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. Neither the name of the University nor the names of its contributors 47 * may be used to endorse or promote products derived from this software 48 * without specific prior written permission. 49 * 50 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 53 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 60 * SUCH DAMAGE. 61 * 62 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 63 * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.20 2003/01/29 22:45:36 hsu Exp $ 64 */ 65 66 #include "opt_inet.h" 67 #include "opt_inet6.h" 68 #include "opt_ipsec.h" 69 #include "opt_tcpdebug.h" 70 71 #include <sys/param.h> 72 #include <sys/systm.h> 73 #include <sys/kernel.h> 74 #include <sys/sysctl.h> 75 #include <sys/mbuf.h> 76 #include <sys/domain.h> 77 #include <sys/protosw.h> 78 #include <sys/socket.h> 79 #include <sys/socketvar.h> 80 #include <sys/in_cksum.h> 81 #include <sys/thread.h> 82 #include <sys/globaldata.h> 83 84 #include <net/if_var.h> 85 #include <net/route.h> 86 #include <net/netmsg2.h> 87 #include <net/netisr2.h> 88 89 #include <netinet/in.h> 90 #include <netinet/in_systm.h> 91 #include <netinet/ip.h> 92 #include <netinet/in_pcb.h> 93 #include <netinet/ip_var.h> 94 #include <netinet6/in6_pcb.h> 95 #include <netinet/ip6.h> 96 #include <netinet6/ip6_var.h> 97 #include <netinet/tcp.h> 98 #define TCPOUTFLAGS 99 #include <netinet/tcp_fsm.h> 100 #include <netinet/tcp_seq.h> 101 #include <netinet/tcp_timer.h> 102 #include <netinet/tcp_timer2.h> 103 #include <netinet/tcp_var.h> 104 #include <netinet/tcpip.h> 105 #ifdef TCPDEBUG 106 #include <netinet/tcp_debug.h> 107 #endif 108 109 #ifdef IPSEC 110 #include <netinet6/ipsec.h> 111 #endif /*IPSEC*/ 112 113 #ifdef FAST_IPSEC 114 #include <netproto/ipsec/ipsec.h> 115 #define IPSEC 116 #endif /*FAST_IPSEC*/ 117 118 #ifdef notyet 119 extern struct mbuf *m_copypack(); 120 #endif 121 122 int path_mtu_discovery = 1; 123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 124 &path_mtu_discovery, 1, "Enable Path MTU Discovery"); 125 126 static int avoid_pure_win_update = 1; 127 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW, 128 &avoid_pure_win_update, 1, "Avoid pure window updates when possible"); 129 130 /* 131 * 1 - enabled for increasing and decreasing the buffer size 132 * 2 - enabled only for increasing the buffer size 133 */ 134 int tcp_do_autosndbuf = 1; 135 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 136 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing"); 137 138 int tcp_autosndbuf_inc = 8*1024; 139 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 140 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer"); 141 142 int tcp_autosndbuf_min = 32768; 143 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_min, CTLFLAG_RW, 144 &tcp_autosndbuf_min, 0, "Min size of automatic send buffer"); 145 146 int tcp_autosndbuf_max = 2*1024*1024; 147 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 148 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer"); 149 150 int tcp_prio_synack = 1; 151 SYSCTL_INT(_net_inet_tcp, OID_AUTO, prio_synack, CTLFLAG_RW, 152 &tcp_prio_synack, 0, "Prioritize SYN, SYN|ACK and pure ACK"); 153 154 static int tcp_idle_cwv = 1; 155 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_cwv, CTLFLAG_RW, 156 &tcp_idle_cwv, 0, 157 "Congestion window validation after idle period (part of RFC2861)"); 158 159 static int tcp_idle_restart = 1; 160 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW, 161 &tcp_idle_restart, 0, "Reset congestion window after idle period"); 162 163 static int tcp_do_tso = 1; 164 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW, 165 &tcp_do_tso, 0, "Enable TCP Segmentation Offload (TSO)"); 166 167 static int tcp_fairsend = 4; 168 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fairsend, CTLFLAG_RW, 169 &tcp_fairsend, 0, 170 "Amount of segments sent before yield to other senders or receivers"); 171 172 static void tcp_idle_cwnd_validate(struct tcpcb *); 173 174 static int tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen); 175 static void tcp_output_sched(struct tcpcb *tp); 176 177 /* 178 * Tcp output routine: figure out what should be sent and send it. 179 */ 180 int 181 tcp_output(struct tcpcb *tp) 182 { 183 struct inpcb * const inp = tp->t_inpcb; 184 struct socket *so = inp->inp_socket; 185 long len, recvwin, sendwin; 186 int nsacked = 0; 187 int off, flags, error = 0; 188 #ifdef TCP_SIGNATURE 189 int sigoff = 0; 190 #endif 191 struct mbuf *m; 192 struct ip *ip; 193 struct tcphdr *th; 194 u_char opt[TCP_MAXOLEN]; 195 unsigned int ipoptlen, optlen, hdrlen; 196 int idle; 197 boolean_t sendalot; 198 struct ip6_hdr *ip6; 199 #ifdef INET6 200 const boolean_t isipv6 = INP_ISIPV6(inp); 201 #else 202 const boolean_t isipv6 = FALSE; 203 #endif 204 boolean_t can_tso = FALSE, use_tso; 205 boolean_t report_sack, idle_cwv = FALSE; 206 u_int segsz, tso_hlen, tso_lenmax = 0; 207 int segcnt = 0; 208 boolean_t need_sched = FALSE; 209 210 KKASSERT(so->so_port == &curthread->td_msgport); 211 212 /* 213 * Determine length of data that should be transmitted, 214 * and flags that will be used. 215 * If there is some data or critical controls (SYN, RST) 216 * to send, then transmit; otherwise, investigate further. 