1 /*- 2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 3 * The Regents of the University of California. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 4. Neither the name of the University nor the names of its contributors 14 * may be used to endorse or promote products derived from this software 15 * without specific prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 * 29 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 30 */ 31 32 #include <sys/cdefs.h> 33 __FBSDID("$FreeBSD$"); 34 35 #include "opt_inet.h" 36 #include "opt_inet6.h" 37 #include "opt_ipsec.h" 38 #include "opt_tcpdebug.h" 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/domain.h> 43 #include <sys/hhook.h> 44 #include <sys/kernel.h> 45 #include <sys/lock.h> 46 #include <sys/mbuf.h> 47 #include <sys/mutex.h> 48 #include <sys/protosw.h> 49 #include <sys/sdt.h> 50 #include <sys/socket.h> 51 #include <sys/socketvar.h> 52 #include <sys/sysctl.h> 53 54 #include <net/if.h> 55 #include <net/route.h> 56 #include <net/vnet.h> 57 58 #include <netinet/cc.h> 59 #include <netinet/in.h> 60 #include <netinet/in_kdtrace.h> 61 #include <netinet/in_systm.h> 62 #include <netinet/ip.h> 63 #include <netinet/in_pcb.h> 64 #include <netinet/ip_var.h> 65 #include <netinet/ip_options.h> 66 #ifdef INET6 67 #include <netinet6/in6_pcb.h> 68 #include <netinet/ip6.h> 69 #include <netinet6/ip6_var.h> 70 #endif 71 #ifdef TCP_RFC7413 72 #include <netinet/tcp_fastopen.h> 73 #endif 74 #define TCPOUTFLAGS 75 #include <netinet/tcp_fsm.h> 76 #include <netinet/tcp_seq.h> 77 #include <netinet/tcp_timer.h> 78 #include <netinet/tcp_var.h> 79 #include <netinet/tcpip.h> 80 #ifdef TCPPCAP 81 #include <netinet/tcp_pcap.h> 82 #endif 83 #ifdef TCPDEBUG 84 #include <netinet/tcp_debug.h> 85 #endif 86 #ifdef TCP_OFFLOAD 87 #include <netinet/tcp_offload.h> 88 #endif 89 90 #ifdef IPSEC 91 #include <netipsec/ipsec.h> 92 #endif /*IPSEC*/ 93 94 #include <machine/in_cksum.h> 95 96 #include <security/mac/mac_framework.h> 97 98 VNET_DEFINE(int, path_mtu_discovery) = 1; 99 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW, 100 &VNET_NAME(path_mtu_discovery), 1, 101 "Enable Path MTU Discovery"); 102 103 VNET_DEFINE(int, tcp_do_tso) = 1; 104 #define V_tcp_do_tso VNET(tcp_do_tso) 105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW, 106 &VNET_NAME(tcp_do_tso), 0, 107 "Enable TCP Segmentation Offload"); 108 109 VNET_DEFINE(int, tcp_sendspace) = 1024*32; 110 #define V_tcp_sendspace VNET(tcp_sendspace) 111 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW, 112 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 113 114 VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 115 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) 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 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) 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 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) 128 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW, 129 &VNET_NAME(tcp_autosndbuf_max), 0, 130 "Max size of automatic send buffer"); 131 132 static void inline hhook_run_tcp_est_out(struct tcpcb *tp, 133 struct tcphdr *th, struct tcpopt *to, 134 long len, int tso); 135 static void inline cc_after_idle(struct tcpcb *tp); 136 137 /* 138 * Wrapper for the TCP established output helper hook. 139 */ 140 static void inline 141 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 142 struct tcpopt *to, long len, int tso) 143 { 144 struct tcp_hhook_data hhook_data; 145 146 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 147 hhook_data.tp = tp; 148 hhook_data.th = th; 149 hhook_data.to = to; 150 hhook_data.len = len; 151 hhook_data.tso = tso; 152 153 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 154 tp->osd); 155 } 156 } 157 158 /* 159 * CC wrapper hook functions 160 */ 161 static void inline 162 cc_after_idle(struct tcpcb *tp) 163 { 164 INP_WLOCK_ASSERT(tp->t_inpcb); 165 166 if (CC_ALGO(tp)->after_idle != NULL) 167 CC_ALGO(tp)->after_idle(tp->ccv); 168 } 169 170 /* 171 * Tcp output routine: figure out what should be sent and send it. 172 */ 173 int 174 tcp_output(struct tcpcb *tp) 175 { 176 struct socket *so = tp->t_inpcb->inp_socket; 177 long len, recwin, sendwin; 178 int off, flags, error = 0; /* Keep compiler happy */ 179 struct mbuf *m; 180 struct ip *ip = NULL; 181 struct ipovly *ipov = NULL; 182 struct tcphdr *th; 183 u_char opt[TCP_MAXOLEN]; 184 unsigned ipoptlen, optlen, hdrlen; 185 #ifdef IPSEC 186 unsigned ipsec_optlen = 0; 187 #endif 188 int idle, sendalot; 189 int sack_rxmit, sack_bytes_rxmt; 190 struct sackhole *p; 191 int tso, mtu; 192 struct tcpopt to; 193 #if 0 194 int maxburst = TCP_MAXBURST; 195 #endif 196 #ifdef INET6 197 struct ip6_hdr *ip6 = NULL; 198 int isipv6; 199 200 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 201 #endif 202 203 INP_WLOCK_ASSERT(tp->t_inpcb); 204 205 #ifdef TCP_OFFLOAD 206 if (tp->t_flags & TF_TOE) 207 return (tcp_offload_output(tp)); 208 #endif 209 210 #ifdef TCP_RFC7413 211 /* 212 * For TFO connections in SYN_RECEIVED, only allow the initial 213 * SYN|ACK and those sent by the retransmit timer. 214 */ 215 if ((tp->t_flags & TF_FASTOPEN) && 216 (tp->t_state == TCPS_SYN_RECEIVED) && 217 SEQ_GT(tp->snd_max, tp->snd_una) && /* inital SYN|ACK sent */ 218 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 219 return (0); 220 #endif 221 /* 222 * Determine length of data that should be transmitted, 223 * and flags that will be used. 224 * If there is some data or critical controls (SYN, RST) 225 * to send, then transmit; otherwise, investigate further. 226 */ 227 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 228 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) 229 cc_after_idle(tp); 230 tp->t_flags &= ~TF_LASTIDLE; 231 if (idle) { 232 if (tp->t_flags & TF_MORETOCOME) { 233 tp->t_flags |= TF_LASTIDLE; 234 idle = 0; 235 } 236 } 237 again: 238 /* 239 * If we've recently taken a timeout, snd_max will be greater than 240 * snd_nxt. There may be SACK information that allows us to avoid 241 * resending already delivered data. Adjust snd_nxt accordingly. 