tcp_output.c revision 1.173.8.1
1 /* $NetBSD: tcp_output.c,v 1.173.8.1 2014/11/03 23:05:59 msaitoh 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 * 93 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 94 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 95 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 96 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 97 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 98 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 99 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 100 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 101 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 102 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 103 * POSSIBILITY OF SUCH DAMAGE. 104 */ 105 106 /* 107 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 108 * The Regents of the University of California. All rights reserved. 109 * 110 * Redistribution and use in source and binary forms, with or without 111 * modification, are permitted provided that the following conditions 112 * are met: 113 * 1. Redistributions of source code must retain the above copyright 114 * notice, this list of conditions and the following disclaimer. 115 * 2. Redistributions in binary form must reproduce the above copyright 116 * notice, this list of conditions and the following disclaimer in the 117 * documentation and/or other materials provided with the distribution. 118 * 3. Neither the name of the University nor the names of its contributors 119 * may be used to endorse or promote products derived from this software 120 * without specific prior written permission. 121 * 122 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 123 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 124 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 125 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 126 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 127 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 128 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 129 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 130 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 131 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 132 * SUCH DAMAGE. 133 * 134 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 135 */ 136 137 #include <sys/cdefs.h> 138 __KERNEL_RCSID(0, "$NetBSD: tcp_output.c,v 1.173.8.1 2014/11/03 23:05:59 msaitoh Exp $"); 139 140 #include "opt_inet.h" 141 #include "opt_ipsec.h" 142 #include "opt_tcp_debug.h" 143 144 #include <sys/param.h> 145 #include <sys/systm.h> 146 #include <sys/malloc.h> 147 #include <sys/mbuf.h> 148 #include <sys/protosw.h> 149 #include <sys/socket.h> 150 #include <sys/socketvar.h> 151 #include <sys/errno.h> 152 #include <sys/domain.h> 153 #include <sys/kernel.h> 154 #ifdef TCP_SIGNATURE 155 #include <sys/md5.h> 156 #endif 157 158 #include <net/if.h> 159 #include <net/route.h> 160 161 #include <netinet/in.h> 162 #include <netinet/in_systm.h> 163 #include <netinet/ip.h> 164 #include <netinet/in_pcb.h> 165 #include <netinet/ip_var.h> 166 167 #ifdef INET6 168 #ifndef INET 169 #include <netinet/in.h> 170 #endif 171 #include <netinet/ip6.h> 172 #include <netinet6/in6_var.h> 173 #include <netinet6/ip6_var.h> 174 #include <netinet6/in6_pcb.h> 175 #include <netinet6/nd6.h> 176 #endif 177 178 #ifdef FAST_IPSEC 179 #include <netipsec/ipsec.h> 180 #include <netipsec/key.h> 181 #ifdef INET6 182 #include <netipsec/ipsec6.h> 183 #endif 184 #endif /* FAST_IPSEC*/ 185 #ifdef KAME_IPSEC 186 #include <netinet6/ipsec.h> 187 #endif 188 189 #include <netinet/tcp.h> 190 #define TCPOUTFLAGS 191 #include <netinet/tcp_fsm.h> 192 #include <netinet/tcp_seq.h> 193 #include <netinet/tcp_timer.h> 194 #include <netinet/tcp_var.h> 195 #include <netinet/tcp_private.h> 196 #include <netinet/tcp_congctl.h> 197 #include <netinet/tcpip.h> 198 #include <netinet/tcp_debug.h> 199 #include <netinet/in_offload.h> 200 #include <netinet6/in6_offload.h> 201 202 #ifdef KAME_IPSEC 203 #include <netkey/key.h> 204 #endif 205 206 #ifdef notyet 207 extern struct mbuf *m_copypack(); 208 #endif 209 210 /* 211 * Knob to enable Congestion Window Monitoring, and control 212 * the burst size it allows. Default burst is 4 packets, per 213 * the Internet draft. 214 */ 215 int tcp_cwm = 0; 216 int tcp_cwm_burstsize = 4; 217 218 int tcp_do_autosndbuf = 1; 219 int tcp_autosndbuf_inc = 8 * 1024; 220 int tcp_autosndbuf_max = 256 * 1024; 221 222 #ifdef TCP_OUTPUT_COUNTERS 223 #include <sys/device.h> 224 225 extern struct evcnt tcp_output_bigheader; 226 extern struct evcnt tcp_output_predict_hit; 227 extern struct evcnt tcp_output_predict_miss; 228 extern struct evcnt tcp_output_copysmall; 229 extern struct evcnt tcp_output_copybig; 230 extern struct evcnt tcp_output_refbig; 231 232 #define TCP_OUTPUT_COUNTER_INCR(ev) (ev)->ev_count++ 233 #else 234 235 #define TCP_OUTPUT_COUNTER_INCR(ev) /* nothing */ 236 237 #endif /* TCP_OUTPUT_COUNTERS */ 238 239 static 240 #ifndef GPROF 241 inline 242 #endif 243 int 244 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep, 245 bool *alwaysfragp) 246 { 247 #ifdef INET 248 struct inpcb *inp = tp->t_inpcb; 249 #endif 250 #ifdef INET6 251 struct in6pcb *in6p = tp->t_in6pcb; 252 #endif 253 struct socket *so = NULL; 254 struct rtentry *rt; 255 struct ifnet *ifp; 256 int size; 257 int hdrlen; 258 int optlen; 259 260 *alwaysfragp = false; 261 262 #ifdef DIAGNOSTIC 263 if (tp->t_inpcb && tp->t_in6pcb) 264 panic("tcp_segsize: both t_inpcb and t_in6pcb are set"); 265 #endif 266 switch (tp->t_family) { 267 #ifdef INET 268 case AF_INET: 269 hdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 270 break; 271 #endif 272 #ifdef INET6 273 case AF_INET6: 274 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 275 break; 276 #endif 277 default: 278 size = tcp_mssdflt; 279 goto out; 280 } 281 282 rt = NULL; 283 #ifdef INET 284 if (inp) { 285 rt = in_pcbrtentry(inp); 286 so = inp->inp_socket; 287 } 288 #endif 289 #ifdef INET6 290 if (in6p) { 291 rt = in6_pcbrtentry(in6p); 292 so = in6p->in6p_socket; 293 } 294 #endif 295 if (rt == NULL) { 296 size = tcp_mssdflt; 297 goto out; 298 } 299 300 ifp = rt->rt_ifp; 301 302 size = tcp_mssdflt; 303 if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) { 304 #ifdef INET6 305 if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) { 306 /* 307 * RFC2460 section 5, last paragraph: if path MTU is 308 * smaller than 1280, use 1280 as packet size and 309 * attach fragment header. 