tcp_output.c revision 1.154
1 /* $NetBSD: tcp_output.c,v 1.154 2007/02/10 09:43:05 degroote 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, 2005, 2006 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 * This code is derived from software contributed to The NetBSD Foundation 80 * by Charles M. Hannum. 81 * This code is derived from software contributed to The NetBSD Foundation 82 * by Rui Paulo. 83 * 84 * Redistribution and use in source and binary forms, with or without 85 * modification, are permitted provided that the following conditions 86 * are met: 87 * 1. Redistributions of source code must retain the above copyright 88 * notice, this list of conditions and the following disclaimer. 89 * 2. Redistributions in binary form must reproduce the above copyright 90 * notice, this list of conditions and the following disclaimer in the 91 * documentation and/or other materials provided with the distribution. 92 * 3. All advertising materials mentioning features or use of this software 93 * must display the following acknowledgement: 94 * This product includes software developed by the NetBSD 95 * Foundation, Inc. and its contributors. 96 * 4. Neither the name of The NetBSD Foundation nor the names of its 97 * contributors may be used to endorse or promote products derived 98 * from this software without specific prior written permission. 99 * 100 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 101 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 102 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 103 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 104 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 105 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 106 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 107 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 108 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 109 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 110 * POSSIBILITY OF SUCH DAMAGE. 111 */ 112 113 /* 114 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 115 * The Regents of the University of California. All rights reserved. 116 * 117 * Redistribution and use in source and binary forms, with or without 118 * modification, are permitted provided that the following conditions 119 * are met: 120 * 1. Redistributions of source code must retain the above copyright 121 * notice, this list of conditions and the following disclaimer. 122 * 2. Redistributions in binary form must reproduce the above copyright 123 * notice, this list of conditions and the following disclaimer in the 124 * documentation and/or other materials provided with the distribution. 125 * 3. 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 1.154 2007/02/10 09:43:05 degroote 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 #ifdef TCP_SIGNATURE 162 #include <sys/md5.h> 163 #endif 164 165 #include <net/if.h> 166 #include <net/route.h> 167 168 #include <netinet/in.h> 169 #include <netinet/in_systm.h> 170 #include <netinet/ip.h> 171 #include <netinet/in_pcb.h> 172 #include <netinet/ip_var.h> 173 174 #ifdef INET6 175 #ifndef INET 176 #include <netinet/in.h> 177 #endif 178 #include <netinet/ip6.h> 179 #include <netinet6/in6_var.h> 180 #include <netinet6/ip6_var.h> 181 #include <netinet6/in6_pcb.h> 182 #include <netinet6/nd6.h> 183 #endif 184 185 #ifdef FAST_IPSEC 186 #include <netipsec/ipsec.h> 187 #include <netipsec/key.h> 188 #ifdef INET6 189 #include <netipsec/ipsec6.h> 190 #endif 191 #endif /* FAST_IPSEC*/ 192 #ifdef IPSEC 193 #include <netinet6/ipsec.h> 194 #endif 195 196 #include <netinet/tcp.h> 197 #define TCPOUTFLAGS 198 #include <netinet/tcp_fsm.h> 199 #include <netinet/tcp_seq.h> 200 #include <netinet/tcp_timer.h> 201 #include <netinet/tcp_var.h> 202 #include <netinet/tcp_congctl.h> 203 #include <netinet/tcpip.h> 204 #include <netinet/tcp_debug.h> 205 #include <netinet/in_offload.h> 206 #include <netinet6/in6_offload.h> 207 208 #ifdef IPSEC 209 #include <netkey/key.h> 210 #endif 211 212 #ifdef notyet 213 extern struct mbuf *m_copypack(); 214 #endif 215 216 /* 217 * Knob to enable Congestion Window Monitoring, and control 218 * the burst size it allows. Default burst is 4 packets, per 219 * the Internet draft. 220 */ 221 int tcp_cwm = 0; 222 int tcp_cwm_burstsize = 4; 223 224 #ifdef TCP_OUTPUT_COUNTERS 225 #include <sys/device.h> 226 227 extern struct evcnt tcp_output_bigheader; 228 extern struct evcnt tcp_output_predict_hit; 229 extern struct evcnt tcp_output_predict_miss; 230 extern struct evcnt tcp_output_copysmall; 231 extern struct evcnt tcp_output_copybig; 232 extern struct evcnt tcp_output_refbig; 233 234 #define TCP_OUTPUT_COUNTER_INCR(ev) (ev)->ev_count++ 235 #else 236 237 #define TCP_OUTPUT_COUNTER_INCR(ev) /* nothing */ 238 239 #endif /* TCP_OUTPUT_COUNTERS */ 240 241 static 242 #ifndef GPROF 243 inline 244 #endif 245 int 246 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep, 247 boolean_t *alwaysfragp) 248 { 249 #ifdef INET 250 struct inpcb *inp = tp->t_inpcb; 251 #endif 252 #ifdef INET6 253 struct in6pcb *in6p = tp->t_in6pcb; 254 #endif 255 struct socket *so = NULL; 256 struct rtentry *rt; 257 struct ifnet *ifp; 258 int size; 259 int hdrlen; 260 int optlen; 261 262 *alwaysfragp = FALSE; 263 264 #ifdef DIAGNOSTIC 265 if (tp->t_inpcb && tp->t_in6pcb) 266 panic("tcp_segsize: both t_inpcb and t_in6pcb are set"); 267 #endif 268 switch (tp->t_family) { 269 #ifdef INET 270 case AF_INET: 271 hdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 272 break; 273 #endif 274 #ifdef INET6 275 case AF_INET6: 276 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 277 break; 278 #endif 279 default: 280 size = tcp_mssdflt; 281 goto out; 282 } 283 284 rt = NULL; 285 #ifdef INET 286 if (inp) { 287 rt = in_pcbrtentry(inp); 288 so = inp->inp_socket; 289 } 290 #endif 291 #ifdef INET6 292 if (in6p) { 293 rt = in6_pcbrtentry(in6p); 294 so = in6p->in6p_socket; 295 } 296 #endif 297 if (rt == NULL) { 298 size = tcp_mssdflt; 299 goto out; 300 } 301 302 ifp = rt->rt_ifp; 303 304 size = tcp_mssdflt; 305 if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) { 306 #ifdef INET6 307 if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) { 308 /* 309 * RFC2460 section 5, last paragraph: if path MTU is 310 * smaller than 1280, use 1280 as packet size and 311 * attach fragment header. 