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/domain.h> 46 #ifdef TCP_HHOOK 47 #include <sys/hhook.h> 48 #endif 49 #include <sys/kernel.h> 50 #ifdef KERN_TLS 51 #include <sys/ktls.h> 52 #endif 53 #include <sys/lock.h> 54 #include <sys/mbuf.h> 55 #include <sys/mutex.h> 56 #include <sys/protosw.h> 57 #include <sys/sdt.h> 58 #include <sys/socket.h> 59 #include <sys/socketvar.h> 60 #include <sys/sysctl.h> 61 62 #include <net/if.h> 63 #include <net/route.h> 64 #include <net/vnet.h> 65 66 #include <netinet/in.h> 67 #include <netinet/in_kdtrace.h> 68 #include <netinet/in_systm.h> 69 #include <netinet/ip.h> 70 #include <netinet/in_pcb.h> 71 #include <netinet/ip_var.h> 72 #include <netinet/ip_options.h> 73 #ifdef INET6 74 #include <netinet6/in6_pcb.h> 75 #include <netinet/ip6.h> 76 #include <netinet6/ip6_var.h> 77 #endif 78 #include <netinet/tcp.h> 79 #define TCPOUTFLAGS 80 #include <netinet/tcp_fsm.h> 81 #include <netinet/tcp_log_buf.h> 82 #include <netinet/tcp_seq.h> 83 #include <netinet/tcp_timer.h> 84 #include <netinet/tcp_var.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 TCPDEBUG 92 #include <netinet/tcp_debug.h> 93 #endif 94 #ifdef TCP_OFFLOAD 95 #include <netinet/tcp_offload.h> 96 #endif 97 98 #include <netipsec/ipsec_support.h> 99 100 #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 static void inline cc_after_idle(struct tcpcb *tp); 151 152 #ifdef TCP_HHOOK 153 /* 154 * Wrapper for the TCP established output helper hook. 155 */ 156 void 157 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 158 struct tcpopt *to, uint32_t len, int tso) 159 { 160 struct tcp_hhook_data hhook_data; 161 162 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 163 hhook_data.tp = tp; 164 hhook_data.th = th; 165 hhook_data.to = to; 166 hhook_data.len = len; 167 hhook_data.tso = tso; 168 169 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 170 tp->osd); 171 } 172 } 173 #endif 174 175 /* 176 * CC wrapper hook functions 177 */ 178 static void inline 179 cc_after_idle(struct tcpcb *tp) 180 { 181 INP_WLOCK_ASSERT(tp->t_inpcb); 182 183 if (CC_ALGO(tp)->after_idle != NULL) 184 CC_ALGO(tp)->after_idle(tp->ccv); 185 } 186 187 /* 188 * Tcp output routine: figure out what should be sent and send it. 189 */ 190 int 191 tcp_output(struct tcpcb *tp) 192 { 193 struct socket *so = tp->t_inpcb->inp_socket; 194 int32_t len; 195 uint32_t recwin, sendwin; 196 int off, flags, 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 #ifdef TCPDEBUG 202 struct ipovly *ipov = NULL; 203 #endif 204 struct tcphdr *th; 205 u_char opt[TCP_MAXOLEN]; 206 unsigned ipoptlen, optlen, hdrlen; 207 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 208 unsigned ipsec_optlen = 0; 209 #endif 210 int idle, sendalot, curticks; 211 int sack_rxmit, sack_bytes_rxmt; 212 struct sackhole *p; 213 int tso, mtu; 214 struct tcpopt to; 215 unsigned int wanted_cookie = 0; 216 unsigned int dont_sendalot = 0; 217 #if 0 218 int maxburst = TCP_MAXBURST; 219 #endif 220 #ifdef INET6 221 struct ip6_hdr *ip6 = NULL; 222 int isipv6; 223 224 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 225 #endif 226 #ifdef KERN_TLS 227 const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0; 228 #else 229 const bool hw_tls = false; 230 #endif 231 232 INP_WLOCK_ASSERT(tp->t_inpcb); 233 234 #ifdef TCP_OFFLOAD 235 if (tp->t_flags & TF_TOE) 236 return (tcp_offload_output(tp)); 237 #endif 238 239 /* 240 * For TFO connections in SYN_SENT or SYN_RECEIVED, 241 * only allow the initial SYN or SYN|ACK and those sent 242 * by the retransmit timer. 243 */ 244 if (IS_FASTOPEN(tp->t_flags) && 245 ((tp->t_state == TCPS_SYN_SENT) || 246 (tp->t_state == TCPS_SYN_RECEIVED)) && 247 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 248 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 249 return (0); 250 251 /* 252 * Determine length of data that should be transmitted, 253 * and flags that will be used. 254 * If there is some data or critical controls (SYN, RST) 255 * to send, then transmit; otherwise, investigate further. 256 */ 257 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 258 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) 259 cc_after_idle(tp); 260 tp->t_flags &= ~TF_LASTIDLE; 261 if (idle) { 262 if (tp->t_flags & TF_MORETOCOME) { 263 tp->t_flags |= TF_LASTIDLE; 264 idle = 0; 265 } 266 } 267 again: 268 /* 269 * If we've recently taken a timeout, snd_max will be greater than 270 * snd_nxt. There may be SACK information that allows us to avoid 271 * resending already delivered data. Adjust snd_nxt accordingly. 272 */ 273 if ((tp->t_flags & TF_SACK_PERMIT) && 274 SEQ_LT(tp->snd_nxt, tp->snd_max)) 275 tcp_sack_adjust(tp); 276 sendalot = 0; 277 tso = 0; 278 mtu = 0; 279 off = tp->snd_nxt - tp->snd_una; 280 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 281 282 flags = tcp_outflags[tp->t_state]; 283 /* 284 * Send any SACK-generated retransmissions. If we're explicitly trying 285 * to send out new data (when sendalot is 1), bypass this function. 286 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 287 * we're replacing a (future) new transmission with a retransmission 288 * now, and we previously incremented snd_cwnd in tcp_input(). 289 */ 290 /* 291 * Still in sack recovery , reset rxmit flag to zero. 292 */ 293 sack_rxmit = 0; 294 sack_bytes_rxmt = 0; 295 len = 0; 296 p = NULL; 297 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 298 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 299 uint32_t cwin; 300 301 cwin = 302 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0); 303 /* Do not retransmit SACK segments beyond snd_recover */ 304 if (SEQ_GT(p->end, tp->snd_recover)) { 305 /* 306 * (At least) part of sack hole extends beyond 307 * snd_recover. Check to see if we can rexmit data 308 * for this hole. 309 */ 310 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 311 /* 312 * Can't rexmit any more data for this hole. 313 * That data will be rexmitted in the next 314 * sack recovery episode, when snd_recover 315 * moves past p->rxmit. 316 */ 317 p = NULL; 318 goto after_sack_rexmit; 319 } else 320 /* Can rexmit part of the current hole */ 321 len = ((int32_t)ulmin(cwin, 322 tp->snd_recover - p->rxmit)); 323 } else 324 len = ((int32_t)ulmin(cwin, p->end - p->rxmit)); 325 off = p->rxmit - tp->snd_una; 326 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 327 __func__, off)); 328 if (len > 0) { 329 sack_rxmit = 1; 330 sendalot = 1; 331 TCPSTAT_INC(tcps_sack_rexmits); 332 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 333 min(len, tp->t_maxseg)); 334 } 335 } 336 after_sack_rexmit: 337 /* 338 * Get standard flags, and add SYN or FIN if requested by 'hidden' 339 * state flags. 340 */ 341 if (tp->t_flags & TF_NEEDFIN) 342 flags |= TH_FIN; 343 if (tp->t_flags & TF_NEEDSYN) 344 flags |= TH_SYN; 345 346 SOCKBUF_LOCK(&so->so_snd); 347 /* 348 * If in persist timeout with window of 0, send 1 byte. 349 * Otherwise, if window is small but nonzero 350 * and timer expired, we will send what we can 351 * and go to transmit state. 352 */ 353 if (tp->t_flags & TF_FORCEDATA) { 354 if (sendwin == 0) { 355 /* 356 * If we still have some data to send, then 357 * clear the FIN bit. Usually this would 358 * happen below when it realizes that we 359 * aren't sending all the data. However, 360 * if we have exactly 1 byte of unsent data, 361 * then it won't clear the FIN bit below, 362 * and if we are in persist state, we wind 363 * up sending the packet without recording 364 * that we sent the FIN bit. 365 * 366 * We can't just blindly clear the FIN bit, 367 * because if we don't have any more data 368 * to send then the probe will be the FIN 369 * itself. 