tcp_output.c revision 1.127
1 /* $NetBSD: tcp_output.c,v 1.127 2005/03/16 00:38:27 yamt 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 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 * 82 * Redistribution and use in source and binary forms, with or without 83 * modification, are permitted provided that the following conditions 84 * are met: 85 * 1. Redistributions of source code must retain the above copyright 86 * notice, this list of conditions and the following disclaimer. 87 * 2. Redistributions in binary form must reproduce the above copyright 88 * notice, this list of conditions and the following disclaimer in the 89 * documentation and/or other materials provided with the distribution. 90 * 3. All advertising materials mentioning features or use of this software 91 * must display the following acknowledgement: 92 * This product includes software developed by the NetBSD 93 * Foundation, Inc. and its contributors. 94 * 4. Neither the name of The NetBSD Foundation nor the names of its 95 * contributors may be used to endorse or promote products derived 96 * from this software without specific prior written permission. 97 * 98 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 99 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 100 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 101 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 102 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 103 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 104 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 105 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 106 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 107 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 108 * POSSIBILITY OF SUCH DAMAGE. 109 */ 110 111 /* 112 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 113 * The Regents of the University of California. All rights reserved. 114 * 115 * Redistribution and use in source and binary forms, with or without 116 * modification, are permitted provided that the following conditions 117 * are met: 118 * 1. Redistributions of source code must retain the above copyright 119 * notice, this list of conditions and the following disclaimer. 120 * 2. Redistributions in binary form must reproduce the above copyright 121 * notice, this list of conditions and the following disclaimer in the 122 * documentation and/or other materials provided with the distribution. 123 * 3. Neither the name of the University nor the names of its contributors 124 * may be used to endorse or promote products derived from this software 125 * without specific prior written permission. 126 * 127 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 128 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 129 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 130 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 131 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 132 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 133 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 134 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 135 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 136 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 137 * SUCH DAMAGE. 138 * 139 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 140 */ 141 142 #include <sys/cdefs.h> 143 __KERNEL_RCSID(0, "$NetBSD: tcp_output.c,v 1.127 2005/03/16 00:38:27 yamt Exp $"); 144 145 #include "opt_inet.h" 146 #include "opt_ipsec.h" 147 #include "opt_tcp_debug.h" 148 149 #include <sys/param.h> 150 #include <sys/systm.h> 151 #include <sys/malloc.h> 152 #include <sys/mbuf.h> 153 #include <sys/protosw.h> 154 #include <sys/socket.h> 155 #include <sys/socketvar.h> 156 #include <sys/errno.h> 157 #include <sys/domain.h> 158 #include <sys/kernel.h> 159 #ifdef TCP_SIGNATURE 160 #include <sys/md5.h> 161 #endif 162 163 #include <net/if.h> 164 #include <net/route.h> 165 166 #include <netinet/in.h> 167 #include <netinet/in_systm.h> 168 #include <netinet/ip.h> 169 #include <netinet/in_pcb.h> 170 #include <netinet/ip_var.h> 171 172 #ifdef INET6 173 #ifndef INET 174 #include <netinet/in.h> 175 #endif 176 #include <netinet/ip6.h> 177 #include <netinet6/in6_var.h> 178 #include <netinet6/ip6_var.h> 179 #include <netinet6/in6_pcb.h> 180 #include <netinet6/nd6.h> 181 #endif 182 183 #ifdef FAST_IPSEC 184 #include <netipsec/ipsec.h> 185 #include <netipsec/key.h> 186 #endif /* FAST_IPSEC*/ 187 #ifdef IPSEC 188 #include <netinet6/ipsec.h> 189 #endif 190 191 #include <netinet/tcp.h> 192 #define TCPOUTFLAGS 193 #include <netinet/tcp_fsm.h> 194 #include <netinet/tcp_seq.h> 195 #include <netinet/tcp_timer.h> 196 #include <netinet/tcp_var.h> 197 #include <netinet/tcpip.h> 198 #include <netinet/tcp_debug.h> 199 200 #ifdef IPSEC 201 #include <netkey/key.h> 202 #endif 203 204 #ifdef notyet 205 extern struct mbuf *m_copypack(); 206 #endif 207 208 /* 209 * Knob to enable Congestion Window Monitoring, and control the 210 * the burst size it allows. Default burst is 4 packets, per 211 * the Internet draft. 