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_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, 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_VNET_INT(_net_inet_tcp, OID_AUTO, tso, 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_VNET_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, 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_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, 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_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, 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_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, 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 < so->so_snd.sb_cc) 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(so->so_snd.sb_cc, sendwin) - off); 352 else { 353 long cwin; 354 355 /* 356 * We are inside of a SACK recovery episode and are 357 * sending new data, having retransmitted all the 358 * data possible in the scoreboard. 359 */ 360 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 361 - off); 362 /* 363 * Don't remove this (len > 0) check ! 364 * We explicitly check for len > 0 here (although it 365 * isn't really necessary), to work around a gcc 366 * optimization issue - to force gcc to compute 367 * len above. Without this check, the computation 368 * of len is bungled by the optimizer. 369 */ 370 if (len > 0) { 371 cwin = tp->snd_cwnd - 372 (tp->snd_nxt - tp->sack_newdata) - 373 sack_bytes_rxmt; 374 if (cwin < 0) 375 cwin = 0; 376 len = lmin(len, cwin); 377 } 378 } 379 } 380 381 /* 382 * Lop off SYN bit if it has already been sent. However, if this 383 * is SYN-SENT state and if segment contains data and if we don't 384 * know that foreign host supports TAO, suppress sending segment. 385 */ 386 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 387 if (tp->t_state != TCPS_SYN_RECEIVED) 388 flags &= ~TH_SYN; 389 off--, len++; 390 } 391 392 /* 393 * Be careful not to send data and/or FIN on SYN segments. 394 * This measure is needed to prevent interoperability problems 395 * with not fully conformant TCP implementations. 396 */ 397 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 398 len = 0; 399 flags &= ~TH_FIN; 400 } 401 402 if (len < 0) { 403 /* 404 * If FIN has been sent but not acked, 405 * but we haven't been called to retransmit, 406 * len will be < 0. Otherwise, window shrank 407 * after we sent into it. If window shrank to 0, 408 * cancel pending retransmit, pull snd_nxt back 409 * to (closed) window, and set the persist timer 410 * if it isn't already going. If the window didn't 411 * close completely, just wait for an ACK. 412 */ 413 len = 0; 414 if (sendwin == 0) { 415 tcp_timer_activate(tp, TT_REXMT, 0); 416 tp->t_rxtshift = 0; 417 tp->snd_nxt = tp->snd_una; 418 if (!tcp_timer_active(tp, TT_PERSIST)) 419 tcp_setpersist(tp); 420 } 421 } 422 423 /* len will be >= 0 after this point. */ 424 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 425 426 /* 427 * Automatic sizing of send socket buffer. Often the send buffer 428 * size is not optimally adjusted to the actual network conditions 429 * at hand (delay bandwidth product). Setting the buffer size too 430 * small limits throughput on links with high bandwidth and high 431 * delay (eg. trans-continental/oceanic links). Setting the 432 * buffer size too big consumes too much real kernel memory, 433 * especially with many connections on busy servers. 434 * 435 * The criteria to step up the send buffer one notch are: 436 * 1. receive window of remote host is larger than send buffer 437 * (with a fudge factor of 5/4th); 438 * 2. send buffer is filled to 7/8th with data (so we actually 439 * have data to make use of it); 440 * 3. send buffer fill has not hit maximal automatic size; 441 * 4. our send window (slow start and cogestion controlled) is 442 * larger than sent but unacknowledged data in send buffer. 443 * 444 * The remote host receive window scaling factor may limit the 445 * growing of the send buffer before it reaches its allowed 446 * maximum. 447 * 448 * It scales directly with slow start or congestion window 449 * and does at most one step per received ACK. This fast 450 * scaling has the drawback of growing the send buffer beyond 451 * what is strictly necessary to make full use of a given 452 * delay*bandwith product. However testing has shown this not 453 * to be much of an problem. At worst we are trading wasting 454 * of available bandwith (the non-use of it) for wasting some 455 * socket buffer memory. 456 * 457 * TODO: Shrink send buffer during idle periods together 458 * with congestion window. Requires another timer. Has to 459 * wait for upcoming tcp timer rewrite. 