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, "UI", 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 tcp_function_block *blk; 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 ((tp = tcp_newtcpcb(inp)) == NULL) { 806 in_pcbdetach(inp); 807 in_pcbfree(inp); 808 sodealloc(so); 809 goto allocfail; 810 } 811 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 812 #ifdef INET6 813 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 814 inp->inp_vflag &= ~INP_IPV4; 815 inp->inp_vflag |= INP_IPV6; 816 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 817 } else { 818 inp->inp_vflag &= ~INP_IPV6; 819 inp->inp_vflag |= INP_IPV4; 820 #endif 821 inp->inp_ip_ttl = sc->sc_ip_ttl; 822 inp->inp_ip_tos = sc->sc_ip_tos; 823 inp->inp_laddr = sc->sc_inc.inc_laddr; 824 #ifdef INET6 825 } 826 #endif 827 828 /* 829 * If there's an mbuf and it has a flowid, then let's initialise the 830 * inp with that particular flowid. 831 */ 832 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 833 inp->inp_flowid = m->m_pkthdr.flowid; 834 inp->inp_flowtype = M_HASHTYPE_GET(m); 835 #ifdef NUMA 836 inp->inp_numa_domain = m->m_pkthdr.numa_domain; 837 #endif 838 } 839 840 inp->inp_lport = sc->sc_inc.inc_lport; 841 #ifdef INET6 842 if (inp->inp_vflag & INP_IPV6PROTO) { 843 struct inpcb *oinp = sotoinpcb(lso); 844 845 /* 846 * Inherit socket options from the listening socket. 847 * Note that in6p_inputopts are not (and should not be) 848 * copied, since it stores previously received options and is 849 * used to detect if each new option is different than the 850 * previous one and hence should be passed to a user. 851 * If we copied in6p_inputopts, a user would not be able to 852 * receive options just after calling the accept system call. 853 */ 854 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 855 if (oinp->in6p_outputopts) 856 inp->in6p_outputopts = 857 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 858 inp->in6p_hops = oinp->in6p_hops; 859 } 860 861 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 862 struct sockaddr_in6 sin6; 863 864 sin6.sin6_family = AF_INET6; 865 sin6.sin6_len = sizeof(sin6); 866 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 867 sin6.sin6_port = sc->sc_inc.inc_fport; 868 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 869 INP_HASH_WLOCK(&V_tcbinfo); 870 error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false); 871 INP_HASH_WUNLOCK(&V_tcbinfo); 872 if (error != 0) 873 goto abort; 874 /* Override flowlabel from in6_pcbconnect. */ 875 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 876 inp->inp_flow |= sc->sc_flowlabel; 877 } 878 #endif /* INET6 */ 879 #if defined(INET) && defined(INET6) 880 else 881 #endif 882 #ifdef INET 883 { 884 struct sockaddr_in sin; 885 886 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 887 888 if (inp->inp_options == NULL) { 889 inp->inp_options = sc->sc_ipopts; 890 sc->sc_ipopts = NULL; 891 } 892 893 sin.sin_family = AF_INET; 894 sin.sin_len = sizeof(sin); 895 sin.sin_addr = sc->sc_inc.inc_faddr; 896 sin.sin_port = sc->sc_inc.inc_fport; 897 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 898 INP_HASH_WLOCK(&V_tcbinfo); 899 error = in_pcbconnect(inp, &sin, thread0.td_ucred, false); 900 INP_HASH_WUNLOCK(&V_tcbinfo); 901 if (error != 0) 902 goto abort; 903 } 904 #endif /* INET */ 905 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 906 /* Copy old policy into new socket's. */ 907 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 908 printf("syncache_socket: could not copy policy\n"); 909 #endif 910 tp->t_state = TCPS_SYN_RECEIVED; 911 tp->iss = sc->sc_iss; 912 tp->irs = sc->sc_irs; 913 tp->t_port = sc->sc_port; 914 tcp_rcvseqinit(tp); 915 tcp_sendseqinit(tp); 916 blk = sototcpcb(lso)->t_fb; 917 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 918 /* 919 * Our parents t_fb was not the default, 920 * we need to release our ref on tp->t_fb and 921 * pickup one on the new entry. 922 */ 923 struct tcp_function_block *rblk; 924 void *ptr = NULL; 925 926 rblk = find_and_ref_tcp_fb(blk); 927 KASSERT(rblk != NULL, 928 ("cannot find blk %p out of syncache?", blk)); 929 930 if (rblk->tfb_tcp_fb_init == NULL || 931 (*rblk->tfb_tcp_fb_init)(tp, &ptr) == 0) { 932 /* Release the old stack */ 933 if (tp->t_fb->tfb_tcp_fb_fini != NULL) 934 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 935 refcount_release(&tp->t_fb->tfb_refcnt); 936 /* Now set in all the pointers */ 937 tp->t_fb = rblk; 938 tp->t_fb_ptr = ptr; 939 } else { 940 /* 941 * Initialization failed. Release the reference count on 942 * the looked up default stack. 943 */ 944 refcount_release(&rblk->tfb_refcnt); 945 } 946 } 947 tp->snd_wl1 = sc->sc_irs; 948 tp->snd_max = tp->iss + 1; 949 tp->snd_nxt = tp->iss + 1; 950 tp->rcv_up = sc->sc_irs + 1; 951 tp->rcv_wnd = sc->sc_wnd; 952 tp->rcv_adv += tp->rcv_wnd; 953 tp->last_ack_sent = tp->rcv_nxt; 954 955 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 956 if (sc->sc_flags & SCF_NOOPT) 957 tp->t_flags |= TF_NOOPT; 958 else { 959 if (sc->sc_flags & SCF_WINSCALE) { 960 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 961 tp->snd_scale = sc->sc_requested_s_scale; 962 tp->request_r_scale = sc->sc_requested_r_scale; 963 } 964 if (sc->sc_flags & SCF_TIMESTAMP) { 965 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 966 tp->ts_recent = sc->sc_tsreflect; 967 tp->ts_recent_age = tcp_ts_getticks(); 968 tp->ts_offset = sc->sc_tsoff; 969 } 970 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 971 if (sc->sc_flags & SCF_SIGNATURE) 972 tp->t_flags |= TF_SIGNATURE; 973 #endif 974 if (sc->sc_flags & SCF_SACK) 975 tp->t_flags |= TF_SACK_PERMIT; 976 } 977 978 tcp_ecn_syncache_socket(tp, sc); 979 980 /* 981 * Set up MSS and get cached values from tcp_hostcache. 982 * This might overwrite some of the defaults we just set. 983 */ 984 tcp_mss(tp, sc->sc_peer_mss); 985 986 /* 987 * If the SYN,ACK was retransmitted, indicate that CWND to be 988 * limited to one segment in cc_conn_init(). 989 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 990 */ 991 if (sc->sc_rxmits > 1) 992 tp->snd_cwnd = 1; 993 994 #ifdef TCP_OFFLOAD 995 /* 996 * Allow a TOE driver to install its hooks. Note that we hold the 997 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 998 * new connection before the TOE driver has done its thing. 999 */ 1000 if (ADDED_BY_TOE(sc)) { 1001 struct toedev *tod = sc->sc_tod; 1002 1003 tod->tod_offload_socket(tod, sc->sc_todctx, so); 1004 } 1005 #endif 1006 #ifdef TCP_BLACKBOX 1007 /* 1008 * Inherit the log state from the listening socket, if 1009 * - the log state of the listening socket is not off and 1010 * - the listening socket was not auto selected from all sessions and 1011 * - a log id is not set on the listening socket. 1012 * This avoids inheriting a log state which was automatically set. 1013 */ 1014 if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) && 1015 ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) && 1016 (sototcpcb(lso)->t_lib == NULL)) { 1017 tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso))); 1018 } 1019 #endif 1020 /* 1021 * Copy and activate timers. 