312 */ 313 size = IPV6_MMTU - hdrlen - sizeof(struct ip6_frag); 314 *alwaysfragp = TRUE; 315 } else 316 size = rt->rt_rmx.rmx_mtu - hdrlen; 317 #else 318 size = rt->rt_rmx.rmx_mtu - hdrlen; 319 #endif 320 } else if (ifp->if_flags & IFF_LOOPBACK) 321 size = ifp->if_mtu - hdrlen; 322 #ifdef INET 323 else if (inp && tp->t_mtudisc) 324 size = ifp->if_mtu - hdrlen; 325 else if (inp && in_localaddr(inp->inp_faddr)) 326 size = ifp->if_mtu - hdrlen; 327 #endif 328 #ifdef INET6 329 else if (in6p) { 330 #ifdef INET 331 if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) { 332 /* mapped addr case */ 333 struct in_addr d; 334 bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d)); 335 if (tp->t_mtudisc || in_localaddr(d)) 336 size = ifp->if_mtu - hdrlen; 337 } else 338 #endif 339 { 340 /* 341 * for IPv6, path MTU discovery is always turned on, 342 * or the node must use packet size <= 1280. 343 */ 344 size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU; 345 size -= hdrlen; 346 } 347 } 348 #endif 349 out: 350 /* 351 * Now we must make room for whatever extra TCP/IP options are in 352 * the packet. 353 */ 354 optlen = tcp_optlen(tp); 355 356 /* 357 * XXX tp->t_ourmss should have the right size, but without this code 358 * fragmentation will occur... need more investigation 359 */ 360 #ifdef INET 361 if (inp) { 362 #if defined(IPSEC) || defined(FAST_IPSEC) 363 if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND)) 364 optlen += ipsec4_hdrsiz_tcp(tp); 365 #endif 366 optlen += ip_optlen(inp); 367 } 368 #endif 369 #ifdef INET6 370 #ifdef INET 371 if (in6p && tp->t_family == AF_INET) { 372 #if defined(IPSEC) || defined(FAST_IPSEC) 373 if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND)) 374 optlen += ipsec4_hdrsiz_tcp(tp); 375 #endif 376 /* XXX size -= ip_optlen(in6p); */ 377 } else 378 #endif 379 if (in6p && tp->t_family == AF_INET6) { 380 #if defined(IPSEC) || defined(FAST_IPSEC) 381 if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND)) 382 optlen += ipsec6_hdrsiz_tcp(tp); 383 #endif 384 optlen += ip6_optlen(in6p); 385 } 386 #endif 387 size -= optlen; 388 389 /* there may not be any room for data if mtu is too small */ 390 if (size < 0) 391 return (EMSGSIZE); 392 393 /* 394 * *rxsegsizep holds *estimated* inbound segment size (estimation 395 * assumes that path MTU is the same for both ways). this is only 396 * for silly window avoidance, do not use the value for other purposes. 397 * 398 * ipseclen is subtracted from both sides, this may not be right. 399 * I'm not quite sure about this (could someone comment). 400 */ 401 *txsegsizep = min(tp->t_peermss - optlen, size); 402 /* 403 * Never send more than half a buffer full. This insures that we can 404 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and 405 * therefore acks will never be delayed unless we run out of data to 406 * transmit. 407 */ 408 if (so) 409 *txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep); 410 *rxsegsizep = min(tp->t_ourmss - optlen, size); 411 412 if (*txsegsizep != tp->t_segsz) { 413 /* 414 * If the new segment size is larger, we don't want to 415 * mess up the congestion window, but if it is smaller 416 * we'll have to reduce the congestion window to ensure 417 * that we don't get into trouble with initial windows 418 * and the rest. In any case, if the segment size 419 * has changed, chances are the path has, too, and 420 * our congestion window will be different. 421 */ 422 if (*txsegsizep < tp->t_segsz) { 423 tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz) 424 * *txsegsizep, *txsegsizep); 425 tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz) 426 * *txsegsizep, *txsegsizep); 427 } 428 tp->t_segsz = *txsegsizep; 429 } 430 431 return (0); 432 } 433 434 static 435 #ifndef GPROF 436 inline 437 #endif 438 int 439 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off, 440 long len, int hdrlen, struct mbuf **mp) 441 { 442 struct mbuf *m, *m0; 443 444 if (tp->t_force && len == 1) 445 tcpstat.tcps_sndprobe++; 446 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 447 tcpstat.tcps_sndrexmitpack++; 448 tcpstat.tcps_sndrexmitbyte += len; 449 } else { 450 tcpstat.tcps_sndpack++; 451 tcpstat.tcps_sndbyte += len; 452 } 453 #ifdef notyet 454 if ((m = m_copypack(so->so_snd.sb_mb, off, 455 (int)len, max_linkhdr + hdrlen)) == 0) 456 return (ENOBUFS); 457 /* 458 * m_copypack left space for our hdr; use it. 459 */ 460 m->m_len += hdrlen; 461 m->m_data -= hdrlen; 462 #else 463 MGETHDR(m, M_DONTWAIT, MT_HEADER); 464 if (__predict_false(m == NULL)) 465 return (ENOBUFS); 466 MCLAIM(m, &tcp_tx_mowner); 467 468 /* 469 * XXX Because other code assumes headers will fit in 470 * XXX one header mbuf. 471 * 472 * (This code should almost *never* be run.) 473 */ 474 if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) { 475 TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader); 476 MCLGET(m, M_DONTWAIT); 477 if ((m->m_flags & M_EXT) == 0) { 478 m_freem(m); 479 return (ENOBUFS); 480 } 481 } 482 483 m->m_data += max_linkhdr; 484 m->m_len = hdrlen; 485 486 /* 487 * To avoid traversing the whole sb_mb chain for correct 488 * data to send, remember last sent mbuf, its offset and 489 * the sent size. When called the next time, see if the 490 * data to send is directly following the previous transfer. 