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 /* 856 * XXXrrs this is quite dangerous, it is possible 857 * for the new function to fail to init. We also 858 * are not asking if the handoff_is_ok though at 859 * the very start thats probalbly ok. 860 */ 861 if (tp->t_fb->tfb_tcp_fb_init) { 862 (*tp->t_fb->tfb_tcp_fb_init)(tp); 863 } 864 } 865 tp->snd_wl1 = sc->sc_irs; 866 tp->snd_max = tp->iss + 1; 867 tp->snd_nxt = tp->iss + 1; 868 tp->rcv_up = sc->sc_irs + 1; 869 tp->rcv_wnd = sc->sc_wnd; 870 tp->rcv_adv += tp->rcv_wnd; 871 tp->last_ack_sent = tp->rcv_nxt; 872 873 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 874 if (sc->sc_flags & SCF_NOOPT) 875 tp->t_flags |= TF_NOOPT; 876 else { 877 if (sc->sc_flags & SCF_WINSCALE) { 878 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 879 tp->snd_scale = sc->sc_requested_s_scale; 880 tp->request_r_scale = sc->sc_requested_r_scale; 881 } 882 if (sc->sc_flags & SCF_TIMESTAMP) { 883 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 884 tp->ts_recent = sc->sc_tsreflect; 885 tp->ts_recent_age = tcp_ts_getticks(); 886 tp->ts_offset = sc->sc_tsoff; 887 } 888 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 889 if (sc->sc_flags & SCF_SIGNATURE) 890 tp->t_flags |= TF_SIGNATURE; 891 #endif 892 if (sc->sc_flags & SCF_SACK) 893 tp->t_flags |= TF_SACK_PERMIT; 894 } 895 896 if (sc->sc_flags & SCF_ECN) 897 tp->t_flags |= TF_ECN_PERMIT; 898 899 /* 900 * Set up MSS and get cached values from tcp_hostcache. 901 * This might overwrite some of the defaults we just set. 902 */ 903 tcp_mss(tp, sc->sc_peer_mss); 904 905 /* 906 * If the SYN,ACK was retransmitted, indicate that CWND to be 907 * limited to one segment in cc_conn_init(). 908 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 909 */ 910 if (sc->sc_rxmits > 1) 911 tp->snd_cwnd = 1; 912 913 #ifdef TCP_OFFLOAD 914 /* 915 * Allow a TOE driver to install its hooks. Note that we hold the 916 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 917 * new connection before the TOE driver has done its thing. 918 */ 919 if (ADDED_BY_TOE(sc)) { 920 struct toedev *tod = sc->sc_tod; 921 922 tod->tod_offload_socket(tod, sc->sc_todctx, so); 923 } 924 #endif 925 /* 926 * Copy and activate timers. 927 */ 928 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 929 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 930 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 931 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 932 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 933 934 TCPSTAT_INC(tcps_accepts); 935 return (so); 936 937 abort: 938 INP_WUNLOCK(inp); 939 abort2: 940 if (so != NULL) 941 soabort(so); 942 return (NULL); 943 } 944 945 /* 946 * This function gets called when we receive an ACK for a 947 * socket in the LISTEN state. We look up the connection 948 * in the syncache, and if its there, we pull it out of 949 * the cache and turn it into a full-blown connection in 950 * the SYN-RECEIVED state. 951 * 952 * On syncache_socket() success the newly created socket 953 * has its underlying inp locked. 954 */ 955 int 956 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 957 struct socket **lsop, struct mbuf *m) 958 { 959 struct syncache *sc; 960 struct syncache_head *sch; 961 struct syncache scs; 962 char *s; 963 964 /* 965 * Global TCP locks are held because we manipulate the PCB lists 966 * and create a new socket. 967 */ 968 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 969 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 970 ("%s: can handle only ACK", __func__)); 971 972 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 973 SCH_LOCK_ASSERT(sch); 974 975 #ifdef INVARIANTS 976 /* 977 * Test code for syncookies comparing the syncache stored 978 * values with the reconstructed values from the cookie. 979 */ 980 if (sc != NULL) 981 syncookie_cmp(inc, sch, sc, th, to, *lsop); 982 #endif 983 984 if (sc == NULL) { 985 /* 986 * There is no syncache entry, so see if this ACK is 987 * a returning syncookie. To do this, first: 988 * A. Check if syncookies are used in case of syncache 989 * overflows 990 * B. See if this socket has had a syncache entry dropped in 991 * the recent past. We don't want to accept a bogus 992 * syncookie if we've never received a SYN or accept it 993 * twice. 994 * C. check that the syncookie is valid. If it is, then 995 * cobble up a fake syncache entry, and return. 996 */ 997 if (!V_tcp_syncookies) { 998 SCH_UNLOCK(sch); 999 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1000 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1001 "segment rejected (syncookies disabled)\n", 1002 s, __func__); 1003 goto failed; 1004 } 1005 if (!V_tcp_syncookiesonly && 1006 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1007 SCH_UNLOCK(sch); 1008 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1009 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1010 "segment rejected (no syncache entry)\n", 1011 s, __func__); 1012 goto failed; 1013 } 1014 bzero(&scs, sizeof(scs)); 1015 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1016 SCH_UNLOCK(sch); 1017 if (sc == NULL) { 1018 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1019 log(LOG_DEBUG, "%s; %s: Segment failed " 1020 "SYNCOOKIE authentication, segment rejected " 1021 "(probably spoofed)\n", s, __func__); 1022 goto failed; 1023 } 1024 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1025 /* If received ACK has MD5 signature, check it. */ 1026 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1027 (!TCPMD5_ENABLED() || 1028 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1029 /* Drop the ACK. */ 1030 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1031 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1032 "MD5 signature doesn't match.