1/*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2001 McAfee, Inc. 5 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG 6 * All rights reserved. 7 * 8 * This software was developed for the FreeBSD Project by Jonathan Lemon 9 * and McAfee Research, the Security Research Division of McAfee, Inc. under 10 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 11 * DARPA CHATS research program. [2001 McAfee, Inc.] 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35#include <sys/cdefs.h> 36__FBSDID("$FreeBSD$"); 37 38#include "opt_inet.h" 39#include "opt_inet6.h" 40#include "opt_ipsec.h" 41#include "opt_pcbgroup.h" 42 43#include <sys/param.h> 44#include <sys/systm.h> 45#include <sys/hash.h> 46#include <sys/refcount.h> 47#include <sys/kernel.h> 48#include <sys/sysctl.h> 49#include <sys/limits.h> 50#include <sys/lock.h> 51#include <sys/mutex.h> 52#include <sys/malloc.h> 53#include <sys/mbuf.h> 54#include <sys/proc.h> /* for proc0 declaration */ 55#include <sys/random.h> 56#include <sys/socket.h> 57#include <sys/socketvar.h> 58#include <sys/syslog.h> 59#include <sys/ucred.h> 60 61#include <sys/md5.h> 62#include <crypto/siphash/siphash.h> 63 64#include <vm/uma.h> 65 66#include <net/if.h> 67#include <net/if_var.h> 68#include <net/route.h> 69#include <net/vnet.h> 70 71#include <netinet/in.h> 72#include <netinet/in_kdtrace.h> 73#include <netinet/in_systm.h> 74#include <netinet/ip.h> 75#include <netinet/in_var.h> 76#include <netinet/in_pcb.h> 77#include <netinet/ip_var.h> 78#include <netinet/ip_options.h> 79#ifdef INET6 80#include <netinet/ip6.h> 81#include <netinet/icmp6.h> 82#include <netinet6/nd6.h> 83#include <netinet6/ip6_var.h> 84#include <netinet6/in6_pcb.h> 85#endif 86#include <netinet/tcp.h> 87#include <netinet/tcp_fastopen.h> 88#include <netinet/tcp_fsm.h> 89#include <netinet/tcp_seq.h> 90#include <netinet/tcp_timer.h> 91#include <netinet/tcp_var.h> 92#include <netinet/tcp_syncache.h> 93#ifdef INET6 94#include <netinet6/tcp6_var.h> 95#endif 96#ifdef TCP_OFFLOAD 97#include <netinet/toecore.h> 98#endif 99#include <netinet/udp.h> 100#include <netinet/udp_var.h> 101 102#include <netipsec/ipsec_support.h> 103 104#include <machine/in_cksum.h> 105 106#include <security/mac/mac_framework.h> 107 108VNET_DEFINE_STATIC(int, tcp_syncookies) = 1; 109#define V_tcp_syncookies VNET(tcp_syncookies) 110SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 111 &VNET_NAME(tcp_syncookies), 0, 112 "Use TCP SYN cookies if the syncache overflows"); 113 114VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0; 115#define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 116SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 117 &VNET_NAME(tcp_syncookiesonly), 0, 118 "Use only TCP SYN cookies"); 119 120VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1; 121#define V_functions_inherit_listen_socket_stack \ 122 VNET(functions_inherit_listen_socket_stack) 123SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack, 124 CTLFLAG_VNET | CTLFLAG_RW, 125 &VNET_NAME(functions_inherit_listen_socket_stack), 0, 126 "Inherit listen socket's stack"); 127 128#ifdef TCP_OFFLOAD 129#define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 130#endif 131 132static void syncache_drop(struct syncache *, struct syncache_head *); 133static void syncache_free(struct syncache *); 134static void syncache_insert(struct syncache *, struct syncache_head *); 135static int syncache_respond(struct syncache *, const struct mbuf *, int); 136static struct socket *syncache_socket(struct syncache *, struct socket *, 137 struct mbuf *m); 138static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 139 int docallout); 140static void syncache_timer(void *); 141 142static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 143 uint8_t *, uintptr_t); 144static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 145static struct syncache 146 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 147 struct syncache *, struct tcphdr *, struct tcpopt *, 148 struct socket *, uint16_t); 149static void syncache_pause(struct in_conninfo *); 150static void syncache_unpause(void *); 151static void syncookie_reseed(void *); 152#ifdef INVARIANTS 153static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 154 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 155 struct socket *lso, uint16_t port); 156#endif 157 158/* 159 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 160 * 3 retransmits corresponds to a timeout with default values of 161 * tcp_rexmit_initial * ( 1 + 162 * tcp_backoff[1] + 163 * tcp_backoff[2] + 164 * tcp_backoff[3]) + 3 * tcp_rexmit_slop, 165 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms, 166 * the odds are that the user has given up attempting to connect by then. 167 */ 168#define SYNCACHE_MAXREXMTS 3 169 170/* Arbitrary values */ 171#define TCP_SYNCACHE_HASHSIZE 512 172#define TCP_SYNCACHE_BUCKETLIMIT 30 173 174VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 175#define V_tcp_syncache VNET(tcp_syncache) 176 177static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, 178 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 179 "TCP SYN cache"); 180 181SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 182 &VNET_NAME(tcp_syncache.bucket_limit), 0, 183 "Per-bucket hash limit for syncache"); 184 185SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 186 &VNET_NAME(tcp_syncache.cache_limit), 0, 187 "Overall entry limit for syncache"); 188 189SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 190 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 191 192SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 193 &VNET_NAME(tcp_syncache.hashsize), 0, 194 "Size of TCP syncache hashtable"); 195 196static int 197sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 198{ 199 int error; 200 u_int new; 201 202 new = V_tcp_syncache.rexmt_limit; 203 error = sysctl_handle_int(oidp, &new, 0, req); 204 if ((error == 0) && (req->newptr != NULL)) { 205 if (new > TCP_MAXRXTSHIFT) 206 error = EINVAL; 207 else 208 V_tcp_syncache.rexmt_limit = new; 209 } 210 return (error); 211} 212 213SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 214 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 215 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 216 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI", 217 "Limit on SYN/ACK retransmissions"); 218 219VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 220SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 221 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 222 "Send reset on socket allocation failure"); 223 224static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 225 226#define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 227#define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 228#define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 229 230/* 231 * Requires the syncache entry to be already removed from the bucket list. 232 */ 233static void 234syncache_free(struct syncache *sc) 235{ 236 237 if (sc->sc_ipopts) 238 (void) m_free(sc->sc_ipopts); 239 if (sc->sc_cred) 240 crfree(sc->sc_cred); 241#ifdef MAC 242 mac_syncache_destroy(&sc->sc_label); 243#endif 244 245 uma_zfree(V_tcp_syncache.zone, sc); 246} 247 248void 249syncache_init(void) 250{ 251 int i; 252 253 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 254 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 255 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 256 V_tcp_syncache.hash_secret = arc4random(); 257 258 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 259 &V_tcp_syncache.hashsize); 260 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 261 &V_tcp_syncache.bucket_limit); 262 if (!powerof2(V_tcp_syncache.hashsize) || 263 V_tcp_syncache.hashsize == 0) { 264 printf("WARNING: syncache hash size is not a power of 2.\n"); 265 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 266 } 267 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 268 269 /* Set limits. */ 270 V_tcp_syncache.cache_limit = 271 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 272 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 273 &V_tcp_syncache.cache_limit); 274 275 /* Allocate the hash table. */ 276 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 277 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 278 279#ifdef VIMAGE 280 V_tcp_syncache.vnet = curvnet; 281#endif 282 283 /* Initialize the hash buckets. */ 284 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 285 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 286 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 287 NULL, MTX_DEF); 288 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 289 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 290 V_tcp_syncache.hashbase[i].sch_length = 0; 291 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 292 V_tcp_syncache.hashbase[i].sch_last_overflow = 293 -(SYNCOOKIE_LIFETIME + 1); 294 } 295 296 /* Create the syncache entry zone. */ 297 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 298 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 299 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 300 V_tcp_syncache.cache_limit); 301 302 /* Start the SYN cookie reseeder callout. */ 303 callout_init(&V_tcp_syncache.secret.reseed, 1); 304 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 305 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 306 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 307 syncookie_reseed, &V_tcp_syncache); 308 309 /* Initialize the pause machinery. */ 310 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF); 311 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx, 312 0); 313 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME; 314 V_tcp_syncache.pause_backoff = 0; 315 V_tcp_syncache.paused = false; 316} 317 318#ifdef VIMAGE 319void 320syncache_destroy(void) 321{ 322 struct syncache_head *sch; 323 struct syncache *sc, *nsc; 324 int i; 325 326 /* 327 * Stop the re-seed timer before freeing resources. No need to 328 * possibly schedule it another time. 329 */ 330 callout_drain(&V_tcp_syncache.secret.reseed); 331 332 /* Stop the SYN cache pause callout. */ 333 mtx_lock(&V_tcp_syncache.pause_mtx); 334 if (callout_stop(&V_tcp_syncache.pause_co) == 0) { 335 mtx_unlock(&V_tcp_syncache.pause_mtx); 336 callout_drain(&V_tcp_syncache.pause_co); 337 } else 338 mtx_unlock(&V_tcp_syncache.pause_mtx); 339 340 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 341 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 342 sch = &V_tcp_syncache.hashbase[i]; 343 callout_drain(&sch->sch_timer); 344 345 SCH_LOCK(sch); 346 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 347 syncache_drop(sc, sch); 348 SCH_UNLOCK(sch); 349 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 350 ("%s: sch->sch_bucket not empty", __func__)); 351 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 352 __func__, sch->sch_length)); 353 mtx_destroy(&sch->sch_mtx); 354 } 355 356 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 357 ("%s: cache_count not 0", __func__)); 358 359 /* Free the allocated global resources. */ 360 uma_zdestroy(V_tcp_syncache.zone); 361 free(V_tcp_syncache.hashbase, M_SYNCACHE); 362 mtx_destroy(&V_tcp_syncache.pause_mtx); 363} 364#endif 365 366/* 367 * Inserts a syncache entry into the specified bucket row. 368 * Locks and unlocks the syncache_head autonomously. 