1022 */ 1023 tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime; 1024 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 1025 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 1026 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1027 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1028 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1029 1030 TCPSTAT_INC(tcps_accepts); 1031 TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN); 1032 1033 if (!solisten_enqueue(so, SS_ISCONNECTED)) 1034 tp->t_flags |= TF_SONOTCONN; 1035 1036 return (so); 1037 1038 allocfail: 1039 /* 1040 * Drop the connection; we will either send a RST or have the peer 1041 * retransmit its SYN again after its RTO and try again. 1042 */ 1043 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 1044 log(LOG_DEBUG, "%s; %s: Socket create failed " 1045 "due to limits or memory shortage\n", 1046 s, __func__); 1047 free(s, M_TCPLOG); 1048 } 1049 TCPSTAT_INC(tcps_listendrop); 1050 return (NULL); 1051 1052 abort: 1053 in_pcbdetach(inp); 1054 in_pcbfree(inp); 1055 sodealloc(so); 1056 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 1057 log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n", 1058 s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "", 1059 error); 1060 free(s, M_TCPLOG); 1061 } 1062 TCPSTAT_INC(tcps_listendrop); 1063 return (NULL); 1064 } 1065 1066 /* 1067 * This function gets called when we receive an ACK for a 1068 * socket in the LISTEN state. We look up the connection 1069 * in the syncache, and if its there, we pull it out of 1070 * the cache and turn it into a full-blown connection in 1071 * the SYN-RECEIVED state. 1072 * 1073 * On syncache_socket() success the newly created socket 1074 * has its underlying inp locked. 1075 */ 1076 int 1077 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1078 struct socket **lsop, struct mbuf *m, uint16_t port) 1079 { 1080 struct syncache *sc; 1081 struct syncache_head *sch; 1082 struct syncache scs; 1083 char *s; 1084 bool locked; 1085 1086 NET_EPOCH_ASSERT(); 1087 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1088 ("%s: can handle only ACK", __func__)); 1089 1090 if (syncache_cookiesonly()) { 1091 sc = NULL; 1092 sch = syncache_hashbucket(inc); 1093 locked = false; 1094 } else { 1095 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1096 locked = true; 1097 SCH_LOCK_ASSERT(sch); 1098 } 1099 1100 #ifdef INVARIANTS 1101 /* 1102 * Test code for syncookies comparing the syncache stored 1103 * values with the reconstructed values from the cookie. 1104 */ 1105 if (sc != NULL) 1106 syncookie_cmp(inc, sch, sc, th, to, *lsop, port); 1107 #endif 1108 1109 if (sc == NULL) { 1110 /* 1111 * There is no syncache entry, so see if this ACK is 1112 * a returning syncookie. To do this, first: 1113 * A. Check if syncookies are used in case of syncache 1114 * overflows 1115 * B. See if this socket has had a syncache entry dropped in 1116 * the recent past. We don't want to accept a bogus 1117 * syncookie if we've never received a SYN or accept it 1118 * twice. 1119 * C. check that the syncookie is valid. If it is, then 1120 * cobble up a fake syncache entry, and return. 1121 */ 1122 if (locked && !V_tcp_syncookies) { 1123 SCH_UNLOCK(sch); 1124 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1125 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1126 "segment rejected (syncookies disabled)\n", 1127 s, __func__); 1128 goto failed; 1129 } 1130 if (locked && !V_tcp_syncookiesonly && 1131 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1132 SCH_UNLOCK(sch); 1133 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1134 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1135 "segment rejected (no syncache entry)\n", 1136 s, __func__); 1137 goto failed; 1138 } 1139 bzero(&scs, sizeof(scs)); 1140 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port); 1141 if (locked) 1142 SCH_UNLOCK(sch); 1143 if (sc == NULL) { 1144 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1145 log(LOG_DEBUG, "%s; %s: Segment failed " 1146 "SYNCOOKIE authentication, segment rejected " 1147 "(probably spoofed)\n", s, __func__); 1148 goto failed; 1149 } 1150 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1151 /* If received ACK has MD5 signature, check it. */ 1152 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1153 (!TCPMD5_ENABLED() || 1154 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1155 /* Drop the ACK. */ 1156 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1157 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1158 "MD5 signature doesn't match.\n", 1159 s, __func__); 1160 free(s, M_TCPLOG); 1161 } 1162 TCPSTAT_INC(tcps_sig_err_sigopt); 1163 return (-1); /* Do not send RST */ 1164 } 1165 #endif /* TCP_SIGNATURE */ 1166 TCPSTATES_INC(TCPS_SYN_RECEIVED); 1167 } else { 1168 if (sc->sc_port != port) { 1169 SCH_UNLOCK(sch); 1170 return (0); 1171 } 1172 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1173 /* 1174 * If listening socket requested TCP digests, check that 1175 * received ACK has signature and it is correct. 1176 * If not, drop the ACK and leave sc entry in th cache, 1177 * because SYN was received with correct signature. 1178 */ 1179 if (sc->sc_flags & SCF_SIGNATURE) { 1180 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1181 /* No signature */ 1182 TCPSTAT_INC(tcps_sig_err_nosigopt); 1183 SCH_UNLOCK(sch); 1184 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1185 log(LOG_DEBUG, "%s; %s: Segment " 1186 "rejected, MD5 signature wasn't " 1187 "provided.\n", s, __func__); 1188 free(s, M_TCPLOG); 1189 } 1190 return (-1); /* Do not send RST */ 1191 } 1192 if (!TCPMD5_ENABLED() || 1193 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1194 /* Doesn't match or no SA */ 1195 SCH_UNLOCK(sch); 1196 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1197 log(LOG_DEBUG, "%s; %s: Segment " 1198 "rejected, MD5 signature doesn't " 1199 "match.\n", s, __func__); 1200 free(s, M_TCPLOG); 1201 } 1202 return (-1); /* Do not send RST */ 1203 } 1204 } 1205 #endif /* TCP_SIGNATURE */ 1206 1207 /* 1208 * RFC 7323 PAWS: If we have a timestamp on this segment and 1209 * it's less than ts_recent, drop it. 1210 * XXXMT: RFC 7323 also requires to send an ACK. 1211 * In tcp_input.c this is only done for TCP segments 1212 * with user data, so be consistent here and just drop 1213 * the segment. 1214 */ 1215 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1216 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1217 SCH_UNLOCK(sch); 1218 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1219 log(LOG_DEBUG, 1220 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1221 "segment dropped\n", s, __func__, 1222 to->to_tsval, sc->sc_tsreflect); 1223 free(s, M_TCPLOG); 1224 } 1225 return (-1); /* Do not send RST */ 1226 } 1227 1228 /* 1229 * If timestamps were not negotiated during SYN/ACK and a 1230 * segment with a timestamp is received, ignore the 1231 * timestamp and process the packet normally. 1232 * See section 3.2 of RFC 7323. 1233 */ 1234 if (!