xref: /csrg-svn/sys/netinet/tcp_output.c (revision 57666)
1 /*
2  * Copyright (c) 1982, 1986, 1988, 1990 Regents of the University of California.
3  * All rights reserved.
4  *
5  * %sccs.include.redist.c%
6  *
7  *	@(#)tcp_output.c	7.26 (Berkeley) 01/22/93
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 tcp_output(tp)
46 	register struct tcpcb *tp;
47 {
48 	register struct socket *so = tp->t_inpcb->inp_socket;
49 	register long len, win;
50 	int off, flags, error;
51 	register struct mbuf *m;
52 	register struct tcpiphdr *ti;
53 	u_char opt[MAX_TCPOPTLEN];
54 	unsigned optlen, hdrlen;
55 	int idle, sendalot;
56 
57 	/*
58 	 * Determine length of data that should be transmitted,
59 	 * and flags that will be used.
60 	 * If there is some data or critical controls (SYN, RST)
61 	 * to send, then transmit; otherwise, investigate further.
62 	 */
63 	idle = (tp->snd_max == tp->snd_una);
64 	if (idle && tp->t_idle >= tp->t_rxtcur)
65 		/*
66 		 * We have been idle for "a while" and no acks are
67 		 * expected to clock out any data we send --
68 		 * slow start to get ack "clock" running again.
69 		 */
70 		tp->snd_cwnd = tp->t_maxseg;
71 again:
72 	sendalot = 0;
73 	off = tp->snd_nxt - tp->snd_una;
74 	win = min(tp->snd_wnd, tp->snd_cwnd);
75 
76 	/*
77 	 * If in persist timeout with window of 0, send 1 byte.
78 	 * Otherwise, if window is small but nonzero
79 	 * and timer expired, we will send what we can
80 	 * and go to transmit state.
81 	 */
82 	if (tp->t_force) {
83 		if (win == 0)
84 			win = 1;
85 		else {
86 			tp->t_timer[TCPT_PERSIST] = 0;
87 			tp->t_rxtshift = 0;
88 		}
89 	}
90 
91 	flags = tcp_outflags[tp->t_state];
92 	len = min(so->so_snd.sb_cc, win) - off;
93 
94 	if (len < 0) {
95 		/*
96 		 * If FIN has been sent but not acked,
97 		 * but we haven't been called to retransmit,
98 		 * len will be -1.  Otherwise, window shrank
99 		 * after we sent into it.  If window shrank to 0,
100 		 * cancel pending retransmit and pull snd_nxt
101 		 * back to (closed) window.  We will enter persist
102 		 * state below.  If the window didn't close completely,
103 		 * just wait for an ACK.
104 		 */
105 		len = 0;
106 		if (win == 0) {
107 			tp->t_timer[TCPT_REXMT] = 0;
108 			tp->snd_nxt = tp->snd_una;
109 		}
110 	}
111 	if (len > tp->t_maxseg) {
112 		len = tp->t_maxseg;
113 		sendalot = 1;
114 	}
115 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
116 		flags &= ~TH_FIN;
117 
118 	win = sbspace(&so->so_rcv);
119 
120 	/*
121 	 * Sender silly window avoidance.  If connection is idle
122 	 * and can send all data, a maximum segment,
123 	 * at least a maximum default-size segment do it,
124 	 * or are forced, do it; otherwise don't bother.
125 	 * If peer's buffer is tiny, then send
126 	 * when window is at least half open.
127 	 * If retransmitting (possibly after persist timer forced us
128 	 * to send into a small window), then must resend.
129 	 */
130 	if (len) {
131 		if (len == tp->t_maxseg)
132 			goto send;
133 		if ((idle || tp->t_flags & TF_NODELAY) &&
134 		    len + off >= so->so_snd.sb_cc)
135 			goto send;
136 		if (tp->t_force)
137 			goto send;
138 		if (len >= tp->max_sndwnd / 2)
139 			goto send;
140 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
141 			goto send;
142 	}
143 
144 	/*
145 	 * Compare available window to amount of window
146 	 * known to peer (as advertised window less
147 	 * next expected input).  If the difference is at least two
148 	 * max size segments, or at least 50% of the maximum possible
149 	 * window, then want to send a window update to peer.
150 	 */
151 	if (win > 0) {
152 		/*
153 		 * "adv" is the amount we can increase the window,
154 		 * taking into account that we are limited by
155 		 * TCP_MAXWIN << tp->rcv_scale.
