1 /* $OpenBSD: tcp_output.c,v 1.96 2011/05/13 14:31:17 oga Exp $ */ 2 /* $NetBSD: tcp_output.c,v 1.16 1997/06/03 16:17:09 kml Exp $ */ 3 4 /* 5 * Copyright (c) 1982, 1986, 1988, 1990, 1993 6 * The Regents of the University of California. All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * @(#)COPYRIGHT 1.1 (NRL) 17 January 1995 33 * 34 * NRL grants permission for redistribution and use in source and binary 35 * forms, with or without modification, of the software and documentation 36 * created at NRL provided that the following conditions are met: 37 * 38 * 1. Redistributions of source code must retain the above copyright 39 * notice, this list of conditions and the following disclaimer. 40 * 2. Redistributions in binary form must reproduce the above copyright 41 * notice, this list of conditions and the following disclaimer in the 42 * documentation and/or other materials provided with the distribution. 43 * 3. All advertising materials mentioning features or use of this software 44 * must display the following acknowledgements: 45 * This product includes software developed by the University of 46 * California, Berkeley and its contributors. 47 * This product includes software developed at the Information 48 * Technology Division, US Naval Research Laboratory. 49 * 4. Neither the name of the NRL nor the names of its contributors 50 * may be used to endorse or promote products derived from this software 51 * without specific prior written permission. 52 * 53 * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS 54 * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 55 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A 56 * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NRL OR 57 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 58 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 59 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 60 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 61 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 62 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 63 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 64 * 65 * The views and conclusions contained in the software and documentation 66 * are those of the authors and should not be interpreted as representing 67 * official policies, either expressed or implied, of the US Naval 68 * Research Laboratory (NRL). 69 */ 70 71 #include <sys/param.h> 72 #include <sys/systm.h> 73 #include <sys/mbuf.h> 74 #include <sys/protosw.h> 75 #include <sys/socket.h> 76 #include <sys/socketvar.h> 77 #include <sys/kernel.h> 78 79 #include <net/route.h> 80 #include <net/if.h> 81 82 #include <netinet/in.h> 83 #include <netinet/in_systm.h> 84 #include <netinet/ip.h> 85 #include <netinet/in_pcb.h> 86 #include <netinet/ip_var.h> 87 #include <netinet/tcp.h> 88 #define TCPOUTFLAGS 89 #include <netinet/tcp_fsm.h> 90 #include <netinet/tcp_seq.h> 91 #include <netinet/tcp_timer.h> 92 #include <netinet/tcp_var.h> 93 #include <netinet/tcpip.h> 94 #include <netinet/tcp_debug.h> 95 96 #ifdef INET6 97 #include <netinet6/tcpipv6.h> 98 #include <netinet6/in6_var.h> 99 #endif /* INET6 */ 100 101 #ifdef notyet 102 extern struct mbuf *m_copypack(); 103 #endif 104 105 #ifdef TCP_SACK 106 extern int tcprexmtthresh; 107 #endif 108 109 #ifdef TCP_SACK 110 #ifdef TCP_SACK_DEBUG 111 void tcp_print_holes(struct tcpcb *tp); 112 113 void 114 tcp_print_holes(struct tcpcb *tp) 115 { 116 struct sackhole *p = tp->snd_holes; 117 if (p == 0) 118 return; 119 printf("Hole report: start--end dups rxmit\n"); 120 while (p) { 121 printf("%x--%x d %d r %x\n", p->start, p->end, p->dups, 122 p->rxmit); 123 p = p->next; 124 } 125 printf("\n"); 126 } 127 #endif /* TCP_SACK_DEBUG */ 128 129 /* 130 * Returns pointer to a sackhole if there are any pending retransmissions; 131 * NULL otherwise. 132 */ 133 struct sackhole * 134 tcp_sack_output(struct tcpcb *tp) 135 { 136 struct sackhole *p; 137 138 if (!tp->sack_enable) 139 return (NULL); 140 p = tp->snd_holes; 141 while (p) { 142 #ifndef TCP_FACK 143 if (p->dups >= tcprexmtthresh && SEQ_LT(p->rxmit, p->end)) { 144 #else 145 /* In FACK, if p->dups is less than tcprexmtthresh, but 146 * snd_fack advances more than tcprextmtthresh * tp->t_maxseg, 147 * tcp_input() will try fast retransmit. This forces output. 