370 */ 371 if (off < sbused(&so->so_snd)) 372 flags &= ~TH_FIN; 373 sendwin = 1; 374 } else { 375 tcp_timer_activate(tp, TT_PERSIST, 0); 376 tp->t_rxtshift = 0; 377 } 378 } 379 380 /* 381 * If snd_nxt == snd_max and we have transmitted a FIN, the 382 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 383 * a negative length. This can also occur when TCP opens up 384 * its congestion window while receiving additional duplicate 385 * acks after fast-retransmit because TCP will reset snd_nxt 386 * to snd_max after the fast-retransmit. 387 * 388 * In the normal retransmit-FIN-only case, however, snd_nxt will 389 * be set to snd_una, the offset will be 0, and the length may 390 * wind up 0. 391 * 392 * If sack_rxmit is true we are retransmitting from the scoreboard 393 * in which case len is already set. 394 */ 395 if (sack_rxmit == 0) { 396 if (sack_bytes_rxmt == 0) 397 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) - 398 off); 399 else { 400 int32_t cwin; 401 402 /* 403 * We are inside of a SACK recovery episode and are 404 * sending new data, having retransmitted all the 405 * data possible in the scoreboard. 406 */ 407 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) - 408 off); 409 /* 410 * Don't remove this (len > 0) check ! 411 * We explicitly check for len > 0 here (although it 412 * isn't really necessary), to work around a gcc 413 * optimization issue - to force gcc to compute 414 * len above. Without this check, the computation 415 * of len is bungled by the optimizer. 416 */ 417 if (len > 0) { 418 cwin = tp->snd_cwnd - 419 (tp->snd_nxt - tp->sack_newdata) - 420 sack_bytes_rxmt; 421 if (cwin < 0) 422 cwin = 0; 423 len = imin(len, cwin); 424 } 425 } 426 } 427 428 /* 429 * Lop off SYN bit if it has already been sent. However, if this 430 * is SYN-SENT state and if segment contains data and if we don't 431 * know that foreign host supports TAO, suppress sending segment. 432 */ 433 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 434 if (tp->t_state != TCPS_SYN_RECEIVED) 435 flags &= ~TH_SYN; 436 /* 437 * When sending additional segments following a TFO SYN|ACK, 438 * do not include the SYN bit. 439 */ 440 if (IS_FASTOPEN(tp->t_flags) && 441 (tp->t_state == TCPS_SYN_RECEIVED)) 442 flags &= ~TH_SYN; 443 off--, len++; 444 } 445 446 /* 447 * Be careful not to send data and/or FIN on SYN segments. 448 * This measure is needed to prevent interoperability problems 449 * with not fully conformant TCP implementations. 450 */ 451 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 452 len = 0; 453 flags &= ~TH_FIN; 454 } 455 456 /* 457 * On TFO sockets, ensure no data is sent in the following cases: 458 * 459 * - When retransmitting SYN|ACK on a passively-created socket 460 * 461 * - When retransmitting SYN on an actively created socket 462 * 463 * - When sending a zero-length cookie (cookie request) on an 464 * actively created socket 465 * 466 * - When the socket is in the CLOSED state (RST is being sent) 467 */ 468 if (IS_FASTOPEN(tp->t_flags) && 469 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 470 ((tp->t_state == TCPS_SYN_SENT) && 471 (tp->t_tfo_client_cookie_len == 0)) || 472 (flags & TH_RST))) 473 len = 0; 474 if (len <= 0) { 475 /* 476 * If FIN has been sent but not acked, 477 * but we haven't been called to retransmit, 478 * len will be < 0. Otherwise, window shrank 479 * after we sent into it. If window shrank to 0, 480 * cancel pending retransmit, pull snd_nxt back 481 * to (closed) window, and set the persist timer 482 * if it isn't already going. If the window didn't 483 * close completely, just wait for an ACK. 484 * 485 * We also do a general check here to ensure that 486 * we will set the persist timer when we have data 487 * to send, but a 0-byte window. This makes sure 488 * the persist timer is set even if the packet 489 * hits one of the "goto send" lines below. 490 */ 491 len = 0; 492 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 493 (off < (int) sbavail(&so->so_snd))) { 494 tcp_timer_activate(tp, TT_REXMT, 0); 495 tp->t_rxtshift = 0; 496 tp->snd_nxt = tp->snd_una; 497 if (!tcp_timer_active(tp, TT_PERSIST)) 498 tcp_setpersist(tp); 499 } 500 } 501 502 /* len will be >= 0 after this point. */ 503 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 504 505 tcp_sndbuf_autoscale(tp, so, sendwin); 506 507 /* 508 * Decide if we can use TCP Segmentation Offloading (if supported by 509 * hardware). 510 * 511 * TSO may only be used if we are in a pure bulk sending state. The 512 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 513 * IP options prevent using TSO. With TSO the TCP header is the same 514 * (except for the sequence number) for all generated packets. This 515 * makes it impossible to transmit any options which vary per generated 516 * segment or packet. 517 * 518 * IPv4 handling has a clear separation of ip options and ip header 519 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 520 * the right thing below to provide length of just ip options and thus 521 * checking for ipoptlen is enough to decide if ip options are present. 522 */ 523 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 524 /* 525 * Pre-calculate here as we save another lookup into the darknesses 526 * of IPsec that way and can actually decide if TSO is ok. 527 */ 528 #ifdef INET6 529 if (isipv6 && IPSEC_ENABLED(ipv6)) 530 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); 531 #ifdef INET 532 else 533 #endif 534 #endif /* INET6 */ 535 #ifdef INET 536 if (IPSEC_ENABLED(ipv4)) 537 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); 538 #endif /* INET */ 539 #endif /* IPSEC */ 540 #ifdef INET6 541 if (isipv6) 542 ipoptlen = ip6_optlen(tp->t_inpcb); 543 else 544 #endif 545 if (tp->t_inpcb->inp_options) 546 ipoptlen = tp->t_inpcb->inp_options->m_len - 547 offsetof(struct ipoption, ipopt_list); 548 else 549 ipoptlen = 0; 550 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 551 ipoptlen += ipsec_optlen; 552 #endif 553 554 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 555 ((tp->t_flags & TF_SIGNATURE) == 0) && 556 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 557 ipoptlen == 0 && !(flags & TH_SYN)) 558 tso = 1; 559 560 if (sack_rxmit) { 561 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd))) 562 flags &= ~TH_FIN; 563 } else { 564 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 565 sbused(&so->so_snd))) 566 flags &= ~TH_FIN; 567 } 568 569 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 570 (long)TCP_MAXWIN << tp->rcv_scale); 571 572 /* 573 * Sender silly window avoidance. We transmit under the following 574 * conditions when len is non-zero: 575 * 576 * - We have a full segment (or more with TSO) 577 * - This is the last buffer in a write()/send() and we are 578 * either idle or running NODELAY 579 * - we've timed out (e.g. persist timer) 580 * - we have more then 1/2 the maximum send window's worth of 581 * data (receiver may be limited the window size) 582 * - we need to retransmit 583 */ 584 if (len) { 585 if (len >= tp->t_maxseg) 586 goto send; 587 /* 588 * NOTE! on localhost connections an 'ack' from the remote 589 * end may occur synchronously with the output and cause 590 * us to flush a buffer queued with moretocome. XXX 591 * 592 * note: the len + off check is almost certainly unnecessary. 