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("$FreeBSD$"); 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/if_var.h> 64 #include <net/route.h> 65 #include <net/vnet.h> 66 67 #include <netinet/in.h> 68 #include <netinet/in_systm.h> 69 #include <netinet/ip.h> 70 #include <netinet/in_var.h> 71 #include <netinet/in_pcb.h> 72 #include <netinet/ip_var.h> 73 #include <netinet/ip_options.h> 74 #ifdef INET6 75 #include <netinet/ip6.h> 76 #include <netinet/icmp6.h> 77 #include <netinet6/nd6.h> 78 #include <netinet6/ip6_var.h> 79 #include <netinet6/in6_pcb.h> 80 #endif 81 #include <netinet/tcp.h> 82 #include <netinet/tcp_fsm.h> 83 #include <netinet/tcp_seq.h> 84 #include <netinet/tcp_timer.h> 85 #include <netinet/tcp_var.h> 86 #include <netinet/tcp_syncache.h> 87 #ifdef INET6 88 #include <netinet6/tcp6_var.h> 89 #endif 90 #ifdef TCP_OFFLOAD 91 #include <netinet/toecore.h> 92 #endif 93 94 #ifdef IPSEC 95 #include <netipsec/ipsec.h> 96 #ifdef INET6 97 #include <netipsec/ipsec6.h> 98 #endif 99 #include <netipsec/key.h> 100 #endif /*IPSEC*/ 101 102 #include <machine/in_cksum.h> 103 104 #include <security/mac/mac_framework.h> 105 106 static VNET_DEFINE(int, tcp_syncookies) = 1; 107 #define V_tcp_syncookies VNET(tcp_syncookies) 108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 109 &VNET_NAME(tcp_syncookies), 0, 110 "Use TCP SYN cookies if the syncache overflows"); 111 112 static VNET_DEFINE(int, tcp_syncookiesonly) = 0; 113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 115 &VNET_NAME(tcp_syncookiesonly), 0, 116 "Use only TCP SYN cookies"); 117 118 #ifdef TCP_OFFLOAD 119 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 120 #endif 121 122 static void syncache_drop(struct syncache *, struct syncache_head *); 123 static void syncache_free(struct syncache *); 124 static void syncache_insert(struct syncache *, struct syncache_head *); 125 static int syncache_respond(struct syncache *, struct syncache_head *, int); 126 static struct socket *syncache_socket(struct syncache *, struct socket *, 127 struct mbuf *m); 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_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 164 &VNET_NAME(tcp_syncache.bucket_limit), 0, 165 "Per-bucket hash limit for syncache"); 166 167 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 168 &VNET_NAME(tcp_syncache.cache_limit), 0, 169 "Overall entry limit for syncache"); 170 171 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 172 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 173 174 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 175 &VNET_NAME(tcp_syncache.hashsize), 0, 176 "Size of TCP syncache hashtable"); 177 178 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW, 179 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 180 "Limit on SYN/ACK retransmissions"); 181 182 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 183 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 184 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 185 "Send reset on socket allocation failure"); 186 187 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 188 189 #define SYNCACHE_HASH(inc, mask) \ 190 ((V_tcp_syncache.hash_secret ^ \ 191 (inc)->inc_faddr.s_addr ^ \ 192 ((inc)->inc_faddr.s_addr >> 16) ^ \ 193 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 194 195 #define SYNCACHE_HASH6(inc, mask) \ 196 ((V_tcp_syncache.hash_secret ^ \ 197 (inc)->inc6_faddr.s6_addr32[0] ^ \ 198 (inc)->inc6_faddr.s6_addr32[3] ^ \ 199 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 200 201 #define ENDPTS_EQ(a, b) ( \ 202 (a)->ie_fport == (b)->ie_fport && \ 203 (a)->ie_lport == (b)->ie_lport && \ 204 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \ 205 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \ 206 ) 207 208 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0) 209 210 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 211 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 212 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 213 214 /* 215 * Requires the syncache entry to be already removed from the bucket list. 216 */ 217 static void 218 syncache_free(struct syncache *sc) 219 { 220 221 if (sc->sc_ipopts) 222 (void) m_free(sc->sc_ipopts); 223 if (sc->sc_cred) 224 crfree(sc->sc_cred); 225 #ifdef MAC 226 mac_syncache_destroy(&sc->sc_label); 227 #endif 228 229 uma_zfree(V_tcp_syncache.zone, sc); 230 } 231 232 void 233 syncache_init(void) 234 { 235 int i; 236 237 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 238 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 239 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 240 V_tcp_syncache.hash_secret = arc4random(); 241 242 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 243 &V_tcp_syncache.hashsize); 244 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 245 &V_tcp_syncache.bucket_limit); 246 if (!powerof2(V_tcp_syncache.hashsize) || 247 V_tcp_syncache.hashsize == 0) { 248 printf("WARNING: syncache hash size is not a power of 2.\n"); 249 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 250 } 251 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 252 253 /* Set limits. */ 254 V_tcp_syncache.cache_limit = 255 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 256 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 257 &V_tcp_syncache.cache_limit); 258 259 /* Allocate the hash table. */ 260 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 261 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 262 263 #ifdef VIMAGE 264 V_tcp_syncache.vnet = curvnet; 265 #endif 266 267 /* Initialize the hash buckets. */ 268 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 269 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 270 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 271 NULL, MTX_DEF); 272 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 273 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 274 V_tcp_syncache.hashbase[i].sch_length = 0; 275 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 276 } 277 278 /* Create the syncache entry zone. */ 279 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 280 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 281 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 282 V_tcp_syncache.cache_limit); 283 284 /* Start the SYN cookie reseeder callout. */ 285 callout_init(&V_tcp_syncache.secret.reseed, 1); 286 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 287 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 288 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 289 syncookie_reseed, &V_tcp_syncache); 290 } 291 292 #ifdef VIMAGE 293 void 294 syncache_destroy(void) 295 { 296 struct syncache_head *sch; 297 struct syncache *sc, *nsc; 298 int i; 299 300 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 301 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 302 303 sch = &V_tcp_syncache.hashbase[i]; 304 callout_drain(&sch->sch_timer); 305 306 SCH_LOCK(sch); 307 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 308 syncache_drop(sc, sch); 309 SCH_UNLOCK(sch); 310 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 311 ("%s: sch->sch_bucket not empty", __func__)); 312 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 313 __func__, sch->sch_length)); 314 mtx_destroy(&sch->sch_mtx); 315 } 316 317 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 318 ("%s: cache_count not 0", __func__)); 319 320 /* Free the allocated global resources. */ 321 uma_zdestroy(V_tcp_syncache.zone); 322 free(V_tcp_syncache.hashbase, M_SYNCACHE); 323 324 callout_drain(&V_tcp_syncache.secret.reseed); 325 } 326 #endif 327 328 /* 329 * Inserts a syncache entry into the specified bucket row. 330 * Locks and unlocks the syncache_head autonomously. 