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 static void inline cc_after_idle(struct tcpcb *tp); 159 160 #ifdef TCP_HHOOK 161 /* 162 * Wrapper for the TCP established output helper hook. 163 */ 164 void 165 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 166 struct tcpopt *to, uint32_t len, int tso) 167 { 168 struct tcp_hhook_data hhook_data; 169 170 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 171 hhook_data.tp = tp; 172 hhook_data.th = th; 173 hhook_data.to = to; 174 hhook_data.len = len; 175 hhook_data.tso = tso; 176 177 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 178 tp->osd); 179 } 180 } 181 #endif 182 183 /* 184 * CC wrapper hook functions 185 */ 186 static void inline 187 cc_after_idle(struct tcpcb *tp) 188 { 189 INP_WLOCK_ASSERT(tp->t_inpcb); 190 191 if (CC_ALGO(tp)->after_idle != NULL) 192 CC_ALGO(tp)->after_idle(tp->ccv); 193 } 194 195 /* 196 * Tcp output routine: figure out what should be sent and send it. 197 */ 198 int 199 tcp_default_output(struct tcpcb *tp) 200 { 201 struct socket *so = tp->t_inpcb->inp_socket; 202 int32_t len; 203 uint32_t recwin, sendwin; 204 uint16_t flags; 205 int off, error = 0; /* Keep compiler happy */ 206 u_int if_hw_tsomaxsegcount = 0; 207 u_int if_hw_tsomaxsegsize = 0; 208 struct mbuf *m; 209 struct ip *ip = NULL; 210 #ifdef TCPDEBUG 211 struct ipovly *ipov = NULL; 212 #endif 213 struct tcphdr *th; 214 u_char opt[TCP_MAXOLEN]; 215 unsigned ipoptlen, optlen, hdrlen, ulen; 216 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 217 unsigned ipsec_optlen = 0; 218 #endif 219 int idle, sendalot, curticks; 220 int sack_rxmit, sack_bytes_rxmt; 221 struct sackhole *p; 222 int tso, mtu; 223 struct tcpopt to; 224 struct udphdr *udp = NULL; 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 = (tp->t_inpcb->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(tp->t_inpcb); 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 off = SEQ_SUB(p->rxmit, tp->snd_una); 341 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 342 __func__, off)); 343 if (len > 0) { 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 - 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_activate(tp, TT_REXMT, 0); 510 tp->t_rxtshift = 0; 511 tp->snd_nxt = tp->snd_una; 512 if (!tcp_timer_active(tp, TT_PERSIST)) 513 tcp_setpersist(tp); 514 } 515 } 516 517 /* len will be >= 0 after this point. */ 518 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 519 520 tcp_sndbuf_autoscale(tp, so, sendwin); 521 522 /* 523 * Decide if we can use TCP Segmentation Offloading (if supported by 524 * hardware). 525 * 526 * TSO may only be used if we are in a pure bulk sending state. The 527 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 528 * IP options prevent using TSO. With TSO the TCP header is the same 529 * (except for the sequence number) for all generated packets. This 530 * makes it impossible to transmit any options which vary per generated 531 * segment or packet. 532 * 533 * IPv4 handling has a clear separation of ip options and ip header 534 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 535 * the right thing below to provide length of just ip options and thus 536 * checking for ipoptlen is enough to decide if ip options are present. 537 */ 538 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 539 /* 540 * Pre-calculate here as we save another lookup into the darknesses 541 * of IPsec that way and can actually decide if TSO is ok. 542 */ 543 #ifdef INET6 544 if (isipv6 && IPSEC_ENABLED(ipv6)) 545 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); 546 #ifdef INET 547 else 548 #endif 549 #endif /* INET6 */ 550 #ifdef INET 551 if (IPSEC_ENABLED(ipv4)) 552 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); 553 #endif /* INET */ 554 #endif /* IPSEC */ 555 #ifdef INET6 556 if (isipv6) 557 ipoptlen = ip6_optlen(tp->t_inpcb); 558 else 559 #endif 560 if (tp->t_inpcb->inp_options) 561 ipoptlen = tp->t_inpcb->inp_options->m_len - 562 offsetof(struct ipoption, ipopt_list); 563 else 564 ipoptlen = 0; 565 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 566 ipoptlen += ipsec_optlen; 567 #endif 568 569 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 570 (tp->t_port == 0) && 571 ((tp->t_flags & TF_SIGNATURE) == 0) && 572 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 573 ipoptlen == 0 && !(flags & TH_SYN)) 574 tso = 1; 575 576 if (sack_rxmit) { 577 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd))) 578 flags &= ~TH_FIN; 579 } else { 580 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 581 sbused(&so->so_snd))) 582 flags &= ~TH_FIN; 583 } 584 585 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 586 (long)TCP_MAXWIN << tp->rcv_scale); 587 588 /* 589 * Sender silly window avoidance. We transmit under the following 590 * conditions when len is non-zero: 591 * 592 * - We have a full segment (or more with TSO) 593 * - This is the last buffer in a write()/send() and we are 594 * either idle or running NODELAY 595 * - we've timed out (e.g. persist timer) 596 * - we have more then 1/2 the maximum send window's worth of 597 * data (receiver may be limited the window size) 598 * - we need to retransmit 599 */ 600 if (len) { 601 if (len >= tp->t_maxseg) 602 goto send; 603 /* 604 * As the TCP header options are now 605 * considered when setting up the initial 606 * window, we would not send the last segment 607 * if we skip considering the option length here. 608 * Note: this may not work when tcp headers change 609 * very dynamically in the future. 610 */ 611 if ((((tp->t_flags & TF_SIGNATURE) ? 612 PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) + 613 ((tp->t_flags & TF_RCVD_TSTMP) ? 614 PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) + 615 len) >= tp->t_maxseg) 616 goto send; 617 /* 618 * NOTE! on localhost connections an 'ack' from the remote 619 * end may occur synchronously with the output and cause 620 * us to flush a buffer queued with moretocome. XXX 621 * 622 * note: the len + off check is almost certainly unnecessary. 623 */ 624 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 625 (idle || (tp->t_flags & TF_NODELAY)) && 626 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) && 627 (tp->t_flags & TF_NOPUSH) == 0) { 628 goto send; 629 } 630 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 631 goto send; 632 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 633 goto send; 634 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 635 goto send; 636 if (sack_rxmit) 637 goto send; 638 } 639 640 /* 641 * Sending of standalone window updates. 642 * 643 * Window updates are important when we close our window due to a 644 * full socket buffer and are opening it again after the application 645 * reads data from it. Once the window has opened again and the 646 * remote end starts to send again the ACK clock takes over and 647 * provides the most current window information. 