148 */ 149 if ((p->dups >= tcprexmtthresh || 150 tp->t_dupacks == tcprexmtthresh) && 151 SEQ_LT(p->rxmit, p->end)) { 152 #endif /* TCP_FACK */ 153 if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */ 154 p = p->next; 155 continue; 156 } 157 #ifdef TCP_SACK_DEBUG 158 if (p) 159 tcp_print_holes(tp); 160 #endif 161 return (p); 162 } 163 p = p->next; 164 } 165 return (NULL); 166 } 167 168 /* 169 * After a timeout, the SACK list may be rebuilt. This SACK information 170 * should be used to avoid retransmitting SACKed data. This function 171 * traverses the SACK list to see if snd_nxt should be moved forward. 172 */ 173 174 void 175 tcp_sack_adjust(struct tcpcb *tp) 176 { 177 struct sackhole *cur = tp->snd_holes; 178 if (cur == NULL) 179 return; /* No holes */ 180 if (SEQ_GEQ(tp->snd_nxt, tp->rcv_lastsack)) 181 return; /* We're already beyond any SACKed blocks */ 182 /* 183 * Two cases for which we want to advance snd_nxt: 184 * i) snd_nxt lies between end of one hole and beginning of another 185 * ii) snd_nxt lies between end of last hole and rcv_lastsack 186 */ 187 while (cur->next) { 188 if (SEQ_LT(tp->snd_nxt, cur->end)) 189 return; 190 if (SEQ_GEQ(tp->snd_nxt, cur->next->start)) 191 cur = cur->next; 192 else { 193 tp->snd_nxt = cur->next->start; 194 return; 195 } 196 } 197 if (SEQ_LT(tp->snd_nxt, cur->end)) 198 return; 199 tp->snd_nxt = tp->rcv_lastsack; 200 return; 201 } 202 #endif /* TCP_SACK */ 203 204 /* 205 * Tcp output routine: figure out what should be sent and send it. 206 */ 207 int 208 tcp_output(struct tcpcb *tp) 209 { 210 struct socket *so = tp->t_inpcb->inp_socket; 211 long len, win, txmaxseg; 212 int off, flags, error; 213 struct mbuf *m; 214 struct tcphdr *th; 215 u_int32_t optbuf[howmany(MAX_TCPOPTLEN, sizeof(u_int32_t))]; 216 u_char *opt = (u_char *)optbuf; 217 unsigned int optlen, hdrlen, packetlen; 218 int idle, sendalot = 0; 219 #ifdef TCP_SACK 220 int i, sack_rxmit = 0; 221 struct sackhole *p; 222 int maxburst = TCP_MAXBURST; 223 #endif 224 #ifdef TCP_SIGNATURE 225 unsigned int sigoff; 226 #endif /* TCP_SIGNATURE */ 227 #ifdef TCP_ECN 228 int needect; 229 #endif 230 231 if (tp->t_flags & TF_BLOCKOUTPUT) { 232 tp->t_flags |= TF_NEEDOUTPUT; 233 return (0); 234 } else 235 tp->t_flags &= ~TF_NEEDOUTPUT; 236 237 #if defined(TCP_SACK) && defined(TCP_SIGNATURE) && defined(DIAGNOSTIC) 238 if (tp->sack_enable && (tp->t_flags & TF_SIGNATURE)) 239 return (EINVAL); 240 #endif /* defined(TCP_SACK) && defined(TCP_SIGNATURE) && defined(DIAGNOSTIC) */ 241 242 /* 243 * Determine length of data that should be transmitted, 244 * and flags that will be used. 245 * If there is some data or critical controls (SYN, RST) 246 * to send, then transmit; otherwise, investigate further. 247 */ 248 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 249 if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) 250 /* 251 * We have been idle for "a while" and no acks are 252 * expected to clock out any data we send -- 253 * slow start to get ack "clock" running again. 254 */ 255 tp->snd_cwnd = 2 * tp->t_maxseg; 256 257 /* remember 'idle' for next invocation of tcp_output */ 258 if (idle && soissending(so)) { 259 tp->t_flags |= TF_LASTIDLE; 260 idle = 0; 261 } else 262 tp->t_flags &= ~TF_LASTIDLE; 263 264 again: 265 #ifdef TCP_SACK 266 /* 267 * If we've recently taken a timeout, snd_max will be greater than 268 * snd_nxt. There may be SACK information that allows us to avoid 269 * resending already delivered data. Adjust snd_nxt accordingly. 270 */ 271 if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max)) 272 tcp_sack_adjust(tp); 273 #endif 274 off = tp->snd_nxt - tp->snd_una; 275 #if defined(TCP_SACK) && defined(TCP_FACK) 276 /* Normally, sendable data is limited by off < tp->snd_cwnd. 277 * But in FACK, sendable data is limited by snd_awnd < snd_cwnd, 278 * regardless of offset. 279 */ 280 if (tp->sack_enable && (tp->t_dupacks > tcprexmtthresh)) 281 win = tp->snd_wnd; 282 else 283 #endif 284 win = ulmin(tp->snd_wnd, tp->snd_cwnd); 285 286 flags = tcp_outflags[tp->t_state]; 287 288 #ifdef TCP_SACK 289 /* 290 * Send any SACK-generated retransmissions. If we're explicitly trying 291 * to send out new data (when sendalot is 1), bypass this function. 292 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 293 * we're replacing a (future) new transmission with a retransmission 294 * now, and we previously incremented snd_cwnd in tcp_input(). 