(sc->sc_flags & SCF_TIMESTAMP) && 1235 (to->to_flags & TOF_TS)) { 1236 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1237 log(LOG_DEBUG, "%s; %s: Timestamp not " 1238 "expected, segment processed normally\n", 1239 s, __func__); 1240 free(s, M_TCPLOG); 1241 s = NULL; 1242 } 1243 } 1244 1245 /* 1246 * If timestamps were negotiated during SYN/ACK and a 1247 * segment without a timestamp is received, silently drop 1248 * the segment, unless the missing timestamps are tolerated. 1249 * See section 3.2 of RFC 7323. 1250 */ 1251 if ((sc->sc_flags & SCF_TIMESTAMP) && 1252 !(to->to_flags & TOF_TS)) { 1253 if (V_tcp_tolerate_missing_ts) { 1254 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1255 log(LOG_DEBUG, 1256 "%s; %s: Timestamp missing, " 1257 "segment processed normally\n", 1258 s, __func__); 1259 free(s, M_TCPLOG); 1260 } 1261 } else { 1262 SCH_UNLOCK(sch); 1263 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1264 log(LOG_DEBUG, 1265 "%s; %s: Timestamp missing, " 1266 "segment silently dropped\n", 1267 s, __func__); 1268 free(s, M_TCPLOG); 1269 } 1270 return (-1); /* Do not send RST */ 1271 } 1272 } 1273 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1274 sch->sch_length--; 1275 #ifdef TCP_OFFLOAD 1276 if (ADDED_BY_TOE(sc)) { 1277 struct toedev *tod = sc->sc_tod; 1278 1279 tod->tod_syncache_removed(tod, sc->sc_todctx); 1280 } 1281 #endif 1282 SCH_UNLOCK(sch); 1283 } 1284 1285 /* 1286 * Segment validation: 1287 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1288 */ 1289 if (th->th_ack != sc->sc_iss + 1) { 1290 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1291 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1292 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1293 goto failed; 1294 } 1295 1296 /* 1297 * The SEQ must fall in the window starting at the received 1298 * initial receive sequence number + 1 (the SYN). 1299 */ 1300 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1301 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1302 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1303 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1304 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1305 goto failed; 1306 } 1307 1308 *lsop = syncache_socket(sc, *lsop, m); 1309 1310 if (__predict_false(*lsop == NULL)) { 1311 TCPSTAT_INC(tcps_sc_aborted); 1312 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1313 } else 1314 TCPSTAT_INC(tcps_sc_completed); 1315 1316 /* how do we find the inp for the new socket? */ 1317 if (sc != &scs) 1318 syncache_free(sc); 1319 return (1); 1320 failed: 1321 if (sc != NULL) { 1322 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1323 if (sc != &scs) 1324 syncache_free(sc); 1325 } 1326 if (s != NULL) 1327 free(s, M_TCPLOG); 1328 *lsop = NULL; 1329 return (0); 1330 } 1331 1332 static struct socket * 1333 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m, 1334 uint64_t response_cookie) 1335 { 1336 struct inpcb *inp; 1337 struct tcpcb *tp; 1338 unsigned int *pending_counter; 1339 struct socket *so; 1340 1341 NET_EPOCH_ASSERT(); 1342 1343 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending; 1344 so = syncache_socket(sc, lso, m); 1345 if (so == NULL) { 1346 TCPSTAT_INC(tcps_sc_aborted); 1347 atomic_subtract_int(pending_counter, 1); 1348 } else { 1349 soisconnected(so); 1350 inp = sotoinpcb(so); 1351 tp = intotcpcb(inp); 1352 tp->t_flags |= TF_FASTOPEN; 1353 tp->t_tfo_cookie.server = response_cookie; 1354 tp->snd_max = tp->iss; 1355 tp->snd_nxt = tp->iss; 1356 tp->t_tfo_pending = pending_counter; 1357 TCPSTATES_INC(TCPS_SYN_RECEIVED); 1358 TCPSTAT_INC(tcps_sc_completed); 1359 } 1360 1361 return (so); 1362 } 1363 1364 /* 1365 * Given a LISTEN socket and an inbound SYN request, add 1366 * this to the syn cache, and send back a segment: 1367 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1368 * to the source. 1369 * 1370 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1371 * Doing so would require that we hold onto the data and deliver it 1372 * to the application. However, if we are the target of a SYN-flood 1373 * DoS attack, an attacker could send data which would eventually 1374 * consume all available buffer space if it were ACKed. By not ACKing 1375 * the data, we avoid this DoS scenario. 1376 * 1377 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1378 * cookie is processed and a new socket is created. In this case, any data 1379 * accompanying the SYN will be queued to the socket by tcp_input() and will 1380 * be ACKed either when the application sends response data or the delayed 1381 * ACK timer expires, whichever comes first. 1382 */ 1383 struct socket * 1384 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1385 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod, 1386 void *todctx, uint8_t iptos, uint16_t port) 1387 { 1388 struct tcpcb *tp; 1389 struct socket *rv = NULL; 1390 struct syncache *sc = NULL; 1391 struct syncache_head *sch; 1392 struct mbuf *ipopts = NULL; 1393 u_int ltflags; 1394 int win, ip_ttl, ip_tos; 1395 char *s; 1396 #ifdef INET6 1397 int autoflowlabel = 0; 1398 #endif 1399 #ifdef MAC 1400 struct label *maclabel; 1401 #endif 1402 struct syncache scs; 1403 struct ucred *cred; 1404 uint64_t tfo_response_cookie; 1405 unsigned int *tfo_pending = NULL; 1406 int tfo_cookie_valid = 0; 1407 int tfo_response_cookie_valid = 0; 1408 bool locked; 1409 1410 INP_RLOCK_ASSERT(inp); /* listen socket */ 1411 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1412 ("%s: unexpected tcp flags", __func__)); 1413 1414 /* 1415 * Combine all so/tp operations very early to drop the INP lock as 1416 * soon as possible. 1417 */ 1418 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1419 tp = sototcpcb(so); 1420 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred); 1421 1422 #ifdef INET6 1423 if (inc->inc_flags & INC_ISIPV6) { 1424 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1425 autoflowlabel = 1; 1426 } 1427 ip_ttl = in6_selecthlim(inp, NULL); 1428 if ((inp->in6p_outputopts == NULL) || 1429 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1430 ip_tos = 0; 1431 } else { 1432 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1433 } 1434 } 1435 #endif 1436 #if defined(INET6) && defined(INET) 1437 else 1438 #endif 1439 #ifdef INET 1440 { 1441 ip_ttl = inp->inp_ip_ttl; 1442 ip_tos = inp->inp_ip_tos; 1443 } 1444 #endif 1445 win = so->sol_sbrcv_hiwat; 1446 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1447 1448 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1449 (tp->t_tfo_pending != NULL) && 1450 (to->to_flags & TOF_FASTOPEN)) { 1451 /* 1452 * Limit the number of pending TFO connections to 1453 * approximately half of the queue limit. This prevents TFO 1454 * SYN floods from starving the service by filling the 1455 * listen queue with bogus TFO connections. 1456 */ 1457 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1458 (so->sol_qlimit / 2)) { 1459 int result; 1460 1461 result = tcp_fastopen_check_cookie(inc, 1462 to->to_tfo_cookie, to->to_tfo_len, 1463 &tfo_response_cookie); 1464 tfo_cookie_valid = (result > 0); 1465 tfo_response_cookie_valid = (result >= 0); 1466 } 1467 1468 /* 1469 * Remember the TFO pending counter as it will have to be 1470 * decremented below if we don't make it to syncache_tfo_expand(). 