1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 5 * The Regents of the University of California. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Neither the name of the University nor the names of its contributors 16 * may be used to endorse or promote products derived from this software 17 * without specific prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 * 31 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 32 */ 33 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD$"); 36 37 #include "opt_inet.h" 38 #include "opt_inet6.h" 39 #include "opt_ipsec.h" 40 #include "opt_kern_tls.h" 41 #include "opt_tcpdebug.h" 42 43 #include <sys/param.h> 44 #include <sys/systm.h> 45 #include <sys/arb.h> 46 #include <sys/domain.h> 47 #ifdef TCP_HHOOK 48 #include <sys/hhook.h> 49 #endif 50 #include <sys/kernel.h> 51 #ifdef KERN_TLS 52 #include <sys/ktls.h> 53 #endif 54 #include <sys/lock.h> 55 #include <sys/mbuf.h> 56 #include <sys/mutex.h> 57 #include <sys/protosw.h> 58 #include <sys/qmath.h> 59 #include <sys/sdt.h> 60 #include <sys/socket.h> 61 #include <sys/socketvar.h> 62 #include <sys/sysctl.h> 63 #include <sys/stats.h> 64 65 #include <net/if.h> 66 #include <net/route.h> 67 #include <net/route/nhop.h> 68 #include <net/vnet.h> 69 70 #include <netinet/in.h> 71 #include <netinet/in_kdtrace.h> 72 #include <netinet/in_systm.h> 73 #include <netinet/ip.h> 74 #include <netinet/in_pcb.h> 75 #include <netinet/ip_var.h> 76 #include <netinet/ip_options.h> 77 #ifdef INET6 78 #include <netinet6/in6_pcb.h> 79 #include <netinet/ip6.h> 80 #include <netinet6/ip6_var.h> 81 #endif 82 #include <netinet/tcp.h> 83 #define TCPOUTFLAGS 84 #include <netinet/tcp_fsm.h> 85 #include <netinet/tcp_log_buf.h> 86 #include <netinet/tcp_seq.h> 87 #include <netinet/tcp_var.h> 88 #include <netinet/tcp_syncache.h> 89 #include <netinet/tcp_timer.h> 90 #include <netinet/tcpip.h> 91 #include <netinet/cc/cc.h> 92 #include <netinet/tcp_fastopen.h> 93 #ifdef TCPPCAP 94 #include <netinet/tcp_pcap.h> 95 #endif 96 #ifdef TCPDEBUG 97 #include <netinet/tcp_debug.h> 98 #endif 99 #ifdef TCP_OFFLOAD 100 #include <netinet/tcp_offload.h> 101 #endif 102 #include <netinet/tcp_ecn.h> 103 104 #include <netipsec/ipsec_support.h> 105 106 #include <netinet/udp.h> 107 #include <netinet/udp_var.h> 108 #include <machine/in_cksum.h> 109 110 #include <security/mac/mac_framework.h> 111 112 VNET_DEFINE(int, path_mtu_discovery) = 1; 113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW, 114 &VNET_NAME(path_mtu_discovery), 1, 115 "Enable Path MTU Discovery"); 116 117 VNET_DEFINE(int, tcp_do_tso) = 1; 118 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW, 119 &VNET_NAME(tcp_do_tso), 0, 120 "Enable TCP Segmentation Offload"); 121 122 VNET_DEFINE(int, tcp_sendspace) = 1024*32; 123 #define V_tcp_sendspace VNET(tcp_sendspace) 124 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW, 125 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 126 127 VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 128 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, 129 &VNET_NAME(tcp_do_autosndbuf), 0, 130 "Enable automatic send buffer sizing"); 131 132 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW, 134 &VNET_NAME(tcp_autosndbuf_inc), 0, 135 "Incrementor step size of automatic send buffer"); 136 137 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; 138 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW, 139 &VNET_NAME(tcp_autosndbuf_max), 0, 140 "Max size of automatic send buffer"); 141 142 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0; 143 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat) 144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW, 145 &VNET_NAME(tcp_sendbuf_auto_lowat), 0, 146 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT"); 147 148 /* 149 * Make sure that either retransmit or persist timer is set for SYN, FIN and 150 * non-ACK. 151 */ 152 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \ 153 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\ 154 tcp_timer_active((tp), TT_REXMT) || \ 155 tcp_timer_active((tp), TT_PERSIST), \ 156 ("neither rexmt nor persist timer is set")) 157 158 static void inline cc_after_idle(struct tcpcb *tp); 159 160 #ifdef TCP_HHOOK 161 /* 162 * Wrapper for the TCP established output helper hook. 163 */ 164 void 165 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 166 struct tcpopt *to, uint32_t len, int tso) 167 { 168 struct tcp_hhook_data hhook_data; 169 170 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 171 hhook_data.tp = tp; 172 hhook_data.th = th; 173 hhook_data.to = to; 174 hhook_data.len = len; 175 hhook_data.tso = tso; 176 177 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 178 tp->osd); 179 } 180 } 181 #endif 182 183 /* 184 * CC wrapper hook functions 185 */ 186 static void inline 187 cc_after_idle(struct tcpcb *tp) 188 { 189 INP_WLOCK_ASSERT(tp->t_inpcb); 190 191 if (CC_ALGO(tp)->after_idle != NULL) 192 CC_ALGO(tp)->after_idle(tp->ccv); 193 } 194 195 /* 196 * Tcp output routine: figure out what should be sent and send it. 197 */ 198 int 199 tcp_default_output(struct tcpcb *tp) 200 { 201 struct socket *so = tp->t_inpcb->inp_socket; 202 int32_t len; 203 uint32_t recwin, sendwin; 204 uint16_t flags; 205 int off, error = 0; /* Keep compiler happy */ 206 u_int if_hw_tsomaxsegcount = 0; 207 u_int if_hw_tsomaxsegsize = 0; 208 struct mbuf *m; 209 struct ip *ip = NULL; 210 #ifdef TCPDEBUG 211 struct ipovly *ipov = NULL; 212 #endif 213 struct tcphdr *th; 214 u_char opt[TCP_MAXOLEN]; 215 unsigned ipoptlen, optlen, hdrlen, ulen; 216 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 217 unsigned ipsec_optlen = 0; 218 #endif 219 int idle, sendalot, curticks; 220 int sack_rxmit, sack_bytes_rxmt; 221 struct sackhole *p; 222 int tso, mtu; 223 struct tcpopt to; 224 struct udphdr *udp = NULL; 225 struct tcp_log_buffer *lgb; 226 unsigned int wanted_cookie = 0; 227 unsigned int dont_sendalot = 0; 228 #if 0 229 int maxburst = TCP_MAXBURST; 230 #endif 231 #ifdef INET6 232 struct ip6_hdr *ip6 = NULL; 233 int isipv6; 234 235 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 236 #endif 237 #ifdef KERN_TLS 238 const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0; 239 #else 240 const bool hw_tls = false; 241 #endif 242 243 NET_EPOCH_ASSERT(); 244 INP_WLOCK_ASSERT(tp->t_inpcb); 245 246 #ifdef TCP_OFFLOAD 247 if (tp->t_flags & TF_TOE) 248 return (tcp_offload_output(tp)); 249 #endif 250 251 /* 252 * For TFO connections in SYN_SENT or SYN_RECEIVED, 253 * only allow the initial SYN or SYN|ACK and those sent 254 * by the retransmit timer. 255 */ 256 if (IS_FASTOPEN(tp->t_flags) && 257 ((tp->t_state == TCPS_SYN_SENT) || 258 (tp->t_state == TCPS_SYN_RECEIVED)) && 259 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 260 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 261 return (0); 262 263 /* 264 * Determine length of data that should be transmitted, 265 * and flags that will be used. 266 * If there is some data or critical controls (SYN, RST) 267 * to send, then transmit; otherwise, investigate further. 268 */ 269 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 270 if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) || 271 (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur)))) 272 cc_after_idle(tp); 273 tp->t_flags &= ~TF_LASTIDLE; 274 if (idle) { 275 if (tp->t_flags & TF_MORETOCOME) { 276 tp->t_flags |= TF_LASTIDLE; 277 idle = 0; 278 } 279 } 280 again: 281 /* 282 * If we've recently taken a timeout, snd_max will be greater than 283 * snd_nxt. There may be SACK information that allows us to avoid 284 * resending already delivered data. Adjust snd_nxt accordingly. 285 */ 286 if ((tp->t_flags & TF_SACK_PERMIT) && 287 SEQ_LT(tp->snd_nxt, tp->snd_max)) 288 tcp_sack_adjust(tp); 289 sendalot = 0; 290 tso = 0; 291 mtu = 0; 292 off = tp->snd_nxt - tp->snd_una; 293 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 294 295 flags = tcp_outflags[tp->t_state]; 296 /* 297 * Send any SACK-generated retransmissions. If we're explicitly trying 298 * to send out new data (when sendalot is 1), bypass this function. 299 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 300 * we're replacing a (future) new transmission with a retransmission 301 * now, and we previously incremented snd_cwnd in tcp_input(). 