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 #ifdef INET6 93 #include <netinet6/tcp6_var.h> 94 #endif 95 #ifdef TCP_OFFLOAD 96 #include <netinet/toecore.h> 97 #endif 98 #include <netinet/udp.h> 99 #include <netinet/udp_var.h> 100 101 #include <netipsec/ipsec_support.h> 102 103 #include <machine/in_cksum.h> 104 105 #include <security/mac/mac_framework.h> 106 107 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1; 108 #define V_tcp_syncookies VNET(tcp_syncookies) 109 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 110 &VNET_NAME(tcp_syncookies), 0, 111 "Use TCP SYN cookies if the syncache overflows"); 112 113 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0; 114 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 116 &VNET_NAME(tcp_syncookiesonly), 0, 117 "Use only TCP SYN cookies"); 118 119 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1; 120 #define V_functions_inherit_listen_socket_stack \ 121 VNET(functions_inherit_listen_socket_stack) 122 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack, 123 CTLFLAG_VNET | CTLFLAG_RW, 124 &VNET_NAME(functions_inherit_listen_socket_stack), 0, 125 "Inherit listen socket's stack"); 126 127 #ifdef TCP_OFFLOAD 128 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 129 #endif 130 131 static void syncache_drop(struct syncache *, struct syncache_head *); 132 static void syncache_free(struct syncache *); 133 static void syncache_insert(struct syncache *, struct syncache_head *); 134 static int syncache_respond(struct syncache *, const struct mbuf *, int); 135 static struct socket *syncache_socket(struct syncache *, struct socket *, 136 struct mbuf *m); 137 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 138 int docallout); 139 static void syncache_timer(void *); 140 141 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 142 uint8_t *, uintptr_t); 143 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 144 static struct syncache 145 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 146 struct syncache *, struct tcphdr *, struct tcpopt *, 147 struct socket *, uint16_t); 148 static void syncache_pause(struct in_conninfo *); 149 static void syncache_unpause(void *); 150 static void syncookie_reseed(void *); 151 #ifdef INVARIANTS 152 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 153 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 154 struct socket *lso, uint16_t port); 155 #endif 156 157 /* 158 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 159 * 3 retransmits corresponds to a timeout with default values of 160 * tcp_rexmit_initial * ( 1 + 161 * tcp_backoff[1] + 162 * tcp_backoff[2] + 163 * tcp_backoff[3]) + 3 * tcp_rexmit_slop, 164 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms, 165 * the odds are that the user has given up attempting to connect by then. 166 */ 167 #define SYNCACHE_MAXREXMTS 3 168 169 /* Arbitrary values */ 170 #define TCP_SYNCACHE_HASHSIZE 512 171 #define TCP_SYNCACHE_BUCKETLIMIT 30 172 173 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 174 #define V_tcp_syncache VNET(tcp_syncache) 175 176 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, 177 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 178 "TCP SYN cache"); 179 180 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 181 &VNET_NAME(tcp_syncache.bucket_limit), 0, 182 "Per-bucket hash limit for syncache"); 183 184 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 185 &VNET_NAME(tcp_syncache.cache_limit), 0, 186 "Overall entry limit for syncache"); 187 188 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 189 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 190 191 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 192 &VNET_NAME(tcp_syncache.hashsize), 0, 193 "Size of TCP syncache hashtable"); 194 195 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET | 196 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0, 197 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) " 198 "and mac(4) checks"); 199 200 static int 201 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 202 { 203 int error; 204 u_int new; 205 206 new = V_tcp_syncache.rexmt_limit; 207 error = sysctl_handle_int(oidp, &new, 0, req); 208 if ((error == 0) && (req->newptr != NULL)) { 209 if (new > TCP_MAXRXTSHIFT) 210 error = EINVAL; 211 else 212 V_tcp_syncache.rexmt_limit = new; 213 } 214 return (error); 215 } 216 217 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 218 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 219 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 220 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI", 221 "Limit on SYN/ACK retransmissions"); 222 223 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 224 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 225 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 226 "Send reset on socket allocation failure"); 227 228 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 229 230 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 231 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 232 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 233 234 /* 235 * Requires the syncache entry to be already removed from the bucket list. 236 */ 237 static void 238 syncache_free(struct syncache *sc) 239 { 240 241 if (sc->sc_ipopts) 242 (void) m_free(sc->sc_ipopts); 243 if (sc->sc_cred) 244 crfree(sc->sc_cred); 245 #ifdef MAC 246 mac_syncache_destroy(&sc->sc_label); 247 #endif 248 249 uma_zfree(V_tcp_syncache.zone, sc); 250 } 251 252 void 253 syncache_init(void) 254 { 255 int i; 256 257 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 258 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 259 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 260 V_tcp_syncache.hash_secret = arc4random(); 261 262 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 263 &V_tcp_syncache.hashsize); 264 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 265 &V_tcp_syncache.bucket_limit); 266 if (!powerof2(V_tcp_syncache.hashsize) || 267 V_tcp_syncache.hashsize == 0) { 268 printf("WARNING: syncache hash size is not a power of 2.\n"); 269 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 270 } 271 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 272 273 /* Set limits. */ 274 V_tcp_syncache.cache_limit = 275 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 276 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 277 &V_tcp_syncache.cache_limit); 278 279 /* Allocate the hash table. */ 280 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 281 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 282 283 #ifdef VIMAGE 284 V_tcp_syncache.vnet = curvnet; 285 #endif 286 287 /* Initialize the hash buckets. */ 288 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 289 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 290 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 291 NULL, MTX_DEF); 292 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 293 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 294 V_tcp_syncache.hashbase[i].sch_length = 0; 295 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 296 V_tcp_syncache.hashbase[i].sch_last_overflow = 297 -(SYNCOOKIE_LIFETIME + 1); 298 } 299 300 /* Create the syncache entry zone. */ 301 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 302 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 303 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 304 V_tcp_syncache.cache_limit); 305 306 /* Start the SYN cookie reseeder callout. */ 307 callout_init(&V_tcp_syncache.secret.reseed, 1); 308 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 309 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 310 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 311 syncookie_reseed, &V_tcp_syncache); 312 313 /* Initialize the pause machinery. */ 314 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF); 315 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx, 316 0); 317 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME; 318 V_tcp_syncache.pause_backoff = 0; 319 V_tcp_syncache.paused = false; 320 } 321 322 #ifdef VIMAGE 323 void 324 syncache_destroy(void) 325 { 326 struct syncache_head *sch; 327 struct syncache *sc, *nsc; 328 int i; 329 330 /* 331 * Stop the re-seed timer before freeing resources. No need to 332 * possibly schedule it another time. 333 */ 334 callout_drain(&V_tcp_syncache.secret.reseed); 335 336 /* Stop the SYN cache pause callout. */ 337 mtx_lock(&V_tcp_syncache.pause_mtx); 338 if (callout_stop(&V_tcp_syncache.pause_co) == 0) { 339 mtx_unlock(&V_tcp_syncache.pause_mtx); 340 callout_drain(&V_tcp_syncache.pause_co); 341 } else 342 mtx_unlock(&V_tcp_syncache.pause_mtx); 343 344 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 345 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 346 sch = &V_tcp_syncache.hashbase[i]; 347 callout_drain(&sch->sch_timer); 348 349 SCH_LOCK(sch); 350 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 351 syncache_drop(sc, sch); 352 SCH_UNLOCK(sch); 353 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 354 ("%s: sch->sch_bucket not empty", __func__)); 355 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 356 __func__, sch->sch_length)); 357 mtx_destroy(&sch->sch_mtx); 358 } 359 360 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 361 ("%s: cache_count not 0", __func__)); 362 363 /* Free the allocated global resources. */ 364 uma_zdestroy(V_tcp_syncache.zone); 365 free(V_tcp_syncache.hashbase, M_SYNCACHE); 366 mtx_destroy(&V_tcp_syncache.pause_mtx); 367 } 368 #endif 369 370 /* 371 * Inserts a syncache entry into the specified bucket row. 372 * Locks and unlocks the syncache_head autonomously. 373 */ 374 static void 375 syncache_insert(struct syncache *sc, struct syncache_head *sch) 376 { 377 struct syncache *sc2; 378 379 SCH_LOCK(sch); 380 381 /* 382 * Make sure that we don't overflow the per-bucket limit. 383 * If the bucket is full, toss the oldest element. 