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 TCPPCAP 78 #include <netinet/tcp_pcap.h> 79 #endif 80 #ifdef TCPDEBUG 81 #include <netinet/tcp_debug.h> 82 #endif 83 #ifdef TCP_OFFLOAD 84 #include <netinet/tcp_offload.h> 85 #endif 86 87 #ifdef IPSEC 88 #include <netipsec/ipsec.h> 89 #endif /*IPSEC*/ 90 91 #include <machine/in_cksum.h> 92 93 #include <security/mac/mac_framework.h> 94 95 VNET_DEFINE(int, path_mtu_discovery) = 1; 96 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW, 97 &VNET_NAME(path_mtu_discovery), 1, 98 "Enable Path MTU Discovery"); 99 100 VNET_DEFINE(int, tcp_do_tso) = 1; 101 #define V_tcp_do_tso VNET(tcp_do_tso) 102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW, 103 &VNET_NAME(tcp_do_tso), 0, 104 "Enable TCP Segmentation Offload"); 105 106 VNET_DEFINE(int, tcp_sendspace) = 1024*32; 107 #define V_tcp_sendspace VNET(tcp_sendspace) 108 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW, 109 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); 110 111 VNET_DEFINE(int, tcp_do_autosndbuf) = 1; 112 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) 113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, 114 &VNET_NAME(tcp_do_autosndbuf), 0, 115 "Enable automatic send buffer sizing"); 116 117 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; 118 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) 119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW, 120 &VNET_NAME(tcp_autosndbuf_inc), 0, 121 "Incrementor step size of automatic send buffer"); 122 123 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; 124 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) 125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW, 126 &VNET_NAME(tcp_autosndbuf_max), 0, 127 "Max size of automatic send buffer"); 128 129 static void inline hhook_run_tcp_est_out(struct tcpcb *tp, 130 struct tcphdr *th, struct tcpopt *to, 131 long len, int tso); 132 static void inline cc_after_idle(struct tcpcb *tp); 133 134 /* 135 * Wrapper for the TCP established output helper hook. 136 */ 137 static void inline 138 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, 139 struct tcpopt *to, long len, int tso) 140 { 141 struct tcp_hhook_data hhook_data; 142 143 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { 144 hhook_data.tp = tp; 145 hhook_data.th = th; 146 hhook_data.to = to; 147 hhook_data.len = len; 148 hhook_data.tso = tso; 149 150 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, 151 tp->osd); 152 } 153 } 154 155 /* 156 * CC wrapper hook functions 157 */ 158 static void inline 159 cc_after_idle(struct tcpcb *tp) 160 { 161 INP_WLOCK_ASSERT(tp->t_inpcb); 162 163 if (CC_ALGO(tp)->after_idle != NULL) 164 CC_ALGO(tp)->after_idle(tp->ccv); 165 } 166 167 /* 168 * Tcp output routine: figure out what should be sent and send it. 169 */ 170 int 171 tcp_output(struct tcpcb *tp) 172 { 173 struct socket *so = tp->t_inpcb->inp_socket; 174 long len, recwin, sendwin; 175 int off, flags, error = 0; /* Keep compiler happy */ 176 struct mbuf *m; 177 struct ip *ip = NULL; 178 struct ipovly *ipov = NULL; 179 struct tcphdr *th; 180 u_char opt[TCP_MAXOLEN]; 181 unsigned ipoptlen, optlen, hdrlen; 182 #ifdef IPSEC 183 unsigned ipsec_optlen = 0; 184 #endif 185 int idle, sendalot; 186 int sack_rxmit, sack_bytes_rxmt; 187 struct sackhole *p; 188 int tso, mtu; 189 struct tcpopt to; 190 #if 0 191 int maxburst = TCP_MAXBURST; 192 #endif 193 #ifdef INET6 194 struct ip6_hdr *ip6 = NULL; 195 int isipv6; 196 197 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 198 #endif 199 200 INP_WLOCK_ASSERT(tp->t_inpcb); 201 202 #ifdef TCP_OFFLOAD 203 if (tp->t_flags & TF_TOE) 204 return (tcp_offload_output(tp)); 205 #endif 206 207 /* 208 * Determine length of data that should be transmitted, 209 * and flags that will be used. 210 * If there is some data or critical controls (SYN, RST) 211 * to send, then transmit; otherwise, investigate further. 212 */ 213 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 214 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) 215 cc_after_idle(tp); 216 tp->t_flags &= ~TF_LASTIDLE; 217 if (idle) { 218 if (tp->t_flags & TF_MORETOCOME) { 219 tp->t_flags |= TF_LASTIDLE; 220 idle = 0; 221 } 222 } 223 again: 224 /* 225 * If we've recently taken a timeout, snd_max will be greater than 226 * snd_nxt. There may be SACK information that allows us to avoid 227 * resending already delivered data. Adjust snd_nxt accordingly. 228 */ 229 if ((tp->t_flags & TF_SACK_PERMIT) && 230 SEQ_LT(tp->snd_nxt, tp->snd_max)) 231 tcp_sack_adjust(tp); 232 sendalot = 0; 233 tso = 0; 234 mtu = 0; 235 off = tp->snd_nxt - tp->snd_una; 236 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 237 238 flags = tcp_outflags[tp->t_state]; 239 /* 240 * Send any SACK-generated retransmissions. If we're explicitly trying 241 * to send out new data (when sendalot is 1), bypass this function. 242 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 243 * we're replacing a (future) new transmission with a retransmission 244 * now, and we previously incremented snd_cwnd in tcp_input(). 245 */ 246 /* 247 * Still in sack recovery , reset rxmit flag to zero. 248 */ 249 sack_rxmit = 0; 250 sack_bytes_rxmt = 0; 251 len = 0; 252 p = NULL; 253 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && 254 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { 255 long cwin; 256 257 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt; 258 if (cwin < 0) 259 cwin = 0; 260 /* Do not retransmit SACK segments beyond snd_recover */ 261 if (SEQ_GT(p->end, tp->snd_recover)) { 262 /* 263 * (At least) part of sack hole extends beyond 264 * snd_recover. Check to see if we can rexmit data 265 * for this hole. 266 */ 267 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { 268 /* 269 * Can't rexmit any more data for this hole. 270 * That data will be rexmitted in the next 271 * sack recovery episode, when snd_recover 272 * moves past p->rxmit. 273 */ 274 p = NULL; 275 goto after_sack_rexmit; 276 } else 277 /* Can rexmit part of the current hole */ 278 len = ((long)ulmin(cwin, 279 tp->snd_recover - p->rxmit)); 280 } else 281 len = ((long)ulmin(cwin, p->end - p->rxmit)); 282 off = p->rxmit - tp->snd_una; 283 KASSERT(off >= 0,("%s: sack block to the left of una : %d", 284 __func__, off)); 285 if (len > 0) { 286 sack_rxmit = 1; 287 sendalot = 1; 288 TCPSTAT_INC(tcps_sack_rexmits); 289 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 290 min(len, tp->t_maxseg)); 291 } 292 } 293 after_sack_rexmit: 294 /* 295 * Get standard flags, and add SYN or FIN if requested by 'hidden' 296 * state flags. 