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/mbuf.h> 46 #include <sys/mutex.h> 47 #include <sys/protosw.h> 48 #include <sys/socket.h> 49 #include <sys/socketvar.h> 50 #include <sys/sysctl.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 #include <netinet/ip_options.h> 60 #ifdef INET6 61 #include <netinet6/in6_pcb.h> 62 #include <netinet/ip6.h> 63 #include <netinet6/ip6_var.h> 64 #endif 65 #include <netinet/tcp.h> 66 #define TCPOUTFLAGS 67 #include <netinet/tcp_fsm.h> 68 #include <netinet/tcp_seq.h> 69 #include <netinet/tcp_timer.h> 70 #include <netinet/tcp_var.h> 71 #include <netinet/tcpip.h> 72 #ifdef TCPDEBUG 73 #include <netinet/tcp_debug.h> 74 #endif 75 76 #ifdef IPSEC 77 #include <netinet6/ipsec.h> 78 #endif /*IPSEC*/ 79 80 #ifdef FAST_IPSEC 81 #include <netipsec/ipsec.h> 82 #define IPSEC 83 #endif /*FAST_IPSEC*/ 84 85 #include <machine/in_cksum.h> 86 87 #include <security/mac/mac_framework.h> 88 89 #ifdef notyet 90 extern struct mbuf *m_copypack(); 91 #endif 92 93 int path_mtu_discovery = 1; 94 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 95 &path_mtu_discovery, 1, "Enable Path MTU Discovery"); 96 97 int ss_fltsz = 1; 98 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW, 99 &ss_fltsz, 1, "Slow start flight size"); 100 101 int ss_fltsz_local = 4; 102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW, 103 &ss_fltsz_local, 1, "Slow start flight size for local networks"); 104 105 int tcp_do_newreno = 1; 106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno, 107 0, "Enable NewReno Algorithms"); 108 109 int tcp_do_tso = 1; 110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW, 111 &tcp_do_tso, 0, "Enable TCP Segmentation Offload"); 112 113 int tcp_do_autosndbuf = 1; 114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 115 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing"); 116 117 int tcp_autosndbuf_inc = 8*1024; 118 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 119 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer"); 120 121 int tcp_autosndbuf_max = 256*1024; 122 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 123 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer"); 124 125 126 /* 127 * Tcp output routine: figure out what should be sent and send it. 128 */ 129 int 130 tcp_output(struct tcpcb *tp) 131 { 132 struct socket *so = tp->t_inpcb->inp_socket; 133 long len, recwin, sendwin; 134 int off, flags, error; 135 #ifdef TCP_SIGNATURE 136 int sigoff = 0; 137 #endif 138 struct mbuf *m; 139 struct ip *ip = NULL; 140 struct ipovly *ipov = NULL; 141 struct tcphdr *th; 142 u_char opt[TCP_MAXOLEN]; 143 unsigned ipoptlen, optlen, hdrlen; 144 int idle, sendalot; 145 int i, sack_rxmit; 146 int sack_bytes_rxmt; 147 struct sackhole *p; 148 int tso = 0; 149 #if 0 150 int maxburst = TCP_MAXBURST; 151 #endif 152 #ifdef INET6 153 struct ip6_hdr *ip6 = NULL; 154 int isipv6; 155 156 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 157 #endif 158 159 INP_LOCK_ASSERT(tp->t_inpcb); 160 161 /* 162 * Determine length of data that should be transmitted, 163 * and flags that will be used. 164 * If there is some data or critical controls (SYN, RST) 165 * to send, then transmit; otherwise, investigate further. 166 */ 167 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 168 if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 169 /* 170 * We have been idle for "a while" and no acks are 171 * expected to clock out any data we send -- 172 * slow start to get ack "clock" running again. 173 * 174 * Set the slow-start flight size depending on whether 175 * this is a local network or not. 176 */ 177 int ss = ss_fltsz; 178 #ifdef INET6 179 if (isipv6) { 180 if (in6_localaddr(&tp->t_inpcb->in6p_faddr)) 181 ss = ss_fltsz_local; 182 } else 183 #endif /* INET6 */ 184 if (in_localaddr(tp->t_inpcb->inp_faddr)) 185 ss = ss_fltsz_local; 186 tp->snd_cwnd = tp->t_maxseg * ss; 187 } 188 tp->t_flags &= ~TF_LASTIDLE; 189 if (idle) { 190 if (tp->t_flags & TF_MORETOCOME) { 191 tp->t_flags |= TF_LASTIDLE; 192 idle = 0; 193 } 194 } 195 again: 196 /* 197 * If we've recently taken a timeout, snd_max will be greater than 198 * snd_nxt. There may be SACK information that allows us to avoid 199 * resending already delivered data. Adjust snd_nxt accordingly. 200 */ 201 if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max)) 202 tcp_sack_adjust(tp); 203 sendalot = 0; 204 off = tp->snd_nxt - tp->snd_una; 205 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 206 sendwin = min(sendwin, tp->snd_bwnd); 207 208 flags = tcp_outflags[tp->t_state]; 209 /* 210 * Send any SACK-generated retransmissions. If we're explicitly trying 211 * to send out new data (when sendalot is 1), bypass this function. 212 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 213 * we're replacing a (future) new transmission with a retransmission 214 * now, and we previously incremented snd_cwnd in tcp_input(). 