tcp_output.c revision 313340
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 313340 2017-02-06 13:32:22Z cy $"); 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#ifdef TCP_RFC7413 73#include <netinet/tcp_fastopen.h> 74#endif 75#define TCPOUTFLAGS 76#include <netinet/tcp_fsm.h> 77#include <netinet/tcp_seq.h> 78#include <netinet/tcp_timer.h> 79#include <netinet/tcp_var.h> 80#include <netinet/tcpip.h> 81#ifdef TCPDEBUG 82#include <netinet/tcp_debug.h> 83#endif 84#ifdef TCP_OFFLOAD 85#include <netinet/tcp_offload.h> 86#endif 87 88#ifdef IPSEC 89#include <netipsec/ipsec.h> 90#endif /*IPSEC*/ 91 92#include <machine/in_cksum.h> 93 94#include <security/mac/mac_framework.h> 95 96VNET_DEFINE(int, path_mtu_discovery) = 1; 97SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 98 &VNET_NAME(path_mtu_discovery), 1, 99 "Enable Path MTU Discovery"); 100 101VNET_DEFINE(int, tcp_do_tso) = 1; 102#define V_tcp_do_tso VNET(tcp_do_tso) 103SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW, 104 &VNET_NAME(tcp_do_tso), 0, 105 "Enable TCP Segmentation Offload"); 106 107VNET_DEFINE(int, tcp_sendspace) = 1024*32; 108#define V_tcp_sendspace VNET(tcp_sendspace) 109SYSCTL_VNET_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_RW, 110 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 111 112VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 113#define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) 114SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 115 &VNET_NAME(tcp_do_autosndbuf), 0, 116 "Enable automatic send buffer sizing"); 117 118VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 119#define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) 120SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 121 &VNET_NAME(tcp_autosndbuf_inc), 0, 122 "Incrementor step size of automatic send buffer"); 123 124VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; 125#define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) 126SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 127 &VNET_NAME(tcp_autosndbuf_max), 0, 128 "Max size of automatic send buffer"); 129 130/* 131 * Make sure that either retransmit or persist timer is set for SYN, FIN and 132 * non-ACK. 133 */ 134#define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \ 135 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\ 136 tcp_timer_active((tp), TT_REXMT) || \ 137 tcp_timer_active((tp), TT_PERSIST), \ 138 ("neither rexmt nor persist timer is set")) 139 140static void inline hhook_run_tcp_est_out(struct tcpcb *tp, 141 struct tcphdr *th, struct tcpopt *to, 142 long len, int tso); 143static void inline cc_after_idle(struct tcpcb *tp); 144 145/* 146 * Wrapper for the TCP established output helper hook. 147 */ 148static void inline 149hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 150 struct tcpopt *to, long len, int tso) 151{ 152 struct tcp_hhook_data hhook_data; 153 154 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 155 hhook_data.tp = tp; 156 hhook_data.th = th; 157 hhook_data.to = to; 158 hhook_data.len = len; 159 hhook_data.tso = tso; 160 161 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 162 tp->osd); 163 } 164} 165 166/* 167 * CC wrapper hook functions 168 */ 169static void inline 170cc_after_idle(struct tcpcb *tp) 171{ 172 INP_WLOCK_ASSERT(tp->t_inpcb); 173 174 if (CC_ALGO(tp)->after_idle != NULL) 175 CC_ALGO(tp)->after_idle(tp->ccv); 176} 177 178/* 179 * Tcp output routine: figure out what should be sent and send it. 180 */ 181int 182tcp_output(struct tcpcb *tp) 183{ 184 struct socket *so = tp->t_inpcb->inp_socket; 185 long len, recwin, sendwin; 186 int off, flags, error = 0; /* Keep compiler happy */ 187 struct mbuf *m; 188 struct ip *ip = NULL; 189 struct ipovly *ipov = NULL; 190 struct tcphdr *th; 191 u_char opt[TCP_MAXOLEN]; 192 unsigned ipoptlen, optlen, hdrlen; 193#ifdef IPSEC 194 unsigned ipsec_optlen = 0; 195#endif 196 int idle, sendalot; 197 int sack_rxmit, sack_bytes_rxmt; 198 struct sackhole *p; 199 int tso, mtu; 200 struct tcpopt to; 201#if 0 202 int maxburst = TCP_MAXBURST; 203#endif 204#ifdef INET6 205 struct ip6_hdr *ip6 = NULL; 206 int isipv6; 207 208 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 209#endif 210 211 INP_WLOCK_ASSERT(tp->t_inpcb); 212 213#ifdef TCP_OFFLOAD 214 if (tp->t_flags & TF_TOE) 215 return (tcp_offload_output(tp)); 216#endif 217 218#ifdef TCP_RFC7413 219 /* 220 * For TFO connections in SYN_RECEIVED, only allow the initial 221 * SYN|ACK and those sent by the retransmit timer. 222 */ 223 if ((tp->t_flags & TF_FASTOPEN) && 224 (tp->t_state == TCPS_SYN_RECEIVED) && 225 SEQ_GT(tp->snd_max, tp->snd_una) && /* inital SYN|ACK sent */ 226 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 227 return (0); 228#endif 229 /* 230 * Determine length of data that should be transmitted, 231 * and flags that will be used. 232 * If there is some data or critical controls (SYN, RST) 233 * to send, then transmit; otherwise, investigate further. 234 */ 235 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 236 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) 237 cc_after_idle(tp); 238 tp->t_flags &= ~TF_LASTIDLE; 239 if (idle) { 240 if (tp->t_flags & TF_MORETOCOME) { 241 tp->t_flags |= TF_LASTIDLE; 242 idle = 0; 243 } 244 } 245again: 246 /* 247 * If we've recently taken a timeout, snd_max will be greater than 248 * snd_nxt. There may be SACK information that allows us to avoid 249 * resending already delivered data. Adjust snd_nxt accordingly. 250 */ 251 if ((tp->t_flags & TF_SACK_PERMIT) && 252 SEQ_LT(tp->snd_nxt, tp->snd_max)) 253 tcp_sack_adjust(tp); 254 sendalot = 0; 255 tso = 0; 256 mtu = 0; 257 off = tp->snd_nxt - tp->snd_una; 258 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 259 260 flags = tcp_outflags[tp->t_state]; 261 /* 262 * Send any SACK-generated retransmissions. If we're explicitly trying 263 * to send out new data (when sendalot is 1), bypass this function. 264 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 265 * we're replacing a (future) new transmission with a retransmission 266 * now, and we previously incremented snd_cwnd in tcp_input(). 267 */ 268 /* 269 * Still in sack recovery , reset rxmit flag to zero. 270 */ 271 sack_rxmit = 0; 272 sack_bytes_rxmt = 0; 273 len = 0; 274 p = NULL; 275 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 276 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 277 long cwin; 278 279 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 280 if (cwin < 0) 281 cwin = 0; 282 /* Do not retransmit SACK segments beyond snd_recover */ 283 if (SEQ_GT(p->end, tp->snd_recover)) { 284 /* 285 * (At least) part of sack hole extends beyond 286 * snd_recover. Check to see if we can rexmit data 287 * for this hole. 288 */ 289 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 290 /* 291 * Can't rexmit any more data for this hole. 292 * That data will be rexmitted in the next 293 * sack recovery episode, when snd_recover 294 * moves past p->rxmit. 