xref: /csrg-svn/sys/netinet/tcp_output.c (revision 30657)
1 /*
2  * Copyright (c) 1982, 1986 Regents of the University of California.
3  * All rights reserved.  The Berkeley software License Agreement
4  * specifies the terms and conditions for redistribution.
5  *
6  *	@(#)tcp_output.c	7.4 (Berkeley) 03/20/87
7  */
8 
9 #include "param.h"
10 #include "systm.h"
11 #include "mbuf.h"
12 #include "protosw.h"
13 #include "socket.h"
14 #include "socketvar.h"
15 #include "errno.h"
16 
17 #include "../net/route.h"
18 
19 #include "in.h"
20 #include "in_pcb.h"
21 #include "in_systm.h"
22 #include "ip.h"
23 #include "ip_var.h"
24 #include "tcp.h"
25 #define	TCPOUTFLAGS
26 #include "tcp_fsm.h"
27 #include "tcp_seq.h"
28 #include "tcp_timer.h"
29 #include "tcp_var.h"
30 #include "tcpip.h"
31 #include "tcp_debug.h"
32 
33 /*
34  * Initial options.
35  */
36 u_char	tcp_initopt[4] = { TCPOPT_MAXSEG, 4, 0x0, 0x0, };
37 
38 /*
39  * Tcp output routine: figure out what should be sent and send it.
40  */
41 tcp_output(tp)
42 	register struct tcpcb *tp;
43 {
44 	register struct socket *so = tp->t_inpcb->inp_socket;
45 	register int len, win;
46 	struct mbuf *m0;
47 	int off, flags, error;
48 	register struct mbuf *m;
49 	register struct tcpiphdr *ti;
50 	u_char *opt;
51 	unsigned optlen = 0;
52 	int idle, sendalot;
53 
54 	/*
55 	 * Determine length of data that should be transmitted,
56 	 * and flags that will be used.
57 	 * If there is some data or critical controls (SYN, RST)
58 	 * to send, then transmit; otherwise, investigate further.
59 	 */
60 	idle = (tp->snd_max == tp->snd_una);
61 again:
62 	sendalot = 0;
63 	off = tp->snd_nxt - tp->snd_una;
64 	win = MIN(tp->snd_wnd, tp->snd_cwnd);
65 
66 	/*
67 	 * If in persist timeout with window of 0, send 1 byte.
68 	 * Otherwise, if window is small but nonzero
69 	 * and timer expired, we will send what we can
70 	 * and go to transmit state.
71 	 */
72 	if (tp->t_force) {
73 		if (win == 0)
74 			win = 1;
75 		else {
76 			tp->t_timer[TCPT_PERSIST] = 0;
77 			tp->t_rxtshift = 0;
78 		}
79 	}
80 
81 	len = MIN(so->so_snd.sb_cc, win) - off;
82 	flags = tcp_outflags[tp->t_state];
83 
84 	if (len < 0) {
85 		/*
86 		 * If FIN has been sent but not acked,
87 		 * but we haven't been called to retransmit,
88 		 * len will be -1; transmit if acking, otherwise no need.
89 		 * Otherwise, window shrank after we sent into it.
90 		 * If window shrank to 0, cancel pending retransmit
91 		 * and pull snd_nxt back to (closed) window.
92 		 * We will enter persist state below.
93 		 * If the window didn't close completely,
94 		 * just wait for an ACK.
95 		 */
96 		if (flags & TH_FIN) {
97 			if (tp->t_flags & TF_ACKNOW)
98 				len = 0;
99 			else
100 				return (0);
101 		} else if (win == 0) {
102 			tp->t_timer[TCPT_REXMT] = 0;
103 			tp->snd_nxt = tp->snd_una;
104 			len = 0;
105 		} else
106 			return (0);
107 	}
108 	if (len > tp->t_maxseg) {
109 		len = tp->t_maxseg;
110 		/*
111 		 * Don't send more than one segment if retransmitting
112 		 * (or persisting, but then we shouldn't be here).
113 		 */
114 		if (tp->t_rxtshift == 0)
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 || len >= TCP_MSS)	/* a lot */
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 35% or more of the
168 	 * maximum possible window, then want to send a window update to peer.
169 	 */
170 	if (win > 0 &&
171 	    ((100*(win-(tp->rcv_adv-tp->rcv_nxt))/so->so_rcv.sb_hiwat) >= 35))
172 		goto send;
173 
174 	/*
175 	 * TCP window updates are not reliable, rather a polling protocol
176 	 * using ``persist'' packets is used to insure receipt of window
177 	 * updates.  The three ``states'' for the output side are:
178 	 *	idle			not doing retransmits or persists
179 	 *	persisting		to move a small or zero window
180 	 *	(re)transmitting	and thereby not persisting
181 	 *
182 	 * tp->t_timer[TCPT_PERSIST]
183 	 *	is set when we are in persist state.