215 */ 216 /* 217 * Still in sack recovery , reset rxmit flag to zero. 218 */ 219 sack_rxmit = 0; 220 sack_bytes_rxmt = 0; 221 len = 0; 222 p = NULL; 223 if (tp->sack_enable && IN_FASTRECOVERY(tp) && 224 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 225 long cwin; 226 227 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 228 if (cwin < 0) 229 cwin = 0; 230 /* Do not retransmit SACK segments beyond snd_recover */ 231 if (SEQ_GT(p->end, tp->snd_recover)) { 232 /* 233 * (At least) part of sack hole extends beyond 234 * snd_recover. Check to see if we can rexmit data 235 * for this hole. 236 */ 237 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 238 /* 239 * Can't rexmit any more data for this hole. 240 * That data will be rexmitted in the next 241 * sack recovery episode, when snd_recover 242 * moves past p->rxmit. 243 */ 244 p = NULL; 245 goto after_sack_rexmit; 246 } else 247 /* Can rexmit part of the current hole */ 248 len = ((long)ulmin(cwin, 249 tp->snd_recover - p->rxmit)); 250 } else 251 len = ((long)ulmin(cwin, p->end - p->rxmit)); 252 off = p->rxmit - tp->snd_una; 253 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 254 __func__, off)); 255 if (len > 0) { 256 sack_rxmit = 1; 257 sendalot = 1; 258 tcpstat.tcps_sack_rexmits++; 259 tcpstat.tcps_sack_rexmit_bytes += 260 min(len, tp->t_maxseg); 261 } 262 } 263 after_sack_rexmit: 264 /* 265 * Get standard flags, and add SYN or FIN if requested by 'hidden' 266 * state flags. 267 */ 268 if (tp->t_flags & TF_NEEDFIN) 269 flags |= TH_FIN; 270 if (tp->t_flags & TF_NEEDSYN) 271 flags |= TH_SYN; 272 273 SOCKBUF_LOCK(&so->so_snd); 274 /* 275 * If in persist timeout with window of 0, send 1 byte. 276 * Otherwise, if window is small but nonzero 277 * and timer expired, we will send what we can 278 * and go to transmit state. 279 */ 280 if (tp->t_flags & TF_FORCEDATA) { 281 if (sendwin == 0) { 282 /* 283 * If we still have some data to send, then 284 * clear the FIN bit. Usually this would 285 * happen below when it realizes that we 286 * aren't sending all the data. However, 287 * if we have exactly 1 byte of unsent data, 288 * then it won't clear the FIN bit below, 289 * and if we are in persist state, we wind 290 * up sending the packet without recording 291 * that we sent the FIN bit. 292 * 293 * We can't just blindly clear the FIN bit, 294 * because if we don't have any more data 295 * to send then the probe will be the FIN 296 * itself. 297 */ 298 if (off < so->so_snd.sb_cc) 299 flags &= ~TH_FIN; 300 sendwin = 1; 301 } else { 302 callout_stop(tp->tt_persist); 303 tp->t_rxtshift = 0; 304 } 305 } 306 307 /* 308 * If snd_nxt == snd_max and we have transmitted a FIN, the 309 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 310 * a negative length. This can also occur when TCP opens up 311 * its congestion window while receiving additional duplicate 312 * acks after fast-retransmit because TCP will reset snd_nxt 313 * to snd_max after the fast-retransmit. 314 * 315 * In the normal retransmit-FIN-only case, however, snd_nxt will 316 * be set to snd_una, the offset will be 0, and the length may 317 * wind up 0. 318 * 319 * If sack_rxmit is true we are retransmitting from the scoreboard 320 * in which case len is already set. 321 */ 322 if (sack_rxmit == 0) { 323 if (sack_bytes_rxmt == 0) 324 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off); 325 else { 326 long cwin; 327 328 /* 329 * We are inside of a SACK recovery episode and are 330 * sending new data, having retransmitted all the 331 * data possible in the scoreboard. 332 */ 333 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 334 - off); 335 /* 336 * Don't remove this (len > 0) check ! 337 * We explicitly check for len > 0 here (although it 338 * isn't really necessary), to work around a gcc 339 * optimization issue - to force gcc to compute 340 * len above. Without this check, the computation 341 * of len is bungled by the optimizer. 342 */ 343 if (len > 0) { 344 cwin = tp->snd_cwnd - 345 (tp->snd_nxt - tp->sack_newdata) - 346 sack_bytes_rxmt; 347 if (cwin < 0) 348 cwin = 0; 349 len = lmin(len, cwin); 350 } 351 } 352 } 353 354 /* 355 * Lop off SYN bit if it has already been sent. However, if this 356 * is SYN-SENT state and if segment contains data and if we don't 357 * know that foreign host supports TAO, suppress sending segment. 358 */ 359 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 360 if (tp->t_state != TCPS_SYN_RECEIVED) 361 flags &= ~TH_SYN; 362 off--, len++; 363 } 364 365 /* 366 * Be careful not to send data and/or FIN on SYN segments. 367 * This measure is needed to prevent interoperability problems 368 * with not fully conformant TCP implementations. 