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 #include <netinet/udp.h> 100 #include <netinet/udp_var.h> 101 102 #include <netipsec/ipsec_support.h> 103 104 #include <machine/in_cksum.h> 105 106 #include <security/mac/mac_framework.h> 107 108 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1; 109 #define V_tcp_syncookies VNET(tcp_syncookies) 110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 111 &VNET_NAME(tcp_syncookies), 0, 112 "Use TCP SYN cookies if the syncache overflows"); 113 114 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0; 115 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 117 &VNET_NAME(tcp_syncookiesonly), 0, 118 "Use only TCP SYN cookies"); 119 120 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1; 121 #define V_functions_inherit_listen_socket_stack \ 122 VNET(functions_inherit_listen_socket_stack) 123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack, 124 CTLFLAG_VNET | CTLFLAG_RW, 125 &VNET_NAME(functions_inherit_listen_socket_stack), 0, 126 "Inherit listen socket's stack"); 127 128 #ifdef TCP_OFFLOAD 129 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 130 #endif 131 132 static void syncache_drop(struct syncache *, struct syncache_head *); 133 static void syncache_free(struct syncache *); 134 static void syncache_insert(struct syncache *, struct syncache_head *); 135 static int syncache_respond(struct syncache *, const struct mbuf *, int); 136 static struct socket *syncache_socket(struct syncache *, struct socket *, 137 struct mbuf *m); 138 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 139 int docallout); 140 static void syncache_timer(void *); 141 142 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 143 uint8_t *, uintptr_t); 144 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 145 static struct syncache 146 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 147 struct syncache *, struct tcphdr *, struct tcpopt *, 148 struct socket *, uint16_t); 149 static void syncache_pause(struct in_conninfo *); 150 static void syncache_unpause(void *); 151 static void syncookie_reseed(void *); 152 #ifdef INVARIANTS 153 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 154 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 155 struct socket *lso, uint16_t port); 156 #endif 157 158 /* 159 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 160 * 3 retransmits corresponds to a timeout with default values of 161 * tcp_rexmit_initial * ( 1 + 162 * tcp_backoff[1] + 163 * tcp_backoff[2] + 164 * tcp_backoff[3]) + 3 * tcp_rexmit_slop, 165 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms, 166 * the odds are that the user has given up attempting to connect by then. 167 */ 168 #define SYNCACHE_MAXREXMTS 3 169 170 /* Arbitrary values */ 171 #define TCP_SYNCACHE_HASHSIZE 512 172 #define TCP_SYNCACHE_BUCKETLIMIT 30 173 174 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 175 #define V_tcp_syncache VNET(tcp_syncache) 176 177 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, 178 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 179 "TCP SYN cache"); 180 181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 182 &VNET_NAME(tcp_syncache.bucket_limit), 0, 183 "Per-bucket hash limit for syncache"); 184 185 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 186 &VNET_NAME(tcp_syncache.cache_limit), 0, 187 "Overall entry limit for syncache"); 188 189 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 190 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 191 192 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 193 &VNET_NAME(tcp_syncache.hashsize), 0, 194 "Size of TCP syncache hashtable"); 195 196 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET | 197 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0, 198 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) " 199 "and mac(4) checks"); 200 201 static int 202 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 203 { 204 int error; 205 u_int new; 206 207 new = V_tcp_syncache.rexmt_limit; 208 error = sysctl_handle_int(oidp, &new, 0, req); 209 if ((error == 0) && (req->newptr != NULL)) { 210 if (new > TCP_MAXRXTSHIFT) 211 error = EINVAL; 212 else 213 V_tcp_syncache.rexmt_limit = new; 214 } 215 return (error); 216 } 217 218 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 219 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 220 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 221 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI", 222 "Limit on SYN/ACK retransmissions"); 223 224 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 225 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 226 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 227 "Send reset on socket allocation failure"); 228 229 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 230 231 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 232 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 233 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 234 235 /* 236 * Requires the syncache entry to be already removed from the bucket list. 237 */ 238 static void 239 syncache_free(struct syncache *sc) 240 { 241 242 if (sc->sc_ipopts) 243 (void) m_free(sc->sc_ipopts); 244 if (sc->sc_cred) 245 crfree(sc->sc_cred); 246 #ifdef MAC 247 mac_syncache_destroy(&sc->sc_label); 248 #endif 249 250 uma_zfree(V_tcp_syncache.zone, sc); 251 } 252 253 void 254 syncache_init(void) 255 { 256 int i; 257 258 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 259 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 260 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 261 V_tcp_syncache.hash_secret = arc4random(); 262 263 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 264 &V_tcp_syncache.hashsize); 265 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 266 &V_tcp_syncache.bucket_limit); 267 if (!powerof2(V_tcp_syncache.hashsize) || 268 V_tcp_syncache.hashsize == 0) { 269 printf("WARNING: syncache hash size is not a power of 2.\n"); 270 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 271 } 272 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 273 274 /* Set limits. */ 275 V_tcp_syncache.cache_limit = 276 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 277 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 278 &V_tcp_syncache.cache_limit); 279 280 /* Allocate the hash table. */ 281 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 282 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 283 284 #ifdef VIMAGE 285 V_tcp_syncache.vnet = curvnet; 286 #endif 287 288 /* Initialize the hash buckets. */ 289 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 290 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 291 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 292 NULL, MTX_DEF); 293 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 294 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 295 V_tcp_syncache.hashbase[i].sch_length = 0; 296 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 297 V_tcp_syncache.hashbase[i].sch_last_overflow = 298 -(SYNCOOKIE_LIFETIME + 1); 299 } 300 301 /* Create the syncache entry zone. */ 302 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 303 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 304 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 305 V_tcp_syncache.cache_limit); 306 307 /* Start the SYN cookie reseeder callout. */ 308 callout_init(&V_tcp_syncache.secret.reseed, 1); 309 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 310 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 311 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 312 syncookie_reseed, &V_tcp_syncache); 313 314 /* Initialize the pause machinery. */ 315 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF); 316 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx, 317 0); 318 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME; 319 V_tcp_syncache.pause_backoff = 0; 320 V_tcp_syncache.paused = false; 321 } 322 323 #ifdef VIMAGE 324 void 325 syncache_destroy(void) 326 { 327 struct syncache_head *sch; 328 struct syncache *sc, *nsc; 329 int i; 330 331 /* 332 * Stop the re-seed timer before freeing resources. No need to 333 * possibly schedule it another time. 334 */ 335 callout_drain(&V_tcp_syncache.secret.reseed); 336 337 /* Stop the SYN cache pause callout. */ 338 mtx_lock(&V_tcp_syncache.pause_mtx); 339 if (callout_stop(&V_tcp_syncache.pause_co) == 0) { 340 mtx_unlock(&V_tcp_syncache.pause_mtx); 341 callout_drain(&V_tcp_syncache.pause_co); 342 } else 343 mtx_unlock(&V_tcp_syncache.pause_mtx); 344 345 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 346 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 347 sch = &V_tcp_syncache.hashbase[i]; 348 callout_drain(&sch->sch_timer); 349 350 SCH_LOCK(sch); 351 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 352 syncache_drop(sc, sch); 353 SCH_UNLOCK(sch); 354 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 355 ("%s: sch->sch_bucket not empty", __func__)); 356 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 357 __func__, sch->sch_length)); 358 mtx_destroy(&sch->sch_mtx); 359 } 360 361 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 362 ("%s: cache_count not 0", __func__)); 363 364 /* Free the allocated global resources. */ 365 uma_zdestroy(V_tcp_syncache.zone); 366 free(V_tcp_syncache.hashbase, M_SYNCACHE); 367 mtx_destroy(&V_tcp_syncache.pause_mtx); 368 } 369 #endif 370 371 /* 372 * Inserts a syncache entry into the specified bucket row. 373 * Locks and unlocks the syncache_head autonomously. 374 */ 375 static void 376 syncache_insert(struct syncache *sc, struct syncache_head *sch) 377 { 378 struct syncache *sc2; 379 380 SCH_LOCK(sch); 381 382 /* 383 * Make sure that we don't overflow the per-bucket limit. 384 * If the bucket is full, toss the oldest element. 