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_inet.h" 72 #include "opt_inet6.h" 73 #include "opt_ipsec.h" 74 #include "opt_tcpdebug.h" 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/kernel.h> 79 #include <sys/sysctl.h> 80 #include <sys/mbuf.h> 81 #include <sys/domain.h> 82 #include <sys/protosw.h> 83 #include <sys/socket.h> 84 #include <sys/socketvar.h> 85 #include <sys/in_cksum.h> 86 #include <sys/thread.h> 87 #include <sys/globaldata.h> 88 89 #include <net/route.h> 90 91 #include <netinet/in.h> 92 #include <netinet/in_systm.h> 93 #include <netinet/ip.h> 94 #include <netinet/in_pcb.h> 95 #include <netinet/ip_var.h> 96 #include <netinet6/in6_pcb.h> 97 #include <netinet/ip6.h> 98 #include <netinet6/ip6_var.h> 99 #include <netinet/tcp.h> 100 #define TCPOUTFLAGS 101 #include <netinet/tcp_fsm.h> 102 #include <netinet/tcp_seq.h> 103 #include <netinet/tcp_timer.h> 104 #include <netinet/tcp_timer2.h> 105 #include <netinet/tcp_var.h> 106 #include <netinet/tcpip.h> 107 #ifdef TCPDEBUG 108 #include <netinet/tcp_debug.h> 109 #endif 110 111 #ifdef IPSEC 112 #include <netinet6/ipsec.h> 113 #endif /*IPSEC*/ 114 115 #ifdef FAST_IPSEC 116 #include <netproto/ipsec/ipsec.h> 117 #define IPSEC 118 #endif /*FAST_IPSEC*/ 119 120 #ifdef notyet 121 extern struct mbuf *m_copypack(); 122 #endif 123 124 int path_mtu_discovery = 0; 125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 126 &path_mtu_discovery, 1, "Enable Path MTU Discovery"); 127 128 static int avoid_pure_win_update = 1; 129 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW, 130 &avoid_pure_win_update, 1, "Avoid pure window updates when possible"); 131 132 int tcp_do_autosndbuf = 1; 133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 134 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing"); 135 136 int tcp_autosndbuf_inc = 8*1024; 137 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 138 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer"); 139 140 int tcp_autosndbuf_max = 2*1024*1024; 141 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 142 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer"); 143 144 /* 145 * Tcp output routine: figure out what should be sent and send it. 146 */ 147 int 148 tcp_output(struct tcpcb *tp) 149 { 150 struct inpcb * const inp = tp->t_inpcb; 151 struct socket *so = inp->inp_socket; 152 long len, recvwin, sendwin; 153 int nsacked = 0; 154 int off, flags, error; 155 #ifdef TCP_SIGNATURE 156 int sigoff = 0; 157 #endif 158 struct mbuf *m; 159 struct ip *ip = NULL; 160 struct ipovly *ipov = NULL; 161 struct tcphdr *th; 162 u_char opt[TCP_MAXOLEN]; 163 unsigned int ipoptlen, optlen, hdrlen; 164 int idle; 165 boolean_t sendalot; 166 struct ip6_hdr *ip6 = NULL; 167 #ifdef INET6 168 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 169 #else 170 const boolean_t isipv6 = FALSE; 171 #endif 172 173 KKASSERT(so->so_port == &curthread->td_msgport); 174 175 /* 176 * Determine length of data that should be transmitted, 177 * and flags that will be used. 178 * If there is some data or critical controls (SYN, RST) 179 * to send, then transmit; otherwise, investigate further. 180 */ 181 182 /* 183 * If we have been idle for a while, the send congestion window 184 * could be no longer representative of the current state of the link. 185 * So unless we are expecting more acks to come in, slow-start from 186 * scratch to re-determine the send congestion window. 