tcp_syncache.c revision 174558
1/*- 2 * Copyright (c) 2001 McAfee, Inc. 3 * Copyright (c) 2006 Andre Oppermann, Internet Business Solutions AG 4 * All rights reserved. 5 * 6 * This software was developed for the FreeBSD Project by Jonathan Lemon 7 * and McAfee Research, the Security Research Division of McAfee, Inc. under 8 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 9 * DARPA CHATS research program. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 */ 32 33#include <sys/cdefs.h> 34__FBSDID("$FreeBSD: head/sys/netinet/tcp_syncache.c 174558 2007-12-12 20:35:59Z kmacy $"); 35 36#include "opt_inet.h" 37#include "opt_inet6.h" 38#include "opt_ipsec.h" 39#include "opt_mac.h" 40 41#include <sys/param.h> 42#include <sys/systm.h> 43#include <sys/kernel.h> 44#include <sys/sysctl.h> 45#include <sys/lock.h> 46#include <sys/mutex.h> 47#include <sys/malloc.h> 48#include <sys/mbuf.h> 49#include <sys/md5.h> 50#include <sys/proc.h> /* for proc0 declaration */ 51#include <sys/random.h> 52#include <sys/socket.h> 53#include <sys/socketvar.h> 54#include <sys/syslog.h> 55 56#include <vm/uma.h> 57 58#include <net/if.h> 59#include <net/route.h> 60 61#include <netinet/in.h> 62#include <netinet/in_systm.h> 63#include <netinet/ip.h> 64#include <netinet/in_var.h> 65#include <netinet/in_pcb.h> 66#include <netinet/ip_var.h> 67#include <netinet/ip_options.h> 68#ifdef INET6 69#include <netinet/ip6.h> 70#include <netinet/icmp6.h> 71#include <netinet6/nd6.h> 72#include <netinet6/ip6_var.h> 73#include <netinet6/in6_pcb.h> 74#endif 75#include <netinet/tcp.h> 76#include <netinet/tcp_fsm.h> 77#include <netinet/tcp_seq.h> 78#include <netinet/tcp_timer.h> 79#include <netinet/tcp_var.h> 80#include <netinet/tcp_syncache.h> 81#include <netinet/tcp_ofld.h> 82#ifdef INET6 83#include <netinet6/tcp6_var.h> 84#endif 85 86#ifdef IPSEC 87#include <netipsec/ipsec.h> 88#ifdef INET6 89#include <netipsec/ipsec6.h> 90#endif 91#include <netipsec/key.h> 92#endif /*IPSEC*/ 93 94#include <machine/in_cksum.h> 95 96#include <security/mac/mac_framework.h> 97 98static int tcp_syncookies = 1; 99SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW, 100 &tcp_syncookies, 0, 101 "Use TCP SYN cookies if the syncache overflows"); 102 103static int tcp_syncookiesonly = 0; 104SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_RW, 105 &tcp_syncookiesonly, 0, 106 "Use only TCP SYN cookies"); 107 108#define SYNCOOKIE_SECRET_SIZE 8 /* dwords */ 109#define SYNCOOKIE_LIFETIME 16 /* seconds */ 110 111struct syncache { 112 TAILQ_ENTRY(syncache) sc_hash; 113 struct in_conninfo sc_inc; /* addresses */ 114 u_long sc_rxttime; /* retransmit time */ 115 u_int16_t sc_rxmits; /* retransmit counter */ 116 117 u_int32_t sc_tsreflect; /* timestamp to reflect */ 118 u_int32_t sc_ts; /* our timestamp to send */ 119 u_int32_t sc_tsoff; /* ts offset w/ syncookies */ 120 u_int32_t sc_flowlabel; /* IPv6 flowlabel */ 121 tcp_seq sc_irs; /* seq from peer */ 122 tcp_seq sc_iss; /* our ISS */ 123 struct mbuf *sc_ipopts; /* source route */ 124 125 u_int16_t sc_peer_mss; /* peer's MSS */ 126 u_int16_t sc_wnd; /* advertised window */ 127 u_int8_t sc_ip_ttl; /* IPv4 TTL */ 128 u_int8_t sc_ip_tos; /* IPv4 TOS */ 129 u_int8_t sc_requested_s_scale:4, 130 sc_requested_r_scale:4; 131 u_int8_t sc_flags; 132#define SCF_NOOPT 0x01 /* no TCP options */ 133#define SCF_WINSCALE 0x02 /* negotiated window scaling */ 134#define SCF_TIMESTAMP 0x04 /* negotiated timestamps */ 135 /* MSS is implicit */ 136#define SCF_UNREACH 0x10 /* icmp unreachable received */ 137#define SCF_SIGNATURE 0x20 /* send MD5 digests */ 138#define SCF_SACK 0x80 /* send SACK option */ 139#ifndef DISABLE_TCP_OFFLOAD 140 struct toe_usrreqs *sc_tu; /* TOE operations */ 141 void *sc_toepcb; /* TOE protocol block */ 142#endif 143#ifdef MAC 144 struct label *sc_label; /* MAC label reference */ 145#endif 146}; 147 148struct syncache_head { 149 struct mtx sch_mtx; 150 TAILQ_HEAD(sch_head, syncache) sch_bucket; 151 struct callout sch_timer; 152 int sch_nextc; 153 u_int sch_length; 154 u_int sch_oddeven; 155 u_int32_t sch_secbits_odd[SYNCOOKIE_SECRET_SIZE]; 156 u_int32_t sch_secbits_even[SYNCOOKIE_SECRET_SIZE]; 157 u_int sch_reseed; /* time_uptime, seconds */ 158}; 159 160static void syncache_drop(struct syncache *, struct syncache_head *); 161static void syncache_free(struct syncache *); 162static void syncache_insert(struct syncache *, struct syncache_head *); 163struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **); 164static int syncache_respond(struct syncache *); 165static struct socket *syncache_socket(struct syncache *, struct socket *, 166 struct mbuf *m); 167static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 168 int docallout); 169static void syncache_timer(void *); 170static void syncookie_generate(struct syncache_head *, struct syncache *, 171 u_int32_t *); 172static struct syncache 173 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 174 struct syncache *, struct tcpopt *, struct tcphdr *, 175 struct socket *); 176 177/* 178 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 179 * 3 retransmits corresponds to a timeout of (1 + 2 + 4 + 8 == 15) seconds, 180 * the odds are that the user has given up attempting to connect by then. 181 */ 182#define SYNCACHE_MAXREXMTS 3 183 184/* Arbitrary values */ 185#define TCP_SYNCACHE_HASHSIZE 512 186#define TCP_SYNCACHE_BUCKETLIMIT 30 187 188struct tcp_syncache { 189 struct syncache_head *hashbase; 190 uma_zone_t zone; 191 u_int hashsize; 192 u_int hashmask; 193 u_int bucket_limit; 194 u_int cache_count; /* XXX: unprotected */ 195 u_int cache_limit; 196 u_int rexmt_limit; 197 u_int hash_secret; 198}; 199static struct tcp_syncache tcp_syncache; 200 201SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache"); 202 203SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN, 204 &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache"); 205 206SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN, 207 &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache"); 208 209SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD, 210 &tcp_syncache.cache_count, 0, "Current number of entries in syncache"); 211 212SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN, 213 &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable"); 214 215SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW, 216 &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions"); 217 218int tcp_sc_rst_sock_fail = 1; 219SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, CTLFLAG_RW, 220 &tcp_sc_rst_sock_fail, 0, "Send reset on socket allocation failure"); 221 222static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 223 224#define SYNCACHE_HASH(inc, mask) \ 225 ((tcp_syncache.hash_secret ^ \ 226 (inc)->inc_faddr.s_addr ^ \ 227 ((inc)->inc_faddr.s_addr >> 16) ^ \ 228 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 229 230#define SYNCACHE_HASH6(inc, mask) \ 231 ((tcp_syncache.hash_secret ^ \ 232 (inc)->inc6_faddr.s6_addr32[0] ^ \ 233 (inc)->inc6_faddr.s6_addr32[3] ^ \ 234 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 235 236#define ENDPTS_EQ(a, b) ( \ 237 (a)->ie_fport == (b)->ie_fport && \ 238 (a)->ie_lport == (b)->ie_lport && \ 239 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \ 240 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \ 241) 242 243#define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0) 244 245#define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 246#define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 247#define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 248 249/* 250 * Requires the syncache entry to be already removed from the bucket list. 251 */ 252static void 253syncache_free(struct syncache *sc) 254{ 255 if (sc->sc_ipopts) 256 (void) m_free(sc->sc_ipopts); 257#ifdef MAC 258 mac_syncache_destroy(&sc->sc_label); 259#endif 260 261 uma_zfree(tcp_syncache.zone, sc); 262} 263 264void 265syncache_init(void) 266{ 267 int i; 268 269 tcp_syncache.cache_count = 0; 270 tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 271 tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 272 tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 273 tcp_syncache.hash_secret = arc4random(); 274 275 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 276 &tcp_syncache.hashsize); 277 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 278 &tcp_syncache.bucket_limit); 279 if (!powerof2(tcp_syncache.hashsize) || tcp_syncache.hashsize == 0) { 280 printf("WARNING: syncache hash size is not a power of 2.\n"); 281 tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 282 } 283 tcp_syncache.hashmask = tcp_syncache.hashsize - 1; 284 285 /* Set limits. */ 286 tcp_syncache.cache_limit = 287 tcp_syncache.hashsize * tcp_syncache.bucket_limit; 288 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 289 &tcp_syncache.cache_limit); 290 291 /* Allocate the hash table. */ 292 MALLOC(tcp_syncache.hashbase, struct syncache_head *, 293 tcp_syncache.hashsize * sizeof(struct syncache_head), 294 M_SYNCACHE, M_WAITOK | M_ZERO); 295 296 /* Initialize the hash buckets. */ 297 for (i = 0; i < tcp_syncache.hashsize; i++) { 298 TAILQ_INIT(&tcp_syncache.hashbase[i].sch_bucket); 299 mtx_init(&tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 300 NULL, MTX_DEF); 301 callout_init_mtx(&tcp_syncache.hashbase[i].sch_timer, 302 &tcp_syncache.hashbase[i].sch_mtx, 0); 303 tcp_syncache.hashbase[i].sch_length = 0; 304 } 305 306 /* Create the syncache entry zone. */ 307 tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 308 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 309 uma_zone_set_max(tcp_syncache.zone, tcp_syncache.cache_limit); 310} 311 312/* 313 * Inserts a syncache entry into the specified bucket row. 314 * Locks and unlocks the syncache_head autonomously. 315 */ 316static void 317syncache_insert(struct syncache *sc, struct syncache_head *sch) 318{ 319 struct syncache *sc2; 320 321 SCH_LOCK(sch); 322 323 /* 324 * Make sure that we don't overflow the per-bucket limit. 325 * If the bucket is full, toss the oldest element. 326 */ 327 if (sch->sch_length >= tcp_syncache.bucket_limit) { 328 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 329 ("sch->sch_length incorrect")); 330 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 331 syncache_drop(sc2, sch); 332 tcpstat.tcps_sc_bucketoverflow++; 333 } 334 335 /* Put it into the bucket. */ 336 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 337 sch->sch_length++; 338 339 /* Reinitialize the bucket row's timer. */ 340 syncache_timeout(sc, sch, 1); 341 342 SCH_UNLOCK(sch); 343 344 tcp_syncache.cache_count++; 345 tcpstat.tcps_sc_added++; 346} 347 348/* 349 * Remove and free entry from syncache bucket row. 350 * Expects locked syncache head. 351 */ 352static void 353syncache_drop(struct syncache *sc, struct syncache_head *sch) 354{ 355 356 SCH_LOCK_ASSERT(sch); 357 358 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 359 sch->sch_length--; 360 361#ifndef DISABLE_TCP_OFFLOAD 362 if (sc->sc_tu) 363 sc->sc_tu->tu_syncache_event(SC_DROP, sc->sc_toepcb); 364#endif 365 syncache_free(sc); 366 tcp_syncache.cache_count--; 367} 368 369/* 370 * Engage/reengage time on bucket row. 371 */ 372static void 373syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 374{ 375 sc->sc_rxttime = ticks + 376 TCPTV_RTOBASE * (tcp_backoff[sc->sc_rxmits]); 377 sc->sc_rxmits++; 378 if (sch->sch_nextc > sc->sc_rxttime) 379 sch->sch_nextc = sc->sc_rxttime; 380 if (!TAILQ_EMPTY(&sch->sch_bucket) && docallout) 381 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 382 syncache_timer, (void *)sch); 383} 384 385/* 386 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 387 * If we have retransmitted an entry the maximum number of times, expire it. 388 * One separate timer for each bucket row. 389 */ 390static void 391syncache_timer(void *xsch) 392{ 393 struct syncache_head *sch = (struct syncache_head *)xsch; 394 struct syncache *sc, *nsc; 395 int tick = ticks; 396 char *s; 397 398 /* NB: syncache_head has already been locked by the callout. */ 399 SCH_LOCK_ASSERT(sch); 400 401 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 402 /* 403 * We do not check if the listen socket still exists 404 * and accept the case where the listen socket may be 405 * gone by the time we resend the SYN/ACK. We do 406 * not expect this to happens often. If it does, 407 * then the RST will be sent by the time the remote 408 * host does the SYN/ACK->ACK. 409 */ 410 if (sc->sc_rxttime > tick) { 411 if (sc->sc_rxttime < sch->sch_nextc) 412 sch->sch_nextc = sc->sc_rxttime; 413 continue; 414 } 415 if (sc->sc_rxmits > tcp_syncache.rexmt_limit) { 416 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 417 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 418 "giving up and removing syncache entry\n", 419 s, __func__); 420 free(s, M_TCPLOG); 421 } 422 syncache_drop(sc, sch); 423 tcpstat.tcps_sc_stale++; 424 continue; 425 } 426 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 427 log(LOG_DEBUG, "%s; %s: Response timeout, " 428 "retransmitting (%u) SYN|ACK\n", 429 s, __func__, sc->sc_rxmits); 430 free(s, M_TCPLOG); 431 } 432 433 (void) syncache_respond(sc); 434 tcpstat.tcps_sc_retransmitted++; 435 syncache_timeout(sc, sch, 0); 436 } 437 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 438 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 439 syncache_timer, (void *)(sch)); 440} 441 442/* 443 * Find an entry in the syncache. 