297 */ 298 if (tp->t_flags & TF_NEEDFIN) 299 flags |= TH_FIN; 300 if (tp->t_flags & TF_NEEDSYN) 301 flags |= TH_SYN; 302 303 SOCKBUF_LOCK(&so->so_snd); 304 /* 305 * If in persist timeout with window of 0, send 1 byte. 306 * Otherwise, if window is small but nonzero 307 * and timer expired, we will send what we can 308 * and go to transmit state. 309 */ 310 if (tp->t_flags & TF_FORCEDATA) { 311 if (sendwin == 0) { 312 /* 313 * If we still have some data to send, then 314 * clear the FIN bit. Usually this would 315 * happen below when it realizes that we 316 * aren't sending all the data. However, 317 * if we have exactly 1 byte of unsent data, 318 * then it won't clear the FIN bit below, 319 * and if we are in persist state, we wind 320 * up sending the packet without recording 321 * that we sent the FIN bit. 322 * 323 * We can't just blindly clear the FIN bit, 324 * because if we don't have any more data 325 * to send then the probe will be the FIN 326 * itself. 327 */ 328 if (off < sbused(&so->so_snd)) 329 flags &= ~TH_FIN; 330 sendwin = 1; 331 } else { 332 tcp_timer_activate(tp, TT_PERSIST, 0); 333 tp->t_rxtshift = 0; 334 } 335 } 336 337 /* 338 * If snd_nxt == snd_max and we have transmitted a FIN, the 339 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in 340 * a negative length. This can also occur when TCP opens up 341 * its congestion window while receiving additional duplicate 342 * acks after fast-retransmit because TCP will reset snd_nxt 343 * to snd_max after the fast-retransmit. 344 * 345 * In the normal retransmit-FIN-only case, however, snd_nxt will 346 * be set to snd_una, the offset will be 0, and the length may 347 * wind up 0. 348 * 349 * If sack_rxmit is true we are retransmitting from the scoreboard 350 * in which case len is already set. 351 */ 352 if (sack_rxmit == 0) { 353 if (sack_bytes_rxmt == 0) 354 len = ((long)ulmin(sbavail(&so->so_snd), sendwin) - 355 off); 356 else { 357 long cwin; 358 359 /* 360 * We are inside of a SACK recovery episode and are 361 * sending new data, having retransmitted all the 362 * data possible in the scoreboard. 363 */ 364 len = ((long)ulmin(sbavail(&so->so_snd), tp->snd_wnd) - 365 off); 366 /* 367 * Don't remove this (len > 0) check ! 368 * We explicitly check for len > 0 here (although it 369 * isn't really necessary), to work around a gcc 370 * optimization issue - to force gcc to compute 371 * len above. Without this check, the computation 372 * of len is bungled by the optimizer. 373 */ 374 if (len > 0) { 375 cwin = tp->snd_cwnd - 376 (tp->snd_nxt - tp->sack_newdata) - 377 sack_bytes_rxmt; 378 if (cwin < 0) 379 cwin = 0; 380 len = lmin(len, cwin); 381 } 382 } 383 } 384 385 /* 386 * Lop off SYN bit if it has already been sent. However, if this 387 * is SYN-SENT state and if segment contains data and if we don't 388 * know that foreign host supports TAO, suppress sending segment. 389 */ 390 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 391 if (tp->t_state != TCPS_SYN_RECEIVED) 392 flags &= ~TH_SYN; 393 off--, len++; 394 } 395 396 /* 397 * Be careful not to send data and/or FIN on SYN segments. 398 * This measure is needed to prevent interoperability problems 399 * with not fully conformant TCP implementations. 400 */ 401 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 402 len = 0; 403 flags &= ~TH_FIN; 404 } 405 406 if (len <= 0) { 407 /* 408 * If FIN has been sent but not acked, 409 * but we haven't been called to retransmit, 410 * len will be < 0. Otherwise, window shrank 411 * after we sent into it. If window shrank to 0, 412 * cancel pending retransmit, pull snd_nxt back 413 * to (closed) window, and set the persist timer 414 * if it isn't already going. If the window didn't 415 * close completely, just wait for an ACK. 416 * 417 * We also do a general check here to ensure that 418 * we will set the persist timer when we have data 419 * to send, but a 0-byte window. This makes sure 420 * the persist timer is set even if the packet 421 * hits one of the "goto send" lines below. 422 */ 423 len = 0; 424 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && 425 (off < (int) sbavail(&so->so_snd))) { 426 tcp_timer_activate(tp, TT_REXMT, 0); 427 tp->t_rxtshift = 0; 428 tp->snd_nxt = tp->snd_una; 429 if (!tcp_timer_active(tp, TT_PERSIST)) 430 tcp_setpersist(tp); 431 } 432 } 433 434 /* len will be >= 0 after this point. */ 435 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 436 437 /* 438 * Automatic sizing of send socket buffer. Often the send buffer 439 * size is not optimally adjusted to the actual network conditions 440 * at hand (delay bandwidth product). Setting the buffer size too 441 * small limits throughput on links with high bandwidth and high 442 * delay (eg. trans-continental/oceanic links). Setting the 443 * buffer size too big consumes too much real kernel memory, 444 * especially with many connections on busy servers. 445 * 446 * The criteria to step up the send buffer one notch are: 447 * 1. receive window of remote host is larger than send buffer 448 * (with a fudge factor of 5/4th); 449 * 2. send buffer is filled to 7/8th with data (so we actually 450 * have data to make use of it); 451 * 3. send buffer fill has not hit maximal automatic size; 452 * 4. our send window (slow start and cogestion controlled) is 453 * larger than sent but unacknowledged data in send buffer. 454 * 455 * The remote host receive window scaling factor may limit the 456 * growing of the send buffer before it reaches its allowed 457 * maximum. 458 * 459 * It scales directly with slow start or congestion window 460 * and does at most one step per received ACK. This fast 461 * scaling has the drawback of growing the send buffer beyond 462 * what is strictly necessary to make full use of a given 463 * delay*bandwith product. However testing has shown this not 464 * to be much of an problem. At worst we are trading wasting 465 * of available bandwith (the non-use of it) for wasting some 466 * socket buffer memory. 467 * 468 * TODO: Shrink send buffer during idle periods together 469 * with congestion window. Requires another timer. Has to 470 * wait for upcoming tcp timer rewrite. 471 * 472 * XXXGL: should there be used sbused() or sbavail()? 