1 /* $NetBSD: tcp_output.c,v 1.88 2002/11/24 10:51:56 scw Exp $ */ 2 3 /* 4 * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Neither the name of the project nor the names of its contributors 16 * may be used to endorse or promote products derived from this software 17 * without specific prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 */ 31 32 /* 33 * @(#)COPYRIGHT 1.1 (NRL) 17 January 1995 34 * 35 * NRL grants permission for redistribution and use in source and binary 36 * forms, with or without modification, of the software and documentation 37 * created at NRL provided that the following conditions are met: 38 * 39 * 1. Redistributions of source code must retain the above copyright 40 * notice, this list of conditions and the following disclaimer. 41 * 2. Redistributions in binary form must reproduce the above copyright 42 * notice, this list of conditions and the following disclaimer in the 43 * documentation and/or other materials provided with the distribution. 44 * 3. All advertising materials mentioning features or use of this software 45 * must display the following acknowledgements: 46 * This product includes software developed by the University of 47 * California, Berkeley and its contributors. 48 * This product includes software developed at the Information 49 * Technology Division, US Naval Research Laboratory. 50 * 4. Neither the name of the NRL nor the names of its contributors 51 * may be used to endorse or promote products derived from this software 52 * without specific prior written permission. 53 * 54 * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS 55 * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 56 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A 57 * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NRL OR 58 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 59 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 60 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 61 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 62 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 63 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 64 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 65 * 66 * The views and conclusions contained in the software and documentation 67 * are those of the authors and should not be interpreted as representing 68 * official policies, either expressed or implied, of the US Naval 69 * Research Laboratory (NRL). 70 */ 71 72 /*- 73 * Copyright (c) 1997, 1998, 2001 The NetBSD Foundation, Inc. 74 * All rights reserved. 75 * 76 * This code is derived from software contributed to The NetBSD Foundation 77 * by Jason R. Thorpe and Kevin M. Lahey of the Numerical Aerospace Simulation 78 * Facility, NASA Ames Research Center. 79 * 80 * Redistribution and use in source and binary forms, with or without 81 * modification, are permitted provided that the following conditions 82 * are met: 83 * 1. Redistributions of source code must retain the above copyright 84 * notice, this list of conditions and the following disclaimer. 85 * 2. Redistributions in binary form must reproduce the above copyright 86 * notice, this list of conditions and the following disclaimer in the 87 * documentation and/or other materials provided with the distribution. 88 * 3. All advertising materials mentioning features or use of this software 89 * must display the following acknowledgement: 90 * This product includes software developed by the NetBSD 91 * Foundation, Inc. and its contributors. 92 * 4. Neither the name of The NetBSD Foundation nor the names of its 93 * contributors may be used to endorse or promote products derived 94 * from this software without specific prior written permission. 95 * 96 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 97 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 98 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 99 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 100 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 101 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 102 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 103 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 104 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 105 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 106 * POSSIBILITY OF SUCH DAMAGE. 107 */ 108 109 /* 110 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 111 * The Regents of the University of California. All rights reserved. 