212 */ 213 int tcp_cwm = 0; 214 int tcp_cwm_burstsize = 4; 215 216 #ifdef TCP_OUTPUT_COUNTERS 217 #include <sys/device.h> 218 219 extern struct evcnt tcp_output_bigheader; 220 extern struct evcnt tcp_output_predict_hit; 221 extern struct evcnt tcp_output_predict_miss; 222 extern struct evcnt tcp_output_copysmall; 223 extern struct evcnt tcp_output_copybig; 224 extern struct evcnt tcp_output_refbig; 225 226 #define TCP_OUTPUT_COUNTER_INCR(ev) (ev)->ev_count++ 227 #else 228 229 #define TCP_OUTPUT_COUNTER_INCR(ev) /* nothing */ 230 231 #endif /* TCP_OUTPUT_COUNTERS */ 232 233 static 234 #ifndef GPROF 235 __inline 236 #endif 237 int 238 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep) 239 { 240 #ifdef INET 241 struct inpcb *inp = tp->t_inpcb; 242 #endif 243 #ifdef INET6 244 struct in6pcb *in6p = tp->t_in6pcb; 245 #endif 246 struct socket *so = NULL; 247 struct rtentry *rt; 248 struct ifnet *ifp; 249 int size; 250 int iphlen; 251 int optlen; 252 253 #ifdef DIAGNOSTIC 254 if (tp->t_inpcb && tp->t_in6pcb) 255 panic("tcp_segsize: both t_inpcb and t_in6pcb are set"); 256 #endif 257 switch (tp->t_family) { 258 #ifdef INET 259 case AF_INET: 260 iphlen = sizeof(struct ip); 261 break; 262 #endif 263 #ifdef INET6 264 case AF_INET6: 265 iphlen = sizeof(struct ip6_hdr); 266 break; 267 #endif 268 default: 269 size = tcp_mssdflt; 270 goto out; 271 } 272 273 rt = NULL; 274 #ifdef INET 275 if (inp) { 276 rt = in_pcbrtentry(inp); 277 so = inp->inp_socket; 278 } 279 #endif 280 #ifdef INET6 281 if (in6p) { 282 rt = in6_pcbrtentry(in6p); 283 so = in6p->in6p_socket; 284 } 285 #endif 286 if (rt == NULL) { 287 size = tcp_mssdflt; 288 goto out; 289 } 290 291 ifp = rt->rt_ifp; 292 293 size = tcp_mssdflt; 294 if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) { 295 #ifdef INET6 296 if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) { 297 /* 298 * RFC2460 section 5, last paragraph: if path MTU is 299 * smaller than 1280, use 1280 as packet size and 300 * attach fragment header. 301 */ 302 size = IPV6_MMTU - iphlen - sizeof(struct ip6_frag) - 303 sizeof(struct tcphdr); 304 } else 305 size = rt->rt_rmx.rmx_mtu - iphlen - 306 sizeof(struct tcphdr); 307 #else 308 size = rt->rt_rmx.rmx_mtu - iphlen - sizeof(struct tcphdr); 309 #endif 310 } else if (ifp->if_flags & IFF_LOOPBACK) 311 size = ifp->if_mtu - iphlen - sizeof(struct tcphdr); 312 #ifdef INET 313 else if (inp && tp->t_mtudisc) 314 size = ifp->if_mtu - iphlen - sizeof(struct tcphdr); 315 else if (inp && in_localaddr(inp->inp_faddr)) 316 size = ifp->if_mtu - iphlen - sizeof(struct tcphdr); 317 #endif 318 #ifdef INET6 319 else if (in6p) { 320 #ifdef INET 321 if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) { 322 /* mapped addr case */ 323 struct in_addr d; 324 bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d)); 325 if (tp->t_mtudisc || in_localaddr(d)) 326 size = ifp->if_mtu - iphlen - sizeof(struct tcphdr); 327 } else 328 #endif 329 { 330 /* 331 * for IPv6, path MTU discovery is always turned on, 332 * or the node must use packet size <= 1280. 333 */ 334 size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU; 335 size -= (iphlen + sizeof(struct tcphdr)); 336 } 337 } 338 #endif 339 out: 340 /* 341 * Now we must make room for whatever extra TCP/IP options are in 342 * the packet. 343 */ 344 optlen = tcp_optlen(tp); 345 346 /* 347 * XXX tp->t_ourmss should have the right size, but without this code 348 * fragmentation will occur... need more investigation 349 */ 350 #ifdef INET 351 if (inp) { 352 #if defined(IPSEC) || defined(FAST_IPSEC) 353 if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND)) 354 optlen += ipsec4_hdrsiz_tcp(tp); 355 #endif 356 optlen += ip_optlen(inp); 357 } 358 #endif 359 #ifdef INET6 360 #ifdef INET 361 if (in6p && tp->t_family == AF_INET) { 362 #if defined(IPSEC) || defined(FAST_IPSEC) 363 if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND)) 364 optlen += ipsec4_hdrsiz_tcp(tp); 365 #endif 366 /* XXX size -= ip_optlen(in6p); */ 367 } else 368 #endif 369 if (in6p && tp->t_family == AF_INET6) { 370 #ifdef IPSEC 371 if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND)) 372 optlen += ipsec6_hdrsiz_tcp(tp); 373 #endif 374 optlen += ip6_optlen(in6p); 375 } 376 #endif 377 size -= optlen; 378 379 /* there may not be any room for data if mtu is too small */ 380 if (size < 0) 381 return (EMSGSIZE); 382 383 /* 384 * *rxsegsizep holds *estimated* inbound segment size (estimation 385 * assumes that path MTU is the same for both ways). this is only 386 * for silly window avoidance, do not use the value for other purposes. 387 * 388 * ipseclen is subtracted from both sides, this may not be right. 389 * I'm not quite sure about this (could someone comment). 390 */ 391 *txsegsizep = min(tp->t_peermss - optlen, size); 392 /* 393 * Never send more than half a buffer full. This insures that we can 394 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and 395 * therefore acks will never be delayed unless we run out of data to 396 * transmit. 397 */ 398 if (so) 399 *txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep); 400 *rxsegsizep = min(tp->t_ourmss - optlen, size); 401 402 if (*txsegsizep != tp->t_segsz) { 403 /* 404 * If the new segment size is larger, we don't want to 405 * mess up the congestion window, but if it is smaller 406 * we'll have to reduce the congestion window to ensure 407 * that we don't get into trouble with initial windows 408 * and the rest. In any case, if the segment size 409 * has changed, chances are the path has, too, and 410 * our congestion window will be different. 