1 /* $OpenBSD: tcp_output.c,v 1.139 2023/07/04 10:48:19 bluhm Exp $ */ 2 /* $NetBSD: tcp_output.c,v 1.16 1997/06/03 16:17:09 kml Exp $ */ 3 4 /* 5 * Copyright (c) 1982, 1986, 1988, 1990, 1993 6 * The Regents of the University of California. All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * @(#)COPYRIGHT 1.1 (NRL) 17 January 1995 33 * 34 * NRL grants permission for redistribution and use in source and binary 35 * forms, with or without modification, of the software and documentation 36 * created at NRL provided that the following conditions are met: 37 * 38 * 1. Redistributions of source code must retain the above copyright 39 * notice, this list of conditions and the following disclaimer. 40 * 2. Redistributions in binary form must reproduce the above copyright 41 * notice, this list of conditions and the following disclaimer in the 42 * documentation and/or other materials provided with the distribution. 43 * 3. All advertising materials mentioning features or use of this software 44 * must display the following acknowledgements: 45 * This product includes software developed by the University of 46 * California, Berkeley and its contributors. 47 * This product includes software developed at the Information 48 * Technology Division, US Naval Research Laboratory. 49 * 4. Neither the name of the NRL nor the names of its contributors 50 * may be used to endorse or promote products derived from this software 51 * without specific prior written permission. 52 * 53 * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS 54 * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 55 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A 56 * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NRL OR 57 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 58 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 59 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 60 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 61 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 62 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 63 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 64 * 65 * The views and conclusions contained in the software and documentation 66 * are those of the authors and should not be interpreted as representing 67 * official policies, either expressed or implied, of the US Naval 68 * Research Laboratory (NRL). 69 */ 70 71 #include "pf.h" 72 #include "stoeplitz.h" 73 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/mbuf.h> 77 #include <sys/protosw.h> 78 #include <sys/socket.h> 79 #include <sys/socketvar.h> 80 #include <sys/kernel.h> 81 82 #include <net/if.h> 83 #include <net/if_var.h> 84 #include <net/route.h> 85 #if NPF > 0 86 #include <net/pfvar.h> 87 #endif 88 89 #include <netinet/in.h> 90 #include <netinet/ip.h> 91 #include <netinet/in_pcb.h> 92 #include <netinet/ip_var.h> 93 #include <netinet/tcp.h> 94 #define TCPOUTFLAGS 95 #include <netinet/tcp_fsm.h> 96 #include <netinet/tcp_seq.h> 97 #include <netinet/tcp_timer.h> 98 #include <netinet/tcp_var.h> 99 #include <netinet/tcp_debug.h> 100 101 #ifdef notyet 102 extern struct mbuf *m_copypack(); 103 #endif 104 105 extern int tcprexmtthresh; 106 107 #ifdef TCP_SACK_DEBUG 108 void tcp_print_holes(struct tcpcb *tp); 109 110 void 111 tcp_print_holes(struct tcpcb *tp) 112 { 113 struct sackhole *p = tp->snd_holes; 114 if (p == NULL) 115 return; 116 printf("Hole report: start--end dups rxmit\n"); 117 while (p) { 118 printf("%x--%x d %d r %x\n", p->start, p->end, p->dups, 119 p->rxmit); 120 p = p->next; 121 } 122 printf("\n"); 123 } 124 #endif /* TCP_SACK_DEBUG */ 125 126 /* 127 * Returns pointer to a sackhole if there are any pending retransmissions; 128 * NULL otherwise. 129 */ 130 struct sackhole * 131 tcp_sack_output(struct tcpcb *tp) 132 { 133 struct sackhole *p; 134 135 if (!tp->sack_enable) 136 return (NULL); 137 p = tp->snd_holes; 138 while (p) { 139 if (p->dups >= tcprexmtthresh && SEQ_LT(p->rxmit, p->end)) { 140 if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */ 141 p = p->next; 142 continue; 143 } 144 #ifdef TCP_SACK_DEBUG 145 if (p) 146 tcp_print_holes(tp); 147 #endif 148 return (p); 149 } 150 p = p->next; 151 } 152 return (NULL); 153 } 154 155 /* 156 * After a timeout, the SACK list may be rebuilt. This SACK information 157 * should be used to avoid retransmitting SACKed data. This function 158 * traverses the SACK list to see if snd_nxt should be moved forward. 