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 #include "opt_pcbgroup.h" 42 43 #include <sys/param.h> 44 #include <sys/systm.h> 45 #include <sys/hash.h> 46 #include <sys/refcount.h> 47 #include <sys/kernel.h> 48 #include <sys/sysctl.h> 49 #include <sys/limits.h> 50 #include <sys/lock.h> 51 #include <sys/mutex.h> 52 #include <sys/malloc.h> 53 #include <sys/mbuf.h> 54 #include <sys/proc.h> /* for proc0 declaration */ 55 #include <sys/random.h> 56 #include <sys/socket.h> 57 #include <sys/socketvar.h> 58 #include <sys/syslog.h> 59 #include <sys/ucred.h> 60 61 #include <sys/md5.h> 62 #include <crypto/siphash/siphash.h> 63 64 #include <vm/uma.h> 65 66 #include <net/if.h> 67 #include <net/if_var.h> 68 #include <net/route.h> 69 #include <net/vnet.h> 70 71 #include <netinet/in.h> 72 #include <netinet/in_kdtrace.h> 73 #include <netinet/in_systm.h> 74 #include <netinet/ip.h> 75 #include <netinet/in_var.h> 76 #include <netinet/in_pcb.h> 77 #include <netinet/ip_var.h> 78 #include <netinet/ip_options.h> 79 #ifdef INET6 80 #include <netinet/ip6.h> 81 #include <netinet/icmp6.h> 82 #include <netinet6/nd6.h> 83 #include <netinet6/ip6_var.h> 84 #include <netinet6/in6_pcb.h> 85 #endif 86 #include <netinet/tcp.h> 87 #include <netinet/tcp_fastopen.h> 88 #include <netinet/tcp_fsm.h> 89 #include <netinet/tcp_seq.h> 90 #include <netinet/tcp_timer.h> 91 #include <netinet/tcp_var.h> 92 #include <netinet/tcp_syncache.h> 93 #ifdef INET6 94 #include <netinet6/tcp6_var.h> 95 #endif 96 #ifdef TCP_OFFLOAD 97 #include <netinet/toecore.h> 98 #endif 99 100 #include <netipsec/ipsec_support.h> 101 102 #include <machine/in_cksum.h> 103 104 #include <security/mac/mac_framework.h> 105 106 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1; 107 #define V_tcp_syncookies VNET(tcp_syncookies) 108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 109 &VNET_NAME(tcp_syncookies), 0, 110 "Use TCP SYN cookies if the syncache overflows"); 111 112 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0; 113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 115 &VNET_NAME(tcp_syncookiesonly), 0, 116 "Use only TCP SYN cookies"); 117 118 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1; 119 #define V_functions_inherit_listen_socket_stack \ 120 VNET(functions_inherit_listen_socket_stack) 121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack, 122 CTLFLAG_VNET | CTLFLAG_RW, 123 &VNET_NAME(functions_inherit_listen_socket_stack), 0, 124 "Inherit listen socket's stack"); 125 126 #ifdef TCP_OFFLOAD 127 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 128 #endif 129 130 static void syncache_drop(struct syncache *, struct syncache_head *); 131 static void syncache_free(struct syncache *); 132 static void syncache_insert(struct syncache *, struct syncache_head *); 133 static int syncache_respond(struct syncache *, const struct mbuf *, int); 134 static struct socket *syncache_socket(struct syncache *, struct socket *, 135 struct mbuf *m); 136 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 137 int docallout); 138 static void syncache_timer(void *); 139 140 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 141 uint8_t *, uintptr_t); 142 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 143 static struct syncache 144 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 145 struct syncache *, struct tcphdr *, struct tcpopt *, 146 struct socket *); 147 static void syncache_pause(struct in_conninfo *); 148 static void syncache_unpause(void *); 149 static void syncookie_reseed(void *); 150 #ifdef INVARIANTS 151 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 152 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 153 struct socket *lso); 154 #endif 155 156 /* 157 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 158 * 3 retransmits corresponds to a timeout with default values of 159 * tcp_rexmit_initial * ( 1 + 160 * tcp_backoff[1] + 161 * tcp_backoff[2] + 162 * tcp_backoff[3]) + 3 * tcp_rexmit_slop, 163 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms, 164 * the odds are that the user has given up attempting to connect by then. 165 */ 166 #define SYNCACHE_MAXREXMTS 3 167 168 /* Arbitrary values */ 169 #define TCP_SYNCACHE_HASHSIZE 512 170 #define TCP_SYNCACHE_BUCKETLIMIT 30 171 172 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 173 #define V_tcp_syncache VNET(tcp_syncache) 174 175 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, 176 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 177 "TCP SYN cache"); 178 179 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 180 &VNET_NAME(tcp_syncache.bucket_limit), 0, 181 "Per-bucket hash limit for syncache"); 182 183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 184 &VNET_NAME(tcp_syncache.cache_limit), 0, 185 "Overall entry limit for syncache"); 186 187 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 188 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 189 190 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 191 &VNET_NAME(tcp_syncache.hashsize), 0, 192 "Size of TCP syncache hashtable"); 193 194 static int 195 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 196 { 197 int error; 198 u_int new; 199 200 new = V_tcp_syncache.rexmt_limit; 201 error = sysctl_handle_int(oidp, &new, 0, req); 202 if ((error == 0) && (req->newptr != NULL)) { 203 if (new > TCP_MAXRXTSHIFT) 204 error = EINVAL; 205 else 206 V_tcp_syncache.rexmt_limit = new; 207 } 208 return (error); 209 } 210 211 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 212 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 213 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 214 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI", 215 "Limit on SYN/ACK retransmissions"); 216 217 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 218 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 219 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 220 "Send reset on socket allocation failure"); 221 222 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 223 224 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 225 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 226 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 227 228 /* 229 * Requires the syncache entry to be already removed from the bucket list. 230 */ 231 static void 232 syncache_free(struct syncache *sc) 233 { 234 235 if (sc->sc_ipopts) 236 (void) m_free(sc->sc_ipopts); 237 if (sc->sc_cred) 238 crfree(sc->sc_cred); 239 #ifdef MAC 240 mac_syncache_destroy(&sc->sc_label); 241 #endif 242 243 uma_zfree(V_tcp_syncache.zone, sc); 244 } 245 246 void 247 syncache_init(void) 248 { 249 int i; 250 251 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 252 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 253 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 254 V_tcp_syncache.hash_secret = arc4random(); 255 256 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 257 &V_tcp_syncache.hashsize); 258 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 259 &V_tcp_syncache.bucket_limit); 260 if (!powerof2(V_tcp_syncache.hashsize) || 261 V_tcp_syncache.hashsize == 0) { 262 printf("WARNING: syncache hash size is not a power of 2.\n"); 263 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 264 } 265 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 266 267 /* Set limits. */ 268 V_tcp_syncache.cache_limit = 269 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 270 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 271 &V_tcp_syncache.cache_limit); 272 273 /* Allocate the hash table. */ 274 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 275 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 276 277 #ifdef VIMAGE 278 V_tcp_syncache.vnet = curvnet; 279 #endif 280 281 /* Initialize the hash buckets. */ 282 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 283 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 284 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 285 NULL, MTX_DEF); 286 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 287 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 288 V_tcp_syncache.hashbase[i].sch_length = 0; 289 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 290 V_tcp_syncache.hashbase[i].sch_last_overflow = 291 -(SYNCOOKIE_LIFETIME + 1); 292 } 293 294 /* Create the syncache entry zone. */ 295 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 296 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 297 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 298 V_tcp_syncache.cache_limit); 299 300 /* Start the SYN cookie reseeder callout. */ 301 callout_init(&V_tcp_syncache.secret.reseed, 1); 302 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 303 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 304 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 305 syncookie_reseed, &V_tcp_syncache); 306 307 /* Initialize the pause machinery. */ 308 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF); 309 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx, 310 0); 311 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME; 312 V_tcp_syncache.pause_backoff = 0; 313 V_tcp_syncache.paused = false; 314 } 315 316 #ifdef VIMAGE 317 void 318 syncache_destroy(void) 319 { 320 struct syncache_head *sch; 321 struct syncache *sc, *nsc; 322 int i; 323 324 /* 325 * Stop the re-seed timer before freeing resources. No need to 326 * possibly schedule it another time. 327 */ 328 callout_drain(&V_tcp_syncache.secret.reseed); 329 330 /* Stop the SYN cache pause callout. */ 331 mtx_lock(&V_tcp_syncache.pause_mtx); 332 if (callout_stop(&V_tcp_syncache.pause_co) == 0) { 333 mtx_unlock(&V_tcp_syncache.pause_mtx); 334 callout_drain(&V_tcp_syncache.pause_co); 335 } else 336 mtx_unlock(&V_tcp_syncache.pause_mtx); 337 338 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 339 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 340 sch = &V_tcp_syncache.hashbase[i]; 341 callout_drain(&sch->sch_timer); 342 343 SCH_LOCK(sch); 344 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 345 syncache_drop(sc, sch); 346 SCH_UNLOCK(sch); 347 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 348 ("%s: sch->sch_bucket not empty", __func__)); 349 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 350 __func__, sch->sch_length)); 351 mtx_destroy(&sch->sch_mtx); 352 } 353 354 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 355 ("%s: cache_count not 0", __func__)); 356 357 /* Free the allocated global resources. */ 358 uma_zdestroy(V_tcp_syncache.zone); 359 free(V_tcp_syncache.hashbase, M_SYNCACHE); 360 mtx_destroy(&V_tcp_syncache.pause_mtx); 361 } 362 #endif 363 364 /* 365 * Inserts a syncache entry into the specified bucket row. 366 * Locks and unlocks the syncache_head autonomously. 