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 <netipsec/ipsec.h> 77 #endif /*IPSEC*/ 78 79 #include <machine/in_cksum.h> 80 81 #include <security/mac/mac_framework.h> 82 83 #ifdef notyet 84 extern struct mbuf *m_copypack(); 85 #endif 86 87 int path_mtu_discovery = 1; 88 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 89 &path_mtu_discovery, 1, "Enable Path MTU Discovery"); 90 91 int ss_fltsz = 1; 92 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW, 93 &ss_fltsz, 1, "Slow start flight size"); 94 95 int ss_fltsz_local = 4; 96 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW, 97 &ss_fltsz_local, 1, "Slow start flight size for local networks"); 98 99 int tcp_do_newreno = 1; 100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, 101 &tcp_do_newreno, 0, "Enable NewReno Algorithms"); 102 103 int tcp_do_tso = 1; 104 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW, 105 &tcp_do_tso, 0, "Enable TCP Segmentation Offload"); 106 107 int tcp_do_autosndbuf = 1; 108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 109 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing"); 110 111 int tcp_autosndbuf_inc = 8*1024; 112 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 113 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer"); 114 115 int tcp_autosndbuf_max = 256*1024; 116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 117 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer"); 118 119 120 /* 121 * Tcp output routine: figure out what should be sent and send it. 122 */ 123 int 124 tcp_output(struct tcpcb *tp) 125 { 126 struct socket *so = tp->t_inpcb->inp_socket; 127 long len, recwin, sendwin; 128 int off, flags, error; 129 #ifdef TCP_SIGNATURE 130 int sigoff = 0; 131 #endif 132 struct mbuf *m; 133 struct ip *ip = NULL; 134 struct ipovly *ipov = NULL; 135 struct tcphdr *th; 136 u_char opt[TCP_MAXOLEN]; 137 unsigned ipoptlen, optlen, hdrlen; 138 int idle, sendalot; 139 int sack_rxmit, sack_bytes_rxmt; 140 struct sackhole *p; 141 int tso = 0; 142 struct tcpopt to; 143 #if 0 144 int maxburst = TCP_MAXBURST; 145 #endif 146 #ifdef INET6 147 struct ip6_hdr *ip6 = NULL; 148 int isipv6; 149 150 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 151 #endif 152 153 INP_LOCK_ASSERT(tp->t_inpcb); 154 155 /* 156 * Determine length of data that should be transmitted, 157 * and flags that will be used. 158 * If there is some data or critical controls (SYN, RST) 159 * to send, then transmit; otherwise, investigate further. 160 */ 161 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 162 if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 163 /* 164 * We have been idle for "a while" and no acks are 165 * expected to clock out any data we send -- 166 * slow start to get ack "clock" running again. 167 * 168 * Set the slow-start flight size depending on whether 169 * this is a local network or not. 170 */ 171 int ss = ss_fltsz; 172 #ifdef INET6 173 if (isipv6) { 174 if (in6_localaddr(&tp->t_inpcb->in6p_faddr)) 175 ss = ss_fltsz_local; 176 } else 177 #endif /* INET6 */ 178 if (in_localaddr(tp->t_inpcb->inp_faddr)) 179 ss = ss_fltsz_local; 180 tp->snd_cwnd = tp->t_maxseg * ss; 181 } 182 tp->t_flags &= ~TF_LASTIDLE; 183 if (idle) { 184 if (tp->t_flags & TF_MORETOCOME) { 185 tp->t_flags |= TF_LASTIDLE; 186 idle = 0; 187 } 188 } 189 again: 190 /* 191 * If we've recently taken a timeout, snd_max will be greater than 192 * snd_nxt. There may be SACK information that allows us to avoid 193 * resending already delivered data. Adjust snd_nxt accordingly. 194 */ 195 if ((tp->t_flags & TF_SACK_PERMIT) && 196 SEQ_LT(tp->snd_nxt, tp->snd_max)) 197 tcp_sack_adjust(tp); 198 sendalot = 0; 199 off = tp->snd_nxt - tp->snd_una; 200 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 201 sendwin = min(sendwin, tp->snd_bwnd); 202 203 flags = tcp_outflags[tp->t_state]; 204 /* 205 * Send any SACK-generated retransmissions. If we're explicitly trying 206 * to send out new data (when sendalot is 1), bypass this function. 207 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 208 * we're replacing a (future) new transmission with a retransmission 209 * now, and we previously incremented snd_cwnd in tcp_input(). 210 */ 211 /* 212 * Still in sack recovery , reset rxmit flag to zero. 213 */ 214 sack_rxmit = 0; 215 sack_bytes_rxmt = 0; 216 len = 0; 217 p = NULL; 218 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) && 219 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 220 long cwin; 221 222 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 223 if (cwin < 0) 224 cwin = 0; 225 /* Do not retransmit SACK segments beyond snd_recover */ 226 if (SEQ_GT(p->end, tp->snd_recover)) { 227 /* 228 * (At least) part of sack hole extends beyond 229 * snd_recover. Check to see if we can rexmit data 230 * for this hole. 231 */ 232 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 233 /* 234 * Can't rexmit any more data for this hole. 235 * That data will be rexmitted in the next 236 * sack recovery episode, when snd_recover 237 * moves past p->rxmit. 