491 * This is important for large TCP windows. 492 */ 493 if (off == 0 || tp->t_lastm == NULL || 494 (tp->t_lastoff + tp->t_lastlen) != off) { 495 TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss); 496 /* 497 * Either a new packet or a retransmit. 498 * Start from the beginning. 499 */ 500 tp->t_lastm = so->so_snd.sb_mb; 501 tp->t_inoff = off; 502 } else { 503 TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit); 504 tp->t_inoff += tp->t_lastlen; 505 } 506 507 /* Traverse forward to next packet */ 508 while (tp->t_inoff > 0) { 509 if (tp->t_lastm == NULL) 510 panic("tp->t_lastm == NULL"); 511 if (tp->t_inoff < tp->t_lastm->m_len) 512 break; 513 tp->t_inoff -= tp->t_lastm->m_len; 514 tp->t_lastm = tp->t_lastm->m_next; 515 } 516 517 tp->t_lastoff = off; 518 tp->t_lastlen = len; 519 m0 = tp->t_lastm; 520 off = tp->t_inoff; 521 522 if (len <= M_TRAILINGSPACE(m)) { 523 m_copydata(m0, off, (int) len, mtod(m, caddr_t) + hdrlen); 524 m->m_len += len; 525 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall); 526 } else { 527 m->m_next = m_copy(m0, off, (int) len); 528 if (m->m_next == NULL) { 529 m_freem(m); 530 return (ENOBUFS); 531 } 532 #ifdef TCP_OUTPUT_COUNTERS 533 if (m->m_next->m_flags & M_EXT) 534 TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig); 535 else 536 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig); 537 #endif /* TCP_OUTPUT_COUNTERS */ 538 } 539 #endif 540 541 *mp = m; 542 return (0); 543 } 544 545 /* 546 * Tcp output routine: figure out what should be sent and send it. 547 */ 548 int 549 tcp_output(struct tcpcb *tp) 550 { 551 struct socket *so; 552 struct route *ro; 553 long len, win; 554 int off, flags, error; 555 struct mbuf *m; 556 struct ip *ip; 557 #ifdef INET6 558 struct ip6_hdr *ip6; 559 #endif 560 struct tcphdr *th; 561 u_char opt[MAX_TCPOPTLEN]; 562 unsigned optlen, hdrlen, packetlen; 563 unsigned int sack_numblks; 564 int idle, sendalot, txsegsize, rxsegsize; 565 int txsegsize_nosack; 566 int maxburst = TCP_MAXBURST; 567 int af; /* address family on the wire */ 568 int iphdrlen; 569 int has_tso4, has_tso6; 570 int has_tso, use_tso; 571 boolean_t alwaysfrag; 572 int sack_rxmit; 573 int sack_bytes_rxmt; 574 struct sackhole *p; 575 #ifdef TCP_SIGNATURE 576 int sigoff = 0; 577 #endif 578 579 #ifdef DIAGNOSTIC 580 if (tp->t_inpcb && tp->t_in6pcb) 581 panic("tcp_output: both t_inpcb and t_in6pcb are set"); 582 #endif 583 so = NULL; 584 ro = NULL; 585 if (tp->t_inpcb) { 586 so = tp->t_inpcb->inp_socket; 587 ro = &tp->t_inpcb->inp_route; 588 } 589 #ifdef INET6 590 else if (tp->t_in6pcb) { 591 so = tp->t_in6pcb->in6p_socket; 592 ro = (struct route *)&tp->t_in6pcb->in6p_route; 593 } 594 #endif 595 596 switch (af = tp->t_family) { 597 #ifdef INET 598 case AF_INET: 599 if (tp->t_inpcb) 600 break; 601 #ifdef INET6 602 /* mapped addr case */ 603 if (tp->t_in6pcb) 604 break; 605 #endif 606 return (EINVAL); 607 #endif 608 #ifdef INET6 609 case AF_INET6: 610 if (tp->t_in6pcb) 611 break; 612 return (EINVAL); 613 #endif 614 default: 615 return (EAFNOSUPPORT); 616 } 617 618 if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag)) 619 return (EMSGSIZE); 620 621 idle = (tp->snd_max == tp->snd_una); 622 623 /* 624 * Determine if we can use TCP segmentation offload: 625 * - If we're using IPv4 626 * - If there is not an IPsec policy that prevents it 627 * - If the interface can do it 628 */ 629 has_tso4 = has_tso6 = FALSE; 630 #if defined(INET) 631 has_tso4 = tp->t_inpcb != NULL && 632 #if defined(IPSEC) || defined(FAST_IPSEC) 633 IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp, 634 IPSEC_DIR_OUTBOUND) && 635 #endif 636 tp->t_inpcb->inp_route.ro_rt != NULL && 637 (tp->t_inpcb->inp_route.ro_rt->rt_ifp->if_capenable & 638 IFCAP_TSOv4) != 0; 639 #endif /* defined(INET) */ 640 #if defined(INET6) 641 has_tso6 = tp->t_in6pcb != NULL && 642 #if defined(IPSEC) || defined(FAST_IPSEC) 643 IPSEC_PCB_SKIP_IPSEC(tp->t_in6pcb->in6p_sp, 644 IPSEC_DIR_OUTBOUND) && 645 #endif 646 tp->t_in6pcb->in6p_route.ro_rt != NULL && 647 (tp->t_in6pcb->in6p_route.ro_rt->rt_ifp->if_capenable & 648 IFCAP_TSOv6) != 0; 649 #endif /* defined(INET6) */ 650 has_tso = (has_tso4 || has_tso6) && !alwaysfrag; 651 652 /* 653 * Restart Window computation. From draft-floyd-incr-init-win-03: 654 * 655 * Optionally, a TCP MAY set the restart window to the 656 * minimum of the value used for the initial window and 657 * the current value of cwnd (in other words, using a 658 * larger value for the restart window should never increase 659 * the size of cwnd). 660 */ 661 if (tcp_cwm) { 662 /* 663 * Hughes/Touch/Heidemann Congestion Window Monitoring. 664 * Count the number of packets currently pending 665 * acknowledgement, and limit our congestion window 666 * to a pre-determined allowed burst size plus that count. 667 * This prevents bursting once all pending packets have 668 * been acknowledged (i.e. transmission is idle). 669 * 670 * XXX Link this to Initial Window? 671 */ 672 tp->snd_cwnd = min(tp->snd_cwnd, 673 (tcp_cwm_burstsize * txsegsize) + 674 (tp->snd_nxt - tp->snd_una)); 675 } else { 676 if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) { 677 /* 678 * We have been idle for "a while" and no acks are 679 * expected to clock out any data we send -- 680 * slow start to get ack "clock" running again. 