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