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 = 0; 155 #ifdef TCP_SIGNATURE 156 int sigoff = 0; 157 #endif 158 struct mbuf *m; 159 struct ip *ip; 160 struct ipovly *ipov; 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; 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 m = NULL; 217 ip = NULL; 218 ipov = NULL; 219 th = NULL; 220 ip6 = NULL; 221 222 /* Make use of SACK information when slow-starting after a RTO. */ 223 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 224 !IN_FASTRECOVERY(tp)) { 225 tcp_seq old_snd_nxt = tp->snd_nxt; 226 227 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt); 228 nsacked += tp->snd_nxt - old_snd_nxt; 229 } 230 231 sendalot = FALSE; 232 off = tp->snd_nxt - tp->snd_una; 233 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked); 234 sendwin = min(sendwin, tp->snd_bwnd); 235 236 flags = tcp_outflags[tp->t_state]; 237 /* 238 * Get standard flags, and add SYN or FIN if requested by 'hidden' 239 * state flags. 240 */ 241 if (tp->t_flags & TF_NEEDFIN) 242 flags |= TH_FIN; 243 if (tp->t_flags & TF_NEEDSYN) 244 flags |= TH_SYN; 245 246 /* 247 * If in persist timeout with window of 0, send 1 byte. 248 * Otherwise, if window is small but nonzero 249 * and timer expired, we will send what we can 250 * and go to transmit state. 251 */ 252 if (tp->t_flags & TF_FORCE) { 253 if (sendwin == 0) { 254 /* 255 * If we still have some data to send, then 256 * clear the FIN bit. Usually this would 257 * happen below when it realizes that we 258 * aren't sending all the data. However, 259 * if we have exactly 1 byte of unsent data, 260 * then it won't clear the FIN bit below, 261 * and if we are in persist state, we wind 262 * up sending the packet without recording 263 * that we sent the FIN bit. 264 * 265 * We can't just blindly clear the FIN bit, 266 * because if we don't have any more data 267 * to send then the probe will be the FIN 268 * itself. 269 */ 270 if (off < so->so_snd.ssb_cc) 271 flags &= ~TH_FIN; 272 sendwin = 1; 273 } else { 274 tcp_callout_stop(tp, tp->tt_persist); 275 tp->t_rxtshift = 0; 276 } 277 } 278 279 /* 280 * If snd_nxt == snd_max and we have transmitted a FIN, the 281 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in 282 * a negative length. This can also occur when TCP opens up 283 * its congestion window while receiving additional duplicate 284 * acks after fast-retransmit because TCP will reset snd_nxt 285 * to snd_max after the fast-retransmit. 286 * 287 * A negative length can also occur when we are in the 288 * TCPS_SYN_RECEIVED state due to a simultanious connect where 289 * our SYN has not been acked yet. 290 * 291 * In the normal retransmit-FIN-only case, however, snd_nxt will 292 * be set to snd_una, the offset will be 0, and the length may 293 * wind up 0. 294 */ 295 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off; 296 297 /* 298 * Lop off SYN bit if it has already been sent. However, if this 299 * is SYN-SENT state and if segment contains data, suppress sending 300 * segment (sending the segment would be an option if we still 301 * did TAO and the remote host supported it). 302 */ 303 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 304 flags &= ~TH_SYN; 305 off--, len++; 306 if (len > 0 && tp->t_state == TCPS_SYN_SENT) 307 return 0; 308 } 309 310 /* 311 * Be careful not to send data and/or FIN on SYN segments. 312 * This measure is needed to prevent interoperability problems 313 * with not fully conformant TCP implementations. 314 */ 315 if (flags & TH_SYN) { 316 len = 0; 317 flags &= ~TH_FIN; 318 } 319 320 if (len < 0) { 321 /* 322 * A negative len can occur if our FIN has been sent but not 323 * acked, or if we are in a simultanious connect in the 324 * TCPS_SYN_RECEIVED state with our SYN sent but not yet 325 * acked. 