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