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