tcp_output.c revision 293896
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/9.3/sys/netinet/tcp_output.c 293896 2016-01-14 09:11:26Z glebius $"); 34 35#include "opt_inet.h" 36#include "opt_inet6.h" 37#include "opt_ipsec.h" 38#include "opt_tcpdebug.h" 39 40#include <sys/param.h> 41#include <sys/systm.h> 42#include <sys/domain.h> 43#include <sys/hhook.h> 44#include <sys/kernel.h> 45#include <sys/lock.h> 46#include <sys/mbuf.h> 47#include <sys/mutex.h> 48#include <sys/protosw.h> 49#include <sys/socket.h> 50#include <sys/socketvar.h> 51#include <sys/sysctl.h> 52 53#include <net/if.h> 54#include <net/route.h> 55#include <net/vnet.h> 56 57#include <netinet/cc.h> 58#include <netinet/in.h> 59#include <netinet/in_systm.h> 60#include <netinet/ip.h> 61#include <netinet/in_pcb.h> 62#include <netinet/ip_var.h> 63#include <netinet/ip_options.h> 64#ifdef INET6 65#include <netinet6/in6_pcb.h> 66#include <netinet/ip6.h> 67#include <netinet6/ip6_var.h> 68#endif 69#define TCPOUTFLAGS 70#include <netinet/tcp_fsm.h> 71#include <netinet/tcp_seq.h> 72#include <netinet/tcp_timer.h> 73#include <netinet/tcp_var.h> 74#include <netinet/tcpip.h> 75#ifdef TCPDEBUG 76#include <netinet/tcp_debug.h> 77#endif 78#ifdef TCP_OFFLOAD 79#include <netinet/tcp_offload.h> 80#endif 81 82#ifdef IPSEC 83#include <netipsec/ipsec.h> 84#endif /*IPSEC*/ 85 86#include <machine/in_cksum.h> 87 88#include <security/mac/mac_framework.h> 89 90VNET_DEFINE(int, path_mtu_discovery) = 1; 91SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 92 &VNET_NAME(path_mtu_discovery), 1, 93 "Enable Path MTU Discovery"); 94 95VNET_DEFINE(int, tcp_do_tso) = 1; 96#define V_tcp_do_tso VNET(tcp_do_tso) 97SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW, 98 &VNET_NAME(tcp_do_tso), 0, 99 "Enable TCP Segmentation Offload"); 100 101VNET_DEFINE(int, tcp_sendspace) = 1024*32; 102#define V_tcp_sendspace VNET(tcp_sendspace) 103SYSCTL_VNET_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_RW, 104 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 105 106VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 107#define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) 108SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 109 &VNET_NAME(tcp_do_autosndbuf), 0, 110 "Enable automatic send buffer sizing"); 111 112VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 113#define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) 114SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 115 &VNET_NAME(tcp_autosndbuf_inc), 0, 116 "Incrementor step size of automatic send buffer"); 117 118VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; 119#define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) 120SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 121 &VNET_NAME(tcp_autosndbuf_max), 0, 122 "Max size of automatic send buffer"); 123 124static void inline hhook_run_tcp_est_out(struct tcpcb *tp, 125 struct tcphdr *th, struct tcpopt *to, 126 long len, int tso); 127static void inline cc_after_idle(struct tcpcb *tp); 128 129/* 130 * Wrapper for the TCP established ouput helper hook. 131 */ 132static void inline 133hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 134 struct tcpopt *to, long len, int tso) 135{ 136 struct tcp_hhook_data hhook_data; 137 138 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 139 hhook_data.tp = tp; 140 hhook_data.th = th; 141 hhook_data.to = to; 142 hhook_data.len = len; 143 hhook_data.tso = tso; 144 145 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 146 tp->osd); 147 } 148} 149 150/* 151 * CC wrapper hook functions 152 */ 153static void inline 154cc_after_idle(struct tcpcb *tp) 155{ 156 INP_WLOCK_ASSERT(tp->t_inpcb); 157 158 if (CC_ALGO(tp)->after_idle != NULL) 159 CC_ALGO(tp)->after_idle(tp->ccv); 160} 161 162/* 163 * Tcp output routine: figure out what should be sent and send it. 164 */ 165int 166tcp_output(struct tcpcb *tp) 167{ 168 struct socket *so = tp->t_inpcb->inp_socket; 169 long len, recwin, sendwin; 170 int off, flags, error = 0; /* Keep compiler happy */ 171 struct mbuf *m; 172 struct ip *ip = NULL; 173 struct ipovly *ipov = NULL; 174 struct tcphdr *th; 175 u_char opt[TCP_MAXOLEN]; 176 unsigned ipoptlen, optlen, hdrlen; 177#ifdef IPSEC 178 unsigned ipsec_optlen = 0; 179#endif 180 int idle, sendalot; 181 int sack_rxmit, sack_bytes_rxmt; 182 struct sackhole *p; 183 int tso, mtu; 184 struct tcpopt to; 185#if 0 186 int maxburst = TCP_MAXBURST; 187#endif 188#ifdef INET6 189 struct ip6_hdr *ip6 = NULL; 190 int isipv6; 191 192 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 193#endif 194 195 INP_WLOCK_ASSERT(tp->t_inpcb); 196 197#ifdef TCP_OFFLOAD 198 if (tp->t_flags & TF_TOE) 199 return (tcp_offload_output(tp)); 200#endif 201 202 /* 203 * Determine length of data that should be transmitted, 204 * and flags that will be used. 205 * If there is some data or critical controls (SYN, RST) 206 * to send, then transmit; otherwise, investigate further. 207 */ 208 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 209 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) 210 cc_after_idle(tp); 211 tp->t_flags &= ~TF_LASTIDLE; 212 if (idle) { 213 if (tp->t_flags & TF_MORETOCOME) { 214 tp->t_flags |= TF_LASTIDLE; 215 idle = 0; 216 } 217 } 218again: 219 /* 220 * If we've recently taken a timeout, snd_max will be greater than 221 * snd_nxt. There may be SACK information that allows us to avoid 222 * resending already delivered data. Adjust snd_nxt accordingly. 223 */ 224 if ((tp->t_flags & TF_SACK_PERMIT) && 225 SEQ_LT(tp->snd_nxt, tp->snd_max)) 226 tcp_sack_adjust(tp); 227 sendalot = 0; 228 tso = 0; 229 mtu = 0; 230 off = tp->snd_nxt - tp->snd_una; 231 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 232 233 flags = tcp_outflags[tp->t_state]; 234 /* 235 * Send any SACK-generated retransmissions. If we're explicitly trying 236 * to send out new data (when sendalot is 1), bypass this function. 237 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 238 * we're replacing a (future) new transmission with a retransmission 239 * now, and we previously incremented snd_cwnd in tcp_input(). 