tcp_syncache.c revision 171652
1239310Sdim/*- 2239310Sdim * Copyright (c) 2001 McAfee, Inc. 3239310Sdim * Copyright (c) 2006 Andre Oppermann, Internet Business Solutions AG 4239310Sdim * All rights reserved. 5239310Sdim * 6239310Sdim * This software was developed for the FreeBSD Project by Jonathan Lemon 7239310Sdim * and McAfee Research, the Security Research Division of McAfee, Inc. under 8239310Sdim * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 9239310Sdim * DARPA CHATS research program. 10239310Sdim * 11239310Sdim * Redistribution and use in source and binary forms, with or without 12239310Sdim * modification, are permitted provided that the following conditions 13239310Sdim * are met: 14239310Sdim * 1. Redistributions of source code must retain the above copyright 15239310Sdim * notice, this list of conditions and the following disclaimer. 16239310Sdim * 2. Redistributions in binary form must reproduce the above copyright 17239310Sdim * notice, this list of conditions and the following disclaimer in the 18239310Sdim * documentation and/or other materials provided with the distribution. 19239310Sdim * 20239310Sdim * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 21239310Sdim * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22239310Sdim * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23263509Sdim * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 24239310Sdim * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25239310Sdim * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26239310Sdim * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27239310Sdim * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28239310Sdim * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29239310Sdim * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30252723Sdim * SUCH DAMAGE. 31252723Sdim * 32252723Sdim * $FreeBSD: head/sys/netinet/tcp_syncache.c 171652 2007-07-29 20:13:22Z bmah $ 33252723Sdim */ 34239310Sdim 35239310Sdim#include "opt_inet.h" 36239310Sdim#include "opt_inet6.h" 37239310Sdim#include "opt_ipsec.h" 38239310Sdim#include "opt_mac.h" 39245431Sdim 40245431Sdim#include <sys/param.h> 41245431Sdim#include <sys/systm.h> 42245431Sdim#include <sys/kernel.h> 43245431Sdim#include <sys/sysctl.h> 44239310Sdim#include <sys/lock.h> 45239310Sdim#include <sys/mutex.h> 46239310Sdim#include <sys/malloc.h> 47239310Sdim#include <sys/mbuf.h> 48239310Sdim#include <sys/md5.h> 49239310Sdim#include <sys/proc.h> /* for proc0 declaration */ 50239310Sdim#include <sys/random.h> 51239310Sdim#include <sys/socket.h> 52239310Sdim#include <sys/socketvar.h> 53#include <sys/syslog.h> 54 55#include <vm/uma.h> 56 57#include <net/if.h> 58#include <net/route.h> 59 60#include <netinet/in.h> 61#include <netinet/in_systm.h> 62#include <netinet/ip.h> 63#include <netinet/in_var.h> 64#include <netinet/in_pcb.h> 65#include <netinet/ip_var.h> 66#include <netinet/ip_options.h> 67#ifdef INET6 68#include <netinet/ip6.h> 69#include <netinet/icmp6.h> 70#include <netinet6/nd6.h> 71#include <netinet6/ip6_var.h> 72#include <netinet6/in6_pcb.h> 73#endif 74#include <netinet/tcp.h> 75#include <netinet/tcp_fsm.h> 76#include <netinet/tcp_seq.h> 77#include <netinet/tcp_timer.h> 78#include <netinet/tcp_var.h> 79#include <netinet/tcp_syncache.h> 80#ifdef INET6 81#include <netinet6/tcp6_var.h> 82#endif 83 84#ifdef IPSEC 85#include <netipsec/ipsec.h> 86#ifdef INET6 87#include <netipsec/ipsec6.h> 88#endif 89#include <netipsec/key.h> 90#endif /*IPSEC*/ 91 92#include <machine/in_cksum.h> 93 94#include <security/mac/mac_framework.h> 95 96static int tcp_syncookies = 1; 97SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW, 98 &tcp_syncookies, 0, 99 "Use TCP SYN cookies if the syncache overflows"); 100 101static int tcp_syncookiesonly = 0; 102SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_RW, 103 &tcp_syncookiesonly, 0, 104 "Use only TCP SYN cookies"); 105 106#define SYNCOOKIE_SECRET_SIZE 8 /* dwords */ 107#define SYNCOOKIE_LIFETIME 16 /* seconds */ 108 109struct syncache { 110 TAILQ_ENTRY(syncache) sc_hash; 111 struct in_conninfo sc_inc; /* addresses */ 112 u_long sc_rxttime; /* retransmit time */ 113 u_int16_t sc_rxmits; /* retransmit counter */ 114 115 u_int32_t sc_tsreflect; /* timestamp to reflect */ 116 u_int32_t sc_ts; /* our timestamp to send */ 117 u_int32_t sc_tsoff; /* ts offset w/ syncookies */ 118 u_int32_t sc_flowlabel; /* IPv6 flowlabel */ 119 tcp_seq sc_irs; /* seq from peer */ 120 tcp_seq sc_iss; /* our ISS */ 121 struct mbuf *sc_ipopts; /* source route */ 122 123 u_int16_t sc_peer_mss; /* peer's MSS */ 124 u_int16_t sc_wnd; /* advertised window */ 125 u_int8_t sc_ip_ttl; /* IPv4 TTL */ 126 u_int8_t sc_ip_tos; /* IPv4 TOS */ 127 u_int8_t sc_requested_s_scale:4, 128 sc_requested_r_scale:4; 129 u_int8_t sc_flags; 130#define SCF_NOOPT 0x01 /* no TCP options */ 131#define SCF_WINSCALE 0x02 /* negotiated window scaling */ 132#define SCF_TIMESTAMP 0x04 /* negotiated timestamps */ 133 /* MSS is implicit */ 134#define SCF_UNREACH 0x10 /* icmp unreachable received */ 135#define SCF_SIGNATURE 0x20 /* send MD5 digests */ 136#define SCF_SACK 0x80 /* send SACK option */ 137#ifdef MAC 138 struct label *sc_label; /* MAC label reference */ 139#endif 140}; 141 142struct syncache_head { 143 struct mtx sch_mtx; 144 TAILQ_HEAD(sch_head, syncache) sch_bucket; 145 struct callout sch_timer; 146 int sch_nextc; 147 u_int sch_length; 148 u_int sch_oddeven; 149 u_int32_t sch_secbits_odd[SYNCOOKIE_SECRET_SIZE]; 150 u_int32_t sch_secbits_even[SYNCOOKIE_SECRET_SIZE]; 151 u_int sch_reseed; /* time_uptime, seconds */ 152}; 153 154static void syncache_drop(struct syncache *, struct syncache_head *); 155static void syncache_free(struct syncache *); 156static void syncache_insert(struct syncache *, struct syncache_head *); 157struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **); 158static int syncache_respond(struct syncache *); 159static struct socket *syncache_socket(struct syncache *, struct socket *, 160 struct mbuf *m); 161static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 162 int docallout); 163static void syncache_timer(void *); 164static void syncookie_generate(struct syncache_head *, struct syncache *, 165 u_int32_t *); 166static struct syncache 167 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 168 struct syncache *, struct tcpopt *, struct tcphdr *, 169 struct socket *); 170 171/* 172 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 173 * 3 retransmits corresponds to a timeout of (1 + 2 + 4 + 8 == 15) seconds, 174 * the odds are that