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 *, 134 const struct mbuf *, int); 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 with default values of 158 * tcp_rexmit_initial * ( 1 + 159 * tcp_backoff[1] + 160 * tcp_backoff[2] + 161 * tcp_backoff[3]) + 3 * tcp_rexmit_slop, 162 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms, 163 * the odds are that the user has given up attempting to connect by then. 164 */ 165 #define SYNCACHE_MAXREXMTS 3 166 167 /* Arbitrary values */ 168 #define TCP_SYNCACHE_HASHSIZE 512 169 #define TCP_SYNCACHE_BUCKETLIMIT 30 170 171 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 172 #define V_tcp_syncache VNET(tcp_syncache) 173 174 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, 175 "TCP SYN cache"); 176 177 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 178 &VNET_NAME(tcp_syncache.bucket_limit), 0, 179 "Per-bucket hash limit for syncache"); 180 181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 182 &VNET_NAME(tcp_syncache.cache_limit), 0, 183 "Overall entry limit for syncache"); 184 185 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 186 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 187 188 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 189 &VNET_NAME(tcp_syncache.hashsize), 0, 190 "Size of TCP syncache hashtable"); 191 192 static int 193 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 194 { 195 int error; 196 u_int new; 197 198 new = V_tcp_syncache.rexmt_limit; 199 error = sysctl_handle_int(oidp, &new, 0, req); 200 if ((error == 0) && (req->newptr != NULL)) { 201 if (new > TCP_MAXRXTSHIFT) 202 error = EINVAL; 203 else 204 V_tcp_syncache.rexmt_limit = new; 205 } 206 return (error); 207 } 208 209 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 210 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW, 211 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 212 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI", 213 "Limit on SYN/ACK retransmissions"); 214 215 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 216 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 217 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 218 "Send reset on socket allocation failure"); 219 220 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 221 222 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 223 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 224 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 225 226 /* 227 * Requires the syncache entry to be already removed from the bucket list. 228 */ 229 static void 230 syncache_free(struct syncache *sc) 231 { 232 233 if (sc->sc_ipopts) 234 (void) m_free(sc->sc_ipopts); 235 if (sc->sc_cred) 236 crfree(sc->sc_cred); 237 #ifdef MAC 238 mac_syncache_destroy(&sc->sc_label); 239 #endif 240 241 uma_zfree(V_tcp_syncache.zone, sc); 242 } 243 244 void 245 syncache_init(void) 246 { 247 int i; 248 249 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 250 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 251 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 252 V_tcp_syncache.hash_secret = arc4random(); 253 254 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 255 &V_tcp_syncache.hashsize); 256 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 257 &V_tcp_syncache.bucket_limit); 258 if (!powerof2(V_tcp_syncache.hashsize) || 259 V_tcp_syncache.hashsize == 0) { 260 printf("WARNING: syncache hash size is not a power of 2.\n"); 261 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 262 } 263 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 264 265 /* Set limits. */ 266 V_tcp_syncache.cache_limit = 267 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 268 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 269 &V_tcp_syncache.cache_limit); 270 271 /* Allocate the hash table. */ 272 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 273 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 274 275 #ifdef VIMAGE 276 V_tcp_syncache.vnet = curvnet; 277 #endif 278 279 /* Initialize the hash buckets. */ 280 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 281 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 282 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 283 NULL, MTX_DEF); 284 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 285 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 286 V_tcp_syncache.hashbase[i].sch_length = 0; 287 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 288 V_tcp_syncache.hashbase[i].sch_last_overflow = 289 -(SYNCOOKIE_LIFETIME + 1); 290 } 291 292 /* Create the syncache entry zone. */ 293 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 294 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 295 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 296 V_tcp_syncache.cache_limit); 297 298 /* Start the SYN cookie reseeder callout. */ 299 callout_init(&V_tcp_syncache.secret.reseed, 1); 300 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 301 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 302 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 303 syncookie_reseed, &V_tcp_syncache); 304 } 305 306 #ifdef VIMAGE 307 void 308 syncache_destroy(void) 309 { 310 struct syncache_head *sch; 311 struct syncache *sc, *nsc; 312 int i; 313 314 /* 315 * Stop the re-seed timer before freeing resources. No need to 316 * possibly schedule it another time. 317 */ 318 callout_drain(&V_tcp_syncache.secret.reseed); 319 320 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 321 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 322 323 sch = &V_tcp_syncache.hashbase[i]; 324 callout_drain(&sch->sch_timer); 325 326 SCH_LOCK(sch); 327 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 328 syncache_drop(sc, sch); 329 SCH_UNLOCK(sch); 330 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 331 ("%s: sch->sch_bucket not empty", __func__)); 332 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 333 __func__, sch->sch_length)); 334 mtx_destroy(&sch->sch_mtx); 335 } 336 337 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 338 ("%s: cache_count not 0", __func__)); 339 340 /* Free the allocated global resources. */ 341 uma_zdestroy(V_tcp_syncache.zone); 342 free(V_tcp_syncache.hashbase, M_SYNCACHE); 343 } 344 #endif 345 346 /* 347 * Inserts a syncache entry into the specified bucket row. 348 * Locks and unlocks the syncache_head autonomously. 349 */ 350 static void 351 syncache_insert(struct syncache *sc, struct syncache_head *sch) 352 { 353 struct syncache *sc2; 354 355 SCH_LOCK(sch); 356 357 /* 358 * Make sure that we don't overflow the per-bucket limit. 359 * If the bucket is full, toss the oldest element. 360 */ 361 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 362 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 363 ("sch->sch_length incorrect")); 364 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 365 sch->sch_last_overflow = time_uptime; 366 syncache_drop(sc2, sch); 367 TCPSTAT_INC(tcps_sc_bucketoverflow); 368 } 369 370 /* Put it into the bucket. */ 371 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 372 sch->sch_length++; 373 374 #ifdef TCP_OFFLOAD 375 if (ADDED_BY_TOE(sc)) { 376 struct toedev *tod = sc->sc_tod; 377 378 tod->tod_syncache_added(tod, sc->sc_todctx); 379 } 380 #endif 381 382 /* Reinitialize the bucket row's timer. */ 383 if (sch->sch_length == 1) 384 sch->sch_nextc = ticks + INT_MAX; 385 syncache_timeout(sc, sch, 1); 386 387 SCH_UNLOCK(sch); 388 389 TCPSTATES_INC(TCPS_SYN_RECEIVED); 390 TCPSTAT_INC(tcps_sc_added); 391 } 392 393 /* 394 * Remove and free entry from syncache bucket row. 395 * Expects locked syncache head. 396 */ 397 static void 398 syncache_drop(struct syncache *sc, struct syncache_head *sch) 399 { 400 401 SCH_LOCK_ASSERT(sch); 402 403 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 404 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 405 sch->sch_length--; 406 407 #ifdef TCP_OFFLOAD 408 if (ADDED_BY_TOE(sc)) { 409 struct toedev *tod = sc->sc_tod; 410 411 tod->tod_syncache_removed(tod, sc->sc_todctx); 412 } 413 #endif 414 415 syncache_free(sc); 416 } 417 418 /* 419 * Engage/reengage time on bucket row. 420 */ 421 static void 422 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 423 { 424 int rexmt; 425 426 if (sc->sc_rxmits == 0) 427 rexmt = tcp_rexmit_initial; 428 else 429 TCPT_RANGESET(rexmt, 430 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], 431 tcp_rexmit_min, TCPTV_REXMTMAX); 432 sc->sc_rxttime = ticks + rexmt; 433 sc->sc_rxmits++; 434 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 435 sch->sch_nextc = sc->sc_rxttime; 436 if (docallout) 437 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 438 syncache_timer, (void *)sch); 439 } 440 } 441 442 /* 443 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 444 * If we have retransmitted an entry the maximum number of times, expire it. 445 * One separate timer for each bucket row. 446 */ 447 static void 448 syncache_timer(void *xsch) 449 { 450 struct syncache_head *sch = (struct syncache_head *)xsch; 451 struct syncache *sc, *nsc; 452 int tick = ticks; 453 char *s; 454 455 CURVNET_SET(sch->sch_sc->vnet); 456 457 /* NB: syncache_head has already been locked by the callout. */ 458 SCH_LOCK_ASSERT(sch); 459 460 /* 461 * In the following cycle we may remove some entries and/or 462 * advance some timeouts, so re-initialize the bucket timer. 