593 */ 594 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 595 (idle || (tp->t_flags & TF_NODELAY)) && 596 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) && 597 (tp->t_flags & TF_NOPUSH) == 0) { 598 goto send; 599 } 600 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 601 goto send; 602 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 603 goto send; 604 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 605 goto send; 606 if (sack_rxmit) 607 goto send; 608 } 609 610 /* 611 * Sending of standalone window updates. 612 * 613 * Window updates are important when we close our window due to a 614 * full socket buffer and are opening it again after the application 615 * reads data from it. Once the window has opened again and the 616 * remote end starts to send again the ACK clock takes over and 617 * provides the most current window information. 618 * 619 * We must avoid the silly window syndrome whereas every read 620 * from the receive buffer, no matter how small, causes a window 621 * update to be sent. We also should avoid sending a flurry of 622 * window updates when the socket buffer had queued a lot of data 623 * and the application is doing small reads. 624 * 625 * Prevent a flurry of pointless window updates by only sending 626 * an update when we can increase the advertized window by more 627 * than 1/4th of the socket buffer capacity. When the buffer is 628 * getting full or is very small be more aggressive and send an 629 * update whenever we can increase by two mss sized segments. 630 * In all other situations the ACK's to new incoming data will 631 * carry further window increases. 632 * 633 * Don't send an independent window update if a delayed 634 * ACK is pending (it will get piggy-backed on it) or the 635 * remote side already has done a half-close and won't send 636 * more data. Skip this if the connection is in T/TCP 637 * half-open state. 638 */ 639 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 640 !(tp->t_flags & TF_DELACK) && 641 !TCPS_HAVERCVDFIN(tp->t_state)) { 642 /* 643 * "adv" is the amount we could increase the window, 644 * taking into account that we are limited by 645 * TCP_MAXWIN << tp->rcv_scale. 646 */ 647 int32_t adv; 648 int oldwin; 649 650 adv = recwin; 651 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 652 oldwin = (tp->rcv_adv - tp->rcv_nxt); 653 adv -= oldwin; 654 } else 655 oldwin = 0; 656 657 /* 658 * If the new window size ends up being the same as or less 659 * than the old size when it is scaled, then don't force 660 * a window update. 661 */ 662 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 663 goto dontupdate; 664 665 if (adv >= (int32_t)(2 * tp->t_maxseg) && 666 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 667 recwin <= (so->so_rcv.sb_hiwat / 8) || 668 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg || 669 adv >= TCP_MAXWIN << tp->rcv_scale)) 670 goto send; 671 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat) 672 goto send; 673 } 674 dontupdate: 675 676 /* 677 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 678 * is also a catch-all for the retransmit timer timeout case. 679 */ 680 if (tp->t_flags & TF_ACKNOW) 681 goto send; 682 if ((flags & TH_RST) || 683 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 684 goto send; 685 if (SEQ_GT(tp->snd_up, tp->snd_una)) 686 goto send; 687 /* 688 * If our state indicates that FIN should be sent 689 * and we have not yet done so, then we need to send. 690 */ 691 if (flags & TH_FIN && 692 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 693 goto send; 694 /* 695 * In SACK, it is possible for tcp_output to fail to send a segment 696 * after the retransmission timer has been turned off. Make sure 697 * that the retransmission timer is set. 698 */ 699 if ((tp->t_flags & TF_SACK_PERMIT) && 700 SEQ_GT(tp->snd_max, tp->snd_una) && 701 !tcp_timer_active(tp, TT_REXMT) && 702 !tcp_timer_active(tp, TT_PERSIST)) { 703 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 704 goto just_return; 705 } 706 /* 707 * TCP window updates are not reliable, rather a polling protocol 708 * using ``persist'' packets is used to insure receipt of window 709 * updates. The three ``states'' for the output side are: 710 * idle not doing retransmits or persists 711 * persisting to move a small or zero window 712 * (re)transmitting and thereby not persisting 713 * 714 * tcp_timer_active(tp, TT_PERSIST) 715 * is true when we are in persist state. 716 * (tp->t_flags & TF_FORCEDATA) 717 * is set when we are called to send a persist packet. 718 * tcp_timer_active(tp, TT_REXMT) 719 * is set when we are retransmitting 720 * The output side is idle when both timers are zero. 721 * 722 * If send window is too small, there is data to transmit, and no 723 * retransmit or persist is pending, then go to persist state. 724 * If nothing happens soon, send when timer expires: 725 * if window is nonzero, transmit what we can, 726 * otherwise force out a byte. 727 */ 728 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && 729 !tcp_timer_active(tp, TT_PERSIST)) { 730 tp->t_rxtshift = 0; 731 tcp_setpersist(tp); 732 } 733 734 /* 735 * No reason to send a segment, just return. 736 */ 737 just_return: 738 SOCKBUF_UNLOCK(&so->so_snd); 739 return (0); 740 741 send: 742 SOCKBUF_LOCK_ASSERT(&so->so_snd); 743 if (len > 0) { 744 if (len >= tp->t_maxseg) 745 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 746 else 747 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 748 } 749 /* 750 * Before ESTABLISHED, force sending of initial options 751 * unless TCP set not to do any options. 752 * NOTE: we assume that the IP/TCP header plus TCP options 753 * always fit in a single mbuf, leaving room for a maximum 754 * link header, i.e. 755 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 756 */ 757 optlen = 0; 758 #ifdef INET6 759 if (isipv6) 760 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 761 else 762 #endif 763 hdrlen = sizeof (struct tcpiphdr); 764 765 /* 766 * Compute options for segment. 767 * We only have to care about SYN and established connection 768 * segments. Options for SYN-ACK segments are handled in TCP 769 * syncache. 770 */ 771 to.to_flags = 0; 772 if ((tp->t_flags & TF_NOOPT) == 0) { 773 /* Maximum segment size. */ 774 if (flags & TH_SYN) { 775 tp->snd_nxt = tp->iss; 776 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 777 to.to_flags |= TOF_MSS; 778 779 /* 780 * On SYN or SYN|ACK transmits on TFO connections, 781 * only include the TFO option if it is not a 782 * retransmit, as the presence of the TFO option may 783 * have caused the original SYN or SYN|ACK to have 784 * been dropped by a middlebox. 