295 */ 296 p = NULL; 297 goto after_sack_rexmit; 298 } else 299 /* Can rexmit part of the current hole */ 300 len = ((long)ulmin(cwin, 301 tp->snd_recover - p->rxmit)); 302 } else 303 len = ((long)ulmin(cwin, p->end - p->rxmit)); 304 off = p->rxmit - tp->snd_una; 305 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 306 __func__, off)); 307 if (len > 0) { 308 sack_rxmit = 1; 309 sendalot = 1; 310 TCPSTAT_INC(tcps_sack_rexmits); 311 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 312 min(len, tp->t_maxseg)); 313 } 314 } 315after_sack_rexmit: 316 /* 317 * Get standard flags, and add SYN or FIN if requested by 'hidden' 318 * state flags. 319 */ 320 if (tp->t_flags & TF_NEEDFIN) 321 flags |= TH_FIN; 322 if (tp->t_flags & TF_NEEDSYN) 323 flags |= TH_SYN; 324 325 SOCKBUF_LOCK(&so->so_snd); 326 /* 327 * If in persist timeout with window of 0, send 1 byte. 328 * Otherwise, if window is small but nonzero 329 * and timer expired, we will send what we can 330 * and go to transmit state. 331 */ 332 if (tp->t_flags & TF_FORCEDATA) { 333 if (sendwin == 0) { 334 /* 335 * If we still have some data to send, then 336 * clear the FIN bit. Usually this would 337 * happen below when it realizes that we 338 * aren't sending all the data. However, 339 * if we have exactly 1 byte of unsent data, 340 * then it won't clear the FIN bit below, 341 * and if we are in persist state, we wind 342 * up sending the packet without recording 343 * that we sent the FIN bit. 344 * 345 * We can't just blindly clear the FIN bit, 346 * because if we don't have any more data 347 * to send then the probe will be the FIN 348 * itself. 349 */ 350 if (off < so->so_snd.sb_cc) 351 flags &= ~TH_FIN; 352 sendwin = 1; 353 } else { 354 tcp_timer_activate(tp, TT_PERSIST, 0); 355 tp->t_rxtshift = 0; 356 } 357 } 358 359 /* 360 * If snd_nxt == snd_max and we have transmitted a FIN, the 361 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 362 * a negative length. This can also occur when TCP opens up 363 * its congestion window while receiving additional duplicate 364 * acks after fast-retransmit because TCP will reset snd_nxt 365 * to snd_max after the fast-retransmit. 366 * 367 * In the normal retransmit-FIN-only case, however, snd_nxt will 368 * be set to snd_una, the offset will be 0, and the length may 369 * wind up 0. 370 * 371 * If sack_rxmit is true we are retransmitting from the scoreboard 372 * in which case len is already set. 373 */ 374 if (sack_rxmit == 0) { 375 if (sack_bytes_rxmt == 0) 376 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off); 377 else { 378 long cwin; 379 380 /* 381 * We are inside of a SACK recovery episode and are 382 * sending new data, having retransmitted all the 383 * data possible in the scoreboard. 384 */ 385 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 386 - off); 387 /* 388 * Don't remove this (len > 0) check ! 389 * We explicitly check for len > 0 here (although it 390 * isn't really necessary), to work around a gcc 391 * optimization issue - to force gcc to compute 392 * len above. Without this check, the computation 393 * of len is bungled by the optimizer. 394 */ 395 if (len > 0) { 396 cwin = tp->snd_cwnd - 397 (tp->snd_nxt - tp->sack_newdata) - 398 sack_bytes_rxmt; 399 if (cwin < 0) 400 cwin = 0; 401 len = lmin(len, cwin); 402 } 403 } 404 } 405 406 /* 407 * Lop off SYN bit if it has already been sent. However, if this 408 * is SYN-SENT state and if segment contains data and if we don't 409 * know that foreign host supports TAO, suppress sending segment. 410 */ 411 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 412 if (tp->t_state != TCPS_SYN_RECEIVED) 413 flags &= ~TH_SYN; 414#ifdef TCP_RFC7413 415 /* 416 * When sending additional segments following a TFO SYN|ACK, 417 * do not include the SYN bit. 418 */ 419 if ((tp->t_flags & TF_FASTOPEN) && 420 (tp->t_state == TCPS_SYN_RECEIVED)) 421 flags &= ~TH_SYN; 422#endif 423 off--, len++; 424 } 425 426 /* 427 * Be careful not to send data and/or FIN on SYN segments. 428 * This measure is needed to prevent interoperability problems 429 * with not fully conformant TCP implementations. 430 */ 431 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 432 len = 0; 433 flags &= ~TH_FIN; 434 } 435 436#ifdef TCP_RFC7413 437 /* 438 * When retransmitting SYN|ACK on a passively-created TFO socket, 439 * don't include data, as the presence of data may have caused the 440 * original SYN|ACK to have been dropped by a middlebox. 441 */ 442 if ((tp->t_flags & TF_FASTOPEN) && 443 (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) || 444 (flags & TH_RST))) 445 len = 0; 446#endif 447 if (len <= 0) { 448 /* 449 * If FIN has been sent but not acked, 450 * but we haven't been called to retransmit, 451 * len will be < 0. Otherwise, window shrank 452 * after we sent into it. If window shrank to 0, 453 * cancel pending retransmit, pull snd_nxt back 454 * to (closed) window, and set the persist timer 455 * if it isn't already going. If the window didn't 456 * close completely, just wait for an ACK. 457 * 458 * We also do a general check here to ensure that 459 * we will set the persist timer when we have data 460 * to send, but a 0-byte window. This makes sure 461 * the persist timer is set even if the packet 462 * hits one of the "goto send" lines below. 463 */ 464 len = 0; 465 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 466 (off < (int) so->so_snd.sb_cc)) { 467 tcp_timer_activate(tp, TT_REXMT, 0); 468 tp->t_rxtshift = 0; 469 tp->snd_nxt = tp->snd_una; 470 if (!tcp_timer_active(tp, TT_PERSIST)) 471 tcp_setpersist(tp); 472 } 473 } 474 475 /* len will be >= 0 after this point. */ 476 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 477 478 /* 479 * Automatic sizing of send socket buffer. Often the send buffer 480 * size is not optimally adjusted to the actual network conditions 481 * at hand (delay bandwidth product). Setting the buffer size too 482 * small limits throughput on links with high bandwidth and high 483 * delay (eg. trans-continental/oceanic links). Setting the 484 * buffer size too big consumes too much real kernel memory, 485 * especially with many connections on busy servers. 486 * 487 * The criteria to step up the send buffer one notch are: 488 * 1. receive window of remote host is larger than send buffer 489 * (with a fudge factor of 5/4th); 490 * 2. send buffer is filled to 7/8th with data (so we actually 491 * have data to make use of it); 492 * 3. send buffer fill has not hit maximal automatic size; 493 * 4. our send window (slow start and cogestion controlled) is 494 * larger than sent but unacknowledged data in send buffer. 