1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2001 McAfee, Inc. 5 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG 6 * All rights reserved. 7 * 8 * This software was developed for the FreeBSD Project by Jonathan Lemon 9 * and McAfee Research, the Security Research Division of McAfee, Inc. under 10 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 11 * DARPA CHATS research program. [2001 McAfee, Inc.] 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include "opt_inet.h" 39 #include "opt_inet6.h" 40 #include "opt_ipsec.h" 41 #include "opt_pcbgroup.h" 42 43 #include <sys/param.h> 44 #include <sys/systm.h> 45 #include <sys/hash.h> 46 #include <sys/refcount.h> 47 #include <sys/kernel.h> 48 #include <sys/sysctl.h> 49 #include <sys/limits.h> 50 #include <sys/lock.h> 51 #include <sys/mutex.h> 52 #include <sys/malloc.h> 53 #include <sys/mbuf.h> 54 #include <sys/proc.h> /* for proc0 declaration */ 55 #include <sys/random.h> 56 #include <sys/socket.h> 57 #include <sys/socketvar.h> 58 #include <sys/syslog.h> 59 #include <sys/ucred.h> 60 61 #include <sys/md5.h> 62 #include <crypto/siphash/siphash.h> 63 64 #include <vm/uma.h> 65 66 #include <net/if.h> 67 #include <net/if_var.h> 68 #include <net/route.h> 69 #include <net/vnet.h> 70 71 #include <netinet/in.h> 72 #include <netinet/in_kdtrace.h> 73 #include <netinet/in_systm.h> 74 #include <netinet/ip.h> 75 #include <netinet/in_var.h> 76 #include <netinet/in_pcb.h> 77 #include <netinet/ip_var.h> 78 #include <netinet/ip_options.h> 79 #ifdef INET6 80 #include <netinet/ip6.h> 81 #include <netinet/icmp6.h> 82 #include <netinet6/nd6.h> 83 #include <netinet6/ip6_var.h> 84 #include <netinet6/in6_pcb.h> 85 #endif 86 #include <netinet/tcp.h> 87 #include <netinet/tcp_fastopen.h> 88 #include <netinet/tcp_fsm.h> 89 #include <netinet/tcp_seq.h> 90 #include <netinet/tcp_timer.h> 91 #include <netinet/tcp_var.h> 92 #include <netinet/tcp_syncache.h> 93 #ifdef INET6 94 #include <netinet6/tcp6_var.h> 95 #endif 96 #ifdef TCP_OFFLOAD 97 #include <netinet/toecore.h> 98 #endif 99 100 #include <netipsec/ipsec_support.h> 101 102 #include <machine/in_cksum.h> 103 104 #include <security/mac/mac_framework.h> 105 106 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1; 107 #define V_tcp_syncookies VNET(tcp_syncookies) 108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 109 &VNET_NAME(tcp_syncookies), 0, 110 "Use TCP SYN cookies if the syncache overflows"); 111 112 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0; 113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 115 &VNET_NAME(tcp_syncookiesonly), 0, 116 "Use only TCP SYN cookies"); 117 118 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1; 119 #define V_functions_inherit_listen_socket_stack \ 120 VNET(functions_inherit_listen_socket_stack) 121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack, 122 CTLFLAG_VNET | CTLFLAG_RW, 123 &VNET_NAME(functions_inherit_listen_socket_stack), 0, 124 "Inherit listen socket's stack"); 125 126 #ifdef TCP_OFFLOAD 127 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 128 #endif 129 130 static void syncache_drop(struct syncache *, struct syncache_head *); 131 static void syncache_free(struct syncache *); 132 static void syncache_insert(struct syncache *, struct syncache_head *); 133 static int syncache_respond(struct syncache *, struct syncache_head *, int, 134 const struct mbuf *); 135 static struct socket *syncache_socket(struct syncache *, struct socket *, 136 struct mbuf *m); 137 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 138 int docallout); 139 static void syncache_timer(void *); 140 141 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 142 uint8_t *, uintptr_t); 143 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 144 static struct syncache 145 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 146 struct syncache *, struct tcphdr *, struct tcpopt *, 147 struct socket *); 148 static void syncookie_reseed(void *); 149 #ifdef INVARIANTS 150 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 151 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 152 struct socket *lso); 153 #endif 154 155 /* 156 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 157 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds, 158 * the odds are that the user has given up attempting to connect by then. 159 */ 160 #define SYNCACHE_MAXREXMTS 3 161 162 /* Arbitrary values */ 163 #define TCP_SYNCACHE_HASHSIZE 512 164 #define TCP_SYNCACHE_BUCKETLIMIT 30 165 166 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 167 #define V_tcp_syncache VNET(tcp_syncache) 168 169 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, 170 "TCP SYN cache"); 171 172 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 173 &VNET_NAME(tcp_syncache.bucket_limit), 0, 174 "Per-bucket hash limit for syncache"); 175 176 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 177 &VNET_NAME(tcp_syncache.cache_limit), 0, 178 "Overall entry limit for syncache"); 179 180 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 181 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 182 183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 184 &VNET_NAME(tcp_syncache.hashsize), 0, 185 "Size of TCP syncache hashtable"); 186 187 static int 188 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 189 { 190 int error; 191 u_int new; 192 193 new = V_tcp_syncache.rexmt_limit; 194 error = sysctl_handle_int(oidp, &new, 0, req); 195 if ((error == 0) && (req->newptr != NULL)) { 196 if (new > TCP_MAXRXTSHIFT) 197 error = EINVAL; 198 else 199 V_tcp_syncache.rexmt_limit = new; 200 } 201 return (error); 202 } 203 204 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 205 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW, 206 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 207 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI", 208 "Limit on SYN/ACK retransmissions"); 209 210 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 211 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 212 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 213 "Send reset on socket allocation failure"); 214 215 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 216 217 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 218 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 219 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 220 221 /* 222 * Requires the syncache entry to be already removed from the bucket list. 223 */ 224 static void 225 syncache_free(struct syncache *sc) 226 { 227 228 if (sc->sc_ipopts) 229 (void) m_free(sc->sc_ipopts); 230 if (sc->sc_cred) 231 crfree(sc->sc_cred); 232 #ifdef MAC 233 mac_syncache_destroy(&sc->sc_label); 234 #endif 235 236 uma_zfree(V_tcp_syncache.zone, sc); 237 } 238 239 void 240 syncache_init(void) 241 { 242 int i; 243 244 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 245 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 246 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 247 V_tcp_syncache.hash_secret = arc4random(); 248 249 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 250 &V_tcp_syncache.hashsize); 251 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 252 &V_tcp_syncache.bucket_limit); 253 if (!powerof2(V_tcp_syncache.hashsize) || 254 V_tcp_syncache.hashsize == 0) { 255 printf("WARNING: syncache hash size is not a power of 2.