1 /* 2 * Copyright (c) 2002, 2003, 2004 Jeffrey M. Hsu. All rights reserved. 3 * Copyright (c) 2002, 2003, 2004 The DragonFly Project. All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Jeffrey M. Hsu. 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 DragonFly Project nor the names of its 17 * contributors may be used to endorse or promote products derived 18 * from this software without specific, prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995 36 * The Regents of the University of California. All rights reserved. 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. Neither the name of the University nor the names of its contributors 47 * may be used to endorse or promote products derived from this software 48 * without specific prior written permission. 49 * 50 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 53 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 60 * SUCH DAMAGE. 61 * 62 * @(#)tcp_input.c 8.12 (Berkeley) 5/24/95 63 * $FreeBSD: src/sys/netinet/tcp_input.c,v 1.107.2.38 2003/05/21 04:46:41 cjc Exp $ 64 */ 65 66 #include "opt_inet.h" 67 #include "opt_inet6.h" 68 #include "opt_ipsec.h" 69 #include "opt_tcpdebug.h" 70 #include "opt_tcp_input.h" 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/kernel.h> 75 #include <sys/sysctl.h> 76 #include <sys/malloc.h> 77 #include <sys/mbuf.h> 78 #include <sys/proc.h> /* for proc0 declaration */ 79 #include <sys/protosw.h> 80 #include <sys/socket.h> 81 #include <sys/socketvar.h> 82 #include <sys/syslog.h> 83 #include <sys/in_cksum.h> 84 85 #include <sys/socketvar2.h> 86 87 #include <machine/cpu.h> /* before tcp_seq.h, for tcp_random18() */ 88 #include <machine/stdarg.h> 89 90 #include <net/if.h> 91 #include <net/route.h> 92 93 #include <netinet/in.h> 94 #include <netinet/in_systm.h> 95 #include <netinet/ip.h> 96 #include <netinet/ip_icmp.h> /* for ICMP_BANDLIM */ 97 #include <netinet/in_var.h> 98 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 99 #include <netinet/in_pcb.h> 100 #include <netinet/ip_var.h> 101 #include <netinet/ip6.h> 102 #include <netinet/icmp6.h> 103 #include <netinet6/nd6.h> 104 #include <netinet6/ip6_var.h> 105 #include <netinet6/in6_pcb.h> 106 #include <netinet/tcp.h> 107 #include <netinet/tcp_fsm.h> 108 #include <netinet/tcp_seq.h> 109 #include <netinet/tcp_timer.h> 110 #include <netinet/tcp_timer2.h> 111 #include <netinet/tcp_var.h> 112 #include <netinet6/tcp6_var.h> 113 #include <netinet/tcpip.h> 114 115 #ifdef TCPDEBUG 116 #include <netinet/tcp_debug.h> 117 118 u_char tcp_saveipgen[40]; /* the size must be of max ip header, now IPv6 */ 119 struct tcphdr tcp_savetcp; 120 #endif 121 122 #ifdef FAST_IPSEC 123 #include <netproto/ipsec/ipsec.h> 124 #include <netproto/ipsec/ipsec6.h> 125 #endif 126 127 #ifdef IPSEC 128 #include <netinet6/ipsec.h> 129 #include <netinet6/ipsec6.h> 130 #include <netproto/key/key.h> 131 #endif 132 133 /* 134 * Limit burst of new packets during SACK based fast recovery 135 * or extended limited transmit. 136 */ 137 #define TCP_SACK_MAXBURST 4 138 139 MALLOC_DEFINE(M_TSEGQ, "tseg_qent", "TCP segment queue entry"); 140 141 static int log_in_vain = 0; 142 SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_RW, 143 &log_in_vain, 0, "Log all incoming TCP connections"); 144 145 static int blackhole = 0; 146 SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_RW, 147 &blackhole, 0, "Do not send RST when dropping refused connections"); 148 149 int tcp_delack_enabled = 1; 150 SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_RW, 151 &tcp_delack_enabled, 0, 152 "Delay ACK to try and piggyback it onto a data packet"); 153 154 #ifdef TCP_DROP_SYNFIN 155 static int drop_synfin = 0; 156 SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_RW, 157 &drop_synfin, 0, "Drop TCP packets with SYN+FIN set"); 158 #endif 159 160 static int tcp_do_limitedtransmit = 1; 161 SYSCTL_INT(_net_inet_tcp, OID_AUTO, limitedtransmit, CTLFLAG_RW, 162 &tcp_do_limitedtransmit, 0, "Enable RFC 3042 (Limited Transmit)"); 163 164 static int tcp_do_early_retransmit = 1; 165 SYSCTL_INT(_net_inet_tcp, OID_AUTO, earlyretransmit, CTLFLAG_RW, 166 &tcp_do_early_retransmit, 0, "Early retransmit"); 167 168 int tcp_aggregate_acks = 1; 169 SYSCTL_INT(_net_inet_tcp, OID_AUTO, aggregate_acks, CTLFLAG_RW, 170 &tcp_aggregate_acks, 0, "Aggregate built-up acks into one ack"); 171 172 static int tcp_do_eifel_detect = 1; 173 SYSCTL_INT(_net_inet_tcp, OID_AUTO, eifel, CTLFLAG_RW, 174 &tcp_do_eifel_detect, 0, "Eifel detection algorithm (RFC 3522)"); 175 176 static int tcp_do_abc = 1; 177 SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc, CTLFLAG_RW, 178 &tcp_do_abc, 0, 179 "TCP Appropriate Byte Counting (RFC 3465)"); 180 181 /* 182 * The following value actually takes range [25ms, 250ms], 183 * given that most modern systems use 1ms ~ 10ms as the unit 184 * of timestamp option. 185 */ 186 static u_int tcp_paws_tolerance = 25; 187 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, paws_tolerance, CTLFLAG_RW, 188 &tcp_paws_tolerance, 0, "RFC1323 PAWS tolerance"); 189 190 /* 191 * Define as tunable for easy testing with SACK on and off. 192 * Warning: do not change setting in the middle of an existing active TCP flow, 193 * else strange things might happen to that flow. 194 */ 195 int tcp_do_sack = 1; 196 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW, 197 &tcp_do_sack, 0, "Enable SACK Algorithms"); 198 199 int tcp_do_smartsack = 1; 200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, smartsack, CTLFLAG_RW, 201 &tcp_do_smartsack, 0, "Enable Smart SACK Algorithms"); 202 203 int tcp_do_rescuesack = 1; 204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rescuesack, CTLFLAG_RW, 205 &tcp_do_rescuesack, 0, "Rescue retransmission for SACK"); 206 207 int tcp_aggressive_rescuesack = 0; 208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rescuesack_agg, CTLFLAG_RW, 209 &tcp_aggressive_rescuesack, 0, "Aggressive rescue retransmission for SACK"); 210 211 static int tcp_force_sackrxt = 1; 212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, force_sackrxt, CTLFLAG_RW, 213 &tcp_force_sackrxt, 0, "Allowed forced SACK retransmit burst"); 214 215 int tcp_do_rfc6675 = 1; 216 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc6675, CTLFLAG_RW, 217 &tcp_do_rfc6675, 0, "Enable RFC6675"); 218 219 int tcp_rfc6675_rxt = 0; 220 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc6675_rxt, CTLFLAG_RW, 221 &tcp_rfc6675_rxt, 0, "Enable RFC6675 retransmit"); 222 223 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0, 224 "TCP Segment Reassembly Queue"); 225 226 int tcp_reass_maxseg = 0; 227 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RD, 228 &tcp_reass_maxseg, 0, 229 "Global maximum number of TCP Segments in Reassembly Queue"); 230 231 int tcp_reass_qsize = 0; 232 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, cursegments, CTLFLAG_RD, 233 &tcp_reass_qsize, 0, 234 "Global number of TCP Segments currently in Reassembly Queue"); 235 236 static int tcp_reass_overflows = 0; 237 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows, CTLFLAG_RD, 238 &tcp_reass_overflows, 0, 239 "Global number of TCP Segment Reassembly Queue Overflows"); 240 241 int tcp_do_autorcvbuf = 1; 242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_RW, 243 &tcp_do_autorcvbuf, 0, "Enable automatic receive buffer sizing"); 244 245 int tcp_autorcvbuf_inc = 16*1024; 246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_inc, CTLFLAG_RW, 247 &tcp_autorcvbuf_inc, 0, 248 "Incrementor step size of automatic receive buffer"); 249 250 int tcp_autorcvbuf_max = 2*1024*1024; 251 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_RW, 252 &tcp_autorcvbuf_max, 0, "Max size of automatic receive buffer"); 253 254 int tcp_sosend_agglim = 2; 255 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_agglim, CTLFLAG_RW, 256 &tcp_sosend_agglim, 0, "TCP sosend mbuf aggregation limit"); 257 258 int tcp_sosend_async = 1; 259 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_async, CTLFLAG_RW, 260 &tcp_sosend_async, 0, "TCP asynchronized pru_send"); 261 262 int tcp_sosend_jcluster = 1; 263 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_jcluster, CTLFLAG_RW, 264 &tcp_sosend_jcluster, 0, "TCP output uses jcluster"); 265 266 static int tcp_ignore_redun_dsack = 1; 267 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ignore_redun_dsack, CTLFLAG_RW, 268 &tcp_ignore_redun_dsack, 0, "Ignore redundant DSACK"); 269 270 static int tcp_reuseport_ext = 1; 271 SYSCTL_INT(_net_inet_tcp, OID_AUTO, reuseport_ext, CTLFLAG_RW, 272 &tcp_reuseport_ext, 0, "SO_REUSEPORT extension"); 273 274 static void tcp_dooptions(struct tcpopt *, u_char *, int, boolean_t, 275 tcp_seq); 276 static void tcp_pulloutofband(struct socket *, 277 struct tcphdr *, struct mbuf *, int); 278 static int tcp_reass(struct tcpcb *, struct tcphdr *, int *, 279 struct mbuf *); 280 static void tcp_xmit_timer(struct tcpcb *, int, tcp_seq); 281 static void tcp_newreno_partial_ack(struct tcpcb *, struct tcphdr *, int); 282 static void tcp_sack_rexmt(struct tcpcb *, boolean_t); 283 static boolean_t tcp_sack_limitedxmit(struct tcpcb *); 284 static int tcp_rmx_msl(const struct tcpcb *); 285 static void tcp_established(struct tcpcb *); 286 static boolean_t tcp_recv_dupack(struct tcpcb *, tcp_seq, u_int); 287 288 /* Neighbor Discovery, Neighbor Unreachability Detection Upper layer hint. */ 289 #ifdef INET6 290 #define ND6_HINT(tp) \ 291 do { \ 292 if ((tp) && (tp)->t_inpcb && \ 293 INP_ISIPV6((tp)->t_inpcb) && \ 294 (tp)->t_inpcb->in6p_route.ro_rt) \ 295 nd6_nud_hint((tp)->t_inpcb->in6p_route.ro_rt, NULL, 0); \ 296 } while (0) 297 #else 298 #define ND6_HINT(tp) 299 #endif 300 301 /* 302 * Indicate whether this ack should be delayed. We can delay the ack if 303 * - delayed acks are enabled and 304 * - there is no delayed ack timer in progress and 305 * - our last ack wasn't a 0-sized window. We never want to delay 306 * the ack that opens up a 0-sized window. 307 */ 308 #define DELAY_ACK(tp) \ 309 (tcp_delack_enabled && !tcp_callout_pending(tp, tp->tt_delack) && \ 310 !(tp->t_flags & TF_RXWIN0SENT)) 311 312 #define acceptable_window_update(tp, th, tiwin) \ 313 (SEQ_LT(tp->snd_wl1, th->th_seq) || \ 314 (tp->snd_wl1 == th->th_seq && \ 315 (SEQ_LT(tp->snd_wl2, th->th_ack) || \ 316 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)))) 317 318 #define iceildiv(n, d) (((n)+(d)-1) / (d)) 319 #define need_early_retransmit(tp, ownd) \ 320 (tcp_do_early_retransmit && \ 321 (tcp_do_eifel_detect && (tp->t_flags & TF_RCVD_TSTMP)) && \ 322 ownd < ((tp->t_rxtthresh + 1) * tp->t_maxseg) && \ 323 tp->t_dupacks + 1 >= iceildiv(ownd, tp->t_maxseg) && \ 324 (!TCP_DO_SACK(tp) || ownd <= tp->t_maxseg || \ 325 tcp_sack_has_sacked(&tp->scb, ownd - tp->t_maxseg))) 326 327 /* 328 * Returns TRUE, if this segment can be merged with the last 329 * pending segment in the reassemble queue and this segment 330 * does not overlap with the pending segment immediately 331 * preceeding the last pending segment. 332 */ 333 static __inline boolean_t 334 tcp_paws_canreasslast(const struct tcpcb *tp, const struct tcphdr *th, int tlen) 335 { 336 const struct tseg_qent *last, *prev; 337 338 last = TAILQ_LAST(&tp->t_segq, tsegqe_head); 339 if (last == NULL) 340 return FALSE; 341 342 /* This segment comes immediately after the last pending segment */ 343 if (last->tqe_th->th_seq + last->tqe_len == th->th_seq) { 344 if (last->tqe_th->th_flags & TH_FIN) { 345 /* No segments should follow segment w/ FIN */ 346 return FALSE; 347 } 348 return TRUE; 349 } 350 351 if (th->th_seq + tlen != last->tqe_th->th_seq) 352 return FALSE; 353 /* This segment comes immediately before the last pending segment */ 354 355 prev = TAILQ_PREV(last, tsegqe_head, tqe_q); 356 if (prev == NULL) { 357 /* 358 * No pending preceeding segment, we assume this segment 359 * could be reassembled. 360 */ 361 return TRUE; 362 } 363 364 /* This segment does not overlap with the preceeding segment */ 365 if (SEQ_GEQ(th->th_seq, prev->tqe_th->th_seq + prev->tqe_len)) 366 return TRUE; 367 368 return FALSE; 369 } 370 371 static __inline void 372 tcp_ncr_update_rxtthresh(struct tcpcb *tp) 373 { 374 int old_rxtthresh = tp->t_rxtthresh; 375 uint32_t ownd = tp->snd_max - tp->snd_una; 376 377 tp->t_rxtthresh = min(tcp_ncr_rxtthresh_max, 378 max(tcprexmtthresh, ((ownd / tp->t_maxseg) >> 1))); 379 if (tp->t_rxtthresh != old_rxtthresh) { 380 tcp_sack_update_lostseq(&tp->scb, tp->snd_una, 381 tp->t_maxseg, tp->t_rxtthresh); 382 } 383 } 384 385 static int 386 tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m) 387 { 388 struct tseg_qent *q; 389 struct tseg_qent *p = NULL; 390 struct tseg_qent *te; 391 struct socket *so = tp->t_inpcb->inp_socket; 392 int flags; 393 394 /* 395 * Call with th == NULL after become established to 396 * force pre-ESTABLISHED data up to user socket. 397 */ 398 if (th == NULL) 399 goto present; 400 401 /* 402 * Limit the number of segments in the reassembly queue to prevent 403 * holding on to too many segments (and thus running out of mbufs). 404 * Make sure to let the missing segment through which caused this 405 * queue. Always keep one global queue entry spare to be able to 406 * process the missing segment. 407 */ 408 if (th->th_seq != tp->rcv_nxt && 409 tcp_reass_qsize + 1 >= tcp_reass_maxseg) { 410 tcp_reass_overflows++; 411 tcpstat.tcps_rcvmemdrop++; 412 m_freem(m); 413 /* no SACK block to report */ 414 tp->reportblk.rblk_start = tp->reportblk.rblk_end; 415 return (0); 416 } 417 418 /* Allocate a new queue entry. */ 419 te = kmalloc(sizeof(struct tseg_qent), M_TSEGQ, M_INTWAIT | M_NULLOK); 420 if (te == NULL) { 421 tcpstat.tcps_rcvmemdrop++; 422 m_freem(m); 423 /* no SACK block to report */ 424 tp->reportblk.rblk_start = tp->reportblk.rblk_end; 425 return (0); 426 } 427 atomic_add_int(&tcp_reass_qsize, 1); 428 429 if (th->th_flags & TH_FIN) 430 tp->t_flags |= TF_QUEDFIN; 431 432 /* 433 * Find a segment which begins after this one does. 434 */ 435 TAILQ_FOREACH(q, &tp->t_segq, tqe_q) { 436 if (SEQ_GT(q->tqe_th->th_seq, th->th_seq)) 437 break; 438 p = q; 439 } 440 441 /* 442 * If there is a preceding segment, it may provide some of 443 * our data already. If so, drop the data from the incoming 444 * segment. If it provides all of our data, drop us. 445 */ 446 if (p != NULL) { 447 tcp_seq_diff_t i; 448 449 /* conversion to int (in i) handles seq wraparound */ 450 i = p->tqe_th->th_seq + p->tqe_len - th->th_seq; 451 if (i > 0) { /* overlaps preceding segment */ 452 tp->sack_flags |= 453 (TSACK_F_DUPSEG | TSACK_F_ENCLOSESEG); 454 /* enclosing block starts w/ preceding segment */ 455 tp->encloseblk.rblk_start = p->tqe_th->th_seq; 456 if (i >= *tlenp) { 457 if (th->th_flags & TH_FIN) 458 p->tqe_th->th_flags |= TH_FIN; 459 460 /* preceding encloses incoming segment */ 461 tp->encloseblk.rblk_end = TCP_SACK_BLKEND( 462 p->tqe_th->th_seq + p->tqe_len, 463 p->tqe_th->th_flags); 464 tcpstat.tcps_rcvduppack++; 465 tcpstat.tcps_rcvdupbyte += *tlenp; 466 m_freem(m); 467 kfree(te, M_TSEGQ); 468 atomic_add_int(&tcp_reass_qsize, -1); 469 /* 470 * Try to present any queued data 471 * at the left window edge to the user. 472 * This is needed after the 3-WHS 473 * completes. 