1 /* 2 * Copyright (c) 2004 Jeffrey M. Hsu. All rights reserved. 3 * Copyright (c) 2004 The DragonFly Project. All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Jeffrey M. Hsu. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of The DragonFly Project nor the names of its 17 * contributors may be used to endorse or promote products derived 18 * from this software without specific, prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 36 * The Regents of the University of California. All rights reserved. 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. All advertising materials mentioning features or use of this software 47 * must display the following acknowledgement: 48 * This product includes software developed by the University of 49 * California, Berkeley and its contributors. 50 * 4. Neither the name of the University 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 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 64 * SUCH DAMAGE. 65 * 66 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 67 * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.20 2003/01/29 22:45:36 hsu Exp $ 68 * $DragonFly: src/sys/netinet/tcp_output.c,v 1.34 2007/04/22 01:13:14 dillon Exp $ 69 */ 70 71 #include "opt_inet6.h" 72 #include "opt_ipsec.h" 73 #include "opt_tcpdebug.h" 74 75 #include <sys/param.h> 76 #include <sys/systm.h> 77 #include <sys/kernel.h> 78 #include <sys/sysctl.h> 79 #include <sys/mbuf.h> 80 #include <sys/domain.h> 81 #include <sys/protosw.h> 82 #include <sys/socket.h> 83 #include <sys/socketvar.h> 84 #include <sys/in_cksum.h> 85 #include <sys/thread.h> 86 #include <sys/globaldata.h> 87 88 #include <net/route.h> 89 90 #include <netinet/in.h> 91 #include <netinet/in_systm.h> 92 #include <netinet/ip.h> 93 #include <netinet/in_pcb.h> 94 #include <netinet/ip_var.h> 95 #include <netinet6/in6_pcb.h> 96 #include <netinet/ip6.h> 97 #include <netinet6/ip6_var.h> 98 #include <netinet/tcp.h> 99 #define TCPOUTFLAGS 100 #include <netinet/tcp_fsm.h> 101 #include <netinet/tcp_seq.h> 102 #include <netinet/tcp_timer.h> 103 #include <netinet/tcp_timer2.h> 104 #include <netinet/tcp_var.h> 105 #include <netinet/tcpip.h> 106 #ifdef TCPDEBUG 107 #include <netinet/tcp_debug.h> 108 #endif 109 110 #ifdef IPSEC 111 #include <netinet6/ipsec.h> 112 #endif /*IPSEC*/ 113 114 #ifdef FAST_IPSEC 115 #include <netproto/ipsec/ipsec.h> 116 #define IPSEC 117 #endif /*FAST_IPSEC*/ 118 119 #ifdef notyet 120 extern struct mbuf *m_copypack(); 121 #endif 122 123 int path_mtu_discovery = 0; 124 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 125 &path_mtu_discovery, 1, "Enable Path MTU Discovery"); 126 127 static int avoid_pure_win_update = 1; 128 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW, 129 &avoid_pure_win_update, 1, "Avoid pure window updates when possible"); 130 131 /* 132 * Tcp output routine: figure out what should be sent and send it. 133 */ 134 int 135 tcp_output(struct tcpcb *tp) 136 { 137 struct inpcb * const inp = tp->t_inpcb; 138 struct socket *so = inp->inp_socket; 139 long len, recvwin, sendwin; 140 int nsacked = 0; 141 int off, flags, error; 142 struct mbuf *m; 143 struct ip *ip = NULL; 144 struct ipovly *ipov = NULL; 145 struct tcphdr *th; 146 u_char opt[TCP_MAXOLEN]; 147 unsigned int ipoptlen, optlen, hdrlen; 148 int idle; 149 boolean_t sendalot; 150 struct ip6_hdr *ip6 = NULL; 151 #ifdef INET6 152 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 153 #else 154 const boolean_t isipv6 = FALSE; 155 #endif 156 struct rmxp_tao *taop; 157 158 /* 159 * Determine length of data that should be transmitted, 160 * and flags that will be used. 161 * If there is some data or critical controls (SYN, RST) 162 * to send, then transmit; otherwise, investigate further. 163 */ 164 165 /* 166 * If we have been idle for a while, the send congestion window 167 * could be no longer representative of the current state of the link. 168 * So unless we are expecting more acks to come in, slow-start from 169 * scratch to re-determine the send congestion window. 