1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 5 * The Regents of the University of California. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Neither the name of the University nor the names of its contributors 16 * may be used to endorse or promote products derived from this software 17 * without specific prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 * 31 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 32 */ 33 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD$"); 36 37 #include "opt_inet.h" 38 #include "opt_inet6.h" 39 #include "opt_ipsec.h" 40 #include "opt_kern_tls.h" 41 #include "opt_tcpdebug.h" 42 43 #include <sys/param.h> 44 #include <sys/systm.h> 45 #include <sys/arb.h> 46 #include <sys/domain.h> 47 #ifdef TCP_HHOOK 48 #include <sys/hhook.h> 49 #endif 50 #include <sys/kernel.h> 51 #ifdef KERN_TLS 52 #include <sys/ktls.h> 53 #endif 54 #include <sys/lock.h> 55 #include <sys/mbuf.h> 56 #include <sys/mutex.h> 57 #include <sys/protosw.h> 58 #include <sys/qmath.h> 59 #include <sys/sdt.h> 60 #include <sys/socket.h> 61 #include <sys/socketvar.h> 62 #include <sys/sysctl.h> 63 #include <sys/stats.h> 64 65 #include <net/if.h> 66 #include <net/route.h> 67 #include <net/route/nhop.h> 68 #include <net/vnet.h> 69 70 #include <netinet/in.h> 71 #include <netinet/in_kdtrace.h> 72 #include <netinet/in_systm.h> 73 #include <netinet/ip.h> 74 #include <netinet/in_pcb.h> 75 #include <netinet/ip_var.h> 76 #include <netinet/ip_options.h> 77 #ifdef INET6 78 #include <netinet6/in6_pcb.h> 79 #include <netinet/ip6.h> 80 #include <netinet6/ip6_var.h> 81 #endif 82 #include <netinet/tcp.h> 83 #define TCPOUTFLAGS 84 #include <netinet/tcp_fsm.h> 85 #include <netinet/tcp_log_buf.h> 86 #include <netinet/tcp_seq.h> 87 #include <netinet/tcp_var.h> 88 #include <netinet/tcp_syncache.h> 89 #include <netinet/tcp_timer.h> 90 #include <netinet/tcpip.h> 91 #include <netinet/cc/cc.h> 92 #include <netinet/tcp_fastopen.h> 93 #ifdef TCPPCAP 94 #include <netinet/tcp_pcap.h> 95 #endif 96 #ifdef TCPDEBUG 97 #include <netinet/tcp_debug.h> 98 #endif 99 #ifdef TCP_OFFLOAD 100 #include <netinet/tcp_offload.h> 101 #endif 102 #include <netinet/tcp_ecn.h> 103 104 #include <netipsec/ipsec_support.h> 105 106 #include <netinet/udp.h> 107 #include <netinet/udp_var.h> 108 #include <machine/in_cksum.h> 109 110 #include <security/mac/mac_framework.h> 111 112 VNET_DEFINE(int, path_mtu_discovery) = 1; 113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW, 114 &VNET_NAME(path_mtu_discovery), 1, 115 "Enable Path MTU Discovery"); 116 117 VNET_DEFINE(int, tcp_do_tso) = 1; 118 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW, 119 &VNET_NAME(tcp_do_tso), 0, 120 "Enable TCP Segmentation Offload"); 121 122 VNET_DEFINE(int, tcp_sendspace) = 1024*32; 123 #define V_tcp_sendspace VNET(tcp_sendspace) 124 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW, 125 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 126 127 VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 128 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, 129 &VNET_NAME(tcp_do_autosndbuf), 0, 130 "Enable automatic send buffer sizing"); 131 132 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW, 134 &VNET_NAME(tcp_autosndbuf_inc), 0, 135 "Incrementor step size of automatic send buffer"); 136 137 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; 138 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW, 139 &VNET_NAME(tcp_autosndbuf_max), 0, 140 "Max size of automatic send buffer"); 141 142 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0; 143 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat) 144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW, 145 &VNET_NAME(tcp_sendbuf_auto_lowat), 0, 146 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT"); 147 148 /* 149 * Make sure that either retransmit or persist timer is set for SYN, FIN and 150 * non-ACK. 151 */ 152 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \ 153 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\ 154 tcp_timer_active((tp), TT_REXMT) || \ 155 tcp_timer_active((tp), TT_PERSIST), \ 156 ("neither rexmt nor persist timer is set")) 157 158 #ifdef TCP_HHOOK 159 /* 160 * Wrapper for the TCP established output helper hook. 161 */ 162 void 163 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 164 struct tcpopt *to, uint32_t len, int tso) 165 { 166 struct tcp_hhook_data hhook_data; 167 168 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 169 hhook_data.tp = tp; 170 hhook_data.th = th; 171 hhook_data.to = to; 172 hhook_data.len = len; 173 hhook_data.tso = tso; 174 175 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 176 tp->osd); 177 } 178 } 179 #endif 180 181 /* 182 * CC wrapper hook functions 183 */ 184 void 185 cc_after_idle(struct tcpcb *tp) 186 { 187 INP_WLOCK_ASSERT(tptoinpcb(tp)); 188 189 if (CC_ALGO(tp)->after_idle != NULL) 190 CC_ALGO(tp)->after_idle(tp->ccv); 191 } 192 193 /* 194 * Tcp output routine: figure out what should be sent and send it. 195 */ 196 int 197 tcp_default_output(struct tcpcb *tp) 198 { 199 struct socket *so = tptosocket(tp); 200 struct inpcb *inp = tptoinpcb(tp); 201 int32_t len; 202 uint32_t recwin, sendwin; 203 uint16_t flags; 204 int off, error = 0; /* Keep compiler happy */ 205 u_int if_hw_tsomaxsegcount = 0; 206 u_int if_hw_tsomaxsegsize = 0; 207 struct mbuf *m; 208 struct ip *ip = NULL; 209 #ifdef TCPDEBUG 210 struct ipovly *ipov = NULL; 211 #endif 212 struct tcphdr *th; 213 u_char opt[TCP_MAXOLEN]; 214 unsigned ipoptlen, optlen, hdrlen, ulen; 215 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 216 unsigned ipsec_optlen = 0; 217 #endif 218 int idle, sendalot, curticks; 219 int sack_rxmit, sack_bytes_rxmt; 220 struct sackhole *p; 221 int tso, mtu; 222 struct tcpopt to; 223 struct udphdr *udp = NULL; 224 struct tcp_log_buffer *lgb; 225 unsigned int wanted_cookie = 0; 226 unsigned int dont_sendalot = 0; 227 #if 0 228 int maxburst = TCP_MAXBURST; 229 #endif 230 #ifdef INET6 231 struct ip6_hdr *ip6 = NULL; 232 int isipv6; 233 234 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 235 #endif 236 #ifdef KERN_TLS 237 const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0; 238 #else 239 const bool hw_tls = false; 240 #endif 241 242 NET_EPOCH_ASSERT(); 243 INP_WLOCK_ASSERT(inp); 244 245 #ifdef TCP_OFFLOAD 246 if (tp->t_flags & TF_TOE) 247 return (tcp_offload_output(tp)); 248 #endif 249 250 /* 251 * For TFO connections in SYN_SENT or SYN_RECEIVED, 252 * only allow the initial SYN or SYN|ACK and those sent 253 * by the retransmit timer. 254 */ 255 if (IS_FASTOPEN(tp->t_flags) && 256 ((tp->t_state == TCPS_SYN_SENT) || 257 (tp->t_state == TCPS_SYN_RECEIVED)) && 258 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 259 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 260 return (0); 261 262 /* 263 * Determine length of data that should be transmitted, 264 * and flags that will be used. 265 * If there is some data or critical controls (SYN, RST) 266 * to send, then transmit; otherwise, investigate further. 267 */ 268 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 269 if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) || 270 (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur)))) 271 cc_after_idle(tp); 272 tp->t_flags &= ~TF_LASTIDLE; 273 if (idle) { 274 if (tp->t_flags & TF_MORETOCOME) { 275 tp->t_flags |= TF_LASTIDLE; 276 idle = 0; 277 } 278 } 279 again: 280 /* 281 * If we've recently taken a timeout, snd_max will be greater than 282 * snd_nxt. There may be SACK information that allows us to avoid 283 * resending already delivered data. Adjust snd_nxt accordingly. 284 */ 285 if ((tp->t_flags & TF_SACK_PERMIT) && 286 SEQ_LT(tp->snd_nxt, tp->snd_max)) 287 tcp_sack_adjust(tp); 288 sendalot = 0; 289 tso = 0; 290 mtu = 0; 291 off = tp->snd_nxt - tp->snd_una; 292 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 293 294 flags = tcp_outflags[tp->t_state]; 295 /* 296 * Send any SACK-generated retransmissions. If we're explicitly trying 297 * to send out new data (when sendalot is 1), bypass this function. 298 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 299 * we're replacing a (future) new transmission with a retransmission 300 * now, and we previously incremented snd_cwnd in tcp_input(). 