473 */ 474 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { 475 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && 476 sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) && 477 sbused(&so->so_snd) < V_tcp_autosndbuf_max && 478 sendwin >= (sbused(&so->so_snd) - 479 (tp->snd_nxt - tp->snd_una))) { 480 if (!sbreserve_locked(&so->so_snd, 481 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, 482 V_tcp_autosndbuf_max), so, curthread)) 483 so->so_snd.sb_flags &= ~SB_AUTOSIZE; 484 } 485 } 486 487 /* 488 * Decide if we can use TCP Segmentation Offloading (if supported by 489 * hardware). 490 * 491 * TSO may only be used if we are in a pure bulk sending state. The 492 * presence of TCP-MD5, SACK retransmits, SACK advertizements and 493 * IP options prevent using TSO. With TSO the TCP header is the same 494 * (except for the sequence number) for all generated packets. This 495 * makes it impossible to transmit any options which vary per generated 496 * segment or packet. 497 */ 498 #ifdef IPSEC 499 /* 500 * Pre-calculate here as we save another lookup into the darknesses 501 * of IPsec that way and can actually decide if TSO is ok. 502 */ 503 ipsec_optlen = ipsec_hdrsiz_tcp(tp); 504 #endif 505 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 506 ((tp->t_flags & TF_SIGNATURE) == 0) && 507 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 508 #ifdef IPSEC 509 ipsec_optlen == 0 && 510 #endif 511 tp->t_inpcb->inp_options == NULL && 512 tp->t_inpcb->in6p_options == NULL) 513 tso = 1; 514 515 if (sack_rxmit) { 516 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd))) 517 flags &= ~TH_FIN; 518 } else { 519 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 520 sbused(&so->so_snd))) 521 flags &= ~TH_FIN; 522 } 523 524 recwin = sbspace(&so->so_rcv); 525 526 /* 527 * Sender silly window avoidance. We transmit under the following 528 * conditions when len is non-zero: 529 * 530 * - We have a full segment (or more with TSO) 531 * - This is the last buffer in a write()/send() and we are 532 * either idle or running NODELAY 533 * - we've timed out (e.g. persist timer) 534 * - we have more then 1/2 the maximum send window's worth of 535 * data (receiver may be limited the window size) 536 * - we need to retransmit 537 */ 538 if (len) { 539 if (len >= tp->t_maxseg) 540 goto send; 541 /* 542 * NOTE! on localhost connections an 'ack' from the remote 543 * end may occur synchronously with the output and cause 544 * us to flush a buffer queued with moretocome. XXX 545 * 546 * note: the len + off check is almost certainly unnecessary. 547 */ 548 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 549 (idle || (tp->t_flags & TF_NODELAY)) && 550 len + off >= sbavail(&so->so_snd) && 551 (tp->t_flags & TF_NOPUSH) == 0) { 552 goto send; 553 } 554 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ 555 goto send; 556 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 557 goto send; 558 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ 559 goto send; 560 if (sack_rxmit) 561 goto send; 562 } 563 564 /* 565 * Sending of standalone window updates. 566 * 567 * Window updates are important when we close our window due to a 568 * full socket buffer and are opening it again after the application 569 * reads data from it. Once the window has opened again and the 570 * remote end starts to send again the ACK clock takes over and 571 * provides the most current window information. 572 * 573 * We must avoid the silly window syndrome whereas every read 574 * from the receive buffer, no matter how small, causes a window 575 * update to be sent. We also should avoid sending a flurry of 576 * window updates when the socket buffer had queued a lot of data 577 * and the application is doing small reads. 578 * 579 * Prevent a flurry of pointless window updates by only sending 580 * an update when we can increase the advertized window by more 581 * than 1/4th of the socket buffer capacity. When the buffer is 582 * getting full or is very small be more aggressive and send an 583 * update whenever we can increase by two mss sized segments. 584 * In all other situations the ACK's to new incoming data will 585 * carry further window increases. 586 * 587 * Don't send an independent window update if a delayed 588 * ACK is pending (it will get piggy-backed on it) or the 589 * remote side already has done a half-close and won't send 590 * more data. Skip this if the connection is in T/TCP 591 * half-open state. 592 */ 593 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 594 !(tp->t_flags & TF_DELACK) && 595 !TCPS_HAVERCVDFIN(tp->t_state)) { 596 /* 597 * "adv" is the amount we could increase the window, 598 * taking into account that we are limited by 599 * TCP_MAXWIN << tp->rcv_scale. 600 */ 601 long adv; 602 int oldwin; 603 604 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 605 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 606 oldwin = (tp->rcv_adv - tp->rcv_nxt); 607 adv -= oldwin; 608 } else 609 oldwin = 0; 610 611 /* 612 * If the new window size ends up being the same as the old 613 * size when it is scaled, then don't force a window update. 614 */ 615 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 616 goto dontupdate; 617 618 if (adv >= (long)(2 * tp->t_maxseg) && 619 (adv >= (long)(so->so_rcv.sb_hiwat / 4) || 620 recwin <= (long)(so->so_rcv.sb_hiwat / 8) || 621 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) 622 goto send; 623 } 624 dontupdate: 625 626 /* 627 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 628 * is also a catch-all for the retransmit timer timeout case. 629 */ 630 if (tp->t_flags & TF_ACKNOW) 631 goto send; 632 if ((flags & TH_RST) || 633 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) 634 goto send; 635 if (SEQ_GT(tp->snd_up, tp->snd_una)) 636 goto send; 637 /* 638 * If our state indicates that FIN should be sent 639 * and we have not yet done so, then we need to send. 640 */ 641 if (flags & TH_FIN && 642 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 643 goto send; 644 /* 645 * In SACK, it is possible for tcp_output to fail to send a segment 646 * after the retransmission timer has been turned off. Make sure 647 * that the retransmission timer is set. 648 */ 649 if ((tp->t_flags & TF_SACK_PERMIT) && 650 SEQ_GT(tp->snd_max, tp->snd_una) && 651 !tcp_timer_active(tp, TT_REXMT) && 652 !tcp_timer_active(tp, TT_PERSIST)) { 653 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 654 goto just_return; 655 } 656 /* 657 * TCP window updates are not reliable, rather a polling protocol 658 * using ``persist'' packets is used to insure receipt of window 659 * updates. The three ``states'' for the output side are: 660 * idle not doing retransmits or persists 661 * persisting to move a small or zero window 662 * (re)transmitting and thereby not persisting 663 * 664 * tcp_timer_active(tp, TT_PERSIST) 665 * is true when we are in persist state. 