785 */ 786 if (IS_FASTOPEN(tp->t_flags) && 787 (tp->t_rxtshift == 0)) { 788 if (tp->t_state == TCPS_SYN_RECEIVED) { 789 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 790 to.to_tfo_cookie = 791 (u_int8_t *)&tp->t_tfo_cookie.server; 792 to.to_flags |= TOF_FASTOPEN; 793 wanted_cookie = 1; 794 } else if (tp->t_state == TCPS_SYN_SENT) { 795 to.to_tfo_len = 796 tp->t_tfo_client_cookie_len; 797 to.to_tfo_cookie = 798 tp->t_tfo_cookie.client; 799 to.to_flags |= TOF_FASTOPEN; 800 wanted_cookie = 1; 801 /* 802 * If we wind up having more data to 803 * send with the SYN than can fit in 804 * one segment, don't send any more 805 * until the SYN|ACK comes back from 806 * the other end. 807 */ 808 dont_sendalot = 1; 809 } 810 } 811 } 812 /* Window scaling. */ 813 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 814 to.to_wscale = tp->request_r_scale; 815 to.to_flags |= TOF_SCALE; 816 } 817 /* Timestamps. */ 818 if ((tp->t_flags & TF_RCVD_TSTMP) || 819 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 820 curticks = tcp_ts_getticks(); 821 to.to_tsval = curticks + tp->ts_offset; 822 to.to_tsecr = tp->ts_recent; 823 to.to_flags |= TOF_TS; 824 if (tp->t_rxtshift == 1) 825 tp->t_badrxtwin = curticks; 826 } 827 828 /* Set receive buffer autosizing timestamp. */ 829 if (tp->rfbuf_ts == 0 && 830 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 831 tp->rfbuf_ts = tcp_ts_getticks(); 832 833 /* Selective ACK's. */ 834 if (tp->t_flags & TF_SACK_PERMIT) { 835 if (flags & TH_SYN) 836 to.to_flags |= TOF_SACKPERM; 837 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 838 (tp->t_flags & TF_SACK_PERMIT) && 839 tp->rcv_numsacks > 0) { 840 to.to_flags |= TOF_SACK; 841 to.to_nsacks = tp->rcv_numsacks; 842 to.to_sacks = (u_char *)tp->sackblks; 843 } 844 } 845 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 846 /* TCP-MD5 (RFC2385). */ 847 /* 848 * Check that TCP_MD5SIG is enabled in tcpcb to 849 * account the size needed to set this TCP option. 850 */ 851 if (tp->t_flags & TF_SIGNATURE) 852 to.to_flags |= TOF_SIGNATURE; 853 #endif /* TCP_SIGNATURE */ 854 855 /* Processing the options. */ 856 hdrlen += optlen = tcp_addoptions(&to, opt); 857 /* 858 * If we wanted a TFO option to be added, but it was unable 859 * to fit, ensure no data is sent. 860 */ 861 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 862 !(to.to_flags & TOF_FASTOPEN)) 863 len = 0; 864 } 865 866 /* 867 * Adjust data length if insertion of options will 868 * bump the packet length beyond the t_maxseg length. 869 * Clear the FIN bit because we cut off the tail of 870 * the segment. 871 */ 872 if (len + optlen + ipoptlen > tp->t_maxseg) { 873 flags &= ~TH_FIN; 874 875 if (tso) { 876 u_int if_hw_tsomax; 877 u_int moff; 878 int max_len; 879 880 /* extract TSO information */ 881 if_hw_tsomax = tp->t_tsomax; 882 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 883 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 884 885 /* 886 * Limit a TSO burst to prevent it from 887 * overflowing or exceeding the maximum length 888 * allowed by the network interface: 889 */ 890 KASSERT(ipoptlen == 0, 891 ("%s: TSO can't do IP options", __func__)); 892 893 /* 894 * Check if we should limit by maximum payload 895 * length: 896 */ 897 if (if_hw_tsomax != 0) { 898 /* compute maximum TSO length */ 899 max_len = (if_hw_tsomax - hdrlen - 900 max_linkhdr); 901 if (max_len <= 0) { 902 len = 0; 903 } else if (len > max_len) { 904 sendalot = 1; 905 len = max_len; 906 } 907 } 908 909 /* 910 * Prevent the last segment from being 911 * fractional unless the send sockbuf can be 912 * emptied: 913 */ 914 max_len = (tp->t_maxseg - optlen); 915 if (((uint32_t)off + (uint32_t)len) < 916 sbavail(&so->so_snd)) { 917 moff = len % max_len; 918 if (moff != 0) { 919 len -= moff; 920 sendalot = 1; 921 } 922 } 923 924 /* 925 * In case there are too many small fragments 926 * don't use TSO: 927 */ 928 if (len <= max_len) { 929 len = max_len; 930 sendalot = 1; 931 tso = 0; 932 } 933 934 /* 935 * Send the FIN in a separate segment 936 * after the bulk sending is done. 937 * We don't trust the TSO implementations 938 * to clear the FIN flag on all but the 939 * last segment. 940 */ 941 if (tp->t_flags & TF_NEEDFIN) 942 sendalot = 1; 943 } else { 944 if (optlen + ipoptlen >= tp->t_maxseg) { 945 /* 946 * Since we don't have enough space to put 947 * the IP header chain and the TCP header in 948 * one packet as required by RFC 7112, don't 949 * send it. Also ensure that at least one 950 * byte of the payload can be put into the 951 * TCP segment. 952 */ 953 SOCKBUF_UNLOCK(&so->so_snd); 954 error = EMSGSIZE; 955 sack_rxmit = 0; 956 goto out; 957 } 958 len = tp->t_maxseg - optlen - ipoptlen; 959 sendalot = 1; 960 if (dont_sendalot) 961 sendalot = 0; 962 } 963 } else 964 tso = 0; 965 966 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 967 ("%s: len > IP_MAXPACKET", __func__)); 968 969 /*#ifdef DIAGNOSTIC*/ 970 #ifdef INET6 971 if (max_linkhdr + hdrlen > MCLBYTES) 972 #else 973 if (max_linkhdr + hdrlen > MHLEN) 974 #endif 975 panic("tcphdr too big"); 976 /*#endif*/ 977 978 /* 979 * This KASSERT is here to catch edge cases at a well defined place. 980 * Before, those had triggered (random) panic conditions further down. 981 */ 982 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 983 984 /* 985 * Grab a header mbuf, attaching a copy of data to 986 * be transmitted, and initialize the header from 987 * the template for sends on this connection. 988 */ 989 if (len) { 990 struct mbuf *mb; 991 struct sockbuf *msb; 992 u_int moff; 993 994 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 995 TCPSTAT_INC(tcps_sndprobe); 996 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 997 tp->t_sndrexmitpack++; 998 TCPSTAT_INC(tcps_sndrexmitpack); 999 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 1000 } else { 1001 TCPSTAT_INC(tcps_sndpack); 1002 TCPSTAT_ADD(tcps_sndbyte, len); 1003 } 1004 #ifdef INET6 1005 if (MHLEN < hdrlen + max_linkhdr) 1006 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1007 else 1008 #endif 1009 m = m_gethdr(M_NOWAIT, MT_DATA); 1010 1011 if (m == NULL) { 1012 SOCKBUF_UNLOCK(&so->so_snd); 1013 error = ENOBUFS; 1014 sack_rxmit = 0; 1015 goto out; 1016 } 1017 1018 m->m_data += max_linkhdr; 1019 m->m_len = hdrlen; 1020 1021 /* 1022 * Start the m_copy functions from the closest mbuf 1023 * to the offset in the socket buffer chain. 1024 */ 1025 mb = sbsndptr_noadv(&so->so_snd, off, &moff); 1026 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 1027 m_copydata(mb, moff, len, 1028 mtod(m, caddr_t) + hdrlen); 1029 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1030 sbsndptr_adv(&so->so_snd, mb, len); 1031 m->m_len += len; 1032 } else { 1033 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1034 msb = NULL; 1035 else 1036 msb = &so->so_snd; 1037 m->m_next = tcp_m_copym(mb, moff, 1038 &len, if_hw_tsomaxsegcount, 1039 if_hw_tsomaxsegsize, msb, hw_tls); 1040 if (len <= (tp->t_maxseg - optlen)) { 1041 /* 1042 * Must have ran out of mbufs for the copy 1043 * shorten it to no longer need tso. Lets 1044 * not put on sendalot since we are low on 1045 * mbufs. 1046 */ 1047 tso = 0; 1048 } 1049 if (m->m_next == NULL) { 1050 SOCKBUF_UNLOCK(&so->so_snd); 1051 (void) m_free(m); 1052 error = ENOBUFS; 1053 sack_rxmit = 0; 1054 goto out; 1055 } 1056 } 1057 1058 /* 1059 * If we're sending everything we've got, set PUSH. 1060 * (This will keep happy those implementations which only 1061 * give data to the user when a buffer fills or 1062 * a PUSH comes in.) 1063 */ 1064 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) && 1065 !(flags & TH_SYN)) 1066 flags |= TH_PUSH; 1067 SOCKBUF_UNLOCK(&so->so_snd); 1068 } else { 1069 SOCKBUF_UNLOCK(&so->so_snd); 1070 if (tp->t_flags & TF_ACKNOW) 1071 TCPSTAT_INC(tcps_sndacks); 1072 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1073 TCPSTAT_INC(tcps_sndctrl); 1074 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1075 TCPSTAT_INC(tcps_sndurg); 1076 else 1077 TCPSTAT_INC(tcps_sndwinup); 1078 1079 m = m_gethdr(M_NOWAIT, MT_DATA); 1080 if (m == NULL) { 1081 error = ENOBUFS; 1082 sack_rxmit = 0; 1083 goto out; 1084 } 1085 #ifdef INET6 1086 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1087 MHLEN >= hdrlen) { 1088 M_ALIGN(m, hdrlen); 1089 } else 1090 #endif 1091 m->m_data += max_linkhdr; 1092 m->m_len = hdrlen; 1093 } 1094 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1095 m->m_pkthdr.rcvif = (struct ifnet *)0; 1096 #ifdef MAC 1097 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1098 #endif 1099 #ifdef INET6 1100 if (isipv6) { 1101 ip6 = mtod(m, struct ip6_hdr *); 1102 th = (struct tcphdr *)(ip6 + 1); 1103 tcpip_fillheaders(tp->t_inpcb, ip6, th); 1104 } else 1105 #endif /* INET6 */ 1106 { 1107 ip = mtod(m, struct ip *); 1108 #ifdef TCPDEBUG 1109 ipov = (struct ipovly *)ip; 1110 #endif 1111 th = (struct tcphdr *)(ip + 1); 1112 tcpip_fillheaders(tp->t_inpcb, ip, th); 1113 } 1114 1115 /* 1116 * Fill in fields, remembering maximum advertised 1117 * window for use in delaying messages about window sizes. 