369 */ 370static void 371syncache_insert(struct syncache *sc, struct syncache_head *sch) 372{ 373 struct syncache *sc2; 374 375 SCH_LOCK(sch); 376 377 /* 378 * Make sure that we don't overflow the per-bucket limit. 379 * If the bucket is full, toss the oldest element. 380 */ 381 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 382 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 383 ("sch->sch_length incorrect")); 384 syncache_pause(&sc->sc_inc); 385 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 386 sch->sch_last_overflow = time_uptime; 387 syncache_drop(sc2, sch); 388 } 389 390 /* Put it into the bucket. */ 391 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 392 sch->sch_length++; 393 394#ifdef TCP_OFFLOAD 395 if (ADDED_BY_TOE(sc)) { 396 struct toedev *tod = sc->sc_tod; 397 398 tod->tod_syncache_added(tod, sc->sc_todctx); 399 } 400#endif 401 402 /* Reinitialize the bucket row's timer. */ 403 if (sch->sch_length == 1) 404 sch->sch_nextc = ticks + INT_MAX; 405 syncache_timeout(sc, sch, 1); 406 407 SCH_UNLOCK(sch); 408 409 TCPSTATES_INC(TCPS_SYN_RECEIVED); 410 TCPSTAT_INC(tcps_sc_added); 411} 412 413/* 414 * Remove and free entry from syncache bucket row. 415 * Expects locked syncache head. 416 */ 417static void 418syncache_drop(struct syncache *sc, struct syncache_head *sch) 419{ 420 421 SCH_LOCK_ASSERT(sch); 422 423 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 424 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 425 sch->sch_length--; 426 427#ifdef TCP_OFFLOAD 428 if (ADDED_BY_TOE(sc)) { 429 struct toedev *tod = sc->sc_tod; 430 431 tod->tod_syncache_removed(tod, sc->sc_todctx); 432 } 433#endif 434 435 syncache_free(sc); 436} 437 438/* 439 * Engage/reengage time on bucket row. 440 */ 441static void 442syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 443{ 444 int rexmt; 445 446 if (sc->sc_rxmits == 0) 447 rexmt = tcp_rexmit_initial; 448 else 449 TCPT_RANGESET(rexmt, 450 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], 451 tcp_rexmit_min, TCPTV_REXMTMAX); 452 sc->sc_rxttime = ticks + rexmt; 453 sc->sc_rxmits++; 454 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 455 sch->sch_nextc = sc->sc_rxttime; 456 if (docallout) 457 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 458 syncache_timer, (void *)sch); 459 } 460} 461 462/* 463 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 464 * If we have retransmitted an entry the maximum number of times, expire it. 465 * One separate timer for each bucket row. 466 */ 467static void 468syncache_timer(void *xsch) 469{ 470 struct syncache_head *sch = (struct syncache_head *)xsch; 471 struct syncache *sc, *nsc; 472 struct epoch_tracker et; 473 int tick = ticks; 474 char *s; 475 bool paused; 476 477 CURVNET_SET(sch->sch_sc->vnet); 478 479 /* NB: syncache_head has already been locked by the callout. */ 480 SCH_LOCK_ASSERT(sch); 481 482 /* 483 * In the following cycle we may remove some entries and/or 484 * advance some timeouts, so re-initialize the bucket timer. 485 */ 486 sch->sch_nextc = tick + INT_MAX; 487 488 /* 489 * If we have paused processing, unconditionally remove 490 * all syncache entries. 491 */ 492 mtx_lock(&V_tcp_syncache.pause_mtx); 493 paused = V_tcp_syncache.paused; 494 mtx_unlock(&V_tcp_syncache.pause_mtx); 495 496 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 497 if (paused) { 498 syncache_drop(sc, sch); 499 continue; 500 } 501 /* 502 * We do not check if the listen socket still exists 503 * and accept the case where the listen socket may be 504 * gone by the time we resend the SYN/ACK. We do 505 * not expect this to happens often. If it does, 506 * then the RST will be sent by the time the remote 507 * host does the SYN/ACK->ACK. 508 */ 509 if (TSTMP_GT(sc->sc_rxttime, tick)) { 510 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 511 sch->sch_nextc = sc->sc_rxttime; 512 continue; 513 } 514 if (sc->sc_rxmits > V_tcp_ecn_maxretries) { 515 sc->sc_flags &= ~SCF_ECN; 516 } 517 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 518 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 519 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 520 "giving up and removing syncache entry\n", 521 s, __func__); 522 free(s, M_TCPLOG); 523 } 524 syncache_drop(sc, sch); 525 TCPSTAT_INC(tcps_sc_stale); 526 continue; 527 } 528 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 529 log(LOG_DEBUG, "%s; %s: Response timeout, " 530 "retransmitting (%u) SYN|ACK\n", 531 s, __func__, sc->sc_rxmits); 532 free(s, M_TCPLOG); 533 } 534 535 NET_EPOCH_ENTER(et); 536 syncache_respond(sc, NULL, TH_SYN|TH_ACK); 537 NET_EPOCH_EXIT(et); 538 TCPSTAT_INC(tcps_sc_retransmitted); 539 syncache_timeout(sc, sch, 0); 540 } 541 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 542 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 543 syncache_timer, (void *)(sch)); 544 CURVNET_RESTORE(); 545} 546 547/* 548 * Returns true if the system is only using cookies at the moment. 549 * This could be due to a sysadmin decision to only use cookies, or it 550 * could be due to the system detecting an attack. 551 */ 552static inline bool 553syncache_cookiesonly(void) 554{ 555 556 return (V_tcp_syncookies && (V_tcp_syncache.paused || 557 V_tcp_syncookiesonly)); 558} 559 560/* 561 * Find the hash bucket for the given connection. 562 */ 563static struct syncache_head * 564syncache_hashbucket(struct in_conninfo *inc) 565{ 566 uint32_t hash; 567 568 /* 569 * The hash is built on foreign port + local port + foreign address. 570 * We rely on the fact that struct in_conninfo starts with 16 bits 571 * of foreign port, then 16 bits of local port then followed by 128 572 * bits of foreign address. In case of IPv4 address, the first 3 573 * 32-bit words of the address always are zeroes. 574 */ 575 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 576 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 577 578 return (&V_tcp_syncache.hashbase[hash]); 579} 580 581/* 582 * Find an entry in the syncache. 583 * Returns always with locked syncache_head plus a matching entry or NULL. 584 */ 585static struct syncache * 586syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 587{ 588 struct syncache *sc; 589 struct syncache_head *sch; 590 591 *schp = sch = syncache_hashbucket(inc); 592 SCH_LOCK(sch); 593 594 /* Circle through bucket row to find matching entry. */ 595 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 596 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 597 sizeof(struct in_endpoints)) == 0) 598 break; 599 600 return (sc); /* Always returns with locked sch. */ 601} 602 603/* 604 * This function is called when we get a RST for a 605 * non-existent connection, so that we can see if the 606 * connection is in the syn cache. If it is, zap it. 607 * If required send a challenge ACK. 608 */ 609void 610syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m, 611 uint16_t port) 612{ 613 struct syncache *sc; 614 struct syncache_head *sch; 615 char *s = NULL; 616 617 if (syncache_cookiesonly()) 618 return; 619 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 620 SCH_LOCK_ASSERT(sch); 621 622 /* 623 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 624 * See RFC 793 page 65, section SEGMENT ARRIVES. 625 */ 626 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 627 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 628 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 629 "FIN flag set, segment ignored\n", s, __func__); 630 TCPSTAT_INC(tcps_badrst); 631 goto done; 632 } 633 634 /* 635 * No corresponding connection was found in syncache. 636 * If syncookies are enabled and possibly exclusively 637 * used, or we are under memory pressure, a valid RST 638 * may not find a syncache entry. In that case we're 639 * done and no SYN|ACK retransmissions will happen. 640 * Otherwise the RST was misdirected or spoofed. 641 */ 642 if (sc == NULL) { 643 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 644 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 645 "syncache entry (possibly syncookie only), " 646 "segment ignored\n", s, __func__); 647 TCPSTAT_INC(tcps_badrst); 648 goto done; 649 } 650 651 /* The remote UDP encaps port does not match. */ 652 if (sc->sc_port != port) { 653 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 654 log(LOG_DEBUG, "%s; %s: Spurious RST with matching " 655 "syncache entry but non-matching UDP encaps port, " 656 "segment ignored\n", s, __func__); 657 TCPSTAT_INC(tcps_badrst); 658 goto done; 659 } 660 661 /* 662 * If the RST bit is set, check the sequence number to see 663 * if this is a valid reset segment. 664 * 665 * RFC 793 page 37: 666 * In all states except SYN-SENT, all reset (RST) segments 667 * are validated by checking their SEQ-fields. A reset is 668 * valid if its sequence number is in the window. 