681 */ 682 int ss = tcp_init_win; 683 #ifdef INET 684 if (tp->t_inpcb && 685 in_localaddr(tp->t_inpcb->inp_faddr)) 686 ss = tcp_init_win_local; 687 #endif 688 #ifdef INET6 689 if (tp->t_in6pcb && 690 in6_localaddr(&tp->t_in6pcb->in6p_faddr)) 691 ss = tcp_init_win_local; 692 #endif 693 tp->snd_cwnd = min(tp->snd_cwnd, 694 TCP_INITIAL_WINDOW(ss, txsegsize)); 695 } 696 } 697 698 txsegsize_nosack = txsegsize; 699 again: 700 use_tso = has_tso; 701 if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) { 702 /* don't duplicate CWR/ECE. */ 703 use_tso = 0; 704 } 705 TCP_REASS_LOCK(tp); 706 sack_numblks = tcp_sack_numblks(tp); 707 if (sack_numblks) { 708 int sackoptlen; 709 710 sackoptlen = TCP_SACK_OPTLEN(sack_numblks); 711 if (sackoptlen > txsegsize_nosack) { 712 sack_numblks = 0; /* give up SACK */ 713 txsegsize = txsegsize_nosack; 714 } else { 715 if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) { 716 /* don't duplicate D-SACK. */ 717 use_tso = 0; 718 } 719 txsegsize = txsegsize_nosack - sackoptlen; 720 } 721 } else { 722 txsegsize = txsegsize_nosack; 723 } 724 725 /* 726 * Determine length of data that should be transmitted, and 727 * flags that should be used. If there is some data or critical 728 * controls (SYN, RST) to send, then transmit; otherwise, 729 * investigate further. 730 * 731 * Readjust SACK information to avoid resending duplicate data. 732 */ 733 if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max)) 734 tcp_sack_adjust(tp); 735 sendalot = 0; 736 off = tp->snd_nxt - tp->snd_una; 737 win = min(tp->snd_wnd, tp->snd_cwnd); 738 739 flags = tcp_outflags[tp->t_state]; 740 741 /* 742 * Send any SACK-generated retransmissions. If we're explicitly trying 743 * to send out new data (when sendalot is 1), bypass this function. 744 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 745 * we're replacing a (future) new transmission with a retransmission 746 * now, and we previously incremented snd_cwnd in tcp_input(). 747 */ 748 /* 749 * Still in sack recovery , reset rxmit flag to zero. 750 */ 751 sack_rxmit = 0; 752 sack_bytes_rxmt = 0; 753 len = 0; 754 p = NULL; 755 do { 756 long cwin; 757 if (!TCP_SACK_ENABLED(tp)) 758 break; 759 if (tp->t_partialacks < 0) 760 break; 761 p = tcp_sack_output(tp, &sack_bytes_rxmt); 762 if (p == NULL) 763 break; 764 765 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 766 if (cwin < 0) 767 cwin = 0; 768 /* Do not retransmit SACK segments beyond snd_recover */ 769 if (SEQ_GT(p->end, tp->snd_recover)) { 770 /* 771 * (At least) part of sack hole extends beyond 772 * snd_recover. Check to see if we can rexmit data 773 * for this hole. 774 */ 775 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 776 /* 777 * Can't rexmit any more data for this hole. 778 * That data will be rexmitted in the next 779 * sack recovery episode, when snd_recover 780 * moves past p->rxmit. 781 */ 782 p = NULL; 783 break; 784 } 785 /* Can rexmit part of the current hole */ 786 len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit)); 787 } else 788 len = ((long)ulmin(cwin, p->end - p->rxmit)); 789 off = p->rxmit - tp->snd_una; 790 if (off + len > so->so_snd.sb_cc) { 791 /* 1 for TH_FIN */ 792 KASSERT(off + len == so->so_snd.sb_cc + 1); 793 KASSERT(p->rxmit + len == tp->snd_max); 794 len = so->so_snd.sb_cc - off; 795 } 796 if (len > 0) { 797 sack_rxmit = 1; 798 sendalot = 1; 799 } 800 } while (/*CONSTCOND*/0); 801 802 /* 803 * If in persist timeout with window of 0, send 1 byte. 804 * Otherwise, if window is small but nonzero 805 * and timer expired, we will send what we can 806 * and go to transmit state. 807 */ 808 if (tp->t_force) { 809 if (win == 0) { 810 /* 811 * If we still have some data to send, then 812 * clear the FIN bit. Usually this would 813 * happen below when it realizes that we 814 * aren't sending all the data. However, 815 * if we have exactly 1 byte of unset data, 816 * then it won't clear the FIN bit below, 817 * and if we are in persist state, we wind 818 * up sending the packet without recording 819 * that we sent the FIN bit. 820 * 821 * We can't just blindly clear the FIN bit, 822 * because if we don't have any more data 823 * to send then the probe will be the FIN 824 * itself. 825 */ 826 if (off < so->so_snd.sb_cc) 827 flags &= ~TH_FIN; 828 win = 1; 829 } else { 830 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 831 tp->t_rxtshift = 0; 832 } 833 } 834 835 if (sack_rxmit == 0) { 836 if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) { 837 long cwin; 838 839 /* 840 * We are inside of a SACK recovery episode and are 841 * sending new data, having retransmitted all the 842 * data possible in the scoreboard. 843 */ 844 if (tp->snd_wnd < so->so_snd.sb_cc) { 845 len = tp->snd_wnd - off; 846 flags &= ~TH_FIN; 847 } else { 848 len = so->so_snd.sb_cc - off; 849 } 850 851 /* 852 * From FreeBSD: 853 * Don't remove this (len > 0) check ! 854 * We explicitly check for len > 0 here (although it 855 * isn't really necessary), to work around a gcc 856 * optimization issue - to force gcc to compute 857 * len above. Without this check, the computation 858 * of len is bungled by the optimizer. 859 */ 860 if (len > 0) { 861 cwin = tp->snd_cwnd - 862 (tp->snd_nxt - tp->sack_newdata) - 863 sack_bytes_rxmt; 864 if (cwin < 0) 865 cwin = 0; 866 if (cwin < len) { 867 len = cwin; 868 flags &= ~TH_FIN; 869 } 870 } 871 } else if (win < so->so_snd.sb_cc) { 872 len = win - off; 873 flags &= ~TH_FIN; 874 } else { 875 len = so->so_snd.sb_cc - off; 876 } 877 } 878 879 if (len < 0) { 880 /* 881 * If FIN has been sent but not acked, 882 * but we haven't been called to retransmit, 883 * len will be -1. Otherwise, window shrank 884 * after we sent into it. If window shrank to 0, 885 * cancel pending retransmit, pull snd_nxt back 886 * to (closed) window, and set the persist timer 887 * if it isn't already going. If the window didn't 888 * close completely, just wait for an ACK. 889 * 890 * If we have a pending FIN, either it has already been 891 * transmitted or it is outside the window, so drop it. 892 * If the FIN has been transmitted, but this is not a 893 * retransmission, then len must be -1. Therefore we also 894 * prevent here the sending of `gratuitous FINs'. This 895 * eliminates the need to check for that case below (e.g. 896 * to back up snd_nxt before the FIN so that the sequence 897 * number is correct). 