217 */ 218 219 /* 220 * If we have been idle for a while, the send congestion window 221 * could be no longer representative of the current state of the 222 * link; need to validate congestion window. However, we should 223 * not perform congestion window validation here, since we could 224 * be asked to send pure ACK. 225 */ 226 if (tp->snd_max == tp->snd_una && 227 (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart) 228 idle_cwv = TRUE; 229 230 /* 231 * Calculate whether the transmit stream was previously idle 232 * and adjust TF_LASTIDLE for the next time. 233 */ 234 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 235 if (idle && (tp->t_flags & TF_MORETOCOME)) 236 tp->t_flags |= TF_LASTIDLE; 237 else 238 tp->t_flags &= ~TF_LASTIDLE; 239 240 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 241 !IN_FASTRECOVERY(tp)) 242 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt); 243 244 /* 245 * Find out whether TSO could be used or not 246 * 247 * For TSO capable devices, the following assumptions apply to 248 * the processing of TCP flags: 249 * - If FIN is set on the large TCP segment, the device must set 250 * FIN on the last segment that it creates from the large TCP 251 * segment. 252 * - If PUSH is set on the large TCP segment, the device must set 253 * PUSH on the last segment that it creates from the large TCP 254 * segment. 255 */ 256 #if !defined(IPSEC) && !defined(FAST_IPSEC) 257 if (tcp_do_tso 258 #ifdef TCP_SIGNATURE 259 && (tp->t_flags & TF_SIGNATURE) == 0 260 #endif 261 ) { 262 if (!isipv6) { 263 struct rtentry *rt = inp->inp_route.ro_rt; 264 265 if (rt != NULL && (rt->rt_flags & RTF_UP) && 266 (rt->rt_ifp->if_hwassist & CSUM_TSO)) { 267 can_tso = TRUE; 268 tso_lenmax = rt->rt_ifp->if_tsolen; 269 } 270 } 271 } 272 #endif /* !IPSEC && !FAST_IPSEC */ 273 274 again: 275 m = NULL; 276 ip = NULL; 277 th = NULL; 278 ip6 = NULL; 279 280 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) == 281 TF_SACK_PERMITTED && 282 (!TAILQ_EMPTY(&tp->t_segq) || 283 tp->reportblk.rblk_start != tp->reportblk.rblk_end)) 284 report_sack = TRUE; 285 else 286 report_sack = FALSE; 287 288 /* Make use of SACK information when slow-starting after a RTO. */ 289 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max && 290 !IN_FASTRECOVERY(tp)) { 291 tcp_seq old_snd_nxt = tp->snd_nxt; 292 293 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt); 294 nsacked += tp->snd_nxt - old_snd_nxt; 295 } 296 297 sendalot = FALSE; 298 off = tp->snd_nxt - tp->snd_una; 299 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked); 300 sendwin = min(sendwin, tp->snd_bwnd); 301 302 flags = tcp_outflags[tp->t_state]; 303 /* 304 * Get standard flags, and add SYN or FIN if requested by 'hidden' 305 * state flags. 306 */ 307 if (tp->t_flags & TF_NEEDFIN) 308 flags |= TH_FIN; 309 if (tp->t_flags & TF_NEEDSYN) 310 flags |= TH_SYN; 311 312 /* 313 * If in persist timeout with window of 0, send 1 byte. 314 * Otherwise, if window is small but nonzero 315 * and timer expired, we will send what we can 316 * and go to transmit state. 317 */ 318 if (tp->t_flags & TF_FORCE) { 319 if (sendwin == 0) { 320 /* 321 * If we still have some data to send, then 322 * clear the FIN bit. Usually this would 323 * happen below when it realizes that we 324 * aren't sending all the data. However, 325 * if we have exactly 1 byte of unsent data, 326 * then it won't clear the FIN bit below, 327 * and if we are in persist state, we wind 328 * up sending the packet without recording 329 * that we sent the FIN bit. 330 * 331 * We can't just blindly clear the FIN bit, 332 * because if we don't have any more data 333 * to send then the probe will be the FIN 334 * itself. 335 */ 336 if (off < so->so_snd.ssb_cc) 337 flags &= ~TH_FIN; 338 sendwin = 1; 339 } else { 340 tcp_callout_stop(tp, tp->tt_persist); 341 tp->t_rxtshift = 0; 342 } 343 } 344 345 /* 346 * If snd_nxt == snd_max and we have transmitted a FIN, the 347 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in 348 * a negative length. This can also occur when TCP opens up 349 * its congestion window while receiving additional duplicate 350 * acks after fast-retransmit because TCP will reset snd_nxt 351 * to snd_max after the fast-retransmit. 352 * 353 * A negative length can also occur when we are in the 354 * TCPS_SYN_RECEIVED state due to a simultanious connect where 355 * our SYN has not been acked yet. 356 * 357 * In the normal retransmit-FIN-only case, however, snd_nxt will 358 * be set to snd_una, the offset will be 0, and the length may 359 * wind up 0. 360 */ 361 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off; 362 363 /* 364 * Lop off SYN bit if it has already been sent. However, if this 365 * is SYN-SENT state and if segment contains data, suppress sending 366 * segment (sending the segment would be an option if we still 367 * did TAO and the remote host supported it). 368 */ 369 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 370 flags &= ~TH_SYN; 371 off--, len++; 372 if (len > 0 && tp->t_state == TCPS_SYN_SENT) { 373 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW); 374 return 0; 375 } 376 } 377 378 /* 379 * Be careful not to send data and/or FIN on SYN segments. 380 * This measure is needed to prevent interoperability problems 381 * with not fully conformant TCP implementations. 382 */ 383 if (flags & TH_SYN) { 384 len = 0; 385 flags &= ~TH_FIN; 386 } 387 388 if (len < 0) { 389 /* 390 * A negative len can occur if our FIN has been sent but not 391 * acked, or if we are in a simultanious connect in the 392 * TCPS_SYN_RECEIVED state with our SYN sent but not yet 393 * acked. 394 * 395 * If our window has contracted to 0 in the FIN case 396 * (which can only occur if we have NOT been called to 397 * retransmit as per code a few paragraphs up) then we 398 * want to shift the retransmit timer over to the 399 * persist timer. 400 * 401 * However, if we are in the TCPS_SYN_RECEIVED state 402 * (the SYN case) we will be in a simultanious connect and 403 * the window may be zero degeneratively. In this case we 404 * do not want to shift to the persist timer after the SYN 405 * or the SYN+ACK transmission. 