1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2001 McAfee, Inc. 5 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG 6 * All rights reserved. 7 * 8 * This software was developed for the FreeBSD Project by Jonathan Lemon 9 * and McAfee Research, the Security Research Division of McAfee, Inc. under 10 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 11 * DARPA CHATS research program. [2001 McAfee, Inc.] 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include "opt_inet.h" 39 #include "opt_inet6.h" 40 #include "opt_ipsec.h" 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/hash.h> 45 #include <sys/refcount.h> 46 #include <sys/kernel.h> 47 #include <sys/sysctl.h> 48 #include <sys/limits.h> 49 #include <sys/lock.h> 50 #include <sys/mutex.h> 51 #include <sys/malloc.h> 52 #include <sys/mbuf.h> 53 #include <sys/proc.h> /* for proc0 declaration */ 54 #include <sys/random.h> 55 #include <sys/socket.h> 56 #include <sys/socketvar.h> 57 #include <sys/syslog.h> 58 #include <sys/ucred.h> 59 60 #include <sys/md5.h> 61 #include <crypto/siphash/siphash.h> 62 63 #include <vm/uma.h> 64 65 #include <net/if.h> 66 #include <net/if_var.h> 67 #include <net/route.h> 68 #include <net/vnet.h> 69 70 #include <netinet/in.h> 71 #include <netinet/in_kdtrace.h> 72 #include <netinet/in_systm.h> 73 #include <netinet/ip.h> 74 #include <netinet/in_var.h> 75 #include <netinet/in_pcb.h> 76 #include <netinet/ip_var.h> 77 #include <netinet/ip_options.h> 78 #ifdef INET6 79 #include <netinet/ip6.h> 80 #include <netinet/icmp6.h> 81 #include <netinet6/nd6.h> 82 #include <netinet6/ip6_var.h> 83 #include <netinet6/in6_pcb.h> 84 #endif 85 #include <netinet/tcp.h> 86 #include <netinet/tcp_fastopen.h> 87 #include <netinet/tcp_fsm.h> 88 #include <netinet/tcp_seq.h> 89 #include <netinet/tcp_timer.h> 90 #include <netinet/tcp_var.h> 91 #include <netinet/tcp_syncache.h> 92 #include <netinet/tcp_ecn.h> 93 #ifdef INET6 94 #include <netinet6/tcp6_var.h> 95 #endif 96 #ifdef TCP_OFFLOAD 97 #include <netinet/toecore.h> 98 #endif 99 #include <netinet/udp.h> 100 #include <netinet/udp_var.h> 101 102 #include <netipsec/ipsec_support.h> 103 104 #include <machine/in_cksum.h> 105 106 #include <security/mac/mac_framework.h> 107 108 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1; 109 #define V_tcp_syncookies VNET(tcp_syncookies) 110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 111 &VNET_NAME(tcp_syncookies), 0, 112 "Use TCP SYN cookies if the syncache overflows"); 113 114 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0; 115 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 117 &VNET_NAME(tcp_syncookiesonly), 0, 118 "Use only TCP SYN cookies"); 119 120 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1; 121 #define V_functions_inherit_listen_socket_stack \ 122 VNET(functions_inherit_listen_socket_stack) 123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack, 124 CTLFLAG_VNET | CTLFLAG_RW, 125 &VNET_NAME(functions_inherit_listen_socket_stack), 0, 126 "Inherit listen socket's stack"); 127 128 #ifdef TCP_OFFLOAD 129 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 130 #endif 131 132 static void syncache_drop(struct syncache *, struct syncache_head *); 133 static void syncache_free(struct syncache *); 134 static void syncache_insert(struct syncache *, struct syncache_head *); 135 static int syncache_respond(struct syncache *, const struct mbuf *, int); 136 static struct socket *syncache_socket(struct syncache *, struct socket *, 137 struct mbuf *m); 138 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 139 int docallout); 140 static void syncache_timer(void *); 141 142 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 143 uint8_t *, uintptr_t); 144 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 145 static struct syncache 146 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 147 struct syncache *, struct tcphdr *, struct tcpopt *, 148 struct socket *, uint16_t); 149 static void syncache_pause(struct in_conninfo *); 150 static void syncache_unpause(void *); 151 static void syncookie_reseed(void *); 152 #ifdef INVARIANTS 153 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 154 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 155 struct socket *lso, uint16_t port); 156 #endif 157 158 /* 159 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 160 * 3 retransmits corresponds to a timeout with default values of 161 * tcp_rexmit_initial * ( 1 + 162 * tcp_backoff[1] + 163 * tcp_backoff[2] + 164 * tcp_backoff[3]) + 3 * tcp_rexmit_slop, 165 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms, 166 * the odds are that the user has given up attempting to connect by then. 167 */ 168 #define SYNCACHE_MAXREXMTS 3 169 170 /* Arbitrary values */ 171 #define TCP_SYNCACHE_HASHSIZE 512 172 #define TCP_SYNCACHE_BUCKETLIMIT 30 173 174 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 175 #define V_tcp_syncache VNET(tcp_syncache) 176 177 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, 178 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 179 "TCP SYN cache"); 180 181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 182 &VNET_NAME(tcp_syncache.bucket_limit), 0, 183 "Per-bucket hash limit for syncache"); 184 185 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 186 &VNET_NAME(tcp_syncache.cache_limit), 0, 187 "Overall entry limit for syncache"); 188 189 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 190 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 191 192 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 193 &VNET_NAME(tcp_syncache.hashsize), 0, 194 "Size of TCP syncache hashtable"); 195 196 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET | 197 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0, 198 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) " 199 "and mac(4) checks"); 200 201 static int 202 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 203 { 204 int error; 205 u_int new; 206 207 new = V_tcp_syncache.rexmt_limit; 208 error = sysctl_handle_int(oidp, &new, 0, req); 209 if ((error == 0) && (req->newptr != NULL)) { 210 if (new > TCP_MAXRXTSHIFT) 211 error = EINVAL; 212 else 213 V_tcp_syncache.rexmt_limit = new; 214 } 215 return (error); 216 } 217 218 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 219 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 220 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 221 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI", 222 "Limit on SYN/ACK retransmissions"); 223 224 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 225 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 226 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 227 "Send reset on socket allocation failure"); 228 229 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 230 231 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 232 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 233 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 234 235 /* 236 * Requires the syncache entry to be already removed from the bucket list. 237 */ 238 static void 239 syncache_free(struct syncache *sc) 240 { 241 242 if (sc->sc_ipopts) 243 (void) m_free(sc->sc_ipopts); 244 if (sc->sc_cred) 245 crfree(sc->sc_cred); 246 #ifdef MAC 247 mac_syncache_destroy(&sc->sc_label); 248 #endif 249 250 uma_zfree(V_tcp_syncache.zone, sc); 251 } 252 253 void 254 syncache_init(void) 255 { 256 int i; 257 258 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 259 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 260 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 261 V_tcp_syncache.hash_secret = arc4random(); 262 263 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 264 &V_tcp_syncache.hashsize); 265 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 266 &V_tcp_syncache.bucket_limit); 267 if (!powerof2(V_tcp_syncache.hashsize) || 268 V_tcp_syncache.hashsize == 0) { 269 printf("WARNING: syncache hash size is not a power of 2.\n"); 270 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 271 } 272 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 273 274 /* Set limits. */ 275 V_tcp_syncache.cache_limit = 276 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 277 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 278 &V_tcp_syncache.cache_limit); 279 280 /* Allocate the hash table. */ 281 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 282 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 283 284 #ifdef VIMAGE 285 V_tcp_syncache.vnet = curvnet; 286 #endif 287 288 /* Initialize the hash buckets. */ 289 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 290 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 291 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 292 NULL, MTX_DEF); 293 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 294 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 295 V_tcp_syncache.hashbase[i].sch_length = 0; 296 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 297 V_tcp_syncache.hashbase[i].sch_last_overflow = 298 -(SYNCOOKIE_LIFETIME + 1); 299 } 300 301 /* Create the syncache entry zone. */ 302 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 303 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 304 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 305 V_tcp_syncache.cache_limit); 306 307 /* Start the SYN cookie reseeder callout. */ 308 callout_init(&V_tcp_syncache.secret.reseed, 1); 309 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 310 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 311 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 312 syncookie_reseed, &V_tcp_syncache); 313 314 /* Initialize the pause machinery. */ 315 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF); 316 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx, 317 0); 318 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME; 319 V_tcp_syncache.pause_backoff = 0; 320 V_tcp_syncache.paused = false; 321 } 322 323 #ifdef VIMAGE 324 void 325 syncache_destroy(void) 326 { 327 struct syncache_head *sch; 328 struct syncache *sc, *nsc; 329 int i; 330 331 /* 332 * Stop the re-seed timer before freeing resources. No need to 333 * possibly schedule it another time. 334 */ 335 callout_drain(&V_tcp_syncache.secret.reseed); 336 337 /* Stop the SYN cache pause callout. */ 338 mtx_lock(&V_tcp_syncache.pause_mtx); 339 if (callout_stop(&V_tcp_syncache.pause_co) == 0) { 340 mtx_unlock(&V_tcp_syncache.pause_mtx); 341 callout_drain(&V_tcp_syncache.pause_co); 342 } else 343 mtx_unlock(&V_tcp_syncache.pause_mtx); 344 345 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 346 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 347 sch = &V_tcp_syncache.hashbase[i]; 348 callout_drain(&sch->sch_timer); 349 350 SCH_LOCK(sch); 351 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 352 syncache_drop(sc, sch); 353 SCH_UNLOCK(sch); 354 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 355 ("%s: sch->sch_bucket not empty", __func__)); 356 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 357 __func__, sch->sch_length)); 358 mtx_destroy(&sch->sch_mtx); 359 } 360 361 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 362 ("%s: cache_count not 0", __func__)); 363 364 /* Free the allocated global resources. */ 365 uma_zdestroy(V_tcp_syncache.zone); 366 free(V_tcp_syncache.hashbase, M_SYNCACHE); 367 mtx_destroy(&V_tcp_syncache.pause_mtx); 368 } 369 #endif 370 371 /* 372 * Inserts a syncache entry into the specified bucket row. 373 * Locks and unlocks the syncache_head autonomously. 