112 * 113 * Redistribution and use in source and binary forms, with or without 114 * modification, are permitted provided that the following conditions 115 * are met: 116 * 1. Redistributions of source code must retain the above copyright 117 * notice, this list of conditions and the following disclaimer. 118 * 2. Redistributions in binary form must reproduce the above copyright 119 * notice, this list of conditions and the following disclaimer in the 120 * documentation and/or other materials provided with the distribution. 121 * 3. All advertising materials mentioning features or use of this software 122 * must display the following acknowledgement: 123 * This product includes software developed by the University of 124 * California, Berkeley and its contributors. 125 * 4. Neither the name of the University nor the names of its contributors 126 * may be used to endorse or promote products derived from this software 127 * without specific prior written permission. 128 * 129 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 130 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 131 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 132 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 133 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 134 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 135 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 136 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 137 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 138 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 139 * SUCH DAMAGE. 140 * 141 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 142 */ 143 144 #include <sys/cdefs.h> 145 __KERNEL_RCSID(0, "$NetBSD: tcp_output.c,v 1.88 2002/11/24 10:51:56 scw Exp $"); 146 147 #include "opt_inet.h" 148 #include "opt_ipsec.h" 149 #include "opt_tcp_debug.h" 150 151 #include <sys/param.h> 152 #include <sys/systm.h> 153 #include <sys/malloc.h> 154 #include <sys/mbuf.h> 155 #include <sys/protosw.h> 156 #include <sys/socket.h> 157 #include <sys/socketvar.h> 158 #include <sys/errno.h> 159 #include <sys/domain.h> 160 #include <sys/kernel.h> 161 162 #include <net/if.h> 163 #include <net/route.h> 164 165 #include <netinet/in.h> 166 #include <netinet/in_systm.h> 167 #include <netinet/ip.h> 168 #include <netinet/in_pcb.h> 169 #include <netinet/ip_var.h> 170 171 #ifdef INET6 172 #ifndef INET 173 #include <netinet/in.h> 174 #endif 175 #include <netinet/ip6.h> 176 #include <netinet6/in6_var.h> 177 #include <netinet6/ip6_var.h> 178 #include <netinet6/in6_pcb.h> 179 #include <netinet6/nd6.h> 180 #endif 181 182 #ifdef IPSEC 183 #include <netinet6/ipsec.h> 184 #endif 185 186 #include <netinet/tcp.h> 187 #define TCPOUTFLAGS 188 #include <netinet/tcp_fsm.h> 189 #include <netinet/tcp_seq.h> 190 #include <netinet/tcp_timer.h> 191 #include <netinet/tcp_var.h> 192 #include <netinet/tcpip.h> 193 #include <netinet/tcp_debug.h> 194 195 #ifdef notyet 196 extern struct mbuf *m_copypack(); 197 #endif 198 199 #define MAX_TCPOPTLEN 32 /* max # bytes that go in options */ 200 201 /* 202 * Knob to enable Congestion Window Monitoring, and control the 203 * the burst size it allows. Default burst is 4 packets, per 204 * the Internet draft. 205 */ 206 int tcp_cwm = 0; 207 int tcp_cwm_burstsize = 4; 208 209 #ifdef TCP_OUTPUT_COUNTERS 210 #include <sys/device.h> 211 212 extern struct evcnt tcp_output_bigheader; 213 extern struct evcnt tcp_output_copysmall; 214 extern struct evcnt tcp_output_copybig; 215 extern struct evcnt tcp_output_refbig; 216 217 #define TCP_OUTPUT_COUNTER_INCR(ev) (ev)->ev_count++ 218 #else 219 220 #define TCP_OUTPUT_COUNTER_INCR(ev) /* nothing */ 221 222 #endif /* TCP_OUTPUT_COUNTERS */ 223 224 static 225 #ifndef GPROF 226 __inline 227 #endif 228 void 229 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep) 230 { 231 #ifdef INET 232 struct inpcb *inp = tp->t_inpcb; 233 #endif 234 #ifdef INET6 235 struct in6pcb *in6p = tp->t_in6pcb; 236 #endif 237 struct socket *so = NULL; 238 struct rtentry *rt; 239 struct ifnet *ifp; 240 int size; 241 int iphlen; 242 int optlen; 243 244 #ifdef DIAGNOSTIC 245 if (tp->t_inpcb && tp->t_in6pcb) 246 panic("tcp_segsize: both t_inpcb and t_in6pcb are set"); 247 #endif 248 switch (tp->t_family) { 249 #ifdef INET 250 case AF_INET: 251 iphlen = sizeof(struct ip); 252 break; 253 #endif 254 #ifdef INET6 255 case AF_INET6: 256 iphlen = sizeof(struct ip6_hdr); 257 break; 258 #endif 259 default: 260 size = tcp_mssdflt; 261 goto out; 262 } 263 264 rt = NULL; 265 #ifdef INET 266 if (inp) { 267 rt = in_pcbrtentry(inp); 268 so = inp->inp_socket; 269 } 270 #endif 271 #ifdef INET6 272 if (in6p) { 273 rt = in6_pcbrtentry(in6p); 