326 * 327 * If our window has contracted to 0 in the FIN case 328 * (which can only occur if we have NOT been called to 329 * retransmit as per code a few paragraphs up) then we 330 * want to shift the retransmit timer over to the 331 * persist timer. 332 * 333 * However, if we are in the TCPS_SYN_RECEIVED state 334 * (the SYN case) we will be in a simultanious connect and 335 * the window may be zero degeneratively. In this case we 336 * do not want to shift to the persist timer after the SYN 337 * or the SYN+ACK transmission. 338 */ 339 len = 0; 340 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) { 341 tcp_callout_stop(tp, tp->tt_rexmt); 342 tp->t_rxtshift = 0; 343 tp->snd_nxt = tp->snd_una; 344 if (!tcp_callout_active(tp, tp->tt_persist)) 345 tcp_setpersist(tp); 346 } 347 } 348 349 KASSERT(len >= 0, ("%s: len < 0", __func__)); 350 /* 351 * Automatic sizing of send socket buffer. Often the send buffer 352 * size is not optimally adjusted to the actual network conditions 353 * at hand (delay bandwidth product). Setting the buffer size too 354 * small limits throughput on links with high bandwidth and high 355 * delay (eg. trans-continental/oceanic links). Setting the 356 * buffer size too big consumes too much real kernel memory, 357 * especially with many connections on busy servers. 358 * 359 * The criteria to step up the send buffer one notch are: 360 * 1. receive window of remote host is larger than send buffer 361 * (with a fudge factor of 5/4th); 362 * 2. send buffer is filled to 7/8th with data (so we actually 363 * have data to make use of it); 364 * 3. send buffer fill has not hit maximal automatic size; 365 * 4. our send window (slow start and cogestion controlled) is 366 * larger than sent but unacknowledged data in send buffer. 367 * 368 * The remote host receive window scaling factor may limit the 369 * growing of the send buffer before it reaches its allowed 370 * maximum. 371 * 372 * It scales directly with slow start or congestion window 373 * and does at most one step per received ACK. This fast 374 * scaling has the drawback of growing the send buffer beyond 375 * what is strictly necessary to make full use of a given 376 * delay*bandwith product. However testing has shown this not 377 * to be much of an problem. At worst we are trading wasting 378 * of available bandwith (the non-use of it) for wasting some 379 * socket buffer memory. 380 * 381 * TODO: Shrink send buffer during idle periods together 382 * with congestion window. Requires another timer. Has to 383 * wait for upcoming tcp timer rewrite. 384 */ 385 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) { 386 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat && 387 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) && 388 so->so_snd.ssb_cc < tcp_autosndbuf_max && 389 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) { 390 u_long newsize; 391 392 newsize = ulmin(so->so_snd.ssb_hiwat + 393 tcp_autosndbuf_inc, 394 tcp_autosndbuf_max); 395 if (!ssb_reserve(&so->so_snd, newsize, so, NULL)) 396 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE); 397 if (newsize >= (TCP_MAXWIN << tp->snd_scale)) 398 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE); 399 } 400 } 401 402 /* 403 * Truncate to the maximum segment length and ensure that FIN is 404 * removed if the length no longer contains the last data byte. 405 */ 406 if (len > tp->t_maxseg) { 407 len = tp->t_maxseg; 408 sendalot = TRUE; 409 } 410 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc)) 411 flags &= ~TH_FIN; 412 413 recvwin = ssb_space(&so->so_rcv); 414 415 /* 416 * Sender silly window avoidance. We transmit under the following 417 * conditions when len is non-zero: 418 * 419 * - We have a full segment 420 * - This is the last buffer in a write()/send() and we are 421 * either idle or running NODELAY 422 * - we've timed out (e.g. persist timer) 423 * - we have more then 1/2 the maximum send window's worth of 424 * data (receiver may be limiting the window size) 425 * - we need to retransmit 426 */ 427 if (len) { 428 if (len == tp->t_maxseg) 429 goto send; 430 /* 431 * NOTE! on localhost connections an 'ack' from the remote 432 * end may occur synchronously with the output and cause 433 * us to flush a buffer queued with moretocome. XXX 434 * 435 * note: the len + off check is almost certainly unnecessary. 436 */ 437 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 438 (idle || (tp->t_flags & TF_NODELAY)) && 439 len + off >= so->so_snd.ssb_cc && 440 !(tp->t_flags & TF_NOPUSH)) { 441 goto send; 442 } 443 if (tp->t_flags & TF_FORCE) /* typ. timeout case */ 444 goto send; 445 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 446 goto send; 447 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 448 goto send; 449 } 450 451 /* 452 * Compare available window to amount of window 453 * known to peer (as advertised window less 454 * next expected input). If the difference is at least two 455 * max size segments, or at least 50% of the maximum possible 456 * window, then want to send a window update to peer. 457 */ 458 if (recvwin > 0) { 459 /* 460 * "adv" is the amount we can increase the window, 461 * taking into account that we are limited by 462 * TCP_MAXWIN << tp->rcv_scale. 463 */ 464 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) - 465 (tp->rcv_adv - tp->rcv_nxt); 466 long hiwat; 467 468 /* 469 * This ack case typically occurs when the user has drained 470 * the TCP socket buffer sufficiently to warrent an ack 471 * containing a 'pure window update'... that is, an ack that 472 * ONLY updates the tcp window. 473 * 474 * It is unclear why we would need to do a pure window update 475 * past 2 segments if we are going to do one at 1/2 the high 476 * water mark anyway, especially since under normal conditions 477 * the user program will drain the socket buffer quickly. 478 * The 2-segment pure window update will often add a large 479 * number of extra, unnecessary acks to the stream. 480 * 481 * avoid_pure_win_update now defaults to 1. 482 */ 483 if (avoid_pure_win_update == 0 || 484 (tp->t_flags & TF_RXRESIZED)) { 485 if (adv >= (long) (2 * tp->t_maxseg)) { 486 goto send; 487 } 488 } 489 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale); 490 if (hiwat > (long)so->so_rcv.ssb_hiwat) 491 hiwat = (long)so->so_rcv.ssb_hiwat; 492 if (adv >= hiwat / 2) 493 goto send; 494 } 495 496 /* 497 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 498 * is also a catch-all for the retransmit timer timeout case. 499 */ 500 if (tp->t_flags & TF_ACKNOW) 501 goto send; 502 if ((flags & TH_RST) || 503 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN))) 504 goto send; 505 if (SEQ_GT(tp->snd_up, tp->snd_una)) 506 goto send; 507 /* 508 * If our state indicates that FIN should be sent 509 * and we have not yet done so, then we need to send. 510 */ 511 if ((flags & TH_FIN) && 512 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una)) 513 goto send; 514 515 /* 516 * TCP window updates are not reliable, rather a polling protocol 517 * using ``persist'' packets is used to insure receipt of window 518 * updates. The three ``states'' for the output side are: 519 * idle not doing retransmits or persists 520 * persisting to move a small or zero window 521 * (re)transmitting and thereby not persisting 522 * 523 * tcp_callout_active(tp, tp->tt_persist) 524 * is true when we are in persist state. 