369 */ 370 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 371 len = 0; 372 flags &= ~TH_FIN; 373 } 374 375 if (len < 0) { 376 /* 377 * If FIN has been sent but not acked, 378 * but we haven't been called to retransmit, 379 * len will be < 0. Otherwise, window shrank 380 * after we sent into it. If window shrank to 0, 381 * cancel pending retransmit, pull snd_nxt back 382 * to (closed) window, and set the persist timer 383 * if it isn't already going. If the window didn't 384 * close completely, just wait for an ACK. 385 */ 386 len = 0; 387 if (sendwin == 0) { 388 callout_stop(tp->tt_rexmt); 389 tp->t_rxtshift = 0; 390 tp->snd_nxt = tp->snd_una; 391 if (!callout_active(tp->tt_persist)) 392 tcp_setpersist(tp); 393 } 394 } 395 396 /* len will be >= 0 after this point. */ 397 KASSERT(len >= 0, ("%s: len < 0", __func__)); 398 399 /* 400 * Automatic sizing of send socket buffer. Often the send buffer 401 * size is not optimally adjusted to the actual network conditions 402 * at hand (delay bandwidth product). Setting the buffer size too 403 * small limits throughput on links with high bandwidth and high 404 * delay (eg. trans-continental/oceanic links). Setting the 405 * buffer size too big consumes too much real kernel memory, 406 * especially with many connections on busy servers. 407 * 408 * The criteria to step up the send buffer one notch are: 409 * 1. receive window of remote host is larger than send buffer 410 * (with a fudge factor of 5/4th); 411 * 2. send buffer is filled to 7/8th with data (so we actually 412 * have data to make use of it); 413 * 3. send buffer fill has not hit maximal automatic size; 414 * 4. our send window (slow start and cogestion controlled) is 415 * larger than sent but unacknowledged data in send buffer. 416 * 417 * The remote host receive window scaling factor may limit the 418 * growing of the send buffer before it reaches its allowed 419 * maximum. 420 * 421 * It scales directly with slow start or congestion window 422 * and does at most one step per received ACK. This fast 423 * scaling has the drawback of growing the send buffer beyond 424 * what is strictly necessary to make full use of a given 425 * delay*bandwith product. However testing has shown this not 426 * to be much of an problem. At worst we are trading wasting 427 * of available bandwith (the non-use of it) for wasting some 428 * socket buffer memory. 429 * 430 * TODO: Shrink send buffer during idle periods together 431 * with congestion window. Requires another timer. Has to 432 * wait for upcoming tcp timer rewrite. 433 */ 434 if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 435 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 436 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 437 so->so_snd.sb_cc < tcp_autosndbuf_max && 438 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 439 if (!sbreserve_locked(&so->so_snd, 440 min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc, 441 tcp_autosndbuf_max), so, curthread)) 442 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 443 } 444 } 445 446 /* 447 * Truncate to the maximum segment length or enable TCP Segmentation 448 * Offloading (if supported by hardware) and ensure that FIN is removed 449 * if the length no longer contains the last data byte. 450 * 451 * TSO may only be used if we are in a pure bulk sending state. The 452 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 453 * IP options prevent using TSO. With TSO the TCP header is the same 454 * (except for the sequence number) for all generated packets. This 455 * makes it impossible to transmit any options which vary per generated 456 * segment or packet. 457 * 458 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and 459 * removal of FIN (if not already catched here) are handled later after 460 * the exact length of the TCP options are known. 461 */ 462 if (len > tp->t_maxseg) { 463 if ((tp->t_flags & TF_TSO) && tcp_do_tso && 464 ((tp->t_flags & TF_SIGNATURE) == 0) && 465 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 466 tp->t_inpcb->inp_options == NULL && 467 tp->t_inpcb->in6p_options == NULL && 468 tp->t_inpcb->inp_sp == NULL) { 469 tso = 1; 470 } else { 471 len = tp->t_maxseg; 472 sendalot = 1; 473 tso = 0; 474 } 475 } 476 if (sack_rxmit) { 477 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 478 flags &= ~TH_FIN; 479 } else { 480 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 481 flags &= ~TH_FIN; 482 } 483 484 recwin = sbspace(&so->so_rcv); 485 486 /* 487 * Sender silly window avoidance. We transmit under the following 488 * conditions when len is non-zero: 489 * 490 * - We have a full segment (or more with TSO) 491 * - This is the last buffer in a write()/send() and we are 492 * either idle or running NODELAY 493 * - we've timed out (e.g. persist timer) 494 * - we have more then 1/2 the maximum send window's worth of 495 * data (receiver may be limited the window size) 496 * - we need to retransmit 497 */ 498 if (len) { 499 if (len >= tp->t_maxseg) 500 goto send; 501 /* 502 * NOTE! on localhost connections an 'ack' from the remote 503 * end may occur synchronously with the output and cause 504 * us to flush a buffer queued with moretocome. XXX 505 * 506 * note: the len + off check is almost certainly unnecessary. 