\n"); 256 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 257 } 258 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 259 260 /* Set limits. */ 261 V_tcp_syncache.cache_limit = 262 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 263 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 264 &V_tcp_syncache.cache_limit); 265 266 /* Allocate the hash table. */ 267 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 268 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 269 270 #ifdef VIMAGE 271 V_tcp_syncache.vnet = curvnet; 272 #endif 273 274 /* Initialize the hash buckets. */ 275 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 276 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 277 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 278 NULL, MTX_DEF); 279 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 280 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 281 V_tcp_syncache.hashbase[i].sch_length = 0; 282 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 283 V_tcp_syncache.hashbase[i].sch_last_overflow = 284 -(SYNCOOKIE_LIFETIME + 1); 285 } 286 287 /* Create the syncache entry zone. */ 288 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 289 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 290 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 291 V_tcp_syncache.cache_limit); 292 293 /* Start the SYN cookie reseeder callout. */ 294 callout_init(&V_tcp_syncache.secret.reseed, 1); 295 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 296 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 297 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 298 syncookie_reseed, &V_tcp_syncache); 299 } 300 301 #ifdef VIMAGE 302 void 303 syncache_destroy(void) 304 { 305 struct syncache_head *sch; 306 struct syncache *sc, *nsc; 307 int i; 308 309 /* 310 * Stop the re-seed timer before freeing resources. No need to 311 * possibly schedule it another time. 312 */ 313 callout_drain(&V_tcp_syncache.secret.reseed); 314 315 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 316 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 317 318 sch = &V_tcp_syncache.hashbase[i]; 319 callout_drain(&sch->sch_timer); 320 321 SCH_LOCK(sch); 322 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 323 syncache_drop(sc, sch); 324 SCH_UNLOCK(sch); 325 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 326 ("%s: sch->sch_bucket not empty", __func__)); 327 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 328 __func__, sch->sch_length)); 329 mtx_destroy(&sch->sch_mtx); 330 } 331 332 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 333 ("%s: cache_count not 0", __func__)); 334 335 /* Free the allocated global resources. */ 336 uma_zdestroy(V_tcp_syncache.zone); 337 free(V_tcp_syncache.hashbase, M_SYNCACHE); 338 } 339 #endif 340 341 /* 342 * Inserts a syncache entry into the specified bucket row. 343 * Locks and unlocks the syncache_head autonomously. 344 */ 345 static void 346 syncache_insert(struct syncache *sc, struct syncache_head *sch) 347 { 348 struct syncache *sc2; 349 350 SCH_LOCK(sch); 351 352 /* 353 * Make sure that we don't overflow the per-bucket limit. 354 * If the bucket is full, toss the oldest element. 355 */ 356 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 357 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 358 ("sch->sch_length incorrect")); 359 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 360 sch->sch_last_overflow = time_uptime; 361 syncache_drop(sc2, sch); 362 TCPSTAT_INC(tcps_sc_bucketoverflow); 363 } 364 365 /* Put it into the bucket. */ 366 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 367 sch->sch_length++; 368 369 #ifdef TCP_OFFLOAD 370 if (ADDED_BY_TOE(sc)) { 371 struct toedev *tod = sc->sc_tod; 372 373 tod->tod_syncache_added(tod, sc->sc_todctx); 374 } 375 #endif 376 377 /* Reinitialize the bucket row's timer. */ 378 if (sch->sch_length == 1) 379 sch->sch_nextc = ticks + INT_MAX; 380 syncache_timeout(sc, sch, 1); 381 382 SCH_UNLOCK(sch); 383 384 TCPSTATES_INC(TCPS_SYN_RECEIVED); 385 TCPSTAT_INC(tcps_sc_added); 386 } 387 388 /* 389 * Remove and free entry from syncache bucket row. 390 * Expects locked syncache head. 391 */ 392 static void 393 syncache_drop(struct syncache *sc, struct syncache_head *sch) 394 { 395 396 SCH_LOCK_ASSERT(sch); 397 398 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 399 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 400 sch->sch_length--; 401 402 #ifdef TCP_OFFLOAD 403 if (ADDED_BY_TOE(sc)) { 404 struct toedev *tod = sc->sc_tod; 405 406 tod->tod_syncache_removed(tod, sc->sc_todctx); 407 } 408 #endif 409 410 syncache_free(sc); 411 } 412 413 /* 414 * Engage/reengage time on bucket row. 415 */ 416 static void 417 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 418 { 419 int rexmt; 420 421 if (sc->sc_rxmits == 0) 422 rexmt = TCPTV_RTOBASE; 423 else 424 TCPT_RANGESET(rexmt, TCPTV_RTOBASE * tcp_syn_backoff[sc->sc_rxmits], 425 tcp_rexmit_min, TCPTV_REXMTMAX); 426 sc->sc_rxttime = ticks + rexmt; 427 sc->sc_rxmits++; 428 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 429 sch->sch_nextc = sc->sc_rxttime; 430 if (docallout) 431 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 432 syncache_timer, (void *)sch); 433 } 434 } 435 436 /* 437 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 438 * If we have retransmitted an entry the maximum number of times, expire it. 439 * One separate timer for each bucket row. 440 */ 441 static void 442 syncache_timer(void *xsch) 443 { 444 struct syncache_head *sch = (struct syncache_head *)xsch; 445 struct syncache *sc, *nsc; 446 int tick = ticks; 447 char *s; 448 449 CURVNET_SET(sch->sch_sc->vnet); 450 451 /* NB: syncache_head has already been locked by the callout. */ 452 SCH_LOCK_ASSERT(sch); 453 454 /* 455 * In the following cycle we may remove some entries and/or 456 * advance some timeouts, so re-initialize the bucket timer. 457 */ 458 sch->sch_nextc = tick + INT_MAX; 459 460 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 461 /* 462 * We do not check if the listen socket still exists 463 * and accept the case where the listen socket may be 464 * gone by the time we resend the SYN/ACK. We do 465 * not expect this to happens often. If it does, 466 * then the RST will be sent by the time the remote 467 * host does the SYN/ACK->ACK. 468 */ 469 if (TSTMP_GT(sc->sc_rxttime, tick)) { 470 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 471 sch->sch_nextc = sc->sc_rxttime; 472 continue; 473 } 474 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 475 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 476 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 477 "giving up and removing syncache entry\n", 478 s, __func__); 479 free(s, M_TCPLOG); 480 } 481 syncache_drop(sc, sch); 482 TCPSTAT_INC(tcps_sc_stale); 483 continue; 484 } 485 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 486 log(LOG_DEBUG, "%s; %s: Response timeout, " 487 "retransmitting (%u) SYN|ACK\n", 488 s, __func__, sc->sc_rxmits); 489 free(s, M_TCPLOG); 490 } 491 492 syncache_respond(sc, sch, 1, NULL); 493 TCPSTAT_INC(tcps_sc_retransmitted); 494 syncache_timeout(sc, sch, 0); 495 } 496 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 497 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 498 syncache_timer, (void *)(sch)); 499 CURVNET_RESTORE(); 500 } 501 502 /* 503 * Find an entry in the syncache. 504 * Returns always with locked syncache_head plus a matching entry or NULL. 505 */ 506 static struct syncache * 507 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 508 { 509 struct syncache *sc; 510 struct syncache_head *sch; 511 uint32_t hash; 512 513 /* 514 * The hash is built on foreign port + local port + foreign address. 515 * We rely on the fact that struct in_conninfo starts with 16 bits 516 * of foreign port, then 16 bits of local port then followed by 128 517 * bits of foreign address. In case of IPv4 address, the first 3 518 * 32-bit words of the address always are zeroes. 