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 #ifdef TCP_RFC7413 75 #include <netinet/tcp_fastopen.h> 76 #endif 77 #include <netinet/tcp.h> 78 #define TCPOUTFLAGS 79 #include <netinet/tcp_fsm.h> 80 #include <netinet/tcp_seq.h> 81 #include <netinet/tcp_timer.h> 82 #include <netinet/tcp_var.h> 83 #include <netinet/tcpip.h> 84 #include <netinet/cc/cc.h> 85 #ifdef TCPPCAP 86 #include <netinet/tcp_pcap.h> 87 #endif 88 #ifdef TCPDEBUG 89 #include <netinet/tcp_debug.h> 90 #endif 91 #ifdef TCP_OFFLOAD 92 #include <netinet/tcp_offload.h> 93 #endif 94 95 #include <netipsec/ipsec_support.h> 96 97 #include <machine/in_cksum.h> 98 99 #include <security/mac/mac_framework.h> 100 101 VNET_DEFINE(int, path_mtu_discovery) = 1; 102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW, 103 &VNET_NAME(path_mtu_discovery), 1, 104 "Enable Path MTU Discovery"); 105 106 VNET_DEFINE(int, tcp_do_tso) = 1; 107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW, 108 &VNET_NAME(tcp_do_tso), 0, 109 "Enable TCP Segmentation Offload"); 110 111 VNET_DEFINE(int, tcp_sendspace) = 1024*32; 112 #define V_tcp_sendspace VNET(tcp_sendspace) 113 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW, 114 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 115 116 VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 117 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, 118 &VNET_NAME(tcp_do_autosndbuf), 0, 119 "Enable automatic send buffer sizing"); 120 121 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 122 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW, 123 &VNET_NAME(tcp_autosndbuf_inc), 0, 124 "Incrementor step size of automatic send buffer"); 125 126 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; 127 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW, 128 &VNET_NAME(tcp_autosndbuf_max), 0, 129 "Max size of automatic send buffer"); 130 131 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0; 132 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat) 133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW, 134 &VNET_NAME(tcp_sendbuf_auto_lowat), 0, 135 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT"); 136 137 /* 138 * Make sure that either retransmit or persist timer is set for SYN, FIN and 139 * non-ACK. 140 */ 141 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \ 142 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\ 143 tcp_timer_active((tp), TT_REXMT) || \ 144 tcp_timer_active((tp), TT_PERSIST), \ 145 ("neither rexmt nor persist timer is set")) 146 147 #ifdef TCP_HHOOK 148 static void inline hhook_run_tcp_est_out(struct tcpcb *tp, 149 struct tcphdr *th, struct tcpopt *to, 150 uint32_t len, int tso); 151 #endif 152 static void inline cc_after_idle(struct tcpcb *tp); 153 154 #ifdef TCP_HHOOK 155 /* 156 * Wrapper for the TCP established output helper hook. 157 */ 158 static void inline 159 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 160 struct tcpopt *to, uint32_t len, int tso) 161 { 162 struct tcp_hhook_data hhook_data; 163 164 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 165 hhook_data.tp = tp; 166 hhook_data.th = th; 167 hhook_data.to = to; 168 hhook_data.len = len; 169 hhook_data.tso = tso; 170 171 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 172 tp->osd); 173 } 174 } 175 #endif 176 177 /* 178 * CC wrapper hook functions 179 */ 180 static void inline 181 cc_after_idle(struct tcpcb *tp) 182 { 183 INP_WLOCK_ASSERT(tp->t_inpcb); 184 185 if (CC_ALGO(tp)->after_idle != NULL) 186 CC_ALGO(tp)->after_idle(tp->ccv); 187 } 188 189 /* 190 * Tcp output routine: figure out what should be sent and send it. 191 */ 192 int 193 tcp_output(struct tcpcb *tp) 194 { 195 struct socket *so = tp->t_inpcb->inp_socket; 196 int32_t len; 197 uint32_t recwin, sendwin; 198 int off, flags, error = 0; /* Keep compiler happy */ 199 struct mbuf *m; 200 struct ip *ip = NULL; 201 struct ipovly *ipov = NULL; 202 struct tcphdr *th; 203 u_char opt[TCP_MAXOLEN]; 204 unsigned ipoptlen, optlen, hdrlen; 205 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 206 unsigned ipsec_optlen = 0; 207 #endif 208 int idle, sendalot; 209 int sack_rxmit, sack_bytes_rxmt; 210 struct sackhole *p; 211 int tso, mtu; 212 struct tcpopt to; 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 #ifdef TCP_RFC7413 231 /* 232 * For TFO connections in SYN_RECEIVED, only allow the initial 233 * SYN|ACK and those sent by the retransmit timer. 234 */ 235 if (IS_FASTOPEN(tp->t_flags) && 236 (tp->t_state == TCPS_SYN_RECEIVED) && 237 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */ 238 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 239 return (0); 240 #endif 241 /* 242 * Determine length of data that should be transmitted, 243 * and flags that will be used. 244 * If there is some data or critical controls (SYN, RST) 245 * to send, then transmit; otherwise, investigate further. 246 */ 247 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 248 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) 249 cc_after_idle(tp); 250 tp->t_flags &= ~TF_LASTIDLE; 251 if (idle) { 252 if (tp->t_flags & TF_MORETOCOME) { 253 tp->t_flags |= TF_LASTIDLE; 254 idle = 0; 255 } 256 } 257 again: 258 /* 259 * If we've recently taken a timeout, snd_max will be greater than 260 * snd_nxt. There may be SACK information that allows us to avoid 261 * resending already delivered data. Adjust snd_nxt accordingly. 262 */ 263 if ((tp->t_flags & TF_SACK_PERMIT) && 264 SEQ_LT(tp->snd_nxt, tp->snd_max)) 265 tcp_sack_adjust(tp); 266 sendalot = 0; 267 tso = 0; 268 mtu = 0; 269 off = tp->snd_nxt - tp->snd_una; 270 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 271 272 flags = tcp_outflags[tp->t_state]; 273 /* 274 * Send any SACK-generated retransmissions. If we're explicitly trying 275 * to send out new data (when sendalot is 1), bypass this function. 276 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 277 * we're replacing a (future) new transmission with a retransmission 278 * now, and we previously incremented snd_cwnd in tcp_input(). 279 */ 280 /* 281 * Still in sack recovery , reset rxmit flag to zero. 282 */ 283 sack_rxmit = 0; 284 sack_bytes_rxmt = 0; 285 len = 0; 286 p = NULL; 287 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 288 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 289 uint32_t cwin; 290 291 cwin = 292 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0); 293 /* Do not retransmit SACK segments beyond snd_recover */ 294 if (SEQ_GT(p->end, tp->snd_recover)) { 295 /* 296 * (At least) part of sack hole extends beyond 297 * snd_recover. Check to see if we can rexmit data 298 * for this hole. 299 */ 300 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 301 /* 302 * Can't rexmit any more data for this hole. 303 * That data will be rexmitted in the next 304 * sack recovery episode, when snd_recover 305 * moves past p->rxmit. 