406 */ 407 len = 0; 408 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) { 409 tcp_callout_stop(tp, tp->tt_rexmt); 410 tp->t_rxtshift = 0; 411 tp->snd_nxt = tp->snd_una; 412 if (!tcp_callout_active(tp, tp->tt_persist)) 413 tcp_setpersist(tp); 414 } 415 } 416 417 KASSERT(len >= 0, ("%s: len < 0", __func__)); 418 /* 419 * Automatic sizing of send socket buffer. Often the send buffer 420 * size is not optimally adjusted to the actual network conditions 421 * at hand (delay bandwidth product). Setting the buffer size too 422 * small limits throughput on links with high bandwidth and high 423 * delay (eg. trans-continental/oceanic links). Setting the 424 * buffer size too big consumes too much real kernel memory, 425 * especially with many connections on busy servers. 426 * 427 * The criteria to step up the send buffer one notch are: 428 * 1. receive window of remote host is larger than send buffer 429 * (with a fudge factor of 5/4th); 430 * 2. hiwat has not significantly exceeded bwnd (inflight) 431 * (bwnd is a maximal value if inflight is disabled). 432 * 3. send buffer is filled to 7/8th with data (so we actually 433 * have data to make use of it); 434 * 4. hiwat has not hit maximal automatic size; 435 * 5. our send window (slow start and cogestion controlled) is 436 * larger than sent but unacknowledged data in send buffer. 437 * 438 * The remote host receive window scaling factor may limit the 439 * growing of the send buffer before it reaches its allowed 440 * maximum. 441 * 442 * It scales directly with slow start or congestion window 443 * and does at most one step per received ACK. This fast 444 * scaling has the drawback of growing the send buffer beyond 445 * what is strictly necessary to make full use of a given 446 * delay*bandwith product. However testing has shown this not 447 * to be much of an problem. At worst we are trading wasting 448 * of available bandwith (the non-use of it) for wasting some 449 * socket buffer memory. 450 * 451 * The criteria for shrinking the buffer is based solely on 452 * the inflight code (snd_bwnd). If inflight is disabled, 453 * the buffer will not be shrinked. Note that snd_bwnd already 454 * has a fudge factor. Our test adds a little hysteresis. 455 */ 456 if (tcp_do_autosndbuf && (so->so_snd.ssb_flags & SSB_AUTOSIZE)) { 457 const int asbinc = tcp_autosndbuf_inc; 458 const int hiwat = so->so_snd.ssb_hiwat; 459 const int lowat = so->so_snd.ssb_lowat; 460 u_long newsize; 461 462 if ((tp->snd_wnd / 4 * 5) >= hiwat && 463 so->so_snd.ssb_cc >= (hiwat / 8 * 7) && 464 hiwat < tp->snd_bwnd + hiwat / 10 && 465 hiwat + asbinc < tcp_autosndbuf_max && 466 hiwat < (TCP_MAXWIN << tp->snd_scale) && 467 sendwin >= (so->so_snd.ssb_cc - 468 (tp->snd_nxt - tp->snd_una))) { 469 newsize = ulmin(hiwat + asbinc, tcp_autosndbuf_max); 470 if (!ssb_reserve(&so->so_snd, newsize, so, NULL)) 471 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE); 472 #if 0 473 if (newsize >= (TCP_MAXWIN << tp->snd_scale)) 474 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE); 475 #endif 476 } else if ((long)tp->snd_bwnd < 477 (long)(hiwat * 3 / 4 - lowat - asbinc) && 478 hiwat > tp->t_maxseg * 2 + asbinc && 479 hiwat + asbinc >= tcp_autosndbuf_min && 480 tcp_do_autosndbuf == 1) { 481 newsize = ulmax(hiwat - asbinc, tp->t_maxseg * 2); 482 ssb_reserve(&so->so_snd, newsize, so, NULL); 483 } 484 } 485 486 /* 487 * Don't use TSO, if: 488 * - Congestion window needs validation 489 * - There are SACK blocks to report 490 * - RST or SYN flags is set 491 * - URG will be set 492 * 493 * XXX 494 * Checking for SYN|RST looks overkill, just to be safe than sorry 495 */ 496 use_tso = can_tso; 497 if (report_sack || idle_cwv || (flags & (TH_RST | TH_SYN))) 498 use_tso = FALSE; 499 if (use_tso) { 500 tcp_seq ugr_nxt = tp->snd_nxt; 501 502 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) && 503 tp->snd_nxt == tp->snd_max) 504 --ugr_nxt; 505 506 if (SEQ_GT(tp->snd_up, ugr_nxt)) 507 use_tso = FALSE; 508 } 509 510 if (use_tso) { 511 /* 512 * Find out segment size and header length for TSO 513 */ 514 error = tcp_tso_getsize(tp, &segsz, &tso_hlen); 515 if (error) 516 use_tso = FALSE; 517 } 518 if (!use_tso) { 519 segsz = tp->t_maxseg; 520 tso_hlen = 0; /* not used */ 521 } 522 523 /* 524 * Truncate to the maximum segment length if not TSO, and ensure that 525 * FIN is removed if the length no longer contains the last data byte. 526 */ 527 if (len > segsz) { 528 if (!use_tso) { 529 len = segsz; 530 ++segcnt; 531 } else { 532 int nsegs; 533 534 if (__predict_false(tso_lenmax < segsz)) 535 tso_lenmax = segsz << 1; 536 537 /* 538 * Truncate TSO transfers to (IP_MAXPACKET - iphlen - 539 * thoff), and make sure that we send equal size 540 * transfers down the stack (rather than big-small- 541 * big-small-...). 542 */ 543 len = min(len, tso_lenmax); 544 nsegs = min(len, (IP_MAXPACKET - tso_hlen)) / segsz; 545 KKASSERT(nsegs > 0); 546 547 len = nsegs * segsz; 548 549 if (len <= segsz) { 550 use_tso = FALSE; 551 ++segcnt; 552 } else { 553 segcnt += nsegs; 554 } 555 } 556 sendalot = TRUE; 557 } else { 558 use_tso = FALSE; 559 if (len > 0) 560 ++segcnt; 561 } 562 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc)) 563 flags &= ~TH_FIN; 564 565 recvwin = ssb_space(&so->so_rcv); 566 567 /* 568 * Sender silly window avoidance. We transmit under the following 569 * conditions when len is non-zero: 570 * 571 * - We have a full segment 572 * - This is the last buffer in a write()/send() and we are 573 * either idle or running NODELAY 574 * - we've timed out (e.g. persist timer) 575 * - we have more then 1/2 the maximum send window's worth of 576 * data (receiver may be limiting the window size) 577 * - we need to retransmit 578 */ 579 if (len) { 580 if (len >= segsz) 581 goto send; 582 /* 583 * NOTE! on localhost connections an 'ack' from the remote 584 * end may occur synchronously with the output and cause 585 * us to flush a buffer queued with moretocome. XXX 586 * 587 * note: the len + off check is almost certainly unnecessary. 