1 /* 2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 3 * The Regents of the University of California. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. All advertising materials mentioning features or use of this software 14 * must display the following acknowledgement: 15 * This product includes software developed by the University of 16 * California, Berkeley and its contributors. 17 * 4. Neither the name of the University nor the names of its contributors 18 * may be used to endorse or promote products derived from this software 19 * without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 * 33 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 34 * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.20 2003/01/29 22:45:36 hsu Exp $ 35 * $DragonFly: src/sys/netinet/tcp_output.c,v 1.7 2003/09/18 18:32:55 hsu Exp $ 36 */ 37 38 #include "opt_inet6.h" 39 #include "opt_ipsec.h" 40 #include "opt_tcpdebug.h" 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/sysctl.h> 46 #include <sys/mbuf.h> 47 #include <sys/domain.h> 48 #include <sys/protosw.h> 49 #include <sys/socket.h> 50 #include <sys/socketvar.h> 51 52 #include <net/route.h> 53 54 #include <netinet/in.h> 55 #include <netinet/in_systm.h> 56 #include <netinet/ip.h> 57 #include <netinet/in_pcb.h> 58 #include <netinet/ip_var.h> 59 #ifdef INET6 60 #include <netinet6/in6_pcb.h> 61 #include <netinet/ip6.h> 62 #include <netinet6/ip6_var.h> 63 #endif 64 #include <netinet/tcp.h> 65 #define TCPOUTFLAGS 66 #include <netinet/tcp_fsm.h> 67 #include <netinet/tcp_seq.h> 68 #include <netinet/tcp_timer.h> 69 #include <netinet/tcp_var.h> 70 #include <netinet/tcpip.h> 71 #ifdef TCPDEBUG 72 #include <netinet/tcp_debug.h> 73 #endif 74 75 #ifdef IPSEC 76 #include <netinet6/ipsec.h> 77 #endif /*IPSEC*/ 78 79 #ifdef FAST_IPSEC 80 #include <netipsec/ipsec.h> 81 #define IPSEC 82 #endif /*FAST_IPSEC*/ 83 84 #include <machine/in_cksum.h> 85 86 #ifdef notyet 87 extern struct mbuf *m_copypack(); 88 #endif 89 90 int path_mtu_discovery = 1; 91 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 92 &path_mtu_discovery, 1, "Enable Path MTU Discovery"); 93 94 int ss_fltsz = 1; 95 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW, 96 &ss_fltsz, 1, "Slow start flight size"); 97 98 int ss_fltsz_local = 4; 99 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW, 100 &ss_fltsz_local, 1, "Slow start flight size for local networks"); 101 102 int tcp_do_newreno = 1; 103 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno, 104 0, "Enable NewReno Algorithms"); 105 106 /* 107 * Tcp output routine: figure out what should be sent and send it. 108 */ 109 int 110 tcp_output(tp) 111 struct tcpcb *tp; 112 { 113 struct socket *so = tp->t_inpcb->inp_socket; 114 long len, recvwin, sendwin; 115 int off, flags, error; 116 struct mbuf *m; 117 struct ip *ip = NULL; 118 struct ipovly *ipov = NULL; 119 #ifdef INET6 120 struct ip6_hdr *ip6 = NULL; 121 #endif /* INET6 */ 122 struct tcphdr *th; 123 u_char opt[TCP_MAXOLEN]; 124 unsigned ipoptlen, optlen, hdrlen; 125 int idle, sendalot; 126 #if 0 127 int maxburst = TCP_MAXBURST; 128 #endif 129 struct rmxp_tao *taop; 130 #ifdef INET6 131 int isipv6; 132 #endif 133 134 #ifdef INET6 135 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 136 #endif 137 138 /* 139 * Determine length of data that should be transmitted, 140 * and flags that will be used. 141 * If there is some data or critical controls (SYN, RST) 142 * to send, then transmit; otherwise, investigate further. 143 */ 144 if ((tp->snd_max == tp->snd_una) && 145 (ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 146 /* 147 * We have been idle for "a while" and no acks are 148 * expected to clock out any data we send -- 149 * slow start to get ack "clock" running again. 