1/*- 2 * Copyright (c) 2001 McAfee, Inc. 3 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG 4 * All rights reserved. 5 * 6 * This software was developed for the FreeBSD Project by Jonathan Lemon 7 * and McAfee Research, the Security Research Division of McAfee, Inc. under 8 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 9 * DARPA CHATS research program. [2001 McAfee, Inc.] 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 */ 32 33#include <sys/cdefs.h> 34__FBSDID("$FreeBSD: stable/10/sys/netinet/tcp_syncache.c 324520 2017-10-11 06:28:46Z sephe $"); 35 36#include "opt_inet.h" 37#include "opt_inet6.h" 38#include "opt_ipsec.h" 39#include "opt_pcbgroup.h" 40 41#include <sys/param.h> 42#include <sys/systm.h> 43#include <sys/kernel.h> 44#include <sys/sysctl.h> 45#include <sys/limits.h> 46#include <sys/lock.h> 47#include <sys/mutex.h> 48#include <sys/malloc.h> 49#include <sys/mbuf.h> 50#include <sys/proc.h> /* for proc0 declaration */ 51#include <sys/random.h> 52#include <sys/socket.h> 53#include <sys/socketvar.h> 54#include <sys/syslog.h> 55#include <sys/ucred.h> 56 57#include <sys/md5.h> 58#include <crypto/siphash/siphash.h> 59 60#include <vm/uma.h> 61 62#include <net/if.h> 63#include <net/route.h> 64#include <net/vnet.h> 65 66#include <netinet/in.h> 67#include <netinet/in_systm.h> 68#include <netinet/ip.h> 69#include <netinet/in_var.h> 70#include <netinet/in_pcb.h> 71#include <netinet/ip_var.h> 72#include <netinet/ip_options.h> 73#ifdef INET6 74#include <netinet/ip6.h> 75#include <netinet/icmp6.h> 76#include <netinet6/nd6.h> 77#include <netinet6/ip6_var.h> 78#include <netinet6/in6_pcb.h> 79#endif 80#include <netinet/tcp.h> 81#ifdef TCP_RFC7413 82#include <netinet/tcp_fastopen.h> 83#endif 84#include <netinet/tcp_fsm.h> 85#include <netinet/tcp_seq.h> 86#include <netinet/tcp_timer.h> 87#include <netinet/tcp_var.h> 88#include <netinet/tcp_syncache.h> 89#ifdef INET6 90#include <netinet6/tcp6_var.h> 91#endif 92#ifdef TCP_OFFLOAD 93#include <netinet/toecore.h> 94#endif 95 96#ifdef IPSEC 97#include <netipsec/ipsec.h> 98#ifdef INET6 99#include <netipsec/ipsec6.h> 100#endif 101#include <netipsec/key.h> 102#endif /*IPSEC*/ 103 104#include <machine/in_cksum.h> 105 106#include <security/mac/mac_framework.h> 107 108static VNET_DEFINE(int, tcp_syncookies) = 1; 109#define V_tcp_syncookies VNET(tcp_syncookies) 110SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW, 111 &VNET_NAME(tcp_syncookies), 0, 112 "Use TCP SYN cookies if the syncache overflows"); 113 114static VNET_DEFINE(int, tcp_syncookiesonly) = 0; 115#define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 116SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_RW, 117 &VNET_NAME(tcp_syncookiesonly), 0, 118 "Use only TCP SYN cookies"); 119 120#ifdef TCP_OFFLOAD 121#define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 122#endif 123 124static void syncache_drop(struct syncache *, struct syncache_head *); 125static void syncache_free(struct syncache *); 126static void syncache_insert(struct syncache *, struct syncache_head *); 127static int syncache_respond(struct syncache *, const struct mbuf *); 128static struct socket *syncache_socket(struct syncache *, struct socket *, 129 struct mbuf *m); 130static int syncache_sysctl_count(SYSCTL_HANDLER_ARGS); 131static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 132 int docallout); 133static void syncache_timer(void *); 134 135static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 136 uint8_t *, uintptr_t); 137static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 138static struct syncache 139 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 140 struct syncache *, struct tcphdr *, struct tcpopt *, 141 struct socket *); 142static void syncookie_reseed(void *); 143#ifdef INVARIANTS 144static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 145 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 146 struct socket *lso); 147#endif 148 149/* 150 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 151 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds, 152 * the odds are that the user has given up attempting to connect by then. 153 */ 154#define SYNCACHE_MAXREXMTS 3 155 156/* Arbitrary values */ 157#define TCP_SYNCACHE_HASHSIZE 512 158#define TCP_SYNCACHE_BUCKETLIMIT 30 159 160static VNET_DEFINE(struct tcp_syncache, tcp_syncache); 161#define V_tcp_syncache VNET(tcp_syncache) 162 163static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, 164 "TCP SYN cache"); 165 166SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN, 167 &VNET_NAME(tcp_syncache.bucket_limit), 0, 168 "Per-bucket hash limit for syncache"); 169 170SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN, 171 &VNET_NAME(tcp_syncache.cache_limit), 0, 172 "Overall entry limit for syncache"); 173 174SYSCTL_VNET_PROC(_net_inet_tcp_syncache, OID_AUTO, count, (CTLTYPE_UINT|CTLFLAG_RD), 175 NULL, 0, &syncache_sysctl_count, "IU", 176 "Current number of entries in syncache"); 177 178SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN, 179 &VNET_NAME(tcp_syncache.hashsize), 0, 180 "Size of TCP syncache hashtable"); 181 182SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW, 183 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 184 "Limit on SYN/ACK retransmissions"); 185 186VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 187SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 188 CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 189 "Send reset on socket allocation failure"); 190 191static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 192 193#define SYNCACHE_HASH(inc, mask) \ 194 ((V_tcp_syncache.hash_secret ^ \ 195 (inc)->inc_faddr.s_addr ^ \ 196 ((inc)->inc_faddr.s_addr >> 16) ^ \ 197 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 198 199#define SYNCACHE_HASH6(inc, mask) \ 200 ((V_tcp_syncache.hash_secret ^ \ 201 (inc)->inc6_faddr.s6_addr32[0] ^ \ 202 (inc)->inc6_faddr.s6_addr32[3] ^ \ 203 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 204 205#define ENDPTS_EQ(a, b) ( \ 206 (a)->ie_fport == (b)->ie_fport && \ 207 (a)->ie_lport == (b)->ie_lport && \ 208 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \ 209 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \ 210) 211 212#define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0) 213 214#define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 215#define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 216#define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 217 218/* 219 * Requires the syncache entry to be already removed from the bucket list. 220 */ 221static void 222syncache_free(struct syncache *sc) 223{ 224 225 if (sc->sc_ipopts) 226 (void) m_free(sc->sc_ipopts); 227 if (sc->sc_cred) 228 crfree(sc->sc_cred); 229#ifdef MAC 230 mac_syncache_destroy(&sc->sc_label); 231#endif 232 233 uma_zfree(V_tcp_syncache.zone, sc); 234} 235 236void 237syncache_init(void) 238{ 239 int i; 240 241 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 242 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 243 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 244 V_tcp_syncache.hash_secret = arc4random(); 245 246 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 247 &V_tcp_syncache.hashsize); 248 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 249 &V_tcp_syncache.bucket_limit); 250 if (!powerof2(V_tcp_syncache.hashsize) || 251 V_tcp_syncache.hashsize == 0) { 252 printf("WARNING: syncache hash size is not a power of 2.