238 */ 239 p = NULL; 240 goto after_sack_rexmit; 241 } else 242 /* Can rexmit part of the current hole */ 243 len = ((long)ulmin(cwin, 244 tp->snd_recover - p->rxmit)); 245 } else 246 len = ((long)ulmin(cwin, p->end - p->rxmit)); 247 off = p->rxmit - tp->snd_una; 248 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 249 __func__, off)); 250 if (len > 0) { 251 sack_rxmit = 1; 252 sendalot = 1; 253 tcpstat.tcps_sack_rexmits++; 254 tcpstat.tcps_sack_rexmit_bytes += 255 min(len, tp->t_maxseg); 256 } 257 } 258 after_sack_rexmit: 259 /* 260 * Get standard flags, and add SYN or FIN if requested by 'hidden' 261 * state flags. 262 */ 263 if (tp->t_flags & TF_NEEDFIN) 264 flags |= TH_FIN; 265 if (tp->t_flags & TF_NEEDSYN) 266 flags |= TH_SYN; 267 268 SOCKBUF_LOCK(&so->so_snd); 269 /* 270 * If in persist timeout with window of 0, send 1 byte. 271 * Otherwise, if window is small but nonzero 272 * and timer expired, we will send what we can 273 * and go to transmit state. 274 */ 275 if (tp->t_flags & TF_FORCEDATA) { 276 if (sendwin == 0) { 277 /* 278 * If we still have some data to send, then 279 * clear the FIN bit. Usually this would 280 * happen below when it realizes that we 281 * aren't sending all the data. However, 282 * if we have exactly 1 byte of unsent data, 283 * then it won't clear the FIN bit below, 284 * and if we are in persist state, we wind 285 * up sending the packet without recording 286 * that we sent the FIN bit. 287 * 288 * We can't just blindly clear the FIN bit, 289 * because if we don't have any more data 290 * to send then the probe will be the FIN 291 * itself. 292 */ 293 if (off < so->so_snd.sb_cc) 294 flags &= ~TH_FIN; 295 sendwin = 1; 296 } else { 297 tcp_timer_activate(tp, TT_PERSIST, 0); 298 tp->t_rxtshift = 0; 299 } 300 } 301 302 /* 303 * If snd_nxt == snd_max and we have transmitted a FIN, the 304 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 305 * a negative length. This can also occur when TCP opens up 306 * its congestion window while receiving additional duplicate 307 * acks after fast-retransmit because TCP will reset snd_nxt 308 * to snd_max after the fast-retransmit. 309 * 310 * In the normal retransmit-FIN-only case, however, snd_nxt will 311 * be set to snd_una, the offset will be 0, and the length may 312 * wind up 0. 313 * 314 * If sack_rxmit is true we are retransmitting from the scoreboard 315 * in which case len is already set. 316 */ 317 if (sack_rxmit == 0) { 318 if (sack_bytes_rxmt == 0) 319 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off); 320 else { 321 long cwin; 322 323 /* 324 * We are inside of a SACK recovery episode and are 325 * sending new data, having retransmitted all the 326 * data possible in the scoreboard. 327 */ 328 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 329 - off); 330 /* 331 * Don't remove this (len > 0) check ! 332 * We explicitly check for len > 0 here (although it 333 * isn't really necessary), to work around a gcc 334 * optimization issue - to force gcc to compute 335 * len above. Without this check, the computation 336 * of len is bungled by the optimizer. 337 */ 338 if (len > 0) { 339 cwin = tp->snd_cwnd - 340 (tp->snd_nxt - tp->sack_newdata) - 341 sack_bytes_rxmt; 342 if (cwin < 0) 343 cwin = 0; 344 len = lmin(len, cwin); 345 } 346 } 347 } 348 349 /* 350 * Lop off SYN bit if it has already been sent. However, if this 351 * is SYN-SENT state and if segment contains data and if we don't 352 * know that foreign host supports TAO, suppress sending segment. 353 */ 354 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 355 if (tp->t_state != TCPS_SYN_RECEIVED) 356 flags &= ~TH_SYN; 357 off--, len++; 358 } 359 360 /* 361 * Be careful not to send data and/or FIN on SYN segments. 362 * This measure is needed to prevent interoperability problems 363 * with not fully conformant TCP implementations. 364 */ 365 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 366 len = 0; 367 flags &= ~TH_FIN; 368 } 369 370 if (len < 0) { 371 /* 372 * If FIN has been sent but not acked, 373 * but we haven't been called to retransmit, 374 * len will be < 0. Otherwise, window shrank 375 * after we sent into it. If window shrank to 0, 376 * cancel pending retransmit, pull snd_nxt back 377 * to (closed) window, and set the persist timer 378 * if it isn't already going. If the window didn't 379 * close completely, just wait for an ACK. 380 */ 381 len = 0; 382 if (sendwin == 0) { 383 tcp_timer_activate(tp, TT_REXMT, 0); 384 tp->t_rxtshift = 0; 385 tp->snd_nxt = tp->snd_una; 386 if (!tcp_timer_active(tp, TT_PERSIST)) 387 tcp_setpersist(tp); 388 } 389 } 390 391 /* len will be >= 0 after this point. */ 392 KASSERT(len >= 0, ("%s: len < 0", __func__)); 393 394 /* 395 * Automatic sizing of send socket buffer. Often the send buffer 396 * size is not optimally adjusted to the actual network conditions 397 * at hand (delay bandwidth product). Setting the buffer size too 398 * small limits throughput on links with high bandwidth and high 399 * delay (eg. trans-continental/oceanic links). Setting the 400 * buffer size too big consumes too much real kernel memory, 401 * especially with many connections on busy servers. 