463 */ 464 sch->sch_nextc = tick + INT_MAX; 465 466 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 467 /* 468 * We do not check if the listen socket still exists 469 * and accept the case where the listen socket may be 470 * gone by the time we resend the SYN/ACK. We do 471 * not expect this to happens often. If it does, 472 * then the RST will be sent by the time the remote 473 * host does the SYN/ACK->ACK. 474 */ 475 if (TSTMP_GT(sc->sc_rxttime, tick)) { 476 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 477 sch->sch_nextc = sc->sc_rxttime; 478 continue; 479 } 480 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 481 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 482 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 483 "giving up and removing syncache entry\n", 484 s, __func__); 485 free(s, M_TCPLOG); 486 } 487 syncache_drop(sc, sch); 488 TCPSTAT_INC(tcps_sc_stale); 489 continue; 490 } 491 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 492 log(LOG_DEBUG, "%s; %s: Response timeout, " 493 "retransmitting (%u) SYN|ACK\n", 494 s, __func__, sc->sc_rxmits); 495 free(s, M_TCPLOG); 496 } 497 498 syncache_respond(sc, sch, NULL, TH_SYN|TH_ACK); 499 TCPSTAT_INC(tcps_sc_retransmitted); 500 syncache_timeout(sc, sch, 0); 501 } 502 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 503 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 504 syncache_timer, (void *)(sch)); 505 CURVNET_RESTORE(); 506 } 507 508 /* 509 * Find an entry in the syncache. 510 * Returns always with locked syncache_head plus a matching entry or NULL. 511 */ 512 static struct syncache * 513 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 514 { 515 struct syncache *sc; 516 struct syncache_head *sch; 517 uint32_t hash; 518 519 /* 520 * The hash is built on foreign port + local port + foreign address. 521 * We rely on the fact that struct in_conninfo starts with 16 bits 522 * of foreign port, then 16 bits of local port then followed by 128 523 * bits of foreign address. In case of IPv4 address, the first 3 524 * 32-bit words of the address always are zeroes. 525 */ 526 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 527 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 528 529 sch = &V_tcp_syncache.hashbase[hash]; 530 *schp = sch; 531 SCH_LOCK(sch); 532 533 /* Circle through bucket row to find matching entry. */ 534 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 535 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 536 sizeof(struct in_endpoints)) == 0) 537 break; 538 539 return (sc); /* Always returns with locked sch. */ 540 } 541 542 /* 543 * This function is called when we get a RST for a 544 * non-existent connection, so that we can see if the 545 * connection is in the syn cache. If it is, zap it. 546 * If required send a challenge ACK. 547 */ 548 void 549 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m) 550 { 551 struct syncache *sc; 552 struct syncache_head *sch; 553 char *s = NULL; 554 555 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 556 SCH_LOCK_ASSERT(sch); 557 558 /* 559 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 560 * See RFC 793 page 65, section SEGMENT ARRIVES. 561 */ 562 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 563 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 564 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 565 "FIN flag set, segment ignored\n", s, __func__); 566 TCPSTAT_INC(tcps_badrst); 567 goto done; 568 } 569 570 /* 571 * No corresponding connection was found in syncache. 572 * If syncookies are enabled and possibly exclusively 573 * used, or we are under memory pressure, a valid RST 574 * may not find a syncache entry. In that case we're 575 * done and no SYN|ACK retransmissions will happen. 576 * Otherwise the RST was misdirected or spoofed. 577 */ 578 if (sc == NULL) { 579 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 580 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 581 "syncache entry (possibly syncookie only), " 582 "segment ignored\n", s, __func__); 583 TCPSTAT_INC(tcps_badrst); 584 goto done; 585 } 586 587 /* 588 * If the RST bit is set, check the sequence number to see 589 * if this is a valid reset segment. 590 * 591 * RFC 793 page 37: 592 * In all states except SYN-SENT, all reset (RST) segments 593 * are validated by checking their SEQ-fields. A reset is 594 * valid if its sequence number is in the window. 595 * 596 * RFC 793 page 69: 597 * There are four cases for the acceptability test for an incoming 598 * segment: 599 * 600 * Segment Receive Test 601 * Length Window 602 * ------- ------- ------------------------------------------- 603 * 0 0 SEG.SEQ = RCV.NXT 604 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 605 * >0 0 not acceptable 606 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 607 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 608 * 609 * Note that when receiving a SYN segment in the LISTEN state, 610 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 611 * described in RFC 793, page 66. 612 */ 613 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 614 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 615 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 616 if (V_tcp_insecure_rst || 617 th->th_seq == sc->sc_irs + 1) { 618 syncache_drop(sc, sch); 619 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 620 log(LOG_DEBUG, 621 "%s; %s: Our SYN|ACK was rejected, " 622 "connection attempt aborted by remote " 623 "endpoint\n", 624 s, __func__); 625 TCPSTAT_INC(tcps_sc_reset); 626 } else { 627 TCPSTAT_INC(tcps_badrst); 628 /* Send challenge ACK. */ 629 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 630 log(LOG_DEBUG, "%s; %s: RST with invalid " 631 " SEQ %u != NXT %u (+WND %u), " 632 "sending challenge ACK\n", 633 s, __func__, 634 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 635 syncache_respond(sc, sch, m, TH_ACK); 636 } 637 } else { 638 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 639 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 640 "NXT %u (+WND %u), segment ignored\n", 641 s, __func__, 642 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 643 TCPSTAT_INC(tcps_badrst); 644 } 645 646 done: 647 if (s != NULL) 648 free(s, M_TCPLOG); 649 SCH_UNLOCK(sch); 650 } 651 652 void 653 syncache_badack(struct in_conninfo *inc) 654 { 655 struct syncache *sc; 656 struct syncache_head *sch; 657 658 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 659 SCH_LOCK_ASSERT(sch); 660 if (sc != NULL) { 661 syncache_drop(sc, sch); 662 TCPSTAT_INC(tcps_sc_badack); 663 } 664 SCH_UNLOCK(sch); 665 } 666 667 void 668 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq) 669 { 670 struct syncache *sc; 671 struct syncache_head *sch; 672 673 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 674 SCH_LOCK_ASSERT(sch); 675 if (sc == NULL) 676 goto done; 677 678 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 679 if (ntohl(th_seq) != sc->sc_iss) 680 goto done; 681 682 /* 683 * If we've rertransmitted 3 times and this is our second error, 684 * we remove the entry. Otherwise, we allow it to continue on. 685 * This prevents us from incorrectly nuking an entry during a 686 * spurious network outage. 687 * 688 * See tcp_notify(). 689 */ 690 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 691 sc->sc_flags |= SCF_UNREACH; 692 goto done; 693 } 694 syncache_drop(sc, sch); 695 TCPSTAT_INC(tcps_sc_unreach); 696 done: 697 SCH_UNLOCK(sch); 698 } 699 700 /* 701 * Build a new TCP socket structure from a syncache entry. 702 * 703 * On success return the newly created socket with its underlying inp locked. 