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 32 #include <sys/cdefs.h> 33 #include "opt_inet.h" 34 #include "opt_inet6.h" 35 #include "opt_ipsec.h" 36 #include "opt_kern_tls.h" 37 38 #include <sys/param.h> 39 #include <sys/systm.h> 40 #include <sys/arb.h> 41 #include <sys/domain.h> 42 #ifdef TCP_HHOOK 43 #include <sys/hhook.h> 44 #endif 45 #include <sys/kernel.h> 46 #ifdef KERN_TLS 47 #include <sys/ktls.h> 48 #endif 49 #include <sys/lock.h> 50 #include <sys/mbuf.h> 51 #include <sys/mutex.h> 52 #include <sys/protosw.h> 53 #include <sys/qmath.h> 54 #include <sys/sdt.h> 55 #include <sys/socket.h> 56 #include <sys/socketvar.h> 57 #include <sys/sysctl.h> 58 #include <sys/stats.h> 59 60 #include <net/if.h> 61 #include <net/route.h> 62 #include <net/route/nhop.h> 63 #include <net/vnet.h> 64 65 #include <netinet/in.h> 66 #include <netinet/in_kdtrace.h> 67 #include <netinet/in_systm.h> 68 #include <netinet/ip.h> 69 #include <netinet/in_pcb.h> 70 #include <netinet/ip_var.h> 71 #include <netinet/ip_options.h> 72 #ifdef INET6 73 #include <netinet6/in6_pcb.h> 74 #include <netinet/ip6.h> 75 #include <netinet6/ip6_var.h> 76 #endif 77 #include <netinet/tcp.h> 78 #define TCPOUTFLAGS 79 #include <netinet/tcp_fsm.h> 80 #include <netinet/tcp_seq.h> 81 #include <netinet/tcp_var.h> 82 #include <netinet/tcp_log_buf.h> 83 #include <netinet/tcp_syncache.h> 84 #include <netinet/tcp_timer.h> 85 #include <netinet/tcpip.h> 86 #include <netinet/cc/cc.h> 87 #include <netinet/tcp_fastopen.h> 88 #ifdef TCPPCAP 89 #include <netinet/tcp_pcap.h> 90 #endif 91 #ifdef TCP_OFFLOAD 92 #include <netinet/tcp_offload.h> 93 #endif 94 #include <netinet/tcp_ecn.h> 95 96 #include <netipsec/ipsec_support.h> 97 98 #include <netinet/udp.h> 99 #include <netinet/udp_var.h> 100 #include <machine/in_cksum.h> 101 102 #include <security/mac/mac_framework.h> 103 104 VNET_DEFINE(int, path_mtu_discovery) = 1; 105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW, 106 &VNET_NAME(path_mtu_discovery), 1, 107 "Enable Path MTU Discovery"); 108 109 VNET_DEFINE(int, tcp_do_tso) = 1; 110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW, 111 &VNET_NAME(tcp_do_tso), 0, 112 "Enable TCP Segmentation Offload"); 113 114 VNET_DEFINE(int, tcp_sendspace) = 1024*32; 115 #define V_tcp_sendspace VNET(tcp_sendspace) 116 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW, 117 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 118 119 VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, 121 &VNET_NAME(tcp_do_autosndbuf), 0, 122 "Enable automatic send buffer sizing"); 123 124 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW, 126 &VNET_NAME(tcp_autosndbuf_inc), 0, 127 "Incrementor step size of automatic send buffer"); 128 129 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; 130 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW, 131 &VNET_NAME(tcp_autosndbuf_max), 0, 132 "Max size of automatic send buffer"); 133 134 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0; 135 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat) 136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW, 137 &VNET_NAME(tcp_sendbuf_auto_lowat), 0, 138 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT"); 139 140 /* 141 * Make sure that either retransmit or persist timer is set for SYN, FIN and 142 * non-ACK. 143 */ 144 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \ 145 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\ 146 tcp_timer_active((tp), TT_REXMT) || \ 147 tcp_timer_active((tp), TT_PERSIST), \ 148 ("neither rexmt nor persist timer is set")) 149 150 #ifdef TCP_HHOOK 151 /* 152 * Wrapper for the TCP established output helper hook. 153 */ 154 void 155 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 156 struct tcpopt *to, uint32_t len, int tso) 157 { 158 struct tcp_hhook_data hhook_data; 159 160 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 161 hhook_data.tp = tp; 162 hhook_data.th = th; 163 hhook_data.to = to; 164 hhook_data.len = len; 165 hhook_data.tso = tso; 166 167 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 168 &tp->t_osd); 169 } 170 } 171 #endif 172 173 /* 174 * CC wrapper hook functions 175 */ 176 void 177 cc_after_idle(struct tcpcb *tp) 178 { 179 INP_WLOCK_ASSERT(tptoinpcb(tp)); 180 181 if (CC_ALGO(tp)->after_idle != NULL) 182 CC_ALGO(tp)->after_idle(&tp->t_ccv); 183 } 184 185 /* 186 * Tcp output routine: figure out what should be sent and send it. 187 */ 188 int 189 tcp_default_output(struct tcpcb *tp) 190 { 191 struct socket *so = tptosocket(tp); 192 struct inpcb *inp = tptoinpcb(tp); 193 int32_t len; 194 uint32_t recwin, sendwin; 195 uint16_t flags; 196 int off, error = 0; /* Keep compiler happy */ 197 u_int if_hw_tsomaxsegcount = 0; 198 u_int if_hw_tsomaxsegsize = 0; 199 struct mbuf *m; 200 struct ip *ip = NULL; 201 struct tcphdr *th; 202 u_char opt[TCP_MAXOLEN]; 203 unsigned ipoptlen, optlen, hdrlen, ulen; 204 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 205 unsigned ipsec_optlen = 0; 206 #endif 207 int idle, sendalot, curticks; 208 int sack_rxmit, sack_bytes_rxmt; 209 struct sackhole *p; 210 int tso, mtu; 211 struct tcpopt to; 212 struct udphdr *udp = NULL; 213 struct tcp_log_buffer *lgb; 214 unsigned int wanted_cookie = 0; 215 unsigned int dont_sendalot = 0; 216 #if 0 217 int maxburst = TCP_MAXBURST; 218 #endif 219 #ifdef INET6 220 struct ip6_hdr *ip6 = NULL; 221 const bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 222 #endif 223 #ifdef KERN_TLS 224 const bool hw_tls = tp->t_nic_ktls_xmit != 0; 225 #else 226 const bool hw_tls = false; 227 #endif 228 229 NET_EPOCH_ASSERT(); 230 INP_WLOCK_ASSERT(inp); 231 232 #ifdef TCP_OFFLOAD 233 if (tp->t_flags & TF_TOE) 234 return (tcp_offload_output(tp)); 235 #endif 236 237 /* 238 * For TFO connections in SYN_SENT or SYN_RECEIVED, 239 * only allow the initial SYN or SYN|ACK and those sent 240 * by the retransmit timer. 241 */ 242 if ((tp->t_flags & TF_FASTOPEN) && 243 ((tp->t_state == TCPS_SYN_SENT) || 244 (tp->t_state == TCPS_SYN_RECEIVED)) && 245 SEQ_GT(tp->snd_max, tp->snd_una) && /* SYN or SYN|ACK sent */ 246 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 247 return (0); 248 249 /* 250 * Determine length of data that should be transmitted, 251 * and flags that will be used. 252 * If there is some data or critical controls (SYN, RST) 253 * to send, then transmit; otherwise, investigate further. 254 */ 255 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 256 if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) || 257 (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur)))) 258 cc_after_idle(tp); 259 tp->t_flags &= ~TF_LASTIDLE; 260 if (idle) { 261 if (tp->t_flags & TF_MORETOCOME) { 262 tp->t_flags |= TF_LASTIDLE; 263 idle = 0; 264 } 265 } 266 again: 267 /* 268 * If we've recently taken a timeout, snd_max will be greater than 269 * snd_nxt. There may be SACK information that allows us to avoid 270 * resending already delivered data. Adjust snd_nxt accordingly. 271 */ 272 if ((tp->t_flags & TF_SACK_PERMIT) && 273 SEQ_LT(tp->snd_nxt, tp->snd_max)) 274 tcp_sack_adjust(tp); 275 sendalot = 0; 276 tso = 0; 277 mtu = 0; 278 off = tp->snd_nxt - tp->snd_una; 279 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 280 281 flags = tcp_outflags[tp->t_state]; 282 /* 283 * Send any SACK-generated retransmissions. If we're explicitly trying 284 * to send out new data (when sendalot is 1), bypass this function. 285 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 286 * we're replacing a (future) new transmission with a retransmission 287 * now, and we previously incremented snd_cwnd in tcp_input(). 288 */ 289 /* 290 * Still in sack recovery , reset rxmit flag to zero. 291 */ 292 sack_rxmit = 0; 293 sack_bytes_rxmt = 0; 294 len = 0; 295 p = NULL; 296 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 297 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 298 uint32_t cwin; 299 300 cwin = 301 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0); 302 /* Do not retransmit SACK segments beyond snd_recover */ 303 if (SEQ_GT(p->end, tp->snd_recover)) { 304 /* 305 * (At least) part of sack hole extends beyond 306 * snd_recover. Check to see if we can rexmit data 307 * for this hole. 308 */ 309 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 310 /* 311 * Can't rexmit any more data for this hole. 312 * That data will be rexmitted in the next 313 * sack recovery episode, when snd_recover 314 * moves past p->rxmit. 