159 */ 160 161 void 162 tcp_sack_adjust(struct tcpcb *tp) 163 { 164 struct sackhole *cur = tp->snd_holes; 165 if (cur == NULL) 166 return; /* No holes */ 167 if (SEQ_GEQ(tp->snd_nxt, tp->rcv_lastsack)) 168 return; /* We're already beyond any SACKed blocks */ 169 /* 170 * Two cases for which we want to advance snd_nxt: 171 * i) snd_nxt lies between end of one hole and beginning of another 172 * ii) snd_nxt lies between end of last hole and rcv_lastsack 173 */ 174 while (cur->next) { 175 if (SEQ_LT(tp->snd_nxt, cur->end)) 176 return; 177 if (SEQ_GEQ(tp->snd_nxt, cur->next->start)) 178 cur = cur->next; 179 else { 180 tp->snd_nxt = cur->next->start; 181 return; 182 } 183 } 184 if (SEQ_LT(tp->snd_nxt, cur->end)) 185 return; 186 tp->snd_nxt = tp->rcv_lastsack; 187 return; 188 } 189 190 /* 191 * Tcp output routine: figure out what should be sent and send it. 192 */ 193 int 194 tcp_output(struct tcpcb *tp) 195 { 196 struct socket *so = tp->t_inpcb->inp_socket; 197 long len, win, txmaxseg; 198 int off, flags, error; 199 struct mbuf *m; 200 struct tcphdr *th; 201 u_int32_t optbuf[howmany(MAX_TCPOPTLEN, sizeof(u_int32_t))]; 202 u_char *opt = (u_char *)optbuf; 203 unsigned int optlen, hdrlen, packetlen; 204 int idle, sendalot = 0; 205 int i, sack_rxmit = 0; 206 struct sackhole *p; 207 uint32_t now; 208 #ifdef TCP_SIGNATURE 209 unsigned int sigoff; 210 #endif /* TCP_SIGNATURE */ 211 #ifdef TCP_ECN 212 int needect; 213 #endif 214 int tso; 215 216 if (tp->t_flags & TF_BLOCKOUTPUT) { 217 tp->t_flags |= TF_NEEDOUTPUT; 218 return (0); 219 } else 220 tp->t_flags &= ~TF_NEEDOUTPUT; 221 222 #if defined(TCP_SIGNATURE) && defined(DIAGNOSTIC) 223 if (tp->sack_enable && (tp->t_flags & TF_SIGNATURE)) 224 return (EINVAL); 225 #endif /* defined(TCP_SIGNATURE) && defined(DIAGNOSTIC) */ 226 227 now = tcp_now(); 228 229 /* 230 * Determine length of data that should be transmitted, 231 * and flags that will be used. 232 * If there is some data or critical controls (SYN, RST) 233 * to send, then transmit; otherwise, investigate further. 234 */ 235 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 236 if (idle && (now - tp->t_rcvtime) >= tp->t_rxtcur) 237 /* 238 * We have been idle for "a while" and no acks are 239 * expected to clock out any data we send -- 240 * slow start to get ack "clock" running again. 241 */ 242 tp->snd_cwnd = 2 * tp->t_maxseg; 243 244 /* remember 'idle' for next invocation of tcp_output */ 245 if (idle && soissending(so)) { 246 tp->t_flags |= TF_LASTIDLE; 247 idle = 0; 248 } else 249 tp->t_flags &= ~TF_LASTIDLE; 250 251 again: 252 /* 253 * If we've recently taken a timeout, snd_max will be greater than 254 * snd_nxt. There may be SACK information that allows us to avoid 255 * resending already delivered data. Adjust snd_nxt accordingly. 256 */ 257 if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max)) 258 tcp_sack_adjust(tp); 259 off = tp->snd_nxt - tp->snd_una; 260 win = ulmin(tp->snd_wnd, tp->snd_cwnd); 261 262 flags = tcp_outflags[tp->t_state]; 263 264 /* 265 * Send any SACK-generated retransmissions. If we're explicitly trying 266 * to send out new data (when sendalot is 1), bypass this function. 267 * If we retransmit in fast recovery mode, decrement snd_cwnd, since 268 * we're replacing a (future) new transmission with a retransmission 269 * now, and we previously incremented snd_cwnd in tcp_input(). 270 */ 271 if (tp->sack_enable && !sendalot) { 272 if (tp->t_dupacks >= tcprexmtthresh && 273 (p = tcp_sack_output(tp))) { 274 off = p->rxmit - tp->snd_una; 275 sack_rxmit = 1; 276 /* Coalesce holes into a single retransmission */ 277 len = min(tp->t_maxseg, p->end - p->rxmit); 278 if (SEQ_LT(tp->snd_una, tp->snd_last)) 279 tp->snd_cwnd -= tp->t_maxseg; 280 } 281 } 282 283 sendalot = 0; 284 tso = 0; 285 /* 286 * If in persist timeout with window of 0, send 1 byte. 287 * Otherwise, if window is small but nonzero 288 * and timer expired, we will send what we can 289 * and go to transmit state. 290 */ 291 if (tp->t_force) { 292 if (win == 0) { 293 /* 294 * If we still have some data to send, then 295 * clear the FIN bit. Usually this would 296 * happen below when it realizes that we 297 * aren't sending all the data. However, 298 * if we have exactly 1 byte of unset data, 299 * then it won't clear the FIN bit below, 300 * and if we are in persist state, we wind 301 * up sending the packet without recording 302 * that we sent the FIN bit. 303 * 304 * We can't just blindly clear the FIN bit, 305 * because if we don't have any more data 306 * to send then the probe will be the FIN 307 * itself. 308 */ 309 if (off < so->so_snd.sb_cc) 310 flags &= ~TH_FIN; 311 win = 1; 312 } else { 313 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 314 tp->t_rxtshift = 0; 315 } 316 } 317 318 if (!sack_rxmit) { 319 len = ulmin(so->so_snd.sb_cc, win) - off; 320 } 321 322 if (len < 0) { 323 /* 324 * If FIN has been sent but not acked, 325 * but we haven't been called to retransmit, 326 * len will be -1. Otherwise, window shrank 327 * after we sent into it. If window shrank to 0, 328 * cancel pending retransmit, pull snd_nxt back 329 * to (closed) window, and set the persist timer 330 * if it isn't already going. If the window didn't 331 * close completely, just wait for an ACK. 332 */ 333 len = 0; 334 if (win == 0) { 335 TCP_TIMER_DISARM(tp, TCPT_REXMT); 336 tp->t_rxtshift = 0; 337 tp->snd_nxt = tp->snd_una; 338 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) 339 tcp_setpersist(tp); 340 } 341 } 342 343 /* 344 * Never send more than half a buffer full. This insures that we can 345 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and 346 * therefore acks will never be delayed unless we run out of data to 347 * transmit. 