1 /*- 2 * Copyright (c) 2001 McAfee, Inc. 3 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG 4 * All rights reserved. 5 * 6 * This software was developed for the FreeBSD Project by Jonathan Lemon 7 * and McAfee Research, the Security Research Division of McAfee, Inc. under 8 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 9 * DARPA CHATS research program. [2001 McAfee, Inc.] 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 */ 32 33 #include <sys/cdefs.h> 34 __FBSDID("$FreeBSD$"); 35 36 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 #include "opt_pcbgroup.h" 40 41 #include <sys/param.h> 42 #include <sys/systm.h> 43 #include <sys/hash.h> 44 #include <sys/refcount.h> 45 #include <sys/kernel.h> 46 #include <sys/sysctl.h> 47 #include <sys/limits.h> 48 #include <sys/lock.h> 49 #include <sys/mutex.h> 50 #include <sys/malloc.h> 51 #include <sys/mbuf.h> 52 #include <sys/proc.h> /* for proc0 declaration */ 53 #include <sys/random.h> 54 #include <sys/socket.h> 55 #include <sys/socketvar.h> 56 #include <sys/syslog.h> 57 #include <sys/ucred.h> 58 59 #include <sys/md5.h> 60 #include <crypto/siphash/siphash.h> 61 62 #include <vm/uma.h> 63 64 #include <net/if.h> 65 #include <net/if_var.h> 66 #include <net/route.h> 67 #include <net/vnet.h> 68 69 #include <netinet/in.h> 70 #include <netinet/in_systm.h> 71 #include <netinet/ip.h> 72 #include <netinet/in_var.h> 73 #include <netinet/in_pcb.h> 74 #include <netinet/ip_var.h> 75 #include <netinet/ip_options.h> 76 #ifdef INET6 77 #include <netinet/ip6.h> 78 #include <netinet/icmp6.h> 79 #include <netinet6/nd6.h> 80 #include <netinet6/ip6_var.h> 81 #include <netinet6/in6_pcb.h> 82 #endif 83 #include <netinet/tcp.h> 84 #ifdef TCP_RFC7413 85 #include <netinet/tcp_fastopen.h> 86 #endif 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->in6p_laddr = sc->sc_inc.inc6_laddr; 699 } else { 700 inp->inp_vflag &= ~INP_IPV6; 701 inp->inp_vflag |= INP_IPV4; 702 #endif 703 inp->inp_laddr = sc->sc_inc.inc_laddr; 704 #ifdef INET6 705 } 706 #endif 707 708 /* 709 * If there's an mbuf and it has a flowid, then let's initialise the 710 * inp with that particular flowid. 711 */ 712 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 713 inp->inp_flowid = m->m_pkthdr.flowid; 714 inp->inp_flowtype = M_HASHTYPE_GET(m); 715 } 716 717 /* 718 * Install in the reservation hash table for now, but don't yet 719 * install a connection group since the full 4-tuple isn't yet 720 * configured. 721 */ 722 inp->inp_lport = sc->sc_inc.inc_lport; 723 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 724 /* 725 * Undo the assignments above if we failed to 726 * put the PCB on the hash lists. 727 */ 728 #ifdef INET6 729 if (sc->sc_inc.inc_flags & INC_ISIPV6) 730 inp->in6p_laddr = in6addr_any; 731 else 732 #endif 733 inp->inp_laddr.s_addr = INADDR_ANY; 734 inp->inp_lport = 0; 735 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 736 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 737 "with error %i\n", 738 s, __func__, error); 739 free(s, M_TCPLOG); 740 } 741 INP_HASH_WUNLOCK(&V_tcbinfo); 742 goto abort; 743 } 744 #ifdef INET6 745 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 746 struct inpcb *oinp = sotoinpcb(lso); 747 struct in6_addr laddr6; 748 struct sockaddr_in6 sin6; 749 /* 750 * Inherit socket options from the listening socket. 751 * Note that in6p_inputopts are not (and should not be) 752 * copied, since it stores previously received options and is 753 * used to detect if each new option is different than the 754 * previous one and hence should be passed to a user. 755 * If we copied in6p_inputopts, a user would not be able to 756 * receive options just after calling the accept system call. 757 */ 758 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 759 if (oinp->in6p_outputopts) 760 inp->in6p_outputopts = 761 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 762 763 sin6.sin6_family = AF_INET6; 764 sin6.sin6_len = sizeof(sin6); 765 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 766 sin6.sin6_port = sc->sc_inc.inc_fport; 767 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 768 laddr6 = inp->in6p_laddr; 769 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 770 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 771 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 772 thread0.td_ucred, m)) != 0) { 773 inp->in6p_laddr = laddr6; 774 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 775 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 776 "with error %i\n", 777 s, __func__, error); 778 free(s, M_TCPLOG); 779 } 780 INP_HASH_WUNLOCK(&V_tcbinfo); 781 goto abort; 782 } 783 /* Override flowlabel from in6_pcbconnect. */ 784 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 785 inp->inp_flow |= sc->sc_flowlabel; 786 } 787 #endif /* INET6 */ 788 #if defined(INET) && defined(INET6) 789 else 790 #endif 791 #ifdef INET 792 { 793 struct in_addr laddr; 794 struct sockaddr_in sin; 795 796 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 797 798 if (inp->inp_options == NULL) { 799 inp->inp_options = sc->sc_ipopts; 800 sc->sc_ipopts = NULL; 801 } 802 803 sin.sin_family = AF_INET; 804 sin.sin_len = sizeof(sin); 805 sin.sin_addr = sc->sc_inc.inc_faddr; 806 sin.sin_port = sc->sc_inc.inc_fport; 807 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 808 laddr = inp->inp_laddr; 809 if (inp->inp_laddr.s_addr == INADDR_ANY) 810 inp->inp_laddr = sc->sc_inc.inc_laddr; 811 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 812 thread0.td_ucred, m)) != 0) { 813 inp->inp_laddr = laddr; 814 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 815 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 816 "with error %i\n", 817 s, __func__, error); 818 free(s, M_TCPLOG); 819 } 820 INP_HASH_WUNLOCK(&V_tcbinfo); 821 goto abort; 822 } 823 } 824 #endif /* INET */ 825 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 826 /* Copy old policy into new socket's. */ 827 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 828 printf("syncache_socket: could not copy policy\n"); 829 #endif 830 INP_HASH_WUNLOCK(&V_tcbinfo); 831 tp = intotcpcb(inp); 832 tcp_state_change(tp, TCPS_SYN_RECEIVED); 833 tp->iss = sc->sc_iss; 834 tp->irs = sc->sc_irs; 835 tcp_rcvseqinit(tp); 836 tcp_sendseqinit(tp); 837 blk = sototcpcb(lso)->t_fb; 838 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 839 /* 840 * Our parents t_fb was not the default, 841 * we need to release our ref on tp->t_fb and 842 * pickup one on the new entry. 