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 */ 31 32#include <sys/cdefs.h> 33__FBSDID("$FreeBSD: releng/10.2/sys/netinet/tcp_output.c 293893 2016-01-14 09:10:46Z glebius $"); 34 35#include "opt_inet.h" 36#include "opt_inet6.h" 37#include "opt_ipsec.h" 38#include "opt_kdtrace.h" 39#include "opt_tcpdebug.h" 40 41#include <sys/param.h> 42#include <sys/systm.h> 43#include <sys/domain.h> 44#include <sys/hhook.h> 45#include <sys/kernel.h> 46#include <sys/lock.h> 47#include <sys/mbuf.h> 48#include <sys/mutex.h> 49#include <sys/protosw.h> 50#include <sys/sdt.h> 51#include <sys/socket.h> 52#include <sys/socketvar.h> 53#include <sys/sysctl.h> 54 55#include <net/if.h> 56#include <net/route.h> 57#include <net/vnet.h> 58 59#include <netinet/cc.h> 60#include <netinet/in.h> 61#include <netinet/in_kdtrace.h> 62#include <netinet/in_systm.h> 63#include <netinet/ip.h> 64#include <netinet/in_pcb.h> 65#include <netinet/ip_var.h> 66#include <netinet/ip_options.h> 67#ifdef INET6 68#include <netinet6/in6_pcb.h> 69#include <netinet/ip6.h> 70#include <netinet6/ip6_var.h> 71#endif 72#define TCPOUTFLAGS 73#include <netinet/tcp_fsm.h> 74#include <netinet/tcp_seq.h> 75#include <netinet/tcp_timer.h> 76#include <netinet/tcp_var.h> 77#include <netinet/tcpip.h> 78#ifdef TCPDEBUG 79#include <netinet/tcp_debug.h> 80#endif 81#ifdef TCP_OFFLOAD 82#include <netinet/tcp_offload.h> 83#endif 84 85#ifdef IPSEC 86#include <netipsec/ipsec.h> 87#endif /*IPSEC*/ 88 89#include <machine/in_cksum.h> 90 91#include <security/mac/mac_framework.h> 92 93VNET_DEFINE(int, path_mtu_discovery) = 1; 94SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 95 &VNET_NAME(path_mtu_discovery), 1, 96 "Enable Path MTU Discovery"); 97 98VNET_DEFINE(int, tcp_do_tso) = 1; 99#define V_tcp_do_tso VNET(tcp_do_tso) 100SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW, 101 &VNET_NAME(tcp_do_tso), 0, 102 "Enable TCP Segmentation Offload"); 103 104VNET_DEFINE(int, tcp_sendspace) = 1024*32; 105#define V_tcp_sendspace VNET(tcp_sendspace) 106SYSCTL_VNET_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_RW, 107 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 108 109VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 110#define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) 111SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 112 &VNET_NAME(tcp_do_autosndbuf), 0, 113 "Enable automatic send buffer sizing"); 114 115VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 116#define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) 117SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 118 &VNET_NAME(tcp_autosndbuf_inc), 0, 119 "Incrementor step size of automatic send buffer"); 120 121VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; 122#define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) 123SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 124 &VNET_NAME(tcp_autosndbuf_max), 0, 125 "Max size of automatic send buffer"); 126 127static void inline hhook_run_tcp_est_out(struct tcpcb *tp, 128 struct tcphdr *th, struct tcpopt *to, 129 long len, int tso); 130static void inline cc_after_idle(struct tcpcb *tp); 131 132/* 133 * Wrapper for the TCP established output helper hook. 134 */ 135static void inline 136hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 137 struct tcpopt *to, long len, int tso) 138{ 139 struct tcp_hhook_data hhook_data; 140 141 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 142 hhook_data.tp = tp; 143 hhook_data.th = th; 144 hhook_data.to = to; 145 hhook_data.len = len; 146 hhook_data.tso = tso; 147 148 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 149 tp->osd); 150 } 151} 152 153/* 154 * CC wrapper hook functions 155 */ 156static void inline 157cc_after_idle(struct tcpcb *tp) 158{ 159 INP_WLOCK_ASSERT(tp->t_inpcb); 160 161 if (CC_ALGO(tp)->after_idle != NULL) 162 CC_ALGO(tp)->after_idle(tp->ccv); 163} 164 165/* 166 * Tcp output routine: figure out what should be sent and send it. 167 */ 168int 169tcp_output(struct tcpcb *tp) 170{ 171 struct socket *so = tp->t_inpcb->inp_socket; 172 long len, recwin, sendwin; 173 int off, flags, error = 0; /* Keep compiler happy */ 174 struct mbuf *m; 175 struct ip *ip = NULL; 176 struct ipovly *ipov = NULL; 177 struct tcphdr *th; 178 u_char opt[TCP_MAXOLEN]; 179 unsigned ipoptlen, optlen, hdrlen; 180#ifdef IPSEC 181 unsigned ipsec_optlen = 0; 182#endif 183 int idle, sendalot; 184 int sack_rxmit, sack_bytes_rxmt; 185 struct sackhole *p; 186 int tso, mtu; 187 struct tcpopt to; 188#if 0 189 int maxburst = TCP_MAXBURST; 190#endif 191#ifdef INET6 192 struct ip6_hdr *ip6 = NULL; 193 int isipv6; 194 195 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 196#endif 197 198 INP_WLOCK_ASSERT(tp->t_inpcb); 199 200#ifdef TCP_OFFLOAD 201 if (tp->t_flags & TF_TOE) 202 return (tcp_offload_output(tp)); 203#endif 204 205 /* 206 * Determine length of data that should be transmitted, 207 * and flags that will be used. 208 * If there is some data or critical controls (SYN, RST) 209 * to send, then transmit; otherwise, investigate further. 210 */ 211 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 212 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) 213 cc_after_idle(tp); 214 tp->t_flags &= ~TF_LASTIDLE; 215 if (idle) { 216 if (tp->t_flags & TF_MORETOCOME) { 217 tp->t_flags |= TF_LASTIDLE; 218 idle = 0; 219 } 220 } 221again: 222 /* 223 * If we've recently taken a timeout, snd_max will be greater than 224 * snd_nxt. There may be SACK information that allows us to avoid 225 * resending already delivered data. Adjust snd_nxt accordingly. 226 */ 227 if ((tp->t_flags & TF_SACK_PERMIT) && 228 SEQ_LT(tp->snd_nxt, tp->snd_max)) 229 tcp_sack_adjust(tp); 230 sendalot = 0; 231 tso = 0; 232 mtu = 0; 233 off = tp->snd_nxt - tp->snd_una; 234 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 235 236 flags = tcp_outflags[tp->t_state]; 237 /* 238 * Send any SACK-generated retransmissions. If we're explicitly trying 239 * to send out new data (when sendalot is 1), bypass this function. 240 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 241 * we're replacing a (future) new transmission with a retransmission 242 * now, and we previously incremented snd_cwnd in tcp_input(). 