588 */ 589 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 590 (idle || (tp->t_flags & TF_NODELAY)) && 591 len + off >= so->so_snd.ssb_cc && 592 !(tp->t_flags & TF_NOPUSH)) { 593 goto send; 594 } 595 if (tp->t_flags & TF_FORCE) /* typ. timeout case */ 596 goto send; 597 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 598 goto send; 599 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 600 goto send; 601 if (tp->t_flags & TF_XMITNOW) 602 goto send; 603 } 604 605 /* 606 * Compare available window to amount of window 607 * known to peer (as advertised window less 608 * next expected input). If the difference is at least two 609 * max size segments, or at least 50% of the maximum possible 610 * window, then want to send a window update to peer. 611 */ 612 if (recvwin > 0) { 613 /* 614 * "adv" is the amount we can increase the window, 615 * taking into account that we are limited by 616 * TCP_MAXWIN << tp->rcv_scale. 617 */ 618 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) - 619 (tp->rcv_adv - tp->rcv_nxt); 620 long hiwat; 621 622 /* 623 * This ack case typically occurs when the user has drained 624 * the TCP socket buffer sufficiently to warrent an ack 625 * containing a 'pure window update'... that is, an ack that 626 * ONLY updates the tcp window. 627 * 628 * It is unclear why we would need to do a pure window update 629 * past 2 segments if we are going to do one at 1/2 the high 630 * water mark anyway, especially since under normal conditions 631 * the user program will drain the socket buffer quickly. 632 * The 2-segment pure window update will often add a large 633 * number of extra, unnecessary acks to the stream. 634 * 635 * avoid_pure_win_update now defaults to 1. 636 */ 637 if (avoid_pure_win_update == 0 || 638 (tp->t_flags & TF_RXRESIZED)) { 639 if (adv >= (long) (2 * segsz)) { 640 goto send; 641 } 642 } 643 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale); 644 if (hiwat > (long)so->so_rcv.ssb_hiwat) 645 hiwat = (long)so->so_rcv.ssb_hiwat; 646 if (adv >= hiwat / 2) 647 goto send; 648 } 649 650 /* 651 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 652 * is also a catch-all for the retransmit timer timeout case. 653 */ 654 if (tp->t_flags & TF_ACKNOW) 655 goto send; 656 if ((flags & TH_RST) || 657 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN))) 658 goto send; 659 if (SEQ_GT(tp->snd_up, tp->snd_una)) 660 goto send; 661 /* 662 * If our state indicates that FIN should be sent 663 * and we have not yet done so, then we need to send. 664 */ 665 if ((flags & TH_FIN) && 666 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una)) 667 goto send; 668 669 /* 670 * TCP window updates are not reliable, rather a polling protocol 671 * using ``persist'' packets is used to insure receipt of window 672 * updates. The three ``states'' for the output side are: 673 * idle not doing retransmits or persists 674 * persisting to move a small or zero window 675 * (re)transmitting and thereby not persisting 676 * 677 * tcp_callout_active(tp, tp->tt_persist) 678 * is true when we are in persist state. 679 * The TF_FORCE flag in tp->t_flags 680 * is set when we are called to send a persist packet. 681 * tcp_callout_active(tp, tp->tt_rexmt) 682 * is set when we are retransmitting 683 * The output side is idle when both timers are zero. 684 * 685 * If send window is too small, there is data to transmit, and no 686 * retransmit or persist is pending, then go to persist state. 687 * 688 * If nothing happens soon, send when timer expires: 689 * if window is nonzero, transmit what we can, otherwise force out 690 * a byte. 691 * 692 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED 693 * with data pending. This situation can occur during a 694 * simultanious connect. 695 */ 696 if (so->so_snd.ssb_cc > 0 && 697 tp->t_state != TCPS_SYN_RECEIVED && 698 !tcp_callout_active(tp, tp->tt_rexmt) && 699 !tcp_callout_active(tp, tp->tt_persist)) { 700 tp->t_rxtshift = 0; 701 tcp_setpersist(tp); 702 } 703 704 /* 705 * No reason to send a segment, just return. 706 */ 707 tp->t_flags &= ~TF_XMITNOW; 708 return (0); 709 710 send: 711 if (need_sched && len > 0) { 712 tcp_output_sched(tp); 713 return 0; 714 } 715 716 /* 717 * Before ESTABLISHED, force sending of initial options 718 * unless TCP set not to do any options. 719 * NOTE: we assume that the IP/TCP header plus TCP options 720 * always fit in a single mbuf, leaving room for a maximum 721 * link header, i.e. 722 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES 723 */ 724 optlen = 0; 725 if (isipv6) 726 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 727 else 728 hdrlen = sizeof(struct tcpiphdr); 729 if (flags & TH_SYN) { 730 tp->snd_nxt = tp->iss; 731 if (!(tp->t_flags & TF_NOOPT)) { 732 u_short mss; 733 734 opt[0] = TCPOPT_MAXSEG; 735 opt[1] = TCPOLEN_MAXSEG; 736 mss = htons((u_short) tcp_mssopt(tp)); 737 memcpy(opt + 2, &mss, sizeof mss); 738 optlen = TCPOLEN_MAXSEG; 739 740 if ((tp->t_flags & TF_REQ_SCALE) && 741 (!(flags & TH_ACK) || 742 (tp->t_flags & TF_RCVD_SCALE))) { 743 *((u_int32_t *)(opt + optlen)) = htonl( 744 TCPOPT_NOP << 24 | 745 TCPOPT_WINDOW << 16 | 746 TCPOLEN_WINDOW << 8 | 747 tp->request_r_scale); 748 optlen += 4; 749 } 750 751 if ((tcp_do_sack && !(flags & TH_ACK)) || 752 tp->t_flags & TF_SACK_PERMITTED) { 753 uint32_t *lp = (uint32_t *)(opt + optlen); 754 755 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED); 756 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED; 757 } 758 } 759 } 760 761 /* 762 * Send a timestamp and echo-reply if this is a SYN and our side 763 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 764 * and our peer have sent timestamps in our SYN's. 765 */ 766 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 767 !