tcp_syncache.c revision 332238
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/11/sys/netinet/tcp_syncache.c 332238 2018-04-07 20:47:25Z tuexen $"); 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/hash.h> 44#include <sys/refcount.h> 45#include <sys/kernel.h> 46#include <sys/sysctl.h> 47#include <sys/limits.h> 48#include <sys/lock.h> 49#include <sys/mutex.h> 50#include <sys/malloc.h> 51#include <sys/mbuf.h> 52#include <sys/proc.h> /* for proc0 declaration */ 53#include <sys/random.h> 54#include <sys/socket.h> 55#include <sys/socketvar.h> 56#include <sys/syslog.h> 57#include <sys/ucred.h> 58 59#include <sys/md5.h> 60#include <crypto/siphash/siphash.h> 61 62#include <vm/uma.h> 63 64#include <net/if.h> 65#include <net/if_var.h> 66#include <net/route.h> 67#include <net/vnet.h> 68 69#include <netinet/in.h> 70#include <netinet/in_systm.h> 71#include <netinet/ip.h> 72#include <netinet/in_var.h> 73#include <netinet/in_pcb.h> 74#include <netinet/ip_var.h> 75#include <netinet/ip_options.h> 76#ifdef INET6 77#include <netinet/ip6.h> 78#include <netinet/icmp6.h> 79#include <netinet6/nd6.h> 80#include <netinet6/ip6_var.h> 81#include <netinet6/in6_pcb.h> 82#endif 83#include <netinet/tcp.h> 84#ifdef TCP_RFC7413 85#include <netinet/tcp_fastopen.h> 86#endif 87#include <netinet/tcp_fsm.h> 88#include <netinet/tcp_seq.h> 89#include <netinet/tcp_timer.h> 90#include <netinet/tcp_var.h> 91#include <netinet/tcp_syncache.h> 92#ifdef INET6 93#include <netinet6/tcp6_var.h> 94#endif 95#ifdef TCP_OFFLOAD 96#include <netinet/toecore.h> 97#endif 98 99#include <netipsec/ipsec_support.h> 100 101#include <machine/in_cksum.h> 102 103#include <security/mac/mac_framework.h> 104 105static VNET_DEFINE(int, tcp_syncookies) = 1; 106#define V_tcp_syncookies VNET(tcp_syncookies) 107SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 108 &VNET_NAME(tcp_syncookies), 0, 109 "Use TCP SYN cookies if the syncache overflows"); 110 111static VNET_DEFINE(int, tcp_syncookiesonly) = 0; 112#define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 113SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 114 &VNET_NAME(tcp_syncookiesonly), 0, 115 "Use only TCP SYN cookies"); 116 117#ifdef TCP_OFFLOAD 118#define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 119#endif 120 121static void syncache_drop(struct syncache *, struct syncache_head *); 122static void syncache_free(struct syncache *); 123static void syncache_insert(struct syncache *, struct syncache_head *); 124static int syncache_respond(struct syncache *, struct syncache_head *, int, 125 const struct mbuf *); 126static struct socket *syncache_socket(struct syncache *, struct socket *, 127 struct mbuf *m); 128static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 129 int docallout); 130static void syncache_timer(void *); 131 132static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 133 uint8_t *, uintptr_t); 134static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 135static struct syncache 136 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 137 struct syncache *, struct tcphdr *, struct tcpopt *, 138 struct socket *); 139static void syncookie_reseed(void *); 140#ifdef INVARIANTS 141static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 142 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 143 struct socket *lso); 144#endif 145 146/* 147 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 148 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds, 149 * the odds are that the user has given up attempting to connect by then. 150 */ 151#define SYNCACHE_MAXREXMTS 3 152 153/* Arbitrary values */ 154#define TCP_SYNCACHE_HASHSIZE 512 155#define TCP_SYNCACHE_BUCKETLIMIT 30 156 157static VNET_DEFINE(struct tcp_syncache, tcp_syncache); 158#define V_tcp_syncache VNET(tcp_syncache) 159 160static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, 161 "TCP SYN cache"); 162 163SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 164 &VNET_NAME(tcp_syncache.bucket_limit), 0, 165 "Per-bucket hash limit for syncache"); 166 167SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 168 &VNET_NAME(tcp_syncache.cache_limit), 0, 169 "Overall entry limit for syncache"); 170 171SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 172 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 173 174SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 175 &VNET_NAME(tcp_syncache.hashsize), 0, 176 "Size of TCP syncache hashtable"); 177 178SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW, 179 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 180 "Limit on SYN/ACK retransmissions"); 181 182VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 183SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 184 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 185 "Send reset on socket allocation failure"); 186 187static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 188 189#define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 190#define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 191#define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 192 193/* 194 * Requires the syncache entry to be already removed from the bucket list. 195 */ 196static void 197syncache_free(struct syncache *sc) 198{ 199 200 if (sc->sc_ipopts) 201 (void) m_free(sc->sc_ipopts); 202 if (sc->sc_cred) 203 crfree(sc->sc_cred); 204#ifdef MAC 205 mac_syncache_destroy(&sc->sc_label); 206#endif 207 208 uma_zfree(V_tcp_syncache.zone, sc); 209} 210 211void 212syncache_init(void) 213{ 214 int i; 215 216 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 217 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 218 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 219 V_tcp_syncache.hash_secret = arc4random(); 220 221 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 222 &V_tcp_syncache.hashsize); 223 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 224 &V_tcp_syncache.bucket_limit); 225 if (!powerof2(V_tcp_syncache.hashsize) || 226 V_tcp_syncache.hashsize == 0) { 227 printf("WARNING: syncache hash size is not a power of 2.\n"); 228 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 229 } 230 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 231 232 /* Set limits. */ 233 V_tcp_syncache.cache_limit = 234 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 235 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 236 &V_tcp_syncache.cache_limit); 237 238 /* Allocate the hash table. */ 239 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 240 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 241 242#ifdef VIMAGE 243 V_tcp_syncache.vnet = curvnet; 244#endif 245 246 /* Initialize the hash buckets. */ 247 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 248 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 249 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 250 NULL, MTX_DEF); 251 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 252 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 253 V_tcp_syncache.hashbase[i].sch_length = 0; 254 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 255 V_tcp_syncache.hashbase[i].sch_last_overflow = 256 -(SYNCOOKIE_LIFETIME + 1); 257 } 258 259 /* Create the syncache entry zone. */ 260 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 261 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 262 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 263 V_tcp_syncache.cache_limit); 264 265 /* Start the SYN cookie reseeder callout. */ 266 callout_init(&V_tcp_syncache.secret.reseed, 1); 267 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 268 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 269 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 270 syncookie_reseed, &V_tcp_syncache); 271} 272 273#ifdef VIMAGE 274void 275syncache_destroy(void) 276{ 277 struct syncache_head *sch; 278 struct syncache *sc, *nsc; 279 int i; 280 281 /* 282 * Stop the re-seed timer before freeing resources. No need to 283 * possibly schedule it another time. 284 */ 285 callout_drain(&V_tcp_syncache.secret.reseed); 286 287 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 288 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 289 290 sch = &V_tcp_syncache.hashbase[i]; 291 callout_drain(&sch->sch_timer); 292 293 SCH_LOCK(sch); 294 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 295 syncache_drop(sc, sch); 296 SCH_UNLOCK(sch); 297 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 298 ("%s: sch->sch_bucket not empty", __func__)); 299 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 300 __func__, sch->sch_length)); 301 mtx_destroy(&sch->sch_mtx); 302 } 303 304 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 305 ("%s: cache_count not 0", __func__)); 306 307 /* Free the allocated global resources. */ 308 uma_zdestroy(V_tcp_syncache.zone); 309 free(V_tcp_syncache.hashbase, M_SYNCACHE); 310} 311#endif 312 313/* 314 * Inserts a syncache entry into the specified bucket row. 315 * Locks and unlocks the syncache_head autonomously. 