460 */ 461 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 462 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 463 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 464 so->so_snd.sb_cc < V_tcp_autosndbuf_max && 465 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 466 if (!sbreserve_locked(&so->so_snd, 467 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 468 V_tcp_autosndbuf_max), so, curthread)) 469 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 470 } 471 } 472 473 /* 474 * Decide if we can use TCP Segmentation Offloading (if supported by 475 * hardware). 476 * 477 * TSO may only be used if we are in a pure bulk sending state. The 478 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 479 * IP options prevent using TSO. With TSO the TCP header is the same 480 * (except for the sequence number) for all generated packets. This 481 * makes it impossible to transmit any options which vary per generated 482 * segment or packet. 483 */ 484 #ifdef IPSEC 485 /* 486 * Pre-calculate here as we save another lookup into the darknesses 487 * of IPsec that way and can actually decide if TSO is ok. 488 */ 489 ipsec_optlen = ipsec_hdrsiz_tcp(tp); 490 #endif 491 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 492 ((tp->t_flags & TF_SIGNATURE) == 0) && 493 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 494 #ifdef IPSEC 495 ipsec_optlen == 0 && 496 #endif 497 tp->t_inpcb->inp_options == NULL && 498 tp->t_inpcb->in6p_options == NULL) 499 tso = 1; 500 501 if (sack_rxmit) { 502 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 503 flags &= ~TH_FIN; 504 } else { 505 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 506 flags &= ~TH_FIN; 507 } 508 509 recwin = sbspace(&so->so_rcv); 510 511 /* 512 * Sender silly window avoidance. We transmit under the following 513 * conditions when len is non-zero: 514 * 515 * - We have a full segment (or more with TSO) 516 * - This is the last buffer in a write()/send() and we are 517 * either idle or running NODELAY 518 * - we've timed out (e.g. persist timer) 519 * - we have more then 1/2 the maximum send window's worth of 520 * data (receiver may be limited the window size) 521 * - we need to retransmit 522 */ 523 if (len) { 524 if (len >= tp->t_maxseg) 525 goto send; 526 /* 527 * NOTE! on localhost connections an 'ack' from the remote 528 * end may occur synchronously with the output and cause 529 * us to flush a buffer queued with moretocome. XXX 530 * 531 * note: the len + off check is almost certainly unnecessary. 532 */ 533 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 534 (idle || (tp->t_flags & TF_NODELAY)) && 535 len + off >= so->so_snd.sb_cc && 536 (tp->t_flags & TF_NOPUSH) == 0) { 537 goto send; 538 } 539 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 540 goto send; 541 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 542 goto send; 543 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 544 goto send; 545 if (sack_rxmit) 546 goto send; 547 } 548 549 /* 550 * Sending of standalone window updates. 551 * 552 * Window updates are important when we close our window due to a 553 * full socket buffer and are opening it again after the application 554 * reads data from it. Once the window has opened again and the 555 * remote end starts to send again the ACK clock takes over and 556 * provides the most current window information. 557 * 558 * We must avoid the silly window syndrome whereas every read 559 * from the receive buffer, no matter how small, causes a window 560 * update to be sent. We also should avoid sending a flurry of 561 * window updates when the socket buffer had queued a lot of data 562 * and the application is doing small reads. 563 * 564 * Prevent a flurry of pointless window updates by only sending 565 * an update when we can increase the advertized window by more 566 * than 1/4th of the socket buffer capacity. When the buffer is 567 * getting full or is very small be more aggressive and send an 568 * update whenever we can increase by two mss sized segments. 569 * In all other situations the ACK's to new incoming data will 570 * carry further window increases. 571 * 572 * Don't send an independent window update if a delayed 573 * ACK is pending (it will get piggy-backed on it) or the 574 * remote side already has done a half-close and won't send 575 * more data. Skip this if the connection is in T/TCP 576 * half-open state. 577 */ 578 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 579 !(tp->t_flags & TF_DELACK) && 580 !TCPS_HAVERCVDFIN(tp->t_state)) { 581 /* 582 * "adv" is the amount we could increase the window, 583 * taking into account that we are limited by 584 * TCP_MAXWIN << tp->rcv_scale. 585 */ 586 long adv; 587 int oldwin; 588 589 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 590 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 591 oldwin = (tp->rcv_adv - tp->rcv_nxt); 592 adv -= oldwin; 593 } else 594 oldwin = 0; 595 596 /* 597 * If the new window size ends up being the same as the old 598 * size when it is scaled, then don't force a window update. 