the user has given up attempting to connect by then. 175 */ 176#define SYNCACHE_MAXREXMTS 3 177 178/* Arbitrary values */ 179#define TCP_SYNCACHE_HASHSIZE 512 180#define TCP_SYNCACHE_BUCKETLIMIT 30 181 182struct tcp_syncache { 183 struct syncache_head *hashbase; 184 uma_zone_t zone; 185 u_int hashsize; 186 u_int hashmask; 187 u_int bucket_limit; 188 u_int cache_count; /* XXX: unprotected */ 189 u_int cache_limit; 190 u_int rexmt_limit; 191 u_int hash_secret; 192}; 193static struct tcp_syncache tcp_syncache; 194 195SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache"); 196 197SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN, 198 &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache"); 199 200SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN, 201 &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache"); 202 203SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD, 204 &tcp_syncache.cache_count, 0, "Current number of entries in syncache"); 205 206SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN, 207 &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable"); 208 209SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW, 210 &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions"); 211 212int tcp_sc_rst_sock_fail = 1; 213SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, CTLFLAG_RW, 214 &tcp_sc_rst_sock_fail, 0, "Send reset on socket allocation failure"); 215 216static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 217 218#define SYNCACHE_HASH(inc, mask) \ 219 ((tcp_syncache.hash_secret ^ \ 220 (inc)->inc_faddr.s_addr ^ \ 221 ((inc)->inc_faddr.s_addr >> 16) ^ \ 222 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 223 224#define SYNCACHE_HASH6(inc, mask) \ 225 ((tcp_syncache.hash_secret ^ \ 226 (inc)->inc6_faddr.s6_addr32[0] ^ \ 227 (inc)->inc6_faddr.s6_addr32[3] ^ \ 228 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 229 230#define ENDPTS_EQ(a, b) ( \ 231 (a)->ie_fport == (b)->ie_fport && \ 232 (a)->ie_lport == (b)->ie_lport && \ 233 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \ 234 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \ 235) 236 237#define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0) 238 239#define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 240#define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 241#define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 242 243/* 244 * Requires the syncache entry to be already removed from the bucket list. 245 */ 246static void 247syncache_free(struct syncache *sc) 248{ 249 if (sc->sc_ipopts) 250 (void) m_free(sc->sc_ipopts); 251#ifdef MAC 252 mac_destroy_syncache(&sc->sc_label); 253#endif 254 255 uma_zfree(tcp_syncache.zone, sc); 256} 257 258void 259syncache_init(void) 260{ 261 int i; 262 263 tcp_syncache.cache_count = 0; 264 tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 265 tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 266 tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 267 tcp_syncache.hash_secret = arc4random(); 268 269 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 270 &tcp_syncache.hashsize); 271 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 272 &tcp_syncache.bucket_limit); 273 if (!powerof2(tcp_syncache.hashsize) || tcp_syncache.hashsize == 0) { 274 printf("WARNING: syncache hash size is not a power of 2.\n"); 275 tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 276 } 277 tcp_syncache.hashmask = tcp_syncache.hashsize - 1; 278 279 /* Set limits. */ 280 tcp_syncache.cache_limit = 281 tcp_syncache.hashsize * tcp_syncache.bucket_limit; 282 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 283 &tcp_syncache.cache_limit); 284 285 /* Allocate the hash table. */ 286 MALLOC(tcp_syncache.hashbase, struct syncache_head *, 287 tcp_syncache.hashsize * sizeof(struct syncache_head), 288 M_SYNCACHE, M_WAITOK | M_ZERO); 289 290 /* Initialize the hash buckets. */ 291 for (i = 0; i < tcp_syncache.hashsize; i++) { 292 TAILQ_INIT(&tcp_syncache.hashbase[i].sch_bucket); 293 mtx_init(&tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 294 NULL, MTX_DEF); 295 callout_init_mtx(&tcp_syncache.hashbase[i].sch_timer, 296 &tcp_syncache.hashbase[i].sch_mtx, 0); 297 tcp_syncache.hashbase[i].sch_length = 0; 298 } 299 300 /* Create the syncache entry zone. */ 301 tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 302 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 303 uma_zone_set_max(tcp_syncache.zone, tcp_syncache.cache_limit); 304} 305 306/* 307 * Inserts a syncache entry into the specified bucket row. 308 * Locks and unlocks the syncache_head autonomously. 309 */ 310static void 311syncache_insert(struct syncache *sc, struct syncache_head *sch) 312{ 313 struct syncache *sc2; 314 315 SCH_LOCK(sch); 316 317 /* 318 * Make sure that we don't overflow the per-bucket limit. 319 * If the bucket is full, toss the oldest element. 320 */ 321 if (sch->sch_length >= tcp_syncache.bucket_limit) { 322 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 323 ("sch->sch_length incorrect")); 324 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 325 syncache_drop(sc2, sch); 326 tcpstat.tcps_sc_bucketoverflow++; 327 } 328 329 /* Put it into the bucket. */ 330 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 331 sch->sch_length++; 332 333 /* Reinitialize the bucket row's timer. */ 334 syncache_timeout(sc, sch, 1); 335 336 SCH_UNLOCK(sch); 337 338 tcp_syncache.cache_count++; 339 tcpstat.tcps_sc_added++; 340} 341 342/* 343 * Remove and free entry from syncache bucket row. 344 * Expects locked syncache head. 345 */ 346static void 347syncache_drop(struct syncache *sc, struct syncache_head *sch) 348{ 349 350 SCH_LOCK_ASSERT(sch); 351 352 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 353 sch->sch_length--; 354 355 syncache_free(sc); 356 tcp_syncache.cache_count--; 357} 358 359/* 360 * Engage/reengage time on bucket row. 361 */ 362static void 363syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 364{ 365 sc->sc_rxttime = ticks + 366 TCPTV_RTOBASE * (tcp_backoff[sc->sc_rxmits]); 367 sc->sc_rxmits++; 368 if (sch->sch_nextc > sc->sc_rxttime) 369 sch->sch_nextc = sc->sc_rxttime; 370 if (!TAILQ_EMPTY(&sch->sch_bucket) && docallout) 371 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 372 syncache_timer, (void *)sch); 373} 374 375/* 376 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 377 * If we have retransmitted an entry the maximum number of times, expire it. 378 * One separate timer for each bucket row. 379 */ 380static void 381syncache_timer(void *xsch) 382{ 383 struct syncache_head *sch = (struct syncache_head *)xsch; 384 struct syncache *sc, *nsc; 385 int tick = ticks; 386 char *s; 387 388 /* NB: syncache_head has already been locked by the callout. */ 389 SCH_LOCK_ASSERT(sch); 390 391 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 392 /* 393 * We do not check if the listen socket still exists 394 * and accept the case where the listen socket may be 395 * gone by the time we resend the SYN/ACK. We do 396 * not expect this to happens often. If it does, 397 * then the RST will be sent by the time the remote 398 * host does the SYN/ACK->ACK. 399 */ 400 if (sc->sc_rxttime > tick) { 401 if (sc->sc_rxttime < sch->sch_nextc) 402 sch->sch_nextc = sc->sc_rxttime; 403 continue; 404 } 405 406 if (sc->sc_rxmits > tcp_syncache.rexmt_limit) { 407 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 408 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 409 "giving up and removing syncache entry\n", 410 s, __func__); 411 free(s, M_TCPLOG); 412 } 413 syncache_drop(sc, sch); 414 tcpstat.tcps_sc_stale++; 415 continue; 416 } 417 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 418 log(LOG_DEBUG, "%s; %s: Response timeout, " 419 "retransmitting (%u) SYN|ACK\n", 420 s, __func__, sc->sc_rxmits); 421 free(s, M_TCPLOG); 422 } 423 424 (void) syncache_respond(sc); 425 tcpstat.tcps_sc_retransmitted++; 426 syncache_timeout(sc, sch, 0); 427 } 428 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 429 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 430 syncache_timer, (void *)(sch)); 431} 432 433/* 434 * Find an entry in the syncache. 435 * Returns always with locked syncache_head plus a matching entry or NULL. 436 */ 437struct syncache * 438syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 439{ 440 struct syncache *sc; 441 struct syncache_head *sch; 442 443#ifdef INET6 444 if (inc->inc_isipv6) { 445 sch = &tcp_syncache.hashbase[ 446 SYNCACHE_HASH6(inc, tcp_syncache.hashmask)]; 447 *schp = sch; 448 449 SCH_LOCK(sch); 450 451 /* Circle through bucket row to find matching entry. */ 452 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 453 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 454 return (sc); 455 } 456 } else 457#endif 458 { 459 sch = &tcp_syncache.hashbase[ 460 SYNCACHE_HASH(inc, tcp_syncache.hashmask)]; 461 *schp = sch; 462 463 SCH_LOCK(sch); 464 465 /* Circle through bucket row to find matching entry. */ 466 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 467#ifdef INET6 468 if (sc->sc_inc.inc_isipv6) 469 continue; 470#endif 471 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 472 return (sc); 473 } 474 } 475 SCH_LOCK_ASSERT(*schp); 476 return (NULL); /* always returns with locked sch */ 477} 478 479/* 480 * This function is called when we get a RST for a 481 * non-existent connection, so that we can see if the 482 * connection is in the syn cache. If it is, zap it. 483 */ 484void 485syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 486{ 487 struct syncache *sc; 488 struct syncache_head *sch; 489 char *s = NULL; 490 491 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 492 SCH_LOCK_ASSERT(sch); 493 494 /* 495 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 496 * See RFC 793 page 65, section SEGMENT ARRIVES. 497 */ 498 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 499 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 500 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 501 "FIN flag set, segment ignored\n", s, __func__); 502 tcpstat.tcps_badrst++; 503 goto done; 504 } 505 506 /* 507 * No corresponding connection was found in syncache. 508 * If syncookies are enabled and possibly exclusively 509 * used, or we are under memory pressure, a valid RST 510 * may not find a syncache entry. In that case we're 511 * done and no SYN|ACK retransmissions will happen. 512 * Otherwise the the RST was misdirected or spoofed. 513 */ 514 if (sc == NULL) { 515 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 516 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 517 "syncache entry (possibly syncookie only), " 518 "segment ignored\n", s, __func__); 519 tcpstat.tcps_badrst++; 520 goto done; 521 } 522 523 /* 524 * If the RST bit is set, check the sequence number to see 525 * if this is a valid reset segment. 526 * RFC 793 page 37: 527 * In all states except SYN-SENT, all reset (RST) segments 528 * are validated by checking their SEQ-fields. A reset is 529 * valid if its sequence number is in the window. 530 * 531 * The sequence number in the reset segment is normally an 532 * echo of our outgoing acknowlegement numbers, but some hosts 533 * send a reset with the sequence number at the rightmost edge 534 * of our receive window, and we have to handle this case. 535 */ 536 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 537 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 538 syncache_drop(sc, sch); 539 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 540 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 541 "connection attempt aborted by remote endpoint\n", 542 s, __func__); 543 tcpstat.tcps_sc_reset++; 544 } else if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 545 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != IRS %u " 546 "(+WND %u), segment ignored\n", 547 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 548 tcpstat.tcps_badrst++; 549 } 550 551done: 552 if (s != NULL) 553 free(s, M_TCPLOG); 554 SCH_UNLOCK(sch); 555} 556 557void 558syncache_badack(struct in_conninfo *inc) 559{ 560 struct syncache *sc; 561 struct syncache_head *sch; 562 563 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 564 SCH_LOCK_ASSERT(sch); 565 if (sc != NULL) { 566 syncache_drop(sc, sch); 567 tcpstat.tcps_sc_badack++; 568 } 569 SCH_UNLOCK(sch); 570} 571 572void 573syncache_unreach(struct in_conninfo *inc, struct tcphdr *th) 574{ 575 struct syncache *sc; 576 struct syncache_head *sch; 577 578 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 579 SCH_LOCK_ASSERT(sch); 580 if (sc == NULL) 581 goto done; 582 583 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 584 if (ntohl(th->th_seq) != sc->sc_iss) 585 goto done; 586 587 /* 588 * If we've rertransmitted 3 times and this is our second error, 589 * we remove the entry. Otherwise, we allow it to continue on. 590 * This prevents us from incorrectly nuking an entry during a 591 * spurious network outage. 592 * 593 * See tcp_notify(). 