384 */ 385 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 386 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 387 ("sch->sch_length incorrect")); 388 syncache_pause(&sc->sc_inc); 389 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 390 sch->sch_last_overflow = time_uptime; 391 syncache_drop(sc2, sch); 392 } 393 394 /* Put it into the bucket. */ 395 TAILQ_INSERT_HEAD(&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_added(tod, sc->sc_todctx); 403 } 404 #endif 405 406 /* Reinitialize the bucket row's timer. */ 407 if (sch->sch_length == 1) 408 sch->sch_nextc = ticks + INT_MAX; 409 syncache_timeout(sc, sch, 1); 410 411 SCH_UNLOCK(sch); 412 413 TCPSTATES_INC(TCPS_SYN_RECEIVED); 414 TCPSTAT_INC(tcps_sc_added); 415 } 416 417 /* 418 * Remove and free entry from syncache bucket row. 419 * Expects locked syncache head. 420 */ 421 static void 422 syncache_drop(struct syncache *sc, struct syncache_head *sch) 423 { 424 425 SCH_LOCK_ASSERT(sch); 426 427 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 428 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 429 sch->sch_length--; 430 431 #ifdef TCP_OFFLOAD 432 if (ADDED_BY_TOE(sc)) { 433 struct toedev *tod = sc->sc_tod; 434 435 tod->tod_syncache_removed(tod, sc->sc_todctx); 436 } 437 #endif 438 439 syncache_free(sc); 440 } 441 442 /* 443 * Engage/reengage time on bucket row. 444 */ 445 static void 446 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 447 { 448 int rexmt; 449 450 if (sc->sc_rxmits == 0) 451 rexmt = tcp_rexmit_initial; 452 else 453 TCPT_RANGESET(rexmt, 454 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], 455 tcp_rexmit_min, TCPTV_REXMTMAX); 456 sc->sc_rxttime = ticks + rexmt; 457 sc->sc_rxmits++; 458 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 459 sch->sch_nextc = sc->sc_rxttime; 460 if (docallout) 461 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 462 syncache_timer, (void *)sch); 463 } 464 } 465 466 /* 467 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 468 * If we have retransmitted an entry the maximum number of times, expire it. 469 * One separate timer for each bucket row. 470 */ 471 static void 472 syncache_timer(void *xsch) 473 { 474 struct syncache_head *sch = (struct syncache_head *)xsch; 475 struct syncache *sc, *nsc; 476 struct epoch_tracker et; 477 int tick = ticks; 478 char *s; 479 bool paused; 480 481 CURVNET_SET(sch->sch_sc->vnet); 482 483 /* NB: syncache_head has already been locked by the callout. */ 484 SCH_LOCK_ASSERT(sch); 485 486 /* 487 * In the following cycle we may remove some entries and/or 488 * advance some timeouts, so re-initialize the bucket timer. 489 */ 490 sch->sch_nextc = tick + INT_MAX; 491 492 /* 493 * If we have paused processing, unconditionally remove 494 * all syncache entries. 495 */ 496 mtx_lock(&V_tcp_syncache.pause_mtx); 497 paused = V_tcp_syncache.paused; 498 mtx_unlock(&V_tcp_syncache.pause_mtx); 499 500 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 501 if (paused) { 502 syncache_drop(sc, sch); 503 continue; 504 } 505 /* 506 * We do not check if the listen socket still exists 507 * and accept the case where the listen socket may be 508 * gone by the time we resend the SYN/ACK. We do 509 * not expect this to happens often. If it does, 510 * then the RST will be sent by the time the remote 511 * host does the SYN/ACK->ACK. 512 */ 513 if (TSTMP_GT(sc->sc_rxttime, tick)) { 514 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 515 sch->sch_nextc = sc->sc_rxttime; 516 continue; 517 } 518 if (sc->sc_rxmits > V_tcp_ecn_maxretries) { 519 sc->sc_flags &= ~SCF_ECN; 520 } 521 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 522 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 523 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 524 "giving up and removing syncache entry\n", 525 s, __func__); 526 free(s, M_TCPLOG); 527 } 528 syncache_drop(sc, sch); 529 TCPSTAT_INC(tcps_sc_stale); 530 continue; 531 } 532 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 533 log(LOG_DEBUG, "%s; %s: Response timeout, " 534 "retransmitting (%u) SYN|ACK\n", 535 s, __func__, sc->sc_rxmits); 536 free(s, M_TCPLOG); 537 } 538 539 NET_EPOCH_ENTER(et); 540 syncache_respond(sc, NULL, TH_SYN|TH_ACK); 541 NET_EPOCH_EXIT(et); 542 TCPSTAT_INC(tcps_sc_retransmitted); 543 syncache_timeout(sc, sch, 0); 544 } 545 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 546 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 547 syncache_timer, (void *)(sch)); 548 CURVNET_RESTORE(); 549 } 550 551 /* 552 * Returns true if the system is only using cookies at the moment. 553 * This could be due to a sysadmin decision to only use cookies, or it 554 * could be due to the system detecting an attack. 555 */ 556 static inline bool 557 syncache_cookiesonly(void) 558 { 559 560 return (V_tcp_syncookies && (V_tcp_syncache.paused || 561 V_tcp_syncookiesonly)); 562 } 563 564 /* 565 * Find the hash bucket for the given connection. 566 */ 567 static struct syncache_head * 568 syncache_hashbucket(struct in_conninfo *inc) 569 { 570 uint32_t hash; 571 572 /* 573 * The hash is built on foreign port + local port + foreign address. 574 * We rely on the fact that struct in_conninfo starts with 16 bits 575 * of foreign port, then 16 bits of local port then followed by 128 576 * bits of foreign address. In case of IPv4 address, the first 3 577 * 32-bit words of the address always are zeroes. 578 */ 579 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 580 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 581 582 return (&V_tcp_syncache.hashbase[hash]); 583 } 584 585 /* 586 * Find an entry in the syncache. 587 * Returns always with locked syncache_head plus a matching entry or NULL. 588 */ 589 static struct syncache * 590 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 591 { 592 struct syncache *sc; 593 struct syncache_head *sch; 594 595 *schp = sch = syncache_hashbucket(inc); 596 SCH_LOCK(sch); 597 598 /* Circle through bucket row to find matching entry. */ 599 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 600 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 601 sizeof(struct in_endpoints)) == 0) 602 break; 603 604 return (sc); /* Always returns with locked sch. */ 605 } 606 607 /* 608 * This function is called when we get a RST for a 609 * non-existent connection, so that we can see if the 610 * connection is in the syn cache. If it is, zap it. 611 * If required send a challenge ACK. 612 */ 613 void 614 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m, 615 uint16_t port) 616 { 617 struct syncache *sc; 618 struct syncache_head *sch; 619 char *s = NULL; 620 621 if (syncache_cookiesonly()) 622 return; 623 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 624 SCH_LOCK_ASSERT(sch); 625 626 /* 627 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 628 * See RFC 793 page 65, section SEGMENT ARRIVES. 629 */ 630 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 631 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 632 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 633 "FIN flag set, segment ignored\n", s, __func__); 634 TCPSTAT_INC(tcps_badrst); 635 goto done; 636 } 637 638 /* 639 * No corresponding connection was found in syncache. 640 * If syncookies are enabled and possibly exclusively 641 * used, or we are under memory pressure, a valid RST 642 * may not find a syncache entry. In that case we're 643 * done and no SYN|ACK retransmissions will happen. 644 * Otherwise the RST was misdirected or spoofed. 645 */ 646 if (sc == NULL) { 647 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 648 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 649 "syncache entry (possibly syncookie only), " 650 "segment ignored\n", s, __func__); 651 TCPSTAT_INC(tcps_badrst); 652 goto done; 653 } 654 655 /* The remote UDP encaps port does not match. */ 656 if (sc->sc_port != port) { 657 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 658 log(LOG_DEBUG, "%s; %s: Spurious RST with matching " 659 "syncache entry but non-matching UDP encaps port, " 660 "segment ignored\n", s, __func__); 661 TCPSTAT_INC(tcps_badrst); 662 goto done; 663 } 664 665 /* 666 * If the RST bit is set, check the sequence number to see 667 * if this is a valid reset segment. 668 * 669 * RFC 793 page 37: 670 * In all states except SYN-SENT, all reset (RST) segments 671 * are validated by checking their SEQ-fields. A reset is 672 * valid if its sequence number is in the window. 673 * 674 * RFC 793 page 69: 675 * There are four cases for the acceptability test for an incoming 676 * segment: 677 * 678 * Segment Receive Test 679 * Length Window 680 * ------- ------- ------------------------------------------- 681 * 0 0 SEG.SEQ = RCV.NXT 682 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 683 * >0 0 not acceptable 684 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 685 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 686 * 687 * Note that when receiving a SYN segment in the LISTEN state, 688 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 689 * described in RFC 793, page 66. 