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_systm.h> 73 #include <netinet/ip.h> 74 #include <netinet/in_var.h> 75 #include <netinet/in_pcb.h> 76 #include <netinet/ip_var.h> 77 #include <netinet/ip_options.h> 78 #ifdef INET6 79 #include <netinet/ip6.h> 80 #include <netinet/icmp6.h> 81 #include <netinet6/nd6.h> 82 #include <netinet6/ip6_var.h> 83 #include <netinet6/in6_pcb.h> 84 #endif 85 #include <netinet/tcp.h> 86 #include <netinet/tcp_fastopen.h> 87 #include <netinet/tcp_fsm.h> 88 #include <netinet/tcp_seq.h> 89 #include <netinet/tcp_timer.h> 90 #include <netinet/tcp_var.h> 91 #include <netinet/tcp_syncache.h> 92 #ifdef INET6 93 #include <netinet6/tcp6_var.h> 94 #endif 95 #ifdef TCP_OFFLOAD 96 #include <netinet/toecore.h> 97 #endif 98 99 #include <netipsec/ipsec_support.h> 100 101 #include <machine/in_cksum.h> 102 103 #include <security/mac/mac_framework.h> 104 105 static VNET_DEFINE(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 static VNET_DEFINE(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 static VNET_DEFINE(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 *, struct syncache_head *, int, 133 const struct mbuf *); 134 static struct socket *syncache_socket(struct syncache *, struct socket *, 135 struct mbuf *m); 136 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 137 int docallout); 138 static void syncache_timer(void *); 139 140 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 141 uint8_t *, uintptr_t); 142 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 143 static struct syncache 144 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 145 struct syncache *, struct tcphdr *, struct tcpopt *, 146 struct socket *); 147 static void syncookie_reseed(void *); 148 #ifdef INVARIANTS 149 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 150 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 151 struct socket *lso); 152 #endif 153 154 /* 155 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 156 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds, 157 * the odds are that the user has given up attempting to connect by then. 158 */ 159 #define SYNCACHE_MAXREXMTS 3 160 161 /* Arbitrary values */ 162 #define TCP_SYNCACHE_HASHSIZE 512 163 #define TCP_SYNCACHE_BUCKETLIMIT 30 164 165 static VNET_DEFINE(struct tcp_syncache, tcp_syncache); 166 #define V_tcp_syncache VNET(tcp_syncache) 167 168 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, 169 "TCP SYN cache"); 170 171 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 172 &VNET_NAME(tcp_syncache.bucket_limit), 0, 173 "Per-bucket hash limit for syncache"); 174 175 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 176 &VNET_NAME(tcp_syncache.cache_limit), 0, 177 "Overall entry limit for syncache"); 178 179 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 180 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 181 182 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 183 &VNET_NAME(tcp_syncache.hashsize), 0, 184 "Size of TCP syncache hashtable"); 185 186 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW, 187 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 188 "Limit on SYN/ACK retransmissions"); 189 190 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 191 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 192 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 193 "Send reset on socket allocation failure"); 194 195 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 196 197 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 198 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 199 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 200 201 /* 202 * Requires the syncache entry to be already removed from the bucket list. 203 */ 204 static void 205 syncache_free(struct syncache *sc) 206 { 207 208 if (sc->sc_ipopts) 209 (void) m_free(sc->sc_ipopts); 210 if (sc->sc_cred) 211 crfree(sc->sc_cred); 212 #ifdef MAC 213 mac_syncache_destroy(&sc->sc_label); 214 #endif 215 216 uma_zfree(V_tcp_syncache.zone, sc); 217 } 218 219 void 220 syncache_init(void) 221 { 222 int i; 223 224 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 225 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 226 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 227 V_tcp_syncache.hash_secret = arc4random(); 228 229 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 230 &V_tcp_syncache.hashsize); 231 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 232 &V_tcp_syncache.bucket_limit); 233 if (!powerof2(V_tcp_syncache.hashsize) || 234 V_tcp_syncache.hashsize == 0) { 235 printf("WARNING: syncache hash size is not a power of 2.\n"); 236 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 237 } 238 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 239 240 /* Set limits. */ 241 V_tcp_syncache.cache_limit = 242 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 243 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 244 &V_tcp_syncache.cache_limit); 245 246 /* Allocate the hash table. */ 247 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 248 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 249 250 #ifdef VIMAGE 251 V_tcp_syncache.vnet = curvnet; 252 #endif 253 254 /* Initialize the hash buckets. */ 255 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 256 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 257 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 258 NULL, MTX_DEF); 259 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 260 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 261 V_tcp_syncache.hashbase[i].sch_length = 0; 262 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 263 V_tcp_syncache.hashbase[i].sch_last_overflow = 264 -(SYNCOOKIE_LIFETIME + 1); 265 } 266 267 /* Create the syncache entry zone. */ 268 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 269 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 270 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 271 V_tcp_syncache.cache_limit); 272 273 /* Start the SYN cookie reseeder callout. */ 274 callout_init(&V_tcp_syncache.secret.reseed, 1); 275 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 276 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 277 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 278 syncookie_reseed, &V_tcp_syncache); 279 } 280 281 #ifdef VIMAGE 282 void 283 syncache_destroy(void) 284 { 285 struct syncache_head *sch; 286 struct syncache *sc, *nsc; 287 int i; 288 289 /* 290 * Stop the re-seed timer before freeing resources. No need to 291 * possibly schedule it another time. 292 */ 293 callout_drain(&V_tcp_syncache.secret.reseed); 294 295 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 296 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 297 298 sch = &V_tcp_syncache.hashbase[i]; 299 callout_drain(&sch->sch_timer); 300 301 SCH_LOCK(sch); 302 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 303 syncache_drop(sc, sch); 304 SCH_UNLOCK(sch); 305 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 306 ("%s: sch->sch_bucket not empty", __func__)); 307 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 308 __func__, sch->sch_length)); 309 mtx_destroy(&sch->sch_mtx); 310 } 311 312 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 313 ("%s: cache_count not 0", __func__)); 314 315 /* Free the allocated global resources. */ 316 uma_zdestroy(V_tcp_syncache.zone); 317 free(V_tcp_syncache.hashbase, M_SYNCACHE); 318 } 319 #endif 320 321 /* 322 * Inserts a syncache entry into the specified bucket row. 323 * Locks and unlocks the syncache_head autonomously. 324 */ 325 static void 326 syncache_insert(struct syncache *sc, struct syncache_head *sch) 327 { 328 struct syncache *sc2; 329 330 SCH_LOCK(sch); 331 332 /* 333 * Make sure that we don't overflow the per-bucket limit. 334 * If the bucket is full, toss the oldest element. 