242 */ 243 if ((tp->t_flags & TF_SACK_PERMIT) && 244 SEQ_LT(tp->snd_nxt, tp->snd_max)) 245 tcp_sack_adjust(tp); 246 sendalot = 0; 247 tso = 0; 248 mtu = 0; 249 off = tp->snd_nxt - tp->snd_una; 250 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 251 252 flags = tcp_outflags[tp->t_state]; 253 /* 254 * Send any SACK-generated retransmissions. If we're explicitly trying 255 * to send out new data (when sendalot is 1), bypass this function. 256 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 257 * we're replacing a (future) new transmission with a retransmission 258 * now, and we previously incremented snd_cwnd in tcp_input(). 259 */ 260 /* 261 * Still in sack recovery , reset rxmit flag to zero. 262 */ 263 sack_rxmit = 0; 264 sack_bytes_rxmt = 0; 265 len = 0; 266 p = NULL; 267 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 268 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 269 long cwin; 270 271 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 272 if (cwin < 0) 273 cwin = 0; 274 /* Do not retransmit SACK segments beyond snd_recover */ 275 if (SEQ_GT(p->end, tp->snd_recover)) { 276 /* 277 * (At least) part of sack hole extends beyond 278 * snd_recover. Check to see if we can rexmit data 279 * for this hole. 280 */ 281 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 282 /* 283 * Can't rexmit any more data for this hole. 284 * That data will be rexmitted in the next 285 * sack recovery episode, when snd_recover 286 * moves past p->rxmit. 287 */ 288 p = NULL; 289 goto after_sack_rexmit; 290 } else 291 /* Can rexmit part of the current hole */ 292 len = ((long)ulmin(cwin, 293 tp->snd_recover - p->rxmit)); 294 } else 295 len = ((long)ulmin(cwin, p->end - p->rxmit)); 296 off = p->rxmit - tp->snd_una; 297 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 298 __func__, off)); 299 if (len > 0) { 300 sack_rxmit = 1; 301 sendalot = 1; 302 TCPSTAT_INC(tcps_sack_rexmits); 303 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 304 min(len, tp->t_maxseg)); 305 } 306 } 307 after_sack_rexmit: 308 /* 309 * Get standard flags, and add SYN or FIN if requested by 'hidden' 310 * state flags. 311 */ 312 if (tp->t_flags & TF_NEEDFIN) 313 flags |= TH_FIN; 314 if (tp->t_flags & TF_NEEDSYN) 315 flags |= TH_SYN; 316 317 SOCKBUF_LOCK(&so->so_snd); 318 /* 319 * If in persist timeout with window of 0, send 1 byte. 320 * Otherwise, if window is small but nonzero 321 * and timer expired, we will send what we can 322 * and go to transmit state. 323 */ 324 if (tp->t_flags & TF_FORCEDATA) { 325 if (sendwin == 0) { 326 /* 327 * If we still have some data to send, then 328 * clear the FIN bit. Usually this would 329 * happen below when it realizes that we 330 * aren't sending all the data. However, 331 * if we have exactly 1 byte of unsent data, 332 * then it won't clear the FIN bit below, 333 * and if we are in persist state, we wind 334 * up sending the packet without recording 335 * that we sent the FIN bit. 336 * 337 * We can't just blindly clear the FIN bit, 338 * because if we don't have any more data 339 * to send then the probe will be the FIN 340 * itself. 341 */ 342 if (off < sbused(&so->so_snd)) 343 flags &= ~TH_FIN; 344 sendwin = 1; 345 } else { 346 tcp_timer_activate(tp, TT_PERSIST, 0); 347 tp->t_rxtshift = 0; 348 } 349 } 350 351 /* 352 * If snd_nxt == snd_max and we have transmitted a FIN, the 353 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 354 * a negative length. This can also occur when TCP opens up 355 * its congestion window while receiving additional duplicate 356 * acks after fast-retransmit because TCP will reset snd_nxt 357 * to snd_max after the fast-retransmit. 358 * 359 * In the normal retransmit-FIN-only case, however, snd_nxt will 360 * be set to snd_una, the offset will be 0, and the length may 361 * wind up 0. 362 * 363 * If sack_rxmit is true we are retransmitting from the scoreboard 364 * in which case len is already set. 365 */ 366 if (sack_rxmit == 0) { 367 if (sack_bytes_rxmt == 0) 368 len = ((long)ulmin(sbavail(&so->so_snd), sendwin) - 369 off); 370 else { 371 long cwin; 372 373 /* 374 * We are inside of a SACK recovery episode and are 375 * sending new data, having retransmitted all the 376 * data possible in the scoreboard. 377 */ 378 len = ((long)ulmin(sbavail(&so->so_snd), tp->snd_wnd) - 379 off); 380 /* 381 * Don't remove this (len > 0) check ! 382 * We explicitly check for len > 0 here (although it 383 * isn't really necessary), to work around a gcc 384 * optimization issue - to force gcc to compute 385 * len above. Without this check, the computation 386 * of len is bungled by the optimizer. 387 */ 388 if (len > 0) { 389 cwin = tp->snd_cwnd - 390 (tp->snd_nxt - tp->sack_newdata) - 391 sack_bytes_rxmt; 392 if (cwin < 0) 393 cwin = 0; 394 len = lmin(len, cwin); 395 } 396 } 397 } 398 399 /* 400 * Lop off SYN bit if it has already been sent. However, if this 401 * is SYN-SENT state and if segment contains data and if we don't 402 * know that foreign host supports TAO, suppress sending segment. 403 */ 404 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 405 if (tp->t_state != TCPS_SYN_RECEIVED) 406 flags &= ~TH_SYN; 407 #ifdef TCP_RFC7413 408 /* 409 * When sending additional segments following a TFO SYN|ACK, 410 * do not include the SYN bit. 411 */ 412 if ((tp->t_flags & TF_FASTOPEN) && 413 (tp->t_state == TCPS_SYN_RECEIVED)) 414 flags &= ~TH_SYN; 415 #endif 416 off--, len++; 417 } 418 419 /* 420 * Be careful not to send data and/or FIN on SYN segments. 421 * This measure is needed to prevent interoperability problems 422 * with not fully conformant TCP implementations. 423 */ 424 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 425 len = 0; 426 flags &= ~TH_FIN; 427 } 428 429 #ifdef TCP_RFC7413 430 /* 431 * When retransmitting SYN|ACK on a passively-created TFO socket, 432 * don't include data, as the presence of data may have caused the 433 * original SYN|ACK to have been dropped by a middlebox. 434 */ 435 if ((tp->t_flags & TF_FASTOPEN) && 436 (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) || 437 (flags & TH_RST))) 438 len = 0; 439 #endif 440 if (len <= 0) { 441 /* 442 * If FIN has been sent but not acked, 443 * but we haven't been called to retransmit, 444 * len will be < 0. Otherwise, window shrank 445 * after we sent into it. If window shrank to 0, 446 * cancel pending retransmit, pull snd_nxt back 447 * to (closed) window, and set the persist timer 448 * if it isn't already going. If the window didn't 449 * close completely, just wait for an ACK. 450 * 451 * We also do a general check here to ensure that 452 * we will set the persist timer when we have data 453 * to send, but a 0-byte window. This makes sure 454 * the persist timer is set even if the packet 455 * hits one of the "goto send" lines below. 