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 * 4. Neither the name of the University nor the names of its contributors 14 * may be used to endorse or promote products derived from this software 15 * without specific prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 * 29 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 30 * $FreeBSD$ 31 */ 32 33 #include "opt_inet.h" 34 #include "opt_inet6.h" 35 #include "opt_ipsec.h" 36 #include "opt_mac.h" 37 #include "opt_tcpdebug.h" 38 #include "opt_tcp_sack.h" 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/domain.h> 43 #include <sys/kernel.h> 44 #include <sys/lock.h> 45 #include <sys/mac.h> 46 #include <sys/mbuf.h> 47 #include <sys/mutex.h> 48 #include <sys/protosw.h> 49 #include <sys/socket.h> 50 #include <sys/socketvar.h> 51 #include <sys/sysctl.h> 52 53 #include <net/route.h> 54 55 #include <netinet/in.h> 56 #include <netinet/in_systm.h> 57 #include <netinet/ip.h> 58 #include <netinet/in_pcb.h> 59 #include <netinet/ip_var.h> 60 #include <netinet/ip_options.h> 61 #ifdef INET6 62 #include <netinet6/in6_pcb.h> 63 #include <netinet/ip6.h> 64 #include <netinet6/ip6_var.h> 65 #endif 66 #include <netinet/tcp.h> 67 #define TCPOUTFLAGS 68 #include <netinet/tcp_fsm.h> 69 #include <netinet/tcp_seq.h> 70 #include <netinet/tcp_timer.h> 71 #include <netinet/tcp_var.h> 72 #include <netinet/tcpip.h> 73 #ifdef TCPDEBUG 74 #include <netinet/tcp_debug.h> 75 #endif 76 77 #ifdef IPSEC 78 #include <netinet6/ipsec.h> 79 #endif /*IPSEC*/ 80 81 #ifdef FAST_IPSEC 82 #include <netipsec/ipsec.h> 83 #define IPSEC 84 #endif /*FAST_IPSEC*/ 85 86 #include <machine/in_cksum.h> 87 88 #ifdef notyet 89 extern struct mbuf *m_copypack(); 90 #endif 91 92 int path_mtu_discovery = 1; 93 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 94 &path_mtu_discovery, 1, "Enable Path MTU Discovery"); 95 96 int ss_fltsz = 1; 97 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW, 98 &ss_fltsz, 1, "Slow start flight size"); 99 100 int ss_fltsz_local = 4; 101 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW, 102 &ss_fltsz_local, 1, "Slow start flight size for local networks"); 103 104 int tcp_do_newreno = 1; 105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno, 106 0, "Enable NewReno Algorithms"); 107 108 /* 109 * Tcp output routine: figure out what should be sent and send it. 110 */ 111 int 112 tcp_output(struct tcpcb *tp) 113 { 114 struct socket *so = tp->t_inpcb->inp_socket; 115 long len, recwin, sendwin; 116 int off, flags, error; 117 #ifdef TCP_SIGNATURE 118 int sigoff = 0; 119 #endif 120 struct mbuf *m; 121 struct ip *ip = NULL; 122 struct ipovly *ipov = NULL; 123 struct tcphdr *th; 124 u_char opt[TCP_MAXOLEN]; 125 unsigned ipoptlen, optlen, hdrlen; 126 int idle, sendalot; 127 int i, sack_rxmit; 128 int sack_bytes_rxmt; 129 struct sackhole *p; 130 #if 0 131 int maxburst = TCP_MAXBURST; 132 #endif 133 #ifdef INET6 134 struct ip6_hdr *ip6 = NULL; 135 int isipv6; 136 137 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 138 #endif 139 140 INP_LOCK_ASSERT(tp->t_inpcb); 141 142 /* 143 * Determine length of data that should be transmitted, 144 * and flags that will be used. 145 * If there is some data or critical controls (SYN, RST) 146 * to send, then transmit; otherwise, investigate further. 147 */ 148 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 149 if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 150 /* 151 * We have been idle for "a while" and no acks are 152 * expected to clock out any data we send -- 153 * slow start to get ack "clock" running again. 154 * 155 * Set the slow-start flight size depending on whether 156 * this is a local network or not. 157 */ 158 int ss = ss_fltsz; 159 #ifdef INET6 160 if (isipv6) { 161 if (in6_localaddr(&tp->t_inpcb->in6p_faddr)) 162 ss = ss_fltsz_local; 163 } else 164 #endif /* INET6 */ 165 if (in_localaddr(tp->t_inpcb->inp_faddr)) 166 ss = ss_fltsz_local; 167 tp->snd_cwnd = tp->t_maxseg * ss; 168 } 169 tp->t_flags &= ~TF_LASTIDLE; 170 if (idle) { 171 if (tp->t_flags & TF_MORETOCOME) { 172 tp->t_flags |= TF_LASTIDLE; 173 idle = 0; 174 } 175 } 176 again: 177 /* 178 * If we've recently taken a timeout, snd_max will be greater than 179 * snd_nxt. There may be SACK information that allows us to avoid 180 * resending already delivered data. Adjust snd_nxt accordingly. 