335 */ 336 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 337 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 338 ("sch->sch_length incorrect")); 339 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 340 sch->sch_last_overflow = time_uptime; 341 syncache_drop(sc2, sch); 342 TCPSTAT_INC(tcps_sc_bucketoverflow); 343 } 344 345 /* Put it into the bucket. */ 346 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 347 sch->sch_length++; 348 349 #ifdef TCP_OFFLOAD 350 if (ADDED_BY_TOE(sc)) { 351 struct toedev *tod = sc->sc_tod; 352 353 tod->tod_syncache_added(tod, sc->sc_todctx); 354 } 355 #endif 356 357 /* Reinitialize the bucket row's timer. */ 358 if (sch->sch_length == 1) 359 sch->sch_nextc = ticks + INT_MAX; 360 syncache_timeout(sc, sch, 1); 361 362 SCH_UNLOCK(sch); 363 364 TCPSTATES_INC(TCPS_SYN_RECEIVED); 365 TCPSTAT_INC(tcps_sc_added); 366 } 367 368 /* 369 * Remove and free entry from syncache bucket row. 370 * Expects locked syncache head. 371 */ 372 static void 373 syncache_drop(struct syncache *sc, struct syncache_head *sch) 374 { 375 376 SCH_LOCK_ASSERT(sch); 377 378 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 379 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 380 sch->sch_length--; 381 382 #ifdef TCP_OFFLOAD 383 if (ADDED_BY_TOE(sc)) { 384 struct toedev *tod = sc->sc_tod; 385 386 tod->tod_syncache_removed(tod, sc->sc_todctx); 387 } 388 #endif 389 390 syncache_free(sc); 391 } 392 393 /* 394 * Engage/reengage time on bucket row. 395 */ 396 static void 397 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 398 { 399 sc->sc_rxttime = ticks + 400 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]); 401 sc->sc_rxmits++; 402 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 403 sch->sch_nextc = sc->sc_rxttime; 404 if (docallout) 405 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 406 syncache_timer, (void *)sch); 407 } 408 } 409 410 /* 411 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 412 * If we have retransmitted an entry the maximum number of times, expire it. 413 * One separate timer for each bucket row. 414 */ 415 static void 416 syncache_timer(void *xsch) 417 { 418 struct syncache_head *sch = (struct syncache_head *)xsch; 419 struct syncache *sc, *nsc; 420 int tick = ticks; 421 char *s; 422 423 CURVNET_SET(sch->sch_sc->vnet); 424 425 /* NB: syncache_head has already been locked by the callout. */ 426 SCH_LOCK_ASSERT(sch); 427 428 /* 429 * In the following cycle we may remove some entries and/or 430 * advance some timeouts, so re-initialize the bucket timer. 431 */ 432 sch->sch_nextc = tick + INT_MAX; 433 434 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 435 /* 436 * We do not check if the listen socket still exists 437 * and accept the case where the listen socket may be 438 * gone by the time we resend the SYN/ACK. We do 439 * not expect this to happens often. If it does, 440 * then the RST will be sent by the time the remote 441 * host does the SYN/ACK->ACK. 442 */ 443 if (TSTMP_GT(sc->sc_rxttime, tick)) { 444 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 445 sch->sch_nextc = sc->sc_rxttime; 446 continue; 447 } 448 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 449 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 450 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 451 "giving up and removing syncache entry\n", 452 s, __func__); 453 free(s, M_TCPLOG); 454 } 455 syncache_drop(sc, sch); 456 TCPSTAT_INC(tcps_sc_stale); 457 continue; 458 } 459 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 460 log(LOG_DEBUG, "%s; %s: Response timeout, " 461 "retransmitting (%u) SYN|ACK\n", 462 s, __func__, sc->sc_rxmits); 463 free(s, M_TCPLOG); 464 } 465 466 syncache_respond(sc, sch, 1, NULL); 467 TCPSTAT_INC(tcps_sc_retransmitted); 468 syncache_timeout(sc, sch, 0); 469 } 470 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 471 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 472 syncache_timer, (void *)(sch)); 473 CURVNET_RESTORE(); 474 } 475 476 /* 477 * Find an entry in the syncache. 478 * Returns always with locked syncache_head plus a matching entry or NULL. 479 */ 480 static struct syncache * 481 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 482 { 483 struct syncache *sc; 484 struct syncache_head *sch; 485 uint32_t hash; 486 487 /* 488 * The hash is built on foreign port + local port + foreign address. 489 * We rely on the fact that struct in_conninfo starts with 16 bits 490 * of foreign port, then 16 bits of local port then followed by 128 491 * bits of foreign address. In case of IPv4 address, the first 3 492 * 32-bit words of the address always are zeroes. 493 */ 494 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 495 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 496 497 sch = &V_tcp_syncache.hashbase[hash]; 498 *schp = sch; 499 SCH_LOCK(sch); 500 501 /* Circle through bucket row to find matching entry. */ 502 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 503 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 504 sizeof(struct in_endpoints)) == 0) 505 break; 506 507 return (sc); /* Always returns with locked sch. */ 508 } 509 510 /* 511 * This function is called when we get a RST for a 512 * non-existent connection, so that we can see if the 513 * connection is in the syn cache. If it is, zap it. 514 */ 515 void 516 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 517 { 518 struct syncache *sc; 519 struct syncache_head *sch; 520 char *s = NULL; 521 522 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 523 SCH_LOCK_ASSERT(sch); 524 525 /* 526 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 527 * See RFC 793 page 65, section SEGMENT ARRIVES. 528 */ 529 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 530 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 531 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 532 "FIN flag set, segment ignored\n", s, __func__); 533 TCPSTAT_INC(tcps_badrst); 534 goto done; 535 } 536 537 /* 538 * No corresponding connection was found in syncache. 539 * If syncookies are enabled and possibly exclusively 540 * used, or we are under memory pressure, a valid RST 541 * may not find a syncache entry. In that case we're 542 * done and no SYN|ACK retransmissions will happen. 543 * Otherwise the RST was misdirected or spoofed. 544 */ 545 if (sc == NULL) { 546 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 547 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 548 "syncache entry (possibly syncookie only), " 549 "segment ignored\n", s, __func__); 550 TCPSTAT_INC(tcps_badrst); 551 goto done; 552 } 553 554 /* 555 * If the RST bit is set, check the sequence number to see 556 * if this is a valid reset segment. 557 * RFC 793 page 37: 558 * In all states except SYN-SENT, all reset (RST) segments 559 * are validated by checking their SEQ-fields. A reset is 560 * valid if its sequence number is in the window. 561 * 562 * The sequence number in the reset segment is normally an 563 * echo of our outgoing acknowlegement numbers, but some hosts 564 * send a reset with the sequence number at the rightmost edge 565 * of our receive window, and we have to handle this case. 566 */ 567 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 568 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 569 syncache_drop(sc, sch); 570 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 571 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 572 "connection attempt aborted by remote endpoint\n", 573 s, __func__); 574 TCPSTAT_INC(tcps_sc_reset); 575 } else { 576 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 577 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 578 "IRS %u (+WND %u), segment ignored\n", 579 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 580 TCPSTAT_INC(tcps_badrst); 581 } 582 583 done: 584 if (s != NULL) 585 free(s, M_TCPLOG); 586 SCH_UNLOCK(sch); 587 } 588 589 void 590 syncache_badack(struct in_conninfo *inc) 591 { 592 struct syncache *sc; 593 struct syncache_head *sch; 594 595 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 596 SCH_LOCK_ASSERT(sch); 597 if (sc != NULL) { 598 syncache_drop(sc, sch); 599 TCPSTAT_INC(tcps_sc_badack); 600 } 601 SCH_UNLOCK(sch); 602 } 603 604 void 605 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq) 606 { 607 struct syncache *sc; 608 struct syncache_head *sch; 609 610 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 611 SCH_LOCK_ASSERT(sch); 612 if (sc == NULL) 613 goto done; 614 615 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 616 if (ntohl(th_seq) != sc->sc_iss) 617 goto done; 618 619 /* 620 * If we've rertransmitted 3 times and this is our second error, 621 * we remove the entry. Otherwise, we allow it to continue on. 622 * This prevents us from incorrectly nuking an entry during a 623 * spurious network outage. 624 * 625 * See tcp_notify(). 626 */ 627 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 628 sc->sc_flags |= SCF_UNREACH; 629 goto done; 630 } 631 syncache_drop(sc, sch); 632 TCPSTAT_INC(tcps_sc_unreach); 633 done: 634 SCH_UNLOCK(sch); 635 } 636 637 /* 638 * Build a new TCP socket structure from a syncache entry. 639 * 640 * On success return the newly created socket with its underlying inp locked. 