1 /* $NetBSD: tcp_output.c,v 1.90 2003/03/01 04:40:28 thorpej 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 1.90 2003/03/01 04:40:28 thorpej 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 MCLAIM(m, &tcp_tx_mowner); 425 426 /* 427 * XXX Because other code assumes headers will fit in 428 * XXX one header mbuf. 429 * 430 * (This code should almost *never* be run.) 431 */ 432 if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) { 433 TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader); 434 MCLGET(m, M_DONTWAIT); 435 if ((m->m_flags & M_EXT) == 0) { 436 m_freem(m); 437 return (ENOBUFS); 438 } 439 } 440 441 m->m_data += max_linkhdr; 442 m->m_len = hdrlen; 443 if (len <= M_TRAILINGSPACE(m)) { 444 m_copydata(so->so_snd.sb_mb, off, (int) len, 445 mtod(m, caddr_t) + hdrlen); 446 m->m_len += len; 447 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall); 448 } else { 449 m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len); 450 if (m->m_next == NULL) { 451 m_freem(m); 452 return (ENOBUFS); 453 } 454 #ifdef TCP_OUTPUT_COUNTERS 455 if (m->m_next->m_flags & M_EXT) 456 TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig); 457 else 458 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig); 459 #endif /* TCP_OUTPUT_COUNTERS */ 460 } 461 #endif 462 463 *mp = m; 464 return (0); 465 } 466 467 /* 468 * Tcp output routine: figure out what should be sent and send it. 469 */ 470 int 471 tcp_output(tp) 472 struct tcpcb *tp; 473 { 474 struct socket *so; 475 struct route *ro; 476 long len, win; 477 int off, flags, error; 478 struct mbuf *m; 479 struct ip *ip; 480 #ifdef INET6 481 struct ip6_hdr *ip6; 482 #endif 483 struct tcphdr *th; 484 u_char opt[MAX_TCPOPTLEN]; 485 unsigned optlen, hdrlen; 486 int idle, sendalot, txsegsize, rxsegsize; 487 int maxburst = TCP_MAXBURST; 488 int af; /* address family on the wire */ 489 int iphdrlen; 490 491 #ifdef DIAGNOSTIC 492 if (tp->t_inpcb && tp->t_in6pcb) 493 panic("tcp_output: both t_inpcb and t_in6pcb are set"); 494 #endif 495 so = NULL; 496 ro = NULL; 497 if (tp->t_inpcb) { 498 so = tp->t_inpcb->inp_socket; 499 ro = &tp->t_inpcb->inp_route; 500 } 501 #ifdef INET6 502 else if (tp->t_in6pcb) { 503 so = tp->t_in6pcb->in6p_socket; 504 ro = (struct route *)&tp->t_in6pcb->in6p_route; 505 } 506 #endif 507 508 switch (af = tp->t_family) { 509 #ifdef INET 510 case AF_INET: 511 if (tp->t_inpcb) 512 break; 513 #ifdef INET6 514 /* mapped addr case */ 515 if (tp->t_in6pcb) 516 break; 517 #endif 518 return EINVAL; 519 #endif 520 #ifdef INET6 521 case AF_INET6: 522 if (tp->t_in6pcb) 523 break; 524 return EINVAL; 525 #endif 526 default: 527 return EAFNOSUPPORT; 528 } 529 530 tcp_segsize(tp, &txsegsize, &rxsegsize); 531 532 idle = (tp->snd_max == tp->snd_una); 533 534 /* 535 * Restart Window computation. From draft-floyd-incr-init-win-03: 536 * 537 * Optionally, a TCP MAY set the restart window to the 538 * minimum of the value used for the initial window and 539 * the current value of cwnd (in other words, using a 540 * larger value for the restart window should never increase 541 * the size of cwnd). 542 */ 543 if (tcp_cwm) { 544 /* 545 * Hughes/Touch/Heidemann Congestion Window Monitoring. 546 * Count the number of packets currently pending 547 * acknowledgement, and limit our congestion window 548 * to a pre-determined allowed burst size plus that count. 