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 } 264 265 /* Create the syncache entry zone. */ 266 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 267 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 268 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 269 V_tcp_syncache.cache_limit); 270 271 /* Start the SYN cookie reseeder callout. */ 272 callout_init(&V_tcp_syncache.secret.reseed, 1); 273 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 274 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 275 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 276 syncookie_reseed, &V_tcp_syncache); 277 } 278 279 #ifdef VIMAGE 280 void 281 syncache_destroy(void) 282 { 283 struct syncache_head *sch; 284 struct syncache *sc, *nsc; 285 int i; 286 287 /* 288 * Stop the re-seed timer before freeing resources. No need to 289 * possibly schedule it another time. 290 */ 291 callout_drain(&V_tcp_syncache.secret.reseed); 292 293 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 294 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 295 296 sch = &V_tcp_syncache.hashbase[i]; 297 callout_drain(&sch->sch_timer); 298 299 SCH_LOCK(sch); 300 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 301 syncache_drop(sc, sch); 302 SCH_UNLOCK(sch); 303 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 304 ("%s: sch->sch_bucket not empty", __func__)); 305 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 306 __func__, sch->sch_length)); 307 mtx_destroy(&sch->sch_mtx); 308 } 309 310 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 311 ("%s: cache_count not 0", __func__)); 312 313 /* Free the allocated global resources. */ 314 uma_zdestroy(V_tcp_syncache.zone); 315 free(V_tcp_syncache.hashbase, M_SYNCACHE); 316 } 317 #endif 318 319 /* 320 * Inserts a syncache entry into the specified bucket row. 321 * Locks and unlocks the syncache_head autonomously. 322 */ 323 static void 324 syncache_insert(struct syncache *sc, struct syncache_head *sch) 325 { 326 struct syncache *sc2; 327 328 SCH_LOCK(sch); 329 330 /* 331 * Make sure that we don't overflow the per-bucket limit. 332 * If the bucket is full, toss the oldest element. 333 */ 334 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 335 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 336 ("sch->sch_length incorrect")); 337 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 338 syncache_drop(sc2, sch); 339 TCPSTAT_INC(tcps_sc_bucketoverflow); 340 } 341 342 /* Put it into the bucket. */ 343 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 344 sch->sch_length++; 345 346 #ifdef TCP_OFFLOAD 347 if (ADDED_BY_TOE(sc)) { 348 struct toedev *tod = sc->sc_tod; 349 350 tod->tod_syncache_added(tod, sc->sc_todctx); 351 } 352 #endif 353 354 /* Reinitialize the bucket row's timer. */ 355 if (sch->sch_length == 1) 356 sch->sch_nextc = ticks + INT_MAX; 357 syncache_timeout(sc, sch, 1); 358 359 SCH_UNLOCK(sch); 360 361 TCPSTATES_INC(TCPS_SYN_RECEIVED); 362 TCPSTAT_INC(tcps_sc_added); 363 } 364 365 /* 366 * Remove and free entry from syncache bucket row. 367 * Expects locked syncache head. 368 */ 369 static void 370 syncache_drop(struct syncache *sc, struct syncache_head *sch) 371 { 372 373 SCH_LOCK_ASSERT(sch); 374 375 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 376 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 377 sch->sch_length--; 378 379 #ifdef TCP_OFFLOAD 380 if (ADDED_BY_TOE(sc)) { 381 struct toedev *tod = sc->sc_tod; 382 383 tod->tod_syncache_removed(tod, sc->sc_todctx); 384 } 385 #endif 386 387 syncache_free(sc); 388 } 389 390 /* 391 * Engage/reengage time on bucket row. 392 */ 393 static void 394 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 395 { 396 sc->sc_rxttime = ticks + 397 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]); 398 sc->sc_rxmits++; 399 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 400 sch->sch_nextc = sc->sc_rxttime; 401 if (docallout) 402 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 403 syncache_timer, (void *)sch); 404 } 405 } 406 407 /* 408 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 409 * If we have retransmitted an entry the maximum number of times, expire it. 410 * One separate timer for each bucket row. 411 */ 412 static void 413 syncache_timer(void *xsch) 414 { 415 struct syncache_head *sch = (struct syncache_head *)xsch; 416 struct syncache *sc, *nsc; 417 int tick = ticks; 418 char *s; 419 420 CURVNET_SET(sch->sch_sc->vnet); 421 422 /* NB: syncache_head has already been locked by the callout. */ 423 SCH_LOCK_ASSERT(sch); 424 425 /* 426 * In the following cycle we may remove some entries and/or 427 * advance some timeouts, so re-initialize the bucket timer. 428 */ 429 sch->sch_nextc = tick + INT_MAX; 430 431 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 432 /* 433 * We do not check if the listen socket still exists 434 * and accept the case where the listen socket may be 435 * gone by the time we resend the SYN/ACK. We do 436 * not expect this to happens often. If it does, 437 * then the RST will be sent by the time the remote 438 * host does the SYN/ACK->ACK. 439 */ 440 if (TSTMP_GT(sc->sc_rxttime, tick)) { 441 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 442 sch->sch_nextc = sc->sc_rxttime; 443 continue; 444 } 445 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 446 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 447 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 448 "giving up and removing syncache entry\n", 449 s, __func__); 450 free(s, M_TCPLOG); 451 } 452 syncache_drop(sc, sch); 453 TCPSTAT_INC(tcps_sc_stale); 454 continue; 455 } 456 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 457 log(LOG_DEBUG, "%s; %s: Response timeout, " 458 "retransmitting (%u) SYN|ACK\n", 459 s, __func__, sc->sc_rxmits); 460 free(s, M_TCPLOG); 461 } 462 463 syncache_respond(sc, sch, 1, NULL); 464 TCPSTAT_INC(tcps_sc_retransmitted); 465 syncache_timeout(sc, sch, 0); 466 } 467 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 468 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 469 syncache_timer, (void *)(sch)); 470 CURVNET_RESTORE(); 471 } 472 473 /* 474 * Find an entry in the syncache. 475 * Returns always with locked syncache_head plus a matching entry or NULL. 476 */ 477 static struct syncache * 478 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 479 { 480 struct syncache *sc; 481 struct syncache_head *sch; 482 uint32_t hash; 483 484 /* 485 * The hash is built on foreign port + local port + foreign address. 486 * We rely on the fact that struct in_conninfo starts with 16 bits 487 * of foreign port, then 16 bits of local port then followed by 128 488 * bits of foreign address. In case of IPv4 address, the first 3 489 * 32-bit words of the address always are zeroes. 490 */ 491 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 492 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 493 494 sch = &V_tcp_syncache.hashbase[hash]; 495 *schp = sch; 496 SCH_LOCK(sch); 497 498 /* Circle through bucket row to find matching entry. */ 499 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 500 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 501 sizeof(struct in_endpoints)) == 0) 502 break; 503 504 return (sc); /* Always returns with locked sch. */ 505 } 506 507 /* 508 * This function is called when we get a RST for a 509 * non-existent connection, so that we can see if the 510 * connection is in the syn cache. If it is, zap it. 511 */ 512 void 513 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 514 { 515 struct syncache *sc; 516 struct syncache_head *sch; 517 char *s = NULL; 518 519 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 520 SCH_LOCK_ASSERT(sch); 521 522 /* 523 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 524 * See RFC 793 page 65, section SEGMENT ARRIVES. 