Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/netinet/tcp_syncache.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
1 /*- 2 * Copyright (c) 2001 McAfee, Inc. 3 * Copyright (c) 2006,2013 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. [2001 McAfee, Inc.] 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: releng/10.2/sys/netinet/tcp_syncache.c 284603 2015-06-19 19:36:21Z hiren $"); 35 36 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 #include "opt_pcbgroup.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/limits.h> 46 #include <sys/lock.h> 47 #include <sys/mutex.h> 48 #include <sys/malloc.h> 49 #include <sys/mbuf.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 #include <sys/ucred.h> 56 57 #include <sys/md5.h> 58 #include <crypto/siphash/siphash.h> 59 60 #include <vm/uma.h> 61 62 #include <net/if.h> 63 #include <net/route.h> 64 #include <net/vnet.h> 65 66 #include <netinet/in.h> 67 #include <netinet/in_systm.h> 68 #include <netinet/ip.h> 69 #include <netinet/in_var.h> 70 #include <netinet/in_pcb.h> 71 #include <netinet/ip_var.h> 72 #include <netinet/ip_options.h> 73 #ifdef INET6 74 #include <netinet/ip6.h> 75 #include <netinet/icmp6.h> 76 #include <netinet6/nd6.h> 77 #include <netinet6/ip6_var.h> 78 #include <netinet6/in6_pcb.h> 79 #endif 80 #include <netinet/tcp.h> 81 #include <netinet/tcp_fsm.h> 82 #include <netinet/tcp_seq.h> 83 #include <netinet/tcp_timer.h> 84 #include <netinet/tcp_var.h> 85 #include <netinet/tcp_syncache.h> 86 #ifdef INET6 87 #include <netinet6/tcp6_var.h> 88 #endif 89 #ifdef TCP_OFFLOAD 90 #include <netinet/toecore.h> 91 #endif 92 93 #ifdef IPSEC 94 #include <netipsec/ipsec.h> 95 #ifdef INET6 96 #include <netipsec/ipsec6.h> 97 #endif 98 #include <netipsec/key.h> 99 #endif /*IPSEC*/ 100 101 #include <machine/in_cksum.h> 102 103 #include <security/mac/mac_framework.h> 104 105 static VNET_DEFINE(int, tcp_syncookies) = 1; 106 #define V_tcp_syncookies VNET(tcp_syncookies) 107 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW, 108 &VNET_NAME(tcp_syncookies), 0, 109 "Use TCP SYN cookies if the syncache overflows"); 110 111 static VNET_DEFINE(int, tcp_syncookiesonly) = 0; 112 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 113 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_RW, 114 &VNET_NAME(tcp_syncookiesonly), 0, 115 "Use only TCP SYN cookies"); 116 117 #ifdef TCP_OFFLOAD 118 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 119 #endif 120 121 static void syncache_drop(struct syncache *, struct syncache_head *); 122 static void syncache_free(struct syncache *); 123 static void syncache_insert(struct syncache *, struct syncache_head *); 124 static int syncache_respond(struct syncache *); 125 static struct socket *syncache_socket(struct syncache *, struct socket *, 126 struct mbuf *m); 127 static int syncache_sysctl_count(SYSCTL_HANDLER_ARGS); 128 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 129 int docallout); 130 static void syncache_timer(void *); 131 132 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 133 uint8_t *, uintptr_t); 134 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 135 static struct syncache 136 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 137 struct syncache *, struct tcphdr *, struct tcpopt *, 138 struct socket *); 139 static void syncookie_reseed(void *); 140 #ifdef INVARIANTS 141 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 142 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 143 struct socket *lso); 144 #endif 145 146 /* 147 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 148 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds, 149 * the odds are that the user has given up attempting to connect by then. 150 */ 151 #define SYNCACHE_MAXREXMTS 3 152 153 /* Arbitrary values */ 154 #define TCP_SYNCACHE_HASHSIZE 512 155 #define TCP_SYNCACHE_BUCKETLIMIT 30 156 157 static VNET_DEFINE(struct tcp_syncache, tcp_syncache); 158 #define V_tcp_syncache VNET(tcp_syncache) 159 160 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, 161 "TCP SYN cache"); 162 163 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN, 164 &VNET_NAME(tcp_syncache.bucket_limit), 0, 165 "Per-bucket hash limit for syncache"); 166 167 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN, 168 &VNET_NAME(tcp_syncache.cache_limit), 0, 169 "Overall entry limit for syncache"); 170 171 SYSCTL_VNET_PROC(_net_inet_tcp_syncache, OID_AUTO, count, (CTLTYPE_UINT|CTLFLAG_RD), 172 NULL, 0, &syncache_sysctl_count, "IU", 173 "Current number of entries in syncache"); 174 175 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN, 176 &VNET_NAME(tcp_syncache.hashsize), 0, 177 "Size of TCP syncache hashtable"); 178 179 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW, 180 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 181 "Limit on SYN/ACK retransmissions"); 182 183 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 184 SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 185 CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 186 "Send reset on socket allocation failure"); 187 188 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 189 190 #define SYNCACHE_HASH(inc, mask) \ 191 ((V_tcp_syncache.hash_secret ^ \ 192 (inc)->inc_faddr.s_addr ^ \ 193 ((inc)->inc_faddr.s_addr >> 16) ^ \ 194 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 195 196 #define SYNCACHE_HASH6(inc, mask) \ 197 ((V_tcp_syncache.hash_secret ^ \ 198 (inc)->inc6_faddr.s6_addr32[0] ^ \ 199 (inc)->inc6_faddr.s6_addr32[3] ^ \ 200 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 201 202 #define ENDPTS_EQ(a, b) ( \ 203 (a)->ie_fport == (b)->ie_fport && \ 204 (a)->ie_lport == (b)->ie_lport && \ 205 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \ 206 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \ 207 ) 208 209 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0) 210 211 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 212 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 213 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 214 215 /* 216 * Requires the syncache entry to be already removed from the bucket list. 217 */ 218 static void 219 syncache_free(struct syncache *sc) 220 { 221 222 if (sc->sc_ipopts) 223 (void) m_free(sc->sc_ipopts); 224 if (sc->sc_cred) 225 crfree(sc->sc_cred); 226 #ifdef MAC 227 mac_syncache_destroy(&sc->sc_label); 228 #endif 229 230 uma_zfree(V_tcp_syncache.zone, sc); 231 } 232 233 void 234 syncache_init(void) 235 { 236 int i; 237 238 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 239 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 240 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 241 V_tcp_syncache.hash_secret = arc4random(); 242 243 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 244 &V_tcp_syncache.hashsize); 245 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 246 &V_tcp_syncache.bucket_limit); 247 if (!powerof2(V_tcp_syncache.hashsize) || 248 V_tcp_syncache.hashsize == 0) { 249 printf("WARNING: syncache hash size is not a power of 2.