tcp_syncache.c revision 309108
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 309108 2016-11-24 14:48:46Z jch $"); 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 } 281 282 /* Create the syncache entry zone. */ 283 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 284 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 285 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 286 V_tcp_syncache.cache_limit); 287 288 /* Start the SYN cookie reseeder callout. */ 289 callout_init(&V_tcp_syncache.secret.reseed, 1); 290 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 291 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 292 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 293 syncookie_reseed, &V_tcp_syncache); 294} 295 296#ifdef VIMAGE 297void 298syncache_destroy(void) 299{ 300 struct syncache_head *sch; 301 struct syncache *sc, *nsc; 302 int i; 303 304 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 305 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 306 307 sch = &V_tcp_syncache.hashbase[i]; 308 callout_drain(&sch->sch_timer); 309 310 SCH_LOCK(sch); 311 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 312 syncache_drop(sc, sch); 313 SCH_UNLOCK(sch); 314 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 315 ("%s: sch->sch_bucket not empty", __func__)); 316 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 317 __func__, sch->sch_length)); 318 mtx_destroy(&sch->sch_mtx); 319 } 320 321 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 322 ("%s: cache_count not 0", __func__)); 323 324 /* Free the allocated global resources. */ 325 uma_zdestroy(V_tcp_syncache.zone); 326 free(V_tcp_syncache.hashbase, M_SYNCACHE); 327 328 callout_drain(&V_tcp_syncache.secret.reseed); 329} 330#endif 331 332static int 333syncache_sysctl_count(SYSCTL_HANDLER_ARGS) 334{ 335 int count; 336 337 count = uma_zone_get_cur(V_tcp_syncache.zone); 338 return (sysctl_handle_int(oidp, &count, 0, req)); 339} 340 341/* 342 * Inserts a syncache entry into the specified bucket row. 343 * Locks and unlocks the syncache_head autonomously. 344 */ 345static void 346syncache_insert(struct syncache *sc, struct syncache_head *sch) 347{ 348 struct syncache *sc2; 349 350 SCH_LOCK(sch); 351 352 /* 353 * Make sure that we don't overflow the per-bucket limit. 354 * If the bucket is full, toss the oldest element. 355 */ 356 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 357 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 358 ("sch->sch_length incorrect")); 359 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 360 syncache_drop(sc2, sch); 361 TCPSTAT_INC(tcps_sc_bucketoverflow); 362 } 363 364 /* Put it into the bucket. */ 365 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 366 sch->sch_length++; 367 368#ifdef TCP_OFFLOAD 369 if (ADDED_BY_TOE(sc)) { 370 struct toedev *tod = sc->sc_tod; 371 372 tod->tod_syncache_added(tod, sc->sc_todctx); 373 } 374#endif 375 376 /* Reinitialize the bucket row's timer. */ 377 if (sch->sch_length == 1) 378 sch->sch_nextc = ticks + INT_MAX; 379 syncache_timeout(sc, sch, 1); 380 381 SCH_UNLOCK(sch); 382 383 TCPSTAT_INC(tcps_sc_added); 384} 385 386/* 387 * Remove and free entry from syncache bucket row. 388 * Expects locked syncache head. 389 */ 390static void 391syncache_drop(struct syncache *sc, struct syncache_head *sch) 392{ 393 394 SCH_LOCK_ASSERT(sch); 395 396 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 397 sch->sch_length--; 398 399#ifdef TCP_OFFLOAD 400 if (ADDED_BY_TOE(sc)) { 401 struct toedev *tod = sc->sc_tod; 402 403 tod->tod_syncache_removed(tod, sc->sc_todctx); 404 } 405#endif 406 407 syncache_free(sc); 408} 409 410/* 411 * Engage/reengage time on bucket row. 412 */ 413static void 414syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 415{ 416 sc->sc_rxttime = ticks + 417 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]); 418 sc->sc_rxmits++; 419 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 420 sch->sch_nextc = sc->sc_rxttime; 421 if (docallout) 422 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 423 syncache_timer, (void *)sch); 424 } 425} 426 427/* 428 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 429 * If we have retransmitted an entry the maximum number of times, expire it. 430 * One separate timer for each bucket row. 431 */ 432static void 433syncache_timer(void *xsch) 434{ 435 struct syncache_head *sch = (struct syncache_head *)xsch; 436 struct syncache *sc, *nsc; 437 int tick = ticks; 438 char *s; 439 440 CURVNET_SET(sch->sch_sc->vnet); 441 442 /* NB: syncache_head has already been locked by the callout. */ 443 SCH_LOCK_ASSERT(sch); 444 445 /* 446 * In the following cycle we may remove some entries and/or 447 * advance some timeouts, so re-initialize the bucket timer. 448 */ 449 sch->sch_nextc = tick + INT_MAX; 450 451 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 452 /* 453 * We do not check if the listen socket still exists 454 * and accept the case where the listen socket may be 455 * gone by the time we resend the SYN/ACK. We do 456 * not expect this to happens often. If it does, 457 * then the RST will be sent by the time the remote 458 * host does the SYN/ACK->ACK. 459 */ 460 if (TSTMP_GT(sc->sc_rxttime, tick)) { 461 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 462 sch->sch_nextc = sc->sc_rxttime; 463 continue; 464 } 465 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 466 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 467 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 468 "giving up and removing syncache entry\n", 469 s, __func__); 470 free(s, M_TCPLOG); 471 } 472 syncache_drop(sc, sch); 473 TCPSTAT_INC(tcps_sc_stale); 474 continue; 475 } 476 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 477 log(LOG_DEBUG, "%s; %s: Response timeout, " 478 "retransmitting (%u) SYN|ACK\n", 479 s, __func__, sc->sc_rxmits); 480 free(s, M_TCPLOG); 481 } 482 483 (void) syncache_respond(sc, NULL); 484 TCPSTAT_INC(tcps_sc_retransmitted); 485 syncache_timeout(sc, sch, 0); 486 } 487 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 488 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 489 syncache_timer, (void *)(sch)); 490 CURVNET_RESTORE(); 491} 492 493/* 494 * Find an entry in the syncache. 495 * Returns always with locked syncache_head plus a matching entry or NULL. 496 */ 497static struct syncache * 498syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 499{ 500 struct syncache *sc; 501 struct syncache_head *sch; 502 503#ifdef INET6 504 if (inc->inc_flags & INC_ISIPV6) { 505 sch = &V_tcp_syncache.hashbase[ 506 SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)]; 507 *schp = sch; 508 509 SCH_LOCK(sch); 510 511 /* Circle through bucket row to find matching entry. */ 512 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 513 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 514 return (sc); 515 } 516 } else 517#endif 518 { 519 sch = &V_tcp_syncache.hashbase[ 520 SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)]; 521 *schp = sch; 522 523 SCH_LOCK(sch); 524 525 /* Circle through bucket row to find matching entry. */ 526 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 527#ifdef INET6 528 if (sc->sc_inc.inc_flags & INC_ISIPV6) 529 continue; 530#endif 531 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 532 return (sc); 533 } 534 } 535 SCH_LOCK_ASSERT(*schp); 536 return (NULL); /* always returns with locked sch */ 537} 538 539/* 540 * This function is called when we get a RST for a 541 * non-existent connection, so that we can see if the 542 * connection is in the syn cache. If it is, zap it. 543 */ 544void 545syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 546{ 547 struct syncache *sc; 548 struct syncache_head *sch; 549 char *s = NULL; 550 551 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 552 SCH_LOCK_ASSERT(sch); 553 554 /* 555 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 556 * See RFC 793 page 65, section SEGMENT ARRIVES. 