1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/hash.h> 43 #include <sys/refcount.h> 44 #include <sys/kernel.h> 45 #include <sys/sysctl.h> 46 #include <sys/limits.h> 47 #include <sys/lock.h> 48 #include <sys/mutex.h> 49 #include <sys/malloc.h> 50 #include <sys/mbuf.h> 51 #include <sys/proc.h> /* for proc0 declaration */ 52 #include <sys/random.h> 53 #include <sys/socket.h> 54 #include <sys/socketvar.h> 55 #include <sys/syslog.h> 56 #include <sys/ucred.h> 57 58 #include <sys/md5.h> 59 #include <crypto/siphash/siphash.h> 60 61 #include <vm/uma.h> 62 63 #include <net/if.h> 64 #include <net/if_var.h> 65 #include <net/route.h> 66 #include <net/vnet.h> 67 68 #include <netinet/in.h> 69 #include <netinet/in_kdtrace.h> 70 #include <netinet/in_systm.h> 71 #include <netinet/ip.h> 72 #include <netinet/in_var.h> 73 #include <netinet/in_pcb.h> 74 #include <netinet/ip_var.h> 75 #include <netinet/ip_options.h> 76 #ifdef INET6 77 #include <netinet/ip6.h> 78 #include <netinet/icmp6.h> 79 #include <netinet6/nd6.h> 80 #include <netinet6/ip6_var.h> 81 #include <netinet6/in6_pcb.h> 82 #endif 83 #include <netinet/tcp.h> 84 #include <netinet/tcp_fastopen.h> 85 #include <netinet/tcp_fsm.h> 86 #include <netinet/tcp_seq.h> 87 #include <netinet/tcp_timer.h> 88 #include <netinet/tcp_var.h> 89 #include <netinet/tcp_syncache.h> 90 #include <netinet/tcp_ecn.h> 91 #ifdef TCP_BLACKBOX 92 #include <netinet/tcp_log_buf.h> 93 #endif 94 #ifdef TCP_OFFLOAD 95 #include <netinet/toecore.h> 96 #endif 97 #include <netinet/udp.h> 98 99 #include <netipsec/ipsec_support.h> 100 101 #include <machine/in_cksum.h> 102 103 #include <security/mac/mac_framework.h> 104 105 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1; 106 #define V_tcp_syncookies VNET(tcp_syncookies) 107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 108 &VNET_NAME(tcp_syncookies), 0, 109 "Use TCP SYN cookies if the syncache overflows"); 110 111 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0; 112 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 114 &VNET_NAME(tcp_syncookiesonly), 0, 115 "Use only TCP SYN cookies"); 116 117 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1; 118 #define V_functions_inherit_listen_socket_stack \ 119 VNET(functions_inherit_listen_socket_stack) 120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack, 121 CTLFLAG_VNET | CTLFLAG_RW, 122 &VNET_NAME(functions_inherit_listen_socket_stack), 0, 123 "Inherit listen socket's stack"); 124 125 #ifdef TCP_OFFLOAD 126 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 127 #endif 128 129 static void syncache_drop(struct syncache *, struct syncache_head *); 130 static void syncache_free(struct syncache *); 131 static void syncache_insert(struct syncache *, struct syncache_head *); 132 static int syncache_respond(struct syncache *, const struct mbuf *, int); 133 static struct socket *syncache_socket(struct syncache *, struct socket *, 134 struct mbuf *m); 135 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 136 int docallout); 137 static void syncache_timer(void *); 138 139 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 140 uint8_t *, uintptr_t); 141 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 142 static struct syncache 143 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 144 struct syncache *, struct tcphdr *, struct tcpopt *, 145 struct socket *, uint16_t); 146 static void syncache_pause(struct in_conninfo *); 147 static void syncache_unpause(void *); 148 static void syncookie_reseed(void *); 149 #ifdef INVARIANTS 150 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 151 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 152 struct socket *lso, uint16_t port); 153 #endif 154 155 /* 156 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 157 * 3 retransmits corresponds to a timeout with default values of 158 * tcp_rexmit_initial * ( 1 + 159 * tcp_backoff[1] + 160 * tcp_backoff[2] + 161 * tcp_backoff[3]) + 3 * tcp_rexmit_slop, 162 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms, 163 * the odds are that the user has given up attempting to connect by then. 164 */ 165 #define SYNCACHE_MAXREXMTS 3 166 167 /* Arbitrary values */ 168 #define TCP_SYNCACHE_HASHSIZE 512 169 #define TCP_SYNCACHE_BUCKETLIMIT 30 170 171 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 172 #define V_tcp_syncache VNET(tcp_syncache) 173 174 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, 175 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 176 "TCP SYN cache"); 177 178 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 179 &VNET_NAME(tcp_syncache.bucket_limit), 0, 180 "Per-bucket hash limit for syncache"); 181 182 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 183 &VNET_NAME(tcp_syncache.cache_limit), 0, 184 "Overall entry limit for syncache"); 185 186 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 187 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 188 189 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 190 &VNET_NAME(tcp_syncache.hashsize), 0, 191 "Size of TCP syncache hashtable"); 192 193 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET | 194 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0, 195 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) " 196 "and mac(4) checks"); 197 198 static int 199 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 200 { 201 int error; 202 u_int new; 203 204 new = V_tcp_syncache.rexmt_limit; 205 error = sysctl_handle_int(oidp, &new, 0, req); 206 if ((error == 0) && (req->newptr != NULL)) { 207 if (new > TCP_MAXRXTSHIFT) 208 error = EINVAL; 209 else 210 V_tcp_syncache.rexmt_limit = new; 211 } 212 return (error); 213 } 214 215 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 216 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 217 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 218 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "IU", 219 "Limit on SYN/ACK retransmissions"); 220 221 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 222 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 223 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 224 "Send reset on socket allocation failure"); 225 226 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 227 228 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 229 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 230 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 231 232 /* 233 * Requires the syncache entry to be already removed from the bucket list. 234 */ 235 static void 236 syncache_free(struct syncache *sc) 237 { 238 239 if (sc->sc_ipopts) 240 (void) m_free(sc->sc_ipopts); 241 if (sc->sc_cred) 242 crfree(sc->sc_cred); 243 #ifdef MAC 244 mac_syncache_destroy(&sc->sc_label); 245 #endif 246 247 uma_zfree(V_tcp_syncache.zone, sc); 248 } 249 250 void 251 syncache_init(void) 252 { 253 int i; 254 255 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 256 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 257 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 258 V_tcp_syncache.hash_secret = arc4random(); 259 260 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 261 &V_tcp_syncache.hashsize); 262 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 263 &V_tcp_syncache.bucket_limit); 264 if (!powerof2(V_tcp_syncache.hashsize) || 265 V_tcp_syncache.hashsize == 0) { 266 printf("WARNING: syncache hash size is not a power of 2.\n"); 267 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 268 } 269 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 270 271 /* Set limits. */ 272 V_tcp_syncache.cache_limit = 273 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 274 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 275 &V_tcp_syncache.cache_limit); 276 277 /* Allocate the hash table. */ 278 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 279 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 280 281 #ifdef VIMAGE 282 V_tcp_syncache.vnet = curvnet; 283 #endif 284 285 /* Initialize the hash buckets. */ 286 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 287 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 288 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 289 NULL, MTX_DEF); 290 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 291 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 292 V_tcp_syncache.hashbase[i].sch_length = 0; 293 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 294 V_tcp_syncache.hashbase[i].sch_last_overflow = 295 -(SYNCOOKIE_LIFETIME + 1); 296 } 297 298 /* Create the syncache entry zone. */ 299 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 300 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 301 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 302 V_tcp_syncache.cache_limit); 303 304 /* Start the SYN cookie reseeder callout. */ 305 callout_init(&V_tcp_syncache.secret.reseed, 1); 306 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 307 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 308 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 309 syncookie_reseed, &V_tcp_syncache); 310 311 /* Initialize the pause machinery. */ 312 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF); 313 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx, 314 0); 315 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME; 316 V_tcp_syncache.pause_backoff = 0; 317 V_tcp_syncache.paused = false; 318 } 319 320 #ifdef VIMAGE 321 void 322 syncache_destroy(void) 323 { 324 struct syncache_head *sch; 325 struct syncache *sc, *nsc; 326 int i; 327 328 /* 329 * Stop the re-seed timer before freeing resources. No need to 330 * possibly schedule it another time. 331 */ 332 callout_drain(&V_tcp_syncache.secret.reseed); 333 334 /* Stop the SYN cache pause callout. */ 335 mtx_lock(&V_tcp_syncache.pause_mtx); 336 if (callout_stop(&V_tcp_syncache.pause_co) == 0) { 337 mtx_unlock(&V_tcp_syncache.pause_mtx); 338 callout_drain(&V_tcp_syncache.pause_co); 339 } else 340 mtx_unlock(&V_tcp_syncache.pause_mtx); 341 342 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 343 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 344 sch = &V_tcp_syncache.hashbase[i]; 345 callout_drain(&sch->sch_timer); 346 347 SCH_LOCK(sch); 348 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 349 syncache_drop(sc, sch); 350 SCH_UNLOCK(sch); 351 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 352 ("%s: sch->sch_bucket not empty", __func__)); 353 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 354 __func__, sch->sch_length)); 355 mtx_destroy(&sch->sch_mtx); 356 } 357 358 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 359 ("%s: cache_count not 0", __func__)); 360 361 /* Free the allocated global resources. */ 362 uma_zdestroy(V_tcp_syncache.zone); 363 free(V_tcp_syncache.hashbase, M_SYNCACHE); 364 mtx_destroy(&V_tcp_syncache.pause_mtx); 365 } 366 #endif 367 368 /* 369 * Inserts a syncache entry into the specified bucket row. 370 * Locks and unlocks the syncache_head autonomously. 