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 int tcp_do_autosndbuf = 1; 132 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 133 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing"); 134 135 int tcp_autosndbuf_inc = 8*1024; 136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 137 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer"); 138 139 int tcp_autosndbuf_max = 16*1024*1024; 140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 141 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer"); 142 143 /* 144 * Tcp output routine: figure out what should be sent and send it. 145 */ 146 int 147 tcp_output(struct tcpcb *tp) 148 { 149 struct inpcb * const inp = tp->t_inpcb; 150 struct socket *so = inp->inp_socket; 151 long len, recvwin, sendwin; 152 int nsacked = 0; 153 int off, flags, error; 154 struct mbuf *m; 155 struct ip *ip = NULL; 156 struct ipovly *ipov = NULL; 157 struct tcphdr *th; 158 u_char opt[TCP_MAXOLEN]; 159 unsigned int ipoptlen, optlen, hdrlen; 160 int idle; 161 boolean_t sendalot; 162 struct ip6_hdr *ip6 = NULL; 163 #ifdef INET6 164 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 165 #else 166 const boolean_t isipv6 = FALSE; 167 #endif 168 struct rmxp_tao *taop; 169 170 /* 171 * Determine length of data that should be transmitted, 172 * and flags that will be used. 173 * If there is some data or critical controls (SYN, RST) 174 * to send, then transmit; otherwise, investigate further. 175 */ 176 177 /* 178 * If we have been idle for a while, the send congestion window 179 * could be no longer representative of the current state of the link. 180 * So unless we are expecting more acks to come in, slow-start from 181 * scratch to re-determine the send congestion window. 182 */ 183 if (tp->snd_max == tp->snd_una && 184 (ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 185 if (tcp_do_rfc3390) { 186 int initial_cwnd = 187 min(4 * tp->t_maxseg, max(2 * tp->t_maxseg, 4380)); 188 189 tp->snd_cwnd = min(tp->snd_cwnd, initial_cwnd); 190 } else { 191 tp->snd_cwnd = tp->t_maxseg; 192 } 193 tp->snd_wacked = 0; 194 } 195 196 /* 197 * Calculate whether the transmit stream was previously idle 198 * and adjust TF_LASTIDLE for the next time. 199 */ 200 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 201 if (idle && (tp->t_flags & TF_MORETOCOME)) 202 tp->t_flags |= TF_LASTIDLE; 203 else 204 tp->t_flags &= ~TF_LASTIDLE; 205 206 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 207 !IN_FASTRECOVERY(tp)) 208 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt); 209 210 again: 211 /* Make use of SACK information when slow-starting after a RTO. */ 212 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 213 !IN_FASTRECOVERY(tp)) { 214 tcp_seq old_snd_nxt = tp->snd_nxt; 215 216 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt); 217 nsacked += tp->snd_nxt - old_snd_nxt; 218 } 219 220 sendalot = FALSE; 221 off = tp->snd_nxt - tp->snd_una; 222 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked); 223 sendwin = min(sendwin, tp->snd_bwnd); 224 225 flags = tcp_outflags[tp->t_state]; 226 /* 227 * Get standard flags, and add SYN or FIN if requested by 'hidden' 228 * state flags. 229 */ 230 if (tp->t_flags & TF_NEEDFIN) 231 flags |= TH_FIN; 232 if (tp->t_flags & TF_NEEDSYN) 233 flags |= TH_SYN; 234 235 /* 236 * If in persist timeout with window of 0, send 1 byte. 237 * Otherwise, if window is small but nonzero 238 * and timer expired, we will send what we can 239 * and go to transmit state. 240 */ 241 if (tp->t_flags & TF_FORCE) { 242 if (sendwin == 0) { 243 /* 244 * If we still have some data to send, then 245 * clear the FIN bit. Usually this would 246 * happen below when it realizes that we 247 * aren't sending all the data. However, 248 * if we have exactly 1 byte of unsent data, 249 * then it won't clear the FIN bit below, 250 * and if we are in persist state, we wind 251 * up sending the packet without recording 252 * that we sent the FIN bit. 253 * 254 * We can't just blindly clear the FIN bit, 255 * because if we don't have any more data 256 * to send then the probe will be the FIN 257 * itself. 