594 */ 595 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 596 sc->sc_flags |= SCF_UNREACH; 597 goto done; 598 } 599 syncache_drop(sc, sch); 600 tcpstat.tcps_sc_unreach++; 601done: 602 SCH_UNLOCK(sch); 603} 604 605/* 606 * Build a new TCP socket structure from a syncache entry. 607 */ 608static struct socket * 609syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 610{ 611 struct inpcb *inp = NULL; 612 struct socket *so; 613 struct tcpcb *tp; 614 char *s; 615 616 NET_ASSERT_GIANT(); 617 INP_INFO_WLOCK_ASSERT(&tcbinfo); 618 619 /* 620 * Ok, create the full blown connection, and set things up 621 * as they would have been set up if we had created the 622 * connection when the SYN arrived. If we can't create 623 * the connection, abort it. 624 */ 625 so = sonewconn(lso, SS_ISCONNECTED); 626 if (so == NULL) { 627 /* 628 * Drop the connection; we will either send a RST or 629 * have the peer retransmit its SYN again after its 630 * RTO and try again. 631 */ 632 tcpstat.tcps_listendrop++; 633 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 634 log(LOG_DEBUG, "%s; %s: Socket create failed " 635 "due to limits or memory shortage\n", 636 s, __func__); 637 free(s, M_TCPLOG); 638 } 639 goto abort2; 640 } 641#ifdef MAC 642 SOCK_LOCK(so); 643 mac_set_socket_peer_from_mbuf(m, so); 644 SOCK_UNLOCK(so); 645#endif 646 647 inp = sotoinpcb(so); 648 INP_LOCK(inp); 649 650 /* Insert new socket into PCB hash list. */ 651 inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6; 652#ifdef INET6 653 if (sc->sc_inc.inc_isipv6) { 654 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 655 } else { 656 inp->inp_vflag &= ~INP_IPV6; 657 inp->inp_vflag |= INP_IPV4; 658#endif 659 inp->inp_laddr = sc->sc_inc.inc_laddr; 660#ifdef INET6 661 } 662#endif 663 inp->inp_lport = sc->sc_inc.inc_lport; 664 if (in_pcbinshash(inp) != 0) { 665 /* 666 * Undo the assignments above if we failed to 667 * put the PCB on the hash lists. 668 */ 669#ifdef INET6 670 if (sc->sc_inc.inc_isipv6) 671 inp->in6p_laddr = in6addr_any; 672 else 673#endif 674 inp->inp_laddr.s_addr = INADDR_ANY; 675 inp->inp_lport = 0; 676 goto abort; 677 } 678#ifdef IPSEC 679 /* Copy old policy into new socket's. */ 680 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp)) 681 printf("syncache_socket: could not copy policy\n"); 682#endif 683#ifdef INET6 684 if (sc->sc_inc.inc_isipv6) { 685 struct inpcb *oinp = sotoinpcb(lso); 686 struct in6_addr laddr6; 687 struct sockaddr_in6 sin6; 688 /* 689 * Inherit socket options from the listening socket. 690 * Note that in6p_inputopts are not (and should not be) 691 * copied, since it stores previously received options and is 692 * used to detect if each new option is different than the 693 * previous one and hence should be passed to a user. 694 * If we copied in6p_inputopts, a user would not be able to 695 * receive options just after calling the accept system call. 696 */ 697 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 698 if (oinp->in6p_outputopts) 699 inp->in6p_outputopts = 700 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 701 702 sin6.sin6_family = AF_INET6; 703 sin6.sin6_len = sizeof(sin6); 704 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 705 sin6.sin6_port = sc->sc_inc.inc_fport; 706 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 707 laddr6 = inp->in6p_laddr; 708 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 709 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 710 if (in6_pcbconnect(inp, (struct sockaddr *)&sin6, 711 thread0.td_ucred)) { 712 inp->in6p_laddr = laddr6; 713 goto abort; 714 } 715 /* Override flowlabel from in6_pcbconnect. */ 716 inp->in6p_flowinfo &= ~IPV6_FLOWLABEL_MASK; 717 inp->in6p_flowinfo |= sc->sc_flowlabel; 718 } else 719#endif 720 { 721 struct in_addr laddr; 722 struct sockaddr_in sin; 723 724 inp->inp_options = ip_srcroute(m); 725 if (inp->inp_options == NULL) { 726 inp->inp_options = sc->sc_ipopts; 727 sc->sc_ipopts = NULL; 728 } 729 730 sin.sin_family = AF_INET; 731 sin.sin_len = sizeof(sin); 732 sin.sin_addr = sc->sc_inc.inc_faddr; 733 sin.sin_port = sc->sc_inc.inc_fport; 734 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 735 laddr = inp->inp_laddr; 736 if (inp->inp_laddr.s_addr == INADDR_ANY) 737 inp->inp_laddr = sc->sc_inc.inc_laddr; 738 if (in_pcbconnect(inp, (struct sockaddr *)&sin, 739 thread0.td_ucred)) { 740 inp->inp_laddr = laddr; 741 goto abort; 742 } 743 } 744 tp = intotcpcb(inp); 745 tp->t_state = TCPS_SYN_RECEIVED; 746 tp->iss = sc->sc_iss; 747 tp->irs = sc->sc_irs; 748 tcp_rcvseqinit(tp); 749 tcp_sendseqinit(tp); 750 tp->snd_wl1 = sc->sc_irs; 751 tp->snd_max = tp->iss + 1; 752 tp->snd_nxt = tp->iss + 1; 753 tp->rcv_up = sc->sc_irs + 1; 754 tp->rcv_wnd = sc->sc_wnd; 755 tp->rcv_adv += tp->rcv_wnd; 756 tp->last_ack_sent = tp->rcv_nxt; 757 758 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 759 if (sc->sc_flags & SCF_NOOPT) 760 tp->t_flags |= TF_NOOPT; 761 else { 762 if (sc->sc_flags & SCF_WINSCALE) { 763 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 764 tp->snd_scale = sc->sc_requested_s_scale; 765 tp->request_r_scale = sc->sc_requested_r_scale; 766 } 767 if (sc->sc_flags & SCF_TIMESTAMP) { 768 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 769 tp->ts_recent = sc->sc_tsreflect; 770 tp->ts_recent_age = ticks; 771 tp->ts_offset = sc->sc_tsoff; 772 } 773#ifdef TCP_SIGNATURE 774 if (sc->sc_flags & SCF_SIGNATURE) 775 tp->t_flags |= TF_SIGNATURE; 776#endif 777 if (sc->sc_flags & SCF_SACK) 778 tp->t_flags |= TF_SACK_PERMIT; 779 } 780 781 /* 782 * Set up MSS and get cached values from tcp_hostcache. 783 * This might overwrite some of the defaults we just set. 784 */ 785 tcp_mss(tp, sc->sc_peer_mss); 786 787 /* 788 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment. 789 */ 790 if (sc->sc_rxmits) 791 tp->snd_cwnd = tp->t_maxseg; 792 tcp_timer_activate(tp, TT_KEEP, tcp_keepinit); 793 794 INP_UNLOCK(inp); 795 796 tcpstat.tcps_accepts++; 797 return (so); 798 799abort: 800 INP_UNLOCK(inp); 801abort2: 802 if (so != NULL) 803 soabort(so); 804 return (NULL); 805} 806 807/* 808 * This function gets called when we receive an ACK for a 809 * socket in the LISTEN state. We look up the connection 810 * in the syncache, and if its there, we pull it out of 811 * the cache and turn it into a full-blown connection in 812 * the SYN-RECEIVED state. 813 */ 814int 815syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 816 struct socket **lsop, struct mbuf *m) 817{ 818 struct syncache *sc; 819 struct syncache_head *sch; 820 struct syncache scs; 821 char *s; 822 823 /* 824 * Global TCP locks are held because we manipulate the PCB lists 825 * and create a new socket. 