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 106VNET_DEFINE_STATIC(int, tcp_syncookies) = 1; 107#define V_tcp_syncookies VNET(tcp_syncookies) 108SYSCTL_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 112VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0; 113#define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 114SYSCTL_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 118VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1; 119#define V_functions_inherit_listen_socket_stack \ 120 VNET(functions_inherit_listen_socket_stack) 121SYSCTL_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 130static void syncache_drop(struct syncache *, struct syncache_head *); 131static void syncache_free(struct syncache *); 132static void syncache_insert(struct syncache *, struct syncache_head *); 133static int syncache_respond(struct syncache *, struct syncache_head *, 134 const struct mbuf *, int); 135static struct socket *syncache_socket(struct syncache *, struct socket *, 136 struct mbuf *m); 137static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 138 int docallout); 139static void syncache_timer(void *); 140 141static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 142 uint8_t *, uintptr_t); 143static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 144static struct syncache 145 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 146 struct syncache *, struct tcphdr *, struct tcpopt *, 147 struct socket *); 148static void syncookie_reseed(void *); 149#ifdef INVARIANTS 150static 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 171VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 172#define V_tcp_syncache VNET(tcp_syncache) 173 174static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, 175 "TCP SYN cache"); 176 177SYSCTL_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 181SYSCTL_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 185SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 186 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 187 188SYSCTL_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 192static int 193sysctl_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 209SYSCTL_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 215VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 216SYSCTL_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 220static 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 */ 229static void 230syncache_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 244void 245syncache_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 307void 308syncache_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 */ 350static void 351syncache_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 */ 397static void 398syncache_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 */ 421static void 422syncache_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 */ 447static void 448syncache_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_ecn_maxretries) { 481 sc->sc_flags &= ~SCF_ECN; 482 } 483 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 484 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 485 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 486 "giving up and removing syncache entry\n", 487 s, __func__); 488 free(s, M_TCPLOG); 489 } 490 syncache_drop(sc, sch); 491 TCPSTAT_INC(tcps_sc_stale); 492 continue; 493 } 494 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 495 log(LOG_DEBUG, "%s; %s: Response timeout, " 496 "retransmitting (%u) SYN|ACK\n", 497 s, __func__, sc->sc_rxmits); 498 free(s, M_TCPLOG); 499 } 500 501 syncache_respond(sc, sch, NULL, TH_SYN|TH_ACK); 502 TCPSTAT_INC(tcps_sc_retransmitted); 503 syncache_timeout(sc, sch, 0); 504 } 505 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 506 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 507 syncache_timer, (void *)(sch)); 508 CURVNET_RESTORE(); 509} 510 511/* 512 * Find an entry in the syncache. 513 * Returns always with locked syncache_head plus a matching entry or NULL. 514 */ 515static struct syncache * 516syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 517{ 518 struct syncache *sc; 519 struct syncache_head *sch; 520 uint32_t hash; 521 522 /* 523 * The hash is built on foreign port + local port + foreign address. 524 * We rely on the fact that struct in_conninfo starts with 16 bits 525 * of foreign port, then 16 bits of local port then followed by 128 526 * bits of foreign address. In case of IPv4 address, the first 3 527 * 32-bit words of the address always are zeroes. 528 */ 529 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 530 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 531 532 sch = &V_tcp_syncache.hashbase[hash]; 533 *schp = sch; 534 SCH_LOCK(sch); 535 536 /* Circle through bucket row to find matching entry. */ 537 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 538 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 539 sizeof(struct in_endpoints)) == 0) 540 break; 541 542 return (sc); /* Always returns with locked sch. */ 543} 544 545/* 546 * This function is called when we get a RST for a 547 * non-existent connection, so that we can see if the 548 * connection is in the syn cache. If it is, zap it. 549 * If required send a challenge ACK. 550 */ 551void 552syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m) 553{ 554 struct syncache *sc; 555 struct syncache_head *sch; 556 char *s = NULL; 557 558 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 559 SCH_LOCK_ASSERT(sch); 560 561 /* 562 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 563 * See RFC 793 page 65, section SEGMENT ARRIVES. 564 */ 565 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 566 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 567 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 568 "FIN flag set, segment ignored\n", s, __func__); 569 TCPSTAT_INC(tcps_badrst); 570 goto done; 571 } 572 573 /* 574 * No corresponding connection was found in syncache. 575 * If syncookies are enabled and possibly exclusively 576 * used, or we are under memory pressure, a valid RST 577 * may not find a syncache entry. In that case we're 578 * done and no SYN|ACK retransmissions will happen. 579 * Otherwise the RST was misdirected or spoofed. 580 */ 581 if (sc == NULL) { 582 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 583 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 584 "syncache entry (possibly syncookie only), " 585 "segment ignored\n", s, __func__); 586 TCPSTAT_INC(tcps_badrst); 587 goto done; 588 } 589 590 /* 591 * If the RST bit is set, check the sequence number to see 592 * if this is a valid reset segment. 593 * 594 * RFC 793 page 37: 595 * In all states except SYN-SENT, all reset (RST) segments 596 * are validated by checking their SEQ-fields. A reset is 597 * valid if its sequence number is in the window. 598 * 599 * RFC 793 page 69: 600 * There are four cases for the acceptability test for an incoming 601 * segment: 602 * 603 * Segment Receive Test 604 * Length Window 605 * ------- ------- ------------------------------------------- 606 * 0 0 SEG.SEQ = RCV.NXT 607 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 608 * >0 0 not acceptable 609 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 610 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 611 * 612 * Note that when receiving a SYN segment in the LISTEN state, 613 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 614 * described in RFC 793, page 66. 