tcp_output.c revision 293897
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 293897 2016-01-14 09:11:42Z glebius $"); 34 35#include "opt_inet.h" 36#include "opt_inet6.h" 37#include "opt_ipsec.h" 38#include "opt_kdtrace.h" 39#include "opt_tcpdebug.h" 40 41#include <sys/param.h> 42#include <sys/systm.h> 43#include <sys/domain.h> 44#include <sys/hhook.h> 45#include <sys/kernel.h> 46#include <sys/lock.h> 47#include <sys/mbuf.h> 48#include <sys/mutex.h> 49#include <sys/protosw.h> 50#include <sys/sdt.h> 51#include <sys/socket.h> 52#include <sys/socketvar.h> 53#include <sys/sysctl.h> 54 55#include <net/if.h> 56#include <net/route.h> 57#include <net/vnet.h> 58 59#include <netinet/cc.h> 60#include <netinet/in.h> 61#include <netinet/in_kdtrace.h> 62#include <netinet/in_systm.h> 63#include <netinet/ip.h> 64#include <netinet/in_pcb.h> 65#include <netinet/ip_var.h> 66#include <netinet/ip_options.h> 67#ifdef INET6 68#include <netinet6/in6_pcb.h> 69#include <netinet/ip6.h> 70#include <netinet6/ip6_var.h> 71#endif 72#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 130static void inline hhook_run_tcp_est_out(struct tcpcb *tp, 131 struct tcphdr *th, struct tcpopt *to, 132 long len, int tso); 133static void inline cc_after_idle(struct tcpcb *tp); 134 135/* 136 * Wrapper for the TCP established output helper hook. 137 */ 138static void inline 139hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 140 struct tcpopt *to, long len, int tso) 141{ 142 struct tcp_hhook_data hhook_data; 143 144 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 145 hhook_data.tp = tp; 146 hhook_data.th = th; 147 hhook_data.to = to; 148 hhook_data.len = len; 149 hhook_data.tso = tso; 150 151 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 152 tp->osd); 153 } 154} 155 156/* 157 * CC wrapper hook functions 158 */ 159static void inline 160cc_after_idle(struct tcpcb *tp) 161{ 162 INP_WLOCK_ASSERT(tp->t_inpcb); 163 164 if (CC_ALGO(tp)->after_idle != NULL) 165 CC_ALGO(tp)->after_idle(tp->ccv); 166} 167 168/* 169 * Tcp output routine: figure out what should be sent and send it. 170 */ 171int 172tcp_output(struct tcpcb *tp) 173{ 174 struct socket *so = tp->t_inpcb->inp_socket; 175 long len, recwin, sendwin; 176 int off, flags, error = 0; /* Keep compiler happy */ 177 struct mbuf *m; 178 struct ip *ip = NULL; 179 struct ipovly *ipov = NULL; 180 struct tcphdr *th; 181 u_char opt[TCP_MAXOLEN]; 182 unsigned ipoptlen, optlen, hdrlen; 183#ifdef IPSEC 184 unsigned ipsec_optlen = 0; 185#endif 186 int idle, sendalot; 187 int sack_rxmit, sack_bytes_rxmt; 188 struct sackhole *p; 189 int tso, mtu; 190 struct tcpopt to; 191#if 0 192 int maxburst = TCP_MAXBURST; 193#endif 194#ifdef INET6 195 struct ip6_hdr *ip6 = NULL; 196 int isipv6; 197 198 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 199#endif 200 201 INP_WLOCK_ASSERT(tp->t_inpcb); 202 203#ifdef TCP_OFFLOAD 204 if (tp->t_flags & TF_TOE) 205 return (tcp_offload_output(tp)); 206#endif 207 208#ifdef TCP_RFC7413 209 /* 210 * For TFO connections in SYN_RECEIVED, only allow the initial 211 * SYN|ACK and those sent by the retransmit timer. 212 */ 213 if ((tp->t_flags & TF_FASTOPEN) && 214 (tp->t_state == TCPS_SYN_RECEIVED) && 215 SEQ_GT(tp->snd_max, tp->snd_una) && /* inital SYN|ACK sent */ 216 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 217 return (0); 218#endif 219 /* 220 * Determine length of data that should be transmitted, 221 * and flags that will be used. 222 * If there is some data or critical controls (SYN, RST) 223 * to send, then transmit; otherwise, investigate further. 224 */ 225 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 226 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) 227 cc_after_idle(tp); 228 tp->t_flags &= ~TF_LASTIDLE; 229 if (idle) { 230 if (tp->t_flags & TF_MORETOCOME) { 231 tp->t_flags |= TF_LASTIDLE; 232 idle = 0; 233 } 234 } 235again: 236 /* 237 * If we've recently taken a timeout, snd_max will be greater than 238 * snd_nxt. There may be SACK information that allows us to avoid 239 * resending already delivered data. Adjust snd_nxt accordingly. 240 */ 241 if ((tp->t_flags & TF_SACK_PERMIT) && 242 SEQ_LT(tp->snd_nxt, tp->snd_max)) 243 tcp_sack_adjust(tp); 244 sendalot = 0; 245 tso = 0; 246 mtu = 0; 247 off = tp->snd_nxt - tp->snd_una; 248 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 249 250 flags = tcp_outflags[tp->t_state]; 251 /* 252 * Send any SACK-generated retransmissions. If we're explicitly trying 253 * to send out new data (when sendalot is 1), bypass this function. 254 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 255 * we're replacing a (future) new transmission with a retransmission 256 * now, and we previously incremented snd_cwnd in tcp_input(). 257 */ 258 /* 259 * Still in sack recovery , reset rxmit flag to zero. 260 */ 261 sack_rxmit = 0; 262 sack_bytes_rxmt = 0; 263 len = 0; 264 p = NULL; 265 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 266 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 267 long cwin; 268 269 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 270 if (cwin < 0) 271 cwin = 0; 272 /* Do not retransmit SACK segments beyond snd_recover */ 273 if (SEQ_GT(p->end, tp->snd_recover)) { 274 /* 275 * (At least) part of sack hole extends beyond 276 * snd_recover. Check to see if we can rexmit data 277 * for this hole. 278 */ 279 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 280 /* 281 * Can't rexmit any more data for this hole. 282 * That data will be rexmitted in the next 283 * sack recovery episode, when snd_recover 284 * moves past p->rxmit. 285 */ 286 p = NULL; 287 goto after_sack_rexmit; 288 } else 289 /* Can rexmit part of the current hole */ 290 len = ((long)ulmin(cwin, 291 tp->snd_recover - p->rxmit)); 292 } else 293 len = ((long)ulmin(cwin, p->end - p->rxmit)); 294 off = p->rxmit - tp->snd_una; 295 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 296 __func__, off)); 297 if (len > 0) { 298 sack_rxmit = 1; 299 sendalot = 1; 300 TCPSTAT_INC(tcps_sack_rexmits); 301 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 302 min(len, tp->t_maxseg)); 303 } 304 } 305after_sack_rexmit: 306 /* 307 * Get standard flags, and add SYN or FIN if requested by 'hidden' 308 * state flags. 309 */ 310 if (tp->t_flags & TF_NEEDFIN) 311 flags |= TH_FIN; 312 if (tp->t_flags & TF_NEEDSYN) 313 flags |= TH_SYN; 314 315 SOCKBUF_LOCK(&so->so_snd); 316 /* 317 * If in persist timeout with window of 0, send 1 byte. 318 * Otherwise, if window is small but nonzero 319 * and timer expired, we will send what we can 320 * and go to transmit state. 