\n"); 250 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 251 } 252 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 253 254 /* Set limits. */ 255 V_tcp_syncache.cache_limit = 256 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 257 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 258 &V_tcp_syncache.cache_limit); 259 260 /* Allocate the hash table. */ 261 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 262 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 263 264 #ifdef VIMAGE 265 V_tcp_syncache.vnet = curvnet; 266 #endif 267 268 /* Initialize the hash buckets. */ 269 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 270 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 271 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 272 NULL, MTX_DEF); 273 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 274 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 275 V_tcp_syncache.hashbase[i].sch_length = 0; 276 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 277 } 278 279 /* Create the syncache entry zone. */ 280 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 281 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 282 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 283 V_tcp_syncache.cache_limit); 284 285 /* Start the SYN cookie reseeder callout. */ 286 callout_init(&V_tcp_syncache.secret.reseed, 1); 287 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 288 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 289 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 290 syncookie_reseed, &V_tcp_syncache); 291 } 292 293 #ifdef VIMAGE 294 void 295 syncache_destroy(void) 296 { 297 struct syncache_head *sch; 298 struct syncache *sc, *nsc; 299 int i; 300 301 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 302 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 303 304 sch = &V_tcp_syncache.hashbase[i]; 305 callout_drain(&sch->sch_timer); 306 307 SCH_LOCK(sch); 308 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 309 syncache_drop(sc, sch); 310 SCH_UNLOCK(sch); 311 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 312 ("%s: sch->sch_bucket not empty", __func__)); 313 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 314 __func__, sch->sch_length)); 315 mtx_destroy(&sch->sch_mtx); 316 } 317 318 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 319 ("%s: cache_count not 0", __func__)); 320 321 /* Free the allocated global resources. */ 322 uma_zdestroy(V_tcp_syncache.zone); 323 free(V_tcp_syncache.hashbase, M_SYNCACHE); 324 325 callout_drain(&V_tcp_syncache.secret.reseed); 326 } 327 #endif 328 329 static int 330 syncache_sysctl_count(SYSCTL_HANDLER_ARGS) 331 { 332 int count; 333 334 count = uma_zone_get_cur(V_tcp_syncache.zone); 335 return (sysctl_handle_int(oidp, &count, 0, req)); 336 } 337 338 /* 339 * Inserts a syncache entry into the specified bucket row. 340 * Locks and unlocks the syncache_head autonomously. 341 */ 342 static void 343 syncache_insert(struct syncache *sc, struct syncache_head *sch) 344 { 345 struct syncache *sc2; 346 347 SCH_LOCK(sch); 348 349 /* 350 * Make sure that we don't overflow the per-bucket limit. 351 * If the bucket is full, toss the oldest element. 352 */ 353 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 354 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 355 ("sch->sch_length incorrect")); 356 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 357 syncache_drop(sc2, sch); 358 TCPSTAT_INC(tcps_sc_bucketoverflow); 359 } 360 361 /* Put it into the bucket. */ 362 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 363 sch->sch_length++; 364 365 #ifdef TCP_OFFLOAD 366 if (ADDED_BY_TOE(sc)) { 367 struct toedev *tod = sc->sc_tod; 368 369 tod->tod_syncache_added(tod, sc->sc_todctx); 370 } 371 #endif 372 373 /* Reinitialize the bucket row's timer. */ 374 if (sch->sch_length == 1) 375 sch->sch_nextc = ticks + INT_MAX; 376 syncache_timeout(sc, sch, 1); 377 378 SCH_UNLOCK(sch); 379 380 TCPSTAT_INC(tcps_sc_added); 381 } 382 383 /* 384 * Remove and free entry from syncache bucket row. 385 * Expects locked syncache head. 386 */ 387 static void 388 syncache_drop(struct syncache *sc, struct syncache_head *sch) 389 { 390 391 SCH_LOCK_ASSERT(sch); 392 393 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 394 sch->sch_length--; 395 396 #ifdef TCP_OFFLOAD 397 if (ADDED_BY_TOE(sc)) { 398 struct toedev *tod = sc->sc_tod; 399 400 tod->tod_syncache_removed(tod, sc->sc_todctx); 401 } 402 #endif 403 404 syncache_free(sc); 405 } 406 407 /* 408 * Engage/reengage time on bucket row. 409 */ 410 static void 411 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 412 { 413 sc->sc_rxttime = ticks + 414 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]); 415 sc->sc_rxmits++; 416 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 417 sch->sch_nextc = sc->sc_rxttime; 418 if (docallout) 419 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 420 syncache_timer, (void *)sch); 421 } 422 } 423 424 /* 425 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 426 * If we have retransmitted an entry the maximum number of times, expire it. 427 * One separate timer for each bucket row. 428 */ 429 static void 430 syncache_timer(void *xsch) 431 { 432 struct syncache_head *sch = (struct syncache_head *)xsch; 433 struct syncache *sc, *nsc; 434 int tick = ticks; 435 char *s; 436 437 CURVNET_SET(sch->sch_sc->vnet); 438 439 /* NB: syncache_head has already been locked by the callout. */ 440 SCH_LOCK_ASSERT(sch); 441 442 /* 443 * In the following cycle we may remove some entries and/or 444 * advance some timeouts, so re-initialize the bucket timer. 445 */ 446 sch->sch_nextc = tick + INT_MAX; 447 448 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 449 /* 450 * We do not check if the listen socket still exists 451 * and accept the case where the listen socket may be 452 * gone by the time we resend the SYN/ACK. We do 453 * not expect this to happens often. If it does, 454 * then the RST will be sent by the time the remote 455 * host does the SYN/ACK->ACK. 456 */ 457 if (TSTMP_GT(sc->sc_rxttime, tick)) { 458 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 459 sch->sch_nextc = sc->sc_rxttime; 460 continue; 461 } 462 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 463 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 464 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 465 "giving up and removing syncache entry\n", 466 s, __func__); 467 free(s, M_TCPLOG); 468 } 469 syncache_drop(sc, sch); 470 TCPSTAT_INC(tcps_sc_stale); 471 continue; 472 } 473 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 474 log(LOG_DEBUG, "%s; %s: Response timeout, " 475 "retransmitting (%u) SYN|ACK\n", 476 s, __func__, sc->sc_rxmits); 477 free(s, M_TCPLOG); 478 } 479 480 (void) syncache_respond(sc); 481 TCPSTAT_INC(tcps_sc_retransmitted); 482 syncache_timeout(sc, sch, 0); 483 } 484 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 485 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 486 syncache_timer, (void *)(sch)); 487 CURVNET_RESTORE(); 488 } 489 490 /* 491 * Find an entry in the syncache. 492 * Returns always with locked syncache_head plus a matching entry or NULL. 493 */ 494 static struct syncache * 495 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 496 { 497 struct syncache *sc; 498 struct syncache_head *sch; 499 500 #ifdef INET6 501 if (inc->inc_flags & INC_ISIPV6) { 502 sch = &V_tcp_syncache.hashbase[ 503 SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)]; 504 *schp = sch; 505 506 SCH_LOCK(sch); 507 508 /* Circle through bucket row to find matching entry. */ 509 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 510 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 511 return (sc); 512 } 513 } else 514 #endif 515 { 516 sch = &V_tcp_syncache.hashbase[ 517 SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)]; 518 *schp = sch; 519 520 SCH_LOCK(sch); 521 522 /* Circle through bucket row to find matching entry. */ 523 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 524 #ifdef INET6 525 if (sc->sc_inc.inc_flags & INC_ISIPV6) 526 continue; 527 #endif 528 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 529 return (sc); 530 } 531 } 532 SCH_LOCK_ASSERT(*schp); 533 return (NULL); /* always returns with locked sch */ 534 } 535 536 /* 537 * This function is called when we get a RST for a 538 * non-existent connection, so that we can see if the 539 * connection is in the syn cache. If it is, zap it. 540 */ 541 void 542 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 543 { 544 struct syncache *sc; 545 struct syncache_head *sch; 546 char *s = NULL; 547 548 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 549 SCH_LOCK_ASSERT(sch); 550 551 /* 552 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 553 * See RFC 793 page 65, section SEGMENT ARRIVES. 