669 * 670 * RFC 793 page 69: 671 * There are four cases for the acceptability test for an incoming 672 * segment: 673 * 674 * Segment Receive Test 675 * Length Window 676 * ------- ------- ------------------------------------------- 677 * 0 0 SEG.SEQ = RCV.NXT 678 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 679 * >0 0 not acceptable 680 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 681 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 682 * 683 * Note that when receiving a SYN segment in the LISTEN state, 684 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 685 * described in RFC 793, page 66. 686 */ 687 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 688 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 689 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 690 if (V_tcp_insecure_rst || 691 th->th_seq == sc->sc_irs + 1) { 692 syncache_drop(sc, sch); 693 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 694 log(LOG_DEBUG, 695 "%s; %s: Our SYN|ACK was rejected, " 696 "connection attempt aborted by remote " 697 "endpoint\n", 698 s, __func__); 699 TCPSTAT_INC(tcps_sc_reset); 700 } else { 701 TCPSTAT_INC(tcps_badrst); 702 /* Send challenge ACK. */ 703 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 704 log(LOG_DEBUG, "%s; %s: RST with invalid " 705 " SEQ %u != NXT %u (+WND %u), " 706 "sending challenge ACK\n", 707 s, __func__, 708 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 709 syncache_respond(sc, m, TH_ACK); 710 } 711 } else { 712 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 713 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 714 "NXT %u (+WND %u), segment ignored\n", 715 s, __func__, 716 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 717 TCPSTAT_INC(tcps_badrst); 718 } 719 720done: 721 if (s != NULL) 722 free(s, M_TCPLOG); 723 SCH_UNLOCK(sch); 724} 725 726void 727syncache_badack(struct in_conninfo *inc, uint16_t port) 728{ 729 struct syncache *sc; 730 struct syncache_head *sch; 731 732 if (syncache_cookiesonly()) 733 return; 734 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 735 SCH_LOCK_ASSERT(sch); 736 if ((sc != NULL) && (sc->sc_port == port)) { 737 syncache_drop(sc, sch); 738 TCPSTAT_INC(tcps_sc_badack); 739 } 740 SCH_UNLOCK(sch); 741} 742 743void 744syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port) 745{ 746 struct syncache *sc; 747 struct syncache_head *sch; 748 749 if (syncache_cookiesonly()) 750 return; 751 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 752 SCH_LOCK_ASSERT(sch); 753 if (sc == NULL) 754 goto done; 755 756 /* If the port != sc_port, then it's a bogus ICMP msg */ 757 if (port != sc->sc_port) 758 goto done; 759 760 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 761 if (ntohl(th_seq) != sc->sc_iss) 762 goto done; 763 764 /* 765 * If we've rertransmitted 3 times and this is our second error, 766 * we remove the entry. Otherwise, we allow it to continue on. 767 * This prevents us from incorrectly nuking an entry during a 768 * spurious network outage. 769 * 770 * See tcp_notify(). 771 */ 772 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 773 sc->sc_flags |= SCF_UNREACH; 774 goto done; 775 } 776 syncache_drop(sc, sch); 777 TCPSTAT_INC(tcps_sc_unreach); 778done: 779 SCH_UNLOCK(sch); 780} 781 782/* 783 * Build a new TCP socket structure from a syncache entry. 784 * 785 * On success return the newly created socket with its underlying inp locked. 786 */ 787static struct socket * 788syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 789{ 790 struct tcp_function_block *blk; 791 struct inpcb *inp = NULL; 792 struct socket *so; 793 struct tcpcb *tp; 794 int error; 795 char *s; 796 797 NET_EPOCH_ASSERT(); 798 799 /* 800 * Ok, create the full blown connection, and set things up 801 * as they would have been set up if we had created the 802 * connection when the SYN arrived. If we can't create 803 * the connection, abort it. 804 */ 805 so = sonewconn(lso, 0); 806 if (so == NULL) { 807 /* 808 * Drop the connection; we will either send a RST or 809 * have the peer retransmit its SYN again after its 810 * RTO and try again. 811 */ 812 TCPSTAT_INC(tcps_listendrop); 813 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 814 log(LOG_DEBUG, "%s; %s: Socket create failed " 815 "due to limits or memory shortage\n", 816 s, __func__); 817 free(s, M_TCPLOG); 818 } 819 goto abort2; 820 } 821#ifdef MAC 822 mac_socketpeer_set_from_mbuf(m, so); 823#endif 824 825 inp = sotoinpcb(so); 826 inp->inp_inc.inc_fibnum = so->so_fibnum; 827 INP_WLOCK(inp); 828 /* 829 * Exclusive pcbinfo lock is not required in syncache socket case even 830 * if two inpcb locks can be acquired simultaneously: 831 * - the inpcb in LISTEN state, 832 * - the newly created inp. 833 * 834 * In this case, an inp cannot be at same time in LISTEN state and 835 * just created by an accept() call. 836 */ 837 INP_HASH_WLOCK(&V_tcbinfo); 838 839 /* Insert new socket into PCB hash list. */ 840 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 841#ifdef INET6 842 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 843 inp->inp_vflag &= ~INP_IPV4; 844 inp->inp_vflag |= INP_IPV6; 845 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 846 } else { 847 inp->inp_vflag &= ~INP_IPV6; 848 inp->inp_vflag |= INP_IPV4; 849#endif 850 inp->inp_ip_ttl = sc->sc_ip_ttl; 851 inp->inp_ip_tos = sc->sc_ip_tos; 852 inp->inp_laddr = sc->sc_inc.inc_laddr; 853#ifdef INET6 854 } 855#endif 856 857 /* 858 * If there's an mbuf and it has a flowid, then let's initialise the 859 * inp with that particular flowid. 860 */ 861 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 862 inp->inp_flowid = m->m_pkthdr.flowid; 863 inp->inp_flowtype = M_HASHTYPE_GET(m); 864#ifdef NUMA 865 inp->inp_numa_domain = m->m_pkthdr.numa_domain; 866#endif 867 } 868 869 inp->inp_lport = sc->sc_inc.inc_lport; 870#ifdef INET6 871 if (inp->inp_vflag & INP_IPV6PROTO) { 872 struct inpcb *oinp = sotoinpcb(lso); 873 874 /* 875 * Inherit socket options from the listening socket. 876 * Note that in6p_inputopts are not (and should not be) 877 * copied, since it stores previously received options and is 878 * used to detect if each new option is different than the 879 * previous one and hence should be passed to a user. 880 * If we copied in6p_inputopts, a user would not be able to 881 * receive options just after calling the accept system call. 882 */ 883 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 884 if (oinp->in6p_outputopts) 885 inp->in6p_outputopts = 886 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 887 inp->in6p_hops = oinp->in6p_hops; 888 } 889 890 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 891 struct in6_addr laddr6; 892 struct sockaddr_in6 sin6; 893 894 sin6.sin6_family = AF_INET6; 895 sin6.sin6_len = sizeof(sin6); 896 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 897 sin6.sin6_port = sc->sc_inc.inc_fport; 898 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 899 laddr6 = inp->in6p_laddr; 900 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 901 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 902 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 903 thread0.td_ucred, m, false)) != 0) { 904 inp->in6p_laddr = laddr6; 905 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 906 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 907 "with error %i\n", 908 s, __func__, error); 909 free(s, M_TCPLOG); 910 } 911 INP_HASH_WUNLOCK(&V_tcbinfo); 912 goto abort; 913 } 914 /* Override flowlabel from in6_pcbconnect. */ 915 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 916 inp->inp_flow |= sc->sc_flowlabel; 917 } 918#endif /* INET6 */ 919#if defined(INET) && defined(INET6) 920 else 921#endif 922#ifdef INET 923 { 924 struct in_addr laddr; 925 struct sockaddr_in sin; 926 927 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 928 929 if (inp->inp_options == NULL) { 930 inp->inp_options = sc->sc_ipopts; 931 sc->sc_ipopts = NULL; 932 } 933 934 sin.sin_family = AF_INET; 935 sin.sin_len = sizeof(sin); 936 sin.sin_addr = sc->sc_inc.inc_faddr; 937 sin.sin_port = sc->sc_inc.inc_fport; 938 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 939 laddr = inp->inp_laddr; 940 if (inp->inp_laddr.s_addr == INADDR_ANY) 941 inp->inp_laddr = sc->sc_inc.inc_laddr; 942 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 943 thread0.td_ucred, m, false)) != 0) { 944 inp->inp_laddr = laddr; 945 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 946 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 947 "with error %i\n", 948 s, __func__, error); 949 free(s, M_TCPLOG); 950 } 951 INP_HASH_WUNLOCK(&V_tcbinfo); 952 goto abort; 953 } 954 } 955#endif /* INET */ 956#if defined(IPSEC) || defined(IPSEC_SUPPORT) 957 /* Copy old policy into new socket's. */ 958 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 959 printf("syncache_socket: could not copy policy\n"); 960#endif 961 INP_HASH_WUNLOCK(&V_tcbinfo); 962 tp = intotcpcb(inp); 963 tcp_state_change(tp, TCPS_SYN_RECEIVED); 964 tp->iss = sc->sc_iss; 965 tp->irs = sc->sc_irs; 966 tp->t_port = sc->sc_port; 967 tcp_rcvseqinit(tp); 968 tcp_sendseqinit(tp); 969 blk = sototcpcb(lso)->t_fb; 970 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 971 /* 972 * Our parents t_fb was not the default, 973 * we need to release our ref on tp->t_fb and 974 * pickup one on the new entry. 975 */ 976 struct tcp_function_block *rblk; 977 978 rblk = find_and_ref_tcp_fb(blk); 979 KASSERT(rblk != NULL, 980 ("cannot find blk %p out of syncache?", blk)); 981 if (tp->t_fb->tfb_tcp_fb_fini) 982 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 983 refcount_release(&tp->t_fb->tfb_refcnt); 984 tp->t_fb = rblk; 985 /* 986 * XXXrrs this is quite dangerous, it is possible 987 * for the new function to fail to init. We also 988 * are not asking if the handoff_is_ok though at 989 * the very start thats probalbly ok. 990 */ 991 if (tp->t_fb->tfb_tcp_fb_init) { 992 (*tp->t_fb->tfb_tcp_fb_init)(tp); 993 } 994 } 995 tp->snd_wl1 = sc->sc_irs; 996 tp->snd_max = tp->iss + 1; 997 tp->snd_nxt = tp->iss + 1; 998 tp->rcv_up = sc->sc_irs + 1; 999 tp->rcv_wnd = sc->sc_wnd; 1000 tp->rcv_adv += tp->rcv_wnd; 1001 tp->last_ack_sent = tp->rcv_nxt; 1002 1003 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 1004 if (sc->sc_flags & SCF_NOOPT) 1005 tp->t_flags |= TF_NOOPT; 1006 else { 1007 if (sc->sc_flags & SCF_WINSCALE) { 1008 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 1009 tp->snd_scale = sc->sc_requested_s_scale; 1010 tp->request_r_scale = sc->sc_requested_r_scale; 1011 } 1012 if (sc->sc_flags & SCF_TIMESTAMP) { 1013 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 1014 tp->ts_recent = sc->sc_tsreflect; 1015 tp->ts_recent_age = tcp_ts_getticks(); 1016 tp->ts_offset = sc->sc_tsoff; 1017 } 1018#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1019 if (sc->sc_flags & SCF_SIGNATURE) 1020 tp->t_flags |= TF_SIGNATURE; 1021#endif 1022 if (sc->sc_flags & SCF_SACK) 1023 tp->t_flags |= TF_SACK_PERMIT; 1024 } 1025 1026 if (sc->sc_flags & SCF_ECN) 1027 tp->t_flags2 |= TF2_ECN_PERMIT; 1028 1029 /* 1030 * Set up MSS and get cached values from tcp_hostcache. 