525 * The TF_FORCE flag in tp->t_flags 526 * is set when we are called to send a persist packet. 527 * tcp_callout_active(tp, tp->tt_rexmt) 528 * is set when we are retransmitting 529 * The output side is idle when both timers are zero. 530 * 531 * If send window is too small, there is data to transmit, and no 532 * retransmit or persist is pending, then go to persist state. 533 * 534 * If nothing happens soon, send when timer expires: 535 * if window is nonzero, transmit what we can, otherwise force out 536 * a byte. 537 * 538 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED 539 * with data pending. This situation can occur during a 540 * simultanious connect. 541 */ 542 if (so->so_snd.ssb_cc > 0 && 543 tp->t_state != TCPS_SYN_RECEIVED && 544 !tcp_callout_active(tp, tp->tt_rexmt) && 545 !tcp_callout_active(tp, tp->tt_persist)) { 546 tp->t_rxtshift = 0; 547 tcp_setpersist(tp); 548 } 549 550 /* 551 * No reason to send a segment, just return. 552 */ 553 return (0); 554 555 send: 556 /* 557 * Before ESTABLISHED, force sending of initial options 558 * unless TCP set not to do any options. 559 * NOTE: we assume that the IP/TCP header plus TCP options 560 * always fit in a single mbuf, leaving room for a maximum 561 * link header, i.e. 562 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES 563 */ 564 optlen = 0; 565 if (isipv6) 566 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 567 else 568 hdrlen = sizeof(struct tcpiphdr); 569 if (flags & TH_SYN) { 570 tp->snd_nxt = tp->iss; 571 if (!(tp->t_flags & TF_NOOPT)) { 572 u_short mss; 573 574 opt[0] = TCPOPT_MAXSEG; 575 opt[1] = TCPOLEN_MAXSEG; 576 mss = htons((u_short) tcp_mssopt(tp)); 577 memcpy(opt + 2, &mss, sizeof mss); 578 optlen = TCPOLEN_MAXSEG; 579 580 if ((tp->t_flags & TF_REQ_SCALE) && 581 (!(flags & TH_ACK) || 582 (tp->t_flags & TF_RCVD_SCALE))) { 583 *((u_int32_t *)(opt + optlen)) = htonl( 584 TCPOPT_NOP << 24 | 585 TCPOPT_WINDOW << 16 | 586 TCPOLEN_WINDOW << 8 | 587 tp->request_r_scale); 588 optlen += 4; 589 } 590 591 if ((tcp_do_sack && !(flags & TH_ACK)) || 592 tp->t_flags & TF_SACK_PERMITTED) { 593 uint32_t *lp = (uint32_t *)(opt + optlen); 594 595 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED); 596 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED; 597 } 598 } 599 } 600 601 /* 602 * Send a timestamp and echo-reply if this is a SYN and our side 603 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 604 * and our peer have sent timestamps in our SYN's. 605 */ 606 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 607 !(flags & TH_RST) && 608 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) { 609 u_int32_t *lp = (u_int32_t *)(opt + optlen); 610 611 /* Form timestamp option as shown in appendix A of RFC 1323. */ 612 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 613 *lp++ = htonl(ticks); 614 *lp = htonl(tp->ts_recent); 615 optlen += TCPOLEN_TSTAMP_APPA; 616 } 617 618 /* Set receive buffer autosizing timestamp. */ 619 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) 620 tp->rfbuf_ts = ticks; 621 622 /* 623 * If this is a SACK connection and we have a block to report, 624 * fill in the SACK blocks in the TCP options. 625 */ 626 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) == 627 TF_SACK_PERMITTED && 628 (!LIST_EMPTY(&tp->t_segq) || 629 tp->reportblk.rblk_start != tp->reportblk.rblk_end)) 630 tcp_sack_fill_report(tp, opt, &optlen); 631 632 #ifdef TCP_SIGNATURE 633 if (tp->t_flags & TF_SIGNATURE) { 634 int i; 635 u_char *bp; 636 /* 637 * Initialize TCP-MD5 option (RFC2385) 638 */ 639 bp = (u_char *)opt + optlen; 640 *bp++ = TCPOPT_SIGNATURE; 641 *bp++ = TCPOLEN_SIGNATURE; 642 sigoff = optlen + 2; 643 for (i = 0; i < TCP_SIGLEN; i++) 644 *bp++ = 0; 645 optlen += TCPOLEN_SIGNATURE; 646 /* 647 * Terminate options list and maintain 32-bit alignment. 