184 	 * tp->t_force
185 	 *	is set when we are called to send a persist packet.
186 	 * tp->t_timer[TCPT_REXMT]
187 	 *	is set when we are retransmitting
188 	 * The output side is idle when both timers are zero.
189 	 *
190 	 * If send window is too small, there is data to transmit, and no
191 	 * retransmit or persist is pending, then go to persist state.
192 	 * If nothing happens soon, send when timer expires:
193 	 * if window is nonzero, transmit what we can,
194 	 * otherwise force out a byte.
195 	 */
196 	if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
197 	    tp->t_timer[TCPT_PERSIST] == 0) {
198 		tp->t_rxtshift = 0;
199 		tcp_setpersist(tp);
200 	}
201 
202 	/*
203 	 * No reason to send a segment, just return.
204 	 */
205 	return (0);
206 
207 send:
208 	/*
209 	 * Grab a header mbuf, attaching a copy of data to
210 	 * be transmitted, and initialize the header from
211 	 * the template for sends on this connection.
212 	 */
213 	MGET(m, M_DONTWAIT, MT_HEADER);
214 	if (m == NULL)
215 		return (ENOBUFS);
216 	m->m_off = MMAXOFF - sizeof (struct tcpiphdr);
217 	m->m_len = sizeof (struct tcpiphdr);
218 	if (len) {
219 		if (tp->t_force && len == 1)
220 			tcpstat.tcps_sndprobe++;
221 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
222 			tcpstat.tcps_sndrexmitpack++;
223 			tcpstat.tcps_sndrexmitbyte += len;
224 		} else {
225 			tcpstat.tcps_sndpack++;
226 			tcpstat.tcps_sndbyte += len;
227 		}
228 		m->m_next = m_copy(so->so_snd.sb_mb, off, len);
229 		if (m->m_next == 0)
230 			len = 0;
231 	} else if (tp->t_flags & TF_ACKNOW)
232 		tcpstat.tcps_sndacks++;
233 	else if (flags & (TH_SYN|TH_FIN|TH_RST))
234 		tcpstat.tcps_sndctrl++;
235 	else if (SEQ_GT(tp->snd_up, tp->snd_una))
236 		tcpstat.tcps_sndurg++;
237 	else
238 		tcpstat.tcps_sndwinup++;
239 
240 	ti = mtod(m, struct tcpiphdr *);
241 	if (tp->t_template == 0)
242 		panic("tcp_output");
243 	bcopy((caddr_t)tp->t_template, (caddr_t)ti, sizeof (struct tcpiphdr));
244 
245 	/*
246 	 * Fill in fields, remembering maximum advertised
247 	 * window for use in delaying messages about window sizes.
248 	 * If resending a FIN, be sure not to use a new sequence number.
249 	 */
250 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && len == 0)
251 		tp->snd_nxt--;
252 	ti->ti_seq = htonl(tp->snd_nxt);
253 	ti->ti_ack = htonl(tp->rcv_nxt);
254 	/*
255 	 * Before ESTABLISHED, force sending of initial options
256 	 * unless TCP set to not do any options.
257 	 */
258 	opt = NULL;
259 	if (tp->t_state < TCPS_ESTABLISHED && (tp->t_flags & TF_NOOPT) == 0) {
260 		u_short mss;
261 
262 		mss = MIN(so->so_rcv.sb_hiwat / 2, tcp_mss(tp));
263 		if (mss > IP_MSS - sizeof(struct tcpiphdr)) {
264 			opt = tcp_initopt;
265 			optlen = sizeof (tcp_initopt);
266 			*(u_short *)(opt + 2) = htons(mss);
267 		}
268 	} else if (tp->t_tcpopt) {
269 		opt = mtod(tp->t_tcpopt, u_char *);
270 		optlen = tp->t_tcpopt->m_len;
271 	}
272 	if (opt) {
273 		m0 = m->m_next;
274 		m->m_next = m_get(M_DONTWAIT, MT_DATA);
275 		if (m->m_next == 0) {
276 			(void) m_free(m);
277 			m_freem(m0);
278 			return (ENOBUFS);
279 		}
280 		m->m_next->m_next = m0;
281 		m0 = m->m_next;
282 		m0->m_len = optlen;
283 		bcopy((caddr_t)opt, mtod(m0, caddr_t), optlen);
284 		opt = (u_char *)(mtod(m0, caddr_t) + optlen);
285 		while (m0->m_len & 0x3) {
286 			*opt++ = TCPOPT_EOL;
287 			m0->m_len++;
288 		}
289 		optlen = m0->m_len;
290 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
291 	}
292 	ti->ti_flags = flags;
293 	/*
294 	 * Calculate receive window.  Don't shrink window,
295 	 * but avoid silly window syndrome.