187 */ 188 if (tp->snd_max == tp->snd_una && 189 (ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 190 if (tcp_do_rfc3390) { 191 int initial_cwnd = 192 min(4 * tp->t_maxseg, max(2 * tp->t_maxseg, 4380)); 193 194 tp->snd_cwnd = min(tp->snd_cwnd, initial_cwnd); 195 } else { 196 tp->snd_cwnd = tp->t_maxseg; 197 } 198 tp->snd_wacked = 0; 199 } 200 201 /* 202 * Calculate whether the transmit stream was previously idle 203 * and adjust TF_LASTIDLE for the next time. 204 */ 205 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 206 if (idle && (tp->t_flags & TF_MORETOCOME)) 207 tp->t_flags |= TF_LASTIDLE; 208 else 209 tp->t_flags &= ~TF_LASTIDLE; 210 211 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 212 !IN_FASTRECOVERY(tp)) 213 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt); 214 215 again: 216 /* Make use of SACK information when slow-starting after a RTO. */ 217 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 218 !IN_FASTRECOVERY(tp)) { 219 tcp_seq old_snd_nxt = tp->snd_nxt; 220 221 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt); 222 nsacked += tp->snd_nxt - old_snd_nxt; 223 } 224 225 sendalot = FALSE; 226 off = tp->snd_nxt - tp->snd_una; 227 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked); 228 sendwin = min(sendwin, tp->snd_bwnd); 229 230 flags = tcp_outflags[tp->t_state]; 231 /* 232 * Get standard flags, and add SYN or FIN if requested by 'hidden' 233 * state flags. 234 */ 235 if (tp->t_flags & TF_NEEDFIN) 236 flags |= TH_FIN; 237 if (tp->t_flags & TF_NEEDSYN) 238 flags |= TH_SYN; 239 240 /* 241 * If in persist timeout with window of 0, send 1 byte. 242 * Otherwise, if window is small but nonzero 243 * and timer expired, we will send what we can 244 * and go to transmit state. 245 */ 246 if (tp->t_flags & TF_FORCE) { 247 if (sendwin == 0) { 248 /* 249 * If we still have some data to send, then 250 * clear the FIN bit. Usually this would 251 * happen below when it realizes that we 252 * aren't sending all the data. However, 253 * if we have exactly 1 byte of unsent data, 254 * then it won't clear the FIN bit below, 255 * and if we are in persist state, we wind 256 * up sending the packet without recording 257 * that we sent the FIN bit. 258 * 259 * We can't just blindly clear the FIN bit, 260 * because if we don't have any more data 261 * to send then the probe will be the FIN 262 * itself. 263 */ 264 if (off < so->so_snd.ssb_cc) 265 flags &= ~TH_FIN; 266 sendwin = 1; 267 } else { 268 tcp_callout_stop(tp, tp->tt_persist); 269 tp->t_rxtshift = 0; 270 } 271 } 272 273 /* 274 * If snd_nxt == snd_max and we have transmitted a FIN, the 275 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in 276 * a negative length. This can also occur when TCP opens up 277 * its congestion window while receiving additional duplicate 278 * acks after fast-retransmit because TCP will reset snd_nxt 279 * to snd_max after the fast-retransmit. 280 * 281 * In the normal retransmit-FIN-only case, however, snd_nxt will 282 * be set to snd_una, the offset will be 0, and the length may 283 * wind up 0. 284 */ 285 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off; 286 287 /* 288 * Lop off SYN bit if it has already been sent. However, if this 289 * is SYN-SENT state and if segment contains data, suppress sending 290 * segment (sending the segment would be an option if we still 291 * did TAO and the remote host supported it). 292 */ 293 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 294 flags &= ~TH_SYN; 295 off--, len++; 296 if (len > 0 && tp->t_state == TCPS_SYN_SENT) 297 return 0; 298 } 299 300 /* 301 * Be careful not to send data and/or FIN on SYN segments. 302 * This measure is needed to prevent interoperability problems 303 * with not fully conformant TCP implementations. 304 */ 305 if (flags & TH_SYN) { 306 len = 0; 307 flags &= ~TH_FIN; 308 } 309 310 if (len < 0) { 311 /* 312 * If FIN has been sent but not acked, 313 * but we haven't been called to retransmit, 314 * len will be < 0. Otherwise, window shrank 315 * after we sent into it. If window shrank to 0, 316 * cancel pending retransmit, pull snd_nxt back 317 * to (closed) window, and set the persist timer 318 * if it isn't already going. If the window didn't 319 * close completely, just wait for an ACK. 320 */ 321 len = 0; 322 if (sendwin == 0) { 323 tcp_callout_stop(tp, tp->tt_rexmt); 324 tp->t_rxtshift = 0; 325 tp->snd_nxt = tp->snd_una; 326 if (!tcp_callout_active(tp, tp->tt_persist)) 327 tcp_setpersist(tp); 328 } 329 } 330 331 KASSERT(len >= 0, ("%s: len < 0", __func__)); 332 /* 333 * Automatic sizing of send socket buffer. Often the send buffer 334 * size is not optimally adjusted to the actual network conditions 335 * at hand (delay bandwidth product). Setting the buffer size too 336 * small limits throughput on links with high bandwidth and high 337 * delay (eg. trans-continental/oceanic links). Setting the 338 * buffer size too big consumes too much real kernel memory, 339 * especially with many connections on busy servers. 