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_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies, 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_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies_only, 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 struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **); 126 static int syncache_respond(struct syncache *); 127 static struct socket *syncache_socket(struct syncache *, struct socket *, 128 struct mbuf *m); 129 static int syncache_sysctl_count(SYSCTL_HANDLER_ARGS); 130 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 131 int docallout); 132 static void syncache_timer(void *); 133 134 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 135 uint8_t *, uintptr_t); 136 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 137 static struct syncache 138 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 139 struct syncache *, struct tcphdr *, struct tcpopt *, 140 struct socket *); 141 static void syncookie_reseed(void *); 142 #ifdef INVARIANTS 143 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 144 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 145 struct socket *lso); 146 #endif 147 148 /* 149 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 150 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds, 151 * the odds are that the user has given up attempting to connect by then. 152 */ 153 #define SYNCACHE_MAXREXMTS 3 154 155 /* Arbitrary values */ 156 #define TCP_SYNCACHE_HASHSIZE 512 157 #define TCP_SYNCACHE_BUCKETLIMIT 30 158 159 static VNET_DEFINE(struct tcp_syncache, tcp_syncache); 160 #define V_tcp_syncache VNET(tcp_syncache) 161 162 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, 163 "TCP SYN cache"); 164 165 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN, 166 &VNET_NAME(tcp_syncache.bucket_limit), 0, 167 "Per-bucket hash limit for syncache"); 168 169 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN, 170 &VNET_NAME(tcp_syncache.cache_limit), 0, 171 "Overall entry limit for syncache"); 172 173 SYSCTL_VNET_PROC(_net_inet_tcp_syncache, OID_AUTO, count, (CTLTYPE_UINT|CTLFLAG_RD), 174 NULL, 0, &syncache_sysctl_count, "IU", 175 "Current number of entries in syncache"); 176 177 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN, 178 &VNET_NAME(tcp_syncache.hashsize), 0, 179 "Size of TCP syncache hashtable"); 180 181 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW, 182 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 183 "Limit on SYN/ACK retransmissions"); 184 185 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 186 SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 187 CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 188 "Send reset on socket allocation failure"); 189 190 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 191 192 #define SYNCACHE_HASH(inc, mask) \ 193 ((V_tcp_syncache.hash_secret ^ \ 194 (inc)->inc_faddr.s_addr ^ \ 195 ((inc)->inc_faddr.s_addr >> 16) ^ \ 196 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 197 198 #define SYNCACHE_HASH6(inc, mask) \ 199 ((V_tcp_syncache.hash_secret ^ \ 200 (inc)->inc6_faddr.s6_addr32[0] ^ \ 201 (inc)->inc6_faddr.s6_addr32[3] ^ \ 202 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 203 204 #define ENDPTS_EQ(a, b) ( \ 205 (a)->ie_fport == (b)->ie_fport && \ 206 (a)->ie_lport == (b)->ie_lport && \ 207 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \ 208 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \ 209 ) 210 211 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0) 212 213 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 214 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 215 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 216 217 /* 218 * Requires the syncache entry to be already removed from the bucket list. 219 */ 220 static void 221 syncache_free(struct syncache *sc) 222 { 223 224 if (sc->sc_ipopts) 225 (void) m_free(sc->sc_ipopts); 226 if (sc->sc_cred) 227 crfree(sc->sc_cred); 228 #ifdef MAC 229 mac_syncache_destroy(&sc->sc_label); 230 #endif 231 232 uma_zfree(V_tcp_syncache.zone, sc); 233 } 234 235 void 236 syncache_init(void) 237 { 238 int i; 239 240 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 241 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 242 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 243 V_tcp_syncache.hash_secret = arc4random(); 244 245 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 246 &V_tcp_syncache.hashsize); 247 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 248 &V_tcp_syncache.bucket_limit); 249 if (!powerof2(V_tcp_syncache.hashsize) || 250 V_tcp_syncache.hashsize == 0) { 251 printf("WARNING: syncache hash size is not a power of 2.\n"); 252 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 253 } 254 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 255 256 /* Set limits. */ 257 V_tcp_syncache.cache_limit = 258 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 259 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 260 &V_tcp_syncache.cache_limit); 261 262 /* Allocate the hash table. */ 263 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 264 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 265 266 #ifdef VIMAGE 267 V_tcp_syncache.vnet = curvnet; 268 #endif 269 270 /* Initialize the hash buckets. */ 271 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 272 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 273 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 274 NULL, MTX_DEF); 275 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 276 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 277 V_tcp_syncache.hashbase[i].sch_length = 0; 278 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 279 } 280 281 /* Create the syncache entry zone. */ 282 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 283 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 284 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 285 V_tcp_syncache.cache_limit); 286 287 /* Start the SYN cookie reseeder callout. */ 288 callout_init(&V_tcp_syncache.secret.reseed, 1); 289 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 290 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 291 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 292 syncookie_reseed, &V_tcp_syncache); 293 } 294 295 #ifdef VIMAGE 296 void 297 syncache_destroy(void) 298 { 299 struct syncache_head *sch; 300 struct syncache *sc, *nsc; 301 int i; 302 303 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 304 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 305 306 sch = &V_tcp_syncache.hashbase[i]; 307 callout_drain(&sch->sch_timer); 308 309 SCH_LOCK(sch); 310 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 311 syncache_drop(sc, sch); 312 SCH_UNLOCK(sch); 313 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 314 ("%s: sch->sch_bucket not empty", __func__)); 315 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 316 __func__, sch->sch_length)); 317 mtx_destroy(&sch->sch_mtx); 318 } 319 320 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 321 ("%s: cache_count not 0", __func__)); 322 323 /* Free the allocated global resources. */ 324 uma_zdestroy(V_tcp_syncache.zone); 325 free(V_tcp_syncache.hashbase, M_SYNCACHE); 326 327 callout_drain(&V_tcp_syncache.secret.reseed); 328 } 329 #endif 330 331 static int 332 syncache_sysctl_count(SYSCTL_HANDLER_ARGS) 333 { 334 int count; 335 336 count = uma_zone_get_cur(V_tcp_syncache.zone); 337 return (sysctl_handle_int(oidp, &count, 0, req)); 338 } 339 340 /* 341 * Inserts a syncache entry into the specified bucket row. 342 * Locks and unlocks the syncache_head autonomously. 