1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2001 McAfee, Inc. 5 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG 6 * All rights reserved. 7 * 8 * This software was developed for the FreeBSD Project by Jonathan Lemon 9 * and McAfee Research, the Security Research Division of McAfee, Inc. under 10 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 11 * DARPA CHATS research program. [2001 McAfee, Inc.] 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include "opt_inet.h" 39 #include "opt_inet6.h" 40 #include "opt_ipsec.h" 41 #include "opt_pcbgroup.h" 42 43 #include <sys/param.h> 44 #include <sys/systm.h> 45 #include <sys/hash.h> 46 #include <sys/refcount.h> 47 #include <sys/kernel.h> 48 #include <sys/sysctl.h> 49 #include <sys/limits.h> 50 #include <sys/lock.h> 51 #include <sys/mutex.h> 52 #include <sys/malloc.h> 53 #include <sys/mbuf.h> 54 #include <sys/proc.h> /* for proc0 declaration */ 55 #include <sys/random.h> 56 #include <sys/socket.h> 57 #include <sys/socketvar.h> 58 #include <sys/syslog.h> 59 #include <sys/ucred.h> 60 61 #include <sys/md5.h> 62 #include <crypto/siphash/siphash.h> 63 64 #include <vm/uma.h> 65 66 #include <net/if.h> 67 #include <net/if_var.h> 68 #include <net/route.h> 69 #include <net/vnet.h> 70 71 #include <netinet/in.h> 72 #include <netinet/in_systm.h> 73 #include <netinet/ip.h> 74 #include <netinet/in_var.h> 75 #include <netinet/in_pcb.h> 76 #include <netinet/ip_var.h> 77 #include <netinet/ip_options.h> 78 #ifdef INET6 79 #include <netinet/ip6.h> 80 #include <netinet/icmp6.h> 81 #include <netinet6/nd6.h> 82 #include <netinet6/ip6_var.h> 83 #include <netinet6/in6_pcb.h> 84 #endif 85 #include <netinet/tcp.h> 86 #include <netinet/tcp_fastopen.h> 87 #include <netinet/tcp_fsm.h> 88 #include <netinet/tcp_seq.h> 89 #include <netinet/tcp_timer.h> 90 #include <netinet/tcp_var.h> 91 #include <netinet/tcp_syncache.h> 92 #ifdef INET6 93 #include <netinet6/tcp6_var.h> 94 #endif 95 #ifdef TCP_OFFLOAD 96 #include <netinet/toecore.h> 97 #endif 98 99 #include <netipsec/ipsec_support.h> 100 101 #include <machine/in_cksum.h> 102 103 #include <security/mac/mac_framework.h> 104 105 static VNET_DEFINE(int, tcp_syncookies) = 1; 106 #define V_tcp_syncookies VNET(tcp_syncookies) 107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 108 &VNET_NAME(tcp_syncookies), 0, 109 "Use TCP SYN cookies if the syncache overflows"); 110 111 static VNET_DEFINE(int, tcp_syncookiesonly) = 0; 112 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 114 &VNET_NAME(tcp_syncookiesonly), 0, 115 "Use only TCP SYN cookies"); 116 117 static VNET_DEFINE(int, functions_inherit_listen_socket_stack) = 1; 118 #define V_functions_inherit_listen_socket_stack \ 119 VNET(functions_inherit_listen_socket_stack) 120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack, 121 CTLFLAG_VNET | CTLFLAG_RW, 122 &VNET_NAME(functions_inherit_listen_socket_stack), 0, 123 "Inherit listen socket's stack"); 124 125 #ifdef TCP_OFFLOAD 126 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 127 #endif 128 129 static void syncache_drop(struct syncache *, struct syncache_head *); 130 static void syncache_free(struct syncache *); 131 static void syncache_insert(struct syncache *, struct syncache_head *); 132 static int syncache_respond(struct syncache *, struct syncache_head *, int, 133 const struct mbuf *); 134 static struct socket *syncache_socket(struct syncache *, struct socket *, 135 struct mbuf *m); 136 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 137 int docallout); 138 static void syncache_timer(void *); 139 140 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 141 uint8_t *, uintptr_t); 142 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 143 static struct syncache 144 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 145 struct syncache *, struct tcphdr *, struct tcpopt *, 146 struct socket *); 147 static void syncookie_reseed(void *); 148 #ifdef INVARIANTS 149 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 150 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 151 struct socket *lso); 152 #endif 153 154 /* 155 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 156 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds, 157 * the odds are that the user has given up attempting to connect by then. 158 */ 159 #define SYNCACHE_MAXREXMTS 3 160 161 /* Arbitrary values */ 162 #define TCP_SYNCACHE_HASHSIZE 512 163 #define TCP_SYNCACHE_BUCKETLIMIT 30 164 165 static VNET_DEFINE(struct tcp_syncache, tcp_syncache); 166 #define V_tcp_syncache VNET(tcp_syncache) 167 168 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, 169 "TCP SYN cache"); 170 171 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 172 &VNET_NAME(tcp_syncache.bucket_limit), 0, 173 "Per-bucket hash limit for syncache"); 174 175 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 176 &VNET_NAME(tcp_syncache.cache_limit), 0, 177 "Overall entry limit for syncache"); 178 179 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 180 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 181 182 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 183 &VNET_NAME(tcp_syncache.hashsize), 0, 184 "Size of TCP syncache hashtable"); 185 186 static int 187 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 188 { 189 int error; 190 u_int new; 191 192 new = V_tcp_syncache.rexmt_limit; 193 error = sysctl_handle_int(oidp, &new, 0, req); 194 if ((error == 0) && (req->newptr != NULL)) { 195 if (new > TCP_MAXRXTSHIFT) 196 error = EINVAL; 197 else 198 V_tcp_syncache.rexmt_limit = new; 199 } 200 return (error); 201 } 202 203 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 204 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW, 205 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 206 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI", 207 "Limit on SYN/ACK retransmissions"); 208 209 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 210 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 211 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 212 "Send reset on socket allocation failure"); 213 214 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 215 216 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 217 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 218 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 219 220 /* 221 * Requires the syncache entry to be already removed from the bucket list. 222 */ 223 static void 224 syncache_free(struct syncache *sc) 225 { 226 227 if (sc->sc_ipopts) 228 (void) m_free(sc->sc_ipopts); 229 if (sc->sc_cred) 230 crfree(sc->sc_cred); 231 #ifdef MAC 232 mac_syncache_destroy(&sc->sc_label); 233 #endif 234 235 uma_zfree(V_tcp_syncache.zone, sc); 236 } 237 238 void 239 syncache_init(void) 240 { 241 int i; 242 243 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 244 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 245 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 246 V_tcp_syncache.hash_secret = arc4random(); 247 248 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 249 &V_tcp_syncache.hashsize); 250 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 251 &V_tcp_syncache.bucket_limit); 252 if (!powerof2(V_tcp_syncache.hashsize) || 253 V_tcp_syncache.hashsize == 0) { 254 printf("WARNING: syncache hash size is not a power of 2.\n"); 255 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 256 } 257 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 258 259 /* Set limits. */ 260 V_tcp_syncache.cache_limit = 261 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 262 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 263 &V_tcp_syncache.cache_limit); 264 265 /* Allocate the hash table. */ 266 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 267 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 268 269 #ifdef VIMAGE 270 V_tcp_syncache.vnet = curvnet; 271 #endif 272 273 /* Initialize the hash buckets. */ 274 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 275 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 276 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 277 NULL, MTX_DEF); 278 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 279 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 280 V_tcp_syncache.hashbase[i].sch_length = 0; 281 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 282 V_tcp_syncache.hashbase[i].sch_last_overflow = 283 -(SYNCOOKIE_LIFETIME + 1); 284 } 285 286 /* Create the syncache entry zone. */ 287 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 288 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 289 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 290 V_tcp_syncache.cache_limit); 291 292 /* Start the SYN cookie reseeder callout. */ 293 callout_init(&V_tcp_syncache.secret.reseed, 1); 294 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 295 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 296 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 297 syncookie_reseed, &V_tcp_syncache); 298 } 299 300 #ifdef VIMAGE 301 void 302 syncache_destroy(void) 303 { 304 struct syncache_head *sch; 305 struct syncache *sc, *nsc; 306 int i; 307 308 /* 309 * Stop the re-seed timer before freeing resources. No need to 310 * possibly schedule it another time. 311 */ 312 callout_drain(&V_tcp_syncache.secret.reseed); 313 314 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 315 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 316 317 sch = &V_tcp_syncache.hashbase[i]; 318 callout_drain(&sch->sch_timer); 319 320 SCH_LOCK(sch); 321 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 322 syncache_drop(sc, sch); 323 SCH_UNLOCK(sch); 324 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 325 ("%s: sch->sch_bucket not empty", __func__)); 326 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 327 __func__, sch->sch_length)); 328 mtx_destroy(&sch->sch_mtx); 329 } 330 331 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 332 ("%s: cache_count not 0", __func__)); 333 334 /* Free the allocated global resources. */ 335 uma_zdestroy(V_tcp_syncache.zone); 336 free(V_tcp_syncache.hashbase, M_SYNCACHE); 337 } 338 #endif 339 340 /* 341 * Inserts a syncache entry into the specified bucket row. 342 * Locks and unlocks the syncache_head autonomously. 