310 */ 311 size = IPV6_MMTU - hdrlen - sizeof(struct ip6_frag); 312 *alwaysfragp = true; 313 } else 314 size = rt->rt_rmx.rmx_mtu - hdrlen; 315 #else 316 size = rt->rt_rmx.rmx_mtu - hdrlen; 317 #endif 318 } else if (ifp->if_flags & IFF_LOOPBACK) 319 size = ifp->if_mtu - hdrlen; 320 #ifdef INET 321 else if (inp && tp->t_mtudisc) 322 size = ifp->if_mtu - hdrlen; 323 else if (inp && in_localaddr(inp->inp_faddr)) 324 size = ifp->if_mtu - hdrlen; 325 #endif 326 #ifdef INET6 327 else if (in6p) { 328 #ifdef INET 329 if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) { 330 /* mapped addr case */ 331 struct in_addr d; 332 bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d)); 333 if (tp->t_mtudisc || in_localaddr(d)) 334 size = ifp->if_mtu - hdrlen; 335 } else 336 #endif 337 { 338 /* 339 * for IPv6, path MTU discovery is always turned on, 340 * or the node must use packet size <= 1280. 341 */ 342 size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU; 343 size -= hdrlen; 344 } 345 } 346 #endif 347 out: 348 /* 349 * Now we must make room for whatever extra TCP/IP options are in 350 * the packet. 351 */ 352 optlen = tcp_optlen(tp); 353 354 /* 355 * XXX tp->t_ourmss should have the right size, but without this code 356 * fragmentation will occur... need more investigation 357 */ 358 #ifdef INET 359 if (inp) { 360 #if defined(KAME_IPSEC) || defined(FAST_IPSEC) 361 if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND)) 362 optlen += ipsec4_hdrsiz_tcp(tp); 363 #endif 364 optlen += ip_optlen(inp); 365 } 366 #endif 367 #ifdef INET6 368 #ifdef INET 369 if (in6p && tp->t_family == AF_INET) { 370 #if defined(KAME_IPSEC) || defined(FAST_IPSEC) 371 if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND)) 372 optlen += ipsec4_hdrsiz_tcp(tp); 373 #endif 374 /* XXX size -= ip_optlen(in6p); */ 375 } else 376 #endif 377 if (in6p && tp->t_family == AF_INET6) { 378 #if defined(KAME_IPSEC) || defined(FAST_IPSEC) 379 if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND)) 380 optlen += ipsec6_hdrsiz_tcp(tp); 381 #endif 382 optlen += ip6_optlen(in6p); 383 } 384 #endif 385 size -= optlen; 386 387 /* there may not be any room for data if mtu is too small */ 388 if (size < 0) 389 return (EMSGSIZE); 390 391 /* 392 * *rxsegsizep holds *estimated* inbound segment size (estimation 393 * assumes that path MTU is the same for both ways). this is only 394 * for silly window avoidance, do not use the value for other purposes. 395 * 396 * ipseclen is subtracted from both sides, this may not be right. 397 * I'm not quite sure about this (could someone comment). 398 */ 399 *txsegsizep = min(tp->t_peermss - optlen, size); 400 /* 401 * Never send more than half a buffer full. This insures that we can 402 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and 403 * therefore acks will never be delayed unless we run out of data to 404 * transmit. 405 */ 406 if (so) 407 *txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep); 408 *rxsegsizep = min(tp->t_ourmss - optlen, size); 409 410 if (*txsegsizep != tp->t_segsz) { 411 /* 412 * If the new segment size is larger, we don't want to 413 * mess up the congestion window, but if it is smaller 414 * we'll have to reduce the congestion window to ensure 415 * that we don't get into trouble with initial windows 416 * and the rest. In any case, if the segment size 417 * has changed, chances are the path has, too, and 418 * our congestion window will be different. 419 */ 420 if (*txsegsizep < tp->t_segsz) { 421 tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz) 422 * *txsegsizep, *txsegsizep); 423 tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz) 424 * *txsegsizep, *txsegsizep); 425 } 426 tp->t_segsz = *txsegsizep; 427 } 428 429 return (0); 430 } 431 432 static 433 #ifndef GPROF 434 inline 435 #endif 436 int 437 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off, 438 long len, int hdrlen, struct mbuf **mp) 439 { 440 struct mbuf *m, *m0; 441 uint64_t *tcps; 442 443 tcps = TCP_STAT_GETREF(); 444 if (tp->t_force && len == 1) 445 tcps[TCP_STAT_SNDPROBE]++; 446 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 447 tcps[TCP_STAT_SNDREXMITPACK]++; 448 tcps[TCP_STAT_SNDREXMITBYTE] += len; 449 } else { 450 tcps[TCP_STAT_SNDPACK]++; 451 tcps[TCP_STAT_SNDBYTE] += len; 452 } 453 TCP_STAT_PUTREF(); 454 #ifdef notyet 455 if ((m = m_copypack(so->so_snd.sb_mb, off, 456 (int)len, max_linkhdr + hdrlen)) == 0) 457 return (ENOBUFS); 458 /* 459 * m_copypack left space for our hdr; use it. 460 */ 461 m->m_len += hdrlen; 462 m->m_data -= hdrlen; 463 #else 464 MGETHDR(m, M_DONTWAIT, MT_HEADER); 465 if (__predict_false(m == NULL)) 466 return (ENOBUFS); 467 MCLAIM(m, &tcp_tx_mowner); 468 469 /* 470 * XXX Because other code assumes headers will fit in 471 * XXX one header mbuf. 472 * 473 * (This code should almost *never* be run.) 474 */ 475 if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) { 476 TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader); 477 MCLGET(m, M_DONTWAIT); 478 if ((m->m_flags & M_EXT) == 0) { 479 m_freem(m); 480 return (ENOBUFS); 481 } 482 } 483 484 m->m_data += max_linkhdr; 485 m->m_len = hdrlen; 486 487 /* 488 * To avoid traversing the whole sb_mb chain for correct 489 * data to send, remember last sent mbuf, its offset and 490 * the sent size. When called the next time, see if the 491 * data to send is directly following the previous transfer. 492 * This is important for large TCP windows. 