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_tcpdebug.h" 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/domain.h> 45 #ifdef TCP_HHOOK 46 #include <sys/hhook.h> 47 #endif 48 #include <sys/kernel.h> 49 #include <sys/lock.h> 50 #include <sys/mbuf.h> 51 #include <sys/mutex.h> 52 #include <sys/protosw.h> 53 #include <sys/sdt.h> 54 #include <sys/socket.h> 55 #include <sys/socketvar.h> 56 #include <sys/sysctl.h> 57 58 #include <net/if.h> 59 #include <net/route.h> 60 #include <net/vnet.h> 61 62 #include <netinet/in.h> 63 #include <netinet/in_kdtrace.h> 64 #include <netinet/in_systm.h> 65 #include <netinet/ip.h> 66 #include <netinet/in_pcb.h> 67 #include <netinet/ip_var.h> 68 #include <netinet/ip_options.h> 69 #ifdef INET6 70 #include <netinet6/in6_pcb.h> 71 #include <netinet/ip6.h> 72 #include <netinet6/ip6_var.h> 73 #endif 74 #include <netinet/tcp.h> 75 #include <netinet/tcp_fsm.h> 76 #include <netinet/tcp_log_buf.h> 77 #include <netinet/tcp_seq.h> 78 #include <netinet/tcp_timer.h> 79 #include <netinet/tcp_var.h> 80 #include <netinet/tcpip.h> 81 #include <netinet/cc/cc.h> 82 #include <netinet/tcp_fastopen.h> 83 #ifdef TCPPCAP 84 #include <netinet/tcp_pcap.h> 85 #endif 86 #ifdef TCPDEBUG 87 #include <netinet/tcp_debug.h> 88 #endif 89 #ifdef TCP_OFFLOAD 90 #include <netinet/tcp_offload.h> 91 #endif 92 93 #include <netipsec/ipsec_support.h> 94 95 #include <machine/in_cksum.h> 96 97 #include <security/mac/mac_framework.h> 98 99 VNET_DEFINE(int, path_mtu_discovery) = 1; 100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW, 101 &VNET_NAME(path_mtu_discovery), 1, 102 "Enable Path MTU Discovery"); 103 104 VNET_DEFINE(int, tcp_do_tso) = 1; 105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW, 106 &VNET_NAME(tcp_do_tso), 0, 107 "Enable TCP Segmentation Offload"); 108 109 VNET_DEFINE(int, tcp_sendspace) = 1024*32; 110 #define V_tcp_sendspace VNET(tcp_sendspace) 111 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW, 112 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 113 114 VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, 116 &VNET_NAME(tcp_do_autosndbuf), 0, 117 "Enable automatic send buffer sizing"); 118 119 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW, 121 &VNET_NAME(tcp_autosndbuf_inc), 0, 122 "Incrementor step size of automatic send buffer"); 123 124 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; 125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW, 126 &VNET_NAME(tcp_autosndbuf_max), 0, 127 "Max size of automatic send buffer"); 128 129 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0; 130 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat) 131 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW, 132 &VNET_NAME(tcp_sendbuf_auto_lowat), 0, 133 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT"); 134 135 /* 136 * Make sure that either retransmit or persist timer is set for SYN, FIN and 137 * non-ACK. 138 */ 139 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \ 140 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\ 141 tcp_timer_active((tp), TT_REXMT) || \ 142 tcp_timer_active((tp), TT_PERSIST), \ 143 ("neither rexmt nor persist timer is set")) 144 145 static void inline cc_after_idle(struct tcpcb *tp); 146 147 #ifdef TCP_HHOOK 148 /* 149 * Wrapper for the TCP established output helper hook. 150 */ 151 void 152 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 153 struct tcpopt *to, uint32_t len, int tso) 154 { 155 struct tcp_hhook_data hhook_data; 156 157 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 158 hhook_data.tp = tp; 159 hhook_data.th = th; 160 hhook_data.to = to; 161 hhook_data.len = len; 162 hhook_data.tso = tso; 163 164 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 165 tp->osd); 166 } 167 } 168 #endif 169 170 /* 171 * Flags used when sending segments in tcp_output. Basic flags (TH_RST, 172 * TH_ACK,TH_SYN,TH_FIN) are totally determined by state, with the proviso 173 * that TH_FIN is sent only if all data queued for output is included in the 174 * segment. 175 */ 176 u_char tcp_outflags[TCP_NSTATES] = { 177 TH_RST|TH_ACK, /* 0, CLOSED */ 178 0, /* 1, LISTEN */ 179 TH_SYN, /* 2, SYN_SENT */ 180 TH_SYN|TH_ACK, /* 3, SYN_RECEIVED */ 181 TH_ACK, /* 4, ESTABLISHED */ 182 TH_ACK, /* 5, CLOSE_WAIT */ 183 TH_FIN|TH_ACK, /* 6, FIN_WAIT_1 */ 184 TH_FIN|TH_ACK, /* 7, CLOSING */ 185 TH_FIN|TH_ACK, /* 8, LAST_ACK */ 186 TH_ACK, /* 9, FIN_WAIT_2 */ 187 TH_ACK, /* 10, TIME_WAIT */ 188 }; 189 190 /* 191 * CC wrapper hook functions 192 */ 193 static void inline 194 cc_after_idle(struct tcpcb *tp) 195 { 196 INP_WLOCK_ASSERT(tp->t_inpcb); 197 198 if (CC_ALGO(tp)->after_idle != NULL) 199 CC_ALGO(tp)->after_idle(tp->ccv); 200 } 201 202 /* 203 * Tcp output routine: figure out what should be sent and send it. 204 */ 205 int 206 tcp_output(struct tcpcb *tp) 207 { 208 struct socket *so = tp->t_inpcb->inp_socket; 209 int32_t len; 210 uint32_t recwin, sendwin; 211 int off, flags, error = 0; /* Keep compiler happy */ 212 u_int if_hw_tsomaxsegcount = 0; 213 u_int if_hw_tsomaxsegsize; 214 struct mbuf *m; 215 struct ip *ip = NULL; 216 #ifdef TCPDEBUG 217 struct ipovly *ipov = NULL; 218 #endif 219 struct tcphdr *th; 220 u_char opt[TCP_MAXOLEN]; 221 unsigned ipoptlen, optlen, hdrlen; 222 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 223 unsigned ipsec_optlen = 0; 224 #endif 225 int idle, sendalot, curticks; 226 int sack_rxmit, sack_bytes_rxmt; 227 struct sackhole *p; 228 int tso, mtu; 229 struct tcpopt to; 230 unsigned int wanted_cookie = 0; 231 unsigned int dont_sendalot = 0; 232 #if 0 233 int maxburst = TCP_MAXBURST; 234 #endif 235 #ifdef INET6 236 struct ip6_hdr *ip6 = NULL; 237 int isipv6; 238 239 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 240 #endif 241 242 INP_WLOCK_ASSERT(tp->t_inpcb); 243 244 #ifdef TCP_OFFLOAD 245 if (tp->t_flags & TF_TOE) 246 return (tcp_offload_output(tp)); 247 #endif 248 249 /* 250 * For TFO connections in SYN_RECEIVED, only allow the initial 251 * SYN|ACK and those sent by the retransmit timer. 252 */ 253 if (IS_FASTOPEN(tp->t_flags) && 254 (tp->t_state == TCPS_SYN_RECEIVED) && 255 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */ 256 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 257 return (0); 258 259 /* 260 * Determine length of data that should be transmitted, 261 * and flags that will be used. 262 * If there is some data or critical controls (SYN, RST) 263 * to send, then transmit; otherwise, investigate further. 264 */ 265 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 266 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) 267 cc_after_idle(tp); 268 tp->t_flags &= ~TF_LASTIDLE; 269 if (idle) { 270 if (tp->t_flags & TF_MORETOCOME) { 271 tp->t_flags |= TF_LASTIDLE; 272 idle = 0; 273 } 274 } 275 again: 276 /* 277 * If we've recently taken a timeout, snd_max will be greater than 278 * snd_nxt. There may be SACK information that allows us to avoid 279 * resending already delivered data. Adjust snd_nxt accordingly. 280 */ 281 if ((tp->t_flags & TF_SACK_PERMIT) && 282 SEQ_LT(tp->snd_nxt, tp->snd_max)) 283 tcp_sack_adjust(tp); 284 sendalot = 0; 285 tso = 0; 286 mtu = 0; 287 off = tp->snd_nxt - tp->snd_una; 288 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 289 290 flags = tcp_outflags[tp->t_state]; 291 /* 292 * Send any SACK-generated retransmissions. If we're explicitly trying 293 * to send out new data (when sendalot is 1), bypass this function. 294 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 295 * we're replacing a (future) new transmission with a retransmission 296 * now, and we previously incremented snd_cwnd in tcp_input(). 