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/kernel.h> 44 #include <sys/lock.h> 45 #include <sys/mbuf.h> 46 #include <sys/mutex.h> 47 #include <sys/protosw.h> 48 #include <sys/socket.h> 49 #include <sys/socketvar.h> 50 #include <sys/sysctl.h> 51 52 #include <net/if.h> 53 #include <net/route.h> 54 #include <net/vnet.h> 55 56 #include <netinet/in.h> 57 #include <netinet/in_systm.h> 58 #include <netinet/ip.h> 59 #include <netinet/in_pcb.h> 60 #include <netinet/ip_var.h> 61 #include <netinet/ip_options.h> 62 #ifdef INET6 63 #include <netinet6/in6_pcb.h> 64 #include <netinet/ip6.h> 65 #include <netinet6/ip6_var.h> 66 #endif 67 #include <netinet/tcp.h> 68 #define TCPOUTFLAGS 69 #include <netinet/tcp_fsm.h> 70 #include <netinet/tcp_seq.h> 71 #include <netinet/tcp_timer.h> 72 #include <netinet/tcp_var.h> 73 #include <netinet/tcpip.h> 74 #ifdef TCPDEBUG 75 #include <netinet/tcp_debug.h> 76 #endif 77 78 #ifdef IPSEC 79 #include <netipsec/ipsec.h> 80 #endif /*IPSEC*/ 81 82 #include <machine/in_cksum.h> 83 84 #include <security/mac/mac_framework.h> 85 86 #ifdef notyet 87 extern struct mbuf *m_copypack(); 88 #endif 89 90 VNET_DEFINE(int, path_mtu_discovery) = 1; 91 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, 92 &VNET_NAME(path_mtu_discovery), 1, 93 "Enable Path MTU Discovery"); 94 95 VNET_DEFINE(int, ss_fltsz) = 1; 96 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW, 97 &VNET_NAME(ss_fltsz), 1, 98 "Slow start flight size"); 99 100 VNET_DEFINE(int, ss_fltsz_local) = 4; 101 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, 102 CTLFLAG_RW, &VNET_NAME(ss_fltsz_local), 1, 103 "Slow start flight size for local networks"); 104 105 VNET_DEFINE(int, tcp_do_newreno) = 1; 106 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, 107 &VNET_NAME(tcp_do_newreno), 0, 108 "Enable NewReno Algorithms"); 109 110 VNET_DEFINE(int, tcp_do_tso) = 1; 111 #define V_tcp_do_tso VNET(tcp_do_tso) 112 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW, 113 &VNET_NAME(tcp_do_tso), 0, 114 "Enable TCP Segmentation Offload"); 115 116 VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 117 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) 118 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW, 119 &VNET_NAME(tcp_do_autosndbuf), 0, 120 "Enable automatic send buffer sizing"); 121 122 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 123 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) 124 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW, 125 &VNET_NAME(tcp_autosndbuf_inc), 0, 126 "Incrementor step size of automatic send buffer"); 127 128 VNET_DEFINE(int, tcp_autosndbuf_max) = 256*1024; 129 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) 130 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW, 131 &VNET_NAME(tcp_autosndbuf_max), 0, 132 "Max size of automatic send buffer"); 133 134 135 /* 136 * Tcp output routine: figure out what should be sent and send it. 137 */ 138 int 139 tcp_output(struct tcpcb *tp) 140 { 141 struct socket *so = tp->t_inpcb->inp_socket; 142 long len, recwin, sendwin; 143 int off, flags, error; 144 struct mbuf *m; 145 struct ip *ip = NULL; 146 struct ipovly *ipov = NULL; 147 struct tcphdr *th; 148 u_char opt[TCP_MAXOLEN]; 149 unsigned ipoptlen, optlen, hdrlen; 150 #ifdef IPSEC 151 unsigned ipsec_optlen = 0; 152 #endif 153 int idle, sendalot; 154 int sack_rxmit, sack_bytes_rxmt; 155 struct sackhole *p; 156 int tso = 0; 157 struct tcpopt to; 158 #if 0 159 int maxburst = TCP_MAXBURST; 160 #endif 161 #ifdef INET6 162 struct ip6_hdr *ip6 = NULL; 163 int isipv6; 164 165 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 166 #endif 167 168 INP_WLOCK_ASSERT(tp->t_inpcb); 169 170 /* 171 * Determine length of data that should be transmitted, 172 * and flags that will be used. 173 * If there is some data or critical controls (SYN, RST) 174 * to send, then transmit; otherwise, investigate further. 175 */ 176 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 177 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) { 178 /* 179 * We have been idle for "a while" and no acks are 180 * expected to clock out any data we send -- 181 * slow start to get ack "clock" running again. 182 * 183 * Set the slow-start flight size depending on whether 184 * this is a local network or not. 185 */ 186 int ss = V_ss_fltsz; 187 #ifdef INET6 188 if (isipv6) { 189 if (in6_localaddr(&tp->t_inpcb->in6p_faddr)) 190 ss = V_ss_fltsz_local; 191 } else 192 #endif /* INET6 */ 193 if (in_localaddr(tp->t_inpcb->inp_faddr)) 194 ss = V_ss_fltsz_local; 195 tp->snd_cwnd = tp->t_maxseg * ss; 196 } 197 tp->t_flags &= ~TF_LASTIDLE; 198 if (idle) { 199 if (tp->t_flags & TF_MORETOCOME) { 200 tp->t_flags |= TF_LASTIDLE; 201 idle = 0; 202 } 203 } 204 again: 205 /* 206 * If we've recently taken a timeout, snd_max will be greater than 207 * snd_nxt. There may be SACK information that allows us to avoid 208 * resending already delivered data. Adjust snd_nxt accordingly. 209 */ 210 if ((tp->t_flags & TF_SACK_PERMIT) && 211 SEQ_LT(tp->snd_nxt, tp->snd_max)) 212 tcp_sack_adjust(tp); 213 sendalot = 0; 214 off = tp->snd_nxt - tp->snd_una; 215 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 216 sendwin = min(sendwin, tp->snd_bwnd); 217 218 flags = tcp_outflags[tp->t_state]; 219 /* 220 * Send any SACK-generated retransmissions. If we're explicitly trying 221 * to send out new data (when sendalot is 1), bypass this function. 222 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 223 * we're replacing a (future) new transmission with a retransmission 224 * now, and we previously incremented snd_cwnd in tcp_input(). 