599 */ 600 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 601 goto dontupdate; 602 603 if (adv >= (long)(2 * tp->t_maxseg) && 604 (adv >= (long)(so->so_rcv.sb_hiwat / 4) || 605 recwin <= (long)(so->so_rcv.sb_hiwat / 8) || 606 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 607 goto send; 608 } 609 dontupdate: 610 611 /* 612 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 613 * is also a catch-all for the retransmit timer timeout case. 614 */ 615 if (tp->t_flags & TF_ACKNOW) 616 goto send; 617 if ((flags & TH_RST) || 618 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 619 goto send; 620 if (SEQ_GT(tp->snd_up, tp->snd_una)) 621 goto send; 622 /* 623 * If our state indicates that FIN should be sent 624 * and we have not yet done so, then we need to send. 625 */ 626 if (flags & TH_FIN && 627 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 628 goto send; 629 /* 630 * In SACK, it is possible for tcp_output to fail to send a segment 631 * after the retransmission timer has been turned off. Make sure 632 * that the retransmission timer is set. 633 */ 634 if ((tp->t_flags & TF_SACK_PERMIT) && 635 SEQ_GT(tp->snd_max, tp->snd_una) && 636 !tcp_timer_active(tp, TT_REXMT) && 637 !tcp_timer_active(tp, TT_PERSIST)) { 638 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 639 goto just_return; 640 } 641 /* 642 * TCP window updates are not reliable, rather a polling protocol 643 * using ``persist'' packets is used to insure receipt of window 644 * updates. The three ``states'' for the output side are: 645 * idle not doing retransmits or persists 646 * persisting to move a small or zero window 647 * (re)transmitting and thereby not persisting 648 * 649 * tcp_timer_active(tp, TT_PERSIST) 650 * is true when we are in persist state. 651 * (tp->t_flags & TF_FORCEDATA) 652 * is set when we are called to send a persist packet. 653 * tcp_timer_active(tp, TT_REXMT) 654 * is set when we are retransmitting 655 * The output side is idle when both timers are zero. 656 * 657 * If send window is too small, there is data to transmit, and no 658 * retransmit or persist is pending, then go to persist state. 659 * If nothing happens soon, send when timer expires: 660 * if window is nonzero, transmit what we can, 661 * otherwise force out a byte. 662 */ 663 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) && 664 !tcp_timer_active(tp, TT_PERSIST)) { 665 tp->t_rxtshift = 0; 666 tcp_setpersist(tp); 667 } 668 669 /* 670 * No reason to send a segment, just return. 671 */ 672 just_return: 673 SOCKBUF_UNLOCK(&so->so_snd); 674 return (0); 675 676 send: 677 SOCKBUF_LOCK_ASSERT(&so->so_snd); 678 /* 679 * Before ESTABLISHED, force sending of initial options 680 * unless TCP set not to do any options. 681 * NOTE: we assume that the IP/TCP header plus TCP options 682 * always fit in a single mbuf, leaving room for a maximum 683 * link header, i.e. 684 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 685 */ 686 optlen = 0; 687 #ifdef INET6 688 if (isipv6) 689 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 690 else 691 #endif 692 hdrlen = sizeof (struct tcpiphdr); 693 694 /* 695 * Compute options for segment. 696 * We only have to care about SYN and established connection 697 * segments. Options for SYN-ACK segments are handled in TCP 698 * syncache. 699 */ 700 if ((tp->t_flags & TF_NOOPT) == 0) { 701 to.to_flags = 0; 702 /* Maximum segment size. */ 703 if (flags & TH_SYN) { 704 tp->snd_nxt = tp->iss; 705 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 706 to.to_flags |= TOF_MSS; 707 } 708 /* Window scaling. */ 709 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 710 to.to_wscale = tp->request_r_scale; 711 to.to_flags |= TOF_SCALE; 712 } 713 /* Timestamps. */ 714 if ((tp->t_flags & TF_RCVD_TSTMP) || 715 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 716 to.to_tsval = tcp_ts_getticks() + tp->ts_offset; 717 to.to_tsecr = tp->ts_recent; 718 to.to_flags |= TOF_TS; 719 /* Set receive buffer autosizing timestamp. */ 720 if (tp->rfbuf_ts == 0 && 721 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 722 tp->rfbuf_ts = tcp_ts_getticks(); 723 } 724 /* Selective ACK's. */ 725 if (tp->t_flags & TF_SACK_PERMIT) { 726 if (flags & TH_SYN) 727 to.to_flags |= TOF_SACKPERM; 728 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 729 (tp->t_flags & TF_SACK_PERMIT) && 730 tp->rcv_numsacks > 0) { 731 to.to_flags |= TOF_SACK; 732 to.to_nsacks = tp->rcv_numsacks; 733 to.to_sacks = (u_char *)tp->sackblks; 734 } 735 } 736 #ifdef TCP_SIGNATURE 737 /* TCP-MD5 (RFC2385). */ 738 if (tp->t_flags & TF_SIGNATURE) 739 to.to_flags |= TOF_SIGNATURE; 740 #endif /* TCP_SIGNATURE */ 741 742 /* Processing the options. */ 743 hdrlen += optlen = tcp_addoptions(&to, opt); 744 } 745 746 #ifdef INET6 747 if (isipv6) 748 ipoptlen = ip6_optlen(tp->t_inpcb); 749 else 750 #endif 751 if (tp->t_inpcb->inp_options) 752 ipoptlen = tp->t_inpcb->inp_options->m_len - 753 offsetof(struct ipoption, ipopt_list); 754 else 755 ipoptlen = 0; 756 #ifdef IPSEC 757 ipoptlen += ipsec_optlen; 758 #endif 759 760 /* 761 * Adjust data length if insertion of options will 762 * bump the packet length beyond the t_maxopd length. 