348 */ 349 txmaxseg = ulmin(so->so_snd.sb_hiwat / 2, tp->t_maxseg); 350 351 if (len > txmaxseg) { 352 if (tcp_do_tso && 353 tp->t_inpcb->inp_options == NULL && 354 tp->t_inpcb->inp_outputopts6 == NULL && 355 #ifdef TCP_SIGNATURE 356 ((tp->t_flags & TF_SIGNATURE) == 0) && 357 #endif 358 len >= 2 * tp->t_maxseg && 359 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 360 !(flags & (TH_SYN|TH_RST|TH_FIN))) { 361 tso = 1; 362 /* avoid small chopped packets */ 363 if (len > (len / tp->t_maxseg) * tp->t_maxseg) { 364 len = (len / tp->t_maxseg) * tp->t_maxseg; 365 sendalot = 1; 366 } 367 } else { 368 len = txmaxseg; 369 sendalot = 1; 370 } 371 } 372 if (off + len < so->so_snd.sb_cc) 373 flags &= ~TH_FIN; 374 375 win = sbspace(so, &so->so_rcv); 376 377 /* 378 * Sender silly window avoidance. If connection is idle 379 * and can send all data, a maximum segment, 380 * at least a maximum default-size segment do it, 381 * or are forced, do it; otherwise don't bother. 382 * If peer's buffer is tiny, then send 383 * when window is at least half open. 384 * If retransmitting (possibly after persist timer forced us 385 * to send into a small window), then must resend. 386 */ 387 if (len) { 388 if (len >= txmaxseg) 389 goto send; 390 if ((idle || (tp->t_flags & TF_NODELAY)) && 391 len + off >= so->so_snd.sb_cc && !soissending(so) && 392 (tp->t_flags & TF_NOPUSH) == 0) 393 goto send; 394 if (tp->t_force) 395 goto send; 396 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) 397 goto send; 398 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 399 goto send; 400 if (sack_rxmit) 401 goto send; 402 } 403 404 /* 405 * Compare available window to amount of window 406 * known to peer (as advertised window less 407 * next expected input). If the difference is at least two 408 * max size segments, or at least 50% of the maximum possible 409 * window, then want to send a window update to peer. 410 */ 411 if (win > 0) { 412 /* 413 * "adv" is the amount we can increase the window, 414 * taking into account that we are limited by 415 * TCP_MAXWIN << tp->rcv_scale. 416 */ 417 long adv = lmin(win, (long)TCP_MAXWIN << tp->rcv_scale) - 418 (tp->rcv_adv - tp->rcv_nxt); 419 420 if (adv >= (long) (2 * tp->t_maxseg)) 421 goto send; 422 if (2 * adv >= (long) so->so_rcv.sb_hiwat) 423 goto send; 424 } 425 426 /* 427 * Send if we owe peer an ACK. 428 */ 429 if (tp->t_flags & TF_ACKNOW) 430 goto send; 431 if (flags & (TH_SYN|TH_RST)) 432 goto send; 433 if (SEQ_GT(tp->snd_up, tp->snd_una)) 434 goto send; 435 /* 436 * If our state indicates that FIN should be sent 437 * and we have not yet done so, or we're retransmitting the FIN, 438 * then we need to send. 439 */ 440 if (flags & TH_FIN && 441 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) 442 goto send; 443 /* 444 * In SACK, it is possible for tcp_output to fail to send a segment 445 * after the retransmission timer has been turned off. Make sure 446 * that the retransmission timer is set. 447 */ 448 if (SEQ_GT(tp->snd_max, tp->snd_una) && 449 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 450 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 451 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 452 return (0); 453 } 454 455 /* 456 * TCP window updates are not reliable, rather a polling protocol 457 * using ``persist'' packets is used to insure receipt of window 458 * updates. The three ``states'' for the output side are: 459 * idle not doing retransmits or persists 460 * persisting to move a small or zero window 461 * (re)transmitting and thereby not persisting 462 * 463 * tp->t_timer[TCPT_PERSIST] 464 * is set when we are in persist state. 465 * tp->t_force 466 * is set when we are called to send a persist packet. 467 * tp->t_timer[TCPT_REXMT] 468 * is set when we are retransmitting 469 * The output side is idle when both timers are zero. 470 * 471 * If send window is too small, there is data to transmit, and no 472 * retransmit or persist is pending, then go to persist state. 473 * If nothing happens soon, send when timer expires: 474 * if window is nonzero, transmit what we can, 475 * otherwise force out a byte. 476 */ 477 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 478 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 479 tp->t_rxtshift = 0; 480 tcp_setpersist(tp); 481 } 482 483 /* 484 * No reason to send a segment, just return. 485 */ 486 return (0); 487 488 send: 489 /* 490 * Before ESTABLISHED, force sending of initial options 491 * unless TCP set not to do any options. 