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 if (tp->t_state != TCPS_SYN_RECEIVED) 342 flags &= ~TH_SYN; 343 off--, len++; 344 } 345 346 /* 347 * Be careful not to send data and/or FIN on SYN segments. 348 * This measure is needed to prevent interoperability problems 349 * with not fully conformant TCP implementations. 350 */ 351 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 352 len = 0; 353 flags &= ~TH_FIN; 354 } 355 356 if (len < 0) { 357 /* 358 * If FIN has been sent but not acked, 359 * but we haven't been called to retransmit, 360 * len will be < 0. Otherwise, window shrank 361 * after we sent into it. If window shrank to 0, 362 * cancel pending retransmit, pull snd_nxt back 363 * to (closed) window, and set the persist timer 364 * if it isn't already going. If the window didn't 365 * close completely, just wait for an ACK. 366 */ 367 len = 0; 368 if (sendwin == 0) { 369 callout_stop(tp->tt_rexmt); 370 tp->t_rxtshift = 0; 371 tp->snd_nxt = tp->snd_una; 372 if (!callout_active(tp->tt_persist)) 373 tcp_setpersist(tp); 374 } 375 } 376 377 /* 378 * len will be >= 0 after this point. Truncate to the maximum 379 * segment length and ensure that FIN is removed if the length 380 * no longer contains the last data byte. 381 */ 382 if (len > tp->t_maxseg) { 383 len = tp->t_maxseg; 384 sendalot = 1; 385 } 386 if (sack_rxmit) { 387 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 388 flags &= ~TH_FIN; 389 } else { 390 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 391 flags &= ~TH_FIN; 392 } 393 394 recwin = sbspace(&so->so_rcv); 395 396 /* 397 * Sender silly window avoidance. We transmit under the following 398 * conditions when len is non-zero: 399 * 400 * - We have a full segment 401 * - This is the last buffer in a write()/send() and we are 402 * either idle or running NODELAY 403 * - we've timed out (e.g. persist timer) 404 * - we have more then 1/2 the maximum send window's worth of 405 * data (receiver may be limited the window size) 406 * - we need to retransmit 407 */ 408 if (len) { 409 if (len == tp->t_maxseg) 410 goto send; 411 /* 412 * NOTE! on localhost connections an 'ack' from the remote 413 * end may occur synchronously with the output and cause 414 * us to flush a buffer queued with moretocome. XXX 415 * 416 * note: the len + off check is almost certainly unnecessary. 417 */ 418 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 419 (idle || (tp->t_flags & TF_NODELAY)) && 420 len + off >= so->so_snd.sb_cc && 421 (tp->t_flags & TF_NOPUSH) == 0) { 422 goto send; 423 } 424 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 425 goto send; 426 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 427 goto send; 428 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 429 goto send; 430 if (sack_rxmit) 431 goto send; 432 } 433 434 /* 435 * Compare available window to amount of window 436 * known to peer (as advertised window less 437 * next expected input). If the difference is at least two 438 * max size segments, or at least 50% of the maximum possible 439 * window, then want to send a window update to peer. 440 * Skip this if the connection is in T/TCP half-open state. 441 */ 442 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) { 443 /* 444 * "adv" is the amount we can increase the window, 445 * taking into account that we are limited by 446 * TCP_MAXWIN << tp->rcv_scale. 