763 * Clear the FIN bit because we cut off the tail of 764 * the segment. 765 */ 766 if (len + optlen + ipoptlen > tp->t_maxopd) { 767 flags &= ~TH_FIN; 768 769 if (tso) { 770 KASSERT(ipoptlen == 0, 771 ("%s: TSO can't do IP options", __func__)); 772 773 /* 774 * Limit a burst to t_tsomax minus IP, 775 * TCP and options length to keep ip->ip_len 776 * from overflowing or exceeding the maximum 777 * length allowed by the network interface. 778 */ 779 if (len > tp->t_tsomax - hdrlen) { 780 len = tp->t_tsomax - hdrlen; 781 sendalot = 1; 782 } 783 784 /* 785 * Prevent the last segment from being 786 * fractional unless the send sockbuf can 787 * be emptied. 788 */ 789 if (sendalot && off + len < so->so_snd.sb_cc) { 790 len -= len % (tp->t_maxopd - optlen); 791 sendalot = 1; 792 } 793 794 /* 795 * Send the FIN in a separate segment 796 * after the bulk sending is done. 797 * We don't trust the TSO implementations 798 * to clear the FIN flag on all but the 799 * last segment. 800 */ 801 if (tp->t_flags & TF_NEEDFIN) 802 sendalot = 1; 803 804 } else { 805 len = tp->t_maxopd - optlen - ipoptlen; 806 sendalot = 1; 807 } 808 } else 809 tso = 0; 810 811 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 812 ("%s: len > IP_MAXPACKET", __func__)); 813 814 /*#ifdef DIAGNOSTIC*/ 815 #ifdef INET6 816 if (max_linkhdr + hdrlen > MCLBYTES) 817 #else 818 if (max_linkhdr + hdrlen > MHLEN) 819 #endif 820 panic("tcphdr too big"); 821 /*#endif*/ 822 823 /* 824 * This KASSERT is here to catch edge cases at a well defined place. 825 * Before, those had triggered (random) panic conditions further down. 826 */ 827 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 828 829 /* 830 * Grab a header mbuf, attaching a copy of data to 831 * be transmitted, and initialize the header from 832 * the template for sends on this connection. 833 */ 834 if (len) { 835 struct mbuf *mb; 836 u_int moff; 837 838 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 839 TCPSTAT_INC(tcps_sndprobe); 840 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 841 tp->t_sndrexmitpack++; 842 TCPSTAT_INC(tcps_sndrexmitpack); 843 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 844 } else { 845 TCPSTAT_INC(tcps_sndpack); 846 TCPSTAT_ADD(tcps_sndbyte, len); 847 } 848 #ifdef INET6 849 if (MHLEN < hdrlen + max_linkhdr) 850 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 851 else 852 #endif 853 m = m_gethdr(M_NOWAIT, MT_DATA); 854 855 if (m == NULL) { 856 SOCKBUF_UNLOCK(&so->so_snd); 857 error = ENOBUFS; 858 sack_rxmit = 0; 859 goto out; 860 } 861 862 m->m_data += max_linkhdr; 863 m->m_len = hdrlen; 864 865 /* 866 * Start the m_copy functions from the closest mbuf 867 * to the offset in the socket buffer chain. 868 */ 869 mb = sbsndptr(&so->so_snd, off, len, &moff); 870 871 if (len <= MHLEN - hdrlen - max_linkhdr) { 872 m_copydata(mb, moff, (int)len, 873 mtod(m, caddr_t) + hdrlen); 874 m->m_len += len; 875 } else { 876 m->m_next = m_copy(mb, moff, (int)len); 877 if (m->m_next == NULL) { 878 SOCKBUF_UNLOCK(&so->so_snd); 879 (void) m_free(m); 880 error = ENOBUFS; 881 sack_rxmit = 0; 882 goto out; 883 } 884 } 885 886 /* 887 * If we're sending everything we've got, set PUSH. 888 * (This will keep happy those implementations which only 889 * give data to the user when a buffer fills or 890 * a PUSH comes in.) 891 */ 892 if (off + len == so->so_snd.sb_cc) 893 flags |= TH_PUSH; 894 SOCKBUF_UNLOCK(&so->so_snd); 895 } else { 896 SOCKBUF_UNLOCK(&so->so_snd); 897 if (tp->t_flags & TF_ACKNOW) 898 TCPSTAT_INC(tcps_sndacks); 899 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 900 TCPSTAT_INC(tcps_sndctrl); 901 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 902 TCPSTAT_INC(tcps_sndurg); 903 else 904 TCPSTAT_INC(tcps_sndwinup); 905 906 m = m_gethdr(M_NOWAIT, MT_DATA); 907 if (m == NULL) { 908 error = ENOBUFS; 909 sack_rxmit = 0; 910 goto out; 911 } 912 #ifdef INET6 913 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 914 MHLEN >= hdrlen) { 915 MH_ALIGN(m, hdrlen); 916 } else 917 #endif 918 m->m_data += max_linkhdr; 919 m->m_len = hdrlen; 920 } 921 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 922 m->m_pkthdr.rcvif = (struct ifnet *)0; 923 #ifdef MAC 924 mac_inpcb_create_mbuf(tp->t_inpcb, m); 925 #endif 926 #ifdef INET6 927 if (isipv6) { 928 ip6 = mtod(m, struct ip6_hdr *); 929 th = (struct tcphdr *)(ip6 + 1); 930 tcpip_fillheaders(tp->t_inpcb, ip6, th); 931 } else 932 #endif /* INET6 */ 933 { 934 ip = mtod(m, struct ip *); 935 ipov = (struct ipovly *)ip; 936 th = (struct tcphdr *)(ip + 1); 937 tcpip_fillheaders(tp->t_inpcb, ip, th); 938 } 939 940 /* 941 * Fill in fields, remembering maximum advertised 942 * window for use in delaying messages about window sizes. 943 * If resending a FIN, be sure not to use a new sequence number. 