243 */ 244 /* 245 * Still in sack recovery , reset rxmit flag to zero. 246 */ 247 sack_rxmit = 0; 248 sack_bytes_rxmt = 0; 249 len = 0; 250 p = NULL; 251 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 252 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 253 long cwin; 254 255 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 256 if (cwin < 0) 257 cwin = 0; 258 /* Do not retransmit SACK segments beyond snd_recover */ 259 if (SEQ_GT(p->end, tp->snd_recover)) { 260 /* 261 * (At least) part of sack hole extends beyond 262 * snd_recover. Check to see if we can rexmit data 263 * for this hole. 264 */ 265 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 266 /* 267 * Can't rexmit any more data for this hole. 268 * That data will be rexmitted in the next 269 * sack recovery episode, when snd_recover 270 * moves past p->rxmit. 271 */ 272 p = NULL; 273 goto after_sack_rexmit; 274 } else 275 /* Can rexmit part of the current hole */ 276 len = ((long)ulmin(cwin, 277 tp->snd_recover - p->rxmit)); 278 } else 279 len = ((long)ulmin(cwin, p->end - p->rxmit)); 280 off = p->rxmit - tp->snd_una; 281 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 282 __func__, off)); 283 if (len > 0) { 284 sack_rxmit = 1; 285 sendalot = 1; 286 TCPSTAT_INC(tcps_sack_rexmits); 287 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 288 min(len, tp->t_maxseg)); 289 } 290 } 291after_sack_rexmit: 292 /* 293 * Get standard flags, and add SYN or FIN if requested by 'hidden' 294 * state flags. 295 */ 296 if (tp->t_flags & TF_NEEDFIN) 297 flags |= TH_FIN; 298 if (tp->t_flags & TF_NEEDSYN) 299 flags |= TH_SYN; 300 301 SOCKBUF_LOCK(&so->so_snd); 302 /* 303 * If in persist timeout with window of 0, send 1 byte. 304 * Otherwise, if window is small but nonzero 305 * and timer expired, we will send what we can 306 * and go to transmit state. 307 */ 308 if (tp->t_flags & TF_FORCEDATA) { 309 if (sendwin == 0) { 310 /* 311 * If we still have some data to send, then 312 * clear the FIN bit. Usually this would 313 * happen below when it realizes that we 314 * aren't sending all the data. However, 315 * if we have exactly 1 byte of unsent data, 316 * then it won't clear the FIN bit below, 317 * and if we are in persist state, we wind 318 * up sending the packet without recording 319 * that we sent the FIN bit. 320 * 321 * We can't just blindly clear the FIN bit, 322 * because if we don't have any more data 323 * to send then the probe will be the FIN 324 * itself. 325 */ 326 if (off < so->so_snd.sb_cc) 327 flags &= ~TH_FIN; 328 sendwin = 1; 329 } else { 330 tcp_timer_activate(tp, TT_PERSIST, 0); 331 tp->t_rxtshift = 0; 332 } 333 } 334 335 /* 336 * If snd_nxt == snd_max and we have transmitted a FIN, the 337 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 338 * a negative length. This can also occur when TCP opens up 339 * its congestion window while receiving additional duplicate 340 * acks after fast-retransmit because TCP will reset snd_nxt 341 * to snd_max after the fast-retransmit. 342 * 343 * In the normal retransmit-FIN-only case, however, snd_nxt will 344 * be set to snd_una, the offset will be 0, and the length may 345 * wind up 0. 346 * 347 * If sack_rxmit is true we are retransmitting from the scoreboard 348 * in which case len is already set. 349 */ 350 if (sack_rxmit == 0) { 351 if (sack_bytes_rxmt == 0) 352 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off); 353 else { 354 long cwin; 355 356 /* 357 * We are inside of a SACK recovery episode and are 358 * sending new data, having retransmitted all the 359 * data possible in the scoreboard. 360 */ 361 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 362 - off); 363 /* 364 * Don't remove this (len > 0) check ! 365 * We explicitly check for len > 0 here (although it 366 * isn't really necessary), to work around a gcc 367 * optimization issue - to force gcc to compute 368 * len above. Without this check, the computation 369 * of len is bungled by the optimizer. 370 */ 371 if (len > 0) { 372 cwin = tp->snd_cwnd - 373 (tp->snd_nxt - tp->sack_newdata) - 374 sack_bytes_rxmt; 375 if (cwin < 0) 376 cwin = 0; 377 len = lmin(len, cwin); 378 } 379 } 380 } 381 382 /* 383 * Lop off SYN bit if it has already been sent. However, if this 384 * is SYN-SENT state and if segment contains data and if we don't 385 * know that foreign host supports TAO, suppress sending segment. 386 */ 387 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 388 if (tp->t_state != TCPS_SYN_RECEIVED) 389 flags &= ~TH_SYN; 390 off--, len++; 391 } 392 393 /* 394 * Be careful not to send data and/or FIN on SYN segments. 395 * This measure is needed to prevent interoperability problems 396 * with not fully conformant TCP implementations. 397 */ 398 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 399 len = 0; 400 flags &= ~TH_FIN; 401 } 402 403 if (len <= 0) { 404 /* 405 * If FIN has been sent but not acked, 406 * but we haven't been called to retransmit, 407 * len will be < 0. Otherwise, window shrank 408 * after we sent into it. If window shrank to 0, 409 * cancel pending retransmit, pull snd_nxt back 410 * to (closed) window, and set the persist timer 411 * if it isn't already going. If the window didn't 412 * close completely, just wait for an ACK. 413 * 414 * We also do a general check here to ensure that 415 * we will set the persist timer when we have data 416 * to send, but a 0-byte window. This makes sure 417 * the persist timer is set even if the packet 418 * hits one of the "goto send" lines below. 419 */ 420 len = 0; 421 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 422 (off < (int) so->so_snd.sb_cc)) { 423 tcp_timer_activate(tp, TT_REXMT, 0); 424 tp->t_rxtshift = 0; 425 tp->snd_nxt = tp->snd_una; 426 if (!tcp_timer_active(tp, TT_PERSIST)) 427 tcp_setpersist(tp); 428 } 429 } 430 431 /* len will be >= 0 after this point. */ 432 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 433 434 /* 435 * Automatic sizing of send socket buffer. Often the send buffer 436 * size is not optimally adjusted to the actual network conditions 437 * at hand (delay bandwidth product). Setting the buffer size too 438 * small limits throughput on links with high bandwidth and high 439 * delay (eg. trans-continental/oceanic links). Setting the 440 * buffer size too big consumes too much real kernel memory, 441 * especially with many connections on busy servers. 442 * 443 * The criteria to step up the send buffer one notch are: 444 * 1. receive window of remote host is larger than send buffer 445 * (with a fudge factor of 5/4th); 446 * 2. send buffer is filled to 7/8th with data (so we actually 447 * have data to make use of it); 448 * 3. send buffer fill has not hit maximal automatic size; 449 * 4. our send window (slow start and cogestion controlled) is 450 * larger than sent but unacknowledged data in send buffer. 