1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2001 McAfee, Inc. 5 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG 6 * All rights reserved. 7 * 8 * This software was developed for the FreeBSD Project by Jonathan Lemon 9 * and McAfee Research, the Security Research Division of McAfee, Inc. under 10 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 11 * DARPA CHATS research program. [2001 McAfee, Inc.] 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include "opt_inet.h" 36 #include "opt_inet6.h" 37 #include "opt_ipsec.h" 38 39 #include <sys/param.h> 40 #include <sys/systm.h> 41 #include <sys/hash.h> 42 #include <sys/refcount.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/if_var.h> 64 #include <net/route.h> 65 #include <net/vnet.h> 66 67 #include <netinet/in.h> 68 #include <netinet/in_kdtrace.h> 69 #include <netinet/in_systm.h> 70 #include <netinet/ip.h> 71 #include <netinet/in_var.h> 72 #include <netinet/in_pcb.h> 73 #include <netinet/in_rss.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 #include <netinet6/in6_rss.h> 83 #endif 84 #include <netinet/tcp.h> 85 #include <netinet/tcp_fastopen.h> 86 #include <netinet/tcp_fsm.h> 87 #include <netinet/tcp_seq.h> 88 #include <netinet/tcp_timer.h> 89 #include <netinet/tcp_var.h> 90 #include <netinet/tcp_syncache.h> 91 #include <netinet/tcp_ecn.h> 92 #ifdef TCP_BLACKBOX 93 #include <netinet/tcp_log_buf.h> 94 #endif 95 #ifdef TCP_OFFLOAD 96 #include <netinet/toecore.h> 97 #endif 98 #include <netinet/udp.h> 99 100 #include <netipsec/ipsec_support.h> 101 102 #include <machine/in_cksum.h> 103 104 #include <security/mac/mac_framework.h> 105 106 VNET_DEFINE_STATIC(bool, tcp_syncookies) = true; 107 #define V_tcp_syncookies VNET(tcp_syncookies) 108 SYSCTL_BOOL(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 109 &VNET_NAME(tcp_syncookies), 0, 110 "Use TCP SYN cookies if the syncache overflows"); 111 112 VNET_DEFINE_STATIC(bool, tcp_syncookiesonly) = false; 113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 114 SYSCTL_BOOL(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 115 &VNET_NAME(tcp_syncookiesonly), 0, 116 "Use only TCP SYN cookies"); 117 118 #ifdef TCP_OFFLOAD 119 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 120 #endif 121 122 static void syncache_drop(struct syncache *, struct syncache_head *); 123 static void syncache_free(struct syncache *); 124 static void syncache_insert(struct syncache *, struct syncache_head *); 125 static int syncache_respond(struct syncache *, int); 126 static void syncache_send_challenge_ack(struct syncache *); 127 static struct socket *syncache_socket(struct syncache *, struct socket *, 128 struct mbuf *m); 129 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 130 int docallout); 131 static void syncache_timer(void *); 132 133 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 134 uint8_t *, uintptr_t); 135 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 136 static bool syncookie_expand(struct in_conninfo *, 137 const struct syncache_head *, struct syncache *, 138 struct tcphdr *, struct tcpopt *, struct socket *, 139 uint16_t); 140 static void syncache_pause(struct in_conninfo *); 141 static void syncache_unpause(void *); 142 static void syncookie_reseed(void *); 143 #ifdef INVARIANTS 144 static void syncookie_cmp(struct in_conninfo *, 145 const struct syncache_head *, struct syncache *, 146 struct tcphdr *, struct tcpopt *, struct socket *, 147 uint16_t); 148 #endif 149 150 /* 151 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 152 * 3 retransmits corresponds to a timeout with default values of 153 * tcp_rexmit_initial * ( 1 + 154 * tcp_backoff[1] + 155 * tcp_backoff[2] + 156 * tcp_backoff[3]) + 3 * tcp_rexmit_slop, 157 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms, 158 * the odds are that the user has given up attempting to connect by then. 159 */ 160 #define SYNCACHE_MAXREXMTS 3 161 162 /* Arbitrary values */ 163 #define TCP_SYNCACHE_HASHSIZE 512 164 #define TCP_SYNCACHE_BUCKETLIMIT 30 165 166 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 167 #define V_tcp_syncache VNET(tcp_syncache) 168 169 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, 170 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 171 "TCP SYN cache"); 172 173 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 174 &VNET_NAME(tcp_syncache.bucket_limit), 0, 175 "Per-bucket hash limit for syncache"); 176 177 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 178 &VNET_NAME(tcp_syncache.cache_limit), 0, 179 "Overall entry limit for syncache"); 180 181 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 182 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 183 184 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 185 &VNET_NAME(tcp_syncache.hashsize), 0, 186 "Size of TCP syncache hashtable"); 187 188 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET | 189 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0, 190 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) " 191 "and mac(4) checks"); 192 193 static int 194 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 195 { 196 int error; 197 u_int new; 198 199 new = V_tcp_syncache.rexmt_limit; 200 error = sysctl_handle_int(oidp, &new, 0, req); 201 if ((error == 0) && (req->newptr != NULL)) { 202 if (new > TCP_MAXRXTSHIFT) 203 error = EINVAL; 204 else 205 V_tcp_syncache.rexmt_limit = new; 206 } 207 return (error); 208 } 209 210 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 211 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 212 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 213 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "IU", 214 "Limit on SYN/ACK retransmissions"); 215 216 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 217 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 218 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 219 "Send reset on socket allocation failure"); 220 221 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 222 223 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 224 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 225 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 226 227 /* 228 * Requires the syncache entry to be already removed from the bucket list. 229 */ 230 static void 231 syncache_free(struct syncache *sc) 232 { 233 234 if (sc->sc_ipopts) 235 (void)m_free(sc->sc_ipopts); 236 if (sc->sc_cred) 237 crfree(sc->sc_cred); 238 #ifdef MAC 239 mac_syncache_destroy(&sc->sc_label); 240 #endif 241 242 uma_zfree(V_tcp_syncache.zone, sc); 243 } 244 245 void 246 syncache_init(void) 247 { 248 int i; 249 250 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 251 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 252 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 253 V_tcp_syncache.hash_secret = arc4random(); 254 255 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 256 &V_tcp_syncache.hashsize); 257 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 258 &V_tcp_syncache.bucket_limit); 259 if (!powerof2(V_tcp_syncache.hashsize) || 260 V_tcp_syncache.hashsize == 0) { 261 printf("WARNING: syncache hash size is not a power of 2.\n"); 262 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 263 } 264 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 265 266 /* Set limits. */ 267 V_tcp_syncache.cache_limit = 268 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 269 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 270 &V_tcp_syncache.cache_limit); 271 272 /* Allocate the hash table. */ 273 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 274 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 275 276 #ifdef VIMAGE 277 V_tcp_syncache.vnet = curvnet; 278 #endif 279 280 /* Initialize the hash buckets. */ 281 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 282 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 283 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 284 NULL, MTX_DEF); 285 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 286 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 287 V_tcp_syncache.hashbase[i].sch_length = 0; 288 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 289 V_tcp_syncache.hashbase[i].sch_last_overflow = 290 -(SYNCOOKIE_LIFETIME + 1); 291 } 292 293 /* Create the syncache entry zone. */ 294 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 295 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 296 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 297 V_tcp_syncache.cache_limit); 298 299 /* Start the SYN cookie reseeder callout. */ 300 callout_init(&V_tcp_syncache.secret.reseed, 1); 301 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 302 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 303 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 304 syncookie_reseed, &V_tcp_syncache); 305 306 /* Initialize the pause machinery. */ 307 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF); 308 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx, 309 0); 310 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME; 311 V_tcp_syncache.pause_backoff = 0; 312 V_tcp_syncache.paused = false; 313 } 314 315 #ifdef VIMAGE 316 void 317 syncache_destroy(void) 318 { 319 struct syncache_head *sch; 320 struct syncache *sc, *nsc; 321 int i; 322 323 /* 324 * Stop the re-seed timer before freeing resources. No need to 325 * possibly schedule it another time. 326 */ 327 callout_drain(&V_tcp_syncache.secret.reseed); 328 329 /* Stop the SYN cache pause callout. */ 330 mtx_lock(&V_tcp_syncache.pause_mtx); 331 if (callout_stop(&V_tcp_syncache.pause_co) == 0) { 332 mtx_unlock(&V_tcp_syncache.pause_mtx); 333 callout_drain(&V_tcp_syncache.pause_co); 334 } else 335 mtx_unlock(&V_tcp_syncache.pause_mtx); 336 337 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 338 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 339 sch = &V_tcp_syncache.hashbase[i]; 340 callout_drain(&sch->sch_timer); 341 342 SCH_LOCK(sch); 343 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 344 syncache_drop(sc, sch); 345 SCH_UNLOCK(sch); 346 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 347 ("%s: sch->sch_bucket not empty", __func__)); 348 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 349 __func__, sch->sch_length)); 350 mtx_destroy(&sch->sch_mtx); 351 } 352 353 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 354 ("%s: cache_count not 0", __func__)); 355 356 /* Free the allocated global resources. */ 357 uma_zdestroy(V_tcp_syncache.zone); 358 free(V_tcp_syncache.hashbase, M_SYNCACHE); 359 mtx_destroy(&V_tcp_syncache.pause_mtx); 360 } 361 #endif 362 363 /* 364 * Inserts a syncache entry into the specified bucket row. 365 * Locks and unlocks the syncache_head autonomously. 