\n", 1033 s, __func__); 1034 free(s, M_TCPLOG); 1035 } 1036 TCPSTAT_INC(tcps_sig_err_sigopt); 1037 return (-1); /* Do not send RST */ 1038 } 1039 #endif /* TCP_SIGNATURE */ 1040 } else { 1041 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1042 /* 1043 * If listening socket requested TCP digests, check that 1044 * received ACK has signature and it is correct. 1045 * If not, drop the ACK and leave sc entry in th cache, 1046 * because SYN was received with correct signature. 1047 */ 1048 if (sc->sc_flags & SCF_SIGNATURE) { 1049 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1050 /* No signature */ 1051 TCPSTAT_INC(tcps_sig_err_nosigopt); 1052 SCH_UNLOCK(sch); 1053 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1054 log(LOG_DEBUG, "%s; %s: Segment " 1055 "rejected, MD5 signature wasn't " 1056 "provided.\n", s, __func__); 1057 free(s, M_TCPLOG); 1058 } 1059 return (-1); /* Do not send RST */ 1060 } 1061 if (!TCPMD5_ENABLED() || 1062 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1063 /* Doesn't match or no SA */ 1064 SCH_UNLOCK(sch); 1065 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1066 log(LOG_DEBUG, "%s; %s: Segment " 1067 "rejected, MD5 signature doesn't " 1068 "match.\n", s, __func__); 1069 free(s, M_TCPLOG); 1070 } 1071 return (-1); /* Do not send RST */ 1072 } 1073 } 1074 #endif /* TCP_SIGNATURE */ 1075 /* 1076 * Pull out the entry to unlock the bucket row. 1077 * 1078 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1079 * tcp_state_change(). The tcpcb is not existent at this 1080 * moment. A new one will be allocated via syncache_socket-> 1081 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1082 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1083 */ 1084 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1085 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1086 sch->sch_length--; 1087 #ifdef TCP_OFFLOAD 1088 if (ADDED_BY_TOE(sc)) { 1089 struct toedev *tod = sc->sc_tod; 1090 1091 tod->tod_syncache_removed(tod, sc->sc_todctx); 1092 } 1093 #endif 1094 SCH_UNLOCK(sch); 1095 } 1096 1097 /* 1098 * Segment validation: 1099 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1100 */ 1101 if (th->th_ack != sc->sc_iss + 1) { 1102 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1103 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1104 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1105 goto failed; 1106 } 1107 1108 /* 1109 * The SEQ must fall in the window starting at the received 1110 * initial receive sequence number + 1 (the SYN). 1111 */ 1112 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1113 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1114 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1115 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1116 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1117 goto failed; 1118 } 1119 1120 /* 1121 * If timestamps were not negotiated during SYN/ACK they 1122 * must not appear on any segment during this session. 1123 */ 1124 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1125 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1126 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1127 "segment rejected\n", s, __func__); 1128 goto failed; 1129 } 1130 1131 /* 1132 * If timestamps were negotiated during SYN/ACK they should 1133 * appear on every segment during this session. 1134 * XXXAO: This is only informal as there have been unverified 1135 * reports of non-compliants stacks. 1136 */ 1137 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1138 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1139 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1140 "no action\n", s, __func__); 1141 free(s, M_TCPLOG); 1142 s = NULL; 1143 } 1144 } 1145 1146 /* 1147 * If timestamps were negotiated, the reflected timestamp 1148 * must be equal to what we actually sent in the SYN|ACK 1149 * except in the case of 0. Some boxes are known for sending 1150 * broken timestamp replies during the 3whs (and potentially 1151 * during the connection also). 1152 * 1153 * Accept the final ACK of 3whs with reflected timestamp of 0 1154 * instead of sending a RST and deleting the syncache entry. 1155 */ 1156 if ((to->to_flags & TOF_TS) && to->to_tsecr && 1157 to->to_tsecr != sc->sc_ts) { 1158 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1159 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1160 "segment rejected\n", 1161 s, __func__, to->to_tsecr, sc->sc_ts); 1162 goto failed; 1163 } 1164 1165 *lsop = syncache_socket(sc, *lsop, m); 1166 1167 if (*lsop == NULL) 1168 TCPSTAT_INC(tcps_sc_aborted); 1169 else 1170 TCPSTAT_INC(tcps_sc_completed); 1171 1172 /* how do we find the inp for the new socket? */ 1173 if (sc != &scs) 1174 syncache_free(sc); 1175 return (1); 1176 failed: 1177 if (sc != NULL && sc != &scs) 1178 syncache_free(sc); 1179 if (s != NULL) 1180 free(s, M_TCPLOG); 1181 *lsop = NULL; 1182 return (0); 1183 } 1184 1185 static void 1186 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1187 uint64_t response_cookie) 1188 { 1189 struct inpcb *inp; 1190 struct tcpcb *tp; 1191 unsigned int *pending_counter; 1192 1193 /* 1194 * Global TCP locks are held because we manipulate the PCB lists 1195 * and create a new socket. 