295 */ 296 if (tp->sack_enable && !sendalot) { 297 if (tp->t_dupacks >= tcprexmtthresh && 298 (p = tcp_sack_output(tp))) { 299 off = p->rxmit - tp->snd_una; 300 sack_rxmit = 1; 301 /* Coalesce holes into a single retransmission */ 302 len = min(tp->t_maxseg, p->end - p->rxmit); 303 #ifndef TCP_FACK 304 /* in FACK, hold snd_cwnd constant during recovery */ 305 if (SEQ_LT(tp->snd_una, tp->snd_last)) 306 tp->snd_cwnd -= tp->t_maxseg; 307 #endif 308 } 309 } 310 #endif /* TCP_SACK */ 311 312 sendalot = 0; 313 /* 314 * If in persist timeout with window of 0, send 1 byte. 315 * Otherwise, if window is small but nonzero 316 * and timer expired, we will send what we can 317 * and go to transmit state. 318 */ 319 if (tp->t_force) { 320 if (win == 0) { 321 /* 322 * If we still have some data to send, then 323 * clear the FIN bit. Usually this would 324 * happen below when it realizes that we 325 * aren't sending all the data. However, 326 * if we have exactly 1 byte of unset data, 327 * then it won't clear the FIN bit below, 328 * and if we are in persist state, we wind 329 * up sending the packet without recording 330 * that we sent the FIN bit. 331 * 332 * We can't just blindly clear the FIN bit, 333 * because if we don't have any more data 334 * to send then the probe will be the FIN 335 * itself. 336 */ 337 if (off < so->so_snd.sb_cc) 338 flags &= ~TH_FIN; 339 win = 1; 340 } else { 341 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 342 tp->t_rxtshift = 0; 343 } 344 } 345 346 #ifdef TCP_SACK 347 if (!sack_rxmit) { 348 #endif 349 len = ulmin(so->so_snd.sb_cc, win) - off; 350 351 #if defined(TCP_SACK) && defined(TCP_FACK) 352 /* 353 * If we're in fast recovery (SEQ_GT(tp->snd_last, tp->snd_una)), and 354 * amount of outstanding data (snd_awnd) is >= snd_cwnd, then 355 * do not send data (like zero window conditions) 356 */ 357 if (tp->sack_enable && len && SEQ_GT(tp->snd_last, tp->snd_una) && 358 (tp->snd_awnd >= tp->snd_cwnd)) 359 len = 0; 360 #endif /* TCP_FACK */ 361 #ifdef TCP_SACK 362 } 363 #endif 364 365 if (len < 0) { 366 /* 367 * If FIN has been sent but not acked, 368 * but we haven't been called to retransmit, 369 * len will be -1. Otherwise, window shrank 370 * after we sent into it. If window shrank to 0, 371 * cancel pending retransmit, pull snd_nxt back 372 * to (closed) window, and set the persist timer 373 * if it isn't already going. If the window didn't 374 * close completely, just wait for an ACK. 375 */ 376 len = 0; 377 if (win == 0) { 378 TCP_TIMER_DISARM(tp, TCPT_REXMT); 379 tp->t_rxtshift = 0; 380 tp->snd_nxt = tp->snd_una; 381 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) 382 tcp_setpersist(tp); 383 } 384 } 385 386 /* 387 * Never send more than half a buffer full. This insures that we can 388 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and 389 * therefore acks will never be delayed unless we run out of data to 390 * transmit. 391 */ 392 txmaxseg = ulmin(so->so_snd.sb_hiwat / 2, tp->t_maxseg); 393 394 if (len > txmaxseg) { 395 len = txmaxseg; 396 sendalot = 1; 397 } 398 if (off + len < so->so_snd.sb_cc) 399 flags &= ~TH_FIN; 400 401 win = sbspace(&so->so_rcv); 402 403 /* 404 * Sender silly window avoidance. If connection is idle 405 * and can send all data, a maximum segment, 406 * at least a maximum default-size segment do it, 407 * or are forced, do it; otherwise don't bother. 408 * If peer's buffer is tiny, then send 409 * when window is at least half open. 410 * If retransmitting (possibly after persist timer forced us 411 * to send into a small window), then must resend. 412 */ 413 if (len) { 414 if (len == txmaxseg) 415 goto send; 416 if ((idle || tp->t_flags & TF_NODELAY) && 417 len + off >= so->so_snd.sb_cc && !soissending(so)) 418 goto send; 419 if (tp->t_force) 420 goto send; 421 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 422 goto send; 423 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 424 goto send; 425 #ifdef TCP_SACK 426 if (sack_rxmit) 427 goto send; 428 #endif 429 } 430 431 /* 432 * Compare available window to amount of window 433 * known to peer (as advertised window less 434 * next expected input). If the difference is at least two 435 * max size segments, or at least 50% of the maximum possible 436 * window, then want to send a window update to peer. 437 */ 438 if (win > 0) { 439 /* 440 * "adv" is the amount we can increase the window, 441 * taking into account that we are limited by 442 * TCP_MAXWIN << tp->rcv_scale. 