367 */ 368 static void 369 syncache_insert(struct syncache *sc, struct syncache_head *sch) 370 { 371 struct syncache *sc2; 372 373 SCH_LOCK(sch); 374 375 /* 376 * Make sure that we don't overflow the per-bucket limit. 377 * If the bucket is full, toss the oldest element. 378 */ 379 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 380 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 381 ("sch->sch_length incorrect")); 382 syncache_pause(&sc->sc_inc); 383 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 384 sch->sch_last_overflow = time_uptime; 385 syncache_drop(sc2, sch); 386 } 387 388 /* Put it into the bucket. */ 389 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 390 sch->sch_length++; 391 392 #ifdef TCP_OFFLOAD 393 if (ADDED_BY_TOE(sc)) { 394 struct toedev *tod = sc->sc_tod; 395 396 tod->tod_syncache_added(tod, sc->sc_todctx); 397 } 398 #endif 399 400 /* Reinitialize the bucket row's timer. */ 401 if (sch->sch_length == 1) 402 sch->sch_nextc = ticks + INT_MAX; 403 syncache_timeout(sc, sch, 1); 404 405 SCH_UNLOCK(sch); 406 407 TCPSTATES_INC(TCPS_SYN_RECEIVED); 408 TCPSTAT_INC(tcps_sc_added); 409 } 410 411 /* 412 * Remove and free entry from syncache bucket row. 413 * Expects locked syncache head. 414 */ 415 static void 416 syncache_drop(struct syncache *sc, struct syncache_head *sch) 417 { 418 419 SCH_LOCK_ASSERT(sch); 420 421 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 422 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 423 sch->sch_length--; 424 425 #ifdef TCP_OFFLOAD 426 if (ADDED_BY_TOE(sc)) { 427 struct toedev *tod = sc->sc_tod; 428 429 tod->tod_syncache_removed(tod, sc->sc_todctx); 430 } 431 #endif 432 433 syncache_free(sc); 434 } 435 436 /* 437 * Engage/reengage time on bucket row. 438 */ 439 static void 440 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 441 { 442 int rexmt; 443 444 if (sc->sc_rxmits == 0) 445 rexmt = tcp_rexmit_initial; 446 else 447 TCPT_RANGESET(rexmt, 448 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], 449 tcp_rexmit_min, TCPTV_REXMTMAX); 450 sc->sc_rxttime = ticks + rexmt; 451 sc->sc_rxmits++; 452 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 453 sch->sch_nextc = sc->sc_rxttime; 454 if (docallout) 455 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 456 syncache_timer, (void *)sch); 457 } 458 } 459 460 /* 461 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 462 * If we have retransmitted an entry the maximum number of times, expire it. 463 * One separate timer for each bucket row. 464 */ 465 static void 466 syncache_timer(void *xsch) 467 { 468 struct syncache_head *sch = (struct syncache_head *)xsch; 469 struct syncache *sc, *nsc; 470 struct epoch_tracker et; 471 int tick = ticks; 472 char *s; 473 bool paused; 474 475 CURVNET_SET(sch->sch_sc->vnet); 476 477 /* NB: syncache_head has already been locked by the callout. */ 478 SCH_LOCK_ASSERT(sch); 479 480 /* 481 * In the following cycle we may remove some entries and/or 482 * advance some timeouts, so re-initialize the bucket timer. 483 */ 484 sch->sch_nextc = tick + INT_MAX; 485 486 /* 487 * If we have paused processing, unconditionally remove 488 * all syncache entries. 489 */ 490 mtx_lock(&V_tcp_syncache.pause_mtx); 491 paused = V_tcp_syncache.paused; 492 mtx_unlock(&V_tcp_syncache.pause_mtx); 493 494 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 495 if (paused) { 496 syncache_drop(sc, sch); 497 continue; 498 } 499 /* 500 * We do not check if the listen socket still exists 501 * and accept the case where the listen socket may be 502 * gone by the time we resend the SYN/ACK. We do 503 * not expect this to happens often. If it does, 504 * then the RST will be sent by the time the remote 505 * host does the SYN/ACK->ACK. 506 */ 507 if (TSTMP_GT(sc->sc_rxttime, tick)) { 508 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 509 sch->sch_nextc = sc->sc_rxttime; 510 continue; 511 } 512 if (sc->sc_rxmits > V_tcp_ecn_maxretries) { 513 sc->sc_flags &= ~SCF_ECN; 514 } 515 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 516 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 517 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 518 "giving up and removing syncache entry\n", 519 s, __func__); 520 free(s, M_TCPLOG); 521 } 522 syncache_drop(sc, sch); 523 TCPSTAT_INC(tcps_sc_stale); 524 continue; 525 } 526 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 527 log(LOG_DEBUG, "%s; %s: Response timeout, " 528 "retransmitting (%u) SYN|ACK\n", 529 s, __func__, sc->sc_rxmits); 530 free(s, M_TCPLOG); 531 } 532 533 NET_EPOCH_ENTER(et); 534 syncache_respond(sc, NULL, TH_SYN|TH_ACK); 535 NET_EPOCH_EXIT(et); 536 TCPSTAT_INC(tcps_sc_retransmitted); 537 syncache_timeout(sc, sch, 0); 538 } 539 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 540 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 541 syncache_timer, (void *)(sch)); 542 CURVNET_RESTORE(); 543 } 544 545 /* 546 * Returns true if the system is only using cookies at the moment. 547 * This could be due to a sysadmin decision to only use cookies, or it 548 * could be due to the system detecting an attack. 549 */ 550 static inline bool 551 syncache_cookiesonly(void) 552 { 553 554 return (V_tcp_syncookies && (V_tcp_syncache.paused || 555 V_tcp_syncookiesonly)); 556 } 557 558 /* 559 * Find the hash bucket for the given connection. 560 */ 561 static struct syncache_head * 562 syncache_hashbucket(struct in_conninfo *inc) 563 { 564 uint32_t hash; 565 566 /* 567 * The hash is built on foreign port + local port + foreign address. 568 * We rely on the fact that struct in_conninfo starts with 16 bits 569 * of foreign port, then 16 bits of local port then followed by 128 570 * bits of foreign address. In case of IPv4 address, the first 3 571 * 32-bit words of the address always are zeroes. 572 */ 573 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 574 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 575 576 return (&V_tcp_syncache.hashbase[hash]); 577 } 578 579 /* 580 * Find an entry in the syncache. 581 * Returns always with locked syncache_head plus a matching entry or NULL. 582 */ 583 static struct syncache * 584 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 585 { 586 struct syncache *sc; 587 struct syncache_head *sch; 588 589 *schp = sch = syncache_hashbucket(inc); 590 SCH_LOCK(sch); 591 592 /* Circle through bucket row to find matching entry. */ 593 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 594 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 595 sizeof(struct in_endpoints)) == 0) 596 break; 597 598 return (sc); /* Always returns with locked sch. */ 599 } 600 601 /* 602 * This function is called when we get a RST for a 603 * non-existent connection, so that we can see if the 604 * connection is in the syn cache. If it is, zap it. 605 * If required send a challenge ACK. 606 */ 607 void 608 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m) 609 { 610 struct syncache *sc; 611 struct syncache_head *sch; 612 char *s = NULL; 613 614 if (syncache_cookiesonly()) 615 return; 616 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 617 SCH_LOCK_ASSERT(sch); 618 619 /* 620 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 621 * See RFC 793 page 65, section SEGMENT ARRIVES. 622 */ 623 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 624 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 625 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 626 "FIN flag set, segment ignored\n", s, __func__); 627 TCPSTAT_INC(tcps_badrst); 628 goto done; 629 } 630 631 /* 632 * No corresponding connection was found in syncache. 633 * If syncookies are enabled and possibly exclusively 634 * used, or we are under memory pressure, a valid RST 635 * may not find a syncache entry. In that case we're 636 * done and no SYN|ACK retransmissions will happen. 637 * Otherwise the RST was misdirected or spoofed. 638 */ 639 if (sc == NULL) { 640 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 641 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 642 "syncache entry (possibly syncookie only), " 643 "segment ignored\n", s, __func__); 644 TCPSTAT_INC(tcps_badrst); 645 goto done; 646 } 647 648 /* 649 * If the RST bit is set, check the sequence number to see 650 * if this is a valid reset segment. 651 * 652 * RFC 793 page 37: 653 * In all states except SYN-SENT, all reset (RST) segments 654 * are validated by checking their SEQ-fields. A reset is 655 * valid if its sequence number is in the window. 