1 /* 2 * Copyright (c) 2004 Jeffrey M. Hsu. All rights reserved. 3 * Copyright (c) 2004 The DragonFly Project. All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Jeffrey M. Hsu. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of The DragonFly Project nor the names of its 17 * contributors may be used to endorse or promote products derived 18 * from this software without specific, prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 36 * The Regents of the University of California. All rights reserved. 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. All advertising materials mentioning features or use of this software 47 * must display the following acknowledgement: 48 * This product includes software developed by the University of 49 * California, Berkeley and its contributors. 50 * 4. Neither the name of the University nor the names of its contributors 51 * may be used to endorse or promote products derived from this software 52 * without specific prior written permission. 53 * 54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 64 * SUCH DAMAGE. 65 * 66 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 67 * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.20 2003/01/29 22:45:36 hsu Exp $ 68 */ 69 70 #include "opt_inet.h" 71 #include "opt_inet6.h" 72 #include "opt_ipsec.h" 73 #include "opt_tcpdebug.h" 74 75 #include <sys/param.h> 76 #include <sys/systm.h> 77 #include <sys/kernel.h> 78 #include <sys/sysctl.h> 79 #include <sys/mbuf.h> 80 #include <sys/domain.h> 81 #include <sys/protosw.h> 82 #include <sys/socket.h> 83 #include <sys/socketvar.h> 84 #include <sys/in_cksum.h> 85 #include <sys/thread.h> 86 #include <sys/globaldata.h> 87 88 #include <net/route.h> 89 90 #include <netinet/in.h> 91 #include <netinet/in_systm.h> 92 #include <netinet/ip.h> 93 #include <netinet/in_pcb.h> 94 #include <netinet/ip_var.h> 95 #include <netinet6/in6_pcb.h> 96 #include <netinet/ip6.h> 97 #include <netinet6/ip6_var.h> 98 #include <netinet/tcp.h> 99 #define TCPOUTFLAGS 100 #include <netinet/tcp_fsm.h> 101 #include <netinet/tcp_seq.h> 102 #include <netinet/tcp_timer.h> 103 #include <netinet/tcp_timer2.h> 104 #include <netinet/tcp_var.h> 105 #include <netinet/tcpip.h> 106 #ifdef TCPDEBUG 107 #include <netinet/tcp_debug.h> 108 #endif 109 110 #ifdef IPSEC 111 #include <netinet6/ipsec.h> 112 #endif /*IPSEC*/ 113 114 #ifdef FAST_IPSEC 115 #include <netproto/ipsec/ipsec.h> 116 #define IPSEC 117 #endif /*FAST_IPSEC*/ 118 119 #ifdef notyet 120 extern struct mbuf *m_copypack(); 121 #endif 122 123 int path_mtu_discovery = 0; 124 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 125 &path_mtu_discovery, 1, "Enable Path MTU Discovery"); 126 127 static int avoid_pure_win_update = 1; 128 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW, 129 &avoid_pure_win_update, 1, "Avoid pure window updates when possible"); 130 131 int tcp_do_autosndbuf = 1; 132 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 133 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing"); 134 135 int tcp_autosndbuf_inc = 8*1024; 136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 137 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer"); 138 139 int tcp_autosndbuf_max = 2*1024*1024; 140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 141 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer"); 142 143 static int tcp_idle_cwv = 1; 144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_cwv, CTLFLAG_RW, 145 &tcp_idle_cwv, 0, 146 "Congestion window validation after idle period (part of RFC2861)"); 147 148 static int tcp_idle_restart = 1; 149 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW, 150 &tcp_idle_restart, 0, "Reset congestion window after idle period"); 151 152 static void tcp_idle_cwnd_validate(struct tcpcb *); 153 154 /* 155 * Tcp output routine: figure out what should be sent and send it. 156 */ 157 int 158 tcp_output(struct tcpcb *tp) 159 { 160 struct inpcb * const inp = tp->t_inpcb; 161 struct socket *so = inp->inp_socket; 162 long len, recvwin, sendwin; 163 int nsacked = 0; 164 int off, flags, error = 0; 165 #ifdef TCP_SIGNATURE 166 int sigoff = 0; 167 #endif 168 struct mbuf *m; 169 struct ip *ip; 170 struct ipovly *ipov; 171 struct tcphdr *th; 172 u_char opt[TCP_MAXOLEN]; 173 unsigned int ipoptlen, optlen, hdrlen; 174 int idle, idle_cwv = 0; 175 boolean_t sendalot; 176 struct ip6_hdr *ip6; 177 #ifdef INET6 178 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 179 #else 180 const boolean_t isipv6 = FALSE; 181 #endif 182 183 KKASSERT(so->so_port == &curthread->td_msgport); 184 185 /* 186 * Determine length of data that should be transmitted, 187 * and flags that will be used. 188 * If there is some data or critical controls (SYN, RST) 189 * to send, then transmit; otherwise, investigate further. 