443 */ 444 long adv = lmin(win, (long)TCP_MAXWIN << tp->rcv_scale) - 445 (tp->rcv_adv - tp->rcv_nxt); 446 447 if (adv >= (long) (2 * tp->t_maxseg)) 448 goto send; 449 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 450 goto send; 451 } 452 453 /* 454 * Send if we owe peer an ACK. 455 */ 456 if (tp->t_flags & TF_ACKNOW) 457 goto send; 458 if (flags & (TH_SYN|TH_RST)) 459 goto send; 460 if (SEQ_GT(tp->snd_up, tp->snd_una)) 461 goto send; 462 /* 463 * If our state indicates that FIN should be sent 464 * and we have not yet done so, or we're retransmitting the FIN, 465 * then we need to send. 466 */ 467 if (flags & TH_FIN && 468 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 469 goto send; 470 #ifdef TCP_SACK 471 /* 472 * In SACK, it is possible for tcp_output to fail to send a segment 473 * after the retransmission timer has been turned off. Make sure 474 * that the retransmission timer is set. 475 */ 476 if (SEQ_GT(tp->snd_max, tp->snd_una) && 477 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 478 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 479 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 480 return (0); 481 } 482 #endif /* TCP_SACK */ 483 484 /* 485 * TCP window updates are not reliable, rather a polling protocol 486 * using ``persist'' packets is used to insure receipt of window 487 * updates. The three ``states'' for the output side are: 488 * idle not doing retransmits or persists 489 * persisting to move a small or zero window 490 * (re)transmitting and thereby not persisting 491 * 492 * tp->t_timer[TCPT_PERSIST] 493 * is set when we are in persist state. 494 * tp->t_force 495 * is set when we are called to send a persist packet. 496 * tp->t_timer[TCPT_REXMT] 497 * is set when we are retransmitting 498 * The output side is idle when both timers are zero. 499 * 500 * If send window is too small, there is data to transmit, and no 501 * retransmit or persist is pending, then go to persist state. 502 * If nothing happens soon, send when timer expires: 503 * if window is nonzero, transmit what we can, 504 * otherwise force out a byte. 505 */ 506 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 507 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 508 tp->t_rxtshift = 0; 509 tcp_setpersist(tp); 510 } 511 512 /* 513 * No reason to send a segment, just return. 514 */ 515 return (0); 516 517 send: 518 /* 519 * Before ESTABLISHED, force sending of initial options 520 * unless TCP set not to do any options. 521 * NOTE: we assume that the IP/TCP header plus TCP options 522 * always fit in a single mbuf, leaving room for a maximum 523 * link header, i.e. 524 * max_linkhdr + sizeof(network header) + sizeof(struct tcphdr + 525 * optlen <= MHLEN 526 */ 527 optlen = 0; 528 529 switch (tp->pf) { 530 case 0: /*default to PF_INET*/ 531 #ifdef INET 532 case PF_INET: 533 hdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 534 break; 535 #endif /* INET */ 536 #ifdef INET6 537 case PF_INET6: 538 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 539 break; 540 #endif /* INET6 */ 541 default: 542 return (EPFNOSUPPORT); 543 } 544 545 if (flags & TH_SYN) { 546 tp->snd_nxt = tp->iss; 547 if ((tp->t_flags & TF_NOOPT) == 0) { 548 u_int16_t mss; 549 550 opt[0] = TCPOPT_MAXSEG; 551 opt[1] = 4; 552 mss = htons((u_int16_t) tcp_mss(tp, 0)); 553 bcopy((caddr_t)&mss, (caddr_t)(opt + 2), sizeof(mss)); 554 optlen = 4; 555 556 if (flags & TH_ACK) 557 tcp_mss_update(tp); 558 #ifdef TCP_SACK 559 /* 560 * If this is the first SYN of connection (not a SYN 561 * ACK), include SACK_PERMIT_HDR option. If this is a 562 * SYN ACK, include SACK_PERMIT_HDR option if peer has 563 * already done so. 564 */ 565 if (tp->sack_enable && ((flags & TH_ACK) == 0 || 566 (tp->t_flags & TF_SACK_PERMIT))) { 567 *((u_int32_t *) (opt + optlen)) = 568 htonl(TCPOPT_SACK_PERMIT_HDR); 569 optlen += 4; 570 } 571 #endif 572 573 if ((tp->t_flags & TF_REQ_SCALE) && 574 ((flags & TH_ACK) == 0 || 575 (tp->t_flags & TF_RCVD_SCALE))) { 576 *((u_int32_t *) (opt + optlen)) = htonl( 577 TCPOPT_NOP << 24 | 578 TCPOPT_WINDOW << 16 | 579 TCPOLEN_WINDOW << 8 | 580 tp->request_r_scale); 581 optlen += 4; 582 } 583 } 584 } 585 586 /* 587 * Send a timestamp and echo-reply if this is a SYN and our side 588 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 589 * and our peer have sent timestamps in our SYN's. 590 */ 591 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 592 (flags & TH_RST) == 0 && 593 ((flags & (TH_SYN|TH_ACK)) == TH_SYN || 594 (tp->t_flags & TF_RCVD_TSTMP))) { 595 u_int32_t *lp = (u_int32_t *)(opt + optlen); 596 597 /* Form timestamp option as shown in appendix A of RFC 1323. */ 598 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 