519 */ 520 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 521 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 522 523 sch = &V_tcp_syncache.hashbase[hash]; 524 *schp = sch; 525 SCH_LOCK(sch); 526 527 /* Circle through bucket row to find matching entry. */ 528 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 529 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 530 sizeof(struct in_endpoints)) == 0) 531 break; 532 533 return (sc); /* Always returns with locked sch. */ 534 } 535 536 /* 537 * This function is called when we get a RST for a 538 * non-existent connection, so that we can see if the 539 * connection is in the syn cache. If it is, zap it. 540 */ 541 void 542 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 543 { 544 struct syncache *sc; 545 struct syncache_head *sch; 546 char *s = NULL; 547 548 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 549 SCH_LOCK_ASSERT(sch); 550 551 /* 552 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 553 * See RFC 793 page 65, section SEGMENT ARRIVES. 554 */ 555 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 556 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 557 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 558 "FIN flag set, segment ignored\n", s, __func__); 559 TCPSTAT_INC(tcps_badrst); 560 goto done; 561 } 562 563 /* 564 * No corresponding connection was found in syncache. 565 * If syncookies are enabled and possibly exclusively 566 * used, or we are under memory pressure, a valid RST 567 * may not find a syncache entry. In that case we're 568 * done and no SYN|ACK retransmissions will happen. 569 * Otherwise the RST was misdirected or spoofed. 570 */ 571 if (sc == NULL) { 572 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 573 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 574 "syncache entry (possibly syncookie only), " 575 "segment ignored\n", s, __func__); 576 TCPSTAT_INC(tcps_badrst); 577 goto done; 578 } 579 580 /* 581 * If the RST bit is set, check the sequence number to see 582 * if this is a valid reset segment. 583 * RFC 793 page 37: 584 * In all states except SYN-SENT, all reset (RST) segments 585 * are validated by checking their SEQ-fields. A reset is 586 * valid if its sequence number is in the window. 587 * 588 * The sequence number in the reset segment is normally an 589 * echo of our outgoing acknowlegement numbers, but some hosts 590 * send a reset with the sequence number at the rightmost edge 591 * of our receive window, and we have to handle this case. 592 */ 593 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 594 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 595 syncache_drop(sc, sch); 596 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 597 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 598 "connection attempt aborted by remote endpoint\n", 599 s, __func__); 600 TCPSTAT_INC(tcps_sc_reset); 601 } else { 602 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 603 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 604 "IRS %u (+WND %u), segment ignored\n", 605 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 606 TCPSTAT_INC(tcps_badrst); 607 } 608 609 done: 610 if (s != NULL) 611 free(s, M_TCPLOG); 612 SCH_UNLOCK(sch); 613 } 614 615 void 616 syncache_badack(struct in_conninfo *inc) 617 { 618 struct syncache *sc; 619 struct syncache_head *sch; 620 621 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 622 SCH_LOCK_ASSERT(sch); 623 if (sc != NULL) { 624 syncache_drop(sc, sch); 625 TCPSTAT_INC(tcps_sc_badack); 626 } 627 SCH_UNLOCK(sch); 628 } 629 630 void 631 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq) 632 { 633 struct syncache *sc; 634 struct syncache_head *sch; 635 636 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 637 SCH_LOCK_ASSERT(sch); 638 if (sc == NULL) 639 goto done; 640 641 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 642 if (ntohl(th_seq) != sc->sc_iss) 643 goto done; 644 645 /* 646 * If we've rertransmitted 3 times and this is our second error, 647 * we remove the entry. Otherwise, we allow it to continue on. 648 * This prevents us from incorrectly nuking an entry during a 649 * spurious network outage. 650 * 651 * See tcp_notify(). 652 */ 653 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 654 sc->sc_flags |= SCF_UNREACH; 655 goto done; 656 } 657 syncache_drop(sc, sch); 658 TCPSTAT_INC(tcps_sc_unreach); 659 done: 660 SCH_UNLOCK(sch); 661 } 662 663 /* 664 * Build a new TCP socket structure from a syncache entry. 665 * 666 * On success return the newly created socket with its underlying inp locked. 667 */ 668 static struct socket * 669 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 670 { 671 struct tcp_function_block *blk; 672 struct inpcb *inp = NULL; 673 struct socket *so; 674 struct tcpcb *tp; 675 int error; 676 char *s; 677 678 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 679 680 /* 681 * Ok, create the full blown connection, and set things up 682 * as they would have been set up if we had created the 683 * connection when the SYN arrived. If we can't create 684 * the connection, abort it. 685 */ 686 so = sonewconn(lso, 0); 687 if (so == NULL) { 688 /* 689 * Drop the connection; we will either send a RST or 690 * have the peer retransmit its SYN again after its 691 * RTO and try again. 692 */ 693 TCPSTAT_INC(tcps_listendrop); 694 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 695 log(LOG_DEBUG, "%s; %s: Socket create failed " 696 "due to limits or memory shortage\n", 697 s, __func__); 698 free(s, M_TCPLOG); 699 } 700 goto abort2; 701 } 702 #ifdef MAC 703 mac_socketpeer_set_from_mbuf(m, so); 704 #endif 705 706 inp = sotoinpcb(so); 707 inp->inp_inc.inc_fibnum = so->so_fibnum; 708 INP_WLOCK(inp); 709 /* 710 * Exclusive pcbinfo lock is not required in syncache socket case even 711 * if two inpcb locks can be acquired simultaneously: 712 * - the inpcb in LISTEN state, 713 * - the newly created inp. 714 * 715 * In this case, an inp cannot be at same time in LISTEN state and 716 * just created by an accept() call. 717 */ 718 INP_HASH_WLOCK(&V_tcbinfo); 719 720 /* Insert new socket into PCB hash list. */ 721 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 722 #ifdef INET6 723 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 724 inp->inp_vflag &= ~INP_IPV4; 725 inp->inp_vflag |= INP_IPV6; 726 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 727 } else { 728 inp->inp_vflag &= ~INP_IPV6; 729 inp->inp_vflag |= INP_IPV4; 730 #endif 731 inp->inp_laddr = sc->sc_inc.inc_laddr; 732 #ifdef INET6 733 } 734 #endif 735 736 /* 737 * If there's an mbuf and it has a flowid, then let's initialise the 738 * inp with that particular flowid. 739 */ 740 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 741 inp->inp_flowid = m->m_pkthdr.flowid; 742 inp->inp_flowtype = M_HASHTYPE_GET(m); 743 } 744 745 /* 746 * Install in the reservation hash table for now, but don't yet 747 * install a connection group since the full 4-tuple isn't yet 748 * configured. 749 */ 750 inp->inp_lport = sc->sc_inc.inc_lport; 751 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 752 /* 753 * Undo the assignments above if we failed to 754 * put the PCB on the hash lists. 755 */ 756 #ifdef INET6 757 if (sc->sc_inc.inc_flags & INC_ISIPV6) 758 inp->in6p_laddr = in6addr_any; 759 else 760 #endif 761 inp->inp_laddr.s_addr = INADDR_ANY; 762 inp->inp_lport = 0; 763 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 764 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 765 "with error %i\n", 766 s, __func__, error); 767 free(s, M_TCPLOG); 768 } 769 INP_HASH_WUNLOCK(&V_tcbinfo); 770 goto abort; 771 } 772 #ifdef INET6 773 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 774 struct inpcb *oinp = sotoinpcb(lso); 775 struct in6_addr laddr6; 776 struct sockaddr_in6 sin6; 777 /* 778 * Inherit socket options from the listening socket. 