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 39 #include <sys/param.h> 40 #include <sys/systm.h> 41 #include <sys/domain.h> 42 #include <sys/kernel.h> 43 #include <sys/lock.h> 44 #include <sys/mbuf.h> 45 #include <sys/mutex.h> 46 #include <sys/protosw.h> 47 #include <sys/socket.h> 48 #include <sys/socketvar.h> 49 #include <sys/sysctl.h> 50 51 #include <net/route.h> 52 53 #include <netinet/in.h> 54 #include <netinet/in_systm.h> 55 #include <netinet/ip.h> 56 #include <netinet/in_pcb.h> 57 #include <netinet/ip_var.h> 58 #include <netinet/ip_options.h> 59 #ifdef INET6 60 #include <netinet6/in6_pcb.h> 61 #include <netinet/ip6.h> 62 #include <netinet6/ip6_var.h> 63 #endif 64 #include <netinet/tcp.h> 65 #define TCPOUTFLAGS 66 #include <netinet/tcp_fsm.h> 67 #include <netinet/tcp_seq.h> 68 #include <netinet/tcp_timer.h> 69 #include <netinet/tcp_var.h> 70 #include <netinet/tcpip.h> 71 #ifdef TCPDEBUG 72 #include <netinet/tcp_debug.h> 73 #endif 74 75 #ifdef IPSEC 76 #include <netinet6/ipsec.h> 77 #endif /*IPSEC*/ 78 79 #ifdef FAST_IPSEC 80 #include <netipsec/ipsec.h> 81 #define IPSEC 82 #endif /*FAST_IPSEC*/ 83 84 #include <machine/in_cksum.h> 85 86 #include <security/mac/mac_framework.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, 106 &tcp_do_newreno, 0, "Enable NewReno Algorithms"); 107 108 int tcp_do_tso = 1; 109 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW, 110 &tcp_do_tso, 0, "Enable TCP Segmentation Offload"); 111 112 int tcp_do_autosndbuf = 1; 113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 114 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing"); 115 116 int tcp_autosndbuf_inc = 8*1024; 117 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 118 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer"); 119 120 int tcp_autosndbuf_max = 256*1024; 121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 122 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer"); 123 124 125 /* 126 * Tcp output routine: figure out what should be sent and send it. 127 */ 128 int 129 tcp_output(struct tcpcb *tp) 130 { 131 struct socket *so = tp->t_inpcb->inp_socket; 132 long len, recwin, sendwin; 133 int off, flags, error; 134 #ifdef TCP_SIGNATURE 135 int sigoff = 0; 136 #endif 137 struct mbuf *m; 138 struct ip *ip = NULL; 139 struct ipovly *ipov = NULL; 140 struct tcphdr *th; 141 u_char opt[TCP_MAXOLEN]; 142 unsigned ipoptlen, optlen, hdrlen; 143 int idle, sendalot; 144 int sack_rxmit, sack_bytes_rxmt; 145 struct sackhole *p; 146 int tso = 0; 147 struct tcpopt to; 148 #if 0 149 int maxburst = TCP_MAXBURST; 150 #endif 151 #ifdef INET6 152 struct ip6_hdr *ip6 = NULL; 153 int isipv6; 154 155 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 156 #endif 157 158 INP_LOCK_ASSERT(tp->t_inpcb); 159 160 /* 161 * Determine length of data that should be transmitted, 162 * and flags that will be used. 163 * If there is some data or critical controls (SYN, RST) 164 * to send, then transmit; otherwise, investigate further. 165 */ 166 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 167 if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 168 /* 169 * We have been idle for "a while" and no acks are 170 * expected to clock out any data we send -- 171 * slow start to get ack "clock" running again. 