557 */ 558 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 559 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 560 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 561 "FIN flag set, segment ignored\n", s, __func__); 562 TCPSTAT_INC(tcps_badrst); 563 goto done; 564 } 565 566 /* 567 * No corresponding connection was found in syncache. 568 * If syncookies are enabled and possibly exclusively 569 * used, or we are under memory pressure, a valid RST 570 * may not find a syncache entry. In that case we're 571 * done and no SYN|ACK retransmissions will happen. 572 * Otherwise the RST was misdirected or spoofed. 573 */ 574 if (sc == NULL) { 575 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 576 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 577 "syncache entry (possibly syncookie only), " 578 "segment ignored\n", s, __func__); 579 TCPSTAT_INC(tcps_badrst); 580 goto done; 581 } 582 583 /* 584 * If the RST bit is set, check the sequence number to see 585 * if this is a valid reset segment. 586 * RFC 793 page 37: 587 * In all states except SYN-SENT, all reset (RST) segments 588 * are validated by checking their SEQ-fields. A reset is 589 * valid if its sequence number is in the window. 590 * 591 * The sequence number in the reset segment is normally an 592 * echo of our outgoing acknowlegement numbers, but some hosts 593 * send a reset with the sequence number at the rightmost edge 594 * of our receive window, and we have to handle this case. 595 */ 596 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 597 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 598 syncache_drop(sc, sch); 599 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 600 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 601 "connection attempt aborted by remote endpoint\n", 602 s, __func__); 603 TCPSTAT_INC(tcps_sc_reset); 604 } else { 605 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 606 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 607 "IRS %u (+WND %u), segment ignored\n", 608 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 609 TCPSTAT_INC(tcps_badrst); 610 } 611 612done: 613 if (s != NULL) 614 free(s, M_TCPLOG); 615 SCH_UNLOCK(sch); 616} 617 618void 619syncache_badack(struct in_conninfo *inc) 620{ 621 struct syncache *sc; 622 struct syncache_head *sch; 623 624 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 625 SCH_LOCK_ASSERT(sch); 626 if (sc != NULL) { 627 syncache_drop(sc, sch); 628 TCPSTAT_INC(tcps_sc_badack); 629 } 630 SCH_UNLOCK(sch); 631} 632 633void 634syncache_unreach(struct in_conninfo *inc, struct tcphdr *th) 635{ 636 struct syncache *sc; 637 struct syncache_head *sch; 638 639 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 640 SCH_LOCK_ASSERT(sch); 641 if (sc == NULL) 642 goto done; 643 644 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 645 if (ntohl(th->th_seq) != sc->sc_iss) 646 goto done; 647 648 /* 649 * If we've rertransmitted 3 times and this is our second error, 650 * we remove the entry. Otherwise, we allow it to continue on. 651 * This prevents us from incorrectly nuking an entry during a 652 * spurious network outage. 653 * 654 * See tcp_notify(). 655 */ 656 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 657 sc->sc_flags |= SCF_UNREACH; 658 goto done; 659 } 660 syncache_drop(sc, sch); 661 TCPSTAT_INC(tcps_sc_unreach); 662done: 663 SCH_UNLOCK(sch); 664} 665 666/* 667 * Build a new TCP socket structure from a syncache entry. 668 * 669 * On success return the newly created socket with its underlying inp locked. 670 */ 671static struct socket * 672syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 673{ 674 struct inpcb *inp = NULL; 675 struct socket *so; 676 struct tcpcb *tp; 677 int error; 678 char *s; 679 680 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 681 682 /* 683 * Ok, create the full blown connection, and set things up 684 * as they would have been set up if we had created the 685 * connection when the SYN arrived. If we can't create 686 * the connection, abort it. 687 */ 688 so = sonewconn(lso, 0); 689 if (so == NULL) { 690 /* 691 * Drop the connection; we will either send a RST or 692 * have the peer retransmit its SYN again after its 693 * RTO and try again. 694 */ 695 TCPSTAT_INC(tcps_listendrop); 696 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 697 log(LOG_DEBUG, "%s; %s: Socket create failed " 698 "due to limits or memory shortage\n", 699 s, __func__); 700 free(s, M_TCPLOG); 701 } 702 goto abort2; 703 } 704#ifdef MAC 705 mac_socketpeer_set_from_mbuf(m, so); 706#endif 707 708 inp = sotoinpcb(so); 709 inp->inp_inc.inc_fibnum = so->so_fibnum; 710 INP_WLOCK(inp); 711 /* 712 * Exclusive pcbinfo lock is not required in syncache socket case even 713 * if two inpcb locks can be acquired simultaneously: 714 * - the inpcb in LISTEN state, 715 * - the newly created inp. 716 * 717 * In this case, an inp cannot be at same time in LISTEN state and 718 * just created by an accept() call. 719 */ 720 INP_HASH_WLOCK(&V_tcbinfo); 721 722 /* Insert new socket into PCB hash list. */ 723 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 724#ifdef INET6 725 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 726 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 727 } else { 728 inp->inp_vflag &= ~INP_IPV6; 729 inp->inp_vflag |= INP_IPV4; 730#endif 731 inp->inp_laddr = sc->sc_inc.inc_laddr; 732#ifdef INET6 733 } 734#endif 735 736 /* 737 * If there's an mbuf and it has a flowid, then let's initialise the 738 * inp with that particular flowid. 739 */ 740 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 741 inp->inp_flowid = m->m_pkthdr.flowid; 742 inp->inp_flowtype = M_HASHTYPE_GET(m); 743 } 744 745 /* 746 * Install in the reservation hash table for now, but don't yet 747 * install a connection group since the full 4-tuple isn't yet 748 * configured. 749 */ 750 inp->inp_lport = sc->sc_inc.inc_lport; 751 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) { 752 /* 753 * Undo the assignments above if we failed to 754 * put the PCB on the hash lists. 755 */ 756#ifdef INET6 757 if (sc->sc_inc.inc_flags & INC_ISIPV6) 758 inp->in6p_laddr = in6addr_any; 759 else 760#endif 761 inp->inp_laddr.s_addr = INADDR_ANY; 762 inp->inp_lport = 0; 763 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 764 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 765 "with error %i\n", 766 s, __func__, error); 767 free(s, M_TCPLOG); 768 } 769 INP_HASH_WUNLOCK(&V_tcbinfo); 770 goto abort; 771 } 772#ifdef IPSEC 773 /* Copy old policy into new socket's. */ 774 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp)) 775 printf("syncache_socket: could not copy policy\n"); 776#endif 777#ifdef INET6 778 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 779 struct inpcb *oinp = sotoinpcb(lso); 780 struct in6_addr laddr6; 781 struct sockaddr_in6 sin6; 782 /* 783 * Inherit socket options from the listening socket. 