525 */ 526 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 527 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 528 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 529 "FIN flag set, segment ignored\n", s, __func__); 530 TCPSTAT_INC(tcps_badrst); 531 goto done; 532 } 533 534 /* 535 * No corresponding connection was found in syncache. 536 * If syncookies are enabled and possibly exclusively 537 * used, or we are under memory pressure, a valid RST 538 * may not find a syncache entry. In that case we're 539 * done and no SYN|ACK retransmissions will happen. 540 * Otherwise the RST was misdirected or spoofed. 541 */ 542 if (sc == NULL) { 543 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 544 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 545 "syncache entry (possibly syncookie only), " 546 "segment ignored\n", s, __func__); 547 TCPSTAT_INC(tcps_badrst); 548 goto done; 549 } 550 551 /* 552 * If the RST bit is set, check the sequence number to see 553 * if this is a valid reset segment. 554 * RFC 793 page 37: 555 * In all states except SYN-SENT, all reset (RST) segments 556 * are validated by checking their SEQ-fields. A reset is 557 * valid if its sequence number is in the window. 558 * 559 * The sequence number in the reset segment is normally an 560 * echo of our outgoing acknowlegement numbers, but some hosts 561 * send a reset with the sequence number at the rightmost edge 562 * of our receive window, and we have to handle this case. 563 */ 564 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 565 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 566 syncache_drop(sc, sch); 567 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 568 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 569 "connection attempt aborted by remote endpoint\n", 570 s, __func__); 571 TCPSTAT_INC(tcps_sc_reset); 572 } else { 573 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 574 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 575 "IRS %u (+WND %u), segment ignored\n", 576 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 577 TCPSTAT_INC(tcps_badrst); 578 } 579 580 done: 581 if (s != NULL) 582 free(s, M_TCPLOG); 583 SCH_UNLOCK(sch); 584 } 585 586 void 587 syncache_badack(struct in_conninfo *inc) 588 { 589 struct syncache *sc; 590 struct syncache_head *sch; 591 592 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 593 SCH_LOCK_ASSERT(sch); 594 if (sc != NULL) { 595 syncache_drop(sc, sch); 596 TCPSTAT_INC(tcps_sc_badack); 597 } 598 SCH_UNLOCK(sch); 599 } 600 601 void 602 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th) 603 { 604 struct syncache *sc; 605 struct syncache_head *sch; 606 607 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 608 SCH_LOCK_ASSERT(sch); 609 if (sc == NULL) 610 goto done; 611 612 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 613 if (ntohl(th->th_seq) != sc->sc_iss) 614 goto done; 615 616 /* 617 * If we've rertransmitted 3 times and this is our second error, 618 * we remove the entry. Otherwise, we allow it to continue on. 619 * This prevents us from incorrectly nuking an entry during a 620 * spurious network outage. 621 * 622 * See tcp_notify(). 623 */ 624 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 625 sc->sc_flags |= SCF_UNREACH; 626 goto done; 627 } 628 syncache_drop(sc, sch); 629 TCPSTAT_INC(tcps_sc_unreach); 630 done: 631 SCH_UNLOCK(sch); 632 } 633 634 /* 635 * Build a new TCP socket structure from a syncache entry. 636 * 637 * On success return the newly created socket with its underlying inp locked. 638 */ 639 static struct socket * 640 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 641 { 642 struct tcp_function_block *blk; 643 struct inpcb *inp = NULL; 644 struct socket *so; 645 struct tcpcb *tp; 646 int error; 647 char *s; 648 649 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 650 651 /* 652 * Ok, create the full blown connection, and set things up 653 * as they would have been set up if we had created the 654 * connection when the SYN arrived. If we can't create 655 * the connection, abort it. 656 */ 657 so = sonewconn(lso, 0); 658 if (so == NULL) { 659 /* 660 * Drop the connection; we will either send a RST or 661 * have the peer retransmit its SYN again after its 662 * RTO and try again. 663 */ 664 TCPSTAT_INC(tcps_listendrop); 665 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 666 log(LOG_DEBUG, "%s; %s: Socket create failed " 667 "due to limits or memory shortage\n", 668 s, __func__); 669 free(s, M_TCPLOG); 670 } 671 goto abort2; 672 } 673 #ifdef MAC 674 mac_socketpeer_set_from_mbuf(m, so); 675 #endif 676 677 inp = sotoinpcb(so); 678 inp->inp_inc.inc_fibnum = so->so_fibnum; 679 INP_WLOCK(inp); 680 /* 681 * Exclusive pcbinfo lock is not required in syncache socket case even 682 * if two inpcb locks can be acquired simultaneously: 683 * - the inpcb in LISTEN state, 684 * - the newly created inp. 685 * 686 * In this case, an inp cannot be at same time in LISTEN state and 687 * just created by an accept() call. 688 */ 689 INP_HASH_WLOCK(&V_tcbinfo); 690 691 /* Insert new socket into PCB hash list. */ 692 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 693 #ifdef INET6 694 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 695 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 696 } else { 697 inp->inp_vflag &= ~INP_IPV6; 698 inp->inp_vflag |= INP_IPV4; 699 #endif 700 inp->inp_laddr = sc->sc_inc.inc_laddr; 701 #ifdef INET6 702 } 703 #endif 704 705 /* 706 * If there's an mbuf and it has a flowid, then let's initialise the 707 * inp with that particular flowid. 708 */ 709 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 710 inp->inp_flowid = m->m_pkthdr.flowid; 711 inp->inp_flowtype = M_HASHTYPE_GET(m); 712 } 713 714 /* 715 * Install in the reservation hash table for now, but don't yet 716 * install a connection group since the full 4-tuple isn't yet 717 * configured. 718 */ 719 inp->inp_lport = sc->sc_inc.inc_lport; 720 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 721 /* 722 * Undo the assignments above if we failed to 723 * put the PCB on the hash lists. 724 */ 725 #ifdef INET6 726 if (sc->sc_inc.inc_flags & INC_ISIPV6) 727 inp->in6p_laddr = in6addr_any; 728 else 729 #endif 730 inp->inp_laddr.s_addr = INADDR_ANY; 731 inp->inp_lport = 0; 732 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 733 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 734 "with error %i\n", 735 s, __func__, error); 736 free(s, M_TCPLOG); 737 } 738 INP_HASH_WUNLOCK(&V_tcbinfo); 739 goto abort; 740 } 741 #ifdef INET6 742 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 743 struct inpcb *oinp = sotoinpcb(lso); 744 struct in6_addr laddr6; 745 struct sockaddr_in6 sin6; 746 /* 747 * Inherit socket options from the listening socket. 