331 */ 332 static void 333 syncache_insert(struct syncache *sc, struct syncache_head *sch) 334 { 335 struct syncache *sc2; 336 337 SCH_LOCK(sch); 338 339 /* 340 * Make sure that we don't overflow the per-bucket limit. 341 * If the bucket is full, toss the oldest element. 342 */ 343 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 344 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 345 ("sch->sch_length incorrect")); 346 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 347 syncache_drop(sc2, sch); 348 TCPSTAT_INC(tcps_sc_bucketoverflow); 349 } 350 351 /* Put it into the bucket. */ 352 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 353 sch->sch_length++; 354 355 #ifdef TCP_OFFLOAD 356 if (ADDED_BY_TOE(sc)) { 357 struct toedev *tod = sc->sc_tod; 358 359 tod->tod_syncache_added(tod, sc->sc_todctx); 360 } 361 #endif 362 363 /* Reinitialize the bucket row's timer. */ 364 if (sch->sch_length == 1) 365 sch->sch_nextc = ticks + INT_MAX; 366 syncache_timeout(sc, sch, 1); 367 368 SCH_UNLOCK(sch); 369 370 TCPSTAT_INC(tcps_sc_added); 371 } 372 373 /* 374 * Remove and free entry from syncache bucket row. 375 * Expects locked syncache head. 376 */ 377 static void 378 syncache_drop(struct syncache *sc, struct syncache_head *sch) 379 { 380 381 SCH_LOCK_ASSERT(sch); 382 383 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 384 sch->sch_length--; 385 386 #ifdef TCP_OFFLOAD 387 if (ADDED_BY_TOE(sc)) { 388 struct toedev *tod = sc->sc_tod; 389 390 tod->tod_syncache_removed(tod, sc->sc_todctx); 391 } 392 #endif 393 394 syncache_free(sc); 395 } 396 397 /* 398 * Engage/reengage time on bucket row. 399 */ 400 static void 401 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 402 { 403 sc->sc_rxttime = ticks + 404 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]); 405 sc->sc_rxmits++; 406 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 407 sch->sch_nextc = sc->sc_rxttime; 408 if (docallout) 409 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 410 syncache_timer, (void *)sch); 411 } 412 } 413 414 /* 415 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 416 * If we have retransmitted an entry the maximum number of times, expire it. 417 * One separate timer for each bucket row. 418 */ 419 static void 420 syncache_timer(void *xsch) 421 { 422 struct syncache_head *sch = (struct syncache_head *)xsch; 423 struct syncache *sc, *nsc; 424 int tick = ticks; 425 char *s; 426 427 CURVNET_SET(sch->sch_sc->vnet); 428 429 /* NB: syncache_head has already been locked by the callout. */ 430 SCH_LOCK_ASSERT(sch); 431 432 /* 433 * In the following cycle we may remove some entries and/or 434 * advance some timeouts, so re-initialize the bucket timer. 435 */ 436 sch->sch_nextc = tick + INT_MAX; 437 438 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 439 /* 440 * We do not check if the listen socket still exists 441 * and accept the case where the listen socket may be 442 * gone by the time we resend the SYN/ACK. We do 443 * not expect this to happens often. If it does, 444 * then the RST will be sent by the time the remote 445 * host does the SYN/ACK->ACK. 446 */ 447 if (TSTMP_GT(sc->sc_rxttime, tick)) { 448 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 449 sch->sch_nextc = sc->sc_rxttime; 450 continue; 451 } 452 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 453 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 454 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 455 "giving up and removing syncache entry\n", 456 s, __func__); 457 free(s, M_TCPLOG); 458 } 459 syncache_drop(sc, sch); 460 TCPSTAT_INC(tcps_sc_stale); 461 continue; 462 } 463 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 464 log(LOG_DEBUG, "%s; %s: Response timeout, " 465 "retransmitting (%u) SYN|ACK\n", 466 s, __func__, sc->sc_rxmits); 467 free(s, M_TCPLOG); 468 } 469 470 syncache_respond(sc, sch, 1); 471 TCPSTAT_INC(tcps_sc_retransmitted); 472 syncache_timeout(sc, sch, 0); 473 } 474 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 475 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 476 syncache_timer, (void *)(sch)); 477 CURVNET_RESTORE(); 478 } 479 480 /* 481 * Find an entry in the syncache. 482 * Returns always with locked syncache_head plus a matching entry or NULL. 483 */ 484 static struct syncache * 485 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 486 { 487 struct syncache *sc; 488 struct syncache_head *sch; 489 490 #ifdef INET6 491 if (inc->inc_flags & INC_ISIPV6) { 492 sch = &V_tcp_syncache.hashbase[ 493 SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)]; 494 *schp = sch; 495 496 SCH_LOCK(sch); 497 498 /* Circle through bucket row to find matching entry. */ 499 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 500 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 501 return (sc); 502 } 503 } else 504 #endif 505 { 506 sch = &V_tcp_syncache.hashbase[ 507 SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)]; 508 *schp = sch; 509 510 SCH_LOCK(sch); 511 512 /* Circle through bucket row to find matching entry. */ 513 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 514 #ifdef INET6 515 if (sc->sc_inc.inc_flags & INC_ISIPV6) 516 continue; 517 #endif 518 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 519 return (sc); 520 } 521 } 522 SCH_LOCK_ASSERT(*schp); 523 return (NULL); /* always returns with locked sch */ 524 } 525 526 /* 527 * This function is called when we get a RST for a 528 * non-existent connection, so that we can see if the 529 * connection is in the syn cache. If it is, zap it. 530 */ 531 void 532 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 533 { 534 struct syncache *sc; 535 struct syncache_head *sch; 536 char *s = NULL; 537 538 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 539 SCH_LOCK_ASSERT(sch); 540 541 /* 542 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 543 * See RFC 793 page 65, section SEGMENT ARRIVES. 544 */ 545 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 546 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 547 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 548 "FIN flag set, segment ignored\n", s, __func__); 549 TCPSTAT_INC(tcps_badrst); 550 goto done; 551 } 552 553 /* 554 * No corresponding connection was found in syncache. 555 * If syncookies are enabled and possibly exclusively 556 * used, or we are under memory pressure, a valid RST 557 * may not find a syncache entry. In that case we're 558 * done and no SYN|ACK retransmissions will happen. 559 * Otherwise the RST was misdirected or spoofed. 560 */ 561 if (sc == NULL) { 562 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 563 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 564 "syncache entry (possibly syncookie only), " 565 "segment ignored\n", s, __func__); 566 TCPSTAT_INC(tcps_badrst); 567 goto done; 568 } 569 570 /* 571 * If the RST bit is set, check the sequence number to see 572 * if this is a valid reset segment. 573 * RFC 793 page 37: 574 * In all states except SYN-SENT, all reset (RST) segments 575 * are validated by checking their SEQ-fields. A reset is 576 * valid if its sequence number is in the window. 