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++) { 340 341 sch = &V_tcp_syncache.hashbase[i]; 342 callout_drain(&sch->sch_timer); 343 344 SCH_LOCK(sch); 345 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 346 syncache_drop(sc, sch); 347 SCH_UNLOCK(sch); 348 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 349 ("%s: sch->sch_bucket not empty", __func__)); 350 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 351 __func__, sch->sch_length)); 352 mtx_destroy(&sch->sch_mtx); 353 } 354 355 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 356 ("%s: cache_count not 0", __func__)); 357 358 /* Free the allocated global resources. */ 359 uma_zdestroy(V_tcp_syncache.zone); 360 free(V_tcp_syncache.hashbase, M_SYNCACHE); 361 mtx_destroy(&V_tcp_syncache.pause_mtx); 362 } 363 #endif 364 365 /* 366 * Inserts a syncache entry into the specified bucket row. 367 * Locks and unlocks the syncache_head autonomously. 368 */ 369 static void 370 syncache_insert(struct syncache *sc, struct syncache_head *sch) 371 { 372 struct syncache *sc2; 373 374 SCH_LOCK(sch); 375 376 /* 377 * Make sure that we don't overflow the per-bucket limit. 378 * If the bucket is full, toss the oldest element. 379 */ 380 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 381 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 382 ("sch->sch_length incorrect")); 383 syncache_pause(&sc->sc_inc); 384 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 385 sch->sch_last_overflow = time_uptime; 386 syncache_drop(sc2, sch); 387 } 388 389 /* Put it into the bucket. */ 390 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 391 sch->sch_length++; 392 393 #ifdef TCP_OFFLOAD 394 if (ADDED_BY_TOE(sc)) { 395 struct toedev *tod = sc->sc_tod; 396 397 tod->tod_syncache_added(tod, sc->sc_todctx); 398 } 399 #endif 400 401 /* Reinitialize the bucket row's timer. */ 402 if (sch->sch_length == 1) 403 sch->sch_nextc = ticks + INT_MAX; 404 syncache_timeout(sc, sch, 1); 405 406 SCH_UNLOCK(sch); 407 408 TCPSTATES_INC(TCPS_SYN_RECEIVED); 409 TCPSTAT_INC(tcps_sc_added); 410 } 411 412 /* 413 * Remove and free entry from syncache bucket row. 414 * Expects locked syncache head. 415 */ 416 static void 417 syncache_drop(struct syncache *sc, struct syncache_head *sch) 418 { 419 420 SCH_LOCK_ASSERT(sch); 421 422 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 423 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 424 sch->sch_length--; 425 426 #ifdef TCP_OFFLOAD 427 if (ADDED_BY_TOE(sc)) { 428 struct toedev *tod = sc->sc_tod; 429 430 tod->tod_syncache_removed(tod, sc->sc_todctx); 431 } 432 #endif 433 434 syncache_free(sc); 435 } 436 437 /* 438 * Engage/reengage time on bucket row. 439 */ 440 static void 441 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 442 { 443 int rexmt; 444 445 if (sc->sc_rxmits == 0) 446 rexmt = tcp_rexmit_initial; 447 else 448 TCPT_RANGESET(rexmt, 449 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], 450 tcp_rexmit_min, TCPTV_REXMTMAX); 451 sc->sc_rxttime = ticks + rexmt; 452 sc->sc_rxmits++; 453 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 454 sch->sch_nextc = sc->sc_rxttime; 455 if (docallout) 456 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 457 syncache_timer, (void *)sch); 458 } 459 } 460 461 /* 462 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 463 * If we have retransmitted an entry the maximum number of times, expire it. 464 * One separate timer for each bucket row. 465 */ 466 static void 467 syncache_timer(void *xsch) 468 { 469 struct syncache_head *sch = (struct syncache_head *)xsch; 470 struct syncache *sc, *nsc; 471 struct epoch_tracker et; 472 int tick = ticks; 473 char *s; 474 bool paused; 475 476 CURVNET_SET(sch->sch_sc->vnet); 477 478 /* NB: syncache_head has already been locked by the callout. */ 479 SCH_LOCK_ASSERT(sch); 480 481 /* 482 * In the following cycle we may remove some entries and/or 483 * advance some timeouts, so re-initialize the bucket timer. 484 */ 485 sch->sch_nextc = tick + INT_MAX; 486 487 /* 488 * If we have paused processing, unconditionally remove 489 * all syncache entries. 490 */ 491 mtx_lock(&V_tcp_syncache.pause_mtx); 492 paused = V_tcp_syncache.paused; 493 mtx_unlock(&V_tcp_syncache.pause_mtx); 494 495 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 496 if (paused) { 497 syncache_drop(sc, sch); 498 continue; 499 } 500 /* 501 * We do not check if the listen socket still exists 502 * and accept the case where the listen socket may be 503 * gone by the time we resend the SYN/ACK. We do 504 * not expect this to happens often. If it does, 505 * then the RST will be sent by the time the remote 506 * host does the SYN/ACK->ACK. 507 */ 508 if (TSTMP_GT(sc->sc_rxttime, tick)) { 509 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 510 sch->sch_nextc = sc->sc_rxttime; 511 continue; 512 } 513 if (sc->sc_rxmits > V_tcp_ecn_maxretries) { 514 sc->sc_flags &= ~SCF_ECN; 515 } 516 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 517 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 518 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 519 "giving up and removing syncache entry\n", 520 s, __func__); 521 free(s, M_TCPLOG); 522 } 523 syncache_drop(sc, sch); 524 TCPSTAT_INC(tcps_sc_stale); 525 continue; 526 } 527 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 528 log(LOG_DEBUG, "%s; %s: Response timeout, " 529 "retransmitting (%u) SYN|ACK\n", 530 s, __func__, sc->sc_rxmits); 531 free(s, M_TCPLOG); 532 } 533 534 NET_EPOCH_ENTER(et); 535 syncache_respond(sc, NULL, TH_SYN|TH_ACK); 536 NET_EPOCH_EXIT(et); 537 TCPSTAT_INC(tcps_sc_retransmitted); 538 syncache_timeout(sc, sch, 0); 539 } 540 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 541 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 542 syncache_timer, (void *)(sch)); 543 CURVNET_RESTORE(); 544 } 545 546 /* 547 * Returns true if the system is only using cookies at the moment. 548 * This could be due to a sysadmin decision to only use cookies, or it 549 * could be due to the system detecting an attack. 550 */ 551 static inline bool 552 syncache_cookiesonly(void) 553 { 554 555 return (V_tcp_syncookies && (V_tcp_syncache.paused || 556 V_tcp_syncookiesonly)); 557 } 558 559 /* 560 * Find the hash bucket for the given connection. 561 */ 562 static struct syncache_head * 563 syncache_hashbucket(struct in_conninfo *inc) 564 { 565 uint32_t hash; 566 567 /* 568 * The hash is built on foreign port + local port + foreign address. 569 * We rely on the fact that struct in_conninfo starts with 16 bits 570 * of foreign port, then 16 bits of local port then followed by 128 571 * bits of foreign address. In case of IPv4 address, the first 3 572 * 32-bit words of the address always are zeroes. 573 */ 574 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 575 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 576 577 return (&V_tcp_syncache.hashbase[hash]); 578 } 579 580 /* 581 * Find an entry in the syncache. 582 * Returns always with locked syncache_head plus a matching entry or NULL. 583 */ 584 static struct syncache * 585 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 586 { 587 struct syncache *sc; 588 struct syncache_head *sch; 589 590 *schp = sch = syncache_hashbucket(inc); 591 SCH_LOCK(sch); 592 593 /* Circle through bucket row to find matching entry. */ 594 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 595 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 596 sizeof(struct in_endpoints)) == 0) 597 break; 598 599 return (sc); /* Always returns with locked sch. */ 600 } 601 602 /* 603 * This function is called when we get a RST for a 604 * non-existent connection, so that we can see if the 605 * connection is in the syn cache. If it is, zap it. 606 * If required send a challenge ACK. 607 */ 608 void 609 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m) 610 { 611 struct syncache *sc; 612 struct syncache_head *sch; 613 char *s = NULL; 614 615 if (syncache_cookiesonly()) 616 return; 617 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 618 SCH_LOCK_ASSERT(sch); 619 620 /* 621 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 622 * See RFC 793 page 65, section SEGMENT ARRIVES. 623 */ 624 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 625 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 626 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 627 "FIN flag set, segment ignored\n", s, __func__); 628 TCPSTAT_INC(tcps_badrst); 629 goto done; 630 } 631 632 /* 633 * No corresponding connection was found in syncache. 