784 * Note that in6p_inputopts are not (and should not be) 785 * copied, since it stores previously received options and is 786 * used to detect if each new option is different than the 787 * previous one and hence should be passed to a user. 788 * If we copied in6p_inputopts, a user would not be able to 789 * receive options just after calling the accept system call. 790 */ 791 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 792 if (oinp->in6p_outputopts) 793 inp->in6p_outputopts = 794 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 795 796 sin6.sin6_family = AF_INET6; 797 sin6.sin6_len = sizeof(sin6); 798 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 799 sin6.sin6_port = sc->sc_inc.inc_fport; 800 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 801 laddr6 = inp->in6p_laddr; 802 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 803 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 804 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, 805 thread0.td_ucred, m)) != 0) { 806 inp->in6p_laddr = laddr6; 807 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 808 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 809 "with error %i\n", 810 s, __func__, error); 811 free(s, M_TCPLOG); 812 } 813 INP_HASH_WUNLOCK(&V_tcbinfo); 814 goto abort; 815 } 816 /* Override flowlabel from in6_pcbconnect. */ 817 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 818 inp->inp_flow |= sc->sc_flowlabel; 819 } 820#endif /* INET6 */ 821#if defined(INET) && defined(INET6) 822 else 823#endif 824#ifdef INET 825 { 826 struct in_addr laddr; 827 struct sockaddr_in sin; 828 829 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 830 831 if (inp->inp_options == NULL) { 832 inp->inp_options = sc->sc_ipopts; 833 sc->sc_ipopts = NULL; 834 } 835 836 sin.sin_family = AF_INET; 837 sin.sin_len = sizeof(sin); 838 sin.sin_addr = sc->sc_inc.inc_faddr; 839 sin.sin_port = sc->sc_inc.inc_fport; 840 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 841 laddr = inp->inp_laddr; 842 if (inp->inp_laddr.s_addr == INADDR_ANY) 843 inp->inp_laddr = sc->sc_inc.inc_laddr; 844 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, 845 thread0.td_ucred, m)) != 0) { 846 inp->inp_laddr = laddr; 847 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 848 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 849 "with error %i\n", 850 s, __func__, error); 851 free(s, M_TCPLOG); 852 } 853 INP_HASH_WUNLOCK(&V_tcbinfo); 854 goto abort; 855 } 856 } 857#endif /* INET */ 858 INP_HASH_WUNLOCK(&V_tcbinfo); 859 tp = intotcpcb(inp); 860 tcp_state_change(tp, TCPS_SYN_RECEIVED); 861 tp->iss = sc->sc_iss; 862 tp->irs = sc->sc_irs; 863 tcp_rcvseqinit(tp); 864 tcp_sendseqinit(tp); 865 tp->snd_wl1 = sc->sc_irs; 866 tp->snd_max = tp->iss + 1; 867 tp->snd_nxt = tp->iss + 1; 868 tp->rcv_up = sc->sc_irs + 1; 869 tp->rcv_wnd = sc->sc_wnd; 870 tp->rcv_adv += tp->rcv_wnd; 871 tp->last_ack_sent = tp->rcv_nxt; 872 873 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 874 if (sc->sc_flags & SCF_NOOPT) 875 tp->t_flags |= TF_NOOPT; 876 else { 877 if (sc->sc_flags & SCF_WINSCALE) { 878 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 879 tp->snd_scale = sc->sc_requested_s_scale; 880 tp->request_r_scale = sc->sc_requested_r_scale; 881 } 882 if (sc->sc_flags & SCF_TIMESTAMP) { 883 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 884 tp->ts_recent = sc->sc_tsreflect; 885 tp->ts_recent_age = tcp_ts_getticks(); 886 tp->ts_offset = sc->sc_tsoff; 887 } 888#ifdef TCP_SIGNATURE 889 if (sc->sc_flags & SCF_SIGNATURE) 890 tp->t_flags |= TF_SIGNATURE; 891#endif 892 if (sc->sc_flags & SCF_SACK) 893 tp->t_flags |= TF_SACK_PERMIT; 894 } 895 896 if (sc->sc_flags & SCF_ECN) 897 tp->t_flags |= TF_ECN_PERMIT; 898 899 /* 900 * Set up MSS and get cached values from tcp_hostcache. 901 * This might overwrite some of the defaults we just set. 902 */ 903 tcp_mss(tp, sc->sc_peer_mss); 904 905 /* 906 * If the SYN,ACK was retransmitted, indicate that CWND to be 907 * limited to one segment in cc_conn_init(). 908 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 909 */ 910 if (sc->sc_rxmits > 1) 911 tp->snd_cwnd = 1; 912 913#ifdef TCP_OFFLOAD 914 /* 915 * Allow a TOE driver to install its hooks. Note that we hold the 916 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 917 * new connection before the TOE driver has done its thing. 918 */ 919 if (ADDED_BY_TOE(sc)) { 920 struct toedev *tod = sc->sc_tod; 921 922 tod->tod_offload_socket(tod, sc->sc_todctx, so); 923 } 924#endif 925 /* 926 * Copy and activate timers. 927 */ 928 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 929 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 930 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 931 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 932 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 933 934 TCPSTAT_INC(tcps_accepts); 935 return (so); 936 937abort: 938 INP_WUNLOCK(inp); 939abort2: 940 if (so != NULL) 941 soabort(so); 942 return (NULL); 943} 944 945/* 946 * This function gets called when we receive an ACK for a 947 * socket in the LISTEN state. We look up the connection 948 * in the syncache, and if its there, we pull it out of 949 * the cache and turn it into a full-blown connection in 950 * the SYN-RECEIVED state. 951 * 952 * On syncache_socket() success the newly created socket 953 * has its underlying inp locked. 954 */ 955int 956syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 957 struct socket **lsop, struct mbuf *m) 958{ 959 struct syncache *sc; 960 struct syncache_head *sch; 961 struct syncache scs; 962 char *s; 963 964 /* 965 * Global TCP locks are held because we manipulate the PCB lists 966 * and create a new socket. 967 */ 968 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 969 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 970 ("%s: can handle only ACK", __func__)); 971 972 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 973 SCH_LOCK_ASSERT(sch); 974 975#ifdef INVARIANTS 976 /* 977 * Test code for syncookies comparing the syncache stored 978 * values with the reconstructed values from the cookie. 979 */ 980 if (sc != NULL) 981 syncookie_cmp(inc, sch, sc, th, to, *lsop); 982#endif 983 984 if (sc == NULL) { 985 /* 986 * There is no syncache entry, so see if this ACK is 987 * a returning syncookie. To do this, first: 988 * A. See if this socket has had a syncache entry dropped in 989 * the past. We don't want to accept a bogus syncookie 990 * if we've never received a SYN. 991 * B. check that the syncookie is valid. If it is, then 992 * cobble up a fake syncache entry, and return. 993 */ 994 if (!V_tcp_syncookies) { 995 SCH_UNLOCK(sch); 996 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 997 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 998 "segment rejected (syncookies disabled)\n", 999 s, __func__); 1000 goto failed; 1001 } 1002 bzero(&scs, sizeof(scs)); 1003 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); 1004 SCH_UNLOCK(sch); 1005 if (sc == NULL) { 1006 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1007 log(LOG_DEBUG, "%s; %s: Segment failed " 1008 "SYNCOOKIE authentication, segment rejected " 1009 "(probably spoofed)\n", s, __func__); 1010 goto failed; 1011 } 1012 } else { 1013 /* Pull out the entry to unlock the bucket row. */ 1014 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1015 sch->sch_length--; 1016#ifdef TCP_OFFLOAD 1017 if (ADDED_BY_TOE(sc)) { 1018 struct toedev *tod = sc->sc_tod; 1019 1020 tod->tod_syncache_removed(tod, sc->sc_todctx); 1021 } 1022#endif 1023 SCH_UNLOCK(sch); 1024 } 1025 1026 /* 1027 * Segment validation: 1028 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1029 */ 1030 if (th->th_ack != sc->sc_iss + 1) { 1031 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1032 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1033 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1034 goto failed; 1035 } 1036 1037 /* 1038 * The SEQ must fall in the window starting at the received 1039 * initial receive sequence number + 1 (the SYN). 