316 */ 317static void 318syncache_insert(struct syncache *sc, struct syncache_head *sch) 319{ 320 struct syncache *sc2; 321 322 SCH_LOCK(sch); 323 324 /* 325 * Make sure that we don't overflow the per-bucket limit. 326 * If the bucket is full, toss the oldest element. 327 */ 328 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 329 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 330 ("sch->sch_length incorrect")); 331 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 332 sch->sch_last_overflow = time_uptime; 333 syncache_drop(sc2, sch); 334 TCPSTAT_INC(tcps_sc_bucketoverflow); 335 } 336 337 /* Put it into the bucket. */ 338 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 339 sch->sch_length++; 340 341#ifdef TCP_OFFLOAD 342 if (ADDED_BY_TOE(sc)) { 343 struct toedev *tod = sc->sc_tod; 344 345 tod->tod_syncache_added(tod, sc->sc_todctx); 346 } 347#endif 348 349 /* Reinitialize the bucket row's timer. */ 350 if (sch->sch_length == 1) 351 sch->sch_nextc = ticks + INT_MAX; 352 syncache_timeout(sc, sch, 1); 353 354 SCH_UNLOCK(sch); 355 356 TCPSTATES_INC(TCPS_SYN_RECEIVED); 357 TCPSTAT_INC(tcps_sc_added); 358} 359 360/* 361 * Remove and free entry from syncache bucket row. 362 * Expects locked syncache head. 363 */ 364static void 365syncache_drop(struct syncache *sc, struct syncache_head *sch) 366{ 367 368 SCH_LOCK_ASSERT(sch); 369 370 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 371 TAILQ_REMOVE(&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_removed(tod, sc->sc_todctx); 379 } 380#endif 381 382 syncache_free(sc); 383} 384 385/* 386 * Engage/reengage time on bucket row. 387 */ 388static void 389syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 390{ 391 sc->sc_rxttime = ticks + 392 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]); 393 sc->sc_rxmits++; 394 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 395 sch->sch_nextc = sc->sc_rxttime; 396 if (docallout) 397 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 398 syncache_timer, (void *)sch); 399 } 400} 401 402/* 403 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 404 * If we have retransmitted an entry the maximum number of times, expire it. 405 * One separate timer for each bucket row. 406 */ 407static void 408syncache_timer(void *xsch) 409{ 410 struct syncache_head *sch = (struct syncache_head *)xsch; 411 struct syncache *sc, *nsc; 412 int tick = ticks; 413 char *s; 414 415 CURVNET_SET(sch->sch_sc->vnet); 416 417 /* NB: syncache_head has already been locked by the callout. */ 418 SCH_LOCK_ASSERT(sch); 419 420 /* 421 * In the following cycle we may remove some entries and/or 422 * advance some timeouts, so re-initialize the bucket timer. 423 */ 424 sch->sch_nextc = tick + INT_MAX; 425 426 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 427 /* 428 * We do not check if the listen socket still exists 429 * and accept the case where the listen socket may be 430 * gone by the time we resend the SYN/ACK. We do 431 * not expect this to happens often. If it does, 432 * then the RST will be sent by the time the remote 433 * host does the SYN/ACK->ACK. 434 */ 435 if (TSTMP_GT(sc->sc_rxttime, tick)) { 436 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 437 sch->sch_nextc = sc->sc_rxttime; 438 continue; 439 } 440 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 441 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 442 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 443 "giving up and removing syncache entry\n", 444 s, __func__); 445 free(s, M_TCPLOG); 446 } 447 syncache_drop(sc, sch); 448 TCPSTAT_INC(tcps_sc_stale); 449 continue; 450 } 451 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 452 log(LOG_DEBUG, "%s; %s: Response timeout, " 453 "retransmitting (%u) SYN|ACK\n", 454 s, __func__, sc->sc_rxmits); 455 free(s, M_TCPLOG); 456 } 457 458 syncache_respond(sc, sch, 1, NULL); 459 TCPSTAT_INC(tcps_sc_retransmitted); 460 syncache_timeout(sc, sch, 0); 461 } 462 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 463 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 464 syncache_timer, (void *)(sch)); 465 CURVNET_RESTORE(); 466} 467 468/* 469 * Find an entry in the syncache. 470 * Returns always with locked syncache_head plus a matching entry or NULL. 471 */ 472static struct syncache * 473syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 474{ 475 struct syncache *sc; 476 struct syncache_head *sch; 477 uint32_t hash; 478 479 /* 480 * The hash is built on foreign port + local port + foreign address. 481 * We rely on the fact that struct in_conninfo starts with 16 bits 482 * of foreign port, then 16 bits of local port then followed by 128 483 * bits of foreign address. In case of IPv4 address, the first 3 484 * 32-bit words of the address always are zeroes. 485 */ 486 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 487 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 488 489 sch = &V_tcp_syncache.hashbase[hash]; 490 *schp = sch; 491 SCH_LOCK(sch); 492 493 /* Circle through bucket row to find matching entry. */ 494 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 495 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 496 sizeof(struct in_endpoints)) == 0) 497 break; 498 499 return (sc); /* Always returns with locked sch. */ 500} 501 502/* 503 * This function is called when we get a RST for a 504 * non-existent connection, so that we can see if the 505 * connection is in the syn cache. If it is, zap it. 506 */ 507void 508syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 509{ 510 struct syncache *sc; 511 struct syncache_head *sch; 512 char *s = NULL; 513 514 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 515 SCH_LOCK_ASSERT(sch); 516 517 /* 518 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 519 * See RFC 793 page 65, section SEGMENT ARRIVES. 520 */ 521 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 522 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 523 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 524 "FIN flag set, segment ignored\n", s, __func__); 525 TCPSTAT_INC(tcps_badrst); 526 goto done; 527 } 528 529 /* 530 * No corresponding connection was found in syncache. 531 * If syncookies are enabled and possibly exclusively 532 * used, or we are under memory pressure, a valid RST 533 * may not find a syncache entry. In that case we're 534 * done and no SYN|ACK retransmissions will happen. 535 * Otherwise the RST was misdirected or spoofed. 536 */ 537 if (sc == NULL) { 538 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 539 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 540 "syncache entry (possibly syncookie only), " 541 "segment ignored\n", s, __func__); 542 TCPSTAT_INC(tcps_badrst); 543 goto done; 544 } 545 546 /* 547 * If the RST bit is set, check the sequence number to see 548 * if this is a valid reset segment. 549 * RFC 793 page 37: 550 * In all states except SYN-SENT, all reset (RST) segments 551 * are validated by checking their SEQ-fields. A reset is 552 * valid if its sequence number is in the window. 553 * 554 * The sequence number in the reset segment is normally an 555 * echo of our outgoing acknowlegement numbers, but some hosts 556 * send a reset with the sequence number at the rightmost edge 557 * of our receive window, and we have to handle this case. 558 */ 559 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 560 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 561 syncache_drop(sc, sch); 562 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 563 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 564 "connection attempt aborted by remote endpoint\n", 565 s, __func__); 566 TCPSTAT_INC(tcps_sc_reset); 567 } else { 568 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 569 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 570 "IRS %u (+WND %u), segment ignored\n", 571 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 572 TCPSTAT_INC(tcps_badrst); 573 } 574 575done: 576 if (s != NULL) 577 free(s, M_TCPLOG); 578 SCH_UNLOCK(sch); 579} 580 581void 582syncache_badack(struct in_conninfo *inc) 583{ 584 struct syncache *sc; 585 struct syncache_head *sch; 586 587 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 588 SCH_LOCK_ASSERT(sch); 589 if (sc != NULL) { 590 syncache_drop(sc, sch); 591 TCPSTAT_INC(tcps_sc_badack); 592 } 593 SCH_UNLOCK(sch); 594} 595 596void 597syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq) 598{ 599 struct syncache *sc; 600 struct syncache_head *sch; 601 602 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 603 SCH_LOCK_ASSERT(sch); 604 if (sc == NULL) 605 goto done; 606 607 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 608 if (ntohl(th_seq) != sc->sc_iss) 609 goto done; 610 611 /* 612 * If we've rertransmitted 3 times and this is our second error, 613 * we remove the entry. Otherwise, we allow it to continue on. 614 * This prevents us from incorrectly nuking an entry during a 615 * spurious network outage. 