181 */ 182 if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max)) 183 tcp_sack_adjust(tp); 184 sendalot = 0; 185 off = tp->snd_nxt - tp->snd_una; 186 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 187 sendwin = min(sendwin, tp->snd_bwnd); 188 189 flags = tcp_outflags[tp->t_state]; 190 /* 191 * Send any SACK-generated retransmissions. If we're explicitly trying 192 * to send out new data (when sendalot is 1), bypass this function. 193 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 194 * we're replacing a (future) new transmission with a retransmission 195 * now, and we previously incremented snd_cwnd in tcp_input(). 196 */ 197 /* 198 * Still in sack recovery , reset rxmit flag to zero. 199 */ 200 sack_rxmit = 0; 201 sack_bytes_rxmt = 0; 202 len = 0; 203 p = NULL; 204 if (tp->sack_enable && IN_FASTRECOVERY(tp) && 205 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 206 long cwin; 207 208 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 209 if (cwin < 0) 210 cwin = 0; 211 /* Do not retransmit SACK segments beyond snd_recover */ 212 if (SEQ_GT(p->end, tp->snd_recover)) { 213 /* 214 * (At least) part of sack hole extends beyond 215 * snd_recover. Check to see if we can rexmit data 216 * for this hole. 217 */ 218 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 219 /* 220 * Can't rexmit any more data for this hole. 221 * That data will be rexmitted in the next 222 * sack recovery episode, when snd_recover 223 * moves past p->rxmit. 224 */ 225 p = NULL; 226 goto after_sack_rexmit; 227 } else 228 /* Can rexmit part of the current hole */ 229 len = ((long)ulmin(cwin, 230 tp->snd_recover - p->rxmit)); 231 } else 232 len = ((long)ulmin(cwin, p->end - p->rxmit)); 233 off = p->rxmit - tp->snd_una; 234 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 235 __func__, off)); 236 if (len > 0) { 237 sack_rxmit = 1; 238 sendalot = 1; 239 tcpstat.tcps_sack_rexmits++; 240 tcpstat.tcps_sack_rexmit_bytes += 241 min(len, tp->t_maxseg); 242 } 243 } 244 after_sack_rexmit: 245 /* 246 * Get standard flags, and add SYN or FIN if requested by 'hidden' 247 * state flags. 248 */ 249 if (tp->t_flags & TF_NEEDFIN) 250 flags |= TH_FIN; 251 if (tp->t_flags & TF_NEEDSYN) 252 flags |= TH_SYN; 253 254 SOCKBUF_LOCK(&so->so_snd); 255 /* 256 * If in persist timeout with window of 0, send 1 byte. 257 * Otherwise, if window is small but nonzero 258 * and timer expired, we will send what we can 259 * and go to transmit state. 260 */ 261 if (tp->t_flags & TF_FORCEDATA) { 262 if (sendwin == 0) { 263 /* 264 * If we still have some data to send, then 265 * clear the FIN bit. Usually this would 266 * happen below when it realizes that we 267 * aren't sending all the data. However, 268 * if we have exactly 1 byte of unsent data, 269 * then it won't clear the FIN bit below, 270 * and if we are in persist state, we wind 271 * up sending the packet without recording 272 * that we sent the FIN bit. 273 * 274 * We can't just blindly clear the FIN bit, 275 * because if we don't have any more data 276 * to send then the probe will be the FIN 277 * itself. 278 */ 279 if (off < so->so_snd.sb_cc) 280 flags &= ~TH_FIN; 281 sendwin = 1; 282 } else { 283 callout_stop(tp->tt_persist); 284 tp->t_rxtshift = 0; 285 } 286 } 287 288 /* 289 * If snd_nxt == snd_max and we have transmitted a FIN, the 290 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 291 * a negative length. This can also occur when TCP opens up 292 * its congestion window while receiving additional duplicate 293 * acks after fast-retransmit because TCP will reset snd_nxt 294 * to snd_max after the fast-retransmit. 295 * 296 * In the normal retransmit-FIN-only case, however, snd_nxt will 297 * be set to snd_una, the offset will be 0, and the length may 298 * wind up 0. 299 * 300 * If sack_rxmit is true we are retransmitting from the scoreboard 301 * in which case len is already set. 302 */ 303 if (sack_rxmit == 0) { 304 if (sack_bytes_rxmt == 0) 305 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off); 306 else { 307 long cwin; 308 309 /* 310 * We are inside of a SACK recovery episode and are 311 * sending new data, having retransmitted all the 312 * data possible in the scoreboard. 313 */ 314 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 315 - off); 316 /* 317 * Don't remove this (len > 0) check ! 