615 */ 616 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 617 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 618 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 619 if (V_tcp_insecure_rst || 620 th->th_seq == sc->sc_irs + 1) { 621 syncache_drop(sc, sch); 622 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 623 log(LOG_DEBUG, 624 "%s; %s: Our SYN|ACK was rejected, " 625 "connection attempt aborted by remote " 626 "endpoint\n", 627 s, __func__); 628 TCPSTAT_INC(tcps_sc_reset); 629 } else { 630 TCPSTAT_INC(tcps_badrst); 631 /* Send challenge ACK. */ 632 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 633 log(LOG_DEBUG, "%s; %s: RST with invalid " 634 " SEQ %u != NXT %u (+WND %u), " 635 "sending challenge ACK\n", 636 s, __func__, 637 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 638 syncache_respond(sc, sch, m, TH_ACK); 639 } 640 } else { 641 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 642 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 643 "NXT %u (+WND %u), segment ignored\n", 644 s, __func__, 645 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 646 TCPSTAT_INC(tcps_badrst); 647 } 648 649done: 650 if (s != NULL) 651 free(s, M_TCPLOG); 652 SCH_UNLOCK(sch); 653} 654 655void 656syncache_badack(struct in_conninfo *inc) 657{ 658 struct syncache *sc; 659 struct syncache_head *sch; 660 661 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 662 SCH_LOCK_ASSERT(sch); 663 if (sc != NULL) { 664 syncache_drop(sc, sch); 665 TCPSTAT_INC(tcps_sc_badack); 666 } 667 SCH_UNLOCK(sch); 668} 669 670void 671syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq) 672{ 673 struct syncache *sc; 674 struct syncache_head *sch; 675 676 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 677 SCH_LOCK_ASSERT(sch); 678 if (sc == NULL) 679 goto done; 680 681 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 682 if (ntohl(th_seq) != sc->sc_iss) 683 goto done; 684 685 /* 686 * If we've rertransmitted 3 times and this is our second error, 687 * we remove the entry. Otherwise, we allow it to continue on. 688 * This prevents us from incorrectly nuking an entry during a 689 * spurious network outage. 690 * 691 * See tcp_notify(). 692 */ 693 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 694 sc->sc_flags |= SCF_UNREACH; 695 goto done; 696 } 697 syncache_drop(sc, sch); 698 TCPSTAT_INC(tcps_sc_unreach); 699done: 700 SCH_UNLOCK(sch); 701} 702 703/* 704 * Build a new TCP socket structure from a syncache entry. 705 * 706 * On success return the newly created socket with its underlying inp locked. 707 */ 708static struct socket * 709syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 710{ 711 struct tcp_function_block *blk; 712 struct inpcb *inp = NULL; 713 struct socket *so; 714 struct tcpcb *tp; 715 int error; 716 char *s; 717 718 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 719 720 /* 721 * Ok, create the full blown connection, and set things up 722 * as they would have been set up if we had created the 723 * connection when the SYN arrived. If we can't create 724 * the connection, abort it. 725 */ 726 so = sonewconn(lso, 0); 727 if (so == NULL) { 728 /* 729 * Drop the connection; we will either send a RST or 730 * have the peer retransmit its SYN again after its 731 * RTO and try again. 732 */ 733 TCPSTAT_INC(tcps_listendrop); 734 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 735 log(LOG_DEBUG, "%s; %s: Socket create failed " 736 "due to limits or memory shortage\n", 737 s, __func__); 738 free(s, M_TCPLOG); 739 } 740 goto abort2; 741 } 742#ifdef MAC 743 mac_socketpeer_set_from_mbuf(m, so); 744#endif 745 746 inp = sotoinpcb(so); 747 inp->inp_inc.inc_fibnum = so->so_fibnum; 748 INP_WLOCK(inp); 749 /* 750 * Exclusive pcbinfo lock is not required in syncache socket case even 751 * if two inpcb locks can be acquired simultaneously: 752 * - the inpcb in LISTEN state, 753 * - the newly created inp. 754 * 755 * In this case, an inp cannot be at same time in LISTEN state and 756 * just created by an accept() call. 757 */ 758 INP_HASH_WLOCK(&V_tcbinfo); 759 760 /* Insert new socket into PCB hash list. */ 761 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 762#ifdef INET6 763 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 764 inp->inp_vflag &= ~INP_IPV4; 765 inp->inp_vflag |= INP_IPV6; 766 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 767 } else { 768 inp->inp_vflag &= ~INP_IPV6; 769 inp->inp_vflag |= INP_IPV4; 770#endif 771 inp->inp_ip_ttl = sc->sc_ip_ttl; 772 inp->inp_ip_tos = sc->sc_ip_tos; 773 inp->inp_laddr = sc->sc_inc.inc_laddr; 774#ifdef INET6 775 } 776#endif 777 778 /* 779 * If there's an mbuf and it has a flowid, then let's initialise the 780 * inp with that particular flowid. 781 */ 782 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 783 inp->inp_flowid = m->m_pkthdr.flowid; 784 inp->inp_flowtype = M_HASHTYPE_GET(m); 785 } 786 787 inp->inp_lport = sc->sc_inc.inc_lport; 788#ifdef INET6 789 if (inp->inp_vflag & INP_IPV6PROTO) { 790 struct inpcb *oinp = sotoinpcb(lso); 791 792 /* 793 * Inherit socket options from the listening socket. 794 * Note that in6p_inputopts are not (and should not be) 795 * copied, since it stores previously received options and is 796 * used to detect if each new option is different than the 797 * previous one and hence should be passed to a user. 798 * If we copied in6p_inputopts, a user would not be able to 799 * receive options just after calling the accept system call. 800 */ 801 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 802 if (oinp->in6p_outputopts) 803 inp->in6p_outputopts = 804 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 805 inp->in6p_hops = oinp->in6p_hops; 806 } 807 808 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 809 struct in6_addr laddr6; 810 struct sockaddr_in6 sin6; 811 812 sin6.sin6_family = AF_INET6; 813 sin6.sin6_len = sizeof(sin6); 814 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 815 sin6.sin6_port = sc->sc_inc.inc_fport; 816 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 817 laddr6 = inp->in6p_laddr; 818 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 819 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 820 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 821 thread0.td_ucred, m, false)) != 0) { 822 inp->in6p_laddr = laddr6; 823 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 824 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 825 "with error %i\n", 826 s, __func__, error); 827 free(s, M_TCPLOG); 828 } 829 INP_HASH_WUNLOCK(&V_tcbinfo); 830 goto abort; 831 } 832 /* Override flowlabel from in6_pcbconnect. */ 833 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 834 inp->inp_flow |= sc->sc_flowlabel; 835 } 836#endif /* INET6 */ 837#if defined(INET) && defined(INET6) 838 else 839#endif 840#ifdef INET 841 { 842 struct in_addr laddr; 843 struct sockaddr_in sin; 844 845 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 846 847 if (inp->inp_options == NULL) { 848 inp->inp_options = sc->sc_ipopts; 849 sc->sc_ipopts = NULL; 850 } 851 852 sin.sin_family = AF_INET; 853 sin.sin_len = sizeof(sin); 854 sin.sin_addr = sc->sc_inc.inc_faddr; 855 sin.sin_port = sc->sc_inc.inc_fport; 856 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 857 laddr = inp->inp_laddr; 858 if (inp->inp_laddr.s_addr == INADDR_ANY) 859 inp->inp_laddr = sc->sc_inc.inc_laddr; 860 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 861 thread0.td_ucred, m, false)) != 0) { 862 inp->inp_laddr = laddr; 863 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 864 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 865 "with error %i\n", 866 s, __func__, error); 867 free(s, M_TCPLOG); 868 } 869 INP_HASH_WUNLOCK(&V_tcbinfo); 870 goto abort; 871 } 872 } 873#endif /* INET */ 874#if defined(IPSEC) || defined(IPSEC_SUPPORT) 875 /* Copy old policy into new socket's. */ 876 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 877 printf("syncache_socket: could not copy policy\n"); 878#endif 879 INP_HASH_WUNLOCK(&V_tcbinfo); 880 tp = intotcpcb(inp); 881 tcp_state_change(tp, TCPS_SYN_RECEIVED); 882 tp->iss = sc->sc_iss; 883 tp->irs = sc->sc_irs; 884 tcp_rcvseqinit(tp); 885 tcp_sendseqinit(tp); 886 blk = sototcpcb(lso)->t_fb; 887 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 888 /* 889 * Our parents t_fb was not the default, 890 * we need to release our ref on tp->t_fb and 891 * pickup one on the new entry. 