297 */ 298 /* 299 * Still in sack recovery , reset rxmit flag to zero. 300 */ 301 sack_rxmit = 0; 302 sack_bytes_rxmt = 0; 303 len = 0; 304 p = NULL; 305 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 306 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 307 uint32_t cwin; 308 309 cwin = 310 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0); 311 /* Do not retransmit SACK segments beyond snd_recover */ 312 if (SEQ_GT(p->end, tp->snd_recover)) { 313 /* 314 * (At least) part of sack hole extends beyond 315 * snd_recover. Check to see if we can rexmit data 316 * for this hole. 317 */ 318 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 319 /* 320 * Can't rexmit any more data for this hole. 321 * That data will be rexmitted in the next 322 * sack recovery episode, when snd_recover 323 * moves past p->rxmit. 324 */ 325 p = NULL; 326 goto after_sack_rexmit; 327 } else 328 /* Can rexmit part of the current hole */ 329 len = ((int32_t)ulmin(cwin, 330 tp->snd_recover - p->rxmit)); 331 } else 332 len = ((int32_t)ulmin(cwin, p->end - p->rxmit)); 333 off = p->rxmit - tp->snd_una; 334 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 335 __func__, off)); 336 if (len > 0) { 337 sack_rxmit = 1; 338 sendalot = 1; 339 TCPSTAT_INC(tcps_sack_rexmits); 340 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 341 min(len, tp->t_maxseg)); 342 } 343 } 344 after_sack_rexmit: 345 /* 346 * Get standard flags, and add SYN or FIN if requested by 'hidden' 347 * state flags. 348 */ 349 if (tp->t_flags & TF_NEEDFIN) 350 flags |= TH_FIN; 351 if (tp->t_flags & TF_NEEDSYN) 352 flags |= TH_SYN; 353 354 SOCKBUF_LOCK(&so->so_snd); 355 /* 356 * If in persist timeout with window of 0, send 1 byte. 357 * Otherwise, if window is small but nonzero 358 * and timer expired, we will send what we can 359 * and go to transmit state. 360 */ 361 if (tp->t_flags & TF_FORCEDATA) { 362 if (sendwin == 0) { 363 /* 364 * If we still have some data to send, then 365 * clear the FIN bit. Usually this would 366 * happen below when it realizes that we 367 * aren't sending all the data. However, 368 * if we have exactly 1 byte of unsent data, 369 * then it won't clear the FIN bit below, 370 * and if we are in persist state, we wind 371 * up sending the packet without recording 372 * that we sent the FIN bit. 373 * 374 * We can't just blindly clear the FIN bit, 375 * because if we don't have any more data 376 * to send then the probe will be the FIN 377 * itself. 378 */ 379 if (off < sbused(&so->so_snd)) 380 flags &= ~TH_FIN; 381 sendwin = 1; 382 } else { 383 tcp_timer_activate(tp, TT_PERSIST, 0); 384 tp->t_rxtshift = 0; 385 } 386 } 387 388 /* 389 * If snd_nxt == snd_max and we have transmitted a FIN, the 390 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 391 * a negative length. This can also occur when TCP opens up 392 * its congestion window while receiving additional duplicate 393 * acks after fast-retransmit because TCP will reset snd_nxt 394 * to snd_max after the fast-retransmit. 395 * 396 * In the normal retransmit-FIN-only case, however, snd_nxt will 397 * be set to snd_una, the offset will be 0, and the length may 398 * wind up 0. 399 * 400 * If sack_rxmit is true we are retransmitting from the scoreboard 401 * in which case len is already set. 402 */ 403 if (sack_rxmit == 0) { 404 if (sack_bytes_rxmt == 0) 405 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) - 406 off); 407 else { 408 int32_t cwin; 409 410 /* 411 * We are inside of a SACK recovery episode and are 412 * sending new data, having retransmitted all the 413 * data possible in the scoreboard. 414 */ 415 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) - 416 off); 417 /* 418 * Don't remove this (len > 0) check ! 419 * We explicitly check for len > 0 here (although it 420 * isn't really necessary), to work around a gcc 421 * optimization issue - to force gcc to compute 422 * len above. Without this check, the computation 423 * of len is bungled by the optimizer. 424 */ 425 if (len > 0) { 426 cwin = tp->snd_cwnd - 427 (tp->snd_nxt - tp->sack_newdata) - 428 sack_bytes_rxmt; 429 if (cwin < 0) 430 cwin = 0; 431 len = imin(len, cwin); 432 } 433 } 434 } 435 436 /* 437 * Lop off SYN bit if it has already been sent. However, if this 438 * is SYN-SENT state and if segment contains data and if we don't 439 * know that foreign host supports TAO, suppress sending segment. 440 */ 441 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 442 if (tp->t_state != TCPS_SYN_RECEIVED) 443 flags &= ~TH_SYN; 444 /* 445 * When sending additional segments following a TFO SYN|ACK, 446 * do not include the SYN bit. 447 */ 448 if (IS_FASTOPEN(tp->t_flags) && 449 (tp->t_state == TCPS_SYN_RECEIVED)) 450 flags &= ~TH_SYN; 451 off--, len++; 452 } 453 454 /* 455 * Be careful not to send data and/or FIN on SYN segments. 456 * This measure is needed to prevent interoperability problems 457 * with not fully conformant TCP implementations. 458 */ 459 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 460 len = 0; 461 flags &= ~TH_FIN; 462 } 463 464 /* 465 * On TFO sockets, ensure no data is sent in the following cases: 466 * 467 * - When retransmitting SYN|ACK on a passively-created socket 468 * 469 * - When retransmitting SYN on an actively created socket 470 * 471 * - When sending a zero-length cookie (cookie request) on an 472 * actively created socket 473 * 474 * - When the socket is in the CLOSED state (RST is being sent) 475 */ 476 if (IS_FASTOPEN(tp->t_flags) && 477 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 478 ((tp->t_state == TCPS_SYN_SENT) && 479 (tp->t_tfo_client_cookie_len == 0)) || 480 (flags & TH_RST))) 481 len = 0; 482 if (len <= 0) { 483 /* 484 * If FIN has been sent but not acked, 485 * but we haven't been called to retransmit, 486 * len will be < 0. Otherwise, window shrank 487 * after we sent into it. If window shrank to 0, 488 * cancel pending retransmit, pull snd_nxt back 489 * to (closed) window, and set the persist timer 490 * if it isn't already going. If the window didn't 491 * close completely, just wait for an ACK. 492 * 493 * We also do a general check here to ensure that 494 * we will set the persist timer when we have data 495 * to send, but a 0-byte window. This makes sure 496 * the persist timer is set even if the packet 497 * hits one of the "goto send" lines below. 498 */ 499 len = 0; 500 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 501 (off < (int) sbavail(&so->so_snd))) { 502 tcp_timer_activate(tp, TT_REXMT, 0); 503 tp->t_rxtshift = 0; 504 tp->snd_nxt = tp->snd_una; 505 if (!tcp_timer_active(tp, TT_PERSIST)) 506 tcp_setpersist(tp); 507 } 508 } 509 510 /* len will be >= 0 after this point. */ 511 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 512 513 tcp_sndbuf_autoscale(tp, so, sendwin); 514 515 /* 516 * Decide if we can use TCP Segmentation Offloading (if supported by 517 * hardware). 518 * 519 * TSO may only be used if we are in a pure bulk sending state. The 520 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 521 * IP options prevent using TSO. With TSO the TCP header is the same 522 * (except for the sequence number) for all generated packets. This 523 * makes it impossible to transmit any options which vary per generated 524 * segment or packet. 525 * 526 * IPv4 handling has a clear separation of ip options and ip header 527 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 528 * the right thing below to provide length of just ip options and thus 529 * checking for ipoptlen is enough to decide if ip options are present. 