1 /* $NetBSD: tcp_output.c,v 1.176 2014/05/30 01:39:03 christos 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 1.176 2014/05/30 01:39:03 christos 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 IPSEC 179 #include <netipsec/ipsec.h> 180 #include <netipsec/key.h> 181 #ifdef INET6 182 #include <netipsec/ipsec6.h> 183 #endif 184 #endif /* IPSEC*/ 185 186 #include <netinet/tcp.h> 187 #define TCPOUTFLAGS 188 #include <netinet/tcp_fsm.h> 189 #include <netinet/tcp_seq.h> 190 #include <netinet/tcp_timer.h> 191 #include <netinet/tcp_var.h> 192 #include <netinet/tcp_private.h> 193 #include <netinet/tcp_congctl.h> 194 #include <netinet/tcpip.h> 195 #include <netinet/tcp_debug.h> 196 #include <netinet/in_offload.h> 197 #include <netinet6/in6_offload.h> 198 199 #ifdef notyet 200 extern struct mbuf *m_copypack(); 201 #endif 202 203 /* 204 * Knob to enable Congestion Window Monitoring, and control 205 * the burst size it allows. Default burst is 4 packets, per 206 * the Internet draft. 207 */ 208 int tcp_cwm = 0; 209 int tcp_cwm_burstsize = 4; 210 211 int tcp_do_autosndbuf = 1; 212 int tcp_autosndbuf_inc = 8 * 1024; 213 int tcp_autosndbuf_max = 256 * 1024; 214 215 #ifdef TCP_OUTPUT_COUNTERS 216 #include <sys/device.h> 217 218 extern struct evcnt tcp_output_bigheader; 219 extern struct evcnt tcp_output_predict_hit; 220 extern struct evcnt tcp_output_predict_miss; 221 extern struct evcnt tcp_output_copysmall; 222 extern struct evcnt tcp_output_copybig; 223 extern struct evcnt tcp_output_refbig; 224 225 #define TCP_OUTPUT_COUNTER_INCR(ev) (ev)->ev_count++ 226 #else 227 228 #define TCP_OUTPUT_COUNTER_INCR(ev) /* nothing */ 229 230 #endif /* TCP_OUTPUT_COUNTERS */ 231 232 static 233 #ifndef GPROF 234 inline 235 #endif 236 int 237 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep, 238 bool *alwaysfragp) 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 hdrlen; 251 int optlen; 252 253 *alwaysfragp = false; 254 255 #ifdef DIAGNOSTIC 256 if (tp->t_inpcb && tp->t_in6pcb) 257 panic("tcp_segsize: both t_inpcb and t_in6pcb are set"); 258 #endif 259 switch (tp->t_family) { 260 #ifdef INET 261 case AF_INET: 262 hdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 263 break; 264 #endif 265 #ifdef INET6 266 case AF_INET6: 267 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 268 break; 269 #endif 270 default: 271 size = tcp_mssdflt; 272 goto out; 273 } 274 275 rt = NULL; 276 #ifdef INET 277 if (inp) { 278 rt = in_pcbrtentry(inp); 279 so = inp->inp_socket; 280 } 281 #endif 282 #ifdef INET6 283 if (in6p) { 284 rt = in6_pcbrtentry(in6p); 285 so = in6p->in6p_socket; 286 } 287 #endif 288 if (rt == NULL) { 289 size = tcp_mssdflt; 290 goto out; 291 } 292 293 ifp = rt->rt_ifp; 294 295 size = tcp_mssdflt; 296 if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) { 297 #ifdef INET6 298 if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) { 299 /* 300 * RFC2460 section 5, last paragraph: if path MTU is 301 * smaller than 1280, use 1280 as packet size and 302 * attach fragment header. 303 */ 304 size = IPV6_MMTU - hdrlen - sizeof(struct ip6_frag); 305 *alwaysfragp = true; 306 } else 307 size = rt->rt_rmx.rmx_mtu - hdrlen; 308 #else 309 size = rt->rt_rmx.rmx_mtu - hdrlen; 310 #endif 311 } else if (ifp->if_flags & IFF_LOOPBACK) 312 size = ifp->if_mtu - hdrlen; 313 #ifdef INET 314 else if (inp && tp->t_mtudisc) 315 size = ifp->if_mtu - hdrlen; 316 else if (inp && in_localaddr(inp->inp_faddr)) 317 size = ifp->if_mtu - hdrlen; 318 #endif 319 #ifdef INET6 320 else if (in6p) { 321 #ifdef INET 322 if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) { 323 /* mapped addr case */ 324 struct in_addr d; 325 bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d)); 326 if (tp->t_mtudisc || in_localaddr(d)) 327 size = ifp->if_mtu - hdrlen; 328 } else 329 #endif 330 { 331 /* 332 * for IPv6, path MTU discovery is always turned on, 333 * or the node must use packet size <= 1280. 334 */ 335 size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU; 336 size -= hdrlen; 337 } 338 } 339 #endif 340 out: 341 /* 342 * Now we must make room for whatever extra TCP/IP options are in 343 * the packet. 344 */ 345 optlen = tcp_optlen(tp); 346 347 /* 348 * XXX tp->t_ourmss should have the right size, but without this code 349 * fragmentation will occur... need more investigation 350 */ 351 #ifdef INET 352 if (inp) { 353 #if defined(IPSEC) 354 if (ipsec_used && 355 !IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND)) 356 optlen += ipsec4_hdrsiz_tcp(tp); 357 #endif 358 optlen += ip_optlen(inp); 359 } 360 #endif 361 #ifdef INET6 362 #ifdef INET 363 if (in6p && tp->t_family == AF_INET) { 364 #if defined(IPSEC) 365 if (ipsec_used && 366 !IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND)) 367 optlen += ipsec4_hdrsiz_tcp(tp); 368 #endif 369 /* XXX size -= ip_optlen(in6p); */ 370 } else 371 #endif 372 if (in6p && tp->t_family == AF_INET6) { 373 #if defined(IPSEC) 374 if (ipsec_used && 375 !IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND)) 376 optlen += ipsec6_hdrsiz_tcp(tp); 377 #endif 378 optlen += ip6_optlen(in6p); 379 } 380 #endif 381 size -= optlen; 382 383 /* there may not be any room for data if mtu is too small */ 384 if (size < 0) 385 return (EMSGSIZE); 386 387 /* 388 * *rxsegsizep holds *estimated* inbound segment size (estimation 389 * assumes that path MTU is the same for both ways). this is only 390 * for silly window avoidance, do not use the value for other purposes. 391 * 392 * ipseclen is subtracted from both sides, this may not be right. 