648 * 649 * We must avoid the silly window syndrome whereas every read 650 * from the receive buffer, no matter how small, causes a window 651 * update to be sent. We also should avoid sending a flurry of 652 * window updates when the socket buffer had queued a lot of data 653 * and the application is doing small reads. 654 * 655 * Prevent a flurry of pointless window updates by only sending 656 * an update when we can increase the advertized window by more 657 * than 1/4th of the socket buffer capacity. When the buffer is 658 * getting full or is very small be more aggressive and send an 659 * update whenever we can increase by two mss sized segments. 660 * In all other situations the ACK's to new incoming data will 661 * carry further window increases. 662 * 663 * Don't send an independent window update if a delayed 664 * ACK is pending (it will get piggy-backed on it) or the 665 * remote side already has done a half-close and won't send 666 * more data. Skip this if the connection is in T/TCP 667 * half-open state. 668 */ 669 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 670 !(tp->t_flags & TF_DELACK) && 671 !TCPS_HAVERCVDFIN(tp->t_state)) { 672 /* 673 * "adv" is the amount we could increase the window, 674 * taking into account that we are limited by 675 * TCP_MAXWIN << tp->rcv_scale. 676 */ 677 int32_t adv; 678 int oldwin; 679 680 adv = recwin; 681 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 682 oldwin = (tp->rcv_adv - tp->rcv_nxt); 683 if (adv > oldwin) 684 adv -= oldwin; 685 else 686 adv = 0; 687 } else 688 oldwin = 0; 689 690 /* 691 * If the new window size ends up being the same as or less 692 * than the old size when it is scaled, then don't force 693 * a window update. 694 */ 695 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 696 goto dontupdate; 697 698 if (adv >= (int32_t)(2 * tp->t_maxseg) && 699 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 700 recwin <= (so->so_rcv.sb_hiwat / 8) || 701 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg || 702 adv >= TCP_MAXWIN << tp->rcv_scale)) 703 goto send; 704 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat) 705 goto send; 706 } 707 dontupdate: 708 709 /* 710 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 711 * is also a catch-all for the retransmit timer timeout case. 712 */ 713 if (tp->t_flags & TF_ACKNOW) 714 goto send; 715 if ((flags & TH_RST) || 716 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 717 goto send; 718 if (SEQ_GT(tp->snd_up, tp->snd_una)) 719 goto send; 720 /* 721 * If our state indicates that FIN should be sent 722 * and we have not yet done so, then we need to send. 723 */ 724 if (flags & TH_FIN && 725 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 726 goto send; 727 /* 728 * In SACK, it is possible for tcp_output to fail to send a segment 729 * after the retransmission timer has been turned off. Make sure 730 * that the retransmission timer is set. 731 */ 732 if ((tp->t_flags & TF_SACK_PERMIT) && 733 SEQ_GT(tp->snd_max, tp->snd_una) && 734 !tcp_timer_active(tp, TT_REXMT) && 735 !tcp_timer_active(tp, TT_PERSIST)) { 736 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 737 goto just_return; 738 } 739 /* 740 * TCP window updates are not reliable, rather a polling protocol 741 * using ``persist'' packets is used to insure receipt of window 742 * updates. The three ``states'' for the output side are: 743 * idle not doing retransmits or persists 744 * persisting to move a small or zero window 745 * (re)transmitting and thereby not persisting 746 * 747 * tcp_timer_active(tp, TT_PERSIST) 748 * is true when we are in persist state. 749 * (tp->t_flags & TF_FORCEDATA) 750 * is set when we are called to send a persist packet. 751 * tcp_timer_active(tp, TT_REXMT) 752 * is set when we are retransmitting 753 * The output side is idle when both timers are zero. 754 * 755 * If send window is too small, there is data to transmit, and no 756 * retransmit or persist is pending, then go to persist state. 757 * If nothing happens soon, send when timer expires: 758 * if window is nonzero, transmit what we can, 759 * otherwise force out a byte. 760 */ 761 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && 762 !tcp_timer_active(tp, TT_PERSIST)) { 763 tp->t_rxtshift = 0; 764 tcp_setpersist(tp); 765 } 766 767 /* 768 * No reason to send a segment, just return. 769 */ 770 just_return: 771 SOCKBUF_UNLOCK(&so->so_snd); 772 return (0); 773 774 send: 775 SOCKBUF_LOCK_ASSERT(&so->so_snd); 776 if (len > 0) { 777 if (len >= tp->t_maxseg) 778 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 779 else 780 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 781 } 782 /* 783 * Before ESTABLISHED, force sending of initial options 784 * unless TCP set not to do any options. 785 * NOTE: we assume that the IP/TCP header plus TCP options 786 * always fit in a single mbuf, leaving room for a maximum 787 * link header, i.e. 788 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 789 */ 790 optlen = 0; 791 #ifdef INET6 792 if (isipv6) 793 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 794 else 795 #endif 796 hdrlen = sizeof (struct tcpiphdr); 797 798 if (flags & TH_SYN) { 799 tp->snd_nxt = tp->iss; 800 } 801 802 /* 803 * Compute options for segment. 804 * We only have to care about SYN and established connection 805 * segments. Options for SYN-ACK segments are handled in TCP 806 * syncache. 807 */ 808 to.to_flags = 0; 809 if ((tp->t_flags & TF_NOOPT) == 0) { 810 /* Maximum segment size. */ 811 if (flags & TH_SYN) { 812 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 813 if (tp->t_port) 814 to.to_mss -= V_tcp_udp_tunneling_overhead; 815 to.to_flags |= TOF_MSS; 816 817 /* 818 * On SYN or SYN|ACK transmits on TFO connections, 819 * only include the TFO option if it is not a 820 * retransmit, as the presence of the TFO option may 821 * have caused the original SYN or SYN|ACK to have 822 * been dropped by a middlebox. 