170 */ 171 if (tp->snd_max == tp->snd_una && 172 (ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 173 if (tcp_do_rfc3390) { 174 int initial_cwnd = 175 min(4 * tp->t_maxseg, max(2 * tp->t_maxseg, 4380)); 176 177 tp->snd_cwnd = min(tp->snd_cwnd, initial_cwnd); 178 } else { 179 tp->snd_cwnd = tp->t_maxseg; 180 } 181 tp->snd_wacked = 0; 182 } 183 184 /* 185 * Calculate whether the transmit stream was previously idle 186 * and adjust TF_LASTIDLE for the next time. 187 */ 188 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 189 if (idle && (tp->t_flags & TF_MORETOCOME)) 190 tp->t_flags |= TF_LASTIDLE; 191 else 192 tp->t_flags &= ~TF_LASTIDLE; 193 194 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 195 !IN_FASTRECOVERY(tp)) 196 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt); 197 198 again: 199 /* Make use of SACK information when slow-starting after a RTO. */ 200 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 201 !IN_FASTRECOVERY(tp)) { 202 tcp_seq old_snd_nxt = tp->snd_nxt; 203 204 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt); 205 nsacked += tp->snd_nxt - old_snd_nxt; 206 } 207 208 sendalot = FALSE; 209 off = tp->snd_nxt - tp->snd_una; 210 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked); 211 sendwin = min(sendwin, tp->snd_bwnd); 212 213 flags = tcp_outflags[tp->t_state]; 214 /* 215 * Get standard flags, and add SYN or FIN if requested by 'hidden' 216 * state flags. 217 */ 218 if (tp->t_flags & TF_NEEDFIN) 219 flags |= TH_FIN; 220 if (tp->t_flags & TF_NEEDSYN) 221 flags |= TH_SYN; 222 223 /* 224 * If in persist timeout with window of 0, send 1 byte. 225 * Otherwise, if window is small but nonzero 226 * and timer expired, we will send what we can 227 * and go to transmit state. 228 */ 229 if (tp->t_flags & TF_FORCE) { 230 if (sendwin == 0) { 231 /* 232 * If we still have some data to send, then 233 * clear the FIN bit. Usually this would 234 * happen below when it realizes that we 235 * aren't sending all the data. However, 236 * if we have exactly 1 byte of unsent data, 237 * then it won't clear the FIN bit below, 238 * and if we are in persist state, we wind 239 * up sending the packet without recording 240 * that we sent the FIN bit. 241 * 242 * We can't just blindly clear the FIN bit, 243 * because if we don't have any more data 244 * to send then the probe will be the FIN 245 * itself. 246 */ 247 if (off < so->so_snd.ssb_cc) 248 flags &= ~TH_FIN; 249 sendwin = 1; 250 } else { 251 tcp_callout_stop(tp, tp->tt_persist); 252 tp->t_rxtshift = 0; 253 } 254 } 255 256 /* 257 * If snd_nxt == snd_max and we have transmitted a FIN, the 258 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in 259 * a negative length. This can also occur when TCP opens up 260 * its congestion window while receiving additional duplicate 261 * acks after fast-retransmit because TCP will reset snd_nxt 262 * to snd_max after the fast-retransmit. 263 * 264 * In the normal retransmit-FIN-only case, however, snd_nxt will 265 * be set to snd_una, the offset will be 0, and the length may 266 * wind up 0. 267 */ 268 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off; 269 270 /* 271 * Lop off SYN bit if it has already been sent. However, if this 272 * is SYN-SENT state and if segment contains data and if we don't 273 * know that foreign host supports TAO, suppress sending segment. 274 */ 275 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 276 flags &= ~TH_SYN; 277 off--, len++; 278 if (len > 0 && tp->t_state == TCPS_SYN_SENT && 279 ((taop = tcp_gettaocache(&inp->inp_inc)) == NULL || 280 taop->tao_ccsent == 0)) 281 return 0; 282 } 283 284 /* 285 * Be careful not to send data and/or FIN on SYN segments 286 * in cases when no CC option will be sent. 287 * This measure is needed to prevent interoperability problems 288 * with not fully conformant TCP implementations. 