366 */ 367 static void 368 syncache_insert(struct syncache *sc, struct syncache_head *sch) 369 { 370 struct syncache *sc2; 371 372 SCH_LOCK(sch); 373 374 /* 375 * Make sure that we don't overflow the per-bucket limit. 376 * If the bucket is full, toss the oldest element. 377 */ 378 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 379 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 380 ("sch->sch_length incorrect")); 381 syncache_pause(&sc->sc_inc); 382 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 383 sch->sch_last_overflow = time_uptime; 384 syncache_drop(sc2, sch); 385 } 386 387 /* Put it into the bucket. */ 388 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 389 sch->sch_length++; 390 391 #ifdef TCP_OFFLOAD 392 if (ADDED_BY_TOE(sc)) { 393 struct toedev *tod = sc->sc_tod; 394 395 tod->tod_syncache_added(tod, sc->sc_todctx); 396 } 397 #endif 398 399 /* Reinitialize the bucket row's timer. */ 400 if (sch->sch_length == 1) 401 sch->sch_nextc = ticks + INT_MAX; 402 syncache_timeout(sc, sch, 1); 403 404 SCH_UNLOCK(sch); 405 406 TCPSTATES_INC(TCPS_SYN_RECEIVED); 407 TCPSTAT_INC(tcps_sc_added); 408 } 409 410 /* 411 * Remove and free entry from syncache bucket row. 412 * Expects locked syncache head. 413 */ 414 static void 415 syncache_drop(struct syncache *sc, struct syncache_head *sch) 416 { 417 418 SCH_LOCK_ASSERT(sch); 419 420 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 421 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 422 sch->sch_length--; 423 424 #ifdef TCP_OFFLOAD 425 if (ADDED_BY_TOE(sc)) { 426 struct toedev *tod = sc->sc_tod; 427 428 tod->tod_syncache_removed(tod, sc->sc_todctx); 429 } 430 #endif 431 432 syncache_free(sc); 433 } 434 435 /* 436 * Engage/reengage time on bucket row. 437 */ 438 static void 439 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 440 { 441 int rexmt; 442 443 if (sc->sc_rxmits == 0) 444 rexmt = tcp_rexmit_initial; 445 else 446 TCPT_RANGESET(rexmt, 447 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], 448 tcp_rexmit_min, tcp_rexmit_max); 449 sc->sc_rxttime = ticks + rexmt; 450 sc->sc_rxmits++; 451 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 452 sch->sch_nextc = sc->sc_rxttime; 453 if (docallout) 454 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 455 syncache_timer, (void *)sch); 456 } 457 } 458 459 /* 460 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 461 * If we have retransmitted an entry the maximum number of times, expire it. 462 * One separate timer for each bucket row. 463 */ 464 static void 465 syncache_timer(void *xsch) 466 { 467 struct syncache_head *sch = (struct syncache_head *)xsch; 468 struct syncache *sc, *nsc; 469 struct epoch_tracker et; 470 int tick = ticks; 471 char *s; 472 bool paused; 473 474 CURVNET_SET(sch->sch_sc->vnet); 475 476 /* NB: syncache_head has already been locked by the callout. */ 477 SCH_LOCK_ASSERT(sch); 478 479 /* 480 * In the following cycle we may remove some entries and/or 481 * advance some timeouts, so re-initialize the bucket timer. 482 */ 483 sch->sch_nextc = tick + INT_MAX; 484 485 /* 486 * If we have paused processing, unconditionally remove 487 * all syncache entries. 488 */ 489 mtx_lock(&V_tcp_syncache.pause_mtx); 490 paused = V_tcp_syncache.paused; 491 mtx_unlock(&V_tcp_syncache.pause_mtx); 492 493 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 494 if (paused) { 495 syncache_drop(sc, sch); 496 continue; 497 } 498 /* 499 * We do not check if the listen socket still exists 500 * and accept the case where the listen socket may be 501 * gone by the time we resend the SYN/ACK. We do 502 * not expect this to happens often. If it does, 503 * then the RST will be sent by the time the remote 504 * host does the SYN/ACK->ACK. 505 */ 506 if (TSTMP_GT(sc->sc_rxttime, tick)) { 507 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 508 sch->sch_nextc = sc->sc_rxttime; 509 continue; 510 } 511 if (sc->sc_rxmits > V_tcp_ecn_maxretries) { 512 sc->sc_flags &= ~SCF_ECN_MASK; 513 } 514 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 515 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 516 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 517 "giving up and removing syncache entry\n", 518 s, __func__); 519 free(s, M_TCPLOG); 520 } 521 syncache_drop(sc, sch); 522 TCPSTAT_INC(tcps_sc_stale); 523 continue; 524 } 525 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 526 log(LOG_DEBUG, "%s; %s: Response timeout, " 527 "retransmitting (%u) SYN|ACK\n", 528 s, __func__, sc->sc_rxmits); 529 free(s, M_TCPLOG); 530 } 531 532 NET_EPOCH_ENTER(et); 533 if (syncache_respond(sc, TH_SYN|TH_ACK) == 0) { 534 syncache_timeout(sc, sch, 0); 535 TCPSTAT_INC(tcps_sndacks); 536 TCPSTAT_INC(tcps_sndtotal); 537 TCPSTAT_INC(tcps_sc_retransmitted); 538 } else { 539 /* 540 * Most likely we are memory constrained, so free 541 * resources. 542 */ 543 syncache_drop(sc, sch); 544 TCPSTAT_INC(tcps_sc_dropped); 545 } 546 NET_EPOCH_EXIT(et); 547 } 548 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 549 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 550 syncache_timer, (void *)(sch)); 551 CURVNET_RESTORE(); 552 } 553 554 /* 555 * Returns true if the system is only using cookies at the moment. 556 * This could be due to a sysadmin decision to only use cookies, or it 557 * could be due to the system detecting an attack. 558 */ 559 static inline bool 560 syncache_cookiesonly(void) 561 { 562 return ((V_tcp_syncookies && V_tcp_syncache.paused) || 563 V_tcp_syncookiesonly); 564 } 565 566 /* 567 * Find the hash bucket for the given connection. 568 */ 569 static struct syncache_head * 570 syncache_hashbucket(struct in_conninfo *inc) 571 { 572 uint32_t hash; 573 574 /* 575 * The hash is built on foreign port + local port + foreign address. 576 * We rely on the fact that struct in_conninfo starts with 16 bits 577 * of foreign port, then 16 bits of local port then followed by 128 578 * bits of foreign address. In case of IPv4 address, the first 3 579 * 32-bit words of the address always are zeroes. 580 */ 581 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 582 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 583 584 return (&V_tcp_syncache.hashbase[hash]); 585 } 586 587 /* 588 * Find an entry in the syncache. 589 * Returns always with locked syncache_head plus a matching entry or NULL. 590 */ 591 static struct syncache * 592 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 593 { 594 struct syncache *sc; 595 struct syncache_head *sch; 596 597 *schp = sch = syncache_hashbucket(inc); 598 SCH_LOCK(sch); 599 600 /* Circle through bucket row to find matching entry. */ 601 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 602 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 603 sizeof(struct in_endpoints)) == 0) 604 break; 605 606 return (sc); /* Always returns with locked sch. */ 607 } 608 609 /* 610 * This function is called when we get a RST for a 611 * non-existent connection, so that we can see if the 612 * connection is in the syn cache. If it is, zap it. 613 * If required send a challenge ACK. 614 */ 615 void 616 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, uint16_t port) 617 { 618 struct syncache *sc; 619 struct syncache_head *sch; 620 char *s = NULL; 621 622 if (syncache_cookiesonly()) 623 return; 624 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 625 SCH_LOCK_ASSERT(sch); 626 627 /* 628 * No corresponding connection was found in syncache. 629 * If syncookies are enabled and possibly exclusively 630 * used, or we are under memory pressure, a valid RST 631 * may not find a syncache entry. In that case we're 632 * done and no SYN|ACK retransmissions will happen. 633 * Otherwise the RST was misdirected or spoofed. 634 */ 635 if (sc == NULL) { 636 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 637 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 638 "syncache entry (possibly syncookie only), " 639 "segment ignored\n", s, __func__); 640 TCPSTAT_INC(tcps_badrst); 641 goto done; 642 } 643 644 /* The remote UDP encaps port does not match. */ 645 if (sc->sc_port != port) { 646 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 647 log(LOG_DEBUG, "%s; %s: Spurious RST with matching " 648 "syncache entry but non-matching UDP encaps port, " 649 "segment ignored\n", s, __func__); 650 TCPSTAT_INC(tcps_badrst); 651 goto done; 652 } 653 654 /* 655 * If the RST bit is set, check the sequence number to see 656 * if this is a valid reset segment. 657 * 658 * RFC 793 page 37: 659 * In all states except SYN-SENT, all reset (RST) segments 660 * are validated by checking their SEQ-fields. A reset is 661 * valid if its sequence number is in the window. 662 * 663 * RFC 793 page 69: 664 * There are four cases for the acceptability test for an incoming 665 * segment: 666 * 667 * Segment Receive Test 668 * Length Window 669 * ------- ------- ------------------------------------------- 670 * 0 0 SEG.SEQ = RCV.NXT 671 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 672 * >0 0 not acceptable 673 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 674 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 675 * 676 * Note that when receiving a SYN segment in the LISTEN state, 677 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 678 * described in RFC 793, page 66. 