371 */ 372 static void 373 syncache_insert(struct syncache *sc, struct syncache_head *sch) 374 { 375 struct syncache *sc2; 376 377 SCH_LOCK(sch); 378 379 /* 380 * Make sure that we don't overflow the per-bucket limit. 381 * If the bucket is full, toss the oldest element. 382 */ 383 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 384 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 385 ("sch->sch_length incorrect")); 386 syncache_pause(&sc->sc_inc); 387 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 388 sch->sch_last_overflow = time_uptime; 389 syncache_drop(sc2, sch); 390 } 391 392 /* Put it into the bucket. */ 393 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 394 sch->sch_length++; 395 396 #ifdef TCP_OFFLOAD 397 if (ADDED_BY_TOE(sc)) { 398 struct toedev *tod = sc->sc_tod; 399 400 tod->tod_syncache_added(tod, sc->sc_todctx); 401 } 402 #endif 403 404 /* Reinitialize the bucket row's timer. */ 405 if (sch->sch_length == 1) 406 sch->sch_nextc = ticks + INT_MAX; 407 syncache_timeout(sc, sch, 1); 408 409 SCH_UNLOCK(sch); 410 411 TCPSTATES_INC(TCPS_SYN_RECEIVED); 412 TCPSTAT_INC(tcps_sc_added); 413 } 414 415 /* 416 * Remove and free entry from syncache bucket row. 417 * Expects locked syncache head. 418 */ 419 static void 420 syncache_drop(struct syncache *sc, struct syncache_head *sch) 421 { 422 423 SCH_LOCK_ASSERT(sch); 424 425 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 426 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 427 sch->sch_length--; 428 429 #ifdef TCP_OFFLOAD 430 if (ADDED_BY_TOE(sc)) { 431 struct toedev *tod = sc->sc_tod; 432 433 tod->tod_syncache_removed(tod, sc->sc_todctx); 434 } 435 #endif 436 437 syncache_free(sc); 438 } 439 440 /* 441 * Engage/reengage time on bucket row. 442 */ 443 static void 444 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 445 { 446 int rexmt; 447 448 if (sc->sc_rxmits == 0) 449 rexmt = tcp_rexmit_initial; 450 else 451 TCPT_RANGESET(rexmt, 452 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], 453 tcp_rexmit_min, TCPTV_REXMTMAX); 454 sc->sc_rxttime = ticks + rexmt; 455 sc->sc_rxmits++; 456 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 457 sch->sch_nextc = sc->sc_rxttime; 458 if (docallout) 459 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 460 syncache_timer, (void *)sch); 461 } 462 } 463 464 /* 465 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 466 * If we have retransmitted an entry the maximum number of times, expire it. 467 * One separate timer for each bucket row. 468 */ 469 static void 470 syncache_timer(void *xsch) 471 { 472 struct syncache_head *sch = (struct syncache_head *)xsch; 473 struct syncache *sc, *nsc; 474 struct epoch_tracker et; 475 int tick = ticks; 476 char *s; 477 bool paused; 478 479 CURVNET_SET(sch->sch_sc->vnet); 480 481 /* NB: syncache_head has already been locked by the callout. */ 482 SCH_LOCK_ASSERT(sch); 483 484 /* 485 * In the following cycle we may remove some entries and/or 486 * advance some timeouts, so re-initialize the bucket timer. 487 */ 488 sch->sch_nextc = tick + INT_MAX; 489 490 /* 491 * If we have paused processing, unconditionally remove 492 * all syncache entries. 493 */ 494 mtx_lock(&V_tcp_syncache.pause_mtx); 495 paused = V_tcp_syncache.paused; 496 mtx_unlock(&V_tcp_syncache.pause_mtx); 497 498 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 499 if (paused) { 500 syncache_drop(sc, sch); 501 continue; 502 } 503 /* 504 * We do not check if the listen socket still exists 505 * and accept the case where the listen socket may be 506 * gone by the time we resend the SYN/ACK. We do 507 * not expect this to happens often. If it does, 508 * then the RST will be sent by the time the remote 509 * host does the SYN/ACK->ACK. 510 */ 511 if (TSTMP_GT(sc->sc_rxttime, tick)) { 512 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 513 sch->sch_nextc = sc->sc_rxttime; 514 continue; 515 } 516 if (sc->sc_rxmits > V_tcp_ecn_maxretries) { 517 sc->sc_flags &= ~SCF_ECN_MASK; 518 } 519 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 520 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 521 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 522 "giving up and removing syncache entry\n", 523 s, __func__); 524 free(s, M_TCPLOG); 525 } 526 syncache_drop(sc, sch); 527 TCPSTAT_INC(tcps_sc_stale); 528 continue; 529 } 530 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 531 log(LOG_DEBUG, "%s; %s: Response timeout, " 532 "retransmitting (%u) SYN|ACK\n", 533 s, __func__, sc->sc_rxmits); 534 free(s, M_TCPLOG); 535 } 536 537 NET_EPOCH_ENTER(et); 538 syncache_respond(sc, NULL, TH_SYN|TH_ACK); 539 NET_EPOCH_EXIT(et); 540 TCPSTAT_INC(tcps_sc_retransmitted); 541 syncache_timeout(sc, sch, 0); 542 } 543 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 544 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 545 syncache_timer, (void *)(sch)); 546 CURVNET_RESTORE(); 547 } 548 549 /* 550 * Returns true if the system is only using cookies at the moment. 551 * This could be due to a sysadmin decision to only use cookies, or it 552 * could be due to the system detecting an attack. 553 */ 554 static inline bool 555 syncache_cookiesonly(void) 556 { 557 558 return (V_tcp_syncookies && (V_tcp_syncache.paused || 559 V_tcp_syncookiesonly)); 560 } 561 562 /* 563 * Find the hash bucket for the given connection. 564 */ 565 static struct syncache_head * 566 syncache_hashbucket(struct in_conninfo *inc) 567 { 568 uint32_t hash; 569 570 /* 571 * The hash is built on foreign port + local port + foreign address. 572 * We rely on the fact that struct in_conninfo starts with 16 bits 573 * of foreign port, then 16 bits of local port then followed by 128 574 * bits of foreign address. In case of IPv4 address, the first 3 575 * 32-bit words of the address always are zeroes. 576 */ 577 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 578 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 579 580 return (&V_tcp_syncache.hashbase[hash]); 581 } 582 583 /* 584 * Find an entry in the syncache. 585 * Returns always with locked syncache_head plus a matching entry or NULL. 586 */ 587 static struct syncache * 588 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 589 { 590 struct syncache *sc; 591 struct syncache_head *sch; 592 593 *schp = sch = syncache_hashbucket(inc); 594 SCH_LOCK(sch); 595 596 /* Circle through bucket row to find matching entry. */ 597 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 598 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 599 sizeof(struct in_endpoints)) == 0) 600 break; 601 602 return (sc); /* Always returns with locked sch. */ 603 } 604 605 /* 606 * This function is called when we get a RST for a 607 * non-existent connection, so that we can see if the 608 * connection is in the syn cache. If it is, zap it. 609 * If required send a challenge ACK. 610 */ 611 void 612 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m, 613 uint16_t port) 614 { 615 struct syncache *sc; 616 struct syncache_head *sch; 617 char *s = NULL; 618 619 if (syncache_cookiesonly()) 620 return; 621 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 622 SCH_LOCK_ASSERT(sch); 623 624 /* 625 * No corresponding connection was found in syncache. 626 * If syncookies are enabled and possibly exclusively 627 * used, or we are under memory pressure, a valid RST 628 * may not find a syncache entry. In that case we're 629 * done and no SYN|ACK retransmissions will happen. 630 * Otherwise the RST was misdirected or spoofed. 631 */ 632 if (sc == NULL) { 633 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 634 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 635 "syncache entry (possibly syncookie only), " 636 "segment ignored\n", s, __func__); 637 TCPSTAT_INC(tcps_badrst); 638 goto done; 639 } 640 641 /* The remote UDP encaps port does not match. */ 642 if (sc->sc_port != port) { 643 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 644 log(LOG_DEBUG, "%s; %s: Spurious RST with matching " 645 "syncache entry but non-matching UDP encaps port, " 646 "segment ignored\n", s, __func__); 647 TCPSTAT_INC(tcps_badrst); 648 goto done; 649 } 650 651 /* 652 * If the RST bit is set, check the sequence number to see 653 * if this is a valid reset segment. 654 * 655 * RFC 793 page 37: 656 * In all states except SYN-SENT, all reset (RST) segments 657 * are validated by checking their SEQ-fields. A reset is 658 * valid if its sequence number is in the window. 659 * 660 * RFC 793 page 69: 661 * There are four cases for the acceptability test for an incoming 662 * segment: 663 * 664 * Segment Receive Test 665 * Length Window 666 * ------- ------- ------------------------------------------- 667 * 0 0 SEG.SEQ = RCV.NXT 668 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 669 * >0 0 not acceptable 670 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 671 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 672 * 673 * Note that when receiving a SYN segment in the LISTEN state, 674 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 675 * described in RFC 793, page 66. 