172 * 173 * Set the slow-start flight size depending on whether 174 * this is a local network or not. 175 */ 176 int ss = ss_fltsz; 177 #ifdef INET6 178 if (isipv6) { 179 if (in6_localaddr(&tp->t_inpcb->in6p_faddr)) 180 ss = ss_fltsz_local; 181 } else 182 #endif /* INET6 */ 183 if (in_localaddr(tp->t_inpcb->inp_faddr)) 184 ss = ss_fltsz_local; 185 tp->snd_cwnd = tp->t_maxseg * ss; 186 } 187 tp->t_flags &= ~TF_LASTIDLE; 188 if (idle) { 189 if (tp->t_flags & TF_MORETOCOME) { 190 tp->t_flags |= TF_LASTIDLE; 191 idle = 0; 192 } 193 } 194 again: 195 /* 196 * If we've recently taken a timeout, snd_max will be greater than 197 * snd_nxt. There may be SACK information that allows us to avoid 198 * resending already delivered data. Adjust snd_nxt accordingly. 199 */ 200 if ((tp->t_flags & TF_SACK_PERMIT) && 201 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->t_flags & TF_SACK_PERMIT) && 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 tcp_timer_activate(tp, TT_PERSIST, 0); 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 tcp_timer_activate(tp, TT_REXMT, 0); 389 tp->t_rxtshift = 0; 390 tp->snd_nxt = tp->snd_una; 391 if (!tcp_timer_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->t_flags & TF_SACK_PERMIT) && 571 SEQ_GT(tp->snd_max, tp->snd_una) && 572 !tcp_timer_active(tp, TT_REXMT) && 573 !tcp_timer_active(tp, TT_PERSIST)) { 574 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 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 * tcp_timer_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 * tcp_timer_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 && !tcp_timer_active(tp, TT_REXMT) && 600 !tcp_timer_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); 629 630 /* 631 * Compute options for segment. 632 * We only have to care about SYN and established connection 633 * segments. Options for SYN-ACK segments are handled in TCP 634 * syncache. 635 */ 636 if ((tp->t_flags & TF_NOOPT) == 0) { 637 to.to_flags = 0; 638 /* Maximum segment size. */ 639 if (flags & TH_SYN) { 640 tp->snd_nxt = tp->iss; 641 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 642 to.to_flags |= TOF_MSS; 643 } 644 /* Window scaling. */ 645 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 646 to.to_wscale = tp->request_r_scale; 647 to.to_flags |= TOF_SCALE; 648 } 649 /* Timestamps. */ 650 if ((tp->t_flags & TF_RCVD_TSTMP) || 651 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 652 to.to_tsval = ticks + tp->ts_offset; 653 to.to_tsecr = tp->ts_recent; 654 to.to_flags |= TOF_TS; 655 /* Set receive buffer autosizing timestamp. */ 656 if (tp->rfbuf_ts == 0 && 657 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 658 tp->rfbuf_ts = ticks; 659 } 660 /* Selective ACK's. */ 661 if (tp->t_flags & TF_SACK_PERMIT) { 662 if (flags & TH_SYN) 663 to.to_flags |= TOF_SACKPERM; 664 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 665 (tp->t_flags & TF_SACK_PERMIT) && 666 tp->rcv_numsacks > 0) { 667 to.to_flags |= TOF_SACK; 668 to.to_nsacks = tp->rcv_numsacks; 669 to.to_sacks = (u_char *)tp->sackblks; 670 } 671 } 672 #ifdef TCP_SIGNATURE 673 /* TCP-MD5 (RFC2385). */ 674 #ifdef INET6 675 if (!isipv6 && (tp->t_flags & TF_SIGNATURE)) 676 #else 677 if (tp->t_flags & TF_SIGNATURE) 678 #endif /* INET6 */ 679 to.to_flags |= TOF_SIGNATURE; 680 #endif /* TCP_SIGNATURE */ 681 682 /* Processing the options. */ 683 hdrlen += optlen = tcp_addoptions(&to, (u_char *)&opt); 684 685 #ifdef TCP_SIGNATURE 686 sigoff = to.to_signature - (u_char *)&to; 687 #endif /* TCP_SIGNATURE */ 688 } 689 690 #ifdef INET6 691 if (isipv6) 692 ipoptlen = ip6_optlen(tp->t_inpcb); 693 else 694 #endif 695 if (tp->t_inpcb->inp_options) 696 ipoptlen = tp->t_inpcb->inp_options->m_len - 697 offsetof(struct ipoption, ipopt_list); 698 else 699 ipoptlen = 0; 700 #ifdef IPSEC 701 ipoptlen += ipsec_hdrsiz_tcp(tp); 702 #endif 703 704 /* 705 * Adjust data length if insertion of options will 706 * bump the packet length beyond the t_maxopd length. 