1040 */ 1041 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1042 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1043 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1044 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1045 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1046 goto failed; 1047 } 1048 1049 /* 1050 * If timestamps were not negotiated during SYN/ACK they 1051 * must not appear on any segment during this session. 1052 */ 1053 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 1054 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1055 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 1056 "segment rejected\n", s, __func__); 1057 goto failed; 1058 } 1059 1060 /* 1061 * If timestamps were negotiated during SYN/ACK they should 1062 * appear on every segment during this session. 1063 * XXXAO: This is only informal as there have been unverified 1064 * reports of non-compliants stacks. 1065 */ 1066 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { 1067 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1068 log(LOG_DEBUG, "%s; %s: Timestamp missing, " 1069 "no action\n", s, __func__); 1070 free(s, M_TCPLOG); 1071 s = NULL; 1072 } 1073 } 1074 1075 /* 1076 * If timestamps were negotiated the reflected timestamp 1077 * must be equal to what we actually sent in the SYN|ACK. 1078 */ 1079 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) { 1080 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1081 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 1082 "segment rejected\n", 1083 s, __func__, to->to_tsecr, sc->sc_ts); 1084 goto failed; 1085 } 1086 1087 *lsop = syncache_socket(sc, *lsop, m); 1088 1089 if (*lsop == NULL) 1090 TCPSTAT_INC(tcps_sc_aborted); 1091 else 1092 TCPSTAT_INC(tcps_sc_completed); 1093 1094/* how do we find the inp for the new socket? */ 1095 if (sc != &scs) 1096 syncache_free(sc); 1097 return (1); 1098failed: 1099 if (sc != NULL && sc != &scs) 1100 syncache_free(sc); 1101 if (s != NULL) 1102 free(s, M_TCPLOG); 1103 *lsop = NULL; 1104 return (0); 1105} 1106 1107#ifdef TCP_RFC7413 1108static void 1109syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m, 1110 uint64_t response_cookie) 1111{ 1112 struct inpcb *inp; 1113 struct tcpcb *tp; 1114 unsigned int *pending_counter; 1115 1116 /* 1117 * Global TCP locks are held because we manipulate the PCB lists 1118 * and create a new socket. 1119 */ 1120 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1121 1122 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending; 1123 *lsop = syncache_socket(sc, *lsop, m); 1124 if (*lsop == NULL) { 1125 TCPSTAT_INC(tcps_sc_aborted); 1126 atomic_subtract_int(pending_counter, 1); 1127 } else { 1128 inp = sotoinpcb(*lsop); 1129 tp = intotcpcb(inp); 1130 tp->t_flags |= TF_FASTOPEN; 1131 tp->t_tfo_cookie = response_cookie; 1132 tp->snd_max = tp->iss; 1133 tp->snd_nxt = tp->iss; 1134 tp->t_tfo_pending = pending_counter; 1135 TCPSTAT_INC(tcps_sc_completed); 1136 } 1137} 1138#endif /* TCP_RFC7413 */ 1139 1140/* 1141 * Given a LISTEN socket and an inbound SYN request, add 1142 * this to the syn cache, and send back a segment: 1143 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1144 * to the source. 1145 * 1146 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1147 * Doing so would require that we hold onto the data and deliver it 1148 * to the application. However, if we are the target of a SYN-flood 1149 * DoS attack, an attacker could send data which would eventually 1150 * consume all available buffer space if it were ACKed. By not ACKing 1151 * the data, we avoid this DoS scenario. 1152 * 1153 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1154 * cookie is processed, V_tcp_fastopen_enabled set to true, and the 1155 * TCP_FASTOPEN socket option is set. In this case, a new socket is created 1156 * and returned via lsop, the mbuf is not freed so that tcp_input() can 1157 * queue its data to the socket, and 1 is returned to indicate the 1158 * TFO-socket-creation path was taken. 1159 */ 1160int 1161syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1162 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod, 1163 void *todctx) 1164{ 1165 struct tcpcb *tp; 1166 struct socket *so; 1167 struct syncache *sc = NULL; 1168 struct syncache_head *sch; 1169 struct mbuf *ipopts = NULL; 1170 u_int ltflags; 1171 int win, sb_hiwat, ip_ttl, ip_tos; 1172 char *s; 1173 int rv = 0; 1174#ifdef INET6 1175 int autoflowlabel = 0; 1176#endif 1177#ifdef MAC 1178 struct label *maclabel; 1179#endif 1180 struct syncache scs; 1181 struct ucred *cred; 1182#ifdef TCP_RFC7413 1183 uint64_t tfo_response_cookie; 1184 int tfo_cookie_valid = 0; 1185 int tfo_response_cookie_valid = 0; 1186#endif 1187 1188 INP_WLOCK_ASSERT(inp); /* listen socket */ 1189 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1190 ("%s: unexpected tcp flags", __func__)); 1191 1192 /* 1193 * Combine all so/tp operations very early to drop the INP lock as 1194 * soon as possible. 1195 */ 1196 so = *lsop; 1197 tp = sototcpcb(so); 1198 cred = crhold(so->so_cred); 1199 1200#ifdef INET6 1201 if ((inc->inc_flags & INC_ISIPV6) && 1202 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1203 autoflowlabel = 1; 1204#endif 1205 ip_ttl = inp->inp_ip_ttl; 1206 ip_tos = inp->inp_ip_tos; 1207 win = sbspace(&so->so_rcv); 1208 sb_hiwat = so->so_rcv.sb_hiwat; 1209 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1210 1211#ifdef TCP_RFC7413 1212 if (V_tcp_fastopen_enabled && (tp->t_flags & TF_FASTOPEN) && 1213 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) { 1214 /* 1215 * Limit the number of pending TFO connections to 1216 * approximately half of the queue limit. This prevents TFO 1217 * SYN floods from starving the service by filling the 1218 * listen queue with bogus TFO connections. 1219 */ 1220 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1221 (so->so_qlimit / 2)) { 1222 int result; 1223 1224 result = tcp_fastopen_check_cookie(inc, 1225 to->to_tfo_cookie, to->to_tfo_len, 1226 &tfo_response_cookie); 1227 tfo_cookie_valid = (result > 0); 1228 tfo_response_cookie_valid = (result >= 0); 1229 } else 1230 atomic_subtract_int(tp->t_tfo_pending, 1); 1231 } 1232#endif 1233 1234 /* By the time we drop the lock these should no longer be used. */ 1235 so = NULL; 1236 tp = NULL; 1237 1238#ifdef MAC 1239 if (mac_syncache_init(&maclabel) != 0) { 1240 INP_WUNLOCK(inp); 1241 goto done; 1242 } else 1243 mac_syncache_create(maclabel, inp); 1244#endif 1245#ifdef TCP_RFC7413 1246 if (!tfo_cookie_valid) 1247#endif 1248 INP_WUNLOCK(inp); 1249 1250 /* 1251 * Remember the IP options, if any. 1252 */ 1253#ifdef INET6 1254 if (!