898 */ 899 len = 0; 900 flags &= ~TH_FIN; 901 if (win == 0) { 902 TCP_TIMER_DISARM(tp, TCPT_REXMT); 903 tp->t_rxtshift = 0; 904 tp->snd_nxt = tp->snd_una; 905 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) 906 tcp_setpersist(tp); 907 } 908 } 909 if (len > txsegsize) { 910 if (use_tso) { 911 /* 912 * Truncate TSO transfers to IP_MAXPACKET, and make 913 * sure that we send equal size transfers down the 914 * stack (rather than big-small-big-small-...). 915 */ 916 #ifdef INET6 917 #if IPV6_MAXPACKET != IP_MAXPACKET 918 #error IPV6_MAXPACKET != IP_MAXPACKET 919 #endif 920 #endif 921 len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize; 922 if (len <= txsegsize) { 923 use_tso = 0; 924 } 925 } else 926 len = txsegsize; 927 flags &= ~TH_FIN; 928 sendalot = 1; 929 } else 930 use_tso = 0; 931 if (sack_rxmit) { 932 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 933 flags &= ~TH_FIN; 934 } 935 936 win = sbspace(&so->so_rcv); 937 938 /* 939 * Sender silly window avoidance. If connection is idle 940 * and can send all data, a maximum segment, 941 * at least a maximum default-size segment do it, 942 * or are forced, do it; otherwise don't bother. 943 * If peer's buffer is tiny, then send 944 * when window is at least half open. 945 * If retransmitting (possibly after persist timer forced us 946 * to send into a small window), then must resend. 947 */ 948 if (len) { 949 if (len >= txsegsize) 950 goto send; 951 if ((so->so_state & SS_MORETOCOME) == 0 && 952 ((idle || tp->t_flags & TF_NODELAY) && 953 len + off >= so->so_snd.sb_cc)) 954 goto send; 955 if (tp->t_force) 956 goto send; 957 if (len >= tp->max_sndwnd / 2) 958 goto send; 959 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 960 goto send; 961 if (sack_rxmit) 962 goto send; 963 } 964 965 /* 966 * Compare available window to amount of window known to peer 967 * (as advertised window less next expected input). If the 968 * difference is at least twice the size of the largest segment 969 * we expect to receive (i.e. two segments) or at least 50% of 970 * the maximum possible window, then want to send a window update 971 * to peer. 972 */ 973 if (win > 0) { 974 /* 975 * "adv" is the amount we can increase the window, 976 * taking into account that we are limited by 977 * TCP_MAXWIN << tp->rcv_scale. 978 */ 979 long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) - 980 (tp->rcv_adv - tp->rcv_nxt); 981 982 if (adv >= (long) (2 * rxsegsize)) 983 goto send; 984 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 985 goto send; 986 } 987 988 /* 989 * Send if we owe peer an ACK. 990 */ 991 if (tp->t_flags & TF_ACKNOW) 992 goto send; 993 if (flags & (TH_SYN|TH_FIN|TH_RST)) 994 goto send; 995 if (SEQ_GT(tp->snd_up, tp->snd_una)) 996 goto send; 997 /* 998 * In SACK, it is possible for tcp_output to fail to send a segment 999 * after the retransmission timer has been turned off. Make sure 1000 * that the retransmission timer is set. 1001 */ 1002 if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) && 1003 !TCP_TIMER_ISARMED(tp, TCPT_REXMT) && 1004 !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1005 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1006 goto just_return; 1007 } 1008 1009 /* 1010 * TCP window updates are not reliable, rather a polling protocol 1011 * using ``persist'' packets is used to insure receipt of window 1012 * updates. The three ``states'' for the output side are: 1013 * idle not doing retransmits or persists 1014 * persisting to move a small or zero window 1015 * (re)transmitting and thereby not persisting 1016 * 1017 * tp->t_timer[TCPT_PERSIST] 1018 * is set when we are in persist state. 1019 * tp->t_force 1020 * is set when we are called to send a persist packet. 1021 * tp->t_timer[TCPT_REXMT] 1022 * is set when we are retransmitting 1023 * The output side is idle when both timers are zero. 1024 * 1025 * If send window is too small, there is data to transmit, and no 1026 * retransmit or persist is pending, then go to persist state. 1027 * If nothing happens soon, send when timer expires: 1028 * if window is nonzero, transmit what we can, 1029 * otherwise force out a byte. 1030 */ 1031 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1032 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1033 tp->t_rxtshift = 0; 1034 tcp_setpersist(tp); 1035 } 1036 1037 /* 1038 * No reason to send a segment, just return. 1039 */ 1040 just_return: 1041 TCP_REASS_UNLOCK(tp); 1042 return (0); 1043 1044 send: 1045 /* 1046 * Before ESTABLISHED, force sending of initial options 1047 * unless TCP set not to do any options. 1048 * NOTE: we assume that the IP/TCP header plus TCP options 1049 * always fit in a single mbuf, leaving room for a maximum 1050 * link header, i.e. 1051 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 1052 */ 1053 optlen = 0; 1054 switch (af) { 1055 #ifdef INET 1056 case AF_INET: 1057 iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 1058 break; 1059 #endif 1060 #ifdef INET6 1061 case AF_INET6: 1062 iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1063 break; 1064 #endif 1065 default: /*pacify gcc*/ 1066 iphdrlen = 0; 1067 break; 1068 } 1069 hdrlen = iphdrlen; 1070 if (flags & TH_SYN) { 1071 struct rtentry *rt; 1072 1073 rt = NULL; 1074 #ifdef INET 1075 if (tp->t_inpcb) 1076 rt = in_pcbrtentry(tp->t_inpcb); 1077 #endif 1078 #ifdef INET6 1079 if (tp->t_in6pcb) 1080 rt = in6_pcbrtentry(tp->t_in6pcb); 1081 #endif 1082 1083 tp->snd_nxt = tp->iss; 1084 tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ? 