SENTFIN;
401 			}
402 		}
403 		tp->snd_nxt += len;
404 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
405 			tp->snd_max = tp->snd_nxt;
406 			/*
407 			 * Time this transmission if not a retransmission and
408 			 * not currently timing anything.
409 			 */
410 			if (tp->t_rtt == 0) {
411 				tp->t_rtt = 1;
412 				tp->t_rtseq = startseq;
413 				tcpstat.tcps_segstimed++;
414 			}
415 		}
416 
417 		/*
418 		 * Set retransmit timer if not currently set,
419 		 * and not doing an ack or a keep-alive probe.
420 		 * Initial value for retransmit timer is smoothed
421 		 * round-trip time + 2 * round-trip time variance.
422 		 * Initialize shift counter which is used for backoff
423 		 * of retransmit time.
424 		 */
425 		if (tp->t_timer[TCPT_REXMT] == 0 &&
426 		    tp->snd_nxt != tp->snd_una) {
427 			tp->t_timer[TCPT_REXMT] = tp->t_rxtcur;
428 			if (tp->t_timer[TCPT_PERSIST]) {
429 				tp->t_timer[TCPT_PERSIST] = 0;
430 				tp->t_rxtshift = 0;
431 			}
432 		}
433 	} else
434 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
435 			tp->snd_max = tp->snd_nxt + len;
436 
437 	/*
438 	 * Trace.
439 	 */
440 	if (so->so_options & SO_DEBUG)
441 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
442 
443 	/*
444 	 * Fill in IP length and desired time to live and
445 	 * send to IP level.  There should be a better way
446 	 * to handle ttl and tos; we could keep them in
447 	 * the template, but need a way to checksum without them.
448 	 */
449 	m->m_pkthdr.len = hdrlen + len;
450 	((struct ip *)ti)->ip_len = m->m_pkthdr.len;
451 	((struct ip *)ti)->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;	/* XXX */
452 	((struct ip *)ti)->ip_tos = tp->t_inpcb->inp_ip.ip_tos;	/* XXX */
453 #if BSD >= 43
454 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
455 	    so->so_options & SO_DONTROUTE);
456 #else
457 	error = ip_output(m, (struct mbuf *)0, &tp->t_inpcb->inp_route,
458 	    so->so_options & SO_DONTROUTE);
459 #endif
460 	if (error) {
461 out:
462 		if (error == ENOBUFS) {
463 			tcp_quench(tp->t_inpcb);
464 			return (0);
465 		}
466 		if ((error == EHOSTUNREACH || error == ENETDOWN)
467 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
468 			tp->t_softerror = error;
469 			return (0);
470 		}
471 		return (error);
472 	}
473 	tcpstat.tcps_sndtotal++;
474 
475 	/*
476 	 * Data sent (as far as we can tell).
477 	 * If this advertises a larger window than any other segment,
478 	 * then remember the size of the advertised window.
479 	 * Any pending ACK has now been sent.
480 	 */
481 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
482 		tp->rcv_adv = tp->rcv_nxt + win;
483 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
484 	if (sendalot)
485 		goto again;
486 	return (0);
487 }
488 
489 tcp_setpersist(tp)
490 	register struct tcpcb *tp;
491 {
492 	register t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
493 
494 	if (tp->t_timer[TCPT_REXMT])
495 		panic("tcp_output REXMT");
496 	/*
497 	 * Start/restart persistance timer.
498 	 */
499 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
500 	    t * tcp_backoff[tp->t_rxtshift],
501 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
502 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
503 		tp->t_rxtshift++;
504 }
505