444 * Returns always with locked syncache_head plus a matching entry or NULL. 445 */ 446struct syncache * 447syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 448{ 449 struct syncache *sc; 450 struct syncache_head *sch; 451 452#ifdef INET6 453 if (inc->inc_isipv6) { 454 sch = &tcp_syncache.hashbase[ 455 SYNCACHE_HASH6(inc, tcp_syncache.hashmask)]; 456 *schp = sch; 457 458 SCH_LOCK(sch); 459 460 /* Circle through bucket row to find matching entry. */ 461 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 462 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 463 return (sc); 464 } 465 } else 466#endif 467 { 468 sch = &tcp_syncache.hashbase[ 469 SYNCACHE_HASH(inc, tcp_syncache.hashmask)]; 470 *schp = sch; 471 472 SCH_LOCK(sch); 473 474 /* Circle through bucket row to find matching entry. */ 475 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 476#ifdef INET6 477 if (sc->sc_inc.inc_isipv6) 478 continue; 479#endif 480 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 481 return (sc); 482 } 483 } 484 SCH_LOCK_ASSERT(*schp); 485 return (NULL); /* always returns with locked sch */ 486} 487 488/* 489 * This function is called when we get a RST for a 490 * non-existent connection, so that we can see if the 491 * connection is in the syn cache. If it is, zap it. 492 */ 493void 494syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 495{ 496 struct syncache *sc; 497 struct syncache_head *sch; 498 char *s = NULL; 499 500 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 501 SCH_LOCK_ASSERT(sch); 502 503 /* 504 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 505 * See RFC 793 page 65, section SEGMENT ARRIVES. 506 */ 507 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 508 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 509 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 510 "FIN flag set, segment ignored\n", s, __func__); 511 tcpstat.tcps_badrst++; 512 goto done; 513 } 514 515 /* 516 * No corresponding connection was found in syncache. 517 * If syncookies are enabled and possibly exclusively 518 * used, or we are under memory pressure, a valid RST 519 * may not find a syncache entry. In that case we're 520 * done and no SYN|ACK retransmissions will happen. 521 * Otherwise the the RST was misdirected or spoofed. 522 */ 523 if (sc == NULL) { 524 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 525 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 526 "syncache entry (possibly syncookie only), " 527 "segment ignored\n", s, __func__); 528 tcpstat.tcps_badrst++; 529 goto done; 530 } 531 532 /* 533 * If the RST bit is set, check the sequence number to see 534 * if this is a valid reset segment. 535 * RFC 793 page 37: 536 * In all states except SYN-SENT, all reset (RST) segments 537 * are validated by checking their SEQ-fields. A reset is 538 * valid if its sequence number is in the window. 539 * 540 * The sequence number in the reset segment is normally an 541 * echo of our outgoing acknowlegement numbers, but some hosts 542 * send a reset with the sequence number at the rightmost edge 543 * of our receive window, and we have to handle this case. 544 */ 545 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 546 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 547 syncache_drop(sc, sch); 548 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 549 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 550 "connection attempt aborted by remote endpoint\n", 551 s, __func__); 552 tcpstat.tcps_sc_reset++; 553 } else if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 554 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != IRS %u " 555 "(+WND %u), segment ignored\n", 556 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 557 tcpstat.tcps_badrst++; 558 } 559 560done: 561 if (s != NULL) 562 free(s, M_TCPLOG); 563 SCH_UNLOCK(sch); 564} 565 566void 567syncache_badack(struct in_conninfo *inc) 568{ 569 struct syncache *sc; 570 struct syncache_head *sch; 571 572 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 573 SCH_LOCK_ASSERT(sch); 574 if (sc != NULL) { 575 syncache_drop(sc, sch); 576 tcpstat.tcps_sc_badack++; 577 } 578 SCH_UNLOCK(sch); 579} 580 581void 582syncache_unreach(struct in_conninfo *inc, struct tcphdr *th) 583{ 584 struct syncache *sc; 585 struct syncache_head *sch; 586 587 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 588 SCH_LOCK_ASSERT(sch); 589 if (sc == NULL) 590 goto done; 591 592 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 593 if (ntohl(th->th_seq) != sc->sc_iss) 594 goto done; 595 596 /* 597 * If we've rertransmitted 3 times and this is our second error, 598 * we remove the entry. Otherwise, we allow it to continue on. 599 * This prevents us from incorrectly nuking an entry during a 600 * spurious network outage. 601 * 602 * See tcp_notify(). 603 */ 604 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 605 sc->sc_flags |= SCF_UNREACH; 606 goto done; 607 } 608 syncache_drop(sc, sch); 609 tcpstat.tcps_sc_unreach++; 610done: 611 SCH_UNLOCK(sch); 612} 613 614/* 615 * Build a new TCP socket structure from a syncache entry. 616 */ 617static struct socket * 618syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 619{ 620 struct inpcb *inp = NULL; 621 struct socket *so; 622 struct tcpcb *tp; 623 char *s; 624 625 INP_INFO_WLOCK_ASSERT(&tcbinfo); 626 627 /* 628 * Ok, create the full blown connection, and set things up 629 * as they would have been set up if we had created the 630 * connection when the SYN arrived. If we can't create 631 * the connection, abort it. 632 */ 633 so = sonewconn(lso, SS_ISCONNECTED); 634 if (so == NULL) { 635 /* 636 * Drop the connection; we will either send a RST or 637 * have the peer retransmit its SYN again after its 638 * RTO and try again. 639 */ 640 tcpstat.tcps_listendrop++; 641 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 642 log(LOG_DEBUG, "%s; %s: Socket create failed " 643 "due to limits or memory shortage\n", 644 s, __func__); 645 free(s, M_TCPLOG); 646 } 647 goto abort2; 648 } 649#ifdef MAC 650 SOCK_LOCK(so); 651 mac_socketpeer_set_from_mbuf(m, so); 652 SOCK_UNLOCK(so); 653#endif 654 655 inp = sotoinpcb(so); 656 INP_LOCK(inp); 657 658 /* Insert new socket into PCB hash list. */ 659 inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6; 660#ifdef INET6 661 if (sc->sc_inc.inc_isipv6) { 662 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 663 } else { 664 inp->inp_vflag &= ~INP_IPV6; 665 inp->inp_vflag |= INP_IPV4; 666#endif 667 inp->inp_laddr = sc->sc_inc.inc_laddr; 668#ifdef INET6 669 } 670#endif 671 inp->inp_lport = sc->sc_inc.inc_lport; 672 if (in_pcbinshash(inp) != 0) { 673 /* 674 * Undo the assignments above if we failed to 675 * put the PCB on the hash lists. 