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