447 */ 448 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) - 449 (tp->rcv_adv - tp->rcv_nxt); 450 451 if (adv >= (long) (2 * tp->t_maxseg)) 452 goto send; 453 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 454 goto send; 455 } 456 457 /* 458 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 459 * is also a catch-all for the retransmit timer timeout case. 460 */ 461 if (tp->t_flags & TF_ACKNOW) 462 goto send; 463 if ((flags & TH_RST) || 464 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 465 goto send; 466 if (SEQ_GT(tp->snd_up, tp->snd_una)) 467 goto send; 468 /* 469 * If our state indicates that FIN should be sent 470 * and we have not yet done so, then we need to send. 471 */ 472 if (flags & TH_FIN && 473 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 474 goto send; 475 /* 476 * In SACK, it is possible for tcp_output to fail to send a segment 477 * after the retransmission timer has been turned off. Make sure 478 * that the retransmission timer is set. 479 */ 480 if (tp->sack_enable && SEQ_GT(tp->snd_max, tp->snd_una) && 481 !callout_active(tp->tt_rexmt) && 482 !callout_active(tp->tt_persist)) { 483 callout_reset(tp->tt_rexmt, tp->t_rxtcur, 484 tcp_timer_rexmt, tp); 485 goto just_return; 486 } 487 /* 488 * TCP window updates are not reliable, rather a polling protocol 489 * using ``persist'' packets is used to insure receipt of window 490 * updates. The three ``states'' for the output side are: 491 * idle not doing retransmits or persists 492 * persisting to move a small or zero window 493 * (re)transmitting and thereby not persisting 494 * 495 * callout_active(tp->tt_persist) 496 * is true when we are in persist state. 497 * (tp->t_flags & TF_FORCEDATA) 498 * is set when we are called to send a persist packet. 499 * callout_active(tp->tt_rexmt) 500 * is set when we are retransmitting 501 * The output side is idle when both timers are zero. 502 * 503 * If send window is too small, there is data to transmit, and no 504 * retransmit or persist is pending, then go to persist state. 505 * If nothing happens soon, send when timer expires: 506 * if window is nonzero, transmit what we can, 507 * otherwise force out a byte. 508 */ 509 if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) && 510 !callout_active(tp->tt_persist)) { 511 tp->t_rxtshift = 0; 512 tcp_setpersist(tp); 513 } 514 515 /* 516 * No reason to send a segment, just return. 517 */ 518 just_return: 519 SOCKBUF_UNLOCK(&so->so_snd); 520 return (0); 521 522 send: 523 SOCKBUF_LOCK_ASSERT(&so->so_snd); 524 /* 525 * Before ESTABLISHED, force sending of initial options 526 * unless TCP set not to do any options. 527 * NOTE: we assume that the IP/TCP header plus TCP options 528 * always fit in a single mbuf, leaving room for a maximum 529 * link header, i.e. 530 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 531 */ 532 optlen = 0; 533 #ifdef INET6 534 if (isipv6) 535 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 536 else 537 #endif 538 hdrlen = sizeof (struct tcpiphdr); 539 if (flags & TH_SYN) { 540 tp->snd_nxt = tp->iss; 541 if ((tp->t_flags & TF_NOOPT) == 0) { 542 u_short mss; 543 544 opt[0] = TCPOPT_MAXSEG; 545 opt[1] = TCPOLEN_MAXSEG; 546 mss = htons((u_short) tcp_mssopt(&tp->t_inpcb->inp_inc)); 547 (void)memcpy(opt + 2, &mss, sizeof(mss)); 548 optlen = TCPOLEN_MAXSEG; 549 550 if ((tp->t_flags & TF_REQ_SCALE) && 551 ((flags & TH_ACK) == 0 || 552 (tp->t_flags & TF_RCVD_SCALE))) { 553 *((u_int32_t *)(opt + optlen)) = htonl( 554 TCPOPT_NOP << 24 | 555 TCPOPT_WINDOW << 16 | 556 TCPOLEN_WINDOW << 8 | 557 tp->request_r_scale); 558 optlen += 4; 559 } 560 } 561 } 562 563 /* 564 * Send a timestamp and echo-reply if this is a SYN and our side 565 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 566 * and our peer have sent timestamps in our SYN's. 567 */ 568 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 569 (flags & TH_RST) == 0 && 570 ((flags & TH_ACK) == 0 || 571 (tp->t_flags & TF_RCVD_TSTMP))) { 572 u_int32_t *lp = (u_int32_t *)(opt + optlen); 573 574 /* Form timestamp option as shown in appendix A of RFC 1323. */ 575 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 576 *lp++ = htonl(ticks); 577 *lp = htonl(tp->ts_recent); 578 optlen += TCPOLEN_TSTAMP_APPA; 579 } 580 581 #ifdef TCP_SIGNATURE 582 #ifdef INET6 583 if (!isipv6) 584 #endif 585 if (tp->t_flags & TF_SIGNATURE) { 586 int i; 587 u_char *bp; 588 589 /* Initialize TCP-MD5 option (RFC2385) */ 590 bp = (u_char *)opt + optlen; 591 *bp++ = TCPOPT_SIGNATURE; 592 *bp++ = TCPOLEN_SIGNATURE; 593 sigoff = optlen + 2; 594 for (i = 0; i < TCP_SIGLEN; i++) 595 *bp++ = 0; 596 optlen += TCPOLEN_SIGNATURE; 597 } 598 #endif /* TCP_SIGNATURE */ 599 600 if (tp->sack_enable && ((tp->t_flags & TF_NOOPT) == 0)) { 601 /* 602 * Tack on the SACK permitted option *last*. 603 * And do padding of options after tacking this on. 604 * This is because of MSS, TS, WinScale and Signatures are 605 * all present, we have just 2 bytes left for the SACK 606 * permitted option, which is just enough. 607 */ 608 /* 609 * If this is the first SYN of connection (not a SYN 610 * ACK), include SACK permitted option. If this is a 611 * SYN ACK, include SACK permitted option if peer has 612 * already done so. This is only for active connect, 613 * since the syncache takes care of the passive connect. 614 */ 615 if ((flags & TH_SYN) && 616 (!(flags & TH_ACK) || (tp->t_flags & TF_SACK_PERMIT))) { 617 u_char *bp; 618 bp = (u_char *)opt + optlen; 619 620 *bp++ = TCPOPT_SACK_PERMITTED; 621 *bp++ = TCPOLEN_SACK_PERMITTED; 622 optlen += TCPOLEN_SACK_PERMITTED; 623 } 624 625 /* 626 * Send SACKs if necessary. This should be the last 627 * option processed. Only as many SACKs are sent as 628 * are permitted by the maximum options size. 629 * 630 * In general, SACK blocks consume 8*n+2 bytes. 631 * So a full size SACK blocks option is 34 bytes 632 * (to generate 4 SACK blocks). At a minimum, 633 * we need 10 bytes (to generate 1 SACK block). 634 * If TCP Timestamps (12 bytes) and TCP Signatures 635 * (18 bytes) are both present, we'll just have 636 * 10 bytes for SACK options 40 - (12 + 18). 637 */ 638 if (TCPS_HAVEESTABLISHED(tp->t_state) && 639 (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0 && 640 MAX_TCPOPTLEN - optlen - 2 >= TCPOLEN_SACK) { 641 int nsack, sackoptlen, padlen; 642 u_char *bp = (u_char *)opt + optlen; 643 u_int32_t *lp; 644 645 nsack = (MAX_TCPOPTLEN - optlen - 2) / TCPOLEN_SACK; 646 nsack = min(nsack, tp->rcv_numsacks); 647 sackoptlen = (2 + nsack * TCPOLEN_SACK); 648 649 /* 650 * First we need to pad options so that the 651 * SACK blocks can start at a 4-byte boundary 652 * (sack option and length are at a 2 byte offset). 653 */ 654 padlen = (MAX_TCPOPTLEN - optlen - sackoptlen) % 4; 655 optlen += padlen; 656 while (padlen-- > 0) 657 *bp++ = TCPOPT_NOP; 658 659 tcpstat.tcps_sack_send_blocks++; 660 *bp++ = TCPOPT_SACK; 661 *bp++ = sackoptlen; 662 lp = (u_int32_t *)bp; 663 for (i = 0; i < nsack; i++) { 664 struct sackblk sack = tp->sackblks[i]; 665 *lp++ = htonl(sack.start); 666 *lp++ = htonl(sack.end); 667 } 668 optlen += sackoptlen; 669 } 670 } 671 672 /* Pad TCP options to a 4 byte boundary */ 673 if (optlen < MAX_TCPOPTLEN && (optlen % sizeof(u_int32_t))) { 674 int pad = sizeof(u_int32_t) - (optlen % sizeof(u_int32_t)); 675 u_char *bp = (u_char *)opt + optlen; 676 677 optlen += pad; 678 while (pad) { 679 *bp++ = TCPOPT_EOL; 680 pad--; 681 } 682 } 683 684 hdrlen += optlen; 685 686 #ifdef INET6 687 if (isipv6) 688 ipoptlen = ip6_optlen(tp->t_inpcb); 689 else 690 #endif 691 if (tp->t_inpcb->inp_options) 692 ipoptlen = tp->t_inpcb->inp_options->m_len - 693 offsetof(struct ipoption, ipopt_list); 694 else 695 ipoptlen = 0; 696 #ifdef IPSEC 697 ipoptlen += ipsec_hdrsiz_tcp(tp); 698 #endif 699 700 /* 701 * Adjust data length if insertion of options will 702 * bump the packet length beyond the t_maxopd length. 