(flags & TH_RST) && 768 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) { 769 u_int32_t *lp = (u_int32_t *)(opt + optlen); 770 771 /* Form timestamp option as shown in appendix A of RFC 1323. */ 772 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 773 *lp++ = htonl(ticks); 774 *lp = htonl(tp->ts_recent); 775 optlen += TCPOLEN_TSTAMP_APPA; 776 } 777 778 /* Set receive buffer autosizing timestamp. */ 779 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) 780 tp->rfbuf_ts = ticks; 781 782 /* 783 * If this is a SACK connection and we have a block to report, 784 * fill in the SACK blocks in the TCP options. 785 */ 786 if (report_sack) 787 tcp_sack_fill_report(tp, opt, &optlen); 788 789 #ifdef TCP_SIGNATURE 790 if (tp->t_flags & TF_SIGNATURE) { 791 int i; 792 u_char *bp; 793 /* 794 * Initialize TCP-MD5 option (RFC2385) 795 */ 796 bp = (u_char *)opt + optlen; 797 *bp++ = TCPOPT_SIGNATURE; 798 *bp++ = TCPOLEN_SIGNATURE; 799 sigoff = optlen + 2; 800 for (i = 0; i < TCP_SIGLEN; i++) 801 *bp++ = 0; 802 optlen += TCPOLEN_SIGNATURE; 803 /* 804 * Terminate options list and maintain 32-bit alignment. 805 */ 806 *bp++ = TCPOPT_NOP; 807 *bp++ = TCPOPT_EOL; 808 optlen += 2; 809 } 810 #endif /* TCP_SIGNATURE */ 811 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options")); 812 hdrlen += optlen; 813 814 if (isipv6) { 815 ipoptlen = ip6_optlen(inp); 816 } else { 817 if (inp->inp_options) { 818 ipoptlen = inp->inp_options->m_len - 819 offsetof(struct ipoption, ipopt_list); 820 } else { 821 ipoptlen = 0; 822 } 823 } 824 #ifdef IPSEC 825 ipoptlen += ipsec_hdrsiz_tcp(tp); 826 #endif 827 828 if (use_tso) { 829 /* TSO segment length must be multiple of segment size */ 830 KASSERT(len >= (2 * segsz) && (len % segsz == 0), 831 ("invalid TSO len %ld, segsz %u", len, segsz)); 832 } else { 833 KASSERT(len <= segsz, 834 ("invalid len %ld, segsz %u", len, segsz)); 835 836 /* 837 * Adjust data length if insertion of options will bump 838 * the packet length beyond the t_maxopd length. Clear 839 * FIN to prevent premature closure since there is still 840 * more data to send after this (now truncated) packet. 841 * 842 * If just the options do not fit we are in a no-win 843 * situation and we treat it as an unreachable host. 844 */ 845 if (len + optlen + ipoptlen > tp->t_maxopd) { 846 if (tp->t_maxopd <= optlen + ipoptlen) { 847 static time_t last_optlen_report; 848 849 if (last_optlen_report != time_uptime) { 850 last_optlen_report = time_uptime; 851 kprintf("tcpcb %p: MSS (%d) too " 852 "small to hold options!\n", 853 tp, tp->t_maxopd); 854 } 855 error = EHOSTUNREACH; 856 goto out; 857 } else { 858 flags &= ~TH_FIN; 859 len = tp->t_maxopd - optlen - ipoptlen; 860 sendalot = TRUE; 861 } 862 } 863 } 864 865 #ifdef INET6 866 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big")); 867 #else 868 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big")); 869 #endif 870 871 /* 872 * Grab a header mbuf, attaching a copy of data to 873 * be transmitted, and initialize the header from 874 * the template for sends on this connection. 875 */ 876 if (len) { 877 if ((tp->t_flags & TF_FORCE) && len == 1) 878 tcpstat.tcps_sndprobe++; 879 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 880 if (tp->snd_nxt == tp->snd_una) 881 tp->snd_max_rexmt = tp->snd_max; 882 if (nsacked) { 883 tcpstat.tcps_sndsackrtopack++; 884 tcpstat.tcps_sndsackrtobyte += len; 885 } 886 tcpstat.tcps_sndrexmitpack++; 887 tcpstat.tcps_sndrexmitbyte += len; 888 } else { 889 tcpstat.tcps_sndpack++; 890 tcpstat.tcps_sndbyte += len; 891 } 892 if (idle_cwv) { 893 idle_cwv = FALSE; 894 tcp_idle_cwnd_validate(tp); 895 } 896 /* Update last send time after CWV */ 897 tp->snd_last = ticks; 898 #ifdef notyet 899 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len, 900 max_linkhdr + hdrlen)) == NULL) { 901 error = ENOBUFS; 902 goto after_th; 903 } 904 /* 905 * m_copypack left space for our hdr; use it. 906 */ 907 m->m_len += hdrlen; 908 m->m_data -= hdrlen; 909 #else 910 #ifndef INET6 911 m = m_gethdr(M_NOWAIT, MT_HEADER); 912 #else 913 m = m_getl(hdrlen + max_linkhdr, M_NOWAIT, MT_HEADER, 914 M_PKTHDR, NULL); 915 #endif 916 if (m == NULL) { 917 error = ENOBUFS; 918 goto after_th; 919 } 920 m->m_data += max_linkhdr; 921 m->m_len = hdrlen; 922 if (len <= MHLEN - hdrlen - max_linkhdr) { 923 m_copydata(so->so_snd.ssb_mb, off, (int) len, 924 mtod(m, caddr_t) + hdrlen); 925 m->m_len += len; 926 } else { 927 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len); 928 if (m->m_next == NULL) { 929 m_free(m); 930 m = NULL; 931 error = ENOBUFS; 932 goto after_th; 933 } 934 } 935 #endif 936 /* 937 * If we're sending everything we've got, set PUSH. 938 * (This will keep happy those implementations which only 939 * give data to the user when a buffer fills or 940 * a PUSH comes in.) 941 */ 942 if (off + len == so->so_snd.ssb_cc) 943 flags |= TH_PUSH; 944 } else { 945 if (tp->t_flags & TF_ACKNOW) 946 tcpstat.tcps_sndacks++; 947 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 948 tcpstat.tcps_sndctrl++; 949 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 950 tcpstat.tcps_sndurg++; 951 else 952 tcpstat.tcps_sndwinup++; 953 954 MGETHDR(m, M_NOWAIT, MT_HEADER); 955 if (m == NULL) { 956 error = ENOBUFS; 957 goto after_th; 958 } 959 if (isipv6 && 960 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN) 961 MH_ALIGN(m, hdrlen); 962 else 963 m->m_data += max_linkhdr; 964 m->m_len = hdrlen; 965 966 /* 967 * Prioritize SYN, SYN|ACK and pure ACK. 968 * Leave FIN and RST as they are. 969 */ 970 if (tcp_prio_synack && (flags & (TH_FIN | TH_RST)) == 0) 971 m->m_flags |= M_PRIO; 972 } 973 m->m_pkthdr.rcvif = NULL; 974 if (isipv6) { 975 ip6 = mtod(m, struct ip6_hdr *); 976 th = (struct tcphdr *)(ip6 + 1); 977 tcp_fillheaders(tp, ip6, th, use_tso); 978 } else { 979 ip = mtod(m, struct ip *); 980 th = (struct tcphdr *)(ip + 1); 981 /* this picks up the pseudo header (w/o the length) */ 982 tcp_fillheaders(tp, ip, th, use_tso); 983 } 984 after_th: 985 /* 986 * Fill in fields, remembering maximum advertised 987 * window for use in delaying messages about window sizes. 