1471 */ 1472 tfo_pending = tp->t_tfo_pending; 1473 } 1474 1475 #ifdef MAC 1476 if (mac_syncache_init(&maclabel) != 0) { 1477 INP_RUNLOCK(inp); 1478 goto done; 1479 } else 1480 mac_syncache_create(maclabel, inp); 1481 #endif 1482 if (!tfo_cookie_valid) 1483 INP_RUNLOCK(inp); 1484 1485 /* 1486 * Remember the IP options, if any. 1487 */ 1488 #ifdef INET6 1489 if (!(inc->inc_flags & INC_ISIPV6)) 1490 #endif 1491 #ifdef INET 1492 ipopts = (m) ? ip_srcroute(m) : NULL; 1493 #else 1494 ipopts = NULL; 1495 #endif 1496 1497 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1498 /* 1499 * When the socket is TCP-MD5 enabled check that, 1500 * - a signed packet is valid 1501 * - a non-signed packet does not have a security association 1502 * 1503 * If a signed packet fails validation or a non-signed packet has a 1504 * security association, the packet will be dropped. 1505 */ 1506 if (ltflags & TF_SIGNATURE) { 1507 if (to->to_flags & TOF_SIGNATURE) { 1508 if (!TCPMD5_ENABLED() || 1509 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1510 goto done; 1511 } else { 1512 if (TCPMD5_ENABLED() && 1513 TCPMD5_INPUT(m, NULL, NULL) != ENOENT) 1514 goto done; 1515 } 1516 } else if (to->to_flags & TOF_SIGNATURE) 1517 goto done; 1518 #endif /* TCP_SIGNATURE */ 1519 /* 1520 * See if we already have an entry for this connection. 1521 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1522 * 1523 * XXX: should the syncache be re-initialized with the contents 1524 * of the new SYN here (which may have different options?) 1525 * 1526 * XXX: We do not check the sequence number to see if this is a 1527 * real retransmit or a new connection attempt. The question is 1528 * how to handle such a case; either ignore it as spoofed, or 1529 * drop the current entry and create a new one? 1530 */ 1531 if (syncache_cookiesonly()) { 1532 sc = NULL; 1533 sch = syncache_hashbucket(inc); 1534 locked = false; 1535 } else { 1536 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1537 locked = true; 1538 SCH_LOCK_ASSERT(sch); 1539 } 1540 if (sc != NULL) { 1541 if (tfo_cookie_valid) 1542 INP_RUNLOCK(inp); 1543 TCPSTAT_INC(tcps_sc_dupsyn); 1544 if (ipopts) { 1545 /* 1546 * If we were remembering a previous source route, 1547 * forget it and use the new one we've been given. 1548 */ 1549 if (sc->sc_ipopts) 1550 (void) m_free(sc->sc_ipopts); 1551 sc->sc_ipopts = ipopts; 1552 } 1553 /* 1554 * Update timestamp if present. 1555 */ 1556 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1557 sc->sc_tsreflect = to->to_tsval; 1558 else 1559 sc->sc_flags &= ~SCF_TIMESTAMP; 1560 /* 1561 * Adjust ECN response if needed, e.g. different 1562 * IP ECN field, or a fallback by the remote host. 1563 */ 1564 if (sc->sc_flags & SCF_ECN_MASK) { 1565 sc->sc_flags &= ~SCF_ECN_MASK; 1566 sc->sc_flags = tcp_ecn_syncache_add(tcp_get_flags(th), iptos); 1567 } 1568 #ifdef MAC 1569 /* 1570 * Since we have already unconditionally allocated label 1571 * storage, free it up. The syncache entry will already 1572 * have an initialized label we can use. 1573 */ 1574 mac_syncache_destroy(&maclabel); 1575 #endif 1576 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1577 /* Retransmit SYN|ACK and reset retransmit count. */ 1578 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1579 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1580 "resetting timer and retransmitting SYN|ACK\n", 1581 s, __func__); 1582 free(s, M_TCPLOG); 1583 } 1584 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1585 sc->sc_rxmits = 0; 1586 syncache_timeout(sc, sch, 1); 1587 TCPSTAT_INC(tcps_sndacks); 1588 TCPSTAT_INC(tcps_sndtotal); 1589 } 1590 SCH_UNLOCK(sch); 1591 goto donenoprobe; 1592 } 1593 1594 if (tfo_cookie_valid) { 1595 bzero(&scs, sizeof(scs)); 1596 sc = &scs; 1597 goto skip_alloc; 1598 } 1599 1600 /* 1601 * Skip allocating a syncache entry if we are just going to discard 1602 * it later. 1603 */ 1604 if (!locked) { 1605 bzero(&scs, sizeof(scs)); 1606 sc = &scs; 1607 } else 1608 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1609 if (sc == NULL) { 1610 /* 1611 * The zone allocator couldn't provide more entries. 1612 * Treat this as if the cache was full; drop the oldest 1613 * entry and insert the new one. 1614 */ 1615 TCPSTAT_INC(tcps_sc_zonefail); 1616 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1617 sch->sch_last_overflow = time_uptime; 1618 syncache_drop(sc, sch); 1619 syncache_pause(inc); 1620 } 1621 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1622 if (sc == NULL) { 1623 if (V_tcp_syncookies) { 1624 bzero(&scs, sizeof(scs)); 1625 sc = &scs; 1626 } else { 1627 KASSERT(locked, 1628 ("%s: bucket unexpectedly unlocked", 1629 __func__)); 1630 SCH_UNLOCK(sch); 1631 if (ipopts) 1632 (void) m_free(ipopts); 1633 goto done; 1634 } 1635 } 1636 } 1637 1638 skip_alloc: 1639 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1640 sc->sc_tfo_cookie = &tfo_response_cookie; 1641 1642 /* 1643 * Fill in the syncache values. 1644 */ 1645 #ifdef MAC 1646 sc->sc_label = maclabel; 1647 #endif 1648 sc->sc_cred = cred; 1649 sc->sc_port = port; 1650 cred = NULL; 1651 sc->sc_ipopts = ipopts; 1652 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1653 sc->sc_ip_tos = ip_tos; 1654 sc->sc_ip_ttl = ip_ttl; 1655 #ifdef TCP_OFFLOAD 1656 sc->sc_tod = tod; 1657 sc->sc_todctx = todctx; 1658 #endif 1659 sc->sc_irs = th->th_seq; 1660 sc->sc_flags = 0; 1661 sc->sc_flowlabel = 0; 1662 1663 /* 1664 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1665 * win was derived from socket earlier in the function. 1666 */ 1667 win = imax(win, 0); 1668 win = imin(win, TCP_MAXWIN); 1669 sc->sc_wnd = win; 1670 1671 if (V_tcp_do_rfc1323 && 1672 !(ltflags & TF_NOOPT)) { 1673 /* 1674 * A timestamp received in a SYN makes 1675 * it ok to send timestamp requests and replies. 1676 */ 1677 if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) { 1678 sc->sc_tsreflect = to->to_tsval; 1679 sc->sc_flags |= SCF_TIMESTAMP; 1680 sc->sc_tsoff = tcp_new_ts_offset(inc); 1681 } 1682 if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) { 1683 int wscale = 0; 1684 1685 /* 1686 * Pick the smallest possible scaling factor that 1687 * will still allow us to scale up to sb_max, aka 1688 * kern.ipc.maxsockbuf. 1689 * 1690 * We do this because there are broken firewalls that 1691 * will corrupt the window scale option, leading to 1692 * the other endpoint believing that our advertised 1693 * window is unscaled. At scale factors larger than 1694 * 5 the unscaled window will drop below 1500 bytes, 1695 * leading to serious problems when traversing these 1696 * broken firewalls. 1697 * 1698 * With the default maxsockbuf of 256K, a scale factor 1699 * of 3 will be chosen by this algorithm. Those who 1700 * choose a larger maxsockbuf should watch out 1701 * for the compatibility problems mentioned above. 1702 * 1703 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1704 * or <SYN,ACK>) segment itself is never scaled. 