676 */ 677 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 678 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 679 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 680 if (V_tcp_insecure_rst || 681 th->th_seq == sc->sc_irs + 1) { 682 syncache_drop(sc, sch); 683 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 684 log(LOG_DEBUG, 685 "%s; %s: Our SYN|ACK was rejected, " 686 "connection attempt aborted by remote " 687 "endpoint\n", 688 s, __func__); 689 TCPSTAT_INC(tcps_sc_reset); 690 } else { 691 TCPSTAT_INC(tcps_badrst); 692 /* Send challenge ACK. */ 693 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 694 log(LOG_DEBUG, "%s; %s: RST with invalid " 695 " SEQ %u != NXT %u (+WND %u), " 696 "sending challenge ACK\n", 697 s, __func__, 698 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 699 syncache_respond(sc, m, TH_ACK); 700 } 701 } else { 702 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 703 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 704 "NXT %u (+WND %u), segment ignored\n", 705 s, __func__, 706 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 707 TCPSTAT_INC(tcps_badrst); 708 } 709 710 done: 711 if (s != NULL) 712 free(s, M_TCPLOG); 713 SCH_UNLOCK(sch); 714 } 715 716 void 717 syncache_badack(struct in_conninfo *inc, uint16_t port) 718 { 719 struct syncache *sc; 720 struct syncache_head *sch; 721 722 if (syncache_cookiesonly()) 723 return; 724 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 725 SCH_LOCK_ASSERT(sch); 726 if ((sc != NULL) && (sc->sc_port == port)) { 727 syncache_drop(sc, sch); 728 TCPSTAT_INC(tcps_sc_badack); 729 } 730 SCH_UNLOCK(sch); 731 } 732 733 void 734 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port) 735 { 736 struct syncache *sc; 737 struct syncache_head *sch; 738 739 if (syncache_cookiesonly()) 740 return; 741 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 742 SCH_LOCK_ASSERT(sch); 743 if (sc == NULL) 744 goto done; 745 746 /* If the port != sc_port, then it's a bogus ICMP msg */ 747 if (port != sc->sc_port) 748 goto done; 749 750 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 751 if (ntohl(th_seq) != sc->sc_iss) 752 goto done; 753 754 /* 755 * If we've rertransmitted 3 times and this is our second error, 756 * we remove the entry. Otherwise, we allow it to continue on. 757 * This prevents us from incorrectly nuking an entry during a 758 * spurious network outage. 759 * 760 * See tcp_notify(). 761 */ 762 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 763 sc->sc_flags |= SCF_UNREACH; 764 goto done; 765 } 766 syncache_drop(sc, sch); 767 TCPSTAT_INC(tcps_sc_unreach); 768 done: 769 SCH_UNLOCK(sch); 770 } 771 772 /* 773 * Build a new TCP socket structure from a syncache entry. 774 * 775 * On success return the newly created socket with its underlying inp locked. 776 */ 777 static struct socket * 778 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 779 { 780 struct tcpcb *listening_tcb; 781 struct inpcb *inp = NULL; 782 struct socket *so; 783 struct tcpcb *tp; 784 int error; 785 char *s; 786 787 NET_EPOCH_ASSERT(); 788 789 /* 790 * Ok, create the full blown connection, and set things up 791 * as they would have been set up if we had created the 792 * connection when the SYN arrived. 793 */ 794 if ((so = solisten_clone(lso)) == NULL) 795 goto allocfail; 796 #ifdef MAC 797 mac_socketpeer_set_from_mbuf(m, so); 798 #endif 799 error = in_pcballoc(so, &V_tcbinfo); 800 if (error) { 801 sodealloc(so); 802 goto allocfail; 803 } 804 inp = sotoinpcb(so); 805 if (V_functions_inherit_listen_socket_stack) 806 listening_tcb = sototcpcb(lso); 807 else 808 listening_tcb = NULL; 809 if ((tp = tcp_newtcpcb(inp, listening_tcb)) == NULL) { 810 in_pcbfree(inp); 811 sodealloc(so); 812 goto allocfail; 813 } 814 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 815 #ifdef INET6 816 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 817 inp->inp_vflag &= ~INP_IPV4; 818 inp->inp_vflag |= INP_IPV6; 819 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 820 } else { 821 inp->inp_vflag &= ~INP_IPV6; 822 inp->inp_vflag |= INP_IPV4; 823 #endif 824 inp->inp_ip_ttl = sc->sc_ip_ttl; 825 inp->inp_ip_tos = sc->sc_ip_tos; 826 inp->inp_laddr = sc->sc_inc.inc_laddr; 827 #ifdef INET6 828 } 829 #endif 830 831 /* 832 * If there's an mbuf and it has a flowid, then let's initialise the 833 * inp with that particular flowid. 834 */ 835 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 836 inp->inp_flowid = m->m_pkthdr.flowid; 837 inp->inp_flowtype = M_HASHTYPE_GET(m); 838 #ifdef NUMA 839 inp->inp_numa_domain = m->m_pkthdr.numa_domain; 840 #endif 841 } 842 843 inp->inp_lport = sc->sc_inc.inc_lport; 844 #ifdef INET6 845 if (inp->inp_vflag & INP_IPV6PROTO) { 846 struct inpcb *oinp = sotoinpcb(lso); 847 848 /* 849 * Inherit socket options from the listening socket. 850 * Note that in6p_inputopts are not (and should not be) 851 * copied, since it stores previously received options and is 852 * used to detect if each new option is different than the 853 * previous one and hence should be passed to a user. 854 * If we copied in6p_inputopts, a user would not be able to 855 * receive options just after calling the accept system call. 856 */ 857 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 858 if (oinp->in6p_outputopts) 859 inp->in6p_outputopts = 860 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 861 inp->in6p_hops = oinp->in6p_hops; 862 } 863 864 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 865 struct sockaddr_in6 sin6; 866 867 sin6.sin6_family = AF_INET6; 868 sin6.sin6_len = sizeof(sin6); 869 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 870 sin6.sin6_port = sc->sc_inc.inc_fport; 871 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 872 INP_HASH_WLOCK(&V_tcbinfo); 873 error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false); 874 INP_HASH_WUNLOCK(&V_tcbinfo); 875 if (error != 0) 876 goto abort; 877 /* Override flowlabel from in6_pcbconnect. */ 878 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 879 inp->inp_flow |= sc->sc_flowlabel; 880 } 881 #endif /* INET6 */ 882 #if defined(INET) && defined(INET6) 883 else 884 #endif 885 #ifdef INET 886 { 887 struct sockaddr_in sin; 888 889 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 890 891 if (inp->inp_options == NULL) { 892 inp->inp_options = sc->sc_ipopts; 893 sc->sc_ipopts = NULL; 894 } 895 896 sin.sin_family = AF_INET; 897 sin.sin_len = sizeof(sin); 898 sin.sin_addr = sc->sc_inc.inc_faddr; 899 sin.sin_port = sc->sc_inc.inc_fport; 900 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 901 INP_HASH_WLOCK(&V_tcbinfo); 902 error = in_pcbconnect(inp, &sin, thread0.td_ucred, false); 903 INP_HASH_WUNLOCK(&V_tcbinfo); 904 if (error != 0) 905 goto abort; 906 } 907 #endif /* INET */ 908 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 909 /* Copy old policy into new socket's. */ 910 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 911 printf("syncache_socket: could not copy policy\n"); 912 #endif 913 tp->t_state = TCPS_SYN_RECEIVED; 914 tp->iss = sc->sc_iss; 915 tp->irs = sc->sc_irs; 916 tp->t_port = sc->sc_port; 917 tcp_rcvseqinit(tp); 918 tcp_sendseqinit(tp); 919 tp->snd_wl1 = sc->sc_irs; 920 tp->snd_max = tp->iss + 1; 921 tp->snd_nxt = tp->iss + 1; 922 tp->rcv_up = sc->sc_irs + 1; 923 tp->rcv_wnd = sc->sc_wnd; 924 tp->rcv_adv += tp->rcv_wnd; 925 tp->last_ack_sent = tp->rcv_nxt; 926 927 tp->t_flags = sototcpcb(lso)->t_flags & 928 (TF_LRD|TF_NOPUSH|TF_NODELAY); 929 if (sc->sc_flags & SCF_NOOPT) 930 tp->t_flags |= TF_NOOPT; 931 else { 932 if (sc->sc_flags & SCF_WINSCALE) { 933 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 934 tp->snd_scale = sc->sc_requested_s_scale; 935 tp->request_r_scale = sc->sc_requested_r_scale; 936 } 937 if (sc->sc_flags & SCF_TIMESTAMP) { 938 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 939 tp->ts_recent = sc->sc_tsreflect; 940 tp->ts_recent_age = tcp_ts_getticks(); 941 tp->ts_offset = sc->sc_tsoff; 942 } 943 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 944 if (sc->sc_flags & SCF_SIGNATURE) 945 tp->t_flags |= TF_SIGNATURE; 946 #endif 947 if (sc->sc_flags & SCF_SACK) 948 tp->t_flags |= TF_SACK_PERMIT; 949 } 950 951 tcp_ecn_syncache_socket(tp, sc); 952 953 /* 954 * Set up MSS and get cached values from tcp_hostcache. 955 * This might overwrite some of the defaults we just set. 956 */ 957 tcp_mss(tp, sc->sc_peer_mss); 958 959 /* 960 * If the SYN,ACK was retransmitted, indicate that CWND to be 961 * limited to one segment in cc_conn_init(). 962 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 963 */ 964 if (sc->sc_rxmits > 1) 965 tp->snd_cwnd = 1; 966 967 #ifdef TCP_OFFLOAD 968 /* 969 * Allow a TOE driver to install its hooks. Note that we hold the 970 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 971 * new connection before the TOE driver has done its thing. 972 */ 973 if (ADDED_BY_TOE(sc)) { 974 struct toedev *tod = sc->sc_tod; 975 976 tod->tod_offload_socket(tod, sc->sc_todctx, so); 977 } 978 #endif 979 #ifdef TCP_BLACKBOX 980 /* 981 * Inherit the log state from the listening socket, if 982 * - the log state of the listening socket is not off and 983 * - the listening socket was not auto selected from all sessions and 984 * - a log id is not set on the listening socket. 985 * This avoids inheriting a log state which was automatically set. 986 */ 987 if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) && 988 ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) && 989 (sototcpcb(lso)->t_lib == NULL)) { 990 tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso))); 991 } 992 #endif 993 /* 994 * Copy and activate timers. 995 */ 996 tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime; 997 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 998 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 999 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1000 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1001 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1002 1003 TCPSTAT_INC(tcps_accepts); 1004 TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN); 1005 1006 if (!solisten_enqueue(so, SS_ISCONNECTED)) 1007 tp->t_flags |= TF_SONOTCONN; 1008 1009 return (so); 1010 1011 allocfail: 1012 /* 1013 * Drop the connection; we will either send a RST or have the peer 1014 * retransmit its SYN again after its RTO and try again. 