302 */ 303 /* 304 * Still in sack recovery , reset rxmit flag to zero. 305 */ 306 sack_rxmit = 0; 307 sack_bytes_rxmt = 0; 308 len = 0; 309 p = NULL; 310 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 311 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 312 uint32_t cwin; 313 314 cwin = 315 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0); 316 /* Do not retransmit SACK segments beyond snd_recover */ 317 if (SEQ_GT(p->end, tp->snd_recover)) { 318 /* 319 * (At least) part of sack hole extends beyond 320 * snd_recover. Check to see if we can rexmit data 321 * for this hole. 322 */ 323 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 324 /* 325 * Can't rexmit any more data for this hole. 326 * That data will be rexmitted in the next 327 * sack recovery episode, when snd_recover 328 * moves past p->rxmit. 329 */ 330 p = NULL; 331 goto after_sack_rexmit; 332 } else { 333 /* Can rexmit part of the current hole */ 334 len = ((int32_t)ulmin(cwin, 335 SEQ_SUB(tp->snd_recover, p->rxmit))); 336 } 337 } else { 338 len = ((int32_t)ulmin(cwin, 339 SEQ_SUB(p->end, p->rxmit))); 340 } 341 if (len > 0) { 342 off = SEQ_SUB(p->rxmit, tp->snd_una); 343 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 344 __func__, off)); 345 sack_rxmit = 1; 346 sendalot = 1; 347 TCPSTAT_INC(tcps_sack_rexmits); 348 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 349 min(len, tcp_maxseg(tp))); 350 } 351 } 352 after_sack_rexmit: 353 /* 354 * Get standard flags, and add SYN or FIN if requested by 'hidden' 355 * state flags. 356 */ 357 if (tp->t_flags & TF_NEEDFIN) 358 flags |= TH_FIN; 359 if (tp->t_flags & TF_NEEDSYN) 360 flags |= TH_SYN; 361 362 SOCKBUF_LOCK(&so->so_snd); 363 /* 364 * If in persist timeout with window of 0, send 1 byte. 365 * Otherwise, if window is small but nonzero 366 * and timer expired, we will send what we can 367 * and go to transmit state. 368 */ 369 if (tp->t_flags & TF_FORCEDATA) { 370 if (sendwin == 0) { 371 /* 372 * If we still have some data to send, then 373 * clear the FIN bit. Usually this would 374 * happen below when it realizes that we 375 * aren't sending all the data. However, 376 * if we have exactly 1 byte of unsent data, 377 * then it won't clear the FIN bit below, 378 * and if we are in persist state, we wind 379 * up sending the packet without recording 380 * that we sent the FIN bit. 381 * 382 * We can't just blindly clear the FIN bit, 383 * because if we don't have any more data 384 * to send then the probe will be the FIN 385 * itself. 386 */ 387 if (off < sbused(&so->so_snd)) 388 flags &= ~TH_FIN; 389 sendwin = 1; 390 } else { 391 tcp_timer_activate(tp, TT_PERSIST, 0); 392 tp->t_rxtshift = 0; 393 } 394 } 395 396 /* 397 * If snd_nxt == snd_max and we have transmitted a FIN, the 398 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 399 * a negative length. This can also occur when TCP opens up 400 * its congestion window while receiving additional duplicate 401 * acks after fast-retransmit because TCP will reset snd_nxt 402 * to snd_max after the fast-retransmit. 403 * 404 * In the normal retransmit-FIN-only case, however, snd_nxt will 405 * be set to snd_una, the offset will be 0, and the length may 406 * wind up 0. 407 * 408 * If sack_rxmit is true we are retransmitting from the scoreboard 409 * in which case len is already set. 410 */ 411 if (sack_rxmit == 0) { 412 if (sack_bytes_rxmt == 0) 413 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) - 414 off); 415 else { 416 int32_t cwin; 417 418 /* 419 * We are inside of a SACK recovery episode and are 420 * sending new data, having retransmitted all the 421 * data possible in the scoreboard. 422 */ 423 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) - 424 off); 425 /* 426 * Don't remove this (len > 0) check ! 427 * We explicitly check for len > 0 here (although it 428 * isn't really necessary), to work around a gcc 429 * optimization issue - to force gcc to compute 430 * len above. Without this check, the computation 431 * of len is bungled by the optimizer. 432 */ 433 if (len > 0) { 434 cwin = tp->snd_cwnd - imax(0, (int32_t) 435 (tp->snd_nxt - tp->snd_recover)) - 436 sack_bytes_rxmt; 437 if (cwin < 0) 438 cwin = 0; 439 len = imin(len, cwin); 440 } 441 } 442 } 443 444 /* 445 * Lop off SYN bit if it has already been sent. However, if this 446 * is SYN-SENT state and if segment contains data and if we don't 447 * know that foreign host supports TAO, suppress sending segment. 448 */ 449 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 450 if (tp->t_state != TCPS_SYN_RECEIVED) 451 flags &= ~TH_SYN; 452 /* 453 * When sending additional segments following a TFO SYN|ACK, 454 * do not include the SYN bit. 455 */ 456 if (IS_FASTOPEN(tp->t_flags) && 457 (tp->t_state == TCPS_SYN_RECEIVED)) 458 flags &= ~TH_SYN; 459 off--, len++; 460 } 461 462 /* 463 * Be careful not to send data and/or FIN on SYN segments. 464 * This measure is needed to prevent interoperability problems 465 * with not fully conformant TCP implementations. 466 */ 467 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 468 len = 0; 469 flags &= ~TH_FIN; 470 } 471 472 /* 473 * On TFO sockets, ensure no data is sent in the following cases: 474 * 475 * - When retransmitting SYN|ACK on a passively-created socket 476 * 477 * - When retransmitting SYN on an actively created socket 478 * 479 * - When sending a zero-length cookie (cookie request) on an 480 * actively created socket 481 * 482 * - When the socket is in the CLOSED state (RST is being sent) 483 */ 484 if (IS_FASTOPEN(tp->t_flags) && 485 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 486 ((tp->t_state == TCPS_SYN_SENT) && 487 (tp->t_tfo_client_cookie_len == 0)) || 488 (flags & TH_RST))) 489 len = 0; 490 if (len <= 0) { 491 /* 492 * If FIN has been sent but not acked, 493 * but we haven't been called to retransmit, 494 * len will be < 0. Otherwise, window shrank 495 * after we sent into it. If window shrank to 0, 496 * cancel pending retransmit, pull snd_nxt back 497 * to (closed) window, and set the persist timer 498 * if it isn't already going. If the window didn't 499 * close completely, just wait for an ACK. 500 * 501 * We also do a general check here to ensure that 502 * we will set the persist timer when we have data 503 * to send, but a 0-byte window. This makes sure 504 * the persist timer is set even if the packet 505 * hits one of the "goto send" lines below. 506 */ 507 len = 0; 508 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 509 (off < (int) sbavail(&so->so_snd)) && 510 !tcp_timer_active(tp, TT_PERSIST)) { 511 tcp_timer_activate(tp, TT_REXMT, 0); 512 tp->t_rxtshift = 0; 513 tp->snd_nxt = tp->snd_una; 514 if (!tcp_timer_active(tp, TT_PERSIST)) 515 tcp_setpersist(tp); 516 } 517 } 518 519 /* len will be >= 0 after this point. */ 520 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 521 522 tcp_sndbuf_autoscale(tp, so, sendwin); 523 524 /* 525 * Decide if we can use TCP Segmentation Offloading (if supported by 526 * hardware). 527 * 528 * TSO may only be used if we are in a pure bulk sending state. The 529 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 530 * IP options prevent using TSO. With TSO the TCP header is the same 531 * (except for the sequence number) for all generated packets. This 532 * makes it impossible to transmit any options which vary per generated 533 * segment or packet. 534 * 535 * IPv4 handling has a clear separation of ip options and ip header 536 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 537 * the right thing below to provide length of just ip options and thus 538 * checking for ipoptlen is enough to decide if ip options are present. 539 */ 540 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 541 /* 542 * Pre-calculate here as we save another lookup into the darknesses 543 * of IPsec that way and can actually decide if TSO is ok. 544 */ 545 #ifdef INET6 546 if (isipv6 && IPSEC_ENABLED(ipv6)) 547 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); 548 #ifdef INET 549 else 550 #endif 551 #endif /* INET6 */ 552 #ifdef INET 553 if (IPSEC_ENABLED(ipv4)) 554 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); 555 #endif /* INET */ 556 #endif /* IPSEC */ 557 #ifdef INET6 558 if (isipv6) 559 ipoptlen = ip6_optlen(tp->t_inpcb); 560 else 561 #endif 562 if (tp->t_inpcb->inp_options) 563 ipoptlen = tp->t_inpcb->inp_options->m_len - 564 offsetof(struct ipoption, ipopt_list); 565 else 566 ipoptlen = 0; 567 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 568 ipoptlen += ipsec_optlen; 569 #endif 570 571 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 572 (tp->t_port == 0) && 573 ((tp->t_flags & TF_SIGNATURE) == 0) && 574 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 575 ipoptlen == 0 && !