156 		 */
157 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
158 			(tp->rcv_adv - tp->rcv_nxt);
159 
160 		if (adv >= (long) (2 * tp->t_maxseg))
161 			goto send;
162 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
163 			goto send;
164 	}
165 
166 	/*
167 	 * Send if we owe peer an ACK.
168 	 */
169 	if (tp->t_flags & TF_ACKNOW)
170 		goto send;
171 	if (flags & (TH_SYN|TH_RST))
172 		goto send;
173 	if (SEQ_GT(tp->snd_up, tp->snd_una))
174 		goto send;
175 	/*
176 	 * If our state indicates that FIN should be sent
177 	 * and we have not yet done so, or we're retransmitting the FIN,
178 	 * then we need to send.
179 	 */
180 	if (flags & TH_FIN &&
181 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
182 		goto send;
183 
184 	/*
185 	 * TCP window updates are not reliable, rather a polling protocol
186 	 * using ``persist'' packets is used to insure receipt of window
187 	 * updates.  The three ``states'' for the output side are:
188 	 *	idle			not doing retransmits or persists
189 	 *	persisting		to move a small or zero window
190 	 *	(re)transmitting	and thereby not persisting
191 	 *
192 	 * tp->t_timer[TCPT_PERSIST]
193 	 *	is set when we are in persist state.
194 	 * tp->t_force
195 	 *	is set when we are called to send a persist packet.
196 	 * tp->t_timer[TCPT_REXMT]
197 	 *	is set when we are retransmitting
198 	 * The output side is idle when both timers are zero.
199 	 *
200 	 * If send window is too small, there is data to transmit, and no
201 	 * retransmit or persist is pending, then go to persist state.
202 	 * If nothing happens soon, send when timer expires:
203 	 * if window is nonzero, transmit what we can,
204 	 * otherwise force out a byte.
205 	 */
206 	if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
207 	    tp->t_timer[TCPT_PERSIST] == 0) {
208 		tp->t_rxtshift = 0;
209 		tcp_setpersist(tp);
210 	}
211 
212 	/*
213 	 * No reason to send a segment, just return.
214 	 */
215 	return (0);
216 
217 send:
218 	/*
219 	 * Before ESTABLISHED, force sending of initial options
220 	 * unless TCP set not to do any options.
221 	 * NOTE: we assume that the IP/TCP header plus TCP options
222 	 * always fit in a single mbuf, leaving room for a maximum
223 	 * link header, i.e.
224 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MHLEN
225 	 */
226 	optlen = 0;
227 	hdrlen = sizeof (struct tcpiphdr);
228 	if (flags & TH_SYN && (tp->t_flags & TF_NOOPT) == 0) {
229 		u_short mss;
230 
231  		opt[0] = TCPOPT_MAXSEG;
232  		opt[1] = 4;
233 		mss = htons((u_short) tcp_mss(tp, 0));
234 		bcopy((caddr_t)&mss, (caddr_t)(opt + 2), sizeof(mss));
235  		optlen = 4;
236 
237  		if ((tp->t_flags & TF_REQ_SCALE) &&
238  		    ((flags & TH_ACK) == 0 || (tp->t_flags & TF_RCVD_SCALE))) {
239  			*((u_long *) (opt + optlen)) = htonl(
240  				TCPOPT_NOP<<24 |
241  				TCPOPT_WINDOW<<16 |
242  				TCPOLEN_WINDOW<<8 |
243  				tp->request_r_scale);
244  			optlen += 4;
245  		}
246  	}
247 
248  	/*
249 	 * Send a timestamp and echo-reply if this is a SYN and our side
250 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
251 	 * and our peer have sent timestamps in our SYN's.
252  	 */
253  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
254  	     (flags & TH_RST) == 0 &&
255  	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
256 	     (tp->t_flags & TF_RCVD_TSTMP))) {
257 		u_long *lp = (u_long *)(opt + optlen);
258 
259  		/* Form timestamp option as shown in appendix A of RFC 1323. */
260  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
261  		*lp++ = htonl(tcp_now);
262  		*lp   = htonl(tp->ts_recent);
263  		optlen += TCPOLEN_TSTAMP_APPA;
264  	}
265 
266  	hdrlen += optlen;
267 
268 	/*
269 	 * Adjust data length if insertion of options will
270 	 * bump the packet length beyond the t_maxseg length.