374 */ 375 static void 376 syncache_insert(struct syncache *sc, struct syncache_head *sch) 377 { 378 struct syncache *sc2; 379 380 SCH_LOCK(sch); 381 382 /* 383 * Make sure that we don't overflow the per-bucket limit. 384 * If the bucket is full, toss the oldest element. 385 */ 386 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 387 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 388 ("sch->sch_length incorrect")); 389 syncache_pause(&sc->sc_inc); 390 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 391 sch->sch_last_overflow = time_uptime; 392 syncache_drop(sc2, sch); 393 } 394 395 /* Put it into the bucket. */ 396 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 397 sch->sch_length++; 398 399 #ifdef TCP_OFFLOAD 400 if (ADDED_BY_TOE(sc)) { 401 struct toedev *tod = sc->sc_tod; 402 403 tod->tod_syncache_added(tod, sc->sc_todctx); 404 } 405 #endif 406 407 /* Reinitialize the bucket row's timer. */ 408 if (sch->sch_length == 1) 409 sch->sch_nextc = ticks + INT_MAX; 410 syncache_timeout(sc, sch, 1); 411 412 SCH_UNLOCK(sch); 413 414 TCPSTATES_INC(TCPS_SYN_RECEIVED); 415 TCPSTAT_INC(tcps_sc_added); 416 } 417 418 /* 419 * Remove and free entry from syncache bucket row. 420 * Expects locked syncache head. 421 */ 422 static void 423 syncache_drop(struct syncache *sc, struct syncache_head *sch) 424 { 425 426 SCH_LOCK_ASSERT(sch); 427 428 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 429 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 430 sch->sch_length--; 431 432 #ifdef TCP_OFFLOAD 433 if (ADDED_BY_TOE(sc)) { 434 struct toedev *tod = sc->sc_tod; 435 436 tod->tod_syncache_removed(tod, sc->sc_todctx); 437 } 438 #endif 439 440 syncache_free(sc); 441 } 442 443 /* 444 * Engage/reengage time on bucket row. 445 */ 446 static void 447 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 448 { 449 int rexmt; 450 451 if (sc->sc_rxmits == 0) 452 rexmt = tcp_rexmit_initial; 453 else 454 TCPT_RANGESET(rexmt, 455 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], 456 tcp_rexmit_min, TCPTV_REXMTMAX); 457 sc->sc_rxttime = ticks + rexmt; 458 sc->sc_rxmits++; 459 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 460 sch->sch_nextc = sc->sc_rxttime; 461 if (docallout) 462 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 463 syncache_timer, (void *)sch); 464 } 465 } 466 467 /* 468 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 469 * If we have retransmitted an entry the maximum number of times, expire it. 470 * One separate timer for each bucket row. 471 */ 472 static void 473 syncache_timer(void *xsch) 474 { 475 struct syncache_head *sch = (struct syncache_head *)xsch; 476 struct syncache *sc, *nsc; 477 struct epoch_tracker et; 478 int tick = ticks; 479 char *s; 480 bool paused; 481 482 CURVNET_SET(sch->sch_sc->vnet); 483 484 /* NB: syncache_head has already been locked by the callout. */ 485 SCH_LOCK_ASSERT(sch); 486 487 /* 488 * In the following cycle we may remove some entries and/or 489 * advance some timeouts, so re-initialize the bucket timer. 490 */ 491 sch->sch_nextc = tick + INT_MAX; 492 493 /* 494 * If we have paused processing, unconditionally remove 495 * all syncache entries. 496 */ 497 mtx_lock(&V_tcp_syncache.pause_mtx); 498 paused = V_tcp_syncache.paused; 499 mtx_unlock(&V_tcp_syncache.pause_mtx); 500 501 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 502 if (paused) { 503 syncache_drop(sc, sch); 504 continue; 505 } 506 /* 507 * We do not check if the listen socket still exists 508 * and accept the case where the listen socket may be 509 * gone by the time we resend the SYN/ACK. We do 510 * not expect this to happens often. If it does, 511 * then the RST will be sent by the time the remote 512 * host does the SYN/ACK->ACK. 513 */ 514 if (TSTMP_GT(sc->sc_rxttime, tick)) { 515 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 516 sch->sch_nextc = sc->sc_rxttime; 517 continue; 518 } 519 if (sc->sc_rxmits > V_tcp_ecn_maxretries) { 520 sc->sc_flags &= ~SCF_ECN; 521 } 522 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 523 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 524 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 525 "giving up and removing syncache entry\n", 526 s, __func__); 527 free(s, M_TCPLOG); 528 } 529 syncache_drop(sc, sch); 530 TCPSTAT_INC(tcps_sc_stale); 531 continue; 532 } 533 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 534 log(LOG_DEBUG, "%s; %s: Response timeout, " 535 "retransmitting (%u) SYN|ACK\n", 536 s, __func__, sc->sc_rxmits); 537 free(s, M_TCPLOG); 538 } 539 540 NET_EPOCH_ENTER(et); 541 syncache_respond(sc, NULL, TH_SYN|TH_ACK); 542 NET_EPOCH_EXIT(et); 543 TCPSTAT_INC(tcps_sc_retransmitted); 544 syncache_timeout(sc, sch, 0); 545 } 546 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 547 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 548 syncache_timer, (void *)(sch)); 549 CURVNET_RESTORE(); 550 } 551 552 /* 553 * Returns true if the system is only using cookies at the moment. 554 * This could be due to a sysadmin decision to only use cookies, or it 555 * could be due to the system detecting an attack. 556 */ 557 static inline bool 558 syncache_cookiesonly(void) 559 { 560 561 return (V_tcp_syncookies && (V_tcp_syncache.paused || 562 V_tcp_syncookiesonly)); 563 } 564 565 /* 566 * Find the hash bucket for the given connection. 567 */ 568 static struct syncache_head * 569 syncache_hashbucket(struct in_conninfo *inc) 570 { 571 uint32_t hash; 572 573 /* 574 * The hash is built on foreign port + local port + foreign address. 575 * We rely on the fact that struct in_conninfo starts with 16 bits 576 * of foreign port, then 16 bits of local port then followed by 128 577 * bits of foreign address. In case of IPv4 address, the first 3 578 * 32-bit words of the address always are zeroes. 579 */ 580 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 581 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 582 583 return (&V_tcp_syncache.hashbase[hash]); 584 } 585 586 /* 587 * Find an entry in the syncache. 588 * Returns always with locked syncache_head plus a matching entry or NULL. 589 */ 590 static struct syncache * 591 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 592 { 593 struct syncache *sc; 594 struct syncache_head *sch; 595 596 *schp = sch = syncache_hashbucket(inc); 597 SCH_LOCK(sch); 598 599 /* Circle through bucket row to find matching entry. */ 600 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 601 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 602 sizeof(struct in_endpoints)) == 0) 603 break; 604 605 return (sc); /* Always returns with locked sch. */ 606 } 607 608 /* 609 * This function is called when we get a RST for a 610 * non-existent connection, so that we can see if the 611 * connection is in the syn cache. If it is, zap it. 612 * If required send a challenge ACK. 613 */ 614 void 615 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m, 616 uint16_t port) 617 { 618 struct syncache *sc; 619 struct syncache_head *sch; 620 char *s = NULL; 621 622 if (syncache_cookiesonly()) 623 return; 624 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 625 SCH_LOCK_ASSERT(sch); 626 627 /* 628 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 629 * See RFC 793 page 65, section SEGMENT ARRIVES. 630 */ 631 if (tcp_get_flags(th) & (TH_ACK|TH_SYN|TH_FIN)) { 632 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 633 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 634 "FIN flag set, segment ignored\n", s, __func__); 635 TCPSTAT_INC(tcps_badrst); 636 goto done; 637 } 638 639 /* 640 * No corresponding connection was found in syncache. 641 * If syncookies are enabled and possibly exclusively 642 * used, or we are under memory pressure, a valid RST 643 * may not find a syncache entry. In that case we're 644 * done and no SYN|ACK retransmissions will happen. 645 * Otherwise the RST was misdirected or spoofed. 646 */ 647 if (sc == NULL) { 648 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 649 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 650 "syncache entry (possibly syncookie only), " 651 "segment ignored\n", s, __func__); 652 TCPSTAT_INC(tcps_badrst); 653 goto done; 654 } 655 656 /* The remote UDP encaps port does not match. */ 657 if (sc->sc_port != port) { 658 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 659 log(LOG_DEBUG, "%s; %s: Spurious RST with matching " 660 "syncache entry but non-matching UDP encaps port, " 661 "segment ignored\n", s, __func__); 662 TCPSTAT_INC(tcps_badrst); 663 goto done; 664 } 665 666 /* 667 * If the RST bit is set, check the sequence number to see 668 * if this is a valid reset segment. 669 * 670 * RFC 793 page 37: 671 * In all states except SYN-SENT, all reset (RST) segments 672 * are validated by checking their SEQ-fields. A reset is 673 * valid if its sequence number is in the window. 