1 /* $OpenBSD: tcp_output.c,v 1.121 2017/06/26 09:32:32 mpi 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 "pf.h" 72 73 #include <sys/param.h> 74 #include <sys/systm.h> 75 #include <sys/mbuf.h> 76 #include <sys/protosw.h> 77 #include <sys/socket.h> 78 #include <sys/socketvar.h> 79 #include <sys/kernel.h> 80 81 #include <net/route.h> 82 83 #include <netinet/in.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/tcp_debug.h> 94 95 #ifdef notyet 96 extern struct mbuf *m_copypack(); 97 #endif 98 99 #ifdef TCP_SACK 100 extern int tcprexmtthresh; 101 #endif 102 103 #ifdef TCP_SACK 104 #ifdef TCP_SACK_DEBUG 105 void tcp_print_holes(struct tcpcb *tp); 106 107 void 108 tcp_print_holes(struct tcpcb *tp) 109 { 110 struct sackhole *p = tp->snd_holes; 111 if (p == NULL) 112 return; 113 printf("Hole report: start--end dups rxmit\n"); 114 while (p) { 115 printf("%x--%x d %d r %x\n", p->start, p->end, p->dups, 116 p->rxmit); 117 p = p->next; 118 } 119 printf("\n"); 120 } 121 #endif /* TCP_SACK_DEBUG */ 122 123 /* 124 * Returns pointer to a sackhole if there are any pending retransmissions; 125 * NULL otherwise. 126 */ 127 struct sackhole * 128 tcp_sack_output(struct tcpcb *tp) 129 { 130 struct sackhole *p; 131 132 if (!tp->sack_enable) 133 return (NULL); 134 p = tp->snd_holes; 135 while (p) { 136 #ifndef TCP_FACK 137 if (p->dups >= tcprexmtthresh && SEQ_LT(p->rxmit, p->end)) { 138 #else 139 /* In FACK, if p->dups is less than tcprexmtthresh, but 140 * snd_fack advances more than tcprextmtthresh * tp->t_maxseg, 141 * tcp_input() will try fast retransmit. This forces output. 142 */ 143 if ((p->dups >= tcprexmtthresh || 144 tp->t_dupacks == tcprexmtthresh) && 145 SEQ_LT(p->rxmit, p->end)) { 146 #endif /* TCP_FACK */ 147 if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */ 148 p = p->next; 149 continue; 150 } 151 #ifdef TCP_SACK_DEBUG 152 if (p) 153 tcp_print_holes(tp); 154 #endif 155 return (p); 156 } 157 p = p->next; 158 } 159 return (NULL); 160 } 161 162 /* 163 * After a timeout, the SACK list may be rebuilt. This SACK information 164 * should be used to avoid retransmitting SACKed data. This function 165 * traverses the SACK list to see if snd_nxt should be moved forward. 166 */ 167 168 void 169 tcp_sack_adjust(struct tcpcb *tp) 170 { 171 struct sackhole *cur = tp->snd_holes; 172 if (cur == NULL) 173 return; /* No holes */ 174 if (SEQ_GEQ(tp->snd_nxt, tp->rcv_lastsack)) 175 return; /* We're already beyond any SACKed blocks */ 176 /* 177 * Two cases for which we want to advance snd_nxt: 178 * i) snd_nxt lies between end of one hole and beginning of another 179 * ii) snd_nxt lies between end of last hole and rcv_lastsack 180 */ 181 while (cur->next) { 182 if (SEQ_LT(tp->snd_nxt, cur->end)) 183 return; 184 if (SEQ_GEQ(tp->snd_nxt, cur->next->start)) 185 cur = cur->next; 186 else { 187 tp->snd_nxt = cur->next->start; 188 return; 189 } 190 } 191 if (SEQ_LT(tp->snd_nxt, cur->end)) 192 return; 193 tp->snd_nxt = tp->rcv_lastsack; 194 return; 195 } 196 #endif /* TCP_SACK */ 197 198 /* 199 * Tcp output routine: figure out what should be sent and send it. 200 */ 201 int 202 tcp_output(struct tcpcb *tp) 203 { 204 struct socket *so = tp->t_inpcb->inp_socket; 205 long len, win, txmaxseg; 206 int off, flags, error; 207 struct mbuf *m; 208 struct tcphdr *th; 209 u_int32_t optbuf[howmany(MAX_TCPOPTLEN, sizeof(u_int32_t))]; 210 u_char *opt = (u_char *)optbuf; 211 unsigned int optlen, hdrlen, packetlen; 212 int idle, sendalot = 0; 213 #ifdef TCP_SACK 214 int i, sack_rxmit = 0; 215 struct sackhole *p; 216 int maxburst = TCP_MAXBURST; 217 #endif 218 #ifdef TCP_SIGNATURE 219 unsigned int sigoff; 220 #endif /* TCP_SIGNATURE */ 221 #ifdef TCP_ECN 222 int needect; 223 #endif 224 225 if (tp->t_flags & TF_BLOCKOUTPUT) { 226 tp->t_flags |= TF_NEEDOUTPUT; 227 return (0); 228 } else 229 tp->t_flags &= ~TF_NEEDOUTPUT; 230 231 #if defined(TCP_SACK) && defined(TCP_SIGNATURE) && defined(DIAGNOSTIC) 232 if (tp->sack_enable && (tp->t_flags & TF_SIGNATURE)) 233 return (EINVAL); 234 #endif /* defined(TCP_SACK) && defined(TCP_SIGNATURE) && defined(DIAGNOSTIC) */ 235 236 /* 237 * Determine length of data that should be transmitted, 238 * and flags that will be used. 239 * If there is some data or critical controls (SYN, RST) 240 * to send, then transmit; otherwise, investigate further. 