xref: /csrg-svn/sys/netinet/tcp_output.c (revision 40685)
1 /*
2  * Copyright (c) 1982, 1986, 1988 Regents of the University of California.
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms are permitted
6  * provided that the above copyright notice and this paragraph are
7  * duplicated in all such forms and that any documentation,
8  * advertising materials, and other materials related to such
9  * distribution and use acknowledge that the software was developed
10  * by the University of California, Berkeley.  The name of the
11  * University may not be used to endorse or promote products derived
12  * from this software without specific prior written permission.
13  * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR
14  * IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED
15  * WARRANTIES OF MERCHANTIBILITY AND FITNESS FOR A PARTICULAR PURPOSE.
16  *
17  *	@(#)tcp_output.c	7.19 (Berkeley) 04/03/90
18  */
19 
20 #include "param.h"
21 #include "systm.h"
22 #include "malloc.h"
23 #include "mbuf.h"
24 #include "protosw.h"
25 #include "socket.h"
26 #include "socketvar.h"
27 #include "errno.h"
28 
29 #include "../net/route.h"
30 
31 #include "in.h"
32 #include "in_systm.h"
33 #include "ip.h"
34 #include "in_pcb.h"
35 #include "ip_var.h"
36 #include "tcp.h"
37 #define	TCPOUTFLAGS
38 #include "tcp_fsm.h"
39 #include "tcp_seq.h"
40 #include "tcp_timer.h"
41 #include "tcp_var.h"
42 #include "tcpip.h"
43 #include "tcp_debug.h"
44 
45 /*
46  * Initial options.
47  */
48 u_char	tcp_initopt[4] = { TCPOPT_MAXSEG, 4, 0x0, 0x0, };
49 
50 /*
51  * Tcp output routine: figure out what should be sent and send it.
52  */
53 tcp_output(tp)
54 	register struct tcpcb *tp;
55 {
56 	register struct socket *so = tp->t_inpcb->inp_socket;
57 	register long len, win;
58 	struct mbuf *m0;
59 	int off, flags, error;
60 	register struct mbuf *m;
61 	register struct tcpiphdr *ti;
62 	u_char *opt;
63 	unsigned optlen = 0;
64 	int idle, sendalot;
65 
66 	/*
67 	 * Determine length of data that should be transmitted,
68 	 * and flags that will be used.
69 	 * If there is some data or critical controls (SYN, RST)
70 	 * to send, then transmit; otherwise, investigate further.
71 	 */
72 	idle = (tp->snd_max == tp->snd_una);
73 again:
74 	sendalot = 0;
75 	off = tp->snd_nxt - tp->snd_una;
76 	win = min(tp->snd_wnd, tp->snd_cwnd);
77 
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 			win = 1;
87 		else {
88 			tp->t_timer[TCPT_PERSIST] = 0;
89 			tp->t_rxtshift = 0;
90 		}
91 	}
92 
93 	len = min(so->so_snd.sb_cc, win) - off;
94 	flags = tcp_outflags[tp->t_state];
95 
96 	if (len < 0) {
97 		/*
98 		 * If FIN has been sent but not acked,
99 		 * but we haven't been called to retransmit,
100 		 * len will be -1.  Otherwise, window shrank
101 		 * after we sent into it.  If window shrank to 0,
102 		 * cancel pending retransmit and pull snd_nxt
103 		 * back to (closed) window.  We will enter persist
104 		 * state below.  If the window didn't close completely,
105 		 * just wait for an ACK.
106 		 */
107 		len = 0;
108 		if (win == 0) {
109 			tp->t_timer[TCPT_REXMT] = 0;
110 			tp->snd_nxt = tp->snd_una;
111 		}
112 	}
113 	if (len > tp->t_maxseg) {
114 		len = tp->t_maxseg;
115 		sendalot = 1;
116 	}
117 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
118 		flags &= ~TH_FIN;
119 	win = sbspace(&so->so_rcv);
120 
121 
122 	/*
123 	 * If our state indicates that FIN should be sent
124 	 * and we have not yet done so, or we're retransmitting the FIN,
125 	 * then we need to send.