301 */ 302 /* 303 * Still in sack recovery , reset rxmit flag to zero. 304 */ 305 sack_rxmit = 0; 306 sack_bytes_rxmt = 0; 307 len = 0; 308 p = NULL; 309 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 310 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 311 uint32_t cwin; 312 313 cwin = 314 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0); 315 /* Do not retransmit SACK segments beyond snd_recover */ 316 if (SEQ_GT(p->end, tp->snd_recover)) { 317 /* 318 * (At least) part of sack hole extends beyond 319 * snd_recover. Check to see if we can rexmit data 320 * for this hole. 321 */ 322 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 323 /* 324 * Can't rexmit any more data for this hole. 325 * That data will be rexmitted in the next 326 * sack recovery episode, when snd_recover 327 * moves past p->rxmit. 328 */ 329 p = NULL; 330 goto after_sack_rexmit; 331 } else { 332 /* Can rexmit part of the current hole */ 333 len = ((int32_t)ulmin(cwin, 334 SEQ_SUB(tp->snd_recover, p->rxmit))); 335 } 336 } else { 337 len = ((int32_t)ulmin(cwin, 338 SEQ_SUB(p->end, p->rxmit))); 339 } 340 if (len > 0) { 341 off = SEQ_SUB(p->rxmit, tp->snd_una); 342 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 343 __func__, off)); 344 sack_rxmit = 1; 345 sendalot = 1; 346 TCPSTAT_INC(tcps_sack_rexmits); 347 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 348 min(len, tcp_maxseg(tp))); 349 } 350 } 351 after_sack_rexmit: 352 /* 353 * Get standard flags, and add SYN or FIN if requested by 'hidden' 354 * state flags. 355 */ 356 if (tp->t_flags & TF_NEEDFIN) 357 flags |= TH_FIN; 358 if (tp->t_flags & TF_NEEDSYN) 359 flags |= TH_SYN; 360 361 SOCKBUF_LOCK(&so->so_snd); 362 /* 363 * If in persist timeout with window of 0, send 1 byte. 364 * Otherwise, if window is small but nonzero 365 * and timer expired, we will send what we can 366 * and go to transmit state. 367 */ 368 if (tp->t_flags & TF_FORCEDATA) { 369 if (sendwin == 0) { 370 /* 371 * If we still have some data to send, then 372 * clear the FIN bit. Usually this would 373 * happen below when it realizes that we 374 * aren't sending all the data. However, 375 * if we have exactly 1 byte of unsent data, 376 * then it won't clear the FIN bit below, 377 * and if we are in persist state, we wind 378 * up sending the packet without recording 379 * that we sent the FIN bit. 380 * 381 * We can't just blindly clear the FIN bit, 382 * because if we don't have any more data 383 * to send then the probe will be the FIN 384 * itself. 385 */ 386 if (off < sbused(&so->so_snd)) 387 flags &= ~TH_FIN; 388 sendwin = 1; 389 } else { 390 tcp_timer_activate(tp, TT_PERSIST, 0); 391 tp->t_rxtshift = 0; 392 } 393 } 394 395 /* 396 * If snd_nxt == snd_max and we have transmitted a FIN, the 397 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 398 * a negative length. This can also occur when TCP opens up 399 * its congestion window while receiving additional duplicate 400 * acks after fast-retransmit because TCP will reset snd_nxt 401 * to snd_max after the fast-retransmit. 402 * 403 * In the normal retransmit-FIN-only case, however, snd_nxt will 404 * be set to snd_una, the offset will be 0, and the length may 405 * wind up 0. 406 * 407 * If sack_rxmit is true we are retransmitting from the scoreboard 408 * in which case len is already set. 409 */ 410 if (sack_rxmit == 0) { 411 if (sack_bytes_rxmt == 0) 412 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) - 413 off); 414 else { 415 int32_t cwin; 416 417 /* 418 * We are inside of a SACK recovery episode and are 419 * sending new data, having retransmitted all the 420 * data possible in the scoreboard. 421 */ 422 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) - 423 off); 424 /* 425 * Don't remove this (len > 0) check ! 426 * We explicitly check for len > 0 here (although it 427 * isn't really necessary), to work around a gcc 428 * optimization issue - to force gcc to compute 429 * len above. Without this check, the computation 430 * of len is bungled by the optimizer. 431 */ 432 if (len > 0) { 433 cwin = tp->snd_cwnd - imax(0, (int32_t) 434 (tp->snd_nxt - tp->snd_recover)) - 435 sack_bytes_rxmt; 436 if (cwin < 0) 437 cwin = 0; 438 len = imin(len, cwin); 439 } 440 } 441 } 442 443 /* 444 * Lop off SYN bit if it has already been sent. However, if this 445 * is SYN-SENT state and if segment contains data and if we don't 446 * know that foreign host supports TAO, suppress sending segment. 447 */ 448 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 449 if (tp->t_state != TCPS_SYN_RECEIVED) 450 flags &= ~TH_SYN; 451 /* 452 * When sending additional segments following a TFO SYN|ACK, 453 * do not include the SYN bit. 454 */ 455 if (IS_FASTOPEN(tp->t_flags) && 456 (tp->t_state == TCPS_SYN_RECEIVED)) 457 flags &= ~TH_SYN; 458 off--, len++; 459 } 460 461 /* 462 * Be careful not to send data and/or FIN on SYN segments. 463 * This measure is needed to prevent interoperability problems 464 * with not fully conformant TCP implementations. 465 */ 466 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 467 len = 0; 468 flags &= ~TH_FIN; 469 } 470 471 /* 472 * On TFO sockets, ensure no data is sent in the following cases: 473 * 474 * - When retransmitting SYN|ACK on a passively-created socket 475 * 476 * - When retransmitting SYN on an actively created socket 477 * 478 * - When sending a zero-length cookie (cookie request) on an 479 * actively created socket 480 * 481 * - When the socket is in the CLOSED state (RST is being sent) 482 */ 483 if (IS_FASTOPEN(tp->t_flags) && 484 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 485 ((tp->t_state == TCPS_SYN_SENT) && 486 (tp->t_tfo_client_cookie_len == 0)) || 487 (flags & TH_RST))) 488 len = 0; 489 if (len <= 0) { 490 /* 491 * If FIN has been sent but not acked, 492 * but we haven't been called to retransmit, 493 * len will be < 0. Otherwise, window shrank 494 * after we sent into it. If window shrank to 0, 495 * cancel pending retransmit, pull snd_nxt back 496 * to (closed) window, and set the persist timer 497 * if it isn't already going. If the window didn't 498 * close completely, just wait for an ACK. 499 * 500 * We also do a general check here to ensure that 501 * we will set the persist timer when we have data 502 * to send, but a 0-byte window. This makes sure 503 * the persist timer is set even if the packet 504 * hits one of the "goto send" lines below. 505 */ 506 len = 0; 507 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 508 (off < (int) sbavail(&so->so_snd)) && 509 !tcp_timer_active(tp, TT_PERSIST)) { 510 tcp_timer_activate(tp, TT_REXMT, 0); 511 tp->t_rxtshift = 0; 512 tp->snd_nxt = tp->snd_una; 513 if (!tcp_timer_active(tp, TT_PERSIST)) 514 tcp_setpersist(tp); 515 } 516 } 517 518 /* len will be >= 0 after this point. */ 519 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 520 521 tcp_sndbuf_autoscale(tp, so, sendwin); 522 523 /* 524 * Decide if we can use TCP Segmentation Offloading (if supported by 525 * hardware). 526 * 527 * TSO may only be used if we are in a pure bulk sending state. The 528 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 529 * IP options prevent using TSO. With TSO the TCP header is the same 530 * (except for the sequence number) for all generated packets. This 531 * makes it impossible to transmit any options which vary per generated 532 * segment or packet. 533 * 534 * IPv4 handling has a clear separation of ip options and ip header 535 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 536 * the right thing below to provide length of just ip options and thus 537 * checking for ipoptlen is enough to decide if ip options are present. 538 */ 539 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 540 /* 541 * Pre-calculate here as we save another lookup into the darknesses 542 * of IPsec that way and can actually decide if TSO is ok. 543 */ 544 #ifdef INET6 545 if (isipv6 && IPSEC_ENABLED(ipv6)) 546 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); 547 #ifdef INET 548 else 549 #endif 550 #endif /* INET6 */ 551 #ifdef INET 552 if (IPSEC_ENABLED(ipv4)) 553 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); 554 #endif /* INET */ 555 #endif /* IPSEC */ 556 #ifdef INET6 557 if (isipv6) 558 ipoptlen = ip6_optlen(inp); 559 else 560 #endif 561 if (inp->inp_options) 562 ipoptlen = inp->inp_options->m_len - 563 offsetof(struct ipoption, ipopt_list); 564 else 565 ipoptlen = 0; 566 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 567 ipoptlen += ipsec_optlen; 568 #endif 569 570 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 571 (tp->t_port == 0) && 572 ((tp->t_flags & TF_SIGNATURE) == 0) && 573 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 574 ipoptlen == 0 && !