944 */ 945 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 946 tp->snd_nxt == tp->snd_max) 947 tp->snd_nxt--; 948 /* 949 * If we are starting a connection, send ECN setup 950 * SYN packet. If we are on a retransmit, we may 951 * resend those bits a number of times as per 952 * RFC 3168. 953 */ 954 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 955 if (tp->t_rxtshift >= 1) { 956 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 957 flags |= TH_ECE|TH_CWR; 958 } else 959 flags |= TH_ECE|TH_CWR; 960 } 961 962 if (tp->t_state == TCPS_ESTABLISHED && 963 (tp->t_flags & TF_ECN_PERMIT)) { 964 /* 965 * If the peer has ECN, mark data packets with 966 * ECN capable transmission (ECT). 967 * Ignore pure ack packets, retransmissions and window probes. 968 */ 969 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 970 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 971 #ifdef INET6 972 if (isipv6) 973 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 974 else 975 #endif 976 ip->ip_tos |= IPTOS_ECN_ECT0; 977 TCPSTAT_INC(tcps_ecn_ect0); 978 } 979 980 /* 981 * Reply with proper ECN notifications. 982 */ 983 if (tp->t_flags & TF_ECN_SND_CWR) { 984 flags |= TH_CWR; 985 tp->t_flags &= ~TF_ECN_SND_CWR; 986 } 987 if (tp->t_flags & TF_ECN_SND_ECE) 988 flags |= TH_ECE; 989 } 990 991 /* 992 * If we are doing retransmissions, then snd_nxt will 993 * not reflect the first unsent octet. For ACK only 994 * packets, we do not want the sequence number of the 995 * retransmitted packet, we want the sequence number 996 * of the next unsent octet. So, if there is no data 997 * (and no SYN or FIN), use snd_max instead of snd_nxt 998 * when filling in ti_seq. But if we are in persist 999 * state, snd_max might reflect one byte beyond the 1000 * right edge of the window, so use snd_nxt in that 1001 * case, since we know we aren't doing a retransmission. 1002 * (retransmit and persist are mutually exclusive...) 1003 */ 1004 if (sack_rxmit == 0) { 1005 if (len || (flags & (TH_SYN|TH_FIN)) || 1006 tcp_timer_active(tp, TT_PERSIST)) 1007 th->th_seq = htonl(tp->snd_nxt); 1008 else 1009 th->th_seq = htonl(tp->snd_max); 1010 } else { 1011 th->th_seq = htonl(p->rxmit); 1012 p->rxmit += len; 1013 tp->sackhint.sack_bytes_rexmit += len; 1014 } 1015 th->th_ack = htonl(tp->rcv_nxt); 1016 if (optlen) { 1017 bcopy(opt, th + 1, optlen); 1018 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1019 } 1020 th->th_flags = flags; 1021 /* 1022 * Calculate receive window. Don't shrink window, 1023 * but avoid silly window syndrome. 1024 */ 1025 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 1026 recwin < (long)tp->t_maxseg) 1027 recwin = 0; 1028 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1029 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 1030 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 1031 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 1032 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 1033 1034 /* 1035 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1036 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1037 * case is handled in syncache. 1038 */ 1039 if (flags & TH_SYN) 1040 th->th_win = htons((u_short) 1041 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1042 else 1043 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1044 1045 /* 1046 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1047 * a 0 window. This may cause the remote transmitter to stall. This 1048 * flag tells soreceive() to disable delayed acknowledgements when 1049 * draining the buffer. This can occur if the receiver is attempting 1050 * to read more data than can be buffered prior to transmitting on 1051 * the connection. 1052 */ 1053 if (th->th_win == 0) { 1054 tp->t_sndzerowin++; 1055 tp->t_flags |= TF_RXWIN0SENT; 1056 } else 1057 tp->t_flags &= ~TF_RXWIN0SENT; 1058 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1059 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1060 th->th_flags |= TH_URG; 1061 } else 1062 /* 1063 * If no urgent pointer to send, then we pull 1064 * the urgent pointer to the left edge of the send window 1065 * so that it doesn't drift into the send window on sequence 1066 * number wraparound. 1067 */ 1068 tp->snd_up = tp->snd_una; /* drag it along */ 1069 1070 #ifdef TCP_SIGNATURE 1071 if (tp->t_flags & TF_SIGNATURE) { 1072 int sigoff = to.to_signature - opt; 1073 tcp_signature_compute(m, 0, len, optlen, 1074 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1075 } 1076 #endif 1077 1078 /* 1079 * Put TCP length in extended header, and then 1080 * checksum extended header and data. 1081 */ 1082 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1083 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1084 #ifdef INET6 1085 if (isipv6) { 1086 /* 1087 * ip6_plen is not need to be filled now, and will be filled 1088 * in ip6_output. 