495 * 496 * The remote host receive window scaling factor may limit the 497 * growing of the send buffer before it reaches its allowed 498 * maximum. 499 * 500 * It scales directly with slow start or congestion window 501 * and does at most one step per received ACK. This fast 502 * scaling has the drawback of growing the send buffer beyond 503 * what is strictly necessary to make full use of a given 504 * delay*bandwith product. However testing has shown this not 505 * to be much of an problem. At worst we are trading wasting 506 * of available bandwith (the non-use of it) for wasting some 507 * socket buffer memory. 508 * 509 * TODO: Shrink send buffer during idle periods together 510 * with congestion window. Requires another timer. Has to 511 * wait for upcoming tcp timer rewrite. 512 */ 513 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 514 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 515 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 516 so->so_snd.sb_cc < V_tcp_autosndbuf_max && 517 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 518 if (!sbreserve_locked(&so->so_snd, 519 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 520 V_tcp_autosndbuf_max), so, curthread)) 521 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 522 } 523 } 524 525 /* 526 * Decide if we can use TCP Segmentation Offloading (if supported by 527 * hardware). 528 * 529 * TSO may only be used if we are in a pure bulk sending state. The 530 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 531 * IP options prevent using TSO. With TSO the TCP header is the same 532 * (except for the sequence number) for all generated packets. This 533 * makes it impossible to transmit any options which vary per generated 534 * segment or packet. 535 */ 536#ifdef IPSEC 537 /* 538 * Pre-calculate here as we save another lookup into the darknesses 539 * of IPsec that way and can actually decide if TSO is ok. 540 */ 541 ipsec_optlen = ipsec_hdrsiz_tcp(tp); 542#endif 543 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 544 ((tp->t_flags & TF_SIGNATURE) == 0) && 545 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 546#ifdef IPSEC 547 ipsec_optlen == 0 && 548#endif 549 tp->t_inpcb->inp_options == NULL && 550 tp->t_inpcb->in6p_options == NULL) 551 tso = 1; 552 553 if (sack_rxmit) { 554 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 555 flags &= ~TH_FIN; 556 } else { 557 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 558 flags &= ~TH_FIN; 559 } 560 561 recwin = sbspace(&so->so_rcv); 562 563 /* 564 * Sender silly window avoidance. We transmit under the following 565 * conditions when len is non-zero: 566 * 567 * - We have a full segment (or more with TSO) 568 * - This is the last buffer in a write()/send() and we are 569 * either idle or running NODELAY 570 * - we've timed out (e.g. persist timer) 571 * - we have more then 1/2 the maximum send window's worth of 572 * data (receiver may be limited the window size) 573 * - we need to retransmit 574 */ 575 if (len) { 576 if (len >= tp->t_maxseg) 577 goto send; 578 /* 579 * NOTE! on localhost connections an 'ack' from the remote 580 * end may occur synchronously with the output and cause 581 * us to flush a buffer queued with moretocome. XXX 582 * 583 * note: the len + off check is almost certainly unnecessary. 584 */ 585 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 586 (idle || (tp->t_flags & TF_NODELAY)) && 587 len + off >= so->so_snd.sb_cc && 588 (tp->t_flags & TF_NOPUSH) == 0) { 589 goto send; 590 } 591 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 592 goto send; 593 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 594 goto send; 595 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 596 goto send; 597 if (sack_rxmit) 598 goto send; 599 } 600 601 /* 602 * Sending of standalone window updates. 603 * 604 * Window updates are important when we close our window due to a 605 * full socket buffer and are opening it again after the application 606 * reads data from it. Once the window has opened again and the 607 * remote end starts to send again the ACK clock takes over and 608 * provides the most current window information. 609 * 610 * We must avoid the silly window syndrome whereas every read 611 * from the receive buffer, no matter how small, causes a window 612 * update to be sent. We also should avoid sending a flurry of 613 * window updates when the socket buffer had queued a lot of data 614 * and the application is doing small reads. 615 * 616 * Prevent a flurry of pointless window updates by only sending 617 * an update when we can increase the advertized window by more 618 * than 1/4th of the socket buffer capacity. When the buffer is 619 * getting full or is very small be more aggressive and send an 620 * update whenever we can increase by two mss sized segments. 621 * In all other situations the ACK's to new incoming data will 622 * carry further window increases. 623 * 624 * Don't send an independent window update if a delayed 625 * ACK is pending (it will get piggy-backed on it) or the 626 * remote side already has done a half-close and won't send 627 * more data. Skip this if the connection is in T/TCP 628 * half-open state. 629 */ 630 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 631 !(tp->t_flags & TF_DELACK) && 632 !TCPS_HAVERCVDFIN(tp->t_state)) { 633 /* 634 * "adv" is the amount we could increase the window, 635 * taking into account that we are limited by 636 * TCP_MAXWIN << tp->rcv_scale. 637 */ 638 long adv; 639 int oldwin; 640 641 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 642 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 643 oldwin = (tp->rcv_adv - tp->rcv_nxt); 644 adv -= oldwin; 645 } else 646 oldwin = 0; 647 648 /* 649 * If the new window size ends up being the same as the old 650 * size when it is scaled, then don't force a window update. 