826 */ 827 INP_INFO_WLOCK_ASSERT(&tcbinfo); 828 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 829 ("%s: can handle only ACK", __func__)); 830 831 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 832 SCH_LOCK_ASSERT(sch); 833 if (sc == NULL) { 834 /* 835 * There is no syncache entry, so see if this ACK is 836 * a returning syncookie. To do this, first: 837 * A. See if this socket has had a syncache entry dropped in 838 * the past. We don't want to accept a bogus syncookie 839 * if we've never received a SYN. 840 * B. check that the syncookie is valid. If it is, then 841 * cobble up a fake syncache entry, and return. 842 */ 843 if (!tcp_syncookies) { 844 SCH_UNLOCK(sch); 845 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 846 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 847 "segment rejected (syncookies disabled)\n", 848 s, __func__); 849 goto failed; 850 } 851 bzero(&scs, sizeof(scs)); 852 sc = syncookie_lookup(inc, sch, &scs, to, th, *lsop); 853 SCH_UNLOCK(sch); 854 if (sc == NULL) { 855 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 856 log(LOG_DEBUG, "%s; %s: Segment failed " 857 "SYNCOOKIE authentication, segment rejected " 858 "(probably spoofed)\n", s, __func__); 859 goto failed; 860 } 861 } else { 862 /* Pull out the entry to unlock the bucket row. */ 863 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 864 sch->sch_length--; 865 tcp_syncache.cache_count--; 866 SCH_UNLOCK(sch); 867 } 868 869 /* 870 * Segment validation: 871 * ACK must match our initial sequence number + 1 (the SYN|ACK). 872 */ 873 if (th->th_ack != sc->sc_iss + 1) { 874 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 875 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 876 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 877 goto failed; 878 } 879 /* 880 * The SEQ must match the received initial receive sequence 881 * number + 1 (the SYN) because we didn't ACK any data that 882 * may have come with the SYN. 883 */ 884 if (th->th_seq != sc->sc_irs + 1) { 885 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 886 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 887 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 888 goto failed; 889 } 890 /* 891 * If timestamps were present in the SYN and we accepted 892 * them in our SYN|ACK we require them to be present from 893 * now on. And vice versa. 894 */ 895 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 896 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 897 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 898 "segment rejected\n", s, __func__); 899 goto failed; 900 } 901 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 902 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 903 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 904 "segment rejected\n", s, __func__); 905 goto failed; 906 } 907 /* 908 * If timestamps were negotiated the reflected timestamp 909 * must be equal to what we actually sent in the SYN|ACK. 910 */ 911 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) { 912 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 913 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 914 "segment rejected\n", 915 s, __func__, to->to_tsecr, sc->sc_ts); 916 goto failed; 917 } 918 919 *lsop = syncache_socket(sc, *lsop, m); 920 921 if (*lsop == NULL) 922 tcpstat.tcps_sc_aborted++; 923 else 924 tcpstat.tcps_sc_completed++; 925 926 if (sc != &scs) 927 syncache_free(sc); 928 return (1); 929failed: 930 if (sc != NULL && sc != &scs) 931 syncache_free(sc); 932 if (s != NULL) 933 free(s, M_TCPLOG); 934 *lsop = NULL; 935 return (0); 936} 937 938/* 939 * Given a LISTEN socket and an inbound SYN request, add 940 * this to the syn cache, and send back a segment: 941 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 942 * to the source. 943 * 944 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 945 * Doing so would require that we hold onto the data and deliver it 946 * to the application. However, if we are the target of a SYN-flood 947 * DoS attack, an attacker could send data which would eventually 948 * consume all available buffer space if it were ACKed. By not ACKing 949 * the data, we avoid this DoS scenario. 950 */ 951void 952syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 953 struct inpcb *inp, struct socket **lsop, struct mbuf *m) 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_init_syncache(&maclabel) != 0) { 999 INP_UNLOCK(inp); 1000 INP_INFO_WUNLOCK(&tcbinfo); 1001 goto done; 1002 } else 1003 mac_init_syncache_from_inpcb(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 tcpstat.tcps_sc_dupsyn++; 1032 if (ipopts) { 1033 /* 1034 * If we were remembering a previous source route, 1035 * forget it and use the new one we've been given. 1036 */ 1037 if (sc->sc_ipopts) 1038 (void) m_free(sc->sc_ipopts); 1039 sc->sc_ipopts = ipopts; 1040 } 1041 /* 1042 * Update timestamp if present. 1043 */ 1044 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1045 sc->sc_tsreflect = to->to_tsval; 1046 else 1047 sc->sc_flags &= ~SCF_TIMESTAMP; 1048#ifdef MAC 1049 /* 1050 * Since we have already unconditionally allocated label 1051 * storage, free it up. The syncache entry will already 1052 * have an initialized label we can use. 1053 */ 1054 mac_destroy_syncache(&maclabel); 1055 KASSERT(sc->sc_label != NULL, 1056 ("%s: label not initialized", __func__)); 1057#endif 1058 /* Retransmit SYN|ACK and reset retransmit count. */ 1059 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1060 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1061 "resetting timer and retransmitting SYN|ACK\n", 1062 s, __func__); 1063 free(s, M_TCPLOG); 1064 } 1065 if (syncache_respond(sc) == 0) { 1066 sc->sc_rxmits = 0; 1067 syncache_timeout(sc, sch, 1); 1068 tcpstat.tcps_sndacks++; 1069 tcpstat.tcps_sndtotal++; 1070 } 1071 SCH_UNLOCK(sch); 1072 goto done; 1073 } 1074 1075 sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT | M_ZERO); 1076 if (sc == NULL) { 1077 /* 1078 * The zone allocator couldn't provide more entries. 1079 * Treat this as if the cache was full; drop the oldest 1080 * entry and insert the new one. 1081 */ 1082 tcpstat.tcps_sc_zonefail++; 1083 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) 1084 syncache_drop(sc, sch); 1085 sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT | M_ZERO); 1086 if (sc == NULL) { 1087 if (tcp_syncookies) { 1088 bzero(&scs, sizeof(scs)); 1089 sc = &scs; 1090 } else { 1091 SCH_UNLOCK(sch); 1092 if (ipopts) 1093 (void) m_free(ipopts); 1094 goto done; 1095 } 1096 } 1097 } 1098 1099 /* 1100 * Fill in the syncache values. 