554 */ 555 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 556 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 557 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 558 "FIN flag set, segment ignored\n", s, __func__); 559 TCPSTAT_INC(tcps_badrst); 560 goto done; 561 } 562 563 /* 564 * No corresponding connection was found in syncache. 565 * If syncookies are enabled and possibly exclusively 566 * used, or we are under memory pressure, a valid RST 567 * may not find a syncache entry. In that case we're 568 * done and no SYN|ACK retransmissions will happen. 569 * Otherwise the RST was misdirected or spoofed. 570 */ 571 if (sc == NULL) { 572 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 573 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 574 "syncache entry (possibly syncookie only), " 575 "segment ignored\n", s, __func__); 576 TCPSTAT_INC(tcps_badrst); 577 goto done; 578 } 579 580 /* 581 * If the RST bit is set, check the sequence number to see 582 * if this is a valid reset segment. 583 * RFC 793 page 37: 584 * In all states except SYN-SENT, all reset (RST) segments 585 * are validated by checking their SEQ-fields. A reset is 586 * valid if its sequence number is in the window. 587 * 588 * The sequence number in the reset segment is normally an 589 * echo of our outgoing acknowlegement numbers, but some hosts 590 * send a reset with the sequence number at the rightmost edge 591 * of our receive window, and we have to handle this case. 592 */ 593 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 594 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 595 syncache_drop(sc, sch); 596 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 597 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 598 "connection attempt aborted by remote endpoint\n", 599 s, __func__); 600 TCPSTAT_INC(tcps_sc_reset); 601 } else { 602 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 603 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 604 "IRS %u (+WND %u), segment ignored\n", 605 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 606 TCPSTAT_INC(tcps_badrst); 607 } 608 609 done: 610 if (s != NULL) 611 free(s, M_TCPLOG); 612 SCH_UNLOCK(sch); 613 } 614 615 void 616 syncache_badack(struct in_conninfo *inc) 617 { 618 struct syncache *sc; 619 struct syncache_head *sch; 620 621 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 622 SCH_LOCK_ASSERT(sch); 623 if (sc != NULL) { 624 syncache_drop(sc, sch); 625 TCPSTAT_INC(tcps_sc_badack); 626 } 627 SCH_UNLOCK(sch); 628 } 629 630 void 631 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th) 632 { 633 struct syncache *sc; 634 struct syncache_head *sch; 635 636 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 637 SCH_LOCK_ASSERT(sch); 638 if (sc == NULL) 639 goto done; 640 641 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 642 if (ntohl(th->th_seq) != sc->sc_iss) 643 goto done; 644 645 /* 646 * If we've rertransmitted 3 times and this is our second error, 647 * we remove the entry. Otherwise, we allow it to continue on. 648 * This prevents us from incorrectly nuking an entry during a 649 * spurious network outage. 650 * 651 * See tcp_notify(). 652 */ 653 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 654 sc->sc_flags |= SCF_UNREACH; 655 goto done; 656 } 657 syncache_drop(sc, sch); 658 TCPSTAT_INC(tcps_sc_unreach); 659 done: 660 SCH_UNLOCK(sch); 661 } 662 663 /* 664 * Build a new TCP socket structure from a syncache entry. 665 */ 666 static struct socket * 667 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 668 { 669 struct inpcb *inp = NULL; 670 struct socket *so; 671 struct tcpcb *tp; 672 int error; 673 char *s; 674 675 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 676 677 /* 678 * Ok, create the full blown connection, and set things up 679 * as they would have been set up if we had created the 680 * connection when the SYN arrived. If we can't create 681 * the connection, abort it. 682 */ 683 so = sonewconn(lso, SS_ISCONNECTED); 684 if (so == NULL) { 685 /* 686 * Drop the connection; we will either send a RST or 687 * have the peer retransmit its SYN again after its 688 * RTO and try again. 689 */ 690 TCPSTAT_INC(tcps_listendrop); 691 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 692 log(LOG_DEBUG, "%s; %s: Socket create failed " 693 "due to limits or memory shortage\n", 694 s, __func__); 695 free(s, M_TCPLOG); 696 } 697 goto abort2; 698 } 699 #ifdef MAC 700 mac_socketpeer_set_from_mbuf(m, so); 701 #endif 702 703 inp = sotoinpcb(so); 704 inp->inp_inc.inc_fibnum = so->so_fibnum; 705 INP_WLOCK(inp); 706 INP_HASH_WLOCK(&V_tcbinfo); 707 708 /* Insert new socket into PCB hash list. */ 709 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 710 #ifdef INET6 711 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 712 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 713 } else { 714 inp->inp_vflag &= ~INP_IPV6; 715 inp->inp_vflag |= INP_IPV4; 716 #endif 717 inp->inp_laddr = sc->sc_inc.inc_laddr; 718 #ifdef INET6 719 } 720 #endif 721 722 /* 723 * If there's an mbuf and it has a flowid, then let's initialise the 724 * inp with that particular flowid. 725 */ 726 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 727 inp->inp_flowid = m->m_pkthdr.flowid; 728 inp->inp_flowtype = M_HASHTYPE_GET(m); 729 } 730 731 /* 732 * Install in the reservation hash table for now, but don't yet 733 * install a connection group since the full 4-tuple isn't yet 734 * configured. 735 */ 736 inp->inp_lport = sc->sc_inc.inc_lport; 737 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 738 /* 739 * Undo the assignments above if we failed to 740 * put the PCB on the hash lists. 741 */ 742 #ifdef INET6 743 if (sc->sc_inc.inc_flags & INC_ISIPV6) 744 inp->in6p_laddr = in6addr_any; 745 else 746 #endif 747 inp->inp_laddr.s_addr = INADDR_ANY; 748 inp->inp_lport = 0; 749 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 750 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 751 "with error %i\n", 752 s, __func__, error); 753 free(s, M_TCPLOG); 754 } 755 INP_HASH_WUNLOCK(&V_tcbinfo); 756 goto abort; 757 } 758 #ifdef IPSEC 759 /* Copy old policy into new socket's. */ 760 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp)) 761 printf("syncache_socket: could not copy policy\n"); 762 #endif 763 #ifdef INET6 764 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 765 struct inpcb *oinp = sotoinpcb(lso); 766 struct in6_addr laddr6; 767 struct sockaddr_in6 sin6; 768 /* 769 * Inherit socket options from the listening socket. 770 * Note that in6p_inputopts are not (and should not be) 771 * copied, since it stores previously received options and is 772 * used to detect if each new option is different than the 773 * previous one and hence should be passed to a user. 774 * If we copied in6p_inputopts, a user would not be able to 775 * receive options just after calling the accept system call. 776 */ 777 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 778 if (oinp->in6p_outputopts) 779 inp->in6p_outputopts = 780 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 781 782 sin6.sin6_family = AF_INET6; 783 sin6.sin6_len = sizeof(sin6); 784 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 785 sin6.sin6_port = sc->sc_inc.inc_fport; 786 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 787 laddr6 = inp->in6p_laddr; 788 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 789 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 790 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 791 thread0.td_ucred, m)) != 0) { 792 inp->in6p_laddr = laddr6; 793 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 794 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 795 "with error %i\n", 796 s, __func__, error); 797 