748 * Note that in6p_inputopts are not (and should not be) 749 * copied, since it stores previously received options and is 750 * used to detect if each new option is different than the 751 * previous one and hence should be passed to a user. 752 * If we copied in6p_inputopts, a user would not be able to 753 * receive options just after calling the accept system call. 754 */ 755 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 756 if (oinp->in6p_outputopts) 757 inp->in6p_outputopts = 758 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 759 760 sin6.sin6_family = AF_INET6; 761 sin6.sin6_len = sizeof(sin6); 762 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 763 sin6.sin6_port = sc->sc_inc.inc_fport; 764 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 765 laddr6 = inp->in6p_laddr; 766 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 767 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 768 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 769 thread0.td_ucred, m)) != 0) { 770 inp->in6p_laddr = laddr6; 771 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 772 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 773 "with error %i\n", 774 s, __func__, error); 775 free(s, M_TCPLOG); 776 } 777 INP_HASH_WUNLOCK(&V_tcbinfo); 778 goto abort; 779 } 780 /* Override flowlabel from in6_pcbconnect. */ 781 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 782 inp->inp_flow |= sc->sc_flowlabel; 783 } 784 #endif /* INET6 */ 785 #if defined(INET) && defined(INET6) 786 else 787 #endif 788 #ifdef INET 789 { 790 struct in_addr laddr; 791 struct sockaddr_in sin; 792 793 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 794 795 if (inp->inp_options == NULL) { 796 inp->inp_options = sc->sc_ipopts; 797 sc->sc_ipopts = NULL; 798 } 799 800 sin.sin_family = AF_INET; 801 sin.sin_len = sizeof(sin); 802 sin.sin_addr = sc->sc_inc.inc_faddr; 803 sin.sin_port = sc->sc_inc.inc_fport; 804 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 805 laddr = inp->inp_laddr; 806 if (inp->inp_laddr.s_addr == INADDR_ANY) 807 inp->inp_laddr = sc->sc_inc.inc_laddr; 808 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 809 thread0.td_ucred, m)) != 0) { 810 inp->inp_laddr = laddr; 811 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 812 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 813 "with error %i\n", 814 s, __func__, error); 815 free(s, M_TCPLOG); 816 } 817 INP_HASH_WUNLOCK(&V_tcbinfo); 818 goto abort; 819 } 820 } 821 #endif /* INET */ 822 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 823 /* Copy old policy into new socket's. */ 824 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 825 printf("syncache_socket: could not copy policy\n"); 826 #endif 827 INP_HASH_WUNLOCK(&V_tcbinfo); 828 tp = intotcpcb(inp); 829 tcp_state_change(tp, TCPS_SYN_RECEIVED); 830 tp->iss = sc->sc_iss; 831 tp->irs = sc->sc_irs; 832 tcp_rcvseqinit(tp); 833 tcp_sendseqinit(tp); 834 blk = sototcpcb(lso)->t_fb; 835 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 836 /* 837 * Our parents t_fb was not the default, 838 * we need to release our ref on tp->t_fb and 839 * pickup one on the new entry. 840 */ 841 struct tcp_function_block *rblk; 842 843 rblk = find_and_ref_tcp_fb(blk); 844 KASSERT(rblk != NULL, 845 ("cannot find blk %p out of syncache?", blk)); 846 if (tp->t_fb->tfb_tcp_fb_fini) 847 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 848 refcount_release(&tp->t_fb->tfb_refcnt); 849 tp->t_fb = rblk; 850 if (tp->t_fb->tfb_tcp_fb_init) { 851 (*tp->t_fb->tfb_tcp_fb_init)(tp); 852 } 853 } 854 tp->snd_wl1 = sc->sc_irs; 855 tp->snd_max = tp->iss + 1; 856 tp->snd_nxt = tp->iss + 1; 857 tp->rcv_up = sc->sc_irs + 1; 858 tp->rcv_wnd = sc->sc_wnd; 859 tp->rcv_adv += tp->rcv_wnd; 860 tp->last_ack_sent = tp->rcv_nxt; 861 862 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 863 if (sc->sc_flags & SCF_NOOPT) 864 tp->t_flags |= TF_NOOPT; 865 else { 866 if (sc->sc_flags & SCF_WINSCALE) { 867 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 868 tp->snd_scale = sc->sc_requested_s_scale; 869 tp->request_r_scale = sc->sc_requested_r_scale; 870 } 871 if (sc->sc_flags & SCF_TIMESTAMP) { 872 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 873 tp->ts_recent = sc->sc_tsreflect; 874 tp->ts_recent_age = tcp_ts_getticks(); 875 tp->ts_offset = sc->sc_tsoff; 876 } 877 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 878 if (sc->sc_flags & SCF_SIGNATURE) 879 tp->t_flags |= TF_SIGNATURE; 880 #endif 881 if (sc->sc_flags & SCF_SACK) 882 tp->t_flags |= TF_SACK_PERMIT; 883 } 884 885 if (sc->sc_flags & SCF_ECN) 886 tp->t_flags |= TF_ECN_PERMIT; 887 888 /* 889 * Set up MSS and get cached values from tcp_hostcache. 890 * This might overwrite some of the defaults we just set. 891 */ 892 tcp_mss(tp, sc->sc_peer_mss); 893 894 /* 895 * If the SYN,ACK was retransmitted, indicate that CWND to be 896 * limited to one segment in cc_conn_init(). 897 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 898 */ 899 if (sc->sc_rxmits > 1) 900 tp->snd_cwnd = 1; 901 902 #ifdef TCP_OFFLOAD 903 /* 904 * Allow a TOE driver to install its hooks. Note that we hold the 905 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 906 * new connection before the TOE driver has done its thing. 907 */ 908 if (ADDED_BY_TOE(sc)) { 909 struct toedev *tod = sc->sc_tod; 910 911 tod->tod_offload_socket(tod, sc->sc_todctx, so); 912 } 913 #endif 914 /* 915 * Copy and activate timers. 916 */ 917 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 918 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 919 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 920 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 921 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 922 923 TCPSTAT_INC(tcps_accepts); 924 return (so); 925 926 abort: 927 INP_WUNLOCK(inp); 928 abort2: 929 if (so != NULL) 930 soabort(so); 931 return (NULL); 932 } 933 934 /* 935 * This function gets called when we receive an ACK for a 936 * socket in the LISTEN state. We look up the connection 937 * in the syncache, and if its there, we pull it out of 938 * the cache and turn it into a full-blown connection in 939 * the SYN-RECEIVED state. 940 * 941 * On syncache_socket() success the newly created socket 942 * has its underlying inp locked. 943 */ 944 int 945 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 946 struct socket **lsop, struct mbuf *m) 947 { 948 struct syncache *sc; 949 struct syncache_head *sch; 950 struct syncache scs; 951 char *s; 952 953 /* 954 * Global TCP locks are held because we manipulate the PCB lists 955 * and create a new socket. 956 */ 957 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 958 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 959 ("%s: can handle only ACK", __func__)); 960 961 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 962 SCH_LOCK_ASSERT(sch); 963 964 #ifdef INVARIANTS 965 /* 966 * Test code for syncookies comparing the syncache stored 967 * values with the reconstructed values from the cookie. 968 */ 969 if (sc != NULL) 970 syncookie_cmp(inc, sch, sc, th, to, *lsop); 971 #endif 972 973 if (sc == NULL) { 974 /* 975 * There is no syncache entry, so see if this ACK is 976 * a returning syncookie. To do this, first: 977 * A. See if this socket has had a syncache entry dropped in 978 * the past. We don't want to accept a bogus syncookie 979 * if we've never received a SYN. 980 * B. check that the syncookie is valid. If it is, then 981 * cobble up a fake syncache entry, and return. 982 */ 983 if (!V_tcp_syncookies) { 984 SCH_UNLOCK(sch); 985 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 986 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 987 "segment rejected (syncookies disabled)\n", 988 s, __func__); 989 goto failed; 990 } 991 bzero(&scs, sizeof(scs)); 992 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 993 SCH_UNLOCK(sch); 994 if (sc == NULL) { 995 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 996 log(LOG_DEBUG, "%s; %s: Segment failed " 997 "SYNCOOKIE authentication, segment rejected " 998 "(probably spoofed)\n", s, __func__); 999 goto failed; 1000 } 1001 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1002 /* If received ACK has MD5 signature, check it. */ 1003 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1004 (!TCPMD5_ENABLED() || 1005 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1006 /* Drop the ACK. */ 1007 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1008 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1009 "MD5 signature doesn't match.\n", 1010 s, __func__); 1011 free(s, M_TCPLOG); 1012 } 1013 TCPSTAT_INC(tcps_sig_err_sigopt); 1014 return (-1); /* Do not send RST */ 1015 } 1016 #endif /* TCP_SIGNATURE */ 1017 } else { 1018 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1019 /* 1020 * If listening socket requested TCP digests, check that 1021 * received ACK has signature and it is correct. 1022 * If not, drop the ACK and leave sc entry in th cache, 1023 * because SYN was received with correct signature. 