656 * 657 * RFC 793 page 69: 658 * There are four cases for the acceptability test for an incoming 659 * segment: 660 * 661 * Segment Receive Test 662 * Length Window 663 * ------- ------- ------------------------------------------- 664 * 0 0 SEG.SEQ = RCV.NXT 665 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 666 * >0 0 not acceptable 667 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 668 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 669 * 670 * Note that when receiving a SYN segment in the LISTEN state, 671 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 672 * described in RFC 793, page 66. 673 */ 674 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 675 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 676 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 677 if (V_tcp_insecure_rst || 678 th->th_seq == sc->sc_irs + 1) { 679 syncache_drop(sc, sch); 680 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 681 log(LOG_DEBUG, 682 "%s; %s: Our SYN|ACK was rejected, " 683 "connection attempt aborted by remote " 684 "endpoint\n", 685 s, __func__); 686 TCPSTAT_INC(tcps_sc_reset); 687 } else { 688 TCPSTAT_INC(tcps_badrst); 689 /* Send challenge ACK. */ 690 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 691 log(LOG_DEBUG, "%s; %s: RST with invalid " 692 " SEQ %u != NXT %u (+WND %u), " 693 "sending challenge ACK\n", 694 s, __func__, 695 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 696 syncache_respond(sc, m, TH_ACK); 697 } 698 } else { 699 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 700 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 701 "NXT %u (+WND %u), segment ignored\n", 702 s, __func__, 703 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 704 TCPSTAT_INC(tcps_badrst); 705 } 706 707 done: 708 if (s != NULL) 709 free(s, M_TCPLOG); 710 SCH_UNLOCK(sch); 711 } 712 713 void 714 syncache_badack(struct in_conninfo *inc) 715 { 716 struct syncache *sc; 717 struct syncache_head *sch; 718 719 if (syncache_cookiesonly()) 720 return; 721 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 722 SCH_LOCK_ASSERT(sch); 723 if (sc != NULL) { 724 syncache_drop(sc, sch); 725 TCPSTAT_INC(tcps_sc_badack); 726 } 727 SCH_UNLOCK(sch); 728 } 729 730 void 731 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq) 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) 741 goto done; 742 743 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 744 if (ntohl(th_seq) != sc->sc_iss) 745 goto done; 746 747 /* 748 * If we've rertransmitted 3 times and this is our second error, 749 * we remove the entry. Otherwise, we allow it to continue on. 750 * This prevents us from incorrectly nuking an entry during a 751 * spurious network outage. 752 * 753 * See tcp_notify(). 754 */ 755 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 756 sc->sc_flags |= SCF_UNREACH; 757 goto done; 758 } 759 syncache_drop(sc, sch); 760 TCPSTAT_INC(tcps_sc_unreach); 761 done: 762 SCH_UNLOCK(sch); 763 } 764 765 /* 766 * Build a new TCP socket structure from a syncache entry. 767 * 768 * On success return the newly created socket with its underlying inp locked. 769 */ 770 static struct socket * 771 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 772 { 773 struct tcp_function_block *blk; 774 struct inpcb *inp = NULL; 775 struct socket *so; 776 struct tcpcb *tp; 777 int error; 778 char *s; 779 780 NET_EPOCH_ASSERT(); 781 782 /* 783 * Ok, create the full blown connection, and set things up 784 * as they would have been set up if we had created the 785 * connection when the SYN arrived. If we can't create 786 * the connection, abort it. 787 */ 788 so = sonewconn(lso, 0); 789 if (so == NULL) { 790 /* 791 * Drop the connection; we will either send a RST or 792 * have the peer retransmit its SYN again after its 793 * RTO and try again. 794 */ 795 TCPSTAT_INC(tcps_listendrop); 796 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 797 log(LOG_DEBUG, "%s; %s: Socket create failed " 798 "due to limits or memory shortage\n", 799 s, __func__); 800 free(s, M_TCPLOG); 801 } 802 goto abort2; 803 } 804 #ifdef MAC 805 mac_socketpeer_set_from_mbuf(m, so); 806 #endif 807 808 inp = sotoinpcb(so); 809 inp->inp_inc.inc_fibnum = so->so_fibnum; 810 INP_WLOCK(inp); 811 /* 812 * Exclusive pcbinfo lock is not required in syncache socket case even 813 * if two inpcb locks can be acquired simultaneously: 814 * - the inpcb in LISTEN state, 815 * - the newly created inp. 816 * 817 * In this case, an inp cannot be at same time in LISTEN state and 818 * just created by an accept() call. 819 */ 820 INP_HASH_WLOCK(&V_tcbinfo); 821 822 /* Insert new socket into PCB hash list. */ 823 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 824 #ifdef INET6 825 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 826 inp->inp_vflag &= ~INP_IPV4; 827 inp->inp_vflag |= INP_IPV6; 828 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 829 } else { 830 inp->inp_vflag &= ~INP_IPV6; 831 inp->inp_vflag |= INP_IPV4; 832 #endif 833 inp->inp_ip_ttl = sc->sc_ip_ttl; 834 inp->inp_ip_tos = sc->sc_ip_tos; 835 inp->inp_laddr = sc->sc_inc.inc_laddr; 836 #ifdef INET6 837 } 838 #endif 839 840 /* 841 * If there's an mbuf and it has a flowid, then let's initialise the 842 * inp with that particular flowid. 843 */ 844 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 845 inp->inp_flowid = m->m_pkthdr.flowid; 846 inp->inp_flowtype = M_HASHTYPE_GET(m); 847 #ifdef NUMA 848 inp->inp_numa_domain = m->m_pkthdr.numa_domain; 849 #endif 850 } 851 852 inp->inp_lport = sc->sc_inc.inc_lport; 853 #ifdef INET6 854 if (inp->inp_vflag & INP_IPV6PROTO) { 855 struct inpcb *oinp = sotoinpcb(lso); 856 857 /* 858 * Inherit socket options from the listening socket. 859 * Note that in6p_inputopts are not (and should not be) 860 * copied, since it stores previously received options and is 861 * used to detect if each new option is different than the 862 * previous one and hence should be passed to a user. 863 * If we copied in6p_inputopts, a user would not be able to 864 * receive options just after calling the accept system call. 865 */ 866 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 867 if (oinp->in6p_outputopts) 868 inp->in6p_outputopts = 869 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 870 inp->in6p_hops = oinp->in6p_hops; 871 } 872 873 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 874 struct in6_addr laddr6; 875 struct sockaddr_in6 sin6; 876 877 sin6.sin6_family = AF_INET6; 878 sin6.sin6_len = sizeof(sin6); 879 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 880 sin6.sin6_port = sc->sc_inc.inc_fport; 881 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 882 laddr6 = inp->in6p_laddr; 883 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 884 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 885 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 886 thread0.td_ucred, m, false)) != 0) { 887 inp->in6p_laddr = laddr6; 888 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 889 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 890 "with error %i\n", 891 s, __func__, error); 892 free(s, M_TCPLOG); 893 } 894 INP_HASH_WUNLOCK(&V_tcbinfo); 895 goto abort; 896 } 897 /* Override flowlabel from in6_pcbconnect. */ 898 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 899 inp->inp_flow |= sc->sc_flowlabel; 900 } 901 #endif /* INET6 */ 902 #if defined(INET) && defined(INET6) 903 else 904 #endif 905 #ifdef INET 906 { 907 struct in_addr laddr; 908 struct sockaddr_in sin; 909 910 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 911 912 if (inp->inp_options == NULL) { 913 inp->inp_options = sc->sc_ipopts; 914 sc->sc_ipopts = NULL; 915 } 916 917 sin.sin_family = AF_INET; 918 sin.sin_len = sizeof(sin); 919 sin.sin_addr = sc->sc_inc.inc_faddr; 920 sin.sin_port = sc->sc_inc.inc_fport; 921 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 922 laddr = inp->inp_laddr; 923 if (inp->inp_laddr.s_addr == INADDR_ANY) 924 inp->inp_laddr = sc->sc_inc.inc_laddr; 925 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 926 thread0.td_ucred, m, false)) != 0) { 927 inp->inp_laddr = laddr; 928 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 929 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 930 "with error %i\n", 931 s, __func__, error); 932 free(s, M_TCPLOG); 933 } 934 INP_HASH_WUNLOCK(&V_tcbinfo); 935 goto abort; 936 } 937 } 938 #endif /* INET */ 939 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 940 /* Copy old policy into new socket's. */ 941 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 942 printf("syncache_socket: could not copy policy\n"); 943 #endif 944 INP_HASH_WUNLOCK(&V_tcbinfo); 945 tp = intotcpcb(inp); 946 tcp_state_change(tp, TCPS_SYN_RECEIVED); 947 tp->iss = sc->sc_iss; 948 tp->irs = sc->sc_irs; 949 tcp_rcvseqinit(tp); 950 tcp_sendseqinit(tp); 951 blk = sototcpcb(lso)->t_fb; 952 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 953 /* 954 * Our parents t_fb was not the default, 955 * we need to release our ref on tp->t_fb and 956 * pickup one on the new entry. 957 */ 958 struct tcp_function_block *rblk; 959 960 rblk = find_and_ref_tcp_fb(blk); 961 KASSERT(rblk != NULL, 962 ("cannot find blk %p out of syncache?", blk)); 963 if (tp->t_fb->tfb_tcp_fb_fini) 964 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 965 refcount_release(&tp->t_fb->tfb_refcnt); 966 tp->t_fb = rblk; 967 /* 968 * XXXrrs this is quite dangerous, it is possible 969 * for the new function to fail to init. We also 970 * are not asking if the handoff_is_ok though at 971 * the very start thats probalbly ok. 972 */ 973 if (tp->t_fb->tfb_tcp_fb_init) { 974 (*tp->t_fb->tfb_tcp_fb_init)(tp); 975 } 976 } 977 tp->snd_wl1 = sc->sc_irs; 978 tp->snd_max = tp->iss + 1; 979 tp->snd_nxt = tp->iss + 1; 980 tp->rcv_up = sc->sc_irs + 1; 981 tp->rcv_wnd = sc->sc_wnd; 982 tp->rcv_adv += tp->rcv_wnd; 983 tp->last_ack_sent = tp->rcv_nxt; 984 985 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 986 if (sc->sc_flags & SCF_NOOPT) 987 tp->t_flags |= TF_NOOPT; 988 else { 989 if (sc->sc_flags & SCF_WINSCALE) { 990 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 991 tp->snd_scale = sc->sc_requested_s_scale; 992 tp->request_r_scale = sc->sc_requested_r_scale; 993 } 994 if (sc->sc_flags & SCF_TIMESTAMP) { 995 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 996 tp->ts_recent = sc->sc_tsreflect; 997 tp->ts_recent_age = tcp_ts_getticks(); 998 tp->ts_offset = sc->sc_tsoff; 999 } 1000 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1001 if (sc->sc_flags & SCF_SIGNATURE) 1002 tp->t_flags |= TF_SIGNATURE; 1003 #endif 1004 if (sc->sc_flags & SCF_SACK) 1005 tp->t_flags |= TF_SACK_PERMIT; 1006 } 1007 1008 if (sc->sc_flags & SCF_ECN) 1009 tp->t_flags2 |= TF2_ECN_PERMIT; 1010 1011 /* 1012 * Set up MSS and get cached values from tcp_hostcache. 