1101 */ 1102#ifdef MAC 1103 sc->sc_label = maclabel; 1104#endif 1105 sc->sc_ipopts = ipopts; 1106 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1107#ifdef INET6 1108 if (!inc->inc_isipv6) 1109#endif 1110 { 1111 sc->sc_ip_tos = ip_tos; 1112 sc->sc_ip_ttl = ip_ttl; 1113 } 1114 1115 sc->sc_irs = th->th_seq; 1116 sc->sc_iss = arc4random(); 1117 sc->sc_flags = 0; 1118 sc->sc_flowlabel = 0; 1119 1120 /* 1121 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1122 * win was derived from socket earlier in the function. 1123 */ 1124 win = imax(win, 0); 1125 win = imin(win, TCP_MAXWIN); 1126 sc->sc_wnd = win; 1127 1128 if (tcp_do_rfc1323) { 1129 /* 1130 * A timestamp received in a SYN makes 1131 * it ok to send timestamp requests and replies. 1132 */ 1133 if (to->to_flags & TOF_TS) { 1134 sc->sc_tsreflect = to->to_tsval; 1135 sc->sc_ts = ticks; 1136 sc->sc_flags |= SCF_TIMESTAMP; 1137 } 1138 if (to->to_flags & TOF_SCALE) { 1139 int wscale = 0; 1140 1141 /* 1142 * Compute proper scaling value from buffer space. 1143 * Leave enough room for the socket buffer to grow 1144 * with auto sizing. This allows us to scale the 1145 * receive buffer over a wide range while not losing 1146 * any efficiency or fine granularity. 1147 * 1148 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1149 * or <SYN,ACK>) segment itself is never scaled. 1150 */ 1151 while (wscale < TCP_MAX_WINSHIFT && 1152 (0x1 << wscale) < tcp_minmss) 1153 wscale++; 1154 sc->sc_requested_r_scale = wscale; 1155 sc->sc_requested_s_scale = to->to_wscale; 1156 sc->sc_flags |= SCF_WINSCALE; 1157 } 1158 } 1159#ifdef TCP_SIGNATURE 1160 /* 1161 * If listening socket requested TCP digests, and received SYN 1162 * contains the option, flag this in the syncache so that 1163 * syncache_respond() will do the right thing with the SYN+ACK. 1164 * XXX: Currently we always record the option by default and will 1165 * attempt to use it in syncache_respond(). 1166 */ 1167 if (to->to_flags & TOF_SIGNATURE) 1168 sc->sc_flags |= SCF_SIGNATURE; 1169#endif 1170 if (to->to_flags & TOF_SACK) 1171 sc->sc_flags |= SCF_SACK; 1172 if (to->to_flags & TOF_MSS) 1173 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1174 if (noopt) 1175 sc->sc_flags |= SCF_NOOPT; 1176 1177 if (tcp_syncookies) { 1178 syncookie_generate(sch, sc, &flowtmp); 1179#ifdef INET6 1180 if (autoflowlabel) 1181 sc->sc_flowlabel = flowtmp; 1182#endif 1183 } else { 1184#ifdef INET6 1185 if (autoflowlabel) 1186 sc->sc_flowlabel = 1187 (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK); 1188#endif 1189 } 1190 SCH_UNLOCK(sch); 1191 1192 /* 1193 * Do a standard 3-way handshake. 1194 */ 1195 if (syncache_respond(sc) == 0) { 1196 if (tcp_syncookies && tcp_syncookiesonly && sc != &scs) 1197 syncache_free(sc); 1198 else if (sc != &scs) 1199 syncache_insert(sc, sch); /* locks and unlocks sch */ 1200 tcpstat.tcps_sndacks++; 1201 tcpstat.tcps_sndtotal++; 1202 } else { 1203 if (sc != &scs) 1204 syncache_free(sc); 1205 tcpstat.tcps_sc_dropped++; 1206 } 1207 1208done: 1209#ifdef MAC 1210 if (sc == &scs) 1211 mac_destroy_syncache(&maclabel); 1212#endif 1213 *lsop = NULL; 1214 m_freem(m); 1215 return; 1216} 1217 1218static int 1219syncache_respond(struct syncache *sc) 1220{ 1221 struct ip *ip = NULL; 1222 struct mbuf *m; 1223 struct tcphdr *th; 1224 int optlen, error; 1225 u_int16_t hlen, tlen, mssopt; 1226 struct tcpopt to; 1227#ifdef INET6 1228 struct ip6_hdr *ip6 = NULL; 1229#endif 1230 1231 hlen = 1232#ifdef INET6 1233 (sc->sc_inc.inc_isipv6) ? sizeof(struct ip6_hdr) : 1234#endif 1235 sizeof(struct ip); 1236 tlen = hlen + sizeof(struct tcphdr); 1237 1238 /* Determine MSS we advertize to other end of connection. */ 1239 mssopt = tcp_mssopt(&sc->sc_inc); 1240 if (sc->sc_peer_mss) 1241 mssopt = max( min(sc->sc_peer_mss, mssopt), tcp_minmss); 1242 1243 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1244 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1245 ("syncache: mbuf too small")); 1246 1247 /* Create the IP+TCP header from scratch. */ 1248 m = m_gethdr(M_DONTWAIT, MT_DATA); 1249 if (m == NULL) 1250 return (ENOBUFS); 1251#ifdef MAC 1252 mac_create_mbuf_from_syncache(sc->sc_label, m); 1253#endif 1254 m->m_data += max_linkhdr; 1255 m->m_len = tlen; 1256 m->m_pkthdr.len = tlen; 1257 m->m_pkthdr.rcvif = NULL; 1258 1259#ifdef INET6 1260 if (sc->sc_inc.inc_isipv6) { 1261 ip6 = mtod(m, struct ip6_hdr *); 1262 ip6->ip6_vfc = IPV6_VERSION; 1263 ip6->ip6_nxt = IPPROTO_TCP; 1264 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1265 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1266 ip6->ip6_plen = htons(tlen - hlen); 1267 /* ip6_hlim is set after checksum */ 1268 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1269 ip6->ip6_flow |= sc->sc_flowlabel; 1270 1271 th = (struct tcphdr *)(ip6 + 1); 1272 } else 1273#endif 1274 { 1275 ip = mtod(m, struct ip *); 1276 ip->ip_v = IPVERSION; 1277 ip->ip_hl = sizeof(struct ip) >> 2; 1278 ip->ip_len = tlen; 1279 ip->ip_id = 0; 1280 ip->ip_off = 0; 1281 ip->ip_sum = 0; 1282 ip->ip_p = IPPROTO_TCP; 1283 ip->ip_src = sc->sc_inc.inc_laddr; 1284 ip->ip_dst = sc->sc_inc.inc_faddr; 1285 ip->ip_ttl = sc->sc_ip_ttl; 1286 ip->ip_tos = sc->sc_ip_tos; 1287 1288 /* 1289 * See if we should do MTU discovery. Route lookups are 1290 * expensive, so we will only unset the DF bit if: 1291 * 1292 * 1) path_mtu_discovery is disabled 1293 * 2) the SCF_UNREACH flag has been set 1294 */ 1295 if (path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1296 ip->ip_off |= IP_DF; 1297 1298 th = (struct tcphdr *)(ip + 1); 1299 } 1300 th->th_sport = sc->sc_inc.inc_lport; 1301 th->th_dport = sc->sc_inc.inc_fport; 1302 1303 th->th_seq = htonl(sc->sc_iss); 1304 th->th_ack = htonl(sc->sc_irs + 1); 1305 th->th_off = sizeof(struct tcphdr) >> 2; 1306 th->th_x2 = 0; 1307 th->th_flags = TH_SYN|TH_ACK; 1308 th->th_win = htons(sc->sc_wnd); 1309 th->th_urp = 0; 1310 1311 /* Tack on the TCP options. */ 1312 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1313 to.to_flags = 0; 1314 1315 to.to_mss = mssopt; 1316 to.to_flags = TOF_MSS; 1317 if (sc->sc_flags & SCF_WINSCALE) { 1318 to.to_wscale = sc->sc_requested_r_scale; 1319 to.to_flags |= TOF_SCALE; 1320 } 1321 if (sc->sc_flags & SCF_TIMESTAMP) { 1322 /* Virgin timestamp or TCP cookie enhanced one. */ 1323 to.to_tsval = sc->sc_ts; 1324 to.to_tsecr = sc->sc_tsreflect; 1325 to.to_flags |= TOF_TS; 1326 } 1327 if (sc->sc_flags & SCF_SACK) 1328 to.to_flags |= TOF_SACKPERM; 1329#ifdef TCP_SIGNATURE 1330 if (sc->sc_flags & SCF_SIGNATURE) 