190 */ 191 192 /* 193 * If we have been idle for a while, the send congestion window 194 * could be no longer representative of the current state of the 195 * link; need to validate congestion window. However, we should 196 * not perform congestion window validation here, since we could 197 * be asked to send pure ACK. 198 */ 199 if (tp->snd_max == tp->snd_una && 200 (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart) 201 idle_cwv = 1; 202 203 /* 204 * Calculate whether the transmit stream was previously idle 205 * and adjust TF_LASTIDLE for the next time. 206 */ 207 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 208 if (idle && (tp->t_flags & TF_MORETOCOME)) 209 tp->t_flags |= TF_LASTIDLE; 210 else 211 tp->t_flags &= ~TF_LASTIDLE; 212 213 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 214 !IN_FASTRECOVERY(tp)) 215 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt); 216 217 again: 218 m = NULL; 219 ip = NULL; 220 ipov = NULL; 221 th = NULL; 222 ip6 = NULL; 223 224 /* Make use of SACK information when slow-starting after a RTO. */ 225 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 226 !IN_FASTRECOVERY(tp)) { 227 tcp_seq old_snd_nxt = tp->snd_nxt; 228 229 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt); 230 nsacked += tp->snd_nxt - old_snd_nxt; 231 } 232 233 sendalot = FALSE; 234 off = tp->snd_nxt - tp->snd_una; 235 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked); 236 sendwin = min(sendwin, tp->snd_bwnd); 237 238 flags = tcp_outflags[tp->t_state]; 239 /* 240 * Get standard flags, and add SYN or FIN if requested by 'hidden' 241 * state flags. 242 */ 243 if (tp->t_flags & TF_NEEDFIN) 244 flags |= TH_FIN; 245 if (tp->t_flags & TF_NEEDSYN) 246 flags |= TH_SYN; 247 248 /* 249 * If in persist timeout with window of 0, send 1 byte. 250 * Otherwise, if window is small but nonzero 251 * and timer expired, we will send what we can 252 * and go to transmit state. 253 */ 254 if (tp->t_flags & TF_FORCE) { 255 if (sendwin == 0) { 256 /* 257 * If we still have some data to send, then 258 * clear the FIN bit. Usually this would 259 * happen below when it realizes that we 260 * aren't sending all the data. However, 261 * if we have exactly 1 byte of unsent data, 262 * then it won't clear the FIN bit below, 263 * and if we are in persist state, we wind 264 * up sending the packet without recording 265 * that we sent the FIN bit. 266 * 267 * We can't just blindly clear the FIN bit, 268 * because if we don't have any more data 269 * to send then the probe will be the FIN 270 * itself. 271 */ 272 if (off < so->so_snd.ssb_cc) 273 flags &= ~TH_FIN; 274 sendwin = 1; 275 } else { 276 tcp_callout_stop(tp, tp->tt_persist); 277 tp->t_rxtshift = 0; 278 } 279 } 280 281 /* 282 * If snd_nxt == snd_max and we have transmitted a FIN, the 283 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in 284 * a negative length. This can also occur when TCP opens up 285 * its congestion window while receiving additional duplicate 286 * acks after fast-retransmit because TCP will reset snd_nxt 287 * to snd_max after the fast-retransmit. 288 * 289 * A negative length can also occur when we are in the 290 * TCPS_SYN_RECEIVED state due to a simultanious connect where 291 * our SYN has not been acked yet. 292 * 293 * In the normal retransmit-FIN-only case, however, snd_nxt will 294 * be set to snd_una, the offset will be 0, and the length may 295 * wind up 0. 296 */ 297 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off; 298 299 /* 300 * Lop off SYN bit if it has already been sent. However, if this 301 * is SYN-SENT state and if segment contains data, suppress sending 302 * segment (sending the segment would be an option if we still 303 * did TAO and the remote host supported it). 304 */ 305 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 306 flags &= ~TH_SYN; 307 off--, len++; 308 if (len > 0 && tp->t_state == TCPS_SYN_SENT) { 309 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW); 310 return 0; 311 } 312 } 313 314 /* 315 * Be careful not to send data and/or FIN on SYN segments. 316 * This measure is needed to prevent interoperability problems 317 * with not fully conformant TCP implementations. 318 */ 319 if (flags & TH_SYN) { 320 len = 0; 321 flags &= ~TH_FIN; 322 } 323 324 if (len < 0) { 325 /* 326 * A negative len can occur if our FIN has been sent but not 327 * acked, or if we are in a simultanious connect in the 328 * TCPS_SYN_RECEIVED state with our SYN sent but not yet 329 * acked. 