492 * NOTE: we assume that the IP/TCP header plus TCP options 493 * always fit in a single mbuf, leaving room for a maximum 494 * link header, i.e. 495 * max_linkhdr + sizeof(network header) + sizeof(struct tcphdr + 496 * optlen <= MHLEN 497 */ 498 optlen = 0; 499 500 switch (tp->pf) { 501 case 0: /*default to PF_INET*/ 502 case PF_INET: 503 hdrlen = sizeof(struct ip) + sizeof(struct tcphdr); 504 break; 505 #ifdef INET6 506 case PF_INET6: 507 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 508 break; 509 #endif /* INET6 */ 510 default: 511 return (EPFNOSUPPORT); 512 } 513 514 if (flags & TH_SYN) { 515 tp->snd_nxt = tp->iss; 516 if ((tp->t_flags & TF_NOOPT) == 0) { 517 u_int16_t mss; 518 519 opt[0] = TCPOPT_MAXSEG; 520 opt[1] = 4; 521 mss = htons((u_int16_t) tcp_mss(tp, 0)); 522 memcpy(opt + 2, &mss, sizeof(mss)); 523 optlen = 4; 524 525 if (flags & TH_ACK) 526 tcp_mss_update(tp); 527 /* 528 * If this is the first SYN of connection (not a SYN 529 * ACK), include SACK_PERMIT_HDR option. If this is a 530 * SYN ACK, include SACK_PERMIT_HDR option if peer has 531 * already done so. 532 */ 533 if (tp->sack_enable && ((flags & TH_ACK) == 0 || 534 (tp->t_flags & TF_SACK_PERMIT))) { 535 *((u_int32_t *) (opt + optlen)) = 536 htonl(TCPOPT_SACK_PERMIT_HDR); 537 optlen += 4; 538 } 539 if ((tp->t_flags & TF_REQ_SCALE) && 540 ((flags & TH_ACK) == 0 || 541 (tp->t_flags & TF_RCVD_SCALE))) { 542 *((u_int32_t *) (opt + optlen)) = htonl( 543 TCPOPT_NOP << 24 | 544 TCPOPT_WINDOW << 16 | 545 TCPOLEN_WINDOW << 8 | 546 tp->request_r_scale); 547 optlen += 4; 548 } 549 } 550 } 551 552 /* 553 * Send a timestamp and echo-reply if this is a SYN and our side 554 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side 555 * and our peer have sent timestamps in our SYN's. 556 */ 557 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 558 (flags & TH_RST) == 0 && 559 ((flags & (TH_SYN|TH_ACK)) == TH_SYN || 560 (tp->t_flags & TF_RCVD_TSTMP))) { 561 u_int32_t *lp = (u_int32_t *)(opt + optlen); 562 563 /* Form timestamp option as shown in appendix A of RFC 1323. */ 564 *lp++ = htonl(TCPOPT_TSTAMP_HDR); 565 *lp++ = htonl(now + tp->ts_modulate); 566 *lp = htonl(tp->ts_recent); 567 optlen += TCPOLEN_TSTAMP_APPA; 568 } 569 /* Set receive buffer autosizing timestamp. */ 570 if (tp->rfbuf_ts == 0) { 571 tp->rfbuf_ts = now; 572 tp->rfbuf_cnt = 0; 573 } 574 575 #ifdef TCP_SIGNATURE 576 if (tp->t_flags & TF_SIGNATURE) { 577 u_int8_t *bp = (u_int8_t *)(opt + optlen); 578 579 /* Send signature option */ 580 *(bp++) = TCPOPT_SIGNATURE; 581 *(bp++) = TCPOLEN_SIGNATURE; 582 sigoff = optlen + 2; 583 584 { 585 unsigned int i; 586 587 for (i = 0; i < 16; i++) 588 *(bp++) = 0; 589 } 590 591 592 /* Pad options list to the next 32 bit boundary and 593 * terminate it. 594 */ 595 *bp++ = TCPOPT_NOP; 596 *bp++ = TCPOPT_NOP; 597 598 optlen += TCPOLEN_SIGLEN; 599 } 600 #endif /* TCP_SIGNATURE */ 601 602 /* 603 * Send SACKs if necessary. This should be the last option processed. 604 * Only as many SACKs are sent as are permitted by the maximum options 605 * size. No more than three SACKs are sent. 606 */ 607 if (tp->sack_enable && tp->t_state == TCPS_ESTABLISHED && 608 (tp->t_flags & (TF_SACK_PERMIT|TF_NOOPT)) == TF_SACK_PERMIT && 609 tp->rcv_numsacks) { 610 u_int32_t *lp = (u_int32_t *)(opt + optlen); 611 u_int32_t *olp = lp++; 612 int count = 0; /* actual number of SACKs inserted */ 613 int maxsack = (MAX_TCPOPTLEN - (optlen + 4))/TCPOLEN_SACK; 614 615 tcpstat_inc(tcps_sack_snd_opts); 616 maxsack = min(maxsack, TCP_MAX_SACK); 617 for (i = 0; (i < tp->rcv_numsacks && count < maxsack); i++) { 618 struct sackblk sack = tp->sackblks[i]; 619 if (sack.start == 0 && sack.end == 0) 620 continue; 621 *lp++ = htonl(sack.start); 622 *lp++ = htonl(sack.end); 623 count++; 624 } 625 *olp = htonl(TCPOPT_SACK_HDR|(TCPOLEN_SACK*count+2)); 626 optlen += TCPOLEN_SACK*count + 4; /* including leading NOPs */ 627 } 628 629 #ifdef DIAGNOSTIC 630 if (optlen > MAX_TCPOPTLEN) 631 panic("tcp_output: options too long"); 632 #endif /* DIAGNOSTIC */ 633 634 hdrlen += optlen; 635 636 /* 637 * Adjust data length if insertion of options will 638 * bump the packet length beyond the t_maxopd length. 639 * Clear the FIN bit because we cut off the tail of 640 * the segment. 641 */ 642 if (len > tp->t_maxopd - optlen) { 643 if (tso) { 644 if (len + hdrlen + max_linkhdr > MAXMCLBYTES) { 645 len = MAXMCLBYTES - hdrlen - max_linkhdr; 646 sendalot = 1; 647 } 648 } else { 649 len = tp->t_maxopd - optlen; 650 sendalot = 1; 651 } 652 flags &= ~TH_FIN; 653 } 654 655 #ifdef DIAGNOSTIC 656 if (max_linkhdr + hdrlen > MCLBYTES) 657 panic("tcphdr too big"); 658 #endif 659 660 /* 661 * Grab a header mbuf, attaching a copy of data to 662 * be transmitted, and initialize the header from 663 * the template for sends on this connection. 