xref: /csrg-svn/sys/netinet/tcp_output.c (revision 10143)
1 /*	tcp_output.c	4.50	83/01/04	*/
2 
3 #include "../h/param.h"
4 #include "../h/systm.h"
5 #include "../h/mbuf.h"
6 #include "../h/protosw.h"
7 #include "../h/socket.h"
8 #include "../h/socketvar.h"
9 #include "../netinet/in.h"
10 #include "../net/route.h"
11 #include "../netinet/in_pcb.h"
12 #include "../netinet/in_systm.h"
13 #include "../netinet/ip.h"
14 #include "../netinet/ip_var.h"
15 #include "../netinet/tcp.h"
16 #define	TCPOUTFLAGS
17 #include "../netinet/tcp_fsm.h"
18 #include "../netinet/tcp_seq.h"
19 #include "../netinet/tcp_timer.h"
20 #include "../netinet/tcp_var.h"
21 #include "../netinet/tcpip.h"
22 #include "../netinet/tcp_debug.h"
23 #include <errno.h>
24 
25 /*
26  * Initial options.
27  */
28 u_char	tcp_initopt[4] = { TCPOPT_MAXSEG, 4, 0x0, 0x0, };
29 
30 /*
31  * Tcp output routine: figure out what should be sent and send it.
32  */
33 tcp_output(tp)
34 	register struct tcpcb *tp;
35 {
36 	register struct socket *so = tp->t_inpcb->inp_socket;
37 	register int len;
38 	struct mbuf *m0;
39 	int off, flags, win, error;
40 	register struct mbuf *m;
41 	register struct tcpiphdr *ti;
42 	u_char *opt;
43 	unsigned optlen = 0;
44 	int sendalot;
45 
46 
47 	/*
48 	 * Determine length of data that should be transmitted,
49 	 * and flags that will be used.
50 	 * If there is some data or critical controls (SYN, RST)
51 	 * to send, then transmit; otherwise, investigate further.
52 	 */
53 again:
54 	sendalot = 0;
55 	off = tp->snd_nxt - tp->snd_una;
56 	len = MIN(so->so_snd.sb_cc, tp->snd_wnd+tp->t_force) - off;
57 	if (len < 0)
58 		return (0);	/* ??? */	/* past FIN */
59 	if (len > tp->t_maxseg) {
60 		len = tp->t_maxseg;
61 		sendalot = 1;
62 	}
63 
64 	flags = tcp_outflags[tp->t_state];
65 	if (tp->snd_nxt + len < tp->snd_una + so->so_snd.sb_cc)
66 		flags &= ~TH_FIN;
67 	if (flags & (TH_SYN|TH_RST|TH_FIN))
68 		goto send;
69 	if (SEQ_GT(tp->snd_up, tp->snd_una))
70 		goto send;
71 
72 	/*
73 	 * Sender silly window avoidance.  If can send all data,
74 	 * a maximum segment, at least 1/4 of window do it,
75 	 * or are forced, do it; otherwise don't bother.
76 	 */
77 	if (len) {
78 		if (len == tp->t_maxseg || off+len >= so->so_snd.sb_cc)
79 			goto send;
80 		if (len * 4 >= tp->snd_wnd)		/* a lot */
81 			goto send;
82 		if (tp->t_force)
83 			goto send;
84 	}
85 
86 	/*
87 	 * Send if we owe peer an ACK.
88 	 */
89 	if (tp->t_flags&TF_ACKNOW)
90 		goto send;
91 
92 
93 	/*
94 	 * Calculate available window in i, and also amount
95 	 * of window known to peer (as advertised window less
96 	 * next expected input.)  If this is 35% or more of the
97 	 * maximum possible window, then want to send a segment to peer.
98 	 */
99 	win = sbspace(&so->so_rcv);
100 	if (win > 0 &&
101 	    ((100*(win-(tp->rcv_adv-tp->rcv_nxt))/so->so_rcv.sb_hiwat) >= 35))
102 		goto send;
103 
104 	/*
105 	 * TCP window updates are not reliable, rather a polling protocol
106 	 * using ``persist'' packets is used to insure receipt of window
107 	 * updates.  The three ``states'' for the output side are:
108 	 *	idle			not doing retransmits or persists
109 	 *	persisting		to move a zero window
110 	 *	(re)transmitting	and thereby not persisting
111 	 *
112 	 * tp->t_timer[TCPT_PERSIST]
113 	 *	is set when we are in persist state.