(flags & TH_SYN)) 576 tso = 1; 577 578 if (sack_rxmit) { 579 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd))) 580 flags &= ~TH_FIN; 581 } else { 582 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 583 sbused(&so->so_snd))) 584 flags &= ~TH_FIN; 585 } 586 587 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 588 (long)TCP_MAXWIN << tp->rcv_scale); 589 590 /* 591 * Sender silly window avoidance. We transmit under the following 592 * conditions when len is non-zero: 593 * 594 * - We have a full segment (or more with TSO) 595 * - This is the last buffer in a write()/send() and we are 596 * either idle or running NODELAY 597 * - we've timed out (e.g. persist timer) 598 * - we have more then 1/2 the maximum send window's worth of 599 * data (receiver may be limited the window size) 600 * - we need to retransmit 601 */ 602 if (len) { 603 if (len >= tp->t_maxseg) 604 goto send; 605 /* 606 * As the TCP header options are now 607 * considered when setting up the initial 608 * window, we would not send the last segment 609 * if we skip considering the option length here. 610 * Note: this may not work when tcp headers change 611 * very dynamically in the future. 612 */ 613 if ((((tp->t_flags & TF_SIGNATURE) ? 614 PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) + 615 ((tp->t_flags & TF_RCVD_TSTMP) ? 616 PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) + 617 len) >= tp->t_maxseg) 618 goto send; 619 /* 620 * NOTE! on localhost connections an 'ack' from the remote 621 * end may occur synchronously with the output and cause 622 * us to flush a buffer queued with moretocome. XXX 623 * 624 * note: the len + off check is almost certainly unnecessary. 625 */ 626 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 627 (idle || (tp->t_flags & TF_NODELAY)) && 628 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) && 629 (tp->t_flags & TF_NOPUSH) == 0) { 630 goto send; 631 } 632 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 633 goto send; 634 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 635 goto send; 636 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 637 goto send; 638 if (sack_rxmit) 639 goto send; 640 } 641 642 /* 643 * Sending of standalone window updates. 644 * 645 * Window updates are important when we close our window due to a 646 * full socket buffer and are opening it again after the application 647 * reads data from it. Once the window has opened again and the 648 * remote end starts to send again the ACK clock takes over and 649 * provides the most current window information. 650 * 651 * We must avoid the silly window syndrome whereas every read 652 * from the receive buffer, no matter how small, causes a window 653 * update to be sent. We also should avoid sending a flurry of 654 * window updates when the socket buffer had queued a lot of data 655 * and the application is doing small reads. 656 * 657 * Prevent a flurry of pointless window updates by only sending 658 * an update when we can increase the advertized window by more 659 * than 1/4th of the socket buffer capacity. When the buffer is 660 * getting full or is very small be more aggressive and send an 661 * update whenever we can increase by two mss sized segments. 662 * In all other situations the ACK's to new incoming data will 663 * carry further window increases. 664 * 665 * Don't send an independent window update if a delayed 666 * ACK is pending (it will get piggy-backed on it) or the 667 * remote side already has done a half-close and won't send 668 * more data. Skip this if the connection is in T/TCP 669 * half-open state. 670 */ 671 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 672 !(tp->t_flags & TF_DELACK) && 673 !TCPS_HAVERCVDFIN(tp->t_state)) { 674 /* 675 * "adv" is the amount we could increase the window, 676 * taking into account that we are limited by 677 * TCP_MAXWIN << tp->rcv_scale. 678 */ 679 int32_t adv; 680 int oldwin; 681 682 adv = recwin; 683 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 684 oldwin = (tp->rcv_adv - tp->rcv_nxt); 685 if (adv > oldwin) 686 adv -= oldwin; 687 else 688 adv = 0; 689 } else 690 oldwin = 0; 691 692 /* 693 * If the new window size ends up being the same as or less 694 * than the old size when it is scaled, then don't force 695 * a window update. 696 */ 697 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 698 goto dontupdate; 699 700 if (adv >= (int32_t)(2 * tp->t_maxseg) && 701 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 702 recwin <= (so->so_rcv.sb_hiwat / 8) || 703 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg || 704 adv >= TCP_MAXWIN << tp->rcv_scale)) 705 goto send; 706 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat) 707 goto send; 708 } 709 dontupdate: 710 711 /* 712 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 713 * is also a catch-all for the retransmit timer timeout case. 714 */ 715 if (tp->t_flags & TF_ACKNOW) 716 goto send; 717 if ((flags & TH_RST) || 718 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 719 goto send; 720 if (SEQ_GT(tp->snd_up, tp->snd_una)) 721 goto send; 722 /* 723 * If our state indicates that FIN should be sent 724 * and we have not yet done so, then we need to send. 725 */ 726 if (flags & TH_FIN && 727 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 728 goto send; 729 /* 730 * In SACK, it is possible for tcp_output to fail to send a segment 731 * after the retransmission timer has been turned off. Make sure 732 * that the retransmission timer is set. 733 */ 734 if ((tp->t_flags & TF_SACK_PERMIT) && 735 SEQ_GT(tp->snd_max, tp->snd_una) && 736 !tcp_timer_active(tp, TT_REXMT) && 737 !tcp_timer_active(tp, TT_PERSIST)) { 738 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp)); 739 goto just_return; 740 } 741 /* 742 * TCP window updates are not reliable, rather a polling protocol 743 * using ``persist'' packets is used to insure receipt of window 744 * updates. The three ``states'' for the output side are: 745 * idle not doing retransmits or persists 746 * persisting to move a small or zero window 747 * (re)transmitting and thereby not persisting 748 * 749 * tcp_timer_active(tp, TT_PERSIST) 750 * is true when we are in persist state. 751 * (tp->t_flags & TF_FORCEDATA) 752 * is set when we are called to send a persist packet. 753 * tcp_timer_active(tp, TT_REXMT) 754 * is set when we are retransmitting 755 * The output side is idle when both timers are zero. 756 * 757 * If send window is too small, there is data to transmit, and no 758 * retransmit or persist is pending, then go to persist state. 759 * If nothing happens soon, send when timer expires: 760 * if window is nonzero, transmit what we can, 761 * otherwise force out a byte. 762 */ 763 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && 764 !tcp_timer_active(tp, TT_PERSIST)) { 765 tp->t_rxtshift = 0; 766 tcp_setpersist(tp); 767 } 768 769 /* 770 * No reason to send a segment, just return. 771 */ 772 just_return: 773 SOCKBUF_UNLOCK(&so->so_snd); 774 return (0); 775 776 send: 777 SOCKBUF_LOCK_ASSERT(&so->so_snd); 778 if (len > 0) { 779 if (len >= tp->t_maxseg) 780 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 781 else 782 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 783 } 784 /* 785 * Before ESTABLISHED, force sending of initial options 786 * unless TCP set not to do any options. 787 * NOTE: we assume that the IP/TCP header plus TCP options 788 * always fit in a single mbuf, leaving room for a maximum 789 * link header, i.e. 790 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 791 */ 792 optlen = 0; 793 #ifdef INET6 794 if (isipv6) 795 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 796 else 797 #endif 798 hdrlen = sizeof (struct tcpiphdr); 799 800 if (flags & TH_SYN) { 801 tp->snd_nxt = tp->iss; 802 } 803 804 /* 805 * Compute options for segment. 806 * We only have to care about SYN and established connection 807 * segments. Options for SYN-ACK segments are handled in TCP 808 * syncache. 809 */ 810 to.to_flags = 0; 811 if ((tp->t_flags & TF_NOOPT) == 0) { 812 /* Maximum segment size. */ 813 if (flags & TH_SYN) { 814 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 815 if (tp->t_port) 816 to.to_mss -= V_tcp_udp_tunneling_overhead; 817 to.to_flags |= TOF_MSS; 818 819 /* 820 * On SYN or SYN|ACK transmits on TFO connections, 821 * only include the TFO option if it is not a 822 * retransmit, as the presence of the TFO option may 823 * have caused the original SYN or SYN|ACK to have 824 * been dropped by a middlebox. 