tcp_output.c revision 274043
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: stable/10/sys/netinet/tcp_output.c 274043 2014-11-03 12:38:29Z hselasky $"); 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 len = 0; 415 if (sendwin == 0) { 416 tcp_timer_activate(tp, TT_REXMT, 0); 417 tp->t_rxtshift = 0; 418 tp->snd_nxt = tp->snd_una; 419 if (!tcp_timer_active(tp, TT_PERSIST)) 420 tcp_setpersist(tp); 421 } 422 } 423 424 /* len will be >= 0 after this point. */ 425 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 426 427 /* 428 * Automatic sizing of send socket buffer. Often the send buffer 429 * size is not optimally adjusted to the actual network conditions 430 * at hand (delay bandwidth product). Setting the buffer size too 431 * small limits throughput on links with high bandwidth and high 432 * delay (eg. trans-continental/oceanic links). Setting the 433 * buffer size too big consumes too much real kernel memory, 434 * especially with many connections on busy servers. 435 * 436 * The criteria to step up the send buffer one notch are: 437 * 1. receive window of remote host is larger than send buffer 438 * (with a fudge factor of 5/4th); 439 * 2. send buffer is filled to 7/8th with data (so we actually 440 * have data to make use of it); 441 * 3. send buffer fill has not hit maximal automatic size; 442 * 4. our send window (slow start and cogestion controlled) is 443 * larger than sent but unacknowledged data in send buffer. 444 * 445 * The remote host receive window scaling factor may limit the 446 * growing of the send buffer before it reaches its allowed 447 * maximum. 448 * 449 * It scales directly with slow start or congestion window 450 * and does at most one step per received ACK. This fast 451 * scaling has the drawback of growing the send buffer beyond 452 * what is strictly necessary to make full use of a given 453 * delay*bandwith product. However testing has shown this not 454 * to be much of an problem. At worst we are trading wasting 455 * of available bandwith (the non-use of it) for wasting some 456 * socket buffer memory. 457 * 458 * TODO: Shrink send buffer during idle periods together 459 * with congestion window. Requires another timer. Has to 460 * wait for upcoming tcp timer rewrite. 461 */ 462 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 463 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 464 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 465 so->so_snd.sb_cc < V_tcp_autosndbuf_max && 466 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 467 if (!sbreserve_locked(&so->so_snd, 468 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 469 V_tcp_autosndbuf_max), so, curthread)) 470 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 471 } 472 } 473 474 /* 475 * Decide if we can use TCP Segmentation Offloading (if supported by 476 * hardware). 477 * 478 * TSO may only be used if we are in a pure bulk sending state. The 479 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 480 * IP options prevent using TSO. With TSO the TCP header is the same 481 * (except for the sequence number) for all generated packets. This 482 * makes it impossible to transmit any options which vary per generated 483 * segment or packet. 484 */ 485#ifdef IPSEC 486 /* 487 * Pre-calculate here as we save another lookup into the darknesses 488 * of IPsec that way and can actually decide if TSO is ok. 489 */ 490 ipsec_optlen = ipsec_hdrsiz_tcp(tp); 491#endif 492 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 493 ((tp->t_flags & TF_SIGNATURE) == 0) && 494 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 495#ifdef IPSEC 496 ipsec_optlen == 0 && 497#endif 498 tp->t_inpcb->inp_options == NULL && 499 tp->t_inpcb->in6p_options == NULL) 500 tso = 1; 501 502 if (sack_rxmit) { 503 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 504 flags &= ~TH_FIN; 505 } else { 506 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 507 flags &= ~TH_FIN; 508 } 509 510 recwin = sbspace(&so->so_rcv); 511 512 /* 513 * Sender silly window avoidance. We transmit under the following 514 * conditions when len is non-zero: 515 * 516 * - We have a full segment (or more with TSO) 517 * - This is the last buffer in a write()/send() and we are 518 * either idle or running NODELAY 519 * - we've timed out (e.g. persist timer) 520 * - we have more then 1/2 the maximum send window's worth of 521 * data (receiver may be limited the window size) 522 * - we need to retransmit 523 */ 524 if (len) { 525 if (len >= tp->t_maxseg) 526 goto send; 527 /* 528 * NOTE! on localhost connections an 'ack' from the remote 529 * end may occur synchronously with the output and cause 530 * us to flush a buffer queued with moretocome. XXX 531 * 532 * note: the len + off check is almost certainly unnecessary. 