1015 */ 1016 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 1017 log(LOG_DEBUG, "%s; %s: Socket create failed " 1018 "due to limits or memory shortage\n", 1019 s, __func__); 1020 free(s, M_TCPLOG); 1021 } 1022 TCPSTAT_INC(tcps_listendrop); 1023 return (NULL); 1024 1025 abort: 1026 tcp_discardcb(tp); 1027 in_pcbfree(inp); 1028 sodealloc(so); 1029 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 1030 log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n", 1031 s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "", 1032 error); 1033 free(s, M_TCPLOG); 1034 } 1035 TCPSTAT_INC(tcps_listendrop); 1036 return (NULL); 1037 } 1038 1039 /* 1040 * This function gets called when we receive an ACK for a 1041 * socket in the LISTEN state. We look up the connection 1042 * in the syncache, and if its there, we pull it out of 1043 * the cache and turn it into a full-blown connection in 1044 * the SYN-RECEIVED state. 1045 * 1046 * On syncache_socket() success the newly created socket 1047 * has its underlying inp locked. 1048 */ 1049 int 1050 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1051 struct socket **lsop, struct mbuf *m, uint16_t port) 1052 { 1053 struct syncache *sc; 1054 struct syncache_head *sch; 1055 struct syncache scs; 1056 char *s; 1057 bool locked; 1058 1059 NET_EPOCH_ASSERT(); 1060 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1061 ("%s: can handle only ACK", __func__)); 1062 1063 if (syncache_cookiesonly()) { 1064 sc = NULL; 1065 sch = syncache_hashbucket(inc); 1066 locked = false; 1067 } else { 1068 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1069 locked = true; 1070 SCH_LOCK_ASSERT(sch); 1071 } 1072 1073 #ifdef INVARIANTS 1074 /* 1075 * Test code for syncookies comparing the syncache stored 1076 * values with the reconstructed values from the cookie. 1077 */ 1078 if (sc != NULL) 1079 syncookie_cmp(inc, sch, sc, th, to, *lsop, port); 1080 #endif 1081 1082 if (sc == NULL) { 1083 /* 1084 * There is no syncache entry, so see if this ACK is 1085 * a returning syncookie. To do this, first: 1086 * A. Check if syncookies are used in case of syncache 1087 * overflows 1088 * B. See if this socket has had a syncache entry dropped in 1089 * the recent past. We don't want to accept a bogus 1090 * syncookie if we've never received a SYN or accept it 1091 * twice. 1092 * C. check that the syncookie is valid. If it is, then 1093 * cobble up a fake syncache entry, and return. 1094 */ 1095 if (locked && !V_tcp_syncookies) { 1096 SCH_UNLOCK(sch); 1097 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1098 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1099 "segment rejected (syncookies disabled)\n", 1100 s, __func__); 1101 goto failed; 1102 } 1103 if (locked && !V_tcp_syncookiesonly && 1104 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1105 SCH_UNLOCK(sch); 1106 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1107 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1108 "segment rejected (no syncache entry)\n", 1109 s, __func__); 1110 goto failed; 1111 } 1112 bzero(&scs, sizeof(scs)); 1113 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port); 1114 if (locked) 1115 SCH_UNLOCK(sch); 1116 if (sc == NULL) { 1117 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1118 log(LOG_DEBUG, "%s; %s: Segment failed " 1119 "SYNCOOKIE authentication, segment rejected " 1120 "(probably spoofed)\n", s, __func__); 1121 goto failed; 1122 } 1123 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1124 /* If received ACK has MD5 signature, check it. */ 1125 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1126 (!TCPMD5_ENABLED() || 1127 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1128 /* Drop the ACK. */ 1129 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1130 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1131 "MD5 signature doesn't match.\n", 1132 s, __func__); 1133 free(s, M_TCPLOG); 1134 } 1135 TCPSTAT_INC(tcps_sig_err_sigopt); 1136 return (-1); /* Do not send RST */ 1137 } 1138 #endif /* TCP_SIGNATURE */ 1139 TCPSTATES_INC(TCPS_SYN_RECEIVED); 1140 } else { 1141 if (sc->sc_port != port) { 1142 SCH_UNLOCK(sch); 1143 return (0); 1144 } 1145 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1146 /* 1147 * If listening socket requested TCP digests, check that 1148 * received ACK has signature and it is correct. 1149 * If not, drop the ACK and leave sc entry in th cache, 1150 * because SYN was received with correct signature. 1151 */ 1152 if (sc->sc_flags & SCF_SIGNATURE) { 1153 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1154 /* No signature */ 1155 TCPSTAT_INC(tcps_sig_err_nosigopt); 1156 SCH_UNLOCK(sch); 1157 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1158 log(LOG_DEBUG, "%s; %s: Segment " 1159 "rejected, MD5 signature wasn't " 1160 "provided.\n", s, __func__); 1161 free(s, M_TCPLOG); 1162 } 1163 return (-1); /* Do not send RST */ 1164 } 1165 if (!TCPMD5_ENABLED() || 1166 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1167 /* Doesn't match or no SA */ 1168 SCH_UNLOCK(sch); 1169 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1170 log(LOG_DEBUG, "%s; %s: Segment " 1171 "rejected, MD5 signature doesn't " 1172 "match.\n", s, __func__); 1173 free(s, M_TCPLOG); 1174 } 1175 return (-1); /* Do not send RST */ 1176 } 1177 } 1178 #endif /* TCP_SIGNATURE */ 1179 1180 /* 1181 * RFC 7323 PAWS: If we have a timestamp on this segment and 1182 * it's less than ts_recent, drop it. 1183 * XXXMT: RFC 7323 also requires to send an ACK. 1184 * In tcp_input.c this is only done for TCP segments 1185 * with user data, so be consistent here and just drop 1186 * the segment. 1187 */ 1188 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1189 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1190 SCH_UNLOCK(sch); 1191 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1192 log(LOG_DEBUG, 1193 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1194 "segment dropped\n", s, __func__, 1195 to->to_tsval, sc->sc_tsreflect); 1196 free(s, M_TCPLOG); 1197 } 1198 return (-1); /* Do not send RST */ 1199 } 1200 1201 /* 1202 * If timestamps were not negotiated during SYN/ACK and a 1203 * segment with a timestamp is received, ignore the 1204 * timestamp and process the packet normally. 1205 * See section 3.2 of RFC 7323. 1206 */ 1207 if (!(sc->sc_flags & SCF_TIMESTAMP) && 1208 (to->to_flags & TOF_TS)) { 1209 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1210 log(LOG_DEBUG, "%s; %s: Timestamp not " 1211 "expected, segment processed normally\n", 1212 s, __func__); 1213 free(s, M_TCPLOG); 1214 s = NULL; 1215 } 1216 } 1217 1218 /* 1219 * If timestamps were negotiated during SYN/ACK and a 1220 * segment without a timestamp is received, silently drop 1221 * the segment, unless the missing timestamps are tolerated. 1222 * See section 3.2 of RFC 7323. 1223 */ 1224 if ((sc->sc_flags & SCF_TIMESTAMP) && 1225 !(to->to_flags & TOF_TS)) { 1226 if (V_tcp_tolerate_missing_ts) { 1227 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1228 log(LOG_DEBUG, 1229 "%s; %s: Timestamp missing, " 1230 "segment processed normally\n", 1231 s, __func__); 1232 free(s, M_TCPLOG); 1233 } 1234 } else { 1235 SCH_UNLOCK(sch); 1236 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1237 log(LOG_DEBUG, 1238 "%s; %s: Timestamp missing, " 1239 "segment silently dropped\n", 1240 s, __func__); 1241 free(s, M_TCPLOG); 1242 } 1243 return (-1); /* Do not send RST */ 1244 } 1245 } 1246 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1247 sch->sch_length--; 1248 #ifdef TCP_OFFLOAD 1249 if (ADDED_BY_TOE(sc)) { 1250 struct toedev *tod = sc->sc_tod; 1251 1252 tod->tod_syncache_removed(tod, sc->sc_todctx); 1253 } 1254 #endif 1255 SCH_UNLOCK(sch); 1256 } 1257 1258 /* 1259 * Segment validation: 1260 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1261 */ 1262 if (th->th_ack != sc->sc_iss + 1) { 1263 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1264 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1265 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1266 goto failed; 1267 } 1268 1269 /* 1270 * The SEQ must fall in the window starting at the received 1271 * initial receive sequence number + 1 (the SYN). 1272 */ 1273 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1274 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1275 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1276 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1277 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1278 goto failed; 1279 } 1280 1281 *lsop = syncache_socket(sc, *lsop, m); 1282 1283 if (__predict_false(*lsop == NULL)) { 1284 TCPSTAT_INC(tcps_sc_aborted); 1285 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1286 } else 1287 TCPSTAT_INC(tcps_sc_completed); 1288 1289 /* how do we find the inp for the new socket? */ 1290 if (sc != &scs) 1291 syncache_free(sc); 1292 return (1); 1293 failed: 1294 if (sc != NULL) { 1295 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1296 if (sc != &scs) 1297 syncache_free(sc); 1298 } 1299 if (s != NULL) 1300 free(s, M_TCPLOG); 1301 *lsop = NULL; 1302 return (0); 1303 } 1304 1305 static struct socket * 1306 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m, 1307 uint64_t response_cookie) 1308 { 1309 struct inpcb *inp; 1310 struct tcpcb *tp; 1311 unsigned int *pending_counter; 1312 struct socket *so; 1313 1314 NET_EPOCH_ASSERT(); 1315 1316 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending; 1317 so = syncache_socket(sc, lso, m); 1318 if (so == NULL) { 1319 TCPSTAT_INC(tcps_sc_aborted); 1320 atomic_subtract_int(pending_counter, 1); 1321 } else { 1322 soisconnected(so); 1323 inp = sotoinpcb(so); 1324 tp = intotcpcb(inp); 1325 tp->t_flags |= TF_FASTOPEN; 1326 tp->t_tfo_cookie.server = response_cookie; 1327 tp->snd_max = tp->iss; 1328 tp->snd_nxt = tp->iss; 1329 tp->t_tfo_pending = pending_counter; 1330 TCPSTATES_INC(TCPS_SYN_RECEIVED); 1331 TCPSTAT_INC(tcps_sc_completed); 1332 } 1333 1334 return (so); 1335 } 1336 1337 /* 1338 * Given a LISTEN socket and an inbound SYN request, add 1339 * this to the syn cache, and send back a segment: 1340 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1341 * to the source. 