150 * 151 * Set the slow-start flight size depending on whether 152 * this is a local network or not. 153 */ 154 if ( 155 #ifdef INET6 156 (isipv6 && in6_localaddr(&tp->t_inpcb->in6p_faddr)) || 157 (!isipv6 && 158 #endif 159 in_localaddr(tp->t_inpcb->inp_faddr) 160 #ifdef INET6 161 ) 162 #endif 163 ) 164 tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local; 165 else 166 tp->snd_cwnd = tp->t_maxseg * ss_fltsz; 167 } 168 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 169 if (idle && (tp->t_flags & TF_MORETOCOME)) 170 tp->t_flags |= TF_LASTIDLE; 171 else 172 tp->t_flags &= ~TF_LASTIDLE; 173 again: 174 sendalot = 0; 175 off = tp->snd_nxt - tp->snd_una; 176 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 177 sendwin = min(sendwin, tp->snd_bwnd); 178 179 flags = tcp_outflags[tp->t_state]; 180 /* 181 * Get standard flags, and add SYN or FIN if requested by 'hidden' 182 * state flags. 183 */ 184 if (tp->t_flags & TF_NEEDFIN) 185 flags |= TH_FIN; 186 if (tp->t_flags & TF_NEEDSYN) 187 flags |= TH_SYN; 188 189 /* 190 * If in persist timeout with window of 0, send 1 byte. 191 * Otherwise, if window is small but nonzero 192 * and timer expired, we will send what we can 193 * and go to transmit state. 194 */ 195 if (tp->t_force) { 196 if (sendwin == 0) { 197 /* 198 * If we still have some data to send, then 199 * clear the FIN bit. Usually this would 200 * happen below when it realizes that we 201 * aren't sending all the data. However, 202 * if we have exactly 1 byte of unsent data, 203 * then it won't clear the FIN bit below, 204 * and if we are in persist state, we wind 205 * up sending the packet without recording 206 * that we sent the FIN bit. 207 * 208 * We can't just blindly clear the FIN bit, 209 * because if we don't have any more data 210 * to send then the probe will be the FIN 211 * itself. 212 */ 213 if (off < so->so_snd.sb_cc) 214 flags &= ~TH_FIN; 215 sendwin = 1; 216 } else { 217 callout_stop(tp->tt_persist); 218 tp->t_rxtshift = 0; 219 } 220 } 221 222 /* 223 * If snd_nxt == snd_max and we have transmitted a FIN, the 224 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 225 * a negative length. This can also occur when TCP opens up 226 * its congestion window while receiving additional duplicate 227 * acks after fast-retransmit because TCP will reset snd_nxt 228 * to snd_max after the fast-retransmit. 229 * 230 * In the normal retransmit-FIN-only case, however, snd_nxt will 231 * be set to snd_una, the offset will be 0, and the length may 232 * wind up 0. 233 */ 234 len = (long)ulmin(so->so_snd.sb_cc, sendwin) - off; 235 236 taop = tcp_gettaocache(&tp->t_inpcb->inp_inc); 237 238 /* 239 * Lop off SYN bit if it has already been sent. However, if this 240 * is SYN-SENT state and if segment contains data and if we don't 241 * know that foreign host supports TAO, suppress sending segment. 242 */ 243 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 244 flags &= ~TH_SYN; 245 off--, len++; 246 if (len > 0 && tp->t_state == TCPS_SYN_SENT && 247 (taop == NULL || taop->tao_ccsent == 0)) 248 return 0; 249 } 250 251 /* 252 * Be careful not to send data and/or FIN on SYN segments 253 * in cases when no CC option will be sent. 254 * This measure is needed to prevent interoperability problems 255 * with not fully conformant TCP implementations. 