533 */ 534 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 535 (idle || (tp->t_flags & TF_NODELAY)) && 536 len + off >= so->so_snd.sb_cc && 537 (tp->t_flags & TF_NOPUSH) == 0) { 538 goto send; 539 } 540 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 541 goto send; 542 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 543 goto send; 544 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 545 goto send; 546 if (sack_rxmit) 547 goto send; 548 } 549 550 /* 551 * Sending of standalone window updates. 552 * 553 * Window updates are important when we close our window due to a 554 * full socket buffer and are opening it again after the application 555 * reads data from it. Once the window has opened again and the 556 * remote end starts to send again the ACK clock takes over and 557 * provides the most current window information. 558 * 559 * We must avoid the silly window syndrome whereas every read 560 * from the receive buffer, no matter how small, causes a window 561 * update to be sent. We also should avoid sending a flurry of 562 * window updates when the socket buffer had queued a lot of data 563 * and the application is doing small reads. 564 * 565 * Prevent a flurry of pointless window updates by only sending 566 * an update when we can increase the advertized window by more 567 * than 1/4th of the socket buffer capacity. When the buffer is 568 * getting full or is very small be more aggressive and send an 569 * update whenever we can increase by two mss sized segments. 570 * In all other situations the ACK's to new incoming data will 571 * carry further window increases. 572 * 573 * Don't send an independent window update if a delayed 574 * ACK is pending (it will get piggy-backed on it) or the 575 * remote side already has done a half-close and won't send 576 * more data. Skip this if the connection is in T/TCP 577 * half-open state. 578 */ 579 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 580 !(tp->t_flags & TF_DELACK) && 581 !TCPS_HAVERCVDFIN(tp->t_state)) { 582 /* 583 * "adv" is the amount we could increase the window, 584 * taking into account that we are limited by 585 * TCP_MAXWIN << tp->rcv_scale. 586 */ 587 long adv; 588 int oldwin; 589 590 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 591 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 592 oldwin = (tp->rcv_adv - tp->rcv_nxt); 593 adv -= oldwin; 594 } else 595 oldwin = 0; 596 597 /* 598 * If the new window size ends up being the same as the old 599 * size when it is scaled, then don't force a window update. 600 */ 601 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 602 goto dontupdate; 603 604 if (adv >= (long)(2 * tp->t_maxseg) && 605 (adv >= (long)(so->so_rcv.sb_hiwat / 4) || 606 recwin <= (long)(so->so_rcv.sb_hiwat / 8) || 607 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 608 goto send; 609 } 610dontupdate: 611 612 /* 613 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 614 * is also a catch-all for the retransmit timer timeout case. 615 */ 616 if (tp->t_flags & TF_ACKNOW) 617 goto send; 618 if ((flags & TH_RST) || 619 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 620 goto send; 621 if (SEQ_GT(tp->snd_up, tp->snd_una)) 622 goto send; 623 /* 624 * If our state indicates that FIN should be sent 625 * and we have not yet done so, then we need to send. 626 */ 627 if (flags & TH_FIN && 628 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 629 goto send; 630 /* 631 * In SACK, it is possible for tcp_output to fail to send a segment 632 * after the retransmission timer has been turned off. Make sure 633 * that the retransmission timer is set. 634 */ 635 if ((tp->t_flags & TF_SACK_PERMIT) && 636 SEQ_GT(tp->snd_max, tp->snd_una) && 637 !tcp_timer_active(tp, TT_REXMT) && 638 !tcp_timer_active(tp, TT_PERSIST)) { 639 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 640 goto just_return; 641 } 642 /* 643 * TCP window updates are not reliable, rather a polling protocol 644 * using ``persist'' packets is used to insure receipt of window 645 * updates. The three ``states'' for the output side are: 646 * idle not doing retransmits or persists 647 * persisting to move a small or zero window 648 * (re)transmitting and thereby not persisting 649 * 650 * tcp_timer_active(tp, TT_PERSIST) 651 * is true when we are in persist state. 652 * (tp->t_flags & TF_FORCEDATA) 653 * is set when we are called to send a persist packet. 654 * tcp_timer_active(tp, TT_REXMT) 655 * is set when we are retransmitting 656 * The output side is idle when both timers are zero. 657 * 658 * If send window is too small, there is data to transmit, and no 659 * retransmit or persist is pending, then go to persist state. 660 * If nothing happens soon, send when timer expires: 661 * if window is nonzero, transmit what we can, 662 * otherwise force out a byte. 663 */ 664 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) && 665 !tcp_timer_active(tp, TT_PERSIST)) { 666 tp->t_rxtshift = 0; 667 tcp_setpersist(tp); 668 } 669 670 /* 671 * No reason to send a segment, just return. 