823 */ 824 if (IS_FASTOPEN(tp->t_flags) && 825 (tp->t_rxtshift == 0)) { 826 if (tp->t_state == TCPS_SYN_RECEIVED) { 827 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 828 to.to_tfo_cookie = 829 (u_int8_t *)&tp->t_tfo_cookie.server; 830 to.to_flags |= TOF_FASTOPEN; 831 wanted_cookie = 1; 832 } else if (tp->t_state == TCPS_SYN_SENT) { 833 to.to_tfo_len = 834 tp->t_tfo_client_cookie_len; 835 to.to_tfo_cookie = 836 tp->t_tfo_cookie.client; 837 to.to_flags |= TOF_FASTOPEN; 838 wanted_cookie = 1; 839 /* 840 * If we wind up having more data to 841 * send with the SYN than can fit in 842 * one segment, don't send any more 843 * until the SYN|ACK comes back from 844 * the other end. 845 */ 846 dont_sendalot = 1; 847 } 848 } 849 } 850 /* Window scaling. */ 851 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 852 to.to_wscale = tp->request_r_scale; 853 to.to_flags |= TOF_SCALE; 854 } 855 /* Timestamps. */ 856 if ((tp->t_flags & TF_RCVD_TSTMP) || 857 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 858 curticks = tcp_ts_getticks(); 859 to.to_tsval = curticks + tp->ts_offset; 860 to.to_tsecr = tp->ts_recent; 861 to.to_flags |= TOF_TS; 862 if (tp->t_rxtshift == 1) 863 tp->t_badrxtwin = curticks; 864 } 865 866 /* Set receive buffer autosizing timestamp. */ 867 if (tp->rfbuf_ts == 0 && 868 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 869 tp->rfbuf_ts = tcp_ts_getticks(); 870 871 /* Selective ACK's. */ 872 if (tp->t_flags & TF_SACK_PERMIT) { 873 if (flags & TH_SYN) 874 to.to_flags |= TOF_SACKPERM; 875 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 876 tp->rcv_numsacks > 0) { 877 to.to_flags |= TOF_SACK; 878 to.to_nsacks = tp->rcv_numsacks; 879 to.to_sacks = (u_char *)tp->sackblks; 880 } 881 } 882 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 883 /* TCP-MD5 (RFC2385). */ 884 /* 885 * Check that TCP_MD5SIG is enabled in tcpcb to 886 * account the size needed to set this TCP option. 887 */ 888 if (tp->t_flags & TF_SIGNATURE) 889 to.to_flags |= TOF_SIGNATURE; 890 #endif /* TCP_SIGNATURE */ 891 892 /* Processing the options. */ 893 hdrlen += optlen = tcp_addoptions(&to, opt); 894 /* 895 * If we wanted a TFO option to be added, but it was unable 896 * to fit, ensure no data is sent. 897 */ 898 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 899 !(to.to_flags & TOF_FASTOPEN)) 900 len = 0; 901 } 902 if (tp->t_port) { 903 if (V_tcp_udp_tunneling_port == 0) { 904 /* The port was removed?? */ 905 SOCKBUF_UNLOCK(&so->so_snd); 906 return (EHOSTUNREACH); 907 } 908 hdrlen += sizeof(struct udphdr); 909 } 910 /* 911 * Adjust data length if insertion of options will 912 * bump the packet length beyond the t_maxseg length. 913 * Clear the FIN bit because we cut off the tail of 914 * the segment. 915 */ 916 if (len + optlen + ipoptlen > tp->t_maxseg) { 917 flags &= ~TH_FIN; 918 919 if (tso) { 920 u_int if_hw_tsomax; 921 u_int moff; 922 int max_len; 923 924 /* extract TSO information */ 925 if_hw_tsomax = tp->t_tsomax; 926 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 927 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 928 929 /* 930 * Limit a TSO burst to prevent it from 931 * overflowing or exceeding the maximum length 932 * allowed by the network interface: 933 */ 934 KASSERT(ipoptlen == 0, 935 ("%s: TSO can't do IP options", __func__)); 936 937 /* 938 * Check if we should limit by maximum payload 939 * length: 940 */ 941 if (if_hw_tsomax != 0) { 942 /* compute maximum TSO length */ 943 max_len = (if_hw_tsomax - hdrlen - 944 max_linkhdr); 945 if (max_len <= 0) { 946 len = 0; 947 } else if (len > max_len) { 948 sendalot = 1; 949 len = max_len; 950 } 951 } 952 953 /* 954 * Prevent the last segment from being 955 * fractional unless the send sockbuf can be 956 * emptied: 957 */ 958 max_len = (tp->t_maxseg - optlen); 959 if (((uint32_t)off + (uint32_t)len) < 960 sbavail(&so->so_snd)) { 961 moff = len % max_len; 962 if (moff != 0) { 963 len -= moff; 964 sendalot = 1; 965 } 966 } 967 968 /* 969 * In case there are too many small fragments 970 * don't use TSO: 971 */ 972 if (len <= max_len) { 973 len = max_len; 974 sendalot = 1; 975 tso = 0; 976 } 977 978 /* 979 * Send the FIN in a separate segment 980 * after the bulk sending is done. 981 * We don't trust the TSO implementations 982 * to clear the FIN flag on all but the 983 * last segment. 984 */ 985 if (tp->t_flags & TF_NEEDFIN) 986 sendalot = 1; 987 } else { 988 if (optlen + ipoptlen >= tp->t_maxseg) { 989 /* 990 * Since we don't have enough space to put 991 * the IP header chain and the TCP header in 992 * one packet as required by RFC 7112, don't 993 * send it. Also ensure that at least one 994 * byte of the payload can be put into the 995 * TCP segment. 996 */ 997 SOCKBUF_UNLOCK(&so->so_snd); 998 error = EMSGSIZE; 999 sack_rxmit = 0; 1000 goto out; 1001 } 1002 len = tp->t_maxseg - optlen - ipoptlen; 1003 sendalot = 1; 1004 if (dont_sendalot) 1005 sendalot = 0; 1006 } 1007 } else 1008 tso = 0; 1009 1010 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 1011 ("%s: len > IP_MAXPACKET", __func__)); 1012 1013 /*#ifdef DIAGNOSTIC*/ 1014 #ifdef INET6 1015 if (max_linkhdr + hdrlen > MCLBYTES) 1016 #else 1017 if (max_linkhdr + hdrlen > MHLEN) 1018 #endif 1019 panic("tcphdr too big"); 1020 /*#endif*/ 1021 1022 /* 1023 * This KASSERT is here to catch edge cases at a well defined place. 1024 * Before, those had triggered (random) panic conditions further down. 1025 */ 1026 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 1027 1028 /* 1029 * Grab a header mbuf, attaching a copy of data to 1030 * be transmitted, and initialize the header from 1031 * the template for sends on this connection. 1032 */ 1033 if (len) { 1034 struct mbuf *mb; 1035 struct sockbuf *msb; 1036 u_int moff; 1037 1038 if ((tp->t_flags & TF_FORCEDATA) && len == 1) { 1039 TCPSTAT_INC(tcps_sndprobe); 1040 #ifdef STATS 1041 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1042 stats_voi_update_abs_u32(tp->t_stats, 1043 VOI_TCP_RETXPB, len); 1044 else 1045 stats_voi_update_abs_u64(tp->t_stats, 1046 VOI_TCP_TXPB, len); 1047 #endif /* STATS */ 1048 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 1049 tp->t_sndrexmitpack++; 1050 TCPSTAT_INC(tcps_sndrexmitpack); 1051 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 1052 #ifdef STATS 1053 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 1054 len); 1055 #endif /* STATS */ 1056 } else { 1057 TCPSTAT_INC(tcps_sndpack); 1058 TCPSTAT_ADD(tcps_sndbyte, len); 1059 #ifdef STATS 1060 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 1061 len); 1062 #endif /* STATS */ 1063 } 1064 #ifdef INET6 1065 if (MHLEN < hdrlen + max_linkhdr) 1066 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1067 else 1068 #endif 1069 m = m_gethdr(M_NOWAIT, MT_DATA); 1070 1071 if (m == NULL) { 1072 SOCKBUF_UNLOCK(&so->so_snd); 1073 error = ENOBUFS; 1074 sack_rxmit = 0; 1075 goto out; 1076 } 1077 1078 m->m_data += max_linkhdr; 1079 m->m_len = hdrlen; 1080 1081 /* 1082 * Start the m_copy functions from the closest mbuf 1083 * to the offset in the socket buffer chain. 