1196 */ 1197 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1198 1199 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1200 *lsop = syncache_socket(sc, *lsop, m); 1201 if (*lsop == NULL) { 1202 TCPSTAT_INC(tcps_sc_aborted); 1203 atomic_subtract_int(pending_counter, 1); 1204 } else { 1205 soisconnected(*lsop); 1206 inp = sotoinpcb(*lsop); 1207 tp = intotcpcb(inp); 1208 tp->t_flags |= TF_FASTOPEN; 1209 tp->t_tfo_cookie.server = response_cookie; 1210 tp->snd_max = tp->iss; 1211 tp->snd_nxt = tp->iss; 1212 tp->t_tfo_pending = pending_counter; 1213 TCPSTAT_INC(tcps_sc_completed); 1214 } 1215 } 1216 1217 /* 1218 * Given a LISTEN socket and an inbound SYN request, add 1219 * this to the syn cache, and send back a segment: 1220 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1221 * to the source. 1222 * 1223 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1224 * Doing so would require that we hold onto the data and deliver it 1225 * to the application. However, if we are the target of a SYN-flood 1226 * DoS attack, an attacker could send data which would eventually 1227 * consume all available buffer space if it were ACKed. By not ACKing 1228 * the data, we avoid this DoS scenario. 1229 * 1230 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1231 * cookie is processed and a new socket is created. In this case, any data 1232 * accompanying the SYN will be queued to the socket by tcp_input() and will 1233 * be ACKed either when the application sends response data or the delayed 1234 * ACK timer expires, whichever comes first. 1235 */ 1236 int 1237 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1238 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1239 void *todctx) 1240 { 1241 struct tcpcb *tp; 1242 struct socket *so; 1243 struct syncache *sc = NULL; 1244 struct syncache_head *sch; 1245 struct mbuf *ipopts = NULL; 1246 u_int ltflags; 1247 int win, ip_ttl, ip_tos; 1248 char *s; 1249 int rv = 0; 1250 #ifdef INET6 1251 int autoflowlabel = 0; 1252 #endif 1253 #ifdef MAC 1254 struct label *maclabel; 1255 #endif 1256 struct syncache scs; 1257 struct ucred *cred; 1258 uint64_t tfo_response_cookie; 1259 unsigned int *tfo_pending = NULL; 1260 int tfo_cookie_valid = 0; 1261 int tfo_response_cookie_valid = 0; 1262 1263 INP_WLOCK_ASSERT(inp); /* listen socket */ 1264 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1265 ("%s: unexpected tcp flags", __func__)); 1266 1267 /* 1268 * Combine all so/tp operations very early to drop the INP lock as 1269 * soon as possible. 1270 */ 1271 so = *lsop; 1272 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1273 tp = sototcpcb(so); 1274 cred = crhold(so->so_cred); 1275 1276 #ifdef INET6 1277 if ((inc->inc_flags & INC_ISIPV6) && 1278 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1279 autoflowlabel = 1; 1280 #endif 1281 ip_ttl = inp->inp_ip_ttl; 1282 ip_tos = inp->inp_ip_tos; 1283 win = so->sol_sbrcv_hiwat; 1284 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1285 1286 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1287 (tp->t_tfo_pending != NULL) && 1288 (to->to_flags & TOF_FASTOPEN)) { 1289 /* 1290 * Limit the number of pending TFO connections to 1291 * approximately half of the queue limit. This prevents TFO 1292 * SYN floods from starving the service by filling the 1293 * listen queue with bogus TFO connections. 1294 */ 1295 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1296 (so->sol_qlimit / 2)) { 1297 int result; 1298 1299 result = tcp_fastopen_check_cookie(inc, 1300 to->to_tfo_cookie, to->to_tfo_len, 1301 &tfo_response_cookie); 1302 tfo_cookie_valid = (result > 0); 1303 tfo_response_cookie_valid = (result >= 0); 1304 } 1305 1306 /* 1307 * Remember the TFO pending counter as it will have to be 1308 * decremented below if we don't make it to syncache_tfo_expand(). 1309 */ 1310 tfo_pending = tp->t_tfo_pending; 1311 } 1312 1313 /* By the time we drop the lock these should no longer be used. */ 1314 so = NULL; 1315 tp = NULL; 1316 1317 #ifdef MAC 1318 if (mac_syncache_init(&maclabel) != 0) { 1319 INP_WUNLOCK(inp); 1320 goto done; 1321 } else 1322 mac_syncache_create(maclabel, inp); 1323 #endif 1324 if (!tfo_cookie_valid) 1325 INP_WUNLOCK(inp); 1326 1327 /* 1328 * Remember the IP options, if any. 1329 */ 1330 #ifdef INET6 1331 if (!(inc->inc_flags & INC_ISIPV6)) 1332 #endif 1333 #ifdef INET 1334 ipopts = (m) ? ip_srcroute(m) : NULL; 1335 #else 1336 ipopts = NULL; 1337 #endif 1338 1339 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1340 /* 1341 * If listening socket requested TCP digests, check that received 1342 * SYN has signature and it is correct. If signature doesn't match 1343 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1344 */ 1345 if (ltflags & TF_SIGNATURE) { 1346 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1347 TCPSTAT_INC(tcps_sig_err_nosigopt); 1348 goto done; 1349 } 1350 if (!TCPMD5_ENABLED() || 1351 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1352 goto done; 1353 } 1354 #endif /* TCP_SIGNATURE */ 1355 /* 1356 * See if we already have an entry for this connection. 1357 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1358 * 1359 * XXX: should the syncache be re-initialized with the contents 1360 * of the new SYN here (which may have different options?) 1361 * 1362 * XXX: We do not check the sequence number to see if this is a 1363 * real retransmit or a new connection attempt. The question is 1364 * how to handle such a case; either ignore it as spoofed, or 1365 * drop the current entry and create a new one? 1366 */ 1367 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1368 SCH_LOCK_ASSERT(sch); 1369 if (sc != NULL) { 1370 if (tfo_cookie_valid) 1371 INP_WUNLOCK(inp); 1372 TCPSTAT_INC(tcps_sc_dupsyn); 1373 if (ipopts) { 1374 /* 1375 * If we were remembering a previous source route, 1376 * forget it and use the new one we've been given. 1377 */ 1378 if (sc->sc_ipopts) 1379 (void) m_free(sc->sc_ipopts); 1380 sc->sc_ipopts = ipopts; 1381 } 1382 /* 1383 * Update timestamp if present. 