703 * Clear the FIN bit because we cut off the tail of 704 * the segment. 705 */ 706 if (len + optlen + ipoptlen > tp->t_maxopd) { 707 /* 708 * If there is still more to send, don't close the connection. 709 */ 710 flags &= ~TH_FIN; 711 len = tp->t_maxopd - optlen - ipoptlen; 712 sendalot = 1; 713 } 714 715 /*#ifdef DIAGNOSTIC*/ 716 #ifdef INET6 717 if (max_linkhdr + hdrlen > MCLBYTES) 718 #else 719 if (max_linkhdr + hdrlen > MHLEN) 720 #endif 721 panic("tcphdr too big"); 722 /*#endif*/ 723 724 /* 725 * Grab a header mbuf, attaching a copy of data to 726 * be transmitted, and initialize the header from 727 * the template for sends on this connection. 728 */ 729 if (len) { 730 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 731 tcpstat.tcps_sndprobe++; 732 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 733 tcpstat.tcps_sndrexmitpack++; 734 tcpstat.tcps_sndrexmitbyte += len; 735 } else { 736 tcpstat.tcps_sndpack++; 737 tcpstat.tcps_sndbyte += len; 738 } 739 #ifdef notyet 740 if ((m = m_copypack(so->so_snd.sb_mb, off, 741 (int)len, max_linkhdr + hdrlen)) == 0) { 742 SOCKBUF_UNLOCK(&so->so_snd); 743 error = ENOBUFS; 744 goto out; 745 } 746 /* 747 * m_copypack left space for our hdr; use it. 748 */ 749 m->m_len += hdrlen; 750 m->m_data -= hdrlen; 751 #else 752 MGETHDR(m, M_DONTWAIT, MT_DATA); 753 if (m == NULL) { 754 SOCKBUF_UNLOCK(&so->so_snd); 755 error = ENOBUFS; 756 goto out; 757 } 758 #ifdef INET6 759 if (MHLEN < hdrlen + max_linkhdr) { 760 MCLGET(m, M_DONTWAIT); 761 if ((m->m_flags & M_EXT) == 0) { 762 SOCKBUF_UNLOCK(&so->so_snd); 763 m_freem(m); 764 error = ENOBUFS; 765 goto out; 766 } 767 } 768 #endif 769 m->m_data += max_linkhdr; 770 m->m_len = hdrlen; 771 if (len <= MHLEN - hdrlen - max_linkhdr) { 772 m_copydata(so->so_snd.sb_mb, off, (int) len, 773 mtod(m, caddr_t) + hdrlen); 774 m->m_len += len; 775 } else { 776 m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len); 777 if (m->m_next == 0) { 778 SOCKBUF_UNLOCK(&so->so_snd); 779 (void) m_free(m); 780 error = ENOBUFS; 781 goto out; 782 } 783 } 784 #endif 785 /* 786 * If we're sending everything we've got, set PUSH. 787 * (This will keep happy those implementations which only 788 * give data to the user when a buffer fills or 789 * a PUSH comes in.) 790 */ 791 if (off + len == so->so_snd.sb_cc) 792 flags |= TH_PUSH; 793 SOCKBUF_UNLOCK(&so->so_snd); 794 } else { 795 SOCKBUF_UNLOCK(&so->so_snd); 796 if (tp->t_flags & TF_ACKNOW) 797 tcpstat.tcps_sndacks++; 798 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 799 tcpstat.tcps_sndctrl++; 800 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 801 tcpstat.tcps_sndurg++; 802 else 803 tcpstat.tcps_sndwinup++; 804 805 MGETHDR(m, M_DONTWAIT, MT_DATA); 806 if (m == NULL) { 807 error = ENOBUFS; 808 goto out; 809 } 810 #ifdef INET6 811 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 812 MHLEN >= hdrlen) { 813 MH_ALIGN(m, hdrlen); 814 } else 815 #endif 816 m->m_data += max_linkhdr; 817 m->m_len = hdrlen; 818 } 819 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 820 m->m_pkthdr.rcvif = (struct ifnet *)0; 821 #ifdef MAC 822 mac_create_mbuf_from_inpcb(tp->t_inpcb, m); 823 #endif 824 #ifdef INET6 825 if (isipv6) { 826 ip6 = mtod(m, struct ip6_hdr *); 827 th = (struct tcphdr *)(ip6 + 1); 828 tcpip_fillheaders(tp->t_inpcb, ip6, th); 829 } else 830 #endif /* INET6 */ 831 { 832 ip = mtod(m, struct ip *); 833 ipov = (struct ipovly *)ip; 834 th = (struct tcphdr *)(ip + 1); 835 tcpip_fillheaders(tp->t_inpcb, ip, th); 836 } 837 838 /* 839 * Fill in fields, remembering maximum advertised 840 * window for use in delaying messages about window sizes. 