147#define V_tcp_syncache VNET(tcp_syncache) 148 149SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache"); 150 151SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN, 152 &VNET_NAME(tcp_syncache.bucket_limit), 0, 153 "Per-bucket hash limit for syncache"); 154 155SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN, 156 &VNET_NAME(tcp_syncache.cache_limit), 0, 157 "Overall entry limit for syncache"); 158 159SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD, 160 &VNET_NAME(tcp_syncache.cache_count), 0, 161 "Current number of entries in syncache"); 162 163SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN, 164 &VNET_NAME(tcp_syncache.hashsize), 0, 165 "Size of TCP syncache hashtable"); 166 167SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW, 168 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 169 "Limit on SYN/ACK retransmissions"); 170 171VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 172SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 173 CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 174 "Send reset on socket allocation failure"); 175 176static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 177 178#define SYNCACHE_HASH(inc, mask) \ 179 ((V_tcp_syncache.hash_secret ^ \ 180 (inc)->inc_faddr.s_addr ^ \ 181 ((inc)->inc_faddr.s_addr >> 16) ^ \ 182 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 183 184#define SYNCACHE_HASH6(inc, mask) \ 185 ((V_tcp_syncache.hash_secret ^ \ 186 (inc)->inc6_faddr.s6_addr32[0] ^ \ 187 (inc)->inc6_faddr.s6_addr32[3] ^ \ 188 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 189 190#define ENDPTS_EQ(a, b) ( \ 191 (a)->ie_fport == (b)->ie_fport && \ 192 (a)->ie_lport == (b)->ie_lport && \ 193 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \ 194 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \ 195) 196 197#define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0) 198 199#define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 200#define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 201#define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 202 203/* 204 * Requires the syncache entry to be already removed from the bucket list. 205 */ 206static void 207syncache_free(struct syncache *sc) 208{ 209 210 if (sc->sc_ipopts) 211 (void) m_free(sc->sc_ipopts); 212 if (sc->sc_cred) 213 crfree(sc->sc_cred); 214#ifdef MAC 215 mac_syncache_destroy(&sc->sc_label); 216#endif 217 218 uma_zfree(V_tcp_syncache.zone, sc); 219} 220 221void 222syncache_init(void) 223{ 224 int i; 225 226 V_tcp_syncache.cache_count = 0; 227 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 228 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 229 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 230 V_tcp_syncache.hash_secret = arc4random(); 231 232 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 233 &V_tcp_syncache.hashsize); 234 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 235 &V_tcp_syncache.bucket_limit); 236 if (!powerof2(V_tcp_syncache.hashsize) || 237 V_tcp_syncache.hashsize == 0) { 238 printf("WARNING: syncache hash size is not a power of 2.\n"); 239 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 240 } 241 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 242 243 /* Set limits. */ 244 V_tcp_syncache.cache_limit = 245 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 246 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 247 &V_tcp_syncache.cache_limit); 248 249 /* Allocate the hash table. */ 250 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 251 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 252 253 /* Initialize the hash buckets. */ 254 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 255#ifdef VIMAGE 256 V_tcp_syncache.hashbase[i].sch_vnet = curvnet; 257#endif 258 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 259 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 260 NULL, MTX_DEF); 261 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 262 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 263 V_tcp_syncache.hashbase[i].sch_length = 0; 264 } 265 266 /* Create the syncache entry zone. */ 267 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 268 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 269 uma_zone_set_max(V_tcp_syncache.zone, V_tcp_syncache.cache_limit); 270} 271 272#ifdef VIMAGE 273void 274syncache_destroy(void) 275{ 276 struct syncache_head *sch; 277 struct syncache *sc, *nsc; 278 int i; 279 280 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 281 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 282 283 sch = &V_tcp_syncache.hashbase[i]; 284 callout_drain(&sch->sch_timer); 285 286 SCH_LOCK(sch); 287 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 288 syncache_drop(sc, sch); 289 SCH_UNLOCK(sch); 290 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 291 ("%s: sch->sch_bucket not empty", __func__)); 292 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 293 __func__, sch->sch_length)); 294 mtx_destroy(&sch->sch_mtx); 295 } 296 297 KASSERT(V_tcp_syncache.cache_count == 0, ("%s: cache_count %d not 0", 298 __func__, V_tcp_syncache.cache_count)); 299 300 /* Free the allocated global resources. */ 301 uma_zdestroy(V_tcp_syncache.zone); 302 free(V_tcp_syncache.hashbase, M_SYNCACHE); 303} 304#endif 305 306/* 307 * Inserts a syncache entry into the specified bucket row. 308 * Locks and unlocks the syncache_head autonomously. 309 */ 310static void 311syncache_insert(struct syncache *sc, struct syncache_head *sch) 312{ 313 struct syncache *sc2; 314 315 SCH_LOCK(sch); 316 317 /* 318 * Make sure that we don't overflow the per-bucket limit. 319 * If the bucket is full, toss the oldest element. 320 */ 321 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 322 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 323 ("sch->sch_length incorrect")); 324 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 325 syncache_drop(sc2, sch); 326 TCPSTAT_INC(tcps_sc_bucketoverflow); 327 } 328 329 /* Put it into the bucket. */ 330 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 331 sch->sch_length++; 332 333 /* Reinitialize the bucket row's timer. */ 334 if (sch->sch_length == 1) 335 sch->sch_nextc = ticks + INT_MAX; 336 syncache_timeout(sc, sch, 1); 337 338 SCH_UNLOCK(sch); 339 340 V_tcp_syncache.cache_count++; 341 TCPSTAT_INC(tcps_sc_added); 342} 343 344/* 345 * Remove and free entry from syncache bucket row. 346 * Expects locked syncache head. 347 */ 348static void 349syncache_drop(struct syncache *sc, struct syncache_head *sch) 350{ 351 352 SCH_LOCK_ASSERT(sch); 353 354 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 355 sch->sch_length--; 356 357#ifndef TCP_OFFLOAD_DISABLE 358 if (sc->sc_tu) 359 sc->sc_tu->tu_syncache_event(TOE_SC_DROP, sc->sc_toepcb); 360#endif 361 syncache_free(sc); 362 V_tcp_syncache.cache_count--; 363} 364 365/* 366 * Engage/reengage time on bucket row. 367 */ 368static void 369syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 370{ 371 sc->sc_rxttime = ticks + 372 TCPTV_RTOBASE * (tcp_backoff[sc->sc_rxmits]); 373 sc->sc_rxmits++; 374 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 375 sch->sch_nextc = sc->sc_rxttime; 376 if (docallout) 377 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 378 syncache_timer, (void *)sch); 379 } 380} 381 382/* 383 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 384 * If we have retransmitted an entry the maximum number of times, expire it. 385 * One separate timer for each bucket row. 386 */ 387static void 388syncache_timer(void *xsch) 389{ 390 struct syncache_head *sch = (struct syncache_head *)xsch; 391 struct syncache *sc, *nsc; 392 int tick = ticks; 393 char *s; 394 395 CURVNET_SET(sch->sch_vnet); 396 397 /* NB: syncache_head has already been locked by the callout. */ 398 SCH_LOCK_ASSERT(sch); 399 400 /* 401 * In the following cycle we may remove some entries and/or 402 * advance some timeouts, so re-initialize the bucket timer. 403 */ 404 sch->sch_nextc = tick + INT_MAX; 405 406 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 407 /* 408 * We do not check if the listen socket still exists 409 * and accept the case where the listen socket may be 410 * gone by the time we resend the SYN/ACK. We do 411 * not expect this to happens often. If it does, 412 * then the RST will be sent by the time the remote 413 * host does the SYN/ACK->ACK. 414 */ 415 if (TSTMP_GT(sc->sc_rxttime, tick)) { 416 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 417 sch->sch_nextc = sc->sc_rxttime; 418 continue; 419 } 420 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 421 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 422 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 423 "giving up and removing syncache entry\n", 424 s, __func__); 425 free(s, M_TCPLOG); 426 } 427 syncache_drop(sc, sch); 428 TCPSTAT_INC(tcps_sc_stale); 429 continue; 430 } 431 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 432 log(LOG_DEBUG, "%s; %s: Response timeout, " 433 "retransmitting (%u) SYN|ACK\n", 434 s, __func__, sc->sc_rxmits); 435 free(s, M_TCPLOG); 436 } 437 438 (void) syncache_respond(sc); 439 TCPSTAT_INC(tcps_sc_retransmitted); 440 syncache_timeout(sc, sch, 0); 441 } 442 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 443 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 444 syncache_timer, (void *)(sch)); 445 CURVNET_RESTORE(); 446} 447 448/* 449 * Find an entry in the syncache. 450 * Returns always with locked syncache_head plus a matching entry or NULL. 451 */ 452struct syncache * 453syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 454{ 455 struct syncache *sc; 456 struct syncache_head *sch; 457 458#ifdef INET6 459 if (inc->inc_flags & INC_ISIPV6) { 460 sch = &V_tcp_syncache.hashbase[ 461 SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)]; 462 *schp = sch; 463 464 SCH_LOCK(sch); 465 466 /* Circle through bucket row to find matching entry. */ 467 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 468 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 469 return (sc); 470 } 471 } else 472#endif 473 { 474 sch = &V_tcp_syncache.hashbase[ 475 SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)]; 476 *schp = sch; 477 478 SCH_LOCK(sch); 479 480 /* Circle through bucket row to find matching entry. */ 481 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 482#ifdef INET6 483 if (sc->sc_inc.inc_flags & INC_ISIPV6) 484 continue; 485#endif 486 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 487 return (sc); 488 } 489 } 490 SCH_LOCK_ASSERT(*schp); 491 return (NULL); /* always returns with locked sch */ 492} 493 494/* 495 * This function is called when we get a RST for a 496 * non-existent connection, so that we can see if the 497 * connection is in the syn cache. If it is, zap it. 498 */ 499void 500syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 501{ 502 struct syncache *sc; 503 struct syncache_head *sch; 504 char *s = NULL; 505 506 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 507 SCH_LOCK_ASSERT(sch); 508 509 /* 510 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 511 * See RFC 793 page 65, section SEGMENT ARRIVES. 512 */ 513 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 514 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 515 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 516 "FIN flag set, segment ignored\n", s, __func__); 517 TCPSTAT_INC(tcps_badrst); 518 goto done; 519 } 520 521 /* 522 * No corresponding connection was found in syncache. 523 * If syncookies are enabled and possibly exclusively 524 * used, or we are under memory pressure, a valid RST 525 * may not find a syncache entry. In that case we're 526 * done and no SYN|ACK retransmissions will happen. 527 * Otherwise the the RST was misdirected or spoofed. 528 */ 529 if (sc == NULL) { 530 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 531 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 532 "syncache entry (possibly syncookie only), " 533 "segment ignored\n", s, __func__); 534 TCPSTAT_INC(tcps_badrst); 535 goto done; 536 } 537 538 /* 539 * If the RST bit is set, check the sequence number to see 540 * if this is a valid reset segment. 541 * RFC 793 page 37: 542 * In all states except SYN-SENT, all reset (RST) segments 543 * are validated by checking their SEQ-fields. A reset is 544 * valid if its sequence number is in the window. 