330 * 331 * If our window has contracted to 0 in the FIN case 332 * (which can only occur if we have NOT been called to 333 * retransmit as per code a few paragraphs up) then we 334 * want to shift the retransmit timer over to the 335 * persist timer. 336 * 337 * However, if we are in the TCPS_SYN_RECEIVED state 338 * (the SYN case) we will be in a simultanious connect and 339 * the window may be zero degeneratively. In this case we 340 * do not want to shift to the persist timer after the SYN 341 * or the SYN+ACK transmission. 342 */ 343 len = 0; 344 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) { 345 tcp_callout_stop(tp, tp->tt_rexmt); 346 tp->t_rxtshift = 0; 347 tp->snd_nxt = tp->snd_una; 348 if (!tcp_callout_active(tp, tp->tt_persist)) 349 tcp_setpersist(tp); 350 } 351 } 352 353 KASSERT(len >= 0, ("%s: len < 0", __func__)); 354 /* 355 * Automatic sizing of send socket buffer. Often the send buffer 356 * size is not optimally adjusted to the actual network conditions 357 * at hand (delay bandwidth product). Setting the buffer size too 358 * small limits throughput on links with high bandwidth and high 359 * delay (eg. trans-continental/oceanic links). Setting the 360 * buffer size too big consumes too much real kernel memory, 361 * especially with many connections on busy servers. 362 * 363 * The criteria to step up the send buffer one notch are: 364 * 1. receive window of remote host is larger than send buffer 365 * (with a fudge factor of 5/4th); 366 * 2. send buffer is filled to 7/8th with data (so we actually 367 * have data to make use of it); 368 * 3. send buffer fill has not hit maximal automatic size; 369 * 4. our send window (slow start and cogestion controlled) is 370 * larger than sent but unacknowledged data in send buffer. 371 * 372 * The remote host receive window scaling factor may limit the 373 * growing of the send buffer before it reaches its allowed 374 * maximum. 375 * 376 * It scales directly with slow start or congestion window 377 * and does at most one step per received ACK. This fast 378 * scaling has the drawback of growing the send buffer beyond 379 * what is strictly necessary to make full use of a given 380 * delay*bandwith product. However testing has shown this not 381 * to be much of an problem. At worst we are trading wasting 382 * of available bandwith (the non-use of it) for wasting some 383 * socket buffer memory. 384 * 385 * TODO: Shrink send buffer during idle periods together 386 * with congestion window. Requires another timer. Has to 387 * wait for upcoming tcp timer rewrite. 388 */ 389 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) { 390 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat && 391 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) && 392 so->so_snd.ssb_cc < tcp_autosndbuf_max && 393 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) { 394 u_long newsize; 395 396 newsize = ulmin(so->so_snd.ssb_hiwat + 397 tcp_autosndbuf_inc, 398 tcp_autosndbuf_max); 399 if (!ssb_reserve(&so->so_snd, newsize, so, NULL)) 400 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE); 401 if (newsize >= (TCP_MAXWIN << tp->snd_scale)) 402 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE); 403 } 404 } 405 406 /* 407 * Truncate to the maximum segment length and ensure that FIN is 408 * removed if the length no longer contains the last data byte. 409 */ 410 if (len > tp->t_maxseg) { 411 len = tp->t_maxseg; 412 sendalot = TRUE; 413 } 414 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc)) 415 flags &= ~TH_FIN; 416 417 recvwin = ssb_space(&so->so_rcv); 418 419 /* 420 * Sender silly window avoidance. We transmit under the following 421 * conditions when len is non-zero: 422 * 423 * - We have a full segment 424 * - This is the last buffer in a write()/send() and we are 425 * either idle or running NODELAY 426 * - we've timed out (e.g. persist timer) 427 * - we have more then 1/2 the maximum send window's worth of 428 * data (receiver may be limiting the window size) 429 * - we need to retransmit 430 */ 431 if (len) { 432 if (len == tp->t_maxseg) 433 goto send; 434 /* 435 * NOTE! on localhost connections an 'ack' from the remote 436 * end may occur synchronously with the output and cause 437 * us to flush a buffer queued with moretocome. XXX 438 * 439 * note: the len + off check is almost certainly unnecessary. 440 */ 441 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 442 (idle || (tp->t_flags & TF_NODELAY)) && 443 len + off >= so->so_snd.ssb_cc && 444 !