456 */ 457 len = 0; 458 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 459 (off < (int) sbavail(&so->so_snd))) { 460 tcp_timer_activate(tp, TT_REXMT, 0); 461 tp->t_rxtshift = 0; 462 tp->snd_nxt = tp->snd_una; 463 if (!tcp_timer_active(tp, TT_PERSIST)) 464 tcp_setpersist(tp); 465 } 466 } 467 468 /* len will be >= 0 after this point. */ 469 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 470 471 /* 472 * Automatic sizing of send socket buffer. Often the send buffer 473 * size is not optimally adjusted to the actual network conditions 474 * at hand (delay bandwidth product). Setting the buffer size too 475 * small limits throughput on links with high bandwidth and high 476 * delay (eg. trans-continental/oceanic links). Setting the 477 * buffer size too big consumes too much real kernel memory, 478 * especially with many connections on busy servers. 479 * 480 * The criteria to step up the send buffer one notch are: 481 * 1. receive window of remote host is larger than send buffer 482 * (with a fudge factor of 5/4th); 483 * 2. send buffer is filled to 7/8th with data (so we actually 484 * have data to make use of it); 485 * 3. send buffer fill has not hit maximal automatic size; 486 * 4. our send window (slow start and cogestion controlled) is 487 * larger than sent but unacknowledged data in send buffer. 488 * 489 * The remote host receive window scaling factor may limit the 490 * growing of the send buffer before it reaches its allowed 491 * maximum. 492 * 493 * It scales directly with slow start or congestion window 494 * and does at most one step per received ACK. This fast 495 * scaling has the drawback of growing the send buffer beyond 496 * what is strictly necessary to make full use of a given 497 * delay*bandwith product. However testing has shown this not 498 * to be much of an problem. At worst we are trading wasting 499 * of available bandwith (the non-use of it) for wasting some 500 * socket buffer memory. 501 * 502 * TODO: Shrink send buffer during idle periods together 503 * with congestion window. Requires another timer. Has to 504 * wait for upcoming tcp timer rewrite. 505 * 506 * XXXGL: should there be used sbused() or sbavail()? 507 */ 508 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 509 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 510 sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) && 511 sbused(&so->so_snd) < V_tcp_autosndbuf_max && 512 sendwin >= (sbused(&so->so_snd) - 513 (tp->snd_nxt - tp->snd_una))) { 514 if (!sbreserve_locked(&so->so_snd, 515 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 516 V_tcp_autosndbuf_max), so, curthread)) 517 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 518 } 519 } 520 521 /* 522 * Decide if we can use TCP Segmentation Offloading (if supported by 523 * hardware). 524 * 525 * TSO may only be used if we are in a pure bulk sending state. The 526 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 527 * IP options prevent using TSO. With TSO the TCP header is the same 528 * (except for the sequence number) for all generated packets. This 529 * makes it impossible to transmit any options which vary per generated 530 * segment or packet. 531 */ 532 #ifdef IPSEC 533 /* 534 * Pre-calculate here as we save another lookup into the darknesses 535 * of IPsec that way and can actually decide if TSO is ok. 536 */ 537 ipsec_optlen = ipsec_hdrsiz_tcp(tp); 538 #endif 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 #ifdef IPSEC 543 ipsec_optlen == 0 && 544 #endif 545 tp->t_inpcb->inp_options == NULL && 546 tp->t_inpcb->in6p_options == NULL) 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 = sbspace(&so->so_rcv); 559 560 /* 561 * Sender silly window avoidance. We transmit under the following 562 * conditions when len is non-zero: 563 * 564 * - We have a full segment (or more with TSO) 565 * - This is the last buffer in a write()/send() and we are 566 * either idle or running NODELAY 567 * - we've timed out (e.g. persist timer) 568 * - we have more then 1/2 the maximum send window's worth of 569 * data (receiver may be limited the window size) 570 * - we need to retransmit 571 */ 572 if (len) { 573 if (len >= tp->t_maxseg) 574 goto send; 575 /* 576 * NOTE! on localhost connections an 'ack' from the remote 577 * end may occur synchronously with the output and cause 578 * us to flush a buffer queued with moretocome. XXX 579 * 580 * note: the len + off check is almost certainly unnecessary. 581 */ 582 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 583 (idle || (tp->t_flags & TF_NODELAY)) && 584 len + off >= sbavail(&so->so_snd) && 585 (tp->t_flags & TF_NOPUSH) == 0) { 586 goto send; 587 } 588 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 589 goto send; 590 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 591 goto send; 592 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 593 goto send; 594 if (sack_rxmit) 595 goto send; 596 } 597 598 /* 599 * Sending of standalone window updates. 600 * 601 * Window updates are important when we close our window due to a 602 * full socket buffer and are opening it again after the application 603 * reads data from it. Once the window has opened again and the 604 * remote end starts to send again the ACK clock takes over and 605 * provides the most current window information. 606 * 607 * We must avoid the silly window syndrome whereas every read 608 * from the receive buffer, no matter how small, causes a window 609 * update to be sent. We also should avoid sending a flurry of 610 * window updates when the socket buffer had queued a lot of data 611 * and the application is doing small reads. 612 * 613 * Prevent a flurry of pointless window updates by only sending 614 * an update when we can increase the advertized window by more 615 * than 1/4th of the socket buffer capacity. When the buffer is 616 * getting full or is very small be more aggressive and send an 617 * update whenever we can increase by two mss sized segments. 618 * In all other situations the ACK's to new incoming data will 619 * carry further window increases. 