258 */ 259 if (off < so->so_snd.ssb_cc) 260 flags &= ~TH_FIN; 261 sendwin = 1; 262 } else { 263 tcp_callout_stop(tp, tp->tt_persist); 264 tp->t_rxtshift = 0; 265 } 266 } 267 268 /* 269 * If snd_nxt == snd_max and we have transmitted a FIN, the 270 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in 271 * a negative length. This can also occur when TCP opens up 272 * its congestion window while receiving additional duplicate 273 * acks after fast-retransmit because TCP will reset snd_nxt 274 * to snd_max after the fast-retransmit. 275 * 276 * In the normal retransmit-FIN-only case, however, snd_nxt will 277 * be set to snd_una, the offset will be 0, and the length may 278 * wind up 0. 279 */ 280 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off; 281 282 /* 283 * Lop off SYN bit if it has already been sent. However, if this 284 * is SYN-SENT state and if segment contains data and if we don't 285 * know that foreign host supports TAO, suppress sending segment. 286 */ 287 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 288 flags &= ~TH_SYN; 289 off--, len++; 290 if (len > 0 && tp->t_state == TCPS_SYN_SENT && 291 ((taop = tcp_gettaocache(&inp->inp_inc)) == NULL || 292 taop->tao_ccsent == 0)) 293 return 0; 294 } 295 296 /* 297 * Be careful not to send data and/or FIN on SYN segments 298 * in cases when no CC option will be sent. 299 * This measure is needed to prevent interoperability problems 300 * with not fully conformant TCP implementations. 301 */ 302 if ((flags & TH_SYN) && 303 ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) || 304 ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) { 305 len = 0; 306 flags &= ~TH_FIN; 307 } 308 309 if (len < 0) { 310 /* 311 * If FIN has been sent but not acked, 312 * but we haven't been called to retransmit, 313 * len will be < 0. Otherwise, window shrank 314 * after we sent into it. If window shrank to 0, 315 * cancel pending retransmit, pull snd_nxt back 316 * to (closed) window, and set the persist timer 317 * if it isn't already going. If the window didn't 318 * close completely, just wait for an ACK. 319 */ 320 len = 0; 321 if (sendwin == 0) { 322 tcp_callout_stop(tp, tp->tt_rexmt); 323 tp->t_rxtshift = 0; 324 tp->snd_nxt = tp->snd_una; 325 if (!tcp_callout_active(tp, tp->tt_persist)) 326 tcp_setpersist(tp); 327 } 328 } 329 330 KASSERT(len >= 0, ("%s: len < 0", __func__)); 331 /* 332 * Automatic sizing of send socket buffer. Often the send buffer 333 * size is not optimally adjusted to the actual network conditions 334 * at hand (delay bandwidth product). Setting the buffer size too 335 * small limits throughput on links with high bandwidth and high 336 * delay (eg. trans-continental/oceanic links). Setting the 337 * buffer size too big consumes too much real kernel memory, 338 * especially with many connections on busy servers. 339 * 340 * The criteria to step up the send buffer one notch are: 341 * 1. receive window of remote host is larger than send buffer 342 * (with a fudge factor of 5/4th); 343 * 2. send buffer is filled to 7/8th with data (so we actually 344 * have data to make use of it); 345 * 3. send buffer fill has not hit maximal automatic size; 346 * 4. our send window (slow start and cogestion controlled) is 347 * larger than sent but unacknowledged data in send buffer. 348 * 349 * The remote host receive window scaling factor may limit the 350 * growing of the send buffer before it reaches its allowed 351 * maximum. 352 * 353 * It scales directly with slow start or congestion window 354 * and does at most one step per received ACK. This fast 355 * scaling has the drawback of growing the send buffer beyond 356 * what is strictly necessary to make full use of a given 357 * delay*bandwith product. However testing has shown this not 358 * to be much of an problem. At worst we are trading wasting 359 * of available bandwith (the non-use of it) for wasting some 360 * socket buffer memory. 