385 */ 386 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 387 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 388 ("sch->sch_length incorrect")); 389 syncache_pause(&sc->sc_inc); 390 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 391 sch->sch_last_overflow = time_uptime; 392 syncache_drop(sc2, sch); 393 } 394 395 /* Put it into the bucket. */ 396 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 397 sch->sch_length++; 398 399 #ifdef TCP_OFFLOAD 400 if (ADDED_BY_TOE(sc)) { 401 struct toedev *tod = sc->sc_tod; 402 403 tod->tod_syncache_added(tod, sc->sc_todctx); 404 } 405 #endif 406 407 /* Reinitialize the bucket row's timer. */ 408 if (sch->sch_length == 1) 409 sch->sch_nextc = ticks + INT_MAX; 410 syncache_timeout(sc, sch, 1); 411 412 SCH_UNLOCK(sch); 413 414 TCPSTATES_INC(TCPS_SYN_RECEIVED); 415 TCPSTAT_INC(tcps_sc_added); 416 } 417 418 /* 419 * Remove and free entry from syncache bucket row. 420 * Expects locked syncache head. 421 */ 422 static void 423 syncache_drop(struct syncache *sc, struct syncache_head *sch) 424 { 425 426 SCH_LOCK_ASSERT(sch); 427 428 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 429 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 430 sch->sch_length--; 431 432 #ifdef TCP_OFFLOAD 433 if (ADDED_BY_TOE(sc)) { 434 struct toedev *tod = sc->sc_tod; 435 436 tod->tod_syncache_removed(tod, sc->sc_todctx); 437 } 438 #endif 439 440 syncache_free(sc); 441 } 442 443 /* 444 * Engage/reengage time on bucket row. 445 */ 446 static void 447 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 448 { 449 int rexmt; 450 451 if (sc->sc_rxmits == 0) 452 rexmt = tcp_rexmit_initial; 453 else 454 TCPT_RANGESET(rexmt, 455 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], 456 tcp_rexmit_min, TCPTV_REXMTMAX); 457 sc->sc_rxttime = ticks + rexmt; 458 sc->sc_rxmits++; 459 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 460 sch->sch_nextc = sc->sc_rxttime; 461 if (docallout) 462 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 463 syncache_timer, (void *)sch); 464 } 465 } 466 467 /* 468 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 469 * If we have retransmitted an entry the maximum number of times, expire it. 470 * One separate timer for each bucket row. 471 */ 472 static void 473 syncache_timer(void *xsch) 474 { 475 struct syncache_head *sch = (struct syncache_head *)xsch; 476 struct syncache *sc, *nsc; 477 struct epoch_tracker et; 478 int tick = ticks; 479 char *s; 480 bool paused; 481 482 CURVNET_SET(sch->sch_sc->vnet); 483 484 /* NB: syncache_head has already been locked by the callout. */ 485 SCH_LOCK_ASSERT(sch); 486 487 /* 488 * In the following cycle we may remove some entries and/or 489 * advance some timeouts, so re-initialize the bucket timer. 490 */ 491 sch->sch_nextc = tick + INT_MAX; 492 493 /* 494 * If we have paused processing, unconditionally remove 495 * all syncache entries. 496 */ 497 mtx_lock(&V_tcp_syncache.pause_mtx); 498 paused = V_tcp_syncache.paused; 499 mtx_unlock(&V_tcp_syncache.pause_mtx); 500 501 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 502 if (paused) { 503 syncache_drop(sc, sch); 504 continue; 505 } 506 /* 507 * We do not check if the listen socket still exists 508 * and accept the case where the listen socket may be 509 * gone by the time we resend the SYN/ACK. We do 510 * not expect this to happens often. If it does, 511 * then the RST will be sent by the time the remote 512 * host does the SYN/ACK->ACK. 513 */ 514 if (TSTMP_GT(sc->sc_rxttime, tick)) { 515 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 516 sch->sch_nextc = sc->sc_rxttime; 517 continue; 518 } 519 if (sc->sc_rxmits > V_tcp_ecn_maxretries) { 520 sc->sc_flags &= ~SCF_ECN; 521 } 522 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 523 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 524 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 525 "giving up and removing syncache entry\n", 526 s, __func__); 527 free(s, M_TCPLOG); 528 } 529 syncache_drop(sc, sch); 530 TCPSTAT_INC(tcps_sc_stale); 531 continue; 532 } 533 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 534 log(LOG_DEBUG, "%s; %s: Response timeout, " 535 "retransmitting (%u) SYN|ACK\n", 536 s, __func__, sc->sc_rxmits); 537 free(s, M_TCPLOG); 538 } 539 540 NET_EPOCH_ENTER(et); 541 syncache_respond(sc, NULL, TH_SYN|TH_ACK); 542 NET_EPOCH_EXIT(et); 543 TCPSTAT_INC(tcps_sc_retransmitted); 544 syncache_timeout(sc, sch, 0); 545 } 546 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 547 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 548 syncache_timer, (void *)(sch)); 549 CURVNET_RESTORE(); 550 } 551 552 /* 553 * Returns true if the system is only using cookies at the moment. 554 * This could be due to a sysadmin decision to only use cookies, or it 555 * could be due to the system detecting an attack. 556 */ 557 static inline bool 558 syncache_cookiesonly(void) 559 { 560 561 return (V_tcp_syncookies && (V_tcp_syncache.paused || 562 V_tcp_syncookiesonly)); 563 } 564 565 /* 566 * Find the hash bucket for the given connection. 567 */ 568 static struct syncache_head * 569 syncache_hashbucket(struct in_conninfo *inc) 570 { 571 uint32_t hash; 572 573 /* 574 * The hash is built on foreign port + local port + foreign address. 575 * We rely on the fact that struct in_conninfo starts with 16 bits 576 * of foreign port, then 16 bits of local port then followed by 128 577 * bits of foreign address. In case of IPv4 address, the first 3 578 * 32-bit words of the address always are zeroes. 579 */ 580 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 581 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 582 583 return (&V_tcp_syncache.hashbase[hash]); 584 } 585 586 /* 587 * Find an entry in the syncache. 588 * Returns always with locked syncache_head plus a matching entry or NULL. 589 */ 590 static struct syncache * 591 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 592 { 593 struct syncache *sc; 594 struct syncache_head *sch; 595 596 *schp = sch = syncache_hashbucket(inc); 597 SCH_LOCK(sch); 598 599 /* Circle through bucket row to find matching entry. */ 600 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 601 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 602 sizeof(struct in_endpoints)) == 0) 603 break; 604 605 return (sc); /* Always returns with locked sch. */ 606 } 607 608 /* 609 * This function is called when we get a RST for a 610 * non-existent connection, so that we can see if the 611 * connection is in the syn cache. If it is, zap it. 612 * If required send a challenge ACK. 613 */ 614 void 615 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m, 616 uint16_t port) 617 { 618 struct syncache *sc; 619 struct syncache_head *sch; 620 char *s = NULL; 621 622 if (syncache_cookiesonly()) 623 return; 624 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 625 SCH_LOCK_ASSERT(sch); 626 627 /* 628 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 629 * See RFC 793 page 65, section SEGMENT ARRIVES. 630 */ 631 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 632 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 633 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 634 "FIN flag set, segment ignored\n", s, __func__); 635 TCPSTAT_INC(tcps_badrst); 636 goto done; 637 } 638 639 /* 640 * No corresponding connection was found in syncache. 641 * If syncookies are enabled and possibly exclusively 642 * used, or we are under memory pressure, a valid RST 643 * may not find a syncache entry. In that case we're 644 * done and no SYN|ACK retransmissions will happen. 645 * Otherwise the RST was misdirected or spoofed. 646 */ 647 if (sc == NULL) { 648 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 649 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 650 "syncache entry (possibly syncookie only), " 651 "segment ignored\n", s, __func__); 652 TCPSTAT_INC(tcps_badrst); 653 goto done; 654 } 655 656 /* The remote UDP encaps port does not match. */ 657 if (sc->sc_port != port) { 658 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 659 log(LOG_DEBUG, "%s; %s: Spurious RST with matching " 660 "syncache entry but non-matching UDP encaps port, " 661 "segment ignored\n", s, __func__); 662 TCPSTAT_INC(tcps_badrst); 663 goto done; 664 } 665 666 /* 667 * If the RST bit is set, check the sequence number to see 668 * if this is a valid reset segment. 669 * 670 * RFC 793 page 37: 671 * In all states except SYN-SENT, all reset (RST) segments 672 * are validated by checking their SEQ-fields. A reset is 673 * valid if its sequence number is in the window. 674 * 675 * RFC 793 page 69: 676 * There are four cases for the acceptability test for an incoming 677 * segment: 678 * 679 * Segment Receive Test 680 * Length Window 681 * ------- ------- ------------------------------------------- 682 * 0 0 SEG.SEQ = RCV.NXT 683 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 684 * >0 0 not acceptable 685 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 686 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 687 * 688 * Note that when receiving a SYN segment in the LISTEN state, 689 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 690 * described in RFC 793, page 66. 