241 */ 242 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 243 if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) 244 /* 245 * We have been idle for "a while" and no acks are 246 * expected to clock out any data we send -- 247 * slow start to get ack "clock" running again. 248 */ 249 tp->snd_cwnd = 2 * tp->t_maxseg; 250 251 /* remember 'idle' for next invocation of tcp_output */ 252 if (idle && soissending(so)) { 253 tp->t_flags |= TF_LASTIDLE; 254 idle = 0; 255 } else 256 tp->t_flags &= ~TF_LASTIDLE; 257 258 again: 259 #ifdef TCP_SACK 260 /* 261 * If we've recently taken a timeout, snd_max will be greater than 262 * snd_nxt. There may be SACK information that allows us to avoid 263 * resending already delivered data. Adjust snd_nxt accordingly. 264 */ 265 if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max)) 266 tcp_sack_adjust(tp); 267 #endif 268 off = tp->snd_nxt - tp->snd_una; 269 #if defined(TCP_SACK) && defined(TCP_FACK) 270 /* Normally, sendable data is limited by off < tp->snd_cwnd. 271 * But in FACK, sendable data is limited by snd_awnd < snd_cwnd, 272 * regardless of offset. 273 */ 274 if (tp->sack_enable && (tp->t_dupacks > tcprexmtthresh)) 275 win = tp->snd_wnd; 276 else 277 #endif 278 win = ulmin(tp->snd_wnd, tp->snd_cwnd); 279 280 flags = tcp_outflags[tp->t_state]; 281 282 #ifdef TCP_SACK 283 /* 284 * Send any SACK-generated retransmissions. If we're explicitly trying 285 * to send out new data (when sendalot is 1), bypass this function. 286 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 287 * we're replacing a (future) new transmission with a retransmission 288 * now, and we previously incremented snd_cwnd in tcp_input(). 289 */ 290 if (tp->sack_enable && !sendalot) { 291 if (tp->t_dupacks >= tcprexmtthresh && 292 (p = tcp_sack_output(tp))) { 293 off = p->rxmit - tp->snd_una; 294 sack_rxmit = 1; 295 /* Coalesce holes into a single retransmission */ 296 len = min(tp->t_maxseg, p->end - p->rxmit); 297 #ifndef TCP_FACK 298 /* in FACK, hold snd_cwnd constant during recovery */ 299 if (SEQ_LT(tp->snd_una, tp->snd_last)) 300 tp->snd_cwnd -= tp->t_maxseg; 301 #endif 302 } 303 } 304 #endif /* TCP_SACK */ 305 306 sendalot = 0; 307 /* 308 * If in persist timeout with window of 0, send 1 byte. 309 * Otherwise, if window is small but nonzero 310 * and timer expired, we will send what we can 311 * and go to transmit state. 312 */ 313 if (tp->t_force) { 314 if (win == 0) { 315 /* 316 * If we still have some data to send, then 317 * clear the FIN bit. Usually this would 318 * happen below when it realizes that we 319 * aren't sending all the data. However, 320 * if we have exactly 1 byte of unset data, 321 * then it won't clear the FIN bit below, 322 * and if we are in persist state, we wind 323 * up sending the packet without recording 324 * that we sent the FIN bit. 325 * 326 * We can't just blindly clear the FIN bit, 327 * because if we don't have any more data 328 * to send then the probe will be the FIN 329 * itself. 