343 */ 344 static void 345 syncache_insert(struct syncache *sc, struct syncache_head *sch) 346 { 347 struct syncache *sc2; 348 349 SCH_LOCK(sch); 350 351 /* 352 * Make sure that we don't overflow the per-bucket limit. 353 * If the bucket is full, toss the oldest element. 354 */ 355 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 356 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 357 ("sch->sch_length incorrect")); 358 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 359 sch->sch_last_overflow = time_uptime; 360 syncache_drop(sc2, sch); 361 TCPSTAT_INC(tcps_sc_bucketoverflow); 362 } 363 364 /* Put it into the bucket. */ 365 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 366 sch->sch_length++; 367 368 #ifdef TCP_OFFLOAD 369 if (ADDED_BY_TOE(sc)) { 370 struct toedev *tod = sc->sc_tod; 371 372 tod->tod_syncache_added(tod, sc->sc_todctx); 373 } 374 #endif 375 376 /* Reinitialize the bucket row's timer. */ 377 if (sch->sch_length == 1) 378 sch->sch_nextc = ticks + INT_MAX; 379 syncache_timeout(sc, sch, 1); 380 381 SCH_UNLOCK(sch); 382 383 TCPSTATES_INC(TCPS_SYN_RECEIVED); 384 TCPSTAT_INC(tcps_sc_added); 385 } 386 387 /* 388 * Remove and free entry from syncache bucket row. 389 * Expects locked syncache head. 390 */ 391 static void 392 syncache_drop(struct syncache *sc, struct syncache_head *sch) 393 { 394 395 SCH_LOCK_ASSERT(sch); 396 397 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 398 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 399 sch->sch_length--; 400 401 #ifdef TCP_OFFLOAD 402 if (ADDED_BY_TOE(sc)) { 403 struct toedev *tod = sc->sc_tod; 404 405 tod->tod_syncache_removed(tod, sc->sc_todctx); 406 } 407 #endif 408 409 syncache_free(sc); 410 } 411 412 /* 413 * Engage/reengage time on bucket row. 414 */ 415 static void 416 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 417 { 418 int rexmt; 419 420 if (sc->sc_rxmits == 0) 421 rexmt = TCPTV_RTOBASE; 422 else 423 TCPT_RANGESET(rexmt, TCPTV_RTOBASE * tcp_syn_backoff[sc->sc_rxmits], 424 tcp_rexmit_min, TCPTV_REXMTMAX); 425 sc->sc_rxttime = ticks + rexmt; 426 sc->sc_rxmits++; 427 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 428 sch->sch_nextc = sc->sc_rxttime; 429 if (docallout) 430 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 431 syncache_timer, (void *)sch); 432 } 433 } 434 435 /* 436 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 437 * If we have retransmitted an entry the maximum number of times, expire it. 438 * One separate timer for each bucket row. 439 */ 440 static void 441 syncache_timer(void *xsch) 442 { 443 struct syncache_head *sch = (struct syncache_head *)xsch; 444 struct syncache *sc, *nsc; 445 int tick = ticks; 446 char *s; 447 448 CURVNET_SET(sch->sch_sc->vnet); 449 450 /* NB: syncache_head has already been locked by the callout. */ 451 SCH_LOCK_ASSERT(sch); 452 453 /* 454 * In the following cycle we may remove some entries and/or 455 * advance some timeouts, so re-initialize the bucket timer. 456 */ 457 sch->sch_nextc = tick + INT_MAX; 458 459 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 460 /* 461 * We do not check if the listen socket still exists 462 * and accept the case where the listen socket may be 463 * gone by the time we resend the SYN/ACK. We do 464 * not expect this to happens often. If it does, 465 * then the RST will be sent by the time the remote 466 * host does the SYN/ACK->ACK. 467 */ 468 if (TSTMP_GT(sc->sc_rxttime, tick)) { 469 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 470 sch->sch_nextc = sc->sc_rxttime; 471 continue; 472 } 473 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 474 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 475 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 476 "giving up and removing syncache entry\n", 477 s, __func__); 478 free(s, M_TCPLOG); 479 } 480 syncache_drop(sc, sch); 481 TCPSTAT_INC(tcps_sc_stale); 482 continue; 483 } 484 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 485 log(LOG_DEBUG, "%s; %s: Response timeout, " 486 "retransmitting (%u) SYN|ACK\n", 487 s, __func__, sc->sc_rxmits); 488 free(s, M_TCPLOG); 489 } 490 491 syncache_respond(sc, sch, 1, NULL); 492 TCPSTAT_INC(tcps_sc_retransmitted); 493 syncache_timeout(sc, sch, 0); 494 } 495 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 496 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 497 syncache_timer, (void *)(sch)); 498 CURVNET_RESTORE(); 499 } 500 501 /* 502 * Find an entry in the syncache. 503 * Returns always with locked syncache_head plus a matching entry or NULL. 504 */ 505 static struct syncache * 506 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 507 { 508 struct syncache *sc; 509 struct syncache_head *sch; 510 uint32_t hash; 511 512 /* 513 * The hash is built on foreign port + local port + foreign address. 514 * We rely on the fact that struct in_conninfo starts with 16 bits 515 * of foreign port, then 16 bits of local port then followed by 128 516 * bits of foreign address. In case of IPv4 address, the first 3 517 * 32-bit words of the address always are zeroes. 518 */ 519 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 520 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 521 522 sch = &V_tcp_syncache.hashbase[hash]; 523 *schp = sch; 524 SCH_LOCK(sch); 525 526 /* Circle through bucket row to find matching entry. */ 527 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 528 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 529 sizeof(struct in_endpoints)) == 0) 530 break; 531 532 return (sc); /* Always returns with locked sch. */ 533 } 534 535 /* 536 * This function is called when we get a RST for a 537 * non-existent connection, so that we can see if the 538 * connection is in the syn cache. If it is, zap it. 539 */ 540 void 541 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 542 { 543 struct syncache *sc; 544 struct syncache_head *sch; 545 char *s = NULL; 546 547 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 548 SCH_LOCK_ASSERT(sch); 549 550 /* 551 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 552 * See RFC 793 page 65, section SEGMENT ARRIVES. 553 */ 554 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 555 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 556 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 557 "FIN flag set, segment ignored\n", s, __func__); 558 TCPSTAT_INC(tcps_badrst); 559 goto done; 560 } 561 562 /* 563 * No corresponding connection was found in syncache. 564 * If syncookies are enabled and possibly exclusively 565 * used, or we are under memory pressure, a valid RST 566 * may not find a syncache entry. In that case we're 567 * done and no SYN|ACK retransmissions will happen. 568 * Otherwise the RST was misdirected or spoofed. 569 */ 570 if (sc == NULL) { 571 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 572 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 573 "syncache entry (possibly syncookie only), " 574 "segment ignored\n", s, __func__); 575 TCPSTAT_INC(tcps_badrst); 576 goto done; 577 } 578 579 /* 580 * If the RST bit is set, check the sequence number to see 581 * if this is a valid reset segment. 582 * RFC 793 page 37: 583 * In all states except SYN-SENT, all reset (RST) segments 584 * are validated by checking their SEQ-fields. A reset is 585 * valid if its sequence number is in the window. 