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 = 2*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 u_long newsize; 372 373 newsize = ulmin(so->so_snd.ssb_hiwat + 374 tcp_autosndbuf_inc, 375 tcp_autosndbuf_max); 376 if (!ssb_reserve(&so->so_snd, newsize, so, NULL)) 377 so->so_snd.ssb_flags &= ~SSB_AUTOSIZE; 378 if (newsize >= (TCP_MAXWIN << tp->snd_scale)) 379 so->so_snd.ssb_flags &= ~SSB_AUTOSIZE; 380 } 381 } 382 383 /* 384 * Truncate to the maximum segment length and ensure that FIN is 385 * removed if the length no longer contains the last data byte. 386 */ 387 if (len > tp->t_maxseg) { 388 len = tp->t_maxseg; 389 sendalot = TRUE; 390 } 391 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc)) 392 flags &= ~TH_FIN; 393 394 recvwin = ssb_space(&so->so_rcv); 395 396 /* 397 * Sender silly window avoidance. We transmit under the following 398 * conditions when len is non-zero: 399 * 400 * - We have a full segment 401 * - This is the last buffer in a write()/send() and we are 402 * either idle or running NODELAY 403 * - we've timed out (e.g. persist timer) 404 * - we have more then 1/2 the maximum send window's worth of 405 * data (receiver may be limiting the window size) 406 * - we need to retransmit 407 */ 408 if (len) { 409 if (len == tp->t_maxseg) 410 goto send; 411 /* 412 * NOTE! on localhost connections an 'ack' from the remote 413 * end may occur synchronously with the output and cause 414 * us to flush a buffer queued with moretocome. XXX 415 * 416 * note: the len + off check is almost certainly unnecessary. 417 */ 418 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 419 (idle || (tp->t_flags & TF_NODELAY)) && 420 len + off >= so->so_snd.ssb_cc && 421 !(tp->t_flags & TF_NOPUSH)) { 422 goto send; 423 } 424 if (tp->t_flags & TF_FORCE) /* typ. timeout case */ 425 goto send; 426 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 427 goto send; 428 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 429 goto send; 430 } 431 432 /* 433 * Compare available window to amount of window 434 * known to peer (as advertised window less 435 * next expected input). If the difference is at least two 436 * max size segments, or at least 50% of the maximum possible 437 * window, then want to send a window update to peer. 438 */ 439 if (recvwin > 0) { 440 /* 441 * "adv" is the amount we can increase the window, 442 * taking into account that we are limited by 443 * TCP_MAXWIN << tp->rcv_scale. 444 */ 445 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) - 446 (tp->rcv_adv - tp->rcv_nxt); 447 long hiwat; 448 449 /* 450 * This ack case typically occurs when the user has drained 451 * the TCP socket buffer sufficiently to warrent an ack 452 * containing a 'pure window update'... that is, an ack that 453 * ONLY updates the tcp window. 454 * 455 * It is unclear why we would need to do a pure window update 456 * past 2 segments if we are going to do one at 1/2 the high 457 * water mark anyway, especially since under normal conditions 458 * the user program will drain the socket buffer quickly. 459 * The 2-segment pure window update will often add a large 460 * number of extra, unnecessary acks to the stream. 461 * 462 * avoid_pure_win_update now defaults to 1. 463 */ 464 if (avoid_pure_win_update == 0 || 465 (tp->t_flags & TF_RXRESIZED)) { 466 if (adv >= (long) (2 * tp->t_maxseg)) { 467 goto send; 468 } 469 } 470 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale); 471 if (hiwat > (long)so->so_rcv.ssb_hiwat) 472 hiwat = (long)so->so_rcv.ssb_hiwat; 473 if (adv >= hiwat / 2) 474 goto send; 475 } 476 477 /* 478 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 479 * is also a catch-all for the retransmit timer timeout case. 480 */ 481 if (tp->t_flags & TF_ACKNOW) 482 goto send; 483 if ((flags & TH_RST) || 484 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN))) 485 goto send; 486 if (SEQ_GT(tp->snd_up, tp->snd_una)) 487 goto