577 * 578 * The sequence number in the reset segment is normally an 579 * echo of our outgoing acknowlegement numbers, but some hosts 580 * send a reset with the sequence number at the rightmost edge 581 * of our receive window, and we have to handle this case. 582 */ 583 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 584 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 585 syncache_drop(sc, sch); 586 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 587 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 588 "connection attempt aborted by remote endpoint\n", 589 s, __func__); 590 TCPSTAT_INC(tcps_sc_reset); 591 } else { 592 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 593 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 594 "IRS %u (+WND %u), segment ignored\n", 595 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 596 TCPSTAT_INC(tcps_badrst); 597 } 598 599 done: 600 if (s != NULL) 601 free(s, M_TCPLOG); 602 SCH_UNLOCK(sch); 603 } 604 605 void 606 syncache_badack(struct in_conninfo *inc) 607 { 608 struct syncache *sc; 609 struct syncache_head *sch; 610 611 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 612 SCH_LOCK_ASSERT(sch); 613 if (sc != NULL) { 614 syncache_drop(sc, sch); 615 TCPSTAT_INC(tcps_sc_badack); 616 } 617 SCH_UNLOCK(sch); 618 } 619 620 void 621 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th) 622 { 623 struct syncache *sc; 624 struct syncache_head *sch; 625 626 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 627 SCH_LOCK_ASSERT(sch); 628 if (sc == NULL) 629 goto done; 630 631 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 632 if (ntohl(th->th_seq) != sc->sc_iss) 633 goto done; 634 635 /* 636 * If we've rertransmitted 3 times and this is our second error, 637 * we remove the entry. Otherwise, we allow it to continue on. 638 * This prevents us from incorrectly nuking an entry during a 639 * spurious network outage. 640 * 641 * See tcp_notify(). 642 */ 643 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 644 sc->sc_flags |= SCF_UNREACH; 645 goto done; 646 } 647 syncache_drop(sc, sch); 648 TCPSTAT_INC(tcps_sc_unreach); 649 done: 650 SCH_UNLOCK(sch); 651 } 652 653 /* 654 * Build a new TCP socket structure from a syncache entry. 655 */ 656 static struct socket * 657 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 658 { 659 struct inpcb *inp = NULL; 660 struct socket *so; 661 struct tcpcb *tp; 662 int error; 663 char *s; 664 665 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 666 667 /* 668 * Ok, create the full blown connection, and set things up 669 * as they would have been set up if we had created the 670 * connection when the SYN arrived. If we can't create 671 * the connection, abort it. 672 */ 673 so = sonewconn(lso, 0); 674 if (so == NULL) { 675 /* 676 * Drop the connection; we will either send a RST or 677 * have the peer retransmit its SYN again after its 678 * RTO and try again. 679 */ 680 TCPSTAT_INC(tcps_listendrop); 681 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 682 log(LOG_DEBUG, "%s; %s: Socket create failed " 683 "due to limits or memory shortage\n", 684 s, __func__); 685 free(s, M_TCPLOG); 686 } 687 goto abort2; 688 } 689 #ifdef MAC 690 mac_socketpeer_set_from_mbuf(m, so); 691 #endif 692 693 inp = sotoinpcb(so); 694 inp->inp_inc.inc_fibnum = so->so_fibnum; 695 INP_WLOCK(inp); 696 INP_HASH_WLOCK(&V_tcbinfo); 697 698 /* Insert new socket into PCB hash list. */ 699 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 700 #ifdef INET6 701 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 702 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 703 } else { 704 inp->inp_vflag &= ~INP_IPV6; 705 inp->inp_vflag |= INP_IPV4; 706 #endif 707 inp->inp_laddr = sc->sc_inc.inc_laddr; 708 #ifdef INET6 709 } 710 #endif 711 712 /* 713 * If there's an mbuf and it has a flowid, then let's initialise the 714 * inp with that particular flowid. 715 */ 716 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 717 inp->inp_flowid = m->m_pkthdr.flowid; 718 inp->inp_flowtype = M_HASHTYPE_GET(m); 719 } 720 721 /* 722 * Install in the reservation hash table for now, but don't yet 723 * install a connection group since the full 4-tuple isn't yet 724 * configured. 725 */ 726 inp->inp_lport = sc->sc_inc.inc_lport; 727 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 728 /* 729 * Undo the assignments above if we failed to 730 * put the PCB on the hash lists. 731 */ 732 #ifdef INET6 733 if (sc->sc_inc.inc_flags & INC_ISIPV6) 734 inp->in6p_laddr = in6addr_any; 735 else 736 #endif 737 inp->inp_laddr.s_addr = INADDR_ANY; 738 inp->inp_lport = 0; 739 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 740 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 741 "with error %i\n", 742 s, __func__, error); 743 free(s, M_TCPLOG); 744 } 745 INP_HASH_WUNLOCK(&V_tcbinfo); 746 goto abort; 747 } 748 #ifdef IPSEC 749 /* Copy old policy into new socket's. */ 750 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp)) 751 printf("syncache_socket: could not copy policy\n"); 752 #endif 753 #ifdef INET6 754 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 755 struct inpcb *oinp = sotoinpcb(lso); 756 struct in6_addr laddr6; 757 struct sockaddr_in6 sin6; 758 /* 759 * Inherit socket options from the listening socket. 760 * Note that in6p_inputopts are not (and should not be) 761 * copied, since it stores previously received options and is 762 * used to detect if each new option is different than the 763 * previous one and hence should be passed to a user. 764 * If we copied in6p_inputopts, a user would not be able to 765 * receive options just after calling the accept system call. 766 */ 767 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 768 if (oinp->in6p_outputopts) 769 inp->in6p_outputopts = 770 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 771 772 sin6.sin6_family = AF_INET6; 773 sin6.sin6_len = sizeof(sin6); 774 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 775 sin6.sin6_port = sc->sc_inc.inc_fport; 776 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 777 laddr6 = inp->in6p_laddr; 778 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 779 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 780 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 781 thread0.td_ucred, m)) != 0) { 782 inp->in6p_laddr = laddr6; 783 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 784 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 785 "with error %i\n", 786 s, __func__, error); 787 free(s, M_TCPLOG); 788 } 789 INP_HASH_WUNLOCK(&V_tcbinfo); 790 goto abort; 791 } 792 /* Override flowlabel from in6_pcbconnect. */ 793 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 794 inp->inp_flow |= sc->sc_flowlabel; 795 } 796 #endif /* INET6 */ 797 #if defined(INET) && defined(INET6) 798 else 799 #endif 800 #ifdef INET 801 { 802 struct in_addr laddr; 803 struct sockaddr_in sin; 804 805 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 806 807 if (inp->inp_options == NULL) { 808 inp->inp_options = sc->sc_ipopts; 809 sc->sc_ipopts = NULL; 810 } 811 812 sin.sin_family = AF_INET; 813 sin.sin_len = sizeof(sin); 814 sin.sin_addr = sc->sc_inc.inc_faddr; 815 sin.sin_port = sc->sc_inc.inc_fport; 816 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 817 laddr = inp->inp_laddr; 818 if (inp->inp_laddr.s_addr == INADDR_ANY) 819 inp->inp_laddr = sc->sc_inc.inc_laddr; 820 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 821 thread0.td_ucred, m)) != 0) { 822 inp->inp_laddr = laddr; 823 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 824 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 825 "with error %i\n", 826 s, __func__, error); 827 free(s, M_TCPLOG); 828 } 829 INP_HASH_WUNLOCK(&V_tcbinfo); 830 goto abort; 831 } 832 } 833 #endif /* INET */ 834 INP_HASH_WUNLOCK(&V_tcbinfo); 835 tp = intotcpcb(inp); 836 tcp_state_change(tp, TCPS_SYN_RECEIVED); 837 tp->iss = sc->sc_iss; 838 tp->irs = sc->sc_irs; 839 tcp_rcvseqinit(tp); 840 tcp_sendseqinit(tp); 841 tp->snd_wl1 = sc->sc_irs; 842 tp->snd_max = tp->iss + 1; 843 tp->snd_nxt = tp->iss + 1; 844 tp->rcv_up = sc->sc_irs + 1; 845 tp->rcv_wnd = sc->sc_wnd; 846 tp->rcv_adv += tp->rcv_wnd; 847 tp->last_ack_sent = tp->rcv_nxt; 848 849 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 850 if (sc->sc_flags & SCF_NOOPT) 851 tp->t_flags |= TF_NOOPT; 852 else { 853 if (sc->sc_flags & SCF_WINSCALE) { 854 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 855 tp->snd_scale = sc->sc_requested_s_scale; 856 tp->request_r_scale = sc->sc_requested_r_scale; 857 } 858 if (sc->sc_flags & SCF_TIMESTAMP) { 859 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 860 tp->ts_recent = sc->sc_tsreflect; 861 tp->ts_recent_age = tcp_ts_getticks(); 862 tp->ts_offset = sc->sc_tsoff; 863 } 864 #ifdef TCP_SIGNATURE 865 if (sc->sc_flags & SCF_SIGNATURE) 866 tp->t_flags |= TF_SIGNATURE; 867 #endif 868 if (sc->sc_flags & SCF_SACK) 869 tp->t_flags |= TF_SACK_PERMIT; 870 } 871 872 if (sc->sc_flags & SCF_ECN) 873 tp->t_flags |= TF_ECN_PERMIT; 874 875 /* 876 * Set up MSS and get cached values from tcp_hostcache. 