118 u_int hashsize;
119 u_int hashmask;
120 u_int bucket_limit;
121 u_int cache_count;
122 u_int cache_limit;
123 u_int rexmt_limit;
124 u_int hash_secret;
125 u_int next_reseed;
126 TAILQ_HEAD(, syncache) timerq[SYNCACHE_MAXREXMTS + 1];
127 struct callout tt_timerq[SYNCACHE_MAXREXMTS + 1];
128};
129static struct tcp_syncache tcp_syncache;
130
131SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache");
132
133SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RD,
134 &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache");
135
136SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RD,
137 &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache");
138
139SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD,
140 &tcp_syncache.cache_count, 0, "Current number of entries in syncache");
141
142SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RD,
143 &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable");
144
145SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
146 &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions");
147
148static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
149
150#define SYNCACHE_HASH(inc, mask) \
151 ((tcp_syncache.hash_secret ^ \
152 (inc)->inc_faddr.s_addr ^ \
153 ((inc)->inc_faddr.s_addr >> 16) ^ \
154 (inc)->inc_fport ^ (inc)->inc_lport) & mask)
155
156#define SYNCACHE_HASH6(inc, mask) \
157 ((tcp_syncache.hash_secret ^ \
158 (inc)->inc6_faddr.s6_addr32[0] ^ \
159 (inc)->inc6_faddr.s6_addr32[3] ^ \
160 (inc)->inc_fport ^ (inc)->inc_lport) & mask)
161
162#define ENDPTS_EQ(a, b) ( \
163 (a)->ie_fport == (b)->ie_fport && \
164 (a)->ie_lport == (b)->ie_lport && \
165 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \
166 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \
167)
168
169#define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
170
171#define SYNCACHE_TIMEOUT(sc, slot) do { \
172 sc->sc_rxtslot = slot; \
173 sc->sc_rxttime = ticks + TCPTV_RTOBASE * tcp_backoff[slot]; \
174 TAILQ_INSERT_TAIL(&tcp_syncache.timerq[slot], sc, sc_timerq); \
175 if (!callout_active(&tcp_syncache.tt_timerq[slot])) \
176 callout_reset(&tcp_syncache.tt_timerq[slot], \
177 TCPTV_RTOBASE * tcp_backoff[slot], \
178 syncache_timer, (void *)((intptr_t)slot)); \
179} while (0)
180
181static void
182syncache_free(struct syncache *sc)
183{
184 struct rtentry *rt;
185
186 if (sc->sc_ipopts)
187 (void) m_free(sc->sc_ipopts);
188#ifdef INET6
189 if (sc->sc_inc.inc_isipv6)
190 rt = sc->sc_route6.ro_rt;
191 else
192#endif
193 rt = sc->sc_route.ro_rt;
194 if (rt != NULL) {
195 /*
196 * If this is the only reference to a protocol cloned
197 * route, remove it immediately.
198 */
199 if (rt->rt_flags & RTF_WASCLONED &&
200 (sc->sc_flags & SCF_KEEPROUTE) == 0 &&
201 rt->rt_refcnt == 1)
202 rtrequest(RTM_DELETE, rt_key(rt),
203 rt->rt_gateway, rt_mask(rt),
204 rt->rt_flags, NULL);
205 RTFREE(rt);
206 }
207 zfree(tcp_syncache.zone, sc);
208}
209
210void
211syncache_init(void)
212{
213 int i;
214
215 tcp_syncache.cache_count = 0;
216 tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
217 tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
218 tcp_syncache.cache_limit =
219 tcp_syncache.hashsize * tcp_syncache.bucket_limit;
220 tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
221 tcp_syncache.next_reseed = 0;
222 tcp_syncache.hash_secret = arc4random();
223
224 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
225 &tcp_syncache.hashsize);
226 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
227 &tcp_syncache.cache_limit);
228 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
229 &tcp_syncache.bucket_limit);
230 if (!powerof2(tcp_syncache.hashsize)) {
231 printf("WARNING: syncache hash size is not a power of 2.\n");
232 tcp_syncache.hashsize = 512; /* safe default */
233 }
234 tcp_syncache.hashmask = tcp_syncache.hashsize - 1;
235
236 /* Allocate the hash table. */
237 MALLOC(tcp_syncache.hashbase, struct syncache_head *,
238 tcp_syncache.hashsize * sizeof(struct syncache_head),
239 M_SYNCACHE, M_WAITOK | M_ZERO);
240
241 /* Initialize the hash buckets. */
242 for (i = 0; i < tcp_syncache.hashsize; i++) {
243 TAILQ_INIT(&tcp_syncache.hashbase[i].sch_bucket);
244 tcp_syncache.hashbase[i].sch_length = 0;
245 }
246
247 /* Initialize the timer queues. */
248 for (i = 0; i <= SYNCACHE_MAXREXMTS; i++) {
249 TAILQ_INIT(&tcp_syncache.timerq[i]);
250 callout_init(&tcp_syncache.tt_timerq[i], 0);
251 }
252
253 /*
254 * Allocate the syncache entries. Allow the zone to allocate one
255 * more entry than cache limit, so a new entry can bump out an
256 * older one.