send; 488 /* 489 * If our state indicates that FIN should be sent 490 * and we have not yet done so, then we need to send. 491 */ 492 if (flags & TH_FIN && 493 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una)) 494 goto send; 495 496 /* 497 * TCP window updates are not reliable, rather a polling protocol 498 * using ``persist'' packets is used to insure receipt of window 499 * updates. The three ``states'' for the output side are: 500 * idle not doing retransmits or persists 501 * persisting to move a small or zero window 502 * (re)transmitting and thereby not persisting 503 * 504 * tcp_callout_active(tp, tp->tt_persist) 505 * is true when we are in persist state. 506 * The TF_FORCE flag in tp->t_flags 507 * is set when we are called to send a persist packet. 508 * tcp_callout_active(tp, tp->tt_rexmt) 509 * is set when we are retransmitting 510 * The output side is idle when both timers are zero. 511 * 512 * If send window is too small, there is data to transmit, and no 513 * retransmit or persist is pending, then go to persist state. 514 * If nothing happens soon, send when timer expires: 515 * if window is nonzero, transmit what we can, 516 * otherwise force out a byte. 517 */ 518 if (so->so_snd.ssb_cc > 0 && 519 !tcp_callout_active(tp, tp->tt_rexmt) && 520 !tcp_callout_active(tp, tp->tt_persist)) { 521 tp->t_rxtshift = 0; 522 tcp_setpersist(tp); 523 } 524 525 /* 526 * No reason to send a segment, just return. 527 */ 528 return (0); 529 530 send: 531 /* 532 * Before ESTABLISHED, force sending of initial options 533 * unless TCP set not to do any options. 534 * NOTE: we assume that the IP/TCP header plus TCP options 535 * always fit in a single mbuf, leaving room for a maximum 536 * link header, i.e. 537 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES 538 */ 539 optlen = 0; 540 if (isipv6) 541 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 542 else 543 hdrlen = sizeof(struct tcpiphdr); 544 if (flags & TH_SYN) { 545 tp->snd_nxt = tp->iss; 546 if (!(tp->t_flags & TF_NOOPT)) { 547 u_short mss; 548 549 opt[0] = TCPOPT_MAXSEG; 550 opt[1] = TCPOLEN_MAXSEG; 551 mss = htons((u_short) tcp_mssopt(tp)); 552 memcpy(opt + 2, &mss, sizeof mss); 553 optlen = TCPOLEN_MAXSEG; 554 555 if ((tp->t_flags & TF_REQ_SCALE) && 556 (!(flags & TH_ACK) || 557 (tp->t_flags & TF_RCVD_SCALE))) { 558 *((u_int32_t *)(opt + optlen)) = htonl( 559 TCPOPT_NOP << 24 | 560 TCPOPT_WINDOW << 16 | 561 TCPOLEN_WINDOW << 8 | 562 tp->request_r_scale); 563 optlen += 4; 564 } 565 566 if ((tcp_do_sack && !(flags & TH_ACK)) || 567 tp->t_flags & TF_SACK_PERMITTED) { 568 uint32_t *lp = (uint32_t *)(opt + optlen); 569 570 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED); 571 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED; 572 } 573 } 574 } 575 576 /* 577 * Send a timestamp and echo-reply if this is a SYN and our side 578 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 579 * and our peer have sent timestamps in our SYN's. 580 */ 581 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 582 !(flags & TH_RST) && 583 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) { 584 u_int32_t *lp = (u_int32_t *)(opt + optlen); 585 586 /* Form timestamp option as shown in appendix A of RFC 1323. */ 587 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 588 *lp++ = htonl(ticks); 589 *lp = htonl(tp->ts_recent); 590 optlen += TCPOLEN_TSTAMP_APPA; 591 } 592 593 /* Set receive buffer autosizing timestamp. */ 594 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) 595 tp->rfbuf_ts = ticks; 596 597 /* 598 * Send `CC-family' options if our side wants to use them (TF_REQ_CC), 599 * options are allowed (!TF_NOOPT) and it's not a RST. 600 */ 601 if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC && 602 !(flags & TH_RST)) { 603 switch (flags & (TH_SYN | TH_ACK)) { 604 /* 605 * This is a normal ACK, send CC if we received CC before 606 * from our peer. 607 */ 608 case TH_ACK: 609 if (!