1118 * If resending a FIN, be sure not to use a new sequence number. 1119 */ 1120 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1121 tp->snd_nxt == tp->snd_max) 1122 tp->snd_nxt--; 1123 /* 1124 * If we are starting a connection, send ECN setup 1125 * SYN packet. If we are on a retransmit, we may 1126 * resend those bits a number of times as per 1127 * RFC 3168. 1128 */ 1129 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { 1130 if (tp->t_rxtshift >= 1) { 1131 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 1132 flags |= TH_ECE|TH_CWR; 1133 } else 1134 flags |= TH_ECE|TH_CWR; 1135 } 1136 1137 if (tp->t_state == TCPS_ESTABLISHED && 1138 (tp->t_flags & TF_ECN_PERMIT)) { 1139 /* 1140 * If the peer has ECN, mark data packets with 1141 * ECN capable transmission (ECT). 1142 * Ignore pure ack packets, retransmissions and window probes. 1143 */ 1144 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1145 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 1146 #ifdef INET6 1147 if (isipv6) 1148 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1149 else 1150 #endif 1151 ip->ip_tos |= IPTOS_ECN_ECT0; 1152 TCPSTAT_INC(tcps_ecn_ect0); 1153 } 1154 1155 /* 1156 * Reply with proper ECN notifications. 1157 */ 1158 if (tp->t_flags & TF_ECN_SND_CWR) { 1159 flags |= TH_CWR; 1160 tp->t_flags &= ~TF_ECN_SND_CWR; 1161 } 1162 if (tp->t_flags & TF_ECN_SND_ECE) 1163 flags |= TH_ECE; 1164 } 1165 1166 /* 1167 * If we are doing retransmissions, then snd_nxt will 1168 * not reflect the first unsent octet. For ACK only 1169 * packets, we do not want the sequence number of the 1170 * retransmitted packet, we want the sequence number 1171 * of the next unsent octet. So, if there is no data 1172 * (and no SYN or FIN), use snd_max instead of snd_nxt 1173 * when filling in ti_seq. But if we are in persist 1174 * state, snd_max might reflect one byte beyond the 1175 * right edge of the window, so use snd_nxt in that 1176 * case, since we know we aren't doing a retransmission. 1177 * (retransmit and persist are mutually exclusive...) 1178 */ 1179 if (sack_rxmit == 0) { 1180 if (len || (flags & (TH_SYN|TH_FIN)) || 1181 tcp_timer_active(tp, TT_PERSIST)) 1182 th->th_seq = htonl(tp->snd_nxt); 1183 else 1184 th->th_seq = htonl(tp->snd_max); 1185 } else { 1186 th->th_seq = htonl(p->rxmit); 1187 p->rxmit += len; 1188 tp->sackhint.sack_bytes_rexmit += len; 1189 } 1190 th->th_ack = htonl(tp->rcv_nxt); 1191 if (optlen) { 1192 bcopy(opt, th + 1, optlen); 1193 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1194 } 1195 th->th_flags = flags; 1196 /* 1197 * Calculate receive window. Don't shrink window, 1198 * but avoid silly window syndrome. 1199 * If a RST segment is sent, advertise a window of zero. 1200 */ 1201 if (flags & TH_RST) { 1202 recwin = 0; 1203 } else { 1204 if (recwin < (so->so_rcv.sb_hiwat / 4) && 1205 recwin < tp->t_maxseg) 1206 recwin = 0; 1207 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1208 recwin < (tp->rcv_adv - tp->rcv_nxt)) 1209 recwin = (tp->rcv_adv - tp->rcv_nxt); 1210 } 1211 /* 1212 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1213 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1214 * case is handled in syncache. 1215 */ 1216 if (flags & TH_SYN) 1217 th->th_win = htons((u_short) 1218 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1219 else 1220 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1221 1222 /* 1223 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1224 * a 0 window. This may cause the remote transmitter to stall. This 1225 * flag tells soreceive() to disable delayed acknowledgements when 1226 * draining the buffer. This can occur if the receiver is attempting 1227 * to read more data than can be buffered prior to transmitting on 1228 * the connection. 1229 */ 1230 if (th->th_win == 0) { 1231 tp->t_sndzerowin++; 1232 tp->t_flags |= TF_RXWIN0SENT; 1233 } else 1234 tp->t_flags &= ~TF_RXWIN0SENT; 1235 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1236 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1237 th->th_flags |= TH_URG; 1238 } else 1239 /* 1240 * If no urgent pointer to send, then we pull 1241 * the urgent pointer to the left edge of the send window 1242 * so that it doesn't drift into the send window on sequence 1243 * number wraparound. 1244 */ 1245 tp->snd_up = tp->snd_una; /* drag it along */ 1246 1247 /* 1248 * Put TCP length in extended header, and then 1249 * checksum extended header and data. 1250 */ 1251 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1252 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1253 1254 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1255 if (to.to_flags & TOF_SIGNATURE) { 1256 /* 1257 * Calculate MD5 signature and put it into the place 1258 * determined before. 1259 * NOTE: since TCP options buffer doesn't point into 1260 * mbuf's data, calculate offset and use it. 1261 */ 1262 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th, 1263 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) { 1264 /* 1265 * Do not send segment if the calculation of MD5 1266 * digest has failed. 1267 */ 1268 m_freem(m); 1269 goto out; 1270 } 1271 } 1272 #endif 1273 #ifdef INET6 1274 if (isipv6) { 1275 /* 1276 * There is no need to fill in ip6_plen right now. 1277 * It will be filled later by ip6_output. 1278 */ 1279 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1280 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1281 optlen + len, IPPROTO_TCP, 0); 1282 } 1283 #endif 1284 #if defined(INET6) && defined(INET) 1285 else 1286 #endif 1287 #ifdef INET 1288 { 1289 m->m_pkthdr.csum_flags = CSUM_TCP; 1290 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1291 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1292 1293 /* IP version must be set here for ipv4/ipv6 checking later */ 1294 KASSERT(ip->ip_v == IPVERSION, 1295 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1296 } 1297 #endif 1298 1299 /* 1300 * Enable TSO and specify the size of the segments. 1301 * The TCP pseudo header checksum is always provided. 1302 */ 1303 if (tso) { 1304 KASSERT(len > tp->t_maxseg - optlen, 1305 ("%s: len <= tso_segsz", __func__)); 1306 m->m_pkthdr.csum_flags |= CSUM_TSO; 1307 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 1308 } 1309 1310 KASSERT(len + hdrlen == m_length(m, NULL), 1311 ("%s: mbuf chain shorter than expected: %d + %u != %u", 1312 __func__, len, hdrlen, m_length(m, NULL))); 1313 1314 #ifdef TCP_HHOOK 1315 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1316 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1317 #endif 1318 1319 #ifdef TCPDEBUG 1320 /* 1321 * Trace. 