451 * 452 * The remote host receive window scaling factor may limit the 453 * growing of the send buffer before it reaches its allowed 454 * maximum. 455 * 456 * It scales directly with slow start or congestion window 457 * and does at most one step per received ACK. This fast 458 * scaling has the drawback of growing the send buffer beyond 459 * what is strictly necessary to make full use of a given 460 * delay*bandwith product. However testing has shown this not 461 * to be much of an problem. At worst we are trading wasting 462 * of available bandwith (the non-use of it) for wasting some 463 * socket buffer memory. 464 * 465 * TODO: Shrink send buffer during idle periods together 466 * with congestion window. Requires another timer. Has to 467 * wait for upcoming tcp timer rewrite. 468 */ 469 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 470 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 471 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 472 so->so_snd.sb_cc < V_tcp_autosndbuf_max && 473 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 474 if (!sbreserve_locked(&so->so_snd, 475 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 476 V_tcp_autosndbuf_max), so, curthread)) 477 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 478 } 479 } 480 481 /* 482 * Decide if we can use TCP Segmentation Offloading (if supported by 483 * hardware). 484 * 485 * TSO may only be used if we are in a pure bulk sending state. The 486 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 487 * IP options prevent using TSO. With TSO the TCP header is the same 488 * (except for the sequence number) for all generated packets. This 489 * makes it impossible to transmit any options which vary per generated 490 * segment or packet. 491 */ 492#ifdef IPSEC 493 /* 494 * Pre-calculate here as we save another lookup into the darknesses 495 * of IPsec that way and can actually decide if TSO is ok. 496 */ 497 ipsec_optlen = ipsec_hdrsiz_tcp(tp); 498#endif 499 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 500 ((tp->t_flags & TF_SIGNATURE) == 0) && 501 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 502#ifdef IPSEC 503 ipsec_optlen == 0 && 504#endif 505 tp->t_inpcb->inp_options == NULL && 506 tp->t_inpcb->in6p_options == NULL) 507 tso = 1; 508 509 if (sack_rxmit) { 510 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 511 flags &= ~TH_FIN; 512 } else { 513 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 514 flags &= ~TH_FIN; 515 } 516 517 recwin = sbspace(&so->so_rcv); 518 519 /* 520 * Sender silly window avoidance. We transmit under the following 521 * conditions when len is non-zero: 522 * 523 * - We have a full segment (or more with TSO) 524 * - This is the last buffer in a write()/send() and we are 525 * either idle or running NODELAY 526 * - we've timed out (e.g. persist timer) 527 * - we have more then 1/2 the maximum send window's worth of 528 * data (receiver may be limited the window size) 529 * - we need to retransmit 530 */ 531 if (len) { 532 if (len >= tp->t_maxseg) 533 goto send; 534 /* 535 * NOTE! on localhost connections an 'ack' from the remote 536 * end may occur synchronously with the output and cause 537 * us to flush a buffer queued with moretocome. XXX 538 * 539 * note: the len + off check is almost certainly unnecessary. 540 */ 541 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 542 (idle || (tp->t_flags & TF_NODELAY)) && 543 len + off >= so->so_snd.sb_cc && 544 (tp->t_flags & TF_NOPUSH) == 0) { 545 goto send; 546 } 547 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 548 goto send; 549 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 550 goto send; 551 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 552 goto send; 553 if (sack_rxmit) 554 goto send; 555 } 556 557 /* 558 * Sending of standalone window updates. 559 * 560 * Window updates are important when we close our window due to a 561 * full socket buffer and are opening it again after the application 562 * reads data from it. Once the window has opened again and the 563 * remote end starts to send again the ACK clock takes over and 564 * provides the most current window information. 565 * 566 * We must avoid the silly window syndrome whereas every read 567 * from the receive buffer, no matter how small, causes a window 568 * update to be sent. We also should avoid sending a flurry of 569 * window updates when the socket buffer had queued a lot of data 570 * and the application is doing small reads. 571 * 572 * Prevent a flurry of pointless window updates by only sending 573 * an update when we can increase the advertized window by more 574 * than 1/4th of the socket buffer capacity. When the buffer is 575 * getting full or is very small be more aggressive and send an 576 * update whenever we can increase by two mss sized segments. 577 * In all other situations the ACK's to new incoming data will 578 * carry further window increases. 579 * 580 * Don't send an independent window update if a delayed 581 * ACK is pending (it will get piggy-backed on it) or the 582 * remote side already has done a half-close and won't send 583 * more data. Skip this if the connection is in T/TCP 584 * half-open state. 585 */ 586 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 587 !(tp->t_flags & TF_DELACK) && 588 !TCPS_HAVERCVDFIN(tp->t_state)) { 589 /* 590 * "adv" is the amount we could increase the window, 591 * taking into account that we are limited by 592 * TCP_MAXWIN << tp->rcv_scale. 593 */ 594 long adv; 595 int oldwin; 596 597 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 598 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 599 oldwin = (tp->rcv_adv - tp->rcv_nxt); 600 adv -= oldwin; 601 } else 602 oldwin = 0; 603 604 /* 605 * If the new window size ends up being the same as the old 606 * size when it is scaled, then don't force a window update. 