(tp->t_flags & TF_RCVD_CC)) 610 break; 611 /*FALLTHROUGH*/ 612 613 /* 614 * We can only get here in T/TCP's SYN_SENT* state, when 615 * we're a sending a non-SYN segment without waiting for 616 * the ACK of our SYN. A check above assures that we only 617 * do this if our peer understands T/TCP. 618 */ 619 case 0: 620 opt[optlen++] = TCPOPT_NOP; 621 opt[optlen++] = TCPOPT_NOP; 622 opt[optlen++] = TCPOPT_CC; 623 opt[optlen++] = TCPOLEN_CC; 624 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 625 optlen += 4; 626 break; 627 628 /* 629 * This is our initial SYN, check whether we have to use 630 * CC or CC.new. 631 */ 632 case TH_SYN: 633 opt[optlen++] = TCPOPT_NOP; 634 opt[optlen++] = TCPOPT_NOP; 635 opt[optlen++] = tp->t_flags & TF_SENDCCNEW ? 636 TCPOPT_CCNEW : TCPOPT_CC; 637 opt[optlen++] = TCPOLEN_CC; 638 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 639 optlen += 4; 640 break; 641 642 /* 643 * This is a SYN,ACK; send CC and CC.echo if we received 644 * CC from our peer. 645 */ 646 case (TH_SYN | TH_ACK): 647 if (tp->t_flags & TF_RCVD_CC) { 648 opt[optlen++] = TCPOPT_NOP; 649 opt[optlen++] = TCPOPT_NOP; 650 opt[optlen++] = TCPOPT_CC; 651 opt[optlen++] = TCPOLEN_CC; 652 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 653 optlen += 4; 654 opt[optlen++] = TCPOPT_NOP; 655 opt[optlen++] = TCPOPT_NOP; 656 opt[optlen++] = TCPOPT_CCECHO; 657 opt[optlen++] = TCPOLEN_CC; 658 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_recv); 659 optlen += 4; 660 } 661 break; 662 } 663 } 664 665 /* 666 * If this is a SACK connection and we have a block to report, 667 * fill in the SACK blocks in the TCP options. 668 */ 669 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) == 670 TF_SACK_PERMITTED && 671 (!LIST_EMPTY(&tp->t_segq) || 672 tp->reportblk.rblk_start != tp->reportblk.rblk_end)) 673 tcp_sack_fill_report(tp, opt, &optlen); 674 675 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options")); 676 hdrlen += optlen; 677 678 if (isipv6) { 679 ipoptlen = ip6_optlen(inp); 680 } else { 681 if (inp->inp_options) { 682 ipoptlen = inp->inp_options->m_len - 683 offsetof(struct ipoption, ipopt_list); 684 } else { 685 ipoptlen = 0; 686 } 687 } 688 #ifdef IPSEC 689 ipoptlen += ipsec_hdrsiz_tcp(tp); 690 #endif 691 692 /* 693 * Adjust data length if insertion of options will bump the packet 694 * length beyond the t_maxopd length. Clear FIN to prevent premature 695 * closure since there is still more data to send after this (now 696 * truncated) packet. 697 * 698 * If just the options do not fit we are in a no-win situation and 699 * we treat it as an unreachable host. 700 */ 701 if (len + optlen + ipoptlen > tp->t_maxopd) { 702 if (tp->t_maxopd <= optlen + ipoptlen) { 703 static time_t last_optlen_report; 704 705 if (last_optlen_report != time_second) { 706 last_optlen_report = time_second; 707 kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd); 708 } 709 error = EHOSTUNREACH; 710 goto out; 711 } else { 712 flags &= ~TH_FIN; 713 len = tp->t_maxopd - optlen - ipoptlen; 714 sendalot = TRUE; 715 } 716 } 717 718 #ifdef INET6 719 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big")); 720 #else 721 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big")); 722 #endif 723 724 /* 725 * Grab a header mbuf, attaching a copy of data to 726 * be transmitted, and initialize the header from 727 * the template for sends on this connection. 