402 * 403 * The criteria to step up the send buffer one notch are: 404 * 1. receive window of remote host is larger than send buffer 405 * (with a fudge factor of 5/4th); 406 * 2. send buffer is filled to 7/8th with data (so we actually 407 * have data to make use of it); 408 * 3. send buffer fill has not hit maximal automatic size; 409 * 4. our send window (slow start and cogestion controlled) is 410 * larger than sent but unacknowledged data in send buffer. 411 * 412 * The remote host receive window scaling factor may limit the 413 * growing of the send buffer before it reaches its allowed 414 * maximum. 415 * 416 * It scales directly with slow start or congestion window 417 * and does at most one step per received ACK. This fast 418 * scaling has the drawback of growing the send buffer beyond 419 * what is strictly necessary to make full use of a given 420 * delay*bandwith product. However testing has shown this not 421 * to be much of an problem. At worst we are trading wasting 422 * of available bandwith (the non-use of it) for wasting some 423 * socket buffer memory. 424 * 425 * TODO: Shrink send buffer during idle periods together 426 * with congestion window. Requires another timer. Has to 427 * wait for upcoming tcp timer rewrite. 428 */ 429 if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 430 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 431 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 432 so->so_snd.sb_cc < tcp_autosndbuf_max && 433 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 434 if (!sbreserve_locked(&so->so_snd, 435 min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc, 436 tcp_autosndbuf_max), so, curthread)) 437 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 438 } 439 } 440 441 /* 442 * Truncate to the maximum segment length or enable TCP Segmentation 443 * Offloading (if supported by hardware) and ensure that FIN is removed 444 * if the length no longer contains the last data byte. 445 * 446 * TSO may only be used if we are in a pure bulk sending state. The 447 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 448 * IP options prevent using TSO. With TSO the TCP header is the same 449 * (except for the sequence number) for all generated packets. This 450 * makes it impossible to transmit any options which vary per generated 451 * segment or packet. 452 * 453 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and 454 * removal of FIN (if not already catched here) are handled later after 455 * the exact length of the TCP options are known. 456 */ 457 if (len > tp->t_maxseg) { 458 if ((tp->t_flags & TF_TSO) && tcp_do_tso && 459 ((tp->t_flags & TF_SIGNATURE) == 0) && 460 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 461 tp->t_inpcb->inp_options == NULL && 462 tp->t_inpcb->in6p_options == NULL && 463 tp->t_inpcb->inp_sp == NULL) { 464 tso = 1; 465 } else { 466 len = tp->t_maxseg; 467 sendalot = 1; 468 tso = 0; 469 } 470 } 471 if (sack_rxmit) { 472 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 473 flags &= ~TH_FIN; 474 } else { 475 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 476 flags &= ~TH_FIN; 477 } 478 479 recwin = sbspace(&so->so_rcv); 480 481 /* 482 * Sender silly window avoidance. We transmit under the following 483 * conditions when len is non-zero: 484 * 485 * - We have a full segment (or more with TSO) 486 * - This is the last buffer in a write()/send() and we are 487 * either idle or running NODELAY 488 * - we've timed out (e.g. persist timer) 489 * - we have more then 1/2 the maximum send window's worth of 490 * data (receiver may be limited the window size) 491 * - we need to retransmit 492 */ 493 if (len) { 494 if (len >= tp->t_maxseg) 495 goto send; 496 /* 497 * NOTE! on localhost connections an 'ack' from the remote 498 * end may occur synchronously with the output and cause 499 * us to flush a buffer queued with moretocome. XXX 500 * 501 * note: the len + off check is almost certainly unnecessary. 502 */ 503 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 504 (idle || (tp->t_flags & TF_NODELAY)) && 505 len + off >= so->so_snd.sb_cc && 506 (tp->t_flags & TF_NOPUSH) == 0) { 507 goto send; 508 } 509 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 510 goto send; 511 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 512 goto send; 513 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 514 goto send; 515 if (sack_rxmit) 516 goto send; 517 } 518 519 /* 520 * Compare available window to amount of window 521 * known to peer (as advertised window less 522 * next expected input). If the difference is at least two 523 * max size segments, or at least 50% of the maximum possible 524 * window, then want to send a window update to peer. 525 * Skip this if the connection is in T/TCP half-open state. 526 * Don't send pure window updates when the peer has closed 527 * the connection and won't ever send more data. 528 */ 529 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 530 !TCPS_HAVERCVDFIN(tp->t_state)) { 531 /* 532 * "adv" is the amount we can increase the window, 533 * taking into account that we are limited by 534 * TCP_MAXWIN << tp->rcv_scale. 