256 */ 257 if ((flags & TH_SYN) && 258 ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) || 259 ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) { 260 len = 0; 261 flags &= ~TH_FIN; 262 } 263 264 if (len < 0) { 265 /* 266 * If FIN has been sent but not acked, 267 * but we haven't been called to retransmit, 268 * len will be < 0. Otherwise, window shrank 269 * after we sent into it. If window shrank to 0, 270 * cancel pending retransmit, pull snd_nxt back 271 * to (closed) window, and set the persist timer 272 * if it isn't already going. If the window didn't 273 * close completely, just wait for an ACK. 274 */ 275 len = 0; 276 if (sendwin == 0) { 277 callout_stop(tp->tt_rexmt); 278 tp->t_rxtshift = 0; 279 tp->snd_nxt = tp->snd_una; 280 if (!callout_active(tp->tt_persist)) 281 tcp_setpersist(tp); 282 } 283 } 284 285 /* 286 * len will be >= 0 after this point. Truncate to the maximum 287 * segment length and ensure that FIN is removed if the length 288 * no longer contains the last data byte. 289 */ 290 if (len > tp->t_maxseg) { 291 len = tp->t_maxseg; 292 sendalot = 1; 293 } 294 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 295 flags &= ~TH_FIN; 296 297 recvwin = sbspace(&so->so_rcv); 298 299 /* 300 * Sender silly window avoidance. We transmit under the following 301 * conditions when len is non-zero: 302 * 303 * - We have a full segment 304 * - This is the last buffer in a write()/send() and we are 305 * either idle or running NODELAY 306 * - we've timed out (e.g. persist timer) 307 * - we have more then 1/2 the maximum send window's worth of 308 * data (receiver may be limited the window size) 309 * - we need to retransmit 310 */ 311 if (len) { 312 if (len == tp->t_maxseg) 313 goto send; 314 /* 315 * NOTE! on localhost connections an 'ack' from the remote 316 * end may occur synchronously with the output and cause 317 * us to flush a buffer queued with moretocome. XXX 318 * 319 * note: the len + off check is almost certainly unnecessary. 320 */ 321 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 322 (idle || (tp->t_flags & TF_NODELAY)) && 323 len + off >= so->so_snd.sb_cc && 324 (tp->t_flags & TF_NOPUSH) == 0) { 325 goto send; 326 } 327 if (tp->t_force) /* typ. timeout case */ 328 goto send; 329 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 330 goto send; 331 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 332 goto send; 333 } 334 335 /* 336 * Compare available window to amount of window 337 * known to peer (as advertised window less 338 * next expected input). If the difference is at least two 339 * max size segments, or at least 50% of the maximum possible 340 * window, then want to send a window update to peer. 341 */ 342 if (recvwin > 0) { 343 /* 344 * "adv" is the amount we can increase the window, 345 * taking into account that we are limited by 346 * TCP_MAXWIN << tp->rcv_scale. 347 */ 348 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) - 349 (tp->rcv_adv - tp->rcv_nxt); 350 351 if (adv >= (long) (2 * tp->t_maxseg)) 352 goto send; 353 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 354 goto send; 355 } 356 357 /* 358 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 359 * is also a catch-all for the retransmit timer timeout case. 360 */ 361 if (tp->t_flags & TF_ACKNOW) 362 goto send; 363 if ((flags & TH_RST) || 364 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 365 goto send; 366 if (SEQ_GT(tp->snd_up, tp->snd_una)) 367 goto send; 368 /* 369 * If our state indicates that FIN should be sent 370 * and we have not yet done so, then we need to send. 371 */ 372 if (flags & TH_FIN && 373 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 374 goto send; 375 376 /* 377 * TCP window updates are not reliable, rather a polling protocol 378 * using ``persist'' packets is used to insure receipt of window 379 * updates. The three ``states'' for the output side are: 380 * idle not doing retransmits or persists 381 * persisting to move a small or zero window 382 * (re)transmitting and thereby not persisting 383 * 384 * callout_active(tp->tt_persist) 385 * is true when we are in persist state. 