728 */ 729 if (len) { 730 if ((tp->t_flags & TF_FORCE) && len == 1) 731 tcpstat.tcps_sndprobe++; 732 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 733 if (tp->snd_nxt == tp->snd_una) 734 tp->snd_max_rexmt = tp->snd_max; 735 tcpstat.tcps_sndrexmitpack++; 736 tcpstat.tcps_sndrexmitbyte += len; 737 } else { 738 tcpstat.tcps_sndpack++; 739 tcpstat.tcps_sndbyte += len; 740 } 741 #ifdef notyet 742 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len, 743 max_linkhdr + hdrlen)) == NULL) { 744 error = ENOBUFS; 745 goto out; 746 } 747 /* 748 * m_copypack left space for our hdr; use it. 749 */ 750 m->m_len += hdrlen; 751 m->m_data -= hdrlen; 752 #else 753 #ifndef INET6 754 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 755 #else 756 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER, 757 M_PKTHDR, NULL); 758 #endif 759 if (m == NULL) { 760 error = ENOBUFS; 761 goto out; 762 } 763 m->m_data += max_linkhdr; 764 m->m_len = hdrlen; 765 if (len <= MHLEN - hdrlen - max_linkhdr) { 766 m_copydata(so->so_snd.ssb_mb, off, (int) len, 767 mtod(m, caddr_t) + hdrlen); 768 m->m_len += len; 769 } else { 770 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len); 771 if (m->m_next == NULL) { 772 m_free(m); 773 error = ENOBUFS; 774 goto out; 775 } 776 } 777 #endif 778 /* 779 * If we're sending everything we've got, set PUSH. 780 * (This will keep happy those implementations which only 781 * give data to the user when a buffer fills or 782 * a PUSH comes in.) 783 */ 784 if (off + len == so->so_snd.ssb_cc) 785 flags |= TH_PUSH; 786 } else { 787 if (tp->t_flags & TF_ACKNOW) 788 tcpstat.tcps_sndacks++; 789 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 790 tcpstat.tcps_sndctrl++; 791 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 792 tcpstat.tcps_sndurg++; 793 else 794 tcpstat.tcps_sndwinup++; 795 796 MGETHDR(m, MB_DONTWAIT, MT_HEADER); 797 if (m == NULL) { 798 error = ENOBUFS; 799 goto out; 800 } 801 if (isipv6 && 802 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN) 803 MH_ALIGN(m, hdrlen); 804 else 805 m->m_data += max_linkhdr; 806 m->m_len = hdrlen; 807 } 808 m->m_pkthdr.rcvif = NULL; 809 if (isipv6) { 810 ip6 = mtod(m, struct ip6_hdr *); 811 th = (struct tcphdr *)(ip6 + 1); 812 tcp_fillheaders(tp, ip6, th); 813 } else { 814 ip = mtod(m, struct ip *); 815 ipov = (struct ipovly *)ip; 816 th = (struct tcphdr *)(ip + 1); 817 /* this picks up the pseudo header (w/o the length) */ 818 tcp_fillheaders(tp, ip, th); 819 } 820 821 /* 822 * Fill in fields, remembering maximum advertised 823 * window for use in delaying messages about window sizes. 824 * If resending a FIN, be sure not to use a new sequence number. 825 */ 826 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 827 tp->snd_nxt == tp->snd_max) 828 tp->snd_nxt--; 829 /* 830 * If we are doing retransmissions, then snd_nxt will 831 * not reflect the first unsent octet. For ACK only 832 * packets, we do not want the sequence number of the 833 * retransmitted packet, we want the sequence number 834 * of the next unsent octet. So, if there is no data 835 * (and no SYN or FIN), use snd_max instead of snd_nxt 836 * when filling in ti_seq. But if we are in persist 837 * state, snd_max might reflect one byte beyond the 838 * right edge of the window, so use snd_nxt in that 839 * case, since we know we aren't doing a retransmission. 840 * (retransmit and persist are mutually exclusive...) 841 */ 842 if (len || (flags & (TH_SYN|TH_FIN)) || 843 tcp_callout_active(tp, tp->tt_persist)) 844 th->th_seq = htonl(tp->snd_nxt); 845 else 846 th->th_seq = htonl(tp->snd_max); 847 th->th_ack = htonl(tp->rcv_nxt); 848 if (optlen) { 849 bcopy(opt, th + 1, optlen); 850 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 851 } 852 th->th_flags = flags; 853 854 /* 855 * Calculate receive window. Don't shrink window, but avoid 856 * silly window syndrome by sending a 0 window if the actual 857 * window is less then one segment. 858 */ 859 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) && 860 recvwin < (long)tp->t_maxseg) 861 recvwin = 0; 862 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt)) 863 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt); 864 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale) 865 recvwin = (long)TCP_MAXWIN << tp->rcv_scale; 866 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale)); 867 868 /* 869 * Adjust the RXWIN0SENT flag - indicate that we have advertised 870 * a 0 window. This may cause the remote transmitter to stall. This 871 * flag tells soreceive() to disable delayed acknowledgements when 872 * draining the buffer. This can occur if the receiver is attempting 873 * to read more data then can be buffered prior to transmitting on 874 * the connection. 