393 * I'm not quite sure about this (could someone comment). 394 */ 395 *txsegsizep = min(tp->t_peermss - optlen, size); 396 /* 397 * Never send more than half a buffer full. This insures that we can 398 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and 399 * therefore acks will never be delayed unless we run out of data to 400 * transmit. 401 */ 402 if (so) 403 *txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep); 404 *rxsegsizep = min(tp->t_ourmss - optlen, size); 405 406 if (*txsegsizep != tp->t_segsz) { 407 /* 408 * If the new segment size is larger, we don't want to 409 * mess up the congestion window, but if it is smaller 410 * we'll have to reduce the congestion window to ensure 411 * that we don't get into trouble with initial windows 412 * and the rest. In any case, if the segment size 413 * has changed, chances are the path has, too, and 414 * our congestion window will be different. 415 */ 416 if (*txsegsizep < tp->t_segsz) { 417 tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz) 418 * *txsegsizep, *txsegsizep); 419 tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz) 420 * *txsegsizep, *txsegsizep); 421 } 422 tp->t_segsz = *txsegsizep; 423 } 424 425 return (0); 426 } 427 428 static 429 #ifndef GPROF 430 inline 431 #endif 432 int 433 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off, 434 long len, int hdrlen, struct mbuf **mp) 435 { 436 struct mbuf *m, *m0; 437 uint64_t *tcps; 438 439 tcps = TCP_STAT_GETREF(); 440 if (tp->t_force && len == 1) 441 tcps[TCP_STAT_SNDPROBE]++; 442 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 443 tcps[TCP_STAT_SNDREXMITPACK]++; 444 tcps[TCP_STAT_SNDREXMITBYTE] += len; 445 } else { 446 tcps[TCP_STAT_SNDPACK]++; 447 tcps[TCP_STAT_SNDBYTE] += len; 448 } 449 TCP_STAT_PUTREF(); 450 #ifdef notyet 451 if ((m = m_copypack(so->so_snd.sb_mb, off, 452 (int)len, max_linkhdr + hdrlen)) == 0) 453 return (ENOBUFS); 454 /* 455 * m_copypack left space for our hdr; use it. 456 */ 457 m->m_len += hdrlen; 458 m->m_data -= hdrlen; 459 #else 460 MGETHDR(m, M_DONTWAIT, MT_HEADER); 461 if (__predict_false(m == NULL)) 462 return (ENOBUFS); 463 MCLAIM(m, &tcp_tx_mowner); 464 465 /* 466 * XXX Because other code assumes headers will fit in 467 * XXX one header mbuf. 468 * 469 * (This code should almost *never* be run.) 470 */ 471 if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) { 472 TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader); 473 MCLGET(m, M_DONTWAIT); 474 if ((m->m_flags & M_EXT) == 0) { 475 m_freem(m); 476 return (ENOBUFS); 477 } 478 } 479 480 m->m_data += max_linkhdr; 481 m->m_len = hdrlen; 482 483 /* 484 * To avoid traversing the whole sb_mb chain for correct 485 * data to send, remember last sent mbuf, its offset and 486 * the sent size. When called the next time, see if the 487 * data to send is directly following the previous transfer. 488 * This is important for large TCP windows. 489 */ 490 if (off == 0 || tp->t_lastm == NULL || 491 (tp->t_lastoff + tp->t_lastlen) != off) { 492 TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss); 493 /* 494 * Either a new packet or a retransmit. 495 * Start from the beginning. 496 */ 497 tp->t_lastm = so->so_snd.sb_mb; 498 tp->t_inoff = off; 499 } else { 500 TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit); 501 tp->t_inoff += tp->t_lastlen; 502 } 503 504 /* Traverse forward to next packet */ 505 while (tp->t_inoff > 0) { 506 if (tp->t_lastm == NULL) 507 panic("tp->t_lastm == NULL"); 508 if (tp->t_inoff < tp->t_lastm->m_len) 509 break; 510 tp->t_inoff -= tp->t_lastm->m_len; 511 tp->t_lastm = tp->t_lastm->m_next; 512 } 513 514 tp->t_lastoff = off; 515 tp->t_lastlen = len; 516 m0 = tp->t_lastm; 517 off = tp->t_inoff; 518 519 if (len <= M_TRAILINGSPACE(m)) { 520 m_copydata(m0, off, (int) len, mtod(m, char *) + hdrlen); 521 m->m_len += len; 522 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall); 523 } else { 524 m->m_next = m_copym(m0, off, (int) len, M_DONTWAIT); 525 if (m->m_next == NULL) { 526 m_freem(m); 527 return (ENOBUFS); 528 } 529 #ifdef TCP_OUTPUT_COUNTERS 530 if (m->m_next->m_flags & M_EXT) 531 TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig); 532 else 533 TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig); 534 #endif /* TCP_OUTPUT_COUNTERS */ 535 } 536 #endif 537 538 *mp = m; 539 return (0); 540 } 541 542 /* 543 * Tcp output routine: figure out what should be sent and send it. 544 */ 545 int 546 tcp_output(struct tcpcb *tp) 547 { 548 struct rtentry *rt; 549 struct socket *so; 550 struct route *ro; 551 long len, win; 552 int off, flags, error; 553 struct mbuf *m; 554 struct ip *ip; 555 #ifdef INET6 556 struct ip6_hdr *ip6; 557 #endif 558 struct tcphdr *th; 559 u_char opt[MAX_TCPOPTLEN]; 560 unsigned optlen, hdrlen, packetlen; 561 unsigned int sack_numblks; 562 int idle, sendalot, txsegsize, rxsegsize; 563 int txsegsize_nosack; 564 int maxburst = TCP_MAXBURST; 565 int af; /* address family on the wire */ 566 int iphdrlen; 567 int has_tso4, has_tso6; 568 int has_tso, use_tso; 569 bool alwaysfrag; 570 int sack_rxmit; 571 int sack_bytes_rxmt; 572 int ecn_tos; 573 struct sackhole *p; 574 #ifdef TCP_SIGNATURE 575 int sigoff = 0; 576 #endif 577 uint64_t *tcps; 578 579 #ifdef DIAGNOSTIC 580 if (tp->t_inpcb && tp->t_in6pcb) 581 panic("tcp_output: both t_inpcb and t_in6pcb are set"); 582 #endif 583 so = NULL; 584 ro = NULL; 585 if (tp->t_inpcb) { 586 so = tp->t_inpcb->inp_socket; 587 ro = &tp->t_inpcb->inp_route; 588 } 589 #ifdef INET6 590 else if (tp->t_in6pcb) { 591 so = tp->t_in6pcb->in6p_socket; 592 ro = &tp->t_in6pcb->in6p_route; 593 } 594 #endif 595 596 switch (af = tp->t_family) { 597 #ifdef INET 598 case AF_INET: 599 if (tp->t_inpcb) 600 break; 601 #ifdef INET6 602 /* mapped addr case */ 603 if (tp->t_in6pcb) 604 break; 605 #endif 606 return (EINVAL); 607 #endif 608 #ifdef INET6 609 case AF_INET6: 610 if (tp->t_in6pcb) 611 break; 612 return (EINVAL); 613 #endif 614 default: 615 return (EAFNOSUPPORT); 616 } 617 618 if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag)) 619 return (EMSGSIZE); 620 621 idle = (tp->snd_max == tp->snd_una); 622 623 /* 624 * Determine if we can use TCP segmentation offload: 625 * - If we're using IPv4 626 * - If there is not an IPsec policy that prevents it 627 * - If the interface can do it 628 */ 629 has_tso4 = has_tso6 = false; 630 #if defined(INET) 631 has_tso4 = tp->t_inpcb != NULL && 632 #if defined(IPSEC) 633 ipsec_used && IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp, 634 IPSEC_DIR_OUTBOUND) && 635 #endif 636 (rt = rtcache_validate(&tp->t_inpcb->inp_route)) != NULL && 637 (rt->rt_ifp->if_capenable & IFCAP_TSOv4) != 0; 638 #endif /* defined(INET) */ 639 #if defined(INET6) 640 has_tso6 = tp->t_in6pcb != NULL && 641 #if defined(IPSEC) 642 ipsec_used && IPSEC_PCB_SKIP_IPSEC(tp->t_in6pcb->in6p_sp, 643 IPSEC_DIR_OUTBOUND) && 644 #endif 645 (rt = rtcache_validate(&tp->t_in6pcb->in6p_route)) != NULL && 646 (rt->rt_ifp->if_capenable & IFCAP_TSOv6) != 0; 647 #endif /* defined(INET6) */ 648 has_tso = (has_tso4 || has_tso6) && !alwaysfrag; 649 650 /* 651 * Restart Window computation. From draft-floyd-incr-init-win-03: 652 * 653 * Optionally, a TCP MAY set the restart window to the 654 * minimum of the value used for the initial window and 655 * the current value of cwnd (in other words, using a 656 * larger value for the restart window should never increase 657 * the size of cwnd). 658 */ 659 if (tcp_cwm) { 660 /* 661 * Hughes/Touch/Heidemann Congestion Window Monitoring. 662 * Count the number of packets currently pending 663 * acknowledgement, and limit our congestion window 664 * to a pre-determined allowed burst size plus that count. 665 * This prevents bursting once all pending packets have 666 * been acknowledged (i.e. transmission is idle). 667 * 668 * XXX Link this to Initial Window? 669 */ 670 tp->snd_cwnd = min(tp->snd_cwnd, 671 (tcp_cwm_burstsize * txsegsize) + 672 (tp->snd_nxt - tp->snd_una)); 673 } else { 674 if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) { 675 /* 676 * We have been idle for "a while" and no acks are 677 * expected to clock out any data we send -- 678 * slow start to get ack "clock" running again. 679 */ 680 int ss = tcp_init_win; 681 #ifdef INET 682 if (tp->t_inpcb && 683 in_localaddr(tp->t_inpcb->inp_faddr)) 684 ss = tcp_init_win_local; 685 #endif 686 #ifdef INET6 687 if (tp->t_in6pcb && 688 in6_localaddr(&tp->t_in6pcb->in6p_faddr)) 689 ss = tcp_init_win_local; 690 #endif 691 tp->snd_cwnd = min(tp->snd_cwnd, 692 TCP_INITIAL_WINDOW(ss, txsegsize)); 693 } 694 } 695 696 txsegsize_nosack = txsegsize; 697 again: 698 ecn_tos = 0; 699 use_tso = has_tso; 700 if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) { 701 /* don't duplicate CWR/ECE. */ 702 use_tso = 0; 703 } 704 TCP_REASS_LOCK(tp); 705 sack_numblks = tcp_sack_numblks(tp); 706 if (sack_numblks) { 707 int sackoptlen; 708 709 sackoptlen = TCP_SACK_OPTLEN(sack_numblks); 710 if (sackoptlen > txsegsize_nosack) { 711 sack_numblks = 0; /* give up SACK */ 712 txsegsize = txsegsize_nosack; 713 } else { 714 if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) { 715 /* don't duplicate D-SACK. */ 716 use_tso = 0; 717 } 718 txsegsize = txsegsize_nosack - sackoptlen; 719 } 720 } else { 721 txsegsize = txsegsize_nosack; 722 } 723 724 /* 725 * Determine length of data that should be transmitted, and 726 * flags that should be used. If there is some data or critical 727 * controls (SYN, RST) to send, then transmit; otherwise, 728 * investigate further. 729 * 730 * Readjust SACK information to avoid resending duplicate data. 731 */ 732 if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max)) 733 tcp_sack_adjust(tp); 734 sendalot = 0; 735 off = tp->snd_nxt - tp->snd_una; 736 win = min(tp->snd_wnd, tp->snd_cwnd); 737 738 flags = tcp_outflags[tp->t_state]; 739 740 /* 741 * Send any SACK-generated retransmissions. If we're explicitly trying 742 * to send out new data (when sendalot is 1), bypass this function. 743 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 744 * we're replacing a (future) new transmission with a retransmission 745 * now, and we previously incremented snd_cwnd in tcp_input(). 746 */ 747 /* 748 * Still in sack recovery , reset rxmit flag to zero. 749 */ 750 sack_rxmit = 0; 751 sack_bytes_rxmt = 0; 752 len = 0; 753 p = NULL; 754 do { 755 long cwin; 756 if (!TCP_SACK_ENABLED(tp)) 757 break; 758 if (tp->t_partialacks < 0) 759 break; 760 p = tcp_sack_output(tp, &sack_bytes_rxmt); 761 if (p == NULL) 762 break; 763 764 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 765 if (cwin < 0) 766 cwin = 0; 767 /* Do not retransmit SACK segments beyond snd_recover */ 768 if (SEQ_GT(p->end, tp->snd_recover)) { 769 /* 770 * (At least) part of sack hole extends beyond 771 * snd_recover. Check to see if we can rexmit data 772 * for this hole. 773 */ 774 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 775 /* 776 * Can't rexmit any more data for this hole. 777 * That data will be rexmitted in the next 778 * sack recovery episode, when snd_recover 779 * moves past p->rxmit. 