607 */ 608 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 609 goto dontupdate; 610 611 if (adv >= (long)(2 * tp->t_maxseg) && 612 (adv >= (long)(so->so_rcv.sb_hiwat / 4) || 613 recwin <= (long)(so->so_rcv.sb_hiwat / 8) || 614 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 615 goto send; 616 } 617dontupdate: 618 619 /* 620 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 621 * is also a catch-all for the retransmit timer timeout case. 622 */ 623 if (tp->t_flags & TF_ACKNOW) 624 goto send; 625 if ((flags & TH_RST) || 626 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 627 goto send; 628 if (SEQ_GT(tp->snd_up, tp->snd_una)) 629 goto send; 630 /* 631 * If our state indicates that FIN should be sent 632 * and we have not yet done so, then we need to send. 633 */ 634 if (flags & TH_FIN && 635 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 636 goto send; 637 /* 638 * In SACK, it is possible for tcp_output to fail to send a segment 639 * after the retransmission timer has been turned off. Make sure 640 * that the retransmission timer is set. 641 */ 642 if ((tp->t_flags & TF_SACK_PERMIT) && 643 SEQ_GT(tp->snd_max, tp->snd_una) && 644 !tcp_timer_active(tp, TT_REXMT) && 645 !tcp_timer_active(tp, TT_PERSIST)) { 646 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 647 goto just_return; 648 } 649 /* 650 * TCP window updates are not reliable, rather a polling protocol 651 * using ``persist'' packets is used to insure receipt of window 652 * updates. The three ``states'' for the output side are: 653 * idle not doing retransmits or persists 654 * persisting to move a small or zero window 655 * (re)transmitting and thereby not persisting 656 * 657 * tcp_timer_active(tp, TT_PERSIST) 658 * is true when we are in persist state. 659 * (tp->t_flags & TF_FORCEDATA) 660 * is set when we are called to send a persist packet. 661 * tcp_timer_active(tp, TT_REXMT) 662 * is set when we are retransmitting 663 * The output side is idle when both timers are zero. 664 * 665 * If send window is too small, there is data to transmit, and no 666 * retransmit or persist is pending, then go to persist state. 667 * If nothing happens soon, send when timer expires: 668 * if window is nonzero, transmit what we can, 669 * otherwise force out a byte. 670 */ 671 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) && 672 !tcp_timer_active(tp, TT_PERSIST)) { 673 tp->t_rxtshift = 0; 674 tcp_setpersist(tp); 675 } 676 677 /* 678 * No reason to send a segment, just return. 679 */ 680just_return: 681 SOCKBUF_UNLOCK(&so->so_snd); 682 return (0); 683 684send: 685 SOCKBUF_LOCK_ASSERT(&so->so_snd); 686 if (len > 0) { 687 if (len >= tp->t_maxseg) 688 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 689 else 690 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 691 } 692 /* 693 * Before ESTABLISHED, force sending of initial options 694 * unless TCP set not to do any options. 695 * NOTE: we assume that the IP/TCP header plus TCP options 696 * always fit in a single mbuf, leaving room for a maximum 697 * link header, i.e. 698 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 699 */ 700 optlen = 0; 701#ifdef INET6 702 if (isipv6) 703 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 704 else 705#endif 706 hdrlen = sizeof (struct tcpiphdr); 707 708 /* 709 * Compute options for segment. 710 * We only have to care about SYN and established connection 711 * segments. Options for SYN-ACK segments are handled in TCP 712 * syncache. 713 */ 714 to.to_flags = 0; 715 if ((tp->t_flags & TF_NOOPT) == 0) { 716 /* Maximum segment size. */ 717 if (flags & TH_SYN) { 718 tp->snd_nxt = tp->iss; 719 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 720 to.to_flags |= TOF_MSS; 721 } 722 /* Window scaling. */ 723 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 724 to.to_wscale = tp->request_r_scale; 725 to.to_flags |= TOF_SCALE; 726 } 727 /* Timestamps. */ 728 if ((tp->t_flags & TF_RCVD_TSTMP) || 729 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 730 to.to_tsval = tcp_ts_getticks() + tp->ts_offset; 731 to.to_tsecr = tp->ts_recent; 732 to.to_flags |= TOF_TS; 733 /* Set receive buffer autosizing timestamp. */ 734 if (tp->rfbuf_ts == 0 && 735 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 736 tp->rfbuf_ts = tcp_ts_getticks(); 737 } 738 /* Selective ACK's. */ 739 if (tp->t_flags & TF_SACK_PERMIT) { 740 if (flags & TH_SYN) 741 to.to_flags |= TOF_SACKPERM; 742 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 743 (tp->t_flags & TF_SACK_PERMIT) && 744 tp->rcv_numsacks > 0) { 745 to.to_flags |= TOF_SACK; 746 to.to_nsacks = tp->rcv_numsacks; 747 to.to_sacks = (u_char *)tp->sackblks; 748 } 749 } 750#ifdef TCP_SIGNATURE 751 /* TCP-MD5 (RFC2385). */ 752 if (tp->t_flags & TF_SIGNATURE) 753 to.to_flags |= TOF_SIGNATURE; 754#endif /* TCP_SIGNATURE */ 755 756 /* Processing the options. */ 757 hdrlen += optlen = tcp_addoptions(&to, opt); 758 } 759 760#ifdef INET6 761 if (isipv6) 762 ipoptlen = ip6_optlen(tp->t_inpcb); 763 else 764#endif 765 if (tp->t_inpcb->inp_options) 766 ipoptlen = tp->t_inpcb->inp_options->m_len - 767 offsetof(struct ipoption, ipopt_list); 768 else 769 ipoptlen = 0; 770#ifdef IPSEC 771 ipoptlen += ipsec_optlen; 772#endif 773 774 /* 775 * Adjust data length if insertion of options will 776 * bump the packet length beyond the t_maxopd length. 777 * Clear the FIN bit because we cut off the tail of 778 * the segment. 779 */ 780 if (len + optlen + ipoptlen > tp->t_maxopd) { 781 flags &= ~TH_FIN; 782 783 if (tso) { 784 u_int if_hw_tsomax; 785 u_int if_hw_tsomaxsegcount; 786 u_int if_hw_tsomaxsegsize; 787 struct mbuf *mb; 788 u_int moff; 789 int max_len; 790 791 /* extract TSO information */ 792 if_hw_tsomax = tp->t_tsomax; 793 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 794 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 795 796 /* 797 * Limit a TSO burst to prevent it from 798 * overflowing or exceeding the maximum length 799 * allowed by the network interface: 800 */ 801 KASSERT(ipoptlen == 0, 802 ("%s: TSO can't do IP options", __func__)); 803 804 /* 805 * Check if we should limit by maximum payload 806 * length: 807 */ 808 if (if_hw_tsomax != 0) { 809 /* compute maximum TSO length */ 810 max_len = (if_hw_tsomax - hdrlen); 811 if (max_len <= 0) { 812 len = 0; 813 } else if (len > max_len) { 814 sendalot = 1; 815 len = max_len; 816 } 817 } 818 819 /* 820 * Check if we should limit by maximum segment 821 * size and count: 822 */ 823 if (if_hw_tsomaxsegcount != 0 && 824 if_hw_tsomaxsegsize != 0) { 825 max_len = 0; 826 mb = sbsndmbuf(&so->so_snd, off, &moff); 827 828 while (mb != NULL && max_len < len) { 829 u_int mlen; 830 u_int frags; 831 832 /* 833 * Get length of mbuf fragment 834 * and how many hardware frags, 835 * rounded up, it would use: 836 */ 837 mlen = (mb->m_len - moff); 838 frags = howmany(mlen, 839 if_hw_tsomaxsegsize); 840 841 /* Handle special case: Zero Length Mbuf */ 842 if (frags == 0) 843 frags = 1; 844 845 /* 846 * Check if the fragment limit 847 * will be reached or exceeded: 848 */ 849 if (frags >= if_hw_tsomaxsegcount) { 850 max_len += min(mlen, 851 if_hw_tsomaxsegcount * 852 if_hw_tsomaxsegsize); 853 break; 854 } 855 max_len += mlen; 856 if_hw_tsomaxsegcount -= frags; 857 moff = 0; 858 mb = mb->m_next; 859 } 860 if (max_len <= 0) { 861 len = 0; 862 } else if (len > max_len) { 863 sendalot = 1; 864 len = max_len; 865 } 866 } 867 868 /* 869 * Prevent the last segment from being 870 * fractional unless the send sockbuf can be 871 * emptied: 872 */ 873 max_len = (tp->t_maxopd - optlen); 874 if ((off + len) < so->so_snd.sb_cc) { 875 moff = len % max_len; 876 if (moff != 0) { 877 len -= moff; 878 sendalot = 1; 879 } 880 } 881 882 /* 883 * In case there are too many small fragments 884 * don't use TSO: 885 */ 886 if (len <= max_len) { 887 len = max_len; 888 sendalot = 1; 889 tso = 0; 890 } 891 892 /* 893 * Send the FIN in a separate segment 894 * after the bulk sending is done. 895 * We don't trust the TSO implementations 896 * to clear the FIN flag on all but the 897 * last segment. 898 */ 899 if (tp->t_flags & TF_NEEDFIN) 900 sendalot = 1; 901 902 } else { 903 len = tp->t_maxopd - optlen - ipoptlen; 904 sendalot = 1; 905 } 906 } else 907 tso = 0; 908 909 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 910 ("%s: len > IP_MAXPACKET", __func__)); 911 912/*#ifdef DIAGNOSTIC*/ 913#ifdef INET6 914 if (max_linkhdr + hdrlen > MCLBYTES) 915#else 916 if (max_linkhdr + hdrlen > MHLEN) 917#endif 918 panic("tcphdr too big"); 919/*#endif*/ 920 921 /* 922 * This KASSERT is here to catch edge cases at a well defined place. 923 * Before, those had triggered (random) panic conditions further down. 924 */ 925 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 926 927 /* 928 * Grab a header mbuf, attaching a copy of data to 929 * be transmitted, and initialize the header from 930 * the template for sends on this connection. 931 */ 932 if (len) { 933 struct mbuf *mb; 934 u_int moff; 935 936 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 937 TCPSTAT_INC(tcps_sndprobe); 938 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 939 tp->t_sndrexmitpack++; 940 TCPSTAT_INC(tcps_sndrexmitpack); 941 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 942 } else { 943 TCPSTAT_INC(tcps_sndpack); 944 TCPSTAT_ADD(tcps_sndbyte, len); 945 } 946#ifdef INET6 947 if (MHLEN < hdrlen + max_linkhdr) 948 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 949 else 950#endif 951 m = m_gethdr(M_NOWAIT, MT_DATA); 952 953 if (m == NULL) { 954 SOCKBUF_UNLOCK(&so->so_snd); 955 error = ENOBUFS; 956 sack_rxmit = 0; 957 goto out; 958 } 959 960 m->m_data += max_linkhdr; 961 m->m_len = hdrlen; 962 963 /* 964 * Start the m_copy functions from the closest mbuf 965 * to the offset in the socket buffer chain. 966 */ 967 mb = sbsndptr(&so->so_snd, off, len, &moff); 968 969 if (len <= MHLEN - hdrlen - max_linkhdr) { 970 m_copydata(mb, moff, (int)len, 971 mtod(m, caddr_t) + hdrlen); 972 m->m_len += len; 973 } else { 974 m->m_next = m_copy(mb, moff, (int)len); 975 if (m->m_next == NULL) { 976 SOCKBUF_UNLOCK(&so->so_snd); 977 (void) m_free(m); 978 error = ENOBUFS; 979 sack_rxmit = 0; 980 goto out; 981 } 982 } 983 984 /* 985 * If we're sending everything we've got, set PUSH. 986 * (This will keep happy those implementations which only 987 * give data to the user when a buffer fills or 988 * a PUSH comes in.) 989 */ 990 if (off + len == so->so_snd.sb_cc) 991 flags |= TH_PUSH; 992 SOCKBUF_UNLOCK(&so->so_snd); 993 } else { 994 SOCKBUF_UNLOCK(&so->so_snd); 995 if (tp->t_flags & TF_ACKNOW) 996 TCPSTAT_INC(tcps_sndacks); 997 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 998 TCPSTAT_INC(tcps_sndctrl); 999 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1000 TCPSTAT_INC(tcps_sndurg); 1001 else 1002 TCPSTAT_INC(tcps_sndwinup); 1003 1004 m = m_gethdr(M_NOWAIT, MT_DATA); 1005 if (m == NULL) { 1006 error = ENOBUFS; 1007 sack_rxmit = 0; 1008 goto out; 1009 } 1010#ifdef INET6 1011 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1012 MHLEN >= hdrlen) { 1013 MH_ALIGN(m, hdrlen); 1014 } else 1015#endif 1016 m->m_data += max_linkhdr; 1017 m->m_len = hdrlen; 1018 } 1019 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1020 m->m_pkthdr.rcvif = (struct ifnet *)0; 1021#ifdef MAC 1022 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1023#endif 1024#ifdef INET6 1025 if (isipv6) { 1026 ip6 = mtod(m, struct ip6_hdr *); 1027 th = (struct tcphdr *)(ip6 + 1); 1028 tcpip_fillheaders(tp->t_inpcb, ip6, th); 1029 } else 1030#endif /* INET6 */ 1031 { 1032 ip = mtod(m, struct ip *); 1033 ipov = (struct ipovly *)ip; 1034 th = (struct tcphdr *)(ip + 1); 1035 tcpip_fillheaders(tp->t_inpcb, ip, th); 1036 } 1037 1038 /* 1039 * Fill in fields, remembering maximum advertised 1040 * window for use in delaying messages about window sizes. 1041 * If resending a FIN, be sure not to use a new sequence number. 1042 */ 1043 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1044 tp->snd_nxt == tp->snd_max) 1045 tp->snd_nxt--; 1046 /* 1047 * If we are starting a connection, send ECN setup 1048 * SYN packet. If we are on a retransmit, we may 1049 * resend those bits a number of times as per 1050 * RFC 3168. 1051 */ 1052 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 1053 if (tp->t_rxtshift >= 1) { 1054 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 1055 flags |= TH_ECE|TH_CWR; 1056 } else 1057 flags |= TH_ECE|TH_CWR; 1058 } 1059 1060 if (tp->t_state == TCPS_ESTABLISHED && 1061 (tp->t_flags & TF_ECN_PERMIT)) { 1062 /* 1063 * If the peer has ECN, mark data packets with 1064 * ECN capable transmission (ECT). 1065 * Ignore pure ack packets, retransmissions and window probes. 1066 */ 1067 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1068 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 1069#ifdef INET6 1070 if (isipv6) 1071 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1072 else 1073#endif 1074 ip->ip_tos |= IPTOS_ECN_ECT0; 1075 TCPSTAT_INC(tcps_ecn_ect0); 1076 } 1077 1078 /* 1079 * Reply with proper ECN notifications. 1080 */ 1081 if (tp->t_flags & TF_ECN_SND_CWR) { 1082 flags |= TH_CWR; 1083 tp->t_flags &= ~TF_ECN_SND_CWR; 1084 } 1085 if (tp->t_flags & TF_ECN_SND_ECE) 1086 flags |= TH_ECE; 1087 } 1088 1089 /* 1090 * If we are doing retransmissions, then snd_nxt will 1091 * not reflect the first unsent octet. For ACK only 1092 * packets, we do not want the sequence number of the 1093 * retransmitted packet, we want the sequence number 1094 * of the next unsent octet. So, if there is no data 1095 * (and no SYN or FIN), use snd_max instead of snd_nxt 1096 * when filling in ti_seq. But if we are in persist 1097 * state, snd_max might reflect one byte beyond the 1098 * right edge of the window, so use snd_nxt in that 1099 * case, since we know we aren't doing a retransmission. 