1013 * This might overwrite some of the defaults we just set. 1014 */ 1015 tcp_mss(tp, sc->sc_peer_mss); 1016 1017 /* 1018 * If the SYN,ACK was retransmitted, indicate that CWND to be 1019 * limited to one segment in cc_conn_init(). 1020 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 1021 */ 1022 if (sc->sc_rxmits > 1) 1023 tp->snd_cwnd = 1; 1024 1025 #ifdef TCP_OFFLOAD 1026 /* 1027 * Allow a TOE driver to install its hooks. Note that we hold the 1028 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 1029 * new connection before the TOE driver has done its thing. 1030 */ 1031 if (ADDED_BY_TOE(sc)) { 1032 struct toedev *tod = sc->sc_tod; 1033 1034 tod->tod_offload_socket(tod, sc->sc_todctx, so); 1035 } 1036 #endif 1037 /* 1038 * Copy and activate timers. 1039 */ 1040 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 1041 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 1042 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1043 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1044 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1045 1046 TCPSTAT_INC(tcps_accepts); 1047 return (so); 1048 1049 abort: 1050 INP_WUNLOCK(inp); 1051 abort2: 1052 if (so != NULL) 1053 soabort(so); 1054 return (NULL); 1055 } 1056 1057 /* 1058 * This function gets called when we receive an ACK for a 1059 * socket in the LISTEN state. We look up the connection 1060 * in the syncache, and if its there, we pull it out of 1061 * the cache and turn it into a full-blown connection in 1062 * the SYN-RECEIVED state. 1063 * 1064 * On syncache_socket() success the newly created socket 1065 * has its underlying inp locked. 1066 */ 1067 int 1068 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1069 struct socket **lsop, struct mbuf *m) 1070 { 1071 struct syncache *sc; 1072 struct syncache_head *sch; 1073 struct syncache scs; 1074 char *s; 1075 bool locked; 1076 1077 NET_EPOCH_ASSERT(); 1078 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1079 ("%s: can handle only ACK", __func__)); 1080 1081 if (syncache_cookiesonly()) { 1082 sc = NULL; 1083 sch = syncache_hashbucket(inc); 1084 locked = false; 1085 } else { 1086 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1087 locked = true; 1088 SCH_LOCK_ASSERT(sch); 1089 } 1090 1091 #ifdef INVARIANTS 1092 /* 1093 * Test code for syncookies comparing the syncache stored 1094 * values with the reconstructed values from the cookie. 1095 */ 1096 if (sc != NULL) 1097 syncookie_cmp(inc, sch, sc, th, to, *lsop); 1098 #endif 1099 1100 if (sc == NULL) { 1101 /* 1102 * There is no syncache entry, so see if this ACK is 1103 * a returning syncookie. To do this, first: 1104 * A. Check if syncookies are used in case of syncache 1105 * overflows 1106 * B. See if this socket has had a syncache entry dropped in 1107 * the recent past. We don't want to accept a bogus 1108 * syncookie if we've never received a SYN or accept it 1109 * twice. 1110 * C. check that the syncookie is valid. If it is, then 1111 * cobble up a fake syncache entry, and return. 1112 */ 1113 if (locked && !V_tcp_syncookies) { 1114 SCH_UNLOCK(sch); 1115 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1116 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1117 "segment rejected (syncookies disabled)\n", 1118 s, __func__); 1119 goto failed; 1120 } 1121 if (locked && !V_tcp_syncookiesonly && 1122 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1123 SCH_UNLOCK(sch); 1124 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1125 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1126 "segment rejected (no syncache entry)\n", 1127 s, __func__); 1128 goto failed; 1129 } 1130 bzero(&scs, sizeof(scs)); 1131 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1132 if (locked) 1133 SCH_UNLOCK(sch); 1134 if (sc == NULL) { 1135 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1136 log(LOG_DEBUG, "%s; %s: Segment failed " 1137 "SYNCOOKIE authentication, segment rejected " 1138 "(probably spoofed)\n", s, __func__); 1139 goto failed; 1140 } 1141 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1142 /* If received ACK has MD5 signature, check it. */ 1143 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1144 (!TCPMD5_ENABLED() || 1145 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1146 /* Drop the ACK. */ 1147 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1148 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1149 "MD5 signature doesn't match.\n", 1150 s, __func__); 1151 free(s, M_TCPLOG); 1152 } 1153 TCPSTAT_INC(tcps_sig_err_sigopt); 1154 return (-1); /* Do not send RST */ 1155 } 1156 #endif /* TCP_SIGNATURE */ 1157 } else { 1158 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1159 /* 1160 * If listening socket requested TCP digests, check that 1161 * received ACK has signature and it is correct. 1162 * If not, drop the ACK and leave sc entry in th cache, 1163 * because SYN was received with correct signature. 1164 */ 1165 if (sc->sc_flags & SCF_SIGNATURE) { 1166 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1167 /* No signature */ 1168 TCPSTAT_INC(tcps_sig_err_nosigopt); 1169 SCH_UNLOCK(sch); 1170 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1171 log(LOG_DEBUG, "%s; %s: Segment " 1172 "rejected, MD5 signature wasn't " 1173 "provided.\n", s, __func__); 1174 free(s, M_TCPLOG); 1175 } 1176 return (-1); /* Do not send RST */ 1177 } 1178 if (!TCPMD5_ENABLED() || 1179 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1180 /* Doesn't match or no SA */ 1181 SCH_UNLOCK(sch); 1182 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1183 log(LOG_DEBUG, "%s; %s: Segment " 1184 "rejected, MD5 signature doesn't " 1185 "match.\n", s, __func__); 1186 free(s, M_TCPLOG); 1187 } 1188 return (-1); /* Do not send RST */ 1189 } 1190 } 1191 #endif /* TCP_SIGNATURE */ 1192 1193 /* 1194 * RFC 7323 PAWS: If we have a timestamp on this segment and 1195 * it's less than ts_recent, drop it. 1196 * XXXMT: RFC 7323 also requires to send an ACK. 1197 * In tcp_input.c this is only done for TCP segments 1198 * with user data, so be consistent here and just drop 1199 * the segment. 1200 */ 1201 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1202 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1203 SCH_UNLOCK(sch); 1204 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1205 log(LOG_DEBUG, 1206 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1207 "segment dropped\n", s, __func__, 1208 to->to_tsval, sc->sc_tsreflect); 1209 free(s, M_TCPLOG); 1210 } 1211 return (-1); /* Do not send RST */ 1212 } 1213 1214 /* 1215 * Pull out the entry to unlock the bucket row. 1216 * 1217 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1218 * tcp_state_change(). The tcpcb is not existent at this 1219 * moment. A new one will be allocated via syncache_socket-> 1220 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1221 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1222 */ 1223 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1224 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1225 sch->sch_length--; 1226 #ifdef TCP_OFFLOAD 1227 if (ADDED_BY_TOE(sc)) { 1228 struct toedev *tod = sc->sc_tod; 1229 1230 tod->tod_syncache_removed(tod, sc->sc_todctx); 1231 } 1232 #endif 1233 SCH_UNLOCK(sch); 1234 } 1235 1236 /* 1237 * Segment validation: 1238 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1239 */ 1240 if (th->th_ack != sc->sc_iss + 1) { 1241 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1242 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1243 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1244 goto failed; 1245 } 1246 1247 /* 1248 * The SEQ must fall in the window starting at the received 1249 * initial receive sequence number + 1 (the SYN). 1250 */ 1251 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1252 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1253 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1254 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1255 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1256 goto failed; 1257 } 1258 1259 /* 1260 * If timestamps were not negotiated during SYN/ACK they 1261 * must not appear on any segment during this session. 1262 */ 1263 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1264 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1265 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1266 "segment rejected\n", s, __func__); 1267 goto failed; 1268 } 1269 1270 /* 1271 * If timestamps were negotiated during SYN/ACK they should 1272 * appear on every segment during this session. 1273 * XXXAO: This is only informal as there have been unverified 1274 * reports of non-compliants stacks. 