\n"); 253 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 254 } 255 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 256 257 /* Set limits. */ 258 V_tcp_syncache.cache_limit = 259 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 260 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 261 &V_tcp_syncache.cache_limit); 262 263 /* Allocate the hash table. */ 264 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 265 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 266 267#ifdef VIMAGE 268 V_tcp_syncache.vnet = curvnet; 269#endif 270 271 /* Initialize the hash buckets. */ 272 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 273 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 274 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 275 NULL, MTX_DEF); 276 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 277 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 278 V_tcp_syncache.hashbase[i].sch_length = 0; 279 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 280 V_tcp_syncache.hashbase[i].sch_last_overflow = 281 -(SYNCOOKIE_LIFETIME + 1); 282 } 283 284 /* Create the syncache entry zone. */ 285 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 286 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 287 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 288 V_tcp_syncache.cache_limit); 289 290 /* Start the SYN cookie reseeder callout. */ 291 callout_init(&V_tcp_syncache.secret.reseed, 1); 292 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 293 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 294 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 295 syncookie_reseed, &V_tcp_syncache); 296} 297 298#ifdef VIMAGE 299void 300syncache_destroy(void) 301{ 302 struct syncache_head *sch; 303 struct syncache *sc, *nsc; 304 int i; 305 306 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 307 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 308 309 sch = &V_tcp_syncache.hashbase[i]; 310 callout_drain(&sch->sch_timer); 311 312 SCH_LOCK(sch); 313 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 314 syncache_drop(sc, sch); 315 SCH_UNLOCK(sch); 316 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 317 ("%s: sch->sch_bucket not empty", __func__)); 318 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 319 __func__, sch->sch_length)); 320 mtx_destroy(&sch->sch_mtx); 321 } 322 323 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 324 ("%s: cache_count not 0", __func__)); 325 326 /* Free the allocated global resources. */ 327 uma_zdestroy(V_tcp_syncache.zone); 328 free(V_tcp_syncache.hashbase, M_SYNCACHE); 329 330 callout_drain(&V_tcp_syncache.secret.reseed); 331} 332#endif 333 334static int 335syncache_sysctl_count(SYSCTL_HANDLER_ARGS) 336{ 337 int count; 338 339 count = uma_zone_get_cur(V_tcp_syncache.zone); 340 return (sysctl_handle_int(oidp, &count, 0, req)); 341} 342 343/* 344 * Inserts a syncache entry into the specified bucket row. 345 * Locks and unlocks the syncache_head autonomously. 346 */ 347static void 348syncache_insert(struct syncache *sc, struct syncache_head *sch) 349{ 350 struct syncache *sc2; 351 352 SCH_LOCK(sch); 353 354 /* 355 * Make sure that we don't overflow the per-bucket limit. 356 * If the bucket is full, toss the oldest element. 357 */ 358 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 359 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 360 ("sch->sch_length incorrect")); 361 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 362 sch->sch_last_overflow = time_uptime; 363 syncache_drop(sc2, sch); 364 TCPSTAT_INC(tcps_sc_bucketoverflow); 365 } 366 367 /* Put it into the bucket. */ 368 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 369 sch->sch_length++; 370 371#ifdef TCP_OFFLOAD 372 if (ADDED_BY_TOE(sc)) { 373 struct toedev *tod = sc->sc_tod; 374 375 tod->tod_syncache_added(tod, sc->sc_todctx); 376 } 377#endif 378 379 /* Reinitialize the bucket row's timer. */ 380 if (sch->sch_length == 1) 381 sch->sch_nextc = ticks + INT_MAX; 382 syncache_timeout(sc, sch, 1); 383 384 SCH_UNLOCK(sch); 385 386 TCPSTAT_INC(tcps_sc_added); 387} 388 389/* 390 * Remove and free entry from syncache bucket row. 391 * Expects locked syncache head. 392 */ 393static void 394syncache_drop(struct syncache *sc, struct syncache_head *sch) 395{ 396 397 SCH_LOCK_ASSERT(sch); 398 399 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 400 sch->sch_length--; 401 402#ifdef TCP_OFFLOAD 403 if (ADDED_BY_TOE(sc)) { 404 struct toedev *tod = sc->sc_tod; 405 406 tod->tod_syncache_removed(tod, sc->sc_todctx); 407 } 408#endif 409 410 syncache_free(sc); 411} 412 413/* 414 * Engage/reengage time on bucket row. 415 */ 416static void 417syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 418{ 419 sc->sc_rxttime = ticks + 420 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]); 421 sc->sc_rxmits++; 422 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 423 sch->sch_nextc = sc->sc_rxttime; 424 if (docallout) 425 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 426 syncache_timer, (void *)sch); 427 } 428} 429 430/* 431 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 432 * If we have retransmitted an entry the maximum number of times, expire it. 433 * One separate timer for each bucket row. 434 */ 435static void 436syncache_timer(void *xsch) 437{ 438 struct syncache_head *sch = (struct syncache_head *)xsch; 439 struct syncache *sc, *nsc; 440 int tick = ticks; 441 char *s; 442 443 CURVNET_SET(sch->sch_sc->vnet); 444 445 /* NB: syncache_head has already been locked by the callout. */ 446 SCH_LOCK_ASSERT(sch); 447 448 /* 449 * In the following cycle we may remove some entries and/or 450 * advance some timeouts, so re-initialize the bucket timer. 451 */ 452 sch->sch_nextc = tick + INT_MAX; 453 454 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 455 /* 456 * We do not check if the listen socket still exists 457 * and accept the case where the listen socket may be 458 * gone by the time we resend the SYN/ACK. We do 459 * not expect this to happens often. If it does, 460 * then the RST will be sent by the time the remote 461 * host does the SYN/ACK->ACK. 462 */ 463 if (TSTMP_GT(sc->sc_rxttime, tick)) { 464 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 465 sch->sch_nextc = sc->sc_rxttime; 466 continue; 467 } 468 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 469 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 470 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 471 "giving up and removing syncache entry\n", 472 s, __func__); 473 free(s, M_TCPLOG); 474 } 475 syncache_drop(sc, sch); 476 TCPSTAT_INC(tcps_sc_stale); 477 continue; 478 } 479 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 480 log(LOG_DEBUG, "%s; %s: Response timeout, " 481 "retransmitting (%u) SYN|ACK\n", 482 s, __func__, sc->sc_rxmits); 483 free(s, M_TCPLOG); 484 } 485 486 (void) syncache_respond(sc, NULL); 487 TCPSTAT_INC(tcps_sc_retransmitted); 488 syncache_timeout(sc, sch, 0); 489 } 490 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 491 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 492 syncache_timer, (void *)(sch)); 493 CURVNET_RESTORE(); 494} 495 496/* 497 * Find an entry in the syncache. 498 * Returns always with locked syncache_head plus a matching entry or NULL. 499 */ 500static struct syncache * 501syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 502{ 503 struct syncache *sc; 504 struct syncache_head *sch; 505 506#ifdef INET6 507 if (inc->inc_flags & INC_ISIPV6) { 508 sch = &V_tcp_syncache.hashbase[ 509 SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)]; 510 *schp = sch; 511 512 SCH_LOCK(sch); 513 514 /* Circle through bucket row to find matching entry. */ 515 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 516 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 517 return (sc); 518 } 519 } else 520#endif 521 { 522 sch = &V_tcp_syncache.hashbase[ 523 SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)]; 524 *schp = sch; 525 526 SCH_LOCK(sch); 527 528 /* Circle through bucket row to find matching entry. */ 529 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 530#ifdef INET6 531 if (sc->sc_inc.inc_flags & INC_ISIPV6) 532 continue; 533#endif 534 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 535 return (sc); 536 } 537 } 538 SCH_LOCK_ASSERT(*schp); 539 return (NULL); /* 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 */ 547void 548syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 549{ 550 struct syncache *sc; 551 struct syncache_head *sch; 552 char *s = NULL; 553 554 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 555 SCH_LOCK_ASSERT(sch); 556 557 /* 558 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 559 * See RFC 793 page 65, section SEGMENT ARRIVES. 560 */ 561 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 562 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 563 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 564 "FIN flag set, segment ignored\n", s, __func__); 565 TCPSTAT_INC(tcps_badrst); 566 goto done; 567 } 568 569 /* 570 * No corresponding connection was found in syncache. 571 * If syncookies are enabled and possibly exclusively 572 * used, or we are under memory pressure, a valid RST 573 * may not find a syncache entry. In that case we're 574 * done and no SYN|ACK retransmissions will happen. 575 * Otherwise the RST was misdirected or spoofed. 