1384 */ 1385 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1386 sc->sc_tsreflect = to->to_tsval; 1387 else 1388 sc->sc_flags &= ~SCF_TIMESTAMP; 1389 #ifdef MAC 1390 /* 1391 * Since we have already unconditionally allocated label 1392 * storage, free it up. The syncache entry will already 1393 * have an initialized label we can use. 1394 */ 1395 mac_syncache_destroy(&maclabel); 1396 #endif 1397 /* Retransmit SYN|ACK and reset retransmit count. */ 1398 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1399 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1400 "resetting timer and retransmitting SYN|ACK\n", 1401 s, __func__); 1402 free(s, M_TCPLOG); 1403 } 1404 if (syncache_respond(sc, sch, 1, m) == 0) { 1405 sc->sc_rxmits = 0; 1406 syncache_timeout(sc, sch, 1); 1407 TCPSTAT_INC(tcps_sndacks); 1408 TCPSTAT_INC(tcps_sndtotal); 1409 } 1410 SCH_UNLOCK(sch); 1411 goto done; 1412 } 1413 1414 if (tfo_cookie_valid) { 1415 bzero(&scs, sizeof(scs)); 1416 sc = &scs; 1417 goto skip_alloc; 1418 } 1419 1420 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1421 if (sc == NULL) { 1422 /* 1423 * The zone allocator couldn't provide more entries. 1424 * Treat this as if the cache was full; drop the oldest 1425 * entry and insert the new one. 1426 */ 1427 TCPSTAT_INC(tcps_sc_zonefail); 1428 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1429 sch->sch_last_overflow = time_uptime; 1430 syncache_drop(sc, sch); 1431 } 1432 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1433 if (sc == NULL) { 1434 if (V_tcp_syncookies) { 1435 bzero(&scs, sizeof(scs)); 1436 sc = &scs; 1437 } else { 1438 SCH_UNLOCK(sch); 1439 if (ipopts) 1440 (void) m_free(ipopts); 1441 goto done; 1442 } 1443 } 1444 } 1445 1446 skip_alloc: 1447 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1448 sc->sc_tfo_cookie = &tfo_response_cookie; 1449 1450 /* 1451 * Fill in the syncache values. 1452 */ 1453 #ifdef MAC 1454 sc->sc_label = maclabel; 1455 #endif 1456 sc->sc_cred = cred; 1457 cred = NULL; 1458 sc->sc_ipopts = ipopts; 1459 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1460 #ifdef INET6 1461 if (!(inc->inc_flags & INC_ISIPV6)) 1462 #endif 1463 { 1464 sc->sc_ip_tos = ip_tos; 1465 sc->sc_ip_ttl = ip_ttl; 1466 } 1467 #ifdef TCP_OFFLOAD 1468 sc->sc_tod = tod; 1469 sc->sc_todctx = todctx; 1470 #endif 1471 sc->sc_irs = th->th_seq; 1472 sc->sc_iss = arc4random(); 1473 sc->sc_flags = 0; 1474 sc->sc_flowlabel = 0; 1475 1476 /* 1477 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1478 * win was derived from socket earlier in the function. 1479 */ 1480 win = imax(win, 0); 1481 win = imin(win, TCP_MAXWIN); 1482 sc->sc_wnd = win; 1483 1484 if (V_tcp_do_rfc1323) { 1485 /* 1486 * A timestamp received in a SYN makes 1487 * it ok to send timestamp requests and replies. 1488 */ 1489 if (to->to_flags & TOF_TS) { 1490 sc->sc_tsreflect = to->to_tsval; 1491 sc->sc_ts = tcp_ts_getticks(); 1492 sc->sc_flags |= SCF_TIMESTAMP; 1493 } 1494 if (to->to_flags & TOF_SCALE) { 1495 int wscale = 0; 1496 1497 /* 1498 * Pick the smallest possible scaling factor that 1499 * will still allow us to scale up to sb_max, aka 1500 * kern.ipc.maxsockbuf. 1501 * 1502 * We do this because there are broken firewalls that 1503 * will corrupt the window scale option, leading to 1504 * the other endpoint believing that our advertised 1505 * window is unscaled. At scale factors larger than 1506 * 5 the unscaled window will drop below 1500 bytes, 1507 * leading to serious problems when traversing these 1508 * broken firewalls. 1509 * 1510 * With the default maxsockbuf of 256K, a scale factor 1511 * of 3 will be chosen by this algorithm. Those who 1512 * choose a larger maxsockbuf should watch out 1513 * for the compatibility problems mentioned above. 1514 * 1515 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1516 * or <SYN,ACK>) segment itself is never scaled. 1517 */ 1518 while (wscale < TCP_MAX_WINSHIFT && 1519 (TCP_MAXWIN << wscale) < sb_max) 1520 wscale++; 1521 sc->sc_requested_r_scale = wscale; 1522 sc->sc_requested_s_scale = to->to_wscale; 1523 sc->sc_flags |= SCF_WINSCALE; 1524 } 1525 } 1526 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1527 /* 1528 * If listening socket requested TCP digests, flag this in the 1529 * syncache so that syncache_respond() will do the right thing 1530 * with the SYN+ACK. 1531 */ 1532 if (ltflags & TF_SIGNATURE) 1533 sc->sc_flags |= SCF_SIGNATURE; 1534 #endif /* TCP_SIGNATURE */ 1535 if (to->to_flags & TOF_SACKPERM) 1536 sc->sc_flags |= SCF_SACK; 1537 if (to->to_flags & TOF_MSS) 1538 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1539 if (ltflags & TF_NOOPT) 1540 sc->sc_flags |= SCF_NOOPT; 1541 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1542 sc->sc_flags |= SCF_ECN; 1543 1544 if (V_tcp_syncookies) 1545 sc->sc_iss = syncookie_generate(sch, sc); 1546 #ifdef INET6 1547 if (autoflowlabel) { 1548 if (V_tcp_syncookies) 1549 sc->sc_flowlabel = sc->sc_iss; 1550 else 1551 sc->sc_flowlabel = ip6_randomflowlabel(); 1552 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1553 } 1554 #endif 1555 SCH_UNLOCK(sch); 1556 1557 if (tfo_cookie_valid) { 1558 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1559 /* INP_WUNLOCK(inp) will be performed by the caller */ 1560 rv = 1; 1561 goto tfo_expanded; 1562 } 1563 1564 /* 1565 * Do a standard 3-way handshake. 