988 * If resending a FIN, be sure not to use a new sequence number. 989 */ 990 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 991 tp->snd_nxt == tp->snd_max) 992 tp->snd_nxt--; 993 994 if (th != NULL) { 995 /* 996 * If we are doing retransmissions, then snd_nxt will 997 * not reflect the first unsent octet. For ACK only 998 * packets, we do not want the sequence number of the 999 * retransmitted packet, we want the sequence number 1000 * of the next unsent octet. So, if there is no data 1001 * (and no SYN or FIN), use snd_max instead of snd_nxt 1002 * when filling in ti_seq. But if we are in persist 1003 * state, snd_max might reflect one byte beyond the 1004 * right edge of the window, so use snd_nxt in that 1005 * case, since we know we aren't doing a retransmission. 1006 * (retransmit and persist are mutually exclusive...) 1007 */ 1008 if (len || (flags & (TH_SYN|TH_FIN)) || 1009 tcp_callout_active(tp, tp->tt_persist)) 1010 th->th_seq = htonl(tp->snd_nxt); 1011 else 1012 th->th_seq = htonl(tp->snd_max); 1013 th->th_ack = htonl(tp->rcv_nxt); 1014 if (optlen) { 1015 bcopy(opt, th + 1, optlen); 1016 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1017 } 1018 th->th_flags = flags; 1019 } 1020 1021 /* 1022 * Calculate receive window. Don't shrink window, but avoid 1023 * silly window syndrome by sending a 0 window if the actual 1024 * window is less then one segment. 1025 */ 1026 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) && 1027 recvwin < (long)segsz) 1028 recvwin = 0; 1029 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt)) 1030 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt); 1031 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale) 1032 recvwin = (long)TCP_MAXWIN << tp->rcv_scale; 1033 1034 /* 1035 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1036 * a 0 window. This may cause the remote transmitter to stall. This 1037 * flag tells soreceive() to disable delayed acknowledgements when 1038 * draining the buffer. This can occur if the receiver is attempting 1039 * to read more data then can be buffered prior to transmitting on 1040 * the connection. 1041 */ 1042 if (recvwin == 0) 1043 tp->t_flags |= TF_RXWIN0SENT; 1044 else 1045 tp->t_flags &= ~TF_RXWIN0SENT; 1046 1047 if (th != NULL) 1048 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale)); 1049 1050 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1051 KASSERT(!use_tso, ("URG with TSO")); 1052 if (th != NULL) { 1053 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1054 th->th_flags |= TH_URG; 1055 } 1056 } else { 1057 /* 1058 * If no urgent pointer to send, then we pull 1059 * the urgent pointer to the left edge of the send window 1060 * so that it doesn't drift into the send window on sequence 1061 * number wraparound. 1062 */ 1063 tp->snd_up = tp->snd_una; /* drag it along */ 1064 } 1065 1066 if (th != NULL) { 1067 #ifdef TCP_SIGNATURE 1068 if (tp->t_flags & TF_SIGNATURE) { 1069 tcpsignature_compute(m, len, optlen, 1070 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1071 } 1072 #endif /* TCP_SIGNATURE */ 1073 1074 /* 1075 * Put TCP length in extended header, and then 1076 * checksum extended header and data. 1077 */ 1078 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1079 if (isipv6) { 1080 /* 1081 * ip6_plen is not need to be filled now, and will be 1082 * filled in ip6_output(). 1083 */ 1084 th->th_sum = in6_cksum(m, IPPROTO_TCP, 1085 sizeof(struct ip6_hdr), 1086 sizeof(struct tcphdr) + optlen + len); 1087 } else { 1088 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen; 1089 if (use_tso) { 1090 m->m_pkthdr.csum_flags = CSUM_TSO; 1091 m->m_pkthdr.tso_segsz = segsz; 1092 } else { 1093 m->m_pkthdr.csum_flags = CSUM_TCP; 1094 m->m_pkthdr.csum_data = 1095 offsetof(struct tcphdr, th_sum); 1096 if (len + optlen) { 1097 th->th_sum = in_addword(th->th_sum, 1098 htons((u_short)(optlen + len))); 1099 } 1100 } 1101 1102 /* 1103 * IP version must be set here for ipv4/ipv6 checking 1104 * later 1105 */ 1106 KASSERT(ip->ip_v == IPVERSION, 1107 ("%s: IP version incorrect: %d", 1108 __func__, ip->ip_v)); 1109 } 1110 } 1111 1112 /* 1113 * In transmit state, time the transmission and arrange for 1114 * the retransmit. In persist state, just set snd_max. 1115 */ 1116 if (!(tp->t_flags & TF_FORCE) || 1117 !tcp_callout_active(tp, tp->tt_persist)) { 1118 tcp_seq startseq = tp->snd_nxt; 1119 1120 /* 1121 * Advance snd_nxt over sequence space of this segment. 1122 */ 1123 if (flags & (TH_SYN | TH_FIN)) { 1124 if (flags & TH_SYN) 1125 tp->snd_nxt++; 1126 if (flags & TH_FIN) { 1127 tp->snd_nxt++; 1128 tp->t_flags |= TF_SENTFIN; 1129 } 1130 } 1131 tp->snd_nxt += len; 1132 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1133 tp->snd_max = tp->snd_nxt; 1134 /* 1135 * Time this transmission if not a retransmission and 1136 * not currently timing anything. 1137 */ 1138 if (tp->t_rtttime == 0) { 1139 tp->t_rtttime = ticks; 1140 tp->t_rtseq = startseq; 1141 tcpstat.tcps_segstimed++; 1142 } 1143 } 1144 1145 /* 1146 * Set retransmit timer if not currently set, 1147 * and not doing a pure ack or a keep-alive probe. 1148 * Initial value for retransmit timer is smoothed 1149 * round-trip time + 2 * round-trip time variance. 