1085 rt->rt_ifp : NULL, af); 1086 if ((tp->t_flags & TF_NOOPT) == 0) { 1087 opt[0] = TCPOPT_MAXSEG; 1088 opt[1] = 4; 1089 opt[2] = (tp->t_ourmss >> 8) & 0xff; 1090 opt[3] = tp->t_ourmss & 0xff; 1091 optlen = 4; 1092 1093 if ((tp->t_flags & TF_REQ_SCALE) && 1094 ((flags & TH_ACK) == 0 || 1095 (tp->t_flags & TF_RCVD_SCALE))) { 1096 *((u_int32_t *) (opt + optlen)) = htonl( 1097 TCPOPT_NOP << 24 | 1098 TCPOPT_WINDOW << 16 | 1099 TCPOLEN_WINDOW << 8 | 1100 tp->request_r_scale); 1101 optlen += 4; 1102 } 1103 if (tcp_do_sack) { 1104 u_int8_t *cp = (u_int8_t *)(opt + optlen); 1105 1106 cp[0] = TCPOPT_SACK_PERMITTED; 1107 cp[1] = 2; 1108 cp[2] = TCPOPT_NOP; 1109 cp[3] = TCPOPT_NOP; 1110 optlen += 4; 1111 } 1112 } 1113 } 1114 1115 /* 1116 * Send a timestamp and echo-reply if this is a SYN and our side 1117 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 1118 * and our peer have sent timestamps in our SYN's. 1119 */ 1120 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 1121 (flags & TH_RST) == 0 && 1122 ((flags & (TH_SYN|TH_ACK)) == TH_SYN || 1123 (tp->t_flags & TF_RCVD_TSTMP))) { 1124 u_int32_t *lp = (u_int32_t *)(opt + optlen); 1125 1126 /* Form timestamp option as shown in appendix A of RFC 1323. */ 1127 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 1128 *lp++ = htonl(TCP_TIMESTAMP(tp)); 1129 *lp = htonl(tp->ts_recent); 1130 optlen += TCPOLEN_TSTAMP_APPA; 1131 } 1132 1133 /* 1134 * Tack on the SACK block if it is necessary. 1135 */ 1136 if (sack_numblks) { 1137 int sack_len; 1138 u_char *bp = (u_char *)(opt + optlen); 1139 u_int32_t *lp = (u_int32_t *)(bp + 4); 1140 struct ipqent *tiqe; 1141 1142 sack_len = sack_numblks * 8 + 2; 1143 bp[0] = TCPOPT_NOP; 1144 bp[1] = TCPOPT_NOP; 1145 bp[2] = TCPOPT_SACK; 1146 bp[3] = sack_len; 1147 if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) { 1148 sack_numblks--; 1149 *lp++ = htonl(tp->rcv_dsack_block.left); 1150 *lp++ = htonl(tp->rcv_dsack_block.right); 1151 tp->rcv_sack_flags &= ~TCPSACK_HAVED; 1152 } 1153 for (tiqe = TAILQ_FIRST(&tp->timeq); 1154 sack_numblks > 0; tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) { 1155 KASSERT(tiqe != NULL); 1156 sack_numblks--; 1157 *lp++ = htonl(tiqe->ipqe_seq); 1158 *lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len + 1159 ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0)); 1160 } 1161 optlen += sack_len + 2; 1162 } 1163 TCP_REASS_UNLOCK(tp); 1164 1165 #ifdef TCP_SIGNATURE 1166 if (tp->t_flags & TF_SIGNATURE) { 1167 u_char *bp; 1168 /* 1169 * Initialize TCP-MD5 option (RFC2385) 1170 */ 1171 bp = (u_char *)opt + optlen; 1172 *bp++ = TCPOPT_SIGNATURE; 1173 *bp++ = TCPOLEN_SIGNATURE; 1174 sigoff = optlen + 2; 1175 bzero(bp, TCP_SIGLEN); 1176 bp += TCP_SIGLEN; 1177 optlen += TCPOLEN_SIGNATURE; 1178 /* 1179 * Terminate options list and maintain 32-bit alignment. 1180 */ 1181 *bp++ = TCPOPT_NOP; 1182 *bp++ = TCPOPT_EOL; 1183 optlen += 2; 1184 } 1185 #endif /* TCP_SIGNATURE */ 1186 1187 hdrlen += optlen; 1188 1189 #ifdef DIAGNOSTIC 1190 if (!use_tso && len > txsegsize) 1191 panic("tcp data to be sent is larger than segment"); 1192 else if (use_tso && len > IP_MAXPACKET) 1193 panic("tcp data to be sent is larger than max TSO size"); 1194 if (max_linkhdr + hdrlen > MCLBYTES) 1195 panic("tcphdr too big"); 1196 #endif 1197 1198 /* 1199 * Grab a header mbuf, attaching a copy of data to 1200 * be transmitted, and initialize the header from 1201 * the template for sends on this connection. 1202 */ 1203 if (len) { 1204 error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m); 1205 if (error) 1206 goto out; 1207 /* 1208 * If we're sending everything we've got, set PUSH. 1209 * (This will keep happy those implementations which only 1210 * give data to the user when a buffer fills or 1211 * a PUSH comes in.) 1212 */ 1213 if (off + len == so->so_snd.sb_cc) 1214 flags |= TH_PUSH; 1215 } else { 1216 if (tp->t_flags & TF_ACKNOW) 1217 tcpstat.tcps_sndacks++; 1218 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1219 tcpstat.tcps_sndctrl++; 1220 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1221 tcpstat.tcps_sndurg++; 1222 else 1223 tcpstat.tcps_sndwinup++; 1224 1225 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1226 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 1227 MCLGET(m, M_DONTWAIT); 1228 if ((m->m_flags & M_EXT) == 0) { 1229 m_freem(m); 1230 m = NULL; 1231 } 1232 } 1233 if (m == NULL) { 1234 error = ENOBUFS; 1235 goto out; 1236 } 1237 MCLAIM(m, &tcp_tx_mowner); 1238 m->m_data += max_linkhdr; 1239 m->m_len = hdrlen; 1240 } 1241 m->m_pkthdr.rcvif = (struct ifnet *)0; 1242 switch (af) { 1243 #ifdef INET 1244 case AF_INET: 1245 ip = mtod(m, struct ip *); 1246 #ifdef INET6 1247 ip6 = NULL; 1248 #endif 1249 th = (struct tcphdr *)(ip + 1); 1250 break; 1251 #endif 1252 #ifdef INET6 1253 case AF_INET6: 1254 ip = NULL; 1255 ip6 = mtod(m, struct ip6_hdr *); 1256 th = (struct tcphdr *)(ip6 + 1); 1257 break; 1258 #endif 1259 default: /*pacify gcc*/ 1260 ip = NULL; 1261 #ifdef INET6 1262 ip6 = NULL; 1263 #endif 1264 th = NULL; 1265 break; 1266 } 1267 if (tp->t_template == 0) 1268 panic("tcp_output"); 1269 if (tp->t_template->m_len < iphdrlen) 1270 panic("tcp_output"); 1271 bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen); 1272 1273 /* 1274 * If we are starting a connection, send ECN setup 1275 * SYN packet. If we are on a retransmit, we may 1276 * resend those bits a number of times as per 1277 * RFC 3168. 