343 */ 344 static void 345 syncache_insert(struct syncache *sc, struct syncache_head *sch) 346 { 347 struct syncache *sc2; 348 349 SCH_LOCK(sch); 350 351 /* 352 * Make sure that we don't overflow the per-bucket limit. 353 * If the bucket is full, toss the oldest element. 354 */ 355 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 356 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 357 ("sch->sch_length incorrect")); 358 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 359 syncache_drop(sc2, sch); 360 TCPSTAT_INC(tcps_sc_bucketoverflow); 361 } 362 363 /* Put it into the bucket. */ 364 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 365 sch->sch_length++; 366 367 #ifdef TCP_OFFLOAD 368 if (ADDED_BY_TOE(sc)) { 369 struct toedev *tod = sc->sc_tod; 370 371 tod->tod_syncache_added(tod, sc->sc_todctx); 372 } 373 #endif 374 375 /* Reinitialize the bucket row's timer. */ 376 if (sch->sch_length == 1) 377 sch->sch_nextc = ticks + INT_MAX; 378 syncache_timeout(sc, sch, 1); 379 380 SCH_UNLOCK(sch); 381 382 TCPSTAT_INC(tcps_sc_added); 383 } 384 385 /* 386 * Remove and free entry from syncache bucket row. 387 * Expects locked syncache head. 388 */ 389 static void 390 syncache_drop(struct syncache *sc, struct syncache_head *sch) 391 { 392 393 SCH_LOCK_ASSERT(sch); 394 395 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 396 sch->sch_length--; 397 398 #ifdef TCP_OFFLOAD 399 if (ADDED_BY_TOE(sc)) { 400 struct toedev *tod = sc->sc_tod; 401 402 tod->tod_syncache_removed(tod, sc->sc_todctx); 403 } 404 #endif 405 406 syncache_free(sc); 407 } 408 409 /* 410 * Engage/reengage time on bucket row. 411 */ 412 static void 413 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 414 { 415 sc->sc_rxttime = ticks + 416 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]); 417 sc->sc_rxmits++; 418 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 419 sch->sch_nextc = sc->sc_rxttime; 420 if (docallout) 421 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 422 syncache_timer, (void *)sch); 423 } 424 } 425 426 /* 427 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 428 * If we have retransmitted an entry the maximum number of times, expire it. 429 * One separate timer for each bucket row. 430 */ 431 static void 432 syncache_timer(void *xsch) 433 { 434 struct syncache_head *sch = (struct syncache_head *)xsch; 435 struct syncache *sc, *nsc; 436 int tick = ticks; 437 char *s; 438 439 CURVNET_SET(sch->sch_sc->vnet); 440 441 /* NB: syncache_head has already been locked by the callout. */ 442 SCH_LOCK_ASSERT(sch); 443 444 /* 445 * In the following cycle we may remove some entries and/or 446 * advance some timeouts, so re-initialize the bucket timer. 447 */ 448 sch->sch_nextc = tick + INT_MAX; 449 450 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 451 /* 452 * We do not check if the listen socket still exists 453 * and accept the case where the listen socket may be 454 * gone by the time we resend the SYN/ACK. We do 455 * not expect this to happens often. If it does, 456 * then the RST will be sent by the time the remote 457 * host does the SYN/ACK->ACK. 458 */ 459 if (TSTMP_GT(sc->sc_rxttime, tick)) { 460 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 461 sch->sch_nextc = sc->sc_rxttime; 462 continue; 463 } 464 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 465 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 466 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 467 "giving up and removing syncache entry\n", 468 s, __func__); 469 free(s, M_TCPLOG); 470 } 471 syncache_drop(sc, sch); 472 TCPSTAT_INC(tcps_sc_stale); 473 continue; 474 } 475 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 476 log(LOG_DEBUG, "%s; %s: Response timeout, " 477 "retransmitting (%u) SYN|ACK\n", 478 s, __func__, sc->sc_rxmits); 479 free(s, M_TCPLOG); 480 } 481 482 (void) syncache_respond(sc); 483 TCPSTAT_INC(tcps_sc_retransmitted); 484 syncache_timeout(sc, sch, 0); 485 } 486 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 487 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 488 syncache_timer, (void *)(sch)); 489 CURVNET_RESTORE(); 490 } 491 492 /* 493 * Find an entry in the syncache. 494 * Returns always with locked syncache_head plus a matching entry or NULL. 495 */ 496 struct syncache * 497 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 498 { 499 struct syncache *sc; 500 struct syncache_head *sch; 501 502 #ifdef INET6 503 if (inc->inc_flags & INC_ISIPV6) { 504 sch = &V_tcp_syncache.hashbase[ 505 SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)]; 506 *schp = sch; 507 508 SCH_LOCK(sch); 509 510 /* Circle through bucket row to find matching entry. */ 511 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 512 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 513 return (sc); 514 } 515 } else 516 #endif 517 { 518 sch = &V_tcp_syncache.hashbase[ 519 SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)]; 520 *schp = sch; 521 522 SCH_LOCK(sch); 523 524 /* Circle through bucket row to find matching entry. */ 525 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 526 #ifdef INET6 527 if (sc->sc_inc.inc_flags & INC_ISIPV6) 528 continue; 529 #endif 530 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 531 return (sc); 532 } 533 } 534 SCH_LOCK_ASSERT(*schp); 535 return (NULL); /* always returns with locked sch */ 536 } 537 538 /* 539 * This function is called when we get a RST for a 540 * non-existent connection, so that we can see if the 541 * connection is in the syn cache. If it is, zap it. 542 */ 543 void 544 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 545 { 546 struct syncache *sc; 547 struct syncache_head *sch; 548 char *s = NULL; 549 550 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 551 SCH_LOCK_ASSERT(sch); 552 553 /* 554 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 555 * See RFC 793 page 65, section SEGMENT ARRIVES. 556 */ 557 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 558 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 559 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 560 "FIN flag set, segment ignored\n", s, __func__); 561 TCPSTAT_INC(tcps_badrst); 562 goto done; 563 } 564 565 /* 566 * No corresponding connection was found in syncache. 567 * If syncookies are enabled and possibly exclusively 568 * used, or we are under memory pressure, a valid RST 569 * may not find a syncache entry. In that case we're 570 * done and no SYN|ACK retransmissions will happen. 571 * Otherwise the RST was misdirected or spoofed. 572 */ 573 if (sc == NULL) { 574 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 575 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 576 "syncache entry (possibly syncookie only), " 577 "segment ignored\n", s, __func__); 578 TCPSTAT_INC(tcps_badrst); 579 goto done; 580 } 581 582 /* 583 * If the RST bit is set, check the sequence number to see 584 * if this is a valid reset segment. 585 * RFC 793 page 37: 586 * In all states except SYN-SENT, all reset (RST) segments 587 * are validated by checking their SEQ-fields. A reset is 588 * valid if its sequence number is in the window. 