707 * Clear the FIN bit because we cut off the tail of 708 * the segment. 709 * 710 * When doing TSO limit a burst to TCP_MAXWIN minus the 711 * IP, TCP and Options length to keep ip->ip_len from 712 * overflowing. Prevent the last segment from being 713 * fractional thus making them all equal sized and set 714 * the flag to continue sending. TSO is disabled when 715 * IP options or IPSEC are present. 716 */ 717 if (len + optlen + ipoptlen > tp->t_maxopd) { 718 flags &= ~TH_FIN; 719 if (tso) { 720 if (len > TCP_MAXWIN - hdrlen - optlen) { 721 len = TCP_MAXWIN - hdrlen - optlen; 722 len = len - (len % (tp->t_maxopd - optlen)); 723 sendalot = 1; 724 } else if (tp->t_flags & TF_NEEDFIN) 725 sendalot = 1; 726 } else { 727 len = tp->t_maxopd - optlen - ipoptlen; 728 sendalot = 1; 729 } 730 } 731 732 /*#ifdef DIAGNOSTIC*/ 733 #ifdef INET6 734 if (max_linkhdr + hdrlen > MCLBYTES) 735 #else 736 if (max_linkhdr + hdrlen > MHLEN) 737 #endif 738 panic("tcphdr too big"); 739 /*#endif*/ 740 741 /* 742 * Grab a header mbuf, attaching a copy of data to 743 * be transmitted, and initialize the header from 744 * the template for sends on this connection. 745 */ 746 if (len) { 747 struct mbuf *mb; 748 u_int moff; 749 750 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 751 tcpstat.tcps_sndprobe++; 752 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 753 tcpstat.tcps_sndrexmitpack++; 754 tcpstat.tcps_sndrexmitbyte += len; 755 } else { 756 tcpstat.tcps_sndpack++; 757 tcpstat.tcps_sndbyte += len; 758 } 759 #ifdef notyet 760 if ((m = m_copypack(so->so_snd.sb_mb, off, 761 (int)len, max_linkhdr + hdrlen)) == 0) { 762 SOCKBUF_UNLOCK(&so->so_snd); 763 error = ENOBUFS; 764 goto out; 765 } 766 /* 767 * m_copypack left space for our hdr; use it. 768 */ 769 m->m_len += hdrlen; 770 m->m_data -= hdrlen; 771 #else 772 MGETHDR(m, M_DONTWAIT, MT_DATA); 773 if (m == NULL) { 774 SOCKBUF_UNLOCK(&so->so_snd); 775 error = ENOBUFS; 776 goto out; 777 } 778 #ifdef INET6 779 if (MHLEN < hdrlen + max_linkhdr) { 780 MCLGET(m, M_DONTWAIT); 781 if ((m->m_flags & M_EXT) == 0) { 782 SOCKBUF_UNLOCK(&so->so_snd); 783 m_freem(m); 784 error = ENOBUFS; 785 goto out; 786 } 787 } 788 #endif 789 m->m_data += max_linkhdr; 790 m->m_len = hdrlen; 791 792 /* 793 * Start the m_copy functions from the closest mbuf 794 * to the offset in the socket buffer chain. 795 */ 796 mb = sbsndptr(&so->so_snd, off, len, &moff); 797 798 if (len <= MHLEN - hdrlen - max_linkhdr) { 799 m_copydata(mb, moff, (int)len, 800 mtod(m, caddr_t) + hdrlen); 801 m->m_len += len; 802 } else { 803 m->m_next = m_copy(mb, moff, (int)len); 804 if (m->m_next == NULL) { 805 SOCKBUF_UNLOCK(&so->so_snd); 806 (void) m_free(m); 807 error = ENOBUFS; 808 goto out; 809 } 810 } 811 #endif 812 /* 813 * If we're sending everything we've got, set PUSH. 814 * (This will keep happy those implementations which only 815 * give data to the user when a buffer fills or 816 * a PUSH comes in.) 817 */ 818 if (off + len == so->so_snd.sb_cc) 819 flags |= TH_PUSH; 820 SOCKBUF_UNLOCK(&so->so_snd); 821 } else { 822 SOCKBUF_UNLOCK(&so->so_snd); 823 if (tp->t_flags & TF_ACKNOW) 