641 */ 642 static struct socket * 643 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 644 { 645 struct tcp_function_block *blk; 646 struct inpcb *inp = NULL; 647 struct socket *so; 648 struct tcpcb *tp; 649 int error; 650 char *s; 651 652 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 653 654 /* 655 * Ok, create the full blown connection, and set things up 656 * as they would have been set up if we had created the 657 * connection when the SYN arrived. If we can't create 658 * the connection, abort it. 659 */ 660 so = sonewconn(lso, 0); 661 if (so == NULL) { 662 /* 663 * Drop the connection; we will either send a RST or 664 * have the peer retransmit its SYN again after its 665 * RTO and try again. 666 */ 667 TCPSTAT_INC(tcps_listendrop); 668 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 669 log(LOG_DEBUG, "%s; %s: Socket create failed " 670 "due to limits or memory shortage\n", 671 s, __func__); 672 free(s, M_TCPLOG); 673 } 674 goto abort2; 675 } 676 #ifdef MAC 677 mac_socketpeer_set_from_mbuf(m, so); 678 #endif 679 680 inp = sotoinpcb(so); 681 inp->inp_inc.inc_fibnum = so->so_fibnum; 682 INP_WLOCK(inp); 683 /* 684 * Exclusive pcbinfo lock is not required in syncache socket case even 685 * if two inpcb locks can be acquired simultaneously: 686 * - the inpcb in LISTEN state, 687 * - the newly created inp. 688 * 689 * In this case, an inp cannot be at same time in LISTEN state and 690 * just created by an accept() call. 691 */ 692 INP_HASH_WLOCK(&V_tcbinfo); 693 694 /* Insert new socket into PCB hash list. */ 695 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 696 #ifdef INET6 697 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 698 inp->inp_vflag &= ~INP_IPV4; 699 inp->inp_vflag |= INP_IPV6; 700 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 701 } else { 702 inp->inp_vflag &= ~INP_IPV6; 703 inp->inp_vflag |= INP_IPV4; 704 #endif 705 inp->inp_laddr = sc->sc_inc.inc_laddr; 706 #ifdef INET6 707 } 708 #endif 709 710 /* 711 * If there's an mbuf and it has a flowid, then let's initialise the 712 * inp with that particular flowid. 713 */ 714 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 715 inp->inp_flowid = m->m_pkthdr.flowid; 716 inp->inp_flowtype = M_HASHTYPE_GET(m); 717 } 718 719 /* 720 * Install in the reservation hash table for now, but don't yet 721 * install a connection group since the full 4-tuple isn't yet 722 * configured. 723 */ 724 inp->inp_lport = sc->sc_inc.inc_lport; 725 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 726 /* 727 * Undo the assignments above if we failed to 728 * put the PCB on the hash lists. 729 */ 730 #ifdef INET6 731 if (sc->sc_inc.inc_flags & INC_ISIPV6) 732 inp->in6p_laddr = in6addr_any; 733 else 734 #endif 735 inp->inp_laddr.s_addr = INADDR_ANY; 736 inp->inp_lport = 0; 737 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 738 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 739 "with error %i\n", 740 s, __func__, error); 741 free(s, M_TCPLOG); 742 } 743 INP_HASH_WUNLOCK(&V_tcbinfo); 744 goto abort; 745 } 746 #ifdef INET6 747 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 748 struct inpcb *oinp = sotoinpcb(lso); 749 struct in6_addr laddr6; 750 struct sockaddr_in6 sin6; 751 /* 752 * Inherit socket options from the listening socket. 753 * Note that in6p_inputopts are not (and should not be) 754 * copied, since it stores previously received options and is 755 * used to detect if each new option is different than the 756 * previous one and hence should be passed to a user. 757 * If we copied in6p_inputopts, a user would not be able to 758 * receive options just after calling the accept system call. 759 */ 760 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 761 if (oinp->in6p_outputopts) 762 inp->in6p_outputopts = 763 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 764 765 sin6.sin6_family = AF_INET6; 766 sin6.sin6_len = sizeof(sin6); 767 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 768 sin6.sin6_port = sc->sc_inc.inc_fport; 769 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 770 laddr6 = inp->in6p_laddr; 771 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 772 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 773 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 774 thread0.td_ucred, m)) != 0) { 775 inp->in6p_laddr = laddr6; 776 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 777 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 778 "with error %i\n", 779 s, __func__, error); 780 free(s, M_TCPLOG); 781 } 782 INP_HASH_WUNLOCK(&V_tcbinfo); 783 goto abort; 784 } 785 /* Override flowlabel from in6_pcbconnect. */ 786 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 787 inp->inp_flow |= sc->sc_flowlabel; 788 } 789 #endif /* INET6 */ 790 #if defined(INET) && defined(INET6) 791 else 792 #endif 793 #ifdef INET 794 { 795 struct in_addr laddr; 796 struct sockaddr_in sin; 797 798 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 799 800 if (inp->inp_options == NULL) { 801 inp->inp_options = sc->sc_ipopts; 802 sc->sc_ipopts = NULL; 803 } 804 805 sin.sin_family = AF_INET; 806 sin.sin_len = sizeof(sin); 807 sin.sin_addr = sc->sc_inc.inc_faddr; 808 sin.sin_port = sc->sc_inc.inc_fport; 809 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 810 laddr = inp->inp_laddr; 811 if (inp->inp_laddr.s_addr == INADDR_ANY) 812 inp->inp_laddr = sc->sc_inc.inc_laddr; 813 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 814 thread0.td_ucred, m)) != 0) { 815 inp->inp_laddr = laddr; 816 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 817 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 818 "with error %i\n", 819 s, __func__, error); 820 free(s, M_TCPLOG); 821 } 822 INP_HASH_WUNLOCK(&V_tcbinfo); 823 goto abort; 824 } 825 } 826 #endif /* INET */ 827 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 828 /* Copy old policy into new socket's. */ 829 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 830 printf("syncache_socket: could not copy policy\n"); 831 #endif 832 INP_HASH_WUNLOCK(&V_tcbinfo); 833 tp = intotcpcb(inp); 834 tcp_state_change(tp, TCPS_SYN_RECEIVED); 835 tp->iss = sc->sc_iss; 836 tp->irs = sc->sc_irs; 837 tcp_rcvseqinit(tp); 838 tcp_sendseqinit(tp); 839 blk = sototcpcb(lso)->t_fb; 840 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 841 /* 842 * Our parents t_fb was not the default, 843 * we need to release our ref on tp->t_fb and 844 * pickup one on the new entry. 845 */ 846 struct tcp_function_block *rblk; 847 848 rblk = find_and_ref_tcp_fb(blk); 849 KASSERT(rblk != NULL, 850 ("cannot find blk %p out of syncache?", blk)); 851 if (tp->t_fb->tfb_tcp_fb_fini) 852 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 853 refcount_release(&tp->t_fb->tfb_refcnt); 854 tp->t_fb = rblk; 855 if (tp->t_fb->tfb_tcp_fb_init) { 856 (*tp->t_fb->tfb_tcp_fb_init)(tp); 857 } 858 } 859 tp->snd_wl1 = sc->sc_irs; 860 tp->snd_max = tp->iss + 1; 861 tp->snd_nxt = tp->iss + 1; 862 tp->rcv_up = sc->sc_irs + 1; 863 tp->rcv_wnd = sc->sc_wnd; 864 tp->rcv_adv += tp->rcv_wnd; 865 tp->last_ack_sent = tp->rcv_nxt; 866 867 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 868 if (sc->sc_flags & SCF_NOOPT) 869 tp->t_flags |= TF_NOOPT; 870 else { 871 if (sc->sc_flags & SCF_WINSCALE) { 872 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 873 tp->snd_scale = sc->sc_requested_s_scale; 874 tp->request_r_scale = sc->sc_requested_r_scale; 875 } 876 if (sc->sc_flags & SCF_TIMESTAMP) { 877 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 878 tp->ts_recent = sc->sc_tsreflect; 879 tp->ts_recent_age = tcp_ts_getticks(); 880 tp->ts_offset = sc->sc_tsoff; 881 } 882 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 883 if (sc->sc_flags & SCF_SIGNATURE) 884 tp->t_flags |= TF_SIGNATURE; 885 #endif 886 if (sc->sc_flags & SCF_SACK) 887 tp->t_flags |= TF_SACK_PERMIT; 888 } 889 890 if (sc->sc_flags & SCF_ECN) 891 tp->t_flags |= TF_ECN_PERMIT; 892 893 /* 894 * Set up MSS and get cached values from tcp_hostcache. 895 * This might overwrite some of the defaults we just set. 896 */ 897 tcp_mss(tp, sc->sc_peer_mss); 898 899 /* 900 * If the SYN,ACK was retransmitted, indicate that CWND to be 901 * limited to one segment in cc_conn_init(). 902 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 903 */ 904 if (sc->sc_rxmits > 1) 905 tp->snd_cwnd = 1; 906 907 #ifdef TCP_OFFLOAD 908 /* 909 * Allow a TOE driver to install its hooks. Note that we hold the 910 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 911 * new connection before the TOE driver has done its thing. 912 */ 913 if (ADDED_BY_TOE(sc)) { 914 struct toedev *tod = sc->sc_tod; 915 916 tod->tod_offload_socket(tod, sc->sc_todctx, so); 917 } 918 #endif 919 /* 920 * Copy and activate timers. 