240 */ 241 /* 242 * Still in sack recovery , reset rxmit flag to zero. 243 */ 244 sack_rxmit = 0; 245 sack_bytes_rxmt = 0; 246 len = 0; 247 p = NULL; 248 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 249 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 250 long cwin; 251 252 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 253 if (cwin < 0) 254 cwin = 0; 255 /* Do not retransmit SACK segments beyond snd_recover */ 256 if (SEQ_GT(p->end, tp->snd_recover)) { 257 /* 258 * (At least) part of sack hole extends beyond 259 * snd_recover. Check to see if we can rexmit data 260 * for this hole. 261 */ 262 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 263 /* 264 * Can't rexmit any more data for this hole. 265 * That data will be rexmitted in the next 266 * sack recovery episode, when snd_recover 267 * moves past p->rxmit. 268 */ 269 p = NULL; 270 goto after_sack_rexmit; 271 } else 272 /* Can rexmit part of the current hole */ 273 len = ((long)ulmin(cwin, 274 tp->snd_recover - p->rxmit)); 275 } else 276 len = ((long)ulmin(cwin, p->end - p->rxmit)); 277 off = p->rxmit - tp->snd_una; 278 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 279 __func__, off)); 280 if (len > 0) { 281 sack_rxmit = 1; 282 sendalot = 1; 283 TCPSTAT_INC(tcps_sack_rexmits); 284 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 285 min(len, tp->t_maxseg)); 286 } 287 } 288after_sack_rexmit: 289 /* 290 * Get standard flags, and add SYN or FIN if requested by 'hidden' 291 * state flags. 292 */ 293 if (tp->t_flags & TF_NEEDFIN) 294 flags |= TH_FIN; 295 if (tp->t_flags & TF_NEEDSYN) 296 flags |= TH_SYN; 297 298 SOCKBUF_LOCK(&so->so_snd); 299 /* 300 * If in persist timeout with window of 0, send 1 byte. 301 * Otherwise, if window is small but nonzero 302 * and timer expired, we will send what we can 303 * and go to transmit state. 304 */ 305 if (tp->t_flags & TF_FORCEDATA) { 306 if (sendwin == 0) { 307 /* 308 * If we still have some data to send, then 309 * clear the FIN bit. Usually this would 310 * happen below when it realizes that we 311 * aren't sending all the data. However, 312 * if we have exactly 1 byte of unsent data, 313 * then it won't clear the FIN bit below, 314 * and if we are in persist state, we wind 315 * up sending the packet without recording 316 * that we sent the FIN bit. 317 * 318 * We can't just blindly clear the FIN bit, 319 * because if we don't have any more data 320 * to send then the probe will be the FIN 321 * itself. 322 */ 323 if (off < so->so_snd.sb_cc) 324 flags &= ~TH_FIN; 325 sendwin = 1; 326 } else { 327 tcp_timer_activate(tp, TT_PERSIST, 0); 328 tp->t_rxtshift = 0; 329 } 330 } 331 332 /* 333 * If snd_nxt == snd_max and we have transmitted a FIN, the 334 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 335 * a negative length. This can also occur when TCP opens up 336 * its congestion window while receiving additional duplicate 337 * acks after fast-retransmit because TCP will reset snd_nxt 338 * to snd_max after the fast-retransmit. 339 * 340 * In the normal retransmit-FIN-only case, however, snd_nxt will 341 * be set to snd_una, the offset will be 0, and the length may 342 * wind up 0. 343 * 344 * If sack_rxmit is true we are retransmitting from the scoreboard 345 * in which case len is already set. 346 */ 347 if (sack_rxmit == 0) { 348 if (sack_bytes_rxmt == 0) 349 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off); 350 else { 351 long cwin; 352 353 /* 354 * We are inside of a SACK recovery episode and are 355 * sending new data, having retransmitted all the 356 * data possible in the scoreboard. 357 */ 358 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 359 - off); 360 /* 361 * Don't remove this (len > 0) check ! 362 * We explicitly check for len > 0 here (although it 363 * isn't really necessary), to work around a gcc 364 * optimization issue - to force gcc to compute 365 * len above. Without this check, the computation 366 * of len is bungled by the optimizer. 367 */ 368 if (len > 0) { 369 cwin = tp->snd_cwnd - 370 (tp->snd_nxt - tp->sack_newdata) - 371 sack_bytes_rxmt; 372 if (cwin < 0) 373 cwin = 0; 374 len = lmin(len, cwin); 375 } 376 } 377 } 378 379 /* 380 * Lop off SYN bit if it has already been sent. However, if this 381 * is SYN-SENT state and if segment contains data and if we don't 382 * know that foreign host supports TAO, suppress sending segment. 383 */ 384 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 385 if (tp->t_state != TCPS_SYN_RECEIVED) 386 flags &= ~TH_SYN; 387 off--, len++; 388 } 389 390 /* 391 * Be careful not to send data and/or FIN on SYN segments. 392 * This measure is needed to prevent interoperability problems 393 * with not fully conformant TCP implementations. 394 */ 395 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 396 len = 0; 397 flags &= ~TH_FIN; 398 } 399 400 if (len <= 0) { 401 /* 402 * If FIN has been sent but not acked, 403 * but we haven't been called to retransmit, 404 * len will be < 0. Otherwise, window shrank 405 * after we sent into it. If window shrank to 0, 406 * cancel pending retransmit, pull snd_nxt back 407 * to (closed) window, and set the persist timer 408 * if it isn't already going. If the window didn't 409 * close completely, just wait for an ACK. 410 * 411 * We also do a general check here to ensure that 412 * we will set the persist timer when we have data 413 * to send, but a 0-byte window. This makes sure 414 * the persist timer is set even if the packet 415 * hits one of the "goto send" lines below. 