549 * This prevents bursting once all pending packets have 550 * been acknowledged (i.e. transmission is idle). 551 * 552 * XXX Link this to Initial Window? 553 */ 554 tp->snd_cwnd = min(tp->snd_cwnd, 555 (tcp_cwm_burstsize * txsegsize) + 556 (tp->snd_nxt - tp->snd_una)); 557 } else { 558 if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) { 559 /* 560 * We have been idle for "a while" and no acks are 561 * expected to clock out any data we send -- 562 * slow start to get ack "clock" running again. 563 */ 564 int ss = tcp_init_win; 565 #ifdef INET 566 if (tp->t_inpcb && 567 in_localaddr(tp->t_inpcb->inp_faddr)) 568 ss = tcp_init_win_local; 569 #endif 570 #ifdef INET6 571 if (tp->t_in6pcb && 572 in6_localaddr(&tp->t_in6pcb->in6p_faddr)) 573 ss = tcp_init_win_local; 574 #endif 575 tp->snd_cwnd = min(tp->snd_cwnd, 576 TCP_INITIAL_WINDOW(ss, txsegsize)); 577 } 578 } 579 580 again: 581 /* 582 * Determine length of data that should be transmitted, and 583 * flags that should be used. If there is some data or critical 584 * controls (SYN, RST) to send, then transmit; otherwise, 585 * investigate further. 586 */ 587 sendalot = 0; 588 off = tp->snd_nxt - tp->snd_una; 589 win = min(tp->snd_wnd, tp->snd_cwnd); 590 591 flags = tcp_outflags[tp->t_state]; 592 /* 593 * If in persist timeout with window of 0, send 1 byte. 594 * Otherwise, if window is small but nonzero 595 * and timer expired, we will send what we can 596 * and go to transmit state. 597 */ 598 if (tp->t_force) { 599 if (win == 0) { 600 /* 601 * If we still have some data to send, then 602 * clear the FIN bit. Usually this would 603 * happen below when it realizes that we 604 * aren't sending all the data. However, 605 * if we have exactly 1 byte of unset data, 606 * then it won't clear the FIN bit below, 607 * and if we are in persist state, we wind 608 * up sending the packet without recording 609 * that we sent the FIN bit. 610 * 611 * We can't just blindly clear the FIN bit, 612 * because if we don't have any more data 613 * to send then the probe will be the FIN 614 * itself. 615 */ 616 if (off < so->so_snd.sb_cc) 617 flags &= ~TH_FIN; 618 win = 1; 619 } else { 620 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 621 tp->t_rxtshift = 0; 622 } 623 } 624 625 if (win < so->so_snd.sb_cc) { 626 len = win - off; 627 flags &= ~TH_FIN; 628 } else 629 len = so->so_snd.sb_cc - off; 630 631 if (len < 0) { 632 /* 633 * If FIN has been sent but not acked, 634 * but we haven't been called to retransmit, 635 * len will be -1. Otherwise, window shrank 636 * after we sent into it. If window shrank to 0, 637 * cancel pending retransmit, pull snd_nxt back 638 * to (closed) window, and set the persist timer 639 * if it isn't already going. If the window didn't 640 * close completely, just wait for an ACK. 641 * 642 * If we have a pending FIN, either it has already been 643 * transmitted or it is outside the window, so drop it. 644 * If the FIN has been transmitted, but this is not a 645 * retransmission, then len must be -1. Therefore we also 646 * prevent here the sending of `gratuitous FINs'. This 647 * eliminates the need to check for that case below (e.g. 648 * to back up snd_nxt before the FIN so that the sequence 649 * number is correct). 