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 #define TCPOUTFLAGS 72 #include <netinet/tcp_fsm.h> 73 #include <netinet/tcp_seq.h> 74 #include <netinet/tcp_timer.h> 75 #include <netinet/tcp_var.h> 76 #include <netinet/tcpip.h> 77 #ifdef TCPDEBUG 78 #include <netinet/tcp_debug.h> 79 #endif 80 #ifdef TCP_OFFLOAD 81 #include <netinet/tcp_offload.h> 82 #endif 83 84 #ifdef IPSEC 85 #include <netipsec/ipsec.h> 86 #endif /*IPSEC*/ 87 88 #include <machine/in_cksum.h> 89 90 #include <security/mac/mac_framework.h> 91 92 VNET_DEFINE(int, path_mtu_discovery) = 1; 93 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW, 94 &VNET_NAME(path_mtu_discovery), 1, 95 "Enable Path MTU Discovery"); 96 97 VNET_DEFINE(int, tcp_do_tso) = 1; 98 #define V_tcp_do_tso VNET(tcp_do_tso) 99 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW, 100 &VNET_NAME(tcp_do_tso), 0, 101 "Enable TCP Segmentation Offload"); 102 103 VNET_DEFINE(int, tcp_sendspace) = 1024*32; 104 #define V_tcp_sendspace VNET(tcp_sendspace) 105 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW, 106 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 107 108 VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 109 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) 110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, 111 &VNET_NAME(tcp_do_autosndbuf), 0, 112 "Enable automatic send buffer sizing"); 113 114 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 115 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) 116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW, 117 &VNET_NAME(tcp_autosndbuf_inc), 0, 118 "Incrementor step size of automatic send buffer"); 119 120 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; 121 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) 122 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW, 123 &VNET_NAME(tcp_autosndbuf_max), 0, 124 "Max size of automatic send buffer"); 125 126 static void inline hhook_run_tcp_est_out(struct tcpcb *tp, 127 struct tcphdr *th, struct tcpopt *to, 128 long len, int tso); 129 static void inline cc_after_idle(struct tcpcb *tp); 130 131 /* 132 * Wrapper for the TCP established output helper hook. 133 */ 134 static void inline 135 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 136 struct tcpopt *to, long len, int tso) 137 { 138 struct tcp_hhook_data hhook_data; 139 140 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 141 hhook_data.tp = tp; 142 hhook_data.th = th; 143 hhook_data.to = to; 144 hhook_data.len = len; 145 hhook_data.tso = tso; 146 147 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 148 tp->osd); 149 } 150 } 151 152 /* 153 * CC wrapper hook functions 154 */ 155 static void inline 156 cc_after_idle(struct tcpcb *tp) 157 { 158 INP_WLOCK_ASSERT(tp->t_inpcb); 159 160 if (CC_ALGO(tp)->after_idle != NULL) 161 CC_ALGO(tp)->after_idle(tp->ccv); 162 } 163 164 /* 165 * Tcp output routine: figure out what should be sent and send it. 166 */ 167 int 168 tcp_output(struct tcpcb *tp) 169 { 170 struct socket *so = tp->t_inpcb->inp_socket; 171 long len, recwin, sendwin; 172 int off, flags, error = 0; /* Keep compiler happy */ 173 struct mbuf *m; 174 struct ip *ip = NULL; 175 struct ipovly *ipov = NULL; 176 struct tcphdr *th; 177 u_char opt[TCP_MAXOLEN]; 178 unsigned ipoptlen, optlen, hdrlen; 179 #ifdef IPSEC 180 unsigned ipsec_optlen = 0; 181 #endif 182 int idle, sendalot; 183 int sack_rxmit, sack_bytes_rxmt; 184 struct sackhole *p; 185 int tso, mtu; 186 struct tcpopt to; 187 #if 0 188 int maxburst = TCP_MAXBURST; 189 #endif 190 #ifdef INET6 191 struct ip6_hdr *ip6 = NULL; 192 int isipv6; 193 194 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 195 #endif 196 197 INP_WLOCK_ASSERT(tp->t_inpcb); 198 199 #ifdef TCP_OFFLOAD 200 if (tp->t_flags & TF_TOE) 201 return (tcp_offload_output(tp)); 202 #endif 203 204 /* 205 * Determine length of data that should be transmitted, 206 * and flags that will be used. 207 * If there is some data or critical controls (SYN, RST) 208 * to send, then transmit; otherwise, investigate further. 209 */ 210 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 211 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) 212 cc_after_idle(tp); 213 tp->t_flags &= ~TF_LASTIDLE; 214 if (idle) { 215 if (tp->t_flags & TF_MORETOCOME) { 216 tp->t_flags |= TF_LASTIDLE; 217 idle = 0; 218 } 219 } 220 again: 221 /* 222 * If we've recently taken a timeout, snd_max will be greater than 223 * snd_nxt. There may be SACK information that allows us to avoid 224 * resending already delivered data. Adjust snd_nxt accordingly. 225 */ 226 if ((tp->t_flags & TF_SACK_PERMIT) && 227 SEQ_LT(tp->snd_nxt, tp->snd_max)) 228 tcp_sack_adjust(tp); 229 sendalot = 0; 230 tso = 0; 231 mtu = 0; 232 off = tp->snd_nxt - tp->snd_una; 233 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 234 235 flags = tcp_outflags[tp->t_state]; 236 /* 237 * Send any SACK-generated retransmissions. If we're explicitly trying 238 * to send out new data (when sendalot is 1), bypass this function. 239 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 240 * we're replacing a (future) new transmission with a retransmission 241 * now, and we previously incremented snd_cwnd in tcp_input(). 242 */ 243 /* 244 * Still in sack recovery , reset rxmit flag to zero. 245 */ 246 sack_rxmit = 0; 247 sack_bytes_rxmt = 0; 248 len = 0; 249 p = NULL; 250 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 251 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 252 long cwin; 253 254 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 255 if (cwin < 0) 256 cwin = 0; 257 /* Do not retransmit SACK segments beyond snd_recover */ 258 if (SEQ_GT(p->end, tp->snd_recover)) { 259 /* 260 * (At least) part of sack hole extends beyond 261 * snd_recover. Check to see if we can rexmit data 262 * for this hole. 263 */ 264 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 265 /* 266 * Can't rexmit any more data for this hole. 267 * That data will be rexmitted in the next 268 * sack recovery episode, when snd_recover 269 * moves past p->rxmit. 270 */ 271 p = NULL; 272 goto after_sack_rexmit; 273 } else 274 /* Can rexmit part of the current hole */ 275 len = ((long)ulmin(cwin, 276 tp->snd_recover - p->rxmit)); 277 } else 278 len = ((long)ulmin(cwin, p->end - p->rxmit)); 279 off = p->rxmit - tp->snd_una; 280 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 281 __func__, off)); 282 if (len > 0) { 283 sack_rxmit = 1; 284 sendalot = 1; 285 TCPSTAT_INC(tcps_sack_rexmits); 286 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 287 min(len, tp->t_maxseg)); 288 } 289 } 290 after_sack_rexmit: 291 /* 292 * Get standard flags, and add SYN or FIN if requested by 'hidden' 293 * state flags. 