664 */ 665 if (len) { 666 if (tp->t_force && len == 1) 667 tcpstat_inc(tcps_sndprobe); 668 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 669 tcpstat_pkt(tcps_sndrexmitpack, tcps_sndrexmitbyte, 670 len); 671 tp->t_sndrexmitpack++; 672 } else { 673 tcpstat_pkt(tcps_sndpack, tcps_sndbyte, len); 674 } 675 #ifdef notyet 676 if ((m = m_copypack(so->so_snd.sb_mb, off, 677 (int)len, max_linkhdr + hdrlen)) == 0) { 678 error = ENOBUFS; 679 goto out; 680 } 681 /* 682 * m_copypack left space for our hdr; use it. 683 */ 684 m->m_len += hdrlen; 685 m->m_data -= hdrlen; 686 #else 687 MGETHDR(m, M_DONTWAIT, MT_HEADER); 688 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 689 MCLGET(m, M_DONTWAIT); 690 if ((m->m_flags & M_EXT) == 0) { 691 m_freem(m); 692 m = NULL; 693 } 694 } 695 if (m == NULL) { 696 error = ENOBUFS; 697 goto out; 698 } 699 m->m_data += max_linkhdr; 700 m->m_len = hdrlen; 701 if (len <= m_trailingspace(m)) { 702 m_copydata(so->so_snd.sb_mb, off, (int) len, 703 mtod(m, caddr_t) + hdrlen); 704 m->m_len += len; 705 } else { 706 m->m_next = m_copym(so->so_snd.sb_mb, off, (int) len, 707 M_NOWAIT); 708 if (m->m_next == 0) { 709 (void) m_free(m); 710 error = ENOBUFS; 711 goto out; 712 } 713 } 714 if (so->so_snd.sb_mb->m_flags & M_PKTHDR) 715 m->m_pkthdr.ph_loopcnt = 716 so->so_snd.sb_mb->m_pkthdr.ph_loopcnt; 717 #endif 718 /* 719 * If we're sending everything we've got, set PUSH. 720 * (This will keep happy those implementations which only 721 * give data to the user when a buffer fills or 722 * a PUSH comes in.) 723 */ 724 if (off + len == so->so_snd.sb_cc && !soissending(so)) 725 flags |= TH_PUSH; 726 tp->t_sndtime = now; 727 } else { 728 if (tp->t_flags & TF_ACKNOW) 729 tcpstat_inc(tcps_sndacks); 730 else if (flags & (TH_SYN|TH_FIN|TH_RST)) 731 tcpstat_inc(tcps_sndctrl); 732 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 733 tcpstat_inc(tcps_sndurg); 734 else 735 tcpstat_inc(tcps_sndwinup); 736 737 MGETHDR(m, M_DONTWAIT, MT_HEADER); 738 if (m != NULL && max_linkhdr + hdrlen > MHLEN) { 739 MCLGET(m, M_DONTWAIT); 740 if ((m->m_flags & M_EXT) == 0) { 741 m_freem(m); 742 m = NULL; 743 } 744 } 745 if (m == NULL) { 746 error = ENOBUFS; 747 goto out; 748 } 749 m->m_data += max_linkhdr; 750 m->m_len = hdrlen; 751 } 752 m->m_pkthdr.ph_ifidx = 0; 753 m->m_pkthdr.len = hdrlen + len; 754 755 /* Enable TSO and specify the size of the resulting segments. */ 756 if (tso) { 757 SET(m->m_pkthdr.csum_flags, M_TCP_TSO); 758 m->m_pkthdr.ph_mss = tp->t_maxseg; 759 } 760 761 if (!tp->t_template) 762 panic("tcp_output"); 763 #ifdef DIAGNOSTIC 764 if (tp->t_template->m_len != hdrlen - optlen) 765 panic("tcp_output: template len != hdrlen - optlen"); 766 #endif /* DIAGNOSTIC */ 767 memcpy(mtod(m, caddr_t), mtod(tp->t_template, caddr_t), 768 tp->t_template->m_len); 769 th = (struct tcphdr *)(mtod(m, caddr_t) + tp->t_template->m_len - 770 sizeof(struct tcphdr)); 771 772 /* 773 * Fill in fields, remembering maximum advertised 774 * window for use in delaying messages about window sizes. 775 * If resending a FIN, be sure not to use a new sequence number. 776 */ 777 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) && 778 (tp->snd_nxt == tp->snd_max)) 779 tp->snd_nxt--; 780 /* 781 * If we are doing retransmissions, then snd_nxt will 782 * not reflect the first unsent octet. For ACK only 783 * packets, we do not want the sequence number of the 784 * retransmitted packet, we want the sequence number 785 * of the next unsent octet. So, if there is no data 786 * (and no SYN or FIN), use snd_max instead of snd_nxt 787 * when filling in ti_seq. But if we are in persist 788 * state, snd_max might reflect one byte beyond the 789 * right edge of the window, so use snd_nxt in that 790 * case, since we know we aren't doing a retransmission. 791 * (retransmit and persist are mutually exclusive...) 792 */ 793 if (len || (flags & (TH_SYN|TH_FIN)) || 794 TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) 795 th->th_seq = htonl(tp->snd_nxt); 796 else 797 th->th_seq = htonl(tp->snd_max); 798 799 if (sack_rxmit) { 800 /* 801 * If sendalot was turned on (due to option stuffing), turn it 802 * off. Properly set th_seq field. Advance the ret'x pointer 803 * by len. 804 */ 805 if (sendalot) 806 sendalot = 0; 807 th->th_seq = htonl(p->rxmit); 808 p->rxmit += len; 809 tcpstat_pkt(tcps_sack_rexmits, tcps_sack_rexmit_bytes, len); 810 } 811 812 th->th_ack = htonl(tp->rcv_nxt); 813 if (optlen) { 814 memcpy(th + 1, opt, optlen); 815 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; 816 } 817 #ifdef TCP_ECN 818 if (tcp_do_ecn) { 819 /* 820 * if we have received congestion experienced segs, 821 * set ECE bit. 822 */ 823 if (tp->t_flags & TF_RCVD_CE) { 824 flags |= TH_ECE; 825 tcpstat_inc(tcps_ecn_sndece); 826 } 827 if (!(tp->t_flags & TF_DISABLE_ECN)) { 828 /* 829 * if this is a SYN seg, set ECE and CWR. 830 * set only ECE for SYN-ACK if peer supports ECN. 831 */ 832 if ((flags & (TH_SYN|TH_ACK)) == TH_SYN) 833 flags |= (TH_ECE|TH_CWR); 834 else if ((tp->t_flags & TF_ECN_PERMIT) && 835 (flags & (TH_SYN|TH_ACK)) == (TH_SYN|TH_ACK)) 836 flags |= TH_ECE; 837 } 838 /* 839 * if we have reduced the congestion window, notify 840 * the peer by setting CWR bit. 841 */ 842 if ((tp->t_flags & TF_ECN_PERMIT) && 843 (tp->t_flags & TF_SEND_CWR)) { 844 flags |= TH_CWR; 845 tp->t_flags &= ~TF_SEND_CWR; 846 tcpstat_inc(tcps_ecn_sndcwr); 847 } 848 } 849 #endif 850 th->th_flags = flags; 851 852 /* 853 * Calculate receive window. Don't shrink window, 854 * but avoid silly window syndrome. 