1031 * This might overwrite some of the defaults we just set. 1032 */ 1033 tcp_mss(tp, sc->sc_peer_mss); 1034 1035 /* 1036 * If the SYN,ACK was retransmitted, indicate that CWND to be 1037 * limited to one segment in cc_conn_init(). 1038 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 1039 */ 1040 if (sc->sc_rxmits > 1) 1041 tp->snd_cwnd = 1; 1042 1043#ifdef TCP_OFFLOAD 1044 /* 1045 * Allow a TOE driver to install its hooks. Note that we hold the 1046 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 1047 * new connection before the TOE driver has done its thing. 1048 */ 1049 if (ADDED_BY_TOE(sc)) { 1050 struct toedev *tod = sc->sc_tod; 1051 1052 tod->tod_offload_socket(tod, sc->sc_todctx, so); 1053 } 1054#endif 1055 /* 1056 * Copy and activate timers. 1057 */ 1058 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 1059 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 1060 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1061 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1062 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1063 1064 TCPSTAT_INC(tcps_accepts); 1065 return (so); 1066 1067abort: 1068 INP_WUNLOCK(inp); 1069abort2: 1070 if (so != NULL) 1071 soabort(so); 1072 return (NULL); 1073} 1074 1075/* 1076 * This function gets called when we receive an ACK for a 1077 * socket in the LISTEN state. We look up the connection 1078 * in the syncache, and if its there, we pull it out of 1079 * the cache and turn it into a full-blown connection in 1080 * the SYN-RECEIVED state. 1081 * 1082 * On syncache_socket() success the newly created socket 1083 * has its underlying inp locked. 1084 */ 1085int 1086syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1087 struct socket **lsop, struct mbuf *m, uint16_t port) 1088{ 1089 struct syncache *sc; 1090 struct syncache_head *sch; 1091 struct syncache scs; 1092 char *s; 1093 bool locked; 1094 1095 NET_EPOCH_ASSERT(); 1096 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1097 ("%s: can handle only ACK", __func__)); 1098 1099 if (syncache_cookiesonly()) { 1100 sc = NULL; 1101 sch = syncache_hashbucket(inc); 1102 locked = false; 1103 } else { 1104 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1105 locked = true; 1106 SCH_LOCK_ASSERT(sch); 1107 } 1108 1109#ifdef INVARIANTS 1110 /* 1111 * Test code for syncookies comparing the syncache stored 1112 * values with the reconstructed values from the cookie. 1113 */ 1114 if (sc != NULL) 1115 syncookie_cmp(inc, sch, sc, th, to, *lsop, port); 1116#endif 1117 1118 if (sc == NULL) { 1119 /* 1120 * There is no syncache entry, so see if this ACK is 1121 * a returning syncookie. To do this, first: 1122 * A. Check if syncookies are used in case of syncache 1123 * overflows 1124 * B. See if this socket has had a syncache entry dropped in 1125 * the recent past. We don't want to accept a bogus 1126 * syncookie if we've never received a SYN or accept it 1127 * twice. 1128 * C. check that the syncookie is valid. If it is, then 1129 * cobble up a fake syncache entry, and return. 1130 */ 1131 if (locked && !V_tcp_syncookies) { 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 (syncookies disabled)\n", 1136 s, __func__); 1137 goto failed; 1138 } 1139 if (locked && !V_tcp_syncookiesonly && 1140 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1141 SCH_UNLOCK(sch); 1142 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1143 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1144 "segment rejected (no syncache entry)\n", 1145 s, __func__); 1146 goto failed; 1147 } 1148 bzero(&scs, sizeof(scs)); 1149 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port); 1150 if (locked) 1151 SCH_UNLOCK(sch); 1152 if (sc == NULL) { 1153 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1154 log(LOG_DEBUG, "%s; %s: Segment failed " 1155 "SYNCOOKIE authentication, segment rejected " 1156 "(probably spoofed)\n", s, __func__); 1157 goto failed; 1158 } 1159#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1160 /* If received ACK has MD5 signature, check it. */ 1161 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1162 (!TCPMD5_ENABLED() || 1163 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1164 /* Drop the ACK. */ 1165 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1166 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1167 "MD5 signature doesn't match.\n", 1168 s, __func__); 1169 free(s, M_TCPLOG); 1170 } 1171 TCPSTAT_INC(tcps_sig_err_sigopt); 1172 return (-1); /* Do not send RST */ 1173 } 1174#endif /* TCP_SIGNATURE */ 1175 } else { 1176 if (sc->sc_port != port) { 1177 SCH_UNLOCK(sch); 1178 return (0); 1179 } 1180#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1181 /* 1182 * If listening socket requested TCP digests, check that 1183 * received ACK has signature and it is correct. 1184 * If not, drop the ACK and leave sc entry in th cache, 1185 * because SYN was received with correct signature. 1186 */ 1187 if (sc->sc_flags & SCF_SIGNATURE) { 1188 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1189 /* No signature */ 1190 TCPSTAT_INC(tcps_sig_err_nosigopt); 1191 SCH_UNLOCK(sch); 1192 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1193 log(LOG_DEBUG, "%s; %s: Segment " 1194 "rejected, MD5 signature wasn't " 1195 "provided.\n", s, __func__); 1196 free(s, M_TCPLOG); 1197 } 1198 return (-1); /* Do not send RST */ 1199 } 1200 if (!TCPMD5_ENABLED() || 1201 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1202 /* Doesn't match or no SA */ 1203 SCH_UNLOCK(sch); 1204 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1205 log(LOG_DEBUG, "%s; %s: Segment " 1206 "rejected, MD5 signature doesn't " 1207 "match.\n", s, __func__); 1208 free(s, M_TCPLOG); 1209 } 1210 return (-1); /* Do not send RST */ 1211 } 1212 } 1213#endif /* TCP_SIGNATURE */ 1214 1215 /* 1216 * RFC 7323 PAWS: If we have a timestamp on this segment and 1217 * it's less than ts_recent, drop it. 1218 * XXXMT: RFC 7323 also requires to send an ACK. 1219 * In tcp_input.c this is only done for TCP segments 1220 * with user data, so be consistent here and just drop 1221 * the segment. 1222 */ 1223 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1224 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1225 SCH_UNLOCK(sch); 1226 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1227 log(LOG_DEBUG, 1228 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1229 "segment dropped\n", s, __func__, 1230 to->to_tsval, sc->sc_tsreflect); 1231 free(s, M_TCPLOG); 1232 } 1233 return (-1); /* Do not send RST */ 1234 } 1235 1236 /* 1237 * If timestamps were not negotiated during SYN/ACK and a 1238 * segment with a timestamp is received, ignore the 1239 * timestamp and process the packet normally. 1240 * See section 3.2 of RFC 7323. 1241 */ 1242 if (!(sc->sc_flags & SCF_TIMESTAMP) && 1243 (to->to_flags & TOF_TS)) { 1244 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1245 log(LOG_DEBUG, "%s; %s: Timestamp not " 1246 "expected, segment processed normally\n", 1247 s, __func__); 1248 free(s, M_TCPLOG); 1249 s = NULL; 1250 } 1251 } 1252 1253 /* 1254 * If timestamps were negotiated during SYN/ACK and a 1255 * segment without a timestamp is received, silently drop 1256 * the segment, unless the missing timestamps are tolerated. 1257 * See section 3.2 of RFC 7323. 1258 */ 1259 if ((sc->sc_flags & SCF_TIMESTAMP) && 1260 !(to->to_flags & TOF_TS)) { 1261 if (V_tcp_tolerate_missing_ts) { 1262 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1263 log(LOG_DEBUG, 1264 "%s; %s: Timestamp missing, " 1265 "segment processed normally\n", 1266 s, __func__); 1267 free(s, M_TCPLOG); 1268 } 1269 } else { 1270 SCH_UNLOCK(sch); 1271 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1272 log(LOG_DEBUG, 1273 "%s; %s: Timestamp missing, " 1274 "segment silently dropped\n", 1275 s, __func__); 1276 free(s, M_TCPLOG); 1277 } 1278 return (-1); /* Do not send RST */ 1279 } 1280 } 1281 1282 /* 1283 * Pull out the entry to unlock the bucket row. 1284 * 1285 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1286 * tcp_state_change(). The tcpcb is not existent at this 1287 * moment. A new one will be allocated via syncache_socket-> 1288 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1289 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1290 */ 1291 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1292 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1293 sch->sch_length--; 1294#ifdef TCP_OFFLOAD 1295 if (ADDED_BY_TOE(sc)) { 1296 struct toedev *tod = sc->sc_tod; 1297 1298 tod->tod_syncache_removed(tod, sc->sc_todctx); 1299 } 1300#endif 1301 SCH_UNLOCK(sch); 1302 } 1303 1304 /* 1305 * Segment validation: 1306 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1307 */ 1308 if (th->th_ack != sc->sc_iss + 1) { 1309 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1310 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1311 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1312 goto failed; 1313 } 1314 1315 /* 1316 * The SEQ must fall in the window starting at the received 1317 * initial receive sequence number + 1 (the SYN). 