1278 */ 1279 if (tp->t_state == TCPS_SYN_SENT && tcp_do_ecn) { 1280 if (tp->t_flags & TF_SYN_REXMT) { 1281 if (tp->t_ecn_retries--) 1282 flags |= TH_ECE|TH_CWR; 1283 } else { 1284 flags |= TH_ECE|TH_CWR; 1285 tp->t_ecn_retries = tcp_ecn_maxretries; 1286 } 1287 } 1288 1289 if (TCP_ECN_ALLOWED(tp)) { 1290 /* 1291 * If the peer has ECN, mark data packets 1292 * ECN capable. Ignore pure ack packets, retransmissions 1293 * and window probes. 1294 */ 1295 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1296 !(tp->t_force && len == 1)) { 1297 switch (af) { 1298 #ifdef INET 1299 case AF_INET: 1300 tp->t_inpcb->inp_ip.ip_tos |= IPTOS_ECN_ECT0; 1301 break; 1302 #endif 1303 #ifdef INET6 1304 case AF_INET6: 1305 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1306 break; 1307 #endif 1308 } 1309 tcpstat.tcps_ecn_ect++; 1310 } 1311 1312 /* 1313 * Reply with proper ECN notifications. 1314 */ 1315 if (tp->t_flags & TF_ECN_SND_CWR) { 1316 flags |= TH_CWR; 1317 tp->t_flags &= ~TF_ECN_SND_CWR; 1318 } 1319 if (tp->t_flags & TF_ECN_SND_ECE) { 1320 flags |= TH_ECE; 1321 } 1322 } 1323 1324 1325 /* 1326 * If we are doing retransmissions, then snd_nxt will 1327 * not reflect the first unsent octet. For ACK only 1328 * packets, we do not want the sequence number of the 1329 * retransmitted packet, we want the sequence number 1330 * of the next unsent octet. So, if there is no data 1331 * (and no SYN or FIN), use snd_max instead of snd_nxt 1332 * when filling in ti_seq. But if we are in persist 1333 * state, snd_max might reflect one byte beyond the 1334 * right edge of the window, so use snd_nxt in that 1335 * case, since we know we aren't doing a retransmission. 1336 * (retransmit and persist are mutually exclusive...) 1337 */ 1338 if (TCP_SACK_ENABLED(tp) && sack_rxmit) { 1339 th->th_seq = htonl(p->rxmit); 1340 p->rxmit += len; 1341 } else { 1342 if (len || (flags & (TH_SYN|TH_FIN)) || 1343 TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) 1344 th->th_seq = htonl(tp->snd_nxt); 1345 else 1346 th->th_seq = htonl(tp->snd_max); 1347 } 1348 th->th_ack = htonl(tp->rcv_nxt); 1349 if (optlen) { 1350 bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen); 1351 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1352 } 1353 th->th_flags = flags; 1354 /* 1355 * Calculate receive window. Don't shrink window, 1356 * but avoid silly window syndrome. 1357 */ 1358 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize) 1359 win = 0; 1360 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 1361 win = (long)TCP_MAXWIN << tp->rcv_scale; 1362 if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt)) 1363 win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt); 1364 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale)); 1365 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1366 u_int32_t urp = tp->snd_up - tp->snd_nxt; 1367 if (urp > IP_MAXPACKET) 1368 urp = IP_MAXPACKET; 1369 th->th_urp = htons((u_int16_t)urp); 1370 th->th_flags |= TH_URG; 1371 } else 1372 /* 1373 * If no urgent pointer to send, then we pull 1374 * the urgent pointer to the left edge of the send window 1375 * so that it doesn't drift into the send window on sequence 1376 * number wraparound. 1377 */ 1378 tp->snd_up = tp->snd_una; /* drag it along */ 1379 1380 #ifdef TCP_SIGNATURE 1381 if (sigoff && (tp->t_flags & TF_SIGNATURE)) { 1382 struct secasvar *sav; 1383 u_int8_t *sigp; 1384 1385 sav = tcp_signature_getsav(m, th); 1386 1387 if (sav == NULL) { 1388 if (m) 1389 m_freem(m); 1390 return (EPERM); 1391 } 1392 1393 m->m_pkthdr.len = hdrlen + len; 1394 sigp = (caddr_t)th + sizeof(*th) + sigoff; 1395 tcp_signature(m, th, (caddr_t)th - mtod(m, caddr_t), sav, sigp); 1396 1397 key_sa_recordxfer(sav, m); 1398 #ifdef FAST_IPSEC 1399 KEY_FREESAV(&sav); 1400 #else 1401 key_freesav(sav); 1402 #endif 1403 } 1404 #endif 1405 1406 /* 1407 * Set ourselves up to be checksummed just before the packet 1408 * hits the wire. 1409 */ 1410 switch (af) { 1411 #ifdef INET 1412 case AF_INET: 1413 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1414 if (use_tso) { 1415 m->m_pkthdr.segsz = txsegsize; 1416 m->m_pkthdr.csum_flags = M_CSUM_TSOv4; 1417 } else { 1418 m->m_pkthdr.csum_flags = M_CSUM_TCPv4; 1419 if (len + optlen) { 1420 /* Fixup the pseudo-header checksum. */ 1421 /* XXXJRT Not IP Jumbogram safe. */ 1422 th->th_sum = in_cksum_addword(th->th_sum, 1423 htons((u_int16_t) (len + optlen))); 1424 } 1425 } 1426 break; 1427 #endif 1428 #ifdef INET6 1429 case AF_INET6: 1430 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1431 if (use_tso) { 1432 m->m_pkthdr.segsz = txsegsize; 1433 m->m_pkthdr.csum_flags = M_CSUM_TSOv6; 1434 } else { 1435 m->m_pkthdr.csum_flags = M_CSUM_TCPv6; 1436 if (len + optlen) { 1437 /* Fixup the pseudo-header checksum. */ 1438 /* XXXJRT: Not IPv6 Jumbogram safe. */ 1439 th->th_sum = in_cksum_addword(th->th_sum, 1440 htons((u_int16_t) (len + optlen))); 1441 } 1442 } 1443 break; 1444 #endif 1445 } 1446 1447 /* 1448 * In transmit state, time the transmission and arrange for 1449 * the retransmit. In persist state, just set snd_max. 1450 */ 1451 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1452 tcp_seq startseq = tp->snd_nxt; 1453 1454 /* 1455 * Advance snd_nxt over sequence space of this segment. 1456 * There are no states in which we send both a SYN and a FIN, 1457 * so we collapse the tests for these flags. 