474 */ 475 goto present; /* ??? */ 476 } 477 m_adj(m, i); 478 *tlenp -= i; 479 th->th_seq += i; 480 /* incoming segment end is enclosing block end */ 481 tp->encloseblk.rblk_end = TCP_SACK_BLKEND( 482 th->th_seq + *tlenp, th->th_flags); 483 /* trim end of reported D-SACK block */ 484 tp->reportblk.rblk_end = th->th_seq; 485 } 486 } 487 tcpstat.tcps_rcvoopack++; 488 tcpstat.tcps_rcvoobyte += *tlenp; 489 490 /* 491 * While we overlap succeeding segments trim them or, 492 * if they are completely covered, dequeue them. 493 */ 494 while (q) { 495 tcp_seq_diff_t i = (th->th_seq + *tlenp) - q->tqe_th->th_seq; 496 tcp_seq qend = q->tqe_th->th_seq + q->tqe_len; 497 tcp_seq qend_sack = TCP_SACK_BLKEND(qend, q->tqe_th->th_flags); 498 struct tseg_qent *nq; 499 500 if (i <= 0) 501 break; 502 if (!(tp->sack_flags & TSACK_F_DUPSEG)) { 503 /* first time through */ 504 tp->sack_flags |= (TSACK_F_DUPSEG | TSACK_F_ENCLOSESEG); 505 tp->encloseblk = tp->reportblk; 506 /* report trailing duplicate D-SACK segment */ 507 tp->reportblk.rblk_start = q->tqe_th->th_seq; 508 } 509 if ((tp->sack_flags & TSACK_F_ENCLOSESEG) && 510 SEQ_GT(qend_sack, tp->encloseblk.rblk_end)) { 511 /* extend enclosing block if one exists */ 512 tp->encloseblk.rblk_end = qend_sack; 513 } 514 if (i < q->tqe_len) { 515 q->tqe_th->th_seq += i; 516 q->tqe_len -= i; 517 m_adj(q->tqe_m, i); 518 break; 519 } 520 521 if (q->tqe_th->th_flags & TH_FIN) 522 th->th_flags |= TH_FIN; 523 524 nq = TAILQ_NEXT(q, tqe_q); 525 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 526 m_freem(q->tqe_m); 527 kfree(q, M_TSEGQ); 528 atomic_add_int(&tcp_reass_qsize, -1); 529 q = nq; 530 } 531 532 /* Insert the new segment queue entry into place. */ 533 te->tqe_m = m; 534 te->tqe_th = th; 535 te->tqe_len = *tlenp; 536 537 /* check if can coalesce with following segment */ 538 if (q != NULL && (th->th_seq + *tlenp == q->tqe_th->th_seq)) { 539 tcp_seq tend_sack; 540 541 te->tqe_len += q->tqe_len; 542 if (q->tqe_th->th_flags & TH_FIN) 543 te->tqe_th->th_flags |= TH_FIN; 544 tend_sack = TCP_SACK_BLKEND(te->tqe_th->th_seq + te->tqe_len, 545 te->tqe_th->th_flags); 546 547 m_cat(te->tqe_m, q->tqe_m); 548 tp->encloseblk.rblk_end = tend_sack; 549 /* 550 * When not reporting a duplicate segment, use 551 * the larger enclosing block as the SACK block. 552 */ 553 if (!(tp->sack_flags & TSACK_F_DUPSEG)) 554 tp->reportblk.rblk_end = tend_sack; 555 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 556 kfree(q, M_TSEGQ); 557 atomic_add_int(&tcp_reass_qsize, -1); 558 } 559 560 if (p == NULL) { 561 TAILQ_INSERT_HEAD(&tp->t_segq, te, tqe_q); 562 } else { 563 /* check if can coalesce with preceding segment */ 564 if (p->tqe_th->th_seq + p->tqe_len == th->th_seq) { 565 if (te->tqe_th->th_flags & TH_FIN) 566 p->tqe_th->th_flags |= TH_FIN; 567 p->tqe_len += te->tqe_len; 568 m_cat(p->tqe_m, te->tqe_m); 569 tp->encloseblk.rblk_start = p->tqe_th->th_seq; 570 /* 571 * When not reporting a duplicate segment, use 572 * the larger enclosing block as the SACK block. 573 */ 574 if (!(tp->sack_flags & TSACK_F_DUPSEG)) 575 tp->reportblk.rblk_start = p->tqe_th->th_seq; 576 kfree(te, M_TSEGQ); 577 atomic_add_int(&tcp_reass_qsize, -1); 578 } else { 579 TAILQ_INSERT_AFTER(&tp->t_segq, p, te, tqe_q); 580 } 581 } 582 583 present: 584 /* 585 * Present data to user, advancing rcv_nxt through 586 * completed sequence space. 587 */ 588 if (!TCPS_HAVEESTABLISHED(tp->t_state)) 589 return (0); 590 q = TAILQ_FIRST(&tp->t_segq); 591 if (q == NULL || q->tqe_th->th_seq != tp->rcv_nxt) 592 return (0); 593 tp->rcv_nxt += q->tqe_len; 594 if (!(tp->sack_flags & TSACK_F_DUPSEG)) { 595 /* no SACK block to report since ACK advanced */ 596 tp->reportblk.rblk_start = tp->reportblk.rblk_end; 597 } 598 /* no enclosing block to report since ACK advanced */ 599 tp->sack_flags &= ~TSACK_F_ENCLOSESEG; 600 flags = q->tqe_th->th_flags & TH_FIN; 601 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 602 KASSERT(TAILQ_EMPTY(&tp->t_segq) || 603 TAILQ_FIRST(&tp->t_segq)->tqe_th->th_seq != tp->rcv_nxt, 604 ("segment not coalesced")); 605 if (so->so_state & SS_CANTRCVMORE) { 606 m_freem(q->tqe_m); 607 } else { 608 lwkt_gettoken(&so->so_rcv.ssb_token); 609 ssb_appendstream(&so->so_rcv, q->tqe_m); 610 lwkt_reltoken(&so->so_rcv.ssb_token); 611 } 612 kfree(q, M_TSEGQ); 613 atomic_add_int(&tcp_reass_qsize, -1); 614 ND6_HINT(tp); 615 sorwakeup(so); 616 return (flags); 617 } 618 619 /* 620 * TCP input routine, follows pages 65-76 of the 621 * protocol specification dated September, 1981 very closely. 622 */ 623 #ifdef INET6 624 int 625 tcp6_input(struct mbuf **mp, int *offp, int proto) 626 { 627 struct mbuf *m = *mp; 628 struct in6_ifaddr *ia6; 629 630 IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), IPPROTO_DONE); 631 632 /* 633 * draft-itojun-ipv6-tcp-to-anycast 634 * better place to put this in? 635 */ 636 ia6 = ip6_getdstifaddr(m); 637 if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) { 638 icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR, 639 offsetof(struct ip6_hdr, ip6_dst)); 640 return (IPPROTO_DONE); 641 } 642 643 tcp_input(mp, offp, proto); 644 return (IPPROTO_DONE); 645 } 646 #endif 647 648 int 649 tcp_input(struct mbuf **mp, int *offp, int proto) 650 { 651 int off0; 652 struct tcphdr *th; 653 struct ip *ip = NULL; 654 struct ipovly *ipov; 655 struct inpcb *inp = NULL; 656 u_char *optp = NULL; 657 int optlen = 0; 658 int tlen, off; 659 int len = 0; 660 int drop_hdrlen; 661 struct tcpcb *tp = NULL; 662 int thflags; 663 struct socket *so = NULL; 664 int todrop, acked; 665 boolean_t ourfinisacked, needoutput = FALSE, delayed_dupack = FALSE; 666 tcp_seq th_dupack = 0; /* XXX gcc warning */ 667 u_int to_flags = 0; /* XXX gcc warning */ 668 u_long tiwin; 669 int recvwin; 670 struct tcpopt to; /* options in this segment */ 671 struct sockaddr_in *next_hop = NULL; 672 int rstreason; /* For badport_bandlim accounting purposes */ 673 int cpu; 674 struct ip6_hdr *ip6 = NULL; 675 struct mbuf *m; 676 #ifdef INET6 677 boolean_t isipv6; 678 #else 679 const boolean_t isipv6 = FALSE; 680 #endif 681 #ifdef TCPDEBUG 682 short ostate = 0; 683 #endif 684 685 off0 = *offp; 686 m = *mp; 687 *mp = NULL; 688 689 tcpstat.tcps_rcvtotal++; 690 691 if (m->m_pkthdr.fw_flags & IPFORWARD_MBUF_TAGGED) { 692 struct m_tag *mtag; 693 694 mtag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL); 695 KKASSERT(mtag != NULL); 696 next_hop = m_tag_data(mtag); 697 } 698 699 #ifdef INET6 700 isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? TRUE : FALSE; 701 #endif 702 703 if (isipv6) { 704 /* IP6_EXTHDR_CHECK() is already done at tcp6_input() */ 705 ip6 = mtod(m, struct ip6_hdr *); 706 tlen = (sizeof *ip6) + ntohs(ip6->ip6_plen) - off0; 707 if (in6_cksum(m, IPPROTO_TCP, off0, tlen)) { 708 tcpstat.tcps_rcvbadsum++; 709 goto drop; 710 } 711 th = (struct tcphdr *)((caddr_t)ip6 + off0); 712 713 /* 714 * Be proactive about unspecified IPv6 address in source. 715 * As we use all-zero to indicate unbounded/unconnected pcb, 716 * unspecified IPv6 address can be used to confuse us. 717 * 718 * Note that packets with unspecified IPv6 destination is 719 * already dropped in ip6_input. 720 */ 721 if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) { 722 /* XXX stat */ 723 goto drop; 724 } 725 } else { 726 /* 727 * Get IP and TCP header together in first mbuf. 728 * Note: IP leaves IP header in first mbuf. 729 */ 730 if (off0 > sizeof(struct ip)) { 731 ip_stripoptions(m); 732 off0 = sizeof(struct ip); 733 } 734 /* already checked and pulled up in ip_demux() */ 735 KASSERT(m->m_len >= sizeof(struct tcpiphdr), 736 ("TCP header not in one mbuf: m->m_len %d", m->m_len)); 737 ip = mtod(m, struct ip *); 738 ipov = (struct ipovly *)ip; 739 th = (struct tcphdr *)((caddr_t)ip + off0); 740 tlen = ip->ip_len; 741 742 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) { 743 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) 744 th->th_sum = m->m_pkthdr.csum_data; 745 else 746 th->th_sum = in_pseudo(ip->ip_src.s_addr, 747 ip->ip_dst.s_addr, 748 htonl(m->m_pkthdr.csum_data + 749 ip->ip_len + 750 IPPROTO_TCP)); 751 th->th_sum ^= 0xffff; 752 } else { 753 /* 754 * Checksum extended TCP header and data. 755 */ 756 len = sizeof(struct ip) + tlen; 757 bzero(ipov->ih_x1, sizeof ipov->ih_x1); 758 ipov->ih_len = (u_short)tlen; 759 ipov->ih_len = htons(ipov->ih_len); 760 th->th_sum = in_cksum(m, len); 761 } 762 if (th->th_sum) { 763 tcpstat.tcps_rcvbadsum++; 764 goto drop; 765 } 766 #ifdef INET6 767 /* Re-initialization for later version check */ 768 ip->ip_v = IPVERSION; 769 #endif 770 } 771 772 /* 773 * Check that TCP offset makes sense, 774 * pull out TCP options and adjust length. XXX 775 */ 776 off = th->th_off << 2; 777 /* already checked and pulled up in ip_demux() */ 778 KASSERT(off >= sizeof(struct tcphdr) && off <= tlen, 779 ("bad TCP data offset %d (tlen %d)", off, tlen)); 780 tlen -= off; /* tlen is used instead of ti->ti_len */ 781 if (off > sizeof(struct tcphdr)) { 782 if (isipv6) { 783 IP6_EXTHDR_CHECK(m, off0, off, IPPROTO_DONE); 784 ip6 = mtod(m, struct ip6_hdr *); 785 th = (struct tcphdr *)((caddr_t)ip6 + off0); 786 } else { 787 /* already pulled up in ip_demux() */ 788 KASSERT(m->m_len >= sizeof(struct ip) + off, 789 ("TCP header and options not in one mbuf: " 790 "m_len %d, off %d", m->m_len, off)); 791 } 792 optlen = off - sizeof(struct tcphdr); 793 optp = (u_char *)(th + 1); 794 } 795 thflags = th->th_flags; 796 797 #ifdef TCP_DROP_SYNFIN 798 /* 799 * If the drop_synfin option is enabled, drop all packets with 800 * both the SYN and FIN bits set. This prevents e.g. nmap from 801 * identifying the TCP/IP stack. 802 * 803 * This is a violation of the TCP specification. 804 */ 805 if (drop_synfin && (thflags & (TH_SYN | TH_FIN)) == (TH_SYN | TH_FIN)) 806 goto drop; 807 #endif 808 809 /* 810 * Convert TCP protocol specific fields to host format. 811 */ 812 th->th_seq = ntohl(th->th_seq); 813 th->th_ack = ntohl(th->th_ack); 814 th->th_win = ntohs(th->th_win); 815 th->th_urp = ntohs(th->th_urp); 816 817 /* 818 * Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options, 819 * until after ip6_savecontrol() is called and before other functions 820 * which don't want those proto headers. 821 * Because ip6_savecontrol() is going to parse the mbuf to 822 * search for data to be passed up to user-land, it wants mbuf 823 * parameters to be unchanged. 824 * XXX: the call of ip6_savecontrol() has been obsoleted based on 825 * latest version of the advanced API (20020110). 826 */ 827 drop_hdrlen = off0 + off; 828 829 /* 830 * Locate pcb for segment. 831 */ 832 findpcb: 833 /* IPFIREWALL_FORWARD section */ 834 if (next_hop != NULL && !isipv6) { /* IPv6 support is not there yet */ 835 /* 836 * Transparently forwarded. Pretend to be the destination. 837 * already got one like this? 838 */ 839 cpu = mycpu->gd_cpuid; 840 inp = in_pcblookup_hash(&tcbinfo[cpu], 841 ip->ip_src, th->th_sport, 842 ip->ip_dst, th->th_dport, 843 0, m->m_pkthdr.rcvif); 844 if (!inp) { 845 /* 846 * It's new. Try to find the ambushing socket. 847 */ 848 849 /* 850 * The rest of the ipfw code stores the port in 851 * host order. XXX 852 * (The IP address is still in network order.) 853 */ 854 in_port_t dport = next_hop->sin_port ? 855 htons(next_hop->sin_port) : 856 th->th_dport; 857 858 cpu = tcp_addrcpu(ip->ip_src.s_addr, th->th_sport, 859 next_hop->sin_addr.s_addr, dport); 860 inp = in_pcblookup_hash(&tcbinfo[cpu], 861 ip->ip_src, th->th_sport, 862 next_hop->sin_addr, dport, 863 1, m->m_pkthdr.rcvif); 864 } 865 } else { 866 if (isipv6) { 867 inp = in6_pcblookup_hash(&tcbinfo[0], 868 &ip6->ip6_src, th->th_sport, 869 &ip6->ip6_dst, th->th_dport, 870 1, m->m_pkthdr.rcvif); 871 } else { 872 cpu = mycpu->gd_cpuid; 873 inp = in_pcblookup_pkthash(&tcbinfo[cpu], 874 ip->ip_src, th->th_sport, 875 ip->ip_dst, th->th_dport, 876 1, m->m_pkthdr.rcvif, 877 tcp_reuseport_ext ? m : NULL); 878 } 879 } 880 881 /* 882 * If the state is CLOSED (i.e., TCB does not exist) then 883 * all data in the incoming segment is discarded. 884 * If the TCB exists but is in CLOSED state, it is embryonic, 885 * but should either do a listen or a connect soon. 886 */ 887 if (inp == NULL) { 888 if (log_in_vain) { 889 #ifdef INET6 890 char dbuf[INET6_ADDRSTRLEN+2], sbuf[INET6_ADDRSTRLEN+2]; 891 #else 892 char dbuf[sizeof "aaa.bbb.ccc.ddd"]; 893 char sbuf[sizeof "aaa.bbb.ccc.ddd"]; 894 #endif 895 if (isipv6) { 896 strcpy(dbuf, "["); 897 strcat(dbuf, ip6_sprintf(&ip6->ip6_dst)); 898 strcat(dbuf, "]"); 899 strcpy(sbuf, "["); 900 strcat(sbuf, ip6_sprintf(&ip6->ip6_src)); 901 strcat(sbuf, "]"); 902 } else { 903 strcpy(dbuf, inet_ntoa(ip->ip_dst)); 904 strcpy(sbuf, inet_ntoa(ip->ip_src)); 905 } 906 switch (log_in_vain) { 907 case 1: 908 if (!