1318 */ 1319 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1320 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1321 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1322 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1323 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1324 goto failed; 1325 } 1326 1327 *lsop = syncache_socket(sc, *lsop, m); 1328 1329 if (*lsop == NULL) 1330 TCPSTAT_INC(tcps_sc_aborted); 1331 else 1332 TCPSTAT_INC(tcps_sc_completed); 1333 1334/* how do we find the inp for the new socket? */ 1335 if (sc != &scs) 1336 syncache_free(sc); 1337 return (1); 1338failed: 1339 if (sc != NULL && sc != &scs) 1340 syncache_free(sc); 1341 if (s != NULL) 1342 free(s, M_TCPLOG); 1343 *lsop = NULL; 1344 return (0); 1345} 1346 1347static void 1348syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1349 uint64_t response_cookie) 1350{ 1351 struct inpcb *inp; 1352 struct tcpcb *tp; 1353 unsigned int *pending_counter; 1354 1355 NET_EPOCH_ASSERT(); 1356 1357 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1358 *lsop = syncache_socket(sc, *lsop, m); 1359 if (*lsop == NULL) { 1360 TCPSTAT_INC(tcps_sc_aborted); 1361 atomic_subtract_int(pending_counter, 1); 1362 } else { 1363 soisconnected(*lsop); 1364 inp = sotoinpcb(*lsop); 1365 tp = intotcpcb(inp); 1366 tp->t_flags |= TF_FASTOPEN; 1367 tp->t_tfo_cookie.server = response_cookie; 1368 tp->snd_max = tp->iss; 1369 tp->snd_nxt = tp->iss; 1370 tp->t_tfo_pending = pending_counter; 1371 TCPSTAT_INC(tcps_sc_completed); 1372 } 1373} 1374 1375/* 1376 * Given a LISTEN socket and an inbound SYN request, add 1377 * this to the syn cache, and send back a segment: 1378 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1379 * to the source. 1380 * 1381 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1382 * Doing so would require that we hold onto the data and deliver it 1383 * to the application. However, if we are the target of a SYN-flood 1384 * DoS attack, an attacker could send data which would eventually 1385 * consume all available buffer space if it were ACKed. By not ACKing 1386 * the data, we avoid this DoS scenario. 1387 * 1388 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1389 * cookie is processed and a new socket is created. In this case, any data 1390 * accompanying the SYN will be queued to the socket by tcp_input() and will 1391 * be ACKed either when the application sends response data or the delayed 1392 * ACK timer expires, whichever comes first. 1393 */ 1394int 1395syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1396 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1397 void *todctx, uint8_t iptos, uint16_t port) 1398{ 1399 struct tcpcb *tp; 1400 struct socket *so; 1401 struct syncache *sc = NULL; 1402 struct syncache_head *sch; 1403 struct mbuf *ipopts = NULL; 1404 u_int ltflags; 1405 int win, ip_ttl, ip_tos; 1406 char *s; 1407 int rv = 0; 1408#ifdef INET6 1409 int autoflowlabel = 0; 1410#endif 1411#ifdef MAC 1412 struct label *maclabel; 1413#endif 1414 struct syncache scs; 1415 struct ucred *cred; 1416 uint64_t tfo_response_cookie; 1417 unsigned int *tfo_pending = NULL; 1418 int tfo_cookie_valid = 0; 1419 int tfo_response_cookie_valid = 0; 1420 bool locked; 1421 1422 INP_WLOCK_ASSERT(inp); /* listen socket */ 1423 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1424 ("%s: unexpected tcp flags", __func__)); 1425 1426 /* 1427 * Combine all so/tp operations very early to drop the INP lock as 1428 * soon as possible. 1429 */ 1430 so = *lsop; 1431 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1432 tp = sototcpcb(so); 1433 cred = crhold(so->so_cred); 1434 1435#ifdef INET6 1436 if (inc->inc_flags & INC_ISIPV6) { 1437 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1438 autoflowlabel = 1; 1439 } 1440 ip_ttl = in6_selecthlim(inp, NULL); 1441 if ((inp->in6p_outputopts == NULL) || 1442 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1443 ip_tos = 0; 1444 } else { 1445 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1446 } 1447 } 1448#endif 1449#if defined(INET6) && defined(INET) 1450 else 1451#endif 1452#ifdef INET 1453 { 1454 ip_ttl = inp->inp_ip_ttl; 1455 ip_tos = inp->inp_ip_tos; 1456 } 1457#endif 1458 win = so->sol_sbrcv_hiwat; 1459 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1460 1461 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1462 (tp->t_tfo_pending != NULL) && 1463 (to->to_flags & TOF_FASTOPEN)) { 1464 /* 1465 * Limit the number of pending TFO connections to 1466 * approximately half of the queue limit. This prevents TFO 1467 * SYN floods from starving the service by filling the 1468 * listen queue with bogus TFO connections. 1469 */ 1470 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1471 (so->sol_qlimit / 2)) { 1472 int result; 1473 1474 result = tcp_fastopen_check_cookie(inc, 1475 to->to_tfo_cookie, to->to_tfo_len, 1476 &tfo_response_cookie); 1477 tfo_cookie_valid = (result > 0); 1478 tfo_response_cookie_valid = (result >= 0); 1479 } 1480 1481 /* 1482 * Remember the TFO pending counter as it will have to be 1483 * decremented below if we don't make it to syncache_tfo_expand(). 1484 */ 1485 tfo_pending = tp->t_tfo_pending; 1486 } 1487 1488 /* By the time we drop the lock these should no longer be used. */ 1489 so = NULL; 1490 tp = NULL; 1491 1492#ifdef MAC 1493 if (mac_syncache_init(&maclabel) != 0) { 1494 INP_WUNLOCK(inp); 1495 goto done; 1496 } else 1497 mac_syncache_create(maclabel, inp); 1498#endif 1499 if (!tfo_cookie_valid) 1500 INP_WUNLOCK(inp); 1501 1502 /* 1503 * Remember the IP options, if any. 1504 */ 1505#ifdef INET6 1506 if (!(inc->inc_flags & INC_ISIPV6)) 1507#endif 1508#ifdef INET 1509 ipopts = (m) ? ip_srcroute(m) : NULL; 1510#else 1511 ipopts = NULL; 1512#endif 1513 1514#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1515 /* 1516 * If listening socket requested TCP digests, check that received 1517 * SYN has signature and it is correct. If signature doesn't match 1518 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1519 */ 1520 if (ltflags & TF_SIGNATURE) { 1521 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1522 TCPSTAT_INC(tcps_sig_err_nosigopt); 1523 goto done; 1524 } 1525 if (!TCPMD5_ENABLED() || 1526 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1527 goto done; 1528 } 1529#endif /* TCP_SIGNATURE */ 1530 /* 1531 * See if we already have an entry for this connection. 1532 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1533 * 1534 * XXX: should the syncache be re-initialized with the contents 1535 * of the new SYN here (which may have different options?) 1536 * 1537 * XXX: We do not check the sequence number to see if this is a 1538 * real retransmit or a new connection attempt. The question is 1539 * how to handle such a case; either ignore it as spoofed, or 1540 * drop the current entry and create a new one? 1541 */ 1542 if (syncache_cookiesonly()) { 1543 sc = NULL; 1544 sch = syncache_hashbucket(inc); 1545 locked = false; 1546 } else { 1547 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1548 locked = true; 1549 SCH_LOCK_ASSERT(sch); 1550 } 1551 if (sc != NULL) { 1552 if (tfo_cookie_valid) 1553 INP_WUNLOCK(inp); 1554 TCPSTAT_INC(tcps_sc_dupsyn); 1555 if (ipopts) { 1556 /* 1557 * If we were remembering a previous source route, 1558 * forget it and use the new one we've been given. 1559 */ 1560 if (sc->sc_ipopts) 1561 (void) m_free(sc->sc_ipopts); 1562 sc->sc_ipopts = ipopts; 1563 } 1564 /* 1565 * Update timestamp if present. 1566 */ 1567 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1568 sc->sc_tsreflect = to->to_tsval; 1569 else 1570 sc->sc_flags &= ~SCF_TIMESTAMP; 1571 /* 1572 * Disable ECN if needed. 1573 */ 1574 if ((sc->sc_flags & SCF_ECN) && 1575 ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) { 1576 sc->sc_flags &= ~SCF_ECN; 1577 } 1578#ifdef MAC 1579 /* 1580 * Since we have already unconditionally allocated label 1581 * storage, free it up. The syncache entry will already 1582 * have an initialized label we can use. 1583 */ 1584 mac_syncache_destroy(&maclabel); 1585#endif 1586 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1587 /* Retransmit SYN|ACK and reset retransmit count. */ 1588 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1589 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1590 "resetting timer and retransmitting SYN|ACK\n", 1591 s, __func__); 1592 free(s, M_TCPLOG); 1593 } 1594 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1595 sc->sc_rxmits = 0; 1596 syncache_timeout(sc, sch, 1); 1597 TCPSTAT_INC(tcps_sndacks); 1598 TCPSTAT_INC(tcps_sndtotal); 1599 } 1600 SCH_UNLOCK(sch); 1601 goto donenoprobe; 1602 } 1603 1604 if (tfo_cookie_valid) { 1605 bzero(&scs, sizeof(scs)); 1606 sc = &scs; 1607 goto skip_alloc; 1608 } 1609 1610 /* 1611 * Skip allocating a syncache entry if we are just going to discard 1612 * it later. 1613 */ 1614 if (!locked) { 1615 bzero(&scs, sizeof(scs)); 1616 sc = &scs; 1617 } else 1618 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1619 if (sc == NULL) { 1620 /* 1621 * The zone allocator couldn't provide more entries. 1622 * Treat this as if the cache was full; drop the oldest 1623 * entry and insert the new one. 1624 */ 1625 TCPSTAT_INC(tcps_sc_zonefail); 1626 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1627 sch->sch_last_overflow = time_uptime; 1628 syncache_drop(sc, sch); 1629 syncache_pause(inc); 1630 } 1631 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1632 if (sc == NULL) { 1633 if (V_tcp_syncookies) { 1634 bzero(&scs, sizeof(scs)); 1635 sc = &scs; 1636 } else { 1637 KASSERT(locked, 1638 ("%s: bucket unexpectedly unlocked", 1639 __func__)); 1640 SCH_UNLOCK(sch); 1641 if (ipopts) 1642 (void) m_free(ipopts); 1643 goto done; 1644 } 1645 } 1646 } 1647 1648skip_alloc: 1649 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1650 sc->sc_tfo_cookie = &tfo_response_cookie; 1651 1652 /* 1653 * Fill in the syncache values. 1654 */ 1655#ifdef MAC 1656 sc->sc_label = maclabel; 1657#endif 1658 sc->sc_cred = cred; 1659 sc->sc_port = port; 1660 cred = NULL; 1661 sc->sc_ipopts = ipopts; 1662 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1663 sc->sc_ip_tos = ip_tos; 1664 sc->sc_ip_ttl = ip_ttl; 1665#ifdef TCP_OFFLOAD 1666 sc->sc_tod = tod; 1667 sc->sc_todctx = todctx; 1668#endif 1669 sc->sc_irs = th->th_seq; 1670 sc->sc_flags = 0; 1671 sc->sc_flowlabel = 0; 1672 1673 /* 1674 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1675 * win was derived from socket earlier in the function. 