1084 */ 1085 mb = sbsndptr_noadv(&so->so_snd, off, &moff); 1086 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 1087 m_copydata(mb, moff, len, 1088 mtod(m, caddr_t) + hdrlen); 1089 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1090 sbsndptr_adv(&so->so_snd, mb, len); 1091 m->m_len += len; 1092 } else { 1093 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1094 msb = NULL; 1095 else 1096 msb = &so->so_snd; 1097 m->m_next = tcp_m_copym(mb, moff, 1098 &len, if_hw_tsomaxsegcount, 1099 if_hw_tsomaxsegsize, msb, hw_tls); 1100 if (len <= (tp->t_maxseg - optlen)) { 1101 /* 1102 * Must have ran out of mbufs for the copy 1103 * shorten it to no longer need tso. Lets 1104 * not put on sendalot since we are low on 1105 * mbufs. 1106 */ 1107 tso = 0; 1108 } 1109 if (m->m_next == NULL) { 1110 SOCKBUF_UNLOCK(&so->so_snd); 1111 (void) m_free(m); 1112 error = ENOBUFS; 1113 sack_rxmit = 0; 1114 goto out; 1115 } 1116 } 1117 1118 /* 1119 * If we're sending everything we've got, set PUSH. 1120 * (This will keep happy those implementations which only 1121 * give data to the user when a buffer fills or 1122 * a PUSH comes in.) 1123 */ 1124 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) && 1125 !(flags & TH_SYN)) 1126 flags |= TH_PUSH; 1127 SOCKBUF_UNLOCK(&so->so_snd); 1128 } else { 1129 SOCKBUF_UNLOCK(&so->so_snd); 1130 if (tp->t_flags & TF_ACKNOW) 1131 TCPSTAT_INC(tcps_sndacks); 1132 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1133 TCPSTAT_INC(tcps_sndctrl); 1134 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1135 TCPSTAT_INC(tcps_sndurg); 1136 else 1137 TCPSTAT_INC(tcps_sndwinup); 1138 1139 m = m_gethdr(M_NOWAIT, MT_DATA); 1140 if (m == NULL) { 1141 error = ENOBUFS; 1142 sack_rxmit = 0; 1143 goto out; 1144 } 1145 #ifdef INET6 1146 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1147 MHLEN >= hdrlen) { 1148 M_ALIGN(m, hdrlen); 1149 } else 1150 #endif 1151 m->m_data += max_linkhdr; 1152 m->m_len = hdrlen; 1153 } 1154 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1155 m->m_pkthdr.rcvif = (struct ifnet *)0; 1156 #ifdef MAC 1157 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1158 #endif 1159 #ifdef INET6 1160 if (isipv6) { 1161 ip6 = mtod(m, struct ip6_hdr *); 1162 if (tp->t_port) { 1163 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); 1164 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1165 udp->uh_dport = tp->t_port; 1166 ulen = hdrlen + len - sizeof(struct ip6_hdr); 1167 udp->uh_ulen = htons(ulen); 1168 th = (struct tcphdr *)(udp + 1); 1169 } else { 1170 th = (struct tcphdr *)(ip6 + 1); 1171 } 1172 tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip6, th); 1173 } else 1174 #endif /* INET6 */ 1175 { 1176 ip = mtod(m, struct ip *); 1177 #ifdef TCPDEBUG 1178 ipov = (struct ipovly *)ip; 1179 #endif 1180 if (tp->t_port) { 1181 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); 1182 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1183 udp->uh_dport = tp->t_port; 1184 ulen = hdrlen + len - sizeof(struct ip); 1185 udp->uh_ulen = htons(ulen); 1186 th = (struct tcphdr *)(udp + 1); 1187 } else 1188 th = (struct tcphdr *)(ip + 1); 1189 tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip, th); 1190 } 1191 1192 /* 1193 * Fill in fields, remembering maximum advertised 1194 * window for use in delaying messages about window sizes. 1195 * If resending a FIN, be sure not to use a new sequence number. 1196 */ 1197 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1198 tp->snd_nxt == tp->snd_max) 1199 tp->snd_nxt--; 1200 /* 1201 * If we are starting a connection, send ECN setup 1202 * SYN packet. If we are on a retransmit, we may 1203 * resend those bits a number of times as per 1204 * RFC 3168. 1205 */ 1206 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 1207 flags |= tcp_ecn_output_syn_sent(tp); 1208 } 1209 /* Also handle parallel SYN for ECN */ 1210 if ((TCPS_HAVERCVDSYN(tp->t_state)) && 1211 (tp->t_flags2 & TF2_ECN_PERMIT)) { 1212 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit); 1213 if ((tp->t_state == TCPS_SYN_RECEIVED) && 1214 (tp->t_flags2 & TF2_ECN_SND_ECE)) 1215 tp->t_flags2 &= ~TF2_ECN_SND_ECE; 1216 #ifdef INET6 1217 if (isipv6) { 1218 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20); 1219 ip6->ip6_flow |= htonl(ect << 20); 1220 } 1221 else 1222 #endif 1223 { 1224 ip->ip_tos &= ~IPTOS_ECN_MASK; 1225 ip->ip_tos |= ect; 1226 } 1227 } 1228 1229 /* 1230 * If we are doing retransmissions, then snd_nxt will 1231 * not reflect the first unsent octet. For ACK only 1232 * packets, we do not want the sequence number of the 1233 * retransmitted packet, we want the sequence number 1234 * of the next unsent octet. So, if there is no data 1235 * (and no SYN or FIN), use snd_max instead of snd_nxt 1236 * when filling in ti_seq. But if we are in persist 1237 * state, snd_max might reflect one byte beyond the 1238 * right edge of the window, so use snd_nxt in that 1239 * case, since we know we aren't doing a retransmission. 1240 * (retransmit and persist are mutually exclusive...) 1241 */ 1242 if (sack_rxmit == 0) { 1243 if (len || (flags & (TH_SYN|TH_FIN)) || 1244 tcp_timer_active(tp, TT_PERSIST)) 1245 th->th_seq = htonl(tp->snd_nxt); 1246 else 1247 th->th_seq = htonl(tp->snd_max); 1248 } else { 1249 th->th_seq = htonl(p->rxmit); 1250 p->rxmit += len; 1251 /* 1252 * Lost Retransmission Detection 1253 * trigger resending of a (then 1254 * still existing) hole, when 1255 * fack acks recoverypoint. 1256 */ 1257 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end)) 1258 p->rxmit = tp->snd_recover; 1259 tp->sackhint.sack_bytes_rexmit += len; 1260 } 1261 if (IN_RECOVERY(tp->t_flags)) { 1262 /* 1263 * Account all bytes transmitted while 1264 * IN_RECOVERY, simplifying PRR and 1265 * Lost Retransmit Detection 1266 */ 1267 tp->sackhint.prr_out += len; 1268 } 1269 th->th_ack = htonl(tp->rcv_nxt); 1270 if (optlen) { 1271 bcopy(opt, th + 1, optlen); 1272 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1273 } 1274 tcp_set_flags(th, flags); 1275 /* 1276 * Calculate receive window. Don't shrink window, 1277 * but avoid silly window syndrome. 1278 * If a RST segment is sent, advertise a window of zero. 1279 */ 1280 if (flags & TH_RST) { 1281 recwin = 0; 1282 } else { 1283 if (recwin < (so->so_rcv.sb_hiwat / 4) && 1284 recwin < tp->t_maxseg) 1285 recwin = 0; 1286 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1287 recwin < (tp->rcv_adv - tp->rcv_nxt)) 1288 recwin = (tp->rcv_adv - tp->rcv_nxt); 1289 } 1290 /* 1291 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1292 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1293 * case is handled in syncache. 1294 */ 1295 if (flags & TH_SYN) 1296 th->th_win = htons((u_short) 1297 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1298 else { 1299 /* Avoid shrinking window with window scaling. */ 1300 recwin = roundup2(recwin, 1 << tp->rcv_scale); 1301 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1302 } 1303 1304 /* 1305 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1306 * a 0 window. This may cause the remote transmitter to stall. This 1307 * flag tells soreceive() to disable delayed acknowledgements when 1308 * draining the buffer. This can occur if the receiver is attempting 1309 * to read more data than can be buffered prior to transmitting on 1310 * the connection. 