361 * 362 * TODO: Shrink send buffer during idle periods together 363 * with congestion window. Requires another timer. Has to 364 * wait for upcoming tcp timer rewrite. 365 */ 366 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) { 367 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat && 368 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) && 369 so->so_snd.ssb_cc < tcp_autosndbuf_max && 370 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) { 371 if (!ssb_reserve(&so->so_snd, 372 min(so->so_snd.ssb_hiwat + tcp_autosndbuf_inc, 373 tcp_autosndbuf_max), so, NULL)) 374 so->so_snd.ssb_flags &= ~SSB_AUTOSIZE; 375 } 376 } 377 378 /* 379 * Truncate to the maximum segment length and ensure that FIN is 380 * removed if the length no longer contains the last data byte. 381 */ 382 if (len > tp->t_maxseg) { 383 len = tp->t_maxseg; 384 sendalot = TRUE; 385 } 386 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc)) 387 flags &= ~TH_FIN; 388 389 recvwin = ssb_space(&so->so_rcv); 390 391 /* 392 * Sender silly window avoidance. We transmit under the following 393 * conditions when len is non-zero: 394 * 395 * - We have a full segment 396 * - This is the last buffer in a write()/send() and we are 397 * either idle or running NODELAY 398 * - we've timed out (e.g. persist timer) 399 * - we have more then 1/2 the maximum send window's worth of 400 * data (receiver may be limiting the window size) 401 * - we need to retransmit 402 */ 403 if (len) { 404 if (len == tp->t_maxseg) 405 goto send; 406 /* 407 * NOTE! on localhost connections an 'ack' from the remote 408 * end may occur synchronously with the output and cause 409 * us to flush a buffer queued with moretocome. XXX 410 * 411 * note: the len + off check is almost certainly unnecessary. 412 */ 413 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 414 (idle || (tp->t_flags & TF_NODELAY)) && 415 len + off >= so->so_snd.ssb_cc && 416 !(tp->t_flags & TF_NOPUSH)) { 417 goto send; 418 } 419 if (tp->t_flags & TF_FORCE) /* typ. timeout case */ 420 goto send; 421 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 422 goto send; 423 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 424 goto send; 425 } 426 427 /* 428 * Compare available window to amount of window 429 * known to peer (as advertised window less 430 * next expected input). If the difference is at least two 431 * max size segments, or at least 50% of the maximum possible 432 * window, then want to send a window update to peer. 433 */ 434 if (recvwin > 0) { 435 /* 436 * "adv" is the amount we can increase the window, 437 * taking into account that we are limited by 438 * TCP_MAXWIN << tp->rcv_scale. 439 */ 440 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) - 441 (tp->rcv_adv - tp->rcv_nxt); 442 443 /* 444 * This ack case typically occurs when the user has drained 445 * the TCP socket buffer sufficiently to warrent an ack 446 * containing a 'pure window update'... that is, an ack that 447 * ONLY updates the tcp window. 448 * 449 * It is unclear why we would need to do a pure window update 450 * past 2 segments if we are going to do one at 1/2 the high 451 * water mark anyway, especially since under normal conditions 452 * the user program will drain the socket buffer quickly. 453 * The 2-segment pure window update will often add a large 454 * number of extra, unnecessary acks to the stream. 455 * 456 * avoid_pure_win_update now defaults to 1. 