651 */ 652 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 653 goto dontupdate; 654 655 if (adv >= (long)(2 * tp->t_maxseg) && 656 (adv >= (long)(so->so_rcv.sb_hiwat / 4) || 657 recwin <= (long)(so->so_rcv.sb_hiwat / 8) || 658 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 659 goto send; 660 } 661dontupdate: 662 663 /* 664 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 665 * is also a catch-all for the retransmit timer timeout case. 666 */ 667 if (tp->t_flags & TF_ACKNOW) 668 goto send; 669 if ((flags & TH_RST) || 670 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 671 goto send; 672 if (SEQ_GT(tp->snd_up, tp->snd_una)) 673 goto send; 674 /* 675 * If our state indicates that FIN should be sent 676 * and we have not yet done so, then we need to send. 677 */ 678 if (flags & TH_FIN && 679 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 680 goto send; 681 /* 682 * In SACK, it is possible for tcp_output to fail to send a segment 683 * after the retransmission timer has been turned off. Make sure 684 * that the retransmission timer is set. 685 */ 686 if ((tp->t_flags & TF_SACK_PERMIT) && 687 SEQ_GT(tp->snd_max, tp->snd_una) && 688 !tcp_timer_active(tp, TT_REXMT) && 689 !tcp_timer_active(tp, TT_PERSIST)) { 690 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 691 goto just_return; 692 } 693 /* 694 * TCP window updates are not reliable, rather a polling protocol 695 * using ``persist'' packets is used to insure receipt of window 696 * updates. The three ``states'' for the output side are: 697 * idle not doing retransmits or persists 698 * persisting to move a small or zero window 699 * (re)transmitting and thereby not persisting 700 * 701 * tcp_timer_active(tp, TT_PERSIST) 702 * is true when we are in persist state. 703 * (tp->t_flags & TF_FORCEDATA) 704 * is set when we are called to send a persist packet. 705 * tcp_timer_active(tp, TT_REXMT) 706 * is set when we are retransmitting 707 * The output side is idle when both timers are zero. 708 * 709 * If send window is too small, there is data to transmit, and no 710 * retransmit or persist is pending, then go to persist state. 711 * If nothing happens soon, send when timer expires: 712 * if window is nonzero, transmit what we can, 713 * otherwise force out a byte. 714 */ 715 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) && 716 !tcp_timer_active(tp, TT_PERSIST)) { 717 tp->t_rxtshift = 0; 718 tcp_setpersist(tp); 719 } 720 721 /* 722 * No reason to send a segment, just return. 723 */ 724just_return: 725 SOCKBUF_UNLOCK(&so->so_snd); 726 return (0); 727 728send: 729 SOCKBUF_LOCK_ASSERT(&so->so_snd); 730 if (len > 0) { 731 if (len >= tp->t_maxseg) 732 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 733 else 734 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 735 } 736 /* 737 * Before ESTABLISHED, force sending of initial options 738 * unless TCP set not to do any options. 739 * NOTE: we assume that the IP/TCP header plus TCP options 740 * always fit in a single mbuf, leaving room for a maximum 741 * link header, i.e. 742 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 743 */ 744 optlen = 0; 745#ifdef INET6 746 if (isipv6) 747 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 748 else 749#endif 750 hdrlen = sizeof (struct tcpiphdr); 751 752 /* 753 * Compute options for segment. 754 * We only have to care about SYN and established connection 755 * segments. Options for SYN-ACK segments are handled in TCP 756 * syncache. 757 */ 758 to.to_flags = 0; 759 if ((tp->t_flags & TF_NOOPT) == 0) { 760 /* Maximum segment size. */ 761 if (flags & TH_SYN) { 762 tp->snd_nxt = tp->iss; 763 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 764 to.to_flags |= TOF_MSS; 765#ifdef TCP_RFC7413 766 /* 767 * Only include the TFO option on the first 768 * transmission of the SYN|ACK on a 769 * passively-created TFO socket, as the presence of 770 * the TFO option may have caused the original 771 * SYN|ACK to have been dropped by a middlebox. 772 */ 773 if ((tp->t_flags & TF_FASTOPEN) && 774 (tp->t_state == TCPS_SYN_RECEIVED) && 775 (tp->t_rxtshift == 0)) { 776 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 777 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie; 778 to.to_flags |= TOF_FASTOPEN; 779 } 780#endif 781 } 782 /* Window scaling. */ 783 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 784 to.to_wscale = tp->request_r_scale; 785 to.to_flags |= TOF_SCALE; 786 } 787 /* Timestamps. */ 788 if ((tp->t_flags & TF_RCVD_TSTMP) || 789 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 790 to.to_tsval = tcp_ts_getticks() + tp->ts_offset; 791 to.to_tsecr = tp->ts_recent; 792 to.to_flags |= TOF_TS; 793 /* Set receive buffer autosizing timestamp. */ 794 if (tp->rfbuf_ts == 0 && 795 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 796 tp->rfbuf_ts = tcp_ts_getticks(); 797 } 798 /* Selective ACK's. */ 799 if (tp->t_flags & TF_SACK_PERMIT) { 800 if (flags & TH_SYN) 801 to.to_flags |= TOF_SACKPERM; 802 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 803 (tp->t_flags & TF_SACK_PERMIT) && 804 tp->rcv_numsacks > 0) { 805 to.to_flags |= TOF_SACK; 806 to.to_nsacks = tp->rcv_numsacks; 807 to.to_sacks = (u_char *)tp->sackblks; 808 } 809 } 810#ifdef TCP_SIGNATURE 811 /* TCP-MD5 (RFC2385). */ 812 if (tp->t_flags & TF_SIGNATURE) 813 to.to_flags |= TOF_SIGNATURE; 814#endif /* TCP_SIGNATURE */ 815 816 /* Processing the options. */ 817 hdrlen += optlen = tcp_addoptions(&to, opt); 818 } 819 820#ifdef INET6 821 if (isipv6) 822 ipoptlen = ip6_optlen(tp->t_inpcb); 823 else 824#endif 825 if (tp->t_inpcb->inp_options) 826 ipoptlen = tp->t_inpcb->inp_options->m_len - 827 offsetof(struct ipoption, ipopt_list); 828 else 829 ipoptlen = 0; 830#ifdef IPSEC 831 ipoptlen += ipsec_optlen; 832#endif 833 834 /* 835 * Adjust data length if insertion of options will 836 * bump the packet length beyond the t_maxopd length. 837 * Clear the FIN bit because we cut off the tail of 838 * the segment. 839 */ 840 if (len + optlen + ipoptlen > tp->t_maxopd) { 841 flags &= ~TH_FIN; 842 843 if (tso) { 844 u_int if_hw_tsomax; 845 u_int if_hw_tsomaxsegcount; 846 u_int if_hw_tsomaxsegsize; 847 struct mbuf *mb; 848 u_int moff; 849 int max_len; 850 851 /* extract TSO information */ 852 if_hw_tsomax = tp->t_tsomax; 853 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 854 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 855 856 /* 857 * Limit a TSO burst to prevent it from 858 * overflowing or exceeding the maximum length 859 * allowed by the network interface: 860 */ 861 KASSERT(ipoptlen == 0, 862 ("%s: TSO can't do IP options", __func__)); 863 864 /* 865 * Check if we should limit by maximum payload 866 * length: 867 */ 868 if (if_hw_tsomax != 0) { 869 /* compute maximum TSO length */ 870 max_len = (if_hw_tsomax - hdrlen - 871 max_linkhdr); 872 if (max_len <= 0) { 873 len = 0; 874 } else if (len > max_len) { 875 sendalot = 1; 876 len = max_len; 877 } 878 } 879 880 /* 881 * Check if we should limit by maximum segment 882 * size and count: 883 */ 884 if (if_hw_tsomaxsegcount != 0 && 885 if_hw_tsomaxsegsize != 0) { 886 /* 887 * Subtract one segment for the LINK 888 * and TCP/IP headers mbuf that will 889 * be prepended to this mbuf chain 890 * after the code in this section 891 * limits the number of mbufs in the 892 * chain to if_hw_tsomaxsegcount. 