648 */ 649 *bp++ = TCPOPT_NOP; 650 *bp++ = TCPOPT_EOL; 651 optlen += 2; 652 } 653 #endif /* TCP_SIGNATURE */ 654 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options")); 655 hdrlen += optlen; 656 657 if (isipv6) { 658 ipoptlen = ip6_optlen(inp); 659 } else { 660 if (inp->inp_options) { 661 ipoptlen = inp->inp_options->m_len - 662 offsetof(struct ipoption, ipopt_list); 663 } else { 664 ipoptlen = 0; 665 } 666 } 667 #ifdef IPSEC 668 ipoptlen += ipsec_hdrsiz_tcp(tp); 669 #endif 670 671 /* 672 * Adjust data length if insertion of options will bump the packet 673 * length beyond the t_maxopd length. Clear FIN to prevent premature 674 * closure since there is still more data to send after this (now 675 * truncated) packet. 676 * 677 * If just the options do not fit we are in a no-win situation and 678 * we treat it as an unreachable host. 679 */ 680 if (len + optlen + ipoptlen > tp->t_maxopd) { 681 if (tp->t_maxopd <= optlen + ipoptlen) { 682 static time_t last_optlen_report; 683 684 if (last_optlen_report != time_second) { 685 last_optlen_report = time_second; 686 kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd); 687 } 688 error = EHOSTUNREACH; 689 goto out; 690 } else { 691 flags &= ~TH_FIN; 692 len = tp->t_maxopd - optlen - ipoptlen; 693 sendalot = TRUE; 694 } 695 } 696 697 #ifdef INET6 698 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big")); 699 #else 700 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big")); 701 #endif 702 703 /* 704 * Grab a header mbuf, attaching a copy of data to 705 * be transmitted, and initialize the header from 706 * the template for sends on this connection. 707 */ 708 if (len) { 709 if ((tp->t_flags & TF_FORCE) && len == 1) 710 tcpstat.tcps_sndprobe++; 711 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 712 if (tp->snd_nxt == tp->snd_una) 713 tp->snd_max_rexmt = tp->snd_max; 714 tcpstat.tcps_sndrexmitpack++; 715 tcpstat.tcps_sndrexmitbyte += len; 716 } else { 717 tcpstat.tcps_sndpack++; 718 tcpstat.tcps_sndbyte += len; 719 } 720 #ifdef notyet 721 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len, 722 max_linkhdr + hdrlen)) == NULL) { 723 error = ENOBUFS; 724 goto after_th; 725 } 726 /* 727 * m_copypack left space for our hdr; use it. 728 */ 729 m->m_len += hdrlen; 730 m->m_data -= hdrlen; 731 #else 732 #ifndef INET6 733 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 734 #else 735 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER, 736 M_PKTHDR, NULL); 737 #endif 738 if (m == NULL) { 739 error = ENOBUFS; 740 goto after_th; 741 } 742 m->m_data += max_linkhdr; 743 m->m_len = hdrlen; 744 if (len <= MHLEN - hdrlen - max_linkhdr) { 745 m_copydata(so->so_snd.ssb_mb, off, (int) len, 746 mtod(m, caddr_t) + hdrlen); 747 m->m_len += len; 748 } else { 749 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len); 750 if (m->m_next == NULL) { 751 m_free(m); 752 m = NULL; 753 error = ENOBUFS; 754 goto after_th; 755 } 756 } 757 #endif 758 /* 759 * If we're sending everything we've got, set PUSH. 760 * (This will keep happy those implementations which only 761 * give data to the user when a buffer fills or 762 * a PUSH comes in.) 763 */ 764 if (off + len == so->so_snd.ssb_cc) 765 flags |= TH_PUSH; 766 } else { 767 if (tp->t_flags & TF_ACKNOW) 768 tcpstat.tcps_sndacks++; 769 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 770 tcpstat.tcps_sndctrl++; 771 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 772 tcpstat.tcps_sndurg++; 773 else 774 tcpstat.tcps_sndwinup++; 775 776 MGETHDR(m, MB_DONTWAIT, MT_HEADER); 777 if (m == NULL) { 778 error = ENOBUFS; 779 goto after_th; 780 } 781 if (isipv6 && 782 