114 	 * tp->t_force
115 	 *	is set when we are called to send a persist packet.
116 	 * tp->t_timer[TCPT_REXMT]
117 	 *	is set when we are retransmitting
118 	 * The output side is idle when both timers are zero.
119 	 *
120 	 * If send window is closed, there is data to transmit, and no
121 	 * retransmit or persist is pending, then go to persist state,
122 	 * arranging to force out a byte to get more current window information
123 	 * if nothing happens soon.
124 	 */
125 	if (tp->snd_wnd == 0 && so->so_snd.sb_cc &&
126 	    tp->t_timer[TCPT_REXMT] == 0 && tp->t_timer[TCPT_PERSIST] == 0) {
127 		tp->t_rxtshift = 0;
128 		tcp_setpersist(tp);
129 	}
130 
131 	/*
132 	 * No reason to send a segment, just return.
133 	 */
134 	return (0);
135 
136 send:
137 	/*
138 	 * Grab a header mbuf, attaching a copy of data to
139 	 * be transmitted, and initialize the header from
140 	 * the template for sends on this connection.
141 	 */
142 	MGET(m, M_DONTWAIT, MT_DATA);
143 	if (m == INADDR_ANY)
144 		return (ENOBUFS);
145 	m->m_off = MMAXOFF - sizeof (struct tcpiphdr);
146 	m->m_len = sizeof (struct tcpiphdr);
147 	if (len) {
148 		m->m_next = m_copy(so->so_snd.sb_mb, off, len);
149 		if (m->m_next == 0)
150 			len = 0;
151 	}
152 	ti = mtod(m, struct tcpiphdr *);
153 	if (tp->t_template == 0)
154 		panic("tcp_output");
155 	bcopy((caddr_t)tp->t_template, (caddr_t)ti, sizeof (struct tcpiphdr));
156 
157 	/*
158 	 * Fill in fields, remembering maximum advertised
159 	 * window for use in delaying messages about window sizes.
160 	 */
161 	ti->ti_seq = tp->snd_nxt;
162 	ti->ti_ack = tp->rcv_nxt;
163 	ti->ti_seq = htonl(ti->ti_seq);
164 	ti->ti_ack = htonl(ti->ti_ack);
165 	/*
166 	 * Before ESTABLISHED, force sending of initial options
167 	 * unless TCP set to not do any options.
168 	 */
169 	if (tp->t_state < TCPS_ESTABLISHED) {
170 		if (tp->t_flags&TF_NOOPT)
171 			goto noopt;
172 		opt = tcp_initopt;
173 		optlen = sizeof (tcp_initopt);
174 		*(u_short *)(opt + 2) = MIN(so->so_rcv.sb_hiwat / 2, 1024);
175 		*(u_short *)(opt + 2) = htons(*(u_short *)(opt + 2));
176 	} else {
177 		if (tp->t_tcpopt == 0)
178 			goto noopt;
179 		opt = mtod(tp->t_tcpopt, u_char *);
180 		optlen = tp->t_tcpopt->m_len;
181 	}
182 	if (opt) {
183 		m0 = m->m_next;
184 		m->m_next = m_get(M_DONTWAIT, MT_DATA);
185 		if (m->m_next == 0) {
186 			(void) m_free(m);
187 			m_freem(m0);
188 			return (ENOBUFS);
189 		}
190 		m->m_next->m_next = m0;
191 		m0 = m->m_next;
192 		m0->m_len = optlen;
193 		bcopy((caddr_t)opt, mtod(m0, caddr_t), optlen);
194 		opt = (u_char *)(mtod(m0, caddr_t) + optlen);
195 		while (m0->m_len & 0x3) {
196 			*opt++ = TCPOPT_EOL;
197 			m0->m_len++;
198 		}
199 		optlen = m0->m_len;
200 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
201 	}
202 noopt:
203 	ti->ti_flags = flags;
204 	win = sbspace(&so->so_rcv);
205 	if (win < so->so_rcv.sb_hiwat / 4)	/* avoid silly window */
206 		win = 0;
207 	if (win > 0)
208 		ti->ti_win = htons((u_short)win);
209 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
210 		ti->ti_urp = tp->snd_up - tp->snd_nxt;
211 		ti->ti_urp = htons(ti->ti_urp);
212 		ti->ti_flags |= TH_URG;
213 	} else
214 		/*
215 		 * If no urgent pointer to send, then we pull
216 		 * the urgent pointer to the left edge of the send window
217 		 * so that it doesn't drift into the send window on sequence
218 		 * number wraparound.