1024 */ 1025 if (sc->sc_flags & SCF_SIGNATURE) { 1026 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1027 /* No signature */ 1028 TCPSTAT_INC(tcps_sig_err_nosigopt); 1029 SCH_UNLOCK(sch); 1030 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1031 log(LOG_DEBUG, "%s; %s: Segment " 1032 "rejected, MD5 signature wasn't " 1033 "provided.\n", s, __func__); 1034 free(s, M_TCPLOG); 1035 } 1036 return (-1); /* Do not send RST */ 1037 } 1038 if (!TCPMD5_ENABLED() || 1039 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1040 /* Doesn't match or no SA */ 1041 SCH_UNLOCK(sch); 1042 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1043 log(LOG_DEBUG, "%s; %s: Segment " 1044 "rejected, MD5 signature doesn't " 1045 "match.\n", s, __func__); 1046 free(s, M_TCPLOG); 1047 } 1048 return (-1); /* Do not send RST */ 1049 } 1050 } 1051 #endif /* TCP_SIGNATURE */ 1052 /* 1053 * Pull out the entry to unlock the bucket row. 1054 * 1055 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1056 * tcp_state_change(). The tcpcb is not existent at this 1057 * moment. A new one will be allocated via syncache_socket-> 1058 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1059 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1060 */ 1061 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1062 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1063 sch->sch_length--; 1064 #ifdef TCP_OFFLOAD 1065 if (ADDED_BY_TOE(sc)) { 1066 struct toedev *tod = sc->sc_tod; 1067 1068 tod->tod_syncache_removed(tod, sc->sc_todctx); 1069 } 1070 #endif 1071 SCH_UNLOCK(sch); 1072 } 1073 1074 /* 1075 * Segment validation: 1076 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1077 */ 1078 if (th->th_ack != sc->sc_iss + 1) { 1079 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1080 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1081 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1082 goto failed; 1083 } 1084 1085 /* 1086 * The SEQ must fall in the window starting at the received 1087 * initial receive sequence number + 1 (the SYN). 1088 */ 1089 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1090 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1091 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1092 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1093 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1094 goto failed; 1095 } 1096 1097 /* 1098 * If timestamps were not negotiated during SYN/ACK they 1099 * must not appear on any segment during this session. 1100 */ 1101 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1102 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1103 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1104 "segment rejected\n", s, __func__); 1105 goto failed; 1106 } 1107 1108 /* 1109 * If timestamps were negotiated during SYN/ACK they should 1110 * appear on every segment during this session. 1111 * XXXAO: This is only informal as there have been unverified 1112 * reports of non-compliants stacks. 1113 */ 1114 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1115 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1116 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1117 "no action\n", s, __func__); 1118 free(s, M_TCPLOG); 1119 s = NULL; 1120 } 1121 } 1122 1123 /* 1124 * If timestamps were negotiated, the reflected timestamp 1125 * must be equal to what we actually sent in the SYN|ACK 1126 * except in the case of 0. Some boxes are known for sending 1127 * broken timestamp replies during the 3whs (and potentially 1128 * during the connection also). 1129 * 1130 * Accept the final ACK of 3whs with reflected timestamp of 0 1131 * instead of sending a RST and deleting the syncache entry. 1132 */ 1133 if ((to->to_flags & TOF_TS) && to->to_tsecr && 1134 to->to_tsecr != sc->sc_ts) { 1135 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1136 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1137 "segment rejected\n", 1138 s, __func__, to->to_tsecr, sc->sc_ts); 1139 goto failed; 1140 } 1141 1142 *lsop = syncache_socket(sc, *lsop, m); 1143 1144 if (*lsop == NULL) 1145 TCPSTAT_INC(tcps_sc_aborted); 1146 else 1147 TCPSTAT_INC(tcps_sc_completed); 1148 1149 /* how do we find the inp for the new socket? */ 1150 if (sc != &scs) 1151 syncache_free(sc); 1152 return (1); 1153 failed: 1154 if (sc != NULL && sc != &scs) 1155 syncache_free(sc); 1156 if (s != NULL) 1157 free(s, M_TCPLOG); 1158 *lsop = NULL; 1159 return (0); 1160 } 1161 1162 #ifdef TCP_RFC7413 1163 static void 1164 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1165 uint64_t response_cookie) 1166 { 1167 struct inpcb *inp; 1168 struct tcpcb *tp; 1169 unsigned int *pending_counter; 1170 1171 /* 1172 * Global TCP locks are held because we manipulate the PCB lists 1173 * and create a new socket. 1174 */ 1175 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1176 1177 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1178 *lsop = syncache_socket(sc, *lsop, m); 1179 if (*lsop == NULL) { 1180 TCPSTAT_INC(tcps_sc_aborted); 1181 atomic_subtract_int(pending_counter, 1); 1182 } else { 1183 inp = sotoinpcb(*lsop); 1184 tp = intotcpcb(inp); 1185 tp->t_flags |= TF_FASTOPEN; 1186 tp->t_tfo_cookie = response_cookie; 1187 tp->snd_max = tp->iss; 1188 tp->snd_nxt = tp->iss; 1189 tp->t_tfo_pending = pending_counter; 1190 TCPSTAT_INC(tcps_sc_completed); 1191 } 1192 } 1193 #endif /* TCP_RFC7413 */ 1194 1195 /* 1196 * Given a LISTEN socket and an inbound SYN request, add 1197 * this to the syn cache, and send back a segment: 1198 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1199 * to the source. 1200 * 1201 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1202 * Doing so would require that we hold onto the data and deliver it 1203 * to the application. However, if we are the target of a SYN-flood 1204 * DoS attack, an attacker could send data which would eventually 1205 * consume all available buffer space if it were ACKed. By not ACKing 1206 * the data, we avoid this DoS scenario. 1207 * 1208 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1209 * cookie is processed and a new socket is created. In this case, any data 1210 * accompanying the SYN will be queued to the socket by tcp_input() and will 1211 * be ACKed either when the application sends response data or the delayed 1212 * ACK timer expires, whichever comes first. 1213 */ 1214 int 1215 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1216 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1217 void *todctx) 1218 { 1219 struct tcpcb *tp; 1220 struct socket *so; 1221 struct syncache *sc = NULL; 1222 struct syncache_head *sch; 1223 struct mbuf *ipopts = NULL; 1224 u_int ltflags; 1225 int win, ip_ttl, ip_tos; 1226 char *s; 1227 int rv = 0; 1228 #ifdef INET6 1229 int autoflowlabel = 0; 1230 #endif 1231 #ifdef MAC 1232 struct label *maclabel; 1233 #endif 1234 struct syncache scs; 1235 struct ucred *cred; 1236 #ifdef TCP_RFC7413 1237 uint64_t tfo_response_cookie; 1238 unsigned int *tfo_pending = NULL; 1239 int tfo_cookie_valid = 0; 1240 int tfo_response_cookie_valid = 0; 1241 #endif 1242 1243 INP_WLOCK_ASSERT(inp); /* listen socket */ 1244 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1245 ("%s: unexpected tcp flags", __func__)); 1246 1247 /* 1248 * Combine all so/tp operations very early to drop the INP lock as 1249 * soon as possible. 1250 */ 1251 so = *lsop; 1252 tp = sototcpcb(so); 1253 cred = crhold(so->so_cred); 1254 1255 #ifdef INET6 1256 if ((inc->inc_flags & INC_ISIPV6) && 1257 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1258 autoflowlabel = 1; 1259 #endif 1260 ip_ttl = inp->inp_ip_ttl; 1261 ip_tos = inp->inp_ip_tos; 1262 win = sbspace(&so->so_rcv); 1263 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1264 1265 #ifdef TCP_RFC7413 1266 if (V_tcp_fastopen_enabled && IS_FASTOPEN(tp->t_flags) && 1267 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) { 1268 /* 1269 * Limit the number of pending TFO connections to 1270 * approximately half of the queue limit. This prevents TFO 1271 * SYN floods from starving the service by filling the 1272 * listen queue with bogus TFO connections. 