1311 */ 1312 if (th->th_win == 0) { 1313 tp->t_sndzerowin++; 1314 tp->t_flags |= TF_RXWIN0SENT; 1315 } else 1316 tp->t_flags &= ~TF_RXWIN0SENT; 1317 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1318 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1319 th->th_flags |= TH_URG; 1320 } else 1321 /* 1322 * If no urgent pointer to send, then we pull 1323 * the urgent pointer to the left edge of the send window 1324 * so that it doesn't drift into the send window on sequence 1325 * number wraparound. 1326 */ 1327 tp->snd_up = tp->snd_una; /* drag it along */ 1328 1329 /* 1330 * Put TCP length in extended header, and then 1331 * checksum extended header and data. 1332 */ 1333 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1334 1335 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1336 if (to.to_flags & TOF_SIGNATURE) { 1337 /* 1338 * Calculate MD5 signature and put it into the place 1339 * determined before. 1340 * NOTE: since TCP options buffer doesn't point into 1341 * mbuf's data, calculate offset and use it. 1342 */ 1343 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th, 1344 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) { 1345 /* 1346 * Do not send segment if the calculation of MD5 1347 * digest has failed. 1348 */ 1349 m_freem(m); 1350 goto out; 1351 } 1352 } 1353 #endif 1354 #ifdef INET6 1355 if (isipv6) { 1356 /* 1357 * There is no need to fill in ip6_plen right now. 1358 * It will be filled later by ip6_output. 1359 */ 1360 if (tp->t_port) { 1361 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 1362 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 1363 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 1364 th->th_sum = htons(0); 1365 UDPSTAT_INC(udps_opackets); 1366 } else { 1367 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1368 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1369 th->th_sum = in6_cksum_pseudo(ip6, 1370 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 1371 0); 1372 } 1373 } 1374 #endif 1375 #if defined(INET6) && defined(INET) 1376 else 1377 #endif 1378 #ifdef INET 1379 { 1380 if (tp->t_port) { 1381 m->m_pkthdr.csum_flags = CSUM_UDP; 1382 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 1383 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 1384 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 1385 th->th_sum = htons(0); 1386 UDPSTAT_INC(udps_opackets); 1387 } else { 1388 m->m_pkthdr.csum_flags = CSUM_TCP; 1389 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1390 th->th_sum = in_pseudo(ip->ip_src.s_addr, 1391 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 1392 IPPROTO_TCP + len + optlen)); 1393 } 1394 1395 /* IP version must be set here for ipv4/ipv6 checking later */ 1396 KASSERT(ip->ip_v == IPVERSION, 1397 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1398 } 1399 #endif 1400 1401 /* 1402 * Enable TSO and specify the size of the segments. 1403 * The TCP pseudo header checksum is always provided. 1404 */ 1405 if (tso) { 1406 KASSERT(len > tp->t_maxseg - optlen, 1407 ("%s: len <= tso_segsz", __func__)); 1408 m->m_pkthdr.csum_flags |= CSUM_TSO; 1409 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 1410 } 1411 1412 KASSERT(len + hdrlen == m_length(m, NULL), 1413 ("%s: mbuf chain shorter than expected: %d + %u != %u", 1414 __func__, len, hdrlen, m_length(m, NULL))); 1415 1416 #ifdef TCP_HHOOK 1417 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1418 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1419 #endif 1420 1421 #ifdef TCPDEBUG 1422 /* 1423 * Trace. 1424 */ 1425 if (so->so_options & SO_DEBUG) { 1426 u_short save = 0; 1427 #ifdef INET6 1428 if (!isipv6) 1429 #endif 1430 { 1431 save = ipov->ih_len; 1432 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1433 } 1434 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1435 #ifdef INET6 1436 if (!isipv6) 1437 #endif 1438 ipov->ih_len = save; 1439 } 1440 #endif /* TCPDEBUG */ 1441 TCP_PROBE3(debug__output, tp, th, m); 1442 1443 /* We're getting ready to send; log now. */ 1444 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 1445 len, NULL, false); 1446 1447 /* 1448 * Fill in IP length and desired time to live and 1449 * send to IP level. There should be a better way 1450 * to handle ttl and tos; we could keep them in 1451 * the template, but need a way to checksum without them. 1452 */ 1453 /* 1454 * m->m_pkthdr.len should have been set before checksum calculation, 1455 * because in6_cksum() need it. 1456 */ 1457 #ifdef INET6 1458 if (isipv6) { 1459 /* 1460 * we separately set hoplimit for every segment, since the 1461 * user might want to change the value via setsockopt. 1462 * Also, desired default hop limit might be changed via 1463 * Neighbor Discovery. 1464 */ 1465 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1466 1467 /* 1468 * Set the packet size here for the benefit of DTrace probes. 1469 * ip6_output() will set it properly; it's supposed to include 1470 * the option header lengths as well. 1471 */ 1472 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1473 1474 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 1475 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1476 else 1477 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1478 1479 if (tp->t_state == TCPS_SYN_SENT) 1480 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1481 1482 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1483 1484 #ifdef TCPPCAP 1485 /* Save packet, if requested. */ 1486 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1487 #endif 1488 1489 /* TODO: IPv6 IP6TOS_ECT bit on */ 1490 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, 1491 &tp->t_inpcb->inp_route6, 1492 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1493 NULL, NULL, tp->t_inpcb); 1494 1495 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL) 1496 mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu; 1497 } 1498 #endif /* INET6 */ 1499 #if defined(INET) && defined(INET6) 1500 else 1501 #endif 1502 #ifdef INET 1503 { 1504 ip->ip_len = htons(m->m_pkthdr.len); 1505 #ifdef INET6 1506 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1507 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1508 #endif /* INET6 */ 1509 /* 1510 * If we do path MTU discovery, then we set DF on every packet. 