599 *lp++ = htonl(tcp_now + tp->ts_modulate); 600 *lp = htonl(tp->ts_recent); 601 optlen += TCPOLEN_TSTAMP_APPA; 602 603 /* Set receive buffer autosizing timestamp. */ 604 if (tp->rfbuf_ts == 0) 605 tp->rfbuf_ts = tcp_now; 606 607 } 608 609 #ifdef TCP_SIGNATURE 610 if (tp->t_flags & TF_SIGNATURE) { 611 u_int8_t *bp = (u_int8_t *)(opt + optlen); 612 613 /* Send signature option */ 614 *(bp++) = TCPOPT_SIGNATURE; 615 *(bp++) = TCPOLEN_SIGNATURE; 616 sigoff = optlen + 2; 617 618 { 619 unsigned int i; 620 621 for (i = 0; i < 16; i++) 622 *(bp++) = 0; 623 } 624 625 626 /* Pad options list to the next 32 bit boundary and 627 * terminate it. 628 */ 629 *bp++ = TCPOPT_NOP; 630 *bp++ = TCPOPT_NOP; 631 632 optlen += TCPOLEN_SIGLEN; 633 } 634 #endif /* TCP_SIGNATURE */ 635 636 #ifdef TCP_SACK 637 /* 638 * Send SACKs if necessary. This should be the last option processed. 639 * Only as many SACKs are sent as are permitted by the maximum options 640 * size. No more than three SACKs are sent. 641 */ 642 if (tp->sack_enable && tp->t_state == TCPS_ESTABLISHED && 643 (tp->t_flags & (TF_SACK_PERMIT|TF_NOOPT)) == TF_SACK_PERMIT && 644 tp->rcv_numsacks) { 645 u_int32_t *lp = (u_int32_t *)(opt + optlen); 646 u_int32_t *olp = lp++; 647 int count = 0; /* actual number of SACKs inserted */ 648 int maxsack = (MAX_TCPOPTLEN - (optlen + 4))/TCPOLEN_SACK; 649 650 tcpstat.tcps_sack_snd_opts++; 651 maxsack = min(maxsack, TCP_MAX_SACK); 652 for (i = 0; (i < tp->rcv_numsacks && count < maxsack); i++) { 653 struct sackblk sack = tp->sackblks[i]; 654 if (sack.start == 0 && sack.end == 0) 655 continue; 656 *lp++ = htonl(sack.start); 657 *lp++ = htonl(sack.end); 658 count++; 659 } 660 *olp = htonl(TCPOPT_SACK_HDR|(TCPOLEN_SACK*count+2)); 661 optlen += TCPOLEN_SACK*count + 4; /* including leading NOPs */ 662 } 663 #endif /* TCP_SACK */ 664 665 #ifdef DIAGNOSTIC 666 if (optlen > MAX_TCPOPTLEN) 667 panic("tcp_output: options too long"); 668 #endif /* DIAGNOSTIC */ 669 670 hdrlen += optlen; 671 672 /* 673 * Adjust data length if insertion of options will 674 * bump the packet length beyond the t_maxopd length. 675 */ 676 if (len > tp->t_maxopd - optlen) { 677 len = tp->t_maxopd - optlen; 678 sendalot = 1; 679 flags &= ~TH_FIN; 680 } 681 682 #ifdef DIAGNOSTIC 683 if (max_linkhdr + hdrlen > MCLBYTES) 684 panic("tcphdr too big"); 685 #endif 686 687 /* 688 * Grab a header mbuf, attaching a copy of data to 689 * be transmitted, and initialize the header from 690 * the template for sends on this connection. 691 */ 692 if (len) { 693 if (tp->t_force && len == 1) 694 tcpstat.tcps_sndprobe++; 695 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 696 tcpstat.tcps_sndrexmitpack++; 697 tcpstat.tcps_sndrexmitbyte += len; 698 } else { 699 tcpstat.tcps_sndpack++; 700 tcpstat.tcps_sndbyte += len; 701 } 702 #ifdef notyet 703 if ((m = m_copypack(so->so_snd.sb_mb, off, 704 (int)len, max_linkhdr + hdrlen)) == 0) { 705 error = ENOBUFS; 706 goto out; 707 } 708 /* 709 * m_copypack left space for our hdr; use it. 710 */ 711 m->m_len += hdrlen; 712 m->m_data -= hdrlen; 713 #else 714 MGETHDR(m, M_DONTWAIT, MT_HEADER); 715 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 716 MCLGET(m, M_DONTWAIT); 717 if ((m->m_flags & M_EXT) == 0) { 718 m_freem(m); 719 m = NULL; 720 } 721 } 722 if (m == NULL) { 723 error = ENOBUFS; 724 goto out; 725 } 726 m->m_data += max_linkhdr; 727 m->m_len = hdrlen; 728 if (len <= M_TRAILINGSPACE(m)) { 729 m_copydata(so->so_snd.sb_mb, off, (int) len, 730 mtod(m, caddr_t) + hdrlen); 731 m->m_len += len; 732 } else { 733 m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len); 734 if (m->m_next == 0) { 735 (void) m_free(m); 736 error = ENOBUFS; 737 goto out; 738 } 739 } 740 #endif 741 /* 742 * If we're sending everything we've got, set PUSH. 743 * (This will keep happy those implementations which only 744 * give data to the user when a buffer fills or 745 * a PUSH comes in.) 746 */ 747 if (off + len == so->so_snd.sb_cc && !soissending(so)) 748 flags |= TH_PUSH; 749 } else { 750 if (tp->t_flags & TF_ACKNOW) 751 tcpstat.tcps_sndacks++; 752 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 753 tcpstat.tcps_sndctrl++; 754 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 755 tcpstat.tcps_sndurg++; 756 else 757 tcpstat.tcps_sndwinup++; 758 759 MGETHDR(m, M_DONTWAIT, MT_HEADER); 760 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 761 MCLGET(m, M_DONTWAIT); 762 if ((m->m_flags & M_EXT) == 0) { 763 m_freem(m); 764 m = NULL; 