315 */ 316 p = NULL; 317 goto after_sack_rexmit; 318 } else { 319 /* Can rexmit part of the current hole */ 320 len = ((int32_t)ulmin(cwin, 321 SEQ_SUB(tp->snd_recover, p->rxmit))); 322 } 323 } else { 324 len = ((int32_t)ulmin(cwin, 325 SEQ_SUB(p->end, p->rxmit))); 326 } 327 if (len > 0) { 328 off = SEQ_SUB(p->rxmit, tp->snd_una); 329 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 330 __func__, off)); 331 sack_rxmit = 1; 332 sendalot = 1; 333 } 334 } 335 after_sack_rexmit: 336 /* 337 * Get standard flags, and add SYN or FIN if requested by 'hidden' 338 * state flags. 339 */ 340 if (tp->t_flags & TF_NEEDFIN) 341 flags |= TH_FIN; 342 if (tp->t_flags & TF_NEEDSYN) 343 flags |= TH_SYN; 344 345 SOCKBUF_LOCK(&so->so_snd); 346 /* 347 * If in persist timeout with window of 0, send 1 byte. 348 * Otherwise, if window is small but nonzero 349 * and timer expired, we will send what we can 350 * and go to transmit state. 351 */ 352 if (tp->t_flags & TF_FORCEDATA) { 353 if (sendwin == 0) { 354 /* 355 * If we still have some data to send, then 356 * clear the FIN bit. Usually this would 357 * happen below when it realizes that we 358 * aren't sending all the data. However, 359 * if we have exactly 1 byte of unsent data, 360 * then it won't clear the FIN bit below, 361 * and if we are in persist state, we wind 362 * up sending the packet without recording 363 * that we sent the FIN bit. 364 * 365 * We can't just blindly clear the FIN bit, 366 * because if we don't have any more data 367 * to send then the probe will be the FIN 368 * itself. 369 */ 370 if (off < sbused(&so->so_snd)) 371 flags &= ~TH_FIN; 372 sendwin = 1; 373 } else { 374 tcp_timer_activate(tp, TT_PERSIST, 0); 375 tp->t_rxtshift = 0; 376 } 377 } 378 379 /* 380 * If snd_nxt == snd_max and we have transmitted a FIN, the 381 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 382 * a negative length. This can also occur when TCP opens up 383 * its congestion window while receiving additional duplicate 384 * acks after fast-retransmit because TCP will reset snd_nxt 385 * to snd_max after the fast-retransmit. 386 * 387 * In the normal retransmit-FIN-only case, however, snd_nxt will 388 * be set to snd_una, the offset will be 0, and the length may 389 * wind up 0. 390 * 391 * If sack_rxmit is true we are retransmitting from the scoreboard 392 * in which case len is already set. 393 */ 394 if (sack_rxmit == 0) { 395 if (sack_bytes_rxmt == 0) { 396 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) - 397 off); 398 } else { 399 int32_t cwin; 400 401 /* 402 * We are inside of a SACK recovery episode and are 403 * sending new data, having retransmitted all the 404 * data possible in the scoreboard. 405 */ 406 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) - 407 off); 408 /* 409 * Don't remove this (len > 0) check ! 410 * We explicitly check for len > 0 here (although it 411 * isn't really necessary), to work around a gcc 412 * optimization issue - to force gcc to compute 413 * len above. Without this check, the computation 414 * of len is bungled by the optimizer. 415 */ 416 if (len > 0) { 417 cwin = tp->snd_cwnd - imax(0, (int32_t) 418 (tp->snd_nxt - tp->snd_recover)) - 419 sack_bytes_rxmt; 420 if (cwin < 0) 421 cwin = 0; 422 len = imin(len, cwin); 423 } 424 } 425 } 426 427 /* 428 * Lop off SYN bit if it has already been sent. However, if this 429 * is SYN-SENT state and if segment contains data and if we don't 430 * know that foreign host supports TAO, suppress sending segment. 431 */ 432 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 433 if (tp->t_state != TCPS_SYN_RECEIVED) 434 flags &= ~TH_SYN; 435 /* 436 * When sending additional segments following a TFO SYN|ACK, 437 * do not include the SYN bit. 438 */ 439 if ((tp->t_flags & TF_FASTOPEN) && 440 (tp->t_state == TCPS_SYN_RECEIVED)) 441 flags &= ~TH_SYN; 442 off--, len++; 443 } 444 445 /* 446 * Be careful not to send data and/or FIN on SYN segments. 447 * This measure is needed to prevent interoperability problems 448 * with not fully conformant TCP implementations. 449 */ 450 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 451 len = 0; 452 flags &= ~TH_FIN; 453 } 454 455 /* 456 * On TFO sockets, ensure no data is sent in the following cases: 457 * 458 * - When retransmitting SYN|ACK on a passively-created socket 459 * 460 * - When retransmitting SYN on an actively created socket 461 * 462 * - When sending a zero-length cookie (cookie request) on an 463 * actively created socket 464 * 465 * - When the socket is in the CLOSED state (RST is being sent) 466 */ 467 if ((tp->t_flags & TF_FASTOPEN) && 468 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 469 ((tp->t_state == TCPS_SYN_SENT) && 470 (tp->t_tfo_client_cookie_len == 0)) || 471 (flags & TH_RST))) 472 len = 0; 473 474 /* Without fast-open there should never be data sent on a SYN. */ 475 if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) { 476 len = 0; 477 } 478 479 if (len <= 0) { 480 /* 481 * If FIN has been sent but not acked, 482 * but we haven't been called to retransmit, 483 * len will be < 0. Otherwise, window shrank 484 * after we sent into it. If window shrank to 0, 485 * cancel pending retransmit, pull snd_nxt back 486 * to (closed) window, and set the persist timer 487 * if it isn't already going. If the window didn't 488 * close completely, just wait for an ACK. 489 * 490 * We also do a general check here to ensure that 491 * we will set the persist timer when we have data 492 * to send, but a 0-byte window. This makes sure 493 * the persist timer is set even if the packet 494 * hits one of the "goto send" lines below. 495 */ 496 len = 0; 497 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 498 (off < (int) sbavail(&so->so_snd)) && 499 !tcp_timer_active(tp, TT_PERSIST)) { 500 tcp_timer_activate(tp, TT_REXMT, 0); 501 tp->t_rxtshift = 0; 502 tp->snd_nxt = tp->snd_una; 503 if (!tcp_timer_active(tp, TT_PERSIST)) 504 tcp_setpersist(tp); 505 } 506 } 507 508 /* len will be >= 0 after this point. */ 509 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 510 511 tcp_sndbuf_autoscale(tp, so, sendwin); 512 513 /* 514 * Decide if we can use TCP Segmentation Offloading (if supported by 515 * hardware). 516 * 517 * TSO may only be used if we are in a pure bulk sending state. The 518 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 519 * IP options prevent using TSO. With TSO the TCP header is the same 520 * (except for the sequence number) for all generated packets. This 521 * makes it impossible to transmit any options which vary per generated 522 * segment or packet. 523 * 524 * IPv4 handling has a clear separation of ip options and ip header 525 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 526 * the right thing below to provide length of just ip options and thus 527 * checking for ipoptlen is enough to decide if ip options are present. 528 */ 529 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 530 /* 531 * Pre-calculate here as we save another lookup into the darknesses 532 * of IPsec that way and can actually decide if TSO is ok. 533 */ 534 #ifdef INET6 535 if (isipv6 && IPSEC_ENABLED(ipv6)) 536 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); 537 #ifdef INET 538 else 539 #endif 540 #endif /* INET6 */ 541 #ifdef INET 542 if (IPSEC_ENABLED(ipv4)) 543 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); 544 #endif /* INET */ 545 #endif /* IPSEC */ 546 #ifdef INET6 547 if (isipv6) 548 ipoptlen = ip6_optlen(inp); 549 else 550 #endif 551 if (inp->inp_options) 552 ipoptlen = inp->inp_options->m_len - 553 offsetof(struct ipoption, ipopt_list); 554 else 555 ipoptlen = 0; 556 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 557 ipoptlen += ipsec_optlen; 558 #endif 559 560 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 561 (tp->t_port == 0) && 562 ((tp->t_flags & TF_SIGNATURE) == 0) && 563 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 564 ipoptlen == 0 && !