841 * If resending a FIN, be sure not to use a new sequence number. 842 */ 843 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 844 tp->snd_nxt == tp->snd_max) 845 tp->snd_nxt--; 846 /* 847 * If we are doing retransmissions, then snd_nxt will 848 * not reflect the first unsent octet. For ACK only 849 * packets, we do not want the sequence number of the 850 * retransmitted packet, we want the sequence number 851 * of the next unsent octet. So, if there is no data 852 * (and no SYN or FIN), use snd_max instead of snd_nxt 853 * when filling in ti_seq. But if we are in persist 854 * state, snd_max might reflect one byte beyond the 855 * right edge of the window, so use snd_nxt in that 856 * case, since we know we aren't doing a retransmission. 857 * (retransmit and persist are mutually exclusive...) 858 */ 859 if (sack_rxmit == 0) { 860 if (len || (flags & (TH_SYN|TH_FIN)) 861 || callout_active(tp->tt_persist)) 862 th->th_seq = htonl(tp->snd_nxt); 863 else 864 th->th_seq = htonl(tp->snd_max); 865 } else { 866 th->th_seq = htonl(p->rxmit); 867 p->rxmit += len; 868 tp->sackhint.sack_bytes_rexmit += len; 869 } 870 th->th_ack = htonl(tp->rcv_nxt); 871 if (optlen) { 872 bcopy(opt, th + 1, optlen); 873 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 874 } 875 th->th_flags = flags; 876 /* 877 * Calculate receive window. Don't shrink window, 878 * but avoid silly window syndrome. 879 */ 880 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 881 recwin < (long)tp->t_maxseg) 882 recwin = 0; 883 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 884 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 885 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 886 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 887 th->th_win = htons((u_short) (recwin >> tp->rcv_scale)); 888 889 890 /* 891 * Adjust the RXWIN0SENT flag - indicate that we have advertised 892 * a 0 window. This may cause the remote transmitter to stall. This 893 * flag tells soreceive() to disable delayed acknowledgements when 894 * draining the buffer. This can occur if the receiver is attempting 895 * to read more data then can be buffered prior to transmitting on 896 * the connection. 897 */ 898 if (recwin == 0) 899 tp->t_flags |= TF_RXWIN0SENT; 900 else 901 tp->t_flags &= ~TF_RXWIN0SENT; 902 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 903 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 904 th->th_flags |= TH_URG; 905 } else 906 /* 907 * If no urgent pointer to send, then we pull 908 * the urgent pointer to the left edge of the send window 909 * so that it doesn't drift into the send window on sequence 910 * number wraparound. 911 */ 912 tp->snd_up = tp->snd_una; /* drag it along */ 913 914 #ifdef TCP_SIGNATURE 915 #ifdef INET6 916 if (!isipv6) 917 #endif 918 if (tp->t_flags & TF_SIGNATURE) 919 tcp_signature_compute(m, sizeof(struct ip), len, optlen, 920 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 921 #endif 922 923 /* 924 * Put TCP length in extended header, and then 925 * checksum extended header and data. 926 */ 927 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 928 #ifdef INET6 929 if (isipv6) 930 /* 931 * ip6_plen is not need to be filled now, and will be filled 932 * in ip6_output. 