620 * 621 * Don't send an independent window update if a delayed 622 * ACK is pending (it will get piggy-backed on it) or the 623 * remote side already has done a half-close and won't send 624 * more data. Skip this if the connection is in T/TCP 625 * half-open state. 626 */ 627 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 628 !(tp->t_flags & TF_DELACK) && 629 !TCPS_HAVERCVDFIN(tp->t_state)) { 630 /* 631 * "adv" is the amount we could increase the window, 632 * taking into account that we are limited by 633 * TCP_MAXWIN << tp->rcv_scale. 634 */ 635 long adv; 636 int oldwin; 637 638 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 639 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 640 oldwin = (tp->rcv_adv - tp->rcv_nxt); 641 adv -= oldwin; 642 } else 643 oldwin = 0; 644 645 /* 646 * If the new window size ends up being the same as the old 647 * size when it is scaled, then don't force a window update. 648 */ 649 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 650 goto dontupdate; 651 652 if (adv >= (long)(2 * tp->t_maxseg) && 653 (adv >= (long)(so->so_rcv.sb_hiwat / 4) || 654 recwin <= (long)(so->so_rcv.sb_hiwat / 8) || 655 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 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 if ((tp->t_flags & TF_NOOPT) == 0) { 756 to.to_flags = 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 ((tp->t_flags & TF_FASTOPEN) && 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 /* Set receive buffer autosizing timestamp. */ 791 if (tp->rfbuf_ts == 0 && 792 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 793 tp->rfbuf_ts = tcp_ts_getticks(); 794 } 795 /* Selective ACK's. */ 796 if (tp->t_flags & TF_SACK_PERMIT) { 797 if (flags & TH_SYN) 798 to.to_flags |= TOF_SACKPERM; 799 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 800 (tp->t_flags & TF_SACK_PERMIT) && 801 tp->rcv_numsacks > 0) { 802 to.to_flags |= TOF_SACK; 803 to.to_nsacks = tp->rcv_numsacks; 804 to.to_sacks = (u_char *)tp->sackblks; 805 } 806 } 807 #ifdef TCP_SIGNATURE 808 /* TCP-MD5 (RFC2385). */ 809 if (tp->t_flags & TF_SIGNATURE) 810 to.to_flags |= TOF_SIGNATURE; 811 #endif /* TCP_SIGNATURE */ 812 813 /* Processing the options. */ 814 hdrlen += optlen = tcp_addoptions(&to, opt); 815 } 816 817 #ifdef INET6 818 if (isipv6) 819 ipoptlen = ip6_optlen(tp->t_inpcb); 820 else 821 #endif 822 if (tp->t_inpcb->inp_options) 823 ipoptlen = tp->t_inpcb->inp_options->m_len - 824 offsetof(struct ipoption, ipopt_list); 825 else 826 ipoptlen = 0; 827 #ifdef IPSEC 828 ipoptlen += ipsec_optlen; 829 #endif 830 831 /* 832 * Adjust data length if insertion of options will 833 * bump the packet length beyond the t_maxseg length. 834 * Clear the FIN bit because we cut off the tail of 835 * the segment. 836 */ 837 if (len + optlen + ipoptlen > tp->t_maxseg) { 838 flags &= ~TH_FIN; 839 840 if (tso) { 841 u_int if_hw_tsomax; 842 u_int if_hw_tsomaxsegcount; 843 u_int if_hw_tsomaxsegsize; 844 struct mbuf *mb; 845 u_int moff; 846 int max_len; 847 848 /* extract TSO information */ 849 if_hw_tsomax = tp->t_tsomax; 850 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 851 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 852 853 /* 854 * Limit a TSO burst to prevent it from 855 * overflowing or exceeding the maximum length 856 * allowed by the network interface: 857 */ 858 KASSERT(ipoptlen == 0, 859 ("%s: TSO can't do IP options", __func__)); 860 861 /* 862 * Check if we should limit by maximum payload 863 * length: 864 */ 865 if (if_hw_tsomax != 0) { 866 /* compute maximum TSO length */ 867 max_len = (if_hw_tsomax - hdrlen - 868 max_linkhdr); 869 if (max_len <= 0) { 870 len = 0; 871 } else if (len > max_len) { 872 sendalot = 1; 873 len = max_len; 874 } 875 } 876 877 /* 878 * Check if we should limit by maximum segment 879 * size and count: 880 */ 881 if (if_hw_tsomaxsegcount != 0 && 882 if_hw_tsomaxsegsize != 0) { 883 /* 884 * Subtract one segment for the LINK 885 * and TCP/IP headers mbuf that will 886 * be prepended to this mbuf chain 887 * after the code in this section 888 * limits the number of mbufs in the 889 * chain to if_hw_tsomaxsegcount. 890 */ 891 if_hw_tsomaxsegcount -= 1; 892 max_len = 0; 893 mb = sbsndmbuf(&so->so_snd, off, &moff); 894 895 while (mb != NULL && max_len < len) { 896 u_int mlen; 897 u_int frags; 898 899 /* 900 * Get length of mbuf fragment 901 * and how many hardware frags, 902 * rounded up, it would use: 903 */ 904 mlen = (mb->m_len - moff); 905 frags = howmany(mlen, 906 if_hw_tsomaxsegsize); 907 908 /* Handle special case: Zero Length Mbuf */ 909 if (frags == 0) 910 frags = 1; 911 912 /* 913 * Check if the fragment limit 914 * will be reached or exceeded: 915 */ 916 if (frags >= if_hw_tsomaxsegcount) { 917 max_len += min(mlen, 918 if_hw_tsomaxsegcount * 919 if_hw_tsomaxsegsize); 920 break; 921 } 922 max_len += mlen; 923 if_hw_tsomaxsegcount -= frags; 924 moff = 0; 925 mb = mb->m_next; 926 } 927 if (max_len <= 0) { 928 len = 0; 929 } else if (len > max_len) { 930 sendalot = 1; 931 len = max_len; 932 } 933 } 934 935 /* 936 * Prevent the last segment from being 937 * fractional unless the send sockbuf can be 938 * emptied: 939 */ 940 max_len = (tp->t_maxseg - optlen); 941 if ((off + len) < sbavail(&so->so_snd)) { 942 moff = len % max_len; 943 if (moff != 0) { 944 len -= moff; 945 sendalot = 1; 946 } 947 } 948 949 /* 950 * In case there are too many small fragments 951 * don't use TSO: 952 */ 953 if (len <= max_len) { 954 len = max_len; 955 sendalot = 1; 956 tso = 0; 957 } 958 959 /* 960 * Send the FIN in a separate segment 961 * after the bulk sending is done. 962 * We don't trust the TSO implementations 963 * to clear the FIN flag on all but the 964 * last segment. 965 */ 966 if (tp->t_flags & TF_NEEDFIN) 967 sendalot = 1; 968 969 } else { 970 len = tp->t_maxseg - optlen - ipoptlen; 971 sendalot = 1; 972 } 973 } else 974 tso = 0; 975 976 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 977 ("%s: len > IP_MAXPACKET", __func__)); 978 979 /*#ifdef DIAGNOSTIC*/ 980 #ifdef INET6 981 if (max_linkhdr + hdrlen > MCLBYTES) 982 #else 983 if (max_linkhdr + hdrlen > MHLEN) 984 #endif 985 panic("tcphdr too big"); 986 /*#endif*/ 987 988 /* 989 * This KASSERT is here to catch edge cases at a well defined place. 