843 */ 844 struct tcp_function_block *rblk; 845 846 rblk = find_and_ref_tcp_fb(blk); 847 KASSERT(rblk != NULL, 848 ("cannot find blk %p out of syncache?", blk)); 849 if (tp->t_fb->tfb_tcp_fb_fini) 850 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 851 refcount_release(&tp->t_fb->tfb_refcnt); 852 tp->t_fb = rblk; 853 if (tp->t_fb->tfb_tcp_fb_init) { 854 (*tp->t_fb->tfb_tcp_fb_init)(tp); 855 } 856 } 857 tp->snd_wl1 = sc->sc_irs; 858 tp->snd_max = tp->iss + 1; 859 tp->snd_nxt = tp->iss + 1; 860 tp->rcv_up = sc->sc_irs + 1; 861 tp->rcv_wnd = sc->sc_wnd; 862 tp->rcv_adv += tp->rcv_wnd; 863 tp->last_ack_sent = tp->rcv_nxt; 864 865 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 866 if (sc->sc_flags & SCF_NOOPT) 867 tp->t_flags |= TF_NOOPT; 868 else { 869 if (sc->sc_flags & SCF_WINSCALE) { 870 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 871 tp->snd_scale = sc->sc_requested_s_scale; 872 tp->request_r_scale = sc->sc_requested_r_scale; 873 } 874 if (sc->sc_flags & SCF_TIMESTAMP) { 875 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 876 tp->ts_recent = sc->sc_tsreflect; 877 tp->ts_recent_age = tcp_ts_getticks(); 878 tp->ts_offset = sc->sc_tsoff; 879 } 880 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 881 if (sc->sc_flags & SCF_SIGNATURE) 882 tp->t_flags |= TF_SIGNATURE; 883 #endif 884 if (sc->sc_flags & SCF_SACK) 885 tp->t_flags |= TF_SACK_PERMIT; 886 } 887 888 if (sc->sc_flags & SCF_ECN) 889 tp->t_flags |= TF_ECN_PERMIT; 890 891 /* 892 * Set up MSS and get cached values from tcp_hostcache. 893 * This might overwrite some of the defaults we just set. 894 */ 895 tcp_mss(tp, sc->sc_peer_mss); 896 897 /* 898 * If the SYN,ACK was retransmitted, indicate that CWND to be 899 * limited to one segment in cc_conn_init(). 900 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 901 */ 902 if (sc->sc_rxmits > 1) 903 tp->snd_cwnd = 1; 904 905 #ifdef TCP_OFFLOAD 906 /* 907 * Allow a TOE driver to install its hooks. Note that we hold the 908 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 909 * new connection before the TOE driver has done its thing. 910 */ 911 if (ADDED_BY_TOE(sc)) { 912 struct toedev *tod = sc->sc_tod; 913 914 tod->tod_offload_socket(tod, sc->sc_todctx, so); 915 } 916 #endif 917 /* 918 * Copy and activate timers. 919 */ 920 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 921 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 922 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 923 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 924 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 925 926 TCPSTAT_INC(tcps_accepts); 927 return (so); 928 929 abort: 930 INP_WUNLOCK(inp); 931 abort2: 932 if (so != NULL) 933 soabort(so); 934 return (NULL); 935 } 936 937 /* 938 * This function gets called when we receive an ACK for a 939 * socket in the LISTEN state. We look up the connection 940 * in the syncache, and if its there, we pull it out of 941 * the cache and turn it into a full-blown connection in 942 * the SYN-RECEIVED state. 943 * 944 * On syncache_socket() success the newly created socket 945 * has its underlying inp locked. 946 */ 947 int 948 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 949 struct socket **lsop, struct mbuf *m) 950 { 951 struct syncache *sc; 952 struct syncache_head *sch; 953 struct syncache scs; 954 char *s; 955 956 /* 957 * Global TCP locks are held because we manipulate the PCB lists 958 * and create a new socket. 959 */ 960 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 961 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 962 ("%s: can handle only ACK", __func__)); 963 964 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 965 SCH_LOCK_ASSERT(sch); 966 967 #ifdef INVARIANTS 968 /* 969 * Test code for syncookies comparing the syncache stored 970 * values with the reconstructed values from the cookie. 971 */ 972 if (sc != NULL) 973 syncookie_cmp(inc, sch, sc, th, to, *lsop); 974 #endif 975 976 if (sc == NULL) { 977 /* 978 * There is no syncache entry, so see if this ACK is 979 * a returning syncookie. To do this, first: 980 * A. Check if syncookies are used in case of syncache 981 * overflows 982 * B. See if this socket has had a syncache entry dropped in 983 * the recent past. We don't want to accept a bogus 984 * syncookie if we've never received a SYN or accept it 985 * twice. 986 * C. check that the syncookie is valid. If it is, then 987 * cobble up a fake syncache entry, and return. 988 */ 989 if (!V_tcp_syncookies) { 990 SCH_UNLOCK(sch); 991 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 992 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 993 "segment rejected (syncookies disabled)\n", 994 s, __func__); 995 goto failed; 996 } 997 if (!V_tcp_syncookiesonly && 998 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 999 SCH_UNLOCK(sch); 1000 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1001 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1002 "segment rejected (no syncache entry)\n", 1003 s, __func__); 1004 goto failed; 1005 } 1006 bzero(&scs, sizeof(scs)); 1007 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1008 SCH_UNLOCK(sch); 1009 if (sc == NULL) { 1010 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1011 log(LOG_DEBUG, "%s; %s: Segment failed " 1012 "SYNCOOKIE authentication, segment rejected " 1013 "(probably spoofed)\n", s, __func__); 1014 goto failed; 1015 } 1016 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1017 /* If received ACK has MD5 signature, check it. */ 1018 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1019 (!TCPMD5_ENABLED() || 1020 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1021 /* Drop the ACK. */ 1022 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1023 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1024 "MD5 signature doesn't match.\n", 1025 s, __func__); 1026 free(s, M_TCPLOG); 1027 } 1028 TCPSTAT_INC(tcps_sig_err_sigopt); 1029 return (-1); /* Do not send RST */ 1030 } 1031 #endif /* TCP_SIGNATURE */ 1032 } else { 1033 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1034 /* 1035 * If listening socket requested TCP digests, check that 1036 * received ACK has signature and it is correct. 1037 * If not, drop the ACK and leave sc entry in th cache, 1038 * because SYN was received with correct signature. 