321 */ 322 if (tp->t_flags & TF_FORCEDATA) { 323 if (sendwin == 0) { 324 /* 325 * If we still have some data to send, then 326 * clear the FIN bit. Usually this would 327 * happen below when it realizes that we 328 * aren't sending all the data. However, 329 * if we have exactly 1 byte of unsent data, 330 * then it won't clear the FIN bit below, 331 * and if we are in persist state, we wind 332 * up sending the packet without recording 333 * that we sent the FIN bit. 334 * 335 * We can't just blindly clear the FIN bit, 336 * because if we don't have any more data 337 * to send then the probe will be the FIN 338 * itself. 339 */ 340 if (off < so->so_snd.sb_cc) 341 flags &= ~TH_FIN; 342 sendwin = 1; 343 } else { 344 tcp_timer_activate(tp, TT_PERSIST, 0); 345 tp->t_rxtshift = 0; 346 } 347 } 348 349 /* 350 * If snd_nxt == snd_max and we have transmitted a FIN, the 351 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 352 * a negative length. This can also occur when TCP opens up 353 * its congestion window while receiving additional duplicate 354 * acks after fast-retransmit because TCP will reset snd_nxt 355 * to snd_max after the fast-retransmit. 356 * 357 * In the normal retransmit-FIN-only case, however, snd_nxt will 358 * be set to snd_una, the offset will be 0, and the length may 359 * wind up 0. 360 * 361 * If sack_rxmit is true we are retransmitting from the scoreboard 362 * in which case len is already set. 363 */ 364 if (sack_rxmit == 0) { 365 if (sack_bytes_rxmt == 0) 366 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off); 367 else { 368 long cwin; 369 370 /* 371 * We are inside of a SACK recovery episode and are 372 * sending new data, having retransmitted all the 373 * data possible in the scoreboard. 374 */ 375 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 376 - off); 377 /* 378 * Don't remove this (len > 0) check ! 379 * We explicitly check for len > 0 here (although it 380 * isn't really necessary), to work around a gcc 381 * optimization issue - to force gcc to compute 382 * len above. Without this check, the computation 383 * of len is bungled by the optimizer. 384 */ 385 if (len > 0) { 386 cwin = tp->snd_cwnd - 387 (tp->snd_nxt - tp->sack_newdata) - 388 sack_bytes_rxmt; 389 if (cwin < 0) 390 cwin = 0; 391 len = lmin(len, cwin); 392 } 393 } 394 } 395 396 /* 397 * Lop off SYN bit if it has already been sent. However, if this 398 * is SYN-SENT state and if segment contains data and if we don't 399 * know that foreign host supports TAO, suppress sending segment. 400 */ 401 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 402 if (tp->t_state != TCPS_SYN_RECEIVED) 403 flags &= ~TH_SYN; 404#ifdef TCP_RFC7413 405 /* 406 * When sending additional segments following a TFO SYN|ACK, 407 * do not include the SYN bit. 408 */ 409 if ((tp->t_flags & TF_FASTOPEN) && 410 (tp->t_state == TCPS_SYN_RECEIVED)) 411 flags &= ~TH_SYN; 412#endif 413 off--, len++; 414 } 415 416 /* 417 * Be careful not to send data and/or FIN on SYN segments. 418 * This measure is needed to prevent interoperability problems 419 * with not fully conformant TCP implementations. 420 */ 421 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 422 len = 0; 423 flags &= ~TH_FIN; 424 } 425 426#ifdef TCP_RFC7413 427 /* 428 * When retransmitting SYN|ACK on a passively-created TFO socket, 429 * don't include data, as the presence of data may have caused the 430 * original SYN|ACK to have been dropped by a middlebox. 431 */ 432 if ((tp->t_flags & TF_FASTOPEN) && 433 (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) || 434 (flags & TH_RST))) 435 len = 0; 436#endif 437 if (len <= 0) { 438 /* 439 * If FIN has been sent but not acked, 440 * but we haven't been called to retransmit, 441 * len will be < 0. Otherwise, window shrank 442 * after we sent into it. If window shrank to 0, 443 * cancel pending retransmit, pull snd_nxt back 444 * to (closed) window, and set the persist timer 445 * if it isn't already going. If the window didn't 446 * close completely, just wait for an ACK. 447 * 448 * We also do a general check here to ensure that 449 * we will set the persist timer when we have data 450 * to send, but a 0-byte window. This makes sure 451 * the persist timer is set even if the packet 452 * hits one of the "goto send" lines below. 453 */ 454 len = 0; 455 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 456 (off < (int) so->so_snd.sb_cc)) { 457 tcp_timer_activate(tp, TT_REXMT, 0); 458 tp->t_rxtshift = 0; 459 tp->snd_nxt = tp->snd_una; 460 if (!tcp_timer_active(tp, TT_PERSIST)) 461 tcp_setpersist(tp); 462 } 463 } 464 465 /* len will be >= 0 after this point. */ 466 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 467 468 /* 469 * Automatic sizing of send socket buffer. Often the send buffer 470 * size is not optimally adjusted to the actual network conditions 471 * at hand (delay bandwidth product). Setting the buffer size too 472 * small limits throughput on links with high bandwidth and high 473 * delay (eg. trans-continental/oceanic links). Setting the 474 * buffer size too big consumes too much real kernel memory, 475 * especially with many connections on busy servers. 476 * 477 * The criteria to step up the send buffer one notch are: 478 * 1. receive window of remote host is larger than send buffer 479 * (with a fudge factor of 5/4th); 480 * 2. send buffer is filled to 7/8th with data (so we actually 481 * have data to make use of it); 482 * 3. send buffer fill has not hit maximal automatic size; 483 * 4. our send window (slow start and cogestion controlled) is 484 * larger than sent but unacknowledged data in send buffer. 485 * 486 * The remote host receive window scaling factor may limit the 487 * growing of the send buffer before it reaches its allowed 488 * maximum. 489 * 490 * It scales directly with slow start or congestion window 491 * and does at most one step per received ACK. This fast 492 * scaling has the drawback of growing the send buffer beyond 493 * what is strictly necessary to make full use of a given 494 * delay*bandwith product. However testing has shown this not 495 * to be much of an problem. At worst we are trading wasting 496 * of available bandwith (the non-use of it) for wasting some 497 * socket buffer memory. 498 * 499 * TODO: Shrink send buffer during idle periods together 500 * with congestion window. Requires another timer. Has to 501 * wait for upcoming tcp timer rewrite. 