535 */ 536 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) - 537 (tp->rcv_adv - tp->rcv_nxt); 538 539 if (adv >= (long) (2 * tp->t_maxseg)) 540 goto send; 541 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 542 goto send; 543 } 544 545 /* 546 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 547 * is also a catch-all for the retransmit timer timeout case. 548 */ 549 if (tp->t_flags & TF_ACKNOW) 550 goto send; 551 if ((flags & TH_RST) || 552 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 553 goto send; 554 if (SEQ_GT(tp->snd_up, tp->snd_una)) 555 goto send; 556 /* 557 * If our state indicates that FIN should be sent 558 * and we have not yet done so, then we need to send. 559 */ 560 if (flags & TH_FIN && 561 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 562 goto send; 563 /* 564 * In SACK, it is possible for tcp_output to fail to send a segment 565 * after the retransmission timer has been turned off. Make sure 566 * that the retransmission timer is set. 567 */ 568 if ((tp->t_flags & TF_SACK_PERMIT) && 569 SEQ_GT(tp->snd_max, tp->snd_una) && 570 !tcp_timer_active(tp, TT_REXMT) && 571 !tcp_timer_active(tp, TT_PERSIST)) { 572 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 573 goto just_return; 574 } 575 /* 576 * TCP window updates are not reliable, rather a polling protocol 577 * using ``persist'' packets is used to insure receipt of window 578 * updates. The three ``states'' for the output side are: 579 * idle not doing retransmits or persists 580 * persisting to move a small or zero window 581 * (re)transmitting and thereby not persisting 582 * 583 * tcp_timer_active(tp, TT_PERSIST) 584 * is true when we are in persist state. 585 * (tp->t_flags & TF_FORCEDATA) 586 * is set when we are called to send a persist packet. 587 * tcp_timer_active(tp, TT_REXMT) 588 * is set when we are retransmitting 589 * The output side is idle when both timers are zero. 590 * 591 * If send window is too small, there is data to transmit, and no 592 * retransmit or persist is pending, then go to persist state. 593 * If nothing happens soon, send when timer expires: 594 * if window is nonzero, transmit what we can, 595 * otherwise force out a byte. 596 */ 597 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) && 598 !tcp_timer_active(tp, TT_PERSIST)) { 599 tp->t_rxtshift = 0; 600 tcp_setpersist(tp); 601 } 602 603 /* 604 * No reason to send a segment, just return. 605 */ 606 just_return: 607 SOCKBUF_UNLOCK(&so->so_snd); 608 return (0); 609 610 send: 611 SOCKBUF_LOCK_ASSERT(&so->so_snd); 612 /* 613 * Before ESTABLISHED, force sending of initial options 614 * unless TCP set not to do any options. 615 * NOTE: we assume that the IP/TCP header plus TCP options 616 * always fit in a single mbuf, leaving room for a maximum 617 * link header, i.e. 618 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 619 */ 620 optlen = 0; 621 #ifdef INET6 622 if (isipv6) 623 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 624 else 625 #endif 626 hdrlen = sizeof (struct tcpiphdr); 627 628 /* 629 * Compute options for segment. 630 * We only have to care about SYN and established connection 631 * segments. Options for SYN-ACK segments are handled in TCP 632 * syncache. 633 */ 634 if ((tp->t_flags & TF_NOOPT) == 0) { 635 to.to_flags = 0; 636 /* Maximum segment size. */ 637 if (flags & TH_SYN) { 638 tp->snd_nxt = tp->iss; 639 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 640 to.to_flags |= TOF_MSS; 641 } 642 /* Window scaling. */ 643 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 644 to.to_wscale = tp->request_r_scale; 645 to.to_flags |= TOF_SCALE; 646 } 647 /* Timestamps. */ 648 if ((tp->t_flags & TF_RCVD_TSTMP) || 649 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 650 to.to_tsval = ticks + tp->ts_offset; 651 to.to_tsecr = tp->ts_recent; 652 to.to_flags |= TOF_TS; 653 /* Set receive buffer autosizing timestamp. */ 654 if (tp->rfbuf_ts == 0 && 655 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 656 tp->rfbuf_ts = ticks; 657 } 658 /* Selective ACK's. */ 659 if (tp->t_flags & TF_SACK_PERMIT) { 660 if (flags & TH_SYN) 661 to.to_flags |= TOF_SACKPERM; 662 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 663 (tp->t_flags & TF_SACK_PERMIT) && 664 tp->rcv_numsacks > 0) { 665 to.to_flags |= TOF_SACK; 666 to.to_nsacks = tp->rcv_numsacks; 667 to.to_sacks = (u_char *)tp->sackblks; 668 } 669 } 670 #ifdef TCP_SIGNATURE 671 /* TCP-MD5 (RFC2385). */ 672 #ifdef INET6 673 if (!isipv6 && (tp->t_flags & TF_SIGNATURE)) 674 #else 675 if (tp->t_flags & TF_SIGNATURE) 676 #endif /* INET6 */ 677 to.to_flags |= TOF_SIGNATURE; 678 #endif /* TCP_SIGNATURE */ 679 680 /* Processing the options. */ 681 hdrlen += optlen = tcp_addoptions(&to, (u_char *)&opt); 682 683 #ifdef TCP_SIGNATURE 684 sigoff = to.to_signature - (u_char *)&to; 685 #endif /* TCP_SIGNATURE */ 686 } 687 688 #ifdef INET6 689 if (isipv6) 690 ipoptlen = ip6_optlen(tp->t_inpcb); 691 else 692 #endif 693 if (tp->t_inpcb->inp_options) 694 ipoptlen = tp->t_inpcb->inp_options->m_len - 695 offsetof(struct ipoption, ipopt_list); 696 else 697 ipoptlen = 0; 698 #ifdef IPSEC 699 ipoptlen += ipsec_hdrsiz_tcp(tp); 700 #endif 701 702 /* 703 * Adjust data length if insertion of options will 704 * bump the packet length beyond the t_maxopd length. 