1089 */ 1090 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1091 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1092 optlen + len, IPPROTO_TCP, 0); 1093 } 1094 #endif 1095 #if defined(INET6) && defined(INET) 1096 else 1097 #endif 1098 #ifdef INET 1099 { 1100 m->m_pkthdr.csum_flags = CSUM_TCP; 1101 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1102 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1103 1104 /* IP version must be set here for ipv4/ipv6 checking later */ 1105 KASSERT(ip->ip_v == IPVERSION, 1106 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1107 } 1108 #endif 1109 1110 /* 1111 * Enable TSO and specify the size of the segments. 1112 * The TCP pseudo header checksum is always provided. 1113 * XXX: Fixme: This is currently not the case for IPv6. 1114 */ 1115 if (tso) { 1116 KASSERT(len > tp->t_maxopd - optlen, 1117 ("%s: len <= tso_segsz", __func__)); 1118 m->m_pkthdr.csum_flags |= CSUM_TSO; 1119 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 1120 } 1121 1122 #ifdef IPSEC 1123 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1124 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u", 1125 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1126 #else 1127 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1128 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u", 1129 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1130 #endif 1131 1132 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1133 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1134 1135 #ifdef TCPDEBUG 1136 /* 1137 * Trace. 1138 */ 1139 if (so->so_options & SO_DEBUG) { 1140 u_short save = 0; 1141 #ifdef INET6 1142 if (!isipv6) 1143 #endif 1144 { 1145 save = ipov->ih_len; 1146 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1147 } 1148 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1149 #ifdef INET6 1150 if (!isipv6) 1151 #endif 1152 ipov->ih_len = save; 1153 } 1154 #endif /* TCPDEBUG */ 1155 1156 /* 1157 * Fill in IP length and desired time to live and 1158 * send to IP level. There should be a better way 1159 * to handle ttl and tos; we could keep them in 1160 * the template, but need a way to checksum without them. 1161 */ 1162 /* 1163 * m->m_pkthdr.len should have been set before cksum calcuration, 1164 * because in6_cksum() need it. 1165 */ 1166 #ifdef INET6 1167 if (isipv6) { 1168 struct route_in6 ro; 1169 1170 bzero(&ro, sizeof(ro)); 1171 /* 1172 * we separately set hoplimit for every segment, since the 1173 * user might want to change the value via setsockopt. 1174 * Also, desired default hop limit might be changed via 1175 * Neighbor Discovery. 1176 */ 1177 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1178 1179 /* 1180 * Set the packet size here for the benefit of DTrace probes. 1181 * ip6_output() will set it properly; it's supposed to include 1182 * the option header lengths as well. 1183 */ 1184 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1185 1186 if (tp->t_state == TCPS_SYN_SENT) 1187 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1188 1189 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1190 1191 /* TODO: IPv6 IP6TOS_ECT bit on */ 1192 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro, 1193 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1194 NULL, NULL, tp->t_inpcb); 1195 1196 if (error == EMSGSIZE && ro.ro_rt != NULL) 1197 mtu = ro.ro_rt->rt_mtu; 1198 RO_RTFREE(&ro); 1199 } 1200 #endif /* INET6 */ 1201 #if defined(INET) && defined(INET6) 1202 else 1203 #endif 1204 #ifdef INET 1205 { 1206 struct route ro; 1207 1208 bzero(&ro, sizeof(ro)); 1209 ip->ip_len = htons(m->m_pkthdr.len); 1210 #ifdef INET6 1211 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1212 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1213 #endif /* INET6 */ 1214 /* 1215 * If we do path MTU discovery, then we set DF on every packet. 1216 * This might not be the best thing to do according to RFC3390 1217 * Section 2. However the tcp hostcache migitates the problem 1218 * so it affects only the first tcp connection with a host. 1219 * 1220 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1221 */ 1222 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) 1223 ip->ip_off |= htons(IP_DF); 1224 1225 if (tp->t_state == TCPS_SYN_SENT) 1226 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1227 1228 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1229 1230 error = ip_output(m, tp->t_inpcb->inp_options, &ro, 1231 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1232 tp->t_inpcb); 1233 1234 if (error == EMSGSIZE && ro.ro_rt != NULL) 1235 mtu = ro.ro_rt->rt_mtu; 1236 RO_RTFREE(&ro); 1237 } 1238 #endif /* INET */ 1239 1240 out: 1241 /* 1242 * In transmit state, time the transmission and arrange for 1243 * the retransmit. In persist state, just set snd_max. 1244 */ 1245 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1246 !tcp_timer_active(tp, TT_PERSIST)) { 1247 tcp_seq startseq = tp->snd_nxt; 1248 1249 /* 1250 * Advance snd_nxt over sequence space of this segment. 