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN) 783 MH_ALIGN(m, hdrlen); 784 else 785 m->m_data += max_linkhdr; 786 m->m_len = hdrlen; 787 } 788 m->m_pkthdr.rcvif = NULL; 789 if (isipv6) { 790 ip6 = mtod(m, struct ip6_hdr *); 791 th = (struct tcphdr *)(ip6 + 1); 792 tcp_fillheaders(tp, ip6, th); 793 } else { 794 ip = mtod(m, struct ip *); 795 ipov = (struct ipovly *)ip; 796 th = (struct tcphdr *)(ip + 1); 797 /* this picks up the pseudo header (w/o the length) */ 798 tcp_fillheaders(tp, ip, th); 799 } 800 after_th: 801 /* 802 * Fill in fields, remembering maximum advertised 803 * window for use in delaying messages about window sizes. 804 * If resending a FIN, be sure not to use a new sequence number. 805 */ 806 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 807 tp->snd_nxt == tp->snd_max) 808 tp->snd_nxt--; 809 810 if (th != NULL) { 811 /* 812 * If we are doing retransmissions, then snd_nxt will 813 * not reflect the first unsent octet. For ACK only 814 * packets, we do not want the sequence number of the 815 * retransmitted packet, we want the sequence number 816 * of the next unsent octet. So, if there is no data 817 * (and no SYN or FIN), use snd_max instead of snd_nxt 818 * when filling in ti_seq. But if we are in persist 819 * state, snd_max might reflect one byte beyond the 820 * right edge of the window, so use snd_nxt in that 821 * case, since we know we aren't doing a retransmission. 822 * (retransmit and persist are mutually exclusive...) 823 */ 824 if (len || (flags & (TH_SYN|TH_FIN)) || 825 tcp_callout_active(tp, tp->tt_persist)) 826 th->th_seq = htonl(tp->snd_nxt); 827 else 828 th->th_seq = htonl(tp->snd_max); 829 th->th_ack = htonl(tp->rcv_nxt); 830 if (optlen) { 831 bcopy(opt, th + 1, optlen); 832 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 833 } 834 th->th_flags = flags; 835 } 836 837 /* 838 * Calculate receive window. Don't shrink window, but avoid 839 * silly window syndrome by sending a 0 window if the actual 840 * window is less then one segment. 841 */ 842 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) && 843 recvwin < (long)tp->t_maxseg) 844 recvwin = 0; 845 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt)) 846 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt); 847 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale) 848 recvwin = (long)TCP_MAXWIN << tp->rcv_scale; 849 850 /* 851 * Adjust the RXWIN0SENT flag - indicate that we have advertised 852 * a 0 window. This may cause the remote transmitter to stall. This 853 * flag tells soreceive() to disable delayed acknowledgements when 854 * draining the buffer. This can occur if the receiver is attempting 855 * to read more data then can be buffered prior to transmitting on 856 * the connection. 857 */ 858 if (recvwin == 0) 859 tp->t_flags |= TF_RXWIN0SENT; 860 else 861 tp->t_flags &= ~TF_RXWIN0SENT; 862 863 if (th != NULL) 864 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale)); 865 866 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 867 if (th != NULL) { 868 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 869 th->th_flags |= TH_URG; 870 } 871 } else { 872 /* 873 * If no urgent pointer to send, then we pull 874 * the urgent pointer to the left edge of the send window 875 * so that it doesn't drift into the send window on sequence 876 * number wraparound. 877 */ 878 tp->snd_up = tp->snd_una; /* drag it along */ 879 } 880 881 if (th != NULL) { 882 #ifdef TCP_SIGNATURE 883 if (tp->t_flags & TF_SIGNATURE) { 884 tcpsignature_compute(m, len, optlen, 885 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 886 } 887 #endif /* TCP_SIGNATURE */ 888 889 /* 890 * Put TCP length in extended header, and then 891 * checksum extended header and data. 892 */ 893 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 894 if (isipv6) { 895 /* 896 * ip6_plen is not need to be filled now, and will be 897 * filled in ip6_output(). 