705 * Clear the FIN bit because we cut off the tail of 706 * the segment. 707 * 708 * When doing TSO limit a burst to TCP_MAXWIN minus the 709 * IP, TCP and Options length to keep ip->ip_len from 710 * overflowing. Prevent the last segment from being 711 * fractional thus making them all equal sized and set 712 * the flag to continue sending. TSO is disabled when 713 * IP options or IPSEC are present. 714 */ 715 if (len + optlen + ipoptlen > tp->t_maxopd) { 716 flags &= ~TH_FIN; 717 if (tso) { 718 if (len > TCP_MAXWIN - hdrlen - optlen) { 719 len = TCP_MAXWIN - hdrlen - optlen; 720 len = len - (len % (tp->t_maxopd - optlen)); 721 sendalot = 1; 722 } else if (tp->t_flags & TF_NEEDFIN) 723 sendalot = 1; 724 } else { 725 len = tp->t_maxopd - optlen - ipoptlen; 726 sendalot = 1; 727 } 728 } 729 730 /*#ifdef DIAGNOSTIC*/ 731 #ifdef INET6 732 if (max_linkhdr + hdrlen > MCLBYTES) 733 #else 734 if (max_linkhdr + hdrlen > MHLEN) 735 #endif 736 panic("tcphdr too big"); 737 /*#endif*/ 738 739 /* 740 * Grab a header mbuf, attaching a copy of data to 741 * be transmitted, and initialize the header from 742 * the template for sends on this connection. 743 */ 744 if (len) { 745 struct mbuf *mb; 746 u_int moff; 747 748 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 749 tcpstat.tcps_sndprobe++; 750 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 751 tcpstat.tcps_sndrexmitpack++; 752 tcpstat.tcps_sndrexmitbyte += len; 753 } else { 754 tcpstat.tcps_sndpack++; 755 tcpstat.tcps_sndbyte += len; 756 } 757 #ifdef notyet 758 if ((m = m_copypack(so->so_snd.sb_mb, off, 759 (int)len, max_linkhdr + hdrlen)) == 0) { 760 SOCKBUF_UNLOCK(&so->so_snd); 761 error = ENOBUFS; 762 goto out; 763 } 764 /* 765 * m_copypack left space for our hdr; use it. 766 */ 767 m->m_len += hdrlen; 768 m->m_data -= hdrlen; 769 #else 770 MGETHDR(m, M_DONTWAIT, MT_DATA); 771 if (m == NULL) { 772 SOCKBUF_UNLOCK(&so->so_snd); 773 error = ENOBUFS; 774 goto out; 775 } 776 #ifdef INET6 777 if (MHLEN < hdrlen + max_linkhdr) { 778 MCLGET(m, M_DONTWAIT); 779 if ((m->m_flags & M_EXT) == 0) { 780 SOCKBUF_UNLOCK(&so->so_snd); 781 m_freem(m); 782 error = ENOBUFS; 783 goto out; 784 } 785 } 786 #endif 787 m->m_data += max_linkhdr; 788 m->m_len = hdrlen; 789 790 /* 791 * Start the m_copy functions from the closest mbuf 792 * to the offset in the socket buffer chain. 793 */ 794 mb = sbsndptr(&so->so_snd, off, len, &moff); 795 796 if (len <= MHLEN - hdrlen - max_linkhdr) { 797 m_copydata(mb, moff, (int)len, 798 mtod(m, caddr_t) + hdrlen); 799 m->m_len += len; 800 } else { 801 m->m_next = m_copy(mb, moff, (int)len); 802 if (m->m_next == NULL) { 803 SOCKBUF_UNLOCK(&so->so_snd); 804 (void) m_free(m); 805 error = ENOBUFS; 806 goto out; 807 } 808 } 809 #endif 810 /* 811 * If we're sending everything we've got, set PUSH. 812 * (This will keep happy those implementations which only 813 * give data to the user when a buffer fills or 814 * a PUSH comes in.) 815 */ 816 if (off + len == so->so_snd.sb_cc) 817 flags |= TH_PUSH; 818 SOCKBUF_UNLOCK(&so->so_snd); 819 } else { 820 SOCKBUF_UNLOCK(&so->so_snd); 821 if (tp->t_flags & TF_ACKNOW) 822 tcpstat.tcps_sndacks++; 823 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 824 tcpstat.tcps_sndctrl++; 825 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 826 tcpstat.tcps_sndurg++; 827 else 828 tcpstat.tcps_sndwinup++; 829 830 MGETHDR(m, M_DONTWAIT, MT_DATA); 831 if (m == NULL) { 832 error = ENOBUFS; 833 goto out; 834 } 835 #ifdef INET6 836 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 837 MHLEN >= hdrlen) { 838 MH_ALIGN(m, hdrlen); 839 } else 840 #endif 841 m->m_data += max_linkhdr; 842 m->m_len = hdrlen; 843 } 844 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 845 m->m_pkthdr.rcvif = (struct ifnet *)0; 846 #ifdef MAC 847 mac_create_mbuf_from_inpcb(tp->t_inpcb, m); 848 #endif 849 #ifdef INET6 850 if (isipv6) { 851 ip6 = mtod(m, struct ip6_hdr *); 852 th = (struct tcphdr *)(ip6 + 1); 853 tcpip_fillheaders(tp->t_inpcb, ip6, th); 854 } else 855 #endif /* INET6 */ 856 { 857 ip = mtod(m, struct ip *); 858 ipov = (struct ipovly *)ip; 859 th = (struct tcphdr *)(ip + 1); 860 tcpip_fillheaders(tp->t_inpcb, ip, th); 861 } 862 863 /* 864 * Fill in fields, remembering maximum advertised 865 * window for use in delaying messages about window sizes. 866 * If resending a FIN, be sure not to use a new sequence number. 867 */ 868 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 869 tp->snd_nxt == tp->snd_max) 870 tp->snd_nxt--; 871 /* 872 * If we are doing retransmissions, then snd_nxt will 873 * not reflect the first unsent octet. For ACK only 874 * packets, we do not want the sequence number of the 875 * retransmitted packet, we want the sequence number 876 * of the next unsent octet. So, if there is no data 877 * (and no SYN or FIN), use snd_max instead of snd_nxt 878 * when filling in ti_seq. But if we are in persist 879 * state, snd_max might reflect one byte beyond the 880 * right edge of the window, so use snd_nxt in that 881 * case, since we know we aren't doing a retransmission. 