576 */ 577 if (sc == NULL) { 578 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 579 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 580 "syncache entry (possibly syncookie only), " 581 "segment ignored\n", s, __func__); 582 TCPSTAT_INC(tcps_badrst); 583 goto done; 584 } 585 586 /* 587 * If the RST bit is set, check the sequence number to see 588 * if this is a valid reset segment. 589 * RFC 793 page 37: 590 * In all states except SYN-SENT, all reset (RST) segments 591 * are validated by checking their SEQ-fields. A reset is 592 * valid if its sequence number is in the window. 593 * 594 * The sequence number in the reset segment is normally an 595 * echo of our outgoing acknowlegement numbers, but some hosts 596 * send a reset with the sequence number at the rightmost edge 597 * of our receive window, and we have to handle this case. 598 */ 599 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 600 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 601 syncache_drop(sc, sch); 602 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 603 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 604 "connection attempt aborted by remote endpoint\n", 605 s, __func__); 606 TCPSTAT_INC(tcps_sc_reset); 607 } else { 608 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 609 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 610 "IRS %u (+WND %u), segment ignored\n", 611 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 612 TCPSTAT_INC(tcps_badrst); 613 } 614 615done: 616 if (s != NULL) 617 free(s, M_TCPLOG); 618 SCH_UNLOCK(sch); 619} 620 621void 622syncache_badack(struct in_conninfo *inc) 623{ 624 struct syncache *sc; 625 struct syncache_head *sch; 626 627 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 628 SCH_LOCK_ASSERT(sch); 629 if (sc != NULL) { 630 syncache_drop(sc, sch); 631 TCPSTAT_INC(tcps_sc_badack); 632 } 633 SCH_UNLOCK(sch); 634} 635 636void 637syncache_unreach(struct in_conninfo *inc, struct tcphdr *th) 638{ 639 struct syncache *sc; 640 struct syncache_head *sch; 641 642 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 643 SCH_LOCK_ASSERT(sch); 644 if (sc == NULL) 645 goto done; 646 647 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 648 if (ntohl(th->th_seq) != sc->sc_iss) 649 goto done; 650 651 /* 652 * If we've rertransmitted 3 times and this is our second error, 653 * we remove the entry. Otherwise, we allow it to continue on. 654 * This prevents us from incorrectly nuking an entry during a 655 * spurious network outage. 656 * 657 * See tcp_notify(). 658 */ 659 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 660 sc->sc_flags |= SCF_UNREACH; 661 goto done; 662 } 663 syncache_drop(sc, sch); 664 TCPSTAT_INC(tcps_sc_unreach); 665done: 666 SCH_UNLOCK(sch); 667} 668 669/* 670 * Build a new TCP socket structure from a syncache entry. 671 * 672 * On success return the newly created socket with its underlying inp locked. 673 */ 674static struct socket * 675syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 676{ 677 struct inpcb *inp = NULL; 678 struct socket *so; 679 struct tcpcb *tp; 680 int error; 681 char *s; 682 683 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 684 685 /* 686 * Ok, create the full blown connection, and set things up 687 * as they would have been set up if we had created the 688 * connection when the SYN arrived. If we can't create 689 * the connection, abort it. 690 */ 691 so = sonewconn(lso, 0); 692 if (so == NULL) { 693 /* 694 * Drop the connection; we will either send a RST or 695 * have the peer retransmit its SYN again after its 696 * RTO and try again. 697 */ 698 TCPSTAT_INC(tcps_listendrop); 699 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 700 log(LOG_DEBUG, "%s; %s: Socket create failed " 701 "due to limits or memory shortage\n", 702 s, __func__); 703 free(s, M_TCPLOG); 704 } 705 goto abort2; 706 } 707#ifdef MAC 708 mac_socketpeer_set_from_mbuf(m, so); 709#endif 710 711 inp = sotoinpcb(so); 712 inp->inp_inc.inc_fibnum = so->so_fibnum; 713 INP_WLOCK(inp); 714 /* 715 * Exclusive pcbinfo lock is not required in syncache socket case even 716 * if two inpcb locks can be acquired simultaneously: 717 * - the inpcb in LISTEN state, 718 * - the newly created inp. 719 * 720 * In this case, an inp cannot be at same time in LISTEN state and 721 * just created by an accept() call. 722 */ 723 INP_HASH_WLOCK(&V_tcbinfo); 724 725 /* Insert new socket into PCB hash list. */ 726 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 727#ifdef INET6 728 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 729 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 730 } else { 731 inp->inp_vflag &= ~INP_IPV6; 732 inp->inp_vflag |= INP_IPV4; 733#endif 734 inp->inp_laddr = sc->sc_inc.inc_laddr; 735#ifdef INET6 736 } 737#endif 738 739 /* 740 * If there's an mbuf and it has a flowid, then let's initialise the 741 * inp with that particular flowid. 742 */ 743 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 744 inp->inp_flowid = m->m_pkthdr.flowid; 745 inp->inp_flowtype = M_HASHTYPE_GET(m); 746 } 747 748 /* 749 * Install in the reservation hash table for now, but don't yet 750 * install a connection group since the full 4-tuple isn't yet 751 * configured. 752 */ 753 inp->inp_lport = sc->sc_inc.inc_lport; 754 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 755 /* 756 * Undo the assignments above if we failed to 757 * put the PCB on the hash lists. 758 */ 759#ifdef INET6 760 if (sc->sc_inc.inc_flags & INC_ISIPV6) 761 inp->in6p_laddr = in6addr_any; 762 else 763#endif 764 inp->inp_laddr.s_addr = INADDR_ANY; 765 inp->inp_lport = 0; 766 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 767 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 768 "with error %i\n", 769 s, __func__, error); 770 free(s, M_TCPLOG); 771 } 772 INP_HASH_WUNLOCK(&V_tcbinfo); 773 goto abort; 774 } 775#ifdef IPSEC 776 /* Copy old policy into new socket's. */ 777 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp)) 778 printf("syncache_socket: could not copy policy\n"); 779#endif 780#ifdef INET6 781 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 782 struct inpcb *oinp = sotoinpcb(lso); 783 struct in6_addr laddr6; 784 struct sockaddr_in6 sin6; 785 /* 786 * Inherit socket options from the listening socket. 787 * Note that in6p_inputopts are not (and should not be) 788 * copied, since it stores previously received options and is 789 * used to detect if each new option is different than the 790 * previous one and hence should be passed to a user. 791 * If we copied in6p_inputopts, a user would not be able to 792 * receive options just after calling the accept system call. 793 */ 794 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 795 if (oinp->in6p_outputopts) 796 inp->in6p_outputopts = 797 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 798 799 sin6.sin6_family = AF_INET6; 800 sin6.sin6_len = sizeof(sin6); 801 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 802 sin6.sin6_port = sc->sc_inc.inc_fport; 803 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 804 laddr6 = inp->in6p_laddr; 805 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 806 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 807 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 808 thread0.td_ucred, m)) != 0) { 809 inp->in6p_laddr = laddr6; 810 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 811 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 812 "with error %i\n", 813 s, __func__, error); 814 free(s, M_TCPLOG); 815 } 816 INP_HASH_WUNLOCK(&V_tcbinfo); 817 goto abort; 818 } 819 /* Override flowlabel from in6_pcbconnect. */ 820 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 821 inp->inp_flow |= sc->sc_flowlabel; 822 } 823#endif /* INET6 */ 824#if defined(INET) && defined(INET6) 825 else 826#endif 827#ifdef INET 828 { 829 struct in_addr laddr; 830 struct sockaddr_in sin; 831 832 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 833 834 if (inp->inp_options == NULL) { 835 inp->inp_options = sc->sc_ipopts; 836 sc->sc_ipopts = NULL; 837 } 838 839 sin.sin_family = AF_INET; 840 sin.sin_len = sizeof(sin); 841 sin.sin_addr = sc->sc_inc.inc_faddr; 842 sin.sin_port = sc->sc_inc.inc_fport; 843 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 844 laddr = inp->inp_laddr; 845 if (inp->inp_laddr.s_addr == INADDR_ANY) 846 inp->inp_laddr = sc->sc_inc.inc_laddr; 847 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 848 thread0.td_ucred, m)) != 0) { 849 inp->inp_laddr = laddr; 850 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 