318 * We explicitly check for len > 0 here (although it 319 * isn't really necessary), to work around a gcc 320 * optimization issue - to force gcc to compute 321 * len above. Without this check, the computation 322 * of len is bungled by the optimizer. 323 */ 324 if (len > 0) { 325 cwin = tp->snd_cwnd - 326 (tp->snd_nxt - tp->sack_newdata) - 327 sack_bytes_rxmt; 328 if (cwin < 0) 329 cwin = 0; 330 len = lmin(len, cwin); 331 } 332 } 333 } 334 335 /* 336 * Lop off SYN bit if it has already been sent. However, if this 337 * is SYN-SENT state and if segment contains data and if we don't 338 * know that foreign host supports TAO, suppress sending segment. 339 */ 340 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 341 flags &= ~TH_SYN; 342 off--, len++; 343 } 344 345 /* 346 * Be careful not to send data and/or FIN on SYN segments. 347 * This measure is needed to prevent interoperability problems 348 * with not fully conformant TCP implementations. 349 */ 350 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 351 len = 0; 352 flags &= ~TH_FIN; 353 } 354 355 if (len < 0) { 356 /* 357 * If FIN has been sent but not acked, 358 * but we haven't been called to retransmit, 359 * len will be < 0. Otherwise, window shrank 360 * after we sent into it. If window shrank to 0, 361 * cancel pending retransmit, pull snd_nxt back 362 * to (closed) window, and set the persist timer 363 * if it isn't already going. If the window didn't 364 * close completely, just wait for an ACK. 365 */ 366 len = 0; 367 if (sendwin == 0) { 368 callout_stop(tp->tt_rexmt); 369 tp->t_rxtshift = 0; 370 tp->snd_nxt = tp->snd_una; 371 if (!callout_active(tp->tt_persist)) 372 tcp_setpersist(tp); 373 } 374 } 375 376 /* 377 * len will be >= 0 after this point. Truncate to the maximum 378 * segment length and ensure that FIN is removed if the length 379 * no longer contains the last data byte. 380 */ 381 if (len > tp->t_maxseg) { 382 len = tp->t_maxseg; 383 sendalot = 1; 384 } 385 if (sack_rxmit) { 386 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 387 flags &= ~TH_FIN; 388 } else { 389 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 390 flags &= ~TH_FIN; 391 } 392 393 recwin = sbspace(&so->so_rcv); 394 395 /* 396 * Sender silly window avoidance. We transmit under the following 397 * conditions when len is non-zero: 398 * 399 * - We have a full segment 400 * - This is the last buffer in a write()/send() and we are 401 * either idle or running NODELAY 402 * - we've timed out (e.g. persist timer) 403 * - we have more then 1/2 the maximum send window's worth of 404 * data (receiver may be limited the window size) 405 * - we need to retransmit 406 */ 407 if (len) { 408 if (len == tp->t_maxseg) 409 goto send; 410 /* 411 * NOTE! on localhost connections an 'ack' from the remote 412 * end may occur synchronously with the output and cause 413 * us to flush a buffer queued with moretocome. XXX 414 * 415 * note: the len + off check is almost certainly unnecessary. 416 */ 417 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 418 (idle || (tp->t_flags & TF_NODELAY)) && 419 len + off >= so->so_snd.sb_cc && 420 (tp->t_flags & TF_NOPUSH) == 0) { 421 goto send; 422 } 423 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 424 goto send; 425 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 426 goto send; 427 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 428 goto send; 429 if (sack_rxmit) 430 goto send; 431 } 432 433 /* 434 * Compare available window to amount of window 435 * known to peer (as advertised window less 436 * next expected input). If the difference is at least two 437 * max size segments, or at least 50% of the maximum possible 438 * window, then want to send a window update to peer. 439 * Skip this if the connection is in T/TCP half-open state. 440 */ 441 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) { 442 /* 443 * "adv" is the amount we can increase the window, 444 * taking into account that we are limited by 445 * TCP_MAXWIN << tp->rcv_scale. 