824 tcpstat.tcps_sndacks++; 825 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 826 tcpstat.tcps_sndctrl++; 827 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 828 tcpstat.tcps_sndurg++; 829 else 830 tcpstat.tcps_sndwinup++; 831 832 MGETHDR(m, M_DONTWAIT, MT_DATA); 833 if (m == NULL) { 834 error = ENOBUFS; 835 goto out; 836 } 837 #ifdef INET6 838 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 839 MHLEN >= hdrlen) { 840 MH_ALIGN(m, hdrlen); 841 } else 842 #endif 843 m->m_data += max_linkhdr; 844 m->m_len = hdrlen; 845 } 846 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 847 m->m_pkthdr.rcvif = (struct ifnet *)0; 848 #ifdef MAC 849 mac_create_mbuf_from_inpcb(tp->t_inpcb, m); 850 #endif 851 #ifdef INET6 852 if (isipv6) { 853 ip6 = mtod(m, struct ip6_hdr *); 854 th = (struct tcphdr *)(ip6 + 1); 855 tcpip_fillheaders(tp->t_inpcb, ip6, th); 856 } else 857 #endif /* INET6 */ 858 { 859 ip = mtod(m, struct ip *); 860 ipov = (struct ipovly *)ip; 861 th = (struct tcphdr *)(ip + 1); 862 tcpip_fillheaders(tp->t_inpcb, ip, th); 863 } 864 865 /* 866 * Fill in fields, remembering maximum advertised 867 * window for use in delaying messages about window sizes. 868 * If resending a FIN, be sure not to use a new sequence number. 869 */ 870 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 871 tp->snd_nxt == tp->snd_max) 872 tp->snd_nxt--; 873 /* 874 * If we are doing retransmissions, then snd_nxt will 875 * not reflect the first unsent octet. For ACK only 876 * packets, we do not want the sequence number of the 877 * retransmitted packet, we want the sequence number 878 * of the next unsent octet. So, if there is no data 879 * (and no SYN or FIN), use snd_max instead of snd_nxt 880 * when filling in ti_seq. But if we are in persist 881 * state, snd_max might reflect one byte beyond the 882 * right edge of the window, so use snd_nxt in that 883 * case, since we know we aren't doing a retransmission. 884 * (retransmit and persist are mutually exclusive...) 885 */ 886 if (sack_rxmit == 0) { 887 if (len || (flags & (TH_SYN|TH_FIN)) || 888 tcp_timer_active(tp, TT_PERSIST)) 889 th->th_seq = htonl(tp->snd_nxt); 890 else 891 th->th_seq = htonl(tp->snd_max); 892 } else { 893 th->th_seq = htonl(p->rxmit); 894 p->rxmit += len; 895 tp->sackhint.sack_bytes_rexmit += len; 896 } 897 th->th_ack = htonl(tp->rcv_nxt); 898 if (optlen) { 899 bcopy(opt, th + 1, optlen); 900 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 901 } 902 th->th_flags = flags; 903 /* 904 * Calculate receive window. Don't shrink window, 905 * but avoid silly window syndrome. 906 * 907 * XXX: RFC1323: The Window field in a SYN (i.e., a <SYN> or 908 * <SYN,ACK>) segment itself is never scaled. 909 */ 910 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 911 recwin < (long)tp->t_maxseg) 912 recwin = 0; 913 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 914 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 915 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 916 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 917 th->th_win = htons((u_short) (recwin >> tp->rcv_scale)); 918 919 920 /* 921 * Adjust the RXWIN0SENT flag - indicate that we have advertised 922 * a 0 window. This may cause the remote transmitter to stall. This 923 * flag tells soreceive() to disable delayed acknowledgements when 924 * draining the buffer. This can occur if the receiver is attempting 925 * to read more data then can be buffered prior to transmitting on 926 * the connection. 