893 */ 894 if_hw_tsomaxsegcount -= 1; 895 max_len = 0; 896 mb = sbsndmbuf(&so->so_snd, off, &moff); 897 898 while (mb != NULL && max_len < len) { 899 u_int mlen; 900 u_int frags; 901 902 /* 903 * Get length of mbuf fragment 904 * and how many hardware frags, 905 * rounded up, it would use: 906 */ 907 mlen = (mb->m_len - moff); 908 frags = howmany(mlen, 909 if_hw_tsomaxsegsize); 910 911 /* Handle special case: Zero Length Mbuf */ 912 if (frags == 0) 913 frags = 1; 914 915 /* 916 * Check if the fragment limit 917 * will be reached or exceeded: 918 */ 919 if (frags >= if_hw_tsomaxsegcount) { 920 max_len += min(mlen, 921 if_hw_tsomaxsegcount * 922 if_hw_tsomaxsegsize); 923 break; 924 } 925 max_len += mlen; 926 if_hw_tsomaxsegcount -= frags; 927 moff = 0; 928 mb = mb->m_next; 929 } 930 if (max_len <= 0) { 931 len = 0; 932 } else if (len > max_len) { 933 sendalot = 1; 934 len = max_len; 935 } 936 } 937 938 /* 939 * Prevent the last segment from being 940 * fractional unless the send sockbuf can be 941 * emptied: 942 */ 943 max_len = (tp->t_maxopd - optlen); 944 if ((off + len) < so->so_snd.sb_cc) { 945 moff = len % max_len; 946 if (moff != 0) { 947 len -= moff; 948 sendalot = 1; 949 } 950 } 951 952 /* 953 * In case there are too many small fragments 954 * don't use TSO: 955 */ 956 if (len <= max_len) { 957 len = max_len; 958 sendalot = 1; 959 tso = 0; 960 } 961 962 /* 963 * Send the FIN in a separate segment 964 * after the bulk sending is done. 965 * We don't trust the TSO implementations 966 * to clear the FIN flag on all but the 967 * last segment. 968 */ 969 if (tp->t_flags & TF_NEEDFIN) 970 sendalot = 1; 971 972 } else { 973 len = tp->t_maxopd - optlen - ipoptlen; 974 sendalot = 1; 975 } 976 } else 977 tso = 0; 978 979 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 980 ("%s: len > IP_MAXPACKET", __func__)); 981 982/*#ifdef DIAGNOSTIC*/ 983#ifdef INET6 984 if (max_linkhdr + hdrlen > MCLBYTES) 985#else 986 if (max_linkhdr + hdrlen > MHLEN) 987#endif 988 panic("tcphdr too big"); 989/*#endif*/ 990 991 /* 992 * This KASSERT is here to catch edge cases at a well defined place. 993 * Before, those had triggered (random) panic conditions further down. 994 */ 995 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 996 997 /* 998 * Grab a header mbuf, attaching a copy of data to 999 * be transmitted, and initialize the header from 1000 * the template for sends on this connection. 1001 */ 1002 if (len) { 1003 struct mbuf *mb; 1004 u_int moff; 1005 1006 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 1007 TCPSTAT_INC(tcps_sndprobe); 1008 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 1009 tp->t_sndrexmitpack++; 1010 TCPSTAT_INC(tcps_sndrexmitpack); 1011 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 1012 } else { 1013 TCPSTAT_INC(tcps_sndpack); 1014 TCPSTAT_ADD(tcps_sndbyte, len); 1015 } 1016#ifdef INET6 1017 if (MHLEN < hdrlen + max_linkhdr) 1018 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1019 else 1020#endif 1021 m = m_gethdr(M_NOWAIT, MT_DATA); 1022 1023 if (m == NULL) { 1024 SOCKBUF_UNLOCK(&so->so_snd); 1025 error = ENOBUFS; 1026 sack_rxmit = 0; 1027 goto out; 1028 } 1029 1030 m->m_data += max_linkhdr; 1031 m->m_len = hdrlen; 1032 1033 /* 1034 * Start the m_copy functions from the closest mbuf 1035 * to the offset in the socket buffer chain. 1036 */ 1037 mb = sbsndptr(&so->so_snd, off, len, &moff); 1038 1039 if (len <= MHLEN - hdrlen - max_linkhdr) { 1040 m_copydata(mb, moff, (int)len, 1041 mtod(m, caddr_t) + hdrlen); 1042 m->m_len += len; 1043 } else { 1044 m->m_next = m_copy(mb, moff, (int)len); 1045 if (m->m_next == NULL) { 1046 SOCKBUF_UNLOCK(&so->so_snd); 1047 (void) m_free(m); 1048 error = ENOBUFS; 1049 sack_rxmit = 0; 1050 goto out; 1051 } 1052 } 1053 1054 /* 1055 * If we're sending everything we've got, set PUSH. 1056 * (This will keep happy those implementations which only 1057 * give data to the user when a buffer fills or 1058 * a PUSH comes in.) 1059 */ 1060 if ((off + len == so->so_snd.sb_cc) && !(flags & TH_SYN)) 1061 flags |= TH_PUSH; 1062 SOCKBUF_UNLOCK(&so->so_snd); 1063 } else { 1064 SOCKBUF_UNLOCK(&so->so_snd); 1065 if (tp->t_flags & TF_ACKNOW) 1066 TCPSTAT_INC(tcps_sndacks); 1067 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1068 TCPSTAT_INC(tcps_sndctrl); 1069 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1070 TCPSTAT_INC(tcps_sndurg); 1071 else 1072 TCPSTAT_INC(tcps_sndwinup); 1073 1074 m = m_gethdr(M_NOWAIT, MT_DATA); 1075 if (m == NULL) { 1076 error = ENOBUFS; 1077 sack_rxmit = 0; 1078 goto out; 1079 } 1080#ifdef INET6 1081 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1082 MHLEN >= hdrlen) { 1083 MH_ALIGN(m, hdrlen); 1084 } else 1085#endif 1086 m->m_data += max_linkhdr; 1087 m->m_len = hdrlen; 1088 } 1089 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1090 m->m_pkthdr.rcvif = (struct ifnet *)0; 1091#ifdef MAC 1092 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1093#endif 1094#ifdef INET6 1095 if (isipv6) { 1096 ip6 = mtod(m, struct ip6_hdr *); 1097 th = (struct tcphdr *)(ip6 + 1); 1098 tcpip_fillheaders(tp->t_inpcb, ip6, th); 1099 } else 1100#endif /* INET6 */ 1101 { 1102 ip = mtod(m, struct ip *); 1103 ipov = (struct ipovly *)ip; 1104 th = (struct tcphdr *)(ip + 1); 1105 tcpip_fillheaders(tp->t_inpcb, ip, th); 1106 } 1107 1108 /* 1109 * Fill in fields, remembering maximum advertised 1110 * window for use in delaying messages about window sizes. 1111 * If resending a FIN, be sure not to use a new sequence number. 1112 */ 1113 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1114 tp->snd_nxt == tp->snd_max) 1115 tp->snd_nxt--; 1116 /* 1117 * If we are starting a connection, send ECN setup 1118 * SYN packet. If we are on a retransmit, we may 1119 * resend those bits a number of times as per 1120 * RFC 3168. 1121 */ 1122 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { 1123 if (tp->t_rxtshift >= 1) { 1124 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 1125 flags |= TH_ECE|TH_CWR; 1126 } else 1127 flags |= TH_ECE|TH_CWR; 1128 } 1129 1130 if (tp->t_state == TCPS_ESTABLISHED && 1131 (tp->t_flags & TF_ECN_PERMIT)) { 1132 /* 1133 * If the peer has ECN, mark data packets with 1134 * ECN capable transmission (ECT). 1135 * Ignore pure ack packets, retransmissions and window probes. 1136 */ 1137 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1138 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 1139#ifdef INET6 1140 if (isipv6) 1141 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1142 else 1143#endif 1144 ip->ip_tos |= IPTOS_ECN_ECT0; 1145 TCPSTAT_INC(tcps_ecn_ect0); 1146 } 1147 1148 /* 1149 * Reply with proper ECN notifications. 