340 * 341 * The criteria to step up the send buffer one notch are: 342 * 1. receive window of remote host is larger than send buffer 343 * (with a fudge factor of 5/4th); 344 * 2. send buffer is filled to 7/8th with data (so we actually 345 * have data to make use of it); 346 * 3. send buffer fill has not hit maximal automatic size; 347 * 4. our send window (slow start and cogestion controlled) is 348 * larger than sent but unacknowledged data in send buffer. 349 * 350 * The remote host receive window scaling factor may limit the 351 * growing of the send buffer before it reaches its allowed 352 * maximum. 353 * 354 * It scales directly with slow start or congestion window 355 * and does at most one step per received ACK. This fast 356 * scaling has the drawback of growing the send buffer beyond 357 * what is strictly necessary to make full use of a given 358 * delay*bandwith product. However testing has shown this not 359 * to be much of an problem. At worst we are trading wasting 360 * of available bandwith (the non-use of it) for wasting some 361 * socket buffer memory. 362 * 363 * TODO: Shrink send buffer during idle periods together 364 * with congestion window. Requires another timer. Has to 365 * wait for upcoming tcp timer rewrite. 366 */ 367 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) { 368 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat && 369 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) && 370 so->so_snd.ssb_cc < tcp_autosndbuf_max && 371 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) { 372 u_long newsize; 373 374 newsize = ulmin(so->so_snd.ssb_hiwat + 375 tcp_autosndbuf_inc, 376 tcp_autosndbuf_max); 377 if (!ssb_reserve(&so->so_snd, newsize, so, NULL)) 378 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE); 379 if (newsize >= (TCP_MAXWIN << tp->snd_scale)) 380 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE); 381 } 382 } 383 384 /* 385 * Truncate to the maximum segment length and ensure that FIN is 386 * removed if the length no longer contains the last data byte. 387 */ 388 if (len > tp->t_maxseg) { 389 len = tp->t_maxseg; 390 sendalot = TRUE; 391 } 392 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc)) 393 flags &= ~TH_FIN; 394 395 recvwin = ssb_space(&so->so_rcv); 396 397 /* 398 * Sender silly window avoidance. We transmit under the following 399 * conditions when len is non-zero: 400 * 401 * - We have a full segment 402 * - This is the last buffer in a write()/send() and we are 403 * either idle or running NODELAY 404 * - we've timed out (e.g. persist timer) 405 * - we have more then 1/2 the maximum send window's worth of 406 * data (receiver may be limiting the window size) 407 * - we need to retransmit 408 */ 409 if (len) { 410 if (len == tp->t_maxseg) 411 goto send; 412 /* 413 * NOTE! on localhost connections an 'ack' from the remote 414 * end may occur synchronously with the output and cause 415 * us to flush a buffer queued with moretocome. XXX 416 * 417 * note: the len + off check is almost certainly unnecessary. 418 */ 419 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 420 (idle || (tp->t_flags & TF_NODELAY)) && 421 len + off >= so->so_snd.ssb_cc && 422 !(tp->t_flags & TF_NOPUSH)) { 423 goto send; 424 } 425 if (tp->t_flags & TF_FORCE) /* typ. timeout case */ 426 goto send; 427 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 428 goto send; 429 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 430 goto send; 431 } 432 433 /* 434 * Compare available window to amount of window 435 * known to peer (as advertised window less 436 * next expected input). If the difference is at least two 437 * max size segments, or at least 50% of the maximum possible 438 * window, then want to send a window update to peer. 439 */ 440 if (recvwin > 0) { 441 /* 442 * "adv" is the amount we can increase the window, 443 * taking into account that we are limited by 444 * TCP_MAXWIN << tp->rcv_scale. 