927 */ 928 if (recwin == 0) 929 tp->t_flags |= TF_RXWIN0SENT; 930 else 931 tp->t_flags &= ~TF_RXWIN0SENT; 932 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 933 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 934 th->th_flags |= TH_URG; 935 } else 936 /* 937 * If no urgent pointer to send, then we pull 938 * the urgent pointer to the left edge of the send window 939 * so that it doesn't drift into the send window on sequence 940 * number wraparound. 941 */ 942 tp->snd_up = tp->snd_una; /* drag it along */ 943 944 #ifdef TCP_SIGNATURE 945 #ifdef INET6 946 if (!isipv6) 947 #endif 948 if (tp->t_flags & TF_SIGNATURE) 949 tcp_signature_compute(m, sizeof(struct ip), len, optlen, 950 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 951 #endif 952 953 /* 954 * Put TCP length in extended header, and then 955 * checksum extended header and data. 956 */ 957 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 958 #ifdef INET6 959 if (isipv6) 960 /* 961 * ip6_plen is not need to be filled now, and will be filled 962 * in ip6_output. 963 */ 964 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 965 sizeof(struct tcphdr) + optlen + len); 966 else 967 #endif /* INET6 */ 968 { 969 m->m_pkthdr.csum_flags = CSUM_TCP; 970 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 971 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 972 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 973 974 /* IP version must be set here for ipv4/ipv6 checking later */ 975 KASSERT(ip->ip_v == IPVERSION, 976 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 977 } 978 979 /* 980 * Enable TSO and specify the size of the segments. 981 * The TCP pseudo header checksum is always provided. 982 * XXX: Fixme: This is currently not the case for IPv6. 983 */ 984 if (tso) { 985 m->m_pkthdr.csum_flags = CSUM_TSO; 986 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 987 } 988 989 /* 990 * In transmit state, time the transmission and arrange for 991 * the retransmit. In persist state, just set snd_max. 992 */ 993 if ((tp->t_flags & TF_FORCEDATA) == 0 || 994 !tcp_timer_active(tp, TT_PERSIST)) { 995 tcp_seq startseq = tp->snd_nxt; 996 997 /* 998 * Advance snd_nxt over sequence space of this segment. 999 */ 1000 if (flags & (TH_SYN|TH_FIN)) { 1001 if (flags & TH_SYN) 1002 tp->snd_nxt++; 1003 if (flags & TH_FIN) { 1004 tp->snd_nxt++; 1005 tp->t_flags |= TF_SENTFIN; 1006 } 1007 } 1008 if (sack_rxmit) 1009 goto timer; 1010 tp->snd_nxt += len; 1011 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1012 tp->snd_max = tp->snd_nxt; 1013 /* 1014 * Time this transmission if not a retransmission and 1015 * not currently timing anything. 1016 */ 1017 if (tp->t_rtttime == 0) { 1018 tp->t_rtttime = ticks; 1019 tp->t_rtseq = startseq; 1020 tcpstat.tcps_segstimed++; 1021 } 1022 } 1023 1024 /* 1025 * Set retransmit timer if not currently set, 1026 * and not doing a pure ack or a keep-alive probe. 1027 * Initial value for retransmit timer is smoothed 1028 * round-trip time + 2 * round-trip time variance. 