(flags & TH_SYN)) 575 tso = 1; 576 577 if (sack_rxmit) { 578 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd))) 579 flags &= ~TH_FIN; 580 } else { 581 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 582 sbused(&so->so_snd))) 583 flags &= ~TH_FIN; 584 } 585 586 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 587 (long)TCP_MAXWIN << tp->rcv_scale); 588 589 /* 590 * Sender silly window avoidance. We transmit under the following 591 * conditions when len is non-zero: 592 * 593 * - We have a full segment (or more with TSO) 594 * - This is the last buffer in a write()/send() and we are 595 * either idle or running NODELAY 596 * - we've timed out (e.g. persist timer) 597 * - we have more then 1/2 the maximum send window's worth of 598 * data (receiver may be limited the window size) 599 * - we need to retransmit 600 */ 601 if (len) { 602 if (len >= tp->t_maxseg) 603 goto send; 604 /* 605 * As the TCP header options are now 606 * considered when setting up the initial 607 * window, we would not send the last segment 608 * if we skip considering the option length here. 609 * Note: this may not work when tcp headers change 610 * very dynamically in the future. 611 */ 612 if ((((tp->t_flags & TF_SIGNATURE) ? 613 PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) + 614 ((tp->t_flags & TF_RCVD_TSTMP) ? 615 PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) + 616 len) >= tp->t_maxseg) 617 goto send; 618 /* 619 * NOTE! on localhost connections an 'ack' from the remote 620 * end may occur synchronously with the output and cause 621 * us to flush a buffer queued with moretocome. XXX 622 * 623 * note: the len + off check is almost certainly unnecessary. 624 */ 625 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 626 (idle || (tp->t_flags & TF_NODELAY)) && 627 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) && 628 (tp->t_flags & TF_NOPUSH) == 0) { 629 goto send; 630 } 631 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 632 goto send; 633 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 634 goto send; 635 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 636 goto send; 637 if (sack_rxmit) 638 goto send; 639 } 640 641 /* 642 * Sending of standalone window updates. 643 * 644 * Window updates are important when we close our window due to a 645 * full socket buffer and are opening it again after the application 646 * reads data from it. Once the window has opened again and the 647 * remote end starts to send again the ACK clock takes over and 648 * provides the most current window information. 649 * 650 * We must avoid the silly window syndrome whereas every read 651 * from the receive buffer, no matter how small, causes a window 652 * update to be sent. We also should avoid sending a flurry of 653 * window updates when the socket buffer had queued a lot of data 654 * and the application is doing small reads. 655 * 656 * Prevent a flurry of pointless window updates by only sending 657 * an update when we can increase the advertized window by more 658 * than 1/4th of the socket buffer capacity. When the buffer is 659 * getting full or is very small be more aggressive and send an 660 * update whenever we can increase by two mss sized segments. 661 * In all other situations the ACK's to new incoming data will 662 * carry further window increases. 663 * 664 * Don't send an independent window update if a delayed 665 * ACK is pending (it will get piggy-backed on it) or the 666 * remote side already has done a half-close and won't send 667 * more data. Skip this if the connection is in T/TCP 668 * half-open state. 669 */ 670 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 671 !(tp->t_flags & TF_DELACK) && 672 !TCPS_HAVERCVDFIN(tp->t_state)) { 673 /* 674 * "adv" is the amount we could increase the window, 675 * taking into account that we are limited by 676 * TCP_MAXWIN << tp->rcv_scale. 677 */ 678 int32_t adv; 679 int oldwin; 680 681 adv = recwin; 682 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 683 oldwin = (tp->rcv_adv - tp->rcv_nxt); 684 if (adv > oldwin) 685 adv -= oldwin; 686 else 687 adv = 0; 688 } else 689 oldwin = 0; 690 691 /* 692 * If the new window size ends up being the same as or less 693 * than the old size when it is scaled, then don't force 694 * a window update. 695 */ 696 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 697 goto dontupdate; 698 699 if (adv >= (int32_t)(2 * tp->t_maxseg) && 700 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 701 recwin <= (so->so_rcv.sb_hiwat / 8) || 702 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg || 703 adv >= TCP_MAXWIN << tp->rcv_scale)) 704 goto send; 705 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat) 706 goto send; 707 } 708 dontupdate: 709 710 /* 711 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 712 * is also a catch-all for the retransmit timer timeout case. 713 */ 714 if (tp->t_flags & TF_ACKNOW) 715 goto send; 716 if ((flags & TH_RST) || 717 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 718 goto send; 719 if (SEQ_GT(tp->snd_up, tp->snd_una)) 720 goto send; 721 /* 722 * If our state indicates that FIN should be sent 723 * and we have not yet done so, then we need to send. 724 */ 725 if (flags & TH_FIN && 726 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 727 goto send; 728 /* 729 * In SACK, it is possible for tcp_output to fail to send a segment 730 * after the retransmission timer has been turned off. Make sure 731 * that the retransmission timer is set. 732 */ 733 if ((tp->t_flags & TF_SACK_PERMIT) && 734 SEQ_GT(tp->snd_max, tp->snd_una) && 735 !tcp_timer_active(tp, TT_REXMT) && 736 !tcp_timer_active(tp, TT_PERSIST)) { 737 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp)); 738 goto just_return; 739 } 740 /* 741 * TCP window updates are not reliable, rather a polling protocol 742 * using ``persist'' packets is used to insure receipt of window 743 * updates. The three ``states'' for the output side are: 744 * idle not doing retransmits or persists 745 * persisting to move a small or zero window 746 * (re)transmitting and thereby not persisting 747 * 748 * tcp_timer_active(tp, TT_PERSIST) 749 * is true when we are in persist state. 750 * (tp->t_flags & TF_FORCEDATA) 751 * is set when we are called to send a persist packet. 752 * tcp_timer_active(tp, TT_REXMT) 753 * is set when we are retransmitting 754 * The output side is idle when both timers are zero. 755 * 756 * If send window is too small, there is data to transmit, and no 757 * retransmit or persist is pending, then go to persist state. 758 * If nothing happens soon, send when timer expires: 759 * if window is nonzero, transmit what we can, 760 * otherwise force out a byte. 761 */ 762 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && 763 !tcp_timer_active(tp, TT_PERSIST)) { 764 tp->t_rxtshift = 0; 765 tcp_setpersist(tp); 766 } 767 768 /* 769 * No reason to send a segment, just return. 770 */ 771 just_return: 772 SOCKBUF_UNLOCK(&so->so_snd); 773 return (0); 774 775 send: 776 SOCKBUF_LOCK_ASSERT(&so->so_snd); 777 if (len > 0) { 778 if (len >= tp->t_maxseg) 779 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 780 else 781 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 782 } 783 /* 784 * Before ESTABLISHED, force sending of initial options 785 * unless TCP set not to do any options. 786 * NOTE: we assume that the IP/TCP header plus TCP options 787 * always fit in a single mbuf, leaving room for a maximum 788 * link header, i.e. 789 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 790 */ 791 optlen = 0; 792 #ifdef INET6 793 if (isipv6) 794 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 795 else 796 #endif 797 hdrlen = sizeof (struct tcpiphdr); 798 799 if (flags & TH_SYN) { 800 tp->snd_nxt = tp->iss; 801 } 802 803 /* 804 * Compute options for segment. 805 * We only have to care about SYN and established connection 806 * segments. Options for SYN-ACK segments are handled in TCP 807 * syncache. 