(thflags & TH_SYN)) 909 break; 910 case 2: 911 log(LOG_INFO, 912 "Connection attempt to TCP %s:%d " 913 "from %s:%d flags:0x%02x\n", 914 dbuf, ntohs(th->th_dport), sbuf, 915 ntohs(th->th_sport), thflags); 916 break; 917 default: 918 break; 919 } 920 } 921 if (blackhole) { 922 switch (blackhole) { 923 case 1: 924 if (thflags & TH_SYN) 925 goto drop; 926 break; 927 case 2: 928 goto drop; 929 default: 930 goto drop; 931 } 932 } 933 rstreason = BANDLIM_RST_CLOSEDPORT; 934 goto dropwithreset; 935 } 936 937 #ifdef IPSEC 938 if (isipv6) { 939 if (ipsec6_in_reject_so(m, inp->inp_socket)) { 940 ipsec6stat.in_polvio++; 941 goto drop; 942 } 943 } else { 944 if (ipsec4_in_reject_so(m, inp->inp_socket)) { 945 ipsecstat.in_polvio++; 946 goto drop; 947 } 948 } 949 #endif 950 #ifdef FAST_IPSEC 951 if (isipv6) { 952 if (ipsec6_in_reject(m, inp)) 953 goto drop; 954 } else { 955 if (ipsec4_in_reject(m, inp)) 956 goto drop; 957 } 958 #endif 959 /* Check the minimum TTL for socket. */ 960 #ifdef INET6 961 if ((isipv6 ? ip6->ip6_hlim : ip->ip_ttl) < inp->inp_ip_minttl) 962 goto drop; 963 #endif 964 965 tp = intotcpcb(inp); 966 KASSERT(tp != NULL, ("tcp_input: tp is NULL")); 967 if (tp->t_state <= TCPS_CLOSED) 968 goto drop; 969 970 so = inp->inp_socket; 971 972 #ifdef TCPDEBUG 973 if (so->so_options & SO_DEBUG) { 974 ostate = tp->t_state; 975 if (isipv6) 976 bcopy(ip6, tcp_saveipgen, sizeof(*ip6)); 977 else 978 bcopy(ip, tcp_saveipgen, sizeof(*ip)); 979 tcp_savetcp = *th; 980 } 981 #endif 982 983 bzero(&to, sizeof to); 984 985 if (so->so_options & SO_ACCEPTCONN) { 986 struct in_conninfo inc; 987 988 #ifdef INET6 989 inc.inc_isipv6 = (isipv6 == TRUE); 990 #endif 991 if (isipv6) { 992 inc.inc6_faddr = ip6->ip6_src; 993 inc.inc6_laddr = ip6->ip6_dst; 994 inc.inc6_route.ro_rt = NULL; /* XXX */ 995 } else { 996 inc.inc_faddr = ip->ip_src; 997 inc.inc_laddr = ip->ip_dst; 998 inc.inc_route.ro_rt = NULL; /* XXX */ 999 } 1000 inc.inc_fport = th->th_sport; 1001 inc.inc_lport = th->th_dport; 1002 1003 /* 1004 * If the state is LISTEN then ignore segment if it contains 1005 * a RST. If the segment contains an ACK then it is bad and 1006 * send a RST. If it does not contain a SYN then it is not 1007 * interesting; drop it. 1008 * 1009 * If the state is SYN_RECEIVED (syncache) and seg contains 1010 * an ACK, but not for our SYN/ACK, send a RST. If the seg 1011 * contains a RST, check the sequence number to see if it 1012 * is a valid reset segment. 1013 */ 1014 if ((thflags & (TH_RST | TH_ACK | TH_SYN)) != TH_SYN) { 1015 if ((thflags & (TH_RST | TH_ACK | TH_SYN)) == TH_ACK) { 1016 if (!syncache_expand(&inc, th, &so, m)) { 1017 /* 1018 * No syncache entry, or ACK was not 1019 * for our SYN/ACK. Send a RST. 1020 */ 1021 tcpstat.tcps_badsyn++; 1022 rstreason = BANDLIM_RST_OPENPORT; 1023 goto dropwithreset; 1024 } 1025 1026 /* 1027 * Could not complete 3-way handshake, 1028 * connection is being closed down, and 1029 * syncache will free mbuf. 1030 */ 1031 if (so == NULL) 1032 return(IPPROTO_DONE); 1033 1034 /* 1035 * We must be in the correct protocol thread 1036 * for this connection. 1037 */ 1038 KKASSERT(so->so_port == &curthread->td_msgport); 1039 1040 /* 1041 * Socket is created in state SYN_RECEIVED. 1042 * Continue processing segment. 1043 */ 1044 inp = so->so_pcb; 1045 tp = intotcpcb(inp); 1046 /* 1047 * This is what would have happened in 1048 * tcp_output() when the SYN,ACK was sent. 1049 */ 1050 tp->snd_up = tp->snd_una; 1051 tp->snd_max = tp->snd_nxt = tp->iss + 1; 1052 tp->last_ack_sent = tp->rcv_nxt; 1053 1054 goto after_listen; 1055 } 1056 if (thflags & TH_RST) { 1057 syncache_chkrst(&inc, th); 1058 goto drop; 1059 } 1060 if (thflags & TH_ACK) { 1061 syncache_badack(&inc); 1062 tcpstat.tcps_badsyn++; 1063 rstreason = BANDLIM_RST_OPENPORT; 1064 goto dropwithreset; 1065 } 1066 goto drop; 1067 } 1068 1069 /* 1070 * Segment's flags are (SYN) or (SYN | FIN). 1071 */ 1072 #ifdef INET6 1073 /* 1074 * If deprecated address is forbidden, 1075 * we do not accept SYN to deprecated interface 1076 * address to prevent any new inbound connection from 1077 * getting established. 1078 * When we do not accept SYN, we send a TCP RST, 1079 * with deprecated source address (instead of dropping 1080 * it). We compromise it as it is much better for peer 1081 * to send a RST, and RST will be the final packet 1082 * for the exchange. 1083 * 1084 * If we do not forbid deprecated addresses, we accept 1085 * the SYN packet. RFC2462 does not suggest dropping 1086 * SYN in this case. 1087 * If we decipher RFC2462 5.5.4, it says like this: 1088 * 1. use of deprecated addr with existing 1089 * communication is okay - "SHOULD continue to be 1090 * used" 1091 * 2. use of it with new communication: 1092 * (2a) "SHOULD NOT be used if alternate address 1093 * with sufficient scope is available" 1094 * (2b) nothing mentioned otherwise. 1095 * Here we fall into (2b) case as we have no choice in 1096 * our source address selection - we must obey the peer. 1097 * 1098 * The wording in RFC2462 is confusing, and there are 1099 * multiple description text for deprecated address 1100 * handling - worse, they are not exactly the same. 1101 * I believe 5.5.4 is the best one, so we follow 5.5.4. 1102 */ 1103 if (isipv6 && !ip6_use_deprecated) { 1104 struct in6_ifaddr *ia6; 1105 1106 if ((ia6 = ip6_getdstifaddr(m)) && 1107 (ia6->ia6_flags & IN6_IFF_DEPRECATED)) { 1108 tp = NULL; 1109 rstreason = BANDLIM_RST_OPENPORT; 1110 goto dropwithreset; 1111 } 1112 } 1113 #endif 1114 /* 1115 * If it is from this socket, drop it, it must be forged. 1116 * Don't bother responding if the destination was a broadcast. 1117 */ 1118 if (th->th_dport == th->th_sport) { 1119 if (isipv6) { 1120 if (IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst, 1121 &ip6->ip6_src)) 1122 goto drop; 1123 } else { 1124 if (ip->ip_dst.s_addr == ip->ip_src.s_addr) 1125 goto drop; 1126 } 1127 } 1128 /* 1129 * RFC1122 4.2.3.10, p. 104: discard bcast/mcast SYN 1130 * 1131 * Note that it is quite possible to receive unicast 1132 * link-layer packets with a broadcast IP address. Use 1133 * in_broadcast() to find them. 1134 */ 1135 if (m->m_flags & (M_BCAST | M_MCAST)) 1136 goto drop; 1137 if (isipv6) { 1138 if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) || 1139 IN6_IS_ADDR_MULTICAST(&ip6->ip6_src)) 1140 goto drop; 1141 } else { 1142 if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) || 1143 IN_MULTICAST(ntohl(ip->ip_src.s_addr)) || 1144 ip->ip_src.s_addr == htonl(INADDR_BROADCAST) || 1145 in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif)) 1146 goto drop; 1147 } 1148 /* 1149 * SYN appears to be valid; create compressed TCP state 1150 * for syncache, or perform t/tcp connection. 1151 */ 1152 if (so->so_qlen <= so->so_qlimit) { 1153 tcp_dooptions(&to, optp, optlen, TRUE, th->th_ack); 1154 if (!syncache_add(&inc, &to, th, so, m)) 1155 goto drop; 1156 1157 /* 1158 * Entry added to syncache, mbuf used to 1159 * send SYN,ACK packet. 1160 */ 1161 return(IPPROTO_DONE); 1162 } 1163 goto drop; 1164 } 1165 1166 after_listen: 1167 /* 1168 * Should not happen - syncache should pick up these connections. 1169 * 1170 * Once we are past handling listen sockets we must be in the 1171 * correct protocol processing thread. 1172 */ 1173 KASSERT(tp->t_state != TCPS_LISTEN, ("tcp_input: TCPS_LISTEN state")); 1174 KKASSERT(so->so_port == &curthread->td_msgport); 1175 1176 /* Unscale the window into a 32-bit value. */ 1177 if (!(thflags & TH_SYN)) 1178 tiwin = th->th_win << tp->snd_scale; 1179 else 1180 tiwin = th->th_win; 1181 1182 /* 1183 * This is the second part of the MSS DoS prevention code (after 1184 * minmss on the sending side) and it deals with too many too small 1185 * tcp packets in a too short timeframe (1 second). 1186 * 1187 * XXX Removed. This code was crap. It does not scale to network 1188 * speed, and default values break NFS. Gone. 1189 */ 1190 /* REMOVED */ 1191 1192 /* 1193 * Segment received on connection. 1194 * 1195 * Reset idle time and keep-alive timer. Don't waste time if less 1196 * then a second has elapsed. 1197 */ 1198 if ((int)(ticks - tp->t_rcvtime) > hz) 1199 tcp_timer_keep_activity(tp, thflags); 1200 1201 /* 1202 * Process options. 1203 * XXX this is tradtitional behavior, may need to be cleaned up. 1204 */ 1205 tcp_dooptions(&to, optp, optlen, (thflags & TH_SYN) != 0, th->th_ack); 1206 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 1207 if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) { 1208 tp->t_flags |= TF_RCVD_SCALE; 1209 tp->snd_scale = to.to_requested_s_scale; 1210 } 1211 1212 /* 1213 * Initial send window; will be updated upon next ACK 1214 */ 1215 tp->snd_wnd = th->th_win; 1216 1217 if (to.to_flags & TOF_TS) { 1218 tp->t_flags |= TF_RCVD_TSTMP; 1219 tp->ts_recent = to.to_tsval; 1220 tp->ts_recent_age = ticks; 1221 } 1222 if (!(to.to_flags & TOF_MSS)) 1223 to.to_mss = 0; 1224 tcp_mss(tp, to.to_mss); 1225 /* 1226 * Only set the TF_SACK_PERMITTED per-connection flag 1227 * if we got a SACK_PERMITTED option from the other side 1228 * and the global tcp_do_sack variable is true. 1229 */ 1230 if (tcp_do_sack && (to.to_flags & TOF_SACK_PERMITTED)) 1231 tp->t_flags |= TF_SACK_PERMITTED; 1232 } 1233 1234 /* 1235 * Header prediction: check for the two common cases 1236 * of a uni-directional data xfer. If the packet has 1237 * no control flags, is in-sequence, the window didn't 1238 * change and we're not retransmitting, it's a 1239 * candidate. If the length is zero and the ack moved 1240 * forward, we're the sender side of the xfer. Just 1241 * free the data acked & wake any higher level process 1242 * that was blocked waiting for space. If the length 1243 * is non-zero and the ack didn't move, we're the 1244 * receiver side. If we're getting packets in-order 1245 * (the reassembly queue is empty), add the data to 1246 * the socket buffer and note that we need a delayed ack. 1247 * Make sure that the hidden state-flags are also off. 1248 * Since we check for TCPS_ESTABLISHED above, it can only 1249 * be TH_NEEDSYN. 1250 */ 1251 if (tp->t_state == TCPS_ESTABLISHED && 1252 (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK && 1253 !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) && 1254 (!(to.to_flags & TOF_TS) || 1255 TSTMP_GEQ(to.to_tsval, tp->ts_recent)) && 1256 th->th_seq == tp->rcv_nxt && 1257 tp->snd_nxt == tp->snd_max) { 1258 1259 /* 1260 * If last ACK falls within this segment's sequence numbers, 1261 * record the timestamp. 1262 * NOTE that the test is modified according to the latest 1263 * proposal of the tcplw@cray.com list (Braden 1993/04/26). 1264 */ 1265 if ((to.to_flags & TOF_TS) && 1266 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 1267 tp->ts_recent_age = ticks; 1268 tp->ts_recent = to.to_tsval; 1269 } 1270 1271 if (tlen == 0) { 1272 if (SEQ_GT(th->th_ack, tp->snd_una) && 1273 SEQ_LEQ(th->th_ack, tp->snd_max) && 1274 tp->snd_cwnd >= tp->snd_wnd && 1275 !IN_FASTRECOVERY(tp)) { 1276 /* 1277 * This is a pure ack for outstanding data. 1278 */ 1279 ++tcpstat.tcps_predack; 1280 /* 1281 * "bad retransmit" recovery 1282 * 1283 * If Eifel detection applies, then 1284 * it is deterministic, so use it 1285 * unconditionally over the old heuristic. 1286 * Otherwise, fall back to the old heuristic. 1287 */ 1288 if (tcp_do_eifel_detect && 1289 (to.to_flags & TOF_TS) && to.to_tsecr && 1290 (tp->rxt_flags & TRXT_F_FIRSTACCACK)) { 1291 /* Eifel detection applicable. */ 1292 if (to.to_tsecr < tp->t_rexmtTS) { 1293 tcp_revert_congestion_state(tp); 1294 ++tcpstat.tcps_eifeldetected; 1295 if (tp->t_rxtshift != 1 || 1296 ticks >= tp->t_badrxtwin) 1297 ++tcpstat.tcps_rttcantdetect; 1298 } 1299 } else if (tp->t_rxtshift == 1 && 1300 ticks < tp->t_badrxtwin) { 1301 tcp_revert_congestion_state(tp); 1302 ++tcpstat.tcps_rttdetected; 1303 } 1304 tp->rxt_flags &= ~(TRXT_F_FIRSTACCACK | 1305 TRXT_F_FASTREXMT | TRXT_F_EARLYREXMT); 1306 /* 1307 * Recalculate the retransmit timer / rtt. 1308 * 1309 * Some machines (certain windows boxes) 1310 * send broken timestamp replies during the 1311 * SYN+ACK phase, ignore timestamps of 0. 1312 */ 1313 if ((to.to_flags & TOF_TS) && to.to_tsecr) { 1314 tcp_xmit_timer(tp, 1315 ticks - to.to_tsecr + 1, 1316 th->th_ack); 1317 } else if (tp->t_rtttime && 1318 SEQ_GT(th->th_ack, tp->t_rtseq)) { 1319 tcp_xmit_timer(tp, 1320 ticks - tp->t_rtttime + 1, 1321 th->th_ack); 1322 } 1323 tcp_xmit_bandwidth_limit(tp, th->th_ack); 1324 acked = th->th_ack - tp->snd_una; 1325 tcpstat.tcps_rcvackpack++; 1326 tcpstat.tcps_rcvackbyte += acked; 1327 sbdrop(&so->so_snd.sb, acked); 1328 tp->snd_recover = th->th_ack - 1; 1329 tp->snd_una = th->th_ack; 1330 tp->t_dupacks = 0; 1331 /* 1332 * Update window information. 1333 */ 1334 if (tiwin != tp->snd_wnd && 1335 acceptable_window_update(tp, th, tiwin)) { 1336 /* keep track of pure window updates */ 1337 if (tp->snd_wl2 == th->th_ack && 1338 tiwin > tp->snd_wnd) 1339 tcpstat.tcps_rcvwinupd++; 1340 tp->snd_wnd = tiwin; 1341 tp->snd_wl1 = th->th_seq; 1342 tp->snd_wl2 = th->th_ack; 1343 if (tp->snd_wnd > tp->max_sndwnd) 1344 tp->max_sndwnd = tp->snd_wnd; 1345 } 1346 m_freem(m); 1347 ND6_HINT(tp); /* some progress has been done */ 1348 /* 1349 * If all outstanding data are acked, stop 1350 * retransmit timer, otherwise restart timer 1351 * using current (possibly backed-off) value. 1352 * If process is waiting for space, 1353 * wakeup/selwakeup/signal. If data 1354 * are ready to send, let tcp_output 1355 * decide between more output or persist. 1356 */ 1357 if (tp->snd_una == tp->snd_max) { 1358 tcp_callout_stop(tp, tp->tt_rexmt); 1359 } else if (!tcp_callout_active(tp, 1360 tp->tt_persist)) { 1361 tcp_callout_reset(tp, tp->tt_rexmt, 1362 tp->t_rxtcur, tcp_timer_rexmt); 1363 } 1364 sowwakeup(so); 1365 if (so->so_snd.ssb_cc > 0 && 1366 !tcp_output_pending(tp)) 1367 tcp_output_fair(tp); 1368 return(IPPROTO_DONE); 1369 } 1370 } else if (tiwin == tp->snd_wnd && 1371 th->th_ack == tp->snd_una && 1372 TAILQ_EMPTY(&tp->t_segq) && 1373 tlen <= ssb_space(&so->so_rcv)) { 1374 u_long newsize = 0; /* automatic sockbuf scaling */ 1375 /* 1376 * This is a pure, in-sequence data packet 1377 * with nothing on the reassembly queue and 1378 * we have enough buffer space to take it. 1379 */ 1380 ++tcpstat.tcps_preddat; 1381 tp->rcv_nxt += tlen; 1382 tcpstat.tcps_rcvpack++; 1383 tcpstat.tcps_rcvbyte += tlen; 1384 ND6_HINT(tp); /* some progress has been done */ 1385 /* 1386 * Automatic sizing of receive socket buffer. Often the send 1387 * buffer size is not optimally adjusted to the actual network 1388 * conditions at hand (delay bandwidth product). Setting the 1389 * buffer size too small limits throughput on links with high 1390 * bandwidth and high delay (eg. trans-continental/oceanic links). 