(flags & TH_SYN)) 565 tso = 1; 566 567 if (SEQ_LT((sack_rxmit ? p->rxmit : tp->snd_nxt) + len, 568 tp->snd_una + sbused(&so->so_snd))) { 569 flags &= ~TH_FIN; 570 } 571 572 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 573 (long)TCP_MAXWIN << tp->rcv_scale); 574 575 /* 576 * Sender silly window avoidance. We transmit under the following 577 * conditions when len is non-zero: 578 * 579 * - We have a full segment (or more with TSO) 580 * - This is the last buffer in a write()/send() and we are 581 * either idle or running NODELAY 582 * - we've timed out (e.g. persist timer) 583 * - we have more then 1/2 the maximum send window's worth of 584 * data (receiver may be limited the window size) 585 * - we need to retransmit 586 */ 587 if (len) { 588 if (len >= tp->t_maxseg) 589 goto send; 590 /* 591 * As the TCP header options are now 592 * considered when setting up the initial 593 * window, we would not send the last segment 594 * if we skip considering the option length here. 595 * Note: this may not work when tcp headers change 596 * very dynamically in the future. 597 */ 598 if ((((tp->t_flags & TF_SIGNATURE) ? 599 PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) + 600 ((tp->t_flags & TF_RCVD_TSTMP) ? 601 PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) + 602 len) >= tp->t_maxseg) 603 goto send; 604 /* 605 * NOTE! on localhost connections an 'ack' from the remote 606 * end may occur synchronously with the output and cause 607 * us to flush a buffer queued with moretocome. XXX 608 * 609 * note: the len + off check is almost certainly unnecessary. 610 */ 611 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 612 (idle || (tp->t_flags & TF_NODELAY)) && 613 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) && 614 (tp->t_flags & TF_NOPUSH) == 0) { 615 goto send; 616 } 617 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 618 goto send; 619 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 620 goto send; 621 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 622 goto send; 623 if (sack_rxmit) 624 goto send; 625 } 626 627 /* 628 * Sending of standalone window updates. 629 * 630 * Window updates are important when we close our window due to a 631 * full socket buffer and are opening it again after the application 632 * reads data from it. Once the window has opened again and the 633 * remote end starts to send again the ACK clock takes over and 634 * provides the most current window information. 635 * 636 * We must avoid the silly window syndrome whereas every read 637 * from the receive buffer, no matter how small, causes a window 638 * update to be sent. We also should avoid sending a flurry of 639 * window updates when the socket buffer had queued a lot of data 640 * and the application is doing small reads. 641 * 642 * Prevent a flurry of pointless window updates by only sending 643 * an update when we can increase the advertized window by more 644 * than 1/4th of the socket buffer capacity. When the buffer is 645 * getting full or is very small be more aggressive and send an 646 * update whenever we can increase by two mss sized segments. 647 * In all other situations the ACK's to new incoming data will 648 * carry further window increases. 649 * 650 * Don't send an independent window update if a delayed 651 * ACK is pending (it will get piggy-backed on it) or the 652 * remote side already has done a half-close and won't send 653 * more data. 654 */ 655 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 656 !(tp->t_flags & TF_DELACK) && 657 !TCPS_HAVERCVDFIN(tp->t_state)) { 658 /* 659 * "adv" is the amount we could increase the window, 660 * taking into account that we are limited by 661 * TCP_MAXWIN << tp->rcv_scale. 662 */ 663 int32_t adv; 664 int oldwin; 665 666 adv = recwin; 667 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 668 oldwin = (tp->rcv_adv - tp->rcv_nxt); 669 if (adv > oldwin) 670 adv -= oldwin; 671 else 672 adv = 0; 673 } else 674 oldwin = 0; 675 676 /* 677 * If the new window size ends up being the same as or less 678 * than the old size when it is scaled, then don't force 679 * a window update. 680 */ 681 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 682 goto dontupdate; 683 684 if (adv >= (int32_t)(2 * tp->t_maxseg) && 685 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 686 recwin <= (so->so_rcv.sb_hiwat / 8) || 687 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg || 688 adv >= TCP_MAXWIN << tp->rcv_scale)) 689 goto send; 690 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat) 691 goto send; 692 } 693 dontupdate: 694 695 /* 696 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 697 * is also a catch-all for the retransmit timer timeout case. 698 */ 699 if (tp->t_flags & TF_ACKNOW) 700 goto send; 701 if ((flags & TH_RST) || 702 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 703 goto send; 704 if (SEQ_GT(tp->snd_up, tp->snd_una)) 705 goto send; 706 /* 707 * If our state indicates that FIN should be sent 708 * and we have not yet done so, then we need to send. 709 */ 710 if (flags & TH_FIN && 711 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 712 goto send; 713 /* 714 * In SACK, it is possible for tcp_output to fail to send a segment 715 * after the retransmission timer has been turned off. Make sure 716 * that the retransmission timer is set. 717 */ 718 if ((tp->t_flags & TF_SACK_PERMIT) && 719 SEQ_GT(tp->snd_max, tp->snd_una) && 720 !tcp_timer_active(tp, TT_REXMT) && 721 !tcp_timer_active(tp, TT_PERSIST)) { 722 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp)); 723 goto just_return; 724 } 725 /* 726 * TCP window updates are not reliable, rather a polling protocol 727 * using ``persist'' packets is used to insure receipt of window 728 * updates. The three ``states'' for the output side are: 729 * idle not doing retransmits or persists 730 * persisting to move a small or zero window 731 * (re)transmitting and thereby not persisting 732 * 733 * tcp_timer_active(tp, TT_PERSIST) 734 * is true when we are in persist state. 735 * (tp->t_flags & TF_FORCEDATA) 736 * is set when we are called to send a persist packet. 737 * tcp_timer_active(tp, TT_REXMT) 738 * is set when we are retransmitting 739 * The output side is idle when both timers are zero. 740 * 741 * If send window is too small, there is data to transmit, and no 742 * retransmit or persist is pending, then go to persist state. 743 * If nothing happens soon, send when timer expires: 744 * if window is nonzero, transmit what we can, 745 * otherwise force out a byte. 746 */ 747 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && 748 !tcp_timer_active(tp, TT_PERSIST)) { 749 tp->t_rxtshift = 0; 750 tcp_setpersist(tp); 751 } 752 753 /* 754 * No reason to send a segment, just return. 755 */ 756 just_return: 757 SOCKBUF_UNLOCK(&so->so_snd); 758 return (0); 759 760 send: 761 SOCKBUF_LOCK_ASSERT(&so->so_snd); 762 if (len > 0) { 763 if (len >= tp->t_maxseg) 764 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 765 else 766 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 767 } 768 /* 769 * Before ESTABLISHED, force sending of initial options 770 * unless TCP set not to do any options. 771 * NOTE: we assume that the IP/TCP header plus TCP options 772 * always fit in a single mbuf, leaving room for a maximum 773 * link header, i.e. 774 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 775 */ 776 optlen = 0; 777 #ifdef INET6 778 if (isipv6) 779 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 780 else 781 #endif 782 hdrlen = sizeof (struct tcpiphdr); 783 784 if (flags & TH_SYN) { 785 tp->snd_nxt = tp->iss; 786 } 787 788 /* 789 * Compute options for segment. 790 * We only have to care about SYN and established connection 791 * segments. Options for SYN-ACK segments are handled in TCP 792 * syncache. 793 */ 794 to.to_flags = 0; 795 if ((tp->t_flags & TF_NOOPT) == 0) { 796 /* Maximum segment size. */ 797 if (flags & TH_SYN) { 798 to.to_mss = tcp_mssopt(&inp->inp_inc); 799 if (tp->t_port) 800 to.to_mss -= V_tcp_udp_tunneling_overhead; 801 to.to_flags |= TOF_MSS; 802 803 /* 804 * On SYN or SYN|ACK transmits on TFO connections, 805 * only include the TFO option if it is not a 806 * retransmit, as the presence of the TFO option may 807 * have caused the original SYN or SYN|ACK to have 808 * been dropped by a middlebox. 