(inc->inc_flags & INC_ISIPV6)) 1255#endif 1256#ifdef INET 1257 ipopts = (m) ? ip_srcroute(m) : NULL; 1258#else 1259 ipopts = NULL; 1260#endif 1261 1262 /* 1263 * See if we already have an entry for this connection. 1264 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1265 * 1266 * XXX: should the syncache be re-initialized with the contents 1267 * of the new SYN here (which may have different options?) 1268 * 1269 * XXX: We do not check the sequence number to see if this is a 1270 * real retransmit or a new connection attempt. The question is 1271 * how to handle such a case; either ignore it as spoofed, or 1272 * drop the current entry and create a new one? 1273 */ 1274 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1275 SCH_LOCK_ASSERT(sch); 1276 if (sc != NULL) { 1277#ifdef TCP_RFC7413 1278 if (tfo_cookie_valid) 1279 INP_WUNLOCK(inp); 1280#endif 1281 TCPSTAT_INC(tcps_sc_dupsyn); 1282 if (ipopts) { 1283 /* 1284 * If we were remembering a previous source route, 1285 * forget it and use the new one we've been given. 1286 */ 1287 if (sc->sc_ipopts) 1288 (void) m_free(sc->sc_ipopts); 1289 sc->sc_ipopts = ipopts; 1290 } 1291 /* 1292 * Update timestamp if present. 1293 */ 1294 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1295 sc->sc_tsreflect = to->to_tsval; 1296 else 1297 sc->sc_flags &= ~SCF_TIMESTAMP; 1298#ifdef MAC 1299 /* 1300 * Since we have already unconditionally allocated label 1301 * storage, free it up. The syncache entry will already 1302 * have an initialized label we can use. 1303 */ 1304 mac_syncache_destroy(&maclabel); 1305#endif 1306 /* Retransmit SYN|ACK and reset retransmit count. */ 1307 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1308 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1309 "resetting timer and retransmitting SYN|ACK\n", 1310 s, __func__); 1311 free(s, M_TCPLOG); 1312 } 1313 if (syncache_respond(sc, m) == 0) { 1314 sc->sc_rxmits = 0; 1315 syncache_timeout(sc, sch, 1); 1316 TCPSTAT_INC(tcps_sndacks); 1317 TCPSTAT_INC(tcps_sndtotal); 1318 } 1319 SCH_UNLOCK(sch); 1320 goto done; 1321 } 1322 1323#ifdef TCP_RFC7413 1324 if (tfo_cookie_valid) { 1325 bzero(&scs, sizeof(scs)); 1326 sc = &scs; 1327 goto skip_alloc; 1328 } 1329#endif 1330 1331 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1332 if (sc == NULL) { 1333 /* 1334 * The zone allocator couldn't provide more entries. 1335 * Treat this as if the cache was full; drop the oldest 1336 * entry and insert the new one. 1337 */ 1338 TCPSTAT_INC(tcps_sc_zonefail); 1339 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) 1340 syncache_drop(sc, sch); 1341 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1342 if (sc == NULL) { 1343 if (V_tcp_syncookies) { 1344 bzero(&scs, sizeof(scs)); 1345 sc = &scs; 1346 } else { 1347 SCH_UNLOCK(sch); 1348 if (ipopts) 1349 (void) m_free(ipopts); 1350 goto done; 1351 } 1352 } 1353 } 1354 1355#ifdef TCP_RFC7413 1356skip_alloc: 1357 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1358 sc->sc_tfo_cookie = &tfo_response_cookie; 1359#endif 1360 1361 /* 1362 * Fill in the syncache values. 1363 */ 1364#ifdef MAC 1365 sc->sc_label = maclabel; 1366#endif 1367 sc->sc_cred = cred; 1368 cred = NULL; 1369 sc->sc_ipopts = ipopts; 1370 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1371#ifdef INET6 1372 if (!(inc->inc_flags & INC_ISIPV6)) 1373#endif 1374 { 1375 sc->sc_ip_tos = ip_tos; 1376 sc->sc_ip_ttl = ip_ttl; 1377 } 1378#ifdef TCP_OFFLOAD 1379 sc->sc_tod = tod; 1380 sc->sc_todctx = todctx; 1381#endif 1382 sc->sc_irs = th->th_seq; 1383 sc->sc_iss = arc4random(); 1384 sc->sc_flags = 0; 1385 sc->sc_flowlabel = 0; 1386 1387 /* 1388 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1389 * win was derived from socket earlier in the function. 1390 */ 1391 win = imax(win, 0); 1392 win = imin(win, TCP_MAXWIN); 1393 sc->sc_wnd = win; 1394 1395 if (V_tcp_do_rfc1323) { 1396 /* 1397 * A timestamp received in a SYN makes 1398 * it ok to send timestamp requests and replies. 1399 */ 1400 if (to->to_flags & TOF_TS) { 1401 sc->sc_tsreflect = to->to_tsval; 1402 sc->sc_ts = tcp_ts_getticks(); 1403 sc->sc_flags |= SCF_TIMESTAMP; 1404 } 1405 if (to->to_flags & TOF_SCALE) { 1406 int wscale = 0; 1407 1408 /* 1409 * Pick the smallest possible scaling factor that 1410 * will still allow us to scale up to sb_max, aka 1411 * kern.ipc.maxsockbuf. 1412 * 1413 * We do this because there are broken firewalls that 1414 * will corrupt the window scale option, leading to 1415 * the other endpoint believing that our advertised 1416 * window is unscaled. At scale factors larger than 1417 * 5 the unscaled window will drop below 1500 bytes, 1418 * leading to serious problems when traversing these 1419 * broken firewalls. 1420 * 1421 * With the default maxsockbuf of 256K, a scale factor 1422 * of 3 will be chosen by this algorithm. Those who 1423 * choose a larger maxsockbuf should watch out 1424 * for the compatiblity problems mentioned above. 1425 * 1426 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1427 * or <SYN,ACK>) segment itself is never scaled. 1428 */ 1429 while (wscale < TCP_MAX_WINSHIFT && 1430 (TCP_MAXWIN << wscale) < sb_max) 1431 wscale++; 1432 sc->sc_requested_r_scale = wscale; 1433 sc->sc_requested_s_scale = to->to_wscale; 1434 sc->sc_flags |= SCF_WINSCALE; 1435 } 1436 } 1437#ifdef TCP_SIGNATURE 1438 /* 1439 * If listening socket requested TCP digests, and received SYN 1440 * contains the option, flag this in the syncache so that 1441 * syncache_respond() will do the right thing with the SYN+ACK. 1442 * XXX: Currently we always record the option by default and will 1443 * attempt to use it in syncache_respond(). 1444 */ 1445 if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE) 1446 sc->sc_flags |= SCF_SIGNATURE; 1447#endif 1448 if (to->to_flags & TOF_SACKPERM) 1449 sc->sc_flags |= SCF_SACK; 1450 if (to->to_flags & TOF_MSS) 1451 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1452 if (ltflags & TF_NOOPT) 1453 sc->sc_flags |= SCF_NOOPT; 1454 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1455 sc->sc_flags |= SCF_ECN; 1456 1457 if (V_tcp_syncookies) 1458 sc->sc_iss = syncookie_generate(sch, sc); 1459#ifdef INET6 1460 if (autoflowlabel) { 1461 if (V_tcp_syncookies) 1462 sc->sc_flowlabel = sc->sc_iss; 1463 else 1464 sc->sc_flowlabel = ip6_randomflowlabel(); 1465 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1466 } 1467#endif 1468 SCH_UNLOCK(sch); 1469 1470#ifdef TCP_RFC7413 1471 if (tfo_cookie_valid) { 1472 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie); 1473 /* INP_WUNLOCK(inp) will be performed by the called */ 1474 rv = 1; 1475 goto tfo_done; 1476 } 1477#endif 1478 1479 /* 1480 * Do a standard 3-way handshake. 