1511 * This might not be the best thing to do according to RFC3390 1512 * Section 2. However the tcp hostcache migitates the problem 1513 * so it affects only the first tcp connection with a host. 1514 * 1515 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1516 */ 1517 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 1518 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1519 if (tp->t_port == 0 || len < V_tcp_minmss) { 1520 ip->ip_off |= htons(IP_DF); 1521 } 1522 } else { 1523 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1524 } 1525 1526 if (tp->t_state == TCPS_SYN_SENT) 1527 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1528 1529 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1530 1531 #ifdef TCPPCAP 1532 /* Save packet, if requested. */ 1533 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1534 #endif 1535 1536 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route, 1537 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1538 tp->t_inpcb); 1539 1540 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL) 1541 mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu; 1542 } 1543 #endif /* INET */ 1544 1545 out: 1546 if (error == 0) 1547 tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls); 1548 /* 1549 * In transmit state, time the transmission and arrange for 1550 * the retransmit. In persist state, just set snd_max. 1551 */ 1552 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1553 !tcp_timer_active(tp, TT_PERSIST)) { 1554 tcp_seq startseq = tp->snd_nxt; 1555 1556 /* 1557 * Advance snd_nxt over sequence space of this segment. 1558 */ 1559 if (flags & (TH_SYN|TH_FIN)) { 1560 if (flags & TH_SYN) 1561 tp->snd_nxt++; 1562 if (flags & TH_FIN) { 1563 tp->snd_nxt++; 1564 tp->t_flags |= TF_SENTFIN; 1565 } 1566 } 1567 if (sack_rxmit) 1568 goto timer; 1569 tp->snd_nxt += len; 1570 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1571 tp->snd_max = tp->snd_nxt; 1572 /* 1573 * Time this transmission if not a retransmission and 1574 * not currently timing anything. 1575 */ 1576 tp->t_sndtime = ticks; 1577 if (tp->t_rtttime == 0) { 1578 tp->t_rtttime = ticks; 1579 tp->t_rtseq = startseq; 1580 TCPSTAT_INC(tcps_segstimed); 1581 } 1582 #ifdef STATS 1583 if (!(tp->t_flags & TF_GPUTINPROG) && len) { 1584 tp->t_flags |= TF_GPUTINPROG; 1585 tp->gput_seq = startseq; 1586 tp->gput_ack = startseq + 1587 ulmin(sbavail(&so->so_snd) - off, sendwin); 1588 tp->gput_ts = tcp_ts_getticks(); 1589 } 1590 #endif /* STATS */ 1591 } 1592 1593 /* 1594 * Set retransmit timer if not currently set, 1595 * and not doing a pure ack or a keep-alive probe. 1596 * Initial value for retransmit timer is smoothed 1597 * round-trip time + 2 * round-trip time variance. 1598 * Initialize shift counter which is used for backoff 1599 * of retransmit time. 1600 */ 1601 timer: 1602 if (!tcp_timer_active(tp, TT_REXMT) && 1603 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1604 (tp->snd_nxt != tp->snd_una))) { 1605 if (tcp_timer_active(tp, TT_PERSIST)) { 1606 tcp_timer_activate(tp, TT_PERSIST, 0); 1607 tp->t_rxtshift = 0; 1608 } 1609 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1610 } else if (len == 0 && sbavail(&so->so_snd) && 1611 !tcp_timer_active(tp, TT_REXMT) && 1612 !tcp_timer_active(tp, TT_PERSIST)) { 1613 /* 1614 * Avoid a situation where we do not set persist timer 1615 * after a zero window condition. For example: 1616 * 1) A -> B: packet with enough data to fill the window 1617 * 2) B -> A: ACK for #1 + new data (0 window 1618 * advertisement) 1619 * 3) A -> B: ACK for #2, 0 len packet 1620 * 1621 * In this case, A will not activate the persist timer, 1622 * because it chose to send a packet. Unless tcp_output 1623 * is called for some other reason (delayed ack timer, 1624 * another input packet from B, socket syscall), A will 1625 * not send zero window probes. 1626 * 1627 * So, if you send a 0-length packet, but there is data 1628 * in the socket buffer, and neither the rexmt or 1629 * persist timer is already set, then activate the 1630 * persist timer. 1631 */ 1632 tp->t_rxtshift = 0; 1633 tcp_setpersist(tp); 1634 } 1635 } else { 1636 /* 1637 * Persist case, update snd_max but since we are in 1638 * persist mode (no window) we do not update snd_nxt. 1639 */ 1640 int xlen = len; 1641 if (flags & TH_SYN) 1642 ++xlen; 1643 if (flags & TH_FIN) { 1644 ++xlen; 1645 tp->t_flags |= TF_SENTFIN; 1646 } 1647 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1648 tp->snd_max = tp->snd_nxt + xlen; 1649 } 1650 if ((error == 0) && 1651 (TCPS_HAVEESTABLISHED(tp->t_state) && 1652 (tp->t_flags & TF_SACK_PERMIT) && 1653 tp->rcv_numsacks > 0)) { 1654 /* Clean up any DSACK's sent */ 1655 tcp_clean_dsack_blocks(tp); 1656 } 1657 if (error) { 1658 /* Record the error. */ 1659 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, 1660 error, 0, NULL, false); 1661 1662 /* 1663 * We know that the packet was lost, so back out the 1664 * sequence number advance, if any. 1665 * 1666 * If the error is EPERM the packet got blocked by the 1667 * local firewall. Normally we should terminate the 1668 * connection but the blocking may have been spurious 1669 * due to a firewall reconfiguration cycle. So we treat 1670 * it like a packet loss and let the retransmit timer and 1671 * timeouts do their work over time. 1672 * XXX: It is a POLA question whether calling tcp_drop right 1673 * away would be the really correct behavior instead. 1674 */ 1675 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1676 !tcp_timer_active(tp, TT_PERSIST)) && 1677 ((flags & TH_SYN) == 0) && 1678 (error != EPERM)) { 1679 if (sack_rxmit) { 1680 p->rxmit -= len; 1681 tp->sackhint.sack_bytes_rexmit -= len; 1682 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1683 ("sackhint bytes rtx >= 0")); 1684 KASSERT((flags & TH_FIN) == 0, 1685 ("error while FIN with SACK rxmit")); 1686 } else { 1687 tp->snd_nxt -= len; 1688 if (flags & TH_FIN) 1689 tp->snd_nxt--; 1690 } 1691 } 1692 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1693 switch (error) { 1694 case EACCES: 1695 case EPERM: 1696 tp->t_softerror = error; 1697 return (error); 1698 case ENOBUFS: 1699 TCP_XMIT_TIMER_ASSERT(tp, len, flags); 1700 tp->snd_cwnd = tp->t_maxseg; 1701 return (0); 1702 case EMSGSIZE: 1703 /* 1704 * For some reason the interface we used initially 1705 * to send segments changed to another or lowered 1706 * its MTU. 1707 * If TSO was active we either got an interface 1708 * without TSO capabilits or TSO was turned off. 1709 * If we obtained mtu from ip_output() then update 1710 * it and try again. 1711 */ 1712 if (tso) 1713 tp->t_flags &= ~TF_TSO; 1714 if (mtu != 0) { 1715 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1716 goto again; 1717 } 1718 return (error); 1719 case EHOSTDOWN: 1720 case EHOSTUNREACH: 1721 case ENETDOWN: 1722 case ENETUNREACH: 1723 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1724 tp->t_softerror = error; 1725 return (0); 1726 } 1727 /* FALLTHROUGH */ 1728 default: 1729 return (error); 1730 } 1731 } 1732 TCPSTAT_INC(tcps_sndtotal); 1733 1734 /* 1735 * Data sent (as far as we can tell). 