446 */ 447 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) - 448 (tp->rcv_adv - tp->rcv_nxt); 449 450 if (adv >= (long) (2 * tp->t_maxseg)) 451 goto send; 452 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 453 goto send; 454 } 455 456 /* 457 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 458 * is also a catch-all for the retransmit timer timeout case. 459 */ 460 if (tp->t_flags & TF_ACKNOW) 461 goto send; 462 if ((flags & TH_RST) || 463 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 464 goto send; 465 if (SEQ_GT(tp->snd_up, tp->snd_una)) 466 goto send; 467 /* 468 * If our state indicates that FIN should be sent 469 * and we have not yet done so, then we need to send. 470 */ 471 if (flags & TH_FIN && 472 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 473 goto send; 474 /* 475 * In SACK, it is possible for tcp_output to fail to send a segment 476 * after the retransmission timer has been turned off. Make sure 477 * that the retransmission timer is set. 478 */ 479 if (tp->sack_enable && SEQ_GT(tp->snd_max, tp->snd_una) && 480 !callout_active(tp->tt_rexmt) && 481 !callout_active(tp->tt_persist)) { 482 callout_reset(tp->tt_rexmt, tp->t_rxtcur, 483 tcp_timer_rexmt, tp); 484 goto just_return; 485 } 486 /* 487 * TCP window updates are not reliable, rather a polling protocol 488 * using ``persist'' packets is used to insure receipt of window 489 * updates. The three ``states'' for the output side are: 490 * idle not doing retransmits or persists 491 * persisting to move a small or zero window 492 * (re)transmitting and thereby not persisting 493 * 494 * callout_active(tp->tt_persist) 495 * is true when we are in persist state. 496 * (tp->t_flags & TF_FORCEDATA) 497 * is set when we are called to send a persist packet. 498 * callout_active(tp->tt_rexmt) 499 * is set when we are retransmitting 500 * The output side is idle when both timers are zero. 501 * 502 * If send window is too small, there is data to transmit, and no 503 * retransmit or persist is pending, then go to persist state. 504 * If nothing happens soon, send when timer expires: 505 * if window is nonzero, transmit what we can, 506 * otherwise force out a byte. 507 */ 508 if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) && 509 !callout_active(tp->tt_persist)) { 510 tp->t_rxtshift = 0; 511 tcp_setpersist(tp); 512 } 513 514 /* 515 * No reason to send a segment, just return. 516 */ 517 just_return: 518 SOCKBUF_UNLOCK(&so->so_snd); 519 return (0); 520 521 send: 522 SOCKBUF_LOCK_ASSERT(&so->so_snd); 523 /* 524 * Before ESTABLISHED, force sending of initial options 525 * unless TCP set not to do any options. 526 * NOTE: we assume that the IP/TCP header plus TCP options 527 * always fit in a single mbuf, leaving room for a maximum 528 * link header, i.e. 529 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 530 */ 531 optlen = 0; 532 #ifdef INET6 533 if (isipv6) 534 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 535 else 536 #endif 537 hdrlen = sizeof (struct tcpiphdr); 538 if (flags & TH_SYN) { 539 tp->snd_nxt = tp->iss; 540 if ((tp->t_flags & TF_NOOPT) == 0) { 541 u_short mss; 542 543 opt[0] = TCPOPT_MAXSEG; 544 opt[1] = TCPOLEN_MAXSEG; 545 mss = htons((u_short) tcp_mssopt(&tp->t_inpcb->inp_inc)); 546 (void)memcpy(opt + 2, &mss, sizeof(mss)); 547 optlen = TCPOLEN_MAXSEG; 548 549 if ((tp->t_flags & TF_REQ_SCALE) && 550 ((flags & TH_ACK) == 0 || 551 (tp->t_flags & TF_RCVD_SCALE))) { 552 *((u_int32_t *)(opt + optlen)) = htonl( 553 TCPOPT_NOP << 24 | 554 TCPOPT_WINDOW << 16 | 555 TCPOLEN_WINDOW << 8 | 556 tp->request_r_scale); 557 optlen += 4; 558 } 559 } 560 } 561 562 /* 563 * Send a timestamp and echo-reply if this is a SYN and our side 564 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 565 * and our peer have sent timestamps in our SYN's. 566 */ 567 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 568 (flags & TH_RST) == 0 && 569 ((flags & TH_ACK) == 0 || 570 (tp->t_flags & TF_RCVD_TSTMP))) { 571 u_int32_t *lp = (u_int32_t *)(opt + optlen); 572 573 /* Form timestamp option as shown in appendix A of RFC 1323. */ 574 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 575 *lp++ = htonl(ticks); 576 *lp = htonl(tp->ts_recent); 577 optlen += TCPOLEN_TSTAMP_APPA; 578 } 579 580 #ifdef TCP_SIGNATURE 581 #ifdef INET6 582 if (!isipv6) 583 #endif 584 if (tp->t_flags & TF_SIGNATURE) { 585 int i; 586 u_char *bp; 587 588 /* Initialize TCP-MD5 option (RFC2385) */ 589 bp = (u_char *)opt + optlen; 590 *bp++ = TCPOPT_SIGNATURE; 591 *bp++ = TCPOLEN_SIGNATURE; 592 sigoff = optlen + 2; 593 for (i = 0; i < TCP_SIGLEN; i++) 594 *bp++ = 0; 595 optlen += TCPOLEN_SIGNATURE; 596 } 597 #endif /* TCP_SIGNATURE */ 598 599 if (tp->sack_enable && ((tp->t_flags & TF_NOOPT) == 0)) { 600 /* 601 * Tack on the SACK permitted option *last*. 