779 * Note that in6p_inputopts are not (and should not be) 780 * copied, since it stores previously received options and is 781 * used to detect if each new option is different than the 782 * previous one and hence should be passed to a user. 783 * If we copied in6p_inputopts, a user would not be able to 784 * receive options just after calling the accept system call. 785 */ 786 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 787 if (oinp->in6p_outputopts) 788 inp->in6p_outputopts = 789 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 790 791 sin6.sin6_family = AF_INET6; 792 sin6.sin6_len = sizeof(sin6); 793 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 794 sin6.sin6_port = sc->sc_inc.inc_fport; 795 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 796 laddr6 = inp->in6p_laddr; 797 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 798 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 799 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 800 thread0.td_ucred, m)) != 0) { 801 inp->in6p_laddr = laddr6; 802 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 803 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 804 "with error %i\n", 805 s, __func__, error); 806 free(s, M_TCPLOG); 807 } 808 INP_HASH_WUNLOCK(&V_tcbinfo); 809 goto abort; 810 } 811 /* Override flowlabel from in6_pcbconnect. */ 812 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 813 inp->inp_flow |= sc->sc_flowlabel; 814 } 815 #endif /* INET6 */ 816 #if defined(INET) && defined(INET6) 817 else 818 #endif 819 #ifdef INET 820 { 821 struct in_addr laddr; 822 struct sockaddr_in sin; 823 824 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 825 826 if (inp->inp_options == NULL) { 827 inp->inp_options = sc->sc_ipopts; 828 sc->sc_ipopts = NULL; 829 } 830 831 sin.sin_family = AF_INET; 832 sin.sin_len = sizeof(sin); 833 sin.sin_addr = sc->sc_inc.inc_faddr; 834 sin.sin_port = sc->sc_inc.inc_fport; 835 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 836 laddr = inp->inp_laddr; 837 if (inp->inp_laddr.s_addr == INADDR_ANY) 838 inp->inp_laddr = sc->sc_inc.inc_laddr; 839 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 840 thread0.td_ucred, m)) != 0) { 841 inp->inp_laddr = laddr; 842 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 843 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 844 "with error %i\n", 845 s, __func__, error); 846 free(s, M_TCPLOG); 847 } 848 INP_HASH_WUNLOCK(&V_tcbinfo); 849 goto abort; 850 } 851 } 852 #endif /* INET */ 853 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 854 /* Copy old policy into new socket's. */ 855 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 856 printf("syncache_socket: could not copy policy\n"); 857 #endif 858 INP_HASH_WUNLOCK(&V_tcbinfo); 859 tp = intotcpcb(inp); 860 tcp_state_change(tp, TCPS_SYN_RECEIVED); 861 tp->iss = sc->sc_iss; 862 tp->irs = sc->sc_irs; 863 tcp_rcvseqinit(tp); 864 tcp_sendseqinit(tp); 865 blk = sototcpcb(lso)->t_fb; 866 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 867 /* 868 * Our parents t_fb was not the default, 869 * we need to release our ref on tp->t_fb and 870 * pickup one on the new entry. 871 */ 872 struct tcp_function_block *rblk; 873 874 rblk = find_and_ref_tcp_fb(blk); 875 KASSERT(rblk != NULL, 876 ("cannot find blk %p out of syncache?", blk)); 877 if (tp->t_fb->tfb_tcp_fb_fini) 878 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 879 refcount_release(&tp->t_fb->tfb_refcnt); 880 tp->t_fb = rblk; 881 /* 882 * XXXrrs this is quite dangerous, it is possible 883 * for the new function to fail to init. We also 884 * are not asking if the handoff_is_ok though at 885 * the very start thats probalbly ok. 886 */ 887 if (tp->t_fb->tfb_tcp_fb_init) { 888 (*tp->t_fb->tfb_tcp_fb_init)(tp); 889 } 890 } 891 tp->snd_wl1 = sc->sc_irs; 892 tp->snd_max = tp->iss + 1; 893 tp->snd_nxt = tp->iss + 1; 894 tp->rcv_up = sc->sc_irs + 1; 895 tp->rcv_wnd = sc->sc_wnd; 896 tp->rcv_adv += tp->rcv_wnd; 897 tp->last_ack_sent = tp->rcv_nxt; 898 899 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 900 if (sc->sc_flags & SCF_NOOPT) 901 tp->t_flags |= TF_NOOPT; 902 else { 903 if (sc->sc_flags & SCF_WINSCALE) { 904 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 905 tp->snd_scale = sc->sc_requested_s_scale; 906 tp->request_r_scale = sc->sc_requested_r_scale; 907 } 908 if (sc->sc_flags & SCF_TIMESTAMP) { 909 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 910 tp->ts_recent = sc->sc_tsreflect; 911 tp->ts_recent_age = tcp_ts_getticks(); 912 tp->ts_offset = sc->sc_tsoff; 913 } 914 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 915 if (sc->sc_flags & SCF_SIGNATURE) 916 tp->t_flags |= TF_SIGNATURE; 917 #endif 918 if (sc->sc_flags & SCF_SACK) 919 tp->t_flags |= TF_SACK_PERMIT; 920 } 921 922 if (sc->sc_flags & SCF_ECN) 923 tp->t_flags |= TF_ECN_PERMIT; 924 925 /* 926 * Set up MSS and get cached values from tcp_hostcache. 927 * This might overwrite some of the defaults we just set. 928 */ 929 tcp_mss(tp, sc->sc_peer_mss); 930 931 /* 932 * If the SYN,ACK was retransmitted, indicate that CWND to be 933 * limited to one segment in cc_conn_init(). 934 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 935 */ 936 if (sc->sc_rxmits > 1) 937 tp->snd_cwnd = 1; 938 939 #ifdef TCP_OFFLOAD 940 /* 941 * Allow a TOE driver to install its hooks. Note that we hold the 942 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 943 * new connection before the TOE driver has done its thing. 944 */ 945 if (ADDED_BY_TOE(sc)) { 946 struct toedev *tod = sc->sc_tod; 947 948 tod->tod_offload_socket(tod, sc->sc_todctx, so); 949 } 950 #endif 951 /* 952 * Copy and activate timers. 953 */ 954 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 955 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 956 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 957 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 958 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 959 960 TCPSTAT_INC(tcps_accepts); 961 return (so); 962 963 abort: 964 INP_WUNLOCK(inp); 965 abort2: 966 if (so != NULL) 967 soabort(so); 968 return (NULL); 969 } 970 971 /* 972 * This function gets called when we receive an ACK for a 973 * socket in the LISTEN state. We look up the connection 974 * in the syncache, and if its there, we pull it out of 975 * the cache and turn it into a full-blown connection in 976 * the SYN-RECEIVED state. 977 * 978 * On syncache_socket() success the newly created socket 979 * has its underlying inp locked. 980 */ 981 int 982 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 983 struct socket **lsop, struct mbuf *m) 984 { 985 struct syncache *sc; 986 struct syncache_head *sch; 987 struct syncache scs; 988 char *s; 989 990 /* 991 * Global TCP locks are held because we manipulate the PCB lists 992 * and create a new socket. 993 */ 994 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 995 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 996 ("%s: can handle only ACK", __func__)); 997 998 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 999 SCH_LOCK_ASSERT(sch); 1000 1001 #ifdef INVARIANTS 1002 /* 1003 * Test code for syncookies comparing the syncache stored 1004 * values with the reconstructed values from the cookie. 1005 */ 1006 if (sc != NULL) 1007 syncookie_cmp(inc, sch, sc, th, to, *lsop); 1008 #endif 1009 1010 if (sc == NULL) { 1011 /* 1012 * There is no syncache entry, so see if this ACK is 1013 * a returning syncookie. To do this, first: 1014 * A. Check if syncookies are used in case of syncache 1015 * overflows 1016 * B. See if this socket has had a syncache entry dropped in 1017 * the recent past. We don't want to accept a bogus 1018 * syncookie if we've never received a SYN or accept it 1019 * twice. 1020 * C. check that the syncookie is valid. If it is, then 1021 * cobble up a fake syncache entry, and return. 