330 */ 331 if (off < so->so_snd.sb_cc) 332 flags &= ~TH_FIN; 333 win = 1; 334 } else { 335 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 336 tp->t_rxtshift = 0; 337 } 338 } 339 340 #ifdef TCP_SACK 341 if (!sack_rxmit) { 342 #endif 343 len = ulmin(so->so_snd.sb_cc, win) - off; 344 345 #if defined(TCP_SACK) && defined(TCP_FACK) 346 /* 347 * If we're in fast recovery (SEQ_GT(tp->snd_last, tp->snd_una)), and 348 * amount of outstanding data (snd_awnd) is >= snd_cwnd, then 349 * do not send data (like zero window conditions) 350 */ 351 if (tp->sack_enable && len && SEQ_GT(tp->snd_last, tp->snd_una) && 352 (tp->snd_awnd >= tp->snd_cwnd)) 353 len = 0; 354 #endif /* TCP_FACK */ 355 #ifdef TCP_SACK 356 } 357 #endif 358 359 if (len < 0) { 360 /* 361 * If FIN has been sent but not acked, 362 * but we haven't been called to retransmit, 363 * len will be -1. Otherwise, window shrank 364 * after we sent into it. If window shrank to 0, 365 * cancel pending retransmit, pull snd_nxt back 366 * to (closed) window, and set the persist timer 367 * if it isn't already going. If the window didn't 368 * close completely, just wait for an ACK. 369 */ 370 len = 0; 371 if (win == 0) { 372 TCP_TIMER_DISARM(tp, TCPT_REXMT); 373 tp->t_rxtshift = 0; 374 tp->snd_nxt = tp->snd_una; 375 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) 376 tcp_setpersist(tp); 377 } 378 } 379 380 /* 381 * Never send more than half a buffer full. This insures that we can 382 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and 383 * therefore acks will never be delayed unless we run out of data to 384 * transmit. 385 */ 386 txmaxseg = ulmin(so->so_snd.sb_hiwat / 2, tp->t_maxseg); 387 388 if (len > txmaxseg) { 389 len = txmaxseg; 390 sendalot = 1; 391 } 392 if (off + len < so->so_snd.sb_cc) 393 flags &= ~TH_FIN; 394 395 win = sbspace(so, &so->so_rcv); 396 397 /* 398 * Sender silly window avoidance. If connection is idle 399 * and can send all data, a maximum segment, 400 * at least a maximum default-size segment do it, 401 * or are forced, do it; otherwise don't bother. 402 * If peer's buffer is tiny, then send 403 * when window is at least half open. 404 * If retransmitting (possibly after persist timer forced us 405 * to send into a small window), then must resend. 406 */ 407 if (len) { 408 if (len == txmaxseg) 409 goto send; 410 if ((idle || (tp->t_flags & TF_NODELAY)) && 411 len + off >= so->so_snd.sb_cc && !soissending(so) && 412 (tp->t_flags & TF_NOPUSH) == 0) 413 goto send; 414 if (tp->t_force) 415 goto send; 416 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 417 goto send; 418 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 419 goto send; 420 #ifdef TCP_SACK 421 if (sack_rxmit) 422 goto send; 423 #endif 424 } 425 426 /* 427 * Compare available window to amount of window 428 * known to peer (as advertised window less 429 * next expected input). If the difference is at least two 430 * max size segments, or at least 50% of the maximum possible 431 * window, then want to send a window update to peer. 432 */ 433 if (win > 0) { 434 /* 435 * "adv" is the amount we can increase the window, 436 * taking into account that we are limited by 437 * TCP_MAXWIN << tp->rcv_scale. 438 */ 439 long adv = lmin(win, (long)TCP_MAXWIN << tp->rcv_scale) - 440 (tp->rcv_adv - tp->rcv_nxt); 441 442 if (adv >= (long) (2 * tp->t_maxseg)) 443 goto send; 444 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 445 goto send; 446 } 447 448 /* 449 * Send if we owe peer an ACK. 450 */ 451 if (tp->t_flags & TF_ACKNOW) 452 goto send; 453 if (flags & (TH_SYN|TH_RST)) 454 goto send; 455 if (SEQ_GT(tp->snd_up, tp->snd_una)) 456 goto send; 457 /* 458 * If our state indicates that FIN should be sent 459 * and we have not yet done so, or we're retransmitting the FIN, 460 * then we need to send. 461 */ 462 if (flags & TH_FIN && 463 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 464 goto send; 465 #ifdef TCP_SACK 466 /* 467 * In SACK, it is possible for tcp_output to fail to send a segment 468 * after the retransmission timer has been turned off. Make sure 469 * that the retransmission timer is set. 470 */ 471 if (SEQ_GT(tp->snd_max, tp->snd_una) && 472 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 473 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 474 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 475 return (0); 476 } 477 #endif /* TCP_SACK */ 478 479 /* 480 * TCP window updates are not reliable, rather a polling protocol 481 * using ``persist'' packets is used to insure receipt of window 482 * updates. The three ``states'' for the output side are: 483 * idle not doing retransmits or persists 484 * persisting to move a small or zero window 485 * (re)transmitting and thereby not persisting 486 * 487 * tp->t_timer[TCPT_PERSIST] 488 * is set when we are in persist state. 