1342 * 1343 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1344 * Doing so would require that we hold onto the data and deliver it 1345 * to the application. However, if we are the target of a SYN-flood 1346 * DoS attack, an attacker could send data which would eventually 1347 * consume all available buffer space if it were ACKed. By not ACKing 1348 * the data, we avoid this DoS scenario. 1349 * 1350 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1351 * cookie is processed and a new socket is created. In this case, any data 1352 * accompanying the SYN will be queued to the socket by tcp_input() and will 1353 * be ACKed either when the application sends response data or the delayed 1354 * ACK timer expires, whichever comes first. 1355 */ 1356 struct socket * 1357 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1358 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod, 1359 void *todctx, uint8_t iptos, uint16_t port) 1360 { 1361 struct tcpcb *tp; 1362 struct socket *rv = NULL; 1363 struct syncache *sc = NULL; 1364 struct syncache_head *sch; 1365 struct mbuf *ipopts = NULL; 1366 u_int ltflags; 1367 int win, ip_ttl, ip_tos; 1368 char *s; 1369 #ifdef INET6 1370 int autoflowlabel = 0; 1371 #endif 1372 #ifdef MAC 1373 struct label *maclabel; 1374 #endif 1375 struct syncache scs; 1376 struct ucred *cred; 1377 uint64_t tfo_response_cookie; 1378 unsigned int *tfo_pending = NULL; 1379 int tfo_cookie_valid = 0; 1380 int tfo_response_cookie_valid = 0; 1381 bool locked; 1382 1383 INP_RLOCK_ASSERT(inp); /* listen socket */ 1384 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1385 ("%s: unexpected tcp flags", __func__)); 1386 1387 /* 1388 * Combine all so/tp operations very early to drop the INP lock as 1389 * soon as possible. 1390 */ 1391 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1392 tp = sototcpcb(so); 1393 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred); 1394 1395 #ifdef INET6 1396 if (inc->inc_flags & INC_ISIPV6) { 1397 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1398 autoflowlabel = 1; 1399 } 1400 ip_ttl = in6_selecthlim(inp, NULL); 1401 if ((inp->in6p_outputopts == NULL) || 1402 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1403 ip_tos = 0; 1404 } else { 1405 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1406 } 1407 } 1408 #endif 1409 #if defined(INET6) && defined(INET) 1410 else 1411 #endif 1412 #ifdef INET 1413 { 1414 ip_ttl = inp->inp_ip_ttl; 1415 ip_tos = inp->inp_ip_tos; 1416 } 1417 #endif 1418 win = so->sol_sbrcv_hiwat; 1419 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1420 1421 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1422 (tp->t_tfo_pending != NULL) && 1423 (to->to_flags & TOF_FASTOPEN)) { 1424 /* 1425 * Limit the number of pending TFO connections to 1426 * approximately half of the queue limit. This prevents TFO 1427 * SYN floods from starving the service by filling the 1428 * listen queue with bogus TFO connections. 1429 */ 1430 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1431 (so->sol_qlimit / 2)) { 1432 int result; 1433 1434 result = tcp_fastopen_check_cookie(inc, 1435 to->to_tfo_cookie, to->to_tfo_len, 1436 &tfo_response_cookie); 1437 tfo_cookie_valid = (result > 0); 1438 tfo_response_cookie_valid = (result >= 0); 1439 } 1440 1441 /* 1442 * Remember the TFO pending counter as it will have to be 1443 * decremented below if we don't make it to syncache_tfo_expand(). 1444 */ 1445 tfo_pending = tp->t_tfo_pending; 1446 } 1447 1448 #ifdef MAC 1449 if (mac_syncache_init(&maclabel) != 0) { 1450 INP_RUNLOCK(inp); 1451 goto done; 1452 } else 1453 mac_syncache_create(maclabel, inp); 1454 #endif 1455 if (!tfo_cookie_valid) 1456 INP_RUNLOCK(inp); 1457 1458 /* 1459 * Remember the IP options, if any. 1460 */ 1461 #ifdef INET6 1462 if (!(inc->inc_flags & INC_ISIPV6)) 1463 #endif 1464 #ifdef INET 1465 ipopts = (m) ? ip_srcroute(m) : NULL; 1466 #else 1467 ipopts = NULL; 1468 #endif 1469 1470 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1471 /* 1472 * When the socket is TCP-MD5 enabled check that, 1473 * - a signed packet is valid 1474 * - a non-signed packet does not have a security association 1475 * 1476 * If a signed packet fails validation or a non-signed packet has a 1477 * security association, the packet will be dropped. 1478 */ 1479 if (ltflags & TF_SIGNATURE) { 1480 if (to->to_flags & TOF_SIGNATURE) { 1481 if (!TCPMD5_ENABLED() || 1482 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1483 goto done; 1484 } else { 1485 if (TCPMD5_ENABLED() && 1486 TCPMD5_INPUT(m, NULL, NULL) != ENOENT) 1487 goto done; 1488 } 1489 } else if (to->to_flags & TOF_SIGNATURE) 1490 goto done; 1491 #endif /* TCP_SIGNATURE */ 1492 /* 1493 * See if we already have an entry for this connection. 1494 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1495 * 1496 * XXX: should the syncache be re-initialized with the contents 1497 * of the new SYN here (which may have different options?) 1498 * 1499 * XXX: We do not check the sequence number to see if this is a 1500 * real retransmit or a new connection attempt. The question is 1501 * how to handle such a case; either ignore it as spoofed, or 1502 * drop the current entry and create a new one? 1503 */ 1504 if (syncache_cookiesonly()) { 1505 sc = NULL; 1506 sch = syncache_hashbucket(inc); 1507 locked = false; 1508 } else { 1509 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1510 locked = true; 1511 SCH_LOCK_ASSERT(sch); 1512 } 1513 if (sc != NULL) { 1514 if (tfo_cookie_valid) 1515 INP_RUNLOCK(inp); 1516 TCPSTAT_INC(tcps_sc_dupsyn); 1517 if (ipopts) { 1518 /* 1519 * If we were remembering a previous source route, 1520 * forget it and use the new one we've been given. 1521 */ 1522 if (sc->sc_ipopts) 1523 (void) m_free(sc->sc_ipopts); 1524 sc->sc_ipopts = ipopts; 1525 } 1526 /* 1527 * Update timestamp if present. 1528 */ 1529 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1530 sc->sc_tsreflect = to->to_tsval; 1531 else 1532 sc->sc_flags &= ~SCF_TIMESTAMP; 1533 /* 1534 * Adjust ECN response if needed, e.g. different 1535 * IP ECN field, or a fallback by the remote host. 1536 */ 1537 if (sc->sc_flags & SCF_ECN_MASK) { 1538 sc->sc_flags &= ~SCF_ECN_MASK; 1539 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos); 1540 } 1541 #ifdef MAC 1542 /* 1543 * Since we have already unconditionally allocated label 1544 * storage, free it up. The syncache entry will already 1545 * have an initialized label we can use. 1546 */ 1547 mac_syncache_destroy(&maclabel); 1548 #endif 1549 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1550 /* Retransmit SYN|ACK and reset retransmit count. */ 1551 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1552 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1553 "resetting timer and retransmitting SYN|ACK\n", 1554 s, __func__); 1555 free(s, M_TCPLOG); 1556 } 1557 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1558 sc->sc_rxmits = 0; 1559 syncache_timeout(sc, sch, 1); 1560 TCPSTAT_INC(tcps_sndacks); 1561 TCPSTAT_INC(tcps_sndtotal); 1562 } 1563 SCH_UNLOCK(sch); 1564 goto donenoprobe; 1565 } 1566 1567 if (tfo_cookie_valid) { 1568 bzero(&scs, sizeof(scs)); 1569 sc = &scs; 1570 goto skip_alloc; 1571 } 1572 1573 /* 1574 * Skip allocating a syncache entry if we are just going to discard 1575 * it later. 1576 */ 1577 if (!locked) { 1578 bzero(&scs, sizeof(scs)); 1579 sc = &scs; 1580 } else 1581 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1582 if (sc == NULL) { 1583 /* 1584 * The zone allocator couldn't provide more entries. 1585 * Treat this as if the cache was full; drop the oldest 1586 * entry and insert the new one. 1587 */ 1588 TCPSTAT_INC(tcps_sc_zonefail); 1589 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1590 sch->sch_last_overflow = time_uptime; 1591 syncache_drop(sc, sch); 1592 syncache_pause(inc); 1593 } 1594 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1595 if (sc == NULL) { 1596 if (V_tcp_syncookies) { 1597 bzero(&scs, sizeof(scs)); 1598 sc = &scs; 1599 } else { 1600 KASSERT(locked, 1601 ("%s: bucket unexpectedly unlocked", 1602 __func__)); 1603 SCH_UNLOCK(sch); 1604 if (ipopts) 1605 (void) m_free(ipopts); 1606 goto done; 1607 } 1608 } 1609 } 1610 1611 skip_alloc: 1612 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1613 sc->sc_tfo_cookie = &tfo_response_cookie; 1614 1615 /* 1616 * Fill in the syncache values. 1617 */ 1618 #ifdef MAC 1619 sc->sc_label = maclabel; 1620 #endif 1621 sc->sc_cred = cred; 1622 sc->sc_port = port; 1623 cred = NULL; 1624 sc->sc_ipopts = ipopts; 1625 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1626 sc->sc_ip_tos = ip_tos; 1627 sc->sc_ip_ttl = ip_ttl; 1628 #ifdef TCP_OFFLOAD 1629 sc->sc_tod = tod; 1630 sc->sc_todctx = todctx; 1631 #endif 1632 sc->sc_irs = th->th_seq; 1633 sc->sc_flags = 0; 1634 sc->sc_flowlabel = 0; 1635 1636 /* 1637 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1638 * win was derived from socket earlier in the function. 1639 */ 1640 win = imax(win, 0); 1641 win = imin(win, TCP_MAXWIN); 1642 sc->sc_wnd = win; 1643 1644 if (V_tcp_do_rfc1323 && 1645 !(ltflags & TF_NOOPT)) { 1646 /* 1647 * A timestamp received in a SYN makes 1648 * it ok to send timestamp requests and replies. 1649 */ 1650 if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) { 1651 sc->sc_tsreflect = to->to_tsval; 1652 sc->sc_flags |= SCF_TIMESTAMP; 1653 sc->sc_tsoff = tcp_new_ts_offset(inc); 1654 } 1655 if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) { 1656 int wscale = 0; 1657 1658 /* 1659 * Pick the smallest possible scaling factor that 1660 * will still allow us to scale up to sb_max, aka 1661 * kern.ipc.maxsockbuf. 