386 * tp->t_force 387 * is set when we are called to send a persist packet. 388 * callout_active(tp->tt_rexmt) 389 * is set when we are retransmitting 390 * The output side is idle when both timers are zero. 391 * 392 * If send window is too small, there is data to transmit, and no 393 * retransmit or persist is pending, then go to persist state. 394 * If nothing happens soon, send when timer expires: 395 * if window is nonzero, transmit what we can, 396 * otherwise force out a byte. 397 */ 398 if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) && 399 !callout_active(tp->tt_persist)) { 400 tp->t_rxtshift = 0; 401 tcp_setpersist(tp); 402 } 403 404 /* 405 * No reason to send a segment, just return. 406 */ 407 return (0); 408 409 send: 410 /* 411 * Before ESTABLISHED, force sending of initial options 412 * unless TCP set not to do any options. 413 * NOTE: we assume that the IP/TCP header plus TCP options 414 * always fit in a single mbuf, leaving room for a maximum 415 * link header, i.e. 416 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 417 */ 418 optlen = 0; 419 #ifdef INET6 420 if (isipv6) 421 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 422 else 423 #endif 424 hdrlen = sizeof (struct tcpiphdr); 425 if (flags & TH_SYN) { 426 tp->snd_nxt = tp->iss; 427 if ((tp->t_flags & TF_NOOPT) == 0) { 428 u_short mss; 429 430 opt[0] = TCPOPT_MAXSEG; 431 opt[1] = TCPOLEN_MAXSEG; 432 mss = htons((u_short) tcp_mssopt(tp)); 433 (void)memcpy(opt + 2, &mss, sizeof(mss)); 434 optlen = TCPOLEN_MAXSEG; 435 436 if ((tp->t_flags & TF_REQ_SCALE) && 437 ((flags & TH_ACK) == 0 || 438 (tp->t_flags & TF_RCVD_SCALE))) { 439 *((u_int32_t *)(opt + optlen)) = htonl( 440 TCPOPT_NOP << 24 | 441 TCPOPT_WINDOW << 16 | 442 TCPOLEN_WINDOW << 8 | 443 tp->request_r_scale); 444 optlen += 4; 445 } 446 } 447 } 448 449 /* 450 * Send a timestamp and echo-reply if this is a SYN and our side 451 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 452 * and our peer have sent timestamps in our SYN's. 453 */ 454 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 455 (flags & TH_RST) == 0 && 456 ((flags & TH_ACK) == 0 || 457 (tp->t_flags & TF_RCVD_TSTMP))) { 458 u_int32_t *lp = (u_int32_t *)(opt + optlen); 459 460 /* Form timestamp option as shown in appendix A of RFC 1323. */ 461 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 462 *lp++ = htonl(ticks); 463 *lp = htonl(tp->ts_recent); 464 optlen += TCPOLEN_TSTAMP_APPA; 465 } 466 467 /* 468 * Send `CC-family' options if our side wants to use them (TF_REQ_CC), 469 * options are allowed (!TF_NOOPT) and it's not a RST. 470 */ 471 if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC && 472 (flags & TH_RST) == 0) { 473 switch (flags & (TH_SYN|TH_ACK)) { 474 /* 475 * This is a normal ACK, send CC if we received CC before 476 * from our peer. 477 */ 478 case TH_ACK: 479 if (!(tp->t_flags & TF_RCVD_CC)) 480 break; 481 /*FALLTHROUGH*/ 482 483 /* 484 * We can only get here in T/TCP's SYN_SENT* state, when 485 * we're a sending a non-SYN segment without waiting for 486 * the ACK of our SYN. A check above assures that we only 487 * do this if our peer understands T/TCP. 488 */ 489 case 0: 490 opt[optlen++] = TCPOPT_NOP; 491 opt[optlen++] = TCPOPT_NOP; 492 opt[optlen++] = TCPOPT_CC; 493 opt[optlen++] = TCPOLEN_CC; 494 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 495 496 optlen += 4; 497 break; 498 499 /* 500 * This is our initial SYN, check whether we have to use 501 * CC or CC.new. 502 */ 503 case TH_SYN: 504 opt[optlen++] = TCPOPT_NOP; 505 opt[optlen++] = TCPOPT_NOP; 506 opt[optlen++] = tp->t_flags & TF_SENDCCNEW ? 