780 */ 781 p = NULL; 782 break; 783 } 784 /* Can rexmit part of the current hole */ 785 len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit)); 786 } else 787 len = ((long)ulmin(cwin, p->end - p->rxmit)); 788 off = p->rxmit - tp->snd_una; 789 if (off + len > so->so_snd.sb_cc) { 790 /* 1 for TH_FIN */ 791 KASSERT(off + len == so->so_snd.sb_cc + 1); 792 KASSERT(p->rxmit + len == tp->snd_max); 793 len = so->so_snd.sb_cc - off; 794 } 795 if (len > 0) { 796 sack_rxmit = 1; 797 sendalot = 1; 798 } 799 } while (/*CONSTCOND*/0); 800 801 /* 802 * If in persist timeout with window of 0, send 1 byte. 803 * Otherwise, if window is small but nonzero 804 * and timer expired, we will send what we can 805 * and go to transmit state. 806 */ 807 if (tp->t_force) { 808 if (win == 0) { 809 /* 810 * If we still have some data to send, then 811 * clear the FIN bit. Usually this would 812 * happen below when it realizes that we 813 * aren't sending all the data. However, 814 * if we have exactly 1 byte of unset data, 815 * then it won't clear the FIN bit below, 816 * and if we are in persist state, we wind 817 * up sending the packet without recording 818 * that we sent the FIN bit. 819 * 820 * We can't just blindly clear the FIN bit, 821 * because if we don't have any more data 822 * to send then the probe will be the FIN 823 * itself. 824 */ 825 if (off < so->so_snd.sb_cc) 826 flags &= ~TH_FIN; 827 win = 1; 828 } else { 829 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 830 tp->t_rxtshift = 0; 831 } 832 } 833 834 if (sack_rxmit == 0) { 835 if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) { 836 long cwin; 837 838 /* 839 * We are inside of a SACK recovery episode and are 840 * sending new data, having retransmitted all the 841 * data possible in the scoreboard. 842 */ 843 if (tp->snd_wnd < so->so_snd.sb_cc) { 844 len = tp->snd_wnd - off; 845 flags &= ~TH_FIN; 846 } else { 847 len = so->so_snd.sb_cc - off; 848 } 849 850 /* 851 * From FreeBSD: 852 * Don't remove this (len > 0) check ! 853 * We explicitly check for len > 0 here (although it 854 * isn't really necessary), to work around a gcc 855 * optimization issue - to force gcc to compute 856 * len above. Without this check, the computation 857 * of len is bungled by the optimizer. 858 */ 859 if (len > 0) { 860 cwin = tp->snd_cwnd - 861 (tp->snd_nxt - tp->sack_newdata) - 862 sack_bytes_rxmt; 863 if (cwin < 0) 864 cwin = 0; 865 if (cwin < len) { 866 len = cwin; 867 flags &= ~TH_FIN; 868 } 869 } 870 } else if (win < so->so_snd.sb_cc) { 871 len = win - off; 872 flags &= ~TH_FIN; 873 } else { 874 len = so->so_snd.sb_cc - off; 875 } 876 } 877 878 if (len < 0) { 879 /* 880 * If FIN has been sent but not acked, 881 * but we haven't been called to retransmit, 882 * len will be -1. Otherwise, window shrank 883 * after we sent into it. If window shrank to 0, 884 * cancel pending retransmit, pull snd_nxt back 885 * to (closed) window, and set the persist timer 886 * if it isn't already going. If the window didn't 887 * close completely, just wait for an ACK. 888 * 889 * If we have a pending FIN, either it has already been 890 * transmitted or it is outside the window, so drop it. 891 * If the FIN has been transmitted, but this is not a 892 * retransmission, then len must be -1. Therefore we also 893 * prevent here the sending of `gratuitous FINs'. This 894 * eliminates the need to check for that case below (e.g. 895 * to back up snd_nxt before the FIN so that the sequence 896 * number is correct). 897 */ 898 len = 0; 899 flags &= ~TH_FIN; 900 if (win == 0) { 901 TCP_TIMER_DISARM(tp, TCPT_REXMT); 902 tp->t_rxtshift = 0; 903 tp->snd_nxt = tp->snd_una; 904 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) 905 tcp_setpersist(tp); 906 } 907 } 908 909 /* 910 * Automatic sizing enables the performance of large buffers 911 * and most of the efficiency of small ones by only allocating 912 * space when it is needed. 913 * 914 * The criteria to step up the send buffer one notch are: 915 * 1. receive window of remote host is larger than send buffer 916 * (with a fudge factor of 5/4th); 917 * 2. send buffer is filled to 7/8th with data (so we actually 918 * have data to make use of it); 919 * 3. send buffer fill has not hit maximal automatic size; 920 * 4. our send window (slow start and cogestion controlled) is 921 * larger than sent but unacknowledged data in send buffer. 922 * 923 * The remote host receive window scaling factor may limit the 924 * growing of the send buffer before it reaches its allowed 925 * maximum. 926 * 927 * It scales directly with slow start or congestion window 928 * and does at most one step per received ACK. This fast 929 * scaling has the drawback of growing the send buffer beyond 930 * what is strictly necessary to make full use of a given 931 * delay*bandwith product. However testing has shown this not 932 * to be much of an problem. At worst we are trading wasting 933 * of available bandwith (the non-use of it) for wasting some 934 * socket buffer memory. 935 * 936 * TODO: Shrink send buffer during idle periods together 937 * with congestion window. Requires another timer. 