882 * (retransmit and persist are mutually exclusive...) 883 */ 884 if (sack_rxmit == 0) { 885 if (len || (flags & (TH_SYN|TH_FIN)) || 886 tcp_timer_active(tp, TT_PERSIST)) 887 th->th_seq = htonl(tp->snd_nxt); 888 else 889 th->th_seq = htonl(tp->snd_max); 890 } else { 891 th->th_seq = htonl(p->rxmit); 892 p->rxmit += len; 893 tp->sackhint.sack_bytes_rexmit += len; 894 } 895 th->th_ack = htonl(tp->rcv_nxt); 896 if (optlen) { 897 bcopy(opt, th + 1, optlen); 898 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 899 } 900 th->th_flags = flags; 901 /* 902 * Calculate receive window. Don't shrink window, 903 * but avoid silly window syndrome. 904 */ 905 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 906 recwin < (long)tp->t_maxseg) 907 recwin = 0; 908 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 909 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 910 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 911 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 912 913 /* 914 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 915 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 916 * case is handled in syncache. 917 */ 918 if (flags & TH_SYN) 919 th->th_win = htons((u_short) 920 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 921 else 922 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 923 924 /* 925 * Adjust the RXWIN0SENT flag - indicate that we have advertised 926 * a 0 window. This may cause the remote transmitter to stall. This 927 * flag tells soreceive() to disable delayed acknowledgements when 928 * draining the buffer. This can occur if the receiver is attempting 929 * to read more data then can be buffered prior to transmitting on 930 * the connection. 931 */ 932 if (recwin == 0) 933 tp->t_flags |= TF_RXWIN0SENT; 934 else 935 tp->t_flags &= ~TF_RXWIN0SENT; 936 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 937 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 938 th->th_flags |= TH_URG; 939 } else 940 /* 941 * If no urgent pointer to send, then we pull 942 * the urgent pointer to the left edge of the send window 943 * so that it doesn't drift into the send window on sequence 944 * number wraparound. 945 */ 946 tp->snd_up = tp->snd_una; /* drag it along */ 947 948 #ifdef TCP_SIGNATURE 949 #ifdef INET6 950 if (!isipv6) 951 #endif 952 if (tp->t_flags & TF_SIGNATURE) 953 tcp_signature_compute(m, sizeof(struct ip), len, optlen, 954 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 955 #endif 956 957 /* 958 * Put TCP length in extended header, and then 959 * checksum extended header and data. 960 */ 961 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 962 #ifdef INET6 963 if (isipv6) 964 /* 965 * ip6_plen is not need to be filled now, and will be filled 966 * in ip6_output. 967 */ 968 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 969 sizeof(struct tcphdr) + optlen + len); 970 else 971 #endif /* INET6 */ 972 { 973 m->m_pkthdr.csum_flags = CSUM_TCP; 974 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 975 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 976 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 977 978 /* IP version must be set here for ipv4/ipv6 checking later */ 979 KASSERT(ip->ip_v == IPVERSION, 980 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 981 } 982 983 /* 984 * Enable TSO and specify the size of the segments. 985 * The TCP pseudo header checksum is always provided. 986 * XXX: Fixme: This is currently not the case for IPv6. 987 */ 988 if (tso) { 989 m->m_pkthdr.csum_flags = CSUM_TSO; 990 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 991 } 992 993 /* 994 * In transmit state, time the transmission and arrange for 995 * the retransmit. In persist state, just set snd_max. 996 */ 997 if ((tp->t_flags & TF_FORCEDATA) == 0 || 998 !tcp_timer_active(tp, TT_PERSIST)) { 999 tcp_seq startseq = tp->snd_nxt; 1000 1001 /* 1002 * Advance snd_nxt over sequence space of this segment. 1003 */ 1004 if (flags & (TH_SYN|TH_FIN)) { 1005 if (flags & TH_SYN) 1006 tp->snd_nxt++; 1007 if (flags & TH_FIN) { 1008 tp->snd_nxt++; 1009 tp->t_flags |= TF_SENTFIN; 1010 } 1011 } 1012 if (sack_rxmit) 1013 goto timer; 1014 tp->snd_nxt += len; 1015 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1016 tp->snd_max = tp->snd_nxt; 1017 /* 1018 * Time this transmission if not a retransmission and 1019 * not currently timing anything. 1020 */ 1021 if (tp->t_rtttime == 0) { 1022 tp->t_rtttime = ticks; 1023 tp->t_rtseq = startseq; 1024 tcpstat.tcps_segstimed++; 1025 } 1026 } 1027 1028 /* 1029 * Set retransmit timer if not currently set, 1030 * and not doing a pure ack or a keep-alive probe. 1031 * Initial value for retransmit timer is smoothed 1032 * round-trip time + 2 * round-trip time variance. 