306 */ 307 p = NULL; 308 goto after_sack_rexmit; 309 } else 310 /* Can rexmit part of the current hole */ 311 len = ((int32_t)ulmin(cwin, 312 tp->snd_recover - p->rxmit)); 313 } else 314 len = ((int32_t)ulmin(cwin, p->end - p->rxmit)); 315 off = p->rxmit - tp->snd_una; 316 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 317 __func__, off)); 318 if (len > 0) { 319 sack_rxmit = 1; 320 sendalot = 1; 321 TCPSTAT_INC(tcps_sack_rexmits); 322 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 323 min(len, tp->t_maxseg)); 324 } 325 } 326 after_sack_rexmit: 327 /* 328 * Get standard flags, and add SYN or FIN if requested by 'hidden' 329 * state flags. 330 */ 331 if (tp->t_flags & TF_NEEDFIN) 332 flags |= TH_FIN; 333 if (tp->t_flags & TF_NEEDSYN) 334 flags |= TH_SYN; 335 336 SOCKBUF_LOCK(&so->so_snd); 337 /* 338 * If in persist timeout with window of 0, send 1 byte. 339 * Otherwise, if window is small but nonzero 340 * and timer expired, we will send what we can 341 * and go to transmit state. 342 */ 343 if (tp->t_flags & TF_FORCEDATA) { 344 if (sendwin == 0) { 345 /* 346 * If we still have some data to send, then 347 * clear the FIN bit. Usually this would 348 * happen below when it realizes that we 349 * aren't sending all the data. However, 350 * if we have exactly 1 byte of unsent data, 351 * then it won't clear the FIN bit below, 352 * and if we are in persist state, we wind 353 * up sending the packet without recording 354 * that we sent the FIN bit. 355 * 356 * We can't just blindly clear the FIN bit, 357 * because if we don't have any more data 358 * to send then the probe will be the FIN 359 * itself. 360 */ 361 if (off < sbused(&so->so_snd)) 362 flags &= ~TH_FIN; 363 sendwin = 1; 364 } else { 365 tcp_timer_activate(tp, TT_PERSIST, 0); 366 tp->t_rxtshift = 0; 367 } 368 } 369 370 /* 371 * If snd_nxt == snd_max and we have transmitted a FIN, the 372 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 373 * a negative length. This can also occur when TCP opens up 374 * its congestion window while receiving additional duplicate 375 * acks after fast-retransmit because TCP will reset snd_nxt 376 * to snd_max after the fast-retransmit. 377 * 378 * In the normal retransmit-FIN-only case, however, snd_nxt will 379 * be set to snd_una, the offset will be 0, and the length may 380 * wind up 0. 381 * 382 * If sack_rxmit is true we are retransmitting from the scoreboard 383 * in which case len is already set. 384 */ 385 if (sack_rxmit == 0) { 386 if (sack_bytes_rxmt == 0) 387 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) - 388 off); 389 else { 390 int32_t cwin; 391 392 /* 393 * We are inside of a SACK recovery episode and are 394 * sending new data, having retransmitted all the 395 * data possible in the scoreboard. 396 */ 397 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) - 398 off); 399 /* 400 * Don't remove this (len > 0) check ! 401 * We explicitly check for len > 0 here (although it 402 * isn't really necessary), to work around a gcc 403 * optimization issue - to force gcc to compute 404 * len above. Without this check, the computation 405 * of len is bungled by the optimizer. 406 */ 407 if (len > 0) { 408 cwin = tp->snd_cwnd - 409 (tp->snd_nxt - tp->sack_newdata) - 410 sack_bytes_rxmt; 411 if (cwin < 0) 412 cwin = 0; 413 len = imin(len, cwin); 414 } 415 } 416 } 417 418 /* 419 * Lop off SYN bit if it has already been sent. However, if this 420 * is SYN-SENT state and if segment contains data and if we don't 421 * know that foreign host supports TAO, suppress sending segment. 422 */ 423 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 424 if (tp->t_state != TCPS_SYN_RECEIVED) 425 flags &= ~TH_SYN; 426 #ifdef TCP_RFC7413 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 #endif 435 off--, len++; 436 } 437 438 /* 439 * Be careful not to send data and/or FIN on SYN segments. 440 * This measure is needed to prevent interoperability problems 441 * with not fully conformant TCP implementations. 442 */ 443 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 444 len = 0; 445 flags &= ~TH_FIN; 446 } 447 448 #ifdef TCP_RFC7413 449 /* 450 * When retransmitting SYN|ACK on a passively-created TFO socket, 451 * don't include data, as the presence of data may have caused the 452 * original SYN|ACK to have been dropped by a middlebox. 453 */ 454 if (IS_FASTOPEN(tp->t_flags) && 455 (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) || 456 (flags & TH_RST))) 457 len = 0; 458 #endif 459 if (len <= 0) { 460 /* 461 * If FIN has been sent but not acked, 462 * but we haven't been called to retransmit, 463 * len will be < 0. Otherwise, window shrank 464 * after we sent into it. If window shrank to 0, 465 * cancel pending retransmit, pull snd_nxt back 466 * to (closed) window, and set the persist timer 467 * if it isn't already going. If the window didn't 468 * close completely, just wait for an ACK. 469 * 470 * We also do a general check here to ensure that 471 * we will set the persist timer when we have data 472 * to send, but a 0-byte window. This makes sure 473 * the persist timer is set even if the packet 474 * hits one of the "goto send" lines below. 475 */ 476 len = 0; 477 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 478 (off < (int) sbavail(&so->so_snd))) { 479 tcp_timer_activate(tp, TT_REXMT, 0); 480 tp->t_rxtshift = 0; 481 tp->snd_nxt = tp->snd_una; 482 if (!tcp_timer_active(tp, TT_PERSIST)) 483 tcp_setpersist(tp); 484 } 485 } 486 487 /* len will be >= 0 after this point. */ 488 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 489 490 tcp_sndbuf_autoscale(tp, so, sendwin); 491 492 /* 493 * Decide if we can use TCP Segmentation Offloading (if supported by 494 * hardware). 495 * 496 * TSO may only be used if we are in a pure bulk sending state. The 497 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 498 * IP options prevent using TSO. With TSO the TCP header is the same 499 * (except for the sequence number) for all generated packets. This 500 * makes it impossible to transmit any options which vary per generated 501 * segment or packet. 502 * 503 * IPv4 handling has a clear separation of ip options and ip header 504 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 505 * the right thing below to provide length of just ip options and thus 506 * checking for ipoptlen is enough to decide if ip options are present. 