679 */ 680 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 681 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 682 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 683 if (V_tcp_insecure_rst || 684 th->th_seq == sc->sc_irs + 1) { 685 syncache_drop(sc, sch); 686 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 687 log(LOG_DEBUG, 688 "%s; %s: Our SYN|ACK was rejected, " 689 "connection attempt aborted by remote " 690 "endpoint\n", 691 s, __func__); 692 TCPSTAT_INC(tcps_sc_reset); 693 } else { 694 TCPSTAT_INC(tcps_badrst); 695 /* Send challenge ACK. */ 696 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 697 log(LOG_DEBUG, "%s; %s: RST with invalid " 698 " SEQ %u != NXT %u (+WND %u), " 699 "sending challenge ACK\n", 700 s, __func__, 701 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 702 syncache_send_challenge_ack(sc); 703 } 704 } else { 705 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 706 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 707 "NXT %u (+WND %u), segment ignored\n", 708 s, __func__, 709 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 710 TCPSTAT_INC(tcps_badrst); 711 } 712 713 done: 714 if (s != NULL) 715 free(s, M_TCPLOG); 716 SCH_UNLOCK(sch); 717 } 718 719 void 720 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port) 721 { 722 struct syncache *sc; 723 struct syncache_head *sch; 724 725 if (syncache_cookiesonly()) 726 return; 727 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 728 SCH_LOCK_ASSERT(sch); 729 if (sc == NULL) 730 goto done; 731 732 /* If the port != sc_port, then it's a bogus ICMP msg */ 733 if (port != sc->sc_port) 734 goto done; 735 736 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 737 if (ntohl(th_seq) != sc->sc_iss) 738 goto done; 739 740 /* 741 * If we've retransmitted 3 times and this is our second error, 742 * we remove the entry. Otherwise, we allow it to continue on. 743 * This prevents us from incorrectly nuking an entry during a 744 * spurious network outage. 745 * 746 * See tcp_notify(). 747 */ 748 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 749 sc->sc_flags |= SCF_UNREACH; 750 goto done; 751 } 752 syncache_drop(sc, sch); 753 TCPSTAT_INC(tcps_sc_unreach); 754 done: 755 SCH_UNLOCK(sch); 756 } 757 758 /* 759 * Build a new TCP socket structure from a syncache entry. 760 * 761 * On success return the newly created socket with its underlying inp locked. 762 */ 763 static struct socket * 764 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 765 { 766 struct inpcb *inp = NULL; 767 struct socket *so; 768 struct tcpcb *tp; 769 int error; 770 char *s; 771 772 NET_EPOCH_ASSERT(); 773 774 /* 775 * Ok, create the full blown connection, and set things up 776 * as they would have been set up if we had created the 777 * connection when the SYN arrived. 778 */ 779 if ((so = solisten_clone(lso)) == NULL) 780 goto allocfail; 781 #ifdef MAC 782 mac_socketpeer_set_from_mbuf(m, so); 783 #endif 784 error = in_pcballoc(so, &V_tcbinfo); 785 if (error) { 786 sodealloc(so); 787 goto allocfail; 788 } 789 inp = sotoinpcb(so); 790 if ((tp = tcp_newtcpcb(inp, sototcpcb(lso))) == NULL) { 791 in_pcbfree(inp); 792 sodealloc(so); 793 goto allocfail; 794 } 795 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 796 #ifdef INET6 797 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 798 inp->inp_vflag &= ~INP_IPV4; 799 inp->inp_vflag |= INP_IPV6; 800 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 801 } else { 802 inp->inp_vflag &= ~INP_IPV6; 803 inp->inp_vflag |= INP_IPV4; 804 #endif 805 inp->inp_ip_ttl = sc->sc_ip_ttl; 806 inp->inp_ip_tos = sc->sc_ip_tos; 807 inp->inp_laddr = sc->sc_inc.inc_laddr; 808 #ifdef INET6 809 } 810 #endif 811 inp->inp_lport = sc->sc_inc.inc_lport; 812 #ifdef INET6 813 if (inp->inp_vflag & INP_IPV6PROTO) { 814 struct inpcb *oinp = sotoinpcb(lso); 815 816 /* 817 * Inherit socket options from the listening socket. 818 * Note that in6p_inputopts are not (and should not be) 819 * copied, since it stores previously received options and is 820 * used to detect if each new option is different than the 821 * previous one and hence should be passed to a user. 822 * If we copied in6p_inputopts, a user would not be able to 823 * receive options just after calling the accept system call. 824 */ 825 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 826 if (oinp->in6p_outputopts) 827 inp->in6p_outputopts = 828 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 829 inp->in6p_hops = oinp->in6p_hops; 830 } 831 832 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 833 struct sockaddr_in6 sin6; 834 835 sin6.sin6_family = AF_INET6; 836 sin6.sin6_len = sizeof(sin6); 837 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 838 sin6.sin6_port = sc->sc_inc.inc_fport; 839 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 840 INP_HASH_WLOCK(&V_tcbinfo); 841 error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false); 842 INP_HASH_WUNLOCK(&V_tcbinfo); 843 if (error != 0) 844 goto abort; 845 /* Override flowlabel from in6_pcbconnect. */ 846 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 847 inp->inp_flow |= sc->sc_flowlabel; 848 } 849 #endif /* INET6 */ 850 #if defined(INET) && defined(INET6) 851 else 852 #endif 853 #ifdef INET 854 { 855 struct sockaddr_in sin; 856 857 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 858 859 if (inp->inp_options == NULL) { 860 inp->inp_options = sc->sc_ipopts; 861 sc->sc_ipopts = NULL; 862 } 863 864 sin.sin_family = AF_INET; 865 sin.sin_len = sizeof(sin); 866 sin.sin_addr = sc->sc_inc.inc_faddr; 867 sin.sin_port = sc->sc_inc.inc_fport; 868 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 869 INP_HASH_WLOCK(&V_tcbinfo); 870 error = in_pcbconnect(inp, &sin, thread0.td_ucred); 871 INP_HASH_WUNLOCK(&V_tcbinfo); 872 if (error != 0) 873 goto abort; 874 } 875 #endif /* INET */ 876 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 877 /* Copy old policy into new socket's. */ 878 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 879 printf("syncache_socket: could not copy policy\n"); 880 #endif 881 if (sc->sc_flowtype != M_HASHTYPE_NONE) { 882 inp->inp_flowid = sc->sc_flowid; 883 inp->inp_flowtype = sc->sc_flowtype; 884 } else { 885 /* assign flowid by software RSS hash */ 886 #ifdef INET6 887 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 888 rss_proto_software_hash_v6(&inp->in6p_faddr, 889 &inp->in6p_laddr, 890 inp->inp_fport, 891 inp->inp_lport, 892 IPPROTO_TCP, 893 &inp->inp_flowid, 894 &inp->inp_flowtype); 895 } else 896 #endif /* INET6 */ 897 { 898 #ifdef INET 899 rss_proto_software_hash_v4(inp->inp_faddr, 900 inp->inp_laddr, 901 inp->inp_fport, 902 inp->inp_lport, 903 IPPROTO_TCP, 904 &inp->inp_flowid, 905 &inp->inp_flowtype); 906 #endif /* INET */ 907 } 908 } 909 #ifdef NUMA 910 inp->inp_numa_domain = sc->sc_numa_domain; 911 #endif 912 913 tp->t_state = TCPS_SYN_RECEIVED; 914 tp->iss = sc->sc_iss; 915 tp->irs = sc->sc_irs; 916 tp->t_port = sc->sc_port; 917 tcp_rcvseqinit(tp); 918 tcp_sendseqinit(tp); 919 tp->snd_wl1 = sc->sc_irs; 920 tp->snd_max = tp->iss + 1; 921 tp->snd_nxt = tp->iss + 1; 922 tp->rcv_up = sc->sc_irs + 1; 923 tp->rcv_wnd = sc->sc_wnd; 924 tp->rcv_adv += tp->rcv_wnd; 925 tp->last_ack_sent = tp->rcv_nxt; 926 927 tp->t_flags = sototcpcb(lso)->t_flags & 928 (TF_LRD|TF_NOPUSH|TF_NODELAY); 929 if (sc->sc_flags & SCF_NOOPT) 930 tp->t_flags |= TF_NOOPT; 931 else { 932 if (sc->sc_flags & SCF_WINSCALE) { 933 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 934 tp->snd_scale = sc->sc_requested_s_scale; 935 tp->request_r_scale = sc->sc_requested_r_scale; 936 } 937 if (sc->sc_flags & SCF_TIMESTAMP) { 938 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 939 tp->ts_recent = sc->sc_tsreflect; 940 tp->ts_recent_age = tcp_ts_getticks(); 941 tp->ts_offset = sc->sc_tsoff; 942 } 943 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 944 if (sc->sc_flags & SCF_SIGNATURE) 945 tp->t_flags |= TF_SIGNATURE; 946 #endif 947 if (sc->sc_flags & SCF_SACK) 948 tp->t_flags |= TF_SACK_PERMIT; 949 } 950 951 tcp_ecn_syncache_socket(tp, sc); 952 953 /* 954 * Set up MSS and get cached values from tcp_hostcache. 955 * This might overwrite some of the defaults we just set. 956 */ 957 tcp_mss(tp, sc->sc_peer_mss); 958 959 /* 960 * If the SYN,ACK was retransmitted, indicate that CWND to be 961 * limited to one segment in cc_conn_init(). 962 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 963 */ 964 if (sc->sc_rxmits > 1) 965 tp->snd_cwnd = 1; 966 967 /* Copy over the challenge ACK state. */ 968 tp->t_challenge_ack_end = sc->sc_challenge_ack_end; 969 tp->t_challenge_ack_cnt = sc->sc_challenge_ack_cnt; 970 971 #ifdef TCP_OFFLOAD 972 /* 973 * Allow a TOE driver to install its hooks. Note that we hold the 974 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 975 * new connection before the TOE driver has done its thing. 976 */ 977 if (ADDED_BY_TOE(sc)) { 978 struct toedev *tod = sc->sc_tod; 979 980 tod->tod_offload_socket(tod, sc->sc_todctx, so); 981 } 982 #endif 983 #ifdef TCP_BLACKBOX 984 /* 985 * Inherit the log state from the listening socket, if 986 * - the log state of the listening socket is not off and 987 * - the listening socket was not auto selected from all sessions and 988 * - a log id is not set on the listening socket. 989 * This avoids inheriting a log state which was automatically set. 990 */ 991 if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) && 992 ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) && 993 (sototcpcb(lso)->t_lib == NULL)) { 994 tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso))); 995 } 996 #endif 997 /* 998 * Copy and activate timers. 999 */ 1000 tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime; 1001 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 1002 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 1003 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1004 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1005 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1006 1007 TCPSTAT_INC(tcps_accepts); 1008 TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN); 1009 1010 if (!solisten_enqueue(so, SS_ISCONNECTED)) 1011 tp->t_flags |= TF_SONOTCONN; 1012 /* Can we inherit anything from the listener? */ 1013 if (tp->t_fb->tfb_inherit != NULL) { 1014 (*tp->t_fb->tfb_inherit)(tp, sotoinpcb(lso)); 1015 } 1016 return (so); 1017 1018 allocfail: 1019 /* 1020 * Drop the connection; we will either send a RST or have the peer 1021 * retransmit its SYN again after its RTO and try again. 