825 */ 826 if (IS_FASTOPEN(tp->t_flags) && 827 (tp->t_rxtshift == 0)) { 828 if (tp->t_state == TCPS_SYN_RECEIVED) { 829 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 830 to.to_tfo_cookie = 831 (u_int8_t *)&tp->t_tfo_cookie.server; 832 to.to_flags |= TOF_FASTOPEN; 833 wanted_cookie = 1; 834 } else if (tp->t_state == TCPS_SYN_SENT) { 835 to.to_tfo_len = 836 tp->t_tfo_client_cookie_len; 837 to.to_tfo_cookie = 838 tp->t_tfo_cookie.client; 839 to.to_flags |= TOF_FASTOPEN; 840 wanted_cookie = 1; 841 /* 842 * If we wind up having more data to 843 * send with the SYN than can fit in 844 * one segment, don't send any more 845 * until the SYN|ACK comes back from 846 * the other end. 847 */ 848 dont_sendalot = 1; 849 } 850 } 851 } 852 /* Window scaling. */ 853 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 854 to.to_wscale = tp->request_r_scale; 855 to.to_flags |= TOF_SCALE; 856 } 857 /* Timestamps. */ 858 if ((tp->t_flags & TF_RCVD_TSTMP) || 859 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 860 curticks = tcp_ts_getticks(); 861 to.to_tsval = curticks + tp->ts_offset; 862 to.to_tsecr = tp->ts_recent; 863 to.to_flags |= TOF_TS; 864 if (tp->t_rxtshift == 1) 865 tp->t_badrxtwin = curticks; 866 } 867 868 /* Set receive buffer autosizing timestamp. */ 869 if (tp->rfbuf_ts == 0 && 870 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 871 tp->rfbuf_ts = tcp_ts_getticks(); 872 873 /* Selective ACK's. */ 874 if (tp->t_flags & TF_SACK_PERMIT) { 875 if (flags & TH_SYN) 876 to.to_flags |= TOF_SACKPERM; 877 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 878 tp->rcv_numsacks > 0) { 879 to.to_flags |= TOF_SACK; 880 to.to_nsacks = tp->rcv_numsacks; 881 to.to_sacks = (u_char *)tp->sackblks; 882 } 883 } 884 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 885 /* TCP-MD5 (RFC2385). */ 886 /* 887 * Check that TCP_MD5SIG is enabled in tcpcb to 888 * account the size needed to set this TCP option. 889 */ 890 if (tp->t_flags & TF_SIGNATURE) 891 to.to_flags |= TOF_SIGNATURE; 892 #endif /* TCP_SIGNATURE */ 893 894 /* Processing the options. */ 895 hdrlen += optlen = tcp_addoptions(&to, opt); 896 /* 897 * If we wanted a TFO option to be added, but it was unable 898 * to fit, ensure no data is sent. 899 */ 900 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 901 !(to.to_flags & TOF_FASTOPEN)) 902 len = 0; 903 } 904 if (tp->t_port) { 905 if (V_tcp_udp_tunneling_port == 0) { 906 /* The port was removed?? */ 907 SOCKBUF_UNLOCK(&so->so_snd); 908 return (EHOSTUNREACH); 909 } 910 hdrlen += sizeof(struct udphdr); 911 } 912 /* 913 * Adjust data length if insertion of options will 914 * bump the packet length beyond the t_maxseg length. 915 * Clear the FIN bit because we cut off the tail of 916 * the segment. 917 */ 918 if (len + optlen + ipoptlen > tp->t_maxseg) { 919 flags &= ~TH_FIN; 920 921 if (tso) { 922 u_int if_hw_tsomax; 923 u_int moff; 924 int max_len; 925 926 /* extract TSO information */ 927 if_hw_tsomax = tp->t_tsomax; 928 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 929 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 930 931 /* 932 * Limit a TSO burst to prevent it from 933 * overflowing or exceeding the maximum length 934 * allowed by the network interface: 935 */ 936 KASSERT(ipoptlen == 0, 937 ("%s: TSO can't do IP options", __func__)); 938 939 /* 940 * Check if we should limit by maximum payload 941 * length: 942 */ 943 if (if_hw_tsomax != 0) { 944 /* compute maximum TSO length */ 945 max_len = (if_hw_tsomax - hdrlen - 946 max_linkhdr); 947 if (max_len <= 0) { 948 len = 0; 949 } else if (len > max_len) { 950 sendalot = 1; 951 len = max_len; 952 } 953 } 954 955 /* 956 * Prevent the last segment from being 957 * fractional unless the send sockbuf can be 958 * emptied: 959 */ 960 max_len = (tp->t_maxseg - optlen); 961 if (((uint32_t)off + (uint32_t)len) < 962 sbavail(&so->so_snd)) { 963 moff = len % max_len; 964 if (moff != 0) { 965 len -= moff; 966 sendalot = 1; 967 } 968 } 969 970 /* 971 * In case there are too many small fragments 972 * don't use TSO: 973 */ 974 if (len <= max_len) { 975 len = max_len; 976 sendalot = 1; 977 tso = 0; 978 } 979 980 /* 981 * Send the FIN in a separate segment 982 * after the bulk sending is done. 983 * We don't trust the TSO implementations 984 * to clear the FIN flag on all but the 985 * last segment. 986 */ 987 if (tp->t_flags & TF_NEEDFIN) 988 sendalot = 1; 989 } else { 990 if (optlen + ipoptlen >= tp->t_maxseg) { 991 /* 992 * Since we don't have enough space to put 993 * the IP header chain and the TCP header in 994 * one packet as required by RFC 7112, don't 995 * send it. Also ensure that at least one 996 * byte of the payload can be put into the 997 * TCP segment. 998 */ 999 SOCKBUF_UNLOCK(&so->so_snd); 1000 error = EMSGSIZE; 1001 sack_rxmit = 0; 1002 goto out; 1003 } 1004 len = tp->t_maxseg - optlen - ipoptlen; 1005 sendalot = 1; 1006 if (dont_sendalot) 1007 sendalot = 0; 1008 } 1009 } else 1010 tso = 0; 1011 1012 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 1013 ("%s: len > IP_MAXPACKET", __func__)); 1014 1015 /*#ifdef DIAGNOSTIC*/ 1016 #ifdef INET6 1017 if (max_linkhdr + hdrlen > MCLBYTES) 1018 #else 1019 if (max_linkhdr + hdrlen > MHLEN) 1020 #endif 1021 panic("tcphdr too big"); 1022 /*#endif*/ 1023 1024 /* 1025 * This KASSERT is here to catch edge cases at a well defined place. 1026 * Before, those had triggered (random) panic conditions further down. 1027 */ 1028 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 1029 1030 /* 1031 * Grab a header mbuf, attaching a copy of data to 1032 * be transmitted, and initialize the header from 1033 * the template for sends on this connection. 1034 */ 1035 if (len) { 1036 struct mbuf *mb; 1037 struct sockbuf *msb; 1038 u_int moff; 1039 1040 if ((tp->t_flags & TF_FORCEDATA) && len == 1) { 1041 TCPSTAT_INC(tcps_sndprobe); 1042 #ifdef STATS 1043 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1044 stats_voi_update_abs_u32(tp->t_stats, 1045 VOI_TCP_RETXPB, len); 1046 else 1047 stats_voi_update_abs_u64(tp->t_stats, 1048 VOI_TCP_TXPB, len); 1049 #endif /* STATS */ 1050 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 1051 tp->t_sndrexmitpack++; 1052 TCPSTAT_INC(tcps_sndrexmitpack); 1053 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 1054 #ifdef STATS 1055 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 1056 len); 1057 #endif /* STATS */ 1058 } else { 1059 TCPSTAT_INC(tcps_sndpack); 1060 TCPSTAT_ADD(tcps_sndbyte, len); 1061 #ifdef STATS 1062 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 1063 len); 1064 #endif /* STATS */ 1065 } 1066 #ifdef INET6 1067 if (MHLEN < hdrlen + max_linkhdr) 1068 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1069 else 1070 #endif 1071 m = m_gethdr(M_NOWAIT, MT_DATA); 1072 1073 if (m == NULL) { 1074 SOCKBUF_UNLOCK(&so->so_snd); 1075 error = ENOBUFS; 1076 sack_rxmit = 0; 1077 goto out; 1078 } 1079 1080 m->m_data += max_linkhdr; 1081 m->m_len = hdrlen; 1082 1083 /* 1084 * Start the m_copy functions from the closest mbuf 1085 * to the offset in the socket buffer chain. 1086 */ 1087 mb = sbsndptr_noadv(&so->so_snd, off, &moff); 1088 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 1089 m_copydata(mb, moff, len, 1090 mtod(m, caddr_t) + hdrlen); 1091 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1092 sbsndptr_adv(&so->so_snd, mb, len); 1093 m->m_len += len; 1094 } else { 1095 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1096 msb = NULL; 1097 else 1098 msb = &so->so_snd; 1099 m->m_next = tcp_m_copym(mb, moff, 1100 &len, if_hw_tsomaxsegcount, 1101 if_hw_tsomaxsegsize, msb, hw_tls); 1102 if (len <= (tp->t_maxseg - optlen)) { 1103 /* 1104 * Must have ran out of mbufs for the copy 1105 * shorten it to no longer need tso. Lets 1106 * not put on sendalot since we are low on 1107 * mbufs. 1108 */ 1109 tso = 0; 1110 } 1111 if (m->m_next == NULL) { 1112 SOCKBUF_UNLOCK(&so->so_snd); 1113 (void) m_free(m); 1114 error = ENOBUFS; 1115 sack_rxmit = 0; 1116 goto out; 1117 } 1118 } 1119 1120 /* 1121 * If we're sending everything we've got, set PUSH. 1122 * (This will keep happy those implementations which only 1123 * give data to the user when a buffer fills or 1124 * a PUSH comes in.) 1125 */ 1126 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) && 1127 !