589 * 590 * The sequence number in the reset segment is normally an 591 * echo of our outgoing acknowlegement numbers, but some hosts 592 * send a reset with the sequence number at the rightmost edge 593 * of our receive window, and we have to handle this case. 594 */ 595 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 596 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 597 syncache_drop(sc, sch); 598 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 599 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 600 "connection attempt aborted by remote endpoint\n", 601 s, __func__); 602 TCPSTAT_INC(tcps_sc_reset); 603 } else { 604 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 605 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 606 "IRS %u (+WND %u), segment ignored\n", 607 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 608 TCPSTAT_INC(tcps_badrst); 609 } 610 611 done: 612 if (s != NULL) 613 free(s, M_TCPLOG); 614 SCH_UNLOCK(sch); 615 } 616 617 void 618 syncache_badack(struct in_conninfo *inc) 619 { 620 struct syncache *sc; 621 struct syncache_head *sch; 622 623 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 624 SCH_LOCK_ASSERT(sch); 625 if (sc != NULL) { 626 syncache_drop(sc, sch); 627 TCPSTAT_INC(tcps_sc_badack); 628 } 629 SCH_UNLOCK(sch); 630 } 631 632 void 633 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th) 634 { 635 struct syncache *sc; 636 struct syncache_head *sch; 637 638 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 639 SCH_LOCK_ASSERT(sch); 640 if (sc == NULL) 641 goto done; 642 643 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 644 if (ntohl(th->th_seq) != sc->sc_iss) 645 goto done; 646 647 /* 648 * If we've rertransmitted 3 times and this is our second error, 649 * we remove the entry. Otherwise, we allow it to continue on. 650 * This prevents us from incorrectly nuking an entry during a 651 * spurious network outage. 652 * 653 * See tcp_notify(). 654 */ 655 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 656 sc->sc_flags |= SCF_UNREACH; 657 goto done; 658 } 659 syncache_drop(sc, sch); 660 TCPSTAT_INC(tcps_sc_unreach); 661 done: 662 SCH_UNLOCK(sch); 663 } 664 665 /* 666 * Build a new TCP socket structure from a syncache entry. 667 */ 668 static struct socket * 669 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 670 { 671 struct inpcb *inp = NULL; 672 struct socket *so; 673 struct tcpcb *tp; 674 int error; 675 char *s; 676 677 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 678 679 /* 680 * Ok, create the full blown connection, and set things up 681 * as they would have been set up if we had created the 682 * connection when the SYN arrived. If we can't create 683 * the connection, abort it. 684 */ 685 so = sonewconn(lso, SS_ISCONNECTED); 686 if (so == NULL) { 687 /* 688 * Drop the connection; we will either send a RST or 689 * have the peer retransmit its SYN again after its 690 * RTO and try again. 691 */ 692 TCPSTAT_INC(tcps_listendrop); 693 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 694 log(LOG_DEBUG, "%s; %s: Socket create failed " 695 "due to limits or memory shortage\n", 696 s, __func__); 697 free(s, M_TCPLOG); 698 } 699 goto abort2; 700 } 701 #ifdef MAC 702 mac_socketpeer_set_from_mbuf(m, so); 703 #endif 704 705 inp = sotoinpcb(so); 706 inp->inp_inc.inc_fibnum = so->so_fibnum; 707 INP_WLOCK(inp); 708 INP_HASH_WLOCK(&V_tcbinfo); 709 710 /* Insert new socket into PCB hash list. */ 711 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 712 #ifdef INET6 713 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 714 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 715 } else { 716 inp->inp_vflag &= ~INP_IPV6; 717 inp->inp_vflag |= INP_IPV4; 718 #endif 719 inp->inp_laddr = sc->sc_inc.inc_laddr; 720 #ifdef INET6 721 } 722 #endif 723 724 /* 725 * If there's an mbuf and it has a flowid, then let's initialise the 726 * inp with that particular flowid. 727 */ 728 if (m != NULL && m->m_flags & M_FLOWID) { 729 inp->inp_flags |= INP_HW_FLOWID; 730 inp->inp_flags &= ~INP_SW_FLOWID; 731 inp->inp_flowid = m->m_pkthdr.flowid; 732 } 733 734 /* 735 * Install in the reservation hash table for now, but don't yet 736 * install a connection group since the full 4-tuple isn't yet 737 * configured. 738 */ 739 inp->inp_lport = sc->sc_inc.inc_lport; 740 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 741 /* 742 * Undo the assignments above if we failed to 743 * put the PCB on the hash lists. 744 */ 745 #ifdef INET6 746 if (sc->sc_inc.inc_flags & INC_ISIPV6) 747 inp->in6p_laddr = in6addr_any; 748 else 749 #endif 750 inp->inp_laddr.s_addr = INADDR_ANY; 751 inp->inp_lport = 0; 752 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 753 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 754 "with error %i\n", 755 s, __func__, error); 756 free(s, M_TCPLOG); 757 } 758 INP_HASH_WUNLOCK(&V_tcbinfo); 759 goto abort; 760 } 761 #ifdef IPSEC 762 /* Copy old policy into new socket's. */ 763 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp)) 764 printf("syncache_socket: could not copy policy\n"); 765 #endif 766 #ifdef INET6 767 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 768 struct inpcb *oinp = sotoinpcb(lso); 769 struct in6_addr laddr6; 770 struct sockaddr_in6 sin6; 771 /* 772 * Inherit socket options from the listening socket. 773 * Note that in6p_inputopts are not (and should not be) 774 * copied, since it stores previously received options and is 775 * used to detect if each new option is different than the 776 * previous one and hence should be passed to a user. 777 * If we copied in6p_inputopts, a user would not be able to 778 * receive options just after calling the accept system call. 779 */ 780 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 781 if (oinp->in6p_outputopts) 782 inp->in6p_outputopts = 783 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 784 785 sin6.sin6_family = AF_INET6; 786 sin6.sin6_len = sizeof(sin6); 787 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 788 sin6.sin6_port = sc->sc_inc.inc_fport; 789 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 790 laddr6 = inp->in6p_laddr; 791 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 792 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 793 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 794 thread0.td_ucred, m)) != 0) { 795 inp->in6p_laddr = laddr6; 796 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 797 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 798 "with error %i\n", 799 s, __func__, error); 800 free(s, M_TCPLOG); 801 } 802 INP_HASH_WUNLOCK(&V_tcbinfo); 803 goto abort; 804 } 805 /* Override flowlabel from in6_pcbconnect. */ 806 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 807 inp->inp_flow |= sc->sc_flowlabel; 808 } 809 #endif /* INET6 */ 810 #if defined(INET) && defined(INET6) 811 else 812 #endif 813 #ifdef INET 814 { 815 struct in_addr laddr; 816 struct sockaddr_in sin; 817 818 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 819 820 if (inp->inp_options == NULL) { 821 inp->inp_options = sc->sc_ipopts; 822 sc->sc_ipopts = NULL; 823 } 824 825 sin.sin_family = AF_INET; 826 sin.sin_len = sizeof(sin); 827 sin.sin_addr = sc->sc_inc.inc_faddr; 828 sin.sin_port = sc->sc_inc.inc_fport; 829 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 830 laddr = inp->inp_laddr; 831 if (inp->inp_laddr.s_addr == INADDR_ANY) 832 inp->inp_laddr = sc->sc_inc.inc_laddr; 833 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 834 thread0.td_ucred, m)) != 0) { 835 inp->inp_laddr = laddr; 836 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 837 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 838 "with error %i\n", 839 s, __func__, error); 840 free(s, M_TCPLOG); 841 } 842 INP_HASH_WUNLOCK(&V_tcbinfo); 843 goto abort; 844 } 845 } 846 #endif /* INET */ 847 INP_HASH_WUNLOCK(&V_tcbinfo); 848 tp = intotcpcb(inp); 849 tcp_state_change(tp, TCPS_SYN_RECEIVED); 850 tp->iss = sc->sc_iss; 