294 */ 295 if (tp->t_flags & TF_NEEDFIN) 296 flags |= TH_FIN; 297 if (tp->t_flags & TF_NEEDSYN) 298 flags |= TH_SYN; 299 300 SOCKBUF_LOCK(&so->so_snd); 301 /* 302 * If in persist timeout with window of 0, send 1 byte. 303 * Otherwise, if window is small but nonzero 304 * and timer expired, we will send what we can 305 * and go to transmit state. 306 */ 307 if (tp->t_flags & TF_FORCEDATA) { 308 if (sendwin == 0) { 309 /* 310 * If we still have some data to send, then 311 * clear the FIN bit. Usually this would 312 * happen below when it realizes that we 313 * aren't sending all the data. However, 314 * if we have exactly 1 byte of unsent data, 315 * then it won't clear the FIN bit below, 316 * and if we are in persist state, we wind 317 * up sending the packet without recording 318 * that we sent the FIN bit. 319 * 320 * We can't just blindly clear the FIN bit, 321 * because if we don't have any more data 322 * to send then the probe will be the FIN 323 * itself. 324 */ 325 if (off < sbused(&so->so_snd)) 326 flags &= ~TH_FIN; 327 sendwin = 1; 328 } else { 329 tcp_timer_activate(tp, TT_PERSIST, 0); 330 tp->t_rxtshift = 0; 331 } 332 } 333 334 /* 335 * If snd_nxt == snd_max and we have transmitted a FIN, the 336 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 337 * a negative length. This can also occur when TCP opens up 338 * its congestion window while receiving additional duplicate 339 * acks after fast-retransmit because TCP will reset snd_nxt 340 * to snd_max after the fast-retransmit. 341 * 342 * In the normal retransmit-FIN-only case, however, snd_nxt will 343 * be set to snd_una, the offset will be 0, and the length may 344 * wind up 0. 345 * 346 * If sack_rxmit is true we are retransmitting from the scoreboard 347 * in which case len is already set. 348 */ 349 if (sack_rxmit == 0) { 350 if (sack_bytes_rxmt == 0) 351 len = ((long)ulmin(sbavail(&so->so_snd), sendwin) - 352 off); 353 else { 354 long cwin; 355 356 /* 357 * We are inside of a SACK recovery episode and are 358 * sending new data, having retransmitted all the 359 * data possible in the scoreboard. 360 */ 361 len = ((long)ulmin(sbavail(&so->so_snd), tp->snd_wnd) - 362 off); 363 /* 364 * Don't remove this (len > 0) check ! 365 * We explicitly check for len > 0 here (although it 366 * isn't really necessary), to work around a gcc 367 * optimization issue - to force gcc to compute 368 * len above. Without this check, the computation 369 * of len is bungled by the optimizer. 370 */ 371 if (len > 0) { 372 cwin = tp->snd_cwnd - 373 (tp->snd_nxt - tp->sack_newdata) - 374 sack_bytes_rxmt; 375 if (cwin < 0) 376 cwin = 0; 377 len = lmin(len, cwin); 378 } 379 } 380 } 381 382 /* 383 * Lop off SYN bit if it has already been sent. However, if this 384 * is SYN-SENT state and if segment contains data and if we don't 385 * know that foreign host supports TAO, suppress sending segment. 386 */ 387 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 388 if (tp->t_state != TCPS_SYN_RECEIVED) 389 flags &= ~TH_SYN; 390 off--, len++; 391 } 392 393 /* 394 * Be careful not to send data and/or FIN on SYN segments. 395 * This measure is needed to prevent interoperability problems 396 * with not fully conformant TCP implementations. 397 */ 398 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 399 len = 0; 400 flags &= ~TH_FIN; 401 } 402 403 if (len <= 0) { 404 /* 405 * If FIN has been sent but not acked, 406 * but we haven't been called to retransmit, 407 * len will be < 0. Otherwise, window shrank 408 * after we sent into it. If window shrank to 0, 409 * cancel pending retransmit, pull snd_nxt back 410 * to (closed) window, and set the persist timer 411 * if it isn't already going. If the window didn't 412 * close completely, just wait for an ACK. 413 * 414 * We also do a general check here to ensure that 415 * we will set the persist timer when we have data 416 * to send, but a 0-byte window. This makes sure 417 * the persist timer is set even if the packet 418 * hits one of the "goto send" lines below. 419 */ 420 len = 0; 421 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 422 (off < (int) sbavail(&so->so_snd))) { 423 tcp_timer_activate(tp, TT_REXMT, 0); 424 tp->t_rxtshift = 0; 425 tp->snd_nxt = tp->snd_una; 426 if (!tcp_timer_active(tp, TT_PERSIST)) 427 tcp_setpersist(tp); 428 } 429 } 430 431 /* len will be >= 0 after this point. */ 432 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 433 434 /* 435 * Automatic sizing of send socket buffer. Often the send buffer 436 * size is not optimally adjusted to the actual network conditions 437 * at hand (delay bandwidth product). Setting the buffer size too 438 * small limits throughput on links with high bandwidth and high 439 * delay (eg. trans-continental/oceanic links). Setting the 440 * buffer size too big consumes too much real kernel memory, 441 * especially with many connections on busy servers. 442 * 443 * The criteria to step up the send buffer one notch are: 444 * 1. receive window of remote host is larger than send buffer 445 * (with a fudge factor of 5/4th); 446 * 2. send buffer is filled to 7/8th with data (so we actually 447 * have data to make use of it); 448 * 3. send buffer fill has not hit maximal automatic size; 449 * 4. our send window (slow start and cogestion controlled) is 450 * larger than sent but unacknowledged data in send buffer. 451 * 452 * The remote host receive window scaling factor may limit the 453 * growing of the send buffer before it reaches its allowed 454 * maximum. 455 * 456 * It scales directly with slow start or congestion window 457 * and does at most one step per received ACK. This fast 458 * scaling has the drawback of growing the send buffer beyond 459 * what is strictly necessary to make full use of a given 460 * delay*bandwith product. However testing has shown this not 461 * to be much of an problem. At worst we are trading wasting 462 * of available bandwith (the non-use of it) for wasting some 463 * socket buffer memory. 464 * 465 * TODO: Shrink send buffer during idle periods together 466 * with congestion window. Requires another timer. Has to 467 * wait for upcoming tcp timer rewrite. 