1331 to.to_flags |= TOF_SIGNATURE; 1332#endif 1333 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1334 1335#ifdef TCP_SIGNATURE 1336 tcp_signature_compute(m, sizeof(struct ip), 0, optlen, 1337 to.to_signature, IPSEC_DIR_OUTBOUND); 1338#endif 1339 1340 /* Adjust headers by option size. */ 1341 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1342 m->m_len += optlen; 1343 m->m_pkthdr.len += optlen; 1344#ifdef INET6 1345 if (sc->sc_inc.inc_isipv6) 1346 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1347 else 1348#endif 1349 ip->ip_len += optlen; 1350 } else 1351 optlen = 0; 1352 1353#ifdef INET6 1354 if (sc->sc_inc.inc_isipv6) { 1355 th->th_sum = 0; 1356 th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, 1357 tlen + optlen - hlen); 1358 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1359 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1360 } else 1361#endif 1362 { 1363 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1364 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1365 m->m_pkthdr.csum_flags = CSUM_TCP; 1366 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1367 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1368 } 1369 return (error); 1370} 1371 1372/* 1373 * The purpose of SYN cookies is to avoid keeping track of all SYN's we 1374 * receive and to be able to handle SYN floods from bogus source addresses 1375 * (where we will never receive any reply). SYN floods try to exhaust all 1376 * our memory and available slots in the SYN cache table to cause a denial 1377 * of service to legitimate users of the local host. 1378 * 1379 * The idea of SYN cookies is to encode and include all necessary information 1380 * about the connection setup state within the SYN-ACK we send back and thus 1381 * to get along without keeping any local state until the ACK to the SYN-ACK 1382 * arrives (if ever). Everything we need to know should be available from 1383 * the information we encoded in the SYN-ACK. 1384 * 1385 * More information about the theory behind SYN cookies and its first 1386 * discussion and specification can be found at: 1387 * http://cr.yp.to/syncookies.html (overview) 1388 * http://cr.yp.to/syncookies/archive (gory details) 1389 * 1390 * This implementation extends the orginal idea and first implementation 1391 * of FreeBSD by using not only the initial sequence number field to store 1392 * information but also the timestamp field if present. This way we can 1393 * keep track of the entire state we need to know to recreate the session in 1394 * its original form. Almost all TCP speakers implement RFC1323 timestamps 1395 * these days. For those that do not we still have to live with the known 1396 * shortcomings of the ISN only SYN cookies. 1397 * 1398 * Cookie layers: 1399 * 1400 * Initial sequence number we send: 1401 * 31|................................|0 1402 * DDDDDDDDDDDDDDDDDDDDDDDDDMMMRRRP 1403 * D = MD5 Digest (first dword) 1404 * M = MSS index 1405 * R = Rotation of secret 1406 * P = Odd or Even secret 1407 * 1408 * The MD5 Digest is computed with over following parameters: 1409 * a) randomly rotated secret 1410 * b) struct in_conninfo containing the remote/local ip/port (IPv4&IPv6) 1411 * c) the received initial sequence number from remote host 1412 * d) the rotation offset and odd/even bit 1413 * 1414 * Timestamp we send: 1415 * 31|................................|0 1416 * DDDDDDDDDDDDDDDDDDDDDDSSSSRRRRA5 1417 * D = MD5 Digest (third dword) (only as filler) 1418 * S = Requested send window scale 1419 * R = Requested receive window scale 1420 * A = SACK allowed 1421 * 5 = TCP-MD5 enabled (not implemented yet) 1422 * XORed with MD5 Digest (forth dword) 1423 * 1424 * The timestamp isn't cryptographically secure and doesn't need to be. 1425 * The double use of the MD5 digest dwords ties it to a specific remote/ 1426 * local host/port, remote initial sequence number and our local time 1427 * limited secret. A received timestamp is reverted (XORed) and then 1428 * the contained MD5 dword is compared to the computed one to ensure the 1429 * timestamp belongs to the SYN-ACK we sent. The other parameters may 1430 * have been tampered with but this isn't different from supplying bogus 1431 * values in the SYN in the first place. 1432 * 1433 * Some problems with SYN cookies remain however: 1434 * Consider the problem of a recreated (and retransmitted) cookie. If the 1435 * original SYN was accepted, the connection is established. The second 1436 * SYN is inflight, and if it arrives with an ISN that falls within the 1437 * receive window, the connection is killed. 1438 * 1439 * Notes: 1440 * A heuristic to determine when to accept syn cookies is not necessary. 1441 * An ACK flood would cause the syncookie verification to be attempted, 1442 * but a SYN flood causes syncookies to be generated. Both are of equal 1443 * cost, so there's no point in trying to optimize the ACK flood case. 1444 * Also, if you don't process certain ACKs for some reason, then all someone 1445 * would have to do is launch a SYN and ACK flood at the same time, which 1446 * would stop cookie verification and defeat the entire purpose of syncookies. 1447 */ 1448static int tcp_sc_msstab[] = { 0, 256, 468, 536, 996, 1452, 1460, 8960 }; 1449 1450static void 1451syncookie_generate(struct syncache_head *sch, struct syncache *sc, 1452 u_int32_t *flowlabel) 1453{ 1454 MD5_CTX ctx; 1455 u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)]; 1456 u_int32_t data; 1457 u_int32_t *secbits; 1458 u_int off, pmss, mss; 1459 int i; 1460 1461 SCH_LOCK_ASSERT(sch); 1462 1463 /* Which of the two secrets to use. */ 1464 secbits = sch->sch_oddeven ? 1465 sch->sch_secbits_odd : sch->sch_secbits_even; 1466 1467 /* Reseed secret if too old. */ 1468 if (sch->sch_reseed < time_uptime) { 1469 sch->sch_oddeven = sch->sch_oddeven ? 0 : 1; /* toggle */ 1470 secbits = sch->sch_oddeven ? 