938 */ 939 if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 940 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 941 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 942 so->so_snd.sb_cc < tcp_autosndbuf_max && 943 win >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 944 if (!sbreserve(&so->so_snd, 945 min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc, 946 tcp_autosndbuf_max), so)) 947 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 948 } 949 } 950 951 if (len > txsegsize) { 952 if (use_tso) { 953 /* 954 * Truncate TSO transfers to IP_MAXPACKET, and make 955 * sure that we send equal size transfers down the 956 * stack (rather than big-small-big-small-...). 957 */ 958 #ifdef INET6 959 CTASSERT(IPV6_MAXPACKET == IP_MAXPACKET); 960 #endif 961 len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize; 962 if (len <= txsegsize) { 963 use_tso = 0; 964 } 965 } else 966 len = txsegsize; 967 flags &= ~TH_FIN; 968 sendalot = 1; 969 } else 970 use_tso = 0; 971 if (sack_rxmit) { 972 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 973 flags &= ~TH_FIN; 974 } 975 976 win = sbspace(&so->so_rcv); 977 978 /* 979 * Sender silly window avoidance. If connection is idle 980 * and can send all data, a maximum segment, 981 * at least a maximum default-size segment do it, 982 * or are forced, do it; otherwise don't bother. 983 * If peer's buffer is tiny, then send 984 * when window is at least half open. 985 * If retransmitting (possibly after persist timer forced us 986 * to send into a small window), then must resend. 987 */ 988 if (len) { 989 if (len >= txsegsize) 990 goto send; 991 if ((so->so_state & SS_MORETOCOME) == 0 && 992 ((idle || tp->t_flags & TF_NODELAY) && 993 len + off >= so->so_snd.sb_cc)) 994 goto send; 995 if (tp->t_force) 996 goto send; 997 if (len >= tp->max_sndwnd / 2) 998 goto send; 999 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1000 goto send; 1001 if (sack_rxmit) 1002 goto send; 1003 } 1004 1005 /* 1006 * Compare available window to amount of window known to peer 1007 * (as advertised window less next expected input). If the 1008 * difference is at least twice the size of the largest segment 1009 * we expect to receive (i.e. two segments) or at least 50% of 1010 * the maximum possible window, then want to send a window update 1011 * to peer. 1012 */ 1013 if (win > 0) { 1014 /* 1015 * "adv" is the amount we can increase the window, 1016 * taking into account that we are limited by 1017 * TCP_MAXWIN << tp->rcv_scale. 1018 */ 1019 long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) - 1020 (tp->rcv_adv - tp->rcv_nxt); 1021 1022 if (adv >= (long) (2 * rxsegsize)) 1023 goto send; 1024 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 1025 goto send; 1026 } 1027 1028 /* 1029 * Send if we owe peer an ACK. 1030 */ 1031 if (tp->t_flags & TF_ACKNOW) 1032 goto send; 1033 if (flags & (TH_SYN|TH_FIN|TH_RST)) 1034 goto send; 1035 if (SEQ_GT(tp->snd_up, tp->snd_una)) 1036 goto send; 1037 /* 1038 * In SACK, it is possible for tcp_output to fail to send a segment 1039 * after the retransmission timer has been turned off. Make sure 1040 * that the retransmission timer is set. 1041 */ 1042 if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) && 1043 !TCP_TIMER_ISARMED(tp, TCPT_REXMT) && 1044 !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1045 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1046 goto just_return; 1047 } 1048 1049 /* 1050 * TCP window updates are not reliable, rather a polling protocol 1051 * using ``persist'' packets is used to insure receipt of window 1052 * updates. The three ``states'' for the output side are: 1053 * idle not doing retransmits or persists 1054 * persisting to move a small or zero window 1055 * (re)transmitting and thereby not persisting 1056 * 1057 * tp->t_timer[TCPT_PERSIST] 1058 * is set when we are in persist state. 1059 * tp->t_force 1060 * is set when we are called to send a persist packet. 1061 * tp->t_timer[TCPT_REXMT] 1062 * is set when we are retransmitting 1063 * The output side is idle when both timers are zero. 1064 * 1065 * If send window is too small, there is data to transmit, and no 1066 * retransmit or persist is pending, then go to persist state. 1067 * If nothing happens soon, send when timer expires: 1068 * if window is nonzero, transmit what we can, 1069 * otherwise force out a byte. 1070 */ 1071 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1072 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1073 tp->t_rxtshift = 0; 1074 tcp_setpersist(tp); 1075 } 1076 1077 /* 1078 * No reason to send a segment, just return. 1079 */ 1080 just_return: 1081 TCP_REASS_UNLOCK(tp); 1082 return (0); 1083 1084 send: 1085 /* 1086 * Before ESTABLISHED, force sending of initial options 1087 * unless TCP set not to do any options. 1088 * NOTE: we assume that the IP/TCP header plus TCP options 1089 * always fit in a single mbuf, leaving room for a maximum 1090 * link header, i.e. 1091 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 1092 */ 1093 optlen = 0; 1094 switch (af) { 1095 #ifdef INET 1096 case AF_INET: 1097 iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 1098 break; 1099 #endif 1100 #ifdef INET6 1101 case AF_INET6: 1102 iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1103 break; 1104 #endif 1105 default: /*pacify gcc*/ 1106 iphdrlen = 0; 1107 break; 1108 } 1109 hdrlen = iphdrlen; 1110 if (flags & TH_SYN) { 1111 struct rtentry *synrt; 1112 1113 synrt = NULL; 1114 #ifdef INET 1115 if (tp->t_inpcb) 1116 synrt = in_pcbrtentry(tp->t_inpcb); 1117 #endif 1118 #ifdef INET6 1119 if (tp->t_in6pcb) 1120 synrt = in6_pcbrtentry(tp->t_in6pcb); 1121 #endif 1122 1123 tp->snd_nxt = tp->iss; 1124 tp->t_ourmss = tcp_mss_to_advertise(synrt != NULL ? 