411 */ 412 if (*txsegsizep < tp->t_segsz) { 413 tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz) 414 * *txsegsizep, *txsegsizep); 415 tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz) 416 * *txsegsizep, *txsegsizep); 417 } 418 tp->t_segsz = *txsegsizep; 419 } 420 421 return (0); 422 } 423 424 static 425 #ifndef GPROF 426 __inline 427 #endif 428 int 429 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off, 430 long len, int hdrlen, struct mbuf **mp) 431 { 432 struct mbuf *m, *m0; 433 434 if (tp->t_force && len == 1) 435 tcpstat.tcps_sndprobe++; 436 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 437 tcpstat.tcps_sndrexmitpack++; 438 tcpstat.tcps_sndrexmitbyte += len; 439 } else { 440 tcpstat.tcps_sndpack++; 441 tcpstat.tcps_sndbyte += len; 442 } 443 #ifdef notyet 444 if ((m = m_copypack(so->so_snd.sb_mb, off, 445 (int)len, max_linkhdr + hdrlen)) == 0) 446 return (ENOBUFS); 447 /* 448 * m_copypack left space for our hdr; use it. 449 */ 450 m->m_len += hdrlen; 451 m->m_data -= hdrlen; 452 #else 453 MGETHDR(m, M_DONTWAIT, MT_HEADER); 454 if (__predict_false(m == NULL)) 455 return (ENOBUFS); 456 MCLAIM(m, &tcp_tx_mowner); 457 458 /* 459 * XXX Because other code assumes headers will fit in 460 * XXX one header mbuf. 461 * 462 * (This code should almost *never* be run.) 463 */ 464 if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) { 465 TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader); 466 MCLGET(m, M_DONTWAIT); 467 if ((m->m_flags & M_EXT) == 0) { 468 m_freem(m); 469 return (ENOBUFS); 470 } 471 } 472 473 m->m_data += max_linkhdr; 474 m->m_len = hdrlen; 475 476 /* 477 * To avoid traversing the whole sb_mb chain for correct 478 * data to send, remember last sent mbuf, its offset and 479 * the sent size. When called the next time, see if the 480 * data to send is directly following the previous transfer. 481 * This is important for large TCP windows. 482 */ 483 if (off == 0 || tp->t_lastm == NULL || 484 (tp->t_lastoff + tp->t_lastlen) != off) { 485 TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss); 486 /* 487 * Either a new packet or a retransmit. 488 * Start from the beginning. 489 */ 490 tp->t_lastm = so->so_snd.sb_mb; 491 tp->t_inoff = off; 492 } else { 493 TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit); 494 tp->t_inoff += tp->t_lastlen; 495 } 496 497 /* Traverse forward to next packet */ 498 while (tp->t_inoff > 0) { 499 if (tp->t_lastm == NULL) 500 panic("tp->t_lastm == NULL"); 501 if (tp->t_inoff < tp->t_lastm->m_len) 502 break; 503 tp->t_inoff -= tp->t_lastm->m_len; 504 tp->t_lastm = tp->t_lastm->m_next; 505 } 506 507 tp->t_lastoff = off; 508 tp->t_lastlen = len; 509 m0 = tp->t_lastm; 510 off = tp->t_inoff; 511 512 if (len <= M_TRAILINGSPACE(m)) { 513 m_copydata(m0, off, (int) len, mtod(m, caddr_t) + hdrlen); 514 m->m_len += len; 515 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall); 516 } else { 517 m->m_next = m_copy(m0, off, (int) len); 518 if (m->m_next == NULL) { 519 m_freem(m); 520 return (ENOBUFS); 521 } 522 #ifdef TCP_OUTPUT_COUNTERS 523 if (m->m_next->m_flags & M_EXT) 524 TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig); 525 else 526 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig); 527 #endif /* TCP_OUTPUT_COUNTERS */ 528 } 529 #endif 530 531 *mp = m; 532 return (0); 533 } 534 535 /* 536 * Tcp output routine: figure out what should be sent and send it. 537 */ 538 int 539 tcp_output(struct tcpcb *tp) 540 { 541 struct socket *so; 542 struct route *ro; 543 long len, win; 544 int off, flags, error; 545 struct mbuf *m; 546 struct ip *ip; 547 #ifdef INET6 548 struct ip6_hdr *ip6; 549 #endif 550 struct tcphdr *th; 551 u_char opt[MAX_TCPOPTLEN]; 552 unsigned optlen, hdrlen; 553 unsigned int sack_optlen; 554 int idle, sendalot, txsegsize, rxsegsize; 555 int txsegsize_nosack; 556 int maxburst = TCP_MAXBURST; 557 int af; /* address family on the wire */ 558 int iphdrlen; 559 int has_tso, use_tso; 560 int sack_rxmit; 561 int sack_bytes_rxmt; 562 struct sackhole *p; 563 #ifdef TCP_SIGNATURE 564 int sigoff = 0; 565 #endif 566 567 #ifdef DIAGNOSTIC 568 if (tp->t_inpcb && tp->t_in6pcb) 569 panic("tcp_output: both t_inpcb and t_in6pcb are set"); 570 #endif 571 so = NULL; 572 ro = NULL; 573 if (tp->t_inpcb) { 574 so = tp->t_inpcb->inp_socket; 575 ro = &tp->t_inpcb->inp_route; 576 } 577 #ifdef INET6 578 else if (tp->t_in6pcb) { 579 so = tp->t_in6pcb->in6p_socket; 580 ro = (struct route *)&tp->t_in6pcb->in6p_route; 581 } 582 #endif 583 584 switch (af = tp->t_family) { 585 #ifdef INET 586 case AF_INET: 587 if (tp->t_inpcb) 588 break; 589 #ifdef INET6 590 /* mapped addr case */ 591 if (tp->t_in6pcb) 592 break; 593 #endif 594 return (EINVAL); 595 #endif 596 #ifdef INET6 597 case AF_INET6: 598 if (tp->t_in6pcb) 599 break; 600 return (EINVAL); 601 #endif 602 default: 603 return (EAFNOSUPPORT); 604 } 605 606 if (tcp_segsize(tp, &txsegsize, &rxsegsize)) 607 return (EMSGSIZE); 608 609 idle = (tp->snd_max == tp->snd_una); 610 611 /* 612 * Determine if we can use TCP segmentation offload: 613 * - If we're using IPv4 614 * - If there is not an IPsec policy that prevents it 615 * - If the interface can do it 616 */ 617 has_tso = tp->t_inpcb != NULL && 618 #if defined(IPSEC) || defined(FAST_IPSEC) 619 IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp, 620 IPSEC_DIR_OUTBOUND) && 621 #endif 622 tp->t_inpcb->inp_route.ro_rt != NULL && 623 (tp->t_inpcb->inp_route.ro_rt->rt_ifp->if_capenable & 624 IFCAP_TSOv4) != 0; 625 626 /* 627 * Restart Window computation. From draft-floyd-incr-init-win-03: 628 * 629 * Optionally, a TCP MAY set the restart window to the 630 * minimum of the value used for the initial window and 631 * the current value of cwnd (in other words, using a 632 * larger value for the restart window should never increase 633 * the size of cwnd). 634 */ 635 if (tcp_cwm) { 636 /* 637 * Hughes/Touch/Heidemann Congestion Window Monitoring. 