126 	 */
127 	if (flags & TH_FIN &&
128 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
129 		goto send;
130 	/*
131 	 * Send if we owe peer an ACK.
132 	 */
133 	if (tp->t_flags & TF_ACKNOW)
134 		goto send;
135 	if (flags & (TH_SYN|TH_RST))
136 		goto send;
137 	if (SEQ_GT(tp->snd_up, tp->snd_una))
138 		goto send;
139 
140 	/*
141 	 * Sender silly window avoidance.  If connection is idle
142 	 * and can send all data, a maximum segment,
143 	 * at least a maximum default-size segment do it,
144 	 * or are forced, do it; otherwise don't bother.
145 	 * If peer's buffer is tiny, then send
146 	 * when window is at least half open.
147 	 * If retransmitting (possibly after persist timer forced us
148 	 * to send into a small window), then must resend.
149 	 */
150 	if (len) {
151 		if (len == tp->t_maxseg)
152 			goto send;
153 		if ((idle || tp->t_flags & TF_NODELAY) &&
154 		    len + off >= so->so_snd.sb_cc)
155 			goto send;
156 		if (tp->t_force)
157 			goto send;
158 		if (len >= tp->max_sndwnd / 2)
159 			goto send;
160 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
161 			goto send;
162 	}
163 
164 	/*
165 	 * Compare available window to amount of window
166 	 * known to peer (as advertised window less
167 	 * next expected input).  If the difference is at least two
168 	 * max size segments, or at least 50% of the maximum possible
169 	 * window, then want to send a window update to peer.
170 	 */
171 	if (win > 0) {
172 		int adv = win - (tp->rcv_adv - tp->rcv_nxt);
173 
174 		if (adv >= 2 * tp->t_maxseg)
175 			goto send;
176 		if (2 * adv >= so->so_rcv.sb_hiwat)
177 			goto send;
178 	}
179 
180 	/*
181 	 * TCP window updates are not reliable, rather a polling protocol
182 	 * using ``persist'' packets is used to insure receipt of window
183 	 * updates.  The three ``states'' for the output side are:
184 	 *	idle			not doing retransmits or persists
185 	 *	persisting		to move a small or zero window
186 	 *	(re)transmitting	and thereby not persisting
187 	 *
188 	 * tp->t_timer[TCPT_PERSIST]
189 	 *	is set when we are in persist state.
190 	 * tp->t_force
191 	 *	is set when we are called to send a persist packet.
192 	 * tp->t_timer[TCPT_REXMT]
193 	 *	is set when we are retransmitting
194 	 * The output side is idle when both timers are zero.
195 	 *
196 	 * If send window is too small, there is data to transmit, and no
197 	 * retransmit or persist is pending, then go to persist state.
198 	 * If nothing happens soon, send when timer expires:
199 	 * if window is nonzero, transmit what we can,
200 	 * otherwise force out a byte.
201 	 */
202 	if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
203 	    tp->t_timer[TCPT_PERSIST] == 0) {
204 		tp->t_rxtshift = 0;
205 		tcp_setpersist(tp);
206 	}
207 
208 	/*
209 	 * No reason to send a segment, just return.
210 	 */
211 	return (0);
212 
213 send:
214 	/*
215 	 * Grab a header mbuf, attaching a copy of data to
216 	 * be transmitted, and initialize the header from
217 	 * the template for sends on this connection.