1029 * Initialize shift counter which is used for backoff 1030 * of retransmit time. 1031 */ 1032 timer: 1033 if (!tcp_timer_active(tp, TT_REXMT) && 1034 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1035 (tp->snd_nxt != tp->snd_una))) { 1036 if (tcp_timer_active(tp, TT_PERSIST)) { 1037 tcp_timer_activate(tp, TT_PERSIST, 0); 1038 tp->t_rxtshift = 0; 1039 } 1040 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1041 } 1042 } else { 1043 /* 1044 * Persist case, update snd_max but since we are in 1045 * persist mode (no window) we do not update snd_nxt. 1046 */ 1047 int xlen = len; 1048 if (flags & TH_SYN) 1049 ++xlen; 1050 if (flags & TH_FIN) { 1051 ++xlen; 1052 tp->t_flags |= TF_SENTFIN; 1053 } 1054 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1055 tp->snd_max = tp->snd_nxt + len; 1056 } 1057 1058 #ifdef TCPDEBUG 1059 /* 1060 * Trace. 1061 */ 1062 if (so->so_options & SO_DEBUG) { 1063 u_short save = 0; 1064 #ifdef INET6 1065 if (!isipv6) 1066 #endif 1067 { 1068 save = ipov->ih_len; 1069 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1070 } 1071 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1072 #ifdef INET6 1073 if (!isipv6) 1074 #endif 1075 ipov->ih_len = save; 1076 } 1077 #endif 1078 1079 /* 1080 * Fill in IP length and desired time to live and 1081 * send to IP level. There should be a better way 1082 * to handle ttl and tos; we could keep them in 1083 * the template, but need a way to checksum without them. 1084 */ 1085 /* 1086 * m->m_pkthdr.len should have been set before cksum calcuration, 1087 * because in6_cksum() need it. 1088 */ 1089 #ifdef INET6 1090 if (isipv6) { 1091 /* 1092 * we separately set hoplimit for every segment, since the 1093 * user might want to change the value via setsockopt. 1094 * Also, desired default hop limit might be changed via 1095 * Neighbor Discovery. 1096 */ 1097 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1098 1099 /* TODO: IPv6 IP6TOS_ECT bit on */ 1100 error = ip6_output(m, 1101 tp->t_inpcb->in6p_outputopts, NULL, 1102 ((so->so_options & SO_DONTROUTE) ? 1103 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb); 1104 } else 1105 #endif /* INET6 */ 1106 { 1107 ip->ip_len = m->m_pkthdr.len; 1108 #ifdef INET6 1109 if (INP_CHECK_SOCKAF(so, AF_INET6)) 1110 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1111 #endif /* INET6 */ 1112 /* 1113 * If we do path MTU discovery, then we set DF on every packet. 1114 * This might not be the best thing to do according to RFC3390 1115 * Section 2. However the tcp hostcache migitates the problem 1116 * so it affects only the first tcp connection with a host. 1117 */ 1118 if (path_mtu_discovery) 1119 ip->ip_off |= IP_DF; 1120 1121 error = ip_output(m, tp->t_inpcb->inp_options, NULL, 1122 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1123 tp->t_inpcb); 1124 } 1125 if (error) { 1126 1127 /* 1128 * We know that the packet was lost, so back out the 1129 * sequence number advance, if any. 1130 * 1131 * If the error is EPERM the packet got blocked by the 1132 * local firewall. Normally we should terminate the 1133 * connection but the blocking may have been spurious 1134 * due to a firewall reconfiguration cycle. So we treat 1135 * it like a packet loss and let the retransmit timer and 1136 * timeouts do their work over time. 1137 * XXX: It is a POLA question whether calling tcp_drop right 1138 * away would be the really correct behavior instead. 