507 */ 508 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 509 (idle || (tp->t_flags & TF_NODELAY)) && 510 len + off >= so->so_snd.sb_cc && 511 (tp->t_flags & TF_NOPUSH) == 0) { 512 goto send; 513 } 514 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 515 goto send; 516 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 517 goto send; 518 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 519 goto send; 520 if (sack_rxmit) 521 goto send; 522 } 523 524 /* 525 * Compare available window to amount of window 526 * known to peer (as advertised window less 527 * next expected input). If the difference is at least two 528 * max size segments, or at least 50% of the maximum possible 529 * window, then want to send a window update to peer. 530 * Skip this if the connection is in T/TCP half-open state. 531 */ 532 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) { 533 /* 534 * "adv" is the amount we can increase the window, 535 * taking into account that we are limited by 536 * TCP_MAXWIN << tp->rcv_scale. 537 */ 538 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) - 539 (tp->rcv_adv - tp->rcv_nxt); 540 541 if (adv >= (long) (2 * tp->t_maxseg)) 542 goto send; 543 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 544 goto send; 545 } 546 547 /* 548 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 549 * is also a catch-all for the retransmit timer timeout case. 550 */ 551 if (tp->t_flags & TF_ACKNOW) 552 goto send; 553 if ((flags & TH_RST) || 554 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 555 goto send; 556 if (SEQ_GT(tp->snd_up, tp->snd_una)) 557 goto send; 558 /* 559 * If our state indicates that FIN should be sent 560 * and we have not yet done so, then we need to send. 561 */ 562 if (flags & TH_FIN && 563 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 564 goto send; 565 /* 566 * In SACK, it is possible for tcp_output to fail to send a segment 567 * after the retransmission timer has been turned off. Make sure 568 * that the retransmission timer is set. 569 */ 570 if (tp->sack_enable && SEQ_GT(tp->snd_max, tp->snd_una) && 571 !callout_active(tp->tt_rexmt) && 572 !callout_active(tp->tt_persist)) { 573 callout_reset(tp->tt_rexmt, tp->t_rxtcur, 574 tcp_timer_rexmt, tp); 575 goto just_return; 576 } 577 /* 578 * TCP window updates are not reliable, rather a polling protocol 579 * using ``persist'' packets is used to insure receipt of window 580 * updates. The three ``states'' for the output side are: 581 * idle not doing retransmits or persists 582 * persisting to move a small or zero window 583 * (re)transmitting and thereby not persisting 584 * 585 * callout_active(tp->tt_persist) 586 * is true when we are in persist state. 587 * (tp->t_flags & TF_FORCEDATA) 588 * is set when we are called to send a persist packet. 589 * callout_active(tp->tt_rexmt) 590 * is set when we are retransmitting 591 * The output side is idle when both timers are zero. 592 * 593 * If send window is too small, there is data to transmit, and no 594 * retransmit or persist is pending, then go to persist state. 595 * If nothing happens soon, send when timer expires: 596 * if window is nonzero, transmit what we can, 597 * otherwise force out a byte. 598 */ 599 if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) && 600 !callout_active(tp->tt_persist)) { 601 tp->t_rxtshift = 0; 602 tcp_setpersist(tp); 603 } 604 605 /* 606 * No reason to send a segment, just return. 607 */ 608 just_return: 609 SOCKBUF_UNLOCK(&so->so_snd); 610 return (0); 611 612 send: 613 SOCKBUF_LOCK_ASSERT(&so->so_snd); 614 /* 615 * Before ESTABLISHED, force sending of initial options 616 * unless TCP set not to do any options. 617 * NOTE: we assume that the IP/TCP header plus TCP options 618 * always fit in a single mbuf, leaving room for a maximum 619 * link header, i.e. 620 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 621 */ 622 optlen = 0; 623 #ifdef INET6 624 if (isipv6) 625 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 626 else 627 #endif 628 hdrlen = sizeof (struct tcpiphdr); 629 if (flags & TH_SYN) { 630 tp->snd_nxt = tp->iss; 631 if ((tp->t_flags & TF_NOOPT) == 0) { 632 u_short mss; 633 634 opt[0] = TCPOPT_MAXSEG; 635 opt[1] = TCPOLEN_MAXSEG; 636 mss = htons((u_short) tcp_mssopt(&tp->t_inpcb->inp_inc)); 637 (void)memcpy(opt + 2, &mss, sizeof(mss)); 638 optlen = TCPOLEN_MAXSEG; 639 640 if ((tp->t_flags & TF_REQ_SCALE) && 641 ((flags & TH_ACK) == 0 || 642 (tp->t_flags & TF_RCVD_SCALE))) { 643 *((u_int32_t *)(opt + optlen)) = htonl( 644 TCPOPT_NOP << 24 | 645 TCPOPT_WINDOW << 16 | 646 TCPOLEN_WINDOW << 8 | 647 tp->request_r_scale); 648 optlen += 4; 649 } 650 } 651 } 652 653 /* 654 * Send a timestamp and echo-reply if this is a SYN and our side 655 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 656 * and our peer have sent timestamps in our SYN's. 657 */ 658 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 659 (flags & TH_RST) == 0 && 660 ((flags & TH_ACK) == 0 || 661 (tp->t_flags & TF_RCVD_TSTMP))) { 662 u_int32_t *lp = (u_int32_t *)(opt + optlen); 663 664 /* Form timestamp option as shown in appendix