1039 */ 1040 if (sc->sc_flags & SCF_SIGNATURE) { 1041 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1042 /* No signature */ 1043 TCPSTAT_INC(tcps_sig_err_nosigopt); 1044 SCH_UNLOCK(sch); 1045 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1046 log(LOG_DEBUG, "%s; %s: Segment " 1047 "rejected, MD5 signature wasn't " 1048 "provided.\n", s, __func__); 1049 free(s, M_TCPLOG); 1050 } 1051 return (-1); /* Do not send RST */ 1052 } 1053 if (!TCPMD5_ENABLED() || 1054 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1055 /* Doesn't match or no SA */ 1056 SCH_UNLOCK(sch); 1057 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1058 log(LOG_DEBUG, "%s; %s: Segment " 1059 "rejected, MD5 signature doesn't " 1060 "match.\n", s, __func__); 1061 free(s, M_TCPLOG); 1062 } 1063 return (-1); /* Do not send RST */ 1064 } 1065 } 1066 #endif /* TCP_SIGNATURE */ 1067 /* 1068 * Pull out the entry to unlock the bucket row. 1069 * 1070 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1071 * tcp_state_change(). The tcpcb is not existent at this 1072 * moment. A new one will be allocated via syncache_socket-> 1073 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1074 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1075 */ 1076 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1077 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1078 sch->sch_length--; 1079 #ifdef TCP_OFFLOAD 1080 if (ADDED_BY_TOE(sc)) { 1081 struct toedev *tod = sc->sc_tod; 1082 1083 tod->tod_syncache_removed(tod, sc->sc_todctx); 1084 } 1085 #endif 1086 SCH_UNLOCK(sch); 1087 } 1088 1089 /* 1090 * Segment validation: 1091 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1092 */ 1093 if (th->th_ack != sc->sc_iss + 1) { 1094 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1095 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1096 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1097 goto failed; 1098 } 1099 1100 /* 1101 * The SEQ must fall in the window starting at the received 1102 * initial receive sequence number + 1 (the SYN). 1103 */ 1104 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1105 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1106 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1107 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1108 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1109 goto failed; 1110 } 1111 1112 /* 1113 * If timestamps were not negotiated during SYN/ACK they 1114 * must not appear on any segment during this session. 1115 */ 1116 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1117 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1118 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1119 "segment rejected\n", s, __func__); 1120 goto failed; 1121 } 1122 1123 /* 1124 * If timestamps were negotiated during SYN/ACK they should 1125 * appear on every segment during this session. 1126 * XXXAO: This is only informal as there have been unverified 1127 * reports of non-compliants stacks. 1128 */ 1129 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1130 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1131 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1132 "no action\n", s, __func__); 1133 free(s, M_TCPLOG); 1134 s = NULL; 1135 } 1136 } 1137 1138 /* 1139 * If timestamps were negotiated, the reflected timestamp 1140 * must be equal to what we actually sent in the SYN|ACK 1141 * except in the case of 0. Some boxes are known for sending 1142 * broken timestamp replies during the 3whs (and potentially 1143 * during the connection also). 1144 * 1145 * Accept the final ACK of 3whs with reflected timestamp of 0 1146 * instead of sending a RST and deleting the syncache entry. 1147 */ 1148 if ((to->to_flags & TOF_TS) && to->to_tsecr && 1149 to->to_tsecr != sc->sc_ts) { 1150 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1151 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1152 "segment rejected\n", 1153 s, __func__, to->to_tsecr, sc->sc_ts); 1154 goto failed; 1155 } 1156 1157 *lsop = syncache_socket(sc, *lsop, m); 1158 1159 if (*lsop == NULL) 1160 TCPSTAT_INC(tcps_sc_aborted); 1161 else 1162 TCPSTAT_INC(tcps_sc_completed); 1163 1164 /* how do we find the inp for the new socket? */ 1165 if (sc != &scs) 1166 syncache_free(sc); 1167 return (1); 1168 failed: 1169 if (sc != NULL && sc != &scs) 1170 syncache_free(sc); 1171 if (s != NULL) 1172 free(s, M_TCPLOG); 1173 *lsop = NULL; 1174 return (0); 1175 } 1176 1177 #ifdef TCP_RFC7413 1178 static void 1179 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1180 uint64_t response_cookie) 1181 { 1182 struct inpcb *inp; 1183 struct tcpcb *tp; 1184 unsigned int *pending_counter; 1185 1186 /* 1187 * Global TCP locks are held because we manipulate the PCB lists 1188 * and create a new socket. 1189 */ 1190 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1191 1192 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1193 *lsop = syncache_socket(sc, *lsop, m); 1194 if (*lsop == NULL) { 1195 TCPSTAT_INC(tcps_sc_aborted); 1196 atomic_subtract_int(pending_counter, 1); 1197 } else { 1198 inp = sotoinpcb(*lsop); 1199 tp = intotcpcb(inp); 1200 tp->t_flags |= TF_FASTOPEN; 1201 tp->t_tfo_cookie = response_cookie; 1202 tp->snd_max = tp->iss; 1203 tp->snd_nxt = tp->iss; 1204 tp->t_tfo_pending = pending_counter; 1205 TCPSTAT_INC(tcps_sc_completed); 1206 } 1207 } 1208 #endif /* TCP_RFC7413 */ 1209 1210 /* 1211 * Given a LISTEN socket and an inbound SYN request, add 1212 * this to the syn cache, and send back a segment: 1213 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1214 * to the source. 1215 * 1216 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1217 * Doing so would require that we hold onto the data and deliver it 1218 * to the application. However, if we are the target of a SYN-flood 1219 * DoS attack, an attacker could send data which would eventually 1220 * consume all available buffer space if it were ACKed. By not ACKing 1221 * the data, we avoid this DoS scenario. 1222 * 1223 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1224 * cookie is processed and a new socket is created. In this case, any data 1225 * accompanying the SYN will be queued to the socket by tcp_input() and will 1226 * be ACKed either when the application sends response data or the delayed 1227 * ACK timer expires, whichever comes first. 