545 * 546 * The sequence number in the reset segment is normally an 547 * echo of our outgoing acknowlegement numbers, but some hosts 548 * send a reset with the sequence number at the rightmost edge 549 * of our receive window, and we have to handle this case. 550 */ 551 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 552 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 553 syncache_drop(sc, sch); 554 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 555 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 556 "connection attempt aborted by remote endpoint\n", 557 s, __func__); 558 TCPSTAT_INC(tcps_sc_reset); 559 } else { 560 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 561 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 562 "IRS %u (+WND %u), segment ignored\n", 563 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 564 TCPSTAT_INC(tcps_badrst); 565 } 566 567done: 568 if (s != NULL) 569 free(s, M_TCPLOG); 570 SCH_UNLOCK(sch); 571} 572 573void 574syncache_badack(struct in_conninfo *inc) 575{ 576 struct syncache *sc; 577 struct syncache_head *sch; 578 579 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 580 SCH_LOCK_ASSERT(sch); 581 if (sc != NULL) { 582 syncache_drop(sc, sch); 583 TCPSTAT_INC(tcps_sc_badack); 584 } 585 SCH_UNLOCK(sch); 586} 587 588void 589syncache_unreach(struct in_conninfo *inc, struct tcphdr *th) 590{ 591 struct syncache *sc; 592 struct syncache_head *sch; 593 594 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 595 SCH_LOCK_ASSERT(sch); 596 if (sc == NULL) 597 goto done; 598 599 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 600 if (ntohl(th->th_seq) != sc->sc_iss) 601 goto done; 602 603 /* 604 * If we've rertransmitted 3 times and this is our second error, 605 * we remove the entry. Otherwise, we allow it to continue on. 606 * This prevents us from incorrectly nuking an entry during a 607 * spurious network outage. 608 * 609 * See tcp_notify(). 610 */ 611 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 612 sc->sc_flags |= SCF_UNREACH; 613 goto done; 614 } 615 syncache_drop(sc, sch); 616 TCPSTAT_INC(tcps_sc_unreach); 617done: 618 SCH_UNLOCK(sch); 619} 620 621/* 622 * Build a new TCP socket structure from a syncache entry. 623 */ 624static struct socket * 625syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 626{ 627 struct inpcb *inp = NULL; 628 struct socket *so; 629 struct tcpcb *tp; 630 int error; 631 char *s; 632 633 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 634 635 /* 636 * Ok, create the full blown connection, and set things up 637 * as they would have been set up if we had created the 638 * connection when the SYN arrived. If we can't create 639 * the connection, abort it. 640 */ 641 so = sonewconn(lso, SS_ISCONNECTED); 642 if (so == NULL) { 643 /* 644 * Drop the connection; we will either send a RST or 645 * have the peer retransmit its SYN again after its 646 * RTO and try again. 647 */ 648 TCPSTAT_INC(tcps_listendrop); 649 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 650 log(LOG_DEBUG, "%s; %s: Socket create failed " 651 "due to limits or memory shortage\n", 652 s, __func__); 653 free(s, M_TCPLOG); 654 } 655 goto abort2; 656 } 657#ifdef MAC 658 mac_socketpeer_set_from_mbuf(m, so); 659#endif 660 661 inp = sotoinpcb(so); 662 inp->inp_inc.inc_fibnum = so->so_fibnum; 663 INP_WLOCK(inp); 664 665 /* Insert new socket into PCB hash list. */ 666 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 667#ifdef INET6 668 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 669 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 670 } else { 671 inp->inp_vflag &= ~INP_IPV6; 672 inp->inp_vflag |= INP_IPV4; 673#endif 674 inp->inp_laddr = sc->sc_inc.inc_laddr; 675#ifdef INET6 676 } 677#endif 678 inp->inp_lport = sc->sc_inc.inc_lport; 679 if ((error = in_pcbinshash(inp)) != 0) { 680 /* 681 * Undo the assignments above if we failed to 682 * put the PCB on the hash lists. 683 */ 684#ifdef INET6 685 if (sc->sc_inc.inc_flags & INC_ISIPV6) 686 inp->in6p_laddr = in6addr_any; 687 else 688#endif 689 inp->inp_laddr.s_addr = INADDR_ANY; 690 inp->inp_lport = 0; 691 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 692 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 693 "with error %i\n", 694 s, __func__, error); 695 free(s, M_TCPLOG); 696 } 697 goto abort; 698 } 699#ifdef IPSEC 700 /* Copy old policy into new socket's. */ 701 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp)) 702 printf("syncache_socket: could not copy policy\n"); 703#endif 704#ifdef INET6 705 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 706 struct inpcb *oinp = sotoinpcb(lso); 707 struct in6_addr laddr6; 708 struct sockaddr_in6 sin6; 709 /* 710 * Inherit socket options from the listening socket. 711 * Note that in6p_inputopts are not (and should not be) 712 * copied, since it stores previously received options and is 713 * used to detect if each new option is different than the 714 * previous one and hence should be passed to a user. 715 * If we copied in6p_inputopts, a user would not be able to 716 * receive options just after calling the accept system call. 717 */ 718 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 719 if (oinp->in6p_outputopts) 720 inp->in6p_outputopts = 721 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 722 723 sin6.sin6_family = AF_INET6; 724 sin6.sin6_len = sizeof(sin6); 725 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 726 sin6.sin6_port = sc->sc_inc.inc_fport; 727 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 728 laddr6 = inp->in6p_laddr; 729 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 730 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 731 if ((error = in6_pcbconnect(inp, (struct sockaddr *)&sin6, 732 thread0.td_ucred)) != 0) { 733 inp->in6p_laddr = laddr6; 734 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 735 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 736 "with error %i\n", 737 s, __func__, error); 738 free(s, M_TCPLOG); 739 } 740 goto abort; 741 } 742 /* Override flowlabel from in6_pcbconnect. */ 743 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 744 inp->inp_flow |= sc->sc_flowlabel; 745 } else 746#endif 747 { 748 struct in_addr laddr; 749 struct sockaddr_in sin; 750 751 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 752 753 if (inp->inp_options == NULL) { 754 inp->inp_options = sc->sc_ipopts; 755 sc->sc_ipopts = NULL; 756 } 757 758 sin.sin_family = AF_INET; 759 sin.sin_len = sizeof(sin); 760 sin.sin_addr = sc->sc_inc.inc_faddr; 761 sin.sin_port = sc->sc_inc.inc_fport; 762 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 763 laddr = inp->inp_laddr; 764 if (inp->inp_laddr.s_addr == INADDR_ANY) 765 inp->inp_laddr = sc->sc_inc.inc_laddr; 766 if ((error = in_pcbconnect(inp, (struct sockaddr *)&sin, 767 thread0.td_ucred)) != 0) { 768 inp->inp_laddr = laddr; 769 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 770 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 771 "with error %i\n", 772 s, __func__, error); 773 free(s, M_TCPLOG); 774 } 775 goto abort; 776 } 777 } 778 tp = intotcpcb(inp); 779 tp->t_state = TCPS_SYN_RECEIVED; 780 tp->iss = sc->sc_iss; 781 tp->irs = sc->sc_irs; 782 tcp_rcvseqinit(tp); 783 tcp_sendseqinit(tp); 784 tp->snd_wl1 = sc->sc_irs; 785 tp->snd_max = tp->iss + 1; 786 tp->snd_nxt = tp->iss + 1; 787 tp->rcv_up = sc->sc_irs + 1; 788 tp->rcv_wnd = sc->sc_wnd; 789 tp->rcv_adv += tp->rcv_wnd; 790 tp->last_ack_sent = tp->rcv_nxt; 791 792 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 793 if (sc->sc_flags & SCF_NOOPT) 794 tp->t_flags |= TF_NOOPT; 795 else { 796 if (sc->sc_flags & SCF_WINSCALE) { 797 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 798 tp->snd_scale = sc->sc_requested_s_scale; 799 tp->request_r_scale = sc->sc_requested_r_scale; 800 } 801 if (sc->sc_flags & SCF_TIMESTAMP) { 802 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 803 tp->ts_recent = sc->sc_tsreflect; 804 tp->ts_recent_age = ticks; 805 tp->ts_offset = sc->sc_tsoff; 806 } 807#ifdef TCP_SIGNATURE 808 if (sc->sc_flags & SCF_SIGNATURE) 809 tp->t_flags |= TF_SIGNATURE; 810#endif 811 if (sc->sc_flags & SCF_SACK) 812 tp->t_flags |= TF_SACK_PERMIT; 813 } 814 815 if (sc->sc_flags & SCF_ECN) 816 tp->t_flags |= TF_ECN_PERMIT; 817 818 /* 819 * Set up MSS and get cached values from tcp_hostcache. 820 * This might overwrite some of the defaults we just set. 821 */ 822 tcp_mss(tp, sc->sc_peer_mss); 823 824 /* 825 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment. 826 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 827 */ 828 if (sc->sc_rxmits > 1) 829 tp->snd_cwnd = tp->t_maxseg; 830 tcp_timer_activate(tp, TT_KEEP, tcp_keepinit); 831 832 INP_WUNLOCK(inp); 833 834 TCPSTAT_INC(tcps_accepts); 835 return (so); 836 837abort: 838 INP_WUNLOCK(inp); 839abort2: 840 if (so != NULL) 841 soabort(so); 842 return (NULL); 843} 844 845/* 846 * This function gets called when we receive an ACK for a 847 * socket in the LISTEN state. We look up the connection 848 * in the syncache, and if its there, we pull it out of 849 * the cache and turn it into a full-blown connection in 850 * the SYN-RECEIVED state. 851 */ 852int 853syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 854 struct socket **lsop, struct mbuf *m) 855{ 856 struct syncache *sc; 857 struct syncache_head *sch; 858 struct syncache scs; 859 char *s; 860 861 /* 862 * Global TCP locks are held because we manipulate the PCB lists 863 * and create a new socket. 864 */ 865 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 866 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 867 ("%s: can handle only ACK", __func__)); 868 869 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 870 SCH_LOCK_ASSERT(sch); 871 if (sc == NULL) { 872 /* 873 * There is no syncache entry, so see if this ACK is 874 * a returning syncookie. To do this, first: 875 * A. See if this socket has had a syncache entry dropped in 876 * the past. We don't want to accept a bogus syncookie 877 * if we've never received a SYN. 878 * B. check that the syncookie is valid. If it is, then 879 * cobble up a fake syncache entry, and return. 880 */ 881 if (!V_tcp_syncookies) { 882 SCH_UNLOCK(sch); 883 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 884 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 885 "segment rejected (syncookies disabled)\n", 886 s, __func__); 887 goto failed; 888 } 889 bzero(&scs, sizeof(scs)); 890 sc = syncookie_lookup(inc, sch, &scs, to, th, *lsop); 891 SCH_UNLOCK(sch); 892 if (sc == NULL) { 893 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 894 log(LOG_DEBUG, "%s; %s: Segment failed " 895 "SYNCOOKIE authentication, segment rejected " 896 "(probably spoofed)\n", s, __func__); 897 goto failed; 898 } 899 } else { 900 /* Pull out the entry to unlock the bucket row. */ 901 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 902 sch->sch_length--; 903 V_tcp_syncache.cache_count--; 904 SCH_UNLOCK(sch); 905 } 906 907 /* 908 * Segment validation: 909 * ACK must match our initial sequence number + 1 (the SYN|ACK). 910 */ 911 if (th->th_ack != sc->sc_iss + 1 && !TOEPCB_ISSET(sc)) { 912 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 913 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 914 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 915 goto failed; 916 } 917 918 /* 919 * The SEQ must fall in the window starting at the received 920 * initial receive sequence number + 1 (the SYN). 921 */ 922 if ((SEQ_LEQ(th->th_seq, sc->sc_irs) || 923 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) && 924 !TOEPCB_ISSET(sc)) { 925 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 926 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 927 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 928 goto failed; 929 } 930 931 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 932 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 933 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 934 "segment rejected\n", s, __func__); 935 goto failed; 936 } 937 /* 938 * If timestamps were negotiated the reflected timestamp 939 * must be equal to what we actually sent in the SYN|ACK. 940 */ 941 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts && 942 !TOEPCB_ISSET(sc)) { 943 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 944 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 945 "segment rejected\n", 946 s, __func__, to->to_tsecr, sc->sc_ts); 947 goto failed; 948 } 949 950 *lsop = syncache_socket(sc, *lsop, m); 951 952 if (*lsop == NULL) 953 TCPSTAT_INC(tcps_sc_aborted); 954 else 955 TCPSTAT_INC(tcps_sc_completed); 956 957/* how do we find the inp for the new socket? */ 958 if (sc != &scs) 959 syncache_free(sc); 960 return (1); 961failed: 962 if (sc != NULL && sc != &scs) 963 syncache_free(sc); 964 if (s != NULL) 965 free(s, M_TCPLOG); 966 *lsop = NULL; 967 return (0); 968} 969 970int 971tcp_offload_syncache_expand(struct in_conninfo *inc, struct toeopt *toeo, 972 struct tcphdr *th, struct socket **lsop, struct mbuf *m) 973{ 974 struct tcpopt to; 975 int rc; 976 977 bzero(&to, sizeof(struct tcpopt)); 978 to.to_mss = toeo->to_mss; 979 to.to_wscale = toeo->to_wscale; 980 to.to_flags = toeo->to_flags; 981 982 INP_INFO_WLOCK(&V_tcbinfo); 983 rc = syncache_expand(inc, &to, th, lsop, m); 984 INP_INFO_WUNLOCK(&V_tcbinfo); 985 986 return (rc); 987} 988 989/* 990 * Given a LISTEN socket and an inbound SYN request, add 991 * this to the syn cache, and send back a segment: 992 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 993 * to the source. 994 * 995 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 996 * Doing so would require that we hold onto the data and deliver it 997 * to the application. However, if we are the target of a SYN-flood 998 * DoS attack, an attacker could send data which would eventually 999 * consume all available buffer space if it were ACKed. By not ACKing 1000 * the data, we avoid this DoS scenario. 