257 */
258 tcp_syncache.cache_limit -= 1;
259 tcp_syncache.zone = zinit("syncache", sizeof(struct syncache),
260 tcp_syncache.cache_limit, ZONE_INTERRUPT, 0);
261}
262
263static void
264syncache_insert(sc, sch)
265 struct syncache *sc;
266 struct syncache_head *sch;
267{
268 struct syncache *sc2;
269 int s, i;
270
271 /*
272 * Make sure that we don't overflow the per-bucket
273 * limit or the total cache size limit.
274 */
275 s = splnet();
276 if (sch->sch_length >= tcp_syncache.bucket_limit) {
277 /*
278 * The bucket is full, toss the oldest element.
279 */
280 sc2 = TAILQ_FIRST(&sch->sch_bucket);
281 sc2->sc_tp->ts_recent = ticks;
282 syncache_drop(sc2, sch);
283 tcpstat.tcps_sc_bucketoverflow++;
284 } else if (tcp_syncache.cache_count >= tcp_syncache.cache_limit) {
285 /*
286 * The cache is full. Toss the oldest entry in the
287 * entire cache. This is the front entry in the
288 * first non-empty timer queue with the largest
289 * timeout value.
290 */
291 for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
292 sc2 = TAILQ_FIRST(&tcp_syncache.timerq[i]);
293 if (sc2 != NULL)
294 break;
295 }
296 sc2->sc_tp->ts_recent = ticks;
297 syncache_drop(sc2, NULL);
298 tcpstat.tcps_sc_cacheoverflow++;
299 }
300
301 /* Initialize the entry's timer. */
302 SYNCACHE_TIMEOUT(sc, 0);
303
304 /* Put it into the bucket. */
305 TAILQ_INSERT_TAIL(&sch->sch_bucket, sc, sc_hash);
306 sch->sch_length++;
307 tcp_syncache.cache_count++;
308 tcpstat.tcps_sc_added++;
309 splx(s);
310}
311
312static void
313syncache_drop(sc, sch)
314 struct syncache *sc;
315 struct syncache_head *sch;
316{
317 int s;
318
319 if (sch == NULL) {
320#ifdef INET6
321 if (sc->sc_inc.inc_isipv6) {
322 sch = &tcp_syncache.hashbase[
323 SYNCACHE_HASH6(&sc->sc_inc, tcp_syncache.hashmask)];
324 } else
325#endif
326 {
327 sch = &tcp_syncache.hashbase[
328 SYNCACHE_HASH(&sc->sc_inc, tcp_syncache.hashmask)];
329 }
330 }
331
332 s = splnet();
333
334 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
335 sch->sch_length--;
336 tcp_syncache.cache_count--;
337
338 TAILQ_REMOVE(&tcp_syncache.timerq[sc->sc_rxtslot], sc, sc_timerq);
339 if (TAILQ_EMPTY(&tcp_syncache.timerq[sc->sc_rxtslot]))
340 callout_stop(&tcp_syncache.tt_timerq[sc->sc_rxtslot]);
341 splx(s);
342
343 syncache_free(sc);
344}
345
346/*
347 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
348 * If we have retransmitted an entry the maximum number of times, expire it.