650 */ 651 len = 0; 652 flags &= ~TH_FIN; 653 if (win == 0) { 654 TCP_TIMER_DISARM(tp, TCPT_REXMT); 655 tp->t_rxtshift = 0; 656 tp->snd_nxt = tp->snd_una; 657 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) 658 tcp_setpersist(tp); 659 } 660 } 661 if (len > txsegsize) { 662 len = txsegsize; 663 flags &= ~TH_FIN; 664 sendalot = 1; 665 } 666 667 win = sbspace(&so->so_rcv); 668 669 /* 670 * Sender silly window avoidance. If connection is idle 671 * and can send all data, a maximum segment, 672 * at least a maximum default-size segment do it, 673 * or are forced, do it; otherwise don't bother. 674 * If peer's buffer is tiny, then send 675 * when window is at least half open. 676 * If retransmitting (possibly after persist timer forced us 677 * to send into a small window), then must resend. 678 */ 679 if (len) { 680 if (len == txsegsize) 681 goto send; 682 if ((so->so_state & SS_MORETOCOME) == 0 && 683 ((idle || tp->t_flags & TF_NODELAY) && 684 len + off >= so->so_snd.sb_cc)) 685 goto send; 686 if (tp->t_force) 687 goto send; 688 if (len >= tp->max_sndwnd / 2) 689 goto send; 690 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 691 goto send; 692 } 693 694 /* 695 * Compare available window to amount of window known to peer 696 * (as advertised window less next expected input). If the 697 * difference is at least twice the size of the largest segment 698 * we expect to receive (i.e. two segments) or at least 50% of 699 * the maximum possible window, then want to send a window update 700 * to peer. 701 */ 702 if (win > 0) { 703 /* 704 * "adv" is the amount we can increase the window, 705 * taking into account that we are limited by 706 * TCP_MAXWIN << tp->rcv_scale. 707 */ 708 long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) - 709 (tp->rcv_adv - tp->rcv_nxt); 710 711 if (adv >= (long) (2 * rxsegsize)) 712 goto send; 713 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 714 goto send; 715 } 716 717 /* 718 * Send if we owe peer an ACK. 719 */ 720 if (tp->t_flags & TF_ACKNOW) 721 goto send; 722 if (flags & (TH_SYN|TH_FIN|TH_RST)) 723 goto send; 724 if (SEQ_GT(tp->snd_up, tp->snd_una)) 725 goto send; 726 727 /* 728 * TCP window updates are not reliable, rather a polling protocol 729 * using ``persist'' packets is used to insure receipt of window 730 * updates. The three ``states'' for the output side are: 731 * idle not doing retransmits or persists 732 * persisting to move a small or zero window 733 * (re)transmitting and thereby not persisting 734 * 735 * tp->t_timer[TCPT_PERSIST] 736 * is set when we are in persist state. 737 * tp->t_force 738 * is set when we are called to send a persist packet. 739 * tp->t_timer[TCPT_REXMT] 740 * is set when we are retransmitting 741 * The output side is idle when both timers are zero. 742 * 743 * If send window is too small, there is data to transmit, and no 744 * retransmit or persist is pending, then go to persist state. 745 * If nothing happens soon, send when timer expires: 746 * if window is nonzero, transmit what we can, 747 * otherwise force out a byte. 748 */ 749 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 750 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 751 tp->t_rxtshift = 0; 752 tcp_setpersist(tp); 753 } 754 755 /* 756 * No reason to send a segment, just return. 757 */ 758 return (0); 759 760 send: 761 /* 762 * Before ESTABLISHED, force sending of initial options 763 * unless TCP set not to do any options. 