1022 */ 1023 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 1024 log(LOG_DEBUG, "%s; %s: Socket create failed " 1025 "due to limits or memory shortage\n", 1026 s, __func__); 1027 free(s, M_TCPLOG); 1028 } 1029 TCPSTAT_INC(tcps_listendrop); 1030 return (NULL); 1031 1032 abort: 1033 tcp_discardcb(tp); 1034 in_pcbfree(inp); 1035 sodealloc(so); 1036 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 1037 log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n", 1038 s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "", 1039 error); 1040 free(s, M_TCPLOG); 1041 } 1042 TCPSTAT_INC(tcps_listendrop); 1043 return (NULL); 1044 } 1045 1046 /* 1047 * This function gets called when we receive an ACK for a 1048 * socket in the LISTEN state. We look up the connection 1049 * in the syncache, and if its there, we pull it out of 1050 * the cache and turn it into a full-blown connection in 1051 * the SYN-RECEIVED state. 1052 * 1053 * On syncache_socket() success the newly created socket 1054 * has its underlying inp locked. 1055 * 1056 * *lsop is updated, if and only if 1 is returned. 1057 */ 1058 int 1059 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1060 struct socket **lsop, struct mbuf *m, uint16_t port) 1061 { 1062 struct syncache *sc; 1063 struct syncache_head *sch; 1064 struct syncache scs; 1065 char *s; 1066 bool locked; 1067 1068 NET_EPOCH_ASSERT(); 1069 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1070 ("%s: can handle only ACK", __func__)); 1071 1072 if (syncache_cookiesonly()) { 1073 sc = NULL; 1074 sch = syncache_hashbucket(inc); 1075 locked = false; 1076 } else { 1077 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1078 locked = true; 1079 SCH_LOCK_ASSERT(sch); 1080 } 1081 1082 #ifdef INVARIANTS 1083 /* 1084 * Test code for syncookies comparing the syncache stored 1085 * values with the reconstructed values from the cookie. 1086 */ 1087 if (sc != NULL) 1088 syncookie_cmp(inc, sch, sc, th, to, *lsop, port); 1089 #endif 1090 1091 if (sc == NULL) { 1092 if (locked) { 1093 /* 1094 * The syncache is currently in use (neither disabled, 1095 * nor paused), but no entry was found. 1096 */ 1097 if (!V_tcp_syncookies) { 1098 /* 1099 * Since no syncookies are used in case of 1100 * a bucket overflow, don't even check for 1101 * a valid syncookie. 1102 */ 1103 SCH_UNLOCK(sch); 1104 TCPSTAT_INC(tcps_sc_spurcookie); 1105 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1106 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1107 "segment rejected " 1108 "(syncookies disabled)\n", 1109 s, __func__); 1110 free(s, M_TCPLOG); 1111 } 1112 return (0); 1113 } 1114 if (sch->sch_last_overflow < 1115 time_uptime - SYNCOOKIE_LIFETIME) { 1116 /* 1117 * Since the bucket did not overflow recently, 1118 * don't even check for a valid syncookie. 1119 */ 1120 SCH_UNLOCK(sch); 1121 TCPSTAT_INC(tcps_sc_spurcookie); 1122 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1123 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1124 "segment rejected " 1125 "(no syncache entry)\n", 1126 s, __func__); 1127 free(s, M_TCPLOG); 1128 } 1129 return (0); 1130 } 1131 SCH_UNLOCK(sch); 1132 } 1133 bzero(&scs, sizeof(scs)); 1134 /* 1135 * Now check, if the syncookie is valid. If it is, create an on 1136 * stack syncache entry. 1137 */ 1138 if (syncookie_expand(inc, sch, &scs, th, to, *lsop, port)) { 1139 sc = &scs; 1140 TCPSTAT_INC(tcps_sc_recvcookie); 1141 } else { 1142 TCPSTAT_INC(tcps_sc_failcookie); 1143 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1144 log(LOG_DEBUG, "%s; %s: Segment failed " 1145 "SYNCOOKIE authentication, segment rejected " 1146 "(probably spoofed)\n", s, __func__); 1147 free(s, M_TCPLOG); 1148 } 1149 return (0); 1150 } 1151 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1152 /* If received ACK has MD5 signature, check it. */ 1153 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1154 (!TCPMD5_ENABLED() || 1155 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1156 /* Drop the ACK. */ 1157 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1158 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1159 "MD5 signature doesn't match.\n", 1160 s, __func__); 1161 free(s, M_TCPLOG); 1162 } 1163 TCPSTAT_INC(tcps_sig_err_sigopt); 1164 return (-1); /* Do not send RST */ 1165 } 1166 #endif /* TCP_SIGNATURE */ 1167 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 1168 sc->sc_flowid = m->m_pkthdr.flowid; 1169 sc->sc_flowtype = M_HASHTYPE_GET(m); 1170 } 1171 #ifdef NUMA 1172 sc->sc_numa_domain = m ? m->m_pkthdr.numa_domain : M_NODOM; 1173 #endif 1174 TCPSTATES_INC(TCPS_SYN_RECEIVED); 1175 } else { 1176 if (sc->sc_port != port) { 1177 SCH_UNLOCK(sch); 1178 return (0); 1179 } 1180 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1181 /* 1182 * If listening socket requested TCP digests, check that 1183 * received ACK has signature and it is correct. 1184 * If not, drop the ACK and leave sc entry in the cache, 1185 * because SYN was received with correct signature. 1186 */ 1187 if (sc->sc_flags & SCF_SIGNATURE) { 1188 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1189 /* No signature */ 1190 TCPSTAT_INC(tcps_sig_err_nosigopt); 1191 SCH_UNLOCK(sch); 1192 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1193 log(LOG_DEBUG, "%s; %s: Segment " 1194 "rejected, MD5 signature wasn't " 1195 "provided.\n", s, __func__); 1196 free(s, M_TCPLOG); 1197 } 1198 return (-1); /* Do not send RST */ 1199 } 1200 if (!TCPMD5_ENABLED() || 1201 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1202 /* Doesn't match or no SA */ 1203 SCH_UNLOCK(sch); 1204 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1205 log(LOG_DEBUG, "%s; %s: Segment " 1206 "rejected, MD5 signature doesn't " 1207 "match.\n", s, __func__); 1208 free(s, M_TCPLOG); 1209 } 1210 return (-1); /* Do not send RST */ 1211 } 1212 } 1213 #endif /* TCP_SIGNATURE */ 1214 1215 /* 1216 * RFC 7323 PAWS: If we have a timestamp on this segment and 1217 * it's less than ts_recent, drop it. 1218 * XXXMT: RFC 7323 also requires to send an ACK. 1219 * In tcp_input.c this is only done for TCP segments 1220 * with user data, so be consistent here and just drop 1221 * the segment. 1222 */ 1223 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1224 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1225 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1226 log(LOG_DEBUG, 1227 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1228 "segment dropped\n", s, __func__, 1229 to->to_tsval, sc->sc_tsreflect); 1230 } 1231 SCH_UNLOCK(sch); 1232 free(s, M_TCPLOG); 1233 return (-1); /* Do not send RST */ 1234 } 1235 1236 /* 1237 * If timestamps were not negotiated during SYN/ACK and a 1238 * segment with a timestamp is received, ignore the 1239 * timestamp and process the packet normally. 1240 * See section 3.2 of RFC 7323. 1241 */ 1242 if (!(sc->sc_flags & SCF_TIMESTAMP) && 1243 (to->to_flags & TOF_TS)) { 1244 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1245 log(LOG_DEBUG, "%s; %s: Timestamp not " 1246 "expected, segment processed normally\n", 1247 s, __func__); 1248 free(s, M_TCPLOG); 1249 } 1250 } 1251 1252 /* 1253 * If timestamps were negotiated during SYN/ACK and a 1254 * segment without a timestamp is received, silently drop 1255 * the segment, unless the missing timestamps are tolerated. 1256 * See section 3.2 of RFC 7323. 1257 */ 1258 if ((sc->sc_flags & SCF_TIMESTAMP) && 1259 !(to->to_flags & TOF_TS)) { 1260 if (V_tcp_tolerate_missing_ts) { 1261 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1262 log(LOG_DEBUG, 1263 "%s; %s: Timestamp missing, " 1264 "segment processed normally\n", 1265 s, __func__); 1266 free(s, M_TCPLOG); 1267 } 1268 } else { 1269 SCH_UNLOCK(sch); 1270 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1271 log(LOG_DEBUG, 1272 "%s; %s: Timestamp missing, " 1273 "segment silently dropped\n", 1274 s, __func__); 1275 free(s, M_TCPLOG); 1276 } 1277 return (-1); /* Do not send RST */ 1278 } 1279 } 1280 1281 /* 1282 * SEG.SEQ validation: 1283 * The SEG.SEQ must be in the window starting at our 1284 * initial receive sequence number + 1. 1285 */ 1286 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1287 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1288 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1289 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, " 1290 "sending challenge ACK\n", 1291 s, __func__, th->th_seq, sc->sc_irs + 1); 1292 syncache_send_challenge_ack(sc); 1293 SCH_UNLOCK(sch); 1294 free(s, M_TCPLOG); 1295 return (-1); /* Do not send RST */ 1296 } 1297 1298 /* 1299 * SEG.ACK validation: 1300 * SEG.ACK must match our initial send sequence number + 1. 1301 */ 1302 if (th->th_ack != sc->sc_iss + 1) { 1303 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1304 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, " 1305 "segment rejected\n", 1306 s, __func__, th->th_ack, sc->sc_iss + 1); 1307 SCH_UNLOCK(sch); 1308 free(s, M_TCPLOG); 1309 return (0); /* Do send RST, do not free sc. */ 1310 } 1311 1312 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1313 sch->sch_length--; 1314 #ifdef TCP_OFFLOAD 1315 if (ADDED_BY_TOE(sc)) { 1316 struct toedev *tod = sc->sc_tod; 1317 1318 tod->tod_syncache_removed(tod, sc->sc_todctx); 1319 } 1320 #endif 1321 SCH_UNLOCK(sch); 1322 } 1323 1324 *lsop = syncache_socket(sc, *lsop, m); 1325 1326 if (__predict_false(*lsop == NULL)) { 1327 TCPSTAT_INC(tcps_sc_aborted); 1328 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1329 } else if (sc != &scs) 1330 TCPSTAT_INC(tcps_sc_completed); 1331 1332 if (sc != &scs) 1333 syncache_free(sc); 1334 return (1); 1335 } 1336 1337 static struct socket * 1338 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m, 1339 uint64_t response_cookie) 1340 { 1341 struct inpcb *inp; 1342 struct tcpcb *tp; 1343 unsigned int *pending_counter; 1344 struct socket *so; 1345 1346 NET_EPOCH_ASSERT(); 1347 1348 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending; 1349 so = syncache_socket(sc, lso, m); 1350 if (so == NULL) { 1351 TCPSTAT_INC(tcps_sc_aborted); 1352 atomic_subtract_int(pending_counter, 1); 1353 } else { 1354 soisconnected(so); 1355 inp = sotoinpcb(so); 1356 tp = intotcpcb(inp); 1357 tp->t_flags |= TF_FASTOPEN; 1358 tp->t_tfo_cookie.server = response_cookie; 1359 tp->snd_max = tp->iss; 1360 tp->snd_nxt = tp->iss; 1361 tp->t_tfo_pending = pending_counter; 1362 TCPSTATES_INC(TCPS_SYN_RECEIVED); 1363 TCPSTAT_INC(tcps_sc_completed); 1364 } 1365 1366 return (so); 1367 } 1368 1369 /* 1370 * Given a LISTEN socket and an inbound SYN request, add 1371 * this to the syn cache, and send back a segment: 1372 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1373 * to the source. 