225 */ 226 /* 227 * Still in sack recovery , reset rxmit flag to zero. 228 */ 229 sack_rxmit = 0; 230 sack_bytes_rxmt = 0; 231 len = 0; 232 p = NULL; 233 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) && 234 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 235 long cwin; 236 237 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 238 if (cwin < 0) 239 cwin = 0; 240 /* Do not retransmit SACK segments beyond snd_recover */ 241 if (SEQ_GT(p->end, tp->snd_recover)) { 242 /* 243 * (At least) part of sack hole extends beyond 244 * snd_recover. Check to see if we can rexmit data 245 * for this hole. 246 */ 247 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 248 /* 249 * Can't rexmit any more data for this hole. 250 * That data will be rexmitted in the next 251 * sack recovery episode, when snd_recover 252 * moves past p->rxmit. 253 */ 254 p = NULL; 255 goto after_sack_rexmit; 256 } else 257 /* Can rexmit part of the current hole */ 258 len = ((long)ulmin(cwin, 259 tp->snd_recover - p->rxmit)); 260 } else 261 len = ((long)ulmin(cwin, p->end - p->rxmit)); 262 off = p->rxmit - tp->snd_una; 263 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 264 __func__, off)); 265 if (len > 0) { 266 sack_rxmit = 1; 267 sendalot = 1; 268 TCPSTAT_INC(tcps_sack_rexmits); 269 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 270 min(len, tp->t_maxseg)); 271 } 272 } 273 after_sack_rexmit: 274 /* 275 * Get standard flags, and add SYN or FIN if requested by 'hidden' 276 * state flags. 277 */ 278 if (tp->t_flags & TF_NEEDFIN) 279 flags |= TH_FIN; 280 if (tp->t_flags & TF_NEEDSYN) 281 flags |= TH_SYN; 282 283 SOCKBUF_LOCK(&so->so_snd); 284 /* 285 * If in persist timeout with window of 0, send 1 byte. 286 * Otherwise, if window is small but nonzero 287 * and timer expired, we will send what we can 288 * and go to transmit state. 289 */ 290 if (tp->t_flags & TF_FORCEDATA) { 291 if (sendwin == 0) { 292 /* 293 * If we still have some data to send, then 294 * clear the FIN bit. Usually this would 295 * happen below when it realizes that we 296 * aren't sending all the data. However, 297 * if we have exactly 1 byte of unsent data, 298 * then it won't clear the FIN bit below, 299 * and if we are in persist state, we wind 300 * up sending the packet without recording 301 * that we sent the FIN bit. 302 * 303 * We can't just blindly clear the FIN bit, 304 * because if we don't have any more data 305 * to send then the probe will be the FIN 306 * itself. 307 */ 308 if (off < so->so_snd.sb_cc) 309 flags &= ~TH_FIN; 310 sendwin = 1; 311 } else { 312 tcp_timer_activate(tp, TT_PERSIST, 0); 313 tp->t_rxtshift = 0; 314 } 315 } 316 317 /* 318 * If snd_nxt == snd_max and we have transmitted a FIN, the 319 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 320 * a negative length. This can also occur when TCP opens up 321 * its congestion window while receiving additional duplicate 322 * acks after fast-retransmit because TCP will reset snd_nxt 323 * to snd_max after the fast-retransmit. 324 * 325 * In the normal retransmit-FIN-only case, however, snd_nxt will 326 * be set to snd_una, the offset will be 0, and the length may 327 * wind up 0. 328 * 329 * If sack_rxmit is true we are retransmitting from the scoreboard 330 * in which case len is already set. 331 */ 332 if (sack_rxmit == 0) { 333 if (sack_bytes_rxmt == 0) 334 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off); 335 else { 336 long cwin; 337 338 /* 339 * We are inside of a SACK recovery episode and are 340 * sending new data, having retransmitted all the 341 * data possible in the scoreboard. 342 */ 343 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 344 - off); 345 /* 346 * Don't remove this (len > 0) check ! 347 * We explicitly check for len > 0 here (although it 348 * isn't really necessary), to work around a gcc 349 * optimization issue - to force gcc to compute 350 * len above. Without this check, the computation 351 * of len is bungled by the optimizer. 352 */ 353 if (len > 0) { 354 cwin = tp->snd_cwnd - 355 (tp->snd_nxt - tp->sack_newdata) - 356 sack_bytes_rxmt; 357 if (cwin < 0) 358 cwin = 0; 359 len = lmin(len, cwin); 360 } 361 } 362 } 363 364 /* 365 * Lop off SYN bit if it has already been sent. However, if this 366 * is SYN-SENT state and if segment contains data and if we don't 367 * know that foreign host supports TAO, suppress sending segment. 368 */ 369 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 370 if (tp->t_state != TCPS_SYN_RECEIVED) 371 flags &= ~TH_SYN; 372 off--, len++; 373 } 374 375 /* 376 * Be careful not to send data and/or FIN on SYN segments. 377 * This measure is needed to prevent interoperability problems 378 * with not fully conformant TCP implementations. 379 */ 380 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 381 len = 0; 382 flags &= ~TH_FIN; 383 } 384 385 if (len < 0) { 386 /* 387 * If FIN has been sent but not acked, 388 * but we haven't been called to retransmit, 389 * len will be < 0. Otherwise, window shrank 390 * after we sent into it. If window shrank to 0, 391 * cancel pending retransmit, pull snd_nxt back 392 * to (closed) window, and set the persist timer 393 * if it isn't already going. If the window didn't 394 * close completely, just wait for an ACK. 395 */ 396 len = 0; 397 if (sendwin == 0) { 398 tcp_timer_activate(tp, TT_REXMT, 0); 399 tp->t_rxtshift = 0; 400 tp->snd_nxt = tp->snd_una; 401 if (!tcp_timer_active(tp, TT_PERSIST)) 402 tcp_setpersist(tp); 403 } 404 } 405 406 /* len will be >= 0 after this point. */ 407 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 408 409 /* 410 * Automatic sizing of send socket buffer. Often the send buffer 411 * size is not optimally adjusted to the actual network conditions 412 * at hand (delay bandwidth product). Setting the buffer size too 413 * small limits throughput on links with high bandwidth and high 414 * delay (eg. trans-continental/oceanic links). Setting the 415 * buffer size too big consumes too much real kernel memory, 416 * especially with many connections on busy servers. 