1676 */ 1677 win = imax(win, 0); 1678 win = imin(win, TCP_MAXWIN); 1679 sc->sc_wnd = win; 1680 1681 if (V_tcp_do_rfc1323 && 1682 !(ltflags & TF_NOOPT)) { 1683 /* 1684 * A timestamp received in a SYN makes 1685 * it ok to send timestamp requests and replies. 1686 */ 1687 if (to->to_flags & TOF_TS) { 1688 sc->sc_tsreflect = to->to_tsval; 1689 sc->sc_flags |= SCF_TIMESTAMP; 1690 sc->sc_tsoff = tcp_new_ts_offset(inc); 1691 } 1692 if (to->to_flags & TOF_SCALE) { 1693 int wscale = 0; 1694 1695 /* 1696 * Pick the smallest possible scaling factor that 1697 * will still allow us to scale up to sb_max, aka 1698 * kern.ipc.maxsockbuf. 1699 * 1700 * We do this because there are broken firewalls that 1701 * will corrupt the window scale option, leading to 1702 * the other endpoint believing that our advertised 1703 * window is unscaled. At scale factors larger than 1704 * 5 the unscaled window will drop below 1500 bytes, 1705 * leading to serious problems when traversing these 1706 * broken firewalls. 1707 * 1708 * With the default maxsockbuf of 256K, a scale factor 1709 * of 3 will be chosen by this algorithm. Those who 1710 * choose a larger maxsockbuf should watch out 1711 * for the compatibility problems mentioned above. 1712 * 1713 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1714 * or <SYN,ACK>) segment itself is never scaled. 1715 */ 1716 while (wscale < TCP_MAX_WINSHIFT && 1717 (TCP_MAXWIN << wscale) < sb_max) 1718 wscale++; 1719 sc->sc_requested_r_scale = wscale; 1720 sc->sc_requested_s_scale = to->to_wscale; 1721 sc->sc_flags |= SCF_WINSCALE; 1722 } 1723 } 1724#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1725 /* 1726 * If listening socket requested TCP digests, flag this in the 1727 * syncache so that syncache_respond() will do the right thing 1728 * with the SYN+ACK. 1729 */ 1730 if (ltflags & TF_SIGNATURE) 1731 sc->sc_flags |= SCF_SIGNATURE; 1732#endif /* TCP_SIGNATURE */ 1733 if (to->to_flags & TOF_SACKPERM) 1734 sc->sc_flags |= SCF_SACK; 1735 if (to->to_flags & TOF_MSS) 1736 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1737 if (ltflags & TF_NOOPT) 1738 sc->sc_flags |= SCF_NOOPT; 1739 if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) && 1740 V_tcp_do_ecn) 1741 sc->sc_flags |= SCF_ECN; 1742 1743 if (V_tcp_syncookies) 1744 sc->sc_iss = syncookie_generate(sch, sc); 1745 else 1746 sc->sc_iss = arc4random(); 1747#ifdef INET6 1748 if (autoflowlabel) { 1749 if (V_tcp_syncookies) 1750 sc->sc_flowlabel = sc->sc_iss; 1751 else 1752 sc->sc_flowlabel = ip6_randomflowlabel(); 1753 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1754 } 1755#endif 1756 if (locked) 1757 SCH_UNLOCK(sch); 1758 1759 if (tfo_cookie_valid) { 1760 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1761 /* INP_WUNLOCK(inp) will be performed by the caller */ 1762 rv = 1; 1763 goto tfo_expanded; 1764 } 1765 1766 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1767 /* 1768 * Do a standard 3-way handshake. 1769 */ 1770 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1771 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1772 syncache_free(sc); 1773 else if (sc != &scs) 1774 syncache_insert(sc, sch); /* locks and unlocks sch */ 1775 TCPSTAT_INC(tcps_sndacks); 1776 TCPSTAT_INC(tcps_sndtotal); 1777 } else { 1778 if (sc != &scs) 1779 syncache_free(sc); 1780 TCPSTAT_INC(tcps_sc_dropped); 1781 } 1782 goto donenoprobe; 1783 1784done: 1785 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1786donenoprobe: 1787 if (m) { 1788 *lsop = NULL; 1789 m_freem(m); 1790 } 1791 /* 1792 * If tfo_pending is not NULL here, then a TFO SYN that did not 1793 * result in a new socket was processed and the associated pending 1794 * counter has not yet been decremented. All such TFO processing paths 1795 * transit this point. 1796 */ 1797 if (tfo_pending != NULL) 1798 tcp_fastopen_decrement_counter(tfo_pending); 1799 1800tfo_expanded: 1801 if (cred != NULL) 1802 crfree(cred); 1803#ifdef MAC 1804 if (sc == &scs) 1805 mac_syncache_destroy(&maclabel); 1806#endif 1807 return (rv); 1808} 1809 1810/* 1811 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1812 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1813 */ 1814static int 1815syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags) 1816{ 1817 struct ip *ip = NULL; 1818 struct mbuf *m; 1819 struct tcphdr *th = NULL; 1820 struct udphdr *udp = NULL; 1821 int optlen, error = 0; /* Make compiler happy */ 1822 u_int16_t hlen, tlen, mssopt, ulen; 1823 struct tcpopt to; 1824#ifdef INET6 1825 struct ip6_hdr *ip6 = NULL; 1826#endif 1827 1828 NET_EPOCH_ASSERT(); 1829 1830 hlen = 1831#ifdef INET6 1832 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1833#endif 1834 sizeof(struct ip); 1835 tlen = hlen + sizeof(struct tcphdr); 1836 if (sc->sc_port) { 1837 tlen += sizeof(struct udphdr); 1838 } 1839 /* Determine MSS we advertize to other end of connection. */ 1840 mssopt = tcp_mssopt(&sc->sc_inc); 1841 if (sc->sc_port) 1842 mssopt -= V_tcp_udp_tunneling_overhead; 1843 mssopt = max(mssopt, V_tcp_minmss); 1844 1845 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1846 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1847 ("syncache: mbuf too small")); 1848 1849 /* Create the IP+TCP header from scratch. */ 1850 m = m_gethdr(M_NOWAIT, MT_DATA); 1851 if (m == NULL) 1852 return (ENOBUFS); 1853#ifdef MAC 1854 mac_syncache_create_mbuf(sc->sc_label, m); 1855#endif 1856 m->m_data += max_linkhdr; 1857 m->m_len = tlen; 1858 m->m_pkthdr.len = tlen; 1859 m->m_pkthdr.rcvif = NULL; 1860 1861#ifdef INET6 1862 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1863 ip6 = mtod(m, struct ip6_hdr *); 1864 ip6->ip6_vfc = IPV6_VERSION; 1865 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1866 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1867 ip6->ip6_plen = htons(tlen - hlen); 1868 /* ip6_hlim is set after checksum */ 1869 /* Zero out traffic class and flow label. */ 1870 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1871 ip6->ip6_flow |= sc->sc_flowlabel; 1872 if (sc->sc_port != 0) { 1873 ip6->ip6_nxt = IPPROTO_UDP; 1874 udp = (struct udphdr *)(ip6 + 1); 1875 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1876 udp->uh_dport = sc->sc_port; 1877 ulen = (tlen - sizeof(struct ip6_hdr)); 1878 th = (struct tcphdr *)(udp + 1); 1879 } else { 1880 ip6->ip6_nxt = IPPROTO_TCP; 1881 th = (struct tcphdr *)(ip6 + 1); 1882 } 1883 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20); 1884 } 1885#endif 1886#if defined(INET6) && defined(INET) 1887 else 1888#endif 1889#ifdef INET 1890 { 1891 ip = mtod(m, struct ip *); 1892 ip->ip_v = IPVERSION; 1893 ip->ip_hl = sizeof(struct ip) >> 2; 1894 ip->ip_len = htons(tlen); 1895 ip->ip_id = 0; 1896 ip->ip_off = 0; 1897 ip->ip_sum = 0; 1898 ip->ip_src = sc->sc_inc.inc_laddr; 1899 ip->ip_dst = sc->sc_inc.inc_faddr; 1900 ip->ip_ttl = sc->sc_ip_ttl; 1901 ip->ip_tos = sc->sc_ip_tos; 1902 1903 /* 1904 * See if we should do MTU discovery. Route lookups are 1905 * expensive, so we will only unset the DF bit if: 1906 * 1907 * 1) path_mtu_discovery is disabled 1908 * 2) the SCF_UNREACH flag has been set 1909 */ 1910 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1911 ip->ip_off |= htons(IP_DF); 1912 if (sc->sc_port == 0) { 1913 ip->ip_p = IPPROTO_TCP; 1914 th = (struct tcphdr *)(ip + 1); 1915 } else { 1916 ip->ip_p = IPPROTO_UDP; 1917 udp = (struct udphdr *)(ip + 1); 1918 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1919 udp->uh_dport = sc->sc_port; 1920 ulen = (tlen - sizeof(struct ip)); 1921 th = (struct tcphdr *)(udp + 1); 1922 } 1923 } 1924#endif /* INET */ 1925 th->th_sport = sc->sc_inc.inc_lport; 1926 th->th_dport = sc->sc_inc.inc_fport; 1927 1928 if (flags & TH_SYN) 1929 th->th_seq = htonl(sc->sc_iss); 1930 else 1931 th->th_seq = htonl(sc->sc_iss + 1); 1932 th->th_ack = htonl(sc->sc_irs + 1); 1933 th->th_off = sizeof(struct tcphdr) >> 2; 1934 th->th_x2 = 0; 1935 th->th_flags = flags; 1936 th->th_win = htons(sc->sc_wnd); 1937 th->th_urp = 0; 1938 1939 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) { 1940 th->th_flags |= TH_ECE; 1941 TCPSTAT_INC(tcps_ecn_shs); 1942 } 1943 1944 /* Tack on the TCP options. */ 1945 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1946 to.to_flags = 0; 1947 1948 if (flags & TH_SYN) { 1949 to.to_mss = mssopt; 1950 to.to_flags = TOF_MSS; 1951 if (sc->sc_flags & SCF_WINSCALE) { 1952 to.to_wscale = sc->sc_requested_r_scale; 1953 to.to_flags |= TOF_SCALE; 1954 } 1955 if (sc->sc_flags & SCF_SACK) 1956 to.to_flags |= TOF_SACKPERM; 1957#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1958 if (sc->sc_flags & SCF_SIGNATURE) 1959 to.to_flags |= TOF_SIGNATURE; 1960#endif 1961 if (sc->sc_tfo_cookie) { 1962 to.to_flags |= TOF_FASTOPEN; 1963 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1964 to.to_tfo_cookie = sc->sc_tfo_cookie; 1965 /* don't send cookie again when retransmitting response */ 1966 sc->sc_tfo_cookie = NULL; 1967 } 1968 } 1969 if (sc->sc_flags & SCF_TIMESTAMP) { 1970 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1971 to.to_tsecr = sc->sc_tsreflect; 1972 to.to_flags |= TOF_TS; 1973 } 1974 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1975 1976 /* Adjust headers by option size. */ 1977 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1978 m->m_len += optlen; 1979 m->m_pkthdr.len += optlen; 1980#ifdef INET6 1981 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1982 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1983 else 1984#endif 1985 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1986#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1987 if (sc->sc_flags & SCF_SIGNATURE) { 1988 KASSERT(to.to_flags & TOF_SIGNATURE, 1989 ("tcp_addoptions() didn't set tcp_signature")); 1990 1991 /* NOTE: to.to_signature is inside of mbuf */ 1992 if (!TCPMD5_ENABLED() || 1993 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1994 m_freem(m); 1995 return (EACCES); 1996 } 1997 } 1998#endif 1999 } else 2000 optlen = 0; 2001 2002 if (udp) { 2003 ulen += optlen; 2004 udp->uh_ulen = htons(ulen); 2005 } 2006 M_SETFIB(m, sc->sc_inc.inc_fibnum); 2007 /* 2008 * If we have peer's SYN