(flags & TH_SYN)) 1128 flags |= TH_PUSH; 1129 SOCKBUF_UNLOCK(&so->so_snd); 1130 } else { 1131 SOCKBUF_UNLOCK(&so->so_snd); 1132 if (tp->t_flags & TF_ACKNOW) 1133 TCPSTAT_INC(tcps_sndacks); 1134 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1135 TCPSTAT_INC(tcps_sndctrl); 1136 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1137 TCPSTAT_INC(tcps_sndurg); 1138 else 1139 TCPSTAT_INC(tcps_sndwinup); 1140 1141 m = m_gethdr(M_NOWAIT, MT_DATA); 1142 if (m == NULL) { 1143 error = ENOBUFS; 1144 sack_rxmit = 0; 1145 goto out; 1146 } 1147 #ifdef INET6 1148 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1149 MHLEN >= hdrlen) { 1150 M_ALIGN(m, hdrlen); 1151 } else 1152 #endif 1153 m->m_data += max_linkhdr; 1154 m->m_len = hdrlen; 1155 } 1156 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1157 m->m_pkthdr.rcvif = (struct ifnet *)0; 1158 #ifdef MAC 1159 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1160 #endif 1161 #ifdef INET6 1162 if (isipv6) { 1163 ip6 = mtod(m, struct ip6_hdr *); 1164 if (tp->t_port) { 1165 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); 1166 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1167 udp->uh_dport = tp->t_port; 1168 ulen = hdrlen + len - sizeof(struct ip6_hdr); 1169 udp->uh_ulen = htons(ulen); 1170 th = (struct tcphdr *)(udp + 1); 1171 } else { 1172 th = (struct tcphdr *)(ip6 + 1); 1173 } 1174 tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip6, th); 1175 } else 1176 #endif /* INET6 */ 1177 { 1178 ip = mtod(m, struct ip *); 1179 #ifdef TCPDEBUG 1180 ipov = (struct ipovly *)ip; 1181 #endif 1182 if (tp->t_port) { 1183 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); 1184 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1185 udp->uh_dport = tp->t_port; 1186 ulen = hdrlen + len - sizeof(struct ip); 1187 udp->uh_ulen = htons(ulen); 1188 th = (struct tcphdr *)(udp + 1); 1189 } else 1190 th = (struct tcphdr *)(ip + 1); 1191 tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip, th); 1192 } 1193 1194 /* 1195 * Fill in fields, remembering maximum advertised 1196 * window for use in delaying messages about window sizes. 1197 * If resending a FIN, be sure not to use a new sequence number. 1198 */ 1199 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1200 tp->snd_nxt == tp->snd_max) 1201 tp->snd_nxt--; 1202 /* 1203 * If we are starting a connection, send ECN setup 1204 * SYN packet. If we are on a retransmit, we may 1205 * resend those bits a number of times as per 1206 * RFC 3168. 1207 */ 1208 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 1209 flags |= tcp_ecn_output_syn_sent(tp); 1210 } 1211 /* Also handle parallel SYN for ECN */ 1212 if ((TCPS_HAVERCVDSYN(tp->t_state)) && 1213 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) { 1214 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit); 1215 if ((tp->t_state == TCPS_SYN_RECEIVED) && 1216 (tp->t_flags2 & TF2_ECN_SND_ECE)) 1217 tp->t_flags2 &= ~TF2_ECN_SND_ECE; 1218 #ifdef INET6 1219 if (isipv6) { 1220 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20); 1221 ip6->ip6_flow |= htonl(ect << 20); 1222 } 1223 else 1224 #endif 1225 { 1226 ip->ip_tos &= ~IPTOS_ECN_MASK; 1227 ip->ip_tos |= ect; 1228 } 1229 } 1230 1231 /* 1232 * If we are doing retransmissions, then snd_nxt will 1233 * not reflect the first unsent octet. For ACK only 1234 * packets, we do not want the sequence number of the 1235 * retransmitted packet, we want the sequence number 1236 * of the next unsent octet. So, if there is no data 1237 * (and no SYN or FIN), use snd_max instead of snd_nxt 1238 * when filling in ti_seq. But if we are in persist 1239 * state, snd_max might reflect one byte beyond the 1240 * right edge of the window, so use snd_nxt in that 1241 * case, since we know we aren't doing a retransmission. 1242 * (retransmit and persist are mutually exclusive...) 1243 */ 1244 if (sack_rxmit == 0) { 1245 if (len || (flags & (TH_SYN|TH_FIN)) || 1246 tcp_timer_active(tp, TT_PERSIST)) 1247 th->th_seq = htonl(tp->snd_nxt); 1248 else 1249 th->th_seq = htonl(tp->snd_max); 1250 } else { 1251 th->th_seq = htonl(p->rxmit); 1252 p->rxmit += len; 1253 /* 1254 * Lost Retransmission Detection 1255 * trigger resending of a (then 1256 * still existing) hole, when 1257 * fack acks recoverypoint. 1258 */ 1259 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end)) 1260 p->rxmit = tp->snd_recover; 1261 tp->sackhint.sack_bytes_rexmit += len; 1262 } 1263 if (IN_RECOVERY(tp->t_flags)) { 1264 /* 1265 * Account all bytes transmitted while 1266 * IN_RECOVERY, simplifying PRR and 1267 * Lost Retransmit Detection 1268 */ 1269 tp->sackhint.prr_out += len; 1270 } 1271 th->th_ack = htonl(tp->rcv_nxt); 1272 if (optlen) { 1273 bcopy(opt, th + 1, optlen); 1274 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1275 } 1276 tcp_set_flags(th, flags); 1277 /* 1278 * Calculate receive window. Don't shrink window, 1279 * but avoid silly window syndrome. 1280 * If a RST segment is sent, advertise a window of zero. 1281 */ 1282 if (flags & TH_RST) { 1283 recwin = 0; 1284 } else { 1285 if (recwin < (so->so_rcv.sb_hiwat / 4) && 1286 recwin < tp->t_maxseg) 1287 recwin = 0; 1288 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1289 recwin < (tp->rcv_adv - tp->rcv_nxt)) 1290 recwin = (tp->rcv_adv - tp->rcv_nxt); 1291 } 1292 /* 1293 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1294 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1295 * case is handled in syncache. 1296 */ 1297 if (flags & TH_SYN) 1298 th->th_win = htons((u_short) 1299 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1300 else { 1301 /* Avoid shrinking window with window scaling. */ 1302 recwin = roundup2(recwin, 1 << tp->rcv_scale); 1303 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1304 } 1305 1306 /* 1307 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1308 * a 0 window. This may cause the remote transmitter to stall. This 1309 * flag tells soreceive() to disable delayed acknowledgements when 1310 * draining the buffer. This can occur if the receiver is attempting 1311 * to read more data than can be buffered prior to transmitting on 1312 * the connection. 1313 */ 1314 if (th->th_win == 0) { 1315 tp->t_sndzerowin++; 1316 tp->t_flags |= TF_RXWIN0SENT; 1317 } else 1318 tp->t_flags &= ~TF_RXWIN0SENT; 1319 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1320 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1321 th->th_flags |= TH_URG; 1322 } else 1323 /* 1324 * If no urgent pointer to send, then we pull 1325 * the urgent pointer to the left edge of the send window 1326 * so that it doesn't drift into the send window on sequence 1327 * number wraparound. 1328 */ 1329 tp->snd_up = tp->snd_una; /* drag it along */ 1330 1331 /* 1332 * Put TCP length in extended header, and then 1333 * checksum extended header and data. 1334 */ 1335 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1336 1337 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1338 if (to.to_flags & TOF_SIGNATURE) { 1339 /* 1340 * Calculate MD5 signature and put it into the place 1341 * determined before. 1342 * NOTE: since TCP options buffer doesn't point into 1343 * mbuf's data, calculate offset and use it. 1344 */ 1345 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th, 1346 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) { 1347 /* 1348 * Do not send segment if the calculation of MD5 1349 * digest has failed. 1350 */ 1351 m_freem(m); 1352 goto out; 1353 } 1354 } 1355 #endif 1356 #ifdef INET6 1357 if (isipv6) { 1358 /* 1359 * There is no need to fill in ip6_plen right now. 1360 * It will be filled later by ip6_output. 1361 */ 1362 if (tp->t_port) { 1363 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 1364 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 1365 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 1366 th->th_sum = htons(0); 1367 UDPSTAT_INC(udps_opackets); 1368 } else { 1369 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1370 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1371 th->th_sum = in6_cksum_pseudo(ip6, 1372 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 1373 0); 1374 } 1375 } 1376 #endif 1377 #if defined(INET6) && defined(INET) 1378 else 1379 #endif 1380 #ifdef INET 1381 { 1382 if (tp->t_port) { 1383 m->m_pkthdr.csum_flags = CSUM_UDP; 1384 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 1385 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 1386 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 1387 th->th_sum = htons(0); 1388 UDPSTAT_INC(udps_opackets); 1389 } else { 1390 m->m_pkthdr.csum_flags = CSUM_TCP; 1391 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1392 th->th_sum = in_pseudo(ip->ip_src.s_addr, 1393 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 1394 IPPROTO_TCP + len + optlen)); 1395 } 1396 1397 /* IP version must be set here for ipv4/ipv6 checking later */ 1398 KASSERT(ip->ip_v == IPVERSION, 1399 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1400 } 1401 #endif 1402 1403 /* 1404 * Enable TSO and specify the size of the segments. 