674 * 675 * RFC 793 page 69: 676 * There are four cases for the acceptability test for an incoming 677 * segment: 678 * 679 * Segment Receive Test 680 * Length Window 681 * ------- ------- ------------------------------------------- 682 * 0 0 SEG.SEQ = RCV.NXT 683 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 684 * >0 0 not acceptable 685 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 686 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 687 * 688 * Note that when receiving a SYN segment in the LISTEN state, 689 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 690 * described in RFC 793, page 66. 691 */ 692 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 693 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 694 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 695 if (V_tcp_insecure_rst || 696 th->th_seq == sc->sc_irs + 1) { 697 syncache_drop(sc, sch); 698 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 699 log(LOG_DEBUG, 700 "%s; %s: Our SYN|ACK was rejected, " 701 "connection attempt aborted by remote " 702 "endpoint\n", 703 s, __func__); 704 TCPSTAT_INC(tcps_sc_reset); 705 } else { 706 TCPSTAT_INC(tcps_badrst); 707 /* Send challenge ACK. */ 708 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 709 log(LOG_DEBUG, "%s; %s: RST with invalid " 710 " SEQ %u != NXT %u (+WND %u), " 711 "sending challenge ACK\n", 712 s, __func__, 713 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 714 syncache_respond(sc, m, TH_ACK); 715 } 716 } else { 717 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 718 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 719 "NXT %u (+WND %u), segment ignored\n", 720 s, __func__, 721 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 722 TCPSTAT_INC(tcps_badrst); 723 } 724 725 done: 726 if (s != NULL) 727 free(s, M_TCPLOG); 728 SCH_UNLOCK(sch); 729 } 730 731 void 732 syncache_badack(struct in_conninfo *inc, uint16_t port) 733 { 734 struct syncache *sc; 735 struct syncache_head *sch; 736 737 if (syncache_cookiesonly()) 738 return; 739 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 740 SCH_LOCK_ASSERT(sch); 741 if ((sc != NULL) && (sc->sc_port == port)) { 742 syncache_drop(sc, sch); 743 TCPSTAT_INC(tcps_sc_badack); 744 } 745 SCH_UNLOCK(sch); 746 } 747 748 void 749 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port) 750 { 751 struct syncache *sc; 752 struct syncache_head *sch; 753 754 if (syncache_cookiesonly()) 755 return; 756 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 757 SCH_LOCK_ASSERT(sch); 758 if (sc == NULL) 759 goto done; 760 761 /* If the port != sc_port, then it's a bogus ICMP msg */ 762 if (port != sc->sc_port) 763 goto done; 764 765 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 766 if (ntohl(th_seq) != sc->sc_iss) 767 goto done; 768 769 /* 770 * If we've rertransmitted 3 times and this is our second error, 771 * we remove the entry. Otherwise, we allow it to continue on. 772 * This prevents us from incorrectly nuking an entry during a 773 * spurious network outage. 774 * 775 * See tcp_notify(). 776 */ 777 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 778 sc->sc_flags |= SCF_UNREACH; 779 goto done; 780 } 781 syncache_drop(sc, sch); 782 TCPSTAT_INC(tcps_sc_unreach); 783 done: 784 SCH_UNLOCK(sch); 785 } 786 787 /* 788 * Build a new TCP socket structure from a syncache entry. 789 * 790 * On success return the newly created socket with its underlying inp locked. 791 */ 792 static struct socket * 793 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 794 { 795 struct tcp_function_block *blk; 796 struct inpcb *inp = NULL; 797 struct socket *so; 798 struct tcpcb *tp; 799 int error; 800 char *s; 801 802 NET_EPOCH_ASSERT(); 803 804 /* 805 * Ok, create the full blown connection, and set things up 806 * as they would have been set up if we had created the 807 * connection when the SYN arrived. 808 */ 809 if ((so = solisten_clone(lso)) == NULL) 810 goto allocfail; 811 #ifdef MAC 812 mac_socketpeer_set_from_mbuf(m, so); 813 #endif 814 error = in_pcballoc(so, &V_tcbinfo); 815 if (error) { 816 sodealloc(so); 817 goto allocfail; 818 } 819 inp = sotoinpcb(so); 820 if ((tp = tcp_newtcpcb(inp)) == NULL) { 821 in_pcbdetach(inp); 822 in_pcbfree(inp); 823 sodealloc(so); 824 goto allocfail; 825 } 826 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 827 #ifdef INET6 828 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 829 inp->inp_vflag &= ~INP_IPV4; 830 inp->inp_vflag |= INP_IPV6; 831 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 832 } else { 833 inp->inp_vflag &= ~INP_IPV6; 834 inp->inp_vflag |= INP_IPV4; 835 #endif 836 inp->inp_ip_ttl = sc->sc_ip_ttl; 837 inp->inp_ip_tos = sc->sc_ip_tos; 838 inp->inp_laddr = sc->sc_inc.inc_laddr; 839 #ifdef INET6 840 } 841 #endif 842 843 /* 844 * If there's an mbuf and it has a flowid, then let's initialise the 845 * inp with that particular flowid. 846 */ 847 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 848 inp->inp_flowid = m->m_pkthdr.flowid; 849 inp->inp_flowtype = M_HASHTYPE_GET(m); 850 #ifdef NUMA 851 inp->inp_numa_domain = m->m_pkthdr.numa_domain; 852 #endif 853 } 854 855 inp->inp_lport = sc->sc_inc.inc_lport; 856 #ifdef INET6 857 if (inp->inp_vflag & INP_IPV6PROTO) { 858 struct inpcb *oinp = sotoinpcb(lso); 859 860 /* 861 * Inherit socket options from the listening socket. 862 * Note that in6p_inputopts are not (and should not be) 863 * copied, since it stores previously received options and is 864 * used to detect if each new option is different than the 865 * previous one and hence should be passed to a user. 866 * If we copied in6p_inputopts, a user would not be able to 867 * receive options just after calling the accept system call. 868 */ 869 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 870 if (oinp->in6p_outputopts) 871 inp->in6p_outputopts = 872 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 873 inp->in6p_hops = oinp->in6p_hops; 874 } 875 876 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 877 struct in6_addr laddr6; 878 struct sockaddr_in6 sin6; 879 880 sin6.sin6_family = AF_INET6; 881 sin6.sin6_len = sizeof(sin6); 882 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 883 sin6.sin6_port = sc->sc_inc.inc_fport; 884 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 885 laddr6 = inp->in6p_laddr; 886 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 887 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 888 INP_HASH_WLOCK(&V_tcbinfo); 889 error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 890 thread0.td_ucred, m, false); 891 INP_HASH_WUNLOCK(&V_tcbinfo); 892 if (error != 0) { 893 inp->in6p_laddr = laddr6; 894 goto abort; 895 } 896 /* Override flowlabel from in6_pcbconnect. */ 897 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 898 inp->inp_flow |= sc->sc_flowlabel; 899 } 900 #endif /* INET6 */ 901 #if defined(INET) && defined(INET6) 902 else 903 #endif 904 #ifdef INET 905 { 906 struct in_addr laddr; 907 struct sockaddr_in sin; 908 909 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 910 911 if (inp->inp_options == NULL) { 912 inp->inp_options = sc->sc_ipopts; 913 sc->sc_ipopts = NULL; 914 } 915 916 sin.sin_family = AF_INET; 917 sin.sin_len = sizeof(sin); 918 sin.sin_addr = sc->sc_inc.inc_faddr; 919 sin.sin_port = sc->sc_inc.inc_fport; 920 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 921 laddr = inp->inp_laddr; 922 if (inp->inp_laddr.s_addr == INADDR_ANY) 923 inp->inp_laddr = sc->sc_inc.inc_laddr; 924 INP_HASH_WLOCK(&V_tcbinfo); 925 error = in_pcbconnect(inp, (struct sockaddr *)&sin, 926 thread0.td_ucred, false); 927 INP_HASH_WUNLOCK(&V_tcbinfo); 928 if (error != 0) { 929 inp->inp_laddr = laddr; 930 goto abort; 931 } 932 } 933 #endif /* INET */ 934 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 935 /* Copy old policy into new socket's. */ 936 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 937 printf("syncache_socket: could not copy policy\n"); 938 #endif 939 tp->t_state = TCPS_SYN_RECEIVED; 940 tp->iss = sc->sc_iss; 941 tp->irs = sc->sc_irs; 942 tp->t_port = sc->sc_port; 943 tcp_rcvseqinit(tp); 944 tcp_sendseqinit(tp); 945 blk = sototcpcb(lso)->t_fb; 946 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 947 /* 948 * Our parents t_fb was not the default, 949 * we need to release our ref on tp->t_fb and 950 * pickup one on the new entry. 951 */ 952 struct tcp_function_block *rblk; 953 954 rblk = find_and_ref_tcp_fb(blk); 955 KASSERT(rblk != NULL, 956 ("cannot find blk %p out of syncache?", blk)); 957 if (tp->t_fb->tfb_tcp_fb_fini) 958 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 959 refcount_release(&tp->t_fb->tfb_refcnt); 960 tp->t_fb = rblk; 961 /* 962 * XXXrrs this is quite dangerous, it is possible 963 * for the new function to fail to init. We also 964 * are not asking if the handoff_is_ok though at 965 * the very start thats probalbly ok. 966 */ 967 if (tp->t_fb->tfb_tcp_fb_init) { 968 (*tp->t_fb->tfb_tcp_fb_init)(tp); 969 } 970 } 971 tp->snd_wl1 = sc->sc_irs; 972 tp->snd_max = tp->iss + 1; 973 tp->snd_nxt = tp->iss + 1; 974 tp->rcv_up = sc->sc_irs + 1; 975 tp->rcv_wnd = sc->sc_wnd; 976 tp->rcv_adv += tp->rcv_wnd; 977 tp->last_ack_sent = tp->rcv_nxt; 978 979 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 980 if (sc->sc_flags & SCF_NOOPT) 981 tp->t_flags |= TF_NOOPT; 982 else { 983 if (sc->sc_flags & SCF_WINSCALE) { 984 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 985 tp->snd_scale = sc->sc_requested_s_scale; 986 tp->request_r_scale = sc->sc_requested_r_scale; 987 } 988 if (sc->sc_flags & SCF_TIMESTAMP) { 989 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 990 tp->ts_recent = sc->sc_tsreflect; 991 tp->ts_recent_age = tcp_ts_getticks(); 992 tp->ts_offset = sc->sc_tsoff; 993 } 994 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 995 if (sc->sc_flags & SCF_SIGNATURE) 996 tp->t_flags |= TF_SIGNATURE; 997 #endif 998 if (sc->sc_flags & SCF_SACK) 999 tp->t_flags |= TF_SACK_PERMIT; 1000 } 1001 1002 tcp_ecn_syncache_socket(tp, sc); 1003 1004 /* 1005 * Set up MSS and get cached values from tcp_hostcache. 1006 * This might overwrite some of the defaults we just set. 1007 */ 1008 tcp_mss(tp, sc->sc_peer_mss); 1009 1010 /* 1011 * If the SYN,ACK was retransmitted, indicate that CWND to be 1012 * limited to one segment in cc_conn_init(). 