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 7.5 (Berkeley) 06/06/87 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 sendalot = 1; 111 } 112 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 113 flags &= ~TH_FIN; 114 win = sbspace(&so->so_rcv); 115 116 117 /* 118 * If our state indicates that FIN should be sent 119 * and we have not yet done so, or we're retransmitting the FIN, 120 * then we need to send. 121 */ 122 if (flags & TH_FIN && 123 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 124 goto send; 125 /* 126 * Send if we owe peer an ACK. 127 */ 128 if (tp->t_flags & TF_ACKNOW) 129 goto send; 130 if (flags & (TH_SYN|TH_RST)) 131 goto send; 132 if (SEQ_GT(tp->snd_up, tp->snd_una)) 133 goto send; 134 135 /* 136 * Sender silly window avoidance. If connection is idle 137 * and can send all data, a maximum segment, 138 * at least a maximum default-size segment do it, 139 * or are forced, do it; otherwise don't bother. 140 * If peer's buffer is tiny, then send 141 * when window is at least half open. 142 * If retransmitting (possibly after persist timer forced us 143 * to send into a small window), then must resend. 144 */ 145 if (len) { 146 if (len == tp->t_maxseg || len >= TCP_MSS) /* a lot */ 147 goto send; 148 if ((idle || tp->t_flags & TF_NODELAY) && 149 len + off >= so->so_snd.sb_cc) 150 goto send; 151 if (tp->t_force) 152 goto send; 153 if (len >= tp->max_sndwnd / 2) 154 goto send; 155 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 156 goto send; 157 } 158 159 /* 160 * Compare available window to amount of window 161 * known to peer (as advertised window less 162 * next expected input.) If the difference is 35% or more of the 163 * maximum possible window, then want to send a window update to peer. 164 */ 165 if (win > 0 && 166 ((100*(win-(tp->rcv_adv-tp->rcv_nxt))/so->so_rcv.sb_hiwat) >= 35)) 167 goto send; 168 169 /* 170 * TCP window updates are not reliable, rather a polling protocol 171 * using ``persist'' packets is used to insure receipt of window 172 * updates. The three ``states'' for the output side are: 173 * idle not doing retransmits or persists 174 * persisting to move a small or zero window 175 * (re)transmitting and thereby not persisting 176 * 177 * tp->t_timer[TCPT_PERSIST] 178 * is set when we are in persist state. 179 * tp->t_force 180 * is set when we are called to send a persist packet. 181 * tp->t_timer[TCPT_REXMT] 182 * is set when we are retransmitting 183 * The output side is idle when both timers are zero. 184 * 185 * If send window is too small, there is data to transmit, and no 186 * retransmit or persist is pending, then go to persist state. 187 * If nothing happens soon, send when timer expires: 188 * if window is nonzero, transmit what we can, 189 * otherwise force out a byte. 190 */ 191 if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 && 192 tp->t_timer[TCPT_PERSIST] == 0) { 193 tp->t_rxtshift = 0; 194 tcp_setpersist(tp); 195 } 196 197 /* 198 * No reason to send a segment, just return. 199 */ 200 return (0); 201 202 send: 203 /* 204 * Grab a header mbuf, attaching a copy of data to 205 * be transmitted, and initialize the header from 206 * the template for sends on this connection. 207 */ 208 MGET(m, M_DONTWAIT, MT_HEADER); 209 if (m == NULL) 210 return (ENOBUFS); 211 m->m_off = MMAXOFF - sizeof (struct tcpiphdr); 212 m->m_len = sizeof (struct tcpiphdr); 213 if (len) { 214 if (tp->t_force && len == 1) 215 tcpstat.tcps_sndprobe++; 216 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 217 tcpstat.tcps_sndrexmitpack++; 218 tcpstat.tcps_sndrexmitbyte += len; 219 } else { 220 tcpstat.tcps_sndpack++; 221 tcpstat.tcps_sndbyte += len; 222 } 223 m->m_next = m_copy(so->so_snd.sb_mb, off, len); 224 if (m->m_next == 0) 225 len = 0; 226 } else if (tp->t_flags & TF_ACKNOW) 227 tcpstat.tcps_sndacks++; 228 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 229 tcpstat.tcps_sndctrl++; 230 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 231 tcpstat.tcps_sndurg++; 232 else 233 tcpstat.tcps_sndwinup++; 234 235 ti = mtod(m, struct tcpiphdr *); 236 if (tp->t_template == 0) 237 panic("tcp_output"); 238 bcopy((caddr_t)tp->t_template, (caddr_t)ti, sizeof (struct tcpiphdr)); 239 240 /* 241 * Fill in fields, remembering maximum advertised 242 * window for use in delaying messages about window sizes. 243 * If resending a FIN, be sure not to use a new sequence number. 244 */ 245 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && len == 0) 246 tp->snd_nxt--; 247 ti->ti_seq = htonl(tp->snd_nxt); 248 ti->ti_ack = htonl(tp->rcv_nxt); 249 /* 250 * Before ESTABLISHED, force sending of initial options 251 * unless TCP set to not do any options. 