877 * This might overwrite some of the defaults we just set. 878 */ 879 tcp_mss(tp, sc->sc_peer_mss); 880 881 /* 882 * If the SYN,ACK was retransmitted, indicate that CWND to be 883 * limited to one segment in cc_conn_init(). 884 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 885 */ 886 if (sc->sc_rxmits > 1) 887 tp->snd_cwnd = 1; 888 889 #ifdef TCP_OFFLOAD 890 /* 891 * Allow a TOE driver to install its hooks. Note that we hold the 892 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 893 * new connection before the TOE driver has done its thing. 894 */ 895 if (ADDED_BY_TOE(sc)) { 896 struct toedev *tod = sc->sc_tod; 897 898 tod->tod_offload_socket(tod, sc->sc_todctx, so); 899 } 900 #endif 901 /* 902 * Copy and activate timers. 903 */ 904 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 905 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 906 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 907 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 908 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 909 910 INP_WUNLOCK(inp); 911 912 soisconnected(so); 913 914 TCPSTAT_INC(tcps_accepts); 915 return (so); 916 917 abort: 918 INP_WUNLOCK(inp); 919 abort2: 920 if (so != NULL) 921 soabort(so); 922 return (NULL); 923 } 924 925 /* 926 * This function gets called when we receive an ACK for a 927 * socket in the LISTEN state. We look up the connection 928 * in the syncache, and if its there, we pull it out of 929 * the cache and turn it into a full-blown connection in 930 * the SYN-RECEIVED state. 931 */ 932 int 933 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 934 struct socket **lsop, struct mbuf *m) 935 { 936 struct syncache *sc; 937 struct syncache_head *sch; 938 struct syncache scs; 939 char *s; 940 941 /* 942 * Global TCP locks are held because we manipulate the PCB lists 943 * and create a new socket. 944 */ 945 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 946 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 947 ("%s: can handle only ACK", __func__)); 948 949 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 950 SCH_LOCK_ASSERT(sch); 951 952 #ifdef INVARIANTS 953 /* 954 * Test code for syncookies comparing the syncache stored 955 * values with the reconstructed values from the cookie. 956 */ 957 if (sc != NULL) 958 syncookie_cmp(inc, sch, sc, th, to, *lsop); 959 #endif 960 961 if (sc == NULL) { 962 /* 963 * There is no syncache entry, so see if this ACK is 964 * a returning syncookie. To do this, first: 965 * A. See if this socket has had a syncache entry dropped in 966 * the past. We don't want to accept a bogus syncookie 967 * if we've never received a SYN. 968 * B. check that the syncookie is valid. If it is, then 969 * cobble up a fake syncache entry, and return. 970 */ 971 if (!V_tcp_syncookies) { 972 SCH_UNLOCK(sch); 973 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 974 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 975 "segment rejected (syncookies disabled)\n", 976 s, __func__); 977 goto failed; 978 } 979 bzero(&scs, sizeof(scs)); 980 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 981 SCH_UNLOCK(sch); 982 if (sc == NULL) { 983 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 984 log(LOG_DEBUG, "%s; %s: Segment failed " 985 "SYNCOOKIE authentication, segment rejected " 986 "(probably spoofed)\n", s, __func__); 987 goto failed; 988 } 989 } else { 990 /* Pull out the entry to unlock the bucket row. */ 991 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 992 sch->sch_length--; 993 #ifdef TCP_OFFLOAD 994 if (ADDED_BY_TOE(sc)) { 995 struct toedev *tod = sc->sc_tod; 996 997 tod->tod_syncache_removed(tod, sc->sc_todctx); 998 } 999 #endif 1000 SCH_UNLOCK(sch); 1001 } 1002 1003 /* 1004 * Segment validation: 1005 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1006 */ 1007 if (th->th_ack != sc->sc_iss + 1) { 1008 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1009 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1010 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1011 goto failed; 1012 } 1013 1014 /* 1015 * The SEQ must fall in the window starting at the received 1016 * initial receive sequence number + 1 (the SYN). 1017 */ 1018 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1019 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1020 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1021 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1022 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1023 goto failed; 1024 } 1025 1026 /* 1027 * If timestamps were not negotiated during SYN/ACK they 1028 * must not appear on any segment during this session. 1029 */ 1030 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1031 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1032 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1033 "segment rejected\n", s, __func__); 1034 goto failed; 1035 } 1036 1037 /* 1038 * If timestamps were negotiated during SYN/ACK they should 1039 * appear on every segment during this session. 1040 * XXXAO: This is only informal as there have been unverified 1041 * reports of non-compliants stacks. 1042 */ 1043 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1044 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1045 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1046 "no action\n", s, __func__); 1047 free(s, M_TCPLOG); 1048 s = NULL; 1049 } 1050 } 1051 1052 /* 1053 * If timestamps were negotiated the reflected timestamp 1054 * must be equal to what we actually sent in the SYN|ACK. 1055 */ 1056 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) { 1057 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1058 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1059 "segment rejected\n", 1060 s, __func__, to->to_tsecr, sc->sc_ts); 1061 goto failed; 1062 } 1063 1064 *lsop = syncache_socket(sc, *lsop, m); 1065 1066 if (*lsop == NULL) 1067 TCPSTAT_INC(tcps_sc_aborted); 1068 else 1069 TCPSTAT_INC(tcps_sc_completed); 1070 1071 /* how do we find the inp for the new socket? */ 1072 if (sc != &scs) 1073 syncache_free(sc); 1074 return (1); 1075 failed: 1076 if (sc != NULL && sc != &scs) 1077 syncache_free(sc); 1078 if (s != NULL) 1079 free(s, M_TCPLOG); 1080 *lsop = NULL; 1081 return (0); 1082 } 1083 1084 /* 1085 * Given a LISTEN socket and an inbound SYN request, add 1086 * this to the syn cache, and send back a segment: 1087 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1088 * to the source. 