1125 synrt->rt_ifp : NULL, af); 1126 if ((tp->t_flags & TF_NOOPT) == 0) { 1127 opt[0] = TCPOPT_MAXSEG; 1128 opt[1] = 4; 1129 opt[2] = (tp->t_ourmss >> 8) & 0xff; 1130 opt[3] = tp->t_ourmss & 0xff; 1131 optlen = 4; 1132 1133 if ((tp->t_flags & TF_REQ_SCALE) && 1134 ((flags & TH_ACK) == 0 || 1135 (tp->t_flags & TF_RCVD_SCALE))) { 1136 *((u_int32_t *) (opt + optlen)) = htonl( 1137 TCPOPT_NOP << 24 | 1138 TCPOPT_WINDOW << 16 | 1139 TCPOLEN_WINDOW << 8 | 1140 tp->request_r_scale); 1141 optlen += 4; 1142 } 1143 if (tcp_do_sack) { 1144 u_int8_t *cp = (u_int8_t *)(opt + optlen); 1145 1146 cp[0] = TCPOPT_SACK_PERMITTED; 1147 cp[1] = 2; 1148 cp[2] = TCPOPT_NOP; 1149 cp[3] = TCPOPT_NOP; 1150 optlen += 4; 1151 } 1152 } 1153 } 1154 1155 /* 1156 * Send a timestamp and echo-reply if this is a SYN and our side 1157 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 1158 * and our peer have sent timestamps in our SYN's. 1159 */ 1160 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 1161 (flags & TH_RST) == 0 && 1162 ((flags & (TH_SYN|TH_ACK)) == TH_SYN || 1163 (tp->t_flags & TF_RCVD_TSTMP))) { 1164 u_int32_t *lp = (u_int32_t *)(opt + optlen); 1165 1166 /* Form timestamp option as shown in appendix A of RFC 1323. */ 1167 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 1168 *lp++ = htonl(TCP_TIMESTAMP(tp)); 1169 *lp = htonl(tp->ts_recent); 1170 optlen += TCPOLEN_TSTAMP_APPA; 1171 1172 /* Set receive buffer autosizing timestamp. */ 1173 if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) 1174 tp->rfbuf_ts = TCP_TIMESTAMP(tp); 1175 } 1176 1177 /* 1178 * Tack on the SACK block if it is necessary. 1179 */ 1180 if (sack_numblks) { 1181 int sack_len; 1182 u_char *bp = (u_char *)(opt + optlen); 1183 u_int32_t *lp = (u_int32_t *)(bp + 4); 1184 struct ipqent *tiqe; 1185 1186 sack_len = sack_numblks * 8 + 2; 1187 bp[0] = TCPOPT_NOP; 1188 bp[1] = TCPOPT_NOP; 1189 bp[2] = TCPOPT_SACK; 1190 bp[3] = sack_len; 1191 if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) { 1192 sack_numblks--; 1193 *lp++ = htonl(tp->rcv_dsack_block.left); 1194 *lp++ = htonl(tp->rcv_dsack_block.right); 1195 tp->rcv_sack_flags &= ~TCPSACK_HAVED; 1196 } 1197 for (tiqe = TAILQ_FIRST(&tp->timeq); 1198 sack_numblks > 0; tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) { 1199 KASSERT(tiqe != NULL); 1200 sack_numblks--; 1201 *lp++ = htonl(tiqe->ipqe_seq); 1202 *lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len + 1203 ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0)); 1204 } 1205 optlen += sack_len + 2; 1206 } 1207 TCP_REASS_UNLOCK(tp); 1208 1209 #ifdef TCP_SIGNATURE 1210 if (tp->t_flags & TF_SIGNATURE) { 1211 u_char *bp; 1212 /* 1213 * Initialize TCP-MD5 option (RFC2385) 1214 */ 1215 bp = (u_char *)opt + optlen; 1216 *bp++ = TCPOPT_SIGNATURE; 1217 *bp++ = TCPOLEN_SIGNATURE; 1218 sigoff = optlen + 2; 1219 memset(bp, 0, TCP_SIGLEN); 1220 bp += TCP_SIGLEN; 1221 optlen += TCPOLEN_SIGNATURE; 1222 /* 1223 * Terminate options list and maintain 32-bit alignment. 1224 */ 1225 *bp++ = TCPOPT_NOP; 1226 *bp++ = TCPOPT_EOL; 1227 optlen += 2; 1228 } 1229 #endif /* TCP_SIGNATURE */ 1230 1231 hdrlen += optlen; 1232 1233 #ifdef DIAGNOSTIC 1234 if (!use_tso && len > txsegsize) 1235 panic("tcp data to be sent is larger than segment"); 1236 else if (use_tso && len > IP_MAXPACKET) 1237 panic("tcp data to be sent is larger than max TSO size"); 1238 if (max_linkhdr + hdrlen > MCLBYTES) 1239 panic("tcphdr too big"); 1240 #endif 1241 1242 /* 1243 * Grab a header mbuf, attaching a copy of data to 1244 * be transmitted, and initialize the header from 1245 * the template for sends on this connection. 1246 */ 1247 if (len) { 1248 error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m); 1249 if (error) 1250 goto out; 1251 /* 1252 * If we're sending everything we've got, set PUSH. 1253 * (This will keep happy those implementations which only 1254 * give data to the user when a buffer fills or 1255 * a PUSH comes in.) 1256 */ 1257 if (off + len == so->so_snd.sb_cc) 1258 flags |= TH_PUSH; 1259 } else { 1260 tcps = TCP_STAT_GETREF(); 1261 if (tp->t_flags & TF_ACKNOW) 1262 tcps[TCP_STAT_SNDACKS]++; 1263 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1264 tcps[TCP_STAT_SNDCTRL]++; 1265 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1266 tcps[TCP_STAT_SNDURG]++; 1267 else 1268 tcps[TCP_STAT_SNDWINUP]++; 1269 TCP_STAT_PUTREF(); 1270 1271 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1272 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 1273 MCLGET(m, M_DONTWAIT); 1274 if ((m->m_flags & M_EXT) == 0) { 1275 m_freem(m); 1276 m = NULL; 1277 } 1278 } 1279 if (m == NULL) { 1280 error = ENOBUFS; 1281 goto out; 1282 } 1283 MCLAIM(m, &tcp_tx_mowner); 1284 m->m_data += max_linkhdr; 1285 m->m_len = hdrlen; 1286 } 1287 m->m_pkthdr.rcvif = NULL; 1288 switch (af) { 1289 #ifdef INET 1290 case AF_INET: 1291 ip = mtod(m, struct ip *); 1292 #ifdef INET6 1293 ip6 = NULL; 1294 #endif 1295 th = (struct tcphdr *)(ip + 1); 1296 break; 1297 #endif 1298 #ifdef INET6 1299 case AF_INET6: 1300 ip = NULL; 1301 ip6 = mtod(m, struct ip6_hdr *); 1302 th = (struct tcphdr *)(ip6 + 1); 1303 break; 1304 #endif 1305 default: /*pacify gcc*/ 1306 ip = NULL; 1307 #ifdef INET6 1308 ip6 = NULL; 1309 #endif 1310 th = NULL; 1311 break; 1312 } 1313 if (tp->t_template == 0) 1314 panic("tcp_output"); 1315 if (tp->t_template->m_len < iphdrlen) 1316 panic("tcp_output"); 1317 bcopy(mtod(tp->t_template, void *), mtod(m, void *), iphdrlen); 1318 1319 /* 1320 * If we are starting a connection, send ECN setup 1321 * SYN packet. If we are on a retransmit, we may 1322 * resend those bits a number of times as per 1323 * RFC 3168. 