xref: /csrg-svn/sys/netinet/tcp_output.c (revision 26834)
1 /*
2  * Copyright (c) 1982 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	6.16 (Berkeley) 03/12/86
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 && off)
74 log(7, "persist offset %d\n", off);
75 		if (win == 0)
76 			win = 1;
77 		else {
78 			tp->t_timer[TCPT_PERSIST] = 0;
79 			tp->t_rxtshift = 0;
80 		}
81 	}
82 
83 	len = MIN(so->so_snd.sb_cc, win) - off;
84 	if (len > tp->t_maxseg) {
85 		len = tp->t_maxseg;
86 		/*
87 		 * Don't send more than one segment if retransmitting
88 		 * (or persisting, but then we shouldn't be here).
89 		 */
90 		if (tp->t_rxtshift == 0)
91 			sendalot = 1;
92 	}
93 
94 	flags = tcp_outflags[tp->t_state];
95 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
96 		flags &= ~TH_FIN;
97 
98 	if (len < 0) {
99 		/*
100 		 * If closing and all data acked, len will be -1;
101 		 * retransmit FIN if necessary.
102 		 * Otherwise, window shrank after we sent into it.
103 		 * If window shrank to 0, cancel pending retransmit
104 		 * and pull snd_nxt back to (closed) window.
105 		 * We will enter persist state below.
106 		 */
107 		if (flags & TH_FIN)
108 			len = 0;
109 		else if (win == 0) {
110 			tp->t_timer[TCPT_REXMT] = 0;
111 			tp->snd_nxt = tp->snd_una;
112 			len = 0;
113 		} else
114 			return (0);
115 	}
116 	win = sbspace(&so->so_rcv);
117 
118 	/*
119 	 * Send if we owe peer an ACK.
120 	 */
121 	if (tp->t_flags & TF_ACKNOW)
122 		goto send;
123 	if (flags & (TH_SYN|TH_RST|TH_FIN))
124 		goto send;
125 	if (SEQ_GT(tp->snd_up, tp->snd_una))
126 		goto send;
127 
128 	/*
129 	 * Sender silly window avoidance.  If connection is idle
130 	 * and can send all data, a maximum segment,
131 	 * at least a maximum default-size segment do it,
132 	 * or are forced, do it; otherwise don't bother.
133 	 * If peer's buffer is tiny, then send
134 	 * when window is at least half open.
135 	 * If retransmitting (possibly after persist timer forced us
136 	 * to send into a small window), then must resend.
137 	 */
138 	if (len) {
139 		if (len == tp->t_maxseg || len >= TCP_MSS)	/* a lot */
140 			goto send;
141 		if ((idle || tp->t_flags & TF_NODELAY) &&
142 		    len + off >= so->so_snd.sb_cc)
143 			goto send;
144 		if (tp->t_force)
145 			goto send;
146 		if (len >= tp->max_sndwnd / 2)
147 			goto send;
148 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
149 			goto send;
150 	}
151 
152 	/*
153 	 * Compare available window to amount of window
154 	 * known to peer (as advertised window less
155 	 * next expected input.)  If the difference is 35% or more of the
156 	 * maximum possible window, then want to send a window update to peer.
157 	 */
158 	if (win > 0 &&
159 	    ((100*(win-(tp->rcv_adv-tp->rcv_nxt))/so->so_rcv.sb_hiwat) >= 35))
160 		goto send;
161 
162 	/*
163 	 * TCP window updates are not reliable, rather a polling protocol
164 	 * using ``persist'' packets is used to insure receipt of window
165 	 * updates.  The three ``states'' for the output side are:
166 	 *	idle			not doing retransmits or persists
167 	 *	persisting		to move a small or zero window
168 	 *	(re)transmitting	and thereby not persisting
169 	 *
170 	 * tp->t_timer[TCPT_PERSIST]
171 	 *	is set when we are in persist state.
172 	 * tp->t_force
173 	 *	is set when we are called to send a persist packet.
174 	 * tp->t_timer[TCPT_REXMT]
175 	 *	is set when we are retransmitting
176 	 * The output side is idle when both timers are zero.