921 */ 922 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 923 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 924 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 925 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 926 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 927 928 TCPSTAT_INC(tcps_accepts); 929 return (so); 930 931 abort: 932 INP_WUNLOCK(inp); 933 abort2: 934 if (so != NULL) 935 soabort(so); 936 return (NULL); 937 } 938 939 /* 940 * This function gets called when we receive an ACK for a 941 * socket in the LISTEN state. We look up the connection 942 * in the syncache, and if its there, we pull it out of 943 * the cache and turn it into a full-blown connection in 944 * the SYN-RECEIVED state. 945 * 946 * On syncache_socket() success the newly created socket 947 * has its underlying inp locked. 948 */ 949 int 950 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 951 struct socket **lsop, struct mbuf *m) 952 { 953 struct syncache *sc; 954 struct syncache_head *sch; 955 struct syncache scs; 956 char *s; 957 958 /* 959 * Global TCP locks are held because we manipulate the PCB lists 960 * and create a new socket. 961 */ 962 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 963 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 964 ("%s: can handle only ACK", __func__)); 965 966 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 967 SCH_LOCK_ASSERT(sch); 968 969 #ifdef INVARIANTS 970 /* 971 * Test code for syncookies comparing the syncache stored 972 * values with the reconstructed values from the cookie. 973 */ 974 if (sc != NULL) 975 syncookie_cmp(inc, sch, sc, th, to, *lsop); 976 #endif 977 978 if (sc == NULL) { 979 /* 980 * There is no syncache entry, so see if this ACK is 981 * a returning syncookie. To do this, first: 982 * A. Check if syncookies are used in case of syncache 983 * overflows 984 * B. See if this socket has had a syncache entry dropped in 985 * the recent past. We don't want to accept a bogus 986 * syncookie if we've never received a SYN or accept it 987 * twice. 988 * C. check that the syncookie is valid. If it is, then 989 * cobble up a fake syncache entry, and return. 990 */ 991 if (!V_tcp_syncookies) { 992 SCH_UNLOCK(sch); 993 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 994 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 995 "segment rejected (syncookies disabled)\n", 996 s, __func__); 997 goto failed; 998 } 999 if (!V_tcp_syncookiesonly && 1000 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1001 SCH_UNLOCK(sch); 1002 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1003 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1004 "segment rejected (no syncache entry)\n", 1005 s, __func__); 1006 goto failed; 1007 } 1008 bzero(&scs, sizeof(scs)); 1009 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1010 SCH_UNLOCK(sch); 1011 if (sc == NULL) { 1012 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1013 log(LOG_DEBUG, "%s; %s: Segment failed " 1014 "SYNCOOKIE authentication, segment rejected " 1015 "(probably spoofed)\n", s, __func__); 1016 goto failed; 1017 } 1018 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1019 /* If received ACK has MD5 signature, check it. */ 1020 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1021 (!TCPMD5_ENABLED() || 1022 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1023 /* Drop the ACK. */ 1024 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1025 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1026 "MD5 signature doesn't match.\n", 1027 s, __func__); 1028 free(s, M_TCPLOG); 1029 } 1030 TCPSTAT_INC(tcps_sig_err_sigopt); 1031 return (-1); /* Do not send RST */ 1032 } 1033 #endif /* TCP_SIGNATURE */ 1034 } else { 1035 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1036 /* 1037 * If listening socket requested TCP digests, check that 1038 * received ACK has signature and it is correct. 1039 * If not, drop the ACK and leave sc entry in th cache, 1040 * because SYN was received with correct signature. 1041 */ 1042 if (sc->sc_flags & SCF_SIGNATURE) { 1043 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1044 /* No signature */ 1045 TCPSTAT_INC(tcps_sig_err_nosigopt); 1046 SCH_UNLOCK(sch); 1047 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1048 log(LOG_DEBUG, "%s; %s: Segment " 1049 "rejected, MD5 signature wasn't " 1050 "provided.\n", s, __func__); 1051 free(s, M_TCPLOG); 1052 } 1053 return (-1); /* Do not send RST */ 1054 } 1055 if (!TCPMD5_ENABLED() || 1056 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1057 /* Doesn't match or no SA */ 1058 SCH_UNLOCK(sch); 1059 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1060 log(LOG_DEBUG, "%s; %s: Segment " 1061 "rejected, MD5 signature doesn't " 1062 "match.\n", s, __func__); 1063 free(s, M_TCPLOG); 1064 } 1065 return (-1); /* Do not send RST */ 1066 } 1067 } 1068 #endif /* TCP_SIGNATURE */ 1069 /* 1070 * Pull out the entry to unlock the bucket row. 1071 * 1072 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1073 * tcp_state_change(). The tcpcb is not existent at this 1074 * moment. A new one will be allocated via syncache_socket-> 1075 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1076 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1077 */ 1078 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1079 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1080 sch->sch_length--; 1081 #ifdef TCP_OFFLOAD 1082 if (ADDED_BY_TOE(sc)) { 1083 struct toedev *tod = sc->sc_tod; 1084 1085 tod->tod_syncache_removed(tod, sc->sc_todctx); 1086 } 1087 #endif 1088 SCH_UNLOCK(sch); 1089 } 1090 1091 /* 1092 * Segment validation: 1093 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1094 */ 1095 if (th->th_ack != sc->sc_iss + 1) { 1096 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1097 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1098 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1099 goto failed; 1100 } 1101 1102 /* 1103 * The SEQ must fall in the window starting at the received 1104 * initial receive sequence number + 1 (the SYN). 1105 */ 1106 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1107 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1108 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1109 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1110 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1111 goto failed; 1112 } 1113 1114 /* 1115 * If timestamps were not negotiated during SYN/ACK they 1116 * must not appear on any segment during this session. 1117 */ 1118 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1119 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1120 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1121 "segment rejected\n", s, __func__); 1122 goto failed; 1123 } 1124 1125 /* 1126 * If timestamps were negotiated during SYN/ACK they should 1127 * appear on every segment during this session. 1128 * XXXAO: This is only informal as there have been unverified 1129 * reports of non-compliants stacks. 1130 */ 1131 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1132 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1133 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1134 "no action\n", s, __func__); 1135 free(s, M_TCPLOG); 1136 s = NULL; 1137 } 1138 } 1139 1140 /* 1141 * If timestamps were negotiated, the reflected timestamp 1142 * must be equal to what we actually sent in the SYN|ACK 1143 * except in the case of 0. Some boxes are known for sending 1144 * broken timestamp replies during the 3whs (and potentially 1145 * during the connection also). 1146 * 1147 * Accept the final ACK of 3whs with reflected timestamp of 0 1148 * instead of sending a RST and deleting the syncache entry. 1149 */ 1150 if ((to->to_flags & TOF_TS) && to->to_tsecr && 1151 to->to_tsecr != sc->sc_ts) { 1152 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1153 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1154 "segment rejected\n", 1155 s, __func__, to->to_tsecr, sc->sc_ts); 1156 goto failed; 1157 } 1158 1159 *lsop = syncache_socket(sc, *lsop, m); 1160 1161 if (*lsop == NULL) 1162 TCPSTAT_INC(tcps_sc_aborted); 1163 else 1164 TCPSTAT_INC(tcps_sc_completed); 1165 1166 /* how do we find the inp for the new socket? */ 1167 if (sc != &scs) 1168 syncache_free(sc); 1169 return (1); 1170 failed: 1171 if (sc != NULL && sc != &scs) 1172 syncache_free(sc); 1173 if (s != NULL) 1174 free(s, M_TCPLOG); 1175 *lsop = NULL; 1176 return (0); 1177 } 1178 1179 static void 1180 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1181 uint64_t response_cookie) 1182 { 1183 struct inpcb *inp; 1184 struct tcpcb *tp; 1185 unsigned int *pending_counter; 1186 1187 /* 1188 * Global TCP locks are held because we manipulate the PCB lists 1189 * and create a new socket. 