616 * 617 * See tcp_notify(). 618 */ 619 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 620 sc->sc_flags |= SCF_UNREACH; 621 goto done; 622 } 623 syncache_drop(sc, sch); 624 TCPSTAT_INC(tcps_sc_unreach); 625done: 626 SCH_UNLOCK(sch); 627} 628 629/* 630 * Build a new TCP socket structure from a syncache entry. 631 * 632 * On success return the newly created socket with its underlying inp locked. 633 */ 634static struct socket * 635syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 636{ 637 struct tcp_function_block *blk; 638 struct inpcb *inp = NULL; 639 struct socket *so; 640 struct tcpcb *tp; 641 int error; 642 char *s; 643 644 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 645 646 /* 647 * Ok, create the full blown connection, and set things up 648 * as they would have been set up if we had created the 649 * connection when the SYN arrived. If we can't create 650 * the connection, abort it. 651 */ 652 so = sonewconn(lso, 0); 653 if (so == NULL) { 654 /* 655 * Drop the connection; we will either send a RST or 656 * have the peer retransmit its SYN again after its 657 * RTO and try again. 658 */ 659 TCPSTAT_INC(tcps_listendrop); 660 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 661 log(LOG_DEBUG, "%s; %s: Socket create failed " 662 "due to limits or memory shortage\n", 663 s, __func__); 664 free(s, M_TCPLOG); 665 } 666 goto abort2; 667 } 668#ifdef MAC 669 mac_socketpeer_set_from_mbuf(m, so); 670#endif 671 672 inp = sotoinpcb(so); 673 inp->inp_inc.inc_fibnum = so->so_fibnum; 674 INP_WLOCK(inp); 675 /* 676 * Exclusive pcbinfo lock is not required in syncache socket case even 677 * if two inpcb locks can be acquired simultaneously: 678 * - the inpcb in LISTEN state, 679 * - the newly created inp. 680 * 681 * In this case, an inp cannot be at same time in LISTEN state and 682 * just created by an accept() call. 683 */ 684 INP_HASH_WLOCK(&V_tcbinfo); 685 686 /* Insert new socket into PCB hash list. */ 687 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 688#ifdef INET6 689 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 690 inp->inp_vflag &= ~INP_IPV4; 691 inp->inp_vflag |= INP_IPV6; 692 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 693 } else { 694 inp->inp_vflag &= ~INP_IPV6; 695 inp->inp_vflag |= INP_IPV4; 696#endif 697 inp->inp_laddr = sc->sc_inc.inc_laddr; 698#ifdef INET6 699 } 700#endif 701 702 /* 703 * If there's an mbuf and it has a flowid, then let's initialise the 704 * inp with that particular flowid. 705 */ 706 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 707 inp->inp_flowid = m->m_pkthdr.flowid; 708 inp->inp_flowtype = M_HASHTYPE_GET(m); 709 } 710 711 /* 712 * Install in the reservation hash table for now, but don't yet 713 * install a connection group since the full 4-tuple isn't yet 714 * configured. 715 */ 716 inp->inp_lport = sc->sc_inc.inc_lport; 717 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 718 /* 719 * Undo the assignments above if we failed to 720 * put the PCB on the hash lists. 721 */ 722#ifdef INET6 723 if (sc->sc_inc.inc_flags & INC_ISIPV6) 724 inp->in6p_laddr = in6addr_any; 725 else 726#endif 727 inp->inp_laddr.s_addr = INADDR_ANY; 728 inp->inp_lport = 0; 729 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 730 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 731 "with error %i\n", 732 s, __func__, error); 733 free(s, M_TCPLOG); 734 } 735 INP_HASH_WUNLOCK(&V_tcbinfo); 736 goto abort; 737 } 738#ifdef INET6 739 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 740 struct inpcb *oinp = sotoinpcb(lso); 741 struct in6_addr laddr6; 742 struct sockaddr_in6 sin6; 743 /* 744 * Inherit socket options from the listening socket. 745 * Note that in6p_inputopts are not (and should not be) 746 * copied, since it stores previously received options and is 747 * used to detect if each new option is different than the 748 * previous one and hence should be passed to a user. 749 * If we copied in6p_inputopts, a user would not be able to 750 * receive options just after calling the accept system call. 751 */ 752 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 753 if (oinp->in6p_outputopts) 754 inp->in6p_outputopts = 755 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 756 757 sin6.sin6_family = AF_INET6; 758 sin6.sin6_len = sizeof(sin6); 759 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 760 sin6.sin6_port = sc->sc_inc.inc_fport; 761 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 762 laddr6 = inp->in6p_laddr; 763 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 764 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 765 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 766 thread0.td_ucred, m)) != 0) { 767 inp->in6p_laddr = laddr6; 768 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 769 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 770 "with error %i\n", 771 s, __func__, error); 772 free(s, M_TCPLOG); 773 } 774 INP_HASH_WUNLOCK(&V_tcbinfo); 775 goto abort; 776 } 777 /* Override flowlabel from in6_pcbconnect. */ 778 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 779 inp->inp_flow |= sc->sc_flowlabel; 780 } 781#endif /* INET6 */ 782#if defined(INET) && defined(INET6) 783 else 784#endif 785#ifdef INET 786 { 787 struct in_addr laddr; 788 struct sockaddr_in sin; 789 790 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 791 792 if (inp->inp_options == NULL) { 793 inp->inp_options = sc->sc_ipopts; 794 sc->sc_ipopts = NULL; 795 } 796 797 sin.sin_family = AF_INET; 798 sin.sin_len = sizeof(sin); 799 sin.sin_addr = sc->sc_inc.inc_faddr; 800 sin.sin_port = sc->sc_inc.inc_fport; 801 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 802 laddr = inp->inp_laddr; 803 if (inp->inp_laddr.s_addr == INADDR_ANY) 804 inp->inp_laddr = sc->sc_inc.inc_laddr; 805 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 806 thread0.td_ucred, m)) != 0) { 807 inp->inp_laddr = laddr; 808 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 809 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 810 "with error %i\n", 811 s, __func__, error); 812 free(s, M_TCPLOG); 813 } 814 INP_HASH_WUNLOCK(&V_tcbinfo); 815 goto abort; 816 } 817 } 818#endif /* INET */ 819#if defined(IPSEC) || defined(IPSEC_SUPPORT) 820 /* Copy old policy into new socket's. */ 821 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 822 printf("syncache_socket: could not copy policy\n"); 823#endif 824 INP_HASH_WUNLOCK(&V_tcbinfo); 825 tp = intotcpcb(inp); 826 tcp_state_change(tp, TCPS_SYN_RECEIVED); 827 tp->iss = sc->sc_iss; 828 tp->irs = sc->sc_irs; 829 tcp_rcvseqinit(tp); 830 tcp_sendseqinit(tp); 831 blk = sototcpcb(lso)->t_fb; 832 if (blk != tp->t_fb) { 833 /* 834 * Our parents t_fb was not the default, 835 * we need to release our ref on tp->t_fb and 836 * pickup one on the new entry. 837 */ 838 struct tcp_function_block *rblk; 839 840 rblk = find_and_ref_tcp_fb(blk); 841 KASSERT(rblk != NULL, 842 ("cannot find blk %p out of syncache?", blk)); 843 if (tp->t_fb->tfb_tcp_fb_fini) 844 (*tp->t_fb->tfb_tcp_fb_fini)(tp); 845 refcount_release(&tp->t_fb->tfb_refcnt); 846 tp->t_fb = rblk; 847 if (tp->t_fb->tfb_tcp_fb_init) { 848 (*tp->t_fb->tfb_tcp_fb_init)(tp); 849 } 850 } 851 tp->snd_wl1 = sc->sc_irs; 852 tp->snd_max = tp->iss + 1; 853 tp->snd_nxt = tp->iss + 1; 854 tp->rcv_up = sc->sc_irs + 1; 855 tp->rcv_wnd = sc->sc_wnd; 856 tp->rcv_adv += tp->rcv_wnd; 857 tp->last_ack_sent = tp->rcv_nxt; 858 859 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 860 if (sc->sc_flags & SCF_NOOPT) 861 tp->t_flags |= TF_NOOPT; 862 else { 863 if (sc->sc_flags & SCF_WINSCALE) { 864 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 865 tp->snd_scale = sc->sc_requested_s_scale; 866 tp->request_r_scale = sc->sc_requested_r_scale; 867 } 868 if (sc->sc_flags & SCF_TIMESTAMP) { 869 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 870 tp->ts_recent = sc->sc_tsreflect; 871 tp->ts_recent_age = tcp_ts_getticks(); 872 tp->ts_offset = sc->sc_tsoff; 873 } 874#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 875 if (sc->sc_flags & SCF_SIGNATURE) 876 tp->t_flags |= TF_SIGNATURE; 877#endif 878 if (sc->sc_flags & SCF_SACK) 879 tp->t_flags |= TF_SACK_PERMIT; 880 } 881 882 if (sc->sc_flags & SCF_ECN) 883 tp->t_flags |= TF_ECN_PERMIT; 884 885 /* 886 * Set up MSS and get cached values from tcp_hostcache. 