690 */ 691 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 692 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 693 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 694 if (V_tcp_insecure_rst || 695 th->th_seq == sc->sc_irs + 1) { 696 syncache_drop(sc, sch); 697 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 698 log(LOG_DEBUG, 699 "%s; %s: Our SYN|ACK was rejected, " 700 "connection attempt aborted by remote " 701 "endpoint\n", 702 s, __func__); 703 TCPSTAT_INC(tcps_sc_reset); 704 } else { 705 TCPSTAT_INC(tcps_badrst); 706 /* Send challenge ACK. */ 707 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 708 log(LOG_DEBUG, "%s; %s: RST with invalid " 709 " SEQ %u != NXT %u (+WND %u), " 710 "sending challenge ACK\n", 711 s, __func__, 712 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 713 syncache_respond(sc, m, TH_ACK); 714 } 715 } else { 716 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 717 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 718 "NXT %u (+WND %u), segment ignored\n", 719 s, __func__, 720 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 721 TCPSTAT_INC(tcps_badrst); 722 } 723 724 done: 725 if (s != NULL) 726 free(s, M_TCPLOG); 727 SCH_UNLOCK(sch); 728 } 729 730 void 731 syncache_badack(struct in_conninfo *inc, uint16_t port) 732 { 733 struct syncache *sc; 734 struct syncache_head *sch; 735 736 if (syncache_cookiesonly()) 737 return; 738 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 739 SCH_LOCK_ASSERT(sch); 740 if ((sc != NULL) && (sc->sc_port == port)) { 741 syncache_drop(sc, sch); 742 TCPSTAT_INC(tcps_sc_badack); 743 } 744 SCH_UNLOCK(sch); 745 } 746 747 void 748 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port) 749 { 750 struct syncache *sc; 751 struct syncache_head *sch; 752 753 if (syncache_cookiesonly()) 754 return; 755 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 756 SCH_LOCK_ASSERT(sch); 757 if (sc == NULL) 758 goto done; 759 760 /* If the port != sc_port, then it's a bogus ICMP msg */ 761 if (port != sc->sc_port) 762 goto done; 763 764 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 765 if (ntohl(th_seq) != sc->sc_iss) 766 goto done; 767 768 /* 769 * If we've rertransmitted 3 times and this is our second error, 770 * we remove the entry. Otherwise, we allow it to continue on. 771 * This prevents us from incorrectly nuking an entry during a 772 * spurious network outage. 773 * 774 * See tcp_notify(). 775 */ 776 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 777 sc->sc_flags |= SCF_UNREACH; 778 goto done; 779 } 780 syncache_drop(sc, sch); 781 TCPSTAT_INC(tcps_sc_unreach); 782 done: 783 SCH_UNLOCK(sch); 784 } 785 786 /* 787 * Build a new TCP socket structure from a syncache entry. 788 * 789 * On success return the newly created socket with its underlying inp locked. 790 */ 791 static struct socket * 792 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 793 { 794 struct tcp_function_block *blk; 795 struct inpcb *inp = NULL; 796 struct socket *so; 797 struct tcpcb *tp; 798 int error; 799 char *s; 800 801 NET_EPOCH_ASSERT(); 802 803 /* 804 * Ok, create the full blown connection, and set things up 805 * as they would have been set up if we had created the 806 * connection when the SYN arrived. If we can't create 807 * the connection, abort it. 808 */ 809 so = sonewconn(lso, 0); 810 if (so == NULL) { 811 /* 812 * Drop the connection; we will either send a RST or 813 * have the peer retransmit its SYN again after its 814 * RTO and try again. 815 */ 816 TCPSTAT_INC(tcps_listendrop); 817 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 818 log(LOG_DEBUG, "%s; %s: Socket create failed " 819 "due to limits or memory shortage\n", 820 s, __func__); 821 free(s, M_TCPLOG); 822 } 823 goto abort2; 824 } 825 #ifdef MAC 826 mac_socketpeer_set_from_mbuf(m, so); 827 #endif 828 829 inp = sotoinpcb(so); 830 inp->inp_inc.inc_fibnum = so->so_fibnum; 831 INP_WLOCK(inp); 832 /* 833 * Exclusive pcbinfo lock is not required in syncache socket case even 834 * if two inpcb locks can be acquired simultaneously: 835 * - the inpcb in LISTEN state, 836 * - the newly created inp. 837 * 838 * In this case, an inp cannot be at same time in LISTEN state and 839 * just created by an accept() call. 840 */ 841 INP_HASH_WLOCK(&V_tcbinfo); 842 843 /* Insert new socket into PCB hash list. */ 844 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 845 #ifdef INET6 846 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 847 inp->inp_vflag &= ~INP_IPV4; 848 inp->inp_vflag |= INP_IPV6; 849 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 850 } else { 851 inp->inp_vflag &= ~INP_IPV6; 852 inp->inp_vflag |= INP_IPV4; 853 #endif 854 inp->inp_ip_ttl = sc->sc_ip_ttl; 855 inp->inp_ip_tos = sc->sc_ip_tos; 856 inp->inp_laddr = sc->sc_inc.inc_laddr; 857 #ifdef INET6 858 } 859 #endif 860 861 /* 862 * If there's an mbuf and it has a flowid, then let's initialise the 863 * inp with that particular flowid. 864 */ 865 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 866 inp->inp_flowid = m->m_pkthdr.flowid; 867 inp->inp_flowtype = M_HASHTYPE_GET(m); 868 #ifdef NUMA 869 inp->inp_numa_domain = m->m_pkthdr.numa_domain; 870 #endif 871 } 872 873 inp->inp_lport = sc->sc_inc.inc_lport; 874 #ifdef INET6 875 if (inp->inp_vflag & INP_IPV6PROTO) { 876 struct inpcb *oinp = sotoinpcb(lso); 877 878 /* 879 * Inherit socket options from the listening socket. 880 * Note that in6p_inputopts are not (and should not be) 881 * copied, since it stores previously received options and is 882 * used to detect if each new option is different than the 883 * previous one and hence should be passed to a user. 884 * If we copied in6p_inputopts, a user would not be able to 885 * receive options just after calling the accept system call. 886 */ 887 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 888 if (oinp->in6p_outputopts) 889 inp->in6p_outputopts = 890 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 891 inp->in6p_hops = oinp->in6p_hops; 892 } 893 894 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 895 struct in6_addr laddr6; 896 struct sockaddr_in6 sin6; 897 898 sin6.sin6_family = AF_INET6; 899 sin6.sin6_len = sizeof(sin6); 900 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 901 sin6.sin6_port = sc->sc_inc.inc_fport; 902 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 903 laddr6 = inp->in6p_laddr; 904 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 905 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 906 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 907 thread0.td_ucred, m, false)) != 0) { 908 inp->in6p_laddr = laddr6; 909 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 910 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 911 "with error %i\n", 912 s, __func__, error); 913 free(s, M_TCPLOG); 914 } 915 INP_HASH_WUNLOCK(&V_tcbinfo); 916 goto abort; 917 } 918 /* Override flowlabel from in6_pcbconnect. */ 919 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 920 inp->inp_flow |= sc->sc_flowlabel; 921 } 922 #endif /* INET6 */ 923 #if defined(INET) && defined(INET6) 924 else 925 #endif 926 #ifdef INET 927 { 928 struct in_addr laddr; 929 struct sockaddr_in sin; 930 931 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 932 933 if (inp->inp_options == NULL) { 934 inp->inp_options = sc->sc_ipopts; 935 sc->sc_ipopts = NULL; 936 } 937 938 sin.sin_family = AF_INET; 939 sin.sin_len = sizeof(sin); 940 sin.sin_addr = sc->sc_inc.inc_faddr; 941 sin.sin_port = sc->sc_inc.inc_fport; 942 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 943 laddr = inp->inp_laddr; 944 if (inp->inp_laddr.s_addr == INADDR_ANY) 945 inp->inp_laddr = sc->sc_inc.inc_laddr; 946 if ((error = in_pcbconnect(inp, (struct sockaddr *)&sin, 947 thread0.td_ucred, false)) != 0) { 948 inp->inp_laddr = laddr; 949 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 950 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 951 "with error %i\n", 952 s, __func__, error); 953 free(s, M_TCPLOG); 954 } 955 INP_HASH_WUNLOCK(&V_tcbinfo); 956 goto abort; 957 } 958 } 959 #endif /* INET */ 960 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 961 /* Copy old policy into new socket's. */ 962 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 963 printf("syncache_socket: could not copy policy\n"); 964 #endif 965 INP_HASH_WUNLOCK(&V_tcbinfo); 966 tp = intotcpcb(inp); 967 tcp_state_change(tp, TCPS_SYN_RECEIVED); 968 tp->iss = sc->sc_iss; 969 tp->irs = sc->sc_irs; 970 tp->t_port = sc->sc_port; 971 tcp_rcvseqinit(tp); 972 tcp_sendseqinit(tp); 973 blk = sototcpcb(lso)->t_fb; 974 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 975 /* 976 * Our parents t_fb was not the default, 977 * we need to release our ref on tp->t_fb and 978 * pickup one on the new entry. 979 */ 980 struct tcp_function_block *rblk; 981 982 rblk = find_and_ref_tcp_fb(blk); 983 KASSERT(rblk != NULL, 984 ("cannot find blk %p out of syncache?", blk)); 985 if (tp->t_fb->tfb_tcp_fb_fini) 986 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 987 refcount_release(&tp->t_fb->tfb_refcnt); 988 tp->t_fb = rblk; 989 /* 990 * XXXrrs this is quite dangerous, it is possible 991 * for the new function to fail to init. We also 992 * are not asking if the handoff_is_ok though at 993 * the very start thats probalbly ok. 994 */ 995 if (tp->t_fb->tfb_tcp_fb_init) { 996 (*tp->t_fb->tfb_tcp_fb_init)(tp); 997 } 998 } 999 tp->snd_wl1 = sc->sc_irs; 1000 tp->snd_max = tp->iss + 1; 1001 tp->snd_nxt = tp->iss + 1; 1002 tp->rcv_up = sc->sc_irs + 1; 1003 tp->rcv_wnd = sc->sc_wnd; 1004 tp->rcv_adv += tp->rcv_wnd; 1005 tp->last_ack_sent = tp->rcv_nxt; 1006 1007 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 1008 if (sc->sc_flags & SCF_NOOPT) 1009 tp->t_flags |= TF_NOOPT; 1010 else { 1011 if (sc->sc_flags & SCF_WINSCALE) { 1012 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 1013 tp->snd_scale = sc->sc_requested_s_scale; 1014 tp->request_r_scale = sc->sc_requested_r_scale; 1015 } 1016 if (sc->sc_flags & SCF_TIMESTAMP) { 1017 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 1018 tp->ts_recent = sc->sc_tsreflect; 1019 tp->ts_recent_age = tcp_ts_getticks(); 1020 tp->ts_offset = sc->sc_tsoff; 1021 } 1022 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1023 if (sc->sc_flags & SCF_SIGNATURE) 1024 tp->t_flags |= TF_SIGNATURE; 1025 #endif 1026 if (sc->sc_flags & SCF_SACK) 1027 tp->t_flags |= TF_SACK_PERMIT; 1028 } 1029 1030 if (sc->sc_flags & SCF_ECN) 1031 tp->t_flags2 |= TF2_ECN_PERMIT; 1032 1033 /* 1034 * Set up MSS and get cached values from tcp_hostcache. 1035 * This might overwrite some of the defaults we just set. 1036 */ 1037 tcp_mss(tp, sc->sc_peer_mss); 1038 1039 /* 1040 * If the SYN,ACK was retransmitted, indicate that CWND to be 1041 * limited to one segment in cc_conn_init(). 