489 * tp->t_force 490 * is set when we are called to send a persist packet. 491 * tp->t_timer[TCPT_REXMT] 492 * is set when we are retransmitting 493 * The output side is idle when both timers are zero. 494 * 495 * If send window is too small, there is data to transmit, and no 496 * retransmit or persist is pending, then go to persist state. 497 * If nothing happens soon, send when timer expires: 498 * if window is nonzero, transmit what we can, 499 * otherwise force out a byte. 500 */ 501 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 502 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 503 tp->t_rxtshift = 0; 504 tcp_setpersist(tp); 505 } 506 507 /* 508 * No reason to send a segment, just return. 509 */ 510 return (0); 511 512 send: 513 /* 514 * Before ESTABLISHED, force sending of initial options 515 * unless TCP set not to do any options. 516 * NOTE: we assume that the IP/TCP header plus TCP options 517 * always fit in a single mbuf, leaving room for a maximum 518 * link header, i.e. 519 * max_linkhdr + sizeof(network header) + sizeof(struct tcphdr + 520 * optlen <= MHLEN 521 */ 522 optlen = 0; 523 524 switch (tp->pf) { 525 case 0: /*default to PF_INET*/ 526 case PF_INET: 527 hdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 528 break; 529 #ifdef INET6 530 case PF_INET6: 531 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 532 break; 533 #endif /* INET6 */ 534 default: 535 return (EPFNOSUPPORT); 536 } 537 538 if (flags & TH_SYN) { 539 tp->snd_nxt = tp->iss; 540 if ((tp->t_flags & TF_NOOPT) == 0) { 541 u_int16_t mss; 542 543 opt[0] = TCPOPT_MAXSEG; 544 opt[1] = 4; 545 mss = htons((u_int16_t) tcp_mss(tp, 0)); 546 memcpy(opt + 2, &mss, sizeof(mss)); 547 optlen = 4; 548 549 if (flags & TH_ACK) 550 tcp_mss_update(tp); 551 #ifdef TCP_SACK 552 /* 553 * If this is the first SYN of connection (not a SYN 554 * ACK), include SACK_PERMIT_HDR option. If this is a 555 * SYN ACK, include SACK_PERMIT_HDR option if peer has 556 * already done so. 557 */ 558 if (tp->sack_enable && ((flags & TH_ACK) == 0 || 559 (tp->t_flags & TF_SACK_PERMIT))) { 560 *((u_int32_t *) (opt + optlen)) = 561 htonl(TCPOPT_SACK_PERMIT_HDR); 562 optlen += 4; 563 } 564 #endif 565 566 if ((tp->t_flags & TF_REQ_SCALE) && 567 ((flags & TH_ACK) == 0 || 568 (tp->t_flags & TF_RCVD_SCALE))) { 569 *((u_int32_t *) (opt + optlen)) = htonl( 570 TCPOPT_NOP << 24 | 571 TCPOPT_WINDOW << 16 | 572 TCPOLEN_WINDOW << 8 | 573 tp->request_r_scale); 574 optlen += 4; 575 } 576 } 577 } 578 579 /* 580 * Send a timestamp and echo-reply if this is a SYN and our side 581 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 582 * and our peer have sent timestamps in our SYN's. 583 */ 584 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 585 (flags & TH_RST) == 0 && 586 ((flags & (TH_SYN|TH_ACK)) == TH_SYN || 587 (tp->t_flags & TF_RCVD_TSTMP))) { 588 u_int32_t *lp = (u_int32_t *)(opt + optlen); 589 590 /* Form timestamp option as shown in appendix A of RFC 1323. */ 591 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 592 *lp++ = htonl(tcp_now + tp->ts_modulate); 593 *lp = htonl(tp->ts_recent); 594 optlen += TCPOLEN_TSTAMP_APPA; 595 596 /* Set receive buffer autosizing timestamp. */ 597 if (tp->rfbuf_ts == 0) 598 tp->rfbuf_ts = tcp_now; 599 600 } 601 602 #ifdef TCP_SIGNATURE 603 if (tp->t_flags & TF_SIGNATURE) { 604 u_int8_t *bp = (u_int8_t *)(opt + optlen); 605 606 /* Send signature option */ 607 *(bp++) = TCPOPT_SIGNATURE; 608 *(bp++) = TCPOLEN_SIGNATURE; 609 sigoff = optlen + 2; 610 611 { 612 unsigned int i; 613 614 for (i = 0; i < 16; i++) 615 *(bp++) = 0; 616 } 617 618 619 /* Pad options list to the next 32 bit boundary and 620 * terminate it. 621 */ 622 *bp++ = TCPOPT_NOP; 623 *bp++ = TCPOPT_NOP; 624 625 optlen += TCPOLEN_SIGLEN; 626 } 627 #endif /* TCP_SIGNATURE */ 628 629 #ifdef TCP_SACK 630 /* 631 * Send SACKs if necessary. This should be the last option processed. 