809 */ 810 if ((tp->t_flags & TF_FASTOPEN) && 811 (tp->t_rxtshift == 0)) { 812 if (tp->t_state == TCPS_SYN_RECEIVED) { 813 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 814 to.to_tfo_cookie = 815 (u_int8_t *)&tp->t_tfo_cookie.server; 816 to.to_flags |= TOF_FASTOPEN; 817 wanted_cookie = 1; 818 } else if (tp->t_state == TCPS_SYN_SENT) { 819 to.to_tfo_len = 820 tp->t_tfo_client_cookie_len; 821 to.to_tfo_cookie = 822 tp->t_tfo_cookie.client; 823 to.to_flags |= TOF_FASTOPEN; 824 wanted_cookie = 1; 825 /* 826 * If we wind up having more data to 827 * send with the SYN than can fit in 828 * one segment, don't send any more 829 * until the SYN|ACK comes back from 830 * the other end. 831 */ 832 dont_sendalot = 1; 833 } 834 } 835 } 836 /* Window scaling. */ 837 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 838 to.to_wscale = tp->request_r_scale; 839 to.to_flags |= TOF_SCALE; 840 } 841 /* Timestamps. */ 842 if ((tp->t_flags & TF_RCVD_TSTMP) || 843 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 844 curticks = tcp_ts_getticks(); 845 to.to_tsval = curticks + tp->ts_offset; 846 to.to_tsecr = tp->ts_recent; 847 to.to_flags |= TOF_TS; 848 if (tp->t_rxtshift == 1) 849 tp->t_badrxtwin = curticks; 850 } 851 852 /* Set receive buffer autosizing timestamp. */ 853 if (tp->rfbuf_ts == 0 && 854 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 855 tp->rfbuf_ts = tcp_ts_getticks(); 856 857 /* Selective ACK's. */ 858 if (tp->t_flags & TF_SACK_PERMIT) { 859 if (flags & TH_SYN) 860 to.to_flags |= TOF_SACKPERM; 861 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 862 tp->rcv_numsacks > 0) { 863 to.to_flags |= TOF_SACK; 864 to.to_nsacks = tp->rcv_numsacks; 865 to.to_sacks = (u_char *)tp->sackblks; 866 } 867 } 868 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 869 /* TCP-MD5 (RFC2385). */ 870 /* 871 * Check that TCP_MD5SIG is enabled in tcpcb to 872 * account the size needed to set this TCP option. 873 */ 874 if (tp->t_flags & TF_SIGNATURE) 875 to.to_flags |= TOF_SIGNATURE; 876 #endif /* TCP_SIGNATURE */ 877 878 /* Processing the options. */ 879 hdrlen += optlen = tcp_addoptions(&to, opt); 880 /* 881 * If we wanted a TFO option to be added, but it was unable 882 * to fit, ensure no data is sent. 883 */ 884 if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie && 885 !(to.to_flags & TOF_FASTOPEN)) 886 len = 0; 887 } 888 if (tp->t_port) { 889 if (V_tcp_udp_tunneling_port == 0) { 890 /* The port was removed?? */ 891 SOCKBUF_UNLOCK(&so->so_snd); 892 return (EHOSTUNREACH); 893 } 894 hdrlen += sizeof(struct udphdr); 895 } 896 /* 897 * Adjust data length if insertion of options will 898 * bump the packet length beyond the t_maxseg length. 899 * Clear the FIN bit because we cut off the tail of 900 * the segment. 901 */ 902 if (len + optlen + ipoptlen > tp->t_maxseg) { 903 flags &= ~TH_FIN; 904 905 if (tso) { 906 u_int if_hw_tsomax; 907 u_int moff; 908 int max_len; 909 910 /* extract TSO information */ 911 if_hw_tsomax = tp->t_tsomax; 912 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 913 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 914 915 /* 916 * Limit a TSO burst to prevent it from 917 * overflowing or exceeding the maximum length 918 * allowed by the network interface: 919 */ 920 KASSERT(ipoptlen == 0, 921 ("%s: TSO can't do IP options", __func__)); 922 923 /* 924 * Check if we should limit by maximum payload 925 * length: 926 */ 927 if (if_hw_tsomax != 0) { 928 /* compute maximum TSO length */ 929 max_len = (if_hw_tsomax - hdrlen - 930 max_linkhdr); 931 if (max_len <= 0) { 932 len = 0; 933 } else if (len > max_len) { 934 sendalot = 1; 935 len = max_len; 936 } 937 } 938 939 /* 940 * Prevent the last segment from being 941 * fractional unless the send sockbuf can be 942 * emptied: 943 */ 944 max_len = (tp->t_maxseg - optlen); 945 if (((uint32_t)off + (uint32_t)len) < 946 sbavail(&so->so_snd)) { 947 moff = len % max_len; 948 if (moff != 0) { 949 len -= moff; 950 sendalot = 1; 951 } 952 } 953 954 /* 955 * In case there are too many small fragments 956 * don't use TSO: 957 */ 958 if (len <= max_len) { 959 len = max_len; 960 sendalot = 1; 961 tso = 0; 962 } 963 964 /* 965 * Send the FIN in a separate segment 966 * after the bulk sending is done. 967 * We don't trust the TSO implementations 968 * to clear the FIN flag on all but the 969 * last segment. 970 */ 971 if (tp->t_flags & TF_NEEDFIN) 972 sendalot = 1; 973 } else { 974 if (optlen + ipoptlen >= tp->t_maxseg) { 975 /* 976 * Since we don't have enough space to put 977 * the IP header chain and the TCP header in 978 * one packet as required by RFC 7112, don't 979 * send it. Also ensure that at least one 980 * byte of the payload can be put into the 981 * TCP segment. 982 */ 983 SOCKBUF_UNLOCK(&so->so_snd); 984 error = EMSGSIZE; 985 sack_rxmit = 0; 986 goto out; 987 } 988 len = tp->t_maxseg - optlen - ipoptlen; 989 sendalot = 1; 990 if (dont_sendalot) 991 sendalot = 0; 992 } 993 } else 994 tso = 0; 995 996 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 997 ("%s: len > IP_MAXPACKET", __func__)); 998 999 /*#ifdef DIAGNOSTIC*/ 1000 #ifdef INET6 1001 if (max_linkhdr + hdrlen > MCLBYTES) 1002 #else 1003 if (max_linkhdr + hdrlen > MHLEN) 1004 #endif 1005 panic("tcphdr too big"); 1006 /*#endif*/ 1007 1008 /* 1009 * This KASSERT is here to catch edge cases at a well defined place. 1010 * Before, those had triggered (random) panic conditions further down. 1011 */ 1012 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 1013 1014 /* 1015 * Grab a header mbuf, attaching a copy of data to 1016 * be transmitted, and initialize the header from 1017 * the template for sends on this connection. 1018 */ 1019 if (len) { 1020 struct mbuf *mb; 1021 struct sockbuf *msb; 1022 u_int moff; 1023 1024 if ((tp->t_flags & TF_FORCEDATA) && len == 1) { 1025 TCPSTAT_INC(tcps_sndprobe); 1026 #ifdef STATS 1027 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1028 stats_voi_update_abs_u32(tp->t_stats, 1029 VOI_TCP_RETXPB, len); 1030 else 1031 stats_voi_update_abs_u64(tp->t_stats, 1032 VOI_TCP_TXPB, len); 1033 #endif /* STATS */ 1034 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 1035 tp->t_sndrexmitpack++; 1036 TCPSTAT_INC(tcps_sndrexmitpack); 1037 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 1038 if (sack_rxmit) { 1039 TCPSTAT_INC(tcps_sack_rexmits); 1040 TCPSTAT_ADD(tcps_sack_rexmit_bytes, len); 1041 } 1042 #ifdef STATS 1043 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 1044 len); 1045 #endif /* STATS */ 1046 } else { 1047 TCPSTAT_INC(tcps_sndpack); 1048 TCPSTAT_ADD(tcps_sndbyte, len); 1049 #ifdef STATS 1050 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 1051 len); 1052 #endif /* STATS */ 1053 } 1054 #ifdef INET6 1055 if (MHLEN < hdrlen + max_linkhdr) 1056 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1057 else 1058 #endif 1059 m = m_gethdr(M_NOWAIT, MT_DATA); 1060 1061 if (m == NULL) { 1062 SOCKBUF_UNLOCK(&so->so_snd); 1063 error = ENOBUFS; 1064 sack_rxmit = 0; 1065 goto out; 1066 } 1067 1068 m->m_data += max_linkhdr; 1069 m->m_len = hdrlen; 1070 1071 /* 1072 * Start the m_copy functions from the closest mbuf 1073 * to the offset in the socket buffer chain. 1074 */ 1075 mb = sbsndptr_noadv(&so->so_snd, off, &moff); 1076 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 1077 m_copydata(mb, moff, len, 1078 mtod(m, caddr_t) + hdrlen); 1079 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1080 sbsndptr_adv(&so->so_snd, mb, len); 1081 m->m_len += len; 1082 } else { 1083 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1084 msb = NULL; 1085 else 1086 msb = &so->so_snd; 1087 m->m_next = tcp_m_copym(mb, moff, 1088 &len, if_hw_tsomaxsegcount, 1089 if_hw_tsomaxsegsize, msb, hw_tls); 1090 if (len <= (tp->t_maxseg - optlen)) { 1091 /* 1092 * Must have ran out of mbufs for the copy 1093 * shorten it to no longer need tso. Lets 1094 * not put on sendalot since we are low on 1095 * mbufs. 1096 */ 1097 tso = 0; 1098 } 1099 if (m->m_next == NULL) { 1100 SOCKBUF_UNLOCK(&so->so_snd); 1101 (void) m_free(m); 1102 error = ENOBUFS; 1103 sack_rxmit = 0; 1104 goto out; 1105 } 1106 } 1107 1108 /* 1109 * If we're sending everything we've got, set PUSH. 1110 * (This will keep happy those implementations which only 1111 * give data to the user when a buffer fills or 1112 * a PUSH comes in.) 1113 */ 1114 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) && 1115 !