634 * If syncookies are enabled and possibly exclusively 635 * used, or we are under memory pressure, a valid RST 636 * may not find a syncache entry. In that case we're 637 * done and no SYN|ACK retransmissions will happen. 638 * Otherwise the RST was misdirected or spoofed. 639 */ 640 if (sc == NULL) { 641 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 642 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 643 "syncache entry (possibly syncookie only), " 644 "segment ignored\n", s, __func__); 645 TCPSTAT_INC(tcps_badrst); 646 goto done; 647 } 648 649 /* 650 * If the RST bit is set, check the sequence number to see 651 * if this is a valid reset segment. 652 * 653 * RFC 793 page 37: 654 * In all states except SYN-SENT, all reset (RST) segments 655 * are validated by checking their SEQ-fields. A reset is 656 * valid if its sequence number is in the window. 657 * 658 * RFC 793 page 69: 659 * There are four cases for the acceptability test for an incoming 660 * segment: 661 * 662 * Segment Receive Test 663 * Length Window 664 * ------- ------- ------------------------------------------- 665 * 0 0 SEG.SEQ = RCV.NXT 666 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 667 * >0 0 not acceptable 668 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 669 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 670 * 671 * Note that when receiving a SYN segment in the LISTEN state, 672 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 673 * described in RFC 793, page 66. 674 */ 675 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 676 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 677 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 678 if (V_tcp_insecure_rst || 679 th->th_seq == sc->sc_irs + 1) { 680 syncache_drop(sc, sch); 681 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 682 log(LOG_DEBUG, 683 "%s; %s: Our SYN|ACK was rejected, " 684 "connection attempt aborted by remote " 685 "endpoint\n", 686 s, __func__); 687 TCPSTAT_INC(tcps_sc_reset); 688 } else { 689 TCPSTAT_INC(tcps_badrst); 690 /* Send challenge ACK. */ 691 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 692 log(LOG_DEBUG, "%s; %s: RST with invalid " 693 " SEQ %u != NXT %u (+WND %u), " 694 "sending challenge ACK\n", 695 s, __func__, 696 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 697 syncache_respond(sc, m, TH_ACK); 698 } 699 } else { 700 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 701 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 702 "NXT %u (+WND %u), segment ignored\n", 703 s, __func__, 704 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 705 TCPSTAT_INC(tcps_badrst); 706 } 707 708 done: 709 if (s != NULL) 710 free(s, M_TCPLOG); 711 SCH_UNLOCK(sch); 712 } 713 714 void 715 syncache_badack(struct in_conninfo *inc) 716 { 717 struct syncache *sc; 718 struct syncache_head *sch; 719 720 if (syncache_cookiesonly()) 721 return; 722 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 723 SCH_LOCK_ASSERT(sch); 724 if (sc != NULL) { 725 syncache_drop(sc, sch); 726 TCPSTAT_INC(tcps_sc_badack); 727 } 728 SCH_UNLOCK(sch); 729 } 730 731 void 732 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq) 733 { 734 struct syncache *sc; 735 struct syncache_head *sch; 736 737 if (syncache_cookiesonly()) 738 return; 739 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 740 SCH_LOCK_ASSERT(sch); 741 if (sc == NULL) 742 goto done; 743 744 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 745 if (ntohl(th_seq) != sc->sc_iss) 746 goto done; 747 748 /* 749 * If we've rertransmitted 3 times and this is our second error, 750 * we remove the entry. Otherwise, we allow it to continue on. 751 * This prevents us from incorrectly nuking an entry during a 752 * spurious network outage. 753 * 754 * See tcp_notify(). 755 */ 756 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 757 sc->sc_flags |= SCF_UNREACH; 758 goto done; 759 } 760 syncache_drop(sc, sch); 761 TCPSTAT_INC(tcps_sc_unreach); 762 done: 763 SCH_UNLOCK(sch); 764 } 765 766 /* 767 * Build a new TCP socket structure from a syncache entry. 768 * 769 * On success return the newly created socket with its underlying inp locked. 770 */ 771 static struct socket * 772 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 773 { 774 struct tcp_function_block *blk; 775 struct inpcb *inp = NULL; 776 struct socket *so; 777 struct tcpcb *tp; 778 int error; 779 char *s; 780 781 NET_EPOCH_ASSERT(); 782 783 /* 784 * Ok, create the full blown connection, and set things up 785 * as they would have been set up if we had created the 786 * connection when the SYN arrived. If we can't create 787 * the connection, abort it. 788 */ 789 so = sonewconn(lso, 0); 790 if (so == NULL) { 791 /* 792 * Drop the connection; we will either send a RST or 793 * have the peer retransmit its SYN again after its 794 * RTO and try again. 795 */ 796 TCPSTAT_INC(tcps_listendrop); 797 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 798 log(LOG_DEBUG, "%s; %s: Socket create failed " 799 "due to limits or memory shortage\n", 800 s, __func__); 801 free(s, M_TCPLOG); 802 } 803 goto abort2; 804 } 805 #ifdef MAC 806 mac_socketpeer_set_from_mbuf(m, so); 807 #endif 808 809 inp = sotoinpcb(so); 810 inp->inp_inc.inc_fibnum = so->so_fibnum; 811 INP_WLOCK(inp); 812 /* 813 * Exclusive pcbinfo lock is not required in syncache socket case even 814 * if two inpcb locks can be acquired simultaneously: 815 * - the inpcb in LISTEN state, 816 * - the newly created inp. 817 * 818 * In this case, an inp cannot be at same time in LISTEN state and 819 * just created by an accept() call. 820 */ 821 INP_HASH_WLOCK(&V_tcbinfo); 822 823 /* Insert new socket into PCB hash list. */ 824 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 825 #ifdef INET6 826 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 827 inp->inp_vflag &= ~INP_IPV4; 828 inp->inp_vflag |= INP_IPV6; 829 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 830 } else { 831 inp->inp_vflag &= ~INP_IPV6; 832 inp->inp_vflag |= INP_IPV4; 833 #endif 834 inp->inp_ip_ttl = sc->sc_ip_ttl; 835 inp->inp_ip_tos = sc->sc_ip_tos; 836 inp->inp_laddr = sc->sc_inc.inc_laddr; 837 #ifdef INET6 838 } 839 #endif 840 841 /* 842 * If there's an mbuf and it has a flowid, then let's initialise the 843 * inp with that particular flowid. 844 */ 845 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 846 inp->inp_flowid = m->m_pkthdr.flowid; 847 inp->inp_flowtype = M_HASHTYPE_GET(m); 848 #ifdef NUMA 849 inp->inp_numa_domain = m->m_pkthdr.numa_domain; 850 #endif 851 } 852 853 inp->inp_lport = sc->sc_inc.inc_lport; 854 #ifdef INET6 855 if (inp->inp_vflag & INP_IPV6PROTO) { 856 struct inpcb *oinp = sotoinpcb(lso); 857 858 /* 859 * Inherit socket options from the listening socket. 860 * Note that in6p_inputopts are not (and should not be) 861 * copied, since it stores previously received options and is 862 * used to detect if each new option is different than the 863 * previous one and hence should be passed to a user. 864 * If we copied in6p_inputopts, a user would not be able to 865 * receive options just after calling the accept system call. 866 */ 867 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 868 if (oinp->in6p_outputopts) 869 inp->in6p_outputopts = 870 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 871 inp->in6p_hops = oinp->in6p_hops; 872 } 873 874 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 875 struct in6_addr laddr6; 876 struct sockaddr_in6 sin6; 877 878 sin6.sin6_family = AF_INET6; 879 sin6.sin6_len = sizeof(sin6); 880 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 881 sin6.sin6_port = sc->sc_inc.inc_fport; 882 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 883 laddr6 = inp->in6p_laddr; 884 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 885 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 886 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 887 thread0.td_ucred, m, false)) != 0) { 888 inp->in6p_laddr = laddr6; 889 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 890 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 891 "with error %i\n", 892 s, __func__, error); 893 free(s, M_TCPLOG); 894 } 895 INP_HASH_WUNLOCK(&V_tcbinfo); 896 goto abort; 897 } 898 /* Override flowlabel from in6_pcbconnect. */ 899 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 900 inp->inp_flow |= sc->sc_flowlabel; 901 } 902 #endif /* INET6 */ 903 #if defined(INET) && defined(INET6) 904 else 905 #endif 906 #ifdef INET 907 { 908 struct in_addr laddr; 909 struct sockaddr_in sin; 910 911 