1139 */ 1140 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1141 !tcp_timer_active(tp, TT_PERSIST)) && 1142 ((flags & TH_SYN) == 0) && 1143 (error != EPERM)) { 1144 if (sack_rxmit) { 1145 p->rxmit -= len; 1146 tp->sackhint.sack_bytes_rexmit -= len; 1147 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1148 ("sackhint bytes rtx >= 0")); 1149 } else 1150 tp->snd_nxt -= len; 1151 } 1152 out: 1153 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1154 switch (error) { 1155 case EPERM: 1156 tp->t_softerror = error; 1157 return (error); 1158 case ENOBUFS: 1159 if (!tcp_timer_active(tp, TT_REXMT) && 1160 !tcp_timer_active(tp, TT_PERSIST)) 1161 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1162 tp->snd_cwnd = tp->t_maxseg; 1163 return (0); 1164 case EMSGSIZE: 1165 /* 1166 * For some reason the interface we used initially 1167 * to send segments changed to another or lowered 1168 * its MTU. 1169 * 1170 * tcp_mtudisc() will find out the new MTU and as 1171 * its last action, initiate retransmission, so it 1172 * is important to not do so here. 1173 * 1174 * If TSO was active we either got an interface 1175 * without TSO capabilits or TSO was turned off. 1176 * Disable it for this connection as too and 1177 * immediatly retry with MSS sized segments generated 1178 * by this function. 1179 */ 1180 if (tso) 1181 tp->t_flags &= ~TF_TSO; 1182 tcp_mtudisc(tp->t_inpcb, 0); 1183 return (0); 1184 case EHOSTDOWN: 1185 case EHOSTUNREACH: 1186 case ENETDOWN: 1187 case ENETUNREACH: 1188 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1189 tp->t_softerror = error; 1190 return (0); 1191 } 1192 /* FALLTHROUGH */ 1193 default: 1194 return (error); 1195 } 1196 } 1197 tcpstat.tcps_sndtotal++; 1198 1199 /* 1200 * Data sent (as far as we can tell). 1201 * If this advertises a larger window than any other segment, 1202 * then remember the size of the advertised window. 1203 * Any pending ACK has now been sent. 1204 */ 1205 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1206 tp->rcv_adv = tp->rcv_nxt + recwin; 1207 tp->last_ack_sent = tp->rcv_nxt; 1208 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1209 if (tcp_timer_active(tp, TT_DELACK)) 1210 tcp_timer_activate(tp, TT_DELACK, 0); 1211 #if 0 1212 /* 1213 * This completely breaks TCP if newreno is turned on. What happens 1214 * is that if delayed-acks are turned on on the receiver, this code 1215 * on the transmitter effectively destroys the TCP window, forcing 1216 * it to four packets (1.5Kx4 = 6K window). 1217 */ 1218 if (sendalot && (!tcp_do_newreno || --maxburst)) 1219 goto again; 1220 #endif 1221 if (sendalot) 1222 goto again; 1223 return (0); 1224 } 1225 1226 void 1227 tcp_setpersist(struct tcpcb *tp) 1228 { 1229 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1230 int tt; 1231 1232 if (tcp_timer_active(tp, TT_REXMT)) 1233 panic("tcp_setpersist: retransmit pending"); 1234 /* 1235 * Start/restart persistance timer. 1236 */ 1237 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1238 TCPTV_PERSMIN, TCPTV_PERSMAX); 1239 tcp_timer_activate(tp, TT_PERSIST, tt); 1240 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1241 tp->t_rxtshift++; 1242 } 1243 1244 /* 1245 * Insert TCP options according to the supplied parameters to the place 1246 * optp in a consistent way. Can handle unaligned destinations. 1247 * 1248 * The order of the option processing is crucial for optimal packing and 1249 * alignment for the scarce option space. 