638 * Count the number of packets currently pending 639 * acknowledgement, and limit our congestion window 640 * to a pre-determined allowed burst size plus that count. 641 * This prevents bursting once all pending packets have 642 * been acknowledged (i.e. transmission is idle). 643 * 644 * XXX Link this to Initial Window? 645 */ 646 tp->snd_cwnd = min(tp->snd_cwnd, 647 (tcp_cwm_burstsize * txsegsize) + 648 (tp->snd_nxt - tp->snd_una)); 649 } else { 650 if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) { 651 /* 652 * We have been idle for "a while" and no acks are 653 * expected to clock out any data we send -- 654 * slow start to get ack "clock" running again. 655 */ 656 int ss = tcp_init_win; 657 #ifdef INET 658 if (tp->t_inpcb && 659 in_localaddr(tp->t_inpcb->inp_faddr)) 660 ss = tcp_init_win_local; 661 #endif 662 #ifdef INET6 663 if (tp->t_in6pcb && 664 in6_localaddr(&tp->t_in6pcb->in6p_faddr)) 665 ss = tcp_init_win_local; 666 #endif 667 tp->snd_cwnd = min(tp->snd_cwnd, 668 TCP_INITIAL_WINDOW(ss, txsegsize)); 669 } 670 } 671 672 txsegsize_nosack = txsegsize; 673 again: 674 sack_optlen = tcp_sack_optlen(tp); 675 if (sack_optlen && (tp->rcv_sack_flags & TCPSACK_HAVED) != 0) { 676 /* don't duplicate D-SACK. */ 677 use_tso = 0; 678 } else { 679 use_tso = has_tso; 680 } 681 txsegsize = txsegsize_nosack - sack_optlen; 682 683 /* 684 * Determine length of data that should be transmitted, and 685 * flags that should be used. If there is some data or critical 686 * controls (SYN, RST) to send, then transmit; otherwise, 687 * investigate further. 688 * 689 * Readjust SACK information to avoid resending duplicate data. 690 */ 691 if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max)) 692 tcp_sack_adjust(tp); 693 sendalot = 0; 694 off = tp->snd_nxt - tp->snd_una; 695 win = min(tp->snd_wnd, tp->snd_cwnd); 696 697 flags = tcp_outflags[tp->t_state]; 698 699 /* 700 * Send any SACK-generated retransmissions. If we're explicitly trying 701 * to send out new data (when sendalot is 1), bypass this function. 702 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 703 * we're replacing a (future) new transmission with a retransmission 704 * now, and we previously incremented snd_cwnd in tcp_input(). 705 */ 706 /* 707 * Still in sack recovery , reset rxmit flag to zero. 708 */ 709 sack_rxmit = 0; 710 sack_bytes_rxmt = 0; 711 len = 0; 712 p = NULL; 713 do { 714 long cwin; 715 if (!TCP_SACK_ENABLED(tp)) 716 break; 717 if (tp->t_partialacks < 0) 718 break; 719 p = tcp_sack_output(tp, &sack_bytes_rxmt); 720 if (p == NULL) 721 break; 722 723 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 724 if (cwin < 0) 725 cwin = 0; 726 /* Do not retransmit SACK segments beyond snd_recover */ 727 if (SEQ_GT(p->end, tp->snd_recover)) { 728 /* 729 * (At least) part of sack hole extends beyond 730 * snd_recover. Check to see if we can rexmit data 731 * for this hole. 732 */ 733 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 734 /* 735 * Can't rexmit any more data for this hole. 736 * That data will be rexmitted in the next 737 * sack recovery episode, when snd_recover 738 * moves past p->rxmit. 739 */ 740 p = NULL; 741 break; 742 } 743 /* Can rexmit part of the current hole */ 744 len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit)); 745 } else 746 len = ((long)ulmin(cwin, p->end - p->rxmit)); 747 off = p->rxmit - tp->snd_una; 748 if (len > 0) { 749 sack_rxmit = 1; 750 sendalot = 1; 751 } 752 } while (/*CONSTCOND*/0); 753 754 /* 755 * If in persist timeout with window of 0, send 1 byte. 756 * Otherwise, if window is small but nonzero 757 * and timer expired, we will send what we can 758 * and go to transmit state. 759 */ 760 if (tp->t_force) { 761 if (win == 0) { 762 /* 763 * If we still have some data to send, then 764 * clear the FIN bit. Usually this would 765 * happen below when it realizes that we 766 * aren't sending all the data. However, 767 * if we have exactly 1 byte of unset data, 768 * then it won't clear the FIN bit below, 769 * and if we are in persist state, we wind 770 * up sending the packet without recording 771 * that we sent the FIN bit. 772 * 773 * We can't just blindly clear the FIN bit, 774 * because if we don't have any more data 775 * to send then the probe will be the FIN 776 * itself. 777 */ 778 if (off < so->so_snd.sb_cc) 779 flags &= ~TH_FIN; 780 win = 1; 781 } else { 782 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 783 tp->t_rxtshift = 0; 784 } 785 } 786 787 if (!TCP_SACK_ENABLED(tp)) { 788 if (win < so->so_snd.sb_cc) { 789 len = win - off; 790 flags &= ~TH_FIN; 791 } else 792 len = so->so_snd.sb_cc - off; 793 } else if (sack_rxmit == 0) { 794 if (sack_bytes_rxmt != 0) { 795 long cwin; 796 797 /* 798 * We are inside of a SACK recovery episode and are 799 * sending new data, having retransmitted all the 800 * data possible in the scoreboard. 801 */ 802 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 803 - off); 804 /* 805 * From FreeBSD: 806 * Don't remove this (len > 0) check ! 807 * We explicitly check for len > 0 here (although it 808 * isn't really necessary), to work around a gcc 809 * optimization issue - to force gcc to compute 810 * len above. Without this check, the computation 811 * of len is bungled by the optimizer. 