252 */ 253 opt = NULL; 254 if (tp->t_state < TCPS_ESTABLISHED && (tp->t_flags & TF_NOOPT) == 0) { 255 u_short mss; 256 257 mss = MIN(so->so_rcv.sb_hiwat / 2, tcp_mss(tp)); 258 if (mss > IP_MSS - sizeof(struct tcpiphdr)) { 259 opt = tcp_initopt; 260 optlen = sizeof (tcp_initopt); 261 *(u_short *)(opt + 2) = htons(mss); 262 } 263 } else if (tp->t_tcpopt) { 264 opt = mtod(tp->t_tcpopt, u_char *); 265 optlen = tp->t_tcpopt->m_len; 266 } 267 if (opt) { 268 m0 = m->m_next; 269 m->m_next = m_get(M_DONTWAIT, MT_DATA); 270 if (m->m_next == 0) { 271 (void) m_free(m); 272 m_freem(m0); 273 return (ENOBUFS); 274 } 275 m->m_next->m_next = m0; 276 m0 = m->m_next; 277 m0->m_len = optlen; 278 bcopy((caddr_t)opt, mtod(m0, caddr_t), optlen); 279 opt = (u_char *)(mtod(m0, caddr_t) + optlen); 280 while (m0->m_len & 0x3) { 281 *opt++ = TCPOPT_EOL; 282 m0->m_len++; 283 } 284 optlen = m0->m_len; 285 ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2; 286 } 287 ti->ti_flags = flags; 288 /* 289 * Calculate receive window. Don't shrink window, 290 * but avoid silly window syndrome. 291 */ 292 if (win < so->so_rcv.sb_hiwat / 4 && win < tp->t_maxseg) 293 win = 0; 294 if (win < (int)(tp->rcv_adv - tp->rcv_nxt)) 295 win = (int)(tp->rcv_adv - tp->rcv_nxt); 296 ti->ti_win = htons((u_short)win); 297 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 298 ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 299 ti->ti_flags |= TH_URG; 300 } else 301 /* 302 * If no urgent pointer to send, then we pull 303 * the urgent pointer to the left edge of the send window 304 * so that it doesn't drift into the send window on sequence 305 * number wraparound. 306 */ 307 tp->snd_up = tp->snd_una; /* drag it along */ 308 /* 309 * If anything to send and we can send it all, set PUSH. 310 * (This will keep happy those implementations which only 311 * give data to the user when a buffer fills or a PUSH comes in.) 312 */ 313 if (len && off+len == so->so_snd.sb_cc) 314 ti->ti_flags |= TH_PUSH; 315 316 /* 317 * Put TCP length in extended header, and then 318 * checksum extended header and data. 319 */ 320 if (len + optlen) 321 ti->ti_len = htons((u_short)(sizeof(struct tcphdr) + 322 optlen + len)); 323 ti->ti_sum = in_cksum(m, sizeof (struct tcpiphdr) + (int)optlen + len); 324 325 /* 326 * In transmit state, time the transmission and arrange for 327 * the retransmit. In persist state, just set snd_max. 328 */ 329 if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) { 330 /* 331 * Advance snd_nxt over sequence space of this segment. 332 */ 333 if (flags & TH_SYN) 334 tp->snd_nxt++; 335 if (flags & TH_FIN) { 336 tp->snd_nxt++; 337 tp->t_flags |= TF_SENTFIN; 338 } 339 tp->snd_nxt += len; 340 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 341 tp->snd_max = tp->snd_nxt; 342 /* 343 * Time this transmission if not a retransmission and 344 * not currently timing anything. 345 */ 346 if (tp->t_rtt == 0) { 347 tp->t_rtt = 1; 348 tp->t_rtseq = tp->snd_nxt - len; 349 tcpstat.tcps_segstimed++; 350 } 351 } 352 353 /* 354 * Set retransmit timer if not currently set, 355 * and not doing an ack or a keep-alive probe. 356 * Initial value for retransmit timer is tcp_beta*tp->t_srtt. 357 * Initialize shift counter which is used for backoff 358 * of retransmit time. 359 */ 360 if (tp->t_timer[TCPT_REXMT] == 0 && 361 tp->snd_nxt != tp->snd_una) { 362 TCPT_RANGESET(tp->t_timer[TCPT_REXMT], 363 tcp_beta * (tp->t_srtt ? tp->t_srtt : TCPTV_SRTTDFLT), 364 TCPTV_MIN, TCPTV_MAX); 365 tp->t_rxtshift = 0; 366 tp->t_timer[TCPT_PERSIST] = 0; 367 } 368 } else 369 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 370 tp->snd_max = tp->snd_nxt + len; 371 372 /* 373 * Trace. 374 */ 375 if (so->so_options & SO_DEBUG) 376 tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0); 377 378 /* 379 * Fill in IP length and desired time to live and 380 * send to IP level. 381 */ 382 ((struct ip *)ti)->ip_len = sizeof (struct tcpiphdr) + optlen + len; 383 ((struct ip *)ti)->ip_ttl = TCP_TTL; 384 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route, 385 so->so_options & SO_DONTROUTE); 386 if (error) 387 return (error); 388 tcpstat.tcps_sndtotal++; 389 390 /* 391 * Data sent (as far as we can tell). 392 * If this advertises a larger window than any other segment, 393 * then remember the size of the advertised window. 394 * Any pending ACK has now been sent. 395 */ 396 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv)) 397 tp->rcv_adv = tp->rcv_nxt + win; 398 tp->t_flags &= ~(TF_ACKNOW|TF_DELACK); 399 if (sendalot) 400 goto again; 401 return (0); 402 } 403 404 tcp_setpersist(tp) 405 register struct tcpcb *tp; 406 { 407 408 if (tp->t_timer[TCPT_REXMT]) 409 panic("tcp_output REXMT"); 410 /* 411 * Start/restart persistance timer. 412 */ 413 TCPT_RANGESET(tp->t_timer[TCPT_PERSIST], 414 ((int)(tcp_beta * tp->t_srtt)) << tp->t_rxtshift, 415 TCPTV_PERSMIN, TCPTV_MAX); 416 tp->t_rxtshift++; 417 if (tp->t_rxtshift >= TCP_MAXRXTSHIFT) 418 tp->t_rxtshift = 0; 419 } 420