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 = NULL; 159 struct ip *ip = NULL; 160 struct ipovly *ipov = NULL; 161 struct tcphdr *th = NULL; 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 * A negative length can also occur when we are in the 282 * TCPS_SYN_RECEIVED state due to a simultanious connect where 283 * our SYN has not been acked yet. 284 * 285 * In the normal retransmit-FIN-only case, however, snd_nxt will 286 * be set to snd_una, the offset will be 0, and the length may 287 * wind up 0. 288 */ 289 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off; 290 291 /* 292 * Lop off SYN bit if it has already been sent. However, if this 293 * is SYN-SENT state and if segment contains data, suppress sending 294 * segment (sending the segment would be an option if we still 295 * did TAO and the remote host supported it). 296 */ 297 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 298 flags &= ~TH_SYN; 299 off--, len++; 300 if (len > 0 && tp->t_state == TCPS_SYN_SENT) 301 return 0; 302 } 303 304 /* 305 * Be careful not to send data and/or FIN on SYN segments. 306 * This measure is needed to prevent interoperability problems 307 * with not fully conformant TCP implementations. 308 */ 309 if (flags & TH_SYN) { 310 len = 0; 311 flags &= ~TH_FIN; 312 } 313 314 if (len < 0) { 315 /* 316 * A negative len can occur if our FIN has been sent but not 317 * acked, or if we are in a simultanious connect in the 318 * TCPS_SYN_RECEIVED state with our SYN sent but not yet 319 * acked. 320 * 321 * If our window has contracted to 0 in the FIN case 322 * (which can only occur if we have NOT been called to 323 * retransmit as per code a few paragraphs up) then we 324 * want to shift the retransmit timer over to the 325 * persist timer. 326 * 327 * However, if we are in the TCPS_SYN_RECEIVED state 328 * (the SYN case) we will be in a simultanious connect and 329 * the window may be zero degeneratively. In this case we 330 * do not want to shift to the persist timer after the SYN 331 * or the SYN+ACK transmission. 332 */ 333 len = 0; 334 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) { 335 tcp_callout_stop(tp, tp->tt_rexmt); 336 tp->t_rxtshift = 0; 337 tp->snd_nxt = tp->snd_una; 338 if (!tcp_callout_active(tp, tp->tt_persist)) 339 tcp_setpersist(tp); 340 } 341 } 342 343 KASSERT(len >= 0, ("%s: len < 0", __func__)); 344 /* 345 * Automatic sizing of send socket buffer. Often the send buffer 346 * size is not optimally adjusted to the actual network conditions 347 * at hand (delay bandwidth product). Setting the buffer size too 348 * small limits throughput on links with high bandwidth and high 349 * delay (eg. trans-continental/oceanic links). Setting the 350 * buffer size too big consumes too much real kernel memory, 351 * especially with many connections on busy servers. 352 * 353 * The criteria to step up the send buffer one notch are: 354 * 1. receive window of remote host is larger than send buffer 355 * (with a fudge factor of 5/4th); 356 * 2. send buffer is filled to 7/8th with data (so we actually 357 * have data to make use of it); 358 * 3. send buffer fill has not hit maximal automatic size; 359 * 4. our send window (slow start and cogestion controlled) is 360 * larger than sent but unacknowledged data in send buffer. 361 * 362 * The remote host receive window scaling factor may limit the 363 * growing of the send buffer before it reaches its allowed 364 * maximum. 365 * 366 * It scales directly with slow start or congestion window 367 * and does at most one step per received ACK. This fast 368 * scaling has the drawback of growing the send buffer beyond 369 * what is strictly necessary to make full use of a given 370 * delay*bandwith product. However testing has shown this not 371 * to be much of an problem. At worst we are trading wasting 372 * of available bandwith (the non-use of it) for wasting some 373 * socket buffer memory. 