887 * This might overwrite some of the defaults we just set. 888 */ 889 tcp_mss(tp, sc->sc_peer_mss); 890 891 /* 892 * If the SYN,ACK was retransmitted, indicate that CWND to be 893 * limited to one segment in cc_conn_init(). 894 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 895 */ 896 if (sc->sc_rxmits > 1) 897 tp->snd_cwnd = 1; 898 899#ifdef TCP_OFFLOAD 900 /* 901 * Allow a TOE driver to install its hooks. Note that we hold the 902 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 903 * new connection before the TOE driver has done its thing. 904 */ 905 if (ADDED_BY_TOE(sc)) { 906 struct toedev *tod = sc->sc_tod; 907 908 tod->tod_offload_socket(tod, sc->sc_todctx, so); 909 } 910#endif 911 /* 912 * Copy and activate timers. 913 */ 914 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 915 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 916 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 917 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 918 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 919 920 TCPSTAT_INC(tcps_accepts); 921 return (so); 922 923abort: 924 INP_WUNLOCK(inp); 925abort2: 926 if (so != NULL) 927 soabort(so); 928 return (NULL); 929} 930 931/* 932 * This function gets called when we receive an ACK for a 933 * socket in the LISTEN state. We look up the connection 934 * in the syncache, and if its there, we pull it out of 935 * the cache and turn it into a full-blown connection in 936 * the SYN-RECEIVED state. 937 * 938 * On syncache_socket() success the newly created socket 939 * has its underlying inp locked. 940 */ 941int 942syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 943 struct socket **lsop, struct mbuf *m) 944{ 945 struct syncache *sc; 946 struct syncache_head *sch; 947 struct syncache scs; 948 char *s; 949 950 /* 951 * Global TCP locks are held because we manipulate the PCB lists 952 * and create a new socket. 953 */ 954 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 955 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 956 ("%s: can handle only ACK", __func__)); 957 958 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 959 SCH_LOCK_ASSERT(sch); 960 961#ifdef INVARIANTS 962 /* 963 * Test code for syncookies comparing the syncache stored 964 * values with the reconstructed values from the cookie. 965 */ 966 if (sc != NULL) 967 syncookie_cmp(inc, sch, sc, th, to, *lsop); 968#endif 969 970 if (sc == NULL) { 971 /* 972 * There is no syncache entry, so see if this ACK is 973 * a returning syncookie. To do this, first: 974 * A. Check if syncookies are used in case of syncache 975 * overflows 976 * B. See if this socket has had a syncache entry dropped in 977 * the recent past. We don't want to accept a bogus 978 * syncookie if we've never received a SYN or accept it 979 * twice. 980 * C. check that the syncookie is valid. If it is, then 981 * cobble up a fake syncache entry, and return. 982 */ 983 if (!V_tcp_syncookies) { 984 SCH_UNLOCK(sch); 985 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 986 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 987 "segment rejected (syncookies disabled)\n", 988 s, __func__); 989 goto failed; 990 } 991 if (!V_tcp_syncookiesonly && 992 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 993 SCH_UNLOCK(sch); 994 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 995 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 996 "segment rejected (no syncache entry)\n", 997 s, __func__); 998 goto failed; 999 } 1000 bzero(&scs, sizeof(scs)); 1001 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1002 SCH_UNLOCK(sch); 1003 if (sc == NULL) { 1004 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1005 log(LOG_DEBUG, "%s; %s: Segment failed " 1006 "SYNCOOKIE authentication, segment rejected " 1007 "(probably spoofed)\n", s, __func__); 1008 goto failed; 1009 } 1010#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1011 /* If received ACK has MD5 signature, check it. */ 1012 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1013 (!TCPMD5_ENABLED() || 1014 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1015 /* Drop the ACK. */ 1016 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1017 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1018 "MD5 signature doesn't match.\n", 1019 s, __func__); 1020 free(s, M_TCPLOG); 1021 } 1022 TCPSTAT_INC(tcps_sig_err_sigopt); 1023 return (-1); /* Do not send RST */ 1024 } 1025#endif /* TCP_SIGNATURE */ 1026 } else { 1027#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1028 /* 1029 * If listening socket requested TCP digests, check that 1030 * received ACK has signature and it is correct. 1031 * If not, drop the ACK and leave sc entry in th cache, 1032 * because SYN was received with correct signature. 1033 */ 1034 if (sc->sc_flags & SCF_SIGNATURE) { 1035 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1036 /* No signature */ 1037 TCPSTAT_INC(tcps_sig_err_nosigopt); 1038 SCH_UNLOCK(sch); 1039 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1040 log(LOG_DEBUG, "%s; %s: Segment " 1041 "rejected, MD5 signature wasn't " 1042 "provided.\n", s, __func__); 1043 free(s, M_TCPLOG); 1044 } 1045 return (-1); /* Do not send RST */ 1046 } 1047 if (!TCPMD5_ENABLED() || 1048 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1049 /* Doesn't match or no SA */ 1050 SCH_UNLOCK(sch); 1051 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1052 log(LOG_DEBUG, "%s; %s: Segment " 1053 "rejected, MD5 signature doesn't " 1054 "match.\n", s, __func__); 1055 free(s, M_TCPLOG); 1056 } 1057 return (-1); /* Do not send RST */ 1058 } 1059 } 1060#endif /* TCP_SIGNATURE */ 1061 /* 1062 * Pull out the entry to unlock the bucket row. 1063 * 1064 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not 1065 * tcp_state_change(). The tcpcb is not existent at this 1066 * moment. A new one will be allocated via syncache_socket-> 1067 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then 1068 * syncache_socket() will change it to TCPS_SYN_RECEIVED. 1069 */ 1070 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1071 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1072 sch->sch_length--; 1073#ifdef TCP_OFFLOAD 1074 if (ADDED_BY_TOE(sc)) { 1075 struct toedev *tod = sc->sc_tod; 1076 1077 tod->tod_syncache_removed(tod, sc->sc_todctx); 1078 } 1079#endif 1080 SCH_UNLOCK(sch); 1081 } 1082 1083 /* 1084 * Segment validation: 1085 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1086 */ 1087 if (th->th_ack != sc->sc_iss + 1) { 1088 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1089 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1090 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1091 goto failed; 1092 } 1093 1094 /* 1095 * The SEQ must fall in the window starting at the received 1096 * initial receive sequence number + 1 (the SYN). 1097 */ 1098 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1099 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1100 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1101 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1102 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1103 goto failed; 1104 } 1105 1106 /* 1107 * If timestamps were not negotiated during SYN/ACK they 1108 * must not appear on any segment during this session. 1109 */ 1110 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1111 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1112 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1113 "segment rejected\n", s, __func__); 1114 goto failed; 1115 } 1116 1117 /* 1118 * If timestamps were negotiated during SYN/ACK they should 1119 * appear on every segment during this session. 1120 * XXXAO: This is only informal as there have been unverified 1121 * reports of non-compliants stacks. 1122 */ 1123 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1124 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1125 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1126 "no action\n", s, __func__); 1127 free(s, M_TCPLOG); 1128 s = NULL; 1129 } 1130 } 1131 1132 /* 1133 * If timestamps were negotiated, the reflected timestamp 1134 * must be equal to what we actually sent in the SYN|ACK 1135 * except in the case of 0. Some boxes are known for sending 1136 * broken timestamp replies during the 3whs (and potentially 1137 * during the connection also). 1138 * 1139 * Accept the final ACK of 3whs with reflected timestamp of 0 1140 * instead of sending a RST and deleting the syncache entry. 