free(s, M_TCPLOG); 798 } 799 INP_HASH_WUNLOCK(&V_tcbinfo); 800 goto abort; 801 } 802 /* Override flowlabel from in6_pcbconnect. */ 803 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 804 inp->inp_flow |= sc->sc_flowlabel; 805 } 806 #endif /* INET6 */ 807 #if defined(INET) && defined(INET6) 808 else 809 #endif 810 #ifdef INET 811 { 812 struct in_addr laddr; 813 struct sockaddr_in sin; 814 815 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 816 817 if (inp->inp_options == NULL) { 818 inp->inp_options = sc->sc_ipopts; 819 sc->sc_ipopts = NULL; 820 } 821 822 sin.sin_family = AF_INET; 823 sin.sin_len = sizeof(sin); 824 sin.sin_addr = sc->sc_inc.inc_faddr; 825 sin.sin_port = sc->sc_inc.inc_fport; 826 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 827 laddr = inp->inp_laddr; 828 if (inp->inp_laddr.s_addr == INADDR_ANY) 829 inp->inp_laddr = sc->sc_inc.inc_laddr; 830 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 831 thread0.td_ucred, m)) != 0) { 832 inp->inp_laddr = laddr; 833 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 834 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 835 "with error %i\n", 836 s, __func__, error); 837 free(s, M_TCPLOG); 838 } 839 INP_HASH_WUNLOCK(&V_tcbinfo); 840 goto abort; 841 } 842 } 843 #endif /* INET */ 844 INP_HASH_WUNLOCK(&V_tcbinfo); 845 tp = intotcpcb(inp); 846 tcp_state_change(tp, TCPS_SYN_RECEIVED); 847 tp->iss = sc->sc_iss; 848 tp->irs = sc->sc_irs; 849 tcp_rcvseqinit(tp); 850 tcp_sendseqinit(tp); 851 tp->snd_wl1 = sc->sc_irs; 852 tp->snd_max = tp->iss + 1; 853 tp->snd_nxt = tp->iss + 1; 854 tp->rcv_up = sc->sc_irs + 1; 855 tp->rcv_wnd = sc->sc_wnd; 856 tp->rcv_adv += tp->rcv_wnd; 857 tp->last_ack_sent = tp->rcv_nxt; 858 859 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 860 if (sc->sc_flags & SCF_NOOPT) 861 tp->t_flags |= TF_NOOPT; 862 else { 863 if (sc->sc_flags & SCF_WINSCALE) { 864 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 865 tp->snd_scale = sc->sc_requested_s_scale; 866 tp->request_r_scale = sc->sc_requested_r_scale; 867 } 868 if (sc->sc_flags & SCF_TIMESTAMP) { 869 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 870 tp->ts_recent = sc->sc_tsreflect; 871 tp->ts_recent_age = tcp_ts_getticks(); 872 tp->ts_offset = sc->sc_tsoff; 873 } 874 #ifdef TCP_SIGNATURE 875 if (sc->sc_flags & SCF_SIGNATURE) 876 tp->t_flags |= TF_SIGNATURE; 877 #endif 878 if (sc->sc_flags & SCF_SACK) 879 tp->t_flags |= TF_SACK_PERMIT; 880 } 881 882 if (sc->sc_flags & SCF_ECN) 883 tp->t_flags |= TF_ECN_PERMIT; 884 885 /* 886 * Set up MSS and get cached values from tcp_hostcache. 887 * This might overwrite some of the defaults we just set. 888 */ 889 tcp_mss(tp, sc->sc_peer_mss); 890 891 /* 892 * If the SYN,ACK was retransmitted, indicate that CWND to be 893 * limited to one segment in cc_conn_init(). 894 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 895 */ 896 if (sc->sc_rxmits > 1) 897 tp->snd_cwnd = 1; 898 899 #ifdef TCP_OFFLOAD 900 /* 901 * Allow a TOE driver to install its hooks. Note that we hold the 902 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 903 * new connection before the TOE driver has done its thing. 904 */ 905 if (ADDED_BY_TOE(sc)) { 906 struct toedev *tod = sc->sc_tod; 907 908 tod->tod_offload_socket(tod, sc->sc_todctx, so); 909 } 910 #endif 911 /* 912 * Copy and activate timers. 913 */ 914 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 915 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 916 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 917 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 918 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 919 920 INP_WUNLOCK(inp); 921 922 TCPSTAT_INC(tcps_accepts); 923 return (so); 924 925 abort: 926 INP_WUNLOCK(inp); 927 abort2: 928 if (so != NULL) 929 soabort(so); 930 return (NULL); 931 } 932 933 /* 934 * This function gets called when we receive an ACK for a 935 * socket in the LISTEN state. We look up the connection 936 * in the syncache, and if its there, we pull it out of 937 * the cache and turn it into a full-blown connection in 938 * the SYN-RECEIVED state. 939 */ 940 int 941 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 942 struct socket **lsop, struct mbuf *m) 943 { 944 struct syncache *sc; 945 struct syncache_head *sch; 946 struct syncache scs; 947 char *s; 948 949 /* 950 * Global TCP locks are held because we manipulate the PCB lists 951 * and create a new socket. 952 */ 953 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 954 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 955 ("%s: can handle only ACK", __func__)); 956 957 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 958 SCH_LOCK_ASSERT(sch); 959 960 #ifdef INVARIANTS 961 /* 962 * Test code for syncookies comparing the syncache stored 963 * values with the reconstructed values from the cookie. 964 */ 965 if (sc != NULL) 966 syncookie_cmp(inc, sch, sc, th, to, *lsop); 967 #endif 968 969 if (sc == NULL) { 970 /* 971 * There is no syncache entry, so see if this ACK is 972 * a returning syncookie. To do this, first: 973 * A. See if this socket has had a syncache entry dropped in 974 * the past. We don't want to accept a bogus syncookie 975 * if we've never received a SYN. 976 * B. check that the syncookie is valid. If it is, then 977 * cobble up a fake syncache entry, and return. 978 */ 979 if (!V_tcp_syncookies) { 980 SCH_UNLOCK(sch); 981 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 982 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 983 "segment rejected (syncookies disabled)\n", 984 s, __func__); 985 goto failed; 986 } 987 bzero(&scs, sizeof(scs)); 988 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 989 SCH_UNLOCK(sch); 990 if (sc == NULL) { 991 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 992 log(LOG_DEBUG, "%s; %s: Segment failed " 993 "SYNCOOKIE authentication, segment rejected " 994 "(probably spoofed)\n", s, __func__); 995 goto failed; 996 } 997 } else { 998 /* Pull out the entry to unlock the bucket row. */ 999 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1000 sch->sch_length--; 1001 #ifdef TCP_OFFLOAD 1002 if (ADDED_BY_TOE(sc)) { 1003 struct toedev *tod = sc->sc_tod; 1004 1005 tod->tod_syncache_removed(tod, sc->sc_todctx); 1006 } 1007 #endif 1008 SCH_UNLOCK(sch); 1009 } 1010 1011 /* 1012 * Segment validation: 1013 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1014 */ 1015 if (th->th_ack != sc->sc_iss + 1) { 1016 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1017 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1018 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1019 goto failed; 1020 } 1021 1022 /* 1023 * The SEQ must fall in the window starting at the received 1024 * initial receive sequence number + 1 (the SYN). 1025 */ 1026 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1027 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1028 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1029 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1030 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1031 goto failed; 1032 } 1033 1034 /* 1035 * If timestamps were not negotiated during SYN/ACK they 1036 * must not appear on any segment during this session. 1037 */ 1038 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1039 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1040 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1041 "segment rejected\n", s, __func__); 1042 goto failed; 1043 } 1044 1045 /* 1046 * If timestamps were negotiated during SYN/ACK they should 1047 * appear on every segment during this session. 1048 * XXXAO: This is only informal as there have been unverified 1049 * reports of non-compliants stacks. 