704 */ 705 static struct socket * 706 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 707 { 708 struct tcp_function_block *blk; 709 struct inpcb *inp = NULL; 710 struct socket *so; 711 struct tcpcb *tp; 712 int error; 713 char *s; 714 715 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 716 717 /* 718 * Ok, create the full blown connection, and set things up 719 * as they would have been set up if we had created the 720 * connection when the SYN arrived. If we can't create 721 * the connection, abort it. 722 */ 723 so = sonewconn(lso, 0); 724 if (so == NULL) { 725 /* 726 * Drop the connection; we will either send a RST or 727 * have the peer retransmit its SYN again after its 728 * RTO and try again. 729 */ 730 TCPSTAT_INC(tcps_listendrop); 731 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 732 log(LOG_DEBUG, "%s; %s: Socket create failed " 733 "due to limits or memory shortage\n", 734 s, __func__); 735 free(s, M_TCPLOG); 736 } 737 goto abort2; 738 } 739 #ifdef MAC 740 mac_socketpeer_set_from_mbuf(m, so); 741 #endif 742 743 inp = sotoinpcb(so); 744 inp->inp_inc.inc_fibnum = so->so_fibnum; 745 INP_WLOCK(inp); 746 /* 747 * Exclusive pcbinfo lock is not required in syncache socket case even 748 * if two inpcb locks can be acquired simultaneously: 749 * - the inpcb in LISTEN state, 750 * - the newly created inp. 751 * 752 * In this case, an inp cannot be at same time in LISTEN state and 753 * just created by an accept() call. 754 */ 755 INP_HASH_WLOCK(&V_tcbinfo); 756 757 /* Insert new socket into PCB hash list. */ 758 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 759 #ifdef INET6 760 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 761 inp->inp_vflag &= ~INP_IPV4; 762 inp->inp_vflag |= INP_IPV6; 763 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 764 } else { 765 inp->inp_vflag &= ~INP_IPV6; 766 inp->inp_vflag |= INP_IPV4; 767 #endif 768 inp->inp_laddr = sc->sc_inc.inc_laddr; 769 #ifdef INET6 770 } 771 #endif 772 773 /* 774 * If there's an mbuf and it has a flowid, then let's initialise the 775 * inp with that particular flowid. 776 */ 777 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 778 inp->inp_flowid = m->m_pkthdr.flowid; 779 inp->inp_flowtype = M_HASHTYPE_GET(m); 780 #ifdef NUMA 781 inp->inp_numa_domain = m->m_pkthdr.numa_domain; 782 #endif 783 } 784 785 /* 786 * Install in the reservation hash table for now, but don't yet 787 * install a connection group since the full 4-tuple isn't yet 788 * configured. 789 */ 790 inp->inp_lport = sc->sc_inc.inc_lport; 791 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 792 /* 793 * Undo the assignments above if we failed to 794 * put the PCB on the hash lists. 795 */ 796 #ifdef INET6 797 if (sc->sc_inc.inc_flags & INC_ISIPV6) 798 inp->in6p_laddr = in6addr_any; 799 else 800 #endif 801 inp->inp_laddr.s_addr = INADDR_ANY; 802 inp->inp_lport = 0; 803 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 804 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 805 "with error %i\n", 806 s, __func__, error); 807 free(s, M_TCPLOG); 808 } 809 INP_HASH_WUNLOCK(&V_tcbinfo); 810 goto abort; 811 } 812 #ifdef INET6 813 if (inp->inp_vflag & INP_IPV6PROTO) { 814 struct inpcb *oinp = sotoinpcb(lso); 815 816 /* 817 * Inherit socket options from the listening socket. 818 * Note that in6p_inputopts are not (and should not be) 819 * copied, since it stores previously received options and is 820 * used to detect if each new option is different than the 821 * previous one and hence should be passed to a user. 822 * If we copied in6p_inputopts, a user would not be able to 823 * receive options just after calling the accept system call. 824 */ 825 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 826 if (oinp->in6p_outputopts) 827 inp->in6p_outputopts = 828 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 829 } 830 831 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 832 struct in6_addr laddr6; 833 struct sockaddr_in6 sin6; 834 835 sin6.sin6_family = AF_INET6; 836 sin6.sin6_len = sizeof(sin6); 837 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 838 sin6.sin6_port = sc->sc_inc.inc_fport; 839 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 840 laddr6 = inp->in6p_laddr; 841 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 842 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 843 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 844 thread0.td_ucred, m)) != 0) { 845 inp->in6p_laddr = laddr6; 846 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 847 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 848 "with error %i\n", 849 s, __func__, error); 850 free(s, M_TCPLOG); 851 } 852 INP_HASH_WUNLOCK(&V_tcbinfo); 853 goto abort; 854 } 855 /* Override flowlabel from in6_pcbconnect. */ 856 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 857 inp->inp_flow |= sc->sc_flowlabel; 858 } 859 #endif /* INET6 */ 860 #if defined(INET) && defined(INET6) 861 else 862 #endif 863 #ifdef INET 864 { 865 struct in_addr laddr; 866 struct sockaddr_in sin; 867 868 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 869 870 if (inp->inp_options == NULL) { 871 inp->inp_options = sc->sc_ipopts; 872 sc->sc_ipopts = NULL; 873 } 874 875 sin.sin_family = AF_INET; 876 sin.sin_len = sizeof(sin); 877 sin.sin_addr = sc->sc_inc.inc_faddr; 878 sin.sin_port = sc->sc_inc.inc_fport; 879 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 880 laddr = inp->inp_laddr; 881 if (inp->inp_laddr.s_addr == INADDR_ANY) 882 inp->inp_laddr = sc->sc_inc.inc_laddr; 883 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 884 thread0.td_ucred, m)) != 0) { 885 inp->inp_laddr = laddr; 886 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 887 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 888 "with error %i\n", 889 s, __func__, error); 890 free(s, M_TCPLOG); 891 } 892 INP_HASH_WUNLOCK(&V_tcbinfo); 893 goto abort; 894 } 895 } 896 #endif /* INET */ 897 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 898 /* Copy old policy into new socket's. */ 899 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 900 printf("syncache_socket: could not copy policy\n"); 901 #endif 902 INP_HASH_WUNLOCK(&V_tcbinfo); 903 tp = intotcpcb(inp); 904 tcp_state_change(tp, TCPS_SYN_RECEIVED); 905 tp->iss = sc->sc_iss; 906 tp->irs = sc->sc_irs; 907 tcp_rcvseqinit(tp); 908 tcp_sendseqinit(tp); 909 blk = sototcpcb(lso)->t_fb; 910 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 911 /* 912 * Our parents t_fb was not the default, 913 * we need to release our ref on tp->t_fb and 914 * pickup one on the new entry. 915 */ 916 struct tcp_function_block *rblk; 917 918 rblk = find_and_ref_tcp_fb(blk); 919 KASSERT(rblk != NULL, 920 ("cannot find blk %p out of syncache?", blk)); 921 if (tp->t_fb->tfb_tcp_fb_fini) 922 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 923 refcount_release(&tp->t_fb->tfb_refcnt); 924 tp->t_fb = rblk; 925 /* 926 * XXXrrs this is quite dangerous, it is possible 927 * for the new function to fail to init. We also 928 * are not asking if the handoff_is_ok though at 929 * the very start thats probalbly ok. 930 */ 931 if (tp->t_fb->tfb_tcp_fb_init) { 932 (*tp->t_fb->tfb_tcp_fb_init)(tp); 933 } 934 } 935 tp->snd_wl1 = sc->sc_irs; 936 tp->snd_max = tp->iss + 1; 937 tp->snd_nxt = tp->iss + 1; 938 tp->rcv_up = sc->sc_irs + 1; 939 tp->rcv_wnd = sc->sc_wnd; 940 tp->rcv_adv += tp->rcv_wnd; 941 tp->last_ack_sent = tp->rcv_nxt; 942 943 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 944 if (sc->sc_flags & SCF_NOOPT) 945 tp->t_flags |= TF_NOOPT; 946 else { 947 if (sc->sc_flags & SCF_WINSCALE) { 948 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 949 tp->snd_scale = sc->sc_requested_s_scale; 950 tp->request_r_scale = sc->sc_requested_r_scale; 951 } 952 if (sc->sc_flags & SCF_TIMESTAMP) { 953 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 954 tp->ts_recent = sc->sc_tsreflect; 955 tp->ts_recent_age = tcp_ts_getticks(); 956 tp->ts_offset = sc->sc_tsoff; 957 } 958 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 959 if (sc->sc_flags & SCF_SIGNATURE) 960 tp->t_flags |= TF_SIGNATURE; 961 #endif 962 if (sc->sc_flags & SCF_SACK) 963 tp->t_flags |= TF_SACK_PERMIT; 964 } 965 966 if (sc->sc_flags & SCF_ECN) 967 tp->t_flags |= TF_ECN_PERMIT; 968 969 /* 970 * Set up MSS and get cached values from tcp_hostcache. 971 * This might overwrite some of the defaults we just set. 972 */ 973 tcp_mss(tp, sc->sc_peer_mss); 974 975 /* 976 * If the SYN,ACK was retransmitted, indicate that CWND to be 977 * limited to one segment in cc_conn_init(). 978 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 979 */ 980 if (sc->sc_rxmits > 1) 981 tp->snd_cwnd = 1; 982 983 #ifdef TCP_OFFLOAD 984 /* 985 * Allow a TOE driver to install its hooks. Note that we hold the 986 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 987 * new connection before the TOE driver has done its thing. 988 */ 989 if (ADDED_BY_TOE(sc)) { 990 struct toedev *tod = sc->sc_tod; 991 992 tod->tod_offload_socket(tod, sc->sc_todctx, so); 993 } 994 #endif 995 /* 996 * Copy and activate timers. 