445 */ 446 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) - 447 (tp->rcv_adv - tp->rcv_nxt); 448 long hiwat; 449 450 /* 451 * This ack case typically occurs when the user has drained 452 * the TCP socket buffer sufficiently to warrent an ack 453 * containing a 'pure window update'... that is, an ack that 454 * ONLY updates the tcp window. 455 * 456 * It is unclear why we would need to do a pure window update 457 * past 2 segments if we are going to do one at 1/2 the high 458 * water mark anyway, especially since under normal conditions 459 * the user program will drain the socket buffer quickly. 460 * The 2-segment pure window update will often add a large 461 * number of extra, unnecessary acks to the stream. 462 * 463 * avoid_pure_win_update now defaults to 1. 464 */ 465 if (avoid_pure_win_update == 0 || 466 (tp->t_flags & TF_RXRESIZED)) { 467 if (adv >= (long) (2 * tp->t_maxseg)) { 468 goto send; 469 } 470 } 471 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale); 472 if (hiwat > (long)so->so_rcv.ssb_hiwat) 473 hiwat = (long)so->so_rcv.ssb_hiwat; 474 if (adv >= hiwat / 2) 475 goto send; 476 } 477 478 /* 479 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 480 * is also a catch-all for the retransmit timer timeout case. 481 */ 482 if (tp->t_flags & TF_ACKNOW) 483 goto send; 484 if ((flags & TH_RST) || 485 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN))) 486 goto send; 487 if (SEQ_GT(tp->snd_up, tp->snd_una)) 488 goto send; 489 /* 490 * If our state indicates that FIN should be sent 491 * and we have not yet done so, then we need to send. 492 */ 493 if (flags & TH_FIN && 494 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una)) 495 goto send; 496 497 /* 498 * TCP window updates are not reliable, rather a polling protocol 499 * using ``persist'' packets is used to insure receipt of window 500 * updates. The three ``states'' for the output side are: 501 * idle not doing retransmits or persists 502 * persisting to move a small or zero window 503 * (re)transmitting and thereby not persisting 504 * 505 * tcp_callout_active(tp, tp->tt_persist) 506 * is true when we are in persist state. 507 * The TF_FORCE flag in tp->t_flags 508 * is set when we are called to send a persist packet. 509 * tcp_callout_active(tp, tp->tt_rexmt) 510 * is set when we are retransmitting 511 * The output side is idle when both timers are zero. 512 * 513 * If send window is too small, there is data to transmit, and no 514 * retransmit or persist is pending, then go to persist state. 515 * If nothing happens soon, send when timer expires: 516 * if window is nonzero, transmit what we can, 517 * otherwise force out a byte. 518 */ 519 if (so->so_snd.ssb_cc > 0 && 520 !tcp_callout_active(tp, tp->tt_rexmt) && 521 !tcp_callout_active(tp, tp->tt_persist)) { 522 tp->t_rxtshift = 0; 523 tcp_setpersist(tp); 524 } 525 526 /* 527 * No reason to send a segment, just return. 528 */ 529 return (0); 530 531 send: 532 /* 533 * Before ESTABLISHED, force sending of initial options 534 * unless TCP set not to do any options. 535 * NOTE: we assume that the IP/TCP header plus TCP options 536 * always fit in a single mbuf, leaving room for a maximum 537 * link header, i.e. 538 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES 539 */ 540 optlen = 0; 541 if (isipv6) 542 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 543 else 544 hdrlen = sizeof(struct tcpiphdr); 545 if (flags & TH_SYN) { 546 tp->snd_nxt = tp->iss; 547 if (!(tp->t_flags & TF_NOOPT)) { 548 u_short mss; 549 550 opt[0] = TCPOPT_MAXSEG; 551 opt[1] = TCPOLEN_MAXSEG; 552 mss = htons((u_short) tcp_mssopt(tp)); 553 memcpy(opt + 2, &mss, sizeof mss); 554 optlen = TCPOLEN_MAXSEG; 555 556 if ((tp->t_flags & TF_REQ_SCALE) && 557 (!