507 */ 508 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 509 /* 510 * Pre-calculate here as we save another lookup into the darknesses 511 * of IPsec that way and can actually decide if TSO is ok. 512 */ 513 #ifdef INET6 514 if (isipv6 && IPSEC_ENABLED(ipv6)) 515 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); 516 #ifdef INET 517 else 518 #endif 519 #endif /* INET6 */ 520 #ifdef INET 521 if (IPSEC_ENABLED(ipv4)) 522 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); 523 #endif /* INET */ 524 #endif /* IPSEC */ 525 #ifdef INET6 526 if (isipv6) 527 ipoptlen = ip6_optlen(tp->t_inpcb); 528 else 529 #endif 530 if (tp->t_inpcb->inp_options) 531 ipoptlen = tp->t_inpcb->inp_options->m_len - 532 offsetof(struct ipoption, ipopt_list); 533 else 534 ipoptlen = 0; 535 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 536 ipoptlen += ipsec_optlen; 537 #endif 538 539 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 540 ((tp->t_flags & TF_SIGNATURE) == 0) && 541 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 542 ipoptlen == 0) 543 tso = 1; 544 545 if (sack_rxmit) { 546 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd))) 547 flags &= ~TH_FIN; 548 } else { 549 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 550 sbused(&so->so_snd))) 551 flags &= ~TH_FIN; 552 } 553 554 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 555 (long)TCP_MAXWIN << tp->rcv_scale); 556 557 /* 558 * Sender silly window avoidance. We transmit under the following 559 * conditions when len is non-zero: 560 * 561 * - We have a full segment (or more with TSO) 562 * - This is the last buffer in a write()/send() and we are 563 * either idle or running NODELAY 564 * - we've timed out (e.g. persist timer) 565 * - we have more then 1/2 the maximum send window's worth of 566 * data (receiver may be limited the window size) 567 * - we need to retransmit 568 */ 569 if (len) { 570 if (len >= tp->t_maxseg) 571 goto send; 572 /* 573 * NOTE! on localhost connections an 'ack' from the remote 574 * end may occur synchronously with the output and cause 575 * us to flush a buffer queued with moretocome. XXX 576 * 577 * note: the len + off check is almost certainly unnecessary. 578 */ 579 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 580 (idle || (tp->t_flags & TF_NODELAY)) && 581 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) && 582 (tp->t_flags & TF_NOPUSH) == 0) { 583 goto send; 584 } 585 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 586 goto send; 587 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 588 goto send; 589 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 590 goto send; 591 if (sack_rxmit) 592 goto send; 593 } 594 595 /* 596 * Sending of standalone window updates. 597 * 598 * Window updates are important when we close our window due to a 599 * full socket buffer and are opening it again after the application 600 * reads data from it. Once the window has opened again and the 601 * remote end starts to send again the ACK clock takes over and 602 * provides the most current window information. 603 * 604 * We must avoid the silly window syndrome whereas every read 605 * from the receive buffer, no matter how small, causes a window 606 * update to be sent. We also should avoid sending a flurry of 607 * window updates when the socket buffer had queued a lot of data 608 * and the application is doing small reads. 609 * 610 * Prevent a flurry of pointless window updates by only sending 611 * an update when we can increase the advertized window by more 612 * than 1/4th of the socket buffer capacity. When the buffer is 613 * getting full or is very small be more aggressive and send an 614 * update whenever we can increase by two mss sized segments. 615 * In all other situations the ACK's to new incoming data will 616 * carry further window increases. 617 * 618 * Don't send an independent window update if a delayed 619 * ACK is pending (it will get piggy-backed on it) or the 620 * remote side already has done a half-close and won't send 621 * more data. Skip this if the connection is in T/TCP 622 * half-open state. 623 */ 624 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 625 !(tp->t_flags & TF_DELACK) && 626 !TCPS_HAVERCVDFIN(tp->t_state)) { 627 /* 628 * "adv" is the amount we could increase the window, 629 * taking into account that we are limited by 630 * TCP_MAXWIN << tp->rcv_scale. 631 */ 632 int32_t adv; 633 int oldwin; 634 635 adv = recwin; 636 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 637 oldwin = (tp->rcv_adv - tp->rcv_nxt); 638 adv -= oldwin; 639 } else 640 oldwin = 0; 641 642 /* 643 * If the new window size ends up being the same as or less 644 * than the old size when it is scaled, then don't force 645 * a window update. 646 */ 647 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 648 goto dontupdate; 649 650 if (adv >= (int32_t)(2 * tp->t_maxseg) && 651 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 652 recwin <= (so->so_rcv.sb_hiwat / 8) || 653 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 654 goto send; 655 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat) 656 goto send; 657 } 658 dontupdate: 659 660 /* 661 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 662 * is also a catch-all for the retransmit timer timeout case. 663 */ 664 if (tp->t_flags & TF_ACKNOW) 665 goto send; 666 if ((flags & TH_RST) || 667 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 668 goto send; 669 if (SEQ_GT(tp->snd_up, tp->snd_una)) 670 goto send; 671 /* 672 * If our state indicates that FIN should be sent 673 * and we have not yet done so, then we need to send. 674 */ 675 if (flags & TH_FIN && 676 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 677 goto send; 678 /* 679 * In SACK, it is possible for tcp_output to fail to send a segment 680 * after the retransmission timer has been turned off. Make sure 681 * that the retransmission timer is set. 682 */ 683 if ((tp->t_flags & TF_SACK_PERMIT) && 684 SEQ_GT(tp->snd_max, tp->snd_una) && 685 !tcp_timer_active(tp, TT_REXMT) && 686 !tcp_timer_active(tp, TT_PERSIST)) { 687 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 688 goto just_return; 689 } 690 /* 691 * TCP window updates are not reliable, rather a polling protocol 692 * using ``persist'' packets is used to insure receipt of window 693 * updates. The three ``states'' for the output side are: 694 * idle not doing retransmits or persists 695 * persisting to move a small or zero window 696 * (re)transmitting and thereby not persisting 697 * 698 * tcp_timer_active(tp, TT_PERSIST) 699 * is true when we are in persist state. 700 * (tp->t_flags & TF_FORCEDATA) 701 * is set when we are called to send a persist packet. 702 * tcp_timer_active(tp, TT_REXMT) 703 * is set when we are retransmitting 704 * The output side is idle when both timers are zero. 705 * 706 * If send window is too small, there is data to transmit, and no 707 * retransmit or persist is pending, then go to persist state. 708 * If nothing happens soon, send when timer expires: 709 * if window is nonzero, transmit what we can, 710 * otherwise force out a byte. 711 */ 712 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && 713 !tcp_timer_active(tp, TT_PERSIST)) { 714 tp->t_rxtshift = 0; 715 tcp_setpersist(tp); 716 } 717 718 /* 719 * No reason to send a segment, just return. 720 */ 721 just_return: 722 SOCKBUF_UNLOCK(&so->so_snd); 723 return (0); 724 725 send: 726 SOCKBUF_LOCK_ASSERT(&so->so_snd); 727 if (len > 0) { 728 if (len >= tp->t_maxseg) 729 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 730 else 731 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 732 } 733 /* 734 * Before ESTABLISHED, force sending of initial options 735 * unless TCP set not to do any options. 