1042 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 1043 */ 1044 if (sc->sc_rxmits > 1) 1045 tp->snd_cwnd = 1; 1046 1047 #ifdef TCP_OFFLOAD 1048 /* 1049 * Allow a TOE driver to install its hooks. Note that we hold the 1050 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 1051 * new connection before the TOE driver has done its thing. 1052 */ 1053 if (ADDED_BY_TOE(sc)) { 1054 struct toedev *tod = sc->sc_tod; 1055 1056 tod->tod_offload_socket(tod, sc->sc_todctx, so); 1057 } 1058 #endif 1059 /* 1060 * Copy and activate timers. 1061 */ 1062 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 1063 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 1064 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1065 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1066 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1067 1068 TCPSTAT_INC(tcps_accepts); 1069 return (so); 1070 1071 abort: 1072 INP_WUNLOCK(inp); 1073 abort2: 1074 if (so != NULL) 1075 soabort(so); 1076 return (NULL); 1077 } 1078 1079 /* 1080 * This function gets called when we receive an ACK for a 1081 * socket in the LISTEN state. We look up the connection 1082 * in the syncache, and if its there, we pull it out of 1083 * the cache and turn it into a full-blown connection in 1084 * the SYN-RECEIVED state. 1085 * 1086 * On syncache_socket() success the newly created socket 1087 * has its underlying inp locked. 1088 */ 1089 int 1090 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1091 struct socket **lsop, struct mbuf *m, uint16_t port) 1092 { 1093 struct syncache *sc; 1094 struct syncache_head *sch; 1095 struct syncache scs; 1096 char *s; 1097 bool locked; 1098 1099 NET_EPOCH_ASSERT(); 1100 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1101 ("%s: can handle only ACK", __func__)); 1102 1103 if (syncache_cookiesonly()) { 1104 sc = NULL; 1105 sch = syncache_hashbucket(inc); 1106 locked = false; 1107 } else { 1108 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1109 locked = true; 1110 SCH_LOCK_ASSERT(sch); 1111 } 1112 1113 #ifdef INVARIANTS 1114 /* 1115 * Test code for syncookies comparing the syncache stored 1116 * values with the reconstructed values from the cookie. 1117 */ 1118 if (sc != NULL) 1119 syncookie_cmp(inc, sch, sc, th, to, *lsop, port); 1120 #endif 1121 1122 if (sc == NULL) { 1123 /* 1124 * There is no syncache entry, so see if this ACK is 1125 * a returning syncookie. To do this, first: 1126 * A. Check if syncookies are used in case of syncache 1127 * overflows 1128 * B. See if this socket has had a syncache entry dropped in 1129 * the recent past. We don't want to accept a bogus 1130 * syncookie if we've never received a SYN or accept it 1131 * twice. 1132 * C. check that the syncookie is valid. If it is, then 1133 * cobble up a fake syncache entry, and return. 1134 */ 1135 if (locked && !V_tcp_syncookies) { 1136 SCH_UNLOCK(sch); 1137 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1138 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1139 "segment rejected (syncookies disabled)\n", 1140 s, __func__); 1141 goto failed; 1142 } 1143 if (locked && !V_tcp_syncookiesonly && 1144 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1145 SCH_UNLOCK(sch); 1146 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1147 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1148 "segment rejected (no syncache entry)\n", 1149 s, __func__); 1150 goto failed; 1151 } 1152 bzero(&scs, sizeof(scs)); 1153 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port); 1154 if (locked) 1155 SCH_UNLOCK(sch); 1156 if (sc == NULL) { 1157 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1158 log(LOG_DEBUG, "%s; %s: Segment failed " 1159 "SYNCOOKIE authentication, segment rejected " 1160 "(probably spoofed)\n", s, __func__); 1161 goto failed; 1162 } 1163 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1164 /* If received ACK has MD5 signature, check it. */ 1165 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1166 (!TCPMD5_ENABLED() || 1167 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1168 /* Drop the ACK. */ 1169 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1170 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1171 "MD5 signature doesn't match.\n", 1172 s, __func__); 1173 free(s, M_TCPLOG); 1174 } 1175 TCPSTAT_INC(tcps_sig_err_sigopt); 1176 return (-1); /* Do not send RST */ 1177 } 1178 #endif /* TCP_SIGNATURE */ 1179 } else { 1180 if (sc->sc_port != port) { 1181 SCH_UNLOCK(sch); 1182 return (0); 1183 } 1184 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1185 /* 1186 * If listening socket requested TCP digests, check that 1187 * received ACK has signature and it is correct. 1188 * If not, drop the ACK and leave sc entry in th cache, 1189 * because SYN was received with correct signature. 1190 */ 1191 if (sc->sc_flags & SCF_SIGNATURE) { 1192 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1193 /* No signature */ 1194 TCPSTAT_INC(tcps_sig_err_nosigopt); 1195 SCH_UNLOCK(sch); 1196 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1197 log(LOG_DEBUG, "%s; %s: Segment " 1198 "rejected, MD5 signature wasn't " 1199 "provided.\n", s, __func__); 1200 free(s, M_TCPLOG); 1201 } 1202 return (-1); /* Do not send RST */ 1203 } 1204 if (!TCPMD5_ENABLED() || 1205 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1206 /* Doesn't match or no SA */ 1207 SCH_UNLOCK(sch); 1208 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1209 log(LOG_DEBUG, "%s; %s: Segment " 1210 "rejected, MD5 signature doesn't " 1211 "match.\n", s, __func__); 1212 free(s, M_TCPLOG); 1213 } 1214 return (-1); /* Do not send RST */ 1215 } 1216 } 1217 #endif /* TCP_SIGNATURE */ 1218 1219 /* 1220 * RFC 7323 PAWS: If we have a timestamp on this segment and 1221 * it's less than ts_recent, drop it. 1222 * XXXMT: RFC 7323 also requires to send an ACK. 1223 * In tcp_input.c this is only done for TCP segments 1224 * with user data, so be consistent here and just drop 1225 * the segment. 1226 */ 1227 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1228 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1229 SCH_UNLOCK(sch); 1230 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1231 log(LOG_DEBUG, 1232 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1233 "segment dropped\n", s, __func__, 1234 to->to_tsval, sc->sc_tsreflect); 1235 free(s, M_TCPLOG); 1236 } 1237 return (-1); /* Do not send RST */ 1238 } 1239 1240 /* 1241 * If timestamps were not negotiated during SYN/ACK and a 1242 * segment with a timestamp is received, ignore the 1243 * timestamp and process the packet normally. 1244 * See section 3.2 of RFC 7323. 1245 */ 1246 if (!(sc->sc_flags & SCF_TIMESTAMP) && 1247 (to->to_flags & TOF_TS)) { 1248 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1249 log(LOG_DEBUG, "%s; %s: Timestamp not " 1250 "expected, segment processed normally\n", 1251 s, __func__); 1252 free(s, M_TCPLOG); 1253 s = NULL; 1254 } 1255 } 1256 1257 /* 1258 * If timestamps were negotiated during SYN/ACK and a 1259 * segment without a timestamp is received, silently drop 1260 * the segment, unless the missing timestamps are tolerated. 1261 * See section 3.2 of RFC 7323. 1262 */ 1263 if ((sc->sc_flags & SCF_TIMESTAMP) && 1264 !(to->to_flags & TOF_TS)) { 1265 if (V_tcp_tolerate_missing_ts) { 1266 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1267 log(LOG_DEBUG, 1268 "%s; %s: Timestamp missing, " 1269 "segment processed normally\n", 1270 s, __func__); 1271 free(s, M_TCPLOG); 1272 } 1273 } else { 1274 SCH_UNLOCK(sch); 1275 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1276 log(LOG_DEBUG, 1277 "%s; %s: Timestamp missing, " 1278 "segment silently dropped\n", 1279 s, __func__); 1280 free(s, M_TCPLOG); 1281 } 1282 return (-1); /* Do not send RST */ 1283 } 1284 } 1285 1286 /* 1287 * Pull out the entry to unlock the bucket row. 1288 * 1289 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1290 * tcp_state_change(). The tcpcb is not existent at this 1291 * moment. A new one will be allocated via syncache_socket-> 1292 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1293 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1294 */ 1295 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1296 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1297 sch->sch_length--; 1298 #ifdef TCP_OFFLOAD 1299 if (ADDED_BY_TOE(sc)) { 1300 struct toedev *tod = sc->sc_tod; 1301 1302 tod->tod_syncache_removed(tod, sc->sc_todctx); 1303 } 1304 #endif 1305 SCH_UNLOCK(sch); 1306 } 1307 1308 /* 1309 * Segment validation: 1310 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1311 */ 1312 if (th->th_ack != sc->sc_iss + 1) { 1313 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1314 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1315 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1316 goto failed; 1317 } 1318 1319 /* 1320 * The SEQ must fall in the window starting at the received 1321 * initial receive sequence number + 1 (the SYN). 1322 */ 1323 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1324 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1325 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1326 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1327 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1328 goto failed; 1329 } 1330 1331 *lsop = syncache_socket(sc, *lsop, m); 1332 1333 if (*lsop == NULL) 1334 TCPSTAT_INC(tcps_sc_aborted); 1335 else 1336 TCPSTAT_INC(tcps_sc_completed); 1337 1338 /* how do we find the inp for the new socket? */ 1339 if (sc != &scs) 1340 syncache_free(sc); 1341 return (1); 1342 failed: 1343 if (sc != NULL && sc != &scs) 1344 syncache_free(sc); 1345 if (s != NULL) 1346 free(s, M_TCPLOG); 1347 *lsop = NULL; 1348 return (0); 1349 } 1350 1351 static struct socket * 1352 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m, 1353 uint64_t response_cookie) 1354 { 1355 struct inpcb *inp; 1356 struct tcpcb *tp; 1357 unsigned int *pending_counter; 1358 struct socket *so; 1359 1360 NET_EPOCH_ASSERT(); 1361 1362 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending; 1363 so = syncache_socket(sc, lso, m); 1364 if (so == NULL) { 1365 TCPSTAT_INC(tcps_sc_aborted); 1366 atomic_subtract_int(pending_counter, 1); 1367 } else { 1368 soisconnected(so); 1369 inp = sotoinpcb(so); 1370 tp = intotcpcb(inp); 1371 tp->t_flags |= TF_FASTOPEN; 1372 tp->t_tfo_cookie.server = response_cookie; 1373 tp->snd_max = tp->iss; 1374 tp->snd_nxt = tp->iss; 1375 tp->t_tfo_pending = pending_counter; 1376 TCPSTAT_INC(tcps_sc_completed); 1377 } 1378 1379 return (so); 1380 } 1381 1382 /* 1383 * Given a LISTEN socket and an inbound SYN request, add 1384 * this to the syn cache, and send back a segment: 1385 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1386 * to the source. 