530 */ 531 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 532 /* 533 * Pre-calculate here as we save another lookup into the darknesses 534 * of IPsec that way and can actually decide if TSO is ok. 535 */ 536 #ifdef INET6 537 if (isipv6 && IPSEC_ENABLED(ipv6)) 538 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); 539 #ifdef INET 540 else 541 #endif 542 #endif /* INET6 */ 543 #ifdef INET 544 if (IPSEC_ENABLED(ipv4)) 545 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); 546 #endif /* INET */ 547 #endif /* IPSEC */ 548 #ifdef INET6 549 if (isipv6) 550 ipoptlen = ip6_optlen(tp->t_inpcb); 551 else 552 #endif 553 if (tp->t_inpcb->inp_options) 554 ipoptlen = tp->t_inpcb->inp_options->m_len - 555 offsetof(struct ipoption, ipopt_list); 556 else 557 ipoptlen = 0; 558 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 559 ipoptlen += ipsec_optlen; 560 #endif 561 562 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 563 ((tp->t_flags & TF_SIGNATURE) == 0) && 564 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 565 ipoptlen == 0 && !(flags & TH_SYN)) 566 tso = 1; 567 568 if (sack_rxmit) { 569 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd))) 570 flags &= ~TH_FIN; 571 } else { 572 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 573 sbused(&so->so_snd))) 574 flags &= ~TH_FIN; 575 } 576 577 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 578 (long)TCP_MAXWIN << tp->rcv_scale); 579 580 /* 581 * Sender silly window avoidance. We transmit under the following 582 * conditions when len is non-zero: 583 * 584 * - We have a full segment (or more with TSO) 585 * - This is the last buffer in a write()/send() and we are 586 * either idle or running NODELAY 587 * - we've timed out (e.g. persist timer) 588 * - we have more then 1/2 the maximum send window's worth of 589 * data (receiver may be limited the window size) 590 * - we need to retransmit 591 */ 592 if (len) { 593 if (len >= tp->t_maxseg) 594 goto send; 595 /* 596 * NOTE! on localhost connections an 'ack' from the remote 597 * end may occur synchronously with the output and cause 598 * us to flush a buffer queued with moretocome. XXX 599 * 600 * note: the len + off check is almost certainly unnecessary. 601 */ 602 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 603 (idle || (tp->t_flags & TF_NODELAY)) && 604 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) && 605 (tp->t_flags & TF_NOPUSH) == 0) { 606 goto send; 607 } 608 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 609 goto send; 610 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 611 goto send; 612 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 613 goto send; 614 if (sack_rxmit) 615 goto send; 616 } 617 618 /* 619 * Sending of standalone window updates. 620 * 621 * Window updates are important when we close our window due to a 622 * full socket buffer and are opening it again after the application 623 * reads data from it. Once the window has opened again and the 624 * remote end starts to send again the ACK clock takes over and 625 * provides the most current window information. 626 * 627 * We must avoid the silly window syndrome whereas every read 628 * from the receive buffer, no matter how small, causes a window 629 * update to be sent. We also should avoid sending a flurry of 630 * window updates when the socket buffer had queued a lot of data 631 * and the application is doing small reads. 632 * 633 * Prevent a flurry of pointless window updates by only sending 634 * an update when we can increase the advertized window by more 635 * than 1/4th of the socket buffer capacity. When the buffer is 636 * getting full or is very small be more aggressive and send an 637 * update whenever we can increase by two mss sized segments. 638 * In all other situations the ACK's to new incoming data will 639 * carry further window increases. 640 * 641 * Don't send an independent window update if a delayed 642 * ACK is pending (it will get piggy-backed on it) or the 643 * remote side already has done a half-close and won't send 644 * more data. Skip this if the connection is in T/TCP 645 * half-open state. 646 */ 647 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 648 !(tp->t_flags & TF_DELACK) && 649 !TCPS_HAVERCVDFIN(tp->t_state)) { 650 /* 651 * "adv" is the amount we could increase the window, 652 * taking into account that we are limited by 653 * TCP_MAXWIN << tp->rcv_scale. 654 */ 655 int32_t adv; 656 int oldwin; 657 658 adv = recwin; 659 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 660 oldwin = (tp->rcv_adv - tp->rcv_nxt); 661 adv -= oldwin; 662 } else 663 oldwin = 0; 664 665 /* 666 * If the new window size ends up being the same as or less 667 * than the old size when it is scaled, then don't force 668 * a window update. 669 */ 670 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 671 goto dontupdate; 672 673 if (adv >= (int32_t)(2 * tp->t_maxseg) && 674 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 675 recwin <= (so->so_rcv.sb_hiwat / 8) || 676 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 677 goto send; 678 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat) 679 goto send; 680 } 681 dontupdate: 682 683 /* 684 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 685 * is also a catch-all for the retransmit timer timeout case. 686 */ 687 if (tp->t_flags & TF_ACKNOW) 688 goto send; 689 if ((flags & TH_RST) || 690 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 691 goto send; 692 if (SEQ_GT(tp->snd_up, tp->snd_una)) 693 goto send; 694 /* 695 * If our state indicates that FIN should be sent 696 * and we have not yet done so, then we need to send. 697 */ 698 if (flags & TH_FIN && 699 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 700 goto send; 701 /* 702 * In SACK, it is possible for tcp_output to fail to send a segment 703 * after the retransmission timer has been turned off. Make sure 704 * that the retransmission timer is set. 705 */ 706 if ((tp->t_flags & TF_SACK_PERMIT) && 707 SEQ_GT(tp->snd_max, tp->snd_una) && 708 !tcp_timer_active(tp, TT_REXMT) && 709 !tcp_timer_active(tp, TT_PERSIST)) { 710 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 711 goto just_return; 712 } 713 /* 714 * TCP window updates are not reliable, rather a polling protocol 715 * using ``persist'' packets is used to insure receipt of window 716 * updates. The three ``states'' for the output side are: 717 * idle not doing retransmits or persists 718 * persisting to move a small or zero window 719 * (re)transmitting and thereby not persisting 720 * 721 * tcp_timer_active(tp, TT_PERSIST) 722 * is true when we are in persist state. 723 * (tp->t_flags & TF_FORCEDATA) 724 * is set when we are called to send a persist packet. 725 * tcp_timer_active(tp, TT_REXMT) 726 * is set when we are retransmitting 727 * The output side is idle when both timers are zero. 728 * 729 * If send window is too small, there is data to transmit, and no 730 * retransmit or persist is pending, then go to persist state. 731 * If nothing happens soon, send when timer expires: 732 * if window is nonzero, transmit what we can, 733 * otherwise force out a byte. 734 */ 735 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && 736 !tcp_timer_active(tp, TT_PERSIST)) { 737 tp->t_rxtshift = 0; 738 tcp_setpersist(tp); 739 } 740 741 /* 742 * No reason to send a segment, just return. 743 */ 744 just_return: 745 SOCKBUF_UNLOCK(&so->so_snd); 746 return (0); 747 748 send: 749 SOCKBUF_LOCK_ASSERT(&so->so_snd); 750 if (len > 0) { 751 if (len >= tp->t_maxseg) 752 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 753 else 754 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 755 } 756 /* 757 * Before ESTABLISHED, force sending of initial options 758 * unless TCP set not to do any options. 