493 */ 494 if (off == 0 || tp->t_lastm == NULL || 495 (tp->t_lastoff + tp->t_lastlen) != off) { 496 TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss); 497 /* 498 * Either a new packet or a retransmit. 499 * Start from the beginning. 500 */ 501 tp->t_lastm = so->so_snd.sb_mb; 502 tp->t_inoff = off; 503 } else { 504 TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit); 505 tp->t_inoff += tp->t_lastlen; 506 } 507 508 /* Traverse forward to next packet */ 509 while (tp->t_inoff > 0) { 510 if (tp->t_lastm == NULL) 511 panic("tp->t_lastm == NULL"); 512 if (tp->t_inoff < tp->t_lastm->m_len) 513 break; 514 tp->t_inoff -= tp->t_lastm->m_len; 515 tp->t_lastm = tp->t_lastm->m_next; 516 } 517 518 tp->t_lastoff = off; 519 tp->t_lastlen = len; 520 m0 = tp->t_lastm; 521 off = tp->t_inoff; 522 523 if (len <= M_TRAILINGSPACE(m)) { 524 m_copydata(m0, off, (int) len, mtod(m, char *) + hdrlen); 525 m->m_len += len; 526 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall); 527 } else { 528 m->m_next = m_copym(m0, off, (int) len, M_DONTWAIT); 529 if (m->m_next == NULL) { 530 m_freem(m); 531 return (ENOBUFS); 532 } 533 #ifdef TCP_OUTPUT_COUNTERS 534 if (m->m_next->m_flags & M_EXT) 535 TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig); 536 else 537 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig); 538 #endif /* TCP_OUTPUT_COUNTERS */ 539 } 540 #endif 541 542 *mp = m; 543 return (0); 544 } 545 546 /* 547 * Tcp output routine: figure out what should be sent and send it. 548 */ 549 int 550 tcp_output(struct tcpcb *tp) 551 { 552 struct rtentry *rt; 553 struct socket *so; 554 struct route *ro; 555 long len, win; 556 int off, flags, error; 557 struct mbuf *m; 558 struct ip *ip; 559 #ifdef INET6 560 struct ip6_hdr *ip6; 561 #endif 562 struct tcphdr *th; 563 u_char opt[MAX_TCPOPTLEN]; 564 #define OPT_FITS(more) ((optlen + (more)) < sizeof(opt)) 565 unsigned optlen, hdrlen, packetlen; 566 unsigned int sack_numblks; 567 int idle, sendalot, txsegsize, rxsegsize; 568 int txsegsize_nosack; 569 int maxburst = TCP_MAXBURST; 570 int af; /* address family on the wire */ 571 int iphdrlen; 572 int has_tso4, has_tso6; 573 int has_tso, use_tso; 574 bool alwaysfrag; 575 int sack_rxmit; 576 int sack_bytes_rxmt; 577 int ecn_tos; 578 struct sackhole *p; 579 #ifdef TCP_SIGNATURE 580 int sigoff = 0; 581 #endif 582 uint64_t *tcps; 583 584 #ifdef DIAGNOSTIC 585 if (tp->t_inpcb && tp->t_in6pcb) 586 panic("tcp_output: both t_inpcb and t_in6pcb are set"); 587 #endif 588 so = NULL; 589 ro = NULL; 590 if (tp->t_inpcb) { 591 so = tp->t_inpcb->inp_socket; 592 ro = &tp->t_inpcb->inp_route; 593 } 594 #ifdef INET6 595 else if (tp->t_in6pcb) { 596 so = tp->t_in6pcb->in6p_socket; 597 ro = &tp->t_in6pcb->in6p_route; 598 } 599 #endif 600 601 switch (af = tp->t_family) { 602 #ifdef INET 603 case AF_INET: 604 if (tp->t_inpcb) 605 break; 606 #ifdef INET6 607 /* mapped addr case */ 608 if (tp->t_in6pcb) 609 break; 610 #endif 611 return (EINVAL); 612 #endif 613 #ifdef INET6 614 case AF_INET6: 615 if (tp->t_in6pcb) 616 break; 617 return (EINVAL); 618 #endif 619 default: 620 return (EAFNOSUPPORT); 621 } 622 623 if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag)) 624 return (EMSGSIZE); 625 626 idle = (tp->snd_max == tp->snd_una); 627 628 /* 629 * Determine if we can use TCP segmentation offload: 630 * - If we're using IPv4 631 * - If there is not an IPsec policy that prevents it 632 * - If the interface can do it 633 */ 634 has_tso4 = has_tso6 = false; 635 #if defined(INET) 636 has_tso4 = tp->t_inpcb != NULL && 637 #if defined(KAME_IPSEC) || defined(FAST_IPSEC) 638 IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp, 639 IPSEC_DIR_OUTBOUND) && 640 #endif 641 (rt = rtcache_validate(&tp->t_inpcb->inp_route)) != NULL && 642 (rt->rt_ifp->if_capenable & IFCAP_TSOv4) != 0; 643 #endif /* defined(INET) */ 644 #if defined(INET6) 645 has_tso6 = tp->t_in6pcb != NULL && 646 #if defined(KAME_IPSEC) || defined(FAST_IPSEC) 647 IPSEC_PCB_SKIP_IPSEC(tp->t_in6pcb->in6p_sp, 648 IPSEC_DIR_OUTBOUND) && 649 #endif 650 (rt = rtcache_validate(&tp->t_in6pcb->in6p_route)) != NULL && 651 (rt->rt_ifp->if_capenable & IFCAP_TSOv6) != 0; 652 #endif /* defined(INET6) */ 653 has_tso = (has_tso4 || has_tso6) && !alwaysfrag; 654 655 /* 656 * Restart Window computation. From draft-floyd-incr-init-win-03: 657 * 658 * Optionally, a TCP MAY set the restart window to the 659 * minimum of the value used for the initial window and 660 * the current value of cwnd (in other words, using a 661 * larger value for the restart window should never increase 662 * the size of cwnd). 663 */ 664 if (tcp_cwm) { 665 /* 666 * Hughes/Touch/Heidemann Congestion Window Monitoring. 667 * Count the number of packets currently pending 668 * acknowledgement, and limit our congestion window 669 * to a pre-determined allowed burst size plus that count. 670 * This prevents bursting once all pending packets have 671 * been acknowledged (i.e. transmission is idle). 672 * 673 * XXX Link this to Initial Window? 674 */ 675 tp->snd_cwnd = min(tp->snd_cwnd, 676 (tcp_cwm_burstsize * txsegsize) + 677 (tp->snd_nxt - tp->snd_una)); 678 } else { 679 if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) { 680 /* 681 * We have been idle for "a while" and no acks are 682 * expected to clock out any data we send -- 683 * slow start to get ack "clock" running again. 684 */ 685 int ss = tcp_init_win; 686 #ifdef INET 687 if (tp->t_inpcb && 688 in_localaddr(tp->t_inpcb->inp_faddr)) 689 ss = tcp_init_win_local; 690 #endif 691 #ifdef INET6 692 if (tp->t_in6pcb && 693 in6_localaddr(&tp->t_in6pcb->in6p_faddr)) 694 ss = tcp_init_win_local; 695 #endif 696 tp->snd_cwnd = min(tp->snd_cwnd, 697 TCP_INITIAL_WINDOW(ss, txsegsize)); 698 } 699 } 700 701 txsegsize_nosack = txsegsize; 702 again: 703 ecn_tos = 0; 704 use_tso = has_tso; 705 if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) { 706 /* don't duplicate CWR/ECE. */ 707 use_tso = 0; 708 } 709 TCP_REASS_LOCK(tp); 710 sack_numblks = tcp_sack_numblks(tp); 711 if (sack_numblks) { 712 int sackoptlen; 713 714 sackoptlen = TCP_SACK_OPTLEN(sack_numblks); 715 if (sackoptlen > txsegsize_nosack) { 716 sack_numblks = 0; /* give up SACK */ 717 txsegsize = txsegsize_nosack; 718 } else { 719 if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) { 720 /* don't duplicate D-SACK. */ 721 use_tso = 0; 722 } 723 txsegsize = txsegsize_nosack - sackoptlen; 724 } 725 } else { 726 txsegsize = txsegsize_nosack; 727 } 728 729 /* 730 * Determine length of data that should be transmitted, and 731 * flags that should be used. If there is some data or critical 732 * controls (SYN, RST) to send, then transmit; otherwise, 733 * investigate further. 