1374 * 1375 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1376 * Doing so would require that we hold onto the data and deliver it 1377 * to the application. However, if we are the target of a SYN-flood 1378 * DoS attack, an attacker could send data which would eventually 1379 * consume all available buffer space if it were ACKed. By not ACKing 1380 * the data, we avoid this DoS scenario. 1381 * 1382 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1383 * cookie is processed and a new socket is created. In this case, any data 1384 * accompanying the SYN will be queued to the socket by tcp_input() and will 1385 * be ACKed either when the application sends response data or the delayed 1386 * ACK timer expires, whichever comes first. 1387 */ 1388 struct socket * 1389 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1390 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod, 1391 void *todctx, uint8_t iptos, uint16_t port) 1392 { 1393 struct tcpcb *tp; 1394 struct socket *rv = NULL; 1395 struct syncache *sc = NULL; 1396 struct ucred *cred; 1397 struct syncache_head *sch; 1398 struct mbuf *ipopts = NULL; 1399 u_int ltflags; 1400 int win, ip_ttl, ip_tos; 1401 char *s; 1402 #ifdef INET6 1403 int autoflowlabel = 0; 1404 #endif 1405 #ifdef MAC 1406 struct label *maclabel = NULL; 1407 #endif 1408 struct syncache scs; 1409 uint64_t tfo_response_cookie; 1410 unsigned int *tfo_pending = NULL; 1411 int tfo_cookie_valid = 0; 1412 int tfo_response_cookie_valid = 0; 1413 bool locked; 1414 1415 INP_RLOCK_ASSERT(inp); /* listen socket */ 1416 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1417 ("%s: unexpected tcp flags", __func__)); 1418 1419 /* 1420 * Combine all so/tp operations very early to drop the INP lock as 1421 * soon as possible. 1422 */ 1423 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1424 tp = sototcpcb(so); 1425 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred); 1426 1427 #ifdef INET6 1428 if (inc->inc_flags & INC_ISIPV6) { 1429 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1430 autoflowlabel = 1; 1431 } 1432 ip_ttl = in6_selecthlim(inp, NULL); 1433 if ((inp->in6p_outputopts == NULL) || 1434 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1435 ip_tos = 0; 1436 } else { 1437 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1438 } 1439 } 1440 #endif 1441 #if defined(INET6) && defined(INET) 1442 else 1443 #endif 1444 #ifdef INET 1445 { 1446 ip_ttl = inp->inp_ip_ttl; 1447 ip_tos = inp->inp_ip_tos; 1448 } 1449 #endif 1450 win = so->sol_sbrcv_hiwat; 1451 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1452 1453 if (V_tcp_fastopen_server_enable && (tp->t_flags & TF_FASTOPEN) && 1454 (tp->t_tfo_pending != NULL) && 1455 (to->to_flags & TOF_FASTOPEN)) { 1456 /* 1457 * Limit the number of pending TFO connections to 1458 * approximately half of the queue limit. This prevents TFO 1459 * SYN floods from starving the service by filling the 1460 * listen queue with bogus TFO connections. 1461 */ 1462 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1463 (so->sol_qlimit / 2)) { 1464 int result; 1465 1466 result = tcp_fastopen_check_cookie(inc, 1467 to->to_tfo_cookie, to->to_tfo_len, 1468 &tfo_response_cookie); 1469 tfo_cookie_valid = (result > 0); 1470 tfo_response_cookie_valid = (result >= 0); 1471 } 1472 1473 /* 1474 * Remember the TFO pending counter as it will have to be 1475 * decremented below if we don't make it to syncache_tfo_expand(). 1476 */ 1477 tfo_pending = tp->t_tfo_pending; 1478 } 1479 1480 #ifdef MAC 1481 if (mac_syncache_init(&maclabel) != 0) { 1482 INP_RUNLOCK(inp); 1483 goto done; 1484 } else 1485 mac_syncache_create(maclabel, inp); 1486 #endif 1487 if (!tfo_cookie_valid) 1488 INP_RUNLOCK(inp); 1489 1490 /* 1491 * Remember the IP options, if any. 1492 */ 1493 #ifdef INET6 1494 if (!(inc->inc_flags & INC_ISIPV6)) 1495 #endif 1496 #ifdef INET 1497 ipopts = (m) ? ip_srcroute(m) : NULL; 1498 #else 1499 ipopts = NULL; 1500 #endif 1501 1502 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1503 /* 1504 * When the socket is TCP-MD5 enabled check that, 1505 * - a signed packet is valid 1506 * - a non-signed packet does not have a security association 1507 * 1508 * If a signed packet fails validation or a non-signed packet has a 1509 * security association, the packet will be dropped. 1510 */ 1511 if (ltflags & TF_SIGNATURE) { 1512 if (to->to_flags & TOF_SIGNATURE) { 1513 if (!TCPMD5_ENABLED() || 1514 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1515 goto done; 1516 } else { 1517 if (TCPMD5_ENABLED() && 1518 TCPMD5_INPUT(m, NULL, NULL) != ENOENT) 1519 goto done; 1520 } 1521 } else if (to->to_flags & TOF_SIGNATURE) 1522 goto done; 1523 #endif /* TCP_SIGNATURE */ 1524 /* 1525 * See if we already have an entry for this connection. 1526 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1527 * 1528 * XXX: should the syncache be re-initialized with the contents 1529 * of the new SYN here (which may have different options?) 1530 * 1531 * XXX: We do not check the sequence number to see if this is a 1532 * real retransmit or a new connection attempt. The question is 1533 * how to handle such a case; either ignore it as spoofed, or 1534 * drop the current entry and create a new one? 1535 */ 1536 if (syncache_cookiesonly()) { 1537 sc = NULL; 1538 sch = syncache_hashbucket(inc); 1539 locked = false; 1540 } else { 1541 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1542 locked = true; 1543 SCH_LOCK_ASSERT(sch); 1544 } 1545 if (sc != NULL) { 1546 if (tfo_cookie_valid) 1547 INP_RUNLOCK(inp); 1548 TCPSTAT_INC(tcps_sc_dupsyn); 1549 if (ipopts) { 1550 /* 1551 * If we were remembering a previous source route, 1552 * forget it and use the new one we've been given. 1553 */ 1554 if (sc->sc_ipopts) 1555 (void)m_free(sc->sc_ipopts); 1556 sc->sc_ipopts = ipopts; 1557 } 1558 /* 1559 * Update timestamp if present. 1560 */ 1561 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1562 sc->sc_tsreflect = to->to_tsval; 1563 else 1564 sc->sc_flags &= ~SCF_TIMESTAMP; 1565 /* 1566 * Adjust ECN response if needed, e.g. different 1567 * IP ECN field, or a fallback by the remote host. 1568 */ 1569 if (sc->sc_flags & SCF_ECN_MASK) { 1570 sc->sc_flags &= ~SCF_ECN_MASK; 1571 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos); 1572 } 1573 #ifdef MAC 1574 /* 1575 * Since we have already unconditionally allocated label 1576 * storage, free it up. The syncache entry will already 1577 * have an initialized label we can use. 1578 */ 1579 mac_syncache_destroy(&maclabel); 1580 #endif 1581 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1582 /* Retransmit SYN|ACK and reset retransmit count. */ 1583 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1584 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1585 "resetting timer and retransmitting SYN|ACK\n", 1586 s, __func__); 1587 free(s, M_TCPLOG); 1588 } 1589 if (syncache_respond(sc, TH_SYN|TH_ACK) == 0) { 1590 sc->sc_rxmits = 0; 1591 syncache_timeout(sc, sch, 1); 1592 TCPSTAT_INC(tcps_sndacks); 1593 TCPSTAT_INC(tcps_sndtotal); 1594 } else { 1595 /* 1596 * Most likely we are memory constrained, so free 1597 * resources. 1598 */ 1599 syncache_drop(sc, sch); 1600 TCPSTAT_INC(tcps_sc_dropped); 1601 } 1602 SCH_UNLOCK(sch); 1603 goto donenoprobe; 1604 } 1605 1606 KASSERT(sc == NULL, ("sc(%p) != NULL", sc)); 1607 /* 1608 * Skip allocating a syncache entry if we are just going to discard 1609 * it later. 1610 */ 1611 if (!locked || tfo_cookie_valid) { 1612 bzero(&scs, sizeof(scs)); 1613 sc = &scs; 1614 } else { 1615 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1616 if (sc == NULL) { 1617 /* 1618 * The zone allocator couldn't provide more entries. 1619 * Treat this as if the cache was full; drop the oldest 1620 * entry and insert the new one. 1621 */ 1622 TCPSTAT_INC(tcps_sc_zonefail); 1623 sc = TAILQ_LAST(&sch->sch_bucket, sch_head); 1624 if (sc != NULL) { 1625 sch->sch_last_overflow = time_uptime; 1626 syncache_drop(sc, sch); 1627 syncache_pause(inc); 1628 } 1629 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1630 if (sc == NULL) { 1631 if (V_tcp_syncookies) { 1632 bzero(&scs, sizeof(scs)); 1633 sc = &scs; 1634 } else { 1635 KASSERT(locked, 1636 ("%s: bucket unexpectedly unlocked", 1637 __func__)); 1638 SCH_UNLOCK(sch); 1639 goto done; 1640 } 1641 } 1642 } 1643 } 1644 1645 KASSERT(sc != NULL, ("sc == NULL")); 1646 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1647 sc->sc_tfo_cookie = &tfo_response_cookie; 1648 1649 /* 1650 * Fill in the syncache values. 1651 */ 1652 #ifdef MAC 1653 sc->sc_label = maclabel; 1654 #endif 1655 /* 1656 * sc_cred is only used in syncache_pcblist() to list TCP endpoints in 1657 * TCPS_SYN_RECEIVED state when V_tcp_syncache.see_other is false. 1658 * Therefore, store the credentials only when needed: 1659 * - sc is allocated from the zone and not using the on stack instance. 1660 * - the sysctl variable net.inet.tcp.syncache.see_other is false. 1661 * The reference count is decremented when a zone allocated sc is 1662 * freed in syncache_free(). 1663 */ 1664 if (sc != &scs && !V_tcp_syncache.see_other) { 1665 sc->sc_cred = cred; 1666 cred = NULL; 1667 } else 1668 sc->sc_cred = NULL; 1669 sc->sc_port = port; 1670 sc->sc_ipopts = ipopts; 1671 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1672 sc->sc_ip_tos = ip_tos; 1673 sc->sc_ip_ttl = ip_ttl; 1674 #ifdef TCP_OFFLOAD 1675 sc->sc_tod = tod; 1676 sc->sc_todctx = todctx; 1677 #endif 1678 sc->sc_irs = th->th_seq; 1679 sc->sc_flags = 0; 1680 sc->sc_flowlabel = 0; 1681 1682 /* 1683 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1684 * win was derived from socket earlier in the function. 1685 */ 1686 win = imax(win, 0); 1687 win = imin(win, TCP_MAXWIN); 1688 sc->sc_wnd = win; 1689 1690 if (V_tcp_do_rfc1323 && 1691 !(ltflags & TF_NOOPT)) { 1692 /* 1693 * A timestamp received in a SYN makes 1694 * it ok to send timestamp requests and replies. 