457 */ 458 if (avoid_pure_win_update == 0) { 459 if (adv >= (long) (2 * tp->t_maxseg)) 460 goto send; 461 } 462 if (2 * adv >= (long) so->so_rcv.ssb_hiwat) 463 goto send; 464 } 465 466 /* 467 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 468 * is also a catch-all for the retransmit timer timeout case. 469 */ 470 if (tp->t_flags & TF_ACKNOW) 471 goto send; 472 if ((flags & TH_RST) || 473 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN))) 474 goto send; 475 if (SEQ_GT(tp->snd_up, tp->snd_una)) 476 goto send; 477 /* 478 * If our state indicates that FIN should be sent 479 * and we have not yet done so, then we need to send. 480 */ 481 if (flags & TH_FIN && 482 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una)) 483 goto send; 484 485 /* 486 * TCP window updates are not reliable, rather a polling protocol 487 * using ``persist'' packets is used to insure receipt of window 488 * updates. The three ``states'' for the output side are: 489 * idle not doing retransmits or persists 490 * persisting to move a small or zero window 491 * (re)transmitting and thereby not persisting 492 * 493 * tcp_callout_active(tp, tp->tt_persist) 494 * is true when we are in persist state. 495 * The TF_FORCE flag in tp->t_flags 496 * is set when we are called to send a persist packet. 497 * tcp_callout_active(tp, tp->tt_rexmt) 498 * is set when we are retransmitting 499 * The output side is idle when both timers are zero. 500 * 501 * If send window is too small, there is data to transmit, and no 502 * retransmit or persist is pending, then go to persist state. 503 * If nothing happens soon, send when timer expires: 504 * if window is nonzero, transmit what we can, 505 * otherwise force out a byte. 506 */ 507 if (so->so_snd.ssb_cc > 0 && 508 !tcp_callout_active(tp, tp->tt_rexmt) && 509 !tcp_callout_active(tp, tp->tt_persist)) { 510 tp->t_rxtshift = 0; 511 tcp_setpersist(tp); 512 } 513 514 /* 515 * No reason to send a segment, just return. 516 */ 517 return (0); 518 519 send: 520 /* 521 * Before ESTABLISHED, force sending of initial options 522 * unless TCP set not to do any options. 523 * NOTE: we assume that the IP/TCP header plus TCP options 524 * always fit in a single mbuf, leaving room for a maximum 525 * link header, i.e. 526 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES 527 */ 528 optlen = 0; 529 if (isipv6) 530 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 531 else 532 hdrlen = sizeof(struct tcpiphdr); 533 if (flags & TH_SYN) { 534 tp->snd_nxt = tp->iss; 535 if (!(tp->t_flags & TF_NOOPT)) { 536 u_short mss; 537 538 opt[0] = TCPOPT_MAXSEG; 539 opt[1] = TCPOLEN_MAXSEG; 540 mss = htons((u_short) tcp_mssopt(tp)); 541 memcpy(opt + 2, &mss, sizeof mss); 542 optlen = TCPOLEN_MAXSEG; 543 544 if ((tp->t_flags & TF_REQ_SCALE) && 545 (!(flags & TH_ACK) || 546 (tp->t_flags & TF_RCVD_SCALE))) { 547 *((u_int32_t *)(opt + optlen)) = htonl( 548 TCPOPT_NOP << 24 | 549 TCPOPT_WINDOW << 16 | 550 TCPOLEN_WINDOW << 8 | 551 tp->request_r_scale); 552 optlen += 4; 553 } 554 555 if ((tcp_do_sack && !(flags & TH_ACK)) || 556 tp->t_flags & TF_SACK_PERMITTED) { 557 uint32_t *lp = (uint32_t *)(opt + optlen); 558 559 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED); 560 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED; 561 } 562 } 563 } 564 565 /* 566 * Send a timestamp and echo-reply if this is a SYN and our side 567 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 568 * and our peer have sent timestamps in our SYN's. 569 */ 570 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 571 !(flags & TH_RST) && 572 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) { 573 u_int32_t *lp = (u_int32_t *)(opt + optlen); 574 575 /* Form timestamp option as shown in appendix A of RFC 1323. */ 576 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 577 *lp++ = htonl(ticks); 578 *lp = htonl(tp->ts_recent); 579 optlen += TCPOLEN_TSTAMP_APPA; 580 } 581 582 /* Set receive buffer autosizing timestamp. */ 583 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) 584 tp->rfbuf_ts = ticks; 585 586 /* 587 * Send `CC-family' options if our side wants to use them (TF_REQ_CC), 588 * options are allowed (!TF_NOOPT) and it's not a RST. 589 */ 590 if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC && 591 !