507 TCPOPT_CCNEW : TCPOPT_CC; 508 opt[optlen++] = TCPOLEN_CC; 509 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send); 510 optlen += 4; 511 break; 512 513 /* 514 * This is a SYN,ACK; send CC and CC.echo if we received 515 * CC from our peer. 516 */ 517 case (TH_SYN|TH_ACK): 518 if (tp->t_flags & TF_RCVD_CC) { 519 opt[optlen++] = TCPOPT_NOP; 520 opt[optlen++] = TCPOPT_NOP; 521 opt[optlen++] = TCPOPT_CC; 522 opt[optlen++] = TCPOLEN_CC; 523 *(u_int32_t *)&opt[optlen] = 524 htonl(tp->cc_send); 525 optlen += 4; 526 opt[optlen++] = TCPOPT_NOP; 527 opt[optlen++] = TCPOPT_NOP; 528 opt[optlen++] = TCPOPT_CCECHO; 529 opt[optlen++] = TCPOLEN_CC; 530 *(u_int32_t *)&opt[optlen] = 531 htonl(tp->cc_recv); 532 optlen += 4; 533 } 534 break; 535 } 536 } 537 538 hdrlen += optlen; 539 540 #ifdef INET6 541 if (isipv6) 542 ipoptlen = ip6_optlen(tp->t_inpcb); 543 else 544 #endif 545 { 546 if (tp->t_inpcb->inp_options) { 547 ipoptlen = tp->t_inpcb->inp_options->m_len - 548 offsetof(struct ipoption, ipopt_list); 549 } else { 550 ipoptlen = 0; 551 } 552 } 553 #ifdef IPSEC 554 ipoptlen += ipsec_hdrsiz_tcp(tp); 555 #endif 556 557 /* 558 * Adjust data length if insertion of options will 559 * bump the packet length beyond the t_maxopd length. 560 * Clear the FIN bit because we cut off the tail of 561 * the segment. 562 */ 563 if (len + optlen + ipoptlen > tp->t_maxopd) { 564 /* 565 * If there is still more to send, don't close the connection. 566 */ 567 flags &= ~TH_FIN; 568 len = tp->t_maxopd - optlen - ipoptlen; 569 sendalot = 1; 570 } 571 572 /*#ifdef DIAGNOSTIC*/ 573 #ifdef INET6 574 if (max_linkhdr + hdrlen > MCLBYTES) 575 panic("tcphdr too big"); 576 #else 577 if (max_linkhdr + hdrlen > MHLEN) 578 panic("tcphdr too big"); 579 #endif 580 /*#endif*/ 581 582 /* 583 * Grab a header mbuf, attaching a copy of data to 584 * be transmitted, and initialize the header from 585 * the template for sends on this connection. 586 */ 587 if (len) { 588 if (tp->t_force && len == 1) 589 tcpstat.tcps_sndprobe++; 590 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 591 tcpstat.tcps_sndrexmitpack++; 592 tcpstat.tcps_sndrexmitbyte += len; 593 } else { 594 tcpstat.tcps_sndpack++; 595 tcpstat.tcps_sndbyte += len; 596 } 597 #ifdef notyet 598 if ((m = m_copypack(so->so_snd.sb_mb, off, 599 (int)len, max_linkhdr + hdrlen)) == 0) { 600 error = ENOBUFS; 601 goto out; 602 } 603 /* 604 * m_copypack left space for our hdr; use it. 605 */ 606 m->m_len += hdrlen; 607 m->m_data -= hdrlen; 608 #else 609 MGETHDR(m, M_DONTWAIT, MT_HEADER); 610 if (m == NULL) { 611 error = ENOBUFS; 612 goto out; 613 } 614 #ifdef INET6 615 if (MHLEN < hdrlen + max_linkhdr) { 616 MCLGET(m, M_DONTWAIT); 617 if ((m->m_flags & M_EXT) == 0) { 618 m_freem(m); 619 error = ENOBUFS; 620 goto out; 621 } 622 } 623 #endif 624 m->m_data += max_linkhdr; 625 m->m_len = hdrlen; 626 if (len <= MHLEN - hdrlen - max_linkhdr) { 627 m_copydata(so->so_snd.sb_mb, off, (int) len, 628 mtod(m, caddr_t) + hdrlen); 629 m->m_len += len; 630 } else { 631 m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len); 632 if (m->m_next == 0) { 633 (void) m_free(m); 634 error = ENOBUFS; 635 goto out; 636 } 637 } 638 #endif 639 /* 640 * If we're sending everything we've got, set PUSH. 641 * (This will keep happy those implementations which only 642 * give data to the user when a buffer fills or 643 * a PUSH comes in.) 644 */ 645 if (off + len == so->so_snd.sb_cc) 646 flags |= TH_PUSH; 647 } else { 