1089 * 1090 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1091 * Doing so would require that we hold onto the data and deliver it 1092 * to the application. However, if we are the target of a SYN-flood 1093 * DoS attack, an attacker could send data which would eventually 1094 * consume all available buffer space if it were ACKed. By not ACKing 1095 * the data, we avoid this DoS scenario. 1096 */ 1097 void 1098 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1099 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1100 void *todctx) 1101 { 1102 struct tcpcb *tp; 1103 struct socket *so; 1104 struct syncache *sc = NULL; 1105 struct syncache_head *sch; 1106 struct mbuf *ipopts = NULL; 1107 u_int ltflags; 1108 int win, sb_hiwat, ip_ttl, ip_tos; 1109 char *s; 1110 #ifdef INET6 1111 int autoflowlabel = 0; 1112 #endif 1113 #ifdef MAC 1114 struct label *maclabel; 1115 #endif 1116 struct syncache scs; 1117 struct ucred *cred; 1118 1119 INP_WLOCK_ASSERT(inp); /* listen socket */ 1120 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1121 ("%s: unexpected tcp flags", __func__)); 1122 1123 /* 1124 * Combine all so/tp operations very early to drop the INP lock as 1125 * soon as possible. 1126 */ 1127 so = *lsop; 1128 tp = sototcpcb(so); 1129 cred = crhold(so->so_cred); 1130 1131 #ifdef INET6 1132 if ((inc->inc_flags & INC_ISIPV6) && 1133 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1134 autoflowlabel = 1; 1135 #endif 1136 ip_ttl = inp->inp_ip_ttl; 1137 ip_tos = inp->inp_ip_tos; 1138 win = sbspace(&so->so_rcv); 1139 sb_hiwat = so->so_rcv.sb_hiwat; 1140 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1141 1142 /* By the time we drop the lock these should no longer be used. */ 1143 so = NULL; 1144 tp = NULL; 1145 1146 #ifdef MAC 1147 if (mac_syncache_init(&maclabel) != 0) { 1148 INP_WUNLOCK(inp); 1149 goto done; 1150 } else 1151 mac_syncache_create(maclabel, inp); 1152 #endif 1153 INP_WUNLOCK(inp); 1154 1155 /* 1156 * Remember the IP options, if any. 1157 */ 1158 #ifdef INET6 1159 if (!(inc->inc_flags & INC_ISIPV6)) 1160 #endif 1161 #ifdef INET 1162 ipopts = (m) ? ip_srcroute(m) : NULL; 1163 #else 1164 ipopts = NULL; 1165 #endif 1166 1167 /* 1168 * See if we already have an entry for this connection. 1169 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1170 * 1171 * XXX: should the syncache be re-initialized with the contents 1172 * of the new SYN here (which may have different options?) 1173 * 1174 * XXX: We do not check the sequence number to see if this is a 1175 * real retransmit or a new connection attempt. The question is 1176 * how to handle such a case; either ignore it as spoofed, or 1177 * drop the current entry and create a new one? 1178 */ 1179 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1180 SCH_LOCK_ASSERT(sch); 1181 if (sc != NULL) { 1182 TCPSTAT_INC(tcps_sc_dupsyn); 1183 if (ipopts) { 1184 /* 1185 * If we were remembering a previous source route, 1186 * forget it and use the new one we've been given. 1187 */ 1188 if (sc->sc_ipopts) 1189 (void) m_free(sc->sc_ipopts); 1190 sc->sc_ipopts = ipopts; 1191 } 1192 /* 1193 * Update timestamp if present. 1194 */ 1195 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1196 sc->sc_tsreflect = to->to_tsval; 1197 else 1198 sc->sc_flags &= ~SCF_TIMESTAMP; 1199 #ifdef MAC 1200 /* 1201 * Since we have already unconditionally allocated label 1202 * storage, free it up. The syncache entry will already 1203 * have an initialized label we can use. 1204 */ 1205 mac_syncache_destroy(&maclabel); 1206 #endif 1207 /* Retransmit SYN|ACK and reset retransmit count. */ 1208 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1209 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1210 "resetting timer and retransmitting SYN|ACK\n", 1211 s, __func__); 1212 free(s, M_TCPLOG); 1213 } 1214 if (syncache_respond(sc, sch, 1) == 0) { 1215 sc->sc_rxmits = 0; 1216 syncache_timeout(sc, sch, 1); 1217 TCPSTAT_INC(tcps_sndacks); 1218 TCPSTAT_INC(tcps_sndtotal); 1219 } 1220 SCH_UNLOCK(sch); 1221 goto done; 1222 } 1223 1224 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1225 if (sc == NULL) { 1226 /* 1227 * The zone allocator couldn't provide more entries. 1228 * Treat this as if the cache was full; drop the oldest 1229 * entry and insert the new one. 1230 */ 1231 TCPSTAT_INC(tcps_sc_zonefail); 1232 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) 1233 syncache_drop(sc, sch); 1234 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1235 if (sc == NULL) { 1236 if (V_tcp_syncookies) { 1237 bzero(&scs, sizeof(scs)); 1238 sc = &scs; 1239 } else { 1240 SCH_UNLOCK(sch); 1241 if (ipopts) 1242 (void) m_free(ipopts); 1243 goto done; 1244 } 1245 } 1246 } 1247 1248 /* 1249 * Fill in the syncache values. 1250 */ 1251 #ifdef MAC 1252 sc->sc_label = maclabel; 1253 #endif 1254 sc->sc_cred = cred; 1255 cred = NULL; 1256 sc->sc_ipopts = ipopts; 1257 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1258 #ifdef INET6 1259 if (!(inc->inc_flags & INC_ISIPV6)) 1260 #endif 1261 { 1262 sc->sc_ip_tos = ip_tos; 1263 sc->sc_ip_ttl = ip_ttl; 1264 } 1265 #ifdef TCP_OFFLOAD 1266 sc->sc_tod = tod; 1267 sc->sc_todctx = todctx; 1268 #endif 1269 sc->sc_irs = th->th_seq; 1270 sc->sc_iss = arc4random(); 1271 sc->sc_flags = 0; 1272 sc->sc_flowlabel = 0; 1273 1274 /* 1275 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1276 * win was derived from socket earlier in the function. 1277 */ 1278 win = imax(win, 0); 1279 win = imin(win, TCP_MAXWIN); 1280 sc->sc_wnd = win; 1281 1282 if (V_tcp_do_rfc1323) { 1283 /* 1284 * A timestamp received in a SYN makes 1285 * it ok to send timestamp requests and replies. 1286 */ 1287 if (to->to_flags & TOF_TS) { 1288 sc->sc_tsreflect = to->to_tsval; 1289 sc->sc_ts = tcp_ts_getticks(); 1290 sc->sc_flags |= SCF_TIMESTAMP; 1291 } 1292 if (to->to_flags & TOF_SCALE) { 1293 int wscale = 0; 1294 1295 /* 1296 * Pick the smallest possible scaling factor that 1297 * will still allow us to scale up to sb_max, aka 1298 * kern.ipc.maxsockbuf. 1299 * 1300 * We do this because there are broken firewalls that 1301 * will corrupt the window scale option, leading to 1302 * the other endpoint believing that our advertised 1303 * window is unscaled. At scale factors larger than 1304 * 5 the unscaled window will drop below 1500 bytes, 1305 * leading to serious problems when traversing these 1306 * broken firewalls. 1307 * 1308 * With the default maxsockbuf of 256K, a scale factor 1309 * of 3 will be chosen by this algorithm. Those who 1310 * choose a larger maxsockbuf should watch out 1311 * for the compatiblity problems mentioned above. 1312 * 1313 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1314 * or <SYN,ACK>) segment itself is never scaled. 1315 */ 1316 while (wscale < TCP_MAX_WINSHIFT && 1317 (TCP_MAXWIN << wscale) < sb_max) 1318 wscale++; 1319 sc->sc_requested_r_scale = wscale; 1320 sc->sc_requested_s_scale = to->to_wscale; 1321 sc->sc_flags |= SCF_WINSCALE; 1322 } 1323 } 1324 #ifdef TCP_SIGNATURE 1325 /* 1326 * If listening socket requested TCP digests, OR received SYN 1327 * contains the option, flag this in the syncache so that 1328 * syncache_respond() will do the right thing with the SYN+ACK. 1329 */ 1330 if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE) 1331 sc->sc_flags |= SCF_SIGNATURE; 1332 #endif 1333 if (to->to_flags & TOF_SACKPERM) 1334 sc->sc_flags |= SCF_SACK; 1335 if (to->to_flags & TOF_MSS) 1336 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1337 if (ltflags & TF_NOOPT) 1338 sc->sc_flags |= SCF_NOOPT; 1339 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1340 sc->sc_flags |= SCF_ECN; 1341 1342 if (V_tcp_syncookies) 1343 sc->sc_iss = syncookie_generate(sch, sc); 1344 #ifdef INET6 1345 if (autoflowlabel) { 1346 if (V_tcp_syncookies) 1347 sc->sc_flowlabel = sc->sc_iss; 1348 else 1349 sc->sc_flowlabel = ip6_randomflowlabel(); 1350 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1351 } 1352 #endif 1353 SCH_UNLOCK(sch); 1354 1355 /* 1356 * Do a standard 3-way handshake. 