586 * 587 * The sequence number in the reset segment is normally an 588 * echo of our outgoing acknowlegement numbers, but some hosts 589 * send a reset with the sequence number at the rightmost edge 590 * of our receive window, and we have to handle this case. 591 */ 592 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 593 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 594 syncache_drop(sc, sch); 595 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 596 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 597 "connection attempt aborted by remote endpoint\n", 598 s, __func__); 599 TCPSTAT_INC(tcps_sc_reset); 600 } else { 601 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 602 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 603 "IRS %u (+WND %u), segment ignored\n", 604 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 605 TCPSTAT_INC(tcps_badrst); 606 } 607 608 done: 609 if (s != NULL) 610 free(s, M_TCPLOG); 611 SCH_UNLOCK(sch); 612 } 613 614 void 615 syncache_badack(struct in_conninfo *inc) 616 { 617 struct syncache *sc; 618 struct syncache_head *sch; 619 620 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 621 SCH_LOCK_ASSERT(sch); 622 if (sc != NULL) { 623 syncache_drop(sc, sch); 624 TCPSTAT_INC(tcps_sc_badack); 625 } 626 SCH_UNLOCK(sch); 627 } 628 629 void 630 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq) 631 { 632 struct syncache *sc; 633 struct syncache_head *sch; 634 635 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 636 SCH_LOCK_ASSERT(sch); 637 if (sc == NULL) 638 goto done; 639 640 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 641 if (ntohl(th_seq) != sc->sc_iss) 642 goto done; 643 644 /* 645 * If we've rertransmitted 3 times and this is our second error, 646 * we remove the entry. Otherwise, we allow it to continue on. 647 * This prevents us from incorrectly nuking an entry during a 648 * spurious network outage. 649 * 650 * See tcp_notify(). 651 */ 652 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 653 sc->sc_flags |= SCF_UNREACH; 654 goto done; 655 } 656 syncache_drop(sc, sch); 657 TCPSTAT_INC(tcps_sc_unreach); 658 done: 659 SCH_UNLOCK(sch); 660 } 661 662 /* 663 * Build a new TCP socket structure from a syncache entry. 664 * 665 * On success return the newly created socket with its underlying inp locked. 666 */ 667 static struct socket * 668 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 669 { 670 struct tcp_function_block *blk; 671 struct inpcb *inp = NULL; 672 struct socket *so; 673 struct tcpcb *tp; 674 int error; 675 char *s; 676 677 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 678 679 /* 680 * Ok, create the full blown connection, and set things up 681 * as they would have been set up if we had created the 682 * connection when the SYN arrived. If we can't create 683 * the connection, abort it. 684 */ 685 so = sonewconn(lso, 0); 686 if (so == NULL) { 687 /* 688 * Drop the connection; we will either send a RST or 689 * have the peer retransmit its SYN again after its 690 * RTO and try again. 691 */ 692 TCPSTAT_INC(tcps_listendrop); 693 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 694 log(LOG_DEBUG, "%s; %s: Socket create failed " 695 "due to limits or memory shortage\n", 696 s, __func__); 697 free(s, M_TCPLOG); 698 } 699 goto abort2; 700 } 701 #ifdef MAC 702 mac_socketpeer_set_from_mbuf(m, so); 703 #endif 704 705 inp = sotoinpcb(so); 706 inp->inp_inc.inc_fibnum = so->so_fibnum; 707 INP_WLOCK(inp); 708 /* 709 * Exclusive pcbinfo lock is not required in syncache socket case even 710 * if two inpcb locks can be acquired simultaneously: 711 * - the inpcb in LISTEN state, 712 * - the newly created inp. 713 * 714 * In this case, an inp cannot be at same time in LISTEN state and 715 * just created by an accept() call. 716 */ 717 INP_HASH_WLOCK(&V_tcbinfo); 718 719 /* Insert new socket into PCB hash list. */ 720 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 721 #ifdef INET6 722 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 723 inp->inp_vflag &= ~INP_IPV4; 724 inp->inp_vflag |= INP_IPV6; 725 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 726 } else { 727 inp->inp_vflag &= ~INP_IPV6; 728 inp->inp_vflag |= INP_IPV4; 729 #endif 730 inp->inp_laddr = sc->sc_inc.inc_laddr; 731 #ifdef INET6 732 } 733 #endif 734 735 /* 736 * If there's an mbuf and it has a flowid, then let's initialise the 737 * inp with that particular flowid. 738 */ 739 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 740 inp->inp_flowid = m->m_pkthdr.flowid; 741 inp->inp_flowtype = M_HASHTYPE_GET(m); 742 } 743 744 /* 745 * Install in the reservation hash table for now, but don't yet 746 * install a connection group since the full 4-tuple isn't yet 747 * configured. 748 */ 749 inp->inp_lport = sc->sc_inc.inc_lport; 750 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 751 /* 752 * Undo the assignments above if we failed to 753 * put the PCB on the hash lists. 754 */ 755 #ifdef INET6 756 if (sc->sc_inc.inc_flags & INC_ISIPV6) 757 inp->in6p_laddr = in6addr_any; 758 else 759 #endif 760 inp->inp_laddr.s_addr = INADDR_ANY; 761 inp->inp_lport = 0; 762 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 763 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 764 "with error %i\n", 765 s, __func__, error); 766 free(s, M_TCPLOG); 767 } 768 INP_HASH_WUNLOCK(&V_tcbinfo); 769 goto abort; 770 } 771 #ifdef INET6 772 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 773 struct inpcb *oinp = sotoinpcb(lso); 774 struct in6_addr laddr6; 775 struct sockaddr_in6 sin6; 776 /* 777 * Inherit socket options from the listening socket. 778 * Note that in6p_inputopts are not (and should not be) 779 * copied, since it stores previously received options and is 780 * used to detect if each new option is different than the 781 * previous one and hence should be passed to a user. 782 * If we copied in6p_inputopts, a user would not be able to 783 * receive options just after calling the accept system call. 784 */ 785 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 786 if (oinp->in6p_outputopts) 787 inp->in6p_outputopts = 788 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 789 790 sin6.sin6_family = AF_INET6; 791 sin6.sin6_len = sizeof(sin6); 792 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 793 sin6.sin6_port = sc->sc_inc.inc_fport; 794 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 795 laddr6 = inp->in6p_laddr; 796 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 797 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 798 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 799 thread0.td_ucred, m)) != 0) { 800 inp->in6p_laddr = laddr6; 801 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 802 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 803 "with error %i\n", 804 s, __func__, error); 805 free(s, M_TCPLOG); 806 } 807 INP_HASH_WUNLOCK(&V_tcbinfo); 808 goto abort; 809 } 810 /* Override flowlabel from in6_pcbconnect. */ 811 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 812 inp->inp_flow |= sc->sc_flowlabel; 813 } 814 #endif /* INET6 */ 815 #if defined(INET) && defined(INET6) 816 else 817 #endif 818 #ifdef INET 819 { 820 struct in_addr laddr; 821 struct sockaddr_in sin; 822 823 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 824 825 if (inp->inp_options == NULL) { 826 inp->inp_options = sc->sc_ipopts; 827 sc->sc_ipopts = NULL; 828 } 829 830 sin.sin_family = AF_INET; 831 sin.sin_len = sizeof(sin); 832 sin.sin_addr = sc->sc_inc.inc_faddr; 833 sin.sin_port = sc->sc_inc.inc_fport; 834 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 835 laddr = inp->inp_laddr; 836 if (inp->inp_laddr.s_addr == INADDR_ANY) 837 inp->inp_laddr = sc->sc_inc.inc_laddr; 838 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 839 thread0.td_ucred, m)) != 0) { 840 inp->inp_laddr = laddr; 841 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 842 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 843 "with error %i\n", 844 s, __func__, error); 845 free(s, M_TCPLOG); 846 } 847 INP_HASH_WUNLOCK(&V_tcbinfo); 848 goto abort; 849 } 850 } 851 #endif /* INET */ 852 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 853 /* Copy old policy into new socket's. */ 854 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 855 printf("syncache_socket: could not copy policy\n"); 856 #endif 857 INP_HASH_WUNLOCK(&V_tcbinfo); 858 tp = intotcpcb(inp); 859 tcp_state_change(tp, TCPS_SYN_RECEIVED); 860 tp->iss = sc->sc_iss; 861 tp->irs = sc->sc_irs; 862 tcp_rcvseqinit(tp); 863 tcp_sendseqinit(tp); 864 blk = sototcpcb(lso)->t_fb; 865 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 866 /* 867 * Our parents t_fb was not the default, 868 * we need to release our ref on tp->t_fb and 869 * pickup one on the new entry. 