1250 * 1251 * The optimal order for a SYN/SYN-ACK segment is: 1252 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1253 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1254 * 1255 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1256 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1257 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1258 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1259 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1260 */ 1261 int 1262 tcp_addoptions(struct tcpopt *to, u_char *optp) 1263 { 1264 u_int mask, optlen = 0; 1265 1266 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1267 if ((to->to_flags & mask) != mask) 1268 continue; 1269 switch (to->to_flags & mask) { 1270 case TOF_MSS: 1271 while (optlen % 4) { 1272 optlen += TCPOLEN_NOP; 1273 *optp++ = TCPOPT_NOP; 1274 } 1275 optlen += TCPOLEN_MAXSEG; 1276 *optp++ = TCPOPT_MAXSEG; 1277 *optp++ = TCPOLEN_MAXSEG; 1278 to->to_mss = htons(to->to_mss); 1279 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1280 optp += sizeof(to->to_mss); 1281 break; 1282 case TOF_SCALE: 1283 while (!optlen || optlen % 2 != 1) { 1284 optlen += TCPOLEN_NOP; 1285 *optp++ = TCPOPT_NOP; 1286 } 1287 optlen += TCPOLEN_WINDOW; 1288 *optp++ = TCPOPT_WINDOW; 1289 *optp++ = TCPOLEN_WINDOW; 1290 *optp++ = to->to_wscale; 1291 break; 1292 case TOF_SACKPERM: 1293 while (optlen % 2) { 1294 optlen += TCPOLEN_NOP; 1295 *optp++ = TCPOPT_NOP; 1296 } 1297 optlen += TCPOLEN_SACK_PERMITTED; 1298 *optp++ = TCPOPT_SACK_PERMITTED; 1299 *optp++ = TCPOLEN_SACK_PERMITTED; 1300 break; 1301 case TOF_TS: 1302 while (!optlen || optlen % 4 != 2) { 1303 optlen += TCPOLEN_NOP; 1304 *optp++ = TCPOPT_NOP; 1305 } 1306 optlen += TCPOLEN_TIMESTAMP; 1307 *optp++ = TCPOPT_TIMESTAMP; 1308 *optp++ = TCPOLEN_TIMESTAMP; 1309 to->to_tsval = htonl(to->to_tsval); 1310 to->to_tsecr = htonl(to->to_tsecr); 1311 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1312 optp += sizeof(to->to_tsval); 1313 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1314 optp += sizeof(to->to_tsecr); 1315 break; 1316 case TOF_SIGNATURE: 1317 { 1318 int siglen = TCPOLEN_SIGNATURE - 2; 1319 1320 while (!optlen || optlen % 4 != 2) { 1321 optlen += TCPOLEN_NOP; 1322 *optp++ = TCPOPT_NOP; 1323 } 1324 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1325 continue; 1326 optlen += TCPOLEN_SIGNATURE; 1327 *optp++ = TCPOPT_SIGNATURE; 1328 *optp++ = TCPOLEN_SIGNATURE; 1329 to->to_signature = optp; 1330 while (siglen--) 1331 *optp++ = 0; 1332 break; 1333 } 1334 case TOF_SACK: 1335 { 1336 int sackblks = 0; 1337 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1338 tcp_seq sack_seq; 1339 1340 while (!optlen || optlen % 4 != 2) { 1341 optlen += TCPOLEN_NOP; 1342 *optp++ = TCPOPT_NOP; 1343 } 1344 if (TCP_MAXOLEN - optlen < 2 + TCPOLEN_SACK) 1345 continue; 1346 optlen += TCPOLEN_SACKHDR; 1347 *optp++ = TCPOPT_SACK; 1348 sackblks = min(to->to_nsacks, 1349 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1350 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1351 while (sackblks--) { 1352 sack_seq = htonl(sack->start); 1353 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1354 optp += sizeof(sack_seq); 1355 sack_seq = htonl(sack->end); 1356 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1357 optp += sizeof(sack_seq); 1358 optlen += TCPOLEN_SACK; 1359 sack++; 1360 } 1361 tcpstat.tcps_sack_send_blocks++; 1362 break; 1363 } 1364 default: 1365 panic("%s: unknown TCP option type", __func__); 1366 break; 1367 } 1368 } 1369 1370 /* Terminate and pad TCP options to a 4 byte boundary. */ 1371 if (optlen % 4) { 1372 optlen += TCPOLEN_EOL; 1373 *optp++ = TCPOPT_EOL; 1374 } 1375 while (optlen % 4) { 1376 optlen += TCPOLEN_NOP; 1377 *optp++ = TCPOPT_NOP; 1378 } 1379 1380 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1381 return (optlen); 1382 } 1383