1322 */ 1323 if (so->so_options & SO_DEBUG) { 1324 u_short save = 0; 1325 #ifdef INET6 1326 if (!isipv6) 1327 #endif 1328 { 1329 save = ipov->ih_len; 1330 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1331 } 1332 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1333 #ifdef INET6 1334 if (!isipv6) 1335 #endif 1336 ipov->ih_len = save; 1337 } 1338 #endif /* TCPDEBUG */ 1339 TCP_PROBE3(debug__output, tp, th, m); 1340 1341 /* We're getting ready to send; log now. */ 1342 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 1343 len, NULL, false); 1344 1345 /* 1346 * Fill in IP length and desired time to live and 1347 * send to IP level. There should be a better way 1348 * to handle ttl and tos; we could keep them in 1349 * the template, but need a way to checksum without them. 1350 */ 1351 /* 1352 * m->m_pkthdr.len should have been set before checksum calculation, 1353 * because in6_cksum() need it. 1354 */ 1355 #ifdef INET6 1356 if (isipv6) { 1357 /* 1358 * we separately set hoplimit for every segment, since the 1359 * user might want to change the value via setsockopt. 1360 * Also, desired default hop limit might be changed via 1361 * Neighbor Discovery. 1362 */ 1363 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1364 1365 /* 1366 * Set the packet size here for the benefit of DTrace probes. 1367 * ip6_output() will set it properly; it's supposed to include 1368 * the option header lengths as well. 1369 */ 1370 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1371 1372 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 1373 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1374 else 1375 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1376 1377 if (tp->t_state == TCPS_SYN_SENT) 1378 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1379 1380 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1381 1382 #ifdef TCPPCAP 1383 /* Save packet, if requested. */ 1384 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1385 #endif 1386 1387 /* TODO: IPv6 IP6TOS_ECT bit on */ 1388 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, 1389 &tp->t_inpcb->inp_route6, 1390 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1391 NULL, NULL, tp->t_inpcb); 1392 1393 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL) 1394 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu; 1395 } 1396 #endif /* INET6 */ 1397 #if defined(INET) && defined(INET6) 1398 else 1399 #endif 1400 #ifdef INET 1401 { 1402 ip->ip_len = htons(m->m_pkthdr.len); 1403 #ifdef INET6 1404 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1405 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1406 #endif /* INET6 */ 1407 /* 1408 * If we do path MTU discovery, then we set DF on every packet. 1409 * This might not be the best thing to do according to RFC3390 1410 * Section 2. However the tcp hostcache migitates the problem 1411 * so it affects only the first tcp connection with a host. 1412 * 1413 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1414 */ 1415 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 1416 ip->ip_off |= htons(IP_DF); 1417 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1418 } else { 1419 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1420 } 1421 1422 if (tp->t_state == TCPS_SYN_SENT) 1423 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1424 1425 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1426 1427 #ifdef TCPPCAP 1428 /* Save packet, if requested. */ 1429 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1430 #endif 1431 1432 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route, 1433 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1434 tp->t_inpcb); 1435 1436 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL) 1437 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu; 1438 } 1439 #endif /* INET */ 1440 1441 out: 1442 /* 1443 * In transmit state, time the transmission and arrange for 1444 * the retransmit. In persist state, just set snd_max. 1445 */ 1446 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1447 !tcp_timer_active(tp, TT_PERSIST)) { 1448 tcp_seq startseq = tp->snd_nxt; 1449 1450 /* 1451 * Advance snd_nxt over sequence space of this segment. 1452 */ 1453 if (flags & (TH_SYN|TH_FIN)) { 1454 if (flags & TH_SYN) 1455 tp->snd_nxt++; 1456 if (flags & TH_FIN) { 1457 tp->snd_nxt++; 1458 tp->t_flags |= TF_SENTFIN; 1459 } 1460 } 1461 if (sack_rxmit) 1462 goto timer; 1463 tp->snd_nxt += len; 1464 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1465 tp->snd_max = tp->snd_nxt; 1466 /* 1467 * Time this transmission if not a retransmission and 1468 * not currently timing anything. 1469 */ 1470 if (tp->t_rtttime == 0) { 1471 tp->t_rtttime = ticks; 1472 tp->t_rtseq = startseq; 1473 TCPSTAT_INC(tcps_segstimed); 1474 } 1475 } 1476 1477 /* 1478 * Set retransmit timer if not currently set, 1479 * and not doing a pure ack or a keep-alive probe. 1480 * Initial value for retransmit timer is smoothed 1481 * round-trip time + 2 * round-trip time variance. 1482 * Initialize shift counter which is used for backoff 1483 * of retransmit time. 1484 */ 1485 timer: 1486 if (!tcp_timer_active(tp, TT_REXMT) && 1487 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1488 (tp->snd_nxt != tp->snd_una))) { 1489 if (tcp_timer_active(tp, TT_PERSIST)) { 1490 tcp_timer_activate(tp, TT_PERSIST, 0); 1491 tp->t_rxtshift = 0; 1492 } 1493 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1494 } else if (len == 0 && sbavail(&so->so_snd) && 1495 !tcp_timer_active(tp, TT_REXMT) && 1496 !tcp_timer_active(tp, TT_PERSIST)) { 1497 /* 1498 * Avoid a situation where we do not set persist timer 1499 * after a zero window condition. For example: 1500 * 1) A -> B: packet with enough data to fill the window 1501 * 2) B -> A: ACK for #1 + new data (0 window 1502 * advertisement) 1503 * 3) A -> B: ACK for #2, 0 len packet 1504 * 1505 * In this case, A will not activate the persist timer, 1506 * because it chose to send a packet. Unless tcp_output 1507 * is called for some other reason (delayed ack timer, 1508 * another input packet from B, socket syscall), A will 1509 * not send zero window probes. 1510 * 1511 * So, if you send a 0-length packet, but there is data 1512 * in the socket buffer, and neither the rexmt or 1513 * persist timer is already set, then activate the 1514 * persist timer. 1515 */ 1516 tp->t_rxtshift = 0; 1517 tcp_setpersist(tp); 1518 } 1519 } else { 1520 /* 1521 * Persist case, update snd_max but since we are in 1522 * persist mode (no window) we do not update snd_nxt. 1523 */ 1524 int xlen = len; 1525 if (flags & TH_SYN) 1526 ++xlen; 1527 if (flags & TH_FIN) { 1528 ++xlen; 1529 tp->t_flags |= TF_SENTFIN; 1530 } 1531 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1532 tp->snd_max = tp->snd_nxt + xlen; 1533 } 1534 if ((error == 0) && 1535 (TCPS_HAVEESTABLISHED(tp->t_state) && 1536 (tp->t_flags & TF_SACK_PERMIT) && 1537 tp->rcv_numsacks > 0)) { 1538 /* Clean up any DSACK's sent */ 1539 tcp_clean_dsack_blocks(tp); 1540 } 1541 if (error) { 1542 /* Record the error. */ 1543 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, 1544 error, 0, NULL, false); 1545 1546 /* 1547 * We know that the packet was lost, so back out the 1548 * sequence number advance, if any. 1549 * 1550 * If the error is EPERM the packet got blocked by the 1551 * local firewall. Normally we should terminate the 1552 * connection but the blocking may have been spurious 1553 * due to a firewall reconfiguration cycle. So we treat 1554 * it like a packet loss and let the retransmit timer and 1555 * timeouts do their work over time. 1556 * XXX: It is a POLA question whether calling tcp_drop right 1557 * away would be the really correct behavior instead. 