1391 * 1392 * On the receive side the socket buffer memory is only rarely 1393 * used to any significant extent. This allows us to be much 1394 * more aggressive in scaling the receive socket buffer. For 1395 * the case that the buffer space is actually used to a large 1396 * extent and we run out of kernel memory we can simply drop 1397 * the new segments; TCP on the sender will just retransmit it 1398 * later. Setting the buffer size too big may only consume too 1399 * much kernel memory if the application doesn't read() from 1400 * the socket or packet loss or reordering makes use of the 1401 * reassembly queue. 1402 * 1403 * The criteria to step up the receive buffer one notch are: 1404 * 1. the number of bytes received during the time it takes 1405 * one timestamp to be reflected back to us (the RTT); 1406 * 2. received bytes per RTT is within seven eighth of the 1407 * current socket buffer size; 1408 * 3. receive buffer size has not hit maximal automatic size; 1409 * 1410 * This algorithm does one step per RTT at most and only if 1411 * we receive a bulk stream w/o packet losses or reorderings. 1412 * Shrinking the buffer during idle times is not necessary as 1413 * it doesn't consume any memory when idle. 1414 * 1415 * TODO: Only step up if the application is actually serving 1416 * the buffer to better manage the socket buffer resources. 1417 */ 1418 if (tcp_do_autorcvbuf && 1419 to.to_tsecr && 1420 (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) { 1421 if (to.to_tsecr > tp->rfbuf_ts && 1422 to.to_tsecr - tp->rfbuf_ts < hz) { 1423 if (tp->rfbuf_cnt > 1424 (so->so_rcv.ssb_hiwat / 8 * 7) && 1425 so->so_rcv.ssb_hiwat < 1426 tcp_autorcvbuf_max) { 1427 newsize = 1428 ulmin(so->so_rcv.ssb_hiwat + 1429 tcp_autorcvbuf_inc, 1430 tcp_autorcvbuf_max); 1431 } 1432 /* Start over with next RTT. */ 1433 tp->rfbuf_ts = 0; 1434 tp->rfbuf_cnt = 0; 1435 } else 1436 tp->rfbuf_cnt += tlen; /* add up */ 1437 } 1438 /* 1439 * Add data to socket buffer. 1440 */ 1441 if (so->so_state & SS_CANTRCVMORE) { 1442 m_freem(m); 1443 } else { 1444 /* 1445 * Set new socket buffer size, give up when 1446 * limit is reached. 1447 * 1448 * Adjusting the size can mess up ACK 1449 * sequencing when pure window updates are 1450 * being avoided (which is the default), 1451 * so force an ack. 1452 */ 1453 lwkt_gettoken(&so->so_rcv.ssb_token); 1454 if (newsize) { 1455 tp->t_flags |= TF_RXRESIZED; 1456 if (!ssb_reserve(&so->so_rcv, newsize, 1457 so, NULL)) { 1458 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE); 1459 } 1460 if (newsize >= 1461 (TCP_MAXWIN << tp->rcv_scale)) { 1462 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE); 1463 } 1464 } 1465 m_adj(m, drop_hdrlen); /* delayed header drop */ 1466 ssb_appendstream(&so->so_rcv, m); 1467 lwkt_reltoken(&so->so_rcv.ssb_token); 1468 } 1469 sorwakeup(so); 1470 /* 1471 * This code is responsible for most of the ACKs 1472 * the TCP stack sends back after receiving a data 1473 * packet. Note that the DELAY_ACK check fails if 1474 * the delack timer is already running, which results 1475 * in an ack being sent every other packet (which is 1476 * what we want). 1477 * 1478 * We then further aggregate acks by not actually 1479 * sending one until the protocol thread has completed 1480 * processing the current backlog of packets. This 1481 * does not delay the ack any further, but allows us 1482 * to take advantage of the packet aggregation that 1483 * high speed NICs do (usually blocks of 8-10 packets) 1484 * to send a single ack rather then four or five acks, 1485 * greatly reducing the ack rate, the return channel 1486 * bandwidth, and the protocol overhead on both ends. 1487 * 1488 * Since this also has the effect of slowing down 1489 * the exponential slow-start ramp-up, systems with 1490 * very large bandwidth-delay products might want 1491 * to turn the feature off. 1492 */ 1493 if (DELAY_ACK(tp)) { 1494 tcp_callout_reset(tp, tp->tt_delack, 1495 tcp_delacktime, tcp_timer_delack); 1496 } else if (tcp_aggregate_acks) { 1497 tp->t_flags |= TF_ACKNOW; 1498 if (!(tp->t_flags & TF_ONOUTPUTQ)) { 1499 tp->t_flags |= TF_ONOUTPUTQ; 1500 tp->tt_cpu = mycpu->gd_cpuid; 1501 TAILQ_INSERT_TAIL( 1502 &tcpcbackq[tp->tt_cpu], 1503 tp, t_outputq); 1504 } 1505 } else { 1506 tp->t_flags |= TF_ACKNOW; 1507 tcp_output(tp); 1508 } 1509 return(IPPROTO_DONE); 1510 } 1511 } 1512 1513 /* 1514 * Calculate amount of space in receive window, 1515 * and then do TCP input processing. 1516 * Receive window is amount of space in rcv queue, 1517 * but not less than advertised window. 1518 */ 1519 recvwin = ssb_space(&so->so_rcv); 1520 if (recvwin < 0) 1521 recvwin = 0; 1522 tp->rcv_wnd = imax(recvwin, (int)(tp->rcv_adv - tp->rcv_nxt)); 1523 1524 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 1525 tp->rfbuf_ts = 0; 1526 tp->rfbuf_cnt = 0; 1527 1528 switch (tp->t_state) { 1529 /* 1530 * If the state is SYN_RECEIVED: 1531 * if seg contains an ACK, but not for our SYN/ACK, send a RST. 1532 */ 1533 case TCPS_SYN_RECEIVED: 1534 if ((thflags & TH_ACK) && 1535 (SEQ_LEQ(th->th_ack, tp->snd_una) || 1536 SEQ_GT(th->th_ack, tp->snd_max))) { 1537 rstreason = BANDLIM_RST_OPENPORT; 1538 goto dropwithreset; 1539 } 1540 break; 1541 1542 /* 1543 * If the state is SYN_SENT: 1544 * if seg contains an ACK, but not for our SYN, drop the input. 1545 * if seg contains a RST, then drop the connection. 1546 * if seg does not contain SYN, then drop it. 1547 * Otherwise this is an acceptable SYN segment 1548 * initialize tp->rcv_nxt and tp->irs 1549 * if seg contains ack then advance tp->snd_una 1550 * if SYN has been acked change to ESTABLISHED else SYN_RCVD state 1551 * arrange for segment to be acked (eventually) 1552 * continue processing rest of data/controls, beginning with URG 1553 */ 1554 case TCPS_SYN_SENT: 1555 if ((thflags & TH_ACK) && 1556 (SEQ_LEQ(th->th_ack, tp->iss) || 1557 SEQ_GT(th->th_ack, tp->snd_max))) { 1558 rstreason = BANDLIM_UNLIMITED; 1559 goto dropwithreset; 1560 } 1561 if (thflags & TH_RST) { 1562 if (thflags & TH_ACK) 1563 tp = tcp_drop(tp, ECONNREFUSED); 1564 goto drop; 1565 } 1566 if (!(thflags & TH_SYN)) 1567 goto drop; 1568 1569 tp->irs = th->th_seq; 1570 tcp_rcvseqinit(tp); 1571 if (thflags & TH_ACK) { 1572 /* Our SYN was acked. */ 1573 tcpstat.tcps_connects++; 1574 soisconnected(so); 1575 /* Do window scaling on this connection? */ 1576 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 1577 (TF_RCVD_SCALE | TF_REQ_SCALE)) 1578 tp->rcv_scale = tp->request_r_scale; 1579 tp->rcv_adv += tp->rcv_wnd; 1580 tp->snd_una++; /* SYN is acked */ 1581 tcp_callout_stop(tp, tp->tt_rexmt); 1582 /* 1583 * If there's data, delay ACK; if there's also a FIN 1584 * ACKNOW will be turned on later. 1585 */ 1586 if (DELAY_ACK(tp) && tlen != 0) { 1587 tcp_callout_reset(tp, tp->tt_delack, 1588 tcp_delacktime, tcp_timer_delack); 1589 } else { 1590 tp->t_flags |= TF_ACKNOW; 1591 } 1592 /* 1593 * Received <SYN,ACK> in SYN_SENT[*] state. 1594 * Transitions: 1595 * SYN_SENT --> ESTABLISHED 1596 * SYN_SENT* --> FIN_WAIT_1 1597 */ 1598 tp->t_starttime = ticks; 1599 if (tp->t_flags & TF_NEEDFIN) { 1600 TCP_STATE_CHANGE(tp, TCPS_FIN_WAIT_1); 1601 tp->t_flags &= ~TF_NEEDFIN; 1602 thflags &= ~TH_SYN; 1603 } else { 1604 tcp_established(tp); 1605 } 1606 } else { 1607 /* 1608 * Received initial SYN in SYN-SENT[*] state => 1609 * simultaneous open. 1610 * Do 3-way handshake: 1611 * SYN-SENT -> SYN-RECEIVED 1612 * SYN-SENT* -> SYN-RECEIVED* 1613 */ 1614 tp->t_flags |= TF_ACKNOW; 1615 tcp_callout_stop(tp, tp->tt_rexmt); 1616 TCP_STATE_CHANGE(tp, TCPS_SYN_RECEIVED); 1617 } 1618 1619 /* 1620 * Advance th->th_seq to correspond to first data byte. 1621 * If data, trim to stay within window, 1622 * dropping FIN if necessary. 1623 */ 1624 th->th_seq++; 1625 if (tlen > tp->rcv_wnd) { 1626 todrop = tlen - tp->rcv_wnd; 1627 m_adj(m, -todrop); 1628 tlen = tp->rcv_wnd; 1629 thflags &= ~TH_FIN; 1630 tcpstat.tcps_rcvpackafterwin++; 1631 tcpstat.tcps_rcvbyteafterwin += todrop; 1632 } 1633 tp->snd_wl1 = th->th_seq - 1; 1634 tp->rcv_up = th->th_seq; 1635 /* 1636 * Client side of transaction: already sent SYN and data. 1637 * If the remote host used T/TCP to validate the SYN, 1638 * our data will be ACK'd; if so, enter normal data segment 1639 * processing in the middle of step 5, ack processing. 1640 * Otherwise, goto step 6. 1641 */ 1642 if (thflags & TH_ACK) 1643 goto process_ACK; 1644 1645 goto step6; 1646 1647 /* 1648 * If the state is LAST_ACK or CLOSING or TIME_WAIT: 1649 * do normal processing (we no longer bother with T/TCP). 1650 */ 1651 case TCPS_LAST_ACK: 1652 case TCPS_CLOSING: 1653 case TCPS_TIME_WAIT: 1654 break; /* continue normal processing */ 1655 } 1656 1657 /* 1658 * States other than LISTEN or SYN_SENT. 1659 * First check the RST flag and sequence number since reset segments 1660 * are exempt from the timestamp and connection count tests. This 1661 * fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix 1662 * below which allowed reset segments in half the sequence space 1663 * to fall though and be processed (which gives forged reset 1664 * segments with a random sequence number a 50 percent chance of 1665 * killing a connection). 1666 * Then check timestamp, if present. 1667 * Then check the connection count, if present. 1668 * Then check that at least some bytes of segment are within 1669 * receive window. If segment begins before rcv_nxt, 1670 * drop leading data (and SYN); if nothing left, just ack. 1671 * 1672 * 1673 * If the RST bit is set, check the sequence number to see 1674 * if this is a valid reset segment. 1675 * RFC 793 page 37: 1676 * In all states except SYN-SENT, all reset (RST) segments 1677 * are validated by checking their SEQ-fields. A reset is 1678 * valid if its sequence number is in the window. 1679 * Note: this does not take into account delayed ACKs, so 1680 * we should test against last_ack_sent instead of rcv_nxt. 1681 * The sequence number in the reset segment is normally an 1682 * echo of our outgoing acknowledgement numbers, but some hosts 1683 * send a reset with the sequence number at the rightmost edge 1684 * of our receive window, and we have to handle this case. 1685 * If we have multiple segments in flight, the intial reset 1686 * segment sequence numbers will be to the left of last_ack_sent, 1687 * but they will eventually catch up. 1688 * In any case, it never made sense to trim reset segments to 1689 * fit the receive window since RFC 1122 says: 1690 * 4.2.2.12 RST Segment: RFC-793 Section 3.4 1691 * 1692 * A TCP SHOULD allow a received RST segment to include data. 1693 * 1694 * DISCUSSION 1695 * It has been suggested that a RST segment could contain 1696 * ASCII text that encoded and explained the cause of the 1697 * RST. No standard has yet been established for such 1698 * data. 1699 * 1700 * If the reset segment passes the sequence number test examine 1701 * the state: 1702 * SYN_RECEIVED STATE: 1703 * If passive open, return to LISTEN state. 1704 * If active open, inform user that connection was refused. 1705 * ESTABLISHED, FIN_WAIT_1, FIN_WAIT_2, CLOSE_WAIT STATES: 1706 * Inform user that connection was reset, and close tcb. 1707 * CLOSING, LAST_ACK STATES: 1708 * Close the tcb. 1709 * TIME_WAIT STATE: 1710 * Drop the segment - see Stevens, vol. 2, p. 964 and 1711 * RFC 1337. 1712 */ 1713 if (thflags & TH_RST) { 1714 if (SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 1715 SEQ_LEQ(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 1716 switch (tp->t_state) { 1717 1718 case TCPS_SYN_RECEIVED: 1719 so->so_error = ECONNREFUSED; 1720 goto close; 1721 1722 case TCPS_ESTABLISHED: 1723 case TCPS_FIN_WAIT_1: 1724 case TCPS_FIN_WAIT_2: 1725 case TCPS_CLOSE_WAIT: 1726 so->so_error = ECONNRESET; 1727 close: 1728 TCP_STATE_CHANGE(tp, TCPS_CLOSED); 1729 tcpstat.tcps_drops++; 1730 tp = tcp_close(tp); 1731 break; 1732 1733 case TCPS_CLOSING: 1734 case TCPS_LAST_ACK: 1735 tp = tcp_close(tp); 1736 break; 1737 1738 case TCPS_TIME_WAIT: 1739 break; 1740 } 1741 } 1742 goto drop; 1743 } 1744 1745 /* 1746 * RFC 1323 PAWS: If we have a timestamp reply on this segment 1747 * and it's less than ts_recent, drop it. 1748 */ 1749 if ((to.to_flags & TOF_TS) && tp->ts_recent != 0 && 1750 TSTMP_LT(to.to_tsval, tp->ts_recent)) { 1751 /* Check to see if ts_recent is over 24 days old. */ 1752 if ((int)(ticks - tp->ts_recent_age) > TCP_PAWS_IDLE) { 1753 /* 1754 * Invalidate ts_recent. If this segment updates 1755 * ts_recent, the age will be reset later and ts_recent 1756 * will get a valid value. If it does not, setting 1757 * ts_recent to zero will at least satisfy the 1758 * requirement that zero be placed in the timestamp 1759 * echo reply when ts_recent isn't valid. The 1760 * age isn't reset until we get a valid ts_recent 1761 * because we don't want out-of-order segments to be 1762 * dropped when ts_recent is old. 1763 */ 1764 tp->ts_recent = 0; 1765 } else if (tcp_paws_tolerance && tlen != 0 && 1766 tp->t_state == TCPS_ESTABLISHED && 1767 (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK&& 1768 !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) && 1769 th->th_ack == tp->snd_una && 1770 tiwin == tp->snd_wnd && 1771 TSTMP_GEQ(to.to_tsval + tcp_paws_tolerance, tp->ts_recent)&& 1772 (th->th_seq == tp->rcv_nxt || 1773 (SEQ_GT(th->th_seq, tp->rcv_nxt) && 1774 tcp_paws_canreasslast(tp, th, tlen)))) { 1775 /* 1776 * This tends to prevent valid new segments from being 1777 * dropped by the reordered segments sent by the fast 1778 * retransmission algorithm on the sending side, i.e. 1779 * the fast retransmitted segment w/ larger timestamp 1780 * arrives earlier than the previously sent new segments 1781 * w/ smaller timestamp. 1782 * 1783 * If following conditions are met, the segment is 1784 * accepted: 1785 * - The segment contains data 1786 * - The connection is established 1787 * - The header does not contain important flags 1788 * - SYN or FIN is not needed 1789 * - It does not acknowledge new data 1790 * - Receive window is not changed 1791 * - The timestamp is within "acceptable" range 1792 * - The new segment is what we are expecting or 1793 * the new segment could be merged w/ the last 1794 * pending segment on the reassemble queue 1795 */ 1796 tcpstat.tcps_pawsaccept++; 1797 tcpstat.tcps_pawsdrop++; 1798 } else { 1799 tcpstat.tcps_rcvduppack++; 1800 tcpstat.tcps_rcvdupbyte += tlen; 1801 tcpstat.tcps_pawsdrop++; 1802 if (tlen) 1803 goto dropafterack; 1804 goto drop; 1805 } 1806 } 1807 1808 /* 1809 * In the SYN-RECEIVED state, validate that the packet belongs to 1810 * this connection before trimming the data to fit the receive 1811 * window. Check the sequence number versus IRS since we know 1812 * the sequence numbers haven't wrapped. This is a partial fix 1813 * for the "LAND" DoS attack. 