468 * 469 * XXXGL: should there be used sbused() or sbavail()? 470 */ 471 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 472 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 473 sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) && 474 sbused(&so->so_snd) < V_tcp_autosndbuf_max && 475 sendwin >= (sbused(&so->so_snd) - 476 (tp->snd_nxt - tp->snd_una))) { 477 if (!sbreserve_locked(&so->so_snd, 478 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 479 V_tcp_autosndbuf_max), so, curthread)) 480 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 481 } 482 } 483 484 /* 485 * Decide if we can use TCP Segmentation Offloading (if supported by 486 * hardware). 487 * 488 * TSO may only be used if we are in a pure bulk sending state. The 489 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 490 * IP options prevent using TSO. With TSO the TCP header is the same 491 * (except for the sequence number) for all generated packets. This 492 * makes it impossible to transmit any options which vary per generated 493 * segment or packet. 494 */ 495 #ifdef IPSEC 496 /* 497 * Pre-calculate here as we save another lookup into the darknesses 498 * of IPsec that way and can actually decide if TSO is ok. 499 */ 500 ipsec_optlen = ipsec_hdrsiz_tcp(tp); 501 #endif 502 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 503 ((tp->t_flags & TF_SIGNATURE) == 0) && 504 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 505 #ifdef IPSEC 506 ipsec_optlen == 0 && 507 #endif 508 tp->t_inpcb->inp_options == NULL && 509 tp->t_inpcb->in6p_options == NULL) 510 tso = 1; 511 512 if (sack_rxmit) { 513 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd))) 514 flags &= ~TH_FIN; 515 } else { 516 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 517 sbused(&so->so_snd))) 518 flags &= ~TH_FIN; 519 } 520 521 recwin = sbspace(&so->so_rcv); 522 523 /* 524 * Sender silly window avoidance. We transmit under the following 525 * conditions when len is non-zero: 526 * 527 * - We have a full segment (or more with TSO) 528 * - This is the last buffer in a write()/send() and we are 529 * either idle or running NODELAY 530 * - we've timed out (e.g. persist timer) 531 * - we have more then 1/2 the maximum send window's worth of 532 * data (receiver may be limited the window size) 533 * - we need to retransmit 534 */ 535 if (len) { 536 if (len >= tp->t_maxseg) 537 goto send; 538 /* 539 * NOTE! on localhost connections an 'ack' from the remote 540 * end may occur synchronously with the output and cause 541 * us to flush a buffer queued with moretocome. XXX 542 * 543 * note: the len + off check is almost certainly unnecessary. 544 */ 545 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 546 (idle || (tp->t_flags & TF_NODELAY)) && 547 len + off >= sbavail(&so->so_snd) && 548 (tp->t_flags & TF_NOPUSH) == 0) { 549 goto send; 550 } 551 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 552 goto send; 553 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 554 goto send; 555 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 556 goto send; 557 if (sack_rxmit) 558 goto send; 559 } 560 561 /* 562 * Sending of standalone window updates. 563 * 564 * Window updates are important when we close our window due to a 565 * full socket buffer and are opening it again after the application 566 * reads data from it. Once the window has opened again and the 567 * remote end starts to send again the ACK clock takes over and 568 * provides the most current window information. 569 * 570 * We must avoid the silly window syndrome whereas every read 571 * from the receive buffer, no matter how small, causes a window 572 * update to be sent. We also should avoid sending a flurry of 573 * window updates when the socket buffer had queued a lot of data 574 * and the application is doing small reads. 575 * 576 * Prevent a flurry of pointless window updates by only sending 577 * an update when we can increase the advertized window by more 578 * than 1/4th of the socket buffer capacity. When the buffer is 579 * getting full or is very small be more aggressive and send an 580 * update whenever we can increase by two mss sized segments. 581 * In all other situations the ACK's to new incoming data will 582 * carry further window increases. 583 * 584 * Don't send an independent window update if a delayed 585 * ACK is pending (it will get piggy-backed on it) or the 586 * remote side already has done a half-close and won't send 587 * more data. Skip this if the connection is in T/TCP 588 * half-open state. 589 */ 590 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 591 !(tp->t_flags & TF_DELACK) && 592 !TCPS_HAVERCVDFIN(tp->t_state)) { 593 /* 594 * "adv" is the amount we could increase the window, 595 * taking into account that we are limited by 596 * TCP_MAXWIN << tp->rcv_scale. 597 */ 598 long adv; 599 int oldwin; 600 601 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 602 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 603 oldwin = (tp->rcv_adv - tp->rcv_nxt); 604 adv -= oldwin; 605 } else 606 oldwin = 0; 607 608 /* 609 * If the new window size ends up being the same as the old 610 * size when it is scaled, then don't force a window update. 611 */ 612 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 613 goto dontupdate; 614 615 if (adv >= (long)(2 * tp->t_maxseg) && 616 (adv >= (long)(so->so_rcv.sb_hiwat / 4) || 617 recwin <= (long)(so->so_rcv.sb_hiwat / 8) || 618 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 619 goto send; 620 } 621 dontupdate: 622 623 /* 624 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 625 * is also a catch-all for the retransmit timer timeout case. 626 */ 627 if (tp->t_flags & TF_ACKNOW) 628 goto send; 629 if ((flags & TH_RST) || 630 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 631 goto send; 632 if (SEQ_GT(tp->snd_up, tp->snd_una)) 633 goto send; 634 /* 635 * If our state indicates that FIN should be sent 636 * and we have not yet done so, then we need to send. 637 */ 638 if (flags & TH_FIN && 639 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 640 goto send; 641 /* 642 * In SACK, it is possible for tcp_output to fail to send a segment 643 * after the retransmission timer has been turned off. Make sure 644 * that the retransmission timer is set. 645 */ 646 if ((tp->t_flags & TF_SACK_PERMIT) && 647 SEQ_GT(tp->snd_max, tp->snd_una) && 648 !tcp_timer_active(tp, TT_REXMT) && 649 !tcp_timer_active(tp, TT_PERSIST)) { 650 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 651 goto just_return; 652 } 653 /* 654 * TCP window updates are not reliable, rather a polling protocol 655 * using ``persist'' packets is used to insure receipt of window 656 * updates. The three ``states'' for the output side are: 657 * idle not doing retransmits or persists 658 * persisting to move a small or zero window 659 * (re)transmitting and thereby not persisting 660 * 661 * tcp_timer_active(tp, TT_PERSIST) 662 * is true when we are in persist state. 