870 */ 871 struct tcp_function_block *rblk; 872 873 rblk = find_and_ref_tcp_fb(blk); 874 KASSERT(rblk != NULL, 875 ("cannot find blk %p out of syncache?", blk)); 876 if (tp->t_fb->tfb_tcp_fb_fini) 877 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 878 refcount_release(&tp->t_fb->tfb_refcnt); 879 tp->t_fb = rblk; 880 /* 881 * XXXrrs this is quite dangerous, it is possible 882 * for the new function to fail to init. We also 883 * are not asking if the handoff_is_ok though at 884 * the very start thats probalbly ok. 885 */ 886 if (tp->t_fb->tfb_tcp_fb_init) { 887 (*tp->t_fb->tfb_tcp_fb_init)(tp); 888 } 889 } 890 tp->snd_wl1 = sc->sc_irs; 891 tp->snd_max = tp->iss + 1; 892 tp->snd_nxt = tp->iss + 1; 893 tp->rcv_up = sc->sc_irs + 1; 894 tp->rcv_wnd = sc->sc_wnd; 895 tp->rcv_adv += tp->rcv_wnd; 896 tp->last_ack_sent = tp->rcv_nxt; 897 898 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 899 if (sc->sc_flags & SCF_NOOPT) 900 tp->t_flags |= TF_NOOPT; 901 else { 902 if (sc->sc_flags & SCF_WINSCALE) { 903 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 904 tp->snd_scale = sc->sc_requested_s_scale; 905 tp->request_r_scale = sc->sc_requested_r_scale; 906 } 907 if (sc->sc_flags & SCF_TIMESTAMP) { 908 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 909 tp->ts_recent = sc->sc_tsreflect; 910 tp->ts_recent_age = tcp_ts_getticks(); 911 tp->ts_offset = sc->sc_tsoff; 912 } 913 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 914 if (sc->sc_flags & SCF_SIGNATURE) 915 tp->t_flags |= TF_SIGNATURE; 916 #endif 917 if (sc->sc_flags & SCF_SACK) 918 tp->t_flags |= TF_SACK_PERMIT; 919 } 920 921 if (sc->sc_flags & SCF_ECN) 922 tp->t_flags |= TF_ECN_PERMIT; 923 924 /* 925 * Set up MSS and get cached values from tcp_hostcache. 926 * This might overwrite some of the defaults we just set. 927 */ 928 tcp_mss(tp, sc->sc_peer_mss); 929 930 /* 931 * If the SYN,ACK was retransmitted, indicate that CWND to be 932 * limited to one segment in cc_conn_init(). 933 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 934 */ 935 if (sc->sc_rxmits > 1) 936 tp->snd_cwnd = 1; 937 938 #ifdef TCP_OFFLOAD 939 /* 940 * Allow a TOE driver to install its hooks. Note that we hold the 941 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 942 * new connection before the TOE driver has done its thing. 943 */ 944 if (ADDED_BY_TOE(sc)) { 945 struct toedev *tod = sc->sc_tod; 946 947 tod->tod_offload_socket(tod, sc->sc_todctx, so); 948 } 949 #endif 950 /* 951 * Copy and activate timers. 952 */ 953 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 954 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 955 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 956 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 957 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 958 959 TCPSTAT_INC(tcps_accepts); 960 return (so); 961 962 abort: 963 INP_WUNLOCK(inp); 964 abort2: 965 if (so != NULL) 966 soabort(so); 967 return (NULL); 968 } 969 970 /* 971 * This function gets called when we receive an ACK for a 972 * socket in the LISTEN state. We look up the connection 973 * in the syncache, and if its there, we pull it out of 974 * the cache and turn it into a full-blown connection in 975 * the SYN-RECEIVED state. 976 * 977 * On syncache_socket() success the newly created socket 978 * has its underlying inp locked. 979 */ 980 int 981 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 982 struct socket **lsop, struct mbuf *m) 983 { 984 struct syncache *sc; 985 struct syncache_head *sch; 986 struct syncache scs; 987 char *s; 988 989 /* 990 * Global TCP locks are held because we manipulate the PCB lists 991 * and create a new socket. 992 */ 993 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 994 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 995 ("%s: can handle only ACK", __func__)); 996 997 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 998 SCH_LOCK_ASSERT(sch); 999 1000 #ifdef INVARIANTS 1001 /* 1002 * Test code for syncookies comparing the syncache stored 1003 * values with the reconstructed values from the cookie. 1004 */ 1005 if (sc != NULL) 1006 syncookie_cmp(inc, sch, sc, th, to, *lsop); 1007 #endif 1008 1009 if (sc == NULL) { 1010 /* 1011 * There is no syncache entry, so see if this ACK is 1012 * a returning syncookie. To do this, first: 1013 * A. Check if syncookies are used in case of syncache 1014 * overflows 1015 * B. See if this socket has had a syncache entry dropped in 1016 * the recent past. We don't want to accept a bogus 1017 * syncookie if we've never received a SYN or accept it 1018 * twice. 1019 * C. check that the syncookie is valid. If it is, then 1020 * cobble up a fake syncache entry, and return. 1021 */ 1022 if (!V_tcp_syncookies) { 1023 SCH_UNLOCK(sch); 1024 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1025 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1026 "segment rejected (syncookies disabled)\n", 1027 s, __func__); 1028 goto failed; 1029 } 1030 if (!V_tcp_syncookiesonly && 1031 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1032 SCH_UNLOCK(sch); 1033 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1034 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1035 "segment rejected (no syncache entry)\n", 1036 s, __func__); 1037 goto failed; 1038 } 1039 bzero(&scs, sizeof(scs)); 1040 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1041 SCH_UNLOCK(sch); 1042 if (sc == NULL) { 1043 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1044 log(LOG_DEBUG, "%s; %s: Segment failed " 1045 "SYNCOOKIE authentication, segment rejected " 1046 "(probably spoofed)\n", s, __func__); 1047 goto failed; 1048 } 1049 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1050 /* If received ACK has MD5 signature, check it. */ 1051 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1052 (!TCPMD5_ENABLED() || 1053 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1054 /* Drop the ACK. */ 1055 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1056 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1057 "MD5 signature doesn't match.\n", 1058 s, __func__); 1059 free(s, M_TCPLOG); 1060 } 1061 TCPSTAT_INC(tcps_sig_err_sigopt); 1062 return (-1); /* Do not send RST */ 1063 } 1064 #endif /* TCP_SIGNATURE */ 1065 } else { 1066 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1067 /* 1068 * If listening socket requested TCP digests, check that 1069 * received ACK has signature and it is correct. 1070 * If not, drop the ACK and leave sc entry in th cache, 1071 * because SYN was received with correct signature. 1072 */ 1073 if (sc->sc_flags & SCF_SIGNATURE) { 1074 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1075 /* No signature */ 1076 TCPSTAT_INC(tcps_sig_err_nosigopt); 1077 SCH_UNLOCK(sch); 1078 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1079 log(LOG_DEBUG, "%s; %s: Segment " 1080 "rejected, MD5 signature wasn't " 1081 "provided.\n", s, __func__); 1082 free(s, M_TCPLOG); 1083 } 1084 return (-1); /* Do not send RST */ 1085 } 1086 if (!TCPMD5_ENABLED() || 1087 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1088 /* Doesn't match or no SA */ 1089 SCH_UNLOCK(sch); 1090 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1091 log(LOG_DEBUG, "%s; %s: Segment " 1092 "rejected, MD5 signature doesn't " 1093 "match.\n", s, __func__); 1094 free(s, M_TCPLOG); 1095 } 1096 return (-1); /* Do not send RST */ 1097 } 1098 } 1099 #endif /* TCP_SIGNATURE */ 1100 /* 1101 * Pull out the entry to unlock the bucket row. 1102 * 1103 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1104 * tcp_state_change(). The tcpcb is not existent at this 1105 * moment. A new one will be allocated via syncache_socket-> 1106 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1107 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1108 */ 1109 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1110 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1111 sch->sch_length--; 1112 #ifdef TCP_OFFLOAD 1113 if (ADDED_BY_TOE(sc)) { 1114 struct toedev *tod = sc->sc_tod; 1115 1116 tod->tod_syncache_removed(tod, sc->sc_todctx); 1117 } 1118 #endif 1119 SCH_UNLOCK(sch); 1120 } 1121 1122 /* 1123 * Segment validation: 1124 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1125 */ 1126 if (th->th_ack != sc->sc_iss + 1) { 1127 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1128 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1129 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1130 goto failed; 1131 } 1132 1133 /* 1134 * The SEQ must fall in the window starting at the received 1135 * initial receive sequence number + 1 (the SYN). 1136 */ 1137 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1138 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1139 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1140 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1141 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1142 goto failed; 1143 } 1144 1145 /* 1146 * If timestamps were not negotiated during SYN/ACK they 1147 * must not appear on any segment during this session. 1148 */ 1149 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1150 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1151 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1152 "segment rejected\n", s, __func__); 1153 goto failed; 1154 } 1155 1156 /* 1157 * If timestamps were negotiated during SYN/ACK they should 1158 * appear on every segment during this session. 1159 * XXXAO: This is only informal as there have been unverified 1160 * reports of non-compliants stacks. 