1481 */ 1482 if (syncache_respond(sc, m) == 0) { 1483 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1484 syncache_free(sc); 1485 else if (sc != &scs) 1486 syncache_insert(sc, sch); /* locks and unlocks sch */ 1487 TCPSTAT_INC(tcps_sndacks); 1488 TCPSTAT_INC(tcps_sndtotal); 1489 } else { 1490 if (sc != &scs) 1491 syncache_free(sc); 1492 TCPSTAT_INC(tcps_sc_dropped); 1493 } 1494 1495done: 1496 if (m) { 1497 *lsop = NULL; 1498 m_freem(m); 1499 } 1500#ifdef TCP_RFC7413 1501tfo_done: 1502#endif 1503 if (cred != NULL) 1504 crfree(cred); 1505#ifdef MAC 1506 if (sc == &scs) 1507 mac_syncache_destroy(&maclabel); 1508#endif 1509 return (rv); 1510} 1511 1512/* 1513 * Send SYN|ACK to the peer. Either in response to the peer's SYN, 1514 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1515 */ 1516static int 1517syncache_respond(struct syncache *sc, const struct mbuf *m0) 1518{ 1519 struct ip *ip = NULL; 1520 struct mbuf *m; 1521 struct tcphdr *th = NULL; 1522 int optlen, error = 0; /* Make compiler happy */ 1523 u_int16_t hlen, tlen, mssopt; 1524 struct tcpopt to; 1525#ifdef INET6 1526 struct ip6_hdr *ip6 = NULL; 1527#endif 1528 1529 hlen = 1530#ifdef INET6 1531 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1532#endif 1533 sizeof(struct ip); 1534 tlen = hlen + sizeof(struct tcphdr); 1535 1536 /* Determine MSS we advertize to other end of connection. */ 1537 mssopt = tcp_mssopt(&sc->sc_inc); 1538 if (sc->sc_peer_mss) 1539 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss); 1540 1541 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1542 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1543 ("syncache: mbuf too small")); 1544 1545 /* Create the IP+TCP header from scratch. */ 1546 m = m_gethdr(M_NOWAIT, MT_DATA); 1547 if (m == NULL) 1548 return (ENOBUFS); 1549#ifdef MAC 1550 mac_syncache_create_mbuf(sc->sc_label, m); 1551#endif 1552 m->m_data += max_linkhdr; 1553 m->m_len = tlen; 1554 m->m_pkthdr.len = tlen; 1555 m->m_pkthdr.rcvif = NULL; 1556 1557#ifdef INET6 1558 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1559 ip6 = mtod(m, struct ip6_hdr *); 1560 ip6->ip6_vfc = IPV6_VERSION; 1561 ip6->ip6_nxt = IPPROTO_TCP; 1562 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1563 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1564 ip6->ip6_plen = htons(tlen - hlen); 1565 /* ip6_hlim is set after checksum */ 1566 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1567 ip6->ip6_flow |= sc->sc_flowlabel; 1568 1569 th = (struct tcphdr *)(ip6 + 1); 1570 } 1571#endif 1572#if defined(INET6) && defined(INET) 1573 else 1574#endif 1575#ifdef INET 1576 { 1577 ip = mtod(m, struct ip *); 1578 ip->ip_v = IPVERSION; 1579 ip->ip_hl = sizeof(struct ip) >> 2; 1580 ip->ip_len = htons(tlen); 1581 ip->ip_id = 0; 1582 ip->ip_off = 0; 1583 ip->ip_sum = 0; 1584 ip->ip_p = IPPROTO_TCP; 1585 ip->ip_src = sc->sc_inc.inc_laddr; 1586 ip->ip_dst = sc->sc_inc.inc_faddr; 1587 ip->ip_ttl = sc->sc_ip_ttl; 1588 ip->ip_tos = sc->sc_ip_tos; 1589 1590 /* 1591 * See if we should do MTU discovery. Route lookups are 1592 * expensive, so we will only unset the DF bit if: 1593 * 1594 * 1) path_mtu_discovery is disabled 1595 * 2) the SCF_UNREACH flag has been set 1596 */ 1597 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1598 ip->ip_off |= htons(IP_DF); 1599 1600 th = (struct tcphdr *)(ip + 1); 1601 } 1602#endif /* INET */ 1603 th->th_sport = sc->sc_inc.inc_lport; 1604 th->th_dport = sc->sc_inc.inc_fport; 1605 1606 th->th_seq = htonl(sc->sc_iss); 1607 th->th_ack = htonl(sc->sc_irs + 1); 1608 th->th_off = sizeof(struct tcphdr) >> 2; 1609 th->th_x2 = 0; 1610 th->th_flags = TH_SYN|TH_ACK; 1611 th->th_win = htons(sc->sc_wnd); 1612 th->th_urp = 0; 1613 1614 if (sc->sc_flags & SCF_ECN) { 1615 th->th_flags |= TH_ECE; 1616 TCPSTAT_INC(tcps_ecn_shs); 1617 } 1618 1619 /* Tack on the TCP options. */ 1620 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1621 to.to_flags = 0; 1622 1623 to.to_mss = mssopt; 1624 to.to_flags = TOF_MSS; 1625 if (sc->sc_flags & SCF_WINSCALE) { 1626 to.to_wscale = sc->sc_requested_r_scale; 1627 to.to_flags |= TOF_SCALE; 1628 } 1629 if (sc->sc_flags & SCF_TIMESTAMP) { 1630 /* Virgin timestamp or TCP cookie enhanced one. */ 1631 to.to_tsval = sc->sc_ts; 1632 to.to_tsecr = sc->sc_tsreflect; 1633 to.to_flags |= TOF_TS; 1634 } 1635 if (sc->sc_flags & SCF_SACK) 1636 to.to_flags |= TOF_SACKPERM; 1637#ifdef TCP_SIGNATURE 1638 if (sc->sc_flags & SCF_SIGNATURE) 1639 to.to_flags |= TOF_SIGNATURE; 1640#endif 1641 1642#ifdef TCP_RFC7413 1643 if (sc->sc_tfo_cookie) { 1644 to.to_flags |= TOF_FASTOPEN; 1645 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1646 to.to_tfo_cookie = sc->sc_tfo_cookie; 1647 /* don't send cookie again when retransmitting response */ 1648 sc->sc_tfo_cookie = NULL; 1649 } 1650#endif 1651 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1652 1653 /* Adjust headers by option size. */ 1654 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1655 m->m_len += optlen; 1656 m->m_pkthdr.len += optlen; 1657 1658#ifdef TCP_SIGNATURE 1659 if (sc->sc_flags & SCF_SIGNATURE) 1660 tcp_signature_compute(m, 0, 0, optlen, 1661 to.to_signature, IPSEC_DIR_OUTBOUND); 1662#endif 1663#ifdef INET6 1664 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1665 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1666 else 1667#endif 1668 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1669 } else 1670 optlen = 0; 1671 1672 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1673 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1674 /* 1675 * If we have peer's SYN and it has a flowid, then let's assign it to 1676 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1677 * to SYN|ACK due to lack of inp here. 1678 */ 1679 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 1680 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 1681 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 1682 } 1683#ifdef INET6 1684 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1685 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 1686 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 1687 IPPROTO_TCP, 0); 1688 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1689#ifdef TCP_OFFLOAD 1690 if (ADDED_BY_TOE(sc)) { 1691 struct toedev *tod = sc->sc_tod; 1692 1693 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1694 1695 return (error); 1696 } 1697#endif 1698 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1699 } 1700#endif 1701#if defined(INET6) && defined(INET) 1702 else 1703#endif 1704#ifdef INET 1705 { 1706 m->m_pkthdr.csum_flags = CSUM_TCP; 1707 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1708 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1709#ifdef TCP_OFFLOAD 1710 if (ADDED_BY_TOE(sc)) { 1711 struct toedev *tod = sc->sc_tod; 1712 1713 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 1714 1715 return (error); 1716 } 1717#endif 1718 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1719 } 1720#endif 1721 return (error); 1722} 1723 1724/* 1725 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 1726 * that exceed the capacity of the syncache by avoiding the storage of any 1727 * of the SYNs we receive. Syncookies defend against blind SYN flooding 1728 * attacks where the attacker does not have access to our responses. 1729 * 1730 * Syncookies encode and include all necessary information about the 1731 * connection setup within the SYN|ACK that we send back. That way we 1732 * can avoid keeping any local state until the ACK to our SYN|ACK returns 1733 * (if ever). Normally the syncache and syncookies are running in parallel 1734 * with the latter taking over when the former is exhausted. When matching 1735 * syncache entry is found the syncookie is ignored. 