(tp->t_flags & TF_NOPUSH)) { 445 goto send; 446 } 447 if (tp->t_flags & TF_FORCE) /* typ. timeout case */ 448 goto send; 449 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 450 goto send; 451 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 452 goto send; 453 if (tp->t_flags & TF_XMITNOW) 454 goto send; 455 } 456 457 /* 458 * Compare available window to amount of window 459 * known to peer (as advertised window less 460 * next expected input). If the difference is at least two 461 * max size segments, or at least 50% of the maximum possible 462 * window, then want to send a window update to peer. 463 */ 464 if (recvwin > 0) { 465 /* 466 * "adv" is the amount we can increase the window, 467 * taking into account that we are limited by 468 * TCP_MAXWIN << tp->rcv_scale. 469 */ 470 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) - 471 (tp->rcv_adv - tp->rcv_nxt); 472 long hiwat; 473 474 /* 475 * This ack case typically occurs when the user has drained 476 * the TCP socket buffer sufficiently to warrent an ack 477 * containing a 'pure window update'... that is, an ack that 478 * ONLY updates the tcp window. 479 * 480 * It is unclear why we would need to do a pure window update 481 * past 2 segments if we are going to do one at 1/2 the high 482 * water mark anyway, especially since under normal conditions 483 * the user program will drain the socket buffer quickly. 484 * The 2-segment pure window update will often add a large 485 * number of extra, unnecessary acks to the stream. 486 * 487 * avoid_pure_win_update now defaults to 1. 488 */ 489 if (avoid_pure_win_update == 0 || 490 (tp->t_flags & TF_RXRESIZED)) { 491 if (adv >= (long) (2 * tp->t_maxseg)) { 492 goto send; 493 } 494 } 495 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale); 496 if (hiwat > (long)so->so_rcv.ssb_hiwat) 497 hiwat = (long)so->so_rcv.ssb_hiwat; 498 if (adv >= hiwat / 2) 499 goto send; 500 } 501 502 /* 503 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 504 * is also a catch-all for the retransmit timer timeout case. 505 */ 506 if (tp->t_flags & TF_ACKNOW) 507 goto send; 508 if ((flags & TH_RST) || 509 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN))) 510 goto send; 511 if (SEQ_GT(tp->snd_up, tp->snd_una)) 512 goto send; 513 /* 514 * If our state indicates that FIN should be sent 515 * and we have not yet done so, then we need to send. 516 */ 517 if ((flags & TH_FIN) && 518 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una)) 519 goto send; 520 521 /* 522 * TCP window updates are not reliable, rather a polling protocol 523 * using ``persist'' packets is used to insure receipt of window 524 * updates. The three ``states'' for the output side are: 525 * idle not doing retransmits or persists 526 * persisting to move a small or zero window 527 * (re)transmitting and thereby not persisting 528 * 529 * tcp_callout_active(tp, tp->tt_persist) 530 * is true when we are in persist state. 531 * The TF_FORCE flag in tp->t_flags 532 * is set when we are called to send a persist packet. 533 * tcp_callout_active(tp, tp->tt_rexmt) 534 * is set when we are retransmitting 535 * The output side is idle when both timers are zero. 536 * 537 * If send window is too small, there is data to transmit, and no 538 * retransmit or persist is pending, then go to persist state. 539 * 540 * If nothing happens soon, send when timer expires: 541 * if window is nonzero, transmit what we can, otherwise force out 542 * a byte. 543 * 544 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED 545 * with data pending. This situation can occur during a 546 * simultanious connect. 547 */ 548 if (so->so_snd.ssb_cc > 0 && 549 tp->t_state != TCPS_SYN_RECEIVED && 550 !tcp_callout_active(tp, tp->tt_rexmt) && 551 !tcp_callout_active(tp, tp->tt_persist)) { 552 tp->t_rxtshift = 0; 553 tcp_setpersist(tp); 554 } 555 556 /* 557 * No reason to send a segment, just return. 558 */ 559 tp->t_flags &= ~TF_XMITNOW; 560 return (0); 561 562 send: 563 /* 564 * Before ESTABLISHED, force sending of initial options 565 * unless TCP set not to do any options. 566 * NOTE: we assume that the IP/TCP header plus TCP options 567 * always fit in a single mbuf, leaving room for a maximum 568 * link header, i.e. 569 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES 570 */ 571 optlen = 0; 572 if (isipv6) 573 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 574 else 575 hdrlen = sizeof(struct tcpiphdr); 576 if (flags & TH_SYN) { 577 tp->snd_nxt = tp->iss; 578 if (!(tp->t_flags & TF_NOOPT)) { 579 u_short mss; 580 581 opt[0] = TCPOPT_MAXSEG; 582 opt[1] = TCPOLEN_MAXSEG; 583 mss = htons((u_short) tcp_mssopt(tp)); 584 memcpy(opt + 2, &mss, sizeof mss); 585 optlen = TCPOLEN_MAXSEG; 586 587 if ((tp->t_flags & TF_REQ_SCALE) && 588 (!(flags & TH_ACK) || 589 (tp->t_flags & TF_RCVD_SCALE))) { 590 *((u_int32_t *)(opt + optlen)) = htonl( 591 TCPOPT_NOP << 24 | 592 TCPOPT_WINDOW << 16 | 593 TCPOLEN_WINDOW << 8 | 594 tp->request_r_scale); 595 optlen += 4; 596 } 597 598 if ((tcp_do_sack && !