812 */ 813 if (len > 0) { 814 cwin = tp->snd_cwnd - 815 (tp->snd_nxt - tp->sack_newdata) - 816 sack_bytes_rxmt; 817 if (cwin < 0) 818 cwin = 0; 819 len = lmin(len, cwin); 820 } 821 } else if (win < so->so_snd.sb_cc) { 822 len = win - off; 823 flags &= ~TH_FIN; 824 } else 825 len = so->so_snd.sb_cc - off; 826 } 827 828 if (len < 0) { 829 /* 830 * If FIN has been sent but not acked, 831 * but we haven't been called to retransmit, 832 * len will be -1. Otherwise, window shrank 833 * after we sent into it. If window shrank to 0, 834 * cancel pending retransmit, pull snd_nxt back 835 * to (closed) window, and set the persist timer 836 * if it isn't already going. If the window didn't 837 * close completely, just wait for an ACK. 838 * 839 * If we have a pending FIN, either it has already been 840 * transmitted or it is outside the window, so drop it. 841 * If the FIN has been transmitted, but this is not a 842 * retransmission, then len must be -1. Therefore we also 843 * prevent here the sending of `gratuitous FINs'. This 844 * eliminates the need to check for that case below (e.g. 845 * to back up snd_nxt before the FIN so that the sequence 846 * number is correct). 847 */ 848 len = 0; 849 flags &= ~TH_FIN; 850 if (win == 0) { 851 TCP_TIMER_DISARM(tp, TCPT_REXMT); 852 tp->t_rxtshift = 0; 853 tp->snd_nxt = tp->snd_una; 854 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) 855 tcp_setpersist(tp); 856 } 857 } 858 if (len > txsegsize) { 859 if (use_tso) { 860 /* 861 * Truncate TSO transfers to IP_MAXPACKET, and make 862 * sure that we send equal size transfers down the 863 * stack (rather than big-small-big-small-...). 864 */ 865 len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize; 866 if (len <= txsegsize) { 867 use_tso = 0; 868 } 869 } else 870 len = txsegsize; 871 flags &= ~TH_FIN; 872 sendalot = 1; 873 } else 874 use_tso = 0; 875 if (sack_rxmit) { 876 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 877 flags &= ~TH_FIN; 878 } 879 880 win = sbspace(&so->so_rcv); 881 882 /* 883 * Sender silly window avoidance. If connection is idle 884 * and can send all data, a maximum segment, 885 * at least a maximum default-size segment do it, 886 * or are forced, do it; otherwise don't bother. 887 * If peer's buffer is tiny, then send 888 * when window is at least half open. 889 * If retransmitting (possibly after persist timer forced us 890 * to send into a small window), then must resend. 891 */ 892 if (len) { 893 if (len >= txsegsize) 894 goto send; 895 if ((so->so_state & SS_MORETOCOME) == 0 && 896 ((idle || tp->t_flags & TF_NODELAY) && 897 len + off >= so->so_snd.sb_cc)) 898 goto send; 899 if (tp->t_force) 900 goto send; 901 if (len >= tp->max_sndwnd / 2) 902 goto send; 903 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 904 goto send; 905 if (sack_rxmit) 906 goto send; 907 } 908 909 /* 910 * Compare available window to amount of window known to peer 911 * (as advertised window less next expected input). If the 912 * difference is at least twice the size of the largest segment 913 * we expect to receive (i.e. two segments) or at least 50% of 914 * the maximum possible window, then want to send a window update 915 * to peer. 916 */ 917 if (win > 0) { 918 /* 919 * "adv" is the amount we can increase the window, 920 * taking into account that we are limited by 921 * TCP_MAXWIN << tp->rcv_scale. 922 */ 923 long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) - 924 (tp->rcv_adv - tp->rcv_nxt); 925 926 if (adv >= (long) (2 * rxsegsize)) 927 goto send; 928 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 929 goto send; 930 } 931 932 /* 933 * Send if we owe peer an ACK. 934 */ 935 if (tp->t_flags & TF_ACKNOW) 936 goto send; 937 if (flags & (TH_SYN|TH_FIN|TH_RST)) 938 goto send; 939 if (SEQ_GT(tp->snd_up, tp->snd_una)) 940 goto send; 941 /* 942 * In SACK, it is possible for tcp_output to fail to send a segment 943 * after the retransmission timer has been turned off. Make sure 944 * that the retransmission timer is set. 945 */ 946 if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) && 947 !TCP_TIMER_ISARMED(tp, TCPT_REXMT) && 948 !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 949 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 950 goto just_return; 951 } 952 953 /* 954 * TCP window updates are not reliable, rather a polling protocol 955 * using ``persist'' packets is used to insure receipt of window 956 * updates. The three ``states'' for the output side are: 957 * idle not doing retransmits or persists 958 * persisting to move a small or zero window 959 * (re)transmitting and thereby not persisting 960 * 961 * tp->t_timer[TCPT_PERSIST] 962 * is set when we are in persist state. 963 * tp->t_force 964 * is set when we are called to send a persist packet. 965 * tp->t_timer[TCPT_REXMT] 966 * is set when we are retransmitting 967 * The output side is idle when both timers are zero. 968 * 969 * If send window is too small, there is data to transmit, and no 970 * retransmit or persist is pending, then go to persist state. 971 * If nothing happens soon, send when timer expires: 972 * if window is nonzero, transmit what we can, 973 * otherwise force out a byte. 974 */ 975 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 976 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 977 tp->t_rxtshift = 0; 978 tcp_setpersist(tp); 979 } 980 981 /* 982 * No reason to send a segment, just return. 983 */ 984 just_return: 985 return (0); 986 987 send: 988 /* 989 * Before ESTABLISHED, force sending of initial options 990 * unless TCP set not to do any options. 