271 	 */
272 	 if (len > tp->t_maxseg - optlen) {
273 		len = tp->t_maxseg - optlen;
274 		sendalot = 1;
275 	 }
276 
277 
278 #ifdef DIAGNOSTIC
279  	if (max_linkhdr + hdrlen > MHLEN)
280 		panic("tcphdr too big");
281 #endif
282 
283 	/*
284 	 * Grab a header mbuf, attaching a copy of data to
285 	 * be transmitted, and initialize the header from
286 	 * the template for sends on this connection.
287 	 */
288 	if (len) {
289 		if (tp->t_force && len == 1)
290 			tcpstat.tcps_sndprobe++;
291 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
292 			tcpstat.tcps_sndrexmitpack++;
293 			tcpstat.tcps_sndrexmitbyte += len;
294 		} else {
295 			tcpstat.tcps_sndpack++;
296 			tcpstat.tcps_sndbyte += len;
297 		}
298 #ifdef notyet
299 		if ((m = m_copypack(so->so_snd.sb_mb, off,
300 		    (int)len, max_linkhdr + hdrlen)) == 0) {
301 			error = ENOBUFS;
302 			goto out;
303 		}
304 		/*
305 		 * m_copypack left space for our hdr; use it.
306 		 */
307 		m->m_len += hdrlen;
308 		m->m_data -= hdrlen;
309 #else
310 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
311 		if (m == NULL) {
312 			error = ENOBUFS;
313 			goto out;
314 		}
315 		m->m_data += max_linkhdr;
316 		m->m_len = hdrlen;
317 		if (len <= MHLEN - hdrlen - max_linkhdr) {
318 			m_copydata(so->so_snd.sb_mb, off, (int) len,
319 			    mtod(m, caddr_t) + hdrlen);
320 			m->m_len += len;
321 		} else {
322 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
323 			if (m->m_next == 0)
324 				len = 0;
325 		}
326 #endif
327 		/*
328 		 * If we're sending everything we've got, set PUSH.
329 		 * (This will keep happy those implementations which only
330 		 * give data to the user when a buffer fills or
331 		 * a PUSH comes in.)
332 		 */
333 		if (off + len == so->so_snd.sb_cc)
334 			flags |= TH_PUSH;
335 	} else {
336 		if (tp->t_flags & TF_ACKNOW)
337 			tcpstat.tcps_sndacks++;
338 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
339 			tcpstat.tcps_sndctrl++;
340 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
341 			tcpstat.tcps_sndurg++;
342 		else
343 			tcpstat.tcps_sndwinup++;
344 
345 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
346 		if (m == NULL) {
347 			error = ENOBUFS;
348 			goto out;
349 		}
350 		m->m_data += max_linkhdr;
351 		m->m_len = hdrlen;
352 	}
353 	m->m_pkthdr.rcvif = (struct ifnet *)0;
354 	ti = mtod(m, struct tcpiphdr *);
355 	if (tp->t_template == 0)
356 		panic("tcp_output");
357 	bcopy((caddr_t)tp->t_template, (caddr_t)ti, sizeof (struct tcpiphdr));
358 
359 	/*
360 	 * Fill in fields, remembering maximum advertised
361 	 * window for use in delaying messages about window sizes.
362 	 * If resending a FIN, be sure not to use a new sequence number.
363 	 */
364 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
365 	    tp->snd_nxt == tp->snd_max)
366 		tp->snd_nxt--;
367 	ti->ti_seq = htonl(tp->snd_nxt);
368 	ti->ti_ack = htonl(tp->rcv_nxt);
369 	if (optlen) {
370 		bcopy((caddr_t)opt, (caddr_t)(ti + 1), optlen);
371 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
372 	}
373 	ti->ti_flags = flags;
374 	/*
375 	 * Calculate receive window.  Don't shrink window,
376 	 * but avoid silly window syndrome.
377 	 */
378 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
379 		win = 0;
380 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
381 		win = (long)TCP_MAXWIN << tp->rcv_scale;
382 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
383 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
384 	ti->ti_win = htons((u_short) (win>>tp->rcv_scale));
385 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
386 		ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
387 		ti->ti_flags |= TH_URG;
388 	} else
389 		/*
390 		 * If no urgent pointer to send, then we pull
391 		 * the urgent pointer to the left edge of the send window
392 		 * so that it doesn't drift into the send window on sequence
393 		 * number wraparound.
394 		 */
395 		tp->snd_up = tp->snd_una;		/* drag it along */
396 
397 	/*
398 	 * Put TCP length in extended header, and then
399 	 * checksum extended header and data.