374 * 375 * TODO: Shrink send buffer during idle periods together 376 * with congestion window. Requires another timer. Has to 377 * wait for upcoming tcp timer rewrite. 378 */ 379 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) { 380 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat && 381 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) && 382 so->so_snd.ssb_cc < tcp_autosndbuf_max && 383 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) { 384 u_long newsize; 385 386 newsize = ulmin(so->so_snd.ssb_hiwat + 387 tcp_autosndbuf_inc, 388 tcp_autosndbuf_max); 389 if (!ssb_reserve(&so->so_snd, newsize, so, NULL)) 390 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE); 391 if (newsize >= (TCP_MAXWIN << tp->snd_scale)) 392 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE); 393 } 394 } 395 396 /* 397 * Truncate to the maximum segment length and ensure that FIN is 398 * removed if the length no longer contains the last data byte. 399 */ 400 if (len > tp->t_maxseg) { 401 len = tp->t_maxseg; 402 sendalot = TRUE; 403 } 404 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc)) 405 flags &= ~TH_FIN; 406 407 recvwin = ssb_space(&so->so_rcv); 408 409 /* 410 * Sender silly window avoidance. We transmit under the following 411 * conditions when len is non-zero: 412 * 413 * - We have a full segment 414 * - This is the last buffer in a write()/send() and we are 415 * either idle or running NODELAY 416 * - we've timed out (e.g. persist timer) 417 * - we have more then 1/2 the maximum send window's worth of 418 * data (receiver may be limiting the window size) 419 * - we need to retransmit 420 */ 421 if (len) { 422 if (len == tp->t_maxseg) 423 goto send; 424 /* 425 * NOTE! on localhost connections an 'ack' from the remote 426 * end may occur synchronously with the output and cause 427 * us to flush a buffer queued with moretocome. XXX 428 * 429 * note: the len + off check is almost certainly unnecessary. 430 */ 431 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 432 (idle || (tp->t_flags & TF_NODELAY)) && 433 len + off >= so->so_snd.ssb_cc && 434 !(tp->t_flags & TF_NOPUSH)) { 435 goto send; 436 } 437 if (tp->t_flags & TF_FORCE) /* typ. timeout case */ 438 goto send; 439 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 440 goto send; 441 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 442 goto send; 443 } 444 445 /* 446 * Compare available window to amount of window 447 * known to peer (as advertised window less 448 * next expected input). If the difference is at least two 449 * max size segments, or at least 50% of the maximum possible 450 * window, then want to send a window update to peer. 451 */ 452 if (recvwin > 0) { 453 /* 454 * "adv" is the amount we can increase the window, 455 * taking into account that we are limited by 456 * TCP_MAXWIN << tp->rcv_scale. 457 */ 458 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) - 459 (tp->rcv_adv - tp->rcv_nxt); 460 long hiwat; 461 462 /* 463 * This ack case typically occurs when the user has drained 464 * the TCP socket buffer sufficiently to warrent an ack 465 * containing a 'pure window update'... that is, an ack that 466 * ONLY updates the tcp window. 467 * 468 * It is unclear why we would need to do a pure window update 469 * past 2 segments if we are going to do one at 1/2 the high 470 * water mark anyway, especially since under normal conditions 471 * the user program will drain the socket buffer quickly. 472 * The 2-segment pure window update will often add a large 473 * number of extra, unnecessary acks to the stream. 474 * 475 * avoid_pure_win_update now defaults to 1. 