1705 */ 1706 while (wscale < TCP_MAX_WINSHIFT && 1707 (TCP_MAXWIN << wscale) < sb_max) 1708 wscale++; 1709 sc->sc_requested_r_scale = wscale; 1710 sc->sc_requested_s_scale = to->to_wscale; 1711 sc->sc_flags |= SCF_WINSCALE; 1712 } 1713 } 1714 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1715 /* 1716 * If incoming packet has an MD5 signature, flag this in the 1717 * syncache so that syncache_respond() will do the right thing 1718 * with the SYN+ACK. 1719 */ 1720 if (to->to_flags & TOF_SIGNATURE) 1721 sc->sc_flags |= SCF_SIGNATURE; 1722 #endif /* TCP_SIGNATURE */ 1723 if (to->to_flags & TOF_SACKPERM) 1724 sc->sc_flags |= SCF_SACK; 1725 if (to->to_flags & TOF_MSS) 1726 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1727 if (ltflags & TF_NOOPT) 1728 sc->sc_flags |= SCF_NOOPT; 1729 /* ECN Handshake */ 1730 if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0) 1731 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos); 1732 1733 if (V_tcp_syncookies) 1734 sc->sc_iss = syncookie_generate(sch, sc); 1735 else 1736 sc->sc_iss = arc4random(); 1737 #ifdef INET6 1738 if (autoflowlabel) { 1739 if (V_tcp_syncookies) 1740 sc->sc_flowlabel = sc->sc_iss; 1741 else 1742 sc->sc_flowlabel = ip6_randomflowlabel(); 1743 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1744 } 1745 #endif 1746 if (locked) 1747 SCH_UNLOCK(sch); 1748 1749 if (tfo_cookie_valid) { 1750 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie); 1751 /* INP_RUNLOCK(inp) will be performed by the caller */ 1752 goto tfo_expanded; 1753 } 1754 1755 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1756 /* 1757 * Do a standard 3-way handshake. 1758 */ 1759 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1760 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1761 syncache_free(sc); 1762 else if (sc != &scs) 1763 syncache_insert(sc, sch); /* locks and unlocks sch */ 1764 TCPSTAT_INC(tcps_sndacks); 1765 TCPSTAT_INC(tcps_sndtotal); 1766 } else { 1767 if (sc != &scs) 1768 syncache_free(sc); 1769 TCPSTAT_INC(tcps_sc_dropped); 1770 } 1771 goto donenoprobe; 1772 1773 done: 1774 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1775 donenoprobe: 1776 if (m) 1777 m_freem(m); 1778 /* 1779 * If tfo_pending is not NULL here, then a TFO SYN that did not 1780 * result in a new socket was processed and the associated pending 1781 * counter has not yet been decremented. All such TFO processing paths 1782 * transit this point. 1783 */ 1784 if (tfo_pending != NULL) 1785 tcp_fastopen_decrement_counter(tfo_pending); 1786 1787 tfo_expanded: 1788 if (cred != NULL) 1789 crfree(cred); 1790 #ifdef MAC 1791 if (sc == &scs) 1792 mac_syncache_destroy(&maclabel); 1793 #endif 1794 return (rv); 1795 } 1796 1797 /* 1798 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1799 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1800 */ 1801 static int 1802 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags) 1803 { 1804 struct ip *ip = NULL; 1805 struct mbuf *m; 1806 struct tcphdr *th = NULL; 1807 struct udphdr *udp = NULL; 1808 int optlen, error = 0; /* Make compiler happy */ 1809 u_int16_t hlen, tlen, mssopt, ulen; 1810 struct tcpopt to; 1811 #ifdef INET6 1812 struct ip6_hdr *ip6 = NULL; 1813 #endif 1814 1815 NET_EPOCH_ASSERT(); 1816 1817 hlen = 1818 #ifdef INET6 1819 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1820 #endif 1821 sizeof(struct ip); 1822 tlen = hlen + sizeof(struct tcphdr); 1823 if (sc->sc_port) { 1824 tlen += sizeof(struct udphdr); 1825 } 1826 /* Determine MSS we advertize to other end of connection. */ 1827 mssopt = tcp_mssopt(&sc->sc_inc); 1828 if (sc->sc_port) 1829 mssopt -= V_tcp_udp_tunneling_overhead; 1830 mssopt = max(mssopt, V_tcp_minmss); 1831 1832 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1833 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1834 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + " 1835 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port, 1836 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN)); 1837 1838 /* Create the IP+TCP header from scratch. */ 1839 m = m_gethdr(M_NOWAIT, MT_DATA); 1840 if (m == NULL) 1841 return (ENOBUFS); 1842 #ifdef MAC 1843 mac_syncache_create_mbuf(sc->sc_label, m); 1844 #endif 1845 m->m_data += max_linkhdr; 1846 m->m_len = tlen; 1847 m->m_pkthdr.len = tlen; 1848 m->m_pkthdr.rcvif = NULL; 1849 1850 #ifdef INET6 1851 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1852 ip6 = mtod(m, struct ip6_hdr *); 1853 ip6->ip6_vfc = IPV6_VERSION; 1854 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1855 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1856 ip6->ip6_plen = htons(tlen - hlen); 1857 /* ip6_hlim is set after checksum */ 1858 /* Zero out traffic class and flow label. */ 1859 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1860 ip6->ip6_flow |= sc->sc_flowlabel; 1861 if (sc->sc_port != 0) { 1862 ip6->ip6_nxt = IPPROTO_UDP; 1863 udp = (struct udphdr *)(ip6 + 1); 1864 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1865 udp->uh_dport = sc->sc_port; 1866 ulen = (tlen - sizeof(struct ip6_hdr)); 1867 th = (struct tcphdr *)(udp + 1); 1868 } else { 1869 ip6->ip6_nxt = IPPROTO_TCP; 1870 th = (struct tcphdr *)(ip6 + 1); 1871 } 1872 ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN); 1873 } 1874 #endif 1875 #if defined(INET6) && defined(INET) 1876 else 1877 #endif 1878 #ifdef INET 1879 { 1880 ip = mtod(m, struct ip *); 1881 ip->ip_v = IPVERSION; 1882 ip->ip_hl = sizeof(struct ip) >> 2; 1883 ip->ip_len = htons(tlen); 1884 ip->ip_id = 0; 1885 ip->ip_off = 0; 1886 ip->ip_sum = 0; 1887 ip->ip_src = sc->sc_inc.inc_laddr; 1888 ip->ip_dst = sc->sc_inc.inc_faddr; 1889 ip->ip_ttl = sc->sc_ip_ttl; 1890 ip->ip_tos = sc->sc_ip_tos; 1891 1892 /* 1893 * See if we should do MTU discovery. Route lookups are 1894 * expensive, so we will only unset the DF bit if: 1895 * 1896 * 1) path_mtu_discovery is disabled 1897 * 2) the SCF_UNREACH flag has been set 1898 */ 1899 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1900 ip->ip_off |= htons(IP_DF); 1901 if (sc->sc_port == 0) { 1902 ip->ip_p = IPPROTO_TCP; 1903 th = (struct tcphdr *)(ip + 1); 1904 } else { 1905 ip->ip_p = IPPROTO_UDP; 1906 udp = (struct udphdr *)(ip + 1); 1907 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1908 udp->uh_dport = sc->sc_port; 1909 ulen = (tlen - sizeof(struct ip)); 1910 th = (struct tcphdr *)(udp + 1); 1911 } 1912 } 1913 #endif /* INET */ 1914 th->th_sport = sc->sc_inc.inc_lport; 1915 th->th_dport = sc->sc_inc.inc_fport; 1916 1917 if (flags & TH_SYN) 1918 th->th_seq = htonl(sc->sc_iss); 1919 else 1920 th->th_seq = htonl(sc->sc_iss + 1); 1921 th->th_ack = htonl(sc->sc_irs + 1); 1922 th->th_off = sizeof(struct tcphdr) >> 2; 1923 th->th_win = htons(sc->sc_wnd); 1924 th->th_urp = 0; 1925 1926 flags = tcp_ecn_syncache_respond(flags, sc); 1927 tcp_set_flags(th, flags); 1928 1929 /* Tack on the TCP options. */ 1930 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1931 to.to_flags = 0; 1932 1933 if (flags & TH_SYN) { 1934 to.to_mss = mssopt; 1935 to.to_flags = TOF_MSS; 1936 if (sc->sc_flags & SCF_WINSCALE) { 1937 to.to_wscale = sc->sc_requested_r_scale; 1938 to.to_flags |= TOF_SCALE; 1939 } 1940 if (sc->sc_flags & SCF_SACK) 1941 to.to_flags |= TOF_SACKPERM; 1942 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1943 if (sc->sc_flags & SCF_SIGNATURE) 1944 to.to_flags |= TOF_SIGNATURE; 1945 #endif 1946 if (sc->sc_tfo_cookie) { 1947 to.to_flags |= TOF_FASTOPEN; 1948 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1949 to.to_tfo_cookie = sc->sc_tfo_cookie; 1950 /* don't send cookie again when retransmitting response */ 1951 sc->sc_tfo_cookie = NULL; 1952 } 