1228 */ 1229 int 1230 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1231 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1232 void *todctx) 1233 { 1234 struct tcpcb *tp; 1235 struct socket *so; 1236 struct syncache *sc = NULL; 1237 struct syncache_head *sch; 1238 struct mbuf *ipopts = NULL; 1239 u_int ltflags; 1240 int win, ip_ttl, ip_tos; 1241 char *s; 1242 int rv = 0; 1243 #ifdef INET6 1244 int autoflowlabel = 0; 1245 #endif 1246 #ifdef MAC 1247 struct label *maclabel; 1248 #endif 1249 struct syncache scs; 1250 struct ucred *cred; 1251 #ifdef TCP_RFC7413 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 #endif 1257 1258 INP_WLOCK_ASSERT(inp); /* listen socket */ 1259 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1260 ("%s: unexpected tcp flags", __func__)); 1261 1262 /* 1263 * Combine all so/tp operations very early to drop the INP lock as 1264 * soon as possible. 1265 */ 1266 so = *lsop; 1267 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1268 tp = sototcpcb(so); 1269 cred = crhold(so->so_cred); 1270 1271 #ifdef INET6 1272 if ((inc->inc_flags & INC_ISIPV6) && 1273 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1274 autoflowlabel = 1; 1275 #endif 1276 ip_ttl = inp->inp_ip_ttl; 1277 ip_tos = inp->inp_ip_tos; 1278 win = so->sol_sbrcv_hiwat; 1279 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1280 1281 #ifdef TCP_RFC7413 1282 if (V_tcp_fastopen_enabled && IS_FASTOPEN(tp->t_flags) && 1283 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) { 1284 /* 1285 * Limit the number of pending TFO connections to 1286 * approximately half of the queue limit. This prevents TFO 1287 * SYN floods from starving the service by filling the 1288 * listen queue with bogus TFO connections. 1289 */ 1290 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1291 (so->sol_qlimit / 2)) { 1292 int result; 1293 1294 result = tcp_fastopen_check_cookie(inc, 1295 to->to_tfo_cookie, to->to_tfo_len, 1296 &tfo_response_cookie); 1297 tfo_cookie_valid = (result > 0); 1298 tfo_response_cookie_valid = (result >= 0); 1299 } 1300 1301 /* 1302 * Remember the TFO pending counter as it will have to be 1303 * decremented below if we don't make it to syncache_tfo_expand(). 1304 */ 1305 tfo_pending = tp->t_tfo_pending; 1306 } 1307 #endif 1308 1309 /* By the time we drop the lock these should no longer be used. */ 1310 so = NULL; 1311 tp = NULL; 1312 1313 #ifdef MAC 1314 if (mac_syncache_init(&maclabel) != 0) { 1315 INP_WUNLOCK(inp); 1316 goto done; 1317 } else 1318 mac_syncache_create(maclabel, inp); 1319 #endif 1320 #ifdef TCP_RFC7413 1321 if (!tfo_cookie_valid) 1322 #endif 1323 INP_WUNLOCK(inp); 1324 1325 /* 1326 * Remember the IP options, if any. 1327 */ 1328 #ifdef INET6 1329 if (!(inc->inc_flags & INC_ISIPV6)) 1330 #endif 1331 #ifdef INET 1332 ipopts = (m) ? ip_srcroute(m) : NULL; 1333 #else 1334 ipopts = NULL; 1335 #endif 1336 1337 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1338 /* 1339 * If listening socket requested TCP digests, check that received 1340 * SYN has signature and it is correct. If signature doesn't match 1341 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1342 */ 1343 if (ltflags & TF_SIGNATURE) { 1344 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1345 TCPSTAT_INC(tcps_sig_err_nosigopt); 1346 goto done; 1347 } 1348 if (!TCPMD5_ENABLED() || 1349 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1350 goto done; 1351 } 1352 #endif /* TCP_SIGNATURE */ 1353 /* 1354 * See if we already have an entry for this connection. 1355 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1356 * 1357 * XXX: should the syncache be re-initialized with the contents 1358 * of the new SYN here (which may have different options?) 1359 * 1360 * XXX: We do not check the sequence number to see if this is a 1361 * real retransmit or a new connection attempt. The question is 1362 * how to handle such a case; either ignore it as spoofed, or 1363 * drop the current entry and create a new one? 1364 */ 1365 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1366 SCH_LOCK_ASSERT(sch); 1367 if (sc != NULL) { 1368 #ifdef TCP_RFC7413 1369 if (tfo_cookie_valid) 1370 INP_WUNLOCK(inp); 1371 #endif 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 #ifdef TCP_RFC7413 1415 if (tfo_cookie_valid) { 1416 bzero(&scs, sizeof(scs)); 1417 sc = &scs; 1418 goto skip_alloc; 1419 } 1420 #endif 1421 1422 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1423 if (sc == NULL) { 1424 /* 1425 * The zone allocator couldn't provide more entries. 1426 * Treat this as if the cache was full; drop the oldest 1427 * entry and insert the new one. 1428 */ 1429 TCPSTAT_INC(tcps_sc_zonefail); 1430 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1431 sch->sch_last_overflow = time_uptime; 1432 syncache_drop(sc, sch); 1433 } 1434 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1435 if (sc == NULL) { 1436 if (V_tcp_syncookies) { 1437 bzero(&scs, sizeof(scs)); 1438 sc = &scs; 1439 } else { 1440 SCH_UNLOCK(sch); 1441 if (ipopts) 1442 (void) m_free(ipopts); 1443 goto done; 1444 } 1445 } 1446 } 1447 1448 #ifdef TCP_RFC7413 1449 skip_alloc: 1450 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1451 sc->sc_tfo_cookie = &tfo_response_cookie; 1452 #endif 1453 1454 /* 1455 * Fill in the syncache values. 1456 */ 1457 #ifdef MAC 1458 sc->sc_label = maclabel; 1459 #endif 1460 sc->sc_cred = cred; 1461 cred = NULL; 1462 sc->sc_ipopts = ipopts; 1463 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1464 #ifdef INET6 1465 if (!(inc->inc_flags & INC_ISIPV6)) 1466 #endif 1467 { 1468 sc->sc_ip_tos = ip_tos; 1469 sc->sc_ip_ttl = ip_ttl; 1470 } 1471 #ifdef TCP_OFFLOAD 1472 sc->sc_tod = tod; 1473 sc->sc_todctx = todctx; 1474 #endif 1475 sc->sc_irs = th->th_seq; 1476 sc->sc_iss = arc4random(); 1477 sc->sc_flags = 0; 1478 sc->sc_flowlabel = 0; 1479 1480 /* 1481 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1482 * win was derived from socket earlier in the function. 1483 */ 1484 win = imax(win, 0); 1485 win = imin(win, TCP_MAXWIN); 1486 sc->sc_wnd = win; 1487 1488 if (V_tcp_do_rfc1323) { 1489 /* 1490 * A timestamp received in a SYN makes 1491 * it ok to send timestamp requests and replies. 