1022 */ 1023 if (!V_tcp_syncookies) { 1024 SCH_UNLOCK(sch); 1025 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1026 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1027 "segment rejected (syncookies disabled)\n", 1028 s, __func__); 1029 goto failed; 1030 } 1031 if (!V_tcp_syncookiesonly && 1032 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1033 SCH_UNLOCK(sch); 1034 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1035 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1036 "segment rejected (no syncache entry)\n", 1037 s, __func__); 1038 goto failed; 1039 } 1040 bzero(&scs, sizeof(scs)); 1041 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1042 SCH_UNLOCK(sch); 1043 if (sc == NULL) { 1044 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1045 log(LOG_DEBUG, "%s; %s: Segment failed " 1046 "SYNCOOKIE authentication, segment rejected " 1047 "(probably spoofed)\n", s, __func__); 1048 goto failed; 1049 } 1050 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1051 /* If received ACK has MD5 signature, check it. */ 1052 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1053 (!TCPMD5_ENABLED() || 1054 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1055 /* Drop the ACK. */ 1056 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1057 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1058 "MD5 signature doesn't match.\n", 1059 s, __func__); 1060 free(s, M_TCPLOG); 1061 } 1062 TCPSTAT_INC(tcps_sig_err_sigopt); 1063 return (-1); /* Do not send RST */ 1064 } 1065 #endif /* TCP_SIGNATURE */ 1066 } else { 1067 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1068 /* 1069 * If listening socket requested TCP digests, check that 1070 * received ACK has signature and it is correct. 1071 * If not, drop the ACK and leave sc entry in th cache, 1072 * because SYN was received with correct signature. 1073 */ 1074 if (sc->sc_flags & SCF_SIGNATURE) { 1075 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1076 /* No signature */ 1077 TCPSTAT_INC(tcps_sig_err_nosigopt); 1078 SCH_UNLOCK(sch); 1079 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1080 log(LOG_DEBUG, "%s; %s: Segment " 1081 "rejected, MD5 signature wasn't " 1082 "provided.\n", s, __func__); 1083 free(s, M_TCPLOG); 1084 } 1085 return (-1); /* Do not send RST */ 1086 } 1087 if (!TCPMD5_ENABLED() || 1088 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1089 /* Doesn't match or no SA */ 1090 SCH_UNLOCK(sch); 1091 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1092 log(LOG_DEBUG, "%s; %s: Segment " 1093 "rejected, MD5 signature doesn't " 1094 "match.\n", s, __func__); 1095 free(s, M_TCPLOG); 1096 } 1097 return (-1); /* Do not send RST */ 1098 } 1099 } 1100 #endif /* TCP_SIGNATURE */ 1101 /* 1102 * Pull out the entry to unlock the bucket row. 1103 * 1104 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1105 * tcp_state_change(). The tcpcb is not existent at this 1106 * moment. A new one will be allocated via syncache_socket-> 1107 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1108 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1109 */ 1110 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1111 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1112 sch->sch_length--; 1113 #ifdef TCP_OFFLOAD 1114 if (ADDED_BY_TOE(sc)) { 1115 struct toedev *tod = sc->sc_tod; 1116 1117 tod->tod_syncache_removed(tod, sc->sc_todctx); 1118 } 1119 #endif 1120 SCH_UNLOCK(sch); 1121 } 1122 1123 /* 1124 * Segment validation: 1125 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1126 */ 1127 if (th->th_ack != sc->sc_iss + 1) { 1128 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1129 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1130 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1131 goto failed; 1132 } 1133 1134 /* 1135 * The SEQ must fall in the window starting at the received 1136 * initial receive sequence number + 1 (the SYN). 1137 */ 1138 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1139 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1140 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1141 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1142 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1143 goto failed; 1144 } 1145 1146 /* 1147 * If timestamps were not negotiated during SYN/ACK they 1148 * must not appear on any segment during this session. 1149 */ 1150 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1151 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1152 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1153 "segment rejected\n", s, __func__); 1154 goto failed; 1155 } 1156 1157 /* 1158 * If timestamps were negotiated during SYN/ACK they should 1159 * appear on every segment during this session. 1160 * XXXAO: This is only informal as there have been unverified 1161 * reports of non-compliants stacks. 1162 */ 1163 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1164 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1165 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1166 "no action\n", s, __func__); 1167 free(s, M_TCPLOG); 1168 s = NULL; 1169 } 1170 } 1171 1172 *lsop = syncache_socket(sc, *lsop, m); 1173 1174 if (*lsop == NULL) 1175 TCPSTAT_INC(tcps_sc_aborted); 1176 else 1177 TCPSTAT_INC(tcps_sc_completed); 1178 1179 /* how do we find the inp for the new socket? */ 1180 if (sc != &scs) 1181 syncache_free(sc); 1182 return (1); 1183 failed: 1184 if (sc != NULL && sc != &scs) 1185 syncache_free(sc); 1186 if (s != NULL) 1187 free(s, M_TCPLOG); 1188 *lsop = NULL; 1189 return (0); 1190 } 1191 1192 static void 1193 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1194 uint64_t response_cookie) 1195 { 1196 struct inpcb *inp; 1197 struct tcpcb *tp; 1198 unsigned int *pending_counter; 1199 1200 /* 1201 * Global TCP locks are held because we manipulate the PCB lists 1202 * and create a new socket. 1203 */ 1204 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1205 1206 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1207 *lsop = syncache_socket(sc, *lsop, m); 1208 if (*lsop == NULL) { 1209 TCPSTAT_INC(tcps_sc_aborted); 1210 atomic_subtract_int(pending_counter, 1); 1211 } else { 1212 soisconnected(*lsop); 1213 inp = sotoinpcb(*lsop); 1214 tp = intotcpcb(inp); 1215 tp->t_flags |= TF_FASTOPEN; 1216 tp->t_tfo_cookie.server = response_cookie; 1217 tp->snd_max = tp->iss; 1218 tp->snd_nxt = tp->iss; 1219 tp->t_tfo_pending = pending_counter; 1220 TCPSTAT_INC(tcps_sc_completed); 1221 } 1222 } 1223 1224 /* 1225 * Given a LISTEN socket and an inbound SYN request, add 1226 * this to the syn cache, and send back a segment: 1227 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1228 * to the source. 1229 * 1230 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1231 * Doing so would require that we hold onto the data and deliver it 1232 * to the application. However, if we are the target of a SYN-flood 1233 * DoS attack, an attacker could send data which would eventually 1234 * consume all available buffer space if it were ACKed. By not ACKing 1235 * the data, we avoid this DoS scenario. 1236 * 1237 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1238 * cookie is processed and a new socket is created. In this case, any data 1239 * accompanying the SYN will be queued to the socket by tcp_input() and will 1240 * be ACKed either when the application sends response data or the delayed 1241 * ACK timer expires, whichever comes first. 