417 * 418 * The criteria to step up the send buffer one notch are: 419 * 1. receive window of remote host is larger than send buffer 420 * (with a fudge factor of 5/4th); 421 * 2. send buffer is filled to 7/8th with data (so we actually 422 * have data to make use of it); 423 * 3. send buffer fill has not hit maximal automatic size; 424 * 4. our send window (slow start and cogestion controlled) is 425 * larger than sent but unacknowledged data in send buffer. 426 * 427 * The remote host receive window scaling factor may limit the 428 * growing of the send buffer before it reaches its allowed 429 * maximum. 430 * 431 * It scales directly with slow start or congestion window 432 * and does at most one step per received ACK. This fast 433 * scaling has the drawback of growing the send buffer beyond 434 * what is strictly necessary to make full use of a given 435 * delay*bandwith product. However testing has shown this not 436 * to be much of an problem. At worst we are trading wasting 437 * of available bandwith (the non-use of it) for wasting some 438 * socket buffer memory. 439 * 440 * TODO: Shrink send buffer during idle periods together 441 * with congestion window. Requires another timer. Has to 442 * wait for upcoming tcp timer rewrite. 443 */ 444 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 445 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 446 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && 447 so->so_snd.sb_cc < V_tcp_autosndbuf_max && 448 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) { 449 if (!sbreserve_locked(&so->so_snd, 450 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 451 V_tcp_autosndbuf_max), so, curthread)) 452 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 453 } 454 } 455 456 /* 457 * Truncate to the maximum segment length or enable TCP Segmentation 458 * Offloading (if supported by hardware) and ensure that FIN is removed 459 * if the length no longer contains the last data byte. 460 * 461 * TSO may only be used if we are in a pure bulk sending state. The 462 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 463 * IP options prevent using TSO. With TSO the TCP header is the same 464 * (except for the sequence number) for all generated packets. This 465 * makes it impossible to transmit any options which vary per generated 466 * segment or packet. 467 * 468 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and 469 * removal of FIN (if not already catched here) are handled later after 470 * the exact length of the TCP options are known. 471 */ 472 #ifdef IPSEC 473 /* 474 * Pre-calculate here as we save another lookup into the darknesses 475 * of IPsec that way and can actually decide if TSO is ok. 476 */ 477 ipsec_optlen = ipsec_hdrsiz_tcp(tp); 478 #endif 479 if (len > tp->t_maxseg) { 480 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && 481 ((tp->t_flags & TF_SIGNATURE) == 0) && 482 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 483 tp->t_inpcb->inp_options == NULL && 484 tp->t_inpcb->in6p_options == NULL 485 #ifdef IPSEC 486 && ipsec_optlen == 0 487 #endif 488 ) { 489 tso = 1; 490 } else { 491 len = tp->t_maxseg; 492 sendalot = 1; 493 tso = 0; 494 } 495 } 496 if (sack_rxmit) { 497 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) 498 flags &= ~TH_FIN; 499 } else { 500 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc)) 501 flags &= ~TH_FIN; 502 } 503 504 recwin = sbspace(&so->so_rcv); 505 506 /* 507 * Sender silly window avoidance. We transmit under the following 508 * conditions when len is non-zero: 509 * 510 * - We have a full segment (or more with TSO) 511 * - This is the last buffer in a write()/send() and we are 512 * either idle or running NODELAY 513 * - we've timed out (e.g. persist timer) 514 * - we have more then 1/2 the maximum send window's worth of 515 * data (receiver may be limited the window size) 516 * - we need to retransmit 517 */ 518 if (len) { 519 if (len >= tp->t_maxseg) 520 goto send; 521 /* 522 * NOTE! on localhost connections an 'ack' from the remote 523 * end may occur synchronously with the output and cause 524 * us to flush a buffer queued with moretocome. XXX 525 * 526 * note: the len + off check is almost certainly unnecessary. 527 */ 528 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 529 (idle || (tp->t_flags & TF_NODELAY)) && 530 len + off >= so->so_snd.sb_cc && 531 (tp->t_flags & TF_NOPUSH) == 0) { 532 goto send; 533 } 534 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 535 goto send; 536 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 537 goto send; 538 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 539 goto send; 540 if (sack_rxmit) 541 goto send; 542 } 543 544 /* 545 * Compare available window to amount of window 546 * known to peer (as advertised window less 547 * next expected input). If the difference is at least two 548 * max size segments, or at least 50% of the maximum possible 549 * window, then want to send a window update to peer. 550 * Skip this if the connection is in T/TCP half-open state. 