177 	 *
178 	 * If send window is too small, there is data to transmit, and no
179 	 * retransmit or persist is pending, then go to persist state.
180 	 * If nothing happens soon, send when timer expires:
181 	 * if window is nonzero, transmit what we can,
182 	 * otherwise force out a byte.
183 	 */
184 	if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
185 	    tp->t_timer[TCPT_PERSIST] == 0) {
186 		tp->t_rxtshift = 0;
187 		tcp_setpersist(tp);
188 	}
189 
190 	/*
191 	 * No reason to send a segment, just return.
192 	 */
193 	return (0);
194 
195 send:
196 	/*
197 	 * Grab a header mbuf, attaching a copy of data to
198 	 * be transmitted, and initialize the header from
199 	 * the template for sends on this connection.
200 	 */
201 	MGET(m, M_DONTWAIT, MT_HEADER);
202 	if (m == NULL)
203 		return (ENOBUFS);
204 	m->m_off = MMAXOFF - sizeof (struct tcpiphdr);
205 	m->m_len = sizeof (struct tcpiphdr);
206 	if (len) {
207 		m->m_next = m_copy(so->so_snd.sb_mb, off, len);
208 		if (m->m_next == 0)
209 			len = 0;
210 	}
211 	ti = mtod(m, struct tcpiphdr *);
212 	if (tp->t_template == 0)
213 		panic("tcp_output");
214 	bcopy((caddr_t)tp->t_template, (caddr_t)ti, sizeof (struct tcpiphdr));
215 
216 	/*
217 	 * Fill in fields, remembering maximum advertised
218 	 * window for use in delaying messages about window sizes.
219 	 */
220 	ti->ti_seq = htonl(tp->snd_nxt);
221 	ti->ti_ack = htonl(tp->rcv_nxt);
222 	/*
223 	 * Before ESTABLISHED, force sending of initial options
224 	 * unless TCP set to not do any options.
225 	 */
226 	opt = NULL;
227 	if (tp->t_state < TCPS_ESTABLISHED && (tp->t_flags & TF_NOOPT) == 0) {
228 		u_short mss;
229 
230 		mss = MIN(so->so_rcv.sb_hiwat / 2, tcp_mss(tp));
231 		if (mss > IP_MSS - sizeof(struct tcpiphdr)) {
232 			opt = tcp_initopt;
233 			optlen = sizeof (tcp_initopt);
234 			*(u_short *)(opt + 2) = htons(mss);
235 		}
236 	} else if (tp->t_tcpopt) {
237 		opt = mtod(tp->t_tcpopt, u_char *);
238 		optlen = tp->t_tcpopt->m_len;
239 	}
240 	if (opt) {
241 		m0 = m->m_next;
242 		m->m_next = m_get(M_DONTWAIT, MT_DATA);
243 		if (m->m_next == 0) {
244 			(void) m_free(m);
245 			m_freem(m0);
246 			return (ENOBUFS);
247 		}
248 		m->m_next->m_next = m0;
249 		m0 = m->m_next;
250 		m0->m_len = optlen;
251 		bcopy((caddr_t)opt, mtod(m0, caddr_t), optlen);
252 		opt = (u_char *)(mtod(m0, caddr_t) + optlen);
253 		while (m0->m_len & 0x3) {
254 			*opt++ = TCPOPT_EOL;
255 			m0->m_len++;
256 		}
257 		optlen = m0->m_len;
258 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
259 	}
260 	ti->ti_flags = flags;
261 	/*
262 	 * Calculate receive window.  Don't shrink window,
263 	 * but avoid silly window syndrome.
264 	 */
265 	if (win < so->so_rcv.sb_hiwat / 4 && win < tp->t_maxseg)
266 		win = 0;
267 	if (win < (int)(tp->rcv_adv - tp->rcv_nxt))
268 		win = (int)(tp->rcv_adv - tp->rcv_nxt);
269 	ti->ti_win = htons((u_short)win);
270 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
271 		ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
272 		ti->ti_flags |= TH_URG;
273 	} else
274 		/*
275 		 * If no urgent pointer to send, then we pull
276 		 * the urgent pointer to the left edge of the send window
277 		 * so that it doesn't drift into the send window on sequence
278 		 * number wraparound.