1324 */ 1325 if (tp->t_state == TCPS_SYN_SENT && tcp_do_ecn) { 1326 if (tp->t_flags & TF_SYN_REXMT) { 1327 if (tp->t_ecn_retries--) 1328 flags |= TH_ECE|TH_CWR; 1329 } else { 1330 flags |= TH_ECE|TH_CWR; 1331 tp->t_ecn_retries = tcp_ecn_maxretries; 1332 } 1333 } 1334 1335 if (TCP_ECN_ALLOWED(tp)) { 1336 /* 1337 * If the peer has ECN, mark data packets 1338 * ECN capable. Ignore pure ack packets, retransmissions 1339 * and window probes. 1340 */ 1341 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1342 !(tp->t_force && len == 1)) { 1343 ecn_tos = IPTOS_ECN_ECT0; 1344 TCP_STATINC(TCP_STAT_ECN_ECT); 1345 } 1346 1347 /* 1348 * Reply with proper ECN notifications. 1349 */ 1350 if (tp->t_flags & TF_ECN_SND_CWR) { 1351 flags |= TH_CWR; 1352 tp->t_flags &= ~TF_ECN_SND_CWR; 1353 } 1354 if (tp->t_flags & TF_ECN_SND_ECE) { 1355 flags |= TH_ECE; 1356 } 1357 } 1358 1359 1360 /* 1361 * If we are doing retransmissions, then snd_nxt will 1362 * not reflect the first unsent octet. For ACK only 1363 * packets, we do not want the sequence number of the 1364 * retransmitted packet, we want the sequence number 1365 * of the next unsent octet. So, if there is no data 1366 * (and no SYN or FIN), use snd_max instead of snd_nxt 1367 * when filling in ti_seq. But if we are in persist 1368 * state, snd_max might reflect one byte beyond the 1369 * right edge of the window, so use snd_nxt in that 1370 * case, since we know we aren't doing a retransmission. 1371 * (retransmit and persist are mutually exclusive...) 1372 */ 1373 if (TCP_SACK_ENABLED(tp) && sack_rxmit) { 1374 th->th_seq = htonl(p->rxmit); 1375 p->rxmit += len; 1376 } else { 1377 if (len || (flags & (TH_SYN|TH_FIN)) || 1378 TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) 1379 th->th_seq = htonl(tp->snd_nxt); 1380 else 1381 th->th_seq = htonl(tp->snd_max); 1382 } 1383 th->th_ack = htonl(tp->rcv_nxt); 1384 if (optlen) { 1385 bcopy((void *)opt, (void *)(th + 1), optlen); 1386 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1387 } 1388 th->th_flags = flags; 1389 /* 1390 * Calculate receive window. Don't shrink window, 1391 * but avoid silly window syndrome. 1392 */ 1393 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize) 1394 win = 0; 1395 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 1396 win = (long)TCP_MAXWIN << tp->rcv_scale; 1397 if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt)) 1398 win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt); 1399 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale)); 1400 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1401 u_int32_t urp = tp->snd_up - tp->snd_nxt; 1402 if (urp > IP_MAXPACKET) 1403 urp = IP_MAXPACKET; 1404 th->th_urp = htons((u_int16_t)urp); 1405 th->th_flags |= TH_URG; 1406 } else 1407 /* 1408 * If no urgent pointer to send, then we pull 1409 * the urgent pointer to the left edge of the send window 1410 * so that it doesn't drift into the send window on sequence 1411 * number wraparound. 1412 */ 1413 tp->snd_up = tp->snd_una; /* drag it along */ 1414 1415 #ifdef TCP_SIGNATURE 1416 if (sigoff && (tp->t_flags & TF_SIGNATURE)) { 1417 struct secasvar *sav; 1418 u_int8_t *sigp; 1419 1420 sav = tcp_signature_getsav(m, th); 1421 1422 if (sav == NULL) { 1423 if (m) 1424 m_freem(m); 1425 return (EPERM); 1426 } 1427 1428 m->m_pkthdr.len = hdrlen + len; 1429 sigp = (char *)th + sizeof(*th) + sigoff; 1430 tcp_signature(m, th, (char *)th - mtod(m, char *), sav, sigp); 1431 1432 key_sa_recordxfer(sav, m); 1433 KEY_FREESAV(&sav); 1434 } 1435 #endif 1436 1437 /* 1438 * Set ourselves up to be checksummed just before the packet 1439 * hits the wire. 1440 */ 1441 switch (af) { 1442 #ifdef INET 1443 case AF_INET: 1444 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1445 if (use_tso) { 1446 m->m_pkthdr.segsz = txsegsize; 1447 m->m_pkthdr.csum_flags = M_CSUM_TSOv4; 1448 } else { 1449 m->m_pkthdr.csum_flags = M_CSUM_TCPv4; 1450 if (len + optlen) { 1451 /* Fixup the pseudo-header checksum. */ 1452 /* XXXJRT Not IP Jumbogram safe. */ 1453 th->th_sum = in_cksum_addword(th->th_sum, 1454 htons((u_int16_t) (len + optlen))); 1455 } 1456 } 1457 break; 1458 #endif 1459 #ifdef INET6 1460 case AF_INET6: 1461 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1462 if (use_tso) { 1463 m->m_pkthdr.segsz = txsegsize; 1464 m->m_pkthdr.csum_flags = M_CSUM_TSOv6; 1465 } else { 1466 m->m_pkthdr.csum_flags = M_CSUM_TCPv6; 1467 if (len + optlen) { 1468 /* Fixup the pseudo-header checksum. */ 1469 /* XXXJRT: Not IPv6 Jumbogram safe. */ 1470 th->th_sum = in_cksum_addword(th->th_sum, 1471 htons((u_int16_t) (len + optlen))); 1472 } 1473 } 1474 break; 1475 #endif 1476 } 1477 1478 /* 1479 * In transmit state, time the transmission and arrange for 1480 * the retransmit. In persist state, just set snd_max. 1481 */ 1482 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1483 tcp_seq startseq = tp->snd_nxt; 1484 1485 /* 1486 * Advance snd_nxt over sequence space of this segment. 1487 * There are no states in which we send both a SYN and a FIN, 1488 * so we collapse the tests for these flags. 1489 */ 1490 if (flags & (TH_SYN|TH_FIN)) 1491 tp->snd_nxt++; 1492 if (sack_rxmit) 1493 goto timer; 1494 tp->snd_nxt += len; 1495 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1496 tp->snd_max = tp->snd_nxt; 1497 /* 1498 * Time this transmission if not a retransmission and 1499 * not currently timing anything. 