274 so = in6p->in6p_socket; 275 } 276 #endif 277 if (rt == NULL) { 278 size = tcp_mssdflt; 279 goto out; 280 } 281 282 ifp = rt->rt_ifp; 283 284 size = tcp_mssdflt; 285 if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) 286 size = rt->rt_rmx.rmx_mtu - iphlen - sizeof(struct tcphdr); 287 else if (ifp->if_flags & IFF_LOOPBACK) 288 size = ifp->if_mtu - iphlen - sizeof(struct tcphdr); 289 #ifdef INET 290 else if (inp && tp->t_mtudisc) 291 size = ifp->if_mtu - iphlen - sizeof(struct tcphdr); 292 else if (inp && in_localaddr(inp->inp_faddr)) 293 size = ifp->if_mtu - iphlen - sizeof(struct tcphdr); 294 #endif 295 #ifdef INET6 296 else if (in6p) { 297 #ifdef INET 298 if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) { 299 /* mapped addr case */ 300 struct in_addr d; 301 bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d)); 302 if (tp->t_mtudisc || in_localaddr(d)) 303 size = ifp->if_mtu - iphlen - sizeof(struct tcphdr); 304 } else 305 #endif 306 { 307 /* 308 * for IPv6, path MTU discovery is always turned on, 309 * or the node must use packet size <= 1280. 310 */ 311 size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU; 312 size -= (iphlen + sizeof(struct tcphdr)); 313 } 314 } 315 #endif 316 out: 317 /* 318 * Now we must make room for whatever extra TCP/IP options are in 319 * the packet. 320 */ 321 optlen = tcp_optlen(tp); 322 323 /* 324 * XXX tp->t_ourmss should have the right size, but without this code 325 * fragmentation will occur... need more investigation 326 */ 327 #ifdef INET 328 if (inp) { 329 #ifdef IPSEC 330 optlen += ipsec4_hdrsiz_tcp(tp); 331 #endif 332 optlen += ip_optlen(inp); 333 } 334 #endif 335 #ifdef INET6 336 #ifdef INET 337 if (in6p && tp->t_family == AF_INET) { 338 #ifdef IPSEC 339 optlen += ipsec4_hdrsiz_tcp(tp); 340 #endif 341 /* XXX size -= ip_optlen(in6p); */ 342 } else 343 #endif 344 if (in6p && tp->t_family == AF_INET6) { 345 #ifdef IPSEC 346 optlen += ipsec6_hdrsiz_tcp(tp); 347 #endif 348 optlen += ip6_optlen(in6p); 349 } 350 #endif 351 size -= optlen; 352 353 /* 354 * *rxsegsizep holds *estimated* inbound segment size (estimation 355 * assumes that path MTU is the same for both ways). this is only 356 * for silly window avoidance, do not use the value for other purposes. 357 * 358 * ipseclen is subtracted from both sides, this may not be right. 359 * I'm not quite sure about this (could someone comment). 360 */ 361 *txsegsizep = min(tp->t_peermss - optlen, size); 362 /* 363 * Never send more than half a buffer full. This insures that we can 364 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and 365 * therefore acks will never be delayed unless we run out of data to 366 * transmit. 367 */ 368 if (so) 369 *txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep); 370 *rxsegsizep = min(tp->t_ourmss - optlen, size); 371 372 if (*txsegsizep != tp->t_segsz) { 373 /* 374 * If the new segment size is larger, we don't want to 375 * mess up the congestion window, but if it is smaller 376 * we'll have to reduce the congestion window to ensure 377 * that we don't get into trouble with initial windows 378 * and the rest. In any case, if the segment size 379 * has changed, chances are the path has, too, and 380 * our congestion window will be different. 381 */ 382 if (*txsegsizep < tp->t_segsz) { 383 tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz) 384 * *txsegsizep, *txsegsizep); 385 tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz) 386 * *txsegsizep, *txsegsizep); 387 } 388 tp->t_segsz = *txsegsizep; 389 } 390 } 391 392 static 393 #ifndef GPROF 394 __inline 395 #endif 396 int 397 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off, 398 long len, int hdrlen, struct mbuf **mp) 399 { 400 struct mbuf *m; 401 402 if (tp->t_force && len == 1) 403 tcpstat.tcps_sndprobe++; 404 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 405 tcpstat.tcps_sndrexmitpack++; 406 tcpstat.tcps_sndrexmitbyte += len; 407 } else { 408 tcpstat.tcps_sndpack++; 409 tcpstat.tcps_sndbyte += len; 410 } 411 #ifdef notyet 412 if ((m = m_copypack(so->so_snd.sb_mb, off, 413 (int)len, max_linkhdr + hdrlen)) == 0) 414 return (ENOBUFS); 415 /* 416 * m_copypack left space for our hdr; use it. 417 */ 418 m->m_len += hdrlen; 419 m->m_data -= hdrlen; 420 #else 421 MGETHDR(m, M_DONTWAIT, MT_HEADER); 422 if (__predict_false(m == NULL)) 423 return (ENOBUFS); 424 425 /* 426 * XXX Because other code assumes headers will fit in 427 * XXX one header mbuf. 428 * 429 * (This code should almost *never* be run.) 430 */ 431 if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) { 432 TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader); 433 MCLGET(m, M_DONTWAIT); 434 if ((m->m_flags & M_EXT) == 0) { 435 m_freem(m); 436 return (ENOBUFS); 437 } 438 } 439 440 m->m_data += max_linkhdr; 441 m->m_len = hdrlen; 442 if (len <= M_TRAILINGSPACE(m)) { 443 m_copydata(so->so_snd.sb_mb, off, (int) len, 444 mtod(m, caddr_t) + hdrlen); 445 m->m_len += len; 446 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall); 447 } else { 448 m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len); 449 if (m->m_next == NULL) { 450 m_freem(m); 451 return (ENOBUFS); 452 } 453 #ifdef TCP_OUTPUT_COUNTERS 454 if (m->m_next->m_flags & M_EXT) 455 TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig); 456 else 457 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig); 458 #endif /* TCP_OUTPUT_COUNTERS */ 459 } 460 #endif 461 462 *mp = m; 463 return (0); 464 } 465 466 /* 467 * Tcp output routine: figure out what should be sent and send it. 468 */ 469 int 470 tcp_output(tp) 471 struct tcpcb *tp; 472 { 473 struct socket *so; 474 struct route *ro; 475 long len, win; 476 int off, flags, error; 477 struct mbuf *m; 478 struct ip *ip; 479 #ifdef INET6 480 struct ip6_hdr *ip6; 481 #endif 482 struct tcphdr *th; 483 u_char opt[MAX_TCPOPTLEN]; 484 unsigned optlen, hdrlen; 485 int idle, sendalot, txsegsize, rxsegsize; 486 int maxburst = TCP_MAXBURST; 487 int af; /* address family on the wire */ 488 int iphdrlen; 489 490 #ifdef DIAGNOSTIC 491 if (tp->t_inpcb && tp->t_in6pcb) 492 panic("tcp_output: both t_inpcb and t_in6pcb are set"); 493 #endif 494 so = NULL; 495 ro = NULL; 496 if (tp->t_inpcb) { 497 so = tp->t_inpcb->inp_socket; 498 ro = &tp->t_inpcb->inp_route; 499 } 500 #ifdef INET6 501 else if (tp->t_in6pcb) { 502 so = tp->t_in6pcb->in6p_socket; 503 ro = (struct route *)&tp->t_in6pcb->in6p_route; 504 } 505 #endif 506 507 switch (af = tp->t_family) { 508 #ifdef INET 509 case AF_INET: 510 if (tp->t_inpcb) 511 break; 512 #ifdef INET6 513 /* mapped addr case */ 514 if (tp->t_in6pcb) 515 break; 516 #endif 517 return EINVAL; 518 #endif 519 #ifdef INET6 520 case AF_INET6: 521 if (tp->t_in6pcb) 522 break; 523 return EINVAL; 524 #endif 525 default: 526 return EAFNOSUPPORT; 527 } 528 529 tcp_segsize(tp, &txsegsize, &rxsegsize); 530 531 idle = (tp->snd_max == tp->snd_una); 532 533 /* 534 * Restart Window computation. From draft-floyd-incr-init-win-03: 535 * 536 * Optionally, a TCP MAY set the restart window to the 537 * minimum of the value used for the initial window and 538 * the current value of cwnd (in other words, using a 539 * larger value for the restart window should never increase 540 * the size of cwnd). 541 */ 542 if (tcp_cwm) { 543 /* 544 * Hughes/Touch/Heidemann Congestion Window Monitoring. 545 * Count the number of packets currently pending 546 * acknowledgement, and limit our congestion window 547 * to a pre-determined allowed burst size plus that count. 548 * This prevents bursting once all pending packets have 549 * been acknowledged (i.e. transmission is idle). 550 * 551 * XXX Link this to Initial Window? 552 */ 553 tp->snd_cwnd = min(tp->snd_cwnd, 554 (tcp_cwm_burstsize * txsegsize) + 555 (tp->snd_nxt - tp->snd_una)); 556 } else { 557 if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) { 558 /* 559 * We have been idle for "a while" and no acks are 560 * expected to clock out any data we send -- 561 * slow start to get ack "clock" running again. 562 */ 563 tp->snd_cwnd = min(tp->snd_cwnd, 564 TCP_INITIAL_WINDOW(tcp_init_win, txsegsize)); 565 } 566 } 567 568 again: 569 /* 570 * Determine length of data that should be transmitted, and 571 * flags that should be used. If there is some data or critical 572 * controls (SYN, RST) to send, then transmit; otherwise, 573 * investigate further. 574 */ 575 sendalot = 0; 576 off = tp->snd_nxt - tp->snd_una; 577 win = min(tp->snd_wnd, tp->snd_cwnd); 578 579 flags = tcp_outflags[tp->t_state]; 580 /* 581 * If in persist timeout with window of 0, send 1 byte. 582 * Otherwise, if window is small but nonzero 583 * and timer expired, we will send what we can 584 * and go to transmit state. 585 */ 586 if (tp->t_force) { 587 if (win == 0) { 588 /* 589 * If we still have some data to send, then 590 * clear the FIN bit. Usually this would 591 * happen below when it realizes that we 592 * aren't sending all the data. However, 593 * if we have exactly 1 byte of unset data, 594 * then it won't clear the FIN bit below, 595 * and if we are in persist state, we wind 596 * up sending the packet without recording 597 * that we sent the FIN bit. 598 * 599 * We can't just blindly clear the FIN bit, 600 * because if we don't have any more data 601 * to send then the probe will be the FIN 602 * itself. 