892 */ 893 struct tcp_function_block *rblk; 894 895 rblk = find_and_ref_tcp_fb(blk); 896 KASSERT(rblk != NULL, 897 ("cannot find blk %p out of syncache?", blk)); 898 if (tp->t_fb->tfb_tcp_fb_fini) 899 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 900 refcount_release(&tp->t_fb->tfb_refcnt); 901 tp->t_fb = rblk; 902 /* 903 * XXXrrs this is quite dangerous, it is possible 904 * for the new function to fail to init. We also 905 * are not asking if the handoff_is_ok though at 906 * the very start thats probalbly ok. 907 */ 908 if (tp->t_fb->tfb_tcp_fb_init) { 909 (*tp->t_fb->tfb_tcp_fb_init)(tp); 910 } 911 } 912 tp->snd_wl1 = sc->sc_irs; 913 tp->snd_max = tp->iss + 1; 914 tp->snd_nxt = tp->iss + 1; 915 tp->rcv_up = sc->sc_irs + 1; 916 tp->rcv_wnd = sc->sc_wnd; 917 tp->rcv_adv += tp->rcv_wnd; 918 tp->last_ack_sent = tp->rcv_nxt; 919 920 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 921 if (sc->sc_flags & SCF_NOOPT) 922 tp->t_flags |= TF_NOOPT; 923 else { 924 if (sc->sc_flags & SCF_WINSCALE) { 925 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 926 tp->snd_scale = sc->sc_requested_s_scale; 927 tp->request_r_scale = sc->sc_requested_r_scale; 928 } 929 if (sc->sc_flags & SCF_TIMESTAMP) { 930 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 931 tp->ts_recent = sc->sc_tsreflect; 932 tp->ts_recent_age = tcp_ts_getticks(); 933 tp->ts_offset = sc->sc_tsoff; 934 } 935#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 936 if (sc->sc_flags & SCF_SIGNATURE) 937 tp->t_flags |= TF_SIGNATURE; 938#endif 939 if (sc->sc_flags & SCF_SACK) 940 tp->t_flags |= TF_SACK_PERMIT; 941 } 942 943 if (sc->sc_flags & SCF_ECN) 944 tp->t_flags |= TF_ECN_PERMIT; 945 946 /* 947 * Set up MSS and get cached values from tcp_hostcache. 948 * This might overwrite some of the defaults we just set. 949 */ 950 tcp_mss(tp, sc->sc_peer_mss); 951 952 /* 953 * If the SYN,ACK was retransmitted, indicate that CWND to be 954 * limited to one segment in cc_conn_init(). 955 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 956 */ 957 if (sc->sc_rxmits > 1) 958 tp->snd_cwnd = 1; 959 960#ifdef TCP_OFFLOAD 961 /* 962 * Allow a TOE driver to install its hooks. Note that we hold the 963 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 964 * new connection before the TOE driver has done its thing. 965 */ 966 if (ADDED_BY_TOE(sc)) { 967 struct toedev *tod = sc->sc_tod; 968 969 tod->tod_offload_socket(tod, sc->sc_todctx, so); 970 } 971#endif 972 /* 973 * Copy and activate timers. 974 */ 975 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 976 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 977 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 978 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 979 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 980 981 TCPSTAT_INC(tcps_accepts); 982 return (so); 983 984abort: 985 INP_WUNLOCK(inp); 986abort2: 987 if (so != NULL) 988 soabort(so); 989 return (NULL); 990} 991 992/* 993 * This function gets called when we receive an ACK for a 994 * socket in the LISTEN state. We look up the connection 995 * in the syncache, and if its there, we pull it out of 996 * the cache and turn it into a full-blown connection in 997 * the SYN-RECEIVED state. 998 * 999 * On syncache_socket() success the newly created socket 1000 * has its underlying inp locked. 1001 */ 1002int 1003syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1004 struct socket **lsop, struct mbuf *m) 1005{ 1006 struct syncache *sc; 1007 struct syncache_head *sch; 1008 struct syncache scs; 1009 char *s; 1010 1011 /* 1012 * Global TCP locks are held because we manipulate the PCB lists 1013 * and create a new socket. 1014 */ 1015 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1016 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1017 ("%s: can handle only ACK", __func__)); 1018 1019 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1020 SCH_LOCK_ASSERT(sch); 1021 1022#ifdef INVARIANTS 1023 /* 1024 * Test code for syncookies comparing the syncache stored 1025 * values with the reconstructed values from the cookie. 1026 */ 1027 if (sc != NULL) 1028 syncookie_cmp(inc, sch, sc, th, to, *lsop); 1029#endif 1030 1031 if (sc == NULL) { 1032 /* 1033 * There is no syncache entry, so see if this ACK is 1034 * a returning syncookie. To do this, first: 1035 * A. Check if syncookies are used in case of syncache 1036 * overflows 1037 * B. See if this socket has had a syncache entry dropped in 1038 * the recent past. We don't want to accept a bogus 1039 * syncookie if we've never received a SYN or accept it 1040 * twice. 1041 * C. check that the syncookie is valid. If it is, then 1042 * cobble up a fake syncache entry, and return. 1043 */ 1044 if (!V_tcp_syncookies) { 1045 SCH_UNLOCK(sch); 1046 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1047 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1048 "segment rejected (syncookies disabled)\n", 1049 s, __func__); 1050 goto failed; 1051 } 1052 if (!V_tcp_syncookiesonly && 1053 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1054 SCH_UNLOCK(sch); 1055 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1056 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1057 "segment rejected (no syncache entry)\n", 1058 s, __func__); 1059 goto failed; 1060 } 1061 bzero(&scs, sizeof(scs)); 1062 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1063 SCH_UNLOCK(sch); 1064 if (sc == NULL) { 1065 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1066 log(LOG_DEBUG, "%s; %s: Segment failed " 1067 "SYNCOOKIE authentication, segment rejected " 1068 "(probably spoofed)\n", s, __func__); 1069 goto failed; 1070 } 1071#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1072 /* If received ACK has MD5 signature, check it. */ 1073 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1074 (!TCPMD5_ENABLED() || 1075 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1076 /* Drop the ACK. */ 1077 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1078 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1079 "MD5 signature doesn't match.\n", 1080 s, __func__); 1081 free(s, M_TCPLOG); 1082 } 1083 TCPSTAT_INC(tcps_sig_err_sigopt); 1084 return (-1); /* Do not send RST */ 1085 } 1086#endif /* TCP_SIGNATURE */ 1087 } else { 1088#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1089 /* 1090 * If listening socket requested TCP digests, check that 1091 * received ACK has signature and it is correct. 1092 * If not, drop the ACK and leave sc entry in th cache, 1093 * because SYN was received with correct signature. 1094 */ 1095 if (sc->sc_flags & SCF_SIGNATURE) { 1096 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1097 /* No signature */ 1098 TCPSTAT_INC(tcps_sig_err_nosigopt); 1099 SCH_UNLOCK(sch); 1100 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1101 log(LOG_DEBUG, "%s; %s: Segment " 1102 "rejected, MD5 signature wasn't " 1103 "provided.\n", s, __func__); 1104 free(s, M_TCPLOG); 1105 } 1106 return (-1); /* Do not send RST */ 1107 } 1108 if (!TCPMD5_ENABLED() || 1109 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1110 /* Doesn't match or no SA */ 1111 SCH_UNLOCK(sch); 1112 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1113 log(LOG_DEBUG, "%s; %s: Segment " 1114 "rejected, MD5 signature doesn't " 1115 "match.\n", s, __func__); 1116 free(s, M_TCPLOG); 1117 } 1118 return (-1); /* Do not send RST */ 1119 } 1120 } 1121#endif /* TCP_SIGNATURE */ 1122 1123 /* 1124 * RFC 7323 PAWS: If we have a timestamp on this segment and 1125 * it's less than ts_recent, drop it. 1126 * XXXMT: RFC 7323 also requires to send an ACK. 1127 * In tcp_input.c this is only done for TCP segments 1128 * with user data, so be consistent here and just drop 1129 * the segment. 