1566 */ 1567 if (syncache_respond(sc, sch, 0, m) == 0) { 1568 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1569 syncache_free(sc); 1570 else if (sc != &scs) 1571 syncache_insert(sc, sch); /* locks and unlocks sch */ 1572 TCPSTAT_INC(tcps_sndacks); 1573 TCPSTAT_INC(tcps_sndtotal); 1574 } else { 1575 if (sc != &scs) 1576 syncache_free(sc); 1577 TCPSTAT_INC(tcps_sc_dropped); 1578 } 1579 1580 done: 1581 if (m) { 1582 *lsop = NULL; 1583 m_freem(m); 1584 } 1585 /* 1586 * If tfo_pending is not NULL here, then a TFO SYN that did not 1587 * result in a new socket was processed and the associated pending 1588 * counter has not yet been decremented. All such TFO processing paths 1589 * transit this point. 1590 */ 1591 if (tfo_pending != NULL) 1592 tcp_fastopen_decrement_counter(tfo_pending); 1593 1594 tfo_expanded: 1595 if (cred != NULL) 1596 crfree(cred); 1597 #ifdef MAC 1598 if (sc == &scs) 1599 mac_syncache_destroy(&maclabel); 1600 #endif 1601 return (rv); 1602 } 1603 1604 /* 1605 * Send SYN|ACK to the peer. Either in response to the peer's SYN, 1606 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1607 */ 1608 static int 1609 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked, 1610 const struct mbuf *m0) 1611 { 1612 struct ip *ip = NULL; 1613 struct mbuf *m; 1614 struct tcphdr *th = NULL; 1615 int optlen, error = 0; /* Make compiler happy */ 1616 u_int16_t hlen, tlen, mssopt; 1617 struct tcpopt to; 1618 #ifdef INET6 1619 struct ip6_hdr *ip6 = NULL; 1620 #endif 1621 hlen = 1622 #ifdef INET6 1623 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1624 #endif 1625 sizeof(struct ip); 1626 tlen = hlen + sizeof(struct tcphdr); 1627 1628 /* Determine MSS we advertize to other end of connection. */ 1629 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); 1630 1631 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1632 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1633 ("syncache: mbuf too small")); 1634 1635 /* Create the IP+TCP header from scratch. */ 1636 m = m_gethdr(M_NOWAIT, MT_DATA); 1637 if (m == NULL) 1638 return (ENOBUFS); 1639 #ifdef MAC 1640 mac_syncache_create_mbuf(sc->sc_label, m); 1641 #endif 1642 m->m_data += max_linkhdr; 1643 m->m_len = tlen; 1644 m->m_pkthdr.len = tlen; 1645 m->m_pkthdr.rcvif = NULL; 1646 1647 #ifdef INET6 1648 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1649 ip6 = mtod(m, struct ip6_hdr *); 1650 ip6->ip6_vfc = IPV6_VERSION; 1651 ip6->ip6_nxt = IPPROTO_TCP; 1652 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1653 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1654 ip6->ip6_plen = htons(tlen - hlen); 1655 /* ip6_hlim is set after checksum */ 1656 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1657 ip6->ip6_flow |= sc->sc_flowlabel; 1658 1659 th = (struct tcphdr *)(ip6 + 1); 1660 } 1661 #endif 1662 #if defined(INET6) && defined(INET) 1663 else 1664 #endif 1665 #ifdef INET 1666 { 1667 ip = mtod(m, struct ip *); 1668 ip->ip_v = IPVERSION; 1669 ip->ip_hl = sizeof(struct ip) >> 2; 1670 ip->ip_len = htons(tlen); 1671 ip->ip_id = 0; 1672 ip->ip_off = 0; 1673 ip->ip_sum = 0; 1674 ip->ip_p = IPPROTO_TCP; 1675 ip->ip_src = sc->sc_inc.inc_laddr; 1676 ip->ip_dst = sc->sc_inc.inc_faddr; 1677 ip->ip_ttl = sc->sc_ip_ttl; 1678 ip->ip_tos = sc->sc_ip_tos; 1679 1680 /* 1681 * See if we should do MTU discovery. Route lookups are 1682 * expensive, so we will only unset the DF bit if: 1683 * 1684 * 1) path_mtu_discovery is disabled 1685 * 2) the SCF_UNREACH flag has been set 1686 */ 1687 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1688 ip->ip_off |= htons(IP_DF); 1689 1690 th = (struct tcphdr *)(ip + 1); 1691 } 1692 #endif /* INET */ 1693 th->th_sport = sc->sc_inc.inc_lport; 1694 th->th_dport = sc->sc_inc.inc_fport; 1695 1696 th->th_seq = htonl(sc->sc_iss); 1697 th->th_ack = htonl(sc->sc_irs + 1); 1698 th->th_off = sizeof(struct tcphdr) >> 2; 1699 th->th_x2 = 0; 1700 th->th_flags = TH_SYN|TH_ACK; 1701 th->th_win = htons(sc->sc_wnd); 1702 th->th_urp = 0; 1703 1704 if (sc->sc_flags & SCF_ECN) { 1705 th->th_flags |= TH_ECE; 1706 TCPSTAT_INC(tcps_ecn_shs); 1707 } 1708 1709 /* Tack on the TCP options. */ 1710 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1711 to.to_flags = 0; 1712 1713 to.to_mss = mssopt; 1714 to.to_flags = TOF_MSS; 1715 if (sc->sc_flags & SCF_WINSCALE) { 1716 to.to_wscale = sc->sc_requested_r_scale; 1717 to.to_flags |= TOF_SCALE; 1718 } 1719 if (sc->sc_flags & SCF_TIMESTAMP) { 1720 /* Virgin timestamp or TCP cookie enhanced one. */ 1721 to.to_tsval = sc->sc_ts; 1722 to.to_tsecr = sc->sc_tsreflect; 1723 to.to_flags |= TOF_TS; 1724 } 1725 if (sc->sc_flags & SCF_SACK) 1726 to.to_flags |= TOF_SACKPERM; 1727 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1728 if (sc->sc_flags & SCF_SIGNATURE) 1729 to.to_flags |= TOF_SIGNATURE; 1730 #endif 1731 if (sc->sc_tfo_cookie) { 1732 to.to_flags |= TOF_FASTOPEN; 1733 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1734 to.to_tfo_cookie = sc->sc_tfo_cookie; 1735 /* don't send cookie again when retransmitting response */ 1736 sc->sc_tfo_cookie = NULL; 1737 } 1738 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1739 1740 /* Adjust headers by option size. */ 1741 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1742 m->m_len += optlen; 1743 m->m_pkthdr.len += optlen; 1744 #ifdef INET6 1745 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1746 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1747 else 1748 #endif 1749 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1750 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1751 if (sc->sc_flags & SCF_SIGNATURE) { 1752 KASSERT(to.to_flags & TOF_SIGNATURE, 1753 ("tcp_addoptions() didn't set tcp_signature")); 1754 1755 /* NOTE: to.to_signature is inside of mbuf */ 1756 if (!TCPMD5_ENABLED() || 1757 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1758 m_freem(m); 1759 return (EACCES); 1760 } 1761 } 1762 #endif 1763 } else 1764 optlen = 0; 1765 1766 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1767 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1768 /* 1769 * If we have peer's SYN and it has a flowid, then let's assign it to 1770 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1771 * to SYN|ACK due to lack of inp here. 1772 */ 1773 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1774 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1775 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1776 } 1777 #ifdef INET6 1778 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1779 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1780 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1781 IPPROTO_TCP, 0); 1782 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1783 #ifdef TCP_OFFLOAD 1784 if (ADDED_BY_TOE(sc)) { 1785 struct toedev *tod = sc->sc_tod; 1786 1787 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1788 1789 return (error); 1790 } 1791 #endif 1792 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1793 } 1794 #endif 1795 #if defined(INET6) && defined(INET) 1796 else 1797 #endif 1798 #ifdef INET 1799 { 1800 m->m_pkthdr.csum_flags = CSUM_TCP; 1801 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1802 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1803 #ifdef TCP_OFFLOAD 1804 if (ADDED_BY_TOE(sc)) { 1805 struct toedev *tod = sc->sc_tod; 1806 1807 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1808 1809 return (error); 1810 } 1811 #endif 1812 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1813 } 1814 #endif 1815 return (error); 1816 } 1817 1818 /* 1819 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1820 * that exceed the capacity of the syncache by avoiding the storage of any 1821 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1822 * attacks where the attacker does not have access to our responses. 1823 * 1824 * Syncookies encode and include all necessary information about the 1825 * connection setup within the SYN|ACK that we send back. That way we 1826 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1827 * (if ever). Normally the syncache and syncookies are running in parallel 1828 * with the latter taking over when the former is exhausted. When matching 1829 * syncache entry is found the syncookie is ignored. 1830 * 1831 * The only reliable information persisting the 3WHS is our initial sequence 1832 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1833 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1834 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1835 * returns and signifies a legitimate connection if it matches the ACK. 1836 * 1837 * The available space of 32 bits to store the hash and to encode the SYN 1838 * option information is very tight and we should have at least 24 bits for 1839 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1840 * 1841 * SYN option information we have to encode to fully restore a connection: 1842 * MSS: is imporant to chose an optimal segment size to avoid IP level 1843 * fragmentation along the path. The common MSS values can be encoded 1844 * in a 3-bit table. Uncommon values are captured by the next lower value 1845 * in the table leading to a slight increase in packetization overhead. 1846 * WSCALE: is necessary to allow large windows to be used for high delay- 1847 * bandwidth product links. Not scaling the window when it was initially 1848 * negotiated is bad for performance as lack of scaling further decreases 1849 * the apparent available send window. We only need to encode the WSCALE 1850 * we received from the remote end. Our end can be recalculated at any 1851 * time. The common WSCALE values can be encoded in a 3-bit table. 1852 * Uncommon values are captured by the next lower value in the table 1853 * making us under-estimate the available window size halving our 1854 * theoretically possible maximum throughput for that connection. 1855 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1856 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1857 * that are included in all segments on a connection. We enable them when 1858 * the ACK has them. 1859 * 1860 * Security of syncookies and attack vectors: 1861 * 1862 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1863 * together with the gloabl secret to make it unique per connection attempt. 1864 * Thus any change of any of those parameters results in a different MAC output 1865 * in an unpredictable way unless a collision is encountered. 24 bits of the 1866 * MAC are embedded into the ISS. 1867 * 1868 * To prevent replay attacks two rotating global secrets are updated with a 1869 * new random value every 15 seconds. The life-time of a syncookie is thus 1870 * 15-30 seconds. 1871 * 1872 * Vector 1: Attacking the secret. This requires finding a weakness in the 1873 * MAC itself or the way it is used here. The attacker can do a chosen plain 1874 * text attack by varying and testing the all parameters under his control. 1875 * The strength depends on the size and randomness of the secret, and the 1876 * cryptographic security of the MAC function. Due to the constant updating 1877 * of the secret the attacker has at most 29.999 seconds to find the secret 1878 * and launch spoofed connections. After that he has to start all over again. 1879 * 1880 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1881 * size an average of 4,823 attempts are required for a 50% chance of success 1882 * to spoof a single syncookie (birthday collision paradox). However the 1883 * attacker is blind and doesn't know if one of his attempts succeeded unless 1884 * he has a side channel to interfere success from. A single connection setup 1885 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1886 * This many attempts are required for each one blind spoofed connection. For 1887 * every additional spoofed connection he has to launch another N attempts. 