603 */ 604 if (off < so->so_snd.sb_cc) 605 flags &= ~TH_FIN; 606 win = 1; 607 } else { 608 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 609 tp->t_rxtshift = 0; 610 } 611 } 612 613 if (win < so->so_snd.sb_cc) { 614 len = win - off; 615 flags &= ~TH_FIN; 616 } else 617 len = so->so_snd.sb_cc - off; 618 619 if (len < 0) { 620 /* 621 * If FIN has been sent but not acked, 622 * but we haven't been called to retransmit, 623 * len will be -1. Otherwise, window shrank 624 * after we sent into it. If window shrank to 0, 625 * cancel pending retransmit, pull snd_nxt back 626 * to (closed) window, and set the persist timer 627 * if it isn't already going. If the window didn't 628 * close completely, just wait for an ACK. 629 * 630 * If we have a pending FIN, either it has already been 631 * transmitted or it is outside the window, so drop it. 632 * If the FIN has been transmitted, but this is not a 633 * retransmission, then len must be -1. Therefore we also 634 * prevent here the sending of `gratuitous FINs'. This 635 * eliminates the need to check for that case below (e.g. 636 * to back up snd_nxt before the FIN so that the sequence 637 * number is correct). 638 */ 639 len = 0; 640 flags &= ~TH_FIN; 641 if (win == 0) { 642 TCP_TIMER_DISARM(tp, TCPT_REXMT); 643 tp->t_rxtshift = 0; 644 tp->snd_nxt = tp->snd_una; 645 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) 646 tcp_setpersist(tp); 647 } 648 } 649 if (len > txsegsize) { 650 len = txsegsize; 651 flags &= ~TH_FIN; 652 sendalot = 1; 653 } 654 655 win = sbspace(&so->so_rcv); 656 657 /* 658 * Sender silly window avoidance. If connection is idle 659 * and can send all data, a maximum segment, 660 * at least a maximum default-size segment do it, 661 * or are forced, do it; otherwise don't bother. 662 * If peer's buffer is tiny, then send 663 * when window is at least half open. 664 * If retransmitting (possibly after persist timer forced us 665 * to send into a small window), then must resend. 666 */ 667 if (len) { 668 if (len == txsegsize) 669 goto send; 670 if ((so->so_state & SS_MORETOCOME) == 0 && 671 ((idle || tp->t_flags & TF_NODELAY) && 672 len + off >= so->so_snd.sb_cc)) 673 goto send; 674 if (tp->t_force) 675 goto send; 676 if (len >= tp->max_sndwnd / 2) 677 goto send; 678 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 679 goto send; 680 } 681 682 /* 683 * Compare available window to amount of window known to peer 684 * (as advertised window less next expected input). If the 685 * difference is at least twice the size of the largest segment 686 * we expect to receive (i.e. two segments) or at least 50% of 687 * the maximum possible window, then want to send a window update 688 * to peer. 689 */ 690 if (win > 0) { 691 /* 692 * "adv" is the amount we can increase the window, 693 * taking into account that we are limited by 694 * TCP_MAXWIN << tp->rcv_scale. 695 */ 696 long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) - 697 (tp->rcv_adv - tp->rcv_nxt); 698 699 if (adv >= (long) (2 * rxsegsize)) 700 goto send; 701 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 702 goto send; 703 } 704 705 /* 706 * Send if we owe peer an ACK. 707 */ 708 if (tp->t_flags & TF_ACKNOW) 709 goto send; 710 if (flags & (TH_SYN|TH_FIN|TH_RST)) 711 goto send; 712 if (SEQ_GT(tp->snd_up, tp->snd_una)) 713 goto send; 714 715 /* 716 * TCP window updates are not reliable, rather a polling protocol 717 * using ``persist'' packets is used to insure receipt of window 718 * updates. The three ``states'' for the output side are: 719 * idle not doing retransmits or persists 720 * persisting to move a small or zero window 721 * (re)transmitting and thereby not persisting 722 * 723 * tp->t_timer[TCPT_PERSIST] 724 * is set when we are in persist state. 725 * tp->t_force 726 * is set when we are called to send a persist packet. 727 * tp->t_timer[TCPT_REXMT] 728 * is set when we are retransmitting 729 * The output side is idle when both timers are zero. 730 * 731 * If send window is too small, there is data to transmit, and no 732 * retransmit or persist is pending, then go to persist state. 733 * If nothing happens soon, send when timer expires: 734 * if window is nonzero, transmit what we can, 735 * otherwise force out a byte. 736 */ 737 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 738 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 739 tp->t_rxtshift = 0; 740 tcp_setpersist(tp); 741 } 742 743 /* 744 * No reason to send a segment, just return. 