(flags & TH_SYN)) 1116 flags |= TH_PUSH; 1117 SOCKBUF_UNLOCK(&so->so_snd); 1118 } else { 1119 SOCKBUF_UNLOCK(&so->so_snd); 1120 if (tp->t_flags & TF_ACKNOW) 1121 TCPSTAT_INC(tcps_sndacks); 1122 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1123 TCPSTAT_INC(tcps_sndctrl); 1124 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1125 TCPSTAT_INC(tcps_sndurg); 1126 else 1127 TCPSTAT_INC(tcps_sndwinup); 1128 1129 m = m_gethdr(M_NOWAIT, MT_DATA); 1130 if (m == NULL) { 1131 error = ENOBUFS; 1132 sack_rxmit = 0; 1133 goto out; 1134 } 1135 #ifdef INET6 1136 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1137 MHLEN >= hdrlen) { 1138 M_ALIGN(m, hdrlen); 1139 } else 1140 #endif 1141 m->m_data += max_linkhdr; 1142 m->m_len = hdrlen; 1143 } 1144 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1145 m->m_pkthdr.rcvif = (struct ifnet *)0; 1146 #ifdef MAC 1147 mac_inpcb_create_mbuf(inp, m); 1148 #endif 1149 #ifdef INET6 1150 if (isipv6) { 1151 ip6 = mtod(m, struct ip6_hdr *); 1152 if (tp->t_port) { 1153 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); 1154 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1155 udp->uh_dport = tp->t_port; 1156 ulen = hdrlen + len - sizeof(struct ip6_hdr); 1157 udp->uh_ulen = htons(ulen); 1158 th = (struct tcphdr *)(udp + 1); 1159 } else { 1160 th = (struct tcphdr *)(ip6 + 1); 1161 } 1162 tcpip_fillheaders(inp, tp->t_port, ip6, th); 1163 } else 1164 #endif /* INET6 */ 1165 { 1166 ip = mtod(m, struct ip *); 1167 if (tp->t_port) { 1168 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); 1169 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1170 udp->uh_dport = tp->t_port; 1171 ulen = hdrlen + len - sizeof(struct ip); 1172 udp->uh_ulen = htons(ulen); 1173 th = (struct tcphdr *)(udp + 1); 1174 } else 1175 th = (struct tcphdr *)(ip + 1); 1176 tcpip_fillheaders(inp, tp->t_port, ip, th); 1177 } 1178 1179 /* 1180 * Fill in fields, remembering maximum advertised 1181 * window for use in delaying messages about window sizes. 1182 * If resending a FIN, be sure not to use a new sequence number. 1183 */ 1184 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1185 tp->snd_nxt == tp->snd_max) 1186 tp->snd_nxt--; 1187 /* 1188 * If we are starting a connection, send ECN setup 1189 * SYN packet. If we are on a retransmit, we may 1190 * resend those bits a number of times as per 1191 * RFC 3168. 1192 */ 1193 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 1194 flags |= tcp_ecn_output_syn_sent(tp); 1195 } 1196 /* Also handle parallel SYN for ECN */ 1197 if ((TCPS_HAVERCVDSYN(tp->t_state)) && 1198 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) { 1199 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit); 1200 if ((tp->t_state == TCPS_SYN_RECEIVED) && 1201 (tp->t_flags2 & TF2_ECN_SND_ECE)) 1202 tp->t_flags2 &= ~TF2_ECN_SND_ECE; 1203 #ifdef INET6 1204 if (isipv6) { 1205 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN); 1206 ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN); 1207 } 1208 else 1209 #endif 1210 { 1211 ip->ip_tos &= ~IPTOS_ECN_MASK; 1212 ip->ip_tos |= ect; 1213 } 1214 } 1215 1216 /* 1217 * If we are doing retransmissions, then snd_nxt will 1218 * not reflect the first unsent octet. For ACK only 1219 * packets, we do not want the sequence number of the 1220 * retransmitted packet, we want the sequence number 1221 * of the next unsent octet. So, if there is no data 1222 * (and no SYN or FIN), use snd_max instead of snd_nxt 1223 * when filling in ti_seq. But if we are in persist 1224 * state, snd_max might reflect one byte beyond the 1225 * right edge of the window, so use snd_nxt in that 1226 * case, since we know we aren't doing a retransmission. 1227 * (retransmit and persist are mutually exclusive...) 1228 */ 1229 if (sack_rxmit == 0) { 1230 if (len || (flags & (TH_SYN|TH_FIN)) || 1231 tcp_timer_active(tp, TT_PERSIST)) 1232 th->th_seq = htonl(tp->snd_nxt); 1233 else 1234 th->th_seq = htonl(tp->snd_max); 1235 } else { 1236 th->th_seq = htonl(p->rxmit); 1237 p->rxmit += len; 1238 /* 1239 * Lost Retransmission Detection 1240 * trigger resending of a (then 1241 * still existing) hole, when 1242 * fack acks recoverypoint. 1243 */ 1244 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end)) 1245 p->rxmit = tp->snd_recover; 1246 tp->sackhint.sack_bytes_rexmit += len; 1247 } 1248 if (IN_RECOVERY(tp->t_flags)) { 1249 /* 1250 * Account all bytes transmitted while 1251 * IN_RECOVERY, simplifying PRR and 1252 * Lost Retransmit Detection 1253 */ 1254 tp->sackhint.prr_out += len; 1255 } 1256 th->th_ack = htonl(tp->rcv_nxt); 1257 if (optlen) { 1258 bcopy(opt, th + 1, optlen); 1259 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1260 } 1261 tcp_set_flags(th, flags); 1262 /* 1263 * Calculate receive window. Don't shrink window, 1264 * but avoid silly window syndrome. 1265 * If a RST segment is sent, advertise a window of zero. 1266 */ 1267 if (flags & TH_RST) { 1268 recwin = 0; 1269 } else { 1270 if (recwin < (so->so_rcv.sb_hiwat / 4) && 1271 recwin < tp->t_maxseg) 1272 recwin = 0; 1273 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1274 recwin < (tp->rcv_adv - tp->rcv_nxt)) 1275 recwin = (tp->rcv_adv - tp->rcv_nxt); 1276 } 1277 /* 1278 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1279 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1280 * case is handled in syncache. 1281 */ 1282 if (flags & TH_SYN) 1283 th->th_win = htons((u_short) 1284 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1285 else { 1286 /* Avoid shrinking window with window scaling. */ 1287 recwin = roundup2(recwin, 1 << tp->rcv_scale); 1288 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1289 } 1290 1291 /* 1292 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1293 * a 0 window. This may cause the remote transmitter to stall. This 1294 * flag tells soreceive() to disable delayed acknowledgements when 1295 * draining the buffer. This can occur if the receiver is attempting 1296 * to read more data than can be buffered prior to transmitting on 1297 * the connection. 1298 */ 1299 if (th->th_win == 0) { 1300 tp->t_sndzerowin++; 1301 tp->t_flags |= TF_RXWIN0SENT; 1302 } else 1303 tp->t_flags &= ~TF_RXWIN0SENT; 1304 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1305 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1306 th->th_flags |= TH_URG; 1307 } else 1308 /* 1309 * If no urgent pointer to send, then we pull 1310 * the urgent pointer to the left edge of the send window 1311 * so that it doesn't drift into the send window on sequence 1312 * number wraparound. 1313 */ 1314 tp->snd_up = tp->snd_una; /* drag it along */ 1315 1316 /* 1317 * Put TCP length in extended header, and then 1318 * checksum extended header and data. 1319 */ 1320 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1321 1322 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1323 if (to.to_flags & TOF_SIGNATURE) { 1324 /* 1325 * Calculate MD5 signature and put it into the place 1326 * determined before. 1327 * NOTE: since TCP options buffer doesn't point into 1328 * mbuf's data, calculate offset and use it. 1329 */ 1330 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th, 1331 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) { 1332 /* 1333 * Do not send segment if the calculation of MD5 1334 * digest has failed. 1335 */ 1336 m_freem(m); 1337 goto out; 1338 } 1339 } 1340 #endif 1341 #ifdef INET6 1342 if (isipv6) { 1343 /* 1344 * There is no need to fill in ip6_plen right now. 1345 * It will be filled later by ip6_output. 