997 */ 998 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 999 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 1000 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1001 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1002 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1003 1004 TCPSTAT_INC(tcps_accepts); 1005 return (so); 1006 1007 abort: 1008 INP_WUNLOCK(inp); 1009 abort2: 1010 if (so != NULL) 1011 soabort(so); 1012 return (NULL); 1013 } 1014 1015 /* 1016 * This function gets called when we receive an ACK for a 1017 * socket in the LISTEN state. We look up the connection 1018 * in the syncache, and if its there, we pull it out of 1019 * the cache and turn it into a full-blown connection in 1020 * the SYN-RECEIVED state. 1021 * 1022 * On syncache_socket() success the newly created socket 1023 * has its underlying inp locked. 1024 */ 1025 int 1026 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1027 struct socket **lsop, struct mbuf *m) 1028 { 1029 struct syncache *sc; 1030 struct syncache_head *sch; 1031 struct syncache scs; 1032 char *s; 1033 1034 /* 1035 * Global TCP locks are held because we manipulate the PCB lists 1036 * and create a new socket. 1037 */ 1038 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1039 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1040 ("%s: can handle only ACK", __func__)); 1041 1042 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1043 SCH_LOCK_ASSERT(sch); 1044 1045 #ifdef INVARIANTS 1046 /* 1047 * Test code for syncookies comparing the syncache stored 1048 * values with the reconstructed values from the cookie. 1049 */ 1050 if (sc != NULL) 1051 syncookie_cmp(inc, sch, sc, th, to, *lsop); 1052 #endif 1053 1054 if (sc == NULL) { 1055 /* 1056 * There is no syncache entry, so see if this ACK is 1057 * a returning syncookie. To do this, first: 1058 * A. Check if syncookies are used in case of syncache 1059 * overflows 1060 * B. See if this socket has had a syncache entry dropped in 1061 * the recent past. We don't want to accept a bogus 1062 * syncookie if we've never received a SYN or accept it 1063 * twice. 1064 * C. check that the syncookie is valid. If it is, then 1065 * cobble up a fake syncache entry, and return. 1066 */ 1067 if (!V_tcp_syncookies) { 1068 SCH_UNLOCK(sch); 1069 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1070 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1071 "segment rejected (syncookies disabled)\n", 1072 s, __func__); 1073 goto failed; 1074 } 1075 if (!V_tcp_syncookiesonly && 1076 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1077 SCH_UNLOCK(sch); 1078 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1079 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1080 "segment rejected (no syncache entry)\n", 1081 s, __func__); 1082 goto failed; 1083 } 1084 bzero(&scs, sizeof(scs)); 1085 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1086 SCH_UNLOCK(sch); 1087 if (sc == NULL) { 1088 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1089 log(LOG_DEBUG, "%s; %s: Segment failed " 1090 "SYNCOOKIE authentication, segment rejected " 1091 "(probably spoofed)\n", s, __func__); 1092 goto failed; 1093 } 1094 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1095 /* If received ACK has MD5 signature, check it. */ 1096 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1097 (!TCPMD5_ENABLED() || 1098 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1099 /* Drop the ACK. */ 1100 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1101 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1102 "MD5 signature doesn't match.\n", 1103 s, __func__); 1104 free(s, M_TCPLOG); 1105 } 1106 TCPSTAT_INC(tcps_sig_err_sigopt); 1107 return (-1); /* Do not send RST */ 1108 } 1109 #endif /* TCP_SIGNATURE */ 1110 } else { 1111 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1112 /* 1113 * If listening socket requested TCP digests, check that 1114 * received ACK has signature and it is correct. 1115 * If not, drop the ACK and leave sc entry in th cache, 1116 * because SYN was received with correct signature. 1117 */ 1118 if (sc->sc_flags & SCF_SIGNATURE) { 1119 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1120 /* No signature */ 1121 TCPSTAT_INC(tcps_sig_err_nosigopt); 1122 SCH_UNLOCK(sch); 1123 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1124 log(LOG_DEBUG, "%s; %s: Segment " 1125 "rejected, MD5 signature wasn't " 1126 "provided.\n", s, __func__); 1127 free(s, M_TCPLOG); 1128 } 1129 return (-1); /* Do not send RST */ 1130 } 1131 if (!TCPMD5_ENABLED() || 1132 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1133 /* Doesn't match or no SA */ 1134 SCH_UNLOCK(sch); 1135 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1136 log(LOG_DEBUG, "%s; %s: Segment " 1137 "rejected, MD5 signature doesn't " 1138 "match.\n", s, __func__); 1139 free(s, M_TCPLOG); 1140 } 1141 return (-1); /* Do not send RST */ 1142 } 1143 } 1144 #endif /* TCP_SIGNATURE */ 1145 1146 /* 1147 * RFC 7323 PAWS: If we have a timestamp on this segment and 1148 * it's less than ts_recent, drop it. 1149 * XXXMT: RFC 7323 also requires to send an ACK. 1150 * In tcp_input.c this is only done for TCP segments 1151 * with user data, so be consistent here and just drop 1152 * the segment. 1153 */ 1154 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1155 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1156 SCH_UNLOCK(sch); 1157 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1158 log(LOG_DEBUG, 1159 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1160 "segment dropped\n", s, __func__, 1161 to->to_tsval, sc->sc_tsreflect); 1162 free(s, M_TCPLOG); 1163 } 1164 return (-1); /* Do not send RST */ 1165 } 1166 1167 /* 1168 * Pull out the entry to unlock the bucket row. 1169 * 1170 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1171 * tcp_state_change(). The tcpcb is not existent at this 1172 * moment. A new one will be allocated via syncache_socket-> 1173 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1174 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1175 */ 1176 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1177 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1178 sch->sch_length--; 1179 #ifdef TCP_OFFLOAD 1180 if (ADDED_BY_TOE(sc)) { 1181 struct toedev *tod = sc->sc_tod; 1182 1183 tod->tod_syncache_removed(tod, sc->sc_todctx); 1184 } 1185 #endif 1186 SCH_UNLOCK(sch); 1187 } 1188 1189 /* 1190 * Segment validation: 1191 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1192 */ 1193 if (th->th_ack != sc->sc_iss + 1) { 1194 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1195 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1196 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1197 goto failed; 1198 } 1199 1200 /* 1201 * The SEQ must fall in the window starting at the received 1202 * initial receive sequence number + 1 (the SYN). 1203 */ 1204 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1205 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1206 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1207 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1208 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1209 goto failed; 1210 } 1211 1212 /* 1213 * If timestamps were not negotiated during SYN/ACK they 1214 * must not appear on any segment during this session. 1215 */ 1216 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1217 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1218 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1219 "segment rejected\n", s, __func__); 1220 goto failed; 1221 } 1222 1223 /* 1224 * If timestamps were negotiated during SYN/ACK they should 1225 * appear on every segment during this session. 1226 * XXXAO: This is only informal as there have been unverified 1227 * reports of non-compliants stacks. 1228 */ 1229 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1230 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1231 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1232 "no action\n", s, __func__); 1233 free(s, M_TCPLOG); 1234 s = NULL; 1235 } 1236 } 1237 1238 *lsop = syncache_socket(sc, *lsop, m); 1239 1240 if (*lsop == NULL) 1241 TCPSTAT_INC(tcps_sc_aborted); 1242 else 1243 TCPSTAT_INC(tcps_sc_completed); 1244 1245 /* how do we find the inp for the new socket? */ 1246 if (sc != &scs) 1247 syncache_free(sc); 1248 return (1); 1249 failed: 1250 if (sc != NULL && sc != &scs) 1251 syncache_free(sc); 1252 if (s != NULL) 1253 free(s, M_TCPLOG); 1254 *lsop = NULL; 1255 return (0); 1256 } 1257 1258 static void 1259 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1260 uint64_t response_cookie) 1261 { 1262 struct inpcb *inp; 1263 struct tcpcb *tp; 1264 unsigned int *pending_counter; 1265 1266 /* 1267 * Global TCP locks are held because we manipulate the PCB lists 1268 * and create a new socket. 