1130 */ 1131 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1132 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1133 SCH_UNLOCK(sch); 1134 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1135 log(LOG_DEBUG, 1136 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1137 "segment dropped\n", s, __func__, 1138 to->to_tsval, sc->sc_tsreflect); 1139 free(s, M_TCPLOG); 1140 } 1141 return (-1); /* Do not send RST */ 1142 } 1143 1144 /* 1145 * If timestamps were not negotiated during SYN/ACK and a 1146 * segment with a timestamp is received, ignore the 1147 * timestamp and process the packet normally. 1148 * See section 3.2 of RFC 7323. 1149 */ 1150 if (!(sc->sc_flags & SCF_TIMESTAMP) && 1151 (to->to_flags & TOF_TS)) { 1152 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1153 log(LOG_DEBUG, "%s; %s: Timestamp not " 1154 "expected, segment processed normally\n", 1155 s, __func__); 1156 free(s, M_TCPLOG); 1157 s = NULL; 1158 } 1159 } 1160 1161 /* 1162 * If timestamps were negotiated during SYN/ACK and a 1163 * segment without a timestamp is received, silently drop 1164 * the segment, unless the missing timestamps are tolerated. 1165 * See section 3.2 of RFC 7323. 1166 */ 1167 if ((sc->sc_flags & SCF_TIMESTAMP) && 1168 !(to->to_flags & TOF_TS)) { 1169 if (V_tcp_tolerate_missing_ts) { 1170 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1171 log(LOG_DEBUG, 1172 "%s; %s: Timestamp missing, " 1173 "segment processed normally\n", 1174 s, __func__); 1175 free(s, M_TCPLOG); 1176 } 1177 } else { 1178 SCH_UNLOCK(sch); 1179 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1180 log(LOG_DEBUG, 1181 "%s; %s: Timestamp missing, " 1182 "segment silently dropped\n", 1183 s, __func__); 1184 free(s, M_TCPLOG); 1185 } 1186 return (-1); /* Do not send RST */ 1187 } 1188 } 1189 1190 /* 1191 * Pull out the entry to unlock the bucket row. 1192 * 1193 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1194 * tcp_state_change(). The tcpcb is not existent at this 1195 * moment. A new one will be allocated via syncache_socket-> 1196 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1197 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1198 */ 1199 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1200 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1201 sch->sch_length--; 1202#ifdef TCP_OFFLOAD 1203 if (ADDED_BY_TOE(sc)) { 1204 struct toedev *tod = sc->sc_tod; 1205 1206 tod->tod_syncache_removed(tod, sc->sc_todctx); 1207 } 1208#endif 1209 SCH_UNLOCK(sch); 1210 } 1211 1212 /* 1213 * Segment validation: 1214 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1215 */ 1216 if (th->th_ack != sc->sc_iss + 1) { 1217 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1218 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1219 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1220 goto failed; 1221 } 1222 1223 /* 1224 * The SEQ must fall in the window starting at the received 1225 * initial receive sequence number + 1 (the SYN). 1226 */ 1227 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1228 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1229 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1230 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1231 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1232 goto failed; 1233 } 1234 1235 *lsop = syncache_socket(sc, *lsop, m); 1236 1237 if (*lsop == NULL) 1238 TCPSTAT_INC(tcps_sc_aborted); 1239 else 1240 TCPSTAT_INC(tcps_sc_completed); 1241 1242/* how do we find the inp for the new socket? */ 1243 if (sc != &scs) 1244 syncache_free(sc); 1245 return (1); 1246failed: 1247 if (sc != NULL && sc != &scs) 1248 syncache_free(sc); 1249 if (s != NULL) 1250 free(s, M_TCPLOG); 1251 *lsop = NULL; 1252 return (0); 1253} 1254 1255static void 1256syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1257 uint64_t response_cookie) 1258{ 1259 struct inpcb *inp; 1260 struct tcpcb *tp; 1261 unsigned int *pending_counter; 1262 1263 /* 1264 * Global TCP locks are held because we manipulate the PCB lists 1265 * and create a new socket. 1266 */ 1267 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1268 1269 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1270 *lsop = syncache_socket(sc, *lsop, m); 1271 if (*lsop == NULL) { 1272 TCPSTAT_INC(tcps_sc_aborted); 1273 atomic_subtract_int(pending_counter, 1); 1274 } else { 1275 soisconnected(*lsop); 1276 inp = sotoinpcb(*lsop); 1277 tp = intotcpcb(inp); 1278 tp->t_flags |= TF_FASTOPEN; 1279 tp->t_tfo_cookie.server = response_cookie; 1280 tp->snd_max = tp->iss; 1281 tp->snd_nxt = tp->iss; 1282 tp->t_tfo_pending = pending_counter; 1283 TCPSTAT_INC(tcps_sc_completed); 1284 } 1285} 1286 1287/* 1288 * Given a LISTEN socket and an inbound SYN request, add 1289 * this to the syn cache, and send back a segment: 1290 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1291 * to the source. 1292 * 1293 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1294 * Doing so would require that we hold onto the data and deliver it 1295 * to the application. However, if we are the target of a SYN-flood 1296 * DoS attack, an attacker could send data which would eventually 1297 * consume all available buffer space if it were ACKed. By not ACKing 1298 * the data, we avoid this DoS scenario. 1299 * 1300 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1301 * cookie is processed and a new socket is created. In this case, any data 1302 * accompanying the SYN will be queued to the socket by tcp_input() and will 1303 * be ACKed either when the application sends response data or the delayed 1304 * ACK timer expires, whichever comes first. 1305 */ 1306int 1307syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1308 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1309 void *todctx, uint8_t iptos) 1310{ 1311 struct tcpcb *tp; 1312 struct socket *so; 1313 struct syncache *sc = NULL; 1314 struct syncache_head *sch; 1315 struct mbuf *ipopts = NULL; 1316 u_int ltflags; 1317 int win, ip_ttl, ip_tos; 1318 char *s; 1319 int rv = 0; 1320#ifdef INET6 1321 int autoflowlabel = 0; 1322#endif 1323#ifdef MAC 1324 struct label *maclabel; 1325#endif 1326 struct syncache scs; 1327 struct ucred *cred; 1328 uint64_t tfo_response_cookie; 1329 unsigned int *tfo_pending = NULL; 1330 int tfo_cookie_valid = 0; 1331 int tfo_response_cookie_valid = 0; 1332 1333 INP_WLOCK_ASSERT(inp); /* listen socket */ 1334 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1335 ("%s: unexpected tcp flags", __func__)); 1336 1337 /* 1338 * Combine all so/tp operations very early to drop the INP lock as 1339 * soon as possible. 1340 */ 1341 so = *lsop; 1342 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1343 tp = sototcpcb(so); 1344 cred = crhold(so->so_cred); 1345 1346#ifdef INET6 1347 if (inc->inc_flags & INC_ISIPV6) { 1348 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1349 autoflowlabel = 1; 1350 } 1351 ip_ttl = in6_selecthlim(inp, NULL); 1352 if ((inp->in6p_outputopts == NULL) || 1353 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1354 ip_tos = 0; 1355 } else { 1356 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1357 } 1358 } 1359#endif 1360#if defined(INET6) && defined(INET) 1361 else 1362#endif 1363#ifdef INET 1364 { 1365 ip_ttl = inp->inp_ip_ttl; 1366 ip_tos = inp->inp_ip_tos; 1367 } 1368#endif 1369 win = so->sol_sbrcv_hiwat; 1370 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1371 1372 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1373 (tp->t_tfo_pending != NULL) && 1374 (to->to_flags & TOF_FASTOPEN)) { 1375 /* 1376 * Limit the number of pending TFO connections to 1377 * approximately half of the queue limit. This prevents TFO 1378 * SYN floods from starving the service by filling the 1379 * listen queue with bogus TFO connections. 