and it has a flowid, then let's assign it to 2009 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 2010 * to SYN|ACK due to lack of inp here. 2011 */ 2012 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 2013 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 2014 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 2015 } 2016#ifdef INET6 2017 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 2018 if (sc->sc_port) { 2019 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 2020 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2021 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, 2022 IPPROTO_UDP, 0); 2023 th->th_sum = htons(0); 2024 } else { 2025 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 2026 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2027 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 2028 IPPROTO_TCP, 0); 2029 } 2030 ip6->ip6_hlim = sc->sc_ip_ttl; 2031#ifdef TCP_OFFLOAD 2032 if (ADDED_BY_TOE(sc)) { 2033 struct toedev *tod = sc->sc_tod; 2034 2035 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2036 2037 return (error); 2038 } 2039#endif 2040 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 2041 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 2042 } 2043#endif 2044#if defined(INET6) && defined(INET) 2045 else 2046#endif 2047#ifdef INET 2048 { 2049 if (sc->sc_port) { 2050 m->m_pkthdr.csum_flags = CSUM_UDP; 2051 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2052 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 2053 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 2054 th->th_sum = htons(0); 2055 } else { 2056 m->m_pkthdr.csum_flags = CSUM_TCP; 2057 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2058 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 2059 htons(tlen + optlen - hlen + IPPROTO_TCP)); 2060 } 2061#ifdef TCP_OFFLOAD 2062 if (ADDED_BY_TOE(sc)) { 2063 struct toedev *tod = sc->sc_tod; 2064 2065 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2066 2067 return (error); 2068 } 2069#endif 2070 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 2071 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 2072 } 2073#endif 2074 return (error); 2075} 2076 2077/* 2078 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 2079 * that exceed the capacity of the syncache by avoiding the storage of any 2080 * of the SYNs we receive. Syncookies defend against blind SYN flooding 2081 * attacks where the attacker does not have access to our responses. 2082 * 2083 * Syncookies encode and include all necessary information about the 2084 * connection setup within the SYN|ACK that we send back. That way we 2085 * can avoid keeping any local state until the ACK to our SYN|ACK returns 2086 * (if ever). Normally the syncache and syncookies are running in parallel 2087 * with the latter taking over when the former is exhausted. When matching 2088 * syncache entry is found the syncookie is ignored. 2089 * 2090 * The only reliable information persisting the 3WHS is our initial sequence 2091 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 2092 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 2093 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 2094 * returns and signifies a legitimate connection if it matches the ACK. 2095 * 2096 * The available space of 32 bits to store the hash and to encode the SYN 2097 * option information is very tight and we should have at least 24 bits for 2098 * the MAC to keep the number of guesses by blind spoofing reasonably high. 2099 * 2100 * SYN option information we have to encode to fully restore a connection: 2101 * MSS: is imporant to chose an optimal segment size to avoid IP level 2102 * fragmentation along the path. The common MSS values can be encoded 2103 * in a 3-bit table. Uncommon values are captured by the next lower value 2104 * in the table leading to a slight increase in packetization overhead. 2105 * WSCALE: is necessary to allow large windows to be used for high delay- 2106 * bandwidth product links. Not scaling the window when it was initially 2107 * negotiated is bad for performance as lack of scaling further decreases 2108 * the apparent available send window. We only need to encode the WSCALE 2109 * we received from the remote end. Our end can be recalculated at any 2110 * time. The common WSCALE values can be encoded in a 3-bit table. 2111 * Uncommon values are captured by the next lower value in the table 2112 * making us under-estimate the available window size halving our 2113 * theoretically possible maximum throughput for that connection. 2114 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 2115 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 2116 * that are included in all segments on a connection. We enable them when 2117 * the ACK has them. 2118 * 2119 * Security of syncookies and attack vectors: 2120 * 2121 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 2122 * together with the gloabl secret to make it unique per connection attempt. 2123 * Thus any change of any of those parameters results in a different MAC output 2124 * in an unpredictable way unless a collision is encountered. 24 bits of the 2125 * MAC are embedded into the ISS. 2126 * 2127 * To prevent replay attacks two rotating global secrets are updated with a 2128 * new random value every 15 seconds. The life-time of a syncookie is thus 2129 * 15-30 seconds. 2130 * 2131 * Vector 1: Attacking the secret. This requires finding a weakness in the 2132 * MAC itself or the way it is used here. The attacker can do a chosen plain 2133 * text attack by varying and testing the all parameters under his control. 2134 * The strength depends on the size and randomness of the secret, and the 2135 * cryptographic security of the MAC function. Due to the constant updating 2136 * of the secret the attacker has at most 29.999 seconds to find the secret 2137 * and launch spoofed connections. After that he has to start all over again. 2138 * 2139 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 2140 * size an average of 4,823 attempts are required for a 50% chance of success 2141 * to spoof a single syncookie (birthday collision paradox). However the 2142 * attacker is blind and doesn't know if one of his attempts succeeded unless 2143 * he has a side channel to interfere success from. A single connection setup 2144 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 2145 * This many attempts are required for each one blind spoofed connection. For 2146 * every additional spoofed connection he has to launch another N attempts. 2147 * Thus for a sustained rate 100 spoofed connections per second approximately 2148 * 1,800,000 packets per second would have to be sent. 2149 * 2150 * NB: The MAC function should be fast so that it doesn't become a CPU 2151 * exhaustion attack vector itself. 2152 * 2153 * References: 2154 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 2155 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 2156 * http://cr.yp.to/syncookies.html (overview) 2157 * http://cr.yp.to/syncookies/archive (details) 2158 * 2159 * 2160 * Schematic construction of a syncookie enabled Initial Sequence Number: 2161 * 0 1 2 3 2162 * 12345678901234567890123456789012 2163 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2164 * 2165 * x 24 MAC (truncated) 2166 * W 3 Send Window Scale index 2167 * M 3 MSS index 2168 * S 1 SACK permitted 2169 * P 1 Odd/even secret 2170 */ 2171 2172/* 2173 * Distribution and probability of certain MSS values. Those in between are 2174 * rounded down to the next lower one. 2175 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2176 * .2% .3% 5% 7% 7% 20% 15% 45% 2177 */ 2178static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2179 2180/* 2181 * Distribution and probability of certain WSCALE values. We have to map the 2182 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2183 * bits based on prevalence of certain values. Where we don't have an exact 2184 * match for are rounded down to the next lower one letting us under-estimate 2185 * the true available window. At the moment this would happen only for the 2186 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2187 * and window size). The absence of the WSCALE option (no scaling in either 2188 * direction) is encoded with index zero. 2189 * [WSCALE values histograms, Allman, 2012] 2190 * X 10 10 35 5 6 14 10% by host 2191 * X 11 4 5 5 18 49 3% by connections 2192 */ 2193static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2194 2195/* 2196 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2197 * and good cryptographic properties. 2198 */ 2199static uint32_t 2200syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2201 uint8_t *secbits, uintptr_t secmod) 2202{ 2203 SIPHASH_CTX ctx; 2204 uint32_t siphash[2]; 2205 2206 SipHash24_Init(&ctx); 2207 SipHash_SetKey(&ctx, secbits); 2208 switch (inc->inc_flags & INC_ISIPV6) { 2209#ifdef INET 2210 case 0: 2211 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2212 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2213 break; 2214#endif 2215#ifdef INET6 2216 case INC_ISIPV6: 2217 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2218 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2219 break; 2220#endif 2221 } 2222 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2223 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2224 SipHash_Update(&ctx, &irs, sizeof(irs)); 2225 SipHash_Update(&ctx, &flags, sizeof(flags)); 2226 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2227 SipHash_Final((u_int8_t *)&siphash, &ctx); 2228 2229 return (siphash[0] ^ siphash[1]); 2230} 2231 2232static tcp_seq 2233syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2234{ 2235 u_int i, secbit, wscale; 2236 uint32_t iss, hash; 2237 uint8_t *secbits; 2238 union syncookie cookie; 2239 2240 cookie.cookie = 0; 2241 2242 /* Map our computed MSS into the 3-bit index. */ 2243 for (i = nitems(tcp_sc_msstab) - 1; 2244 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2245 i--) 2246 ; 2247 cookie.flags.mss_idx = i; 2248 2249 /* 2250 * Map the send window scale into the 3-bit index but only if 2251 * the wscale option was received. 