1033 * Initialize shift counter which is used for backoff 1034 * of retransmit time. 1035 */ 1036 timer: 1037 if (!tcp_timer_active(tp, TT_REXMT) && 1038 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1039 (tp->snd_nxt != tp->snd_una))) { 1040 if (tcp_timer_active(tp, TT_PERSIST)) { 1041 tcp_timer_activate(tp, TT_PERSIST, 0); 1042 tp->t_rxtshift = 0; 1043 } 1044 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1045 } 1046 } else { 1047 /* 1048 * Persist case, update snd_max but since we are in 1049 * persist mode (no window) we do not update snd_nxt. 1050 */ 1051 int xlen = len; 1052 if (flags & TH_SYN) 1053 ++xlen; 1054 if (flags & TH_FIN) { 1055 ++xlen; 1056 tp->t_flags |= TF_SENTFIN; 1057 } 1058 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1059 tp->snd_max = tp->snd_nxt + len; 1060 } 1061 1062 #ifdef TCPDEBUG 1063 /* 1064 * Trace. 1065 */ 1066 if (so->so_options & SO_DEBUG) { 1067 u_short save = 0; 1068 #ifdef INET6 1069 if (!isipv6) 1070 #endif 1071 { 1072 save = ipov->ih_len; 1073 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1074 } 1075 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1076 #ifdef INET6 1077 if (!isipv6) 1078 #endif 1079 ipov->ih_len = save; 1080 } 1081 #endif 1082 1083 /* 1084 * Fill in IP length and desired time to live and 1085 * send to IP level. There should be a better way 1086 * to handle ttl and tos; we could keep them in 1087 * the template, but need a way to checksum without them. 1088 */ 1089 /* 1090 * m->m_pkthdr.len should have been set before cksum calcuration, 1091 * because in6_cksum() need it. 1092 */ 1093 #ifdef INET6 1094 if (isipv6) { 1095 /* 1096 * we separately set hoplimit for every segment, since the 1097 * user might want to change the value via setsockopt. 1098 * Also, desired default hop limit might be changed via 1099 * Neighbor Discovery. 1100 */ 1101 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1102 1103 /* TODO: IPv6 IP6TOS_ECT bit on */ 1104 error = ip6_output(m, 1105 tp->t_inpcb->in6p_outputopts, NULL, 1106 ((so->so_options & SO_DONTROUTE) ? 1107 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb); 1108 } else 1109 #endif /* INET6 */ 1110 { 1111 ip->ip_len = m->m_pkthdr.len; 1112 #ifdef INET6 1113 if (INP_CHECK_SOCKAF(so, AF_INET6)) 1114 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1115 #endif /* INET6 */ 1116 /* 1117 * If we do path MTU discovery, then we set DF on every packet. 1118 * This might not be the best thing to do according to RFC3390 1119 * Section 2. However the tcp hostcache migitates the problem 1120 * so it affects only the first tcp connection with a host. 1121 */ 1122 if (path_mtu_discovery) 1123 ip->ip_off |= IP_DF; 1124 1125 error = ip_output(m, tp->t_inpcb->inp_options, NULL, 1126 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1127 tp->t_inpcb); 1128 } 1129 if (error) { 1130 1131 /* 1132 * We know that the packet was lost, so back out the 1133 * sequence number advance, if any. 1134 * 1135 * If the error is EPERM the packet got blocked by the 1136 * local firewall. Normally we should terminate the 1137 * connection but the blocking may have been spurious 1138 * due to a firewall reconfiguration cycle. So we treat 1139 * it like a packet loss and let the retransmit timer and 1140 * timeouts do their work over time. 1141 * XXX: It is a POLA question whether calling tcp_drop right 1142 * away would be the really correct behavior instead. 1143 */ 1144 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1145 !tcp_timer_active(tp, TT_PERSIST)) && 1146 ((flags & TH_SYN) == 0) && 1147 (error != EPERM)) { 1148 if (sack_rxmit) { 1149 p->rxmit -= len; 1150 tp->sackhint.sack_bytes_rexmit -= len; 1151 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1152 ("sackhint bytes rtx >= 0")); 1153 } else 1154 tp->snd_nxt -= len; 1155 } 1156 out: 1157 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1158 switch (error) { 1159 case EPERM: 1160 tp->t_softerror = error; 1161 return (error); 1162 case ENOBUFS: 1163 if (!tcp_timer_active(tp, TT_REXMT) && 1164 !tcp_timer_active(tp, TT_PERSIST)) 1165 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1166 tp->snd_cwnd = tp->t_maxseg; 1167 return (0); 1168 case EMSGSIZE: 1169 /* 1170 * For some reason the interface we used initially 1171 * to send segments changed to another or lowered 1172 * its MTU. 1173 * 1174 * tcp_mtudisc() will find out the new MTU and as 1175 * its last action, initiate retransmission, so it 1176 * is important to not do so here. 1177 * 1178 * If TSO was active we either got an interface 1179 * without TSO capabilits or TSO was turned off. 1180 * Disable it for this connection as too and 1181 * immediatly retry with MSS sized segments generated 1182 * by this function. 1183 */ 1184 if (tso) 1185 tp->t_flags &= ~TF_TSO; 1186 tcp_mtudisc(tp->t_inpcb, 0); 1187 return (0); 1188 case EHOSTDOWN: 1189 case EHOSTUNREACH: 1190 case ENETDOWN: 1191 case ENETUNREACH: 1192 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1193 tp->t_softerror = error; 1194 return (0); 1195 } 1196 /* FALLTHROUGH */ 1197 default: 1198 return (error); 1199 } 1200 } 1201 tcpstat.tcps_sndtotal++; 1202 1203 /* 1204 * Data sent (as far as we can tell). 1205 * If this advertises a larger window than any other segment, 1206 * then remember the size of the advertised window. 