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 912 913 if (inp->inp_options == NULL) { 914 inp->inp_options = sc->sc_ipopts; 915 sc->sc_ipopts = NULL; 916 } 917 918 sin.sin_family = AF_INET; 919 sin.sin_len = sizeof(sin); 920 sin.sin_addr = sc->sc_inc.inc_faddr; 921 sin.sin_port = sc->sc_inc.inc_fport; 922 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 923 laddr = inp->inp_laddr; 924 if (inp->inp_laddr.s_addr == INADDR_ANY) 925 inp->inp_laddr = sc->sc_inc.inc_laddr; 926 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 927 thread0.td_ucred, m, false)) != 0) { 928 inp->inp_laddr = laddr; 929 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 930 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 931 "with error %i\n", 932 s, __func__, error); 933 free(s, M_TCPLOG); 934 } 935 INP_HASH_WUNLOCK(&V_tcbinfo); 936 goto abort; 937 } 938 } 939 #endif /* INET */ 940 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 941 /* Copy old policy into new socket's. */ 942 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 943 printf("syncache_socket: could not copy policy\n"); 944 #endif 945 INP_HASH_WUNLOCK(&V_tcbinfo); 946 tp = intotcpcb(inp); 947 tcp_state_change(tp, TCPS_SYN_RECEIVED); 948 tp->iss = sc->sc_iss; 949 tp->irs = sc->sc_irs; 950 tcp_rcvseqinit(tp); 951 tcp_sendseqinit(tp); 952 blk = sototcpcb(lso)->t_fb; 953 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 954 /* 955 * Our parents t_fb was not the default, 956 * we need to release our ref on tp->t_fb and 957 * pickup one on the new entry. 958 */ 959 struct tcp_function_block *rblk; 960 961 rblk = find_and_ref_tcp_fb(blk); 962 KASSERT(rblk != NULL, 963 ("cannot find blk %p out of syncache?", blk)); 964 if (tp->t_fb->tfb_tcp_fb_fini) 965 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 966 refcount_release(&tp->t_fb->tfb_refcnt); 967 tp->t_fb = rblk; 968 /* 969 * XXXrrs this is quite dangerous, it is possible 970 * for the new function to fail to init. We also 971 * are not asking if the handoff_is_ok though at 972 * the very start thats probalbly ok. 973 */ 974 if (tp->t_fb->tfb_tcp_fb_init) { 975 (*tp->t_fb->tfb_tcp_fb_init)(tp); 976 } 977 } 978 tp->snd_wl1 = sc->sc_irs; 979 tp->snd_max = tp->iss + 1; 980 tp->snd_nxt = tp->iss + 1; 981 tp->rcv_up = sc->sc_irs + 1; 982 tp->rcv_wnd = sc->sc_wnd; 983 tp->rcv_adv += tp->rcv_wnd; 984 tp->last_ack_sent = tp->rcv_nxt; 985 986 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 987 if (sc->sc_flags & SCF_NOOPT) 988 tp->t_flags |= TF_NOOPT; 989 else { 990 if (sc->sc_flags & SCF_WINSCALE) { 991 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 992 tp->snd_scale = sc->sc_requested_s_scale; 993 tp->request_r_scale = sc->sc_requested_r_scale; 994 } 995 if (sc->sc_flags & SCF_TIMESTAMP) { 996 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 997 tp->ts_recent = sc->sc_tsreflect; 998 tp->ts_recent_age = tcp_ts_getticks(); 999 tp->ts_offset = sc->sc_tsoff; 1000 } 1001 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1002 if (sc->sc_flags & SCF_SIGNATURE) 1003 tp->t_flags |= TF_SIGNATURE; 1004 #endif 1005 if (sc->sc_flags & SCF_SACK) 1006 tp->t_flags |= TF_SACK_PERMIT; 1007 } 1008 1009 if (sc->sc_flags & SCF_ECN) 1010 tp->t_flags2 |= TF2_ECN_PERMIT; 1011 1012 /* 1013 * Set up MSS and get cached values from tcp_hostcache. 1014 * This might overwrite some of the defaults we just set. 1015 */ 1016 tcp_mss(tp, sc->sc_peer_mss); 1017 1018 /* 1019 * If the SYN,ACK was retransmitted, indicate that CWND to be 1020 * limited to one segment in cc_conn_init(). 1021 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 1022 */ 1023 if (sc->sc_rxmits > 1) 1024 tp->snd_cwnd = 1; 1025 1026 #ifdef TCP_OFFLOAD 1027 /* 1028 * Allow a TOE driver to install its hooks. Note that we hold the 1029 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 1030 * new connection before the TOE driver has done its thing. 1031 */ 1032 if (ADDED_BY_TOE(sc)) { 1033 struct toedev *tod = sc->sc_tod; 1034 1035 tod->tod_offload_socket(tod, sc->sc_todctx, so); 1036 } 1037 #endif 1038 /* 1039 * Copy and activate timers. 1040 */ 1041 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 1042 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 1043 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1044 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1045 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1046 1047 TCPSTAT_INC(tcps_accepts); 1048 return (so); 1049 1050 abort: 1051 INP_WUNLOCK(inp); 1052 abort2: 1053 if (so != NULL) 1054 soabort(so); 1055 return (NULL); 1056 } 1057 1058 /* 1059 * This function gets called when we receive an ACK for a 1060 * socket in the LISTEN state. We look up the connection 1061 * in the syncache, and if its there, we pull it out of 1062 * the cache and turn it into a full-blown connection in 1063 * the SYN-RECEIVED state. 1064 * 1065 * On syncache_socket() success the newly created socket 1066 * has its underlying inp locked. 1067 */ 1068 int 1069 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1070 struct socket **lsop, struct mbuf *m) 1071 { 1072 struct syncache *sc; 1073 struct syncache_head *sch; 1074 struct syncache scs; 1075 char *s; 1076 bool locked; 1077 1078 NET_EPOCH_ASSERT(); 1079 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1080 ("%s: can handle only ACK", __func__)); 1081 1082 if (syncache_cookiesonly()) { 1083 sc = NULL; 1084 sch = syncache_hashbucket(inc); 1085 locked = false; 1086 } else { 1087 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1088 locked = true; 1089 SCH_LOCK_ASSERT(sch); 1090 } 1091 1092 #ifdef INVARIANTS 1093 /* 1094 * Test code for syncookies comparing the syncache stored 1095 * values with the reconstructed values from the cookie. 1096 */ 1097 if (sc != NULL) 1098 syncookie_cmp(inc, sch, sc, th, to, *lsop); 1099 #endif 1100 1101 if (sc == NULL) { 1102 /* 1103 * There is no syncache entry, so see if this ACK is 1104 * a returning syncookie. To do this, first: 1105 * A. Check if syncookies are used in case of syncache 1106 * overflows 1107 * B. See if this socket has had a syncache entry dropped in 1108 * the recent past. We don't want to accept a bogus 1109 * syncookie if we've never received a SYN or accept it 1110 * twice. 1111 * C. check that the syncookie is valid. If it is, then 1112 * cobble up a fake syncache entry, and return. 1113 */ 1114 if (locked && !V_tcp_syncookies) { 1115 SCH_UNLOCK(sch); 1116 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1117 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1118 "segment rejected (syncookies disabled)\n", 1119 s, __func__); 1120 goto failed; 1121 } 1122 if (locked && !V_tcp_syncookiesonly && 1123 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1124 SCH_UNLOCK(sch); 1125 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1126 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1127 "segment rejected (no syncache entry)\n", 1128 s, __func__); 1129 goto failed; 1130 } 1131 bzero(&scs, sizeof(scs)); 1132 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1133 if (locked) 1134 SCH_UNLOCK(sch); 1135 if (sc == NULL) { 1136 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1137 log(LOG_DEBUG, "%s; %s: Segment failed " 1138 "SYNCOOKIE authentication, segment rejected " 1139 "(probably spoofed)\n", s, __func__); 1140 goto failed; 1141 } 1142 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1143 /* If received ACK has MD5 signature, check it. */ 1144 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1145 (!TCPMD5_ENABLED() || 1146 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1147 /* Drop the ACK. */ 1148 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1149 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1150 "MD5 signature doesn't match.\n", 1151 s, __func__); 1152 free(s, M_TCPLOG); 1153 } 1154 TCPSTAT_INC(tcps_sig_err_sigopt); 1155 return (-1); /* Do not send RST */ 1156 } 1157 #endif /* TCP_SIGNATURE */ 1158 } else { 1159 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1160 /* 1161 * If listening socket requested TCP digests, check that 1162 * received ACK has signature and it is correct. 1163 * If not, drop the ACK and leave sc entry in th cache, 1164 * because SYN was received with correct signature. 1165 */ 1166 if (sc->sc_flags & SCF_SIGNATURE) { 1167 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1168 /* No signature */ 1169 TCPSTAT_INC(tcps_sig_err_nosigopt); 1170 SCH_UNLOCK(sch); 1171 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1172 log(LOG_DEBUG, "%s; %s: Segment " 1173 "rejected, MD5 signature wasn't " 1174 "provided.