1888 * Thus for a sustained rate 100 spoofed connections per second approximately 1889 * 1,800,000 packets per second would have to be sent. 1890 * 1891 * NB: The MAC function should be fast so that it doesn't become a CPU 1892 * exhaustion attack vector itself. 1893 * 1894 * References: 1895 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1896 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1897 * http://cr.yp.to/syncookies.html (overview) 1898 * http://cr.yp.to/syncookies/archive (details) 1899 * 1900 * 1901 * Schematic construction of a syncookie enabled Initial Sequence Number: 1902 * 0 1 2 3 1903 * 12345678901234567890123456789012 1904 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1905 * 1906 * x 24 MAC (truncated) 1907 * W 3 Send Window Scale index 1908 * M 3 MSS index 1909 * S 1 SACK permitted 1910 * P 1 Odd/even secret 1911 */ 1912 1913 /* 1914 * Distribution and probability of certain MSS values. Those in between are 1915 * rounded down to the next lower one. 1916 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1917 * .2% .3% 5% 7% 7% 20% 15% 45% 1918 */ 1919 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1920 1921 /* 1922 * Distribution and probability of certain WSCALE values. We have to map the 1923 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1924 * bits based on prevalence of certain values. Where we don't have an exact 1925 * match for are rounded down to the next lower one letting us under-estimate 1926 * the true available window. At the moment this would happen only for the 1927 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1928 * and window size). The absence of the WSCALE option (no scaling in either 1929 * direction) is encoded with index zero. 1930 * [WSCALE values histograms, Allman, 2012] 1931 * X 10 10 35 5 6 14 10% by host 1932 * X 11 4 5 5 18 49 3% by connections 1933 */ 1934 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1935 1936 /* 1937 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1938 * and good cryptographic properties. 1939 */ 1940 static uint32_t 1941 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1942 uint8_t *secbits, uintptr_t secmod) 1943 { 1944 SIPHASH_CTX ctx; 1945 uint32_t siphash[2]; 1946 1947 SipHash24_Init(&ctx); 1948 SipHash_SetKey(&ctx, secbits); 1949 switch (inc->inc_flags & INC_ISIPV6) { 1950 #ifdef INET 1951 case 0: 1952 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1953 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1954 break; 1955 #endif 1956 #ifdef INET6 1957 case INC_ISIPV6: 1958 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1959 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1960 break; 1961 #endif 1962 } 1963 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1964 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1965 SipHash_Update(&ctx, &irs, sizeof(irs)); 1966 SipHash_Update(&ctx, &flags, sizeof(flags)); 1967 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1968 SipHash_Final((u_int8_t *)&siphash, &ctx); 1969 1970 return (siphash[0] ^ siphash[1]); 1971 } 1972 1973 static tcp_seq 1974 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 1975 { 1976 u_int i, secbit, wscale; 1977 uint32_t iss, hash; 1978 uint8_t *secbits; 1979 union syncookie cookie; 1980 1981 SCH_LOCK_ASSERT(sch); 1982 1983 cookie.cookie = 0; 1984 1985 /* Map our computed MSS into the 3-bit index. */ 1986 for (i = nitems(tcp_sc_msstab) - 1; 1987 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 1988 i--) 1989 ; 1990 cookie.flags.mss_idx = i; 1991 1992 /* 1993 * Map the send window scale into the 3-bit index but only if 1994 * the wscale option was received. 1995 */ 1996 if (sc->sc_flags & SCF_WINSCALE) { 1997 wscale = sc->sc_requested_s_scale; 1998 for (i = nitems(tcp_sc_wstab) - 1; 1999 tcp_sc_wstab[i] > wscale && i > 0; 2000 i--) 2001 ; 2002 cookie.flags.wscale_idx = i; 2003 } 2004 2005 /* Can we do SACK? */ 2006 if (sc->sc_flags & SCF_SACK) 2007 cookie.flags.sack_ok = 1; 2008 2009 /* Which of the two secrets to use. */ 2010 secbit = sch->sch_sc->secret.oddeven & 0x1; 2011 cookie.flags.odd_even = secbit; 2012 2013 secbits = sch->sch_sc->secret.key[secbit]; 2014 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2015 (uintptr_t)sch); 2016 2017 /* 2018 * Put the flags into the hash and XOR them to get better ISS number 2019 * variance. This doesn't enhance the cryptographic strength and is 2020 * done to prevent the 8 cookie bits from showing up directly on the 2021 * wire. 2022 */ 2023 iss = hash & ~0xff; 2024 iss |= cookie.cookie ^ (hash >> 24); 2025 2026 /* Randomize the timestamp. */ 2027 if (sc->sc_flags & SCF_TIMESTAMP) { 2028 sc->sc_ts = arc4random(); 2029 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 2030 } 2031 2032 TCPSTAT_INC(tcps_sc_sendcookie); 2033 return (iss); 2034 } 2035 2036 static struct syncache * 2037 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2038 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2039 struct socket *lso) 2040 { 2041 uint32_t hash; 2042 uint8_t *secbits; 2043 tcp_seq ack, seq; 2044 int wnd, wscale = 0; 2045 union syncookie cookie; 2046 2047 SCH_LOCK_ASSERT(sch); 2048 2049 /* 2050 * Pull information out of SYN-ACK/ACK and revert sequence number 2051 * advances. 2052 */ 2053 ack = th->th_ack - 1; 2054 seq = th->th_seq - 1; 2055 2056 /* 2057 * Unpack the flags containing enough information to restore the 2058 * connection. 