602 * And do padding of options after tacking this on. 603 * This is because of MSS, TS, WinScale and Signatures are 604 * all present, we have just 2 bytes left for the SACK 605 * permitted option, which is just enough. 606 */ 607 /* 608 * If this is the first SYN of connection (not a SYN 609 * ACK), include SACK permitted option. If this is a 610 * SYN ACK, include SACK permitted option if peer has 611 * already done so. This is only for active connect, 612 * since the syncache takes care of the passive connect. 613 */ 614 if ((flags & TH_SYN) && 615 (!(flags & TH_ACK) || (tp->t_flags & TF_SACK_PERMIT))) { 616 u_char *bp; 617 bp = (u_char *)opt + optlen; 618 619 *bp++ = TCPOPT_SACK_PERMITTED; 620 *bp++ = TCPOLEN_SACK_PERMITTED; 621 optlen += TCPOLEN_SACK_PERMITTED; 622 } 623 624 /* 625 * Send SACKs if necessary. This should be the last 626 * option processed. Only as many SACKs are sent as 627 * are permitted by the maximum options size. 628 * 629 * In general, SACK blocks consume 8*n+2 bytes. 630 * So a full size SACK blocks option is 34 bytes 631 * (to generate 4 SACK blocks). At a minimum, 632 * we need 10 bytes (to generate 1 SACK block). 633 * If TCP Timestamps (12 bytes) and TCP Signatures 634 * (18 bytes) are both present, we'll just have 635 * 10 bytes for SACK options 40 - (12 + 18). 636 */ 637 if (TCPS_HAVEESTABLISHED(tp->t_state) && 638 (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0 && 639 MAX_TCPOPTLEN - optlen - 2 >= TCPOLEN_SACK) { 640 int nsack, sackoptlen, padlen; 641 u_char *bp = (u_char *)opt + optlen; 642 u_int32_t *lp; 643 644 nsack = (MAX_TCPOPTLEN - optlen - 2) / TCPOLEN_SACK; 645 nsack = min(nsack, tp->rcv_numsacks); 646 sackoptlen = (2 + nsack * TCPOLEN_SACK); 647 648 /* 649 * First we need to pad options so that the 650 * SACK blocks can start at a 4-byte boundary 651 * (sack option and length are at a 2 byte offset). 652 */ 653 padlen = (MAX_TCPOPTLEN - optlen - sackoptlen) % 4; 654 optlen += padlen; 655 while (padlen-- > 0) 656 *bp++ = TCPOPT_NOP; 657 658 tcpstat.tcps_sack_send_blocks++; 659 *bp++ = TCPOPT_SACK; 660 *bp++ = sackoptlen; 661 lp = (u_int32_t *)bp; 662 for (i = 0; i < nsack; i++) { 663 struct sackblk sack = tp->sackblks[i]; 664 *lp++ = htonl(sack.start); 665 *lp++ = htonl(sack.end); 666 } 667 optlen += sackoptlen; 668 } 669 } 670 671 /* Pad TCP options to a 4 byte boundary */ 672 if (optlen < MAX_TCPOPTLEN && (optlen % sizeof(u_int32_t))) { 673 int pad = sizeof(u_int32_t) - (optlen % sizeof(u_int32_t)); 674 u_char *bp = (u_char *)opt + optlen; 675 676 optlen += pad; 677 while (pad) { 678 *bp++ = TCPOPT_EOL; 679 pad--; 680 } 681 } 682 683 hdrlen += optlen; 684 685 #ifdef INET6 686 if (isipv6) 687 ipoptlen = ip6_optlen(tp->t_inpcb); 688 else 689 #endif 690 if (tp->t_inpcb->inp_options) 691 ipoptlen = tp->t_inpcb->inp_options->m_len - 692 offsetof(struct ipoption, ipopt_list); 693 else 694 ipoptlen = 0; 695 #ifdef IPSEC 696 ipoptlen += ipsec_hdrsiz_tcp(tp); 697 #endif 698 699 /* 700 * Adjust data length if insertion of options will 701 * bump the packet length beyond the t_maxopd length. 702 * Clear the FIN bit because we cut off the tail of 703 * the segment. 704 */ 705 if (len + optlen + ipoptlen > tp->t_maxopd) { 706 /* 707 * If there is still more to send, don't close the connection. 708 */ 709 flags &= ~TH_FIN; 710 len = tp->t_maxopd - optlen - ipoptlen; 711 sendalot = 1; 712 } 713 714 /*#ifdef DIAGNOSTIC*/ 715 #ifdef INET6 716 if (max_linkhdr + hdrlen > MCLBYTES) 717 #else 718 if (max_linkhdr + hdrlen > MHLEN) 719 #endif 720 panic("tcphdr too big"); 721 /*#endif*/ 722 723 /* 724 * Grab a header mbuf, attaching a copy of data to 725 * be transmitted, and initialize the header from 726 * the template for sends on this connection. 