632 * Only as many SACKs are sent as are permitted by the maximum options 633 * size. No more than three SACKs are sent. 634 */ 635 if (tp->sack_enable && tp->t_state == TCPS_ESTABLISHED && 636 (tp->t_flags & (TF_SACK_PERMIT|TF_NOOPT)) == TF_SACK_PERMIT && 637 tp->rcv_numsacks) { 638 u_int32_t *lp = (u_int32_t *)(opt + optlen); 639 u_int32_t *olp = lp++; 640 int count = 0; /* actual number of SACKs inserted */ 641 int maxsack = (MAX_TCPOPTLEN - (optlen + 4))/TCPOLEN_SACK; 642 643 tcpstat_inc(tcps_sack_snd_opts); 644 maxsack = min(maxsack, TCP_MAX_SACK); 645 for (i = 0; (i < tp->rcv_numsacks && count < maxsack); i++) { 646 struct sackblk sack = tp->sackblks[i]; 647 if (sack.start == 0 && sack.end == 0) 648 continue; 649 *lp++ = htonl(sack.start); 650 *lp++ = htonl(sack.end); 651 count++; 652 } 653 *olp = htonl(TCPOPT_SACK_HDR|(TCPOLEN_SACK*count+2)); 654 optlen += TCPOLEN_SACK*count + 4; /* including leading NOPs */ 655 } 656 #endif /* TCP_SACK */ 657 658 #ifdef DIAGNOSTIC 659 if (optlen > MAX_TCPOPTLEN) 660 panic("tcp_output: options too long"); 661 #endif /* DIAGNOSTIC */ 662 663 hdrlen += optlen; 664 665 /* 666 * Adjust data length if insertion of options will 667 * bump the packet length beyond the t_maxopd length. 668 */ 669 if (len > tp->t_maxopd - optlen) { 670 len = tp->t_maxopd - optlen; 671 sendalot = 1; 672 flags &= ~TH_FIN; 673 } 674 675 #ifdef DIAGNOSTIC 676 if (max_linkhdr + hdrlen > MCLBYTES) 677 panic("tcphdr too big"); 678 #endif 679 680 /* 681 * Grab a header mbuf, attaching a copy of data to 682 * be transmitted, and initialize the header from 683 * the template for sends on this connection. 684 */ 685 if (len) { 686 if (tp->t_force && len == 1) 687 tcpstat_inc(tcps_sndprobe); 688 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 689 tcpstat_pkt(tcps_sndrexmitpack, tcps_sndrexmitbyte, 690 len); 691 } else { 692 tcpstat_pkt(tcps_sndpack, tcps_sndbyte, len); 693 } 694 #ifdef notyet 695 if ((m = m_copypack(so->so_snd.sb_mb, off, 696 (int)len, max_linkhdr + hdrlen)) == 0) { 697 error = ENOBUFS; 698 goto out; 699 } 700 /* 701 * m_copypack left space for our hdr; use it. 702 */ 703 m->m_len += hdrlen; 704 m->m_data -= hdrlen; 705 #else 706 MGETHDR(m, M_DONTWAIT, MT_HEADER); 707 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 708 MCLGET(m, M_DONTWAIT); 709 if ((m->m_flags & M_EXT) == 0) { 710 m_freem(m); 711 m = NULL; 712 } 713 } 714 if (m == NULL) { 715 error = ENOBUFS; 716 goto out; 717 } 718 m->m_data += max_linkhdr; 719 m->m_len = hdrlen; 720 if (len <= M_TRAILINGSPACE(m)) { 721 m_copydata(so->so_snd.sb_mb, off, (int) len, 722 mtod(m, caddr_t) + hdrlen); 723 m->m_len += len; 724 } else { 725 m->m_next = m_copym(so->so_snd.sb_mb, off, (int) len, 726 M_NOWAIT); 727 if (m->m_next == 0) { 728 (void) m_free(m); 729 error = ENOBUFS; 730 goto out; 731 } 732 } 733 if (so->so_snd.sb_mb->m_flags & M_PKTHDR) 734 m->m_pkthdr.ph_loopcnt = 735 so->so_snd.sb_mb->m_pkthdr.ph_loopcnt; 736 #endif 737 /* 738 * If we're sending everything we've got, set PUSH. 739 * (This will keep happy those implementations which only 740 * give data to the user when a buffer fills or 741 * a PUSH comes in.) 742 */ 743 if (off + len == so->so_snd.sb_cc && !soissending(so)) 744 flags |= TH_PUSH; 745 } else { 746 if (tp->t_flags & TF_ACKNOW) 747 tcpstat_inc(tcps_sndacks); 748 