736 * NOTE: we assume that the IP/TCP header plus TCP options 737 * always fit in a single mbuf, leaving room for a maximum 738 * link header, i.e. 739 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 740 */ 741 optlen = 0; 742 #ifdef INET6 743 if (isipv6) 744 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 745 else 746 #endif 747 hdrlen = sizeof (struct tcpiphdr); 748 749 /* 750 * Compute options for segment. 751 * We only have to care about SYN and established connection 752 * segments. Options for SYN-ACK segments are handled in TCP 753 * syncache. 754 */ 755 to.to_flags = 0; 756 if ((tp->t_flags & TF_NOOPT) == 0) { 757 /* Maximum segment size. */ 758 if (flags & TH_SYN) { 759 tp->snd_nxt = tp->iss; 760 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 761 to.to_flags |= TOF_MSS; 762 #ifdef TCP_RFC7413 763 /* 764 * Only include the TFO option on the first 765 * transmission of the SYN|ACK on a 766 * passively-created TFO socket, as the presence of 767 * the TFO option may have caused the original 768 * SYN|ACK to have been dropped by a middlebox. 769 */ 770 if (IS_FASTOPEN(tp->t_flags) && 771 (tp->t_state == TCPS_SYN_RECEIVED) && 772 (tp->t_rxtshift == 0)) { 773 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 774 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie; 775 to.to_flags |= TOF_FASTOPEN; 776 } 777 #endif 778 } 779 /* Window scaling. */ 780 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 781 to.to_wscale = tp->request_r_scale; 782 to.to_flags |= TOF_SCALE; 783 } 784 /* Timestamps. */ 785 if ((tp->t_flags & TF_RCVD_TSTMP) || 786 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 787 to.to_tsval = tcp_ts_getticks() + tp->ts_offset; 788 to.to_tsecr = tp->ts_recent; 789 to.to_flags |= TOF_TS; 790 } 791 792 /* Set receive buffer autosizing timestamp. */ 793 if (tp->rfbuf_ts == 0 && 794 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 795 tp->rfbuf_ts = tcp_ts_getticks(); 796 797 /* Selective ACK's. */ 798 if (tp->t_flags & TF_SACK_PERMIT) { 799 if (flags & TH_SYN) 800 to.to_flags |= TOF_SACKPERM; 801 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 802 (tp->t_flags & TF_SACK_PERMIT) && 803 tp->rcv_numsacks > 0) { 804 to.to_flags |= TOF_SACK; 805 to.to_nsacks = tp->rcv_numsacks; 806 to.to_sacks = (u_char *)tp->sackblks; 807 } 808 } 809 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 810 /* TCP-MD5 (RFC2385). */ 811 /* 812 * Check that TCP_MD5SIG is enabled in tcpcb to 813 * account the size needed to set this TCP option. 814 */ 815 if (tp->t_flags & TF_SIGNATURE) 816 to.to_flags |= TOF_SIGNATURE; 817 #endif /* TCP_SIGNATURE */ 818 819 /* Processing the options. */ 820 hdrlen += optlen = tcp_addoptions(&to, opt); 821 } 822 823 /* 824 * Adjust data length if insertion of options will 825 * bump the packet length beyond the t_maxseg length. 826 * Clear the FIN bit because we cut off the tail of 827 * the segment. 828 */ 829 if (len + optlen + ipoptlen > tp->t_maxseg) { 830 flags &= ~TH_FIN; 831 832 if (tso) { 833 u_int if_hw_tsomax; 834 u_int if_hw_tsomaxsegcount; 835 u_int if_hw_tsomaxsegsize; 836 struct mbuf *mb; 837 u_int moff; 838 int max_len; 839 840 /* extract TSO information */ 841 if_hw_tsomax = tp->t_tsomax; 842 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 843 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 844 845 /* 846 * Limit a TSO burst to prevent it from 847 * overflowing or exceeding the maximum length 848 * allowed by the network interface: 849 */ 850 KASSERT(ipoptlen == 0, 851 ("%s: TSO can't do IP options", __func__)); 852 853 /* 854 * Check if we should limit by maximum payload 855 * length: 856 */ 857 if (if_hw_tsomax != 0) { 858 /* compute maximum TSO length */ 859 max_len = (if_hw_tsomax - hdrlen - 860 max_linkhdr); 861 if (max_len <= 0) { 862 len = 0; 863 } else if (len > max_len) { 864 sendalot = 1; 865 len = max_len; 866 } 867 } 868 869 /* 870 * Check if we should limit by maximum segment 871 * size and count: 872 */ 873 if (if_hw_tsomaxsegcount != 0 && 874 if_hw_tsomaxsegsize != 0) { 875 /* 876 * Subtract one segment for the LINK 877 * and TCP/IP headers mbuf that will 878 * be prepended to this mbuf chain 879 * after the code in this section 880 * limits the number of mbufs in the 881 * chain to if_hw_tsomaxsegcount. 882 */ 883 if_hw_tsomaxsegcount -= 1; 884 max_len = 0; 885 mb = sbsndmbuf(&so->so_snd, off, &moff); 886 887 while (mb != NULL && max_len < len) { 888 u_int mlen; 889 u_int frags; 890 891 /* 892 * Get length of mbuf fragment 893 * and how many hardware frags, 894 * rounded up, it would use: 895 */ 896 mlen = (mb->m_len - moff); 897 frags = howmany(mlen, 898 if_hw_tsomaxsegsize); 899 900 /* Handle special case: Zero Length Mbuf */ 901 if (frags == 0) 902 frags = 1; 903 904 /* 905 * Check if the fragment limit 906 * will be reached or exceeded: 907 */ 908 if (frags >= if_hw_tsomaxsegcount) { 909 max_len += min(mlen, 910 if_hw_tsomaxsegcount * 911 if_hw_tsomaxsegsize); 912 break; 913 } 914 max_len += mlen; 915 if_hw_tsomaxsegcount -= frags; 916 moff = 0; 917 mb = mb->m_next; 918 } 919 if (max_len <= 0) { 920 len = 0; 921 } else if (len > max_len) { 922 sendalot = 1; 923 len = max_len; 924 } 925 } 926 927 /* 928 * Prevent the last segment from being 929 * fractional unless the send sockbuf can be 930 * emptied: 931 */ 932 max_len = (tp->t_maxseg - optlen); 933 if (((uint32_t)off + (uint32_t)len) < 934 sbavail(&so->so_snd)) { 935 moff = len % max_len; 936 if (moff != 0) { 937 len -= moff; 938 sendalot = 1; 939 } 940 } 941 942 /* 943 * In case there are too many small fragments 944 * don't use TSO: 945 */ 946 if (len <= max_len) { 947 len = max_len; 948 sendalot = 1; 949 tso = 0; 950 } 951 952 /* 953 * Send the FIN in a separate segment 954 * after the bulk sending is done. 955 * We don't trust the TSO implementations 956 * to clear the FIN flag on all but the 957 * last segment. 958 */ 959 if (tp->t_flags & TF_NEEDFIN) 960 sendalot = 1; 961 962 } else { 963 len = tp->t_maxseg - optlen - ipoptlen; 964 sendalot = 1; 965 } 966 } else 967 tso = 0; 968 969 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 970 ("%s: len > IP_MAXPACKET", __func__)); 971 972 /*#ifdef DIAGNOSTIC*/ 973 #ifdef INET6 974 if (max_linkhdr + hdrlen > MCLBYTES) 975 #else 976 if (max_linkhdr + hdrlen > MHLEN) 977 #endif 978 panic("tcphdr too big"); 979 /*#endif*/ 980 981 /* 982 * This KASSERT is here to catch edge cases at a well defined place. 983 * Before, those had triggered (random) panic conditions further down. 984 */ 985 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 986 987 /* 988 * Grab a header mbuf, attaching a copy of data to 989 * be transmitted, and initialize the header from 990 * the template for sends on this connection. 