1273 */ 1274 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1275 (so->so_qlimit / 2)) { 1276 int result; 1277 1278 result = tcp_fastopen_check_cookie(inc, 1279 to->to_tfo_cookie, to->to_tfo_len, 1280 &tfo_response_cookie); 1281 tfo_cookie_valid = (result > 0); 1282 tfo_response_cookie_valid = (result >= 0); 1283 } 1284 1285 /* 1286 * Remember the TFO pending counter as it will have to be 1287 * decremented below if we don't make it to syncache_tfo_expand(). 1288 */ 1289 tfo_pending = tp->t_tfo_pending; 1290 } 1291 #endif 1292 1293 /* By the time we drop the lock these should no longer be used. */ 1294 so = NULL; 1295 tp = NULL; 1296 1297 #ifdef MAC 1298 if (mac_syncache_init(&maclabel) != 0) { 1299 INP_WUNLOCK(inp); 1300 goto done; 1301 } else 1302 mac_syncache_create(maclabel, inp); 1303 #endif 1304 #ifdef TCP_RFC7413 1305 if (!tfo_cookie_valid) 1306 #endif 1307 INP_WUNLOCK(inp); 1308 1309 /* 1310 * Remember the IP options, if any. 1311 */ 1312 #ifdef INET6 1313 if (!(inc->inc_flags & INC_ISIPV6)) 1314 #endif 1315 #ifdef INET 1316 ipopts = (m) ? ip_srcroute(m) : NULL; 1317 #else 1318 ipopts = NULL; 1319 #endif 1320 1321 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1322 /* 1323 * If listening socket requested TCP digests, check that received 1324 * SYN has signature and it is correct. If signature doesn't match 1325 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1326 */ 1327 if (ltflags & TF_SIGNATURE) { 1328 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1329 TCPSTAT_INC(tcps_sig_err_nosigopt); 1330 goto done; 1331 } 1332 if (!TCPMD5_ENABLED() || 1333 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1334 goto done; 1335 } 1336 #endif /* TCP_SIGNATURE */ 1337 /* 1338 * See if we already have an entry for this connection. 1339 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1340 * 1341 * XXX: should the syncache be re-initialized with the contents 1342 * of the new SYN here (which may have different options?) 1343 * 1344 * XXX: We do not check the sequence number to see if this is a 1345 * real retransmit or a new connection attempt. The question is 1346 * how to handle such a case; either ignore it as spoofed, or 1347 * drop the current entry and create a new one? 1348 */ 1349 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1350 SCH_LOCK_ASSERT(sch); 1351 if (sc != NULL) { 1352 #ifdef TCP_RFC7413 1353 if (tfo_cookie_valid) 1354 INP_WUNLOCK(inp); 1355 #endif 1356 TCPSTAT_INC(tcps_sc_dupsyn); 1357 if (ipopts) { 1358 /* 1359 * If we were remembering a previous source route, 1360 * forget it and use the new one we've been given. 1361 */ 1362 if (sc->sc_ipopts) 1363 (void) m_free(sc->sc_ipopts); 1364 sc->sc_ipopts = ipopts; 1365 } 1366 /* 1367 * Update timestamp if present. 1368 */ 1369 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1370 sc->sc_tsreflect = to->to_tsval; 1371 else 1372 sc->sc_flags &= ~SCF_TIMESTAMP; 1373 #ifdef MAC 1374 /* 1375 * Since we have already unconditionally allocated label 1376 * storage, free it up. The syncache entry will already 1377 * have an initialized label we can use. 1378 */ 1379 mac_syncache_destroy(&maclabel); 1380 #endif 1381 /* Retransmit SYN|ACK and reset retransmit count. */ 1382 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1383 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1384 "resetting timer and retransmitting SYN|ACK\n", 1385 s, __func__); 1386 free(s, M_TCPLOG); 1387 } 1388 if (syncache_respond(sc, sch, 1, m) == 0) { 1389 sc->sc_rxmits = 0; 1390 syncache_timeout(sc, sch, 1); 1391 TCPSTAT_INC(tcps_sndacks); 1392 TCPSTAT_INC(tcps_sndtotal); 1393 } 1394 SCH_UNLOCK(sch); 1395 goto done; 1396 } 1397 1398 #ifdef TCP_RFC7413 1399 if (tfo_cookie_valid) { 1400 bzero(&scs, sizeof(scs)); 1401 sc = &scs; 1402 goto skip_alloc; 1403 } 1404 #endif 1405 1406 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1407 if (sc == NULL) { 1408 /* 1409 * The zone allocator couldn't provide more entries. 1410 * Treat this as if the cache was full; drop the oldest 1411 * entry and insert the new one. 1412 */ 1413 TCPSTAT_INC(tcps_sc_zonefail); 1414 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) 1415 syncache_drop(sc, sch); 1416 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1417 if (sc == NULL) { 1418 if (V_tcp_syncookies) { 1419 bzero(&scs, sizeof(scs)); 1420 sc = &scs; 1421 } else { 1422 SCH_UNLOCK(sch); 1423 if (ipopts) 1424 (void) m_free(ipopts); 1425 goto done; 1426 } 1427 } 1428 } 1429 1430 #ifdef TCP_RFC7413 1431 skip_alloc: 1432 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1433 sc->sc_tfo_cookie = &tfo_response_cookie; 1434 #endif 1435 1436 /* 1437 * Fill in the syncache values. 1438 */ 1439 #ifdef MAC 1440 sc->sc_label = maclabel; 1441 #endif 1442 sc->sc_cred = cred; 1443 cred = NULL; 1444 sc->sc_ipopts = ipopts; 1445 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1446 #ifdef INET6 1447 if (!(inc->inc_flags & INC_ISIPV6)) 1448 #endif 1449 { 1450 sc->sc_ip_tos = ip_tos; 1451 sc->sc_ip_ttl = ip_ttl; 1452 } 1453 #ifdef TCP_OFFLOAD 1454 sc->sc_tod = tod; 1455 sc->sc_todctx = todctx; 1456 #endif 1457 sc->sc_irs = th->th_seq; 1458 sc->sc_iss = arc4random(); 1459 sc->sc_flags = 0; 1460 sc->sc_flowlabel = 0; 1461 1462 /* 1463 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1464 * win was derived from socket earlier in the function. 1465 */ 1466 win = imax(win, 0); 1467 win = imin(win, TCP_MAXWIN); 1468 sc->sc_wnd = win; 1469 1470 if (V_tcp_do_rfc1323) { 1471 /* 1472 * A timestamp received in a SYN makes 1473 * it ok to send timestamp requests and replies. 1474 */ 1475 if (to->to_flags & TOF_TS) { 1476 sc->sc_tsreflect = to->to_tsval; 1477 sc->sc_ts = tcp_ts_getticks(); 1478 sc->sc_flags |= SCF_TIMESTAMP; 1479 } 1480 if (to->to_flags & TOF_SCALE) { 1481 int wscale = 0; 1482 1483 /* 1484 * Pick the smallest possible scaling factor that 1485 * will still allow us to scale up to sb_max, aka 1486 * kern.ipc.maxsockbuf. 1487 * 1488 * We do this because there are broken firewalls that 1489 * will corrupt the window scale option, leading to 1490 * the other endpoint believing that our advertised 1491 * window is unscaled. At scale factors larger than 1492 * 5 the unscaled window will drop below 1500 bytes, 1493 * leading to serious problems when traversing these 1494 * broken firewalls. 1495 * 1496 * With the default maxsockbuf of 256K, a scale factor 1497 * of 3 will be chosen by this algorithm. Those who 1498 * choose a larger maxsockbuf should watch out 1499 * for the compatibility problems mentioned above. 1500 * 1501 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1502 * or <SYN,ACK>) segment itself is never scaled. 1503 */ 1504 while (wscale < TCP_MAX_WINSHIFT && 1505 (TCP_MAXWIN << wscale) < sb_max) 1506 wscale++; 1507 sc->sc_requested_r_scale = wscale; 1508 sc->sc_requested_s_scale = to->to_wscale; 1509 sc->sc_flags |= SCF_WINSCALE; 1510 } 1511 } 1512 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1513 /* 1514 * If listening socket requested TCP digests, flag this in the 1515 * syncache so that syncache_respond() will do the right thing 1516 * with the SYN+ACK. 1517 */ 1518 if (ltflags & TF_SIGNATURE) 1519 sc->sc_flags |= SCF_SIGNATURE; 1520 #endif /* TCP_SIGNATURE */ 1521 if (to->to_flags & TOF_SACKPERM) 1522 sc->sc_flags |= SCF_SACK; 1523 if (to->to_flags & TOF_MSS) 1524 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1525 if (ltflags & TF_NOOPT) 1526 sc->sc_flags |= SCF_NOOPT; 1527 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1528 sc->sc_flags |= SCF_ECN; 1529 1530 if (V_tcp_syncookies) 1531 sc->sc_iss = syncookie_generate(sch, sc); 1532 #ifdef INET6 1533 if (autoflowlabel) { 1534 if (V_tcp_syncookies) 1535 sc->sc_flowlabel = sc->sc_iss; 1536 else 1537 sc->sc_flowlabel = ip6_randomflowlabel(); 1538 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1539 } 1540 #endif 1541 SCH_UNLOCK(sch); 1542 1543 #ifdef TCP_RFC7413 1544 if (tfo_cookie_valid) { 1545 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1546 /* INP_WUNLOCK(inp) will be performed by the caller */ 1547 rv = 1; 1548 goto tfo_expanded; 1549 } 1550 #endif 1551 1552 /* 1553 * Do a standard 3-way handshake. 