1161 */ 1162 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1163 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1164 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1165 "no action\n", s, __func__); 1166 free(s, M_TCPLOG); 1167 s = NULL; 1168 } 1169 } 1170 1171 /* 1172 * If timestamps were negotiated, the reflected timestamp 1173 * must be equal to what we actually sent in the SYN|ACK 1174 * except in the case of 0. Some boxes are known for sending 1175 * broken timestamp replies during the 3whs (and potentially 1176 * during the connection also). 1177 * 1178 * Accept the final ACK of 3whs with reflected timestamp of 0 1179 * instead of sending a RST and deleting the syncache entry. 1180 */ 1181 if ((to->to_flags & TOF_TS) && to->to_tsecr && 1182 to->to_tsecr != sc->sc_ts) { 1183 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1184 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1185 "segment rejected\n", 1186 s, __func__, to->to_tsecr, sc->sc_ts); 1187 goto failed; 1188 } 1189 1190 *lsop = syncache_socket(sc, *lsop, m); 1191 1192 if (*lsop == NULL) 1193 TCPSTAT_INC(tcps_sc_aborted); 1194 else 1195 TCPSTAT_INC(tcps_sc_completed); 1196 1197 /* how do we find the inp for the new socket? */ 1198 if (sc != &scs) 1199 syncache_free(sc); 1200 return (1); 1201 failed: 1202 if (sc != NULL && sc != &scs) 1203 syncache_free(sc); 1204 if (s != NULL) 1205 free(s, M_TCPLOG); 1206 *lsop = NULL; 1207 return (0); 1208 } 1209 1210 static void 1211 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1212 uint64_t response_cookie) 1213 { 1214 struct inpcb *inp; 1215 struct tcpcb *tp; 1216 unsigned int *pending_counter; 1217 1218 /* 1219 * Global TCP locks are held because we manipulate the PCB lists 1220 * and create a new socket. 1221 */ 1222 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1223 1224 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1225 *lsop = syncache_socket(sc, *lsop, m); 1226 if (*lsop == NULL) { 1227 TCPSTAT_INC(tcps_sc_aborted); 1228 atomic_subtract_int(pending_counter, 1); 1229 } else { 1230 soisconnected(*lsop); 1231 inp = sotoinpcb(*lsop); 1232 tp = intotcpcb(inp); 1233 tp->t_flags |= TF_FASTOPEN; 1234 tp->t_tfo_cookie.server = response_cookie; 1235 tp->snd_max = tp->iss; 1236 tp->snd_nxt = tp->iss; 1237 tp->t_tfo_pending = pending_counter; 1238 TCPSTAT_INC(tcps_sc_completed); 1239 } 1240 } 1241 1242 /* 1243 * Given a LISTEN socket and an inbound SYN request, add 1244 * this to the syn cache, and send back a segment: 1245 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1246 * to the source. 1247 * 1248 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1249 * Doing so would require that we hold onto the data and deliver it 1250 * to the application. However, if we are the target of a SYN-flood 1251 * DoS attack, an attacker could send data which would eventually 1252 * consume all available buffer space if it were ACKed. By not ACKing 1253 * the data, we avoid this DoS scenario. 1254 * 1255 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1256 * cookie is processed and a new socket is created. In this case, any data 1257 * accompanying the SYN will be queued to the socket by tcp_input() and will 1258 * be ACKed either when the application sends response data or the delayed 1259 * ACK timer expires, whichever comes first. 1260 */ 1261 int 1262 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1263 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1264 void *todctx) 1265 { 1266 struct tcpcb *tp; 1267 struct socket *so; 1268 struct syncache *sc = NULL; 1269 struct syncache_head *sch; 1270 struct mbuf *ipopts = NULL; 1271 u_int ltflags; 1272 int win, ip_ttl, ip_tos; 1273 char *s; 1274 int rv = 0; 1275 #ifdef INET6 1276 int autoflowlabel = 0; 1277 #endif 1278 #ifdef MAC 1279 struct label *maclabel; 1280 #endif 1281 struct syncache scs; 1282 struct ucred *cred; 1283 uint64_t tfo_response_cookie; 1284 unsigned int *tfo_pending = NULL; 1285 int tfo_cookie_valid = 0; 1286 int tfo_response_cookie_valid = 0; 1287 1288 INP_WLOCK_ASSERT(inp); /* listen socket */ 1289 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1290 ("%s: unexpected tcp flags", __func__)); 1291 1292 /* 1293 * Combine all so/tp operations very early to drop the INP lock as 1294 * soon as possible. 1295 */ 1296 so = *lsop; 1297 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1298 tp = sototcpcb(so); 1299 cred = crhold(so->so_cred); 1300 1301 #ifdef INET6 1302 if ((inc->inc_flags & INC_ISIPV6) && 1303 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1304 autoflowlabel = 1; 1305 #endif 1306 ip_ttl = inp->inp_ip_ttl; 1307 ip_tos = inp->inp_ip_tos; 1308 win = so->sol_sbrcv_hiwat; 1309 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1310 1311 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1312 (tp->t_tfo_pending != NULL) && 1313 (to->to_flags & TOF_FASTOPEN)) { 1314 /* 1315 * Limit the number of pending TFO connections to 1316 * approximately half of the queue limit. This prevents TFO 1317 * SYN floods from starving the service by filling the 1318 * listen queue with bogus TFO connections. 1319 */ 1320 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1321 (so->sol_qlimit / 2)) { 1322 int result; 1323 1324 result = tcp_fastopen_check_cookie(inc, 1325 to->to_tfo_cookie, to->to_tfo_len, 1326 &tfo_response_cookie); 1327 tfo_cookie_valid = (result > 0); 1328 tfo_response_cookie_valid = (result >= 0); 1329 } 1330 1331 /* 1332 * Remember the TFO pending counter as it will have to be 1333 * decremented below if we don't make it to syncache_tfo_expand(). 1334 */ 1335 tfo_pending = tp->t_tfo_pending; 1336 } 1337 1338 /* By the time we drop the lock these should no longer be used. */ 1339 so = NULL; 1340 tp = NULL; 1341 1342 #ifdef MAC 1343 if (mac_syncache_init(&maclabel) != 0) { 1344 INP_WUNLOCK(inp); 1345 goto done; 1346 } else 1347 mac_syncache_create(maclabel, inp); 1348 #endif 1349 if (!tfo_cookie_valid) 1350 INP_WUNLOCK(inp); 1351 1352 /* 1353 * Remember the IP options, if any. 1354 */ 1355 #ifdef INET6 1356 if (!(inc->inc_flags & INC_ISIPV6)) 1357 #endif 1358 #ifdef INET 1359 ipopts = (m) ? ip_srcroute(m) : NULL; 1360 #else 1361 ipopts = NULL; 1362 #endif 1363 1364 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1365 /* 1366 * If listening socket requested TCP digests, check that received 1367 * SYN has signature and it is correct. If signature doesn't match 1368 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1369 */ 1370 if (ltflags & TF_SIGNATURE) { 1371 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1372 TCPSTAT_INC(tcps_sig_err_nosigopt); 1373 goto done; 1374 } 1375 if (!TCPMD5_ENABLED() || 1376 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1377 goto done; 1378 } 1379 #endif /* TCP_SIGNATURE */ 1380 /* 1381 * See if we already have an entry for this connection. 1382 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1383 * 1384 * XXX: should the syncache be re-initialized with the contents 1385 * of the new SYN here (which may have different options?) 1386 * 1387 * XXX: We do not check the sequence number to see if this is a 1388 * real retransmit or a new connection attempt. The question is 1389 * how to handle such a case; either ignore it as spoofed, or 1390 * drop the current entry and create a new one? 1391 */ 1392 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1393 SCH_LOCK_ASSERT(sch); 1394 if (sc != NULL) { 1395 if (tfo_cookie_valid) 1396 INP_WUNLOCK(inp); 1397 TCPSTAT_INC(tcps_sc_dupsyn); 1398 if (ipopts) { 1399 /* 1400 * If we were remembering a previous source route, 1401 * forget it and use the new one we've been given. 1402 */ 1403 if (sc->sc_ipopts) 1404 (void) m_free(sc->sc_ipopts); 1405 sc->sc_ipopts = ipopts; 1406 } 1407 /* 1408 * Update timestamp if present. 1409 */ 1410 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1411 sc->sc_tsreflect = to->to_tsval; 1412 else 1413 sc->sc_flags &= ~SCF_TIMESTAMP; 1414 #ifdef MAC 1415 /* 1416 * Since we have already unconditionally allocated label 1417 * storage, free it up. The syncache entry will already 1418 * have an initialized label we can use. 1419 */ 1420 mac_syncache_destroy(&maclabel); 1421 #endif 1422 /* Retransmit SYN|ACK and reset retransmit count. */ 1423 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1424 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1425 "resetting timer and retransmitting SYN|ACK\n", 1426 s, __func__); 1427 free(s, M_TCPLOG); 1428 } 1429 if (syncache_respond(sc, sch, 1, m) == 0) { 1430 sc->sc_rxmits = 0; 1431 syncache_timeout(sc, sch, 1); 1432 TCPSTAT_INC(tcps_sndacks); 1433 TCPSTAT_INC(tcps_sndtotal); 1434 } 1435 SCH_UNLOCK(sch); 1436 goto done; 1437 } 1438 1439 if (tfo_cookie_valid) { 1440 bzero(&scs, sizeof(scs)); 1441 sc = &scs; 1442 goto skip_alloc; 1443 } 1444 1445 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1446 if (sc == NULL) { 1447 /* 1448 * The zone allocator couldn't provide more entries. 