764 * NOTE: we assume that the IP/TCP header plus TCP options 765 * always fit in a single mbuf, leaving room for a maximum 766 * link header, i.e. 767 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 768 */ 769 optlen = 0; 770 switch (af) { 771 #ifdef INET 772 case AF_INET: 773 iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 774 break; 775 #endif 776 #ifdef INET6 777 case AF_INET6: 778 iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 779 break; 780 #endif 781 default: /*pacify gcc*/ 782 iphdrlen = 0; 783 break; 784 } 785 hdrlen = iphdrlen; 786 if (flags & TH_SYN) { 787 struct rtentry *rt; 788 789 rt = NULL; 790 #ifdef INET 791 if (tp->t_inpcb) 792 rt = in_pcbrtentry(tp->t_inpcb); 793 #endif 794 #ifdef INET6 795 if (tp->t_in6pcb) 796 rt = in6_pcbrtentry(tp->t_in6pcb); 797 #endif 798 799 tp->snd_nxt = tp->iss; 800 tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ? 801 rt->rt_ifp : NULL, af); 802 if ((tp->t_flags & TF_NOOPT) == 0) { 803 opt[0] = TCPOPT_MAXSEG; 804 opt[1] = 4; 805 opt[2] = (tp->t_ourmss >> 8) & 0xff; 806 opt[3] = tp->t_ourmss & 0xff; 807 optlen = 4; 808 809 if ((tp->t_flags & TF_REQ_SCALE) && 810 ((flags & TH_ACK) == 0 || 811 (tp->t_flags & TF_RCVD_SCALE))) { 812 *((u_int32_t *) (opt + optlen)) = htonl( 813 TCPOPT_NOP << 24 | 814 TCPOPT_WINDOW << 16 | 815 TCPOLEN_WINDOW << 8 | 816 tp->request_r_scale); 817 optlen += 4; 818 } 819 } 820 } 821 822 /* 823 * Send a timestamp and echo-reply if this is a SYN and our side 824 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 825 * and our peer have sent timestamps in our SYN's. 826 */ 827 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 828 (flags & TH_RST) == 0 && 829 ((flags & (TH_SYN|TH_ACK)) == TH_SYN || 830 (tp->t_flags & TF_RCVD_TSTMP))) { 831 u_int32_t *lp = (u_int32_t *)(opt + optlen); 832 833 /* Form timestamp option as shown in appendix A of RFC 1323. */ 834 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 835 *lp++ = htonl(TCP_TIMESTAMP(tp)); 836 *lp = htonl(tp->ts_recent); 837 optlen += TCPOLEN_TSTAMP_APPA; 838 } 839 840 hdrlen += optlen; 841 842 #ifdef DIAGNOSTIC 843 if (len > txsegsize) 844 panic("tcp data to be sent is larger than segment"); 845 if (max_linkhdr + hdrlen > MCLBYTES) 846 panic("tcphdr too big"); 847 #endif 848 849 /* 850 * Grab a header mbuf, attaching a copy of data to 851 * be transmitted, and initialize the header from 852 * the template for sends on this connection. 853 */ 854 if (len) { 855 error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m); 856 if (error) 857 goto out; 858 /* 859 * If we're sending everything we've got, set PUSH. 860 * (This will keep happy those implementations which only 861 * give data to the user when a buffer fills or 862 * a PUSH comes in.) 863 */ 864 if (off + len == so->so_snd.sb_cc) 865 flags |= TH_PUSH; 866 } else { 867 if (tp->t_flags & TF_ACKNOW) 868 tcpstat.tcps_sndacks++; 869 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 870 tcpstat.tcps_sndctrl++; 871 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 872 tcpstat.tcps_sndurg++; 873 else 874 tcpstat.tcps_sndwinup++; 875 876 MGETHDR(m, M_DONTWAIT, MT_HEADER); 877 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 878 MCLGET(m, M_DONTWAIT); 879 if ((m->m_flags & M_EXT) == 0) { 880 m_freem(m); 881 m = NULL; 882 } 883 } 884 if (m == NULL) { 885 error = ENOBUFS; 886 goto out; 887 } 888 MCLAIM(m, &tcp_tx_mowner); 889 m->m_data += max_linkhdr; 890 m->m_len = hdrlen; 891 } 892 m->m_pkthdr.rcvif = (struct ifnet *)0; 893 switch (af) { 894 #ifdef INET 895 case AF_INET: 896 ip = mtod(m, struct ip *); 897 #ifdef INET6 898 ip6 = NULL; 899 #endif 900 th = (struct tcphdr *)(ip + 1); 901 break; 902 #endif 903 #ifdef INET6 904 case AF_INET6: 905 ip = NULL; 906 ip6 = mtod(m, struct ip6_hdr *); 907 th = (struct tcphdr *)(ip6 + 1); 908 break; 909 #endif 910 default: /*pacify gcc*/ 911 ip = NULL; 912 #ifdef INET6 913 ip6 = NULL; 914 #endif 915 th = NULL; 916 break; 917 } 918 if (tp->t_template == 0) 919 panic("tcp_output"); 920 if (tp->t_template->m_len < iphdrlen) 921 panic("tcp_output"); 922 bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen); 923 924 /* 925 * If we are doing retransmissions, then snd_nxt will 926 * not reflect the first unsent octet. For ACK only 927 * packets, we do not want the sequence number of the 928 * retransmitted packet, we want the sequence number 929 * of the next unsent octet. So, if there is no data 930 * (and no SYN or FIN), use snd_max instead of snd_nxt 931 * when filling in ti_seq. But if we are in persist 932 * state, snd_max might reflect one byte beyond the 933 * right edge of the window, so use snd_nxt in that 934 * case, since we know we aren't doing a retransmission. 935 * (retransmit and persist are mutually exclusive...) 936 */ 937 if (len || (flags & (TH_SYN|TH_FIN)) || 938 TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) 939 th->th_seq = htonl(tp->snd_nxt); 940 else 941 th->th_seq = htonl(tp->snd_max); 942 th->th_ack = htonl(tp->rcv_nxt); 943 if (optlen) { 944 bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen); 945 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 946 } 947 th->th_flags = flags; 948 /* 949 * Calculate receive window. Don't shrink window, 950 * but avoid silly window syndrome. 951 */ 952 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize) 953 win = 0; 954 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 955 win = (long)TCP_MAXWIN << tp->rcv_scale; 956 if (win < (long)(tp->rcv_adv - tp->rcv_nxt)) 957 win = (long)(tp->rcv_adv - tp->rcv_nxt); 958 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale)); 959 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 960 u_int32_t urp = tp->snd_up - tp->snd_nxt; 961 if (urp > IP_MAXPACKET) 962 urp = IP_MAXPACKET; 963 th->th_urp = htons((u_int16_t)urp); 964 th->th_flags |= TH_URG; 965 } else 966 /* 967 * If no urgent pointer to send, then we pull 968 * the urgent pointer to the left edge of the send window 969 * so that it doesn't drift into the send window on sequence 970 * number wraparound. 971 */ 972 tp->snd_up = tp->snd_una; /* drag it along */ 973 974 /* 975 * Set ourselves up to be checksummed just before the packet 976 * hits the wire. 977 */ 978 switch (af) { 979 #ifdef INET 980 case AF_INET: 981 m->m_pkthdr.csum_flags = M_CSUM_TCPv4; 982 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 983 if (len + optlen) { 984 /* Fixup the pseudo-header checksum. */ 985 /* XXXJRT Not IP Jumbogram safe. */ 986 th->th_sum = in_cksum_addword(th->th_sum, 987 htons((u_int16_t) (len + optlen))); 988 } 989 break; 990 #endif 991 #ifdef INET6 992 case AF_INET6: 993 /* 994 * XXX Actually delaying the checksum is Hard 995 * XXX (well, maybe not for Itojun, but it is 996 * XXX for me), but we can still take advantage 997 * XXX of the cached pseudo-header checksum. 