1150 * Initialize shift counter which is used for backoff 1151 * of retransmit time. 1152 */ 1153 if (!tcp_callout_active(tp, tp->tt_rexmt) && 1154 tp->snd_nxt != tp->snd_una) { 1155 if (tcp_callout_active(tp, tp->tt_persist)) { 1156 tcp_callout_stop(tp, tp->tt_persist); 1157 tp->t_rxtshift = 0; 1158 } 1159 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur, 1160 tcp_timer_rexmt); 1161 } else if (len == 0 && so->so_snd.ssb_cc && 1162 tp->t_state > TCPS_SYN_RECEIVED && 1163 !tcp_callout_active(tp, tp->tt_rexmt) && 1164 !tcp_callout_active(tp, tp->tt_persist)) { 1165 /* 1166 * Avoid a situation where we do not set persist timer 1167 * after a zero window condition. For example: 1168 * 1) A -> B: packet with enough data to fill the window 1169 * 2) B -> A: ACK for #1 + new data (0 window 1170 * advertisement) 1171 * 3) A -> B: ACK for #2, 0 len packet 1172 * 1173 * In this case, A will not activate the persist timer, 1174 * because it chose to send a packet. Unless tcp_output 1175 * is called for some other reason (delayed ack timer, 1176 * another input packet from B, socket syscall), A will 1177 * not send zero window probes. 1178 * 1179 * So, if you send a 0-length packet, but there is data 1180 * in the socket buffer, and neither the rexmt or 1181 * persist timer is already set, then activate the 1182 * persist timer. 1183 */ 1184 tp->t_rxtshift = 0; 1185 tcp_setpersist(tp); 1186 } 1187 } else { 1188 /* 1189 * Persist case, update snd_max but since we are in 1190 * persist mode (no window) we do not update snd_nxt. 1191 */ 1192 int xlen = len; 1193 if (flags & TH_SYN) 1194 panic("tcp_output: persist timer to send SYN"); 1195 if (flags & TH_FIN) { 1196 ++xlen; 1197 tp->t_flags |= TF_SENTFIN; 1198 } 1199 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1200 tp->snd_max = tp->snd_nxt + xlen; 1201 } 1202 1203 if (th != NULL) { 1204 #ifdef TCPDEBUG 1205 /* Trace. */ 1206 if (so->so_options & SO_DEBUG) { 1207 tcp_trace(TA_OUTPUT, tp->t_state, tp, 1208 mtod(m, void *), th, 0); 1209 } 1210 #endif 1211 1212 /* 1213 * Fill in IP length and desired time to live and 1214 * send to IP level. There should be a better way 1215 * to handle ttl and tos; we could keep them in 1216 * the template, but need a way to checksum without them. 1217 */ 1218 /* 1219 * m->m_pkthdr.len should have been set before cksum 1220 * calcuration, because in6_cksum() need it. 1221 */ 1222 if (isipv6) { 1223 /* 1224 * we separately set hoplimit for every segment, 1225 * since the user might want to change the value 1226 * via setsockopt. Also, desired default hop 1227 * limit might be changed via Neighbor Discovery. 1228 */ 1229 ip6->ip6_hlim = in6_selecthlim(inp, 1230 (inp->in6p_route.ro_rt ? 1231 inp->in6p_route.ro_rt->rt_ifp : NULL)); 1232 1233 /* TODO: IPv6 IP6TOS_ECT bit on */ 1234 error = ip6_output(m, inp->in6p_outputopts, 1235 &inp->in6p_route, (so->so_options & SO_DONTROUTE), 1236 NULL, NULL, inp); 1237 } else { 1238 struct rtentry *rt; 1239 1240 KASSERT(!INP_CHECK_SOCKAF(so, AF_INET6), ("inet6 pcb")); 1241 1242 ip->ip_len = m->m_pkthdr.len; 1243 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */ 1244 ip->ip_tos = inp->inp_ip_tos; /* XXX */ 1245 /* 1246 * See if we should do MTU discovery. 1247 * We do it only if the following are true: 1248 * 1) we have a valid route to the destination 1249 * 2) the MTU is not locked (if it is, 1250 * then discovery has been disabled) 1251 */ 1252 if (path_mtu_discovery && 1253 (rt = inp->inp_route.ro_rt) && 1254 (rt->rt_flags & RTF_UP) && 1255 !(rt->rt_rmx.rmx_locks & RTV_MTU)) 1256 ip->ip_off |= IP_DF; 1257 1258 KASSERT(inp->inp_flags & INP_HASH, 1259 ("inpcb has no hash")); 1260 m_sethash(m, inp->inp_hashval); 1261 error = ip_output(m, inp->inp_options, &inp->inp_route, 1262 (so->so_options & SO_DONTROUTE) | 1263 IP_DEBUGROUTE, NULL, inp); 1264 } 1265 } else { 1266 KASSERT(error != 0, ("no error, but th not set")); 1267 } 1268 if (error) { 1269 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW); 1270 1271 /* 1272 * We know that the packet was lost, so back out the 1273 * sequence number advance, if any. 1274 */ 1275 if (!(tp->t_flags & TF_FORCE) || 1276 !tcp_callout_active(tp, tp->tt_persist)) { 1277 /* 1278 * No need to check for TH_FIN here because 1279 * the TF_SENTFIN flag handles that case. 1280 */ 1281 if (!(flags & TH_SYN)) 1282 tp->snd_nxt -= len; 1283 } 1284 1285 out: 1286 if (error == ENOBUFS) { 1287 KASSERT((len == 0 && (flags & (TH_SYN | TH_FIN)) == 0) || 1288 tcp_callout_active(tp, tp->tt_rexmt) || 1289 tcp_callout_active(tp, tp->tt_persist), 1290 ("neither rexmt nor persist timer is set")); 1291 return (0); 1292 } 1293 if (error == EMSGSIZE) { 1294 /* 1295 * ip_output() will have already fixed the route 1296 * for us. tcp_mtudisc() will, as its last action, 1297 * initiate retransmission, so it is important to 1298 * not do so here. 1299 */ 1300 tcp_mtudisc(inp, 0); 1301 return 0; 1302 } 1303 if ((error == EHOSTUNREACH || error == ENETDOWN) && 1304 TCPS_HAVERCVDSYN(tp->t_state)) { 1305 tp->t_softerror = error; 1306 return (0); 1307 } 1308 return (error); 1309 } 1310 tcpstat.tcps_sndtotal++; 1311 1312 /* 1313 * Data sent (as far as we can tell). 1314 * 1315 * If this advertises a larger window than any other segment, 1316 * then remember the size of the advertised window. 1317 * 1318 * Any pending ACK has now been sent. 1319 */ 1320 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) { 1321 tp->rcv_adv = tp->rcv_nxt + recvwin; 1322 tp->t_flags &= ~TF_RXRESIZED; 1323 } 1324 tp->last_ack_sent = tp->rcv_nxt; 1325 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW); 1326 if (tcp_delack_enabled) 1327 tcp_callout_stop(tp, tp->tt_delack); 1328 if (sendalot) { 1329 if (tcp_fairsend > 0 && (tp->t_flags & TF_FAIRSEND) && 1330 segcnt >= tcp_fairsend) 1331 need_sched = TRUE; 1332 goto again; 1333 } 1334 return (0); 1335 } 1336 1337 void 1338 tcp_setpersist(struct tcpcb *tp) 1339 { 1340 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1341 int tt; 1342 1343 if (tp->t_state == TCPS_SYN_SENT || 1344 tp->t_state == TCPS_SYN_RECEIVED) { 1345 panic("tcp_setpersist: not established yet, current %s", 1346 tp->t_state == TCPS_SYN_SENT ? 