1141 */ 1142 if ((to->to_flags & TOF_TS) && to->to_tsecr && 1143 to->to_tsecr != sc->sc_ts) { 1144 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1145 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1146 "segment rejected\n", 1147 s, __func__, to->to_tsecr, sc->sc_ts); 1148 goto failed; 1149 } 1150 1151 *lsop = syncache_socket(sc, *lsop, m); 1152 1153 if (*lsop == NULL) 1154 TCPSTAT_INC(tcps_sc_aborted); 1155 else 1156 TCPSTAT_INC(tcps_sc_completed); 1157 1158/* how do we find the inp for the new socket? */ 1159 if (sc != &scs) 1160 syncache_free(sc); 1161 return (1); 1162failed: 1163 if (sc != NULL && sc != &scs) 1164 syncache_free(sc); 1165 if (s != NULL) 1166 free(s, M_TCPLOG); 1167 *lsop = NULL; 1168 return (0); 1169} 1170 1171#ifdef TCP_RFC7413 1172static void 1173syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1174 uint64_t response_cookie) 1175{ 1176 struct inpcb *inp; 1177 struct tcpcb *tp; 1178 unsigned int *pending_counter; 1179 1180 /* 1181 * Global TCP locks are held because we manipulate the PCB lists 1182 * and create a new socket. 1183 */ 1184 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1185 1186 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1187 *lsop = syncache_socket(sc, *lsop, m); 1188 if (*lsop == NULL) { 1189 TCPSTAT_INC(tcps_sc_aborted); 1190 atomic_subtract_int(pending_counter, 1); 1191 } else { 1192 inp = sotoinpcb(*lsop); 1193 tp = intotcpcb(inp); 1194 tp->t_flags |= TF_FASTOPEN; 1195 tp->t_tfo_cookie = response_cookie; 1196 tp->snd_max = tp->iss; 1197 tp->snd_nxt = tp->iss; 1198 tp->t_tfo_pending = pending_counter; 1199 TCPSTAT_INC(tcps_sc_completed); 1200 } 1201} 1202#endif /* TCP_RFC7413 */ 1203 1204/* 1205 * Given a LISTEN socket and an inbound SYN request, add 1206 * this to the syn cache, and send back a segment: 1207 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1208 * to the source. 1209 * 1210 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1211 * Doing so would require that we hold onto the data and deliver it 1212 * to the application. However, if we are the target of a SYN-flood 1213 * DoS attack, an attacker could send data which would eventually 1214 * consume all available buffer space if it were ACKed. By not ACKing 1215 * the data, we avoid this DoS scenario. 1216 * 1217 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1218 * cookie is processed, V_tcp_fastopen_enabled set to true, and the 1219 * TCP_FASTOPEN socket option is set. In this case, a new socket is created 1220 * and returned via lsop, the mbuf is not freed so that tcp_input() can 1221 * queue its data to the socket, and 1 is returned to indicate the 1222 * TFO-socket-creation path was taken. 1223 */ 1224int 1225syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1226 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1227 void *todctx) 1228{ 1229 struct tcpcb *tp; 1230 struct socket *so; 1231 struct syncache *sc = NULL; 1232 struct syncache_head *sch; 1233 struct mbuf *ipopts = NULL; 1234 u_int ltflags; 1235 int win, sb_hiwat, ip_ttl, ip_tos; 1236 char *s; 1237 int rv = 0; 1238#ifdef INET6 1239 int autoflowlabel = 0; 1240#endif 1241#ifdef MAC 1242 struct label *maclabel; 1243#endif 1244 struct syncache scs; 1245 struct ucred *cred; 1246#ifdef TCP_RFC7413 1247 uint64_t tfo_response_cookie; 1248 int tfo_cookie_valid = 0; 1249 int tfo_response_cookie_valid = 0; 1250#endif 1251 1252 INP_WLOCK_ASSERT(inp); /* listen socket */ 1253 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1254 ("%s: unexpected tcp flags", __func__)); 1255 1256 /* 1257 * Combine all so/tp operations very early to drop the INP lock as 1258 * soon as possible. 1259 */ 1260 so = *lsop; 1261 tp = sototcpcb(so); 1262 cred = crhold(so->so_cred); 1263 1264#ifdef INET6 1265 if ((inc->inc_flags & INC_ISIPV6) && 1266 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1267 autoflowlabel = 1; 1268#endif 1269 ip_ttl = inp->inp_ip_ttl; 1270 ip_tos = inp->inp_ip_tos; 1271 win = sbspace(&so->so_rcv); 1272 sb_hiwat = so->so_rcv.sb_hiwat; 1273 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1274 1275#ifdef TCP_RFC7413 1276 if (V_tcp_fastopen_enabled && (tp->t_flags & TF_FASTOPEN) && 1277 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) { 1278 /* 1279 * Limit the number of pending TFO connections to 1280 * approximately half of the queue limit. This prevents TFO 1281 * SYN floods from starving the service by filling the 1282 * listen queue with bogus TFO connections. 1283 */ 1284 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1285 (so->so_qlimit / 2)) { 1286 int result; 1287 1288 result = tcp_fastopen_check_cookie(inc, 1289 to->to_tfo_cookie, to->to_tfo_len, 1290 &tfo_response_cookie); 1291 tfo_cookie_valid = (result > 0); 1292 tfo_response_cookie_valid = (result >= 0); 1293 } else 1294 atomic_subtract_int(tp->t_tfo_pending, 1); 1295 } 1296#endif 1297 1298 /* By the time we drop the lock these should no longer be used. */ 1299 so = NULL; 1300 tp = NULL; 1301 1302#ifdef MAC 1303 if (mac_syncache_init(&maclabel) != 0) { 1304 INP_WUNLOCK(inp); 1305 goto done; 1306 } else 1307 mac_syncache_create(maclabel, inp); 1308#endif 1309#ifdef TCP_RFC7413 1310 if (!tfo_cookie_valid) 1311#endif 1312 INP_WUNLOCK(inp); 1313 1314 /* 1315 * Remember the IP options, if any. 1316 */ 1317#ifdef INET6 1318 if (!(inc->inc_flags & INC_ISIPV6)) 1319#endif 1320#ifdef INET 1321 ipopts = (m) ? ip_srcroute(m) : NULL; 1322#else 1323 ipopts = NULL; 1324#endif 1325 1326#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1327 /* 1328 * If listening socket requested TCP digests, check that received 1329 * SYN has signature and it is correct. If signature doesn't match 1330 * or TCP_SIGNATURE support isn't enabled, drop the packet. 1331 */ 1332 if (ltflags & TF_SIGNATURE) { 1333 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1334 TCPSTAT_INC(tcps_sig_err_nosigopt); 1335 goto done; 1336 } 1337 if (!TCPMD5_ENABLED() || 1338 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1339 goto done; 1340 } 1341#endif /* TCP_SIGNATURE */ 1342 /* 1343 * See if we already have an entry for this connection. 1344 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1345 * 1346 * XXX: should the syncache be re-initialized with the contents 1347 * of the new SYN here (which may have different options?) 1348 * 1349 * XXX: We do not check the sequence number to see if this is a 1350 * real retransmit or a new connection attempt. The question is 1351 * how to handle such a case; either ignore it as spoofed, or 1352 * drop the current entry and create a new one? 1353 */ 1354 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1355 SCH_LOCK_ASSERT(sch); 1356 if (sc != NULL) { 1357#ifdef TCP_RFC7413 1358 if (tfo_cookie_valid) 1359 INP_WUNLOCK(inp); 1360#endif 1361 TCPSTAT_INC(tcps_sc_dupsyn); 1362 if (ipopts) { 1363 /* 1364 * If we were remembering a previous source route, 1365 * forget it and use the new one we've been given. 1366 */ 1367 if (sc->sc_ipopts) 1368 (void) m_free(sc->sc_ipopts); 1369 sc->sc_ipopts = ipopts; 1370 } 1371 /* 1372 * Update timestamp if present. 1373 */ 1374 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1375 sc->sc_tsreflect = to->to_tsval; 1376 else 1377 sc->sc_flags &= ~SCF_TIMESTAMP; 1378#ifdef MAC 1379 /* 1380 * Since we have already unconditionally allocated label 1381 * storage, free it up. The syncache entry will already 1382 * have an initialized label we can use. 1383 */ 1384 mac_syncache_destroy(&maclabel); 1385#endif 1386 /* Retransmit SYN|ACK and reset retransmit count. */ 1387 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1388 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1389 "resetting timer and retransmitting SYN|ACK\n", 1390 s, __func__); 1391 free(s, M_TCPLOG); 1392 } 1393 if (syncache_respond(sc, sch, 1, m) == 0) { 1394 sc->sc_rxmits = 0; 1395 syncache_timeout(sc, sch, 1); 1396 TCPSTAT_INC(tcps_sndacks); 1397 TCPSTAT_INC(tcps_sndtotal); 1398 } 1399 SCH_UNLOCK(sch); 1400 goto done; 1401 } 1402 1403#ifdef TCP_RFC7413 1404 if (tfo_cookie_valid) { 1405 bzero(&scs, sizeof(scs)); 1406 sc = &scs; 1407 goto skip_alloc; 1408 } 1409#endif 1410 1411 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1412 if (sc == NULL) { 1413 /* 1414 * The zone allocator couldn't provide more entries. 1415 * Treat this as if the cache was full; drop the oldest 1416 * entry and insert the new one. 1417 */ 1418 TCPSTAT_INC(tcps_sc_zonefail); 1419 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1420 sch->sch_last_overflow = time_uptime; 1421 syncache_drop(sc, sch); 1422 } 1423 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1424 if (sc == NULL) { 1425 if (V_tcp_syncookies) { 1426 bzero(&scs, sizeof(scs)); 1427 sc = &scs; 1428 } else { 1429 SCH_UNLOCK(sch); 1430 if (ipopts) 1431 (void) m_free(ipopts); 1432 goto done; 1433 } 1434 } 1435 } 1436 1437#ifdef TCP_RFC7413 1438skip_alloc: 1439 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1440 sc->sc_tfo_cookie = &tfo_response_cookie; 1441#endif 1442 1443 /* 1444 * Fill in the syncache values. 