1013 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 1014 */ 1015 if (sc->sc_rxmits > 1) 1016 tp->snd_cwnd = 1; 1017 1018 #ifdef TCP_OFFLOAD 1019 /* 1020 * Allow a TOE driver to install its hooks. Note that we hold the 1021 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 1022 * new connection before the TOE driver has done its thing. 1023 */ 1024 if (ADDED_BY_TOE(sc)) { 1025 struct toedev *tod = sc->sc_tod; 1026 1027 tod->tod_offload_socket(tod, sc->sc_todctx, so); 1028 } 1029 #endif 1030 /* 1031 * Copy and activate timers. 1032 */ 1033 tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime; 1034 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 1035 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 1036 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1037 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1038 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1039 1040 TCPSTAT_INC(tcps_accepts); 1041 TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN); 1042 1043 if (!solisten_enqueue(so, SS_ISCONNECTED)) 1044 tp->t_flags |= TF_SONOTCONN; 1045 1046 return (so); 1047 1048 allocfail: 1049 /* 1050 * Drop the connection; we will either send a RST or have the peer 1051 * retransmit its SYN again after its RTO and try again. 1052 */ 1053 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 1054 log(LOG_DEBUG, "%s; %s: Socket create failed " 1055 "due to limits or memory shortage\n", 1056 s, __func__); 1057 free(s, M_TCPLOG); 1058 } 1059 TCPSTAT_INC(tcps_listendrop); 1060 return (NULL); 1061 1062 abort: 1063 in_pcbdetach(inp); 1064 in_pcbfree(inp); 1065 sodealloc(so); 1066 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 1067 log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n", 1068 s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "", 1069 error); 1070 free(s, M_TCPLOG); 1071 } 1072 TCPSTAT_INC(tcps_listendrop); 1073 return (NULL); 1074 } 1075 1076 /* 1077 * This function gets called when we receive an ACK for a 1078 * socket in the LISTEN state. We look up the connection 1079 * in the syncache, and if its there, we pull it out of 1080 * the cache and turn it into a full-blown connection in 1081 * the SYN-RECEIVED state. 1082 * 1083 * On syncache_socket() success the newly created socket 1084 * has its underlying inp locked. 1085 */ 1086 int 1087 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1088 struct socket **lsop, struct mbuf *m, uint16_t port) 1089 { 1090 struct syncache *sc; 1091 struct syncache_head *sch; 1092 struct syncache scs; 1093 char *s; 1094 bool locked; 1095 1096 NET_EPOCH_ASSERT(); 1097 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1098 ("%s: can handle only ACK", __func__)); 1099 1100 if (syncache_cookiesonly()) { 1101 sc = NULL; 1102 sch = syncache_hashbucket(inc); 1103 locked = false; 1104 } else { 1105 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1106 locked = true; 1107 SCH_LOCK_ASSERT(sch); 1108 } 1109 1110 #ifdef INVARIANTS 1111 /* 1112 * Test code for syncookies comparing the syncache stored 1113 * values with the reconstructed values from the cookie. 1114 */ 1115 if (sc != NULL) 1116 syncookie_cmp(inc, sch, sc, th, to, *lsop, port); 1117 #endif 1118 1119 if (sc == NULL) { 1120 /* 1121 * There is no syncache entry, so see if this ACK is 1122 * a returning syncookie. To do this, first: 1123 * A. Check if syncookies are used in case of syncache 1124 * overflows 1125 * B. See if this socket has had a syncache entry dropped in 1126 * the recent past. We don't want to accept a bogus 1127 * syncookie if we've never received a SYN or accept it 1128 * twice. 1129 * C. check that the syncookie is valid. If it is, then 1130 * cobble up a fake syncache entry, and return. 1131 */ 1132 if (locked && !V_tcp_syncookies) { 1133 SCH_UNLOCK(sch); 1134 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1135 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1136 "segment rejected (syncookies disabled)\n", 1137 s, __func__); 1138 goto failed; 1139 } 1140 if (locked && !V_tcp_syncookiesonly && 1141 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1142 SCH_UNLOCK(sch); 1143 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1144 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1145 "segment rejected (no syncache entry)\n", 1146 s, __func__); 1147 goto failed; 1148 } 1149 bzero(&scs, sizeof(scs)); 1150 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port); 1151 if (locked) 1152 SCH_UNLOCK(sch); 1153 if (sc == NULL) { 1154 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1155 log(LOG_DEBUG, "%s; %s: Segment failed " 1156 "SYNCOOKIE authentication, segment rejected " 1157 "(probably spoofed)\n", s, __func__); 1158 goto failed; 1159 } 1160 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1161 /* If received ACK has MD5 signature, check it. */ 1162 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1163 (!TCPMD5_ENABLED() || 1164 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1165 /* Drop the ACK. */ 1166 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1167 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1168 "MD5 signature doesn't match.\n", 1169 s, __func__); 1170 free(s, M_TCPLOG); 1171 } 1172 TCPSTAT_INC(tcps_sig_err_sigopt); 1173 return (-1); /* Do not send RST */ 1174 } 1175 #endif /* TCP_SIGNATURE */ 1176 TCPSTATES_INC(TCPS_SYN_RECEIVED); 1177 } else { 1178 if (sc->sc_port != port) { 1179 SCH_UNLOCK(sch); 1180 return (0); 1181 } 1182 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1183 /* 1184 * If listening socket requested TCP digests, check that 1185 * received ACK has signature and it is correct. 1186 * If not, drop the ACK and leave sc entry in th cache, 1187 * because SYN was received with correct signature. 1188 */ 1189 if (sc->sc_flags & SCF_SIGNATURE) { 1190 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1191 /* No signature */ 1192 TCPSTAT_INC(tcps_sig_err_nosigopt); 1193 SCH_UNLOCK(sch); 1194 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1195 log(LOG_DEBUG, "%s; %s: Segment " 1196 "rejected, MD5 signature wasn't " 1197 "provided.\n", s, __func__); 1198 free(s, M_TCPLOG); 1199 } 1200 return (-1); /* Do not send RST */ 1201 } 1202 if (!TCPMD5_ENABLED() || 1203 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1204 /* Doesn't match or no SA */ 1205 SCH_UNLOCK(sch); 1206 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1207 log(LOG_DEBUG, "%s; %s: Segment " 1208 "rejected, MD5 signature doesn't " 1209 "match.\n", s, __func__); 1210 free(s, M_TCPLOG); 1211 } 1212 return (-1); /* Do not send RST */ 1213 } 1214 } 1215 #endif /* TCP_SIGNATURE */ 1216 1217 /* 1218 * RFC 7323 PAWS: If we have a timestamp on this segment and 1219 * it's less than ts_recent, drop it. 1220 * XXXMT: RFC 7323 also requires to send an ACK. 1221 * In tcp_input.c this is only done for TCP segments 1222 * with user data, so be consistent here and just drop 1223 * the segment. 1224 */ 1225 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1226 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1227 SCH_UNLOCK(sch); 1228 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1229 log(LOG_DEBUG, 1230 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1231 "segment dropped\n", s, __func__, 1232 to->to_tsval, sc->sc_tsreflect); 1233 free(s, M_TCPLOG); 1234 } 1235 return (-1); /* Do not send RST */ 1236 } 1237 1238 /* 1239 * If timestamps were not negotiated during SYN/ACK and a 1240 * segment with a timestamp is received, ignore the 1241 * timestamp and process the packet normally. 1242 * See section 3.2 of RFC 7323. 1243 */ 1244 if (!(sc->sc_flags & SCF_TIMESTAMP) && 1245 (to->to_flags & TOF_TS)) { 1246 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1247 log(LOG_DEBUG, "%s; %s: Timestamp not " 1248 "expected, segment processed normally\n", 1249 s, __func__); 1250 free(s, M_TCPLOG); 1251 s = NULL; 1252 } 1253 } 1254 1255 /* 1256 * If timestamps were negotiated during SYN/ACK and a 1257 * segment without a timestamp is received, silently drop 1258 * the segment, unless the missing timestamps are tolerated. 1259 * See section 3.2 of RFC 7323. 1260 */ 1261 if ((sc->sc_flags & SCF_TIMESTAMP) && 1262 !(to->to_flags & TOF_TS)) { 1263 if (V_tcp_tolerate_missing_ts) { 1264 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1265 log(LOG_DEBUG, 1266 "%s; %s: Timestamp missing, " 1267 "segment processed normally\n", 1268 s, __func__); 1269 free(s, M_TCPLOG); 1270 } 1271 } else { 1272 SCH_UNLOCK(sch); 1273 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1274 log(LOG_DEBUG, 1275 "%s; %s: Timestamp missing, " 1276 "segment silently dropped\n", 1277 s, __func__); 1278 free(s, M_TCPLOG); 1279 } 1280 return (-1); /* Do not send RST */ 1281 } 1282 } 1283 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1284 sch->sch_length--; 1285 #ifdef TCP_OFFLOAD 1286 if (ADDED_BY_TOE(sc)) { 1287 struct toedev *tod = sc->sc_tod; 1288 1289 tod->tod_syncache_removed(tod, sc->sc_todctx); 1290 } 1291 #endif 1292 SCH_UNLOCK(sch); 1293 } 1294 1295 /* 1296 * Segment validation: 1297 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1298 */ 1299 if (th->th_ack != sc->sc_iss + 1) { 1300 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1301 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1302 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1303 goto failed; 1304 } 1305 1306 /* 1307 * The SEQ must fall in the window starting at the received 1308 * initial receive sequence number + 1 (the SYN). 