990 * Before, those had triggered (random) panic conditions further down. 991 */ 992 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 993 994 /* 995 * Grab a header mbuf, attaching a copy of data to 996 * be transmitted, and initialize the header from 997 * the template for sends on this connection. 998 */ 999 if (len) { 1000 struct mbuf *mb; 1001 u_int moff; 1002 1003 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 1004 TCPSTAT_INC(tcps_sndprobe); 1005 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 1006 tp->t_sndrexmitpack++; 1007 TCPSTAT_INC(tcps_sndrexmitpack); 1008 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 1009 } else { 1010 TCPSTAT_INC(tcps_sndpack); 1011 TCPSTAT_ADD(tcps_sndbyte, len); 1012 } 1013 #ifdef INET6 1014 if (MHLEN < hdrlen + max_linkhdr) 1015 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1016 else 1017 #endif 1018 m = m_gethdr(M_NOWAIT, MT_DATA); 1019 1020 if (m == NULL) { 1021 SOCKBUF_UNLOCK(&so->so_snd); 1022 error = ENOBUFS; 1023 sack_rxmit = 0; 1024 goto out; 1025 } 1026 1027 m->m_data += max_linkhdr; 1028 m->m_len = hdrlen; 1029 1030 /* 1031 * Start the m_copy functions from the closest mbuf 1032 * to the offset in the socket buffer chain. 1033 */ 1034 mb = sbsndptr(&so->so_snd, off, len, &moff); 1035 1036 if (len <= MHLEN - hdrlen - max_linkhdr) { 1037 m_copydata(mb, moff, (int)len, 1038 mtod(m, caddr_t) + hdrlen); 1039 m->m_len += len; 1040 } else { 1041 m->m_next = m_copy(mb, moff, (int)len); 1042 if (m->m_next == NULL) { 1043 SOCKBUF_UNLOCK(&so->so_snd); 1044 (void) m_free(m); 1045 error = ENOBUFS; 1046 sack_rxmit = 0; 1047 goto out; 1048 } 1049 } 1050 1051 /* 1052 * If we're sending everything we've got, set PUSH. 1053 * (This will keep happy those implementations which only 1054 * give data to the user when a buffer fills or 1055 * a PUSH comes in.) 1056 */ 1057 if ((off + len == sbused(&so->so_snd)) && !(flags & TH_SYN)) 1058 flags |= TH_PUSH; 1059 SOCKBUF_UNLOCK(&so->so_snd); 1060 } else { 1061 SOCKBUF_UNLOCK(&so->so_snd); 1062 if (tp->t_flags & TF_ACKNOW) 1063 TCPSTAT_INC(tcps_sndacks); 1064 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1065 TCPSTAT_INC(tcps_sndctrl); 1066 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1067 TCPSTAT_INC(tcps_sndurg); 1068 else 1069 TCPSTAT_INC(tcps_sndwinup); 1070 1071 m = m_gethdr(M_NOWAIT, MT_DATA); 1072 if (m == NULL) { 1073 error = ENOBUFS; 1074 sack_rxmit = 0; 1075 goto out; 1076 } 1077 #ifdef INET6 1078 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1079 MHLEN >= hdrlen) { 1080 M_ALIGN(m, hdrlen); 1081 } else 1082 #endif 1083 m->m_data += max_linkhdr; 1084 m->m_len = hdrlen; 1085 } 1086 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1087 m->m_pkthdr.rcvif = (struct ifnet *)0; 1088 #ifdef MAC 1089 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1090 #endif 1091 #ifdef INET6 1092 if (isipv6) { 1093 ip6 = mtod(m, struct ip6_hdr *); 1094 th = (struct tcphdr *)(ip6 + 1); 1095 tcpip_fillheaders(tp->t_inpcb, ip6, th); 1096 } else 1097 #endif /* INET6 */ 1098 { 1099 ip = mtod(m, struct ip *); 1100 ipov = (struct ipovly *)ip; 1101 th = (struct tcphdr *)(ip + 1); 1102 tcpip_fillheaders(tp->t_inpcb, ip, th); 1103 } 1104 1105 /* 1106 * Fill in fields, remembering maximum advertised 1107 * window for use in delaying messages about window sizes. 1108 * If resending a FIN, be sure not to use a new sequence number. 1109 */ 1110 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1111 tp->snd_nxt == tp->snd_max) 1112 tp->snd_nxt--; 1113 /* 1114 * If we are starting a connection, send ECN setup 1115 * SYN packet. If we are on a retransmit, we may 1116 * resend those bits a number of times as per 1117 * RFC 3168. 1118 */ 1119 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 1120 if (tp->t_rxtshift >= 1) { 1121 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 1122 flags |= TH_ECE|TH_CWR; 1123 } else 1124 flags |= TH_ECE|TH_CWR; 1125 } 1126 1127 if (tp->t_state == TCPS_ESTABLISHED && 1128 (tp->t_flags & TF_ECN_PERMIT)) { 1129 /* 1130 * If the peer has ECN, mark data packets with 1131 * ECN capable transmission (ECT). 1132 * Ignore pure ack packets, retransmissions and window probes. 1133 */ 1134 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1135 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 1136 #ifdef INET6 1137 if (isipv6) 1138 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1139 else 1140 #endif 1141 ip->ip_tos |= IPTOS_ECN_ECT0; 1142 TCPSTAT_INC(tcps_ecn_ect0); 1143 } 1144 1145 /* 1146 * Reply with proper ECN notifications. 1147 */ 1148 if (tp->t_flags & TF_ECN_SND_CWR) { 1149 flags |= TH_CWR; 1150 tp->t_flags &= ~TF_ECN_SND_CWR; 1151 } 1152 if (tp->t_flags & TF_ECN_SND_ECE) 1153 flags |= TH_ECE; 1154 } 1155 1156 /* 1157 * If we are doing retransmissions, then snd_nxt will 1158 * not reflect the first unsent octet. For ACK only 1159 * packets, we do not want the sequence number of the 1160 * retransmitted packet, we want the sequence number 1161 * of the next unsent octet. So, if there is no data 1162 * (and no SYN or FIN), use snd_max instead of snd_nxt 1163 * when filling in ti_seq. But if we are in persist 1164 * state, snd_max might reflect one byte beyond the 1165 * right edge of the window, so use snd_nxt in that 1166 * case, since we know we aren't doing a retransmission. 1167 * (retransmit and persist are mutually exclusive...) 1168 */ 1169 if (sack_rxmit == 0) { 1170 if (len || (flags & (TH_SYN|TH_FIN)) || 1171 tcp_timer_active(tp, TT_PERSIST)) 1172 th->th_seq = htonl(tp->snd_nxt); 1173 else 1174 th->th_seq = htonl(tp->snd_max); 1175 } else { 1176 th->th_seq = htonl(p->rxmit); 1177 p->rxmit += len; 1178 tp->sackhint.sack_bytes_rexmit += len; 1179 } 1180 th->th_ack = htonl(tp->rcv_nxt); 1181 if (optlen) { 1182 bcopy(opt, th + 1, optlen); 1183 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1184 } 1185 th->th_flags = flags; 1186 /* 1187 * Calculate receive window. Don't shrink window, 1188 * but avoid silly window syndrome. 1189 */ 1190 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 1191 recwin < (long)tp->t_maxseg) 1192 recwin = 0; 1193 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1194 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 1195 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 1196 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 1197 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 1198 1199 /* 1200 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1201 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1202 * case is handled in syncache. 