1445 */ 1446#ifdef MAC 1447 sc->sc_label = maclabel; 1448#endif 1449 sc->sc_cred = cred; 1450 cred = NULL; 1451 sc->sc_ipopts = ipopts; 1452 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1453#ifdef INET6 1454 if (!(inc->inc_flags & INC_ISIPV6)) 1455#endif 1456 { 1457 sc->sc_ip_tos = ip_tos; 1458 sc->sc_ip_ttl = ip_ttl; 1459 } 1460#ifdef TCP_OFFLOAD 1461 sc->sc_tod = tod; 1462 sc->sc_todctx = todctx; 1463#endif 1464 sc->sc_irs = th->th_seq; 1465 sc->sc_iss = arc4random(); 1466 sc->sc_flags = 0; 1467 sc->sc_flowlabel = 0; 1468 1469 /* 1470 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1471 * win was derived from socket earlier in the function. 1472 */ 1473 win = imax(win, 0); 1474 win = imin(win, TCP_MAXWIN); 1475 sc->sc_wnd = win; 1476 1477 if (V_tcp_do_rfc1323) { 1478 /* 1479 * A timestamp received in a SYN makes 1480 * it ok to send timestamp requests and replies. 1481 */ 1482 if (to->to_flags & TOF_TS) { 1483 sc->sc_tsreflect = to->to_tsval; 1484 sc->sc_ts = tcp_ts_getticks(); 1485 sc->sc_flags |= SCF_TIMESTAMP; 1486 } 1487 if (to->to_flags & TOF_SCALE) { 1488 int wscale = 0; 1489 1490 /* 1491 * Pick the smallest possible scaling factor that 1492 * will still allow us to scale up to sb_max, aka 1493 * kern.ipc.maxsockbuf. 1494 * 1495 * We do this because there are broken firewalls that 1496 * will corrupt the window scale option, leading to 1497 * the other endpoint believing that our advertised 1498 * window is unscaled. At scale factors larger than 1499 * 5 the unscaled window will drop below 1500 bytes, 1500 * leading to serious problems when traversing these 1501 * broken firewalls. 1502 * 1503 * With the default maxsockbuf of 256K, a scale factor 1504 * of 3 will be chosen by this algorithm. Those who 1505 * choose a larger maxsockbuf should watch out 1506 * for the compatibility problems mentioned above. 1507 * 1508 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1509 * or <SYN,ACK>) segment itself is never scaled. 1510 */ 1511 while (wscale < TCP_MAX_WINSHIFT && 1512 (TCP_MAXWIN << wscale) < sb_max) 1513 wscale++; 1514 sc->sc_requested_r_scale = wscale; 1515 sc->sc_requested_s_scale = to->to_wscale; 1516 sc->sc_flags |= SCF_WINSCALE; 1517 } 1518 } 1519#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1520 /* 1521 * If listening socket requested TCP digests, flag this in the 1522 * syncache so that syncache_respond() will do the right thing 1523 * with the SYN+ACK. 1524 */ 1525 if (ltflags & TF_SIGNATURE) 1526 sc->sc_flags |= SCF_SIGNATURE; 1527#endif /* TCP_SIGNATURE */ 1528 if (to->to_flags & TOF_SACKPERM) 1529 sc->sc_flags |= SCF_SACK; 1530 if (to->to_flags & TOF_MSS) 1531 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1532 if (ltflags & TF_NOOPT) 1533 sc->sc_flags |= SCF_NOOPT; 1534 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1535 sc->sc_flags |= SCF_ECN; 1536 1537 if (V_tcp_syncookies) 1538 sc->sc_iss = syncookie_generate(sch, sc); 1539#ifdef INET6 1540 if (autoflowlabel) { 1541 if (V_tcp_syncookies) 1542 sc->sc_flowlabel = sc->sc_iss; 1543 else 1544 sc->sc_flowlabel = ip6_randomflowlabel(); 1545 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1546 } 1547#endif 1548 SCH_UNLOCK(sch); 1549 1550#ifdef TCP_RFC7413 1551 if (tfo_cookie_valid) { 1552 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1553 /* INP_WUNLOCK(inp) will be performed by the called */ 1554 rv = 1; 1555 goto tfo_done; 1556 } 1557#endif 1558 1559 /* 1560 * Do a standard 3-way handshake. 1561 */ 1562 if (syncache_respond(sc, sch, 0, m) == 0) { 1563 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1564 syncache_free(sc); 1565 else if (sc != &scs) 1566 syncache_insert(sc, sch); /* locks and unlocks sch */ 1567 TCPSTAT_INC(tcps_sndacks); 1568 TCPSTAT_INC(tcps_sndtotal); 1569 } else { 1570 if (sc != &scs) 1571 syncache_free(sc); 1572 TCPSTAT_INC(tcps_sc_dropped); 1573 } 1574 1575done: 1576 if (m) { 1577 *lsop = NULL; 1578 m_freem(m); 1579 } 1580#ifdef TCP_RFC7413 1581tfo_done: 1582#endif 1583 if (cred != NULL) 1584 crfree(cred); 1585#ifdef MAC 1586 if (sc == &scs) 1587 mac_syncache_destroy(&maclabel); 1588#endif 1589 return (rv); 1590} 1591 1592/* 1593 * Send SYN|ACK to the peer. Either in response to the peer's SYN, 1594 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1595 */ 1596static int 1597syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked, 1598 const struct mbuf *m0) 1599{ 1600 struct ip *ip = NULL; 1601 struct mbuf *m; 1602 struct tcphdr *th = NULL; 1603 int optlen, error = 0; /* Make compiler happy */ 1604 u_int16_t hlen, tlen, mssopt; 1605 struct tcpopt to; 1606#ifdef INET6 1607 struct ip6_hdr *ip6 = NULL; 1608#endif 1609 hlen = 1610#ifdef INET6 1611 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1612#endif 1613 sizeof(struct ip); 1614 tlen = hlen + sizeof(struct tcphdr); 1615 1616 /* Determine MSS we advertize to other end of connection. */ 1617 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); 1618 1619 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1620 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1621 ("syncache: mbuf too small")); 1622 1623 /* Create the IP+TCP header from scratch. */ 1624 m = m_gethdr(M_NOWAIT, MT_DATA); 1625 if (m == NULL) 1626 return (ENOBUFS); 1627#ifdef MAC 1628 mac_syncache_create_mbuf(sc->sc_label, m); 1629#endif 1630 m->m_data += max_linkhdr; 1631 m->m_len = tlen; 1632 m->m_pkthdr.len = tlen; 1633 m->m_pkthdr.rcvif = NULL; 1634 1635#ifdef INET6 1636 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1637 ip6 = mtod(m, struct ip6_hdr *); 1638 ip6->ip6_vfc = IPV6_VERSION; 1639 ip6->ip6_nxt = IPPROTO_TCP; 1640 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1641 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1642 ip6->ip6_plen = htons(tlen - hlen); 1643 /* ip6_hlim is set after checksum */ 1644 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1645 ip6->ip6_flow |= sc->sc_flowlabel; 1646 1647 th = (struct tcphdr *)(ip6 + 1); 1648 } 1649#endif 1650#if defined(INET6) && defined(INET) 1651 else 1652#endif 1653#ifdef INET 1654 { 1655 ip = mtod(m, struct ip *); 1656 ip->ip_v = IPVERSION; 1657 ip->ip_hl = sizeof(struct ip) >> 2; 1658 ip->ip_len = htons(tlen); 1659 ip->ip_id = 0; 1660 ip->ip_off = 0; 1661 ip->ip_sum = 0; 1662 ip->ip_p = IPPROTO_TCP; 1663 ip->ip_src = sc->sc_inc.inc_laddr; 1664 ip->ip_dst = sc->sc_inc.inc_faddr; 1665 ip->ip_ttl = sc->sc_ip_ttl; 1666 ip->ip_tos = sc->sc_ip_tos; 1667 1668 /* 1669 * See if we should do MTU discovery. Route lookups are 1670 * expensive, so we will only unset the DF bit if: 1671 * 1672 * 1) path_mtu_discovery is disabled 1673 * 2) the SCF_UNREACH flag has been set 1674 */ 1675 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1676 ip->ip_off |= htons(IP_DF); 1677 1678 th = (struct tcphdr *)(ip + 1); 1679 } 1680#endif /* INET */ 1681 th->th_sport = sc->sc_inc.inc_lport; 1682 th->th_dport = sc->sc_inc.inc_fport; 1683 1684 th->th_seq = htonl(sc->sc_iss); 1685 th->th_ack = htonl(sc->sc_irs + 1); 1686 th->th_off = sizeof(struct tcphdr) >> 2; 1687 th->th_x2 = 0; 1688 th->th_flags = TH_SYN|TH_ACK; 1689 th->th_win = htons(sc->sc_wnd); 1690 th->th_urp = 0; 1691 1692 if (sc->sc_flags & SCF_ECN) { 1693 th->th_flags |= TH_ECE; 1694 TCPSTAT_INC(tcps_ecn_shs); 1695 } 1696 1697 /* Tack on the TCP options. */ 1698 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1699 to.to_flags = 0; 1700 1701 to.to_mss = mssopt; 1702 to.to_flags = TOF_MSS; 1703 if (sc->sc_flags & SCF_WINSCALE) { 1704 to.to_wscale = sc->sc_requested_r_scale; 1705 to.to_flags |= TOF_SCALE; 1706 } 1707 if (sc->sc_flags & SCF_TIMESTAMP) { 1708 /* Virgin timestamp or TCP cookie enhanced one. */ 1709 to.to_tsval = sc->sc_ts; 1710 to.to_tsecr = sc->sc_tsreflect; 1711 to.to_flags |= TOF_TS; 1712 } 1713 if (sc->sc_flags & SCF_SACK) 1714 to.to_flags |= TOF_SACKPERM; 1715#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1716 if (sc->sc_flags & SCF_SIGNATURE) 1717 to.to_flags |= TOF_SIGNATURE; 1718#endif 1719#ifdef TCP_RFC7413 1720 if (sc->sc_tfo_cookie) { 1721 to.to_flags |= TOF_FASTOPEN; 1722 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1723 to.to_tfo_cookie = sc->sc_tfo_cookie; 1724 /* don't send cookie again when retransmitting response */ 1725 sc->sc_tfo_cookie = NULL; 1726 } 1727#endif 1728 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1729 1730 /* Adjust headers by option size. */ 1731 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1732 m->m_len += optlen; 1733 m->m_pkthdr.len += optlen; 1734#ifdef INET6 1735 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1736 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1737 