1695 */ 1696 if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) { 1697 sc->sc_tsreflect = to->to_tsval; 1698 sc->sc_flags |= SCF_TIMESTAMP; 1699 sc->sc_tsoff = tcp_new_ts_offset(inc); 1700 } 1701 if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) { 1702 u_int wscale = 0; 1703 1704 /* 1705 * Pick the smallest possible scaling factor that 1706 * will still allow us to scale up to sb_max, aka 1707 * kern.ipc.maxsockbuf. 1708 * 1709 * We do this because there are broken firewalls that 1710 * will corrupt the window scale option, leading to 1711 * the other endpoint believing that our advertised 1712 * window is unscaled. At scale factors larger than 1713 * 5 the unscaled window will drop below 1500 bytes, 1714 * leading to serious problems when traversing these 1715 * broken firewalls. 1716 * 1717 * With the default maxsockbuf of 256K, a scale factor 1718 * of 3 will be chosen by this algorithm. Those who 1719 * choose a larger maxsockbuf should watch out 1720 * for the compatibility problems mentioned above. 1721 * 1722 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1723 * or <SYN,ACK>) segment itself is never scaled. 1724 */ 1725 while (wscale < TCP_MAX_WINSHIFT && 1726 (TCP_MAXWIN << wscale) < sb_max) 1727 wscale++; 1728 sc->sc_requested_r_scale = wscale; 1729 sc->sc_requested_s_scale = to->to_wscale; 1730 sc->sc_flags |= SCF_WINSCALE; 1731 } 1732 } 1733 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1734 /* 1735 * If incoming packet has an MD5 signature, flag this in the 1736 * syncache so that syncache_respond() will do the right thing 1737 * with the SYN+ACK. 1738 */ 1739 if (to->to_flags & TOF_SIGNATURE) 1740 sc->sc_flags |= SCF_SIGNATURE; 1741 #endif /* TCP_SIGNATURE */ 1742 if (to->to_flags & TOF_SACKPERM) 1743 sc->sc_flags |= SCF_SACK; 1744 if (to->to_flags & TOF_MSS) 1745 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1746 if (ltflags & TF_NOOPT) 1747 sc->sc_flags |= SCF_NOOPT; 1748 /* ECN Handshake */ 1749 if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0) 1750 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos); 1751 1752 if (V_tcp_syncookies || V_tcp_syncookiesonly) 1753 sc->sc_iss = syncookie_generate(sch, sc); 1754 else 1755 sc->sc_iss = arc4random(); 1756 #ifdef INET6 1757 if (autoflowlabel) { 1758 if (V_tcp_syncookies || V_tcp_syncookiesonly) 1759 sc->sc_flowlabel = sc->sc_iss; 1760 else 1761 sc->sc_flowlabel = ip6_randomflowlabel(); 1762 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1763 } 1764 #endif 1765 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 1766 sc->sc_flowid = m->m_pkthdr.flowid; 1767 sc->sc_flowtype = M_HASHTYPE_GET(m); 1768 } 1769 #ifdef NUMA 1770 sc->sc_numa_domain = m ? m->m_pkthdr.numa_domain : M_NODOM; 1771 #endif 1772 if (locked) 1773 SCH_UNLOCK(sch); 1774 1775 if (tfo_cookie_valid) { 1776 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie); 1777 /* INP_RUNLOCK(inp) will be performed by the caller */ 1778 goto tfo_expanded; 1779 } 1780 1781 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1782 /* 1783 * Do a standard 3-way handshake. 1784 */ 1785 if (syncache_respond(sc, TH_SYN|TH_ACK) == 0) { 1786 if (sc != &scs) 1787 syncache_insert(sc, sch); /* locks and unlocks sch */ 1788 TCPSTAT_INC(tcps_sndacks); 1789 TCPSTAT_INC(tcps_sndtotal); 1790 } else { 1791 /* 1792 * Most likely we are memory constrained, so free resources. 1793 */ 1794 if (sc != &scs) 1795 syncache_free(sc); 1796 TCPSTAT_INC(tcps_sc_dropped); 1797 } 1798 goto donenoprobe; 1799 1800 done: 1801 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1802 donenoprobe: 1803 if (m) 1804 m_freem(m); 1805 /* 1806 * If tfo_pending is not NULL here, then a TFO SYN that did not 1807 * result in a new socket was processed and the associated pending 1808 * counter has not yet been decremented. All such TFO processing paths 1809 * transit this point. 1810 */ 1811 if (tfo_pending != NULL) 1812 tcp_fastopen_decrement_counter(tfo_pending); 1813 1814 tfo_expanded: 1815 if (cred != NULL) 1816 crfree(cred); 1817 if (sc == NULL || sc == &scs) { 1818 #ifdef MAC 1819 mac_syncache_destroy(&maclabel); 1820 #endif 1821 if (ipopts) 1822 (void)m_free(ipopts); 1823 } 1824 return (rv); 1825 } 1826 1827 /* 1828 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment 1829 * or upon 3WHS ACK timeout. 1830 */ 1831 static int 1832 syncache_respond(struct syncache *sc, int flags) 1833 { 1834 struct ip *ip = NULL; 1835 struct mbuf *m; 1836 struct tcphdr *th = NULL; 1837 struct udphdr *udp = NULL; 1838 int optlen, error = 0; /* Make compiler happy */ 1839 u_int16_t hlen, tlen, mssopt, ulen; 1840 struct tcpopt to; 1841 #ifdef INET6 1842 struct ip6_hdr *ip6 = NULL; 1843 #endif 1844 1845 NET_EPOCH_ASSERT(); 1846 1847 hlen = 1848 #ifdef INET6 1849 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1850 #endif 1851 sizeof(struct ip); 1852 tlen = hlen + sizeof(struct tcphdr); 1853 if (sc->sc_port) { 1854 tlen += sizeof(struct udphdr); 1855 } 1856 /* Determine MSS we advertize to other end of connection. */ 1857 mssopt = tcp_mssopt(&sc->sc_inc); 1858 if (sc->sc_port) 1859 mssopt -= V_tcp_udp_tunneling_overhead; 1860 mssopt = max(mssopt, V_tcp_minmss); 1861 1862 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1863 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1864 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + " 1865 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port, 1866 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN)); 1867 1868 /* Create the IP+TCP header from scratch. */ 1869 m = m_gethdr(M_NOWAIT, MT_DATA); 1870 if (m == NULL) 1871 return (ENOBUFS); 1872 #ifdef MAC 1873 mac_syncache_create_mbuf(sc->sc_label, m); 1874 #endif 1875 m->m_data += max_linkhdr; 1876 m->m_len = tlen; 1877 m->m_pkthdr.len = tlen; 1878 m->m_pkthdr.rcvif = NULL; 1879 1880 #ifdef INET6 1881 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1882 ip6 = mtod(m, struct ip6_hdr *); 1883 ip6->ip6_vfc = IPV6_VERSION; 1884 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1885 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1886 ip6->ip6_plen = htons(tlen - hlen); 1887 /* ip6_hlim is set after checksum */ 1888 /* Zero out traffic class and flow label. */ 1889 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1890 ip6->ip6_flow |= sc->sc_flowlabel; 1891 if (sc->sc_port != 0) { 1892 ip6->ip6_nxt = IPPROTO_UDP; 1893 udp = (struct udphdr *)(ip6 + 1); 1894 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1895 udp->uh_dport = sc->sc_port; 1896 ulen = (tlen - sizeof(struct ip6_hdr)); 1897 th = (struct tcphdr *)(udp + 1); 1898 } else { 1899 ip6->ip6_nxt = IPPROTO_TCP; 1900 th = (struct tcphdr *)(ip6 + 1); 1901 } 1902 ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN); 1903 } 1904 #endif 1905 #if defined(INET6) && defined(INET) 1906 else 1907 #endif 1908 #ifdef INET 1909 { 1910 ip = mtod(m, struct ip *); 1911 ip->ip_v = IPVERSION; 1912 ip->ip_hl = sizeof(struct ip) >> 2; 1913 ip->ip_len = htons(tlen); 1914 ip->ip_id = 0; 1915 ip->ip_off = 0; 1916 ip->ip_sum = 0; 1917 ip->ip_src = sc->sc_inc.inc_laddr; 1918 ip->ip_dst = sc->sc_inc.inc_faddr; 1919 ip->ip_ttl = sc->sc_ip_ttl; 1920 ip->ip_tos = sc->sc_ip_tos; 1921 1922 /* 1923 * See if we should do MTU discovery. Route lookups are 1924 * expensive, so we will only unset the DF bit if: 1925 * 1926 * 1) path_mtu_discovery is disabled 1927 * 2) the SCF_UNREACH flag has been set 1928 */ 1929 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1930 ip->ip_off |= htons(IP_DF); 1931 if (sc->sc_port == 0) { 1932 ip->ip_p = IPPROTO_TCP; 1933 th = (struct tcphdr *)(ip + 1); 1934 } else { 1935 ip->ip_p = IPPROTO_UDP; 1936 udp = (struct udphdr *)(ip + 1); 1937 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1938 udp->uh_dport = sc->sc_port; 1939 ulen = (tlen - sizeof(struct ip)); 1940 th = (struct tcphdr *)(udp + 1); 1941 } 1942 } 1943 #endif /* INET */ 1944 th->th_sport = sc->sc_inc.inc_lport; 1945 th->th_dport = sc->sc_inc.inc_fport; 1946 1947 if (flags & TH_SYN) 1948 th->th_seq = htonl(sc->sc_iss); 1949 else 1950 th->th_seq = htonl(sc->sc_iss + 1); 1951 th->th_ack = htonl(sc->sc_irs + 1); 1952 th->th_off = sizeof(struct tcphdr) >> 2; 1953 th->th_win = htons(sc->sc_wnd); 1954 th->th_urp = 0; 1955 1956 flags = tcp_ecn_syncache_respond(flags, sc); 1957 tcp_set_flags(th, flags); 1958 1959 /* Tack on the TCP options. */ 1960 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1961 to.to_flags = 0; 1962 1963 if (flags & TH_SYN) { 1964 to.to_mss = mssopt; 1965 to.to_flags = TOF_MSS; 1966 if (sc->sc_flags & SCF_WINSCALE) { 1967 to.to_wscale = sc->sc_requested_r_scale; 1968 to.to_flags |= TOF_SCALE; 1969 } 1970 if (sc->sc_flags & SCF_SACK) 1971 to.to_flags |= TOF_SACKPERM; 1972 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1973 if (sc->sc_flags & SCF_SIGNATURE) 1974 to.to_flags |= TOF_SIGNATURE; 1975 #endif 1976 if (sc->sc_tfo_cookie) { 1977 to.to_flags |= TOF_FASTOPEN; 1978 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1979 to.to_tfo_cookie = sc->sc_tfo_cookie; 1980 /* don't send cookie again when retransmitting response */ 1981 sc->sc_tfo_cookie = NULL; 1982 } 1983 } 1984 if (sc->sc_flags & SCF_TIMESTAMP) { 1985 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1986 to.to_tsecr = sc->sc_tsreflect; 1987 to.to_flags |= TOF_TS; 1988 } 1989 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1990 1991 /* Adjust headers by option size. */ 1992 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1993 m->m_len += optlen; 1994 m->m_pkthdr.len += optlen; 1995 #ifdef INET6 1996 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1997 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1998 else 1999 #endif 2000 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 2001 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 2002 if (sc->sc_flags & SCF_SIGNATURE) { 2003 KASSERT(to.to_flags & TOF_SIGNATURE, 2004 ("tcp_addoptions() didn't set tcp_signature")); 2005 2006 /* NOTE: to.to_signature is inside of mbuf */ 2007 if (!TCPMD5_ENABLED() || 2008 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 2009 m_freem(m); 2010 return (EACCES); 2011 } 2012 } 2013 #endif 2014 } else 2015 optlen = 0; 2016 2017 if (udp) { 2018 ulen += optlen; 2019 udp->uh_ulen = htons(ulen); 2020 } 2021 M_SETFIB(m, sc->sc_inc.inc_fibnum); 2022 m->m_pkthdr.flowid = sc->sc_flowid; 2023 M_HASHTYPE_SET(m, sc->sc_flowtype); 2024 #ifdef NUMA 2025 m->m_pkthdr.numa_domain = sc->sc_numa_domain; 2026 #endif 2027 #ifdef INET6 2028 