1458 */ 1459 if (flags & (TH_SYN|TH_FIN)) 1460 tp->snd_nxt++; 1461 if (sack_rxmit) 1462 goto timer; 1463 tp->snd_nxt += len; 1464 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1465 tp->snd_max = tp->snd_nxt; 1466 /* 1467 * Time this transmission if not a retransmission and 1468 * not currently timing anything. 1469 */ 1470 if (tp->t_rtttime == 0) { 1471 tp->t_rtttime = tcp_now; 1472 tp->t_rtseq = startseq; 1473 tcpstat.tcps_segstimed++; 1474 } 1475 } 1476 1477 /* 1478 * Set retransmit timer if not currently set, 1479 * and not doing an ack or a keep-alive probe. 1480 * Initial value for retransmit timer is smoothed 1481 * round-trip time + 2 * round-trip time variance. 1482 * Initialize shift counter which is used for backoff 1483 * of retransmit time. 1484 */ 1485 timer: 1486 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1487 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1488 tp->snd_nxt != tp->snd_una)) { 1489 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1490 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 1491 tp->t_rxtshift = 0; 1492 } 1493 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1494 } 1495 } else 1496 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 1497 tp->snd_max = tp->snd_nxt + len; 1498 1499 #ifdef TCP_DEBUG 1500 /* 1501 * Trace. 1502 */ 1503 if (so->so_options & SO_DEBUG) 1504 tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0); 1505 #endif 1506 1507 /* 1508 * Fill in IP length and desired time to live and 1509 * send to IP level. There should be a better way 1510 * to handle ttl and tos; we could keep them in 1511 * the template, but need a way to checksum without them. 1512 */ 1513 m->m_pkthdr.len = hdrlen + len; 1514 1515 switch (af) { 1516 #ifdef INET 1517 case AF_INET: 1518 ip->ip_len = htons(m->m_pkthdr.len); 1519 packetlen = m->m_pkthdr.len; 1520 if (tp->t_inpcb) { 1521 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl; 1522 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos; 1523 } 1524 #ifdef INET6 1525 else if (tp->t_in6pcb) { 1526 ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/ 1527 ip->ip_tos = 0; /*XXX*/ 1528 } 1529 #endif 1530 break; 1531 #endif 1532 #ifdef INET6 1533 case AF_INET6: 1534 packetlen = m->m_pkthdr.len; 1535 ip6->ip6_nxt = IPPROTO_TCP; 1536 if (tp->t_in6pcb) { 1537 /* 1538 * we separately set hoplimit for every segment, since 1539 * the user might want to change the value via 1540 * setsockopt. Also, desired default hop limit might 1541 * be changed via Neighbor Discovery. 1542 */ 1543 ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb, 1544 ro->ro_rt ? ro->ro_rt->rt_ifp : NULL); 1545 } 1546 /* ip6->ip6_flow = ??? */ 1547 /* ip6_plen will be filled in ip6_output(). */ 1548 break; 1549 #endif 1550 default: /*pacify gcc*/ 1551 packetlen = 0; 1552 break; 1553 } 1554 1555 switch (af) { 1556 #ifdef INET 1557 case AF_INET: 1558 { 1559 struct mbuf *opts; 1560 1561 if (tp->t_inpcb) 1562 opts = tp->t_inpcb->inp_options; 1563 else 1564 opts = NULL; 1565 error = ip_output(m, opts, ro, 1566 (tp->t_mtudisc ? IP_MTUDISC : 0) | 1567 (so->so_options & SO_DONTROUTE), 1568 (struct ip_moptions *)0, so); 1569 break; 1570 } 1571 #endif 1572 #ifdef INET6 1573 case AF_INET6: 1574 { 1575 struct ip6_pktopts *opts; 1576 1577 if (tp->t_in6pcb) 1578 opts = tp->t_in6pcb->in6p_outputopts; 1579 else 1580 opts = NULL; 1581 error = ip6_output(m, opts, (struct route_in6 *)ro, 1582 so->so_options & SO_DONTROUTE, 1583 (struct ip6_moptions *)0, so, NULL); 1584 break; 1585 } 1586 #endif 1587 default: 1588 error = EAFNOSUPPORT; 1589 break; 1590 } 1591 if (error) { 1592 out: 1593 if (error == ENOBUFS) { 1594 tcpstat.tcps_selfquench++; 1595 #ifdef INET 1596 if (tp->t_inpcb) 1597 tcp_quench(tp->t_inpcb, 0); 1598 #endif 1599 #ifdef INET6 1600 if (tp->t_in6pcb) 1601 tcp6_quench(tp->t_in6pcb, 0); 1602 #endif 1603 error = 0; 1604 } else if ((error == EHOSTUNREACH || error == ENETDOWN) && 1605 TCPS_HAVERCVDSYN(tp->t_state)) { 1606 tp->t_softerror = error; 1607 error = 0; 1608 } 1609 1610 /* Back out the seqence number advance. */ 1611 if (sack_rxmit) 1612 p->rxmit -= len; 1613 1614 /* Restart the delayed ACK timer, if necessary. */ 1615 if (tp->t_flags & TF_DELACK) 1616 TCP_RESTART_DELACK(tp); 1617 1618 return (error); 1619 } 1620 1621 if (packetlen > tp->t_pmtud_mtu_sent) 1622 tp->t_pmtud_mtu_sent = packetlen; 1623 1624 tcpstat.tcps_sndtotal++; 1625 if (tp->t_flags & TF_DELACK) 1626 tcpstat.tcps_delack++; 1627 1628 /* 1629 * Data sent (as far as we can tell). 1630 * If this advertises a larger window than any other segment, 1631 * then remember the size of the advertised window. 1632 * Any pending ACK has now been sent. 1633 */ 1634 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv)) 1635 tp->rcv_adv = tp->rcv_nxt + win; 1636 tp->last_ack_sent = tp->rcv_nxt; 1637 tp->t_flags &= ~TF_ACKNOW; 1638 TCP_CLEAR_DELACK(tp); 1639 #ifdef DIAGNOSTIC 1640 if (maxburst < 0) 1641 printf("tcp_output: maxburst exceeded by %d\n", -maxburst); 1642 #endif 1643 if (sendalot && (tp->t_congctl == &tcp_reno_ctl || --maxburst)) 1644 goto again; 1645 return (0); 1646 } 1647 1648 void 1649 tcp_setpersist(struct tcpcb *tp) 1650 { 1651 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2); 1652 int nticks; 1653 1654 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT)) 1655 panic("tcp_output REXMT"); 1656 /* 1657 * Start/restart persistance timer. 1658 */ 1659 if (t < tp->t_rttmin) 1660 t = tp->t_rttmin; 1661 TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift], 1662 TCPTV_PERSMIN, TCPTV_PERSMAX); 1663 TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks); 1664 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1665 tp->t_rxtshift++; 1666 } 1667
Indexes created Tue Sep 23 14:10:03 GMT 2025