1001 */ 1002static void 1003_syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1004 struct inpcb *inp, struct socket **lsop, struct mbuf *m, 1005 struct toe_usrreqs *tu, void *toepcb) 1006{ 1007 struct tcpcb *tp; 1008 struct socket *so; 1009 struct syncache *sc = NULL; 1010 struct syncache_head *sch; 1011 struct mbuf *ipopts = NULL; 1012 u_int32_t flowtmp; 1013 int win, sb_hiwat, ip_ttl, ip_tos, noopt; 1014 char *s; 1015#ifdef INET6 1016 int autoflowlabel = 0; 1017#endif 1018#ifdef MAC 1019 struct label *maclabel; 1020#endif 1021 struct syncache scs; 1022 struct ucred *cred; 1023 1024 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 1025 INP_WLOCK_ASSERT(inp); /* listen socket */ 1026 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1027 ("%s: unexpected tcp flags", __func__)); 1028 1029 /* 1030 * Combine all so/tp operations very early to drop the INP lock as 1031 * soon as possible. 1032 */ 1033 so = *lsop; 1034 tp = sototcpcb(so); 1035 cred = crhold(so->so_cred); 1036 1037#ifdef INET6 1038 if ((inc->inc_flags & INC_ISIPV6) && 1039 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1040 autoflowlabel = 1; 1041#endif 1042 ip_ttl = inp->inp_ip_ttl; 1043 ip_tos = inp->inp_ip_tos; 1044 win = sbspace(&so->so_rcv); 1045 sb_hiwat = so->so_rcv.sb_hiwat; 1046 noopt = (tp->t_flags & TF_NOOPT); 1047 1048 /* By the time we drop the lock these should no longer be used. */ 1049 so = NULL; 1050 tp = NULL; 1051 1052#ifdef MAC 1053 if (mac_syncache_init(&maclabel) != 0) { 1054 INP_WUNLOCK(inp); 1055 INP_INFO_WUNLOCK(&V_tcbinfo); 1056 goto done; 1057 } else 1058 mac_syncache_create(maclabel, inp); 1059#endif 1060 INP_WUNLOCK(inp); 1061 INP_INFO_WUNLOCK(&V_tcbinfo); 1062 1063 /* 1064 * Remember the IP options, if any. 1065 */ 1066#ifdef INET6 1067 if (!(inc->inc_flags & INC_ISIPV6)) 1068#endif 1069 ipopts = (m) ? ip_srcroute(m) : NULL; 1070 1071 /* 1072 * See if we already have an entry for this connection. 1073 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1074 * 1075 * XXX: should the syncache be re-initialized with the contents 1076 * of the new SYN here (which may have different options?) 1077 * 1078 * XXX: We do not check the sequence number to see if this is a 1079 * real retransmit or a new connection attempt. The question is 1080 * how to handle such a case; either ignore it as spoofed, or 1081 * drop the current entry and create a new one? 1082 */ 1083 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1084 SCH_LOCK_ASSERT(sch); 1085 if (sc != NULL) { 1086#ifndef TCP_OFFLOAD_DISABLE 1087 if (sc->sc_tu) 1088 sc->sc_tu->tu_syncache_event(TOE_SC_ENTRY_PRESENT, 1089 sc->sc_toepcb); 1090#endif 1091 TCPSTAT_INC(tcps_sc_dupsyn); 1092 if (ipopts) { 1093 /* 1094 * If we were remembering a previous source route, 1095 * forget it and use the new one we've been given. 1096 */ 1097 if (sc->sc_ipopts) 1098 (void) m_free(sc->sc_ipopts); 1099 sc->sc_ipopts = ipopts; 1100 } 1101 /* 1102 * Update timestamp if present. 1103 */ 1104 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1105 sc->sc_tsreflect = to->to_tsval; 1106 else 1107 sc->sc_flags &= ~SCF_TIMESTAMP; 1108#ifdef MAC 1109 /* 1110 * Since we have already unconditionally allocated label 1111 * storage, free it up. The syncache entry will already 1112 * have an initialized label we can use. 1113 */ 1114 mac_syncache_destroy(&maclabel); 1115#endif 1116 /* Retransmit SYN|ACK and reset retransmit count. */ 1117 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1118 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1119 "resetting timer and retransmitting SYN|ACK\n", 1120 s, __func__); 1121 free(s, M_TCPLOG); 1122 } 1123 if (!TOEPCB_ISSET(sc) && syncache_respond(sc) == 0) { 1124 sc->sc_rxmits = 0; 1125 syncache_timeout(sc, sch, 1); 1126 TCPSTAT_INC(tcps_sndacks); 1127 TCPSTAT_INC(tcps_sndtotal); 1128 } 1129 SCH_UNLOCK(sch); 1130 goto done; 1131 } 1132 1133 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1134 if (sc == NULL) { 1135 /* 1136 * The zone allocator couldn't provide more entries. 1137 * Treat this as if the cache was full; drop the oldest 1138 * entry and insert the new one. 1139 */ 1140 TCPSTAT_INC(tcps_sc_zonefail); 1141 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) 1142 syncache_drop(sc, sch); 1143 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1144 if (sc == NULL) { 1145 if (V_tcp_syncookies) { 1146 bzero(&scs, sizeof(scs)); 1147 sc = &scs; 1148 } else { 1149 SCH_UNLOCK(sch); 1150 if (ipopts) 1151 (void) m_free(ipopts); 1152 goto done; 1153 } 1154 } 1155 } 1156 1157 /* 1158 * Fill in the syncache values. 1159 */ 1160#ifdef MAC 1161 sc->sc_label = maclabel; 1162#endif 1163 sc->sc_cred = cred; 1164 cred = NULL; 1165 sc->sc_ipopts = ipopts; 1166 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1167#ifdef INET6 1168 if (!(inc->inc_flags & INC_ISIPV6)) 1169#endif 1170 { 1171 sc->sc_ip_tos = ip_tos; 1172 sc->sc_ip_ttl = ip_ttl; 1173 } 1174#ifndef TCP_OFFLOAD_DISABLE 1175 sc->sc_tu = tu; 1176 sc->sc_toepcb = toepcb; 1177#endif 1178 sc->sc_irs = th->th_seq; 1179 sc->sc_iss = arc4random(); 1180 sc->sc_flags = 0; 1181 sc->sc_flowlabel = 0; 1182 1183 /* 1184 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1185 * win was derived from socket earlier in the function. 1186 */ 1187 win = imax(win, 0); 1188 win = imin(win, TCP_MAXWIN); 1189 sc->sc_wnd = win; 1190 1191 if (V_tcp_do_rfc1323) { 1192 /* 1193 * A timestamp received in a SYN makes 1194 * it ok to send timestamp requests and replies. 1195 */ 1196 if (to->to_flags & TOF_TS) { 1197 sc->sc_tsreflect = to->to_tsval; 1198 sc->sc_ts = ticks; 1199 sc->sc_flags |= SCF_TIMESTAMP; 1200 } 1201 if (to->to_flags & TOF_SCALE) { 1202 int wscale = 0; 1203 1204 /* 1205 * Pick the smallest possible scaling factor that 1206 * will still allow us to scale up to sb_max, aka 1207 * kern.ipc.maxsockbuf. 1208 * 1209 * We do this because there are broken firewalls that 1210 * will corrupt the window scale option, leading to 1211 * the other endpoint believing that our advertised 1212 * window is unscaled. At scale factors larger than 1213 * 5 the unscaled window will drop below 1500 bytes, 1214 * leading to serious problems when traversing these 1215 * broken firewalls. 1216 * 1217 * With the default maxsockbuf of 256K, a scale factor 1218 * of 3 will be chosen by this algorithm. Those who 1219 * choose a larger maxsockbuf should watch out 1220 * for the compatiblity problems mentioned above. 1221 * 1222 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1223 * or <SYN,ACK>) segment itself is never scaled. 1224 */ 1225 while (wscale < TCP_MAX_WINSHIFT && 1226 (TCP_MAXWIN << wscale) < sb_max) 1227 wscale++; 1228 sc->sc_requested_r_scale = wscale; 1229 sc->sc_requested_s_scale = to->to_wscale; 1230 sc->sc_flags |= SCF_WINSCALE; 1231 } 1232 } 1233#ifdef TCP_SIGNATURE 1234 /* 1235 * If listening socket requested TCP digests, and received SYN 1236 * contains the option, flag this in the syncache so that 1237 * syncache_respond() will do the right thing with the SYN+ACK. 1238 * XXX: Currently we always record the option by default and will 1239 * attempt to use it in syncache_respond(). 1240 */ 1241 if (to->to_flags & TOF_SIGNATURE) 1242 sc->sc_flags |= SCF_SIGNATURE; 1243#endif 1244 if (to->to_flags & TOF_SACKPERM) 1245 sc->sc_flags |= SCF_SACK; 1246 if (to->to_flags & TOF_MSS) 1247 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1248 if (noopt) 1249 sc->sc_flags |= SCF_NOOPT; 1250 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1251 sc->sc_flags |= SCF_ECN; 1252 1253 if (V_tcp_syncookies) { 1254 syncookie_generate(sch, sc, &flowtmp); 1255#ifdef INET6 1256 if (autoflowlabel) 1257 sc->sc_flowlabel = flowtmp; 1258#endif 1259 } else { 1260#ifdef INET6 1261 if (autoflowlabel) 1262 sc->sc_flowlabel = 1263 (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK); 1264#endif 1265 } 1266 SCH_UNLOCK(sch); 1267 1268 /* 1269 * Do a standard 3-way handshake. 1270 */ 1271 if (TOEPCB_ISSET(sc) || syncache_respond(sc) == 0) { 1272 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1273 syncache_free(sc); 1274 else if (sc != &scs) 1275 syncache_insert(sc, sch); /* locks and unlocks sch */ 1276 TCPSTAT_INC(tcps_sndacks); 1277 TCPSTAT_INC(tcps_sndtotal); 1278 } else { 1279 if (sc != &scs) 1280 syncache_free(sc); 1281 TCPSTAT_INC(tcps_sc_dropped); 1282 } 1283 1284done: 1285 if (cred != NULL) 1286 crfree(cred); 1287#ifdef MAC 1288 if (sc == &scs) 1289 mac_syncache_destroy(&maclabel); 1290#endif 1291 if (m) { 1292 1293 *lsop = NULL; 1294 m_freem(m); 1295 } 1296} 1297 1298static int 1299syncache_respond(struct syncache *sc) 1300{ 1301 struct ip *ip = NULL; 1302 struct mbuf *m; 1303 struct tcphdr *th; 1304 int optlen, error; 1305 u_int16_t hlen, tlen, mssopt; 1306 struct tcpopt to; 1307#ifdef INET6 1308 struct ip6_hdr *ip6 = NULL; 1309#endif 1310 1311 hlen = 1312#ifdef INET6 1313 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1314#endif 1315 sizeof(struct ip); 1316 tlen = hlen + sizeof(struct tcphdr); 1317 1318 /* Determine MSS we advertize to other end of connection. */ 1319 mssopt = tcp_mssopt(&sc->sc_inc); 1320 if (sc->sc_peer_mss) 1321 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss); 1322 1323 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1324 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1325 ("syncache: mbuf too small")); 1326 1327 /* Create the IP+TCP header from scratch. */ 1328 m = m_gethdr(M_DONTWAIT, MT_DATA); 1329 if (m == NULL) 1330 return (ENOBUFS); 1331#ifdef MAC 1332 mac_syncache_create_mbuf(sc->sc_label, m); 1333#endif 1334 m->m_data += max_linkhdr; 1335 m->m_len = tlen; 1336 m->m_pkthdr.len = tlen; 1337 m->m_pkthdr.rcvif = NULL; 1338 1339#ifdef INET6 1340 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1341 ip6 = mtod(m, struct ip6_hdr *); 1342 ip6->ip6_vfc = IPV6_VERSION; 1343 ip6->ip6_nxt = IPPROTO_TCP; 1344 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1345 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1346 ip6->ip6_plen = htons(tlen - hlen); 1347 /* ip6_hlim is set after checksum */ 1348 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1349 ip6->ip6_flow |= sc->sc_flowlabel; 1350 1351 th = (struct tcphdr *)(ip6 + 1); 1352 } else 1353#endif 1354 { 1355 ip = mtod(m, struct ip *); 1356 ip->ip_v = IPVERSION; 1357 ip->ip_hl = sizeof(struct ip) >> 2; 1358 ip->ip_len = tlen; 1359 ip->ip_id = 0; 1360 ip->ip_off = 0; 1361 ip->ip_sum = 0; 1362 ip->ip_p = IPPROTO_TCP; 1363 ip->ip_src = sc->sc_inc.inc_laddr; 1364 ip->ip_dst = sc->sc_inc.inc_faddr; 1365 ip->ip_ttl = sc->sc_ip_ttl; 1366 ip->ip_tos = sc->sc_ip_tos; 1367 1368 /* 1369 * See if we should do MTU discovery. Route lookups are 1370 * expensive, so we will only unset the DF bit if: 1371 * 1372 * 1) path_mtu_discovery is disabled 1373 * 2) the SCF_UNREACH flag has been set 1374 */ 1375 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1376 ip->ip_off |= IP_DF; 1377 1378 th = (struct tcphdr *)(ip + 1); 1379 } 1380 th->th_sport = sc->sc_inc.inc_lport; 1381 th->th_dport = sc->sc_inc.inc_fport; 1382 1383 th->th_seq = htonl(sc->sc_iss); 1384 th->th_ack = htonl(sc->sc_irs + 1); 1385 th->th_off = sizeof(struct tcphdr) >> 2; 1386 th->th_x2 = 0; 1387 th->th_flags = TH_SYN|TH_ACK; 1388 th->th_win = htons(sc->sc_wnd); 1389 th->th_urp = 0; 1390 1391 if (sc->sc_flags & SCF_ECN) { 1392 th->th_flags |= TH_ECE; 1393 TCPSTAT_INC(tcps_ecn_shs); 1394 } 1395 1396 /* Tack on the TCP options. */ 1397 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1398 to.to_flags = 0; 1399 1400 to.to_mss = mssopt; 1401 to.to_flags = TOF_MSS; 1402 if (sc->sc_flags & SCF_WINSCALE) { 1403 to.to_wscale = sc->sc_requested_r_scale; 1404 to.to_flags |= TOF_SCALE; 1405 } 1406 if (sc->sc_flags & SCF_TIMESTAMP) { 1407 /* Virgin timestamp or TCP cookie enhanced one. */ 1408 to.to_tsval = sc->sc_ts; 1409 to.to_tsecr = sc->sc_tsreflect; 1410 to.to_flags |= TOF_TS; 1411 } 1412 if (sc->sc_flags & SCF_SACK) 1413 to.to_flags |= TOF_SACKPERM; 1414#ifdef TCP_SIGNATURE 1415 if (sc->sc_flags & SCF_SIGNATURE) 1416 to.to_flags |= TOF_SIGNATURE; 1417#endif 1418 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1419 1420 /* Adjust headers by option size. */ 1421 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1422 m->m_len += optlen; 1423 m->m_pkthdr.len += optlen; 1424 1425#ifdef TCP_SIGNATURE 1426 if (sc->sc_flags & SCF_SIGNATURE) 1427 tcp_signature_compute(m, 0, 0, optlen, 1428 to.to_signature, IPSEC_DIR_OUTBOUND); 1429#endif 1430#ifdef INET6 1431 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1432 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1433 else 1434#endif 1435 ip->ip_len += optlen; 1436 } else 1437 optlen = 0; 1438 1439 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1440#ifdef INET6 1441 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1442 th->th_sum = 0; 1443 th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, 1444 tlen + optlen - hlen); 1445 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1446 