476 */ 477 if (avoid_pure_win_update == 0 || 478 (tp->t_flags & TF_RXRESIZED)) { 479 if (adv >= (long) (2 * tp->t_maxseg)) { 480 goto send; 481 } 482 } 483 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale); 484 if (hiwat > (long)so->so_rcv.ssb_hiwat) 485 hiwat = (long)so->so_rcv.ssb_hiwat; 486 if (adv >= hiwat / 2) 487 goto send; 488 } 489 490 /* 491 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 492 * is also a catch-all for the retransmit timer timeout case. 493 */ 494 if (tp->t_flags & TF_ACKNOW) 495 goto send; 496 if ((flags & TH_RST) || 497 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN))) 498 goto send; 499 if (SEQ_GT(tp->snd_up, tp->snd_una)) 500 goto send; 501 /* 502 * If our state indicates that FIN should be sent 503 * and we have not yet done so, then we need to send. 504 */ 505 if ((flags & TH_FIN) && 506 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una)) 507 goto send; 508 509 /* 510 * TCP window updates are not reliable, rather a polling protocol 511 * using ``persist'' packets is used to insure receipt of window 512 * updates. The three ``states'' for the output side are: 513 * idle not doing retransmits or persists 514 * persisting to move a small or zero window 515 * (re)transmitting and thereby not persisting 516 * 517 * tcp_callout_active(tp, tp->tt_persist) 518 * is true when we are in persist state. 519 * The TF_FORCE flag in tp->t_flags 520 * is set when we are called to send a persist packet. 521 * tcp_callout_active(tp, tp->tt_rexmt) 522 * is set when we are retransmitting 523 * The output side is idle when both timers are zero. 524 * 525 * If send window is too small, there is data to transmit, and no 526 * retransmit or persist is pending, then go to persist state. 527 * 528 * If nothing happens soon, send when timer expires: 529 * if window is nonzero, transmit what we can, otherwise force out 530 * a byte. 531 * 532 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED 533 * with data pending. This situation can occur during a 534 * simultanious connect. 535 */ 536 if (so->so_snd.ssb_cc > 0 && 537 tp->t_state != TCPS_SYN_RECEIVED && 538 !tcp_callout_active(tp, tp->tt_rexmt) && 539 !tcp_callout_active(tp, tp->tt_persist)) { 540 tp->t_rxtshift = 0; 541 tcp_setpersist(tp); 542 } 543 544 /* 545 * No reason to send a segment, just return. 546 */ 547 return (0); 548 549 send: 550 /* 551 * Before ESTABLISHED, force sending of initial options 552 * unless TCP set not to do any options. 553 * NOTE: we assume that the IP/TCP header plus TCP options 554 * always fit in a single mbuf, leaving room for a maximum 555 * link header, i.e. 556 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES 557 */ 558 optlen = 0; 559 if (isipv6) 560 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 561 else 562 hdrlen = sizeof(struct tcpiphdr); 563 if (flags & TH_SYN) { 564 tp->snd_nxt = tp->iss; 565 if (!(tp->t_flags & TF_NOOPT)) { 566 u_short mss; 567 568 opt[0] = TCPOPT_MAXSEG; 569 opt[1] = TCPOLEN_MAXSEG; 570 mss = htons((u_short) tcp_mssopt(tp)); 571 memcpy(opt + 2, &mss, sizeof mss); 572 optlen = TCPOLEN_MAXSEG; 573 574 if ((tp->t_flags & TF_REQ_SCALE) && 575 (!(flags & TH_ACK) || 576 (tp->t_flags & TF_RCVD_SCALE))) { 577 *((u_int32_t *)(opt + optlen)) = htonl( 578 TCPOPT_NOP << 24 | 579 TCPOPT_WINDOW << 16 | 580 TCPOLEN_WINDOW << 8 | 581 tp->request_r_scale); 582 optlen += 4; 583 } 584 585 if ((tcp_do_sack && !(flags & TH_ACK)) || 586 tp->t_flags & TF_SACK_PERMITTED) { 587 uint32_t *lp = (uint32_t *)(opt + optlen); 588 589 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED); 590 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED; 591 } 592 } 593 } 594 595 /* 596 * Send a timestamp and echo-reply if this is a SYN and our side 597 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 598 * and our peer have sent timestamps in our SYN's. 599 */ 600 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 601 !(flags & TH_RST) && 602 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) { 603 u_int32_t *lp = (u_int32_t *)(opt + optlen); 604 605 /* Form timestamp option as shown in appendix A of RFC 1323. */ 606 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 607 *lp++ = htonl(ticks); 608 *lp = htonl(tp->ts_recent); 609 optlen += TCPOLEN_TSTAMP_APPA; 610 } 611 612 /* Set receive buffer autosizing timestamp. */ 613 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) 614 tp->rfbuf_ts = ticks; 615 616 /* 617 * If this is a SACK connection and we have a block to report, 618 * fill in the SACK blocks in the TCP options. 