1269 */ 1270 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1271 1272 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1273 *lsop = syncache_socket(sc, *lsop, m); 1274 if (*lsop == NULL) { 1275 TCPSTAT_INC(tcps_sc_aborted); 1276 atomic_subtract_int(pending_counter, 1); 1277 } else { 1278 soisconnected(*lsop); 1279 inp = sotoinpcb(*lsop); 1280 tp = intotcpcb(inp); 1281 tp->t_flags |= TF_FASTOPEN; 1282 tp->t_tfo_cookie.server = response_cookie; 1283 tp->snd_max = tp->iss; 1284 tp->snd_nxt = tp->iss; 1285 tp->t_tfo_pending = pending_counter; 1286 TCPSTAT_INC(tcps_sc_completed); 1287 } 1288 } 1289 1290 /* 1291 * Given a LISTEN socket and an inbound SYN request, add 1292 * this to the syn cache, and send back a segment: 1293 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1294 * to the source. 1295 * 1296 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1297 * Doing so would require that we hold onto the data and deliver it 1298 * to the application. However, if we are the target of a SYN-flood 1299 * DoS attack, an attacker could send data which would eventually 1300 * consume all available buffer space if it were ACKed. By not ACKing 1301 * the data, we avoid this DoS scenario. 1302 * 1303 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1304 * cookie is processed and a new socket is created. In this case, any data 1305 * accompanying the SYN will be queued to the socket by tcp_input() and will 1306 * be ACKed either when the application sends response data or the delayed 1307 * ACK timer expires, whichever comes first. 1308 */ 1309 int 1310 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1311 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1312 void *todctx) 1313 { 1314 struct tcpcb *tp; 1315 struct socket *so; 1316 struct syncache *sc = NULL; 1317 struct syncache_head *sch; 1318 struct mbuf *ipopts = NULL; 1319 u_int ltflags; 1320 int win, ip_ttl, ip_tos; 1321 char *s; 1322 int rv = 0; 1323 #ifdef INET6 1324 int autoflowlabel = 0; 1325 #endif 1326 #ifdef MAC 1327 struct label *maclabel; 1328 #endif 1329 struct syncache scs; 1330 struct ucred *cred; 1331 uint64_t tfo_response_cookie; 1332 unsigned int *tfo_pending = NULL; 1333 int tfo_cookie_valid = 0; 1334 int tfo_response_cookie_valid = 0; 1335 1336 INP_WLOCK_ASSERT(inp); /* listen socket */ 1337 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1338 ("%s: unexpected tcp flags", __func__)); 1339 1340 /* 1341 * Combine all so/tp operations very early to drop the INP lock as 1342 * soon as possible. 1343 */ 1344 so = *lsop; 1345 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1346 tp = sototcpcb(so); 1347 cred = crhold(so->so_cred); 1348 1349 #ifdef INET6 1350 if ((inc->inc_flags & INC_ISIPV6) && 1351 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1352 autoflowlabel = 1; 1353 #endif 1354 ip_ttl = inp->inp_ip_ttl; 1355 ip_tos = inp->inp_ip_tos; 1356 win = so->sol_sbrcv_hiwat; 1357 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1358 1359 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1360 (tp->t_tfo_pending != NULL) && 1361 (to->to_flags & TOF_FASTOPEN)) { 1362 /* 1363 * Limit the number of pending TFO connections to 1364 * approximately half of the queue limit. This prevents TFO 1365 * SYN floods from starving the service by filling the 1366 * listen queue with bogus TFO connections. 1367 */ 1368 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1369 (so->sol_qlimit / 2)) { 1370 int result; 1371 1372 result = tcp_fastopen_check_cookie(inc, 1373 to->to_tfo_cookie, to->to_tfo_len, 1374 &tfo_response_cookie); 1375 tfo_cookie_valid = (result > 0); 1376 tfo_response_cookie_valid = (result >= 0); 1377 } 1378 1379 /* 1380 * Remember the TFO pending counter as it will have to be 1381 * decremented below if we don't make it to syncache_tfo_expand(). 1382 */ 1383 tfo_pending = tp->t_tfo_pending; 1384 } 1385 1386 /* By the time we drop the lock these should no longer be used. */ 1387 so = NULL; 1388 tp = NULL; 1389 1390 #ifdef MAC 1391 if (mac_syncache_init(&maclabel) != 0) { 1392 INP_WUNLOCK(inp); 1393 goto done; 1394 } else 1395 mac_syncache_create(maclabel, inp); 1396 #endif 1397 if (!tfo_cookie_valid) 1398 INP_WUNLOCK(inp); 1399 1400 /* 1401 * Remember the IP options, if any. 1402 */ 1403 #ifdef INET6 1404 if (!(inc->inc_flags & INC_ISIPV6)) 1405 #endif 1406 #ifdef INET 1407 ipopts = (m) ? ip_srcroute(m) : NULL; 1408 #else 1409 ipopts = NULL; 1410 #endif 1411 1412 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1413 /* 1414 * If listening socket requested TCP digests, check that received 1415 * SYN has signature and it is correct. If signature doesn't match 1416 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1417 */ 1418 if (ltflags & TF_SIGNATURE) { 1419 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1420 TCPSTAT_INC(tcps_sig_err_nosigopt); 1421 goto done; 1422 } 1423 if (!TCPMD5_ENABLED() || 1424 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1425 goto done; 1426 } 1427 #endif /* TCP_SIGNATURE */ 1428 /* 1429 * See if we already have an entry for this connection. 1430 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1431 * 1432 * XXX: should the syncache be re-initialized with the contents 1433 * of the new SYN here (which may have different options?) 1434 * 1435 * XXX: We do not check the sequence number to see if this is a 1436 * real retransmit or a new connection attempt. The question is 1437 * how to handle such a case; either ignore it as spoofed, or 1438 * drop the current entry and create a new one? 1439 */ 1440 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1441 SCH_LOCK_ASSERT(sch); 1442 if (sc != NULL) { 1443 if (tfo_cookie_valid) 1444 INP_WUNLOCK(inp); 1445 TCPSTAT_INC(tcps_sc_dupsyn); 1446 if (ipopts) { 1447 /* 1448 * If we were remembering a previous source route, 1449 * forget it and use the new one we've been given. 1450 */ 1451 if (sc->sc_ipopts) 1452 (void) m_free(sc->sc_ipopts); 1453 sc->sc_ipopts = ipopts; 1454 } 1455 /* 1456 * Update timestamp if present. 1457 */ 1458 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1459 sc->sc_tsreflect = to->to_tsval; 1460 else 1461 sc->sc_flags &= ~SCF_TIMESTAMP; 1462 #ifdef MAC 1463 /* 1464 * Since we have already unconditionally allocated label 1465 * storage, free it up. The syncache entry will already 1466 * have an initialized label we can use. 1467 */ 1468 mac_syncache_destroy(&maclabel); 1469 #endif 1470 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1471 /* Retransmit SYN|ACK and reset retransmit count. */ 1472 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1473 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1474 "resetting timer and retransmitting SYN|ACK\n", 1475 s, __func__); 1476 free(s, M_TCPLOG); 1477 } 1478 if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) { 1479 sc->sc_rxmits = 0; 1480 syncache_timeout(sc, sch, 1); 1481 TCPSTAT_INC(tcps_sndacks); 1482 TCPSTAT_INC(tcps_sndtotal); 1483 } 1484 SCH_UNLOCK(sch); 1485 goto donenoprobe; 1486 } 1487 1488 if (tfo_cookie_valid) { 1489 bzero(&scs, sizeof(scs)); 1490 sc = &scs; 1491 goto skip_alloc; 1492 } 1493 1494 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1495 if (sc == NULL) { 1496 /* 1497 * The zone allocator couldn't provide more entries. 1498 * Treat this as if the cache was full; drop the oldest 1499 * entry and insert the new one. 1500 */ 1501 TCPSTAT_INC(tcps_sc_zonefail); 1502 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1503 sch->sch_last_overflow = time_uptime; 1504 syncache_drop(sc, sch); 1505 } 1506 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1507 if (sc == NULL) { 1508 if (V_tcp_syncookies) { 1509 bzero(&scs, sizeof(scs)); 1510 sc = &scs; 1511 } else { 1512 SCH_UNLOCK(sch); 1513 if (ipopts) 1514 (void) m_free(ipopts); 1515 goto done; 1516 } 1517 } 1518 } 1519 1520 skip_alloc: 1521 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1522 sc->sc_tfo_cookie = &tfo_response_cookie; 1523 1524 /* 1525 * Fill in the syncache values. 1526 */ 1527 #ifdef MAC 1528 sc->sc_label = maclabel; 1529 #endif 1530 sc->sc_cred = cred; 1531 cred = NULL; 1532 sc->sc_ipopts = ipopts; 1533 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1534 #ifdef INET6 1535 if (!(inc->inc_flags & INC_ISIPV6)) 1536 #endif 1537 { 1538 sc->sc_ip_tos = ip_tos; 1539 sc->sc_ip_ttl = ip_ttl; 1540 } 1541 #ifdef TCP_OFFLOAD 1542 sc->sc_tod = tod; 1543 sc->sc_todctx = todctx; 1544 #endif 1545 sc->sc_irs = th->th_seq; 1546 sc->sc_iss = arc4random(); 1547 sc->sc_flags = 0; 1548 sc->sc_flowlabel = 0; 1549 1550 /* 1551 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1552 * win was derived from socket earlier in the function. 