1387 * 1388 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1389 * Doing so would require that we hold onto the data and deliver it 1390 * to the application. However, if we are the target of a SYN-flood 1391 * DoS attack, an attacker could send data which would eventually 1392 * consume all available buffer space if it were ACKed. By not ACKing 1393 * the data, we avoid this DoS scenario. 1394 * 1395 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1396 * cookie is processed and a new socket is created. In this case, any data 1397 * accompanying the SYN will be queued to the socket by tcp_input() and will 1398 * be ACKed either when the application sends response data or the delayed 1399 * ACK timer expires, whichever comes first. 1400 */ 1401 struct socket * 1402 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1403 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod, 1404 void *todctx, uint8_t iptos, uint16_t port) 1405 { 1406 struct tcpcb *tp; 1407 struct socket *rv = NULL; 1408 struct syncache *sc = NULL; 1409 struct syncache_head *sch; 1410 struct mbuf *ipopts = NULL; 1411 u_int ltflags; 1412 int win, ip_ttl, ip_tos; 1413 char *s; 1414 #ifdef INET6 1415 int autoflowlabel = 0; 1416 #endif 1417 #ifdef MAC 1418 struct label *maclabel; 1419 #endif 1420 struct syncache scs; 1421 struct ucred *cred; 1422 uint64_t tfo_response_cookie; 1423 unsigned int *tfo_pending = NULL; 1424 int tfo_cookie_valid = 0; 1425 int tfo_response_cookie_valid = 0; 1426 bool locked; 1427 1428 INP_RLOCK_ASSERT(inp); /* listen socket */ 1429 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1430 ("%s: unexpected tcp flags", __func__)); 1431 1432 /* 1433 * Combine all so/tp operations very early to drop the INP lock as 1434 * soon as possible. 1435 */ 1436 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1437 tp = sototcpcb(so); 1438 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred); 1439 1440 #ifdef INET6 1441 if (inc->inc_flags & INC_ISIPV6) { 1442 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1443 autoflowlabel = 1; 1444 } 1445 ip_ttl = in6_selecthlim(inp, NULL); 1446 if ((inp->in6p_outputopts == NULL) || 1447 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1448 ip_tos = 0; 1449 } else { 1450 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1451 } 1452 } 1453 #endif 1454 #if defined(INET6) && defined(INET) 1455 else 1456 #endif 1457 #ifdef INET 1458 { 1459 ip_ttl = inp->inp_ip_ttl; 1460 ip_tos = inp->inp_ip_tos; 1461 } 1462 #endif 1463 win = so->sol_sbrcv_hiwat; 1464 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1465 1466 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1467 (tp->t_tfo_pending != NULL) && 1468 (to->to_flags & TOF_FASTOPEN)) { 1469 /* 1470 * Limit the number of pending TFO connections to 1471 * approximately half of the queue limit. This prevents TFO 1472 * SYN floods from starving the service by filling the 1473 * listen queue with bogus TFO connections. 1474 */ 1475 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1476 (so->sol_qlimit / 2)) { 1477 int result; 1478 1479 result = tcp_fastopen_check_cookie(inc, 1480 to->to_tfo_cookie, to->to_tfo_len, 1481 &tfo_response_cookie); 1482 tfo_cookie_valid = (result > 0); 1483 tfo_response_cookie_valid = (result >= 0); 1484 } 1485 1486 /* 1487 * Remember the TFO pending counter as it will have to be 1488 * decremented below if we don't make it to syncache_tfo_expand(). 1489 */ 1490 tfo_pending = tp->t_tfo_pending; 1491 } 1492 1493 #ifdef MAC 1494 if (mac_syncache_init(&maclabel) != 0) { 1495 INP_RUNLOCK(inp); 1496 goto done; 1497 } else 1498 mac_syncache_create(maclabel, inp); 1499 #endif 1500 if (!tfo_cookie_valid) 1501 INP_RUNLOCK(inp); 1502 1503 /* 1504 * Remember the IP options, if any. 1505 */ 1506 #ifdef INET6 1507 if (!(inc->inc_flags & INC_ISIPV6)) 1508 #endif 1509 #ifdef INET 1510 ipopts = (m) ? ip_srcroute(m) : NULL; 1511 #else 1512 ipopts = NULL; 1513 #endif 1514 1515 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1516 /* 1517 * If listening socket requested TCP digests, check that received 1518 * SYN has signature and it is correct. If signature doesn't match 1519 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1520 */ 1521 if (ltflags & TF_SIGNATURE) { 1522 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1523 TCPSTAT_INC(tcps_sig_err_nosigopt); 1524 goto done; 1525 } 1526 if (!TCPMD5_ENABLED() || 1527 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1528 goto done; 1529 } 1530 #endif /* TCP_SIGNATURE */ 1531 /* 1532 * See if we already have an entry for this connection. 1533 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1534 * 1535 * XXX: should the syncache be re-initialized with the contents 1536 * of the new SYN here (which may have different options?) 1537 * 1538 * XXX: We do not check the sequence number to see if this is a 1539 * real retransmit or a new connection attempt. The question is 1540 * how to handle such a case; either ignore it as spoofed, or 1541 * drop the current entry and create a new one? 1542 */ 1543 if (syncache_cookiesonly()) { 1544 sc = NULL; 1545 sch = syncache_hashbucket(inc); 1546 locked = false; 1547 } else { 1548 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1549 locked = true; 1550 SCH_LOCK_ASSERT(sch); 1551 } 1552 if (sc != NULL) { 1553 if (tfo_cookie_valid) 1554 INP_RUNLOCK(inp); 1555 TCPSTAT_INC(tcps_sc_dupsyn); 1556 if (ipopts) { 1557 /* 1558 * If we were remembering a previous source route, 1559 * forget it and use the new one we've been given. 1560 */ 1561 if (sc->sc_ipopts) 1562 (void) m_free(sc->sc_ipopts); 1563 sc->sc_ipopts = ipopts; 1564 } 1565 /* 1566 * Update timestamp if present. 1567 */ 1568 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1569 sc->sc_tsreflect = to->to_tsval; 1570 else 1571 sc->sc_flags &= ~SCF_TIMESTAMP; 1572 /* 1573 * Disable ECN if needed. 1574 */ 1575 if ((sc->sc_flags & SCF_ECN) && 1576 ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) { 1577 sc->sc_flags &= ~SCF_ECN; 1578 } 1579 #ifdef MAC 1580 /* 1581 * Since we have already unconditionally allocated label 1582 * storage, free it up. The syncache entry will already 1583 * have an initialized label we can use. 1584 */ 1585 mac_syncache_destroy(&maclabel); 1586 #endif 1587 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1588 /* Retransmit SYN|ACK and reset retransmit count. */ 1589 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1590 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1591 "resetting timer and retransmitting SYN|ACK\n", 1592 s, __func__); 1593 free(s, M_TCPLOG); 1594 } 1595 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1596 sc->sc_rxmits = 0; 1597 syncache_timeout(sc, sch, 1); 1598 TCPSTAT_INC(tcps_sndacks); 1599 TCPSTAT_INC(tcps_sndtotal); 1600 } 1601 SCH_UNLOCK(sch); 1602 goto donenoprobe; 1603 } 1604 1605 if (tfo_cookie_valid) { 1606 bzero(&scs, sizeof(scs)); 1607 sc = &scs; 1608 goto skip_alloc; 1609 } 1610 1611 /* 1612 * Skip allocating a syncache entry if we are just going to discard 1613 * it later. 1614 */ 1615 if (!locked) { 1616 bzero(&scs, sizeof(scs)); 1617 sc = &scs; 1618 } else 1619 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1620 if (sc == NULL) { 1621 /* 1622 * The zone allocator couldn't provide more entries. 1623 * Treat this as if the cache was full; drop the oldest 1624 * entry and insert the new one. 1625 */ 1626 TCPSTAT_INC(tcps_sc_zonefail); 1627 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1628 sch->sch_last_overflow = time_uptime; 1629 syncache_drop(sc, sch); 1630 syncache_pause(inc); 1631 } 1632 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1633 if (sc == NULL) { 1634 if (V_tcp_syncookies) { 1635 bzero(&scs, sizeof(scs)); 1636 sc = &scs; 1637 } else { 1638 KASSERT(locked, 1639 ("%s: bucket unexpectedly unlocked", 1640 __func__)); 1641 SCH_UNLOCK(sch); 1642 if (ipopts) 1643 (void) m_free(ipopts); 1644 goto done; 1645 } 1646 } 1647 } 1648 1649 skip_alloc: 1650 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1651 sc->sc_tfo_cookie = &tfo_response_cookie; 1652 1653 /* 1654 * Fill in the syncache values. 1655 */ 1656 #ifdef MAC 1657 sc->sc_label = maclabel; 1658 #endif 1659 sc->sc_cred = cred; 1660 sc->sc_port = port; 1661 cred = NULL; 1662 sc->sc_ipopts = ipopts; 1663 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1664 sc->sc_ip_tos = ip_tos; 1665 sc->sc_ip_ttl = ip_ttl; 1666 #ifdef TCP_OFFLOAD 1667 sc->sc_tod = tod; 1668 sc->sc_todctx = todctx; 1669 #endif 1670 sc->sc_irs = th->th_seq; 1671 sc->sc_flags = 0; 1672 sc->sc_flowlabel = 0; 1673 1674 /* 1675 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1676 * win was derived from socket earlier in the function. 1677 */ 1678 win = imax(win, 0); 1679 win = imin(win, TCP_MAXWIN); 1680 sc->sc_wnd = win; 1681 1682 if (V_tcp_do_rfc1323 && 1683 !(ltflags & TF_NOOPT)) { 1684 /* 1685 * A timestamp received in a SYN makes 1686 * it ok to send timestamp requests and replies. 1687 */ 1688 if (to->to_flags & TOF_TS) { 1689 sc->sc_tsreflect = to->to_tsval; 1690 sc->sc_flags |= SCF_TIMESTAMP; 1691 sc->sc_tsoff = tcp_new_ts_offset(inc); 1692 } 1693 if (to->to_flags & TOF_SCALE) { 1694 int wscale = 0; 1695 1696 /* 1697 * Pick the smallest possible scaling factor that 1698 * will still allow us to scale up to sb_max, aka 1699 * kern.ipc.maxsockbuf. 