875 */ 876 if (recvwin == 0) 877 tp->t_flags |= TF_RXWIN0SENT; 878 else 879 tp->t_flags &= ~TF_RXWIN0SENT; 880 881 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 882 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 883 th->th_flags |= TH_URG; 884 } else { 885 /* 886 * If no urgent pointer to send, then we pull 887 * the urgent pointer to the left edge of the send window 888 * so that it doesn't drift into the send window on sequence 889 * number wraparound. 890 */ 891 tp->snd_up = tp->snd_una; /* drag it along */ 892 } 893 894 /* 895 * Put TCP length in extended header, and then 896 * checksum extended header and data. 897 */ 898 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 899 if (isipv6) { 900 /* 901 * ip6_plen is not need to be filled now, and will be filled 902 * in ip6_output(). 903 */ 904 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 905 sizeof(struct tcphdr) + optlen + len); 906 } else { 907 m->m_pkthdr.csum_flags = CSUM_TCP; 908 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 909 if (len + optlen) 910 th->th_sum = in_addword(th->th_sum, 911 htons((u_short)(optlen + len))); 912 913 /* IP version must be set here for ipv4/ipv6 checking later */ 914 KASSERT(ip->ip_v == IPVERSION, 915 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 916 } 917 918 /* 919 * In transmit state, time the transmission and arrange for 920 * the retransmit. In persist state, just set snd_max. 921 */ 922 if (!(tp->t_flags & TF_FORCE) || 923 !tcp_callout_active(tp, tp->tt_persist)) { 924 tcp_seq startseq = tp->snd_nxt; 925 926 /* 927 * Advance snd_nxt over sequence space of this segment. 928 */ 929 if (flags & (TH_SYN | TH_FIN)) { 930 if (flags & TH_SYN) 931 tp->snd_nxt++; 932 if (flags & TH_FIN) { 933 tp->snd_nxt++; 934 tp->t_flags |= TF_SENTFIN; 935 } 936 } 937 tp->snd_nxt += len; 938 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 939 tp->snd_max = tp->snd_nxt; 940 /* 941 * Time this transmission if not a retransmission and 942 * not currently timing anything. 943 */ 944 if (tp->t_rtttime == 0) { 945 tp->t_rtttime = ticks; 946 tp->t_rtseq = startseq; 947 tcpstat.tcps_segstimed++; 948 } 949 } 950 951 /* 952 * Set retransmit timer if not currently set, 953 * and not doing a pure ack or a keep-alive probe. 954 * Initial value for retransmit timer is smoothed 955 * round-trip time + 2 * round-trip time variance. 956 * Initialize shift counter which is used for backoff 957 * of retransmit time. 958 */ 959 if (!tcp_callout_active(tp, tp->tt_rexmt) && 960 tp->snd_nxt != tp->snd_una) { 961 if (tcp_callout_active(tp, tp->tt_persist)) { 962 tcp_callout_stop(tp, tp->tt_persist); 963 tp->t_rxtshift = 0; 964 } 965 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur, 966 tcp_timer_rexmt); 967 } 968 } else { 969 /* 970 * Persist case, update snd_max but since we are in 971 * persist mode (no window) we do not update snd_nxt. 972 */ 973 int xlen = len; 974 if (flags & TH_SYN) 975 ++xlen; 976 if (flags & TH_FIN) { 977 ++xlen; 978 tp->t_flags |= TF_SENTFIN; 979 } 980 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 981 tp->snd_max = tp->snd_nxt + xlen; 982 } 983 984 #ifdef TCPDEBUG 985 /* 986 * Trace. 987 */ 988 if (so->so_options & SO_DEBUG) 989 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 990 #endif 991 992 /* 993 * Fill in IP length and desired time to live and 994 * send to IP level. There should be a better way 995 * to handle ttl and tos; we could keep them in 996 * the template, but need a way to checksum without them. 997 */ 998 /* 999 * m->m_pkthdr.len should have been set before cksum calcuration, 1000 * because in6_cksum() need it. 1001 */ 1002 if (isipv6) { 1003 /* 1004 * we separately set hoplimit for every segment, since the 1005 * user might want to change the value via setsockopt. 