666 * (tp->t_flags & TF_FORCEDATA) 667 * is set when we are called to send a persist packet. 668 * tcp_timer_active(tp, TT_REXMT) 669 * is set when we are retransmitting 670 * The output side is idle when both timers are zero. 671 * 672 * If send window is too small, there is data to transmit, and no 673 * retransmit or persist is pending, then go to persist state. 674 * If nothing happens soon, send when timer expires: 675 * if window is nonzero, transmit what we can, 676 * otherwise force out a byte. 677 */ 678 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && 679 !tcp_timer_active(tp, TT_PERSIST)) { 680 tp->t_rxtshift = 0; 681 tcp_setpersist(tp); 682 } 683 684 /* 685 * No reason to send a segment, just return. 686 */ 687 just_return: 688 SOCKBUF_UNLOCK(&so->so_snd); 689 return (0); 690 691 send: 692 SOCKBUF_LOCK_ASSERT(&so->so_snd); 693 if (len > 0) { 694 if (len >= tp->t_maxseg) 695 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 696 else 697 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 698 } 699 /* 700 * Before ESTABLISHED, force sending of initial options 701 * unless TCP set not to do any options. 702 * NOTE: we assume that the IP/TCP header plus TCP options 703 * always fit in a single mbuf, leaving room for a maximum 704 * link header, i.e. 705 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES 706 */ 707 optlen = 0; 708 #ifdef INET6 709 if (isipv6) 710 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 711 else 712 #endif 713 hdrlen = sizeof (struct tcpiphdr); 714 715 /* 716 * Compute options for segment. 717 * We only have to care about SYN and established connection 718 * segments. Options for SYN-ACK segments are handled in TCP 719 * syncache. 720 */ 721 if ((tp->t_flags & TF_NOOPT) == 0) { 722 to.to_flags = 0; 723 /* Maximum segment size. */ 724 if (flags & TH_SYN) { 725 tp->snd_nxt = tp->iss; 726 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); 727 to.to_flags |= TOF_MSS; 728 } 729 /* Window scaling. */ 730 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 731 to.to_wscale = tp->request_r_scale; 732 to.to_flags |= TOF_SCALE; 733 } 734 /* Timestamps. */ 735 if ((tp->t_flags & TF_RCVD_TSTMP) || 736 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 737 to.to_tsval = tcp_ts_getticks() + tp->ts_offset; 738 to.to_tsecr = tp->ts_recent; 739 to.to_flags |= TOF_TS; 740 /* Set receive buffer autosizing timestamp. */ 741 if (tp->rfbuf_ts == 0 && 742 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 743 tp->rfbuf_ts = tcp_ts_getticks(); 744 } 745 /* Selective ACK's. */ 746 if (tp->t_flags & TF_SACK_PERMIT) { 747 if (flags & TH_SYN) 748 to.to_flags |= TOF_SACKPERM; 749 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 750 (tp->t_flags & TF_SACK_PERMIT) && 751 tp->rcv_numsacks > 0) { 752 to.to_flags |= TOF_SACK; 753 to.to_nsacks = tp->rcv_numsacks; 754 to.to_sacks = (u_char *)tp->sackblks; 755 } 756 } 757 #ifdef TCP_SIGNATURE 758 /* TCP-MD5 (RFC2385). */ 759 if (tp->t_flags & TF_SIGNATURE) 760 to.to_flags |= TOF_SIGNATURE; 761 #endif /* TCP_SIGNATURE */ 762 763 /* Processing the options. */ 764 hdrlen += optlen = tcp_addoptions(&to, opt); 765 } 766 767 #ifdef INET6 768 if (isipv6) 769 ipoptlen = ip6_optlen(tp->t_inpcb); 770 else 771 #endif 772 if (tp->t_inpcb->inp_options) 773 ipoptlen = tp->t_inpcb->inp_options->m_len - 774 offsetof(struct ipoption, ipopt_list); 775 else 776 ipoptlen = 0; 777 #ifdef IPSEC 778 ipoptlen += ipsec_optlen; 779 #endif 780 781 /* 782 * Adjust data length if insertion of options will 783 * bump the packet length beyond the t_maxopd length. 784 * Clear the FIN bit because we cut off the tail of 785 * the segment. 786 */ 787 if (len + optlen + ipoptlen > tp->t_maxopd) { 788 flags &= ~TH_FIN; 789 790 if (tso) { 791 u_int if_hw_tsomax; 792 u_int if_hw_tsomaxsegcount; 793 u_int if_hw_tsomaxsegsize; 794 struct mbuf *mb; 795 u_int moff; 796 int max_len; 797 798 /* extract TSO information */ 799 if_hw_tsomax = tp->t_tsomax; 800 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 801 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 802 803 /* 804 * Limit a TSO burst to prevent it from 805 * overflowing or exceeding the maximum length 806 * allowed by the network interface: 807 */ 808 KASSERT(ipoptlen == 0, 809 ("%s: TSO can't do IP options", __func__)); 810 811 /* 812 * Check if we should limit by maximum payload 813 * length: 814 */ 815 if (if_hw_tsomax != 0) { 816 /* compute maximum TSO length */ 817 max_len = (if_hw_tsomax - hdrlen - 818 max_linkhdr); 819 if (max_len <= 0) { 820 len = 0; 821 } else if (len > max_len) { 822 sendalot = 1; 823 len = max_len; 824 } 825 } 826 827 /* 828 * Check if we should limit by maximum segment 829 * size and count: 830 */ 831 if (if_hw_tsomaxsegcount != 0 && 832 if_hw_tsomaxsegsize != 0) { 833 /* 834 * Subtract one segment for the LINK 835 * and TCP/IP headers mbuf that will 836 * be prepended to this mbuf chain 837 * after the code in this section 838 * limits the number of mbufs in the 839 * chain to if_hw_tsomaxsegcount. 840 */ 841 if_hw_tsomaxsegcount -= 1; 842 max_len = 0; 843 mb = sbsndmbuf(&so->so_snd, off, &moff); 844 845 while (mb != NULL && max_len < len) { 846 u_int mlen; 847 u_int frags; 848 849 /* 850 * Get length of mbuf fragment 851 * and how many hardware frags, 852 * rounded up, it would use: 853 */ 854 mlen = (mb->m_len - moff); 855 frags = howmany(mlen, 856 if_hw_tsomaxsegsize); 857 858 /* Handle special case: Zero Length Mbuf */ 859 if (frags == 0) 860 frags = 1; 861 862 /* 863 * Check if the fragment limit 864 * will be reached or exceeded: 865 */ 866 if (frags >= if_hw_tsomaxsegcount) { 867 max_len += min(mlen, 868 if_hw_tsomaxsegcount * 869 if_hw_tsomaxsegsize); 870 break; 871 } 872 max_len += mlen; 873 if_hw_tsomaxsegcount -= frags; 874 moff = 0; 875 mb = mb->m_next; 876 } 877 if (max_len <= 0) { 878 len = 0; 879 } else if (len > max_len) { 880 sendalot = 1; 881 len = max_len; 882 } 883 } 884 885 /* 886 * Prevent the last segment from being 887 * fractional unless the send sockbuf can be 888 * emptied: 889 */ 890 max_len = (tp->t_maxopd - optlen); 891 if ((off + len) < sbavail(&so->so_snd)) { 892 moff = len % max_len; 893 if (moff != 0) { 894 len -= moff; 895 sendalot = 1; 896 } 897 } 898 899 /* 900 * In case there are too many small fragments 901 * don't use TSO: 902 */ 903 if (len <= max_len) { 904 len = max_len; 905 sendalot = 1; 906 tso = 0; 907 } 908 909 /* 910 * Send the FIN in a separate segment 911 * after the bulk sending is done. 912 * We don't trust the TSO implementations 913 * to clear the FIN flag on all but the 914 * last segment. 