416 */ 417 len = 0; 418 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 419 (off < (int) so->so_snd.sb_cc)) { 420 tcp_timer_activate(tp, TT_REXMT, 0); 421 tp->t_rxtshift = 0; 422 tp->snd_nxt = tp->snd_una; 423 if (!tcp_timer_active(tp, TT_PERSIST)) 424 tcp_setpersist(tp); 425 } 426 } 427 428 /* len will be >= 0 after this point. */ 429 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 430 431 /* 432 * Automatic sizing of send socket buffer. Often the send buffer 433 * size is not optimally adjusted to the actual network conditions 434 * at hand (delay bandwidth product). Setting the buffer size too 435 * small limits throughput on links with high bandwidth and high 436 * delay (eg. trans-continental/oceanic links). Setting the 437 * buffer size too big consumes too much real kernel memory, 438 * especially with many connections on busy servers. 439 * 440 * The criteria to step up the send buffer one notch are: 441 * 1. receive window of remote host is larger than send buffer 442 * (with a fudge factor of 5/4th); 443 * 2. send buffer is filled to 7/8th with data (so we actually 444 * have data to make use of it); 445 * 3. send buffer fill has not hit maximal automatic size; 446 * 4. our send window (slow start and cogestion controlled) is 447 * larger than sent but unacknowledged data in send buffer. 448 * 449 * The remote host receive window scaling factor may limit the 450 * growing of the send buffer before it reaches its allowed 451 * maximum. 452 * 453 * It scales directly with slow start or congestion window 454 * and does at most one step per received ACK. This fast 455 * scaling has the drawback of growing the send buffer beyond 456 * what is strictly necessary to make full use of a given 457 * delay*bandwith product. However testing has shown this not 458 * to be much of an problem. At worst we are trading wasting 459 * of available bandwith (the non-use of it) for wasting some 460 * socket buffer memory. 461 * 462 * TODO: Shrink send buffer during idle periods together 463 * with congestion window. Requires another timer. Has to 464 * wait for upcoming tcp timer rewrite. 465 */ 466 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 467 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 468 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 469 so->so_snd.sb_cc < V_tcp_autosndbuf_max && 470 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 471 if (!sbreserve_locked(&so->so_snd, 472 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 473 V_tcp_autosndbuf_max), so, curthread)) 474 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 475 } 476 } 477 478 /* 479 * Decide if we can use TCP Segmentation Offloading (if supported by 480 * hardware). 481 * 482 * TSO may only be used if we are in a pure bulk sending state. The 483 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 484 * IP options prevent using TSO. With TSO the TCP header is the same 485 * (except for the sequence number) for all generated packets. This 486 * makes it impossible to transmit any options which vary per generated 487 * segment or packet. 488 */ 489#ifdef IPSEC 490 /* 491 * Pre-calculate here as we save another lookup into the darknesses 492 * of IPsec that way and can actually decide if TSO is ok. 493 */ 494 ipsec_optlen = ipsec_hdrsiz_tcp(tp); 495#endif 496 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 497 ((tp->t_flags & TF_SIGNATURE) == 0) && 498 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 499#ifdef IPSEC 500 ipsec_optlen == 0 && 501#endif 502 tp->t_inpcb->inp_options == NULL && 503 tp->t_inpcb->in6p_options == NULL) 504 tso = 1; 505 506 if (sack_rxmit) { 507 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 508 flags &= ~TH_FIN; 509 } else { 510 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 511 flags &= ~TH_FIN; 512 } 513 514 recwin = sbspace(&so->so_rcv); 515 516 /* 517 * Sender silly window avoidance. We transmit under the following 518 * conditions when len is non-zero: 519 * 520 * - We have a full segment (or more with TSO) 521 * - This is the last buffer in a write()/send() and we are 522 * either idle or running NODELAY 523 * - we've timed out (e.g. persist timer) 524 * - we have more then 1/2 the maximum send window's worth of 525 * data (receiver may be limited the window size) 526 * - we need to retransmit 527 */ 528 if (len) { 529 if (len >= tp->t_maxseg) 530 goto send; 531 /* 532 * NOTE! on localhost connections an 'ack' from the remote 533 * end may occur synchronously with the output and cause 534 * us to flush a buffer queued with moretocome. XXX 535 * 536 * note: the len + off check is almost certainly unnecessary. 537 */ 538 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 539 (idle || (tp->t_flags & TF_NODELAY)) && 540 len + off >= so->so_snd.sb_cc && 541 (tp->t_flags & TF_NOPUSH) == 0) { 542 goto send; 543 } 544 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 545 goto send; 546 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 547 goto send; 548 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 549 goto send; 550 if (sack_rxmit) 551 goto send; 552 } 553 554 /* 555 * Sending of standalone window updates. 556 * 557 * Window updates are important when we close our window due to a 558 * full socket buffer and are opening it again after the application 559 * reads data from it. Once the window has opened again and the 560 * remote end starts to send again the ACK clock takes over and 561 * provides the most current window information. 562 * 563 * We must avoid the silly window syndrome whereas every read 564 * from the receive buffer, no matter how small, causes a window 565 * update to be sent. We also should avoid sending a flurry of 566 * window updates when the socket buffer had queued a lot of data 567 * and the application is doing small reads. 568 * 569 * Prevent a flurry of pointless window updates by only sending 570 * an update when we can increase the advertized window by more 571 * than 1/4th of the socket buffer capacity. When the buffer is 572 * getting full or is very small be more aggressive and send an 573 * update whenever we can increase by two mss sized segments. 574 * In all other situations the ACK's to new incoming data will 575 * carry further window increases. 