648 if (tp->t_flags & TF_ACKNOW) 649 tcpstat.tcps_sndacks++; 650 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 651 tcpstat.tcps_sndctrl++; 652 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 653 tcpstat.tcps_sndurg++; 654 else 655 tcpstat.tcps_sndwinup++; 656 657 MGETHDR(m, M_DONTWAIT, MT_HEADER); 658 if (m == NULL) { 659 error = ENOBUFS; 660 goto out; 661 } 662 #ifdef INET6 663 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 664 MHLEN >= hdrlen) { 665 MH_ALIGN(m, hdrlen); 666 } else 667 #endif 668 m->m_data += max_linkhdr; 669 m->m_len = hdrlen; 670 } 671 m->m_pkthdr.rcvif = (struct ifnet *)0; 672 #ifdef INET6 673 if (isipv6) { 674 ip6 = mtod(m, struct ip6_hdr *); 675 th = (struct tcphdr *)(ip6 + 1); 676 tcp_fillheaders(tp, ip6, th); 677 } else 678 #endif /* INET6 */ 679 { 680 ip = mtod(m, struct ip *); 681 ipov = (struct ipovly *)ip; 682 th = (struct tcphdr *)(ip + 1); 683 /* this picks up the pseudo header (w/o the length) */ 684 tcp_fillheaders(tp, ip, th); 685 } 686 687 /* 688 * Fill in fields, remembering maximum advertised 689 * window for use in delaying messages about window sizes. 690 * If resending a FIN, be sure not to use a new sequence number. 691 */ 692 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 693 tp->snd_nxt == tp->snd_max) 694 tp->snd_nxt--; 695 /* 696 * If we are doing retransmissions, then snd_nxt will 697 * not reflect the first unsent octet. For ACK only 698 * packets, we do not want the sequence number of the 699 * retransmitted packet, we want the sequence number 700 * of the next unsent octet. So, if there is no data 701 * (and no SYN or FIN), use snd_max instead of snd_nxt 702 * when filling in ti_seq. But if we are in persist 703 * state, snd_max might reflect one byte beyond the 704 * right edge of the window, so use snd_nxt in that 705 * case, since we know we aren't doing a retransmission. 706 * (retransmit and persist are mutually exclusive...) 707 */ 708 if (len || (flags & (TH_SYN|TH_FIN)) || callout_active(tp->tt_persist)) 709 th->th_seq = htonl(tp->snd_nxt); 710 else 711 th->th_seq = htonl(tp->snd_max); 712 th->th_ack = htonl(tp->rcv_nxt); 713 if (optlen) { 714 bcopy(opt, th + 1, optlen); 715 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 716 } 717 th->th_flags = flags; 718 /* 719 * Calculate receive window. Don't shrink window, 720 * but avoid silly window syndrome. 721 */ 722 if (recvwin < (long)(so->so_rcv.sb_hiwat / 4) && 723 recvwin < (long)tp->t_maxseg) 724 recvwin = 0; 725 if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 726 recvwin = (long)(tp->rcv_adv - tp->rcv_nxt); 727 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale) 728 recvwin = (long)TCP_MAXWIN << tp->rcv_scale; 729 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale)); 730 731 /* 732 * Adjust the RXWIN0SENT flag - indicate that we have advertised 733 * a 0 window. This may cause the remote transmitter to stall. This 734 * flag tells soreceive() to disable delayed acknowledgements when 735 * draining the buffer. This can occur if the receiver is attempting 736 * to read more data then can be buffered prior to transmitting on 737 * the connection. 738 */ 739 if (recvwin == 0) 740 tp->t_flags |= TF_RXWIN0SENT; 741 else 742 tp->t_flags &= ~TF_RXWIN0SENT; 743 744 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 745 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 746 th->th_flags |= TH_URG; 747 } else 748 /* 749 * If no urgent pointer to send, then we pull 750 * the urgent pointer to the left edge of the send window 751 * so that it doesn't drift into the send window on sequence 752 * number wraparound. 