915 */ 916 if (tp->t_flags & TF_NEEDFIN) 917 sendalot = 1; 918 919 } else { 920 len = tp->t_maxopd - optlen - ipoptlen; 921 sendalot = 1; 922 } 923 } else 924 tso = 0; 925 926 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 927 ("%s: len > IP_MAXPACKET", __func__)); 928 929 /*#ifdef DIAGNOSTIC*/ 930 #ifdef INET6 931 if (max_linkhdr + hdrlen > MCLBYTES) 932 #else 933 if (max_linkhdr + hdrlen > MHLEN) 934 #endif 935 panic("tcphdr too big"); 936 /*#endif*/ 937 938 /* 939 * This KASSERT is here to catch edge cases at a well defined place. 940 * Before, those had triggered (random) panic conditions further down. 941 */ 942 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 943 944 /* 945 * Grab a header mbuf, attaching a copy of data to 946 * be transmitted, and initialize the header from 947 * the template for sends on this connection. 948 */ 949 if (len) { 950 struct mbuf *mb; 951 u_int moff; 952 953 if ((tp->t_flags & TF_FORCEDATA) && len == 1) 954 TCPSTAT_INC(tcps_sndprobe); 955 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 956 tp->t_sndrexmitpack++; 957 TCPSTAT_INC(tcps_sndrexmitpack); 958 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 959 } else { 960 TCPSTAT_INC(tcps_sndpack); 961 TCPSTAT_ADD(tcps_sndbyte, len); 962 } 963 #ifdef INET6 964 if (MHLEN < hdrlen + max_linkhdr) 965 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 966 else 967 #endif 968 m = m_gethdr(M_NOWAIT, MT_DATA); 969 970 if (m == NULL) { 971 SOCKBUF_UNLOCK(&so->so_snd); 972 error = ENOBUFS; 973 sack_rxmit = 0; 974 goto out; 975 } 976 977 m->m_data += max_linkhdr; 978 m->m_len = hdrlen; 979 980 /* 981 * Start the m_copy functions from the closest mbuf 982 * to the offset in the socket buffer chain. 983 */ 984 mb = sbsndptr(&so->so_snd, off, len, &moff); 985 986 if (len <= MHLEN - hdrlen - max_linkhdr) { 987 m_copydata(mb, moff, (int)len, 988 mtod(m, caddr_t) + hdrlen); 989 m->m_len += len; 990 } else { 991 m->m_next = m_copy(mb, moff, (int)len); 992 if (m->m_next == NULL) { 993 SOCKBUF_UNLOCK(&so->so_snd); 994 (void) m_free(m); 995 error = ENOBUFS; 996 sack_rxmit = 0; 997 goto out; 998 } 999 } 1000 1001 /* 1002 * If we're sending everything we've got, set PUSH. 1003 * (This will keep happy those implementations which only 1004 * give data to the user when a buffer fills or 1005 * a PUSH comes in.) 1006 */ 1007 if (off + len == sbused(&so->so_snd)) 1008 flags |= TH_PUSH; 1009 SOCKBUF_UNLOCK(&so->so_snd); 1010 } else { 1011 SOCKBUF_UNLOCK(&so->so_snd); 1012 if (tp->t_flags & TF_ACKNOW) 1013 TCPSTAT_INC(tcps_sndacks); 1014 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 1015 TCPSTAT_INC(tcps_sndctrl); 1016 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 1017 TCPSTAT_INC(tcps_sndurg); 1018 else 1019 TCPSTAT_INC(tcps_sndwinup); 1020 1021 m = m_gethdr(M_NOWAIT, MT_DATA); 1022 if (m == NULL) { 1023 error = ENOBUFS; 1024 sack_rxmit = 0; 1025 goto out; 1026 } 1027 #ifdef INET6 1028 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 1029 MHLEN >= hdrlen) { 1030 M_ALIGN(m, hdrlen); 1031 } else 1032 #endif 1033 m->m_data += max_linkhdr; 1034 m->m_len = hdrlen; 1035 } 1036 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); 1037 m->m_pkthdr.rcvif = (struct ifnet *)0; 1038 #ifdef MAC 1039 mac_inpcb_create_mbuf(tp->t_inpcb, m); 1040 #endif 1041 #ifdef INET6 1042 if (isipv6) { 1043 ip6 = mtod(m, struct ip6_hdr *); 1044 th = (struct tcphdr *)(ip6 + 1); 1045 tcpip_fillheaders(tp->t_inpcb, ip6, th); 1046 } else 1047 #endif /* INET6 */ 1048 { 1049 ip = mtod(m, struct ip *); 1050 ipov = (struct ipovly *)ip; 1051 th = (struct tcphdr *)(ip + 1); 1052 tcpip_fillheaders(tp->t_inpcb, ip, th); 1053 } 1054 1055 /* 1056 * Fill in fields, remembering maximum advertised 1057 * window for use in delaying messages about window sizes. 1058 * If resending a FIN, be sure not to use a new sequence number. 1059 */ 1060 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 1061 tp->snd_nxt == tp->snd_max) 1062 tp->snd_nxt--; 1063 /* 1064 * If we are starting a connection, send ECN setup 1065 * SYN packet. If we are on a retransmit, we may 1066 * resend those bits a number of times as per 1067 * RFC 3168. 1068 */ 1069 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { 1070 if (tp->t_rxtshift >= 1) { 1071 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 1072 flags |= TH_ECE|TH_CWR; 1073 } else 1074 flags |= TH_ECE|TH_CWR; 1075 } 1076 1077 if (tp->t_state == TCPS_ESTABLISHED && 1078 (tp->t_flags & TF_ECN_PERMIT)) { 1079 /* 1080 * If the peer has ECN, mark data packets with 1081 * ECN capable transmission (ECT). 1082 * Ignore pure ack packets, retransmissions and window probes. 1083 */ 1084 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 1085 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 1086 #ifdef INET6 1087 if (isipv6) 1088 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1089 else 1090 #endif 1091 ip->ip_tos |= IPTOS_ECN_ECT0; 1092 TCPSTAT_INC(tcps_ecn_ect0); 1093 } 1094 1095 /* 1096 * Reply with proper ECN notifications. 1097 */ 1098 if (tp->t_flags & TF_ECN_SND_CWR) { 1099 flags |= TH_CWR; 1100 tp->t_flags &= ~TF_ECN_SND_CWR; 1101 } 1102 if (tp->t_flags & TF_ECN_SND_ECE) 1103 flags |= TH_ECE; 1104 } 1105 1106 /* 1107 * If we are doing retransmissions, then snd_nxt will 1108 * not reflect the first unsent octet. For ACK only 1109 * packets, we do not want the sequence number of the 1110 * retransmitted packet, we want the sequence number 1111 * of the next unsent octet. So, if there is no data 1112 * (and no SYN or FIN), use snd_max instead of snd_nxt 1113 * when filling in ti_seq. But if we are in persist 1114 * state, snd_max might reflect one byte beyond the 1115 * right edge of the window, so use snd_nxt in that 1116 * case, since we know we aren't doing a retransmission. 1117 * (retransmit and persist are mutually exclusive...) 