1449 * Treat this as if the cache was full; drop the oldest 1450 * entry and insert the new one. 1451 */ 1452 TCPSTAT_INC(tcps_sc_zonefail); 1453 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1454 sch->sch_last_overflow = time_uptime; 1455 syncache_drop(sc, sch); 1456 } 1457 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1458 if (sc == NULL) { 1459 if (V_tcp_syncookies) { 1460 bzero(&scs, sizeof(scs)); 1461 sc = &scs; 1462 } else { 1463 SCH_UNLOCK(sch); 1464 if (ipopts) 1465 (void) m_free(ipopts); 1466 goto done; 1467 } 1468 } 1469 } 1470 1471 skip_alloc: 1472 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1473 sc->sc_tfo_cookie = &tfo_response_cookie; 1474 1475 /* 1476 * Fill in the syncache values. 1477 */ 1478 #ifdef MAC 1479 sc->sc_label = maclabel; 1480 #endif 1481 sc->sc_cred = cred; 1482 cred = NULL; 1483 sc->sc_ipopts = ipopts; 1484 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1485 #ifdef INET6 1486 if (!(inc->inc_flags & INC_ISIPV6)) 1487 #endif 1488 { 1489 sc->sc_ip_tos = ip_tos; 1490 sc->sc_ip_ttl = ip_ttl; 1491 } 1492 #ifdef TCP_OFFLOAD 1493 sc->sc_tod = tod; 1494 sc->sc_todctx = todctx; 1495 #endif 1496 sc->sc_irs = th->th_seq; 1497 sc->sc_iss = arc4random(); 1498 sc->sc_flags = 0; 1499 sc->sc_flowlabel = 0; 1500 1501 /* 1502 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1503 * win was derived from socket earlier in the function. 1504 */ 1505 win = imax(win, 0); 1506 win = imin(win, TCP_MAXWIN); 1507 sc->sc_wnd = win; 1508 1509 if (V_tcp_do_rfc1323) { 1510 /* 1511 * A timestamp received in a SYN makes 1512 * it ok to send timestamp requests and replies. 1513 */ 1514 if (to->to_flags & TOF_TS) { 1515 sc->sc_tsreflect = to->to_tsval; 1516 sc->sc_ts = tcp_ts_getticks(); 1517 sc->sc_flags |= SCF_TIMESTAMP; 1518 } 1519 if (to->to_flags & TOF_SCALE) { 1520 int wscale = 0; 1521 1522 /* 1523 * Pick the smallest possible scaling factor that 1524 * will still allow us to scale up to sb_max, aka 1525 * kern.ipc.maxsockbuf. 1526 * 1527 * We do this because there are broken firewalls that 1528 * will corrupt the window scale option, leading to 1529 * the other endpoint believing that our advertised 1530 * window is unscaled. At scale factors larger than 1531 * 5 the unscaled window will drop below 1500 bytes, 1532 * leading to serious problems when traversing these 1533 * broken firewalls. 1534 * 1535 * With the default maxsockbuf of 256K, a scale factor 1536 * of 3 will be chosen by this algorithm. Those who 1537 * choose a larger maxsockbuf should watch out 1538 * for the compatibility problems mentioned above. 1539 * 1540 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1541 * or <SYN,ACK>) segment itself is never scaled. 1542 */ 1543 while (wscale < TCP_MAX_WINSHIFT && 1544 (TCP_MAXWIN << wscale) < sb_max) 1545 wscale++; 1546 sc->sc_requested_r_scale = wscale; 1547 sc->sc_requested_s_scale = to->to_wscale; 1548 sc->sc_flags |= SCF_WINSCALE; 1549 } 1550 } 1551 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1552 /* 1553 * If listening socket requested TCP digests, flag this in the 1554 * syncache so that syncache_respond() will do the right thing 1555 * with the SYN+ACK. 1556 */ 1557 if (ltflags & TF_SIGNATURE) 1558 sc->sc_flags |= SCF_SIGNATURE; 1559 #endif /* TCP_SIGNATURE */ 1560 if (to->to_flags & TOF_SACKPERM) 1561 sc->sc_flags |= SCF_SACK; 1562 if (to->to_flags & TOF_MSS) 1563 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1564 if (ltflags & TF_NOOPT) 1565 sc->sc_flags |= SCF_NOOPT; 1566 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1567 sc->sc_flags |= SCF_ECN; 1568 1569 if (V_tcp_syncookies) 1570 sc->sc_iss = syncookie_generate(sch, sc); 1571 #ifdef INET6 1572 if (autoflowlabel) { 1573 if (V_tcp_syncookies) 1574 sc->sc_flowlabel = sc->sc_iss; 1575 else 1576 sc->sc_flowlabel = ip6_randomflowlabel(); 1577 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1578 } 1579 #endif 1580 SCH_UNLOCK(sch); 1581 1582 if (tfo_cookie_valid) { 1583 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1584 /* INP_WUNLOCK(inp) will be performed by the caller */ 1585 rv = 1; 1586 goto tfo_expanded; 1587 } 1588 1589 /* 1590 * Do a standard 3-way handshake. 1591 */ 1592 if (syncache_respond(sc, sch, 0, m) == 0) { 1593 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1594 syncache_free(sc); 1595 else if (sc != &scs) 1596 syncache_insert(sc, sch); /* locks and unlocks sch */ 1597 TCPSTAT_INC(tcps_sndacks); 1598 TCPSTAT_INC(tcps_sndtotal); 1599 } else { 1600 if (sc != &scs) 1601 syncache_free(sc); 1602 TCPSTAT_INC(tcps_sc_dropped); 1603 } 1604 1605 done: 1606 if (m) { 1607 *lsop = NULL; 1608 m_freem(m); 1609 } 1610 /* 1611 * If tfo_pending is not NULL here, then a TFO SYN that did not 1612 * result in a new socket was processed and the associated pending 1613 * counter has not yet been decremented. All such TFO processing paths 1614 * transit this point. 1615 */ 1616 if (tfo_pending != NULL) 1617 tcp_fastopen_decrement_counter(tfo_pending); 1618 1619 tfo_expanded: 1620 if (cred != NULL) 1621 crfree(cred); 1622 #ifdef MAC 1623 if (sc == &scs) 1624 mac_syncache_destroy(&maclabel); 1625 #endif 1626 return (rv); 1627 } 1628 1629 /* 1630 * Send SYN|ACK to the peer. Either in response to the peer's SYN, 1631 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1632 */ 1633 static int 1634 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked, 1635 const struct mbuf *m0) 1636 { 1637 struct ip *ip = NULL; 1638 struct mbuf *m; 1639 struct tcphdr *th = NULL; 1640 int optlen, error = 0; /* Make compiler happy */ 1641 u_int16_t hlen, tlen, mssopt; 1642 struct tcpopt to; 1643 #ifdef INET6 1644 struct ip6_hdr *ip6 = NULL; 1645 #endif 1646 hlen = 1647 #ifdef INET6 1648 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1649 #endif 1650 sizeof(struct ip); 1651 tlen = hlen + sizeof(struct tcphdr); 1652 1653 /* Determine MSS we advertize to other end of connection. */ 1654 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); 1655 1656 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1657 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1658 ("syncache: mbuf too small")); 1659 1660 /* Create the IP+TCP header from scratch. */ 1661 m = m_gethdr(M_NOWAIT, MT_DATA); 1662 if (m == NULL) 1663 return (ENOBUFS); 1664 #ifdef MAC 1665 mac_syncache_create_mbuf(sc->sc_label, m); 1666 #endif 1667 m->m_data += max_linkhdr; 1668 m->m_len = tlen; 1669 m->m_pkthdr.len = tlen; 1670 m->m_pkthdr.rcvif = NULL; 1671 1672 #ifdef INET6 1673 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1674 ip6 = mtod(m, struct ip6_hdr *); 1675 ip6->ip6_vfc = IPV6_VERSION; 1676 ip6->ip6_nxt = IPPROTO_TCP; 1677 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1678 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1679 ip6->ip6_plen = htons(tlen - hlen); 1680 /* ip6_hlim is set after checksum */ 1681 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1682 ip6->ip6_flow |= sc->sc_flowlabel; 1683 1684 th = (struct tcphdr *)(ip6 + 1); 1685 } 1686 #endif 1687 #if defined(INET6) && defined(INET) 1688 else 1689 #endif 1690 #ifdef INET 1691 { 1692 ip = mtod(m, struct ip *); 1693 ip->ip_v = IPVERSION; 1694 ip->ip_hl = sizeof(struct ip) >> 2; 1695 ip->ip_len = htons(tlen); 1696 ip->ip_id = 0; 1697 ip->ip_off = 0; 1698 ip->ip_sum = 0; 1699 ip->ip_p = IPPROTO_TCP; 1700 ip->ip_src = sc->sc_inc.inc_laddr; 1701 ip->ip_dst = sc->sc_inc.inc_faddr; 1702 ip->ip_ttl = sc->sc_ip_ttl; 1703 ip->ip_tos = sc->sc_ip_tos; 1704 1705 /* 1706 * See if we should do MTU discovery. Route lookups are 1707 * expensive, so we will only unset the DF bit if: 1708 * 1709 * 1) path_mtu_discovery is disabled 1710 * 2) the SCF_UNREACH flag has been set 1711 */ 1712 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1713 ip->ip_off |= htons(IP_DF); 1714 1715 th = (struct tcphdr *)(ip + 1); 1716 } 1717 #endif /* INET */ 1718 th->th_sport = sc->sc_inc.inc_lport; 1719 th->th_dport = sc->sc_inc.inc_fport; 1720 1721 th->th_seq = htonl(sc->sc_iss); 1722 th->th_ack = htonl(sc->sc_irs + 1); 1723 th->th_off = sizeof(struct tcphdr) >> 2; 1724 th->th_x2 = 0; 1725 th->th_flags = TH_SYN|TH_ACK; 1726 th->th_win = htons(sc->sc_wnd); 1727 th->th_urp = 0; 1728 1729 if (sc->sc_flags & SCF_ECN) { 1730 th->th_flags |= TH_ECE; 1731 TCPSTAT_INC(tcps_ecn_shs); 1732 } 1733 1734 /* Tack on the TCP options. */ 1735 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1736 to.to_flags = 0; 1737 1738 to.to_mss = mssopt; 1739 to.to_flags = TOF_MSS; 1740 if (sc->sc_flags & SCF_WINSCALE) { 1741 to.to_wscale = sc->sc_requested_r_scale; 1742 to.to_flags |= TOF_SCALE; 1743 } 1744 if (sc->sc_flags & SCF_TIMESTAMP) { 1745 /* Virgin timestamp or TCP cookie enhanced one. */ 1746 to.to_tsval = sc->sc_ts; 1747 to.to_tsecr = sc->sc_tsreflect; 1748 to.to_flags |= TOF_TS; 1749 } 1750 if (sc->sc_flags & SCF_SACK) 1751 to.to_flags |= TOF_SACKPERM; 1752 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1753 if (sc->sc_flags & SCF_SIGNATURE) 1754 to.to_flags |= TOF_SIGNATURE; 1755 #endif 1756 if (sc->sc_tfo_cookie) { 1757 to.to_flags |= TOF_FASTOPEN; 1758 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1759 