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