1100 * (retransmit and persist are mutually exclusive...) 1101 */ 1102 if (sack_rxmit == 0) { 1103 if (len || (flags & (TH_SYN|TH_FIN)) || 1104 tcp_timer_active(tp, TT_PERSIST)) 1105 th->th_seq = htonl(tp->snd_nxt); 1106 else 1107 th->th_seq = htonl(tp->snd_max); 1108 } else { 1109 th->th_seq = htonl(p->rxmit); 1110 p->rxmit += len; 1111 tp->sackhint.sack_bytes_rexmit += len; 1112 } 1113 th->th_ack = htonl(tp->rcv_nxt); 1114 if (optlen) { 1115 bcopy(opt, th + 1, optlen); 1116 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1117 } 1118 th->th_flags = flags; 1119 /* 1120 * Calculate receive window. Don't shrink window, 1121 * but avoid silly window syndrome. 1122 */ 1123 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 1124 recwin < (long)tp->t_maxseg) 1125 recwin = 0; 1126 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1127 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 1128 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 1129 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 1130 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 1131 1132 /* 1133 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1134 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1135 * case is handled in syncache. 1136 */ 1137 if (flags & TH_SYN) 1138 th->th_win = htons((u_short) 1139 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1140 else 1141 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1142 1143 /* 1144 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1145 * a 0 window. This may cause the remote transmitter to stall. This 1146 * flag tells soreceive() to disable delayed acknowledgements when 1147 * draining the buffer. This can occur if the receiver is attempting 1148 * to read more data than can be buffered prior to transmitting on 1149 * the connection. 1150 */ 1151 if (th->th_win == 0) { 1152 tp->t_sndzerowin++; 1153 tp->t_flags |= TF_RXWIN0SENT; 1154 } else 1155 tp->t_flags &= ~TF_RXWIN0SENT; 1156 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1157 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1158 th->th_flags |= TH_URG; 1159 } else 1160 /* 1161 * If no urgent pointer to send, then we pull 1162 * the urgent pointer to the left edge of the send window 1163 * so that it doesn't drift into the send window on sequence 1164 * number wraparound. 1165 */ 1166 tp->snd_up = tp->snd_una; /* drag it along */ 1167 1168#ifdef TCP_SIGNATURE 1169 if (to.to_flags & TOF_SIGNATURE) { 1170 int sigoff = to.to_signature - opt; 1171 tcp_signature_compute(m, 0, len, optlen, 1172 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1173 } 1174#endif 1175 1176 /* 1177 * Put TCP length in extended header, and then 1178 * checksum extended header and data. 1179 */ 1180 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1181 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1182#ifdef INET6 1183 if (isipv6) { 1184 /* 1185 * ip6_plen is not need to be filled now, and will be filled 1186 * in ip6_output. 1187 */ 1188 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1189 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1190 optlen + len, IPPROTO_TCP, 0); 1191 } 1192#endif 1193#if defined(INET6) && defined(INET) 1194 else 1195#endif 1196#ifdef INET 1197 { 1198 m->m_pkthdr.csum_flags = CSUM_TCP; 1199 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1200 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1201 1202 /* IP version must be set here for ipv4/ipv6 checking later */ 1203 KASSERT(ip->ip_v == IPVERSION, 1204 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1205 } 1206#endif 1207 1208 /* 1209 * Enable TSO and specify the size of the segments. 1210 * The TCP pseudo header checksum is always provided. 1211 * XXX: Fixme: This is currently not the case for IPv6. 1212 */ 1213 if (tso) { 1214 KASSERT(len > tp->t_maxopd - optlen, 1215 ("%s: len <= tso_segsz", __func__)); 1216 m->m_pkthdr.csum_flags |= CSUM_TSO; 1217 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 1218 } 1219 1220#ifdef IPSEC 1221 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1222 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u", 1223 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1224#else 1225 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1226 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u", 1227 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1228#endif 1229 1230 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1231 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1232 1233#ifdef TCPDEBUG 1234 /* 1235 * Trace. 1236 */ 1237 if (so->so_options & SO_DEBUG) { 1238 u_short save = 0; 1239#ifdef INET6 1240 if (!isipv6) 1241#endif 1242 { 1243 save = ipov->ih_len; 1244 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1245 } 1246 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1247#ifdef INET6 1248 if (!isipv6) 1249#endif 1250 ipov->ih_len = save; 1251 } 1252#endif /* TCPDEBUG */ 1253 1254 /* 1255 * Fill in IP length and desired time to live and 1256 * send to IP level. There should be a better way 1257 * to handle ttl and tos; we could keep them in 1258 * the template, but need a way to checksum without them. 1259 */ 1260 /* 1261 * m->m_pkthdr.len should have been set before cksum calcuration, 1262 * because in6_cksum() need it. 1263 */ 1264#ifdef INET6 1265 if (isipv6) { 1266 struct route_in6 ro; 1267 1268 bzero(&ro, sizeof(ro)); 1269 /* 1270 * we separately set hoplimit for every segment, since the 1271 * user might want to change the value via setsockopt. 1272 * Also, desired default hop limit might be changed via 1273 * Neighbor Discovery. 1274 */ 1275 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1276 1277 /* 1278 * Set the packet size here for the benefit of DTrace probes. 