759 * NOTE: we assume that the IP/TCP header plus TCP options 760 * always fit in a single mbuf, leaving room for a maximum 761 * link header, i.e. 762 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 763 */ 764 optlen = 0; 765 #ifdef INET6 766 if (isipv6) 767 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 768 else 769 #endif 770 hdrlen = sizeof (struct tcpiphdr); 771 772 /* 773 * Compute options for segment. 774 * We only have to care about SYN and established connection 775 * segments. Options for SYN-ACK segments are handled in TCP 776 * syncache. 777 */ 778 to.to_flags = 0; 779 if ((tp->t_flags & TF_NOOPT) == 0) { 780 /* Maximum segment size. */ 781 if (flags & TH_SYN) { 782 tp->snd_nxt = tp->iss; 783 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 784 to.to_flags |= TOF_MSS; 785 786 /* 787 * On SYN or SYN|ACK transmits on TFO connections, 788 * only include the TFO option if it is not a 789 * retransmit, as the presence of the TFO option may 790 * have caused the original SYN or SYN|ACK to have 791 * been dropped by a middlebox. 792 */ 793 if (IS_FASTOPEN(tp->t_flags) && 794 (tp->t_rxtshift == 0)) { 795 if (tp->t_state == TCPS_SYN_RECEIVED) { 796 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 797 to.to_tfo_cookie = 798 (u_int8_t *)&tp->t_tfo_cookie.server; 799 to.to_flags |= TOF_FASTOPEN; 800 wanted_cookie = 1; 801 } else if (tp->t_state == TCPS_SYN_SENT) { 802 to.to_tfo_len = 803 tp->t_tfo_client_cookie_len; 804 to.to_tfo_cookie = 805 tp->t_tfo_cookie.client; 806 to.to_flags |= TOF_FASTOPEN; 807 wanted_cookie = 1; 808 /* 809 * If we wind up having more data to 810 * send with the SYN than can fit in 811 * one segment, don't send any more 812 * until the SYN|ACK comes back from 813 * the other end. 814 */ 815 dont_sendalot = 1; 816 } 817 } 818 } 819 /* Window scaling. */ 820 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 821 to.to_wscale = tp->request_r_scale; 822 to.to_flags |= TOF_SCALE; 823 } 824 /* Timestamps. */ 825 if ((tp->t_flags & TF_RCVD_TSTMP) || 826 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 827 curticks = tcp_ts_getticks(); 828 to.to_tsval = curticks + tp->ts_offset; 829 to.to_tsecr = tp->ts_recent; 830 to.to_flags |= TOF_TS; 831 if (tp->t_rxtshift == 1) 832 tp->t_badrxtwin = curticks; 833 } 834 835 /* Set receive buffer autosizing timestamp. */ 836 if (tp->rfbuf_ts == 0 && 837 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 838 tp->rfbuf_ts = tcp_ts_getticks(); 839 840 /* Selective ACK's. */ 841 if (tp->t_flags & TF_SACK_PERMIT) { 842 if (flags & TH_SYN) 843 to.to_flags |= TOF_SACKPERM; 844 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 845 (tp->t_flags & TF_SACK_PERMIT) && 846 tp->rcv_numsacks > 0) { 847 to.to_flags |= TOF_SACK; 848 to.to_nsacks = tp->rcv_numsacks; 849 to.to_sacks = (u_char *)tp->sackblks; 850 } 851 } 852 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 853 /* TCP-MD5 (RFC2385). */ 854 /* 855 * Check that TCP_MD5SIG is enabled in tcpcb to 856 * account the size needed to set this TCP option. 857 */ 858 if (tp->t_flags & TF_SIGNATURE) 859 to.to_flags |= TOF_SIGNATURE; 860 #endif /* TCP_SIGNATURE */ 861 862 /* Processing the options. */ 863 hdrlen += optlen = tcp_addoptions(&to, opt); 864 /* 865 * If we wanted a TFO option to be added, but it was unable 866 * to fit, ensure no data is sent. 867 */ 868 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 869 !(to.to_flags & TOF_FASTOPEN)) 870 len = 0; 871 } 872 873 /* 874 * Adjust data length if insertion of options will 875 * bump the packet length beyond the t_maxseg length. 876 * Clear the FIN bit because we cut off the tail of 877 * the segment. 878 */ 879 if (len + optlen + ipoptlen > tp->t_maxseg) { 880 flags &= ~TH_FIN; 881 882 if (tso) { 883 u_int if_hw_tsomax; 884 u_int moff; 885 int max_len; 886 887 /* extract TSO information */ 888 if_hw_tsomax = tp->t_tsomax; 889 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 890 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 891 892 /* 893 * Limit a TSO burst to prevent it from 894 * overflowing or exceeding the maximum length 895 * allowed by the network interface: 896 */ 897 KASSERT(ipoptlen == 0, 898 ("%s: TSO can't do IP options", __func__)); 899 900 /* 901 * Check if we should limit by maximum payload 902 * length: 903 */ 904 if (if_hw_tsomax != 0) { 905 /* compute maximum TSO length */ 906 max_len = (if_hw_tsomax - hdrlen - 907 max_linkhdr); 908 if (max_len <= 0) { 909 len = 0; 910 } else if (len > max_len) { 911 sendalot = 1; 912 len = max_len; 913 } 914 } 915 /* 916 * Prevent the last segment from being 917 * fractional unless the send sockbuf can be 918 * emptied: 919 */ 920 max_len = (tp->t_maxseg - optlen); 921 if (((uint32_t)off + (uint32_t)len) < 922 sbavail(&so->so_snd)) { 923 moff = len % max_len; 924 if (moff != 0) { 925 len -= moff; 926 sendalot = 1; 927 } 928 } 929 930 /* 931 * In case there are too many small fragments 932 * don't use TSO: 933 */ 934 if (len <= max_len) { 935 len = max_len; 936 sendalot = 1; 937 tso = 0; 938 } 939 940 /* 941 * Send the FIN in a separate segment 942 * after the bulk sending is done. 943 * We don't trust the TSO implementations 944 * to clear the FIN flag on all but the 945 * last segment. 946 */ 947 if (tp->t_flags & TF_NEEDFIN) 948 sendalot = 1; 949 } else { 950 len = tp->t_maxseg - optlen - ipoptlen; 951 sendalot = 1; 952 if (dont_sendalot) 953 sendalot = 0; 954 } 955 } else 956 tso = 0; 957 958 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 959 ("%s: len > IP_MAXPACKET", __func__)); 960 961 /*#ifdef DIAGNOSTIC*/ 962 #ifdef INET6 963 if (max_linkhdr + hdrlen > MCLBYTES) 964 #else 965 if (max_linkhdr + hdrlen > MHLEN) 966 #endif 967 panic("tcphdr too big"); 968 /*#endif*/ 969 970 /* 971 * This KASSERT is here to catch edge cases at a well defined place. 972 * Before, those had triggered (random) panic conditions further down. 973 */ 974 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 975 976 /* 977 * Grab a header mbuf, attaching a copy of data to 978 * be transmitted, and initialize the header from 979 * the template for sends on this connection. 980 */ 981 if (len) { 982 struct mbuf *mb; 983 struct sockbuf *msb; 984 u_int moff; 985 986 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 987 TCPSTAT_INC(tcps_sndprobe); 988 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 989 tp->t_sndrexmitpack++; 990 TCPSTAT_INC(tcps_sndrexmitpack); 991 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 992 } else { 993 TCPSTAT_INC(tcps_sndpack); 994 TCPSTAT_ADD(tcps_sndbyte, len); 995 } 996 #ifdef INET6 997 if (MHLEN < hdrlen + max_linkhdr) 998 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 999 else 1000 #endif 1001 m = m_gethdr(M_NOWAIT, MT_DATA); 1002 1003 if (m == NULL) { 1004 SOCKBUF_UNLOCK(&so->so_snd); 1005 error = ENOBUFS; 1006 sack_rxmit = 0; 1007 goto out; 1008 } 1009 1010 m->m_data += max_linkhdr; 1011 m->m_len = hdrlen; 1012 1013 /* 1014 * Start the m_copy functions from the closest mbuf 1015 * to the offset in the socket buffer chain. 1016 */ 1017 mb = sbsndptr_noadv(&so->so_snd, off, &moff); 1018 if (len <= MHLEN - hdrlen - max_linkhdr) { 1019 m_copydata(mb, moff, len, 1020 mtod(m, caddr_t) + hdrlen); 1021 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1022 sbsndptr_adv(&so->so_snd, mb, len); 1023 m->m_len += len; 1024 } else { 1025 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 1026 msb = NULL; 1027 else 1028 msb = &so->so_snd; 1029 m->m_next = tcp_m_copym(mb, moff, 1030 &len, if_hw_tsomaxsegcount, 1031 if_hw_tsomaxsegsize, msb); 1032 if (len <= (tp->t_maxseg - optlen)) { 1033 /* 1034 * Must have ran out of mbufs for the copy 1035 * shorten it to no longer need tso. Lets 1036 * not put on sendalot since we are low on 1037 * mbufs. 