2252 */ 2253 if (sc->sc_flags & SCF_WINSCALE) { 2254 wscale = sc->sc_requested_s_scale; 2255 for (i = nitems(tcp_sc_wstab) - 1; 2256 tcp_sc_wstab[i] > wscale && i > 0; 2257 i--) 2258 ; 2259 cookie.flags.wscale_idx = i; 2260 } 2261 2262 /* Can we do SACK? */ 2263 if (sc->sc_flags & SCF_SACK) 2264 cookie.flags.sack_ok = 1; 2265 2266 /* Which of the two secrets to use. */ 2267 secbit = V_tcp_syncache.secret.oddeven & 0x1; 2268 cookie.flags.odd_even = secbit; 2269 2270 secbits = V_tcp_syncache.secret.key[secbit]; 2271 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2272 (uintptr_t)sch); 2273 2274 /* 2275 * Put the flags into the hash and XOR them to get better ISS number 2276 * variance. This doesn't enhance the cryptographic strength and is 2277 * done to prevent the 8 cookie bits from showing up directly on the 2278 * wire. 2279 */ 2280 iss = hash & ~0xff; 2281 iss |= cookie.cookie ^ (hash >> 24); 2282 2283 TCPSTAT_INC(tcps_sc_sendcookie); 2284 return (iss); 2285} 2286 2287static struct syncache * 2288syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2289 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2290 struct socket *lso, uint16_t port) 2291{ 2292 uint32_t hash; 2293 uint8_t *secbits; 2294 tcp_seq ack, seq; 2295 int wnd, wscale = 0; 2296 union syncookie cookie; 2297 2298 /* 2299 * Pull information out of SYN-ACK/ACK and revert sequence number 2300 * advances. 2301 */ 2302 ack = th->th_ack - 1; 2303 seq = th->th_seq - 1; 2304 2305 /* 2306 * Unpack the flags containing enough information to restore the 2307 * connection. 2308 */ 2309 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2310 2311 /* Which of the two secrets to use. */ 2312 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; 2313 2314 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2315 2316 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2317 if ((ack & ~0xff) != (hash & ~0xff)) 2318 return (NULL); 2319 2320 /* Fill in the syncache values. */ 2321 sc->sc_flags = 0; 2322 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2323 sc->sc_ipopts = NULL; 2324 2325 sc->sc_irs = seq; 2326 sc->sc_iss = ack; 2327 2328 switch (inc->inc_flags & INC_ISIPV6) { 2329#ifdef INET 2330 case 0: 2331 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2332 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2333 break; 2334#endif 2335#ifdef INET6 2336 case INC_ISIPV6: 2337 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2338 sc->sc_flowlabel = 2339 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2340 break; 2341#endif 2342 } 2343 2344 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2345 2346 /* We can simply recompute receive window scale we sent earlier. */ 2347 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2348 wscale++; 2349 2350 /* Only use wscale if it was enabled in the orignal SYN. */ 2351 if (cookie.flags.wscale_idx > 0) { 2352 sc->sc_requested_r_scale = wscale; 2353 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2354 sc->sc_flags |= SCF_WINSCALE; 2355 } 2356 2357 wnd = lso->sol_sbrcv_hiwat; 2358 wnd = imax(wnd, 0); 2359 wnd = imin(wnd, TCP_MAXWIN); 2360 sc->sc_wnd = wnd; 2361 2362 if (cookie.flags.sack_ok) 2363 sc->sc_flags |= SCF_SACK; 2364 2365 if (to->to_flags & TOF_TS) { 2366 sc->sc_flags |= SCF_TIMESTAMP; 2367 sc->sc_tsreflect = to->to_tsval; 2368 sc->sc_tsoff = tcp_new_ts_offset(inc); 2369 } 2370 2371 if (to->to_flags & TOF_SIGNATURE) 2372 sc->sc_flags |= SCF_SIGNATURE; 2373 2374 sc->sc_rxmits = 0; 2375 2376 sc->sc_port = port; 2377 2378 TCPSTAT_INC(tcps_sc_recvcookie); 2379 return (sc); 2380} 2381 2382#ifdef INVARIANTS 2383static int 2384syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2385 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2386 struct socket *lso, uint16_t port) 2387{ 2388 struct syncache scs, *scx; 2389 char *s; 2390 2391 bzero(&scs, sizeof(scs)); 2392 scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port); 2393 2394 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2395 return (0); 2396 2397 if (scx != NULL) { 2398 if (sc->sc_peer_mss != scx->sc_peer_mss) 2399 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2400 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2401 2402 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2403 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2404 s, __func__, sc->sc_requested_r_scale, 2405 scx->sc_requested_r_scale); 2406 2407 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2408 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2409 s, __func__, sc->sc_requested_s_scale, 2410 scx->sc_requested_s_scale); 2411 2412 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2413 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2414 } 2415 2416 if (s != NULL) 2417 free(s, M_TCPLOG); 2418 return (0); 2419} 2420#endif /* INVARIANTS */ 2421 2422static void 2423syncookie_reseed(void *arg) 2424{ 2425 struct tcp_syncache *sc = arg; 2426 uint8_t *secbits; 2427 int secbit; 2428 2429 /* 2430 * Reseeding the secret doesn't have to be protected by a lock. 2431 * It only must be ensured that the new random values are visible 2432 * to all CPUs in a SMP environment. The atomic with release 2433 * semantics ensures that. 2434 */ 2435 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2436 secbits = sc->secret.key[secbit]; 2437 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2438 atomic_add_rel_int(&sc->secret.oddeven, 1); 2439 2440 /* Reschedule ourself. */ 2441 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2442} 2443 2444/* 2445 * We have overflowed a bucket. Let's pause dealing with the syncache. 2446 * This function will increment the bucketoverflow statistics appropriately 2447 * (once per pause when pausing is enabled; otherwise, once per overflow). 2448 */ 2449static void 2450syncache_pause(struct in_conninfo *inc) 2451{ 2452 time_t delta; 2453 const char *s; 2454 2455 /* XXX: 2456 * 2. Add sysctl read here so we don't get the benefit of this 2457 * change without the new sysctl. 2458 */ 2459 2460 /* 2461 * Try an unlocked read. If we already know that another thread 2462 * has activated the feature, there is no need to proceed. 2463 */ 2464 if (V_tcp_syncache.paused) 2465 return; 2466 2467 /* Are cookied enabled? If not, we can't pause. */ 2468 if (!V_tcp_syncookies) { 2469 TCPSTAT_INC(tcps_sc_bucketoverflow); 2470 return; 2471 } 2472 2473 /* 2474 * We may be the first thread to find an overflow. Get the lock 2475 * and evaluate if we need to take action. 2476 */ 2477 mtx_lock(&V_tcp_syncache.pause_mtx); 2478 if (V_tcp_syncache.paused) { 2479 mtx_unlock(&V_tcp_syncache.pause_mtx); 2480 return; 2481 } 2482 2483 /* Activate protection. */ 2484 V_tcp_syncache.paused = true; 2485 TCPSTAT_INC(tcps_sc_bucketoverflow); 2486 2487 /* 2488 * Determine the last backoff time. If we are seeing a re-newed 2489 * attack within that same time after last reactivating the syncache, 2490 * consider it an extension of the same attack. 2491 */ 2492 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; 2493 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { 2494 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { 2495 delta <<= 1; 2496 V_tcp_syncache.pause_backoff++; 2497 } 2498 } else { 2499 delta = TCP_SYNCACHE_PAUSE_TIME; 2500 V_tcp_syncache.pause_backoff = 0; 2501 } 2502 2503 /* Log a warning, including IP addresses, if able. */ 2504 if (inc != NULL) 2505 s = tcp_log_addrs(inc, NULL, NULL, NULL); 2506 else 2507 s = (const char *)NULL; 2508 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " 2509 "the next %lld seconds%s%s%s\n", (long long)delta, 2510 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", 2511 (s != NULL) ? ")" : ""); 2512 free(__DECONST(void *, s), M_TCPLOG); 2513 2514 /* Use the calculated delta to set a new pause time. */ 2515 V_tcp_syncache.pause_until = time_uptime + delta; 2516 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, 2517 &V_tcp_syncache); 2518 mtx_unlock(&V_tcp_syncache.pause_mtx); 2519} 2520 2521/* Evaluate whether we need to unpause. */ 2522static void 2523syncache_unpause(void *arg) 2524{ 2525 struct tcp_syncache *sc; 2526 time_t delta; 2527 2528 sc = arg; 2529 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); 2530 callout_deactivate(&sc->pause_co); 2531 2532 /* 2533 * Check to make sure we are not running early. If the pause 2534 * time has expired, then deactivate the protection. 2535 */ 2536 if ((delta = sc->pause_until - time_uptime) > 0) 2537 callout_schedule(&sc->pause_co, delta * hz); 2538 else 2539 sc->paused = false; 2540} 2541 2542/* 2543 * Exports the syncache entries to userland so that netstat can display 2544 * them alongside the other sockets. This function is intended to be 2545 * called only from tcp_pcblist. 2546 * 2547 * Due to concurrency on an active system, the number of pcbs exported 2548 * may have no relation to max_pcbs. max_pcbs merely indicates the 2549 * amount of space the caller allocated for this function to use. 2550 */ 2551int 2552syncache_pcblist(struct sysctl_req *req) 2553{ 2554 struct xtcpcb xt; 2555 struct syncache *sc; 2556 struct syncache_head *sch; 2557 int error, i; 2558 2559 bzero(&xt, sizeof(xt)); 2560 xt.xt_len = sizeof(xt); 2561 xt.t_state = TCPS_SYN_RECEIVED; 2562 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2563 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2564 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2565 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2566 2567 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 2568 sch = &V_tcp_syncache.hashbase[i]; 2569 SCH_LOCK(sch); 2570 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2571 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2572 continue; 2573 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2574 xt.xt_inp.inp_vflag = INP_IPV6; 2575 else 2576 xt.xt_inp.inp_vflag = INP_IPV4; 2577 xt.xt_encaps_port = sc->sc_port; 2578 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2579 sizeof (struct in_conninfo)); 2580 error = SYSCTL_OUT(req, &xt, sizeof xt); 2581 if (error) { 2582 SCH_UNLOCK(sch); 2583 return (0); 2584 } 2585 } 2586 SCH_UNLOCK(sch); 2587 } 2588 2589 return (0); 2590} 2591