663 * (tp->t_flags & TF_FORCEDATA) 664 * is set when we are called to send a persist packet. 665 * tcp_timer_active(tp, TT_REXMT) 666 * is set when we are retransmitting 667 * The output side is idle when both timers are zero. 668 * 669 * If send window is too small, there is data to transmit, and no 670 * retransmit or persist is pending, then go to persist state. 671 * If nothing happens soon, send when timer expires: 672 * if window is nonzero, transmit what we can, 673 * otherwise force out a byte. 674 */ 675 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && 676 !tcp_timer_active(tp, TT_PERSIST)) { 677 tp->t_rxtshift = 0; 678 tcp_setpersist(tp); 679 } 680 681 /* 682 * No reason to send a segment, just return. 683 */ 684 just_return: 685 SOCKBUF_UNLOCK(&so->so_snd); 686 return (0); 687 688 send: 689 SOCKBUF_LOCK_ASSERT(&so->so_snd); 690 if (len > 0) { 691 if (len >= tp->t_maxseg) 692 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 693 else 694 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 695 } 696 /* 697 * Before ESTABLISHED, force sending of initial options 698 * unless TCP set not to do any options. 699 * NOTE: we assume that the IP/TCP header plus TCP options 700 * always fit in a single mbuf, leaving room for a maximum 701 * link header, i.e. 702 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 703 */ 704 optlen = 0; 705 #ifdef INET6 706 if (isipv6) 707 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 708 else 709 #endif 710 hdrlen = sizeof (struct tcpiphdr); 711 712 /* 713 * Compute options for segment. 714 * We only have to care about SYN and established connection 715 * segments. Options for SYN-ACK segments are handled in TCP 716 * syncache. 717 */ 718 if ((tp->t_flags & TF_NOOPT) == 0) { 719 to.to_flags = 0; 720 /* Maximum segment size. */ 721 if (flags & TH_SYN) { 722 tp->snd_nxt = tp->iss; 723 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 724 to.to_flags |= TOF_MSS; 725 } 726 /* Window scaling. */ 727 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 728 to.to_wscale = tp->request_r_scale; 729 to.to_flags |= TOF_SCALE; 730 } 731 /* Timestamps. */ 732 if ((tp->t_flags & TF_RCVD_TSTMP) || 733 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 734 to.to_tsval = tcp_ts_getticks() + tp->ts_offset; 735 to.to_tsecr = tp->ts_recent; 736 to.to_flags |= TOF_TS; 737 /* Set receive buffer autosizing timestamp. */ 738 if (tp->rfbuf_ts == 0 && 739 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 740 tp->rfbuf_ts = tcp_ts_getticks(); 741 } 742 /* Selective ACK's. */ 743 if (tp->t_flags & TF_SACK_PERMIT) { 744 if (flags & TH_SYN) 745 to.to_flags |= TOF_SACKPERM; 746 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 747 (tp->t_flags & TF_SACK_PERMIT) && 748 tp->rcv_numsacks > 0) { 749 to.to_flags |= TOF_SACK; 750 to.to_nsacks = tp->rcv_numsacks; 751 to.to_sacks = (u_char *)tp->sackblks; 752 } 753 } 754 #ifdef TCP_SIGNATURE 755 /* TCP-MD5 (RFC2385). */ 756 if (tp->t_flags & TF_SIGNATURE) 757 to.to_flags |= TOF_SIGNATURE; 758 #endif /* TCP_SIGNATURE */ 759 760 /* Processing the options. */ 761 hdrlen += optlen = tcp_addoptions(&to, opt); 762 } 763 764 #ifdef INET6 765 if (isipv6) 766 ipoptlen = ip6_optlen(tp->t_inpcb); 767 else 768 #endif 769 if (tp->t_inpcb->inp_options) 770 ipoptlen = tp->t_inpcb->inp_options->m_len - 771 offsetof(struct ipoption, ipopt_list); 772 else 773 ipoptlen = 0; 774 #ifdef IPSEC 775 ipoptlen += ipsec_optlen; 776 #endif 777 778 /* 779 * Adjust data length if insertion of options will 780 * bump the packet length beyond the t_maxopd length. 781 * Clear the FIN bit because we cut off the tail of 782 * the segment. 783 */ 784 if (len + optlen + ipoptlen > tp->t_maxopd) { 785 flags &= ~TH_FIN; 786 787 if (tso) { 788 u_int if_hw_tsomax; 789 u_int if_hw_tsomaxsegcount; 790 u_int if_hw_tsomaxsegsize; 791 struct mbuf *mb; 792 u_int moff; 793 int max_len; 794 795 /* extract TSO information */ 796 if_hw_tsomax = tp->t_tsomax; 797 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 798 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 799 800 /* 801 * Limit a TSO burst to prevent it from 802 * overflowing or exceeding the maximum length 803 * allowed by the network interface: 804 */ 805 KASSERT(ipoptlen == 0, 806 ("%s: TSO can't do IP options", __func__)); 807 808 /* 809 * Check if we should limit by maximum payload 810 * length: 811 */ 812 if (if_hw_tsomax != 0) { 813 /* compute maximum TSO length */ 814 max_len = (if_hw_tsomax - hdrlen); 815 if (max_len <= 0) { 816 len = 0; 817 } else if (len > max_len) { 818 sendalot = 1; 819 len = max_len; 820 } 821 } 822 823 /* 824 * Check if we should limit by maximum segment 825 * size and count: 826 */ 827 if (if_hw_tsomaxsegcount != 0 && 828 if_hw_tsomaxsegsize != 0) { 829 max_len = 0; 830 mb = sbsndmbuf(&so->so_snd, off, &moff); 831 832 while (mb != NULL && max_len < len) { 833 u_int mlen; 834 u_int frags; 835 836 /* 837 * Get length of mbuf fragment 838 * and how many hardware frags, 839 * rounded up, it would use: 840 */ 841 mlen = (mb->m_len - moff); 842 frags = howmany(mlen, 843 if_hw_tsomaxsegsize); 844 845 /* Handle special case: Zero Length Mbuf */ 846 if (frags == 0) 847 frags = 1; 848 849 /* 850 * Check if the fragment limit 851 * will be reached or exceeded: 852 */ 853 if (frags >= if_hw_tsomaxsegcount) { 854 max_len += min(mlen, 855 if_hw_tsomaxsegcount * 856 if_hw_tsomaxsegsize); 857 break; 858 } 859 max_len += mlen; 860 if_hw_tsomaxsegcount -= frags; 861 moff = 0; 862 mb = mb->m_next; 863 } 864 if (max_len <= 0) { 865 len = 0; 866 } else if (len > max_len) { 867 sendalot = 1; 868 len = max_len; 869 } 870 } 871 872 /* 873 * Prevent the last segment from being 874 * fractional unless the send sockbuf can be 875 * emptied: 876 */ 877 max_len = (tp->t_maxopd - optlen); 878 if ((off + len) < sbavail(&so->so_snd)) { 879 moff = len % max_len; 880 if (moff != 0) { 881 len -= moff; 882 sendalot = 1; 883 } 884 } 885 886 /* 887 * In case there are too many small fragments 888 * don't use TSO: 889 */ 890 if (len <= max_len) { 891 len = max_len; 892 sendalot = 1; 893 tso = 0; 894 } 895 896 /* 897 * Send the FIN in a separate segment 898 * after the bulk sending is done. 899 * We don't trust the TSO implementations 900 * to clear the FIN flag on all but the 901 * last segment. 