1038 */ 1039 tso = 0; 1040 } 1041 if (m->m_next == NULL) { 1042 SOCKBUF_UNLOCK(&so->so_snd); 1043 (void) m_free(m); 1044 error = ENOBUFS; 1045 sack_rxmit = 0; 1046 goto out; 1047 } 1048 } 1049 1050 /* 1051 * If we're sending everything we've got, set PUSH. 1052 * (This will keep happy those implementations which only 1053 * give data to the user when a buffer fills or 1054 * a PUSH comes in.) 1055 */ 1056 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) && 1057 !(flags & TH_SYN)) 1058 flags |= TH_PUSH; 1059 SOCKBUF_UNLOCK(&so->so_snd); 1060 } else { 1061 SOCKBUF_UNLOCK(&so->so_snd); 1062 if (tp->t_flags & TF_ACKNOW) 1063 TCPSTAT_INC(tcps_sndacks); 1064 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1065 TCPSTAT_INC(tcps_sndctrl); 1066 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1067 TCPSTAT_INC(tcps_sndurg); 1068 else 1069 TCPSTAT_INC(tcps_sndwinup); 1070 1071 m = m_gethdr(M_NOWAIT, MT_DATA); 1072 if (m == NULL) { 1073 error = ENOBUFS; 1074 sack_rxmit = 0; 1075 goto out; 1076 } 1077 #ifdef INET6 1078 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1079 MHLEN >= hdrlen) { 1080 M_ALIGN(m, hdrlen); 1081 } else 1082 #endif 1083 m->m_data += max_linkhdr; 1084 m->m_len = hdrlen; 1085 } 1086 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1087 m->m_pkthdr.rcvif = (struct ifnet *)0; 1088 #ifdef MAC 1089 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1090 #endif 1091 #ifdef INET6 1092 if (isipv6) { 1093 ip6 = mtod(m, struct ip6_hdr *); 1094 th = (struct tcphdr *)(ip6 + 1); 1095 tcpip_fillheaders(tp->t_inpcb, ip6, th); 1096 } else 1097 #endif /* INET6 */ 1098 { 1099 ip = mtod(m, struct ip *); 1100 #ifdef TCPDEBUG 1101 ipov = (struct ipovly *)ip; 1102 #endif 1103 th = (struct tcphdr *)(ip + 1); 1104 tcpip_fillheaders(tp->t_inpcb, ip, th); 1105 } 1106 1107 /* 1108 * Fill in fields, remembering maximum advertised 1109 * window for use in delaying messages about window sizes. 1110 * If resending a FIN, be sure not to use a new sequence number. 1111 */ 1112 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1113 tp->snd_nxt == tp->snd_max) 1114 tp->snd_nxt--; 1115 /* 1116 * If we are starting a connection, send ECN setup 1117 * SYN packet. If we are on a retransmit, we may 1118 * resend those bits a number of times as per 1119 * RFC 3168. 1120 */ 1121 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { 1122 if (tp->t_rxtshift >= 1) { 1123 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 1124 flags |= TH_ECE|TH_CWR; 1125 } else 1126 flags |= TH_ECE|TH_CWR; 1127 } 1128 1129 if (tp->t_state == TCPS_ESTABLISHED && 1130 (tp->t_flags & TF_ECN_PERMIT)) { 1131 /* 1132 * If the peer has ECN, mark data packets with 1133 * ECN capable transmission (ECT). 1134 * Ignore pure ack packets, retransmissions and window probes. 1135 */ 1136 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1137 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 1138 #ifdef INET6 1139 if (isipv6) 1140 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1141 else 1142 #endif 1143 ip->ip_tos |= IPTOS_ECN_ECT0; 1144 TCPSTAT_INC(tcps_ecn_ect0); 1145 } 1146 1147 /* 1148 * Reply with proper ECN notifications. 1149 */ 1150 if (tp->t_flags & TF_ECN_SND_CWR) { 1151 flags |= TH_CWR; 1152 tp->t_flags &= ~TF_ECN_SND_CWR; 1153 } 1154 if (tp->t_flags & TF_ECN_SND_ECE) 1155 flags |= TH_ECE; 1156 } 1157 1158 /* 1159 * If we are doing retransmissions, then snd_nxt will 1160 * not reflect the first unsent octet. For ACK only 1161 * packets, we do not want the sequence number of the 1162 * retransmitted packet, we want the sequence number 1163 * of the next unsent octet. So, if there is no data 1164 * (and no SYN or FIN), use snd_max instead of snd_nxt 1165 * when filling in ti_seq. But if we are in persist 1166 * state, snd_max might reflect one byte beyond the 1167 * right edge of the window, so use snd_nxt in that 1168 * case, since we know we aren't doing a retransmission. 1169 * (retransmit and persist are mutually exclusive...) 1170 */ 1171 if (sack_rxmit == 0) { 1172 if (len || (flags & (TH_SYN|TH_FIN)) || 1173 tcp_timer_active(tp, TT_PERSIST)) 1174 th->th_seq = htonl(tp->snd_nxt); 1175 else 1176 th->th_seq = htonl(tp->snd_max); 1177 } else { 1178 th->th_seq = htonl(p->rxmit); 1179 p->rxmit += len; 1180 tp->sackhint.sack_bytes_rexmit += len; 1181 } 1182 th->th_ack = htonl(tp->rcv_nxt); 1183 if (optlen) { 1184 bcopy(opt, th + 1, optlen); 1185 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1186 } 1187 th->th_flags = flags; 1188 /* 1189 * Calculate receive window. Don't shrink window, 1190 * but avoid silly window syndrome. 1191 */ 1192 if (recwin < (so->so_rcv.sb_hiwat / 4) && 1193 recwin < tp->t_maxseg) 1194 recwin = 0; 1195 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1196 recwin < (tp->rcv_adv - tp->rcv_nxt)) 1197 recwin = (tp->rcv_adv - tp->rcv_nxt); 1198 1199 /* 1200 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1201 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1202 * case is handled in syncache. 1203 */ 1204 if (flags & TH_SYN) 1205 th->th_win = htons((u_short) 1206 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1207 else 1208 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1209 1210 /* 1211 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1212 * a 0 window. This may cause the remote transmitter to stall. This 1213 * flag tells soreceive() to disable delayed acknowledgements when 1214 * draining the buffer. This can occur if the receiver is attempting 1215 * to read more data than can be buffered prior to transmitting on 1216 * the connection. 1217 */ 1218 if (th->th_win == 0) { 1219 tp->t_sndzerowin++; 1220 tp->t_flags |= TF_RXWIN0SENT; 1221 } else 1222 tp->t_flags &= ~TF_RXWIN0SENT; 1223 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1224 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1225 th->th_flags |= TH_URG; 1226 } else 1227 /* 1228 * If no urgent pointer to send, then we pull 1229 * the urgent pointer to the left edge of the send window 1230 * so that it doesn't drift into the send window on sequence 1231 * number wraparound. 1232 */ 1233 tp->snd_up = tp->snd_una; /* drag it along */ 1234 1235 /* 1236 * Put TCP length in extended header, and then 1237 * checksum extended header and data. 1238 */ 1239 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1240 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1241 1242 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1243 if (to.to_flags & TOF_SIGNATURE) { 1244 /* 1245 * Calculate MD5 signature and put it into the place 1246 * determined before. 1247 * NOTE: since TCP options buffer doesn't point into 1248 * mbuf's data, calculate offset and use it. 1249 */ 1250 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th, 1251 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) { 1252 /* 1253 * Do not send segment if the calculation of MD5 1254 * digest has failed. 1255 */ 1256 m_freem(m); 1257 goto out; 1258 } 1259 } 1260 #endif 1261 #ifdef INET6 1262 if (isipv6) { 1263 /* 1264 * There is no need to fill in ip6_plen right now. 1265 * It will be filled later by ip6_output. 1266 */ 1267 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1268 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1269 optlen + len, IPPROTO_TCP, 0); 1270 } 1271 #endif 1272 #if defined(INET6) && defined(INET) 1273 else 1274 #endif 1275 #ifdef INET 1276 { 1277 m->m_pkthdr.csum_flags = CSUM_TCP; 1278 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1279 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1280 1281 /* IP version must be set here for ipv4/ipv6 checking later */ 1282 KASSERT(ip->ip_v == IPVERSION, 1283 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1284 } 1285 #endif 1286 1287 /* 1288 * Enable TSO and specify the size of the segments. 