1118 */ 1119 if (sack_rxmit == 0) { 1120 if (len || (flags & (TH_SYN|TH_FIN)) || 1121 tcp_timer_active(tp, TT_PERSIST)) 1122 th->th_seq = htonl(tp->snd_nxt); 1123 else 1124 th->th_seq = htonl(tp->snd_max); 1125 } else { 1126 th->th_seq = htonl(p->rxmit); 1127 p->rxmit += len; 1128 tp->sackhint.sack_bytes_rexmit += len; 1129 } 1130 th->th_ack = htonl(tp->rcv_nxt); 1131 if (optlen) { 1132 bcopy(opt, th + 1, optlen); 1133 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 1134 } 1135 th->th_flags = flags; 1136 /* 1137 * Calculate receive window. Don't shrink window, 1138 * but avoid silly window syndrome. 1139 */ 1140 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 1141 recwin < (long)tp->t_maxseg) 1142 recwin = 0; 1143 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 1144 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 1145 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 1146 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 1147 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 1148 1149 /* 1150 * According to RFC1323 the window field in a SYN (i.e., a <SYN> 1151 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> 1152 * case is handled in syncache. 1153 */ 1154 if (flags & TH_SYN) 1155 th->th_win = htons((u_short) 1156 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 1157 else 1158 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 1159 1160 /* 1161 * Adjust the RXWIN0SENT flag - indicate that we have advertised 1162 * a 0 window. This may cause the remote transmitter to stall. This 1163 * flag tells soreceive() to disable delayed acknowledgements when 1164 * draining the buffer. This can occur if the receiver is attempting 1165 * to read more data than can be buffered prior to transmitting on 1166 * the connection. 1167 */ 1168 if (th->th_win == 0) { 1169 tp->t_sndzerowin++; 1170 tp->t_flags |= TF_RXWIN0SENT; 1171 } else 1172 tp->t_flags &= ~TF_RXWIN0SENT; 1173 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 1174 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 1175 th->th_flags |= TH_URG; 1176 } else 1177 /* 1178 * If no urgent pointer to send, then we pull 1179 * the urgent pointer to the left edge of the send window 1180 * so that it doesn't drift into the send window on sequence 1181 * number wraparound. 1182 */ 1183 tp->snd_up = tp->snd_una; /* drag it along */ 1184 1185 #ifdef TCP_SIGNATURE 1186 if (tp->t_flags & TF_SIGNATURE) { 1187 int sigoff = to.to_signature - opt; 1188 tcp_signature_compute(m, 0, len, optlen, 1189 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND); 1190 } 1191 #endif 1192 1193 /* 1194 * Put TCP length in extended header, and then 1195 * checksum extended header and data. 1196 */ 1197 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 1198 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1199 #ifdef INET6 1200 if (isipv6) { 1201 /* 1202 * ip6_plen is not need to be filled now, and will be filled 1203 * in ip6_output. 1204 */ 1205 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1206 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 1207 optlen + len, IPPROTO_TCP, 0); 1208 } 1209 #endif 1210 #if defined(INET6) && defined(INET) 1211 else 1212 #endif 1213 #ifdef INET 1214 { 1215 m->m_pkthdr.csum_flags = CSUM_TCP; 1216 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1217 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); 1218 1219 /* IP version must be set here for ipv4/ipv6 checking later */ 1220 KASSERT(ip->ip_v == IPVERSION, 1221 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 1222 } 1223 #endif 1224 1225 /* 1226 * Enable TSO and specify the size of the segments. 1227 * The TCP pseudo header checksum is always provided. 1228 */ 1229 if (tso) { 1230 KASSERT(len > tp->t_maxopd - optlen, 1231 ("%s: len <= tso_segsz", __func__)); 1232 m->m_pkthdr.csum_flags |= CSUM_TSO; 1233 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; 1234 } 1235 1236 #ifdef IPSEC 1237 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 1238 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u", 1239 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 1240 #else 1241 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 1242 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u", 1243 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 1244 #endif 1245 1246 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 1247 hhook_run_tcp_est_out(tp, th, &to, len, tso); 1248 1249 #ifdef TCPDEBUG 1250 /* 1251 * Trace. 1252 */ 1253 if (so->so_options & SO_DEBUG) { 1254 u_short save = 0; 1255 #ifdef INET6 1256 if (!isipv6) 1257 #endif 1258 { 1259 save = ipov->ih_len; 1260 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); 1261 } 1262 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 1263 #ifdef INET6 1264 if (!isipv6) 1265 #endif 1266 ipov->ih_len = save; 1267 } 1268 #endif /* TCPDEBUG */ 1269 TCP_PROBE3(debug__input, tp, th, mtod(m, const char *)); 1270 1271 /* 1272 * Fill in IP length and desired time to live and 1273 * send to IP level. There should be a better way 1274 * to handle ttl and tos; we could keep them in 1275 * the template, but need a way to checksum without them. 1276 */ 1277 /* 1278 * m->m_pkthdr.len should have been set before checksum calculation, 1279 * because in6_cksum() need it. 1280 */ 1281 #ifdef INET6 1282 if (isipv6) { 1283 struct route_in6 ro; 1284 1285 bzero(&ro, sizeof(ro)); 1286 /* 1287 * we separately set hoplimit for every segment, since the 1288 * user might want to change the value via setsockopt. 1289 * Also, desired default hop limit might be changed via 1290 * Neighbor Discovery. 1291 */ 1292 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); 1293 1294 /* 1295 * Set the packet size here for the benefit of DTrace probes. 1296 * ip6_output() will set it properly; it's supposed to include 1297 * the option header lengths as well. 1298 */ 1299 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 1300 1301 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) 1302 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1303 else 1304 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1305 1306 if (tp->t_state == TCPS_SYN_SENT) 1307 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 1308 1309 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 1310 1311 #ifdef TCPPCAP 1312 /* Save packet, if requested. */ 1313 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1314 #endif 1315 1316 /* TODO: IPv6 IP6TOS_ECT bit on */ 1317 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro, 1318 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 1319 NULL, NULL, tp->t_inpcb); 1320 1321 if (error == EMSGSIZE && ro.ro_rt != NULL) 1322 mtu = ro.ro_rt->rt_mtu; 1323 RO_RTFREE(&ro); 1324 } 1325 #endif /* INET6 */ 1326 #if defined(INET) && defined(INET6) 1327 else 1328 #endif 1329 #ifdef INET 1330 { 1331 struct route ro; 1332 1333 bzero(&ro, sizeof(ro)); 1334 ip->ip_len = htons(m->m_pkthdr.len); 1335 #ifdef INET6 1336 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) 1337 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); 1338 #endif /* INET6 */ 1339 /* 1340 * If we do path MTU discovery, then we set DF on every packet. 