576 * 577 * Don't send an independent window update if a delayed 578 * ACK is pending (it will get piggy-backed on it) or the 579 * remote side already has done a half-close and won't send 580 * more data. Skip this if the connection is in T/TCP 581 * half-open state. 582 */ 583 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 584 !(tp->t_flags & TF_DELACK) && 585 !TCPS_HAVERCVDFIN(tp->t_state)) { 586 /* 587 * "adv" is the amount we could increase the window, 588 * taking into account that we are limited by 589 * TCP_MAXWIN << tp->rcv_scale. 590 */ 591 long adv; 592 int oldwin; 593 594 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 595 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 596 oldwin = (tp->rcv_adv - tp->rcv_nxt); 597 adv -= oldwin; 598 } else 599 oldwin = 0; 600 601 /* 602 * If the new window size ends up being the same as the old 603 * size when it is scaled, then don't force a window update. 604 */ 605 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 606 goto dontupdate; 607 608 if (adv >= (long)(2 * tp->t_maxseg) && 609 (adv >= (long)(so->so_rcv.sb_hiwat / 4) || 610 recwin <= (long)(so->so_rcv.sb_hiwat / 8) || 611 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 612 goto send; 613 } 614dontupdate: 615 616 /* 617 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 618 * is also a catch-all for the retransmit timer timeout case. 619 */ 620 if (tp->t_flags & TF_ACKNOW) 621 goto send; 622 if ((flags & TH_RST) || 623 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 624 goto send; 625 if (SEQ_GT(tp->snd_up, tp->snd_una)) 626 goto send; 627 /* 628 * If our state indicates that FIN should be sent 629 * and we have not yet done so, then we need to send. 630 */ 631 if (flags & TH_FIN && 632 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 633 goto send; 634 /* 635 * In SACK, it is possible for tcp_output to fail to send a segment 636 * after the retransmission timer has been turned off. Make sure 637 * that the retransmission timer is set. 638 */ 639 if ((tp->t_flags & TF_SACK_PERMIT) && 640 SEQ_GT(tp->snd_max, tp->snd_una) && 641 !tcp_timer_active(tp, TT_REXMT) && 642 !tcp_timer_active(tp, TT_PERSIST)) { 643 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 644 goto just_return; 645 } 646 /* 647 * TCP window updates are not reliable, rather a polling protocol 648 * using ``persist'' packets is used to insure receipt of window 649 * updates. The three ``states'' for the output side are: 650 * idle not doing retransmits or persists 651 * persisting to move a small or zero window 652 * (re)transmitting and thereby not persisting 653 * 654 * tcp_timer_active(tp, TT_PERSIST) 655 * is true when we are in persist state. 656 * (tp->t_flags & TF_FORCEDATA) 657 * is set when we are called to send a persist packet. 658 * tcp_timer_active(tp, TT_REXMT) 659 * is set when we are retransmitting 660 * The output side is idle when both timers are zero. 661 * 662 * If send window is too small, there is data to transmit, and no 663 * retransmit or persist is pending, then go to persist state. 664 * If nothing happens soon, send when timer expires: 665 * if window is nonzero, transmit what we can, 666 * otherwise force out a byte. 667 */ 668 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) && 669 !tcp_timer_active(tp, TT_PERSIST)) { 670 tp->t_rxtshift = 0; 671 tcp_setpersist(tp); 672 } 673 674 /* 675 * No reason to send a segment, just return. 676 */ 677just_return: 678 SOCKBUF_UNLOCK(&so->so_snd); 679 return (0); 680 681send: 682 SOCKBUF_LOCK_ASSERT(&so->so_snd); 683 /* 684 * Before ESTABLISHED, force sending of initial options 685 * unless TCP set not to do any options. 686 * NOTE: we assume that the IP/TCP header plus TCP options 687 * always fit in a single mbuf, leaving room for a maximum 688 * link header, i.e. 689 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 690 */ 691 optlen = 0; 692#ifdef INET6 693 if (isipv6) 694 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 695 else 696#endif 697 hdrlen = sizeof (struct tcpiphdr); 698 699 /* 700 * Compute options for segment. 701 * We only have to care about SYN and established connection 702 * segments. Options for SYN-ACK segments are handled in TCP 703 * syncache. 704 */ 705 to.to_flags = 0; 706 if ((tp->t_flags & TF_NOOPT) == 0) { 707 /* Maximum segment size. */ 708 if (flags & TH_SYN) { 709 tp->snd_nxt = tp->iss; 710 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 711 to.to_flags |= TOF_MSS; 712 } 713 /* Window scaling. */ 714 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 715 to.to_wscale = tp->request_r_scale; 716 to.to_flags |= TOF_SCALE; 717 } 718 /* Timestamps. */ 719 if ((tp->t_flags & TF_RCVD_TSTMP) || 720 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 721 to.to_tsval = tcp_ts_getticks() + tp->ts_offset; 722 to.to_tsecr = tp->ts_recent; 723 to.to_flags |= TOF_TS; 724 /* Set receive buffer autosizing timestamp. */ 725 if (tp->rfbuf_ts == 0 && 726 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 727 tp->rfbuf_ts = tcp_ts_getticks(); 728 } 729 /* Selective ACK's. */ 730 if (tp->t_flags & TF_SACK_PERMIT) { 731 if (flags & TH_SYN) 732 to.to_flags |= TOF_SACKPERM; 733 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 734 (tp->t_flags & TF_SACK_PERMIT) && 735 tp->rcv_numsacks > 0) { 736 to.to_flags |= TOF_SACK; 737 to.to_nsacks = tp->rcv_numsacks; 738 to.to_sacks = (u_char *)tp->sackblks; 739 } 740 } 741#ifdef TCP_SIGNATURE 742 /* TCP-MD5 (RFC2385). */ 743 if (tp->t_flags & TF_SIGNATURE) 744 to.to_flags |= TOF_SIGNATURE; 745#endif /* TCP_SIGNATURE */ 746 747 /* Processing the options. */ 748 hdrlen += optlen = tcp_addoptions(&to, opt); 749 } 750 751#ifdef INET6 752 if (isipv6) 753 ipoptlen = ip6_optlen(tp->t_inpcb); 754 else 755#endif 756 if (tp->t_inpcb->inp_options) 757 ipoptlen = tp->t_inpcb->inp_options->m_len - 758 offsetof(struct ipoption, ipopt_list); 759 else 760 ipoptlen = 0; 761#ifdef IPSEC 762 ipoptlen += ipsec_optlen; 763#endif 764 765 /* 766 * Adjust data length if insertion of options will 767 * bump the packet length beyond the t_maxopd length. 