998 */ 999 /* equals to hdrlen + len */ 1000 m->m_pkthdr.len = sizeof(struct ip6_hdr) 1001 + sizeof(struct tcphdr) + optlen + len; 1002 #ifdef notyet 1003 m->m_pkthdr.csum_flags = M_CSUM_TCPv6; 1004 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1005 #endif 1006 if (len + optlen) { 1007 /* Fixup the pseudo-header checksum. */ 1008 /* XXXJRT: Not IPv6 Jumbogram safe. */ 1009 th->th_sum = in_cksum_addword(th->th_sum, 1010 htons((u_int16_t) (len + optlen))); 1011 } 1012 #ifndef notyet 1013 th->th_sum = in6_cksum(m, 0, sizeof(struct ip6_hdr), 1014 sizeof(struct tcphdr) + optlen + len); 1015 #endif 1016 break; 1017 #endif 1018 } 1019 1020 /* 1021 * In transmit state, time the transmission and arrange for 1022 * the retransmit. In persist state, just set snd_max. 1023 */ 1024 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1025 tcp_seq startseq = tp->snd_nxt; 1026 1027 /* 1028 * Advance snd_nxt over sequence space of this segment. 1029 * There are no states in which we send both a SYN and a FIN, 1030 * so we collapse the tests for these flags. 1031 */ 1032 if (flags & (TH_SYN|TH_FIN)) 1033 tp->snd_nxt++; 1034 tp->snd_nxt += len; 1035 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1036 tp->snd_max = tp->snd_nxt; 1037 /* 1038 * Time this transmission if not a retransmission and 1039 * not currently timing anything. 1040 */ 1041 if (tp->t_rtttime == 0) { 1042 tp->t_rtttime = tcp_now; 1043 tp->t_rtseq = startseq; 1044 tcpstat.tcps_segstimed++; 1045 } 1046 } 1047 1048 /* 1049 * Set retransmit timer if not currently set, 1050 * and not doing an ack or a keep-alive probe. 1051 * Initial value for retransmit timer is smoothed 1052 * round-trip time + 2 * round-trip time variance. 1053 * Initialize shift counter which is used for backoff 1054 * of retransmit time. 1055 */ 1056 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1057 tp->snd_nxt != tp->snd_una) { 1058 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1059 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1060 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 1061 tp->t_rxtshift = 0; 1062 } 1063 } 1064 } else 1065 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 1066 tp->snd_max = tp->snd_nxt + len; 1067 1068 #ifdef TCP_DEBUG 1069 /* 1070 * Trace. 1071 */ 1072 if (so->so_options & SO_DEBUG) { 1073 /* 1074 * need to recover version # field, which was overwritten 1075 * on ip_cksum computation. 1076 */ 1077 struct ip *sip; 1078 sip = mtod(m, struct ip *); 1079 switch (af) { 1080 #ifdef INET 1081 case AF_INET: 1082 sip->ip_v = 4; 1083 break; 1084 #endif 1085 #ifdef INET6 1086 case AF_INET6: 1087 sip->ip_v = 6; 1088 break; 1089 #endif 1090 } 1091 tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0); 1092 } 1093 #endif 1094 1095 /* 1096 * Fill in IP length and desired time to live and 1097 * send to IP level. There should be a better way 1098 * to handle ttl and tos; we could keep them in 1099 * the template, but need a way to checksum without them. 1100 */ 1101 m->m_pkthdr.len = hdrlen + len; 1102 1103 switch (af) { 1104 #ifdef INET 1105 case AF_INET: 1106 ip->ip_len = htons(m->m_pkthdr.len); 1107 if (tp->t_inpcb) { 1108 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl; 1109 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos; 1110 } 1111 #ifdef INET6 1112 else if (tp->t_in6pcb) { 1113 ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/ 1114 ip->ip_tos = 0; /*XXX*/ 1115 } 1116 #endif 1117 break; 1118 #endif 1119 #ifdef INET6 1120 case AF_INET6: 1121 ip6->ip6_nxt = IPPROTO_TCP; 1122 if (tp->t_in6pcb) { 1123 /* 1124 * we separately set hoplimit for every segment, since 1125 * the user might want to change the value via 1126 * setsockopt. Also, desired default hop limit might 1127 * be changed via Neighbor Discovery. 