898 */ 899 th->th_sum = in6_cksum(m, IPPROTO_TCP, 900 sizeof(struct ip6_hdr), 901 sizeof(struct tcphdr) + optlen + len); 902 } else { 903 m->m_pkthdr.csum_flags = CSUM_TCP; 904 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 905 if (len + optlen) { 906 th->th_sum = in_addword(th->th_sum, 907 htons((u_short)(optlen + len))); 908 } 909 910 /* 911 * IP version must be set here for ipv4/ipv6 checking 912 * later 913 */ 914 KASSERT(ip->ip_v == IPVERSION, 915 ("%s: IP version incorrect: %d", 916 __func__, ip->ip_v)); 917 } 918 } 919 920 /* 921 * In transmit state, time the transmission and arrange for 922 * the retransmit. In persist state, just set snd_max. 923 */ 924 if (!(tp->t_flags & TF_FORCE) || 925 !tcp_callout_active(tp, tp->tt_persist)) { 926 tcp_seq startseq = tp->snd_nxt; 927 928 /* 929 * Advance snd_nxt over sequence space of this segment. 930 */ 931 if (flags & (TH_SYN | TH_FIN)) { 932 if (flags & TH_SYN) 933 tp->snd_nxt++; 934 if (flags & TH_FIN) { 935 tp->snd_nxt++; 936 tp->t_flags |= TF_SENTFIN; 937 } 938 } 939 tp->snd_nxt += len; 940 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 941 tp->snd_max = tp->snd_nxt; 942 /* 943 * Time this transmission if not a retransmission and 944 * not currently timing anything. 945 */ 946 if (tp->t_rtttime == 0) { 947 tp->t_rtttime = ticks; 948 tp->t_rtseq = startseq; 949 tcpstat.tcps_segstimed++; 950 } 951 } 952 953 /* 954 * Set retransmit timer if not currently set, 955 * and not doing a pure ack or a keep-alive probe. 956 * Initial value for retransmit timer is smoothed 957 * round-trip time + 2 * round-trip time variance. 958 * Initialize shift counter which is used for backoff 959 * of retransmit time. 960 */ 961 if (!tcp_callout_active(tp, tp->tt_rexmt) && 962 tp->snd_nxt != tp->snd_una) { 963 if (tcp_callout_active(tp, tp->tt_persist)) { 964 tcp_callout_stop(tp, tp->tt_persist); 965 tp->t_rxtshift = 0; 966 } 967 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur, 968 tcp_timer_rexmt); 969 } 970 } else { 971 /* 972 * Persist case, update snd_max but since we are in 973 * persist mode (no window) we do not update snd_nxt. 974 */ 975 int xlen = len; 976 if (flags & TH_SYN) 977 panic("tcp_output: persist timer to send SYN\n"); 978 if (flags & TH_FIN) { 979 ++xlen; 980 tp->t_flags |= TF_SENTFIN; 981 } 982 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 983 tp->snd_max = tp->snd_nxt + xlen; 984 } 985 986 if (th != NULL) { 987 #ifdef TCPDEBUG 988 /* Trace. */ 989 if (so->so_options & SO_DEBUG) { 990 tcp_trace(TA_OUTPUT, tp->t_state, tp, 991 mtod(m, void *), th, 0); 992 } 993 #endif 994 995 /* 996 * Fill in IP length and desired time to live and 997 * send to IP level. There should be a better way 998 * to handle ttl and tos; we could keep them in 999 * the template, but need a way to checksum without them. 1000 */ 1001 /* 1002 * m->m_pkthdr.len should have been set before cksum 1003 * calcuration, because in6_cksum() need it. 1004 */ 1005 if (isipv6) { 1006 /* 1007 * we separately set hoplimit for every segment, 1008 * since the user might want to change the value 1009 * via setsockopt. Also, desired default hop 1010 * limit might be changed via Neighbor Discovery. 1011 */ 1012 ip6->ip6_hlim = in6_selecthlim(inp, 1013 (inp->in6p_route.ro_rt ? 1014 inp->in6p_route.ro_rt->rt_ifp : NULL)); 1015 1016 /* TODO: IPv6 IP6TOS_ECT bit on */ 1017 error = ip6_output(m, inp->in6p_outputopts, 1018 &inp->in6p_route, (so->so_options & SO_DONTROUTE), 1019 NULL, NULL, inp); 1020 } else { 1021 struct rtentry *rt; 1022 ip->ip_len = m->m_pkthdr.len; 1023 #ifdef INET6 1024 if (INP_CHECK_SOCKAF(so, AF_INET6)) 1025 ip->ip_ttl = in6_selecthlim(inp, 1026 (inp->in6p_route.ro_rt ? 