(flags & TH_ACK) || 558 (tp->t_flags & TF_RCVD_SCALE))) { 559 *((u_int32_t *)(opt + optlen)) = htonl( 560 TCPOPT_NOP << 24 | 561 TCPOPT_WINDOW << 16 | 562 TCPOLEN_WINDOW << 8 | 563 tp->request_r_scale); 564 optlen += 4; 565 } 566 567 if ((tcp_do_sack && !(flags & TH_ACK)) || 568 tp->t_flags & TF_SACK_PERMITTED) { 569 uint32_t *lp = (uint32_t *)(opt + optlen); 570 571 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED); 572 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED; 573 } 574 } 575 } 576 577 /* 578 * Send a timestamp and echo-reply if this is a SYN and our side 579 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 580 * and our peer have sent timestamps in our SYN's. 581 */ 582 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 583 !(flags & TH_RST) && 584 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) { 585 u_int32_t *lp = (u_int32_t *)(opt + optlen); 586 587 /* Form timestamp option as shown in appendix A of RFC 1323. */ 588 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 589 *lp++ = htonl(ticks); 590 *lp = htonl(tp->ts_recent); 591 optlen += TCPOLEN_TSTAMP_APPA; 592 } 593 594 /* Set receive buffer autosizing timestamp. */ 595 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) 596 tp->rfbuf_ts = ticks; 597 598 /* 599 * If this is a SACK connection and we have a block to report, 600 * fill in the SACK blocks in the TCP options. 601 */ 602 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) == 603 TF_SACK_PERMITTED && 604 (!LIST_EMPTY(&tp->t_segq) || 605 tp->reportblk.rblk_start != tp->reportblk.rblk_end)) 606 tcp_sack_fill_report(tp, opt, &optlen); 607 608 #ifdef TCP_SIGNATURE 609 if (tp->t_flags & TF_SIGNATURE) { 610 int i; 611 u_char *bp; 612 /* 613 * Initialize TCP-MD5 option (RFC2385) 614 */ 615 bp = (u_char *)opt + optlen; 616 *bp++ = TCPOPT_SIGNATURE; 617 *bp++ = TCPOLEN_SIGNATURE; 618 sigoff = optlen + 2; 619 for (i = 0; i < TCP_SIGLEN; i++) 620 *bp++ = 0; 621 optlen += TCPOLEN_SIGNATURE; 622 /* 623 * Terminate options list and maintain 32-bit alignment. 624 */ 625 *bp++ = TCPOPT_NOP; 626 *bp++ = TCPOPT_EOL; 627 optlen += 2; 628 } 629 #endif /* TCP_SIGNATURE */ 630 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options")); 631 hdrlen += optlen; 632 633 if (isipv6) { 634 ipoptlen = ip6_optlen(inp); 635 } else { 636 if (inp->inp_options) { 637 ipoptlen = inp->inp_options->m_len - 638 offsetof(struct ipoption, ipopt_list); 639 } else { 640 ipoptlen = 0; 641 } 642 } 643 #ifdef IPSEC 644 ipoptlen += ipsec_hdrsiz_tcp(tp); 645 #endif 646 647 /* 648 * Adjust data length if insertion of options will bump the packet 649 * length beyond the t_maxopd length. Clear FIN to prevent premature 650 * closure since there is still more data to send after this (now 651 * truncated) packet. 652 * 653 * If just the options do not fit we are in a no-win situation and 654 * we treat it as an unreachable host. 655 */ 656 if (len + optlen + ipoptlen > tp->t_maxopd) { 657 if (tp->t_maxopd <= optlen + ipoptlen) { 658 static time_t last_optlen_report; 659 660 if (last_optlen_report != time_second) { 661 last_optlen_report = time_second; 662 kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd); 663 } 664 error = EHOSTUNREACH; 665 goto out; 666 } else { 667 flags &= ~TH_FIN; 668 len = tp->t_maxopd - optlen - ipoptlen; 669 sendalot = TRUE; 670 } 671 } 672 673 #ifdef INET6 674 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big")); 675 #else 676 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big")); 677 #endif 678 679 /* 680 * Grab a header mbuf, attaching a copy of data to 681 * be transmitted, and initialize the header from 682 * the template for sends on this connection. 