1207 * Any pending ACK has now been sent. 1208 */ 1209 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1210 tp->rcv_adv = tp->rcv_nxt + recwin; 1211 tp->last_ack_sent = tp->rcv_nxt; 1212 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1213 if (tcp_timer_active(tp, TT_DELACK)) 1214 tcp_timer_activate(tp, TT_DELACK, 0); 1215 #if 0 1216 /* 1217 * This completely breaks TCP if newreno is turned on. What happens 1218 * is that if delayed-acks are turned on on the receiver, this code 1219 * on the transmitter effectively destroys the TCP window, forcing 1220 * it to four packets (1.5Kx4 = 6K window). 1221 */ 1222 if (sendalot && (!tcp_do_newreno || --maxburst)) 1223 goto again; 1224 #endif 1225 if (sendalot) 1226 goto again; 1227 return (0); 1228 } 1229 1230 void 1231 tcp_setpersist(struct tcpcb *tp) 1232 { 1233 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1234 int tt; 1235 1236 if (tcp_timer_active(tp, TT_REXMT)) 1237 panic("tcp_setpersist: retransmit pending"); 1238 /* 1239 * Start/restart persistance timer. 1240 */ 1241 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1242 TCPTV_PERSMIN, TCPTV_PERSMAX); 1243 tcp_timer_activate(tp, TT_PERSIST, tt); 1244 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1245 tp->t_rxtshift++; 1246 } 1247 1248 /* 1249 * Insert TCP options according to the supplied parameters to the place 1250 * optp in a consistent way. Can handle unaligned destinations. 1251 * 1252 * The order of the option processing is crucial for optimal packing and 1253 * alignment for the scarce option space. 1254 * 1255 * The optimal order for a SYN/SYN-ACK segment is: 1256 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1257 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1258 * 1259 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1260 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1261 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1262 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1263 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1264 */ 1265 int 1266 tcp_addoptions(struct tcpopt *to, u_char *optp) 1267 { 1268 u_int mask, optlen = 0; 1269 1270 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1271 if ((to->to_flags & mask) != mask) 1272 continue; 1273 switch (to->to_flags & mask) { 1274 case TOF_MSS: 1275 while (optlen % 4) { 1276 optlen += TCPOLEN_NOP; 1277 *optp++ = TCPOPT_NOP; 1278 } 1279 optlen += TCPOLEN_MAXSEG; 1280 *optp++ = TCPOPT_MAXSEG; 1281 *optp++ = TCPOLEN_MAXSEG; 1282 to->to_mss = htons(to->to_mss); 1283 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1284 optp += sizeof(to->to_mss); 1285 break; 1286 case TOF_SCALE: 1287 while (!optlen || optlen % 2 != 1) { 1288 optlen += TCPOLEN_NOP; 1289 *optp++ = TCPOPT_NOP; 1290 } 1291 optlen += TCPOLEN_WINDOW; 1292 *optp++ = TCPOPT_WINDOW; 1293 *optp++ = TCPOLEN_WINDOW; 1294 *optp++ = to->to_wscale; 1295 break; 1296 case TOF_SACKPERM: 1297 while (optlen % 2) { 1298 optlen += TCPOLEN_NOP; 1299 *optp++ = TCPOPT_NOP; 1300 } 1301 optlen += TCPOLEN_SACK_PERMITTED; 1302 *optp++ = TCPOPT_SACK_PERMITTED; 1303 *optp++ = TCPOLEN_SACK_PERMITTED; 1304 break; 1305 case TOF_TS: 1306 while (!optlen || optlen % 4 != 2) { 1307 optlen += TCPOLEN_NOP; 1308 *optp++ = TCPOPT_NOP; 1309 } 1310 optlen += TCPOLEN_TIMESTAMP; 1311 *optp++ = TCPOPT_TIMESTAMP; 1312 *optp++ = TCPOLEN_TIMESTAMP; 1313 to->to_tsval = htonl(to->to_tsval); 1314 to->to_tsecr = htonl(to->to_tsecr); 1315 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1316 optp += sizeof(to->to_tsval); 1317 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1318 optp += sizeof(to->to_tsecr); 1319 break; 1320 case TOF_SIGNATURE: 1321 { 1322 int siglen = TCPOLEN_SIGNATURE - 2; 1323 1324 while (!optlen || optlen % 4 != 2) { 1325 optlen += TCPOLEN_NOP; 1326 *optp++ = TCPOPT_NOP; 1327 } 1328 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1329 continue; 1330 optlen += TCPOLEN_SIGNATURE; 1331 *optp++ = TCPOPT_SIGNATURE; 1332 *optp++ = TCPOLEN_SIGNATURE; 1333 to->to_signature = optp; 1334 while (siglen--) 1335 *optp++ = 0; 1336 break; 1337 } 1338 case TOF_SACK: 1339 { 1340 int sackblks = 0; 1341 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1342 tcp_seq sack_seq; 1343 1344 while (!optlen || optlen % 4 != 2) { 1345 optlen += TCPOLEN_NOP; 1346 *optp++ = TCPOPT_NOP; 1347 } 1348 if (TCP_MAXOLEN - optlen < 2 + TCPOLEN_SACK) 1349 continue; 1350 optlen += TCPOLEN_SACKHDR; 1351 *optp++ = TCPOPT_SACK; 1352 sackblks = min(to->to_nsacks, 1353 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1354 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1355 while (sackblks--) { 1356 sack_seq = htonl(sack->start); 1357 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1358 optp += sizeof(sack_seq); 1359 sack_seq = htonl(sack->end); 1360 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1361 optp += sizeof(sack_seq); 1362 optlen += TCPOLEN_SACK; 1363 sack++; 1364 } 1365 tcpstat.tcps_sack_send_blocks++; 1366 break; 1367 } 1368 default: 1369 panic("%s: unknown TCP option type", __func__); 1370 break; 1371 } 1372 } 1373 1374 /* Terminate and pad TCP options to a 4 byte boundary. */ 1375 if (optlen % 4) { 1376 optlen += TCPOLEN_EOL; 1377 *optp++ = TCPOPT_EOL; 1378 } 1379 while (optlen % 4) { 1380 optlen += TCPOLEN_NOP; 1381 *optp++ = TCPOPT_NOP; 1382 } 1383 1384 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1385 return (optlen); 1386 } 1387