672 */ 673just_return: 674 SOCKBUF_UNLOCK(&so->so_snd); 675 return (0); 676 677send: 678 SOCKBUF_LOCK_ASSERT(&so->so_snd); 679 if (len > 0) { 680 if (len >= tp->t_maxseg) 681 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 682 else 683 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 684 } 685 /* 686 * Before ESTABLISHED, force sending of initial options 687 * unless TCP set not to do any options. 688 * NOTE: we assume that the IP/TCP header plus TCP options 689 * always fit in a single mbuf, leaving room for a maximum 690 * link header, i.e. 691 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 692 */ 693 optlen = 0; 694#ifdef INET6 695 if (isipv6) 696 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 697 else 698#endif 699 hdrlen = sizeof (struct tcpiphdr); 700 701 /* 702 * Compute options for segment. 703 * We only have to care about SYN and established connection 704 * segments. Options for SYN-ACK segments are handled in TCP 705 * syncache. 706 */ 707 if ((tp->t_flags & TF_NOOPT) == 0) { 708 to.to_flags = 0; 709 /* Maximum segment size. */ 710 if (flags & TH_SYN) { 711 tp->snd_nxt = tp->iss; 712 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 713 to.to_flags |= TOF_MSS; 714 } 715 /* Window scaling. */ 716 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 717 to.to_wscale = tp->request_r_scale; 718 to.to_flags |= TOF_SCALE; 719 } 720 /* Timestamps. */ 721 if ((tp->t_flags & TF_RCVD_TSTMP) || 722 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 723 to.to_tsval = tcp_ts_getticks() + tp->ts_offset; 724 to.to_tsecr = tp->ts_recent; 725 to.to_flags |= TOF_TS; 726 /* Set receive buffer autosizing timestamp. */ 727 if (tp->rfbuf_ts == 0 && 728 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 729 tp->rfbuf_ts = tcp_ts_getticks(); 730 } 731 /* Selective ACK's. */ 732 if (tp->t_flags & TF_SACK_PERMIT) { 733 if (flags & TH_SYN) 734 to.to_flags |= TOF_SACKPERM; 735 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 736 (tp->t_flags & TF_SACK_PERMIT) && 737 tp->rcv_numsacks > 0) { 738 to.to_flags |= TOF_SACK; 739 to.to_nsacks = tp->rcv_numsacks; 740 to.to_sacks = (u_char *)tp->sackblks; 741 } 742 } 743#ifdef TCP_SIGNATURE 744 /* TCP-MD5 (RFC2385). */ 745 if (tp->t_flags & TF_SIGNATURE) 746 to.to_flags |= TOF_SIGNATURE; 747#endif /* TCP_SIGNATURE */ 748 749 /* Processing the options. */ 750 hdrlen += optlen = tcp_addoptions(&to, opt); 751 } 752 753#ifdef INET6 754 if (isipv6) 755 ipoptlen = ip6_optlen(tp->t_inpcb); 756 else 757#endif 758 if (tp->t_inpcb->inp_options) 759 ipoptlen = tp->t_inpcb->inp_options->m_len - 760 offsetof(struct ipoption, ipopt_list); 761 else 762 ipoptlen = 0; 763#ifdef IPSEC 764 ipoptlen += ipsec_optlen; 765#endif 766 767 /* 768 * Adjust data length if insertion of options will 769 * bump the packet length beyond the t_maxopd length. 770 * Clear the FIN bit because we cut off the tail of 771 * the segment. 772 */ 773 if (len + optlen + ipoptlen > tp->t_maxopd) { 774 flags &= ~TH_FIN; 775 776 if (tso) { 777 u_int if_hw_tsomax; 778 u_int if_hw_tsomaxsegcount; 779 u_int if_hw_tsomaxsegsize; 780 struct mbuf *mb; 781 u_int moff; 782 int max_len; 783 784 /* extract TSO information */ 785 if_hw_tsomax = tp->t_tsomax; 786 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 787 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 788 789 /* 790 * Limit a TSO burst to prevent it from 791 * overflowing or exceeding the maximum length 792 * allowed by the network interface: 793 */ 794 KASSERT(ipoptlen == 0, 795 ("%s: TSO can't do IP options", __func__)); 796 797 /* 798 * Check if we should limit by maximum payload 799 * length: 800 */ 801 if (if_hw_tsomax != 0) { 802 /* compute maximum TSO length */ 803 max_len = (if_hw_tsomax - hdrlen); 804 if (max_len <= 0) { 805 len = 0; 806 } else if (len > (u_int)max_len) { 807 sendalot = 1; 808 len = (u_int)max_len; 809 } 810 } 811 812 /* 813 * Check if we should limit by maximum segment 814 * size and count: 815 */ 816 if (if_hw_tsomaxsegcount != 0 && 817 if_hw_tsomaxsegsize != 0) { 818 max_len = 0; 819 mb = sbsndmbuf(&so->so_snd, off, &moff); 820 821 while (mb != NULL && (u_int)max_len < len) { 822 u_int mlen; 823 u_int frags; 824 825 /* 826 * Get length of mbuf fragment 827 * and how many hardware frags, 828 * rounded up, it would use: 829 */ 830 mlen = (mb->m_len - moff); 831 frags = howmany(mlen, 832 if_hw_tsomaxsegsize); 833 834 /* Handle special case: Zero Length Mbuf */ 835 if (frags == 0) 836 frags = 1; 837 838 /* 839 * Check if the fragment limit 840 * will be reached or exceeded: 841 */ 842 if (frags >= if_hw_tsomaxsegcount) { 843 max_len += min(mlen, 844 if_hw_tsomaxsegcount * 845 if_hw_tsomaxsegsize); 846 break; 847 } 848 max_len += mlen; 849 if_hw_tsomaxsegcount -= frags; 850 moff = 0; 851 mb = mb->m_next; 852 } 853 if (max_len <= 0) { 854 len = 0; 855 } else if (len > (u_int)max_len) { 856 sendalot = 1; 857 len = (u_int)max_len; 858 } 859 } 860 861 /* 862 * Prevent the last segment from being 863 * fractional unless the send sockbuf can be 864 * emptied: 865 */ 866 max_len = (tp->t_maxopd - optlen); 867 if ((off + len) < so->so_snd.sb_cc) { 868 moff = len % (u_int)max_len; 869 if (moff != 0) { 870 len -= moff; 871 sendalot = 1; 872 } 873 } 874 875 /* 876 * In case there are too many small fragments 877 * don't use TSO: 878 */ 879 if (len <= (u_int)max_len) { 880 len = (u_int)max_len; 881 sendalot = 1; 882 tso = 0; 883 } 884 885 /* 886 * Send the FIN in a separate segment 887 * after the bulk sending is done. 888 * We don't trust the TSO implementations 889 * to clear the FIN flag on all but the 890 * last segment. 