\n", s, __func__); 1175 free(s, M_TCPLOG); 1176 } 1177 return (-1); /* Do not send RST */ 1178 } 1179 if (!TCPMD5_ENABLED() || 1180 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1181 /* Doesn't match or no SA */ 1182 SCH_UNLOCK(sch); 1183 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1184 log(LOG_DEBUG, "%s; %s: Segment " 1185 "rejected, MD5 signature doesn't " 1186 "match.\n", s, __func__); 1187 free(s, M_TCPLOG); 1188 } 1189 return (-1); /* Do not send RST */ 1190 } 1191 } 1192 #endif /* TCP_SIGNATURE */ 1193 1194 /* 1195 * RFC 7323 PAWS: If we have a timestamp on this segment and 1196 * it's less than ts_recent, drop it. 1197 * XXXMT: RFC 7323 also requires to send an ACK. 1198 * In tcp_input.c this is only done for TCP segments 1199 * with user data, so be consistent here and just drop 1200 * the segment. 1201 */ 1202 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1203 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1204 SCH_UNLOCK(sch); 1205 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1206 log(LOG_DEBUG, 1207 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1208 "segment dropped\n", s, __func__, 1209 to->to_tsval, sc->sc_tsreflect); 1210 free(s, M_TCPLOG); 1211 } 1212 return (-1); /* Do not send RST */ 1213 } 1214 1215 /* 1216 * Pull out the entry to unlock the bucket row. 1217 * 1218 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1219 * tcp_state_change(). The tcpcb is not existent at this 1220 * moment. A new one will be allocated via syncache_socket-> 1221 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1222 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1223 */ 1224 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1225 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1226 sch->sch_length--; 1227 #ifdef TCP_OFFLOAD 1228 if (ADDED_BY_TOE(sc)) { 1229 struct toedev *tod = sc->sc_tod; 1230 1231 tod->tod_syncache_removed(tod, sc->sc_todctx); 1232 } 1233 #endif 1234 SCH_UNLOCK(sch); 1235 } 1236 1237 /* 1238 * Segment validation: 1239 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1240 */ 1241 if (th->th_ack != sc->sc_iss + 1) { 1242 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1243 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1244 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1245 goto failed; 1246 } 1247 1248 /* 1249 * The SEQ must fall in the window starting at the received 1250 * initial receive sequence number + 1 (the SYN). 1251 */ 1252 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1253 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1254 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1255 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1256 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1257 goto failed; 1258 } 1259 1260 /* 1261 * If timestamps were not negotiated during SYN/ACK they 1262 * must not appear on any segment during this session. 1263 */ 1264 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1265 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1266 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1267 "segment rejected\n", s, __func__); 1268 goto failed; 1269 } 1270 1271 /* 1272 * If timestamps were negotiated during SYN/ACK they should 1273 * appear on every segment during this session. 1274 * XXXAO: This is only informal as there have been unverified 1275 * reports of non-compliants stacks. 1276 */ 1277 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1278 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1279 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1280 "no action\n", s, __func__); 1281 free(s, M_TCPLOG); 1282 s = NULL; 1283 } 1284 } 1285 1286 *lsop = syncache_socket(sc, *lsop, m); 1287 1288 if (*lsop == NULL) 1289 TCPSTAT_INC(tcps_sc_aborted); 1290 else 1291 TCPSTAT_INC(tcps_sc_completed); 1292 1293 /* how do we find the inp for the new socket? */ 1294 if (sc != &scs) 1295 syncache_free(sc); 1296 return (1); 1297 failed: 1298 if (sc != NULL && sc != &scs) 1299 syncache_free(sc); 1300 if (s != NULL) 1301 free(s, M_TCPLOG); 1302 *lsop = NULL; 1303 return (0); 1304 } 1305 1306 static void 1307 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1308 uint64_t response_cookie) 1309 { 1310 struct inpcb *inp; 1311 struct tcpcb *tp; 1312 unsigned int *pending_counter; 1313 1314 NET_EPOCH_ASSERT(); 1315 1316 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1317 *lsop = syncache_socket(sc, *lsop, m); 1318 if (*lsop == NULL) { 1319 TCPSTAT_INC(tcps_sc_aborted); 1320 atomic_subtract_int(pending_counter, 1); 1321 } else { 1322 soisconnected(*lsop); 1323 inp = sotoinpcb(*lsop); 1324 tp = intotcpcb(inp); 1325 tp->t_flags |= TF_FASTOPEN; 1326 tp->t_tfo_cookie.server = response_cookie; 1327 tp->snd_max = tp->iss; 1328 tp->snd_nxt = tp->iss; 1329 tp->t_tfo_pending = pending_counter; 1330 TCPSTAT_INC(tcps_sc_completed); 1331 } 1332 } 1333 1334 /* 1335 * Given a LISTEN socket and an inbound SYN request, add 1336 * this to the syn cache, and send back a segment: 1337 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1338 * to the source. 1339 * 1340 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1341 * Doing so would require that we hold onto the data and deliver it 1342 * to the application. However, if we are the target of a SYN-flood 1343 * DoS attack, an attacker could send data which would eventually 1344 * consume all available buffer space if it were ACKed. By not ACKing 1345 * the data, we avoid this DoS scenario. 1346 * 1347 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1348 * cookie is processed and a new socket is created. In this case, any data 1349 * accompanying the SYN will be queued to the socket by tcp_input() and will 1350 * be ACKed either when the application sends response data or the delayed 1351 * ACK timer expires, whichever comes first. 1352 */ 1353 int 1354 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1355 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1356 void *todctx, uint8_t iptos) 1357 { 1358 struct tcpcb *tp; 1359 struct socket *so; 1360 struct syncache *sc = NULL; 1361 struct syncache_head *sch; 1362 struct mbuf *ipopts = NULL; 1363 u_int ltflags; 1364 int win, ip_ttl, ip_tos; 1365 char *s; 1366 int rv = 0; 1367 #ifdef INET6 1368 int autoflowlabel = 0; 1369 #endif 1370 #ifdef MAC 1371 struct label *maclabel; 1372 #endif 1373 struct syncache scs; 1374 struct ucred *cred; 1375 uint64_t tfo_response_cookie; 1376 unsigned int *tfo_pending = NULL; 1377 int tfo_cookie_valid = 0; 1378 int tfo_response_cookie_valid = 0; 1379 bool locked; 1380 1381 INP_WLOCK_ASSERT(inp); /* listen socket */ 1382 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1383 ("%s: unexpected tcp flags", __func__)); 1384 1385 /* 1386 * Combine all so/tp operations very early to drop the INP lock as 1387 * soon as possible. 1388 */ 1389 so = *lsop; 1390 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1391 tp = sototcpcb(so); 1392 cred = crhold(so->so_cred); 1393 1394 #ifdef INET6 1395 if (inc->inc_flags & INC_ISIPV6) { 1396 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1397 autoflowlabel = 1; 1398 } 1399 ip_ttl = in6_selecthlim(inp, NULL); 1400 if ((inp->in6p_outputopts == NULL) || 1401 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1402 ip_tos = 0; 1403 } else { 1404 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1405 } 1406 } 1407 #endif 1408 #if defined(INET6) && defined(INET) 1409 else 1410 #endif 1411 #ifdef INET 1412 { 1413 ip_ttl = inp->inp_ip_ttl; 1414 ip_tos = inp->inp_ip_tos; 1415 } 1416 #endif 1417 win = so->sol_sbrcv_hiwat; 1418 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1419 1420 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1421 (tp->t_tfo_pending != NULL) && 1422 (to->to_flags & TOF_FASTOPEN)) { 1423 /* 1424 * Limit the number of pending TFO connections to 1425 * approximately half of the queue limit. This prevents TFO 1426 * SYN floods from starving the service by filling the 1427 * listen queue with bogus TFO connections. 1428 */ 1429 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1430 (so->sol_qlimit / 2)) { 1431 int result; 1432 1433 result = tcp_fastopen_check_cookie(inc, 1434 to->to_tfo_cookie, to->to_tfo_len, 1435 &tfo_response_cookie); 1436 tfo_cookie_valid = (result > 0); 1437 tfo_response_cookie_valid = (result >= 0); 1438 } 1439 1440 /* 1441 * Remember the TFO pending counter as it will have to be 1442 * decremented below if we don't make it to syncache_tfo_expand(). 