808 */ 809 to.to_flags = 0; 810 if ((tp->t_flags & TF_NOOPT) == 0) { 811 /* Maximum segment size. */ 812 if (flags & TH_SYN) { 813 to.to_mss = tcp_mssopt(&inp->inp_inc); 814 if (tp->t_port) 815 to.to_mss -= V_tcp_udp_tunneling_overhead; 816 to.to_flags |= TOF_MSS; 817 818 /* 819 * On SYN or SYN|ACK transmits on TFO connections, 820 * only include the TFO option if it is not a 821 * retransmit, as the presence of the TFO option may 822 * have caused the original SYN or SYN|ACK to have 823 * been dropped by a middlebox. 824 */ 825 if (IS_FASTOPEN(tp->t_flags) && 826 (tp->t_rxtshift == 0)) { 827 if (tp->t_state == TCPS_SYN_RECEIVED) { 828 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 829 to.to_tfo_cookie = 830 (u_int8_t *)&tp->t_tfo_cookie.server; 831 to.to_flags |= TOF_FASTOPEN; 832 wanted_cookie = 1; 833 } else if (tp->t_state == TCPS_SYN_SENT) { 834 to.to_tfo_len = 835 tp->t_tfo_client_cookie_len; 836 to.to_tfo_cookie = 837 tp->t_tfo_cookie.client; 838 to.to_flags |= TOF_FASTOPEN; 839 wanted_cookie = 1; 840 /* 841 * If we wind up having more data to 842 * send with the SYN than can fit in 843 * one segment, don't send any more 844 * until the SYN|ACK comes back from 845 * the other end. 846 */ 847 dont_sendalot = 1; 848 } 849 } 850 } 851 /* Window scaling. */ 852 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 853 to.to_wscale = tp->request_r_scale; 854 to.to_flags |= TOF_SCALE; 855 } 856 /* Timestamps. */ 857 if ((tp->t_flags & TF_RCVD_TSTMP) || 858 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 859 curticks = tcp_ts_getticks(); 860 to.to_tsval = curticks + tp->ts_offset; 861 to.to_tsecr = tp->ts_recent; 862 to.to_flags |= TOF_TS; 863 if (tp->t_rxtshift == 1) 864 tp->t_badrxtwin = curticks; 865 } 866 867 /* Set receive buffer autosizing timestamp. */ 868 if (tp->rfbuf_ts == 0 && 869 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 870 tp->rfbuf_ts = tcp_ts_getticks(); 871 872 /* Selective ACK's. */ 873 if (tp->t_flags & TF_SACK_PERMIT) { 874 if (flags & TH_SYN) 875 to.to_flags |= TOF_SACKPERM; 876 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 877 tp->rcv_numsacks > 0) { 878 to.to_flags |= TOF_SACK; 879 to.to_nsacks = tp->rcv_numsacks; 880 to.to_sacks = (u_char *)tp->sackblks; 881 } 882 } 883 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 884 /* TCP-MD5 (RFC2385). */ 885 /* 886 * Check that TCP_MD5SIG is enabled in tcpcb to 887 * account the size needed to set this TCP option. 888 */ 889 if (tp->t_flags & TF_SIGNATURE) 890 to.to_flags |= TOF_SIGNATURE; 891 #endif /* TCP_SIGNATURE */ 892 893 /* Processing the options. */ 894 hdrlen += optlen = tcp_addoptions(&to, opt); 895 /* 896 * If we wanted a TFO option to be added, but it was unable 897 * to fit, ensure no data is sent. 898 */ 899 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 900 !(to.to_flags & TOF_FASTOPEN)) 901 len = 0; 902 } 903 if (tp->t_port) { 904 if (V_tcp_udp_tunneling_port == 0) { 905 /* The port was removed?? */ 906 SOCKBUF_UNLOCK(&so->so_snd); 907 return (EHOSTUNREACH); 908 } 909 hdrlen += sizeof(struct udphdr); 910 } 911 /* 912 * Adjust data length if insertion of options will 913 * bump the packet length beyond the t_maxseg length. 914 * Clear the FIN bit because we cut off the tail of 915 * the segment. 916 */ 917 if (len + optlen + ipoptlen > tp->t_maxseg) { 918 flags &= ~TH_FIN; 919 920 if (tso) { 921 u_int if_hw_tsomax; 922 u_int moff; 923 int max_len; 924 925 /* extract TSO information */ 926 if_hw_tsomax = tp->t_tsomax; 927 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 928 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 929 930 /* 931 * Limit a TSO burst to prevent it from 932 * overflowing or exceeding the maximum length 933 * allowed by the network interface: 934 */ 935 KASSERT(ipoptlen == 0, 936 ("%s: TSO can't do IP options", __func__)); 937 938 /* 939 * Check if we should limit by maximum payload 940 * length: 941 */ 942 if (if_hw_tsomax != 0) { 943 /* compute maximum TSO length */ 944 max_len = (if_hw_tsomax - hdrlen - 945 max_linkhdr); 946 if (max_len <= 0) { 947 len = 0; 948 } else if (len > max_len) { 949 sendalot = 1; 950 len = max_len; 951 } 952 } 953 954 /* 955 * Prevent the last segment from being 956 * fractional unless the send sockbuf can be 957 * emptied: 958 */ 959 max_len = (tp->t_maxseg - optlen); 960 if (((uint32_t)off + (uint32_t)len) < 961 sbavail(&so->so_snd)) { 962 moff = len % max_len; 963 if (moff != 0) { 964 len -= moff; 965 sendalot = 1; 966 } 967 } 968 969 /* 970 * In case there are too many small fragments 971 * don't use TSO: 972 */ 973 if (len <= max_len) { 974 len = max_len; 975 sendalot = 1; 976 tso = 0; 977 } 978 979 /* 980 * Send the FIN in a separate segment 981 * after the bulk sending is done. 982 * We don't trust the TSO implementations 983 * to clear the FIN flag on all but the 984 * last segment. 985 */ 986 if (tp->t_flags & TF_NEEDFIN) 987 sendalot = 1; 988 } else { 989 if (optlen + ipoptlen >= tp->t_maxseg) { 990 /* 991 * Since we don't have enough space to put 992 * the IP header chain and the TCP header in 993 * one packet as required by RFC 7112, don't 994 * send it. Also ensure that at least one 995 * byte of the payload can be put into the 996 * TCP segment. 997 */ 998 SOCKBUF_UNLOCK(&so->so_snd); 999 error = EMSGSIZE; 1000 sack_rxmit = 0; 1001 goto out; 1002 } 1003 len = tp->t_maxseg - optlen - ipoptlen; 1004 sendalot = 1; 1005 if (dont_sendalot) 1006 sendalot = 0; 1007 } 1008 } else 1009 tso = 0; 1010 1011 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 1012 ("%s: len > IP_MAXPACKET", __func__)); 1013 1014 /*#ifdef DIAGNOSTIC*/ 1015 #ifdef INET6 1016 if (max_linkhdr + hdrlen > MCLBYTES) 1017 #else 1018 if (max_linkhdr + hdrlen > MHLEN) 1019 #endif 1020 panic("tcphdr too big"); 1021 /*#endif*/ 1022 1023 /* 1024 * This KASSERT is here to catch edge cases at a well defined place. 1025 * Before, those had triggered (random) panic conditions further down. 1026 */ 1027 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 1028 1029 /* 1030 * Grab a header mbuf, attaching a copy of data to 1031 * be transmitted, and initialize the header from 1032 * the template for sends on this connection. 1033 */ 1034 if (len) { 1035 struct mbuf *mb; 1036 struct sockbuf *msb; 1037 u_int moff; 1038 1039 if ((tp->t_flags & TF_FORCEDATA) && len == 1) { 1040 TCPSTAT_INC(tcps_sndprobe); 1041 #ifdef STATS 1042 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1043 stats_voi_update_abs_u32(tp->t_stats, 1044 VOI_TCP_RETXPB, len); 1045 else 1046 stats_voi_update_abs_u64(tp->t_stats, 1047 VOI_TCP_TXPB, len); 1048 #endif /* STATS */ 1049 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 1050 tp->t_sndrexmitpack++; 1051 TCPSTAT_INC(tcps_sndrexmitpack); 1052 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 1053 #ifdef STATS 1054 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 1055 len); 1056 #endif /* STATS */ 1057 } else { 1058 TCPSTAT_INC(tcps_sndpack); 1059 TCPSTAT_ADD(tcps_sndbyte, len); 1060 #ifdef STATS 1061 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 1062 len); 1063 #endif /* STATS */ 1064 } 1065 #ifdef INET6 1066 if (MHLEN < hdrlen + max_linkhdr) 1067 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1068 else 1069 #endif 1070 m = m_gethdr(M_NOWAIT, MT_DATA); 1071 1072 if (m == NULL) { 1073 SOCKBUF_UNLOCK(&so->so_snd); 1074 error = ENOBUFS; 1075 sack_rxmit = 0; 1076 goto out; 1077 } 1078 1079 m->m_data += max_linkhdr; 1080 m->m_len = hdrlen; 1081 1082 /* 1083 * Start the m_copy functions from the closest mbuf 1084 * to the offset in the socket buffer chain. 1085 */ 1086 mb = sbsndptr_noadv(&so->so_snd, off, &moff); 1087 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 1088 m_copydata(mb, moff, len, 1089 mtod(m, caddr_t) + hdrlen); 1090 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1091 sbsndptr_adv(&so->so_snd, mb, len); 1092 m->m_len += len; 1093 } else { 1094 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1095 msb = NULL; 1096 else 1097 msb = &so->so_snd; 1098 m->m_next = tcp_m_copym(mb, moff, 1099 &len, if_hw_tsomaxsegcount, 1100 if_hw_tsomaxsegsize, msb, hw_tls); 1101 if (len <= (tp->t_maxseg - optlen)) { 1102 /* 1103 * Must have ran out of mbufs for the copy 1104 * shorten it to no longer need tso. Lets 1105 * not put on sendalot since we are low on 1106 * mbufs. 1107 */ 1108 tso = 0; 1109 } 1110 if (m->m_next == NULL) { 1111 SOCKBUF_UNLOCK(&so->so_snd); 1112 (void) m_free(m); 1113 error = ENOBUFS; 1114 sack_rxmit = 0; 1115 goto out; 1116 } 1117 } 1118 1119 /* 1120 * If we're sending everything we've got, set PUSH. 1121 * (This will keep happy those implementations which only 1122 * give data to the user when a buffer fills or 1123 * a PUSH comes in.) 1124 */ 1125 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) && 1126 !