1190 */ 1191 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1192 1193 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1194 *lsop = syncache_socket(sc, *lsop, m); 1195 if (*lsop == NULL) { 1196 TCPSTAT_INC(tcps_sc_aborted); 1197 atomic_subtract_int(pending_counter, 1); 1198 } else { 1199 soisconnected(*lsop); 1200 inp = sotoinpcb(*lsop); 1201 tp = intotcpcb(inp); 1202 tp->t_flags |= TF_FASTOPEN; 1203 tp->t_tfo_cookie.server = response_cookie; 1204 tp->snd_max = tp->iss; 1205 tp->snd_nxt = tp->iss; 1206 tp->t_tfo_pending = pending_counter; 1207 TCPSTAT_INC(tcps_sc_completed); 1208 } 1209 } 1210 1211 /* 1212 * Given a LISTEN socket and an inbound SYN request, add 1213 * this to the syn cache, and send back a segment: 1214 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1215 * to the source. 1216 * 1217 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1218 * Doing so would require that we hold onto the data and deliver it 1219 * to the application. However, if we are the target of a SYN-flood 1220 * DoS attack, an attacker could send data which would eventually 1221 * consume all available buffer space if it were ACKed. By not ACKing 1222 * the data, we avoid this DoS scenario. 1223 * 1224 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1225 * cookie is processed and a new socket is created. In this case, any data 1226 * accompanying the SYN will be queued to the socket by tcp_input() and will 1227 * be ACKed either when the application sends response data or the delayed 1228 * ACK timer expires, whichever comes first. 1229 */ 1230 int 1231 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1232 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1233 void *todctx) 1234 { 1235 struct tcpcb *tp; 1236 struct socket *so; 1237 struct syncache *sc = NULL; 1238 struct syncache_head *sch; 1239 struct mbuf *ipopts = NULL; 1240 u_int ltflags; 1241 int win, ip_ttl, ip_tos; 1242 char *s; 1243 int rv = 0; 1244 #ifdef INET6 1245 int autoflowlabel = 0; 1246 #endif 1247 #ifdef MAC 1248 struct label *maclabel; 1249 #endif 1250 struct syncache scs; 1251 struct ucred *cred; 1252 uint64_t tfo_response_cookie; 1253 unsigned int *tfo_pending = NULL; 1254 int tfo_cookie_valid = 0; 1255 int tfo_response_cookie_valid = 0; 1256 1257 INP_WLOCK_ASSERT(inp); /* listen socket */ 1258 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1259 ("%s: unexpected tcp flags", __func__)); 1260 1261 /* 1262 * Combine all so/tp operations very early to drop the INP lock as 1263 * soon as possible. 1264 */ 1265 so = *lsop; 1266 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1267 tp = sototcpcb(so); 1268 cred = crhold(so->so_cred); 1269 1270 #ifdef INET6 1271 if ((inc->inc_flags & INC_ISIPV6) && 1272 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1273 autoflowlabel = 1; 1274 #endif 1275 ip_ttl = inp->inp_ip_ttl; 1276 ip_tos = inp->inp_ip_tos; 1277 win = so->sol_sbrcv_hiwat; 1278 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1279 1280 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1281 (tp->t_tfo_pending != NULL) && 1282 (to->to_flags & TOF_FASTOPEN)) { 1283 /* 1284 * Limit the number of pending TFO connections to 1285 * approximately half of the queue limit. This prevents TFO 1286 * SYN floods from starving the service by filling the 1287 * listen queue with bogus TFO connections. 1288 */ 1289 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1290 (so->sol_qlimit / 2)) { 1291 int result; 1292 1293 result = tcp_fastopen_check_cookie(inc, 1294 to->to_tfo_cookie, to->to_tfo_len, 1295 &tfo_response_cookie); 1296 tfo_cookie_valid = (result > 0); 1297 tfo_response_cookie_valid = (result >= 0); 1298 } 1299 1300 /* 1301 * Remember the TFO pending counter as it will have to be 1302 * decremented below if we don't make it to syncache_tfo_expand(). 1303 */ 1304 tfo_pending = tp->t_tfo_pending; 1305 } 1306 1307 /* By the time we drop the lock these should no longer be used. */ 1308 so = NULL; 1309 tp = NULL; 1310 1311 #ifdef MAC 1312 if (mac_syncache_init(&maclabel) != 0) { 1313 INP_WUNLOCK(inp); 1314 goto done; 1315 } else 1316 mac_syncache_create(maclabel, inp); 1317 #endif 1318 if (!tfo_cookie_valid) 1319 INP_WUNLOCK(inp); 1320 1321 /* 1322 * Remember the IP options, if any. 1323 */ 1324 #ifdef INET6 1325 if (!(inc->inc_flags & INC_ISIPV6)) 1326 #endif 1327 #ifdef INET 1328 ipopts = (m) ? ip_srcroute(m) : NULL; 1329 #else 1330 ipopts = NULL; 1331 #endif 1332 1333 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1334 /* 1335 * If listening socket requested TCP digests, check that received 1336 * SYN has signature and it is correct. If signature doesn't match 1337 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1338 */ 1339 if (ltflags & TF_SIGNATURE) { 1340 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1341 TCPSTAT_INC(tcps_sig_err_nosigopt); 1342 goto done; 1343 } 1344 if (!TCPMD5_ENABLED() || 1345 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1346 goto done; 1347 } 1348 #endif /* TCP_SIGNATURE */ 1349 /* 1350 * See if we already have an entry for this connection. 1351 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1352 * 1353 * XXX: should the syncache be re-initialized with the contents 1354 * of the new SYN here (which may have different options?) 1355 * 1356 * XXX: We do not check the sequence number to see if this is a 1357 * real retransmit or a new connection attempt. The question is 1358 * how to handle such a case; either ignore it as spoofed, or 1359 * drop the current entry and create a new one? 1360 */ 1361 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1362 SCH_LOCK_ASSERT(sch); 1363 if (sc != NULL) { 1364 if (tfo_cookie_valid) 1365 INP_WUNLOCK(inp); 1366 TCPSTAT_INC(tcps_sc_dupsyn); 1367 if (ipopts) { 1368 /* 1369 * If we were remembering a previous source route, 1370 * forget it and use the new one we've been given. 1371 */ 1372 if (sc->sc_ipopts) 1373 (void) m_free(sc->sc_ipopts); 1374 sc->sc_ipopts = ipopts; 1375 } 1376 /* 1377 * Update timestamp if present. 1378 */ 1379 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1380 sc->sc_tsreflect = to->to_tsval; 1381 else 1382 sc->sc_flags &= ~SCF_TIMESTAMP; 1383 #ifdef MAC 1384 /* 1385 * Since we have already unconditionally allocated label 1386 * storage, free it up. The syncache entry will already 1387 * have an initialized label we can use. 1388 */ 1389 mac_syncache_destroy(&maclabel); 1390 #endif 1391 /* Retransmit SYN|ACK and reset retransmit count. */ 1392 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1393 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1394 "resetting timer and retransmitting SYN|ACK\n", 1395 s, __func__); 1396 free(s, M_TCPLOG); 1397 } 1398 if (syncache_respond(sc, sch, 1, m) == 0) { 1399 sc->sc_rxmits = 0; 1400 syncache_timeout(sc, sch, 1); 1401 TCPSTAT_INC(tcps_sndacks); 1402 TCPSTAT_INC(tcps_sndtotal); 1403 } 1404 SCH_UNLOCK(sch); 1405 goto done; 1406 } 1407 1408 if (tfo_cookie_valid) { 1409 bzero(&scs, sizeof(scs)); 1410 sc = &scs; 1411 goto skip_alloc; 1412 } 1413 1414 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1415 if (sc == NULL) { 1416 /* 1417 * The zone allocator couldn't provide more entries. 1418 * Treat this as if the cache was full; drop the oldest 1419 * entry and insert the new one. 1420 */ 1421 TCPSTAT_INC(tcps_sc_zonefail); 1422 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1423 sch->sch_last_overflow = time_uptime; 1424 syncache_drop(sc, sch); 1425 } 1426 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1427 if (sc == NULL) { 1428 if (V_tcp_syncookies) { 1429 bzero(&scs, sizeof(scs)); 1430 sc = &scs; 1431 } else { 1432 SCH_UNLOCK(sch); 1433 if (ipopts) 1434 (void) m_free(ipopts); 1435 goto done; 1436 } 1437 } 1438 } 1439 1440 skip_alloc: 1441 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1442 sc->sc_tfo_cookie = &tfo_response_cookie; 1443 1444 /* 1445 * Fill in the syncache values. 1446 */ 1447 #ifdef MAC 1448 sc->sc_label = maclabel; 1449 #endif 1450 sc->sc_cred = cred; 1451 cred = NULL; 1452 sc->sc_ipopts = ipopts; 1453 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1454 #ifdef INET6 1455 if (!(inc->inc_flags & INC_ISIPV6)) 1456 #endif 1457 { 1458 sc->sc_ip_tos = ip_tos; 1459 sc->sc_ip_ttl = ip_ttl; 1460 } 1461 #ifdef TCP_OFFLOAD 1462 sc->sc_tod = tod; 1463 sc->sc_todctx = todctx; 1464 #endif 1465 sc->sc_irs = th->th_seq; 1466 sc->sc_iss = arc4random(); 1467 sc->sc_flags = 0; 1468 sc->sc_flowlabel = 0; 1469 1470 /* 1471 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1472 * win was derived from socket earlier in the function. 