1150 */ 1151 if (tp->t_flags & TF_ECN_SND_CWR) { 1152 flags |= TH_CWR; 1153 tp->t_flags &= ~TF_ECN_SND_CWR; 1154 } 1155 if (tp->t_flags & TF_ECN_SND_ECE) 1156 flags |= TH_ECE; 1157 } 1158 1159 /* 1160 * If we are doing retransmissions, then snd_nxt will 1161 * not reflect the first unsent octet. For ACK only 1162 * packets, we do not want the sequence number of the 1163 * retransmitted packet, we want the sequence number 1164 * of the next unsent octet. So, if there is no data 1165 * (and no SYN or FIN), use snd_max instead of snd_nxt 1166 * when filling in ti_seq. But if we are in persist 1167 * state, snd_max might reflect one byte beyond the 1168 * right edge of the window, so use snd_nxt in that 1169 * case, since we know we aren't doing a retransmission. 1170 * (retransmit and persist are mutually exclusive...) 1171 */ 1172 if (sack_rxmit == 0) { 1173 if (len || (flags & (TH_SYN|TH_FIN)) || 1174 tcp_timer_active(tp, TT_PERSIST)) 1175 th->th_seq = htonl(tp->snd_nxt); 1176 else 1177 th->th_seq = htonl(tp->snd_max); 1178 } else { 1179 th->th_seq = htonl(p->rxmit); 1180 p->rxmit += len; 1181 tp->sackhint.sack_bytes_rexmit += len; 1182 } 1183 th->th_ack = htonl(tp->rcv_nxt); 1184 if (optlen) { 1185 bcopy(opt, th + 1, optlen); 1186 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1187 } 1188 th->th_flags = flags; 1189 /* 1190 * Calculate receive window. Don't shrink window, 1191 * but avoid silly window syndrome. 1192 */ 1193 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 1194 recwin < (long)tp->t_maxseg) 1195 recwin = 0; 1196 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1197 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 1198 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 1199 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 1200 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 1201 1202 /* 1203 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1204 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1205 * case is handled in syncache. 1206 */ 1207 if (flags & TH_SYN) 1208 th->th_win = htons((u_short) 1209 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1210 else 1211 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1212 1213 /* 1214 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1215 * a 0 window. This may cause the remote transmitter to stall. This 1216 * flag tells soreceive() to disable delayed acknowledgements when 1217 * draining the buffer. This can occur if the receiver is attempting 1218 * to read more data than can be buffered prior to transmitting on 1219 * the connection. 1220 */ 1221 if (th->th_win == 0) { 1222 tp->t_sndzerowin++; 1223 tp->t_flags |= TF_RXWIN0SENT; 1224 } else 1225 tp->t_flags &= ~TF_RXWIN0SENT; 1226 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1227 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1228 th->th_flags |= TH_URG; 1229 } else 1230 /* 1231 * If no urgent pointer to send, then we pull 1232 * the urgent pointer to the left edge of the send window 1233 * so that it doesn't drift into the send window on sequence 1234 * number wraparound. 1235 */ 1236 tp->snd_up = tp->snd_una; /* drag it along */ 1237 1238#ifdef TCP_SIGNATURE 1239 if (to.to_flags & TOF_SIGNATURE) { 1240 int sigoff = to.to_signature - opt; 1241 tcp_signature_compute(m, 0, len, optlen, 1242 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1243 } 1244#endif 1245 1246 /* 1247 * Put TCP length in extended header, and then 1248 * checksum extended header and data. 1249 */ 1250 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1251 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1252#ifdef INET6 1253 if (isipv6) { 1254 /* 1255 * There is no need to fill in ip6_plen right now. 1256 * It will be filled later by ip6_output. 1257 */ 1258 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1259 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1260 optlen + len, IPPROTO_TCP, 0); 1261 } 1262#endif 1263#if defined(INET6) && defined(INET) 1264 else 1265#endif 1266#ifdef INET 1267 { 1268 m->m_pkthdr.csum_flags = CSUM_TCP; 1269 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1270 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1271 1272 /* IP version must be set here for ipv4/ipv6 checking later */ 1273 KASSERT(ip->ip_v == IPVERSION, 1274 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1275 } 1276#endif 1277 1278 /* 1279 * Enable TSO and specify the size of the segments. 1280 * The TCP pseudo header checksum is always provided. 1281 * XXX: Fixme: This is currently not the case for IPv6. 1282 */ 1283 if (tso) { 1284 KASSERT(len > tp->t_maxopd - optlen, 1285 ("%s: len <= tso_segsz", __func__)); 1286 m->m_pkthdr.csum_flags |= CSUM_TSO; 1287 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 1288 } 1289 1290#ifdef IPSEC 1291 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1292 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u", 1293 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1294#else 1295 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1296 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u", 1297 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1298#endif 1299 1300 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1301 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1302 1303#ifdef TCPDEBUG 1304 /* 1305 * Trace. 1306 */ 1307 if (so->so_options & SO_DEBUG) { 1308 u_short save = 0; 1309#ifdef INET6 1310 if (!isipv6) 1311#endif 1312 { 1313 save = ipov->ih_len; 1314 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1315 } 1316 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1317#ifdef INET6 1318 if (!isipv6) 1319#endif 1320 ipov->ih_len = save; 1321 } 1322#endif /* TCPDEBUG */ 1323 1324 /* 1325 * Fill in IP length and desired time to live and 1326 * send to IP level. There should be a better way 1327 * to handle ttl and tos; we could keep them in 1328 * the template, but need a way to checksum without them. 1329 */ 1330 /* 1331 * m->m_pkthdr.len should have been set before cksum calcuration, 1332 * because in6_cksum() need it. 1333 */ 1334#ifdef INET6 1335 if (isipv6) { 1336 struct route_in6 ro; 1337 1338 bzero(&ro, sizeof(ro)); 1339 /* 1340 * we separately set hoplimit for every segment, since the 1341 * user might want to change the value via setsockopt. 1342 * Also, desired default hop limit might be changed via 1343 * Neighbor Discovery. 1344 */ 1345 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1346 1347 /* 1348 * Set the packet size here for the benefit of DTrace probes. 