551 * Don't send pure window updates when the peer has closed 552 * the connection and won't ever send more data. 553 */ 554 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 555 !TCPS_HAVERCVDFIN(tp->t_state)) { 556 /* 557 * "adv" is the amount we can increase the window, 558 * taking into account that we are limited by 559 * TCP_MAXWIN << tp->rcv_scale. 560 */ 561 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) - 562 (tp->rcv_adv - tp->rcv_nxt); 563 564 if (adv >= (long) (2 * tp->t_maxseg)) 565 goto send; 566 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 567 goto send; 568 } 569 570 /* 571 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 572 * is also a catch-all for the retransmit timer timeout case. 573 */ 574 if (tp->t_flags & TF_ACKNOW) 575 goto send; 576 if ((flags & TH_RST) || 577 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 578 goto send; 579 if (SEQ_GT(tp->snd_up, tp->snd_una)) 580 goto send; 581 /* 582 * If our state indicates that FIN should be sent 583 * and we have not yet done so, then we need to send. 584 */ 585 if (flags & TH_FIN && 586 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 587 goto send; 588 /* 589 * In SACK, it is possible for tcp_output to fail to send a segment 590 * after the retransmission timer has been turned off. Make sure 591 * that the retransmission timer is set. 592 */ 593 if ((tp->t_flags & TF_SACK_PERMIT) && 594 SEQ_GT(tp->snd_max, tp->snd_una) && 595 !tcp_timer_active(tp, TT_REXMT) && 596 !tcp_timer_active(tp, TT_PERSIST)) { 597 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 598 goto just_return; 599 } 600 /* 601 * TCP window updates are not reliable, rather a polling protocol 602 * using ``persist'' packets is used to insure receipt of window 603 * updates. The three ``states'' for the output side are: 604 * idle not doing retransmits or persists 605 * persisting to move a small or zero window 606 * (re)transmitting and thereby not persisting 607 * 608 * tcp_timer_active(tp, TT_PERSIST) 609 * is true when we are in persist state. 610 * (tp->t_flags & TF_FORCEDATA) 611 * is set when we are called to send a persist packet. 612 * tcp_timer_active(tp, TT_REXMT) 613 * is set when we are retransmitting 614 * The output side is idle when both timers are zero. 615 * 616 * If send window is too small, there is data to transmit, and no 617 * retransmit or persist is pending, then go to persist state. 618 * If nothing happens soon, send when timer expires: 619 * if window is nonzero, transmit what we can, 620 * otherwise force out a byte. 621 */ 622 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) && 623 !tcp_timer_active(tp, TT_PERSIST)) { 624 tp->t_rxtshift = 0; 625 tcp_setpersist(tp); 626 } 627 628 /* 629 * No reason to send a segment, just return. 630 */ 631 just_return: 632 SOCKBUF_UNLOCK(&so->so_snd); 633 return (0); 634 635 send: 636 SOCKBUF_LOCK_ASSERT(&so->so_snd); 637 /* 638 * Before ESTABLISHED, force sending of initial options 639 * unless TCP set not to do any options. 640 * NOTE: we assume that the IP/TCP header plus TCP options 641 * always fit in a single mbuf, leaving room for a maximum 642 * link header, i.e. 643 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 644 */ 645 optlen = 0; 646 #ifdef INET6 647 if (isipv6) 648 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 649 else 650 #endif 651 hdrlen = sizeof (struct tcpiphdr); 652 653 /* 654 * Compute options for segment. 655 * We only have to care about SYN and established connection 656 * segments. Options for SYN-ACK segments are handled in TCP 657 * syncache. 658 */ 659 if ((tp->t_flags & TF_NOOPT) == 0) { 660 to.to_flags = 0; 661 /* Maximum segment size. */ 662 if (flags & TH_SYN) { 663 tp->snd_nxt = tp->iss; 664 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 665 to.to_flags |= TOF_MSS; 666 } 667 /* Window scaling. */ 668 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 669 to.to_wscale = tp->request_r_scale; 670 to.to_flags |= TOF_SCALE; 671 } 672 /* Timestamps. */ 673 if ((tp->t_flags & TF_RCVD_TSTMP) || 674 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 675 to.to_tsval = ticks + tp->ts_offset; 676 to.to_tsecr = tp->ts_recent; 677 to.to_flags |= TOF_TS; 678 /* Set receive buffer autosizing timestamp. */ 679 if (tp->rfbuf_ts == 0 && 680 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 681 tp->rfbuf_ts = ticks; 682 } 683 /* Selective ACK's. */ 684 if (tp->t_flags & TF_SACK_PERMIT) { 685 if (flags & TH_SYN) 686 to.to_flags |= TOF_SACKPERM; 687 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 688 (tp->t_flags & TF_SACK_PERMIT) && 689 tp->rcv_numsacks > 0) { 690 to.to_flags |= TOF_SACK; 691 to.to_nsacks = tp->rcv_numsacks; 692 to.to_sacks = (u_char *)tp->sackblks; 693 } 694 } 695 #ifdef TCP_SIGNATURE 696 /* TCP-MD5 (RFC2385). */ 697 if (tp->t_flags & TF_SIGNATURE) 698 to.to_flags |= TOF_SIGNATURE; 699 #endif /* TCP_SIGNATURE */ 700 701 /* Processing the options. */ 702 hdrlen += optlen = tcp_addoptions(&to, opt); 703 } 704 705 #ifdef INET6 706 if (isipv6) 707 ipoptlen = ip6_optlen(tp->t_inpcb); 708 else 709 #endif 710 if (tp->t_inpcb->inp_options) 711 ipoptlen = tp->t_inpcb->inp_options->m_len - 712 offsetof(struct ipoption, ipopt_list); 713 else 714 ipoptlen = 0; 715 #ifdef IPSEC 716 ipoptlen += ipsec_optlen; 717 #endif 718 719 /* 720 * Adjust data length if insertion of options will 721 * bump the packet length beyond the t_maxopd length. 722 * Clear the FIN bit because we cut off the tail of 723 * the segment. 724 * 725 * When doing TSO limit a burst to TCP_MAXWIN minus the 726 * IP, TCP and Options length to keep ip->ip_len from 727 * overflowing. Prevent the last segment from being 728 * fractional thus making them all equal sized and set 729 * the flag to continue sending. TSO is disabled when 730 * IP options or IPSEC are present. 