1473 */ 1474 win = imax(win, 0); 1475 win = imin(win, TCP_MAXWIN); 1476 sc->sc_wnd = win; 1477 1478 if (V_tcp_do_rfc1323) { 1479 /* 1480 * A timestamp received in a SYN makes 1481 * it ok to send timestamp requests and replies. 1482 */ 1483 if (to->to_flags & TOF_TS) { 1484 sc->sc_tsreflect = to->to_tsval; 1485 sc->sc_ts = tcp_ts_getticks(); 1486 sc->sc_flags |= SCF_TIMESTAMP; 1487 } 1488 if (to->to_flags & TOF_SCALE) { 1489 int wscale = 0; 1490 1491 /* 1492 * Pick the smallest possible scaling factor that 1493 * will still allow us to scale up to sb_max, aka 1494 * kern.ipc.maxsockbuf. 1495 * 1496 * We do this because there are broken firewalls that 1497 * will corrupt the window scale option, leading to 1498 * the other endpoint believing that our advertised 1499 * window is unscaled. At scale factors larger than 1500 * 5 the unscaled window will drop below 1500 bytes, 1501 * leading to serious problems when traversing these 1502 * broken firewalls. 1503 * 1504 * With the default maxsockbuf of 256K, a scale factor 1505 * of 3 will be chosen by this algorithm. Those who 1506 * choose a larger maxsockbuf should watch out 1507 * for the compatibility problems mentioned above. 1508 * 1509 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1510 * or <SYN,ACK>) segment itself is never scaled. 1511 */ 1512 while (wscale < TCP_MAX_WINSHIFT && 1513 (TCP_MAXWIN << wscale) < sb_max) 1514 wscale++; 1515 sc->sc_requested_r_scale = wscale; 1516 sc->sc_requested_s_scale = to->to_wscale; 1517 sc->sc_flags |= SCF_WINSCALE; 1518 } 1519 } 1520 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1521 /* 1522 * If listening socket requested TCP digests, flag this in the 1523 * syncache so that syncache_respond() will do the right thing 1524 * with the SYN+ACK. 1525 */ 1526 if (ltflags & TF_SIGNATURE) 1527 sc->sc_flags |= SCF_SIGNATURE; 1528 #endif /* TCP_SIGNATURE */ 1529 if (to->to_flags & TOF_SACKPERM) 1530 sc->sc_flags |= SCF_SACK; 1531 if (to->to_flags & TOF_MSS) 1532 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1533 if (ltflags & TF_NOOPT) 1534 sc->sc_flags |= SCF_NOOPT; 1535 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1536 sc->sc_flags |= SCF_ECN; 1537 1538 if (V_tcp_syncookies) 1539 sc->sc_iss = syncookie_generate(sch, sc); 1540 #ifdef INET6 1541 if (autoflowlabel) { 1542 if (V_tcp_syncookies) 1543 sc->sc_flowlabel = sc->sc_iss; 1544 else 1545 sc->sc_flowlabel = ip6_randomflowlabel(); 1546 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1547 } 1548 #endif 1549 SCH_UNLOCK(sch); 1550 1551 if (tfo_cookie_valid) { 1552 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1553 /* INP_WUNLOCK(inp) will be performed by the caller */ 1554 rv = 1; 1555 goto tfo_expanded; 1556 } 1557 1558 /* 1559 * Do a standard 3-way handshake. 1560 */ 1561 if (syncache_respond(sc, sch, 0, m) == 0) { 1562 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1563 syncache_free(sc); 1564 else if (sc != &scs) 1565 syncache_insert(sc, sch); /* locks and unlocks sch */ 1566 TCPSTAT_INC(tcps_sndacks); 1567 TCPSTAT_INC(tcps_sndtotal); 1568 } else { 1569 if (sc != &scs) 1570 syncache_free(sc); 1571 TCPSTAT_INC(tcps_sc_dropped); 1572 } 1573 1574 done: 1575 if (m) { 1576 *lsop = NULL; 1577 m_freem(m); 1578 } 1579 /* 1580 * If tfo_pending is not NULL here, then a TFO SYN that did not 1581 * result in a new socket was processed and the associated pending 1582 * counter has not yet been decremented. All such TFO processing paths 1583 * transit this point. 1584 */ 1585 if (tfo_pending != NULL) 1586 tcp_fastopen_decrement_counter(tfo_pending); 1587 1588 tfo_expanded: 1589 if (cred != NULL) 1590 crfree(cred); 1591 #ifdef MAC 1592 if (sc == &scs) 1593 mac_syncache_destroy(&maclabel); 1594 #endif 1595 return (rv); 1596 } 1597 1598 /* 1599 * Send SYN|ACK to the peer. Either in response to the peer's SYN, 1600 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1601 */ 1602 static int 1603 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked, 1604 const struct mbuf *m0) 1605 { 1606 struct ip *ip = NULL; 1607 struct mbuf *m; 1608 struct tcphdr *th = NULL; 1609 int optlen, error = 0; /* Make compiler happy */ 1610 u_int16_t hlen, tlen, mssopt; 1611 struct tcpopt to; 1612 #ifdef INET6 1613 struct ip6_hdr *ip6 = NULL; 1614 #endif 1615 hlen = 1616 #ifdef INET6 1617 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1618 #endif 1619 sizeof(struct ip); 1620 tlen = hlen + sizeof(struct tcphdr); 1621 1622 /* Determine MSS we advertize to other end of connection. */ 1623 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); 1624 1625 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1626 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1627 ("syncache: mbuf too small")); 1628 1629 /* Create the IP+TCP header from scratch. */ 1630 m = m_gethdr(M_NOWAIT, MT_DATA); 1631 if (m == NULL) 1632 return (ENOBUFS); 1633 #ifdef MAC 1634 mac_syncache_create_mbuf(sc->sc_label, m); 1635 #endif 1636 m->m_data += max_linkhdr; 1637 m->m_len = tlen; 1638 m->m_pkthdr.len = tlen; 1639 m->m_pkthdr.rcvif = NULL; 1640 1641 #ifdef INET6 1642 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1643 ip6 = mtod(m, struct ip6_hdr *); 1644 ip6->ip6_vfc = IPV6_VERSION; 1645 ip6->ip6_nxt = IPPROTO_TCP; 1646 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1647 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1648 ip6->ip6_plen = htons(tlen - hlen); 1649 /* ip6_hlim is set after checksum */ 1650 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1651 ip6->ip6_flow |= sc->sc_flowlabel; 1652 1653 th = (struct tcphdr *)(ip6 + 1); 1654 } 1655 #endif 1656 #if defined(INET6) && defined(INET) 1657 else 1658 #endif 1659 #ifdef INET 1660 { 1661 ip = mtod(m, struct ip *); 1662 ip->ip_v = IPVERSION; 1663 ip->ip_hl = sizeof(struct ip) >> 2; 1664 ip->ip_len = htons(tlen); 1665 ip->ip_id = 0; 1666 ip->ip_off = 0; 1667 ip->ip_sum = 0; 1668 ip->ip_p = IPPROTO_TCP; 1669 ip->ip_src = sc->sc_inc.inc_laddr; 1670 ip->ip_dst = sc->sc_inc.inc_faddr; 1671 ip->ip_ttl = sc->sc_ip_ttl; 1672 ip->ip_tos = sc->sc_ip_tos; 1673 1674 /* 1675 * See if we should do MTU discovery. Route lookups are 1676 * expensive, so we will only unset the DF bit if: 1677 * 1678 * 1) path_mtu_discovery is disabled 1679 * 2) the SCF_UNREACH flag has been set 1680 */ 1681 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1682 ip->ip_off |= htons(IP_DF); 1683 1684 th = (struct tcphdr *)(ip + 1); 1685 } 1686 #endif /* INET */ 1687 th->th_sport = sc->sc_inc.inc_lport; 1688 th->th_dport = sc->sc_inc.inc_fport; 1689 1690 th->th_seq = htonl(sc->sc_iss); 1691 th->th_ack = htonl(sc->sc_irs + 1); 1692 th->th_off = sizeof(struct tcphdr) >> 2; 1693 th->th_x2 = 0; 1694 th->th_flags = TH_SYN|TH_ACK; 1695 th->th_win = htons(sc->sc_wnd); 1696 th->th_urp = 0; 1697 1698 if (sc->sc_flags & SCF_ECN) { 1699 th->th_flags |= TH_ECE; 1700 TCPSTAT_INC(tcps_ecn_shs); 1701 } 1702 1703 /* Tack on the TCP options. */ 1704 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1705 to.to_flags = 0; 1706 1707 to.to_mss = mssopt; 1708 to.to_flags = TOF_MSS; 1709 if (sc->sc_flags & SCF_WINSCALE) { 1710 to.to_wscale = sc->sc_requested_r_scale; 1711 to.to_flags |= TOF_SCALE; 1712 } 1713 if (sc->sc_flags & SCF_TIMESTAMP) { 1714 /* Virgin timestamp or TCP cookie enhanced one. */ 1715 to.to_tsval = sc->sc_ts; 1716 to.to_tsecr = sc->sc_tsreflect; 1717 to.to_flags |= TOF_TS; 1718 } 1719 if (sc->sc_flags & SCF_SACK) 1720 to.to_flags |= TOF_SACKPERM; 1721 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1722 if (sc->sc_flags & SCF_SIGNATURE) 1723 to.to_flags |= TOF_SIGNATURE; 1724 #endif 1725 if (sc->sc_tfo_cookie) { 1726 to.to_flags |= TOF_FASTOPEN; 1727 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1728 to.to_tfo_cookie = sc->sc_tfo_cookie; 1729 /* don't send cookie again when retransmitting response */ 1730 sc->sc_tfo_cookie = NULL; 1731 } 1732 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1733 1734 /* Adjust headers by option size. */ 1735 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1736 m->m_len += optlen; 1737 m->m_pkthdr.len += optlen; 1738 #ifdef INET6 1739 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1740 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1741 else 1742 #endif 1743 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1744 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1745 if (sc->sc_flags & SCF_SIGNATURE) { 1746 KASSERT(to.to_flags & TOF_SIGNATURE, 1747 ("tcp_addoptions() didn't