1405 * The TCP pseudo header checksum is always provided. 1406 */ 1407 if (tso) { 1408 KASSERT(len > tp->t_maxseg - optlen, 1409 ("%s: len <= tso_segsz", __func__)); 1410 m->m_pkthdr.csum_flags |= CSUM_TSO; 1411 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 1412 } 1413 1414 KASSERT(len + hdrlen == m_length(m, NULL), 1415 ("%s: mbuf chain shorter than expected: %d + %u != %u", 1416 __func__, len, hdrlen, m_length(m, NULL))); 1417 1418 #ifdef TCP_HHOOK 1419 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1420 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1421 #endif 1422 1423 #ifdef TCPDEBUG 1424 /* 1425 * Trace. 1426 */ 1427 if (so->so_options & SO_DEBUG) { 1428 u_short save = 0; 1429 #ifdef INET6 1430 if (!isipv6) 1431 #endif 1432 { 1433 save = ipov->ih_len; 1434 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1435 } 1436 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1437 #ifdef INET6 1438 if (!isipv6) 1439 #endif 1440 ipov->ih_len = save; 1441 } 1442 #endif /* TCPDEBUG */ 1443 TCP_PROBE3(debug__output, tp, th, m); 1444 1445 /* We're getting ready to send; log now. */ 1446 /* XXXMT: We are not honoring verbose logging. */ 1447 if (tp->t_logstate != TCP_LOG_STATE_OFF) 1448 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, 1449 TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0, 1450 NULL); 1451 else 1452 lgb = NULL; 1453 1454 /* 1455 * Fill in IP length and desired time to live and 1456 * send to IP level. There should be a better way 1457 * to handle ttl and tos; we could keep them in 1458 * the template, but need a way to checksum without them. 1459 */ 1460 /* 1461 * m->m_pkthdr.len should have been set before checksum calculation, 1462 * because in6_cksum() need it. 1463 */ 1464 #ifdef INET6 1465 if (isipv6) { 1466 /* 1467 * we separately set hoplimit for every segment, since the 1468 * user might want to change the value via setsockopt. 1469 * Also, desired default hop limit might be changed via 1470 * Neighbor Discovery. 1471 */ 1472 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1473 1474 /* 1475 * Set the packet size here for the benefit of DTrace probes. 1476 * ip6_output() will set it properly; it's supposed to include 1477 * the option header lengths as well. 1478 */ 1479 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1480 1481 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 1482 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1483 else 1484 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1485 1486 if (tp->t_state == TCPS_SYN_SENT) 1487 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1488 1489 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1490 1491 #ifdef TCPPCAP 1492 /* Save packet, if requested. */ 1493 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1494 #endif 1495 1496 /* TODO: IPv6 IP6TOS_ECT bit on */ 1497 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, 1498 &tp->t_inpcb->inp_route6, 1499 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1500 NULL, NULL, tp->t_inpcb); 1501 1502 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL) 1503 mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu; 1504 } 1505 #endif /* INET6 */ 1506 #if defined(INET) && defined(INET6) 1507 else 1508 #endif 1509 #ifdef INET 1510 { 1511 ip->ip_len = htons(m->m_pkthdr.len); 1512 #ifdef INET6 1513 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1514 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1515 #endif /* INET6 */ 1516 /* 1517 * If we do path MTU discovery, then we set DF on every packet. 1518 * This might not be the best thing to do according to RFC3390 1519 * Section 2. However the tcp hostcache migitates the problem 1520 * so it affects only the first tcp connection with a host. 1521 * 1522 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1523 */ 1524 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 1525 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1526 if (tp->t_port == 0 || len < V_tcp_minmss) { 1527 ip->ip_off |= htons(IP_DF); 1528 } 1529 } else { 1530 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1531 } 1532 1533 if (tp->t_state == TCPS_SYN_SENT) 1534 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1535 1536 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1537 1538 #ifdef TCPPCAP 1539 /* Save packet, if requested. */ 1540 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1541 #endif 1542 1543 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route, 1544 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1545 tp->t_inpcb); 1546 1547 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL) 1548 mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu; 1549 } 1550 #endif /* INET */ 1551 1552 if (lgb != NULL) { 1553 lgb->tlb_errno = error; 1554 lgb = NULL; 1555 } 1556 out: 1557 if (error == 0) 1558 tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls); 1559 /* 1560 * In transmit state, time the transmission and arrange for 1561 * the retransmit. In persist state, just set snd_max. 1562 */ 1563 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1564 !tcp_timer_active(tp, TT_PERSIST)) { 1565 tcp_seq startseq = tp->snd_nxt; 1566 1567 /* 1568 * Advance snd_nxt over sequence space of this segment. 1569 */ 1570 if (flags & (TH_SYN|TH_FIN)) { 1571 if (flags & TH_SYN) 1572 tp->snd_nxt++; 1573 if (flags & TH_FIN) { 1574 tp->snd_nxt++; 1575 tp->t_flags |= TF_SENTFIN; 1576 } 1577 } 1578 if (sack_rxmit) 1579 goto timer; 1580 tp->snd_nxt += len; 1581 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1582 /* 1583 * Update "made progress" indication if we just 1584 * added new data to an empty socket buffer. 1585 */ 1586 if (tp->snd_una == tp->snd_max) 1587 tp->t_acktime = ticks; 1588 tp->snd_max = tp->snd_nxt; 1589 /* 1590 * Time this transmission if not a retransmission and 1591 * not currently timing anything. 1592 */ 1593 tp->t_sndtime = ticks; 1594 if (tp->t_rtttime == 0) { 1595 tp->t_rtttime = ticks; 1596 tp->t_rtseq = startseq; 1597 TCPSTAT_INC(tcps_segstimed); 1598 } 1599 #ifdef STATS 1600 if (!(tp->t_flags & TF_GPUTINPROG) && len) { 1601 tp->t_flags |= TF_GPUTINPROG; 1602 tp->gput_seq = startseq; 1603 tp->gput_ack = startseq + 1604 ulmin(sbavail(&so->so_snd) - off, sendwin); 1605 tp->gput_ts = tcp_ts_getticks(); 1606 } 1607 #endif /* STATS */ 1608 } 1609 1610 /* 1611 * Set retransmit timer if not currently set, 1612 * and not doing a pure ack or a keep-alive probe. 1613 * Initial value for retransmit timer is smoothed 1614 * round-trip time + 2 * round-trip time variance. 1615 * Initialize shift counter which is used for backoff 1616 * of retransmit time. 1617 */ 1618 timer: 1619 if (!tcp_timer_active(tp, TT_REXMT) && 1620 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1621 (tp->snd_nxt != tp->snd_una))) { 1622 if (tcp_timer_active(tp, TT_PERSIST)) { 1623 tcp_timer_activate(tp, TT_PERSIST, 0); 1624 tp->t_rxtshift = 0; 1625 } 1626 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp)); 1627 } else if (len == 0 && sbavail(&so->so_snd) && 1628 !tcp_timer_active(tp, TT_REXMT) && 1629 !tcp_timer_active(tp, TT_PERSIST)) { 1630 /* 1631 * Avoid a situation where we do not set persist timer 1632 * after a zero window condition. For example: 1633 * 1) A -> B: packet with enough data to fill the window 1634 * 2) B -> A: ACK for #1 + new data (0 window 1635 * advertisement) 1636 * 3) A -> B: ACK for #2, 0 len packet 1637 * 1638 * In this case, A will not activate the persist timer, 1639 * because it chose to send a packet. Unless tcp_output 1640 * is called for some other reason (delayed ack timer, 1641 * another input packet from B, socket syscall), A will 1642 * not send zero window probes. 