1443 */ 1444 tfo_pending = tp->t_tfo_pending; 1445 } 1446 1447 /* By the time we drop the lock these should no longer be used. */ 1448 so = NULL; 1449 tp = NULL; 1450 1451 #ifdef MAC 1452 if (mac_syncache_init(&maclabel) != 0) { 1453 INP_WUNLOCK(inp); 1454 goto done; 1455 } else 1456 mac_syncache_create(maclabel, inp); 1457 #endif 1458 if (!tfo_cookie_valid) 1459 INP_WUNLOCK(inp); 1460 1461 /* 1462 * Remember the IP options, if any. 1463 */ 1464 #ifdef INET6 1465 if (!(inc->inc_flags & INC_ISIPV6)) 1466 #endif 1467 #ifdef INET 1468 ipopts = (m) ? ip_srcroute(m) : NULL; 1469 #else 1470 ipopts = NULL; 1471 #endif 1472 1473 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1474 /* 1475 * If listening socket requested TCP digests, check that received 1476 * SYN has signature and it is correct. If signature doesn't match 1477 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1478 */ 1479 if (ltflags & TF_SIGNATURE) { 1480 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1481 TCPSTAT_INC(tcps_sig_err_nosigopt); 1482 goto done; 1483 } 1484 if (!TCPMD5_ENABLED() || 1485 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1486 goto done; 1487 } 1488 #endif /* TCP_SIGNATURE */ 1489 /* 1490 * See if we already have an entry for this connection. 1491 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1492 * 1493 * XXX: should the syncache be re-initialized with the contents 1494 * of the new SYN here (which may have different options?) 1495 * 1496 * XXX: We do not check the sequence number to see if this is a 1497 * real retransmit or a new connection attempt. The question is 1498 * how to handle such a case; either ignore it as spoofed, or 1499 * drop the current entry and create a new one? 1500 */ 1501 if (syncache_cookiesonly()) { 1502 sc = NULL; 1503 sch = syncache_hashbucket(inc); 1504 locked = false; 1505 } else { 1506 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1507 locked = true; 1508 SCH_LOCK_ASSERT(sch); 1509 } 1510 if (sc != NULL) { 1511 if (tfo_cookie_valid) 1512 INP_WUNLOCK(inp); 1513 TCPSTAT_INC(tcps_sc_dupsyn); 1514 if (ipopts) { 1515 /* 1516 * If we were remembering a previous source route, 1517 * forget it and use the new one we've been given. 1518 */ 1519 if (sc->sc_ipopts) 1520 (void) m_free(sc->sc_ipopts); 1521 sc->sc_ipopts = ipopts; 1522 } 1523 /* 1524 * Update timestamp if present. 1525 */ 1526 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1527 sc->sc_tsreflect = to->to_tsval; 1528 else 1529 sc->sc_flags &= ~SCF_TIMESTAMP; 1530 /* 1531 * Disable ECN if needed. 1532 */ 1533 if ((sc->sc_flags & SCF_ECN) && 1534 ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) { 1535 sc->sc_flags &= ~SCF_ECN; 1536 } 1537 #ifdef MAC 1538 /* 1539 * Since we have already unconditionally allocated label 1540 * storage, free it up. The syncache entry will already 1541 * have an initialized label we can use. 1542 */ 1543 mac_syncache_destroy(&maclabel); 1544 #endif 1545 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1546 /* Retransmit SYN|ACK and reset retransmit count. */ 1547 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1548 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1549 "resetting timer and retransmitting SYN|ACK\n", 1550 s, __func__); 1551 free(s, M_TCPLOG); 1552 } 1553 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1554 sc->sc_rxmits = 0; 1555 syncache_timeout(sc, sch, 1); 1556 TCPSTAT_INC(tcps_sndacks); 1557 TCPSTAT_INC(tcps_sndtotal); 1558 } 1559 SCH_UNLOCK(sch); 1560 goto donenoprobe; 1561 } 1562 1563 if (tfo_cookie_valid) { 1564 bzero(&scs, sizeof(scs)); 1565 sc = &scs; 1566 goto skip_alloc; 1567 } 1568 1569 /* 1570 * Skip allocating a syncache entry if we are just going to discard 1571 * it later. 1572 */ 1573 if (!locked) { 1574 bzero(&scs, sizeof(scs)); 1575 sc = &scs; 1576 } else 1577 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1578 if (sc == NULL) { 1579 /* 1580 * The zone allocator couldn't provide more entries. 1581 * Treat this as if the cache was full; drop the oldest 1582 * entry and insert the new one. 1583 */ 1584 TCPSTAT_INC(tcps_sc_zonefail); 1585 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1586 sch->sch_last_overflow = time_uptime; 1587 syncache_drop(sc, sch); 1588 syncache_pause(inc); 1589 } 1590 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1591 if (sc == NULL) { 1592 if (V_tcp_syncookies) { 1593 bzero(&scs, sizeof(scs)); 1594 sc = &scs; 1595 } else { 1596 KASSERT(locked, 1597 ("%s: bucket unexpectedly unlocked", 1598 __func__)); 1599 SCH_UNLOCK(sch); 1600 if (ipopts) 1601 (void) m_free(ipopts); 1602 goto done; 1603 } 1604 } 1605 } 1606 1607 skip_alloc: 1608 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1609 sc->sc_tfo_cookie = &tfo_response_cookie; 1610 1611 /* 1612 * Fill in the syncache values. 1613 */ 1614 #ifdef MAC 1615 sc->sc_label = maclabel; 1616 #endif 1617 sc->sc_cred = cred; 1618 cred = NULL; 1619 sc->sc_ipopts = ipopts; 1620 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1621 sc->sc_ip_tos = ip_tos; 1622 sc->sc_ip_ttl = ip_ttl; 1623 #ifdef TCP_OFFLOAD 1624 sc->sc_tod = tod; 1625 sc->sc_todctx = todctx; 1626 #endif 1627 sc->sc_irs = th->th_seq; 1628 sc->sc_flags = 0; 1629 sc->sc_flowlabel = 0; 1630 1631 /* 1632 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1633 * win was derived from socket earlier in the function. 1634 */ 1635 win = imax(win, 0); 1636 win = imin(win, TCP_MAXWIN); 1637 sc->sc_wnd = win; 1638 1639 if (V_tcp_do_rfc1323) { 1640 /* 1641 * A timestamp received in a SYN makes 1642 * it ok to send timestamp requests and replies. 1643 */ 1644 if (to->to_flags & TOF_TS) { 1645 sc->sc_tsreflect = to->to_tsval; 1646 sc->sc_flags |= SCF_TIMESTAMP; 1647 sc->sc_tsoff = tcp_new_ts_offset(inc); 1648 } 1649 if (to->to_flags & TOF_SCALE) { 1650 int wscale = 0; 1651 1652 /* 1653 * Pick the smallest possible scaling factor that 1654 * will still allow us to scale up to sb_max, aka 1655 * kern.ipc.maxsockbuf. 1656 * 1657 * We do this because there are broken firewalls that 1658 * will corrupt the window scale option, leading to 1659 * the other endpoint believing that our advertised 1660 * window is unscaled. At scale factors larger than 1661 * 5 the unscaled window will drop below 1500 bytes, 1662 * leading to serious problems when traversing these 1663 * broken firewalls. 1664 * 1665 * With the default maxsockbuf of 256K, a scale factor 1666 * of 3 will be chosen by this algorithm. Those who 1667 * choose a larger maxsockbuf should watch out 1668 * for the compatibility problems mentioned above. 1669 * 1670 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1671 * or <SYN,ACK>) segment itself is never scaled. 1672 */ 1673 while (wscale < TCP_MAX_WINSHIFT && 1674 (TCP_MAXWIN << wscale) < sb_max) 1675 wscale++; 1676 sc->sc_requested_r_scale = wscale; 1677 sc->sc_requested_s_scale = to->to_wscale; 1678 sc->sc_flags |= SCF_WINSCALE; 1679 } 1680 } 1681 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1682 /* 1683 * If listening socket requested TCP digests, flag this in the 1684 * syncache so that syncache_respond() will do the right thing 1685 * with the SYN+ACK. 1686 */ 1687 if (ltflags & TF_SIGNATURE) 1688 sc->sc_flags |= SCF_SIGNATURE; 1689 #endif /* TCP_SIGNATURE */ 1690 if (to->to_flags & TOF_SACKPERM) 1691 sc->sc_flags |= SCF_SACK; 1692 if (to->to_flags & TOF_MSS) 1693 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1694 if (ltflags & TF_NOOPT) 1695 sc->sc_flags |= SCF_NOOPT; 1696 if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) && 1697 V_tcp_do_ecn) 1698 sc->sc_flags |= SCF_ECN; 1699 1700 if (V_tcp_syncookies) 1701 sc->sc_iss = syncookie_generate(sch, sc); 1702 else 1703 sc->sc_iss = arc4random(); 1704 #ifdef INET6 1705 if (autoflowlabel) { 1706 if (V_tcp_syncookies) 1707 sc->sc_flowlabel = sc->sc_iss; 1708 else 1709 sc->sc_flowlabel = ip6_randomflowlabel(); 1710 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1711 } 1712 #endif 1713 if (locked) 1714 SCH_UNLOCK(sch); 1715 1716 if (tfo_cookie_valid) { 1717 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1718 /* INP_WUNLOCK(inp) will be performed by the caller */ 1719 rv = 1; 1720 goto tfo_expanded; 1721 } 1722 1723 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1724 /* 1725 * Do a standard 3-way handshake. 