else 1738#endif 1739 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1740#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1741 if (sc->sc_flags & SCF_SIGNATURE) { 1742 KASSERT(to.to_flags & TOF_SIGNATURE, 1743 ("tcp_addoptions() didn't set tcp_signature")); 1744 1745 /* NOTE: to.to_signature is inside of mbuf */ 1746 if (!TCPMD5_ENABLED() || 1747 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1748 m_freem(m); 1749 return (EACCES); 1750 } 1751 } 1752#endif 1753 } else 1754 optlen = 0; 1755 1756 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1757 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1758 /* 1759 * If we have peer's SYN and it has a flowid, then let's assign it to 1760 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1761 * to SYN|ACK due to lack of inp here. 1762 */ 1763 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1764 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1765 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1766 } 1767#ifdef INET6 1768 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1769 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1770 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1771 IPPROTO_TCP, 0); 1772 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1773#ifdef TCP_OFFLOAD 1774 if (ADDED_BY_TOE(sc)) { 1775 struct toedev *tod = sc->sc_tod; 1776 1777 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1778 1779 return (error); 1780 } 1781#endif 1782 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1783 } 1784#endif 1785#if defined(INET6) && defined(INET) 1786 else 1787#endif 1788#ifdef INET 1789 { 1790 m->m_pkthdr.csum_flags = CSUM_TCP; 1791 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1792 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1793#ifdef TCP_OFFLOAD 1794 if (ADDED_BY_TOE(sc)) { 1795 struct toedev *tod = sc->sc_tod; 1796 1797 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1798 1799 return (error); 1800 } 1801#endif 1802 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1803 } 1804#endif 1805 return (error); 1806} 1807 1808/* 1809 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1810 * that exceed the capacity of the syncache by avoiding the storage of any 1811 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1812 * attacks where the attacker does not have access to our responses. 1813 * 1814 * Syncookies encode and include all necessary information about the 1815 * connection setup within the SYN|ACK that we send back. That way we 1816 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1817 * (if ever). Normally the syncache and syncookies are running in parallel 1818 * with the latter taking over when the former is exhausted. When matching 1819 * syncache entry is found the syncookie is ignored. 1820 * 1821 * The only reliable information persisting the 3WHS is our initial sequence 1822 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1823 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1824 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1825 * returns and signifies a legitimate connection if it matches the ACK. 1826 * 1827 * The available space of 32 bits to store the hash and to encode the SYN 1828 * option information is very tight and we should have at least 24 bits for 1829 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1830 * 1831 * SYN option information we have to encode to fully restore a connection: 1832 * MSS: is imporant to chose an optimal segment size to avoid IP level 1833 * fragmentation along the path. The common MSS values can be encoded 1834 * in a 3-bit table. Uncommon values are captured by the next lower value 1835 * in the table leading to a slight increase in packetization overhead. 1836 * WSCALE: is necessary to allow large windows to be used for high delay- 1837 * bandwidth product links. Not scaling the window when it was initially 1838 * negotiated is bad for performance as lack of scaling further decreases 1839 * the apparent available send window. We only need to encode the WSCALE 1840 * we received from the remote end. Our end can be recalculated at any 1841 * time. The common WSCALE values can be encoded in a 3-bit table. 1842 * Uncommon values are captured by the next lower value in the table 1843 * making us under-estimate the available window size halving our 1844 * theoretically possible maximum throughput for that connection. 1845 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1846 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1847 * that are included in all segments on a connection. We enable them when 1848 * the ACK has them. 1849 * 1850 * Security of syncookies and attack vectors: 1851 * 1852 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1853 * together with the gloabl secret to make it unique per connection attempt. 1854 * Thus any change of any of those parameters results in a different MAC output 1855 * in an unpredictable way unless a collision is encountered. 24 bits of the 1856 * MAC are embedded into the ISS. 1857 * 1858 * To prevent replay attacks two rotating global secrets are updated with a 1859 * new random value every 15 seconds. The life-time of a syncookie is thus 1860 * 15-30 seconds. 1861 * 1862 * Vector 1: Attacking the secret. This requires finding a weakness in the 1863 * MAC itself or the way it is used here. The attacker can do a chosen plain 1864 * text attack by varying and testing the all parameters under his control. 1865 * The strength depends on the size and randomness of the secret, and the 1866 * cryptographic security of the MAC function. Due to the constant updating 1867 * of the secret the attacker has at most 29.999 seconds to find the secret 1868 * and launch spoofed connections. After that he has to start all over again. 1869 * 1870 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1871 * size an average of 4,823 attempts are required for a 50% chance of success 1872 * to spoof a single syncookie (birthday collision paradox). However the 1873 * attacker is blind and doesn't know if one of his attempts succeeded unless 1874 * he has a side channel to interfere success from. A single connection setup 1875 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1876 * This many attempts are required for each one blind spoofed connection. For 1877 * every additional spoofed connection he has to launch another N attempts. 1878 * Thus for a sustained rate 100 spoofed connections per second approximately 1879 * 1,800,000 packets per second would have to be sent. 1880 * 1881 * NB: The MAC function should be fast so that it doesn't become a CPU 1882 * exhaustion attack vector itself. 1883 * 1884 * References: 1885 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1886 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1887 * http://cr.yp.to/syncookies.html (overview) 1888 * http://cr.yp.to/syncookies/archive (details) 1889 * 1890 * 1891 * Schematic construction of a syncookie enabled Initial Sequence Number: 1892 * 0 1 2 3 1893 * 12345678901234567890123456789012 1894 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1895 * 1896 * x 24 MAC (truncated) 1897 * W 3 Send Window Scale index 1898 * M 3 MSS index 1899 * S 1 SACK permitted 1900 * P 1 Odd/even secret 1901 */ 1902 1903/* 1904 * Distribution and probability of certain MSS values. Those in between are 1905 * rounded down to the next lower one. 1906 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1907 * .2% .3% 5% 7% 7% 20% 15% 45% 1908 */ 1909static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1910 1911/* 1912 * Distribution and probability of certain WSCALE values. We have to map the 1913 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1914 * bits based on prevalence of certain values. Where we don't have an exact 1915 * match for are rounded down to the next lower one letting us under-estimate 1916 * the true available window. At the moment this would happen only for the 1917 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1918 * and window size). The absence of the WSCALE option (no scaling in either 1919 * direction) is encoded with index zero. 1920 * [WSCALE values histograms, Allman, 2012] 1921 * X 10 10 35 5 6 14 10% by host 1922 * X 11 4 5 5 18 49 3% by connections 1923 */ 1924static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1925 1926/* 1927 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1928 * and good cryptographic properties. 