218 	 */
219 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
220 	if (m == NULL)
221 		return (ENOBUFS);
222 	m->m_data += max_linkhdr;
223 	m->m_len = sizeof (struct tcpiphdr);
224 	m->m_pkthdr.rcvif = (struct ifnet *)0;
225 	ti = mtod(m, struct tcpiphdr *);
226 	if (len) {
227 		if (tp->t_force && len == 1)
228 			tcpstat.tcps_sndprobe++;
229 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
230 			tcpstat.tcps_sndrexmitpack++;
231 			tcpstat.tcps_sndrexmitbyte += len;
232 		} else {
233 			tcpstat.tcps_sndpack++;
234 			tcpstat.tcps_sndbyte += len;
235 		}
236 		if (len <= MHLEN - sizeof (struct tcpiphdr) - max_linkhdr) {
237 			m_copydata(so->so_snd.sb_mb, off, (int) len,
238 			    mtod(m, caddr_t) + sizeof(struct tcpiphdr));
239 			m->m_len += len;
240 		} else {
241 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
242 			if (m->m_next == 0)
243 				len = 0;
244 		}
245 	} else if (tp->t_flags & TF_ACKNOW)
246 		tcpstat.tcps_sndacks++;
247 	else if (flags & (TH_SYN|TH_FIN|TH_RST))
248 		tcpstat.tcps_sndctrl++;
249 	else if (SEQ_GT(tp->snd_up, tp->snd_una))
250 		tcpstat.tcps_sndurg++;
251 	else
252 		tcpstat.tcps_sndwinup++;
253 
254 	if (tp->t_template == 0)
255 		panic("tcp_output");
256 	bcopy((caddr_t)tp->t_template, (caddr_t)ti, sizeof (struct tcpiphdr));
257 
258 	/*
259 	 * Fill in fields, remembering maximum advertised
260 	 * window for use in delaying messages about window sizes.
261 	 * If resending a FIN, be sure not to use a new sequence number.
262 	 */
263 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
264 	    tp->snd_nxt == tp->snd_max)
265 		tp->snd_nxt--;
266 	ti->ti_seq = htonl(tp->snd_nxt);
267 	ti->ti_ack = htonl(tp->rcv_nxt);
268 	/*
269 	 * Before ESTABLISHED, force sending of initial options
270 	 * unless TCP set to not do any options.
271 	 */
272 	opt = NULL;
273 	if (flags & TH_SYN && (tp->t_flags & TF_NOOPT) == 0) {
274 		u_short mss;
275 
276 		mss = min(so->so_rcv.sb_hiwat / 2, tcp_mss(tp));
277 		if (mss > IP_MSS - sizeof(struct tcpiphdr)) {
278 			opt = tcp_initopt;
279 			optlen = sizeof (tcp_initopt);
280 			*(u_short *)(opt + 2) = htons(mss);
281 		}
282 	}
283 	if (opt) {
284 		m0 = m->m_next;
285 		m->m_next = m_get(M_DONTWAIT, MT_DATA);
286 		if (m->m_next == 0) {
287 			(void) m_free(m);
288 			m_freem(m0);
289 			return (ENOBUFS);
290 		}
291 		m->m_next->m_next = m0;
292 		m0 = m->m_next;
293 		m0->m_len = optlen;
294 		bcopy((caddr_t)opt, mtod(m0, caddr_t), optlen);
295 		opt = (u_char *)(mtod(m0, caddr_t) + optlen);
296 		while (m0->m_len & 0x3) {
297 			*opt++ = TCPOPT_EOL;
298 			m0->m_len++;
299 		}
300 		optlen = m0->m_len;
301 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
302 	}
303 	ti->ti_flags = flags;
304 	/*
305 	 * Calculate receive window.  Don't shrink window,
306 	 * but avoid silly window syndrome.
307 	 */
308 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
309 		win = 0;
310 	if (win > IP_MAXPACKET)
311 		win = IP_MAXPACKET;
312 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
313 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
314 	ti->ti_win = htons((u_short)win);
315 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
316 		ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
317 		ti->ti_flags |= TH_URG;
318 	} else
319 		/*
320 		 * If no urgent pointer to send, then we pull
321 		 * the urgent pointer to the left edge of the send window
322 		 * so that it doesn't drift into the send window on sequence
323 		 * number wraparound.
324 		 */
325 		tp->snd_up = tp->snd_una;		/* drag it along */
326 	/*
327 	 * If anything to send and we can send it all, set PUSH.
328 	 * (This will keep happy those implementations which only
329 	 * give data to the user when a buffer fills or a PUSH comes in.)