1006 * Also, desired default hop limit might be changed via 1007 * Neighbor Discovery. 1008 */ 1009 ip6->ip6_hlim = in6_selecthlim(inp, 1010 (inp->in6p_route.ro_rt ? 1011 inp->in6p_route.ro_rt->rt_ifp : NULL)); 1012 1013 /* TODO: IPv6 IP6TOS_ECT bit on */ 1014 error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route, 1015 (so->so_options & SO_DONTROUTE), NULL, NULL, 1016 inp); 1017 } else { 1018 struct rtentry *rt; 1019 ip->ip_len = m->m_pkthdr.len; 1020 #ifdef INET6 1021 if (INP_CHECK_SOCKAF(so, AF_INET6)) 1022 ip->ip_ttl = in6_selecthlim(inp, 1023 (inp->in6p_route.ro_rt ? 1024 inp->in6p_route.ro_rt->rt_ifp : NULL)); 1025 else 1026 #endif 1027 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */ 1028 1029 ip->ip_tos = inp->inp_ip_tos; /* XXX */ 1030 /* 1031 * See if we should do MTU discovery. 1032 * We do it only if the following are true: 1033 * 1) we have a valid route to the destination 1034 * 2) the MTU is not locked (if it is, 1035 * then discovery has been disabled) 1036 */ 1037 if (path_mtu_discovery && 1038 (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) && 1039 !(rt->rt_rmx.rmx_locks & RTV_MTU)) 1040 ip->ip_off |= IP_DF; 1041 1042 error = ip_output(m, inp->inp_options, &inp->inp_route, 1043 (so->so_options & SO_DONTROUTE) | 1044 IP_DEBUGROUTE, NULL, inp); 1045 } 1046 if (error) { 1047 1048 /* 1049 * We know that the packet was lost, so back out the 1050 * sequence number advance, if any. 1051 */ 1052 if (!(tp->t_flags & TF_FORCE) || 1053 !tcp_callout_active(tp, tp->tt_persist)) { 1054 /* 1055 * No need to check for TH_FIN here because 1056 * the TF_SENTFIN flag handles that case. 1057 */ 1058 if (!(flags & TH_SYN)) 1059 tp->snd_nxt -= len; 1060 } 1061 1062 out: 1063 if (error == ENOBUFS) { 1064 /* 1065 * If we can't send, make sure there is something 1066 * to get us going again later. Persist state 1067 * is not necessarily right, but it is close enough. 1068 */ 1069 if (!tcp_callout_active(tp, tp->tt_rexmt) && 1070 !tcp_callout_active(tp, tp->tt_persist)) { 1071 tp->t_rxtshift = 0; 1072 tcp_setpersist(tp); 1073 } 1074 tcp_quench(inp, 0); 1075 return (0); 1076 } 1077 if (error == EMSGSIZE) { 1078 /* 1079 * ip_output() will have already fixed the route 1080 * for us. tcp_mtudisc() will, as its last action, 1081 * initiate retransmission, so it is important to 1082 * not do so here. 1083 */ 1084 tcp_mtudisc(inp, 0); 1085 return 0; 1086 } 1087 if ((error == EHOSTUNREACH || error == ENETDOWN) && 1088 TCPS_HAVERCVDSYN(tp->t_state)) { 1089 tp->t_softerror = error; 1090 return (0); 1091 } 1092 return (error); 1093 } 1094 tcpstat.tcps_sndtotal++; 1095 1096 /* 1097 * Data sent (as far as we can tell). 1098 * 1099 * If this advertises a larger window than any other segment, 1100 * then remember the size of the advertised window. 1101 * 1102 * Any pending ACK has now been sent. 1103 */ 1104 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) { 1105 tp->rcv_adv = tp->rcv_nxt + recvwin; 1106 tp->t_flags &= ~TF_RXRESIZED; 1107 } 1108 tp->last_ack_sent = tp->rcv_nxt; 1109 tp->t_flags &= ~TF_ACKNOW; 1110 if (tcp_delack_enabled) 1111 tcp_callout_stop(tp, tp->tt_delack); 1112 if (sendalot) 1113 goto again; 1114 return (0); 1115 } 1116 1117 void 1118 tcp_setpersist(struct tcpcb *tp) 1119 { 1120 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1121 int tt; 1122 1123 if (tcp_callout_active(tp, tp->tt_rexmt)) 1124 panic("tcp_setpersist: retransmit pending"); 1125 /* 1126 * Start/restart persistance timer. 1127 */ 1128 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN, 1129 TCPTV_PERSMAX); 1130 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist); 1131 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1132 tp->t_rxtshift++; 1133 } 1134