933 */ 934 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 935 sizeof(struct tcphdr) + optlen + len); 936 else 937 #endif /* INET6 */ 938 { 939 m->m_pkthdr.csum_flags = CSUM_TCP; 940 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 941 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 942 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 943 944 /* IP version must be set here for ipv4/ipv6 checking later */ 945 KASSERT(ip->ip_v == IPVERSION, 946 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 947 } 948 949 /* 950 * In transmit state, time the transmission and arrange for 951 * the retransmit. In persist state, just set snd_max. 952 */ 953 if ((tp->t_flags & TF_FORCEDATA) == 0 || 954 !callout_active(tp->tt_persist)) { 955 tcp_seq startseq = tp->snd_nxt; 956 957 /* 958 * Advance snd_nxt over sequence space of this segment. 959 */ 960 if (flags & (TH_SYN|TH_FIN)) { 961 if (flags & TH_SYN) 962 tp->snd_nxt++; 963 if (flags & TH_FIN) { 964 tp->snd_nxt++; 965 tp->t_flags |= TF_SENTFIN; 966 } 967 } 968 if (sack_rxmit) 969 goto timer; 970 tp->snd_nxt += len; 971 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 972 tp->snd_max = tp->snd_nxt; 973 /* 974 * Time this transmission if not a retransmission and 975 * not currently timing anything. 976 */ 977 if (tp->t_rtttime == 0) { 978 tp->t_rtttime = ticks; 979 tp->t_rtseq = startseq; 980 tcpstat.tcps_segstimed++; 981 } 982 } 983 984 /* 985 * Set retransmit timer if not currently set, 986 * and not doing a pure ack or a keep-alive probe. 987 * Initial value for retransmit timer is smoothed 988 * round-trip time + 2 * round-trip time variance. 989 * Initialize shift counter which is used for backoff 990 * of retransmit time. 991 */ 992 timer: 993 if (!callout_active(tp->tt_rexmt) && 994 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 995 (tp->snd_nxt != tp->snd_una))) { 996 if (callout_active(tp->tt_persist)) { 997 callout_stop(tp->tt_persist); 998 tp->t_rxtshift = 0; 999 } 1000 callout_reset(tp->tt_rexmt, tp->t_rxtcur, 1001 tcp_timer_rexmt, tp); 1002 } 1003 } else { 1004 /* 1005 * Persist case, update snd_max but since we are in 1006 * persist mode (no window) we do not update snd_nxt. 1007 */ 1008 int xlen = len; 1009 if (flags & TH_SYN) 1010 ++xlen; 1011 if (flags & TH_FIN) { 1012 ++xlen; 1013 tp->t_flags |= TF_SENTFIN; 1014 } 1015 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1016 tp->snd_max = tp->snd_nxt + len; 1017 } 1018 1019 #ifdef TCPDEBUG 1020 /* 1021 * Trace. 1022 */ 1023 if (so->so_options & SO_DEBUG) { 1024 u_short save = 0; 1025 #ifdef INET6 1026 if (!isipv6) 1027 #endif 1028 { 1029 save = ipov->ih_len; 1030 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1031 } 1032 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1033 #ifdef INET6 1034 if (!isipv6) 1035 #endif 1036 ipov->ih_len = save; 1037 } 1038 #endif 1039 1040 /* 1041 * Fill in IP length and desired time to live and 1042 * send to IP level. There should be a better way 1043 * to handle ttl and tos; we could keep them in 1044 * the template, but need a way to checksum without them. 1045 */ 1046 /* 1047 * m->m_pkthdr.len should have been set before cksum calcuration, 1048 * because in6_cksum() need it. 1049 */ 1050 #ifdef INET6 1051 if (isipv6) { 1052 /* 1053 * we separately set hoplimit for every segment, since the 1054 * user might want to change the value via setsockopt. 1055 * Also, desired default hop limit might be changed via 1056 * Neighbor Discovery. 1057 */ 1058 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1059 1060 /* TODO: IPv6 IP6TOS_ECT bit on */ 1061 error = ip6_output(m, 1062 tp->t_inpcb->in6p_outputopts, NULL, 1063 ((so->so_options & SO_DONTROUTE) ? 