1380 */ 1381 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1382 (so->sol_qlimit / 2)) { 1383 int result; 1384 1385 result = tcp_fastopen_check_cookie(inc, 1386 to->to_tfo_cookie, to->to_tfo_len, 1387 &tfo_response_cookie); 1388 tfo_cookie_valid = (result > 0); 1389 tfo_response_cookie_valid = (result >= 0); 1390 } 1391 1392 /* 1393 * Remember the TFO pending counter as it will have to be 1394 * decremented below if we don't make it to syncache_tfo_expand(). 1395 */ 1396 tfo_pending = tp->t_tfo_pending; 1397 } 1398 1399 /* By the time we drop the lock these should no longer be used. */ 1400 so = NULL; 1401 tp = NULL; 1402 1403#ifdef MAC 1404 if (mac_syncache_init(&maclabel) != 0) { 1405 INP_WUNLOCK(inp); 1406 goto done; 1407 } else 1408 mac_syncache_create(maclabel, inp); 1409#endif 1410 if (!tfo_cookie_valid) 1411 INP_WUNLOCK(inp); 1412 1413 /* 1414 * Remember the IP options, if any. 1415 */ 1416#ifdef INET6 1417 if (!(inc->inc_flags & INC_ISIPV6)) 1418#endif 1419#ifdef INET 1420 ipopts = (m) ? ip_srcroute(m) : NULL; 1421#else 1422 ipopts = NULL; 1423#endif 1424 1425#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1426 /* 1427 * If listening socket requested TCP digests, check that received 1428 * SYN has signature and it is correct. If signature doesn't match 1429 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1430 */ 1431 if (ltflags & TF_SIGNATURE) { 1432 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1433 TCPSTAT_INC(tcps_sig_err_nosigopt); 1434 goto done; 1435 } 1436 if (!TCPMD5_ENABLED() || 1437 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1438 goto done; 1439 } 1440#endif /* TCP_SIGNATURE */ 1441 /* 1442 * See if we already have an entry for this connection. 1443 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1444 * 1445 * XXX: should the syncache be re-initialized with the contents 1446 * of the new SYN here (which may have different options?) 1447 * 1448 * XXX: We do not check the sequence number to see if this is a 1449 * real retransmit or a new connection attempt. The question is 1450 * how to handle such a case; either ignore it as spoofed, or 1451 * drop the current entry and create a new one? 1452 */ 1453 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1454 SCH_LOCK_ASSERT(sch); 1455 if (sc != NULL) { 1456 if (tfo_cookie_valid) 1457 INP_WUNLOCK(inp); 1458 TCPSTAT_INC(tcps_sc_dupsyn); 1459 if (ipopts) { 1460 /* 1461 * If we were remembering a previous source route, 1462 * forget it and use the new one we've been given. 1463 */ 1464 if (sc->sc_ipopts) 1465 (void) m_free(sc->sc_ipopts); 1466 sc->sc_ipopts = ipopts; 1467 } 1468 /* 1469 * Update timestamp if present. 1470 */ 1471 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1472 sc->sc_tsreflect = to->to_tsval; 1473 else 1474 sc->sc_flags &= ~SCF_TIMESTAMP; 1475 /* 1476 * Disable ECN if needed. 1477 */ 1478 if ((sc->sc_flags & SCF_ECN) && 1479 ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) { 1480 sc->sc_flags &= ~SCF_ECN; 1481 } 1482#ifdef MAC 1483 /* 1484 * Since we have already unconditionally allocated label 1485 * storage, free it up. The syncache entry will already 1486 * have an initialized label we can use. 1487 */ 1488 mac_syncache_destroy(&maclabel); 1489#endif 1490 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1491 /* Retransmit SYN|ACK and reset retransmit count. */ 1492 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1493 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1494 "resetting timer and retransmitting SYN|ACK\n", 1495 s, __func__); 1496 free(s, M_TCPLOG); 1497 } 1498 if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) { 1499 sc->sc_rxmits = 0; 1500 syncache_timeout(sc, sch, 1); 1501 TCPSTAT_INC(tcps_sndacks); 1502 TCPSTAT_INC(tcps_sndtotal); 1503 } 1504 SCH_UNLOCK(sch); 1505 goto donenoprobe; 1506 } 1507 1508 if (tfo_cookie_valid) { 1509 bzero(&scs, sizeof(scs)); 1510 sc = &scs; 1511 goto skip_alloc; 1512 } 1513 1514 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1515 if (sc == NULL) { 1516 /* 1517 * The zone allocator couldn't provide more entries. 1518 * Treat this as if the cache was full; drop the oldest 1519 * entry and insert the new one. 1520 */ 1521 TCPSTAT_INC(tcps_sc_zonefail); 1522 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1523 sch->sch_last_overflow = time_uptime; 1524 syncache_drop(sc, sch); 1525 } 1526 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1527 if (sc == NULL) { 1528 if (V_tcp_syncookies) { 1529 bzero(&scs, sizeof(scs)); 1530 sc = &scs; 1531 } else { 1532 SCH_UNLOCK(sch); 1533 if (ipopts) 1534 (void) m_free(ipopts); 1535 goto done; 1536 } 1537 } 1538 } 1539 1540skip_alloc: 1541 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1542 sc->sc_tfo_cookie = &tfo_response_cookie; 1543 1544 /* 1545 * Fill in the syncache values. 1546 */ 1547#ifdef MAC 1548 sc->sc_label = maclabel; 1549#endif 1550 sc->sc_cred = cred; 1551 cred = NULL; 1552 sc->sc_ipopts = ipopts; 1553 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1554 sc->sc_ip_tos = ip_tos; 1555 sc->sc_ip_ttl = ip_ttl; 1556#ifdef TCP_OFFLOAD 1557 sc->sc_tod = tod; 1558 sc->sc_todctx = todctx; 1559#endif 1560 sc->sc_irs = th->th_seq; 1561 sc->sc_iss = arc4random(); 1562 sc->sc_flags = 0; 1563 sc->sc_flowlabel = 0; 1564 1565 /* 1566 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1567 * win was derived from socket earlier in the function. 1568 */ 1569 win = imax(win, 0); 1570 win = imin(win, TCP_MAXWIN); 1571 sc->sc_wnd = win; 1572 1573 if (V_tcp_do_rfc1323) { 1574 /* 1575 * A timestamp received in a SYN makes 1576 * it ok to send timestamp requests and replies. 1577 */ 1578 if (to->to_flags & TOF_TS) { 1579 sc->sc_tsreflect = to->to_tsval; 1580 sc->sc_flags |= SCF_TIMESTAMP; 1581 sc->sc_tsoff = tcp_new_ts_offset(inc); 1582 } 1583 if (to->to_flags & TOF_SCALE) { 1584 int wscale = 0; 1585 1586 /* 1587 * Pick the smallest possible scaling factor that 1588 * will still allow us to scale up to sb_max, aka 1589 * kern.ipc.maxsockbuf. 1590 * 1591 * We do this because there are broken firewalls that 1592 * will corrupt the window scale option, leading to 1593 * the other endpoint believing that our advertised 1594 * window is unscaled. At scale factors larger than 1595 * 5 the unscaled window will drop below 1500 bytes, 1596 * leading to serious problems when traversing these 1597 * broken firewalls. 1598 * 1599 * With the default maxsockbuf of 256K, a scale factor 1600 * of 3 will be chosen by this algorithm. Those who 1601 * choose a larger maxsockbuf should watch out 1602 * for the compatibility problems mentioned above. 1603 * 1604 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1605 * or <SYN,ACK>) segment itself is never scaled. 1606 */ 1607 while (wscale < TCP_MAX_WINSHIFT && 1608 (TCP_MAXWIN << wscale) < sb_max) 1609 wscale++; 1610 sc->sc_requested_r_scale = wscale; 1611 sc->sc_requested_s_scale = to->to_wscale; 1612 sc->sc_flags |= SCF_WINSCALE; 1613 } 1614 } 1615#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1616 /* 1617 * If listening socket requested TCP digests, flag this in the 1618 * syncache so that syncache_respond() will do the right thing 1619 * with the SYN+ACK. 1620 */ 1621 if (ltflags & TF_SIGNATURE) 1622 sc->sc_flags |= SCF_SIGNATURE; 1623#endif /* TCP_SIGNATURE */ 1624 if (to->to_flags & TOF_SACKPERM) 1625 sc->sc_flags |= SCF_SACK; 1626 if (to->to_flags & TOF_MSS) 1627 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1628 if (ltflags & TF_NOOPT) 1629 sc->sc_flags |= SCF_NOOPT; 1630 if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) && 1631 V_tcp_do_ecn) 1632 sc->sc_flags |= SCF_ECN; 1633 1634 if (V_tcp_syncookies) 1635 sc->sc_iss = syncookie_generate(sch, sc); 1636#ifdef INET6 1637 if (autoflowlabel) { 1638 if (V_tcp_syncookies) 1639 sc->sc_flowlabel = sc->sc_iss; 1640 else 1641 sc->sc_flowlabel = ip6_randomflowlabel(); 1642 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1643 } 1644#endif 1645 SCH_UNLOCK(sch); 1646 1647 if (tfo_cookie_valid) { 1648 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1649 /* INP_WUNLOCK(inp) will be performed by the caller */ 1650 rv = 1; 1651 goto tfo_expanded; 1652 } 1653 1654 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1655 /* 1656 * Do a standard 3-way handshake. 