330 	 */
331 	if (len && off+len == so->so_snd.sb_cc)
332 		ti->ti_flags |= TH_PUSH;
333 
334 	/*
335 	 * Put TCP length in extended header, and then
336 	 * checksum extended header and data.
337 	 */
338 	if (len + optlen)
339 		ti->ti_len = htons((u_short)(sizeof(struct tcphdr) +
340 		    optlen + len));
341 	ti->ti_sum = in_cksum(m,
342 	    (int)(sizeof (struct tcpiphdr) + (int)optlen + len));
343 
344 	/*
345 	 * In transmit state, time the transmission and arrange for
346 	 * the retransmit.  In persist state, just set snd_max.
347 	 */
348 	if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) {
349 		tcp_seq startseq = tp->snd_nxt;
350 
351 		/*
352 		 * Advance snd_nxt over sequence space of this segment.
353 		 */
354 		if (flags & TH_SYN)
355 			tp->snd_nxt++;
356 		if (flags & TH_FIN) {
357 			tp->snd_nxt++;
358 			tp->t_flags |= TF_SENTFIN;
359 		}
360 		tp->snd_nxt += len;
361 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
362 			tp->snd_max = tp->snd_nxt;
363 			/*
364 			 * Time this transmission if not a retransmission and
365 			 * not currently timing anything.
366 			 */
367 			if (tp->t_rtt == 0) {
368 				tp->t_rtt = 1;
369 				tp->t_rtseq = startseq;
370 				tcpstat.tcps_segstimed++;
371 			}
372 		}
373 
374 		/*
375 		 * Set retransmit timer if not currently set,
376 		 * and not doing an ack or a keep-alive probe.
377 		 * Initial value for retransmit timer is smoothed
378 		 * round-trip time + 2 * round-trip time variance.
379 		 * Initialize shift counter which is used for backoff
380 		 * of retransmit time.
381 		 */
382 		if (tp->t_timer[TCPT_REXMT] == 0 &&
383 		    tp->snd_nxt != tp->snd_una) {
384 			tp->t_timer[TCPT_REXMT] = tp->t_rxtcur;
385 			if (tp->t_timer[TCPT_PERSIST]) {
386 				tp->t_timer[TCPT_PERSIST] = 0;
387 				tp->t_rxtshift = 0;
388 			}
389 		}
390 	} else
391 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
392 			tp->snd_max = tp->snd_nxt + len;
393 
394 	/*
395 	 * Trace.
396 	 */
397 	if (so->so_options & SO_DEBUG)
398 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
399 
400 	/*
401 	 * Fill in IP length and desired time to live and
402 	 * send to IP level.
403 	 */
404 	((struct ip *)ti)->ip_len = sizeof (struct tcpiphdr) + optlen + len;
405 	if (m->m_flags & M_PKTHDR)
406 		m->m_pkthdr.len = ((struct ip *)ti)->ip_len;
407 	((struct ip *)ti)->ip_ttl = tcp_ttl;
408 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
409 	    so->so_options & SO_DONTROUTE);
410 	if (error) {
411 		if (error == ENOBUFS) {
412 			tcp_quench(tp->t_inpcb);
413 			return (0);
414 		}
415 		return (error);
416 	}
417 	tcpstat.tcps_sndtotal++;
418 
419 	/*
420 	 * Data sent (as far as we can tell).
421 	 * If this advertises a larger window than any other segment,
422 	 * then remember the size of the advertised window.
423 	 * Any pending ACK has now been sent.
424 	 */
425 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
426 		tp->rcv_adv = tp->rcv_nxt + win;
427 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
428 	if (sendalot)
429 		goto again;
430 	return (0);
431 }
432 
433 tcp_setpersist(tp)
434 	register struct tcpcb *tp;
435 {
436 	register t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
437 
438 	if (tp->t_timer[TCPT_REXMT])
439 		panic("tcp_output REXMT");
440 	/*
441 	 * Start/restart persistance timer.
442 	 */
443 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
444 	    t * tcp_backoff[tp->t_rxtshift],
445 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
446 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
447 		tp->t_rxtshift++;
448 }
449