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 2029 if (sc->sc_port) { 2030 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 2031 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2032 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, 2033 IPPROTO_UDP, 0); 2034 th->th_sum = htons(0); 2035 } else { 2036 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 2037 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2038 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 2039 IPPROTO_TCP, 0); 2040 } 2041 ip6->ip6_hlim = sc->sc_ip_ttl; 2042 #ifdef TCP_OFFLOAD 2043 if (ADDED_BY_TOE(sc)) { 2044 struct toedev *tod = sc->sc_tod; 2045 2046 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2047 2048 return (error); 2049 } 2050 #endif 2051 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 2052 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 2053 } 2054 #endif 2055 #if defined(INET6) && defined(INET) 2056 else 2057 #endif 2058 #ifdef INET 2059 { 2060 if (sc->sc_port) { 2061 m->m_pkthdr.csum_flags = CSUM_UDP; 2062 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2063 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 2064 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 2065 th->th_sum = htons(0); 2066 } else { 2067 m->m_pkthdr.csum_flags = CSUM_TCP; 2068 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2069 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 2070 htons(tlen + optlen - hlen + IPPROTO_TCP)); 2071 } 2072 #ifdef TCP_OFFLOAD 2073 if (ADDED_BY_TOE(sc)) { 2074 struct toedev *tod = sc->sc_tod; 2075 2076 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2077 2078 return (error); 2079 } 2080 #endif 2081 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 2082 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 2083 } 2084 #endif 2085 return (error); 2086 } 2087 2088 static void 2089 syncache_send_challenge_ack(struct syncache *sc) 2090 { 2091 if (tcp_challenge_ack_check(&sc->sc_challenge_ack_end, 2092 &sc->sc_challenge_ack_cnt)) { 2093 if (syncache_respond(sc, TH_ACK) == 0) { 2094 TCPSTAT_INC(tcps_sndacks); 2095 TCPSTAT_INC(tcps_sndtotal); 2096 } 2097 } 2098 } 2099 2100 /* 2101 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 2102 * that exceed the capacity of the syncache by avoiding the storage of any 2103 * of the SYNs we receive. Syncookies defend against blind SYN flooding 2104 * attacks where the attacker does not have access to our responses. 2105 * 2106 * Syncookies encode and include all necessary information about the 2107 * connection setup within the SYN|ACK that we send back. That way we 2108 * can avoid keeping any local state until the ACK to our SYN|ACK returns 2109 * (if ever). Normally the syncache and syncookies are running in parallel 2110 * with the latter taking over when the former is exhausted. When matching 2111 * syncache entry is found the syncookie is ignored. 2112 * 2113 * The only reliable information persisting the 3WHS is our initial sequence 2114 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 2115 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 2116 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 2117 * returns and signifies a legitimate connection if it matches the ACK. 2118 * 2119 * The available space of 32 bits to store the hash and to encode the SYN 2120 * option information is very tight and we should have at least 24 bits for 2121 * the MAC to keep the number of guesses by blind spoofing reasonably high. 2122 * 2123 * SYN option information we have to encode to fully restore a connection: 2124 * MSS: is imporant to chose an optimal segment size to avoid IP level 2125 * fragmentation along the path. The common MSS values can be encoded 2126 * in a 3-bit table. Uncommon values are captured by the next lower value 2127 * in the table leading to a slight increase in packetization overhead. 2128 * WSCALE: is necessary to allow large windows to be used for high delay- 2129 * bandwidth product links. Not scaling the window when it was initially 2130 * negotiated is bad for performance as lack of scaling further decreases 2131 * the apparent available send window. We only need to encode the WSCALE 2132 * we received from the remote end. Our end can be recalculated at any 2133 * time. The common WSCALE values can be encoded in a 3-bit table. 2134 * Uncommon values are captured by the next lower value in the table 2135 * making us under-estimate the available window size halving our 2136 * theoretically possible maximum throughput for that connection. 2137 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 2138 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 2139 * that are included in all segments on a connection. We enable them when 2140 * the ACK has them. 2141 * 2142 * Security of syncookies and attack vectors: 2143 * 2144 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 2145 * together with the gloabl secret to make it unique per connection attempt. 2146 * Thus any change of any of those parameters results in a different MAC output 2147 * in an unpredictable way unless a collision is encountered. 24 bits of the 2148 * MAC are embedded into the ISS. 2149 * 2150 * To prevent replay attacks two rotating global secrets are updated with a 2151 * new random value every 15 seconds. The life-time of a syncookie is thus 2152 * 15-30 seconds. 2153 * 2154 * Vector 1: Attacking the secret. This requires finding a weakness in the 2155 * MAC itself or the way it is used here. The attacker can do a chosen plain 2156 * text attack by varying and testing the all parameters under his control. 2157 * The strength depends on the size and randomness of the secret, and the 2158 * cryptographic security of the MAC function. Due to the constant updating 2159 * of the secret the attacker has at most 29.999 seconds to find the secret 2160 * and launch spoofed connections. After that he has to start all over again. 2161 * 2162 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 2163 * size an average of 4,823 attempts are required for a 50% chance of success 2164 * to spoof a single syncookie (birthday collision paradox). However the 2165 * attacker is blind and doesn't know if one of his attempts succeeded unless 2166 * he has a side channel to interfere success from. A single connection setup 2167 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 2168 * This many attempts are required for each one blind spoofed connection. For 2169 * every additional spoofed connection he has to launch another N attempts. 2170 * Thus for a sustained rate 100 spoofed connections per second approximately 2171 * 1,800,000 packets per second would have to be sent. 2172 * 2173 * NB: The MAC function should be fast so that it doesn't become a CPU 2174 * exhaustion attack vector itself. 2175 * 2176 * References: 2177 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 2178 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 2179 * http://cr.yp.to/syncookies.html (overview) 2180 * http://cr.yp.to/syncookies/archive (details) 2181 * 2182 * 2183 * Schematic construction of a syncookie enabled Initial Sequence Number: 2184 * 0 1 2 3 2185 * 12345678901234567890123456789012 2186 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2187 * 2188 * x 24 MAC (truncated) 2189 * W 3 Send Window Scale index 2190 * M 3 MSS index 2191 * S 1 SACK permitted 2192 * P 1 Odd/even secret 2193 */ 2194 2195 /* 2196 * Distribution and probability of certain MSS values. Those in between are 2197 * rounded down to the next lower one. 2198 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2199 * .2% .3% 5% 7% 7% 20% 15% 45% 2200 */ 2201 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2202 2203 /* 2204 * Distribution and probability of certain WSCALE values. We have to map the 2205 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2206 * bits based on prevalence of certain values. Where we don't have an exact 2207 * match for are rounded down to the next lower one letting us under-estimate 2208 * the true available window. At the moment this would happen only for the 2209 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2210 * and window size). The absence of the WSCALE option (no scaling in either 2211 * direction) is encoded with index zero. 2212 * [WSCALE values histograms, Allman, 2012] 2213 * X 10 10 35 5 6 14 10% by host 2214 * X 11 4 5 5 18 49 3% by connections 2215 */ 2216 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2217 2218 /* 2219 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2220 * and good cryptographic properties. 2221 */ 2222 static uint32_t 2223 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2224 uint8_t *secbits, uintptr_t secmod) 2225 { 2226 SIPHASH_CTX ctx; 2227 uint32_t siphash[2]; 2228 2229 SipHash24_Init(&ctx); 2230 SipHash_SetKey(&ctx, secbits); 2231 switch (inc->inc_flags & INC_ISIPV6) { 2232 #ifdef INET 2233 case 0: 2234 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2235 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2236 break; 2237 #endif 2238 #ifdef INET6 2239 case INC_ISIPV6: 2240 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2241 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2242 break; 2243 #endif 2244 } 2245 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2246 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2247 SipHash_Update(&ctx, &irs, sizeof(irs)); 2248 SipHash_Update(&ctx, &flags, sizeof(flags)); 2249 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2250 SipHash_Final((u_int8_t *)&siphash, &ctx); 2251 2252 return (siphash[0] ^ siphash[1]); 2253 } 2254 2255 static tcp_seq 2256 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2257 { 2258 u_int i, secbit, wscale; 2259 uint32_t iss, hash; 2260 uint8_t *secbits; 2261 union syncookie cookie; 2262 2263 cookie.cookie = 0; 2264 2265 /* Map our computed MSS into the 3-bit index. */ 2266 for (i = nitems(tcp_sc_msstab) - 1; 2267 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2268 i--) 2269 ; 2270 cookie.flags.mss_idx = i; 2271 2272 /* 2273 * Map the send window scale into the 3-bit index but only if 2274 * the wscale option was received. 