349 */
350static void
351syncache_timer(xslot)
352 void *xslot;
353{
354 intptr_t slot = (intptr_t)xslot;
355 struct syncache *sc, *nsc;
356 struct inpcb *inp;
357 int s;
358
359 s = splnet();
360 if (callout_pending(&tcp_syncache.tt_timerq[slot]) ||
361 !callout_active(&tcp_syncache.tt_timerq[slot])) {
362 splx(s);
363 return;
364 }
365 callout_deactivate(&tcp_syncache.tt_timerq[slot]);
366
367 nsc = TAILQ_FIRST(&tcp_syncache.timerq[slot]);
368 while (nsc != NULL) {
369 if (ticks < nsc->sc_rxttime)
370 break;
371 sc = nsc;
372 nsc = TAILQ_NEXT(sc, sc_timerq);
373 inp = sc->sc_tp->t_inpcb;
374 if (slot == SYNCACHE_MAXREXMTS ||
375 slot >= tcp_syncache.rexmt_limit ||
376 inp->inp_gencnt != sc->sc_inp_gencnt) {
377 syncache_drop(sc, NULL);
378 tcpstat.tcps_sc_stale++;
379 continue;
380 }
381 (void) syncache_respond(sc, NULL);
382 tcpstat.tcps_sc_retransmitted++;
383 TAILQ_REMOVE(&tcp_syncache.timerq[slot], sc, sc_timerq);
384 SYNCACHE_TIMEOUT(sc, slot + 1);
385 }
386 if (nsc != NULL)
387 callout_reset(&tcp_syncache.tt_timerq[slot],
388 nsc->sc_rxttime - ticks, syncache_timer, (void *)(slot));
389 splx(s);
390}
391
392/*
393 * Find an entry in the syncache.
394 */
395struct syncache *
396syncache_lookup(inc, schp)
397 struct in_conninfo *inc;
398 struct syncache_head **schp;
399{
400 struct syncache *sc;
401 struct syncache_head *sch;
402 int s;
403
404#ifdef INET6
405 if (inc->inc_isipv6) {
406 sch = &tcp_syncache.hashbase[
407 SYNCACHE_HASH6(inc, tcp_syncache.hashmask)];
408 *schp = sch;
409 s = splnet();
410 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
411 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) {
412 splx(s);
413 return (sc);
414 }
415 }
416 splx(s);
417 } else
418#endif
419 {
420 sch = &tcp_syncache.hashbase[
421 SYNCACHE_HASH(inc, tcp_syncache.hashmask)];
422 *schp = sch;
423 s = splnet();
424 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
425#ifdef INET6
426 if (sc->sc_inc.inc_isipv6)
427 continue;
428#endif
429 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) {
430 splx(s);
431 return (sc);
432 }
433 }
434 splx(s);
435 }
436 return (NULL);
437}
438
439/*
440 * This function is called when we get a RST for a
441 * non-existent connection, so that we can see if the
442 * connection is in the syn cache. If it is, zap it.
443 */
444void
445syncache_chkrst(inc, th)
446 struct in_conninfo *inc;
447 struct tcphdr *th;
448{
449 struct syncache *sc;
450 struct syncache_head *sch;
451
452 sc = syncache_lookup(inc, &sch);
453 if (sc == NULL)
454 return;
455 /*
456 * If the RST bit is set, check the sequence number to see
457 * if this is a valid reset segment.
458 * RFC 793 page 37:
459 * In all states except SYN-SENT, all reset (RST) segments
460 * are validated by checking their SEQ-fields. A reset is
461 * valid if its sequence number is in the window.
462 *
463 * The sequence number in the reset segment is normally an
464 * echo of our outgoing acknowlegement numbers, but some hosts
465 * send a reset with the sequence number at the rightmost edge
466 * of our receive window, and we have to handle this case.