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1447 } else 1448#endif 1449 { 1450 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1451 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1452 m->m_pkthdr.csum_flags = CSUM_TCP; 1453 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1454 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1455 } 1456 return (error); 1457} 1458 1459void 1460syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1461 struct inpcb *inp, struct socket **lsop, struct mbuf *m) 1462{ 1463 _syncache_add(inc, to, th, inp, lsop, m, NULL, NULL); 1464} 1465 1466void 1467tcp_offload_syncache_add(struct in_conninfo *inc, struct toeopt *toeo, 1468 struct tcphdr *th, struct inpcb *inp, struct socket **lsop, 1469 struct toe_usrreqs *tu, void *toepcb) 1470{ 1471 struct tcpopt to; 1472 1473 bzero(&to, sizeof(struct tcpopt)); 1474 to.to_mss = toeo->to_mss; 1475 to.to_wscale = toeo->to_wscale; 1476 to.to_flags = toeo->to_flags; 1477 1478 INP_INFO_WLOCK(&V_tcbinfo); 1479 INP_WLOCK(inp); 1480 1481 _syncache_add(inc, &to, th, inp, lsop, NULL, tu, toepcb); 1482} 1483 1484/* 1485 * The purpose of SYN cookies is to avoid keeping track of all SYN's we 1486 * receive and to be able to handle SYN floods from bogus source addresses 1487 * (where we will never receive any reply). SYN floods try to exhaust all 1488 * our memory and available slots in the SYN cache table to cause a denial 1489 * of service to legitimate users of the local host. 1490 * 1491 * The idea of SYN cookies is to encode and include all necessary information 1492 * about the connection setup state within the SYN-ACK we send back and thus 1493 * to get along without keeping any local state until the ACK to the SYN-ACK 1494 * arrives (if ever). Everything we need to know should be available from 1495 * the information we encoded in the SYN-ACK. 1496 * 1497 * More information about the theory behind SYN cookies and its first 1498 * discussion and specification can be found at: 1499 * http://cr.yp.to/syncookies.html (overview) 1500 * http://cr.yp.to/syncookies/archive (gory details) 1501 * 1502 * This implementation extends the orginal idea and first implementation 1503 * of FreeBSD by using not only the initial sequence number field to store 1504 * information but also the timestamp field if present. This way we can 1505 * keep track of the entire state we need to know to recreate the session in 1506 * its original form. Almost all TCP speakers implement RFC1323 timestamps 1507 * these days. For those that do not we still have to live with the known 1508 * shortcomings of the ISN only SYN cookies. 1509 * 1510 * Cookie layers: 1511 * 1512 * Initial sequence number we send: 1513 * 31|................................|0 1514 * DDDDDDDDDDDDDDDDDDDDDDDDDMMMRRRP 1515 * D = MD5 Digest (first dword) 1516 * M = MSS index 1517 * R = Rotation of secret 1518 * P = Odd or Even secret 1519 * 1520 * The MD5 Digest is computed with over following parameters: 1521 * a) randomly rotated secret 1522 * b) struct in_conninfo containing the remote/local ip/port (IPv4&IPv6) 1523 * c) the received initial sequence number from remote host 1524 * d) the rotation offset and odd/even bit 1525 * 1526 * Timestamp we send: 1527 * 31|................................|0 1528 * DDDDDDDDDDDDDDDDDDDDDDSSSSRRRRA5 1529 * D = MD5 Digest (third dword) (only as filler) 1530 * S = Requested send window scale 1531 * R = Requested receive window scale 1532 * A = SACK allowed 1533 * 5 = TCP-MD5 enabled (not implemented yet) 1534 * XORed with MD5 Digest (forth dword) 1535 * 1536 * The timestamp isn't cryptographically secure and doesn't need to be. 1537 * The double use of the MD5 digest dwords ties it to a specific remote/ 1538 * local host/port, remote initial sequence number and our local time 1539 * limited secret. A received timestamp is reverted (XORed) and then 1540 * the contained MD5 dword is compared to the computed one to ensure the 1541 * timestamp belongs to the SYN-ACK we sent. The other parameters may 1542 * have been tampered with but this isn't different from supplying bogus 1543 * values in the SYN in the first place. 1544 * 1545 * Some problems with SYN cookies remain however: 1546 * Consider the problem of a recreated (and retransmitted) cookie. If the 1547 * original SYN was accepted, the connection is established. The second 1548 * SYN is inflight, and if it arrives with an ISN that falls within the 1549 * receive window, the connection is killed. 1550 * 1551 * Notes: 1552 * A heuristic to determine when to accept syn cookies is not necessary. 1553 * An ACK flood would cause the syncookie verification to be attempted, 1554 * but a SYN flood causes syncookies to be generated. Both are of equal 1555 * cost, so there's no point in trying to optimize the ACK flood case. 1556 * Also, if you don't process certain ACKs for some reason, then all someone 1557 * would have to do is launch a SYN and ACK flood at the same time, which 1558 * would stop cookie verification and defeat the entire purpose of syncookies. 1559 */ 1560static int tcp_sc_msstab[] = { 0, 256, 468, 536, 996, 1452, 1460, 8960 }; 1561 1562static void 1563syncookie_generate(struct syncache_head *sch, struct syncache *sc, 1564 u_int32_t *flowlabel) 1565{ 1566 MD5_CTX ctx; 1567 u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)]; 1568 u_int32_t data; 1569 u_int32_t *secbits; 1570 u_int off, pmss, mss; 1571 int i; 1572 1573 SCH_LOCK_ASSERT(sch); 1574 1575 /* Which of the two secrets to use. */ 1576 secbits = sch->sch_oddeven ? 1577 sch->sch_secbits_odd : sch->sch_secbits_even; 1578 1579 /* Reseed secret if too old. */ 1580 if (sch->sch_reseed < time_uptime) { 1581 sch->sch_oddeven = sch->sch_oddeven ? 0 : 1; /* toggle */ 1582 secbits = sch->sch_oddeven ? 1583 sch->sch_secbits_odd : sch->sch_secbits_even; 1584 for (i = 0; i < SYNCOOKIE_SECRET_SIZE; i++) 1585 secbits[i] = arc4random(); 1586 sch->sch_reseed = time_uptime + SYNCOOKIE_LIFETIME; 1587 } 1588 1589 /* Secret rotation offset. */ 1590 off = sc->sc_iss & 0x7; /* iss was randomized before */ 1591 1592 /* Maximum segment size calculation. */ 1593 pmss = 1594 max( min(sc->sc_peer_mss, tcp_mssopt(&sc->sc_inc)), V_tcp_minmss); 1595 for (mss = sizeof(tcp_sc_msstab) / sizeof(int) - 1; mss > 0; mss--) 1596 if (tcp_sc_msstab[mss] <= pmss) 1597 break; 1598 1599 /* Fold parameters and MD5 digest into the ISN we will send. */ 1600 data = sch->sch_oddeven;/* odd or even secret, 1 bit */ 1601 data |= off << 1; /* secret offset, derived from iss, 3 bits */ 1602 data |= mss << 4; /* mss, 3 bits */ 1603 1604 MD5Init(&ctx); 1605 MD5Update(&ctx, ((u_int8_t *)secbits) + off, 1606 SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off); 1607 MD5Update(&ctx, secbits, off); 1608 MD5Update(&ctx, &sc->sc_inc, sizeof(sc->sc_inc)); 1609 MD5Update(&ctx, &sc->sc_irs, sizeof(sc->sc_irs)); 1610 MD5Update(&ctx, &data, sizeof(data)); 1611 MD5Final((u_int8_t *)&md5_buffer, &ctx); 1612 1613 data |= (md5_buffer[0] << 7); 1614 sc->sc_iss = data; 1615 1616#ifdef INET6 1617 *flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK; 1618#endif 1619 1620 /* Additional parameters are stored in the timestamp if present. */ 1621 if (sc->sc_flags & SCF_TIMESTAMP) { 1622 data = ((sc->sc_flags & SCF_SIGNATURE) ? 1 : 0); /* TCP-MD5, 1 bit */ 1623 data |= ((sc->sc_flags & SCF_SACK) ? 1 : 0) << 1; /* SACK, 1 bit */ 1624 data |= sc->sc_requested_s_scale << 2; /* SWIN scale, 4 bits */ 1625 data |= sc->sc_requested_r_scale << 6; /* RWIN scale, 4 bits */ 1626 data |= md5_buffer[2] << 10; /* more digest bits */ 1627 data ^= md5_buffer[3]; 1628 sc->sc_ts = data; 1629 sc->sc_tsoff = data - ticks; /* after XOR */ 1630 } 1631 1632 TCPSTAT_INC(tcps_sc_sendcookie); 1633} 1634 1635static struct syncache * 1636syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 1637 struct syncache *sc, struct tcpopt *to, struct tcphdr *th, 1638 struct socket *so) 1639{ 1640 MD5_CTX ctx; 1641 u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)]; 1642 u_int32_t data = 0; 1643 u_int32_t *secbits; 1644 tcp_seq ack, seq; 1645 int off, mss, wnd, flags; 1646 1647 SCH_LOCK_ASSERT(sch); 1648 1649 /* 1650 * Pull information out of SYN-ACK/ACK and 1651 * revert sequence number advances. 1652 */ 1653 ack = th->th_ack - 1; 1654 seq = th->th_seq - 1; 1655 off = (ack >> 1) & 0x7; 1656 mss = (ack >> 4) & 0x7; 1657 flags = ack & 0x7f; 1658 1659 /* Which of the two secrets to use. */ 1660 secbits = (flags & 0x1) ? sch->sch_secbits_odd : sch->sch_secbits_even; 1661 1662 /* 1663 * The secret wasn't updated for the lifetime of a syncookie, 1664 * so this SYN-ACK/ACK is either too old (replay) or totally bogus. 1665 */ 1666 if (sch->sch_reseed + SYNCOOKIE_LIFETIME < time_uptime) { 1667 return (NULL); 1668 } 1669 1670 /* Recompute the digest so we can compare it. */ 1671 MD5Init(&ctx); 1672 MD5Update(&ctx, ((u_int8_t *)secbits) + off, 1673 SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off); 1674 MD5Update(&ctx, secbits, off); 1675 MD5Update(&ctx, inc, sizeof(*inc)); 1676 MD5Update(&ctx, &seq, sizeof(seq)); 1677 MD5Update(&ctx, &flags, sizeof(flags)); 1678 MD5Final((u_int8_t *)&md5_buffer, &ctx); 1679 1680 /* Does the digest part of or ACK'ed ISS match? */ 1681 if ((ack & (~0x7f)) != (md5_buffer[0] << 7)) 1682 return (NULL); 1683 1684 /* Does the digest part of our reflected timestamp match? */ 1685 if (to->to_flags & TOF_TS) { 1686 data = md5_buffer[3] ^ to->to_tsecr; 1687 if ((data & (~0x3ff)) != (md5_buffer[2] << 10)) 1688 return (NULL); 1689 } 1690 1691 /* Fill in the syncache values. */ 1692 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1693 sc->sc_ipopts = NULL; 1694 1695 sc->sc_irs = seq; 1696 sc->sc_iss = ack; 1697 1698#ifdef INET6 1699 if (inc->inc_flags & INC_ISIPV6) { 1700 if (sotoinpcb(so)->inp_flags & IN6P_AUTOFLOWLABEL) 1701 sc->sc_flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK; 1702 } else 1703#endif 1704 { 1705 sc->sc_ip_ttl = sotoinpcb(so)->inp_ip_ttl; 1706 sc->sc_ip_tos = sotoinpcb(so)->inp_ip_tos; 1707 } 1708 1709 /* Additional parameters that were encoded in the timestamp. */ 1710 if (data) { 1711 sc->sc_flags |= SCF_TIMESTAMP; 1712 sc->sc_tsreflect = to->to_tsval; 1713 sc->sc_ts = to->to_tsecr; 1714 sc->sc_tsoff = to->to_tsecr - ticks; 1715 sc->sc_flags |= (data & 0x1) ? SCF_SIGNATURE : 0; 1716 sc->sc_flags |= ((data >> 1) & 0x1) ? SCF_SACK : 0; 1717 sc->sc_requested_s_scale = min((data >> 2) & 0xf, 1718 TCP_MAX_WINSHIFT); 1719 sc->sc_requested_r_scale = min((data >> 6) & 0xf, 1720 TCP_MAX_WINSHIFT); 1721 if (sc->sc_requested_s_scale || sc->sc_requested_r_scale) 1722 sc->sc_flags |= SCF_WINSCALE; 1723 } else 1724 sc->sc_flags |= SCF_NOOPT; 1725 1726 wnd = sbspace(&so->so_rcv); 1727 wnd = imax(wnd, 0); 1728 wnd = imin(wnd, TCP_MAXWIN); 1729 sc->sc_wnd = wnd; 1730 1731 sc->sc_rxmits = 0; 1732 sc->sc_peer_mss = tcp_sc_msstab[mss]; 1733 1734 TCPSTAT_INC(tcps_sc_recvcookie); 1735 return (sc); 1736} 1737 1738/* 1739 * Returns the current number of syncache entries. This number 1740 * will probably change before you get around to calling 1741 * syncache_pcblist. 1742 */ 1743 1744int 1745syncache_pcbcount(void) 1746{ 1747 struct syncache_head *sch; 1748 int count, i; 1749 1750 for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 1751 /* No need to lock for a read. */ 1752 sch = &V_tcp_syncache.hashbase[i]; 1753 count += sch->sch_length; 1754 } 1755 return count; 1756} 1757 1758/* 1759 * Exports the syncache entries to userland so that netstat can display 1760 * them alongside the other sockets. This function is intended to be 1761 * called only from tcp_pcblist. 1762 * 1763 * Due to concurrency on an active system, the number of pcbs exported 1764 * may have no relation to max_pcbs. max_pcbs merely indicates the 1765 * amount of space the caller allocated for this function to use. 1766 */ 1767int 1768syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 1769{ 1770 struct xtcpcb xt; 1771 struct syncache *sc; 1772 struct syncache_head *sch; 1773 int count, error, i; 1774 1775 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 1776 sch = &V_tcp_syncache.hashbase[i]; 1777 SCH_LOCK(sch); 1778 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 1779 if (count >= max_pcbs) { 1780 SCH_UNLOCK(sch); 1781 goto exit; 1782 } 1783 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 1784 continue; 1785 bzero(&xt, sizeof(xt)); 1786 xt.xt_len = sizeof(xt); 1787 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1788 xt.xt_inp.inp_vflag = INP_IPV6; 1789 else 1790 xt.xt_inp.inp_vflag = INP_IPV4; 1791 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 1792 xt.xt_tp.t_inpcb = &xt.xt_inp; 1793 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 1794 xt.xt_socket.xso_protocol = IPPROTO_TCP; 1795 xt.xt_socket.xso_len = sizeof (struct xsocket); 1796 xt.xt_socket.so_type = SOCK_STREAM; 1797 xt.xt_socket.so_state = SS_ISCONNECTING; 1798 error = SYSCTL_OUT(req, &xt, sizeof xt); 1799 if (error) { 1800 SCH_UNLOCK(sch); 1801 goto exit; 1802 } 1803 count++; 1804 } 1805 SCH_UNLOCK(sch); 1806 } 1807exit: 1808 *pcbs_exported = count; 1809 return error; 1810}