(flags & TH_SYN)) 1127 flags |= TH_PUSH; 1128 SOCKBUF_UNLOCK(&so->so_snd); 1129 } else { 1130 SOCKBUF_UNLOCK(&so->so_snd); 1131 if (tp->t_flags & TF_ACKNOW) 1132 TCPSTAT_INC(tcps_sndacks); 1133 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1134 TCPSTAT_INC(tcps_sndctrl); 1135 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1136 TCPSTAT_INC(tcps_sndurg); 1137 else 1138 TCPSTAT_INC(tcps_sndwinup); 1139 1140 m = m_gethdr(M_NOWAIT, MT_DATA); 1141 if (m == NULL) { 1142 error = ENOBUFS; 1143 sack_rxmit = 0; 1144 goto out; 1145 } 1146 #ifdef INET6 1147 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1148 MHLEN >= hdrlen) { 1149 M_ALIGN(m, hdrlen); 1150 } else 1151 #endif 1152 m->m_data += max_linkhdr; 1153 m->m_len = hdrlen; 1154 } 1155 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1156 m->m_pkthdr.rcvif = (struct ifnet *)0; 1157 #ifdef MAC 1158 mac_inpcb_create_mbuf(inp, m); 1159 #endif 1160 #ifdef INET6 1161 if (isipv6) { 1162 ip6 = mtod(m, struct ip6_hdr *); 1163 if (tp->t_port) { 1164 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); 1165 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1166 udp->uh_dport = tp->t_port; 1167 ulen = hdrlen + len - sizeof(struct ip6_hdr); 1168 udp->uh_ulen = htons(ulen); 1169 th = (struct tcphdr *)(udp + 1); 1170 } else { 1171 th = (struct tcphdr *)(ip6 + 1); 1172 } 1173 tcpip_fillheaders(inp, tp->t_port, ip6, th); 1174 } else 1175 #endif /* INET6 */ 1176 { 1177 ip = mtod(m, struct ip *); 1178 #ifdef TCPDEBUG 1179 ipov = (struct ipovly *)ip; 1180 #endif 1181 if (tp->t_port) { 1182 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); 1183 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1184 udp->uh_dport = tp->t_port; 1185 ulen = hdrlen + len - sizeof(struct ip); 1186 udp->uh_ulen = htons(ulen); 1187 th = (struct tcphdr *)(udp + 1); 1188 } else 1189 th = (struct tcphdr *)(ip + 1); 1190 tcpip_fillheaders(inp, tp->t_port, ip, th); 1191 } 1192 1193 /* 1194 * Fill in fields, remembering maximum advertised 1195 * window for use in delaying messages about window sizes. 1196 * If resending a FIN, be sure not to use a new sequence number. 1197 */ 1198 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1199 tp->snd_nxt == tp->snd_max) 1200 tp->snd_nxt--; 1201 /* 1202 * If we are starting a connection, send ECN setup 1203 * SYN packet. If we are on a retransmit, we may 1204 * resend those bits a number of times as per 1205 * RFC 3168. 1206 */ 1207 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 1208 flags |= tcp_ecn_output_syn_sent(tp); 1209 } 1210 /* Also handle parallel SYN for ECN */ 1211 if ((TCPS_HAVERCVDSYN(tp->t_state)) && 1212 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) { 1213 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit); 1214 if ((tp->t_state == TCPS_SYN_RECEIVED) && 1215 (tp->t_flags2 & TF2_ECN_SND_ECE)) 1216 tp->t_flags2 &= ~TF2_ECN_SND_ECE; 1217 #ifdef INET6 1218 if (isipv6) { 1219 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20); 1220 ip6->ip6_flow |= htonl(ect << 20); 1221 } 1222 else 1223 #endif 1224 { 1225 ip->ip_tos &= ~IPTOS_ECN_MASK; 1226 ip->ip_tos |= ect; 1227 } 1228 } 1229 1230 /* 1231 * If we are doing retransmissions, then snd_nxt will 1232 * not reflect the first unsent octet. For ACK only 1233 * packets, we do not want the sequence number of the 1234 * retransmitted packet, we want the sequence number 1235 * of the next unsent octet. So, if there is no data 1236 * (and no SYN or FIN), use snd_max instead of snd_nxt 1237 * when filling in ti_seq. But if we are in persist 1238 * state, snd_max might reflect one byte beyond the 1239 * right edge of the window, so use snd_nxt in that 1240 * case, since we know we aren't doing a retransmission. 1241 * (retransmit and persist are mutually exclusive...) 1242 */ 1243 if (sack_rxmit == 0) { 1244 if (len || (flags & (TH_SYN|TH_FIN)) || 1245 tcp_timer_active(tp, TT_PERSIST)) 1246 th->th_seq = htonl(tp->snd_nxt); 1247 else 1248 th->th_seq = htonl(tp->snd_max); 1249 } else { 1250 th->th_seq = htonl(p->rxmit); 1251 p->rxmit += len; 1252 /* 1253 * Lost Retransmission Detection 1254 * trigger resending of a (then 1255 * still existing) hole, when 1256 * fack acks recoverypoint. 1257 */ 1258 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end)) 1259 p->rxmit = tp->snd_recover; 1260 tp->sackhint.sack_bytes_rexmit += len; 1261 } 1262 if (IN_RECOVERY(tp->t_flags)) { 1263 /* 1264 * Account all bytes transmitted while 1265 * IN_RECOVERY, simplifying PRR and 1266 * Lost Retransmit Detection 1267 */ 1268 tp->sackhint.prr_out += len; 1269 } 1270 th->th_ack = htonl(tp->rcv_nxt); 1271 if (optlen) { 1272 bcopy(opt, th + 1, optlen); 1273 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1274 } 1275 tcp_set_flags(th, flags); 1276 /* 1277 * Calculate receive window. Don't shrink window, 1278 * but avoid silly window syndrome. 1279 * If a RST segment is sent, advertise a window of zero. 1280 */ 1281 if (flags & TH_RST) { 1282 recwin = 0; 1283 } else { 1284 if (recwin < (so->so_rcv.sb_hiwat / 4) && 1285 recwin < tp->t_maxseg) 1286 recwin = 0; 1287 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1288 recwin < (tp->rcv_adv - tp->rcv_nxt)) 1289 recwin = (tp->rcv_adv - tp->rcv_nxt); 1290 } 1291 /* 1292 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1293 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1294 * case is handled in syncache. 1295 */ 1296 if (flags & TH_SYN) 1297 th->th_win = htons((u_short) 1298 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1299 else { 1300 /* Avoid shrinking window with window scaling. */ 1301 recwin = roundup2(recwin, 1 << tp->rcv_scale); 1302 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1303 } 1304 1305 /* 1306 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1307 * a 0 window. This may cause the remote transmitter to stall. This 1308 * flag tells soreceive() to disable delayed acknowledgements when 1309 * draining the buffer. This can occur if the receiver is attempting 1310 * to read more data than can be buffered prior to transmitting on 1311 * the connection. 1312 */ 1313 if (th->th_win == 0) { 1314 tp->t_sndzerowin++; 1315 tp->t_flags |= TF_RXWIN0SENT; 1316 } else 1317 tp->t_flags &= ~TF_RXWIN0SENT; 1318 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1319 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1320 th->th_flags |= TH_URG; 1321 } else 1322 /* 1323 * If no urgent pointer to send, then we pull 1324 * the urgent pointer to the left edge of the send window 1325 * so that it doesn't drift into the send window on sequence 1326 * number wraparound. 1327 */ 1328 tp->snd_up = tp->snd_una; /* drag it along */ 1329 1330 /* 1331 * Put TCP length in extended header, and then 1332 * checksum extended header and data. 1333 */ 1334 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1335 1336 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1337 if (to.to_flags & TOF_SIGNATURE) { 1338 /* 1339 * Calculate MD5 signature and put it into the place 1340 * determined before. 1341 * NOTE: since TCP options buffer doesn't point into 1342 * mbuf's data, calculate offset and use it. 1343 */ 1344 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th, 1345 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) { 1346 /* 1347 * Do not send segment if the calculation of MD5 1348 * digest has failed. 1349 */ 1350 m_freem(m); 1351 goto out; 1352 } 1353 } 1354 #endif 1355 #ifdef INET6 1356 if (isipv6) { 1357 /* 1358 * There is no need to fill in ip6_plen right now. 1359 * It will be filled later by ip6_output. 