691 */ 692 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 693 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 694 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 695 if (V_tcp_insecure_rst || 696 th->th_seq == sc->sc_irs + 1) { 697 syncache_drop(sc, sch); 698 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 699 log(LOG_DEBUG, 700 "%s; %s: Our SYN|ACK was rejected, " 701 "connection attempt aborted by remote " 702 "endpoint\n", 703 s, __func__); 704 TCPSTAT_INC(tcps_sc_reset); 705 } else { 706 TCPSTAT_INC(tcps_badrst); 707 /* Send challenge ACK. */ 708 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 709 log(LOG_DEBUG, "%s; %s: RST with invalid " 710 " SEQ %u != NXT %u (+WND %u), " 711 "sending challenge ACK\n", 712 s, __func__, 713 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 714 syncache_respond(sc, m, TH_ACK); 715 } 716 } else { 717 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 718 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 719 "NXT %u (+WND %u), segment ignored\n", 720 s, __func__, 721 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 722 TCPSTAT_INC(tcps_badrst); 723 } 724 725 done: 726 if (s != NULL) 727 free(s, M_TCPLOG); 728 SCH_UNLOCK(sch); 729 } 730 731 void 732 syncache_badack(struct in_conninfo *inc, uint16_t port) 733 { 734 struct syncache *sc; 735 struct syncache_head *sch; 736 737 if (syncache_cookiesonly()) 738 return; 739 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 740 SCH_LOCK_ASSERT(sch); 741 if ((sc != NULL) && (sc->sc_port == port)) { 742 syncache_drop(sc, sch); 743 TCPSTAT_INC(tcps_sc_badack); 744 } 745 SCH_UNLOCK(sch); 746 } 747 748 void 749 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port) 750 { 751 struct syncache *sc; 752 struct syncache_head *sch; 753 754 if (syncache_cookiesonly()) 755 return; 756 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 757 SCH_LOCK_ASSERT(sch); 758 if (sc == NULL) 759 goto done; 760 761 /* If the port != sc_port, then it's a bogus ICMP msg */ 762 if (port != sc->sc_port) 763 goto done; 764 765 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 766 if (ntohl(th_seq) != sc->sc_iss) 767 goto done; 768 769 /* 770 * If we've rertransmitted 3 times and this is our second error, 771 * we remove the entry. Otherwise, we allow it to continue on. 772 * This prevents us from incorrectly nuking an entry during a 773 * spurious network outage. 774 * 775 * See tcp_notify(). 776 */ 777 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 778 sc->sc_flags |= SCF_UNREACH; 779 goto done; 780 } 781 syncache_drop(sc, sch); 782 TCPSTAT_INC(tcps_sc_unreach); 783 done: 784 SCH_UNLOCK(sch); 785 } 786 787 /* 788 * Build a new TCP socket structure from a syncache entry. 789 * 790 * On success return the newly created socket with its underlying inp locked. 791 */ 792 static struct socket * 793 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 794 { 795 struct tcp_function_block *blk; 796 struct inpcb *inp = NULL; 797 struct socket *so; 798 struct tcpcb *tp; 799 int error; 800 char *s; 801 802 NET_EPOCH_ASSERT(); 803 804 /* 805 * Ok, create the full blown connection, and set things up 806 * as they would have been set up if we had created the 807 * connection when the SYN arrived. If we can't create 808 * the connection, abort it. 809 */ 810 so = sonewconn(lso, 0); 811 if (so == NULL) { 812 /* 813 * Drop the connection; we will either send a RST or 814 * have the peer retransmit its SYN again after its 815 * RTO and try again. 816 */ 817 TCPSTAT_INC(tcps_listendrop); 818 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 819 log(LOG_DEBUG, "%s; %s: Socket create failed " 820 "due to limits or memory shortage\n", 821 s, __func__); 822 free(s, M_TCPLOG); 823 } 824 goto abort2; 825 } 826 #ifdef MAC 827 mac_socketpeer_set_from_mbuf(m, so); 828 #endif 829 830 inp = sotoinpcb(so); 831 inp->inp_inc.inc_fibnum = so->so_fibnum; 832 INP_WLOCK(inp); 833 /* 834 * Exclusive pcbinfo lock is not required in syncache socket case even 835 * if two inpcb locks can be acquired simultaneously: 836 * - the inpcb in LISTEN state, 837 * - the newly created inp. 838 * 839 * In this case, an inp cannot be at same time in LISTEN state and 840 * just created by an accept() call. 841 */ 842 INP_HASH_WLOCK(&V_tcbinfo); 843 844 /* Insert new socket into PCB hash list. */ 845 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 846 #ifdef INET6 847 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 848 inp->inp_vflag &= ~INP_IPV4; 849 inp->inp_vflag |= INP_IPV6; 850 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 851 } else { 852 inp->inp_vflag &= ~INP_IPV6; 853 inp->inp_vflag |= INP_IPV4; 854 #endif 855 inp->inp_ip_ttl = sc->sc_ip_ttl; 856 inp->inp_ip_tos = sc->sc_ip_tos; 857 inp->inp_laddr = sc->sc_inc.inc_laddr; 858 #ifdef INET6 859 } 860 #endif 861 862 /* 863 * If there's an mbuf and it has a flowid, then let's initialise the 864 * inp with that particular flowid. 865 */ 866 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 867 inp->inp_flowid = m->m_pkthdr.flowid; 868 inp->inp_flowtype = M_HASHTYPE_GET(m); 869 #ifdef NUMA 870 inp->inp_numa_domain = m->m_pkthdr.numa_domain; 871 #endif 872 } 873 874 inp->inp_lport = sc->sc_inc.inc_lport; 875 #ifdef INET6 876 if (inp->inp_vflag & INP_IPV6PROTO) { 877 struct inpcb *oinp = sotoinpcb(lso); 878 879 /* 880 * Inherit socket options from the listening socket. 881 * Note that in6p_inputopts are not (and should not be) 882 * copied, since it stores previously received options and is 883 * used to detect if each new option is different than the 884 * previous one and hence should be passed to a user. 885 * If we copied in6p_inputopts, a user would not be able to 886 * receive options just after calling the accept system call. 887 */ 888 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 889 if (oinp->in6p_outputopts) 890 inp->in6p_outputopts = 891 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 892 inp->in6p_hops = oinp->in6p_hops; 893 } 894 895 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 896 struct in6_addr laddr6; 897 struct sockaddr_in6 sin6; 898 899 sin6.sin6_family = AF_INET6; 900 sin6.sin6_len = sizeof(sin6); 901 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 902 sin6.sin6_port = sc->sc_inc.inc_fport; 903 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 904 laddr6 = inp->in6p_laddr; 905 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 906 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 907 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 908 thread0.td_ucred, m, false)) != 0) { 909 inp->in6p_laddr = laddr6; 910 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 911 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 912 "with error %i\n", 913 s, __func__, error); 914 free(s, M_TCPLOG); 915 } 916 INP_HASH_WUNLOCK(&V_tcbinfo); 917 goto abort; 918 } 919 /* Override flowlabel from in6_pcbconnect. */ 920 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 921 inp->inp_flow |= sc->sc_flowlabel; 922 } 923 #endif /* INET6 */ 924 #if defined(INET) && defined(INET6) 925 else 926 #endif 927 #ifdef INET 928 { 929 struct in_addr laddr; 930 struct sockaddr_in sin; 931 932 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 933 934 if (inp->inp_options == NULL) { 935 inp->inp_options = sc->sc_ipopts; 936 sc->sc_ipopts = NULL; 937 } 938 939 sin.sin_family = AF_INET; 940 sin.sin_len = sizeof(sin); 941 sin.sin_addr = sc->sc_inc.inc_faddr; 942 sin.sin_port = sc->sc_inc.inc_fport; 943 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 944 laddr = inp->inp_laddr; 945 if (inp->inp_laddr.s_addr == INADDR_ANY) 946 inp->inp_laddr = sc->sc_inc.inc_laddr; 947 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 948 thread0.td_ucred, m, false)) != 0) { 949 inp->inp_laddr = laddr; 950 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 951 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 952 "with error %i\n", 953 s, __func__, error); 954 free(s, M_TCPLOG); 955 } 956 INP_HASH_WUNLOCK(&V_tcbinfo); 957 goto abort; 958 } 959 } 960 #endif /* INET */ 961 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 962 /* Copy old policy into new socket's. */ 963 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 964 printf("syncache_socket: could not copy policy\n"); 965 #endif 966 INP_HASH_WUNLOCK(&V_tcbinfo); 967 tp = intotcpcb(inp); 968 tcp_state_change(tp, TCPS_SYN_RECEIVED); 969 tp->iss = sc->sc_iss; 970 tp->irs = sc->sc_irs; 971 tp->t_port = sc->sc_port; 972 tcp_rcvseqinit(tp); 973 tcp_sendseqinit(tp); 974 blk = sototcpcb(lso)->t_fb; 975 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 976 /* 977 * Our parents t_fb was not the default, 978 * we need to release our ref on tp->t_fb and 979 * pickup one on the new entry. 980 */ 981 struct tcp_function_block *rblk; 982 983 rblk = find_and_ref_tcp_fb(blk); 984 KASSERT(rblk != NULL, 985 ("cannot find blk %p out of syncache?", blk)); 986 if (tp->t_fb->tfb_tcp_fb_fini) 987 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 988 refcount_release(&tp->t_fb->tfb_refcnt); 989 tp->t_fb = rblk; 990 /* 991 * XXXrrs this is quite dangerous, it is possible 992 * for the new function to fail to init. We also 993 * are not asking if the handoff_is_ok though at 994 * the very start thats probalbly ok. 995 */ 996 if (tp->t_fb->tfb_tcp_fb_init) { 997 (*tp->t_fb->tfb_tcp_fb_init)(tp); 998 } 999 } 1000 tp->snd_wl1 = sc->sc_irs; 1001 tp->snd_max = tp->iss + 1; 1002 tp->snd_nxt = tp->iss + 1; 1003 tp->rcv_up = sc->sc_irs + 1; 1004 tp->rcv_wnd = sc->sc_wnd; 1005 tp->rcv_adv += tp->rcv_wnd; 1006 tp->last_ack_sent = tp->rcv_nxt; 1007 1008 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 1009 if (sc->sc_flags & SCF_NOOPT) 1010 tp->t_flags |= TF_NOOPT; 1011 else { 1012 if (sc->sc_flags & SCF_WINSCALE) { 1013 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 1014 tp->snd_scale = sc->sc_requested_s_scale; 1015 tp->request_r_scale = sc->sc_requested_r_scale; 1016 } 1017 if (sc->sc_flags & SCF_TIMESTAMP) { 1018 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 1019 tp->ts_recent = sc->sc_tsreflect; 1020 tp->ts_recent_age = tcp_ts_getticks(); 1021 tp->ts_offset = sc->sc_tsoff; 1022 } 1023 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1024 if (sc->sc_flags & SCF_SIGNATURE) 1025 tp->t_flags |= TF_SIGNATURE; 1026 #endif 1027 if (sc->sc_flags & SCF_SACK) 1028 tp->t_flags |= TF_SACK_PERMIT; 1029 } 1030 1031 if (sc->sc_flags & SCF_ECN) 1032 tp->t_flags2 |= TF2_ECN_PERMIT; 1033 1034 /* 1035 * Set up MSS and get cached values from tcp_hostcache. 1036 * This might overwrite some of the defaults we just set. 1037 */ 1038 tcp_mss(tp, sc->sc_peer_mss); 1039 1040 /* 1041 * If the SYN,ACK was retransmitted, indicate that CWND to be 1042 * limited to one segment in cc_conn_init(). 