289 */ 290 if ((flags & TH_SYN) && 291 ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) || 292 ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) { 293 len = 0; 294 flags &= ~TH_FIN; 295 } 296 297 if (len < 0) { 298 /* 299 * If FIN has been sent but not acked, 300 * but we haven't been called to retransmit, 301 * len will be < 0. Otherwise, window shrank 302 * after we sent into it. If window shrank to 0, 303 * cancel pending retransmit, pull snd_nxt back 304 * to (closed) window, and set the persist timer 305 * if it isn't already going. If the window didn't 306 * close completely, just wait for an ACK. 307 */ 308 len = 0; 309 if (sendwin == 0) { 310 tcp_callout_stop(tp, tp->tt_rexmt); 311 tp->t_rxtshift = 0; 312 tp->snd_nxt = tp->snd_una; 313 if (!tcp_callout_active(tp, tp->tt_persist)) 314 tcp_setpersist(tp); 315 } 316 } 317 318 /* 319 * len will be >= 0 after this point. Truncate to the maximum 320 * segment length and ensure that FIN is removed if the length 321 * no longer contains the last data byte. 322 */ 323 if (len > tp->t_maxseg) { 324 len = tp->t_maxseg; 325 sendalot = TRUE; 326 } 327 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc)) 328 flags &= ~TH_FIN; 329 330 recvwin = ssb_space(&so->so_rcv); 331 332 /* 333 * Sender silly window avoidance. We transmit under the following 334 * conditions when len is non-zero: 335 * 336 * - We have a full segment 337 * - This is the last buffer in a write()/send() and we are 338 * either idle or running NODELAY 339 * - we've timed out (e.g. persist timer) 340 * - we have more then 1/2 the maximum send window's worth of 341 * data (receiver may be limiting the window size) 342 * - we need to retransmit 343 */ 344 if (len) { 345 if (len == tp->t_maxseg) 346 goto send; 347 /* 348 * NOTE! on localhost connections an 'ack' from the remote 349 * end may occur synchronously with the output and cause 350 * us to flush a buffer queued with moretocome. XXX 351 * 352 * note: the len + off check is almost certainly unnecessary. 353 */ 354 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 355 (idle || (tp->t_flags & TF_NODELAY)) && 356 len + off >= so->so_snd.ssb_cc && 357 !(tp->t_flags & TF_NOPUSH)) { 358 goto send; 359 } 360 if (tp->t_flags & TF_FORCE) /* typ. timeout case */ 361 goto send; 362 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 363 goto send; 364 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 365 goto send; 366 } 367 368 /* 369 * Compare available window to amount of window 370 * known to peer (as advertised window less 371 * next expected input). If the difference is at least two 372 * max size segments, or at least 50% of the maximum possible 373 * window, then want to send a window update to peer. 374 */ 375 if (recvwin > 0) { 376 /* 377 * "adv" is the amount we can increase the window, 378 * taking into account that we are limited by 379 * TCP_MAXWIN << tp->rcv_scale. 380 */ 381 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) - 382 (tp->rcv_adv - tp->rcv_nxt); 383 384 /* 385 * This ack case typically occurs when the user has drained 386 * the TCP socket buffer sufficiently to warrent an ack 387 * containing a 'pure window update'... that is, an ack that 388 * ONLY updates the tcp window. 389 * 390 * It is unclear why we would need to do a pure window update 391 * past 2 segments if we are going to do one at 1/2 the high 392 * water mark anyway, especially since under normal conditions 393 * the user program will drain the socket buffer quickly. 394 * The 2-segment pure window update will often add a large 395 * number of extra, unnecessary acks to the stream. 396 * 397 * avoid_pure_win_update now defaults to 1. 