991 * NOTE: we assume that the IP/TCP header plus TCP options 992 * always fit in a single mbuf, leaving room for a maximum 993 * link header, i.e. 994 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 995 */ 996 optlen = 0; 997 switch (af) { 998 #ifdef INET 999 case AF_INET: 1000 iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 1001 break; 1002 #endif 1003 #ifdef INET6 1004 case AF_INET6: 1005 iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1006 break; 1007 #endif 1008 default: /*pacify gcc*/ 1009 iphdrlen = 0; 1010 break; 1011 } 1012 hdrlen = iphdrlen; 1013 if (flags & TH_SYN) { 1014 struct rtentry *rt; 1015 1016 rt = NULL; 1017 #ifdef INET 1018 if (tp->t_inpcb) 1019 rt = in_pcbrtentry(tp->t_inpcb); 1020 #endif 1021 #ifdef INET6 1022 if (tp->t_in6pcb) 1023 rt = in6_pcbrtentry(tp->t_in6pcb); 1024 #endif 1025 1026 tp->snd_nxt = tp->iss; 1027 tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ? 1028 rt->rt_ifp : NULL, af); 1029 if ((tp->t_flags & TF_NOOPT) == 0) { 1030 opt[0] = TCPOPT_MAXSEG; 1031 opt[1] = 4; 1032 opt[2] = (tp->t_ourmss >> 8) & 0xff; 1033 opt[3] = tp->t_ourmss & 0xff; 1034 optlen = 4; 1035 1036 if ((tp->t_flags & TF_REQ_SCALE) && 1037 ((flags & TH_ACK) == 0 || 1038 (tp->t_flags & TF_RCVD_SCALE))) { 1039 *((u_int32_t *) (opt + optlen)) = htonl( 1040 TCPOPT_NOP << 24 | 1041 TCPOPT_WINDOW << 16 | 1042 TCPOLEN_WINDOW << 8 | 1043 tp->request_r_scale); 1044 optlen += 4; 1045 } 1046 if (tcp_do_sack) { 1047 u_int8_t *p = (u_int8_t *)(opt + optlen); 1048 1049 p[0] = TCPOPT_SACK_PERMITTED; 1050 p[1] = 2; 1051 p[2] = TCPOPT_NOP; 1052 p[3] = TCPOPT_NOP; 1053 optlen += 4; 1054 } 1055 } 1056 } 1057 1058 /* 1059 * Send a timestamp and echo-reply if this is a SYN and our side 1060 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 1061 * and our peer have sent timestamps in our SYN's. 1062 */ 1063 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 1064 (flags & TH_RST) == 0 && 1065 ((flags & (TH_SYN|TH_ACK)) == TH_SYN || 1066 (tp->t_flags & TF_RCVD_TSTMP))) { 1067 u_int32_t *lp = (u_int32_t *)(opt + optlen); 1068 1069 /* Form timestamp option as shown in appendix A of RFC 1323. */ 1070 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 1071 *lp++ = htonl(TCP_TIMESTAMP(tp)); 1072 *lp = htonl(tp->ts_recent); 1073 optlen += TCPOLEN_TSTAMP_APPA; 1074 } 1075 1076 /* 1077 * Tack on the SACK block if it is necessary. 1078 */ 1079 if (sack_optlen) { 1080 int sack_len, i; 1081 u_char *bp = (u_char *)(opt + optlen); 1082 u_int32_t *lp = (u_int32_t *)(bp + 4); 1083 1084 sack_len = tp->rcv_sack_num * 8 + 2; 1085 bp[0] = TCPOPT_NOP; 1086 bp[1] = TCPOPT_NOP; 1087 bp[2] = TCPOPT_SACK; 1088 bp[3] = sack_len; 1089 for (i = 0; i < tp->rcv_sack_num; i++) { 1090 *lp++ = htonl(tp->rcv_sack_block[i].left); 1091 *lp++ = htonl(tp->rcv_sack_block[i].right); 1092 } 1093 if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) { 1094 tp->rcv_sack_flags &= ~TCPSACK_HAVED; 1095 tcp_update_sack_list(tp); 1096 } 1097 KASSERT(sack_len + 2 == sack_optlen); 1098 optlen += sack_len + 2; 1099 } 1100 1101 #ifdef TCP_SIGNATURE 1102 #if defined(INET6) && defined(FAST_IPSEC) 1103 if (tp->t_family == AF_INET) 1104 #endif 1105 if (tp->t_flags & TF_SIGNATURE) { 1106 u_char *bp; 1107 /* 1108 * Initialize TCP-MD5 option (RFC2385) 1109 */ 1110 bp = (u_char *)opt + optlen; 1111 *bp++ = TCPOPT_SIGNATURE; 1112 *bp++ = TCPOLEN_SIGNATURE; 1113 sigoff = optlen + 2; 1114 bzero(bp, TCP_SIGLEN); 1115 bp += TCP_SIGLEN; 1116 optlen += TCPOLEN_SIGNATURE; 1117 /* 1118 * Terminate options list and maintain 32-bit alignment. 1119 */ 1120 *bp++ = TCPOPT_NOP; 1121 *bp++ = TCPOPT_EOL; 1122 optlen += 2; 1123 } 1124 #endif /* TCP_SIGNATURE */ 1125 1126 hdrlen += optlen; 1127 1128 #ifdef DIAGNOSTIC 1129 if (!use_tso && len > txsegsize) 1130 panic("tcp data to be sent is larger than segment"); 1131 else if (use_tso && len > IP_MAXPACKET) 1132 panic("tcp data to be sent is larger than max TSO size"); 1133 if (max_linkhdr + hdrlen > MCLBYTES) 1134 panic("tcphdr too big"); 1135 #endif 1136 1137 /* 1138 * Grab a header mbuf, attaching a copy of data to 1139 * be transmitted, and initialize the header from 1140 * the template for sends on this connection. 1141 */ 1142 if (len) { 1143 error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m); 1144 if (error) 1145 goto out; 1146 /* 1147 * If we're sending everything we've got, set PUSH. 1148 * (This will keep happy those implementations which only 1149 * give data to the user when a buffer fills or 1150 * a PUSH comes in.) 1151 */ 1152 if (off + len == so->so_snd.sb_cc) 1153 flags |= TH_PUSH; 1154 } else { 1155 if (tp->t_flags & TF_ACKNOW) 1156 tcpstat.tcps_sndacks++; 1157 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1158 tcpstat.tcps_sndctrl++; 1159 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1160 tcpstat.tcps_sndurg++; 1161 else 1162 tcpstat.tcps_sndwinup++; 1163 1164 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1165 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 1166 MCLGET(m, M_DONTWAIT); 1167 if ((m->m_flags & M_EXT) == 0) { 1168 m_freem(m); 1169 m = NULL; 1170 } 1171 } 1172 if (m == NULL) { 1173 error = ENOBUFS; 1174 goto out; 1175 } 1176 MCLAIM(m, &tcp_tx_mowner); 1177 m->m_data += max_linkhdr; 1178 m->m_len = hdrlen; 1179 } 1180 m->m_pkthdr.rcvif = (struct ifnet *)0; 1181 switch (af) { 1182 #ifdef INET 1183 case AF_INET: 1184 ip = mtod(m, struct ip *); 1185 #ifdef INET6 1186 ip6 = NULL; 1187 #endif 1188 th = (struct tcphdr *)(ip + 1); 1189 break; 1190 #endif 1191 #ifdef INET6 1192 case AF_INET6: 1193 ip = NULL; 1194 ip6 = mtod(m, struct ip6_hdr *); 1195 th = (struct tcphdr *)(ip6 + 1); 1196 break; 1197 #endif 1198 default: /*pacify gcc*/ 1199 ip = NULL; 1200 #ifdef INET6 1201 ip6 = NULL; 1202 #endif 1203 th = NULL; 1204 break; 1205 } 1206 if (tp->t_template == 0) 1207 panic("tcp_output"); 1208 if (tp->t_template->m_len < iphdrlen) 1209 panic("tcp_output"); 1210 bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen); 1211 1212 /* 1213 * If we are doing retransmissions, then snd_nxt will 1214 * not reflect the first unsent octet. For ACK only 1215 * packets, we do not want the sequence number of the 1216 * retransmitted packet, we want the sequence number 1217 * of the next unsent octet. So, if there is no data 1218 * (and no SYN or FIN), use snd_max instead of snd_nxt 1219 * when filling in ti_seq. But if we are in persist 1220 * state, snd_max might reflect one byte beyond the 1221 * right edge of the window, so use snd_nxt in that 1222 * case, since we know we aren't doing a retransmission. 