1279 * ip6_output() will set it properly; it's supposed to include 1280 * the option header lengths as well. 1281 */ 1282 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1283 1284 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) 1285 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1286 else 1287 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1288 1289 if (tp->t_state == TCPS_SYN_SENT) 1290 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1291 1292 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1293 1294 /* TODO: IPv6 IP6TOS_ECT bit on */ 1295 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro, 1296 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1297 NULL, NULL, tp->t_inpcb); 1298 1299 if (error == EMSGSIZE && ro.ro_rt != NULL) 1300 mtu = ro.ro_rt->rt_mtu; 1301 RO_RTFREE(&ro); 1302 } 1303#endif /* INET6 */ 1304#if defined(INET) && defined(INET6) 1305 else 1306#endif 1307#ifdef INET 1308 { 1309 struct route ro; 1310 1311 bzero(&ro, sizeof(ro)); 1312 ip->ip_len = htons(m->m_pkthdr.len); 1313#ifdef INET6 1314 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1315 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1316#endif /* INET6 */ 1317 /* 1318 * If we do path MTU discovery, then we set DF on every packet. 1319 * This might not be the best thing to do according to RFC3390 1320 * Section 2. However the tcp hostcache migitates the problem 1321 * so it affects only the first tcp connection with a host. 1322 * 1323 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1324 */ 1325 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) { 1326 ip->ip_off |= htons(IP_DF); 1327 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1328 } else { 1329 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1330 } 1331 1332 if (tp->t_state == TCPS_SYN_SENT) 1333 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1334 1335 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1336 1337 error = ip_output(m, tp->t_inpcb->inp_options, &ro, 1338 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1339 tp->t_inpcb); 1340 1341 if (error == EMSGSIZE && ro.ro_rt != NULL) 1342 mtu = ro.ro_rt->rt_mtu; 1343 RO_RTFREE(&ro); 1344 } 1345#endif /* INET */ 1346 1347out: 1348 /* 1349 * In transmit state, time the transmission and arrange for 1350 * the retransmit. In persist state, just set snd_max. 1351 */ 1352 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1353 !tcp_timer_active(tp, TT_PERSIST)) { 1354 tcp_seq startseq = tp->snd_nxt; 1355 1356 /* 1357 * Advance snd_nxt over sequence space of this segment. 1358 */ 1359 if (flags & (TH_SYN|TH_FIN)) { 1360 if (flags & TH_SYN) 1361 tp->snd_nxt++; 1362 if (flags & TH_FIN) { 1363 tp->snd_nxt++; 1364 tp->t_flags |= TF_SENTFIN; 1365 } 1366 } 1367 if (sack_rxmit) 1368 goto timer; 1369 tp->snd_nxt += len; 1370 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1371 tp->snd_max = tp->snd_nxt; 1372 /* 1373 * Time this transmission if not a retransmission and 1374 * not currently timing anything. 1375 */ 1376 if (tp->t_rtttime == 0) { 1377 tp->t_rtttime = ticks; 1378 tp->t_rtseq = startseq; 1379 TCPSTAT_INC(tcps_segstimed); 1380 } 1381 } 1382 1383 /* 1384 * Set retransmit timer if not currently set, 1385 * and not doing a pure ack or a keep-alive probe. 1386 * Initial value for retransmit timer is smoothed 1387 * round-trip time + 2 * round-trip time variance. 1388 * Initialize shift counter which is used for backoff 1389 * of retransmit time. 1390 */ 1391timer: 1392 if (!tcp_timer_active(tp, TT_REXMT) && 1393 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1394 (tp->snd_nxt != tp->snd_una))) { 1395 if (tcp_timer_active(tp, TT_PERSIST)) { 1396 tcp_timer_activate(tp, TT_PERSIST, 0); 1397 tp->t_rxtshift = 0; 1398 } 1399 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1400 } else if (len == 0 && so->so_snd.sb_cc && 1401 !tcp_timer_active(tp, TT_REXMT) && 1402 !tcp_timer_active(tp, TT_PERSIST)) { 1403 /* 1404 * Avoid a situation where we do not set persist timer 1405 * after a zero window condition. For example: 1406 * 1) A -> B: packet with enough data to fill the window 1407 * 2) B -> A: ACK for #1 + new data (0 window 1408 * advertisement) 1409 * 3) A -> B: ACK for #2, 0 len packet 1410 * 1411 * In this case, A will not activate the persist timer, 1412 * because it chose to send a packet. Unless tcp_output 1413 * is called for some other reason (delayed ack timer, 1414 * another input packet from B, socket syscall), A will 1415 * not send zero window probes. 1416 * 1417 * So, if you send a 0-length packet, but there is data 1418 * in the socket buffer, and neither the rexmt or 1419 * persist timer is already set, then activate the 1420 * persist timer. 1421 */ 1422 tp->t_rxtshift = 0; 1423 tcp_setpersist(tp); 1424 } 1425 } else { 1426 /* 1427 * Persist case, update snd_max but since we are in 1428 * persist mode (no window) we do not update snd_nxt. 1429 */ 1430 int xlen = len; 1431 if (flags & TH_SYN) 1432 ++xlen; 1433 if (flags & TH_FIN) { 1434 ++xlen; 1435 tp->t_flags |= TF_SENTFIN; 1436 } 1437 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1438 tp->snd_max = tp->snd_nxt + len; 1439 } 1440 1441 if (error) { 1442 1443 /* 1444 * We know that the packet was lost, so back out the 1445 * sequence number advance, if any. 1446 * 1447 * If the error is EPERM the packet got blocked by the 1448 * local firewall. Normally we should terminate the 1449 * connection but the blocking may have been spurious 1450 * due to a firewall reconfiguration cycle. So we treat 1451 * it like a packet loss and let the retransmit timer and 1452 * timeouts do their work over time. 1453 * XXX: It is a POLA question whether calling tcp_drop right 1454 * away would be the really correct behavior instead. 1455 */ 1456 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1457 !tcp_timer_active(tp, TT_PERSIST)) && 1458 ((flags & TH_SYN) == 0) && 1459 (error != EPERM)) { 1460 if (sack_rxmit) { 1461 p->rxmit -= len; 1462 tp->sackhint.sack_bytes_rexmit -= len; 1463 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1464 ("sackhint bytes rtx >= 0")); 1465 } else 1466 tp->snd_nxt -= len; 1467 } 1468 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1469 switch (error) { 1470 case EPERM: 1471 tp->t_softerror = error; 1472 return (error); 1473 case ENOBUFS: 1474 if (!tcp_timer_active(tp, TT_REXMT) && 1475 !tcp_timer_active(tp, TT_PERSIST)) 1476 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1477 tp->snd_cwnd = tp->t_maxseg; 1478 return (0); 1479 case EMSGSIZE: 1480 /* 1481 * For some reason the interface we used initially 1482 * to send segments changed to another or lowered 1483 * its MTU. 1484 * If TSO was active we either got an interface 1485 * without TSO capabilits or TSO was turned off. 1486 * If we obtained mtu from ip_output() then update 1487 * it and try again. 