991 */ 992 if (len) { 993 struct mbuf *mb; 994 u_int moff; 995 996 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 997 TCPSTAT_INC(tcps_sndprobe); 998 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 999 tp->t_sndrexmitpack++; 1000 TCPSTAT_INC(tcps_sndrexmitpack); 1001 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 1002 } else { 1003 TCPSTAT_INC(tcps_sndpack); 1004 TCPSTAT_ADD(tcps_sndbyte, len); 1005 } 1006 #ifdef INET6 1007 if (MHLEN < hdrlen + max_linkhdr) 1008 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1009 else 1010 #endif 1011 m = m_gethdr(M_NOWAIT, MT_DATA); 1012 1013 if (m == NULL) { 1014 SOCKBUF_UNLOCK(&so->so_snd); 1015 error = ENOBUFS; 1016 sack_rxmit = 0; 1017 goto out; 1018 } 1019 1020 m->m_data += max_linkhdr; 1021 m->m_len = hdrlen; 1022 1023 /* 1024 * Start the m_copy functions from the closest mbuf 1025 * to the offset in the socket buffer chain. 1026 */ 1027 mb = sbsndptr(&so->so_snd, off, len, &moff); 1028 1029 if (len <= MHLEN - hdrlen - max_linkhdr) { 1030 m_copydata(mb, moff, len, 1031 mtod(m, caddr_t) + hdrlen); 1032 m->m_len += len; 1033 } else { 1034 m->m_next = m_copym(mb, moff, len, M_NOWAIT); 1035 if (m->m_next == NULL) { 1036 SOCKBUF_UNLOCK(&so->so_snd); 1037 (void) m_free(m); 1038 error = ENOBUFS; 1039 sack_rxmit = 0; 1040 goto out; 1041 } 1042 } 1043 1044 /* 1045 * If we're sending everything we've got, set PUSH. 1046 * (This will keep happy those implementations which only 1047 * give data to the user when a buffer fills or 1048 * a PUSH comes in.) 1049 */ 1050 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) && 1051 !(flags & TH_SYN)) 1052 flags |= TH_PUSH; 1053 SOCKBUF_UNLOCK(&so->so_snd); 1054 } else { 1055 SOCKBUF_UNLOCK(&so->so_snd); 1056 if (tp->t_flags & TF_ACKNOW) 1057 TCPSTAT_INC(tcps_sndacks); 1058 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1059 TCPSTAT_INC(tcps_sndctrl); 1060 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1061 TCPSTAT_INC(tcps_sndurg); 1062 else 1063 TCPSTAT_INC(tcps_sndwinup); 1064 1065 m = m_gethdr(M_NOWAIT, MT_DATA); 1066 if (m == NULL) { 1067 error = ENOBUFS; 1068 sack_rxmit = 0; 1069 goto out; 1070 } 1071 #ifdef INET6 1072 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1073 MHLEN >= hdrlen) { 1074 M_ALIGN(m, hdrlen); 1075 } else 1076 #endif 1077 m->m_data += max_linkhdr; 1078 m->m_len = hdrlen; 1079 } 1080 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1081 m->m_pkthdr.rcvif = (struct ifnet *)0; 1082 #ifdef MAC 1083 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1084 #endif 1085 #ifdef INET6 1086 if (isipv6) { 1087 ip6 = mtod(m, struct ip6_hdr *); 1088 th = (struct tcphdr *)(ip6 + 1); 1089 tcpip_fillheaders(tp->t_inpcb, ip6, th); 1090 } else 1091 #endif /* INET6 */ 1092 { 1093 ip = mtod(m, struct ip *); 1094 ipov = (struct ipovly *)ip; 1095 th = (struct tcphdr *)(ip + 1); 1096 tcpip_fillheaders(tp->t_inpcb, ip, th); 1097 } 1098 1099 /* 1100 * Fill in fields, remembering maximum advertised 1101 * window for use in delaying messages about window sizes. 1102 * If resending a FIN, be sure not to use a new sequence number. 1103 */ 1104 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1105 tp->snd_nxt == tp->snd_max) 1106 tp->snd_nxt--; 1107 /* 1108 * If we are starting a connection, send ECN setup 1109 * SYN packet. If we are on a retransmit, we may 1110 * resend those bits a number of times as per 1111 * RFC 3168. 1112 */ 1113 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { 1114 if (tp->t_rxtshift >= 1) { 1115 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 1116 flags |= TH_ECE|TH_CWR; 1117 } else 1118 flags |= TH_ECE|TH_CWR; 1119 } 1120 1121 if (tp->t_state == TCPS_ESTABLISHED && 1122 (tp->t_flags & TF_ECN_PERMIT)) { 1123 /* 1124 * If the peer has ECN, mark data packets with 1125 * ECN capable transmission (ECT). 1126 * Ignore pure ack packets, retransmissions and window probes. 1127 */ 1128 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1129 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 1130 #ifdef INET6 1131 if (isipv6) 1132 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1133 else 1134 #endif 1135 ip->ip_tos |= IPTOS_ECN_ECT0; 1136 TCPSTAT_INC(tcps_ecn_ect0); 1137 } 1138 1139 /* 1140 * Reply with proper ECN notifications. 1141 */ 1142 if (tp->t_flags & TF_ECN_SND_CWR) { 1143 flags |= TH_CWR; 1144 tp->t_flags &= ~TF_ECN_SND_CWR; 1145 } 1146 if (tp->t_flags & TF_ECN_SND_ECE) 1147 flags |= TH_ECE; 1148 } 1149 1150 /* 1151 * If we are doing retransmissions, then snd_nxt will 1152 * not reflect the first unsent octet. For ACK only 1153 * packets, we do not want the sequence number of the 1154 * retransmitted packet, we want the sequence number 1155 * of the next unsent octet. So, if there is no data 1156 * (and no SYN or FIN), use snd_max instead of snd_nxt 1157 * when filling in ti_seq. But if we are in persist 1158 * state, snd_max might reflect one byte beyond the 1159 * right edge of the window, so use snd_nxt in that 1160 * case, since we know we aren't doing a retransmission. 1161 * (retransmit and persist are mutually exclusive...) 1162 */ 1163 if (sack_rxmit == 0) { 1164 if (len || (flags & (TH_SYN|TH_FIN)) || 1165 tcp_timer_active(tp, TT_PERSIST)) 1166 th->th_seq = htonl(tp->snd_nxt); 1167 else 1168 th->th_seq = htonl(tp->snd_max); 1169 } else { 1170 th->th_seq = htonl(p->rxmit); 1171 p->rxmit += len; 1172 tp->sackhint.sack_bytes_rexmit += len; 1173 } 1174 th->th_ack = htonl(tp->rcv_nxt); 1175 if (optlen) { 1176 bcopy(opt, th + 1, optlen); 1177 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1178 } 1179 th->th_flags = flags; 1180 /* 1181 * Calculate receive window. Don't shrink window, 1182 * but avoid silly window syndrome. 1183 */ 1184 if (recwin < (so->so_rcv.sb_hiwat / 4) && 1185 recwin < tp->t_maxseg) 1186 recwin = 0; 1187 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1188 recwin < (tp->rcv_adv - tp->rcv_nxt)) 1189 recwin = (tp->rcv_adv - tp->rcv_nxt); 1190 1191 /* 1192 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1193 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1194 * case is handled in syncache. 1195 */ 1196 if (flags & TH_SYN) 1197 th->th_win = htons((u_short) 1198 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1199 else 1200 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1201 1202 /* 1203 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1204 * a 0 window. This may cause the remote transmitter to stall. This 1205 * flag tells soreceive() to disable delayed acknowledgements when 1206 * draining the buffer. This can occur if the receiver is attempting 1207 * to read more data than can be buffered prior to transmitting on 1208 * the connection. 1209 */ 1210 if (th->th_win == 0) { 1211 tp->t_sndzerowin++; 1212 tp->t_flags |= TF_RXWIN0SENT; 1213 } else 1214 tp->t_flags &= ~TF_RXWIN0SENT; 1215 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1216 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1217 th->th_flags |= TH_URG; 1218 } else 1219 /* 1220 * If no urgent pointer to send, then we pull 1221 * the urgent pointer to the left edge of the send window 1222 * so that it doesn't drift into the send window on sequence 1223 * number wraparound. 