2059 */ 2060 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2061 2062 /* Which of the two secrets to use. */ 2063 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 2064 2065 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2066 2067 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2068 if ((ack & ~0xff) != (hash & ~0xff)) 2069 return (NULL); 2070 2071 /* Fill in the syncache values. */ 2072 sc->sc_flags = 0; 2073 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2074 sc->sc_ipopts = NULL; 2075 2076 sc->sc_irs = seq; 2077 sc->sc_iss = ack; 2078 2079 switch (inc->inc_flags & INC_ISIPV6) { 2080 #ifdef INET 2081 case 0: 2082 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2083 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2084 break; 2085 #endif 2086 #ifdef INET6 2087 case INC_ISIPV6: 2088 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2089 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 2090 break; 2091 #endif 2092 } 2093 2094 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2095 2096 /* We can simply recompute receive window scale we sent earlier. */ 2097 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2098 wscale++; 2099 2100 /* Only use wscale if it was enabled in the orignal SYN. */ 2101 if (cookie.flags.wscale_idx > 0) { 2102 sc->sc_requested_r_scale = wscale; 2103 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2104 sc->sc_flags |= SCF_WINSCALE; 2105 } 2106 2107 wnd = lso->sol_sbrcv_hiwat; 2108 wnd = imax(wnd, 0); 2109 wnd = imin(wnd, TCP_MAXWIN); 2110 sc->sc_wnd = wnd; 2111 2112 if (cookie.flags.sack_ok) 2113 sc->sc_flags |= SCF_SACK; 2114 2115 if (to->to_flags & TOF_TS) { 2116 sc->sc_flags |= SCF_TIMESTAMP; 2117 sc->sc_tsreflect = to->to_tsval; 2118 sc->sc_ts = to->to_tsecr; 2119 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 2120 } 2121 2122 if (to->to_flags & TOF_SIGNATURE) 2123 sc->sc_flags |= SCF_SIGNATURE; 2124 2125 sc->sc_rxmits = 0; 2126 2127 TCPSTAT_INC(tcps_sc_recvcookie); 2128 return (sc); 2129 } 2130 2131 #ifdef INVARIANTS 2132 static int 2133 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2134 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2135 struct socket *lso) 2136 { 2137 struct syncache scs, *scx; 2138 char *s; 2139 2140 bzero(&scs, sizeof(scs)); 2141 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2142 2143 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2144 return (0); 2145 2146 if (scx != NULL) { 2147 if (sc->sc_peer_mss != scx->sc_peer_mss) 2148 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2149 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2150 2151 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2152 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2153 s, __func__, sc->sc_requested_r_scale, 2154 scx->sc_requested_r_scale); 2155 2156 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2157 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2158 s, __func__, sc->sc_requested_s_scale, 2159 scx->sc_requested_s_scale); 2160 2161 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2162 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2163 } 2164 2165 if (s != NULL) 2166 free(s, M_TCPLOG); 2167 return (0); 2168 } 2169 #endif /* INVARIANTS */ 2170 2171 static void 2172 syncookie_reseed(void *arg) 2173 { 2174 struct tcp_syncache *sc = arg; 2175 uint8_t *secbits; 2176 int secbit; 2177 2178 /* 2179 * Reseeding the secret doesn't have to be protected by a lock. 2180 * It only must be ensured that the new random values are visible 2181 * to all CPUs in a SMP environment. The atomic with release 2182 * semantics ensures that. 2183 */ 2184 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2185 secbits = sc->secret.key[secbit]; 2186 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2187 atomic_add_rel_int(&sc->secret.oddeven, 1); 2188 2189 /* Reschedule ourself. */ 2190 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2191 } 2192 2193 /* 2194 * Exports the syncache entries to userland so that netstat can display 2195 * them alongside the other sockets. This function is intended to be 2196 * called only from tcp_pcblist. 2197 * 2198 * Due to concurrency on an active system, the number of pcbs exported 2199 * may have no relation to max_pcbs. max_pcbs merely indicates the 2200 * amount of space the caller allocated for this function to use. 2201 */ 2202 int 2203 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2204 { 2205 struct xtcpcb xt; 2206 struct syncache *sc; 2207 struct syncache_head *sch; 2208 int count, error, i; 2209 2210 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2211 sch = &V_tcp_syncache.hashbase[i]; 2212 SCH_LOCK(sch); 2213 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2214 if (count >= max_pcbs) { 2215 SCH_UNLOCK(sch); 2216 goto exit; 2217 } 2218 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2219 continue; 2220 bzero(&xt, sizeof(xt)); 2221 xt.xt_len = sizeof(xt); 2222 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2223 xt.xt_inp.inp_vflag = INP_IPV6; 2224 else 2225 xt.xt_inp.inp_vflag = INP_IPV4; 2226 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2227 sizeof (struct in_conninfo)); 2228 xt.t_state = TCPS_SYN_RECEIVED; 2229 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2230 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2231 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2232 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2233 error = SYSCTL_OUT(req, &xt, sizeof xt); 2234 if (error) { 2235 SCH_UNLOCK(sch); 2236 goto exit; 2237 } 2238 count++; 2239 } 2240 SCH_UNLOCK(sch); 2241 } 2242 exit: 2243 *pcbs_exported = count; 2244 return error; 2245 } 2246