A of RFC 1323. */ 665 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 666 *lp++ = htonl(ticks + tp->ts_offset); 667 *lp = htonl(tp->ts_recent); 668 optlen += TCPOLEN_TSTAMP_APPA; 669 } 670 671 /* Set receive buffer autosizing timestamp. */ 672 if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) 673 tp->rfbuf_ts = ticks; 674 675 #ifdef TCP_SIGNATURE 676 #ifdef INET6 677 if (!isipv6) 678 #endif 679 if (tp->t_flags & TF_SIGNATURE) { 680 int i; 681 u_char *bp; 682 683 /* Initialize TCP-MD5 option (RFC2385) */ 684 bp = (u_char *)opt + optlen; 685 *bp++ = TCPOPT_SIGNATURE; 686 *bp++ = TCPOLEN_SIGNATURE; 687 sigoff = optlen + 2; 688 for (i = 0; i < TCP_SIGLEN; i++) 689 *bp++ = 0; 690 optlen += TCPOLEN_SIGNATURE; 691 } 692 #endif /* TCP_SIGNATURE */ 693 694 if (tp->sack_enable && ((tp->t_flags & TF_NOOPT) == 0)) { 695 /* 696 * Tack on the SACK permitted option *last*. 697 * And do padding of options after tacking this on. 698 * This is because of MSS, TS, WinScale and Signatures are 699 * all present, we have just 2 bytes left for the SACK 700 * permitted option, which is just enough. 701 */ 702 /* 703 * If this is the first SYN of connection (not a SYN 704 * ACK), include SACK permitted option. If this is a 705 * SYN ACK, include SACK permitted option if peer has 706 * already done so. This is only for active connect, 707 * since the syncache takes care of the passive connect. 708 */ 709 if ((flags & TH_SYN) && 710 (!(flags & TH_ACK) || (tp->t_flags & TF_SACK_PERMIT))) { 711 u_char *bp; 712 bp = (u_char *)opt + optlen; 713 714 *bp++ = TCPOPT_SACK_PERMITTED; 715 *bp++ = TCPOLEN_SACK_PERMITTED; 716 optlen += TCPOLEN_SACK_PERMITTED; 717 } 718 719 /* 720 * Send SACKs if necessary. This should be the last 721 * option processed. Only as many SACKs are sent as 722 * are permitted by the maximum options size. 723 * 724 * In general, SACK blocks consume 8*n+2 bytes. 725 * So a full size SACK blocks option is 34 bytes 726 * (to generate 4 SACK blocks). At a minimum, 727 * we need 10 bytes (to generate 1 SACK block). 728 * If TCP Timestamps (12 bytes) and TCP Signatures 729 * (18 bytes) are both present, we'll just have 730 * 10 bytes for SACK options 40 - (12 + 18). 731 */ 732 if (TCPS_HAVEESTABLISHED(tp->t_state) && 733 (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0 && 734 MAX_TCPOPTLEN - optlen - 2 >= TCPOLEN_SACK) { 735 int nsack, sackoptlen, padlen; 736 u_char *bp = (u_char *)opt + optlen; 737 u_int32_t *lp; 738 739 nsack = (MAX_TCPOPTLEN - optlen - 2) / TCPOLEN_SACK; 740 nsack = min(nsack, tp->rcv_numsacks); 741 sackoptlen = (2 + nsack * TCPOLEN_SACK); 742 743 /* 744 * First we need to pad options so that the 745 * SACK blocks can start at a 4-byte boundary 746 * (sack option and length are at a 2 byte offset). 747 */ 748 padlen = (MAX_TCPOPTLEN - optlen - sackoptlen) % 4; 749 optlen += padlen; 750 while (padlen-- > 0) 751 *bp++ = TCPOPT_NOP; 752 753 tcpstat.tcps_sack_send_blocks++; 754 *bp++ = TCPOPT_SACK; 755 *bp++ = sackoptlen; 756 lp = (u_int32_t *)bp; 757 for (i = 0; i < nsack; i++) { 758 struct sackblk sack = tp->sackblks[i]; 759 *lp++ = htonl(sack.start); 760 *lp++ = htonl(sack.end); 761 } 762 optlen += sackoptlen; 763 } 764 } 765 766 /* Pad TCP options to a 4 byte boundary */ 767 if (optlen < MAX_TCPOPTLEN && (optlen % sizeof(u_int32_t))) { 768 int pad = sizeof(u_int32_t) - (optlen % sizeof(u_int32_t)); 769 u_char *bp = (u_char *)opt + optlen; 770 771 optlen += pad; 772 while (pad) { 773 *bp++ = TCPOPT_EOL; 774 pad--; 775 } 776 } 777 778 hdrlen += optlen; 779 780 #ifdef INET6 781 if (isipv6) 782 ipoptlen = ip6_optlen(tp->t_inpcb); 783 else 784 #endif 785 if (tp->t_inpcb->inp_options) 786 ipoptlen = tp->t_inpcb->inp_options->m_len - 787 offsetof(struct ipoption, ipopt_list); 788 else 789 ipoptlen = 0; 790 #ifdef IPSEC 791 ipoptlen += ipsec_hdrsiz_tcp(tp); 792 #endif 793 794 /* 795 * Adjust data length if insertion of options will 796 * bump the packet length beyond the t_maxopd length. 797 * Clear the FIN bit because we cut off the tail of 798 * the segment. 799 * 800 * When doing TSO limit a burst to TCP_MAXWIN minus the 801 * IP, TCP and Options length to keep ip->ip_len from 802 * overflowing. Prevent the last segment from being 803 * fractional thus making them all equal sized and set 804 * the flag to continue sending. 805 */ 806 if (len + optlen + ipoptlen > tp->t_maxopd) { 807 flags &= ~TH_FIN; 808 if (tso) { 809 if (len > TCP_MAXWIN - hdrlen) { 810 len = TCP_MAXWIN - hdrlen; 811 len = len - (len % (tp->t_maxopd - optlen)); 812 sendalot = 1; 813 } else if (tp->t_flags & TF_NEEDFIN) 814 sendalot = 1; 815 } else { 816 len = tp->t_maxopd - optlen - ipoptlen; 817 sendalot = 1; 818 } 819 } 820 821 /*#ifdef DIAGNOSTIC*/ 822 #ifdef INET6 823 if (max_linkhdr + hdrlen > MCLBYTES) 824 #else 825 if (max_linkhdr + hdrlen > MHLEN) 826 #endif 827 panic("tcphdr too big"); 828 /*#endif*/ 829 830 /* 831 * Grab a header mbuf, attaching a copy of data to 832 * be transmitted, and initialize the header from 833 * the template for sends on this connection. 