745 */ 746 return (0); 747 748 send: 749 /* 750 * Before ESTABLISHED, force sending of initial options 751 * unless TCP set not to do any options. 752 * NOTE: we assume that the IP/TCP header plus TCP options 753 * always fit in a single mbuf, leaving room for a maximum 754 * link header, i.e. 755 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 756 */ 757 optlen = 0; 758 switch (af) { 759 #ifdef INET 760 case AF_INET: 761 iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 762 break; 763 #endif 764 #ifdef INET6 765 case AF_INET6: 766 iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 767 break; 768 #endif 769 default: /*pacify gcc*/ 770 iphdrlen = 0; 771 break; 772 } 773 hdrlen = iphdrlen; 774 if (flags & TH_SYN) { 775 struct rtentry *rt; 776 777 rt = NULL; 778 #ifdef INET 779 if (tp->t_inpcb) 780 rt = in_pcbrtentry(tp->t_inpcb); 781 #endif 782 #ifdef INET6 783 if (tp->t_in6pcb) 784 rt = in6_pcbrtentry(tp->t_in6pcb); 785 #endif 786 787 tp->snd_nxt = tp->iss; 788 tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ? 789 rt->rt_ifp : NULL, af); 790 if ((tp->t_flags & TF_NOOPT) == 0) { 791 opt[0] = TCPOPT_MAXSEG; 792 opt[1] = 4; 793 opt[2] = (tp->t_ourmss >> 8) & 0xff; 794 opt[3] = tp->t_ourmss & 0xff; 795 optlen = 4; 796 797 if ((tp->t_flags & TF_REQ_SCALE) && 798 ((flags & TH_ACK) == 0 || 799 (tp->t_flags & TF_RCVD_SCALE))) { 800 *((u_int32_t *) (opt + optlen)) = htonl( 801 TCPOPT_NOP << 24 | 802 TCPOPT_WINDOW << 16 | 803 TCPOLEN_WINDOW << 8 | 804 tp->request_r_scale); 805 optlen += 4; 806 } 807 } 808 } 809 810 /* 811 * Send a timestamp and echo-reply if this is a SYN and our side 812 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 813 * and our peer have sent timestamps in our SYN's. 814 */ 815 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 816 (flags & TH_RST) == 0 && 817 ((flags & (TH_SYN|TH_ACK)) == TH_SYN || 818 (tp->t_flags & TF_RCVD_TSTMP))) { 819 u_int32_t *lp = (u_int32_t *)(opt + optlen); 820 821 /* Form timestamp option as shown in appendix A of RFC 1323. */ 822 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 823 *lp++ = htonl(TCP_TIMESTAMP(tp)); 824 *lp = htonl(tp->ts_recent); 825 optlen += TCPOLEN_TSTAMP_APPA; 826 } 827 828 hdrlen += optlen; 829 830 #ifdef DIAGNOSTIC 831 if (len > txsegsize) 832 panic("tcp data to be sent is larger than segment"); 833 if (max_linkhdr + hdrlen > MCLBYTES) 834 panic("tcphdr too big"); 835 #endif 836 837 /* 838 * Grab a header mbuf, attaching a copy of data to 839 * be transmitted, and initialize the header from 840 * the template for sends on this connection. 841 */ 842 if (len) { 843 error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m); 844 if (error) 845 goto out; 846 /* 847 * If we're sending everything we've got, set PUSH. 848 * (This will keep happy those implementations which only 849 * give data to the user when a buffer fills or 850 * a PUSH comes in.) 851 */ 852 if (off + len == so->so_snd.sb_cc) 853 flags |= TH_PUSH; 854 } else { 855 if (tp->t_flags & TF_ACKNOW) 856 tcpstat.tcps_sndacks++; 857 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 858 tcpstat.tcps_sndctrl++; 859 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 860 tcpstat.tcps_sndurg++; 861 else 862 tcpstat.tcps_sndwinup++; 863 864 MGETHDR(m, M_DONTWAIT, MT_HEADER); 865 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 866 MCLGET(m, M_DONTWAIT); 867 if ((m->m_flags & M_EXT) == 0) { 868 m_freem(m); 869 m = NULL; 870 } 871 } 872 if (m == NULL) { 873 error = ENOBUFS; 874 goto out; 875 } 876 m->m_data += max_linkhdr; 877 m->m_len = hdrlen; 878 } 879 m->m_pkthdr.rcvif = (struct ifnet *)0; 880 switch (af) { 881 #ifdef INET 882 case AF_INET: 883 ip = mtod(m, struct ip *); 884 #ifdef INET6 885 ip6 = NULL; 886 #endif 887 th = (struct tcphdr *)(ip + 1); 888 break; 889 #endif 890 #ifdef INET6 891 case AF_INET6: 892 ip = NULL; 893 ip6 = mtod(m, struct ip6_hdr *); 894 th = (struct tcphdr *)(ip6 + 1); 895 break; 896 #endif 897 default: /*pacify gcc*/ 898 ip = NULL; 899 #ifdef INET6 900 ip6 = NULL; 901 #endif 902 th = NULL; 903 break; 904 } 905 if (tp->t_template == 0) 906 panic("tcp_output"); 907 if (tp->t_template->m_len < iphdrlen) 908 panic("tcp_output"); 909 bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen); 910 911 /* 912 * If we are doing retransmissions, then snd_nxt will 913 * not reflect the first unsent octet. For ACK only 914 * packets, we do not want the sequence number of the 915 * retransmitted packet, we want the sequence number 916 * of the next unsent octet. So, if there is no data 917 * (and no SYN or FIN), use snd_max instead of snd_nxt 918 * when filling in ti_seq. But if we are in persist 919 * state, snd_max might reflect one byte beyond the 920 * right edge of the window, so use snd_nxt in that 921 * case, since we know we aren't doing a retransmission. 