1814 */ 1815 if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) { 1816 rstreason = BANDLIM_RST_OPENPORT; 1817 goto dropwithreset; 1818 } 1819 1820 todrop = tp->rcv_nxt - th->th_seq; 1821 if (todrop > 0) { 1822 if (TCP_DO_SACK(tp)) { 1823 /* Report duplicate segment at head of packet. */ 1824 tp->reportblk.rblk_start = th->th_seq; 1825 tp->reportblk.rblk_end = TCP_SACK_BLKEND( 1826 th->th_seq + tlen, thflags); 1827 if (SEQ_GT(tp->reportblk.rblk_end, tp->rcv_nxt)) 1828 tp->reportblk.rblk_end = tp->rcv_nxt; 1829 tp->sack_flags |= (TSACK_F_DUPSEG | TSACK_F_SACKLEFT); 1830 tp->t_flags |= TF_ACKNOW; 1831 } 1832 if (thflags & TH_SYN) { 1833 thflags &= ~TH_SYN; 1834 th->th_seq++; 1835 if (th->th_urp > 1) 1836 th->th_urp--; 1837 else 1838 thflags &= ~TH_URG; 1839 todrop--; 1840 } 1841 /* 1842 * Following if statement from Stevens, vol. 2, p. 960. 1843 */ 1844 if (todrop > tlen || 1845 (todrop == tlen && !(thflags & TH_FIN))) { 1846 /* 1847 * Any valid FIN must be to the left of the window. 1848 * At this point the FIN must be a duplicate or out 1849 * of sequence; drop it. 1850 */ 1851 thflags &= ~TH_FIN; 1852 1853 /* 1854 * Send an ACK to resynchronize and drop any data. 1855 * But keep on processing for RST or ACK. 1856 */ 1857 tp->t_flags |= TF_ACKNOW; 1858 todrop = tlen; 1859 tcpstat.tcps_rcvduppack++; 1860 tcpstat.tcps_rcvdupbyte += todrop; 1861 } else { 1862 tcpstat.tcps_rcvpartduppack++; 1863 tcpstat.tcps_rcvpartdupbyte += todrop; 1864 } 1865 drop_hdrlen += todrop; /* drop from the top afterwards */ 1866 th->th_seq += todrop; 1867 tlen -= todrop; 1868 if (th->th_urp > todrop) 1869 th->th_urp -= todrop; 1870 else { 1871 thflags &= ~TH_URG; 1872 th->th_urp = 0; 1873 } 1874 } 1875 1876 /* 1877 * If new data are received on a connection after the 1878 * user processes are gone, then RST the other end. 1879 */ 1880 if ((so->so_state & SS_NOFDREF) && 1881 tp->t_state > TCPS_CLOSE_WAIT && tlen) { 1882 tp = tcp_close(tp); 1883 tcpstat.tcps_rcvafterclose++; 1884 rstreason = BANDLIM_UNLIMITED; 1885 goto dropwithreset; 1886 } 1887 1888 /* 1889 * If segment ends after window, drop trailing data 1890 * (and PUSH and FIN); if nothing left, just ACK. 1891 */ 1892 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 1893 if (todrop > 0) { 1894 tcpstat.tcps_rcvpackafterwin++; 1895 if (todrop >= tlen) { 1896 tcpstat.tcps_rcvbyteafterwin += tlen; 1897 /* 1898 * If a new connection request is received 1899 * while in TIME_WAIT, drop the old connection 1900 * and start over if the sequence numbers 1901 * are above the previous ones. 1902 */ 1903 if (thflags & TH_SYN && 1904 tp->t_state == TCPS_TIME_WAIT && 1905 SEQ_GT(th->th_seq, tp->rcv_nxt)) { 1906 tp = tcp_close(tp); 1907 goto findpcb; 1908 } 1909 /* 1910 * If window is closed can only take segments at 1911 * window edge, and have to drop data and PUSH from 1912 * incoming segments. Continue processing, but 1913 * remember to ack. Otherwise, drop segment 1914 * and ack. 1915 */ 1916 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 1917 tp->t_flags |= TF_ACKNOW; 1918 tcpstat.tcps_rcvwinprobe++; 1919 } else 1920 goto dropafterack; 1921 } else 1922 tcpstat.tcps_rcvbyteafterwin += todrop; 1923 m_adj(m, -todrop); 1924 tlen -= todrop; 1925 thflags &= ~(TH_PUSH | TH_FIN); 1926 } 1927 1928 /* 1929 * If last ACK falls within this segment's sequence numbers, 1930 * record its timestamp. 1931 * NOTE: 1932 * 1) That the test incorporates suggestions from the latest 1933 * proposal of the tcplw@cray.com list (Braden 1993/04/26). 1934 * 2) That updating only on newer timestamps interferes with 1935 * our earlier PAWS tests, so this check should be solely 1936 * predicated on the sequence space of this segment. 1937 * 3) That we modify the segment boundary check to be 1938 * Last.ACK.Sent <= SEG.SEQ + SEG.LEN 1939 * instead of RFC1323's 1940 * Last.ACK.Sent < SEG.SEQ + SEG.LEN, 1941 * This modified check allows us to overcome RFC1323's 1942 * limitations as described in Stevens TCP/IP Illustrated 1943 * Vol. 2 p.869. In such cases, we can still calculate the 1944 * RTT correctly when RCV.NXT == Last.ACK.Sent. 1945 */ 1946 if ((to.to_flags & TOF_TS) && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 1947 SEQ_LEQ(tp->last_ack_sent, (th->th_seq + tlen 1948 + ((thflags & TH_SYN) != 0) 1949 + ((thflags & TH_FIN) != 0)))) { 1950 tp->ts_recent_age = ticks; 1951 tp->ts_recent = to.to_tsval; 1952 } 1953 1954 /* 1955 * If a SYN is in the window, then this is an 1956 * error and we send an RST and drop the connection. 1957 */ 1958 if (thflags & TH_SYN) { 1959 tp = tcp_drop(tp, ECONNRESET); 1960 rstreason = BANDLIM_UNLIMITED; 1961 goto dropwithreset; 1962 } 1963 1964 /* 1965 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN 1966 * flag is on (half-synchronized state), then queue data for 1967 * later processing; else drop segment and return. 1968 */ 1969 if (!(thflags & TH_ACK)) { 1970 if (tp->t_state == TCPS_SYN_RECEIVED || 1971 (tp->t_flags & TF_NEEDSYN)) 1972 goto step6; 1973 else 1974 goto drop; 1975 } 1976 1977 /* 1978 * Ack processing. 1979 */ 1980 switch (tp->t_state) { 1981 /* 1982 * In SYN_RECEIVED state, the ACK acknowledges our SYN, so enter 1983 * ESTABLISHED state and continue processing. 1984 * The ACK was checked above. 1985 */ 1986 case TCPS_SYN_RECEIVED: 1987 1988 tcpstat.tcps_connects++; 1989 soisconnected(so); 1990 /* Do window scaling? */ 1991 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 1992 (TF_RCVD_SCALE | TF_REQ_SCALE)) 1993 tp->rcv_scale = tp->request_r_scale; 1994 /* 1995 * Make transitions: 1996 * SYN-RECEIVED -> ESTABLISHED 1997 * SYN-RECEIVED* -> FIN-WAIT-1 1998 */ 1999 tp->t_starttime = ticks; 2000 if (tp->t_flags & TF_NEEDFIN) { 2001 TCP_STATE_CHANGE(tp, TCPS_FIN_WAIT_1); 2002 tp->t_flags &= ~TF_NEEDFIN; 2003 } else { 2004 tcp_established(tp); 2005 } 2006 /* 2007 * If segment contains data or ACK, will call tcp_reass() 2008 * later; if not, do so now to pass queued data to user. 2009 */ 2010 if (tlen == 0 && !(thflags & TH_FIN)) 2011 tcp_reass(tp, NULL, NULL, NULL); 2012 /* fall into ... */ 2013 2014 /* 2015 * In ESTABLISHED state: drop duplicate ACKs; ACK out of range 2016 * ACKs. If the ack is in the range 2017 * tp->snd_una < th->th_ack <= tp->snd_max 2018 * then advance tp->snd_una to th->th_ack and drop 2019 * data from the retransmission queue. If this ACK reflects 2020 * more up to date window information we update our window information. 2021 */ 2022 case TCPS_ESTABLISHED: 2023 case TCPS_FIN_WAIT_1: 2024 case TCPS_FIN_WAIT_2: 2025 case TCPS_CLOSE_WAIT: 2026 case TCPS_CLOSING: 2027 case TCPS_LAST_ACK: 2028 case TCPS_TIME_WAIT: 2029 2030 if (SEQ_LEQ(th->th_ack, tp->snd_una)) { 2031 boolean_t maynotdup = FALSE; 2032 2033 if (TCP_DO_SACK(tp)) 2034 tcp_sack_update_scoreboard(tp, &to); 2035 2036 if (tlen != 0 || tiwin != tp->snd_wnd || 2037 ((thflags & TH_FIN) && !(tp->t_flags & TF_SAWFIN))) 2038 maynotdup = TRUE; 2039 2040 if (!tcp_callout_active(tp, tp->tt_rexmt) || 2041 th->th_ack != tp->snd_una) { 2042 if (!maynotdup) 2043 tcpstat.tcps_rcvdupack++; 2044 tp->t_dupacks = 0; 2045 break; 2046 } 2047 2048 #define DELAY_DUPACK \ 2049 do { \ 2050 delayed_dupack = TRUE; \ 2051 th_dupack = th->th_ack; \ 2052 to_flags = to.to_flags; \ 2053 } while (0) 2054 if (maynotdup) { 2055 if (!tcp_do_rfc6675 || 2056 !TCP_DO_SACK(tp) || 2057 (to.to_flags & 2058 (TOF_SACK | TOF_SACK_REDUNDANT)) 2059 != TOF_SACK) { 2060 tp->t_dupacks = 0; 2061 } else { 2062 DELAY_DUPACK; 2063 } 2064 break; 2065 } 2066 if ((thflags & TH_FIN) && !(tp->t_flags & TF_QUEDFIN)) { 2067 /* 2068 * This could happen, if the reassemable 2069 * queue overflew or was drained. Don't 2070 * drop this FIN here; defer the duplicated 2071 * ACK processing until this FIN gets queued. 2072 */ 2073 DELAY_DUPACK; 2074 break; 2075 } 2076 #undef DELAY_DUPACK 2077 2078 if (tcp_recv_dupack(tp, th->th_ack, to.to_flags)) 2079 goto drop; 2080 else 2081 break; 2082 } 2083 2084 KASSERT(SEQ_GT(th->th_ack, tp->snd_una), ("th_ack <= snd_una")); 2085 tp->t_dupacks = 0; 2086 if (SEQ_GT(th->th_ack, tp->snd_max)) { 2087 /* 2088 * Detected optimistic ACK attack. 2089 * Force slow-start to de-synchronize attack. 2090 */ 2091 tp->snd_cwnd = tp->t_maxseg; 2092 tp->snd_wacked = 0; 2093 2094 tcpstat.tcps_rcvacktoomuch++; 2095 goto dropafterack; 2096 } 2097 /* 2098 * If we reach this point, ACK is not a duplicate, 2099 * i.e., it ACKs something we sent. 2100 */ 2101 if (tp->t_flags & TF_NEEDSYN) { 2102 /* 2103 * T/TCP: Connection was half-synchronized, and our 2104 * SYN has been ACK'd (so connection is now fully 2105 * synchronized). Go to non-starred state, 2106 * increment snd_una for ACK of SYN, and check if 2107 * we can do window scaling. 2108 */ 2109 tp->t_flags &= ~TF_NEEDSYN; 2110 tp->snd_una++; 2111 /* Do window scaling? */ 2112 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 2113 (TF_RCVD_SCALE | TF_REQ_SCALE)) 2114 tp->rcv_scale = tp->request_r_scale; 2115 } 2116 2117 process_ACK: 2118 acked = th->th_ack - tp->snd_una; 2119 tcpstat.tcps_rcvackpack++; 2120 tcpstat.tcps_rcvackbyte += acked; 2121 2122 if (tcp_do_eifel_detect && acked > 0 && 2123 (to.to_flags & TOF_TS) && (to.to_tsecr != 0) && 2124 (tp->rxt_flags & TRXT_F_FIRSTACCACK)) { 2125 /* Eifel detection applicable. */ 2126 if (to.to_tsecr < tp->t_rexmtTS) { 2127 ++tcpstat.tcps_eifeldetected; 2128 tcp_revert_congestion_state(tp); 2129 if (tp->t_rxtshift != 1 || 2130 ticks >= tp->t_badrxtwin) 2131 ++tcpstat.tcps_rttcantdetect; 2132 } 2133 } else if (tp->t_rxtshift == 1 && ticks < tp->t_badrxtwin) { 2134 /* 2135 * If we just performed our first retransmit, 2136 * and the ACK arrives within our recovery window, 2137 * then it was a mistake to do the retransmit 2138 * in the first place. Recover our original cwnd 2139 * and ssthresh, and proceed to transmit where we 2140 * left off. 2141 */ 2142 tcp_revert_congestion_state(tp); 2143 ++tcpstat.tcps_rttdetected; 2144 } 2145 2146 /* 2147 * If we have a timestamp reply, update smoothed 2148 * round trip time. If no timestamp is present but 2149 * transmit timer is running and timed sequence 2150 * number was acked, update smoothed round trip time. 2151 * Since we now have an rtt measurement, cancel the 2152 * timer backoff (cf., Phil Karn's retransmit alg.). 2153 * Recompute the initial retransmit timer. 2154 * 2155 * Some machines (certain windows boxes) send broken 2156 * timestamp replies during the SYN+ACK phase, ignore 2157 * timestamps of 0. 2158 */ 2159 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) 2160 tcp_xmit_timer(tp, ticks - to.to_tsecr + 1, 2161 th->th_ack); 2162 else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq)) 2163 tcp_xmit_timer(tp, ticks - tp->t_rtttime + 1, 2164 th->th_ack); 2165 tcp_xmit_bandwidth_limit(tp, th->th_ack); 2166 2167 /* 2168 * If no data (only SYN) was ACK'd, 2169 * skip rest of ACK processing. 2170 */ 2171 if (acked == 0) 2172 goto step6; 2173 2174 /* Stop looking for an acceptable ACK since one was received. */ 2175 tp->rxt_flags &= ~(TRXT_F_FIRSTACCACK | 2176 TRXT_F_FASTREXMT | TRXT_F_EARLYREXMT); 2177 2178 if (acked > so->so_snd.ssb_cc) { 2179 tp->snd_wnd -= so->so_snd.ssb_cc; 2180 sbdrop(&so->so_snd.sb, (int)so->so_snd.ssb_cc); 2181 ourfinisacked = TRUE; 2182 } else { 2183 sbdrop(&so->so_snd.sb, acked); 2184 tp->snd_wnd -= acked; 2185 ourfinisacked = FALSE; 2186 } 2187 sowwakeup(so); 2188 2189 /* 2190 * Update window information. 2191 */ 2192 if (acceptable_window_update(tp, th, tiwin)) { 2193 /* keep track of pure window updates */ 2194 if (tlen == 0 && tp->snd_wl2 == th->th_ack && 2195 tiwin > tp->snd_wnd) 2196 tcpstat.tcps_rcvwinupd++; 2197 tp->snd_wnd = tiwin; 2198 tp->snd_wl1 = th->th_seq; 2199 tp->snd_wl2 = th->th_ack; 2200 if (tp->snd_wnd > tp->max_sndwnd) 2201 tp->max_sndwnd = tp->snd_wnd; 2202 needoutput = TRUE; 2203 } 2204 2205 tp->snd_una = th->th_ack; 2206 if (TCP_DO_SACK(tp)) 2207 tcp_sack_update_scoreboard(tp, &to); 2208 if (IN_FASTRECOVERY(tp)) { 2209 if (SEQ_GEQ(th->th_ack, tp->snd_recover)) { 2210 EXIT_FASTRECOVERY(tp); 2211 needoutput = TRUE; 2212 /* 2213 * If the congestion window was inflated 2214 * to account for the other side's 2215 * cached packets, retract it. 2216 */ 2217 if (!TCP_DO_SACK(tp)) 2218 tp->snd_cwnd = tp->snd_ssthresh; 2219 2220 /* 2221 * Window inflation should have left us 2222 * with approximately snd_ssthresh outstanding 2223 * data. But, in case we would be inclined 2224 * to send a burst, better do it using 2225 * slow start. 2226 */ 2227 if (SEQ_GT(th->th_ack + tp->snd_cwnd, 2228 tp->snd_max + 2 * tp->t_maxseg)) 2229 tp->snd_cwnd = 2230 (tp->snd_max - tp->snd_una) + 2231 2 * tp->t_maxseg; 2232 2233 tp->snd_wacked = 0; 2234 } else { 2235 if (TCP_DO_SACK(tp)) { 2236 tp->snd_max_rexmt = tp->snd_max; 2237 tcp_sack_rexmt(tp, 2238 tp->snd_una == tp->rexmt_high); 2239 } else { 2240 tcp_newreno_partial_ack(tp, th, acked); 2241 } 2242 needoutput = FALSE; 2243 } 2244 } else { 2245 /* 2246 * Open the congestion window. When in slow-start, 2247 * open exponentially: maxseg per packet. Otherwise, 2248 * open linearly: maxseg per window. 2249 */ 2250 if (tp->snd_cwnd <= tp->snd_ssthresh) { 2251 u_int abc_sslimit = 2252 (SEQ_LT(tp->snd_nxt, tp->snd_max) ? 2253 tp->t_maxseg : 2 * tp->t_maxseg); 2254 2255 /* slow-start */ 2256 tp->snd_cwnd += tcp_do_abc ? 2257 min(acked, abc_sslimit) : tp->t_maxseg; 2258 } else { 2259 /* linear increase */ 2260 tp->snd_wacked += tcp_do_abc ? acked : 2261 tp->t_maxseg; 2262 if (tp->snd_wacked >= tp->snd_cwnd) { 2263 tp->snd_wacked -= tp->snd_cwnd; 2264 tp->snd_cwnd += tp->t_maxseg; 2265 } 2266 } 2267 tp->snd_cwnd = min(tp->snd_cwnd, 2268 TCP_MAXWIN << tp->snd_scale); 2269 tp->snd_recover = th->th_ack - 1; 2270 } 2271 if (SEQ_LT(tp->snd_nxt, tp->snd_una)) 2272 tp->snd_nxt = tp->snd_una; 2273 2274 /* 2275 * If all outstanding data is acked, stop retransmit 2276 * timer and remember to restart (more output or persist). 2277 * If there is more data to be acked, restart retransmit 2278 * timer, using current (possibly backed-off) value. 