734 * 735 * Readjust SACK information to avoid resending duplicate data. 736 */ 737 if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max)) 738 tcp_sack_adjust(tp); 739 sendalot = 0; 740 off = tp->snd_nxt - tp->snd_una; 741 win = min(tp->snd_wnd, tp->snd_cwnd); 742 743 flags = tcp_outflags[tp->t_state]; 744 745 /* 746 * Send any SACK-generated retransmissions. If we're explicitly trying 747 * to send out new data (when sendalot is 1), bypass this function. 748 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 749 * we're replacing a (future) new transmission with a retransmission 750 * now, and we previously incremented snd_cwnd in tcp_input(). 751 */ 752 /* 753 * Still in sack recovery , reset rxmit flag to zero. 754 */ 755 sack_rxmit = 0; 756 sack_bytes_rxmt = 0; 757 len = 0; 758 p = NULL; 759 do { 760 long cwin; 761 if (!TCP_SACK_ENABLED(tp)) 762 break; 763 if (tp->t_partialacks < 0) 764 break; 765 p = tcp_sack_output(tp, &sack_bytes_rxmt); 766 if (p == NULL) 767 break; 768 769 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 770 if (cwin < 0) 771 cwin = 0; 772 /* Do not retransmit SACK segments beyond snd_recover */ 773 if (SEQ_GT(p->end, tp->snd_recover)) { 774 /* 775 * (At least) part of sack hole extends beyond 776 * snd_recover. Check to see if we can rexmit data 777 * for this hole. 778 */ 779 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 780 /* 781 * Can't rexmit any more data for this hole. 782 * That data will be rexmitted in the next 783 * sack recovery episode, when snd_recover 784 * moves past p->rxmit. 785 */ 786 p = NULL; 787 break; 788 } 789 /* Can rexmit part of the current hole */ 790 len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit)); 791 } else 792 len = ((long)ulmin(cwin, p->end - p->rxmit)); 793 off = p->rxmit - tp->snd_una; 794 if (off + len > so->so_snd.sb_cc) { 795 /* 1 for TH_FIN */ 796 KASSERT(off + len == so->so_snd.sb_cc + 1); 797 KASSERT(p->rxmit + len == tp->snd_max); 798 len = so->so_snd.sb_cc - off; 799 } 800 if (len > 0) { 801 sack_rxmit = 1; 802 sendalot = 1; 803 } 804 } while (/*CONSTCOND*/0); 805 806 /* 807 * If in persist timeout with window of 0, send 1 byte. 808 * Otherwise, if window is small but nonzero 809 * and timer expired, we will send what we can 810 * and go to transmit state. 811 */ 812 if (tp->t_force) { 813 if (win == 0) { 814 /* 815 * If we still have some data to send, then 816 * clear the FIN bit. Usually this would 817 * happen below when it realizes that we 818 * aren't sending all the data. However, 819 * if we have exactly 1 byte of unset data, 820 * then it won't clear the FIN bit below, 821 * and if we are in persist state, we wind 822 * up sending the packet without recording 823 * that we sent the FIN bit. 824 * 825 * We can't just blindly clear the FIN bit, 826 * because if we don't have any more data 827 * to send then the probe will be the FIN 828 * itself. 829 */ 830 if (off < so->so_snd.sb_cc) 831 flags &= ~TH_FIN; 832 win = 1; 833 } else { 834 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 835 tp->t_rxtshift = 0; 836 } 837 } 838 839 if (sack_rxmit == 0) { 840 if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) { 841 long cwin; 842 843 /* 844 * We are inside of a SACK recovery episode and are 845 * sending new data, having retransmitted all the 846 * data possible in the scoreboard. 847 */ 848 if (tp->snd_wnd < so->so_snd.sb_cc) { 849 len = tp->snd_wnd - off; 850 flags &= ~TH_FIN; 851 } else { 852 len = so->so_snd.sb_cc - off; 853 } 854 855 /* 856 * From FreeBSD: 857 * Don't remove this (len > 0) check ! 858 * We explicitly check for len > 0 here (although it 859 * isn't really necessary), to work around a gcc 860 * optimization issue - to force gcc to compute 861 * len above. Without this check, the computation 862 * of len is bungled by the optimizer. 863 */ 864 if (len > 0) { 865 cwin = tp->snd_cwnd - 866 (tp->snd_nxt - tp->sack_newdata) - 867 sack_bytes_rxmt; 868 if (cwin < 0) 869 cwin = 0; 870 if (cwin < len) { 871 len = cwin; 872 flags &= ~TH_FIN; 873 } 874 } 875 } else if (win < so->so_snd.sb_cc) { 876 len = win - off; 877 flags &= ~TH_FIN; 878 } else { 879 len = so->so_snd.sb_cc - off; 880 } 881 } 882 883 if (len < 0) { 884 /* 885 * If FIN has been sent but not acked, 886 * but we haven't been called to retransmit, 887 * len will be -1. Otherwise, window shrank 888 * after we sent into it. If window shrank to 0, 889 * cancel pending retransmit, pull snd_nxt back 890 * to (closed) window, and set the persist timer 891 * if it isn't already going. If the window didn't 892 * close completely, just wait for an ACK. 893 * 894 * If we have a pending FIN, either it has already been 895 * transmitted or it is outside the window, so drop it. 896 * If the FIN has been transmitted, but this is not a 897 * retransmission, then len must be -1. Therefore we also 898 * prevent here the sending of `gratuitous FINs'. This 899 * eliminates the need to check for that case below (e.g. 900 * to back up snd_nxt before the FIN so that the sequence 901 * number is correct). 902 */ 903 len = 0; 904 flags &= ~TH_FIN; 905 if (win == 0) { 906 TCP_TIMER_DISARM(tp, TCPT_REXMT); 907 tp->t_rxtshift = 0; 908 tp->snd_nxt = tp->snd_una; 909 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) 910 tcp_setpersist(tp); 911 } 912 } 913 914 /* 915 * Automatic sizing enables the performance of large buffers 916 * and most of the efficiency of small ones by only allocating 917 * space when it is needed. 