1736 * If this advertises a larger window than any other segment, 1737 * then remember the size of the advertised window. 1738 * Any pending ACK has now been sent. 1739 */ 1740 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1741 tp->rcv_adv = tp->rcv_nxt + recwin; 1742 tp->last_ack_sent = tp->rcv_nxt; 1743 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1744 if (tcp_timer_active(tp, TT_DELACK)) 1745 tcp_timer_activate(tp, TT_DELACK, 0); 1746 #if 0 1747 /* 1748 * This completely breaks TCP if newreno is turned on. What happens 1749 * is that if delayed-acks are turned on on the receiver, this code 1750 * on the transmitter effectively destroys the TCP window, forcing 1751 * it to four packets (1.5Kx4 = 6K window). 1752 */ 1753 if (sendalot && --maxburst) 1754 goto again; 1755 #endif 1756 if (sendalot) 1757 goto again; 1758 return (0); 1759 } 1760 1761 void 1762 tcp_setpersist(struct tcpcb *tp) 1763 { 1764 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1765 int tt; 1766 1767 tp->t_flags &= ~TF_PREVVALID; 1768 if (tcp_timer_active(tp, TT_REXMT)) 1769 panic("tcp_setpersist: retransmit pending"); 1770 /* 1771 * Start/restart persistence timer. 1772 */ 1773 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1774 tcp_persmin, tcp_persmax); 1775 tcp_timer_activate(tp, TT_PERSIST, tt); 1776 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1777 tp->t_rxtshift++; 1778 } 1779 1780 /* 1781 * Insert TCP options according to the supplied parameters to the place 1782 * optp in a consistent way. Can handle unaligned destinations. 1783 * 1784 * The order of the option processing is crucial for optimal packing and 1785 * alignment for the scarce option space. 1786 * 1787 * The optimal order for a SYN/SYN-ACK segment is: 1788 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1789 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1790 * 1791 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1792 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1793 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1794 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1795 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1796 */ 1797 int 1798 tcp_addoptions(struct tcpopt *to, u_char *optp) 1799 { 1800 u_int32_t mask, optlen = 0; 1801 1802 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1803 if ((to->to_flags & mask) != mask) 1804 continue; 1805 if (optlen == TCP_MAXOLEN) 1806 break; 1807 switch (to->to_flags & mask) { 1808 case TOF_MSS: 1809 while (optlen % 4) { 1810 optlen += TCPOLEN_NOP; 1811 *optp++ = TCPOPT_NOP; 1812 } 1813 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1814 continue; 1815 optlen += TCPOLEN_MAXSEG; 1816 *optp++ = TCPOPT_MAXSEG; 1817 *optp++ = TCPOLEN_MAXSEG; 1818 to->to_mss = htons(to->to_mss); 1819 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1820 optp += sizeof(to->to_mss); 1821 break; 1822 case TOF_SCALE: 1823 while (!optlen || optlen % 2 != 1) { 1824 optlen += TCPOLEN_NOP; 1825 *optp++ = TCPOPT_NOP; 1826 } 1827 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1828 continue; 1829 optlen += TCPOLEN_WINDOW; 1830 *optp++ = TCPOPT_WINDOW; 1831 *optp++ = TCPOLEN_WINDOW; 1832 *optp++ = to->to_wscale; 1833 break; 1834 case TOF_SACKPERM: 1835 while (optlen % 2) { 1836 optlen += TCPOLEN_NOP; 1837 *optp++ = TCPOPT_NOP; 1838 } 1839 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1840 continue; 1841 optlen += TCPOLEN_SACK_PERMITTED; 1842 *optp++ = TCPOPT_SACK_PERMITTED; 1843 *optp++ = TCPOLEN_SACK_PERMITTED; 1844 break; 1845 case TOF_TS: 1846 while (!optlen || optlen % 4 != 2) { 1847 optlen += TCPOLEN_NOP; 1848 *optp++ = TCPOPT_NOP; 1849 } 1850 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1851 continue; 1852 optlen += TCPOLEN_TIMESTAMP; 1853 *optp++ = TCPOPT_TIMESTAMP; 1854 *optp++ = TCPOLEN_TIMESTAMP; 1855 to->to_tsval = htonl(to->to_tsval); 1856 to->to_tsecr = htonl(to->to_tsecr); 1857 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1858 optp += sizeof(to->to_tsval); 1859 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1860 optp += sizeof(to->to_tsecr); 1861 break; 1862 case TOF_SIGNATURE: 1863 { 1864 int siglen = TCPOLEN_SIGNATURE - 2; 1865 1866 while (!optlen || optlen % 4 != 2) { 1867 optlen += TCPOLEN_NOP; 1868 *optp++ = TCPOPT_NOP; 1869 } 1870 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) { 1871 to->to_flags &= ~TOF_SIGNATURE; 1872 continue; 1873 } 1874 optlen += TCPOLEN_SIGNATURE; 1875 *optp++ = TCPOPT_SIGNATURE; 1876 *optp++ = TCPOLEN_SIGNATURE; 1877 to->to_signature = optp; 1878 while (siglen--) 1879 *optp++ = 0; 1880 break; 1881 } 1882 case TOF_SACK: 1883 { 1884 int sackblks = 0; 1885 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1886 tcp_seq sack_seq; 1887 1888 while (!optlen || optlen % 4 != 2) { 1889 optlen += TCPOLEN_NOP; 1890 *optp++ = TCPOPT_NOP; 1891 } 1892 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1893 continue; 1894 optlen += TCPOLEN_SACKHDR; 1895 *optp++ = TCPOPT_SACK; 1896 sackblks = min(to->to_nsacks, 1897 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1898 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1899 while (sackblks--) { 1900 sack_seq = htonl(sack->start); 1901 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1902 optp += sizeof(sack_seq); 1903 sack_seq = htonl(sack->end); 1904 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1905 optp += sizeof(sack_seq); 1906 optlen += TCPOLEN_SACK; 1907 sack++; 1908 } 1909 TCPSTAT_INC(tcps_sack_send_blocks); 1910 break; 1911 } 1912 case TOF_FASTOPEN: 1913 { 1914 int total_len; 1915 1916 /* XXX is there any point to aligning this option? */ 1917 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; 1918 if (TCP_MAXOLEN - optlen < total_len) { 1919 to->to_flags &= ~TOF_FASTOPEN; 1920 continue; 1921 } 1922 *optp++ = TCPOPT_FAST_OPEN; 1923 *optp++ = total_len; 1924 if (to->to_tfo_len > 0) { 1925 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); 1926 optp += to->to_tfo_len; 1927 } 1928 optlen += total_len; 1929 break; 1930 } 1931 default: 1932 panic("%s: unknown TCP option type", __func__); 1933 break; 1934 } 1935 } 1936 1937 /* Terminate and pad TCP options to a 4 byte boundary. */ 1938 if (optlen % 4) { 1939 optlen += TCPOLEN_EOL; 1940 *optp++ = TCPOPT_EOL; 1941 } 1942 /* 1943 * According to RFC 793 (STD0007): 1944 * "The content of the header beyond the End-of-Option option 1945 * must be header padding (i.e., zero)." 1946 * and later: "The padding is composed of zeros." 