296 	 */
297 	if (win < so->so_rcv.sb_hiwat / 4 && win < tp->t_maxseg)
298 		win = 0;
299 	if (win < (int)(tp->rcv_adv - tp->rcv_nxt))
300 		win = (int)(tp->rcv_adv - tp->rcv_nxt);
301 	ti->ti_win = htons((u_short)win);
302 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
303 		ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
304 		ti->ti_flags |= TH_URG;
305 	} else
306 		/*
307 		 * If no urgent pointer to send, then we pull
308 		 * the urgent pointer to the left edge of the send window
309 		 * so that it doesn't drift into the send window on sequence
310 		 * number wraparound.
311 		 */
312 		tp->snd_up = tp->snd_una;		/* drag it along */
313 	/*
314 	 * If anything to send and we can send it all, set PUSH.
315 	 * (This will keep happy those implementations which only
316 	 * give data to the user when a buffer fills or a PUSH comes in.)
317 	 */
318 	if (len && off+len == so->so_snd.sb_cc)
319 		ti->ti_flags |= TH_PUSH;
320 
321 	/*
322 	 * Put TCP length in extended header, and then
323 	 * checksum extended header and data.
324 	 */
325 	if (len + optlen)
326 		ti->ti_len = htons((u_short)(sizeof(struct tcphdr) +
327 		    optlen + len));
328 	ti->ti_sum = in_cksum(m, sizeof (struct tcpiphdr) + (int)optlen + len);
329 
330 	/*
331 	 * In transmit state, time the transmission and arrange for
332 	 * the retransmit.  In persist state, just set snd_max.
333 	 */
334 	if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) {
335 		/*
336 		 * Advance snd_nxt over sequence space of this segment.
337 		 */
338 		if (flags & TH_SYN)
339 			tp->snd_nxt++;
340 		if (flags & TH_FIN) {
341 			tp->snd_nxt++;
342 			tp->t_flags |= TF_SENTFIN;
343 		}
344 		tp->snd_nxt += len;
345 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
346 			tp->snd_max = tp->snd_nxt;
347 			/*
348 			 * Time this transmission if not a retransmission and
349 			 * not currently timing anything.
350 			 */
351 			if (tp->t_rtt == 0) {
352 				tp->t_rtt = 1;
353 				tp->t_rtseq = tp->snd_nxt - len;
354 				tcpstat.tcps_segstimed++;
355 			}
356 		}
357 
358 		/*
359 		 * Set retransmit timer if not currently set,
360 		 * and not doing an ack or a keep-alive probe.
361 		 * Initial value for retransmit timer is tcp_beta*tp->t_srtt.
362 		 * Initialize shift counter which is used for backoff
363 		 * of retransmit time.
364 		 */
365 		if (tp->t_timer[TCPT_REXMT] == 0 &&
366 		    tp->snd_nxt != tp->snd_una) {
367 			TCPT_RANGESET(tp->t_timer[TCPT_REXMT],
368 			  tcp_beta * (tp->t_srtt ? tp->t_srtt : TCPTV_SRTTDFLT),
369 			  TCPTV_MIN, TCPTV_MAX);
370 			tp->t_rxtshift = 0;
371 			tp->t_timer[TCPT_PERSIST] = 0;
372 		}
373 	} else
374 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
375 			tp->snd_max = tp->snd_nxt + len;
376 
377 	/*
378 	 * Trace.
379 	 */
380 	if (so->so_options & SO_DEBUG)
381 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
382 
383 	/*
384 	 * Fill in IP length and desired time to live and
385 	 * send to IP level.
386 	 */
387 	((struct ip *)ti)->ip_len = sizeof (struct tcpiphdr) + optlen + len;
388 	((struct ip *)ti)->ip_ttl = TCP_TTL;
389 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
390 	    so->so_options & SO_DONTROUTE);
391 	if (error)
392 		return (error);
393 	tcpstat.tcps_sndtotal++;
394 
395 	/*
396 	 * Data sent (as far as we can tell).
397 	 * If this advertises a larger window than any other segment,
398 	 * then remember the size of the advertised window.
399 	 * Any pending ACK has now been sent.
400 	 */
401 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
402 		tp->rcv_adv = tp->rcv_nxt + win;
403 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
404 	if (sendalot)
405 		goto again;
406 	return (0);
407 }
408 
409 tcp_setpersist(tp)
410 	register struct tcpcb *tp;
411 {
412 
413 	if (tp->t_timer[TCPT_REXMT])
414 		panic("tcp_output REXMT");
415 	/*
416 	 * Start/restart persistance timer.
417 	 */
418 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
419 	    ((int)(tcp_beta * tp->t_srtt)) << tp->t_rxtshift,
420 	    TCPTV_PERSMIN, TCPTV_MAX);
421 	tp->t_rxtshift++;
422 	if (tp->t_rxtshift >= TCP_MAXRXTSHIFT)
423 		tp->t_rxtshift = 0;
424 }
425