918 * 919 * The criteria to step up the send buffer one notch are: 920 * 1. receive window of remote host is larger than send buffer 921 * (with a fudge factor of 5/4th); 922 * 2. send buffer is filled to 7/8th with data (so we actually 923 * have data to make use of it); 924 * 3. send buffer fill has not hit maximal automatic size; 925 * 4. our send window (slow start and cogestion controlled) is 926 * larger than sent but unacknowledged data in send buffer. 927 * 928 * The remote host receive window scaling factor may limit the 929 * growing of the send buffer before it reaches its allowed 930 * maximum. 931 * 932 * It scales directly with slow start or congestion window 933 * and does at most one step per received ACK. This fast 934 * scaling has the drawback of growing the send buffer beyond 935 * what is strictly necessary to make full use of a given 936 * delay*bandwith product. However testing has shown this not 937 * to be much of an problem. At worst we are trading wasting 938 * of available bandwith (the non-use of it) for wasting some 939 * socket buffer memory. 940 * 941 * TODO: Shrink send buffer during idle periods together 942 * with congestion window. Requires another timer. 943 */ 944 if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 945 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 946 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 947 so->so_snd.sb_cc < tcp_autosndbuf_max && 948 win >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 949 if (!sbreserve(&so->so_snd, 950 min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc, 951 tcp_autosndbuf_max), so)) 952 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 953 } 954 } 955 956 if (len > txsegsize) { 957 if (use_tso) { 958 /* 959 * Truncate TSO transfers to IP_MAXPACKET, and make 960 * sure that we send equal size transfers down the 961 * stack (rather than big-small-big-small-...). 962 */ 963 #ifdef INET6 964 CTASSERT(IPV6_MAXPACKET == IP_MAXPACKET); 965 #endif 966 len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize; 967 if (len <= txsegsize) { 968 use_tso = 0; 969 } 970 } else 971 len = txsegsize; 972 flags &= ~TH_FIN; 973 sendalot = 1; 974 } else 975 use_tso = 0; 976 if (sack_rxmit) { 977 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 978 flags &= ~TH_FIN; 979 } 980 981 win = sbspace(&so->so_rcv); 982 983 /* 984 * Sender silly window avoidance. If connection is idle 985 * and can send all data, a maximum segment, 986 * at least a maximum default-size segment do it, 987 * or are forced, do it; otherwise don't bother. 988 * If peer's buffer is tiny, then send 989 * when window is at least half open. 990 * If retransmitting (possibly after persist timer forced us 991 * to send into a small window), then must resend. 992 */ 993 if (len) { 994 if (len >= txsegsize) 995 goto send; 996 if ((so->so_state & SS_MORETOCOME) == 0 && 997 ((idle || tp->t_flags & TF_NODELAY) && 998 len + off >= so->so_snd.sb_cc)) 999 goto send; 1000 if (tp->t_force) 1001 goto send; 1002 if (len >= tp->max_sndwnd / 2) 1003 goto send; 1004 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1005 goto send; 1006 if (sack_rxmit) 1007 goto send; 1008 } 1009 1010 /* 1011 * Compare available window to amount of window known to peer 1012 * (as advertised window less next expected input). If the 1013 * difference is at least twice the size of the largest segment 1014 * we expect to receive (i.e. two segments) or at least 50% of 1015 * the maximum possible window, then want to send a window update 1016 * to peer. 1017 */ 1018 if (win > 0) { 1019 /* 1020 * "adv" is the amount we can increase the window, 1021 * taking into account that we are limited by 1022 * TCP_MAXWIN << tp->rcv_scale. 1023 */ 1024 long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) - 1025 (tp->rcv_adv - tp->rcv_nxt); 1026 1027 if (adv >= (long) (2 * rxsegsize)) 1028 goto send; 1029 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 1030 goto send; 1031 } 1032 1033 /* 1034 * Send if we owe peer an ACK. 1035 */ 1036 if (tp->t_flags & TF_ACKNOW) 1037 goto send; 1038 if (flags & (TH_SYN|TH_FIN|TH_RST)) 1039 goto send; 1040 if (SEQ_GT(tp->snd_up, tp->snd_una)) 1041 goto send; 1042 /* 1043 * In SACK, it is possible for tcp_output to fail to send a segment 1044 * after the retransmission timer has been turned off. Make sure 1045 * that the retransmission timer is set. 1046 */ 1047 if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) && 1048 !TCP_TIMER_ISARMED(tp, TCPT_REXMT) && 1049 !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1050 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1051 goto just_return; 1052 } 1053 1054 /* 1055 * TCP window updates are not reliable, rather a polling protocol 1056 * using ``persist'' packets is used to insure receipt of window 1057 * updates. The three ``states'' for the output side are: 1058 * idle not doing retransmits or persists 1059 * persisting to move a small or zero window 1060 * (re)transmitting and thereby not persisting 1061 * 1062 * tp->t_timer[TCPT_PERSIST] 1063 * is set when we are in persist state. 1064 * tp->t_force 1065 * is set when we are called to send a persist packet. 1066 * tp->t_timer[TCPT_REXMT] 1067 * is set when we are retransmitting 1068 * The output side is idle when both timers are zero. 1069 * 1070 * If send window is too small, there is data to transmit, and no 1071 * retransmit or persist is pending, then go to persist state. 1072 * If nothing happens soon, send when timer expires: 1073 * if window is nonzero, transmit what we can, 1074 * otherwise force out a byte. 1075 */ 1076 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1077 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1078 tp->t_rxtshift = 0; 1079 tcp_setpersist(tp); 1080 } 1081 1082 /* 1083 * No reason to send a segment, just return. 