1128 */ 1129 ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb, 1130 ro->ro_rt ? ro->ro_rt->rt_ifp : NULL); 1131 } 1132 /* ip6->ip6_flow = ??? */ 1133 /* ip6_plen will be filled in ip6_output(). */ 1134 break; 1135 #endif 1136 } 1137 1138 #ifdef IPSEC 1139 if (ipsec_setsocket(m, so) != 0) { 1140 m_freem(m); 1141 error = ENOBUFS; 1142 goto out; 1143 } 1144 #endif /*IPSEC*/ 1145 1146 switch (af) { 1147 #ifdef INET 1148 case AF_INET: 1149 { 1150 struct mbuf *opts; 1151 1152 if (tp->t_inpcb) 1153 opts = tp->t_inpcb->inp_options; 1154 else 1155 opts = NULL; 1156 error = ip_output(m, opts, ro, 1157 (tp->t_mtudisc ? IP_MTUDISC : 0) | 1158 (so->so_options & SO_DONTROUTE), 1159 0); 1160 break; 1161 } 1162 #endif 1163 #ifdef INET6 1164 case AF_INET6: 1165 { 1166 struct ip6_pktopts *opts; 1167 1168 if (tp->t_in6pcb) 1169 opts = tp->t_in6pcb->in6p_outputopts; 1170 else 1171 opts = NULL; 1172 error = ip6_output(m, opts, (struct route_in6 *)ro, 1173 so->so_options & SO_DONTROUTE, 0, NULL); 1174 break; 1175 } 1176 #endif 1177 default: 1178 error = EAFNOSUPPORT; 1179 break; 1180 } 1181 if (error) { 1182 out: 1183 if (error == ENOBUFS) { 1184 tcpstat.tcps_selfquench++; 1185 #ifdef INET 1186 if (tp->t_inpcb) 1187 tcp_quench(tp->t_inpcb, 0); 1188 #endif 1189 #ifdef INET6 1190 if (tp->t_in6pcb) 1191 tcp6_quench(tp->t_in6pcb, 0); 1192 #endif 1193 error = 0; 1194 } else if ((error == EHOSTUNREACH || error == ENETDOWN) && 1195 TCPS_HAVERCVDSYN(tp->t_state)) { 1196 tp->t_softerror = error; 1197 error = 0; 1198 } 1199 1200 /* Restart the delayed ACK timer, if necessary. */ 1201 if (tp->t_flags & TF_DELACK) 1202 TCP_RESTART_DELACK(tp); 1203 1204 return (error); 1205 } 1206 tcpstat.tcps_sndtotal++; 1207 if (tp->t_flags & TF_DELACK) 1208 tcpstat.tcps_delack++; 1209 1210 /* 1211 * Data sent (as far as we can tell). 1212 * If this advertises a larger window than any other segment, 1213 * then remember the size of the advertised window. 1214 * Any pending ACK has now been sent. 1215 */ 1216 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv)) 1217 tp->rcv_adv = tp->rcv_nxt + win; 1218 tp->last_ack_sent = tp->rcv_nxt; 1219 tp->t_flags &= ~TF_ACKNOW; 1220 TCP_CLEAR_DELACK(tp); 1221 #ifdef DIAGNOSTIC 1222 if (maxburst < 0) 1223 printf("tcp_output: maxburst exceeded by %d\n", -maxburst); 1224 #endif 1225 if (sendalot && (!tcp_do_newreno || --maxburst)) 1226 goto again; 1227 return (0); 1228 } 1229 1230 void 1231 tcp_setpersist(tp) 1232 struct tcpcb *tp; 1233 { 1234 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2); 1235 int nticks; 1236 1237 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT)) 1238 panic("tcp_output REXMT"); 1239 /* 1240 * Start/restart persistance timer. 1241 */ 1242 if (t < tp->t_rttmin) 1243 t = tp->t_rttmin; 1244 TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift], 1245 TCPTV_PERSMIN, TCPTV_PERSMAX); 1246 TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks); 1247 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1248 tp->t_rxtshift++; 1249 } 1250