502 */ 503 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 504 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 505 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 506 so->so_snd.sb_cc < V_tcp_autosndbuf_max && 507 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 508 if (!sbreserve_locked(&so->so_snd, 509 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 510 V_tcp_autosndbuf_max), so, curthread)) 511 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 512 } 513 } 514 515 /* 516 * Decide if we can use TCP Segmentation Offloading (if supported by 517 * hardware). 518 * 519 * TSO may only be used if we are in a pure bulk sending state. The 520 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 521 * IP options prevent using TSO. With TSO the TCP header is the same 522 * (except for the sequence number) for all generated packets. This 523 * makes it impossible to transmit any options which vary per generated 524 * segment or packet. 525 */ 526#ifdef IPSEC 527 /* 528 * Pre-calculate here as we save another lookup into the darknesses 529 * of IPsec that way and can actually decide if TSO is ok. 530 */ 531 ipsec_optlen = ipsec_hdrsiz_tcp(tp); 532#endif 533 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 534 ((tp->t_flags & TF_SIGNATURE) == 0) && 535 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 536#ifdef IPSEC 537 ipsec_optlen == 0 && 538#endif 539 tp->t_inpcb->inp_options == NULL && 540 tp->t_inpcb->in6p_options == NULL) 541 tso = 1; 542 543 if (sack_rxmit) { 544 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 545 flags &= ~TH_FIN; 546 } else { 547 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 548 flags &= ~TH_FIN; 549 } 550 551 recwin = sbspace(&so->so_rcv); 552 553 /* 554 * Sender silly window avoidance. We transmit under the following 555 * conditions when len is non-zero: 556 * 557 * - We have a full segment (or more with TSO) 558 * - This is the last buffer in a write()/send() and we are 559 * either idle or running NODELAY 560 * - we've timed out (e.g. persist timer) 561 * - we have more then 1/2 the maximum send window's worth of 562 * data (receiver may be limited the window size) 563 * - we need to retransmit 564 */ 565 if (len) { 566 if (len >= tp->t_maxseg) 567 goto send; 568 /* 569 * NOTE! on localhost connections an 'ack' from the remote 570 * end may occur synchronously with the output and cause 571 * us to flush a buffer queued with moretocome. XXX 572 * 573 * note: the len + off check is almost certainly unnecessary. 574 */ 575 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 576 (idle || (tp->t_flags & TF_NODELAY)) && 577 len + off >= so->so_snd.sb_cc && 578 (tp->t_flags & TF_NOPUSH) == 0) { 579 goto send; 580 } 581 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 582 goto send; 583 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 584 goto send; 585 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 586 goto send; 587 if (sack_rxmit) 588 goto send; 589 } 590 591 /* 592 * Sending of standalone window updates. 593 * 594 * Window updates are important when we close our window due to a 595 * full socket buffer and are opening it again after the application 596 * reads data from it. Once the window has opened again and the 597 * remote end starts to send again the ACK clock takes over and 598 * provides the most current window information. 599 * 600 * We must avoid the silly window syndrome whereas every read 601 * from the receive buffer, no matter how small, causes a window 602 * update to be sent. We also should avoid sending a flurry of 603 * window updates when the socket buffer had queued a lot of data 604 * and the application is doing small reads. 605 * 606 * Prevent a flurry of pointless window updates by only sending 607 * an update when we can increase the advertized window by more 608 * than 1/4th of the socket buffer capacity. When the buffer is 609 * getting full or is very small be more aggressive and send an 610 * update whenever we can increase by two mss sized segments. 611 * In all other situations the ACK's to new incoming data will 612 * carry further window increases. 613 * 614 * Don't send an independent window update if a delayed 615 * ACK is pending (it will get piggy-backed on it) or the 616 * remote side already has done a half-close and won't send 617 * more data. Skip this if the connection is in T/TCP 618 * half-open state. 619 */ 620 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 621 !(tp->t_flags & TF_DELACK) && 622 !TCPS_HAVERCVDFIN(tp->t_state)) { 623 /* 624 * "adv" is the amount we could increase the window, 625 * taking into account that we are limited by 626 * TCP_MAXWIN << tp->rcv_scale. 627 */ 628 long adv; 629 int oldwin; 630 631 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 632 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 633 oldwin = (tp->rcv_adv - tp->rcv_nxt); 634 adv -= oldwin; 635 } else 636 oldwin = 0; 637 638 /* 639 * If the new window size ends up being the same as the old 640 * size when it is scaled, then don't force a window update. 641 */ 642 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 643 goto dontupdate; 644 645 if (adv >= (long)(2 * tp->t_maxseg) && 646 (adv >= (long)(so->so_rcv.sb_hiwat / 4) || 647 recwin <= (long)(so->so_rcv.sb_hiwat / 8) || 648 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 649 goto send; 650 } 651dontupdate: 652 653 /* 654 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 655 * is also a catch-all for the retransmit timer timeout case. 656 */ 657 if (tp->t_flags & TF_ACKNOW) 658 goto send; 659 if ((flags & TH_RST) || 660 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 661 goto send; 662 if (SEQ_GT(tp->snd_up, tp->snd_una)) 663 goto send; 664 /* 665 * If our state indicates that FIN should be sent 666 * and we have not yet done so, then we need to send. 667 */ 668 if (flags & TH_FIN && 669 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 670 goto send; 671 /* 672 * In SACK, it is possible for tcp_output to fail to send a segment 673 * after the retransmission timer has been turned off. Make sure 674 * that the retransmission timer is set. 675 */ 676 if ((tp->t_flags & TF_SACK_PERMIT) && 677 SEQ_GT(tp->snd_max, tp->snd_una) && 678 !tcp_timer_active(tp, TT_REXMT) && 679 !tcp_timer_active(tp, TT_PERSIST)) { 680 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 681 goto just_return; 682 } 683 /* 684 * TCP window updates are not reliable, rather a polling protocol 685 * using ``persist'' packets is used to insure receipt of window 686 * updates. The three ``states'' for the output side are: 687 * idle not doing retransmits or persists 688 * persisting to move a small or zero window 689 * (re)transmitting and thereby not persisting 690 * 691 * tcp_timer_active(tp, TT_PERSIST) 692 * is true when we are in persist state. 