1349 * ip6_output() will set it properly; it's supposed to include 1350 * the option header lengths as well. 1351 */ 1352 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1353 1354 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) 1355 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1356 else 1357 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1358 1359 if (tp->t_state == TCPS_SYN_SENT) 1360 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1361 1362 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1363 1364 /* TODO: IPv6 IP6TOS_ECT bit on */ 1365 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro, 1366 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1367 NULL, NULL, tp->t_inpcb); 1368 1369 if (error == EMSGSIZE && ro.ro_rt != NULL) 1370 mtu = ro.ro_rt->rt_mtu; 1371 RO_RTFREE(&ro); 1372 } 1373#endif /* INET6 */ 1374#if defined(INET) && defined(INET6) 1375 else 1376#endif 1377#ifdef INET 1378 { 1379 struct route ro; 1380 1381 bzero(&ro, sizeof(ro)); 1382 ip->ip_len = htons(m->m_pkthdr.len); 1383#ifdef INET6 1384 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1385 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1386#endif /* INET6 */ 1387 /* 1388 * If we do path MTU discovery, then we set DF on every packet. 1389 * This might not be the best thing to do according to RFC3390 1390 * Section 2. However the tcp hostcache migitates the problem 1391 * so it affects only the first tcp connection with a host. 1392 * 1393 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1394 */ 1395 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) { 1396 ip->ip_off |= htons(IP_DF); 1397 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1398 } else { 1399 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1400 } 1401 1402 if (tp->t_state == TCPS_SYN_SENT) 1403 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1404 1405 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1406 1407 error = ip_output(m, tp->t_inpcb->inp_options, &ro, 1408 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1409 tp->t_inpcb); 1410 1411 if (error == EMSGSIZE && ro.ro_rt != NULL) 1412 mtu = ro.ro_rt->rt_mtu; 1413 RO_RTFREE(&ro); 1414 } 1415#endif /* INET */ 1416 1417out: 1418 /* 1419 * In transmit state, time the transmission and arrange for 1420 * the retransmit. In persist state, just set snd_max. 1421 */ 1422 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1423 !tcp_timer_active(tp, TT_PERSIST)) { 1424 tcp_seq startseq = tp->snd_nxt; 1425 1426 /* 1427 * Advance snd_nxt over sequence space of this segment. 1428 */ 1429 if (flags & (TH_SYN|TH_FIN)) { 1430 if (flags & TH_SYN) 1431 tp->snd_nxt++; 1432 if (flags & TH_FIN) { 1433 tp->snd_nxt++; 1434 tp->t_flags |= TF_SENTFIN; 1435 } 1436 } 1437 if (sack_rxmit) 1438 goto timer; 1439 tp->snd_nxt += len; 1440 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1441 tp->snd_max = tp->snd_nxt; 1442 /* 1443 * Time this transmission if not a retransmission and 1444 * not currently timing anything. 1445 */ 1446 if (tp->t_rtttime == 0) { 1447 tp->t_rtttime = ticks; 1448 tp->t_rtseq = startseq; 1449 TCPSTAT_INC(tcps_segstimed); 1450 } 1451 } 1452 1453 /* 1454 * Set retransmit timer if not currently set, 1455 * and not doing a pure ack or a keep-alive probe. 1456 * Initial value for retransmit timer is smoothed 1457 * round-trip time + 2 * round-trip time variance. 1458 * Initialize shift counter which is used for backoff 1459 * of retransmit time. 1460 */ 1461timer: 1462 if (!tcp_timer_active(tp, TT_REXMT) && 1463 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1464 (tp->snd_nxt != tp->snd_una))) { 1465 if (tcp_timer_active(tp, TT_PERSIST)) { 1466 tcp_timer_activate(tp, TT_PERSIST, 0); 1467 tp->t_rxtshift = 0; 1468 } 1469 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1470 } else if (len == 0 && so->so_snd.sb_cc && 1471 !tcp_timer_active(tp, TT_REXMT) && 1472 !tcp_timer_active(tp, TT_PERSIST)) { 1473 /* 1474 * Avoid a situation where we do not set persist timer 1475 * after a zero window condition. For example: 1476 * 1) A -> B: packet with enough data to fill the window 1477 * 2) B -> A: ACK for #1 + new data (0 window 1478 * advertisement) 1479 * 3) A -> B: ACK for #2, 0 len packet 1480 * 1481 * In this case, A will not activate the persist timer, 1482 * because it chose to send a packet. Unless tcp_output 1483 * is called for some other reason (delayed ack timer, 1484 * another input packet from B, socket syscall), A will 1485 * not send zero window probes. 1486 * 1487 * So, if you send a 0-length packet, but there is data 1488 * in the socket buffer, and neither the rexmt or 1489 * persist timer is already set, then activate the 1490 * persist timer. 1491 */ 1492 tp->t_rxtshift = 0; 1493 tcp_setpersist(tp); 1494 } 1495 } else { 1496 /* 1497 * Persist case, update snd_max but since we are in 1498 * persist mode (no window) we do not update snd_nxt. 1499 */ 1500 int xlen = len; 1501 if (flags & TH_SYN) 1502 ++xlen; 1503 if (flags & TH_FIN) { 1504 ++xlen; 1505 tp->t_flags |= TF_SENTFIN; 1506 } 1507 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1508 tp->snd_max = tp->snd_nxt + len; 1509 } 1510 1511 if (error) { 1512 1513 /* 1514 * We know that the packet was lost, so back out the 1515 * sequence number advance, if any. 1516 * 1517 * If the error is EPERM the packet got blocked by the 1518 * local firewall. Normally we should terminate the 1519 * connection but the blocking may have been spurious 1520 * due to a firewall reconfiguration cycle. So we treat 1521 * it like a packet loss and let the retransmit timer and 1522 * timeouts do their work over time. 1523 * XXX: It is a POLA question whether calling tcp_drop right 1524 * away would be the really correct behavior instead. 1525 */ 1526 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1527 !tcp_timer_active(tp, TT_PERSIST)) && 1528 ((flags & TH_SYN) == 0) && 1529 (error != EPERM)) { 1530 if (sack_rxmit) { 1531 p->rxmit -= len; 1532 tp->sackhint.sack_bytes_rexmit -= len; 1533 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1534 ("sackhint bytes rtx >= 0")); 1535 } else 1536 tp->snd_nxt -= len; 1537 } 1538 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1539 switch (error) { 1540 case EPERM: 1541 tp->t_softerror = error; 1542 return (error); 1543 case ENOBUFS: 1544 TCP_XMIT_TIMER_ASSERT(tp, len, flags); 1545 tp->snd_cwnd = tp->t_maxseg; 1546 return (0); 1547 case EMSGSIZE: 1548 /* 1549 * For some reason the interface we used initially 1550 * to send segments changed to another or lowered 1551 * its MTU. 1552 * If TSO was active we either got an interface 1553 * without TSO capabilits or TSO was turned off. 1554 * If we obtained mtu from ip_output() then update 1555 * it and try again. 1556 */ 1557 if (tso) 1558 tp->t_flags &= ~TF_TSO; 1559 if (mtu != 0) { 1560 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1561 goto again; 1562 } 1563 return (error); 1564 case EHOSTDOWN: 1565 case EHOSTUNREACH: 1566 case ENETDOWN: 1567 case ENETUNREACH: 1568 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1569 tp->t_softerror = error; 1570 return (0); 1571 } 1572 /* FALLTHROUGH */ 1573 default: 1574 return (error); 1575 } 1576 } 1577 TCPSTAT_INC(tcps_sndtotal); 1578 1579 /* 1580 * Data sent (as far as we can tell). 