1275 */ 1276 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1277 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1278 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1279 "no action\n", s, __func__); 1280 free(s, M_TCPLOG); 1281 s = NULL; 1282 } 1283 } 1284 1285 *lsop = syncache_socket(sc, *lsop, m); 1286 1287 if (*lsop == NULL) 1288 TCPSTAT_INC(tcps_sc_aborted); 1289 else 1290 TCPSTAT_INC(tcps_sc_completed); 1291 1292 /* how do we find the inp for the new socket? */ 1293 if (sc != &scs) 1294 syncache_free(sc); 1295 return (1); 1296 failed: 1297 if (sc != NULL && sc != &scs) 1298 syncache_free(sc); 1299 if (s != NULL) 1300 free(s, M_TCPLOG); 1301 *lsop = NULL; 1302 return (0); 1303 } 1304 1305 static void 1306 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1307 uint64_t response_cookie) 1308 { 1309 struct inpcb *inp; 1310 struct tcpcb *tp; 1311 unsigned int *pending_counter; 1312 1313 NET_EPOCH_ASSERT(); 1314 1315 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1316 *lsop = syncache_socket(sc, *lsop, m); 1317 if (*lsop == NULL) { 1318 TCPSTAT_INC(tcps_sc_aborted); 1319 atomic_subtract_int(pending_counter, 1); 1320 } else { 1321 soisconnected(*lsop); 1322 inp = sotoinpcb(*lsop); 1323 tp = intotcpcb(inp); 1324 tp->t_flags |= TF_FASTOPEN; 1325 tp->t_tfo_cookie.server = response_cookie; 1326 tp->snd_max = tp->iss; 1327 tp->snd_nxt = tp->iss; 1328 tp->t_tfo_pending = pending_counter; 1329 TCPSTAT_INC(tcps_sc_completed); 1330 } 1331 } 1332 1333 /* 1334 * Given a LISTEN socket and an inbound SYN request, add 1335 * this to the syn cache, and send back a segment: 1336 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1337 * to the source. 1338 * 1339 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1340 * Doing so would require that we hold onto the data and deliver it 1341 * to the application. However, if we are the target of a SYN-flood 1342 * DoS attack, an attacker could send data which would eventually 1343 * consume all available buffer space if it were ACKed. By not ACKing 1344 * the data, we avoid this DoS scenario. 1345 * 1346 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1347 * cookie is processed and a new socket is created. In this case, any data 1348 * accompanying the SYN will be queued to the socket by tcp_input() and will 1349 * be ACKed either when the application sends response data or the delayed 1350 * ACK timer expires, whichever comes first. 1351 */ 1352 int 1353 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1354 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1355 void *todctx, uint8_t iptos) 1356 { 1357 struct tcpcb *tp; 1358 struct socket *so; 1359 struct syncache *sc = NULL; 1360 struct syncache_head *sch; 1361 struct mbuf *ipopts = NULL; 1362 u_int ltflags; 1363 int win, ip_ttl, ip_tos; 1364 char *s; 1365 int rv = 0; 1366 #ifdef INET6 1367 int autoflowlabel = 0; 1368 #endif 1369 #ifdef MAC 1370 struct label *maclabel; 1371 #endif 1372 struct syncache scs; 1373 struct ucred *cred; 1374 uint64_t tfo_response_cookie; 1375 unsigned int *tfo_pending = NULL; 1376 int tfo_cookie_valid = 0; 1377 int tfo_response_cookie_valid = 0; 1378 bool locked; 1379 1380 INP_WLOCK_ASSERT(inp); /* listen socket */ 1381 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1382 ("%s: unexpected tcp flags", __func__)); 1383 1384 /* 1385 * Combine all so/tp operations very early to drop the INP lock as 1386 * soon as possible. 1387 */ 1388 so = *lsop; 1389 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1390 tp = sototcpcb(so); 1391 cred = crhold(so->so_cred); 1392 1393 #ifdef INET6 1394 if (inc->inc_flags & INC_ISIPV6) { 1395 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1396 autoflowlabel = 1; 1397 } 1398 ip_ttl = in6_selecthlim(inp, NULL); 1399 if ((inp->in6p_outputopts == NULL) || 1400 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1401 ip_tos = 0; 1402 } else { 1403 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1404 } 1405 } 1406 #endif 1407 #if defined(INET6) && defined(INET) 1408 else 1409 #endif 1410 #ifdef INET 1411 { 1412 ip_ttl = inp->inp_ip_ttl; 1413 ip_tos = inp->inp_ip_tos; 1414 } 1415 #endif 1416 win = so->sol_sbrcv_hiwat; 1417 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1418 1419 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1420 (tp->t_tfo_pending != NULL) && 1421 (to->to_flags & TOF_FASTOPEN)) { 1422 /* 1423 * Limit the number of pending TFO connections to 1424 * approximately half of the queue limit. This prevents TFO 1425 * SYN floods from starving the service by filling the 1426 * listen queue with bogus TFO connections. 1427 */ 1428 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1429 (so->sol_qlimit / 2)) { 1430 int result; 1431 1432 result = tcp_fastopen_check_cookie(inc, 1433 to->to_tfo_cookie, to->to_tfo_len, 1434 &tfo_response_cookie); 1435 tfo_cookie_valid = (result > 0); 1436 tfo_response_cookie_valid = (result >= 0); 1437 } 1438 1439 /* 1440 * Remember the TFO pending counter as it will have to be 1441 * decremented below if we don't make it to syncache_tfo_expand(). 1442 */ 1443 tfo_pending = tp->t_tfo_pending; 1444 } 1445 1446 /* By the time we drop the lock these should no longer be used. */ 1447 so = NULL; 1448 tp = NULL; 1449 1450 #ifdef MAC 1451 if (mac_syncache_init(&maclabel) != 0) { 1452 INP_WUNLOCK(inp); 1453 goto done; 1454 } else 1455 mac_syncache_create(maclabel, inp); 1456 #endif 1457 if (!tfo_cookie_valid) 1458 INP_WUNLOCK(inp); 1459 1460 /* 1461 * Remember the IP options, if any. 1462 */ 1463 #ifdef INET6 1464 if (!(inc->inc_flags & INC_ISIPV6)) 1465 #endif 1466 #ifdef INET 1467 ipopts = (m) ? ip_srcroute(m) : NULL; 1468 #else 1469 ipopts = NULL; 1470 #endif 1471 1472 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1473 /* 1474 * If listening socket requested TCP digests, check that received 1475 * SYN has signature and it is correct. If signature doesn't match 1476 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1477 */ 1478 if (ltflags & TF_SIGNATURE) { 1479 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1480 TCPSTAT_INC(tcps_sig_err_nosigopt); 1481 goto done; 1482 } 1483 if (!TCPMD5_ENABLED() || 1484 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1485 goto done; 1486 } 1487 #endif /* TCP_SIGNATURE */ 1488 /* 1489 * See if we already have an entry for this connection. 1490 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1491 * 1492 * XXX: should the syncache be re-initialized with the contents 1493 * of the new SYN here (which may have different options?) 1494 * 1495 * XXX: We do not check the sequence number to see if this is a 1496 * real retransmit or a new connection attempt. The question is 1497 * how to handle such a case; either ignore it as spoofed, or 1498 * drop the current entry and create a new one? 1499 */ 1500 if (syncache_cookiesonly()) { 1501 sc = NULL; 1502 sch = syncache_hashbucket(inc); 1503 locked = false; 1504 } else { 1505 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1506 locked = true; 1507 SCH_LOCK_ASSERT(sch); 1508 } 1509 if (sc != NULL) { 1510 if (tfo_cookie_valid) 1511 INP_WUNLOCK(inp); 1512 TCPSTAT_INC(tcps_sc_dupsyn); 1513 if (ipopts) { 1514 /* 1515 * If we were remembering a previous source route, 1516 * forget it and use the new one we've been given. 1517 */ 1518 if (sc->sc_ipopts) 1519 (void) m_free(sc->sc_ipopts); 1520 sc->sc_ipopts = ipopts; 1521 } 1522 /* 1523 * Update timestamp if present. 1524 */ 1525 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1526 sc->sc_tsreflect = to->to_tsval; 1527 else 1528 sc->sc_flags &= ~SCF_TIMESTAMP; 1529 /* 1530 * Disable ECN if needed. 1531 */ 1532 if ((sc->sc_flags & SCF_ECN) && 1533 ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) { 1534 sc->sc_flags &= ~SCF_ECN; 1535 } 1536 #ifdef MAC 1537 /* 1538 * Since we have already unconditionally allocated label 1539 * storage, free it up. The syncache entry will already 1540 * have an initialized label we can use. 1541 */ 1542 mac_syncache_destroy(&maclabel); 1543 #endif 1544 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1545 /* Retransmit SYN|ACK and reset retransmit count. */ 1546 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1547 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1548 "resetting timer and retransmitting SYN|ACK\n", 1549 s, __func__); 1550 free(s, M_TCPLOG); 1551 } 1552 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1553 sc->sc_rxmits = 0; 1554 syncache_timeout(sc, sch, 1); 1555 TCPSTAT_INC(tcps_sndacks); 1556 TCPSTAT_INC(tcps_sndtotal); 1557 } 1558 SCH_UNLOCK(sch); 1559 goto donenoprobe; 1560 } 1561 1562 if (tfo_cookie_valid) { 1563 bzero(&scs, sizeof(scs)); 1564 sc = &scs; 1565 goto skip_alloc; 1566 } 1567 1568 /* 1569 * Skip allocating a syncache entry if we are just going to discard 1570 * it later. 1571 */ 1572 if (!locked) { 1573 bzero(&scs, sizeof(scs)); 1574 sc = &scs; 1575 } else 1576 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1577 if (sc == NULL) { 1578 /* 1579 * The zone allocator couldn't provide more entries. 1580 * Treat this as if the cache was full; drop the oldest 1581 * entry and insert the new one. 1582 */ 1583 TCPSTAT_INC(tcps_sc_zonefail); 1584 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1585 sch->sch_last_overflow = time_uptime; 1586 syncache_drop(sc, sch); 1587 syncache_pause(inc); 1588 } 1589 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1590 if (sc == NULL) { 1591 if (V_tcp_syncookies) { 1592 bzero(&scs, sizeof(scs)); 1593 sc = &scs; 1594 } else { 1595 KASSERT(locked, 1596 ("%s: bucket unexpectedly unlocked", 1597 __func__)); 1598 SCH_UNLOCK(sch); 1599 if (ipopts) 1600 (void) m_free(ipopts); 1601 goto done; 1602 } 1603 } 1604 } 1605 1606 skip_alloc: 1607 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1608 sc->sc_tfo_cookie = &tfo_response_cookie; 1609 1610 /* 1611 * Fill in the syncache values. 