1203 */ 1204 if (flags & TH_SYN) 1205 th->th_win = htons((u_short) 1206 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1207 else 1208 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1209 1210 /* 1211 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1212 * a 0 window. This may cause the remote transmitter to stall. This 1213 * flag tells soreceive() to disable delayed acknowledgements when 1214 * draining the buffer. This can occur if the receiver is attempting 1215 * to read more data than can be buffered prior to transmitting on 1216 * the connection. 1217 */ 1218 if (th->th_win == 0) { 1219 tp->t_sndzerowin++; 1220 tp->t_flags |= TF_RXWIN0SENT; 1221 } else 1222 tp->t_flags &= ~TF_RXWIN0SENT; 1223 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1224 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1225 th->th_flags |= TH_URG; 1226 } else 1227 /* 1228 * If no urgent pointer to send, then we pull 1229 * the urgent pointer to the left edge of the send window 1230 * so that it doesn't drift into the send window on sequence 1231 * number wraparound. 1232 */ 1233 tp->snd_up = tp->snd_una; /* drag it along */ 1234 1235 #ifdef TCP_SIGNATURE 1236 if (tp->t_flags & TF_SIGNATURE) { 1237 int sigoff = to.to_signature - opt; 1238 tcp_signature_compute(m, 0, len, optlen, 1239 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1240 } 1241 #endif 1242 1243 /* 1244 * Put TCP length in extended header, and then 1245 * checksum extended header and data. 1246 */ 1247 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1248 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1249 #ifdef INET6 1250 if (isipv6) { 1251 /* 1252 * ip6_plen is not need to be filled now, and will be filled 1253 * in ip6_output. 1254 */ 1255 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1256 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1257 optlen + len, IPPROTO_TCP, 0); 1258 } 1259 #endif 1260 #if defined(INET6) && defined(INET) 1261 else 1262 #endif 1263 #ifdef INET 1264 { 1265 m->m_pkthdr.csum_flags = CSUM_TCP; 1266 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1267 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1268 1269 /* IP version must be set here for ipv4/ipv6 checking later */ 1270 KASSERT(ip->ip_v == IPVERSION, 1271 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1272 } 1273 #endif 1274 1275 /* 1276 * Enable TSO and specify the size of the segments. 1277 * The TCP pseudo header checksum is always provided. 1278 */ 1279 if (tso) { 1280 KASSERT(len > tp->t_maxseg - optlen, 1281 ("%s: len <= tso_segsz", __func__)); 1282 m->m_pkthdr.csum_flags |= CSUM_TSO; 1283 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 1284 } 1285 1286 #ifdef IPSEC 1287 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1288 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u", 1289 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1290 #else 1291 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1292 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u", 1293 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1294 #endif 1295 1296 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1297 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1298 1299 #ifdef TCPDEBUG 1300 /* 1301 * Trace. 1302 */ 1303 if (so->so_options & SO_DEBUG) { 1304 u_short save = 0; 1305 #ifdef INET6 1306 if (!isipv6) 1307 #endif 1308 { 1309 save = ipov->ih_len; 1310 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1311 } 1312 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1313 #ifdef INET6 1314 if (!isipv6) 1315 #endif 1316 ipov->ih_len = save; 1317 } 1318 #endif /* TCPDEBUG */ 1319 TCP_PROBE3(debug__input, tp, th, mtod(m, const char *)); 1320 1321 /* 1322 * Fill in IP length and desired time to live and 1323 * send to IP level. There should be a better way 1324 * to handle ttl and tos; we could keep them in 1325 * the template, but need a way to checksum without them. 1326 */ 1327 /* 1328 * m->m_pkthdr.len should have been set before checksum calculation, 1329 * because in6_cksum() need it. 1330 */ 1331 #ifdef INET6 1332 if (isipv6) { 1333 struct route_in6 ro; 1334 1335 bzero(&ro, sizeof(ro)); 1336 /* 1337 * we separately set hoplimit for every segment, since the 1338 * user might want to change the value via setsockopt. 1339 * Also, desired default hop limit might be changed via 1340 * Neighbor Discovery. 1341 */ 1342 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1343 1344 /* 1345 * Set the packet size here for the benefit of DTrace probes. 1346 * ip6_output() will set it properly; it's supposed to include 1347 * the option header lengths as well. 1348 */ 1349 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1350 1351 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 1352 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1353 else 1354 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1355 1356 if (tp->t_state == TCPS_SYN_SENT) 1357 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1358 1359 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1360 1361 #ifdef TCPPCAP 1362 /* Save packet, if requested. */ 1363 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1364 #endif 1365 1366 /* TODO: IPv6 IP6TOS_ECT bit on */ 1367 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro, 1368 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1369 NULL, NULL, tp->t_inpcb); 1370 1371 if (error == EMSGSIZE && ro.ro_rt != NULL) 1372 mtu = ro.ro_rt->rt_mtu; 1373 RO_RTFREE(&ro); 1374 } 1375 #endif /* INET6 */ 1376 #if defined(INET) && defined(INET6) 1377 else 1378 #endif 1379 #ifdef INET 1380 { 1381 struct route ro; 1382 1383 bzero(&ro, sizeof(ro)); 1384 ip->ip_len = htons(m->m_pkthdr.len); 1385 #ifdef INET6 1386 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1387 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1388 #endif /* INET6 */ 1389 /* 1390 * If we do path MTU discovery, then we set DF on every packet. 