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 7.10 (Berkeley) 10/09/87 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) 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 int adv = win - (tp->rcv_adv - tp->rcv_nxt); 167 168 if (100 * adv / so->so_rcv.sb_hiwat >= 35) 169 goto send; 170 if (adv >= 2 * tp->t_maxseg && so->so_rcv.sb_cc == 0) 171 goto send; 172 } 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 && 251 tp->snd_nxt == tp->snd_max) 252 tp->snd_nxt--; 253 ti->ti_seq = htonl(tp->snd_nxt); 254 ti->ti_ack = htonl(tp->rcv_nxt); 255 /* 256 * Before ESTABLISHED, force sending of initial options 257 * unless TCP set to not do any options. 258 */ 259 opt = NULL; 260 if (flags & TH_SYN && (tp->t_flags & TF_NOOPT) == 0) { 261 u_short mss; 262 263 mss = MIN(so->so_rcv.sb_hiwat / 2, tcp_mss(tp)); 264 if (mss > IP_MSS - sizeof(struct tcpiphdr)) { 265 opt = tcp_initopt; 266 optlen = sizeof (tcp_initopt); 267 *(u_short *)(opt + 2) = htons(mss); 268 } 269 } 270 if (opt) { 271 m0 = m->m_next; 272 m->m_next = m_get(M_DONTWAIT, MT_DATA); 273 if (m->m_next == 0) { 274 (void) m_free(m); 275 m_freem(m0); 276 return (ENOBUFS); 277 } 278 m->m_next->m_next = m0; 279 m0 = m->m_next; 280 m0->m_len = optlen; 281 bcopy((caddr_t)opt, mtod(m0, caddr_t), optlen); 282 opt = (u_char *)(mtod(m0, caddr_t) + optlen); 283 while (m0->m_len & 0x3) { 284 *opt++ = TCPOPT_EOL; 285 m0->m_len++; 286 } 287 optlen = m0->m_len; 288 ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2; 289 } 290 ti->ti_flags = flags; 291 /* 292 * Calculate receive window. Don't shrink window, 293 * but avoid silly window syndrome. 294 */ 295 if (win < so->so_rcv.sb_hiwat / 4 && win < tp->t_maxseg) 296 win = 0; 297 if (win < (int)(tp->rcv_adv - tp->rcv_nxt)) 298 win = (int)(tp->rcv_adv - tp->rcv_nxt); 299 ti->ti_win = htons((u_short)win); 300 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 301 ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 302 ti->ti_flags |= TH_URG; 303 } else 304 /* 305 * If no urgent pointer to send, then we pull 306 * the urgent pointer to the left edge of the send window 307 * so that it doesn't drift into the send window on sequence 308 * number wraparound. 309 */ 310 tp->snd_up = tp->snd_una; /* drag it along */ 311 /* 312 * If anything to send and we can send it all, set PUSH. 313 * (This will keep happy those implementations which only 314 * give data to the user when a buffer fills or a PUSH comes in.) 315 */ 316 if (len && off+len == so->so_snd.sb_cc) 317 ti->ti_flags |= TH_PUSH; 318 319 /* 320 * Put TCP length in extended header, and then 321 * checksum extended header and data. 322 */ 323 if (len + optlen) 324 ti->ti_len = htons((u_short)(sizeof(struct tcphdr) + 325 optlen + len)); 326 ti->ti_sum = in_cksum(m, sizeof (struct tcpiphdr) + (int)optlen + len); 327 328 /* 329 * In transmit state, time the transmission and arrange for 330 * the retransmit. In persist state, just set snd_max. 331 */ 332 if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) { 333 tcp_seq startseq = tp->snd_nxt; 334 335 /* 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 = startseq; 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 smoothed 362 * round-trip time + 2 * round-trip time variance. 363 * Initialize shift counter which is used for backoff 364 * of retransmit time. 365 */ 366 if (tp->t_timer[TCPT_REXMT] == 0 && 367 tp->snd_nxt != tp->snd_una) { 368 tp->t_timer[TCPT_REXMT] = tp->t_rxtcur; 369 if (tp->t_timer[TCPT_PERSIST]) { 370 tp->t_timer[TCPT_PERSIST] = 0; 371 tp->t_rxtshift = 0; 372 } 373 } 374 } else 375 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 376 tp->snd_max = tp->snd_nxt + len; 377 378 /* 379 * Trace. 380 */ 381 if (so->so_options & SO_DEBUG) 382 tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0); 383 384 /* 385 * Fill in IP length and desired time to live and 386 * send to IP level. 387 */ 388 ((struct ip *)ti)->ip_len = sizeof (struct tcpiphdr) + optlen + len; 389 ((struct ip *)ti)->ip_ttl = TCP_TTL; 390 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route, 391 so->so_options & SO_DONTROUTE); 392 if (error) 393 return (error); 394 tcpstat.tcps_sndtotal++; 395 396 /* 397 * Data sent (as far as we can tell). 398 * If this advertises a larger window than any other segment, 399 * then remember the size of the advertised window. 400 * Any pending ACK has now been sent. 401 */ 402 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv)) 403 tp->rcv_adv = tp->rcv_nxt + win; 404 tp->t_flags &= ~(TF_ACKNOW|TF_DELACK); 405 if (sendalot) 406 goto again; 407 return (0); 408 } 409 410 tcp_setpersist(tp) 411 register struct tcpcb *tp; 412 { 413 register t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 414 415 if (tp->t_timer[TCPT_REXMT]) 416 panic("tcp_output REXMT"); 417 /* 418 * Start/restart persistance timer. 419 */ 420 TCPT_RANGESET(tp->t_timer[TCPT_PERSIST], 421 t * tcp_backoff[tp->t_rxtshift], 422 TCPTV_PERSMIN, TCPTV_PERSMAX); 423 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 424 tp->t_rxtshift++; 425 } 426