1084 */ 1085 just_return: 1086 TCP_REASS_UNLOCK(tp); 1087 return (0); 1088 1089 send: 1090 /* 1091 * Before ESTABLISHED, force sending of initial options 1092 * unless TCP set not to do any options. 1093 * NOTE: we assume that the IP/TCP header plus TCP options 1094 * always fit in a single mbuf, leaving room for a maximum 1095 * link header, i.e. 1096 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 1097 */ 1098 optlen = 0; 1099 switch (af) { 1100 #ifdef INET 1101 case AF_INET: 1102 iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 1103 break; 1104 #endif 1105 #ifdef INET6 1106 case AF_INET6: 1107 iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1108 break; 1109 #endif 1110 default: /*pacify gcc*/ 1111 iphdrlen = 0; 1112 break; 1113 } 1114 hdrlen = iphdrlen; 1115 if (flags & TH_SYN) { 1116 struct rtentry *synrt; 1117 1118 synrt = NULL; 1119 #ifdef INET 1120 if (tp->t_inpcb) 1121 synrt = in_pcbrtentry(tp->t_inpcb); 1122 #endif 1123 #ifdef INET6 1124 if (tp->t_in6pcb) 1125 synrt = in6_pcbrtentry(tp->t_in6pcb); 1126 #endif 1127 1128 tp->snd_nxt = tp->iss; 1129 tp->t_ourmss = tcp_mss_to_advertise(synrt != NULL ? 1130 synrt->rt_ifp : NULL, af); 1131 if ((tp->t_flags & TF_NOOPT) == 0 && OPT_FITS(4)) { 1132 opt[0] = TCPOPT_MAXSEG; 1133 opt[1] = 4; 1134 opt[2] = (tp->t_ourmss >> 8) & 0xff; 1135 opt[3] = tp->t_ourmss & 0xff; 1136 optlen = 4; 1137 1138 if ((tp->t_flags & TF_REQ_SCALE) && 1139 ((flags & TH_ACK) == 0 || 1140 (tp->t_flags & TF_RCVD_SCALE)) && 1141 OPT_FITS(4)) { 1142 *((u_int32_t *) (opt + optlen)) = htonl( 1143 TCPOPT_NOP << 24 | 1144 TCPOPT_WINDOW << 16 | 1145 TCPOLEN_WINDOW << 8 | 1146 tp->request_r_scale); 1147 optlen += 4; 1148 } 1149 if (tcp_do_sack && OPT_FITS(4)) { 1150 u_int8_t *cp = (u_int8_t *)(opt + optlen); 1151 1152 cp[0] = TCPOPT_SACK_PERMITTED; 1153 cp[1] = 2; 1154 cp[2] = TCPOPT_NOP; 1155 cp[3] = TCPOPT_NOP; 1156 optlen += 4; 1157 } 1158 } 1159 } 1160 1161 /* 1162 * Send a timestamp and echo-reply if this is a SYN and our side 1163 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 1164 * and our peer have sent timestamps in our SYN's. 1165 */ 1166 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 1167 (flags & TH_RST) == 0 && 1168 ((flags & (TH_SYN|TH_ACK)) == TH_SYN || 1169 (tp->t_flags & TF_RCVD_TSTMP)) && OPT_FITS(TCPOLEN_TSTAMP_APPA)) { 1170 u_int32_t *lp = (u_int32_t *)(opt + optlen); 1171 1172 /* Form timestamp option as shown in appendix A of RFC 1323. */ 1173 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 1174 *lp++ = htonl(TCP_TIMESTAMP(tp)); 1175 *lp = htonl(tp->ts_recent); 1176 optlen += TCPOLEN_TSTAMP_APPA; 1177 1178 /* Set receive buffer autosizing timestamp. */ 1179 if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) 1180 tp->rfbuf_ts = TCP_TIMESTAMP(tp); 1181 } 1182 1183 /* 1184 * Tack on the SACK block if it is necessary. 1185 */ 1186 if (sack_numblks) { 1187 int sack_len; 1188 u_char *bp = (u_char *)(opt + optlen); 1189 u_int32_t *lp = (u_int32_t *)(bp + 4); 1190 struct ipqent *tiqe; 1191 1192 sack_len = sack_numblks * 8 + 2; 1193 if (OPT_FITS(sack_len + 2)) { 1194 bp[0] = TCPOPT_NOP; 1195 bp[1] = TCPOPT_NOP; 1196 bp[2] = TCPOPT_SACK; 1197 bp[3] = sack_len; 1198 if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) { 1199 sack_numblks--; 1200 *lp++ = htonl(tp->rcv_dsack_block.left); 1201 *lp++ = htonl(tp->rcv_dsack_block.right); 1202 tp->rcv_sack_flags &= ~TCPSACK_HAVED; 1203 } 1204 for (tiqe = TAILQ_FIRST(&tp->timeq); 1205 sack_numblks > 0; 1206 tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) { 1207 KASSERT(tiqe != NULL); 1208 sack_numblks--; 1209 *lp++ = htonl(tiqe->ipqe_seq); 1210 *lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len + 1211 ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0)); 1212 } 1213 optlen += sack_len + 2; 1214 } 1215 } 1216 TCP_REASS_UNLOCK(tp); 1217 1218 #ifdef TCP_SIGNATURE 1219 if ((tp->t_flags & TF_SIGNATURE) && OPT_FITS(TCPOLEN_SIGNATURE + 2)) { 1220 u_char *bp; 1221 /* 1222 * Initialize TCP-MD5 option (RFC2385) 1223 */ 1224 bp = (u_char *)opt + optlen; 1225 *bp++ = TCPOPT_SIGNATURE; 1226 *bp++ = TCPOLEN_SIGNATURE; 1227 sigoff = optlen + 2; 1228 memset(bp, 0, TCP_SIGLEN); 1229 bp += TCP_SIGLEN; 1230 optlen += TCPOLEN_SIGNATURE; 1231 /* 1232 * Terminate options list and maintain 32-bit alignment. 1233 */ 1234 *bp++ = TCPOPT_NOP; 1235 *bp++ = TCPOPT_EOL; 1236 optlen += 2; 1237 } 1238 #endif /* TCP_SIGNATURE */ 1239 1240 hdrlen += optlen; 1241 1242 #ifdef DIAGNOSTIC 1243 if (!use_tso && len > txsegsize) 1244 panic("tcp data to be sent is larger than segment"); 1245 else if (use_tso && len > IP_MAXPACKET) 1246 panic("tcp data to be sent is larger than max TSO size"); 1247 if (max_linkhdr + hdrlen > MCLBYTES) 1248 panic("tcphdr too big"); 1249 #endif 1250 1251 /* 1252 * Grab a header mbuf, attaching a copy of data to 1253 * be transmitted, and initialize the header from 1254 * the template for sends on this connection. 1255 */ 1256 if (len) { 1257 error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m); 1258 if (error) 1259 goto out; 1260 /* 1261 * If we're sending everything we've got, set PUSH. 1262 * (This will keep happy those implementations which only 1263 * give data to the user when a buffer fills or 1264 * a PUSH comes in.) 1265 */ 1266 if (off + len == so->so_snd.sb_cc) 1267 flags |= TH_PUSH; 1268 } else { 1269 tcps = TCP_STAT_GETREF(); 1270 if (tp->t_flags & TF_ACKNOW) 1271 tcps[TCP_STAT_SNDACKS]++; 1272 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1273 tcps[TCP_STAT_SNDCTRL]++; 1274 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1275 tcps[TCP_STAT_SNDURG]++; 1276 else 1277 tcps[TCP_STAT_SNDWINUP]++; 1278 TCP_STAT_PUTREF(); 1279 1280 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1281 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 1282 MCLGET(m, M_DONTWAIT); 1283 if ((m->m_flags & M_EXT) == 0) { 1284 m_freem(m); 1285 m = NULL; 1286 } 1287 } 1288 if (m == NULL) { 1289 error = ENOBUFS; 1290 goto out; 1291 } 1292 