1309 */ 1310 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1311 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1312 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1313 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1314 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1315 goto failed; 1316 } 1317 1318 *lsop = syncache_socket(sc, *lsop, m); 1319 1320 if (*lsop == NULL) 1321 TCPSTAT_INC(tcps_sc_aborted); 1322 else 1323 TCPSTAT_INC(tcps_sc_completed); 1324 1325 /* how do we find the inp for the new socket? */ 1326 if (sc != &scs) 1327 syncache_free(sc); 1328 return (1); 1329 failed: 1330 if (sc != NULL) { 1331 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1332 if (sc != &scs) 1333 syncache_free(sc); 1334 } 1335 if (s != NULL) 1336 free(s, M_TCPLOG); 1337 *lsop = NULL; 1338 return (0); 1339 } 1340 1341 static struct socket * 1342 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m, 1343 uint64_t response_cookie) 1344 { 1345 struct inpcb *inp; 1346 struct tcpcb *tp; 1347 unsigned int *pending_counter; 1348 struct socket *so; 1349 1350 NET_EPOCH_ASSERT(); 1351 1352 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending; 1353 so = syncache_socket(sc, lso, m); 1354 if (so == NULL) { 1355 TCPSTAT_INC(tcps_sc_aborted); 1356 atomic_subtract_int(pending_counter, 1); 1357 } else { 1358 soisconnected(so); 1359 inp = sotoinpcb(so); 1360 tp = intotcpcb(inp); 1361 tp->t_flags |= TF_FASTOPEN; 1362 tp->t_tfo_cookie.server = response_cookie; 1363 tp->snd_max = tp->iss; 1364 tp->snd_nxt = tp->iss; 1365 tp->t_tfo_pending = pending_counter; 1366 TCPSTATES_INC(TCPS_SYN_RECEIVED); 1367 TCPSTAT_INC(tcps_sc_completed); 1368 } 1369 1370 return (so); 1371 } 1372 1373 /* 1374 * Given a LISTEN socket and an inbound SYN request, add 1375 * this to the syn cache, and send back a segment: 1376 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1377 * to the source. 1378 * 1379 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1380 * Doing so would require that we hold onto the data and deliver it 1381 * to the application. However, if we are the target of a SYN-flood 1382 * DoS attack, an attacker could send data which would eventually 1383 * consume all available buffer space if it were ACKed. By not ACKing 1384 * the data, we avoid this DoS scenario. 1385 * 1386 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1387 * cookie is processed and a new socket is created. In this case, any data 1388 * accompanying the SYN will be queued to the socket by tcp_input() and will 1389 * be ACKed either when the application sends response data or the delayed 1390 * ACK timer expires, whichever comes first. 1391 */ 1392 struct socket * 1393 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1394 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod, 1395 void *todctx, uint8_t iptos, uint16_t port) 1396 { 1397 struct tcpcb *tp; 1398 struct socket *rv = NULL; 1399 struct syncache *sc = NULL; 1400 struct syncache_head *sch; 1401 struct mbuf *ipopts = NULL; 1402 u_int ltflags; 1403 int win, ip_ttl, ip_tos; 1404 char *s; 1405 #ifdef INET6 1406 int autoflowlabel = 0; 1407 #endif 1408 #ifdef MAC 1409 struct label *maclabel; 1410 #endif 1411 struct syncache scs; 1412 struct ucred *cred; 1413 uint64_t tfo_response_cookie; 1414 unsigned int *tfo_pending = NULL; 1415 int tfo_cookie_valid = 0; 1416 int tfo_response_cookie_valid = 0; 1417 bool locked; 1418 1419 INP_RLOCK_ASSERT(inp); /* listen socket */ 1420 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1421 ("%s: unexpected tcp flags", __func__)); 1422 1423 /* 1424 * Combine all so/tp operations very early to drop the INP lock as 1425 * soon as possible. 1426 */ 1427 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1428 tp = sototcpcb(so); 1429 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred); 1430 1431 #ifdef INET6 1432 if (inc->inc_flags & INC_ISIPV6) { 1433 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1434 autoflowlabel = 1; 1435 } 1436 ip_ttl = in6_selecthlim(inp, NULL); 1437 if ((inp->in6p_outputopts == NULL) || 1438 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1439 ip_tos = 0; 1440 } else { 1441 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1442 } 1443 } 1444 #endif 1445 #if defined(INET6) && defined(INET) 1446 else 1447 #endif 1448 #ifdef INET 1449 { 1450 ip_ttl = inp->inp_ip_ttl; 1451 ip_tos = inp->inp_ip_tos; 1452 } 1453 #endif 1454 win = so->sol_sbrcv_hiwat; 1455 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1456 1457 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1458 (tp->t_tfo_pending != NULL) && 1459 (to->to_flags & TOF_FASTOPEN)) { 1460 /* 1461 * Limit the number of pending TFO connections to 1462 * approximately half of the queue limit. This prevents TFO 1463 * SYN floods from starving the service by filling the 1464 * listen queue with bogus TFO connections. 1465 */ 1466 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1467 (so->sol_qlimit / 2)) { 1468 int result; 1469 1470 result = tcp_fastopen_check_cookie(inc, 1471 to->to_tfo_cookie, to->to_tfo_len, 1472 &tfo_response_cookie); 1473 tfo_cookie_valid = (result > 0); 1474 tfo_response_cookie_valid = (result >= 0); 1475 } 1476 1477 /* 1478 * Remember the TFO pending counter as it will have to be 1479 * decremented below if we don't make it to syncache_tfo_expand(). 1480 */ 1481 tfo_pending = tp->t_tfo_pending; 1482 } 1483 1484 #ifdef MAC 1485 if (mac_syncache_init(&maclabel) != 0) { 1486 INP_RUNLOCK(inp); 1487 goto done; 1488 } else 1489 mac_syncache_create(maclabel, inp); 1490 #endif 1491 if (!tfo_cookie_valid) 1492 INP_RUNLOCK(inp); 1493 1494 /* 1495 * Remember the IP options, if any. 1496 */ 1497 #ifdef INET6 1498 if (!(inc->inc_flags & INC_ISIPV6)) 1499 #endif 1500 #ifdef INET 1501 ipopts = (m) ? ip_srcroute(m) : NULL; 1502 #else 1503 ipopts = NULL; 1504 #endif 1505 1506 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1507 /* 1508 * When the socket is TCP-MD5 enabled check that, 1509 * - a signed packet is valid 1510 * - a non-signed packet does not have a security association 1511 * 1512 * If a signed packet fails validation or a non-signed packet has a 1513 * security association, the packet will be dropped. 1514 */ 1515 if (ltflags & TF_SIGNATURE) { 1516 if (to->to_flags & TOF_SIGNATURE) { 1517 if (!TCPMD5_ENABLED() || 1518 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1519 goto done; 1520 } else { 1521 if (TCPMD5_ENABLED() && 1522 TCPMD5_INPUT(m, NULL, NULL) != ENOENT) 1523 goto done; 1524 } 1525 } else if (to->to_flags & TOF_SIGNATURE) 1526 goto done; 1527 #endif /* TCP_SIGNATURE */ 1528 /* 1529 * See if we already have an entry for this connection. 1530 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1531 * 1532 * XXX: should the syncache be re-initialized with the contents 1533 * of the new SYN here (which may have different options?) 1534 * 1535 * XXX: We do not check the sequence number to see if this is a 1536 * real retransmit or a new connection attempt. The question is 1537 * how to handle such a case; either ignore it as spoofed, or 1538 * drop the current entry and create a new one? 1539 */ 1540 if (syncache_cookiesonly()) { 1541 sc = NULL; 1542 sch = syncache_hashbucket(inc); 1543 locked = false; 1544 } else { 1545 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1546 locked = true; 1547 SCH_LOCK_ASSERT(sch); 1548 } 1549 if (sc != NULL) { 1550 if (tfo_cookie_valid) 1551 INP_RUNLOCK(inp); 1552 TCPSTAT_INC(tcps_sc_dupsyn); 1553 if (ipopts) { 1554 /* 1555 * If we were remembering a previous source route, 1556 * forget it and use the new one we've been given. 1557 */ 1558 if (sc->sc_ipopts) 1559 (void) m_free(sc->sc_ipopts); 1560 sc->sc_ipopts = ipopts; 1561 } 1562 /* 1563 * Update timestamp if present. 1564 */ 1565 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1566 sc->sc_tsreflect = to->to_tsval; 1567 else 1568 sc->sc_flags &= ~SCF_TIMESTAMP; 1569 /* 1570 * Disable ECN if needed. 1571 */ 1572 if ((sc->sc_flags & SCF_ECN) && 1573 ((tcp_get_flags(th) & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) { 1574 sc->sc_flags &= ~SCF_ECN; 1575 } 1576 #ifdef MAC 1577 /* 1578 * Since we have already unconditionally allocated label 1579 * storage, free it up. The syncache entry will already 1580 * have an initialized label we can use. 1581 */ 1582 mac_syncache_destroy(&maclabel); 1583 #endif 1584 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1585 /* Retransmit SYN|ACK and reset retransmit count. */ 1586 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1587 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1588 "resetting timer and retransmitting SYN|ACK\n", 1589 s, __func__); 1590 free(s, M_TCPLOG); 1591 } 1592 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1593 sc->sc_rxmits = 0; 1594 syncache_timeout(sc, sch, 1); 1595 TCPSTAT_INC(tcps_sndacks); 1596 TCPSTAT_INC(tcps_sndtotal); 1597 } 1598 SCH_UNLOCK(sch); 1599 goto donenoprobe; 1600 } 1601 1602 if (tfo_cookie_valid) { 1603 bzero(&scs, sizeof(scs)); 1604 sc = &scs; 1605 goto skip_alloc; 1606 } 1607 1608 /* 1609 * Skip allocating a syncache entry if we are just going to discard 1610 * it later. 1611 */ 1612 if (!locked) { 1613 bzero(&scs, sizeof(scs)); 1614 sc = &scs; 1615 } else 1616 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1617 if (sc == NULL) { 1618 /* 1619 * The zone allocator couldn't provide more entries. 1620 * Treat this as if the cache was full; drop the oldest 1621 * entry and insert the new one. 1622 */ 1623 TCPSTAT_INC(tcps_sc_zonefail); 1624 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1625 sch->sch_last_overflow = time_uptime; 1626 syncache_drop(sc, sch); 1627 syncache_pause(inc); 1628 } 1629 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1630 if (sc == NULL) { 1631 if (V_tcp_syncookies) { 1632 bzero(&scs, sizeof(scs)); 1633 sc = &scs; 1634 } else { 1635 KASSERT(locked, 1636 ("%s: bucket unexpectedly unlocked", 1637 __func__)); 1638 SCH_UNLOCK(sch); 1639 if (ipopts) 1640 (void) m_free(ipopts); 1641 goto done; 1642 } 1643 } 1644 } 1645 1646 skip_alloc: 1647 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1648 sc->sc_tfo_cookie = &tfo_response_cookie; 1649 1650 /* 1651 * Fill in the syncache values. 1652 */ 1653 #ifdef MAC 1654 sc->sc_label = maclabel; 1655 #endif 1656 sc->sc_cred = cred; 1657 sc->sc_port = port; 1658 cred = NULL; 1659 sc->sc_ipopts = ipopts; 1660 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1661 sc->sc_ip_tos = ip_tos; 1662 sc->sc_ip_ttl = ip_ttl; 1663 #ifdef TCP_OFFLOAD 1664 sc->sc_tod = tod; 1665 sc->sc_todctx = todctx; 1666 #endif 1667 sc->sc_irs = th->th_seq; 1668 sc->sc_flags = 0; 1669 sc->sc_flowlabel = 0; 1670 1671 /* 1672 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1673 * win was derived from socket earlier in the function. 