1581 * If this advertises a larger window than any other segment, 1582 * then remember the size of the advertised window. 1583 * Any pending ACK has now been sent. 1584 */ 1585 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1586 tp->rcv_adv = tp->rcv_nxt + recwin; 1587 tp->last_ack_sent = tp->rcv_nxt; 1588 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1589 if (tcp_timer_active(tp, TT_DELACK)) 1590 tcp_timer_activate(tp, TT_DELACK, 0); 1591#if 0 1592 /* 1593 * This completely breaks TCP if newreno is turned on. What happens 1594 * is that if delayed-acks are turned on on the receiver, this code 1595 * on the transmitter effectively destroys the TCP window, forcing 1596 * it to four packets (1.5Kx4 = 6K window). 1597 */ 1598 if (sendalot && --maxburst) 1599 goto again; 1600#endif 1601 if (sendalot) 1602 goto again; 1603 return (0); 1604} 1605 1606void 1607tcp_setpersist(struct tcpcb *tp) 1608{ 1609 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1610 int tt; 1611 1612 tp->t_flags &= ~TF_PREVVALID; 1613 if (tcp_timer_active(tp, TT_REXMT)) 1614 panic("tcp_setpersist: retransmit pending"); 1615 /* 1616 * Start/restart persistance timer. 1617 */ 1618 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1619 tcp_persmin, tcp_persmax); 1620 tcp_timer_activate(tp, TT_PERSIST, tt); 1621 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1622 tp->t_rxtshift++; 1623} 1624 1625/* 1626 * Insert TCP options according to the supplied parameters to the place 1627 * optp in a consistent way. Can handle unaligned destinations. 1628 * 1629 * The order of the option processing is crucial for optimal packing and 1630 * alignment for the scarce option space. 1631 * 1632 * The optimal order for a SYN/SYN-ACK segment is: 1633 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1634 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1635 * 1636 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1637 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1638 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1639 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1640 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1641 */ 1642int 1643tcp_addoptions(struct tcpopt *to, u_char *optp) 1644{ 1645 u_int mask, optlen = 0; 1646 1647 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1648 if ((to->to_flags & mask) != mask) 1649 continue; 1650 if (optlen == TCP_MAXOLEN) 1651 break; 1652 switch (to->to_flags & mask) { 1653 case TOF_MSS: 1654 while (optlen % 4) { 1655 optlen += TCPOLEN_NOP; 1656 *optp++ = TCPOPT_NOP; 1657 } 1658 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1659 continue; 1660 optlen += TCPOLEN_MAXSEG; 1661 *optp++ = TCPOPT_MAXSEG; 1662 *optp++ = TCPOLEN_MAXSEG; 1663 to->to_mss = htons(to->to_mss); 1664 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1665 optp += sizeof(to->to_mss); 1666 break; 1667 case TOF_SCALE: 1668 while (!optlen || optlen % 2 != 1) { 1669 optlen += TCPOLEN_NOP; 1670 *optp++ = TCPOPT_NOP; 1671 } 1672 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1673 continue; 1674 optlen += TCPOLEN_WINDOW; 1675 *optp++ = TCPOPT_WINDOW; 1676 *optp++ = TCPOLEN_WINDOW; 1677 *optp++ = to->to_wscale; 1678 break; 1679 case TOF_SACKPERM: 1680 while (optlen % 2) { 1681 optlen += TCPOLEN_NOP; 1682 *optp++ = TCPOPT_NOP; 1683 } 1684 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1685 continue; 1686 optlen += TCPOLEN_SACK_PERMITTED; 1687 *optp++ = TCPOPT_SACK_PERMITTED; 1688 *optp++ = TCPOLEN_SACK_PERMITTED; 1689 break; 1690 case TOF_TS: 1691 while (!optlen || optlen % 4 != 2) { 1692 optlen += TCPOLEN_NOP; 1693 *optp++ = TCPOPT_NOP; 1694 } 1695 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1696 continue; 1697 optlen += TCPOLEN_TIMESTAMP; 1698 *optp++ = TCPOPT_TIMESTAMP; 1699 *optp++ = TCPOLEN_TIMESTAMP; 1700 to->to_tsval = htonl(to->to_tsval); 1701 to->to_tsecr = htonl(to->to_tsecr); 1702 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1703 optp += sizeof(to->to_tsval); 1704 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1705 optp += sizeof(to->to_tsecr); 1706 break; 1707#ifdef TCP_SIGNATURE 1708 case TOF_SIGNATURE: 1709 { 1710 int siglen = TCPOLEN_SIGNATURE - 2; 1711 1712 while (!optlen || optlen % 4 != 2) { 1713 optlen += TCPOLEN_NOP; 1714 *optp++ = TCPOPT_NOP; 1715 } 1716 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1717 continue; 1718 optlen += TCPOLEN_SIGNATURE; 1719 *optp++ = TCPOPT_SIGNATURE; 1720 *optp++ = TCPOLEN_SIGNATURE; 1721 to->to_signature = optp; 1722 while (siglen--) 1723 *optp++ = 0; 1724 break; 1725 } 1726#endif 1727 case TOF_SACK: 1728 { 1729 int sackblks = 0; 1730 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1731 tcp_seq sack_seq; 1732 1733 while (!optlen || optlen % 4 != 2) { 1734 optlen += TCPOLEN_NOP; 1735 *optp++ = TCPOPT_NOP; 1736 } 1737 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1738 continue; 1739 optlen += TCPOLEN_SACKHDR; 1740 *optp++ = TCPOPT_SACK; 1741 sackblks = min(to->to_nsacks, 1742 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1743 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1744 while (sackblks--) { 1745 sack_seq = htonl(sack->start); 1746 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1747 optp += sizeof(sack_seq); 1748 sack_seq = htonl(sack->end); 1749 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1750 optp += sizeof(sack_seq); 1751 optlen += TCPOLEN_SACK; 1752 sack++; 1753 } 1754 TCPSTAT_INC(tcps_sack_send_blocks); 1755 break; 1756 } 1757#ifdef TCP_RFC7413 1758 case TOF_FASTOPEN: 1759 { 1760 int total_len; 1761 1762 /* XXX is there any point to aligning this option? */ 1763 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; 1764 if (TCP_MAXOLEN - optlen < total_len) 1765 continue; 1766 *optp++ = TCPOPT_FAST_OPEN; 1767 *optp++ = total_len; 1768 if (to->to_tfo_len > 0) { 1769 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); 1770 optp += to->to_tfo_len; 1771 } 1772 optlen += total_len; 1773 break; 1774 } 1775#endif 1776 default: 1777 panic("%s: unknown TCP option type", __func__); 1778 break; 1779 } 1780 } 1781 1782 /* Terminate and pad TCP options to a 4 byte boundary. */ 1783 if (optlen % 4) { 1784 optlen += TCPOLEN_EOL; 1785 *optp++ = TCPOPT_EOL; 1786 } 1787 /* 1788 * According to RFC 793 (STD0007): 1789 * "The content of the header beyond the End-of-Option option 1790 * must be header padding (i.e., zero)." 1791 * and later: "The padding is composed of zeros." 1792 */ 1793 while (optlen % 4) { 1794 optlen += TCPOLEN_PAD; 1795 *optp++ = TCPOPT_PAD; 1796 } 1797 1798 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1799 return (optlen); 1800} 1801