1953 } 1954 if (sc->sc_flags & SCF_TIMESTAMP) { 1955 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1956 to.to_tsecr = sc->sc_tsreflect; 1957 to.to_flags |= TOF_TS; 1958 } 1959 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1960 1961 /* Adjust headers by option size. */ 1962 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1963 m->m_len += optlen; 1964 m->m_pkthdr.len += optlen; 1965 #ifdef INET6 1966 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1967 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1968 else 1969 #endif 1970 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1971 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1972 if (sc->sc_flags & SCF_SIGNATURE) { 1973 KASSERT(to.to_flags & TOF_SIGNATURE, 1974 ("tcp_addoptions() didn't set tcp_signature")); 1975 1976 /* NOTE: to.to_signature is inside of mbuf */ 1977 if (!TCPMD5_ENABLED() || 1978 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1979 m_freem(m); 1980 return (EACCES); 1981 } 1982 } 1983 #endif 1984 } else 1985 optlen = 0; 1986 1987 if (udp) { 1988 ulen += optlen; 1989 udp->uh_ulen = htons(ulen); 1990 } 1991 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1992 /* 1993 * If we have peer's SYN and it has a flowid, then let's assign it to 1994 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1995 * to SYN|ACK due to lack of inp here. 1996 */ 1997 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1998 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1999 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 2000 } 2001 #ifdef INET6 2002 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 2003 if (sc->sc_port) { 2004 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 2005 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2006 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, 2007 IPPROTO_UDP, 0); 2008 th->th_sum = htons(0); 2009 } else { 2010 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 2011 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2012 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 2013 IPPROTO_TCP, 0); 2014 } 2015 ip6->ip6_hlim = sc->sc_ip_ttl; 2016 #ifdef TCP_OFFLOAD 2017 if (ADDED_BY_TOE(sc)) { 2018 struct toedev *tod = sc->sc_tod; 2019 2020 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2021 2022 return (error); 2023 } 2024 #endif 2025 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 2026 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 2027 } 2028 #endif 2029 #if defined(INET6) && defined(INET) 2030 else 2031 #endif 2032 #ifdef INET 2033 { 2034 if (sc->sc_port) { 2035 m->m_pkthdr.csum_flags = CSUM_UDP; 2036 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2037 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 2038 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 2039 th->th_sum = htons(0); 2040 } else { 2041 m->m_pkthdr.csum_flags = CSUM_TCP; 2042 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2043 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 2044 htons(tlen + optlen - hlen + IPPROTO_TCP)); 2045 } 2046 #ifdef TCP_OFFLOAD 2047 if (ADDED_BY_TOE(sc)) { 2048 struct toedev *tod = sc->sc_tod; 2049 2050 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2051 2052 return (error); 2053 } 2054 #endif 2055 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 2056 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 2057 } 2058 #endif 2059 return (error); 2060 } 2061 2062 /* 2063 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 2064 * that exceed the capacity of the syncache by avoiding the storage of any 2065 * of the SYNs we receive. Syncookies defend against blind SYN flooding 2066 * attacks where the attacker does not have access to our responses. 2067 * 2068 * Syncookies encode and include all necessary information about the 2069 * connection setup within the SYN|ACK that we send back. That way we 2070 * can avoid keeping any local state until the ACK to our SYN|ACK returns 2071 * (if ever). Normally the syncache and syncookies are running in parallel 2072 * with the latter taking over when the former is exhausted. When matching 2073 * syncache entry is found the syncookie is ignored. 2074 * 2075 * The only reliable information persisting the 3WHS is our initial sequence 2076 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 2077 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 2078 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 2079 * returns and signifies a legitimate connection if it matches the ACK. 2080 * 2081 * The available space of 32 bits to store the hash and to encode the SYN 2082 * option information is very tight and we should have at least 24 bits for 2083 * the MAC to keep the number of guesses by blind spoofing reasonably high. 2084 * 2085 * SYN option information we have to encode to fully restore a connection: 2086 * MSS: is imporant to chose an optimal segment size to avoid IP level 2087 * fragmentation along the path. The common MSS values can be encoded 2088 * in a 3-bit table. Uncommon values are captured by the next lower value 2089 * in the table leading to a slight increase in packetization overhead. 2090 * WSCALE: is necessary to allow large windows to be used for high delay- 2091 * bandwidth product links. Not scaling the window when it was initially 2092 * negotiated is bad for performance as lack of scaling further decreases 2093 * the apparent available send window. We only need to encode the WSCALE 2094 * we received from the remote end. Our end can be recalculated at any 2095 * time. The common WSCALE values can be encoded in a 3-bit table. 2096 * Uncommon values are captured by the next lower value in the table 2097 * making us under-estimate the available window size halving our 2098 * theoretically possible maximum throughput for that connection. 2099 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 2100 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 2101 * that are included in all segments on a connection. We enable them when 2102 * the ACK has them. 2103 * 2104 * Security of syncookies and attack vectors: 2105 * 2106 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 2107 * together with the gloabl secret to make it unique per connection attempt. 2108 * Thus any change of any of those parameters results in a different MAC output 2109 * in an unpredictable way unless a collision is encountered. 24 bits of the 2110 * MAC are embedded into the ISS. 2111 * 2112 * To prevent replay attacks two rotating global secrets are updated with a 2113 * new random value every 15 seconds. The life-time of a syncookie is thus 2114 * 15-30 seconds. 2115 * 2116 * Vector 1: Attacking the secret. This requires finding a weakness in the 2117 * MAC itself or the way it is used here. The attacker can do a chosen plain 2118 * text attack by varying and testing the all parameters under his control. 2119 * The strength depends on the size and randomness of the secret, and the 2120 * cryptographic security of the MAC function. Due to the constant updating 2121 * of the secret the attacker has at most 29.999 seconds to find the secret 2122 * and launch spoofed connections. After that he has to start all over again. 