467 */
468 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
469 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
470 syncache_drop(sc, sch);
471 tcpstat.tcps_sc_reset++;
472 }
473}
474
475void
476syncache_badack(inc)
477 struct in_conninfo *inc;
478{
479 struct syncache *sc;
480 struct syncache_head *sch;
481
482 sc = syncache_lookup(inc, &sch);
483 if (sc != NULL) {
484 syncache_drop(sc, sch);
485 tcpstat.tcps_sc_badack++;
486 }
487}
488
489void
490syncache_unreach(inc, th)
491 struct in_conninfo *inc;
492 struct tcphdr *th;
493{
494 struct syncache *sc;
495 struct syncache_head *sch;
496
497 /* we are called at splnet() here */
498 sc = syncache_lookup(inc, &sch);
499 if (sc == NULL)
500 return;
501
502 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
503 if (ntohl(th->th_seq) != sc->sc_iss)
504 return;
505
506 /*
507 * If we've rertransmitted 3 times and this is our second error,
508 * we remove the entry. Otherwise, we allow it to continue on.
509 * This prevents us from incorrectly nuking an entry during a
510 * spurious network outage.
511 *
512 * See tcp_notify().
513 */
514 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxtslot < 3) {
515 sc->sc_flags |= SCF_UNREACH;
516 return;
517 }
518 syncache_drop(sc, sch);
519 tcpstat.tcps_sc_unreach++;
520}
521
522/*
523 * Build a new TCP socket structure from a syncache entry.
524 */
525static struct socket *
526syncache_socket(sc, lso)
527 struct syncache *sc;
528 struct socket *lso;
529{
530 struct inpcb *inp = NULL;
531 struct socket *so;
532 struct tcpcb *tp;
533
534 /*
535 * Ok, create the full blown connection, and set things up
536 * as they would have been set up if we had created the
537 * connection when the SYN arrived. If we can't create
538 * the connection, abort it.
539 */
540 so = sonewconn(lso, SS_ISCONNECTED);
541 if (so == NULL) {
542 /*
543 * Drop the connection; we will send a RST if the peer
544 * retransmits the ACK,
545 */
546 tcpstat.tcps_listendrop++;
547 goto abort;
548 }
549
550 inp = sotoinpcb(so);
551
552 /*
553 * Insert new socket into hash list.
554 */
555 inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6;
556#ifdef INET6
557 if (sc->sc_inc.inc_isipv6) {
558 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
559 } else {
560 inp->inp_vflag &= ~INP_IPV6;
561 inp->inp_vflag |= INP_IPV4;
562#endif
563 inp->inp_laddr = sc->sc_inc.inc_laddr;
564#ifdef INET6
565 }
566#endif
567 inp->inp_lport = sc->sc_inc.inc_lport;
568 if (in_pcbinshash(inp) != 0) {
569 /*
570 * Undo the assignments above if we failed to
571 * put the PCB on the hash lists.
572 */
573#ifdef INET6
574 if (sc->sc_inc.inc_isipv6)
575 inp->in6p_laddr = in6addr_any;
576 else
577#endif
578 inp->inp_laddr.s_addr = INADDR_ANY;
579 inp->inp_lport = 0;
580 goto abort;
581 }
582#ifdef IPSEC
583 /* copy old policy into new socket's */
584 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
585 printf("syncache_expand: could not copy policy\n");
586#endif
587#ifdef INET6
588 if (sc->sc_inc.inc_isipv6) {
589 struct inpcb *oinp = sotoinpcb(lso);
590 struct in6_addr laddr6;
591 struct sockaddr_in6 *sin6;
592 /*
593 * Inherit socket options from the listening socket.
594 * Note that in6p_inputopts are not (and should not be)
595 * copied, since it stores previously received options and is
596 * used to detect if each new option is different than the
597 * previous one and hence should be passed to a user.
598 * If we copied in6p_inputopts, a user would not be able to
599 * receive options just after calling the accept system call.