1043 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 1044 */ 1045 if (sc->sc_rxmits > 1) 1046 tp->snd_cwnd = 1; 1047 1048 #ifdef TCP_OFFLOAD 1049 /* 1050 * Allow a TOE driver to install its hooks. Note that we hold the 1051 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 1052 * new connection before the TOE driver has done its thing. 1053 */ 1054 if (ADDED_BY_TOE(sc)) { 1055 struct toedev *tod = sc->sc_tod; 1056 1057 tod->tod_offload_socket(tod, sc->sc_todctx, so); 1058 } 1059 #endif 1060 /* 1061 * Copy and activate timers. 1062 */ 1063 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 1064 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 1065 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1066 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1067 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1068 1069 TCPSTAT_INC(tcps_accepts); 1070 return (so); 1071 1072 abort: 1073 INP_WUNLOCK(inp); 1074 abort2: 1075 if (so != NULL) 1076 soabort(so); 1077 return (NULL); 1078 } 1079 1080 /* 1081 * This function gets called when we receive an ACK for a 1082 * socket in the LISTEN state. We look up the connection 1083 * in the syncache, and if its there, we pull it out of 1084 * the cache and turn it into a full-blown connection in 1085 * the SYN-RECEIVED state. 1086 * 1087 * On syncache_socket() success the newly created socket 1088 * has its underlying inp locked. 1089 */ 1090 int 1091 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1092 struct socket **lsop, struct mbuf *m, uint16_t port) 1093 { 1094 struct syncache *sc; 1095 struct syncache_head *sch; 1096 struct syncache scs; 1097 char *s; 1098 bool locked; 1099 1100 NET_EPOCH_ASSERT(); 1101 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1102 ("%s: can handle only ACK", __func__)); 1103 1104 if (syncache_cookiesonly()) { 1105 sc = NULL; 1106 sch = syncache_hashbucket(inc); 1107 locked = false; 1108 } else { 1109 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1110 locked = true; 1111 SCH_LOCK_ASSERT(sch); 1112 } 1113 1114 #ifdef INVARIANTS 1115 /* 1116 * Test code for syncookies comparing the syncache stored 1117 * values with the reconstructed values from the cookie. 1118 */ 1119 if (sc != NULL) 1120 syncookie_cmp(inc, sch, sc, th, to, *lsop, port); 1121 #endif 1122 1123 if (sc == NULL) { 1124 /* 1125 * There is no syncache entry, so see if this ACK is 1126 * a returning syncookie. To do this, first: 1127 * A. Check if syncookies are used in case of syncache 1128 * overflows 1129 * B. See if this socket has had a syncache entry dropped in 1130 * the recent past. We don't want to accept a bogus 1131 * syncookie if we've never received a SYN or accept it 1132 * twice. 1133 * C. check that the syncookie is valid. If it is, then 1134 * cobble up a fake syncache entry, and return. 1135 */ 1136 if (locked && !V_tcp_syncookies) { 1137 SCH_UNLOCK(sch); 1138 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1139 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1140 "segment rejected (syncookies disabled)\n", 1141 s, __func__); 1142 goto failed; 1143 } 1144 if (locked && !V_tcp_syncookiesonly && 1145 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1146 SCH_UNLOCK(sch); 1147 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1148 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1149 "segment rejected (no syncache entry)\n", 1150 s, __func__); 1151 goto failed; 1152 } 1153 bzero(&scs, sizeof(scs)); 1154 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port); 1155 if (locked) 1156 SCH_UNLOCK(sch); 1157 if (sc == NULL) { 1158 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1159 log(LOG_DEBUG, "%s; %s: Segment failed " 1160 "SYNCOOKIE authentication, segment rejected " 1161 "(probably spoofed)\n", s, __func__); 1162 goto failed; 1163 } 1164 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1165 /* If received ACK has MD5 signature, check it. */ 1166 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1167 (!TCPMD5_ENABLED() || 1168 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1169 /* Drop the ACK. */ 1170 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1171 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1172 "MD5 signature doesn't match.\n", 1173 s, __func__); 1174 free(s, M_TCPLOG); 1175 } 1176 TCPSTAT_INC(tcps_sig_err_sigopt); 1177 return (-1); /* Do not send RST */ 1178 } 1179 #endif /* TCP_SIGNATURE */ 1180 } else { 1181 if (sc->sc_port != port) { 1182 SCH_UNLOCK(sch); 1183 return (0); 1184 } 1185 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1186 /* 1187 * If listening socket requested TCP digests, check that 1188 * received ACK has signature and it is correct. 1189 * If not, drop the ACK and leave sc entry in th cache, 1190 * because SYN was received with correct signature. 1191 */ 1192 if (sc->sc_flags & SCF_SIGNATURE) { 1193 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1194 /* No signature */ 1195 TCPSTAT_INC(tcps_sig_err_nosigopt); 1196 SCH_UNLOCK(sch); 1197 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1198 log(LOG_DEBUG, "%s; %s: Segment " 1199 "rejected, MD5 signature wasn't " 1200 "provided.\n", s, __func__); 1201 free(s, M_TCPLOG); 1202 } 1203 return (-1); /* Do not send RST */ 1204 } 1205 if (!TCPMD5_ENABLED() || 1206 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1207 /* Doesn't match or no SA */ 1208 SCH_UNLOCK(sch); 1209 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1210 log(LOG_DEBUG, "%s; %s: Segment " 1211 "rejected, MD5 signature doesn't " 1212 "match.\n", s, __func__); 1213 free(s, M_TCPLOG); 1214 } 1215 return (-1); /* Do not send RST */ 1216 } 1217 } 1218 #endif /* TCP_SIGNATURE */ 1219 1220 /* 1221 * RFC 7323 PAWS: If we have a timestamp on this segment and 1222 * it's less than ts_recent, drop it. 1223 * XXXMT: RFC 7323 also requires to send an ACK. 1224 * In tcp_input.c this is only done for TCP segments 1225 * with user data, so be consistent here and just drop 1226 * the segment. 1227 */ 1228 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1229 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1230 SCH_UNLOCK(sch); 1231 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1232 log(LOG_DEBUG, 1233 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1234 "segment dropped\n", s, __func__, 1235 to->to_tsval, sc->sc_tsreflect); 1236 free(s, M_TCPLOG); 1237 } 1238 return (-1); /* Do not send RST */ 1239 } 1240 1241 /* 1242 * If timestamps were not negotiated during SYN/ACK and a 1243 * segment with a timestamp is received, ignore the 1244 * timestamp and process the packet normally. 1245 * See section 3.2 of RFC 7323. 1246 */ 1247 if (!(sc->sc_flags & SCF_TIMESTAMP) && 1248 (to->to_flags & TOF_TS)) { 1249 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1250 log(LOG_DEBUG, "%s; %s: Timestamp not " 1251 "expected, segment processed normally\n", 1252 s, __func__); 1253 free(s, M_TCPLOG); 1254 s = NULL; 1255 } 1256 } 1257 1258 /* 1259 * If timestamps were negotiated during SYN/ACK and a 1260 * segment without a timestamp is received, silently drop 1261 * the segment, unless the missing timestamps are tolerated. 1262 * See section 3.2 of RFC 7323. 1263 */ 1264 if ((sc->sc_flags & SCF_TIMESTAMP) && 1265 !(to->to_flags & TOF_TS)) { 1266 if (V_tcp_tolerate_missing_ts) { 1267 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1268 log(LOG_DEBUG, 1269 "%s; %s: Timestamp missing, " 1270 "segment processed normally\n", 1271 s, __func__); 1272 free(s, M_TCPLOG); 1273 } 1274 } else { 1275 SCH_UNLOCK(sch); 1276 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1277 log(LOG_DEBUG, 1278 "%s; %s: Timestamp missing, " 1279 "segment silently dropped\n", 1280 s, __func__); 1281 free(s, M_TCPLOG); 1282 } 1283 return (-1); /* Do not send RST */ 1284 } 1285 } 1286 1287 /* 1288 * Pull out the entry to unlock the bucket row. 1289 * 1290 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1291 * tcp_state_change(). The tcpcb is not existent at this 1292 * moment. A new one will be allocated via syncache_socket-> 1293 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1294 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1295 */ 1296 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1297 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1298 sch->sch_length--; 1299 #ifdef TCP_OFFLOAD 1300 if (ADDED_BY_TOE(sc)) { 1301 struct toedev *tod = sc->sc_tod; 1302 1303 tod->tod_syncache_removed(tod, sc->sc_todctx); 1304 } 1305 #endif 1306 SCH_UNLOCK(sch); 1307 } 1308 1309 /* 1310 * Segment validation: 1311 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1312 */ 1313 if (th->th_ack != sc->sc_iss + 1) { 1314 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1315 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1316 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1317 goto failed; 1318 } 1319 1320 /* 1321 * The SEQ must fall in the window starting at the received 1322 * initial receive sequence number + 1 (the SYN). 1323 */ 1324 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1325 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1326 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1327 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1328 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1329 goto failed; 1330 } 1331 1332 *lsop = syncache_socket(sc, *lsop, m); 1333 1334 if (*lsop == NULL) 1335 TCPSTAT_INC(tcps_sc_aborted); 1336 else 1337 TCPSTAT_INC(tcps_sc_completed); 1338 1339 /* how do we find the inp for the new socket? */ 1340 if (sc != &scs) 1341 syncache_free(sc); 1342 return (1); 1343 failed: 1344 if (sc != NULL && sc != &scs) 1345 syncache_free(sc); 1346 if (s != NULL) 1347 free(s, M_TCPLOG); 1348 *lsop = NULL; 1349 return (0); 1350 } 1351 1352 static struct socket * 1353 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m, 1354 uint64_t response_cookie) 1355 { 1356 struct inpcb *inp; 1357 struct tcpcb *tp; 1358 unsigned int *pending_counter; 1359 struct socket *so; 1360 1361 NET_EPOCH_ASSERT(); 1362 1363 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending; 1364 so = syncache_socket(sc, lso, m); 1365 if (so == NULL) { 1366 TCPSTAT_INC(tcps_sc_aborted); 1367 atomic_subtract_int(pending_counter, 1); 1368 } else { 1369 soisconnected(so); 1370 inp = sotoinpcb(so); 1371 tp = intotcpcb(inp); 1372 tp->t_flags |= TF_FASTOPEN; 1373 tp->t_tfo_cookie.server = response_cookie; 1374 tp->snd_max = tp->iss; 1375 tp->snd_nxt = tp->iss; 1376 tp->t_tfo_pending = pending_counter; 1377 TCPSTAT_INC(tcps_sc_completed); 1378 } 1379 1380 return (so); 1381 } 1382 1383 /* 1384 * Given a LISTEN socket and an inbound SYN request, add 1385 * this to the syn cache, and send back a segment: 1386 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1387 * to the source. 