1027 inp->in6p_route.ro_rt->rt_ifp : NULL)); 1028 else 1029 #endif 1030 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */ 1031 1032 ip->ip_tos = inp->inp_ip_tos; /* XXX */ 1033 /* 1034 * See if we should do MTU discovery. 1035 * We do it only if the following are true: 1036 * 1) we have a valid route to the destination 1037 * 2) the MTU is not locked (if it is, 1038 * then discovery has been disabled) 1039 */ 1040 if (path_mtu_discovery && 1041 (rt = inp->inp_route.ro_rt) && 1042 (rt->rt_flags & RTF_UP) && 1043 !(rt->rt_rmx.rmx_locks & RTV_MTU)) 1044 ip->ip_off |= IP_DF; 1045 1046 error = ip_output(m, inp->inp_options, &inp->inp_route, 1047 (so->so_options & SO_DONTROUTE) | 1048 IP_DEBUGROUTE, NULL, inp); 1049 } 1050 } else { 1051 KASSERT(error != 0, ("no error, but th not set\n")); 1052 } 1053 if (error) { 1054 1055 /* 1056 * We know that the packet was lost, so back out the 1057 * sequence number advance, if any. 1058 */ 1059 if (!(tp->t_flags & TF_FORCE) || 1060 !tcp_callout_active(tp, tp->tt_persist)) { 1061 /* 1062 * No need to check for TH_FIN here because 1063 * the TF_SENTFIN flag handles that case. 1064 */ 1065 if (!(flags & TH_SYN)) 1066 tp->snd_nxt -= len; 1067 } 1068 1069 out: 1070 if (error == ENOBUFS) { 1071 /* 1072 * If we can't send, make sure there is something 1073 * to get us going again later. 1074 * 1075 * The persist timer isn't necessarily allowed in all 1076 * states, use the rexmt timer. 1077 */ 1078 if (!tcp_callout_active(tp, tp->tt_rexmt) && 1079 !tcp_callout_active(tp, tp->tt_persist)) { 1080 tcp_callout_reset(tp, tp->tt_rexmt, 1081 tp->t_rxtcur, 1082 tcp_timer_rexmt); 1083 #if 0 1084 tp->t_rxtshift = 0; 1085 tcp_setpersist(tp); 1086 #endif 1087 } 1088 tcp_quench(inp, 0); 1089 return (0); 1090 } 1091 if (error == EMSGSIZE) { 1092 /* 1093 * ip_output() will have already fixed the route 1094 * for us. tcp_mtudisc() will, as its last action, 1095 * initiate retransmission, so it is important to 1096 * not do so here. 1097 */ 1098 tcp_mtudisc(inp, 0); 1099 return 0; 1100 } 1101 if ((error == EHOSTUNREACH || error == ENETDOWN) && 1102 TCPS_HAVERCVDSYN(tp->t_state)) { 1103 tp->t_softerror = error; 1104 return (0); 1105 } 1106 return (error); 1107 } 1108 tcpstat.tcps_sndtotal++; 1109 1110 /* 1111 * Data sent (as far as we can tell). 1112 * 1113 * If this advertises a larger window than any other segment, 1114 * then remember the size of the advertised window. 1115 * 1116 * Any pending ACK has now been sent. 1117 */ 1118 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) { 1119 tp->rcv_adv = tp->rcv_nxt + recvwin; 1120 tp->t_flags &= ~TF_RXRESIZED; 1121 } 1122 tp->last_ack_sent = tp->rcv_nxt; 1123 tp->t_flags &= ~TF_ACKNOW; 1124 if (tcp_delack_enabled) 1125 tcp_callout_stop(tp, tp->tt_delack); 1126 if (sendalot) 1127 goto again; 1128 return (0); 1129 } 1130 1131 void 1132 tcp_setpersist(struct tcpcb *tp) 1133 { 1134 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1135 int tt; 1136 1137 if (tp->t_state == TCPS_SYN_SENT || 1138 tp->t_state == TCPS_SYN_RECEIVED) { 1139 panic("tcp_setpersist: not established yet, current %s\n", 1140 tp->t_state == TCPS_SYN_SENT ? 1141 "SYN_SENT" : "SYN_RECEIVED"); 1142 } 1143 1144 if (tcp_callout_active(tp, tp->tt_rexmt)) 1145 panic("tcp_setpersist: retransmit pending"); 1146 /* 1147 * Start/restart persistance timer. 1148 */ 1149 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN, 1150 TCPTV_PERSMAX); 1151 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist); 1152 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1153 tp->t_rxtshift++; 1154 } 1155