set tcp_signature")); 1748 1749 /* NOTE: to.to_signature is inside of mbuf */ 1750 if (!TCPMD5_ENABLED() || 1751 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1752 m_freem(m); 1753 return (EACCES); 1754 } 1755 } 1756 #endif 1757 } else 1758 optlen = 0; 1759 1760 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1761 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1762 /* 1763 * If we have peer's SYN and it has a flowid, then let's assign it to 1764 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1765 * to SYN|ACK due to lack of inp here. 1766 */ 1767 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1768 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1769 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1770 } 1771 #ifdef INET6 1772 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1773 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1774 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1775 IPPROTO_TCP, 0); 1776 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1777 #ifdef TCP_OFFLOAD 1778 if (ADDED_BY_TOE(sc)) { 1779 struct toedev *tod = sc->sc_tod; 1780 1781 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1782 1783 return (error); 1784 } 1785 #endif 1786 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1787 } 1788 #endif 1789 #if defined(INET6) && defined(INET) 1790 else 1791 #endif 1792 #ifdef INET 1793 { 1794 m->m_pkthdr.csum_flags = CSUM_TCP; 1795 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1796 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1797 #ifdef TCP_OFFLOAD 1798 if (ADDED_BY_TOE(sc)) { 1799 struct toedev *tod = sc->sc_tod; 1800 1801 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1802 1803 return (error); 1804 } 1805 #endif 1806 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1807 } 1808 #endif 1809 return (error); 1810 } 1811 1812 /* 1813 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1814 * that exceed the capacity of the syncache by avoiding the storage of any 1815 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1816 * attacks where the attacker does not have access to our responses. 1817 * 1818 * Syncookies encode and include all necessary information about the 1819 * connection setup within the SYN|ACK that we send back. That way we 1820 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1821 * (if ever). Normally the syncache and syncookies are running in parallel 1822 * with the latter taking over when the former is exhausted. When matching 1823 * syncache entry is found the syncookie is ignored. 1824 * 1825 * The only reliable information persisting the 3WHS is our initial sequence 1826 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1827 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1828 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1829 * returns and signifies a legitimate connection if it matches the ACK. 1830 * 1831 * The available space of 32 bits to store the hash and to encode the SYN 1832 * option information is very tight and we should have at least 24 bits for 1833 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1834 * 1835 * SYN option information we have to encode to fully restore a connection: 1836 * MSS: is imporant to chose an optimal segment size to avoid IP level 1837 * fragmentation along the path. The common MSS values can be encoded 1838 * in a 3-bit table. Uncommon values are captured by the next lower value 1839 * in the table leading to a slight increase in packetization overhead. 1840 * WSCALE: is necessary to allow large windows to be used for high delay- 1841 * bandwidth product links. Not scaling the window when it was initially 1842 * negotiated is bad for performance as lack of scaling further decreases 1843 * the apparent available send window. We only need to encode the WSCALE 1844 * we received from the remote end. Our end can be recalculated at any 1845 * time. The common WSCALE values can be encoded in a 3-bit table. 1846 * Uncommon values are captured by the next lower value in the table 1847 * making us under-estimate the available window size halving our 1848 * theoretically possible maximum throughput for that connection. 1849 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1850 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1851 * that are included in all segments on a connection. We enable them when 1852 * the ACK has them. 1853 * 1854 * Security of syncookies and attack vectors: 1855 * 1856 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1857 * together with the gloabl secret to make it unique per connection attempt. 1858 * Thus any change of any of those parameters results in a different MAC output 1859 * in an unpredictable way unless a collision is encountered. 24 bits of the 1860 * MAC are embedded into the ISS. 1861 * 1862 * To prevent replay attacks two rotating global secrets are updated with a 1863 * new random value every 15 seconds. The life-time of a syncookie is thus 1864 * 15-30 seconds. 1865 * 1866 * Vector 1: Attacking the secret. This requires finding a weakness in the 1867 * MAC itself or the way it is used here. The attacker can do a chosen plain 1868 * text attack by varying and testing the all parameters under his control. 1869 * The strength depends on the size and randomness of the secret, and the 1870 * cryptographic security of the MAC function. Due to the constant updating 1871 * of the secret the attacker has at most 29.999 seconds to find the secret 1872 * and launch spoofed connections. After that he has to start all over again. 1873 * 1874 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1875 * size an average of 4,823 attempts are required for a 50% chance of success 1876 * to spoof a single syncookie (birthday collision paradox). However the 1877 * attacker is blind and doesn't know if one of his attempts succeeded unless 1878 * he has a side channel to interfere success from. A single connection setup 1879 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1880 * This many attempts are required for each one blind spoofed connection. For 1881 * every additional spoofed connection he has to launch another N attempts. 1882 * Thus for a sustained rate 100 spoofed connections per second approximately 1883 * 1,800,000 packets per second would have to be sent. 1884 * 1885 * NB: The MAC function should be fast so that it doesn't become a CPU 1886 * exhaustion attack vector itself. 1887 * 1888 * References: 1889 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1890 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1891 * http://cr.yp.to/syncookies.html (overview) 1892 * http://cr.yp.to/syncookies/archive (details) 1893 * 1894 * 1895 * Schematic construction of a syncookie enabled Initial Sequence Number: 1896 * 0 1 2 3 1897 * 12345678901234567890123456789012 1898 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1899 * 1900 * x 24 MAC (truncated) 1901 * W 3 Send Window Scale index 1902 * M 3 MSS index 1903 * S 1 SACK permitted 1904 * P 1 Odd/even secret 1905 */ 1906 1907 /* 1908 * Distribution and probability of certain MSS values. Those in between are 1909 * rounded down to the next lower one. 1910 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1911 * .2% .3% 5% 7% 7% 20% 15% 45% 1912 */ 1913 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1914 1915 /* 1916 * Distribution and probability of certain WSCALE values. We have to map the 1917 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1918 * bits based on prevalence of certain values. Where we don't have an exact 1919 * match for are rounded down to the next lower one letting us under-estimate 1920 * the true available window. At the moment this would happen only for the 1921 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1922 * and window size). The absence of the WSCALE option (no scaling in either 1923 * direction) is encoded with index zero. 1924 * [WSCALE values histograms, Allman, 2012] 1925 * X 10 10 35 5 6 14 10% by host 1926 * X 11 4 5 5 18 49 3% by connections 1927 */ 1928 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1929 1930 /* 1931 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1932 * and good cryptographic properties. 