1492 */ 1493 if (to->to_flags & TOF_TS) { 1494 sc->sc_tsreflect = to->to_tsval; 1495 sc->sc_ts = tcp_ts_getticks(); 1496 sc->sc_flags |= SCF_TIMESTAMP; 1497 } 1498 if (to->to_flags & TOF_SCALE) { 1499 int wscale = 0; 1500 1501 /* 1502 * Pick the smallest possible scaling factor that 1503 * will still allow us to scale up to sb_max, aka 1504 * kern.ipc.maxsockbuf. 1505 * 1506 * We do this because there are broken firewalls that 1507 * will corrupt the window scale option, leading to 1508 * the other endpoint believing that our advertised 1509 * window is unscaled. At scale factors larger than 1510 * 5 the unscaled window will drop below 1500 bytes, 1511 * leading to serious problems when traversing these 1512 * broken firewalls. 1513 * 1514 * With the default maxsockbuf of 256K, a scale factor 1515 * of 3 will be chosen by this algorithm. Those who 1516 * choose a larger maxsockbuf should watch out 1517 * for the compatibility problems mentioned above. 1518 * 1519 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1520 * or <SYN,ACK>) segment itself is never scaled. 1521 */ 1522 while (wscale < TCP_MAX_WINSHIFT && 1523 (TCP_MAXWIN << wscale) < sb_max) 1524 wscale++; 1525 sc->sc_requested_r_scale = wscale; 1526 sc->sc_requested_s_scale = to->to_wscale; 1527 sc->sc_flags |= SCF_WINSCALE; 1528 } 1529 } 1530 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1531 /* 1532 * If listening socket requested TCP digests, flag this in the 1533 * syncache so that syncache_respond() will do the right thing 1534 * with the SYN+ACK. 1535 */ 1536 if (ltflags & TF_SIGNATURE) 1537 sc->sc_flags |= SCF_SIGNATURE; 1538 #endif /* TCP_SIGNATURE */ 1539 if (to->to_flags & TOF_SACKPERM) 1540 sc->sc_flags |= SCF_SACK; 1541 if (to->to_flags & TOF_MSS) 1542 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1543 if (ltflags & TF_NOOPT) 1544 sc->sc_flags |= SCF_NOOPT; 1545 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1546 sc->sc_flags |= SCF_ECN; 1547 1548 if (V_tcp_syncookies) 1549 sc->sc_iss = syncookie_generate(sch, sc); 1550 #ifdef INET6 1551 if (autoflowlabel) { 1552 if (V_tcp_syncookies) 1553 sc->sc_flowlabel = sc->sc_iss; 1554 else 1555 sc->sc_flowlabel = ip6_randomflowlabel(); 1556 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1557 } 1558 #endif 1559 SCH_UNLOCK(sch); 1560 1561 #ifdef TCP_RFC7413 1562 if (tfo_cookie_valid) { 1563 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1564 /* INP_WUNLOCK(inp) will be performed by the caller */ 1565 rv = 1; 1566 goto tfo_expanded; 1567 } 1568 #endif 1569 1570 /* 1571 * Do a standard 3-way handshake. 1572 */ 1573 if (syncache_respond(sc, sch, 0, m) == 0) { 1574 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1575 syncache_free(sc); 1576 else if (sc != &scs) 1577 syncache_insert(sc, sch); /* locks and unlocks sch */ 1578 TCPSTAT_INC(tcps_sndacks); 1579 TCPSTAT_INC(tcps_sndtotal); 1580 } else { 1581 if (sc != &scs) 1582 syncache_free(sc); 1583 TCPSTAT_INC(tcps_sc_dropped); 1584 } 1585 1586 done: 1587 if (m) { 1588 *lsop = NULL; 1589 m_freem(m); 1590 } 1591 #ifdef TCP_RFC7413 1592 /* 1593 * If tfo_pending is not NULL here, then a TFO SYN that did not 1594 * result in a new socket was processed and the associated pending 1595 * counter has not yet been decremented. All such TFO processing paths 1596 * transit this point. 1597 */ 1598 if (tfo_pending != NULL) 1599 tcp_fastopen_decrement_counter(tfo_pending); 1600 1601 tfo_expanded: 1602 #endif 1603 if (cred != NULL) 1604 crfree(cred); 1605 #ifdef MAC 1606 if (sc == &scs) 1607 mac_syncache_destroy(&maclabel); 1608 #endif 1609 return (rv); 1610 } 1611 1612 /* 1613 * Send SYN|ACK to the peer. Either in response to the peer's SYN, 1614 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1615 */ 1616 static int 1617 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked, 1618 const struct mbuf *m0) 1619 { 1620 struct ip *ip = NULL; 1621 struct mbuf *m; 1622 struct tcphdr *th = NULL; 1623 int optlen, error = 0; /* Make compiler happy */ 1624 u_int16_t hlen, tlen, mssopt; 1625 struct tcpopt to; 1626 #ifdef INET6 1627 struct ip6_hdr *ip6 = NULL; 1628 #endif 1629 hlen = 1630 #ifdef INET6 1631 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1632 #endif 1633 sizeof(struct ip); 1634 tlen = hlen + sizeof(struct tcphdr); 1635 1636 /* Determine MSS we advertize to other end of connection. */ 1637 mssopt = tcp_mssopt(&sc->sc_inc); 1638 if (sc->sc_peer_mss) 1639 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss); 1640 1641 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1642 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1643 ("syncache: mbuf too small")); 1644 1645 /* Create the IP+TCP header from scratch. */ 1646 m = m_gethdr(M_NOWAIT, MT_DATA); 1647 if (m == NULL) 1648 return (ENOBUFS); 1649 #ifdef MAC 1650 mac_syncache_create_mbuf(sc->sc_label, m); 1651 #endif 1652 m->m_data += max_linkhdr; 1653 m->m_len = tlen; 1654 m->m_pkthdr.len = tlen; 1655 m->m_pkthdr.rcvif = NULL; 1656 1657 #ifdef INET6 1658 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1659 ip6 = mtod(m, struct ip6_hdr *); 1660 ip6->ip6_vfc = IPV6_VERSION; 1661 ip6->ip6_nxt = IPPROTO_TCP; 1662 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1663 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1664 ip6->ip6_plen = htons(tlen - hlen); 1665 /* ip6_hlim is set after checksum */ 1666 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1667 ip6->ip6_flow |= sc->sc_flowlabel; 1668 1669 th = (struct tcphdr *)(ip6 + 1); 1670 } 1671 #endif 1672 #if defined(INET6) && defined(INET) 1673 else 1674 #endif 1675 #ifdef INET 1676 { 1677 ip = mtod(m, struct ip *); 1678 ip->ip_v = IPVERSION; 1679 ip->ip_hl = sizeof(struct ip) >> 2; 1680 ip->ip_len = htons(tlen); 1681 ip->ip_id = 0; 1682 ip->ip_off = 0; 1683 ip->ip_sum = 0; 1684 ip->ip_p = IPPROTO_TCP; 1685 ip->ip_src = sc->sc_inc.inc_laddr; 1686 ip->ip_dst = sc->sc_inc.inc_faddr; 1687 ip->ip_ttl = sc->sc_ip_ttl; 1688 ip->ip_tos = sc->sc_ip_tos; 1689 1690 /* 1691 * See if we should do MTU discovery. Route lookups are 1692 * expensive, so we will only unset the DF bit if: 1693 * 1694 * 1) path_mtu_discovery is disabled 1695 * 2) the SCF_UNREACH flag has been set 1696 */ 1697 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1698 ip->ip_off |= htons(IP_DF); 1699 1700 th = (struct tcphdr *)(ip + 1); 1701 } 1702 #endif /* INET */ 1703 th->th_sport = sc->sc_inc.inc_lport; 1704 th->th_dport = sc->sc_inc.inc_fport; 1705 1706 th->th_seq = htonl(sc->sc_iss); 1707 th->th_ack = htonl(sc->sc_irs + 1); 1708 th->th_off = sizeof(struct