693 * (tp->t_flags & TF_FORCEDATA) 694 * is set when we are called to send a persist packet. 695 * tcp_timer_active(tp, TT_REXMT) 696 * is set when we are retransmitting 697 * The output side is idle when both timers are zero. 698 * 699 * If send window is too small, there is data to transmit, and no 700 * retransmit or persist is pending, then go to persist state. 701 * If nothing happens soon, send when timer expires: 702 * if window is nonzero, transmit what we can, 703 * otherwise force out a byte. 704 */ 705 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) && 706 !tcp_timer_active(tp, TT_PERSIST)) { 707 tp->t_rxtshift = 0; 708 tcp_setpersist(tp); 709 } 710 711 /* 712 * No reason to send a segment, just return. 713 */ 714just_return: 715 SOCKBUF_UNLOCK(&so->so_snd); 716 return (0); 717 718send: 719 SOCKBUF_LOCK_ASSERT(&so->so_snd); 720 if (len > 0) { 721 if (len >= tp->t_maxseg) 722 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 723 else 724 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 725 } 726 /* 727 * Before ESTABLISHED, force sending of initial options 728 * unless TCP set not to do any options. 729 * NOTE: we assume that the IP/TCP header plus TCP options 730 * always fit in a single mbuf, leaving room for a maximum 731 * link header, i.e. 732 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 733 */ 734 optlen = 0; 735#ifdef INET6 736 if (isipv6) 737 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 738 else 739#endif 740 hdrlen = sizeof (struct tcpiphdr); 741 742 /* 743 * Compute options for segment. 744 * We only have to care about SYN and established connection 745 * segments. Options for SYN-ACK segments are handled in TCP 746 * syncache. 747 */ 748 to.to_flags = 0; 749 if ((tp->t_flags & TF_NOOPT) == 0) { 750 /* Maximum segment size. */ 751 if (flags & TH_SYN) { 752 tp->snd_nxt = tp->iss; 753 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 754 to.to_flags |= TOF_MSS; 755#ifdef TCP_RFC7413 756 /* 757 * Only include the TFO option on the first 758 * transmission of the SYN|ACK on a 759 * passively-created TFO socket, as the presence of 760 * the TFO option may have caused the original 761 * SYN|ACK to have been dropped by a middlebox. 762 */ 763 if ((tp->t_flags & TF_FASTOPEN) && 764 (tp->t_state == TCPS_SYN_RECEIVED) && 765 (tp->t_rxtshift == 0)) { 766 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 767 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie; 768 to.to_flags |= TOF_FASTOPEN; 769 } 770#endif 771 } 772 /* Window scaling. */ 773 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 774 to.to_wscale = tp->request_r_scale; 775 to.to_flags |= TOF_SCALE; 776 } 777 /* Timestamps. */ 778 if ((tp->t_flags & TF_RCVD_TSTMP) || 779 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 780 to.to_tsval = tcp_ts_getticks() + tp->ts_offset; 781 to.to_tsecr = tp->ts_recent; 782 to.to_flags |= TOF_TS; 783 /* Set receive buffer autosizing timestamp. */ 784 if (tp->rfbuf_ts == 0 && 785 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 786 tp->rfbuf_ts = tcp_ts_getticks(); 787 } 788 /* Selective ACK's. */ 789 if (tp->t_flags & TF_SACK_PERMIT) { 790 if (flags & TH_SYN) 791 to.to_flags |= TOF_SACKPERM; 792 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 793 (tp->t_flags & TF_SACK_PERMIT) && 794 tp->rcv_numsacks > 0) { 795 to.to_flags |= TOF_SACK; 796 to.to_nsacks = tp->rcv_numsacks; 797 to.to_sacks = (u_char *)tp->sackblks; 798 } 799 } 800#ifdef TCP_SIGNATURE 801 /* TCP-MD5 (RFC2385). */ 802 if (tp->t_flags & TF_SIGNATURE) 803 to.to_flags |= TOF_SIGNATURE; 804#endif /* TCP_SIGNATURE */ 805 806 /* Processing the options. */ 807 hdrlen += optlen = tcp_addoptions(&to, opt); 808 } 809 810#ifdef INET6 811 if (isipv6) 812 ipoptlen = ip6_optlen(tp->t_inpcb); 813 else 814#endif 815 if (tp->t_inpcb->inp_options) 816 ipoptlen = tp->t_inpcb->inp_options->m_len - 817 offsetof(struct ipoption, ipopt_list); 818 else 819 ipoptlen = 0; 820#ifdef IPSEC 821 ipoptlen += ipsec_optlen; 822#endif 823 824 /* 825 * Adjust data length if insertion of options will 826 * bump the packet length beyond the t_maxopd length. 827 * Clear the FIN bit because we cut off the tail of 828 * the segment. 829 */ 830 if (len + optlen + ipoptlen > tp->t_maxopd) { 831 flags &= ~TH_FIN; 832 833 if (tso) { 834 u_int if_hw_tsomax; 835 u_int if_hw_tsomaxsegcount; 836 u_int if_hw_tsomaxsegsize; 837 struct mbuf *mb; 838 u_int moff; 839 int max_len; 840 841 /* extract TSO information */ 842 if_hw_tsomax = tp->t_tsomax; 843 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 844 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 845 846 /* 847 * Limit a TSO burst to prevent it from 848 * overflowing or exceeding the maximum length 849 * allowed by the network interface: 850 */ 851 KASSERT(ipoptlen == 0, 852 ("%s: TSO can't do IP options", __func__)); 853 854 /* 855 * Check if we should limit by maximum payload 856 * length: 857 */ 858 if (if_hw_tsomax != 0) { 859 /* compute maximum TSO length */ 860 max_len = (if_hw_tsomax - hdrlen - 861 max_linkhdr); 862 if (max_len <= 0) { 863 len = 0; 864 } else if (len > max_len) { 865 sendalot = 1; 866 len = max_len; 867 } 868 } 869 870 /* 871 * Check if we should limit by maximum segment 872 * size and count: 873 */ 874 if (if_hw_tsomaxsegcount != 0 && 875 if_hw_tsomaxsegsize != 0) { 876 /* 877 * Subtract one segment for the LINK 878 * and TCP/IP headers mbuf that will 879 * be prepended to this mbuf chain 880 * after the code in this section 881 * limits the number of mbufs in the 882 * chain to if_hw_tsomaxsegcount. 883 */ 884 if_hw_tsomaxsegcount -= 1; 885 max_len = 0; 886 mb = sbsndmbuf(&so->so_snd, off, &moff); 887 888 while (mb != NULL && max_len < len) { 889 u_int mlen; 890 u_int frags; 891 892 /* 893 * Get length of mbuf fragment 894 * and how many hardware frags, 895 * rounded up, it would use: 896 */ 897 mlen = (mb->m_len - moff); 898 frags = howmany(mlen, 899 if_hw_tsomaxsegsize); 900 901 /* Handle special case: Zero Length Mbuf */ 902 if (frags == 0) 903 frags = 1; 904 905 /* 906 * Check if the fragment limit 907 * will be reached or exceeded: 908 */ 909 if (frags >= if_hw_tsomaxsegcount) { 910 max_len += min(mlen, 911 if_hw_tsomaxsegcount * 912 if_hw_tsomaxsegsize); 913 break; 914 } 915 max_len += mlen; 916 if_hw_tsomaxsegcount -= frags; 917 moff = 0; 918 mb = mb->m_next; 919 } 920 if (max_len <= 0) { 921 len = 0; 922 } else if (len > max_len) { 923 sendalot = 1; 924 len = max_len; 925 } 926 } 927 928 /* 929 * Prevent the last segment from being 930 * fractional unless the send sockbuf can be 931 * emptied: 932 */ 933 max_len = (tp->t_maxopd - optlen); 934 if ((off + len) < so->so_snd.sb_cc) { 935 moff = len % max_len; 936 if (moff != 0) { 937 len -= moff; 938 sendalot = 1; 939 } 940 } 941 942 /* 943 * In case there are too many small fragments 944 * don't use TSO: 945 */ 946 if (len <= max_len) { 947 len = max_len; 948 sendalot = 1; 949 tso = 0; 950 } 951 952 /* 953 * Send the FIN in a separate segment 954 * after the bulk sending is done. 955 * We don't trust the TSO implementations 956 * to clear the FIN flag on all but the 957 * last segment. 