765 } 766 } 767 if (m == NULL) { 768 error = ENOBUFS; 769 goto out; 770 } 771 m->m_data += max_linkhdr; 772 m->m_len = hdrlen; 773 } 774 m->m_pkthdr.rcvif = (struct ifnet *)0; 775 m->m_pkthdr.len = hdrlen + len; 776 777 if (!tp->t_template) 778 panic("tcp_output"); 779 #ifdef DIAGNOSTIC 780 if (tp->t_template->m_len != hdrlen - optlen) 781 panic("tcp_output: template len != hdrlen - optlen"); 782 #endif /* DIAGNOSTIC */ 783 bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), 784 tp->t_template->m_len); 785 th = (struct tcphdr *)(mtod(m, caddr_t) + tp->t_template->m_len - 786 sizeof(struct tcphdr)); 787 788 /* 789 * Fill in fields, remembering maximum advertised 790 * window for use in delaying messages about window sizes. 791 * If resending a FIN, be sure not to use a new sequence number. 792 */ 793 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) && 794 (tp->snd_nxt == tp->snd_max)) 795 tp->snd_nxt--; 796 /* 797 * If we are doing retransmissions, then snd_nxt will 798 * not reflect the first unsent octet. For ACK only 799 * packets, we do not want the sequence number of the 800 * retransmitted packet, we want the sequence number 801 * of the next unsent octet. So, if there is no data 802 * (and no SYN or FIN), use snd_max instead of snd_nxt 803 * when filling in ti_seq. But if we are in persist 804 * state, snd_max might reflect one byte beyond the 805 * right edge of the window, so use snd_nxt in that 806 * case, since we know we aren't doing a retransmission. 807 * (retransmit and persist are mutually exclusive...) 808 */ 809 if (len || (flags & (TH_SYN|TH_FIN)) || TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) 810 th->th_seq = htonl(tp->snd_nxt); 811 else 812 th->th_seq = htonl(tp->snd_max); 813 814 #ifdef TCP_SACK 815 if (sack_rxmit) { 816 /* 817 * If sendalot was turned on (due to option stuffing), turn it 818 * off. Properly set th_seq field. Advance the ret'x pointer 819 * by len. 820 */ 821 if (sendalot) 822 sendalot = 0; 823 th->th_seq = htonl(p->rxmit); 824 p->rxmit += len; 825 #if defined(TCP_SACK) && defined(TCP_FACK) 826 tp->retran_data += len; 827 #endif /* TCP_FACK */ 828 tcpstat.tcps_sack_rexmits++; 829 tcpstat.tcps_sack_rexmit_bytes += len; 830 } 831 #endif /* TCP_SACK */ 832 833 th->th_ack = htonl(tp->rcv_nxt); 834 if (optlen) { 835 bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen); 836 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 837 } 838 #ifdef TCP_ECN 839 if (tcp_do_ecn) { 840 /* 841 * if we have received congestion experienced segs, 842 * set ECE bit. 843 */ 844 if (tp->t_flags & TF_RCVD_CE) { 845 flags |= TH_ECE; 846 tcpstat.tcps_ecn_sndece++; 847 } 848 if (!(tp->t_flags & TF_DISABLE_ECN)) { 849 /* 850 * if this is a SYN seg, set ECE and CWR. 851 * set only ECE for SYN-ACK if peer supports ECN. 852 */ 853 if ((flags & (TH_SYN|TH_ACK)) == TH_SYN) 854 flags |= (TH_ECE|TH_CWR); 855 else if ((tp->t_flags & TF_ECN_PERMIT) && 856 (flags & (TH_SYN|TH_ACK)) == (TH_SYN|TH_ACK)) 857 flags |= TH_ECE; 858 } 859 /* 860 * if we have reduced the congestion window, notify 861 * the peer by setting CWR bit. 862 */ 863 if ((tp->t_flags & TF_ECN_PERMIT) && 864 (tp->t_flags & TF_SEND_CWR)) { 865 flags |= TH_CWR; 866 tp->t_flags &= ~TF_SEND_CWR; 867 tcpstat.tcps_ecn_sndcwr++; 868 } 869 } 870 #endif 871 th->th_flags = flags; 872 873 /* 874 * Calculate receive window. Don't shrink window, 875 * but avoid silly window syndrome. 876 */ 877 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg) 878 win = 0; 879 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 880 win = (long)TCP_MAXWIN << tp->rcv_scale; 881 if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt)) 882 win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt); 883 if (flags & TH_RST) 884 win = 0; 885 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale)); 886 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 887 u_int32_t urp = tp->snd_up - tp->snd_nxt; 888 if (urp > IP_MAXPACKET) 889 urp = IP_MAXPACKET; 890 th->th_urp = htons((u_int16_t)urp); 891 th->th_flags |= TH_URG; 892 } else 893 /* 894 * If no urgent pointer to send, then we pull 895 * the urgent pointer to the left edge of the send window 896 * so that it doesn't drift into the send window on sequence 897 * number wraparound. 