1360 */ 1361 if (tp->t_port) { 1362 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 1363 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 1364 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 1365 th->th_sum = htons(0); 1366 UDPSTAT_INC(udps_opackets); 1367 } else { 1368 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1369 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1370 th->th_sum = in6_cksum_pseudo(ip6, 1371 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 1372 0); 1373 } 1374 } 1375 #endif 1376 #if defined(INET6) && defined(INET) 1377 else 1378 #endif 1379 #ifdef INET 1380 { 1381 if (tp->t_port) { 1382 m->m_pkthdr.csum_flags = CSUM_UDP; 1383 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 1384 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 1385 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 1386 th->th_sum = htons(0); 1387 UDPSTAT_INC(udps_opackets); 1388 } else { 1389 m->m_pkthdr.csum_flags = CSUM_TCP; 1390 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1391 th->th_sum = in_pseudo(ip->ip_src.s_addr, 1392 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 1393 IPPROTO_TCP + len + optlen)); 1394 } 1395 1396 /* IP version must be set here for ipv4/ipv6 checking later */ 1397 KASSERT(ip->ip_v == IPVERSION, 1398 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1399 } 1400 #endif 1401 1402 /* 1403 * Enable TSO and specify the size of the segments. 1404 * The TCP pseudo header checksum is always provided. 1405 */ 1406 if (tso) { 1407 KASSERT(len > tp->t_maxseg - optlen, 1408 ("%s: len <= tso_segsz", __func__)); 1409 m->m_pkthdr.csum_flags |= CSUM_TSO; 1410 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 1411 } 1412 1413 KASSERT(len + hdrlen == m_length(m, NULL), 1414 ("%s: mbuf chain shorter than expected: %d + %u != %u", 1415 __func__, len, hdrlen, m_length(m, NULL))); 1416 1417 #ifdef TCP_HHOOK 1418 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1419 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1420 #endif 1421 1422 #ifdef TCPDEBUG 1423 /* 1424 * Trace. 1425 */ 1426 if (so->so_options & SO_DEBUG) { 1427 u_short save = 0; 1428 #ifdef INET6 1429 if (!isipv6) 1430 #endif 1431 { 1432 save = ipov->ih_len; 1433 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1434 } 1435 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1436 #ifdef INET6 1437 if (!isipv6) 1438 #endif 1439 ipov->ih_len = save; 1440 } 1441 #endif /* TCPDEBUG */ 1442 TCP_PROBE3(debug__output, tp, th, m); 1443 1444 /* We're getting ready to send; log now. */ 1445 /* XXXMT: We are not honoring verbose logging. */ 1446 if (tp->t_logstate != TCP_LOG_STATE_OFF) 1447 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, 1448 TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0, 1449 NULL); 1450 else 1451 lgb = NULL; 1452 1453 /* 1454 * Fill in IP length and desired time to live and 1455 * send to IP level. There should be a better way 1456 * to handle ttl and tos; we could keep them in 1457 * the template, but need a way to checksum without them. 1458 */ 1459 /* 1460 * m->m_pkthdr.len should have been set before checksum calculation, 1461 * because in6_cksum() need it. 1462 */ 1463 #ifdef INET6 1464 if (isipv6) { 1465 /* 1466 * we separately set hoplimit for every segment, since the 1467 * user might want to change the value via setsockopt. 1468 * Also, desired default hop limit might be changed via 1469 * Neighbor Discovery. 1470 */ 1471 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 1472 1473 /* 1474 * Set the packet size here for the benefit of DTrace probes. 1475 * ip6_output() will set it properly; it's supposed to include 1476 * the option header lengths as well. 1477 */ 1478 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1479 1480 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 1481 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1482 else 1483 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1484 1485 if (tp->t_state == TCPS_SYN_SENT) 1486 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1487 1488 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1489 1490 #ifdef TCPPCAP 1491 /* Save packet, if requested. */ 1492 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1493 #endif 1494 1495 /* TODO: IPv6 IP6TOS_ECT bit on */ 1496 error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6, 1497 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1498 NULL, NULL, inp); 1499 1500 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) 1501 mtu = inp->inp_route6.ro_nh->nh_mtu; 1502 } 1503 #endif /* INET6 */ 1504 #if defined(INET) && defined(INET6) 1505 else 1506 #endif 1507 #ifdef INET 1508 { 1509 ip->ip_len = htons(m->m_pkthdr.len); 1510 #ifdef INET6 1511 if (inp->inp_vflag & INP_IPV6PROTO) 1512 ip->ip_ttl = in6_selecthlim(inp, NULL); 1513 #endif /* INET6 */ 1514 /* 1515 * If we do path MTU discovery, then we set DF on every packet. 1516 * This might not be the best thing to do according to RFC3390 1517 * Section 2. However the tcp hostcache migitates the problem 1518 * so it affects only the first tcp connection with a host. 1519 * 1520 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1521 */ 1522 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 1523 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1524 if (tp->t_port == 0 || len < V_tcp_minmss) { 1525 ip->ip_off |= htons(IP_DF); 1526 } 1527 } else { 1528 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1529 } 1530 1531 if (tp->t_state == TCPS_SYN_SENT) 1532 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1533 1534 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1535 1536 #ifdef TCPPCAP 1537 /* Save packet, if requested. */ 1538 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1539 #endif 1540 1541 error = ip_output(m, inp->inp_options, &inp->inp_route, 1542 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp); 1543 1544 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) 1545 mtu = inp->inp_route.ro_nh->nh_mtu; 1546 } 1547 #endif /* INET */ 1548 1549 if (lgb != NULL) { 1550 lgb->tlb_errno = error; 1551 lgb = NULL; 1552 } 1553 out: 1554 if (error == 0) 1555 tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls); 1556 /* 1557 * In transmit state, time the transmission and arrange for 1558 * the retransmit. In persist state, just set snd_max. 1559 */ 1560 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1561 !tcp_timer_active(tp, TT_PERSIST)) { 1562 tcp_seq startseq = tp->snd_nxt; 1563 1564 /* 1565 * Advance snd_nxt over sequence space of this segment. 1566 */ 1567 if (flags & (TH_SYN|TH_FIN)) { 1568 if (flags & TH_SYN) 1569 tp->snd_nxt++; 1570 if (flags & TH_FIN) { 1571 tp->snd_nxt++; 1572 tp->t_flags |= TF_SENTFIN; 1573 } 1574 } 1575 if (sack_rxmit) 1576 goto timer; 1577 tp->snd_nxt += len; 1578 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1579 /* 1580 * Update "made progress" indication if we just 1581 * added new data to an empty socket buffer. 1582 */ 1583 if (tp->snd_una == tp->snd_max) 1584 tp->t_acktime = ticks; 1585 tp->snd_max = tp->snd_nxt; 1586 /* 1587 * Time this transmission if not a retransmission and 1588 * not currently timing anything. 1589 */ 1590 tp->t_sndtime = ticks; 1591 if (tp->t_rtttime == 0) { 1592 tp->t_rtttime = ticks; 1593 tp->t_rtseq = startseq; 1594 TCPSTAT_INC(tcps_segstimed); 1595 } 1596 #ifdef STATS 1597 if (!(tp->t_flags & TF_GPUTINPROG) && len) { 1598 tp->t_flags |= TF_GPUTINPROG; 1599 tp->gput_seq = startseq; 1600 tp->gput_ack = startseq + 1601 ulmin(sbavail(&so->so_snd) - off, sendwin); 1602 tp->gput_ts = tcp_ts_getticks(); 1603 } 1604 #endif /* STATS */ 1605 } 1606 1607 /* 1608 * Set retransmit timer if not currently set, 1609 * and not doing a pure ack or a keep-alive probe. 1610 * Initial value for retransmit timer is smoothed 1611 * round-trip time + 2 * round-trip time variance. 1612 * Initialize shift counter which is used for backoff 1613 * of retransmit time. 1614 */ 1615 timer: 1616 if (!tcp_timer_active(tp, TT_REXMT) && 1617 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1618 (tp->snd_nxt != tp->snd_una))) { 1619 if (tcp_timer_active(tp, TT_PERSIST)) { 1620 tcp_timer_activate(tp, TT_PERSIST, 0); 1621 tp->t_rxtshift = 0; 1622 } 1623 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp)); 1624 } else if (len == 0 && sbavail(&so->so_snd) && 1625 !tcp_timer_active(tp, TT_REXMT) && 1626 !tcp_timer_active(tp, TT_PERSIST)) { 1627 /* 1628 * Avoid a situation where we do not set persist timer 1629 * after a zero window condition. For example: 1630 * 1) A -> B: packet with enough data to fill the window 1631 * 2) B -> A: ACK for #1 + new data (0 window 1632 * advertisement) 1633 * 3) A -> B: ACK for #2, 0 len packet 1634 * 1635 * In this case, A will not activate the persist timer, 1636 * because it chose to send a packet. Unless tcp_output 1637 * is called for some other reason (delayed ack timer, 1638 * another input packet from B, socket syscall), A will 1639 * not send zero window probes. 1640 * 1641 * So, if you send a 0-length packet, but there is data 1642 * in the socket buffer, and neither the rexmt or 1643 * persist timer is already set, then activate the 1644 * persist timer. 