958 */ 959 if (tp->t_flags & TF_NEEDFIN) 960 sendalot = 1; 961 962 } else { 963 len = tp->t_maxopd - optlen - ipoptlen; 964 sendalot = 1; 965 } 966 } else 967 tso = 0; 968 969 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 970 ("%s: len > IP_MAXPACKET", __func__)); 971 972/*#ifdef DIAGNOSTIC*/ 973#ifdef INET6 974 if (max_linkhdr + hdrlen > MCLBYTES) 975#else 976 if (max_linkhdr + hdrlen > MHLEN) 977#endif 978 panic("tcphdr too big"); 979/*#endif*/ 980 981 /* 982 * This KASSERT is here to catch edge cases at a well defined place. 983 * Before, those had triggered (random) panic conditions further down. 984 */ 985 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 986 987 /* 988 * Grab a header mbuf, attaching a copy of data to 989 * be transmitted, and initialize the header from 990 * the template for sends on this connection. 991 */ 992 if (len) { 993 struct mbuf *mb; 994 u_int moff; 995 996 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 997 TCPSTAT_INC(tcps_sndprobe); 998 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 999 tp->t_sndrexmitpack++; 1000 TCPSTAT_INC(tcps_sndrexmitpack); 1001 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 1002 } else { 1003 TCPSTAT_INC(tcps_sndpack); 1004 TCPSTAT_ADD(tcps_sndbyte, len); 1005 } 1006#ifdef INET6 1007 if (MHLEN < hdrlen + max_linkhdr) 1008 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1009 else 1010#endif 1011 m = m_gethdr(M_NOWAIT, MT_DATA); 1012 1013 if (m == NULL) { 1014 SOCKBUF_UNLOCK(&so->so_snd); 1015 error = ENOBUFS; 1016 sack_rxmit = 0; 1017 goto out; 1018 } 1019 1020 m->m_data += max_linkhdr; 1021 m->m_len = hdrlen; 1022 1023 /* 1024 * Start the m_copy functions from the closest mbuf 1025 * to the offset in the socket buffer chain. 1026 */ 1027 mb = sbsndptr(&so->so_snd, off, len, &moff); 1028 1029 if (len <= MHLEN - hdrlen - max_linkhdr) { 1030 m_copydata(mb, moff, (int)len, 1031 mtod(m, caddr_t) + hdrlen); 1032 m->m_len += len; 1033 } else { 1034 m->m_next = m_copy(mb, moff, (int)len); 1035 if (m->m_next == NULL) { 1036 SOCKBUF_UNLOCK(&so->so_snd); 1037 (void) m_free(m); 1038 error = ENOBUFS; 1039 sack_rxmit = 0; 1040 goto out; 1041 } 1042 } 1043 1044 /* 1045 * If we're sending everything we've got, set PUSH. 1046 * (This will keep happy those implementations which only 1047 * give data to the user when a buffer fills or 1048 * a PUSH comes in.) 1049 */ 1050 if ((off + len == so->so_snd.sb_cc) && !(flags & TH_SYN)) 1051 flags |= TH_PUSH; 1052 SOCKBUF_UNLOCK(&so->so_snd); 1053 } else { 1054 SOCKBUF_UNLOCK(&so->so_snd); 1055 if (tp->t_flags & TF_ACKNOW) 1056 TCPSTAT_INC(tcps_sndacks); 1057 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1058 TCPSTAT_INC(tcps_sndctrl); 1059 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1060 TCPSTAT_INC(tcps_sndurg); 1061 else 1062 TCPSTAT_INC(tcps_sndwinup); 1063 1064 m = m_gethdr(M_NOWAIT, MT_DATA); 1065 if (m == NULL) { 1066 error = ENOBUFS; 1067 sack_rxmit = 0; 1068 goto out; 1069 } 1070#ifdef INET6 1071 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1072 MHLEN >= hdrlen) { 1073 MH_ALIGN(m, hdrlen); 1074 } else 1075#endif 1076 m->m_data += max_linkhdr; 1077 m->m_len = hdrlen; 1078 } 1079 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1080 m->m_pkthdr.rcvif = (struct ifnet *)0; 1081#ifdef MAC 1082 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1083#endif 1084#ifdef INET6 1085 if (isipv6) { 1086 ip6 = mtod(m, struct ip6_hdr *); 1087 th = (struct tcphdr *)(ip6 + 1); 1088 tcpip_fillheaders(tp->t_inpcb, ip6, th); 1089 } else 1090#endif /* INET6 */ 1091 { 1092 ip = mtod(m, struct ip *); 1093 ipov = (struct ipovly *)ip; 1094 th = (struct tcphdr *)(ip + 1); 1095 tcpip_fillheaders(tp->t_inpcb, ip, th); 1096 } 1097 1098 /* 1099 * Fill in fields, remembering maximum advertised 1100 * window for use in delaying messages about window sizes. 1101 * If resending a FIN, be sure not to use a new sequence number. 1102 */ 1103 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1104 tp->snd_nxt == tp->snd_max) 1105 tp->snd_nxt--; 1106 /* 1107 * If we are starting a connection, send ECN setup 1108 * SYN packet. If we are on a retransmit, we may 1109 * resend those bits a number of times as per 1110 * RFC 3168. 1111 */ 1112 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 1113 if (tp->t_rxtshift >= 1) { 1114 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 1115 flags |= TH_ECE|TH_CWR; 1116 } else 1117 flags |= TH_ECE|TH_CWR; 1118 } 1119 1120 if (tp->t_state == TCPS_ESTABLISHED && 1121 (tp->t_flags & TF_ECN_PERMIT)) { 1122 /* 1123 * If the peer has ECN, mark data packets with 1124 * ECN capable transmission (ECT). 1125 * Ignore pure ack packets, retransmissions and window probes. 1126 */ 1127 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1128 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 1129#ifdef INET6 1130 if (isipv6) 1131 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1132 else 1133#endif 1134 ip->ip_tos |= IPTOS_ECN_ECT0; 1135 TCPSTAT_INC(tcps_ecn_ect0); 1136 } 1137 1138 /* 1139 * Reply with proper ECN notifications. 1140 */ 1141 if (tp->t_flags & TF_ECN_SND_CWR) { 1142 flags |= TH_CWR; 1143 tp->t_flags &= ~TF_ECN_SND_CWR; 1144 } 1145 if (tp->t_flags & TF_ECN_SND_ECE) 1146 flags |= TH_ECE; 1147 } 1148 1149 /* 1150 * If we are doing retransmissions, then snd_nxt will 1151 * not reflect the first unsent octet. For ACK only 1152 * packets, we do not want the sequence number of the 1153 * retransmitted packet, we want the sequence number 1154 * of the next unsent octet. So, if there is no data 1155 * (and no SYN or FIN), use snd_max instead of snd_nxt 1156 * when filling in ti_seq. But if we are in persist 1157 * state, snd_max might reflect one byte beyond the 1158 * right edge of the window, so use snd_nxt in that 1159 * case, since we know we aren't doing a retransmission. 1160 * (retransmit and persist are mutually exclusive...) 