851 tp->irs = sc->sc_irs; 852 tcp_rcvseqinit(tp); 853 tcp_sendseqinit(tp); 854 tp->snd_wl1 = sc->sc_irs; 855 tp->snd_max = tp->iss + 1; 856 tp->snd_nxt = tp->iss + 1; 857 tp->rcv_up = sc->sc_irs + 1; 858 tp->rcv_wnd = sc->sc_wnd; 859 tp->rcv_adv += tp->rcv_wnd; 860 tp->last_ack_sent = tp->rcv_nxt; 861 862 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 863 if (sc->sc_flags & SCF_NOOPT) 864 tp->t_flags |= TF_NOOPT; 865 else { 866 if (sc->sc_flags & SCF_WINSCALE) { 867 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 868 tp->snd_scale = sc->sc_requested_s_scale; 869 tp->request_r_scale = sc->sc_requested_r_scale; 870 } 871 if (sc->sc_flags & SCF_TIMESTAMP) { 872 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 873 tp->ts_recent = sc->sc_tsreflect; 874 tp->ts_recent_age = tcp_ts_getticks(); 875 tp->ts_offset = sc->sc_tsoff; 876 } 877 #ifdef TCP_SIGNATURE 878 if (sc->sc_flags & SCF_SIGNATURE) 879 tp->t_flags |= TF_SIGNATURE; 880 #endif 881 if (sc->sc_flags & SCF_SACK) 882 tp->t_flags |= TF_SACK_PERMIT; 883 } 884 885 if (sc->sc_flags & SCF_ECN) 886 tp->t_flags |= TF_ECN_PERMIT; 887 888 /* 889 * Set up MSS and get cached values from tcp_hostcache. 890 * This might overwrite some of the defaults we just set. 891 */ 892 tcp_mss(tp, sc->sc_peer_mss); 893 894 /* 895 * If the SYN,ACK was retransmitted, indicate that CWND to be 896 * limited to one segment in cc_conn_init(). 897 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 898 */ 899 if (sc->sc_rxmits > 1) 900 tp->snd_cwnd = 1; 901 902 #ifdef TCP_OFFLOAD 903 /* 904 * Allow a TOE driver to install its hooks. Note that we hold the 905 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 906 * new connection before the TOE driver has done its thing. 907 */ 908 if (ADDED_BY_TOE(sc)) { 909 struct toedev *tod = sc->sc_tod; 910 911 tod->tod_offload_socket(tod, sc->sc_todctx, so); 912 } 913 #endif 914 /* 915 * Copy and activate timers. 916 */ 917 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 918 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 919 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 920 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 921 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 922 923 INP_WUNLOCK(inp); 924 925 TCPSTAT_INC(tcps_accepts); 926 return (so); 927 928 abort: 929 INP_WUNLOCK(inp); 930 abort2: 931 if (so != NULL) 932 soabort(so); 933 return (NULL); 934 } 935 936 /* 937 * This function gets called when we receive an ACK for a 938 * socket in the LISTEN state. We look up the connection 939 * in the syncache, and if its there, we pull it out of 940 * the cache and turn it into a full-blown connection in 941 * the SYN-RECEIVED state. 942 */ 943 int 944 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 945 struct socket **lsop, struct mbuf *m) 946 { 947 struct syncache *sc; 948 struct syncache_head *sch; 949 struct syncache scs; 950 char *s; 951 952 /* 953 * Global TCP locks are held because we manipulate the PCB lists 954 * and create a new socket. 955 */ 956 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 957 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 958 ("%s: can handle only ACK", __func__)); 959 960 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 961 SCH_LOCK_ASSERT(sch); 962 963 #ifdef INVARIANTS 964 /* 965 * Test code for syncookies comparing the syncache stored 966 * values with the reconstructed values from the cookie. 967 */ 968 if (sc != NULL) 969 syncookie_cmp(inc, sch, sc, th, to, *lsop); 970 #endif 971 972 if (sc == NULL) { 973 /* 974 * There is no syncache entry, so see if this ACK is 975 * a returning syncookie. To do this, first: 976 * A. See if this socket has had a syncache entry dropped in 977 * the past. We don't want to accept a bogus syncookie 978 * if we've never received a SYN. 979 * B. check that the syncookie is valid. If it is, then 980 * cobble up a fake syncache entry, and return. 981 */ 982 if (!V_tcp_syncookies) { 983 SCH_UNLOCK(sch); 984 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 985 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 986 "segment rejected (syncookies disabled)\n", 987 s, __func__); 988 goto failed; 989 } 990 bzero(&scs, sizeof(scs)); 991 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 992 SCH_UNLOCK(sch); 993 if (sc == NULL) { 994 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 995 log(LOG_DEBUG, "%s; %s: Segment failed " 996 "SYNCOOKIE authentication, segment rejected " 997 "(probably spoofed)\n", s, __func__); 998 goto failed; 999 } 1000 } else { 1001 /* Pull out the entry to unlock the bucket row. */ 1002 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1003 sch->sch_length--; 1004 #ifdef TCP_OFFLOAD 1005 if (ADDED_BY_TOE(sc)) { 1006 struct toedev *tod = sc->sc_tod; 1007 1008 tod->tod_syncache_removed(tod, sc->sc_todctx); 1009 } 1010 #endif 1011 SCH_UNLOCK(sch); 1012 } 1013 1014 /* 1015 * Segment validation: 1016 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1017 */ 1018 if (th->th_ack != sc->sc_iss + 1) { 1019 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1020 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1021 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1022 goto failed; 1023 } 1024 1025 /* 1026 * The SEQ must fall in the window starting at the received 1027 * initial receive sequence number + 1 (the SYN). 1028 */ 1029 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1030 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1031 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1032 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1033 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1034 goto failed; 1035 } 1036 1037 /* 1038 * If timestamps were not negotiated during SYN/ACK they 1039 * must not appear on any segment during this session. 1040 */ 1041 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1042 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1043 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1044 "segment rejected\n", s, __func__); 1045 goto failed; 1046 } 1047 1048 /* 1049 * If timestamps were negotiated during SYN/ACK they should 1050 * appear on every segment during this session. 1051 * XXXAO: This is only informal as there have been unverified 1052 * reports of non-compliants stacks. 1053 */ 1054 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1055 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1056 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1057 "no action\n", s, __func__); 1058 free(s, M_TCPLOG); 1059 s = NULL; 1060 } 1061 } 1062 1063 /* 1064 * If timestamps were negotiated the reflected timestamp 1065 * must be equal to what we actually sent in the SYN|ACK. 1066 */ 1067 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) { 1068 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1069 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1070 "segment rejected\n", 1071 s, __func__, to->to_tsecr, sc->sc_ts); 1072 goto failed; 1073 } 1074 1075 *lsop = syncache_socket(sc, *lsop, m); 1076 1077 if (*lsop == NULL) 1078 TCPSTAT_INC(tcps_sc_aborted); 1079 else 1080 TCPSTAT_INC(tcps_sc_completed); 1081 1082 /* how do we find the inp for the new socket? */ 1083 if (sc != &scs) 1084 syncache_free(sc); 1085 return (1); 1086 failed: 1087 if (sc != NULL && sc != &scs) 1088 syncache_free(sc); 1089 if (s != NULL) 1090 free(s, M_TCPLOG); 1091 *lsop = NULL; 1092 return (0); 1093 } 1094 1095 /* 1096 * Given a LISTEN socket and an inbound SYN request, add 1097 * this to the syn cache, and send back a segment: 1098 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1099 * to the source. 