1289 * The TCP pseudo header checksum is always provided. 1290 */ 1291 if (tso) { 1292 KASSERT(len > tp->t_maxseg - optlen, 1293 ("%s: len <= tso_segsz", __func__)); 1294 m->m_pkthdr.csum_flags |= CSUM_TSO; 1295 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 1296 } 1297 1298 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 1299 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1300 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u", 1301 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1302 #else 1303 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1304 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u", 1305 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1306 #endif 1307 1308 #ifdef TCP_HHOOK 1309 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1310 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1311 #endif 1312 1313 #ifdef TCPDEBUG 1314 /* 1315 * Trace. 1316 */ 1317 if (so->so_options & SO_DEBUG) { 1318 u_short save = 0; 1319 #ifdef INET6 1320 if (!isipv6) 1321 #endif 1322 { 1323 save = ipov->ih_len; 1324 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1325 } 1326 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1327 #ifdef INET6 1328 if (!isipv6) 1329 #endif 1330 ipov->ih_len = save; 1331 } 1332 #endif /* TCPDEBUG */ 1333 TCP_PROBE3(debug__output, tp, th, m); 1334 1335 /* We're getting ready to send; log now. */ 1336 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 1337 len, NULL, false); 1338 1339 /* 1340 * Fill in IP length and desired time to live and 1341 * send to IP level. There should be a better way 1342 * to handle ttl and tos; we could keep them in 1343 * the template, but need a way to checksum without them. 1344 */ 1345 /* 1346 * m->m_pkthdr.len should have been set before checksum calculation, 1347 * because in6_cksum() need it. 1348 */ 1349 #ifdef INET6 1350 if (isipv6) { 1351 /* 1352 * we separately set hoplimit for every segment, since the 1353 * user might want to change the value via setsockopt. 1354 * Also, desired default hop limit might be changed via 1355 * Neighbor Discovery. 1356 */ 1357 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1358 1359 /* 1360 * Set the packet size here for the benefit of DTrace probes. 1361 * ip6_output() will set it properly; it's supposed to include 1362 * the option header lengths as well. 1363 */ 1364 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1365 1366 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 1367 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1368 else 1369 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1370 1371 if (tp->t_state == TCPS_SYN_SENT) 1372 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1373 1374 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1375 1376 #ifdef TCPPCAP 1377 /* Save packet, if requested. */ 1378 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1379 #endif 1380 1381 /* TODO: IPv6 IP6TOS_ECT bit on */ 1382 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, 1383 &tp->t_inpcb->inp_route6, 1384 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1385 NULL, NULL, tp->t_inpcb); 1386 1387 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL) 1388 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu; 1389 } 1390 #endif /* INET6 */ 1391 #if defined(INET) && defined(INET6) 1392 else 1393 #endif 1394 #ifdef INET 1395 { 1396 ip->ip_len = htons(m->m_pkthdr.len); 1397 #ifdef INET6 1398 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1399 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1400 #endif /* INET6 */ 1401 /* 1402 * If we do path MTU discovery, then we set DF on every packet. 1403 * This might not be the best thing to do according to RFC3390 1404 * Section 2. However the tcp hostcache migitates the problem 1405 * so it affects only the first tcp connection with a host. 1406 * 1407 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1408 */ 1409 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 1410 ip->ip_off |= htons(IP_DF); 1411 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1412 } else { 1413 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1414 } 1415 1416 if (tp->t_state == TCPS_SYN_SENT) 1417 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1418 1419 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1420 1421 #ifdef TCPPCAP 1422 /* Save packet, if requested. */ 1423 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1424 #endif 1425 1426 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route, 1427 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1428 tp->t_inpcb); 1429 1430 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL) 1431 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu; 1432 } 1433 #endif /* INET */ 1434 1435 out: 1436 /* 1437 * In transmit state, time the transmission and arrange for 1438 * the retransmit. In persist state, just set snd_max. 1439 */ 1440 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1441 !tcp_timer_active(tp, TT_PERSIST)) { 1442 tcp_seq startseq = tp->snd_nxt; 1443 1444 /* 1445 * Advance snd_nxt over sequence space of this segment. 1446 */ 1447 if (flags & (TH_SYN|TH_FIN)) { 1448 if (flags & TH_SYN) 1449 tp->snd_nxt++; 1450 if (flags & TH_FIN) { 1451 tp->snd_nxt++; 1452 tp->t_flags |= TF_SENTFIN; 1453 } 1454 } 1455 if (sack_rxmit) 1456 goto timer; 1457 tp->snd_nxt += len; 1458 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1459 tp->snd_max = tp->snd_nxt; 1460 /* 1461 * Time this transmission if not a retransmission and 1462 * not currently timing anything. 1463 */ 1464 if (tp->t_rtttime == 0) { 1465 tp->t_rtttime = ticks; 1466 tp->t_rtseq = startseq; 1467 TCPSTAT_INC(tcps_segstimed); 1468 } 1469 } 1470 1471 /* 1472 * Set retransmit timer if not currently set, 1473 * and not doing a pure ack or a keep-alive probe. 1474 * Initial value for retransmit timer is smoothed 1475 * round-trip time + 2 * round-trip time variance. 1476 * Initialize shift counter which is used for backoff 1477 * of retransmit time. 1478 */ 1479 timer: 1480 if (!tcp_timer_active(tp, TT_REXMT) && 1481 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1482 (tp->snd_nxt != tp->snd_una))) { 1483 if (tcp_timer_active(tp, TT_PERSIST)) { 1484 tcp_timer_activate(tp, TT_PERSIST, 0); 1485 tp->t_rxtshift = 0; 1486 } 1487 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1488 } else if (len == 0 && sbavail(&so->so_snd) && 1489 !tcp_timer_active(tp, TT_REXMT) && 1490 !tcp_timer_active(tp, TT_PERSIST)) { 1491 /* 1492 * Avoid a situation where we do not set persist timer 1493 * after a zero window condition. For example: 1494 * 1) A -> B: packet with enough data to fill the window 1495 * 2) B -> A: ACK for #1 + new data (0 window 1496 * advertisement) 1497 * 3) A -> B: ACK for #2, 0 len packet 1498 * 1499 * In this case, A will not activate the persist timer, 1500 * because it chose to send a packet. Unless tcp_output 1501 * is called for some other reason (delayed ack timer, 1502 * another input packet from B, socket syscall), A will 1503 * not send zero window probes. 1504 * 1505 * So, if you send a 0-length packet, but there is data 1506 * in the socket buffer, and neither the rexmt or 1507 * persist timer is already set, then activate the 1508 * persist timer. 