753 */ 754 tp->snd_up = tp->snd_una; /* drag it along */ 755 756 /* 757 * Put TCP length in extended header, and then 758 * checksum extended header and data. 759 */ 760 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 761 #ifdef INET6 762 if (isipv6) 763 /* 764 * ip6_plen is not need to be filled now, and will be filled 765 * in ip6_output. 766 */ 767 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 768 sizeof(struct tcphdr) + optlen + len); 769 else 770 #endif /* INET6 */ 771 { 772 m->m_pkthdr.csum_flags = CSUM_TCP; 773 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 774 if (len + optlen) 775 th->th_sum = in_addword(th->th_sum, 776 htons((u_short)(optlen + len))); 777 778 /* IP version must be set here for ipv4/ipv6 checking later */ 779 KASSERT(ip->ip_v == IPVERSION, 780 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 781 } 782 783 /* 784 * In transmit state, time the transmission and arrange for 785 * the retransmit. In persist state, just set snd_max. 786 */ 787 if (tp->t_force == 0 || !callout_active(tp->tt_persist)) { 788 tcp_seq startseq = tp->snd_nxt; 789 790 /* 791 * Advance snd_nxt over sequence space of this segment. 792 */ 793 if (flags & (TH_SYN|TH_FIN)) { 794 if (flags & TH_SYN) 795 tp->snd_nxt++; 796 if (flags & TH_FIN) { 797 tp->snd_nxt++; 798 tp->t_flags |= TF_SENTFIN; 799 } 800 } 801 tp->snd_nxt += len; 802 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 803 tp->snd_max = tp->snd_nxt; 804 /* 805 * Time this transmission if not a retransmission and 806 * not currently timing anything. 807 */ 808 if (tp->t_rtttime == 0) { 809 tp->t_rtttime = ticks; 810 tp->t_rtseq = startseq; 811 tcpstat.tcps_segstimed++; 812 } 813 } 814 815 /* 816 * Set retransmit timer if not currently set, 817 * and not doing a pure ack or a keep-alive probe. 818 * Initial value for retransmit timer is smoothed 819 * round-trip time + 2 * round-trip time variance. 820 * Initialize shift counter which is used for backoff 821 * of retransmit time. 822 */ 823 if (!callout_active(tp->tt_rexmt) && 824 tp->snd_nxt != tp->snd_una) { 825 if (callout_active(tp->tt_persist)) { 826 callout_stop(tp->tt_persist); 827 tp->t_rxtshift = 0; 828 } 829 callout_reset(tp->tt_rexmt, tp->t_rxtcur, 830 tcp_timer_rexmt, tp); 831 } 832 } else { 833 /* 834 * Persist case, update snd_max but since we are in 835 * persist mode (no window) we do not update snd_nxt. 836 */ 837 int xlen = len; 838 if (flags & TH_SYN) 839 ++xlen; 840 if (flags & TH_FIN) { 841 ++xlen; 842 tp->t_flags |= TF_SENTFIN; 843 } 844 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 845 tp->snd_max = tp->snd_nxt + xlen; 846 } 847 848 #ifdef TCPDEBUG 849 /* 850 * Trace. 851 */ 852 if (so->so_options & SO_DEBUG) 853 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 854 #endif 855 856 /* 857 * Fill in IP length and desired time to live and 858 * send to IP level. There should be a better way 859 * to handle ttl and tos; we could keep them in 860 * the template, but need a way to checksum without them. 861 */ 862 /* 863 * m->m_pkthdr.len should have been set before cksum calcuration, 864 * because in6_cksum() need it. 865 */ 866 #ifdef INET6 867 if (isipv6) { 868 /* 869 * we separately set hoplimit for every segment, since the 870 * user might want to change the value via setsockopt. 871 * Also, desired default hop limit might be changed via 872 * Neighbor Discovery. 873 */ 874 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, 875 tp->t_inpcb->in6p_route.ro_rt ? 