1554 */ 1555 if (syncache_respond(sc, sch, 0, m) == 0) { 1556 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1557 syncache_free(sc); 1558 else if (sc != &scs) 1559 syncache_insert(sc, sch); /* locks and unlocks sch */ 1560 TCPSTAT_INC(tcps_sndacks); 1561 TCPSTAT_INC(tcps_sndtotal); 1562 } else { 1563 if (sc != &scs) 1564 syncache_free(sc); 1565 TCPSTAT_INC(tcps_sc_dropped); 1566 } 1567 1568 done: 1569 if (m) { 1570 *lsop = NULL; 1571 m_freem(m); 1572 } 1573 #ifdef TCP_RFC7413 1574 /* 1575 * If tfo_pending is not NULL here, then a TFO SYN that did not 1576 * result in a new socket was processed and the associated pending 1577 * counter has not yet been decremented. All such TFO processing paths 1578 * transit this point. 1579 */ 1580 if (tfo_pending != NULL) 1581 tcp_fastopen_decrement_counter(tfo_pending); 1582 1583 tfo_expanded: 1584 #endif 1585 if (cred != NULL) 1586 crfree(cred); 1587 #ifdef MAC 1588 if (sc == &scs) 1589 mac_syncache_destroy(&maclabel); 1590 #endif 1591 return (rv); 1592 } 1593 1594 /* 1595 * Send SYN|ACK to the peer. Either in response to the peer's SYN, 1596 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1597 */ 1598 static int 1599 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked, 1600 const struct mbuf *m0) 1601 { 1602 struct ip *ip = NULL; 1603 struct mbuf *m; 1604 struct tcphdr *th = NULL; 1605 int optlen, error = 0; /* Make compiler happy */ 1606 u_int16_t hlen, tlen, mssopt; 1607 struct tcpopt to; 1608 #ifdef INET6 1609 struct ip6_hdr *ip6 = NULL; 1610 #endif 1611 hlen = 1612 #ifdef INET6 1613 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1614 #endif 1615 sizeof(struct ip); 1616 tlen = hlen + sizeof(struct tcphdr); 1617 1618 /* Determine MSS we advertize to other end of connection. */ 1619 mssopt = tcp_mssopt(&sc->sc_inc); 1620 if (sc->sc_peer_mss) 1621 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss); 1622 1623 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1624 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1625 ("syncache: mbuf too small")); 1626 1627 /* Create the IP+TCP header from scratch. */ 1628 m = m_gethdr(M_NOWAIT, MT_DATA); 1629 if (m == NULL) 1630 return (ENOBUFS); 1631 #ifdef MAC 1632 mac_syncache_create_mbuf(sc->sc_label, m); 1633 #endif 1634 m->m_data += max_linkhdr; 1635 m->m_len = tlen; 1636 m->m_pkthdr.len = tlen; 1637 m->m_pkthdr.rcvif = NULL; 1638 1639 #ifdef INET6 1640 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1641 ip6 = mtod(m, struct ip6_hdr *); 1642 ip6->ip6_vfc = IPV6_VERSION; 1643 ip6->ip6_nxt = IPPROTO_TCP; 1644 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1645 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1646 ip6->ip6_plen = htons(tlen - hlen); 1647 /* ip6_hlim is set after checksum */ 1648 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1649 ip6->ip6_flow |= sc->sc_flowlabel; 1650 1651 th = (struct tcphdr *)(ip6 + 1); 1652 } 1653 #endif 1654 #if defined(INET6) && defined(INET) 1655 else 1656 #endif 1657 #ifdef INET 1658 { 1659 ip = mtod(m, struct ip *); 1660 ip->ip_v = IPVERSION; 1661 ip->ip_hl = sizeof(struct ip) >> 2; 1662 ip->ip_len = htons(tlen); 1663 ip->ip_id = 0; 1664 ip->ip_off = 0; 1665 ip->ip_sum = 0; 1666 ip->ip_p = IPPROTO_TCP; 1667 ip->ip_src = sc->sc_inc.inc_laddr; 1668 ip->ip_dst = sc->sc_inc.inc_faddr; 1669 ip->ip_ttl = sc->sc_ip_ttl; 1670 ip->ip_tos = sc->sc_ip_tos; 1671 1672 /* 1673 * See if we should do MTU discovery. Route lookups are 1674 * expensive, so we will only unset the DF bit if: 1675 * 1676 * 1) path_mtu_discovery is disabled 1677 * 2) the SCF_UNREACH flag has been set 1678 */ 1679 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1680 ip->ip_off |= htons(IP_DF); 1681 1682 th = (struct tcphdr *)(ip + 1); 1683 } 1684 #endif /* INET */ 1685 th->th_sport = sc->sc_inc.inc_lport; 1686 th->th_dport = sc->sc_inc.inc_fport; 1687 1688 th->th_seq = htonl(sc->sc_iss); 1689 th->th_ack = htonl(sc->sc_irs + 1); 1690 th->th_off = sizeof(struct tcphdr) >> 2; 1691 th->th_x2 = 0; 1692 th->th_flags = TH_SYN|TH_ACK; 1693 th->th_win = htons(sc->sc_wnd); 1694 th->th_urp = 0; 1695 1696 if (sc->sc_flags & SCF_ECN) { 1697 th->th_flags |= TH_ECE; 1698 TCPSTAT_INC(tcps_ecn_shs); 1699 } 1700 1701 /* Tack on the TCP options. */ 1702 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1703 to.to_flags = 0; 1704 1705 to.to_mss = mssopt; 1706 to.to_flags = TOF_MSS; 1707 if (sc->sc_flags & SCF_WINSCALE) { 1708 to.to_wscale = sc->sc_requested_r_scale; 1709 to.to_flags |= TOF_SCALE; 1710 } 1711 if (sc->sc_flags & SCF_TIMESTAMP) { 1712 /* Virgin timestamp or TCP cookie enhanced one. */ 1713 to.to_tsval = sc->sc_ts; 1714 to.to_tsecr = sc->sc_tsreflect; 1715 to.to_flags |= TOF_TS; 1716 } 1717 if (sc->sc_flags & SCF_SACK) 1718 to.to_flags |= TOF_SACKPERM; 1719 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1720 if (sc->sc_flags & SCF_SIGNATURE) 1721 to.to_flags |= TOF_SIGNATURE; 1722 #endif 1723 #ifdef TCP_RFC7413 1724 if (sc->sc_tfo_cookie) { 1725 to.to_flags |= TOF_FASTOPEN; 1726 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1727 to.to_tfo_cookie = sc->sc_tfo_cookie; 1728 /* don't send cookie again when retransmitting response */ 1729 sc->sc_tfo_cookie = NULL; 1730 } 1731 #endif 1732 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1733 1734 /* Adjust headers by option size. */ 1735 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1736 m->m_len += optlen; 1737 m->m_pkthdr.len += optlen; 1738 #ifdef INET6 1739 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1740 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1741 else 1742 #endif 1743 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1744 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1745 if (sc->sc_flags & SCF_SIGNATURE) { 1746 KASSERT(to.to_flags & TOF_SIGNATURE, 1747 ("tcp_addoptions() didn't set tcp_signature")); 1748 1749 /* NOTE: to.to_signature is inside of mbuf */ 1750 if (!TCPMD5_ENABLED() || 1751 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1752 m_freem(m); 1753 return (EACCES); 1754 } 1755 } 1756 #endif 1757 } else 1758 optlen = 0; 1759 1760 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1761 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1762 /* 1763 * If we have peer's SYN and it has a flowid, then let's assign it to 1764 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1765 * to SYN|ACK due to lack of inp here. 1766 */ 1767 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1768 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1769 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1770 } 1771 #ifdef INET6 1772 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1773 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1774 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1775 IPPROTO_TCP, 0); 1776 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1777 #ifdef TCP_OFFLOAD 1778 if (ADDED_BY_TOE(sc)) { 1779 struct toedev *tod = sc->sc_tod; 1780 1781 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1782 1783 return (error); 1784 } 1785 #endif 1786 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1787 } 1788 #endif 1789 #if defined(INET6) && defined(INET) 1790 else 1791 #endif 1792 #ifdef INET 1793 { 1794 m->m_pkthdr.csum_flags = CSUM_TCP; 1795 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1796 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1797 #ifdef TCP_OFFLOAD 1798 if (ADDED_BY_TOE(sc)) { 1799 struct toedev *tod = sc->sc_tod; 1800 1801 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1802 1803 return (error); 1804 } 1805 #endif 1806 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1807 } 1808 #endif 1809 return (error); 1810 } 1811 1812 /* 1813 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1814 * that exceed the capacity of the syncache by avoiding the storage of any 1815 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1816 * attacks where the attacker does not have access to our responses. 