902 */ 903 if (tp->t_flags & TF_NEEDFIN) 904 sendalot = 1; 905 906 } else { 907 len = tp->t_maxopd - optlen - ipoptlen; 908 sendalot = 1; 909 } 910 } else 911 tso = 0; 912 913 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 914 ("%s: len > IP_MAXPACKET", __func__)); 915 916 /*#ifdef DIAGNOSTIC*/ 917 #ifdef INET6 918 if (max_linkhdr + hdrlen > MCLBYTES) 919 #else 920 if (max_linkhdr + hdrlen > MHLEN) 921 #endif 922 panic("tcphdr too big"); 923 /*#endif*/ 924 925 /* 926 * This KASSERT is here to catch edge cases at a well defined place. 927 * Before, those had triggered (random) panic conditions further down. 928 */ 929 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 930 931 /* 932 * Grab a header mbuf, attaching a copy of data to 933 * be transmitted, and initialize the header from 934 * the template for sends on this connection. 935 */ 936 if (len) { 937 struct mbuf *mb; 938 u_int moff; 939 940 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 941 TCPSTAT_INC(tcps_sndprobe); 942 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 943 tp->t_sndrexmitpack++; 944 TCPSTAT_INC(tcps_sndrexmitpack); 945 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 946 } else { 947 TCPSTAT_INC(tcps_sndpack); 948 TCPSTAT_ADD(tcps_sndbyte, len); 949 } 950 #ifdef INET6 951 if (MHLEN < hdrlen + max_linkhdr) 952 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 953 else 954 #endif 955 m = m_gethdr(M_NOWAIT, MT_DATA); 956 957 if (m == NULL) { 958 SOCKBUF_UNLOCK(&so->so_snd); 959 error = ENOBUFS; 960 sack_rxmit = 0; 961 goto out; 962 } 963 964 m->m_data += max_linkhdr; 965 m->m_len = hdrlen; 966 967 /* 968 * Start the m_copy functions from the closest mbuf 969 * to the offset in the socket buffer chain. 970 */ 971 mb = sbsndptr(&so->so_snd, off, len, &moff); 972 973 if (len <= MHLEN - hdrlen - max_linkhdr) { 974 m_copydata(mb, moff, (int)len, 975 mtod(m, caddr_t) + hdrlen); 976 m->m_len += len; 977 } else { 978 m->m_next = m_copy(mb, moff, (int)len); 979 if (m->m_next == NULL) { 980 SOCKBUF_UNLOCK(&so->so_snd); 981 (void) m_free(m); 982 error = ENOBUFS; 983 sack_rxmit = 0; 984 goto out; 985 } 986 } 987 988 /* 989 * If we're sending everything we've got, set PUSH. 990 * (This will keep happy those implementations which only 991 * give data to the user when a buffer fills or 992 * a PUSH comes in.) 993 */ 994 if (off + len == sbused(&so->so_snd)) 995 flags |= TH_PUSH; 996 SOCKBUF_UNLOCK(&so->so_snd); 997 } else { 998 SOCKBUF_UNLOCK(&so->so_snd); 999 if (tp->t_flags & TF_ACKNOW) 1000 TCPSTAT_INC(tcps_sndacks); 1001 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1002 TCPSTAT_INC(tcps_sndctrl); 1003 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1004 TCPSTAT_INC(tcps_sndurg); 1005 else 1006 TCPSTAT_INC(tcps_sndwinup); 1007 1008 m = m_gethdr(M_NOWAIT, MT_DATA); 1009 if (m == NULL) { 1010 error = ENOBUFS; 1011 sack_rxmit = 0; 1012 goto out; 1013 } 1014 #ifdef INET6 1015 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1016 MHLEN >= hdrlen) { 1017 M_ALIGN(m, hdrlen); 1018 } else 1019 #endif 1020 m->m_data += max_linkhdr; 1021 m->m_len = hdrlen; 1022 } 1023 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1024 m->m_pkthdr.rcvif = (struct ifnet *)0; 1025 #ifdef MAC 1026 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1027 #endif 1028 #ifdef INET6 1029 if (isipv6) { 1030 ip6 = mtod(m, struct ip6_hdr *); 1031 th = (struct tcphdr *)(ip6 + 1); 1032 tcpip_fillheaders(tp->t_inpcb, ip6, th); 1033 } else 1034 #endif /* INET6 */ 1035 { 1036 ip = mtod(m, struct ip *); 1037 ipov = (struct ipovly *)ip; 1038 th = (struct tcphdr *)(ip + 1); 1039 tcpip_fillheaders(tp->t_inpcb, ip, th); 1040 } 1041 1042 /* 1043 * Fill in fields, remembering maximum advertised 1044 * window for use in delaying messages about window sizes. 1045 * If resending a FIN, be sure not to use a new sequence number. 1046 */ 1047 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1048 tp->snd_nxt == tp->snd_max) 1049 tp->snd_nxt--; 1050 /* 1051 * If we are starting a connection, send ECN setup 1052 * SYN packet. If we are on a retransmit, we may 1053 * resend those bits a number of times as per 1054 * RFC 3168. 1055 */ 1056 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 1057 if (tp->t_rxtshift >= 1) { 1058 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 1059 flags |= TH_ECE|TH_CWR; 1060 } else 1061 flags |= TH_ECE|TH_CWR; 1062 } 1063 1064 if (tp->t_state == TCPS_ESTABLISHED && 1065 (tp->t_flags & TF_ECN_PERMIT)) { 1066 /* 1067 * If the peer has ECN, mark data packets with 1068 * ECN capable transmission (ECT). 1069 * Ignore pure ack packets, retransmissions and window probes. 1070 */ 1071 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1072 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 1073 #ifdef INET6 1074 if (isipv6) 1075 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1076 else 1077 #endif 1078 ip->ip_tos |= IPTOS_ECN_ECT0; 1079 TCPSTAT_INC(tcps_ecn_ect0); 1080 } 1081 1082 /* 1083 * Reply with proper ECN notifications. 1084 */ 1085 if (tp->t_flags & TF_ECN_SND_CWR) { 1086 flags |= TH_CWR; 1087 tp->t_flags &= ~TF_ECN_SND_CWR; 1088 } 1089 if (tp->t_flags & TF_ECN_SND_ECE) 1090 flags |= TH_ECE; 1091 } 1092 1093 /* 1094 * If we are doing retransmissions, then snd_nxt will 1095 * not reflect the first unsent octet. For ACK only 1096 * packets, we do not want the sequence number of the 1097 * retransmitted packet, we want the sequence number 1098 * of the next unsent octet. So, if there is no data 1099 * (and no SYN or FIN), use snd_max instead of snd_nxt 1100 * when filling in ti_seq. But if we are in persist 1101 * state, snd_max might reflect one byte beyond the 1102 * right edge of the window, so use snd_nxt in that 1103 * case, since we know we aren't doing a retransmission. 1104 * (retransmit and persist are mutually exclusive...) 