xref: /csrg-svn/sys/netinet/tcp_output.c (revision 17273)
1 /*	tcp_output.c	6.5	84/10/19	*/
2 
3 #include "param.h"
4 #include "systm.h"
5 #include "mbuf.h"
6 #include "protosw.h"
7 #include "socket.h"
8 #include "socketvar.h"
9 #include "errno.h"
10 
11 #include "../net/route.h"
12 
13 #include "in.h"
14 #include "in_pcb.h"
15 #include "in_systm.h"
16 #include "ip.h"
17 #include "ip_var.h"
18 #include "tcp.h"
19 #define	TCPOUTFLAGS
20 #include "tcp_fsm.h"
21 #include "tcp_seq.h"
22 #include "tcp_timer.h"
23 #include "tcp_var.h"
24 #include "tcpip.h"
25 #include "tcp_debug.h"
26 
27 /*
28  * Initial options.
29  */
30 u_char	tcp_initopt[4] = { TCPOPT_MAXSEG, 4, 0x0, 0x0, };
31 
32 /*
33  * Tcp output routine: figure out what should be sent and send it.
34  */
35 tcp_output(tp)
36 	register struct tcpcb *tp;
37 {
38 	register struct socket *so = tp->t_inpcb->inp_socket;
39 	register int len;
40 	struct mbuf *m0;
41 	int off, flags, win, error;
42 	register struct mbuf *m;
43 	register struct tcpiphdr *ti;
44 	u_char *opt;
45 	unsigned optlen = 0;
46 	int sendalot;
47 
48 	/*
49 	 * Determine length of data that should be transmitted,
50 	 * and flags that will be used.
51 	 * If there is some data or critical controls (SYN, RST)
52 	 * to send, then transmit; otherwise, investigate further.
53 	 */
54 again:
55 	sendalot = 0;
56 	off = tp->snd_nxt - tp->snd_una;
57 	len = MIN(so->so_snd.sb_cc, tp->snd_wnd+tp->t_force) - off;
58 	if (len < 0)
59 		return (0);	/* ??? */	/* past FIN */
60 	if (len > tp->t_maxseg) {
61 		len = tp->t_maxseg;
62 		sendalot = 1;
63 	}
64 
65 	flags = tcp_outflags[tp->t_state];
66 	if (tp->snd_nxt + len < tp->snd_una + so->so_snd.sb_cc)
67 		flags &= ~TH_FIN;
68 	if (flags & (TH_SYN|TH_RST|TH_FIN))
69 		goto send;
70 	if (SEQ_GT(tp->snd_up, tp->snd_una))
71 		goto send;
72 
73 	/*
74 	 * Sender silly window avoidance.  If can send all data,
75 	 * a maximum segment, at least 1/4 of window do it,
76 	 * or are forced, do it; otherwise don't bother.
77 	 */
78 	if (len) {
79 		if (len == tp->t_maxseg || off+len >= so->so_snd.sb_cc)
80 			goto send;
81 		if (len * 4 >= tp->snd_wnd)		/* a lot */
82 			goto send;
83 		if (tp->t_force)
84 			goto send;
85 	}
86 
87 	/*
88 	 * Send if we owe peer an ACK.
89 	 */
90 	if (tp->t_flags&TF_ACKNOW)
91 		goto send;
92 
93 
94 	/*
95 	 * Calculate available window in i, and also amount
96 	 * of window known to peer (as advertised window less
97 	 * next expected input.)  If this is 35% or more of the
98 	 * maximum possible window, then want to send a segment to peer.
99 	 */
100 	win = sbspace(&so->so_rcv);
101 	if (win > 0 &&
102 	    ((100*(win-(tp->rcv_adv-tp->rcv_nxt))/so->so_rcv.sb_hiwat) >= 35))
103 		goto send;
104 
105 	/*
106 	 * TCP window updates are not reliable, rather a polling protocol
107 	 * using ``persist'' packets is used to insure receipt of window
108 	 * updates.  The three ``states'' for the output side are:
109 	 *	idle			not doing retransmits or persists
110 	 *	persisting		to move a zero window
111 	 *	(re)transmitting	and thereby not persisting
112 	 *
113 	 * tp->t_timer[TCPT_PERSIST]
114 	 *	is set when we are in persist state.
115 	 * tp->t_force
116 	 *	is set when we are called to send a persist packet.