876 tp->t_inpcb->in6p_route.ro_rt->rt_ifp 877 : NULL); 878 879 /* TODO: IPv6 IP6TOS_ECT bit on */ 880 error = ip6_output(m, 881 tp->t_inpcb->in6p_outputopts, 882 &tp->t_inpcb->in6p_route, 883 (so->so_options & SO_DONTROUTE), NULL, NULL, 884 tp->t_inpcb); 885 } else 886 #endif /* INET6 */ 887 { 888 struct rtentry *rt; 889 ip->ip_len = m->m_pkthdr.len; 890 #ifdef INET6 891 if (INP_CHECK_SOCKAF(so, AF_INET6)) 892 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, 893 tp->t_inpcb->in6p_route.ro_rt ? 894 tp->t_inpcb->in6p_route.ro_rt->rt_ifp 895 : NULL); 896 else 897 #endif /* INET6 */ 898 ip->ip_ttl = tp->t_inpcb->inp_ip_ttl; /* XXX */ 899 ip->ip_tos = tp->t_inpcb->inp_ip_tos; /* XXX */ 900 /* 901 * See if we should do MTU discovery. We do it only if the following 902 * are true: 903 * 1) we have a valid route to the destination 904 * 2) the MTU is not locked (if it is, then discovery has been 905 * disabled) 906 */ 907 if (path_mtu_discovery && (rt = tp->t_inpcb->inp_route.ro_rt) && 908 (rt->rt_flags & RTF_UP) && !(rt->rt_rmx.rmx_locks & RTV_MTU)) 909 ip->ip_off |= IP_DF; 910 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route, 911 (so->so_options & SO_DONTROUTE), 0, tp->t_inpcb); 912 } 913 if (error) { 914 915 /* 916 * We know that the packet was lost, so back out the 917 * sequence number advance, if any. 918 */ 919 if (tp->t_force == 0 || !callout_active(tp->tt_persist)) { 920 /* 921 * No need to check for TH_FIN here because 922 * the TF_SENTFIN flag handles that case. 923 */ 924 if ((flags & TH_SYN) == 0) 925 tp->snd_nxt -= len; 926 } 927 928 out: 929 if (error == ENOBUFS) { 930 /* 931 * If we can't send, make sure there is something 932 * to get us going again later. Persist state 933 * is not necessarily right, but it is close enough. 934 */ 935 if (!callout_active(tp->tt_rexmt) && 936 !callout_active(tp->tt_persist)) { 937 tp->t_rxtshift = 0; 938 tcp_setpersist(tp); 939 } 940 tcp_quench(tp->t_inpcb, 0); 941 return (0); 942 } 943 if (error == EMSGSIZE) { 944 /* 945 * ip_output() will have already fixed the route 946 * for us. tcp_mtudisc() will, as its last action, 947 * initiate retransmission, so it is important to 948 * not do so here. 949 */ 950 tcp_mtudisc(tp->t_inpcb, 0); 951 return 0; 952 } 953 if ((error == EHOSTUNREACH || error == ENETDOWN) 954 && TCPS_HAVERCVDSYN(tp->t_state)) { 955 tp->t_softerror = error; 956 return (0); 957 } 958 return (error); 959 } 960 tcpstat.tcps_sndtotal++; 961 962 /* 963 * Data sent (as far as we can tell). 964 * If this advertises a larger window than any other segment, 965 * then remember the size of the advertised window. 966 * Any pending ACK has now been sent. 967 */ 968 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) 969 tp->rcv_adv = tp->rcv_nxt + recvwin; 970 tp->last_ack_sent = tp->rcv_nxt; 971 tp->t_flags &= ~TF_ACKNOW; 972 if (tcp_delack_enabled) 973 callout_stop(tp->tt_delack); 974 #if 0 975 /* 976 * This completely breaks TCP if newreno is turned on. What happens 977 * is that if delayed-acks are turned on on the receiver, this code 978 * on the transmitter effectively destroys the TCP window, forcing 979 * it to four packets (1.5Kx4 = 6K window). 980 */ 981 if (sendalot && (!tcp_do_newreno || --maxburst)) 982 goto again; 983 #endif 984 if (sendalot) 985 goto again; 986 return (0); 987 } 988 989 void 990 tcp_setpersist(tp) 991 struct tcpcb *tp; 992 { 993 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 994 int tt; 995 996 if (callout_active(tp->tt_rexmt)) 997 panic("tcp_setpersist: retransmit pending"); 998 /* 999 * Start/restart persistance timer. 1000 */ 1001 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1002 TCPTV_PERSMIN, TCPTV_PERSMAX); 1003 callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp); 1004 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1005 tp->t_rxtshift++; 1006 } 1007