398 */ 399 if (avoid_pure_win_update == 0) { 400 if (adv >= (long) (2 * tp->t_maxseg)) 401 goto send; 402 } 403 if (2 * adv >= (long) so->so_rcv.ssb_hiwat) 404 goto send; 405 } 406 407 /* 408 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 409 * is also a catch-all for the retransmit timer timeout case. 410 */ 411 if (tp->t_flags & TF_ACKNOW) 412 goto send; 413 if ((flags & TH_RST) || 414 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN))) 415 goto send; 416 if (SEQ_GT(tp->snd_up, tp->snd_una)) 417 goto send; 418 /* 419 * If our state indicates that FIN should be sent 420 * and we have not yet done so, then we need to send. 421 */ 422 if (flags & TH_FIN && 423 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una)) 424 goto send; 425 426 /* 427 * TCP window updates are not reliable, rather a polling protocol 428 * using ``persist'' packets is used to insure receipt of window 429 * updates. The three ``states'' for the output side are: 430 * idle not doing retransmits or persists 431 * persisting to move a small or zero window 432 * (re)transmitting and thereby not persisting 433 * 434 * tcp_callout_active(tp, tp->tt_persist) 435 * is true when we are in persist state. 436 * The TF_FORCE flag in tp->t_flags 437 * is set when we are called to send a persist packet. 438 * tcp_callout_active(tp, tp->tt_rexmt) 439 * is set when we are retransmitting 440 * The output side is idle when both timers are zero. 441 * 442 * If send window is too small, there is data to transmit, and no 443 * retransmit or persist is pending, then go to persist state. 444 * If nothing happens soon, send when timer expires: 445 * if window is nonzero, transmit what we can, 446 * otherwise force out a byte. 447 */ 448 if (so->so_snd.ssb_cc > 0 && 449 !tcp_callout_active(tp, tp->tt_rexmt) && 450 !tcp_callout_active(tp, tp->tt_persist)) { 451 tp->t_rxtshift = 0; 452 tcp_setpersist(tp); 453 } 454 455 /* 456 * No reason to send a segment, just return. 457 */ 458 return (0); 459 460 send: 461 /* 462 * Before ESTABLISHED, force sending of initial options 463 * unless TCP set not to do any options. 464 * NOTE: we assume that the IP/TCP header plus TCP options 465 * always fit in a single mbuf, leaving room for a maximum 466 * link header, i.e. 467 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES 468 */ 469 optlen = 0; 470 if (isipv6) 471 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 472 else 473 hdrlen = sizeof(struct tcpiphdr); 474 if (flags & TH_SYN) { 475 tp->snd_nxt = tp->iss; 476 if (!(tp->t_flags & TF_NOOPT)) { 477 u_short mss; 478 479 opt[0] = TCPOPT_MAXSEG; 480 opt[1] = TCPOLEN_MAXSEG; 481 mss = htons((u_short) tcp_mssopt(tp)); 482 memcpy(opt + 2, &mss, sizeof mss); 483 optlen = TCPOLEN_MAXSEG; 484 485 if ((tp->t_flags & TF_REQ_SCALE) && 486 (!(flags & TH_ACK) || 487 (tp->t_flags & TF_RCVD_SCALE))) { 488 *((u_int32_t *)(opt + optlen)) = htonl( 489 TCPOPT_NOP << 24 | 490 TCPOPT_WINDOW << 16 | 491 TCPOLEN_WINDOW << 8 | 492 tp->request_r_scale); 493 optlen += 4; 494 } 495 496 if ((tcp_do_sack && !(flags & TH_ACK)) || 497 tp->t_flags & TF_SACK_PERMITTED) { 498 uint32_t *lp = (uint32_t *)(opt + optlen); 499 500 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED); 501 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED; 502 } 503 } 504 } 505 506 /* 507 * Send a timestamp and echo-reply if this is a SYN and our side 508 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 509 * and our peer have sent timestamps in our SYN's. 510 */ 511 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 512 !(flags & TH_RST) && 513 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) { 514 u_int32_t *lp = (u_int32_t *)(opt + optlen); 515 516 /* Form timestamp option as shown in appendix A of RFC 1323. */ 517 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 518 *lp++ = htonl(ticks); 519 *lp = htonl(tp->ts_recent); 520 optlen += TCPOLEN_TSTAMP_APPA; 521 } 522 523 /* 524 * Send `CC-family' options if our side wants to use them (TF_REQ_CC), 525 * options are allowed (!TF_NOOPT) and it's not a RST. 526 */ 527 if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC && 528 !