1388 * 1389 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1390 * Doing so would require that we hold onto the data and deliver it 1391 * to the application. However, if we are the target of a SYN-flood 1392 * DoS attack, an attacker could send data which would eventually 1393 * consume all available buffer space if it were ACKed. By not ACKing 1394 * the data, we avoid this DoS scenario. 1395 * 1396 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1397 * cookie is processed and a new socket is created. In this case, any data 1398 * accompanying the SYN will be queued to the socket by tcp_input() and will 1399 * be ACKed either when the application sends response data or the delayed 1400 * ACK timer expires, whichever comes first. 1401 */ 1402 struct socket * 1403 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1404 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod, 1405 void *todctx, uint8_t iptos, uint16_t port) 1406 { 1407 struct tcpcb *tp; 1408 struct socket *rv = NULL; 1409 struct syncache *sc = NULL; 1410 struct syncache_head *sch; 1411 struct mbuf *ipopts = NULL; 1412 u_int ltflags; 1413 int win, ip_ttl, ip_tos; 1414 char *s; 1415 #ifdef INET6 1416 int autoflowlabel = 0; 1417 #endif 1418 #ifdef MAC 1419 struct label *maclabel; 1420 #endif 1421 struct syncache scs; 1422 struct ucred *cred; 1423 uint64_t tfo_response_cookie; 1424 unsigned int *tfo_pending = NULL; 1425 int tfo_cookie_valid = 0; 1426 int tfo_response_cookie_valid = 0; 1427 bool locked; 1428 1429 INP_RLOCK_ASSERT(inp); /* listen socket */ 1430 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1431 ("%s: unexpected tcp flags", __func__)); 1432 1433 /* 1434 * Combine all so/tp operations very early to drop the INP lock as 1435 * soon as possible. 1436 */ 1437 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1438 tp = sototcpcb(so); 1439 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred); 1440 1441 #ifdef INET6 1442 if (inc->inc_flags & INC_ISIPV6) { 1443 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1444 autoflowlabel = 1; 1445 } 1446 ip_ttl = in6_selecthlim(inp, NULL); 1447 if ((inp->in6p_outputopts == NULL) || 1448 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1449 ip_tos = 0; 1450 } else { 1451 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1452 } 1453 } 1454 #endif 1455 #if defined(INET6) && defined(INET) 1456 else 1457 #endif 1458 #ifdef INET 1459 { 1460 ip_ttl = inp->inp_ip_ttl; 1461 ip_tos = inp->inp_ip_tos; 1462 } 1463 #endif 1464 win = so->sol_sbrcv_hiwat; 1465 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1466 1467 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1468 (tp->t_tfo_pending != NULL) && 1469 (to->to_flags & TOF_FASTOPEN)) { 1470 /* 1471 * Limit the number of pending TFO connections to 1472 * approximately half of the queue limit. This prevents TFO 1473 * SYN floods from starving the service by filling the 1474 * listen queue with bogus TFO connections. 1475 */ 1476 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1477 (so->sol_qlimit / 2)) { 1478 int result; 1479 1480 result = tcp_fastopen_check_cookie(inc, 1481 to->to_tfo_cookie, to->to_tfo_len, 1482 &tfo_response_cookie); 1483 tfo_cookie_valid = (result > 0); 1484 tfo_response_cookie_valid = (result >= 0); 1485 } 1486 1487 /* 1488 * Remember the TFO pending counter as it will have to be 1489 * decremented below if we don't make it to syncache_tfo_expand(). 1490 */ 1491 tfo_pending = tp->t_tfo_pending; 1492 } 1493 1494 /* By the time we drop the lock these should no longer be used. */ 1495 so = NULL; 1496 tp = NULL; 1497 1498 #ifdef MAC 1499 if (mac_syncache_init(&maclabel) != 0) { 1500 INP_RUNLOCK(inp); 1501 goto done; 1502 } else 1503 mac_syncache_create(maclabel, inp); 1504 #endif 1505 if (!tfo_cookie_valid) 1506 INP_RUNLOCK(inp); 1507 1508 /* 1509 * Remember the IP options, if any. 1510 */ 1511 #ifdef INET6 1512 if (!(inc->inc_flags & INC_ISIPV6)) 1513 #endif 1514 #ifdef INET 1515 ipopts = (m) ? ip_srcroute(m) : NULL; 1516 #else 1517 ipopts = NULL; 1518 #endif 1519 1520 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1521 /* 1522 * If listening socket requested TCP digests, check that received 1523 * SYN has signature and it is correct. If signature doesn't match 1524 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1525 */ 1526 if (ltflags & TF_SIGNATURE) { 1527 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1528 TCPSTAT_INC(tcps_sig_err_nosigopt); 1529 goto done; 1530 } 1531 if (!TCPMD5_ENABLED() || 1532 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1533 goto done; 1534 } 1535 #endif /* TCP_SIGNATURE */ 1536 /* 1537 * See if we already have an entry for this connection. 1538 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1539 * 1540 * XXX: should the syncache be re-initialized with the contents 1541 * of the new SYN here (which may have different options?) 1542 * 1543 * XXX: We do not check the sequence number to see if this is a 1544 * real retransmit or a new connection attempt. The question is 1545 * how to handle such a case; either ignore it as spoofed, or 1546 * drop the current entry and create a new one? 1547 */ 1548 if (syncache_cookiesonly()) { 1549 sc = NULL; 1550 sch = syncache_hashbucket(inc); 1551 locked = false; 1552 } else { 1553 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1554 locked = true; 1555 SCH_LOCK_ASSERT(sch); 1556 } 1557 if (sc != NULL) { 1558 if (tfo_cookie_valid) 1559 INP_RUNLOCK(inp); 1560 TCPSTAT_INC(tcps_sc_dupsyn); 1561 if (ipopts) { 1562 /* 1563 * If we were remembering a previous source route, 1564 * forget it and use the new one we've been given. 1565 */ 1566 if (sc->sc_ipopts) 1567 (void) m_free(sc->sc_ipopts); 1568 sc->sc_ipopts = ipopts; 1569 } 1570 /* 1571 * Update timestamp if present. 1572 */ 1573 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1574 sc->sc_tsreflect = to->to_tsval; 1575 else 1576 sc->sc_flags &= ~SCF_TIMESTAMP; 1577 /* 1578 * Disable ECN if needed. 1579 */ 1580 if ((sc->sc_flags & SCF_ECN) && 1581 ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) { 1582 sc->sc_flags &= ~SCF_ECN; 1583 } 1584 #ifdef MAC 1585 /* 1586 * Since we have already unconditionally allocated label 1587 * storage, free it up. The syncache entry will already 1588 * have an initialized label we can use. 1589 */ 1590 mac_syncache_destroy(&maclabel); 1591 #endif 1592 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1593 /* Retransmit SYN|ACK and reset retransmit count. */ 1594 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1595 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1596 "resetting timer and retransmitting SYN|ACK\n", 1597 s, __func__); 1598 free(s, M_TCPLOG); 1599 } 1600 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1601 sc->sc_rxmits = 0; 1602 syncache_timeout(sc, sch, 1); 1603 TCPSTAT_INC(tcps_sndacks); 1604 TCPSTAT_INC(tcps_sndtotal); 1605 } 1606 SCH_UNLOCK(sch); 1607 goto donenoprobe; 1608 } 1609 1610 if (tfo_cookie_valid) { 1611 bzero(&scs, sizeof(scs)); 1612 sc = &scs; 1613 goto skip_alloc; 1614 } 1615 1616 /* 1617 * Skip allocating a syncache entry if we are just going to discard 1618 * it later. 1619 */ 1620 if (!locked) { 1621 bzero(&scs, sizeof(scs)); 1622 sc = &scs; 1623 } else 1624 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1625 if (sc == NULL) { 1626 /* 1627 * The zone allocator couldn't provide more entries. 1628 * Treat this as if the cache was full; drop the oldest 1629 * entry and insert the new one. 1630 */ 1631 TCPSTAT_INC(tcps_sc_zonefail); 1632 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1633 sch->sch_last_overflow = time_uptime; 1634 syncache_drop(sc, sch); 1635 syncache_pause(inc); 1636 } 1637 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1638 if (sc == NULL) { 1639 if (V_tcp_syncookies) { 1640 bzero(&scs, sizeof(scs)); 1641 sc = &scs; 1642 } else { 1643 KASSERT(locked, 1644 ("%s: bucket unexpectedly unlocked", 1645 __func__)); 1646 SCH_UNLOCK(sch); 1647 if (ipopts) 1648 (void) m_free(ipopts); 1649 goto done; 1650 } 1651 } 1652 } 1653 1654 skip_alloc: 1655 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1656 sc->sc_tfo_cookie = &tfo_response_cookie; 1657 1658 /* 1659 * Fill in the syncache values. 1660 */ 1661 #ifdef MAC 1662 sc->sc_label = maclabel; 1663 #endif 1664 sc->sc_cred = cred; 1665 sc->sc_port = port; 1666 cred = NULL; 1667 sc->sc_ipopts = ipopts; 1668 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1669 sc->sc_ip_tos = ip_tos; 1670 sc->sc_ip_ttl = ip_ttl; 1671 #ifdef TCP_OFFLOAD 1672 sc->sc_tod = tod; 1673 sc->sc_todctx = todctx; 1674 #endif 1675 sc->sc_irs = th->th_seq; 1676 sc->sc_flags = 0; 1677 sc->sc_flowlabel = 0; 1678 1679 /* 1680 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1681 * win was derived from socket earlier in the function. 1682 */ 1683 win = imax(win, 0); 1684 win = imin(win, TCP_MAXWIN); 1685 sc->sc_wnd = win; 1686 1687 if (V_tcp_do_rfc1323 && 1688 !(ltflags & TF_NOOPT)) { 1689 /* 1690 * A timestamp received in a SYN makes 1691 * it ok to send timestamp requests and replies. 1692 */ 1693 if (to->to_flags & TOF_TS) { 1694 sc->sc_tsreflect = to->to_tsval; 1695 sc->sc_flags |= SCF_TIMESTAMP; 1696 sc->sc_tsoff = tcp_new_ts_offset(inc); 1697 } 1698 if (to->to_flags & TOF_SCALE) { 1699 int wscale = 0; 1700 1701 /* 1702 * Pick the smallest possible scaling factor that 1703 * will still allow us to scale up to sb_max, aka 1704 * kern.ipc.maxsockbuf. 