1251 */ 1252 if (flags & (TH_SYN|TH_FIN)) { 1253 if (flags & TH_SYN) 1254 tp->snd_nxt++; 1255 if (flags & TH_FIN) { 1256 tp->snd_nxt++; 1257 tp->t_flags |= TF_SENTFIN; 1258 } 1259 } 1260 if (sack_rxmit) 1261 goto timer; 1262 tp->snd_nxt += len; 1263 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1264 tp->snd_max = tp->snd_nxt; 1265 /* 1266 * Time this transmission if not a retransmission and 1267 * not currently timing anything. 1268 */ 1269 if (tp->t_rtttime == 0) { 1270 tp->t_rtttime = ticks; 1271 tp->t_rtseq = startseq; 1272 TCPSTAT_INC(tcps_segstimed); 1273 } 1274 } 1275 1276 /* 1277 * Set retransmit timer if not currently set, 1278 * and not doing a pure ack or a keep-alive probe. 1279 * Initial value for retransmit timer is smoothed 1280 * round-trip time + 2 * round-trip time variance. 1281 * Initialize shift counter which is used for backoff 1282 * of retransmit time. 1283 */ 1284 timer: 1285 if (!tcp_timer_active(tp, TT_REXMT) && 1286 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1287 (tp->snd_nxt != tp->snd_una))) { 1288 if (tcp_timer_active(tp, TT_PERSIST)) { 1289 tcp_timer_activate(tp, TT_PERSIST, 0); 1290 tp->t_rxtshift = 0; 1291 } 1292 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1293 } 1294 } else { 1295 /* 1296 * Persist case, update snd_max but since we are in 1297 * persist mode (no window) we do not update snd_nxt. 1298 */ 1299 int xlen = len; 1300 if (flags & TH_SYN) 1301 ++xlen; 1302 if (flags & TH_FIN) { 1303 ++xlen; 1304 tp->t_flags |= TF_SENTFIN; 1305 } 1306 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1307 tp->snd_max = tp->snd_nxt + len; 1308 } 1309 1310 if (error) { 1311 1312 /* 1313 * We know that the packet was lost, so back out the 1314 * sequence number advance, if any. 1315 * 1316 * If the error is EPERM the packet got blocked by the 1317 * local firewall. Normally we should terminate the 1318 * connection but the blocking may have been spurious 1319 * due to a firewall reconfiguration cycle. So we treat 1320 * it like a packet loss and let the retransmit timer and 1321 * timeouts do their work over time. 1322 * XXX: It is a POLA question whether calling tcp_drop right 1323 * away would be the really correct behavior instead. 1324 */ 1325 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1326 !tcp_timer_active(tp, TT_PERSIST)) && 1327 ((flags & TH_SYN) == 0) && 1328 (error != EPERM)) { 1329 if (sack_rxmit) { 1330 p->rxmit -= len; 1331 tp->sackhint.sack_bytes_rexmit -= len; 1332 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1333 ("sackhint bytes rtx >= 0")); 1334 } else 1335 tp->snd_nxt -= len; 1336 } 1337 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1338 switch (error) { 1339 case EPERM: 1340 tp->t_softerror = error; 1341 return (error); 1342 case ENOBUFS: 1343 if (!tcp_timer_active(tp, TT_REXMT) && 1344 !tcp_timer_active(tp, TT_PERSIST)) 1345 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1346 tp->snd_cwnd = tp->t_maxseg; 1347 return (0); 1348 case EMSGSIZE: 1349 /* 1350 * For some reason the interface we used initially 1351 * to send segments changed to another or lowered 1352 * its MTU. 1353 * If TSO was active we either got an interface 1354 * without TSO capabilits or TSO was turned off. 1355 * If we obtained mtu from ip_output() then update 1356 * it and try again. 1357 */ 1358 if (tso) 1359 tp->t_flags &= ~TF_TSO; 1360 if (mtu != 0) { 1361 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1362 goto again; 1363 } 1364 return (error); 1365 case EHOSTDOWN: 1366 case EHOSTUNREACH: 1367 case ENETDOWN: 1368 case ENETUNREACH: 1369 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1370 tp->t_softerror = error; 1371 return (0); 1372 } 1373 /* FALLTHROUGH */ 1374 default: 1375 return (error); 1376 } 1377 } 1378 TCPSTAT_INC(tcps_sndtotal); 1379 1380 /* 1381 * Data sent (as far as we can tell). 1382 * If this advertises a larger window than any other segment, 1383 * then remember the size of the advertised window. 1384 * Any pending ACK has now been sent. 1385 */ 1386 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1387 tp->rcv_adv = tp->rcv_nxt + recwin; 1388 tp->last_ack_sent = tp->rcv_nxt; 1389 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1390 if (tcp_timer_active(tp, TT_DELACK)) 1391 tcp_timer_activate(tp, TT_DELACK, 0); 1392 #if 0 1393 /* 1394 * This completely breaks TCP if newreno is turned on. What happens 1395 * is that if delayed-acks are turned on on the receiver, this code 1396 * on the transmitter effectively destroys the TCP window, forcing 1397 * it to four packets (1.5Kx4 = 6K window). 1398 */ 1399 if (sendalot && --maxburst) 1400 goto again; 1401 #endif 1402 if (sendalot) 1403 goto again; 1404 return (0); 1405 } 1406 1407 void 1408 tcp_setpersist(struct tcpcb *tp) 1409 { 1410 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1411 int tt; 1412 1413 tp->t_flags &= ~TF_PREVVALID; 1414 if (tcp_timer_active(tp, TT_REXMT)) 1415 panic("tcp_setpersist: retransmit pending"); 1416 /* 1417 * Start/restart persistance timer. 