(flags & TH_ACK)) || 599 tp->t_flags & TF_SACK_PERMITTED) { 600 uint32_t *lp = (uint32_t *)(opt + optlen); 601 602 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED); 603 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED; 604 } 605 } 606 } 607 608 /* 609 * Send a timestamp and echo-reply if this is a SYN and our side 610 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 611 * and our peer have sent timestamps in our SYN's. 612 */ 613 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 614 !(flags & TH_RST) && 615 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) { 616 u_int32_t *lp = (u_int32_t *)(opt + optlen); 617 618 /* Form timestamp option as shown in appendix A of RFC 1323. */ 619 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 620 *lp++ = htonl(ticks); 621 *lp = htonl(tp->ts_recent); 622 optlen += TCPOLEN_TSTAMP_APPA; 623 } 624 625 /* Set receive buffer autosizing timestamp. */ 626 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) 627 tp->rfbuf_ts = ticks; 628 629 /* 630 * If this is a SACK connection and we have a block to report, 631 * fill in the SACK blocks in the TCP options. 632 */ 633 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) == 634 TF_SACK_PERMITTED && 635 (!TAILQ_EMPTY(&tp->t_segq) || 636 tp->reportblk.rblk_start != tp->reportblk.rblk_end)) 637 tcp_sack_fill_report(tp, opt, &optlen); 638 639 #ifdef TCP_SIGNATURE 640 if (tp->t_flags & TF_SIGNATURE) { 641 int i; 642 u_char *bp; 643 /* 644 * Initialize TCP-MD5 option (RFC2385) 645 */ 646 bp = (u_char *)opt + optlen; 647 *bp++ = TCPOPT_SIGNATURE; 648 *bp++ = TCPOLEN_SIGNATURE; 649 sigoff = optlen + 2; 650 for (i = 0; i < TCP_SIGLEN; i++) 651 *bp++ = 0; 652 optlen += TCPOLEN_SIGNATURE; 653 /* 654 * Terminate options list and maintain 32-bit alignment. 655 */ 656 *bp++ = TCPOPT_NOP; 657 *bp++ = TCPOPT_EOL; 658 optlen += 2; 659 } 660 #endif /* TCP_SIGNATURE */ 661 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options")); 662 hdrlen += optlen; 663 664 if (isipv6) { 665 ipoptlen = ip6_optlen(inp); 666 } else { 667 if (inp->inp_options) { 668 ipoptlen = inp->inp_options->m_len - 669 offsetof(struct ipoption, ipopt_list); 670 } else { 671 ipoptlen = 0; 672 } 673 } 674 #ifdef IPSEC 675 ipoptlen += ipsec_hdrsiz_tcp(tp); 676 #endif 677 678 /* 679 * Adjust data length if insertion of options will bump the packet 680 * length beyond the t_maxopd length. Clear FIN to prevent premature 681 * closure since there is still more data to send after this (now 682 * truncated) packet. 683 * 684 * If just the options do not fit we are in a no-win situation and 685 * we treat it as an unreachable host. 686 */ 687 if (len + optlen + ipoptlen > tp->t_maxopd) { 688 if (tp->t_maxopd <= optlen + ipoptlen) { 689 static time_t last_optlen_report; 690 691 if (last_optlen_report != time_second) { 692 last_optlen_report = time_second; 693 kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd); 694 } 695 error = EHOSTUNREACH; 696 goto out; 697 } else { 698 flags &= ~TH_FIN; 699 len = tp->t_maxopd - optlen - ipoptlen; 700 sendalot = TRUE; 701 } 702 } 703 704 #ifdef INET6 705 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big")); 706 #else 707 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big")); 708 #endif 709 710 /* 711 * Grab a header mbuf, attaching a copy of data to 712 * be transmitted, and initialize the header from 713 * the template for sends on this connection. 714 */ 715 if (len) { 716 if ((tp->t_flags & TF_FORCE) && len == 1) 717 tcpstat.tcps_sndprobe++; 718 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 719 if (tp->snd_nxt == tp->snd_una) 720 tp->snd_max_rexmt = tp->snd_max; 721 if (nsacked) { 722 tcpstat.tcps_sndsackrtopack++; 723 tcpstat.tcps_sndsackrtobyte += len; 724 } 725 tcpstat.tcps_sndrexmitpack++; 726 tcpstat.tcps_sndrexmitbyte += len; 727 } else { 728 tcpstat.tcps_sndpack++; 729 tcpstat.tcps_sndbyte += len; 730 } 731 if (idle_cwv) { 732 idle_cwv = 0; 733 tcp_idle_cwnd_validate(tp); 734 } 735 /* Update last send time after CWV */ 736 tp->snd_last = ticks; 737 #ifdef notyet 738 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len, 739 max_linkhdr + hdrlen)) == NULL) { 740 error = ENOBUFS; 741 goto after_th; 742 } 743 /* 744 * m_copypack left space for our hdr; use it. 745 */ 746 m->m_len += hdrlen; 747 m->m_data -= hdrlen; 748 #else 749 #ifndef INET6 750 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 751 #else 752 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER, 753 M_PKTHDR, NULL); 754 #endif 755 if (m == NULL) { 756 error = ENOBUFS; 