1643 * 1644 * So, if you send a 0-length packet, but there is data 1645 * in the socket buffer, and neither the rexmt or 1646 * persist timer is already set, then activate the 1647 * persist timer. 1648 */ 1649 tp->t_rxtshift = 0; 1650 tcp_setpersist(tp); 1651 } 1652 } else { 1653 /* 1654 * Persist case, update snd_max but since we are in 1655 * persist mode (no window) we do not update snd_nxt. 1656 */ 1657 int xlen = len; 1658 if (flags & TH_SYN) 1659 ++xlen; 1660 if (flags & TH_FIN) { 1661 ++xlen; 1662 tp->t_flags |= TF_SENTFIN; 1663 } 1664 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1665 tp->snd_max = tp->snd_nxt + xlen; 1666 } 1667 if ((error == 0) && 1668 (TCPS_HAVEESTABLISHED(tp->t_state) && 1669 (tp->t_flags & TF_SACK_PERMIT) && 1670 tp->rcv_numsacks > 0)) { 1671 /* Clean up any DSACK's sent */ 1672 tcp_clean_dsack_blocks(tp); 1673 } 1674 if (error) { 1675 /* 1676 * We know that the packet was lost, so back out the 1677 * sequence number advance, if any. 1678 * 1679 * If the error is EPERM the packet got blocked by the 1680 * local firewall. Normally we should terminate the 1681 * connection but the blocking may have been spurious 1682 * due to a firewall reconfiguration cycle. So we treat 1683 * it like a packet loss and let the retransmit timer and 1684 * timeouts do their work over time. 1685 * XXX: It is a POLA question whether calling tcp_drop right 1686 * away would be the really correct behavior instead. 1687 */ 1688 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1689 !tcp_timer_active(tp, TT_PERSIST)) && 1690 ((flags & TH_SYN) == 0) && 1691 (error != EPERM)) { 1692 if (sack_rxmit) { 1693 p->rxmit -= len; 1694 tp->sackhint.sack_bytes_rexmit -= len; 1695 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1696 ("sackhint bytes rtx >= 0")); 1697 KASSERT((flags & TH_FIN) == 0, 1698 ("error while FIN with SACK rxmit")); 1699 } else { 1700 tp->snd_nxt -= len; 1701 if (flags & TH_FIN) 1702 tp->snd_nxt--; 1703 } 1704 } 1705 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1706 switch (error) { 1707 case EACCES: 1708 case EPERM: 1709 tp->t_softerror = error; 1710 return (error); 1711 case ENOBUFS: 1712 TCP_XMIT_TIMER_ASSERT(tp, len, flags); 1713 tp->snd_cwnd = tp->t_maxseg; 1714 return (0); 1715 case EMSGSIZE: 1716 /* 1717 * For some reason the interface we used initially 1718 * to send segments changed to another or lowered 1719 * its MTU. 1720 * If TSO was active we either got an interface 1721 * without TSO capabilits or TSO was turned off. 1722 * If we obtained mtu from ip_output() then update 1723 * it and try again. 1724 */ 1725 if (tso) 1726 tp->t_flags &= ~TF_TSO; 1727 if (mtu != 0) { 1728 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1729 goto again; 1730 } 1731 return (error); 1732 case EHOSTDOWN: 1733 case EHOSTUNREACH: 1734 case ENETDOWN: 1735 case ENETUNREACH: 1736 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1737 tp->t_softerror = error; 1738 return (0); 1739 } 1740 /* FALLTHROUGH */ 1741 default: 1742 return (error); 1743 } 1744 } 1745 TCPSTAT_INC(tcps_sndtotal); 1746 1747 /* 1748 * Data sent (as far as we can tell). 1749 * If this advertises a larger window than any other segment, 1750 * then remember the size of the advertised window. 1751 * Any pending ACK has now been sent. 1752 */ 1753 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1754 tp->rcv_adv = tp->rcv_nxt + recwin; 1755 tp->last_ack_sent = tp->rcv_nxt; 1756 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1757 if (tcp_timer_active(tp, TT_DELACK)) 1758 tcp_timer_activate(tp, TT_DELACK, 0); 1759 #if 0 1760 /* 1761 * This completely breaks TCP if newreno is turned on. What happens 1762 * is that if delayed-acks are turned on on the receiver, this code 1763 * on the transmitter effectively destroys the TCP window, forcing 1764 * it to four packets (1.5Kx4 = 6K window). 1765 */ 1766 if (sendalot && --maxburst) 1767 goto again; 1768 #endif 1769 if (sendalot) 1770 goto again; 1771 return (0); 1772 } 1773 1774 void 1775 tcp_setpersist(struct tcpcb *tp) 1776 { 1777 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1778 int tt; 1779 int maxunacktime; 1780 1781 tp->t_flags &= ~TF_PREVVALID; 1782 if (tcp_timer_active(tp, TT_REXMT)) 1783 panic("tcp_setpersist: retransmit pending"); 1784 /* 1785 * If the state is already closed, don't bother. 1786 */ 1787 if (tp->t_state == TCPS_CLOSED) 1788 return; 1789 1790 /* 1791 * Start/restart persistence timer. 1792 */ 1793 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1794 tcp_persmin, tcp_persmax); 1795 if (TP_MAXUNACKTIME(tp) && tp->t_acktime) { 1796 maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks; 1797 if (maxunacktime < 1) 1798 maxunacktime = 1; 1799 if (maxunacktime < tt) 1800 tt = maxunacktime; 1801 } 1802 tcp_timer_activate(tp, TT_PERSIST, tt); 1803 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1804 tp->t_rxtshift++; 1805 } 1806 1807 /* 1808 * Insert TCP options according to the supplied parameters to the place 1809 * optp in a consistent way. Can handle unaligned destinations. 1810 * 1811 * The order of the option processing is crucial for optimal packing and 1812 * alignment for the scarce option space. 1813 * 1814 * The optimal order for a SYN/SYN-ACK segment is: 1815 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1816 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1817 * 1818 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1819 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1820 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1821 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1822 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1823 */ 1824 int 1825 tcp_addoptions(struct tcpopt *to, u_char *optp) 1826 { 1827 u_int32_t mask, optlen = 0; 1828 1829 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1830 if ((to->to_flags & mask) != mask) 1831 continue; 1832 if (optlen == TCP_MAXOLEN) 1833 break; 1834 switch (to->to_flags & mask) { 1835 case TOF_MSS: 1836 while (optlen % 4) { 1837 optlen += TCPOLEN_NOP; 1838 *optp++ = TCPOPT_NOP; 1839 } 1840 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1841 continue; 1842 optlen += TCPOLEN_MAXSEG; 1843 *optp++ = TCPOPT_MAXSEG; 1844 *optp++ = TCPOLEN_MAXSEG; 1845 to->to_mss = htons(to->to_mss); 1846 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1847 optp += sizeof(to->to_mss); 1848 break; 1849 case TOF_SCALE: 1850 while (!optlen || optlen % 2 != 1) { 1851 optlen += TCPOLEN_NOP; 1852 *optp++ = TCPOPT_NOP; 1853 } 1854 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1855 continue; 1856 optlen += TCPOLEN_WINDOW; 1857 *optp++ = TCPOPT_WINDOW; 1858 *optp++ = TCPOLEN_WINDOW; 1859 *optp++ = to->to_wscale; 1860 break; 1861 case TOF_SACKPERM: 1862 while (optlen % 2) { 1863 optlen += TCPOLEN_NOP; 1864 *optp++ = TCPOPT_NOP; 1865 } 1866 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1867 continue; 1868 optlen += TCPOLEN_SACK_PERMITTED; 1869 *optp++ = TCPOPT_SACK_PERMITTED; 1870 *optp++ = TCPOLEN_SACK_PERMITTED; 1871 break; 1872 case TOF_TS: 1873 while (!optlen || optlen % 4 != 2) { 1874 optlen += TCPOLEN_NOP; 1875 *optp++ = TCPOPT_NOP; 1876 } 1877 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1878 continue; 1879 optlen += TCPOLEN_TIMESTAMP; 1880 *optp++ = TCPOPT_TIMESTAMP; 1881 *optp++ = TCPOLEN_TIMESTAMP; 1882 to->to_tsval = htonl(to->to_tsval); 1883 to->to_tsecr = htonl(to->to_tsecr); 1884 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1885 optp += sizeof(to->to_tsval); 1886 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1887 optp += sizeof(to->to_tsecr); 1888 break; 1889 case TOF_SIGNATURE: 1890 { 1891 int siglen = TCPOLEN_SIGNATURE - 2; 1892 1893 while (!optlen || optlen % 4 != 2) { 1894 optlen += TCPOLEN_NOP; 1895 *optp++ = TCPOPT_NOP; 1896 } 1897 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) { 1898 to->to_flags &= ~TOF_SIGNATURE; 1899 continue; 1900 } 1901 optlen += TCPOLEN_SIGNATURE; 1902 *optp++ = TCPOPT_SIGNATURE; 1903 *optp++ = TCPOLEN_SIGNATURE; 1904 to->to_signature = optp; 1905 while (siglen--) 1906 *optp++ = 0; 1907 break; 1908 } 1909 case TOF_SACK: 1910 { 1911 int sackblks = 0; 1912 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1913 tcp_seq sack_seq; 1914 1915 while (!optlen || optlen % 4 != 2) { 1916 optlen += TCPOLEN_NOP; 1917 *optp++ = TCPOPT_NOP; 1918 } 1919 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1920 continue; 1921 optlen += TCPOLEN_SACKHDR; 1922 *optp++ = TCPOPT_SACK; 1923 sackblks = min(to->to_nsacks, 1924 