1105 */ 1106 if (sack_rxmit == 0) { 1107 if (len || (flags & (TH_SYN|TH_FIN)) || 1108 tcp_timer_active(tp, TT_PERSIST)) 1109 th->th_seq = htonl(tp->snd_nxt); 1110 else 1111 th->th_seq = htonl(tp->snd_max); 1112 } else { 1113 th->th_seq = htonl(p->rxmit); 1114 p->rxmit += len; 1115 tp->sackhint.sack_bytes_rexmit += len; 1116 } 1117 th->th_ack = htonl(tp->rcv_nxt); 1118 if (optlen) { 1119 bcopy(opt, th + 1, optlen); 1120 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1121 } 1122 th->th_flags = flags; 1123 /* 1124 * Calculate receive window. Don't shrink window, 1125 * but avoid silly window syndrome. 1126 */ 1127 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 1128 recwin < (long)tp->t_maxseg) 1129 recwin = 0; 1130 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1131 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 1132 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 1133 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 1134 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 1135 1136 /* 1137 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1138 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1139 * case is handled in syncache. 1140 */ 1141 if (flags & TH_SYN) 1142 th->th_win = htons((u_short) 1143 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1144 else 1145 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1146 1147 /* 1148 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1149 * a 0 window. This may cause the remote transmitter to stall. This 1150 * flag tells soreceive() to disable delayed acknowledgements when 1151 * draining the buffer. This can occur if the receiver is attempting 1152 * to read more data than can be buffered prior to transmitting on 1153 * the connection. 1154 */ 1155 if (th->th_win == 0) { 1156 tp->t_sndzerowin++; 1157 tp->t_flags |= TF_RXWIN0SENT; 1158 } else 1159 tp->t_flags &= ~TF_RXWIN0SENT; 1160 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1161 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1162 th->th_flags |= TH_URG; 1163 } else 1164 /* 1165 * If no urgent pointer to send, then we pull 1166 * the urgent pointer to the left edge of the send window 1167 * so that it doesn't drift into the send window on sequence 1168 * number wraparound. 1169 */ 1170 tp->snd_up = tp->snd_una; /* drag it along */ 1171 1172 #ifdef TCP_SIGNATURE 1173 if (tp->t_flags & TF_SIGNATURE) { 1174 int sigoff = to.to_signature - opt; 1175 tcp_signature_compute(m, 0, len, optlen, 1176 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1177 } 1178 #endif 1179 1180 /* 1181 * Put TCP length in extended header, and then 1182 * checksum extended header and data. 1183 */ 1184 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1185 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1186 #ifdef INET6 1187 if (isipv6) { 1188 /* 1189 * ip6_plen is not need to be filled now, and will be filled 1190 * in ip6_output. 1191 */ 1192 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1193 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1194 optlen + len, IPPROTO_TCP, 0); 1195 } 1196 #endif 1197 #if defined(INET6) && defined(INET) 1198 else 1199 #endif 1200 #ifdef INET 1201 { 1202 m->m_pkthdr.csum_flags = CSUM_TCP; 1203 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1204 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1205 1206 /* IP version must be set here for ipv4/ipv6 checking later */ 1207 KASSERT(ip->ip_v == IPVERSION, 1208 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1209 } 1210 #endif 1211 1212 /* 1213 * Enable TSO and specify the size of the segments. 1214 * The TCP pseudo header checksum is always provided. 1215 * XXX: Fixme: This is currently not the case for IPv6. 1216 */ 1217 if (tso) { 1218 KASSERT(len > tp->t_maxopd - optlen, 1219 ("%s: len <= tso_segsz", __func__)); 1220 m->m_pkthdr.csum_flags |= CSUM_TSO; 1221 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 1222 } 1223 1224 #ifdef IPSEC 1225 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1226 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u", 1227 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1228 #else 1229 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1230 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u", 1231 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1232 #endif 1233 1234 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1235 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1236 1237 #ifdef TCPDEBUG 1238 /* 1239 * Trace. 1240 */ 1241 if (so->so_options & SO_DEBUG) { 1242 u_short save = 0; 1243 #ifdef INET6 1244 if (!isipv6) 1245 #endif 1246 { 1247 save = ipov->ih_len; 1248 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1249 } 1250 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1251 #ifdef INET6 1252 if (!isipv6) 1253 #endif 1254 ipov->ih_len = save; 1255 } 1256 #endif /* TCPDEBUG */ 1257 1258 /* 1259 * Fill in IP length and desired time to live and 1260 * send to IP level. There should be a better way 1261 * to handle ttl and tos; we could keep them in 1262 * the template, but need a way to checksum without them. 1263 */ 1264 /* 1265 * m->m_pkthdr.len should have been set before checksum calculation, 1266 * because in6_cksum() need it. 1267 */ 1268 #ifdef INET6 1269 if (isipv6) { 1270 struct route_in6 ro; 1271 1272 bzero(&ro, sizeof(ro)); 1273 /* 1274 * we separately set hoplimit for every segment, since the 1275 * user might want to change the value via setsockopt. 1276 * Also, desired default hop limit might be changed via 1277 * Neighbor Discovery. 1278 */ 1279 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1280 1281 /* 1282 * Set the packet size here for the benefit of DTrace probes. 1283 * ip6_output() will set it properly; it's supposed to include 1284 * the option header lengths as well. 1285 */ 1286 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1287 1288 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) 1289 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1290 else 1291 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1292 1293 if (tp->t_state == TCPS_SYN_SENT) 1294 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1295 1296 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1297 1298 /* TODO: IPv6 IP6TOS_ECT bit on */ 1299 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro, 1300 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1301 NULL, NULL, tp->t_inpcb); 1302 1303 if (error == EMSGSIZE && ro.ro_rt != NULL) 1304 mtu = ro.ro_rt->rt_mtu; 1305 RO_RTFREE(&ro); 1306 } 1307 #endif /* INET6 */ 1308 #if defined(INET) && defined(INET6) 1309 else 1310 #endif 1311 #ifdef INET 1312 { 1313 struct route ro; 1314 1315 bzero(&ro, sizeof(ro)); 1316 ip->ip_len = htons(m->m_pkthdr.len); 1317 #ifdef INET6 1318 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1319 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1320 #endif /* INET6 */ 1321 /* 1322 * If we do path MTU discovery, then we set DF on every packet. 