834 */ 835 if (len) { 836 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 837 tcpstat.tcps_sndprobe++; 838 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 839 tcpstat.tcps_sndrexmitpack++; 840 tcpstat.tcps_sndrexmitbyte += len; 841 } else { 842 tcpstat.tcps_sndpack++; 843 tcpstat.tcps_sndbyte += len; 844 } 845 #ifdef notyet 846 if ((m = m_copypack(so->so_snd.sb_mb, off, 847 (int)len, max_linkhdr + hdrlen)) == 0) { 848 SOCKBUF_UNLOCK(&so->so_snd); 849 error = ENOBUFS; 850 goto out; 851 } 852 /* 853 * m_copypack left space for our hdr; use it. 854 */ 855 m->m_len += hdrlen; 856 m->m_data -= hdrlen; 857 #else 858 MGETHDR(m, M_DONTWAIT, MT_DATA); 859 if (m == NULL) { 860 SOCKBUF_UNLOCK(&so->so_snd); 861 error = ENOBUFS; 862 goto out; 863 } 864 #ifdef INET6 865 if (MHLEN < hdrlen + max_linkhdr) { 866 MCLGET(m, M_DONTWAIT); 867 if ((m->m_flags & M_EXT) == 0) { 868 SOCKBUF_UNLOCK(&so->so_snd); 869 m_freem(m); 870 error = ENOBUFS; 871 goto out; 872 } 873 } 874 #endif 875 m->m_data += max_linkhdr; 876 m->m_len = hdrlen; 877 if (len <= MHLEN - hdrlen - max_linkhdr) { 878 m_copydata(so->so_snd.sb_mb, off, (int) len, 879 mtod(m, caddr_t) + hdrlen); 880 m->m_len += len; 881 } else { 882 m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len); 883 if (m->m_next == 0) { 884 SOCKBUF_UNLOCK(&so->so_snd); 885 (void) m_free(m); 886 error = ENOBUFS; 887 goto out; 888 } 889 } 890 #endif 891 /* 892 * If we're sending everything we've got, set PUSH. 893 * (This will keep happy those implementations which only 894 * give data to the user when a buffer fills or 895 * a PUSH comes in.) 896 */ 897 if (off + len == so->so_snd.sb_cc) 898 flags |= TH_PUSH; 899 SOCKBUF_UNLOCK(&so->so_snd); 900 } else { 901 SOCKBUF_UNLOCK(&so->so_snd); 902 if (tp->t_flags & TF_ACKNOW) 903 tcpstat.tcps_sndacks++; 904 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 905 tcpstat.tcps_sndctrl++; 906 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 907 tcpstat.tcps_sndurg++; 908 else 909 tcpstat.tcps_sndwinup++; 910 911 MGETHDR(m, M_DONTWAIT, MT_DATA); 912 if (m == NULL) { 913 error = ENOBUFS; 914 goto out; 915 } 916 #ifdef INET6 917 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 918 MHLEN >= hdrlen) { 919 MH_ALIGN(m, hdrlen); 920 } else 921 #endif 922 m->m_data += max_linkhdr; 923 m->m_len = hdrlen; 924 } 925 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 926 m->m_pkthdr.rcvif = (struct ifnet *)0; 927 #ifdef MAC 928 mac_create_mbuf_from_inpcb(tp->t_inpcb, m); 929 #endif 930 #ifdef INET6 931 if (isipv6) { 932 ip6 = mtod(m, struct ip6_hdr *); 933 th = (struct tcphdr *)(ip6 + 1); 934 tcpip_fillheaders(tp->t_inpcb, ip6, th); 935 } else 936 #endif /* INET6 */ 937 { 938 ip = mtod(m, struct ip *); 939 ipov = (struct ipovly *)ip; 940 th = (struct tcphdr *)(ip + 1); 941 tcpip_fillheaders(tp->t_inpcb, ip, th); 942 } 943 944 /* 945 * Fill in fields, remembering maximum advertised 946 * window for use in delaying messages about window sizes. 947 * If resending a FIN, be sure not to use a new sequence number. 948 */ 949 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 950 tp->snd_nxt == tp->snd_max) 951 tp->snd_nxt--; 952 /* 953 * If we are doing retransmissions, then snd_nxt will 954 * not reflect the first unsent octet. For ACK only 955 * packets, we do not want the sequence number of the 956 * retransmitted packet, we want the sequence number 957 * of the next unsent octet. So, if there is no data 958 * (and no SYN or FIN), use snd_max instead of snd_nxt 959 * when filling in ti_seq. But if we are in persist 960 * state, snd_max might reflect one byte beyond the 961 * right edge of the window, so use snd_nxt in that 962 * case, since we know we aren't doing a retransmission. 963 * (retransmit and persist are mutually exclusive...) 964 */ 965 if (sack_rxmit == 0) { 966 if (len || (flags & (TH_SYN|TH_FIN)) 967 || callout_active(tp->tt_persist)) 968 th->th_seq = htonl(tp->snd_nxt); 969 else 970 th->th_seq = htonl(tp->snd_max); 971 } else { 972 th->th_seq = htonl(p->rxmit); 973 p->rxmit += len; 974 tp->sackhint.sack_bytes_rexmit += len; 975 } 976 th->th_ack = htonl(tp->rcv_nxt); 977 if (optlen) { 978 bcopy(opt, th + 1, optlen); 979 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 980 } 981 th->th_flags = flags; 982 /* 983 * Calculate receive window. Don't shrink window, 984 * but avoid silly window syndrome. 985 */ 986 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 987 recwin < (long)tp->t_maxseg) 988 recwin = 0; 989 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 990 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 991 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 992 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 993 th->th_win = htons((u_short) (recwin >> tp->rcv_scale)); 994 995 996 /* 997 * Adjust the RXWIN0SENT flag - indicate that we have advertised 998 * a 0 window. This may cause the remote transmitter to stall. This 999 * flag tells soreceive() to disable delayed acknowledgements when 1000 * draining the buffer. This can occur if the receiver is attempting 1001 * to read more data then can be buffered prior to transmitting on 1002 * the connection. 