757 goto after_th; 758 } 759 m->m_data += max_linkhdr; 760 m->m_len = hdrlen; 761 if (len <= MHLEN - hdrlen - max_linkhdr) { 762 m_copydata(so->so_snd.ssb_mb, off, (int) len, 763 mtod(m, caddr_t) + hdrlen); 764 m->m_len += len; 765 } else { 766 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len); 767 if (m->m_next == NULL) { 768 m_free(m); 769 m = NULL; 770 error = ENOBUFS; 771 goto after_th; 772 } 773 } 774 #endif 775 /* 776 * If we're sending everything we've got, set PUSH. 777 * (This will keep happy those implementations which only 778 * give data to the user when a buffer fills or 779 * a PUSH comes in.) 780 */ 781 if (off + len == so->so_snd.ssb_cc) 782 flags |= TH_PUSH; 783 } else { 784 if (tp->t_flags & TF_ACKNOW) 785 tcpstat.tcps_sndacks++; 786 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 787 tcpstat.tcps_sndctrl++; 788 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 789 tcpstat.tcps_sndurg++; 790 else 791 tcpstat.tcps_sndwinup++; 792 793 MGETHDR(m, MB_DONTWAIT, MT_HEADER); 794 if (m == NULL) { 795 error = ENOBUFS; 796 goto after_th; 797 } 798 if (isipv6 && 799 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN) 800 MH_ALIGN(m, hdrlen); 801 else 802 m->m_data += max_linkhdr; 803 m->m_len = hdrlen; 804 } 805 m->m_pkthdr.rcvif = NULL; 806 if (isipv6) { 807 ip6 = mtod(m, struct ip6_hdr *); 808 th = (struct tcphdr *)(ip6 + 1); 809 tcp_fillheaders(tp, ip6, th); 810 } else { 811 ip = mtod(m, struct ip *); 812 ipov = (struct ipovly *)ip; 813 th = (struct tcphdr *)(ip + 1); 814 /* this picks up the pseudo header (w/o the length) */ 815 tcp_fillheaders(tp, ip, th); 816 } 817 after_th: 818 /* 819 * Fill in fields, remembering maximum advertised 820 * window for use in delaying messages about window sizes. 821 * If resending a FIN, be sure not to use a new sequence number. 822 */ 823 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 824 tp->snd_nxt == tp->snd_max) 825 tp->snd_nxt--; 826 827 if (th != NULL) { 828 /* 829 * If we are doing retransmissions, then snd_nxt will 830 * not reflect the first unsent octet. For ACK only 831 * packets, we do not want the sequence number of the 832 * retransmitted packet, we want the sequence number 833 * of the next unsent octet. So, if there is no data 834 * (and no SYN or FIN), use snd_max instead of snd_nxt 835 * when filling in ti_seq. But if we are in persist 836 * state, snd_max might reflect one byte beyond the 837 * right edge of the window, so use snd_nxt in that 838 * case, since we know we aren't doing a retransmission. 839 * (retransmit and persist are mutually exclusive...) 840 */ 841 if (len || (flags & (TH_SYN|TH_FIN)) || 842 tcp_callout_active(tp, tp->tt_persist)) 843 th->th_seq = htonl(tp->snd_nxt); 844 else 845 th->th_seq = htonl(tp->snd_max); 846 th->th_ack = htonl(tp->rcv_nxt); 847 if (optlen) { 848 bcopy(opt, th + 1, optlen); 849 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 850 } 851 th->th_flags = flags; 852 } 853 854 /* 855 * Calculate receive window. Don't shrink window, but avoid 856 * silly window syndrome by sending a 0 window if the actual 857 * window is less then one segment. 858 */ 859 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) && 860 recvwin < (long)tp->t_maxseg) 861 recvwin = 0; 862 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt)) 863 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt); 864 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale) 865 recvwin = (long)TCP_MAXWIN << tp->rcv_scale; 866 867 /* 868 * Adjust the RXWIN0SENT flag - indicate that we have advertised 869 * a 0 window. This may cause the remote transmitter to stall. This 870 * flag tells soreceive() to disable delayed acknowledgements when 871 * draining the buffer. This can occur if the receiver is attempting 872 * to read more data then can be buffered prior to transmitting on 873 * the connection. 874 */ 875 if (recvwin == 0) 876 tp->t_flags |= TF_RXWIN0SENT; 877 else 878 tp->t_flags &= ~TF_RXWIN0SENT; 879 880 if (th != NULL) 881 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale)); 882 883 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 884 if (th != NULL) { 885 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 886 th->th_flags |= TH_URG; 887 } 888 } else { 889 /* 890 * If no urgent pointer to send, then we pull 891 * the urgent pointer to the left edge of the send window 892 * so that it doesn't drift into the send window on sequence 893 * number wraparound. 