1612 */ 1613 #ifdef MAC 1614 sc->sc_label = maclabel; 1615 #endif 1616 sc->sc_cred = cred; 1617 cred = NULL; 1618 sc->sc_ipopts = ipopts; 1619 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1620 sc->sc_ip_tos = ip_tos; 1621 sc->sc_ip_ttl = ip_ttl; 1622 #ifdef TCP_OFFLOAD 1623 sc->sc_tod = tod; 1624 sc->sc_todctx = todctx; 1625 #endif 1626 sc->sc_irs = th->th_seq; 1627 sc->sc_flags = 0; 1628 sc->sc_flowlabel = 0; 1629 1630 /* 1631 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1632 * win was derived from socket earlier in the function. 1633 */ 1634 win = imax(win, 0); 1635 win = imin(win, TCP_MAXWIN); 1636 sc->sc_wnd = win; 1637 1638 if (V_tcp_do_rfc1323) { 1639 /* 1640 * A timestamp received in a SYN makes 1641 * it ok to send timestamp requests and replies. 1642 */ 1643 if (to->to_flags & TOF_TS) { 1644 sc->sc_tsreflect = to->to_tsval; 1645 sc->sc_flags |= SCF_TIMESTAMP; 1646 sc->sc_tsoff = tcp_new_ts_offset(inc); 1647 } 1648 if (to->to_flags & TOF_SCALE) { 1649 int wscale = 0; 1650 1651 /* 1652 * Pick the smallest possible scaling factor that 1653 * will still allow us to scale up to sb_max, aka 1654 * kern.ipc.maxsockbuf. 1655 * 1656 * We do this because there are broken firewalls that 1657 * will corrupt the window scale option, leading to 1658 * the other endpoint believing that our advertised 1659 * window is unscaled. At scale factors larger than 1660 * 5 the unscaled window will drop below 1500 bytes, 1661 * leading to serious problems when traversing these 1662 * broken firewalls. 1663 * 1664 * With the default maxsockbuf of 256K, a scale factor 1665 * of 3 will be chosen by this algorithm. Those who 1666 * choose a larger maxsockbuf should watch out 1667 * for the compatibility problems mentioned above. 1668 * 1669 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1670 * or <SYN,ACK>) segment itself is never scaled. 1671 */ 1672 while (wscale < TCP_MAX_WINSHIFT && 1673 (TCP_MAXWIN << wscale) < sb_max) 1674 wscale++; 1675 sc->sc_requested_r_scale = wscale; 1676 sc->sc_requested_s_scale = to->to_wscale; 1677 sc->sc_flags |= SCF_WINSCALE; 1678 } 1679 } 1680 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1681 /* 1682 * If listening socket requested TCP digests, flag this in the 1683 * syncache so that syncache_respond() will do the right thing 1684 * with the SYN+ACK. 1685 */ 1686 if (ltflags & TF_SIGNATURE) 1687 sc->sc_flags |= SCF_SIGNATURE; 1688 #endif /* TCP_SIGNATURE */ 1689 if (to->to_flags & TOF_SACKPERM) 1690 sc->sc_flags |= SCF_SACK; 1691 if (to->to_flags & TOF_MSS) 1692 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1693 if (ltflags & TF_NOOPT) 1694 sc->sc_flags |= SCF_NOOPT; 1695 if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) && 1696 V_tcp_do_ecn) 1697 sc->sc_flags |= SCF_ECN; 1698 1699 if (V_tcp_syncookies) 1700 sc->sc_iss = syncookie_generate(sch, sc); 1701 else 1702 sc->sc_iss = arc4random(); 1703 #ifdef INET6 1704 if (autoflowlabel) { 1705 if (V_tcp_syncookies) 1706 sc->sc_flowlabel = sc->sc_iss; 1707 else 1708 sc->sc_flowlabel = ip6_randomflowlabel(); 1709 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1710 } 1711 #endif 1712 if (locked) 1713 SCH_UNLOCK(sch); 1714 1715 if (tfo_cookie_valid) { 1716 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1717 /* INP_WUNLOCK(inp) will be performed by the caller */ 1718 rv = 1; 1719 goto tfo_expanded; 1720 } 1721 1722 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1723 /* 1724 * Do a standard 3-way handshake. 1725 */ 1726 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1727 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1728 syncache_free(sc); 1729 else if (sc != &scs) 1730 syncache_insert(sc, sch); /* locks and unlocks sch */ 1731 TCPSTAT_INC(tcps_sndacks); 1732 TCPSTAT_INC(tcps_sndtotal); 1733 } else { 1734 if (sc != &scs) 1735 syncache_free(sc); 1736 TCPSTAT_INC(tcps_sc_dropped); 1737 } 1738 goto donenoprobe; 1739 1740 done: 1741 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1742 donenoprobe: 1743 if (m) { 1744 *lsop = NULL; 1745 m_freem(m); 1746 } 1747 /* 1748 * If tfo_pending is not NULL here, then a TFO SYN that did not 1749 * result in a new socket was processed and the associated pending 1750 * counter has not yet been decremented. All such TFO processing paths 1751 * transit this point. 1752 */ 1753 if (tfo_pending != NULL) 1754 tcp_fastopen_decrement_counter(tfo_pending); 1755 1756 tfo_expanded: 1757 if (cred != NULL) 1758 crfree(cred); 1759 #ifdef MAC 1760 if (sc == &scs) 1761 mac_syncache_destroy(&maclabel); 1762 #endif 1763 return (rv); 1764 } 1765 1766 /* 1767 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1768 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1769 */ 1770 static int 1771 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags) 1772 { 1773 struct ip *ip = NULL; 1774 struct mbuf *m; 1775 struct tcphdr *th = NULL; 1776 int optlen, error = 0; /* Make compiler happy */ 1777 u_int16_t hlen, tlen, mssopt; 1778 struct tcpopt to; 1779 #ifdef INET6 1780 struct ip6_hdr *ip6 = NULL; 1781 #endif 1782 1783 NET_EPOCH_ASSERT(); 1784 1785 hlen = 1786 #ifdef INET6 1787 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1788 #endif 1789 sizeof(struct ip); 1790 tlen = hlen + sizeof(struct tcphdr); 1791 1792 /* Determine MSS we advertize to other end of connection. */ 1793 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); 1794 1795 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1796 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1797 ("syncache: mbuf too small")); 1798 1799 /* Create the IP+TCP header from scratch. */ 1800 m = m_gethdr(M_NOWAIT, MT_DATA); 1801 if (m == NULL) 1802 return (ENOBUFS); 1803 #ifdef MAC 1804 mac_syncache_create_mbuf(sc->sc_label, m); 1805 #endif 1806 m->m_data += max_linkhdr; 1807 m->m_len = tlen; 1808 m->m_pkthdr.len = tlen; 1809 m->m_pkthdr.rcvif = NULL; 1810 1811 #ifdef INET6 1812 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1813 ip6 = mtod(m, struct ip6_hdr *); 1814 ip6->ip6_vfc = IPV6_VERSION; 1815 ip6->ip6_nxt = IPPROTO_TCP; 1816 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1817 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1818 ip6->ip6_plen = htons(tlen - hlen); 1819 /* ip6_hlim is set after checksum */ 1820 /* Zero out traffic class and flow label. */ 1821 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1822 ip6->ip6_flow |= sc->sc_flowlabel; 1823 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20); 1824 1825 th = (struct tcphdr *)(ip6 + 1); 1826 } 1827 #endif 1828 #if defined(INET6) && defined(INET) 1829 else 1830 #endif 1831 #ifdef INET 1832 { 1833 ip = mtod(m, struct ip *); 1834 ip->ip_v = IPVERSION; 1835 ip->ip_hl = sizeof(struct ip) >> 2; 1836 ip->ip_len = htons(tlen); 1837 ip->ip_id = 0; 1838 ip->ip_off = 0; 1839 ip->ip_sum = 0; 1840 ip->ip_p = IPPROTO_TCP; 1841 ip->ip_src = sc->sc_inc.inc_laddr; 1842 ip->ip_dst = sc->sc_inc.inc_faddr; 1843 ip->ip_ttl = sc->sc_ip_ttl; 1844 ip->ip_tos = sc->sc_ip_tos; 1845 1846 /* 1847 * See if we should do MTU discovery. Route lookups are 1848 * expensive, so we will only unset the DF bit if: 1849 * 1850 * 1) path_mtu_discovery is disabled 1851 * 2) the SCF_UNREACH flag has been set 1852 */ 1853 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1854 ip->ip_off |= htons(IP_DF); 1855 1856 th = (struct tcphdr *)(ip + 1); 1857 } 1858 #endif /* INET */ 1859 th->th_sport = sc->sc_inc.inc_lport; 1860 th->th_dport = sc->sc_inc.inc_fport; 1861 1862 if (flags & TH_SYN) 1863 th->th_seq = htonl(sc->sc_iss); 1864 else 1865 th->th_seq = htonl(sc->sc_iss + 1); 1866 th->th_ack = htonl(sc->sc_irs + 1); 1867 th->th_off = sizeof(struct tcphdr) >> 2; 1868 th->th_x2 = 0; 1869 th->th_flags = flags; 1870 th->th_win = htons(sc->sc_wnd); 1871 th->th_urp = 0; 1872 1873 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) { 1874 th->th_flags |= TH_ECE; 1875 TCPSTAT_INC(tcps_ecn_shs); 1876 } 1877 1878 /* Tack on the TCP options. */ 1879 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1880 to.to_flags = 0; 1881 1882 if (flags & TH_SYN) { 1883 to.to_mss = mssopt; 1884 to.to_flags = TOF_MSS; 1885 if (sc->sc_flags & SCF_WINSCALE) { 1886 to.to_wscale = sc->sc_requested_r_scale; 1887 to.to_flags |= TOF_SCALE; 1888 } 1889 if (sc->sc_flags & SCF_SACK) 1890 to.to_flags |= TOF_SACKPERM; 1891 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1892 if (sc->sc_flags & SCF_SIGNATURE) 1893 to.to_flags |= TOF_SIGNATURE; 1894 #endif 1895 if (sc->sc_tfo_cookie) { 1896 to.to_flags |= TOF_FASTOPEN; 1897 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1898 to.to_tfo_cookie = sc->sc_tfo_cookie; 1899 /* don't send cookie again when retransmitting response */ 1900 sc->sc_tfo_cookie = NULL; 1901 } 1902 } 1903 if (sc->sc_flags & SCF_TIMESTAMP) { 1904 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1905 to.to_tsecr = sc->sc_tsreflect; 1906 to.to_flags |= TOF_TS; 1907 } 1908 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1909 1910 /* Adjust headers by option size. */ 1911 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1912 m->m_len += optlen; 1913 m->m_pkthdr.len += optlen; 1914 #ifdef INET6 1915 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1916 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1917 else 1918 #endif 1919 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1920 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1921 if (sc->sc_flags & SCF_SIGNATURE) { 1922 KASSERT(to.to_flags & TOF_SIGNATURE, 1923 ("tcp_addoptions() didn't set tcp_signature")); 1924 1925 /* NOTE: to.to_signature is inside of mbuf */ 1926 if (!TCPMD5_ENABLED() || 1927 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1928 m_freem(m); 1929 return (EACCES); 1930 } 1931 } 1932 #endif 1933 } else 1934 optlen = 0; 1935 1936 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1937 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1938 /* 1939 * If we have peer's SYN and it has a flowid, then let's assign it to 1940 