1341 * This might not be the best thing to do according to RFC3390 1342 * Section 2. However the tcp hostcache migitates the problem 1343 * so it affects only the first tcp connection with a host. 1344 * 1345 * NB: Don't set DF on small MTU/MSS to have a safe fallback. 1346 */ 1347 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) { 1348 ip->ip_off |= htons(IP_DF); 1349 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 1350 } else { 1351 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 1352 } 1353 1354 if (tp->t_state == TCPS_SYN_SENT) 1355 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 1356 1357 TCP_PROBE5(send, NULL, tp, ip, tp, th); 1358 1359 #ifdef TCPPCAP 1360 /* Save packet, if requested. */ 1361 tcp_pcap_add(th, m, &(tp->t_outpkts)); 1362 #endif 1363 1364 error = ip_output(m, tp->t_inpcb->inp_options, &ro, 1365 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 1366 tp->t_inpcb); 1367 1368 if (error == EMSGSIZE && ro.ro_rt != NULL) 1369 mtu = ro.ro_rt->rt_mtu; 1370 RO_RTFREE(&ro); 1371 } 1372 #endif /* INET */ 1373 1374 out: 1375 /* 1376 * In transmit state, time the transmission and arrange for 1377 * the retransmit. In persist state, just set snd_max. 1378 */ 1379 if ((tp->t_flags & TF_FORCEDATA) == 0 || 1380 !tcp_timer_active(tp, TT_PERSIST)) { 1381 tcp_seq startseq = tp->snd_nxt; 1382 1383 /* 1384 * Advance snd_nxt over sequence space of this segment. 1385 */ 1386 if (flags & (TH_SYN|TH_FIN)) { 1387 if (flags & TH_SYN) 1388 tp->snd_nxt++; 1389 if (flags & TH_FIN) { 1390 tp->snd_nxt++; 1391 tp->t_flags |= TF_SENTFIN; 1392 } 1393 } 1394 if (sack_rxmit) 1395 goto timer; 1396 tp->snd_nxt += len; 1397 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 1398 tp->snd_max = tp->snd_nxt; 1399 /* 1400 * Time this transmission if not a retransmission and 1401 * not currently timing anything. 1402 */ 1403 if (tp->t_rtttime == 0) { 1404 tp->t_rtttime = ticks; 1405 tp->t_rtseq = startseq; 1406 TCPSTAT_INC(tcps_segstimed); 1407 } 1408 } 1409 1410 /* 1411 * Set retransmit timer if not currently set, 1412 * and not doing a pure ack or a keep-alive probe. 1413 * Initial value for retransmit timer is smoothed 1414 * round-trip time + 2 * round-trip time variance. 1415 * Initialize shift counter which is used for backoff 1416 * of retransmit time. 1417 */ 1418 timer: 1419 if (!tcp_timer_active(tp, TT_REXMT) && 1420 ((sack_rxmit && tp->snd_nxt != tp->snd_max) || 1421 (tp->snd_nxt != tp->snd_una))) { 1422 if (tcp_timer_active(tp, TT_PERSIST)) { 1423 tcp_timer_activate(tp, TT_PERSIST, 0); 1424 tp->t_rxtshift = 0; 1425 } 1426 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1427 } else if (len == 0 && sbavail(&so->so_snd) && 1428 !tcp_timer_active(tp, TT_REXMT) && 1429 !tcp_timer_active(tp, TT_PERSIST)) { 1430 /* 1431 * Avoid a situation where we do not set persist timer 1432 * after a zero window condition. For example: 1433 * 1) A -> B: packet with enough data to fill the window 1434 * 2) B -> A: ACK for #1 + new data (0 window 1435 * advertisement) 1436 * 3) A -> B: ACK for #2, 0 len packet 1437 * 1438 * In this case, A will not activate the persist timer, 1439 * because it chose to send a packet. Unless tcp_output 1440 * is called for some other reason (delayed ack timer, 1441 * another input packet from B, socket syscall), A will 1442 * not send zero window probes. 1443 * 1444 * So, if you send a 0-length packet, but there is data 1445 * in the socket buffer, and neither the rexmt or 1446 * persist timer is already set, then activate the 1447 * persist timer. 1448 */ 1449 tp->t_rxtshift = 0; 1450 tcp_setpersist(tp); 1451 } 1452 } else { 1453 /* 1454 * Persist case, update snd_max but since we are in 1455 * persist mode (no window) we do not update snd_nxt. 1456 */ 1457 int xlen = len; 1458 if (flags & TH_SYN) 1459 ++xlen; 1460 if (flags & TH_FIN) { 1461 ++xlen; 1462 tp->t_flags |= TF_SENTFIN; 1463 } 1464 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) 1465 tp->snd_max = tp->snd_nxt + len; 1466 } 1467 1468 if (error) { 1469 1470 /* 1471 * We know that the packet was lost, so back out the 1472 * sequence number advance, if any. 1473 * 1474 * If the error is EPERM the packet got blocked by the 1475 * local firewall. Normally we should terminate the 1476 * connection but the blocking may have been spurious 1477 * due to a firewall reconfiguration cycle. So we treat 1478 * it like a packet loss and let the retransmit timer and 1479 * timeouts do their work over time. 1480 * XXX: It is a POLA question whether calling tcp_drop right 1481 * away would be the really correct behavior instead. 1482 */ 1483 if (((tp->t_flags & TF_FORCEDATA) == 0 || 1484 !tcp_timer_active(tp, TT_PERSIST)) && 1485 ((flags & TH_SYN) == 0) && 1486 (error != EPERM)) { 1487 if (sack_rxmit) { 1488 p->rxmit -= len; 1489 tp->sackhint.sack_bytes_rexmit -= len; 1490 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 1491 ("sackhint bytes rtx >= 0")); 1492 } else 1493 tp->snd_nxt -= len; 1494 } 1495 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ 1496 switch (error) { 1497 case EPERM: 1498 tp->t_softerror = error; 1499 return (error); 1500 case ENOBUFS: 1501 if (!tcp_timer_active(tp, TT_REXMT) && 1502 !tcp_timer_active(tp, TT_PERSIST)) 1503 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); 1504 tp->snd_cwnd = tp->t_maxseg; 1505 return (0); 1506 case EMSGSIZE: 1507 /* 1508 * For some reason the interface we used initially 1509 * to send segments changed to another or lowered 1510 * its MTU. 1511 * If TSO was active we either got an interface 1512 * without TSO capabilits or TSO was turned off. 1513 * If we obtained mtu from ip_output() then update 1514 * it and try again. 