891 */ 892 if (tp->t_flags & TF_NEEDFIN) 893 sendalot = 1; 894 895 } else { 896 len = tp->t_maxopd - optlen - ipoptlen; 897 sendalot = 1; 898 } 899 } else 900 tso = 0; 901 902 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 903 ("%s: len > IP_MAXPACKET", __func__)); 904 905/*#ifdef DIAGNOSTIC*/ 906#ifdef INET6 907 if (max_linkhdr + hdrlen > MCLBYTES) 908#else 909 if (max_linkhdr + hdrlen > MHLEN) 910#endif 911 panic("tcphdr too big"); 912/*#endif*/ 913 914 /* 915 * This KASSERT is here to catch edge cases at a well defined place. 916 * Before, those had triggered (random) panic conditions further down. 917 */ 918 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 919 920 /* 921 * Grab a header mbuf, attaching a copy of data to 922 * be transmitted, and initialize the header from 923 * the template for sends on this connection. 924 */ 925 if (len) { 926 struct mbuf *mb; 927 u_int moff; 928 929 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 930 TCPSTAT_INC(tcps_sndprobe); 931 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 932 tp->t_sndrexmitpack++; 933 TCPSTAT_INC(tcps_sndrexmitpack); 934 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 935 } else { 936 TCPSTAT_INC(tcps_sndpack); 937 TCPSTAT_ADD(tcps_sndbyte, len); 938 } 939#ifdef INET6 940 if (MHLEN < hdrlen + max_linkhdr) 941 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 942 else 943#endif 944 m = m_gethdr(M_NOWAIT, MT_DATA); 945 946 if (m == NULL) { 947 SOCKBUF_UNLOCK(&so->so_snd); 948 error = ENOBUFS; 949 sack_rxmit = 0; 950 goto out; 951 } 952 953 m->m_data += max_linkhdr; 954 m->m_len = hdrlen; 955 956 /* 957 * Start the m_copy functions from the closest mbuf 958 * to the offset in the socket buffer chain. 959 */ 960 mb = sbsndptr(&so->so_snd, off, len, &moff); 961 962 if (len <= MHLEN - hdrlen - max_linkhdr) { 963 m_copydata(mb, moff, (int)len, 964 mtod(m, caddr_t) + hdrlen); 965 m->m_len += len; 966 } else { 967 m->m_next = m_copy(mb, moff, (int)len); 968 if (m->m_next == NULL) { 969 SOCKBUF_UNLOCK(&so->so_snd); 970 (void) m_free(m); 971 error = ENOBUFS; 972 sack_rxmit = 0; 973 goto out; 974 } 975 } 976 977 /* 978 * If we're sending everything we've got, set PUSH. 979 * (This will keep happy those implementations which only 980 * give data to the user when a buffer fills or 981 * a PUSH comes in.) 982 */ 983 if (off + len == so->so_snd.sb_cc) 984 flags |= TH_PUSH; 985 SOCKBUF_UNLOCK(&so->so_snd); 986 } else { 987 SOCKBUF_UNLOCK(&so->so_snd); 988 if (tp->t_flags & TF_ACKNOW) 989 TCPSTAT_INC(tcps_sndacks); 990 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 991 TCPSTAT_INC(tcps_sndctrl); 992 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 993 TCPSTAT_INC(tcps_sndurg); 994 else 995 TCPSTAT_INC(tcps_sndwinup); 996 997 m = m_gethdr(M_NOWAIT, MT_DATA); 998 if (m == NULL) { 999 error = ENOBUFS; 1000 sack_rxmit = 0; 1001 goto out; 1002 } 1003#ifdef INET6 1004 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1005 MHLEN >= hdrlen) { 1006 MH_ALIGN(m, hdrlen); 1007 } else 1008#endif 1009 m->m_data += max_linkhdr; 1010 m->m_len = hdrlen; 1011 } 1012 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1013 m->m_pkthdr.rcvif = (struct ifnet *)0; 1014#ifdef MAC 1015 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1016#endif 1017#ifdef INET6 1018 if (isipv6) { 1019 ip6 = mtod(m, struct ip6_hdr *); 1020 th = (struct tcphdr *)(ip6 + 1); 1021 tcpip_fillheaders(tp->t_inpcb, ip6, th); 1022 } else 1023#endif /* INET6 */ 1024 { 1025 ip = mtod(m, struct ip *); 1026 ipov = (struct ipovly *)ip; 1027 th = (struct tcphdr *)(ip + 1); 1028 tcpip_fillheaders(tp->t_inpcb, ip, th); 1029 } 1030 1031 /* 1032 * Fill in fields, remembering maximum advertised 1033 * window for use in delaying messages about window sizes. 1034 * If resending a FIN, be sure not to use a new sequence number. 1035 */ 1036 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1037 tp->snd_nxt == tp->snd_max) 1038 tp->snd_nxt--; 1039 /* 1040 * If we are starting a connection, send ECN setup 1041 * SYN packet. If we are on a retransmit, we may 1042 * resend those bits a number of times as per 1043 * RFC 3168. 