727 */ 728 if (len) { 729 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 730 tcpstat.tcps_sndprobe++; 731 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 732 tcpstat.tcps_sndrexmitpack++; 733 tcpstat.tcps_sndrexmitbyte += len; 734 } else { 735 tcpstat.tcps_sndpack++; 736 tcpstat.tcps_sndbyte += len; 737 } 738 #ifdef notyet 739 if ((m = m_copypack(so->so_snd.sb_mb, off, 740 (int)len, max_linkhdr + hdrlen)) == 0) { 741 SOCKBUF_UNLOCK(&so->so_snd); 742 error = ENOBUFS; 743 goto out; 744 } 745 /* 746 * m_copypack left space for our hdr; use it. 747 */ 748 m->m_len += hdrlen; 749 m->m_data -= hdrlen; 750 #else 751 MGETHDR(m, M_DONTWAIT, MT_DATA); 752 if (m == NULL) { 753 SOCKBUF_UNLOCK(&so->so_snd); 754 error = ENOBUFS; 755 goto out; 756 } 757 #ifdef INET6 758 if (MHLEN < hdrlen + max_linkhdr) { 759 MCLGET(m, M_DONTWAIT); 760 if ((m->m_flags & M_EXT) == 0) { 761 SOCKBUF_UNLOCK(&so->so_snd); 762 m_freem(m); 763 error = ENOBUFS; 764 goto out; 765 } 766 } 767 #endif 768 m->m_data += max_linkhdr; 769 m->m_len = hdrlen; 770 if (len <= MHLEN - hdrlen - max_linkhdr) { 771 m_copydata(so->so_snd.sb_mb, off, (int) len, 772 mtod(m, caddr_t) + hdrlen); 773 m->m_len += len; 774 } else { 775 m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len); 776 if (m->m_next == 0) { 777 SOCKBUF_UNLOCK(&so->so_snd); 778 (void) m_free(m); 779 error = ENOBUFS; 780 goto out; 781 } 782 } 783 #endif 784 /* 785 * If we're sending everything we've got, set PUSH. 786 * (This will keep happy those implementations which only 787 * give data to the user when a buffer fills or 788 * a PUSH comes in.) 789 */ 790 if (off + len == so->so_snd.sb_cc) 791 flags |= TH_PUSH; 792 SOCKBUF_UNLOCK(&so->so_snd); 793 } else { 794 SOCKBUF_UNLOCK(&so->so_snd); 795 if (tp->t_flags & TF_ACKNOW) 796 tcpstat.tcps_sndacks++; 797 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 798 tcpstat.tcps_sndctrl++; 799 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 800 tcpstat.tcps_sndurg++; 801 else 802 tcpstat.tcps_sndwinup++; 803 804 MGETHDR(m, M_DONTWAIT, MT_DATA); 805 if (m == NULL) { 806 error = ENOBUFS; 807 goto out; 808 } 809 #ifdef INET6 810 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 811 MHLEN >= hdrlen) { 812 MH_ALIGN(m, hdrlen); 813 } else 814 #endif 815 m->m_data += max_linkhdr; 816 m->m_len = hdrlen; 817 } 818 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 819 m->m_pkthdr.rcvif = (struct ifnet *)0; 820 #ifdef MAC 821 mac_create_mbuf_from_inpcb(tp->t_inpcb, m); 822 #endif 823 #ifdef INET6 824 if (isipv6) { 825 ip6 = mtod(m, struct ip6_hdr *); 826 th = (struct tcphdr *)(ip6 + 1); 827 tcpip_fillheaders(tp->t_inpcb, ip6, th); 828 } else 829 #endif /* INET6 */ 830 { 831 ip = mtod(m, struct ip *); 832 ipov = (struct ipovly *)ip; 833 th = (struct tcphdr *)(ip + 1); 834 tcpip_fillheaders(tp->t_inpcb, ip, th); 835 } 836 837 /* 838 * Fill in fields, remembering maximum advertised 839 * window for use in delaying messages about window sizes. 840 * If resending a FIN, be sure not to use a new sequence number. 841 */ 842 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 843 tp->snd_nxt == tp->snd_max) 844 tp->snd_nxt--; 845 /* 846 * If we are doing retransmissions, then snd_nxt will 847 * not reflect the first unsent octet. For ACK only 848 * packets, we do not want the sequence number of the 849 * retransmitted packet, we want the sequence number 850 * of the next unsent octet. So, if there is no data 851 * (and no SYN or FIN), use snd_max instead of snd_nxt 852 * when filling in ti_seq. But if we are in persist 853 * state, snd_max might reflect one byte beyond the 854 * right edge of the window, so use snd_nxt in that 855 * case, since we know we aren't doing a retransmission. 856 * (retransmit and persist are mutually exclusive...) 857 */ 858 if (sack_rxmit == 0) { 859 if (len || (flags & (TH_SYN|TH_FIN)) 860 || callout_active(tp->tt_persist)) 861 th->th_seq = htonl(tp->snd_nxt); 862 else 863 th->th_seq = htonl(tp->snd_max); 864 } else { 865 th->th_seq = htonl(p->rxmit); 866 p->rxmit += len; 867 tp->sackhint.sack_bytes_rexmit += len; 868 } 869 th->th_ack = htonl(tp->rcv_nxt); 870 if (optlen) { 871 bcopy(opt, th + 1, optlen); 872 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 873 } 874 th->th_flags = flags; 875 /* 876 * Calculate receive window. Don't shrink window, 877 * but avoid silly window syndrome. 878 */ 879 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 880 recwin < (long)tp->t_maxseg) 881 recwin = 0; 882 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 883 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 884 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 885 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 886 th->th_win = htons((u_short) (recwin >> tp->rcv_scale)); 887 888 889 /* 890 * Adjust the RXWIN0SENT flag - indicate that we have advertised 891 * a 0 window. This may cause the remote transmitter to stall. This 892 * flag tells soreceive() to disable delayed acknowledgements when 893 * draining the buffer. This can occur if the receiver is attempting 894 * to read more data then can be buffered prior to transmitting on 895 * the connection. 