MCLAIM(m, &tcp_tx_mowner); 1293 m->m_data += max_linkhdr; 1294 m->m_len = hdrlen; 1295 } 1296 m->m_pkthdr.rcvif = NULL; 1297 switch (af) { 1298 #ifdef INET 1299 case AF_INET: 1300 ip = mtod(m, struct ip *); 1301 #ifdef INET6 1302 ip6 = NULL; 1303 #endif 1304 th = (struct tcphdr *)(ip + 1); 1305 break; 1306 #endif 1307 #ifdef INET6 1308 case AF_INET6: 1309 ip = NULL; 1310 ip6 = mtod(m, struct ip6_hdr *); 1311 th = (struct tcphdr *)(ip6 + 1); 1312 break; 1313 #endif 1314 default: /*pacify gcc*/ 1315 ip = NULL; 1316 #ifdef INET6 1317 ip6 = NULL; 1318 #endif 1319 th = NULL; 1320 break; 1321 } 1322 if (tp->t_template == 0) 1323 panic("tcp_output"); 1324 if (tp->t_template->m_len < iphdrlen) 1325 panic("tcp_output"); 1326 bcopy(mtod(tp->t_template, void *), mtod(m, void *), iphdrlen); 1327 1328 /* 1329 * If we are starting a connection, send ECN setup 1330 * SYN packet. If we are on a retransmit, we may 1331 * resend those bits a number of times as per 1332 * RFC 3168. 1333 */ 1334 if (tp->t_state == TCPS_SYN_SENT && tcp_do_ecn) { 1335 if (tp->t_flags & TF_SYN_REXMT) { 1336 if (tp->t_ecn_retries--) 1337 flags |= TH_ECE|TH_CWR; 1338 } else { 1339 flags |= TH_ECE|TH_CWR; 1340 tp->t_ecn_retries = tcp_ecn_maxretries; 1341 } 1342 } 1343 1344 if (TCP_ECN_ALLOWED(tp)) { 1345 /* 1346 * If the peer has ECN, mark data packets 1347 * ECN capable. Ignore pure ack packets, retransmissions 1348 * and window probes. 1349 */ 1350 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1351 !(tp->t_force && len == 1)) { 1352 ecn_tos = IPTOS_ECN_ECT0; 1353 TCP_STATINC(TCP_STAT_ECN_ECT); 1354 } 1355 1356 /* 1357 * Reply with proper ECN notifications. 1358 */ 1359 if (tp->t_flags & TF_ECN_SND_CWR) { 1360 flags |= TH_CWR; 1361 tp->t_flags &= ~TF_ECN_SND_CWR; 1362 } 1363 if (tp->t_flags & TF_ECN_SND_ECE) { 1364 flags |= TH_ECE; 1365 } 1366 } 1367 1368 1369 /* 1370 * If we are doing retransmissions, then snd_nxt will 1371 * not reflect the first unsent octet. For ACK only 1372 * packets, we do not want the sequence number of the 1373 * retransmitted packet, we want the sequence number 1374 * of the next unsent octet. So, if there is no data 1375 * (and no SYN or FIN), use snd_max instead of snd_nxt 1376 * when filling in ti_seq. But if we are in persist 1377 * state, snd_max might reflect one byte beyond the 1378 * right edge of the window, so use snd_nxt in that 1379 * case, since we know we aren't doing a retransmission. 1380 * (retransmit and persist are mutually exclusive...) 1381 */ 1382 if (TCP_SACK_ENABLED(tp) && sack_rxmit) { 1383 th->th_seq = htonl(p->rxmit); 1384 p->rxmit += len; 1385 } else { 1386 if (len || (flags & (TH_SYN|TH_FIN)) || 1387 TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) 1388 th->th_seq = htonl(tp->snd_nxt); 1389 else 1390 th->th_seq = htonl(tp->snd_max); 1391 } 1392 th->th_ack = htonl(tp->rcv_nxt); 1393 if (optlen) { 1394 bcopy((void *)opt, (void *)(th + 1), optlen); 1395 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1396 } 1397 th->th_flags = flags; 1398 /* 1399 * Calculate receive window. Don't shrink window, 1400 * but avoid silly window syndrome. 1401 */ 1402 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize) 1403 win = 0; 1404 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 1405 win = (long)TCP_MAXWIN << tp->rcv_scale; 1406 if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt)) 1407 win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt); 1408 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale)); 1409 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1410 u_int32_t urp = tp->snd_up - tp->snd_nxt; 1411 if (urp > IP_MAXPACKET) 1412 urp = IP_MAXPACKET; 1413 th->th_urp = htons((u_int16_t)urp); 1414 th->th_flags |= TH_URG; 1415 } else 1416 /* 1417 * If no urgent pointer to send, then we pull 1418 * the urgent pointer to the left edge of the send window 1419 * so that it doesn't drift into the send window on sequence 1420 * number wraparound. 1421 */ 1422 tp->snd_up = tp->snd_una; /* drag it along */ 1423 1424 #ifdef TCP_SIGNATURE 1425 if (sigoff && (tp->t_flags & TF_SIGNATURE)) { 1426 struct secasvar *sav; 1427 u_int8_t *sigp; 1428 1429 sav = tcp_signature_getsav(m, th); 1430 1431 if (sav == NULL) { 1432 if (m) 1433 m_freem(m); 1434 return (EPERM); 1435 } 1436 1437 m->m_pkthdr.len = hdrlen + len; 1438 sigp = (char *)th + sizeof(*th) + sigoff; 1439 tcp_signature(m, th, (char *)th - mtod(m, char *), sav, sigp); 1440 1441 key_sa_recordxfer(sav, m); 1442 #ifdef FAST_IPSEC 1443 KEY_FREESAV(&sav); 1444 #else 1445 key_freesav(sav); 1446 #endif 1447 } 1448 #endif 1449 1450 /* 1451 * Set ourselves up to be checksummed just before the packet 1452 * hits the wire. 1453 */ 1454 switch (af) { 1455 #ifdef INET 1456 case AF_INET: 1457 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1458 if (use_tso) { 1459 m->m_pkthdr.segsz = txsegsize; 1460 m->m_pkthdr.csum_flags = M_CSUM_TSOv4; 1461 } else { 1462 m->m_pkthdr.csum_flags = M_CSUM_TCPv4; 1463 if (len + optlen) { 1464 /* Fixup the pseudo-header checksum. */ 1465 /* XXXJRT Not IP Jumbogram safe. */ 1466 th->th_sum = in_cksum_addword(th->th_sum, 1467 htons((u_int16_t) (len + optlen))); 1468 } 1469 } 1470 break; 1471 #endif 1472 #ifdef INET6 1473 case AF_INET6: 1474 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1475 if (use_tso) { 1476 m->m_pkthdr.segsz = txsegsize; 1477 m->m_pkthdr.csum_flags = M_CSUM_TSOv6; 1478 } else { 1479 m->m_pkthdr.csum_flags = M_CSUM_TCPv6; 1480 if (len + optlen) { 1481 /* Fixup the pseudo-header checksum. */ 1482 /* XXXJRT: Not IPv6 Jumbogram safe. */ 1483 th->th_sum = in_cksum_addword(th->th_sum, 1484 htons((u_int16_t) (len + optlen))); 1485 } 1486 } 1487 break; 1488 #endif 1489 } 1490 1491 /* 1492 * In transmit state, time the transmission and arrange for 1493 * the retransmit. In persist state, just set snd_max. 1494 */ 1495 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1496 tcp_seq startseq = tp->snd_nxt; 1497 1498 /* 1499 * Advance snd_nxt over sequence space of this segment. 