1050 */ 1051 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1052 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1053 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1054 "no action\n", s, __func__); 1055 free(s, M_TCPLOG); 1056 s = NULL; 1057 } 1058 } 1059 1060 /* 1061 * If timestamps were negotiated the reflected timestamp 1062 * must be equal to what we actually sent in the SYN|ACK. 1063 */ 1064 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) { 1065 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1066 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1067 "segment rejected\n", 1068 s, __func__, to->to_tsecr, sc->sc_ts); 1069 goto failed; 1070 } 1071 1072 *lsop = syncache_socket(sc, *lsop, m); 1073 1074 if (*lsop == NULL) 1075 TCPSTAT_INC(tcps_sc_aborted); 1076 else 1077 TCPSTAT_INC(tcps_sc_completed); 1078 1079 /* how do we find the inp for the new socket? */ 1080 if (sc != &scs) 1081 syncache_free(sc); 1082 return (1); 1083 failed: 1084 if (sc != NULL && sc != &scs) 1085 syncache_free(sc); 1086 if (s != NULL) 1087 free(s, M_TCPLOG); 1088 *lsop = NULL; 1089 return (0); 1090 } 1091 1092 /* 1093 * Given a LISTEN socket and an inbound SYN request, add 1094 * this to the syn cache, and send back a segment: 1095 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1096 * to the source. 1097 * 1098 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1099 * Doing so would require that we hold onto the data and deliver it 1100 * to the application. However, if we are the target of a SYN-flood 1101 * DoS attack, an attacker could send data which would eventually 1102 * consume all available buffer space if it were ACKed. By not ACKing 1103 * the data, we avoid this DoS scenario. 1104 */ 1105 void 1106 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1107 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1108 void *todctx) 1109 { 1110 struct tcpcb *tp; 1111 struct socket *so; 1112 struct syncache *sc = NULL; 1113 struct syncache_head *sch; 1114 struct mbuf *ipopts = NULL; 1115 u_int ltflags; 1116 int win, sb_hiwat, ip_ttl, ip_tos; 1117 char *s; 1118 #ifdef INET6 1119 int autoflowlabel = 0; 1120 #endif 1121 #ifdef MAC 1122 struct label *maclabel; 1123 #endif 1124 struct syncache scs; 1125 struct ucred *cred; 1126 1127 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 1128 INP_WLOCK_ASSERT(inp); /* listen socket */ 1129 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1130 ("%s: unexpected tcp flags", __func__)); 1131 1132 /* 1133 * Combine all so/tp operations very early to drop the INP lock as 1134 * soon as possible. 1135 */ 1136 so = *lsop; 1137 tp = sototcpcb(so); 1138 cred = crhold(so->so_cred); 1139 1140 #ifdef INET6 1141 if ((inc->inc_flags & INC_ISIPV6) && 1142 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1143 autoflowlabel = 1; 1144 #endif 1145 ip_ttl = inp->inp_ip_ttl; 1146 ip_tos = inp->inp_ip_tos; 1147 win = sbspace(&so->so_rcv); 1148 sb_hiwat = so->so_rcv.sb_hiwat; 1149 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1150 1151 /* By the time we drop the lock these should no longer be used. */ 1152 so = NULL; 1153 tp = NULL; 1154 1155 #ifdef MAC 1156 if (mac_syncache_init(&maclabel) != 0) { 1157 INP_WUNLOCK(inp); 1158 INP_INFO_WUNLOCK(&V_tcbinfo); 1159 goto done; 1160 } else 1161 mac_syncache_create(maclabel, inp); 1162 #endif 1163 INP_WUNLOCK(inp); 1164 INP_INFO_WUNLOCK(&V_tcbinfo); 1165 1166 /* 1167 * Remember the IP options, if any. 1168 */ 1169 #ifdef INET6 1170 if (!(inc->inc_flags & INC_ISIPV6)) 1171 #endif 1172 #ifdef INET 1173 ipopts = (m) ? ip_srcroute(m) : NULL; 1174 #else 1175 ipopts = NULL; 1176 #endif 1177 1178 /* 1179 * See if we already have an entry for this connection. 1180 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1181 * 1182 * XXX: should the syncache be re-initialized with the contents 1183 * of the new SYN here (which may have different options?) 1184 * 1185 * XXX: We do not check the sequence number to see if this is a 1186 * real retransmit or a new connection attempt. The question is 1187 * how to handle such a case; either ignore it as spoofed, or 1188 * drop the current entry and create a new one? 1189 */ 1190 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1191 SCH_LOCK_ASSERT(sch); 1192 if (sc != NULL) { 1193 TCPSTAT_INC(tcps_sc_dupsyn); 1194 if (ipopts) { 1195 /* 1196 * If we were remembering a previous source route, 1197 * forget it and use the new one we've been given. 1198 */ 1199 if (sc->sc_ipopts) 1200 (void) m_free(sc->sc_ipopts); 1201 sc->sc_ipopts = ipopts; 1202 } 1203 /* 1204 * Update timestamp if present. 1205 */ 1206 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1207 sc->sc_tsreflect = to->to_tsval; 1208 else 1209 sc->sc_flags &= ~SCF_TIMESTAMP; 1210 #ifdef MAC 1211 /* 1212 * Since we have already unconditionally allocated label 1213 * storage, free it up. The syncache entry will already 1214 * have an initialized label we can use. 1215 */ 1216 mac_syncache_destroy(&maclabel); 1217 #endif 1218 /* Retransmit SYN|ACK and reset retransmit count. */ 1219 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1220 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1221 "resetting timer and retransmitting SYN|ACK\n", 1222 s, __func__); 1223 free(s, M_TCPLOG); 1224 } 1225 if (syncache_respond(sc) == 0) { 1226 sc->sc_rxmits = 0; 1227 syncache_timeout(sc, sch, 1); 1228 TCPSTAT_INC(tcps_sndacks); 1229 TCPSTAT_INC(tcps_sndtotal); 1230 } 1231 SCH_UNLOCK(sch); 1232 goto done; 1233 } 1234 1235 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1236 if (sc == NULL) { 1237 /* 1238 * The zone allocator couldn't provide more entries. 1239 * Treat this as if the cache was full; drop the oldest 1240 * entry and insert the new one. 1241 */ 1242 TCPSTAT_INC(tcps_sc_zonefail); 1243 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) 1244 syncache_drop(sc, sch); 1245 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1246 if (sc == NULL) { 1247 if (V_tcp_syncookies) { 1248 bzero(&scs, sizeof(scs)); 1249 sc = &scs; 1250 } else { 1251 SCH_UNLOCK(sch); 1252 if (ipopts) 1253 (void) m_free(ipopts); 1254 goto done; 1255 } 1256 } 1257 } 1258 1259 /* 1260 * Fill in the syncache values. 1261 */ 1262 #ifdef MAC 1263 sc->sc_label = maclabel; 1264 #endif 1265 sc->sc_cred = cred; 1266 cred = NULL; 1267 sc->sc_ipopts = ipopts; 1268 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1269 #ifdef INET6 1270 if (!(inc->inc_flags & INC_ISIPV6)) 1271 #endif 1272 { 1273 sc->sc_ip_tos = ip_tos; 1274 sc->sc_ip_ttl = ip_ttl; 1275 } 1276 #ifdef TCP_OFFLOAD 1277 sc->sc_tod = tod; 1278 sc->sc_todctx = todctx; 1279 #endif 1280 sc->sc_irs = th->th_seq; 1281 sc->sc_iss = arc4random(); 1282 sc->sc_flags = 0; 1283 sc->sc_flowlabel = 0; 1284 1285 /* 1286 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1287 * win was derived from socket earlier in the function. 1288 */ 1289 win = imax(win, 0); 1290 win = imin(win, TCP_MAXWIN); 1291 sc->sc_wnd = win; 1292 1293 if (V_tcp_do_rfc1323) { 1294 /* 1295 * A timestamp received in a SYN makes 1296 * it ok to send timestamp requests and replies. 1297 */ 1298 if (to->to_flags & TOF_TS) { 1299 sc->sc_tsreflect = to->to_tsval; 1300 sc->sc_ts = tcp_ts_getticks(); 1301 sc->sc_flags |= SCF_TIMESTAMP; 1302 } 1303 if (to->to_flags & TOF_SCALE) { 1304 int wscale = 0; 1305 1306 /* 1307 * Pick the smallest possible scaling factor that 1308 * will still allow us to scale up to sb_max, aka 1309 * kern.ipc.maxsockbuf. 1310 * 1311 * We do this because there are broken firewalls that 1312 * will corrupt the window scale option, leading to 1313 * the other endpoint believing that our advertised 1314 * window is unscaled. At scale factors larger than 1315 * 5 the unscaled window will drop below 1500 bytes, 1316 * leading to serious problems when traversing these 1317 * broken firewalls. 1318 * 1319 * With the default maxsockbuf of 256K, a scale factor 1320 * of 3 will be chosen by this algorithm. Those who 1321 * choose a larger maxsockbuf should watch out 1322 * for the compatiblity problems mentioned above. 