1817 * 1818 * Syncookies encode and include all necessary information about the 1819 * connection setup within the SYN|ACK that we send back. That way we 1820 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1821 * (if ever). Normally the syncache and syncookies are running in parallel 1822 * with the latter taking over when the former is exhausted. When matching 1823 * syncache entry is found the syncookie is ignored. 1824 * 1825 * The only reliable information persisting the 3WHS is our initial sequence 1826 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1827 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1828 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1829 * returns and signifies a legitimate connection if it matches the ACK. 1830 * 1831 * The available space of 32 bits to store the hash and to encode the SYN 1832 * option information is very tight and we should have at least 24 bits for 1833 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1834 * 1835 * SYN option information we have to encode to fully restore a connection: 1836 * MSS: is imporant to chose an optimal segment size to avoid IP level 1837 * fragmentation along the path. The common MSS values can be encoded 1838 * in a 3-bit table. Uncommon values are captured by the next lower value 1839 * in the table leading to a slight increase in packetization overhead. 1840 * WSCALE: is necessary to allow large windows to be used for high delay- 1841 * bandwidth product links. Not scaling the window when it was initially 1842 * negotiated is bad for performance as lack of scaling further decreases 1843 * the apparent available send window. We only need to encode the WSCALE 1844 * we received from the remote end. Our end can be recalculated at any 1845 * time. The common WSCALE values can be encoded in a 3-bit table. 1846 * Uncommon values are captured by the next lower value in the table 1847 * making us under-estimate the available window size halving our 1848 * theoretically possible maximum throughput for that connection. 1849 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1850 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1851 * that are included in all segments on a connection. We enable them when 1852 * the ACK has them. 1853 * 1854 * Security of syncookies and attack vectors: 1855 * 1856 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1857 * together with the gloabl secret to make it unique per connection attempt. 1858 * Thus any change of any of those parameters results in a different MAC output 1859 * in an unpredictable way unless a collision is encountered. 24 bits of the 1860 * MAC are embedded into the ISS. 1861 * 1862 * To prevent replay attacks two rotating global secrets are updated with a 1863 * new random value every 15 seconds. The life-time of a syncookie is thus 1864 * 15-30 seconds. 1865 * 1866 * Vector 1: Attacking the secret. This requires finding a weakness in the 1867 * MAC itself or the way it is used here. The attacker can do a chosen plain 1868 * text attack by varying and testing the all parameters under his control. 1869 * The strength depends on the size and randomness of the secret, and the 1870 * cryptographic security of the MAC function. Due to the constant updating 1871 * of the secret the attacker has at most 29.999 seconds to find the secret 1872 * and launch spoofed connections. After that he has to start all over again. 1873 * 1874 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1875 * size an average of 4,823 attempts are required for a 50% chance of success 1876 * to spoof a single syncookie (birthday collision paradox). However the 1877 * attacker is blind and doesn't know if one of his attempts succeeded unless 1878 * he has a side channel to interfere success from. A single connection setup 1879 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1880 * This many attempts are required for each one blind spoofed connection. For 1881 * every additional spoofed connection he has to launch another N attempts. 1882 * Thus for a sustained rate 100 spoofed connections per second approximately 1883 * 1,800,000 packets per second would have to be sent. 1884 * 1885 * NB: The MAC function should be fast so that it doesn't become a CPU 1886 * exhaustion attack vector itself. 1887 * 1888 * References: 1889 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1890 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1891 * http://cr.yp.to/syncookies.html (overview) 1892 * http://cr.yp.to/syncookies/archive (details) 1893 * 1894 * 1895 * Schematic construction of a syncookie enabled Initial Sequence Number: 1896 * 0 1 2 3 1897 * 12345678901234567890123456789012 1898 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1899 * 1900 * x 24 MAC (truncated) 1901 * W 3 Send Window Scale index 1902 * M 3 MSS index 1903 * S 1 SACK permitted 1904 * P 1 Odd/even secret 1905 */ 1906 1907 /* 1908 * Distribution and probability of certain MSS values. Those in between are 1909 * rounded down to the next lower one. 1910 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1911 * .2% .3% 5% 7% 7% 20% 15% 45% 1912 */ 1913 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1914 1915 /* 1916 * Distribution and probability of certain WSCALE values. We have to map the 1917 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1918 * bits based on prevalence of certain values. Where we don't have an exact 1919 * match for are rounded down to the next lower one letting us under-estimate 1920 * the true available window. At the moment this would happen only for the 1921 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1922 * and window size). The absence of the WSCALE option (no scaling in either 1923 * direction) is encoded with index zero. 1924 * [WSCALE values histograms, Allman, 2012] 1925 * X 10 10 35 5 6 14 10% by host 1926 * X 11 4 5 5 18 49 3% by connections 1927 */ 1928 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1929 1930 /* 1931 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1932 * and good cryptographic properties. 1933 */ 1934 static uint32_t 1935 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1936 uint8_t *secbits, uintptr_t secmod) 1937 { 1938 SIPHASH_CTX ctx; 1939 uint32_t siphash[2]; 1940 1941 SipHash24_Init(&ctx); 1942 SipHash_SetKey(&ctx, secbits); 1943 switch (inc->inc_flags & INC_ISIPV6) { 1944 #ifdef INET 1945 case 0: 1946 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1947 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1948 break; 1949 #endif 1950 #ifdef INET6 1951 case INC_ISIPV6: 1952 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1953 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1954 break; 1955 #endif 1956 } 1957 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1958 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1959 SipHash_Update(&ctx, &irs, sizeof(irs)); 1960 SipHash_Update(&ctx, &flags, sizeof(flags)); 1961 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1962 SipHash_Final((u_int8_t *)&siphash, &ctx); 1963 1964 return (siphash[0] ^ siphash[1]); 1965 } 1966 1967 static tcp_seq 1968 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 1969 { 1970 u_int i, mss, secbit, wscale; 1971 uint32_t iss, hash; 1972 uint8_t *secbits; 1973 union syncookie cookie; 1974 1975 SCH_LOCK_ASSERT(sch); 1976 1977 cookie.cookie = 0; 1978 1979 /* Map our computed MSS into the 3-bit index. */ 1980 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss)); 1981 for (i = nitems(tcp_sc_msstab) - 1; tcp_sc_msstab[i] > mss && i > 0; 1982 i--) 1983 ; 1984 cookie.flags.mss_idx = i; 1985 1986 /* 1987 * Map the send window scale into the 3-bit index but only if 1988 * the wscale option was received. 