1418 */ 1419 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1420 TCPTV_PERSMIN, TCPTV_PERSMAX); 1421 tcp_timer_activate(tp, TT_PERSIST, tt); 1422 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1423 tp->t_rxtshift++; 1424 } 1425 1426 /* 1427 * Insert TCP options according to the supplied parameters to the place 1428 * optp in a consistent way. Can handle unaligned destinations. 1429 * 1430 * The order of the option processing is crucial for optimal packing and 1431 * alignment for the scarce option space. 1432 * 1433 * The optimal order for a SYN/SYN-ACK segment is: 1434 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1435 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1436 * 1437 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1438 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1439 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1440 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1441 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1442 */ 1443 int 1444 tcp_addoptions(struct tcpopt *to, u_char *optp) 1445 { 1446 u_int mask, optlen = 0; 1447 1448 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1449 if ((to->to_flags & mask) != mask) 1450 continue; 1451 if (optlen == TCP_MAXOLEN) 1452 break; 1453 switch (to->to_flags & mask) { 1454 case TOF_MSS: 1455 while (optlen % 4) { 1456 optlen += TCPOLEN_NOP; 1457 *optp++ = TCPOPT_NOP; 1458 } 1459 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1460 continue; 1461 optlen += TCPOLEN_MAXSEG; 1462 *optp++ = TCPOPT_MAXSEG; 1463 *optp++ = TCPOLEN_MAXSEG; 1464 to->to_mss = htons(to->to_mss); 1465 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1466 optp += sizeof(to->to_mss); 1467 break; 1468 case TOF_SCALE: 1469 while (!optlen || optlen % 2 != 1) { 1470 optlen += TCPOLEN_NOP; 1471 *optp++ = TCPOPT_NOP; 1472 } 1473 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1474 continue; 1475 optlen += TCPOLEN_WINDOW; 1476 *optp++ = TCPOPT_WINDOW; 1477 *optp++ = TCPOLEN_WINDOW; 1478 *optp++ = to->to_wscale; 1479 break; 1480 case TOF_SACKPERM: 1481 while (optlen % 2) { 1482 optlen += TCPOLEN_NOP; 1483 *optp++ = TCPOPT_NOP; 1484 } 1485 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1486 continue; 1487 optlen += TCPOLEN_SACK_PERMITTED; 1488 *optp++ = TCPOPT_SACK_PERMITTED; 1489 *optp++ = TCPOLEN_SACK_PERMITTED; 1490 break; 1491 case TOF_TS: 1492 while (!optlen || optlen % 4 != 2) { 1493 optlen += TCPOLEN_NOP; 1494 *optp++ = TCPOPT_NOP; 1495 } 1496 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1497 continue; 1498 optlen += TCPOLEN_TIMESTAMP; 1499 *optp++ = TCPOPT_TIMESTAMP; 1500 *optp++ = TCPOLEN_TIMESTAMP; 1501 to->to_tsval = htonl(to->to_tsval); 1502 to->to_tsecr = htonl(to->to_tsecr); 1503 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1504 optp += sizeof(to->to_tsval); 1505 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1506 optp += sizeof(to->to_tsecr); 1507 break; 1508 case TOF_SIGNATURE: 1509 { 1510 int siglen = TCPOLEN_SIGNATURE - 2; 1511 1512 while (!optlen || optlen % 4 != 2) { 1513 optlen += TCPOLEN_NOP; 1514 *optp++ = TCPOPT_NOP; 1515 } 1516 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1517 continue; 1518 optlen += TCPOLEN_SIGNATURE; 1519 *optp++ = TCPOPT_SIGNATURE; 1520 *optp++ = TCPOLEN_SIGNATURE; 1521 to->to_signature = optp; 1522 while (siglen--) 1523 *optp++ = 0; 1524 break; 1525 } 1526 case TOF_SACK: 1527 { 1528 int sackblks = 0; 1529 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1530 tcp_seq sack_seq; 1531 1532 while (!optlen || optlen % 4 != 2) { 1533 optlen += TCPOLEN_NOP; 1534 *optp++ = TCPOPT_NOP; 1535 } 1536 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1537 continue; 1538 optlen += TCPOLEN_SACKHDR; 1539 *optp++ = TCPOPT_SACK; 1540 sackblks = min(to->to_nsacks, 1541 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1542 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1543 while (sackblks--) { 1544 sack_seq = htonl(sack->start); 1545 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1546 optp += sizeof(sack_seq); 1547 sack_seq = htonl(sack->end); 1548 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1549 optp += sizeof(sack_seq); 1550 optlen += TCPOLEN_SACK; 1551 sack++; 1552 } 1553 TCPSTAT_INC(tcps_sack_send_blocks); 1554 break; 1555 } 1556 default: 1557 panic("%s: unknown TCP option type", __func__); 1558 break; 1559 } 1560 } 1561 1562 /* Terminate and pad TCP options to a 4 byte boundary. */ 1563 if (optlen % 4) { 1564 optlen += TCPOLEN_EOL; 1565 *optp++ = TCPOPT_EOL; 1566 } 1567 /* 1568 * According to RFC 793 (STD0007): 1569 * "The content of the header beyond the End-of-Option option 1570 * must be header padding (i.e., zero)." 1571 * and later: "The padding is composed of zeros." 1572 */ 1573 while (optlen % 4) { 1574 optlen += TCPOLEN_PAD; 1575 *optp++ = TCPOPT_PAD; 1576 } 1577 1578 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1579 return (optlen); 1580 } 1581