1726 */ 1727 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1728 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1729 syncache_free(sc); 1730 else if (sc != &scs) 1731 syncache_insert(sc, sch); /* locks and unlocks sch */ 1732 TCPSTAT_INC(tcps_sndacks); 1733 TCPSTAT_INC(tcps_sndtotal); 1734 } else { 1735 if (sc != &scs) 1736 syncache_free(sc); 1737 TCPSTAT_INC(tcps_sc_dropped); 1738 } 1739 goto donenoprobe; 1740 1741 done: 1742 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1743 donenoprobe: 1744 if (m) { 1745 *lsop = NULL; 1746 m_freem(m); 1747 } 1748 /* 1749 * If tfo_pending is not NULL here, then a TFO SYN that did not 1750 * result in a new socket was processed and the associated pending 1751 * counter has not yet been decremented. All such TFO processing paths 1752 * transit this point. 1753 */ 1754 if (tfo_pending != NULL) 1755 tcp_fastopen_decrement_counter(tfo_pending); 1756 1757 tfo_expanded: 1758 if (cred != NULL) 1759 crfree(cred); 1760 #ifdef MAC 1761 if (sc == &scs) 1762 mac_syncache_destroy(&maclabel); 1763 #endif 1764 return (rv); 1765 } 1766 1767 /* 1768 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1769 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1770 */ 1771 static int 1772 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags) 1773 { 1774 struct ip *ip = NULL; 1775 struct mbuf *m; 1776 struct tcphdr *th = NULL; 1777 int optlen, error = 0; /* Make compiler happy */ 1778 u_int16_t hlen, tlen, mssopt; 1779 struct tcpopt to; 1780 #ifdef INET6 1781 struct ip6_hdr *ip6 = NULL; 1782 #endif 1783 1784 NET_EPOCH_ASSERT(); 1785 1786 hlen = 1787 #ifdef INET6 1788 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1789 #endif 1790 sizeof(struct ip); 1791 tlen = hlen + sizeof(struct tcphdr); 1792 1793 /* Determine MSS we advertize to other end of connection. */ 1794 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); 1795 1796 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1797 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1798 ("syncache: mbuf too small")); 1799 1800 /* Create the IP+TCP header from scratch. */ 1801 m = m_gethdr(M_NOWAIT, MT_DATA); 1802 if (m == NULL) 1803 return (ENOBUFS); 1804 #ifdef MAC 1805 mac_syncache_create_mbuf(sc->sc_label, m); 1806 #endif 1807 m->m_data += max_linkhdr; 1808 m->m_len = tlen; 1809 m->m_pkthdr.len = tlen; 1810 m->m_pkthdr.rcvif = NULL; 1811 1812 #ifdef INET6 1813 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1814 ip6 = mtod(m, struct ip6_hdr *); 1815 ip6->ip6_vfc = IPV6_VERSION; 1816 ip6->ip6_nxt = IPPROTO_TCP; 1817 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1818 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1819 ip6->ip6_plen = htons(tlen - hlen); 1820 /* ip6_hlim is set after checksum */ 1821 /* Zero out traffic class and flow label. */ 1822 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1823 ip6->ip6_flow |= sc->sc_flowlabel; 1824 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20); 1825 1826 th = (struct tcphdr *)(ip6 + 1); 1827 } 1828 #endif 1829 #if defined(INET6) && defined(INET) 1830 else 1831 #endif 1832 #ifdef INET 1833 { 1834 ip = mtod(m, struct ip *); 1835 ip->ip_v = IPVERSION; 1836 ip->ip_hl = sizeof(struct ip) >> 2; 1837 ip->ip_len = htons(tlen); 1838 ip->ip_id = 0; 1839 ip->ip_off = 0; 1840 ip->ip_sum = 0; 1841 ip->ip_p = IPPROTO_TCP; 1842 ip->ip_src = sc->sc_inc.inc_laddr; 1843 ip->ip_dst = sc->sc_inc.inc_faddr; 1844 ip->ip_ttl = sc->sc_ip_ttl; 1845 ip->ip_tos = sc->sc_ip_tos; 1846 1847 /* 1848 * See if we should do MTU discovery. Route lookups are 1849 * expensive, so we will only unset the DF bit if: 1850 * 1851 * 1) path_mtu_discovery is disabled 1852 * 2) the SCF_UNREACH flag has been set 1853 */ 1854 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1855 ip->ip_off |= htons(IP_DF); 1856 1857 th = (struct tcphdr *)(ip + 1); 1858 } 1859 #endif /* INET */ 1860 th->th_sport = sc->sc_inc.inc_lport; 1861 th->th_dport = sc->sc_inc.inc_fport; 1862 1863 if (flags & TH_SYN) 1864 th->th_seq = htonl(sc->sc_iss); 1865 else 1866 th->th_seq = htonl(sc->sc_iss + 1); 1867 th->th_ack = htonl(sc->sc_irs + 1); 1868 th->th_off = sizeof(struct tcphdr) >> 2; 1869 th->th_x2 = 0; 1870 th->th_flags = flags; 1871 th->th_win = htons(sc->sc_wnd); 1872 th->th_urp = 0; 1873 1874 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) { 1875 th->th_flags |= TH_ECE; 1876 TCPSTAT_INC(tcps_ecn_shs); 1877 } 1878 1879 /* Tack on the TCP options. */ 1880 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1881 to.to_flags = 0; 1882 1883 if (flags & TH_SYN) { 1884 to.to_mss = mssopt; 1885 to.to_flags = TOF_MSS; 1886 if (sc->sc_flags & SCF_WINSCALE) { 1887 to.to_wscale = sc->sc_requested_r_scale; 1888 to.to_flags |= TOF_SCALE; 1889 } 1890 if (sc->sc_flags & SCF_SACK) 1891 to.to_flags |= TOF_SACKPERM; 1892 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1893 if (sc->sc_flags & SCF_SIGNATURE) 1894 to.to_flags |= TOF_SIGNATURE; 1895 #endif 1896 if (sc->sc_tfo_cookie) { 1897 to.to_flags |= TOF_FASTOPEN; 1898 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1899 to.to_tfo_cookie = sc->sc_tfo_cookie; 1900 /* don't send cookie again when retransmitting response */ 1901 sc->sc_tfo_cookie = NULL; 1902 } 1903 } 1904 if (sc->sc_flags & SCF_TIMESTAMP) { 1905 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1906 to.to_tsecr = sc->sc_tsreflect; 1907 to.to_flags |= TOF_TS; 1908 } 1909 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1910 1911 /* Adjust headers by option size. */ 1912 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1913 m->m_len += optlen; 1914 m->m_pkthdr.len += optlen; 1915 #ifdef INET6 1916 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1917 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1918 else 1919 #endif 1920 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1921 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1922 if (sc->sc_flags & SCF_SIGNATURE) { 1923 KASSERT(to.to_flags & TOF_SIGNATURE, 1924 ("tcp_addoptions() didn't set tcp_signature")); 1925 1926 /* NOTE: to.to_signature is inside of mbuf */ 1927 if (!TCPMD5_ENABLED() || 1928 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1929 m_freem(m); 1930 return (EACCES); 1931 } 1932 } 1933 #endif 1934 } else 1935 optlen = 0; 1936 1937 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1938 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1939 /* 1940 * If we have peer's SYN and it has a flowid, then let's assign it to 1941 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1942 * to SYN|ACK due to lack of inp here. 1943 */ 1944 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1945 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1946 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1947 } 1948 #ifdef INET6 1949 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1950 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1951 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1952 IPPROTO_TCP, 0); 1953 ip6->ip6_hlim = sc->sc_ip_ttl; 1954 #ifdef TCP_OFFLOAD 1955 if (ADDED_BY_TOE(sc)) { 1956 struct toedev *tod = sc->sc_tod; 1957 1958 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1959 1960 return (error); 1961 } 1962 #endif 1963 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 1964 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1965 } 1966 #endif 1967 #if defined(INET6) && defined(INET) 1968 else 1969 #endif 1970 #ifdef INET 1971 { 1972 m->m_pkthdr.csum_flags = CSUM_TCP; 1973 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1974 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1975 #ifdef TCP_OFFLOAD 1976 if (ADDED_BY_TOE(sc)) { 1977 struct toedev *tod = sc->sc_tod; 1978 1979 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1980 1981 return (error); 1982 } 1983 #endif 1984 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 1985 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1986 } 1987 #endif 1988 return (error); 1989 } 1990 1991 /* 1992 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1993 * that exceed the capacity of the syncache by avoiding the storage of any 1994 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1995 * attacks where the attacker does not have access to our responses. 1996 * 1997 * Syncookies encode and include all necessary information about the 1998 * connection setup within the SYN|ACK that we send back. That way we 1999 * can avoid keeping any local state until the ACK to our SYN|ACK returns 2000 * (if ever). Normally the syncache and syncookies are running in parallel 2001 * with the latter taking over when the former is exhausted. When matching 2002 * syncache entry is found the syncookie is ignored. 