2279 */ 2280 if (th->th_ack == tp->snd_max) { 2281 tcp_callout_stop(tp, tp->tt_rexmt); 2282 needoutput = TRUE; 2283 } else if (!tcp_callout_active(tp, tp->tt_persist)) { 2284 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur, 2285 tcp_timer_rexmt); 2286 } 2287 2288 switch (tp->t_state) { 2289 /* 2290 * In FIN_WAIT_1 STATE in addition to the processing 2291 * for the ESTABLISHED state if our FIN is now acknowledged 2292 * then enter FIN_WAIT_2. 2293 */ 2294 case TCPS_FIN_WAIT_1: 2295 if (ourfinisacked) { 2296 /* 2297 * If we can't receive any more 2298 * data, then closing user can proceed. 2299 * Starting the timer is contrary to the 2300 * specification, but if we don't get a FIN 2301 * we'll hang forever. 2302 */ 2303 if (so->so_state & SS_CANTRCVMORE) { 2304 soisdisconnected(so); 2305 tcp_callout_reset(tp, tp->tt_2msl, 2306 tp->t_maxidle, tcp_timer_2msl); 2307 } 2308 TCP_STATE_CHANGE(tp, TCPS_FIN_WAIT_2); 2309 } 2310 break; 2311 2312 /* 2313 * In CLOSING STATE in addition to the processing for 2314 * the ESTABLISHED state if the ACK acknowledges our FIN 2315 * then enter the TIME-WAIT state, otherwise ignore 2316 * the segment. 2317 */ 2318 case TCPS_CLOSING: 2319 if (ourfinisacked) { 2320 TCP_STATE_CHANGE(tp, TCPS_TIME_WAIT); 2321 tcp_canceltimers(tp); 2322 tcp_callout_reset(tp, tp->tt_2msl, 2323 2 * tcp_rmx_msl(tp), 2324 tcp_timer_2msl); 2325 soisdisconnected(so); 2326 } 2327 break; 2328 2329 /* 2330 * In LAST_ACK, we may still be waiting for data to drain 2331 * and/or to be acked, as well as for the ack of our FIN. 2332 * If our FIN is now acknowledged, delete the TCB, 2333 * enter the closed state and return. 2334 */ 2335 case TCPS_LAST_ACK: 2336 if (ourfinisacked) { 2337 tp = tcp_close(tp); 2338 goto drop; 2339 } 2340 break; 2341 2342 /* 2343 * In TIME_WAIT state the only thing that should arrive 2344 * is a retransmission of the remote FIN. Acknowledge 2345 * it and restart the finack timer. 2346 */ 2347 case TCPS_TIME_WAIT: 2348 tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp), 2349 tcp_timer_2msl); 2350 goto dropafterack; 2351 } 2352 } 2353 2354 step6: 2355 /* 2356 * Update window information. 2357 * Don't look at window if no ACK: TAC's send garbage on first SYN. 2358 */ 2359 if ((thflags & TH_ACK) && 2360 acceptable_window_update(tp, th, tiwin)) { 2361 /* keep track of pure window updates */ 2362 if (tlen == 0 && tp->snd_wl2 == th->th_ack && 2363 tiwin > tp->snd_wnd) 2364 tcpstat.tcps_rcvwinupd++; 2365 tp->snd_wnd = tiwin; 2366 tp->snd_wl1 = th->th_seq; 2367 tp->snd_wl2 = th->th_ack; 2368 if (tp->snd_wnd > tp->max_sndwnd) 2369 tp->max_sndwnd = tp->snd_wnd; 2370 needoutput = TRUE; 2371 } 2372 2373 /* 2374 * Process segments with URG. 2375 */ 2376 if ((thflags & TH_URG) && th->th_urp && 2377 !TCPS_HAVERCVDFIN(tp->t_state)) { 2378 /* 2379 * This is a kludge, but if we receive and accept 2380 * random urgent pointers, we'll crash in 2381 * soreceive. It's hard to imagine someone 2382 * actually wanting to send this much urgent data. 2383 */ 2384 if (th->th_urp + so->so_rcv.ssb_cc > sb_max) { 2385 th->th_urp = 0; /* XXX */ 2386 thflags &= ~TH_URG; /* XXX */ 2387 goto dodata; /* XXX */ 2388 } 2389 /* 2390 * If this segment advances the known urgent pointer, 2391 * then mark the data stream. This should not happen 2392 * in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since 2393 * a FIN has been received from the remote side. 2394 * In these states we ignore the URG. 2395 * 2396 * According to RFC961 (Assigned Protocols), 2397 * the urgent pointer points to the last octet 2398 * of urgent data. We continue, however, 2399 * to consider it to indicate the first octet 2400 * of data past the urgent section as the original 2401 * spec states (in one of two places). 2402 */ 2403 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) { 2404 tp->rcv_up = th->th_seq + th->th_urp; 2405 so->so_oobmark = so->so_rcv.ssb_cc + 2406 (tp->rcv_up - tp->rcv_nxt) - 1; 2407 if (so->so_oobmark == 0) 2408 sosetstate(so, SS_RCVATMARK); 2409 sohasoutofband(so); 2410 tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); 2411 } 2412 /* 2413 * Remove out of band data so doesn't get presented to user. 2414 * This can happen independent of advancing the URG pointer, 2415 * but if two URG's are pending at once, some out-of-band 2416 * data may creep in... ick. 2417 */ 2418 if (th->th_urp <= (u_long)tlen && 2419 !(so->so_options & SO_OOBINLINE)) { 2420 /* hdr drop is delayed */ 2421 tcp_pulloutofband(so, th, m, drop_hdrlen); 2422 } 2423 } else { 2424 /* 2425 * If no out of band data is expected, 2426 * pull receive urgent pointer along 2427 * with the receive window. 2428 */ 2429 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) 2430 tp->rcv_up = tp->rcv_nxt; 2431 } 2432 2433 dodata: /* XXX */ 2434 /* 2435 * Process the segment text, merging it into the TCP sequencing queue, 2436 * and arranging for acknowledgment of receipt if necessary. 2437 * This process logically involves adjusting tp->rcv_wnd as data 2438 * is presented to the user (this happens in tcp_usrreq.c, 2439 * case PRU_RCVD). If a FIN has already been received on this 2440 * connection then we just ignore the text. 2441 */ 2442 if ((tlen || (thflags & TH_FIN)) && !TCPS_HAVERCVDFIN(tp->t_state)) { 2443 if (thflags & TH_FIN) 2444 tp->t_flags |= TF_SAWFIN; 2445 m_adj(m, drop_hdrlen); /* delayed header drop */ 2446 /* 2447 * Insert segment which includes th into TCP reassembly queue 2448 * with control block tp. Set thflags to whether reassembly now 2449 * includes a segment with FIN. This handles the common case 2450 * inline (segment is the next to be received on an established 2451 * connection, and the queue is empty), avoiding linkage into 2452 * and removal from the queue and repetition of various 2453 * conversions. 2454 * Set DELACK for segments received in order, but ack 2455 * immediately when segments are out of order (so 2456 * fast retransmit can work). 2457 */ 2458 if (th->th_seq == tp->rcv_nxt && 2459 TAILQ_EMPTY(&tp->t_segq) && 2460 TCPS_HAVEESTABLISHED(tp->t_state)) { 2461 if (thflags & TH_FIN) 2462 tp->t_flags |= TF_QUEDFIN; 2463 if (DELAY_ACK(tp)) { 2464 tcp_callout_reset(tp, tp->tt_delack, 2465 tcp_delacktime, tcp_timer_delack); 2466 } else { 2467 tp->t_flags |= TF_ACKNOW; 2468 } 2469 tp->rcv_nxt += tlen; 2470 thflags = th->th_flags & TH_FIN; 2471 tcpstat.tcps_rcvpack++; 2472 tcpstat.tcps_rcvbyte += tlen; 2473 ND6_HINT(tp); 2474 if (so->so_state & SS_CANTRCVMORE) { 2475 m_freem(m); 2476 } else { 2477 lwkt_gettoken(&so->so_rcv.ssb_token); 2478 ssb_appendstream(&so->so_rcv, m); 2479 lwkt_reltoken(&so->so_rcv.ssb_token); 2480 } 2481 sorwakeup(so); 2482 } else { 2483 if (!(tp->sack_flags & TSACK_F_DUPSEG)) { 2484 /* Initialize SACK report block. */ 2485 tp->reportblk.rblk_start = th->th_seq; 2486 tp->reportblk.rblk_end = TCP_SACK_BLKEND( 2487 th->th_seq + tlen, thflags); 2488 } 2489 thflags = tcp_reass(tp, th, &tlen, m); 2490 tp->t_flags |= TF_ACKNOW; 2491 } 2492 2493 /* 2494 * Note the amount of data that peer has sent into 2495 * our window, in order to estimate the sender's 2496 * buffer size. 2497 */ 2498 len = so->so_rcv.ssb_hiwat - (tp->rcv_adv - tp->rcv_nxt); 2499 } else { 2500 m_freem(m); 2501 thflags &= ~TH_FIN; 2502 } 2503 2504 /* 2505 * If FIN is received ACK the FIN and let the user know 2506 * that the connection is closing. 2507 */ 2508 if (thflags & TH_FIN) { 2509 if (!TCPS_HAVERCVDFIN(tp->t_state)) { 2510 socantrcvmore(so); 2511 /* 2512 * If connection is half-synchronized 2513 * (ie NEEDSYN flag on) then delay ACK, 2514 * so it may be piggybacked when SYN is sent. 2515 * Otherwise, since we received a FIN then no 2516 * more input can be expected, send ACK now. 2517 */ 2518 if (DELAY_ACK(tp) && (tp->t_flags & TF_NEEDSYN)) { 2519 tcp_callout_reset(tp, tp->tt_delack, 2520 tcp_delacktime, tcp_timer_delack); 2521 } else { 2522 tp->t_flags |= TF_ACKNOW; 2523 } 2524 tp->rcv_nxt++; 2525 } 2526 2527 switch (tp->t_state) { 2528 /* 2529 * In SYN_RECEIVED and ESTABLISHED STATES 2530 * enter the CLOSE_WAIT state. 2531 */ 2532 case TCPS_SYN_RECEIVED: 2533 tp->t_starttime = ticks; 2534 /*FALLTHROUGH*/ 2535 case TCPS_ESTABLISHED: 2536 TCP_STATE_CHANGE(tp, TCPS_CLOSE_WAIT); 2537 break; 2538 2539 /* 2540 * If still in FIN_WAIT_1 STATE FIN has not been acked so 2541 * enter the CLOSING state. 2542 */ 2543 case TCPS_FIN_WAIT_1: 2544 TCP_STATE_CHANGE(tp, TCPS_CLOSING); 2545 break; 2546 2547 /* 2548 * In FIN_WAIT_2 state enter the TIME_WAIT state, 2549 * starting the time-wait timer, turning off the other 2550 * standard timers. 2551 */ 2552 case TCPS_FIN_WAIT_2: 2553 TCP_STATE_CHANGE(tp, TCPS_TIME_WAIT); 2554 tcp_canceltimers(tp); 2555 tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp), 2556 tcp_timer_2msl); 2557 soisdisconnected(so); 2558 break; 2559 2560 /* 2561 * In TIME_WAIT state restart the 2 MSL time_wait timer. 2562 */ 2563 case TCPS_TIME_WAIT: 2564 tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp), 2565 tcp_timer_2msl); 2566 break; 2567 } 2568 } 2569 2570 #ifdef TCPDEBUG 2571 if (so->so_options & SO_DEBUG) 2572 tcp_trace(TA_INPUT, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0); 2573 #endif 2574 2575 /* 2576 * Delayed duplicated ACK processing 2577 */ 2578 if (delayed_dupack && tcp_recv_dupack(tp, th_dupack, to_flags)) 2579 needoutput = FALSE; 2580 2581 /* 2582 * Return any desired output. 2583 */ 2584 if ((tp->t_flags & TF_ACKNOW) || 2585 (needoutput && tcp_sack_report_needed(tp))) { 2586 tcp_output_cancel(tp); 2587 tcp_output_fair(tp); 2588 } else if (needoutput && !tcp_output_pending(tp)) { 2589 tcp_output_fair(tp); 2590 } 2591 tcp_sack_report_cleanup(tp); 2592 return(IPPROTO_DONE); 2593 2594 dropafterack: 2595 /* 2596 * Generate an ACK dropping incoming segment if it occupies 2597 * sequence space, where the ACK reflects our state. 2598 * 2599 * We can now skip the test for the RST flag since all 2600 * paths to this code happen after packets containing 2601 * RST have been dropped. 2602 * 2603 * In the SYN-RECEIVED state, don't send an ACK unless the 2604 * segment we received passes the SYN-RECEIVED ACK test. 2605 * If it fails send a RST. This breaks the loop in the 2606 * "LAND" DoS attack, and also prevents an ACK storm 2607 * between two listening ports that have been sent forged 2608 * SYN segments, each with the source address of the other. 2609 */ 2610 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 2611 (SEQ_GT(tp->snd_una, th->th_ack) || 2612 SEQ_GT(th->th_ack, tp->snd_max)) ) { 2613 rstreason = BANDLIM_RST_OPENPORT; 2614 goto dropwithreset; 2615 } 2616 #ifdef TCPDEBUG 2617 if (so->so_options & SO_DEBUG) 2618 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0); 2619 #endif 2620 m_freem(m); 2621 tp->t_flags |= TF_ACKNOW; 2622 tcp_output(tp); 2623 tcp_sack_report_cleanup(tp); 2624 return(IPPROTO_DONE); 2625 2626 dropwithreset: 2627 /* 2628 * Generate a RST, dropping incoming segment. 2629 * Make ACK acceptable to originator of segment. 2630 * Don't bother to respond if destination was broadcast/multicast. 2631 */ 2632 if ((thflags & TH_RST) || m->m_flags & (M_BCAST | M_MCAST)) 2633 goto drop; 2634 if (isipv6) { 2635 if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) || 2636 IN6_IS_ADDR_MULTICAST(&ip6->ip6_src)) 2637 goto drop; 2638 } else { 2639 if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) || 2640 IN_MULTICAST(ntohl(ip->ip_src.s_addr)) || 2641 ip->ip_src.s_addr == htonl(INADDR_BROADCAST) || 2642 in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif)) 2643 goto drop; 2644 } 2645 /* IPv6 anycast check is done at tcp6_input() */ 2646 2647 /* 2648 * Perform bandwidth limiting. 2649 */ 2650 #ifdef ICMP_BANDLIM 2651 if (badport_bandlim(rstreason) < 0) 2652 goto drop; 2653 #endif 2654 2655 #ifdef TCPDEBUG 2656 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 2657 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0); 2658 #endif 2659 if (thflags & TH_ACK) 2660 /* mtod() below is safe as long as hdr dropping is delayed */ 2661 tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack, 2662 TH_RST); 2663 else { 2664 if (thflags & TH_SYN) 2665 tlen++; 2666 /* mtod() below is safe as long as hdr dropping is delayed */ 2667 tcp_respond(tp, mtod(m, void *), th, m, th->th_seq + tlen, 2668 (tcp_seq)0, TH_RST | TH_ACK); 2669 } 2670 if (tp != NULL) 2671 tcp_sack_report_cleanup(tp); 2672 return(IPPROTO_DONE); 2673 2674 drop: 2675 /* 2676 * Drop space held by incoming segment and return. 2677 */ 2678 #ifdef TCPDEBUG 2679 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 2680 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0); 2681 #endif 2682 m_freem(m); 2683 if (tp != NULL) 2684 tcp_sack_report_cleanup(tp); 2685 return(IPPROTO_DONE); 2686 } 2687 2688 /* 2689 * Parse TCP options and place in tcpopt. 2690 */ 2691 static void 2692 tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, boolean_t is_syn, 2693 tcp_seq ack) 2694 { 2695 int opt, optlen, i; 2696 2697 to->to_flags = 0; 2698 for (; cnt > 0; cnt -= optlen, cp += optlen) { 2699 opt = cp[0]; 2700 if (opt == TCPOPT_EOL) 2701 break; 2702 if (opt == TCPOPT_NOP) 2703 optlen = 1; 2704 else { 2705 if (cnt < 2) 2706 break; 2707 optlen = cp[1]; 2708 if (optlen < 2 || optlen > cnt) 2709 break; 2710 } 2711 switch (opt) { 2712 case TCPOPT_MAXSEG: 2713 if (optlen != TCPOLEN_MAXSEG) 2714 continue; 2715 if (!is_syn) 2716 continue; 2717 to->to_flags |= TOF_MSS; 2718 bcopy(cp + 2, &to->to_mss, sizeof to->to_mss); 2719 to->to_mss = ntohs(to->to_mss); 2720 break; 2721 case TCPOPT_WINDOW: 2722 if (optlen != TCPOLEN_WINDOW) 2723 continue; 2724 if (!is_syn) 2725 continue; 2726 to->to_flags |= TOF_SCALE; 2727 to->to_requested_s_scale = min(cp[2], TCP_MAX_WINSHIFT); 2728 break; 2729 case TCPOPT_TIMESTAMP: 2730 if (optlen != TCPOLEN_TIMESTAMP) 2731 continue; 2732 to->to_flags |= TOF_TS; 2733 bcopy(cp + 2, &to->to_tsval, sizeof to->to_tsval); 2734 to->to_tsval = ntohl(to->to_tsval); 2735 bcopy(cp + 6, &to->to_tsecr, sizeof to->to_tsecr); 2736 to->to_tsecr = ntohl(to->to_tsecr); 2737 /* 2738 * If echoed timestamp is later than the current time, 2739 * fall back to non RFC1323 RTT calculation. 