855 */ 856 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg) 857 win = 0; 858 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 859 win = (long)TCP_MAXWIN << tp->rcv_scale; 860 if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt)) 861 win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt); 862 if (flags & TH_RST) 863 win = 0; 864 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale)); 865 if (th->th_win == 0) 866 tp->t_sndzerowin++; 867 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 868 u_int32_t urp = tp->snd_up - tp->snd_nxt; 869 if (urp > IP_MAXPACKET) 870 urp = IP_MAXPACKET; 871 th->th_urp = htons((u_int16_t)urp); 872 th->th_flags |= TH_URG; 873 } else 874 /* 875 * If no urgent pointer to send, then we pull 876 * the urgent pointer to the left edge of the send window 877 * so that it doesn't drift into the send window on sequence 878 * number wraparound. 879 */ 880 tp->snd_up = tp->snd_una; /* drag it along */ 881 882 #ifdef TCP_SIGNATURE 883 if (tp->t_flags & TF_SIGNATURE) { 884 int iphlen; 885 union sockaddr_union src, dst; 886 struct tdb *tdb; 887 888 bzero(&src, sizeof(union sockaddr_union)); 889 bzero(&dst, sizeof(union sockaddr_union)); 890 891 switch (tp->pf) { 892 case 0: /*default to PF_INET*/ 893 case AF_INET: 894 iphlen = sizeof(struct ip); 895 src.sa.sa_len = sizeof(struct sockaddr_in); 896 src.sa.sa_family = AF_INET; 897 src.sin.sin_addr = mtod(m, struct ip *)->ip_src; 898 dst.sa.sa_len = sizeof(struct sockaddr_in); 899 dst.sa.sa_family = AF_INET; 900 dst.sin.sin_addr = mtod(m, struct ip *)->ip_dst; 901 break; 902 #ifdef INET6 903 case AF_INET6: 904 iphlen = sizeof(struct ip6_hdr); 905 src.sa.sa_len = sizeof(struct sockaddr_in6); 906 src.sa.sa_family = AF_INET6; 907 src.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_src; 908 dst.sa.sa_len = sizeof(struct sockaddr_in6); 909 dst.sa.sa_family = AF_INET6; 910 dst.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_dst; 911 break; 912 #endif /* INET6 */ 913 } 914 915 tdb = gettdbbysrcdst(rtable_l2(tp->t_inpcb->inp_rtableid), 916 0, &src, &dst, IPPROTO_TCP); 917 if (tdb == NULL) { 918 m_freem(m); 919 return (EPERM); 920 } 921 922 if (tcp_signature(tdb, tp->pf, m, th, iphlen, 0, 923 mtod(m, caddr_t) + hdrlen - optlen + sigoff) < 0) { 924 m_freem(m); 925 tdb_unref(tdb); 926 return (EINVAL); 927 } 928 tdb_unref(tdb); 929 } 930 #endif /* TCP_SIGNATURE */ 931 932 /* Defer checksumming until later (ip_output() or hardware) */ 933 m->m_pkthdr.csum_flags |= M_TCP_CSUM_OUT; 934 935 /* 936 * In transmit state, time the transmission and arrange for 937 * the retransmit. In persist state, just set snd_max. 938 */ 939 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 940 tcp_seq startseq = tp->snd_nxt; 941 942 /* 943 * Advance snd_nxt over sequence space of this segment. 944 */ 945 if (flags & (TH_SYN|TH_FIN)) { 946 if (flags & TH_SYN) 947 tp->snd_nxt++; 948 if (flags & TH_FIN) { 949 tp->snd_nxt++; 950 tp->t_flags |= TF_SENTFIN; 951 } 952 } 953 if (tp->sack_enable) { 954 if (sack_rxmit && (p->rxmit != tp->snd_nxt)) { 955 goto timer; 956 } 957 } 958 tp->snd_nxt += len; 959 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 960 tp->snd_max = tp->snd_nxt; 961 /* 962 * Time this transmission if not a retransmission and 963 * not currently timing anything. 964 */ 965 if (tp->t_rtttime == 0) { 966 tp->t_rtttime = now; 967 tp->t_rtseq = startseq; 968 tcpstat_inc(tcps_segstimed); 969 } 970 } 971 972 /* 973 * Set retransmit timer if not currently set, 974 * and not doing an ack or a keep-alive probe. 975 * Initial value for retransmit timer is smoothed 976 * round-trip time + 2 * round-trip time variance. 977 * Initialize shift counter which is used for backoff 978 * of retransmit time. 979 */ 980 timer: 981 if (tp->sack_enable && sack_rxmit && 982 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 983 tp->snd_nxt != tp->snd_max) { 984 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 985 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 986 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 987 tp->t_rxtshift = 0; 988 } 989 } 990 991 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 992 tp->snd_nxt != tp->snd_una) { 993 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur); 994 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) { 995 TCP_TIMER_DISARM(tp, TCPT_PERSIST); 996 tp->t_rxtshift = 0; 997 } 998 } 999 1000 if (len == 0 && so->so_snd.sb_cc && 1001 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 && 1002 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) { 1003 /* 1004 * Avoid a situation where we do not set persist timer 1005 * after a zero window condition. For example: 1006 * 1) A -> B: packet with enough data to fill the window 1007 * 2) B -> A: ACK for #1 + new data (0 window 1008 * advertisement) 1009 * 3) A -> B: ACK for #2, 0 len packet 1010 * 1011 * In this case, A will not activate the persist timer, 1012 * because it chose to send a packet. Unless tcp_output 1013 * is called for some other reason (delayed ack timer, 1014 * another input packet from B, socket syscall), A will 1015 * not send zero window probes. 1016 * 1017 * So, if you send a 0-length packet, but there is data 1018 * in the socket buffer, and neither the rexmt or 1019 * persist timer is already set, then activate the 1020 * persist timer. 