(flags & TH_RST)) { 529 switch (flags & (TH_SYN | TH_ACK)) { 530 /* 531 * This is a normal ACK, send CC if we received CC before 532 * from our peer. 533 */ 534 case TH_ACK: 535 if (!(tp->t_flags & TF_RCVD_CC)) 536 break; 537 /*FALLTHROUGH*/ 538 539 /* 540 * We can only get here in T/TCP's SYN_SENT* state, when 541 * we're a sending a non-SYN segment without waiting for 542 * the ACK of our SYN. A check above assures that we only 543 * do this if our peer understands T/TCP. 544 */ 545 case 0: 546 opt[optlen++] = TCPOPT_NOP; 547 opt[optlen++] = TCPOPT_NOP; 548 opt[optlen++] = TCPOPT_CC; 549 opt[optlen++] = TCPOLEN_CC; 550 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 551 optlen += 4; 552 break; 553 554 /* 555 * This is our initial SYN, check whether we have to use 556 * CC or CC.new. 557 */ 558 case TH_SYN: 559 opt[optlen++] = TCPOPT_NOP; 560 opt[optlen++] = TCPOPT_NOP; 561 opt[optlen++] = tp->t_flags & TF_SENDCCNEW ? 562 TCPOPT_CCNEW : TCPOPT_CC; 563 opt[optlen++] = TCPOLEN_CC; 564 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 565 optlen += 4; 566 break; 567 568 /* 569 * This is a SYN,ACK; send CC and CC.echo if we received 570 * CC from our peer. 571 */ 572 case (TH_SYN | TH_ACK): 573 if (tp->t_flags & TF_RCVD_CC) { 574 opt[optlen++] = TCPOPT_NOP; 575 opt[optlen++] = TCPOPT_NOP; 576 opt[optlen++] = TCPOPT_CC; 577 opt[optlen++] = TCPOLEN_CC; 578 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 579 optlen += 4; 580 opt[optlen++] = TCPOPT_NOP; 581 opt[optlen++] = TCPOPT_NOP; 582 opt[optlen++] = TCPOPT_CCECHO; 583 opt[optlen++] = TCPOLEN_CC; 584 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_recv); 585 optlen += 4; 586 } 587 break; 588 } 589 } 590 591 /* 592 * If this is a SACK connection and we have a block to report, 593 * fill in the SACK blocks in the TCP options. 594 */ 595 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) == 596 TF_SACK_PERMITTED && 597 (!LIST_EMPTY(&tp->t_segq) || 598 tp->reportblk.rblk_start != tp->reportblk.rblk_end)) 599 tcp_sack_fill_report(tp, opt, &optlen); 600 601 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options")); 602 hdrlen += optlen; 603 604 if (isipv6) { 605 ipoptlen = ip6_optlen(inp); 606 } else { 607 if (inp->inp_options) { 608 ipoptlen = inp->inp_options->m_len - 609 offsetof(struct ipoption, ipopt_list); 610 } else { 611 ipoptlen = 0; 612 } 613 } 614 #ifdef IPSEC 615 ipoptlen += ipsec_hdrsiz_tcp(tp); 616 #endif 617 618 /* 619 * Adjust data length if insertion of options will bump the packet 620 * length beyond the t_maxopd length. Clear FIN to prevent premature 621 * closure since there is still more data to send after this (now 622 * truncated) packet. 623 * 624 * If just the options do not fit we are in a no-win situation and 625 * we treat it as an unreachable host. 626 */ 627 if (len + optlen + ipoptlen > tp->t_maxopd) { 628 if (tp->t_maxopd <= optlen + ipoptlen) { 629 static time_t last_optlen_report; 630 631 if (last_optlen_report != time_second) { 632 last_optlen_report = time_second; 633 kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd); 634 } 635 error = EHOSTUNREACH; 636 goto out; 637 } else { 638 flags &= ~TH_FIN; 639 len = tp->t_maxopd - optlen - ipoptlen; 640 sendalot = TRUE; 641 } 642 } 643 644 #ifdef INET6 645 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big")); 646 #else 647 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big")); 648 #endif 649 650 /* 651 * Grab a header mbuf, attaching a copy of data to 652 * be transmitted, and initialize the header from 653 * the template for sends on this connection. 