1044 */ 1045 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 1046 if (tp->t_rxtshift >= 1) { 1047 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 1048 flags |= TH_ECE|TH_CWR; 1049 } else 1050 flags |= TH_ECE|TH_CWR; 1051 } 1052 1053 if (tp->t_state == TCPS_ESTABLISHED && 1054 (tp->t_flags & TF_ECN_PERMIT)) { 1055 /* 1056 * If the peer has ECN, mark data packets with 1057 * ECN capable transmission (ECT). 1058 * Ignore pure ack packets, retransmissions and window probes. 1059 */ 1060 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1061 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 1062#ifdef INET6 1063 if (isipv6) 1064 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1065 else 1066#endif 1067 ip->ip_tos |= IPTOS_ECN_ECT0; 1068 TCPSTAT_INC(tcps_ecn_ect0); 1069 } 1070 1071 /* 1072 * Reply with proper ECN notifications. 1073 */ 1074 if (tp->t_flags & TF_ECN_SND_CWR) { 1075 flags |= TH_CWR; 1076 tp->t_flags &= ~TF_ECN_SND_CWR; 1077 } 1078 if (tp->t_flags & TF_ECN_SND_ECE) 1079 flags |= TH_ECE; 1080 } 1081 1082 /* 1083 * If we are doing retransmissions, then snd_nxt will 1084 * not reflect the first unsent octet. For ACK only 1085 * packets, we do not want the sequence number of the 1086 * retransmitted packet, we want the sequence number 1087 * of the next unsent octet. So, if there is no data 1088 * (and no SYN or FIN), use snd_max instead of snd_nxt 1089 * when filling in ti_seq. But if we are in persist 1090 * state, snd_max might reflect one byte beyond the 1091 * right edge of the window, so use snd_nxt in that 1092 * case, since we know we aren't doing a retransmission. 1093 * (retransmit and persist are mutually exclusive...) 1094 */ 1095 if (sack_rxmit == 0) { 1096 if (len || (flags & (TH_SYN|TH_FIN)) || 1097 tcp_timer_active(tp, TT_PERSIST)) 1098 th->th_seq = htonl(tp->snd_nxt); 1099 else 1100 th->th_seq = htonl(tp->snd_max); 1101 } else { 1102 th->th_seq = htonl(p->rxmit); 1103 p->rxmit += len; 1104 tp->sackhint.sack_bytes_rexmit += len; 1105 } 1106 th->th_ack = htonl(tp->rcv_nxt); 1107 if (optlen) { 1108 bcopy(opt, th + 1, optlen); 1109 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1110 } 1111 th->th_flags = flags; 1112 /* 1113 * Calculate receive window. Don't shrink window, 1114 * but avoid silly window syndrome. 1115 */ 1116 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 1117 recwin < (long)tp->t_maxseg) 1118 recwin = 0; 1119 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1120 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 1121 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 1122 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 1123 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 1124 1125 /* 1126 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1127 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1128 * case is handled in syncache. 1129 */ 1130 if (flags & TH_SYN) 1131 th->th_win = htons((u_short) 1132 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1133 else 1134 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1135 1136 /* 1137 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1138 * a 0 window. This may cause the remote transmitter to stall. This 1139 * flag tells soreceive() to disable delayed acknowledgements when 1140 * draining the buffer. This can occur if the receiver is attempting 1141 * to read more data than can be buffered prior to transmitting on 1142 * the connection. 1143 */ 1144 if (th->th_win == 0) { 1145 tp->t_sndzerowin++; 1146 tp->t_flags |= TF_RXWIN0SENT; 1147 } else 1148 tp->t_flags &= ~TF_RXWIN0SENT; 1149 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1150 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1151 th->th_flags |= TH_URG; 1152 } else 1153 /* 1154 * If no urgent pointer to send, then we pull 1155 * the urgent pointer to the left edge of the send window 1156 * so that it doesn't drift into the send window on sequence 1157 * number wraparound. 1158 */ 1159 tp->snd_up = tp->snd_una; /* drag it along */ 1160 1161#ifdef TCP_SIGNATURE 1162 if (tp->t_flags & TF_SIGNATURE) { 1163 int sigoff = to.to_signature - opt; 1164 tcp_signature_compute(m, 0, len, optlen, 1165 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1166 } 1167#endif 1168 1169 /* 1170 * Put TCP length in extended header, and then 1171 * checksum extended header and data. 1172 */ 1173 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1174 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1175#ifdef INET6 1176 if (isipv6) { 1177 /* 1178 * ip6_plen is not need to be filled now, and will be filled 1179 * in ip6_output. 1180 */ 1181 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1182 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1183 optlen + len, IPPROTO_TCP, 0); 1184 } 1185#endif 1186#if defined(INET6) && defined(INET) 1187 else 1188#endif 1189#ifdef INET 1190 { 1191 m->m_pkthdr.csum_flags = CSUM_TCP; 1192 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1193 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1194 1195 /* IP version must be set here for ipv4/ipv6 checking later */ 1196 KASSERT(ip->ip_v == IPVERSION, 1197 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1198 } 1199#endif 1200 1201 /* 1202 * Enable TSO and specify the size of the segments. 