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1941 * to SYN|ACK due to lack of inp here. 1942 */ 1943 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1944 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1945 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1946 } 1947 #ifdef INET6 1948 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1949 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1950 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1951 IPPROTO_TCP, 0); 1952 ip6->ip6_hlim = sc->sc_ip_ttl; 1953 #ifdef TCP_OFFLOAD 1954 if (ADDED_BY_TOE(sc)) { 1955 struct toedev *tod = sc->sc_tod; 1956 1957 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1958 1959 return (error); 1960 } 1961 #endif 1962 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 1963 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1964 } 1965 #endif 1966 #if defined(INET6) && defined(INET) 1967 else 1968 #endif 1969 #ifdef INET 1970 { 1971 m->m_pkthdr.csum_flags = CSUM_TCP; 1972 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1973 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1974 #ifdef TCP_OFFLOAD 1975 if (ADDED_BY_TOE(sc)) { 1976 struct toedev *tod = sc->sc_tod; 1977 1978 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1979 1980 return (error); 1981 } 1982 #endif 1983 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 1984 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1985 } 1986 #endif 1987 return (error); 1988 } 1989 1990 /* 1991 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1992 * that exceed the capacity of the syncache by avoiding the storage of any 1993 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1994 * attacks where the attacker does not have access to our responses. 1995 * 1996 * Syncookies encode and include all necessary information about the 1997 * connection setup within the SYN|ACK that we send back. That way we 1998 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1999 * (if ever). Normally the syncache and syncookies are running in parallel 2000 * with the latter taking over when the former is exhausted. When matching 2001 * syncache entry is found the syncookie is ignored. 2002 * 2003 * The only reliable information persisting the 3WHS is our initial sequence 2004 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 2005 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 2006 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 2007 * returns and signifies a legitimate connection if it matches the ACK. 2008 * 2009 * The available space of 32 bits to store the hash and to encode the SYN 2010 * option information is very tight and we should have at least 24 bits for 2011 * the MAC to keep the number of guesses by blind spoofing reasonably high. 2012 * 2013 * SYN option information we have to encode to fully restore a connection: 2014 * MSS: is imporant to chose an optimal segment size to avoid IP level 2015 * fragmentation along the path. The common MSS values can be encoded 2016 * in a 3-bit table. Uncommon values are captured by the next lower value 2017 * in the table leading to a slight increase in packetization overhead. 2018 * WSCALE: is necessary to allow large windows to be used for high delay- 2019 * bandwidth product links. Not scaling the window when it was initially 2020 * negotiated is bad for performance as lack of scaling further decreases 2021 * the apparent available send window. We only need to encode the WSCALE 2022 * we received from the remote end. Our end can be recalculated at any 2023 * time. The common WSCALE values can be encoded in a 3-bit table. 2024 * Uncommon values are captured by the next lower value in the table 2025 * making us under-estimate the available window size halving our 2026 * theoretically possible maximum throughput for that connection. 2027 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 2028 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 2029 * that are included in all segments on a connection. We enable them when 2030 * the ACK has them. 2031 * 2032 * Security of syncookies and attack vectors: 2033 * 2034 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 2035 * together with the gloabl secret to make it unique per connection attempt. 2036 * Thus any change of any of those parameters results in a different MAC output 2037 * in an unpredictable way unless a collision is encountered. 24 bits of the 2038 * MAC are embedded into the ISS. 2039 * 2040 * To prevent replay attacks two rotating global secrets are updated with a 2041 * new random value every 15 seconds. The life-time of a syncookie is thus 2042 * 15-30 seconds. 2043 * 2044 * Vector 1: Attacking the secret. This requires finding a weakness in the 2045 * MAC itself or the way it is used here. The attacker can do a chosen plain 2046 * text attack by varying and testing the all parameters under his control. 2047 * The strength depends on the size and randomness of the secret, and the 2048 * cryptographic security of the MAC function. Due to the constant updating 2049 * of the secret the attacker has at most 29.999 seconds to find the secret 2050 * and launch spoofed connections. After that he has to start all over again. 2051 * 2052 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 2053 * size an average of 4,823 attempts are required for a 50% chance of success 2054 * to spoof a single syncookie (birthday collision paradox). However the 2055 * attacker is blind and doesn't know if one of his attempts succeeded unless 2056 * he has a side channel to interfere success from. A single connection setup 2057 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 2058 * This many attempts are required for each one blind spoofed connection. For 2059 * every additional spoofed connection he has to launch another N attempts. 2060 * Thus for a sustained rate 100 spoofed connections per second approximately 2061 * 1,800,000 packets per second would have to be sent. 2062 * 2063 * NB: The MAC function should be fast so that it doesn't become a CPU 2064 * exhaustion attack vector itself. 2065 * 2066 * References: 2067 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 2068 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 2069 * http://cr.yp.to/syncookies.html (overview) 2070 * http://cr.yp.to/syncookies/archive (details) 2071 * 2072 * 2073 * Schematic construction of a syncookie enabled Initial Sequence Number: 2074 * 0 1 2 3 2075 * 12345678901234567890123456789012 2076 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2077 * 2078 * x 24 MAC (truncated) 2079 * W 3 Send Window Scale index 2080 * M 3 MSS index 2081 * S 1 SACK permitted 2082 * P 1 Odd/even secret 2083 */ 2084 2085 /* 2086 * Distribution and probability of certain MSS values. Those in between are 2087 * rounded down to the next lower one. 2088 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2089 * .2% .3% 5% 7% 7% 20% 15% 45% 2090 */ 2091 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2092 2093 /* 2094 * Distribution and probability of certain WSCALE values. We have to map the 2095 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2096 * bits based on prevalence of certain values. Where we don't have an exact 2097 * match for are rounded down to the next lower one letting us under-estimate 2098 * the true available window. At the moment this would happen only for the 2099 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2100 * and window size). The absence of the WSCALE option (no scaling in either 2101 * direction) is encoded with index zero. 2102 * [WSCALE values histograms, Allman, 2012] 2103 * X 10 10 35 5 6 14 10% by host 2104 * X 11 4 5 5 18 49 3% by connections 2105 */ 2106 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2107 2108 /* 2109 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2110 * and good cryptographic properties. 2111 */ 2112 static uint32_t 2113 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2114 uint8_t *secbits, uintptr_t secmod) 2115 { 2116 SIPHASH_CTX ctx; 2117 uint32_t siphash[2]; 2118 2119 SipHash24_Init(&ctx); 2120 SipHash_SetKey(&ctx, secbits); 2121 switch (inc->inc_flags & INC_ISIPV6) { 2122 #ifdef INET 2123 case 0: 2124 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2125 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2126 break; 2127 #endif 2128 #ifdef INET6 2129 case INC_ISIPV6: 2130 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2131 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2132 break; 2133 #endif 2134 } 2135 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2136 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2137 SipHash_Update(&ctx, &irs, sizeof(irs)); 2138 SipHash_Update(&ctx, &flags, sizeof(flags)); 2139 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2140 SipHash_Final((u_int8_t *)&siphash, &ctx); 2141 2142 return (siphash[0] ^ siphash[1]); 2143 } 2144 2145 static tcp_seq 2146 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2147 { 2148 u_int i, secbit, wscale; 2149 uint32_t iss, hash; 2150 uint8_t *secbits; 2151 union syncookie cookie; 2152 2153 cookie.cookie = 0; 2154 2155 /* Map our computed MSS into the 3-bit index. */ 2156 for (i = nitems(tcp_sc_msstab) - 1; 2157 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2158 i--) 2159 ; 2160 cookie.flags.mss_idx = i; 2161 2162 /* 2163 * Map the send window scale into the 3-bit index but only if 2164 * the wscale option was received. 