| 147#define V_tcp_syncache VNET(tcp_syncache) 148 149SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache"); 150 151SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN, 152 &VNET_NAME(tcp_syncache.bucket_limit), 0, 153 "Per-bucket hash limit for syncache"); 154 155SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN, 156 &VNET_NAME(tcp_syncache.cache_limit), 0, 157 "Overall entry limit for syncache"); 158 159SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD, 160 &VNET_NAME(tcp_syncache.cache_count), 0, 161 "Current number of entries in syncache"); 162 163SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN, 164 &VNET_NAME(tcp_syncache.hashsize), 0, 165 "Size of TCP syncache hashtable"); 166 167SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW, 168 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 169 "Limit on SYN/ACK retransmissions"); 170 171VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 172SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 173 CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 174 "Send reset on socket allocation failure"); 175 176static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 177 178#define SYNCACHE_HASH(inc, mask) \ 179 ((V_tcp_syncache.hash_secret ^ \ 180 (inc)->inc_faddr.s_addr ^ \ 181 ((inc)->inc_faddr.s_addr >> 16) ^ \ 182 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 183 184#define SYNCACHE_HASH6(inc, mask) \ 185 ((V_tcp_syncache.hash_secret ^ \ 186 (inc)->inc6_faddr.s6_addr32[0] ^ \ 187 (inc)->inc6_faddr.s6_addr32[3] ^ \ 188 (inc)->inc_fport ^ (inc)->inc_lport) & mask) 189 190#define ENDPTS_EQ(a, b) ( \ 191 (a)->ie_fport == (b)->ie_fport && \ 192 (a)->ie_lport == (b)->ie_lport && \ 193 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \ 194 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \ 195) 196 197#define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0) 198 199#define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 200#define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 201#define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 202 203/* 204 * Requires the syncache entry to be already removed from the bucket list. 205 */ 206static void 207syncache_free(struct syncache *sc) 208{ 209 210 if (sc->sc_ipopts) 211 (void) m_free(sc->sc_ipopts); 212 if (sc->sc_cred) 213 crfree(sc->sc_cred); 214#ifdef MAC 215 mac_syncache_destroy(&sc->sc_label); 216#endif 217 218 uma_zfree(V_tcp_syncache.zone, sc); 219} 220 221void 222syncache_init(void) 223{ 224 int i; 225 226 V_tcp_syncache.cache_count = 0; 227 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 228 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 229 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 230 V_tcp_syncache.hash_secret = arc4random(); 231 232 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 233 &V_tcp_syncache.hashsize); 234 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 235 &V_tcp_syncache.bucket_limit); 236 if (!powerof2(V_tcp_syncache.hashsize) || 237 V_tcp_syncache.hashsize == 0) { 238 printf("WARNING: syncache hash size is not a power of 2.\n"); 239 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 240 } 241 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 242 243 /* Set limits. */ 244 V_tcp_syncache.cache_limit = 245 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 246 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 247 &V_tcp_syncache.cache_limit); 248 249 /* Allocate the hash table. */ 250 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 251 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 252 253 /* Initialize the hash buckets. */ 254 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 255#ifdef VIMAGE 256 V_tcp_syncache.hashbase[i].sch_vnet = curvnet; 257#endif 258 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 259 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 260 NULL, MTX_DEF); 261 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 262 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 263 V_tcp_syncache.hashbase[i].sch_length = 0; 264 } 265 266 /* Create the syncache entry zone. */ 267 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 268 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 269 uma_zone_set_max(V_tcp_syncache.zone, V_tcp_syncache.cache_limit); 270} 271 272#ifdef VIMAGE 273void 274syncache_destroy(void) 275{ 276 struct syncache_head *sch; 277 struct syncache *sc, *nsc; 278 int i; 279 280 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 281 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 282 283 sch = &V_tcp_syncache.hashbase[i]; 284 callout_drain(&sch->sch_timer); 285 286 SCH_LOCK(sch); 287 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 288 syncache_drop(sc, sch); 289 SCH_UNLOCK(sch); 290 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 291 ("%s: sch->sch_bucket not empty", __func__)); 292 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 293 __func__, sch->sch_length)); 294 mtx_destroy(&sch->sch_mtx); 295 } 296 297 KASSERT(V_tcp_syncache.cache_count == 0, ("%s: cache_count %d not 0", 298 __func__, V_tcp_syncache.cache_count)); 299 300 /* Free the allocated global resources. */ 301 uma_zdestroy(V_tcp_syncache.zone); 302 free(V_tcp_syncache.hashbase, M_SYNCACHE); 303} 304#endif 305 306/* 307 * Inserts a syncache entry into the specified bucket row. 308 * Locks and unlocks the syncache_head autonomously. 309 */ 310static void 311syncache_insert(struct syncache *sc, struct syncache_head *sch) 312{ 313 struct syncache *sc2; 314 315 SCH_LOCK(sch); 316 317 /* 318 * Make sure that we don't overflow the per-bucket limit. 319 * If the bucket is full, toss the oldest element. 320 */ 321 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 322 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 323 ("sch->sch_length incorrect")); 324 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 325 syncache_drop(sc2, sch); 326 TCPSTAT_INC(tcps_sc_bucketoverflow); 327 } 328 329 /* Put it into the bucket. */ 330 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 331 sch->sch_length++; 332 333 /* Reinitialize the bucket row's timer. */ 334 if (sch->sch_length == 1) 335 sch->sch_nextc = ticks + INT_MAX; 336 syncache_timeout(sc, sch, 1); 337 338 SCH_UNLOCK(sch); 339 340 V_tcp_syncache.cache_count++; 341 TCPSTAT_INC(tcps_sc_added); 342} 343 344/* 345 * Remove and free entry from syncache bucket row. 346 * Expects locked syncache head. 347 */ 348static void 349syncache_drop(struct syncache *sc, struct syncache_head *sch) 350{ 351 352 SCH_LOCK_ASSERT(sch); 353 354 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 355 sch->sch_length--; 356 357#ifndef TCP_OFFLOAD_DISABLE 358 if (sc->sc_tu) 359 sc->sc_tu->tu_syncache_event(TOE_SC_DROP, sc->sc_toepcb); 360#endif 361 syncache_free(sc); 362 V_tcp_syncache.cache_count--; 363} 364 365/* 366 * Engage/reengage time on bucket row. 367 */ 368static void 369syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 370{ 371 sc->sc_rxttime = ticks + 372 TCPTV_RTOBASE * (tcp_backoff[sc->sc_rxmits]); 373 sc->sc_rxmits++; 374 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 375 sch->sch_nextc = sc->sc_rxttime; 376 if (docallout) 377 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 378 syncache_timer, (void *)sch); 379 } 380} 381 382/* 383 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 384 * If we have retransmitted an entry the maximum number of times, expire it. 385 * One separate timer for each bucket row. 386 */ 387static void 388syncache_timer(void *xsch) 389{ 390 struct syncache_head *sch = (struct syncache_head *)xsch; 391 struct syncache *sc, *nsc; 392 int tick = ticks; 393 char *s; 394 395 CURVNET_SET(sch->sch_vnet); 396 397 /* NB: syncache_head has already been locked by the callout. */ 398 SCH_LOCK_ASSERT(sch); 399 400 /* 401 * In the following cycle we may remove some entries and/or 402 * advance some timeouts, so re-initialize the bucket timer. 403 */ 404 sch->sch_nextc = tick + INT_MAX; 405 406 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 407 /* 408 * We do not check if the listen socket still exists 409 * and accept the case where the listen socket may be 410 * gone by the time we resend the SYN/ACK. We do 411 * not expect this to happens often. If it does, 412 * then the RST will be sent by the time the remote 413 * host does the SYN/ACK->ACK. 414 */ 415 if (TSTMP_GT(sc->sc_rxttime, tick)) { 416 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 417 sch->sch_nextc = sc->sc_rxttime; 418 continue; 419 } 420 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 421 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 422 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 423 "giving up and removing syncache entry\n", 424 s, __func__); 425 free(s, M_TCPLOG); 426 } 427 syncache_drop(sc, sch); 428 TCPSTAT_INC(tcps_sc_stale); 429 continue; 430 } 431 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 432 log(LOG_DEBUG, "%s; %s: Response timeout, " 433 "retransmitting (%u) SYN|ACK\n", 434 s, __func__, sc->sc_rxmits); 435 free(s, M_TCPLOG); 436 } 437 438 (void) syncache_respond(sc); 439 TCPSTAT_INC(tcps_sc_retransmitted); 440 syncache_timeout(sc, sch, 0); 441 } 442 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 443 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 444 syncache_timer, (void *)(sch)); 445 CURVNET_RESTORE(); 446} 447 448/* 449 * Find an entry in the syncache. 450 * Returns always with locked syncache_head plus a matching entry or NULL. 451 */ 452struct syncache * 453syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 454{ 455 struct syncache *sc; 456 struct syncache_head *sch; 457 458#ifdef INET6 459 if (inc->inc_flags & INC_ISIPV6) { 460 sch = &V_tcp_syncache.hashbase[ 461 SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)]; 462 *schp = sch; 463 464 SCH_LOCK(sch); 465 466 /* Circle through bucket row to find matching entry. */ 467 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 468 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 469 return (sc); 470 } 471 } else 472#endif 473 { 474 sch = &V_tcp_syncache.hashbase[ 475 SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)]; 476 *schp = sch; 477 478 SCH_LOCK(sch); 479 480 /* Circle through bucket row to find matching entry. */ 481 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 482#ifdef INET6 483 if (sc->sc_inc.inc_flags & INC_ISIPV6) 484 continue; 485#endif 486 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) 487 return (sc); 488 } 489 } 490 SCH_LOCK_ASSERT(*schp); 491 return (NULL); /* always returns with locked sch */ 492} 493 494/* 495 * This function is called when we get a RST for a 496 * non-existent connection, so that we can see if the 497 * connection is in the syn cache. If it is, zap it. 498 */ 499void 500syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th) 501{ 502 struct syncache *sc; 503 struct syncache_head *sch; 504 char *s = NULL; 505 506 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 507 SCH_LOCK_ASSERT(sch); 508 509 /* 510 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 511 * See RFC 793 page 65, section SEGMENT ARRIVES. 512 */ 513 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { 514 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 515 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 516 "FIN flag set, segment ignored\n", s, __func__); 517 TCPSTAT_INC(tcps_badrst); 518 goto done; 519 } 520 521 /* 522 * No corresponding connection was found in syncache. 523 * If syncookies are enabled and possibly exclusively 524 * used, or we are under memory pressure, a valid RST 525 * may not find a syncache entry. In that case we're 526 * done and no SYN|ACK retransmissions will happen. 527 * Otherwise the the RST was misdirected or spoofed. 528 */ 529 if (sc == NULL) { 530 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 531 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 532 "syncache entry (possibly syncookie only), " 533 "segment ignored\n", s, __func__); 534 TCPSTAT_INC(tcps_badrst); 535 goto done; 536 } 537 538 /* 539 * If the RST bit is set, check the sequence number to see 540 * if this is a valid reset segment. 541 * RFC 793 page 37: 542 * In all states except SYN-SENT, all reset (RST) segments 543 * are validated by checking their SEQ-fields. A reset is 544 * valid if its sequence number is in the window. 