1346 */ 1347 if (tp->t_port) { 1348 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 1349 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 1350 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 1351 th->th_sum = htons(0); 1352 UDPSTAT_INC(udps_opackets); 1353 } else { 1354 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1355 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1356 th->th_sum = in6_cksum_pseudo(ip6, 1357 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 1358 0); 1359 } 1360 } 1361 #endif 1362 #if defined(INET6) && defined(INET) 1363 else 1364 #endif 1365 #ifdef INET 1366 { 1367 if (tp->t_port) { 1368 m->m_pkthdr.csum_flags = CSUM_UDP; 1369 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 1370 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 1371 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 1372 th->th_sum = htons(0); 1373 UDPSTAT_INC(udps_opackets); 1374 } else { 1375 m->m_pkthdr.csum_flags = CSUM_TCP; 1376 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1377 th->th_sum = in_pseudo(ip->ip_src.s_addr, 1378 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 1379 IPPROTO_TCP + len + optlen)); 1380 } 1381 1382 /* IP version must be set here for ipv4/ipv6 checking later */ 1383 KASSERT(ip->ip_v == IPVERSION, 1384 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1385 } 1386 #endif 1387 1388 /* 1389 * Enable TSO and specify the size of the segments. 1390 * The TCP pseudo header checksum is always provided. 1391 */ 1392 if (tso) { 1393 KASSERT(len > tp->t_maxseg - optlen, 1394 ("%s: len <= tso_segsz", __func__)); 1395 m->m_pkthdr.csum_flags |= CSUM_TSO; 1396 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 1397 } 1398 1399 KASSERT(len + hdrlen == m_length(m, NULL), 1400 ("%s: mbuf chain shorter than expected: %d + %u != %u", 1401 __func__, len, hdrlen, m_length(m, NULL))); 1402 1403 #ifdef TCP_HHOOK 1404 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1405 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1406 #endif 1407 1408 TCP_PROBE3(debug__output, tp, th, m); 1409 1410 /* We're getting ready to send; log now. */ 1411 /* XXXMT: We are not honoring verbose logging. */ 1412 1413 if (tcp_bblogging_on(tp)) 1414 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, 1415 TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0, 1416 NULL); 1417 else 1418 lgb = NULL; 1419 1420 /* 1421 * Fill in IP length and desired time to live and 1422 * send to IP level. There should be a better way 1423 * to handle ttl and tos; we could keep them in 1424 * the template, but need a way to checksum without them. 1425 */ 1426 /* 1427 * m->m_pkthdr.len should have been set before checksum calculation, 1428 * because in6_cksum() need it. 1429 */ 1430 #ifdef INET6 1431 if (isipv6) { 1432 /* 1433 * we separately set hoplimit for every segment, since the 1434 * user might want to change the value via setsockopt. 1435 * Also, desired default hop limit might be changed via 1436 * Neighbor Discovery. 1437 */ 1438 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 1439 1440 /* 1441 * Set the packet size here for the benefit of DTrace probes. 1442 * ip6_output() will set it properly; it's supposed to include 1443 * the option header lengths as well. 1444 */ 1445 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1446 1447 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 1448 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1449 else 1450 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1451 1452 if (tp->t_state == TCPS_SYN_SENT) 1453 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1454 1455 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1456 1457 #ifdef TCPPCAP 1458 /* Save packet, if requested. */ 1459 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1460 #endif 1461 1462 /* TODO: IPv6 IP6TOS_ECT bit on */ 1463 error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6, 1464 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1465 NULL, NULL, inp); 1466 1467 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) 1468 mtu = inp->inp_route6.ro_nh->nh_mtu; 1469 } 1470 #endif /* INET6 */ 1471 #if defined(INET) && defined(INET6) 1472 else 1473 #endif 1474 #ifdef INET 1475 { 1476 ip->ip_len = htons(m->m_pkthdr.len); 1477 #ifdef INET6 1478 if (inp->inp_vflag & INP_IPV6PROTO) 1479 ip->ip_ttl = in6_selecthlim(inp, NULL); 1480 #endif /* INET6 */ 1481 /* 1482 * If we do path MTU discovery, then we set DF on every packet. 1483 * This might not be the best thing to do according to RFC3390 1484 * Section 2. However the tcp hostcache migitates the problem 1485 * so it affects only the first tcp connection with a host. 1486 * 1487 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1488 */ 1489 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 1490 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1491 if (tp->t_port == 0 || len < V_tcp_minmss) { 1492 ip->ip_off |= htons(IP_DF); 1493 } 1494 } else { 1495 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1496 } 1497 1498 if (tp->t_state == TCPS_SYN_SENT) 1499 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1500 1501 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1502 1503 #ifdef TCPPCAP 1504 /* Save packet, if requested. */ 1505 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1506 #endif 1507 1508 error = ip_output(m, inp->inp_options, &inp->inp_route, 1509 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp); 1510 1511 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) 1512 mtu = inp->inp_route.ro_nh->nh_mtu; 1513 } 1514 #endif /* INET */ 1515 1516 if (lgb != NULL) { 1517 lgb->tlb_errno = error; 1518 lgb = NULL; 1519 } 1520 out: 1521 if (error == 0) 1522 tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls); 1523 /* 1524 * In transmit state, time the transmission and arrange for 1525 * the retransmit. In persist state, just set snd_max. In a closed 1526 * state just return. 1527 */ 1528 if (flags & TH_RST) { 1529 TCPSTAT_INC(tcps_sndtotal); 1530 return (0); 1531 } else if ((tp->t_flags & TF_FORCEDATA) == 0 || 1532 !tcp_timer_active(tp, TT_PERSIST)) { 1533 tcp_seq startseq = tp->snd_nxt; 1534 1535 /* 1536 * Advance snd_nxt over sequence space of this segment. 1537 */ 1538 if (flags & (TH_SYN|TH_FIN)) { 1539 if (flags & TH_SYN) 1540 tp->snd_nxt++; 1541 if (flags & TH_FIN) { 1542 tp->snd_nxt++; 1543 tp->t_flags |= TF_SENTFIN; 1544 } 1545 } 1546 if (sack_rxmit) 1547 goto timer; 1548 tp->snd_nxt += len; 1549 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1550 /* 1551 * Update "made progress" indication if we just 1552 * added new data to an empty socket buffer. 1553 */ 1554 if (tp->snd_una == tp->snd_max) 1555 tp->t_acktime = ticks; 1556 tp->snd_max = tp->snd_nxt; 1557 /* 1558 * Time this transmission if not a retransmission and 1559 * not currently timing anything. 1560 */ 1561 tp->t_sndtime = ticks; 1562 if (tp->t_rtttime == 0) { 1563 tp->t_rtttime = ticks; 1564 tp->t_rtseq = startseq; 1565 TCPSTAT_INC(tcps_segstimed); 1566 } 1567 #ifdef STATS 1568 if (!(tp->t_flags & TF_GPUTINPROG) && len) { 1569 tp->t_flags |= TF_GPUTINPROG; 1570 tp->gput_seq = startseq; 1571 tp->gput_ack = startseq + 1572 ulmin(sbavail(&so->so_snd) - off, sendwin); 1573 tp->gput_ts = tcp_ts_getticks(); 1574 } 1575 #endif /* STATS */ 1576 } 1577 1578 /* 1579 * Set retransmit timer if not currently set, 1580 * and not doing a pure ack or a keep-alive probe. 1581 * Initial value for retransmit timer is smoothed 1582 * round-trip time + 2 * round-trip time variance. 1583 * Initialize shift counter which is used for backoff 1584 * of retransmit time. 1585 */ 1586 timer: 1587 if (!tcp_timer_active(tp, TT_REXMT) && 1588 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1589 (tp->snd_nxt != tp->snd_una))) { 1590 if (tcp_timer_active(tp, TT_PERSIST)) { 1591 tcp_timer_activate(tp, TT_PERSIST, 0); 1592 tp->t_rxtshift = 0; 1593 } 1594 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp)); 1595 } else if (len == 0 && sbavail(&so->so_snd) && 1596 !tcp_timer_active(tp, TT_REXMT) && 1597 !tcp_timer_active(tp, TT_PERSIST)) { 1598 /* 1599 * Avoid a situation where we do not set persist timer 1600 * after a zero window condition. For example: 1601 * 1) A -> B: packet with enough data to fill the window 1602 * 2) B -> A: ACK for #1 + new data (0 window 1603 * advertisement) 1604 * 3) A -> B: ACK for #2, 0 len packet 1605 * 1606 * In this case, A will not activate the persist timer, 1607 * because it chose to send a packet. Unless tcp_output 1608 * is called for some other reason (delayed ack timer, 1609 * another input packet from B, socket syscall), A will 1610 * not send zero window probes. 1611 * 1612 * So, if you send a 0-length packet, but there is data 1613 * in the socket buffer, and neither the rexmt or 1614 * persist timer is already set, then activate the 1615 * persist timer. 