1064 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb); 1065 } else 1066 #endif /* INET6 */ 1067 { 1068 ip->ip_len = m->m_pkthdr.len; 1069 #ifdef INET6 1070 if (INP_CHECK_SOCKAF(so, AF_INET6)) 1071 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1072 #endif /* INET6 */ 1073 /* 1074 * If we do path MTU discovery, then we set DF on every packet. 1075 * This might not be the best thing to do according to RFC3390 1076 * Section 2. However the tcp hostcache migitates the problem 1077 * so it affects only the first tcp connection with a host. 1078 */ 1079 if (path_mtu_discovery) 1080 ip->ip_off |= IP_DF; 1081 1082 error = ip_output(m, tp->t_inpcb->inp_options, NULL, 1083 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1084 tp->t_inpcb); 1085 } 1086 if (error) { 1087 1088 /* 1089 * We know that the packet was lost, so back out the 1090 * sequence number advance, if any. 1091 */ 1092 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1093 !callout_active(tp->tt_persist)) { 1094 /* 1095 * No need to check for TH_FIN here because 1096 * the TF_SENTFIN flag handles that case. 1097 */ 1098 if ((flags & TH_SYN) == 0) { 1099 if (sack_rxmit) { 1100 p->rxmit -= len; 1101 tp->sackhint.sack_bytes_rexmit -= len; 1102 KASSERT(tp->sackhint.sack_bytes_rexmit 1103 >= 0, 1104 ("sackhint bytes rtx >= 0")); 1105 } else 1106 tp->snd_nxt -= len; 1107 } 1108 } 1109 1110 out: 1111 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1112 if (error == ENOBUFS) { 1113 if (!callout_active(tp->tt_rexmt) && 1114 !callout_active(tp->tt_persist)) 1115 callout_reset(tp->tt_rexmt, tp->t_rxtcur, 1116 tcp_timer_rexmt, tp); 1117 tp->snd_cwnd = tp->t_maxseg; 1118 return (0); 1119 } 1120 if (error == EMSGSIZE) { 1121 /* 1122 * ip_output() will have already fixed the route 1123 * for us. tcp_mtudisc() will, as its last action, 1124 * initiate retransmission, so it is important to 1125 * not do so here. 1126 */ 1127 tcp_mtudisc(tp->t_inpcb, 0); 1128 return 0; 1129 } 1130 if ((error == EHOSTUNREACH || error == ENETDOWN) 1131 && TCPS_HAVERCVDSYN(tp->t_state)) { 1132 tp->t_softerror = error; 1133 return (0); 1134 } 1135 return (error); 1136 } 1137 tcpstat.tcps_sndtotal++; 1138 1139 /* 1140 * Data sent (as far as we can tell). 1141 * If this advertises a larger window than any other segment, 1142 * then remember the size of the advertised window. 1143 * Any pending ACK has now been sent. 1144 */ 1145 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1146 tp->rcv_adv = tp->rcv_nxt + recwin; 1147 tp->last_ack_sent = tp->rcv_nxt; 1148 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1149 if (callout_active(tp->tt_delack)) 1150 callout_stop(tp->tt_delack); 1151 #if 0 1152 /* 1153 * This completely breaks TCP if newreno is turned on. What happens 1154 * is that if delayed-acks are turned on on the receiver, this code 1155 * on the transmitter effectively destroys the TCP window, forcing 1156 * it to four packets (1.5Kx4 = 6K window). 1157 */ 1158 if (sendalot && (!tcp_do_newreno || --maxburst)) 1159 goto again; 1160 #endif 1161 if (sendalot) 1162 goto again; 1163 return (0); 1164 } 1165 1166 void 1167 tcp_setpersist(tp) 1168 register struct tcpcb *tp; 1169 { 1170 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1171 int tt; 1172 1173 if (callout_active(tp->tt_rexmt)) 1174 panic("tcp_setpersist: retransmit pending"); 1175 /* 1176 * Start/restart persistance timer. 1177 */ 1178 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1179 TCPTV_PERSMIN, TCPTV_PERSMAX); 1180 callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp); 1181 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1182 tp->t_rxtshift++; 1183 } 1184