1161 */ 1162 if (sack_rxmit == 0) { 1163 if (len || (flags & (TH_SYN|TH_FIN)) || 1164 tcp_timer_active(tp, TT_PERSIST)) 1165 th->th_seq = htonl(tp->snd_nxt); 1166 else 1167 th->th_seq = htonl(tp->snd_max); 1168 } else { 1169 th->th_seq = htonl(p->rxmit); 1170 p->rxmit += len; 1171 tp->sackhint.sack_bytes_rexmit += len; 1172 } 1173 th->th_ack = htonl(tp->rcv_nxt); 1174 if (optlen) { 1175 bcopy(opt, th + 1, optlen); 1176 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1177 } 1178 th->th_flags = flags; 1179 /* 1180 * Calculate receive window. Don't shrink window, 1181 * but avoid silly window syndrome. 1182 */ 1183 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 1184 recwin < (long)tp->t_maxseg) 1185 recwin = 0; 1186 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1187 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 1188 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 1189 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 1190 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 1191 1192 /* 1193 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1194 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1195 * case is handled in syncache. 1196 */ 1197 if (flags & TH_SYN) 1198 th->th_win = htons((u_short) 1199 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1200 else 1201 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1202 1203 /* 1204 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1205 * a 0 window. This may cause the remote transmitter to stall. This 1206 * flag tells soreceive() to disable delayed acknowledgements when 1207 * draining the buffer. This can occur if the receiver is attempting 1208 * to read more data than can be buffered prior to transmitting on 1209 * the connection. 1210 */ 1211 if (th->th_win == 0) { 1212 tp->t_sndzerowin++; 1213 tp->t_flags |= TF_RXWIN0SENT; 1214 } else 1215 tp->t_flags &= ~TF_RXWIN0SENT; 1216 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1217 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1218 th->th_flags |= TH_URG; 1219 } else 1220 /* 1221 * If no urgent pointer to send, then we pull 1222 * the urgent pointer to the left edge of the send window 1223 * so that it doesn't drift into the send window on sequence 1224 * number wraparound. 1225 */ 1226 tp->snd_up = tp->snd_una; /* drag it along */ 1227 1228#ifdef TCP_SIGNATURE 1229 if (to.to_flags & TOF_SIGNATURE) { 1230 int sigoff = to.to_signature - opt; 1231 tcp_signature_compute(m, 0, len, optlen, 1232 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1233 } 1234#endif 1235 1236 /* 1237 * Put TCP length in extended header, and then 1238 * checksum extended header and data. 1239 */ 1240 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1241 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1242#ifdef INET6 1243 if (isipv6) { 1244 /* 1245 * ip6_plen is not need to be filled now, and will be filled 1246 * in ip6_output. 1247 */ 1248 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1249 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1250 optlen + len, IPPROTO_TCP, 0); 1251 } 1252#endif 1253#if defined(INET6) && defined(INET) 1254 else 1255#endif 1256#ifdef INET 1257 { 1258 m->m_pkthdr.csum_flags = CSUM_TCP; 1259 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1260 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1261 1262 /* IP version must be set here for ipv4/ipv6 checking later */ 1263 KASSERT(ip->ip_v == IPVERSION, 1264 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1265 } 1266#endif 1267 1268 /* 1269 * Enable TSO and specify the size of the segments. 1270 * The TCP pseudo header checksum is always provided. 1271 * XXX: Fixme: This is currently not the case for IPv6. 1272 */ 1273 if (tso) { 1274 KASSERT(len > tp->t_maxopd - optlen, 1275 ("%s: len <= tso_segsz", __func__)); 1276 m->m_pkthdr.csum_flags |= CSUM_TSO; 1277 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 1278 } 1279 1280#ifdef IPSEC 1281 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1282 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u", 1283 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1284#else 1285 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1286 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u", 1287 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1288#endif 1289 1290 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1291 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1292 1293#ifdef TCPDEBUG 1294 /* 1295 * Trace. 1296 */ 1297 if (so->so_options & SO_DEBUG) { 1298 u_short save = 0; 1299#ifdef INET6 1300 if (!isipv6) 1301#endif 1302 { 1303 save = ipov->ih_len; 1304 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1305 } 1306 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1307#ifdef INET6 1308 if (!isipv6) 1309#endif 1310 ipov->ih_len = save; 1311 } 1312#endif /* TCPDEBUG */ 1313 1314 /* 1315 * Fill in IP length and desired time to live and 1316 * send to IP level. There should be a better way 1317 * to handle ttl and tos; we could keep them in 1318 * the template, but need a way to checksum without them. 1319 */ 1320 /* 1321 * m->m_pkthdr.len should have been set before cksum calcuration, 1322 * because in6_cksum() need it. 1323 */ 1324#ifdef INET6 1325 if (isipv6) { 1326 struct route_in6 ro; 1327 1328 bzero(&ro, sizeof(ro)); 1329 /* 1330 * we separately set hoplimit for every segment, since the 1331 * user might want to change the value via setsockopt. 1332 * Also, desired default hop limit might be changed via 1333 * Neighbor Discovery. 1334 */ 1335 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1336 1337 /* 1338 * Set the packet size here for the benefit of DTrace probes. 1339 * ip6_output() will set it properly; it's supposed to include 1340 * the option header lengths as well. 1341 */ 1342 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1343 1344 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) 1345 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1346 else 1347 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1348 1349 if (tp->t_state == TCPS_SYN_SENT) 1350 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1351 1352 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1353 1354 /* TODO: IPv6 IP6TOS_ECT bit on */ 1355 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro, 1356 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1357 NULL, NULL, tp->t_inpcb); 1358 1359 if (error == EMSGSIZE && ro.ro_rt != NULL) 1360 mtu = ro.ro_rt->rt_mtu; 1361 RO_RTFREE(&ro); 1362 } 1363#endif /* INET6 */ 1364#if defined(INET) && defined(INET6) 1365 else 1366#endif 1367#ifdef INET 1368 { 1369 struct route ro; 1370 1371 bzero(&ro, sizeof(ro)); 1372 ip->ip_len = htons(m->m_pkthdr.len); 1373#ifdef INET6 1374 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1375 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1376#endif /* INET6 */ 1377 /* 1378 * If we do path MTU discovery, then we set DF on every packet. 