1203 * The TCP pseudo header checksum is always provided. 1204 * XXX: Fixme: This is currently not the case for IPv6. 1205 */ 1206 if (tso) { 1207 KASSERT(len > tp->t_maxopd - optlen, 1208 ("%s: len <= tso_segsz", __func__)); 1209 m->m_pkthdr.csum_flags |= CSUM_TSO; 1210 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 1211 } 1212 1213#ifdef IPSEC 1214 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1215 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u", 1216 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1217#else 1218 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1219 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u", 1220 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1221#endif 1222 1223 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1224 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1225 1226#ifdef TCPDEBUG 1227 /* 1228 * Trace. 1229 */ 1230 if (so->so_options & SO_DEBUG) { 1231 u_short save = 0; 1232#ifdef INET6 1233 if (!isipv6) 1234#endif 1235 { 1236 save = ipov->ih_len; 1237 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1238 } 1239 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1240#ifdef INET6 1241 if (!isipv6) 1242#endif 1243 ipov->ih_len = save; 1244 } 1245#endif /* TCPDEBUG */ 1246 1247 /* 1248 * Fill in IP length and desired time to live and 1249 * send to IP level. There should be a better way 1250 * to handle ttl and tos; we could keep them in 1251 * the template, but need a way to checksum without them. 1252 */ 1253 /* 1254 * m->m_pkthdr.len should have been set before cksum calcuration, 1255 * because in6_cksum() need it. 1256 */ 1257#ifdef INET6 1258 if (isipv6) { 1259 struct route_in6 ro; 1260 1261 bzero(&ro, sizeof(ro)); 1262 /* 1263 * we separately set hoplimit for every segment, since the 1264 * user might want to change the value via setsockopt. 1265 * Also, desired default hop limit might be changed via 1266 * Neighbor Discovery. 1267 */ 1268 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1269 1270 /* 1271 * Set the packet size here for the benefit of DTrace probes. 1272 * ip6_output() will set it properly; it's supposed to include 1273 * the option header lengths as well. 1274 */ 1275 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1276 1277 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) 1278 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1279 else 1280 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1281 1282 if (tp->t_state == TCPS_SYN_SENT) 1283 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1284 1285 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1286 1287 /* TODO: IPv6 IP6TOS_ECT bit on */ 1288 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro, 1289 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1290 NULL, NULL, tp->t_inpcb); 1291 1292 if (error == EMSGSIZE && ro.ro_rt != NULL) 1293 mtu = ro.ro_rt->rt_mtu; 1294 RO_RTFREE(&ro); 1295 } 1296#endif /* INET6 */ 1297#if defined(INET) && defined(INET6) 1298 else 1299#endif 1300#ifdef INET 1301 { 1302 struct route ro; 1303 1304 bzero(&ro, sizeof(ro)); 1305 ip->ip_len = htons(m->m_pkthdr.len); 1306#ifdef INET6 1307 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1308 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1309#endif /* INET6 */ 1310 /* 1311 * If we do path MTU discovery, then we set DF on every packet. 1312 * This might not be the best thing to do according to RFC3390 1313 * Section 2. However the tcp hostcache migitates the problem 1314 * so it affects only the first tcp connection with a host. 1315 * 1316 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1317 */ 1318 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) { 1319 ip->ip_off |= htons(IP_DF); 1320 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1321 } else { 1322 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1323 } 1324 1325 if (tp->t_state == TCPS_SYN_SENT) 1326 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1327 1328 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1329 1330 error = ip_output(m, tp->t_inpcb->inp_options, &ro, 1331 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1332 tp->t_inpcb); 1333 1334 if (error == EMSGSIZE && ro.ro_rt != NULL) 1335 mtu = ro.ro_rt->rt_mtu; 1336 RO_RTFREE(&ro); 1337 } 1338#endif /* INET */ 1339 1340out: 1341 /* 1342 * In transmit state, time the transmission and arrange for 1343 * the retransmit. In persist state, just set snd_max. 1344 */ 1345 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1346 !tcp_timer_active(tp, TT_PERSIST)) { 1347 tcp_seq startseq = tp->snd_nxt; 1348 1349 /* 1350 * Advance snd_nxt over sequence space of this segment. 