400 	 */
401 	if (len + optlen)
402 		ti->ti_len = htons((u_short)(sizeof (struct tcphdr) +
403 		    optlen + len));
404 	ti->ti_sum = in_cksum(m, (int)(hdrlen + len));
405 
406 	/*
407 	 * In transmit state, time the transmission and arrange for
408 	 * the retransmit.  In persist state, just set snd_max.
409 	 */
410 	if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) {
411 		tcp_seq startseq = tp->snd_nxt;
412 
413 		/*
414 		 * Advance snd_nxt over sequence space of this segment.
415 		 */
416 		if (flags & (TH_SYN|TH_FIN)) {
417 			if (flags & TH_SYN)
418 				tp->snd_nxt++;
419 			if (flags & TH_FIN) {
420 				tp->snd_nxt++;
421 				tp->t_flags |= TF_SENTFIN;
422 			}
423 		}
424 		tp->snd_nxt += len;
425 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
426 			tp->snd_max = tp->snd_nxt;
427 			/*
428 			 * Time this transmission if not a retransmission and
429 			 * not currently timing anything.
430 			 */
431 			if (tp->t_rtt == 0) {
432 				tp->t_rtt = 1;
433 				tp->t_rtseq = startseq;
434 				tcpstat.tcps_segstimed++;
435 			}
436 		}
437 
438 		/*
439 		 * Set retransmit timer if not currently set,
440 		 * and not doing an ack or a keep-alive probe.
441 		 * Initial value for retransmit timer is smoothed
442 		 * round-trip time + 2 * round-trip time variance.
443 		 * Initialize shift counter which is used for backoff
444 		 * of retransmit time.
445 		 */
446 		if (tp->t_timer[TCPT_REXMT] == 0 &&
447 		    tp->snd_nxt != tp->snd_una) {
448 			tp->t_timer[TCPT_REXMT] = tp->t_rxtcur;
449 			if (tp->t_timer[TCPT_PERSIST]) {
450 				tp->t_timer[TCPT_PERSIST] = 0;
451 				tp->t_rxtshift = 0;
452 			}
453 		}
454 	} else
455 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
456 			tp->snd_max = tp->snd_nxt + len;
457 
458 	/*
459 	 * Trace.
460 	 */
461 	if (so->so_options & SO_DEBUG)
462 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
463 
464 	/*
465 	 * Fill in IP length and desired time to live and
466 	 * send to IP level.  There should be a better way
467 	 * to handle ttl and tos; we could keep them in
468 	 * the template, but need a way to checksum without them.
469 	 */
470 	m->m_pkthdr.len = hdrlen + len;
471 	((struct ip *)ti)->ip_len = m->m_pkthdr.len;
472 	((struct ip *)ti)->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;	/* XXX */
473 	((struct ip *)ti)->ip_tos = tp->t_inpcb->inp_ip.ip_tos;	/* XXX */
474 #if BSD >= 43
475 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
476 	    so->so_options & SO_DONTROUTE, 0);
477 #else
478 	error = ip_output(m, (struct mbuf *)0, &tp->t_inpcb->inp_route,
479 	    so->so_options & SO_DONTROUTE);
480 #endif
481 	if (error) {
482 out:
483 		if (error == ENOBUFS) {
484 			tcp_quench(tp->t_inpcb);
485 			return (0);
486 		}
487 		if ((error == EHOSTUNREACH || error == ENETDOWN)
488 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
489 			tp->t_softerror = error;
490 			return (0);
491 		}
492 		return (error);
493 	}
494 	tcpstat.tcps_sndtotal++;
495 
496 	/*
497 	 * Data sent (as far as we can tell).
498 	 * If this advertises a larger window than any other segment,
499 	 * then remember the size of the advertised window.
500 	 * Any pending ACK has now been sent.
501 	 */
502 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
503 		tp->rcv_adv = tp->rcv_nxt + win;
504 	tp->last_ack_sent = tp->rcv_nxt;
505 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
506 	if (sendalot)
507 		goto again;
508 	return (0);
509 }
510 
511 tcp_setpersist(tp)
512 	register struct tcpcb *tp;
513 {
514 	register t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
515 
516 	if (tp->t_timer[TCPT_REXMT])
517 		panic("tcp_output REXMT");
518 	/*
519 	 * Start/restart persistance timer.
520 	 */
521 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
522 	    t * tcp_backoff[tp->t_rxtshift],
523 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
524 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
525 		tp->t_rxtshift++;
526 }
527