1500 * There are no states in which we send both a SYN and a FIN, 1501 * so we collapse the tests for these flags. 1502 */ 1503 if (flags & (TH_SYN|TH_FIN)) 1504 tp->snd_nxt++; 1505 if (sack_rxmit) 1506 goto timer; 1507 tp->snd_nxt += len; 1508 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1509 tp->snd_max = tp->snd_nxt; 1510 /* 1511 * Time this transmission if not a retransmission and 1512 * not currently timing anything. 1513 */ 1514 if (tp->t_rtttime == 0) { 1515 tp->t_rtttime = tcp_now; 1516 tp->t_rtseq = startseq; 1517 TCP_STATINC(TCP_STAT_SEGSTIMED); 1518 } 1519 } 1520 1521 /* 1522 * Set retransmit timer if not currently set, 1523 * and not doing an ack or a keep-alive probe. 1524 * Initial value for retransmit timer is smoothed 1525 * round-trip time + 2 * round-trip time variance. 1526 * Initialize shift counter which is used for backoff 1527 * of retransmit time. 1528 */ 1529 timer: 1530 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1531 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1532 tp->snd_nxt != tp->snd_una)) { 1533 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1534 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 1535 tp->t_rxtshift = 0; 1536 } 1537 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1538 } 1539 } else 1540 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 1541 tp->snd_max = tp->snd_nxt + len; 1542 1543 #ifdef TCP_DEBUG 1544 /* 1545 * Trace. 1546 */ 1547 if (so->so_options & SO_DEBUG) 1548 tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0); 1549 #endif 1550 1551 /* 1552 * Fill in IP length and desired time to live and 1553 * send to IP level. There should be a better way 1554 * to handle ttl and tos; we could keep them in 1555 * the template, but need a way to checksum without them. 1556 */ 1557 m->m_pkthdr.len = hdrlen + len; 1558 1559 switch (af) { 1560 #ifdef INET 1561 case AF_INET: 1562 ip->ip_len = htons(m->m_pkthdr.len); 1563 packetlen = m->m_pkthdr.len; 1564 if (tp->t_inpcb) { 1565 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl; 1566 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos | ecn_tos; 1567 } 1568 #ifdef INET6 1569 else if (tp->t_in6pcb) { 1570 ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/ 1571 ip->ip_tos = ecn_tos; /*XXX*/ 1572 } 1573 #endif 1574 break; 1575 #endif 1576 #ifdef INET6 1577 case AF_INET6: 1578 packetlen = m->m_pkthdr.len; 1579 ip6->ip6_nxt = IPPROTO_TCP; 1580 if (tp->t_in6pcb) { 1581 /* 1582 * we separately set hoplimit for every segment, since 1583 * the user might want to change the value via 1584 * setsockopt. Also, desired default hop limit might 1585 * be changed via Neighbor Discovery. 1586 */ 1587 ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb, 1588 (rt = rtcache_validate(ro)) != NULL ? rt->rt_ifp 1589 : NULL); 1590 } 1591 ip6->ip6_flow |= htonl(ecn_tos << 20); 1592 /* ip6->ip6_flow = ??? (from template) */ 1593 /* ip6_plen will be filled in ip6_output(). */ 1594 break; 1595 #endif 1596 default: /*pacify gcc*/ 1597 packetlen = 0; 1598 break; 1599 } 1600 1601 switch (af) { 1602 #ifdef INET 1603 case AF_INET: 1604 { 1605 struct mbuf *opts; 1606 1607 if (tp->t_inpcb) 1608 opts = tp->t_inpcb->inp_options; 1609 else 1610 opts = NULL; 1611 error = ip_output(m, opts, ro, 1612 (tp->t_mtudisc ? IP_MTUDISC : 0) | 1613 (so->so_options & SO_DONTROUTE), NULL, so); 1614 break; 1615 } 1616 #endif 1617 #ifdef INET6 1618 case AF_INET6: 1619 { 1620 struct ip6_pktopts *opts; 1621 1622 if (tp->t_in6pcb) 1623 opts = tp->t_in6pcb->in6p_outputopts; 1624 else 1625 opts = NULL; 1626 error = ip6_output(m, opts, ro, so->so_options & SO_DONTROUTE, 1627 NULL, so, NULL); 1628 break; 1629 } 1630 #endif 1631 default: 1632 error = EAFNOSUPPORT; 1633 break; 1634 } 1635 if (error) { 1636 out: 1637 if (error == ENOBUFS) { 1638 TCP_STATINC(TCP_STAT_SELFQUENCH); 1639 #ifdef INET 1640 if (tp->t_inpcb) 1641 tcp_quench(tp->t_inpcb, 0); 1642 #endif 1643 #ifdef INET6 1644 if (tp->t_in6pcb) 1645 tcp6_quench(tp->t_in6pcb, 0); 1646 #endif 1647 error = 0; 1648 } else if ((error == EHOSTUNREACH || error == ENETDOWN) && 1649 TCPS_HAVERCVDSYN(tp->t_state)) { 1650 tp->t_softerror = error; 1651 error = 0; 1652 } 1653 1654 /* Back out the seqence number advance. */ 1655 if (sack_rxmit) 1656 p->rxmit -= len; 1657 1658 /* Restart the delayed ACK timer, if necessary. */ 1659 if (tp->t_flags & TF_DELACK) 1660 TCP_RESTART_DELACK(tp); 1661 1662 return (error); 1663 } 1664 1665 if (packetlen > tp->t_pmtud_mtu_sent) 1666 tp->t_pmtud_mtu_sent = packetlen; 1667 1668 tcps = TCP_STAT_GETREF(); 1669 tcps[TCP_STAT_SNDTOTAL]++; 1670 if (tp->t_flags & TF_DELACK) 1671 tcps[TCP_STAT_DELACK]++; 1672 TCP_STAT_PUTREF(); 1673 1674 /* 1675 * Data sent (as far as we can tell). 1676 * If this advertises a larger window than any other segment, 1677 * then remember the size of the advertised window. 1678 * Any pending ACK has now been sent. 1679 */ 1680 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv)) 1681 tp->rcv_adv = tp->rcv_nxt + win; 1682 tp->last_ack_sent = tp->rcv_nxt; 1683 tp->t_flags &= ~TF_ACKNOW; 1684 TCP_CLEAR_DELACK(tp); 1685 #ifdef DIAGNOSTIC 1686 if (maxburst < 0) 1687 printf("tcp_output: maxburst exceeded by %d\n", -maxburst); 1688 #endif 1689 if (sendalot && (tp->t_congctl == &tcp_reno_ctl || --maxburst)) 1690 goto again; 1691 return (0); 1692 } 1693 1694 void 1695 tcp_setpersist(struct tcpcb *tp) 1696 { 1697 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2); 1698 int nticks; 1699 1700 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT)) 1701 panic("tcp_output REXMT"); 1702 /* 1703 * Start/restart persistance timer. 1704 */ 1705 if (t < tp->t_rttmin) 1706 t = tp->t_rttmin; 1707 TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift], 1708 TCPTV_PERSMIN, TCPTV_PERSMAX); 1709 TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks); 1710 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1711 tp->t_rxtshift++; 1712 } 1713
Indexes created Tue Sep 23 07:09:52 GMT 2025