1553 */ 1554 win = imax(win, 0); 1555 win = imin(win, TCP_MAXWIN); 1556 sc->sc_wnd = win; 1557 1558 if (V_tcp_do_rfc1323) { 1559 /* 1560 * A timestamp received in a SYN makes 1561 * it ok to send timestamp requests and replies. 1562 */ 1563 if (to->to_flags & TOF_TS) { 1564 sc->sc_tsreflect = to->to_tsval; 1565 sc->sc_flags |= SCF_TIMESTAMP; 1566 sc->sc_tsoff = tcp_new_ts_offset(inc); 1567 } 1568 if (to->to_flags & TOF_SCALE) { 1569 int wscale = 0; 1570 1571 /* 1572 * Pick the smallest possible scaling factor that 1573 * will still allow us to scale up to sb_max, aka 1574 * kern.ipc.maxsockbuf. 1575 * 1576 * We do this because there are broken firewalls that 1577 * will corrupt the window scale option, leading to 1578 * the other endpoint believing that our advertised 1579 * window is unscaled. At scale factors larger than 1580 * 5 the unscaled window will drop below 1500 bytes, 1581 * leading to serious problems when traversing these 1582 * broken firewalls. 1583 * 1584 * With the default maxsockbuf of 256K, a scale factor 1585 * of 3 will be chosen by this algorithm. Those who 1586 * choose a larger maxsockbuf should watch out 1587 * for the compatibility problems mentioned above. 1588 * 1589 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1590 * or <SYN,ACK>) segment itself is never scaled. 1591 */ 1592 while (wscale < TCP_MAX_WINSHIFT && 1593 (TCP_MAXWIN << wscale) < sb_max) 1594 wscale++; 1595 sc->sc_requested_r_scale = wscale; 1596 sc->sc_requested_s_scale = to->to_wscale; 1597 sc->sc_flags |= SCF_WINSCALE; 1598 } 1599 } 1600 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1601 /* 1602 * If listening socket requested TCP digests, flag this in the 1603 * syncache so that syncache_respond() will do the right thing 1604 * with the SYN+ACK. 1605 */ 1606 if (ltflags & TF_SIGNATURE) 1607 sc->sc_flags |= SCF_SIGNATURE; 1608 #endif /* TCP_SIGNATURE */ 1609 if (to->to_flags & TOF_SACKPERM) 1610 sc->sc_flags |= SCF_SACK; 1611 if (to->to_flags & TOF_MSS) 1612 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1613 if (ltflags & TF_NOOPT) 1614 sc->sc_flags |= SCF_NOOPT; 1615 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1616 sc->sc_flags |= SCF_ECN; 1617 1618 if (V_tcp_syncookies) 1619 sc->sc_iss = syncookie_generate(sch, sc); 1620 #ifdef INET6 1621 if (autoflowlabel) { 1622 if (V_tcp_syncookies) 1623 sc->sc_flowlabel = sc->sc_iss; 1624 else 1625 sc->sc_flowlabel = ip6_randomflowlabel(); 1626 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1627 } 1628 #endif 1629 SCH_UNLOCK(sch); 1630 1631 if (tfo_cookie_valid) { 1632 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1633 /* INP_WUNLOCK(inp) will be performed by the caller */ 1634 rv = 1; 1635 goto tfo_expanded; 1636 } 1637 1638 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1639 /* 1640 * Do a standard 3-way handshake. 1641 */ 1642 if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) { 1643 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1644 syncache_free(sc); 1645 else if (sc != &scs) 1646 syncache_insert(sc, sch); /* locks and unlocks sch */ 1647 TCPSTAT_INC(tcps_sndacks); 1648 TCPSTAT_INC(tcps_sndtotal); 1649 } else { 1650 if (sc != &scs) 1651 syncache_free(sc); 1652 TCPSTAT_INC(tcps_sc_dropped); 1653 } 1654 goto donenoprobe; 1655 1656 done: 1657 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1658 donenoprobe: 1659 if (m) { 1660 *lsop = NULL; 1661 m_freem(m); 1662 } 1663 /* 1664 * If tfo_pending is not NULL here, then a TFO SYN that did not 1665 * result in a new socket was processed and the associated pending 1666 * counter has not yet been decremented. All such TFO processing paths 1667 * transit this point. 1668 */ 1669 if (tfo_pending != NULL) 1670 tcp_fastopen_decrement_counter(tfo_pending); 1671 1672 tfo_expanded: 1673 if (cred != NULL) 1674 crfree(cred); 1675 #ifdef MAC 1676 if (sc == &scs) 1677 mac_syncache_destroy(&maclabel); 1678 #endif 1679 return (rv); 1680 } 1681 1682 /* 1683 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1684 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1685 */ 1686 static int 1687 syncache_respond(struct syncache *sc, struct syncache_head *sch, 1688 const struct mbuf *m0, int flags) 1689 { 1690 struct ip *ip = NULL; 1691 struct mbuf *m; 1692 struct tcphdr *th = NULL; 1693 int optlen, error = 0; /* Make compiler happy */ 1694 u_int16_t hlen, tlen, mssopt; 1695 struct tcpopt to; 1696 #ifdef INET6 1697 struct ip6_hdr *ip6 = NULL; 1698 #endif 1699 hlen = 1700 #ifdef INET6 1701 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1702 #endif 1703 sizeof(struct ip); 1704 tlen = hlen + sizeof(struct tcphdr); 1705 1706 /* Determine MSS we advertize to other end of connection. */ 1707 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); 1708 1709 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1710 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1711 ("syncache: mbuf too small")); 1712 1713 /* Create the IP+TCP header from scratch. */ 1714 m = m_gethdr(M_NOWAIT, MT_DATA); 1715 if (m == NULL) 1716 return (ENOBUFS); 1717 #ifdef MAC 1718 mac_syncache_create_mbuf(sc->sc_label, m); 1719 #endif 1720 m->m_data += max_linkhdr; 1721 m->m_len = tlen; 1722 m->m_pkthdr.len = tlen; 1723 m->m_pkthdr.rcvif = NULL; 1724 1725 #ifdef INET6 1726 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1727 ip6 = mtod(m, struct ip6_hdr *); 1728 ip6->ip6_vfc = IPV6_VERSION; 1729 ip6->ip6_nxt = IPPROTO_TCP; 1730 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1731 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1732 ip6->ip6_plen = htons(tlen - hlen); 1733 /* ip6_hlim is set after checksum */ 1734 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1735 ip6->ip6_flow |= sc->sc_flowlabel; 1736 1737 th = (struct tcphdr *)(ip6 + 1); 1738 } 1739 #endif 1740 #if defined(INET6) && defined(INET) 1741 else 1742 #endif 1743 #ifdef INET 1744 { 1745 ip = mtod(m, struct ip *); 1746 ip->ip_v = IPVERSION; 1747 ip->ip_hl = sizeof(struct ip) >> 2; 1748 ip->ip_len = htons(tlen); 1749 ip->ip_id = 0; 1750 ip->ip_off = 0; 1751 ip->ip_sum = 0; 1752 ip->ip_p = IPPROTO_TCP; 1753 ip->ip_src = sc->sc_inc.inc_laddr; 1754 ip->ip_dst = sc->sc_inc.inc_faddr; 1755 ip->ip_ttl = sc->sc_ip_ttl; 1756 ip->ip_tos = sc->sc_ip_tos; 1757 1758 /* 1759 * See if we should do MTU discovery. Route lookups are 1760 * expensive, so we will only unset the DF bit if: 1761 * 1762 * 1) path_mtu_discovery is disabled 1763 * 2) the SCF_UNREACH flag has been set 1764 */ 1765 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1766 ip->ip_off |= htons(IP_DF); 1767 1768 th = (struct tcphdr *)(ip + 1); 1769 } 1770 #endif /* INET */ 1771 th->th_sport = sc->sc_inc.inc_lport; 1772 th->th_dport = sc->sc_inc.inc_fport; 1773 1774 if (flags & TH_SYN) 1775 th->th_seq = htonl(sc->sc_iss); 1776 else 1777 th->th_seq = htonl(sc->sc_iss + 1); 1778 th->th_ack = htonl(sc->sc_irs + 1); 1779 th->th_off = sizeof(struct tcphdr) >> 2; 1780 th->th_x2 = 0; 1781 th->th_flags = flags; 1782 th->th_win = htons(sc->sc_wnd); 1783 th->th_urp = 0; 1784 1785 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) { 1786 th->th_flags |= TH_ECE; 1787 TCPSTAT_INC(tcps_ecn_shs); 1788 } 1789 1790 /* Tack on the TCP options. */ 1791 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1792 to.to_flags = 0; 1793 1794 if (flags & TH_SYN) { 1795 to.to_mss = mssopt; 1796 to.to_flags = TOF_MSS; 1797 if (sc->sc_flags & SCF_WINSCALE) { 1798 to.to_wscale = sc->sc_requested_r_scale; 1799 to.to_flags |= TOF_SCALE; 1800 } 1801 if (sc->sc_flags & SCF_SACK) 1802 to.to_flags |= TOF_SACKPERM; 1803 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1804 if (sc->sc_flags & SCF_SIGNATURE) 1805 to.to_flags |= TOF_SIGNATURE; 1806 #endif 1807 if (sc->sc_tfo_cookie) { 1808 to.to_flags |= TOF_FASTOPEN; 1809 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1810 to.to_tfo_cookie = sc->sc_tfo_cookie; 1811 /* don't send cookie again when retransmitting response */ 1812 sc->sc_tfo_cookie = NULL; 1813 } 1814 } 1815 if (sc->sc_flags & SCF_TIMESTAMP) { 1816 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1817 to.to_tsecr = sc->sc_tsreflect; 1818 to.to_flags |= TOF_TS; 1819 } 1820 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1821 1822 /* Adjust headers by option size. */ 1823 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1824 m->m_len += optlen; 1825 m->m_pkthdr.len += optlen; 1826 #ifdef INET6 1827 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1828 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1829 