851 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 852 "with error %i\n", 853 s, __func__, error); 854 free(s, M_TCPLOG); 855 } 856 INP_HASH_WUNLOCK(&V_tcbinfo); 857 goto abort; 858 } 859 } 860#endif /* INET */ 861 INP_HASH_WUNLOCK(&V_tcbinfo); 862 tp = intotcpcb(inp); 863 tcp_state_change(tp, TCPS_SYN_RECEIVED); 864 tp->iss = sc->sc_iss; 865 tp->irs = sc->sc_irs; 866 tcp_rcvseqinit(tp); 867 tcp_sendseqinit(tp); 868 tp->snd_wl1 = sc->sc_irs; 869 tp->snd_max = tp->iss + 1; 870 tp->snd_nxt = tp->iss + 1; 871 tp->rcv_up = sc->sc_irs + 1; 872 tp->rcv_wnd = sc->sc_wnd; 873 tp->rcv_adv += tp->rcv_wnd; 874 tp->last_ack_sent = tp->rcv_nxt; 875 876 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 877 if (sc->sc_flags & SCF_NOOPT) 878 tp->t_flags |= TF_NOOPT; 879 else { 880 if (sc->sc_flags & SCF_WINSCALE) { 881 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 882 tp->snd_scale = sc->sc_requested_s_scale; 883 tp->request_r_scale = sc->sc_requested_r_scale; 884 } 885 if (sc->sc_flags & SCF_TIMESTAMP) { 886 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 887 tp->ts_recent = sc->sc_tsreflect; 888 tp->ts_recent_age = tcp_ts_getticks(); 889 tp->ts_offset = sc->sc_tsoff; 890 } 891#ifdef TCP_SIGNATURE 892 if (sc->sc_flags & SCF_SIGNATURE) 893 tp->t_flags |= TF_SIGNATURE; 894#endif 895 if (sc->sc_flags & SCF_SACK) 896 tp->t_flags |= TF_SACK_PERMIT; 897 } 898 899 if (sc->sc_flags & SCF_ECN) 900 tp->t_flags |= TF_ECN_PERMIT; 901 902 /* 903 * Set up MSS and get cached values from tcp_hostcache. 904 * This might overwrite some of the defaults we just set. 905 */ 906 tcp_mss(tp, sc->sc_peer_mss); 907 908 /* 909 * If the SYN,ACK was retransmitted, indicate that CWND to be 910 * limited to one segment in cc_conn_init(). 911 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 912 */ 913 if (sc->sc_rxmits > 1) 914 tp->snd_cwnd = 1; 915 916#ifdef TCP_OFFLOAD 917 /* 918 * Allow a TOE driver to install its hooks. Note that we hold the 919 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 920 * new connection before the TOE driver has done its thing. 921 */ 922 if (ADDED_BY_TOE(sc)) { 923 struct toedev *tod = sc->sc_tod; 924 925 tod->tod_offload_socket(tod, sc->sc_todctx, so); 926 } 927#endif 928 /* 929 * Copy and activate timers. 930 */ 931 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 932 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 933 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 934 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 935 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 936 937 TCPSTAT_INC(tcps_accepts); 938 return (so); 939 940abort: 941 INP_WUNLOCK(inp); 942abort2: 943 if (so != NULL) 944 soabort(so); 945 return (NULL); 946} 947 948/* 949 * This function gets called when we receive an ACK for a 950 * socket in the LISTEN state. We look up the connection 951 * in the syncache, and if its there, we pull it out of 952 * the cache and turn it into a full-blown connection in 953 * the SYN-RECEIVED state. 954 * 955 * On syncache_socket() success the newly created socket 956 * has its underlying inp locked. 957 */ 958int 959syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 960 struct socket **lsop, struct mbuf *m) 961{ 962 struct syncache *sc; 963 struct syncache_head *sch; 964 struct syncache scs; 965 char *s; 966 967 /* 968 * Global TCP locks are held because we manipulate the PCB lists 969 * and create a new socket. 970 */ 971 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 972 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 973 ("%s: can handle only ACK", __func__)); 974 975 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 976 SCH_LOCK_ASSERT(sch); 977 978#ifdef INVARIANTS 979 /* 980 * Test code for syncookies comparing the syncache stored 981 * values with the reconstructed values from the cookie. 982 */ 983 if (sc != NULL) 984 syncookie_cmp(inc, sch, sc, th, to, *lsop); 985#endif 986 987 if (sc == NULL) { 988 /* 989 * There is no syncache entry, so see if this ACK is 990 * a returning syncookie. To do this, first: 991 * A. Check if syncookies are used in case of syncache 992 * overflows 993 * B. See if this socket has had a syncache entry dropped in 994 * the recent past. We don't want to accept a bogus 995 * syncookie if we've never received a SYN or accept it 996 * twice. 997 * C. check that the syncookie is valid. If it is, then 998 * cobble up a fake syncache entry, and return. 999 */ 1000 if (!V_tcp_syncookies) { 1001 SCH_UNLOCK(sch); 1002 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1003 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1004 "segment rejected (syncookies disabled)\n", 1005 s, __func__); 1006 goto failed; 1007 } 1008 if (!V_tcp_syncookiesonly && 1009 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1010 SCH_UNLOCK(sch); 1011 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1012 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1013 "segment rejected (no syncache entry)\n", 1014 s, __func__); 1015 goto failed; 1016 } 1017 bzero(&scs, sizeof(scs)); 1018 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1019 SCH_UNLOCK(sch); 1020 if (sc == NULL) { 1021 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1022 log(LOG_DEBUG, "%s; %s: Segment failed " 1023 "SYNCOOKIE authentication, segment rejected " 1024 "(probably spoofed)\n", s, __func__); 1025 goto failed; 1026 } 1027 } else { 1028 /* Pull out the entry to unlock the bucket row. */ 1029 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1030 sch->sch_length--; 1031#ifdef TCP_OFFLOAD 1032 if (ADDED_BY_TOE(sc)) { 1033 struct toedev *tod = sc->sc_tod; 1034 1035 tod->tod_syncache_removed(tod, sc->sc_todctx); 1036 } 1037#endif 1038 SCH_UNLOCK(sch); 1039 } 1040 1041 /* 1042 * Segment validation: 1043 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1044 */ 1045 if (th->th_ack != sc->sc_iss + 1) { 1046 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1047 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1048 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1049 goto failed; 1050 } 1051 1052 /* 1053 * The SEQ must fall in the window starting at the received 1054 * initial receive sequence number + 1 (the SYN). 1055 */ 1056 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1057 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1058 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1059 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1060 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1061 goto failed; 1062 } 1063 1064 /* 1065 * If timestamps were not negotiated during SYN/ACK they 1066 * must not appear on any segment during this session. 1067 */ 1068 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1069 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1070 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1071 "segment rejected\n", s, __func__); 1072 goto failed; 1073 } 1074 1075 /* 1076 * If timestamps were negotiated during SYN/ACK they should 1077 * appear on every segment during this session. 1078 * XXXAO: This is only informal as there have been unverified 1079 * reports of non-compliants stacks. 1080 */ 1081 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1082 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1083 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1084 "no action\n", s, __func__); 1085 free(s, M_TCPLOG); 1086 s = NULL; 1087 } 1088 } 1089 1090 /* 1091 * If timestamps were negotiated the reflected timestamp 1092 * must be equal to what we actually sent in the SYN|ACK. 1093 */ 1094 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) { 1095 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1096 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1097 "segment rejected\n", 1098 s, __func__, to->to_tsecr, sc->sc_ts); 1099 goto failed; 1100 } 1101 1102 *lsop = syncache_socket(sc, *lsop, m); 1103 1104 if (*lsop == NULL) 1105 TCPSTAT_INC(tcps_sc_aborted); 1106 else 1107 TCPSTAT_INC(tcps_sc_completed); 1108 1109/* how do we find the inp for the new socket? */ 1110 if (sc != &scs) 1111 syncache_free(sc); 1112 return (1); 1113failed: 1114 if (sc != NULL && sc != &scs) 1115 syncache_free(sc); 1116 if (s != NULL) 1117 free(s, M_TCPLOG); 1118 *lsop = NULL; 1119 return (0); 1120} 1121 1122#ifdef TCP_RFC7413 1123static void 1124syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1125 uint64_t response_cookie) 1126{ 1127 struct inpcb *inp; 1128 struct tcpcb *tp; 1129 unsigned int *pending_counter; 1130 1131 /* 1132 * Global TCP locks are held because we manipulate the PCB lists 1133 * and create a new socket. 