1488 */ 1489 if (tso) 1490 tp->t_flags &= ~TF_TSO; 1491 if (mtu != 0) { 1492 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1493 goto again; 1494 } 1495 return (error); 1496 case EHOSTDOWN: 1497 case EHOSTUNREACH: 1498 case ENETDOWN: 1499 case ENETUNREACH: 1500 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1501 tp->t_softerror = error; 1502 return (0); 1503 } 1504 /* FALLTHROUGH */ 1505 default: 1506 return (error); 1507 } 1508 } 1509 TCPSTAT_INC(tcps_sndtotal); 1510 1511 /* 1512 * Data sent (as far as we can tell). 1513 * If this advertises a larger window than any other segment, 1514 * then remember the size of the advertised window. 1515 * Any pending ACK has now been sent. 1516 */ 1517 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1518 tp->rcv_adv = tp->rcv_nxt + recwin; 1519 tp->last_ack_sent = tp->rcv_nxt; 1520 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1521 if (tcp_timer_active(tp, TT_DELACK)) 1522 tcp_timer_activate(tp, TT_DELACK, 0); 1523#if 0 1524 /* 1525 * This completely breaks TCP if newreno is turned on. What happens 1526 * is that if delayed-acks are turned on on the receiver, this code 1527 * on the transmitter effectively destroys the TCP window, forcing 1528 * it to four packets (1.5Kx4 = 6K window). 1529 */ 1530 if (sendalot && --maxburst) 1531 goto again; 1532#endif 1533 if (sendalot) 1534 goto again; 1535 return (0); 1536} 1537 1538void 1539tcp_setpersist(struct tcpcb *tp) 1540{ 1541 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1542 int tt; 1543 1544 tp->t_flags &= ~TF_PREVVALID; 1545 if (tcp_timer_active(tp, TT_REXMT)) 1546 panic("tcp_setpersist: retransmit pending"); 1547 /* 1548 * Start/restart persistance timer. 1549 */ 1550 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1551 TCPTV_PERSMIN, TCPTV_PERSMAX); 1552 tcp_timer_activate(tp, TT_PERSIST, tt); 1553 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1554 tp->t_rxtshift++; 1555} 1556 1557/* 1558 * Insert TCP options according to the supplied parameters to the place 1559 * optp in a consistent way. Can handle unaligned destinations. 1560 * 1561 * The order of the option processing is crucial for optimal packing and 1562 * alignment for the scarce option space. 1563 * 1564 * The optimal order for a SYN/SYN-ACK segment is: 1565 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1566 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1567 * 1568 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1569 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1570 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1571 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1572 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1573 */ 1574int 1575tcp_addoptions(struct tcpopt *to, u_char *optp) 1576{ 1577 u_int mask, optlen = 0; 1578 1579 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1580 if ((to->to_flags & mask) != mask) 1581 continue; 1582 if (optlen == TCP_MAXOLEN) 1583 break; 1584 switch (to->to_flags & mask) { 1585 case TOF_MSS: 1586 while (optlen % 4) { 1587 optlen += TCPOLEN_NOP; 1588 *optp++ = TCPOPT_NOP; 1589 } 1590 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1591 continue; 1592 optlen += TCPOLEN_MAXSEG; 1593 *optp++ = TCPOPT_MAXSEG; 1594 *optp++ = TCPOLEN_MAXSEG; 1595 to->to_mss = htons(to->to_mss); 1596 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1597 optp += sizeof(to->to_mss); 1598 break; 1599 case TOF_SCALE: 1600 while (!optlen || optlen % 2 != 1) { 1601 optlen += TCPOLEN_NOP; 1602 *optp++ = TCPOPT_NOP; 1603 } 1604 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1605 continue; 1606 optlen += TCPOLEN_WINDOW; 1607 *optp++ = TCPOPT_WINDOW; 1608 *optp++ = TCPOLEN_WINDOW; 1609 *optp++ = to->to_wscale; 1610 break; 1611 case TOF_SACKPERM: 1612 while (optlen % 2) { 1613 optlen += TCPOLEN_NOP; 1614 *optp++ = TCPOPT_NOP; 1615 } 1616 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1617 continue; 1618 optlen += TCPOLEN_SACK_PERMITTED; 1619 *optp++ = TCPOPT_SACK_PERMITTED; 1620 *optp++ = TCPOLEN_SACK_PERMITTED; 1621 break; 1622 case TOF_TS: 1623 while (!optlen || optlen % 4 != 2) { 1624 optlen += TCPOLEN_NOP; 1625 *optp++ = TCPOPT_NOP; 1626 } 1627 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1628 continue; 1629 optlen += TCPOLEN_TIMESTAMP; 1630 *optp++ = TCPOPT_TIMESTAMP; 1631 *optp++ = TCPOLEN_TIMESTAMP; 1632 to->to_tsval = htonl(to->to_tsval); 1633 to->to_tsecr = htonl(to->to_tsecr); 1634 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1635 optp += sizeof(to->to_tsval); 1636 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1637 optp += sizeof(to->to_tsecr); 1638 break; 1639#ifdef TCP_SIGNATURE 1640 case TOF_SIGNATURE: 1641 { 1642 int siglen = TCPOLEN_SIGNATURE - 2; 1643 1644 while (!optlen || optlen % 4 != 2) { 1645 optlen += TCPOLEN_NOP; 1646 *optp++ = TCPOPT_NOP; 1647 } 1648 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1649 continue; 1650 optlen += TCPOLEN_SIGNATURE; 1651 *optp++ = TCPOPT_SIGNATURE; 1652 *optp++ = TCPOLEN_SIGNATURE; 1653 to->to_signature = optp; 1654 while (siglen--) 1655 *optp++ = 0; 1656 break; 1657 } 1658#endif 1659 case TOF_SACK: 1660 { 1661 int sackblks = 0; 1662 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1663 tcp_seq sack_seq; 1664 1665 while (!optlen || optlen % 4 != 2) { 1666 optlen += TCPOLEN_NOP; 1667 *optp++ = TCPOPT_NOP; 1668 } 1669 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1670 continue; 1671 optlen += TCPOLEN_SACKHDR; 1672 *optp++ = TCPOPT_SACK; 1673 sackblks = min(to->to_nsacks, 1674 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1675 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1676 while (sackblks--) { 1677 sack_seq = htonl(sack->start); 1678 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1679 optp += sizeof(sack_seq); 1680 sack_seq = htonl(sack->end); 1681 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1682 optp += sizeof(sack_seq); 1683 optlen += TCPOLEN_SACK; 1684 sack++; 1685 } 1686 TCPSTAT_INC(tcps_sack_send_blocks); 1687 break; 1688 } 1689 default: 1690 panic("%s: unknown TCP option type", __func__); 1691 break; 1692 } 1693 } 1694 1695 /* Terminate and pad TCP options to a 4 byte boundary. */ 1696 if (optlen % 4) { 1697 optlen += TCPOLEN_EOL; 1698 *optp++ = TCPOPT_EOL; 1699 } 1700 /* 1701 * According to RFC 793 (STD0007): 1702 * "The content of the header beyond the End-of-Option option 1703 * must be header padding (i.e., zero)." 1704 * and later: "The padding is composed of zeros." 1705 */ 1706 while (optlen % 4) { 1707 optlen += TCPOLEN_PAD; 1708 *optp++ = TCPOPT_PAD; 1709 } 1710 1711 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1712 return (optlen); 1713} 1714