768 * Clear the FIN bit because we cut off the tail of 769 * the segment. 770 */ 771 if (len + optlen + ipoptlen > tp->t_maxopd) { 772 flags &= ~TH_FIN; 773 774 if (tso) { 775 KASSERT(ipoptlen == 0, 776 ("%s: TSO can't do IP options", __func__)); 777 778 /* 779 * Limit a burst to t_tsomax minus IP, 780 * TCP and options length to keep ip->ip_len 781 * from overflowing or exceeding the maximum 782 * length allowed by the network interface. 783 */ 784 if (len > tp->t_tsomax - hdrlen) { 785 len = tp->t_tsomax - hdrlen; 786 sendalot = 1; 787 } 788 789 /* 790 * Prevent the last segment from being 791 * fractional unless the send sockbuf can 792 * be emptied. 793 */ 794 if (sendalot && off + len < so->so_snd.sb_cc) { 795 len -= len % (tp->t_maxopd - optlen); 796 sendalot = 1; 797 } 798 799 /* 800 * Send the FIN in a separate segment 801 * after the bulk sending is done. 802 * We don't trust the TSO implementations 803 * to clear the FIN flag on all but the 804 * last segment. 805 */ 806 if (tp->t_flags & TF_NEEDFIN) 807 sendalot = 1; 808 809 } else { 810 len = tp->t_maxopd - optlen - ipoptlen; 811 sendalot = 1; 812 } 813 } else 814 tso = 0; 815 816 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 817 ("%s: len > IP_MAXPACKET", __func__)); 818 819/*#ifdef DIAGNOSTIC*/ 820#ifdef INET6 821 if (max_linkhdr + hdrlen > MCLBYTES) 822#else 823 if (max_linkhdr + hdrlen > MHLEN) 824#endif 825 panic("tcphdr too big"); 826/*#endif*/ 827 828 /* 829 * This KASSERT is here to catch edge cases at a well defined place. 830 * Before, those had triggered (random) panic conditions further down. 831 */ 832 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 833 834 /* 835 * Grab a header mbuf, attaching a copy of data to 836 * be transmitted, and initialize the header from 837 * the template for sends on this connection. 838 */ 839 if (len) { 840 struct mbuf *mb; 841 u_int moff; 842 843 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 844 TCPSTAT_INC(tcps_sndprobe); 845 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 846 tp->t_sndrexmitpack++; 847 TCPSTAT_INC(tcps_sndrexmitpack); 848 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 849 } else { 850 TCPSTAT_INC(tcps_sndpack); 851 TCPSTAT_ADD(tcps_sndbyte, len); 852 } 853 MGETHDR(m, M_DONTWAIT, MT_DATA); 854 if (m == NULL) { 855 SOCKBUF_UNLOCK(&so->so_snd); 856 error = ENOBUFS; 857 goto out; 858 } 859#ifdef INET6 860 if (MHLEN < hdrlen + max_linkhdr) { 861 MCLGET(m, M_DONTWAIT); 862 if ((m->m_flags & M_EXT) == 0) { 863 SOCKBUF_UNLOCK(&so->so_snd); 864 m_freem(m); 865 error = ENOBUFS; 866 goto out; 867 } 868 } 869#endif 870 m->m_data += max_linkhdr; 871 m->m_len = hdrlen; 872 873 /* 874 * Start the m_copy functions from the closest mbuf 875 * to the offset in the socket buffer chain. 876 */ 877 mb = sbsndptr(&so->so_snd, off, len, &moff); 878 879 if (len <= MHLEN - hdrlen - max_linkhdr) { 880 m_copydata(mb, moff, (int)len, 881 mtod(m, caddr_t) + hdrlen); 882 m->m_len += len; 883 } else { 884 m->m_next = m_copy(mb, moff, (int)len); 885 if (m->m_next == NULL) { 886 SOCKBUF_UNLOCK(&so->so_snd); 887 (void) m_free(m); 888 error = ENOBUFS; 889 goto out; 890 } 891 } 892 893 /* 894 * If we're sending everything we've got, set PUSH. 895 * (This will keep happy those implementations which only 896 * give data to the user when a buffer fills or 897 * a PUSH comes in.) 898 */ 899 if (off + len == so->so_snd.sb_cc) 900 flags |= TH_PUSH; 901 SOCKBUF_UNLOCK(&so->so_snd); 902 } else { 903 SOCKBUF_UNLOCK(&so->so_snd); 904 if (tp->t_flags & TF_ACKNOW) 905 TCPSTAT_INC(tcps_sndacks); 906 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 907 TCPSTAT_INC(tcps_sndctrl); 908 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 909 TCPSTAT_INC(tcps_sndurg); 910 else 911 TCPSTAT_INC(tcps_sndwinup); 912 913 MGETHDR(m, M_DONTWAIT, MT_DATA); 914 if (m == NULL) { 915 error = ENOBUFS; 916 goto out; 917 } 918#ifdef INET6 919 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 920 MHLEN >= hdrlen) { 921 MH_ALIGN(m, hdrlen); 922 } else 923#endif 924 m->m_data += max_linkhdr; 925 m->m_len = hdrlen; 926 } 927 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 928 m->m_pkthdr.rcvif = (struct ifnet *)0; 929#ifdef MAC 930 mac_inpcb_create_mbuf(tp->t_inpcb, m); 931#endif 932#ifdef INET6 933 if (isipv6) { 934 ip6 = mtod(m, struct ip6_hdr *); 935 th = (struct tcphdr *)(ip6 + 1); 936 tcpip_fillheaders(tp->t_inpcb, ip6, th); 937 } else 938#endif /* INET6 */ 939 { 940 ip = mtod(m, struct ip *); 941 ipov = (struct ipovly *)ip; 942 th = (struct tcphdr *)(ip + 1); 943 tcpip_fillheaders(tp->t_inpcb, ip, th); 944 } 945 946 /* 947 * Fill in fields, remembering maximum advertised 948 * window for use in delaying messages about window sizes. 949 * If resending a FIN, be sure not to use a new sequence number. 950 */ 951 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 952 tp->snd_nxt == tp->snd_max) 953 tp->snd_nxt--; 954 /* 955 * If we are starting a connection, send ECN setup 956 * SYN packet. If we are on a retransmit, we may 957 * resend those bits a number of times as per 958 * RFC 3168. 959 */ 960 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 961 if (tp->t_rxtshift >= 1) { 962 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 963 flags |= TH_ECE|TH_CWR; 964 } else 965 flags |= TH_ECE|TH_CWR; 966 } 967 968 if (tp->t_state == TCPS_ESTABLISHED && 969 (tp->t_flags & TF_ECN_PERMIT)) { 970 /* 971 * If the peer has ECN, mark data packets with 972 * ECN capable transmission (ECT). 973 * Ignore pure ack packets, retransmissions and window probes. 974 */ 975 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 976 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 977#ifdef INET6 978 if (isipv6) 979 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 980 else 981#endif 982 ip->ip_tos |= IPTOS_ECN_ECT0; 983 TCPSTAT_INC(tcps_ecn_ect0); 984 } 985 986 /* 987 * Reply with proper ECN notifications. 