1736 * 1737 * The only reliable information persisting the 3WHS is our inital sequence 1738 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 1739 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 1740 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 1741 * returns and signifies a legitimate connection if it matches the ACK. 1742 * 1743 * The available space of 32 bits to store the hash and to encode the SYN 1744 * option information is very tight and we should have at least 24 bits for 1745 * the MAC to keep the number of guesses by blind spoofing reasonably high. 1746 * 1747 * SYN option information we have to encode to fully restore a connection: 1748 * MSS: is imporant to chose an optimal segment size to avoid IP level 1749 * fragmentation along the path. The common MSS values can be encoded 1750 * in a 3-bit table. Uncommon values are captured by the next lower value 1751 * in the table leading to a slight increase in packetization overhead. 1752 * WSCALE: is necessary to allow large windows to be used for high delay- 1753 * bandwidth product links. Not scaling the window when it was initially 1754 * negotiated is bad for performance as lack of scaling further decreases 1755 * the apparent available send window. We only need to encode the WSCALE 1756 * we received from the remote end. Our end can be recalculated at any 1757 * time. The common WSCALE values can be encoded in a 3-bit table. 1758 * Uncommon values are captured by the next lower value in the table 1759 * making us under-estimate the available window size halving our 1760 * theoretically possible maximum throughput for that connection. 1761 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 1762 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 1763 * that are included in all segments on a connection. We enable them when 1764 * the ACK has them. 1765 * 1766 * Security of syncookies and attack vectors: 1767 * 1768 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 1769 * together with the gloabl secret to make it unique per connection attempt. 1770 * Thus any change of any of those parameters results in a different MAC output 1771 * in an unpredictable way unless a collision is encountered. 24 bits of the 1772 * MAC are embedded into the ISS. 1773 * 1774 * To prevent replay attacks two rotating global secrets are updated with a 1775 * new random value every 15 seconds. The life-time of a syncookie is thus 1776 * 15-30 seconds. 1777 * 1778 * Vector 1: Attacking the secret. This requires finding a weakness in the 1779 * MAC itself or the way it is used here. The attacker can do a chosen plain 1780 * text attack by varying and testing the all parameters under his control. 1781 * The strength depends on the size and randomness of the secret, and the 1782 * cryptographic security of the MAC function. Due to the constant updating 1783 * of the secret the attacker has at most 29.999 seconds to find the secret 1784 * and launch spoofed connections. After that he has to start all over again. 1785 * 1786 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 1787 * size an average of 4,823 attempts are required for a 50% chance of success 1788 * to spoof a single syncookie (birthday collision paradox). However the 1789 * attacker is blind and doesn't know if one of his attempts succeeded unless 1790 * he has a side channel to interfere success from. A single connection setup 1791 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 1792 * This many attempts are required for each one blind spoofed connection. For 1793 * every additional spoofed connection he has to launch another N attempts. 1794 * Thus for a sustained rate 100 spoofed connections per second approximately 1795 * 1,800,000 packets per second would have to be sent. 1796 * 1797 * NB: The MAC function should be fast so that it doesn't become a CPU 1798 * exhaustion attack vector itself. 1799 * 1800 * References: 1801 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 1802 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 1803 * http://cr.yp.to/syncookies.html (overview) 1804 * http://cr.yp.to/syncookies/archive (details) 1805 * 1806 * 1807 * Schematic construction of a syncookie enabled Initial Sequence Number: 1808 * 0 1 2 3 1809 * 12345678901234567890123456789012 1810 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 1811 * 1812 * x 24 MAC (truncated) 1813 * W 3 Send Window Scale index 1814 * M 3 MSS index 1815 * S 1 SACK permitted 1816 * P 1 Odd/even secret 1817 */ 1818 1819/* 1820 * Distribution and probability of certain MSS values. Those in between are 1821 * rounded down to the next lower one. 1822 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 1823 * .2% .3% 5% 7% 7% 20% 15% 45% 1824 */ 1825static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 1826 1827/* 1828 * Distribution and probability of certain WSCALE values. We have to map the 1829 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 1830 * bits based on prevalence of certain values. Where we don't have an exact 1831 * match for are rounded down to the next lower one letting us under-estimate 1832 * the true available window. At the moment this would happen only for the 1833 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 1834 * and window size). The absence of the WSCALE option (no scaling in either 1835 * direction) is encoded with index zero. 1836 * [WSCALE values histograms, Allman, 2012] 1837 * X 10 10 35 5 6 14 10% by host 1838 * X 11 4 5 5 18 49 3% by connections 1839 */ 1840static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 1841 1842/* 1843 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 1844 * and good cryptographic properties. 1845 */ 1846static uint32_t 1847syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 1848 uint8_t *secbits, uintptr_t secmod) 1849{ 1850 SIPHASH_CTX ctx; 1851 uint32_t siphash[2]; 1852 1853 SipHash24_Init(&ctx); 1854 SipHash_SetKey(&ctx, secbits); 1855 switch (inc->inc_flags & INC_ISIPV6) { 1856#ifdef INET 1857 case 0: 1858 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 1859 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 1860 break; 1861#endif 1862#ifdef INET6 1863 case INC_ISIPV6: 1864 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 1865 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 1866 break; 1867#endif 1868 } 1869 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 1870 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 1871 SipHash_Update(&ctx, &irs, sizeof(irs)); 1872 SipHash_Update(&ctx, &flags, sizeof(flags)); 1873 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 1874 SipHash_Final((u_int8_t *)&siphash, &ctx); 1875 1876 return (siphash[0] ^ siphash[1]); 1877} 1878 1879static tcp_seq 1880syncookie_generate(struct syncache_head *sch, struct syncache *sc) 1881{ 1882 u_int i, mss, secbit, wscale; 1883 uint32_t iss, hash; 1884 uint8_t *secbits; 1885 union syncookie cookie; 1886 1887 SCH_LOCK_ASSERT(sch); 1888 1889 cookie.cookie = 0; 1890 1891 /* Map our computed MSS into the 3-bit index. */ 1892 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss)); 1893 for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1; 1894 tcp_sc_msstab[i] > mss && i > 0; 1895 i--) 1896 ; 1897 cookie.flags.mss_idx = i; 1898 1899 /* 1900 * Map the send window scale into the 3-bit index but only if 1901 * the wscale option was received. 