1705 * 1706 * We do this because there are broken firewalls that 1707 * will corrupt the window scale option, leading to 1708 * the other endpoint believing that our advertised 1709 * window is unscaled. At scale factors larger than 1710 * 5 the unscaled window will drop below 1500 bytes, 1711 * leading to serious problems when traversing these 1712 * broken firewalls. 1713 * 1714 * With the default maxsockbuf of 256K, a scale factor 1715 * of 3 will be chosen by this algorithm. Those who 1716 * choose a larger maxsockbuf should watch out 1717 * for the compatibility problems mentioned above. 1718 * 1719 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1720 * or <SYN,ACK>) segment itself is never scaled. 1721 */ 1722 while (wscale < TCP_MAX_WINSHIFT && 1723 (TCP_MAXWIN << wscale) < sb_max) 1724 wscale++; 1725 sc->sc_requested_r_scale = wscale; 1726 sc->sc_requested_s_scale = to->to_wscale; 1727 sc->sc_flags |= SCF_WINSCALE; 1728 } 1729 } 1730 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1731 /* 1732 * If listening socket requested TCP digests, flag this in the 1733 * syncache so that syncache_respond() will do the right thing 1734 * with the SYN+ACK. 1735 */ 1736 if (ltflags & TF_SIGNATURE) 1737 sc->sc_flags |= SCF_SIGNATURE; 1738 #endif /* TCP_SIGNATURE */ 1739 if (to->to_flags & TOF_SACKPERM) 1740 sc->sc_flags |= SCF_SACK; 1741 if (to->to_flags & TOF_MSS) 1742 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1743 if (ltflags & TF_NOOPT) 1744 sc->sc_flags |= SCF_NOOPT; 1745 if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) && 1746 V_tcp_do_ecn) 1747 sc->sc_flags |= SCF_ECN; 1748 1749 if (V_tcp_syncookies) 1750 sc->sc_iss = syncookie_generate(sch, sc); 1751 else 1752 sc->sc_iss = arc4random(); 1753 #ifdef INET6 1754 if (autoflowlabel) { 1755 if (V_tcp_syncookies) 1756 sc->sc_flowlabel = sc->sc_iss; 1757 else 1758 sc->sc_flowlabel = ip6_randomflowlabel(); 1759 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1760 } 1761 #endif 1762 if (locked) 1763 SCH_UNLOCK(sch); 1764 1765 if (tfo_cookie_valid) { 1766 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie); 1767 /* INP_RUNLOCK(inp) will be performed by the caller */ 1768 goto tfo_expanded; 1769 } 1770 1771 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1772 /* 1773 * Do a standard 3-way handshake. 1774 */ 1775 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1776 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1777 syncache_free(sc); 1778 else if (sc != &scs) 1779 syncache_insert(sc, sch); /* locks and unlocks sch */ 1780 TCPSTAT_INC(tcps_sndacks); 1781 TCPSTAT_INC(tcps_sndtotal); 1782 } else { 1783 if (sc != &scs) 1784 syncache_free(sc); 1785 TCPSTAT_INC(tcps_sc_dropped); 1786 } 1787 goto donenoprobe; 1788 1789 done: 1790 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1791 donenoprobe: 1792 if (m) 1793 m_freem(m); 1794 /* 1795 * If tfo_pending is not NULL here, then a TFO SYN that did not 1796 * result in a new socket was processed and the associated pending 1797 * counter has not yet been decremented. All such TFO processing paths 1798 * transit this point. 1799 */ 1800 if (tfo_pending != NULL) 1801 tcp_fastopen_decrement_counter(tfo_pending); 1802 1803 tfo_expanded: 1804 if (cred != NULL) 1805 crfree(cred); 1806 #ifdef MAC 1807 if (sc == &scs) 1808 mac_syncache_destroy(&maclabel); 1809 #endif 1810 return (rv); 1811 } 1812 1813 /* 1814 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1815 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1816 */ 1817 static int 1818 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags) 1819 { 1820 struct ip *ip = NULL; 1821 struct mbuf *m; 1822 struct tcphdr *th = NULL; 1823 struct udphdr *udp = NULL; 1824 int optlen, error = 0; /* Make compiler happy */ 1825 u_int16_t hlen, tlen, mssopt, ulen; 1826 struct tcpopt to; 1827 #ifdef INET6 1828 struct ip6_hdr *ip6 = NULL; 1829 #endif 1830 1831 NET_EPOCH_ASSERT(); 1832 1833 hlen = 1834 #ifdef INET6 1835 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1836 #endif 1837 sizeof(struct ip); 1838 tlen = hlen + sizeof(struct tcphdr); 1839 if (sc->sc_port) { 1840 tlen += sizeof(struct udphdr); 1841 } 1842 /* Determine MSS we advertize to other end of connection. */ 1843 mssopt = tcp_mssopt(&sc->sc_inc); 1844 if (sc->sc_port) 1845 mssopt -= V_tcp_udp_tunneling_overhead; 1846 mssopt = max(mssopt, V_tcp_minmss); 1847 1848 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1849 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1850 ("syncache: mbuf too small")); 1851 1852 /* Create the IP+TCP header from scratch. */ 1853 m = m_gethdr(M_NOWAIT, MT_DATA); 1854 if (m == NULL) 1855 return (ENOBUFS); 1856 #ifdef MAC 1857 mac_syncache_create_mbuf(sc->sc_label, m); 1858 #endif 1859 m->m_data += max_linkhdr; 1860 m->m_len = tlen; 1861 m->m_pkthdr.len = tlen; 1862 m->m_pkthdr.rcvif = NULL; 1863 1864 #ifdef INET6 1865 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1866 ip6 = mtod(m, struct ip6_hdr *); 1867 ip6->ip6_vfc = IPV6_VERSION; 1868 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1869 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1870 ip6->ip6_plen = htons(tlen - hlen); 1871 /* ip6_hlim is set after checksum */ 1872 /* Zero out traffic class and flow label. */ 1873 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1874 ip6->ip6_flow |= sc->sc_flowlabel; 1875 if (sc->sc_port != 0) { 1876 ip6->ip6_nxt = IPPROTO_UDP; 1877 udp = (struct udphdr *)(ip6 + 1); 1878 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1879 udp->uh_dport = sc->sc_port; 1880 ulen = (tlen - sizeof(struct ip6_hdr)); 1881 th = (struct tcphdr *)(udp + 1); 1882 } else { 1883 ip6->ip6_nxt = IPPROTO_TCP; 1884 th = (struct tcphdr *)(ip6 + 1); 1885 } 1886 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20); 1887 } 1888 #endif 1889 #if defined(INET6) && defined(INET) 1890 else 1891 #endif 1892 #ifdef INET 1893 { 1894 ip = mtod(m, struct ip *); 1895 ip->ip_v = IPVERSION; 1896 ip->ip_hl = sizeof(struct ip) >> 2; 1897 ip->ip_len = htons(tlen); 1898 ip->ip_id = 0; 1899 ip->ip_off = 0; 1900 ip->ip_sum = 0; 1901 ip->ip_src = sc->sc_inc.inc_laddr; 1902 ip->ip_dst = sc->sc_inc.inc_faddr; 1903 ip->ip_ttl = sc->sc_ip_ttl; 1904 ip->ip_tos = sc->sc_ip_tos; 1905 1906 /* 1907 * See if we should do MTU discovery. Route lookups are 1908 * expensive, so we will only unset the DF bit if: 1909 * 1910 * 1) path_mtu_discovery is disabled 1911 * 2) the SCF_UNREACH flag has been set 1912 */ 1913 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1914 ip->ip_off |= htons(IP_DF); 1915 if (sc->sc_port == 0) { 1916 ip->ip_p = IPPROTO_TCP; 1917 th = (struct tcphdr *)(ip + 1); 1918 } else { 1919 ip->ip_p = IPPROTO_UDP; 1920 udp = (struct udphdr *)(ip + 1); 1921 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1922 udp->uh_dport = sc->sc_port; 1923 ulen = (tlen - sizeof(struct ip)); 1924 th = (struct tcphdr *)(udp + 1); 1925 } 1926 } 1927 #endif /* INET */ 1928 th->th_sport = sc->sc_inc.inc_lport; 1929 th->th_dport = sc->sc_inc.inc_fport; 1930 1931 if (flags & TH_SYN) 1932 th->th_seq = htonl(sc->sc_iss); 1933 else 1934 th->th_seq = htonl(sc->sc_iss + 1); 1935 th->th_ack = htonl(sc->sc_irs + 1); 1936 th->th_off = sizeof(struct tcphdr) >> 2; 1937 th->th_x2 = 0; 1938 th->th_flags = flags; 1939 th->th_win = htons(sc->sc_wnd); 1940 th->th_urp = 0; 1941 1942 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) { 1943 th->th_flags |= TH_ECE; 1944 TCPSTAT_INC(tcps_ecn_shs); 1945 } 1946 1947 /* Tack on the TCP options. */ 1948 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1949 to.to_flags = 0; 1950 1951 if (flags & TH_SYN) { 1952 to.to_mss = mssopt; 1953 to.to_flags = TOF_MSS; 1954 if (sc->sc_flags & SCF_WINSCALE) { 1955 to.to_wscale = sc->sc_requested_r_scale; 1956 to.to_flags |= TOF_SCALE; 1957 } 1958 if (sc->sc_flags & SCF_SACK) 1959 to.to_flags |= TOF_SACKPERM; 1960 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1961 if (sc->sc_flags & SCF_SIGNATURE) 1962 to.to_flags |= TOF_SIGNATURE; 1963 #endif 1964 if (sc->sc_tfo_cookie) { 1965 to.to_flags |= TOF_FASTOPEN; 1966 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1967 to.to_tfo_cookie = sc->sc_tfo_cookie; 1968 /* don't send cookie again when retransmitting response */ 1969 sc->sc_tfo_cookie = NULL; 1970 } 1971 } 1972 if (sc->sc_flags & SCF_TIMESTAMP) { 1973 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1974 to.to_tsecr = sc->sc_tsreflect; 1975 to.to_flags |= TOF_TS; 1976 } 1977 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1978 1979 /* Adjust headers by option size. */ 1980 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1981 m->m_len += optlen; 1982 m->m_pkthdr.len += optlen; 1983 #ifdef INET6 1984 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1985 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1986 else 1987 #endif 1988 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1989 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1990 if (sc->sc_flags & SCF_SIGNATURE) { 1991 KASSERT(to.to_flags & TOF_SIGNATURE, 1992 ("tcp_addoptions() didn't set tcp_signature")); 1993 1994 /* NOTE: to.to_signature is inside of mbuf */ 1995 if (!TCPMD5_ENABLED() || 1996 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1997 m_freem(m); 1998 return (EACCES); 1999 } 2000 } 2001 #endif 2002 } else 2003 optlen = 0; 2004 2005 if (udp) { 2006 ulen += optlen; 2007 udp->uh_ulen = htons(ulen); 2008 } 2009 M_SETFIB(m, sc->sc_inc.inc_fibnum); 2010 /* 2011 * If we have peer's SYN and it has a flowid, then let's assign it to 2012 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 2013 * to SYN|ACK due to lack of inp here. 2014 */ 2015 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 2016 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 2017 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 2018 } 2019 #ifdef INET6 2020 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 2021 if (sc->sc_port) { 2022 