922 * (retransmit and persist are mutually exclusive...) 923 */ 924 if (len || (flags & (TH_SYN|TH_FIN)) || 925 TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) 926 th->th_seq = htonl(tp->snd_nxt); 927 else 928 th->th_seq = htonl(tp->snd_max); 929 th->th_ack = htonl(tp->rcv_nxt); 930 if (optlen) { 931 bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen); 932 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 933 } 934 th->th_flags = flags; 935 /* 936 * Calculate receive window. Don't shrink window, 937 * but avoid silly window syndrome. 938 */ 939 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize) 940 win = 0; 941 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 942 win = (long)TCP_MAXWIN << tp->rcv_scale; 943 if (win < (long)(tp->rcv_adv - tp->rcv_nxt)) 944 win = (long)(tp->rcv_adv - tp->rcv_nxt); 945 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale)); 946 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 947 u_int32_t urp = tp->snd_up - tp->snd_nxt; 948 if (urp > IP_MAXPACKET) 949 urp = IP_MAXPACKET; 950 th->th_urp = htons((u_int16_t)urp); 951 th->th_flags |= TH_URG; 952 } else 953 /* 954 * If no urgent pointer to send, then we pull 955 * the urgent pointer to the left edge of the send window 956 * so that it doesn't drift into the send window on sequence 957 * number wraparound. 958 */ 959 tp->snd_up = tp->snd_una; /* drag it along */ 960 961 /* 962 * Set ourselves up to be checksummed just before the packet 963 * hits the wire. 964 */ 965 switch (af) { 966 #ifdef INET 967 case AF_INET: 968 m->m_pkthdr.csum_flags = M_CSUM_TCPv4; 969 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 970 if (len + optlen) { 971 /* Fixup the pseudo-header checksum. */ 972 /* XXXJRT Not IP Jumbogram safe. */ 973 th->th_sum = in_cksum_addword(th->th_sum, 974 htons((u_int16_t) (len + optlen))); 975 } 976 break; 977 #endif 978 #ifdef INET6 979 case AF_INET6: 980 /* 981 * XXX Actually delaying the checksum is Hard 982 * XXX (well, maybe not for Itojun, but it is 983 * XXX for me), but we can still take advantage 984 * XXX of the cached pseudo-header checksum. 985 */ 986 /* equals to hdrlen + len */ 987 m->m_pkthdr.len = sizeof(struct ip6_hdr) 988 + sizeof(struct tcphdr) + optlen + len; 989 #ifdef notyet 990 m->m_pkthdr.csum_flags = M_CSUM_TCPv6; 991 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 992 #endif 993 if (len + optlen) { 994 /* Fixup the pseudo-header checksum. */ 995 /* XXXJRT: Not IPv6 Jumbogram safe. */ 996 th->th_sum = in_cksum_addword(th->th_sum, 997 htons((u_int16_t) (len + optlen))); 998 } 999 #ifndef notyet 1000 th->th_sum = in6_cksum(m, 0, sizeof(struct ip6_hdr), 1001 sizeof(struct tcphdr) + optlen + len); 1002 #endif 1003 break; 1004 #endif 1005 } 1006 1007 /* 1008 * In transmit state, time the transmission and arrange for 1009 * the retransmit. In persist state, just set snd_max. 1010 */ 1011 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1012 tcp_seq startseq = tp->snd_nxt; 1013 1014 /* 1015 * Advance snd_nxt over sequence space of this segment. 1016 * There are no states in which we send both a SYN and a FIN, 1017 * so we collapse the tests for these flags. 1018 */ 1019 if (flags & (TH_SYN|TH_FIN)) 1020 tp->snd_nxt++; 1021 tp->snd_nxt += len; 1022 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1023 tp->snd_max = tp->snd_nxt; 1024 /* 1025 * Time this transmission if not a retransmission and 1026 * not currently timing anything. 1027 */ 1028 if (tp->t_rtttime == 0) { 1029 tp->t_rtttime = tcp_now; 1030 tp->t_rtseq = startseq; 1031 tcpstat.tcps_segstimed++; 1032 } 1033 } 1034 1035 /* 1036 * Set retransmit timer if not currently set, 1037 * and not doing an ack or a keep-alive probe. 1038 * Initial value for retransmit timer is smoothed 1039 * round-trip time + 2 * round-trip time variance. 1040 * Initialize shift counter which is used for backoff 1041 * of retransmit time. 1042 */ 1043 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1044 tp->snd_nxt != tp->snd_una) { 1045 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1046 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1047 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 1048 tp->t_rxtshift = 0; 1049 } 1050 } 1051 } else 1052 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 1053 tp->snd_max = tp->snd_nxt + len; 1054 1055 #ifdef TCP_DEBUG 1056 /* 1057 * Trace. 