1223 * (retransmit and persist are mutually exclusive...) 1224 */ 1225 if (TCP_SACK_ENABLED(tp) && sack_rxmit) { 1226 th->th_seq = htonl(p->rxmit); 1227 p->rxmit += len; 1228 } else { 1229 if (len || (flags & (TH_SYN|TH_FIN)) || 1230 TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) 1231 th->th_seq = htonl(tp->snd_nxt); 1232 else 1233 th->th_seq = htonl(tp->snd_max); 1234 } 1235 th->th_ack = htonl(tp->rcv_nxt); 1236 if (optlen) { 1237 bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen); 1238 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1239 } 1240 th->th_flags = flags; 1241 /* 1242 * Calculate receive window. Don't shrink window, 1243 * but avoid silly window syndrome. 1244 */ 1245 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize) 1246 win = 0; 1247 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 1248 win = (long)TCP_MAXWIN << tp->rcv_scale; 1249 if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt)) 1250 win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt); 1251 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale)); 1252 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1253 u_int32_t urp = tp->snd_up - tp->snd_nxt; 1254 if (urp > IP_MAXPACKET) 1255 urp = IP_MAXPACKET; 1256 th->th_urp = htons((u_int16_t)urp); 1257 th->th_flags |= TH_URG; 1258 } else 1259 /* 1260 * If no urgent pointer to send, then we pull 1261 * the urgent pointer to the left edge of the send window 1262 * so that it doesn't drift into the send window on sequence 1263 * number wraparound. 1264 */ 1265 tp->snd_up = tp->snd_una; /* drag it along */ 1266 1267 #ifdef TCP_SIGNATURE 1268 #if defined(INET6) && defined(FAST_IPSEC) 1269 if (tp->t_family == AF_INET) /* XXX */ 1270 #endif 1271 if (sigoff && (tp->t_flags & TF_SIGNATURE)) { 1272 struct secasvar *sav; 1273 u_int8_t *sigp; 1274 1275 sav = tcp_signature_getsav(m, th); 1276 1277 if (sav == NULL) { 1278 if (m) 1279 m_freem(m); 1280 return (EPERM); 1281 } 1282 1283 m->m_pkthdr.len = hdrlen + len; 1284 sigp = (caddr_t)th + sizeof(*th) + sigoff; 1285 tcp_signature(m, th, (caddr_t)th - mtod(m, caddr_t), sav, sigp); 1286 1287 key_sa_recordxfer(sav, m); 1288 #ifdef FAST_IPSEC 1289 KEY_FREESAV(&sav); 1290 #else 1291 key_freesav(sav); 1292 #endif 1293 } 1294 #endif 1295 1296 /* 1297 * Set ourselves up to be checksummed just before the packet 1298 * hits the wire. Maybe skip checksums on loopback interfaces. 1299 */ 1300 switch (af) { 1301 #ifdef INET 1302 case AF_INET: 1303 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1304 if (use_tso) { 1305 m->m_pkthdr.segsz = txsegsize; 1306 m->m_pkthdr.csum_flags = M_CSUM_TSOv4; 1307 } else { 1308 if (__predict_true(ro->ro_rt == NULL || 1309 !(ro->ro_rt->rt_ifp->if_flags & 1310 IFF_LOOPBACK) || 1311 tcp_do_loopback_cksum)) 1312 m->m_pkthdr.csum_flags = M_CSUM_TCPv4; 1313 else 1314 m->m_pkthdr.csum_flags = 0; 1315 if (len + optlen) { 1316 /* Fixup the pseudo-header checksum. */ 1317 /* XXXJRT Not IP Jumbogram safe. */ 1318 th->th_sum = in_cksum_addword(th->th_sum, 1319 htons((u_int16_t) (len + optlen))); 1320 } 1321 } 1322 break; 1323 #endif 1324 #ifdef INET6 1325 case AF_INET6: 1326 /* 1327 * XXX Actually delaying the checksum is Hard 1328 * XXX (well, maybe not for Itojun, but it is 1329 * XXX for me), but we can still take advantage 1330 * XXX of the cached pseudo-header checksum. 1331 */ 1332 /* equals to hdrlen + len */ 1333 m->m_pkthdr.len = sizeof(struct ip6_hdr) 1334 + sizeof(struct tcphdr) + optlen + len; 1335 #ifdef notyet 1336 if (__predict_true(ro->ro_rt == NULL || 1337 !(ro->ro_rt->rt_ifp->if_flags & 1338 IFF_LOOPBACK) || 1339 tcp_do_loopback_cksum)) 1340 m->m_pkthdr.csum_flags = M_CSUM_TCPv6; 1341 else 1342 m->m_pkthdr.csum_flags = 0; 1343 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1344 #endif 1345 if (len + optlen) { 1346 /* Fixup the pseudo-header checksum. */ 1347 /* XXXJRT: Not IPv6 Jumbogram safe. */ 1348 th->th_sum = in_cksum_addword(th->th_sum, 1349 htons((u_int16_t) (len + optlen))); 1350 } 1351 #ifndef notyet 1352 if (__predict_true(ro->ro_rt == NULL || 1353 !(ro->ro_rt->rt_ifp->if_flags & 1354 IFF_LOOPBACK) || 1355 tcp_do_loopback_cksum)) 1356 th->th_sum = in6_cksum(m, 0, sizeof(struct ip6_hdr), 1357 sizeof(struct tcphdr) + optlen + len); 1358 #endif 1359 break; 1360 #endif 1361 } 1362 1363 /* 1364 * In transmit state, time the transmission and arrange for 1365 * the retransmit. In persist state, just set snd_max. 1366 */ 1367 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1368 tcp_seq startseq = tp->snd_nxt; 1369 1370 /* 1371 * Advance snd_nxt over sequence space of this segment. 1372 * There are no states in which we send both a SYN and a FIN, 1373 * so we collapse the tests for these flags. 