1134 */ 1135 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1136 1137 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1138 *lsop = syncache_socket(sc, *lsop, m); 1139 if (*lsop == NULL) { 1140 TCPSTAT_INC(tcps_sc_aborted); 1141 atomic_subtract_int(pending_counter, 1); 1142 } else { 1143 inp = sotoinpcb(*lsop); 1144 tp = intotcpcb(inp); 1145 tp->t_flags |= TF_FASTOPEN; 1146 tp->t_tfo_cookie = response_cookie; 1147 tp->snd_max = tp->iss; 1148 tp->snd_nxt = tp->iss; 1149 tp->t_tfo_pending = pending_counter; 1150 TCPSTAT_INC(tcps_sc_completed); 1151 } 1152} 1153#endif /* TCP_RFC7413 */ 1154 1155/* 1156 * Given a LISTEN socket and an inbound SYN request, add 1157 * this to the syn cache, and send back a segment: 1158 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1159 * to the source. 1160 * 1161 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1162 * Doing so would require that we hold onto the data and deliver it 1163 * to the application. However, if we are the target of a SYN-flood 1164 * DoS attack, an attacker could send data which would eventually 1165 * consume all available buffer space if it were ACKed. By not ACKing 1166 * the data, we avoid this DoS scenario. 1167 * 1168 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1169 * cookie is processed, V_tcp_fastopen_enabled set to true, and the 1170 * TCP_FASTOPEN socket option is set. In this case, a new socket is created 1171 * and returned via lsop, the mbuf is not freed so that tcp_input() can 1172 * queue its data to the socket, and 1 is returned to indicate the 1173 * TFO-socket-creation path was taken. 1174 */ 1175int 1176syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1177 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1178 void *todctx) 1179{ 1180 struct tcpcb *tp; 1181 struct socket *so; 1182 struct syncache *sc = NULL; 1183 struct syncache_head *sch; 1184 struct mbuf *ipopts = NULL; 1185 u_int ltflags; 1186 int win, sb_hiwat, ip_ttl, ip_tos; 1187 char *s; 1188 int rv = 0; 1189#ifdef INET6 1190 int autoflowlabel = 0; 1191#endif 1192#ifdef MAC 1193 struct label *maclabel; 1194#endif 1195 struct syncache scs; 1196 struct ucred *cred; 1197#ifdef TCP_RFC7413 1198 uint64_t tfo_response_cookie; 1199 int tfo_cookie_valid = 0; 1200 int tfo_response_cookie_valid = 0; 1201#endif 1202 1203 INP_WLOCK_ASSERT(inp); /* listen socket */ 1204 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1205 ("%s: unexpected tcp flags", __func__)); 1206 1207 /* 1208 * Combine all so/tp operations very early to drop the INP lock as 1209 * soon as possible. 1210 */ 1211 so = *lsop; 1212 tp = sototcpcb(so); 1213 cred = crhold(so->so_cred); 1214 1215#ifdef INET6 1216 if ((inc->inc_flags & INC_ISIPV6) && 1217 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1218 autoflowlabel = 1; 1219#endif 1220 ip_ttl = inp->inp_ip_ttl; 1221 ip_tos = inp->inp_ip_tos; 1222 win = sbspace(&so->so_rcv); 1223 sb_hiwat = so->so_rcv.sb_hiwat; 1224 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1225 1226#ifdef TCP_RFC7413 1227 if (V_tcp_fastopen_enabled && (tp->t_flags & TF_FASTOPEN) && 1228 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) { 1229 /* 1230 * Limit the number of pending TFO connections to 1231 * approximately half of the queue limit. This prevents TFO 1232 * SYN floods from starving the service by filling the 1233 * listen queue with bogus TFO connections. 1234 */ 1235 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1236 (so->so_qlimit / 2)) { 1237 int result; 1238 1239 result = tcp_fastopen_check_cookie(inc, 1240 to->to_tfo_cookie, to->to_tfo_len, 1241 &tfo_response_cookie); 1242 tfo_cookie_valid = (result > 0); 1243 tfo_response_cookie_valid = (result >= 0); 1244 } else 1245 atomic_subtract_int(tp->t_tfo_pending, 1); 1246 } 1247#endif 1248 1249 /* By the time we drop the lock these should no longer be used. */ 1250 so = NULL; 1251 tp = NULL; 1252 1253#ifdef MAC 1254 if (mac_syncache_init(&maclabel) != 0) { 1255 INP_WUNLOCK(inp); 1256 goto done; 1257 } else 1258 mac_syncache_create(maclabel, inp); 1259#endif 1260#ifdef TCP_RFC7413 1261 if (!tfo_cookie_valid) 1262#endif 1263 INP_WUNLOCK(inp); 1264 1265 /* 1266 * Remember the IP options, if any. 1267 */ 1268#ifdef INET6 1269 if (!(inc->inc_flags & INC_ISIPV6)) 1270#endif 1271#ifdef INET 1272 ipopts = (m) ? ip_srcroute(m) : NULL; 1273#else 1274 ipopts = NULL; 1275#endif 1276 1277 /* 1278 * See if we already have an entry for this connection. 1279 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1280 * 1281 * XXX: should the syncache be re-initialized with the contents 1282 * of the new SYN here (which may have different options?) 1283 * 1284 * XXX: We do not check the sequence number to see if this is a 1285 * real retransmit or a new connection attempt. The question is 1286 * how to handle such a case; either ignore it as spoofed, or 1287 * drop the current entry and create a new one? 1288 */ 1289 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1290 SCH_LOCK_ASSERT(sch); 1291 if (sc != NULL) { 1292#ifdef TCP_RFC7413 1293 if (tfo_cookie_valid) 1294 INP_WUNLOCK(inp); 1295#endif 1296 TCPSTAT_INC(tcps_sc_dupsyn); 1297 if (ipopts) { 1298 /* 1299 * If we were remembering a previous source route, 1300 * forget it and use the new one we've been given. 1301 */ 1302 if (sc->sc_ipopts) 1303 (void) m_free(sc->sc_ipopts); 1304 sc->sc_ipopts = ipopts; 1305 } 1306 /* 1307 * Update timestamp if present. 1308 */ 1309 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1310 sc->sc_tsreflect = to->to_tsval; 1311 else 1312 sc->sc_flags &= ~SCF_TIMESTAMP; 1313#ifdef MAC 1314 /* 1315 * Since we have already unconditionally allocated label 1316 * storage, free it up. The syncache entry will already 1317 * have an initialized label we can use. 1318 */ 1319 mac_syncache_destroy(&maclabel); 1320#endif 1321 /* Retransmit SYN|ACK and reset retransmit count. */ 1322 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1323 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1324 "resetting timer and retransmitting SYN|ACK\n", 1325 s, __func__); 1326 free(s, M_TCPLOG); 1327 } 1328 if (syncache_respond(sc, m) == 0) { 1329 sc->sc_rxmits = 0; 1330 syncache_timeout(sc, sch, 1); 1331 TCPSTAT_INC(tcps_sndacks); 1332 TCPSTAT_INC(tcps_sndtotal); 1333 } 1334 SCH_UNLOCK(sch); 1335 goto done; 1336 } 1337 1338#ifdef TCP_RFC7413 1339 if (tfo_cookie_valid) { 1340 bzero(&scs, sizeof(scs)); 1341 sc = &scs; 1342 goto skip_alloc; 1343 } 1344#endif 1345 1346 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1347 if (sc == NULL) { 1348 /* 1349 * The zone allocator couldn't provide more entries. 1350 * Treat this as if the cache was full; drop the oldest 1351 * entry and insert the new one. 1352 */ 1353 TCPSTAT_INC(tcps_sc_zonefail); 1354 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1355 sch->sch_last_overflow = time_uptime; 1356 syncache_drop(sc, sch); 1357 } 1358 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1359 if (sc == NULL) { 1360 if (V_tcp_syncookies) { 1361 bzero(&scs, sizeof(scs)); 1362 sc = &scs; 1363 } else { 1364 SCH_UNLOCK(sch); 1365 if (ipopts) 1366 (void) m_free(ipopts); 1367 goto done; 1368 } 1369 } 1370 } 1371 1372#ifdef TCP_RFC7413 1373skip_alloc: 1374 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1375 sc->sc_tfo_cookie = &tfo_response_cookie; 1376#endif 1377 1378 /* 1379 * Fill in the syncache values. 1380 */ 1381#ifdef MAC 1382 sc->sc_label = maclabel; 1383#endif 1384 sc->sc_cred = cred; 1385 cred = NULL; 1386 sc->sc_ipopts = ipopts; 1387 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1388#ifdef INET6 1389 if (!(inc->inc_flags & INC_ISIPV6)) 1390#endif 1391 { 1392 sc->sc_ip_tos = ip_tos; 1393 sc->sc_ip_ttl = ip_ttl; 1394 } 1395#ifdef TCP_OFFLOAD 1396 sc->sc_tod = tod; 1397 sc->sc_todctx = todctx; 1398#endif 1399 sc->sc_irs = th->th_seq; 1400 sc->sc_iss = arc4random(); 1401 sc->sc_flags = 0; 1402 sc->sc_flowlabel = 0; 1403 1404 /* 1405 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1406 * win was derived from socket earlier in the function. 1407 */ 1408 win = imax(win, 0); 1409 win = imin(win, TCP_MAXWIN); 1410 sc->sc_wnd = win; 1411 1412 if (V_tcp_do_rfc1323) { 1413 /* 1414 * A timestamp received in a SYN makes 1415 * it ok to send timestamp requests and replies. 