988 */ 989 if (tp->t_flags & TF_ECN_SND_CWR) { 990 flags |= TH_CWR; 991 tp->t_flags &= ~TF_ECN_SND_CWR; 992 } 993 if (tp->t_flags & TF_ECN_SND_ECE) 994 flags |= TH_ECE; 995 } 996 997 /* 998 * If we are doing retransmissions, then snd_nxt will 999 * not reflect the first unsent octet. For ACK only 1000 * packets, we do not want the sequence number of the 1001 * retransmitted packet, we want the sequence number 1002 * of the next unsent octet. So, if there is no data 1003 * (and no SYN or FIN), use snd_max instead of snd_nxt 1004 * when filling in ti_seq. But if we are in persist 1005 * state, snd_max might reflect one byte beyond the 1006 * right edge of the window, so use snd_nxt in that 1007 * case, since we know we aren't doing a retransmission. 1008 * (retransmit and persist are mutually exclusive...) 1009 */ 1010 if (sack_rxmit == 0) { 1011 if (len || (flags & (TH_SYN|TH_FIN)) || 1012 tcp_timer_active(tp, TT_PERSIST)) 1013 th->th_seq = htonl(tp->snd_nxt); 1014 else 1015 th->th_seq = htonl(tp->snd_max); 1016 } else { 1017 th->th_seq = htonl(p->rxmit); 1018 p->rxmit += len; 1019 tp->sackhint.sack_bytes_rexmit += len; 1020 } 1021 th->th_ack = htonl(tp->rcv_nxt); 1022 if (optlen) { 1023 bcopy(opt, th + 1, optlen); 1024 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1025 } 1026 th->th_flags = flags; 1027 /* 1028 * Calculate receive window. Don't shrink window, 1029 * but avoid silly window syndrome. 1030 */ 1031 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 1032 recwin < (long)tp->t_maxseg) 1033 recwin = 0; 1034 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1035 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 1036 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 1037 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 1038 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 1039 1040 /* 1041 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1042 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1043 * case is handled in syncache. 1044 */ 1045 if (flags & TH_SYN) 1046 th->th_win = htons((u_short) 1047 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1048 else 1049 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1050 1051 /* 1052 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1053 * a 0 window. This may cause the remote transmitter to stall. This 1054 * flag tells soreceive() to disable delayed acknowledgements when 1055 * draining the buffer. This can occur if the receiver is attempting 1056 * to read more data than can be buffered prior to transmitting on 1057 * the connection. 1058 */ 1059 if (th->th_win == 0) { 1060 tp->t_sndzerowin++; 1061 tp->t_flags |= TF_RXWIN0SENT; 1062 } else 1063 tp->t_flags &= ~TF_RXWIN0SENT; 1064 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1065 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1066 th->th_flags |= TH_URG; 1067 } else 1068 /* 1069 * If no urgent pointer to send, then we pull 1070 * the urgent pointer to the left edge of the send window 1071 * so that it doesn't drift into the send window on sequence 1072 * number wraparound. 1073 */ 1074 tp->snd_up = tp->snd_una; /* drag it along */ 1075 1076#ifdef TCP_SIGNATURE 1077 if (to.to_flags & TOF_SIGNATURE) { 1078 int sigoff = to.to_signature - opt; 1079 tcp_signature_compute(m, 0, len, optlen, 1080 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1081 } 1082#endif 1083 1084 /* 1085 * Put TCP length in extended header, and then 1086 * checksum extended header and data. 1087 */ 1088 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1089 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1090#ifdef INET6 1091 if (isipv6) { 1092 /* 1093 * ip6_plen is not need to be filled now, and will be filled 1094 * in ip6_output. 1095 */ 1096 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1097 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1098 optlen + len, IPPROTO_TCP, 0); 1099 } 1100#endif 1101#if defined(INET6) && defined(INET) 1102 else 1103#endif 1104#ifdef INET 1105 { 1106 m->m_pkthdr.csum_flags = CSUM_TCP; 1107 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1108 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1109 1110 /* IP version must be set here for ipv4/ipv6 checking later */ 1111 KASSERT(ip->ip_v == IPVERSION, 1112 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1113 } 1114#endif 1115 1116 /* 1117 * Enable TSO and specify the size of the segments. 1118 * The TCP pseudo header checksum is always provided. 1119 * XXX: Fixme: This is currently not the case for IPv6. 1120 */ 1121 if (tso) { 1122 KASSERT(len > tp->t_maxopd - optlen, 1123 ("%s: len <= tso_segsz", __func__)); 1124 m->m_pkthdr.csum_flags |= CSUM_TSO; 1125 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 1126 } 1127 1128#ifdef IPSEC 1129 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1130 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u", 1131 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1132#else 1133 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1134 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u", 1135 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1136#endif 1137 1138 /* 1139 * In transmit state, time the transmission and arrange for 1140 * the retransmit. In persist state, just set snd_max. 1141 */ 1142 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1143 !tcp_timer_active(tp, TT_PERSIST)) { 1144 tcp_seq startseq = tp->snd_nxt; 1145 1146 /* 1147 * Advance snd_nxt over sequence space of this segment. 1148 */ 1149 if (flags & (TH_SYN|TH_FIN)) { 1150 if (flags & TH_SYN) 1151 tp->snd_nxt++; 1152 if (flags & TH_FIN) { 1153 tp->snd_nxt++; 1154 tp->t_flags |= TF_SENTFIN; 1155 } 1156 } 1157 if (sack_rxmit) 1158 goto timer; 1159 tp->snd_nxt += len; 1160 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1161 tp->snd_max = tp->snd_nxt; 1162 /* 1163 * Time this transmission if not a retransmission and 1164 * not currently timing anything. 1165 */ 1166 if (tp->t_rtttime == 0) { 1167 tp->t_rtttime = ticks; 1168 tp->t_rtseq = startseq; 1169 TCPSTAT_INC(tcps_segstimed); 1170 } 1171 } 1172 1173 /* 1174 * Set retransmit timer if not currently set, 1175 * and not doing a pure ack or a keep-alive probe. 