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