894 */ 895 tp->snd_up = tp->snd_una; /* drag it along */ 896 } 897 898 if (th != NULL) { 899 #ifdef TCP_SIGNATURE 900 if (tp->t_flags & TF_SIGNATURE) { 901 tcpsignature_compute(m, len, optlen, 902 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 903 } 904 #endif /* TCP_SIGNATURE */ 905 906 /* 907 * Put TCP length in extended header, and then 908 * checksum extended header and data. 909 */ 910 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 911 if (isipv6) { 912 /* 913 * ip6_plen is not need to be filled now, and will be 914 * filled in ip6_output(). 915 */ 916 th->th_sum = in6_cksum(m, IPPROTO_TCP, 917 sizeof(struct ip6_hdr), 918 sizeof(struct tcphdr) + optlen + len); 919 } else { 920 m->m_pkthdr.csum_flags = CSUM_TCP; 921 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 922 if (len + optlen) { 923 th->th_sum = in_addword(th->th_sum, 924 htons((u_short)(optlen + len))); 925 } 926 927 /* 928 * IP version must be set here for ipv4/ipv6 checking 929 * later 930 */ 931 KASSERT(ip->ip_v == IPVERSION, 932 ("%s: IP version incorrect: %d", 933 __func__, ip->ip_v)); 934 } 935 } 936 937 /* 938 * In transmit state, time the transmission and arrange for 939 * the retransmit. In persist state, just set snd_max. 940 */ 941 if (!(tp->t_flags & TF_FORCE) || 942 !tcp_callout_active(tp, tp->tt_persist)) { 943 tcp_seq startseq = tp->snd_nxt; 944 945 /* 946 * Advance snd_nxt over sequence space of this segment. 947 */ 948 if (flags & (TH_SYN | TH_FIN)) { 949 if (flags & TH_SYN) 950 tp->snd_nxt++; 951 if (flags & TH_FIN) { 952 tp->snd_nxt++; 953 tp->t_flags |= TF_SENTFIN; 954 } 955 } 956 tp->snd_nxt += len; 957 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 958 tp->snd_max = tp->snd_nxt; 959 /* 960 * Time this transmission if not a retransmission and 961 * not currently timing anything. 962 */ 963 if (tp->t_rtttime == 0) { 964 tp->t_rtttime = ticks; 965 tp->t_rtseq = startseq; 966 tcpstat.tcps_segstimed++; 967 } 968 } 969 970 /* 971 * Set retransmit timer if not currently set, 972 * and not doing a pure ack or a keep-alive probe. 973 * Initial value for retransmit timer is smoothed 974 * round-trip time + 2 * round-trip time variance. 975 * Initialize shift counter which is used for backoff 976 * of retransmit time. 977 */ 978 if (!tcp_callout_active(tp, tp->tt_rexmt) && 979 tp->snd_nxt != tp->snd_una) { 980 if (tcp_callout_active(tp, tp->tt_persist)) { 981 tcp_callout_stop(tp, tp->tt_persist); 982 tp->t_rxtshift = 0; 983 } 984 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur, 985 tcp_timer_rexmt); 986 } 987 } else { 988 /* 989 * Persist case, update snd_max but since we are in 990 * persist mode (no window) we do not update snd_nxt. 991 */ 992 int xlen = len; 993 if (flags & TH_SYN) 994 panic("tcp_output: persist timer to send SYN"); 995 if (flags & TH_FIN) { 996 ++xlen; 997 tp->t_flags |= TF_SENTFIN; 998 } 999 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1000 tp->snd_max = tp->snd_nxt + xlen; 1001 } 1002 1003 if (th != NULL) { 1004 #ifdef TCPDEBUG 1005 /* Trace. */ 1006 if (so->so_options & SO_DEBUG) { 1007 tcp_trace(TA_OUTPUT, tp->t_state, tp, 1008 mtod(m, void *), th, 0); 1009 } 1010 #endif 1011 1012 /* 1013 * Fill in IP length and desired time to live and 1014 * send to IP level. There should be a better way 1015 * to handle ttl and tos; we could keep them in 1016 * the template, but need a way to checksum without them. 1017 */ 1018 /* 1019 * m->m_pkthdr.len should have been set before cksum 1020 * calcuration, because in6_cksum() need it. 1021 */ 1022 if (isipv6) { 1023 /* 1024 * we separately set hoplimit for every segment, 1025 * since the user might want to change the value 1026 * via setsockopt. Also, desired default hop 1027 * limit might be changed via Neighbor Discovery. 1028 */ 1029 ip6->ip6_hlim = in6_selecthlim(inp, 1030 (inp->in6p_route.ro_rt ? 1031 inp->in6p_route.ro_rt->rt_ifp : NULL)); 1032 1033 /* TODO: IPv6 IP6TOS_ECT bit on */ 1034 error = ip6_output(m, inp->in6p_outputopts, 1035 &inp->in6p_route, (so->so_options & SO_DONTROUTE), 1036 NULL, NULL, inp); 1037 } else { 1038 struct rtentry *rt; 1039 ip->ip_len = m->m_pkthdr.len; 1040 #ifdef INET6 1041 if (INP_CHECK_SOCKAF(so, AF_INET6)) 1042 ip->ip_ttl = in6_selecthlim(inp, 1043 (inp->in6p_route.ro_rt ? 1044 inp->in6p_route.ro_rt->rt_ifp : NULL)); 1045 else 1046 #endif 1047 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */ 1048 1049 ip->ip_tos = inp->inp_ip_tos; /* XXX */ 1050 /* 1051 * See if we should do MTU discovery. 1052 * We do it only if the following are true: 1053 * 1) we have a valid route to the destination 1054 * 2) the MTU is not locked (if it is, 1055 * then discovery has been disabled) 1056 */ 1057 if (path_mtu_discovery && 1058 (rt = inp->inp_route.ro_rt) && 1059 (rt->rt_flags & RTF_UP) && 1060 !