2123 * 2124 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 2125 * size an average of 4,823 attempts are required for a 50% chance of success 2126 * to spoof a single syncookie (birthday collision paradox). However the 2127 * attacker is blind and doesn't know if one of his attempts succeeded unless 2128 * he has a side channel to interfere success from. A single connection setup 2129 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 2130 * This many attempts are required for each one blind spoofed connection. For 2131 * every additional spoofed connection he has to launch another N attempts. 2132 * Thus for a sustained rate 100 spoofed connections per second approximately 2133 * 1,800,000 packets per second would have to be sent. 2134 * 2135 * NB: The MAC function should be fast so that it doesn't become a CPU 2136 * exhaustion attack vector itself. 2137 * 2138 * References: 2139 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 2140 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 2141 * http://cr.yp.to/syncookies.html (overview) 2142 * http://cr.yp.to/syncookies/archive (details) 2143 * 2144 * 2145 * Schematic construction of a syncookie enabled Initial Sequence Number: 2146 * 0 1 2 3 2147 * 12345678901234567890123456789012 2148 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2149 * 2150 * x 24 MAC (truncated) 2151 * W 3 Send Window Scale index 2152 * M 3 MSS index 2153 * S 1 SACK permitted 2154 * P 1 Odd/even secret 2155 */ 2156 2157 /* 2158 * Distribution and probability of certain MSS values. Those in between are 2159 * rounded down to the next lower one. 2160 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2161 * .2% .3% 5% 7% 7% 20% 15% 45% 2162 */ 2163 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2164 2165 /* 2166 * Distribution and probability of certain WSCALE values. We have to map the 2167 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2168 * bits based on prevalence of certain values. Where we don't have an exact 2169 * match for are rounded down to the next lower one letting us under-estimate 2170 * the true available window. At the moment this would happen only for the 2171 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2172 * and window size). The absence of the WSCALE option (no scaling in either 2173 * direction) is encoded with index zero. 2174 * [WSCALE values histograms, Allman, 2012] 2175 * X 10 10 35 5 6 14 10% by host 2176 * X 11 4 5 5 18 49 3% by connections 2177 */ 2178 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2179 2180 /* 2181 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2182 * and good cryptographic properties. 2183 */ 2184 static uint32_t 2185 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2186 uint8_t *secbits, uintptr_t secmod) 2187 { 2188 SIPHASH_CTX ctx; 2189 uint32_t siphash[2]; 2190 2191 SipHash24_Init(&ctx); 2192 SipHash_SetKey(&ctx, secbits); 2193 switch (inc->inc_flags & INC_ISIPV6) { 2194 #ifdef INET 2195 case 0: 2196 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2197 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2198 break; 2199 #endif 2200 #ifdef INET6 2201 case INC_ISIPV6: 2202 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2203 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2204 break; 2205 #endif 2206 } 2207 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2208 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2209 SipHash_Update(&ctx, &irs, sizeof(irs)); 2210 SipHash_Update(&ctx, &flags, sizeof(flags)); 2211 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2212 SipHash_Final((u_int8_t *)&siphash, &ctx); 2213 2214 return (siphash[0] ^ siphash[1]); 2215 } 2216 2217 static tcp_seq 2218 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2219 { 2220 u_int i, secbit, wscale; 2221 uint32_t iss, hash; 2222 uint8_t *secbits; 2223 union syncookie cookie; 2224 2225 cookie.cookie = 0; 2226 2227 /* Map our computed MSS into the 3-bit index. */ 2228 for (i = nitems(tcp_sc_msstab) - 1; 2229 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2230 i--) 2231 ; 2232 cookie.flags.mss_idx = i; 2233 2234 /* 2235 * Map the send window scale into the 3-bit index but only if 2236 * the wscale option was received. 2237 */ 2238 if (sc->sc_flags & SCF_WINSCALE) { 2239 wscale = sc->sc_requested_s_scale; 2240 for (i = nitems(tcp_sc_wstab) - 1; 2241 tcp_sc_wstab[i] > wscale && i > 0; 2242 i--) 2243 ; 2244 cookie.flags.wscale_idx = i; 2245 } 2246 2247 /* Can we do SACK? */ 2248 if (sc->sc_flags & SCF_SACK) 2249 cookie.flags.sack_ok = 1; 2250 2251 /* Which of the two secrets to use. */ 2252 secbit = V_tcp_syncache.secret.oddeven & 0x1; 2253 cookie.flags.odd_even = secbit; 2254 2255 secbits = V_tcp_syncache.secret.key[secbit]; 2256 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2257 (uintptr_t)sch); 2258 2259 /* 2260 * Put the flags into the hash and XOR them to get better ISS number 2261 * variance. This doesn't enhance the cryptographic strength and is 2262 * done to prevent the 8 cookie bits from showing up directly on the 2263 * wire. 2264 */ 2265 iss = hash & ~0xff; 2266 iss |= cookie.cookie ^ (hash >> 24); 2267 2268 TCPSTAT_INC(tcps_sc_sendcookie); 2269 return (iss); 2270 } 2271 2272 static struct syncache * 2273 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2274 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2275 struct socket *lso, uint16_t port) 2276 { 2277 uint32_t hash; 2278 uint8_t *secbits; 2279 tcp_seq ack, seq; 2280 int wnd, wscale = 0; 2281 union syncookie cookie; 2282 2283 /* 2284 * Pull information out of SYN-ACK/ACK and revert sequence number 2285 * advances. 2286 */ 2287 ack = th->th_ack - 1; 2288 seq = th->th_seq - 1; 2289 2290 /* 2291 * Unpack the flags containing enough information to restore the 2292 * connection. 2293 */ 2294 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2295 2296 /* Which of the two secrets to use. */ 2297 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; 2298 2299 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2300 2301 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2302 if ((ack & ~0xff) != (hash & ~0xff)) 2303 return (NULL); 2304 2305 /* Fill in the syncache values. */ 2306 sc->sc_flags = 0; 2307 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2308 sc->sc_ipopts = NULL; 2309 2310 sc->sc_irs = seq; 2311 sc->sc_iss = ack; 2312 2313 switch (inc->inc_flags & INC_ISIPV6) { 2314 #ifdef INET 2315 case 0: 2316 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2317 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2318 break; 2319 #endif 2320 #ifdef INET6 2321 case INC_ISIPV6: 2322 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2323 sc->sc_flowlabel = 2324 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2325 break; 2326 #endif 2327 } 2328 2329 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2330 2331 /* We can simply recompute receive window scale we sent earlier. */ 2332 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2333 wscale++; 2334 2335 /* Only use wscale if it was enabled in the orignal SYN. */ 2336 if (cookie.flags.wscale_idx > 0) { 2337 sc->sc_requested_r_scale = wscale; 2338 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2339 sc->sc_flags |= SCF_WINSCALE; 