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 2023 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2024 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, 2025 IPPROTO_UDP, 0); 2026 th->th_sum = htons(0); 2027 } else { 2028 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 2029 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2030 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 2031 IPPROTO_TCP, 0); 2032 } 2033 ip6->ip6_hlim = sc->sc_ip_ttl; 2034 #ifdef TCP_OFFLOAD 2035 if (ADDED_BY_TOE(sc)) { 2036 struct toedev *tod = sc->sc_tod; 2037 2038 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2039 2040 return (error); 2041 } 2042 #endif 2043 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 2044 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 2045 } 2046 #endif 2047 #if defined(INET6) && defined(INET) 2048 else 2049 #endif 2050 #ifdef INET 2051 { 2052 if (sc->sc_port) { 2053 m->m_pkthdr.csum_flags = CSUM_UDP; 2054 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2055 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 2056 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 2057 th->th_sum = htons(0); 2058 } else { 2059 m->m_pkthdr.csum_flags = CSUM_TCP; 2060 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2061 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 2062 htons(tlen + optlen - hlen + IPPROTO_TCP)); 2063 } 2064 #ifdef TCP_OFFLOAD 2065 if (ADDED_BY_TOE(sc)) { 2066 struct toedev *tod = sc->sc_tod; 2067 2068 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2069 2070 return (error); 2071 } 2072 #endif 2073 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 2074 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 2075 } 2076 #endif 2077 return (error); 2078 } 2079 2080 /* 2081 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 2082 * that exceed the capacity of the syncache by avoiding the storage of any 2083 * of the SYNs we receive. Syncookies defend against blind SYN flooding 2084 * attacks where the attacker does not have access to our responses. 2085 * 2086 * Syncookies encode and include all necessary information about the 2087 * connection setup within the SYN|ACK that we send back. That way we 2088 * can avoid keeping any local state until the ACK to our SYN|ACK returns 2089 * (if ever). Normally the syncache and syncookies are running in parallel 2090 * with the latter taking over when the former is exhausted. When matching 2091 * syncache entry is found the syncookie is ignored. 2092 * 2093 * The only reliable information persisting the 3WHS is our initial sequence 2094 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 2095 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 2096 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 2097 * returns and signifies a legitimate connection if it matches the ACK. 2098 * 2099 * The available space of 32 bits to store the hash and to encode the SYN 2100 * option information is very tight and we should have at least 24 bits for 2101 * the MAC to keep the number of guesses by blind spoofing reasonably high. 2102 * 2103 * SYN option information we have to encode to fully restore a connection: 2104 * MSS: is imporant to chose an optimal segment size to avoid IP level 2105 * fragmentation along the path. The common MSS values can be encoded 2106 * in a 3-bit table. Uncommon values are captured by the next lower value 2107 * in the table leading to a slight increase in packetization overhead. 2108 * WSCALE: is necessary to allow large windows to be used for high delay- 2109 * bandwidth product links. Not scaling the window when it was initially 2110 * negotiated is bad for performance as lack of scaling further decreases 2111 * the apparent available send window. We only need to encode the WSCALE 2112 * we received from the remote end. Our end can be recalculated at any 2113 * time. The common WSCALE values can be encoded in a 3-bit table. 2114 * Uncommon values are captured by the next lower value in the table 2115 * making us under-estimate the available window size halving our 2116 * theoretically possible maximum throughput for that connection. 2117 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 2118 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 2119 * that are included in all segments on a connection. We enable them when 2120 * the ACK has them. 2121 * 2122 * Security of syncookies and attack vectors: 2123 * 2124 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 2125 * together with the gloabl secret to make it unique per connection attempt. 2126 * Thus any change of any of those parameters results in a different MAC output 2127 * in an unpredictable way unless a collision is encountered. 24 bits of the 2128 * MAC are embedded into the ISS. 2129 * 2130 * To prevent replay attacks two rotating global secrets are updated with a 2131 * new random value every 15 seconds. The life-time of a syncookie is thus 2132 * 15-30 seconds. 2133 * 2134 * Vector 1: Attacking the secret. This requires finding a weakness in the 2135 * MAC itself or the way it is used here. The attacker can do a chosen plain 2136 * text attack by varying and testing the all parameters under his control. 2137 * The strength depends on the size and randomness of the secret, and the 2138 * cryptographic security of the MAC function. Due to the constant updating 2139 * of the secret the attacker has at most 29.999 seconds to find the secret 2140 * and launch spoofed connections. After that he has to start all over again. 2141 * 2142 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 2143 * size an average of 4,823 attempts are required for a 50% chance of success 2144 * to spoof a single syncookie (birthday collision paradox). However the 2145 * attacker is blind and doesn't know if one of his attempts succeeded unless 2146 * he has a side channel to interfere success from. A single connection setup 2147 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 2148 * This many attempts are required for each one blind spoofed connection. For 2149 * every additional spoofed connection he has to launch another N attempts. 2150 * Thus for a sustained rate 100 spoofed connections per second approximately 2151 * 1,800,000 packets per second would have to be sent. 2152 * 2153 * NB: The MAC function should be fast so that it doesn't become a CPU 2154 * exhaustion attack vector itself. 2155 * 2156 * References: 2157 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 2158 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 2159 * http://cr.yp.to/syncookies.html (overview) 2160 * http://cr.yp.to/syncookies/archive (details) 2161 * 2162 * 2163 * Schematic construction of a syncookie enabled Initial Sequence Number: 2164 * 0 1 2 3 2165 * 12345678901234567890123456789012 2166 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2167 * 2168 * x 24 MAC (truncated) 2169 * W 3 Send Window Scale index 2170 * M 3 MSS index 2171 * S 1 SACK permitted 2172 * P 1 Odd/even secret 2173 */ 2174 2175 /* 2176 * Distribution and probability of certain MSS values. Those in between are 2177 * rounded down to the next lower one. 2178 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2179 * .2% .3% 5% 7% 7% 20% 15% 45% 2180 */ 2181 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2182 2183 /* 2184 * Distribution and probability of certain WSCALE values. We have to map the 2185 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2186 * bits based on prevalence of certain values. Where we don't have an exact 2187 * match for are rounded down to the next lower one letting us under-estimate 2188 * the true available window. At the moment this would happen only for the 2189 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2190 * and window size). The absence of the WSCALE option (no scaling in either 2191 * direction) is encoded with index zero. 2192 * [WSCALE values histograms, Allman, 2012] 2193 * X 10 10 35 5 6 14 10% by host 2194 * X 11 4 5 5 18 49 3% by connections 2195 */ 2196 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2197 2198 /* 2199 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2200 * and good cryptographic properties. 2201 */ 2202 static uint32_t 2203 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2204 uint8_t *secbits, uintptr_t secmod) 2205 { 2206 SIPHASH_CTX ctx; 2207 uint32_t siphash[2]; 2208 2209 SipHash24_Init(&ctx); 2210 SipHash_SetKey(&ctx, secbits); 2211 switch (inc->inc_flags & INC_ISIPV6) { 2212 #ifdef INET 2213 case 0: 2214 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2215 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2216 break; 2217 #endif 2218 #ifdef INET6 2219 case INC_ISIPV6: 2220 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2221 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2222 break; 2223 #endif 2224 } 2225 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2226 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2227 SipHash_Update(&ctx, &irs, sizeof(irs)); 2228 SipHash_Update(&ctx, &flags, sizeof(flags)); 2229 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2230 SipHash_Final((u_int8_t *)&siphash, &ctx); 2231 2232 return (siphash[0] ^ siphash[1]); 2233 } 2234 2235 static tcp_seq 2236 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2237 { 2238 u_int i, secbit, wscale; 2239 uint32_t iss, hash; 2240 uint8_t *secbits; 2241 union syncookie cookie; 2242 2243 cookie.cookie = 0; 2244 2245 /* Map our computed MSS into the 3-bit index. */ 2246 for (i = nitems(tcp_sc_msstab) - 1; 2247 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2248 i--) 2249 ; 2250 cookie.flags.mss_idx = i; 2251 2252 /* 2253 * Map the send window scale into the 3-bit index but only if 2254 * the wscale option was received. 