731 */ 732 if (len + optlen + ipoptlen > tp->t_maxopd) { 733 flags &= ~TH_FIN; 734 if (tso) { 735 if (len > TCP_MAXWIN - hdrlen - optlen) { 736 len = TCP_MAXWIN - hdrlen - optlen; 737 len = len - (len % (tp->t_maxopd - optlen)); 738 sendalot = 1; 739 } else if (tp->t_flags & TF_NEEDFIN) 740 sendalot = 1; 741 } else { 742 len = tp->t_maxopd - optlen - ipoptlen; 743 sendalot = 1; 744 } 745 } 746 747 /*#ifdef DIAGNOSTIC*/ 748 #ifdef INET6 749 if (max_linkhdr + hdrlen > MCLBYTES) 750 #else 751 if (max_linkhdr + hdrlen > MHLEN) 752 #endif 753 panic("tcphdr too big"); 754 /*#endif*/ 755 756 /* 757 * This KASSERT is here to catch edge cases at a well defined place. 758 * Before, those had triggered (random) panic conditions further down. 759 */ 760 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 761 762 /* 763 * Grab a header mbuf, attaching a copy of data to 764 * be transmitted, and initialize the header from 765 * the template for sends on this connection. 766 */ 767 if (len) { 768 struct mbuf *mb; 769 u_int moff; 770 771 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 772 TCPSTAT_INC(tcps_sndprobe); 773 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 774 TCPSTAT_INC(tcps_sndrexmitpack); 775 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 776 } else { 777 TCPSTAT_INC(tcps_sndpack); 778 TCPSTAT_ADD(tcps_sndbyte, len); 779 } 780 #ifdef notyet 781 if ((m = m_copypack(so->so_snd.sb_mb, off, 782 (int)len, max_linkhdr + hdrlen)) == 0) { 783 SOCKBUF_UNLOCK(&so->so_snd); 784 error = ENOBUFS; 785 goto out; 786 } 787 /* 788 * m_copypack left space for our hdr; use it. 789 */ 790 m->m_len += hdrlen; 791 m->m_data -= hdrlen; 792 #else 793 MGETHDR(m, M_DONTWAIT, MT_DATA); 794 if (m == NULL) { 795 SOCKBUF_UNLOCK(&so->so_snd); 796 error = ENOBUFS; 797 goto out; 798 } 799 #ifdef INET6 800 if (MHLEN < hdrlen + max_linkhdr) { 801 MCLGET(m, M_DONTWAIT); 802 if ((m->m_flags & M_EXT) == 0) { 803 SOCKBUF_UNLOCK(&so->so_snd); 804 m_freem(m); 805 error = ENOBUFS; 806 goto out; 807 } 808 } 809 #endif 810 m->m_data += max_linkhdr; 811 m->m_len = hdrlen; 812 813 /* 814 * Start the m_copy functions from the closest mbuf 815 * to the offset in the socket buffer chain. 816 */ 817 mb = sbsndptr(&so->so_snd, off, len, &moff); 818 819 if (len <= MHLEN - hdrlen - max_linkhdr) { 820 m_copydata(mb, moff, (int)len, 821 mtod(m, caddr_t) + hdrlen); 822 m->m_len += len; 823 } else { 824 m->m_next = m_copy(mb, moff, (int)len); 825 if (m->m_next == NULL) { 826 SOCKBUF_UNLOCK(&so->so_snd); 827 (void) m_free(m); 828 error = ENOBUFS; 829 goto out; 830 } 831 } 832 #endif 833 /* 834 * If we're sending everything we've got, set PUSH. 835 * (This will keep happy those implementations which only 836 * give data to the user when a buffer fills or 837 * a PUSH comes in.) 838 */ 839 if (off + len == so->so_snd.sb_cc) 840 flags |= TH_PUSH; 841 SOCKBUF_UNLOCK(&so->so_snd); 842 } else { 843 SOCKBUF_UNLOCK(&so->so_snd); 844 if (tp->t_flags & TF_ACKNOW) 845 TCPSTAT_INC(tcps_sndacks); 846 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 847 TCPSTAT_INC(tcps_sndctrl); 848 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 849 TCPSTAT_INC(tcps_sndurg); 850 else 851 TCPSTAT_INC(tcps_sndwinup); 852 853 MGETHDR(m, M_DONTWAIT, MT_DATA); 854 if (m == NULL) { 855 error = ENOBUFS; 856 goto out; 857 } 858 #ifdef INET6 859 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 860 MHLEN >= hdrlen) { 861 MH_ALIGN(m, hdrlen); 862 } else 863 #endif 864 m->m_data += max_linkhdr; 865 m->m_len = hdrlen; 866 } 867 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 868 m->m_pkthdr.rcvif = (struct ifnet *)0; 869 #ifdef MAC 870 mac_inpcb_create_mbuf(tp->t_inpcb, m); 871 #endif 872 #ifdef INET6 873 if (isipv6) { 874 ip6 = mtod(m, struct ip6_hdr *); 875 th = (struct tcphdr *)(ip6 + 1); 876 tcpip_fillheaders(tp->t_inpcb, ip6, th); 877 } else 878 #endif /* INET6 */ 879 { 880 ip = mtod(m, struct ip *); 881 ipov = (struct ipovly *)ip; 882 th = (struct tcphdr *)(ip + 1); 883 tcpip_fillheaders(tp->t_inpcb, ip, th); 884 } 885 886 /* 887 * Fill in fields, remembering maximum advertised 888 * window for use in delaying messages about window sizes. 889 * If resending a FIN, be sure not to use a new sequence number. 890 */ 891 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 892 tp->snd_nxt == tp->snd_max) 893 tp->snd_nxt--; 894 /* 895 * If we are starting a connection, send ECN setup 896 * SYN packet. If we are on a retransmit, we may 897 * resend those bits a number of times as per 898 * RFC 3168. 899 */ 900 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 901 if (tp->t_rxtshift >= 1) { 902 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 903 flags |= TH_ECE|TH_CWR; 904 } else 905 flags |= TH_ECE|TH_CWR; 906 } 907 908 if (tp->t_state == TCPS_ESTABLISHED && 909 (tp->t_flags & TF_ECN_PERMIT)) { 910 /* 911 * If the peer has ECN, mark data packets with 912 * ECN capable transmission (ECT). 913 * Ignore pure ack packets, retransmissions and window probes. 914 */ 915 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 916 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 917 #ifdef INET6 918 if (isipv6) 919 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 920 else 921 #endif 922 ip->ip_tos |= IPTOS_ECN_ECT0; 923 TCPSTAT_INC(tcps_ecn_ect0); 924 } 925 926 /* 927 * Reply with proper ECN notifications. 