1515 */ 1516 if (tso) 1517 tp->t_flags &= ~TF_TSO; 1518 if (mtu != 0) { 1519 tcp_mss_update(tp, -1, mtu, NULL, NULL); 1520 goto again; 1521 } 1522 return (error); 1523 case EHOSTDOWN: 1524 case EHOSTUNREACH: 1525 case ENETDOWN: 1526 case ENETUNREACH: 1527 if (TCPS_HAVERCVDSYN(tp->t_state)) { 1528 tp->t_softerror = error; 1529 return (0); 1530 } 1531 /* FALLTHROUGH */ 1532 default: 1533 return (error); 1534 } 1535 } 1536 TCPSTAT_INC(tcps_sndtotal); 1537 1538 /* 1539 * Data sent (as far as we can tell). 1540 * If this advertises a larger window than any other segment, 1541 * then remember the size of the advertised window. 1542 * Any pending ACK has now been sent. 1543 */ 1544 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 1545 tp->rcv_adv = tp->rcv_nxt + recwin; 1546 tp->last_ack_sent = tp->rcv_nxt; 1547 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 1548 if (tcp_timer_active(tp, TT_DELACK)) 1549 tcp_timer_activate(tp, TT_DELACK, 0); 1550 #if 0 1551 /* 1552 * This completely breaks TCP if newreno is turned on. What happens 1553 * is that if delayed-acks are turned on on the receiver, this code 1554 * on the transmitter effectively destroys the TCP window, forcing 1555 * it to four packets (1.5Kx4 = 6K window). 1556 */ 1557 if (sendalot && --maxburst) 1558 goto again; 1559 #endif 1560 if (sendalot) 1561 goto again; 1562 return (0); 1563 } 1564 1565 void 1566 tcp_setpersist(struct tcpcb *tp) 1567 { 1568 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; 1569 int tt; 1570 1571 tp->t_flags &= ~TF_PREVVALID; 1572 if (tcp_timer_active(tp, TT_REXMT)) 1573 panic("tcp_setpersist: retransmit pending"); 1574 /* 1575 * Start/restart persistance timer. 1576 */ 1577 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 1578 TCPTV_PERSMIN, TCPTV_PERSMAX); 1579 tcp_timer_activate(tp, TT_PERSIST, tt); 1580 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1581 tp->t_rxtshift++; 1582 } 1583 1584 /* 1585 * Insert TCP options according to the supplied parameters to the place 1586 * optp in a consistent way. Can handle unaligned destinations. 1587 * 1588 * The order of the option processing is crucial for optimal packing and 1589 * alignment for the scarce option space. 1590 * 1591 * The optimal order for a SYN/SYN-ACK segment is: 1592 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + 1593 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. 1594 * 1595 * The SACK options should be last. SACK blocks consume 8*n+2 bytes. 1596 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). 1597 * At minimum we need 10 bytes (to generate 1 SACK block). If both 1598 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, 1599 * we only have 10 bytes for SACK options (40 - (12 + 18)). 1600 */ 1601 int 1602 tcp_addoptions(struct tcpopt *to, u_char *optp) 1603 { 1604 u_int mask, optlen = 0; 1605 1606 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { 1607 if ((to->to_flags & mask) != mask) 1608 continue; 1609 if (optlen == TCP_MAXOLEN) 1610 break; 1611 switch (to->to_flags & mask) { 1612 case TOF_MSS: 1613 while (optlen % 4) { 1614 optlen += TCPOLEN_NOP; 1615 *optp++ = TCPOPT_NOP; 1616 } 1617 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) 1618 continue; 1619 optlen += TCPOLEN_MAXSEG; 1620 *optp++ = TCPOPT_MAXSEG; 1621 *optp++ = TCPOLEN_MAXSEG; 1622 to->to_mss = htons(to->to_mss); 1623 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); 1624 optp += sizeof(to->to_mss); 1625 break; 1626 case TOF_SCALE: 1627 while (!optlen || optlen % 2 != 1) { 1628 optlen += TCPOLEN_NOP; 1629 *optp++ = TCPOPT_NOP; 1630 } 1631 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) 1632 continue; 1633 optlen += TCPOLEN_WINDOW; 1634 *optp++ = TCPOPT_WINDOW; 1635 *optp++ = TCPOLEN_WINDOW; 1636 *optp++ = to->to_wscale; 1637 break; 1638 case TOF_SACKPERM: 1639 while (optlen % 2) { 1640 optlen += TCPOLEN_NOP; 1641 *optp++ = TCPOPT_NOP; 1642 } 1643 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) 1644 continue; 1645 optlen += TCPOLEN_SACK_PERMITTED; 1646 *optp++ = TCPOPT_SACK_PERMITTED; 1647 *optp++ = TCPOLEN_SACK_PERMITTED; 1648 break; 1649 case TOF_TS: 1650 while (!optlen || optlen % 4 != 2) { 1651 optlen += TCPOLEN_NOP; 1652 *optp++ = TCPOPT_NOP; 1653 } 1654 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) 1655 continue; 1656 optlen += TCPOLEN_TIMESTAMP; 1657 *optp++ = TCPOPT_TIMESTAMP; 1658 *optp++ = TCPOLEN_TIMESTAMP; 1659 to->to_tsval = htonl(to->to_tsval); 1660 to->to_tsecr = htonl(to->to_tsecr); 1661 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); 1662 optp += sizeof(to->to_tsval); 1663 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); 1664 optp += sizeof(to->to_tsecr); 1665 break; 1666 case TOF_SIGNATURE: 1667 { 1668 int siglen = TCPOLEN_SIGNATURE - 2; 1669 1670 while (!optlen || optlen % 4 != 2) { 1671 optlen += TCPOLEN_NOP; 1672 *optp++ = TCPOPT_NOP; 1673 } 1674 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) 1675 continue; 1676 optlen += TCPOLEN_SIGNATURE; 1677 *optp++ = TCPOPT_SIGNATURE; 1678 *optp++ = TCPOLEN_SIGNATURE; 1679 to->to_signature = optp; 1680 while (siglen--) 1681 *optp++ = 0; 1682 break; 1683 } 1684 case TOF_SACK: 1685 { 1686 int sackblks = 0; 1687 struct sackblk *sack = (struct sackblk *)to->to_sacks; 1688 tcp_seq sack_seq; 1689 1690 while (!optlen || optlen % 4 != 2) { 1691 optlen += TCPOLEN_NOP; 1692 *optp++ = TCPOPT_NOP; 1693 } 1694 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) 1695 continue; 1696 optlen += TCPOLEN_SACKHDR; 1697 *optp++ = TCPOPT_SACK; 1698 sackblks = min(to->to_nsacks, 1699 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); 1700 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; 1701 while (sackblks--) { 1702 sack_seq = htonl(sack->start); 1703 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1704 optp += sizeof(sack_seq); 1705 sack_seq = htonl(sack->end); 1706 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); 1707 optp += sizeof(sack_seq); 1708 optlen += TCPOLEN_SACK; 1709 sack++; 1710 } 1711 TCPSTAT_INC(tcps_sack_send_blocks); 1712 break; 1713 } 1714 default: 1715 panic("%s: unknown TCP option type", __func__); 1716 break; 1717 } 1718 } 1719 1720 /* Terminate and pad TCP options to a 4 byte boundary. */ 1721 if (optlen % 4) { 1722 optlen += TCPOLEN_EOL; 1723 *optp++ = TCPOPT_EOL; 1724 } 1725 /* 1726 * According to RFC 793 (STD0007): 1727 * "The content of the header beyond the End-of-Option option 1728 * must be header padding (i.e., zero)." 1729 * and later: "The padding is composed of zeros." 1730 */ 1731 while (optlen % 4) { 1732 optlen += TCPOLEN_PAD; 1733 *optp++ = TCPOPT_PAD; 1734 } 1735 1736 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); 1737 return (optlen); 1738 } 1739