1003 */ 1004 if (recwin == 0) 1005 tp->t_flags |= TF_RXWIN0SENT; 1006 else 1007 tp->t_flags &= ~TF_RXWIN0SENT; 1008 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1009 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1010 th->th_flags |= TH_URG; 1011 } else 1012 /* 1013 * If no urgent pointer to send, then we pull 1014 * the urgent pointer to the left edge of the send window 1015 * so that it doesn't drift into the send window on sequence 1016 * number wraparound. 1017 */ 1018 tp->snd_up = tp->snd_una; /* drag it along */ 1019 1020 #ifdef TCP_SIGNATURE 1021 #ifdef INET6 1022 if (!isipv6) 1023 #endif 1024 if (tp->t_flags & TF_SIGNATURE) 1025 tcp_signature_compute(m, sizeof(struct ip), len, optlen, 1026 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1027 #endif 1028 1029 /* 1030 * Put TCP length in extended header, and then 1031 * checksum extended header and data. 1032 */ 1033 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1034 #ifdef INET6 1035 if (isipv6) 1036 /* 1037 * ip6_plen is not need to be filled now, and will be filled 1038 * in ip6_output. 1039 */ 1040 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 1041 sizeof(struct tcphdr) + optlen + len); 1042 else 1043 #endif /* INET6 */ 1044 { 1045 m->m_pkthdr.csum_flags = CSUM_TCP; 1046 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1047 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1048 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1049 1050 /* IP version must be set here for ipv4/ipv6 checking later */ 1051 KASSERT(ip->ip_v == IPVERSION, 1052 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1053 } 1054 1055 /* 1056 * Enable TSO and specify the size of the segments. 1057 * The TCP pseudo header checksum is always provided. 1058 * XXX: Fixme: This is currently not the case for IPv6. 1059 */ 1060 if (tso) { 1061 m->m_pkthdr.csum_flags = CSUM_TSO; 1062 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 1063 } 1064 1065 /* 1066 * In transmit state, time the transmission and arrange for 1067 * the retransmit. In persist state, just set snd_max. 1068 */ 1069 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1070 !callout_active(tp->tt_persist)) { 1071 tcp_seq startseq = tp->snd_nxt; 1072 1073 /* 1074 * Advance snd_nxt over sequence space of this segment. 1075 */ 1076 if (flags & (TH_SYN|TH_FIN)) { 1077 if (flags & TH_SYN) 1078 tp->snd_nxt++; 1079 if (flags & TH_FIN) { 1080 tp->snd_nxt++; 1081 tp->t_flags |= TF_SENTFIN; 1082 } 1083 } 1084 if (sack_rxmit) 1085 goto timer; 1086 tp->snd_nxt += len; 1087 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1088 tp->snd_max = tp->snd_nxt; 1089 /* 1090 * Time this transmission if not a retransmission and 1091 * not currently timing anything. 1092 */ 1093 if (tp->t_rtttime == 0) { 1094 tp->t_rtttime = ticks; 1095 tp->t_rtseq = startseq; 1096 tcpstat.tcps_segstimed++; 1097 } 1098 } 1099 1100 /* 1101 * Set retransmit timer if not currently set, 1102 * and not doing a pure ack or a keep-alive probe. 1103 * Initial value for retransmit timer is smoothed 1104 * round-trip time + 2 * round-trip time variance. 1105 * Initialize shift counter which is used for backoff 1106 * of retransmit time. 1107 */ 1108 timer: 1109 if (!callout_active(tp->tt_rexmt) && 1110 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1111 (tp->snd_nxt != tp->snd_una))) { 1112 if (callout_active(tp->tt_persist)) { 1113 callout_stop(tp->tt_persist); 1114 tp->t_rxtshift = 0; 1115 } 1116 callout_reset(tp->tt_rexmt, tp->t_rxtcur, 1117 tcp_timer_rexmt, tp); 1118 } 1119 } else { 1120 /* 1121 * Persist case, update snd_max but since we are in 1122 * persist mode (no window) we do not update snd_nxt. 1123 */ 1124 int xlen = len; 1125 if (flags & TH_SYN) 1126 ++xlen; 1127 if (flags & TH_FIN) { 1128 ++xlen; 1129 tp->t_flags |= TF_SENTFIN; 1130 } 1131 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1132 tp->snd_max = tp->snd_nxt + len; 1133 } 1134 1135 #ifdef TCPDEBUG 1136 /* 1137 * Trace. 1138 */ 1139 if (so->so_options & SO_DEBUG) { 1140 u_short save = 0; 1141 #ifdef INET6 1142 if (!isipv6) 1143 #endif 1144 { 1145 save = ipov->ih_len; 1146 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1147 } 1148 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1149 #ifdef INET6 1150 if (!isipv6) 1151 #endif 1152 ipov->ih_len = save; 1153 } 1154 #endif 1155 1156 /* 1157 * Fill in IP length and desired time to live and 1158 * send to IP level. There should be a better way 1159 * to handle ttl and tos; we could keep them in 1160 * the template, but need a way to checksum without them. 1161 */ 1162 /* 1163 * m->m_pkthdr.len should have been set before cksum calcuration, 1164 * because in6_cksum() need it. 1165 */ 1166 #ifdef INET6 1167 if (isipv6) { 1168 /* 1169 * we separately set hoplimit for every segment, since the 1170 * user might want to change the value via setsockopt. 