676 */ 677#ifdef INET6 678 if (sc->sc_inc.inc_isipv6) 679 inp->in6p_laddr = in6addr_any; 680 else 681#endif 682 inp->inp_laddr.s_addr = INADDR_ANY; 683 inp->inp_lport = 0; 684 goto abort; 685 } 686#ifdef IPSEC 687 /* Copy old policy into new socket's. */ 688 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp)) 689 printf("syncache_socket: could not copy policy\n"); 690#endif 691#ifdef INET6 692 if (sc->sc_inc.inc_isipv6) { 693 struct inpcb *oinp = sotoinpcb(lso); 694 struct in6_addr laddr6; 695 struct sockaddr_in6 sin6; 696 /* 697 * Inherit socket options from the listening socket. 698 * Note that in6p_inputopts are not (and should not be) 699 * copied, since it stores previously received options and is 700 * used to detect if each new option is different than the 701 * previous one and hence should be passed to a user. 702 * If we copied in6p_inputopts, a user would not be able to 703 * receive options just after calling the accept system call. 704 */ 705 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 706 if (oinp->in6p_outputopts) 707 inp->in6p_outputopts = 708 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 709 710 sin6.sin6_family = AF_INET6; 711 sin6.sin6_len = sizeof(sin6); 712 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 713 sin6.sin6_port = sc->sc_inc.inc_fport; 714 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 715 laddr6 = inp->in6p_laddr; 716 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 717 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 718 if (in6_pcbconnect(inp, (struct sockaddr *)&sin6, 719 thread0.td_ucred)) { 720 inp->in6p_laddr = laddr6; 721 goto abort; 722 } 723 /* Override flowlabel from in6_pcbconnect. */ 724 inp->in6p_flowinfo &= ~IPV6_FLOWLABEL_MASK; 725 inp->in6p_flowinfo |= sc->sc_flowlabel; 726 } else 727#endif 728 { 729 struct in_addr laddr; 730 struct sockaddr_in sin; 731 732 inp->inp_options = ip_srcroute(m); 733 if (inp->inp_options == NULL) { 734 inp->inp_options = sc->sc_ipopts; 735 sc->sc_ipopts = NULL; 736 } 737 738 sin.sin_family = AF_INET; 739 sin.sin_len = sizeof(sin); 740 sin.sin_addr = sc->sc_inc.inc_faddr; 741 sin.sin_port = sc->sc_inc.inc_fport; 742 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 743 laddr = inp->inp_laddr; 744 if (inp->inp_laddr.s_addr == INADDR_ANY) 745 inp->inp_laddr = sc->sc_inc.inc_laddr; 746 if (in_pcbconnect(inp, (struct sockaddr *)&sin, 747 thread0.td_ucred)) { 748 inp->inp_laddr = laddr; 749 goto abort; 750 } 751 } 752 tp = intotcpcb(inp); 753 tp->t_state = TCPS_SYN_RECEIVED; 754 tp->iss = sc->sc_iss; 755 tp->irs = sc->sc_irs; 756 tcp_rcvseqinit(tp); 757 tcp_sendseqinit(tp); 758 tp->snd_wl1 = sc->sc_irs; 759 tp->snd_max = tp->iss + 1; 760 tp->snd_nxt = tp->iss + 1; 761 tp->rcv_up = sc->sc_irs + 1; 762 tp->rcv_wnd = sc->sc_wnd; 763 tp->rcv_adv += tp->rcv_wnd; 764 tp->last_ack_sent = tp->rcv_nxt; 765 766 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 767 if (sc->sc_flags & SCF_NOOPT) 768 tp->t_flags |= TF_NOOPT; 769 else { 770 if (sc->sc_flags & SCF_WINSCALE) { 771 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 772 tp->snd_scale = sc->sc_requested_s_scale; 773 tp->request_r_scale = sc->sc_requested_r_scale; 774 } 775 if (sc->sc_flags & SCF_TIMESTAMP) { 776 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 777 tp->ts_recent = sc->sc_tsreflect; 778 tp->ts_recent_age = ticks; 779 tp->ts_offset = sc->sc_tsoff; 780 } 781#ifdef TCP_SIGNATURE 782 if (sc->sc_flags & SCF_SIGNATURE) 783 tp->t_flags |= TF_SIGNATURE; 784#endif 785 if (sc->sc_flags & SCF_SACK) 786 tp->t_flags |= TF_SACK_PERMIT; 787 } 788 789 /* 790 * Set up MSS and get cached values from tcp_hostcache. 791 * This might overwrite some of the defaults we just set. 792 */ 793 tcp_mss(tp, sc->sc_peer_mss); 794 795 /* 796 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment. 797 */ 798 if (sc->sc_rxmits) 799 tp->snd_cwnd = tp->t_maxseg; 800 tcp_timer_activate(tp, TT_KEEP, tcp_keepinit); 801 802 INP_UNLOCK(inp); 803 804 tcpstat.tcps_accepts++; 805 return (so); 806 807abort: 808 INP_UNLOCK(inp); 809abort2: 810 if (so != NULL) 811 soabort(so); 812 return (NULL); 813} 814 815/* 816 * This function gets called when we receive an ACK for a 817 * socket in the LISTEN state. We look up the connection 818 * in the syncache, and if its there, we pull it out of 819 * the cache and turn it into a full-blown connection in 820 * the SYN-RECEIVED state. 821 */ 822int 823syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 824 struct socket **lsop, struct mbuf *m) 825{ 826 struct syncache *sc; 827 struct syncache_head *sch; 828 struct syncache scs; 829 char *s; 830 831 /* 832 * Global TCP locks are held because we manipulate the PCB lists 833 * and create a new socket. 834 */ 835 INP_INFO_WLOCK_ASSERT(&tcbinfo); 836 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 837 ("%s: can handle only ACK", __func__)); 838 839 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 840 SCH_LOCK_ASSERT(sch); 841 if (sc == NULL) { 842 /* 843 * There is no syncache entry, so see if this ACK is 844 * a returning syncookie. To do this, first: 845 * A. See if this socket has had a syncache entry dropped in 846 * the past. We don't want to accept a bogus syncookie 847 * if we've never received a SYN. 848 * B. check that the syncookie is valid. If it is, then 849 * cobble up a fake syncache entry, and return. 850 */ 851 if (!tcp_syncookies) { 852 SCH_UNLOCK(sch); 853 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 854 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 855 "segment rejected (syncookies disabled)\n", 856 s, __func__); 857 goto failed; 858 } 859 bzero(&scs, sizeof(scs)); 860 sc = syncookie_lookup(inc, sch, &scs, to, th, *lsop); 861 SCH_UNLOCK(sch); 862 if (sc == NULL) { 863 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 864 log(LOG_DEBUG, "%s; %s: Segment failed " 865 "SYNCOOKIE authentication, segment rejected " 866 "(probably spoofed)\n", s, __func__); 867 goto failed; 868 } 869 } else { 870 /* Pull out the entry to unlock the bucket row. */ 871 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 872 sch->sch_length--; 873 tcp_syncache.cache_count--; 874 SCH_UNLOCK(sch); 875 } 876 877 /* 878 * Segment validation: 879 * ACK must match our initial sequence number + 1 (the SYN|ACK). 