1509 */ 1510 tp->t_rxtshift = 0; 1511 tcp_setpersist(tp); 1512 } 1513 } else { 1514 /* 1515 * Persist case, update snd_max but since we are in 1516 * persist mode (no window) we do not update snd_nxt. 1517 */ 1518 int xlen = len; 1519 if (flags & TH_SYN) 1520 ++xlen; 1521 if (flags & TH_FIN) { 1522 ++xlen; 1523 tp->t_flags |= TF_SENTFIN; 1524 } 1525 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1526 tp->snd_max = tp->snd_nxt + xlen; 1527 } 1528 1529 if (error) { 1530 /* Record the error. */ 1531 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, 1532 error, 0, NULL, false); 1533 1534 /* 1535 * We know that the packet was lost, so back out the 1536 * sequence number advance, if any. 1537 * 1538 * If the error is EPERM the packet got blocked by the 1539 * local firewall. Normally we should terminate the 1540 * connection but the blocking may have been spurious 1541 * due to a firewall reconfiguration cycle. So we treat 1542 * it like a packet loss and let the retransmit timer and 1543 * timeouts do their work over time. 1544 * XXX: It is a POLA question whether calling tcp_drop right 1545 * away would be the really correct behavior instead. 1546 */ 1547 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1548 !tcp_timer_active(tp, TT_PERSIST)) && 1549 ((flags & TH_SYN) == 0) && 1550 (error != EPERM)) { 1551 if (sack_rxmit) { 1552 p->rxmit -= len; 1553 tp->sackhint.sack_bytes_rexmit -= len; 1554 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1555 ("sackhint bytes rtx >= 0")); 1556 } else 1557 tp->snd_nxt -= len; 1558 } 1559 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1560 switch (error) { 1561 case EACCES: 1562 case EPERM: 1563 tp->t_softerror = error; 1564 return (error); 1565 case ENOBUFS: 1566 TCP_XMIT_TIMER_ASSERT(tp, len, flags); 1567 tp->snd_cwnd = tp->t_maxseg; 1568 return (0); 1569 case EMSGSIZE: 1570 /* 1571 * For some reason the interface we used initially 1572 * to send segments changed to another or lowered 1573 * its MTU. 1574 * If TSO was active we either got an interface 1575 * without TSO capabilits or TSO was turned off. 1576 * If we obtained mtu from ip_output() then update 1577 * it and try again. 1578 */ 1579 if (tso) 1580 tp->t_flags &= ~TF_TSO; 1581 if (mtu != 0) { 1582 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1583 goto again; 1584 } 1585 return (error); 1586 case EHOSTDOWN: 1587 case EHOSTUNREACH: 1588 case ENETDOWN: 1589 case ENETUNREACH: 1590 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1591 tp->t_softerror = error; 1592 return (0); 1593 } 1594 /* FALLTHROUGH */ 1595 default: 1596 return (error); 1597 } 1598 } 1599 TCPSTAT_INC(tcps_sndtotal); 1600 1601 /* 1602 * Data sent (as far as we can tell). 1603 * If this advertises a larger window than any other segment, 1604 * then remember the size of the advertised window. 1605 * Any pending ACK has now been sent. 1606 */ 1607 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1608 tp->rcv_adv = tp->rcv_nxt + recwin; 1609 tp->last_ack_sent = tp->rcv_nxt; 1610 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1611 if (tcp_timer_active(tp, TT_DELACK)) 1612 tcp_timer_activate(tp, TT_DELACK, 0); 1613 #if 0 1614 /* 1615 * This completely breaks TCP if newreno is turned on. What happens 1616 * is that if delayed-acks are turned on on the receiver, this code 1617 * on the transmitter effectively destroys the TCP window, forcing 1618 * it to four packets (1.5Kx4 = 6K window). 1619 */ 1620 if (sendalot && --maxburst) 1621 goto again; 1622 #endif 1623 if (sendalot) 1624 goto again; 1625 return (0); 1626 } 1627 1628 void 1629 tcp_setpersist(struct tcpcb *tp) 1630 { 1631 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1632 int tt; 1633 1634 tp->t_flags &= ~TF_PREVVALID; 1635 if (tcp_timer_active(tp, TT_REXMT)) 1636 panic("tcp_setpersist: retransmit pending"); 1637 /* 1638 * Start/restart persistence timer. 1639 */ 1640 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1641 tcp_persmin, tcp_persmax); 1642 tcp_timer_activate(tp, TT_PERSIST, tt); 1643 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1644 tp->t_rxtshift++; 1645 } 1646 1647 /* 1648 * Insert TCP options according to the supplied parameters to the place 1649 * optp in a consistent way. Can handle unaligned destinations. 1650 * 1651 * The order of the option processing is crucial for optimal packing and 1652 * alignment for the scarce option space. 1653 * 1654 * The optimal order for a SYN/SYN-ACK segment is: 1655 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1656 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1657 * 1658 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1659 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1660 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1661 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1662 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1663 */ 1664 int 1665 tcp_addoptions(struct tcpopt *to, u_char *optp) 1666 { 1667 u_int32_t mask, optlen = 0; 1668 1669 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1670 if ((to->to_flags & mask) != mask) 1671 continue; 1672 if (optlen == TCP_MAXOLEN) 1673 break; 1674 switch (to->to_flags & mask) { 1675 case TOF_MSS: 1676 while (optlen % 4) { 1677 optlen += TCPOLEN_NOP; 1678 *optp++ = TCPOPT_NOP; 1679 } 1680 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1681 continue; 1682 optlen += TCPOLEN_MAXSEG; 1683 *optp++ = TCPOPT_MAXSEG; 1684 *optp++ = TCPOLEN_MAXSEG; 1685 to->to_mss = htons(to->to_mss); 1686 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1687 optp += sizeof(to->to_mss); 1688 break; 1689 case TOF_SCALE: 1690 while (!optlen || optlen % 2 != 1) { 1691 optlen += TCPOLEN_NOP; 1692 *optp++ = TCPOPT_NOP; 1693 } 1694 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1695 continue; 1696 optlen += TCPOLEN_WINDOW; 1697 *optp++ = TCPOPT_WINDOW; 1698 *optp++ = TCPOLEN_WINDOW; 1699 *optp++ = to->to_wscale; 1700 break; 1701 case TOF_SACKPERM: 1702 while (optlen % 2) { 1703 optlen += TCPOLEN_NOP; 1704 *optp++ = TCPOPT_NOP; 1705 } 1706 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1707 continue; 1708 optlen += TCPOLEN_SACK_PERMITTED; 1709 *optp++ = TCPOPT_SACK_PERMITTED; 1710 *optp++ = TCPOLEN_SACK_PERMITTED; 1711 break; 1712 case TOF_TS: 1713 while (!optlen || optlen % 4 != 2) { 1714 optlen += TCPOLEN_NOP; 1715 *optp++ = TCPOPT_NOP; 1716 } 1717 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1718 continue; 1719 optlen += TCPOLEN_TIMESTAMP; 1720 *optp++ = TCPOPT_TIMESTAMP; 1721 *optp++ = TCPOLEN_TIMESTAMP; 1722 to->to_tsval = htonl(to->to_tsval); 1723 to->to_tsecr = htonl(to->to_tsecr); 1724 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1725 optp += sizeof(to->to_tsval); 1726 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1727 optp += sizeof(to->to_tsecr); 1728 break; 1729 case TOF_SIGNATURE: 1730 { 1731 int siglen = TCPOLEN_SIGNATURE - 2; 1732 1733 while (!optlen || optlen % 4 != 2) { 1734 optlen += TCPOLEN_NOP; 1735 *optp++ = TCPOPT_NOP; 1736 } 1737 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) { 1738 to->to_flags &= ~TOF_SIGNATURE; 1739 continue; 1740 } 1741 optlen += TCPOLEN_SIGNATURE; 1742 *optp++ = TCPOPT_SIGNATURE; 1743 *optp++ = TCPOLEN_SIGNATURE; 1744 to->to_signature = optp; 1745 while (siglen--) 1746 *optp++ = 0; 1747 break; 1748 } 1749 case TOF_SACK: 1750 { 1751 int sackblks = 0; 1752 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1753 tcp_seq sack_seq; 1754 1755 while (!optlen || optlen % 4 != 2) { 1756 optlen += TCPOLEN_NOP; 1757 *optp++ = TCPOPT_NOP; 1758 } 1759 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1760 continue; 1761 optlen += TCPOLEN_SACKHDR; 1762 *optp++ = TCPOPT_SACK; 1763 sackblks = min(to->to_nsacks, 1764 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1765 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1766 while (sackblks--) { 1767 sack_seq = htonl(sack->start); 1768 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1769 optp += sizeof(sack_seq); 1770 sack_seq = htonl(sack->end); 1771 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1772 optp += sizeof(sack_seq); 1773 optlen += TCPOLEN_SACK; 1774 sack++; 1775 } 1776 TCPSTAT_INC(tcps_sack_send_blocks); 1777 break; 1778 } 1779 case TOF_FASTOPEN: 1780 { 1781 int total_len; 1782 1783 /* XXX is there any point to aligning this option? */ 1784 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; 1785 if (TCP_MAXOLEN - optlen < total_len) { 1786 to->to_flags &= ~TOF_FASTOPEN; 1787 continue; 1788 } 1789 *optp++ = TCPOPT_FAST_OPEN; 1790 *optp++ = total_len; 1791 if (to->to_tfo_len > 0) { 1792 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); 1793 optp += to->to_tfo_len; 1794 } 1795 optlen += total_len; 1796 break; 1797 } 1798 default: 1799 panic("%s: unknown TCP option type", __func__); 1800 break; 1801 } 1802 } 1803 1804 /* Terminate and pad TCP options to a 4 byte boundary. */ 1805 if (optlen % 4) { 1806 optlen += TCPOLEN_EOL; 1807 *optp++ = TCPOPT_EOL; 1808 } 1809 /* 1810 * According to RFC 793 (STD0007): 1811 * "The content of the header beyond the End-of-Option option 1812 * must be header padding (i.e., zero)." 1813 * and later: "The padding is composed of zeros." 1814 */ 1815 while (optlen % 4) { 1816 optlen += TCPOLEN_PAD; 1817 *optp++ = TCPOPT_PAD; 1818 } 1819 1820 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1821 return (optlen); 1822 } 1823 1824 /* 1825 * This is a copy of m_copym(), taking the TSO segment size/limit 1826 * constraints into account, and advancing the sndptr as it goes. 1827 */ 1828 struct mbuf * 1829 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen, 1830 int32_t seglimit, int32_t segsize, struct sockbuf *sb) 1831 { 1832 struct mbuf *n, **np; 1833 struct mbuf *top; 1834 int32_t off = off0; 1835 int32_t len = *plen; 1836 int32_t fragsize; 1837 int32_t len_cp = 0; 1838 int32_t *pkthdrlen; 1839 uint32_t mlen, frags; 1840 bool copyhdr; 1841 1842 1843 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off)); 1844 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len)); 1845 if (off == 0 && m->m_flags & M_PKTHDR) 1846 copyhdr = true; 1847 else 1848 copyhdr = false; 1849 while (off > 0) { 1850 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain")); 1851 if (off < m->m_len) 1852 break; 1853 off -= m->m_len; 1854 if ((sb) && (m == sb->sb_sndptr)) { 1855 sb->sb_sndptroff += m->m_len; 1856 sb->sb_sndptr = m->m_next; 1857 } 1858 m = m->m_next; 1859 } 1860 np = ⊤ 1861 top = NULL; 1862 pkthdrlen = NULL; 1863 while (len > 0) { 1864 if (m == NULL) { 1865 KASSERT(len == M_COPYALL, 1866 ("tcp_m_copym, length > size of mbuf chain")); 1867 *plen = len_cp; 1868 if (pkthdrlen != NULL) 1869 *pkthdrlen = len_cp; 1870 break; 1871 } 1872 mlen = min(len, m->m_len - off); 1873 if (seglimit) { 1874 /* 1875 * For M_NOMAP mbufs, add 3 segments 1876 * + 1 in case we are crossing page boundaries 1877 * + 2 in case the TLS hdr/trailer are used 1878 * It is cheaper to just add the segments 1879 * than it is to take the cache miss to look 1880 * at the mbuf ext_pgs state in detail. 1881 */ 1882 if (m->m_flags & M_NOMAP) { 1883 fragsize = min(segsize, PAGE_SIZE); 1884 frags = 3; 1885 } else { 1886 fragsize = segsize; 1887 frags = 0; 1888 } 1889 1890 /* Break if we really can't fit anymore. */ 1891 if ((frags + 1) >= seglimit) { 1892 *plen = len_cp; 1893 if (pkthdrlen != NULL) 1894 *pkthdrlen = len_cp; 1895 break; 1896 } 1897 1898 /* 1899 * Reduce size if you can't copy the whole 1900 * mbuf. If we can't copy the whole mbuf, also 1901 * adjust len so the loop will end after this 1902 * mbuf. 1903 */ 1904 if ((frags + howmany(mlen, fragsize)) >= seglimit) { 1905 mlen = (seglimit - frags - 1) * fragsize; 1906 len = mlen; 1907 *plen = len_cp + len; 1908 if (pkthdrlen != NULL) 1909 *pkthdrlen = *plen; 1910 } 1911 frags += howmany(mlen, fragsize); 1912 if (frags == 0) 1913 frags++; 1914 seglimit -= frags; 1915 KASSERT(seglimit > 0, 1916 ("%s: seglimit went too low", __func__)); 1917 } 1918 if (copyhdr) 1919 n = m_gethdr(M_NOWAIT, m->m_type); 1920 else 1921 n = m_get(M_NOWAIT, m->m_type); 1922 *np = n; 1923 if (n == NULL) 1924 goto nospace; 1925 if (copyhdr) { 1926 if (!m_dup_pkthdr(n, m, M_NOWAIT)) 1927 goto nospace; 1928 if (len == M_COPYALL) 1929 n->m_pkthdr.len -= off0; 1930 else 1931 n->m_pkthdr.len = len; 1932 pkthdrlen = &n->m_pkthdr.len; 1933 copyhdr = false; 1934 } 1935 n->m_len = mlen; 1936 len_cp += n->m_len; 1937 if (m->m_flags & M_EXT) { 1938 n->m_data = m->m_data + off; 1939 mb_dupcl(n, m); 1940 } else 1941 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 1942 (u_int)n->m_len); 1943 1944 if (sb && (sb->sb_sndptr == m) && 1945 ((n->m_len + off) >= m->m_len) && m->m_next) { 1946 sb->sb_sndptroff += m->m_len; 1947 sb->sb_sndptr = m->m_next; 1948 } 1949 off = 0; 1950 if (len != M_COPYALL) { 1951 len -= n->m_len; 1952 } 1953 m = m->m_next; 1954 np = &n->m_next; 1955 } 1956 return (top); 1957 nospace: 1958 m_freem(top); 1959 return (NULL); 1960 } 1961 1962 void 1963 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin) 1964 { 1965 1966 /* 1967 * Automatic sizing of send socket buffer. Often the send buffer 1968 * size is not optimally adjusted to the actual network conditions 1969 * at hand (delay bandwidth product). Setting the buffer size too 1970 * small limits throughput on links with high bandwidth and high 1971 * delay (eg. trans-continental/oceanic links). Setting the 1972 * buffer size too big consumes too much real kernel memory, 1973 * especially with many connections on busy servers. 1974 * 1975 * The criteria to step up the send buffer one notch are: 1976 * 1. receive window of remote host is larger than send buffer 1977 * (with a fudge factor of 5/4th); 1978 * 2. send buffer is filled to 7/8th with data (so we actually 1979 * have data to make use of it); 1980 * 3. send buffer fill has not hit maximal automatic size; 1981 * 4. our send window (slow start and cogestion controlled) is 1982 * larger than sent but unacknowledged data in send buffer. 1983 * 1984 * The remote host receive window scaling factor may limit the 1985 * growing of the send buffer before it reaches its allowed 1986 * maximum. 1987 * 1988 * It scales directly with slow start or congestion window 1989 * and does at most one step per received ACK. This fast 1990 * scaling has the drawback of growing the send buffer beyond 1991 * what is strictly necessary to make full use of a given 1992 * delay*bandwidth product. However testing has shown this not 1993 * to be much of an problem. At worst we are trading wasting 1994 * of available bandwidth (the non-use of it) for wasting some 1995 * socket buffer memory. 1996 * 1997 * TODO: Shrink send buffer during idle periods together 1998 * with congestion window. Requires another timer. Has to 1999 * wait for upcoming tcp timer rewrite. 2000 * 2001 * XXXGL: should there be used sbused() or sbavail()? 2002 */ 2003 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 2004 int lowat; 2005 2006 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0; 2007 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat && 2008 sbused(&so->so_snd) >= 2009 (so->so_snd.sb_hiwat / 8 * 7) - lowat && 2010 sbused(&so->so_snd) < V_tcp_autosndbuf_max && 2011 sendwin >= (sbused(&so->so_snd) - 2012 (tp->snd_nxt - tp->snd_una))) { 2013 if (!sbreserve_locked(&so->so_snd, 2014 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 2015 V_tcp_autosndbuf_max), so, curthread)) 2016 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 2017 } 2018 } 2019 } 2020