tcphdr) >> 2; 1709 th->th_x2 = 0; 1710 th->th_flags = TH_SYN|TH_ACK; 1711 th->th_win = htons(sc->sc_wnd); 1712 th->th_urp = 0; 1713 1714 if (sc->sc_flags & SCF_ECN) { 1715 th->th_flags |= TH_ECE; 1716 TCPSTAT_INC(tcps_ecn_shs); 1717 } 1718 1719 /* Tack on the TCP options. */ 1720 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1721 to.to_flags = 0; 1722 1723 to.to_mss = mssopt; 1724 to.to_flags = TOF_MSS; 1725 if (sc->sc_flags & SCF_WINSCALE) { 1726 to.to_wscale = sc->sc_requested_r_scale; 1727 to.to_flags |= TOF_SCALE; 1728 } 1729 if (sc->sc_flags & SCF_TIMESTAMP) { 1730 /* Virgin timestamp or TCP cookie enhanced one. */ 1731 to.to_tsval = sc->sc_ts; 1732 to.to_tsecr = sc->sc_tsreflect; 1733 to.to_flags |= TOF_TS; 1734 } 1735 if (sc->sc_flags & SCF_SACK) 1736 to.to_flags |= TOF_SACKPERM; 1737 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1738 if (sc->sc_flags & SCF_SIGNATURE) 1739 to.to_flags |= TOF_SIGNATURE; 1740 #endif 1741 #ifdef TCP_RFC7413 1742 if (sc->sc_tfo_cookie) { 1743 to.to_flags |= TOF_FASTOPEN; 1744 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1745 to.to_tfo_cookie = sc->sc_tfo_cookie; 1746 /* don't send cookie again when retransmitting response */ 1747 sc->sc_tfo_cookie = NULL; 1748 } 1749 #endif 1750 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1751 1752 /* Adjust headers by option size. */ 1753 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1754 m->m_len += optlen; 1755 m->m_pkthdr.len += optlen; 1756 #ifdef INET6 1757 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1758 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1759 else 1760 #endif 1761 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1762 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1763 if (sc->sc_flags & SCF_SIGNATURE) { 1764 KASSERT(to.to_flags & TOF_SIGNATURE, 1765 ("tcp_addoptions() didn't set tcp_signature")); 1766 1767 /* NOTE: to.to_signature is inside of mbuf */ 1768 if (!TCPMD5_ENABLED() || 1769 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1770 m_freem(m); 1771 return (EACCES); 1772 } 1773 } 1774 #endif 1775 } else 1776 optlen = 0; 1777 1778 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1779 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1780 /* 1781 * If we have peer's SYN and it has a flowid, then let's assign it to 1782 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1783 * to SYN|ACK due to lack of inp here. 1784 */ 1785 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1786 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1787 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1788 } 1789 #ifdef INET6 1790 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1791 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1792 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1793 IPPROTO_TCP, 0); 1794 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1795 #ifdef TCP_OFFLOAD 1796 if (ADDED_BY_TOE(sc)) { 1797 struct toedev *tod = sc->sc_tod; 1798 1799 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1800 1801 return (error); 1802 } 1803 #endif 1804 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1805 } 1806 #endif 1807 #if defined(INET6) && defined(INET) 1808 else 1809 #endif 1810 #ifdef INET 1811 { 1812 m->m_pkthdr.csum_flags = CSUM_TCP; 1813 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1814 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1815 #ifdef TCP_OFFLOAD 1816 if (ADDED_BY_TOE(sc)) { 1817 struct toedev *tod = sc->sc_tod; 1818 1819 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1820 1821 return (error); 1822 } 1823 #endif 1824 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1825 } 1826 #endif 1827 return (error); 1828 } 1829 1830 /* 1831 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1832 * that exceed the capacity of the syncache by avoiding the storage of any 1833 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1834 * attacks where the attacker does not have access to our responses. 1835 * 1836 * Syncookies encode and include all necessary information about the 1837 * connection setup within the SYN|ACK that we send back. That way we 1838 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1839 * (if ever). Normally the syncache and syncookies are running in parallel 1840 * with the latter taking over when the former is exhausted. When matching 1841 * syncache entry is found the syncookie is ignored. 1842 * 1843 * The only reliable information persisting the 3WHS is our initial sequence 1844 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1845 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1846 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1847 * returns and signifies a legitimate connection if it matches the ACK. 1848 * 1849 * The available space of 32 bits to store the hash and to encode the SYN 1850 * option information is very tight and we should have at least 24 bits for 1851 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1852 * 1853 * SYN option information we have to encode to fully restore a connection: 1854 * MSS: is imporant to chose an optimal segment size to avoid IP level 1855 * fragmentation along the path. The common MSS values can be encoded 1856 * in a 3-bit table. Uncommon values are captured by the next lower value 1857 * in the table leading to a slight increase in packetization overhead. 1858 * WSCALE: is necessary to allow large windows to be used for high delay- 1859 * bandwidth product links. Not scaling the window when it was initially 1860 * negotiated is bad for performance as lack of scaling further decreases 1861 * the apparent available send window. We only need to encode the WSCALE 1862 * we received from the remote end. Our end can be recalculated at any 1863 * time. The common WSCALE values can be encoded in a 3-bit table. 1864 * Uncommon values are captured by the next lower value in the table 1865 * making us under-estimate the available window size halving our 1866 * theoretically possible maximum throughput for that connection. 1867 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1868 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1869 * that are included in all segments on a connection. We enable them when 1870 * the ACK has them. 1871 * 1872 * Security of syncookies and attack vectors: 1873 * 1874 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1875 * together with the gloabl secret to make it unique per connection attempt. 