896 */ 897 if (recwin == 0) 898 tp->t_flags |= TF_RXWIN0SENT; 899 else 900 tp->t_flags &= ~TF_RXWIN0SENT; 901 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 902 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 903 th->th_flags |= TH_URG; 904 } else 905 /* 906 * If no urgent pointer to send, then we pull 907 * the urgent pointer to the left edge of the send window 908 * so that it doesn't drift into the send window on sequence 909 * number wraparound. 910 */ 911 tp->snd_up = tp->snd_una; /* drag it along */ 912 913 #ifdef TCP_SIGNATURE 914 #ifdef INET6 915 if (!isipv6) 916 #endif 917 if (tp->t_flags & TF_SIGNATURE) 918 tcp_signature_compute(m, sizeof(struct ip), len, optlen, 919 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 920 #endif 921 922 /* 923 * Put TCP length in extended header, and then 924 * checksum extended header and data. 925 */ 926 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 927 #ifdef INET6 928 if (isipv6) 929 /* 930 * ip6_plen is not need to be filled now, and will be filled 931 * in ip6_output. 932 */ 933 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 934 sizeof(struct tcphdr) + optlen + len); 935 else 936 #endif /* INET6 */ 937 { 938 m->m_pkthdr.csum_flags = CSUM_TCP; 939 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 940 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 941 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 942 943 /* IP version must be set here for ipv4/ipv6 checking later */ 944 KASSERT(ip->ip_v == IPVERSION, 945 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 946 } 947 948 /* 949 * In transmit state, time the transmission and arrange for 950 * the retransmit. In persist state, just set snd_max. 951 */ 952 if ((tp->t_flags & TF_FORCEDATA) == 0 || 953 !callout_active(tp->tt_persist)) { 954 tcp_seq startseq = tp->snd_nxt; 955 956 /* 957 * Advance snd_nxt over sequence space of this segment. 958 */ 959 if (flags & (TH_SYN|TH_FIN)) { 960 if (flags & TH_SYN) 961 tp->snd_nxt++; 962 if (flags & TH_FIN) { 963 tp->snd_nxt++; 964 tp->t_flags |= TF_SENTFIN; 965 } 966 } 967 if (sack_rxmit) 968 goto timer; 969 tp->snd_nxt += len; 970 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 971 tp->snd_max = tp->snd_nxt; 972 /* 973 * Time this transmission if not a retransmission and 974 * not currently timing anything. 975 */ 976 if (tp->t_rtttime == 0) { 977 tp->t_rtttime = ticks; 978 tp->t_rtseq = startseq; 979 tcpstat.tcps_segstimed++; 980 } 981 } 982 983 /* 984 * Set retransmit timer if not currently set, 985 * and not doing a pure ack or a keep-alive probe. 986 * Initial value for retransmit timer is smoothed 987 * round-trip time + 2 * round-trip time variance. 988 * Initialize shift counter which is used for backoff 989 * of retransmit time. 990 */ 991 timer: 992 if (!callout_active(tp->tt_rexmt) && 993 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 994 (tp->snd_nxt != tp->snd_una))) { 995 if (callout_active(tp->tt_persist)) { 996 callout_stop(tp->tt_persist); 997 tp->t_rxtshift = 0; 998 } 999 callout_reset(tp->tt_rexmt, tp->t_rxtcur, 1000 tcp_timer_rexmt, tp); 1001 } 1002 } else { 1003 /* 1004 * Persist case, update snd_max but since we are in 1005 * persist mode (no window) we do not update snd_nxt. 1006 */ 1007 int xlen = len; 1008 if (flags & TH_SYN) 1009 ++xlen; 1010 if (flags & TH_FIN) { 1011 ++xlen; 1012 tp->t_flags |= TF_SENTFIN; 1013 } 1014 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1015 tp->snd_max = tp->snd_nxt + len; 1016 } 1017 1018 #ifdef TCPDEBUG 1019 /* 1020 * Trace. 1021 */ 1022 if (so->so_options & SO_DEBUG) { 1023 u_short save = 0; 1024 #ifdef INET6 1025 if (!isipv6) 1026 #endif 1027 { 1028 save = ipov->ih_len; 1029 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1030 } 1031 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1032 #ifdef INET6 1033 if (!isipv6) 1034 #endif 1035 ipov->ih_len = save; 1036 } 1037 #endif 1038 1039 /* 1040 * Fill in IP length and desired time to live and 1041 * send to IP level. There should be a better way 1042 * to handle ttl and tos; we could keep them in 1043 * the template, but need a way to checksum without them. 1044 */ 1045 /* 1046 * m->m_pkthdr.len should have been set before cksum calcuration, 1047 * because in6_cksum() need it. 1048 */ 1049 #ifdef INET6 1050 if (isipv6) { 1051 /* 1052 * we separately set hoplimit for every segment, since the 1053 * user might want to change the value via setsockopt. 