1100 * 1101 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1102 * Doing so would require that we hold onto the data and deliver it 1103 * to the application. However, if we are the target of a SYN-flood 1104 * DoS attack, an attacker could send data which would eventually 1105 * consume all available buffer space if it were ACKed. By not ACKing 1106 * the data, we avoid this DoS scenario. 1107 */ 1108 void 1109 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1110 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1111 void *todctx) 1112 { 1113 struct tcpcb *tp; 1114 struct socket *so; 1115 struct syncache *sc = NULL; 1116 struct syncache_head *sch; 1117 struct mbuf *ipopts = NULL; 1118 u_int ltflags; 1119 int win, sb_hiwat, ip_ttl, ip_tos; 1120 char *s; 1121 #ifdef INET6 1122 int autoflowlabel = 0; 1123 #endif 1124 #ifdef MAC 1125 struct label *maclabel; 1126 #endif 1127 struct syncache scs; 1128 struct ucred *cred; 1129 1130 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 1131 INP_WLOCK_ASSERT(inp); /* listen socket */ 1132 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1133 ("%s: unexpected tcp flags", __func__)); 1134 1135 /* 1136 * Combine all so/tp operations very early to drop the INP lock as 1137 * soon as possible. 1138 */ 1139 so = *lsop; 1140 tp = sototcpcb(so); 1141 cred = crhold(so->so_cred); 1142 1143 #ifdef INET6 1144 if ((inc->inc_flags & INC_ISIPV6) && 1145 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1146 autoflowlabel = 1; 1147 #endif 1148 ip_ttl = inp->inp_ip_ttl; 1149 ip_tos = inp->inp_ip_tos; 1150 win = sbspace(&so->so_rcv); 1151 sb_hiwat = so->so_rcv.sb_hiwat; 1152 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1153 1154 /* By the time we drop the lock these should no longer be used. */ 1155 so = NULL; 1156 tp = NULL; 1157 1158 #ifdef MAC 1159 if (mac_syncache_init(&maclabel) != 0) { 1160 INP_WUNLOCK(inp); 1161 INP_INFO_WUNLOCK(&V_tcbinfo); 1162 goto done; 1163 } else 1164 mac_syncache_create(maclabel, inp); 1165 #endif 1166 INP_WUNLOCK(inp); 1167 INP_INFO_WUNLOCK(&V_tcbinfo); 1168 1169 /* 1170 * Remember the IP options, if any. 1171 */ 1172 #ifdef INET6 1173 if (!(inc->inc_flags & INC_ISIPV6)) 1174 #endif 1175 #ifdef INET 1176 ipopts = (m) ? ip_srcroute(m) : NULL; 1177 #else 1178 ipopts = NULL; 1179 #endif 1180 1181 /* 1182 * See if we already have an entry for this connection. 1183 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1184 * 1185 * XXX: should the syncache be re-initialized with the contents 1186 * of the new SYN here (which may have different options?) 1187 * 1188 * XXX: We do not check the sequence number to see if this is a 1189 * real retransmit or a new connection attempt. The question is 1190 * how to handle such a case; either ignore it as spoofed, or 1191 * drop the current entry and create a new one? 1192 */ 1193 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1194 SCH_LOCK_ASSERT(sch); 1195 if (sc != NULL) { 1196 TCPSTAT_INC(tcps_sc_dupsyn); 1197 if (ipopts) { 1198 /* 1199 * If we were remembering a previous source route, 1200 * forget it and use the new one we've been given. 1201 */ 1202 if (sc->sc_ipopts) 1203 (void) m_free(sc->sc_ipopts); 1204 sc->sc_ipopts = ipopts; 1205 } 1206 /* 1207 * Update timestamp if present. 1208 */ 1209 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1210 sc->sc_tsreflect = to->to_tsval; 1211 else 1212 sc->sc_flags &= ~SCF_TIMESTAMP; 1213 #ifdef MAC 1214 /* 1215 * Since we have already unconditionally allocated label 1216 * storage, free it up. The syncache entry will already 1217 * have an initialized label we can use. 1218 */ 1219 mac_syncache_destroy(&maclabel); 1220 #endif 1221 /* Retransmit SYN|ACK and reset retransmit count. */ 1222 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1223 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1224 "resetting timer and retransmitting SYN|ACK\n", 1225 s, __func__); 1226 free(s, M_TCPLOG); 1227 } 1228 if (syncache_respond(sc) == 0) { 1229 sc->sc_rxmits = 0; 1230 syncache_timeout(sc, sch, 1); 1231 TCPSTAT_INC(tcps_sndacks); 1232 TCPSTAT_INC(tcps_sndtotal); 1233 } 1234 SCH_UNLOCK(sch); 1235 goto done; 1236 } 1237 1238 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1239 if (sc == NULL) { 1240 /* 1241 * The zone allocator couldn't provide more entries. 1242 * Treat this as if the cache was full; drop the oldest 1243 * entry and insert the new one. 1244 */ 1245 TCPSTAT_INC(tcps_sc_zonefail); 1246 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) 1247 syncache_drop(sc, sch); 1248 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1249 if (sc == NULL) { 1250 if (V_tcp_syncookies) { 1251 bzero(&scs, sizeof(scs)); 1252 sc = &scs; 1253 } else { 1254 SCH_UNLOCK(sch); 1255 if (ipopts) 1256 (void) m_free(ipopts); 1257 goto done; 1258 } 1259 } 1260 } 1261 1262 /* 1263 * Fill in the syncache values. 1264 */ 1265 #ifdef MAC 1266 sc->sc_label = maclabel; 1267 #endif 1268 sc->sc_cred = cred; 1269 cred = NULL; 1270 sc->sc_ipopts = ipopts; 1271 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1272 #ifdef INET6 1273 if (!(inc->inc_flags & INC_ISIPV6)) 1274 #endif 1275 { 1276 sc->sc_ip_tos = ip_tos; 1277 sc->sc_ip_ttl = ip_ttl; 1278 } 1279 #ifdef TCP_OFFLOAD 1280 sc->sc_tod = tod; 1281 sc->sc_todctx = todctx; 1282 #endif 1283 sc->sc_irs = th->th_seq; 1284 sc->sc_iss = arc4random(); 1285 sc->sc_flags = 0; 1286 sc->sc_flowlabel = 0; 1287 1288 /* 1289 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1290 * win was derived from socket earlier in the function. 1291 */ 1292 win = imax(win, 0); 1293 win = imin(win, TCP_MAXWIN); 1294 sc->sc_wnd = win; 1295 1296 if (V_tcp_do_rfc1323) { 1297 /* 1298 * A timestamp received in a SYN makes 1299 * it ok to send timestamp requests and replies. 1300 */ 1301 if (to->to_flags & TOF_TS) { 1302 sc->sc_tsreflect = to->to_tsval; 1303 sc->sc_ts = tcp_ts_getticks(); 1304 sc->sc_flags |= SCF_TIMESTAMP; 1305 } 1306 if (to->to_flags & TOF_SCALE) { 1307 int wscale = 0; 1308 1309 /* 1310 * Pick the smallest possible scaling factor that 1311 * will still allow us to scale up to sb_max, aka 1312 * kern.ipc.maxsockbuf. 1313 * 1314 * We do this because there are broken firewalls that 1315 * will corrupt the window scale option, leading to 1316 * the other endpoint believing that our advertised 1317 * window is unscaled. At scale factors larger than 1318 * 5 the unscaled window will drop below 1500 bytes, 1319 * leading to serious problems when traversing these 1320 * broken firewalls. 1321 * 1322 * With the default maxsockbuf of 256K, a scale factor 1323 * of 3 will be chosen by this algorithm. Those who 1324 * choose a larger maxsockbuf should watch out 1325 * for the compatiblity problems mentioned above. 1326 * 1327 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1328 * or <SYN,ACK>) segment itself is never scaled. 1329 */ 1330 while (wscale < TCP_MAX_WINSHIFT && 1331 (TCP_MAXWIN << wscale) < sb_max) 1332 wscale++; 1333 sc->sc_requested_r_scale = wscale; 1334 sc->sc_requested_s_scale = to->to_wscale; 1335 sc->sc_flags |= SCF_WINSCALE; 1336 } 1337 } 1338 #ifdef TCP_SIGNATURE 1339 /* 1340 * If listening socket requested TCP digests, and received SYN 1341 * contains the option, flag this in the syncache so that 1342 * syncache_respond() will do the right thing with the SYN+ACK. 1343 * XXX: Currently we always record the option by default and will 1344 * attempt to use it in syncache_respond(). 