1242 */ 1243 int 1244 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1245 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1246 void *todctx) 1247 { 1248 struct tcpcb *tp; 1249 struct socket *so; 1250 struct syncache *sc = NULL; 1251 struct syncache_head *sch; 1252 struct mbuf *ipopts = NULL; 1253 u_int ltflags; 1254 int win, ip_ttl, ip_tos; 1255 char *s; 1256 int rv = 0; 1257 #ifdef INET6 1258 int autoflowlabel = 0; 1259 #endif 1260 #ifdef MAC 1261 struct label *maclabel; 1262 #endif 1263 struct syncache scs; 1264 struct ucred *cred; 1265 uint64_t tfo_response_cookie; 1266 unsigned int *tfo_pending = NULL; 1267 int tfo_cookie_valid = 0; 1268 int tfo_response_cookie_valid = 0; 1269 1270 INP_WLOCK_ASSERT(inp); /* listen socket */ 1271 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1272 ("%s: unexpected tcp flags", __func__)); 1273 1274 /* 1275 * Combine all so/tp operations very early to drop the INP lock as 1276 * soon as possible. 1277 */ 1278 so = *lsop; 1279 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1280 tp = sototcpcb(so); 1281 cred = crhold(so->so_cred); 1282 1283 #ifdef INET6 1284 if ((inc->inc_flags & INC_ISIPV6) && 1285 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1286 autoflowlabel = 1; 1287 #endif 1288 ip_ttl = inp->inp_ip_ttl; 1289 ip_tos = inp->inp_ip_tos; 1290 win = so->sol_sbrcv_hiwat; 1291 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1292 1293 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1294 (tp->t_tfo_pending != NULL) && 1295 (to->to_flags & TOF_FASTOPEN)) { 1296 /* 1297 * Limit the number of pending TFO connections to 1298 * approximately half of the queue limit. This prevents TFO 1299 * SYN floods from starving the service by filling the 1300 * listen queue with bogus TFO connections. 1301 */ 1302 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1303 (so->sol_qlimit / 2)) { 1304 int result; 1305 1306 result = tcp_fastopen_check_cookie(inc, 1307 to->to_tfo_cookie, to->to_tfo_len, 1308 &tfo_response_cookie); 1309 tfo_cookie_valid = (result > 0); 1310 tfo_response_cookie_valid = (result >= 0); 1311 } 1312 1313 /* 1314 * Remember the TFO pending counter as it will have to be 1315 * decremented below if we don't make it to syncache_tfo_expand(). 1316 */ 1317 tfo_pending = tp->t_tfo_pending; 1318 } 1319 1320 /* By the time we drop the lock these should no longer be used. */ 1321 so = NULL; 1322 tp = NULL; 1323 1324 #ifdef MAC 1325 if (mac_syncache_init(&maclabel) != 0) { 1326 INP_WUNLOCK(inp); 1327 goto done; 1328 } else 1329 mac_syncache_create(maclabel, inp); 1330 #endif 1331 if (!tfo_cookie_valid) 1332 INP_WUNLOCK(inp); 1333 1334 /* 1335 * Remember the IP options, if any. 1336 */ 1337 #ifdef INET6 1338 if (!(inc->inc_flags & INC_ISIPV6)) 1339 #endif 1340 #ifdef INET 1341 ipopts = (m) ? ip_srcroute(m) : NULL; 1342 #else 1343 ipopts = NULL; 1344 #endif 1345 1346 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1347 /* 1348 * If listening socket requested TCP digests, check that received 1349 * SYN has signature and it is correct. If signature doesn't match 1350 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1351 */ 1352 if (ltflags & TF_SIGNATURE) { 1353 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1354 TCPSTAT_INC(tcps_sig_err_nosigopt); 1355 goto done; 1356 } 1357 if (!TCPMD5_ENABLED() || 1358 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1359 goto done; 1360 } 1361 #endif /* TCP_SIGNATURE */ 1362 /* 1363 * See if we already have an entry for this connection. 1364 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1365 * 1366 * XXX: should the syncache be re-initialized with the contents 1367 * of the new SYN here (which may have different options?) 1368 * 1369 * XXX: We do not check the sequence number to see if this is a 1370 * real retransmit or a new connection attempt. The question is 1371 * how to handle such a case; either ignore it as spoofed, or 1372 * drop the current entry and create a new one? 1373 */ 1374 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1375 SCH_LOCK_ASSERT(sch); 1376 if (sc != NULL) { 1377 if (tfo_cookie_valid) 1378 INP_WUNLOCK(inp); 1379 TCPSTAT_INC(tcps_sc_dupsyn); 1380 if (ipopts) { 1381 /* 1382 * If we were remembering a previous source route, 1383 * forget it and use the new one we've been given. 1384 */ 1385 if (sc->sc_ipopts) 1386 (void) m_free(sc->sc_ipopts); 1387 sc->sc_ipopts = ipopts; 1388 } 1389 /* 1390 * Update timestamp if present. 1391 */ 1392 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1393 sc->sc_tsreflect = to->to_tsval; 1394 else 1395 sc->sc_flags &= ~SCF_TIMESTAMP; 1396 #ifdef MAC 1397 /* 1398 * Since we have already unconditionally allocated label 1399 * storage, free it up. The syncache entry will already 1400 * have an initialized label we can use. 1401 */ 1402 mac_syncache_destroy(&maclabel); 1403 #endif 1404 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1405 /* Retransmit SYN|ACK and reset retransmit count. */ 1406 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1407 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1408 "resetting timer and retransmitting SYN|ACK\n", 1409 s, __func__); 1410 free(s, M_TCPLOG); 1411 } 1412 if (syncache_respond(sc, sch, 1, m) == 0) { 1413 sc->sc_rxmits = 0; 1414 syncache_timeout(sc, sch, 1); 1415 TCPSTAT_INC(tcps_sndacks); 1416 TCPSTAT_INC(tcps_sndtotal); 1417 } 1418 SCH_UNLOCK(sch); 1419 goto donenoprobe; 1420 } 1421 1422 if (tfo_cookie_valid) { 1423 bzero(&scs, sizeof(scs)); 1424 sc = &scs; 1425 goto skip_alloc; 1426 } 1427 1428 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1429 if (sc == NULL) { 1430 /* 1431 * The zone allocator couldn't provide more entries. 1432 * Treat this as if the cache was full; drop the oldest 1433 * entry and insert the new one. 1434 */ 1435 TCPSTAT_INC(tcps_sc_zonefail); 1436 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1437 sch->sch_last_overflow = time_uptime; 1438 syncache_drop(sc, sch); 1439 } 1440 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1441 if (sc == NULL) { 1442 if (V_tcp_syncookies) { 1443 bzero(&scs, sizeof(scs)); 1444 sc = &scs; 1445 } else { 1446 SCH_UNLOCK(sch); 1447 if (ipopts) 1448 (void) m_free(ipopts); 1449 goto done; 1450 } 1451 } 1452 } 1453 1454 skip_alloc: 1455 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1456 sc->sc_tfo_cookie = &tfo_response_cookie; 1457 1458 /* 1459 * Fill in the syncache values. 1460 */ 1461 #ifdef MAC 1462 sc->sc_label = maclabel; 1463 #endif 1464 sc->sc_cred = cred; 1465 cred = NULL; 1466 sc->sc_ipopts = ipopts; 1467 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1468 #ifdef INET6 1469 if (!(inc->inc_flags & INC_ISIPV6)) 1470 #endif 1471 { 1472 sc->sc_ip_tos = ip_tos; 1473 sc->sc_ip_ttl = ip_ttl; 1474 } 1475 #ifdef TCP_OFFLOAD 1476 sc->sc_tod = tod; 1477 sc->sc_todctx = todctx; 1478 #endif 1479 sc->sc_irs = th->th_seq; 1480 sc->sc_iss = arc4random(); 1481 sc->sc_flags = 0; 1482 sc->sc_flowlabel = 0; 1483 1484 /* 1485 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1486 * win was derived from socket earlier in the function. 1487 */ 1488 win = imax(win, 0); 1489 win = imin(win, TCP_MAXWIN); 1490 sc->sc_wnd = win; 1491 1492 if (V_tcp_do_rfc1323) { 1493 /* 1494 * A timestamp received in a SYN makes 1495 * it ok to send timestamp requests and replies. 1496 */ 1497 if (to->to_flags & TOF_TS) { 1498 sc->sc_tsreflect = to->to_tsval; 1499 sc->sc_flags |= SCF_TIMESTAMP; 1500 } 1501 if (to->to_flags & TOF_SCALE) { 1502 int wscale = 0; 1503 1504 /* 1505 * Pick the smallest possible scaling factor that 1506 * will still allow us to scale up to sb_max, aka 1507 * kern.ipc.maxsockbuf. 1508 * 1509 * We do this because there are broken firewalls that 1510 * will corrupt the window scale option, leading to 1511 * the other endpoint believing that our advertised 1512 * window is unscaled. At scale factors larger than 1513 * 5 the unscaled window will drop below 1500 bytes, 1514 * leading to serious problems when traversing these 1515 * broken firewalls. 1516 * 1517 * With the default maxsockbuf of 256K, a scale factor 1518 * of 3 will be chosen by this algorithm. Those who 1519 * choose a larger maxsockbuf should watch out 1520 * for the compatibility problems mentioned above. 