(flags & TH_RST)) { 592 switch (flags & (TH_SYN | TH_ACK)) { 593 /* 594 * This is a normal ACK, send CC if we received CC before 595 * from our peer. 596 */ 597 case TH_ACK: 598 if (!(tp->t_flags & TF_RCVD_CC)) 599 break; 600 /*FALLTHROUGH*/ 601 602 /* 603 * We can only get here in T/TCP's SYN_SENT* state, when 604 * we're a sending a non-SYN segment without waiting for 605 * the ACK of our SYN. A check above assures that we only 606 * do this if our peer understands T/TCP. 607 */ 608 case 0: 609 opt[optlen++] = TCPOPT_NOP; 610 opt[optlen++] = TCPOPT_NOP; 611 opt[optlen++] = TCPOPT_CC; 612 opt[optlen++] = TCPOLEN_CC; 613 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 614 optlen += 4; 615 break; 616 617 /* 618 * This is our initial SYN, check whether we have to use 619 * CC or CC.new. 620 */ 621 case TH_SYN: 622 opt[optlen++] = TCPOPT_NOP; 623 opt[optlen++] = TCPOPT_NOP; 624 opt[optlen++] = tp->t_flags & TF_SENDCCNEW ? 625 TCPOPT_CCNEW : TCPOPT_CC; 626 opt[optlen++] = TCPOLEN_CC; 627 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 628 optlen += 4; 629 break; 630 631 /* 632 * This is a SYN,ACK; send CC and CC.echo if we received 633 * CC from our peer. 634 */ 635 case (TH_SYN | TH_ACK): 636 if (tp->t_flags & TF_RCVD_CC) { 637 opt[optlen++] = TCPOPT_NOP; 638 opt[optlen++] = TCPOPT_NOP; 639 opt[optlen++] = TCPOPT_CC; 640 opt[optlen++] = TCPOLEN_CC; 641 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 642 optlen += 4; 643 opt[optlen++] = TCPOPT_NOP; 644 opt[optlen++] = TCPOPT_NOP; 645 opt[optlen++] = TCPOPT_CCECHO; 646 opt[optlen++] = TCPOLEN_CC; 647 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_recv); 648 optlen += 4; 649 } 650 break; 651 } 652 } 653 654 /* 655 * If this is a SACK connection and we have a block to report, 656 * fill in the SACK blocks in the TCP options. 657 */ 658 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) == 659 TF_SACK_PERMITTED && 660 (!LIST_EMPTY(&tp->t_segq) || 661 tp->reportblk.rblk_start != tp->reportblk.rblk_end)) 662 tcp_sack_fill_report(tp, opt, &optlen); 663 664 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options")); 665 hdrlen += optlen; 666 667 if (isipv6) { 668 ipoptlen = ip6_optlen(inp); 669 } else { 670 if (inp->inp_options) { 671 ipoptlen = inp->inp_options->m_len - 672 offsetof(struct ipoption, ipopt_list); 673 } else { 674 ipoptlen = 0; 675 } 676 } 677 #ifdef IPSEC 678 ipoptlen += ipsec_hdrsiz_tcp(tp); 679 #endif 680 681 /* 682 * Adjust data length if insertion of options will bump the packet 683 * length beyond the t_maxopd length. Clear FIN to prevent premature 684 * closure since there is still more data to send after this (now 685 * truncated) packet. 686 * 687 * If just the options do not fit we are in a no-win situation and 688 * we treat it as an unreachable host. 689 */ 690 if (len + optlen + ipoptlen > tp->t_maxopd) { 691 if (tp->t_maxopd <= optlen + ipoptlen) { 692 static time_t last_optlen_report; 693 694 if (last_optlen_report != time_second) { 695 last_optlen_report = time_second; 696 kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd); 697 } 698 error = EHOSTUNREACH; 699 goto out; 700 } else { 701 flags &= ~TH_FIN; 702 len = tp->t_maxopd - optlen - ipoptlen; 703 sendalot = TRUE; 704 } 705 } 706 707 #ifdef INET6 708 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big")); 709 #else 710 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big")); 711 #endif 712 713 /* 714 * Grab a header mbuf, attaching a copy of data to 715 * be transmitted, and initialize the header from 716 * the template for sends on this connection. 