219 		 */
220 		tp->snd_up = tp->snd_una;		/* drag it along */
221 	/*
222 	 * If anything to send and we can send it all, set PUSH.
223 	 * (This will keep happy those implementations which only
224 	 * give data to the user when a buffer fills or a PUSH comes in.)
225 	 */
226 	if (len && off+len == so->so_snd.sb_cc)
227 		ti->ti_flags |= TH_PUSH;
228 
229 	/*
230 	 * Put TCP length in extended header, and then
231 	 * checksum extended header and data.
232 	 */
233 	if (len + optlen) {
234 		ti->ti_len = sizeof (struct tcphdr) + optlen + len;
235 		ti->ti_len = htons((u_short)ti->ti_len);
236 	}
237 	ti->ti_sum = in_cksum(m, sizeof (struct tcpiphdr) + (int)optlen + len);
238 
239 	/*
240 	 * In transmit state, time the transmission and arrange for
241 	 * the retransmit.  In persist state, reset persist time for
242 	 * next persist.
243 	 */
244 	if (tp->t_force == 0) {
245 		/*
246 		 * Advance snd_nxt over sequence space of this segment.
247 		 */
248 		if (flags & (TH_SYN|TH_FIN))
249 			tp->snd_nxt++;
250 		tp->snd_nxt += len;
251 		if (SEQ_GT(tp->snd_nxt, tp->snd_max))
252 			tp->snd_max = tp->snd_nxt;
253 
254 		/*
255 		 * Time this transmission if not a retransmission and
256 		 * not currently timing anything.
257 		 */
258 		if (SEQ_GT(tp->snd_nxt, tp->snd_max) && tp->t_rtt == 0) {
259 			tp->t_rtt = 1;
260 			tp->t_rtseq = tp->snd_nxt - len;
261 		}
262 
263 		/*
264 		 * Set retransmit timer if not currently set.
265 		 * Initial value for retransmit timer to tcp_beta*tp->t_srtt.
266 		 * Initialize shift counter which is used for exponential
267 		 * backoff of retransmit time.
268 		 */
269 		if (tp->t_timer[TCPT_REXMT] == 0 &&
270 		    tp->snd_nxt != tp->snd_una) {
271 			TCPT_RANGESET(tp->t_timer[TCPT_REXMT],
272 			    tcp_beta * tp->t_srtt, TCPTV_MIN, TCPTV_MAX);
273 			tp->t_rtt = 0;
274 			tp->t_rxtshift = 0;
275 		}
276 		tp->t_timer[TCPT_PERSIST] = 0;
277 	} else {
278 		if (SEQ_GT(tp->snd_una+1, tp->snd_max))
279 			tp->snd_max = tp->snd_una+1;
280 	}
281 
282 	/*
283 	 * Trace.
284 	 */
285 	if (so->so_options & SO_DEBUG)
286 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
287 
288 	/*
289 	 * Fill in IP length and desired time to live and
290 	 * send to IP level.
291 	 */
292 	((struct ip *)ti)->ip_len = sizeof (struct tcpiphdr) + optlen + len;
293 	((struct ip *)ti)->ip_ttl = TCP_TTL;
294 	if (error = ip_output(m, tp->t_ipopt, (so->so_options & SO_DONTROUTE) ?
295 	    &routetoif : &tp->t_inpcb->inp_route, 0))
296 		return (error);
297 
298 	/*
299 	 * Data sent (as far as we can tell).
300 	 * If this advertises a larger window than any other segment,
301 	 * then remember the size of the advertised window.
302 	 * Drop send for purpose of ACK requirements.
303 	 */
304 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
305 		tp->rcv_adv = tp->rcv_nxt + win;
306 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
307 	if (sendalot && tp->t_force == 0)
308 		goto again;
309 	return (0);
310 }
311 
312 tcp_setpersist(tp)
313 	register struct tcpcb *tp;
314 {
315 
316 	if (tp->t_timer[TCPT_REXMT])
317 		panic("tcp_output REXMT");
318 	/*
319 	 * Start/restart persistance timer.
320 	 */
321 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
322 	    ((int)(tcp_beta * tp->t_srtt)) << tp->t_rxtshift,
323 	    TCPTV_PERSMIN, TCPTV_MAX);
324 	tp->t_rxtshift++;
325 	if (tp->t_rxtshift >= TCP_MAXRXTSHIFT)
326 		tp->t_rxtshift = 0;
327 }
328