1662 * 1663 * We do this because there are broken firewalls that 1664 * will corrupt the window scale option, leading to 1665 * the other endpoint believing that our advertised 1666 * window is unscaled. At scale factors larger than 1667 * 5 the unscaled window will drop below 1500 bytes, 1668 * leading to serious problems when traversing these 1669 * broken firewalls. 1670 * 1671 * With the default maxsockbuf of 256K, a scale factor 1672 * of 3 will be chosen by this algorithm. Those who 1673 * choose a larger maxsockbuf should watch out 1674 * for the compatibility problems mentioned above. 1675 * 1676 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1677 * or <SYN,ACK>) segment itself is never scaled. 1678 */ 1679 while (wscale < TCP_MAX_WINSHIFT && 1680 (TCP_MAXWIN << wscale) < sb_max) 1681 wscale++; 1682 sc->sc_requested_r_scale = wscale; 1683 sc->sc_requested_s_scale = to->to_wscale; 1684 sc->sc_flags |= SCF_WINSCALE; 1685 } 1686 } 1687 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1688 /* 1689 * If incoming packet has an MD5 signature, flag this in the 1690 * syncache so that syncache_respond() will do the right thing 1691 * with the SYN+ACK. 1692 */ 1693 if (to->to_flags & TOF_SIGNATURE) 1694 sc->sc_flags |= SCF_SIGNATURE; 1695 #endif /* TCP_SIGNATURE */ 1696 if (to->to_flags & TOF_SACKPERM) 1697 sc->sc_flags |= SCF_SACK; 1698 if (to->to_flags & TOF_MSS) 1699 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1700 if (ltflags & TF_NOOPT) 1701 sc->sc_flags |= SCF_NOOPT; 1702 /* ECN Handshake */ 1703 if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0) 1704 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos); 1705 1706 if (V_tcp_syncookies) 1707 sc->sc_iss = syncookie_generate(sch, sc); 1708 else 1709 sc->sc_iss = arc4random(); 1710 #ifdef INET6 1711 if (autoflowlabel) { 1712 if (V_tcp_syncookies) 1713 sc->sc_flowlabel = sc->sc_iss; 1714 else 1715 sc->sc_flowlabel = ip6_randomflowlabel(); 1716 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1717 } 1718 #endif 1719 if (locked) 1720 SCH_UNLOCK(sch); 1721 1722 if (tfo_cookie_valid) { 1723 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie); 1724 /* INP_RUNLOCK(inp) will be performed by the caller */ 1725 goto tfo_expanded; 1726 } 1727 1728 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1729 /* 1730 * Do a standard 3-way handshake. 1731 */ 1732 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1733 if (sc != &scs) 1734 syncache_insert(sc, sch); /* locks and unlocks sch */ 1735 TCPSTAT_INC(tcps_sndacks); 1736 TCPSTAT_INC(tcps_sndtotal); 1737 } else { 1738 if (sc != &scs) 1739 syncache_free(sc); 1740 TCPSTAT_INC(tcps_sc_dropped); 1741 } 1742 goto donenoprobe; 1743 1744 done: 1745 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1746 donenoprobe: 1747 if (m) 1748 m_freem(m); 1749 /* 1750 * If tfo_pending is not NULL here, then a TFO SYN that did not 1751 * result in a new socket was processed and the associated pending 1752 * counter has not yet been decremented. All such TFO processing paths 1753 * transit this point. 1754 */ 1755 if (tfo_pending != NULL) 1756 tcp_fastopen_decrement_counter(tfo_pending); 1757 1758 tfo_expanded: 1759 if (cred != NULL) 1760 crfree(cred); 1761 #ifdef MAC 1762 if (sc == &scs) 1763 mac_syncache_destroy(&maclabel); 1764 #endif 1765 return (rv); 1766 } 1767 1768 /* 1769 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1770 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1771 */ 1772 static int 1773 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags) 1774 { 1775 struct ip *ip = NULL; 1776 struct mbuf *m; 1777 struct tcphdr *th = NULL; 1778 struct udphdr *udp = NULL; 1779 int optlen, error = 0; /* Make compiler happy */ 1780 u_int16_t hlen, tlen, mssopt, ulen; 1781 struct tcpopt to; 1782 #ifdef INET6 1783 struct ip6_hdr *ip6 = NULL; 1784 #endif 1785 1786 NET_EPOCH_ASSERT(); 1787 1788 hlen = 1789 #ifdef INET6 1790 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1791 #endif 1792 sizeof(struct ip); 1793 tlen = hlen + sizeof(struct tcphdr); 1794 if (sc->sc_port) { 1795 tlen += sizeof(struct udphdr); 1796 } 1797 /* Determine MSS we advertize to other end of connection. */ 1798 mssopt = tcp_mssopt(&sc->sc_inc); 1799 if (sc->sc_port) 1800 mssopt -= V_tcp_udp_tunneling_overhead; 1801 mssopt = max(mssopt, V_tcp_minmss); 1802 1803 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1804 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1805 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + " 1806 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port, 1807 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN)); 1808 1809 /* Create the IP+TCP header from scratch. */ 1810 m = m_gethdr(M_NOWAIT, MT_DATA); 1811 if (m == NULL) 1812 return (ENOBUFS); 1813 #ifdef MAC 1814 mac_syncache_create_mbuf(sc->sc_label, m); 1815 #endif 1816 m->m_data += max_linkhdr; 1817 m->m_len = tlen; 1818 m->m_pkthdr.len = tlen; 1819 m->m_pkthdr.rcvif = NULL; 1820 1821 #ifdef INET6 1822 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1823 ip6 = mtod(m, struct ip6_hdr *); 1824 ip6->ip6_vfc = IPV6_VERSION; 1825 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1826 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1827 ip6->ip6_plen = htons(tlen - hlen); 1828 /* ip6_hlim is set after checksum */ 1829 /* Zero out traffic class and flow label. */ 1830 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1831 ip6->ip6_flow |= sc->sc_flowlabel; 1832 if (sc->sc_port != 0) { 1833 ip6->ip6_nxt = IPPROTO_UDP; 1834 udp = (struct udphdr *)(ip6 + 1); 1835 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1836 udp->uh_dport = sc->sc_port; 1837 ulen = (tlen - sizeof(struct ip6_hdr)); 1838 th = (struct tcphdr *)(udp + 1); 1839 } else { 1840 ip6->ip6_nxt = IPPROTO_TCP; 1841 th = (struct tcphdr *)(ip6 + 1); 1842 } 1843 ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN); 1844 } 1845 #endif 1846 #if defined(INET6) && defined(INET) 1847 else 1848 #endif 1849 #ifdef INET 1850 { 1851 ip = mtod(m, struct ip *); 1852 ip->ip_v = IPVERSION; 1853 ip->ip_hl = sizeof(struct ip) >> 2; 1854 ip->ip_len = htons(tlen); 1855 ip->ip_id = 0; 1856 ip->ip_off = 0; 1857 ip->ip_sum = 0; 1858 ip->ip_src = sc->sc_inc.inc_laddr; 1859 ip->ip_dst = sc->sc_inc.inc_faddr; 1860 ip->ip_ttl = sc->sc_ip_ttl; 1861 ip->ip_tos = sc->sc_ip_tos; 1862 1863 /* 1864 * See if we should do MTU discovery. Route lookups are 1865 * expensive, so we will only unset the DF bit if: 1866 * 1867 * 1) path_mtu_discovery is disabled 1868 * 2) the SCF_UNREACH flag has been set 1869 */ 1870 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1871 ip->ip_off |= htons(IP_DF); 1872 if (sc->sc_port == 0) { 1873 ip->ip_p = IPPROTO_TCP; 1874 th = (struct tcphdr *)(ip + 1); 1875 } else { 1876 ip->ip_p = IPPROTO_UDP; 1877 udp = (struct udphdr *)(ip + 1); 1878 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1879 udp->uh_dport = sc->sc_port; 1880 ulen = (tlen - sizeof(struct ip)); 1881 th = (struct tcphdr *)(udp + 1); 1882 } 1883 } 1884 #endif /* INET */ 1885 th->th_sport = sc->sc_inc.inc_lport; 1886 th->th_dport = sc->sc_inc.inc_fport; 1887 1888 if (flags & TH_SYN) 1889 th->th_seq = htonl(sc->sc_iss); 1890 else 1891 th->th_seq = htonl(sc->sc_iss + 1); 1892 th->th_ack = htonl(sc->sc_irs + 1); 1893 th->th_off = sizeof(struct tcphdr) >> 2; 1894 th->th_win = htons(sc->sc_wnd); 1895 th->th_urp = 0; 1896 1897 flags = tcp_ecn_syncache_respond(flags, sc); 1898 tcp_set_flags(th, flags); 1899 1900 /* Tack on the TCP options. */ 1901 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1902 to.to_flags = 0; 1903 1904 if (flags & TH_SYN) { 1905 to.to_mss = mssopt; 1906 to.to_flags = TOF_MSS; 1907 if (sc->sc_flags & SCF_WINSCALE) { 1908 to.to_wscale = sc->sc_requested_r_scale; 1909 to.to_flags |= TOF_SCALE; 1910 } 1911 if (sc->sc_flags & SCF_SACK) 1912 to.to_flags |= TOF_SACKPERM; 1913 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1914 if (sc->sc_flags & SCF_SIGNATURE) 1915 to.to_flags |= TOF_SIGNATURE; 1916 #endif 1917 if (sc->sc_tfo_cookie) { 1918 to.to_flags |= TOF_FASTOPEN; 1919 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1920 to.to_tfo_cookie = sc->sc_tfo_cookie; 1921 /* don't send cookie again when retransmitting response */ 1922 sc->sc_tfo_cookie = NULL; 1923 } 1924 } 1925 if (sc->sc_flags & SCF_TIMESTAMP) { 1926 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1927 to.to_tsecr = sc->sc_tsreflect; 1928 to.to_flags |= TOF_TS; 1929 } 1930 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1931 1932 /* Adjust headers by option size. */ 1933 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1934 m->m_len += optlen; 1935 m->m_pkthdr.len += optlen; 1936 #ifdef INET6 1937 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1938 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1939 else 1940 #endif 1941 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1942 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1943 if (sc->sc_flags & SCF_SIGNATURE) { 1944 KASSERT(to.to_flags & TOF_SIGNATURE, 1945 ("tcp_addoptions() didn't set tcp_signature")); 1946 1947 /* NOTE: to.to_signature is inside of mbuf */ 1948 if (!TCPMD5_ENABLED() || 1949 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1950 m_freem(m); 1951 return (EACCES); 1952 } 1953 } 1954 #endif 1955 } else 1956 optlen = 0; 1957 1958 if (udp) { 1959 ulen += optlen; 1960 udp->uh_ulen = htons(ulen); 1961 } 1962 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1963 /* 1964 * If we have peer's SYN and it has a flowid, then let's assign it to 1965 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1966 * to SYN|ACK due to lack of inp here. 1967 */ 1968 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1969 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1970 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1971 } 1972 #ifdef INET6 1973 