1989 */ 1990 if (sc->sc_flags & SCF_WINSCALE) { 1991 wscale = sc->sc_requested_s_scale; 1992 for (i = nitems(tcp_sc_wstab) - 1; 1993 tcp_sc_wstab[i] > wscale && i > 0; 1994 i--) 1995 ; 1996 cookie.flags.wscale_idx = i; 1997 } 1998 1999 /* Can we do SACK? */ 2000 if (sc->sc_flags & SCF_SACK) 2001 cookie.flags.sack_ok = 1; 2002 2003 /* Which of the two secrets to use. */ 2004 secbit = sch->sch_sc->secret.oddeven & 0x1; 2005 cookie.flags.odd_even = secbit; 2006 2007 secbits = sch->sch_sc->secret.key[secbit]; 2008 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2009 (uintptr_t)sch); 2010 2011 /* 2012 * Put the flags into the hash and XOR them to get better ISS number 2013 * variance. This doesn't enhance the cryptographic strength and is 2014 * done to prevent the 8 cookie bits from showing up directly on the 2015 * wire. 2016 */ 2017 iss = hash & ~0xff; 2018 iss |= cookie.cookie ^ (hash >> 24); 2019 2020 /* Randomize the timestamp. */ 2021 if (sc->sc_flags & SCF_TIMESTAMP) { 2022 sc->sc_ts = arc4random(); 2023 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 2024 } 2025 2026 TCPSTAT_INC(tcps_sc_sendcookie); 2027 return (iss); 2028 } 2029 2030 static struct syncache * 2031 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2032 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2033 struct socket *lso) 2034 { 2035 uint32_t hash; 2036 uint8_t *secbits; 2037 tcp_seq ack, seq; 2038 int wnd, wscale = 0; 2039 union syncookie cookie; 2040 2041 SCH_LOCK_ASSERT(sch); 2042 2043 /* 2044 * Pull information out of SYN-ACK/ACK and revert sequence number 2045 * advances. 2046 */ 2047 ack = th->th_ack - 1; 2048 seq = th->th_seq - 1; 2049 2050 /* 2051 * Unpack the flags containing enough information to restore the 2052 * connection. 2053 */ 2054 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2055 2056 /* Which of the two secrets to use. */ 2057 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 2058 2059 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2060 2061 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2062 if ((ack & ~0xff) != (hash & ~0xff)) 2063 return (NULL); 2064 2065 /* Fill in the syncache values. */ 2066 sc->sc_flags = 0; 2067 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2068 sc->sc_ipopts = NULL; 2069 2070 sc->sc_irs = seq; 2071 sc->sc_iss = ack; 2072 2073 switch (inc->inc_flags & INC_ISIPV6) { 2074 #ifdef INET 2075 case 0: 2076 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2077 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2078 break; 2079 #endif 2080 #ifdef INET6 2081 case INC_ISIPV6: 2082 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2083 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 2084 break; 2085 #endif 2086 } 2087 2088 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2089 2090 /* We can simply recompute receive window scale we sent earlier. */ 2091 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2092 wscale++; 2093 2094 /* Only use wscale if it was enabled in the orignal SYN. */ 2095 if (cookie.flags.wscale_idx > 0) { 2096 sc->sc_requested_r_scale = wscale; 2097 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2098 sc->sc_flags |= SCF_WINSCALE; 2099 } 2100 2101 wnd = sbspace(&lso->so_rcv); 2102 wnd = imax(wnd, 0); 2103 wnd = imin(wnd, TCP_MAXWIN); 2104 sc->sc_wnd = wnd; 2105 2106 if (cookie.flags.sack_ok) 2107 sc->sc_flags |= SCF_SACK; 2108 2109 if (to->to_flags & TOF_TS) { 2110 sc->sc_flags |= SCF_TIMESTAMP; 2111 sc->sc_tsreflect = to->to_tsval; 2112 sc->sc_ts = to->to_tsecr; 2113 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 2114 } 2115 2116 if (to->to_flags & TOF_SIGNATURE) 2117 sc->sc_flags |= SCF_SIGNATURE; 2118 2119 sc->sc_rxmits = 0; 2120 2121 TCPSTAT_INC(tcps_sc_recvcookie); 2122 return (sc); 2123 } 2124 2125 #ifdef INVARIANTS 2126 static int 2127 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2128 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2129 struct socket *lso) 2130 { 2131 struct syncache scs, *scx; 2132 char *s; 2133 2134 bzero(&scs, sizeof(scs)); 2135 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2136 2137 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2138 return (0); 2139 2140 if (scx != NULL) { 2141 if (sc->sc_peer_mss != scx->sc_peer_mss) 2142 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2143 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2144 2145 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2146 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2147 s, __func__, sc->sc_requested_r_scale, 2148 scx->sc_requested_r_scale); 2149 2150 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2151 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2152 s, __func__, sc->sc_requested_s_scale, 2153 scx->sc_requested_s_scale); 2154 2155 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2156 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2157 } 2158 2159 if (s != NULL) 2160 free(s, M_TCPLOG); 2161 return (0); 2162 } 2163 #endif /* INVARIANTS */ 2164 2165 static void 2166 syncookie_reseed(void *arg) 2167 { 2168 struct tcp_syncache *sc = arg; 2169 uint8_t *secbits; 2170 int secbit; 2171 2172 /* 2173 * Reseeding the secret doesn't have to be protected by a lock. 2174 * It only must be ensured that the new random values are visible 2175 * to all CPUs in a SMP environment. The atomic with release 2176 * semantics ensures that. 2177 */ 2178 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2179 secbits = sc->secret.key[secbit]; 2180 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2181 atomic_add_rel_int(&sc->secret.oddeven, 1); 2182 2183 /* Reschedule ourself. */ 2184 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2185 } 2186 2187 /* 2188 * Exports the syncache entries to userland so that netstat can display 2189 * them alongside the other sockets. This function is intended to be 2190 * called only from tcp_pcblist. 2191 * 2192 * Due to concurrency on an active system, the number of pcbs exported 2193 * may have no relation to max_pcbs. max_pcbs merely indicates the 2194 * amount of space the caller allocated for this function to use. 2195 */ 2196 int 2197 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2198 { 2199 struct xtcpcb xt; 2200 struct syncache *sc; 2201 struct syncache_head *sch; 2202 int count, error, i; 2203 2204 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2205 sch = &V_tcp_syncache.hashbase[i]; 2206 SCH_LOCK(sch); 2207 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2208 if (count >= max_pcbs) { 2209 SCH_UNLOCK(sch); 2210 goto exit; 2211 } 2212 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2213 continue; 2214 bzero(&xt, sizeof(xt)); 2215 xt.xt_len = sizeof(xt); 2216 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2217 xt.xt_inp.inp_vflag = INP_IPV6; 2218 else 2219 xt.xt_inp.inp_vflag = INP_IPV4; 2220 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 2221 xt.xt_tp.t_inpcb = &xt.xt_inp; 2222 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 2223 xt.xt_socket.xso_protocol = IPPROTO_TCP; 2224 xt.xt_socket.xso_len = sizeof (struct xsocket); 2225 xt.xt_socket.so_type = SOCK_STREAM; 2226 xt.xt_socket.so_state = SS_ISCONNECTING; 2227 error = SYSCTL_OUT(req, &xt, sizeof xt); 2228 if (error) { 2229 SCH_UNLOCK(sch); 2230 goto exit; 2231 } 2232 count++; 2233 } 2234 SCH_UNLOCK(sch); 2235 } 2236 exit: 2237 *pcbs_exported = count; 2238 return error; 2239 } 2240