else 1830 #endif 1831 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1832 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1833 if (sc->sc_flags & SCF_SIGNATURE) { 1834 KASSERT(to.to_flags & TOF_SIGNATURE, 1835 ("tcp_addoptions() didn't set tcp_signature")); 1836 1837 /* NOTE: to.to_signature is inside of mbuf */ 1838 if (!TCPMD5_ENABLED() || 1839 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1840 m_freem(m); 1841 return (EACCES); 1842 } 1843 } 1844 #endif 1845 } else 1846 optlen = 0; 1847 1848 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1849 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1850 /* 1851 * If we have peer's SYN and it has a flowid, then let's assign it to 1852 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1853 * to SYN|ACK due to lack of inp here. 1854 */ 1855 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1856 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1857 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1858 } 1859 #ifdef INET6 1860 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1861 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1862 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1863 IPPROTO_TCP, 0); 1864 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1865 #ifdef TCP_OFFLOAD 1866 if (ADDED_BY_TOE(sc)) { 1867 struct toedev *tod = sc->sc_tod; 1868 1869 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1870 1871 return (error); 1872 } 1873 #endif 1874 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 1875 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1876 } 1877 #endif 1878 #if defined(INET6) && defined(INET) 1879 else 1880 #endif 1881 #ifdef INET 1882 { 1883 m->m_pkthdr.csum_flags = CSUM_TCP; 1884 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1885 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1886 #ifdef TCP_OFFLOAD 1887 if (ADDED_BY_TOE(sc)) { 1888 struct toedev *tod = sc->sc_tod; 1889 1890 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1891 1892 return (error); 1893 } 1894 #endif 1895 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 1896 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1897 } 1898 #endif 1899 return (error); 1900 } 1901 1902 /* 1903 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1904 * that exceed the capacity of the syncache by avoiding the storage of any 1905 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1906 * attacks where the attacker does not have access to our responses. 1907 * 1908 * Syncookies encode and include all necessary information about the 1909 * connection setup within the SYN|ACK that we send back. That way we 1910 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1911 * (if ever). Normally the syncache and syncookies are running in parallel 1912 * with the latter taking over when the former is exhausted. When matching 1913 * syncache entry is found the syncookie is ignored. 1914 * 1915 * The only reliable information persisting the 3WHS is our initial sequence 1916 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1917 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1918 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1919 * returns and signifies a legitimate connection if it matches the ACK. 1920 * 1921 * The available space of 32 bits to store the hash and to encode the SYN 1922 * option information is very tight and we should have at least 24 bits for 1923 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1924 * 1925 * SYN option information we have to encode to fully restore a connection: 1926 * MSS: is imporant to chose an optimal segment size to avoid IP level 1927 * fragmentation along the path. The common MSS values can be encoded 1928 * in a 3-bit table. Uncommon values are captured by the next lower value 1929 * in the table leading to a slight increase in packetization overhead. 1930 * WSCALE: is necessary to allow large windows to be used for high delay- 1931 * bandwidth product links. Not scaling the window when it was initially 1932 * negotiated is bad for performance as lack of scaling further decreases 1933 * the apparent available send window. We only need to encode the WSCALE 1934 * we received from the remote end. Our end can be recalculated at any 1935 * time. The common WSCALE values can be encoded in a 3-bit table. 1936 * Uncommon values are captured by the next lower value in the table 1937 * making us under-estimate the available window size halving our 1938 * theoretically possible maximum throughput for that connection. 1939 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1940 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1941 * that are included in all segments on a connection. We enable them when 1942 * the ACK has them. 1943 * 1944 * Security of syncookies and attack vectors: 1945 * 1946 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1947 * together with the gloabl secret to make it unique per connection attempt. 1948 * Thus any change of any of those parameters results in a different MAC output 1949 * in an unpredictable way unless a collision is encountered. 24 bits of the 1950 * MAC are embedded into the ISS. 1951 * 1952 * To prevent replay attacks two rotating global secrets are updated with a 1953 * new random value every 15 seconds. The life-time of a syncookie is thus 1954 * 15-30 seconds. 1955 * 1956 * Vector 1: Attacking the secret. This requires finding a weakness in the 1957 * MAC itself or the way it is used here. The attacker can do a chosen plain 1958 * text attack by varying and testing the all parameters under his control. 1959 * The strength depends on the size and randomness of the secret, and the 1960 * cryptographic security of the MAC function. Due to the constant updating 1961 * of the secret the attacker has at most 29.999 seconds to find the secret 1962 * and launch spoofed connections. After that he has to start all over again. 1963 * 1964 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1965 * size an average of 4,823 attempts are required for a 50% chance of success 1966 * to spoof a single syncookie (birthday collision paradox). However the 1967 * attacker is blind and doesn't know if one of his attempts succeeded unless 1968 * he has a side channel to interfere success from. A single connection setup 1969 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1970 * This many attempts are required for each one blind spoofed connection. For 1971 * every additional spoofed connection he has to launch another N attempts. 1972 * Thus for a sustained rate 100 spoofed connections per second approximately 1973 * 1,800,000 packets per second would have to be sent. 1974 * 1975 * NB: The MAC function should be fast so that it doesn't become a CPU 1976 * exhaustion attack vector itself. 1977 * 1978 * References: 1979 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1980 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1981 * http://cr.yp.to/syncookies.html (overview) 1982 * http://cr.yp.to/syncookies/archive (details) 1983 * 1984 * 1985 * Schematic construction of a syncookie enabled Initial Sequence Number: 1986 * 0 1 2 3 1987 * 12345678901234567890123456789012 1988 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1989 * 1990 * x 24 MAC (truncated) 1991 * W 3 Send Window Scale index 1992 * M 3 MSS index 1993 * S 1 SACK permitted 1994 * P 1 Odd/even secret 1995 */ 1996 1997 /* 1998 * Distribution and probability of certain MSS values. Those in between are 1999 * rounded down to the next lower one. 2000 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2001 * .2% .3% 5% 7% 7% 20% 15% 45% 2002 */ 2003 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2004 2005 /* 2006 * Distribution and probability of certain WSCALE values. We have to map the 2007 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2008 * bits based on prevalence of certain values. Where we don't have an exact 2009 * match for are rounded down to the next lower one letting us under-estimate 2010 * the true available window. At the moment this would happen only for the 2011 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2012 * and window size). The absence of the WSCALE option (no scaling in either 2013 * direction) is encoded with index zero. 2014 * [WSCALE values histograms, Allman, 2012] 2015 * X 10 10 35 5 6 14 10% by host 2016 * X 11 4 5 5 18 49 3% by connections 2017 */ 2018 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2019 2020 /* 2021 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2022 * and good cryptographic properties. 