545 * 546 * The sequence number in the reset segment is normally an 547 * echo of our outgoing acknowlegement numbers, but some hosts 548 * send a reset with the sequence number at the rightmost edge 549 * of our receive window, and we have to handle this case. 550 */ 551 if (SEQ_GEQ(th->th_seq, sc->sc_irs) && 552 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 553 syncache_drop(sc, sch); 554 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 555 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " 556 "connection attempt aborted by remote endpoint\n", 557 s, __func__); 558 TCPSTAT_INC(tcps_sc_reset); 559 } else { 560 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 561 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 562 "IRS %u (+WND %u), segment ignored\n", 563 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd); 564 TCPSTAT_INC(tcps_badrst); 565 } 566 567done: 568 if (s != NULL) 569 free(s, M_TCPLOG); 570 SCH_UNLOCK(sch); 571} 572 573void 574syncache_badack(struct in_conninfo *inc) 575{ 576 struct syncache *sc; 577 struct syncache_head *sch; 578 579 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 580 SCH_LOCK_ASSERT(sch); 581 if (sc != NULL) { 582 syncache_drop(sc, sch); 583 TCPSTAT_INC(tcps_sc_badack); 584 } 585 SCH_UNLOCK(sch); 586} 587 588void 589syncache_unreach(struct in_conninfo *inc, struct tcphdr *th) 590{ 591 struct syncache *sc; 592 struct syncache_head *sch; 593 594 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 595 SCH_LOCK_ASSERT(sch); 596 if (sc == NULL) 597 goto done; 598 599 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 600 if (ntohl(th->th_seq) != sc->sc_iss) 601 goto done; 602 603 /* 604 * If we've rertransmitted 3 times and this is our second error, 605 * we remove the entry. Otherwise, we allow it to continue on. 606 * This prevents us from incorrectly nuking an entry during a 607 * spurious network outage. 608 * 609 * See tcp_notify(). 610 */ 611 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 612 sc->sc_flags |= SCF_UNREACH; 613 goto done; 614 } 615 syncache_drop(sc, sch); 616 TCPSTAT_INC(tcps_sc_unreach); 617done: 618 SCH_UNLOCK(sch); 619} 620 621/* 622 * Build a new TCP socket structure from a syncache entry. 623 */ 624static struct socket * 625syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 626{ 627 struct inpcb *inp = NULL; 628 struct socket *so; 629 struct tcpcb *tp; 630 int error; 631 char *s; 632 633 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 634 635 /* 636 * Ok, create the full blown connection, and set things up 637 * as they would have been set up if we had created the 638 * connection when the SYN arrived. If we can't create 639 * the connection, abort it. 640 */ 641 so = sonewconn(lso, SS_ISCONNECTED); 642 if (so == NULL) { 643 /* 644 * Drop the connection; we will either send a RST or 645 * have the peer retransmit its SYN again after its 646 * RTO and try again. 647 */ 648 TCPSTAT_INC(tcps_listendrop); 649 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 650 log(LOG_DEBUG, "%s; %s: Socket create failed " 651 "due to limits or memory shortage\n", 652 s, __func__); 653 free(s, M_TCPLOG); 654 } 655 goto abort2; 656 } 657#ifdef MAC 658 mac_socketpeer_set_from_mbuf(m, so); 659#endif 660 661 inp = sotoinpcb(so); 662 inp->inp_inc.inc_fibnum = so->so_fibnum; 663 INP_WLOCK(inp); 664 665 /* Insert new socket into PCB hash list. */ 666 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 667#ifdef INET6 668 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 669 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 670 } else { 671 inp->inp_vflag &= ~INP_IPV6; 672 inp->inp_vflag |= INP_IPV4; 673#endif 674 inp->inp_laddr = sc->sc_inc.inc_laddr; 675#ifdef INET6 676 } 677#endif 678 inp->inp_lport = sc->sc_inc.inc_lport; 679 if ((error = in_pcbinshash(inp)) != 0) { 680 /* 681 * Undo the assignments above if we failed to 682 * put the PCB on the hash lists. 683 */ 684#ifdef INET6 685 if (sc->sc_inc.inc_flags & INC_ISIPV6) 686 inp->in6p_laddr = in6addr_any; 687 else 688#endif 689 inp->inp_laddr.s_addr = INADDR_ANY; 690 inp->inp_lport = 0; 691 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 692 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed " 693 "with error %i\n", 694 s, __func__, error); 695 free(s, M_TCPLOG); 696 } 697 goto abort; 698 } 699#ifdef IPSEC 700 /* Copy old policy into new socket's. */ 701 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp)) 702 printf("syncache_socket: could not copy policy\n"); 703#endif 704#ifdef INET6 705 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 706 struct inpcb *oinp = sotoinpcb(lso); 707 struct in6_addr laddr6; 708 struct sockaddr_in6 sin6; 709 /* 710 * Inherit socket options from the listening socket. 711 * Note that in6p_inputopts are not (and should not be) 712 * copied, since it stores previously received options and is 713 * used to detect if each new option is different than the 714 * previous one and hence should be passed to a user. 715 * If we copied in6p_inputopts, a user would not be able to 716 * receive options just after calling the accept system call. 717 */ 718 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 719 if (oinp->in6p_outputopts) 720 inp->in6p_outputopts = 721 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 722 723 sin6.sin6_family = AF_INET6; 724 sin6.sin6_len = sizeof(sin6); 725 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 726 sin6.sin6_port = sc->sc_inc.inc_fport; 727 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 728 laddr6 = inp->in6p_laddr; 729 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 730 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 731 if ((error = in6_pcbconnect(inp, (struct sockaddr *)&sin6, 732 thread0.td_ucred)) != 0) { 733 inp->in6p_laddr = laddr6; 734 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 735 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " 736 "with error %i\n", 737 s, __func__, error); 738 free(s, M_TCPLOG); 739 } 740 goto abort; 741 } 742 /* Override flowlabel from in6_pcbconnect. */ 743 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 744 inp->inp_flow |= sc->sc_flowlabel; 745 } else 746#endif 747 { 748 struct in_addr laddr; 749 struct sockaddr_in sin; 750 751 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 752 753 if (inp->inp_options == NULL) { 754 inp->inp_options = sc->sc_ipopts; 755 sc->sc_ipopts = NULL; 756 } 757 758 sin.sin_family = AF_INET; 759 sin.sin_len = sizeof(sin); 760 sin.sin_addr = sc->sc_inc.inc_faddr; 761 sin.sin_port = sc->sc_inc.inc_fport; 762 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 763 laddr = inp->inp_laddr; 764 if (inp->inp_laddr.s_addr == INADDR_ANY) 765 inp->inp_laddr = sc->sc_inc.inc_laddr; 766 if ((error = in_pcbconnect(inp, (struct sockaddr *)&sin, 767 thread0.td_ucred)) != 0) { 768 inp->inp_laddr = laddr; 769 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 770 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " 771 "with error %i\n", 772 s, __func__, error); 773 free(s, M_TCPLOG); 774 } 775 goto abort; 776 } 777 } 778 tp = intotcpcb(inp); 779 tp->t_state = TCPS_SYN_RECEIVED; 780 tp->iss = sc->sc_iss; 781 tp->irs = sc->sc_irs; 782 tcp_rcvseqinit(tp); 783 tcp_sendseqinit(tp); 784 tp->snd_wl1 = sc->sc_irs; 785 tp->snd_max = tp->iss + 1; 786 tp->snd_nxt = tp->iss + 1; 787 tp->rcv_up = sc->sc_irs + 1; 788 tp->rcv_wnd = sc->sc_wnd; 789 tp->rcv_adv += tp->rcv_wnd; 790 tp->last_ack_sent = tp->rcv_nxt; 791 792 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 793 if (sc->sc_flags & SCF_NOOPT) 794 tp->t_flags |= TF_NOOPT; 795 else { 796 if (sc->sc_flags & SCF_WINSCALE) { 797 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 798 tp->snd_scale = sc->sc_requested_s_scale; 799 tp->request_r_scale = sc->sc_requested_r_scale; 800 } 801 if (sc->sc_flags & SCF_TIMESTAMP) { 802 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 803 tp->ts_recent = sc->sc_tsreflect; 804 tp->ts_recent_age = ticks; 805 tp->ts_offset = sc->sc_tsoff; 806 } 807#ifdef TCP_SIGNATURE 808 if (sc->sc_flags & SCF_SIGNATURE) 809 tp->t_flags |= TF_SIGNATURE; 810#endif 811 if (sc->sc_flags & SCF_SACK) 812 tp->t_flags |= TF_SACK_PERMIT; 813 } 814 815 if (sc->sc_flags & SCF_ECN) 816 tp->t_flags |= TF_ECN_PERMIT; 817 818 /* 819 * Set up MSS and get cached values from tcp_hostcache. 820 * This might overwrite some of the defaults we just set. 821 */ 822 tcp_mss(tp, sc->sc_peer_mss); 823 824 /* 825 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment. 826 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 827 */ 828 if (sc->sc_rxmits > 1) 829 tp->snd_cwnd = tp->t_maxseg; 830 tcp_timer_activate(tp, TT_KEEP, tcp_keepinit); 831 832 INP_WUNLOCK(inp); 833 834 TCPSTAT_INC(tcps_accepts); 835 return (so); 836 837abort: 838 INP_WUNLOCK(inp); 839abort2: 840 if (so != NULL) 841 soabort(so); 842 return (NULL); 843} 844 845/* 846 * This function gets called when we receive an ACK for a 847 * socket in the LISTEN state. We look up the connection 848 * in the syncache, and if its there, we pull it out of 849 * the cache and turn it into a full-blown connection in 850 * the SYN-RECEIVED state. 851 */ 852int 853syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 854 struct socket **lsop, struct mbuf *m) 855{ 856 struct syncache *sc; 857 struct syncache_head *sch; 858 struct syncache scs; 859 char *s; 860 861 /* 862 * Global TCP locks are held because we manipulate the PCB lists 863 * and create a new socket. 864 */ 865 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 866 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 867 ("%s: can handle only ACK", __func__)); 868 869 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 870 SCH_LOCK_ASSERT(sch); 871 if (sc == NULL) { 872 /* 873 * There is no syncache entry, so see if this ACK is 874 * a returning syncookie. To do this, first: 875 * A. See if this socket has had a syncache entry dropped in 876 * the past. We don't want to accept a bogus syncookie 877 * if we've never received a SYN. 878 * B. check that the syncookie is valid. If it is, then 879 * cobble up a fake syncache entry, and return. 880 */ 881 if (!V_tcp_syncookies) { 882 SCH_UNLOCK(sch); 883 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 884 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 885 "segment rejected (syncookies disabled)\n", 886 s, __func__); 887 goto failed; 888 } 889 bzero(&scs, sizeof(scs)); 890 sc = syncookie_lookup(inc, sch, &scs, to, th, *lsop); 891 SCH_UNLOCK(sch); 892 if (sc == NULL) { 893 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 894 log(LOG_DEBUG, "%s; %s: Segment failed " 895 "SYNCOOKIE authentication, segment rejected " 896 "(probably spoofed)\n", s, __func__); 897 goto failed; 898 } 899 } else { 900 /* Pull out the entry to unlock the bucket row. */ 901 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 902 sch->sch_length--; 903 V_tcp_syncache.cache_count--; 904 SCH_UNLOCK(sch); 905 } 906 907 /* 908 * Segment validation: 909 * ACK must match our initial sequence number + 1 (the SYN|ACK). 910 */ 911 if (th->th_ack != sc->sc_iss + 1 && !TOEPCB_ISSET(sc)) { 912 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 913 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 914 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 915 goto failed; 916 } 917 918 /* 919 * The SEQ must fall in the window starting at the received 920 * initial receive sequence number + 1 (the SYN). 921 */ 922 if ((SEQ_LEQ(th->th_seq, sc->sc_irs) || 923 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) && 924 !TOEPCB_ISSET(sc)) { 925 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 926 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 927 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 928 goto failed; 929 } 930 931 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { 932 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 933 log(LOG_DEBUG, "%s; %s: Timestamp not expected, " 934 "segment rejected\n", s, __func__); 935 goto failed; 936 } 937 /* 938 * If timestamps were negotiated the reflected timestamp 939 * must be equal to what we actually sent in the SYN|ACK. 940 */ 941 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts && 942 !TOEPCB_ISSET(sc)) { 943 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 944 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, " 945 "segment rejected\n", 946 s, __func__, to->to_tsecr, sc->sc_ts); 947 goto failed; 948 } 949 950 *lsop = syncache_socket(sc, *lsop, m); 951 952 if (*lsop == NULL) 953 TCPSTAT_INC(tcps_sc_aborted); 954 else 955 TCPSTAT_INC(tcps_sc_completed); 956 957/* how do we find the inp for the new socket? */ 958 if (sc != &scs) 959 syncache_free(sc); 960 return (1); 961failed: 962 if (sc != NULL && sc != &scs) 963 syncache_free(sc); 964 if (s != NULL) 965 free(s, M_TCPLOG); 966 *lsop = NULL; 967 return (0); 968} 969 970int 971tcp_offload_syncache_expand(struct in_conninfo *inc, struct toeopt *toeo, 972 struct tcphdr *th, struct socket **lsop, struct mbuf *m) 973{ 974 struct tcpopt to; 975 int rc; 976 977 bzero(&to, sizeof(struct tcpopt)); 978 to.to_mss = toeo->to_mss; 979 to.to_wscale = toeo->to_wscale; 980 to.to_flags = toeo->to_flags; 981 982 INP_INFO_WLOCK(&V_tcbinfo); 983 rc = syncache_expand(inc, &to, th, lsop, m); 984 INP_INFO_WUNLOCK(&V_tcbinfo); 985 986 return (rc); 987} 988 989/* 990 * Given a LISTEN socket and an inbound SYN request, add 991 * this to the syn cache, and send back a segment: 992 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 993 * to the source. 994 * 995 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 996 * Doing so would require that we hold onto the data and deliver it 997 * to the application. However, if we are the target of a SYN-flood 998 * DoS attack, an attacker could send data which would eventually 999 * consume all available buffer space if it were ACKed. By not ACKing 1000 * the data, we avoid this DoS scenario. 