1224 */ 1225 tp->snd_up = tp->snd_una; /* drag it along */ 1226 1227 /* 1228 * Put TCP length in extended header, and then 1229 * checksum extended header and data. 1230 */ 1231 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1232 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1233 1234 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1235 if (to.to_flags & TOF_SIGNATURE) { 1236 /* 1237 * Calculate MD5 signature and put it into the place 1238 * determined before. 1239 * NOTE: since TCP options buffer doesn't point into 1240 * mbuf's data, calculate offset and use it. 1241 */ 1242 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 1243 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 1244 /* 1245 * Do not send segment if the calculation of MD5 1246 * digest has failed. 1247 */ 1248 goto out; 1249 } 1250 } 1251 #endif 1252 #ifdef INET6 1253 if (isipv6) { 1254 /* 1255 * There is no need to fill in ip6_plen right now. 1256 * It will be filled later by ip6_output. 1257 */ 1258 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1259 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1260 optlen + len, IPPROTO_TCP, 0); 1261 } 1262 #endif 1263 #if defined(INET6) && defined(INET) 1264 else 1265 #endif 1266 #ifdef INET 1267 { 1268 m->m_pkthdr.csum_flags = CSUM_TCP; 1269 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1270 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1271 1272 /* IP version must be set here for ipv4/ipv6 checking later */ 1273 KASSERT(ip->ip_v == IPVERSION, 1274 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1275 } 1276 #endif 1277 1278 /* 1279 * Enable TSO and specify the size of the segments. 1280 * The TCP pseudo header checksum is always provided. 1281 */ 1282 if (tso) { 1283 KASSERT(len > tp->t_maxseg - optlen, 1284 ("%s: len <= tso_segsz", __func__)); 1285 m->m_pkthdr.csum_flags |= CSUM_TSO; 1286 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 1287 } 1288 1289 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 1290 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1291 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u", 1292 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1293 #else 1294 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1295 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u", 1296 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1297 #endif 1298 1299 #ifdef TCP_HHOOK 1300 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1301 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1302 #endif 1303 1304 #ifdef TCPDEBUG 1305 /* 1306 * Trace. 1307 */ 1308 if (so->so_options & SO_DEBUG) { 1309 u_short save = 0; 1310 #ifdef INET6 1311 if (!isipv6) 1312 #endif 1313 { 1314 save = ipov->ih_len; 1315 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1316 } 1317 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1318 #ifdef INET6 1319 if (!isipv6) 1320 #endif 1321 ipov->ih_len = save; 1322 } 1323 #endif /* TCPDEBUG */ 1324 TCP_PROBE3(debug__output, tp, th, m); 1325 1326 /* 1327 * Fill in IP length and desired time to live and 1328 * send to IP level. There should be a better way 1329 * to handle ttl and tos; we could keep them in 1330 * the template, but need a way to checksum without them. 1331 */ 1332 /* 1333 * m->m_pkthdr.len should have been set before checksum calculation, 1334 * because in6_cksum() need it. 1335 */ 1336 #ifdef INET6 1337 if (isipv6) { 1338 /* 1339 * we separately set hoplimit for every segment, since the 1340 * user might want to change the value via setsockopt. 1341 * Also, desired default hop limit might be changed via 1342 * Neighbor Discovery. 1343 */ 1344 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1345 1346 /* 1347 * Set the packet size here for the benefit of DTrace probes. 1348 * ip6_output() will set it properly; it's supposed to include 1349 * the option header lengths as well. 1350 */ 1351 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1352 1353 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 1354 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1355 else 1356 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1357 1358 if (tp->t_state == TCPS_SYN_SENT) 1359 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1360 1361 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1362 1363 #ifdef TCPPCAP 1364 /* Save packet, if requested. */ 1365 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1366 #endif 1367 1368 /* TODO: IPv6 IP6TOS_ECT bit on */ 1369 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, 1370 &tp->t_inpcb->inp_route6, 1371 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1372 NULL, NULL, tp->t_inpcb); 1373 1374 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL) 1375 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu; 1376 } 1377 #endif /* INET6 */ 1378 #if defined(INET) && defined(INET6) 1379 else 1380 #endif 1381 #ifdef INET 1382 { 1383 ip->ip_len = htons(m->m_pkthdr.len); 1384 #ifdef INET6 1385 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1386 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1387 #endif /* INET6 */ 1388 /* 1389 * If we do path MTU discovery, then we set DF on every packet. 1390 * This might not be the best thing to do according to RFC3390 1391 * Section 2. However the tcp hostcache migitates the problem 1392 * so it affects only the first tcp connection with a host. 1393 * 1394 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1395 */ 1396 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 1397 ip->ip_off |= htons(IP_DF); 1398 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1399 } else { 1400 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1401 } 1402 1403 if (tp->t_state == TCPS_SYN_SENT) 1404 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1405 1406 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1407 1408 #ifdef TCPPCAP 1409 /* Save packet, if requested. */ 1410 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1411 #endif 1412 1413 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route, 1414 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1415 tp->t_inpcb); 1416 1417 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL) 1418 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu; 1419 } 1420 #endif /* INET */ 1421 1422 out: 1423 /* 1424 * In transmit state, time the transmission and arrange for 1425 * the retransmit. In persist state, just set snd_max. 1426 */ 1427 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1428 !tcp_timer_active(tp, TT_PERSIST)) { 1429 tcp_seq startseq = tp->snd_nxt; 1430 1431 /* 1432 * Advance snd_nxt over sequence space of this segment. 