880 */ 881 if (th->th_ack != sc->sc_iss + 1 && sc->sc_toepcb == NULL) { 882 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 883 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 884 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 885 goto failed; 886 } 887 /* 888 * The SEQ must match the received initial receive sequence 889 * number + 1 (the SYN) because we didn't ACK any data that 890 * may have come with the SYN. 891 */ 892 if (th->th_seq != sc->sc_irs + 1 && sc->sc_toepcb == NULL) { 893 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 894 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 895 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 896 goto failed; 897 } 898 899 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 900 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 901 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 902 "segment rejected\n", s, __func__); 903 goto failed; 904 } 905 /* 906 * If timestamps were negotiated the reflected timestamp 907 * must be equal to what we actually sent in the SYN|ACK. 908 */ 909 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts && 910 sc->sc_toepcb == NULL) { 911 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 912 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 913 "segment rejected\n", 914 s, __func__, to->to_tsecr, sc->sc_ts); 915 goto failed; 916 } 917 918 *lsop = syncache_socket(sc, *lsop, m); 919 920 if (*lsop == NULL) 921 tcpstat.tcps_sc_aborted++; 922 else 923 tcpstat.tcps_sc_completed++; 924 925 if (sc != &scs) 926 syncache_free(sc); 927 return (1); 928failed: 929 if (sc != NULL && sc != &scs) 930 syncache_free(sc); 931 if (s != NULL) 932 free(s, M_TCPLOG); 933 *lsop = NULL; 934 return (0); 935} 936 937/* 938 * Given a LISTEN socket and an inbound SYN request, add 939 * this to the syn cache, and send back a segment: 940 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 941 * to the source. 942 * 943 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 944 * Doing so would require that we hold onto the data and deliver it 945 * to the application. However, if we are the target of a SYN-flood 946 * DoS attack, an attacker could send data which would eventually 947 * consume all available buffer space if it were ACKed. By not ACKing 948 * the data, we avoid this DoS scenario. 949 */ 950static void 951_syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 952 struct inpcb *inp, struct socket **lsop, struct mbuf *m, 953 struct toe_usrreqs *tu, void *toepcb) 954{ 955 struct tcpcb *tp; 956 struct socket *so; 957 struct syncache *sc = NULL; 958 struct syncache_head *sch; 959 struct mbuf *ipopts = NULL; 960 u_int32_t flowtmp; 961 int win, sb_hiwat, ip_ttl, ip_tos, noopt; 962 char *s; 963#ifdef INET6 964 int autoflowlabel = 0; 965#endif 966#ifdef MAC 967 struct label *maclabel; 968#endif 969 struct syncache scs; 970 971 INP_INFO_WLOCK_ASSERT(&tcbinfo); 972 INP_LOCK_ASSERT(inp); /* listen socket */ 973 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 974 ("%s: unexpected tcp flags", __func__)); 975 976 /* 977 * Combine all so/tp operations very early to drop the INP lock as 978 * soon as possible. 979 */ 980 so = *lsop; 981 tp = sototcpcb(so); 982 983#ifdef INET6 984 if (inc->inc_isipv6 && 985 (inp->in6p_flags & IN6P_AUTOFLOWLABEL)) 986 autoflowlabel = 1; 987#endif 988 ip_ttl = inp->inp_ip_ttl; 989 ip_tos = inp->inp_ip_tos; 990 win = sbspace(&so->so_rcv); 991 sb_hiwat = so->so_rcv.sb_hiwat; 992 noopt = (tp->t_flags & TF_NOOPT); 993 994 so = NULL; 995 tp = NULL; 996 997#ifdef MAC 998 if (mac_syncache_init(&maclabel) != 0) { 999 INP_UNLOCK(inp); 1000 INP_INFO_WUNLOCK(&tcbinfo); 1001 goto done; 1002 } else 1003 mac_syncache_create(maclabel, inp); 1004#endif 1005 INP_UNLOCK(inp); 1006 INP_INFO_WUNLOCK(&tcbinfo); 1007 1008 /* 1009 * Remember the IP options, if any. 1010 */ 1011#ifdef INET6 1012 if (!inc->inc_isipv6) 1013#endif 1014 ipopts = ip_srcroute(m); 1015 1016 /* 1017 * See if we already have an entry for this connection. 1018 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1019 * 1020 * XXX: should the syncache be re-initialized with the contents 1021 * of the new SYN here (which may have different options?) 1022 * 1023 * XXX: We do not check the sequence number to see if this is a 1024 * real retransmit or a new connection attempt. The question is 1025 * how to handle such a case; either ignore it as spoofed, or 1026 * drop the current entry and create a new one? 1027 */ 1028 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1029 SCH_LOCK_ASSERT(sch); 1030 if (sc != NULL) { 1031#ifndef DISABLE_TCP_OFFLOAD 1032 if (sc->sc_tu) 1033 sc->sc_tu->tu_syncache_event(SC_ENTRY_PRESENT, 1034 sc->sc_toepcb); 1035#endif 1036 tcpstat.tcps_sc_dupsyn++; 1037 if (ipopts) { 1038 /* 1039 * If we were remembering a previous source route, 1040 * forget it and use the new one we've been given. 1041 */ 1042 if (sc->sc_ipopts) 1043 (void) m_free(sc->sc_ipopts); 1044 sc->sc_ipopts = ipopts; 1045 } 1046 /* 1047 * Update timestamp if present. 1048 */ 1049 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1050 sc->sc_tsreflect = to->to_tsval; 1051 else 1052 sc->sc_flags &= ~SCF_TIMESTAMP; 1053#ifdef MAC 1054 /* 1055 * Since we have already unconditionally allocated label 1056 * storage, free it up. The syncache entry will already 1057 * have an initialized label we can use. 1058 */ 1059 mac_syncache_destroy(&maclabel); 1060 KASSERT(sc->sc_label != NULL, 1061 ("%s: label not initialized", __func__)); 1062#endif 1063 /* Retransmit SYN|ACK and reset retransmit count. */ 1064 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1065 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1066 "resetting timer and retransmitting SYN|ACK\n", 1067 s, __func__); 1068 free(s, M_TCPLOG); 1069 } 1070 if ((sc->sc_toepcb == NULL) && syncache_respond(sc) == 0) { 1071 sc->sc_rxmits = 0; 1072 syncache_timeout(sc, sch, 1); 1073 tcpstat.tcps_sndacks++; 1074 tcpstat.tcps_sndtotal++; 1075 } 1076 SCH_UNLOCK(sch); 1077 goto done; 1078 } 1079 1080 sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT | M_ZERO); 1081 if (sc == NULL) { 1082 /* 1083 * The zone allocator couldn't provide more entries. 1084 * Treat this as if the cache was full; drop the oldest 1085 * entry and insert the new one. 1086 */ 1087 tcpstat.tcps_sc_zonefail++; 1088 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) 1089 syncache_drop(sc, sch); 1090 sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT | M_ZERO); 1091 if (sc == NULL) { 1092 if (tcp_syncookies) { 1093 bzero(&scs, sizeof(scs)); 1094 sc = &scs; 1095 } else { 1096 SCH_UNLOCK(sch); 1097 if (ipopts) 1098 (void) m_free(ipopts); 1099 goto done; 1100 } 1101 } 1102 } 1103 1104 /* 1105 * Fill in the syncache values. 1106 */ 1107#ifdef MAC 1108 sc->sc_label = maclabel; 1109#endif 1110 sc->sc_ipopts = ipopts; 1111 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1112#ifdef INET6 1113 if (!inc->inc_isipv6) 1114#endif 1115 { 1116 sc->sc_ip_tos = ip_tos; 1117 sc->sc_ip_ttl = ip_ttl; 1118 } 1119#ifndef DISABLE_TCP_OFFLOAD 1120 sc->sc_tu = tu; 1121 sc->sc_toepcb = toepcb; 1122#endif 1123 sc->sc_irs = th->th_seq; 1124 sc->sc_iss = arc4random(); 1125 sc->sc_flags = 0; 1126 sc->sc_flowlabel = 0; 1127 1128 /* 1129 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1130 * win was derived from socket earlier in the function. 