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1974 if (sc->sc_port) { 1975 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 1976 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 1977 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, 1978 IPPROTO_UDP, 0); 1979 th->th_sum = htons(0); 1980 } else { 1981 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1982 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1983 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1984 IPPROTO_TCP, 0); 1985 } 1986 ip6->ip6_hlim = sc->sc_ip_ttl; 1987 #ifdef TCP_OFFLOAD 1988 if (ADDED_BY_TOE(sc)) { 1989 struct toedev *tod = sc->sc_tod; 1990 1991 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1992 1993 return (error); 1994 } 1995 #endif 1996 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 1997 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1998 } 1999 #endif 2000 #if defined(INET6) && defined(INET) 2001 else 2002 #endif 2003 #ifdef INET 2004 { 2005 if (sc->sc_port) { 2006 m->m_pkthdr.csum_flags = CSUM_UDP; 2007 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2008 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 2009 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 2010 th->th_sum = htons(0); 2011 } else { 2012 m->m_pkthdr.csum_flags = CSUM_TCP; 2013 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2014 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 2015 htons(tlen + optlen - hlen + IPPROTO_TCP)); 2016 } 2017 #ifdef TCP_OFFLOAD 2018 if (ADDED_BY_TOE(sc)) { 2019 struct toedev *tod = sc->sc_tod; 2020 2021 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2022 2023 return (error); 2024 } 2025 #endif 2026 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 2027 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 2028 } 2029 #endif 2030 return (error); 2031 } 2032 2033 /* 2034 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 2035 * that exceed the capacity of the syncache by avoiding the storage of any 2036 * of the SYNs we receive. Syncookies defend against blind SYN flooding 2037 * attacks where the attacker does not have access to our responses. 2038 * 2039 * Syncookies encode and include all necessary information about the 2040 * connection setup within the SYN|ACK that we send back. That way we 2041 * can avoid keeping any local state until the ACK to our SYN|ACK returns 2042 * (if ever). Normally the syncache and syncookies are running in parallel 2043 * with the latter taking over when the former is exhausted. When matching 2044 * syncache entry is found the syncookie is ignored. 2045 * 2046 * The only reliable information persisting the 3WHS is our initial sequence 2047 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 2048 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 2049 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 2050 * returns and signifies a legitimate connection if it matches the ACK. 2051 * 2052 * The available space of 32 bits to store the hash and to encode the SYN 2053 * option information is very tight and we should have at least 24 bits for 2054 * the MAC to keep the number of guesses by blind spoofing reasonably high. 2055 * 2056 * SYN option information we have to encode to fully restore a connection: 2057 * MSS: is imporant to chose an optimal segment size to avoid IP level 2058 * fragmentation along the path. The common MSS values can be encoded 2059 * in a 3-bit table. Uncommon values are captured by the next lower value 2060 * in the table leading to a slight increase in packetization overhead. 2061 * WSCALE: is necessary to allow large windows to be used for high delay- 2062 * bandwidth product links. Not scaling the window when it was initially 2063 * negotiated is bad for performance as lack of scaling further decreases 2064 * the apparent available send window. We only need to encode the WSCALE 2065 * we received from the remote end. Our end can be recalculated at any 2066 * time. The common WSCALE values can be encoded in a 3-bit table. 2067 * Uncommon values are captured by the next lower value in the table 2068 * making us under-estimate the available window size halving our 2069 * theoretically possible maximum throughput for that connection. 2070 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 2071 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 2072 * that are included in all segments on a connection. We enable them when 2073 * the ACK has them. 2074 * 2075 * Security of syncookies and attack vectors: 2076 * 2077 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 2078 * together with the gloabl secret to make it unique per connection attempt. 2079 * Thus any change of any of those parameters results in a different MAC output 2080 * in an unpredictable way unless a collision is encountered. 24 bits of the 2081 * MAC are embedded into the ISS. 2082 * 2083 * To prevent replay attacks two rotating global secrets are updated with a 2084 * new random value every 15 seconds. The life-time of a syncookie is thus 2085 * 15-30 seconds. 2086 * 2087 * Vector 1: Attacking the secret. This requires finding a weakness in the 2088 * MAC itself or the way it is used here. The attacker can do a chosen plain 2089 * text attack by varying and testing the all parameters under his control. 2090 * The strength depends on the size and randomness of the secret, and the 2091 * cryptographic security of the MAC function. Due to the constant updating 2092 * of the secret the attacker has at most 29.999 seconds to find the secret 2093 * and launch spoofed connections. After that he has to start all over again. 2094 * 2095 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 2096 * size an average of 4,823 attempts are required for a 50% chance of success 2097 * to spoof a single syncookie (birthday collision paradox). However the 2098 * attacker is blind and doesn't know if one of his attempts succeeded unless 2099 * he has a side channel to interfere success from. A single connection setup 2100 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 2101 * This many attempts are required for each one blind spoofed connection. For 2102 * every additional spoofed connection he has to launch another N attempts. 2103 * Thus for a sustained rate 100 spoofed connections per second approximately 2104 * 1,800,000 packets per second would have to be sent. 2105 * 2106 * NB: The MAC function should be fast so that it doesn't become a CPU 2107 * exhaustion attack vector itself. 2108 * 2109 * References: 2110 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 2111 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 2112 * http://cr.yp.to/syncookies.html (overview) 2113 * http://cr.yp.to/syncookies/archive (details) 2114 * 2115 * 2116 * Schematic construction of a syncookie enabled Initial Sequence Number: 2117 * 0 1 2 3 2118 * 12345678901234567890123456789012 2119 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2120 * 2121 * x 24 MAC (truncated) 2122 * W 3 Send Window Scale index 2123 * M 3 MSS index 2124 * S 1 SACK permitted 2125 * P 1 Odd/even secret 2126 */ 2127 2128 /* 2129 * Distribution and probability of certain MSS values. Those in between are 2130 * rounded down to the next lower one. 2131 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2132 * .2% .3% 5% 7% 7% 20% 15% 45% 2133 */ 2134 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2135 2136 /* 2137 * Distribution and probability of certain WSCALE values. We have to map the 2138 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2139 * bits based on prevalence of certain values. Where we don't have an exact 2140 * match for are rounded down to the next lower one letting us under-estimate 2141 * the true available window. At the moment this would happen only for the 2142 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2143 * and window size). The absence of the WSCALE option (no scaling in either 2144 * direction) is encoded with index zero. 2145 * [WSCALE values histograms, Allman, 2012] 2146 * X 10 10 35 5 6 14 10% by host 2147 * X 11 4 5 5 18 49 3% by connections 2148 */ 2149 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2150 2151 /* 2152 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2153 * and good cryptographic properties. 2154 */ 2155 static uint32_t 2156 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2157 uint8_t *secbits, uintptr_t secmod) 2158 { 2159 SIPHASH_CTX ctx; 2160 uint32_t siphash[2]; 2161 2162 SipHash24_Init(&ctx); 2163 SipHash_SetKey(&ctx, secbits); 2164 switch (inc->inc_flags & INC_ISIPV6) { 2165 #ifdef INET 2166 case 0: 2167 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2168 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2169 break; 2170 #endif 2171 #ifdef INET6 2172 case INC_ISIPV6: 2173 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2174 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2175 break; 2176 #endif 2177 } 2178 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2179 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2180 SipHash_Update(&ctx, &irs, sizeof(irs)); 2181 SipHash_Update(&ctx, &flags, sizeof(flags)); 2182 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2183 SipHash_Final((u_int8_t *)&siphash, &ctx); 2184 2185 return (siphash[0] ^ siphash[1]); 2186 } 2187 2188 static tcp_seq 2189 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2190 { 2191 u_int i, secbit, wscale; 2192 uint32_t iss, hash; 2193 uint8_t *secbits; 2194 union syncookie cookie; 2195 2196 cookie.cookie = 0; 2197 2198 /* Map our computed MSS into the 3-bit index. */ 2199 for (i = nitems(tcp_sc_msstab) - 1; 2200 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2201 i--) 2202 ; 2203 cookie.flags.mss_idx = i; 2204 2205 /* 2206 * Map the send window scale into the 3-bit index but only if 2207 * the wscale option was received. 