1521 * 1522 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1523 * or <SYN,ACK>) segment itself is never scaled. 1524 */ 1525 while (wscale < TCP_MAX_WINSHIFT && 1526 (TCP_MAXWIN << wscale) < sb_max) 1527 wscale++; 1528 sc->sc_requested_r_scale = wscale; 1529 sc->sc_requested_s_scale = to->to_wscale; 1530 sc->sc_flags |= SCF_WINSCALE; 1531 } 1532 } 1533 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1534 /* 1535 * If listening socket requested TCP digests, flag this in the 1536 * syncache so that syncache_respond() will do the right thing 1537 * with the SYN+ACK. 1538 */ 1539 if (ltflags & TF_SIGNATURE) 1540 sc->sc_flags |= SCF_SIGNATURE; 1541 #endif /* TCP_SIGNATURE */ 1542 if (to->to_flags & TOF_SACKPERM) 1543 sc->sc_flags |= SCF_SACK; 1544 if (to->to_flags & TOF_MSS) 1545 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1546 if (ltflags & TF_NOOPT) 1547 sc->sc_flags |= SCF_NOOPT; 1548 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1549 sc->sc_flags |= SCF_ECN; 1550 1551 if (V_tcp_syncookies) 1552 sc->sc_iss = syncookie_generate(sch, sc); 1553 #ifdef INET6 1554 if (autoflowlabel) { 1555 if (V_tcp_syncookies) 1556 sc->sc_flowlabel = sc->sc_iss; 1557 else 1558 sc->sc_flowlabel = ip6_randomflowlabel(); 1559 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1560 } 1561 #endif 1562 SCH_UNLOCK(sch); 1563 1564 if (tfo_cookie_valid) { 1565 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1566 /* INP_WUNLOCK(inp) will be performed by the caller */ 1567 rv = 1; 1568 goto tfo_expanded; 1569 } 1570 1571 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1572 /* 1573 * Do a standard 3-way handshake. 1574 */ 1575 if (syncache_respond(sc, sch, 0, m) == 0) { 1576 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1577 syncache_free(sc); 1578 else if (sc != &scs) 1579 syncache_insert(sc, sch); /* locks and unlocks sch */ 1580 TCPSTAT_INC(tcps_sndacks); 1581 TCPSTAT_INC(tcps_sndtotal); 1582 } else { 1583 if (sc != &scs) 1584 syncache_free(sc); 1585 TCPSTAT_INC(tcps_sc_dropped); 1586 } 1587 goto donenoprobe; 1588 1589 done: 1590 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1591 donenoprobe: 1592 if (m) { 1593 *lsop = NULL; 1594 m_freem(m); 1595 } 1596 /* 1597 * If tfo_pending is not NULL here, then a TFO SYN that did not 1598 * result in a new socket was processed and the associated pending 1599 * counter has not yet been decremented. All such TFO processing paths 1600 * transit this point. 1601 */ 1602 if (tfo_pending != NULL) 1603 tcp_fastopen_decrement_counter(tfo_pending); 1604 1605 tfo_expanded: 1606 if (cred != NULL) 1607 crfree(cred); 1608 #ifdef MAC 1609 if (sc == &scs) 1610 mac_syncache_destroy(&maclabel); 1611 #endif 1612 return (rv); 1613 } 1614 1615 /* 1616 * Send SYN|ACK to the peer. Either in response to the peer's SYN, 1617 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1618 */ 1619 static int 1620 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked, 1621 const struct mbuf *m0) 1622 { 1623 struct ip *ip = NULL; 1624 struct mbuf *m; 1625 struct tcphdr *th = NULL; 1626 int optlen, error = 0; /* Make compiler happy */ 1627 u_int16_t hlen, tlen, mssopt; 1628 struct tcpopt to; 1629 #ifdef INET6 1630 struct ip6_hdr *ip6 = NULL; 1631 #endif 1632 hlen = 1633 #ifdef INET6 1634 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1635 #endif 1636 sizeof(struct ip); 1637 tlen = hlen + sizeof(struct tcphdr); 1638 1639 /* Determine MSS we advertize to other end of connection. */ 1640 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); 1641 1642 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1643 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1644 ("syncache: mbuf too small")); 1645 1646 /* Create the IP+TCP header from scratch. */ 1647 m = m_gethdr(M_NOWAIT, MT_DATA); 1648 if (m == NULL) 1649 return (ENOBUFS); 1650 #ifdef MAC 1651 mac_syncache_create_mbuf(sc->sc_label, m); 1652 #endif 1653 m->m_data += max_linkhdr; 1654 m->m_len = tlen; 1655 m->m_pkthdr.len = tlen; 1656 m->m_pkthdr.rcvif = NULL; 1657 1658 #ifdef INET6 1659 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1660 ip6 = mtod(m, struct ip6_hdr *); 1661 ip6->ip6_vfc = IPV6_VERSION; 1662 ip6->ip6_nxt = IPPROTO_TCP; 1663 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1664 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1665 ip6->ip6_plen = htons(tlen - hlen); 1666 /* ip6_hlim is set after checksum */ 1667 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1668 ip6->ip6_flow |= sc->sc_flowlabel; 1669 1670 th = (struct tcphdr *)(ip6 + 1); 1671 } 1672 #endif 1673 #if defined(INET6) && defined(INET) 1674 else 1675 #endif 1676 #ifdef INET 1677 { 1678 ip = mtod(m, struct ip *); 1679 ip->ip_v = IPVERSION; 1680 ip->ip_hl = sizeof(struct ip) >> 2; 1681 ip->ip_len = htons(tlen); 1682 ip->ip_id = 0; 1683 ip->ip_off = 0; 1684 ip->ip_sum = 0; 1685 ip->ip_p = IPPROTO_TCP; 1686 ip->ip_src = sc->sc_inc.inc_laddr; 1687 ip->ip_dst = sc->sc_inc.inc_faddr; 1688 ip->ip_ttl = sc->sc_ip_ttl; 1689 ip->ip_tos = sc->sc_ip_tos; 1690 1691 /* 1692 * See if we should do MTU discovery. Route lookups are 1693 * expensive, so we will only unset the DF bit if: 1694 * 1695 * 1) path_mtu_discovery is disabled 1696 * 2) the SCF_UNREACH flag has been set 1697 */ 1698 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1699 ip->ip_off |= htons(IP_DF); 1700 1701 th = (struct tcphdr *)(ip + 1); 1702 } 1703 #endif /* INET */ 1704 th->th_sport = sc->sc_inc.inc_lport; 1705 th->th_dport = sc->sc_inc.inc_fport; 1706 1707 th->th_seq = htonl(sc->sc_iss); 1708 th->th_ack = htonl(sc->sc_irs + 1); 1709 th->th_off = sizeof(struct tcphdr) >> 2; 1710 th->th_x2 = 0; 1711 th->th_flags = TH_SYN|TH_ACK; 1712 th->th_win = htons(sc->sc_wnd); 1713 th->th_urp = 0; 1714 1715 if (sc->sc_flags & SCF_ECN) { 1716 th->th_flags |= TH_ECE; 1717 TCPSTAT_INC(tcps_ecn_shs); 1718 } 1719 1720 /* Tack on the TCP options. */ 1721 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1722 to.to_flags = 0; 1723 1724 to.to_mss = mssopt; 1725 to.to_flags = TOF_MSS; 1726 if (sc->sc_flags & SCF_WINSCALE) { 1727 to.to_wscale = sc->sc_requested_r_scale; 1728 to.to_flags |= TOF_SCALE; 1729 } 1730 if (sc->sc_flags & SCF_TIMESTAMP) { 1731 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 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 if (sc->sc_tfo_cookie) { 1742 to.to_flags |= TOF_FASTOPEN; 1743 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1744 to.to_tfo_cookie = sc->sc_tfo_cookie; 1745 /* don't send cookie again when retransmitting response */ 1746 sc->sc_tfo_cookie = NULL; 1747 } 1748 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1749 1750 /* Adjust headers by option size. */ 1751 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1752 m->m_len += optlen; 1753 m->m_pkthdr.len += optlen; 1754 #ifdef INET6 1755 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1756 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1757 else 1758 #endif 1759 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1760 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1761 if (sc->sc_flags & SCF_SIGNATURE) { 1762 KASSERT(to.to_flags & TOF_SIGNATURE, 1763 ("tcp_addoptions() didn't set tcp_signature")); 1764 1765 /* NOTE: to.to_signature is inside of mbuf */ 1766 if (!TCPMD5_ENABLED() || 1767 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1768 m_freem(m); 1769 return (EACCES); 1770 } 1771 } 1772 #endif 1773 } else 1774 optlen = 0; 1775 1776 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1777 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1778 /* 1779 * If we have peer's SYN and it has a flowid, then let's assign it to 1780 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1781 * to SYN|ACK due to lack of inp here. 1782 */ 1783 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1784 