1645 */ 1646 tp->t_rxtshift = 0; 1647 tcp_setpersist(tp); 1648 } 1649 } else { 1650 /* 1651 * Persist case, update snd_max but since we are in 1652 * persist mode (no window) we do not update snd_nxt. 1653 */ 1654 int xlen = len; 1655 if (flags & TH_SYN) 1656 ++xlen; 1657 if (flags & TH_FIN) { 1658 ++xlen; 1659 tp->t_flags |= TF_SENTFIN; 1660 } 1661 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1662 tp->snd_max = tp->snd_nxt + xlen; 1663 } 1664 if ((error == 0) && 1665 (TCPS_HAVEESTABLISHED(tp->t_state) && 1666 (tp->t_flags & TF_SACK_PERMIT) && 1667 tp->rcv_numsacks > 0)) { 1668 /* Clean up any DSACK's sent */ 1669 tcp_clean_dsack_blocks(tp); 1670 } 1671 if (error) { 1672 /* 1673 * We know that the packet was lost, so back out the 1674 * sequence number advance, if any. 1675 * 1676 * If the error is EPERM the packet got blocked by the 1677 * local firewall. Normally we should terminate the 1678 * connection but the blocking may have been spurious 1679 * due to a firewall reconfiguration cycle. So we treat 1680 * it like a packet loss and let the retransmit timer and 1681 * timeouts do their work over time. 1682 * XXX: It is a POLA question whether calling tcp_drop right 1683 * away would be the really correct behavior instead. 1684 */ 1685 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1686 !tcp_timer_active(tp, TT_PERSIST)) && 1687 ((flags & TH_SYN) == 0) && 1688 (error != EPERM)) { 1689 if (sack_rxmit) { 1690 p->rxmit -= len; 1691 tp->sackhint.sack_bytes_rexmit -= len; 1692 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1693 ("sackhint bytes rtx >= 0")); 1694 KASSERT((flags & TH_FIN) == 0, 1695 ("error while FIN with SACK rxmit")); 1696 } else { 1697 tp->snd_nxt -= len; 1698 if (flags & TH_FIN) 1699 tp->snd_nxt--; 1700 } 1701 } 1702 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1703 switch (error) { 1704 case EACCES: 1705 case EPERM: 1706 tp->t_softerror = error; 1707 return (error); 1708 case ENOBUFS: 1709 TCP_XMIT_TIMER_ASSERT(tp, len, flags); 1710 tp->snd_cwnd = tp->t_maxseg; 1711 return (0); 1712 case EMSGSIZE: 1713 /* 1714 * For some reason the interface we used initially 1715 * to send segments changed to another or lowered 1716 * its MTU. 1717 * If TSO was active we either got an interface 1718 * without TSO capabilits or TSO was turned off. 1719 * If we obtained mtu from ip_output() then update 1720 * it and try again. 1721 */ 1722 if (tso) 1723 tp->t_flags &= ~TF_TSO; 1724 if (mtu != 0) { 1725 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1726 goto again; 1727 } 1728 return (error); 1729 case EHOSTDOWN: 1730 case EHOSTUNREACH: 1731 case ENETDOWN: 1732 case ENETUNREACH: 1733 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1734 tp->t_softerror = error; 1735 return (0); 1736 } 1737 /* FALLTHROUGH */ 1738 default: 1739 return (error); 1740 } 1741 } 1742 TCPSTAT_INC(tcps_sndtotal); 1743 1744 /* 1745 * Data sent (as far as we can tell). 1746 * If this advertises a larger window than any other segment, 1747 * then remember the size of the advertised window. 1748 * Any pending ACK has now been sent. 1749 */ 1750 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1751 tp->rcv_adv = tp->rcv_nxt + recwin; 1752 tp->last_ack_sent = tp->rcv_nxt; 1753 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1754 if (tcp_timer_active(tp, TT_DELACK)) 1755 tcp_timer_activate(tp, TT_DELACK, 0); 1756 #if 0 1757 /* 1758 * This completely breaks TCP if newreno is turned on. What happens 1759 * is that if delayed-acks are turned on on the receiver, this code 1760 * on the transmitter effectively destroys the TCP window, forcing 1761 * it to four packets (1.5Kx4 = 6K window). 1762 */ 1763 if (sendalot && --maxburst) 1764 goto again; 1765 #endif 1766 if (sendalot) 1767 goto again; 1768 return (0); 1769 } 1770 1771 void 1772 tcp_setpersist(struct tcpcb *tp) 1773 { 1774 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1775 int tt; 1776 int maxunacktime; 1777 1778 tp->t_flags &= ~TF_PREVVALID; 1779 if (tcp_timer_active(tp, TT_REXMT)) 1780 panic("tcp_setpersist: retransmit pending"); 1781 /* 1782 * If the state is already closed, don't bother. 1783 */ 1784 if (tp->t_state == TCPS_CLOSED) 1785 return; 1786 1787 /* 1788 * Start/restart persistence timer. 1789 */ 1790 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1791 tcp_persmin, tcp_persmax); 1792 if (TP_MAXUNACKTIME(tp) && tp->t_acktime) { 1793 maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks; 1794 if (maxunacktime < 1) 1795 maxunacktime = 1; 1796 if (maxunacktime < tt) 1797 tt = maxunacktime; 1798 } 1799 tcp_timer_activate(tp, TT_PERSIST, tt); 1800 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1801 tp->t_rxtshift++; 1802 } 1803 1804 /* 1805 * Insert TCP options according to the supplied parameters to the place 1806 * optp in a consistent way. Can handle unaligned destinations. 1807 * 1808 * The order of the option processing is crucial for optimal packing and 1809 * alignment for the scarce option space. 1810 * 1811 * The optimal order for a SYN/SYN-ACK segment is: 1812 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1813 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1814 * 1815 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1816 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1817 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1818 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1819 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1820 */ 1821 int 1822 tcp_addoptions(struct tcpopt *to, u_char *optp) 1823 { 1824 u_int32_t mask, optlen = 0; 1825 1826 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1827 if ((to->to_flags & mask) != mask) 1828 continue; 1829 if (optlen == TCP_MAXOLEN) 1830 break; 1831 switch (to->to_flags & mask) { 1832 case TOF_MSS: 1833 while (optlen % 4) { 1834 optlen += TCPOLEN_NOP; 1835 *optp++ = TCPOPT_NOP; 1836 } 1837 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1838 continue; 1839 optlen += TCPOLEN_MAXSEG; 1840 *optp++ = TCPOPT_MAXSEG; 1841 *optp++ = TCPOLEN_MAXSEG; 1842 to->to_mss = htons(to->to_mss); 1843 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1844 optp += sizeof(to->to_mss); 1845 break; 1846 case TOF_SCALE: 1847 while (!optlen || optlen % 2 != 1) { 1848 optlen += TCPOLEN_NOP; 1849 *optp++ = TCPOPT_NOP; 1850 } 1851 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1852 continue; 1853 optlen += TCPOLEN_WINDOW; 1854 *optp++ = TCPOPT_WINDOW; 1855 *optp++ = TCPOLEN_WINDOW; 1856 *optp++ = to->to_wscale; 1857 break; 1858 case TOF_SACKPERM: 1859 while (optlen % 2) { 1860 optlen += TCPOLEN_NOP; 1861 *optp++ = TCPOPT_NOP; 1862 } 1863 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1864 continue; 1865 optlen += TCPOLEN_SACK_PERMITTED; 1866 *optp++ = TCPOPT_SACK_PERMITTED; 1867 *optp++ = TCPOLEN_SACK_PERMITTED; 1868 break; 1869 case TOF_TS: 1870 while (!optlen || optlen % 4 != 2) { 1871 optlen += TCPOLEN_NOP; 1872 *optp++ = TCPOPT_NOP; 1873 } 1874 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1875 continue; 1876 optlen += TCPOLEN_TIMESTAMP; 1877 *optp++ = TCPOPT_TIMESTAMP; 1878 *optp++ = TCPOLEN_TIMESTAMP; 1879 to->to_tsval = htonl(to->to_tsval); 1880 to->to_tsecr = htonl(to->to_tsecr); 1881 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1882 optp += sizeof(to->to_tsval); 1883 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1884 optp += sizeof(to->to_tsecr); 1885 break; 1886 case TOF_SIGNATURE: 1887 { 1888 int siglen = TCPOLEN_SIGNATURE - 2; 1889 1890 while (!optlen || optlen % 4 != 2) { 1891 optlen += TCPOLEN_NOP; 1892 *optp++ = TCPOPT_NOP; 1893 } 1894 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) { 1895 to->to_flags &= ~TOF_SIGNATURE; 1896 continue; 1897 } 1898 optlen += TCPOLEN_SIGNATURE; 1899 *optp++ = TCPOPT_SIGNATURE; 1900 *optp++ = TCPOLEN_SIGNATURE; 1901 to->to_signature = optp; 1902 while (siglen--) 1903 *optp++ = 0; 1904 break; 1905 } 1906 case TOF_SACK: 1907 { 1908 int sackblks = 0; 1909 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1910 tcp_seq sack_seq; 1911 1912 while (!optlen || optlen % 4 != 2) { 1913 optlen += TCPOLEN_NOP; 1914 *optp++ = TCPOPT_NOP; 1915 } 1916 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1917 continue; 1918 optlen += TCPOLEN_SACKHDR; 1919 *optp++ = TCPOPT_SACK; 1920 sackblks = min(to->to_nsacks, 1921 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1922 