1345 */ 1346 if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE) 1347 sc->sc_flags |= SCF_SIGNATURE; 1348 #endif 1349 if (to->to_flags & TOF_SACKPERM) 1350 sc->sc_flags |= SCF_SACK; 1351 if (to->to_flags & TOF_MSS) 1352 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1353 if (ltflags & TF_NOOPT) 1354 sc->sc_flags |= SCF_NOOPT; 1355 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1356 sc->sc_flags |= SCF_ECN; 1357 1358 if (V_tcp_syncookies) 1359 sc->sc_iss = syncookie_generate(sch, sc); 1360 #ifdef INET6 1361 if (autoflowlabel) { 1362 if (V_tcp_syncookies) 1363 sc->sc_flowlabel = sc->sc_iss; 1364 else 1365 sc->sc_flowlabel = ip6_randomflowlabel(); 1366 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1367 } 1368 #endif 1369 SCH_UNLOCK(sch); 1370 1371 /* 1372 * Do a standard 3-way handshake. 1373 */ 1374 if (syncache_respond(sc) == 0) { 1375 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1376 syncache_free(sc); 1377 else if (sc != &scs) 1378 syncache_insert(sc, sch); /* locks and unlocks sch */ 1379 TCPSTAT_INC(tcps_sndacks); 1380 TCPSTAT_INC(tcps_sndtotal); 1381 } else { 1382 if (sc != &scs) 1383 syncache_free(sc); 1384 TCPSTAT_INC(tcps_sc_dropped); 1385 } 1386 1387 done: 1388 if (cred != NULL) 1389 crfree(cred); 1390 #ifdef MAC 1391 if (sc == &scs) 1392 mac_syncache_destroy(&maclabel); 1393 #endif 1394 if (m) { 1395 1396 *lsop = NULL; 1397 m_freem(m); 1398 } 1399 } 1400 1401 static int 1402 syncache_respond(struct syncache *sc) 1403 { 1404 struct ip *ip = NULL; 1405 struct mbuf *m; 1406 struct tcphdr *th = NULL; 1407 int optlen, error = 0; /* Make compiler happy */ 1408 u_int16_t hlen, tlen, mssopt; 1409 struct tcpopt to; 1410 #ifdef INET6 1411 struct ip6_hdr *ip6 = NULL; 1412 #endif 1413 1414 hlen = 1415 #ifdef INET6 1416 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1417 #endif 1418 sizeof(struct ip); 1419 tlen = hlen + sizeof(struct tcphdr); 1420 1421 /* Determine MSS we advertize to other end of connection. */ 1422 mssopt = tcp_mssopt(&sc->sc_inc); 1423 if (sc->sc_peer_mss) 1424 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss); 1425 1426 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1427 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1428 ("syncache: mbuf too small")); 1429 1430 /* Create the IP+TCP header from scratch. */ 1431 m = m_gethdr(M_NOWAIT, MT_DATA); 1432 if (m == NULL) 1433 return (ENOBUFS); 1434 #ifdef MAC 1435 mac_syncache_create_mbuf(sc->sc_label, m); 1436 #endif 1437 m->m_data += max_linkhdr; 1438 m->m_len = tlen; 1439 m->m_pkthdr.len = tlen; 1440 m->m_pkthdr.rcvif = NULL; 1441 1442 #ifdef INET6 1443 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1444 ip6 = mtod(m, struct ip6_hdr *); 1445 ip6->ip6_vfc = IPV6_VERSION; 1446 ip6->ip6_nxt = IPPROTO_TCP; 1447 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1448 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1449 ip6->ip6_plen = htons(tlen - hlen); 1450 /* ip6_hlim is set after checksum */ 1451 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1452 ip6->ip6_flow |= sc->sc_flowlabel; 1453 1454 th = (struct tcphdr *)(ip6 + 1); 1455 } 1456 #endif 1457 #if defined(INET6) && defined(INET) 1458 else 1459 #endif 1460 #ifdef INET 1461 { 1462 ip = mtod(m, struct ip *); 1463 ip->ip_v = IPVERSION; 1464 ip->ip_hl = sizeof(struct ip) >> 2; 1465 ip->ip_len = htons(tlen); 1466 ip->ip_id = 0; 1467 ip->ip_off = 0; 1468 ip->ip_sum = 0; 1469 ip->ip_p = IPPROTO_TCP; 1470 ip->ip_src = sc->sc_inc.inc_laddr; 1471 ip->ip_dst = sc->sc_inc.inc_faddr; 1472 ip->ip_ttl = sc->sc_ip_ttl; 1473 ip->ip_tos = sc->sc_ip_tos; 1474 1475 /* 1476 * See if we should do MTU discovery. Route lookups are 1477 * expensive, so we will only unset the DF bit if: 1478 * 1479 * 1) path_mtu_discovery is disabled 1480 * 2) the SCF_UNREACH flag has been set 1481 */ 1482 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1483 ip->ip_off |= htons(IP_DF); 1484 1485 th = (struct tcphdr *)(ip + 1); 1486 } 1487 #endif /* INET */ 1488 th->th_sport = sc->sc_inc.inc_lport; 1489 th->th_dport = sc->sc_inc.inc_fport; 1490 1491 th->th_seq = htonl(sc->sc_iss); 1492 th->th_ack = htonl(sc->sc_irs + 1); 1493 th->th_off = sizeof(struct tcphdr) >> 2; 1494 th->th_x2 = 0; 1495 th->th_flags = TH_SYN|TH_ACK; 1496 th->th_win = htons(sc->sc_wnd); 1497 th->th_urp = 0; 1498 1499 if (sc->sc_flags & SCF_ECN) { 1500 th->th_flags |= TH_ECE; 1501 TCPSTAT_INC(tcps_ecn_shs); 1502 } 1503 1504 /* Tack on the TCP options. */ 1505 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1506 to.to_flags = 0; 1507 1508 to.to_mss = mssopt; 1509 to.to_flags = TOF_MSS; 1510 if (sc->sc_flags & SCF_WINSCALE) { 1511 to.to_wscale = sc->sc_requested_r_scale; 1512 to.to_flags |= TOF_SCALE; 1513 } 1514 if (sc->sc_flags & SCF_TIMESTAMP) { 1515 /* Virgin timestamp or TCP cookie enhanced one. */ 1516 to.to_tsval = sc->sc_ts; 1517 to.to_tsecr = sc->sc_tsreflect; 1518 to.to_flags |= TOF_TS; 1519 } 1520 if (sc->sc_flags & SCF_SACK) 1521 to.to_flags |= TOF_SACKPERM; 1522 #ifdef TCP_SIGNATURE 1523 if (sc->sc_flags & SCF_SIGNATURE) 1524 to.to_flags |= TOF_SIGNATURE; 1525 #endif 1526 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1527 1528 /* Adjust headers by option size. */ 1529 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1530 m->m_len += optlen; 1531 m->m_pkthdr.len += optlen; 1532 1533 #ifdef TCP_SIGNATURE 1534 if (sc->sc_flags & SCF_SIGNATURE) 1535 tcp_signature_compute(m, 0, 0, optlen, 1536 to.to_signature, IPSEC_DIR_OUTBOUND); 1537 #endif 1538 #ifdef INET6 1539 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1540 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1541 else 1542 #endif 1543 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1544 } else 1545 optlen = 0; 1546 1547 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1548 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1549 #ifdef INET6 1550 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1551 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1552 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1553 IPPROTO_TCP, 0); 1554 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1555 #ifdef TCP_OFFLOAD 1556 if (ADDED_BY_TOE(sc)) { 1557 struct toedev *tod = sc->sc_tod; 1558 1559 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1560 1561 return (error); 1562 } 1563 #endif 1564 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1565 } 1566 #endif 1567 #if defined(INET6) && defined(INET) 1568 else 1569 #endif 1570 #ifdef INET 1571 { 1572 m->m_pkthdr.csum_flags = CSUM_TCP; 1573 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1574 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1575 #ifdef TCP_OFFLOAD 1576 if (ADDED_BY_TOE(sc)) { 1577 struct toedev *tod = sc->sc_tod; 1578 1579 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1580 1581 return (error); 1582 } 1583 #endif 1584 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1585 } 1586 #endif 1587 return (error); 1588 } 1589 1590 /* 1591 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1592 * that exceed the capacity of the syncache by avoiding the storage of any 1593 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1594 * attacks where the attacker does not have access to our responses. 1595 * 1596 * Syncookies encode and include all necessary information about the 1597 * connection setup within the SYN|ACK that we send back. That way we 1598 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1599 * (if ever). Normally the syncache and syncookies are running in parallel 1600 * with the latter taking over when the former is exhausted. When matching 1601 * syncache entry is found the syncookie is ignored. 