1357 */ 1358 if (syncache_respond(sc, sch, 0) == 0) { 1359 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1360 syncache_free(sc); 1361 else if (sc != &scs) 1362 syncache_insert(sc, sch); /* locks and unlocks sch */ 1363 TCPSTAT_INC(tcps_sndacks); 1364 TCPSTAT_INC(tcps_sndtotal); 1365 } else { 1366 if (sc != &scs) 1367 syncache_free(sc); 1368 TCPSTAT_INC(tcps_sc_dropped); 1369 } 1370 1371 done: 1372 if (cred != NULL) 1373 crfree(cred); 1374 #ifdef MAC 1375 if (sc == &scs) 1376 mac_syncache_destroy(&maclabel); 1377 #endif 1378 if (m) { 1379 1380 *lsop = NULL; 1381 m_freem(m); 1382 } 1383 } 1384 1385 static int 1386 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked) 1387 { 1388 struct ip *ip = NULL; 1389 struct mbuf *m; 1390 struct tcphdr *th = NULL; 1391 int optlen, error = 0; /* Make compiler happy */ 1392 u_int16_t hlen, tlen, mssopt; 1393 struct tcpopt to; 1394 #ifdef INET6 1395 struct ip6_hdr *ip6 = NULL; 1396 #endif 1397 #ifdef TCP_SIGNATURE 1398 struct secasvar *sav; 1399 #endif 1400 1401 hlen = 1402 #ifdef INET6 1403 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1404 #endif 1405 sizeof(struct ip); 1406 tlen = hlen + sizeof(struct tcphdr); 1407 1408 /* Determine MSS we advertize to other end of connection. */ 1409 mssopt = tcp_mssopt(&sc->sc_inc); 1410 if (sc->sc_peer_mss) 1411 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss); 1412 1413 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1414 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1415 ("syncache: mbuf too small")); 1416 1417 /* Create the IP+TCP header from scratch. */ 1418 m = m_gethdr(M_NOWAIT, MT_DATA); 1419 if (m == NULL) 1420 return (ENOBUFS); 1421 #ifdef MAC 1422 mac_syncache_create_mbuf(sc->sc_label, m); 1423 #endif 1424 m->m_data += max_linkhdr; 1425 m->m_len = tlen; 1426 m->m_pkthdr.len = tlen; 1427 m->m_pkthdr.rcvif = NULL; 1428 1429 #ifdef INET6 1430 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1431 ip6 = mtod(m, struct ip6_hdr *); 1432 ip6->ip6_vfc = IPV6_VERSION; 1433 ip6->ip6_nxt = IPPROTO_TCP; 1434 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1435 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1436 ip6->ip6_plen = htons(tlen - hlen); 1437 /* ip6_hlim is set after checksum */ 1438 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1439 ip6->ip6_flow |= sc->sc_flowlabel; 1440 1441 th = (struct tcphdr *)(ip6 + 1); 1442 } 1443 #endif 1444 #if defined(INET6) && defined(INET) 1445 else 1446 #endif 1447 #ifdef INET 1448 { 1449 ip = mtod(m, struct ip *); 1450 ip->ip_v = IPVERSION; 1451 ip->ip_hl = sizeof(struct ip) >> 2; 1452 ip->ip_len = htons(tlen); 1453 ip->ip_id = 0; 1454 ip->ip_off = 0; 1455 ip->ip_sum = 0; 1456 ip->ip_p = IPPROTO_TCP; 1457 ip->ip_src = sc->sc_inc.inc_laddr; 1458 ip->ip_dst = sc->sc_inc.inc_faddr; 1459 ip->ip_ttl = sc->sc_ip_ttl; 1460 ip->ip_tos = sc->sc_ip_tos; 1461 1462 /* 1463 * See if we should do MTU discovery. Route lookups are 1464 * expensive, so we will only unset the DF bit if: 1465 * 1466 * 1) path_mtu_discovery is disabled 1467 * 2) the SCF_UNREACH flag has been set 1468 */ 1469 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1470 ip->ip_off |= htons(IP_DF); 1471 1472 th = (struct tcphdr *)(ip + 1); 1473 } 1474 #endif /* INET */ 1475 th->th_sport = sc->sc_inc.inc_lport; 1476 th->th_dport = sc->sc_inc.inc_fport; 1477 1478 th->th_seq = htonl(sc->sc_iss); 1479 th->th_ack = htonl(sc->sc_irs + 1); 1480 th->th_off = sizeof(struct tcphdr) >> 2; 1481 th->th_x2 = 0; 1482 th->th_flags = TH_SYN|TH_ACK; 1483 th->th_win = htons(sc->sc_wnd); 1484 th->th_urp = 0; 1485 1486 if (sc->sc_flags & SCF_ECN) { 1487 th->th_flags |= TH_ECE; 1488 TCPSTAT_INC(tcps_ecn_shs); 1489 } 1490 1491 /* Tack on the TCP options. */ 1492 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1493 to.to_flags = 0; 1494 1495 to.to_mss = mssopt; 1496 to.to_flags = TOF_MSS; 1497 if (sc->sc_flags & SCF_WINSCALE) { 1498 to.to_wscale = sc->sc_requested_r_scale; 1499 to.to_flags |= TOF_SCALE; 1500 } 1501 if (sc->sc_flags & SCF_TIMESTAMP) { 1502 /* Virgin timestamp or TCP cookie enhanced one. */ 1503 to.to_tsval = sc->sc_ts; 1504 to.to_tsecr = sc->sc_tsreflect; 1505 to.to_flags |= TOF_TS; 1506 } 1507 if (sc->sc_flags & SCF_SACK) 1508 to.to_flags |= TOF_SACKPERM; 1509 #ifdef TCP_SIGNATURE 1510 sav = NULL; 1511 if (sc->sc_flags & SCF_SIGNATURE) { 1512 sav = tcp_get_sav(m, IPSEC_DIR_OUTBOUND); 1513 if (sav != NULL) 1514 to.to_flags |= TOF_SIGNATURE; 1515 else { 1516 1517 /* 1518 * We've got SCF_SIGNATURE flag 1519 * inherited from listening socket, 1520 * but no SADB key for given source 1521 * address. Assume signature is not 1522 * required and remove signature flag 1523 * instead of silently dropping 1524 * connection. 1525 */ 1526 if (locked == 0) 1527 SCH_LOCK(sch); 1528 sc->sc_flags &= ~SCF_SIGNATURE; 1529 if (locked == 0) 1530 SCH_UNLOCK(sch); 1531 } 1532 } 1533 #endif 1534 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1535 1536 /* Adjust headers by option size. */ 1537 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1538 m->m_len += optlen; 1539 m->m_pkthdr.len += optlen; 1540 1541 #ifdef TCP_SIGNATURE 1542 if (sc->sc_flags & SCF_SIGNATURE) 1543 tcp_signature_do_compute(m, 0, optlen, 1544 to.to_signature, sav); 1545 #endif 1546 #ifdef INET6 1547 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1548 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1549 else 1550 #endif 1551 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1552 } else 1553 optlen = 0; 1554 1555 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1556 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1557 #ifdef INET6 1558 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1559 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1560 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1561 IPPROTO_TCP, 0); 1562 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1563 #ifdef TCP_OFFLOAD 1564 if (ADDED_BY_TOE(sc)) { 1565 struct toedev *tod = sc->sc_tod; 1566 1567 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1568 1569 return (error); 1570 } 1571 #endif 1572 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1573 } 1574 #endif 1575 #if defined(INET6) && defined(INET) 1576 else 1577 #endif 1578 #ifdef INET 1579 { 1580 m->m_pkthdr.csum_flags = CSUM_TCP; 1581 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1582 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1583 #ifdef TCP_OFFLOAD 1584 if (ADDED_BY_TOE(sc)) { 1585 struct toedev *tod = sc->sc_tod; 1586 1587 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1588 1589 return (error); 1590 } 1591 #endif 1592 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1593 } 1594 #endif 1595 return (error); 1596 } 1597 1598 /* 1599 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1600 * that exceed the capacity of the syncache by avoiding the storage of any 1601 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1602 * attacks where the attacker does not have access to our responses. 1603 * 1604 * Syncookies encode and include all necessary information about the 1605 * connection setup within the SYN|ACK that we send back. That way we 1606 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1607 * (if ever). Normally the syncache and syncookies are running in parallel 1608 * with the latter taking over when the former is exhausted. When matching 1609 * syncache entry is found the syncookie is ignored. 