1391 * This might not be the best thing to do according to RFC3390 1392 * Section 2. However the tcp hostcache migitates the problem 1393 * so it affects only the first tcp connection with a host. 1394 * 1395 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1396 */ 1397 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 1398 ip->ip_off |= htons(IP_DF); 1399 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1400 } else { 1401 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1402 } 1403 1404 if (tp->t_state == TCPS_SYN_SENT) 1405 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1406 1407 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1408 1409 #ifdef TCPPCAP 1410 /* Save packet, if requested. */ 1411 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1412 #endif 1413 1414 error = ip_output(m, tp->t_inpcb->inp_options, &ro, 1415 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1416 tp->t_inpcb); 1417 1418 if (error == EMSGSIZE && ro.ro_rt != NULL) 1419 mtu = ro.ro_rt->rt_mtu; 1420 RO_RTFREE(&ro); 1421 } 1422 #endif /* INET */ 1423 1424 out: 1425 /* 1426 * In transmit state, time the transmission and arrange for 1427 * the retransmit. In persist state, just set snd_max. 1428 */ 1429 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1430 !tcp_timer_active(tp, TT_PERSIST)) { 1431 tcp_seq startseq = tp->snd_nxt; 1432 1433 /* 1434 * Advance snd_nxt over sequence space of this segment. 1435 */ 1436 if (flags & (TH_SYN|TH_FIN)) { 1437 if (flags & TH_SYN) 1438 tp->snd_nxt++; 1439 if (flags & TH_FIN) { 1440 tp->snd_nxt++; 1441 tp->t_flags |= TF_SENTFIN; 1442 } 1443 } 1444 if (sack_rxmit) 1445 goto timer; 1446 tp->snd_nxt += len; 1447 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1448 tp->snd_max = tp->snd_nxt; 1449 /* 1450 * Time this transmission if not a retransmission and 1451 * not currently timing anything. 1452 */ 1453 if (tp->t_rtttime == 0) { 1454 tp->t_rtttime = ticks; 1455 tp->t_rtseq = startseq; 1456 TCPSTAT_INC(tcps_segstimed); 1457 } 1458 } 1459 1460 /* 1461 * Set retransmit timer if not currently set, 1462 * and not doing a pure ack or a keep-alive probe. 1463 * Initial value for retransmit timer is smoothed 1464 * round-trip time + 2 * round-trip time variance. 1465 * Initialize shift counter which is used for backoff 1466 * of retransmit time. 1467 */ 1468 timer: 1469 if (!tcp_timer_active(tp, TT_REXMT) && 1470 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1471 (tp->snd_nxt != tp->snd_una))) { 1472 if (tcp_timer_active(tp, TT_PERSIST)) { 1473 tcp_timer_activate(tp, TT_PERSIST, 0); 1474 tp->t_rxtshift = 0; 1475 } 1476 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1477 } else if (len == 0 && sbavail(&so->so_snd) && 1478 !tcp_timer_active(tp, TT_REXMT) && 1479 !tcp_timer_active(tp, TT_PERSIST)) { 1480 /* 1481 * Avoid a situation where we do not set persist timer 1482 * after a zero window condition. For example: 1483 * 1) A -> B: packet with enough data to fill the window 1484 * 2) B -> A: ACK for #1 + new data (0 window 1485 * advertisement) 1486 * 3) A -> B: ACK for #2, 0 len packet 1487 * 1488 * In this case, A will not activate the persist timer, 1489 * because it chose to send a packet. Unless tcp_output 1490 * is called for some other reason (delayed ack timer, 1491 * another input packet from B, socket syscall), A will 1492 * not send zero window probes. 1493 * 1494 * So, if you send a 0-length packet, but there is data 1495 * in the socket buffer, and neither the rexmt or 1496 * persist timer is already set, then activate the 1497 * persist timer. 1498 */ 1499 tp->t_rxtshift = 0; 1500 tcp_setpersist(tp); 1501 } 1502 } else { 1503 /* 1504 * Persist case, update snd_max but since we are in 1505 * persist mode (no window) we do not update snd_nxt. 1506 */ 1507 int xlen = len; 1508 if (flags & TH_SYN) 1509 ++xlen; 1510 if (flags & TH_FIN) { 1511 ++xlen; 1512 tp->t_flags |= TF_SENTFIN; 1513 } 1514 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1515 tp->snd_max = tp->snd_nxt + len; 1516 } 1517 1518 if (error) { 1519 1520 /* 1521 * We know that the packet was lost, so back out the 1522 * sequence number advance, if any. 1523 * 1524 * If the error is EPERM the packet got blocked by the 1525 * local firewall. Normally we should terminate the 1526 * connection but the blocking may have been spurious 1527 * due to a firewall reconfiguration cycle. So we treat 1528 * it like a packet loss and let the retransmit timer and 1529 * timeouts do their work over time. 1530 * XXX: It is a POLA question whether calling tcp_drop right 1531 * away would be the really correct behavior instead. 1532 */ 1533 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1534 !tcp_timer_active(tp, TT_PERSIST)) && 1535 ((flags & TH_SYN) == 0) && 1536 (error != EPERM)) { 1537 if (sack_rxmit) { 1538 p->rxmit -= len; 1539 tp->sackhint.sack_bytes_rexmit -= len; 1540 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1541 ("sackhint bytes rtx >= 0")); 1542 } else 1543 tp->snd_nxt -= len; 1544 } 1545 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1546 switch (error) { 1547 case EPERM: 1548 tp->t_softerror = error; 1549 return (error); 1550 case ENOBUFS: 1551 if (!tcp_timer_active(tp, TT_REXMT) && 1552 !tcp_timer_active(tp, TT_PERSIST)) 1553 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1554 tp->snd_cwnd = tp->t_maxseg; 1555 return (0); 1556 case EMSGSIZE: 1557 /* 1558 * For some reason the interface we used initially 1559 * to send segments changed to another or lowered 1560 * its MTU. 1561 * If TSO was active we either got an interface 1562 * without TSO capabilits or TSO was turned off. 1563 * If we obtained mtu from ip_output() then update 1564 * it and try again. 1565 */ 1566 if (tso) 1567 tp->t_flags &= ~TF_TSO; 1568 if (mtu != 0) { 1569 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1570 goto again; 1571 } 1572 return (error); 1573 case EHOSTDOWN: 1574 case EHOSTUNREACH: 1575 case ENETDOWN: 1576 case ENETUNREACH: 1577 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1578 tp->t_softerror = error; 1579 return (0); 1580 } 1581 /* FALLTHROUGH */ 1582 default: 1583 return (error); 1584 } 1585 } 1586 TCPSTAT_INC(tcps_sndtotal); 1587 1588 /* 1589 * Data sent (as far as we can tell). 1590 * If this advertises a larger window than any other segment, 1591 * then remember the size of the advertised window. 