1323 * 1324 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1325 * or <SYN,ACK>) segment itself is never scaled. 1326 */ 1327 while (wscale < TCP_MAX_WINSHIFT && 1328 (TCP_MAXWIN << wscale) < sb_max) 1329 wscale++; 1330 sc->sc_requested_r_scale = wscale; 1331 sc->sc_requested_s_scale = to->to_wscale; 1332 sc->sc_flags |= SCF_WINSCALE; 1333 } 1334 } 1335 #ifdef TCP_SIGNATURE 1336 /* 1337 * If listening socket requested TCP digests, and received SYN 1338 * contains the option, flag this in the syncache so that 1339 * syncache_respond() will do the right thing with the SYN+ACK. 1340 * XXX: Currently we always record the option by default and will 1341 * attempt to use it in syncache_respond(). 1342 */ 1343 if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE) 1344 sc->sc_flags |= SCF_SIGNATURE; 1345 #endif 1346 if (to->to_flags & TOF_SACKPERM) 1347 sc->sc_flags |= SCF_SACK; 1348 if (to->to_flags & TOF_MSS) 1349 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1350 if (ltflags & TF_NOOPT) 1351 sc->sc_flags |= SCF_NOOPT; 1352 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1353 sc->sc_flags |= SCF_ECN; 1354 1355 if (V_tcp_syncookies) 1356 sc->sc_iss = syncookie_generate(sch, sc); 1357 #ifdef INET6 1358 if (autoflowlabel) { 1359 if (V_tcp_syncookies) 1360 sc->sc_flowlabel = sc->sc_iss; 1361 else 1362 sc->sc_flowlabel = ip6_randomflowlabel(); 1363 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1364 } 1365 #endif 1366 SCH_UNLOCK(sch); 1367 1368 /* 1369 * Do a standard 3-way handshake. 1370 */ 1371 if (syncache_respond(sc) == 0) { 1372 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1373 syncache_free(sc); 1374 else if (sc != &scs) 1375 syncache_insert(sc, sch); /* locks and unlocks sch */ 1376 TCPSTAT_INC(tcps_sndacks); 1377 TCPSTAT_INC(tcps_sndtotal); 1378 } else { 1379 if (sc != &scs) 1380 syncache_free(sc); 1381 TCPSTAT_INC(tcps_sc_dropped); 1382 } 1383 1384 done: 1385 if (cred != NULL) 1386 crfree(cred); 1387 #ifdef MAC 1388 if (sc == &scs) 1389 mac_syncache_destroy(&maclabel); 1390 #endif 1391 if (m) { 1392 1393 *lsop = NULL; 1394 m_freem(m); 1395 } 1396 } 1397 1398 static int 1399 syncache_respond(struct syncache *sc) 1400 { 1401 struct ip *ip = NULL; 1402 struct mbuf *m; 1403 struct tcphdr *th = NULL; 1404 int optlen, error = 0; /* Make compiler happy */ 1405 u_int16_t hlen, tlen, mssopt; 1406 struct tcpopt to; 1407 #ifdef INET6 1408 struct ip6_hdr *ip6 = NULL; 1409 #endif 1410 1411 hlen = 1412 #ifdef INET6 1413 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1414 #endif 1415 sizeof(struct ip); 1416 tlen = hlen + sizeof(struct tcphdr); 1417 1418 /* Determine MSS we advertize to other end of connection. */ 1419 mssopt = tcp_mssopt(&sc->sc_inc); 1420 if (sc->sc_peer_mss) 1421 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss); 1422 1423 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1424 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1425 ("syncache: mbuf too small")); 1426 1427 /* Create the IP+TCP header from scratch. */ 1428 m = m_gethdr(M_NOWAIT, MT_DATA); 1429 if (m == NULL) 1430 return (ENOBUFS); 1431 #ifdef MAC 1432 mac_syncache_create_mbuf(sc->sc_label, m); 1433 #endif 1434 m->m_data += max_linkhdr; 1435 m->m_len = tlen; 1436 m->m_pkthdr.len = tlen; 1437 m->m_pkthdr.rcvif = NULL; 1438 1439 #ifdef INET6 1440 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1441 ip6 = mtod(m, struct ip6_hdr *); 1442 ip6->ip6_vfc = IPV6_VERSION; 1443 ip6->ip6_nxt = IPPROTO_TCP; 1444 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1445 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1446 ip6->ip6_plen = htons(tlen - hlen); 1447 /* ip6_hlim is set after checksum */ 1448 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1449 ip6->ip6_flow |= sc->sc_flowlabel; 1450 1451 th = (struct tcphdr *)(ip6 + 1); 1452 } 1453 #endif 1454 #if defined(INET6) && defined(INET) 1455 else 1456 #endif 1457 #ifdef INET 1458 { 1459 ip = mtod(m, struct ip *); 1460 ip->ip_v = IPVERSION; 1461 ip->ip_hl = sizeof(struct ip) >> 2; 1462 ip->ip_len = htons(tlen); 1463 ip->ip_id = 0; 1464 ip->ip_off = 0; 1465 ip->ip_sum = 0; 1466 ip->ip_p = IPPROTO_TCP; 1467 ip->ip_src = sc->sc_inc.inc_laddr; 1468 ip->ip_dst = sc->sc_inc.inc_faddr; 1469 ip->ip_ttl = sc->sc_ip_ttl; 1470 ip->ip_tos = sc->sc_ip_tos; 1471 1472 /* 1473 * See if we should do MTU discovery. Route lookups are 1474 * expensive, so we will only unset the DF bit if: 1475 * 1476 * 1) path_mtu_discovery is disabled 1477 * 2) the SCF_UNREACH flag has been set 1478 */ 1479 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1480 ip->ip_off |= htons(IP_DF); 1481 1482 th = (struct tcphdr *)(ip + 1); 1483 } 1484 #endif /* INET */ 1485 th->th_sport = sc->sc_inc.inc_lport; 1486 th->th_dport = sc->sc_inc.inc_fport; 1487 1488 th->th_seq = htonl(sc->sc_iss); 1489 th->th_ack = htonl(sc->sc_irs + 1); 1490 th->th_off = sizeof(struct tcphdr) >> 2; 1491 th->th_x2 = 0; 1492 th->th_flags = TH_SYN|TH_ACK; 1493 th->th_win = htons(sc->sc_wnd); 1494 th->th_urp = 0; 1495 1496 if (sc->sc_flags & SCF_ECN) { 1497 th->th_flags |= TH_ECE; 1498 TCPSTAT_INC(tcps_ecn_shs); 1499 } 1500 1501 /* Tack on the TCP options. */ 1502 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1503 to.to_flags = 0; 1504 1505 to.to_mss = mssopt; 1506 to.to_flags = TOF_MSS; 1507 if (sc->sc_flags & SCF_WINSCALE) { 1508 to.to_wscale = sc->sc_requested_r_scale; 1509 to.to_flags |= TOF_SCALE; 1510 } 1511 if (sc->sc_flags & SCF_TIMESTAMP) { 1512 /* Virgin timestamp or TCP cookie enhanced one. */ 1513 to.to_tsval = sc->sc_ts; 1514 to.to_tsecr = sc->sc_tsreflect; 1515 to.to_flags |= TOF_TS; 1516 } 1517 if (sc->sc_flags & SCF_SACK) 1518 to.to_flags |= TOF_SACKPERM; 1519 #ifdef TCP_SIGNATURE 1520 if (sc->sc_flags & SCF_SIGNATURE) 1521 to.to_flags |= TOF_SIGNATURE; 1522 #endif 1523 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1524 1525 /* Adjust headers by option size. */ 1526 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1527 m->m_len += optlen; 1528 m->m_pkthdr.len += optlen; 1529 1530 #ifdef TCP_SIGNATURE 1531 if (sc->sc_flags & SCF_SIGNATURE) 1532 tcp_signature_compute(m, 0, 0, optlen, 1533 to.to_signature, IPSEC_DIR_OUTBOUND); 1534 #endif 1535 #ifdef INET6 1536 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1537 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1538 else 1539 #endif 1540 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1541 } else 1542 optlen = 0; 1543 1544 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1545 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1546 #ifdef INET6 1547 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1548 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1549 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1550 IPPROTO_TCP, 0); 1551 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1552 #ifdef TCP_OFFLOAD 1553 if (ADDED_BY_TOE(sc)) { 1554 struct toedev *tod = sc->sc_tod; 1555 1556 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1557 1558 return (error); 1559 } 1560 #endif 1561 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1562 } 1563 #endif 1564 #if defined(INET6) && defined(INET) 1565 else 1566 #endif 1567 #ifdef INET 1568 { 1569 m->m_pkthdr.csum_flags = CSUM_TCP; 1570 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1571 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1572 #ifdef TCP_OFFLOAD 1573 if (ADDED_BY_TOE(sc)) { 1574 struct toedev *tod = sc->sc_tod; 1575 1576 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1577 1578 return (error); 1579 } 1580 #endif 1581 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1582 } 1583 #endif 1584 return (error); 1585 } 1586 1587 /* 1588 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1589 * that exceed the capacity of the syncache by avoiding the storage of any 1590 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1591 * attacks where the attacker does not have access to our responses. 