117 	 * tp->t_timer[TCPT_REXMT]
118 	 *	is set when we are retransmitting
119 	 * The output side is idle when both timers are zero.
120 	 *
121 	 * If send window is closed, there is data to transmit, and no
122 	 * retransmit or persist is pending, then go to persist state,
123 	 * arranging to force out a byte to get more current window information
124 	 * if nothing happens soon.
125 	 */
126 	if (tp->snd_wnd == 0 && so->so_snd.sb_cc &&
127 	    tp->t_timer[TCPT_REXMT] == 0 && tp->t_timer[TCPT_PERSIST] == 0) {
128 		tp->t_rxtshift = 0;
129 		tcp_setpersist(tp);
130 	}
131 
132 	/*
133 	 * No reason to send a segment, just return.
134 	 */
135 	return (0);
136 
137 send:
138 	/*
139 	 * Grab a header mbuf, attaching a copy of data to
140 	 * be transmitted, and initialize the header from
141 	 * the template for sends on this connection.
142 	 */
143 	MGET(m, M_DONTWAIT, MT_HEADER);
144 	if (m == NULL)
145 		return (ENOBUFS);
146 	m->m_off = MMAXOFF - sizeof (struct tcpiphdr);
147 	m->m_len = sizeof (struct tcpiphdr);
148 	if (len) {
149 		m->m_next = m_copy(so->so_snd.sb_mb, off, len);
150 		if (m->m_next == 0)
151 			len = 0;
152 	}
153 	ti = mtod(m, struct tcpiphdr *);
154 	if (tp->t_template == 0)
155 		panic("tcp_output");
156 	bcopy((caddr_t)tp->t_template, (caddr_t)ti, sizeof (struct tcpiphdr));
157 
158 	/*
159 	 * Fill in fields, remembering maximum advertised
160 	 * window for use in delaying messages about window sizes.
161 	 */
162 	ti->ti_seq = tp->snd_nxt;
163 	ti->ti_ack = tp->rcv_nxt;
164 	ti->ti_seq = htonl(ti->ti_seq);
165 	ti->ti_ack = htonl(ti->ti_ack);
166 	/*
167 	 * Before ESTABLISHED, force sending of initial options
168 	 * unless TCP set to not do any options.
169 	 */
170 	if (tp->t_state < TCPS_ESTABLISHED) {
171 		int mss;
172 
173 		if (tp->t_flags&TF_NOOPT)
174 			goto noopt;
175 		mss = MIN(so->so_rcv.sb_hiwat / 2, tcp_mss(tp));
176 		if (mss <= IP_MSS - sizeof(struct tcpiphdr))
177 			goto noopt;
178 		opt = tcp_initopt;
179 		optlen = sizeof (tcp_initopt);
180 		*(u_short *)(opt + 2) = htons(mss);
181 	} else {
182 		if (tp->t_tcpopt == 0)
183 			goto noopt;
184 		opt = mtod(tp->t_tcpopt, u_char *);
185 		optlen = tp->t_tcpopt->m_len;
186 	}
187 	if (opt) {
188 		m0 = m->m_next;
189 		m->m_next = m_get(M_DONTWAIT, MT_DATA);
190 		if (m->m_next == 0) {
191 			(void) m_free(m);
192 			m_freem(m0);
193 			return (ENOBUFS);
194 		}
195 		m->m_next->m_next = m0;
196 		m0 = m->m_next;
197 		m0->m_len = optlen;
198 		bcopy((caddr_t)opt, mtod(m0, caddr_t), optlen);
199 		opt = (u_char *)(mtod(m0, caddr_t) + optlen);
200 		while (m0->m_len & 0x3) {
201 			*opt++ = TCPOPT_EOL;
202 			m0->m_len++;
203 		}
204 		optlen = m0->m_len;
205 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
206 	}
207 noopt:
208 	ti->ti_flags = flags;
209 	win = sbspace(&so->so_rcv);
210 	if (win < so->so_rcv.sb_hiwat / 4)	/* avoid silly window */
211 		win = 0;
212 	if (win > 0)
213 		ti->ti_win = htons((u_short)win);
214 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
215 		ti->ti_urp = tp->snd_up - tp->snd_nxt;
216 		ti->ti_urp = htons(ti->ti_urp);
217 		ti->ti_flags |= TH_URG;
218 	} else
219 		/*
220 		 * If no urgent pointer to send, then we pull
221 		 * the urgent pointer to the left edge of the send window
222 		 * so that it doesn't drift into the send window on sequence
223 		 * number wraparound.