1131 */ 1132 win = imax(win, 0); 1133 win = imin(win, TCP_MAXWIN); 1134 sc->sc_wnd = win; 1135 1136 if (tcp_do_rfc1323) { 1137 /* 1138 * A timestamp received in a SYN makes 1139 * it ok to send timestamp requests and replies. 1140 */ 1141 if (to->to_flags & TOF_TS) { 1142 sc->sc_tsreflect = to->to_tsval; 1143 sc->sc_ts = ticks; 1144 sc->sc_flags |= SCF_TIMESTAMP; 1145 } 1146 if (to->to_flags & TOF_SCALE) { 1147 int wscale = 0; 1148 1149 /* 1150 * Pick the smallest possible scaling factor that 1151 * will still allow us to scale up to sb_max, aka 1152 * kern.ipc.maxsockbuf. 1153 * 1154 * We do this because there are broken firewalls that 1155 * will corrupt the window scale option, leading to 1156 * the other endpoint believing that our advertised 1157 * window is unscaled. At scale factors larger than 1158 * 5 the unscaled window will drop below 1500 bytes, 1159 * leading to serious problems when traversing these 1160 * broken firewalls. 1161 * 1162 * With the default maxsockbuf of 256K, a scale factor 1163 * of 3 will be chosen by this algorithm. Those who 1164 * choose a larger maxsockbuf should watch out 1165 * for the compatiblity problems mentioned above. 1166 * 1167 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1168 * or <SYN,ACK>) segment itself is never scaled. 1169 */ 1170 while (wscale < TCP_MAX_WINSHIFT && 1171 (TCP_MAXWIN << wscale) < sb_max) 1172 wscale++; 1173 sc->sc_requested_r_scale = wscale; 1174 sc->sc_requested_s_scale = to->to_wscale; 1175 sc->sc_flags |= SCF_WINSCALE; 1176 } 1177 } 1178#ifdef TCP_SIGNATURE 1179 /* 1180 * If listening socket requested TCP digests, and received SYN 1181 * contains the option, flag this in the syncache so that 1182 * syncache_respond() will do the right thing with the SYN+ACK. 1183 * XXX: Currently we always record the option by default and will 1184 * attempt to use it in syncache_respond(). 1185 */ 1186 if (to->to_flags & TOF_SIGNATURE) 1187 sc->sc_flags |= SCF_SIGNATURE; 1188#endif 1189 if (to->to_flags & TOF_SACKPERM) 1190 sc->sc_flags |= SCF_SACK; 1191 if (to->to_flags & TOF_MSS) 1192 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1193 if (noopt) 1194 sc->sc_flags |= SCF_NOOPT; 1195 1196 if (tcp_syncookies) { 1197 syncookie_generate(sch, sc, &flowtmp); 1198#ifdef INET6 1199 if (autoflowlabel) 1200 sc->sc_flowlabel = flowtmp; 1201#endif 1202 } else { 1203#ifdef INET6 1204 if (autoflowlabel) 1205 sc->sc_flowlabel = 1206 (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK); 1207#endif 1208 } 1209 SCH_UNLOCK(sch); 1210 1211 /* 1212 * Do a standard 3-way handshake. 1213 */ 1214 if (sc->sc_toepcb || syncache_respond(sc) == 0) { 1215 if (tcp_syncookies && tcp_syncookiesonly && sc != &scs) 1216 syncache_free(sc); 1217 else if (sc != &scs) 1218 syncache_insert(sc, sch); /* locks and unlocks sch */ 1219 tcpstat.tcps_sndacks++; 1220 tcpstat.tcps_sndtotal++; 1221 } else { 1222 if (sc != &scs) 1223 syncache_free(sc); 1224 tcpstat.tcps_sc_dropped++; 1225 } 1226 1227done: 1228#ifdef MAC 1229 if (sc == &scs) 1230 mac_syncache_destroy(&maclabel); 1231#endif 1232 if (m) { 1233 1234 *lsop = NULL; 1235 m_freem(m); 1236 } 1237 return; 1238} 1239 1240static int 1241syncache_respond(struct syncache *sc) 1242{ 1243 struct ip *ip = NULL; 1244 struct mbuf *m; 1245 struct tcphdr *th; 1246 int optlen, error; 1247 u_int16_t hlen, tlen, mssopt; 1248 struct tcpopt to; 1249#ifdef INET6 1250 struct ip6_hdr *ip6 = NULL; 1251#endif 1252 1253 hlen = 1254#ifdef INET6 1255 (sc->sc_inc.inc_isipv6) ? sizeof(struct ip6_hdr) : 1256#endif 1257 sizeof(struct ip); 1258 tlen = hlen + sizeof(struct tcphdr); 1259 1260 /* Determine MSS we advertize to other end of connection. */ 1261 mssopt = tcp_mssopt(&sc->sc_inc); 1262 if (sc->sc_peer_mss) 1263 mssopt = max( min(sc->sc_peer_mss, mssopt), tcp_minmss); 1264 1265 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1266 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1267 ("syncache: mbuf too small")); 1268 1269 /* Create the IP+TCP header from scratch. */ 1270 m = m_gethdr(M_DONTWAIT, MT_DATA); 1271 if (m == NULL) 1272 return (ENOBUFS); 1273#ifdef MAC 1274 mac_syncache_create_mbuf(sc->sc_label, m); 1275#endif 1276 m->m_data += max_linkhdr; 1277 m->m_len = tlen; 1278 m->m_pkthdr.len = tlen; 1279 m->m_pkthdr.rcvif = NULL; 1280 1281#ifdef INET6 1282 if (sc->sc_inc.inc_isipv6) { 1283 ip6 = mtod(m, struct ip6_hdr *); 1284 ip6->ip6_vfc = IPV6_VERSION; 1285 ip6->ip6_nxt = IPPROTO_TCP; 1286 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1287 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1288 ip6->ip6_plen = htons(tlen - hlen); 1289 /* ip6_hlim is set after checksum */ 1290 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1291 ip6->ip6_flow |= sc->sc_flowlabel; 1292 1293 th = (struct tcphdr *)(ip6 + 1); 1294 } else 1295#endif 1296 { 1297 ip = mtod(m, struct ip *); 1298 ip->ip_v = IPVERSION; 1299 ip->ip_hl = sizeof(struct ip) >> 2; 1300 ip->ip_len = tlen; 1301 ip->ip_id = 0; 1302 ip->ip_off = 0; 1303 ip->ip_sum = 0; 1304 ip->ip_p = IPPROTO_TCP; 1305 ip->ip_src = sc->sc_inc.inc_laddr; 1306 ip->ip_dst = sc->sc_inc.inc_faddr; 1307 ip->ip_ttl = sc->sc_ip_ttl; 1308 ip->ip_tos = sc->sc_ip_tos; 1309 1310 /* 1311 * See if we should do MTU discovery. Route lookups are 1312 * expensive, so we will only unset the DF bit if: 1313 * 1314 * 1) path_mtu_discovery is disabled 1315 * 2) the SCF_UNREACH flag has been set 1316 */ 1317 if (path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1318 ip->ip_off |= IP_DF; 1319 1320 th = (struct tcphdr *)(ip + 1); 1321 } 1322 th->th_sport = sc->sc_inc.inc_lport; 1323 th->th_dport = sc->sc_inc.inc_fport; 1324 1325 th->th_seq = htonl(sc->sc_iss); 1326 th->th_ack = htonl(sc->sc_irs + 1); 1327 th->th_off = sizeof(struct tcphdr) >> 2; 1328 th->th_x2 = 0; 1329 th->th_flags = TH_SYN|TH_ACK; 1330 th->th_win = htons(sc->sc_wnd); 1331 th->th_urp = 0; 1332 1333 /* Tack on the TCP options. */ 1334 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1335 to.to_flags = 0; 1336 1337 to.to_mss = mssopt; 1338 to.to_flags = TOF_MSS; 1339 if (sc->sc_flags & SCF_WINSCALE) { 1340 to.to_wscale = sc->sc_requested_r_scale; 1341 to.to_flags |= TOF_SCALE; 1342 } 1343 if (sc->sc_flags & SCF_TIMESTAMP) { 1344 /* Virgin timestamp or TCP cookie enhanced one. */ 1345 to.to_tsval = sc->sc_ts; 1346 to.to_tsecr = sc->sc_tsreflect; 1347 to.to_flags |= TOF_TS; 1348 } 1349 if (sc->sc_flags & SCF_SACK) 1350 to.to_flags |= TOF_SACKPERM; 1351#ifdef TCP_SIGNATURE 1352 if (sc->sc_flags & SCF_SIGNATURE) 1353 to.to_flags |= TOF_SIGNATURE; 1354#endif 1355 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1356 1357 /* Adjust