2255 */ 2256 if (sc->sc_flags & SCF_WINSCALE) { 2257 wscale = sc->sc_requested_s_scale; 2258 for (i = nitems(tcp_sc_wstab) - 1; 2259 tcp_sc_wstab[i] > wscale && i > 0; 2260 i--) 2261 ; 2262 cookie.flags.wscale_idx = i; 2263 } 2264 2265 /* Can we do SACK? */ 2266 if (sc->sc_flags & SCF_SACK) 2267 cookie.flags.sack_ok = 1; 2268 2269 /* Which of the two secrets to use. */ 2270 secbit = V_tcp_syncache.secret.oddeven & 0x1; 2271 cookie.flags.odd_even = secbit; 2272 2273 secbits = V_tcp_syncache.secret.key[secbit]; 2274 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2275 (uintptr_t)sch); 2276 2277 /* 2278 * Put the flags into the hash and XOR them to get better ISS number 2279 * variance. This doesn't enhance the cryptographic strength and is 2280 * done to prevent the 8 cookie bits from showing up directly on the 2281 * wire. 2282 */ 2283 iss = hash & ~0xff; 2284 iss |= cookie.cookie ^ (hash >> 24); 2285 2286 TCPSTAT_INC(tcps_sc_sendcookie); 2287 return (iss); 2288 } 2289 2290 static struct syncache * 2291 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2292 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2293 struct socket *lso, uint16_t port) 2294 { 2295 uint32_t hash; 2296 uint8_t *secbits; 2297 tcp_seq ack, seq; 2298 int wnd, wscale = 0; 2299 union syncookie cookie; 2300 2301 /* 2302 * Pull information out of SYN-ACK/ACK and revert sequence number 2303 * advances. 2304 */ 2305 ack = th->th_ack - 1; 2306 seq = th->th_seq - 1; 2307 2308 /* 2309 * Unpack the flags containing enough information to restore the 2310 * connection. 2311 */ 2312 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2313 2314 /* Which of the two secrets to use. */ 2315 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; 2316 2317 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2318 2319 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2320 if ((ack & ~0xff) != (hash & ~0xff)) 2321 return (NULL); 2322 2323 /* Fill in the syncache values. */ 2324 sc->sc_flags = 0; 2325 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2326 sc->sc_ipopts = NULL; 2327 2328 sc->sc_irs = seq; 2329 sc->sc_iss = ack; 2330 2331 switch (inc->inc_flags & INC_ISIPV6) { 2332 #ifdef INET 2333 case 0: 2334 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2335 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2336 break; 2337 #endif 2338 #ifdef INET6 2339 case INC_ISIPV6: 2340 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2341 sc->sc_flowlabel = 2342 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2343 break; 2344 #endif 2345 } 2346 2347 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2348 2349 /* We can simply recompute receive window scale we sent earlier. */ 2350 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2351 wscale++; 2352 2353 /* Only use wscale if it was enabled in the orignal SYN. */ 2354 if (cookie.flags.wscale_idx > 0) { 2355 sc->sc_requested_r_scale = wscale; 2356 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2357 sc->sc_flags |= SCF_WINSCALE; 2358 } 2359 2360 wnd = lso->sol_sbrcv_hiwat; 2361 wnd = imax(wnd, 0); 2362 wnd = imin(wnd, TCP_MAXWIN); 2363 sc->sc_wnd = wnd; 2364 2365 if (cookie.flags.sack_ok) 2366 sc->sc_flags |= SCF_SACK; 2367 2368 if (to->to_flags & TOF_TS) { 2369 sc->sc_flags |= SCF_TIMESTAMP; 2370 sc->sc_tsreflect = to->to_tsval; 2371 sc->sc_tsoff = tcp_new_ts_offset(inc); 2372 } 2373 2374 if (to->to_flags & TOF_SIGNATURE) 2375 sc->sc_flags |= SCF_SIGNATURE; 2376 2377 sc->sc_rxmits = 0; 2378 2379 sc->sc_port = port; 2380 2381 TCPSTAT_INC(tcps_sc_recvcookie); 2382 return (sc); 2383 } 2384 2385 #ifdef INVARIANTS 2386 static int 2387 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2388 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2389 struct socket *lso, uint16_t port) 2390 { 2391 struct syncache scs, *scx; 2392 char *s; 2393 2394 bzero(&scs, sizeof(scs)); 2395 scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port); 2396 2397 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2398 return (0); 2399 2400 if (scx != NULL) { 2401 if (sc->sc_peer_mss != scx->sc_peer_mss) 2402 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2403 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2404 2405 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2406 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2407 s, __func__, sc->sc_requested_r_scale, 2408 scx->sc_requested_r_scale); 2409 2410 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2411 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2412 s, __func__, sc->sc_requested_s_scale, 2413 scx->sc_requested_s_scale); 2414 2415 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2416 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2417 } 2418 2419 if (s != NULL) 2420 free(s, M_TCPLOG); 2421 return (0); 2422 } 2423 #endif /* INVARIANTS */ 2424 2425 static void 2426 syncookie_reseed(void *arg) 2427 { 2428 struct tcp_syncache *sc = arg; 2429 uint8_t *secbits; 2430 int secbit; 2431 2432 /* 2433 * Reseeding the secret doesn't have to be protected by a lock. 2434 * It only must be ensured that the new random values are visible 2435 * to all CPUs in a SMP environment. The atomic with release 2436 * semantics ensures that. 2437 */ 2438 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2439 secbits = sc->secret.key[secbit]; 2440 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2441 atomic_add_rel_int(&sc->secret.oddeven, 1); 2442 2443 /* Reschedule ourself. */ 2444 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2445 } 2446 2447 /* 2448 * We have overflowed a bucket. Let's pause dealing with the syncache. 2449 * This function will increment the bucketoverflow statistics appropriately 2450 * (once per pause when pausing is enabled; otherwise, once per overflow). 2451 */ 2452 static void 2453 syncache_pause(struct in_conninfo *inc) 2454 { 2455 time_t delta; 2456 const char *s; 2457 2458 /* XXX: 2459 * 2. Add sysctl read here so we don't get the benefit of this 2460 * change without the new sysctl. 2461 */ 2462 2463 /* 2464 * Try an unlocked read. If we already know that another thread 2465 * has activated the feature, there is no need to proceed. 2466 */ 2467 if (V_tcp_syncache.paused) 2468 return; 2469 2470 /* Are cookied enabled? If not, we can't pause. */ 2471 if (!V_tcp_syncookies) { 2472 TCPSTAT_INC(tcps_sc_bucketoverflow); 2473 return; 2474 } 2475 2476 /* 2477 * We may be the first thread to find an overflow. Get the lock 2478 * and evaluate if we need to take action. 2479 */ 2480 mtx_lock(&V_tcp_syncache.pause_mtx); 2481 if (V_tcp_syncache.paused) { 2482 mtx_unlock(&V_tcp_syncache.pause_mtx); 2483 return; 2484 } 2485 2486 /* Activate protection. */ 2487 V_tcp_syncache.paused = true; 2488 TCPSTAT_INC(tcps_sc_bucketoverflow); 2489 2490 /* 2491 * Determine the last backoff time. If we are seeing a re-newed 2492 * attack within that same time after last reactivating the syncache, 2493 * consider it an extension of the same attack. 2494 */ 2495 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; 2496 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { 2497 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { 2498 delta <<= 1; 2499 V_tcp_syncache.pause_backoff++; 2500 } 2501 } else { 2502 delta = TCP_SYNCACHE_PAUSE_TIME; 2503 V_tcp_syncache.pause_backoff = 0; 2504 } 2505 2506 /* Log a warning, including IP addresses, if able. */ 2507 if (inc != NULL) 2508 s = tcp_log_addrs(inc, NULL, NULL, NULL); 2509 else 2510 s = (const char *)NULL; 2511 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " 2512 "the next %lld seconds%s%s%s\n", (long long)delta, 2513 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", 2514 (s != NULL) ? ")" : ""); 2515 free(__DECONST(void *, s), M_TCPLOG); 2516 2517 /* Use the calculated delta to set a new pause time. */ 2518 V_tcp_syncache.pause_until = time_uptime + delta; 2519 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, 2520 &V_tcp_syncache); 2521 mtx_unlock(&V_tcp_syncache.pause_mtx); 2522 } 2523 2524 /* Evaluate whether we need to unpause. */ 2525 static void 2526 syncache_unpause(void *arg) 2527 { 2528 struct tcp_syncache *sc; 2529 time_t delta; 2530 2531 sc = arg; 2532 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); 2533 callout_deactivate(&sc->pause_co); 2534 2535 /* 2536 * Check to make sure we are not running early. If the pause 2537 * time has expired, then deactivate the protection. 2538 */ 2539 if ((delta = sc->pause_until - time_uptime) > 0) 2540 callout_schedule(&sc->pause_co, delta * hz); 2541 else 2542 sc->paused = false; 2543 } 2544 2545 /* 2546 * Exports the syncache entries to userland so that netstat can display 2547 * them alongside the other sockets. This function is intended to be 2548 * called only from tcp_pcblist. 2549 * 2550 * Due to concurrency on an active system, the number of pcbs exported 2551 * may have no relation to max_pcbs. max_pcbs merely indicates the 2552 * amount of space the caller allocated for this function to use. 2553 */ 2554 int 2555 syncache_pcblist(struct sysctl_req *req) 2556 { 2557 struct xtcpcb xt; 2558 struct syncache *sc; 2559 struct syncache_head *sch; 2560 int error, i; 2561 2562 bzero(&xt, sizeof(xt)); 2563 xt.xt_len = sizeof(xt); 2564 xt.t_state = TCPS_SYN_RECEIVED; 2565 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2566 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2567 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2568 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2569 2570 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 2571 sch = &V_tcp_syncache.hashbase[i]; 2572 SCH_LOCK(sch); 2573 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2574 if (sc->sc_cred != NULL && 2575 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2576 continue; 2577 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2578 xt.xt_inp.inp_vflag = INP_IPV6; 2579 else 2580 xt.xt_inp.inp_vflag = INP_IPV4; 2581 xt.xt_encaps_port = sc->sc_port; 2582 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2583 sizeof (struct in_conninfo)); 2584 error = SYSCTL_OUT(req, &xt, sizeof xt); 2585 if (error) { 2586 SCH_UNLOCK(sch); 2587 return (0); 2588 } 2589 } 2590 SCH_UNLOCK(sch); 2591 } 2592 2593 return (0); 2594 } 2595