1558 */ 1559 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1560 !tcp_timer_active(tp, TT_PERSIST)) && 1561 ((flags & TH_SYN) == 0) && 1562 (error != EPERM)) { 1563 if (sack_rxmit) { 1564 p->rxmit -= len; 1565 tp->sackhint.sack_bytes_rexmit -= len; 1566 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1567 ("sackhint bytes rtx >= 0")); 1568 } else 1569 tp->snd_nxt -= len; 1570 } 1571 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1572 switch (error) { 1573 case EACCES: 1574 case EPERM: 1575 tp->t_softerror = error; 1576 return (error); 1577 case ENOBUFS: 1578 TCP_XMIT_TIMER_ASSERT(tp, len, flags); 1579 tp->snd_cwnd = tp->t_maxseg; 1580 return (0); 1581 case EMSGSIZE: 1582 /* 1583 * For some reason the interface we used initially 1584 * to send segments changed to another or lowered 1585 * its MTU. 1586 * If TSO was active we either got an interface 1587 * without TSO capabilits or TSO was turned off. 1588 * If we obtained mtu from ip_output() then update 1589 * it and try again. 1590 */ 1591 if (tso) 1592 tp->t_flags &= ~TF_TSO; 1593 if (mtu != 0) { 1594 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1595 goto again; 1596 } 1597 return (error); 1598 case EHOSTDOWN: 1599 case EHOSTUNREACH: 1600 case ENETDOWN: 1601 case ENETUNREACH: 1602 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1603 tp->t_softerror = error; 1604 return (0); 1605 } 1606 /* FALLTHROUGH */ 1607 default: 1608 return (error); 1609 } 1610 } 1611 TCPSTAT_INC(tcps_sndtotal); 1612 1613 /* 1614 * Data sent (as far as we can tell). 1615 * If this advertises a larger window than any other segment, 1616 * then remember the size of the advertised window. 1617 * Any pending ACK has now been sent. 1618 */ 1619 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1620 tp->rcv_adv = tp->rcv_nxt + recwin; 1621 tp->last_ack_sent = tp->rcv_nxt; 1622 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1623 if (tcp_timer_active(tp, TT_DELACK)) 1624 tcp_timer_activate(tp, TT_DELACK, 0); 1625 #if 0 1626 /* 1627 * This completely breaks TCP if newreno is turned on. What happens 1628 * is that if delayed-acks are turned on on the receiver, this code 1629 * on the transmitter effectively destroys the TCP window, forcing 1630 * it to four packets (1.5Kx4 = 6K window). 1631 */ 1632 if (sendalot && --maxburst) 1633 goto again; 1634 #endif 1635 if (sendalot) 1636 goto again; 1637 return (0); 1638 } 1639 1640 void 1641 tcp_setpersist(struct tcpcb *tp) 1642 { 1643 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1644 int tt; 1645 1646 tp->t_flags &= ~TF_PREVVALID; 1647 if (tcp_timer_active(tp, TT_REXMT)) 1648 panic("tcp_setpersist: retransmit pending"); 1649 /* 1650 * Start/restart persistence timer. 1651 */ 1652 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1653 tcp_persmin, tcp_persmax); 1654 tcp_timer_activate(tp, TT_PERSIST, tt); 1655 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1656 tp->t_rxtshift++; 1657 } 1658 1659 /* 1660 * Insert TCP options according to the supplied parameters to the place 1661 * optp in a consistent way. Can handle unaligned destinations. 1662 * 1663 * The order of the option processing is crucial for optimal packing and 1664 * alignment for the scarce option space. 1665 * 1666 * The optimal order for a SYN/SYN-ACK segment is: 1667 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1668 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1669 * 1670 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1671 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1672 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1673 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1674 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1675 */ 1676 int 1677 tcp_addoptions(struct tcpopt *to, u_char *optp) 1678 { 1679 u_int32_t mask, optlen = 0; 1680 1681 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1682 if ((to->to_flags & mask) != mask) 1683 continue; 1684 if (optlen == TCP_MAXOLEN) 1685 break; 1686 switch (to->to_flags & mask) { 1687 case TOF_MSS: 1688 while (optlen % 4) { 1689 optlen += TCPOLEN_NOP; 1690 *optp++ = TCPOPT_NOP; 1691 } 1692 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1693 continue; 1694 optlen += TCPOLEN_MAXSEG; 1695 *optp++ = TCPOPT_MAXSEG; 1696 *optp++ = TCPOLEN_MAXSEG; 1697 to->to_mss = htons(to->to_mss); 1698 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1699 optp += sizeof(to->to_mss); 1700 break; 1701 case TOF_SCALE: 1702 while (!optlen || optlen % 2 != 1) { 1703 optlen += TCPOLEN_NOP; 1704 *optp++ = TCPOPT_NOP; 1705 } 1706 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1707 continue; 1708 optlen += TCPOLEN_WINDOW; 1709 *optp++ = TCPOPT_WINDOW; 1710 *optp++ = TCPOLEN_WINDOW; 1711 *optp++ = to->to_wscale; 1712 break; 1713 case TOF_SACKPERM: 1714 while (optlen % 2) { 1715 optlen += TCPOLEN_NOP; 1716 *optp++ = TCPOPT_NOP; 1717 } 1718 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1719 continue; 1720 optlen += TCPOLEN_SACK_PERMITTED; 1721 *optp++ = TCPOPT_SACK_PERMITTED; 1722 *optp++ = TCPOLEN_SACK_PERMITTED; 1723 break; 1724 case TOF_TS: 1725 while (!optlen || optlen % 4 != 2) { 1726 optlen += TCPOLEN_NOP; 1727 *optp++ = TCPOPT_NOP; 1728 } 1729 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1730 continue; 1731 optlen += TCPOLEN_TIMESTAMP; 1732 *optp++ = TCPOPT_TIMESTAMP; 1733 *optp++ = TCPOLEN_TIMESTAMP; 1734 to->to_tsval = htonl(to->to_tsval); 1735 to->to_tsecr = htonl(to->to_tsecr); 1736 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1737 optp += sizeof(to->to_tsval); 1738 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1739 optp += sizeof(to->to_tsecr); 1740 break; 1741 case TOF_SIGNATURE: 1742 { 1743 int siglen = TCPOLEN_SIGNATURE - 2; 1744 1745 while (!optlen || optlen % 4 != 2) { 1746 optlen += TCPOLEN_NOP; 1747 *optp++ = TCPOPT_NOP; 1748 } 1749 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) { 1750 to->to_flags &= ~TOF_SIGNATURE; 1751 continue; 1752 } 1753 optlen += TCPOLEN_SIGNATURE; 1754 *optp++ = TCPOPT_SIGNATURE; 1755 *optp++ = TCPOLEN_SIGNATURE; 1756 to->to_signature = optp; 1757 while (siglen--) 1758 *optp++ = 0; 1759 break; 1760 } 1761 case TOF_SACK: 1762 { 1763 int sackblks = 0; 1764 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1765 tcp_seq sack_seq; 1766 1767 while (!optlen || optlen % 4 != 2) { 1768 optlen += TCPOLEN_NOP; 1769 *optp++ = TCPOPT_NOP; 1770 } 1771 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1772 continue; 1773 optlen += TCPOLEN_SACKHDR; 1774 *optp++ = TCPOPT_SACK; 1775 sackblks = min(to->to_nsacks, 1776 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1777 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1778 while (sackblks--) { 1779 sack_seq = htonl(sack->start); 1780 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1781 optp += sizeof(sack_seq); 1782 sack_seq = htonl(sack->end); 1783 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1784 optp += sizeof(sack_seq); 1785 optlen += TCPOLEN_SACK; 1786 sack++; 1787 } 1788 TCPSTAT_INC(tcps_sack_send_blocks); 1789 break; 1790 } 1791 case TOF_FASTOPEN: 1792 { 1793 int total_len; 1794 1795 /* XXX is there any point to aligning this option? */ 1796 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; 1797 if (TCP_MAXOLEN - optlen < total_len) { 1798 to->to_flags &= ~TOF_FASTOPEN; 1799 continue; 1800 } 1801 *optp++ = TCPOPT_FAST_OPEN; 1802 *optp++ = total_len; 1803 if (to->to_tfo_len > 0) { 1804 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); 1805 optp += to->to_tfo_len; 1806 } 1807 optlen += total_len; 1808 break; 1809 } 1810 default: 1811 panic("%s: unknown TCP option type", __func__); 1812 break; 1813 } 1814 } 1815 1816 /* Terminate and pad TCP options to a 4 byte boundary. */ 1817 if (optlen % 4) { 1818 optlen += TCPOLEN_EOL; 1819 *optp++ = TCPOPT_EOL; 1820 } 1821 /* 1822 * According to RFC 793 (STD0007): 1823 * "The content of the header beyond the End-of-Option option 1824 * must be header padding (i.e., zero)." 