600 */
601 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
602 if (oinp->in6p_outputopts)
603 inp->in6p_outputopts =
604 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
605 inp->in6p_route = sc->sc_route6;
606 sc->sc_route6.ro_rt = NULL;
607
608 MALLOC(sin6, struct sockaddr_in6 *, sizeof *sin6,
609 M_SONAME, M_NOWAIT | M_ZERO);
610 if (sin6 == NULL)
611 goto abort;
612 sin6->sin6_family = AF_INET6;
613 sin6->sin6_len = sizeof(*sin6);
614 sin6->sin6_addr = sc->sc_inc.inc6_faddr;
615 sin6->sin6_port = sc->sc_inc.inc_fport;
616 laddr6 = inp->in6p_laddr;
617 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
618 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
619 if (in6_pcbconnect(inp, (struct sockaddr *)sin6, &thread0)) {
620 inp->in6p_laddr = laddr6;
621 FREE(sin6, M_SONAME);
622 goto abort;
623 }
624 FREE(sin6, M_SONAME);
625 } else
626#endif
627 {
628 struct in_addr laddr;
629 struct sockaddr_in *sin;
630
631 inp->inp_options = ip_srcroute();
632 if (inp->inp_options == NULL) {
633 inp->inp_options = sc->sc_ipopts;
634 sc->sc_ipopts = NULL;
635 }
636 inp->inp_route = sc->sc_route;
637 sc->sc_route.ro_rt = NULL;
638
639 MALLOC(sin, struct sockaddr_in *, sizeof *sin,
640 M_SONAME, M_NOWAIT | M_ZERO);
641 if (sin == NULL)
642 goto abort;
643 sin->sin_family = AF_INET;
644 sin->sin_len = sizeof(*sin);
645 sin->sin_addr = sc->sc_inc.inc_faddr;
646 sin->sin_port = sc->sc_inc.inc_fport;
647 bzero((caddr_t)sin->sin_zero, sizeof(sin->sin_zero));
648 laddr = inp->inp_laddr;
649 if (inp->inp_laddr.s_addr == INADDR_ANY)
650 inp->inp_laddr = sc->sc_inc.inc_laddr;
651 if (in_pcbconnect(inp, (struct sockaddr *)sin, &thread0)) {
652 inp->inp_laddr = laddr;
653 FREE(sin, M_SONAME);
654 goto abort;
655 }
656 FREE(sin, M_SONAME);
657 }
658
659 tp = intotcpcb(inp);
660 tp->t_state = TCPS_SYN_RECEIVED;
661 tp->iss = sc->sc_iss;
662 tp->irs = sc->sc_irs;
663 tcp_rcvseqinit(tp);
664 tcp_sendseqinit(tp);
665 tp->snd_wl1 = sc->sc_irs;
666 tp->rcv_up = sc->sc_irs + 1;
667 tp->rcv_wnd = sc->sc_wnd;
668 tp->rcv_adv += tp->rcv_wnd;
669
670 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
671 if (sc->sc_flags & SCF_NOOPT)
672 tp->t_flags |= TF_NOOPT;
673 if (sc->sc_flags & SCF_WINSCALE) {
674 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
675 tp->requested_s_scale = sc->sc_requested_s_scale;
676 tp->request_r_scale = sc->sc_request_r_scale;
677 }
678 if (sc->sc_flags & SCF_TIMESTAMP) {
679 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
680 tp->ts_recent = sc->sc_tsrecent;
681 tp->ts_recent_age = ticks;
682 }
683 if (sc->sc_flags & SCF_CC) {
684 /*
685 * Initialization of the tcpcb for transaction;
686 * set SND.WND = SEG.WND,
687 * initialize CCsend and CCrecv.
688 */
689 tp->t_flags |= TF_REQ_CC|TF_RCVD_CC;
690 tp->cc_send = sc->sc_cc_send;
691 tp->cc_recv = sc->sc_cc_recv;
692 }
693
694 tcp_mss(tp, sc->sc_peer_mss);
695
696 /*
697 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment.
698 */
699 if (sc->sc_rxtslot != 0)
700 tp->snd_cwnd = tp->t_maxseg;
701 callout_reset(tp->tt_keep, tcp_keepinit, tcp_timer_keep, tp);
702
703 tcpstat.tcps_accepts++;
704 return (so);
705
706abort:
707 if (so != NULL)
708 (void) soabort(so);
709 return (NULL);
710}
711
712/*
713 * This function gets called when we receive an ACK for a
714 * socket in the LISTEN state. We look up the connection
715 * in the syncache, and if its there, we pull it out of
716 * the cache and turn it into a full-blown connection in
717 * the SYN-RECEIVED state.