1929 */ 1930static uint32_t 1931syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1932 uint8_t *secbits, uintptr_t secmod) 1933{ 1934 SIPHASH_CTX ctx; 1935 uint32_t siphash[2]; 1936 1937 SipHash24_Init(&ctx); 1938 SipHash_SetKey(&ctx, secbits); 1939 switch (inc->inc_flags & INC_ISIPV6) { 1940#ifdef INET 1941 case 0: 1942 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1943 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1944 break; 1945#endif 1946#ifdef INET6 1947 case INC_ISIPV6: 1948 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1949 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1950 break; 1951#endif 1952 } 1953 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1954 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1955 SipHash_Update(&ctx, &irs, sizeof(irs)); 1956 SipHash_Update(&ctx, &flags, sizeof(flags)); 1957 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1958 SipHash_Final((u_int8_t *)&siphash, &ctx); 1959 1960 return (siphash[0] ^ siphash[1]); 1961} 1962 1963static tcp_seq 1964syncookie_generate(struct syncache_head *sch, struct syncache *sc) 1965{ 1966 u_int i, secbit, wscale; 1967 uint32_t iss, hash; 1968 uint8_t *secbits; 1969 union syncookie cookie; 1970 1971 SCH_LOCK_ASSERT(sch); 1972 1973 cookie.cookie = 0; 1974 1975 /* Map our computed MSS into the 3-bit index. */ 1976 for (i = nitems(tcp_sc_msstab) - 1; 1977 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 1978 i--) 1979 ; 1980 cookie.flags.mss_idx = i; 1981 1982 /* 1983 * Map the send window scale into the 3-bit index but only if 1984 * the wscale option was received. 1985 */ 1986 if (sc->sc_flags & SCF_WINSCALE) { 1987 wscale = sc->sc_requested_s_scale; 1988 for (i = nitems(tcp_sc_wstab) - 1; 1989 tcp_sc_wstab[i] > wscale && i > 0; 1990 i--) 1991 ; 1992 cookie.flags.wscale_idx = i; 1993 } 1994 1995 /* Can we do SACK? */ 1996 if (sc->sc_flags & SCF_SACK) 1997 cookie.flags.sack_ok = 1; 1998 1999 /* Which of the two secrets to use. */ 2000 secbit = sch->sch_sc->secret.oddeven & 0x1; 2001 cookie.flags.odd_even = secbit; 2002 2003 secbits = sch->sch_sc->secret.key[secbit]; 2004 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2005 (uintptr_t)sch); 2006 2007 /* 2008 * Put the flags into the hash and XOR them to get better ISS number 2009 * variance. This doesn't enhance the cryptographic strength and is 2010 * done to prevent the 8 cookie bits from showing up directly on the 2011 * wire. 2012 */ 2013 iss = hash & ~0xff; 2014 iss |= cookie.cookie ^ (hash >> 24); 2015 2016 /* Randomize the timestamp. */ 2017 if (sc->sc_flags & SCF_TIMESTAMP) { 2018 sc->sc_ts = arc4random(); 2019 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 2020 } 2021 2022 TCPSTAT_INC(tcps_sc_sendcookie); 2023 return (iss); 2024} 2025 2026static struct syncache * 2027syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2028 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2029 struct socket *lso) 2030{ 2031 uint32_t hash; 2032 uint8_t *secbits; 2033 tcp_seq ack, seq; 2034 int wnd, wscale = 0; 2035 union syncookie cookie; 2036 2037 SCH_LOCK_ASSERT(sch); 2038 2039 /* 2040 * Pull information out of SYN-ACK/ACK and revert sequence number 2041 * advances. 2042 */ 2043 ack = th->th_ack - 1; 2044 seq = th->th_seq - 1; 2045 2046 /* 2047 * Unpack the flags containing enough information to restore the 2048 * connection. 2049 */ 2050 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2051 2052 /* Which of the two secrets to use. */ 2053 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 2054 2055 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2056 2057 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2058 if ((ack & ~0xff) != (hash & ~0xff)) 2059 return (NULL); 2060 2061 /* Fill in the syncache values. */ 2062 sc->sc_flags = 0; 2063 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2064 sc->sc_ipopts = NULL; 2065 2066 sc->sc_irs = seq; 2067 sc->sc_iss = ack; 2068 2069 switch (inc->inc_flags & INC_ISIPV6) { 2070#ifdef INET 2071 case 0: 2072 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2073 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2074 break; 2075#endif 2076#ifdef INET6 2077 case INC_ISIPV6: 2078 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2079 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 2080 break; 2081#endif 2082 } 2083 2084 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2085 2086 /* We can simply recompute receive window scale we sent earlier. */ 2087 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2088 wscale++; 2089 2090 /* Only use wscale if it was enabled in the orignal SYN. */ 2091 if (cookie.flags.wscale_idx > 0) { 2092 sc->sc_requested_r_scale = wscale; 2093 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2094 sc->sc_flags |= SCF_WINSCALE; 2095 } 2096 2097 wnd = sbspace(&lso->so_rcv); 2098 wnd = imax(wnd, 0); 2099 wnd = imin(wnd, TCP_MAXWIN); 2100 sc->sc_wnd = wnd; 2101 2102 if (cookie.flags.sack_ok) 2103 sc->sc_flags |= SCF_SACK; 2104 2105 if (to->to_flags & TOF_TS) { 2106 sc->sc_flags |= SCF_TIMESTAMP; 2107 sc->sc_tsreflect = to->to_tsval; 2108 sc->sc_ts = to->to_tsecr; 2109 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 2110 } 2111 2112 if (to->to_flags & TOF_SIGNATURE) 2113 sc->sc_flags |= SCF_SIGNATURE; 2114 2115 sc->sc_rxmits = 0; 2116 2117 TCPSTAT_INC(tcps_sc_recvcookie); 2118 return (sc); 2119} 2120 2121#ifdef INVARIANTS 2122static int 2123syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2124 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2125 struct socket *lso) 2126{ 2127 struct syncache scs, *scx; 2128 char *s; 2129 2130 bzero(&scs, sizeof(scs)); 2131 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2132 2133 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2134 return (0); 2135 2136 if (scx != NULL) { 2137 if (sc->sc_peer_mss != scx->sc_peer_mss) 2138 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2139 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2140 2141 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2142 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2143 s, __func__, sc->sc_requested_r_scale, 2144 scx->sc_requested_r_scale); 2145 2146 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2147 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2148 s, __func__, sc->sc_requested_s_scale, 2149 scx->sc_requested_s_scale); 2150 2151 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2152 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2153 } 2154 2155 if (s != NULL) 2156 free(s, M_TCPLOG); 2157 return (0); 2158} 2159#endif /* INVARIANTS */ 2160 2161static void 2162syncookie_reseed(void *arg) 2163{ 2164 struct tcp_syncache *sc = arg; 2165 uint8_t *secbits; 2166 int secbit; 2167 2168 /* 2169 * Reseeding the secret doesn't have to be protected by a lock. 2170 * It only must be ensured that the new random values are visible 2171 * to all CPUs in a SMP environment. The atomic with release 2172 * semantics ensures that. 2173 */ 2174 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2175 secbits = sc->secret.key[secbit]; 2176 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2177 atomic_add_rel_int(&sc->secret.oddeven, 1); 2178 2179 /* Reschedule ourself. */ 2180 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2181} 2182 2183/* 2184 * Exports the syncache entries to userland so that netstat can display 2185 * them alongside the other sockets. This function is intended to be 2186 * called only from tcp_pcblist. 2187 * 2188 * Due to concurrency on an active system, the number of pcbs exported 2189 * may have no relation to max_pcbs. max_pcbs merely indicates the 2190 * amount of space the caller allocated for this function to use. 2191 */ 2192int 2193syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2194{ 2195 struct xtcpcb xt; 2196 struct syncache *sc; 2197 struct syncache_head *sch; 2198 int count, error, i; 2199 2200 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2201 sch = &V_tcp_syncache.hashbase[i]; 2202 SCH_LOCK(sch); 2203 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2204 if (count >= max_pcbs) { 2205 SCH_UNLOCK(sch); 2206 goto exit; 2207 } 2208 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2209 continue; 2210 bzero(&xt, sizeof(xt)); 2211 xt.xt_len = sizeof(xt); 2212 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2213 xt.xt_inp.inp_vflag = INP_IPV6; 2214 else 2215 xt.xt_inp.inp_vflag = INP_IPV4; 2216 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 2217 xt.xt_tp.t_inpcb = &xt.xt_inp; 2218 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 2219 xt.xt_socket.xso_protocol = IPPROTO_TCP; 2220 xt.xt_socket.xso_len = sizeof (struct xsocket); 2221 xt.xt_socket.so_type = SOCK_STREAM; 2222 xt.xt_socket.so_state = SS_ISCONNECTING; 2223 error = SYSCTL_OUT(req, &xt, sizeof xt); 2224 if (error) { 2225 SCH_UNLOCK(sch); 2226 goto exit; 2227 } 2228 count++; 2229 } 2230 SCH_UNLOCK(sch); 2231 } 2232exit: 2233 *pcbs_exported = count; 2234 return error; 2235} 2236