279 		 */
280 		tp->snd_up = tp->snd_una;		/* drag it along */
281 	/*
282 	 * If anything to send and we can send it all, set PUSH.
283 	 * (This will keep happy those implementations which only
284 	 * give data to the user when a buffer fills or a PUSH comes in.)
285 	 */
286 	if (len && off+len == so->so_snd.sb_cc)
287 		ti->ti_flags |= TH_PUSH;
288 
289 	/*
290 	 * Put TCP length in extended header, and then
291 	 * checksum extended header and data.
292 	 */
293 	if (len + optlen)
294 		ti->ti_len = htons((u_short)(sizeof(struct tcphdr) +
295 		    optlen + len));
296 	ti->ti_sum = in_cksum(m, sizeof (struct tcpiphdr) + (int)optlen + len);
297 
298 	/*
299 	 * In transmit state, time the transmission and arrange for
300 	 * the retransmit.  In persist state, just set snd_max.
301 	 */
302 	if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) {
303 		/*
304 		 * Advance snd_nxt over sequence space of this segment.
305 		 */
306 		if (flags & (TH_SYN|TH_FIN))
307 			tp->snd_nxt++;
308 		tp->snd_nxt += len;
309 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
310 			tp->snd_max = tp->snd_nxt;
311 			/*
312 			 * Time this transmission if not a retransmission and
313 			 * not currently timing anything.
314 			 */
315 			if (tp->t_rtt == 0) {
316 				tp->t_rtt = 1;
317 				tp->t_rtseq = tp->snd_nxt - len;
318 			}
319 		}
320 
321 		/*
322 		 * Set retransmit timer if not currently set,
323 		 * and not doing an ack or a keep-alive probe.
324 		 * Initial value for retransmit timer is tcp_beta*tp->t_srtt.
325 		 * Initialize shift counter which is used for backoff
326 		 * of retransmit time.
327 		 */
328 		if (tp->t_timer[TCPT_REXMT] == 0 &&
329 		    tp->snd_nxt != tp->snd_una) {
330 			TCPT_RANGESET(tp->t_timer[TCPT_REXMT],
331 			    tcp_beta * tp->t_srtt, TCPTV_MIN, TCPTV_MAX);
332 			tp->t_rxtshift = 0;
333 			tp->t_timer[TCPT_PERSIST] = 0;
334 		}
335 	} else
336 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
337 			tp->snd_max = tp->snd_nxt + len;
338 
339 	/*
340 	 * Trace.
341 	 */
342 	if (so->so_options & SO_DEBUG)
343 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
344 
345 	/*
346 	 * Fill in IP length and desired time to live and
347 	 * send to IP level.
348 	 */
349 	((struct ip *)ti)->ip_len = sizeof (struct tcpiphdr) + optlen + len;
350 	((struct ip *)ti)->ip_ttl = TCP_TTL;
351 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
352 	    so->so_options & SO_DONTROUTE);
353 	if (error)
354 		return (error);
355 
356 	/*
357 	 * Data sent (as far as we can tell).
358 	 * If this advertises a larger window than any other segment,
359 	 * then remember the size of the advertised window.
360 	 * Any pending ACK has now been sent.
361 	 */
362 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
363 		tp->rcv_adv = tp->rcv_nxt + win;
364 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
365 	if (sendalot)
366 		goto again;
367 	return (0);
368 }
369 
370 tcp_setpersist(tp)
371 	register struct tcpcb *tp;
372 {
373 
374 	if (tp->t_timer[TCPT_REXMT])
375 		panic("tcp_output REXMT");
376 	/*
377 	 * Start/restart persistance timer.
378 	 */
379 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
380 	    ((int)(tcp_beta * tp->t_srtt)) << tp->t_rxtshift,
381 	    TCPTV_PERSMIN, TCPTV_MAX);
382 	tp->t_rxtshift++;
383 	if (tp->t_rxtshift >= TCP_MAXRXTSHIFT)
384 		tp->t_rxtshift = 0;
385 }
386