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1925 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1926 while (sackblks--) { 1927 sack_seq = htonl(sack->start); 1928 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1929 optp += sizeof(sack_seq); 1930 sack_seq = htonl(sack->end); 1931 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1932 optp += sizeof(sack_seq); 1933 optlen += TCPOLEN_SACK; 1934 sack++; 1935 } 1936 TCPSTAT_INC(tcps_sack_send_blocks); 1937 break; 1938 } 1939 case TOF_FASTOPEN: 1940 { 1941 int total_len; 1942 1943 /* XXX is there any point to aligning this option? */ 1944 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; 1945 if (TCP_MAXOLEN - optlen < total_len) { 1946 to->to_flags &= ~TOF_FASTOPEN; 1947 continue; 1948 } 1949 *optp++ = TCPOPT_FAST_OPEN; 1950 *optp++ = total_len; 1951 if (to->to_tfo_len > 0) { 1952 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); 1953 optp += to->to_tfo_len; 1954 } 1955 optlen += total_len; 1956 break; 1957 } 1958 default: 1959 panic("%s: unknown TCP option type", __func__); 1960 break; 1961 } 1962 } 1963 1964 /* Terminate and pad TCP options to a 4 byte boundary. */ 1965 if (optlen % 4) { 1966 optlen += TCPOLEN_EOL; 1967 *optp++ = TCPOPT_EOL; 1968 } 1969 /* 1970 * According to RFC 793 (STD0007): 1971 * "The content of the header beyond the End-of-Option option 1972 * must be header padding (i.e., zero)." 1973 * and later: "The padding is composed of zeros." 1974 */ 1975 while (optlen % 4) { 1976 optlen += TCPOLEN_PAD; 1977 *optp++ = TCPOPT_PAD; 1978 } 1979 1980 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1981 return (optlen); 1982 } 1983 1984 /* 1985 * This is a copy of m_copym(), taking the TSO segment size/limit 1986 * constraints into account, and advancing the sndptr as it goes. 1987 */ 1988 struct mbuf * 1989 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen, 1990 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls) 1991 { 1992 #ifdef KERN_TLS 1993 struct ktls_session *tls, *ntls; 1994 struct mbuf *start __diagused; 1995 #endif 1996 struct mbuf *n, **np; 1997 struct mbuf *top; 1998 int32_t off = off0; 1999 int32_t len = *plen; 2000 int32_t fragsize; 2001 int32_t len_cp = 0; 2002 int32_t *pkthdrlen; 2003 uint32_t mlen, frags; 2004 bool copyhdr; 2005 2006 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off)); 2007 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len)); 2008 if (off == 0 && m->m_flags & M_PKTHDR) 2009 copyhdr = true; 2010 else 2011 copyhdr = false; 2012 while (off > 0) { 2013 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain")); 2014 if (off < m->m_len) 2015 break; 2016 off -= m->m_len; 2017 if ((sb) && (m == sb->sb_sndptr)) { 2018 sb->sb_sndptroff += m->m_len; 2019 sb->sb_sndptr = m->m_next; 2020 } 2021 m = m->m_next; 2022 } 2023 np = ⊤ 2024 top = NULL; 2025 pkthdrlen = NULL; 2026 #ifdef KERN_TLS 2027 if (hw_tls && (m->m_flags & M_EXTPG)) 2028 tls = m->m_epg_tls; 2029 else 2030 tls = NULL; 2031 start = m; 2032 #endif 2033 while (len > 0) { 2034 if (m == NULL) { 2035 KASSERT(len == M_COPYALL, 2036 ("tcp_m_copym, length > size of mbuf chain")); 2037 *plen = len_cp; 2038 if (pkthdrlen != NULL) 2039 *pkthdrlen = len_cp; 2040 break; 2041 } 2042 #ifdef KERN_TLS 2043 if (hw_tls) { 2044 if (m->m_flags & M_EXTPG) 2045 ntls = m->m_epg_tls; 2046 else 2047 ntls = NULL; 2048 2049 /* 2050 * Avoid mixing TLS records with handshake 2051 * data or TLS records from different 2052 * sessions. 2053 */ 2054 if (tls != ntls) { 2055 MPASS(m != start); 2056 *plen = len_cp; 2057 if (pkthdrlen != NULL) 2058 *pkthdrlen = len_cp; 2059 break; 2060 } 2061 } 2062 #endif 2063 mlen = min(len, m->m_len - off); 2064 if (seglimit) { 2065 /* 2066 * For M_EXTPG mbufs, add 3 segments 2067 * + 1 in case we are crossing page boundaries 2068 * + 2 in case the TLS hdr/trailer are used 2069 * It is cheaper to just add the segments 2070 * than it is to take the cache miss to look 2071 * at the mbuf ext_pgs state in detail. 2072 */ 2073 if (m->m_flags & M_EXTPG) { 2074 fragsize = min(segsize, PAGE_SIZE); 2075 frags = 3; 2076 } else { 2077 fragsize = segsize; 2078 frags = 0; 2079 } 2080 2081 /* Break if we really can't fit anymore. */ 2082 if ((frags + 1) >= seglimit) { 2083 *plen = len_cp; 2084 if (pkthdrlen != NULL) 2085 *pkthdrlen = len_cp; 2086 break; 2087 } 2088 2089 /* 2090 * Reduce size if you can't copy the whole 2091 * mbuf. If we can't copy the whole mbuf, also 2092 * adjust len so the loop will end after this 2093 * mbuf. 2094 */ 2095 if ((frags + howmany(mlen, fragsize)) >= seglimit) { 2096 mlen = (seglimit - frags - 1) * fragsize; 2097 len = mlen; 2098 *plen = len_cp + len; 2099 if (pkthdrlen != NULL) 2100 *pkthdrlen = *plen; 2101 } 2102 frags += howmany(mlen, fragsize); 2103 if (frags == 0) 2104 frags++; 2105 seglimit -= frags; 2106 KASSERT(seglimit > 0, 2107 ("%s: seglimit went too low", __func__)); 2108 } 2109 if (copyhdr) 2110 n = m_gethdr(M_NOWAIT, m->m_type); 2111 else 2112 n = m_get(M_NOWAIT, m->m_type); 2113 *np = n; 2114 if (n == NULL) 2115 goto nospace; 2116 if (copyhdr) { 2117 if (!m_dup_pkthdr(n, m, M_NOWAIT)) 2118 goto nospace; 2119 if (len == M_COPYALL) 2120 n->m_pkthdr.len -= off0; 2121 else 2122 n->m_pkthdr.len = len; 2123 pkthdrlen = &n->m_pkthdr.len; 2124 copyhdr = false; 2125 } 2126 n->m_len = mlen; 2127 len_cp += n->m_len; 2128 if (m->m_flags & (M_EXT|M_EXTPG)) { 2129 n->m_data = m->m_data + off; 2130 mb_dupcl(n, m); 2131 } else 2132 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 2133 (u_int)n->m_len); 2134 2135 if (sb && (sb->sb_sndptr == m) && 2136 ((n->m_len + off) >= m->m_len) && m->m_next) { 2137 sb->sb_sndptroff += m->m_len; 2138 sb->sb_sndptr = m->m_next; 2139 } 2140 off = 0; 2141 if (len != M_COPYALL) { 2142 len -= n->m_len; 2143 } 2144 m = m->m_next; 2145 np = &n->m_next; 2146 } 2147 return (top); 2148 nospace: 2149 m_freem(top); 2150 return (NULL); 2151 } 2152 2153 void 2154 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin) 2155 { 2156 2157 /* 2158 * Automatic sizing of send socket buffer. Often the send buffer 2159 * size is not optimally adjusted to the actual network conditions 2160 * at hand (delay bandwidth product). Setting the buffer size too 2161 * small limits throughput on links with high bandwidth and high 2162 * delay (eg. trans-continental/oceanic links). Setting the 2163 * buffer size too big consumes too much real kernel memory, 2164 * especially with many connections on busy servers. 2165 * 2166 * The criteria to step up the send buffer one notch are: 2167 * 1. receive window of remote host is larger than send buffer 2168 * (with a fudge factor of 5/4th); 2169 * 2. send buffer is filled to 7/8th with data (so we actually 2170 * have data to make use of it); 2171 * 3. send buffer fill has not hit maximal automatic size; 2172 * 4. our send window (slow start and cogestion controlled) is 2173 * larger than sent but unacknowledged data in send buffer. 2174 * 2175 * The remote host receive window scaling factor may limit the 2176 * growing of the send buffer before it reaches its allowed 2177 * maximum. 2178 * 2179 * It scales directly with slow start or congestion window 2180 * and does at most one step per received ACK. This fast 2181 * scaling has the drawback of growing the send buffer beyond 2182 * what is strictly necessary to make full use of a given 2183 * delay*bandwidth product. However testing has shown this not 2184 * to be much of an problem. At worst we are trading wasting 2185 * of available bandwidth (the non-use of it) for wasting some 2186 * socket buffer memory. 2187 * 2188 * TODO: Shrink send buffer during idle periods together 2189 * with congestion window. Requires another timer. Has to 2190 * wait for upcoming tcp timer rewrite. 2191 * 2192 * XXXGL: should there be used sbused() or sbavail()? 2193 */ 2194 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 2195 int lowat; 2196 2197 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0; 2198 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat && 2199 sbused(&so->so_snd) >= 2200 (so->so_snd.sb_hiwat / 8 * 7) - lowat && 2201 sbused(&so->so_snd) < V_tcp_autosndbuf_max && 2202 sendwin >= (sbused(&so->so_snd) - 2203 (tp->snd_nxt - tp->snd_una))) { 2204 if (!sbreserve_locked(so, SO_SND, 2205 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 2206 V_tcp_autosndbuf_max), curthread)) 2207 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 2208 } 2209 } 2210 } 2211