1700 * 1701 * We do this because there are broken firewalls that 1702 * will corrupt the window scale option, leading to 1703 * the other endpoint believing that our advertised 1704 * window is unscaled. At scale factors larger than 1705 * 5 the unscaled window will drop below 1500 bytes, 1706 * leading to serious problems when traversing these 1707 * broken firewalls. 1708 * 1709 * With the default maxsockbuf of 256K, a scale factor 1710 * of 3 will be chosen by this algorithm. Those who 1711 * choose a larger maxsockbuf should watch out 1712 * for the compatibility problems mentioned above. 1713 * 1714 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1715 * or <SYN,ACK>) segment itself is never scaled. 1716 */ 1717 while (wscale < TCP_MAX_WINSHIFT && 1718 (TCP_MAXWIN << wscale) < sb_max) 1719 wscale++; 1720 sc->sc_requested_r_scale = wscale; 1721 sc->sc_requested_s_scale = to->to_wscale; 1722 sc->sc_flags |= SCF_WINSCALE; 1723 } 1724 } 1725 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1726 /* 1727 * If listening socket requested TCP digests, flag this in the 1728 * syncache so that syncache_respond() will do the right thing 1729 * with the SYN+ACK. 1730 */ 1731 if (ltflags & TF_SIGNATURE) 1732 sc->sc_flags |= SCF_SIGNATURE; 1733 #endif /* TCP_SIGNATURE */ 1734 if (to->to_flags & TOF_SACKPERM) 1735 sc->sc_flags |= SCF_SACK; 1736 if (to->to_flags & TOF_MSS) 1737 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1738 if (ltflags & TF_NOOPT) 1739 sc->sc_flags |= SCF_NOOPT; 1740 if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) && 1741 V_tcp_do_ecn) 1742 sc->sc_flags |= SCF_ECN; 1743 1744 if (V_tcp_syncookies) 1745 sc->sc_iss = syncookie_generate(sch, sc); 1746 else 1747 sc->sc_iss = arc4random(); 1748 #ifdef INET6 1749 if (autoflowlabel) { 1750 if (V_tcp_syncookies) 1751 sc->sc_flowlabel = sc->sc_iss; 1752 else 1753 sc->sc_flowlabel = ip6_randomflowlabel(); 1754 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1755 } 1756 #endif 1757 if (locked) 1758 SCH_UNLOCK(sch); 1759 1760 if (tfo_cookie_valid) { 1761 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie); 1762 /* INP_RUNLOCK(inp) will be performed by the caller */ 1763 goto tfo_expanded; 1764 } 1765 1766 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1767 /* 1768 * Do a standard 3-way handshake. 1769 */ 1770 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1771 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1772 syncache_free(sc); 1773 else if (sc != &scs) 1774 syncache_insert(sc, sch); /* locks and unlocks sch */ 1775 TCPSTAT_INC(tcps_sndacks); 1776 TCPSTAT_INC(tcps_sndtotal); 1777 } else { 1778 if (sc != &scs) 1779 syncache_free(sc); 1780 TCPSTAT_INC(tcps_sc_dropped); 1781 } 1782 goto donenoprobe; 1783 1784 done: 1785 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1786 donenoprobe: 1787 if (m) 1788 m_freem(m); 1789 /* 1790 * If tfo_pending is not NULL here, then a TFO SYN that did not 1791 * result in a new socket was processed and the associated pending 1792 * counter has not yet been decremented. All such TFO processing paths 1793 * transit this point. 1794 */ 1795 if (tfo_pending != NULL) 1796 tcp_fastopen_decrement_counter(tfo_pending); 1797 1798 tfo_expanded: 1799 if (cred != NULL) 1800 crfree(cred); 1801 #ifdef MAC 1802 if (sc == &scs) 1803 mac_syncache_destroy(&maclabel); 1804 #endif 1805 return (rv); 1806 } 1807 1808 /* 1809 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1810 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1811 */ 1812 static int 1813 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags) 1814 { 1815 struct ip *ip = NULL; 1816 struct mbuf *m; 1817 struct tcphdr *th = NULL; 1818 struct udphdr *udp = NULL; 1819 int optlen, error = 0; /* Make compiler happy */ 1820 u_int16_t hlen, tlen, mssopt, ulen; 1821 struct tcpopt to; 1822 #ifdef INET6 1823 struct ip6_hdr *ip6 = NULL; 1824 #endif 1825 1826 NET_EPOCH_ASSERT(); 1827 1828 hlen = 1829 #ifdef INET6 1830 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1831 #endif 1832 sizeof(struct ip); 1833 tlen = hlen + sizeof(struct tcphdr); 1834 if (sc->sc_port) { 1835 tlen += sizeof(struct udphdr); 1836 } 1837 /* Determine MSS we advertize to other end of connection. */ 1838 mssopt = tcp_mssopt(&sc->sc_inc); 1839 if (sc->sc_port) 1840 mssopt -= V_tcp_udp_tunneling_overhead; 1841 mssopt = max(mssopt, V_tcp_minmss); 1842 1843 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1844 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1845 ("syncache: mbuf too small")); 1846 1847 /* Create the IP+TCP header from scratch. */ 1848 m = m_gethdr(M_NOWAIT, MT_DATA); 1849 if (m == NULL) 1850 return (ENOBUFS); 1851 #ifdef MAC 1852 mac_syncache_create_mbuf(sc->sc_label, m); 1853 #endif 1854 m->m_data += max_linkhdr; 1855 m->m_len = tlen; 1856 m->m_pkthdr.len = tlen; 1857 m->m_pkthdr.rcvif = NULL; 1858 1859 #ifdef INET6 1860 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1861 ip6 = mtod(m, struct ip6_hdr *); 1862 ip6->ip6_vfc = IPV6_VERSION; 1863 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1864 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1865 ip6->ip6_plen = htons(tlen - hlen); 1866 /* ip6_hlim is set after checksum */ 1867 /* Zero out traffic class and flow label. */ 1868 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1869 ip6->ip6_flow |= sc->sc_flowlabel; 1870 if (sc->sc_port != 0) { 1871 ip6->ip6_nxt = IPPROTO_UDP; 1872 udp = (struct udphdr *)(ip6 + 1); 1873 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1874 udp->uh_dport = sc->sc_port; 1875 ulen = (tlen - sizeof(struct ip6_hdr)); 1876 th = (struct tcphdr *)(udp + 1); 1877 } else { 1878 ip6->ip6_nxt = IPPROTO_TCP; 1879 th = (struct tcphdr *)(ip6 + 1); 1880 } 1881 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20); 1882 } 1883 #endif 1884 #if defined(INET6) && defined(INET) 1885 else 1886 #endif 1887 #ifdef INET 1888 { 1889 ip = mtod(m, struct ip *); 1890 ip->ip_v = IPVERSION; 1891 ip->ip_hl = sizeof(struct ip) >> 2; 1892 ip->ip_len = htons(tlen); 1893 ip->ip_id = 0; 1894 ip->ip_off = 0; 1895 ip->ip_sum = 0; 1896 ip->ip_src = sc->sc_inc.inc_laddr; 1897 ip->ip_dst = sc->sc_inc.inc_faddr; 1898 ip->ip_ttl = sc->sc_ip_ttl; 1899 ip->ip_tos = sc->sc_ip_tos; 1900 1901 /* 1902 * See if we should do MTU discovery. Route lookups are 1903 * expensive, so we will only unset the DF bit if: 1904 * 1905 * 1) path_mtu_discovery is disabled 1906 * 2) the SCF_UNREACH flag has been set 1907 */ 1908 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1909 ip->ip_off |= htons(IP_DF); 1910 if (sc->sc_port == 0) { 1911 ip->ip_p = IPPROTO_TCP; 1912 th = (struct tcphdr *)(ip + 1); 1913 } else { 1914 ip->ip_p = IPPROTO_UDP; 1915 udp = (struct udphdr *)(ip + 1); 1916 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1917 udp->uh_dport = sc->sc_port; 1918 ulen = (tlen - sizeof(struct ip)); 1919 th = (struct tcphdr *)(udp + 1); 1920 } 1921 } 1922 #endif /* INET */ 1923 th->th_sport = sc->sc_inc.inc_lport; 1924 th->th_dport = sc->sc_inc.inc_fport; 1925 1926 if (flags & TH_SYN) 1927 th->th_seq = htonl(sc->sc_iss); 1928 else 1929 th->th_seq = htonl(sc->sc_iss + 1); 1930 th->th_ack = htonl(sc->sc_irs + 1); 1931 th->th_off = sizeof(struct tcphdr) >> 2; 1932 th->th_x2 = 0; 1933 th->th_flags = flags; 1934 th->th_win = htons(sc->sc_wnd); 1935 th->th_urp = 0; 1936 1937 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) { 1938 th->th_flags |= TH_ECE; 1939 TCPSTAT_INC(tcps_ecn_shs); 1940 } 1941 1942 /* Tack on the TCP options. */ 1943 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1944 to.to_flags = 0; 1945 1946 if (flags & TH_SYN) { 1947 to.to_mss = mssopt; 1948 to.to_flags = TOF_MSS; 1949 if (sc->sc_flags & SCF_WINSCALE) { 1950 to.to_wscale = sc->sc_requested_r_scale; 1951 to.to_flags |= TOF_SCALE; 1952 } 1953 if (sc->sc_flags & SCF_SACK) 1954 to.to_flags |= TOF_SACKPERM; 1955 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1956 if (sc->sc_flags & SCF_SIGNATURE) 1957 to.to_flags |= TOF_SIGNATURE; 1958 #endif 1959 if (sc->sc_tfo_cookie) { 1960 to.to_flags |= TOF_FASTOPEN; 1961 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1962 to.to_tfo_cookie = sc->sc_tfo_cookie; 1963 /* don't send cookie again when retransmitting response */ 1964 sc->sc_tfo_cookie = NULL; 1965 } 1966 } 1967 if (sc->sc_flags & SCF_TIMESTAMP) { 1968 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1969 to.to_tsecr = sc->sc_tsreflect; 1970 to.to_flags |= TOF_TS; 1971 } 1972 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1973 1974 /* Adjust headers by option size. */ 1975 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1976 m->m_len += optlen; 1977 m->m_pkthdr.len += optlen; 1978 #ifdef INET6 1979 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1980 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1981 else 1982 #endif 1983 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1984 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1985 if (sc->sc_flags & SCF_SIGNATURE) { 1986 KASSERT(to.to_flags & TOF_SIGNATURE, 1987 ("tcp_addoptions() didn't set tcp_signature")); 1988 1989 /* NOTE: to.to_signature is inside of mbuf */ 1990 if (!TCPMD5_ENABLED() || 1991 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1992 m_freem(m); 1993 return (EACCES); 1994 } 1995 } 1996 #endif 1997 } else 1998 optlen = 0; 1999 2000 if (udp) { 2001 ulen += optlen; 2002 udp->uh_ulen = htons(ulen); 2003 } 2004 M_SETFIB(m, sc->sc_inc.inc_fibnum); 2005 /* 2006 * If we have peer's SYN and it has a flowid, then let's assign it to 2007 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 2008 * to SYN|ACK due to lack of inp here. 2009 */ 2010 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 2011 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 2012 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 2013 } 2014 #ifdef INET6 2015 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 2016 if (sc->sc_port) { 2017 