1176 * Initial value for retransmit timer is smoothed 1177 * round-trip time + 2 * round-trip time variance. 1178 * Initialize shift counter which is used for backoff 1179 * of retransmit time. 1180 */ 1181timer: 1182 if (!tcp_timer_active(tp, TT_REXMT) && 1183 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1184 (tp->snd_nxt != tp->snd_una))) { 1185 if (tcp_timer_active(tp, TT_PERSIST)) { 1186 tcp_timer_activate(tp, TT_PERSIST, 0); 1187 tp->t_rxtshift = 0; 1188 } 1189 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1190 } 1191 } else { 1192 /* 1193 * Persist case, update snd_max but since we are in 1194 * persist mode (no window) we do not update snd_nxt. 1195 */ 1196 int xlen = len; 1197 if (flags & TH_SYN) 1198 ++xlen; 1199 if (flags & TH_FIN) { 1200 ++xlen; 1201 tp->t_flags |= TF_SENTFIN; 1202 } 1203 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1204 tp->snd_max = tp->snd_nxt + len; 1205 } 1206 1207 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1208 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1209 1210#ifdef TCPDEBUG 1211 /* 1212 * Trace. 1213 */ 1214 if (so->so_options & SO_DEBUG) { 1215 u_short save = 0; 1216#ifdef INET6 1217 if (!isipv6) 1218#endif 1219 { 1220 save = ipov->ih_len; 1221 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1222 } 1223 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1224#ifdef INET6 1225 if (!isipv6) 1226#endif 1227 ipov->ih_len = save; 1228 } 1229#endif /* TCPDEBUG */ 1230 1231 /* 1232 * Fill in IP length and desired time to live and 1233 * send to IP level. There should be a better way 1234 * to handle ttl and tos; we could keep them in 1235 * the template, but need a way to checksum without them. 1236 */ 1237 /* 1238 * m->m_pkthdr.len should have been set before cksum calcuration, 1239 * because in6_cksum() need it. 1240 */ 1241#ifdef INET6 1242 if (isipv6) { 1243 struct route_in6 ro; 1244 1245 bzero(&ro, sizeof(ro)); 1246 /* 1247 * we separately set hoplimit for every segment, since the 1248 * user might want to change the value via setsockopt. 1249 * Also, desired default hop limit might be changed via 1250 * Neighbor Discovery. 1251 */ 1252 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1253 1254 /* TODO: IPv6 IP6TOS_ECT bit on */ 1255 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro, 1256 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1257 NULL, NULL, tp->t_inpcb); 1258 1259 if (error == EMSGSIZE && ro.ro_rt != NULL) 1260 mtu = ro.ro_rt->rt_rmx.rmx_mtu; 1261 RO_RTFREE(&ro); 1262 } 1263#endif /* INET6 */ 1264#if defined(INET) && defined(INET6) 1265 else 1266#endif 1267#ifdef INET 1268 { 1269 struct route ro; 1270 1271 bzero(&ro, sizeof(ro)); 1272 ip->ip_len = m->m_pkthdr.len; 1273#ifdef INET6 1274 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1275 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1276#endif /* INET6 */ 1277 /* 1278 * If we do path MTU discovery, then we set DF on every packet. 1279 * This might not be the best thing to do according to RFC3390 1280 * Section 2. However the tcp hostcache migitates the problem 1281 * so it affects only the first tcp connection with a host. 1282 * 1283 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1284 */ 1285 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) 1286 ip->ip_off |= IP_DF; 1287 1288 error = ip_output(m, tp->t_inpcb->inp_options, &ro, 1289 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1290 tp->t_inpcb); 1291 1292 if (error == EMSGSIZE && ro.ro_rt != NULL) 1293 mtu = ro.ro_rt->rt_rmx.rmx_mtu; 1294 RO_RTFREE(&ro); 1295 } 1296#endif /* INET */ 1297 if (error) { 1298 1299 /* 1300 * We know that the packet was lost, so back out the 1301 * sequence number advance, if any. 1302 * 1303 * If the error is EPERM the packet got blocked by the 1304 * local firewall. Normally we should terminate the 1305 * connection but the blocking may have been spurious 1306 * due to a firewall reconfiguration cycle. So we treat 1307 * it like a packet loss and let the retransmit timer and 1308 * timeouts do their work over time. 1309 * XXX: It is a POLA question whether calling tcp_drop right 1310 * away would be the really correct behavior instead. 1311 */ 1312 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1313 !tcp_timer_active(tp, TT_PERSIST)) && 1314 ((flags & TH_SYN) == 0) && 1315 (error != EPERM)) { 1316 if (sack_rxmit) { 1317 p->rxmit -= len; 1318 tp->sackhint.sack_bytes_rexmit -= len; 1319 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1320 ("sackhint bytes rtx >= 0")); 1321 } else 1322 tp->snd_nxt -= len; 1323 } 1324out: 1325 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1326 switch (error) { 1327 case EPERM: 1328 tp->t_softerror = error; 1329 return (error); 1330 case ENOBUFS: 1331 if (!tcp_timer_active(tp, TT_REXMT) && 1332 !tcp_timer_active(tp, TT_PERSIST)) 1333 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1334 tp->snd_cwnd = tp->t_maxseg; 1335 return (0); 1336 case EMSGSIZE: 1337 /* 1338 * For some reason the interface we used initially 1339 * to send segments changed to another or lowered 1340 * its MTU. 1341 * If TSO was active we either got an interface 1342 * without TSO capabilits or TSO was turned off. 1343 * If we obtained mtu from ip_output() then update 1344 * it and try again. 1345 */ 1346 if (tso) 1347 tp->t_flags &= ~TF_TSO; 1348 if (mtu != 0) { 1349 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1350 goto again; 1351 } 1352 return (error); 1353 case EHOSTDOWN: 1354 case EHOSTUNREACH: 1355 case ENETDOWN: 1356 case ENETUNREACH: 1357 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1358 tp->t_softerror = error; 1359 return (0); 1360 } 1361 /* FALLTHROUGH */ 1362 default: 1363 return (error); 1364 } 1365 } 1366 TCPSTAT_INC(tcps_sndtotal); 1367 1368 /* 1369 * Data sent (as far as we can tell). 1370 * If this advertises a larger window than any other segment, 1371 * then remember the size of the advertised window. 