1592 * 1593 * Syncookies encode and include all necessary information about the 1594 * connection setup within the SYN|ACK that we send back. That way we 1595 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1596 * (if ever). Normally the syncache and syncookies are running in parallel 1597 * with the latter taking over when the former is exhausted. When matching 1598 * syncache entry is found the syncookie is ignored. 1599 * 1600 * The only reliable information persisting the 3WHS is our inital sequence 1601 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1602 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1603 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1604 * returns and signifies a legitimate connection if it matches the ACK. 1605 * 1606 * The available space of 32 bits to store the hash and to encode the SYN 1607 * option information is very tight and we should have at least 24 bits for 1608 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1609 * 1610 * SYN option information we have to encode to fully restore a connection: 1611 * MSS: is imporant to chose an optimal segment size to avoid IP level 1612 * fragmentation along the path. The common MSS values can be encoded 1613 * in a 3-bit table. Uncommon values are captured by the next lower value 1614 * in the table leading to a slight increase in packetization overhead. 1615 * WSCALE: is necessary to allow large windows to be used for high delay- 1616 * bandwidth product links. Not scaling the window when it was initially 1617 * negotiated is bad for performance as lack of scaling further decreases 1618 * the apparent available send window. We only need to encode the WSCALE 1619 * we received from the remote end. Our end can be recalculated at any 1620 * time. The common WSCALE values can be encoded in a 3-bit table. 1621 * Uncommon values are captured by the next lower value in the table 1622 * making us under-estimate the available window size halving our 1623 * theoretically possible maximum throughput for that connection. 1624 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1625 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1626 * that are included in all segments on a connection. We enable them when 1627 * the ACK has them. 1628 * 1629 * Security of syncookies and attack vectors: 1630 * 1631 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1632 * together with the gloabl secret to make it unique per connection attempt. 1633 * Thus any change of any of those parameters results in a different MAC output 1634 * in an unpredictable way unless a collision is encountered. 24 bits of the 1635 * MAC are embedded into the ISS. 1636 * 1637 * To prevent replay attacks two rotating global secrets are updated with a 1638 * new random value every 15 seconds. The life-time of a syncookie is thus 1639 * 15-30 seconds. 1640 * 1641 * Vector 1: Attacking the secret. This requires finding a weakness in the 1642 * MAC itself or the way it is used here. The attacker can do a chosen plain 1643 * text attack by varying and testing the all parameters under his control. 1644 * The strength depends on the size and randomness of the secret, and the 1645 * cryptographic security of the MAC function. Due to the constant updating 1646 * of the secret the attacker has at most 29.999 seconds to find the secret 1647 * and launch spoofed connections. After that he has to start all over again. 1648 * 1649 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1650 * size an average of 4,823 attempts are required for a 50% chance of success 1651 * to spoof a single syncookie (birthday collision paradox). However the 1652 * attacker is blind and doesn't know if one of his attempts succeeded unless 1653 * he has a side channel to interfere success from. A single connection setup 1654 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1655 * This many attempts are required for each one blind spoofed connection. For 1656 * every additional spoofed connection he has to launch another N attempts. 1657 * Thus for a sustained rate 100 spoofed connections per second approximately 1658 * 1,800,000 packets per second would have to be sent. 1659 * 1660 * NB: The MAC function should be fast so that it doesn't become a CPU 1661 * exhaustion attack vector itself. 1662 * 1663 * References: 1664 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1665 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1666 * http://cr.yp.to/syncookies.html (overview) 1667 * http://cr.yp.to/syncookies/archive (details) 1668 * 1669 * 1670 * Schematic construction of a syncookie enabled Initial Sequence Number: 1671 * 0 1 2 3 1672 * 12345678901234567890123456789012 1673 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1674 * 1675 * x 24 MAC (truncated) 1676 * W 3 Send Window Scale index 1677 * M 3 MSS index 1678 * S 1 SACK permitted 1679 * P 1 Odd/even secret 1680 */ 1681 1682 /* 1683 * Distribution and probability of certain MSS values. Those in between are 1684 * rounded down to the next lower one. 1685 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1686 * .2% .3% 5% 7% 7% 20% 15% 45% 1687 */ 1688 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1689 1690 /* 1691 * Distribution and probability of certain WSCALE values. We have to map the 1692 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1693 * bits based on prevalence of certain values. Where we don't have an exact 1694 * match for are rounded down to the next lower one letting us under-estimate 1695 * the true available window. At the moment this would happen only for the 1696 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1697 * and window size). The absence of the WSCALE option (no scaling in either 1698 * direction) is encoded with index zero. 1699 * [WSCALE values histograms, Allman, 2012] 1700 * X 10 10 35 5 6 14 10% by host 1701 * X 11 4 5 5 18 49 3% by connections 1702 */ 1703 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1704 1705 /* 1706 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1707 * and good cryptographic properties. 1708 */ 1709 static uint32_t 1710 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1711 uint8_t *secbits, uintptr_t secmod) 1712 { 1713 SIPHASH_CTX ctx; 1714 uint32_t siphash[2]; 1715 1716 SipHash24_Init(&ctx); 1717 SipHash_SetKey(&ctx, secbits); 1718 switch (inc->inc_flags & INC_ISIPV6) { 1719 #ifdef INET 1720 case 0: 1721 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1722 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1723 break; 1724 #endif 1725 #ifdef INET6 1726 case INC_ISIPV6: 1727 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1728 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1729 break; 1730 #endif 1731 } 1732 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1733 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1734 SipHash_Update(&ctx, &flags, sizeof(flags)); 1735 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1736 SipHash_Final((u_int8_t *)&siphash, &ctx); 1737 1738 return (siphash[0] ^ siphash[1]); 1739 } 1740 1741 static tcp_seq 1742 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 1743 { 1744 u_int i, mss, secbit, wscale; 1745 uint32_t iss, hash; 1746 uint8_t *secbits; 1747 union syncookie cookie; 1748 1749 SCH_LOCK_ASSERT(sch); 1750 1751 cookie.cookie = 0; 1752 1753 /* Map our computed MSS into the 3-bit index. */ 1754 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss)); 1755 for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1; 1756 tcp_sc_msstab[i] > mss && i > 0; 1757 i--) 1758 ; 1759 cookie.flags.mss_idx = i; 1760 1761 /* 1762 * Map the send window scale into the 3-bit index but only if 1763 * the wscale option was received. 