619 */ 620 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) == 621 TF_SACK_PERMITTED && 622 (!LIST_EMPTY(&tp->t_segq) || 623 tp->reportblk.rblk_start != tp->reportblk.rblk_end)) 624 tcp_sack_fill_report(tp, opt, &optlen); 625 626 #ifdef TCP_SIGNATURE 627 if (tp->t_flags & TF_SIGNATURE) { 628 int i; 629 u_char *bp; 630 /* 631 * Initialize TCP-MD5 option (RFC2385) 632 */ 633 bp = (u_char *)opt + optlen; 634 *bp++ = TCPOPT_SIGNATURE; 635 *bp++ = TCPOLEN_SIGNATURE; 636 sigoff = optlen + 2; 637 for (i = 0; i < TCP_SIGLEN; i++) 638 *bp++ = 0; 639 optlen += TCPOLEN_SIGNATURE; 640 /* 641 * Terminate options list and maintain 32-bit alignment. 642 */ 643 *bp++ = TCPOPT_NOP; 644 *bp++ = TCPOPT_EOL; 645 optlen += 2; 646 } 647 #endif /* TCP_SIGNATURE */ 648 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options")); 649 hdrlen += optlen; 650 651 if (isipv6) { 652 ipoptlen = ip6_optlen(inp); 653 } else { 654 if (inp->inp_options) { 655 ipoptlen = inp->inp_options->m_len - 656 offsetof(struct ipoption, ipopt_list); 657 } else { 658 ipoptlen = 0; 659 } 660 } 661 #ifdef IPSEC 662 ipoptlen += ipsec_hdrsiz_tcp(tp); 663 #endif 664 665 /* 666 * Adjust data length if insertion of options will bump the packet 667 * length beyond the t_maxopd length. Clear FIN to prevent premature 668 * closure since there is still more data to send after this (now 669 * truncated) packet. 670 * 671 * If just the options do not fit we are in a no-win situation and 672 * we treat it as an unreachable host. 673 */ 674 if (len + optlen + ipoptlen > tp->t_maxopd) { 675 if (tp->t_maxopd <= optlen + ipoptlen) { 676 static time_t last_optlen_report; 677 678 if (last_optlen_report != time_second) { 679 last_optlen_report = time_second; 680 kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd); 681 } 682 error = EHOSTUNREACH; 683 goto out; 684 } else { 685 flags &= ~TH_FIN; 686 len = tp->t_maxopd - optlen - ipoptlen; 687 sendalot = TRUE; 688 } 689 } 690 691 #ifdef INET6 692 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big")); 693 #else 694 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big")); 695 #endif 696 697 /* 698 * Grab a header mbuf, attaching a copy of data to 699 * be transmitted, and initialize the header from 700 * the template for sends on this connection. 701 */ 702 if (len) { 703 if ((tp->t_flags & TF_FORCE) && len == 1) 704 tcpstat.tcps_sndprobe++; 705 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 706 if (tp->snd_nxt == tp->snd_una) 707 tp->snd_max_rexmt = tp->snd_max; 708 tcpstat.tcps_sndrexmitpack++; 709 tcpstat.tcps_sndrexmitbyte += len; 710 } else { 711 tcpstat.tcps_sndpack++; 712 tcpstat.tcps_sndbyte += len; 713 } 714 #ifdef notyet 715 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len, 716 max_linkhdr + hdrlen)) == NULL) { 717 error = ENOBUFS; 718 goto after_th; 719 } 720 /* 721 * m_copypack left space for our hdr; use it. 722 */ 723 m->m_len += hdrlen; 724 m->m_data -= hdrlen; 725 #else 726 #ifndef INET6 727 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 728 #else 729 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER, 730 M_PKTHDR, NULL); 731 #endif 732 if (m == NULL) { 733 error = ENOBUFS; 734 goto after_th; 735 } 736 m->m_data += max_linkhdr; 737 m->m_len = hdrlen; 738 if (len <= MHLEN - hdrlen - max_linkhdr) { 739 m_copydata(so->so_snd.ssb_mb, off, (int) len, 740 mtod(m, caddr_t) + hdrlen); 741 m->m_len += len; 742 } else { 743 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len); 744 if (m->m_next == NULL) { 745 m_free(m); 746 m = NULL; 747 error = ENOBUFS; 748 goto after_th; 749 } 750 } 751 #endif 752 /* 753 * If we're sending everything we've got, set PUSH. 754 * (This will keep happy those implementations which only 755 * give data to the user when a buffer fills or 756 * a PUSH comes in.) 757 */ 758 if (off + len == so->so_snd.ssb_cc) 759 flags |= TH_PUSH; 760 } else { 761 if (tp->t_flags & TF_ACKNOW) 762 tcpstat.tcps_sndacks++; 763 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 764 tcpstat.tcps_sndctrl++; 765 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 766 tcpstat.tcps_sndurg++; 767 else 768 tcpstat.tcps_sndwinup++; 769 770 MGETHDR(m, MB_DONTWAIT, MT_HEADER); 771 if (m == NULL) { 772 error = ENOBUFS; 773 goto after_th; 774 } 775 if (isipv6 && 776 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN) 777 MH_ALIGN(m, hdrlen); 778 else 779 m->m_data += max_linkhdr; 780 m->m_len = hdrlen; 781 } 782 m->m_pkthdr.rcvif = NULL; 783 if (isipv6) { 784 ip6 = mtod(m, struct ip6_hdr *); 785 th = (struct tcphdr *)(ip6 + 1); 786 tcp_fillheaders(tp, ip6, th); 787 } else { 788 ip = mtod(m, struct ip *); 789 ipov = (struct ipovly *)ip; 790 th = (struct tcphdr *)(ip + 1); 791 /* this picks up the pseudo header (w/o the length) */ 792 tcp_fillheaders(tp, ip, th); 793 } 794 after_th: 795 /* 796 * Fill in fields, remembering maximum advertised 797 * window for use in delaying messages about window sizes. 798 * If resending a FIN, be sure not to use a new sequence number. 799 */ 800 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 801 tp->snd_nxt == tp->snd_max) 802 tp->snd_nxt--; 803 804 if (th != NULL) { 805 /* 806 * If we are doing retransmissions, then snd_nxt will 807 * not reflect the first unsent octet. For ACK only 808 * packets, we do not want the sequence number of the 809 * retransmitted packet, we want the sequence number 810 * of the next unsent octet. So, if there is no data 811 * (and no SYN or FIN), use snd_max instead of snd_nxt 812 * when filling in ti_seq. But if we are in persist 813 * state, snd_max might reflect one byte beyond the 814 * right edge of the window, so use snd_nxt in that 815 * case, since we know we aren't doing a retransmission. 816 * (retransmit and persist are mutually exclusive...) 817 */ 818 if (len || (flags & (TH_SYN|TH_FIN)) || 819 tcp_callout_active(tp, tp->tt_persist)) 820 th->th_seq = htonl(tp->snd_nxt); 821 else 822 th->th_seq = htonl(tp->snd_max); 823 th->th_ack = htonl(tp->rcv_nxt); 824 if (optlen) { 825 bcopy(opt, th + 1, optlen); 826 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 827 } 828 th->th_flags = flags; 829 } 830 831 /* 832 * Calculate receive window. Don't shrink window, but avoid 833 * silly window syndrome by sending a 0 window if the actual 834 * window is less then one segment. 835 */ 836 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) && 837 recvwin < (long)tp->t_maxseg) 838 recvwin = 0; 839 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt)) 840 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt); 841 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale) 842 recvwin = (long)TCP_MAXWIN << tp->rcv_scale; 843 844 /* 845 * Adjust the RXWIN0SENT flag - indicate that we have advertised 846 * a 0 window. This may cause the remote transmitter to stall. This 847 * flag tells soreceive() to disable delayed acknowledgements when 848 * draining the buffer. This can occur if the receiver is attempting 849 * to read more data then can be buffered prior to transmitting on 850 * the connection. 851 */ 852 if (recvwin == 0) 853 tp->t_flags |= TF_RXWIN0SENT; 854 else 855 tp->t_flags &= ~TF_RXWIN0SENT; 856 857 if (th != NULL) 858 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale)); 859 860 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 861 if (th != NULL) { 862 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 863 th->th_flags |= TH_URG; 864 } 865 } else { 866 /* 867 * If no urgent pointer to send, then we pull 868 * the urgent pointer to the left edge of the send window 869 * so that it doesn't drift into the send window on sequence 870 * number wraparound. 