2208 */ 2209 if (sc->sc_flags & SCF_WINSCALE) { 2210 wscale = sc->sc_requested_s_scale; 2211 for (i = nitems(tcp_sc_wstab) - 1; 2212 tcp_sc_wstab[i] > wscale && i > 0; 2213 i--) 2214 ; 2215 cookie.flags.wscale_idx = i; 2216 } 2217 2218 /* Can we do SACK? */ 2219 if (sc->sc_flags & SCF_SACK) 2220 cookie.flags.sack_ok = 1; 2221 2222 /* Which of the two secrets to use. */ 2223 secbit = V_tcp_syncache.secret.oddeven & 0x1; 2224 cookie.flags.odd_even = secbit; 2225 2226 secbits = V_tcp_syncache.secret.key[secbit]; 2227 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2228 (uintptr_t)sch); 2229 2230 /* 2231 * Put the flags into the hash and XOR them to get better ISS number 2232 * variance. This doesn't enhance the cryptographic strength and is 2233 * done to prevent the 8 cookie bits from showing up directly on the 2234 * wire. 2235 */ 2236 iss = hash & ~0xff; 2237 iss |= cookie.cookie ^ (hash >> 24); 2238 2239 TCPSTAT_INC(tcps_sc_sendcookie); 2240 return (iss); 2241 } 2242 2243 static struct syncache * 2244 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2245 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2246 struct socket *lso, uint16_t port) 2247 { 2248 uint32_t hash; 2249 uint8_t *secbits; 2250 tcp_seq ack, seq; 2251 int wnd, wscale = 0; 2252 union syncookie cookie; 2253 2254 /* 2255 * Pull information out of SYN-ACK/ACK and revert sequence number 2256 * advances. 2257 */ 2258 ack = th->th_ack - 1; 2259 seq = th->th_seq - 1; 2260 2261 /* 2262 * Unpack the flags containing enough information to restore the 2263 * connection. 2264 */ 2265 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2266 2267 /* Which of the two secrets to use. */ 2268 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; 2269 2270 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2271 2272 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2273 if ((ack & ~0xff) != (hash & ~0xff)) 2274 return (NULL); 2275 2276 /* Fill in the syncache values. */ 2277 sc->sc_flags = 0; 2278 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2279 sc->sc_ipopts = NULL; 2280 2281 sc->sc_irs = seq; 2282 sc->sc_iss = ack; 2283 2284 switch (inc->inc_flags & INC_ISIPV6) { 2285 #ifdef INET 2286 case 0: 2287 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2288 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2289 break; 2290 #endif 2291 #ifdef INET6 2292 case INC_ISIPV6: 2293 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2294 sc->sc_flowlabel = 2295 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2296 break; 2297 #endif 2298 } 2299 2300 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2301 2302 /* We can simply recompute receive window scale we sent earlier. */ 2303 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2304 wscale++; 2305 2306 /* Only use wscale if it was enabled in the orignal SYN. */ 2307 if (cookie.flags.wscale_idx > 0) { 2308 sc->sc_requested_r_scale = wscale; 2309 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2310 sc->sc_flags |= SCF_WINSCALE; 2311 } 2312 2313 wnd = lso->sol_sbrcv_hiwat; 2314 wnd = imax(wnd, 0); 2315 wnd = imin(wnd, TCP_MAXWIN); 2316 sc->sc_wnd = wnd; 2317 2318 if (cookie.flags.sack_ok) 2319 sc->sc_flags |= SCF_SACK; 2320 2321 if (to->to_flags & TOF_TS) { 2322 sc->sc_flags |= SCF_TIMESTAMP; 2323 sc->sc_tsreflect = to->to_tsval; 2324 sc->sc_tsoff = tcp_new_ts_offset(inc); 2325 } 2326 2327 if (to->to_flags & TOF_SIGNATURE) 2328 sc->sc_flags |= SCF_SIGNATURE; 2329 2330 sc->sc_rxmits = 0; 2331 2332 sc->sc_port = port; 2333 2334 TCPSTAT_INC(tcps_sc_recvcookie); 2335 return (sc); 2336 } 2337 2338 #ifdef INVARIANTS 2339 static int 2340 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2341 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2342 struct socket *lso, uint16_t port) 2343 { 2344 struct syncache scs, *scx; 2345 char *s; 2346 2347 bzero(&scs, sizeof(scs)); 2348 scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port); 2349 2350 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2351 return (0); 2352 2353 if (scx != NULL) { 2354 if (sc->sc_peer_mss != scx->sc_peer_mss) 2355 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2356 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2357 2358 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2359 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2360 s, __func__, sc->sc_requested_r_scale, 2361 scx->sc_requested_r_scale); 2362 2363 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2364 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2365 s, __func__, sc->sc_requested_s_scale, 2366 scx->sc_requested_s_scale); 2367 2368 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2369 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2370 } 2371 2372 if (s != NULL) 2373 free(s, M_TCPLOG); 2374 return (0); 2375 } 2376 #endif /* INVARIANTS */ 2377 2378 static void 2379 syncookie_reseed(void *arg) 2380 { 2381 struct tcp_syncache *sc = arg; 2382 uint8_t *secbits; 2383 int secbit; 2384 2385 /* 2386 * Reseeding the secret doesn't have to be protected by a lock. 2387 * It only must be ensured that the new random values are visible 2388 * to all CPUs in a SMP environment. The atomic with release 2389 * semantics ensures that. 2390 */ 2391 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2392 secbits = sc->secret.key[secbit]; 2393 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2394 atomic_add_rel_int(&sc->secret.oddeven, 1); 2395 2396 /* Reschedule ourself. */ 2397 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2398 } 2399 2400 /* 2401 * We have overflowed a bucket. Let's pause dealing with the syncache. 2402 * This function will increment the bucketoverflow statistics appropriately 2403 * (once per pause when pausing is enabled; otherwise, once per overflow). 2404 */ 2405 static void 2406 syncache_pause(struct in_conninfo *inc) 2407 { 2408 time_t delta; 2409 const char *s; 2410 2411 /* XXX: 2412 * 2. Add sysctl read here so we don't get the benefit of this 2413 * change without the new sysctl. 2414 */ 2415 2416 /* 2417 * Try an unlocked read. If we already know that another thread 2418 * has activated the feature, there is no need to proceed. 2419 */ 2420 if (V_tcp_syncache.paused) 2421 return; 2422 2423 /* Are cookied enabled? If not, we can't pause. */ 2424 if (!V_tcp_syncookies) { 2425 TCPSTAT_INC(tcps_sc_bucketoverflow); 2426 return; 2427 } 2428 2429 /* 2430 * We may be the first thread to find an overflow. Get the lock 2431 * and evaluate if we need to take action. 2432 */ 2433 mtx_lock(&V_tcp_syncache.pause_mtx); 2434 if (V_tcp_syncache.paused) { 2435 mtx_unlock(&V_tcp_syncache.pause_mtx); 2436 return; 2437 } 2438 2439 /* Activate protection. */ 2440 V_tcp_syncache.paused = true; 2441 TCPSTAT_INC(tcps_sc_bucketoverflow); 2442 2443 /* 2444 * Determine the last backoff time. If we are seeing a re-newed 2445 * attack within that same time after last reactivating the syncache, 2446 * consider it an extension of the same attack. 2447 */ 2448 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; 2449 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { 2450 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { 2451 delta <<= 1; 2452 V_tcp_syncache.pause_backoff++; 2453 } 2454 } else { 2455 delta = TCP_SYNCACHE_PAUSE_TIME; 2456 V_tcp_syncache.pause_backoff = 0; 2457 } 2458 2459 /* Log a warning, including IP addresses, if able. */ 2460 if (inc != NULL) 2461 s = tcp_log_addrs(inc, NULL, NULL, NULL); 2462 else 2463 s = (const char *)NULL; 2464 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " 2465 "the next %lld seconds%s%s%s\n", (long long)delta, 2466 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", 2467 (s != NULL) ? ")" : ""); 2468 free(__DECONST(void *, s), M_TCPLOG); 2469 2470 /* Use the calculated delta to set a new pause time. */ 2471 V_tcp_syncache.pause_until = time_uptime + delta; 2472 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, 2473 &V_tcp_syncache); 2474 mtx_unlock(&V_tcp_syncache.pause_mtx); 2475 } 2476 2477 /* Evaluate whether we need to unpause. */ 2478 static void 2479 syncache_unpause(void *arg) 2480 { 2481 struct tcp_syncache *sc; 2482 time_t delta; 2483 2484 sc = arg; 2485 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); 2486 callout_deactivate(&sc->pause_co); 2487 2488 /* 2489 * Check to make sure we are not running early. If the pause 2490 * time has expired, then deactivate the protection. 2491 */ 2492 if ((delta = sc->pause_until - time_uptime) > 0) 2493 callout_schedule(&sc->pause_co, delta * hz); 2494 else 2495 sc->paused = false; 2496 } 2497 2498 /* 2499 * Exports the syncache entries to userland so that netstat can display 2500 * them alongside the other sockets. This function is intended to be 2501 * called only from tcp_pcblist. 2502 * 2503 * Due to concurrency on an active system, the number of pcbs exported 2504 * may have no relation to max_pcbs. max_pcbs merely indicates the 2505 * amount of space the caller allocated for this function to use. 2506 */ 2507 int 2508 syncache_pcblist(struct sysctl_req *req) 2509 { 2510 struct xtcpcb xt; 2511 struct syncache *sc; 2512 struct syncache_head *sch; 2513 int error, i; 2514 2515 bzero(&xt, sizeof(xt)); 2516 xt.xt_len = sizeof(xt); 2517 xt.t_state = TCPS_SYN_RECEIVED; 2518 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2519 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2520 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2521 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2522 2523 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 2524 sch = &V_tcp_syncache.hashbase[i]; 2525 SCH_LOCK(sch); 2526 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2527 if (sc->sc_cred != NULL && 2528 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2529 continue; 2530 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2531 xt.xt_inp.inp_vflag = INP_IPV6; 2532 else 2533 xt.xt_inp.inp_vflag = INP_IPV4; 2534 xt.xt_encaps_port = sc->sc_port; 2535 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2536 sizeof (struct in_conninfo)); 2537 error = SYSCTL_OUT(req, &xt, sizeof xt); 2538 if (error) { 2539 SCH_UNLOCK(sch); 2540 return (0); 2541 } 2542 } 2543 SCH_UNLOCK(sch); 2544 } 2545 2546 return (0); 2547 } 2548