717 */ 718 if (len) { 719 if ((tp->t_flags & TF_FORCE) && len == 1) 720 tcpstat.tcps_sndprobe++; 721 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 722 if (tp->snd_nxt == tp->snd_una) 723 tp->snd_max_rexmt = tp->snd_max; 724 tcpstat.tcps_sndrexmitpack++; 725 tcpstat.tcps_sndrexmitbyte += len; 726 } else { 727 tcpstat.tcps_sndpack++; 728 tcpstat.tcps_sndbyte += len; 729 } 730 #ifdef notyet 731 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len, 732 max_linkhdr + hdrlen)) == NULL) { 733 error = ENOBUFS; 734 goto out; 735 } 736 /* 737 * m_copypack left space for our hdr; use it. 738 */ 739 m->m_len += hdrlen; 740 m->m_data -= hdrlen; 741 #else 742 #ifndef INET6 743 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 744 #else 745 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER, 746 M_PKTHDR, NULL); 747 #endif 748 if (m == NULL) { 749 error = ENOBUFS; 750 goto out; 751 } 752 m->m_data += max_linkhdr; 753 m->m_len = hdrlen; 754 if (len <= MHLEN - hdrlen - max_linkhdr) { 755 m_copydata(so->so_snd.ssb_mb, off, (int) len, 756 mtod(m, caddr_t) + hdrlen); 757 m->m_len += len; 758 } else { 759 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len); 760 if (m->m_next == NULL) { 761 m_free(m); 762 error = ENOBUFS; 763 goto out; 764 } 765 } 766 #endif 767 /* 768 * If we're sending everything we've got, set PUSH. 769 * (This will keep happy those implementations which only 770 * give data to the user when a buffer fills or 771 * a PUSH comes in.) 772 */ 773 if (off + len == so->so_snd.ssb_cc) 774 flags |= TH_PUSH; 775 } else { 776 if (tp->t_flags & TF_ACKNOW) 777 tcpstat.tcps_sndacks++; 778 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 779 tcpstat.tcps_sndctrl++; 780 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 781 tcpstat.tcps_sndurg++; 782 else 783 tcpstat.tcps_sndwinup++; 784 785 MGETHDR(m, MB_DONTWAIT, MT_HEADER); 786 if (m == NULL) { 787 error = ENOBUFS; 788 goto out; 789 } 790 if (isipv6 && 791 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN) 792 MH_ALIGN(m, hdrlen); 793 else 794 m->m_data += max_linkhdr; 795 m->m_len = hdrlen; 796 } 797 m->m_pkthdr.rcvif = NULL; 798 if (isipv6) { 799 ip6 = mtod(m, struct ip6_hdr *); 800 th = (struct tcphdr *)(ip6 + 1); 801 tcp_fillheaders(tp, ip6, th); 802 } else { 803 ip = mtod(m, struct ip *); 804 ipov = (struct ipovly *)ip; 805 th = (struct tcphdr *)(ip + 1); 806 /* this picks up the pseudo header (w/o the length) */ 807 tcp_fillheaders(tp, ip, th); 808 } 809 810 /* 811 * Fill in fields, remembering maximum advertised 812 * window for use in delaying messages about window sizes. 813 * If resending a FIN, be sure not to use a new sequence number. 814 */ 815 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 816 tp->snd_nxt == tp->snd_max) 817 tp->snd_nxt--; 818 /* 819 * If we are doing retransmissions, then snd_nxt will 820 * not reflect the first unsent octet. For ACK only 821 * packets, we do not want the sequence number of the 822 * retransmitted packet, we want the sequence number 823 * of the next unsent octet. So, if there is no data 824 * (and no SYN or FIN), use snd_max instead of snd_nxt 825 * when filling in ti_seq. But if we are in persist 826 * state, snd_max might reflect one byte beyond the 827 * right edge of the window, so use snd_nxt in that 828 * case, since we know we aren't doing a retransmission. 829 * (retransmit and persist are mutually exclusive...) 830 */ 831 if (len || (flags & (TH_SYN|TH_FIN)) || 832 tcp_callout_active(tp, tp->tt_persist)) 833 th->th_seq = htonl(tp->snd_nxt); 834 else 835 th->th_seq = htonl(tp->snd_max); 836 th->th_ack = htonl(tp->rcv_nxt); 837 if (optlen) { 838 bcopy(opt, th + 1, optlen); 839 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 840 } 841 th->th_flags = flags; 842 /* 843 * Calculate receive window. Don't shrink window, 844 * but avoid silly window syndrome. 