871 */ 872 tp->snd_up = tp->snd_una; /* drag it along */ 873 } 874 875 if (th != NULL) { 876 #ifdef TCP_SIGNATURE 877 if (tp->t_flags & TF_SIGNATURE) { 878 tcpsignature_compute(m, len, optlen, 879 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 880 } 881 #endif /* TCP_SIGNATURE */ 882 883 /* 884 * Put TCP length in extended header, and then 885 * checksum extended header and data. 886 */ 887 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 888 if (isipv6) { 889 /* 890 * ip6_plen is not need to be filled now, and will be 891 * filled in ip6_output(). 892 */ 893 th->th_sum = in6_cksum(m, IPPROTO_TCP, 894 sizeof(struct ip6_hdr), 895 sizeof(struct tcphdr) + optlen + len); 896 } else { 897 m->m_pkthdr.csum_flags = CSUM_TCP; 898 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 899 if (len + optlen) { 900 th->th_sum = in_addword(th->th_sum, 901 htons((u_short)(optlen + len))); 902 } 903 904 /* 905 * IP version must be set here for ipv4/ipv6 checking 906 * later 907 */ 908 KASSERT(ip->ip_v == IPVERSION, 909 ("%s: IP version incorrect: %d", 910 __func__, ip->ip_v)); 911 } 912 } 913 914 /* 915 * In transmit state, time the transmission and arrange for 916 * the retransmit. In persist state, just set snd_max. 917 */ 918 if (!(tp->t_flags & TF_FORCE) || 919 !tcp_callout_active(tp, tp->tt_persist)) { 920 tcp_seq startseq = tp->snd_nxt; 921 922 /* 923 * Advance snd_nxt over sequence space of this segment. 924 */ 925 if (flags & (TH_SYN | TH_FIN)) { 926 if (flags & TH_SYN) 927 tp->snd_nxt++; 928 if (flags & TH_FIN) { 929 tp->snd_nxt++; 930 tp->t_flags |= TF_SENTFIN; 931 } 932 } 933 tp->snd_nxt += len; 934 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 935 tp->snd_max = tp->snd_nxt; 936 /* 937 * Time this transmission if not a retransmission and 938 * not currently timing anything. 939 */ 940 if (tp->t_rtttime == 0) { 941 tp->t_rtttime = ticks; 942 tp->t_rtseq = startseq; 943 tcpstat.tcps_segstimed++; 944 } 945 } 946 947 /* 948 * Set retransmit timer if not currently set, 949 * and not doing a pure ack or a keep-alive probe. 950 * Initial value for retransmit timer is smoothed 951 * round-trip time + 2 * round-trip time variance. 952 * Initialize shift counter which is used for backoff 953 * of retransmit time. 954 */ 955 if (!tcp_callout_active(tp, tp->tt_rexmt) && 956 tp->snd_nxt != tp->snd_una) { 957 if (tcp_callout_active(tp, tp->tt_persist)) { 958 tcp_callout_stop(tp, tp->tt_persist); 959 tp->t_rxtshift = 0; 960 } 961 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur, 962 tcp_timer_rexmt); 963 } 964 } else { 965 /* 966 * Persist case, update snd_max but since we are in 967 * persist mode (no window) we do not update snd_nxt. 968 */ 969 int xlen = len; 970 if (flags & TH_SYN) 971 panic("tcp_output: persist timer to send SYN\n"); 972 if (flags & TH_FIN) { 973 ++xlen; 974 tp->t_flags |= TF_SENTFIN; 975 } 976 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 977 tp->snd_max = tp->snd_nxt + xlen; 978 } 979 980 if (th != NULL) { 981 #ifdef TCPDEBUG 982 /* Trace. */ 983 if (so->so_options & SO_DEBUG) { 984 tcp_trace(TA_OUTPUT, tp->t_state, tp, 985 mtod(m, void *), th, 0); 986 } 987 #endif 988 989 /* 990 * Fill in IP length and desired time to live and 991 * send to IP level. There should be a better way 992 * to handle ttl and tos; we could keep them in 993 * the template, but need a way to checksum without them. 994 */ 995 /* 996 * m->m_pkthdr.len should have been set before cksum 997 * calcuration, because in6_cksum() need it. 998 */ 999 if (isipv6) { 1000 /* 1001 * we separately set hoplimit for every segment, 1002 * since the user might want to change the value 1003 * via setsockopt. Also, desired default hop 1004 * limit might be changed via Neighbor Discovery. 1005 */ 1006 ip6->ip6_hlim = in6_selecthlim(inp, 1007 (inp->in6p_route.ro_rt ? 