2003 * 2004 * The only reliable information persisting the 3WHS is our initial sequence 2005 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 2006 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 2007 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 2008 * returns and signifies a legitimate connection if it matches the ACK. 2009 * 2010 * The available space of 32 bits to store the hash and to encode the SYN 2011 * option information is very tight and we should have at least 24 bits for 2012 * the MAC to keep the number of guesses by blind spoofing reasonably high. 2013 * 2014 * SYN option information we have to encode to fully restore a connection: 2015 * MSS: is imporant to chose an optimal segment size to avoid IP level 2016 * fragmentation along the path. The common MSS values can be encoded 2017 * in a 3-bit table. Uncommon values are captured by the next lower value 2018 * in the table leading to a slight increase in packetization overhead. 2019 * WSCALE: is necessary to allow large windows to be used for high delay- 2020 * bandwidth product links. Not scaling the window when it was initially 2021 * negotiated is bad for performance as lack of scaling further decreases 2022 * the apparent available send window. We only need to encode the WSCALE 2023 * we received from the remote end. Our end can be recalculated at any 2024 * time. The common WSCALE values can be encoded in a 3-bit table. 2025 * Uncommon values are captured by the next lower value in the table 2026 * making us under-estimate the available window size halving our 2027 * theoretically possible maximum throughput for that connection. 2028 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 2029 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 2030 * that are included in all segments on a connection. We enable them when 2031 * the ACK has them. 2032 * 2033 * Security of syncookies and attack vectors: 2034 * 2035 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 2036 * together with the gloabl secret to make it unique per connection attempt. 2037 * Thus any change of any of those parameters results in a different MAC output 2038 * in an unpredictable way unless a collision is encountered. 24 bits of the 2039 * MAC are embedded into the ISS. 2040 * 2041 * To prevent replay attacks two rotating global secrets are updated with a 2042 * new random value every 15 seconds. The life-time of a syncookie is thus 2043 * 15-30 seconds. 2044 * 2045 * Vector 1: Attacking the secret. This requires finding a weakness in the 2046 * MAC itself or the way it is used here. The attacker can do a chosen plain 2047 * text attack by varying and testing the all parameters under his control. 2048 * The strength depends on the size and randomness of the secret, and the 2049 * cryptographic security of the MAC function. Due to the constant updating 2050 * of the secret the attacker has at most 29.999 seconds to find the secret 2051 * and launch spoofed connections. After that he has to start all over again. 2052 * 2053 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 2054 * size an average of 4,823 attempts are required for a 50% chance of success 2055 * to spoof a single syncookie (birthday collision paradox). However the 2056 * attacker is blind and doesn't know if one of his attempts succeeded unless 2057 * he has a side channel to interfere success from. A single connection setup 2058 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 2059 * This many attempts are required for each one blind spoofed connection. For 2060 * every additional spoofed connection he has to launch another N attempts. 2061 * Thus for a sustained rate 100 spoofed connections per second approximately 2062 * 1,800,000 packets per second would have to be sent. 2063 * 2064 * NB: The MAC function should be fast so that it doesn't become a CPU 2065 * exhaustion attack vector itself. 2066 * 2067 * References: 2068 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 2069 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 2070 * http://cr.yp.to/syncookies.html (overview) 2071 * http://cr.yp.to/syncookies/archive (details) 2072 * 2073 * 2074 * Schematic construction of a syncookie enabled Initial Sequence Number: 2075 * 0 1 2 3 2076 * 12345678901234567890123456789012 2077 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2078 * 2079 * x 24 MAC (truncated) 2080 * W 3 Send Window Scale index 2081 * M 3 MSS index 2082 * S 1 SACK permitted 2083 * P 1 Odd/even secret 2084 */ 2085 2086 /* 2087 * Distribution and probability of certain MSS values. Those in between are 2088 * rounded down to the next lower one. 2089 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2090 * .2% .3% 5% 7% 7% 20% 15% 45% 2091 */ 2092 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2093 2094 /* 2095 * Distribution and probability of certain WSCALE values. We have to map the 2096 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2097 * bits based on prevalence of certain values. Where we don't have an exact 2098 * match for are rounded down to the next lower one letting us under-estimate 2099 * the true available window. At the moment this would happen only for the 2100 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2101 * and window size). The absence of the WSCALE option (no scaling in either 2102 * direction) is encoded with index zero. 2103 * [WSCALE values histograms, Allman, 2012] 2104 * X 10 10 35 5 6 14 10% by host 2105 * X 11 4 5 5 18 49 3% by connections 2106 */ 2107 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2108 2109 /* 2110 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2111 * and good cryptographic properties. 2112 */ 2113 static uint32_t 2114 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2115 uint8_t *secbits, uintptr_t secmod) 2116 { 2117 SIPHASH_CTX ctx; 2118 uint32_t siphash[2]; 2119 2120 SipHash24_Init(&ctx); 2121 SipHash_SetKey(&ctx, secbits); 2122 switch (inc->inc_flags & INC_ISIPV6) { 2123 #ifdef INET 2124 case 0: 2125 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2126 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2127 break; 2128 #endif 2129 #ifdef INET6 2130 case INC_ISIPV6: 2131 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2132 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2133 break; 2134 #endif 2135 } 2136 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2137 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2138 SipHash_Update(&ctx, &irs, sizeof(irs)); 2139 SipHash_Update(&ctx, &flags, sizeof(flags)); 2140 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2141 SipHash_Final((u_int8_t *)&siphash, &ctx); 2142 2143 return (siphash[0] ^ siphash[1]); 2144 } 2145 2146 static tcp_seq 2147 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2148 { 2149 u_int i, secbit, wscale; 2150 uint32_t iss, hash; 2151 uint8_t *secbits; 2152 union syncookie cookie; 2153 2154 cookie.cookie = 0; 2155 2156 /* Map our computed MSS into the 3-bit index. */ 2157 for (i = nitems(tcp_sc_msstab) - 1; 2158 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2159 i--) 2160 ; 2161 cookie.flags.mss_idx = i; 2162 2163 /* 2164 * Map the send window scale into the 3-bit index but only if 2165 * the wscale option was received. 2166 */ 2167 if (sc->sc_flags & SCF_WINSCALE) { 2168 wscale = sc->sc_requested_s_scale; 2169 for (i = nitems(tcp_sc_wstab) - 1; 2170 tcp_sc_wstab[i] > wscale && i > 0; 2171 i--) 2172 ; 2173 cookie.flags.wscale_idx = i; 2174 } 2175 2176 /* Can we do SACK? */ 2177 if (sc->sc_flags & SCF_SACK) 2178 cookie.flags.sack_ok = 1; 2179 2180 /* Which of the two secrets to use. */ 2181 secbit = V_tcp_syncache.secret.oddeven & 0x1; 2182 cookie.flags.odd_even = secbit; 2183 2184 secbits = V_tcp_syncache.secret.key[secbit]; 2185 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2186 (uintptr_t)sch); 2187 2188 /* 2189 * Put the flags into the hash and XOR them to get better ISS number 2190 * variance. This doesn't enhance the cryptographic strength and is 2191 * done to prevent the 8 cookie bits from showing up directly on the 2192 * wire. 2193 */ 2194 iss = hash & ~0xff; 2195 iss |= cookie.cookie ^ (hash >> 24); 2196 2197 TCPSTAT_INC(tcps_sc_sendcookie); 2198 return (iss); 2199 } 2200 2201 static struct syncache * 2202 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2203 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2204 struct socket *lso) 2205 { 2206 uint32_t hash; 2207 uint8_t *secbits; 2208 tcp_seq ack, seq; 2209 int wnd, wscale = 0; 2210 union syncookie cookie; 2211 2212 /* 2213 * Pull information out of SYN-ACK/ACK and revert sequence number 2214 * advances. 2215 */ 2216 ack = th->th_ack - 1; 2217 seq = th->th_seq - 1; 2218 2219 /* 2220 * Unpack the flags containing enough information to restore the 2221 * connection. 