1602 * 1603 * The only reliable information persisting the 3WHS is our inital sequence 1604 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1605 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1606 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1607 * returns and signifies a legitimate connection if it matches the ACK. 1608 * 1609 * The available space of 32 bits to store the hash and to encode the SYN 1610 * option information is very tight and we should have at least 24 bits for 1611 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1612 * 1613 * SYN option information we have to encode to fully restore a connection: 1614 * MSS: is imporant to chose an optimal segment size to avoid IP level 1615 * fragmentation along the path. The common MSS values can be encoded 1616 * in a 3-bit table. Uncommon values are captured by the next lower value 1617 * in the table leading to a slight increase in packetization overhead. 1618 * WSCALE: is necessary to allow large windows to be used for high delay- 1619 * bandwidth product links. Not scaling the window when it was initially 1620 * negotiated is bad for performance as lack of scaling further decreases 1621 * the apparent available send window. We only need to encode the WSCALE 1622 * we received from the remote end. Our end can be recalculated at any 1623 * time. The common WSCALE values can be encoded in a 3-bit table. 1624 * Uncommon values are captured by the next lower value in the table 1625 * making us under-estimate the available window size halving our 1626 * theoretically possible maximum throughput for that connection. 1627 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1628 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1629 * that are included in all segments on a connection. We enable them when 1630 * the ACK has them. 1631 * 1632 * Security of syncookies and attack vectors: 1633 * 1634 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1635 * together with the gloabl secret to make it unique per connection attempt. 1636 * Thus any change of any of those parameters results in a different MAC output 1637 * in an unpredictable way unless a collision is encountered. 24 bits of the 1638 * MAC are embedded into the ISS. 1639 * 1640 * To prevent replay attacks two rotating global secrets are updated with a 1641 * new random value every 15 seconds. The life-time of a syncookie is thus 1642 * 15-30 seconds. 1643 * 1644 * Vector 1: Attacking the secret. This requires finding a weakness in the 1645 * MAC itself or the way it is used here. The attacker can do a chosen plain 1646 * text attack by varying and testing the all parameters under his control. 1647 * The strength depends on the size and randomness of the secret, and the 1648 * cryptographic security of the MAC function. Due to the constant updating 1649 * of the secret the attacker has at most 29.999 seconds to find the secret 1650 * and launch spoofed connections. After that he has to start all over again. 1651 * 1652 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1653 * size an average of 4,823 attempts are required for a 50% chance of success 1654 * to spoof a single syncookie (birthday collision paradox). However the 1655 * attacker is blind and doesn't know if one of his attempts succeeded unless 1656 * he has a side channel to interfere success from. A single connection setup 1657 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1658 * This many attempts are required for each one blind spoofed connection. For 1659 * every additional spoofed connection he has to launch another N attempts. 1660 * Thus for a sustained rate 100 spoofed connections per second approximately 1661 * 1,800,000 packets per second would have to be sent. 1662 * 1663 * NB: The MAC function should be fast so that it doesn't become a CPU 1664 * exhaustion attack vector itself. 1665 * 1666 * References: 1667 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1668 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1669 * http://cr.yp.to/syncookies.html (overview) 1670 * http://cr.yp.to/syncookies/archive (details) 1671 * 1672 * 1673 * Schematic construction of a syncookie enabled Initial Sequence Number: 1674 * 0 1 2 3 1675 * 12345678901234567890123456789012 1676 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1677 * 1678 * x 24 MAC (truncated) 1679 * W 3 Send Window Scale index 1680 * M 3 MSS index 1681 * S 1 SACK permitted 1682 * P 1 Odd/even secret 1683 */ 1684 1685 /* 1686 * Distribution and probability of certain MSS values. Those in between are 1687 * rounded down to the next lower one. 1688 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1689 * .2% .3% 5% 7% 7% 20% 15% 45% 1690 */ 1691 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1692 1693 /* 1694 * Distribution and probability of certain WSCALE values. We have to map the 1695 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1696 * bits based on prevalence of certain values. Where we don't have an exact 1697 * match for are rounded down to the next lower one letting us under-estimate 1698 * the true available window. At the moment this would happen only for the 1699 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1700 * and window size). The absence of the WSCALE option (no scaling in either 1701 * direction) is encoded with index zero. 1702 * [WSCALE values histograms, Allman, 2012] 1703 * X 10 10 35 5 6 14 10% by host 1704 * X 11 4 5 5 18 49 3% by connections 1705 */ 1706 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1707 1708 /* 1709 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1710 * and good cryptographic properties. 1711 */ 1712 static uint32_t 1713 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1714 uint8_t *secbits, uintptr_t secmod) 1715 { 1716 SIPHASH_CTX ctx; 1717 uint32_t siphash[2]; 1718 1719 SipHash24_Init(&ctx); 1720 SipHash_SetKey(&ctx, secbits); 1721 switch (inc->inc_flags & INC_ISIPV6) { 1722 #ifdef INET 1723 case 0: 1724 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1725 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1726 break; 1727 #endif 1728 #ifdef INET6 1729 case INC_ISIPV6: 1730 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1731 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1732 break; 1733 #endif 1734 } 1735 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1736 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1737 SipHash_Update(&ctx, &flags, sizeof(flags)); 1738 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1739 SipHash_Final((u_int8_t *)&siphash, &ctx); 1740 1741 return (siphash[0] ^ siphash[1]); 1742 } 1743 1744 static tcp_seq 1745 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 1746 { 1747 u_int i, mss, secbit, wscale; 1748 uint32_t iss, hash; 1749 uint8_t *secbits; 1750 union syncookie cookie; 1751 1752 SCH_LOCK_ASSERT(sch); 1753 1754 cookie.cookie = 0; 1755 1756 /* Map our computed MSS into the 3-bit index. */ 1757 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss)); 1758 for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1; 1759 tcp_sc_msstab[i] > mss && i > 0; 1760 i--) 1761 ; 1762 cookie.flags.mss_idx = i; 1763 1764 /* 1765 * Map the send window scale into the 3-bit index but only if 1766 * the wscale option was received. 