1657 */ 1658 if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) { 1659 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1660 syncache_free(sc); 1661 else if (sc != &scs) 1662 syncache_insert(sc, sch); /* locks and unlocks sch */ 1663 TCPSTAT_INC(tcps_sndacks); 1664 TCPSTAT_INC(tcps_sndtotal); 1665 } else { 1666 if (sc != &scs) 1667 syncache_free(sc); 1668 TCPSTAT_INC(tcps_sc_dropped); 1669 } 1670 goto donenoprobe; 1671 1672done: 1673 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1674donenoprobe: 1675 if (m) { 1676 *lsop = NULL; 1677 m_freem(m); 1678 } 1679 /* 1680 * If tfo_pending is not NULL here, then a TFO SYN that did not 1681 * result in a new socket was processed and the associated pending 1682 * counter has not yet been decremented. All such TFO processing paths 1683 * transit this point. 1684 */ 1685 if (tfo_pending != NULL) 1686 tcp_fastopen_decrement_counter(tfo_pending); 1687 1688tfo_expanded: 1689 if (cred != NULL) 1690 crfree(cred); 1691#ifdef MAC 1692 if (sc == &scs) 1693 mac_syncache_destroy(&maclabel); 1694#endif 1695 return (rv); 1696} 1697 1698/* 1699 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1700 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1701 */ 1702static int 1703syncache_respond(struct syncache *sc, struct syncache_head *sch, 1704 const struct mbuf *m0, int flags) 1705{ 1706 struct ip *ip = NULL; 1707 struct mbuf *m; 1708 struct tcphdr *th = NULL; 1709 int optlen, error = 0; /* Make compiler happy */ 1710 u_int16_t hlen, tlen, mssopt; 1711 struct tcpopt to; 1712#ifdef INET6 1713 struct ip6_hdr *ip6 = NULL; 1714#endif 1715 hlen = 1716#ifdef INET6 1717 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1718#endif 1719 sizeof(struct ip); 1720 tlen = hlen + sizeof(struct tcphdr); 1721 1722 /* Determine MSS we advertize to other end of connection. */ 1723 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); 1724 1725 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1726 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1727 ("syncache: mbuf too small")); 1728 1729 /* Create the IP+TCP header from scratch. */ 1730 m = m_gethdr(M_NOWAIT, MT_DATA); 1731 if (m == NULL) 1732 return (ENOBUFS); 1733#ifdef MAC 1734 mac_syncache_create_mbuf(sc->sc_label, m); 1735#endif 1736 m->m_data += max_linkhdr; 1737 m->m_len = tlen; 1738 m->m_pkthdr.len = tlen; 1739 m->m_pkthdr.rcvif = NULL; 1740 1741#ifdef INET6 1742 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1743 ip6 = mtod(m, struct ip6_hdr *); 1744 ip6->ip6_vfc = IPV6_VERSION; 1745 ip6->ip6_nxt = IPPROTO_TCP; 1746 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1747 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1748 ip6->ip6_plen = htons(tlen - hlen); 1749 /* ip6_hlim is set after checksum */ 1750 /* Zero out traffic class and flow label. */ 1751 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1752 ip6->ip6_flow |= sc->sc_flowlabel; 1753 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20); 1754 1755 th = (struct tcphdr *)(ip6 + 1); 1756 } 1757#endif 1758#if defined(INET6) && defined(INET) 1759 else 1760#endif 1761#ifdef INET 1762 { 1763 ip = mtod(m, struct ip *); 1764 ip->ip_v = IPVERSION; 1765 ip->ip_hl = sizeof(struct ip) >> 2; 1766 ip->ip_len = htons(tlen); 1767 ip->ip_id = 0; 1768 ip->ip_off = 0; 1769 ip->ip_sum = 0; 1770 ip->ip_p = IPPROTO_TCP; 1771 ip->ip_src = sc->sc_inc.inc_laddr; 1772 ip->ip_dst = sc->sc_inc.inc_faddr; 1773 ip->ip_ttl = sc->sc_ip_ttl; 1774 ip->ip_tos = sc->sc_ip_tos; 1775 1776 /* 1777 * See if we should do MTU discovery. Route lookups are 1778 * expensive, so we will only unset the DF bit if: 1779 * 1780 * 1) path_mtu_discovery is disabled 1781 * 2) the SCF_UNREACH flag has been set 1782 */ 1783 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1784 ip->ip_off |= htons(IP_DF); 1785 1786 th = (struct tcphdr *)(ip + 1); 1787 } 1788#endif /* INET */ 1789 th->th_sport = sc->sc_inc.inc_lport; 1790 th->th_dport = sc->sc_inc.inc_fport; 1791 1792 if (flags & TH_SYN) 1793 th->th_seq = htonl(sc->sc_iss); 1794 else 1795 th->th_seq = htonl(sc->sc_iss + 1); 1796 th->th_ack = htonl(sc->sc_irs + 1); 1797 th->th_off = sizeof(struct tcphdr) >> 2; 1798 th->th_x2 = 0; 1799 th->th_flags = flags; 1800 th->th_win = htons(sc->sc_wnd); 1801 th->th_urp = 0; 1802 1803 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) { 1804 th->th_flags |= TH_ECE; 1805 TCPSTAT_INC(tcps_ecn_shs); 1806 } 1807 1808 /* Tack on the TCP options. */ 1809 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1810 to.to_flags = 0; 1811 1812 if (flags & TH_SYN) { 1813 to.to_mss = mssopt; 1814 to.to_flags = TOF_MSS; 1815 if (sc->sc_flags & SCF_WINSCALE) { 1816 to.to_wscale = sc->sc_requested_r_scale; 1817 to.to_flags |= TOF_SCALE; 1818 } 1819 if (sc->sc_flags & SCF_SACK) 1820 to.to_flags |= TOF_SACKPERM; 1821#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1822 if (sc->sc_flags & SCF_SIGNATURE) 1823 to.to_flags |= TOF_SIGNATURE; 1824#endif 1825 if (sc->sc_tfo_cookie) { 1826 to.to_flags |= TOF_FASTOPEN; 1827 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1828 to.to_tfo_cookie = sc->sc_tfo_cookie; 1829 /* don't send cookie again when retransmitting response */ 1830 sc->sc_tfo_cookie = NULL; 1831 } 1832 } 1833 if (sc->sc_flags & SCF_TIMESTAMP) { 1834 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1835 to.to_tsecr = sc->sc_tsreflect; 1836 to.to_flags |= TOF_TS; 1837 } 1838 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1839 1840 /* Adjust headers by option size. */ 1841 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1842 m->m_len += optlen; 1843 m->m_pkthdr.len += optlen; 1844#ifdef INET6 1845 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1846 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1847 else 1848#endif 1849 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1850#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1851 if (sc->sc_flags & SCF_SIGNATURE) { 1852 KASSERT(to.to_flags & TOF_SIGNATURE, 1853 ("tcp_addoptions() didn't set tcp_signature")); 1854 1855 /* NOTE: to.to_signature is inside of mbuf */ 1856 if (!TCPMD5_ENABLED() || 1857 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1858 m_freem(m); 1859 return (EACCES); 1860 } 1861 } 1862#endif 1863 } else 1864 optlen = 0; 1865 1866 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1867 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1868 /* 1869 * If we have peer's SYN and it has a flowid, then let's assign it to 1870 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1871 * to SYN|ACK due to lack of inp here. 1872 */ 1873 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1874 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1875 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1876 } 1877#ifdef INET6 1878 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1879 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1880 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1881 IPPROTO_TCP, 0); 1882 ip6->ip6_hlim = sc->sc_ip_ttl; 1883#ifdef TCP_OFFLOAD 1884 if (ADDED_BY_TOE(sc)) { 1885 struct toedev *tod = sc->sc_tod; 1886 1887 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1888 1889 return (error); 1890 } 1891#endif 1892 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 1893 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1894 } 1895#endif 1896#if defined(INET6) && defined(INET) 1897 else 1898#endif 1899#ifdef INET 1900 { 1901 m->m_pkthdr.csum_flags = CSUM_TCP; 1902 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1903 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1904#ifdef TCP_OFFLOAD 1905 if (ADDED_BY_TOE(sc)) { 1906 struct toedev *tod = sc->sc_tod; 1907 1908 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1909 1910 return (error); 1911 } 1912#endif 1913 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 1914 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1915 } 1916#endif 1917 return (error); 1918} 1919 1920/* 1921 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1922 * that exceed the capacity of the syncache by avoiding the storage of any 1923 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1924 * attacks where the attacker does not have access to our responses. 