1500 */ 1501 if (tp->t_rtttime == 0) { 1502 tp->t_rtttime = tcp_now; 1503 tp->t_rtseq = startseq; 1504 TCP_STATINC(TCP_STAT_SEGSTIMED); 1505 } 1506 } 1507 1508 /* 1509 * Set retransmit timer if not currently set, 1510 * and not doing an ack or a keep-alive probe. 1511 * Initial value for retransmit timer is smoothed 1512 * round-trip time + 2 * round-trip time variance. 1513 * Initialize shift counter which is used for backoff 1514 * of retransmit time. 1515 */ 1516 timer: 1517 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1518 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1519 tp->snd_nxt != tp->snd_una)) { 1520 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 1521 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 1522 tp->t_rxtshift = 0; 1523 } 1524 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 1525 } 1526 } else 1527 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 1528 tp->snd_max = tp->snd_nxt + len; 1529 1530 #ifdef TCP_DEBUG 1531 /* 1532 * Trace. 1533 */ 1534 if (so->so_options & SO_DEBUG) 1535 tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0); 1536 #endif 1537 1538 /* 1539 * Fill in IP length and desired time to live and 1540 * send to IP level. There should be a better way 1541 * to handle ttl and tos; we could keep them in 1542 * the template, but need a way to checksum without them. 1543 */ 1544 m->m_pkthdr.len = hdrlen + len; 1545 1546 switch (af) { 1547 #ifdef INET 1548 case AF_INET: 1549 ip->ip_len = htons(m->m_pkthdr.len); 1550 packetlen = m->m_pkthdr.len; 1551 if (tp->t_inpcb) { 1552 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl; 1553 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos | ecn_tos; 1554 } 1555 #ifdef INET6 1556 else if (tp->t_in6pcb) { 1557 ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/ 1558 ip->ip_tos = ecn_tos; /*XXX*/ 1559 } 1560 #endif 1561 break; 1562 #endif 1563 #ifdef INET6 1564 case AF_INET6: 1565 packetlen = m->m_pkthdr.len; 1566 ip6->ip6_nxt = IPPROTO_TCP; 1567 if (tp->t_in6pcb) { 1568 /* 1569 * we separately set hoplimit for every segment, since 1570 * the user might want to change the value via 1571 * setsockopt. Also, desired default hop limit might 1572 * be changed via Neighbor Discovery. 1573 */ 1574 ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb, 1575 (rt = rtcache_validate(ro)) != NULL ? rt->rt_ifp 1576 : NULL); 1577 } 1578 ip6->ip6_flow |= htonl(ecn_tos << 20); 1579 /* ip6->ip6_flow = ??? (from template) */ 1580 /* ip6_plen will be filled in ip6_output(). */ 1581 break; 1582 #endif 1583 default: /*pacify gcc*/ 1584 packetlen = 0; 1585 break; 1586 } 1587 1588 switch (af) { 1589 #ifdef INET 1590 case AF_INET: 1591 { 1592 struct mbuf *opts; 1593 1594 if (tp->t_inpcb) 1595 opts = tp->t_inpcb->inp_options; 1596 else 1597 opts = NULL; 1598 error = ip_output(m, opts, ro, 1599 (tp->t_mtudisc ? IP_MTUDISC : 0) | 1600 (so->so_options & SO_DONTROUTE), NULL, so); 1601 break; 1602 } 1603 #endif 1604 #ifdef INET6 1605 case AF_INET6: 1606 { 1607 struct ip6_pktopts *opts; 1608 1609 if (tp->t_in6pcb) 1610 opts = tp->t_in6pcb->in6p_outputopts; 1611 else 1612 opts = NULL; 1613 error = ip6_output(m, opts, ro, so->so_options & SO_DONTROUTE, 1614 NULL, so, NULL); 1615 break; 1616 } 1617 #endif 1618 default: 1619 error = EAFNOSUPPORT; 1620 break; 1621 } 1622 if (error) { 1623 out: 1624 if (error == ENOBUFS) { 1625 TCP_STATINC(TCP_STAT_SELFQUENCH); 1626 #ifdef INET 1627 if (tp->t_inpcb) 1628 tcp_quench(tp->t_inpcb, 0); 1629 #endif 1630 #ifdef INET6 1631 if (tp->t_in6pcb) 1632 tcp6_quench(tp->t_in6pcb, 0); 1633 #endif 1634 error = 0; 1635 } else if ((error == EHOSTUNREACH || error == ENETDOWN) && 1636 TCPS_HAVERCVDSYN(tp->t_state)) { 1637 tp->t_softerror = error; 1638 error = 0; 1639 } 1640 1641 /* Back out the seqence number advance. */ 1642 if (sack_rxmit) 1643 p->rxmit -= len; 1644 1645 /* Restart the delayed ACK timer, if necessary. */ 1646 if (tp->t_flags & TF_DELACK) 1647 TCP_RESTART_DELACK(tp); 1648 1649 return (error); 1650 } 1651 1652 if (packetlen > tp->t_pmtud_mtu_sent) 1653 tp->t_pmtud_mtu_sent = packetlen; 1654 1655 tcps = TCP_STAT_GETREF(); 1656 tcps[TCP_STAT_SNDTOTAL]++; 1657 if (tp->t_flags & TF_DELACK) 1658 tcps[TCP_STAT_DELACK]++; 1659 TCP_STAT_PUTREF(); 1660 1661 /* 1662 * Data sent (as far as we can tell). 1663 * If this advertises a larger window than any other segment, 1664 * then remember the size of the advertised window. 1665 * Any pending ACK has now been sent. 1666 */ 1667 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv)) 1668 tp->rcv_adv = tp->rcv_nxt + win; 1669 tp->last_ack_sent = tp->rcv_nxt; 1670 tp->t_flags &= ~TF_ACKNOW; 1671 TCP_CLEAR_DELACK(tp); 1672 #ifdef DIAGNOSTIC 1673 if (maxburst < 0) 1674 printf("tcp_output: maxburst exceeded by %d\n", -maxburst); 1675 #endif 1676 if (sendalot && (tp->t_congctl == &tcp_reno_ctl || --maxburst)) 1677 goto again; 1678 return (0); 1679 } 1680 1681 void 1682 tcp_setpersist(struct tcpcb *tp) 1683 { 1684 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2); 1685 int nticks; 1686 1687 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT)) 1688 panic("tcp_output REXMT"); 1689 /* 1690 * Start/restart persistance timer. 1691 */ 1692 if (t < tp->t_rttmin) 1693 t = tp->t_rttmin; 1694 TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift], 1695 TCPTV_PERSMIN, TCPTV_PERSMAX); 1696 TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks); 1697 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1698 tp->t_rxtshift++; 1699 } 1700