1876 * Thus any change of any of those parameters results in a different MAC output 1877 * in an unpredictable way unless a collision is encountered. 24 bits of the 1878 * MAC are embedded into the ISS. 1879 * 1880 * To prevent replay attacks two rotating global secrets are updated with a 1881 * new random value every 15 seconds. The life-time of a syncookie is thus 1882 * 15-30 seconds. 1883 * 1884 * Vector 1: Attacking the secret. This requires finding a weakness in the 1885 * MAC itself or the way it is used here. The attacker can do a chosen plain 1886 * text attack by varying and testing the all parameters under his control. 1887 * The strength depends on the size and randomness of the secret, and the 1888 * cryptographic security of the MAC function. Due to the constant updating 1889 * of the secret the attacker has at most 29.999 seconds to find the secret 1890 * and launch spoofed connections. After that he has to start all over again. 1891 * 1892 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1893 * size an average of 4,823 attempts are required for a 50% chance of success 1894 * to spoof a single syncookie (birthday collision paradox). However the 1895 * attacker is blind and doesn't know if one of his attempts succeeded unless 1896 * he has a side channel to interfere success from. A single connection setup 1897 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1898 * This many attempts are required for each one blind spoofed connection. For 1899 * every additional spoofed connection he has to launch another N attempts. 1900 * Thus for a sustained rate 100 spoofed connections per second approximately 1901 * 1,800,000 packets per second would have to be sent. 1902 * 1903 * NB: The MAC function should be fast so that it doesn't become a CPU 1904 * exhaustion attack vector itself. 1905 * 1906 * References: 1907 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1908 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1909 * http://cr.yp.to/syncookies.html (overview) 1910 * http://cr.yp.to/syncookies/archive (details) 1911 * 1912 * 1913 * Schematic construction of a syncookie enabled Initial Sequence Number: 1914 * 0 1 2 3 1915 * 12345678901234567890123456789012 1916 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1917 * 1918 * x 24 MAC (truncated) 1919 * W 3 Send Window Scale index 1920 * M 3 MSS index 1921 * S 1 SACK permitted 1922 * P 1 Odd/even secret 1923 */ 1924 1925 /* 1926 * Distribution and probability of certain MSS values. Those in between are 1927 * rounded down to the next lower one. 1928 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1929 * .2% .3% 5% 7% 7% 20% 15% 45% 1930 */ 1931 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1932 1933 /* 1934 * Distribution and probability of certain WSCALE values. We have to map the 1935 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1936 * bits based on prevalence of certain values. Where we don't have an exact 1937 * match for are rounded down to the next lower one letting us under-estimate 1938 * the true available window. At the moment this would happen only for the 1939 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1940 * and window size). The absence of the WSCALE option (no scaling in either 1941 * direction) is encoded with index zero. 1942 * [WSCALE values histograms, Allman, 2012] 1943 * X 10 10 35 5 6 14 10% by host 1944 * X 11 4 5 5 18 49 3% by connections 1945 */ 1946 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1947 1948 /* 1949 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1950 * and good cryptographic properties. 1951 */ 1952 static uint32_t 1953 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1954 uint8_t *secbits, uintptr_t secmod) 1955 { 1956 SIPHASH_CTX ctx; 1957 uint32_t siphash[2]; 1958 1959 SipHash24_Init(&ctx); 1960 SipHash_SetKey(&ctx, secbits); 1961 switch (inc->inc_flags & INC_ISIPV6) { 1962 #ifdef INET 1963 case 0: 1964 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1965 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1966 break; 1967 #endif 1968 #ifdef INET6 1969 case INC_ISIPV6: 1970 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1971 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1972 break; 1973 #endif 1974 } 1975 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1976 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1977 SipHash_Update(&ctx, &irs, sizeof(irs)); 1978 SipHash_Update(&ctx, &flags, sizeof(flags)); 1979 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1980 SipHash_Final((u_int8_t *)&siphash, &ctx); 1981 1982 return (siphash[0] ^ siphash[1]); 1983 } 1984 1985 static tcp_seq 1986 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 1987 { 1988 u_int i, mss, secbit, wscale; 1989 uint32_t iss, hash; 1990 uint8_t *secbits; 1991 union syncookie cookie; 1992 1993 SCH_LOCK_ASSERT(sch); 1994 1995 cookie.cookie = 0; 1996 1997 /* Map our computed MSS into the 3-bit index. */ 1998 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss)); 1999 for (i = nitems(tcp_sc_msstab) - 1; tcp_sc_msstab[i] > mss && i > 0; 2000 i--) 2001 ; 2002 cookie.flags.mss_idx = i; 2003 2004 /* 2005 * Map the send window scale into the 3-bit index but only if 2006 * the wscale option was received. 2007 */ 2008 if (sc->sc_flags & SCF_WINSCALE) { 2009 wscale = sc->sc_requested_s_scale; 2010 for (i = nitems(tcp_sc_wstab) - 1; 2011 tcp_sc_wstab[i] > wscale && i > 0; 2012 i--) 2013 ; 2014 cookie.flags.wscale_idx = i; 2015 } 2016 2017 /* Can we do SACK? */ 2018 if (sc->sc_flags & SCF_SACK) 2019 cookie.flags.sack_ok = 1; 2020 2021 /* Which of the two secrets to use. */ 2022 secbit = sch->sch_sc->secret.oddeven & 0x1; 2023 cookie.flags.odd_even = secbit; 2024 2025 secbits = sch->sch_sc->secret.key[secbit]; 2026 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2027 (uintptr_t)sch); 2028 2029 /* 2030 * Put the flags into the hash and XOR them to get better ISS number 2031 * variance. This doesn't enhance the cryptographic strength and is 2032 * done to prevent the 8 cookie bits from showing up directly on the 2033 * wire. 2034 */ 2035 iss = hash & ~0xff; 2036 iss |= cookie.cookie ^ (hash >> 24); 2037 2038 /* Randomize the timestamp. */ 2039 if (sc->sc_flags & SCF_TIMESTAMP) { 2040 sc->sc_ts = arc4random(); 2041 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 2042 } 2043 2044 TCPSTAT_INC(tcps_sc_sendcookie); 2045 return (iss); 2046 } 2047 2048 static struct syncache * 2049 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2050 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2051 struct socket *lso) 2052 { 2053 uint32_t hash; 2054 uint8_t *secbits; 2055 tcp_seq ack, seq; 2056 int wnd, wscale = 0; 2057 union syncookie cookie; 2058 2059 SCH_LOCK_ASSERT(sch); 2060 2061 /* 2062 * Pull information out of SYN-ACK/ACK and revert sequence number 2063 * advances. 