headers by option size. */ 1358 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1359 m->m_len += optlen; 1360 m->m_pkthdr.len += optlen; 1361 1362#ifdef TCP_SIGNATURE 1363 if (sc->sc_flags & SCF_SIGNATURE) 1364 tcp_signature_compute(m, sizeof(struct ip), 0, optlen, 1365 to.to_signature, IPSEC_DIR_OUTBOUND); 1366#endif 1367#ifdef INET6 1368 if (sc->sc_inc.inc_isipv6) 1369 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1370 else 1371#endif 1372 ip->ip_len += optlen; 1373 } else 1374 optlen = 0; 1375 1376#ifdef INET6 1377 if (sc->sc_inc.inc_isipv6) { 1378 th->th_sum = 0; 1379 th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, 1380 tlen + optlen - hlen); 1381 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1382 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1383 } else 1384#endif 1385 { 1386 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1387 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1388 m->m_pkthdr.csum_flags = CSUM_TCP; 1389 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1390 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1391 } 1392 return (error); 1393} 1394 1395void 1396syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1397 struct inpcb *inp, struct socket **lsop, struct mbuf *m) 1398{ 1399 _syncache_add(inc, to, th, inp, lsop, m, NULL, NULL); 1400} 1401 1402void 1403syncache_offload_add(struct in_conninfo *inc, struct tcpopt *to, 1404 struct tcphdr *th, struct inpcb *inp, struct socket **lsop, 1405 struct toe_usrreqs *tu, void *toepcb) 1406{ 1407 _syncache_add(inc, to, th, inp, lsop, NULL, tu, toepcb); 1408} 1409 1410/* 1411 * The purpose of SYN cookies is to avoid keeping track of all SYN's we 1412 * receive and to be able to handle SYN floods from bogus source addresses 1413 * (where we will never receive any reply). SYN floods try to exhaust all 1414 * our memory and available slots in the SYN cache table to cause a denial 1415 * of service to legitimate users of the local host. 1416 * 1417 * The idea of SYN cookies is to encode and include all necessary information 1418 * about the connection setup state within the SYN-ACK we send back and thus 1419 * to get along without keeping any local state until the ACK to the SYN-ACK 1420 * arrives (if ever). Everything we need to know should be available from 1421 * the information we encoded in the SYN-ACK. 1422 * 1423 * More information about the theory behind SYN cookies and its first 1424 * discussion and specification can be found at: 1425 * http://cr.yp.to/syncookies.html (overview) 1426 * http://cr.yp.to/syncookies/archive (gory details) 1427 * 1428 * This implementation extends the orginal idea and first implementation 1429 * of FreeBSD by using not only the initial sequence number field to store 1430 * information but also the timestamp field if present. This way we can 1431 * keep track of the entire state we need to know to recreate the session in 1432 * its original form. Almost all TCP speakers implement RFC1323 timestamps 1433 * these days. For those that do not we still have to live with the known 1434 * shortcomings of the ISN only SYN cookies. 1435 * 1436 * Cookie layers: 1437 * 1438 * Initial sequence number we send: 1439 * 31|................................|0 1440 * DDDDDDDDDDDDDDDDDDDDDDDDDMMMRRRP 1441 * D = MD5 Digest (first dword) 1442 * M = MSS index 1443 * R = Rotation of secret 1444 * P = Odd or Even secret 1445 * 1446 * The MD5 Digest is computed with over following parameters: 1447 * a) randomly rotated secret 1448 * b) struct in_conninfo containing the remote/local ip/port (IPv4&IPv6) 1449 * c) the received initial sequence number from remote host 1450 * d) the rotation offset and odd/even bit 1451 * 1452 * Timestamp we send: 1453 * 31|................................|0 1454 * DDDDDDDDDDDDDDDDDDDDDDSSSSRRRRA5 1455 * D = MD5 Digest (third dword) (only as filler) 1456 * S = Requested send window scale 1457 * R = Requested receive window scale 1458 * A = SACK allowed 1459 * 5 = TCP-MD5 enabled (not implemented yet) 1460 * XORed with MD5 Digest (forth dword) 1461 * 1462 * The timestamp isn't cryptographically secure and doesn't need to be. 1463 * The double use of the MD5 digest dwords ties it to a specific remote/ 1464 * local host/port, remote initial sequence number and our local time 1465 * limited secret. A received timestamp is reverted (XORed) and then 1466 * the contained MD5 dword is compared to the computed one to ensure the 1467 * timestamp belongs to the SYN-ACK we sent. The other parameters may 1468 * have been tampered with but this isn't different from supplying bogus 1469 * values in the SYN in the first place. 1470 * 1471 * Some problems with SYN cookies remain however: 1472 * Consider the problem of a recreated (and retransmitted) cookie. If the 1473 * original SYN was accepted, the connection is established. The second 1474 * SYN is inflight, and if it arrives with an ISN that falls within the 1475 * receive window, the connection is killed. 1476 * 1477 * Notes: 1478 * A heuristic to determine when to accept syn cookies is not necessary. 1479 * An ACK flood would cause the syncookie verification to be attempted, 1480 * but a SYN flood causes syncookies to be generated. Both are of equal 1481 * cost, so there's no point in trying to optimize the ACK flood case. 1482 * Also, if you don't process certain ACKs for some reason, then all someone 1483 * would have to do is launch a SYN and ACK flood at the same time, which 1484 * would stop cookie verification and defeat the entire purpose of syncookies. 1485 */ 1486static int tcp_sc_msstab[] = { 0, 256, 468, 536, 996, 1452, 1460, 8960 }; 1487 1488static void 1489syncookie_generate(struct syncache_head *sch, struct syncache *sc, 1490 u_int32_t *flowlabel) 1491{ 1492 MD5_CTX ctx; 1493 u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)]; 1494 u_int32_t data; 1495 u_int32_t *secbits; 1496 u_int off, pmss, mss; 1497 int i; 1498 1499 SCH_LOCK_ASSERT(sch); 1500 1501 /* Which of the two secrets to use. */ 1502 secbits = sch->sch_oddeven ? 1503 sch->sch_secbits_odd : sch->sch_secbits_even; 1504 1505 /* Reseed secret if too old. */ 1506 if (sch->sch_reseed < time_uptime) { 1507 sch->sch_oddeven = sch->sch_oddeven ? 0 : 1; /* toggle */ 1508 secbits = sch->sch_oddeven ? 