1021 */ 1022 tp->t_rxtshift = 0; 1023 tcp_setpersist(tp); 1024 } 1025 } else 1026 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max)) 1027 tp->snd_max = tp->snd_nxt + len; 1028 1029 tcp_update_sndspace(tp); 1030 1031 /* 1032 * Trace. 1033 */ 1034 if (so->so_options & SO_DEBUG) 1035 tcp_trace(TA_OUTPUT, tp->t_state, tp, tp, mtod(m, caddr_t), 0, 1036 len); 1037 1038 /* 1039 * Fill in IP length and desired time to live and 1040 * send to IP level. There should be a better way 1041 * to handle ttl and tos; we could keep them in 1042 * the template, but need a way to checksum without them. 1043 */ 1044 1045 #ifdef TCP_ECN 1046 /* 1047 * if peer is ECN capable, set the ECT bit in the IP header. 1048 * but don't set ECT for a pure ack, a retransmit or a window probe. 1049 */ 1050 needect = 0; 1051 if (tcp_do_ecn && (tp->t_flags & TF_ECN_PERMIT)) { 1052 if (len == 0 || SEQ_LT(tp->snd_nxt, tp->snd_max) || 1053 (tp->t_force && len == 1)) { 1054 /* don't set ECT */ 1055 } else { 1056 needect = 1; 1057 tcpstat_inc(tcps_ecn_sndect); 1058 } 1059 } 1060 #endif 1061 1062 /* force routing table */ 1063 m->m_pkthdr.ph_rtableid = tp->t_inpcb->inp_rtableid; 1064 1065 #if NPF > 0 1066 pf_mbuf_link_inpcb(m, tp->t_inpcb); 1067 #endif 1068 1069 switch (tp->pf) { 1070 case 0: /*default to PF_INET*/ 1071 case AF_INET: 1072 { 1073 struct ip *ip; 1074 1075 ip = mtod(m, struct ip *); 1076 ip->ip_len = htons(m->m_pkthdr.len); 1077 packetlen = m->m_pkthdr.len; 1078 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl; 1079 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos; 1080 #ifdef TCP_ECN 1081 if (needect) 1082 ip->ip_tos |= IPTOS_ECN_ECT0; 1083 #endif 1084 } 1085 #if NSTOEPLITZ > 0 1086 m->m_pkthdr.ph_flowid = tp->t_inpcb->inp_flowid; 1087 SET(m->m_pkthdr.csum_flags, M_FLOWID); 1088 #endif 1089 error = ip_output(m, tp->t_inpcb->inp_options, 1090 &tp->t_inpcb->inp_route, 1091 (ip_mtudisc ? IP_MTUDISC : 0), NULL, tp->t_inpcb, 0); 1092 break; 1093 #ifdef INET6 1094 case AF_INET6: 1095 { 1096 struct ip6_hdr *ip6; 1097 1098 ip6 = mtod(m, struct ip6_hdr *); 1099 ip6->ip6_plen = m->m_pkthdr.len - 1100 sizeof(struct ip6_hdr); 1101 packetlen = m->m_pkthdr.len; 1102 ip6->ip6_nxt = IPPROTO_TCP; 1103 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb); 1104 #ifdef TCP_ECN 1105 if (needect) 1106 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 1107 #endif 1108 } 1109 error = ip6_output(m, tp->t_inpcb->inp_outputopts6, 1110 &tp->t_inpcb->inp_route6, 0, NULL, tp->t_inpcb); 1111 break; 1112 #endif /* INET6 */ 1113 } 1114 1115 if (error) { 1116 out: 1117 if (error == ENOBUFS) { 1118 /* 1119 * If the interface queue is full, or IP cannot 1120 * get an mbuf, trigger TCP slow start. 1121 */ 1122 tp->snd_cwnd = tp->t_maxseg; 1123 return (0); 1124 } 1125 if (error == EMSGSIZE) { 1126 /* 1127 * ip_output() will have already fixed the route 1128 * for us. tcp_mtudisc() will, as its last action, 1129 * initiate retransmission, so it is important to 1130 * not do so here. 1131 */ 1132 tcp_mtudisc(tp->t_inpcb, -1); 1133 return (0); 1134 } 1135 if ((error == EHOSTUNREACH || error == ENETDOWN) && 1136 TCPS_HAVERCVDSYN(tp->t_state)) { 1137 tp->t_softerror = error; 1138 return (0); 1139 } 1140 1141 /* Restart the delayed ACK timer, if necessary. */ 1142 if (TCP_TIMER_ISARMED(tp, TCPT_DELACK)) 1143 TCP_TIMER_ARM(tp, TCPT_DELACK, tcp_delack_msecs); 1144 1145 return (error); 1146 } 1147 1148 if (packetlen > tp->t_pmtud_mtu_sent) 1149 tp->t_pmtud_mtu_sent = packetlen; 1150 1151 tcpstat_inc(tcps_sndtotal); 1152 if (TCP_TIMER_ISARMED(tp, TCPT_DELACK)) 1153 tcpstat_inc(tcps_delack); 1154 1155 /* 1156 * Data sent (as far as we can tell). 1157 * If this advertises a larger window than any other segment, 1158 * then remember the size of the advertised window. 1159 * Any pending ACK has now been sent. 1160 */ 1161 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv)) 1162 tp->rcv_adv = tp->rcv_nxt + win; 1163 tp->last_ack_sent = tp->rcv_nxt; 1164 tp->t_sndacktime = now; 1165 tp->t_flags &= ~TF_ACKNOW; 1166 TCP_TIMER_DISARM(tp, TCPT_DELACK); 1167 if (sendalot) 1168 goto again; 1169 return (0); 1170 } 1171 1172 void 1173 tcp_setpersist(struct tcpcb *tp) 1174 { 1175 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + TCP_RTT_BASE_SHIFT); 1176 int msec; 1177 1178 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT)) 1179 panic("tcp_output REXMT"); 1180 /* 1181 * Start/restart persistence timer. 