2064 */ 2065 ack = th->th_ack - 1; 2066 seq = th->th_seq - 1; 2067 2068 /* 2069 * Unpack the flags containing enough information to restore the 2070 * connection. 2071 */ 2072 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2073 2074 /* Which of the two secrets to use. */ 2075 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 2076 2077 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2078 2079 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2080 if ((ack & ~0xff) != (hash & ~0xff)) 2081 return (NULL); 2082 2083 /* Fill in the syncache values. */ 2084 sc->sc_flags = 0; 2085 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2086 sc->sc_ipopts = NULL; 2087 2088 sc->sc_irs = seq; 2089 sc->sc_iss = ack; 2090 2091 switch (inc->inc_flags & INC_ISIPV6) { 2092 #ifdef INET 2093 case 0: 2094 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2095 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2096 break; 2097 #endif 2098 #ifdef INET6 2099 case INC_ISIPV6: 2100 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2101 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 2102 break; 2103 #endif 2104 } 2105 2106 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2107 2108 /* We can simply recompute receive window scale we sent earlier. */ 2109 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2110 wscale++; 2111 2112 /* Only use wscale if it was enabled in the orignal SYN. */ 2113 if (cookie.flags.wscale_idx > 0) { 2114 sc->sc_requested_r_scale = wscale; 2115 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2116 sc->sc_flags |= SCF_WINSCALE; 2117 } 2118 2119 wnd = lso->sol_sbrcv_hiwat; 2120 wnd = imax(wnd, 0); 2121 wnd = imin(wnd, TCP_MAXWIN); 2122 sc->sc_wnd = wnd; 2123 2124 if (cookie.flags.sack_ok) 2125 sc->sc_flags |= SCF_SACK; 2126 2127 if (to->to_flags & TOF_TS) { 2128 sc->sc_flags |= SCF_TIMESTAMP; 2129 sc->sc_tsreflect = to->to_tsval; 2130 sc->sc_ts = to->to_tsecr; 2131 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 2132 } 2133 2134 if (to->to_flags & TOF_SIGNATURE) 2135 sc->sc_flags |= SCF_SIGNATURE; 2136 2137 sc->sc_rxmits = 0; 2138 2139 TCPSTAT_INC(tcps_sc_recvcookie); 2140 return (sc); 2141 } 2142 2143 #ifdef INVARIANTS 2144 static int 2145 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2146 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2147 struct socket *lso) 2148 { 2149 struct syncache scs, *scx; 2150 char *s; 2151 2152 bzero(&scs, sizeof(scs)); 2153 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2154 2155 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2156 return (0); 2157 2158 if (scx != NULL) { 2159 if (sc->sc_peer_mss != scx->sc_peer_mss) 2160 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2161 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2162 2163 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2164 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2165 s, __func__, sc->sc_requested_r_scale, 2166 scx->sc_requested_r_scale); 2167 2168 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2169 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2170 s, __func__, sc->sc_requested_s_scale, 2171 scx->sc_requested_s_scale); 2172 2173 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2174 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2175 } 2176 2177 if (s != NULL) 2178 free(s, M_TCPLOG); 2179 return (0); 2180 } 2181 #endif /* INVARIANTS */ 2182 2183 static void 2184 syncookie_reseed(void *arg) 2185 { 2186 struct tcp_syncache *sc = arg; 2187 uint8_t *secbits; 2188 int secbit; 2189 2190 /* 2191 * Reseeding the secret doesn't have to be protected by a lock. 2192 * It only must be ensured that the new random values are visible 2193 * to all CPUs in a SMP environment. The atomic with release 2194 * semantics ensures that. 2195 */ 2196 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2197 secbits = sc->secret.key[secbit]; 2198 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2199 atomic_add_rel_int(&sc->secret.oddeven, 1); 2200 2201 /* Reschedule ourself. */ 2202 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2203 } 2204 2205 /* 2206 * Exports the syncache entries to userland so that netstat can display 2207 * them alongside the other sockets. This function is intended to be 2208 * called only from tcp_pcblist. 2209 * 2210 * Due to concurrency on an active system, the number of pcbs exported 2211 * may have no relation to max_pcbs. max_pcbs merely indicates the 2212 * amount of space the caller allocated for this function to use. 2213 */ 2214 int 2215 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2216 { 2217 struct xtcpcb xt; 2218 struct syncache *sc; 2219 struct syncache_head *sch; 2220 int count, error, i; 2221 2222 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2223 sch = &V_tcp_syncache.hashbase[i]; 2224 SCH_LOCK(sch); 2225 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2226 if (count >= max_pcbs) { 2227 SCH_UNLOCK(sch); 2228 goto exit; 2229 } 2230 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2231 continue; 2232 bzero(&xt, sizeof(xt)); 2233 xt.xt_len = sizeof(xt); 2234 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2235 xt.xt_inp.inp_vflag = INP_IPV6; 2236 else 2237 xt.xt_inp.inp_vflag = INP_IPV4; 2238 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2239 sizeof (struct in_conninfo)); 2240 xt.t_state = TCPS_SYN_RECEIVED; 2241 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2242 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2243 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2244 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2245 error = SYSCTL_OUT(req, &xt, sizeof xt); 2246 if (error) { 2247 SCH_UNLOCK(sch); 2248 goto exit; 2249 } 2250 count++; 2251 } 2252 SCH_UNLOCK(sch); 2253 } 2254 exit: 2255 *pcbs_exported = count; 2256 return error; 2257 } 2258