2740 */ 2741 if (to->to_tsecr != 0 && TSTMP_GT(to->to_tsecr, ticks)) 2742 to->to_tsecr = 0; 2743 break; 2744 case TCPOPT_SACK_PERMITTED: 2745 if (optlen != TCPOLEN_SACK_PERMITTED) 2746 continue; 2747 if (!is_syn) 2748 continue; 2749 to->to_flags |= TOF_SACK_PERMITTED; 2750 break; 2751 case TCPOPT_SACK: 2752 if ((optlen - 2) & 0x07) /* not multiple of 8 */ 2753 continue; 2754 to->to_nsackblocks = (optlen - 2) / 8; 2755 to->to_sackblocks = (struct raw_sackblock *) (cp + 2); 2756 to->to_flags |= TOF_SACK; 2757 for (i = 0; i < to->to_nsackblocks; i++) { 2758 struct raw_sackblock *r = &to->to_sackblocks[i]; 2759 2760 r->rblk_start = ntohl(r->rblk_start); 2761 r->rblk_end = ntohl(r->rblk_end); 2762 2763 if (SEQ_LEQ(r->rblk_end, r->rblk_start)) { 2764 /* 2765 * Invalid SACK block; discard all 2766 * SACK blocks 2767 */ 2768 tcpstat.tcps_rcvbadsackopt++; 2769 to->to_nsackblocks = 0; 2770 to->to_sackblocks = NULL; 2771 to->to_flags &= ~TOF_SACK; 2772 break; 2773 } 2774 } 2775 if ((to->to_flags & TOF_SACK) && 2776 tcp_sack_ndsack_blocks(to->to_sackblocks, 2777 to->to_nsackblocks, ack)) 2778 to->to_flags |= TOF_DSACK; 2779 break; 2780 #ifdef TCP_SIGNATURE 2781 /* 2782 * XXX In order to reply to a host which has set the 2783 * TCP_SIGNATURE option in its initial SYN, we have to 2784 * record the fact that the option was observed here 2785 * for the syncache code to perform the correct response. 2786 */ 2787 case TCPOPT_SIGNATURE: 2788 if (optlen != TCPOLEN_SIGNATURE) 2789 continue; 2790 to->to_flags |= (TOF_SIGNATURE | TOF_SIGLEN); 2791 break; 2792 #endif /* TCP_SIGNATURE */ 2793 default: 2794 continue; 2795 } 2796 } 2797 } 2798 2799 /* 2800 * Pull out of band byte out of a segment so 2801 * it doesn't appear in the user's data queue. 2802 * It is still reflected in the segment length for 2803 * sequencing purposes. 2804 * "off" is the delayed to be dropped hdrlen. 2805 */ 2806 static void 2807 tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m, int off) 2808 { 2809 int cnt = off + th->th_urp - 1; 2810 2811 while (cnt >= 0) { 2812 if (m->m_len > cnt) { 2813 char *cp = mtod(m, caddr_t) + cnt; 2814 struct tcpcb *tp = sototcpcb(so); 2815 2816 tp->t_iobc = *cp; 2817 tp->t_oobflags |= TCPOOB_HAVEDATA; 2818 bcopy(cp + 1, cp, m->m_len - cnt - 1); 2819 m->m_len--; 2820 if (m->m_flags & M_PKTHDR) 2821 m->m_pkthdr.len--; 2822 return; 2823 } 2824 cnt -= m->m_len; 2825 m = m->m_next; 2826 if (m == NULL) 2827 break; 2828 } 2829 panic("tcp_pulloutofband"); 2830 } 2831 2832 /* 2833 * Collect new round-trip time estimate and update averages and current 2834 * timeout. 2835 */ 2836 static void 2837 tcp_xmit_timer(struct tcpcb *tp, int rtt, tcp_seq ack) 2838 { 2839 int rebaserto = 0; 2840 2841 tcpstat.tcps_rttupdated++; 2842 tp->t_rttupdated++; 2843 if ((tp->rxt_flags & TRXT_F_REBASERTO) && 2844 SEQ_GT(ack, tp->snd_max_prev)) { 2845 #ifdef DEBUG_EIFEL_RESPONSE 2846 kprintf("srtt/rttvar, prev %d/%d, cur %d/%d, ", 2847 tp->t_srtt_prev, tp->t_rttvar_prev, 2848 tp->t_srtt, tp->t_rttvar); 2849 #endif 2850 2851 tcpstat.tcps_eifelresponse++; 2852 rebaserto = 1; 2853 tp->rxt_flags &= ~TRXT_F_REBASERTO; 2854 tp->t_srtt = max(tp->t_srtt_prev, (rtt << TCP_RTT_SHIFT)); 2855 tp->t_rttvar = max(tp->t_rttvar_prev, 2856 (rtt << (TCP_RTTVAR_SHIFT - 1))); 2857 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) 2858 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 2859 2860 #ifdef DEBUG_EIFEL_RESPONSE 2861 kprintf("new %d/%d ", tp->t_srtt, tp->t_rttvar); 2862 #endif 2863 } else if (tp->t_srtt != 0) { 2864 int delta; 2865 2866 /* 2867 * srtt is stored as fixed point with 5 bits after the 2868 * binary point (i.e., scaled by 32). The following magic 2869 * is equivalent to the smoothing algorithm in rfc793 with 2870 * an alpha of .875 (srtt = rtt/32 + srtt*31/32 in fixed 2871 * point). Adjust rtt to origin 0. 2872 */ 2873 delta = ((rtt - 1) << TCP_DELTA_SHIFT) 2874 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 2875 2876 if ((tp->t_srtt += delta) <= 0) 2877 tp->t_srtt = 1; 2878 2879 /* 2880 * We accumulate a smoothed rtt variance (actually, a 2881 * smoothed mean difference), then set the retransmit 2882 * timer to smoothed rtt + 4 times the smoothed variance. 2883 * rttvar is stored as fixed point with 4 bits after the 2884 * binary point (scaled by 16). The following is 2885 * equivalent to rfc793 smoothing with an alpha of .75 2886 * (rttvar = rttvar*3/4 + |delta| / 4). This replaces 2887 * rfc793's wired-in beta. 2888 */ 2889 if (delta < 0) 2890 delta = -delta; 2891 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 2892 if ((tp->t_rttvar += delta) <= 0) 2893 tp->t_rttvar = 1; 2894 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) 2895 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 2896 } else { 2897 /* 2898 * No rtt measurement yet - use the unsmoothed rtt. 2899 * Set the variance to half the rtt (so our first 2900 * retransmit happens at 3*rtt). 2901 */ 2902 tp->t_srtt = rtt << TCP_RTT_SHIFT; 2903 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); 2904 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 2905 } 2906 tp->t_rtttime = 0; 2907 tp->t_rxtshift = 0; 2908 2909 #ifdef DEBUG_EIFEL_RESPONSE 2910 if (rebaserto) { 2911 kprintf("| rxtcur prev %d, old %d, ", 2912 tp->t_rxtcur_prev, tp->t_rxtcur); 2913 } 2914 #endif 2915 2916 /* 2917 * the retransmit should happen at rtt + 4 * rttvar. 2918 * Because of the way we do the smoothing, srtt and rttvar 2919 * will each average +1/2 tick of bias. When we compute 2920 * the retransmit timer, we want 1/2 tick of rounding and 2921 * 1 extra tick because of +-1/2 tick uncertainty in the 2922 * firing of the timer. The bias will give us exactly the 2923 * 1.5 tick we need. But, because the bias is 2924 * statistical, we have to test that we don't drop below 2925 * the minimum feasible timer (which is 2 ticks). 2926 */ 2927 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 2928 max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX); 2929 2930 if (rebaserto) { 2931 if (tp->t_rxtcur < tp->t_rxtcur_prev + tcp_eifel_rtoinc) { 2932 /* 2933 * RFC4015 requires that the new RTO is at least 2934 * 2*G (tcp_eifel_rtoinc) greater then the RTO 2935 * (t_rxtcur_prev) when the spurious retransmit 2936 * timeout happens. 2937 * 2938 * The above condition could be true, if the SRTT 2939 * and RTTVAR used to calculate t_rxtcur_prev 2940 * resulted in a value less than t_rttmin. So 2941 * simply increasing SRTT by tcp_eifel_rtoinc when 2942 * preparing for the Eifel response could not ensure 2943 * that the new RTO will be tcp_eifel_rtoinc greater 2944 * t_rxtcur_prev. 2945 */ 2946 tp->t_rxtcur = tp->t_rxtcur_prev + tcp_eifel_rtoinc; 2947 } 2948 #ifdef DEBUG_EIFEL_RESPONSE 2949 kprintf("new %d\n", tp->t_rxtcur); 2950 #endif 2951 } 2952 2953 /* 2954 * We received an ack for a packet that wasn't retransmitted; 2955 * it is probably safe to discard any error indications we've 2956 * received recently. This isn't quite right, but close enough 2957 * for now (a route might have failed after we sent a segment, 2958 * and the return path might not be symmetrical). 2959 */ 2960 tp->t_softerror = 0; 2961 } 2962 2963 /* 2964 * Determine a reasonable value for maxseg size. 2965 * If the route is known, check route for mtu. 2966 * If none, use an mss that can be handled on the outgoing 2967 * interface without forcing IP to fragment; if bigger than 2968 * an mbuf cluster (MCLBYTES), round down to nearest multiple of MCLBYTES 2969 * to utilize large mbufs. If no route is found, route has no mtu, 2970 * or the destination isn't local, use a default, hopefully conservative 2971 * size (usually 512 or the default IP max size, but no more than the mtu 2972 * of the interface), as we can't discover anything about intervening 2973 * gateways or networks. We also initialize the congestion/slow start 2974 * window to be a single segment if the destination isn't local. 2975 * While looking at the routing entry, we also initialize other path-dependent 2976 * parameters from pre-set or cached values in the routing entry. 2977 * 2978 * Also take into account the space needed for options that we 2979 * send regularly. Make maxseg shorter by that amount to assure 2980 * that we can send maxseg amount of data even when the options 2981 * are present. Store the upper limit of the length of options plus 2982 * data in maxopd. 2983 * 2984 * NOTE that this routine is only called when we process an incoming 2985 * segment, for outgoing segments only tcp_mssopt is called. 2986 */ 2987 void 2988 tcp_mss(struct tcpcb *tp, int offer) 2989 { 2990 struct rtentry *rt; 2991 struct ifnet *ifp; 2992 int rtt, mss; 2993 u_long bufsize; 2994 struct inpcb *inp = tp->t_inpcb; 2995 struct socket *so; 2996 #ifdef INET6 2997 boolean_t isipv6 = INP_ISIPV6(inp); 2998 size_t min_protoh = isipv6 ? 2999 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 3000 sizeof(struct tcpiphdr); 3001 #else 3002 const boolean_t isipv6 = FALSE; 3003 const size_t min_protoh = sizeof(struct tcpiphdr); 3004 #endif 3005 3006 if (isipv6) 3007 rt = tcp_rtlookup6(&inp->inp_inc); 3008 else 3009 rt = tcp_rtlookup(&inp->inp_inc); 3010 if (rt == NULL) { 3011 tp->t_maxopd = tp->t_maxseg = 3012 (isipv6 ? tcp_v6mssdflt : tcp_mssdflt); 3013 return; 3014 } 3015 ifp = rt->rt_ifp; 3016 so = inp->inp_socket; 3017 3018 /* 3019 * Offer == 0 means that there was no MSS on the SYN segment, 3020 * in this case we use either the interface mtu or tcp_mssdflt. 3021 * 3022 * An offer which is too large will be cut down later. 3023 */ 3024 if (offer == 0) { 3025 if (isipv6) { 3026 if (in6_localaddr(&inp->in6p_faddr)) 3027 offer = IN6_LINKMTU(rt->rt_ifp) - min_protoh; 3028 else 3029 offer = tcp_v6mssdflt; 3030 } else { 3031 if (in_localaddr(inp->inp_faddr)) 3032 offer = ifp->if_mtu - min_protoh; 3033 else 3034 offer = tcp_mssdflt; 3035 } 3036 } 3037 3038 /* 3039 * Prevent DoS attack with too small MSS. Round up 3040 * to at least minmss. 3041 * 3042 * Sanity check: make sure that maxopd will be large 3043 * enough to allow some data on segments even is the 3044 * all the option space is used (40bytes). Otherwise 3045 * funny things may happen in tcp_output. 3046 */ 3047 offer = max(offer, tcp_minmss); 3048 offer = max(offer, 64); 3049 3050 rt->rt_rmx.rmx_mssopt = offer; 3051 3052 /* 3053 * While we're here, check if there's an initial rtt 3054 * or rttvar. Convert from the route-table units 3055 * to scaled multiples of the slow timeout timer. 3056 */ 3057 if (tp->t_srtt == 0 && (rtt = rt->rt_rmx.rmx_rtt)) { 3058 /* 3059 * XXX the lock bit for RTT indicates that the value 3060 * is also a minimum value; this is subject to time. 3061 */ 3062 if (rt->rt_rmx.rmx_locks & RTV_RTT) 3063 tp->t_rttmin = rtt / (RTM_RTTUNIT / hz); 3064 tp->t_srtt = rtt / (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 3065 tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE; 3066 tcpstat.tcps_usedrtt++; 3067 if (rt->rt_rmx.rmx_rttvar) { 3068 tp->t_rttvar = rt->rt_rmx.rmx_rttvar / 3069 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 3070 tcpstat.tcps_usedrttvar++; 3071 } else { 3072 /* default variation is +- 1 rtt */ 3073 tp->t_rttvar = 3074 tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE; 3075 } 3076 TCPT_RANGESET(tp->t_rxtcur, 3077 ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1, 3078 tp->t_rttmin, TCPTV_REXMTMAX); 3079 } 3080 3081 /* 3082 * if there's an mtu associated with the route, use it 3083 * else, use the link mtu. Take the smaller of mss or offer 3084 * as our final mss. 3085 */ 3086 if (rt->rt_rmx.rmx_mtu) { 3087 mss = rt->rt_rmx.rmx_mtu; 3088 } else { 3089 if (isipv6) 3090 mss = IN6_LINKMTU(rt->rt_ifp); 3091 else 3092 mss = ifp->if_mtu; 3093 } 3094 mss -= min_protoh; 3095 mss = min(mss, offer); 3096 3097 /* 3098 * maxopd stores the maximum length of data AND options 3099 * in a segment; maxseg is the amount of data in a normal 3100 * segment. We need to store this value (maxopd) apart 3101 * from maxseg, because now every segment carries options 3102 * and thus we normally have somewhat less data in segments. 3103 */ 3104 tp->t_maxopd = mss; 3105 3106 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 3107 ((tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)) 3108 mss -= TCPOLEN_TSTAMP_APPA; 3109 3110 #if (MCLBYTES & (MCLBYTES - 1)) == 0 3111 if (mss > MCLBYTES) 3112 mss &= ~(MCLBYTES-1); 3113 #else 3114 if (mss > MCLBYTES) 3115 mss = mss / MCLBYTES * MCLBYTES; 3116 #endif 3117 /* 3118 * If there's a pipesize, change the socket buffer 3119 * to that size. Make the socket buffers an integral 3120 * number of mss units; if the mss is larger than 3121 * the socket buffer, decrease the mss. 3122 */ 3123 #ifdef RTV_SPIPE 3124 if ((bufsize = rt->rt_rmx.rmx_sendpipe) == 0) 3125 #endif 3126 bufsize = so->so_snd.ssb_hiwat; 3127 if (bufsize < mss) 3128 mss = bufsize; 3129 else { 3130 bufsize = roundup(bufsize, mss); 3131 if (bufsize > sb_max) 3132 bufsize = sb_max; 3133 if (bufsize > so->so_snd.ssb_hiwat) 3134 ssb_reserve(&so->so_snd, bufsize, so, NULL); 3135 } 3136 tp->t_maxseg = mss; 3137 3138 #ifdef RTV_RPIPE 3139 if ((bufsize = rt->rt_rmx.rmx_recvpipe) == 0) 3140 #endif 3141 bufsize = so->so_rcv.ssb_hiwat; 3142 if (bufsize > mss) { 3143 bufsize = roundup(bufsize, mss); 3144 if (bufsize > sb_max) 3145 bufsize = sb_max; 3146 if (bufsize > so->so_rcv.ssb_hiwat) { 3147 lwkt_gettoken(&so->so_rcv.ssb_token); 3148 ssb_reserve(&so->so_rcv, bufsize, so, NULL); 3149 lwkt_reltoken(&so->so_rcv.ssb_token); 3150 } 3151 } 3152 3153 /* 3154 * Set the slow-start flight size 3155 * 3156 * NOTE: t_maxseg must have been configured! 3157 */ 3158 tp->snd_cwnd = tcp_initial_window(tp); 3159 3160 if (rt->rt_rmx.rmx_ssthresh) { 3161 /* 3162 * There's some sort of gateway or interface 3163 * buffer limit on the path. Use this to set 3164 * the slow start threshhold, but set the 3165 * threshold to no less than 2*mss. 3166 */ 3167 tp->snd_ssthresh = max(2 * mss, rt->rt_rmx.rmx_ssthresh); 3168 tcpstat.tcps_usedssthresh++; 3169 } 3170 } 3171 3172 /* 3173 * Determine the MSS option to send on an outgoing SYN. 3174 */ 3175 int 3176 tcp_mssopt(struct tcpcb *tp) 3177 { 3178 struct rtentry *rt; 3179 #ifdef INET6 3180 boolean_t isipv6 = INP_ISIPV6(tp->t_inpcb); 3181 int min_protoh = isipv6 ? 