1433 */ 1434 if (flags & (TH_SYN|TH_FIN)) { 1435 if (flags & TH_SYN) 1436 tp->snd_nxt++; 1437 if (flags & TH_FIN) { 1438 tp->snd_nxt++; 1439 tp->t_flags |= TF_SENTFIN; 1440 } 1441 } 1442 if (sack_rxmit) 1443 goto timer; 1444 tp->snd_nxt += len; 1445 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1446 tp->snd_max = tp->snd_nxt; 1447 /* 1448 * Time this transmission if not a retransmission and 1449 * not currently timing anything. 1450 */ 1451 if (tp->t_rtttime == 0) { 1452 tp->t_rtttime = ticks; 1453 tp->t_rtseq = startseq; 1454 TCPSTAT_INC(tcps_segstimed); 1455 } 1456 } 1457 1458 /* 1459 * Set retransmit timer if not currently set, 1460 * and not doing a pure ack or a keep-alive probe. 1461 * Initial value for retransmit timer is smoothed 1462 * round-trip time + 2 * round-trip time variance. 1463 * Initialize shift counter which is used for backoff 1464 * of retransmit time. 1465 */ 1466 timer: 1467 if (!tcp_timer_active(tp, TT_REXMT) && 1468 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1469 (tp->snd_nxt != tp->snd_una))) { 1470 if (tcp_timer_active(tp, TT_PERSIST)) { 1471 tcp_timer_activate(tp, TT_PERSIST, 0); 1472 tp->t_rxtshift = 0; 1473 } 1474 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1475 } else if (len == 0 && sbavail(&so->so_snd) && 1476 !tcp_timer_active(tp, TT_REXMT) && 1477 !tcp_timer_active(tp, TT_PERSIST)) { 1478 /* 1479 * Avoid a situation where we do not set persist timer 1480 * after a zero window condition. For example: 1481 * 1) A -> B: packet with enough data to fill the window 1482 * 2) B -> A: ACK for #1 + new data (0 window 1483 * advertisement) 1484 * 3) A -> B: ACK for #2, 0 len packet 1485 * 1486 * In this case, A will not activate the persist timer, 1487 * because it chose to send a packet. Unless tcp_output 1488 * is called for some other reason (delayed ack timer, 1489 * another input packet from B, socket syscall), A will 1490 * not send zero window probes. 1491 * 1492 * So, if you send a 0-length packet, but there is data 1493 * in the socket buffer, and neither the rexmt or 1494 * persist timer is already set, then activate the 1495 * persist timer. 1496 */ 1497 tp->t_rxtshift = 0; 1498 tcp_setpersist(tp); 1499 } 1500 } else { 1501 /* 1502 * Persist case, update snd_max but since we are in 1503 * persist mode (no window) we do not update snd_nxt. 1504 */ 1505 int xlen = len; 1506 if (flags & TH_SYN) 1507 ++xlen; 1508 if (flags & TH_FIN) { 1509 ++xlen; 1510 tp->t_flags |= TF_SENTFIN; 1511 } 1512 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1513 tp->snd_max = tp->snd_nxt + xlen; 1514 } 1515 1516 if (error) { 1517 1518 /* 1519 * We know that the packet was lost, so back out the 1520 * sequence number advance, if any. 1521 * 1522 * If the error is EPERM the packet got blocked by the 1523 * local firewall. Normally we should terminate the 1524 * connection but the blocking may have been spurious 1525 * due to a firewall reconfiguration cycle. So we treat 1526 * it like a packet loss and let the retransmit timer and 1527 * timeouts do their work over time. 1528 * XXX: It is a POLA question whether calling tcp_drop right 1529 * away would be the really correct behavior instead. 1530 */ 1531 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1532 !tcp_timer_active(tp, TT_PERSIST)) && 1533 ((flags & TH_SYN) == 0) && 1534 (error != EPERM)) { 1535 if (sack_rxmit) { 1536 p->rxmit -= len; 1537 tp->sackhint.sack_bytes_rexmit -= len; 1538 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1539 ("sackhint bytes rtx >= 0")); 1540 } else 1541 tp->snd_nxt -= len; 1542 } 1543 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1544 switch (error) { 1545 case EACCES: 1546 tp->t_softerror = error; 1547 return (0); 1548 case EPERM: 1549 tp->t_softerror = error; 1550 return (error); 1551 case ENOBUFS: 1552 TCP_XMIT_TIMER_ASSERT(tp, len, flags); 1553 tp->snd_cwnd = tp->t_maxseg; 1554 return (0); 1555 case EMSGSIZE: 1556 /* 1557 * For some reason the interface we used initially 1558 * to send segments changed to another or lowered 1559 * its MTU. 1560 * If TSO was active we either got an interface 1561 * without TSO capabilits or TSO was turned off. 1562 * If we obtained mtu from ip_output() then update 1563 * it and try again. 1564 */ 1565 if (tso) 1566 tp->t_flags &= ~TF_TSO; 1567 if (mtu != 0) { 1568 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1569 goto again; 1570 } 1571 return (error); 1572 case EHOSTDOWN: 1573 case EHOSTUNREACH: 1574 case ENETDOWN: 1575 case ENETUNREACH: 1576 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1577 tp->t_softerror = error; 1578 return (0); 1579 } 1580 /* FALLTHROUGH */ 1581 default: 1582 return (error); 1583 } 1584 } 1585 TCPSTAT_INC(tcps_sndtotal); 1586 1587 /* 1588 * Data sent (as far as we can tell). 1589 * If this advertises a larger window than any other segment, 1590 * then remember the size of the advertised window. 1591 * Any pending ACK has now been sent. 1592 */ 1593 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1594 tp->rcv_adv = tp->rcv_nxt + recwin; 1595 tp->last_ack_sent = tp->rcv_nxt; 1596 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1597 if (tcp_timer_active(tp, TT_DELACK)) 1598 tcp_timer_activate(tp, TT_DELACK, 0); 1599 #if 0 1600 /* 1601 * This completely breaks TCP if newreno is turned on. What happens 1602 * is that if delayed-acks are turned on on the receiver, this code 1603 * on the transmitter effectively destroys the TCP window, forcing 1604 * it to four packets (1.5Kx4 = 6K window). 1605 */ 1606 if (sendalot && --maxburst) 1607 goto again; 1608 #endif 1609 if (sendalot) 1610 goto again; 1611 return (0); 1612 } 1613 1614 void 1615 tcp_setpersist(struct tcpcb *tp) 1616 { 1617 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1618 int tt; 1619 1620 tp->t_flags &= ~TF_PREVVALID; 1621 if (tcp_timer_active(tp, TT_REXMT)) 1622 panic("tcp_setpersist: retransmit pending"); 1623 /* 1624 * Start/restart persistence timer. 1625 */ 1626 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1627 tcp_persmin, tcp_persmax); 1628 tcp_timer_activate(tp, TT_PERSIST, tt); 1629 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1630 tp->t_rxtshift++; 1631 } 1632 1633 /* 1634 * Insert TCP options according to the supplied parameters to the place 1635 * optp in a consistent way. Can handle unaligned destinations. 1636 * 1637 * The order of the option processing is crucial for optimal packing and 1638 * alignment for the scarce option space. 