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 749 tcpstat_inc(tcps_sndctrl); 750 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 751 tcpstat_inc(tcps_sndurg); 752 else 753 tcpstat_inc(tcps_sndwinup); 754 755 MGETHDR(m, M_DONTWAIT, MT_HEADER); 756 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 757 MCLGET(m, M_DONTWAIT); 758 if ((m->m_flags & M_EXT) == 0) { 759 m_freem(m); 760 m = NULL; 761 } 762 } 763 if (m == NULL) { 764 error = ENOBUFS; 765 goto out; 766 } 767 m->m_data += max_linkhdr; 768 m->m_len = hdrlen; 769 } 770 m->m_pkthdr.ph_ifidx = 0; 771 m->m_pkthdr.len = hdrlen + len; 772 773 if (!tp->t_template) 774 panic("tcp_output"); 775 #ifdef DIAGNOSTIC 776 if (tp->t_template->m_len != hdrlen - optlen) 777 panic("tcp_output: template len != hdrlen - optlen"); 778 #endif /* DIAGNOSTIC */ 779 memcpy(mtod(m, caddr_t), mtod(tp->t_template, caddr_t), 780 tp->t_template->m_len); 781 th = (struct tcphdr *)(mtod(m, caddr_t) + tp->t_template->m_len - 782 sizeof(struct tcphdr)); 783 784 /* 785 * Fill in fields, remembering maximum advertised 786 * window for use in delaying messages about window sizes. 787 * If resending a FIN, be sure not to use a new sequence number. 788 */ 789 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) && 790 (tp->snd_nxt == tp->snd_max)) 791 tp->snd_nxt--; 792 /* 793 * If we are doing retransmissions, then snd_nxt will 794 * not reflect the first unsent octet. For ACK only 795 * packets, we do not want the sequence number of the 796 * retransmitted packet, we want the sequence number 797 * of the next unsent octet. So, if there is no data 798 * (and no SYN or FIN), use snd_max instead of snd_nxt 799 * when filling in ti_seq. But if we are in persist 800 * state, snd_max might reflect one byte beyond the 801 * right edge of the window, so use snd_nxt in that 802 * case, since we know we aren't doing a retransmission. 803 * (retransmit and persist are mutually exclusive...) 804 */ 805 if (len || (flags & (TH_SYN|TH_FIN)) || TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) 806 th->th_seq = htonl(tp->snd_nxt); 807 else 808 th->th_seq = htonl(tp->snd_max); 809 810 #ifdef TCP_SACK 811 if (sack_rxmit) { 812 /* 813 * If sendalot was turned on (due to option stuffing), turn it 814 * off. Properly set th_seq field. Advance the ret'x pointer 815 * by len. 816 */ 817 if (sendalot) 818 sendalot = 0; 819 th->th_seq = htonl(p->rxmit); 820 p->rxmit += len; 821 #if defined(TCP_SACK) && defined(TCP_FACK) 822 tp->retran_data += len; 823 #endif /* TCP_FACK */ 824 tcpstat_pkt(tcps_sack_rexmits, tcps_sack_rexmit_bytes, len); 825 } 826 #endif /* TCP_SACK */ 827 828 th->th_ack = htonl(tp->rcv_nxt); 829 if (optlen) { 830 memcpy(th + 1, opt, optlen); 831 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 832 } 833 #ifdef TCP_ECN 834 if (tcp_do_ecn) { 835 /* 836 * if we have received congestion experienced segs, 837 * set ECE bit. 838 */ 839 if (tp->t_flags & TF_RCVD_CE) { 840 flags |= TH_ECE; 841 tcpstat_inc(tcps_ecn_sndece); 842 } 843 if (!(tp->t_flags & TF_DISABLE_ECN)) { 844 /* 845 * if this is a SYN seg, set ECE and CWR. 846 * set only ECE for SYN-ACK if peer supports ECN. 847 */ 848 if ((flags & (TH_SYN|TH_ACK)) == TH_SYN) 849 flags |= (TH_ECE|TH_CWR); 850 else if ((tp->t_flags & TF_ECN_PERMIT) && 851 (flags & (TH_SYN|TH_ACK)) == (TH_SYN|TH_ACK)) 852 flags |= TH_ECE; 853 } 854 /* 855 * if we have reduced the congestion window, notify 856 * the peer by setting CWR bit. 857 */ 858 if ((tp->t_flags & TF_ECN_PERMIT) && 859 (tp->t_flags & TF_SEND_CWR)) { 860 flags |= TH_CWR; 861 tp->t_flags &= ~TF_SEND_CWR; 862 tcpstat_inc(tcps_ecn_sndcwr); 863 } 864 } 865 #endif 866 th->th_flags = flags; 867 868 /* 869 * Calculate receive window. Don't shrink window, 870 * but avoid silly window syndrome. 