1674 */ 1675 win = imax(win, 0); 1676 win = imin(win, TCP_MAXWIN); 1677 sc->sc_wnd = win; 1678 1679 if (V_tcp_do_rfc1323 && 1680 !(ltflags & TF_NOOPT)) { 1681 /* 1682 * A timestamp received in a SYN makes 1683 * it ok to send timestamp requests and replies. 1684 */ 1685 if (to->to_flags & TOF_TS) { 1686 sc->sc_tsreflect = to->to_tsval; 1687 sc->sc_flags |= SCF_TIMESTAMP; 1688 sc->sc_tsoff = tcp_new_ts_offset(inc); 1689 } 1690 if (to->to_flags & TOF_SCALE) { 1691 int wscale = 0; 1692 1693 /* 1694 * Pick the smallest possible scaling factor that 1695 * will still allow us to scale up to sb_max, aka 1696 * kern.ipc.maxsockbuf. 1697 * 1698 * We do this because there are broken firewalls that 1699 * will corrupt the window scale option, leading to 1700 * the other endpoint believing that our advertised 1701 * window is unscaled. At scale factors larger than 1702 * 5 the unscaled window will drop below 1500 bytes, 1703 * leading to serious problems when traversing these 1704 * broken firewalls. 1705 * 1706 * With the default maxsockbuf of 256K, a scale factor 1707 * of 3 will be chosen by this algorithm. Those who 1708 * choose a larger maxsockbuf should watch out 1709 * for the compatibility problems mentioned above. 1710 * 1711 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1712 * or <SYN,ACK>) segment itself is never scaled. 1713 */ 1714 while (wscale < TCP_MAX_WINSHIFT && 1715 (TCP_MAXWIN << wscale) < sb_max) 1716 wscale++; 1717 sc->sc_requested_r_scale = wscale; 1718 sc->sc_requested_s_scale = to->to_wscale; 1719 sc->sc_flags |= SCF_WINSCALE; 1720 } 1721 } 1722 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1723 /* 1724 * If incoming packet has an MD5 signature, flag this in the 1725 * syncache so that syncache_respond() will do the right thing 1726 * with the SYN+ACK. 1727 */ 1728 if (to->to_flags & TOF_SIGNATURE) 1729 sc->sc_flags |= SCF_SIGNATURE; 1730 #endif /* TCP_SIGNATURE */ 1731 if (to->to_flags & TOF_SACKPERM) 1732 sc->sc_flags |= SCF_SACK; 1733 if (to->to_flags & TOF_MSS) 1734 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1735 if (ltflags & TF_NOOPT) 1736 sc->sc_flags |= SCF_NOOPT; 1737 /* ECN Handshake */ 1738 if (V_tcp_do_ecn) 1739 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos); 1740 1741 if (V_tcp_syncookies) 1742 sc->sc_iss = syncookie_generate(sch, sc); 1743 else 1744 sc->sc_iss = arc4random(); 1745 #ifdef INET6 1746 if (autoflowlabel) { 1747 if (V_tcp_syncookies) 1748 sc->sc_flowlabel = sc->sc_iss; 1749 else 1750 sc->sc_flowlabel = ip6_randomflowlabel(); 1751 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1752 } 1753 #endif 1754 if (locked) 1755 SCH_UNLOCK(sch); 1756 1757 if (tfo_cookie_valid) { 1758 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie); 1759 /* INP_RUNLOCK(inp) will be performed by the caller */ 1760 goto tfo_expanded; 1761 } 1762 1763 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1764 /* 1765 * Do a standard 3-way handshake. 1766 */ 1767 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1768 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1769 syncache_free(sc); 1770 else if (sc != &scs) 1771 syncache_insert(sc, sch); /* locks and unlocks sch */ 1772 TCPSTAT_INC(tcps_sndacks); 1773 TCPSTAT_INC(tcps_sndtotal); 1774 } else { 1775 if (sc != &scs) 1776 syncache_free(sc); 1777 TCPSTAT_INC(tcps_sc_dropped); 1778 } 1779 goto donenoprobe; 1780 1781 done: 1782 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1783 donenoprobe: 1784 if (m) 1785 m_freem(m); 1786 /* 1787 * If tfo_pending is not NULL here, then a TFO SYN that did not 1788 * result in a new socket was processed and the associated pending 1789 * counter has not yet been decremented. All such TFO processing paths 1790 * transit this point. 1791 */ 1792 if (tfo_pending != NULL) 1793 tcp_fastopen_decrement_counter(tfo_pending); 1794 1795 tfo_expanded: 1796 if (cred != NULL) 1797 crfree(cred); 1798 #ifdef MAC 1799 if (sc == &scs) 1800 mac_syncache_destroy(&maclabel); 1801 #endif 1802 return (rv); 1803 } 1804 1805 /* 1806 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1807 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1808 */ 1809 static int 1810 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags) 1811 { 1812 struct ip *ip = NULL; 1813 struct mbuf *m; 1814 struct tcphdr *th = NULL; 1815 struct udphdr *udp = NULL; 1816 int optlen, error = 0; /* Make compiler happy */ 1817 u_int16_t hlen, tlen, mssopt, ulen; 1818 struct tcpopt to; 1819 #ifdef INET6 1820 struct ip6_hdr *ip6 = NULL; 1821 #endif 1822 1823 NET_EPOCH_ASSERT(); 1824 1825 hlen = 1826 #ifdef INET6 1827 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1828 #endif 1829 sizeof(struct ip); 1830 tlen = hlen + sizeof(struct tcphdr); 1831 if (sc->sc_port) { 1832 tlen += sizeof(struct udphdr); 1833 } 1834 /* Determine MSS we advertize to other end of connection. */ 1835 mssopt = tcp_mssopt(&sc->sc_inc); 1836 if (sc->sc_port) 1837 mssopt -= V_tcp_udp_tunneling_overhead; 1838 mssopt = max(mssopt, V_tcp_minmss); 1839 1840 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1841 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1842 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + " 1843 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port, 1844 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN)); 1845 1846 /* Create the IP+TCP header from scratch. */ 1847 m = m_gethdr(M_NOWAIT, MT_DATA); 1848 if (m == NULL) 1849 return (ENOBUFS); 1850 #ifdef MAC 1851 mac_syncache_create_mbuf(sc->sc_label, m); 1852 #endif 1853 m->m_data += max_linkhdr; 1854 m->m_len = tlen; 1855 m->m_pkthdr.len = tlen; 1856 m->m_pkthdr.rcvif = NULL; 1857 1858 #ifdef INET6 1859 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1860 ip6 = mtod(m, struct ip6_hdr *); 1861 ip6->ip6_vfc = IPV6_VERSION; 1862 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1863 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1864 ip6->ip6_plen = htons(tlen - hlen); 1865 /* ip6_hlim is set after checksum */ 1866 /* Zero out traffic class and flow label. */ 1867 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1868 ip6->ip6_flow |= sc->sc_flowlabel; 1869 if (sc->sc_port != 0) { 1870 ip6->ip6_nxt = IPPROTO_UDP; 1871 udp = (struct udphdr *)(ip6 + 1); 1872 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1873 udp->uh_dport = sc->sc_port; 1874 ulen = (tlen - sizeof(struct ip6_hdr)); 1875 th = (struct tcphdr *)(udp + 1); 1876 } else { 1877 ip6->ip6_nxt = IPPROTO_TCP; 1878 th = (struct tcphdr *)(ip6 + 1); 1879 } 1880 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20); 1881 } 1882 #endif 1883 #if defined(INET6) && defined(INET) 1884 else 1885 #endif 1886 #ifdef INET 1887 { 1888 ip = mtod(m, struct ip *); 1889 ip->ip_v = IPVERSION; 1890 ip->ip_hl = sizeof(struct ip) >> 2; 1891 ip->ip_len = htons(tlen); 1892 ip->ip_id = 0; 1893 ip->ip_off = 0; 1894 ip->ip_sum = 0; 1895 ip->ip_src = sc->sc_inc.inc_laddr; 1896 ip->ip_dst = sc->sc_inc.inc_faddr; 1897 ip->ip_ttl = sc->sc_ip_ttl; 1898 ip->ip_tos = sc->sc_ip_tos; 1899 1900 /* 1901 * See if we should do MTU discovery. Route lookups are 1902 * expensive, so we will only unset the DF bit if: 1903 * 1904 * 1) path_mtu_discovery is disabled 1905 * 2) the SCF_UNREACH flag has been set 1906 */ 1907 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1908 ip->ip_off |= htons(IP_DF); 1909 if (sc->sc_port == 0) { 1910 ip->ip_p = IPPROTO_TCP; 1911 th = (struct tcphdr *)(ip + 1); 1912 } else { 1913 ip->ip_p = IPPROTO_UDP; 1914 udp = (struct udphdr *)(ip + 1); 1915 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1916 udp->uh_dport = sc->sc_port; 1917 ulen = (tlen - sizeof(struct ip)); 1918 th = (struct tcphdr *)(udp + 1); 1919 } 1920 } 1921 #endif /* INET */ 1922 th->th_sport = sc->sc_inc.inc_lport; 1923 th->th_dport = sc->sc_inc.inc_fport; 1924 1925 if (flags & TH_SYN) 1926 th->th_seq = htonl(sc->sc_iss); 1927 else 1928 th->th_seq = htonl(sc->sc_iss + 1); 1929 th->th_ack = htonl(sc->sc_irs + 1); 1930 th->th_off = sizeof(struct tcphdr) >> 2; 1931 th->th_win = htons(sc->sc_wnd); 1932 th->th_urp = 0; 1933 1934 flags = tcp_ecn_syncache_respond(flags, sc); 1935 tcp_set_flags(th, flags); 1936 1937 /* Tack on the TCP options. */ 1938 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1939 to.to_flags = 0; 1940 1941 if (flags & TH_SYN) { 1942 to.to_mss = mssopt; 1943 to.to_flags = TOF_MSS; 1944 if (sc->sc_flags & SCF_WINSCALE) { 1945 to.to_wscale = sc->sc_requested_r_scale; 1946 to.to_flags |= TOF_SCALE; 1947 } 1948 if (sc->sc_flags & SCF_SACK) 1949 to.to_flags |= TOF_SACKPERM; 1950 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1951 if (sc->sc_flags & SCF_SIGNATURE) 1952 to.to_flags |= TOF_SIGNATURE; 1953 #endif 1954 if (sc->sc_tfo_cookie) { 1955 to.to_flags |= TOF_FASTOPEN; 1956 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1957 to.to_tfo_cookie = sc->sc_tfo_cookie; 1958 /* don't send cookie again when retransmitting response */ 1959 sc->sc_tfo_cookie = NULL; 1960 } 1961 } 1962 if (sc->sc_flags & SCF_TIMESTAMP) { 1963 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1964 to.to_tsecr = sc->sc_tsreflect; 1965 to.to_flags |= TOF_TS; 1966 } 1967 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1968 1969 /* Adjust headers by option size. */ 1970 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1971 m->m_len += optlen; 1972 m->m_pkthdr.len += optlen; 1973 #ifdef INET6 1974 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1975 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1976 else 1977 #endif 1978 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1979 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1980 if (sc->sc_flags & SCF_SIGNATURE) { 1981 KASSERT(to.to_flags & TOF_SIGNATURE, 1982 ("tcp_addoptions() didn't set tcp_signature")); 1983 1984 /* NOTE: to.to_signature is inside of mbuf */ 1985 if (!TCPMD5_ENABLED() || 1986 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1987 m_freem(m); 1988 return (EACCES); 1989 } 1990 } 1991 #endif 1992 } else 1993 optlen = 0; 1994 1995 if (udp) { 1996 ulen += optlen; 1997 udp->uh_ulen = htons(ulen); 1998 } 1999 M_SETFIB(m, sc->sc_inc.inc_fibnum); 2000 /* 2001 * If we have peer's SYN and it has a flowid, then let's