3182 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 3183 sizeof(struct tcpiphdr); 3184 #else 3185 const boolean_t isipv6 = FALSE; 3186 const size_t min_protoh = sizeof(struct tcpiphdr); 3187 #endif 3188 3189 if (isipv6) 3190 rt = tcp_rtlookup6(&tp->t_inpcb->inp_inc); 3191 else 3192 rt = tcp_rtlookup(&tp->t_inpcb->inp_inc); 3193 if (rt == NULL) 3194 return (isipv6 ? tcp_v6mssdflt : tcp_mssdflt); 3195 3196 #ifdef INET6 3197 return ((isipv6 ? IN6_LINKMTU(rt->rt_ifp) : rt->rt_ifp->if_mtu) - 3198 min_protoh); 3199 #else 3200 return (rt->rt_ifp->if_mtu - min_protoh); 3201 #endif 3202 } 3203 3204 /* 3205 * When a partial ack arrives, force the retransmission of the 3206 * next unacknowledged segment. Do not exit Fast Recovery. 3207 * 3208 * Implement the Slow-but-Steady variant of NewReno by restarting the 3209 * the retransmission timer. Turn it off here so it can be restarted 3210 * later in tcp_output(). 3211 */ 3212 static void 3213 tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th, int acked) 3214 { 3215 tcp_seq old_snd_nxt = tp->snd_nxt; 3216 u_long ocwnd = tp->snd_cwnd; 3217 3218 tcp_callout_stop(tp, tp->tt_rexmt); 3219 tp->t_rtttime = 0; 3220 tp->snd_nxt = th->th_ack; 3221 /* Set snd_cwnd to one segment beyond acknowledged offset. */ 3222 tp->snd_cwnd = tp->t_maxseg; 3223 tp->t_flags |= TF_ACKNOW; 3224 tcp_output(tp); 3225 if (SEQ_GT(old_snd_nxt, tp->snd_nxt)) 3226 tp->snd_nxt = old_snd_nxt; 3227 /* partial window deflation */ 3228 if (ocwnd > acked) 3229 tp->snd_cwnd = ocwnd - acked + tp->t_maxseg; 3230 else 3231 tp->snd_cwnd = tp->t_maxseg; 3232 } 3233 3234 /* 3235 * In contrast to the Slow-but-Steady NewReno variant, 3236 * we do not reset the retransmission timer for SACK retransmissions, 3237 * except when retransmitting snd_una. 3238 */ 3239 static void 3240 tcp_sack_rexmt(struct tcpcb *tp, boolean_t force) 3241 { 3242 tcp_seq old_snd_nxt = tp->snd_nxt; 3243 u_long ocwnd = tp->snd_cwnd; 3244 uint32_t pipe; 3245 int nseg = 0; /* consecutive new segments */ 3246 int nseg_rexmt = 0; /* retransmitted segments */ 3247 int maxrexmt = 0; 3248 3249 if (force) { 3250 uint32_t unsacked = tcp_sack_first_unsacked_len(tp); 3251 3252 /* 3253 * Try to fill the first hole in the receiver's 3254 * reassemble queue. 3255 */ 3256 maxrexmt = howmany(unsacked, tp->t_maxseg); 3257 if (maxrexmt > tcp_force_sackrxt) 3258 maxrexmt = tcp_force_sackrxt; 3259 } 3260 3261 tp->t_rtttime = 0; 3262 pipe = tcp_sack_compute_pipe(tp); 3263 while (((tcp_seq_diff_t)(ocwnd - pipe) >= (tcp_seq_diff_t)tp->t_maxseg 3264 || (force && nseg_rexmt < maxrexmt && nseg == 0)) && 3265 (!tcp_do_smartsack || nseg < TCP_SACK_MAXBURST)) { 3266 tcp_seq old_snd_max, old_rexmt_high, nextrexmt; 3267 uint32_t sent, seglen; 3268 boolean_t rescue; 3269 int error; 3270 3271 old_rexmt_high = tp->rexmt_high; 3272 if (!tcp_sack_nextseg(tp, &nextrexmt, &seglen, &rescue)) { 3273 tp->rexmt_high = old_rexmt_high; 3274 break; 3275 } 3276 3277 /* 3278 * If the next tranmission is a rescue retranmission, 3279 * we check whether we have already sent some data 3280 * (either new segments or retransmitted segments) 3281 * into the the network or not. Since the idea of rescue 3282 * retransmission is to sustain ACK clock, as long as 3283 * some segments are in the network, ACK clock will be 3284 * kept ticking. 3285 */ 3286 if (rescue && (nseg_rexmt > 0 || nseg > 0)) { 3287 tp->rexmt_high = old_rexmt_high; 3288 break; 3289 } 3290 3291 if (nextrexmt == tp->snd_max) 3292 ++nseg; 3293 else 3294 ++nseg_rexmt; 3295 tp->snd_nxt = nextrexmt; 3296 tp->snd_cwnd = nextrexmt - tp->snd_una + seglen; 3297 old_snd_max = tp->snd_max; 3298 if (nextrexmt == tp->snd_una) 3299 tcp_callout_stop(tp, tp->tt_rexmt); 3300 tp->t_flags |= TF_XMITNOW; 3301 error = tcp_output(tp); 3302 if (error != 0) { 3303 tp->rexmt_high = old_rexmt_high; 3304 break; 3305 } 3306 sent = tp->snd_nxt - nextrexmt; 3307 if (sent <= 0) { 3308 tp->rexmt_high = old_rexmt_high; 3309 break; 3310 } 3311 pipe += sent; 3312 tcpstat.tcps_sndsackpack++; 3313 tcpstat.tcps_sndsackbyte += sent; 3314 3315 if (rescue) { 3316 tcpstat.tcps_sackrescue++; 3317 tp->rexmt_rescue = tp->snd_nxt; 3318 tp->sack_flags |= TSACK_F_SACKRESCUED; 3319 break; 3320 } 3321 if (SEQ_LT(nextrexmt, old_snd_max) && 3322 SEQ_LT(tp->rexmt_high, tp->snd_nxt)) { 3323 tp->rexmt_high = seq_min(tp->snd_nxt, old_snd_max); 3324 if (tcp_aggressive_rescuesack && 3325 (tp->sack_flags & TSACK_F_SACKRESCUED) && 3326 SEQ_LT(tp->rexmt_rescue, tp->rexmt_high)) { 3327 /* Drag RescueRxt along with HighRxt */ 3328 tp->rexmt_rescue = tp->rexmt_high; 3329 } 3330 } 3331 } 3332 if (SEQ_GT(old_snd_nxt, tp->snd_nxt)) 3333 tp->snd_nxt = old_snd_nxt; 3334 tp->snd_cwnd = ocwnd; 3335 } 3336 3337 /* 3338 * Return TRUE, if some new segments are sent 3339 */ 3340 static boolean_t 3341 tcp_sack_limitedxmit(struct tcpcb *tp) 3342 { 3343 tcp_seq oldsndnxt = tp->snd_nxt; 3344 tcp_seq oldsndmax = tp->snd_max; 3345 u_long ocwnd = tp->snd_cwnd; 3346 uint32_t pipe, sent; 3347 boolean_t ret = FALSE; 3348 tcp_seq_diff_t cwnd_left; 3349 tcp_seq next; 3350 3351 tp->rexmt_high = tp->snd_una - 1; 3352 pipe = tcp_sack_compute_pipe(tp); 3353 cwnd_left = (tcp_seq_diff_t)(ocwnd - pipe); 3354 if (cwnd_left < (tcp_seq_diff_t)tp->t_maxseg) 3355 return FALSE; 3356 3357 if (tcp_do_smartsack) 3358 cwnd_left = ulmin(cwnd_left, tp->t_maxseg * TCP_SACK_MAXBURST); 3359 3360 next = tp->snd_nxt = tp->snd_max; 3361 tp->snd_cwnd = tp->snd_nxt - tp->snd_una + 3362 rounddown(cwnd_left, tp->t_maxseg); 3363 3364 tp->t_flags |= TF_XMITNOW; 3365 tcp_output(tp); 3366 3367 sent = tp->snd_nxt - next; 3368 if (sent > 0) { 3369 tcpstat.tcps_sndlimited += howmany(sent, tp->t_maxseg); 3370 ret = TRUE; 3371 } 3372 3373 if (SEQ_LT(oldsndnxt, oldsndmax)) { 3374 KASSERT(SEQ_GEQ(oldsndnxt, tp->snd_una), 3375 ("snd_una moved in other threads")); 3376 tp->snd_nxt = oldsndnxt; 3377 } 3378 tp->snd_cwnd = ocwnd; 3379 3380 if (ret && TCP_DO_NCR(tp)) 3381 tcp_ncr_update_rxtthresh(tp); 3382 3383 return ret; 3384 } 3385 3386 /* 3387 * Reset idle time and keep-alive timer, typically called when a valid 3388 * tcp packet is received but may also be called when FASTKEEP is set 3389 * to prevent the previous long-timeout from calculating to a drop. 3390 * 3391 * Only update t_rcvtime for non-SYN packets. 3392 * 3393 * Handle the case where one side thinks the connection is established 3394 * but the other side has, say, rebooted without cleaning out the 3395 * connection. The SYNs could be construed as an attack and wind 3396 * up ignored, but in case it isn't an attack we can validate the 3397 * connection by forcing a keepalive. 3398 */ 3399 void 3400 tcp_timer_keep_activity(struct tcpcb *tp, int thflags) 3401 { 3402 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 3403 if ((thflags & (TH_SYN | TH_ACK)) == TH_SYN) { 3404 tp->t_flags |= TF_KEEPALIVE; 3405 tcp_callout_reset(tp, tp->tt_keep, hz / 2, 3406 tcp_timer_keep); 3407 } else { 3408 tp->t_rcvtime = ticks; 3409 tp->t_flags &= ~TF_KEEPALIVE; 3410 tcp_callout_reset(tp, tp->tt_keep, 3411 tp->t_keepidle, 3412 tcp_timer_keep); 3413 } 3414 } 3415 } 3416 3417 static int 3418 tcp_rmx_msl(const struct tcpcb *tp) 3419 { 3420 struct rtentry *rt; 3421 struct inpcb *inp = tp->t_inpcb; 3422 int msl; 3423 #ifdef INET6 3424 boolean_t isipv6 = INP_ISIPV6(inp); 3425 #else 3426 const boolean_t isipv6 = FALSE; 3427 #endif 3428 3429 if (isipv6) 3430 rt = tcp_rtlookup6(&inp->inp_inc); 3431 else 3432 rt = tcp_rtlookup(&inp->inp_inc); 3433 if (rt == NULL || rt->rt_rmx.rmx_msl == 0) 3434 return tcp_msl; 3435 3436 msl = (rt->rt_rmx.rmx_msl * hz) / 1000; 3437 if (msl == 0) 3438 msl = 1; 3439 3440 return msl; 3441 } 3442 3443 static void 3444 tcp_established(struct tcpcb *tp) 3445 { 3446 TCP_STATE_CHANGE(tp, TCPS_ESTABLISHED); 3447 tcp_callout_reset(tp, tp->tt_keep, tp->t_keepidle, tcp_timer_keep); 3448 3449 if (tp->t_rxtsyn > 0) { 3450 /* 3451 * RFC6298: 3452 * "If the timer expires awaiting the ACK of a SYN segment 3453 * and the TCP implementation is using an RTO less than 3 3454 * seconds, the RTO MUST be re-initialized to 3 seconds 3455 * when data transmission begins" 3456 */ 3457 if (tp->t_rxtcur < TCPTV_RTOBASE3) 3458 tp->t_rxtcur = TCPTV_RTOBASE3; 3459 } 3460 } 3461 3462 /* 3463 * Returns TRUE, if the ACK should be dropped 3464 */ 3465 static boolean_t 3466 tcp_recv_dupack(struct tcpcb *tp, tcp_seq th_ack, u_int to_flags) 3467 { 3468 boolean_t fast_sack_rexmt = TRUE; 3469 3470 tcpstat.tcps_rcvdupack++; 3471 3472 /* 3473 * We have outstanding data (other than a window probe), 3474 * this is a completely duplicate ack (ie, window info 3475 * didn't change), the ack is the biggest we've seen and 3476 * we've seen exactly our rexmt threshhold of them, so 3477 * assume a packet has been dropped and retransmit it. 3478 * Kludge snd_nxt & the congestion window so we send only 3479 * this one packet. 3480 */ 3481 if (IN_FASTRECOVERY(tp)) { 3482 if (TCP_DO_SACK(tp)) { 3483 boolean_t force = FALSE; 3484 3485 if (tp->snd_una == tp->rexmt_high && 3486 (to_flags & (TOF_SACK | TOF_SACK_REDUNDANT)) == 3487 TOF_SACK) { 3488 /* 3489 * New segments got SACKed and 3490 * no retransmit yet. 3491 */ 3492 force = TRUE; 3493 } 3494 3495 /* No artifical cwnd inflation. */ 3496 tcp_sack_rexmt(tp, force); 3497 } else { 3498 /* 3499 * Dup acks mean that packets have left 3500 * the network (they're now cached at the 3501 * receiver) so bump cwnd by the amount in 3502 * the receiver to keep a constant cwnd 3503 * packets in the network. 3504 */ 3505 tp->snd_cwnd += tp->t_maxseg; 3506 tcp_output(tp); 3507 } 3508 return TRUE; 3509 } else if (SEQ_LT(th_ack, tp->snd_recover)) { 3510 tp->t_dupacks = 0; 3511 return FALSE; 3512 } else if (tcp_ignore_redun_dsack && TCP_DO_SACK(tp) && 3513 (to_flags & (TOF_DSACK | TOF_SACK_REDUNDANT)) == 3514 (TOF_DSACK | TOF_SACK_REDUNDANT)) { 3515 /* 3516 * If the ACK carries DSACK and other SACK blocks 3517 * carry information that we have already known, 3518 * don't count this ACK as duplicate ACK. This 3519 * prevents spurious early retransmit and fast 3520 * retransmit. This also meets the requirement of 3521 * RFC3042 that new segments should not be sent if 3522 * the SACK blocks do not contain new information 3523 * (XXX we actually loosen the requirment that only 3524 * DSACK is checked here). 3525 * 3526 * This kind of ACKs are usually sent after spurious 3527 * retransmit. 3528 */ 3529 /* Do nothing; don't change t_dupacks */ 3530 return TRUE; 3531 } else if (tp->t_dupacks == 0 && TCP_DO_NCR(tp)) { 3532 tcp_ncr_update_rxtthresh(tp); 3533 } 3534 3535 if (++tp->t_dupacks == tp->t_rxtthresh) { 3536 tcp_seq old_snd_nxt; 3537 u_int win; 3538 3539 fastretransmit: 3540 if (tcp_do_eifel_detect && (tp->t_flags & TF_RCVD_TSTMP)) { 3541 tcp_save_congestion_state(tp); 3542 tp->rxt_flags |= TRXT_F_FASTREXMT; 3543 } 3544 /* 3545 * We know we're losing at the current window size, 3546 * so do congestion avoidance: set ssthresh to half 3547 * the current window and pull our congestion window 3548 * back to the new ssthresh. 3549 */ 3550 win = min(tp->snd_wnd, tp->snd_cwnd) / 2 / tp->t_maxseg; 3551 if (win < 2) 3552 win = 2; 3553 tp->snd_ssthresh = win * tp->t_maxseg; 3554 ENTER_FASTRECOVERY(tp); 3555 tp->snd_recover = tp->snd_max; 3556 tcp_callout_stop(tp, tp->tt_rexmt); 3557 tp->t_rtttime = 0; 3558 old_snd_nxt = tp->snd_nxt; 3559 tp->snd_nxt = th_ack; 3560 if (TCP_DO_SACK(tp)) { 3561 uint32_t rxtlen; 3562 3563 rxtlen = tcp_sack_first_unsacked_len(tp); 3564 if (rxtlen > tp->t_maxseg) 3565 rxtlen = tp->t_maxseg; 3566 tp->snd_cwnd = rxtlen; 3567 } else { 3568 tp->snd_cwnd = tp->t_maxseg; 3569 } 3570 tcp_output(tp); 3571 ++tcpstat.tcps_sndfastrexmit; 3572 tp->snd_cwnd = tp->snd_ssthresh; 3573 tp->rexmt_high = tp->snd_nxt; 3574 tp->sack_flags &= ~TSACK_F_SACKRESCUED; 3575 if (SEQ_GT(old_snd_nxt, tp->snd_nxt)) 3576 tp->snd_nxt = old_snd_nxt; 3577 KASSERT(tp->snd_limited <= 2, ("tp->snd_limited too big")); 3578 if (TCP_DO_SACK(tp)) { 3579 if (fast_sack_rexmt) 3580 tcp_sack_rexmt(tp, FALSE); 3581 } else { 3582 tp->snd_cwnd += tp->t_maxseg * 3583 (tp->t_dupacks - tp->snd_limited); 3584 } 3585 } else if ((tcp_do_rfc6675 && TCP_DO_SACK(tp)) || TCP_DO_NCR(tp)) { 3586 /* 3587 * The RFC6675 recommends to reduce the byte threshold, 3588 * and enter fast retransmit if IsLost(snd_una). However, 3589 * if we use IsLost(snd_una) based fast retransmit here, 3590 * segments reordering will cause spurious retransmit. So 3591 * we defer the IsLost(snd_una) based fast retransmit until 3592 * the extended limited transmit can't send any segments and 3593 * early retransmit can't be done. 3594 */ 3595 if (tcp_rfc6675_rxt && tcp_do_rfc6675 && 3596 tcp_sack_islost(&tp->scb, tp->snd_una)) 3597 goto fastretransmit; 3598 3599 if (tcp_do_limitedtransmit || TCP_DO_NCR(tp)) { 3600 if (!tcp_sack_limitedxmit(tp)) { 3601 /* outstanding data */ 3602 uint32_t ownd = tp->snd_max - tp->snd_una; 3603 3604 if (need_early_retransmit(tp, ownd)) { 3605 ++tcpstat.tcps_sndearlyrexmit; 3606 tp->rxt_flags |= TRXT_F_EARLYREXMT; 3607 goto fastretransmit; 3608 } else if (tcp_do_rfc6675 && 3609 tcp_sack_islost(&tp->scb, tp->snd_una)) { 3610 fast_sack_rexmt = FALSE; 3611 goto fastretransmit; 3612 } 3613 } 3614 } 3615 } else if (tcp_do_limitedtransmit) { 3616 u_long oldcwnd = tp->snd_cwnd; 3617 tcp_seq oldsndmax = tp->snd_max; 3618 tcp_seq oldsndnxt = tp->snd_nxt; 3619 /* outstanding data */ 3620 uint32_t ownd = tp->snd_max - tp->snd_una; 3621 u_int sent; 3622 3623 KASSERT(tp->t_dupacks == 1 || tp->t_dupacks == 2, 3624 ("dupacks not 1 or 2")); 3625 if (tp->t_dupacks == 1) 3626 tp->snd_limited = 0; 3627 tp->snd_nxt = tp->snd_max; 3628 tp->snd_cwnd = ownd + 3629 (tp->t_dupacks - tp->snd_limited) * tp->t_maxseg; 3630 tp->t_flags |= TF_XMITNOW; 3631 tcp_output(tp); 3632 3633 if (SEQ_LT(oldsndnxt, oldsndmax)) { 3634 KASSERT(SEQ_GEQ(oldsndnxt, tp->snd_una), 3635 ("snd_una moved in other threads")); 3636 tp->snd_nxt = oldsndnxt; 3637 } 3638 tp->snd_cwnd = oldcwnd; 3639 sent = tp->snd_max - oldsndmax; 3640 if (sent > tp->t_maxseg) { 3641 KASSERT((tp->t_dupacks == 2 && tp->snd_limited == 0) || 3642 (sent == tp->t_maxseg + 1 && 3643 (tp->t_flags & TF_SENTFIN)), 3644 ("sent too much")); 3645 KASSERT(sent <= tp->t_maxseg * 2, 3646 ("sent too many segments")); 3647 tp->snd_limited = 2; 3648 tcpstat.tcps_sndlimited += 2; 3649 } else if (sent > 0) { 3650 ++tp->snd_limited; 3651 ++tcpstat.tcps_sndlimited; 3652 } else if (need_early_retransmit(tp, ownd)) { 3653 ++tcpstat.tcps_sndearlyrexmit; 3654 tp->rxt_flags |= TRXT_F_EARLYREXMT; 3655 goto fastretransmit; 3656 } 3657 } 3658 return TRUE; 3659 } 3660