654 */ 655 if (len) { 656 if ((tp->t_flags & TF_FORCE) && len == 1) 657 tcpstat.tcps_sndprobe++; 658 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 659 if (tp->snd_nxt == tp->snd_una) 660 tp->snd_max_rexmt = tp->snd_max; 661 tcpstat.tcps_sndrexmitpack++; 662 tcpstat.tcps_sndrexmitbyte += len; 663 } else { 664 tcpstat.tcps_sndpack++; 665 tcpstat.tcps_sndbyte += len; 666 } 667 #ifdef notyet 668 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len, 669 max_linkhdr + hdrlen)) == NULL) { 670 error = ENOBUFS; 671 goto out; 672 } 673 /* 674 * m_copypack left space for our hdr; use it. 675 */ 676 m->m_len += hdrlen; 677 m->m_data -= hdrlen; 678 #else 679 #ifndef INET6 680 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 681 #else 682 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER, 683 M_PKTHDR, NULL); 684 #endif 685 if (m == NULL) { 686 error = ENOBUFS; 687 goto out; 688 } 689 m->m_data += max_linkhdr; 690 m->m_len = hdrlen; 691 if (len <= MHLEN - hdrlen - max_linkhdr) { 692 m_copydata(so->so_snd.ssb_mb, off, (int) len, 693 mtod(m, caddr_t) + hdrlen); 694 m->m_len += len; 695 } else { 696 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len); 697 if (m->m_next == NULL) { 698 m_free(m); 699 error = ENOBUFS; 700 goto out; 701 } 702 } 703 #endif 704 /* 705 * If we're sending everything we've got, set PUSH. 706 * (This will keep happy those implementations which only 707 * give data to the user when a buffer fills or 708 * a PUSH comes in.) 709 */ 710 if (off + len == so->so_snd.ssb_cc) 711 flags |= TH_PUSH; 712 } else { 713 if (tp->t_flags & TF_ACKNOW) 714 tcpstat.tcps_sndacks++; 715 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 716 tcpstat.tcps_sndctrl++; 717 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 718 tcpstat.tcps_sndurg++; 719 else 720 tcpstat.tcps_sndwinup++; 721 722 MGETHDR(m, MB_DONTWAIT, MT_HEADER); 723 if (m == NULL) { 724 error = ENOBUFS; 725 goto out; 726 } 727 if (isipv6 && 728 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN) 729 MH_ALIGN(m, hdrlen); 730 else 731 m->m_data += max_linkhdr; 732 m->m_len = hdrlen; 733 } 734 m->m_pkthdr.rcvif = (struct ifnet *)0; 735 if (isipv6) { 736 ip6 = mtod(m, struct ip6_hdr *); 737 th = (struct tcphdr *)(ip6 + 1); 738 tcp_fillheaders(tp, ip6, th); 739 } else { 740 ip = mtod(m, struct ip *); 741 ipov = (struct ipovly *)ip; 742 th = (struct tcphdr *)(ip + 1); 743 /* this picks up the pseudo header (w/o the length) */ 744 tcp_fillheaders(tp, ip, th); 745 } 746 747 /* 748 * Fill in fields, remembering maximum advertised 749 * window for use in delaying messages about window sizes. 750 * If resending a FIN, be sure not to use a new sequence number. 751 */ 752 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 753 tp->snd_nxt == tp->snd_max) 754 tp->snd_nxt--; 755 /* 756 * If we are doing retransmissions, then snd_nxt will 757 * not reflect the first unsent octet. For ACK only 758 * packets, we do not want the sequence number of the 759 * retransmitted packet, we want the sequence number 760 * of the next unsent octet. So, if there is no data 761 * (and no SYN or FIN), use snd_max instead of snd_nxt 762 * when filling in ti_seq. But if we are in persist 763 * state, snd_max might reflect one byte beyond the 764 * right edge of the window, so use snd_nxt in that 765 * case, since we know we aren't doing a retransmission. 766 * (retransmit and persist are mutually exclusive...) 767 */ 768 if (len || (flags & (TH_SYN|TH_FIN)) || 769 tcp_callout_active(tp, tp->tt_persist)) 770 th->th_seq = htonl(tp->snd_nxt); 771 else 772 th->th_seq = htonl(tp->snd_max); 773 th->th_ack = htonl(tp->rcv_nxt); 774 if (optlen) { 775 bcopy(opt, th + 1, optlen); 776 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 777 } 778 th->th_flags = flags; 779 /* 780 * Calculate receive window. Don't shrink window, 781 * but avoid silly window syndrome. 