2023 */ 2024 static uint32_t 2025 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2026 uint8_t *secbits, uintptr_t secmod) 2027 { 2028 SIPHASH_CTX ctx; 2029 uint32_t siphash[2]; 2030 2031 SipHash24_Init(&ctx); 2032 SipHash_SetKey(&ctx, secbits); 2033 switch (inc->inc_flags & INC_ISIPV6) { 2034 #ifdef INET 2035 case 0: 2036 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2037 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2038 break; 2039 #endif 2040 #ifdef INET6 2041 case INC_ISIPV6: 2042 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2043 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2044 break; 2045 #endif 2046 } 2047 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2048 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2049 SipHash_Update(&ctx, &irs, sizeof(irs)); 2050 SipHash_Update(&ctx, &flags, sizeof(flags)); 2051 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2052 SipHash_Final((u_int8_t *)&siphash, &ctx); 2053 2054 return (siphash[0] ^ siphash[1]); 2055 } 2056 2057 static tcp_seq 2058 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2059 { 2060 u_int i, secbit, wscale; 2061 uint32_t iss, hash; 2062 uint8_t *secbits; 2063 union syncookie cookie; 2064 2065 SCH_LOCK_ASSERT(sch); 2066 2067 cookie.cookie = 0; 2068 2069 /* Map our computed MSS into the 3-bit index. */ 2070 for (i = nitems(tcp_sc_msstab) - 1; 2071 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2072 i--) 2073 ; 2074 cookie.flags.mss_idx = i; 2075 2076 /* 2077 * Map the send window scale into the 3-bit index but only if 2078 * the wscale option was received. 2079 */ 2080 if (sc->sc_flags & SCF_WINSCALE) { 2081 wscale = sc->sc_requested_s_scale; 2082 for (i = nitems(tcp_sc_wstab) - 1; 2083 tcp_sc_wstab[i] > wscale && i > 0; 2084 i--) 2085 ; 2086 cookie.flags.wscale_idx = i; 2087 } 2088 2089 /* Can we do SACK? */ 2090 if (sc->sc_flags & SCF_SACK) 2091 cookie.flags.sack_ok = 1; 2092 2093 /* Which of the two secrets to use. */ 2094 secbit = sch->sch_sc->secret.oddeven & 0x1; 2095 cookie.flags.odd_even = secbit; 2096 2097 secbits = sch->sch_sc->secret.key[secbit]; 2098 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2099 (uintptr_t)sch); 2100 2101 /* 2102 * Put the flags into the hash and XOR them to get better ISS number 2103 * variance. This doesn't enhance the cryptographic strength and is 2104 * done to prevent the 8 cookie bits from showing up directly on the 2105 * wire. 2106 */ 2107 iss = hash & ~0xff; 2108 iss |= cookie.cookie ^ (hash >> 24); 2109 2110 TCPSTAT_INC(tcps_sc_sendcookie); 2111 return (iss); 2112 } 2113 2114 static struct syncache * 2115 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2116 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2117 struct socket *lso) 2118 { 2119 uint32_t hash; 2120 uint8_t *secbits; 2121 tcp_seq ack, seq; 2122 int wnd, wscale = 0; 2123 union syncookie cookie; 2124 2125 SCH_LOCK_ASSERT(sch); 2126 2127 /* 2128 * Pull information out of SYN-ACK/ACK and revert sequence number 2129 * advances. 2130 */ 2131 ack = th->th_ack - 1; 2132 seq = th->th_seq - 1; 2133 2134 /* 2135 * Unpack the flags containing enough information to restore the 2136 * connection. 2137 */ 2138 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2139 2140 /* Which of the two secrets to use. */ 2141 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 2142 2143 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2144 2145 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2146 if ((ack & ~0xff) != (hash & ~0xff)) 2147 return (NULL); 2148 2149 /* Fill in the syncache values. */ 2150 sc->sc_flags = 0; 2151 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2152 sc->sc_ipopts = NULL; 2153 2154 sc->sc_irs = seq; 2155 sc->sc_iss = ack; 2156 2157 switch (inc->inc_flags & INC_ISIPV6) { 2158 #ifdef INET 2159 case 0: 2160 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2161 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2162 break; 2163 #endif 2164 #ifdef INET6 2165 case INC_ISIPV6: 2166 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2167 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 2168 break; 2169 #endif 2170 } 2171 2172 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2173 2174 /* We can simply recompute receive window scale we sent earlier. */ 2175 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2176 wscale++; 2177 2178 /* Only use wscale if it was enabled in the orignal SYN. */ 2179 if (cookie.flags.wscale_idx > 0) { 2180 sc->sc_requested_r_scale = wscale; 2181 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2182 sc->sc_flags |= SCF_WINSCALE; 2183 } 2184 2185 wnd = lso->sol_sbrcv_hiwat; 2186 wnd = imax(wnd, 0); 2187 wnd = imin(wnd, TCP_MAXWIN); 2188 sc->sc_wnd = wnd; 2189 2190 if (cookie.flags.sack_ok) 2191 sc->sc_flags |= SCF_SACK; 2192 2193 if (to->to_flags & TOF_TS) { 2194 sc->sc_flags |= SCF_TIMESTAMP; 2195 sc->sc_tsreflect = to->to_tsval; 2196 sc->sc_tsoff = tcp_new_ts_offset(inc); 2197 } 2198 2199 if (to->to_flags & TOF_SIGNATURE) 2200 sc->sc_flags |= SCF_SIGNATURE; 2201 2202 sc->sc_rxmits = 0; 2203 2204 TCPSTAT_INC(tcps_sc_recvcookie); 2205 return (sc); 2206 } 2207 2208 #ifdef INVARIANTS 2209 static int 2210 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2211 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2212 struct socket *lso) 2213 { 2214 struct syncache scs, *scx; 2215 char *s; 2216 2217 bzero(&scs, sizeof(scs)); 2218 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2219 2220 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2221 return (0); 2222 2223 if (scx != NULL) { 2224 if (sc->sc_peer_mss != scx->sc_peer_mss) 2225 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2226 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2227 2228 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2229 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2230 s, __func__, sc->sc_requested_r_scale, 2231 scx->sc_requested_r_scale); 2232 2233 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2234 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2235 s, __func__, sc->sc_requested_s_scale, 2236 scx->sc_requested_s_scale); 2237 2238 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2239 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2240 } 2241 2242 if (s != NULL) 2243 free(s, M_TCPLOG); 2244 return (0); 2245 } 2246 #endif /* INVARIANTS */ 2247 2248 static void 2249 syncookie_reseed(void *arg) 2250 { 2251 struct tcp_syncache *sc = arg; 2252 uint8_t *secbits; 2253 int secbit; 2254 2255 /* 2256 * Reseeding the secret doesn't have to be protected by a lock. 2257 * It only must be ensured that the new random values are visible 2258 * to all CPUs in a SMP environment. The atomic with release 2259 * semantics ensures that. 2260 */ 2261 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2262 secbits = sc->secret.key[secbit]; 2263 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2264 atomic_add_rel_int(&sc->secret.oddeven, 1); 2265 2266 /* Reschedule ourself. */ 2267 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2268 } 2269 2270 /* 2271 * Exports the syncache entries to userland so that netstat can display 2272 * them alongside the other sockets. This function is intended to be 2273 * called only from tcp_pcblist. 2274 * 2275 * Due to concurrency on an active system, the number of pcbs exported 2276 * may have no relation to max_pcbs. max_pcbs merely indicates the 2277 * amount of space the caller allocated for this function to use. 2278 */ 2279 int 2280 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2281 { 2282 struct xtcpcb xt; 2283 struct syncache *sc; 2284 struct syncache_head *sch; 2285 int count, error, i; 2286 2287 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2288 sch = &V_tcp_syncache.hashbase[i]; 2289 SCH_LOCK(sch); 2290 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2291 if (count >= max_pcbs) { 2292 SCH_UNLOCK(sch); 2293 goto exit; 2294 } 2295 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2296 continue; 2297 bzero(&xt, sizeof(xt)); 2298 xt.xt_len = sizeof(xt); 2299 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2300 xt.xt_inp.inp_vflag = INP_IPV6; 2301 else 2302 xt.xt_inp.inp_vflag = INP_IPV4; 2303 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2304 sizeof (struct in_conninfo)); 2305 xt.t_state = TCPS_SYN_RECEIVED; 2306 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2307 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2308 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2309 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2310 error = SYSCTL_OUT(req, &xt, sizeof xt); 2311 if (error) { 2312 SCH_UNLOCK(sch); 2313 goto exit; 2314 } 2315 count++; 2316 } 2317 SCH_UNLOCK(sch); 2318 } 2319 exit: 2320 *pcbs_exported = count; 2321 return error; 2322 } 2323