1767 */ 1768 if (sc->sc_flags & SCF_WINSCALE) { 1769 wscale = sc->sc_requested_s_scale; 1770 for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1; 1771 tcp_sc_wstab[i] > wscale && i > 0; 1772 i--) 1773 ; 1774 cookie.flags.wscale_idx = i; 1775 } 1776 1777 /* Can we do SACK? */ 1778 if (sc->sc_flags & SCF_SACK) 1779 cookie.flags.sack_ok = 1; 1780 1781 /* Which of the two secrets to use. */ 1782 secbit = sch->sch_sc->secret.oddeven & 0x1; 1783 cookie.flags.odd_even = secbit; 1784 1785 secbits = sch->sch_sc->secret.key[secbit]; 1786 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 1787 (uintptr_t)sch); 1788 1789 /* 1790 * Put the flags into the hash and XOR them to get better ISS number 1791 * variance. This doesn't enhance the cryptographic strength and is 1792 * done to prevent the 8 cookie bits from showing up directly on the 1793 * wire. 1794 */ 1795 iss = hash & ~0xff; 1796 iss |= cookie.cookie ^ (hash >> 24); 1797 1798 /* Randomize the timestamp. */ 1799 if (sc->sc_flags & SCF_TIMESTAMP) { 1800 sc->sc_ts = arc4random(); 1801 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 1802 } 1803 1804 TCPSTAT_INC(tcps_sc_sendcookie); 1805 return (iss); 1806 } 1807 1808 static struct syncache * 1809 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 1810 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 1811 struct socket *lso) 1812 { 1813 uint32_t hash; 1814 uint8_t *secbits; 1815 tcp_seq ack, seq; 1816 int wnd, wscale = 0; 1817 union syncookie cookie; 1818 1819 SCH_LOCK_ASSERT(sch); 1820 1821 /* 1822 * Pull information out of SYN-ACK/ACK and revert sequence number 1823 * advances. 1824 */ 1825 ack = th->th_ack - 1; 1826 seq = th->th_seq - 1; 1827 1828 /* 1829 * Unpack the flags containing enough information to restore the 1830 * connection. 1831 */ 1832 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 1833 1834 /* Which of the two secrets to use. */ 1835 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 1836 1837 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 1838 1839 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 1840 if ((ack & ~0xff) != (hash & ~0xff)) 1841 return (NULL); 1842 1843 /* Fill in the syncache values. */ 1844 sc->sc_flags = 0; 1845 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1846 sc->sc_ipopts = NULL; 1847 1848 sc->sc_irs = seq; 1849 sc->sc_iss = ack; 1850 1851 switch (inc->inc_flags & INC_ISIPV6) { 1852 #ifdef INET 1853 case 0: 1854 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 1855 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 1856 break; 1857 #endif 1858 #ifdef INET6 1859 case INC_ISIPV6: 1860 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 1861 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 1862 break; 1863 #endif 1864 } 1865 1866 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 1867 1868 /* We can simply recompute receive window scale we sent earlier. */ 1869 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 1870 wscale++; 1871 1872 /* Only use wscale if it was enabled in the orignal SYN. */ 1873 if (cookie.flags.wscale_idx > 0) { 1874 sc->sc_requested_r_scale = wscale; 1875 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 1876 sc->sc_flags |= SCF_WINSCALE; 1877 } 1878 1879 wnd = sbspace(&lso->so_rcv); 1880 wnd = imax(wnd, 0); 1881 wnd = imin(wnd, TCP_MAXWIN); 1882 sc->sc_wnd = wnd; 1883 1884 if (cookie.flags.sack_ok) 1885 sc->sc_flags |= SCF_SACK; 1886 1887 if (to->to_flags & TOF_TS) { 1888 sc->sc_flags |= SCF_TIMESTAMP; 1889 sc->sc_tsreflect = to->to_tsval; 1890 sc->sc_ts = to->to_tsecr; 1891 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 1892 } 1893 1894 if (to->to_flags & TOF_SIGNATURE) 1895 sc->sc_flags |= SCF_SIGNATURE; 1896 1897 sc->sc_rxmits = 0; 1898 1899 TCPSTAT_INC(tcps_sc_recvcookie); 1900 return (sc); 1901 } 1902 1903 #ifdef INVARIANTS 1904 static int 1905 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 1906 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 1907 struct socket *lso) 1908 { 1909 struct syncache scs, *scx; 1910 char *s; 1911 1912 bzero(&scs, sizeof(scs)); 1913 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 1914 1915 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 1916 return (0); 1917 1918 if (scx != NULL) { 1919 if (sc->sc_peer_mss != scx->sc_peer_mss) 1920 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 1921 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 1922 1923 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 1924 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 1925 s, __func__, sc->sc_requested_r_scale, 1926 scx->sc_requested_r_scale); 1927 1928 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 1929 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 1930 s, __func__, sc->sc_requested_s_scale, 1931 scx->sc_requested_s_scale); 1932 1933 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 1934 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 1935 } 1936 1937 if (s != NULL) 1938 free(s, M_TCPLOG); 1939 return (0); 1940 } 1941 #endif /* INVARIANTS */ 1942 1943 static void 1944 syncookie_reseed(void *arg) 1945 { 1946 struct tcp_syncache *sc = arg; 1947 uint8_t *secbits; 1948 int secbit; 1949 1950 /* 1951 * Reseeding the secret doesn't have to be protected by a lock. 1952 * It only must be ensured that the new random values are visible 1953 * to all CPUs in a SMP environment. The atomic with release 1954 * semantics ensures that. 1955 */ 1956 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 1957 secbits = sc->secret.key[secbit]; 1958 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 1959 atomic_add_rel_int(&sc->secret.oddeven, 1); 1960 1961 /* Reschedule ourself. */ 1962 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 1963 } 1964 1965 /* 1966 * Returns the current number of syncache entries. This number 1967 * will probably change before you get around to calling 1968 * syncache_pcblist. 1969 */ 1970 int 1971 syncache_pcbcount(void) 1972 { 1973 struct syncache_head *sch; 1974 int count, i; 1975 1976 for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 1977 /* No need to lock for a read. */ 1978 sch = &V_tcp_syncache.hashbase[i]; 1979 count += sch->sch_length; 1980 } 1981 return count; 1982 } 1983 1984 /* 1985 * Exports the syncache entries to userland so that netstat can display 1986 * them alongside the other sockets. This function is intended to be 1987 * called only from tcp_pcblist. 1988 * 1989 * Due to concurrency on an active system, the number of pcbs exported 1990 * may have no relation to max_pcbs. max_pcbs merely indicates the 1991 * amount of space the caller allocated for this function to use. 1992 */ 1993 int 1994 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 1995 { 1996 struct xtcpcb xt; 1997 struct syncache *sc; 1998 struct syncache_head *sch; 1999 int count, error, i; 2000 2001 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2002 sch = &V_tcp_syncache.hashbase[i]; 2003 SCH_LOCK(sch); 2004 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2005 if (count >= max_pcbs) { 2006 SCH_UNLOCK(sch); 2007 goto exit; 2008 } 2009 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2010 continue; 2011 bzero(&xt, sizeof(xt)); 2012 xt.xt_len = sizeof(xt); 2013 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2014 xt.xt_inp.inp_vflag = INP_IPV6; 2015 else 2016 xt.xt_inp.inp_vflag = INP_IPV4; 2017 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 2018 xt.xt_tp.t_inpcb = &xt.xt_inp; 2019 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 2020 xt.xt_socket.xso_protocol = IPPROTO_TCP; 2021 xt.xt_socket.xso_len = sizeof (struct xsocket); 2022 xt.xt_socket.so_type = SOCK_STREAM; 2023 xt.xt_socket.so_state = SS_ISCONNECTING; 2024 error = SYSCTL_OUT(req, &xt, sizeof xt); 2025 if (error) { 2026 SCH_UNLOCK(sch); 2027 goto exit; 2028 } 2029 count++; 2030 } 2031 SCH_UNLOCK(sch); 2032 } 2033 exit: 2034 *pcbs_exported = count; 2035 return error; 2036 } 2037