1351 */ 1352 if (flags & (TH_SYN|TH_FIN)) { 1353 if (flags & TH_SYN) 1354 tp->snd_nxt++; 1355 if (flags & TH_FIN) { 1356 tp->snd_nxt++; 1357 tp->t_flags |= TF_SENTFIN; 1358 } 1359 } 1360 if (sack_rxmit) 1361 goto timer; 1362 tp->snd_nxt += len; 1363 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1364 tp->snd_max = tp->snd_nxt; 1365 /* 1366 * Time this transmission if not a retransmission and 1367 * not currently timing anything. 1368 */ 1369 if (tp->t_rtttime == 0) { 1370 tp->t_rtttime = ticks; 1371 tp->t_rtseq = startseq; 1372 TCPSTAT_INC(tcps_segstimed); 1373 } 1374 } 1375 1376 /* 1377 * Set retransmit timer if not currently set, 1378 * and not doing a pure ack or a keep-alive probe. 1379 * Initial value for retransmit timer is smoothed 1380 * round-trip time + 2 * round-trip time variance. 1381 * Initialize shift counter which is used for backoff 1382 * of retransmit time. 1383 */ 1384timer: 1385 if (!tcp_timer_active(tp, TT_REXMT) && 1386 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1387 (tp->snd_nxt != tp->snd_una))) { 1388 if (tcp_timer_active(tp, TT_PERSIST)) { 1389 tcp_timer_activate(tp, TT_PERSIST, 0); 1390 tp->t_rxtshift = 0; 1391 } 1392 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1393 } 1394 } else { 1395 /* 1396 * Persist case, update snd_max but since we are in 1397 * persist mode (no window) we do not update snd_nxt. 1398 */ 1399 int xlen = len; 1400 if (flags & TH_SYN) 1401 ++xlen; 1402 if (flags & TH_FIN) { 1403 ++xlen; 1404 tp->t_flags |= TF_SENTFIN; 1405 } 1406 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1407 tp->snd_max = tp->snd_nxt + len; 1408 } 1409 1410 if (error) { 1411 1412 /* 1413 * We know that the packet was lost, so back out the 1414 * sequence number advance, if any. 1415 * 1416 * If the error is EPERM the packet got blocked by the 1417 * local firewall. Normally we should terminate the 1418 * connection but the blocking may have been spurious 1419 * due to a firewall reconfiguration cycle. So we treat 1420 * it like a packet loss and let the retransmit timer and 1421 * timeouts do their work over time. 1422 * XXX: It is a POLA question whether calling tcp_drop right 1423 * away would be the really correct behavior instead. 1424 */ 1425 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1426 !tcp_timer_active(tp, TT_PERSIST)) && 1427 ((flags & TH_SYN) == 0) && 1428 (error != EPERM)) { 1429 if (sack_rxmit) { 1430 p->rxmit -= len; 1431 tp->sackhint.sack_bytes_rexmit -= len; 1432 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1433 ("sackhint bytes rtx >= 0")); 1434 } else 1435 tp->snd_nxt -= len; 1436 } 1437 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1438 switch (error) { 1439 case EPERM: 1440 tp->t_softerror = error; 1441 return (error); 1442 case ENOBUFS: 1443 if (!tcp_timer_active(tp, TT_REXMT) && 1444 !tcp_timer_active(tp, TT_PERSIST)) 1445 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1446 tp->snd_cwnd = tp->t_maxseg; 1447 return (0); 1448 case EMSGSIZE: 1449 /* 1450 * For some reason the interface we used initially 1451 * to send segments changed to another or lowered 1452 * its MTU. 1453 * If TSO was active we either got an interface 1454 * without TSO capabilits or TSO was turned off. 1455 * If we obtained mtu from ip_output() then update 1456 * it and try again. 1457 */ 1458 if (tso) 1459 tp->t_flags &= ~TF_TSO; 1460 if (mtu != 0) { 1461 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1462 goto again; 1463 } 1464 return (error); 1465 case EHOSTDOWN: 1466 case EHOSTUNREACH: 1467 case ENETDOWN: 1468 case ENETUNREACH: 1469 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1470 tp->t_softerror = error; 1471 return (0); 1472 } 1473 /* FALLTHROUGH */ 1474 default: 1475 return (error); 1476 } 1477 } 1478 TCPSTAT_INC(tcps_sndtotal); 1479 1480 /* 1481 * Data sent (as far as we can tell). 1482 * If this advertises a larger window than any other segment, 1483 * then remember the size of the advertised window. 1484 * Any pending ACK has now been sent. 1485 */ 1486 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1487 tp->rcv_adv = tp->rcv_nxt + recwin; 1488 tp->last_ack_sent = tp->rcv_nxt; 1489 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1490 if (tcp_timer_active(tp, TT_DELACK)) 1491 tcp_timer_activate(tp, TT_DELACK, 0); 1492#if 0 1493 /* 1494 * This completely breaks TCP if newreno is turned on. What happens 1495 * is that if delayed-acks are turned on on the receiver, this code 1496 * on the transmitter effectively destroys the TCP window, forcing 1497 * it to four packets (1.5Kx4 = 6K window). 1498 */ 1499 if (sendalot && --maxburst) 1500 goto again; 1501#endif 1502 if (sendalot) 1503 goto again; 1504 return (0); 1505} 1506 1507void 1508tcp_setpersist(struct tcpcb *tp) 1509{ 1510 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1511 int tt; 1512 1513 tp->t_flags &= ~TF_PREVVALID; 1514 if (tcp_timer_active(tp, TT_REXMT)) 1515 panic("tcp_setpersist: retransmit pending"); 1516 /* 1517 * Start/restart persistance timer. 