1610 * 1611 * The only reliable information persisting the 3WHS is our inital sequence 1612 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1613 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1614 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1615 * returns and signifies a legitimate connection if it matches the ACK. 1616 * 1617 * The available space of 32 bits to store the hash and to encode the SYN 1618 * option information is very tight and we should have at least 24 bits for 1619 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1620 * 1621 * SYN option information we have to encode to fully restore a connection: 1622 * MSS: is imporant to chose an optimal segment size to avoid IP level 1623 * fragmentation along the path. The common MSS values can be encoded 1624 * in a 3-bit table. Uncommon values are captured by the next lower value 1625 * in the table leading to a slight increase in packetization overhead. 1626 * WSCALE: is necessary to allow large windows to be used for high delay- 1627 * bandwidth product links. Not scaling the window when it was initially 1628 * negotiated is bad for performance as lack of scaling further decreases 1629 * the apparent available send window. We only need to encode the WSCALE 1630 * we received from the remote end. Our end can be recalculated at any 1631 * time. The common WSCALE values can be encoded in a 3-bit table. 1632 * Uncommon values are captured by the next lower value in the table 1633 * making us under-estimate the available window size halving our 1634 * theoretically possible maximum throughput for that connection. 1635 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1636 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1637 * that are included in all segments on a connection. We enable them when 1638 * the ACK has them. 1639 * 1640 * Security of syncookies and attack vectors: 1641 * 1642 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1643 * together with the gloabl secret to make it unique per connection attempt. 1644 * Thus any change of any of those parameters results in a different MAC output 1645 * in an unpredictable way unless a collision is encountered. 24 bits of the 1646 * MAC are embedded into the ISS. 1647 * 1648 * To prevent replay attacks two rotating global secrets are updated with a 1649 * new random value every 15 seconds. The life-time of a syncookie is thus 1650 * 15-30 seconds. 1651 * 1652 * Vector 1: Attacking the secret. This requires finding a weakness in the 1653 * MAC itself or the way it is used here. The attacker can do a chosen plain 1654 * text attack by varying and testing the all parameters under his control. 1655 * The strength depends on the size and randomness of the secret, and the 1656 * cryptographic security of the MAC function. Due to the constant updating 1657 * of the secret the attacker has at most 29.999 seconds to find the secret 1658 * and launch spoofed connections. After that he has to start all over again. 1659 * 1660 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1661 * size an average of 4,823 attempts are required for a 50% chance of success 1662 * to spoof a single syncookie (birthday collision paradox). However the 1663 * attacker is blind and doesn't know if one of his attempts succeeded unless 1664 * he has a side channel to interfere success from. A single connection setup 1665 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1666 * This many attempts are required for each one blind spoofed connection. For 1667 * every additional spoofed connection he has to launch another N attempts. 1668 * Thus for a sustained rate 100 spoofed connections per second approximately 1669 * 1,800,000 packets per second would have to be sent. 1670 * 1671 * NB: The MAC function should be fast so that it doesn't become a CPU 1672 * exhaustion attack vector itself. 1673 * 1674 * References: 1675 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1676 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1677 * http://cr.yp.to/syncookies.html (overview) 1678 * http://cr.yp.to/syncookies/archive (details) 1679 * 1680 * 1681 * Schematic construction of a syncookie enabled Initial Sequence Number: 1682 * 0 1 2 3 1683 * 12345678901234567890123456789012 1684 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1685 * 1686 * x 24 MAC (truncated) 1687 * W 3 Send Window Scale index 1688 * M 3 MSS index 1689 * S 1 SACK permitted 1690 * P 1 Odd/even secret 1691 */ 1692 1693 /* 1694 * Distribution and probability of certain MSS values. Those in between are 1695 * rounded down to the next lower one. 1696 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1697 * .2% .3% 5% 7% 7% 20% 15% 45% 1698 */ 1699 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1700 1701 /* 1702 * Distribution and probability of certain WSCALE values. We have to map the 1703 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1704 * bits based on prevalence of certain values. Where we don't have an exact 1705 * match for are rounded down to the next lower one letting us under-estimate 1706 * the true available window. At the moment this would happen only for the 1707 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1708 * and window size). The absence of the WSCALE option (no scaling in either 1709 * direction) is encoded with index zero. 1710 * [WSCALE values histograms, Allman, 2012] 1711 * X 10 10 35 5 6 14 10% by host 1712 * X 11 4 5 5 18 49 3% by connections 1713 */ 1714 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1715 1716 /* 1717 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1718 * and good cryptographic properties. 1719 */ 1720 static uint32_t 1721 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1722 uint8_t *secbits, uintptr_t secmod) 1723 { 1724 SIPHASH_CTX ctx; 1725 uint32_t siphash[2]; 1726 1727 SipHash24_Init(&ctx); 1728 SipHash_SetKey(&ctx, secbits); 1729 switch (inc->inc_flags & INC_ISIPV6) { 1730 #ifdef INET 1731 case 0: 1732 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1733 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1734 break; 1735 #endif 1736 #ifdef INET6 1737 case INC_ISIPV6: 1738 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1739 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1740 break; 1741 #endif 1742 } 1743 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1744 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1745 SipHash_Update(&ctx, &flags, sizeof(flags)); 1746 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1747 SipHash_Final((u_int8_t *)&siphash, &ctx); 1748 1749 return (siphash[0] ^ siphash[1]); 1750 } 1751 1752 static tcp_seq 1753 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 1754 { 1755 u_int i, mss, secbit, wscale; 1756 uint32_t iss, hash; 1757 uint8_t *secbits; 1758 union syncookie cookie; 1759 1760 SCH_LOCK_ASSERT(sch); 1761 1762 cookie.cookie = 0; 1763 1764 /* Map our computed MSS into the 3-bit index. */ 1765 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss)); 1766 for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1; 1767 tcp_sc_msstab[i] > mss && i > 0; 1768 i--) 1769 ; 1770 cookie.flags.mss_idx = i; 1771 1772 /* 1773 * Map the send window scale into the 3-bit index but only if 1774 * the wscale option was received. 