1825 * and later: "The padding is composed of zeros." 1826 */ 1827 while (optlen % 4) { 1828 optlen += TCPOLEN_PAD; 1829 *optp++ = TCPOPT_PAD; 1830 } 1831 1832 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1833 return (optlen); 1834 } 1835 1836 /* 1837 * This is a copy of m_copym(), taking the TSO segment size/limit 1838 * constraints into account, and advancing the sndptr as it goes. 1839 */ 1840 struct mbuf * 1841 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen, 1842 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls) 1843 { 1844 #ifdef KERN_TLS 1845 struct ktls_session *tls, *ntls; 1846 struct mbuf *start; 1847 #endif 1848 struct mbuf *n, **np; 1849 struct mbuf *top; 1850 int32_t off = off0; 1851 int32_t len = *plen; 1852 int32_t fragsize; 1853 int32_t len_cp = 0; 1854 int32_t *pkthdrlen; 1855 uint32_t mlen, frags; 1856 bool copyhdr; 1857 1858 1859 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off)); 1860 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len)); 1861 if (off == 0 && m->m_flags & M_PKTHDR) 1862 copyhdr = true; 1863 else 1864 copyhdr = false; 1865 while (off > 0) { 1866 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain")); 1867 if (off < m->m_len) 1868 break; 1869 off -= m->m_len; 1870 if ((sb) && (m == sb->sb_sndptr)) { 1871 sb->sb_sndptroff += m->m_len; 1872 sb->sb_sndptr = m->m_next; 1873 } 1874 m = m->m_next; 1875 } 1876 np = ⊤ 1877 top = NULL; 1878 pkthdrlen = NULL; 1879 #ifdef KERN_TLS 1880 if (m->m_flags & M_NOMAP) 1881 tls = m->m_ext.ext_pgs->tls; 1882 else 1883 tls = NULL; 1884 start = m; 1885 #endif 1886 while (len > 0) { 1887 if (m == NULL) { 1888 KASSERT(len == M_COPYALL, 1889 ("tcp_m_copym, length > size of mbuf chain")); 1890 *plen = len_cp; 1891 if (pkthdrlen != NULL) 1892 *pkthdrlen = len_cp; 1893 break; 1894 } 1895 #ifdef KERN_TLS 1896 if (hw_tls) { 1897 if (m->m_flags & M_NOMAP) 1898 ntls = m->m_ext.ext_pgs->tls; 1899 else 1900 ntls = NULL; 1901 1902 /* 1903 * Avoid mixing TLS records with handshake 1904 * data or TLS records from different 1905 * sessions. 1906 */ 1907 if (tls != ntls) { 1908 MPASS(m != start); 1909 *plen = len_cp; 1910 if (pkthdrlen != NULL) 1911 *pkthdrlen = len_cp; 1912 break; 1913 } 1914 1915 /* 1916 * Don't end a send in the middle of a TLS 1917 * record if it spans multiple TLS records. 1918 */ 1919 if (tls != NULL && (m != start) && len < m->m_len) { 1920 *plen = len_cp; 1921 if (pkthdrlen != NULL) 1922 *pkthdrlen = len_cp; 1923 break; 1924 } 1925 } 1926 #endif 1927 mlen = min(len, m->m_len - off); 1928 if (seglimit) { 1929 /* 1930 * For M_NOMAP mbufs, add 3 segments 1931 * + 1 in case we are crossing page boundaries 1932 * + 2 in case the TLS hdr/trailer are used 1933 * It is cheaper to just add the segments 1934 * than it is to take the cache miss to look 1935 * at the mbuf ext_pgs state in detail. 1936 */ 1937 if (m->m_flags & M_NOMAP) { 1938 fragsize = min(segsize, PAGE_SIZE); 1939 frags = 3; 1940 } else { 1941 fragsize = segsize; 1942 frags = 0; 1943 } 1944 1945 /* Break if we really can't fit anymore. */ 1946 if ((frags + 1) >= seglimit) { 1947 *plen = len_cp; 1948 if (pkthdrlen != NULL) 1949 *pkthdrlen = len_cp; 1950 break; 1951 } 1952 1953 /* 1954 * Reduce size if you can't copy the whole 1955 * mbuf. If we can't copy the whole mbuf, also 1956 * adjust len so the loop will end after this 1957 * mbuf. 1958 */ 1959 if ((frags + howmany(mlen, fragsize)) >= seglimit) { 1960 mlen = (seglimit - frags - 1) * fragsize; 1961 len = mlen; 1962 *plen = len_cp + len; 1963 if (pkthdrlen != NULL) 1964 *pkthdrlen = *plen; 1965 } 1966 frags += howmany(mlen, fragsize); 1967 if (frags == 0) 1968 frags++; 1969 seglimit -= frags; 1970 KASSERT(seglimit > 0, 1971 ("%s: seglimit went too low", __func__)); 1972 } 1973 if (copyhdr) 1974 n = m_gethdr(M_NOWAIT, m->m_type); 1975 else 1976 n = m_get(M_NOWAIT, m->m_type); 1977 *np = n; 1978 if (n == NULL) 1979 goto nospace; 1980 if (copyhdr) { 1981 if (!m_dup_pkthdr(n, m, M_NOWAIT)) 1982 goto nospace; 1983 if (len == M_COPYALL) 1984 n->m_pkthdr.len -= off0; 1985 else 1986 n->m_pkthdr.len = len; 1987 pkthdrlen = &n->m_pkthdr.len; 1988 copyhdr = false; 1989 } 1990 n->m_len = mlen; 1991 len_cp += n->m_len; 1992 if (m->m_flags & M_EXT) { 1993 n->m_data = m->m_data + off; 1994 mb_dupcl(n, m); 1995 } else 1996 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 1997 (u_int)n->m_len); 1998 1999 if (sb && (sb->sb_sndptr == m) && 2000 ((n->m_len + off) >= m->m_len) && m->m_next) { 2001 sb->sb_sndptroff += m->m_len; 2002 sb->sb_sndptr = m->m_next; 2003 } 2004 off = 0; 2005 if (len != M_COPYALL) { 2006 len -= n->m_len; 2007 } 2008 m = m->m_next; 2009 np = &n->m_next; 2010 } 2011 return (top); 2012 nospace: 2013 m_freem(top); 2014 return (NULL); 2015 } 2016 2017 void 2018 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin) 2019 { 2020 2021 /* 2022 * Automatic sizing of send socket buffer. Often the send buffer 2023 * size is not optimally adjusted to the actual network conditions 2024 * at hand (delay bandwidth product). Setting the buffer size too 2025 * small limits throughput on links with high bandwidth and high 2026 * delay (eg. trans-continental/oceanic links). Setting the 2027 * buffer size too big consumes too much real kernel memory, 2028 * especially with many connections on busy servers. 2029 * 2030 * The criteria to step up the send buffer one notch are: 2031 * 1. receive window of remote host is larger than send buffer 2032 * (with a fudge factor of 5/4th); 2033 * 2. send buffer is filled to 7/8th with data (so we actually 2034 * have data to make use of it); 2035 * 3. send buffer fill has not hit maximal automatic size; 2036 * 4. our send window (slow start and cogestion controlled) is 2037 * larger than sent but unacknowledged data in send buffer. 2038 * 2039 * The remote host receive window scaling factor may limit the 2040 * growing of the send buffer before it reaches its allowed 2041 * maximum. 2042 * 2043 * It scales directly with slow start or congestion window 2044 * and does at most one step per received ACK. This fast 2045 * scaling has the drawback of growing the send buffer beyond 2046 * what is strictly necessary to make full use of a given 2047 * delay*bandwidth product. However testing has shown this not 2048 * to be much of an problem. At worst we are trading wasting 2049 * of available bandwidth (the non-use of it) for wasting some 2050 * socket buffer memory. 2051 * 2052 * TODO: Shrink send buffer during idle periods together 2053 * with congestion window. Requires another timer. Has to 2054 * wait for upcoming tcp timer rewrite. 2055 * 2056 * XXXGL: should there be used sbused() or sbavail()? 2057 */ 2058 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 2059 int lowat; 2060 2061 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0; 2062 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat && 2063 sbused(&so->so_snd) >= 2064 (so->so_snd.sb_hiwat / 8 * 7) - lowat && 2065 sbused(&so->so_snd) < V_tcp_autosndbuf_max && 2066 sendwin >= (sbused(&so->so_snd) - 2067 (tp->snd_nxt - tp->snd_una))) { 2068 if (!sbreserve_locked(&so->so_snd, 2069 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 2070 V_tcp_autosndbuf_max), so, curthread)) 2071 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 2072 } 2073 } 2074 } 2075