1518 */ 1519 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1520 TCPTV_PERSMIN, TCPTV_PERSMAX); 1521 tcp_timer_activate(tp, TT_PERSIST, tt); 1522 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1523 tp->t_rxtshift++; 1524} 1525 1526/* 1527 * Insert TCP options according to the supplied parameters to the place 1528 * optp in a consistent way. Can handle unaligned destinations. 1529 * 1530 * The order of the option processing is crucial for optimal packing and 1531 * alignment for the scarce option space. 1532 * 1533 * The optimal order for a SYN/SYN-ACK segment is: 1534 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1535 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1536 * 1537 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1538 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1539 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1540 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1541 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1542 */ 1543int 1544tcp_addoptions(struct tcpopt *to, u_char *optp) 1545{ 1546 u_int mask, optlen = 0; 1547 1548 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1549 if ((to->to_flags & mask) != mask) 1550 continue; 1551 if (optlen == TCP_MAXOLEN) 1552 break; 1553 switch (to->to_flags & mask) { 1554 case TOF_MSS: 1555 while (optlen % 4) { 1556 optlen += TCPOLEN_NOP; 1557 *optp++ = TCPOPT_NOP; 1558 } 1559 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1560 continue; 1561 optlen += TCPOLEN_MAXSEG; 1562 *optp++ = TCPOPT_MAXSEG; 1563 *optp++ = TCPOLEN_MAXSEG; 1564 to->to_mss = htons(to->to_mss); 1565 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1566 optp += sizeof(to->to_mss); 1567 break; 1568 case TOF_SCALE: 1569 while (!optlen || optlen % 2 != 1) { 1570 optlen += TCPOLEN_NOP; 1571 *optp++ = TCPOPT_NOP; 1572 } 1573 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1574 continue; 1575 optlen += TCPOLEN_WINDOW; 1576 *optp++ = TCPOPT_WINDOW; 1577 *optp++ = TCPOLEN_WINDOW; 1578 *optp++ = to->to_wscale; 1579 break; 1580 case TOF_SACKPERM: 1581 while (optlen % 2) { 1582 optlen += TCPOLEN_NOP; 1583 *optp++ = TCPOPT_NOP; 1584 } 1585 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1586 continue; 1587 optlen += TCPOLEN_SACK_PERMITTED; 1588 *optp++ = TCPOPT_SACK_PERMITTED; 1589 *optp++ = TCPOLEN_SACK_PERMITTED; 1590 break; 1591 case TOF_TS: 1592 while (!optlen || optlen % 4 != 2) { 1593 optlen += TCPOLEN_NOP; 1594 *optp++ = TCPOPT_NOP; 1595 } 1596 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1597 continue; 1598 optlen += TCPOLEN_TIMESTAMP; 1599 *optp++ = TCPOPT_TIMESTAMP; 1600 *optp++ = TCPOLEN_TIMESTAMP; 1601 to->to_tsval = htonl(to->to_tsval); 1602 to->to_tsecr = htonl(to->to_tsecr); 1603 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1604 optp += sizeof(to->to_tsval); 1605 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1606 optp += sizeof(to->to_tsecr); 1607 break; 1608 case TOF_SIGNATURE: 1609 { 1610 int siglen = TCPOLEN_SIGNATURE - 2; 1611 1612 while (!optlen || optlen % 4 != 2) { 1613 optlen += TCPOLEN_NOP; 1614 *optp++ = TCPOPT_NOP; 1615 } 1616 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1617 continue; 1618 optlen += TCPOLEN_SIGNATURE; 1619 *optp++ = TCPOPT_SIGNATURE; 1620 *optp++ = TCPOLEN_SIGNATURE; 1621 to->to_signature = optp; 1622 while (siglen--) 1623 *optp++ = 0; 1624 break; 1625 } 1626 case TOF_SACK: 1627 { 1628 int sackblks = 0; 1629 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1630 tcp_seq sack_seq; 1631 1632 while (!optlen || optlen % 4 != 2) { 1633 optlen += TCPOLEN_NOP; 1634 *optp++ = TCPOPT_NOP; 1635 } 1636 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1637 continue; 1638 optlen += TCPOLEN_SACKHDR; 1639 *optp++ = TCPOPT_SACK; 1640 sackblks = min(to->to_nsacks, 1641 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1642 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1643 while (sackblks--) { 1644 sack_seq = htonl(sack->start); 1645 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1646 optp += sizeof(sack_seq); 1647 sack_seq = htonl(sack->end); 1648 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1649 optp += sizeof(sack_seq); 1650 optlen += TCPOLEN_SACK; 1651 sack++; 1652 } 1653 TCPSTAT_INC(tcps_sack_send_blocks); 1654 break; 1655 } 1656 default: 1657 panic("%s: unknown TCP option type", __func__); 1658 break; 1659 } 1660 } 1661 1662 /* Terminate and pad TCP options to a 4 byte boundary. */ 1663 if (optlen % 4) { 1664 optlen += TCPOLEN_EOL; 1665 *optp++ = TCPOPT_EOL; 1666 } 1667 /* 1668 * According to RFC 793 (STD0007): 1669 * "The content of the header beyond the End-of-Option option 1670 * must be header padding (i.e., zero)." 1671 * and later: "The padding is composed of zeros." 1672 */ 1673 while (optlen % 4) { 1674 optlen += TCPOLEN_PAD; 1675 *optp++ = TCPOPT_PAD; 1676 } 1677 1678 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1679 return (optlen); 1680} 1681