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 2018 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2019 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, 2020 IPPROTO_UDP, 0); 2021 th->th_sum = htons(0); 2022 } else { 2023 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 2024 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2025 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 2026 IPPROTO_TCP, 0); 2027 } 2028 ip6->ip6_hlim = sc->sc_ip_ttl; 2029 #ifdef TCP_OFFLOAD 2030 if (ADDED_BY_TOE(sc)) { 2031 struct toedev *tod = sc->sc_tod; 2032 2033 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2034 2035 return (error); 2036 } 2037 #endif 2038 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 2039 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 2040 } 2041 #endif 2042 #if defined(INET6) && defined(INET) 2043 else 2044 #endif 2045 #ifdef INET 2046 { 2047 if (sc->sc_port) { 2048 m->m_pkthdr.csum_flags = CSUM_UDP; 2049 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2050 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 2051 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 2052 th->th_sum = htons(0); 2053 } else { 2054 m->m_pkthdr.csum_flags = CSUM_TCP; 2055 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2056 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 2057 htons(tlen + optlen - hlen + IPPROTO_TCP)); 2058 } 2059 #ifdef TCP_OFFLOAD 2060 if (ADDED_BY_TOE(sc)) { 2061 struct toedev *tod = sc->sc_tod; 2062 2063 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2064 2065 return (error); 2066 } 2067 #endif 2068 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 2069 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 2070 } 2071 #endif 2072 return (error); 2073 } 2074 2075 /* 2076 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 2077 * that exceed the capacity of the syncache by avoiding the storage of any 2078 * of the SYNs we receive. Syncookies defend against blind SYN flooding 2079 * attacks where the attacker does not have access to our responses. 2080 * 2081 * Syncookies encode and include all necessary information about the 2082 * connection setup within the SYN|ACK that we send back. That way we 2083 * can avoid keeping any local state until the ACK to our SYN|ACK returns 2084 * (if ever). Normally the syncache and syncookies are running in parallel 2085 * with the latter taking over when the former is exhausted. When matching 2086 * syncache entry is found the syncookie is ignored. 2087 * 2088 * The only reliable information persisting the 3WHS is our initial sequence 2089 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 2090 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 2091 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 2092 * returns and signifies a legitimate connection if it matches the ACK. 2093 * 2094 * The available space of 32 bits to store the hash and to encode the SYN 2095 * option information is very tight and we should have at least 24 bits for 2096 * the MAC to keep the number of guesses by blind spoofing reasonably high. 2097 * 2098 * SYN option information we have to encode to fully restore a connection: 2099 * MSS: is imporant to chose an optimal segment size to avoid IP level 2100 * fragmentation along the path. The common MSS values can be encoded 2101 * in a 3-bit table. Uncommon values are captured by the next lower value 2102 * in the table leading to a slight increase in packetization overhead. 2103 * WSCALE: is necessary to allow large windows to be used for high delay- 2104 * bandwidth product links. Not scaling the window when it was initially 2105 * negotiated is bad for performance as lack of scaling further decreases 2106 * the apparent available send window. We only need to encode the WSCALE 2107 * we received from the remote end. Our end can be recalculated at any 2108 * time. The common WSCALE values can be encoded in a 3-bit table. 2109 * Uncommon values are captured by the next lower value in the table 2110 * making us under-estimate the available window size halving our 2111 * theoretically possible maximum throughput for that connection. 2112 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 2113 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 2114 * that are included in all segments on a connection. We enable them when 2115 * the ACK has them. 2116 * 2117 * Security of syncookies and attack vectors: 2118 * 2119 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 2120 * together with the gloabl secret to make it unique per connection attempt. 2121 * Thus any change of any of those parameters results in a different MAC output 2122 * in an unpredictable way unless a collision is encountered. 24 bits of the 2123 * MAC are embedded into the ISS. 2124 * 2125 * To prevent replay attacks two rotating global secrets are updated with a 2126 * new random value every 15 seconds. The life-time of a syncookie is thus 2127 * 15-30 seconds. 2128 * 2129 * Vector 1: Attacking the secret. This requires finding a weakness in the 2130 * MAC itself or the way it is used here. The attacker can do a chosen plain 2131 * text attack by varying and testing the all parameters under his control. 2132 * The strength depends on the size and randomness of the secret, and the 2133 * cryptographic security of the MAC function. Due to the constant updating 2134 * of the secret the attacker has at most 29.999 seconds to find the secret 2135 * and launch spoofed connections. After that he has to start all over again. 2136 * 2137 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 2138 * size an average of 4,823 attempts are required for a 50% chance of success 2139 * to spoof a single syncookie (birthday collision paradox). However the 2140 * attacker is blind and doesn't know if one of his attempts succeeded unless 2141 * he has a side channel to interfere success from. A single connection setup 2142 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 2143 * This many attempts are required for each one blind spoofed connection. For 2144 * every additional spoofed connection he has to launch another N attempts. 2145 * Thus for a sustained rate 100 spoofed connections per second approximately 2146 * 1,800,000 packets per second would have to be sent. 2147 * 2148 * NB: The MAC function should be fast so that it doesn't become a CPU 2149 * exhaustion attack vector itself. 2150 * 2151 * References: 2152 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 2153 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 2154 * http://cr.yp.to/syncookies.html (overview) 2155 * http://cr.yp.to/syncookies/archive (details) 2156 * 2157 * 2158 * Schematic construction of a syncookie enabled Initial Sequence Number: 2159 * 0 1 2 3 2160 * 12345678901234567890123456789012 2161 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2162 * 2163 * x 24 MAC (truncated) 2164 * W 3 Send Window Scale index 2165 * M 3 MSS index 2166 * S 1 SACK permitted 2167 * P 1 Odd/even secret 2168 */ 2169 2170 /* 2171 * Distribution and probability of certain MSS values. Those in between are 2172 * rounded down to the next lower one. 2173 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2174 * .2% .3% 5% 7% 7% 20% 15% 45% 2175 */ 2176 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2177 2178 /* 2179 * Distribution and probability of certain WSCALE values. We have to map the 2180 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2181 * bits based on prevalence of certain values. Where we don't have an exact 2182 * match for are rounded down to the next lower one letting us under-estimate 2183 * the true available window. At the moment this would happen only for the 2184 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2185 * and window size). The absence of the WSCALE option (no scaling in either 2186 * direction) is encoded with index zero. 2187 * [WSCALE values histograms, Allman, 2012] 2188 * X 10 10 35 5 6 14 10% by host 2189 * X 11 4 5 5 18 49 3% by connections 2190 */ 2191 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2192 2193 /* 2194 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2195 * and good cryptographic properties. 2196 */ 2197 static uint32_t 2198 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2199 uint8_t *secbits, uintptr_t secmod) 2200 { 2201 SIPHASH_CTX ctx; 2202 uint32_t siphash[2]; 2203 2204 SipHash24_Init(&ctx); 2205 SipHash_SetKey(&ctx, secbits); 2206 switch (inc->inc_flags & INC_ISIPV6) { 2207 #ifdef INET 2208 case 0: 2209 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2210 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2211 break; 2212 #endif 2213 #ifdef INET6 2214 case INC_ISIPV6: 2215 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2216 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2217 break; 2218 #endif 2219 } 2220 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2221 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2222 SipHash_Update(&ctx, &irs, sizeof(irs)); 2223 SipHash_Update(&ctx, &flags, sizeof(flags)); 2224 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2225 SipHash_Final((u_int8_t *)&siphash, &ctx); 2226 2227 return (siphash[0] ^ siphash[1]); 2228 } 2229 2230 static tcp_seq 2231 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2232 { 2233 u_int i, secbit, wscale; 2234 uint32_t iss, hash; 2235 uint8_t *secbits; 2236 union syncookie cookie; 2237 2238 cookie.cookie = 0; 2239 2240 /* Map our computed MSS into the 3-bit index. */ 2241 for (i = nitems(tcp_sc_msstab) - 1; 2242 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2243 i--) 2244 ; 2245 cookie.flags.mss_idx = i; 2246 2247 /* 2248 * Map the send window scale into the 3-bit index but only if 2249 * the wscale option was received. 