871 */ 872 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg) 873 win = 0; 874 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 875 win = (long)TCP_MAXWIN << tp->rcv_scale; 876 if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt)) 877 win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt); 878 if (flags & TH_RST) 879 win = 0; 880 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale)); 881 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 882 u_int32_t urp = tp->snd_up - tp->snd_nxt; 883 if (urp > IP_MAXPACKET) 884 urp = IP_MAXPACKET; 885 th->th_urp = htons((u_int16_t)urp); 886 th->th_flags |= TH_URG; 887 } else 888 /* 889 * If no urgent pointer to send, then we pull 890 * the urgent pointer to the left edge of the send window 891 * so that it doesn't drift into the send window on sequence 892 * number wraparound. 893 */ 894 tp->snd_up = tp->snd_una; /* drag it along */ 895 896 #ifdef TCP_SIGNATURE 897 if (tp->t_flags & TF_SIGNATURE) { 898 int iphlen; 899 union sockaddr_union src, dst; 900 struct tdb *tdb; 901 902 bzero(&src, sizeof(union sockaddr_union)); 903 bzero(&dst, sizeof(union sockaddr_union)); 904 905 switch (tp->pf) { 906 case 0: /*default to PF_INET*/ 907 case AF_INET: 908 iphlen = sizeof(struct ip); 909 src.sa.sa_len = sizeof(struct sockaddr_in); 910 src.sa.sa_family = AF_INET; 911 src.sin.sin_addr = mtod(m, struct ip *)->ip_src; 912 dst.sa.sa_len = sizeof(struct sockaddr_in); 913 dst.sa.sa_family = AF_INET; 914 dst.sin.sin_addr = mtod(m, struct ip *)->ip_dst; 915 break; 916 #ifdef INET6 917 case AF_INET6: 918 iphlen = sizeof(struct ip6_hdr); 919 src.sa.sa_len = sizeof(struct sockaddr_in6); 920 src.sa.sa_family = AF_INET6; 921 src.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_src; 922 dst.sa.sa_len = sizeof(struct sockaddr_in6); 923 dst.sa.sa_family = AF_INET6; 924 dst.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_dst; 925 break; 926 #endif /* INET6 */ 927 } 928 929 tdb = gettdbbysrcdst(rtable_l2(tp->t_inpcb->inp_rtableid), 930 0, &src, &dst, IPPROTO_TCP); 931 if (tdb == NULL) { 932 m_freem(m); 933 return (EPERM); 934 } 935 936 if (tcp_signature(tdb, tp->pf, m, th, iphlen, 0, 937 mtod(m, caddr_t) + hdrlen - optlen + sigoff) < 0) { 938 m_freem(m); 939 return (EINVAL); 940 } 941 } 942 #endif /* TCP_SIGNATURE */ 943 944 /* Defer checksumming until later (ip_output() or hardware) */ 945 m->m_pkthdr.csum_flags |= M_TCP_CSUM_OUT; 946 947 /* 948 * In transmit state, time the transmission and arrange for 949 * the retransmit. In persist state, just set snd_max. 950 */ 951 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 952 tcp_seq startseq = tp->snd_nxt; 953 954 /* 955 * Advance snd_nxt over sequence space of this segment. 956 */ 957 if (flags & (TH_SYN|TH_FIN)) { 958 if (flags & TH_SYN) 959 tp->snd_nxt++; 960 if (flags & TH_FIN) { 961 tp->snd_nxt++; 962 tp->t_flags |= TF_SENTFIN; 963 } 964 } 965 #ifdef TCP_SACK 966 if (tp->sack_enable) { 967 if (sack_rxmit && (p->rxmit != tp->snd_nxt)) { 968 goto timer; 969 } 970 } 971 #endif 972 tp->snd_nxt += len; 973 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 974 tp->snd_max = tp->snd_nxt; 975 /* 976 * Time this transmission if not a retransmission and 977 * not currently timing anything. 978 */ 979 if (tp->t_rtttime == 0) { 980 tp->t_rtttime = tcp_now; 981 tp->t_rtseq = startseq; 982 tcpstat_inc(tcps_segstimed); 983 } 984 } 985 986 /* 987 * Set retransmit timer if not currently set, 988 * and not doing an ack or a keep-alive probe. 989 * Initial value for retransmit timer is smoothed 990 * round-trip time + 2 * round-trip time variance. 