1054 * Also, desired default hop limit might be changed via 1055 * Neighbor Discovery. 1056 */ 1057 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1058 1059 /* TODO: IPv6 IP6TOS_ECT bit on */ 1060 error = ip6_output(m, 1061 tp->t_inpcb->in6p_outputopts, NULL, 1062 ((so->so_options & SO_DONTROUTE) ? 1063 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb); 1064 } else 1065 #endif /* INET6 */ 1066 { 1067 ip->ip_len = m->m_pkthdr.len; 1068 #ifdef INET6 1069 if (INP_CHECK_SOCKAF(so, AF_INET6)) 1070 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1071 #endif /* INET6 */ 1072 /* 1073 * If we do path MTU discovery, then we set DF on every packet. 1074 * This might not be the best thing to do according to RFC3390 1075 * Section 2. However the tcp hostcache migitates the problem 1076 * so it affects only the first tcp connection with a host. 1077 */ 1078 if (path_mtu_discovery) 1079 ip->ip_off |= IP_DF; 1080 1081 error = ip_output(m, tp->t_inpcb->inp_options, NULL, 1082 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1083 tp->t_inpcb); 1084 } 1085 if (error) { 1086 1087 /* 1088 * We know that the packet was lost, so back out the 1089 * sequence number advance, if any. 1090 */ 1091 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1092 !callout_active(tp->tt_persist)) { 1093 /* 1094 * No need to check for TH_FIN here because 1095 * the TF_SENTFIN flag handles that case. 1096 */ 1097 if ((flags & TH_SYN) == 0) { 1098 if (sack_rxmit) { 1099 p->rxmit -= len; 1100 tp->sackhint.sack_bytes_rexmit -= len; 1101 KASSERT(tp->sackhint.sack_bytes_rexmit 1102 >= 0, 1103 ("sackhint bytes rtx >= 0")); 1104 } else 1105 tp->snd_nxt -= len; 1106 } 1107 } 1108 1109 out: 1110 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1111 if (error == ENOBUFS) { 1112 if (!callout_active(tp->tt_rexmt) && 1113 !callout_active(tp->tt_persist)) 1114 callout_reset(tp->tt_rexmt, tp->t_rxtcur, 1115 tcp_timer_rexmt, tp); 1116 tp->snd_cwnd = tp->t_maxseg; 1117 return (0); 1118 } 1119 if (error == EMSGSIZE) { 1120 /* 1121 * ip_output() will have already fixed the route 1122 * for us. tcp_mtudisc() will, as its last action, 1123 * initiate retransmission, so it is important to 1124 * not do so here. 1125 */ 1126 tcp_mtudisc(tp->t_inpcb, 0); 1127 return 0; 1128 } 1129 if ((error == EHOSTUNREACH || error == ENETDOWN) 1130 && TCPS_HAVERCVDSYN(tp->t_state)) { 1131 tp->t_softerror = error; 1132 return (0); 1133 } 1134 return (error); 1135 } 1136 tcpstat.tcps_sndtotal++; 1137 1138 /* 1139 * Data sent (as far as we can tell). 1140 * If this advertises a larger window than any other segment, 1141 * then remember the size of the advertised window. 1142 * Any pending ACK has now been sent. 1143 */ 1144 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1145 tp->rcv_adv = tp->rcv_nxt + recwin; 1146 tp->last_ack_sent = tp->rcv_nxt; 1147 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1148 if (callout_active(tp->tt_delack)) 1149 callout_stop(tp->tt_delack); 1150 #if 0 1151 /* 1152 * This completely breaks TCP if newreno is turned on. What happens 1153 * is that if delayed-acks are turned on on the receiver, this code 1154 * on the transmitter effectively destroys the TCP window, forcing 1155 * it to four packets (1.5Kx4 = 6K window). 1156 */ 1157 if (sendalot && (!tcp_do_newreno || --maxburst)) 1158 goto again; 1159 #endif 1160 if (sendalot) 1161 goto again; 1162 return (0); 1163 } 1164 1165 void 1166 tcp_setpersist(tp) 1167 register struct tcpcb *tp; 1168 { 1169 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1170 int tt; 1171 1172 if (callout_active(tp->tt_rexmt)) 1173 panic("tcp_setpersist: retransmit pending"); 1174 /* 1175 * Start/restart persistance timer. 1176 */ 1177 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1178 TCPTV_PERSMIN, TCPTV_PERSMAX); 1179 callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp); 1180 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1181 tp->t_rxtshift++; 1182 } 1183