1379 * This might not be the best thing to do according to RFC3390 1380 * Section 2. However the tcp hostcache migitates the problem 1381 * so it affects only the first tcp connection with a host. 1382 * 1383 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1384 */ 1385 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) { 1386 ip->ip_off |= htons(IP_DF); 1387 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1388 } else { 1389 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1390 } 1391 1392 if (tp->t_state == TCPS_SYN_SENT) 1393 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1394 1395 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1396 1397 error = ip_output(m, tp->t_inpcb->inp_options, &ro, 1398 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1399 tp->t_inpcb); 1400 1401 if (error == EMSGSIZE && ro.ro_rt != NULL) 1402 mtu = ro.ro_rt->rt_mtu; 1403 RO_RTFREE(&ro); 1404 } 1405#endif /* INET */ 1406 1407out: 1408 /* 1409 * In transmit state, time the transmission and arrange for 1410 * the retransmit. In persist state, just set snd_max. 1411 */ 1412 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1413 !tcp_timer_active(tp, TT_PERSIST)) { 1414 tcp_seq startseq = tp->snd_nxt; 1415 1416 /* 1417 * Advance snd_nxt over sequence space of this segment. 1418 */ 1419 if (flags & (TH_SYN|TH_FIN)) { 1420 if (flags & TH_SYN) 1421 tp->snd_nxt++; 1422 if (flags & TH_FIN) { 1423 tp->snd_nxt++; 1424 tp->t_flags |= TF_SENTFIN; 1425 } 1426 } 1427 if (sack_rxmit) 1428 goto timer; 1429 tp->snd_nxt += len; 1430 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1431 tp->snd_max = tp->snd_nxt; 1432 /* 1433 * Time this transmission if not a retransmission and 1434 * not currently timing anything. 1435 */ 1436 if (tp->t_rtttime == 0) { 1437 tp->t_rtttime = ticks; 1438 tp->t_rtseq = startseq; 1439 TCPSTAT_INC(tcps_segstimed); 1440 } 1441 } 1442 1443 /* 1444 * Set retransmit timer if not currently set, 1445 * and not doing a pure ack or a keep-alive probe. 1446 * Initial value for retransmit timer is smoothed 1447 * round-trip time + 2 * round-trip time variance. 1448 * Initialize shift counter which is used for backoff 1449 * of retransmit time. 1450 */ 1451timer: 1452 if (!tcp_timer_active(tp, TT_REXMT) && 1453 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1454 (tp->snd_nxt != tp->snd_una))) { 1455 if (tcp_timer_active(tp, TT_PERSIST)) { 1456 tcp_timer_activate(tp, TT_PERSIST, 0); 1457 tp->t_rxtshift = 0; 1458 } 1459 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1460 } else if (len == 0 && so->so_snd.sb_cc && 1461 !tcp_timer_active(tp, TT_REXMT) && 1462 !tcp_timer_active(tp, TT_PERSIST)) { 1463 /* 1464 * Avoid a situation where we do not set persist timer 1465 * after a zero window condition. For example: 1466 * 1) A -> B: packet with enough data to fill the window 1467 * 2) B -> A: ACK for #1 + new data (0 window 1468 * advertisement) 1469 * 3) A -> B: ACK for #2, 0 len packet 1470 * 1471 * In this case, A will not activate the persist timer, 1472 * because it chose to send a packet. Unless tcp_output 1473 * is called for some other reason (delayed ack timer, 1474 * another input packet from B, socket syscall), A will 1475 * not send zero window probes. 1476 * 1477 * So, if you send a 0-length packet, but there is data 1478 * in the socket buffer, and neither the rexmt or 1479 * persist timer is already set, then activate the 1480 * persist timer. 1481 */ 1482 tp->t_rxtshift = 0; 1483 tcp_setpersist(tp); 1484 } 1485 } else { 1486 /* 1487 * Persist case, update snd_max but since we are in 1488 * persist mode (no window) we do not update snd_nxt. 1489 */ 1490 int xlen = len; 1491 if (flags & TH_SYN) 1492 ++xlen; 1493 if (flags & TH_FIN) { 1494 ++xlen; 1495 tp->t_flags |= TF_SENTFIN; 1496 } 1497 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1498 tp->snd_max = tp->snd_nxt + len; 1499 } 1500 1501 if (error) { 1502 1503 /* 1504 * We know that the packet was lost, so back out the 1505 * sequence number advance, if any. 1506 * 1507 * If the error is EPERM the packet got blocked by the 1508 * local firewall. Normally we should terminate the 1509 * connection but the blocking may have been spurious 1510 * due to a firewall reconfiguration cycle. So we treat 1511 * it like a packet loss and let the retransmit timer and 1512 * timeouts do their work over time. 1513 * XXX: It is a POLA question whether calling tcp_drop right 1514 * away would be the really correct behavior instead. 1515 */ 1516 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1517 !tcp_timer_active(tp, TT_PERSIST)) && 1518 ((flags & TH_SYN) == 0) && 1519 (error != EPERM)) { 1520 if (sack_rxmit) { 1521 p->rxmit -= len; 1522 tp->sackhint.sack_bytes_rexmit -= len; 1523 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1524 ("sackhint bytes rtx >= 0")); 1525 } else 1526 tp->snd_nxt -= len; 1527 } 1528 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1529 switch (error) { 1530 case EPERM: 1531 tp->t_softerror = error; 1532 return (error); 1533 case ENOBUFS: 1534 if (!tcp_timer_active(tp, TT_REXMT) && 1535 !tcp_timer_active(tp, TT_PERSIST)) 1536 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1537 tp->snd_cwnd = tp->t_maxseg; 1538 return (0); 1539 case EMSGSIZE: 1540 /* 1541 * For some reason the interface we used initially 1542 * to send segments changed to another or lowered 1543 * its MTU. 1544 * If TSO was active we either got an interface 1545 * without TSO capabilits or TSO was turned off. 1546 * If we obtained mtu from ip_output() then update 1547 * it and try again. 1548 */ 1549 if (tso) 1550 tp->t_flags &= ~TF_TSO; 1551 if (mtu != 0) { 1552 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1553 goto again; 1554 } 1555 return (error); 1556 case EHOSTDOWN: 1557 case EHOSTUNREACH: 1558 case ENETDOWN: 1559 case ENETUNREACH: 1560 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1561 tp->t_softerror = error; 1562 return (0); 1563 } 1564 /* FALLTHROUGH */ 1565 default: 1566 return (error); 1567 } 1568 } 1569 TCPSTAT_INC(tcps_sndtotal); 1570 1571 /* 1572 * Data sent (as far as we can tell). 