845 */ 846 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) && 847 recvwin < (long)tp->t_maxseg) 848 recvwin = 0; 849 if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 850 recvwin = (long)(tp->rcv_adv - tp->rcv_nxt); 851 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale) 852 recvwin = (long)TCP_MAXWIN << tp->rcv_scale; 853 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale)); 854 855 /* 856 * Adjust the RXWIN0SENT flag - indicate that we have advertised 857 * a 0 window. This may cause the remote transmitter to stall. This 858 * flag tells soreceive() to disable delayed acknowledgements when 859 * draining the buffer. This can occur if the receiver is attempting 860 * to read more data then can be buffered prior to transmitting on 861 * the connection. 862 */ 863 if (recvwin == 0) 864 tp->t_flags |= TF_RXWIN0SENT; 865 else 866 tp->t_flags &= ~TF_RXWIN0SENT; 867 868 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 869 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 870 th->th_flags |= TH_URG; 871 } else { 872 /* 873 * If no urgent pointer to send, then we pull 874 * the urgent pointer to the left edge of the send window 875 * so that it doesn't drift into the send window on sequence 876 * number wraparound. 877 */ 878 tp->snd_up = tp->snd_una; /* drag it along */ 879 } 880 881 /* 882 * Put TCP length in extended header, and then 883 * checksum extended header and data. 884 */ 885 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 886 if (isipv6) { 887 /* 888 * ip6_plen is not need to be filled now, and will be filled 889 * in ip6_output(). 890 */ 891 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 892 sizeof(struct tcphdr) + optlen + len); 893 } else { 894 m->m_pkthdr.csum_flags = CSUM_TCP; 895 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 896 if (len + optlen) 897 th->th_sum = in_addword(th->th_sum, 898 htons((u_short)(optlen + len))); 899 900 /* IP version must be set here for ipv4/ipv6 checking later */ 901 KASSERT(ip->ip_v == IPVERSION, 902 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 903 } 904 905 /* 906 * In transmit state, time the transmission and arrange for 907 * the retransmit. In persist state, just set snd_max. 908 */ 909 if (!(tp->t_flags & TF_FORCE) || 910 !tcp_callout_active(tp, tp->tt_persist)) { 911 tcp_seq startseq = tp->snd_nxt; 912 913 /* 914 * Advance snd_nxt over sequence space of this segment. 915 */ 916 if (flags & (TH_SYN | TH_FIN)) { 917 if (flags & TH_SYN) 918 tp->snd_nxt++; 919 if (flags & TH_FIN) { 920 tp->snd_nxt++; 921 tp->t_flags |= TF_SENTFIN; 922 } 923 } 924 tp->snd_nxt += len; 925 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 926 tp->snd_max = tp->snd_nxt; 927 /* 928 * Time this transmission if not a retransmission and 929 * not currently timing anything. 930 */ 931 if (tp->t_rtttime == 0) { 932 tp->t_rtttime = ticks; 933 tp->t_rtseq = startseq; 934 tcpstat.tcps_segstimed++; 935 } 936 } 937 938 /* 939 * Set retransmit timer if not currently set, 940 * and not doing a pure ack or a keep-alive probe. 941 * Initial value for retransmit timer is smoothed 942 * round-trip time + 2 * round-trip time variance. 943 * Initialize shift counter which is used for backoff 944 * of retransmit time. 945 */ 946 if (!tcp_callout_active(tp, tp->tt_rexmt) && 947 tp->snd_nxt != tp->snd_una) { 948 if (tcp_callout_active(tp, tp->tt_persist)) { 949 tcp_callout_stop(tp, tp->tt_persist); 950 tp->t_rxtshift = 0; 951 } 952 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur, 953 tcp_timer_rexmt); 954 } 955 } else { 956 /* 957 * Persist case, update snd_max but since we are in 958 * persist mode (no window) we do not update snd_nxt. 959 */ 960 int xlen = len; 961 if (flags & TH_SYN) 962 ++xlen; 963 if (flags & TH_FIN) { 964 ++xlen; 965 tp->t_flags |= TF_SENTFIN; 966 } 967 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 968 tp->snd_max = tp->snd_nxt + xlen; 969 } 970 971 #ifdef TCPDEBUG 972 /* 973 * Trace. 974 */ 975 if (so->so_options & SO_DEBUG) 976 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 977 #endif 978 979 /* 980 * Fill in IP length and desired time to live and 981 * send to IP level. There should be a better way 982 * to handle ttl and tos; we could keep them in 983 * the template, but need a way to checksum without them. 984 */ 985 /* 986 * m->m_pkthdr.len should have been set before cksum calcuration, 987 * because in6_cksum() need it. 988 */ 989 if (isipv6) { 990 /* 991 * we separately set hoplimit for every segment, since the 992 * user might want to change the value via setsockopt. 