1616 */ 1617 tp->t_rxtshift = 0; 1618 tcp_setpersist(tp); 1619 } 1620 } else { 1621 /* 1622 * Persist case, update snd_max but since we are in 1623 * persist mode (no window) we do not update snd_nxt. 1624 */ 1625 int xlen = len; 1626 if (flags & TH_SYN) 1627 ++xlen; 1628 if (flags & TH_FIN) { 1629 ++xlen; 1630 tp->t_flags |= TF_SENTFIN; 1631 } 1632 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1633 tp->snd_max = tp->snd_nxt + xlen; 1634 } 1635 if ((error == 0) && 1636 (TCPS_HAVEESTABLISHED(tp->t_state) && 1637 (tp->t_flags & TF_SACK_PERMIT) && 1638 tp->rcv_numsacks > 0)) { 1639 /* Clean up any DSACK's sent */ 1640 tcp_clean_dsack_blocks(tp); 1641 } 1642 if (error) { 1643 /* 1644 * We know that the packet was lost, so back out the 1645 * sequence number advance, if any. 1646 * 1647 * If the error is EPERM the packet got blocked by the 1648 * local firewall. Normally we should terminate the 1649 * connection but the blocking may have been spurious 1650 * due to a firewall reconfiguration cycle. So we treat 1651 * it like a packet loss and let the retransmit timer and 1652 * timeouts do their work over time. 1653 * XXX: It is a POLA question whether calling tcp_drop right 1654 * away would be the really correct behavior instead. 1655 */ 1656 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1657 !tcp_timer_active(tp, TT_PERSIST)) && 1658 ((flags & TH_SYN) == 0) && 1659 (error != EPERM)) { 1660 if (sack_rxmit) { 1661 p->rxmit = SEQ_MIN(p->end, p->rxmit) - len; 1662 tp->sackhint.sack_bytes_rexmit -= len; 1663 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1664 ("sackhint bytes rtx >= 0")); 1665 KASSERT((flags & TH_FIN) == 0, 1666 ("error while FIN with SACK rxmit")); 1667 } else { 1668 tp->snd_nxt -= len; 1669 if (flags & TH_FIN) 1670 tp->snd_nxt--; 1671 } 1672 if (IN_RECOVERY(tp->t_flags)) 1673 tp->sackhint.prr_out -= len; 1674 } 1675 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1676 switch (error) { 1677 case EACCES: 1678 case EPERM: 1679 tp->t_softerror = error; 1680 return (error); 1681 case ENOBUFS: 1682 TCP_XMIT_TIMER_ASSERT(tp, len, flags); 1683 tp->snd_cwnd = tp->t_maxseg; 1684 return (0); 1685 case EMSGSIZE: 1686 /* 1687 * For some reason the interface we used initially 1688 * to send segments changed to another or lowered 1689 * its MTU. 1690 * If TSO was active we either got an interface 1691 * without TSO capabilits or TSO was turned off. 1692 * If we obtained mtu from ip_output() then update 1693 * it and try again. 1694 */ 1695 if (tso) 1696 tp->t_flags &= ~TF_TSO; 1697 if (mtu != 0) { 1698 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1699 goto again; 1700 } 1701 return (error); 1702 case EHOSTDOWN: 1703 case EHOSTUNREACH: 1704 case ENETDOWN: 1705 case ENETUNREACH: 1706 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1707 tp->t_softerror = error; 1708 return (0); 1709 } 1710 /* FALLTHROUGH */ 1711 default: 1712 return (error); 1713 } 1714 } 1715 TCPSTAT_INC(tcps_sndtotal); 1716 1717 /* 1718 * Data sent (as far as we can tell). 1719 * If this advertises a larger window than any other segment, 1720 * then remember the size of the advertised window. 1721 * Any pending ACK has now been sent. 1722 */ 1723 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1724 tp->rcv_adv = tp->rcv_nxt + recwin; 1725 tp->last_ack_sent = tp->rcv_nxt; 1726 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1727 if (tcp_timer_active(tp, TT_DELACK)) 1728 tcp_timer_activate(tp, TT_DELACK, 0); 1729 if (sendalot) 1730 goto again; 1731 return (0); 1732 } 1733 1734 void 1735 tcp_setpersist(struct tcpcb *tp) 1736 { 1737 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1738 int tt; 1739 int maxunacktime; 1740 1741 tp->t_flags &= ~TF_PREVVALID; 1742 if (tcp_timer_active(tp, TT_REXMT)) 1743 panic("tcp_setpersist: retransmit pending"); 1744 /* 1745 * If the state is already closed, don't bother. 1746 */ 1747 if (tp->t_state == TCPS_CLOSED) 1748 return; 1749 1750 /* 1751 * Start/restart persistence timer. 1752 */ 1753 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1754 tcp_persmin, tcp_persmax); 1755 if (TP_MAXUNACKTIME(tp) && tp->t_acktime) { 1756 maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks; 1757 if (maxunacktime < 1) 1758 maxunacktime = 1; 1759 if (maxunacktime < tt) 1760 tt = maxunacktime; 1761 } 1762 tcp_timer_activate(tp, TT_PERSIST, tt); 1763 if (tp->t_rxtshift < V_tcp_retries) 1764 tp->t_rxtshift++; 1765 } 1766 1767 /* 1768 * Insert TCP options according to the supplied parameters to the place 1769 * optp in a consistent way. Can handle unaligned destinations. 1770 * 1771 * The order of the option processing is crucial for optimal packing and 1772 * alignment for the scarce option space. 1773 * 1774 * The optimal order for a SYN/SYN-ACK segment is: 1775 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1776 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1777 * 1778 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1779 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1780 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1781 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1782 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1783 */ 1784 int 1785 tcp_addoptions(struct tcpopt *to, u_char *optp) 1786 { 1787 u_int32_t mask, optlen = 0; 1788 1789 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1790 if ((to->to_flags & mask) != mask) 1791 continue; 1792 if (optlen == TCP_MAXOLEN) 1793 break; 1794 switch (to->to_flags & mask) { 1795 case TOF_MSS: 1796 while (optlen % 4) { 1797 optlen += TCPOLEN_NOP; 1798 *optp++ = TCPOPT_NOP; 1799 } 1800 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1801 continue; 1802 optlen += TCPOLEN_MAXSEG; 1803 *optp++ = TCPOPT_MAXSEG; 1804 *optp++ = TCPOLEN_MAXSEG; 1805 to->to_mss = htons(to->to_mss); 1806 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1807 optp += sizeof(to->to_mss); 1808 break; 1809 case TOF_SCALE: 1810 while (!optlen || optlen % 2 != 1) { 1811 optlen += TCPOLEN_NOP; 1812 *optp++ = TCPOPT_NOP; 1813 } 1814 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1815 continue; 1816 optlen += TCPOLEN_WINDOW; 1817 *optp++ = TCPOPT_WINDOW; 1818 *optp++ = TCPOLEN_WINDOW; 1819 *optp++ = to->to_wscale; 1820 break; 1821 case TOF_SACKPERM: 1822 while (optlen % 2) { 1823 optlen += TCPOLEN_NOP; 1824 *optp++ = TCPOPT_NOP; 1825 } 1826 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1827 continue; 1828 optlen += TCPOLEN_SACK_PERMITTED; 1829 *optp++ = TCPOPT_SACK_PERMITTED; 1830 *optp++ = TCPOLEN_SACK_PERMITTED; 1831 break; 1832 case TOF_TS: 1833 while (!optlen || optlen % 4 != 2) { 1834 optlen += TCPOLEN_NOP; 1835 *optp++ = TCPOPT_NOP; 1836 } 1837 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1838 continue; 1839 optlen += TCPOLEN_TIMESTAMP; 1840 *optp++ = TCPOPT_TIMESTAMP; 1841 *optp++ = TCPOLEN_TIMESTAMP; 1842 to->to_tsval = htonl(to->to_tsval); 1843 to->to_tsecr = htonl(to->to_tsecr); 1844 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1845 optp += sizeof(to->to_tsval); 1846 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1847 optp += sizeof(to->to_tsecr); 1848 break; 1849 case TOF_SIGNATURE: 1850 { 1851 int siglen = TCPOLEN_SIGNATURE - 2; 1852 1853 while (!optlen || optlen % 4 != 2) { 1854 optlen += TCPOLEN_NOP; 1855 *optp++ = TCPOPT_NOP; 1856 } 1857 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) { 1858 to->to_flags &= ~TOF_SIGNATURE; 1859 continue; 1860 } 1861 optlen += TCPOLEN_SIGNATURE; 1862 *optp++ = TCPOPT_SIGNATURE; 1863 *optp++ = TCPOLEN_SIGNATURE; 1864 to->to_signature = optp; 1865 while (siglen--) 1866 *optp++ = 0; 1867 break; 1868 } 1869 case TOF_SACK: 1870 { 1871 int sackblks = 0; 1872 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1873 tcp_seq sack_seq; 1874 1875 while (!optlen || optlen % 4 != 2) { 1876 optlen += TCPOLEN_NOP; 1877 *optp++ = TCPOPT_NOP; 1878 } 1879 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1880 continue; 1881 optlen += TCPOLEN_SACKHDR; 1882 *optp++ = TCPOPT_SACK; 1883 sackblks = min(to->to_nsacks, 1884 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1885 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1886 while (sackblks--) { 1887 sack_seq = htonl(sack->start); 1888 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1889 optp += sizeof(sack_seq); 1890 sack_seq = htonl(sack->end); 1891 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1892 optp += sizeof(sack_seq); 1893 optlen += TCPOLEN_SACK; 1894 sack++; 1895 } 1896 TCPSTAT_INC(tcps_sack_send_blocks); 1897 break; 1898 } 1899 case TOF_FASTOPEN: 1900 { 1901 int total_len; 1902 1903 /* XXX is there any point to aligning this option? */ 1904 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; 1905 if (TCP_MAXOLEN - optlen < total_len) { 1906 to->to_flags &= ~TOF_FASTOPEN; 1907 continue; 1908 } 1909 *optp++ = TCPOPT_FAST_OPEN; 1910 *optp++ = total_len; 1911 if (to->to_tfo_len > 0) { 1912 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); 1913 optp += to->to_tfo_len; 1914 } 1915 optlen += total_len; 1916 break; 1917 } 1918 default: 1919 panic("%s: unknown TCP option type", __func__); 1920 break; 1921 } 1922 } 1923 1924 /* Terminate and pad TCP options to a 4 byte boundary. */ 1925 if (optlen % 4) { 1926 optlen += TCPOLEN_EOL; 1927 *optp++ = TCPOPT_EOL; 1928 } 1929 /* 1930 * According to RFC 793 (STD0007): 1931 * "The content of the header beyond the End-of-Option option 1932 * must be header padding (i.e., zero)." 