1001 */ 1002static void 1003_syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1004 struct inpcb *inp, struct socket **lsop, struct mbuf *m, 1005 struct toe_usrreqs *tu, void *toepcb) 1006{ 1007 struct tcpcb *tp; 1008 struct socket *so; 1009 struct syncache *sc = NULL; 1010 struct syncache_head *sch; 1011 struct mbuf *ipopts = NULL; 1012 u_int32_t flowtmp; 1013 int win, sb_hiwat, ip_ttl, ip_tos, noopt; 1014 char *s; 1015#ifdef INET6 1016 int autoflowlabel = 0; 1017#endif 1018#ifdef MAC 1019 struct label *maclabel; 1020#endif 1021 struct syncache scs; 1022 struct ucred *cred; 1023 1024 INP_INFO_WLOCK_ASSERT(&V_tcbinfo); 1025 INP_WLOCK_ASSERT(inp); /* listen socket */ 1026 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1027 ("%s: unexpected tcp flags", __func__)); 1028 1029 /* 1030 * Combine all so/tp operations very early to drop the INP lock as 1031 * soon as possible. 1032 */ 1033 so = *lsop; 1034 tp = sototcpcb(so); 1035 cred = crhold(so->so_cred); 1036 1037#ifdef INET6 1038 if ((inc->inc_flags & INC_ISIPV6) && 1039 (inp->inp_flags & IN6P_AUTOFLOWLABEL)) 1040 autoflowlabel = 1; 1041#endif 1042 ip_ttl = inp->inp_ip_ttl; 1043 ip_tos = inp->inp_ip_tos; 1044 win = sbspace(&so->so_rcv); 1045 sb_hiwat = so->so_rcv.sb_hiwat; 1046 noopt = (tp->t_flags & TF_NOOPT); 1047 1048 /* By the time we drop the lock these should no longer be used. */ 1049 so = NULL; 1050 tp = NULL; 1051 1052#ifdef MAC 1053 if (mac_syncache_init(&maclabel) != 0) { 1054 INP_WUNLOCK(inp); 1055 INP_INFO_WUNLOCK(&V_tcbinfo); 1056 goto done; 1057 } else 1058 mac_syncache_create(maclabel, inp); 1059#endif 1060 INP_WUNLOCK(inp); 1061 INP_INFO_WUNLOCK(&V_tcbinfo); 1062 1063 /* 1064 * Remember the IP options, if any. 1065 */ 1066#ifdef INET6 1067 if (!(inc->inc_flags & INC_ISIPV6)) 1068#endif 1069 ipopts = (m) ? ip_srcroute(m) : NULL; 1070 1071 /* 1072 * See if we already have an entry for this connection. 1073 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1074 * 1075 * XXX: should the syncache be re-initialized with the contents 1076 * of the new SYN here (which may have different options?) 1077 * 1078 * XXX: We do not check the sequence number to see if this is a 1079 * real retransmit or a new connection attempt. The question is 1080 * how to handle such a case; either ignore it as spoofed, or 1081 * drop the current entry and create a new one? 1082 */ 1083 sc = syncache_lookup(inc, &sch); /* returns locked entry */ 1084 SCH_LOCK_ASSERT(sch); 1085 if (sc != NULL) { 1086#ifndef TCP_OFFLOAD_DISABLE 1087 if (sc->sc_tu) 1088 sc->sc_tu->tu_syncache_event(TOE_SC_ENTRY_PRESENT, 1089 sc->sc_toepcb); 1090#endif 1091 TCPSTAT_INC(tcps_sc_dupsyn); 1092 if (ipopts) { 1093 /* 1094 * If we were remembering a previous source route, 1095 * forget it and use the new one we've been given. 1096 */ 1097 if (sc->sc_ipopts) 1098 (void) m_free(sc->sc_ipopts); 1099 sc->sc_ipopts = ipopts; 1100 } 1101 /* 1102 * Update timestamp if present. 1103 */ 1104 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1105 sc->sc_tsreflect = to->to_tsval; 1106 else 1107 sc->sc_flags &= ~SCF_TIMESTAMP; 1108#ifdef MAC 1109 /* 1110 * Since we have already unconditionally allocated label 1111 * storage, free it up. The syncache entry will already 1112 * have an initialized label we can use. 1113 */ 1114 mac_syncache_destroy(&maclabel); 1115#endif 1116 /* Retransmit SYN|ACK and reset retransmit count. */ 1117 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1118 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1119 "resetting timer and retransmitting SYN|ACK\n", 1120 s, __func__); 1121 free(s, M_TCPLOG); 1122 } 1123 if (!TOEPCB_ISSET(sc) && syncache_respond(sc) == 0) { 1124 sc->sc_rxmits = 0; 1125 syncache_timeout(sc, sch, 1); 1126 TCPSTAT_INC(tcps_sndacks); 1127 TCPSTAT_INC(tcps_sndtotal); 1128 } 1129 SCH_UNLOCK(sch); 1130 goto done; 1131 } 1132 1133 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1134 if (sc == NULL) { 1135 /* 1136 * The zone allocator couldn't provide more entries. 1137 * Treat this as if the cache was full; drop the oldest 1138 * entry and insert the new one. 1139 */ 1140 TCPSTAT_INC(tcps_sc_zonefail); 1141 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) 1142 syncache_drop(sc, sch); 1143 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1144 if (sc == NULL) { 1145 if (V_tcp_syncookies) { 1146 bzero(&scs, sizeof(scs)); 1147 sc = &scs; 1148 } else { 1149 SCH_UNLOCK(sch); 1150 if (ipopts) 1151 (void) m_free(ipopts); 1152 goto done; 1153 } 1154 } 1155 } 1156 1157 /* 1158 * Fill in the syncache values. 1159 */ 1160#ifdef MAC 1161 sc->sc_label = maclabel; 1162#endif 1163 sc->sc_cred = cred; 1164 cred = NULL; 1165 sc->sc_ipopts = ipopts; 1166 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1167#ifdef INET6 1168 if (!(inc->inc_flags & INC_ISIPV6)) 1169#endif 1170 { 1171 sc->sc_ip_tos = ip_tos; 1172 sc->sc_ip_ttl = ip_ttl; 1173 } 1174#ifndef TCP_OFFLOAD_DISABLE 1175 sc->sc_tu = tu; 1176 sc->sc_toepcb = toepcb; 1177#endif 1178 sc->sc_irs = th->th_seq; 1179 sc->sc_iss = arc4random(); 1180 sc->sc_flags = 0; 1181 sc->sc_flowlabel = 0; 1182 1183 /* 1184 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1185 * win was derived from socket earlier in the function. 1186 */ 1187 win = imax(win, 0); 1188 win = imin(win, TCP_MAXWIN); 1189 sc->sc_wnd = win; 1190 1191 if (V_tcp_do_rfc1323) { 1192 /* 1193 * A timestamp received in a SYN makes 1194 * it ok to send timestamp requests and replies. 1195 */ 1196 if (to->to_flags & TOF_TS) { 1197 sc->sc_tsreflect = to->to_tsval; 1198 sc->sc_ts = ticks; 1199 sc->sc_flags |= SCF_TIMESTAMP; 1200 } 1201 if (to->to_flags & TOF_SCALE) { 1202 int wscale = 0; 1203 1204 /* 1205 * Pick the smallest possible scaling factor that 1206 * will still allow us to scale up to sb_max, aka 1207 * kern.ipc.maxsockbuf. 1208 * 1209 * We do this because there are broken firewalls that 1210 * will corrupt the window scale option, leading to 1211 * the other endpoint believing that our advertised 1212 * window is unscaled. At scale factors larger than 1213 * 5 the unscaled window will drop below 1500 bytes, 1214 * leading to serious problems when traversing these 1215 * broken firewalls. 1216 * 1217 * With the default maxsockbuf of 256K, a scale factor 1218 * of 3 will be chosen by this algorithm. Those who 1219 * choose a larger maxsockbuf should watch out 1220 * for the compatiblity problems mentioned above. 1221 * 1222 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1223 * or <SYN,ACK>) segment itself is never scaled. 1224 */ 1225 while (wscale < TCP_MAX_WINSHIFT && 1226 (TCP_MAXWIN << wscale) < sb_max) 1227 wscale++; 1228 sc->sc_requested_r_scale = wscale; 1229 sc->sc_requested_s_scale = to->to_wscale; 1230 sc->sc_flags |= SCF_WINSCALE; 1231 } 1232 } 1233#ifdef TCP_SIGNATURE 1234 /* 1235 * If listening socket requested TCP digests, and received SYN 1236 * contains the option, flag this in the syncache so that 1237 * syncache_respond() will do the right thing with the SYN+ACK. 1238 * XXX: Currently we always record the option by default and will 1239 * attempt to use it in syncache_respond(). 1240 */ 1241 if (to->to_flags & TOF_SIGNATURE) 1242 sc->sc_flags |= SCF_SIGNATURE; 1243#endif 1244 if (to->to_flags & TOF_SACKPERM) 1245 sc->sc_flags |= SCF_SACK; 1246 if (to->to_flags & TOF_MSS) 1247 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1248 if (noopt) 1249 sc->sc_flags |= SCF_NOOPT; 1250 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn) 1251 sc->sc_flags |= SCF_ECN; 1252 1253 if (V_tcp_syncookies) { 1254 syncookie_generate(sch, sc, &flowtmp); 1255#ifdef INET6 1256 if (autoflowlabel) 1257 sc->sc_flowlabel = flowtmp; 1258#endif 1259 } else { 1260#ifdef INET6 1261 if (autoflowlabel) 1262 sc->sc_flowlabel = 1263 (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK); 1264#endif 1265 } 1266 SCH_UNLOCK(sch); 1267 1268 /* 1269 * Do a standard 3-way handshake. 1270 */ 1271 if (TOEPCB_ISSET(sc) || syncache_respond(sc) == 0) { 1272 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1273 syncache_free(sc); 1274 else if (sc != &scs) 1275 syncache_insert(sc, sch); /* locks and unlocks sch */ 1276 TCPSTAT_INC(tcps_sndacks); 1277 TCPSTAT_INC(tcps_sndtotal); 1278 } else { 1279 if (sc != &scs) 1280 syncache_free(sc); 1281 TCPSTAT_INC(tcps_sc_dropped); 1282 } 1283 1284done: 1285 if (cred != NULL) 1286 crfree(cred); 1287#ifdef MAC 1288 if (sc == &scs) 1289 mac_syncache_destroy(&maclabel); 1290#endif 1291 if (m) { 1292 1293 *lsop = NULL; 1294 m_freem(m); 1295 } 1296} 1297 1298static int 1299syncache_respond(struct syncache *sc) 1300{ 1301 struct ip *ip = NULL; 1302 struct mbuf *m; 1303 struct tcphdr *th; 1304 int optlen, error; 1305 u_int16_t hlen, tlen, mssopt; 1306 struct tcpopt to; 1307#ifdef INET6 1308 struct ip6_hdr *ip6 = NULL; 1309#endif 1310 1311 hlen = 1312#ifdef INET6 1313 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1314#endif 1315 sizeof(struct ip); 1316 tlen = hlen + sizeof(struct tcphdr); 1317 1318 /* Determine MSS we advertize to other end of connection. */ 1319 mssopt = tcp_mssopt(&sc->sc_inc); 1320 if (sc->sc_peer_mss) 1321 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss); 1322 1323 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1324 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1325 ("syncache: mbuf too small")); 1326 1327 /* Create the IP+TCP header from scratch. */ 1328 m = m_gethdr(M_DONTWAIT, MT_DATA); 1329 if (m == NULL) 1330 return (ENOBUFS); 1331#ifdef MAC 1332 mac_syncache_create_mbuf(sc->sc_label, m); 1333#endif 1334 m->m_data += max_linkhdr; 1335 m->m_len = tlen; 1336 m->m_pkthdr.len = tlen; 1337 m->m_pkthdr.rcvif = NULL; 1338 1339#ifdef INET6 1340 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1341 ip6 = mtod(m, struct ip6_hdr *); 1342 ip6->ip6_vfc = IPV6_VERSION; 1343 ip6->ip6_nxt = IPPROTO_TCP; 1344 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1345 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1346 ip6->ip6_plen = htons(tlen - hlen); 1347 /* ip6_hlim is set after checksum */ 1348 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK; 1349 ip6->ip6_flow |= sc->sc_flowlabel; 1350 1351 th = (struct tcphdr *)(ip6 + 1); 1352 } else 1353#endif 1354 { 1355 ip = mtod(m, struct ip *); 1356 ip->ip_v = IPVERSION; 1357 ip->ip_hl = sizeof(struct ip) >> 2; 1358 ip->ip_len = tlen; 1359 ip->ip_id = 0; 1360 ip->ip_off = 0; 1361 ip->ip_sum = 0; 1362 ip->ip_p = IPPROTO_TCP; 1363 ip->ip_src = sc->sc_inc.inc_laddr; 1364 ip->ip_dst = sc->sc_inc.inc_faddr; 1365 ip->ip_ttl = sc->sc_ip_ttl; 1366 ip->ip_tos = sc->sc_ip_tos; 1367 1368 /* 1369 * See if we should do MTU discovery. Route lookups are 1370 * expensive, so we will only unset the DF bit if: 1371 * 1372 * 1) path_mtu_discovery is disabled 1373 * 2) the SCF_UNREACH flag has been set 1374 */ 1375 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1376 ip->ip_off |= IP_DF; 1377 1378 th = (struct tcphdr *)(ip + 1); 1379 } 1380 th->th_sport = sc->sc_inc.inc_lport; 1381 th->th_dport = sc->sc_inc.inc_fport; 1382 1383 th->th_seq = htonl(sc->sc_iss); 1384 th->th_ack = htonl(sc->sc_irs + 1); 1385 th->th_off = sizeof(struct tcphdr) >> 2; 1386 th->th_x2 = 0; 1387 th->th_flags = TH_SYN|TH_ACK; 1388 th->th_win = htons(sc->sc_wnd); 1389 th->th_urp = 0; 1390 1391 if (sc->sc_flags & SCF_ECN) { 1392 th->th_flags |= TH_ECE; 1393 TCPSTAT_INC(tcps_ecn_shs); 1394 } 1395 1396 /* Tack on the TCP options. */ 1397 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1398 to.to_flags = 0; 1399 1400 to.to_mss = mssopt; 1401 to.to_flags = TOF_MSS; 1402 if (sc->sc_flags & SCF_WINSCALE) { 1403 to.to_wscale = sc->sc_requested_r_scale; 1404 to.to_flags |= TOF_SCALE; 1405 } 1406 if (sc->sc_flags & SCF_TIMESTAMP) { 1407 /* Virgin timestamp or TCP cookie enhanced one. */ 1408 to.to_tsval = sc->sc_ts; 1409 to.to_tsecr = sc->sc_tsreflect; 1410 to.to_flags |= TOF_TS; 1411 } 1412 if (sc->sc_flags & SCF_SACK) 1413 to.to_flags |= TOF_SACKPERM; 1414#ifdef TCP_SIGNATURE 1415 if (sc->sc_flags & SCF_SIGNATURE) 1416 to.to_flags |= TOF_SIGNATURE; 1417#endif 1418 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1419 1420 /* Adjust headers by option size. */ 1421 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1422 m->m_len += optlen; 1423 m->m_pkthdr.len += optlen; 1424 1425#ifdef TCP_SIGNATURE 1426 if (sc->sc_flags & SCF_SIGNATURE) 1427 tcp_signature_compute(m, 0, 0, optlen, 1428 to.to_signature, IPSEC_DIR_OUTBOUND); 1429#endif 1430#ifdef INET6 1431 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1432 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1433 else 1434#endif 1435 ip->ip_len += optlen; 1436 } else 1437 optlen = 0; 1438 1439 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1440#ifdef INET6 1441 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1442 th->th_sum = 0; 1443 th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, 1444 tlen + optlen - hlen); 1445 ip6->ip6_hlim = in6_selecthlim(NULL, NULL); 1446 