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