224 		 */
225 		tp->snd_up = tp->snd_una;		/* drag it along */
226 	/*
227 	 * If anything to send and we can send it all, set PUSH.
228 	 * (This will keep happy those implementations which only
229 	 * give data to the user when a buffer fills or a PUSH comes in.)
230 	 */
231 	if (len && off+len == so->so_snd.sb_cc)
232 		ti->ti_flags |= TH_PUSH;
233 
234 	/*
235 	 * Put TCP length in extended header, and then
236 	 * checksum extended header and data.
237 	 */
238 	if (len + optlen) {
239 		ti->ti_len = sizeof (struct tcphdr) + optlen + len;
240 		ti->ti_len = htons((u_short)ti->ti_len);
241 	}
242 	ti->ti_sum = in_cksum(m, sizeof (struct tcpiphdr) + (int)optlen + len);
243 
244 	/*
245 	 * In transmit state, time the transmission and arrange for
246 	 * the retransmit.  In persist state, reset persist time for
247 	 * next persist.
248 	 */
249 	if (tp->t_force == 0) {
250 		/*
251 		 * Advance snd_nxt over sequence space of this segment.
252 		 */
253 		if (flags & (TH_SYN|TH_FIN))
254 			tp->snd_nxt++;
255 		tp->snd_nxt += len;
256 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
257 			tp->snd_max = tp->snd_nxt;
258 			/*
259 			 * Time this transmission if not a retransmission and
260 			 * not currently timing anything.
261 			 */
262 			if (tp->t_rtt == 0) {
263 				tp->t_rtt = 1;
264 				tp->t_rtseq = tp->snd_nxt - len;
265 			}
266 		}
267 
268 		/*
269 		 * Set retransmit timer if not currently set.
270 		 * Initial value for retransmit timer to tcp_beta*tp->t_srtt.
271 		 * Initialize shift counter which is used for exponential
272 		 * backoff of retransmit time.
273 		 */
274 		if (tp->t_timer[TCPT_REXMT] == 0 &&
275 		    tp->snd_nxt != tp->snd_una) {
276 			TCPT_RANGESET(tp->t_timer[TCPT_REXMT],
277 			    tcp_beta * tp->t_srtt, TCPTV_MIN, TCPTV_MAX);
278 			tp->t_rxtshift = 0;
279 		}
280 		tp->t_timer[TCPT_PERSIST] = 0;
281 	} else {
282 		if (SEQ_GT(tp->snd_una+1, tp->snd_max))
283 			tp->snd_max = tp->snd_una+1;
284 	}
285 
286 	/*
287 	 * Trace.
288 	 */
289 	if (so->so_options & SO_DEBUG)
290 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
291 
292 	/*
293 	 * Fill in IP length and desired time to live and
294 	 * send to IP level.
295 	 */
296 	((struct ip *)ti)->ip_len = sizeof (struct tcpiphdr) + optlen + len;
297 	((struct ip *)ti)->ip_ttl = TCP_TTL;
298 	if (so->so_options & SO_DONTROUTE)
299 		error =
300 		   ip_output(m, tp->t_ipopt, (struct route *)0, IP_ROUTETOIF);
301 	else
302 		error = ip_output(m, tp->t_ipopt, &tp->t_inpcb->inp_route, 0);
303 	if (error)
304 		return (error);
305 
306 	/*
307 	 * Data sent (as far as we can tell).
308 	 * If this advertises a larger window than any other segment,
309 	 * then remember the size of the advertised window.
310 	 * Drop send for purpose of ACK requirements.
311 	 */
312 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
313 		tp->rcv_adv = tp->rcv_nxt + win;
314 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
315 	if (sendalot && tp->t_force == 0)
316 		goto again;
317 	return (0);
318 }
319 
320 tcp_setpersist(tp)
321 	register struct tcpcb *tp;
322 {
323 
324 	if (tp->t_timer[TCPT_REXMT])
325 		panic("tcp_output REXMT");
326 	/*
327 	 * Start/restart persistance timer.
328 	 */
329 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
330 	    ((int)(tcp_beta * tp->t_srtt)) << tp->t_rxtshift,
331 	    TCPTV_PERSMIN, TCPTV_MAX);
332 	tp->t_rxtshift++;
333 	if (tp->t_rxtshift >= TCP_MAXRXTSHIFT)
334 		tp->t_rxtshift = 0;
335 }
336