928 */ 929 if (tp->t_flags & TF_ECN_SND_CWR) { 930 flags |= TH_CWR; 931 tp->t_flags &= ~TF_ECN_SND_CWR; 932 } 933 if (tp->t_flags & TF_ECN_SND_ECE) 934 flags |= TH_ECE; 935 } 936 937 /* 938 * If we are doing retransmissions, then snd_nxt will 939 * not reflect the first unsent octet. For ACK only 940 * packets, we do not want the sequence number of the 941 * retransmitted packet, we want the sequence number 942 * of the next unsent octet. So, if there is no data 943 * (and no SYN or FIN), use snd_max instead of snd_nxt 944 * when filling in ti_seq. But if we are in persist 945 * state, snd_max might reflect one byte beyond the 946 * right edge of the window, so use snd_nxt in that 947 * case, since we know we aren't doing a retransmission. 948 * (retransmit and persist are mutually exclusive...) 949 */ 950 if (sack_rxmit == 0) { 951 if (len || (flags & (TH_SYN|TH_FIN)) || 952 tcp_timer_active(tp, TT_PERSIST)) 953 th->th_seq = htonl(tp->snd_nxt); 954 else 955 th->th_seq = htonl(tp->snd_max); 956 } else { 957 th->th_seq = htonl(p->rxmit); 958 p->rxmit += len; 959 tp->sackhint.sack_bytes_rexmit += len; 960 } 961 th->th_ack = htonl(tp->rcv_nxt); 962 if (optlen) { 963 bcopy(opt, th + 1, optlen); 964 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 965 } 966 th->th_flags = flags; 967 /* 968 * Calculate receive window. Don't shrink window, 969 * but avoid silly window syndrome. 970 */ 971 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 972 recwin < (long)tp->t_maxseg) 973 recwin = 0; 974 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 975 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 976 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 977 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 978 979 /* 980 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 981 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 982 * case is handled in syncache. 983 */ 984 if (flags & TH_SYN) 985 th->th_win = htons((u_short) 986 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 987 else 988 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 989 990 /* 991 * Adjust the RXWIN0SENT flag - indicate that we have advertised 992 * a 0 window. This may cause the remote transmitter to stall. This 993 * flag tells soreceive() to disable delayed acknowledgements when 994 * draining the buffer. This can occur if the receiver is attempting 995 * to read more data than can be buffered prior to transmitting on 996 * the connection. 997 */ 998 if (th->th_win == 0) 999 tp->t_flags |= TF_RXWIN0SENT; 1000 else 1001 tp->t_flags &= ~TF_RXWIN0SENT; 1002 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1003 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1004 th->th_flags |= TH_URG; 1005 } else 1006 /* 1007 * If no urgent pointer to send, then we pull 1008 * the urgent pointer to the left edge of the send window 1009 * so that it doesn't drift into the send window on sequence 1010 * number wraparound. 1011 */ 1012 tp->snd_up = tp->snd_una; /* drag it along */ 1013 1014 #ifdef TCP_SIGNATURE 1015 if (tp->t_flags & TF_SIGNATURE) { 1016 int sigoff = to.to_signature - opt; 1017 tcp_signature_compute(m, 0, len, optlen, 1018 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1019 } 1020 #endif 1021 1022 /* 1023 * Put TCP length in extended header, and then 1024 * checksum extended header and data. 1025 */ 1026 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1027 #ifdef INET6 1028 if (isipv6) 1029 /* 1030 * ip6_plen is not need to be filled now, and will be filled 1031 * in ip6_output. 1032 */ 1033 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), 1034 sizeof(struct tcphdr) + optlen + len); 1035 else 1036 #endif /* INET6 */ 1037 { 1038 m->m_pkthdr.csum_flags = CSUM_TCP; 1039 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1040 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1041 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1042 1043 /* IP version must be set here for ipv4/ipv6 checking later */ 1044 KASSERT(ip->ip_v == IPVERSION, 1045 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1046 } 1047 1048 /* 1049 * Enable TSO and specify the size of the segments. 1050 * The TCP pseudo header checksum is always provided. 1051 * XXX: Fixme: This is currently not the case for IPv6. 1052 */ 1053 if (tso) { 1054 m->m_pkthdr.csum_flags |= CSUM_TSO; 1055 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 1056 } 1057 1058 /* 1059 * In transmit state, time the transmission and arrange for 1060 * the retransmit. In persist state, just set snd_max. 1061 */ 1062 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1063 !tcp_timer_active(tp, TT_PERSIST)) { 1064 tcp_seq startseq = tp->snd_nxt; 1065 1066 /* 1067 * Advance snd_nxt over sequence space of this segment. 1068 */ 1069 if (flags & (TH_SYN|TH_FIN)) { 1070 if (flags & TH_SYN) 1071 tp->snd_nxt++; 1072 if (flags & TH_FIN) { 1073 tp->snd_nxt++; 1074 tp->t_flags |= TF_SENTFIN; 1075 } 1076 } 1077 if (sack_rxmit) 1078 goto timer; 1079 tp->snd_nxt += len; 1080 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1081 tp->snd_max = tp->snd_nxt; 1082 /* 1083 * Time this transmission if not a retransmission and 1084 * not currently timing anything. 1085 */ 1086 if (tp->t_rtttime == 0) { 1087 tp->t_rtttime = ticks; 1088 tp->t_rtseq = startseq; 1089 TCPSTAT_INC(tcps_segstimed); 1090 } 1091 } 1092 1093 /* 1094 * Set retransmit timer if not currently set, 1095 * and not doing a pure ack or a keep-alive probe. 1096 * Initial value for retransmit timer is smoothed 1097 * round-trip time + 2 * round-trip time variance. 