1374 */ 1375 if (flags & (TH_SYN|TH_FIN)) 1376 tp->snd_nxt++; 1377 if (sack_rxmit) 1378 goto timer; 1379 tp->snd_nxt += len; 1380 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1381 tp->snd_max = tp->snd_nxt; 1382 /* 1383 * Time this transmission if not a retransmission and 1384 * not currently timing anything. 1385 */ 1386 if (tp->t_rtttime == 0) { 1387 tp->t_rtttime = tcp_now; 1388 tp->t_rtseq = startseq; 1389 tcpstat.tcps_segstimed++; 1390 } 1391 } 1392 1393 /* 1394 * Set retransmit timer if not currently set, 1395 * and not doing an ack or a keep-alive probe. 1396 * Initial value for retransmit timer is smoothed 1397 * round-trip time + 2 * round-trip time variance. 1398 * Initialize shift counter which is used for backoff 1399 * of retransmit time. 1400 */ 1401 timer: 1402 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1403 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1404 tp->snd_nxt != tp->snd_una)) { 1405 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1406 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 1407 tp->t_rxtshift = 0; 1408 } 1409 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1410 } 1411 } else 1412 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 1413 tp->snd_max = tp->snd_nxt + len; 1414 1415 #ifdef TCP_DEBUG 1416 /* 1417 * Trace. 1418 */ 1419 if (so->so_options & SO_DEBUG) 1420 tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0); 1421 #endif 1422 1423 /* 1424 * Fill in IP length and desired time to live and 1425 * send to IP level. There should be a better way 1426 * to handle ttl and tos; we could keep them in 1427 * the template, but need a way to checksum without them. 1428 */ 1429 m->m_pkthdr.len = hdrlen + len; 1430 1431 switch (af) { 1432 #ifdef INET 1433 case AF_INET: 1434 ip->ip_len = htons(m->m_pkthdr.len); 1435 if (tp->t_inpcb) { 1436 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl; 1437 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos; 1438 } 1439 #ifdef INET6 1440 else if (tp->t_in6pcb) { 1441 ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/ 1442 ip->ip_tos = 0; /*XXX*/ 1443 } 1444 #endif 1445 break; 1446 #endif 1447 #ifdef INET6 1448 case AF_INET6: 1449 ip6->ip6_nxt = IPPROTO_TCP; 1450 if (tp->t_in6pcb) { 1451 /* 1452 * we separately set hoplimit for every segment, since 1453 * the user might want to change the value via 1454 * setsockopt. Also, desired default hop limit might 1455 * be changed via Neighbor Discovery. 1456 */ 1457 ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb, 1458 ro->ro_rt ? ro->ro_rt->rt_ifp : NULL); 1459 } 1460 /* ip6->ip6_flow = ??? */ 1461 /* ip6_plen will be filled in ip6_output(). */ 1462 break; 1463 #endif 1464 } 1465 1466 switch (af) { 1467 #ifdef INET 1468 case AF_INET: 1469 { 1470 struct mbuf *opts; 1471 1472 if (tp->t_inpcb) 1473 opts = tp->t_inpcb->inp_options; 1474 else 1475 opts = NULL; 1476 error = ip_output(m, opts, ro, 1477 (tp->t_mtudisc ? IP_MTUDISC : 0) | 1478 (so->so_options & SO_DONTROUTE), 1479 (struct ip_moptions *)0, so); 1480 break; 1481 } 1482 #endif 1483 #ifdef INET6 1484 case AF_INET6: 1485 { 1486 struct ip6_pktopts *opts; 1487 1488 if (tp->t_in6pcb) 1489 opts = tp->t_in6pcb->in6p_outputopts; 1490 else 1491 opts = NULL; 1492 error = ip6_output(m, opts, (struct route_in6 *)ro, 1493 so->so_options & SO_DONTROUTE, 1494 (struct ip6_moptions *)0, so, NULL); 1495 break; 1496 } 1497 #endif 1498 default: 1499 error = EAFNOSUPPORT; 1500 break; 1501 } 1502 if (error) { 1503 out: 1504 if (error == ENOBUFS) { 1505 tcpstat.tcps_selfquench++; 1506 #ifdef INET 1507 if (tp->t_inpcb) 1508 tcp_quench(tp->t_inpcb, 0); 1509 #endif 1510 #ifdef INET6 1511 if (tp->t_in6pcb) 1512 tcp6_quench(tp->t_in6pcb, 0); 1513 #endif 1514 error = 0; 1515 } else if ((error == EHOSTUNREACH || error == ENETDOWN) && 1516 TCPS_HAVERCVDSYN(tp->t_state)) { 1517 tp->t_softerror = error; 1518 error = 0; 1519 } 1520 1521 /* Back out the seqence number advance. */ 1522 if (sack_rxmit) 1523 p->rxmit -= len; 1524 1525 /* Restart the delayed ACK timer, if necessary. */ 1526 if (tp->t_flags & TF_DELACK) 1527 TCP_RESTART_DELACK(tp); 1528 1529 return (error); 1530 } 1531 tcpstat.tcps_sndtotal++; 1532 if (tp->t_flags & TF_DELACK) 1533 tcpstat.tcps_delack++; 1534 1535 /* 1536 * Data sent (as far as we can tell). 1537 * If this advertises a larger window than any other segment, 1538 * then remember the size of the advertised window. 1539 * Any pending ACK has now been sent. 1540 */ 1541 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv)) 1542 tp->rcv_adv = tp->rcv_nxt + win; 1543 tp->last_ack_sent = tp->rcv_nxt; 1544 tp->t_flags &= ~TF_ACKNOW; 1545 TCP_CLEAR_DELACK(tp); 1546 #ifdef DIAGNOSTIC 1547 if (maxburst < 0) 1548 printf("tcp_output: maxburst exceeded by %d\n", -maxburst); 1549 #endif 1550 if (sendalot && (!tcp_do_newreno || --maxburst)) 1551 goto again; 1552 return (0); 1553 } 1554 1555 void 1556 tcp_setpersist(struct tcpcb *tp) 1557 { 1558 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2); 1559 int nticks; 1560 1561 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT)) 1562 panic("tcp_output REXMT"); 1563 /* 1564 * Start/restart persistance timer. 1565 */ 1566 if (t < tp->t_rttmin) 1567 t = tp->t_rttmin; 1568 TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift], 1569 TCPTV_PERSMIN, TCPTV_PERSMAX); 1570 TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks); 1571 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1572 tp->t_rxtshift++; 1573 } 1574
Indexes created Tue Sep 23 14:10:03 GMT 2025