683 */ 684 if (len) { 685 if ((tp->t_flags & TF_FORCE) && len == 1) 686 tcpstat.tcps_sndprobe++; 687 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 688 if (tp->snd_nxt == tp->snd_una) 689 tp->snd_max_rexmt = tp->snd_max; 690 tcpstat.tcps_sndrexmitpack++; 691 tcpstat.tcps_sndrexmitbyte += len; 692 } else { 693 tcpstat.tcps_sndpack++; 694 tcpstat.tcps_sndbyte += len; 695 } 696 #ifdef notyet 697 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len, 698 max_linkhdr + hdrlen)) == NULL) { 699 error = ENOBUFS; 700 goto out; 701 } 702 /* 703 * m_copypack left space for our hdr; use it. 704 */ 705 m->m_len += hdrlen; 706 m->m_data -= hdrlen; 707 #else 708 #ifndef INET6 709 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 710 #else 711 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER, 712 M_PKTHDR, NULL); 713 #endif 714 if (m == NULL) { 715 error = ENOBUFS; 716 goto out; 717 } 718 m->m_data += max_linkhdr; 719 m->m_len = hdrlen; 720 if (len <= MHLEN - hdrlen - max_linkhdr) { 721 m_copydata(so->so_snd.ssb_mb, off, (int) len, 722 mtod(m, caddr_t) + hdrlen); 723 m->m_len += len; 724 } else { 725 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len); 726 if (m->m_next == NULL) { 727 m_free(m); 728 error = ENOBUFS; 729 goto out; 730 } 731 } 732 #endif 733 /* 734 * If we're sending everything we've got, set PUSH. 735 * (This will keep happy those implementations which only 736 * give data to the user when a buffer fills or 737 * a PUSH comes in.) 738 */ 739 if (off + len == so->so_snd.ssb_cc) 740 flags |= TH_PUSH; 741 } else { 742 if (tp->t_flags & TF_ACKNOW) 743 tcpstat.tcps_sndacks++; 744 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 745 tcpstat.tcps_sndctrl++; 746 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 747 tcpstat.tcps_sndurg++; 748 else 749 tcpstat.tcps_sndwinup++; 750 751 MGETHDR(m, MB_DONTWAIT, MT_HEADER); 752 if (m == NULL) { 753 error = ENOBUFS; 754 goto out; 755 } 756 if (isipv6 && 757 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN) 758 MH_ALIGN(m, hdrlen); 759 else 760 m->m_data += max_linkhdr; 761 m->m_len = hdrlen; 762 } 763 m->m_pkthdr.rcvif = NULL; 764 if (isipv6) { 765 ip6 = mtod(m, struct ip6_hdr *); 766 th = (struct tcphdr *)(ip6 + 1); 767 tcp_fillheaders(tp, ip6, th); 768 } else { 769 ip = mtod(m, struct ip *); 770 ipov = (struct ipovly *)ip; 771 th = (struct tcphdr *)(ip + 1); 772 /* this picks up the pseudo header (w/o the length) */ 773 tcp_fillheaders(tp, ip, th); 774 } 775 776 /* 777 * Fill in fields, remembering maximum advertised 778 * window for use in delaying messages about window sizes. 779 * If resending a FIN, be sure not to use a new sequence number. 780 */ 781 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 782 tp->snd_nxt == tp->snd_max) 783 tp->snd_nxt--; 784 /* 785 * If we are doing retransmissions, then snd_nxt will 786 * not reflect the first unsent octet. For ACK only 787 * packets, we do not want the sequence number of the 788 * retransmitted packet, we want the sequence number 789 * of the next unsent octet. So, if there is no data 790 * (and no SYN or FIN), use snd_max instead of snd_nxt 791 * when filling in ti_seq. But if we are in persist 792 * state, snd_max might reflect one byte beyond the 793 * right edge of the window, so use snd_nxt in that 794 * case, since we know we aren't doing a retransmission. 795 * (retransmit and persist are mutually exclusive...) 796 */ 797 if (len || (flags & (TH_SYN|TH_FIN)) || 798 tcp_callout_active(tp, tp->tt_persist)) 799 th->th_seq = htonl(tp->snd_nxt); 800 else 801 th->th_seq = htonl(tp->snd_max); 802 th->th_ack = htonl(tp->rcv_nxt); 803 if (optlen) { 804 bcopy(opt, th + 1, optlen); 805 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 806 } 807 th->th_flags = flags; 808 809 /* 810 * Calculate receive window. Don't shrink window, but avoid 811 * silly window syndrome by sending a 0 window if the actual 812 * window is less then one segment. 813 */ 814 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) && 815 recvwin < (long)tp->t_maxseg) 816 recvwin = 0; 817 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt)) 818 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt); 819 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale) 820 recvwin = (long)TCP_MAXWIN << tp->rcv_scale; 821 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale)); 822 823 /* 824 * Adjust the RXWIN0SENT flag - indicate that we have advertised 825 * a 0 window. This may cause the remote transmitter to stall. This 826 * flag tells soreceive() to disable delayed acknowledgements when 827 * draining the buffer. This can occur if the receiver is attempting 828 * to read more data then can be buffered prior to transmitting on 829 * the connection. 