assign it to 2002 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 2003 * to SYN|ACK due to lack of inp here. 2004 */ 2005 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 2006 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 2007 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 2008 } 2009 #ifdef INET6 2010 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 2011 if (sc->sc_port) { 2012 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 2013 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2014 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, 2015 IPPROTO_UDP, 0); 2016 th->th_sum = htons(0); 2017 } else { 2018 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 2019 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2020 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 2021 IPPROTO_TCP, 0); 2022 } 2023 ip6->ip6_hlim = sc->sc_ip_ttl; 2024 #ifdef TCP_OFFLOAD 2025 if (ADDED_BY_TOE(sc)) { 2026 struct toedev *tod = sc->sc_tod; 2027 2028 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2029 2030 return (error); 2031 } 2032 #endif 2033 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 2034 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 2035 } 2036 #endif 2037 #if defined(INET6) && defined(INET) 2038 else 2039 #endif 2040 #ifdef INET 2041 { 2042 if (sc->sc_port) { 2043 m->m_pkthdr.csum_flags = CSUM_UDP; 2044 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2045 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 2046 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 2047 th->th_sum = htons(0); 2048 } else { 2049 m->m_pkthdr.csum_flags = CSUM_TCP; 2050 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2051 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 2052 htons(tlen + optlen - hlen + IPPROTO_TCP)); 2053 } 2054 #ifdef TCP_OFFLOAD 2055 if (ADDED_BY_TOE(sc)) { 2056 struct toedev *tod = sc->sc_tod; 2057 2058 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2059 2060 return (error); 2061 } 2062 #endif 2063 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 2064 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 2065 } 2066 #endif 2067 return (error); 2068 } 2069 2070 /* 2071 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 2072 * that exceed the capacity of the syncache by avoiding the storage of any 2073 * of the SYNs we receive. Syncookies defend against blind SYN flooding 2074 * attacks where the attacker does not have access to our responses. 2075 * 2076 * Syncookies encode and include all necessary information about the 2077 * connection setup within the SYN|ACK that we send back. That way we 2078 * can avoid keeping any local state until the ACK to our SYN|ACK returns 2079 * (if ever). Normally the syncache and syncookies are running in parallel 2080 * with the latter taking over when the former is exhausted. When matching 2081 * syncache entry is found the syncookie is ignored. 2082 * 2083 * The only reliable information persisting the 3WHS is our initial sequence 2084 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 2085 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 2086 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 2087 * returns and signifies a legitimate connection if it matches the ACK. 2088 * 2089 * The available space of 32 bits to store the hash and to encode the SYN 2090 * option information is very tight and we should have at least 24 bits for 2091 * the MAC to keep the number of guesses by blind spoofing reasonably high. 2092 * 2093 * SYN option information we have to encode to fully restore a connection: 2094 * MSS: is imporant to chose an optimal segment size to avoid IP level 2095 * fragmentation along the path. The common MSS values can be encoded 2096 * in a 3-bit table. Uncommon values are captured by the next lower value 2097 * in the table leading to a slight increase in packetization overhead. 2098 * WSCALE: is necessary to allow large windows to be used for high delay- 2099 * bandwidth product links. Not scaling the window when it was initially 2100 * negotiated is bad for performance as lack of scaling further decreases 2101 * the apparent available send window. We only need to encode the WSCALE 2102 * we received from the remote end. Our end can be recalculated at any 2103 * time. The common WSCALE values can be encoded in a 3-bit table. 2104 * Uncommon values are captured by the next lower value in the table 2105 * making us under-estimate the available window size halving our 2106 * theoretically possible maximum throughput for that connection. 2107 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 2108 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 2109 * that are included in all segments on a connection. We enable them when 2110 * the ACK has them. 2111 * 2112 * Security of syncookies and attack vectors: 2113 * 2114 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 2115 * together with the gloabl secret to make it unique per connection attempt. 2116 * Thus any change of any of those parameters results in a different MAC output 2117 * in an unpredictable way unless a collision is encountered. 24 bits of the 2118 * MAC are embedded into the ISS. 2119 * 2120 * To prevent replay attacks two rotating global secrets are updated with a 2121 * new random value every 15 seconds. The life-time of a syncookie is thus 2122 * 15-30 seconds. 2123 * 2124 * Vector 1: Attacking the secret. This requires finding a weakness in the 2125 * MAC itself or the way it is used here. The attacker can do a chosen plain 2126 * text attack by varying and testing the all parameters under his control. 2127 * The strength depends on the size and randomness of the secret, and the 2128 * cryptographic security of the MAC function. Due to the constant updating 2129 * of the secret the attacker has at most 29.999 seconds to find the secret 2130 * and launch spoofed connections. After that he has to start all over again. 2131 * 2132 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 2133 * size an average of 4,823 attempts are required for a 50% chance of success 2134 * to spoof a single syncookie (birthday collision paradox). However the 2135 * attacker is blind and doesn't know if one of his attempts succeeded unless 2136 * he has a side channel to interfere success from. A single connection setup 2137 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 2138 * This many attempts are required for each one blind spoofed connection. For 2139 * every additional spoofed connection he has to launch another N attempts. 2140 * Thus for a sustained rate 100 spoofed connections per second approximately 2141 * 1,800,000 packets per second would have to be sent. 2142 * 2143 * NB: The MAC function should be fast so that it doesn't become a CPU 2144 * exhaustion attack vector itself. 2145 * 2146 * References: 2147 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 2148 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 2149 * http://cr.yp.to/syncookies.html (overview) 2150 * http://cr.yp.to/syncookies/archive (details) 2151 * 2152 * 2153 * Schematic construction of a syncookie enabled Initial Sequence Number: 2154 * 0 1 2 3 2155 * 12345678901234567890123456789012 2156 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2157 * 2158 * x 24 MAC (truncated) 2159 * W 3 Send Window Scale index 2160 * M 3 MSS index 2161 * S 1 SACK permitted 2162 * P 1 Odd/even secret 2163 */ 2164 2165 /* 2166 * Distribution and probability of certain MSS values. Those in between are 2167 * rounded down to the next lower one. 2168 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2169 * .2% .3% 5% 7% 7% 20% 15% 45% 2170 */ 2171 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2172 2173 /* 2174 * Distribution and probability of certain WSCALE values. We have to map the 2175 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2176 * bits based on prevalence of certain values. Where we don't have an exact 2177 * match for are rounded down to the next lower one letting us under-estimate 2178 * the true available window. At the moment this would happen only for the 2179 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2180 * and window size). The absence of the WSCALE option (no scaling in either 2181 * direction) is encoded with index zero. 2182 * [WSCALE values histograms, Allman, 2012] 2183 * X 10 10 35 5 6 14 10% by host 2184 * X 11 4 5 5 18 49 3% by connections 2185 */ 2186 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2187 2188 /* 2189 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2190 * and good cryptographic properties. 2191 */ 2192 static uint32_t 2193 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2194 uint8_t *secbits, uintptr_t secmod) 2195 { 2196 SIPHASH_CTX ctx; 2197 uint32_t siphash[2]; 2198 2199 SipHash24_Init(&ctx); 2200 SipHash_SetKey(&ctx, secbits); 2201 switch (inc->inc_flags & INC_ISIPV6) { 2202 #ifdef INET 2203 case 0: 2204 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2205 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2206 break; 2207 #endif 2208 #ifdef INET6 2209 case INC_ISIPV6: 2210 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2211 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2212 break; 2213 #endif 2214 } 2215 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2216 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2217 SipHash_Update(&ctx, &irs, sizeof(irs)); 2218 SipHash_Update(&ctx, &flags, sizeof(flags)); 2219 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2220 SipHash_Final((u_int8_t *)&siphash, &ctx); 2221 2222 return (siphash[0] ^ siphash[1]); 2223 } 2224 2225 static tcp_seq 2226 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2227 { 2228 u_int i, secbit, wscale; 2229 uint32_t iss, hash; 2230 uint8_t *secbits; 2231 union syncookie cookie; 2232 2233 cookie.cookie = 0; 2234 2235 /* Map our computed MSS into the 3-bit index. */ 2236 for (i = nitems(tcp_sc_msstab) - 1; 2237 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2238 i--) 2239 ; 2240 cookie.flags.mss_idx = i; 2241 2242 /* 2243 * Map the send window scale into the 3-bit index but only if 2244 * the wscale option was received. 