1347 "SYN_SENT" : "SYN_RECEIVED"); 1348 } 1349 1350 if (tcp_callout_active(tp, tp->tt_rexmt)) 1351 panic("tcp_setpersist: retransmit pending"); 1352 /* 1353 * Start/restart persistance timer. 1354 */ 1355 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN, 1356 TCPTV_PERSMAX); 1357 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist); 1358 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1359 tp->t_rxtshift++; 1360 } 1361 1362 static void 1363 tcp_idle_cwnd_validate(struct tcpcb *tp) 1364 { 1365 u_long initial_cwnd = tcp_initial_window(tp); 1366 u_long min_cwnd; 1367 1368 tcpstat.tcps_sndidle++; 1369 1370 /* According to RFC5681: RW=min(IW,cwnd) */ 1371 min_cwnd = min(tp->snd_cwnd, initial_cwnd); 1372 1373 if (tcp_idle_cwv) { 1374 u_long idle_time, decay_cwnd; 1375 1376 /* 1377 * RFC2861, but only after idle period. 1378 */ 1379 1380 /* 1381 * Before the congestion window is reduced, ssthresh 1382 * is set to the maximum of its current value and 3/4 1383 * cwnd. If the sender then has more data to send 1384 * than the decayed cwnd allows, the TCP will slow- 1385 * start (perform exponential increase) at least 1386 * half-way back up to the old value of cwnd. 1387 */ 1388 tp->snd_ssthresh = max(tp->snd_ssthresh, 1389 (3 * tp->snd_cwnd) / 4); 1390 1391 /* 1392 * Decay the congestion window by half for every RTT 1393 * that the flow remains inactive. 1394 * 1395 * The difference between our implementation and 1396 * RFC2861 is that we don't allow cwnd to go below 1397 * the value allowed by RFC5681 (min_cwnd). 1398 */ 1399 idle_time = ticks - tp->snd_last; 1400 decay_cwnd = tp->snd_cwnd; 1401 while (idle_time >= tp->t_rxtcur && 1402 decay_cwnd > min_cwnd) { 1403 decay_cwnd >>= 1; 1404 idle_time -= tp->t_rxtcur; 1405 } 1406 tp->snd_cwnd = max(decay_cwnd, min_cwnd); 1407 } else { 1408 /* 1409 * Slow-start from scratch to re-determine the send 1410 * congestion window. 1411 */ 1412 tp->snd_cwnd = min_cwnd; 1413 } 1414 1415 /* Restart ABC counting during congestion avoidance */ 1416 tp->snd_wacked = 0; 1417 } 1418 1419 static int 1420 tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen0) 1421 { 1422 struct inpcb * const inp = tp->t_inpcb; 1423 #ifdef INET6 1424 const boolean_t isipv6 = INP_ISIPV6(inp); 1425 #else 1426 const boolean_t isipv6 = FALSE; 1427 #endif 1428 unsigned int ipoptlen, optlen; 1429 u_int hlen; 1430 1431 hlen = sizeof(struct ip) + sizeof(struct tcphdr); 1432 1433 if (isipv6) { 1434 ipoptlen = ip6_optlen(inp); 1435 } else { 1436 if (inp->inp_options) { 1437 ipoptlen = inp->inp_options->m_len - 1438 offsetof(struct ipoption, ipopt_list); 1439 } else { 1440 ipoptlen = 0; 1441 } 1442 } 1443 #ifdef IPSEC 1444 ipoptlen += ipsec_hdrsiz_tcp(tp); 1445 #endif 1446 hlen += ipoptlen; 1447 1448 optlen = 0; 1449 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 1450 (tp->t_flags & TF_RCVD_TSTMP)) 1451 optlen += TCPOLEN_TSTAMP_APPA; 1452 hlen += optlen; 1453 1454 if (tp->t_maxopd <= optlen + ipoptlen) 1455 return EHOSTUNREACH; 1456 1457 *segsz = tp->t_maxopd - optlen - ipoptlen; 1458 *hlen0 = hlen; 1459 return 0; 1460 } 1461 1462 static void 1463 tcp_output_sched_handler(netmsg_t nmsg) 1464 { 1465 struct tcpcb *tp = nmsg->lmsg.u.ms_resultp; 1466 1467 /* Reply ASAP */ 1468 crit_enter(); 1469 lwkt_replymsg(&nmsg->lmsg, 0); 1470 crit_exit(); 1471 1472 tcp_output_fair(tp); 1473 } 1474 1475 void 1476 tcp_output_init(struct tcpcb *tp) 1477 { 1478 netmsg_init(tp->tt_sndmore, NULL, &netisr_adone_rport, MSGF_DROPABLE, 1479 tcp_output_sched_handler); 1480 tp->tt_sndmore->lmsg.u.ms_resultp = tp; 1481 } 1482 1483 void 1484 tcp_output_cancel(struct tcpcb *tp) 1485 { 1486 /* 1487 * This message is still pending to be processed; 1488 * drop it. Optimized. 1489 */ 1490 crit_enter(); 1491 if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0) { 1492 lwkt_dropmsg(&tp->tt_sndmore->lmsg); 1493 } 1494 crit_exit(); 1495 } 1496 1497 boolean_t 1498 tcp_output_pending(struct tcpcb *tp) 1499 { 1500 if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0) 1501 return TRUE; 1502 else 1503 return FALSE; 1504 } 1505 1506 static void 1507 tcp_output_sched(struct tcpcb *tp) 1508 { 1509 crit_enter(); 1510 if (tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) 1511 lwkt_sendmsg(netisr_cpuport(mycpuid), &tp->tt_sndmore->lmsg); 1512 crit_exit(); 1513 } 1514 1515 /* 1516 * Fairsend 1517 * 1518 * Yield to other senders or receivers on the same netisr if the current 1519 * TCP stream has sent tcp_fairsend segments and is going to burst more 1520 * segments. Bursting large amount of segements in a single TCP stream 1521 * could delay other senders' segments and receivers' ACKs quite a lot, 1522 * if others segments and ACKs are queued on to the same hardware transmit 1523 * queue; thus cause unfairness between senders and suppress receiving 1524 * performance. 1525 * 1526 * Fairsend should be performed at the places that do not affect segment 1527 * sending during congestion control, e.g. 1528 * - User requested output 1529 * - ACK input triggered output 1530 * 1531 * NOTE: 1532 * For devices that are TSO capable, their TSO aggregation size limit could 1533 * affect fairsend. 1534 */ 1535 int 1536 tcp_output_fair(struct tcpcb *tp) 1537 { 1538 int ret; 1539 1540 tp->t_flags |= TF_FAIRSEND; 1541 ret = tcp_output(tp); 1542 tp->t_flags &= ~TF_FAIRSEND; 1543 1544 return ret; 1545 } 1546