to.to_tfo_cookie = sc->sc_tfo_cookie; 1760 /* don't send cookie again when retransmitting response */ 1761 sc->sc_tfo_cookie = NULL; 1762 } 1763 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1764 1765 /* Adjust headers by option size. */ 1766 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1767 m->m_len += optlen; 1768 m->m_pkthdr.len += optlen; 1769 #ifdef INET6 1770 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1771 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1772 else 1773 #endif 1774 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1775 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1776 if (sc->sc_flags & SCF_SIGNATURE) { 1777 KASSERT(to.to_flags & TOF_SIGNATURE, 1778 ("tcp_addoptions() didn't set tcp_signature")); 1779 1780 /* NOTE: to.to_signature is inside of mbuf */ 1781 if (!TCPMD5_ENABLED() || 1782 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1783 m_freem(m); 1784 return (EACCES); 1785 } 1786 } 1787 #endif 1788 } else 1789 optlen = 0; 1790 1791 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1792 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1793 /* 1794 * If we have peer's SYN and it has a flowid, then let's assign it to 1795 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1796 * to SYN|ACK due to lack of inp here. 1797 */ 1798 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1799 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1800 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1801 } 1802 #ifdef INET6 1803 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1804 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1805 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1806 IPPROTO_TCP, 0); 1807 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1808 #ifdef TCP_OFFLOAD 1809 if (ADDED_BY_TOE(sc)) { 1810 struct toedev *tod = sc->sc_tod; 1811 1812 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1813 1814 return (error); 1815 } 1816 #endif 1817 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1818 } 1819 #endif 1820 #if defined(INET6) && defined(INET) 1821 else 1822 #endif 1823 #ifdef INET 1824 { 1825 m->m_pkthdr.csum_flags = CSUM_TCP; 1826 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1827 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1828 #ifdef TCP_OFFLOAD 1829 if (ADDED_BY_TOE(sc)) { 1830 struct toedev *tod = sc->sc_tod; 1831 1832 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1833 1834 return (error); 1835 } 1836 #endif 1837 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1838 } 1839 #endif 1840 return (error); 1841 } 1842 1843 /* 1844 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1845 * that exceed the capacity of the syncache by avoiding the storage of any 1846 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1847 * attacks where the attacker does not have access to our responses. 1848 * 1849 * Syncookies encode and include all necessary information about the 1850 * connection setup within the SYN|ACK that we send back. That way we 1851 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1852 * (if ever). Normally the syncache and syncookies are running in parallel 1853 * with the latter taking over when the former is exhausted. When matching 1854 * syncache entry is found the syncookie is ignored. 1855 * 1856 * The only reliable information persisting the 3WHS is our initial sequence 1857 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1858 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1859 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1860 * returns and signifies a legitimate connection if it matches the ACK. 1861 * 1862 * The available space of 32 bits to store the hash and to encode the SYN 1863 * option information is very tight and we should have at least 24 bits for 1864 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1865 * 1866 * SYN option information we have to encode to fully restore a connection: 1867 * MSS: is imporant to chose an optimal segment size to avoid IP level 1868 * fragmentation along the path. The common MSS values can be encoded 1869 * in a 3-bit table. Uncommon values are captured by the next lower value 1870 * in the table leading to a slight increase in packetization overhead. 1871 * WSCALE: is necessary to allow large windows to be used for high delay- 1872 * bandwidth product links. Not scaling the window when it was initially 1873 * negotiated is bad for performance as lack of scaling further decreases 1874 * the apparent available send window. We only need to encode the WSCALE 1875 * we received from the remote end. Our end can be recalculated at any 1876 * time. The common WSCALE values can be encoded in a 3-bit table. 1877 * Uncommon values are captured by the next lower value in the table 1878 * making us under-estimate the available window size halving our 1879 * theoretically possible maximum throughput for that connection. 1880 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1881 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1882 * that are included in all segments on a connection. We enable them when 1883 * the ACK has them. 1884 * 1885 * Security of syncookies and attack vectors: 1886 * 1887 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1888 * together with the gloabl secret to make it unique per connection attempt. 1889 * Thus any change of any of those parameters results in a different MAC output 1890 * in an unpredictable way unless a collision is encountered. 24 bits of the 1891 * MAC are embedded into the ISS. 1892 * 1893 * To prevent replay attacks two rotating global secrets are updated with a 1894 * new random value every 15 seconds. The life-time of a syncookie is thus 1895 * 15-30 seconds. 1896 * 1897 * Vector 1: Attacking the secret. This requires finding a weakness in the 1898 * MAC itself or the way it is used here. The attacker can do a chosen plain 1899 * text attack by varying and testing the all parameters under his control. 1900 * The strength depends on the size and randomness of the secret, and the 1901 * cryptographic security of the MAC function. Due to the constant updating 1902 * of the secret the attacker has at most 29.999 seconds to find the secret 1903 * and launch spoofed connections. After that he has to start all over again. 1904 * 1905 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1906 * size an average of 4,823 attempts are required for a 50% chance of success 1907 * to spoof a single syncookie (birthday collision paradox). However the 1908 * attacker is blind and doesn't know if one of his attempts succeeded unless 1909 * he has a side channel to interfere success from. A single connection setup 1910 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1911 * This many attempts are required for each one blind spoofed connection. For 1912 * every additional spoofed connection he has to launch another N attempts. 1913 * Thus for a sustained rate 100 spoofed connections per second approximately 1914 * 1,800,000 packets per second would have to be sent. 1915 * 1916 * NB: The MAC function should be fast so that it doesn't become a CPU 1917 * exhaustion attack vector itself. 1918 * 1919 * References: 1920 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1921 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1922 * http://cr.yp.to/syncookies.html (overview) 1923 * http://cr.yp.to/syncookies/archive (details) 1924 * 1925 * 1926 * Schematic construction of a syncookie enabled Initial Sequence Number: 1927 * 0 1 2 3 1928 * 12345678901234567890123456789012 1929 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1930 * 1931 * x 24 MAC (truncated) 1932 * W 3 Send Window Scale index 1933 * M 3 MSS index 1934 * S 1 SACK permitted 1935 * P 1 Odd/even secret 1936 */ 1937 1938 /* 1939 * Distribution and probability of certain MSS values. Those in between are 1940 * rounded down to the next lower one. 1941 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1942 * .2% .3% 5% 7% 7% 20% 15% 45% 1943 */ 1944 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1945 1946 /* 1947 * Distribution and probability of certain WSCALE values. We have to map the 1948 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1949 * bits based on prevalence of certain values. Where we don't have an exact 1950 * match for are rounded down to the next lower one letting us under-estimate 1951 * the true available window. At the moment this would happen only for the 1952 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1953 * and window size). The absence of the WSCALE option (no scaling in either 1954 * direction) is encoded with index zero. 1955 * [WSCALE values histograms, Allman, 2012] 1956 * X 10 10 35 5 6 14 10% by host 1957 * X 11 4 5 5 18 49 3% by connections 1958 */ 1959 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1960 1961 /* 1962 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1963 * and good cryptographic properties. 