1098 * Initialize shift counter which is used for backoff 1099 * of retransmit time. 1100 */ 1101 timer: 1102 if (!tcp_timer_active(tp, TT_REXMT) && 1103 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1104 (tp->snd_nxt != tp->snd_una))) { 1105 if (tcp_timer_active(tp, TT_PERSIST)) { 1106 tcp_timer_activate(tp, TT_PERSIST, 0); 1107 tp->t_rxtshift = 0; 1108 } 1109 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1110 } 1111 } else { 1112 /* 1113 * Persist case, update snd_max but since we are in 1114 * persist mode (no window) we do not update snd_nxt. 1115 */ 1116 int xlen = len; 1117 if (flags & TH_SYN) 1118 ++xlen; 1119 if (flags & TH_FIN) { 1120 ++xlen; 1121 tp->t_flags |= TF_SENTFIN; 1122 } 1123 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1124 tp->snd_max = tp->snd_nxt + len; 1125 } 1126 1127 #ifdef TCPDEBUG 1128 /* 1129 * Trace. 1130 */ 1131 if (so->so_options & SO_DEBUG) { 1132 u_short save = 0; 1133 #ifdef INET6 1134 if (!isipv6) 1135 #endif 1136 { 1137 save = ipov->ih_len; 1138 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1139 } 1140 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1141 #ifdef INET6 1142 if (!isipv6) 1143 #endif 1144 ipov->ih_len = save; 1145 } 1146 #endif 1147 1148 /* 1149 * Fill in IP length and desired time to live and 1150 * send to IP level. There should be a better way 1151 * to handle ttl and tos; we could keep them in 1152 * the template, but need a way to checksum without them. 1153 */ 1154 /* 1155 * m->m_pkthdr.len should have been set before cksum calcuration, 1156 * because in6_cksum() need it. 1157 */ 1158 #ifdef INET6 1159 if (isipv6) { 1160 /* 1161 * we separately set hoplimit for every segment, since the 1162 * user might want to change the value via setsockopt. 1163 * Also, desired default hop limit might be changed via 1164 * Neighbor Discovery. 1165 */ 1166 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1167 1168 /* TODO: IPv6 IP6TOS_ECT bit on */ 1169 error = ip6_output(m, 1170 tp->t_inpcb->in6p_outputopts, NULL, 1171 ((so->so_options & SO_DONTROUTE) ? 1172 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb); 1173 } else 1174 #endif /* INET6 */ 1175 { 1176 ip->ip_len = m->m_pkthdr.len; 1177 #ifdef INET6 1178 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1179 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1180 #endif /* INET6 */ 1181 /* 1182 * If we do path MTU discovery, then we set DF on every packet. 1183 * This might not be the best thing to do according to RFC3390 1184 * Section 2. However the tcp hostcache migitates the problem 1185 * so it affects only the first tcp connection with a host. 1186 */ 1187 if (V_path_mtu_discovery) 1188 ip->ip_off |= IP_DF; 1189 1190 error = ip_output(m, tp->t_inpcb->inp_options, NULL, 1191 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1192 tp->t_inpcb); 1193 } 1194 if (error) { 1195 1196 /* 1197 * We know that the packet was lost, so back out the 1198 * sequence number advance, if any. 1199 * 1200 * If the error is EPERM the packet got blocked by the 1201 * local firewall. Normally we should terminate the 1202 * connection but the blocking may have been spurious 1203 * due to a firewall reconfiguration cycle. So we treat 1204 * it like a packet loss and let the retransmit timer and 1205 * timeouts do their work over time. 1206 * XXX: It is a POLA question whether calling tcp_drop right 1207 * away would be the really correct behavior instead. 1208 */ 1209 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1210 !tcp_timer_active(tp, TT_PERSIST)) && 1211 ((flags & TH_SYN) == 0) && 1212 (error != EPERM)) { 1213 if (sack_rxmit) { 1214 p->rxmit -= len; 1215 tp->sackhint.sack_bytes_rexmit -= len; 1216 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1217 ("sackhint bytes rtx >= 0")); 1218 } else 1219 tp->snd_nxt -= len; 1220 } 1221 out: 1222 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1223 switch (error) { 1224 case EPERM: 1225 tp->t_softerror = error; 1226 return (error); 1227 case ENOBUFS: 1228 if (!tcp_timer_active(tp, TT_REXMT) && 1229 !tcp_timer_active(tp, TT_PERSIST)) 1230 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1231 tp->snd_cwnd = tp->t_maxseg; 1232 return (0); 1233 case EMSGSIZE: 1234 /* 1235 * For some reason the interface we used initially 1236 * to send segments changed to another or lowered 1237 * its MTU. 1238 * 1239 * tcp_mtudisc() will find out the new MTU and as 1240 * its last action, initiate retransmission, so it 1241 * is important to not do so here. 1242 * 1243 * If TSO was active we either got an interface 1244 * without TSO capabilits or TSO was turned off. 1245 * Disable it for this connection as too and 1246 * immediatly retry with MSS sized segments generated 1247 * by this function. 1248 */ 1249 if (tso) 1250 tp->t_flags &= ~TF_TSO; 1251 tcp_mtudisc(tp->t_inpcb, 0); 1252 return (0); 1253 case EHOSTDOWN: 1254 case EHOSTUNREACH: 1255 case ENETDOWN: 1256 case ENETUNREACH: 1257 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1258 tp->t_softerror = error; 1259 return (0); 1260 } 1261 /* FALLTHROUGH */ 1262 default: 1263 return (error); 1264 } 1265 } 1266 TCPSTAT_INC(tcps_sndtotal); 1267 1268 /* 1269 * Data sent (as far as we can tell). 1270 * If this advertises a larger window than any other segment, 1271 * then remember the size of the advertised window. 1272 * Any pending ACK has now been sent. 1273 */ 1274 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1275 tp->rcv_adv = tp->rcv_nxt + recwin; 1276 tp->last_ack_sent = tp->rcv_nxt; 1277 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1278 if (tcp_timer_active(tp, TT_DELACK)) 1279 tcp_timer_activate(tp, TT_DELACK, 0); 1280 #if 0 1281 /* 1282 * This completely breaks TCP if newreno is turned on. What happens 1283 * is that if delayed-acks are turned on on the receiver, this code 1284 * on the transmitter effectively destroys the TCP window, forcing 1285 * it to four packets (1.5Kx4 = 6K window). 