1933 */ 1934 static uint32_t 1935 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1936 uint8_t *secbits, uintptr_t secmod) 1937 { 1938 SIPHASH_CTX ctx; 1939 uint32_t siphash[2]; 1940 1941 SipHash24_Init(&ctx); 1942 SipHash_SetKey(&ctx, secbits); 1943 switch (inc->inc_flags & INC_ISIPV6) { 1944 #ifdef INET 1945 case 0: 1946 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1947 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1948 break; 1949 #endif 1950 #ifdef INET6 1951 case INC_ISIPV6: 1952 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1953 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1954 break; 1955 #endif 1956 } 1957 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1958 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1959 SipHash_Update(&ctx, &irs, sizeof(irs)); 1960 SipHash_Update(&ctx, &flags, sizeof(flags)); 1961 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1962 SipHash_Final((u_int8_t *)&siphash, &ctx); 1963 1964 return (siphash[0] ^ siphash[1]); 1965 } 1966 1967 static tcp_seq 1968 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 1969 { 1970 u_int i, secbit, wscale; 1971 uint32_t iss, hash; 1972 uint8_t *secbits; 1973 union syncookie cookie; 1974 1975 SCH_LOCK_ASSERT(sch); 1976 1977 cookie.cookie = 0; 1978 1979 /* Map our computed MSS into the 3-bit index. */ 1980 for (i = nitems(tcp_sc_msstab) - 1; 1981 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 1982 i--) 1983 ; 1984 cookie.flags.mss_idx = i; 1985 1986 /* 1987 * Map the send window scale into the 3-bit index but only if 1988 * the wscale option was received. 1989 */ 1990 if (sc->sc_flags & SCF_WINSCALE) { 1991 wscale = sc->sc_requested_s_scale; 1992 for (i = nitems(tcp_sc_wstab) - 1; 1993 tcp_sc_wstab[i] > wscale && i > 0; 1994 i--) 1995 ; 1996 cookie.flags.wscale_idx = i; 1997 } 1998 1999 /* Can we do SACK? */ 2000 if (sc->sc_flags & SCF_SACK) 2001 cookie.flags.sack_ok = 1; 2002 2003 /* Which of the two secrets to use. */ 2004 secbit = sch->sch_sc->secret.oddeven & 0x1; 2005 cookie.flags.odd_even = secbit; 2006 2007 secbits = sch->sch_sc->secret.key[secbit]; 2008 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2009 (uintptr_t)sch); 2010 2011 /* 2012 * Put the flags into the hash and XOR them to get better ISS number 2013 * variance. This doesn't enhance the cryptographic strength and is 2014 * done to prevent the 8 cookie bits from showing up directly on the 2015 * wire. 2016 */ 2017 iss = hash & ~0xff; 2018 iss |= cookie.cookie ^ (hash >> 24); 2019 2020 /* Randomize the timestamp. */ 2021 if (sc->sc_flags & SCF_TIMESTAMP) { 2022 sc->sc_ts = arc4random(); 2023 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 2024 } 2025 2026 TCPSTAT_INC(tcps_sc_sendcookie); 2027 return (iss); 2028 } 2029 2030 static struct syncache * 2031 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2032 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2033 struct socket *lso) 2034 { 2035 uint32_t hash; 2036 uint8_t *secbits; 2037 tcp_seq ack, seq; 2038 int wnd, wscale = 0; 2039 union syncookie cookie; 2040 2041 SCH_LOCK_ASSERT(sch); 2042 2043 /* 2044 * Pull information out of SYN-ACK/ACK and revert sequence number 2045 * advances. 2046 */ 2047 ack = th->th_ack - 1; 2048 seq = th->th_seq - 1; 2049 2050 /* 2051 * Unpack the flags containing enough information to restore the 2052 * connection. 2053 */ 2054 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2055 2056 /* Which of the two secrets to use. */ 2057 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 2058 2059 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2060 2061 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2062 if ((ack & ~0xff) != (hash & ~0xff)) 2063 return (NULL); 2064 2065 /* Fill in the syncache values. */ 2066 sc->sc_flags = 0; 2067 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2068 sc->sc_ipopts = NULL; 2069 2070 sc->sc_irs = seq; 2071 sc->sc_iss = ack; 2072 2073 switch (inc->inc_flags & INC_ISIPV6) { 2074 #ifdef INET 2075 case 0: 2076 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2077 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2078 break; 2079 #endif 2080 #ifdef INET6 2081 case INC_ISIPV6: 2082 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2083 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 2084 break; 2085 #endif 2086 } 2087 2088 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2089 2090 /* We can simply recompute receive window scale we sent earlier. */ 2091 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2092 wscale++; 2093 2094 /* Only use wscale if it was enabled in the orignal SYN. */ 2095 if (cookie.flags.wscale_idx > 0) { 2096 sc->sc_requested_r_scale = wscale; 2097 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2098 sc->sc_flags |= SCF_WINSCALE; 2099 } 2100 2101 wnd = lso->sol_sbrcv_hiwat; 2102 wnd = imax(wnd, 0); 2103 wnd = imin(wnd, TCP_MAXWIN); 2104 sc->sc_wnd = wnd; 2105 2106 if (cookie.flags.sack_ok) 2107 sc->sc_flags |= SCF_SACK; 2108 2109 if (to->to_flags & TOF_TS) { 2110 sc->sc_flags |= SCF_TIMESTAMP; 2111 sc->sc_tsreflect = to->to_tsval; 2112 sc->sc_ts = to->to_tsecr; 2113 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 2114 } 2115 2116 if (to->to_flags & TOF_SIGNATURE) 2117 sc->sc_flags |= SCF_SIGNATURE; 2118 2119 sc->sc_rxmits = 0; 2120 2121 TCPSTAT_INC(tcps_sc_recvcookie); 2122 return (sc); 2123 } 2124 2125 #ifdef INVARIANTS 2126 static int 2127 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2128 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2129 struct socket *lso) 2130 { 2131 struct syncache scs, *scx; 2132 char *s; 2133 2134 bzero(&scs, sizeof(scs)); 2135 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2136 2137 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2138 return (0); 2139 2140 if (scx != NULL) { 2141 if (sc->sc_peer_mss != scx->sc_peer_mss) 2142 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2143 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2144 2145 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2146 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2147 s, __func__, sc->sc_requested_r_scale, 2148 scx->sc_requested_r_scale); 2149 2150 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2151 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2152 s, __func__, sc->sc_requested_s_scale, 2153 scx->sc_requested_s_scale); 2154 2155 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2156 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2157 } 2158 2159 if (s != NULL) 2160 free(s, M_TCPLOG); 2161 return (0); 2162 } 2163 #endif /* INVARIANTS */ 2164 2165 static void 2166 syncookie_reseed(void *arg) 2167 { 2168 struct tcp_syncache *sc = arg; 2169 uint8_t *secbits; 2170 int secbit; 2171 2172 /* 2173 * Reseeding the secret doesn't have to be protected by a lock. 2174 * It only must be ensured that the new random values are visible 2175 * to all CPUs in a SMP environment. The atomic with release 2176 * semantics ensures that. 2177 */ 2178 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2179 secbits = sc->secret.key[secbit]; 2180 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2181 atomic_add_rel_int(&sc->secret.oddeven, 1); 2182 2183 /* Reschedule ourself. */ 2184 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2185 } 2186 2187 /* 2188 * Exports the syncache entries to userland so that netstat can display 2189 * them alongside the other sockets. This function is intended to be 2190 * called only from tcp_pcblist. 2191 * 2192 * Due to concurrency on an active system, the number of pcbs exported 2193 * may have no relation to max_pcbs. max_pcbs merely indicates the 2194 * amount of space the caller allocated for this function to use. 2195 */ 2196 int 2197 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2198 { 2199 struct xtcpcb xt; 2200 struct syncache *sc; 2201 struct syncache_head *sch; 2202 int count, error, i; 2203 2204 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2205 sch = &V_tcp_syncache.hashbase[i]; 2206 SCH_LOCK(sch); 2207 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2208 if (count >= max_pcbs) { 2209 SCH_UNLOCK(sch); 2210 goto exit; 2211 } 2212 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2213 continue; 2214 bzero(&xt, sizeof(xt)); 2215 xt.xt_len = sizeof(xt); 2216 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2217 xt.xt_inp.inp_vflag = INP_IPV6; 2218 else 2219 xt.xt_inp.inp_vflag = INP_IPV4; 2220 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2221 sizeof (struct in_conninfo)); 2222 xt.t_state = TCPS_SYN_RECEIVED; 2223 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2224 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2225 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2226 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2227 error = SYSCTL_OUT(req, &xt, sizeof xt); 2228 if (error) { 2229 SCH_UNLOCK(sch); 2230 goto exit; 2231 } 2232 count++; 2233 } 2234 SCH_UNLOCK(sch); 2235 } 2236 exit: 2237 *pcbs_exported = count; 2238 return error; 2239 } 2240