991 * Initialize shift counter which is used for backoff 992 * of retransmit time. 993 */ 994 #ifdef TCP_SACK 995 timer: 996 if (tp->sack_enable && sack_rxmit && 997 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 998 tp->snd_nxt != tp->snd_max) { 999 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1000 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1001 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 1002 tp->t_rxtshift = 0; 1003 } 1004 } 1005 #endif 1006 1007 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1008 tp->snd_nxt != tp->snd_una) { 1009 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1010 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1011 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 1012 tp->t_rxtshift = 0; 1013 } 1014 } 1015 1016 if (len == 0 && so->so_snd.sb_cc && 1017 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1018 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1019 /* 1020 * Avoid a situation where we do not set persist timer 1021 * after a zero window condition. For example: 1022 * 1) A -> B: packet with enough data to fill the window 1023 * 2) B -> A: ACK for #1 + new data (0 window 1024 * advertisement) 1025 * 3) A -> B: ACK for #2, 0 len packet 1026 * 1027 * In this case, A will not activate the persist timer, 1028 * because it chose to send a packet. Unless tcp_output 1029 * is called for some other reason (delayed ack timer, 1030 * another input packet from B, socket syscall), A will 1031 * not send zero window probes. 1032 * 1033 * So, if you send a 0-length packet, but there is data 1034 * in the socket buffer, and neither the rexmt or 1035 * persist timer is already set, then activate the 1036 * persist timer. 1037 */ 1038 tp->t_rxtshift = 0; 1039 tcp_setpersist(tp); 1040 } 1041 } else 1042 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 1043 tp->snd_max = tp->snd_nxt + len; 1044 1045 tcp_update_sndspace(tp); 1046 1047 /* 1048 * Trace. 1049 */ 1050 if (so->so_options & SO_DEBUG) 1051 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, caddr_t), 0, 1052 len); 1053 1054 /* 1055 * Fill in IP length and desired time to live and 1056 * send to IP level. There should be a better way 1057 * to handle ttl and tos; we could keep them in 1058 * the template, but need a way to checksum without them. 1059 */ 1060 1061 #ifdef TCP_ECN 1062 /* 1063 * if peer is ECN capable, set the ECT bit in the IP header. 1064 * but don't set ECT for a pure ack, a retransmit or a window probe. 1065 */ 1066 needect = 0; 1067 if (tcp_do_ecn && (tp->t_flags & TF_ECN_PERMIT)) { 1068 if (len == 0 || SEQ_LT(tp->snd_nxt, tp->snd_max) || 1069 (tp->t_force && len == 1)) { 1070 /* don't set ECT */ 1071 } else { 1072 needect = 1; 1073 tcpstat_inc(tcps_ecn_sndect); 1074 } 1075 } 1076 #endif 1077 1078 /* force routing table */ 1079 m->m_pkthdr.ph_rtableid = tp->t_inpcb->inp_rtableid; 1080 1081 #if NPF > 0 1082 m->m_pkthdr.pf.inp = tp->t_inpcb; 1083 #endif 1084 1085 switch (tp->pf) { 1086 case 0: /*default to PF_INET*/ 1087 case AF_INET: 1088 { 1089 struct ip *ip; 1090 1091 ip = mtod(m, struct ip *); 1092 ip->ip_len = htons(m->m_pkthdr.len); 1093 packetlen = m->m_pkthdr.len; 1094 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl; 1095 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos; 1096 #ifdef TCP_ECN 1097 if (needect) 1098 ip->ip_tos |= IPTOS_ECN_ECT0; 1099 #endif 1100 } 1101 error = ip_output(m, tp->t_inpcb->inp_options, 1102 &tp->t_inpcb->inp_route, 1103 (ip_mtudisc ? IP_MTUDISC : 0), NULL, tp->t_inpcb, 0); 1104 break; 1105 #ifdef INET6 1106 case AF_INET6: 1107 { 1108 struct ip6_hdr *ip6; 1109 1110 ip6 = mtod(m, struct ip6_hdr *); 1111 ip6->ip6_plen = m->m_pkthdr.len - 1112 sizeof(struct ip6_hdr); 1113 packetlen = m->m_pkthdr.len; 1114 ip6->ip6_nxt = IPPROTO_TCP; 1115 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb); 1116 #ifdef TCP_ECN 1117 if (needect) 1118 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1119 #endif 1120 } 1121 error = ip6_output(m, tp->t_inpcb->inp_outputopts6, 1122 &tp->t_inpcb->inp_route6, 1123 0, NULL, 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_inc(tcps_sndtotal); 1173 if (tp->t_flags & TF_DELACK) 1174 tcpstat_inc(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