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1785 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1786 } 1787 #ifdef INET6 1788 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1789 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1790 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1791 IPPROTO_TCP, 0); 1792 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1793 #ifdef TCP_OFFLOAD 1794 if (ADDED_BY_TOE(sc)) { 1795 struct toedev *tod = sc->sc_tod; 1796 1797 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1798 1799 return (error); 1800 } 1801 #endif 1802 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 1803 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1804 } 1805 #endif 1806 #if defined(INET6) && defined(INET) 1807 else 1808 #endif 1809 #ifdef INET 1810 { 1811 m->m_pkthdr.csum_flags = CSUM_TCP; 1812 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1813 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1814 #ifdef TCP_OFFLOAD 1815 if (ADDED_BY_TOE(sc)) { 1816 struct toedev *tod = sc->sc_tod; 1817 1818 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1819 1820 return (error); 1821 } 1822 #endif 1823 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 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, 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 for (i = nitems(tcp_sc_msstab) - 1; 1999 tcp_sc_msstab[i] > sc->sc_peer_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_tsoff = arc4random() - tcp_ts_getticks(); 2041 } 2042 2043 TCPSTAT_INC(tcps_sc_sendcookie); 2044 return (iss); 2045 } 2046 2047 static struct syncache * 2048 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2049 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2050 struct socket *lso) 2051 { 2052 uint32_t hash; 2053 uint8_t *secbits; 2054 tcp_seq ack, seq; 2055 int wnd, wscale = 0; 2056 union syncookie cookie; 2057 2058 SCH_LOCK_ASSERT(sch); 2059 2060 /* 2061 * Pull information out of SYN-ACK/ACK and revert sequence number 2062 * advances. 2063 */ 2064 ack = th->th_ack - 1; 2065 seq = th->th_seq - 1; 2066 2067 /* 2068 * Unpack the flags containing enough information to restore the 2069 * connection. 2070 */ 2071 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2072 2073 /* Which of the two secrets to use. */ 2074 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 2075 2076 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2077 2078 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2079 if ((ack & ~0xff) != (hash & ~0xff)) 2080 return (NULL); 2081 2082 /* Fill in the syncache values. */ 2083 sc->sc_flags = 0; 2084 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2085 sc->sc_ipopts = NULL; 2086 2087 sc->sc_irs = seq; 2088 sc->sc_iss = ack; 2089 2090 switch (inc->inc_flags & INC_ISIPV6) { 2091 #ifdef INET 2092 case 0: 2093 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2094 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2095 break; 2096 #endif 2097 #ifdef INET6 2098 case INC_ISIPV6: 2099 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2100 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 2101 break; 2102 #endif 2103 } 2104 2105 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2106 2107 /* We can simply recompute receive window scale we sent earlier. */ 2108 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2109 wscale++; 2110 2111 /* Only use wscale if it was enabled in the orignal SYN. */ 2112 if (cookie.flags.wscale_idx > 0) { 2113 sc->sc_requested_r_scale = wscale; 2114 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2115 sc->sc_flags |= SCF_WINSCALE; 2116 } 2117 2118 wnd = lso->sol_sbrcv_hiwat; 2119 wnd = imax(wnd, 0); 2120 wnd = imin(wnd, TCP_MAXWIN); 2121 sc->sc_wnd = wnd; 2122 2123 if (cookie.flags.sack_ok) 2124 sc->sc_flags |= SCF_SACK; 2125 2126 if (to->to_flags & TOF_TS) { 2127 sc->sc_flags |= SCF_TIMESTAMP; 2128 sc->sc_tsreflect = to->to_tsval; 2129 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 2130 } 2131 2132 if (to->to_flags & TOF_SIGNATURE) 2133 sc->sc_flags |= SCF_SIGNATURE; 2134 2135 sc->sc_rxmits = 0; 2136 2137 TCPSTAT_INC(tcps_sc_recvcookie); 2138 return (sc); 2139 } 2140 2141 #ifdef INVARIANTS 2142 static int 2143 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2144 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2145 struct socket *lso) 2146 { 2147 struct syncache scs, *scx; 2148 char *s; 2149 2150 bzero(&scs, sizeof(scs)); 2151 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2152 2153 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2154 return (0); 2155 2156 if (scx != NULL) { 2157 if (sc->sc_peer_mss != scx->sc_peer_mss) 2158 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2159 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2160 2161 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2162 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2163 s, __func__, sc->sc_requested_r_scale, 2164 scx->sc_requested_r_scale); 2165 2166 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2167 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2168 s, __func__, sc->sc_requested_s_scale, 2169 scx->sc_requested_s_scale); 2170 2171 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2172 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2173 } 2174 2175 if (s != NULL) 2176 free(s, M_TCPLOG); 2177 return (0); 2178 } 2179 #endif /* INVARIANTS */ 2180 2181 static void 2182 syncookie_reseed(void *arg) 2183 { 2184 struct tcp_syncache *sc = arg; 2185 uint8_t *secbits; 2186 int secbit; 2187 2188 /* 2189 * Reseeding the secret doesn't have to be protected by a lock. 2190 * It only must be ensured that the new random values are visible 2191 * to all CPUs in a SMP environment. The atomic with release 2192 * semantics ensures that. 2193 */ 2194 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2195 secbits = sc->secret.key[secbit]; 2196 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2197 atomic_add_rel_int(&sc->secret.oddeven, 1); 2198 2199 /* Reschedule ourself. */ 2200 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2201 } 2202 2203 /* 2204 * Exports the syncache entries to userland so that netstat can display 2205 * them alongside the other sockets. This function is intended to be 2206 * called only from tcp_pcblist. 2207 * 2208 * Due to concurrency on an active system, the number of pcbs exported 2209 * may have no relation to max_pcbs. max_pcbs merely indicates the 2210 * amount of space the caller allocated for this function to use. 2211 */ 2212 int 2213 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2214 { 2215 struct xtcpcb xt; 2216 struct syncache *sc; 2217 struct syncache_head *sch; 2218 int count, error, i; 2219 2220 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2221 sch = &V_tcp_syncache.hashbase[i]; 2222 SCH_LOCK(sch); 2223 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2224 if (count >= max_pcbs) { 2225 SCH_UNLOCK(sch); 2226 goto exit; 2227 } 2228 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2229 continue; 2230 bzero(&xt, sizeof(xt)); 2231 xt.xt_len = sizeof(xt); 2232 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2233 xt.xt_inp.inp_vflag = INP_IPV6; 2234 else 2235 xt.xt_inp.inp_vflag = INP_IPV4; 2236 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2237 sizeof (struct in_conninfo)); 2238 xt.t_state = TCPS_SYN_RECEIVED; 2239 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2240 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2241 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2242 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2243 error = SYSCTL_OUT(req, &xt, sizeof xt); 2244 if (error) { 2245 SCH_UNLOCK(sch); 2246 goto exit; 2247 } 2248 count++; 2249 } 2250 SCH_UNLOCK(sch); 2251 } 2252 exit: 2253 *pcbs_exported = count; 2254 return error; 2255 } 2256