782 */ 783 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) && 784 recvwin < (long)tp->t_maxseg) 785 recvwin = 0; 786 if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 787 recvwin = (long)(tp->rcv_adv - tp->rcv_nxt); 788 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale) 789 recvwin = (long)TCP_MAXWIN << tp->rcv_scale; 790 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale)); 791 792 /* 793 * Adjust the RXWIN0SENT flag - indicate that we have advertised 794 * a 0 window. This may cause the remote transmitter to stall. This 795 * flag tells soreceive() to disable delayed acknowledgements when 796 * draining the buffer. This can occur if the receiver is attempting 797 * to read more data then can be buffered prior to transmitting on 798 * the connection. 799 */ 800 if (recvwin == 0) 801 tp->t_flags |= TF_RXWIN0SENT; 802 else 803 tp->t_flags &= ~TF_RXWIN0SENT; 804 805 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 806 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 807 th->th_flags |= TH_URG; 808 } else { 809 /* 810 * If no urgent pointer to send, then we pull 811 * the urgent pointer to the left edge of the send window 812 * so that it doesn't drift into the send window on sequence 813 * number wraparound. 814 */ 815 tp->snd_up = tp->snd_una; /* drag it along */ 816 } 817 818 /* 819 * Put TCP length in extended header, and then 820 * checksum extended header and data. 821 */ 822 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 823 if (isipv6) { 824 /* 825 * ip6_plen is not need to be filled now, and will be filled 826 * in ip6_output(). 827 */ 828 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 829 sizeof(struct tcphdr) + optlen + len); 830 } else { 831 m->m_pkthdr.csum_flags = CSUM_TCP; 832 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 833 if (len + optlen) 834 th->th_sum = in_addword(th->th_sum, 835 htons((u_short)(optlen + len))); 836 837 /* IP version must be set here for ipv4/ipv6 checking later */ 838 KASSERT(ip->ip_v == IPVERSION, 839 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 840 } 841 842 /* 843 * In transmit state, time the transmission and arrange for 844 * the retransmit. In persist state, just set snd_max. 845 */ 846 if (!(tp->t_flags & TF_FORCE) || 847 !tcp_callout_active(tp, tp->tt_persist)) { 848 tcp_seq startseq = tp->snd_nxt; 849 850 /* 851 * Advance snd_nxt over sequence space of this segment. 852 */ 853 if (flags & (TH_SYN | TH_FIN)) { 854 if (flags & TH_SYN) 855 tp->snd_nxt++; 856 if (flags & TH_FIN) { 857 tp->snd_nxt++; 858 tp->t_flags |= TF_SENTFIN; 859 } 860 } 861 tp->snd_nxt += len; 862 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 863 tp->snd_max = tp->snd_nxt; 864 /* 865 * Time this transmission if not a retransmission and 866 * not currently timing anything. 867 */ 868 if (tp->t_rtttime == 0) { 869 tp->t_rtttime = ticks; 870 tp->t_rtseq = startseq; 871 tcpstat.tcps_segstimed++; 872 } 873 } 874 875 /* 876 * Set retransmit timer if not currently set, 877 * and not doing a pure ack or a keep-alive probe. 878 * Initial value for retransmit timer is smoothed 879 * round-trip time + 2 * round-trip time variance. 880 * Initialize shift counter which is used for backoff 881 * of retransmit time. 882 */ 883 if (!tcp_callout_active(tp, tp->tt_rexmt) && 884 tp->snd_nxt != tp->snd_una) { 885 if (tcp_callout_active(tp, tp->tt_persist)) { 886 tcp_callout_stop(tp, tp->tt_persist); 887 tp->t_rxtshift = 0; 888 } 889 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur, 890 tcp_timer_rexmt); 891 } 892 } else { 893 /* 894 * Persist case, update snd_max but since we are in 895 * persist mode (no window) we do not update snd_nxt. 896 */ 897 int xlen = len; 898 if (flags & TH_SYN) 899 ++xlen; 900 if (flags & TH_FIN) { 901 ++xlen; 902 tp->t_flags |= TF_SENTFIN; 903 } 904 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 905 tp->snd_max = tp->snd_nxt + xlen; 906 } 907 908 #ifdef TCPDEBUG 909 /* 910 * Trace. 