898 */ 899 tp->snd_up = tp->snd_una; /* drag it along */ 900 901 #ifdef TCP_SIGNATURE 902 if (tp->t_flags & TF_SIGNATURE) { 903 int iphlen; 904 union sockaddr_union src, dst; 905 struct tdb *tdb; 906 907 bzero(&src, sizeof(union sockaddr_union)); 908 bzero(&dst, sizeof(union sockaddr_union)); 909 910 switch (tp->pf) { 911 case 0: /*default to PF_INET*/ 912 #ifdef INET 913 case AF_INET: 914 iphlen = sizeof(struct ip); 915 src.sa.sa_len = sizeof(struct sockaddr_in); 916 src.sa.sa_family = AF_INET; 917 src.sin.sin_addr = mtod(m, struct ip *)->ip_src; 918 dst.sa.sa_len = sizeof(struct sockaddr_in); 919 dst.sa.sa_family = AF_INET; 920 dst.sin.sin_addr = mtod(m, struct ip *)->ip_dst; 921 break; 922 #endif /* INET */ 923 #ifdef INET6 924 case AF_INET6: 925 iphlen = sizeof(struct ip6_hdr); 926 src.sa.sa_len = sizeof(struct sockaddr_in6); 927 src.sa.sa_family = AF_INET6; 928 src.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_src; 929 dst.sa.sa_len = sizeof(struct sockaddr_in6); 930 dst.sa.sa_family = AF_INET6; 931 dst.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_dst; 932 break; 933 #endif /* INET6 */ 934 } 935 936 /* XXX gettdbbysrcdst() should really be called at spltdb(). */ 937 /* XXX this is splsoftnet(), currently they are the same. */ 938 tdb = gettdbbysrcdst(rtable_l2(tp->t_inpcb->inp_rtableid), 939 0, &src, &dst, IPPROTO_TCP); 940 if (tdb == NULL) 941 return (EPERM); 942 943 if (tcp_signature(tdb, tp->pf, m, th, iphlen, 0, 944 mtod(m, caddr_t) + hdrlen - optlen + sigoff) < 0) 945 return (EINVAL); 946 } 947 #endif /* TCP_SIGNATURE */ 948 949 /* 950 * Put TCP length in extended header, and then 951 * checksum extended header and data. 952 */ 953 switch (tp->pf) { 954 case 0: /*default to PF_INET*/ 955 #ifdef INET 956 case AF_INET: 957 /* Defer checksumming until later (ip_output() or hardware) */ 958 m->m_pkthdr.csum_flags |= M_TCP_CSUM_OUT; 959 if (len + optlen) 960 th->th_sum = in_cksum_addword(th->th_sum, 961 htons((u_int16_t)(len + optlen))); 962 break; 963 #endif /* INET */ 964 #ifdef INET6 965 case AF_INET6: 966 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 967 hdrlen - sizeof(struct ip6_hdr) + len); 968 break; 969 #endif /* INET6 */ 970 } 971 972 /* 973 * In transmit state, time the transmission and arrange for 974 * the retransmit. In persist state, just set snd_max. 975 */ 976 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 977 tcp_seq startseq = tp->snd_nxt; 978 979 /* 980 * Advance snd_nxt over sequence space of this segment. 981 */ 982 if (flags & (TH_SYN|TH_FIN)) { 983 if (flags & TH_SYN) 984 tp->snd_nxt++; 985 if (flags & TH_FIN) { 986 tp->snd_nxt++; 987 tp->t_flags |= TF_SENTFIN; 988 } 989 } 990 #ifdef TCP_SACK 991 if (tp->sack_enable) { 992 if (sack_rxmit && (p->rxmit != tp->snd_nxt)) { 993 goto timer; 994 } 995 } 996 #endif 997 tp->snd_nxt += len; 998 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 999 tp->snd_max = tp->snd_nxt; 1000 /* 1001 * Time this transmission if not a retransmission and 1002 * not currently timing anything. 1003 */ 1004 if (tp->t_rtttime == 0) { 1005 tp->t_rtttime = tcp_now; 1006 tp->t_rtseq = startseq; 1007 tcpstat.tcps_segstimed++; 1008 } 1009 } 1010 1011 /* 1012 * Set retransmit timer if not currently set, 1013 * and not doing an ack or a keep-alive probe. 1014 * Initial value for retransmit timer is smoothed 1015 * round-trip time + 2 * round-trip time variance. 1016 * Initialize shift counter which is used for backoff 1017 * of retransmit time. 1018 */ 1019 #ifdef TCP_SACK 1020 timer: 1021 if (tp->sack_enable && sack_rxmit && 1022 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1023 tp->snd_nxt != tp->snd_max) { 1024 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1025 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1026 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 1027 tp->t_rxtshift = 0; 1028 } 1029 } 1030 #endif 1031 1032 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1033 tp->snd_nxt != tp->snd_una) { 1034 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1035 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1036 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 1037 tp->t_rxtshift = 0; 1038 } 1039 } 1040 } else 1041 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 1042 tp->snd_max = tp->snd_nxt + len; 1043 1044 tcp_update_sndspace(tp); 1045 1046 /* 1047 * Trace. 1048 */ 1049 if (so->so_options & SO_DEBUG) 1050 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, caddr_t), 0, 1051 len); 1052 1053 /* 1054 * Fill in IP length and desired time to live and 1055 * send to IP level. There should be a better way 1056 * to handle ttl and tos; we could keep them in 1057 * the template, but need a way to checksum without them. 1058 */ 1059 1060 #ifdef TCP_ECN 1061 /* 1062 * if peer is ECN capable, set the ECT bit in the IP header. 