1639 * 1640 * The optimal order for a SYN/SYN-ACK segment is: 1641 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1642 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1643 * 1644 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1645 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1646 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1647 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1648 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1649 */ 1650 int 1651 tcp_addoptions(struct tcpopt *to, u_char *optp) 1652 { 1653 u_int32_t mask, optlen = 0; 1654 1655 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1656 if ((to->to_flags & mask) != mask) 1657 continue; 1658 if (optlen == TCP_MAXOLEN) 1659 break; 1660 switch (to->to_flags & mask) { 1661 case TOF_MSS: 1662 while (optlen % 4) { 1663 optlen += TCPOLEN_NOP; 1664 *optp++ = TCPOPT_NOP; 1665 } 1666 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1667 continue; 1668 optlen += TCPOLEN_MAXSEG; 1669 *optp++ = TCPOPT_MAXSEG; 1670 *optp++ = TCPOLEN_MAXSEG; 1671 to->to_mss = htons(to->to_mss); 1672 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1673 optp += sizeof(to->to_mss); 1674 break; 1675 case TOF_SCALE: 1676 while (!optlen || optlen % 2 != 1) { 1677 optlen += TCPOLEN_NOP; 1678 *optp++ = TCPOPT_NOP; 1679 } 1680 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1681 continue; 1682 optlen += TCPOLEN_WINDOW; 1683 *optp++ = TCPOPT_WINDOW; 1684 *optp++ = TCPOLEN_WINDOW; 1685 *optp++ = to->to_wscale; 1686 break; 1687 case TOF_SACKPERM: 1688 while (optlen % 2) { 1689 optlen += TCPOLEN_NOP; 1690 *optp++ = TCPOPT_NOP; 1691 } 1692 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1693 continue; 1694 optlen += TCPOLEN_SACK_PERMITTED; 1695 *optp++ = TCPOPT_SACK_PERMITTED; 1696 *optp++ = TCPOLEN_SACK_PERMITTED; 1697 break; 1698 case TOF_TS: 1699 while (!optlen || optlen % 4 != 2) { 1700 optlen += TCPOLEN_NOP; 1701 *optp++ = TCPOPT_NOP; 1702 } 1703 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1704 continue; 1705 optlen += TCPOLEN_TIMESTAMP; 1706 *optp++ = TCPOPT_TIMESTAMP; 1707 *optp++ = TCPOLEN_TIMESTAMP; 1708 to->to_tsval = htonl(to->to_tsval); 1709 to->to_tsecr = htonl(to->to_tsecr); 1710 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1711 optp += sizeof(to->to_tsval); 1712 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1713 optp += sizeof(to->to_tsecr); 1714 break; 1715 case TOF_SIGNATURE: 1716 { 1717 int siglen = TCPOLEN_SIGNATURE - 2; 1718 1719 while (!optlen || optlen % 4 != 2) { 1720 optlen += TCPOLEN_NOP; 1721 *optp++ = TCPOPT_NOP; 1722 } 1723 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) { 1724 to->to_flags &= ~TOF_SIGNATURE; 1725 continue; 1726 } 1727 optlen += TCPOLEN_SIGNATURE; 1728 *optp++ = TCPOPT_SIGNATURE; 1729 *optp++ = TCPOLEN_SIGNATURE; 1730 to->to_signature = optp; 1731 while (siglen--) 1732 *optp++ = 0; 1733 break; 1734 } 1735 case TOF_SACK: 1736 { 1737 int sackblks = 0; 1738 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1739 tcp_seq sack_seq; 1740 1741 while (!optlen || optlen % 4 != 2) { 1742 optlen += TCPOLEN_NOP; 1743 *optp++ = TCPOPT_NOP; 1744 } 1745 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1746 continue; 1747 optlen += TCPOLEN_SACKHDR; 1748 *optp++ = TCPOPT_SACK; 1749 sackblks = min(to->to_nsacks, 1750 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1751 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1752 while (sackblks--) { 1753 sack_seq = htonl(sack->start); 1754 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1755 optp += sizeof(sack_seq); 1756 sack_seq = htonl(sack->end); 1757 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1758 optp += sizeof(sack_seq); 1759 optlen += TCPOLEN_SACK; 1760 sack++; 1761 } 1762 TCPSTAT_INC(tcps_sack_send_blocks); 1763 break; 1764 } 1765 #ifdef TCP_RFC7413 1766 case TOF_FASTOPEN: 1767 { 1768 int total_len; 1769 1770 /* XXX is there any point to aligning this option? */ 1771 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; 1772 if (TCP_MAXOLEN - optlen < total_len) 1773 continue; 1774 *optp++ = TCPOPT_FAST_OPEN; 1775 *optp++ = total_len; 1776 if (to->to_tfo_len > 0) { 1777 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); 1778 optp += to->to_tfo_len; 1779 } 1780 optlen += total_len; 1781 break; 1782 } 1783 #endif 1784 default: 1785 panic("%s: unknown TCP option type", __func__); 1786 break; 1787 } 1788 } 1789 1790 /* Terminate and pad TCP options to a 4 byte boundary. */ 1791 if (optlen % 4) { 1792 optlen += TCPOLEN_EOL; 1793 *optp++ = TCPOPT_EOL; 1794 } 1795 /* 1796 * According to RFC 793 (STD0007): 1797 * "The content of the header beyond the End-of-Option option 1798 * must be header padding (i.e., zero)." 1799 * and later: "The padding is composed of zeros." 1800 */ 1801 while (optlen % 4) { 1802 optlen += TCPOLEN_PAD; 1803 *optp++ = TCPOPT_PAD; 1804 } 1805 1806 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1807 return (optlen); 1808 } 1809 1810 void 1811 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin) 1812 { 1813 1814 /* 1815 * Automatic sizing of send socket buffer. Often the send buffer 1816 * size is not optimally adjusted to the actual network conditions 1817 * at hand (delay bandwidth product). Setting the buffer size too 1818 * small limits throughput on links with high bandwidth and high 1819 * delay (eg. trans-continental/oceanic links). Setting the 1820 * buffer size too big consumes too much real kernel memory, 1821 * especially with many connections on busy servers. 1822 * 1823 * The criteria to step up the send buffer one notch are: 1824 * 1. receive window of remote host is larger than send buffer 1825 * (with a fudge factor of 5/4th); 1826 * 2. send buffer is filled to 7/8th with data (so we actually 1827 * have data to make use of it); 1828 * 3. send buffer fill has not hit maximal automatic size; 1829 * 4. our send window (slow start and cogestion controlled) is 1830 * larger than sent but unacknowledged data in send buffer. 1831 * 1832 * The remote host receive window scaling factor may limit the 1833 * growing of the send buffer before it reaches its allowed 1834 * maximum. 1835 * 1836 * It scales directly with slow start or congestion window 1837 * and does at most one step per received ACK. This fast 1838 * scaling has the drawback of growing the send buffer beyond 1839 * what is strictly necessary to make full use of a given 1840 * delay*bandwidth product. However testing has shown this not 1841 * to be much of an problem. At worst we are trading wasting 1842 * of available bandwidth (the non-use of it) for wasting some 1843 * socket buffer memory. 1844 * 1845 * TODO: Shrink send buffer during idle periods together 1846 * with congestion window. Requires another timer. Has to 1847 * wait for upcoming tcp timer rewrite. 1848 * 1849 * XXXGL: should there be used sbused() or sbavail()? 1850 */ 1851 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 1852 int lowat; 1853 1854 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0; 1855 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat && 1856 sbused(&so->so_snd) >= 1857 (so->so_snd.sb_hiwat / 8 * 7) - lowat && 1858 sbused(&so->so_snd) < V_tcp_autosndbuf_max && 1859 sendwin >= (sbused(&so->so_snd) - 1860 (tp->snd_nxt - tp->snd_una))) { 1861 if (!sbreserve_locked(&so->so_snd, 1862 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 1863 V_tcp_autosndbuf_max), so, curthread)) 1864 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 1865 } 1866 } 1867 } 1868