1323 * This might not be the best thing to do according to RFC3390 1324 * Section 2. However the tcp hostcache migitates the problem 1325 * so it affects only the first tcp connection with a host. 1326 * 1327 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1328 */ 1329 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) { 1330 ip->ip_off |= htons(IP_DF); 1331 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1332 } else { 1333 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1334 } 1335 1336 if (tp->t_state == TCPS_SYN_SENT) 1337 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1338 1339 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1340 1341 error = ip_output(m, tp->t_inpcb->inp_options, &ro, 1342 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1343 tp->t_inpcb); 1344 1345 if (error == EMSGSIZE && ro.ro_rt != NULL) 1346 mtu = ro.ro_rt->rt_mtu; 1347 RO_RTFREE(&ro); 1348 } 1349 #endif /* INET */ 1350 1351 out: 1352 /* 1353 * In transmit state, time the transmission and arrange for 1354 * the retransmit. In persist state, just set snd_max. 1355 */ 1356 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1357 !tcp_timer_active(tp, TT_PERSIST)) { 1358 tcp_seq startseq = tp->snd_nxt; 1359 1360 /* 1361 * Advance snd_nxt over sequence space of this segment. 1362 */ 1363 if (flags & (TH_SYN|TH_FIN)) { 1364 if (flags & TH_SYN) 1365 tp->snd_nxt++; 1366 if (flags & TH_FIN) { 1367 tp->snd_nxt++; 1368 tp->t_flags |= TF_SENTFIN; 1369 } 1370 } 1371 if (sack_rxmit) 1372 goto timer; 1373 tp->snd_nxt += len; 1374 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1375 tp->snd_max = tp->snd_nxt; 1376 /* 1377 * Time this transmission if not a retransmission and 1378 * not currently timing anything. 1379 */ 1380 if (tp->t_rtttime == 0) { 1381 tp->t_rtttime = ticks; 1382 tp->t_rtseq = startseq; 1383 TCPSTAT_INC(tcps_segstimed); 1384 } 1385 } 1386 1387 /* 1388 * Set retransmit timer if not currently set, 1389 * and not doing a pure ack or a keep-alive probe. 1390 * Initial value for retransmit timer is smoothed 1391 * round-trip time + 2 * round-trip time variance. 1392 * Initialize shift counter which is used for backoff 1393 * of retransmit time. 1394 */ 1395 timer: 1396 if (!tcp_timer_active(tp, TT_REXMT) && 1397 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1398 (tp->snd_nxt != tp->snd_una))) { 1399 if (tcp_timer_active(tp, TT_PERSIST)) { 1400 tcp_timer_activate(tp, TT_PERSIST, 0); 1401 tp->t_rxtshift = 0; 1402 } 1403 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1404 } else if (len == 0 && sbavail(&so->so_snd) && 1405 !tcp_timer_active(tp, TT_REXMT) && 1406 !tcp_timer_active(tp, TT_PERSIST)) { 1407 /* 1408 * Avoid a situation where we do not set persist timer 1409 * after a zero window condition. For example: 1410 * 1) A -> B: packet with enough data to fill the window 1411 * 2) B -> A: ACK for #1 + new data (0 window 1412 * advertisement) 1413 * 3) A -> B: ACK for #2, 0 len packet 1414 * 1415 * In this case, A will not activate the persist timer, 1416 * because it chose to send a packet. Unless tcp_output 1417 * is called for some other reason (delayed ack timer, 1418 * another input packet from B, socket syscall), A will 1419 * not send zero window probes. 1420 * 1421 * So, if you send a 0-length packet, but there is data 1422 * in the socket buffer, and neither the rexmt or 1423 * persist timer is already set, then activate the 1424 * persist timer. 1425 */ 1426 tp->t_rxtshift = 0; 1427 tcp_setpersist(tp); 1428 } 1429 } else { 1430 /* 1431 * Persist case, update snd_max but since we are in 1432 * persist mode (no window) we do not update snd_nxt. 1433 */ 1434 int xlen = len; 1435 if (flags & TH_SYN) 1436 ++xlen; 1437 if (flags & TH_FIN) { 1438 ++xlen; 1439 tp->t_flags |= TF_SENTFIN; 1440 } 1441 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1442 tp->snd_max = tp->snd_nxt + len; 1443 } 1444 1445 if (error) { 1446 1447 /* 1448 * We know that the packet was lost, so back out the 1449 * sequence number advance, if any. 1450 * 1451 * If the error is EPERM the packet got blocked by the 1452 * local firewall. Normally we should terminate the 1453 * connection but the blocking may have been spurious 1454 * due to a firewall reconfiguration cycle. So we treat 1455 * it like a packet loss and let the retransmit timer and 1456 * timeouts do their work over time. 1457 * XXX: It is a POLA question whether calling tcp_drop right 1458 * away would be the really correct behavior instead. 1459 */ 1460 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1461 !tcp_timer_active(tp, TT_PERSIST)) && 1462 ((flags & TH_SYN) == 0) && 1463 (error != EPERM)) { 1464 if (sack_rxmit) { 1465 p->rxmit -= len; 1466 tp->sackhint.sack_bytes_rexmit -= len; 1467 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1468 ("sackhint bytes rtx >= 0")); 1469 } else 1470 tp->snd_nxt -= len; 1471 } 1472 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1473 switch (error) { 1474 case EPERM: 1475 tp->t_softerror = error; 1476 return (error); 1477 case ENOBUFS: 1478 if (!tcp_timer_active(tp, TT_REXMT) && 1479 !tcp_timer_active(tp, TT_PERSIST)) 1480 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1481 tp->snd_cwnd = tp->t_maxseg; 1482 return (0); 1483 case EMSGSIZE: 1484 /* 1485 * For some reason the interface we used initially 1486 * to send segments changed to another or lowered 1487 * its MTU. 1488 * If TSO was active we either got an interface 1489 * without TSO capabilits or TSO was turned off. 1490 * If we obtained mtu from ip_output() then update 1491 * it and try again. 