1933 * and later: "The padding is composed of zeros." 1934 */ 1935 while (optlen % 4) { 1936 optlen += TCPOLEN_PAD; 1937 *optp++ = TCPOPT_PAD; 1938 } 1939 1940 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1941 return (optlen); 1942 } 1943 1944 /* 1945 * This is a copy of m_copym(), taking the TSO segment size/limit 1946 * constraints into account, and advancing the sndptr as it goes. 1947 */ 1948 struct mbuf * 1949 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen, 1950 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls) 1951 { 1952 #ifdef KERN_TLS 1953 struct ktls_session *tls, *ntls; 1954 struct mbuf *start __diagused; 1955 #endif 1956 struct mbuf *n, **np; 1957 struct mbuf *top; 1958 int32_t off = off0; 1959 int32_t len = *plen; 1960 int32_t fragsize; 1961 int32_t len_cp = 0; 1962 int32_t *pkthdrlen; 1963 uint32_t mlen, frags; 1964 bool copyhdr; 1965 1966 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off)); 1967 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len)); 1968 if (off == 0 && m->m_flags & M_PKTHDR) 1969 copyhdr = true; 1970 else 1971 copyhdr = false; 1972 while (off > 0) { 1973 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain")); 1974 if (off < m->m_len) 1975 break; 1976 off -= m->m_len; 1977 if ((sb) && (m == sb->sb_sndptr)) { 1978 sb->sb_sndptroff += m->m_len; 1979 sb->sb_sndptr = m->m_next; 1980 } 1981 m = m->m_next; 1982 } 1983 np = ⊤ 1984 top = NULL; 1985 pkthdrlen = NULL; 1986 #ifdef KERN_TLS 1987 if (hw_tls && (m->m_flags & M_EXTPG)) 1988 tls = m->m_epg_tls; 1989 else 1990 tls = NULL; 1991 start = m; 1992 #endif 1993 while (len > 0) { 1994 if (m == NULL) { 1995 KASSERT(len == M_COPYALL, 1996 ("tcp_m_copym, length > size of mbuf chain")); 1997 *plen = len_cp; 1998 if (pkthdrlen != NULL) 1999 *pkthdrlen = len_cp; 2000 break; 2001 } 2002 #ifdef KERN_TLS 2003 if (hw_tls) { 2004 if (m->m_flags & M_EXTPG) 2005 ntls = m->m_epg_tls; 2006 else 2007 ntls = NULL; 2008 2009 /* 2010 * Avoid mixing TLS records with handshake 2011 * data or TLS records from different 2012 * sessions. 2013 */ 2014 if (tls != ntls) { 2015 MPASS(m != start); 2016 *plen = len_cp; 2017 if (pkthdrlen != NULL) 2018 *pkthdrlen = len_cp; 2019 break; 2020 } 2021 } 2022 #endif 2023 mlen = min(len, m->m_len - off); 2024 if (seglimit) { 2025 /* 2026 * For M_EXTPG mbufs, add 3 segments 2027 * + 1 in case we are crossing page boundaries 2028 * + 2 in case the TLS hdr/trailer are used 2029 * It is cheaper to just add the segments 2030 * than it is to take the cache miss to look 2031 * at the mbuf ext_pgs state in detail. 2032 */ 2033 if (m->m_flags & M_EXTPG) { 2034 fragsize = min(segsize, PAGE_SIZE); 2035 frags = 3; 2036 } else { 2037 fragsize = segsize; 2038 frags = 0; 2039 } 2040 2041 /* Break if we really can't fit anymore. */ 2042 if ((frags + 1) >= seglimit) { 2043 *plen = len_cp; 2044 if (pkthdrlen != NULL) 2045 *pkthdrlen = len_cp; 2046 break; 2047 } 2048 2049 /* 2050 * Reduce size if you can't copy the whole 2051 * mbuf. If we can't copy the whole mbuf, also 2052 * adjust len so the loop will end after this 2053 * mbuf. 2054 */ 2055 if ((frags + howmany(mlen, fragsize)) >= seglimit) { 2056 mlen = (seglimit - frags - 1) * fragsize; 2057 len = mlen; 2058 *plen = len_cp + len; 2059 if (pkthdrlen != NULL) 2060 *pkthdrlen = *plen; 2061 } 2062 frags += howmany(mlen, fragsize); 2063 if (frags == 0) 2064 frags++; 2065 seglimit -= frags; 2066 KASSERT(seglimit > 0, 2067 ("%s: seglimit went too low", __func__)); 2068 } 2069 if (copyhdr) 2070 n = m_gethdr(M_NOWAIT, m->m_type); 2071 else 2072 n = m_get(M_NOWAIT, m->m_type); 2073 *np = n; 2074 if (n == NULL) 2075 goto nospace; 2076 if (copyhdr) { 2077 if (!m_dup_pkthdr(n, m, M_NOWAIT)) 2078 goto nospace; 2079 if (len == M_COPYALL) 2080 n->m_pkthdr.len -= off0; 2081 else 2082 n->m_pkthdr.len = len; 2083 pkthdrlen = &n->m_pkthdr.len; 2084 copyhdr = false; 2085 } 2086 n->m_len = mlen; 2087 len_cp += n->m_len; 2088 if (m->m_flags & (M_EXT | M_EXTPG)) { 2089 n->m_data = m->m_data + off; 2090 mb_dupcl(n, m); 2091 } else 2092 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 2093 (u_int)n->m_len); 2094 2095 if (sb && (sb->sb_sndptr == m) && 2096 ((n->m_len + off) >= m->m_len) && m->m_next) { 2097 sb->sb_sndptroff += m->m_len; 2098 sb->sb_sndptr = m->m_next; 2099 } 2100 off = 0; 2101 if (len != M_COPYALL) { 2102 len -= n->m_len; 2103 } 2104 m = m->m_next; 2105 np = &n->m_next; 2106 } 2107 return (top); 2108 nospace: 2109 m_freem(top); 2110 return (NULL); 2111 } 2112 2113 void 2114 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin) 2115 { 2116 2117 /* 2118 * Automatic sizing of send socket buffer. Often the send buffer 2119 * size is not optimally adjusted to the actual network conditions 2120 * at hand (delay bandwidth product). Setting the buffer size too 2121 * small limits throughput on links with high bandwidth and high 2122 * delay (eg. trans-continental/oceanic links). Setting the 2123 * buffer size too big consumes too much real kernel memory, 2124 * especially with many connections on busy servers. 2125 * 2126 * The criteria to step up the send buffer one notch are: 2127 * 1. receive window of remote host is larger than send buffer 2128 * (with a fudge factor of 5/4th); 2129 * 2. send buffer is filled to 7/8th with data (so we actually 2130 * have data to make use of it); 2131 * 3. send buffer fill has not hit maximal automatic size; 2132 * 4. our send window (slow start and cogestion controlled) is 2133 * larger than sent but unacknowledged data in send buffer. 2134 * 2135 * The remote host receive window scaling factor may limit the 2136 * growing of the send buffer before it reaches its allowed 2137 * maximum. 2138 * 2139 * It scales directly with slow start or congestion window 2140 * and does at most one step per received ACK. This fast 2141 * scaling has the drawback of growing the send buffer beyond 2142 * what is strictly necessary to make full use of a given 2143 * delay*bandwidth product. However testing has shown this not 2144 * to be much of an problem. At worst we are trading wasting 2145 * of available bandwidth (the non-use of it) for wasting some 2146 * socket buffer memory. 2147 * 2148 * TODO: Shrink send buffer during idle periods together 2149 * with congestion window. Requires another timer. Has to 2150 * wait for upcoming tcp timer rewrite. 2151 * 2152 * XXXGL: should there be used sbused() or sbavail()? 2153 */ 2154 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 2155 int lowat; 2156 2157 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0; 2158 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat && 2159 sbused(&so->so_snd) >= 2160 (so->so_snd.sb_hiwat / 8 * 7) - lowat && 2161 sbused(&so->so_snd) < V_tcp_autosndbuf_max && 2162 sendwin >= (sbused(&so->so_snd) - 2163 (tp->snd_nxt - tp->snd_una))) { 2164 if (!sbreserve_locked(so, SO_SND, 2165 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 2166 V_tcp_autosndbuf_max), curthread)) 2167 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 2168 } 2169 } 2170 } 2171