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1923 while (sackblks--) { 1924 sack_seq = htonl(sack->start); 1925 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1926 optp += sizeof(sack_seq); 1927 sack_seq = htonl(sack->end); 1928 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1929 optp += sizeof(sack_seq); 1930 optlen += TCPOLEN_SACK; 1931 sack++; 1932 } 1933 TCPSTAT_INC(tcps_sack_send_blocks); 1934 break; 1935 } 1936 case TOF_FASTOPEN: 1937 { 1938 int total_len; 1939 1940 /* XXX is there any point to aligning this option? */ 1941 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; 1942 if (TCP_MAXOLEN - optlen < total_len) { 1943 to->to_flags &= ~TOF_FASTOPEN; 1944 continue; 1945 } 1946 *optp++ = TCPOPT_FAST_OPEN; 1947 *optp++ = total_len; 1948 if (to->to_tfo_len > 0) { 1949 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); 1950 optp += to->to_tfo_len; 1951 } 1952 optlen += total_len; 1953 break; 1954 } 1955 default: 1956 panic("%s: unknown TCP option type", __func__); 1957 break; 1958 } 1959 } 1960 1961 /* Terminate and pad TCP options to a 4 byte boundary. */ 1962 if (optlen % 4) { 1963 optlen += TCPOLEN_EOL; 1964 *optp++ = TCPOPT_EOL; 1965 } 1966 /* 1967 * According to RFC 793 (STD0007): 1968 * "The content of the header beyond the End-of-Option option 1969 * must be header padding (i.e., zero)." 1970 * and later: "The padding is composed of zeros." 1971 */ 1972 while (optlen % 4) { 1973 optlen += TCPOLEN_PAD; 1974 *optp++ = TCPOPT_PAD; 1975 } 1976 1977 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1978 return (optlen); 1979 } 1980 1981 /* 1982 * This is a copy of m_copym(), taking the TSO segment size/limit 1983 * constraints into account, and advancing the sndptr as it goes. 1984 */ 1985 struct mbuf * 1986 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen, 1987 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls) 1988 { 1989 #ifdef KERN_TLS 1990 struct ktls_session *tls, *ntls; 1991 struct mbuf *start __diagused; 1992 #endif 1993 struct mbuf *n, **np; 1994 struct mbuf *top; 1995 int32_t off = off0; 1996 int32_t len = *plen; 1997 int32_t fragsize; 1998 int32_t len_cp = 0; 1999 int32_t *pkthdrlen; 2000 uint32_t mlen, frags; 2001 bool copyhdr; 2002 2003 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off)); 2004 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len)); 2005 if (off == 0 && m->m_flags & M_PKTHDR) 2006 copyhdr = true; 2007 else 2008 copyhdr = false; 2009 while (off > 0) { 2010 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain")); 2011 if (off < m->m_len) 2012 break; 2013 off -= m->m_len; 2014 if ((sb) && (m == sb->sb_sndptr)) { 2015 sb->sb_sndptroff += m->m_len; 2016 sb->sb_sndptr = m->m_next; 2017 } 2018 m = m->m_next; 2019 } 2020 np = ⊤ 2021 top = NULL; 2022 pkthdrlen = NULL; 2023 #ifdef KERN_TLS 2024 if (hw_tls && (m->m_flags & M_EXTPG)) 2025 tls = m->m_epg_tls; 2026 else 2027 tls = NULL; 2028 start = m; 2029 #endif 2030 while (len > 0) { 2031 if (m == NULL) { 2032 KASSERT(len == M_COPYALL, 2033 ("tcp_m_copym, length > size of mbuf chain")); 2034 *plen = len_cp; 2035 if (pkthdrlen != NULL) 2036 *pkthdrlen = len_cp; 2037 break; 2038 } 2039 #ifdef KERN_TLS 2040 if (hw_tls) { 2041 if (m->m_flags & M_EXTPG) 2042 ntls = m->m_epg_tls; 2043 else 2044 ntls = NULL; 2045 2046 /* 2047 * Avoid mixing TLS records with handshake 2048 * data or TLS records from different 2049 * sessions. 2050 */ 2051 if (tls != ntls) { 2052 MPASS(m != start); 2053 *plen = len_cp; 2054 if (pkthdrlen != NULL) 2055 *pkthdrlen = len_cp; 2056 break; 2057 } 2058 } 2059 #endif 2060 mlen = min(len, m->m_len - off); 2061 if (seglimit) { 2062 /* 2063 * For M_EXTPG mbufs, add 3 segments 2064 * + 1 in case we are crossing page boundaries 2065 * + 2 in case the TLS hdr/trailer are used 2066 * It is cheaper to just add the segments 2067 * than it is to take the cache miss to look 2068 * at the mbuf ext_pgs state in detail. 2069 */ 2070 if (m->m_flags & M_EXTPG) { 2071 fragsize = min(segsize, PAGE_SIZE); 2072 frags = 3; 2073 } else { 2074 fragsize = segsize; 2075 frags = 0; 2076 } 2077 2078 /* Break if we really can't fit anymore. */ 2079 if ((frags + 1) >= seglimit) { 2080 *plen = len_cp; 2081 if (pkthdrlen != NULL) 2082 *pkthdrlen = len_cp; 2083 break; 2084 } 2085 2086 /* 2087 * Reduce size if you can't copy the whole 2088 * mbuf. If we can't copy the whole mbuf, also 2089 * adjust len so the loop will end after this 2090 * mbuf. 2091 */ 2092 if ((frags + howmany(mlen, fragsize)) >= seglimit) { 2093 mlen = (seglimit - frags - 1) * fragsize; 2094 len = mlen; 2095 *plen = len_cp + len; 2096 if (pkthdrlen != NULL) 2097 *pkthdrlen = *plen; 2098 } 2099 frags += howmany(mlen, fragsize); 2100 if (frags == 0) 2101 frags++; 2102 seglimit -= frags; 2103 KASSERT(seglimit > 0, 2104 ("%s: seglimit went too low", __func__)); 2105 } 2106 if (copyhdr) 2107 n = m_gethdr(M_NOWAIT, m->m_type); 2108 else 2109 n = m_get(M_NOWAIT, m->m_type); 2110 *np = n; 2111 if (n == NULL) 2112 goto nospace; 2113 if (copyhdr) { 2114 if (!m_dup_pkthdr(n, m, M_NOWAIT)) 2115 goto nospace; 2116 if (len == M_COPYALL) 2117 n->m_pkthdr.len -= off0; 2118 else 2119 n->m_pkthdr.len = len; 2120 pkthdrlen = &n->m_pkthdr.len; 2121 copyhdr = false; 2122 } 2123 n->m_len = mlen; 2124 len_cp += n->m_len; 2125 if (m->m_flags & (M_EXT|M_EXTPG)) { 2126 n->m_data = m->m_data + off; 2127 mb_dupcl(n, m); 2128 } else 2129 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 2130 (u_int)n->m_len); 2131 2132 if (sb && (sb->sb_sndptr == m) && 2133 ((n->m_len + off) >= m->m_len) && m->m_next) { 2134 sb->sb_sndptroff += m->m_len; 2135 sb->sb_sndptr = m->m_next; 2136 } 2137 off = 0; 2138 if (len != M_COPYALL) { 2139 len -= n->m_len; 2140 } 2141 m = m->m_next; 2142 np = &n->m_next; 2143 } 2144 return (top); 2145 nospace: 2146 m_freem(top); 2147 return (NULL); 2148 } 2149 2150 void 2151 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin) 2152 { 2153 2154 /* 2155 * Automatic sizing of send socket buffer. Often the send buffer 2156 * size is not optimally adjusted to the actual network conditions 2157 * at hand (delay bandwidth product). Setting the buffer size too 2158 * small limits throughput on links with high bandwidth and high 2159 * delay (eg. trans-continental/oceanic links). Setting the 2160 * buffer size too big consumes too much real kernel memory, 2161 * especially with many connections on busy servers. 2162 * 2163 * The criteria to step up the send buffer one notch are: 2164 * 1. receive window of remote host is larger than send buffer 2165 * (with a fudge factor of 5/4th); 2166 * 2. send buffer is filled to 7/8th with data (so we actually 2167 * have data to make use of it); 2168 * 3. send buffer fill has not hit maximal automatic size; 2169 * 4. our send window (slow start and cogestion controlled) is 2170 * larger than sent but unacknowledged data in send buffer. 2171 * 2172 * The remote host receive window scaling factor may limit the 2173 * growing of the send buffer before it reaches its allowed 2174 * maximum. 2175 * 2176 * It scales directly with slow start or congestion window 2177 * and does at most one step per received ACK. This fast 2178 * scaling has the drawback of growing the send buffer beyond 2179 * what is strictly necessary to make full use of a given 2180 * delay*bandwidth product. However testing has shown this not 2181 * to be much of an problem. At worst we are trading wasting 2182 * of available bandwidth (the non-use of it) for wasting some 2183 * socket buffer memory. 2184 * 2185 * TODO: Shrink send buffer during idle periods together 2186 * with congestion window. Requires another timer. Has to 2187 * wait for upcoming tcp timer rewrite. 2188 * 2189 * XXXGL: should there be used sbused() or sbavail()? 2190 */ 2191 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 2192 int lowat; 2193 2194 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0; 2195 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat && 2196 sbused(&so->so_snd) >= 2197 (so->so_snd.sb_hiwat / 8 * 7) - lowat && 2198 sbused(&so->so_snd) < V_tcp_autosndbuf_max && 2199 sendwin >= (sbused(&so->so_snd) - 2200 (tp->snd_nxt - tp->snd_una))) { 2201 if (!sbreserve_locked(so, SO_SND, 2202 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 2203 V_tcp_autosndbuf_max), curthread)) 2204 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 2205 } 2206 } 2207 } 2208