1764 */ 1765 if (sc->sc_flags & SCF_WINSCALE) { 1766 wscale = sc->sc_requested_s_scale; 1767 for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1; 1768 tcp_sc_wstab[i] > wscale && i > 0; 1769 i--) 1770 ; 1771 cookie.flags.wscale_idx = i; 1772 } 1773 1774 /* Can we do SACK? */ 1775 if (sc->sc_flags & SCF_SACK) 1776 cookie.flags.sack_ok = 1; 1777 1778 /* Which of the two secrets to use. */ 1779 secbit = sch->sch_sc->secret.oddeven & 0x1; 1780 cookie.flags.odd_even = secbit; 1781 1782 secbits = sch->sch_sc->secret.key[secbit]; 1783 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 1784 (uintptr_t)sch); 1785 1786 /* 1787 * Put the flags into the hash and XOR them to get better ISS number 1788 * variance. This doesn't enhance the cryptographic strength and is 1789 * done to prevent the 8 cookie bits from showing up directly on the 1790 * wire. 1791 */ 1792 iss = hash & ~0xff; 1793 iss |= cookie.cookie ^ (hash >> 24); 1794 1795 /* Randomize the timestamp. */ 1796 if (sc->sc_flags & SCF_TIMESTAMP) { 1797 sc->sc_ts = arc4random(); 1798 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 1799 } 1800 1801 TCPSTAT_INC(tcps_sc_sendcookie); 1802 return (iss); 1803 } 1804 1805 static struct syncache * 1806 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 1807 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 1808 struct socket *lso) 1809 { 1810 uint32_t hash; 1811 uint8_t *secbits; 1812 tcp_seq ack, seq; 1813 int wnd, wscale = 0; 1814 union syncookie cookie; 1815 1816 SCH_LOCK_ASSERT(sch); 1817 1818 /* 1819 * Pull information out of SYN-ACK/ACK and revert sequence number 1820 * advances. 1821 */ 1822 ack = th->th_ack - 1; 1823 seq = th->th_seq - 1; 1824 1825 /* 1826 * Unpack the flags containing enough information to restore the 1827 * connection. 1828 */ 1829 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 1830 1831 /* Which of the two secrets to use. */ 1832 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 1833 1834 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 1835 1836 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 1837 if ((ack & ~0xff) != (hash & ~0xff)) 1838 return (NULL); 1839 1840 /* Fill in the syncache values. */ 1841 sc->sc_flags = 0; 1842 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1843 sc->sc_ipopts = NULL; 1844 1845 sc->sc_irs = seq; 1846 sc->sc_iss = ack; 1847 1848 switch (inc->inc_flags & INC_ISIPV6) { 1849 #ifdef INET 1850 case 0: 1851 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 1852 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 1853 break; 1854 #endif 1855 #ifdef INET6 1856 case INC_ISIPV6: 1857 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 1858 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 1859 break; 1860 #endif 1861 } 1862 1863 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 1864 1865 /* We can simply recompute receive window scale we sent earlier. */ 1866 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 1867 wscale++; 1868 1869 /* Only use wscale if it was enabled in the orignal SYN. */ 1870 if (cookie.flags.wscale_idx > 0) { 1871 sc->sc_requested_r_scale = wscale; 1872 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 1873 sc->sc_flags |= SCF_WINSCALE; 1874 } 1875 1876 wnd = sbspace(&lso->so_rcv); 1877 wnd = imax(wnd, 0); 1878 wnd = imin(wnd, TCP_MAXWIN); 1879 sc->sc_wnd = wnd; 1880 1881 if (cookie.flags.sack_ok) 1882 sc->sc_flags |= SCF_SACK; 1883 1884 if (to->to_flags & TOF_TS) { 1885 sc->sc_flags |= SCF_TIMESTAMP; 1886 sc->sc_tsreflect = to->to_tsval; 1887 sc->sc_ts = to->to_tsecr; 1888 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 1889 } 1890 1891 if (to->to_flags & TOF_SIGNATURE) 1892 sc->sc_flags |= SCF_SIGNATURE; 1893 1894 sc->sc_rxmits = 0; 1895 1896 TCPSTAT_INC(tcps_sc_recvcookie); 1897 return (sc); 1898 } 1899 1900 #ifdef INVARIANTS 1901 static int 1902 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 1903 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 1904 struct socket *lso) 1905 { 1906 struct syncache scs, *scx; 1907 char *s; 1908 1909 bzero(&scs, sizeof(scs)); 1910 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 1911 1912 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 1913 return (0); 1914 1915 if (scx != NULL) { 1916 if (sc->sc_peer_mss != scx->sc_peer_mss) 1917 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 1918 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 1919 1920 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 1921 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 1922 s, __func__, sc->sc_requested_r_scale, 1923 scx->sc_requested_r_scale); 1924 1925 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 1926 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 1927 s, __func__, sc->sc_requested_s_scale, 1928 scx->sc_requested_s_scale); 1929 1930 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 1931 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 1932 } 1933 1934 if (s != NULL) 1935 free(s, M_TCPLOG); 1936 return (0); 1937 } 1938 #endif /* INVARIANTS */ 1939 1940 static void 1941 syncookie_reseed(void *arg) 1942 { 1943 struct tcp_syncache *sc = arg; 1944 uint8_t *secbits; 1945 int secbit; 1946 1947 /* 1948 * Reseeding the secret doesn't have to be protected by a lock. 1949 * It only must be ensured that the new random values are visible 1950 * to all CPUs in a SMP environment. The atomic with release 1951 * semantics ensures that. 1952 */ 1953 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 1954 secbits = sc->secret.key[secbit]; 1955 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 1956 atomic_add_rel_int(&sc->secret.oddeven, 1); 1957 1958 /* Reschedule ourself. */ 1959 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 1960 } 1961 1962 /* 1963 * Returns the current number of syncache entries. This number 1964 * will probably change before you get around to calling 1965 * syncache_pcblist. 1966 */ 1967 int 1968 syncache_pcbcount(void) 1969 { 1970 struct syncache_head *sch; 1971 int count, i; 1972 1973 for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 1974 /* No need to lock for a read. */ 1975 sch = &V_tcp_syncache.hashbase[i]; 1976 count += sch->sch_length; 1977 } 1978 return count; 1979 } 1980 1981 /* 1982 * Exports the syncache entries to userland so that netstat can display 1983 * them alongside the other sockets. This function is intended to be 1984 * called only from tcp_pcblist. 1985 * 1986 * Due to concurrency on an active system, the number of pcbs exported 1987 * may have no relation to max_pcbs. max_pcbs merely indicates the 1988 * amount of space the caller allocated for this function to use. 1989 */ 1990 int 1991 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 1992 { 1993 struct xtcpcb xt; 1994 struct syncache *sc; 1995 struct syncache_head *sch; 1996 int count, error, i; 1997 1998 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 1999 sch = &V_tcp_syncache.hashbase[i]; 2000 SCH_LOCK(sch); 2001 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2002 if (count >= max_pcbs) { 2003 SCH_UNLOCK(sch); 2004 goto exit; 2005 } 2006 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2007 continue; 2008 bzero(&xt, sizeof(xt)); 2009 xt.xt_len = sizeof(xt); 2010 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2011 xt.xt_inp.inp_vflag = INP_IPV6; 2012 else 2013 xt.xt_inp.inp_vflag = INP_IPV4; 2014 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 2015 xt.xt_tp.t_inpcb = &xt.xt_inp; 2016 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 2017 xt.xt_socket.xso_protocol = IPPROTO_TCP; 2018 xt.xt_socket.xso_len = sizeof (struct xsocket); 2019 xt.xt_socket.so_type = SOCK_STREAM; 2020 xt.xt_socket.so_state = SS_ISCONNECTING; 2021 error = SYSCTL_OUT(req, &xt, sizeof xt); 2022 if (error) { 2023 SCH_UNLOCK(sch); 2024 goto exit; 2025 } 2026 count++; 2027 } 2028 SCH_UNLOCK(sch); 2029 } 2030 exit: 2031 *pcbs_exported = count; 2032 return error; 2033 }
Cache object: 3b24daa4701696d0b16162b4e5f3ab07
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]