1573 * If this advertises a larger window than any other segment, 1574 * then remember the size of the advertised window. 1575 * Any pending ACK has now been sent. 1576 */ 1577 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1578 tp->rcv_adv = tp->rcv_nxt + recwin; 1579 tp->last_ack_sent = tp->rcv_nxt; 1580 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1581 if (tcp_timer_active(tp, TT_DELACK)) 1582 tcp_timer_activate(tp, TT_DELACK, 0); 1583#if 0 1584 /* 1585 * This completely breaks TCP if newreno is turned on. What happens 1586 * is that if delayed-acks are turned on on the receiver, this code 1587 * on the transmitter effectively destroys the TCP window, forcing 1588 * it to four packets (1.5Kx4 = 6K window). 1589 */ 1590 if (sendalot && --maxburst) 1591 goto again; 1592#endif 1593 if (sendalot) 1594 goto again; 1595 return (0); 1596} 1597 1598void 1599tcp_setpersist(struct tcpcb *tp) 1600{ 1601 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1602 int tt; 1603 1604 tp->t_flags &= ~TF_PREVVALID; 1605 if (tcp_timer_active(tp, TT_REXMT)) 1606 panic("tcp_setpersist: retransmit pending"); 1607 /* 1608 * Start/restart persistance timer. 1609 */ 1610 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1611 TCPTV_PERSMIN, TCPTV_PERSMAX); 1612 tcp_timer_activate(tp, TT_PERSIST, tt); 1613 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1614 tp->t_rxtshift++; 1615} 1616 1617/* 1618 * Insert TCP options according to the supplied parameters to the place 1619 * optp in a consistent way. Can handle unaligned destinations. 1620 * 1621 * The order of the option processing is crucial for optimal packing and 1622 * alignment for the scarce option space. 1623 * 1624 * The optimal order for a SYN/SYN-ACK segment is: 1625 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1626 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1627 * 1628 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1629 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1630 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1631 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1632 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1633 */ 1634int 1635tcp_addoptions(struct tcpopt *to, u_char *optp) 1636{ 1637 u_int mask, optlen = 0; 1638 1639 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1640 if ((to->to_flags & mask) != mask) 1641 continue; 1642 if (optlen == TCP_MAXOLEN) 1643 break; 1644 switch (to->to_flags & mask) { 1645 case TOF_MSS: 1646 while (optlen % 4) { 1647 optlen += TCPOLEN_NOP; 1648 *optp++ = TCPOPT_NOP; 1649 } 1650 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1651 continue; 1652 optlen += TCPOLEN_MAXSEG; 1653 *optp++ = TCPOPT_MAXSEG; 1654 *optp++ = TCPOLEN_MAXSEG; 1655 to->to_mss = htons(to->to_mss); 1656 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1657 optp += sizeof(to->to_mss); 1658 break; 1659 case TOF_SCALE: 1660 while (!optlen || optlen % 2 != 1) { 1661 optlen += TCPOLEN_NOP; 1662 *optp++ = TCPOPT_NOP; 1663 } 1664 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1665 continue; 1666 optlen += TCPOLEN_WINDOW; 1667 *optp++ = TCPOPT_WINDOW; 1668 *optp++ = TCPOLEN_WINDOW; 1669 *optp++ = to->to_wscale; 1670 break; 1671 case TOF_SACKPERM: 1672 while (optlen % 2) { 1673 optlen += TCPOLEN_NOP; 1674 *optp++ = TCPOPT_NOP; 1675 } 1676 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1677 continue; 1678 optlen += TCPOLEN_SACK_PERMITTED; 1679 *optp++ = TCPOPT_SACK_PERMITTED; 1680 *optp++ = TCPOLEN_SACK_PERMITTED; 1681 break; 1682 case TOF_TS: 1683 while (!optlen || optlen % 4 != 2) { 1684 optlen += TCPOLEN_NOP; 1685 *optp++ = TCPOPT_NOP; 1686 } 1687 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1688 continue; 1689 optlen += TCPOLEN_TIMESTAMP; 1690 *optp++ = TCPOPT_TIMESTAMP; 1691 *optp++ = TCPOLEN_TIMESTAMP; 1692 to->to_tsval = htonl(to->to_tsval); 1693 to->to_tsecr = htonl(to->to_tsecr); 1694 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1695 optp += sizeof(to->to_tsval); 1696 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1697 optp += sizeof(to->to_tsecr); 1698 break; 1699#ifdef TCP_SIGNATURE 1700 case TOF_SIGNATURE: 1701 { 1702 int siglen = TCPOLEN_SIGNATURE - 2; 1703 1704 while (!optlen || optlen % 4 != 2) { 1705 optlen += TCPOLEN_NOP; 1706 *optp++ = TCPOPT_NOP; 1707 } 1708 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1709 continue; 1710 optlen += TCPOLEN_SIGNATURE; 1711 *optp++ = TCPOPT_SIGNATURE; 1712 *optp++ = TCPOLEN_SIGNATURE; 1713 to->to_signature = optp; 1714 while (siglen--) 1715 *optp++ = 0; 1716 break; 1717 } 1718#endif 1719 case TOF_SACK: 1720 { 1721 int sackblks = 0; 1722 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1723 tcp_seq sack_seq; 1724 1725 while (!optlen || optlen % 4 != 2) { 1726 optlen += TCPOLEN_NOP; 1727 *optp++ = TCPOPT_NOP; 1728 } 1729 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1730 continue; 1731 optlen += TCPOLEN_SACKHDR; 1732 *optp++ = TCPOPT_SACK; 1733 sackblks = min(to->to_nsacks, 1734 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1735 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1736 while (sackblks--) { 1737 sack_seq = htonl(sack->start); 1738 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1739 optp += sizeof(sack_seq); 1740 sack_seq = htonl(sack->end); 1741 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1742 optp += sizeof(sack_seq); 1743 optlen += TCPOLEN_SACK; 1744 sack++; 1745 } 1746 TCPSTAT_INC(tcps_sack_send_blocks); 1747 break; 1748 } 1749#ifdef TCP_RFC7413 1750 case TOF_FASTOPEN: 1751 { 1752 int total_len; 1753 1754 /* XXX is there any point to aligning this option? */ 1755 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; 1756 if (TCP_MAXOLEN - optlen < total_len) 1757 continue; 1758 *optp++ = TCPOPT_FAST_OPEN; 1759 *optp++ = total_len; 1760 if (to->to_tfo_len > 0) { 1761 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); 1762 optp += to->to_tfo_len; 1763 } 1764 optlen += total_len; 1765 break; 1766 } 1767#endif 1768 default: 1769 panic("%s: unknown TCP option type", __func__); 1770 break; 1771 } 1772 } 1773 1774 /* Terminate and pad TCP options to a 4 byte boundary. */ 1775 if (optlen % 4) { 1776 optlen += TCPOLEN_EOL; 1777 *optp++ = TCPOPT_EOL; 1778 } 1779 /* 1780 * According to RFC 793 (STD0007): 1781 * "The content of the header beyond the End-of-Option option 1782 * must be header padding (i.e., zero)." 1783 * and later: "The padding is composed of zeros." 1784 */ 1785 while (optlen % 4) { 1786 optlen += TCPOLEN_PAD; 1787 *optp++ = TCPOPT_PAD; 1788 } 1789 1790 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1791 return (optlen); 1792} 1793