1286 */ 1287 if (sendalot && (!V_tcp_do_newreno || --maxburst)) 1288 goto again; 1289 #endif 1290 if (sendalot) 1291 goto again; 1292 return (0); 1293 } 1294 1295 void 1296 tcp_setpersist(struct tcpcb *tp) 1297 { 1298 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1299 int tt; 1300 1301 if (tcp_timer_active(tp, TT_REXMT)) 1302 panic("tcp_setpersist: retransmit pending"); 1303 /* 1304 * Start/restart persistance timer. 1305 */ 1306 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1307 TCPTV_PERSMIN, TCPTV_PERSMAX); 1308 tcp_timer_activate(tp, TT_PERSIST, tt); 1309 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1310 tp->t_rxtshift++; 1311 } 1312 1313 /* 1314 * Insert TCP options according to the supplied parameters to the place 1315 * optp in a consistent way. Can handle unaligned destinations. 1316 * 1317 * The order of the option processing is crucial for optimal packing and 1318 * alignment for the scarce option space. 1319 * 1320 * The optimal order for a SYN/SYN-ACK segment is: 1321 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1322 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1323 * 1324 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1325 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1326 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1327 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1328 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1329 */ 1330 int 1331 tcp_addoptions(struct tcpopt *to, u_char *optp) 1332 { 1333 u_int mask, optlen = 0; 1334 1335 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1336 if ((to->to_flags & mask) != mask) 1337 continue; 1338 if (optlen == TCP_MAXOLEN) 1339 break; 1340 switch (to->to_flags & mask) { 1341 case TOF_MSS: 1342 while (optlen % 4) { 1343 optlen += TCPOLEN_NOP; 1344 *optp++ = TCPOPT_NOP; 1345 } 1346 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1347 continue; 1348 optlen += TCPOLEN_MAXSEG; 1349 *optp++ = TCPOPT_MAXSEG; 1350 *optp++ = TCPOLEN_MAXSEG; 1351 to->to_mss = htons(to->to_mss); 1352 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1353 optp += sizeof(to->to_mss); 1354 break; 1355 case TOF_SCALE: 1356 while (!optlen || optlen % 2 != 1) { 1357 optlen += TCPOLEN_NOP; 1358 *optp++ = TCPOPT_NOP; 1359 } 1360 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1361 continue; 1362 optlen += TCPOLEN_WINDOW; 1363 *optp++ = TCPOPT_WINDOW; 1364 *optp++ = TCPOLEN_WINDOW; 1365 *optp++ = to->to_wscale; 1366 break; 1367 case TOF_SACKPERM: 1368 while (optlen % 2) { 1369 optlen += TCPOLEN_NOP; 1370 *optp++ = TCPOPT_NOP; 1371 } 1372 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1373 continue; 1374 optlen += TCPOLEN_SACK_PERMITTED; 1375 *optp++ = TCPOPT_SACK_PERMITTED; 1376 *optp++ = TCPOLEN_SACK_PERMITTED; 1377 break; 1378 case TOF_TS: 1379 while (!optlen || optlen % 4 != 2) { 1380 optlen += TCPOLEN_NOP; 1381 *optp++ = TCPOPT_NOP; 1382 } 1383 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1384 continue; 1385 optlen += TCPOLEN_TIMESTAMP; 1386 *optp++ = TCPOPT_TIMESTAMP; 1387 *optp++ = TCPOLEN_TIMESTAMP; 1388 to->to_tsval = htonl(to->to_tsval); 1389 to->to_tsecr = htonl(to->to_tsecr); 1390 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1391 optp += sizeof(to->to_tsval); 1392 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1393 optp += sizeof(to->to_tsecr); 1394 break; 1395 case TOF_SIGNATURE: 1396 { 1397 int siglen = TCPOLEN_SIGNATURE - 2; 1398 1399 while (!optlen || optlen % 4 != 2) { 1400 optlen += TCPOLEN_NOP; 1401 *optp++ = TCPOPT_NOP; 1402 } 1403 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1404 continue; 1405 optlen += TCPOLEN_SIGNATURE; 1406 *optp++ = TCPOPT_SIGNATURE; 1407 *optp++ = TCPOLEN_SIGNATURE; 1408 to->to_signature = optp; 1409 while (siglen--) 1410 *optp++ = 0; 1411 break; 1412 } 1413 case TOF_SACK: 1414 { 1415 int sackblks = 0; 1416 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1417 tcp_seq sack_seq; 1418 1419 while (!optlen || optlen % 4 != 2) { 1420 optlen += TCPOLEN_NOP; 1421 *optp++ = TCPOPT_NOP; 1422 } 1423 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1424 continue; 1425 optlen += TCPOLEN_SACKHDR; 1426 *optp++ = TCPOPT_SACK; 1427 sackblks = min(to->to_nsacks, 1428 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1429 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1430 while (sackblks--) { 1431 sack_seq = htonl(sack->start); 1432 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1433 optp += sizeof(sack_seq); 1434 sack_seq = htonl(sack->end); 1435 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1436 optp += sizeof(sack_seq); 1437 optlen += TCPOLEN_SACK; 1438 sack++; 1439 } 1440 TCPSTAT_INC(tcps_sack_send_blocks); 1441 break; 1442 } 1443 default: 1444 panic("%s: unknown TCP option type", __func__); 1445 break; 1446 } 1447 } 1448 1449 /* Terminate and pad TCP options to a 4 byte boundary. */ 1450 if (optlen % 4) { 1451 optlen += TCPOLEN_EOL; 1452 *optp++ = TCPOPT_EOL; 1453 } 1454 /* 1455 * According to RFC 793 (STD0007): 1456 * "The content of the header beyond the End-of-Option option 1457 * must be header padding (i.e., zero)." 1458 * and later: "The padding is composed of zeros." 1459 */ 1460 while (optlen % 4) { 1461 optlen += TCPOLEN_PAD; 1462 *optp++ = TCPOPT_PAD; 1463 } 1464 1465 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1466 return (optlen); 1467 } 1468