911 */ 912 if (so->so_options & SO_DEBUG) 913 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 914 #endif 915 916 /* 917 * Fill in IP length and desired time to live and 918 * send to IP level. There should be a better way 919 * to handle ttl and tos; we could keep them in 920 * the template, but need a way to checksum without them. 921 */ 922 /* 923 * m->m_pkthdr.len should have been set before cksum calcuration, 924 * because in6_cksum() need it. 925 */ 926 if (isipv6) { 927 /* 928 * we separately set hoplimit for every segment, since the 929 * user might want to change the value via setsockopt. 930 * Also, desired default hop limit might be changed via 931 * Neighbor Discovery. 932 */ 933 ip6->ip6_hlim = in6_selecthlim(inp, 934 (inp->in6p_route.ro_rt ? 935 inp->in6p_route.ro_rt->rt_ifp : NULL)); 936 937 /* TODO: IPv6 IP6TOS_ECT bit on */ 938 error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route, 939 (so->so_options & SO_DONTROUTE), NULL, NULL, 940 inp); 941 } else { 942 struct rtentry *rt; 943 ip->ip_len = m->m_pkthdr.len; 944 #ifdef INET6 945 if (INP_CHECK_SOCKAF(so, AF_INET6)) 946 ip->ip_ttl = in6_selecthlim(inp, 947 (inp->in6p_route.ro_rt ? 948 inp->in6p_route.ro_rt->rt_ifp : NULL)); 949 else 950 #endif 951 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */ 952 953 ip->ip_tos = inp->inp_ip_tos; /* XXX */ 954 /* 955 * See if we should do MTU discovery. 956 * We do it only if the following are true: 957 * 1) we have a valid route to the destination 958 * 2) the MTU is not locked (if it is, 959 * then discovery has been disabled) 960 */ 961 if (path_mtu_discovery && 962 (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) && 963 !(rt->rt_rmx.rmx_locks & RTV_MTU)) 964 ip->ip_off |= IP_DF; 965 966 error = ip_output(m, inp->inp_options, &inp->inp_route, 967 (so->so_options & SO_DONTROUTE) | 968 IP_DEBUGROUTE, NULL, inp); 969 } 970 if (error) { 971 972 /* 973 * We know that the packet was lost, so back out the 974 * sequence number advance, if any. 975 */ 976 if (!(tp->t_flags & TF_FORCE) || 977 !tcp_callout_active(tp, tp->tt_persist)) { 978 /* 979 * No need to check for TH_FIN here because 980 * the TF_SENTFIN flag handles that case. 981 */ 982 if (!(flags & TH_SYN)) 983 tp->snd_nxt -= len; 984 } 985 986 out: 987 if (error == ENOBUFS) { 988 /* 989 * If we can't send, make sure there is something 990 * to get us going again later. Persist state 991 * is not necessarily right, but it is close enough. 992 */ 993 if (!tcp_callout_active(tp, tp->tt_rexmt) && 994 !tcp_callout_active(tp, tp->tt_persist)) { 995 tp->t_rxtshift = 0; 996 tcp_setpersist(tp); 997 } 998 tcp_quench(inp, 0); 999 return (0); 1000 } 1001 if (error == EMSGSIZE) { 1002 /* 1003 * ip_output() will have already fixed the route 1004 * for us. tcp_mtudisc() will, as its last action, 1005 * initiate retransmission, so it is important to 1006 * not do so here. 1007 */ 1008 tcp_mtudisc(inp, 0); 1009 return 0; 1010 } 1011 if ((error == EHOSTUNREACH || error == ENETDOWN) && 1012 TCPS_HAVERCVDSYN(tp->t_state)) { 1013 tp->t_softerror = error; 1014 return (0); 1015 } 1016 return (error); 1017 } 1018 tcpstat.tcps_sndtotal++; 1019 1020 /* 1021 * Data sent (as far as we can tell). 1022 * If this advertises a larger window than any other segment, 1023 * then remember the size of the advertised window. 1024 * Any pending ACK has now been sent. 1025 */ 1026 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) 1027 tp->rcv_adv = tp->rcv_nxt + recvwin; 1028 tp->last_ack_sent = tp->rcv_nxt; 1029 tp->t_flags &= ~TF_ACKNOW; 1030 if (tcp_delack_enabled) 1031 tcp_callout_stop(tp, tp->tt_delack); 1032 if (sendalot) 1033 goto again; 1034 return (0); 1035 } 1036 1037 void 1038 tcp_setpersist(struct tcpcb *tp) 1039 { 1040 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1041 int tt; 1042 1043 if (tcp_callout_active(tp, tp->tt_rexmt)) 1044 panic("tcp_setpersist: retransmit pending"); 1045 /* 1046 * Start/restart persistance timer. 1047 */ 1048 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN, 1049 TCPTV_PERSMAX); 1050 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist); 1051 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1052 tp->t_rxtshift++; 1053 } 1054