1925 * 1926 * Syncookies encode and include all necessary information about the 1927 * connection setup within the SYN|ACK that we send back. That way we 1928 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1929 * (if ever). Normally the syncache and syncookies are running in parallel 1930 * with the latter taking over when the former is exhausted. When matching 1931 * syncache entry is found the syncookie is ignored. 1932 * 1933 * The only reliable information persisting the 3WHS is our initial sequence 1934 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1935 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1936 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1937 * returns and signifies a legitimate connection if it matches the ACK. 1938 * 1939 * The available space of 32 bits to store the hash and to encode the SYN 1940 * option information is very tight and we should have at least 24 bits for 1941 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1942 * 1943 * SYN option information we have to encode to fully restore a connection: 1944 * MSS: is imporant to chose an optimal segment size to avoid IP level 1945 * fragmentation along the path. The common MSS values can be encoded 1946 * in a 3-bit table. Uncommon values are captured by the next lower value 1947 * in the table leading to a slight increase in packetization overhead. 1948 * WSCALE: is necessary to allow large windows to be used for high delay- 1949 * bandwidth product links. Not scaling the window when it was initially 1950 * negotiated is bad for performance as lack of scaling further decreases 1951 * the apparent available send window. We only need to encode the WSCALE 1952 * we received from the remote end. Our end can be recalculated at any 1953 * time. The common WSCALE values can be encoded in a 3-bit table. 1954 * Uncommon values are captured by the next lower value in the table 1955 * making us under-estimate the available window size halving our 1956 * theoretically possible maximum throughput for that connection. 1957 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1958 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1959 * that are included in all segments on a connection. We enable them when 1960 * the ACK has them. 1961 * 1962 * Security of syncookies and attack vectors: 1963 * 1964 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1965 * together with the gloabl secret to make it unique per connection attempt. 1966 * Thus any change of any of those parameters results in a different MAC output 1967 * in an unpredictable way unless a collision is encountered. 24 bits of the 1968 * MAC are embedded into the ISS. 1969 * 1970 * To prevent replay attacks two rotating global secrets are updated with a 1971 * new random value every 15 seconds. The life-time of a syncookie is thus 1972 * 15-30 seconds. 1973 * 1974 * Vector 1: Attacking the secret. This requires finding a weakness in the 1975 * MAC itself or the way it is used here. The attacker can do a chosen plain 1976 * text attack by varying and testing the all parameters under his control. 1977 * The strength depends on the size and randomness of the secret, and the 1978 * cryptographic security of the MAC function. Due to the constant updating 1979 * of the secret the attacker has at most 29.999 seconds to find the secret 1980 * and launch spoofed connections. After that he has to start all over again. 1981 * 1982 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1983 * size an average of 4,823 attempts are required for a 50% chance of success 1984 * to spoof a single syncookie (birthday collision paradox). However the 1985 * attacker is blind and doesn't know if one of his attempts succeeded unless 1986 * he has a side channel to interfere success from. A single connection setup 1987 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1988 * This many attempts are required for each one blind spoofed connection. For 1989 * every additional spoofed connection he has to launch another N attempts. 1990 * Thus for a sustained rate 100 spoofed connections per second approximately 1991 * 1,800,000 packets per second would have to be sent. 1992 * 1993 * NB: The MAC function should be fast so that it doesn't become a CPU 1994 * exhaustion attack vector itself. 1995 * 1996 * References: 1997 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1998 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1999 * http://cr.yp.to/syncookies.html (overview) 2000 * http://cr.yp.to/syncookies/archive (details) 2001 * 2002 * 2003 * Schematic construction of a syncookie enabled Initial Sequence Number: 2004 * 0 1 2 3 2005 * 12345678901234567890123456789012 2006 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2007 * 2008 * x 24 MAC (truncated) 2009 * W 3 Send Window Scale index 2010 * M 3 MSS index 2011 * S 1 SACK permitted 2012 * P 1 Odd/even secret 2013 */ 2014 2015/* 2016 * Distribution and probability of certain MSS values. Those in between are 2017 * rounded down to the next lower one. 2018 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2019 * .2% .3% 5% 7% 7% 20% 15% 45% 2020 */ 2021static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2022 2023/* 2024 * Distribution and probability of certain WSCALE values. We have to map the 2025 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2026 * bits based on prevalence of certain values. Where we don't have an exact 2027 * match for are rounded down to the next lower one letting us under-estimate 2028 * the true available window. At the moment this would happen only for the 2029 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2030 * and window size). The absence of the WSCALE option (no scaling in either 2031 * direction) is encoded with index zero. 2032 * [WSCALE values histograms, Allman, 2012] 2033 * X 10 10 35 5 6 14 10% by host 2034 * X 11 4 5 5 18 49 3% by connections 2035 */ 2036static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2037 2038/* 2039 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2040 * and good cryptographic properties. 2041 */ 2042static uint32_t 2043syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2044 uint8_t *secbits, uintptr_t secmod) 2045{ 2046 SIPHASH_CTX ctx; 2047 uint32_t siphash[2]; 2048 2049 SipHash24_Init(&ctx); 2050 SipHash_SetKey(&ctx, secbits); 2051 switch (inc->inc_flags & INC_ISIPV6) { 2052#ifdef INET 2053 case 0: 2054 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2055 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2056 break; 2057#endif 2058#ifdef INET6 2059 case INC_ISIPV6: 2060 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2061 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2062 break; 2063#endif 2064 } 2065 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2066 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2067 SipHash_Update(&ctx, &irs, sizeof(irs)); 2068 SipHash_Update(&ctx, &flags, sizeof(flags)); 2069 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2070 SipHash_Final((u_int8_t *)&siphash, &ctx); 2071 2072 return (siphash[0] ^ siphash[1]); 2073} 2074 2075static tcp_seq 2076syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2077{ 2078 u_int i, secbit, wscale; 2079 uint32_t iss, hash; 2080 uint8_t *secbits; 2081 union syncookie cookie; 2082 2083 SCH_LOCK_ASSERT(sch); 2084 2085 cookie.cookie = 0; 2086 2087 /* Map our computed MSS into the 3-bit index. */ 2088 for (i = nitems(tcp_sc_msstab) - 1; 2089 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2090 i--) 2091 ; 2092 cookie.flags.mss_idx = i; 2093 2094 /* 2095 * Map the send window scale into the 3-bit index but only if 2096 * the wscale option was received. 