2222 */ 2223 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2224 2225 /* Which of the two secrets to use. */ 2226 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; 2227 2228 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2229 2230 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2231 if ((ack & ~0xff) != (hash & ~0xff)) 2232 return (NULL); 2233 2234 /* Fill in the syncache values. */ 2235 sc->sc_flags = 0; 2236 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2237 sc->sc_ipopts = NULL; 2238 2239 sc->sc_irs = seq; 2240 sc->sc_iss = ack; 2241 2242 switch (inc->inc_flags & INC_ISIPV6) { 2243 #ifdef INET 2244 case 0: 2245 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2246 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2247 break; 2248 #endif 2249 #ifdef INET6 2250 case INC_ISIPV6: 2251 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2252 sc->sc_flowlabel = 2253 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2254 break; 2255 #endif 2256 } 2257 2258 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2259 2260 /* We can simply recompute receive window scale we sent earlier. */ 2261 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2262 wscale++; 2263 2264 /* Only use wscale if it was enabled in the orignal SYN. */ 2265 if (cookie.flags.wscale_idx > 0) { 2266 sc->sc_requested_r_scale = wscale; 2267 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2268 sc->sc_flags |= SCF_WINSCALE; 2269 } 2270 2271 wnd = lso->sol_sbrcv_hiwat; 2272 wnd = imax(wnd, 0); 2273 wnd = imin(wnd, TCP_MAXWIN); 2274 sc->sc_wnd = wnd; 2275 2276 if (cookie.flags.sack_ok) 2277 sc->sc_flags |= SCF_SACK; 2278 2279 if (to->to_flags & TOF_TS) { 2280 sc->sc_flags |= SCF_TIMESTAMP; 2281 sc->sc_tsreflect = to->to_tsval; 2282 sc->sc_tsoff = tcp_new_ts_offset(inc); 2283 } 2284 2285 if (to->to_flags & TOF_SIGNATURE) 2286 sc->sc_flags |= SCF_SIGNATURE; 2287 2288 sc->sc_rxmits = 0; 2289 2290 TCPSTAT_INC(tcps_sc_recvcookie); 2291 return (sc); 2292 } 2293 2294 #ifdef INVARIANTS 2295 static int 2296 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2297 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2298 struct socket *lso) 2299 { 2300 struct syncache scs, *scx; 2301 char *s; 2302 2303 bzero(&scs, sizeof(scs)); 2304 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2305 2306 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2307 return (0); 2308 2309 if (scx != NULL) { 2310 if (sc->sc_peer_mss != scx->sc_peer_mss) 2311 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2312 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2313 2314 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2315 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2316 s, __func__, sc->sc_requested_r_scale, 2317 scx->sc_requested_r_scale); 2318 2319 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2320 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2321 s, __func__, sc->sc_requested_s_scale, 2322 scx->sc_requested_s_scale); 2323 2324 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2325 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2326 } 2327 2328 if (s != NULL) 2329 free(s, M_TCPLOG); 2330 return (0); 2331 } 2332 #endif /* INVARIANTS */ 2333 2334 static void 2335 syncookie_reseed(void *arg) 2336 { 2337 struct tcp_syncache *sc = arg; 2338 uint8_t *secbits; 2339 int secbit; 2340 2341 /* 2342 * Reseeding the secret doesn't have to be protected by a lock. 2343 * It only must be ensured that the new random values are visible 2344 * to all CPUs in a SMP environment. The atomic with release 2345 * semantics ensures that. 2346 */ 2347 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2348 secbits = sc->secret.key[secbit]; 2349 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2350 atomic_add_rel_int(&sc->secret.oddeven, 1); 2351 2352 /* Reschedule ourself. */ 2353 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2354 } 2355 2356 /* 2357 * We have overflowed a bucket. Let's pause dealing with the syncache. 2358 * This function will increment the bucketoverflow statistics appropriately 2359 * (once per pause when pausing is enabled; otherwise, once per overflow). 2360 */ 2361 static void 2362 syncache_pause(struct in_conninfo *inc) 2363 { 2364 time_t delta; 2365 const char *s; 2366 2367 /* XXX: 2368 * 2. Add sysctl read here so we don't get the benefit of this 2369 * change without the new sysctl. 2370 */ 2371 2372 /* 2373 * Try an unlocked read. If we already know that another thread 2374 * has activated the feature, there is no need to proceed. 2375 */ 2376 if (V_tcp_syncache.paused) 2377 return; 2378 2379 /* Are cookied enabled? If not, we can't pause. */ 2380 if (!V_tcp_syncookies) { 2381 TCPSTAT_INC(tcps_sc_bucketoverflow); 2382 return; 2383 } 2384 2385 /* 2386 * We may be the first thread to find an overflow. Get the lock 2387 * and evaluate if we need to take action. 2388 */ 2389 mtx_lock(&V_tcp_syncache.pause_mtx); 2390 if (V_tcp_syncache.paused) { 2391 mtx_unlock(&V_tcp_syncache.pause_mtx); 2392 return; 2393 } 2394 2395 /* Activate protection. */ 2396 V_tcp_syncache.paused = true; 2397 TCPSTAT_INC(tcps_sc_bucketoverflow); 2398 2399 /* 2400 * Determine the last backoff time. If we are seeing a re-newed 2401 * attack within that same time after last reactivating the syncache, 2402 * consider it an extension of the same attack. 2403 */ 2404 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; 2405 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { 2406 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { 2407 delta <<= 1; 2408 V_tcp_syncache.pause_backoff++; 2409 } 2410 } else { 2411 delta = TCP_SYNCACHE_PAUSE_TIME; 2412 V_tcp_syncache.pause_backoff = 0; 2413 } 2414 2415 /* Log a warning, including IP addresses, if able. */ 2416 if (inc != NULL) 2417 s = tcp_log_addrs(inc, NULL, NULL, NULL); 2418 else 2419 s = (const char *)NULL; 2420 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " 2421 "the next %lld seconds%s%s%s\n", (long long)delta, 2422 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", 2423 (s != NULL) ? ")" : ""); 2424 free(__DECONST(void *, s), M_TCPLOG); 2425 2426 /* Use the calculated delta to set a new pause time. */ 2427 V_tcp_syncache.pause_until = time_uptime + delta; 2428 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, 2429 &V_tcp_syncache); 2430 mtx_unlock(&V_tcp_syncache.pause_mtx); 2431 } 2432 2433 /* Evaluate whether we need to unpause. */ 2434 static void 2435 syncache_unpause(void *arg) 2436 { 2437 struct tcp_syncache *sc; 2438 time_t delta; 2439 2440 sc = arg; 2441 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); 2442 callout_deactivate(&sc->pause_co); 2443 2444 /* 2445 * Check to make sure we are not running early. If the pause 2446 * time has expired, then deactivate the protection. 2447 */ 2448 if ((delta = sc->pause_until - time_uptime) > 0) 2449 callout_schedule(&sc->pause_co, delta * hz); 2450 else 2451 sc->paused = false; 2452 } 2453 2454 /* 2455 * Exports the syncache entries to userland so that netstat can display 2456 * them alongside the other sockets. This function is intended to be 2457 * called only from tcp_pcblist. 2458 * 2459 * Due to concurrency on an active system, the number of pcbs exported 2460 * may have no relation to max_pcbs. max_pcbs merely indicates the 2461 * amount of space the caller allocated for this function to use. 2462 */ 2463 int 2464 syncache_pcblist(struct sysctl_req *req) 2465 { 2466 struct xtcpcb xt; 2467 struct syncache *sc; 2468 struct syncache_head *sch; 2469 int error, i; 2470 2471 bzero(&xt, sizeof(xt)); 2472 xt.xt_len = sizeof(xt); 2473 xt.t_state = TCPS_SYN_RECEIVED; 2474 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2475 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2476 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2477 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2478 2479 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 2480 sch = &V_tcp_syncache.hashbase[i]; 2481 SCH_LOCK(sch); 2482 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2483 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2484 continue; 2485 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2486 xt.xt_inp.inp_vflag = INP_IPV6; 2487 else 2488 xt.xt_inp.inp_vflag = INP_IPV4; 2489 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2490 sizeof (struct in_conninfo)); 2491 error = SYSCTL_OUT(req, &xt, sizeof xt); 2492 if (error) { 2493 SCH_UNLOCK(sch); 2494 return (0); 2495 } 2496 } 2497 SCH_UNLOCK(sch); 2498 } 2499 2500 return (0); 2501 } 2502