1058 */ 1059 if (so->so_options & SO_DEBUG) { 1060 /* 1061 * need to recover version # field, which was overwritten 1062 * on ip_cksum computation. 1063 */ 1064 struct ip *sip; 1065 sip = mtod(m, struct ip *); 1066 switch (af) { 1067 #ifdef INET 1068 case AF_INET: 1069 sip->ip_v = 4; 1070 break; 1071 #endif 1072 #ifdef INET6 1073 case AF_INET6: 1074 sip->ip_v = 6; 1075 break; 1076 #endif 1077 } 1078 tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0); 1079 } 1080 #endif 1081 1082 /* 1083 * Fill in IP length and desired time to live and 1084 * send to IP level. There should be a better way 1085 * to handle ttl and tos; we could keep them in 1086 * the template, but need a way to checksum without them. 1087 */ 1088 m->m_pkthdr.len = hdrlen + len; 1089 1090 switch (af) { 1091 #ifdef INET 1092 case AF_INET: 1093 ip->ip_len = htons(m->m_pkthdr.len); 1094 if (tp->t_inpcb) { 1095 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl; 1096 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos; 1097 } 1098 #ifdef INET6 1099 else if (tp->t_in6pcb) { 1100 ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/ 1101 ip->ip_tos = 0; /*XXX*/ 1102 } 1103 #endif 1104 break; 1105 #endif 1106 #ifdef INET6 1107 case AF_INET6: 1108 ip6->ip6_nxt = IPPROTO_TCP; 1109 if (tp->t_in6pcb) { 1110 /* 1111 * we separately set hoplimit for every segment, since 1112 * the user might want to change the value via 1113 * setsockopt. Also, desired default hop limit might 1114 * be changed via Neighbor Discovery. 1115 */ 1116 ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb, 1117 ro->ro_rt ? ro->ro_rt->rt_ifp : NULL); 1118 } 1119 /* ip6->ip6_flow = ??? */ 1120 /* ip6_plen will be filled in ip6_output(). */ 1121 break; 1122 #endif 1123 } 1124 1125 #ifdef IPSEC 1126 if (ipsec_setsocket(m, so) != 0) { 1127 m_freem(m); 1128 error = ENOBUFS; 1129 goto out; 1130 } 1131 #endif /*IPSEC*/ 1132 1133 switch (af) { 1134 #ifdef INET 1135 case AF_INET: 1136 { 1137 struct mbuf *opts; 1138 1139 if (tp->t_inpcb) 1140 opts = tp->t_inpcb->inp_options; 1141 else 1142 opts = NULL; 1143 error = ip_output(m, opts, ro, 1144 (tp->t_mtudisc ? IP_MTUDISC : 0) | 1145 (so->so_options & SO_DONTROUTE), 1146 0); 1147 break; 1148 } 1149 #endif 1150 #ifdef INET6 1151 case AF_INET6: 1152 { 1153 struct ip6_pktopts *opts; 1154 1155 if (tp->t_in6pcb) 1156 opts = tp->t_in6pcb->in6p_outputopts; 1157 else 1158 opts = NULL; 1159 error = ip6_output(m, opts, (struct route_in6 *)ro, 1160 so->so_options & SO_DONTROUTE, 0, NULL); 1161 break; 1162 } 1163 #endif 1164 default: 1165 error = EAFNOSUPPORT; 1166 break; 1167 } 1168 if (error) { 1169 out: 1170 if (error == ENOBUFS) { 1171 tcpstat.tcps_selfquench++; 1172 #ifdef INET 1173 if (tp->t_inpcb) 1174 tcp_quench(tp->t_inpcb, 0); 1175 #endif 1176 #ifdef INET6 1177 if (tp->t_in6pcb) 1178 tcp6_quench(tp->t_in6pcb, 0); 1179 #endif 1180 error = 0; 1181 } else if ((error == EHOSTUNREACH || error == ENETDOWN) && 1182 TCPS_HAVERCVDSYN(tp->t_state)) { 1183 tp->t_softerror = error; 1184 error = 0; 1185 } 1186 1187 /* Restart the delayed ACK timer, if necessary. */ 1188 if (tp->t_flags & TF_DELACK) 1189 TCP_RESTART_DELACK(tp); 1190 1191 return (error); 1192 } 1193 tcpstat.tcps_sndtotal++; 1194 if (tp->t_flags & TF_DELACK) 1195 tcpstat.tcps_delack++; 1196 1197 /* 1198 * Data sent (as far as we can tell). 1199 * If this advertises a larger window than any other segment, 1200 * then remember the size of the advertised window. 1201 * Any pending ACK has now been sent. 1202 */ 1203 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv)) 1204 tp->rcv_adv = tp->rcv_nxt + win; 1205 tp->last_ack_sent = tp->rcv_nxt; 1206 tp->t_flags &= ~TF_ACKNOW; 1207 TCP_CLEAR_DELACK(tp); 1208 #ifdef DIAGNOSTIC 1209 if (maxburst < 0) 1210 printf("tcp_output: maxburst exceeded by %d\n", -maxburst); 1211 #endif 1212 if (sendalot && (!tcp_do_newreno || --maxburst)) 1213 goto again; 1214 return (0); 1215 } 1216 1217 void 1218 tcp_setpersist(tp) 1219 struct tcpcb *tp; 1220 { 1221 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2); 1222 int nticks; 1223 1224 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT)) 1225 panic("tcp_output REXMT"); 1226 /* 1227 * Start/restart persistance timer. 1228 */ 1229 if (t < tp->t_rttmin) 1230 t = tp->t_rttmin; 1231 TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift], 1232 TCPTV_PERSMIN, TCPTV_PERSMAX); 1233 TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks); 1234 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1235 tp->t_rxtshift++; 1236 } 1237