993 * Also, desired default hop limit might be changed via 994 * Neighbor Discovery. 995 */ 996 ip6->ip6_hlim = in6_selecthlim(inp, 997 (inp->in6p_route.ro_rt ? 998 inp->in6p_route.ro_rt->rt_ifp : NULL)); 999 1000 /* TODO: IPv6 IP6TOS_ECT bit on */ 1001 error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route, 1002 (so->so_options & SO_DONTROUTE), NULL, NULL, 1003 inp); 1004 } else { 1005 struct rtentry *rt; 1006 ip->ip_len = m->m_pkthdr.len; 1007 #ifdef INET6 1008 if (INP_CHECK_SOCKAF(so, AF_INET6)) 1009 ip->ip_ttl = in6_selecthlim(inp, 1010 (inp->in6p_route.ro_rt ? 1011 inp->in6p_route.ro_rt->rt_ifp : NULL)); 1012 else 1013 #endif 1014 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */ 1015 1016 ip->ip_tos = inp->inp_ip_tos; /* XXX */ 1017 /* 1018 * See if we should do MTU discovery. 1019 * We do it only if the following are true: 1020 * 1) we have a valid route to the destination 1021 * 2) the MTU is not locked (if it is, 1022 * then discovery has been disabled) 1023 */ 1024 if (path_mtu_discovery && 1025 (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) && 1026 !(rt->rt_rmx.rmx_locks & RTV_MTU)) 1027 ip->ip_off |= IP_DF; 1028 1029 error = ip_output(m, inp->inp_options, &inp->inp_route, 1030 (so->so_options & SO_DONTROUTE) | 1031 IP_DEBUGROUTE, NULL, inp); 1032 } 1033 if (error) { 1034 1035 /* 1036 * We know that the packet was lost, so back out the 1037 * sequence number advance, if any. 1038 */ 1039 if (!(tp->t_flags & TF_FORCE) || 1040 !tcp_callout_active(tp, tp->tt_persist)) { 1041 /* 1042 * No need to check for TH_FIN here because 1043 * the TF_SENTFIN flag handles that case. 1044 */ 1045 if (!(flags & TH_SYN)) 1046 tp->snd_nxt -= len; 1047 } 1048 1049 out: 1050 if (error == ENOBUFS) { 1051 /* 1052 * If we can't send, make sure there is something 1053 * to get us going again later. Persist state 1054 * is not necessarily right, but it is close enough. 1055 */ 1056 if (!tcp_callout_active(tp, tp->tt_rexmt) && 1057 !tcp_callout_active(tp, tp->tt_persist)) { 1058 tp->t_rxtshift = 0; 1059 tcp_setpersist(tp); 1060 } 1061 tcp_quench(inp, 0); 1062 return (0); 1063 } 1064 if (error == EMSGSIZE) { 1065 /* 1066 * ip_output() will have already fixed the route 1067 * for us. tcp_mtudisc() will, as its last action, 1068 * initiate retransmission, so it is important to 1069 * not do so here. 1070 */ 1071 tcp_mtudisc(inp, 0); 1072 return 0; 1073 } 1074 if ((error == EHOSTUNREACH || error == ENETDOWN) && 1075 TCPS_HAVERCVDSYN(tp->t_state)) { 1076 tp->t_softerror = error; 1077 return (0); 1078 } 1079 return (error); 1080 } 1081 tcpstat.tcps_sndtotal++; 1082 1083 /* 1084 * Data sent (as far as we can tell). 1085 * If this advertises a larger window than any other segment, 1086 * then remember the size of the advertised window. 1087 * Any pending ACK has now been sent. 1088 */ 1089 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) 1090 tp->rcv_adv = tp->rcv_nxt + recvwin; 1091 tp->last_ack_sent = tp->rcv_nxt; 1092 tp->t_flags &= ~TF_ACKNOW; 1093 if (tcp_delack_enabled) 1094 tcp_callout_stop(tp, tp->tt_delack); 1095 if (sendalot) 1096 goto again; 1097 return (0); 1098 } 1099 1100 void 1101 tcp_setpersist(struct tcpcb *tp) 1102 { 1103 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1104 int tt; 1105 1106 if (tcp_callout_active(tp, tp->tt_rexmt)) 1107 panic("tcp_setpersist: retransmit pending"); 1108 /* 1109 * Start/restart persistance timer. 1110 */ 1111 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN, 1112 TCPTV_PERSMAX); 1113 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist); 1114 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1115 tp->t_rxtshift++; 1116 } 1117