1008 inp->in6p_route.ro_rt->rt_ifp : NULL)); 1009 1010 /* TODO: IPv6 IP6TOS_ECT bit on */ 1011 error = ip6_output(m, inp->in6p_outputopts, 1012 &inp->in6p_route, (so->so_options & SO_DONTROUTE), 1013 NULL, NULL, inp); 1014 } else { 1015 struct rtentry *rt; 1016 ip->ip_len = m->m_pkthdr.len; 1017 #ifdef INET6 1018 if (INP_CHECK_SOCKAF(so, AF_INET6)) 1019 ip->ip_ttl = in6_selecthlim(inp, 1020 (inp->in6p_route.ro_rt ? 1021 inp->in6p_route.ro_rt->rt_ifp : NULL)); 1022 else 1023 #endif 1024 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */ 1025 1026 ip->ip_tos = inp->inp_ip_tos; /* XXX */ 1027 /* 1028 * See if we should do MTU discovery. 1029 * We do it only if the following are true: 1030 * 1) we have a valid route to the destination 1031 * 2) the MTU is not locked (if it is, 1032 * then discovery has been disabled) 1033 */ 1034 if (path_mtu_discovery && 1035 (rt = inp->inp_route.ro_rt) && 1036 (rt->rt_flags & RTF_UP) && 1037 !(rt->rt_rmx.rmx_locks & RTV_MTU)) 1038 ip->ip_off |= IP_DF; 1039 1040 error = ip_output(m, inp->inp_options, &inp->inp_route, 1041 (so->so_options & SO_DONTROUTE) | 1042 IP_DEBUGROUTE, NULL, inp); 1043 } 1044 } else { 1045 KASSERT(error != 0, ("no error, but th not set\n")); 1046 } 1047 if (error) { 1048 1049 /* 1050 * We know that the packet was lost, so back out the 1051 * sequence number advance, if any. 1052 */ 1053 if (!(tp->t_flags & TF_FORCE) || 1054 !tcp_callout_active(tp, tp->tt_persist)) { 1055 /* 1056 * No need to check for TH_FIN here because 1057 * the TF_SENTFIN flag handles that case. 1058 */ 1059 if (!(flags & TH_SYN)) 1060 tp->snd_nxt -= len; 1061 } 1062 1063 out: 1064 if (error == ENOBUFS) { 1065 /* 1066 * If we can't send, make sure there is something 1067 * to get us going again later. 1068 * 1069 * The persist timer isn't necessarily allowed in all 1070 * states, use the rexmt timer. 1071 */ 1072 if (!tcp_callout_active(tp, tp->tt_rexmt) && 1073 !tcp_callout_active(tp, tp->tt_persist)) { 1074 tcp_callout_reset(tp, tp->tt_rexmt, 1075 tp->t_rxtcur, 1076 tcp_timer_rexmt); 1077 #if 0 1078 tp->t_rxtshift = 0; 1079 tcp_setpersist(tp); 1080 #endif 1081 } 1082 tcp_quench(inp, 0); 1083 return (0); 1084 } 1085 if (error == EMSGSIZE) { 1086 /* 1087 * ip_output() will have already fixed the route 1088 * for us. tcp_mtudisc() will, as its last action, 1089 * initiate retransmission, so it is important to 1090 * not do so here. 1091 */ 1092 tcp_mtudisc(inp, 0); 1093 return 0; 1094 } 1095 if ((error == EHOSTUNREACH || error == ENETDOWN) && 1096 TCPS_HAVERCVDSYN(tp->t_state)) { 1097 tp->t_softerror = error; 1098 return (0); 1099 } 1100 return (error); 1101 } 1102 tcpstat.tcps_sndtotal++; 1103 1104 /* 1105 * Data sent (as far as we can tell). 1106 * 1107 * If this advertises a larger window than any other segment, 1108 * then remember the size of the advertised window. 1109 * 1110 * Any pending ACK has now been sent. 1111 */ 1112 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) { 1113 tp->rcv_adv = tp->rcv_nxt + recvwin; 1114 tp->t_flags &= ~TF_RXRESIZED; 1115 } 1116 tp->last_ack_sent = tp->rcv_nxt; 1117 tp->t_flags &= ~TF_ACKNOW; 1118 if (tcp_delack_enabled) 1119 tcp_callout_stop(tp, tp->tt_delack); 1120 if (sendalot) { 1121 th = NULL; 1122 goto again; 1123 } 1124 return (0); 1125 } 1126 1127 void 1128 tcp_setpersist(struct tcpcb *tp) 1129 { 1130 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1131 int tt; 1132 1133 if (tp->t_state == TCPS_SYN_SENT || 1134 tp->t_state == TCPS_SYN_RECEIVED) { 1135 panic("tcp_setpersist: not established yet, current %s\n", 1136 tp->t_state == TCPS_SYN_SENT ? 1137 "SYN_SENT" : "SYN_RECEIVED"); 1138 } 1139 1140 if (tcp_callout_active(tp, tp->tt_rexmt)) 1141 panic("tcp_setpersist: retransmit pending"); 1142 /* 1143 * Start/restart persistance timer. 1144 */ 1145 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN, 1146 TCPTV_PERSMAX); 1147 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist); 1148 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1149 tp->t_rxtshift++; 1150 } 1151