1964 */ 1965 static uint32_t 1966 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1967 uint8_t *secbits, uintptr_t secmod) 1968 { 1969 SIPHASH_CTX ctx; 1970 uint32_t siphash[2]; 1971 1972 SipHash24_Init(&ctx); 1973 SipHash_SetKey(&ctx, secbits); 1974 switch (inc->inc_flags & INC_ISIPV6) { 1975 #ifdef INET 1976 case 0: 1977 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1978 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1979 break; 1980 #endif 1981 #ifdef INET6 1982 case INC_ISIPV6: 1983 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1984 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1985 break; 1986 #endif 1987 } 1988 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1989 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1990 SipHash_Update(&ctx, &irs, sizeof(irs)); 1991 SipHash_Update(&ctx, &flags, sizeof(flags)); 1992 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1993 SipHash_Final((u_int8_t *)&siphash, &ctx); 1994 1995 return (siphash[0] ^ siphash[1]); 1996 } 1997 1998 static tcp_seq 1999 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2000 { 2001 u_int i, secbit, wscale; 2002 uint32_t iss, hash; 2003 uint8_t *secbits; 2004 union syncookie cookie; 2005 2006 SCH_LOCK_ASSERT(sch); 2007 2008 cookie.cookie = 0; 2009 2010 /* Map our computed MSS into the 3-bit index. */ 2011 for (i = nitems(tcp_sc_msstab) - 1; 2012 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2013 i--) 2014 ; 2015 cookie.flags.mss_idx = i; 2016 2017 /* 2018 * Map the send window scale into the 3-bit index but only if 2019 * the wscale option was received. 2020 */ 2021 if (sc->sc_flags & SCF_WINSCALE) { 2022 wscale = sc->sc_requested_s_scale; 2023 for (i = nitems(tcp_sc_wstab) - 1; 2024 tcp_sc_wstab[i] > wscale && i > 0; 2025 i--) 2026 ; 2027 cookie.flags.wscale_idx = i; 2028 } 2029 2030 /* Can we do SACK? */ 2031 if (sc->sc_flags & SCF_SACK) 2032 cookie.flags.sack_ok = 1; 2033 2034 /* Which of the two secrets to use. */ 2035 secbit = sch->sch_sc->secret.oddeven & 0x1; 2036 cookie.flags.odd_even = secbit; 2037 2038 secbits = sch->sch_sc->secret.key[secbit]; 2039 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2040 (uintptr_t)sch); 2041 2042 /* 2043 * Put the flags into the hash and XOR them to get better ISS number 2044 * variance. This doesn't enhance the cryptographic strength and is 2045 * done to prevent the 8 cookie bits from showing up directly on the 2046 * wire. 2047 */ 2048 iss = hash & ~0xff; 2049 iss |= cookie.cookie ^ (hash >> 24); 2050 2051 /* Randomize the timestamp. */ 2052 if (sc->sc_flags & SCF_TIMESTAMP) { 2053 sc->sc_ts = arc4random(); 2054 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 2055 } 2056 2057 TCPSTAT_INC(tcps_sc_sendcookie); 2058 return (iss); 2059 } 2060 2061 static struct syncache * 2062 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2063 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2064 struct socket *lso) 2065 { 2066 uint32_t hash; 2067 uint8_t *secbits; 2068 tcp_seq ack, seq; 2069 int wnd, wscale = 0; 2070 union syncookie cookie; 2071 2072 SCH_LOCK_ASSERT(sch); 2073 2074 /* 2075 * Pull information out of SYN-ACK/ACK and revert sequence number 2076 * advances. 2077 */ 2078 ack = th->th_ack - 1; 2079 seq = th->th_seq - 1; 2080 2081 /* 2082 * Unpack the flags containing enough information to restore the 2083 * connection. 2084 */ 2085 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2086 2087 /* Which of the two secrets to use. */ 2088 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 2089 2090 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2091 2092 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2093 if ((ack & ~0xff) != (hash & ~0xff)) 2094 return (NULL); 2095 2096 /* Fill in the syncache values. */ 2097 sc->sc_flags = 0; 2098 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2099 sc->sc_ipopts = NULL; 2100 2101 sc->sc_irs = seq; 2102 sc->sc_iss = ack; 2103 2104 switch (inc->inc_flags & INC_ISIPV6) { 2105 #ifdef INET 2106 case 0: 2107 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2108 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2109 break; 2110 #endif 2111 #ifdef INET6 2112 case INC_ISIPV6: 2113 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2114 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 2115 break; 2116 #endif 2117 } 2118 2119 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2120 2121 /* We can simply recompute receive window scale we sent earlier. */ 2122 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2123 wscale++; 2124 2125 /* Only use wscale if it was enabled in the orignal SYN. */ 2126 if (cookie.flags.wscale_idx > 0) { 2127 sc->sc_requested_r_scale = wscale; 2128 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2129 sc->sc_flags |= SCF_WINSCALE; 2130 } 2131 2132 wnd = lso->sol_sbrcv_hiwat; 2133 wnd = imax(wnd, 0); 2134 wnd = imin(wnd, TCP_MAXWIN); 2135 sc->sc_wnd = wnd; 2136 2137 if (cookie.flags.sack_ok) 2138 sc->sc_flags |= SCF_SACK; 2139 2140 if (to->to_flags & TOF_TS) { 2141 sc->sc_flags |= SCF_TIMESTAMP; 2142 sc->sc_tsreflect = to->to_tsval; 2143 sc->sc_ts = to->to_tsecr; 2144 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 2145 } 2146 2147 if (to->to_flags & TOF_SIGNATURE) 2148 sc->sc_flags |= SCF_SIGNATURE; 2149 2150 sc->sc_rxmits = 0; 2151 2152 TCPSTAT_INC(tcps_sc_recvcookie); 2153 return (sc); 2154 } 2155 2156 #ifdef INVARIANTS 2157 static int 2158 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2159 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2160 struct socket *lso) 2161 { 2162 struct syncache scs, *scx; 2163 char *s; 2164 2165 bzero(&scs, sizeof(scs)); 2166 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2167 2168 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2169 return (0); 2170 2171 if (scx != NULL) { 2172 if (sc->sc_peer_mss != scx->sc_peer_mss) 2173 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2174 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2175 2176 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2177 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2178 s, __func__, sc->sc_requested_r_scale, 2179 scx->sc_requested_r_scale); 2180 2181 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2182 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2183 s, __func__, sc->sc_requested_s_scale, 2184 scx->sc_requested_s_scale); 2185 2186 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2187 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2188 } 2189 2190 if (s != NULL) 2191 free(s, M_TCPLOG); 2192 return (0); 2193 } 2194 #endif /* INVARIANTS */ 2195 2196 static void 2197 syncookie_reseed(void *arg) 2198 { 2199 struct tcp_syncache *sc = arg; 2200 uint8_t *secbits; 2201 int secbit; 2202 2203 /* 2204 * Reseeding the secret doesn't have to be protected by a lock. 2205 * It only must be ensured that the new random values are visible 2206 * to all CPUs in a SMP environment. The atomic with release 2207 * semantics ensures that. 2208 */ 2209 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2210 secbits = sc->secret.key[secbit]; 2211 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2212 atomic_add_rel_int(&sc->secret.oddeven, 1); 2213 2214 /* Reschedule ourself. */ 2215 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2216 } 2217 2218 /* 2219 * Exports the syncache entries to userland so that netstat can display 2220 * them alongside the other sockets. This function is intended to be 2221 * called only from tcp_pcblist. 2222 * 2223 * Due to concurrency on an active system, the number of pcbs exported 2224 * may have no relation to max_pcbs. max_pcbs merely indicates the 2225 * amount of space the caller allocated for this function to use. 2226 */ 2227 int 2228 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2229 { 2230 struct xtcpcb xt; 2231 struct syncache *sc; 2232 struct syncache_head *sch; 2233 int count, error, i; 2234 2235 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2236 sch = &V_tcp_syncache.hashbase[i]; 2237 SCH_LOCK(sch); 2238 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2239 if (count >= max_pcbs) { 2240 SCH_UNLOCK(sch); 2241 goto exit; 2242 } 2243 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2244 continue; 2245 bzero(&xt, sizeof(xt)); 2246 xt.xt_len = sizeof(xt); 2247 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2248 xt.xt_inp.inp_vflag = INP_IPV6; 2249 else 2250 xt.xt_inp.inp_vflag = INP_IPV4; 2251 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2252 sizeof (struct in_conninfo)); 2253 xt.t_state = TCPS_SYN_RECEIVED; 2254 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2255 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2256 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2257 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2258 error = SYSCTL_OUT(req, &xt, sizeof xt); 2259 if (error) { 2260 SCH_UNLOCK(sch); 2261 goto exit; 2262 } 2263 count++; 2264 } 2265 SCH_UNLOCK(sch); 2266 } 2267 exit: 2268 *pcbs_exported = count; 2269 return error; 2270 } 2271