1063 * but don't set ECT for a pure ack, a retransmit or a window probe. 1064 */ 1065 needect = 0; 1066 if (tcp_do_ecn && (tp->t_flags & TF_ECN_PERMIT)) { 1067 if (len == 0 || SEQ_LT(tp->snd_nxt, tp->snd_max) || 1068 (tp->t_force && len == 1)) { 1069 /* don't set ECT */ 1070 } else { 1071 needect = 1; 1072 tcpstat.tcps_ecn_sndect++; 1073 } 1074 } 1075 #endif 1076 1077 /* force routing domain */ 1078 m->m_pkthdr.rdomain = tp->t_inpcb->inp_rtableid; 1079 1080 switch (tp->pf) { 1081 case 0: /*default to PF_INET*/ 1082 #ifdef INET 1083 case AF_INET: 1084 { 1085 struct ip *ip; 1086 1087 ip = mtod(m, struct ip *); 1088 ip->ip_len = htons(m->m_pkthdr.len); 1089 packetlen = m->m_pkthdr.len; 1090 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl; 1091 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos; 1092 #ifdef TCP_ECN 1093 if (needect) 1094 ip->ip_tos |= IPTOS_ECN_ECT0; 1095 #endif 1096 } 1097 error = ip_output(m, tp->t_inpcb->inp_options, 1098 &tp->t_inpcb->inp_route, 1099 (ip_mtudisc ? IP_MTUDISC : 0) | 1100 (so->so_options & SO_DONTROUTE), 1101 (void *)NULL, tp->t_inpcb); 1102 break; 1103 #endif /* INET */ 1104 #ifdef INET6 1105 case AF_INET6: 1106 { 1107 struct ip6_hdr *ip6; 1108 1109 ip6 = mtod(m, struct ip6_hdr *); 1110 ip6->ip6_plen = m->m_pkthdr.len - 1111 sizeof(struct ip6_hdr); 1112 packetlen = m->m_pkthdr.len; 1113 ip6->ip6_nxt = IPPROTO_TCP; 1114 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1115 #ifdef TCP_ECN 1116 if (needect) 1117 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1118 #endif 1119 } 1120 error = ip6_output(m, tp->t_inpcb->inp_outputopts6, 1121 &tp->t_inpcb->inp_route6, 1122 (so->so_options & SO_DONTROUTE), NULL, NULL, 1123 tp->t_inpcb); 1124 break; 1125 #endif /* INET6 */ 1126 } 1127 1128 #if defined(TCP_SACK) && defined(TCP_FACK) 1129 /* Update snd_awnd to reflect the new data that was sent. */ 1130 tp->snd_awnd = tcp_seq_subtract(tp->snd_max, tp->snd_fack) + 1131 tp->retran_data; 1132 #endif /* defined(TCP_SACK) && defined(TCP_FACK) */ 1133 1134 if (error) { 1135 out: 1136 if (error == ENOBUFS) { 1137 /* 1138 * If the interface queue is full, or IP cannot 1139 * get an mbuf, trigger TCP slow start. 1140 */ 1141 tp->snd_cwnd = tp->t_maxseg; 1142 return (0); 1143 } 1144 if (error == EMSGSIZE) { 1145 /* 1146 * ip_output() will have already fixed the route 1147 * for us. tcp_mtudisc() will, as its last action, 1148 * initiate retransmission, so it is important to 1149 * not do so here. 1150 */ 1151 tcp_mtudisc(tp->t_inpcb, -1); 1152 return (0); 1153 } 1154 if (error == EACCES) /* translate pf(4) error for userland */ 1155 error = EHOSTUNREACH; 1156 if ((error == EHOSTUNREACH || error == ENETDOWN) && 1157 TCPS_HAVERCVDSYN(tp->t_state)) { 1158 tp->t_softerror = error; 1159 return (0); 1160 } 1161 1162 /* Restart the delayed ACK timer, if necessary. */ 1163 if (tp->t_flags & TF_DELACK) 1164 TCP_RESTART_DELACK(tp); 1165 1166 return (error); 1167 } 1168 1169 if (packetlen > tp->t_pmtud_mtu_sent) 1170 tp->t_pmtud_mtu_sent = packetlen; 1171 1172 tcpstat.tcps_sndtotal++; 1173 if (tp->t_flags & TF_DELACK) 1174 tcpstat.tcps_delack++; 1175 1176 /* 1177 * Data sent (as far as we can tell). 1178 * If this advertises a larger window than any other segment, 1179 * then remember the size of the advertised window. 1180 * Any pending ACK has now been sent. 1181 */ 1182 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv)) 1183 tp->rcv_adv = tp->rcv_nxt + win; 1184 tp->last_ack_sent = tp->rcv_nxt; 1185 tp->t_flags &= ~TF_ACKNOW; 1186 TCP_CLEAR_DELACK(tp); 1187 #if defined(TCP_SACK) 1188 if (sendalot && --maxburst) 1189 #else 1190 if (sendalot) 1191 #endif 1192 goto again; 1193 return (0); 1194 } 1195 1196 void 1197 tcp_setpersist(struct tcpcb *tp) 1198 { 1199 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + TCP_RTT_BASE_SHIFT); 1200 int nticks; 1201 1202 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT)) 1203 panic("tcp_output REXMT"); 1204 /* 1205 * Start/restart persistence timer. 1206 */ 1207 if (t < tp->t_rttmin) 1208 t = tp->t_rttmin; 1209 TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift], 1210 TCPTV_PERSMIN, TCPTV_PERSMAX); 1211 TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks); 1212 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1213 tp->t_rxtshift++; 1214 } 1215