830 */ 831 if (recvwin == 0) 832 tp->t_flags |= TF_RXWIN0SENT; 833 else 834 tp->t_flags &= ~TF_RXWIN0SENT; 835 836 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 837 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 838 th->th_flags |= TH_URG; 839 } else { 840 /* 841 * If no urgent pointer to send, then we pull 842 * the urgent pointer to the left edge of the send window 843 * so that it doesn't drift into the send window on sequence 844 * number wraparound. 845 */ 846 tp->snd_up = tp->snd_una; /* drag it along */ 847 } 848 849 #ifdef TCP_SIGNATURE 850 if (tp->t_flags & TF_SIGNATURE) 851 tcpsignature_compute(m, len, optlen, 852 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 853 #endif /* TCP_SIGNATURE */ 854 855 /* 856 * Put TCP length in extended header, and then 857 * checksum extended header and data. 858 */ 859 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 860 if (isipv6) { 861 /* 862 * ip6_plen is not need to be filled now, and will be filled 863 * in ip6_output(). 864 */ 865 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 866 sizeof(struct tcphdr) + optlen + len); 867 } else { 868 m->m_pkthdr.csum_flags = CSUM_TCP; 869 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 870 if (len + optlen) 871 th->th_sum = in_addword(th->th_sum, 872 htons((u_short)(optlen + len))); 873 874 /* IP version must be set here for ipv4/ipv6 checking later */ 875 KASSERT(ip->ip_v == IPVERSION, 876 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 877 } 878 879 /* 880 * In transmit state, time the transmission and arrange for 881 * the retransmit. In persist state, just set snd_max. 882 */ 883 if (!(tp->t_flags & TF_FORCE) || 884 !tcp_callout_active(tp, tp->tt_persist)) { 885 tcp_seq startseq = tp->snd_nxt; 886 887 /* 888 * Advance snd_nxt over sequence space of this segment. 889 */ 890 if (flags & (TH_SYN | TH_FIN)) { 891 if (flags & TH_SYN) 892 tp->snd_nxt++; 893 if (flags & TH_FIN) { 894 tp->snd_nxt++; 895 tp->t_flags |= TF_SENTFIN; 896 } 897 } 898 tp->snd_nxt += len; 899 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 900 tp->snd_max = tp->snd_nxt; 901 /* 902 * Time this transmission if not a retransmission and 903 * not currently timing anything. 904 */ 905 if (tp->t_rtttime == 0) { 906 tp->t_rtttime = ticks; 907 tp->t_rtseq = startseq; 908 tcpstat.tcps_segstimed++; 909 } 910 } 911 912 /* 913 * Set retransmit timer if not currently set, 914 * and not doing a pure ack or a keep-alive probe. 915 * Initial value for retransmit timer is smoothed 916 * round-trip time + 2 * round-trip time variance. 917 * Initialize shift counter which is used for backoff 918 * of retransmit time. 919 */ 920 if (!tcp_callout_active(tp, tp->tt_rexmt) && 921 tp->snd_nxt != tp->snd_una) { 922 if (tcp_callout_active(tp, tp->tt_persist)) { 923 tcp_callout_stop(tp, tp->tt_persist); 924 tp->t_rxtshift = 0; 925 } 926 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur, 927 tcp_timer_rexmt); 928 } 929 } else { 930 /* 931 * Persist case, update snd_max but since we are in 932 * persist mode (no window) we do not update snd_nxt. 933 */ 934 int xlen = len; 935 if (flags & TH_SYN) 936 ++xlen; 937 if (flags & TH_FIN) { 938 ++xlen; 939 tp->t_flags |= TF_SENTFIN; 940 } 941 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 942 tp->snd_max = tp->snd_nxt + xlen; 943 } 944 945 #ifdef TCPDEBUG 946 /* 947 * Trace. 