1182 */ 1183 if (t < tp->t_rttmin) 1184 t = tp->t_rttmin; 1185 TCPT_RANGESET(msec, t * tcp_backoff[tp->t_rxtshift], 1186 TCPTV_PERSMIN, TCPTV_PERSMAX); 1187 TCP_TIMER_ARM(tp, TCPT_PERSIST, msec); 1188 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 1189 tp->t_rxtshift++; 1190 } 1191 1192 int 1193 tcp_chopper(struct mbuf *m0, struct mbuf_list *ml, struct ifnet *ifp, 1194 u_int mss) 1195 { 1196 struct ip *ip = NULL; 1197 #ifdef INET6 1198 struct ip6_hdr *ip6 = NULL; 1199 #endif 1200 struct tcphdr *th; 1201 int firstlen, iphlen, hlen, tlen, off; 1202 int error; 1203 1204 ml_init(ml); 1205 ml_enqueue(ml, m0); 1206 1207 ip = mtod(m0, struct ip *); 1208 switch (ip->ip_v) { 1209 case 4: 1210 iphlen = ip->ip_hl << 2; 1211 if (ISSET(ip->ip_off, htons(IP_OFFMASK | IP_MF)) || 1212 iphlen != sizeof(struct ip) || ip->ip_p != IPPROTO_TCP) { 1213 /* only TCP without fragment or IP option supported */ 1214 error = EPROTOTYPE; 1215 goto bad; 1216 } 1217 break; 1218 #ifdef INET6 1219 case 6: 1220 ip = NULL; 1221 ip6 = mtod(m0, struct ip6_hdr *); 1222 iphlen = sizeof(struct ip6_hdr); 1223 if (ip6->ip6_nxt != IPPROTO_TCP) { 1224 /* only TCP without IPv6 header chain supported */ 1225 error = EPROTOTYPE; 1226 goto bad; 1227 } 1228 break; 1229 #endif 1230 default: 1231 panic("%s: unknown ip version %d", __func__, ip->ip_v); 1232 } 1233 1234 tlen = m0->m_pkthdr.len; 1235 if (tlen < iphlen + sizeof(struct tcphdr)) { 1236 error = ENOPROTOOPT; 1237 goto bad; 1238 } 1239 /* IP and TCP header should be contiguous, this check is paranoia */ 1240 if (m0->m_len < iphlen + sizeof(*th)) { 1241 ml_dequeue(ml); 1242 if ((m0 = m_pullup(m0, iphlen + sizeof(*th))) == NULL) { 1243 error = ENOBUFS; 1244 goto bad; 1245 } 1246 ml_enqueue(ml, m0); 1247 } 1248 th = (struct tcphdr *)(mtod(m0, caddr_t) + iphlen); 1249 hlen = iphlen + (th->th_off << 2); 1250 if (tlen < hlen) { 1251 error = ENOPROTOOPT; 1252 goto bad; 1253 } 1254 firstlen = MIN(tlen - hlen, mss); 1255 1256 CLR(m0->m_pkthdr.csum_flags, M_TCP_TSO); 1257 1258 /* 1259 * Loop through length of payload after first segment, 1260 * make new header and copy data of each part and link onto chain. 1261 */ 1262 for (off = hlen + firstlen; off < tlen; off += mss) { 1263 struct mbuf *m; 1264 struct tcphdr *mhth; 1265 int len; 1266 1267 len = MIN(tlen - off, mss); 1268 1269 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1270 if (m == NULL) { 1271 error = ENOBUFS; 1272 goto bad; 1273 } 1274 ml_enqueue(ml, m); 1275 if ((error = m_dup_pkthdr(m, m0, M_DONTWAIT)) != 0) 1276 goto bad; 1277 1278 /* IP and TCP header to the end, space for link layer header */ 1279 m->m_len = hlen; 1280 m_align(m, hlen); 1281 1282 /* copy and adjust TCP header */ 1283 mhth = (struct tcphdr *)(mtod(m, caddr_t) + iphlen); 1284 memcpy(mhth, th, hlen - iphlen); 1285 mhth->th_seq = htonl(ntohl(th->th_seq) + (off - hlen)); 1286 if (off + len < tlen) 1287 CLR(mhth->th_flags, TH_PUSH|TH_FIN); 1288 1289 /* add mbuf chain with payload */ 1290 m->m_pkthdr.len = hlen + len; 1291 if ((m->m_next = m_copym(m0, off, len, M_DONTWAIT)) == NULL) { 1292 error = ENOBUFS; 1293 goto bad; 1294 } 1295 1296 /* copy and adjust IP header, calculate checksum */ 1297 SET(m->m_pkthdr.csum_flags, M_TCP_CSUM_OUT); 1298 if (ip) { 1299 struct ip *mhip; 1300 1301 mhip = mtod(m, struct ip *); 1302 *mhip = *ip; 1303 mhip->ip_len = htons(hlen + len); 1304 mhip->ip_id = htons(ip_randomid()); 1305 in_hdr_cksum_out(m, ifp); 1306 in_proto_cksum_out(m, ifp); 1307 } 1308 #ifdef INET6 1309 if (ip6) { 1310 struct ip6_hdr *mhip6; 1311 1312 mhip6 = mtod(m, struct ip6_hdr *); 1313 *mhip6 = *ip6; 1314 mhip6->ip6_plen = htons(hlen - iphlen + len); 1315 in6_proto_cksum_out(m, ifp); 1316 } 1317 #endif 1318 } 1319 1320 /* 1321 * Update first segment by trimming what's been copied out 1322 * and updating header, then send each segment (in order). 1323 */ 1324 if (hlen + firstlen < tlen) { 1325 m_adj(m0, hlen + firstlen - tlen); 1326 CLR(th->th_flags, TH_PUSH|TH_FIN); 1327 } 1328 /* adjust IP header, calculate checksum */ 1329 SET(m0->m_pkthdr.csum_flags, M_TCP_CSUM_OUT); 1330 if (ip) { 1331 ip->ip_len = htons(m0->m_pkthdr.len); 1332 in_hdr_cksum_out(m0, ifp); 1333 in_proto_cksum_out(m0, ifp); 1334 } 1335 #ifdef INET6 1336 if (ip6) { 1337 ip6->ip6_plen = htons(m0->m_pkthdr.len - iphlen); 1338 in6_proto_cksum_out(m0, ifp); 1339 } 1340 #endif 1341 1342 tcpstat_add(tcps_outpkttso, ml_len(ml)); 1343 return 0; 1344 1345 bad: 1346 tcpstat_inc(tcps_outbadtso); 1347 ml_purge(ml); 1348 return error; 1349 } 1350 1351 int 1352 tcp_if_output_tso(struct ifnet *ifp, struct mbuf **mp, struct sockaddr *dst, 1353 struct rtentry *rt, uint32_t ifcap, u_int mtu) 1354 { 1355 struct mbuf_list ml; 1356 int error; 1357 1358 /* caller must fail later or fragment */ 1359 if (!ISSET((*mp)->m_pkthdr.csum_flags, M_TCP_TSO)) 1360 return 0; 1361 if ((*mp)->m_pkthdr.ph_mss > mtu) { 1362 CLR((*mp)->m_pkthdr.csum_flags, M_TCP_TSO); 1363 return 0; 1364 } 1365 1366 /* network interface hardware will do TSO */ 1367 if (in_ifcap_cksum(*mp, ifp, ifcap)) { 1368 if (ISSET(ifcap, IFCAP_TSOv4)) { 1369 in_hdr_cksum_out(*mp, ifp); 1370 in_proto_cksum_out(*mp, ifp); 1371 } 1372 #ifdef INET6 1373 if (ISSET(ifcap, IFCAP_TSOv6)) 1374 in6_proto_cksum_out(*mp, ifp); 1375 #endif 1376 error = ifp->if_output(ifp, *mp, dst, rt); 1377 if (!error) 1378 tcpstat_inc(tcps_outhwtso); 1379 goto done; 1380 } 1381 1382 /* as fallback do TSO in software */ 1383 if ((error = tcp_chopper(*mp, &ml, ifp, (*mp)->m_pkthdr.ph_mss)) || 1384 (error = if_output_ml(ifp, &ml, dst, rt))) 1385 goto done; 1386 tcpstat_inc(tcps_outswtso); 1387 1388 done: 1389 *mp = NULL; 1390 return error; 1391 } 1392