1509 sch->sch_secbits_odd : sch->sch_secbits_even; 1510 for (i = 0; i < SYNCOOKIE_SECRET_SIZE; i++) 1511 secbits[i] = arc4random(); 1512 sch->sch_reseed = time_uptime + SYNCOOKIE_LIFETIME; 1513 } 1514 1515 /* Secret rotation offset. */ 1516 off = sc->sc_iss & 0x7; /* iss was randomized before */ 1517 1518 /* Maximum segment size calculation. */ 1519 pmss = max( min(sc->sc_peer_mss, tcp_mssopt(&sc->sc_inc)), tcp_minmss); 1520 for (mss = sizeof(tcp_sc_msstab) / sizeof(int) - 1; mss > 0; mss--) 1521 if (tcp_sc_msstab[mss] <= pmss) 1522 break; 1523 1524 /* Fold parameters and MD5 digest into the ISN we will send. */ 1525 data = sch->sch_oddeven;/* odd or even secret, 1 bit */ 1526 data |= off << 1; /* secret offset, derived from iss, 3 bits */ 1527 data |= mss << 4; /* mss, 3 bits */ 1528 1529 MD5Init(&ctx); 1530 MD5Update(&ctx, ((u_int8_t *)secbits) + off, 1531 SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off); 1532 MD5Update(&ctx, secbits, off); 1533 MD5Update(&ctx, &sc->sc_inc, sizeof(sc->sc_inc)); 1534 MD5Update(&ctx, &sc->sc_irs, sizeof(sc->sc_irs)); 1535 MD5Update(&ctx, &data, sizeof(data)); 1536 MD5Final((u_int8_t *)&md5_buffer, &ctx); 1537 1538 data |= (md5_buffer[0] << 7); 1539 sc->sc_iss = data; 1540 1541#ifdef INET6 1542 *flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK; 1543#endif 1544 1545 /* Additional parameters are stored in the timestamp if present. */ 1546 if (sc->sc_flags & SCF_TIMESTAMP) { 1547 data = ((sc->sc_flags & SCF_SIGNATURE) ? 1 : 0); /* TCP-MD5, 1 bit */ 1548 data |= ((sc->sc_flags & SCF_SACK) ? 1 : 0) << 1; /* SACK, 1 bit */ 1549 data |= sc->sc_requested_s_scale << 2; /* SWIN scale, 4 bits */ 1550 data |= sc->sc_requested_r_scale << 6; /* RWIN scale, 4 bits */ 1551 data |= md5_buffer[2] << 10; /* more digest bits */ 1552 data ^= md5_buffer[3]; 1553 sc->sc_ts = data; 1554 sc->sc_tsoff = data - ticks; /* after XOR */ 1555 } 1556 1557 tcpstat.tcps_sc_sendcookie++; 1558 return; 1559} 1560 1561static struct syncache * 1562syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 1563 struct syncache *sc, struct tcpopt *to, struct tcphdr *th, 1564 struct socket *so) 1565{ 1566 MD5_CTX ctx; 1567 u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)]; 1568 u_int32_t data = 0; 1569 u_int32_t *secbits; 1570 tcp_seq ack, seq; 1571 int off, mss, wnd, flags; 1572 1573 SCH_LOCK_ASSERT(sch); 1574 1575 /* 1576 * Pull information out of SYN-ACK/ACK and 1577 * revert sequence number advances. 1578 */ 1579 ack = th->th_ack - 1; 1580 seq = th->th_seq - 1; 1581 off = (ack >> 1) & 0x7; 1582 mss = (ack >> 4) & 0x7; 1583 flags = ack & 0x7f; 1584 1585 /* Which of the two secrets to use. */ 1586 secbits = (flags & 0x1) ? sch->sch_secbits_odd : sch->sch_secbits_even; 1587 1588 /* 1589 * The secret wasn't updated for the lifetime of a syncookie, 1590 * so this SYN-ACK/ACK is either too old (replay) or totally bogus. 1591 */ 1592 if (sch->sch_reseed < time_uptime) { 1593 return (NULL); 1594 } 1595 1596 /* Recompute the digest so we can compare it. */ 1597 MD5Init(&ctx); 1598 MD5Update(&ctx, ((u_int8_t *)secbits) + off, 1599 SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off); 1600 MD5Update(&ctx, secbits, off); 1601 MD5Update(&ctx, inc, sizeof(*inc)); 1602 MD5Update(&ctx, &seq, sizeof(seq)); 1603 MD5Update(&ctx, &flags, sizeof(flags)); 1604 MD5Final((u_int8_t *)&md5_buffer, &ctx); 1605 1606 /* Does the digest part of or ACK'ed ISS match? */ 1607 if ((ack & (~0x7f)) != (md5_buffer[0] << 7)) 1608 return (NULL); 1609 1610 /* Does the digest part of our reflected timestamp match? */ 1611 if (to->to_flags & TOF_TS) { 1612 data = md5_buffer[3] ^ to->to_tsecr; 1613 if ((data & (~0x3ff)) != (md5_buffer[2] << 10)) 1614 return (NULL); 1615 } 1616 1617 /* Fill in the syncache values. */ 1618 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1619 sc->sc_ipopts = NULL; 1620 1621 sc->sc_irs = seq; 1622 sc->sc_iss = ack; 1623 1624#ifdef INET6 1625 if (inc->inc_isipv6) { 1626 if (sotoinpcb(so)->in6p_flags & IN6P_AUTOFLOWLABEL) 1627 sc->sc_flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK; 1628 } else 1629#endif 1630 { 1631 sc->sc_ip_ttl = sotoinpcb(so)->inp_ip_ttl; 1632 sc->sc_ip_tos = sotoinpcb(so)->inp_ip_tos; 1633 } 1634 1635 /* Additional parameters that were encoded in the timestamp. */ 1636 if (data) { 1637 sc->sc_flags |= SCF_TIMESTAMP; 1638 sc->sc_tsreflect = to->to_tsval; 1639 sc->sc_ts = to->to_tsecr; 1640 sc->sc_tsoff = to->to_tsecr - ticks; 1641 sc->sc_flags |= (data & 0x1) ? SCF_SIGNATURE : 0; 1642 sc->sc_flags |= ((data >> 1) & 0x1) ? SCF_SACK : 0; 1643 sc->sc_requested_s_scale = min((data >> 2) & 0xf, 1644 TCP_MAX_WINSHIFT); 1645 sc->sc_requested_r_scale = min((data >> 6) & 0xf, 1646 TCP_MAX_WINSHIFT); 1647 if (sc->sc_requested_s_scale || sc->sc_requested_r_scale) 1648 sc->sc_flags |= SCF_WINSCALE; 1649 } else 1650 sc->sc_flags |= SCF_NOOPT; 1651 1652 wnd = sbspace(&so->so_rcv); 1653 wnd = imax(wnd, 0); 1654 wnd = imin(wnd, TCP_MAXWIN); 1655 sc->sc_wnd = wnd; 1656 1657 sc->sc_rxmits = 0; 1658 sc->sc_peer_mss = tcp_sc_msstab[mss]; 1659 1660 tcpstat.tcps_sc_recvcookie++; 1661 return (sc); 1662} 1663 1664/* 1665 * Returns the current number of syncache entries. This number 1666 * will probably change before you get around to calling 1667 * syncache_pcblist. 1668 */ 1669 1670int 1671syncache_pcbcount(void) 1672{ 1673 struct syncache_head *sch; 1674 int count, i; 1675 1676 for (count = 0, i = 0; i < tcp_syncache.hashsize; i++) { 1677 /* No need to lock for a read. */ 1678 sch = &tcp_syncache.hashbase[i]; 1679 count += sch->sch_length; 1680 } 1681 return count; 1682} 1683 1684/* 1685 * Exports the syncache entries to userland so that netstat can display 1686 * them alongside the other sockets. This function is intended to be 1687 * called only from tcp_pcblist. 1688 * 1689 * Due to concurrency on an active system, the number of pcbs exported 1690 * may have no relation to max_pcbs. max_pcbs merely indicates the 1691 * amount of space the caller allocated for this function to use. 1692 */ 1693int 1694syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 1695{ 1696 struct xtcpcb xt; 1697 struct syncache *sc; 1698 struct syncache_head *sch; 1699 int count, error, i; 1700 1701 for (count = 0, error = 0, i = 0; i < tcp_syncache.hashsize; i++) { 1702 sch = &tcp_syncache.hashbase[i]; 1703 SCH_LOCK(sch); 1704 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 1705 if (count >= max_pcbs) { 1706 SCH_UNLOCK(sch); 1707 goto exit; 1708 } 1709 bzero(&xt, sizeof(xt)); 1710 xt.xt_len = sizeof(xt); 1711 if (sc->sc_inc.inc_isipv6) 1712 xt.xt_inp.inp_vflag = INP_IPV6; 1713 else 1714 xt.xt_inp.inp_vflag = INP_IPV4; 1715 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 1716 xt.xt_tp.t_inpcb = &xt.xt_inp; 1717 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 1718 xt.xt_socket.xso_protocol = IPPROTO_TCP; 1719 xt.xt_socket.xso_len = sizeof (struct xsocket); 1720 xt.xt_socket.so_type = SOCK_STREAM; 1721 xt.xt_socket.so_state = SS_ISCONNECTING; 1722 error = SYSCTL_OUT(req, &xt, sizeof xt); 1723 if (error) { 1724 SCH_UNLOCK(sch); 1725 goto exit; 1726 } 1727 count++; 1728 } 1729 SCH_UNLOCK(sch); 1730 } 1731exit: 1732 *pcbs_exported = count; 1733 return error; 1734} 1735