2165 */ 2166 if (sc->sc_flags & SCF_WINSCALE) { 2167 wscale = sc->sc_requested_s_scale; 2168 for (i = nitems(tcp_sc_wstab) - 1; 2169 tcp_sc_wstab[i] > wscale && i > 0; 2170 i--) 2171 ; 2172 cookie.flags.wscale_idx = i; 2173 } 2174 2175 /* Can we do SACK? */ 2176 if (sc->sc_flags & SCF_SACK) 2177 cookie.flags.sack_ok = 1; 2178 2179 /* Which of the two secrets to use. */ 2180 secbit = V_tcp_syncache.secret.oddeven & 0x1; 2181 cookie.flags.odd_even = secbit; 2182 2183 secbits = V_tcp_syncache.secret.key[secbit]; 2184 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2185 (uintptr_t)sch); 2186 2187 /* 2188 * Put the flags into the hash and XOR them to get better ISS number 2189 * variance. This doesn't enhance the cryptographic strength and is 2190 * done to prevent the 8 cookie bits from showing up directly on the 2191 * wire. 2192 */ 2193 iss = hash & ~0xff; 2194 iss |= cookie.cookie ^ (hash >> 24); 2195 2196 TCPSTAT_INC(tcps_sc_sendcookie); 2197 return (iss); 2198 } 2199 2200 static struct syncache * 2201 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2202 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2203 struct socket *lso) 2204 { 2205 uint32_t hash; 2206 uint8_t *secbits; 2207 tcp_seq ack, seq; 2208 int wnd, wscale = 0; 2209 union syncookie cookie; 2210 2211 /* 2212 * Pull information out of SYN-ACK/ACK and revert sequence number 2213 * advances. 2214 */ 2215 ack = th->th_ack - 1; 2216 seq = th->th_seq - 1; 2217 2218 /* 2219 * Unpack the flags containing enough information to restore the 2220 * connection. 2221 */ 2222 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2223 2224 /* Which of the two secrets to use. */ 2225 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; 2226 2227 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2228 2229 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2230 if ((ack & ~0xff) != (hash & ~0xff)) 2231 return (NULL); 2232 2233 /* Fill in the syncache values. */ 2234 sc->sc_flags = 0; 2235 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2236 sc->sc_ipopts = NULL; 2237 2238 sc->sc_irs = seq; 2239 sc->sc_iss = ack; 2240 2241 switch (inc->inc_flags & INC_ISIPV6) { 2242 #ifdef INET 2243 case 0: 2244 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2245 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2246 break; 2247 #endif 2248 #ifdef INET6 2249 case INC_ISIPV6: 2250 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2251 sc->sc_flowlabel = 2252 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2253 break; 2254 #endif 2255 } 2256 2257 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2258 2259 /* We can simply recompute receive window scale we sent earlier. */ 2260 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2261 wscale++; 2262 2263 /* Only use wscale if it was enabled in the orignal SYN. */ 2264 if (cookie.flags.wscale_idx > 0) { 2265 sc->sc_requested_r_scale = wscale; 2266 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2267 sc->sc_flags |= SCF_WINSCALE; 2268 } 2269 2270 wnd = lso->sol_sbrcv_hiwat; 2271 wnd = imax(wnd, 0); 2272 wnd = imin(wnd, TCP_MAXWIN); 2273 sc->sc_wnd = wnd; 2274 2275 if (cookie.flags.sack_ok) 2276 sc->sc_flags |= SCF_SACK; 2277 2278 if (to->to_flags & TOF_TS) { 2279 sc->sc_flags |= SCF_TIMESTAMP; 2280 sc->sc_tsreflect = to->to_tsval; 2281 sc->sc_tsoff = tcp_new_ts_offset(inc); 2282 } 2283 2284 if (to->to_flags & TOF_SIGNATURE) 2285 sc->sc_flags |= SCF_SIGNATURE; 2286 2287 sc->sc_rxmits = 0; 2288 2289 TCPSTAT_INC(tcps_sc_recvcookie); 2290 return (sc); 2291 } 2292 2293 #ifdef INVARIANTS 2294 static int 2295 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2296 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2297 struct socket *lso) 2298 { 2299 struct syncache scs, *scx; 2300 char *s; 2301 2302 bzero(&scs, sizeof(scs)); 2303 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2304 2305 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2306 return (0); 2307 2308 if (scx != NULL) { 2309 if (sc->sc_peer_mss != scx->sc_peer_mss) 2310 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2311 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2312 2313 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2314 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2315 s, __func__, sc->sc_requested_r_scale, 2316 scx->sc_requested_r_scale); 2317 2318 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2319 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2320 s, __func__, sc->sc_requested_s_scale, 2321 scx->sc_requested_s_scale); 2322 2323 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2324 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2325 } 2326 2327 if (s != NULL) 2328 free(s, M_TCPLOG); 2329 return (0); 2330 } 2331 #endif /* INVARIANTS */ 2332 2333 static void 2334 syncookie_reseed(void *arg) 2335 { 2336 struct tcp_syncache *sc = arg; 2337 uint8_t *secbits; 2338 int secbit; 2339 2340 /* 2341 * Reseeding the secret doesn't have to be protected by a lock. 2342 * It only must be ensured that the new random values are visible 2343 * to all CPUs in a SMP environment. The atomic with release 2344 * semantics ensures that. 2345 */ 2346 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2347 secbits = sc->secret.key[secbit]; 2348 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2349 atomic_add_rel_int(&sc->secret.oddeven, 1); 2350 2351 /* Reschedule ourself. */ 2352 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2353 } 2354 2355 /* 2356 * We have overflowed a bucket. Let's pause dealing with the syncache. 2357 * This function will increment the bucketoverflow statistics appropriately 2358 * (once per pause when pausing is enabled; otherwise, once per overflow). 2359 */ 2360 static void 2361 syncache_pause(struct in_conninfo *inc) 2362 { 2363 time_t delta; 2364 const char *s; 2365 2366 /* XXX: 2367 * 2. Add sysctl read here so we don't get the benefit of this 2368 * change without the new sysctl. 2369 */ 2370 2371 /* 2372 * Try an unlocked read. If we already know that another thread 2373 * has activated the feature, there is no need to proceed. 2374 */ 2375 if (V_tcp_syncache.paused) 2376 return; 2377 2378 /* Are cookied enabled? If not, we can't pause. */ 2379 if (!V_tcp_syncookies) { 2380 TCPSTAT_INC(tcps_sc_bucketoverflow); 2381 return; 2382 } 2383 2384 /* 2385 * We may be the first thread to find an overflow. Get the lock 2386 * and evaluate if we need to take action. 2387 */ 2388 mtx_lock(&V_tcp_syncache.pause_mtx); 2389 if (V_tcp_syncache.paused) { 2390 mtx_unlock(&V_tcp_syncache.pause_mtx); 2391 return; 2392 } 2393 2394 /* Activate protection. */ 2395 V_tcp_syncache.paused = true; 2396 TCPSTAT_INC(tcps_sc_bucketoverflow); 2397 2398 /* 2399 * Determine the last backoff time. If we are seeing a re-newed 2400 * attack within that same time after last reactivating the syncache, 2401 * consider it an extension of the same attack. 2402 */ 2403 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; 2404 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { 2405 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { 2406 delta <<= 1; 2407 V_tcp_syncache.pause_backoff++; 2408 } 2409 } else { 2410 delta = TCP_SYNCACHE_PAUSE_TIME; 2411 V_tcp_syncache.pause_backoff = 0; 2412 } 2413 2414 /* Log a warning, including IP addresses, if able. */ 2415 if (inc != NULL) 2416 s = tcp_log_addrs(inc, NULL, NULL, NULL); 2417 else 2418 s = (const char *)NULL; 2419 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " 2420 "the next %lld seconds%s%s%s\n", (long long)delta, 2421 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", 2422 (s != NULL) ? ")" : ""); 2423 free(__DECONST(void *, s), M_TCPLOG); 2424 2425 /* Use the calculated delta to set a new pause time. */ 2426 V_tcp_syncache.pause_until = time_uptime + delta; 2427 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, 2428 &V_tcp_syncache); 2429 mtx_unlock(&V_tcp_syncache.pause_mtx); 2430 } 2431 2432 /* Evaluate whether we need to unpause. */ 2433 static void 2434 syncache_unpause(void *arg) 2435 { 2436 struct tcp_syncache *sc; 2437 time_t delta; 2438 2439 sc = arg; 2440 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); 2441 callout_deactivate(&sc->pause_co); 2442 2443 /* 2444 * Check to make sure we are not running early. If the pause 2445 * time has expired, then deactivate the protection. 2446 */ 2447 if ((delta = sc->pause_until - time_uptime) > 0) 2448 callout_schedule(&sc->pause_co, delta * hz); 2449 else 2450 sc->paused = false; 2451 } 2452 2453 /* 2454 * Exports the syncache entries to userland so that netstat can display 2455 * them alongside the other sockets. This function is intended to be 2456 * called only from tcp_pcblist. 2457 * 2458 * Due to concurrency on an active system, the number of pcbs exported 2459 * may have no relation to max_pcbs. max_pcbs merely indicates the 2460 * amount of space the caller allocated for this function to use. 2461 */ 2462 int 2463 syncache_pcblist(struct sysctl_req *req) 2464 { 2465 struct xtcpcb xt; 2466 struct syncache *sc; 2467 struct syncache_head *sch; 2468 int error, i; 2469 2470 bzero(&xt, sizeof(xt)); 2471 xt.xt_len = sizeof(xt); 2472 xt.t_state = TCPS_SYN_RECEIVED; 2473 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2474 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2475 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2476 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2477 2478 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 2479 sch = &V_tcp_syncache.hashbase[i]; 2480 SCH_LOCK(sch); 2481 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2482 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2483 continue; 2484 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2485 xt.xt_inp.inp_vflag = INP_IPV6; 2486 else 2487 xt.xt_inp.inp_vflag = INP_IPV4; 2488 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2489 sizeof (struct in_conninfo)); 2490 error = SYSCTL_OUT(req, &xt, sizeof xt); 2491 if (error) { 2492 SCH_UNLOCK(sch); 2493 return (0); 2494 } 2495 } 2496 SCH_UNLOCK(sch); 2497 } 2498 2499 return (0); 2500 } 2501