1471 sch->sch_secbits_odd : sch->sch_secbits_even; 1472 for (i = 0; i < SYNCOOKIE_SECRET_SIZE; i++) 1473 secbits[i] = arc4random(); 1474 sch->sch_reseed = time_uptime + SYNCOOKIE_LIFETIME; 1475 } 1476 1477 /* Secret rotation offset. */ 1478 off = sc->sc_iss & 0x7; /* iss was randomized before */ 1479 1480 /* Maximum segment size calculation. */ 1481 pmss = max( min(sc->sc_peer_mss, tcp_mssopt(&sc->sc_inc)), tcp_minmss); 1482 for (mss = sizeof(tcp_sc_msstab) / sizeof(int) - 1; mss > 0; mss--) 1483 if (tcp_sc_msstab[mss] <= pmss) 1484 break; 1485 1486 /* Fold parameters and MD5 digest into the ISN we will send. */ 1487 data = sch->sch_oddeven;/* odd or even secret, 1 bit */ 1488 data |= off << 1; /* secret offset, derived from iss, 3 bits */ 1489 data |= mss << 4; /* mss, 3 bits */ 1490 1491 MD5Init(&ctx); 1492 MD5Update(&ctx, ((u_int8_t *)secbits) + off, 1493 SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off); 1494 MD5Update(&ctx, secbits, off); 1495 MD5Update(&ctx, &sc->sc_inc, sizeof(sc->sc_inc)); 1496 MD5Update(&ctx, &sc->sc_irs, sizeof(sc->sc_irs)); 1497 MD5Update(&ctx, &data, sizeof(data)); 1498 MD5Final((u_int8_t *)&md5_buffer, &ctx); 1499 1500 data |= (md5_buffer[0] << 7); 1501 sc->sc_iss = data; 1502 1503#ifdef INET6 1504 *flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK; 1505#endif 1506 1507 /* Additional parameters are stored in the timestamp if present. */ 1508 if (sc->sc_flags & SCF_TIMESTAMP) { 1509 data = ((sc->sc_flags & SCF_SIGNATURE) ? 1 : 0); /* TCP-MD5, 1 bit */ 1510 data |= ((sc->sc_flags & SCF_SACK) ? 1 : 0) << 1; /* SACK, 1 bit */ 1511 data |= sc->sc_requested_s_scale << 2; /* SWIN scale, 4 bits */ 1512 data |= sc->sc_requested_r_scale << 6; /* RWIN scale, 4 bits */ 1513 data |= md5_buffer[2] << 10; /* more digest bits */ 1514 data ^= md5_buffer[3]; 1515 sc->sc_ts = data; 1516 sc->sc_tsoff = data - ticks; /* after XOR */ 1517 } 1518 1519 tcpstat.tcps_sc_sendcookie++; 1520 return; 1521} 1522 1523static struct syncache * 1524syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 1525 struct syncache *sc, struct tcpopt *to, struct tcphdr *th, 1526 struct socket *so) 1527{ 1528 MD5_CTX ctx; 1529 u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)]; 1530 u_int32_t data = 0; 1531 u_int32_t *secbits; 1532 tcp_seq ack, seq; 1533 int off, mss, wnd, flags; 1534 1535 SCH_LOCK_ASSERT(sch); 1536 1537 /* 1538 * Pull information out of SYN-ACK/ACK and 1539 * revert sequence number advances. 1540 */ 1541 ack = th->th_ack - 1; 1542 seq = th->th_seq - 1; 1543 off = (ack >> 1) & 0x7; 1544 mss = (ack >> 4) & 0x7; 1545 flags = ack & 0x7f; 1546 1547 /* Which of the two secrets to use. */ 1548 secbits = (flags & 0x1) ? sch->sch_secbits_odd : sch->sch_secbits_even; 1549 1550 /* 1551 * The secret wasn't updated for the lifetime of a syncookie, 1552 * so this SYN-ACK/ACK is either too old (replay) or totally bogus. 1553 */ 1554 if (sch->sch_reseed < time_uptime) { 1555 return (NULL); 1556 } 1557 1558 /* Recompute the digest so we can compare it. */ 1559 MD5Init(&ctx); 1560 MD5Update(&ctx, ((u_int8_t *)secbits) + off, 1561 SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off); 1562 MD5Update(&ctx, secbits, off); 1563 MD5Update(&ctx, inc, sizeof(*inc)); 1564 MD5Update(&ctx, &seq, sizeof(seq)); 1565 MD5Update(&ctx, &flags, sizeof(flags)); 1566 MD5Final((u_int8_t *)&md5_buffer, &ctx); 1567 1568 /* Does the digest part of or ACK'ed ISS match? */ 1569 if ((ack & (~0x7f)) != (md5_buffer[0] << 7)) 1570 return (NULL); 1571 1572 /* Does the digest part of our reflected timestamp match? */ 1573 if (to->to_flags & TOF_TS) { 1574 data = md5_buffer[3] ^ to->to_tsecr; 1575 if ((data & (~0x3ff)) != (md5_buffer[2] << 10)) 1576 return (NULL); 1577 } 1578 1579 /* Fill in the syncache values. */ 1580 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1581 sc->sc_ipopts = NULL; 1582 1583 sc->sc_irs = seq; 1584 sc->sc_iss = ack; 1585 1586#ifdef INET6 1587 if (inc->inc_isipv6) { 1588 if (sotoinpcb(so)->in6p_flags & IN6P_AUTOFLOWLABEL) 1589 sc->sc_flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK; 1590 } else 1591#endif 1592 { 1593 sc->sc_ip_ttl = sotoinpcb(so)->inp_ip_ttl; 1594 sc->sc_ip_tos = sotoinpcb(so)->inp_ip_tos; 1595 } 1596 1597 /* Additional parameters that were encoded in the timestamp. */ 1598 if (data) { 1599 sc->sc_flags |= SCF_TIMESTAMP; 1600 sc->sc_tsreflect = to->to_tsval; 1601 sc->sc_ts = to->to_tsecr; 1602 sc->sc_tsoff = to->to_tsecr - ticks; 1603 sc->sc_flags |= (data & 0x1) ? SCF_SIGNATURE : 0; 1604 sc->sc_flags |= ((data >> 1) & 0x1) ? SCF_SACK : 0; 1605 sc->sc_requested_s_scale = min((data >> 2) & 0xf, 1606 TCP_MAX_WINSHIFT); 1607 sc->sc_requested_r_scale = min((data >> 6) & 0xf, 1608 TCP_MAX_WINSHIFT); 1609 if (sc->sc_requested_s_scale || sc->sc_requested_r_scale) 1610 sc->sc_flags |= SCF_WINSCALE; 1611 } else 1612 sc->sc_flags |= SCF_NOOPT; 1613 1614 wnd = sbspace(&so->so_rcv); 1615 wnd = imax(wnd, 0); 1616 wnd = imin(wnd, TCP_MAXWIN); 1617 sc->sc_wnd = wnd; 1618 1619 sc->sc_rxmits = 0; 1620 sc->sc_peer_mss = tcp_sc_msstab[mss]; 1621 1622 tcpstat.tcps_sc_recvcookie++; 1623 return (sc); 1624} 1625 1626/* 1627 * Returns the current number of syncache entries. This number 1628 * will probably change before you get around to calling 1629 * syncache_pcblist. 1630 */ 1631 1632int 1633syncache_pcbcount(void) 1634{ 1635 struct syncache_head *sch; 1636 int count, i; 1637 1638 for (count = 0, i = 0; i < tcp_syncache.hashsize; i++) { 1639 /* No need to lock for a read. */ 1640 sch = &tcp_syncache.hashbase[i]; 1641 count += sch->sch_length; 1642 } 1643 return count; 1644} 1645 1646/* 1647 * Exports the syncache entries to userland so that netstat can display 1648 * them alongside the other sockets. This function is intended to be 1649 * called only from tcp_pcblist. 1650 * 1651 * Due to concurrency on an active system, the number of pcbs exported 1652 * may have no relation to max_pcbs. max_pcbs merely indicates the 1653 * amount of space the caller allocated for this function to use. 1654 */ 1655int 1656syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 1657{ 1658 struct xtcpcb xt; 1659 struct syncache *sc; 1660 struct syncache_head *sch; 1661 int count, error, i; 1662 1663 for (count = 0, error = 0, i = 0; i < tcp_syncache.hashsize; i++) { 1664 sch = &tcp_syncache.hashbase[i]; 1665 SCH_LOCK(sch); 1666 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 1667 if (count >= max_pcbs) { 1668 SCH_UNLOCK(sch); 1669 goto exit; 1670 } 1671 bzero(&xt, sizeof(xt)); 1672 xt.xt_len = sizeof(xt); 1673 if (sc->sc_inc.inc_isipv6) 1674 xt.xt_inp.inp_vflag = INP_IPV6; 1675 else 1676 xt.xt_inp.inp_vflag = INP_IPV4; 1677 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 1678 xt.xt_tp.t_inpcb = &xt.xt_inp; 1679 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 1680 xt.xt_socket.xso_protocol = IPPROTO_TCP; 1681 xt.xt_socket.xso_len = sizeof (struct xsocket); 1682 xt.xt_socket.so_type = SOCK_STREAM; 1683 xt.xt_socket.so_state = SS_ISCONNECTING; 1684 error = SYSCTL_OUT(req, &xt, sizeof xt); 1685 if (error) { 1686 SCH_UNLOCK(sch); 1687 goto exit; 1688 } 1689 count++; 1690 } 1691 SCH_UNLOCK(sch); 1692 } 1693exit: 1694 *pcbs_exported = count; 1695 return error; 1696} 1697 1698