1775 */ 1776 if (sc->sc_flags & SCF_WINSCALE) { 1777 wscale = sc->sc_requested_s_scale; 1778 for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1; 1779 tcp_sc_wstab[i] > wscale && i > 0; 1780 i--) 1781 ; 1782 cookie.flags.wscale_idx = i; 1783 } 1784 1785 /* Can we do SACK? */ 1786 if (sc->sc_flags & SCF_SACK) 1787 cookie.flags.sack_ok = 1; 1788 1789 /* Which of the two secrets to use. */ 1790 secbit = sch->sch_sc->secret.oddeven & 0x1; 1791 cookie.flags.odd_even = secbit; 1792 1793 secbits = sch->sch_sc->secret.key[secbit]; 1794 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 1795 (uintptr_t)sch); 1796 1797 /* 1798 * Put the flags into the hash and XOR them to get better ISS number 1799 * variance. This doesn't enhance the cryptographic strength and is 1800 * done to prevent the 8 cookie bits from showing up directly on the 1801 * wire. 1802 */ 1803 iss = hash & ~0xff; 1804 iss |= cookie.cookie ^ (hash >> 24); 1805 1806 /* Randomize the timestamp. */ 1807 if (sc->sc_flags & SCF_TIMESTAMP) { 1808 sc->sc_ts = arc4random(); 1809 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 1810 } 1811 1812 TCPSTAT_INC(tcps_sc_sendcookie); 1813 return (iss); 1814 } 1815 1816 static struct syncache * 1817 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 1818 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 1819 struct socket *lso) 1820 { 1821 uint32_t hash; 1822 uint8_t *secbits; 1823 tcp_seq ack, seq; 1824 int wnd, wscale = 0; 1825 union syncookie cookie; 1826 1827 SCH_LOCK_ASSERT(sch); 1828 1829 /* 1830 * Pull information out of SYN-ACK/ACK and revert sequence number 1831 * advances. 1832 */ 1833 ack = th->th_ack - 1; 1834 seq = th->th_seq - 1; 1835 1836 /* 1837 * Unpack the flags containing enough information to restore the 1838 * connection. 1839 */ 1840 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 1841 1842 /* Which of the two secrets to use. */ 1843 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 1844 1845 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 1846 1847 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 1848 if ((ack & ~0xff) != (hash & ~0xff)) 1849 return (NULL); 1850 1851 /* Fill in the syncache values. */ 1852 sc->sc_flags = 0; 1853 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1854 sc->sc_ipopts = NULL; 1855 1856 sc->sc_irs = seq; 1857 sc->sc_iss = ack; 1858 1859 switch (inc->inc_flags & INC_ISIPV6) { 1860 #ifdef INET 1861 case 0: 1862 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 1863 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 1864 break; 1865 #endif 1866 #ifdef INET6 1867 case INC_ISIPV6: 1868 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 1869 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 1870 break; 1871 #endif 1872 } 1873 1874 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 1875 1876 /* We can simply recompute receive window scale we sent earlier. */ 1877 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 1878 wscale++; 1879 1880 /* Only use wscale if it was enabled in the orignal SYN. */ 1881 if (cookie.flags.wscale_idx > 0) { 1882 sc->sc_requested_r_scale = wscale; 1883 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 1884 sc->sc_flags |= SCF_WINSCALE; 1885 } 1886 1887 wnd = sbspace(&lso->so_rcv); 1888 wnd = imax(wnd, 0); 1889 wnd = imin(wnd, TCP_MAXWIN); 1890 sc->sc_wnd = wnd; 1891 1892 if (cookie.flags.sack_ok) 1893 sc->sc_flags |= SCF_SACK; 1894 1895 if (to->to_flags & TOF_TS) { 1896 sc->sc_flags |= SCF_TIMESTAMP; 1897 sc->sc_tsreflect = to->to_tsval; 1898 sc->sc_ts = to->to_tsecr; 1899 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 1900 } 1901 1902 if (to->to_flags & TOF_SIGNATURE) 1903 sc->sc_flags |= SCF_SIGNATURE; 1904 1905 sc->sc_rxmits = 0; 1906 1907 TCPSTAT_INC(tcps_sc_recvcookie); 1908 return (sc); 1909 } 1910 1911 #ifdef INVARIANTS 1912 static int 1913 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 1914 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 1915 struct socket *lso) 1916 { 1917 struct syncache scs, *scx; 1918 char *s; 1919 1920 bzero(&scs, sizeof(scs)); 1921 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 1922 1923 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 1924 return (0); 1925 1926 if (scx != NULL) { 1927 if (sc->sc_peer_mss != scx->sc_peer_mss) 1928 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 1929 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 1930 1931 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 1932 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 1933 s, __func__, sc->sc_requested_r_scale, 1934 scx->sc_requested_r_scale); 1935 1936 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 1937 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 1938 s, __func__, sc->sc_requested_s_scale, 1939 scx->sc_requested_s_scale); 1940 1941 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 1942 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 1943 } 1944 1945 if (s != NULL) 1946 free(s, M_TCPLOG); 1947 return (0); 1948 } 1949 #endif /* INVARIANTS */ 1950 1951 static void 1952 syncookie_reseed(void *arg) 1953 { 1954 struct tcp_syncache *sc = arg; 1955 uint8_t *secbits; 1956 int secbit; 1957 1958 /* 1959 * Reseeding the secret doesn't have to be protected by a lock. 1960 * It only must be ensured that the new random values are visible 1961 * to all CPUs in a SMP environment. The atomic with release 1962 * semantics ensures that. 1963 */ 1964 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 1965 secbits = sc->secret.key[secbit]; 1966 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 1967 atomic_add_rel_int(&sc->secret.oddeven, 1); 1968 1969 /* Reschedule ourself. */ 1970 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 1971 } 1972 1973 /* 1974 * Returns the current number of syncache entries. This number 1975 * will probably change before you get around to calling 1976 * syncache_pcblist. 1977 */ 1978 int 1979 syncache_pcbcount(void) 1980 { 1981 struct syncache_head *sch; 1982 int count, i; 1983 1984 for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 1985 /* No need to lock for a read. */ 1986 sch = &V_tcp_syncache.hashbase[i]; 1987 count += sch->sch_length; 1988 } 1989 return count; 1990 } 1991 1992 /* 1993 * Exports the syncache entries to userland so that netstat can display 1994 * them alongside the other sockets. This function is intended to be 1995 * called only from tcp_pcblist. 1996 * 1997 * Due to concurrency on an active system, the number of pcbs exported 1998 * may have no relation to max_pcbs. max_pcbs merely indicates the 1999 * amount of space the caller allocated for this function to use. 2000 */ 2001 int 2002 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2003 { 2004 struct xtcpcb xt; 2005 struct syncache *sc; 2006 struct syncache_head *sch; 2007 int count, error, i; 2008 2009 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2010 sch = &V_tcp_syncache.hashbase[i]; 2011 SCH_LOCK(sch); 2012 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2013 if (count >= max_pcbs) { 2014 SCH_UNLOCK(sch); 2015 goto exit; 2016 } 2017 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2018 continue; 2019 bzero(&xt, sizeof(xt)); 2020 xt.xt_len = sizeof(xt); 2021 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2022 xt.xt_inp.inp_vflag = INP_IPV6; 2023 else 2024 xt.xt_inp.inp_vflag = INP_IPV4; 2025 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 2026 xt.xt_tp.t_inpcb = &xt.xt_inp; 2027 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 2028 xt.xt_socket.xso_protocol = IPPROTO_TCP; 2029 xt.xt_socket.xso_len = sizeof (struct xsocket); 2030 xt.xt_socket.so_type = SOCK_STREAM; 2031 xt.xt_socket.so_state = SS_ISCONNECTING; 2032 error = SYSCTL_OUT(req, &xt, sizeof xt); 2033 if (error) { 2034 SCH_UNLOCK(sch); 2035 goto exit; 2036 } 2037 count++; 2038 } 2039 SCH_UNLOCK(sch); 2040 } 2041 exit: 2042 *pcbs_exported = count; 2043 return error; 2044 } 2045