2250 */ 2251 if (sc->sc_flags & SCF_WINSCALE) { 2252 wscale = sc->sc_requested_s_scale; 2253 for (i = nitems(tcp_sc_wstab) - 1; 2254 tcp_sc_wstab[i] > wscale && i > 0; 2255 i--) 2256 ; 2257 cookie.flags.wscale_idx = i; 2258 } 2259 2260 /* Can we do SACK? */ 2261 if (sc->sc_flags & SCF_SACK) 2262 cookie.flags.sack_ok = 1; 2263 2264 /* Which of the two secrets to use. */ 2265 secbit = V_tcp_syncache.secret.oddeven & 0x1; 2266 cookie.flags.odd_even = secbit; 2267 2268 secbits = V_tcp_syncache.secret.key[secbit]; 2269 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2270 (uintptr_t)sch); 2271 2272 /* 2273 * Put the flags into the hash and XOR them to get better ISS number 2274 * variance. This doesn't enhance the cryptographic strength and is 2275 * done to prevent the 8 cookie bits from showing up directly on the 2276 * wire. 2277 */ 2278 iss = hash & ~0xff; 2279 iss |= cookie.cookie ^ (hash >> 24); 2280 2281 TCPSTAT_INC(tcps_sc_sendcookie); 2282 return (iss); 2283 } 2284 2285 static struct syncache * 2286 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2287 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2288 struct socket *lso, uint16_t port) 2289 { 2290 uint32_t hash; 2291 uint8_t *secbits; 2292 tcp_seq ack, seq; 2293 int wnd, wscale = 0; 2294 union syncookie cookie; 2295 2296 /* 2297 * Pull information out of SYN-ACK/ACK and revert sequence number 2298 * advances. 2299 */ 2300 ack = th->th_ack - 1; 2301 seq = th->th_seq - 1; 2302 2303 /* 2304 * Unpack the flags containing enough information to restore the 2305 * connection. 2306 */ 2307 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2308 2309 /* Which of the two secrets to use. */ 2310 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; 2311 2312 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2313 2314 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2315 if ((ack & ~0xff) != (hash & ~0xff)) 2316 return (NULL); 2317 2318 /* Fill in the syncache values. */ 2319 sc->sc_flags = 0; 2320 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2321 sc->sc_ipopts = NULL; 2322 2323 sc->sc_irs = seq; 2324 sc->sc_iss = ack; 2325 2326 switch (inc->inc_flags & INC_ISIPV6) { 2327 #ifdef INET 2328 case 0: 2329 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2330 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2331 break; 2332 #endif 2333 #ifdef INET6 2334 case INC_ISIPV6: 2335 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2336 sc->sc_flowlabel = 2337 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2338 break; 2339 #endif 2340 } 2341 2342 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2343 2344 /* We can simply recompute receive window scale we sent earlier. */ 2345 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2346 wscale++; 2347 2348 /* Only use wscale if it was enabled in the orignal SYN. */ 2349 if (cookie.flags.wscale_idx > 0) { 2350 sc->sc_requested_r_scale = wscale; 2351 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2352 sc->sc_flags |= SCF_WINSCALE; 2353 } 2354 2355 wnd = lso->sol_sbrcv_hiwat; 2356 wnd = imax(wnd, 0); 2357 wnd = imin(wnd, TCP_MAXWIN); 2358 sc->sc_wnd = wnd; 2359 2360 if (cookie.flags.sack_ok) 2361 sc->sc_flags |= SCF_SACK; 2362 2363 if (to->to_flags & TOF_TS) { 2364 sc->sc_flags |= SCF_TIMESTAMP; 2365 sc->sc_tsreflect = to->to_tsval; 2366 sc->sc_tsoff = tcp_new_ts_offset(inc); 2367 } 2368 2369 if (to->to_flags & TOF_SIGNATURE) 2370 sc->sc_flags |= SCF_SIGNATURE; 2371 2372 sc->sc_rxmits = 0; 2373 2374 sc->sc_port = port; 2375 2376 TCPSTAT_INC(tcps_sc_recvcookie); 2377 return (sc); 2378 } 2379 2380 #ifdef INVARIANTS 2381 static int 2382 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2383 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2384 struct socket *lso, uint16_t port) 2385 { 2386 struct syncache scs, *scx; 2387 char *s; 2388 2389 bzero(&scs, sizeof(scs)); 2390 scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port); 2391 2392 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2393 return (0); 2394 2395 if (scx != NULL) { 2396 if (sc->sc_peer_mss != scx->sc_peer_mss) 2397 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2398 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2399 2400 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2401 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2402 s, __func__, sc->sc_requested_r_scale, 2403 scx->sc_requested_r_scale); 2404 2405 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2406 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2407 s, __func__, sc->sc_requested_s_scale, 2408 scx->sc_requested_s_scale); 2409 2410 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2411 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2412 } 2413 2414 if (s != NULL) 2415 free(s, M_TCPLOG); 2416 return (0); 2417 } 2418 #endif /* INVARIANTS */ 2419 2420 static void 2421 syncookie_reseed(void *arg) 2422 { 2423 struct tcp_syncache *sc = arg; 2424 uint8_t *secbits; 2425 int secbit; 2426 2427 /* 2428 * Reseeding the secret doesn't have to be protected by a lock. 2429 * It only must be ensured that the new random values are visible 2430 * to all CPUs in a SMP environment. The atomic with release 2431 * semantics ensures that. 2432 */ 2433 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2434 secbits = sc->secret.key[secbit]; 2435 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2436 atomic_add_rel_int(&sc->secret.oddeven, 1); 2437 2438 /* Reschedule ourself. */ 2439 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2440 } 2441 2442 /* 2443 * We have overflowed a bucket. Let's pause dealing with the syncache. 2444 * This function will increment the bucketoverflow statistics appropriately 2445 * (once per pause when pausing is enabled; otherwise, once per overflow). 2446 */ 2447 static void 2448 syncache_pause(struct in_conninfo *inc) 2449 { 2450 time_t delta; 2451 const char *s; 2452 2453 /* XXX: 2454 * 2. Add sysctl read here so we don't get the benefit of this 2455 * change without the new sysctl. 2456 */ 2457 2458 /* 2459 * Try an unlocked read. If we already know that another thread 2460 * has activated the feature, there is no need to proceed. 2461 */ 2462 if (V_tcp_syncache.paused) 2463 return; 2464 2465 /* Are cookied enabled? If not, we can't pause. */ 2466 if (!V_tcp_syncookies) { 2467 TCPSTAT_INC(tcps_sc_bucketoverflow); 2468 return; 2469 } 2470 2471 /* 2472 * We may be the first thread to find an overflow. Get the lock 2473 * and evaluate if we need to take action. 2474 */ 2475 mtx_lock(&V_tcp_syncache.pause_mtx); 2476 if (V_tcp_syncache.paused) { 2477 mtx_unlock(&V_tcp_syncache.pause_mtx); 2478 return; 2479 } 2480 2481 /* Activate protection. */ 2482 V_tcp_syncache.paused = true; 2483 TCPSTAT_INC(tcps_sc_bucketoverflow); 2484 2485 /* 2486 * Determine the last backoff time. If we are seeing a re-newed 2487 * attack within that same time after last reactivating the syncache, 2488 * consider it an extension of the same attack. 2489 */ 2490 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; 2491 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { 2492 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { 2493 delta <<= 1; 2494 V_tcp_syncache.pause_backoff++; 2495 } 2496 } else { 2497 delta = TCP_SYNCACHE_PAUSE_TIME; 2498 V_tcp_syncache.pause_backoff = 0; 2499 } 2500 2501 /* Log a warning, including IP addresses, if able. */ 2502 if (inc != NULL) 2503 s = tcp_log_addrs(inc, NULL, NULL, NULL); 2504 else 2505 s = (const char *)NULL; 2506 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " 2507 "the next %lld seconds%s%s%s\n", (long long)delta, 2508 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", 2509 (s != NULL) ? ")" : ""); 2510 free(__DECONST(void *, s), M_TCPLOG); 2511 2512 /* Use the calculated delta to set a new pause time. */ 2513 V_tcp_syncache.pause_until = time_uptime + delta; 2514 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, 2515 &V_tcp_syncache); 2516 mtx_unlock(&V_tcp_syncache.pause_mtx); 2517 } 2518 2519 /* Evaluate whether we need to unpause. */ 2520 static void 2521 syncache_unpause(void *arg) 2522 { 2523 struct tcp_syncache *sc; 2524 time_t delta; 2525 2526 sc = arg; 2527 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); 2528 callout_deactivate(&sc->pause_co); 2529 2530 /* 2531 * Check to make sure we are not running early. If the pause 2532 * time has expired, then deactivate the protection. 2533 */ 2534 if ((delta = sc->pause_until - time_uptime) > 0) 2535 callout_schedule(&sc->pause_co, delta * hz); 2536 else 2537 sc->paused = false; 2538 } 2539 2540 /* 2541 * Exports the syncache entries to userland so that netstat can display 2542 * them alongside the other sockets. This function is intended to be 2543 * called only from tcp_pcblist. 2544 * 2545 * Due to concurrency on an active system, the number of pcbs exported 2546 * may have no relation to max_pcbs. max_pcbs merely indicates the 2547 * amount of space the caller allocated for this function to use. 2548 */ 2549 int 2550 syncache_pcblist(struct sysctl_req *req) 2551 { 2552 struct xtcpcb xt; 2553 struct syncache *sc; 2554 struct syncache_head *sch; 2555 int error, i; 2556 2557 bzero(&xt, sizeof(xt)); 2558 xt.xt_len = sizeof(xt); 2559 xt.t_state = TCPS_SYN_RECEIVED; 2560 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2561 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2562 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2563 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2564 2565 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 2566 sch = &V_tcp_syncache.hashbase[i]; 2567 SCH_LOCK(sch); 2568 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2569 if (sc->sc_cred != NULL && 2570 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2571 continue; 2572 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2573 xt.xt_inp.inp_vflag = INP_IPV6; 2574 else 2575 xt.xt_inp.inp_vflag = INP_IPV4; 2576 xt.xt_encaps_port = sc->sc_port; 2577 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2578 sizeof (struct in_conninfo)); 2579 error = SYSCTL_OUT(req, &xt, sizeof xt); 2580 if (error) { 2581 SCH_UNLOCK(sch); 2582 return (0); 2583 } 2584 } 2585 SCH_UNLOCK(sch); 2586 } 2587 2588 return (0); 2589 } 2590