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 1447 } else 1448#endif 1449 { 1450 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 1451 htons(tlen + optlen - hlen + IPPROTO_TCP)); 1452 m->m_pkthdr.csum_flags = CSUM_TCP; 1453 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 1454 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 1455 } 1456 return (error); 1457} 1458 1459void 1460syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1461 struct inpcb *inp, struct socket **lsop, struct mbuf *m) 1462{ 1463 _syncache_add(inc, to, th, inp, lsop, m, NULL, NULL); 1464} 1465 1466void 1467tcp_offload_syncache_add(struct in_conninfo *inc, struct toeopt *toeo, 1468 struct tcphdr *th, struct inpcb *inp, struct socket **lsop, 1469 struct toe_usrreqs *tu, void *toepcb) 1470{ 1471 struct tcpopt to; 1472 1473 bzero(&to, sizeof(struct tcpopt)); 1474 to.to_mss = toeo->to_mss; 1475 to.to_wscale = toeo->to_wscale; 1476 to.to_flags = toeo->to_flags; 1477 1478 INP_INFO_WLOCK(&V_tcbinfo); 1479 INP_WLOCK(inp); 1480 1481 _syncache_add(inc, &to, th, inp, lsop, NULL, tu, toepcb); 1482} 1483 1484/* 1485 * The purpose of SYN cookies is to avoid keeping track of all SYN's we 1486 * receive and to be able to handle SYN floods from bogus source addresses 1487 * (where we will never receive any reply). SYN floods try to exhaust all 1488 * our memory and available slots in the SYN cache table to cause a denial 1489 * of service to legitimate users of the local host. 1490 * 1491 * The idea of SYN cookies is to encode and include all necessary information 1492 * about the connection setup state within the SYN-ACK we send back and thus 1493 * to get along without keeping any local state until the ACK to the SYN-ACK 1494 * arrives (if ever). Everything we need to know should be available from 1495 * the information we encoded in the SYN-ACK. 1496 * 1497 * More information about the theory behind SYN cookies and its first 1498 * discussion and specification can be found at: 1499 * http://cr.yp.to/syncookies.html (overview) 1500 * http://cr.yp.to/syncookies/archive (gory details) 1501 * 1502 * This implementation extends the orginal idea and first implementation 1503 * of FreeBSD by using not only the initial sequence number field to store 1504 * information but also the timestamp field if present. This way we can 1505 * keep track of the entire state we need to know to recreate the session in 1506 * its original form. Almost all TCP speakers implement RFC1323 timestamps 1507 * these days. For those that do not we still have to live with the known 1508 * shortcomings of the ISN only SYN cookies. 1509 * 1510 * Cookie layers: 1511 * 1512 * Initial sequence number we send: 1513 * 31|................................|0 1514 * DDDDDDDDDDDDDDDDDDDDDDDDDMMMRRRP 1515 * D = MD5 Digest (first dword) 1516 * M = MSS index 1517 * R = Rotation of secret 1518 * P = Odd or Even secret 1519 * 1520 * The MD5 Digest is computed with over following parameters: 1521 * a) randomly rotated secret 1522 * b) struct in_conninfo containing the remote/local ip/port (IPv4&IPv6) 1523 * c) the received initial sequence number from remote host 1524 * d) the rotation offset and odd/even bit 1525 * 1526 * Timestamp we send: 1527 * 31|................................|0 1528 * DDDDDDDDDDDDDDDDDDDDDDSSSSRRRRA5 1529 * D = MD5 Digest (third dword) (only as filler) 1530 * S = Requested send window scale 1531 * R = Requested receive window scale 1532 * A = SACK allowed 1533 * 5 = TCP-MD5 enabled (not implemented yet) 1534 * XORed with MD5 Digest (forth dword) 1535 * 1536 * The timestamp isn't cryptographically secure and doesn't need to be. 1537 * The double use of the MD5 digest dwords ties it to a specific remote/ 1538 * local host/port, remote initial sequence number and our local time 1539 * limited secret. A received timestamp is reverted (XORed) and then 1540 * the contained MD5 dword is compared to the computed one to ensure the 1541 * timestamp belongs to the SYN-ACK we sent. The other parameters may 1542 * have been tampered with but this isn't different from supplying bogus 1543 * values in the SYN in the first place. 1544 * 1545 * Some problems with SYN cookies remain however: 1546 * Consider the problem of a recreated (and retransmitted) cookie. If the 1547 * original SYN was accepted, the connection is established. The second 1548 * SYN is inflight, and if it arrives with an ISN that falls within the 1549 * receive window, the connection is killed. 1550 * 1551 * Notes: 1552 * A heuristic to determine when to accept syn cookies is not necessary. 1553 * An ACK flood would cause the syncookie verification to be attempted, 1554 * but a SYN flood causes syncookies to be generated. Both are of equal 1555 * cost, so there's no point in trying to optimize the ACK flood case. 1556 * Also, if you don't process certain ACKs for some reason, then all someone 1557 * would have to do is launch a SYN and ACK flood at the same time, which 1558 * would stop cookie verification and defeat the entire purpose of syncookies. 1559 */ 1560static int tcp_sc_msstab[] = { 0, 256, 468, 536, 996, 1452, 1460, 8960 }; 1561 1562static void 1563syncookie_generate(struct syncache_head *sch, struct syncache *sc, 1564 u_int32_t *flowlabel) 1565{ 1566 MD5_CTX ctx; 1567 u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)]; 1568 u_int32_t data; 1569 u_int32_t *secbits; 1570 u_int off, pmss, mss; 1571 int i; 1572 1573 SCH_LOCK_ASSERT(sch); 1574 1575 /* Which of the two secrets to use. */ 1576 secbits = sch->sch_oddeven ? 1577 sch->sch_secbits_odd : sch->sch_secbits_even; 1578 1579 /* Reseed secret if too old. */ 1580 if (sch->sch_reseed < time_uptime) { 1581 sch->sch_oddeven = sch->sch_oddeven ? 0 : 1; /* toggle */ 1582 secbits = sch->sch_oddeven ? 1583 sch->sch_secbits_odd : sch->sch_secbits_even; 1584 for (i = 0; i < SYNCOOKIE_SECRET_SIZE; i++) 1585 secbits[i] = arc4random(); 1586 sch->sch_reseed = time_uptime + SYNCOOKIE_LIFETIME; 1587 } 1588 1589 /* Secret rotation offset. */ 1590 off = sc->sc_iss & 0x7; /* iss was randomized before */ 1591 1592 /* Maximum segment size calculation. */ 1593 pmss = 1594 max( min(sc->sc_peer_mss, tcp_mssopt(&sc->sc_inc)), V_tcp_minmss); 1595 for (mss = sizeof(tcp_sc_msstab) / sizeof(int) - 1; mss > 0; mss--) 1596 if (tcp_sc_msstab[mss] <= pmss) 1597 break; 1598 1599 /* Fold parameters and MD5 digest into the ISN we will send. */ 1600 data = sch->sch_oddeven;/* odd or even secret, 1 bit */ 1601 data |= off << 1; /* secret offset, derived from iss, 3 bits */ 1602 data |= mss << 4; /* mss, 3 bits */ 1603 1604 MD5Init(&ctx); 1605 MD5Update(&ctx, ((u_int8_t *)secbits) + off, 1606 SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off); 1607 MD5Update(&ctx, secbits, off); 1608 MD5Update(&ctx, &sc->sc_inc, sizeof(sc->sc_inc)); 1609 MD5Update(&ctx, &sc->sc_irs, sizeof(sc->sc_irs)); 1610 MD5Update(&ctx, &data, sizeof(data)); 1611 MD5Final((u_int8_t *)&md5_buffer, &ctx); 1612 1613 data |= (md5_buffer[0] << 7); 1614 sc->sc_iss = data; 1615 1616#ifdef INET6 1617 *flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK; 1618#endif 1619 1620 /* Additional parameters are stored in the timestamp if present. */ 1621 if (sc->sc_flags & SCF_TIMESTAMP) { 1622 data = ((sc->sc_flags & SCF_SIGNATURE) ? 1 : 0); /* TCP-MD5, 1 bit */ 1623 data |= ((sc->sc_flags & SCF_SACK) ? 1 : 0) << 1; /* SACK, 1 bit */ 1624 data |= sc->sc_requested_s_scale << 2; /* SWIN scale, 4 bits */ 1625 data |= sc->sc_requested_r_scale << 6; /* RWIN scale, 4 bits */ 1626 data |= md5_buffer[2] << 10; /* more digest bits */ 1627 data ^= md5_buffer[3]; 1628 sc->sc_ts = data; 1629 sc->sc_tsoff = data - ticks; /* after XOR */ 1630 } 1631 1632 TCPSTAT_INC(tcps_sc_sendcookie); 1633} 1634 1635static struct syncache * 1636syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 1637 struct syncache *sc, struct tcpopt *to, struct tcphdr *th, 1638 struct socket *so) 1639{ 1640 MD5_CTX ctx; 1641 u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)]; 1642 u_int32_t data = 0; 1643 u_int32_t *secbits; 1644 tcp_seq ack, seq; 1645 int off, mss, wnd, flags; 1646 1647 SCH_LOCK_ASSERT(sch); 1648 1649 /* 1650 * Pull information out of SYN-ACK/ACK and 1651 * revert sequence number advances. 1652 */ 1653 ack = th->th_ack - 1; 1654 seq = th->th_seq - 1; 1655 off = (ack >> 1) & 0x7; 1656 mss = (ack >> 4) & 0x7; 1657 flags = ack & 0x7f; 1658 1659 /* Which of the two secrets to use. */ 1660 secbits = (flags & 0x1) ? sch->sch_secbits_odd : sch->sch_secbits_even; 1661 1662 /* 1663 * The secret wasn't updated for the lifetime of a syncookie, 1664 * so this SYN-ACK/ACK is either too old (replay) or totally bogus. 1665 */ 1666 if (sch->sch_reseed + SYNCOOKIE_LIFETIME < time_uptime) { 1667 return (NULL); 1668 } 1669 1670 /* Recompute the digest so we can compare it. */ 1671 MD5Init(&ctx); 1672 MD5Update(&ctx, ((u_int8_t *)secbits) + off, 1673 SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off); 1674 MD5Update(&ctx, secbits, off); 1675 MD5Update(&ctx, inc, sizeof(*inc)); 1676 MD5Update(&ctx, &seq, sizeof(seq)); 1677 MD5Update(&ctx, &flags, sizeof(flags)); 1678 MD5Final((u_int8_t *)&md5_buffer, &ctx); 1679 1680 /* Does the digest part of or ACK'ed ISS match? */ 1681 if ((ack & (~0x7f)) != (md5_buffer[0] << 7)) 1682 return (NULL); 1683 1684 /* Does the digest part of our reflected timestamp match? */ 1685 if (to->to_flags & TOF_TS) { 1686 data = md5_buffer[3] ^ to->to_tsecr; 1687 if ((data & (~0x3ff)) != (md5_buffer[2] << 10)) 1688 return (NULL); 1689 } 1690 1691 /* Fill in the syncache values. */ 1692 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1693 sc->sc_ipopts = NULL; 1694 1695 sc->sc_irs = seq; 1696 sc->sc_iss = ack; 1697 1698#ifdef INET6 1699 if (inc->inc_flags & INC_ISIPV6) { 1700 if (sotoinpcb(so)->inp_flags & IN6P_AUTOFLOWLABEL) 1701 sc->sc_flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK; 1702 } else 1703#endif 1704 { 1705 sc->sc_ip_ttl = sotoinpcb(so)->inp_ip_ttl; 1706 sc->sc_ip_tos = sotoinpcb(so)->inp_ip_tos; 1707 } 1708 1709 /* Additional parameters that were encoded in the timestamp. */ 1710 if (data) { 1711 sc->sc_flags |= SCF_TIMESTAMP; 1712 sc->sc_tsreflect = to->to_tsval; 1713 sc->sc_ts = to->to_tsecr; 1714 sc->sc_tsoff = to->to_tsecr - ticks; 1715 sc->sc_flags |= (data & 0x1) ? SCF_SIGNATURE : 0; 1716 sc->sc_flags |= ((data >> 1) & 0x1) ? SCF_SACK : 0; 1717 sc->sc_requested_s_scale = min((data >> 2) & 0xf, 1718 TCP_MAX_WINSHIFT); 1719 sc->sc_requested_r_scale = min((data >> 6) & 0xf, 1720 TCP_MAX_WINSHIFT); 1721 if (sc->sc_requested_s_scale || sc->sc_requested_r_scale) 1722 sc->sc_flags |= SCF_WINSCALE; 1723 } else 1724 sc->sc_flags |= SCF_NOOPT; 1725 1726 wnd = sbspace(&so->so_rcv); 1727 wnd = imax(wnd, 0); 1728 wnd = imin(wnd, TCP_MAXWIN); 1729 sc->sc_wnd = wnd; 1730 1731 sc->sc_rxmits = 0; 1732 sc->sc_peer_mss = tcp_sc_msstab[mss]; 1733 1734 TCPSTAT_INC(tcps_sc_recvcookie); 1735 return (sc); 1736} 1737 1738/* 1739 * Returns the current number of syncache entries. This number 1740 * will probably change before you get around to calling 1741 * syncache_pcblist. 1742 */ 1743 1744int 1745syncache_pcbcount(void) 1746{ 1747 struct syncache_head *sch; 1748 int count, i; 1749 1750 for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 1751 /* No need to lock for a read. */ 1752 sch = &V_tcp_syncache.hashbase[i]; 1753 count += sch->sch_length; 1754 } 1755 return count; 1756} 1757 1758/* 1759 * Exports the syncache entries to userland so that netstat can display 1760 * them alongside the other sockets. This function is intended to be 1761 * called only from tcp_pcblist. 1762 * 1763 * Due to concurrency on an active system, the number of pcbs exported 1764 * may have no relation to max_pcbs. max_pcbs merely indicates the 1765 * amount of space the caller allocated for this function to use. 1766 */ 1767int 1768syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported) 1769{ 1770 struct xtcpcb xt; 1771 struct syncache *sc; 1772 struct syncache_head *sch; 1773 int count, error, i; 1774 1775 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) { 1776 sch = &V_tcp_syncache.hashbase[i]; 1777 SCH_LOCK(sch); 1778 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 1779 if (count >= max_pcbs) { 1780 SCH_UNLOCK(sch); 1781 goto exit; 1782 } 1783 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 1784 continue; 1785 bzero(&xt, sizeof(xt)); 1786 xt.xt_len = sizeof(xt); 1787 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1788 xt.xt_inp.inp_vflag = INP_IPV6; 1789 else 1790 xt.xt_inp.inp_vflag = INP_IPV4; 1791 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); 1792 xt.xt_tp.t_inpcb = &xt.xt_inp; 1793 xt.xt_tp.t_state = TCPS_SYN_RECEIVED; 1794 xt.xt_socket.xso_protocol = IPPROTO_TCP; 1795 xt.xt_socket.xso_len = sizeof (struct xsocket); 1796 xt.xt_socket.so_type = SOCK_STREAM; 1797 xt.xt_socket.so_state = SS_ISCONNECTING; 1798 error = SYSCTL_OUT(req, &xt, sizeof xt); 1799 if (error) { 1800 SCH_UNLOCK(sch); 1801 goto exit; 1802 } 1803 count++; 1804 } 1805 SCH_UNLOCK(sch); 1806 } 1807exit: 1808 *pcbs_exported = count; 1809 return error; 1810}
|