1947 */ 1948 while (optlen % 4) { 1949 optlen += TCPOLEN_PAD; 1950 *optp++ = TCPOPT_PAD; 1951 } 1952 1953 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1954 return (optlen); 1955 } 1956 1957 /* 1958 * This is a copy of m_copym(), taking the TSO segment size/limit 1959 * constraints into account, and advancing the sndptr as it goes. 1960 */ 1961 struct mbuf * 1962 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen, 1963 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls) 1964 { 1965 #ifdef KERN_TLS 1966 struct ktls_session *tls, *ntls; 1967 struct mbuf *start __diagused; 1968 #endif 1969 struct mbuf *n, **np; 1970 struct mbuf *top; 1971 int32_t off = off0; 1972 int32_t len = *plen; 1973 int32_t fragsize; 1974 int32_t len_cp = 0; 1975 int32_t *pkthdrlen; 1976 uint32_t mlen, frags; 1977 bool copyhdr; 1978 1979 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off)); 1980 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len)); 1981 if (off == 0 && m->m_flags & M_PKTHDR) 1982 copyhdr = true; 1983 else 1984 copyhdr = false; 1985 while (off > 0) { 1986 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain")); 1987 if (off < m->m_len) 1988 break; 1989 off -= m->m_len; 1990 if ((sb) && (m == sb->sb_sndptr)) { 1991 sb->sb_sndptroff += m->m_len; 1992 sb->sb_sndptr = m->m_next; 1993 } 1994 m = m->m_next; 1995 } 1996 np = ⊤ 1997 top = NULL; 1998 pkthdrlen = NULL; 1999 #ifdef KERN_TLS 2000 if (hw_tls && (m->m_flags & M_EXTPG)) 2001 tls = m->m_epg_tls; 2002 else 2003 tls = NULL; 2004 start = m; 2005 #endif 2006 while (len > 0) { 2007 if (m == NULL) { 2008 KASSERT(len == M_COPYALL, 2009 ("tcp_m_copym, length > size of mbuf chain")); 2010 *plen = len_cp; 2011 if (pkthdrlen != NULL) 2012 *pkthdrlen = len_cp; 2013 break; 2014 } 2015 #ifdef KERN_TLS 2016 if (hw_tls) { 2017 if (m->m_flags & M_EXTPG) 2018 ntls = m->m_epg_tls; 2019 else 2020 ntls = NULL; 2021 2022 /* 2023 * Avoid mixing TLS records with handshake 2024 * data or TLS records from different 2025 * sessions. 2026 */ 2027 if (tls != ntls) { 2028 MPASS(m != start); 2029 *plen = len_cp; 2030 if (pkthdrlen != NULL) 2031 *pkthdrlen = len_cp; 2032 break; 2033 } 2034 } 2035 #endif 2036 mlen = min(len, m->m_len - off); 2037 if (seglimit) { 2038 /* 2039 * For M_EXTPG mbufs, add 3 segments 2040 * + 1 in case we are crossing page boundaries 2041 * + 2 in case the TLS hdr/trailer are used 2042 * It is cheaper to just add the segments 2043 * than it is to take the cache miss to look 2044 * at the mbuf ext_pgs state in detail. 2045 */ 2046 if (m->m_flags & M_EXTPG) { 2047 fragsize = min(segsize, PAGE_SIZE); 2048 frags = 3; 2049 } else { 2050 fragsize = segsize; 2051 frags = 0; 2052 } 2053 2054 /* Break if we really can't fit anymore. */ 2055 if ((frags + 1) >= seglimit) { 2056 *plen = len_cp; 2057 if (pkthdrlen != NULL) 2058 *pkthdrlen = len_cp; 2059 break; 2060 } 2061 2062 /* 2063 * Reduce size if you can't copy the whole 2064 * mbuf. If we can't copy the whole mbuf, also 2065 * adjust len so the loop will end after this 2066 * mbuf. 2067 */ 2068 if ((frags + howmany(mlen, fragsize)) >= seglimit) { 2069 mlen = (seglimit - frags - 1) * fragsize; 2070 len = mlen; 2071 *plen = len_cp + len; 2072 if (pkthdrlen != NULL) 2073 *pkthdrlen = *plen; 2074 } 2075 frags += howmany(mlen, fragsize); 2076 if (frags == 0) 2077 frags++; 2078 seglimit -= frags; 2079 KASSERT(seglimit > 0, 2080 ("%s: seglimit went too low", __func__)); 2081 } 2082 if (copyhdr) 2083 n = m_gethdr(M_NOWAIT, m->m_type); 2084 else 2085 n = m_get(M_NOWAIT, m->m_type); 2086 *np = n; 2087 if (n == NULL) 2088 goto nospace; 2089 if (copyhdr) { 2090 if (!m_dup_pkthdr(n, m, M_NOWAIT)) 2091 goto nospace; 2092 if (len == M_COPYALL) 2093 n->m_pkthdr.len -= off0; 2094 else 2095 n->m_pkthdr.len = len; 2096 pkthdrlen = &n->m_pkthdr.len; 2097 copyhdr = false; 2098 } 2099 n->m_len = mlen; 2100 len_cp += n->m_len; 2101 if (m->m_flags & (M_EXT|M_EXTPG)) { 2102 n->m_data = m->m_data + off; 2103 mb_dupcl(n, m); 2104 } else 2105 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 2106 (u_int)n->m_len); 2107 2108 if (sb && (sb->sb_sndptr == m) && 2109 ((n->m_len + off) >= m->m_len) && m->m_next) { 2110 sb->sb_sndptroff += m->m_len; 2111 sb->sb_sndptr = m->m_next; 2112 } 2113 off = 0; 2114 if (len != M_COPYALL) { 2115 len -= n->m_len; 2116 } 2117 m = m->m_next; 2118 np = &n->m_next; 2119 } 2120 return (top); 2121 nospace: 2122 m_freem(top); 2123 return (NULL); 2124 } 2125 2126 void 2127 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin) 2128 { 2129 2130 /* 2131 * Automatic sizing of send socket buffer. Often the send buffer 2132 * size is not optimally adjusted to the actual network conditions 2133 * at hand (delay bandwidth product). Setting the buffer size too 2134 * small limits throughput on links with high bandwidth and high 2135 * delay (eg. trans-continental/oceanic links). Setting the 2136 * buffer size too big consumes too much real kernel memory, 2137 * especially with many connections on busy servers. 2138 * 2139 * The criteria to step up the send buffer one notch are: 2140 * 1. receive window of remote host is larger than send buffer 2141 * (with a fudge factor of 5/4th); 2142 * 2. send buffer is filled to 7/8th with data (so we actually 2143 * have data to make use of it); 2144 * 3. send buffer fill has not hit maximal automatic size; 2145 * 4. our send window (slow start and cogestion controlled) is 2146 * larger than sent but unacknowledged data in send buffer. 2147 * 2148 * The remote host receive window scaling factor may limit the 2149 * growing of the send buffer before it reaches its allowed 2150 * maximum. 2151 * 2152 * It scales directly with slow start or congestion window 2153 * and does at most one step per received ACK. This fast 2154 * scaling has the drawback of growing the send buffer beyond 2155 * what is strictly necessary to make full use of a given 2156 * delay*bandwidth product. However testing has shown this not 2157 * to be much of an problem. At worst we are trading wasting 2158 * of available bandwidth (the non-use of it) for wasting some 2159 * socket buffer memory. 2160 * 2161 * TODO: Shrink send buffer during idle periods together 2162 * with congestion window. Requires another timer. Has to 2163 * wait for upcoming tcp timer rewrite. 2164 * 2165 * XXXGL: should there be used sbused() or sbavail()? 2166 */ 2167 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 2168 int lowat; 2169 2170 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0; 2171 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat && 2172 sbused(&so->so_snd) >= 2173 (so->so_snd.sb_hiwat / 8 * 7) - lowat && 2174 sbused(&so->so_snd) < V_tcp_autosndbuf_max && 2175 sendwin >= (sbused(&so->so_snd) - 2176 (tp->snd_nxt - tp->snd_una))) { 2177 if (!sbreserve_locked(so, SO_SND, 2178 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 2179 V_tcp_autosndbuf_max), curthread)) 2180 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 2181 } 2182 } 2183 } 2184