1416 */ 1417 if (to->to_flags & TOF_TS) { 1418 sc->sc_tsreflect = to->to_tsval; 1419 sc->sc_ts = tcp_ts_getticks(); 1420 sc->sc_flags |= SCF_TIMESTAMP; 1421 } 1422 if (to->to_flags & TOF_SCALE) { 1423 int wscale = 0; 1424 1425 /* 1426 * Pick the smallest possible scaling factor that 1427 * will still allow us to scale up to sb_max, aka 1428 * kern.ipc.maxsockbuf. 1429 * 1430 * We do this because there are broken firewalls that 1431 * will corrupt the window scale option, leading to 1432 * the other endpoint believing that our advertised 1433 * window is unscaled. At scale factors larger than 1434 * 5 the unscaled window will drop below 1500 bytes, 1435 * leading to serious problems when traversing these 1436 * broken firewalls. 1437 * 1438 * With the default maxsockbuf of 256K, a scale factor 1439 * of 3 will be chosen by this algorithm. Those who 1440 * choose a larger maxsockbuf should watch out 1441 * for the compatiblity problems mentioned above. 1442 * 1443 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1444 * or <SYN,ACK>) segment itself is never scaled. 1445 */ 1446 while (wscale < TCP_MAX_WINSHIFT && 1447 (TCP_MAXWIN << wscale) < sb_max) 1448 wscale++; 1449 sc->sc_requested_r_scale = wscale; 1450 sc->sc_requested_s_scale = to->to_wscale; 1451 sc->sc_flags |= SCF_WINSCALE; 1452 } 1453 } 1454#ifdef TCP_SIGNATURE 1455 /* 1456 * If listening socket requested TCP digests, and received SYN 1457 * contains the option, flag this in the syncache so that 1458 * syncache_respond() will do the right thing with the SYN+ACK. 1459 * XXX: Currently we always record the option by default and will 1460 * attempt to use it in syncache_respond(). 1461 */ 1462 if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE) 1463 sc->sc_flags |= SCF_SIGNATURE; 1464#endif 1465 if (to->to_flags & TOF_SACKPERM) 1466 sc->sc_flags |= SCF_SACK; 1467 if (to->to_flags & TOF_MSS) 1468 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1469 if (ltflags & TF_NOOPT) 1470 sc->sc_flags |= SCF_NOOPT; 1471 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1472 sc->sc_flags |= SCF_ECN; 1473 1474 if (V_tcp_syncookies) 1475 sc->sc_iss = syncookie_generate(sch, sc); 1476#ifdef INET6 1477 if (autoflowlabel) { 1478 if (V_tcp_syncookies) 1479 sc->sc_flowlabel = sc->sc_iss; 1480 else 1481 sc->sc_flowlabel = ip6_randomflowlabel(); 1482 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1483 } 1484#endif 1485 SCH_UNLOCK(sch); 1486 1487#ifdef TCP_RFC7413 1488 if (tfo_cookie_valid) { 1489 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1490 /* INP_WUNLOCK(inp) will be performed by the called */ 1491 rv = 1; 1492 goto tfo_done; 1493 } 1494#endif 1495 1496 /* 1497 * Do a standard 3-way handshake. 1498 */ 1499 if (syncache_respond(sc, m) == 0) { 1500 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1501 syncache_free(sc); 1502 else if (sc != &scs) 1503 syncache_insert(sc, sch); /* locks and unlocks sch */ 1504 TCPSTAT_INC(tcps_sndacks); 1505 TCPSTAT_INC(tcps_sndtotal); 1506 } else { 1507 if (sc != &scs) 1508 syncache_free(sc); 1509 TCPSTAT_INC(tcps_sc_dropped); 1510 } 1511 1512done: 1513 if (m) { 1514 *lsop = NULL; 1515 m_freem(m); 1516 } 1517#ifdef TCP_RFC7413 1518tfo_done: 1519#endif 1520 if (cred != NULL) 1521 crfree(cred); 1522#ifdef MAC 1523 if (sc == &scs) 1524 mac_syncache_destroy(&maclabel); 1525#endif 1526 return (rv); 1527} 1528 1529/* 1530 * Send SYN|ACK to the peer. Either in response to the peer's SYN, 1531 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1532 */ 1533static int 1534syncache_respond(struct syncache *sc, const struct mbuf *m0) 1535{ 1536 struct ip *ip = NULL; 1537 struct mbuf *m; 1538 struct tcphdr *th = NULL; 1539 int optlen, error = 0; /* Make compiler happy */ 1540 u_int16_t hlen, tlen, mssopt; 1541 struct tcpopt to; 1542#ifdef INET6 1543 struct ip6_hdr *ip6 = NULL; 1544#endif 1545 1546 hlen = 1547#ifdef INET6 1548 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1549#endif 1550 sizeof(struct ip); 1551 tlen = hlen + sizeof(struct tcphdr); 1552 1553 /* Determine MSS we advertize to other end of connection. */ 1554 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); 1555 1556 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1557 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1558 ("syncache: mbuf too small")); 1559 1560 /* Create the IP+TCP header from scratch. */ 1561 m = m_gethdr(M_NOWAIT, MT_DATA); 1562 if (m == NULL) 1563 return (ENOBUFS); 1564#ifdef MAC 1565 mac_syncache_create_mbuf(sc->sc_label, m); 1566#endif 1567 m->m_data += max_linkhdr; 1568 m->m_len = tlen; 1569 m->m_pkthdr.len = tlen; 1570 m->m_pkthdr.rcvif = NULL; 1571 1572#ifdef INET6 1573 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1574 ip6 = mtod(m, struct ip6_hdr *); 1575 ip6->ip6_vfc = IPV6_VERSION; 1576 ip6->ip6_nxt = IPPROTO_TCP; 1577 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1578 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1579 ip6->ip6_plen = htons(tlen - hlen); 1580 /* ip6_hlim is set after checksum */ 1581 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1582 ip6->ip6_flow |= sc->sc_flowlabel; 1583 1584 th = (struct tcphdr *)(ip6 + 1); 1585 } 1586#endif 1587#if defined(INET6) && defined(INET) 1588 else 1589#endif 1590#ifdef INET 1591 { 1592 ip = mtod(m, struct ip *); 1593 ip->ip_v = IPVERSION; 1594 ip->ip_hl = sizeof(struct ip) >> 2; 1595 ip->ip_len = htons(tlen); 1596 ip->ip_id = 0; 1597 ip->ip_off = 0; 1598 ip->ip_sum = 0; 1599 ip->ip_p = IPPROTO_TCP; 1600 ip->ip_src = sc->sc_inc.inc_laddr; 1601 ip->ip_dst = sc->sc_inc.inc_faddr; 1602 ip->ip_ttl = sc->sc_ip_ttl; 1603 ip->ip_tos = sc->sc_ip_tos; 1604 1605 /* 1606 * See if we should do MTU discovery. Route lookups are 1607 * expensive, so we will only unset the DF bit if: 1608 * 1609 * 1) path_mtu_discovery is disabled 1610 * 2) the SCF_UNREACH flag has been set 1611 */ 1612 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1613 ip->ip_off |= htons(IP_DF); 1614 1615 th = (struct tcphdr *)(ip + 1); 1616 } 1617#endif /* INET */ 1618 th->th_sport = sc->sc_inc.inc_lport; 1619 th->th_dport = sc->sc_inc.inc_fport; 1620 1621 th->th_seq = htonl(sc->sc_iss); 1622 th->th_ack = htonl(sc->sc_irs + 1); 1623 th->th_off = sizeof(struct tcphdr) >> 2; 1624 th->th_x2 = 0; 1625 th->th_flags = TH_SYN|TH_ACK; 1626 th->th_win = htons(sc->sc_wnd); 1627 th->th_urp = 0; 1628 1629 if (sc->sc_flags & SCF_ECN) { 1630 th->th_flags |= TH_ECE; 1631 TCPSTAT_INC(tcps_ecn_shs); 1632 } 1633 1634 /* Tack on the TCP options. */ 1635 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1636 to.to_flags = 0; 1637 1638 to.to_mss = mssopt; 1639 to.to_flags = TOF_MSS; 1640 if (sc->sc_flags & SCF_WINSCALE) { 1641 to.to_wscale = sc->sc_requested_r_scale; 1642 to.to_flags |= TOF_SCALE; 1643 } 1644 if (sc->sc_flags & SCF_TIMESTAMP) { 1645 /* Virgin timestamp or TCP cookie enhanced one. */ 1646 to.to_tsval = sc->sc_ts; 1647 to.to_tsecr = sc->sc_tsreflect; 1648 to.to_flags |= TOF_TS; 1649 } 1650 if (sc->sc_flags & SCF_SACK) 1651 to.to_flags |= TOF_SACKPERM; 1652#ifdef TCP_SIGNATURE 1653 if (sc->sc_flags & SCF_SIGNATURE) 1654 to.to_flags |= TOF_SIGNATURE; 1655#endif 1656 1657#ifdef TCP_RFC7413 1658 if (sc->sc_tfo_cookie) { 1659 to.to_flags |= TOF_FASTOPEN; 1660 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1661 to.to_tfo_cookie = sc->sc_tfo_cookie; 1662 /* don't send cookie again when retransmitting response */ 1663 sc->sc_tfo_cookie = NULL; 1664 } 1665#endif 1666 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1667 1668 /* Adjust headers by option size. */ 1669 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1670 m->m_len += optlen; 1671 m->m_pkthdr.len += optlen; 1672 1673#ifdef TCP_SIGNATURE 1674 if (sc->sc_flags & SCF_SIGNATURE) 1675 tcp_signature_compute(m, 0, 0, optlen, 1676 to.to_signature, IPSEC_DIR_OUTBOUND); 1677#endif 1678#ifdef INET6 1679 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1680 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1681 else 1682#endif 1683 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1684 } else 1685 optlen = 0; 1686 1687 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1688 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1689 /* 1690 * If we have peer's SYN and it has a flowid, then let's assign it to 1691 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1692 * to SYN|ACK due to lack of inp here. 1693 */ 1694 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1695 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1696 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1697 } 