1372 * Any pending ACK has now been sent. 1373 */ 1374 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1375 tp->rcv_adv = tp->rcv_nxt + recwin; 1376 tp->last_ack_sent = tp->rcv_nxt; 1377 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1378 if (tcp_timer_active(tp, TT_DELACK)) 1379 tcp_timer_activate(tp, TT_DELACK, 0); 1380#if 0 1381 /* 1382 * This completely breaks TCP if newreno is turned on. What happens 1383 * is that if delayed-acks are turned on on the receiver, this code 1384 * on the transmitter effectively destroys the TCP window, forcing 1385 * it to four packets (1.5Kx4 = 6K window). 1386 */ 1387 if (sendalot && --maxburst) 1388 goto again; 1389#endif 1390 if (sendalot) 1391 goto again; 1392 return (0); 1393} 1394 1395void 1396tcp_setpersist(struct tcpcb *tp) 1397{ 1398 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1399 int tt; 1400 1401 tp->t_flags &= ~TF_PREVVALID; 1402 if (tcp_timer_active(tp, TT_REXMT)) 1403 panic("tcp_setpersist: retransmit pending"); 1404 /* 1405 * Start/restart persistance timer. 1406 */ 1407 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1408 TCPTV_PERSMIN, TCPTV_PERSMAX); 1409 tcp_timer_activate(tp, TT_PERSIST, tt); 1410 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1411 tp->t_rxtshift++; 1412} 1413 1414/* 1415 * Insert TCP options according to the supplied parameters to the place 1416 * optp in a consistent way. Can handle unaligned destinations. 1417 * 1418 * The order of the option processing is crucial for optimal packing and 1419 * alignment for the scarce option space. 1420 * 1421 * The optimal order for a SYN/SYN-ACK segment is: 1422 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1423 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1424 * 1425 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1426 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1427 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1428 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1429 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1430 */ 1431int 1432tcp_addoptions(struct tcpopt *to, u_char *optp) 1433{ 1434 u_int mask, optlen = 0; 1435 1436 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1437 if ((to->to_flags & mask) != mask) 1438 continue; 1439 if (optlen == TCP_MAXOLEN) 1440 break; 1441 switch (to->to_flags & mask) { 1442 case TOF_MSS: 1443 while (optlen % 4) { 1444 optlen += TCPOLEN_NOP; 1445 *optp++ = TCPOPT_NOP; 1446 } 1447 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1448 continue; 1449 optlen += TCPOLEN_MAXSEG; 1450 *optp++ = TCPOPT_MAXSEG; 1451 *optp++ = TCPOLEN_MAXSEG; 1452 to->to_mss = htons(to->to_mss); 1453 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1454 optp += sizeof(to->to_mss); 1455 break; 1456 case TOF_SCALE: 1457 while (!optlen || optlen % 2 != 1) { 1458 optlen += TCPOLEN_NOP; 1459 *optp++ = TCPOPT_NOP; 1460 } 1461 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1462 continue; 1463 optlen += TCPOLEN_WINDOW; 1464 *optp++ = TCPOPT_WINDOW; 1465 *optp++ = TCPOLEN_WINDOW; 1466 *optp++ = to->to_wscale; 1467 break; 1468 case TOF_SACKPERM: 1469 while (optlen % 2) { 1470 optlen += TCPOLEN_NOP; 1471 *optp++ = TCPOPT_NOP; 1472 } 1473 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1474 continue; 1475 optlen += TCPOLEN_SACK_PERMITTED; 1476 *optp++ = TCPOPT_SACK_PERMITTED; 1477 *optp++ = TCPOLEN_SACK_PERMITTED; 1478 break; 1479 case TOF_TS: 1480 while (!optlen || optlen % 4 != 2) { 1481 optlen += TCPOLEN_NOP; 1482 *optp++ = TCPOPT_NOP; 1483 } 1484 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1485 continue; 1486 optlen += TCPOLEN_TIMESTAMP; 1487 *optp++ = TCPOPT_TIMESTAMP; 1488 *optp++ = TCPOLEN_TIMESTAMP; 1489 to->to_tsval = htonl(to->to_tsval); 1490 to->to_tsecr = htonl(to->to_tsecr); 1491 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1492 optp += sizeof(to->to_tsval); 1493 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1494 optp += sizeof(to->to_tsecr); 1495 break; 1496#ifdef TCP_SIGNATURE 1497 case TOF_SIGNATURE: 1498 { 1499 int siglen = TCPOLEN_SIGNATURE - 2; 1500 1501 while (!optlen || optlen % 4 != 2) { 1502 optlen += TCPOLEN_NOP; 1503 *optp++ = TCPOPT_NOP; 1504 } 1505 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1506 continue; 1507 optlen += TCPOLEN_SIGNATURE; 1508 *optp++ = TCPOPT_SIGNATURE; 1509 *optp++ = TCPOLEN_SIGNATURE; 1510 to->to_signature = optp; 1511 while (siglen--) 1512 *optp++ = 0; 1513 break; 1514 } 1515#endif 1516 case TOF_SACK: 1517 { 1518 int sackblks = 0; 1519 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1520 tcp_seq sack_seq; 1521 1522 while (!optlen || optlen % 4 != 2) { 1523 optlen += TCPOLEN_NOP; 1524 *optp++ = TCPOPT_NOP; 1525 } 1526 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1527 continue; 1528 optlen += TCPOLEN_SACKHDR; 1529 *optp++ = TCPOPT_SACK; 1530 sackblks = min(to->to_nsacks, 1531 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1532 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1533 while (sackblks--) { 1534 sack_seq = htonl(sack->start); 1535 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1536 optp += sizeof(sack_seq); 1537 sack_seq = htonl(sack->end); 1538 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1539 optp += sizeof(sack_seq); 1540 optlen += TCPOLEN_SACK; 1541 sack++; 1542 } 1543 TCPSTAT_INC(tcps_sack_send_blocks); 1544 break; 1545 } 1546 default: 1547 panic("%s: unknown TCP option type", __func__); 1548 break; 1549 } 1550 } 1551 1552 /* Terminate and pad TCP options to a 4 byte boundary. */ 1553 if (optlen % 4) { 1554 optlen += TCPOLEN_EOL; 1555 *optp++ = TCPOPT_EOL; 1556 } 1557 /* 1558 * According to RFC 793 (STD0007): 1559 * "The content of the header beyond the End-of-Option option 1560 * must be header padding (i.e., zero)." 1561 * and later: "The padding is composed of zeros." 1562 */ 1563 while (optlen % 4) { 1564 optlen += TCPOLEN_PAD; 1565 *optp++ = TCPOPT_PAD; 1566 } 1567 1568 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1569 return (optlen); 1570} 1571