948 */ 949 if (so->so_options & SO_DEBUG) 950 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 951 #endif 952 953 /* 954 * Fill in IP length and desired time to live and 955 * send to IP level. There should be a better way 956 * to handle ttl and tos; we could keep them in 957 * the template, but need a way to checksum without them. 958 */ 959 /* 960 * m->m_pkthdr.len should have been set before cksum calcuration, 961 * because in6_cksum() need it. 962 */ 963 if (isipv6) { 964 /* 965 * we separately set hoplimit for every segment, since the 966 * user might want to change the value via setsockopt. 967 * Also, desired default hop limit might be changed via 968 * Neighbor Discovery. 969 */ 970 ip6->ip6_hlim = in6_selecthlim(inp, 971 (inp->in6p_route.ro_rt ? 972 inp->in6p_route.ro_rt->rt_ifp : NULL)); 973 974 /* TODO: IPv6 IP6TOS_ECT bit on */ 975 error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route, 976 (so->so_options & SO_DONTROUTE), NULL, NULL, 977 inp); 978 } else { 979 struct rtentry *rt; 980 ip->ip_len = m->m_pkthdr.len; 981 #ifdef INET6 982 if (INP_CHECK_SOCKAF(so, AF_INET6)) 983 ip->ip_ttl = in6_selecthlim(inp, 984 (inp->in6p_route.ro_rt ? 985 inp->in6p_route.ro_rt->rt_ifp : NULL)); 986 else 987 #endif 988 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */ 989 990 ip->ip_tos = inp->inp_ip_tos; /* XXX */ 991 /* 992 * See if we should do MTU discovery. 993 * We do it only if the following are true: 994 * 1) we have a valid route to the destination 995 * 2) the MTU is not locked (if it is, 996 * then discovery has been disabled) 997 */ 998 if (path_mtu_discovery && 999 (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) && 1000 !(rt->rt_rmx.rmx_locks & RTV_MTU)) 1001 ip->ip_off |= IP_DF; 1002 1003 error = ip_output(m, inp->inp_options, &inp->inp_route, 1004 (so->so_options & SO_DONTROUTE) | 1005 IP_DEBUGROUTE, NULL, inp); 1006 } 1007 if (error) { 1008 1009 /* 1010 * We know that the packet was lost, so back out the 1011 * sequence number advance, if any. 1012 */ 1013 if (!(tp->t_flags & TF_FORCE) || 1014 !tcp_callout_active(tp, tp->tt_persist)) { 1015 /* 1016 * No need to check for TH_FIN here because 1017 * the TF_SENTFIN flag handles that case. 1018 */ 1019 if (!(flags & TH_SYN)) 1020 tp->snd_nxt -= len; 1021 } 1022 1023 out: 1024 if (error == ENOBUFS) { 1025 /* 1026 * If we can't send, make sure there is something 1027 * to get us going again later. Persist state 1028 * is not necessarily right, but it is close enough. 1029 */ 1030 if (!tcp_callout_active(tp, tp->tt_rexmt) && 1031 !tcp_callout_active(tp, tp->tt_persist)) { 1032 tp->t_rxtshift = 0; 1033 tcp_setpersist(tp); 1034 } 1035 tcp_quench(inp, 0); 1036 return (0); 1037 } 1038 if (error == EMSGSIZE) { 1039 /* 1040 * ip_output() will have already fixed the route 1041 * for us. tcp_mtudisc() will, as its last action, 1042 * initiate retransmission, so it is important to 1043 * not do so here. 1044 */ 1045 tcp_mtudisc(inp, 0); 1046 return 0; 1047 } 1048 if ((error == EHOSTUNREACH || error == ENETDOWN) && 1049 TCPS_HAVERCVDSYN(tp->t_state)) { 1050 tp->t_softerror = error; 1051 return (0); 1052 } 1053 return (error); 1054 } 1055 tcpstat.tcps_sndtotal++; 1056 1057 /* 1058 * Data sent (as far as we can tell). 1059 * 1060 * If this advertises a larger window than any other segment, 1061 * then remember the size of the advertised window. 1062 * 1063 * Any pending ACK has now been sent. 1064 */ 1065 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) { 1066 tp->rcv_adv = tp->rcv_nxt + recvwin; 1067 tp->t_flags &= ~TF_RXRESIZED; 1068 } 1069 tp->last_ack_sent = tp->rcv_nxt; 1070 tp->t_flags &= ~TF_ACKNOW; 1071 if (tcp_delack_enabled) 1072 tcp_callout_stop(tp, tp->tt_delack); 1073 if (sendalot) 1074 goto again; 1075 return (0); 1076 } 1077 1078 void 1079 tcp_setpersist(struct tcpcb *tp) 1080 { 1081 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1082 int tt; 1083 1084 if (tcp_callout_active(tp, tp->tt_rexmt)) 1085 panic("tcp_setpersist: retransmit pending"); 1086 /* 1087 * Start/restart persistance timer. 1088 */ 1089 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN, 1090 TCPTV_PERSMAX); 1091 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist); 1092 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1093 tp->t_rxtshift++; 1094 } 1095