1698#ifdef INET6 1699 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1700 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1701 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1702 IPPROTO_TCP, 0); 1703 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1704#ifdef TCP_OFFLOAD 1705 if (ADDED_BY_TOE(sc)) { 1706 struct toedev *tod = sc->sc_tod; 1707 1708 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1709 1710 return (error); 1711 } 1712#endif 1713 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1714 } 1715#endif 1716#if defined(INET6) && defined(INET) 1717 else 1718#endif 1719#ifdef INET 1720 { 1721 m->m_pkthdr.csum_flags = CSUM_TCP; 1722 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1723 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1724#ifdef TCP_OFFLOAD 1725 if (ADDED_BY_TOE(sc)) { 1726 struct toedev *tod = sc->sc_tod; 1727 1728 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1729 1730 return (error); 1731 } 1732#endif 1733 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1734 } 1735#endif 1736 return (error); 1737} 1738 1739/* 1740 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1741 * that exceed the capacity of the syncache by avoiding the storage of any 1742 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1743 * attacks where the attacker does not have access to our responses. 1744 * 1745 * Syncookies encode and include all necessary information about the 1746 * connection setup within the SYN|ACK that we send back. That way we 1747 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1748 * (if ever). Normally the syncache and syncookies are running in parallel 1749 * with the latter taking over when the former is exhausted. When matching 1750 * syncache entry is found the syncookie is ignored. 1751 * 1752 * The only reliable information persisting the 3WHS is our inital sequence 1753 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1754 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1755 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1756 * returns and signifies a legitimate connection if it matches the ACK. 1757 * 1758 * The available space of 32 bits to store the hash and to encode the SYN 1759 * option information is very tight and we should have at least 24 bits for 1760 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1761 * 1762 * SYN option information we have to encode to fully restore a connection: 1763 * MSS: is imporant to chose an optimal segment size to avoid IP level 1764 * fragmentation along the path. The common MSS values can be encoded 1765 * in a 3-bit table. Uncommon values are captured by the next lower value 1766 * in the table leading to a slight increase in packetization overhead. 1767 * WSCALE: is necessary to allow large windows to be used for high delay- 1768 * bandwidth product links. Not scaling the window when it was initially 1769 * negotiated is bad for performance as lack of scaling further decreases 1770 * the apparent available send window. We only need to encode the WSCALE 1771 * we received from the remote end. Our end can be recalculated at any 1772 * time. The common WSCALE values can be encoded in a 3-bit table. 1773 * Uncommon values are captured by the next lower value in the table 1774 * making us under-estimate the available window size halving our 1775 * theoretically possible maximum throughput for that connection. 1776 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1777 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1778 * that are included in all segments on a connection. We enable them when 1779 * the ACK has them. 1780 * 1781 * Security of syncookies and attack vectors: 1782 * 1783 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1784 * together with the gloabl secret to make it unique per connection attempt. 1785 * Thus any change of any of those parameters results in a different MAC output 1786 * in an unpredictable way unless a collision is encountered. 24 bits of the 1787 * MAC are embedded into the ISS. 1788 * 1789 * To prevent replay attacks two rotating global secrets are updated with a 1790 * new random value every 15 seconds. The life-time of a syncookie is thus 1791 * 15-30 seconds. 1792 * 1793 * Vector 1: Attacking the secret. This requires finding a weakness in the 1794 * MAC itself or the way it is used here. The attacker can do a chosen plain 1795 * text attack by varying and testing the all parameters under his control. 1796 * The strength depends on the size and randomness of the secret, and the 1797 * cryptographic security of the MAC function. Due to the constant updating 1798 * of the secret the attacker has at most 29.999 seconds to find the secret 1799 * and launch spoofed connections. After that he has to start all over again. 1800 * 1801 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1802 * size an average of 4,823 attempts are required for a 50% chance of success 1803 * to spoof a single syncookie (birthday collision paradox). However the 1804 * attacker is blind and doesn't know if one of his attempts succeeded unless 1805 * he has a side channel to interfere success from. A single connection setup 1806 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1807 * This many attempts are required for each one blind spoofed connection. For 1808 * every additional spoofed connection he has to launch another N attempts. 1809 * Thus for a sustained rate 100 spoofed connections per second approximately 1810 * 1,800,000 packets per second would have to be sent. 1811 * 1812 * NB: The MAC function should be fast so that it doesn't become a CPU 1813 * exhaustion attack vector itself. 1814 * 1815 * References: 1816 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1817 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1818 * http://cr.yp.to/syncookies.html (overview) 1819 * http://cr.yp.to/syncookies/archive (details) 1820 * 1821 * 1822 * Schematic construction of a syncookie enabled Initial Sequence Number: 1823 * 0 1 2 3 1824 * 12345678901234567890123456789012 1825 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1826 * 1827 * x 24 MAC (truncated) 1828 * W 3 Send Window Scale index 1829 * M 3 MSS index 1830 * S 1 SACK permitted 1831 * P 1 Odd/even secret 1832 */ 1833 1834/* 1835 * Distribution and probability of certain MSS values. Those in between are 1836 * rounded down to the next lower one. 1837 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1838 * .2% .3% 5% 7% 7% 20% 15% 45% 1839 */ 1840static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1841 1842/* 1843 * Distribution and probability of certain WSCALE values. We have to map the 1844 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1845 * bits based on prevalence of certain values. Where we don't have an exact 1846 * match for are rounded down to the next lower one letting us under-estimate 1847 * the true available window. At the moment this would happen only for the 1848 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1849 * and window size). The absence of the WSCALE option (no scaling in either 1850 * direction) is encoded with index zero. 1851 * [WSCALE values histograms, Allman, 2012] 1852 * X 10 10 35 5 6 14 10% by host 1853 * X 11 4 5 5 18 49 3% by connections 1854 */ 1855static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1856 1857/* 1858 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1859 * and good cryptographic properties. 1860 */ 1861static uint32_t 1862syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1863 uint8_t *secbits, uintptr_t secmod) 1864{ 1865 SIPHASH_CTX ctx; 1866 uint32_t siphash[2]; 1867 1868 SipHash24_Init(&ctx); 1869 SipHash_SetKey(&ctx, secbits); 1870 switch (inc->inc_flags & INC_ISIPV6) { 1871#ifdef INET 1872 case 0: 1873 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1874 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1875 break; 1876#endif 1877#ifdef INET6 1878 case INC_ISIPV6: 1879 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1880 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1881 break; 1882#endif 1883 } 1884 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1885 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1886 SipHash_Update(&ctx, &irs, sizeof(irs)); 1887 SipHash_Update(&ctx, &flags, sizeof(flags)); 1888 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1889 SipHash_Final((u_int8_t *)&siphash, &ctx); 1890 1891 return (siphash[0] ^ siphash[1]); 1892} 1893 1894static tcp_seq 1895syncookie_generate(struct syncache_head *sch, struct syncache *sc) 1896{ 1897 u_int i, secbit, wscale; 1898 uint32_t iss, hash; 1899 uint8_t *secbits; 1900 union syncookie cookie; 1901 1902 SCH_LOCK_ASSERT(sch); 1903 1904 cookie.cookie = 0; 1905 1906 /* Map our computed MSS into the 3-bit index. */ 1907 for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1; 1908 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 1909 i--) 1910 ; 1911 cookie.flags.mss_idx = i; 1912 1913 /* 1914 * Map the send window scale into the 3-bit index but only if 1915 * the wscale option was received. 