2245 */ 2246 if (sc->sc_flags & SCF_WINSCALE) { 2247 wscale = sc->sc_requested_s_scale; 2248 for (i = nitems(tcp_sc_wstab) - 1; 2249 tcp_sc_wstab[i] > wscale && i > 0; 2250 i--) 2251 ; 2252 cookie.flags.wscale_idx = i; 2253 } 2254 2255 /* Can we do SACK? */ 2256 if (sc->sc_flags & SCF_SACK) 2257 cookie.flags.sack_ok = 1; 2258 2259 /* Which of the two secrets to use. */ 2260 secbit = V_tcp_syncache.secret.oddeven & 0x1; 2261 cookie.flags.odd_even = secbit; 2262 2263 secbits = V_tcp_syncache.secret.key[secbit]; 2264 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2265 (uintptr_t)sch); 2266 2267 /* 2268 * Put the flags into the hash and XOR them to get better ISS number 2269 * variance. This doesn't enhance the cryptographic strength and is 2270 * done to prevent the 8 cookie bits from showing up directly on the 2271 * wire. 2272 */ 2273 iss = hash & ~0xff; 2274 iss |= cookie.cookie ^ (hash >> 24); 2275 2276 TCPSTAT_INC(tcps_sc_sendcookie); 2277 return (iss); 2278 } 2279 2280 static struct syncache * 2281 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2282 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2283 struct socket *lso, uint16_t port) 2284 { 2285 uint32_t hash; 2286 uint8_t *secbits; 2287 tcp_seq ack, seq; 2288 int wnd, wscale = 0; 2289 union syncookie cookie; 2290 2291 /* 2292 * Pull information out of SYN-ACK/ACK and revert sequence number 2293 * advances. 2294 */ 2295 ack = th->th_ack - 1; 2296 seq = th->th_seq - 1; 2297 2298 /* 2299 * Unpack the flags containing enough information to restore the 2300 * connection. 2301 */ 2302 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2303 2304 /* Which of the two secrets to use. */ 2305 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; 2306 2307 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2308 2309 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2310 if ((ack & ~0xff) != (hash & ~0xff)) 2311 return (NULL); 2312 2313 /* Fill in the syncache values. */ 2314 sc->sc_flags = 0; 2315 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2316 sc->sc_ipopts = NULL; 2317 2318 sc->sc_irs = seq; 2319 sc->sc_iss = ack; 2320 2321 switch (inc->inc_flags & INC_ISIPV6) { 2322 #ifdef INET 2323 case 0: 2324 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2325 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2326 break; 2327 #endif 2328 #ifdef INET6 2329 case INC_ISIPV6: 2330 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2331 sc->sc_flowlabel = 2332 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2333 break; 2334 #endif 2335 } 2336 2337 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2338 2339 /* We can simply recompute receive window scale we sent earlier. */ 2340 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2341 wscale++; 2342 2343 /* Only use wscale if it was enabled in the orignal SYN. */ 2344 if (cookie.flags.wscale_idx > 0) { 2345 sc->sc_requested_r_scale = wscale; 2346 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2347 sc->sc_flags |= SCF_WINSCALE; 2348 } 2349 2350 wnd = lso->sol_sbrcv_hiwat; 2351 wnd = imax(wnd, 0); 2352 wnd = imin(wnd, TCP_MAXWIN); 2353 sc->sc_wnd = wnd; 2354 2355 if (cookie.flags.sack_ok) 2356 sc->sc_flags |= SCF_SACK; 2357 2358 if (to->to_flags & TOF_TS) { 2359 sc->sc_flags |= SCF_TIMESTAMP; 2360 sc->sc_tsreflect = to->to_tsval; 2361 sc->sc_tsoff = tcp_new_ts_offset(inc); 2362 } 2363 2364 if (to->to_flags & TOF_SIGNATURE) 2365 sc->sc_flags |= SCF_SIGNATURE; 2366 2367 sc->sc_rxmits = 0; 2368 2369 sc->sc_port = port; 2370 2371 TCPSTAT_INC(tcps_sc_recvcookie); 2372 return (sc); 2373 } 2374 2375 #ifdef INVARIANTS 2376 static int 2377 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2378 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2379 struct socket *lso, uint16_t port) 2380 { 2381 struct syncache scs, *scx; 2382 char *s; 2383 2384 bzero(&scs, sizeof(scs)); 2385 scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port); 2386 2387 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2388 return (0); 2389 2390 if (scx != NULL) { 2391 if (sc->sc_peer_mss != scx->sc_peer_mss) 2392 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2393 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2394 2395 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2396 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2397 s, __func__, sc->sc_requested_r_scale, 2398 scx->sc_requested_r_scale); 2399 2400 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2401 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2402 s, __func__, sc->sc_requested_s_scale, 2403 scx->sc_requested_s_scale); 2404 2405 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2406 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2407 } 2408 2409 if (s != NULL) 2410 free(s, M_TCPLOG); 2411 return (0); 2412 } 2413 #endif /* INVARIANTS */ 2414 2415 static void 2416 syncookie_reseed(void *arg) 2417 { 2418 struct tcp_syncache *sc = arg; 2419 uint8_t *secbits; 2420 int secbit; 2421 2422 /* 2423 * Reseeding the secret doesn't have to be protected by a lock. 2424 * It only must be ensured that the new random values are visible 2425 * to all CPUs in a SMP environment. The atomic with release 2426 * semantics ensures that. 2427 */ 2428 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2429 secbits = sc->secret.key[secbit]; 2430 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2431 atomic_add_rel_int(&sc->secret.oddeven, 1); 2432 2433 /* Reschedule ourself. */ 2434 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2435 } 2436 2437 /* 2438 * We have overflowed a bucket. Let's pause dealing with the syncache. 2439 * This function will increment the bucketoverflow statistics appropriately 2440 * (once per pause when pausing is enabled; otherwise, once per overflow). 2441 */ 2442 static void 2443 syncache_pause(struct in_conninfo *inc) 2444 { 2445 time_t delta; 2446 const char *s; 2447 2448 /* XXX: 2449 * 2. Add sysctl read here so we don't get the benefit of this 2450 * change without the new sysctl. 2451 */ 2452 2453 /* 2454 * Try an unlocked read. If we already know that another thread 2455 * has activated the feature, there is no need to proceed. 2456 */ 2457 if (V_tcp_syncache.paused) 2458 return; 2459 2460 /* Are cookied enabled? If not, we can't pause. */ 2461 if (!V_tcp_syncookies) { 2462 TCPSTAT_INC(tcps_sc_bucketoverflow); 2463 return; 2464 } 2465 2466 /* 2467 * We may be the first thread to find an overflow. Get the lock 2468 * and evaluate if we need to take action. 2469 */ 2470 mtx_lock(&V_tcp_syncache.pause_mtx); 2471 if (V_tcp_syncache.paused) { 2472 mtx_unlock(&V_tcp_syncache.pause_mtx); 2473 return; 2474 } 2475 2476 /* Activate protection. */ 2477 V_tcp_syncache.paused = true; 2478 TCPSTAT_INC(tcps_sc_bucketoverflow); 2479 2480 /* 2481 * Determine the last backoff time. If we are seeing a re-newed 2482 * attack within that same time after last reactivating the syncache, 2483 * consider it an extension of the same attack. 2484 */ 2485 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; 2486 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { 2487 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { 2488 delta <<= 1; 2489 V_tcp_syncache.pause_backoff++; 2490 } 2491 } else { 2492 delta = TCP_SYNCACHE_PAUSE_TIME; 2493 V_tcp_syncache.pause_backoff = 0; 2494 } 2495 2496 /* Log a warning, including IP addresses, if able. */ 2497 if (inc != NULL) 2498 s = tcp_log_addrs(inc, NULL, NULL, NULL); 2499 else 2500 s = (const char *)NULL; 2501 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " 2502 "the next %lld seconds%s%s%s\n", (long long)delta, 2503 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", 2504 (s != NULL) ? ")" : ""); 2505 free(__DECONST(void *, s), M_TCPLOG); 2506 2507 /* Use the calculated delta to set a new pause time. */ 2508 V_tcp_syncache.pause_until = time_uptime + delta; 2509 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, 2510 &V_tcp_syncache); 2511 mtx_unlock(&V_tcp_syncache.pause_mtx); 2512 } 2513 2514 /* Evaluate whether we need to unpause. */ 2515 static void 2516 syncache_unpause(void *arg) 2517 { 2518 struct tcp_syncache *sc; 2519 time_t delta; 2520 2521 sc = arg; 2522 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); 2523 callout_deactivate(&sc->pause_co); 2524 2525 /* 2526 * Check to make sure we are not running early. If the pause 2527 * time has expired, then deactivate the protection. 2528 */ 2529 if ((delta = sc->pause_until - time_uptime) > 0) 2530 callout_schedule(&sc->pause_co, delta * hz); 2531 else 2532 sc->paused = false; 2533 } 2534 2535 /* 2536 * Exports the syncache entries to userland so that netstat can display 2537 * them alongside the other sockets. This function is intended to be 2538 * called only from tcp_pcblist. 2539 * 2540 * Due to concurrency on an active system, the number of pcbs exported 2541 * may have no relation to max_pcbs. max_pcbs merely indicates the 2542 * amount of space the caller allocated for this function to use. 2543 */ 2544 int 2545 syncache_pcblist(struct sysctl_req *req) 2546 { 2547 struct xtcpcb xt; 2548 struct syncache *sc; 2549 struct syncache_head *sch; 2550 int error, i; 2551 2552 bzero(&xt, sizeof(xt)); 2553 xt.xt_len = sizeof(xt); 2554 xt.t_state = TCPS_SYN_RECEIVED; 2555 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2556 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2557 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2558 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2559 2560 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 2561 sch = &V_tcp_syncache.hashbase[i]; 2562 SCH_LOCK(sch); 2563 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2564 if (sc->sc_cred != NULL && 2565 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2566 continue; 2567 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2568 xt.xt_inp.inp_vflag = INP_IPV6; 2569 else 2570 xt.xt_inp.inp_vflag = INP_IPV4; 2571 xt.xt_encaps_port = sc->sc_port; 2572 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2573 sizeof (struct in_conninfo)); 2574 error = SYSCTL_OUT(req, &xt, sizeof xt); 2575 if (error) { 2576 SCH_UNLOCK(sch); 2577 return (0); 2578 } 2579 } 2580 SCH_UNLOCK(sch); 2581 } 2582 2583 return (0); 2584 } 2585