2340 } 2341 2342 wnd = lso->sol_sbrcv_hiwat; 2343 wnd = imax(wnd, 0); 2344 wnd = imin(wnd, TCP_MAXWIN); 2345 sc->sc_wnd = wnd; 2346 2347 if (cookie.flags.sack_ok) 2348 sc->sc_flags |= SCF_SACK; 2349 2350 if (to->to_flags & TOF_TS) { 2351 sc->sc_flags |= SCF_TIMESTAMP; 2352 sc->sc_tsreflect = to->to_tsval; 2353 sc->sc_tsoff = tcp_new_ts_offset(inc); 2354 } 2355 2356 if (to->to_flags & TOF_SIGNATURE) 2357 sc->sc_flags |= SCF_SIGNATURE; 2358 2359 sc->sc_rxmits = 0; 2360 2361 sc->sc_port = port; 2362 2363 TCPSTAT_INC(tcps_sc_recvcookie); 2364 return (sc); 2365 } 2366 2367 #ifdef INVARIANTS 2368 static int 2369 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2370 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2371 struct socket *lso, uint16_t port) 2372 { 2373 struct syncache scs, *scx; 2374 char *s; 2375 2376 bzero(&scs, sizeof(scs)); 2377 scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port); 2378 2379 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2380 return (0); 2381 2382 if (scx != NULL) { 2383 if (sc->sc_peer_mss != scx->sc_peer_mss) 2384 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2385 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2386 2387 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2388 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2389 s, __func__, sc->sc_requested_r_scale, 2390 scx->sc_requested_r_scale); 2391 2392 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2393 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2394 s, __func__, sc->sc_requested_s_scale, 2395 scx->sc_requested_s_scale); 2396 2397 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2398 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2399 } 2400 2401 if (s != NULL) 2402 free(s, M_TCPLOG); 2403 return (0); 2404 } 2405 #endif /* INVARIANTS */ 2406 2407 static void 2408 syncookie_reseed(void *arg) 2409 { 2410 struct tcp_syncache *sc = arg; 2411 uint8_t *secbits; 2412 int secbit; 2413 2414 /* 2415 * Reseeding the secret doesn't have to be protected by a lock. 2416 * It only must be ensured that the new random values are visible 2417 * to all CPUs in a SMP environment. The atomic with release 2418 * semantics ensures that. 2419 */ 2420 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2421 secbits = sc->secret.key[secbit]; 2422 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2423 atomic_add_rel_int(&sc->secret.oddeven, 1); 2424 2425 /* Reschedule ourself. */ 2426 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2427 } 2428 2429 /* 2430 * We have overflowed a bucket. Let's pause dealing with the syncache. 2431 * This function will increment the bucketoverflow statistics appropriately 2432 * (once per pause when pausing is enabled; otherwise, once per overflow). 2433 */ 2434 static void 2435 syncache_pause(struct in_conninfo *inc) 2436 { 2437 time_t delta; 2438 const char *s; 2439 2440 /* XXX: 2441 * 2. Add sysctl read here so we don't get the benefit of this 2442 * change without the new sysctl. 2443 */ 2444 2445 /* 2446 * Try an unlocked read. If we already know that another thread 2447 * has activated the feature, there is no need to proceed. 2448 */ 2449 if (V_tcp_syncache.paused) 2450 return; 2451 2452 /* Are cookied enabled? If not, we can't pause. */ 2453 if (!V_tcp_syncookies) { 2454 TCPSTAT_INC(tcps_sc_bucketoverflow); 2455 return; 2456 } 2457 2458 /* 2459 * We may be the first thread to find an overflow. Get the lock 2460 * and evaluate if we need to take action. 2461 */ 2462 mtx_lock(&V_tcp_syncache.pause_mtx); 2463 if (V_tcp_syncache.paused) { 2464 mtx_unlock(&V_tcp_syncache.pause_mtx); 2465 return; 2466 } 2467 2468 /* Activate protection. */ 2469 V_tcp_syncache.paused = true; 2470 TCPSTAT_INC(tcps_sc_bucketoverflow); 2471 2472 /* 2473 * Determine the last backoff time. If we are seeing a re-newed 2474 * attack within that same time after last reactivating the syncache, 2475 * consider it an extension of the same attack. 2476 */ 2477 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; 2478 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { 2479 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { 2480 delta <<= 1; 2481 V_tcp_syncache.pause_backoff++; 2482 } 2483 } else { 2484 delta = TCP_SYNCACHE_PAUSE_TIME; 2485 V_tcp_syncache.pause_backoff = 0; 2486 } 2487 2488 /* Log a warning, including IP addresses, if able. */ 2489 if (inc != NULL) 2490 s = tcp_log_addrs(inc, NULL, NULL, NULL); 2491 else 2492 s = (const char *)NULL; 2493 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " 2494 "the next %lld seconds%s%s%s\n", (long long)delta, 2495 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", 2496 (s != NULL) ? ")" : ""); 2497 free(__DECONST(void *, s), M_TCPLOG); 2498 2499 /* Use the calculated delta to set a new pause time. */ 2500 V_tcp_syncache.pause_until = time_uptime + delta; 2501 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, 2502 &V_tcp_syncache); 2503 mtx_unlock(&V_tcp_syncache.pause_mtx); 2504 } 2505 2506 /* Evaluate whether we need to unpause. */ 2507 static void 2508 syncache_unpause(void *arg) 2509 { 2510 struct tcp_syncache *sc; 2511 time_t delta; 2512 2513 sc = arg; 2514 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); 2515 callout_deactivate(&sc->pause_co); 2516 2517 /* 2518 * Check to make sure we are not running early. If the pause 2519 * time has expired, then deactivate the protection. 2520 */ 2521 if ((delta = sc->pause_until - time_uptime) > 0) 2522 callout_schedule(&sc->pause_co, delta * hz); 2523 else 2524 sc->paused = false; 2525 } 2526 2527 /* 2528 * Exports the syncache entries to userland so that netstat can display 2529 * them alongside the other sockets. This function is intended to be 2530 * called only from tcp_pcblist. 2531 * 2532 * Due to concurrency on an active system, the number of pcbs exported 2533 * may have no relation to max_pcbs. max_pcbs merely indicates the 2534 * amount of space the caller allocated for this function to use. 2535 */ 2536 int 2537 syncache_pcblist(struct sysctl_req *req) 2538 { 2539 struct xtcpcb xt; 2540 struct syncache *sc; 2541 struct syncache_head *sch; 2542 int error, i; 2543 2544 bzero(&xt, sizeof(xt)); 2545 xt.xt_len = sizeof(xt); 2546 xt.t_state = TCPS_SYN_RECEIVED; 2547 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2548 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2549 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2550 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2551 2552 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 2553 sch = &V_tcp_syncache.hashbase[i]; 2554 SCH_LOCK(sch); 2555 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2556 if (sc->sc_cred != NULL && 2557 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2558 continue; 2559 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2560 xt.xt_inp.inp_vflag = INP_IPV6; 2561 else 2562 xt.xt_inp.inp_vflag = INP_IPV4; 2563 xt.xt_encaps_port = sc->sc_port; 2564 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2565 sizeof (struct in_conninfo)); 2566 error = SYSCTL_OUT(req, &xt, sizeof xt); 2567 if (error) { 2568 SCH_UNLOCK(sch); 2569 return (0); 2570 } 2571 } 2572 SCH_UNLOCK(sch); 2573 } 2574 2575 return (0); 2576 } 2577