1902 */ 1903 if (sc->sc_flags & SCF_WINSCALE) { 1904 wscale = sc->sc_requested_s_scale; 1905 for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1; 1906 tcp_sc_wstab[i] > wscale && i > 0; 1907 i--) 1908 ; 1909 cookie.flags.wscale_idx = i; 1910 } 1911 1912 /* Can we do SACK? */ 1913 if (sc->sc_flags & SCF_SACK) 1914 cookie.flags.sack_ok = 1; 1915 1916 /* Which of the two secrets to use. */ 1917 secbit = sch->sch_sc->secret.oddeven & 0x1; 1918 cookie.flags.odd_even = secbit; 1919 1920 secbits = sch->sch_sc->secret.key[secbit]; 1921 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 1922 (uintptr_t)sch); 1923 1924 /* 1925 * Put the flags into the hash and XOR them to get better ISS number 1926 * variance. This doesn't enhance the cryptographic strength and is 1927 * done to prevent the 8 cookie bits from showing up directly on the 1928 * wire. 1929 */ 1930 iss = hash & ~0xff; 1931 iss |= cookie.cookie ^ (hash >> 24); 1932 1933 /* Randomize the timestamp. */ 1934 if (sc->sc_flags & SCF_TIMESTAMP) { 1935 sc->sc_ts = arc4random(); 1936 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks(); 1937 } 1938 1939 TCPSTAT_INC(tcps_sc_sendcookie); 1940 return (iss); 1941} 1942 1943static struct syncache * 1944syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 1945 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 1946 struct socket *lso) 1947{ 1948 uint32_t hash; 1949 uint8_t *secbits; 1950 tcp_seq ack, seq; 1951 int wnd, wscale = 0; 1952 union syncookie cookie; 1953 1954 SCH_LOCK_ASSERT(sch); 1955 1956 /* 1957 * Pull information out of SYN-ACK/ACK and revert sequence number 1958 * advances. 1959 */ 1960 ack = th->th_ack - 1; 1961 seq = th->th_seq - 1; 1962 1963 /* 1964 * Unpack the flags containing enough information to restore the 1965 * connection. 1966 */ 1967 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 1968 1969 /* Which of the two secrets to use. */ 1970 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even]; 1971 1972 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 1973 1974 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 1975 if ((ack & ~0xff) != (hash & ~0xff)) 1976 return (NULL); 1977 1978 /* Fill in the syncache values. */ 1979 sc->sc_flags = 0; 1980 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1981 sc->sc_ipopts = NULL; 1982 1983 sc->sc_irs = seq; 1984 sc->sc_iss = ack; 1985 1986 switch (inc->inc_flags & INC_ISIPV6) { 1987#ifdef INET 1988 case 0: 1989 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 1990 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 1991 break; 1992#endif 1993#ifdef INET6 1994 case INC_ISIPV6: 1995 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 1996 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK; 1997 break; 1998#endif 1999 } 2000 2001 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2002 2003 /* We can simply recompute receive window scale we sent earlier. */ 2004 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2005 wscale++; 2006 2007 /* Only use wscale if it was enabled in the orignal SYN. */ 2008 if (cookie.flags.wscale_idx > 0) { 2009 sc->sc_requested_r_scale = wscale; 2010 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2011 sc->sc_flags |= SCF_WINSCALE; 2012 } 2013 2014 wnd = sbspace(&lso->so_rcv); 2015 wnd = imax(wnd, 0); 2016 wnd = imin(wnd, TCP_MAXWIN); 2017 sc->sc_wnd = wnd; 2018 2019 if (cookie.flags.sack_ok) 2020 sc->sc_flags |= SCF_SACK; 2021 2022 if (to->to_flags & TOF_TS) { 2023 sc->sc_flags |= SCF_TIMESTAMP; 2024 sc->sc_tsreflect = to->to_tsval; 2025 sc->sc_ts = to->to_tsecr; 2026 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks(); 2027 } 2028 2029 if (to->to_flags & TOF_SIGNATURE) 2030 sc->sc_flags |= SCF_SIGNATURE; 2031 2032 sc->sc_rxmits = 0; 2033 2034 TCPSTAT_INC(tcps_sc_recvcookie); 2035 return (sc); 2036} 2037 2038#ifdef INVARIANTS 2039static int 2040syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2041 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2042 struct socket *lso) 2043{ 2044 struct syncache scs, *scx; 2045 char *s; 2046 2047 bzero(&scs, sizeof(scs)); 2048 scx = syncookie_lookup(inc, sch, &scs, th, to, lso); 2049 2050 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2051 return (0); 2052 2053 if (scx != NULL) { 2054 if (sc->sc_peer_mss != scx->sc_peer_mss) 2055 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2056 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2057 2058 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2059 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2060 s, __func__, sc->sc_requested_r_scale, 2061 scx->sc_requested_r_scale); 2062 2063 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2064 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2065 s, __func__, sc->sc_requested_s_scale, 2066 scx->sc_requested_s_scale); 2067 2068 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2069 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2070 } 2071 2072 if (s != NULL) 2073 free(s, M_TCPLOG); 2074 return (0); 2075} 2076#endif /* INVARIANTS */ 2077 2078static void 2079syncookie_reseed(void *arg) 2080{ 2081 struct tcp_syncache *sc = arg; 2082 uint8_t *secbits; 2083 int secbit; 2084 2085 /* 2086 * Reseeding the secret doesn't have to be protected by a lock. 2087 * It only must be ensured that the new random values are visible 2088 * to all CPUs in a SMP environment. The atomic with release 2089 * semantics ensures that. 2090 */ 2091 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2092 secbits = sc->secret.key[secbit]; 2093 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2094 atomic_add_rel_int(&sc->secret.oddeven, 1); 2095 2096 /* Reschedule ourself. */ 2097 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2098} 2099 2100/* 2101 * Returns the current number of syncache entries. This number 2102 * will probably change before you get around to calling 2103 * syncache_pcblist. 2104 */ 2105int 2106syncache_pcbcount(void) 2107{ 2108 struct syncache_head *sch; 2109 int count, i; 2110 2111 for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2112 /* No need to lock for a read. */ 2113 sch = &V_tcp_syncache.hashbase[i]; 2114 count += sch->sch_length; 2115 } 2116 return count; 2117} 2118 2119/* 2120 * Exports the syncache entries to userland so that netstat can display 2121 * them alongside the other sockets. This function is intended to be 2122 * called only from tcp_pcblist. 2123 * 2124 * Due to concurrency on an active system, the number of pcbs exported 2125 * may have no relation to max_pcbs. max_pcbs merely indicates the 2126 * amount of space the caller allocated for this function to use. 2127 */ 2128int 2129syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 2130{ 2131 struct xtcpcb xt; 2132 struct syncache *sc; 2133 struct syncache_head *sch; 2134 int count, error, i; 2135 2136 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 2137 sch = &V_tcp_syncache.hashbase[i]; 2138 SCH_LOCK(sch); 2139 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2140 if (count >= max_pcbs) { 2141 SCH_UNLOCK(sch); 2142 goto exit; 2143 } 2144 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2145 continue; 2146 bzero(&xt, sizeof(xt)); 2147 xt.xt_len = sizeof(xt); 2148 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2149 xt.xt_inp.inp_vflag = INP_IPV6; 2150 else 2151 xt.xt_inp.inp_vflag = INP_IPV4; 2152 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 2153 xt.xt_tp.t_inpcb = &xt.xt_inp; 2154 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 2155 xt.xt_socket.xso_protocol = IPPROTO_TCP; 2156 xt.xt_socket.xso_len = sizeof (struct xsocket); 2157 xt.xt_socket.so_type = SOCK_STREAM; 2158 xt.xt_socket.so_state = SS_ISCONNECTING; 2159 error = SYSCTL_OUT(req, &xt, sizeof xt); 2160 if (error) { 2161 SCH_UNLOCK(sch); 2162 goto exit; 2163 } 2164 count++; 2165 } 2166 SCH_UNLOCK(sch); 2167 } 2168exit: 2169 *pcbs_exported = count; 2170 return error; 2171} 2172