2097 */ 2098 if (sc->sc_flags & SCF_WINSCALE) { 2099 wscale = sc->sc_requested_s_scale; 2100 for (i = nitems(tcp_sc_wstab) - 1; 2101 tcp_sc_wstab[i] > wscale && i > 0; 2102 i--) 2103 ; 2104 cookie.flags.wscale_idx = i; 2105 } 2106 2107 /* Can we do SACK? */ 2108 if (sc->sc_flags & SCF_SACK) 2109 cookie.flags.sack_ok = 1; 2110 2111 /* Which of the two secrets to use. */ 2112 secbit = sch->sch_sc->secret.oddeven & 0x1; 2113 cookie.flags.odd_even = secbit; 2114 2115 secbits = sch->sch_sc->secret.key[secbit]; 2116 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2117 (uintptr_t)sch); 2118 2119 /* 2120 * Put the flags into the hash and XOR them to get better ISS number 2121 * variance. This doesn't enhance the cryptographic strength and is 2122 * done to prevent the 8 cookie bits from showing up directly on the 2123 * wire. 2124 */ 2125 iss = hash & ~0xff; 2126 iss |= cookie.cookie ^ (hash >> 24); 2127 2128 TCPSTAT_INC(tcps_sc_sendcookie); 2129 return (iss); 2130} 2131 2132static struct syncache * 2133syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2134 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2135 struct socket *lso) 2136{ 2137 uint32_t hash; 2138 uint8_t *secbits; 2139 tcp_seq ack, seq; 2140 int wnd, wscale = 0; 2141 union syncookie cookie; 2142 2143 SCH_LOCK_ASSERT(sch); 2144 2145 /* 2146 * Pull information out of SYN-ACK/ACK and revert sequence number 2147 * advances. 2148 */ 2149 ack = th->th_ack - 1; 2150 seq = th->th_seq - 1; 2151 2152 /* 2153 * Unpack the flags containing enough information to restore the 2154 * connection. 2155 */ 2156 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2157 2158 /* Which of the two secrets to use. */ 2159 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 2160 2161 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2162 2163 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2164 if ((ack & ~0xff) != (hash & ~0xff)) 2165 return (NULL); 2166 2167 /* Fill in the syncache values. */ 2168 sc->sc_flags = 0; 2169 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2170 sc->sc_ipopts = NULL; 2171 2172 sc->sc_irs = seq; 2173 sc->sc_iss = ack; 2174 2175 switch (inc->inc_flags & INC_ISIPV6) { 2176#ifdef INET 2177 case 0: 2178 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2179 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2180 break; 2181#endif 2182#ifdef INET6 2183 case INC_ISIPV6: 2184 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2185 sc->sc_flowlabel = 2186 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2187 break; 2188#endif 2189 } 2190 2191 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2192 2193 /* We can simply recompute receive window scale we sent earlier. */ 2194 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2195 wscale++; 2196 2197 /* Only use wscale if it was enabled in the orignal SYN. */ 2198 if (cookie.flags.wscale_idx > 0) { 2199 sc->sc_requested_r_scale = wscale; 2200 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2201 sc->sc_flags |= SCF_WINSCALE; 2202 } 2203 2204 wnd = lso->sol_sbrcv_hiwat; 2205 wnd = imax(wnd, 0); 2206 wnd = imin(wnd, TCP_MAXWIN); 2207 sc->sc_wnd = wnd; 2208 2209 if (cookie.flags.sack_ok) 2210 sc->sc_flags |= SCF_SACK; 2211 2212 if (to->to_flags & TOF_TS) { 2213 sc->sc_flags |= SCF_TIMESTAMP; 2214 sc->sc_tsreflect = to->to_tsval; 2215 sc->sc_tsoff = tcp_new_ts_offset(inc); 2216 } 2217 2218 if (to->to_flags & TOF_SIGNATURE) 2219 sc->sc_flags |= SCF_SIGNATURE; 2220 2221 sc->sc_rxmits = 0; 2222 2223 TCPSTAT_INC(tcps_sc_recvcookie); 2224 return (sc); 2225} 2226 2227#ifdef INVARIANTS 2228static int 2229syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2230 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2231 struct socket *lso) 2232{ 2233 struct syncache scs, *scx; 2234 char *s; 2235 2236 bzero(&scs, sizeof(scs)); 2237 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2238 2239 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2240 return (0); 2241 2242 if (scx != NULL) { 2243 if (sc->sc_peer_mss != scx->sc_peer_mss) 2244 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2245 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2246 2247 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2248 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2249 s, __func__, sc->sc_requested_r_scale, 2250 scx->sc_requested_r_scale); 2251 2252 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2253 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2254 s, __func__, sc->sc_requested_s_scale, 2255 scx->sc_requested_s_scale); 2256 2257 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2258 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2259 } 2260 2261 if (s != NULL) 2262 free(s, M_TCPLOG); 2263 return (0); 2264} 2265#endif /* INVARIANTS */ 2266 2267static void 2268syncookie_reseed(void *arg) 2269{ 2270 struct tcp_syncache *sc = arg; 2271 uint8_t *secbits; 2272 int secbit; 2273 2274 /* 2275 * Reseeding the secret doesn't have to be protected by a lock. 2276 * It only must be ensured that the new random values are visible 2277 * to all CPUs in a SMP environment. The atomic with release 2278 * semantics ensures that. 2279 */ 2280 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2281 secbits = sc->secret.key[secbit]; 2282 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2283 atomic_add_rel_int(&sc->secret.oddeven, 1); 2284 2285 /* Reschedule ourself. */ 2286 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2287} 2288 2289/* 2290 * Exports the syncache entries to userland so that netstat can display 2291 * them alongside the other sockets. This function is intended to be 2292 * called only from tcp_pcblist. 2293 * 2294 * Due to concurrency on an active system, the number of pcbs exported 2295 * may have no relation to max_pcbs. max_pcbs merely indicates the 2296 * amount of space the caller allocated for this function to use. 2297 */ 2298int 2299syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2300{ 2301 struct xtcpcb xt; 2302 struct syncache *sc; 2303 struct syncache_head *sch; 2304 int count, error, i; 2305 2306 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2307 sch = &V_tcp_syncache.hashbase[i]; 2308 SCH_LOCK(sch); 2309 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2310 if (count >= max_pcbs) { 2311 SCH_UNLOCK(sch); 2312 goto exit; 2313 } 2314 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2315 continue; 2316 bzero(&xt, sizeof(xt)); 2317 xt.xt_len = sizeof(xt); 2318 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2319 xt.xt_inp.inp_vflag = INP_IPV6; 2320 else 2321 xt.xt_inp.inp_vflag = INP_IPV4; 2322 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2323 sizeof (struct in_conninfo)); 2324 xt.t_state = TCPS_SYN_RECEIVED; 2325 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2326 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2327 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2328 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2329 error = SYSCTL_OUT(req, &xt, sizeof xt); 2330 if (error) { 2331 SCH_UNLOCK(sch); 2332 goto exit; 2333 } 2334 count++; 2335 } 2336 SCH_UNLOCK(sch); 2337 } 2338exit: 2339 *pcbs_exported = count; 2340 return error; 2341} 2342