(rt->rt_rmx.rmx_locks & RTV_MTU)) 1061 ip->ip_off |= IP_DF; 1062 1063 error = ip_output(m, inp->inp_options, &inp->inp_route, 1064 (so->so_options & SO_DONTROUTE) | 1065 IP_DEBUGROUTE, NULL, inp); 1066 } 1067 } else { 1068 KASSERT(error != 0, ("no error, but th not set")); 1069 } 1070 if (error) { 1071 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW); 1072 1073 /* 1074 * We know that the packet was lost, so back out the 1075 * sequence number advance, if any. 1076 */ 1077 if (!(tp->t_flags & TF_FORCE) || 1078 !tcp_callout_active(tp, tp->tt_persist)) { 1079 /* 1080 * No need to check for TH_FIN here because 1081 * the TF_SENTFIN flag handles that case. 1082 */ 1083 if (!(flags & TH_SYN)) 1084 tp->snd_nxt -= len; 1085 } 1086 1087 out: 1088 if (error == ENOBUFS) { 1089 /* 1090 * If we can't send, make sure there is something 1091 * to get us going again later. 1092 * 1093 * The persist timer isn't necessarily allowed in all 1094 * states, use the rexmt timer. 1095 */ 1096 if (!tcp_callout_active(tp, tp->tt_rexmt) && 1097 !tcp_callout_active(tp, tp->tt_persist)) { 1098 tcp_callout_reset(tp, tp->tt_rexmt, 1099 tp->t_rxtcur, 1100 tcp_timer_rexmt); 1101 #if 0 1102 tp->t_rxtshift = 0; 1103 tcp_setpersist(tp); 1104 #endif 1105 } 1106 tcp_quench(inp, 0); 1107 return (0); 1108 } 1109 if (error == EMSGSIZE) { 1110 /* 1111 * ip_output() will have already fixed the route 1112 * for us. tcp_mtudisc() will, as its last action, 1113 * initiate retransmission, so it is important to 1114 * not do so here. 1115 */ 1116 tcp_mtudisc(inp, 0); 1117 return 0; 1118 } 1119 if ((error == EHOSTUNREACH || error == ENETDOWN) && 1120 TCPS_HAVERCVDSYN(tp->t_state)) { 1121 tp->t_softerror = error; 1122 return (0); 1123 } 1124 return (error); 1125 } 1126 tcpstat.tcps_sndtotal++; 1127 1128 /* 1129 * Data sent (as far as we can tell). 1130 * 1131 * If this advertises a larger window than any other segment, 1132 * then remember the size of the advertised window. 1133 * 1134 * Any pending ACK has now been sent. 1135 */ 1136 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) { 1137 tp->rcv_adv = tp->rcv_nxt + recvwin; 1138 tp->t_flags &= ~TF_RXRESIZED; 1139 } 1140 tp->last_ack_sent = tp->rcv_nxt; 1141 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW); 1142 if (tcp_delack_enabled) 1143 tcp_callout_stop(tp, tp->tt_delack); 1144 if (sendalot) 1145 goto again; 1146 return (0); 1147 } 1148 1149 void 1150 tcp_setpersist(struct tcpcb *tp) 1151 { 1152 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1153 int tt; 1154 1155 if (tp->t_state == TCPS_SYN_SENT || 1156 tp->t_state == TCPS_SYN_RECEIVED) { 1157 panic("tcp_setpersist: not established yet, current %s", 1158 tp->t_state == TCPS_SYN_SENT ? 1159 "SYN_SENT" : "SYN_RECEIVED"); 1160 } 1161 1162 if (tcp_callout_active(tp, tp->tt_rexmt)) 1163 panic("tcp_setpersist: retransmit pending"); 1164 /* 1165 * Start/restart persistance timer. 1166 */ 1167 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN, 1168 TCPTV_PERSMAX); 1169 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist); 1170 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1171 tp->t_rxtshift++; 1172 } 1173 1174 static void 1175 tcp_idle_cwnd_validate(struct tcpcb *tp) 1176 { 1177 u_long initial_cwnd = tcp_initial_window(tp); 1178 u_long min_cwnd; 1179 1180 tcpstat.tcps_sndidle++; 1181 1182 /* According to RFC5681: RW=min(IW,cwnd) */ 1183 min_cwnd = min(tp->snd_cwnd, initial_cwnd); 1184 1185 if (tcp_idle_cwv) { 1186 u_long idle_time, decay_cwnd; 1187 1188 /* 1189 * RFC2861, but only after idle period. 1190 */ 1191 1192 /* 1193 * Before the congestion window is reduced, ssthresh 1194 * is set to the maximum of its current value and 3/4 1195 * cwnd. If the sender then has more data to send 1196 * than the decayed cwnd allows, the TCP will slow- 1197 * start (perform exponential increase) at least 1198 * half-way back up to the old value of cwnd. 1199 */ 1200 tp->snd_ssthresh = max(tp->snd_ssthresh, 1201 (3 * tp->snd_cwnd) / 4); 1202 1203 /* 1204 * Decay the congestion window by half for every RTT 1205 * that the flow remains inactive. 1206 * 1207 * The difference between our implementation and 1208 * RFC2861 is that we don't allow cwnd to go below 1209 * the value allowed by RFC5681 (min_cwnd). 1210 */ 1211 idle_time = ticks - tp->snd_last; 1212 decay_cwnd = tp->snd_cwnd; 1213 while (idle_time >= tp->t_rxtcur && 1214 decay_cwnd > min_cwnd) { 1215 decay_cwnd >>= 1; 1216 idle_time -= tp->t_rxtcur; 1217 } 1218 tp->snd_cwnd = max(decay_cwnd, min_cwnd); 1219 } else { 1220 /* 1221 * Slow-start from scratch to re-determine the send 1222 * congestion window. 1223 */ 1224 tp->snd_cwnd = min_cwnd; 1225 } 1226 1227 /* Restart ABC counting during congestion avoidance */ 1228 tp->snd_wacked = 0; 1229 } 1230