1916 */ 1917 if (sc->sc_flags & SCF_WINSCALE) { 1918 wscale = sc->sc_requested_s_scale; 1919 for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1; 1920 tcp_sc_wstab[i] > wscale && i > 0; 1921 i--) 1922 ; 1923 cookie.flags.wscale_idx = i; 1924 } 1925 1926 /* Can we do SACK? */ 1927 if (sc->sc_flags & SCF_SACK) 1928 cookie.flags.sack_ok = 1; 1929 1930 /* Which of the two secrets to use. */ 1931 secbit = sch->sch_sc->secret.oddeven & 0x1; 1932 cookie.flags.odd_even = secbit; 1933 1934 secbits = sch->sch_sc->secret.key[secbit]; 1935 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 1936 (uintptr_t)sch); 1937 1938 /* 1939 * Put the flags into the hash and XOR them to get better ISS number 1940 * variance. This doesn't enhance the cryptographic strength and is 1941 * done to prevent the 8 cookie bits from showing up directly on the 1942 * wire. 1943 */ 1944 iss = hash & ~0xff; 1945 iss |= cookie.cookie ^ (hash >> 24); 1946 1947 /* Randomize the timestamp. */ 1948 if (sc->sc_flags & SCF_TIMESTAMP) { 1949 sc->sc_ts = arc4random(); 1950 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 1951 } 1952 1953 TCPSTAT_INC(tcps_sc_sendcookie); 1954 return (iss); 1955} 1956 1957static struct syncache * 1958syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 1959 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 1960 struct socket *lso) 1961{ 1962 uint32_t hash; 1963 uint8_t *secbits; 1964 tcp_seq ack, seq; 1965 int wnd, wscale = 0; 1966 union syncookie cookie; 1967 1968 SCH_LOCK_ASSERT(sch); 1969 1970 /* 1971 * Pull information out of SYN-ACK/ACK and revert sequence number 1972 * advances. 1973 */ 1974 ack = th->th_ack - 1; 1975 seq = th->th_seq - 1; 1976 1977 /* 1978 * Unpack the flags containing enough information to restore the 1979 * connection. 1980 */ 1981 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 1982 1983 /* Which of the two secrets to use. */ 1984 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 1985 1986 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 1987 1988 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 1989 if ((ack & ~0xff) != (hash & ~0xff)) 1990 return (NULL); 1991 1992 /* Fill in the syncache values. */ 1993 sc->sc_flags = 0; 1994 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1995 sc->sc_ipopts = NULL; 1996 1997 sc->sc_irs = seq; 1998 sc->sc_iss = ack; 1999 2000 switch (inc->inc_flags & INC_ISIPV6) { 2001#ifdef INET 2002 case 0: 2003 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2004 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2005 break; 2006#endif 2007#ifdef INET6 2008 case INC_ISIPV6: 2009 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2010 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 2011 break; 2012#endif 2013 } 2014 2015 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2016 2017 /* We can simply recompute receive window scale we sent earlier. */ 2018 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2019 wscale++; 2020 2021 /* Only use wscale if it was enabled in the orignal SYN. */ 2022 if (cookie.flags.wscale_idx > 0) { 2023 sc->sc_requested_r_scale = wscale; 2024 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2025 sc->sc_flags |= SCF_WINSCALE; 2026 } 2027 2028 wnd = sbspace(&lso->so_rcv); 2029 wnd = imax(wnd, 0); 2030 wnd = imin(wnd, TCP_MAXWIN); 2031 sc->sc_wnd = wnd; 2032 2033 if (cookie.flags.sack_ok) 2034 sc->sc_flags |= SCF_SACK; 2035 2036 if (to->to_flags & TOF_TS) { 2037 sc->sc_flags |= SCF_TIMESTAMP; 2038 sc->sc_tsreflect = to->to_tsval; 2039 sc->sc_ts = to->to_tsecr; 2040 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 2041 } 2042 2043 if (to->to_flags & TOF_SIGNATURE) 2044 sc->sc_flags |= SCF_SIGNATURE; 2045 2046 sc->sc_rxmits = 0; 2047 2048 TCPSTAT_INC(tcps_sc_recvcookie); 2049 return (sc); 2050} 2051 2052#ifdef INVARIANTS 2053static int 2054syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2055 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2056 struct socket *lso) 2057{ 2058 struct syncache scs, *scx; 2059 char *s; 2060 2061 bzero(&scs, sizeof(scs)); 2062 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2063 2064 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2065 return (0); 2066 2067 if (scx != NULL) { 2068 if (sc->sc_peer_mss != scx->sc_peer_mss) 2069 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2070 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2071 2072 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2073 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2074 s, __func__, sc->sc_requested_r_scale, 2075 scx->sc_requested_r_scale); 2076 2077 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2078 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2079 s, __func__, sc->sc_requested_s_scale, 2080 scx->sc_requested_s_scale); 2081 2082 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2083 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2084 } 2085 2086 if (s != NULL) 2087 free(s, M_TCPLOG); 2088 return (0); 2089} 2090#endif /* INVARIANTS */ 2091 2092static void 2093syncookie_reseed(void *arg) 2094{ 2095 struct tcp_syncache *sc = arg; 2096 uint8_t *secbits; 2097 int secbit; 2098 2099 /* 2100 * Reseeding the secret doesn't have to be protected by a lock. 2101 * It only must be ensured that the new random values are visible 2102 * to all CPUs in a SMP environment. The atomic with release 2103 * semantics ensures that. 2104 */ 2105 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2106 secbits = sc->secret.key[secbit]; 2107 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2108 atomic_add_rel_int(&sc->secret.oddeven, 1); 2109 2110 /* Reschedule ourself. */ 2111 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2112} 2113 2114/* 2115 * Returns the current number of syncache entries. This number 2116 * will probably change before you get around to calling 2117 * syncache_pcblist. 2118 */ 2119int 2120syncache_pcbcount(void) 2121{ 2122 struct syncache_head *sch; 2123 int count, i; 2124 2125 for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2126 /* No need to lock for a read. */ 2127 sch = &V_tcp_syncache.hashbase[i]; 2128 count += sch->sch_length; 2129 } 2130 return count; 2131} 2132 2133/* 2134 * Exports the syncache entries to userland so that netstat can display 2135 * them alongside the other sockets. This function is intended to be 2136 * called only from tcp_pcblist. 2137 * 2138 * Due to concurrency on an active system, the number of pcbs exported 2139 * may have no relation to max_pcbs. max_pcbs merely indicates the 2140 * amount of space the caller allocated for this function to use. 2141 */ 2142int 2143syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2144{ 2145 struct xtcpcb xt; 2146 struct syncache *sc; 2147 struct syncache_head *sch; 2148 int count, error, i; 2149 2150 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2151 sch = &V_tcp_syncache.hashbase[i]; 2152 SCH_LOCK(sch); 2153 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2154 if (count >= max_pcbs) { 2155 SCH_UNLOCK(sch); 2156 goto exit; 2157 } 2158 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2159 continue; 2160 bzero(&xt, sizeof(xt)); 2161 xt.xt_len = sizeof(xt); 2162 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2163 xt.xt_inp.inp_vflag = INP_IPV6; 2164 else 2165 xt.xt_inp.inp_vflag = INP_IPV4; 2166 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 2167 xt.xt_tp.t_inpcb = &xt.xt_inp; 2168 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 2169 xt.xt_socket.xso_protocol = IPPROTO_TCP; 2170 xt.xt_socket.xso_len = sizeof (struct xsocket); 2171 xt.xt_socket.so_type = SOCK_STREAM; 2172 xt.xt_socket.so_state = SS_ISCONNECTING; 2173 error = SYSCTL_OUT(req, &xt, sizeof xt); 2174 if (error) { 2175 SCH_UNLOCK(sch); 2176 goto exit; 2177 } 2178 count++; 2179 } 2180 SCH_UNLOCK(sch); 2181 } 2182exit: 2183 *pcbs_exported = count; 2184 return error; 2185} 2186