1171 * Also, desired default hop limit might be changed via 1172 * Neighbor Discovery. 1173 */ 1174 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1175 1176 /* TODO: IPv6 IP6TOS_ECT bit on */ 1177 error = ip6_output(m, 1178 tp->t_inpcb->in6p_outputopts, NULL, 1179 ((so->so_options & SO_DONTROUTE) ? 1180 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb); 1181 } else 1182 #endif /* INET6 */ 1183 { 1184 ip->ip_len = m->m_pkthdr.len; 1185 #ifdef INET6 1186 if (INP_CHECK_SOCKAF(so, AF_INET6)) 1187 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1188 #endif /* INET6 */ 1189 /* 1190 * If we do path MTU discovery, then we set DF on every packet. 1191 * This might not be the best thing to do according to RFC3390 1192 * Section 2. However the tcp hostcache migitates the problem 1193 * so it affects only the first tcp connection with a host. 1194 */ 1195 if (path_mtu_discovery) 1196 ip->ip_off |= IP_DF; 1197 1198 error = ip_output(m, tp->t_inpcb->inp_options, NULL, 1199 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1200 tp->t_inpcb); 1201 } 1202 if (error) { 1203 1204 /* 1205 * We know that the packet was lost, so back out the 1206 * sequence number advance, if any. 1207 * 1208 * If the error is EPERM the packet got blocked by the 1209 * local firewall. Normally we should terminate the 1210 * connection but the blocking may have been spurious 1211 * due to a firewall reconfiguration cycle. So we treat 1212 * it like a packet loss and let the retransmit timer and 1213 * timeouts do their work over time. 1214 * XXX: It is a POLA question whether calling tcp_drop right 1215 * away would be the really correct behavior instead. 1216 */ 1217 if (error != EPERM && ((tp->t_flags & TF_FORCEDATA) == 0 || 1218 !callout_active(tp->tt_persist))) { 1219 /* 1220 * No need to check for TH_FIN here because 1221 * the TF_SENTFIN flag handles that case. 1222 */ 1223 if ((flags & TH_SYN) == 0) { 1224 if (sack_rxmit) { 1225 p->rxmit -= len; 1226 tp->sackhint.sack_bytes_rexmit -= len; 1227 KASSERT(tp->sackhint.sack_bytes_rexmit 1228 >= 0, 1229 ("sackhint bytes rtx >= 0")); 1230 } else 1231 tp->snd_nxt -= len; 1232 } 1233 } 1234 if (error == EPERM) { 1235 tp->t_softerror = error; 1236 return (error); 1237 } 1238 1239 out: 1240 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1241 if (error == ENOBUFS) { 1242 if (!callout_active(tp->tt_rexmt) && 1243 !callout_active(tp->tt_persist)) 1244 callout_reset(tp->tt_rexmt, tp->t_rxtcur, 1245 tcp_timer_rexmt, tp); 1246 tp->snd_cwnd = tp->t_maxseg; 1247 return (0); 1248 } 1249 if (error == EMSGSIZE) { 1250 /* 1251 * For some reason the interface we used initially 1252 * to send segments changed to another or lowered 1253 * its MTU. 1254 * 1255 * tcp_mtudisc() will find out the new MTU and as 1256 * its last action, initiate retransmission, so it 1257 * is important to not do so here. 1258 * 1259 * If TSO was active we either got an interface 1260 * without TSO capabilits or TSO was turned off. 1261 * Disable it for this connection as too and 1262 * immediatly retry with MSS sized segments generated 1263 * by this function. 1264 */ 1265 if (tso) 1266 tp->t_flags &= ~TF_TSO; 1267 tcp_mtudisc(tp->t_inpcb, 0); 1268 return 0; 1269 } 1270 if ((error == EHOSTUNREACH || error == ENETDOWN) 1271 && TCPS_HAVERCVDSYN(tp->t_state)) { 1272 tp->t_softerror = error; 1273 return (0); 1274 } 1275 return (error); 1276 } 1277 tcpstat.tcps_sndtotal++; 1278 1279 /* 1280 * Data sent (as far as we can tell). 1281 * If this advertises a larger window than any other segment, 1282 * then remember the size of the advertised window. 1283 * Any pending ACK has now been sent. 1284 */ 1285 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1286 tp->rcv_adv = tp->rcv_nxt + recwin; 1287 tp->last_ack_sent = tp->rcv_nxt; 1288 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1289 if (callout_active(tp->tt_delack)) 1290 callout_stop(tp->tt_delack); 1291 #if 0 1292 /* 1293 * This completely breaks TCP if newreno is turned on. What happens 1294 * is that if delayed-acks are turned on on the receiver, this code 1295 * on the transmitter effectively destroys the TCP window, forcing 1296 * it to four packets (1.5Kx4 = 6K window). 1297 */ 1298 if (sendalot && (!tcp_do_newreno || --maxburst)) 1299 goto again; 1300 #endif 1301 if (sendalot) 1302 goto again; 1303 return (0); 1304 } 1305 1306 void 1307 tcp_setpersist(tp) 1308 register struct tcpcb *tp; 1309 { 1310 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1311 int tt; 1312 1313 if (callout_active(tp->tt_rexmt)) 1314 panic("tcp_setpersist: retransmit pending"); 1315 /* 1316 * Start/restart persistance timer. 1317 */ 1318 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1319 TCPTV_PERSMIN, TCPTV_PERSMAX); 1320 callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp); 1321 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1322 tp->t_rxtshift++; 1323 } 1324