1492 */ 1493 if (tso) 1494 tp->t_flags &= ~TF_TSO; 1495 if (mtu != 0) { 1496 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1497 goto again; 1498 } 1499 return (error); 1500 case EHOSTDOWN: 1501 case EHOSTUNREACH: 1502 case ENETDOWN: 1503 case ENETUNREACH: 1504 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1505 tp->t_softerror = error; 1506 return (0); 1507 } 1508 /* FALLTHROUGH */ 1509 default: 1510 return (error); 1511 } 1512 } 1513 TCPSTAT_INC(tcps_sndtotal); 1514 1515 /* 1516 * Data sent (as far as we can tell). 1517 * If this advertises a larger window than any other segment, 1518 * then remember the size of the advertised window. 1519 * Any pending ACK has now been sent. 1520 */ 1521 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1522 tp->rcv_adv = tp->rcv_nxt + recwin; 1523 tp->last_ack_sent = tp->rcv_nxt; 1524 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1525 if (tcp_timer_active(tp, TT_DELACK)) 1526 tcp_timer_activate(tp, TT_DELACK, 0); 1527 #if 0 1528 /* 1529 * This completely breaks TCP if newreno is turned on. What happens 1530 * is that if delayed-acks are turned on on the receiver, this code 1531 * on the transmitter effectively destroys the TCP window, forcing 1532 * it to four packets (1.5Kx4 = 6K window). 1533 */ 1534 if (sendalot && --maxburst) 1535 goto again; 1536 #endif 1537 if (sendalot) 1538 goto again; 1539 return (0); 1540 } 1541 1542 void 1543 tcp_setpersist(struct tcpcb *tp) 1544 { 1545 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1546 int tt; 1547 1548 tp->t_flags &= ~TF_PREVVALID; 1549 if (tcp_timer_active(tp, TT_REXMT)) 1550 panic("tcp_setpersist: retransmit pending"); 1551 /* 1552 * Start/restart persistance timer. 1553 */ 1554 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1555 TCPTV_PERSMIN, TCPTV_PERSMAX); 1556 tcp_timer_activate(tp, TT_PERSIST, tt); 1557 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1558 tp->t_rxtshift++; 1559 } 1560 1561 /* 1562 * Insert TCP options according to the supplied parameters to the place 1563 * optp in a consistent way. Can handle unaligned destinations. 1564 * 1565 * The order of the option processing is crucial for optimal packing and 1566 * alignment for the scarce option space. 1567 * 1568 * The optimal order for a SYN/SYN-ACK segment is: 1569 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1570 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1571 * 1572 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1573 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1574 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1575 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1576 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1577 */ 1578 int 1579 tcp_addoptions(struct tcpopt *to, u_char *optp) 1580 { 1581 u_int mask, optlen = 0; 1582 1583 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1584 if ((to->to_flags & mask) != mask) 1585 continue; 1586 if (optlen == TCP_MAXOLEN) 1587 break; 1588 switch (to->to_flags & mask) { 1589 case TOF_MSS: 1590 while (optlen % 4) { 1591 optlen += TCPOLEN_NOP; 1592 *optp++ = TCPOPT_NOP; 1593 } 1594 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1595 continue; 1596 optlen += TCPOLEN_MAXSEG; 1597 *optp++ = TCPOPT_MAXSEG; 1598 *optp++ = TCPOLEN_MAXSEG; 1599 to->to_mss = htons(to->to_mss); 1600 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1601 optp += sizeof(to->to_mss); 1602 break; 1603 case TOF_SCALE: 1604 while (!optlen || optlen % 2 != 1) { 1605 optlen += TCPOLEN_NOP; 1606 *optp++ = TCPOPT_NOP; 1607 } 1608 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1609 continue; 1610 optlen += TCPOLEN_WINDOW; 1611 *optp++ = TCPOPT_WINDOW; 1612 *optp++ = TCPOLEN_WINDOW; 1613 *optp++ = to->to_wscale; 1614 break; 1615 case TOF_SACKPERM: 1616 while (optlen % 2) { 1617 optlen += TCPOLEN_NOP; 1618 *optp++ = TCPOPT_NOP; 1619 } 1620 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1621 continue; 1622 optlen += TCPOLEN_SACK_PERMITTED; 1623 *optp++ = TCPOPT_SACK_PERMITTED; 1624 *optp++ = TCPOLEN_SACK_PERMITTED; 1625 break; 1626 case TOF_TS: 1627 while (!optlen || optlen % 4 != 2) { 1628 optlen += TCPOLEN_NOP; 1629 *optp++ = TCPOPT_NOP; 1630 } 1631 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1632 continue; 1633 optlen += TCPOLEN_TIMESTAMP; 1634 *optp++ = TCPOPT_TIMESTAMP; 1635 *optp++ = TCPOLEN_TIMESTAMP; 1636 to->to_tsval = htonl(to->to_tsval); 1637 to->to_tsecr = htonl(to->to_tsecr); 1638 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1639 optp += sizeof(to->to_tsval); 1640 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1641 optp += sizeof(to->to_tsecr); 1642 break; 1643 case TOF_SIGNATURE: 1644 { 1645 int siglen = TCPOLEN_SIGNATURE - 2; 1646 1647 while (!optlen || optlen % 4 != 2) { 1648 optlen += TCPOLEN_NOP; 1649 *optp++ = TCPOPT_NOP; 1650 } 1651 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1652 continue; 1653 optlen += TCPOLEN_SIGNATURE; 1654 *optp++ = TCPOPT_SIGNATURE; 1655 *optp++ = TCPOLEN_SIGNATURE; 1656 to->to_signature = optp; 1657 while (siglen--) 1658 *optp++ = 0; 1659 break; 1660 } 1661 case TOF_SACK: 1662 { 1663 int sackblks = 0; 1664 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1665 tcp_seq sack_seq; 1666 1667 while (!optlen || optlen % 4 != 2) { 1668 optlen += TCPOLEN_NOP; 1669 *optp++ = TCPOPT_NOP; 1670 } 1671 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1672 continue; 1673 optlen += TCPOLEN_SACKHDR; 1674 *optp++ = TCPOPT_SACK; 1675 sackblks = min(to->to_nsacks, 1676 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1677 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1678 while (sackblks--) { 1679 sack_seq = htonl(sack->start); 1680 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1681 optp += sizeof(sack_seq); 1682 sack_seq = htonl(sack->end); 1683 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1684 optp += sizeof(sack_seq); 1685 optlen += TCPOLEN_SACK; 1686 sack++; 1687 } 1688 TCPSTAT_INC(tcps_sack_send_blocks); 1689 break; 1690 } 1691 default: 1692 panic("%s: unknown TCP option type", __func__); 1693 break; 1694 } 1695 } 1696 1697 /* Terminate and pad TCP options to a 4 byte boundary. */ 1698 if (optlen % 4) { 1699 optlen += TCPOLEN_EOL; 1700 *optp++ = TCPOPT_EOL; 1701 } 1702 /* 1703 * According to RFC 793 (STD0007): 1704 * "The content of the header beyond the End-of-Option option 1705 * must be header padding (i.e., zero)." 1706 * and later: "The padding is composed of zeros." 1707 */ 1708 while (optlen % 4) { 1709 optlen += TCPOLEN_PAD; 1710 *optp++ = TCPOPT_PAD; 1711 } 1712 1713 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1714 return (optlen); 1715 } 1716