1592 * Any pending ACK has now been sent. 1593 */ 1594 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1595 tp->rcv_adv = tp->rcv_nxt + recwin; 1596 tp->last_ack_sent = tp->rcv_nxt; 1597 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1598 if (tcp_timer_active(tp, TT_DELACK)) 1599 tcp_timer_activate(tp, TT_DELACK, 0); 1600 #if 0 1601 /* 1602 * This completely breaks TCP if newreno is turned on. What happens 1603 * is that if delayed-acks are turned on on the receiver, this code 1604 * on the transmitter effectively destroys the TCP window, forcing 1605 * it to four packets (1.5Kx4 = 6K window). 1606 */ 1607 if (sendalot && --maxburst) 1608 goto again; 1609 #endif 1610 if (sendalot) 1611 goto again; 1612 return (0); 1613 } 1614 1615 void 1616 tcp_setpersist(struct tcpcb *tp) 1617 { 1618 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1619 int tt; 1620 1621 tp->t_flags &= ~TF_PREVVALID; 1622 if (tcp_timer_active(tp, TT_REXMT)) 1623 panic("tcp_setpersist: retransmit pending"); 1624 /* 1625 * Start/restart persistance timer. 1626 */ 1627 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1628 TCPTV_PERSMIN, TCPTV_PERSMAX); 1629 tcp_timer_activate(tp, TT_PERSIST, tt); 1630 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1631 tp->t_rxtshift++; 1632 } 1633 1634 /* 1635 * Insert TCP options according to the supplied parameters to the place 1636 * optp in a consistent way. Can handle unaligned destinations. 1637 * 1638 * The order of the option processing is crucial for optimal packing and 1639 * alignment for the scarce option space. 1640 * 1641 * The optimal order for a SYN/SYN-ACK segment is: 1642 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1643 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1644 * 1645 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1646 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1647 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1648 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1649 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1650 */ 1651 int 1652 tcp_addoptions(struct tcpopt *to, u_char *optp) 1653 { 1654 u_int mask, optlen = 0; 1655 1656 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1657 if ((to->to_flags & mask) != mask) 1658 continue; 1659 if (optlen == TCP_MAXOLEN) 1660 break; 1661 switch (to->to_flags & mask) { 1662 case TOF_MSS: 1663 while (optlen % 4) { 1664 optlen += TCPOLEN_NOP; 1665 *optp++ = TCPOPT_NOP; 1666 } 1667 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1668 continue; 1669 optlen += TCPOLEN_MAXSEG; 1670 *optp++ = TCPOPT_MAXSEG; 1671 *optp++ = TCPOLEN_MAXSEG; 1672 to->to_mss = htons(to->to_mss); 1673 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1674 optp += sizeof(to->to_mss); 1675 break; 1676 case TOF_SCALE: 1677 while (!optlen || optlen % 2 != 1) { 1678 optlen += TCPOLEN_NOP; 1679 *optp++ = TCPOPT_NOP; 1680 } 1681 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1682 continue; 1683 optlen += TCPOLEN_WINDOW; 1684 *optp++ = TCPOPT_WINDOW; 1685 *optp++ = TCPOLEN_WINDOW; 1686 *optp++ = to->to_wscale; 1687 break; 1688 case TOF_SACKPERM: 1689 while (optlen % 2) { 1690 optlen += TCPOLEN_NOP; 1691 *optp++ = TCPOPT_NOP; 1692 } 1693 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1694 continue; 1695 optlen += TCPOLEN_SACK_PERMITTED; 1696 *optp++ = TCPOPT_SACK_PERMITTED; 1697 *optp++ = TCPOLEN_SACK_PERMITTED; 1698 break; 1699 case TOF_TS: 1700 while (!optlen || optlen % 4 != 2) { 1701 optlen += TCPOLEN_NOP; 1702 *optp++ = TCPOPT_NOP; 1703 } 1704 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1705 continue; 1706 optlen += TCPOLEN_TIMESTAMP; 1707 *optp++ = TCPOPT_TIMESTAMP; 1708 *optp++ = TCPOLEN_TIMESTAMP; 1709 to->to_tsval = htonl(to->to_tsval); 1710 to->to_tsecr = htonl(to->to_tsecr); 1711 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1712 optp += sizeof(to->to_tsval); 1713 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1714 optp += sizeof(to->to_tsecr); 1715 break; 1716 case TOF_SIGNATURE: 1717 { 1718 int siglen = TCPOLEN_SIGNATURE - 2; 1719 1720 while (!optlen || optlen % 4 != 2) { 1721 optlen += TCPOLEN_NOP; 1722 *optp++ = TCPOPT_NOP; 1723 } 1724 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1725 continue; 1726 optlen += TCPOLEN_SIGNATURE; 1727 *optp++ = TCPOPT_SIGNATURE; 1728 *optp++ = TCPOLEN_SIGNATURE; 1729 to->to_signature = optp; 1730 while (siglen--) 1731 *optp++ = 0; 1732 break; 1733 } 1734 case TOF_SACK: 1735 { 1736 int sackblks = 0; 1737 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1738 tcp_seq sack_seq; 1739 1740 while (!optlen || optlen % 4 != 2) { 1741 optlen += TCPOLEN_NOP; 1742 *optp++ = TCPOPT_NOP; 1743 } 1744 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1745 continue; 1746 optlen += TCPOLEN_SACKHDR; 1747 *optp++ = TCPOPT_SACK; 1748 sackblks = min(to->to_nsacks, 1749 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1750 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1751 while (sackblks--) { 1752 sack_seq = htonl(sack->start); 1753 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1754 optp += sizeof(sack_seq); 1755 sack_seq = htonl(sack->end); 1756 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1757 optp += sizeof(sack_seq); 1758 optlen += TCPOLEN_SACK; 1759 sack++; 1760 } 1761 TCPSTAT_INC(tcps_sack_send_blocks); 1762 break; 1763 } 1764 #ifdef TCP_RFC7413 1765 case TOF_FASTOPEN: 1766 { 1767 int total_len; 1768 1769 /* XXX is there any point to aligning this option? */ 1770 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; 1771 if (TCP_MAXOLEN - optlen < total_len) 1772 continue; 1773 *optp++ = TCPOPT_FAST_OPEN; 1774 *optp++ = total_len; 1775 if (to->to_tfo_len > 0) { 1776 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); 1777 optp += to->to_tfo_len; 1778 } 1779 optlen += total_len; 1780 break; 1781 } 1782 #endif 1783 default: 1784 panic("%s: unknown TCP option type", __func__); 1785 break; 1786 } 1787 } 1788 1789 /* Terminate and pad TCP options to a 4 byte boundary. */ 1790 if (optlen % 4) { 1791 optlen += TCPOLEN_EOL; 1792 *optp++ = TCPOPT_EOL; 1793 } 1794 /* 1795 * According to RFC 793 (STD0007): 1796 * "The content of the header beyond the End-of-Option option 1797 * must be header padding (i.e., zero)." 1798 * and later: "The padding is composed of zeros." 1799 */ 1800 while (optlen % 4) { 1801 optlen += TCPOLEN_PAD; 1802 *optp++ = TCPOPT_PAD; 1803 } 1804 1805 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1806 return (optlen); 1807 } 1808