2275 */ 2276 if (sc->sc_flags & SCF_WINSCALE) { 2277 wscale = sc->sc_requested_s_scale; 2278 for (i = nitems(tcp_sc_wstab) - 1; 2279 tcp_sc_wstab[i] > wscale && i > 0; 2280 i--) 2281 ; 2282 cookie.flags.wscale_idx = i; 2283 } 2284 2285 /* Can we do SACK? */ 2286 if (sc->sc_flags & SCF_SACK) 2287 cookie.flags.sack_ok = 1; 2288 2289 /* Which of the two secrets to use. */ 2290 secbit = V_tcp_syncache.secret.oddeven & 0x1; 2291 cookie.flags.odd_even = secbit; 2292 2293 secbits = V_tcp_syncache.secret.key[secbit]; 2294 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2295 (uintptr_t)sch); 2296 2297 /* 2298 * Put the flags into the hash and XOR them to get better ISS number 2299 * variance. This doesn't enhance the cryptographic strength and is 2300 * done to prevent the 8 cookie bits from showing up directly on the 2301 * wire. 2302 */ 2303 iss = hash & ~0xff; 2304 iss |= cookie.cookie ^ (hash >> 24); 2305 2306 TCPSTAT_INC(tcps_sc_sendcookie); 2307 return (iss); 2308 } 2309 2310 static bool 2311 syncookie_expand(struct in_conninfo *inc, const struct syncache_head *sch, 2312 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2313 struct socket *lso, uint16_t port) 2314 { 2315 uint32_t hash; 2316 uint8_t *secbits; 2317 tcp_seq ack, seq; 2318 int wnd; 2319 union syncookie cookie; 2320 2321 /* 2322 * Pull information out of SYN-ACK/ACK and revert sequence number 2323 * advances. 2324 */ 2325 ack = th->th_ack - 1; 2326 seq = th->th_seq - 1; 2327 2328 /* 2329 * Unpack the flags containing enough information to restore the 2330 * connection. 2331 */ 2332 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2333 2334 /* Which of the two secrets to use. */ 2335 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; 2336 2337 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2338 2339 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2340 if ((ack & ~0xff) != (hash & ~0xff)) 2341 return (false); 2342 2343 /* Fill in the syncache values. */ 2344 sc->sc_flags = 0; 2345 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2346 sc->sc_ipopts = NULL; 2347 2348 sc->sc_irs = seq; 2349 sc->sc_iss = ack; 2350 2351 switch (inc->inc_flags & INC_ISIPV6) { 2352 #ifdef INET 2353 case 0: 2354 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2355 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2356 break; 2357 #endif 2358 #ifdef INET6 2359 case INC_ISIPV6: 2360 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2361 sc->sc_flowlabel = 2362 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2363 break; 2364 #endif 2365 } 2366 2367 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2368 2369 /* Only use wscale if it was enabled in the orignal SYN. */ 2370 if (cookie.flags.wscale_idx > 0) { 2371 u_int wscale = 0; 2372 2373 /* Recompute the receive window scale that was sent earlier. */ 2374 while (wscale < TCP_MAX_WINSHIFT && 2375 (TCP_MAXWIN << wscale) < sb_max) 2376 wscale++; 2377 sc->sc_requested_r_scale = wscale; 2378 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2379 sc->sc_flags |= SCF_WINSCALE; 2380 } 2381 2382 wnd = lso->sol_sbrcv_hiwat; 2383 wnd = imax(wnd, 0); 2384 wnd = imin(wnd, TCP_MAXWIN); 2385 sc->sc_wnd = wnd; 2386 2387 if (cookie.flags.sack_ok) 2388 sc->sc_flags |= SCF_SACK; 2389 2390 if (to->to_flags & TOF_TS) { 2391 sc->sc_flags |= SCF_TIMESTAMP; 2392 sc->sc_tsreflect = to->to_tsval; 2393 sc->sc_tsoff = tcp_new_ts_offset(inc); 2394 } 2395 2396 if (to->to_flags & TOF_SIGNATURE) 2397 sc->sc_flags |= SCF_SIGNATURE; 2398 2399 sc->sc_rxmits = 0; 2400 2401 sc->sc_port = port; 2402 2403 return (true); 2404 } 2405 2406 #ifdef INVARIANTS 2407 static void 2408 syncookie_cmp(struct in_conninfo *inc, const struct syncache_head *sch, 2409 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2410 struct socket *lso, uint16_t port) 2411 { 2412 struct syncache scs; 2413 char *s; 2414 2415 bzero(&scs, sizeof(scs)); 2416 if (syncookie_expand(inc, sch, &scs, th, to, lso, port) && 2417 (sc->sc_peer_mss != scs.sc_peer_mss || 2418 sc->sc_requested_r_scale != scs.sc_requested_r_scale || 2419 sc->sc_requested_s_scale != scs.sc_requested_s_scale || 2420 (sc->sc_flags & SCF_SACK) != (scs.sc_flags & SCF_SACK))) { 2421 2422 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2423 return; 2424 2425 if (sc->sc_peer_mss != scs.sc_peer_mss) 2426 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2427 s, __func__, sc->sc_peer_mss, scs.sc_peer_mss); 2428 2429 if (sc->sc_requested_r_scale != scs.sc_requested_r_scale) 2430 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2431 s, __func__, sc->sc_requested_r_scale, 2432 scs.sc_requested_r_scale); 2433 2434 if (sc->sc_requested_s_scale != scs.sc_requested_s_scale) 2435 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2436 s, __func__, sc->sc_requested_s_scale, 2437 scs.sc_requested_s_scale); 2438 2439 if ((sc->sc_flags & SCF_SACK) != (scs.sc_flags & SCF_SACK)) 2440 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2441 2442 free(s, M_TCPLOG); 2443 } 2444 } 2445 #endif /* INVARIANTS */ 2446 2447 static void 2448 syncookie_reseed(void *arg) 2449 { 2450 struct tcp_syncache *sc = arg; 2451 uint8_t *secbits; 2452 int secbit; 2453 2454 /* 2455 * Reseeding the secret doesn't have to be protected by a lock. 2456 * It only must be ensured that the new random values are visible 2457 * to all CPUs in a SMP environment. The atomic with release 2458 * semantics ensures that. 2459 */ 2460 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2461 secbits = sc->secret.key[secbit]; 2462 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2463 atomic_add_rel_int(&sc->secret.oddeven, 1); 2464 2465 /* Reschedule ourself. */ 2466 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2467 } 2468 2469 /* 2470 * We have overflowed a bucket. Let's pause dealing with the syncache. 2471 * This function will increment the bucketoverflow statistics appropriately 2472 * (once per pause when pausing is enabled; otherwise, once per overflow). 2473 */ 2474 static void 2475 syncache_pause(struct in_conninfo *inc) 2476 { 2477 time_t delta; 2478 const char *s; 2479 2480 /* XXX: 2481 * 2. Add sysctl read here so we don't get the benefit of this 2482 * change without the new sysctl. 2483 */ 2484 2485 /* 2486 * Try an unlocked read. If we already know that another thread 2487 * has activated the feature, there is no need to proceed. 2488 */ 2489 if (V_tcp_syncache.paused) 2490 return; 2491 2492 /* Are cookied enabled? If not, we can't pause. */ 2493 if (!V_tcp_syncookies) { 2494 TCPSTAT_INC(tcps_sc_bucketoverflow); 2495 return; 2496 } 2497 2498 /* 2499 * We may be the first thread to find an overflow. Get the lock 2500 * and evaluate if we need to take action. 2501 */ 2502 mtx_lock(&V_tcp_syncache.pause_mtx); 2503 if (V_tcp_syncache.paused) { 2504 mtx_unlock(&V_tcp_syncache.pause_mtx); 2505 return; 2506 } 2507 2508 /* Activate protection. */ 2509 V_tcp_syncache.paused = true; 2510 TCPSTAT_INC(tcps_sc_bucketoverflow); 2511 2512 /* 2513 * Determine the last backoff time. If we are seeing a re-newed 2514 * attack within that same time after last reactivating the syncache, 2515 * consider it an extension of the same attack. 2516 */ 2517 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; 2518 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { 2519 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { 2520 delta <<= 1; 2521 V_tcp_syncache.pause_backoff++; 2522 } 2523 } else { 2524 delta = TCP_SYNCACHE_PAUSE_TIME; 2525 V_tcp_syncache.pause_backoff = 0; 2526 } 2527 2528 /* Log a warning, including IP addresses, if able. */ 2529 if (inc != NULL) 2530 s = tcp_log_addrs(inc, NULL, NULL, NULL); 2531 else 2532 s = (const char *)NULL; 2533 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " 2534 "the next %lld seconds%s%s%s\n", (long long)delta, 2535 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", 2536 (s != NULL) ? ")" : ""); 2537 free(__DECONST(void *, s), M_TCPLOG); 2538 2539 /* Use the calculated delta to set a new pause time. */ 2540 V_tcp_syncache.pause_until = time_uptime + delta; 2541 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, 2542 &V_tcp_syncache); 2543 mtx_unlock(&V_tcp_syncache.pause_mtx); 2544 } 2545 2546 /* Evaluate whether we need to unpause. */ 2547 static void 2548 syncache_unpause(void *arg) 2549 { 2550 struct tcp_syncache *sc; 2551 time_t delta; 2552 2553 sc = arg; 2554 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); 2555 callout_deactivate(&sc->pause_co); 2556 2557 /* 2558 * Check to make sure we are not running early. If the pause 2559 * time has expired, then deactivate the protection. 2560 */ 2561 if ((delta = sc->pause_until - time_uptime) > 0) 2562 callout_schedule(&sc->pause_co, delta * hz); 2563 else 2564 sc->paused = false; 2565 } 2566 2567 /* 2568 * Exports the syncache entries to userland so that netstat can display 2569 * them alongside the other sockets. This function is intended to be 2570 * called only from tcp_pcblist. 2571 * 2572 * Due to concurrency on an active system, the number of pcbs exported 2573 * may have no relation to max_pcbs. max_pcbs merely indicates the 2574 * amount of space the caller allocated for this function to use. 2575 */ 2576 int 2577 syncache_pcblist(struct sysctl_req *req) 2578 { 2579 struct xtcpcb xt; 2580 struct syncache *sc; 2581 struct syncache_head *sch; 2582 int error, i; 2583 2584 bzero(&xt, sizeof(xt)); 2585 xt.xt_len = sizeof(xt); 2586 xt.t_state = TCPS_SYN_RECEIVED; 2587 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2588 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2589 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2590 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2591 2592 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 2593 sch = &V_tcp_syncache.hashbase[i]; 2594 SCH_LOCK(sch); 2595 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2596 if (sc->sc_cred != NULL && 2597 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2598 continue; 2599 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2600 xt.xt_inp.inp_vflag = INP_IPV6; 2601 else 2602 xt.xt_inp.inp_vflag = INP_IPV4; 2603 xt.xt_encaps_port = sc->sc_port; 2604 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2605 sizeof (struct in_conninfo)); 2606 error = SYSCTL_OUT(req, &xt, sizeof xt); 2607 if (error) { 2608 SCH_UNLOCK(sch); 2609 return (0); 2610 } 2611 } 2612 SCH_UNLOCK(sch); 2613 } 2614 2615 return (0); 2616 } 2617