718 */
719int
720syncache_expand(inc, th, sop, m)
721 struct in_conninfo *inc;
722 struct tcphdr *th;
723 struct socket **sop;
724 struct mbuf *m;
725{
726 struct syncache *sc;
727 struct syncache_head *sch;
728 struct socket *so;
729
730 sc = syncache_lookup(inc, &sch);
731 if (sc == NULL) {
732 /*
733 * There is no syncache entry, so see if this ACK is
734 * a returning syncookie. To do this, first:
735 * A. See if this socket has had a syncache entry dropped in
736 * the past. We don't want to accept a bogus syncookie
737 * if we've never received a SYN.
738 * B. check that the syncookie is valid. If it is, then
739 * cobble up a fake syncache entry, and return.
740 */
741 if (!tcp_syncookies)
742 return (0);
743 sc = syncookie_lookup(inc, th, *sop);
744 if (sc == NULL)
745 return (0);
746 sch = NULL;
747 tcpstat.tcps_sc_recvcookie++;
748 }
749
750 /*
751 * If seg contains an ACK, but not for our SYN/ACK, send a RST.
752 */
753 if (th->th_ack != sc->sc_iss + 1)
754 return (0);
755
756 so = syncache_socket(sc, *sop);
757 if (so == NULL) {
758#if 0
759resetandabort:
760 /* XXXjlemon check this - is this correct? */
761 (void) tcp_respond(NULL, m, m, th,
762 th->th_seq + tlen, (tcp_seq)0, TH_RST|TH_ACK);
763#endif
764 m_freem(m); /* XXX only needed for above */
765 tcpstat.tcps_sc_aborted++;
766 } else {
767 sc->sc_flags |= SCF_KEEPROUTE;
768 tcpstat.tcps_sc_completed++;
769 }
770 if (sch == NULL)
771 syncache_free(sc);
772 else
773 syncache_drop(sc, sch);
774 *sop = so;
775 return (1);
776}
777
778/*
779 * Given a LISTEN socket and an inbound SYN request, add
780 * this to the syn cache, and send back a segment:
781 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
782 * to the source.
783 *
784 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
785 * Doing so would require that we hold onto the data and deliver it
786 * to the application. However, if we are the target of a SYN-flood
787 * DoS attack, an attacker could send data which would eventually
788 * consume all available buffer space if it were ACKed. By not ACKing
789 * the data, we avoid this DoS scenario.
790 */
791int
792syncache_add(inc, to, th, sop, m)
793 struct in_conninfo *inc;
794 struct tcpopt *to;
795 struct tcphdr *th;
796 struct socket **sop;
797 struct mbuf *m;
798{
799 struct tcpcb *tp;
800 struct socket *so;
801 struct syncache *sc = NULL;
802 struct syncache_head *sch;
803 struct mbuf *ipopts = NULL;
804 struct rmxp_tao *taop;
805 int i, s, win;
806
807 so = *sop;
808 tp = sototcpcb(so);
809
810 /*
811 * Remember the IP options, if any.
812 */
813#ifdef INET6
814 if (!inc->inc_isipv6)
815#endif
816 ipopts = ip_srcroute();
817
818 /*
819 * See if we already have an entry for this connection.
820 * If we do, resend the SYN,ACK, and reset the retransmit timer.
821 *
822 * XXX
823 * should the syncache be re-initialized with the contents
824 * of the new SYN here (which may have different options?)
825 */
826 sc = syncache_lookup(inc, &sch);
827 if (sc != NULL) {
828 tcpstat.tcps_sc_dupsyn++;
829 if (ipopts) {
830 /*
831 * If we were remembering a previous source route,
832 * forget it and use the new one we've been given.
833 */
834 if (sc->sc_ipopts)
835 (void) m_free(sc->sc_ipopts);
836 sc->sc_ipopts = ipopts;
837 }
838 /*
839 * Update timestamp if present.
840 */
841 if (sc->sc_flags & SCF_TIMESTAMP)
842 sc->sc_tsrecent = to->to_tsval;
843 /*
844 * PCB may have changed, pick up new values.
845 */
846 sc->sc_tp = tp;
847 sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
848 if (syncache_respond(sc, m) == 0) {
849 s = splnet();
850 TAILQ_REMOVE(&tcp_syncache.timerq[sc->sc_rxtslot],
851 sc, sc_timerq);
852 SYNCACHE_TIMEOUT(sc, sc->sc_rxtslot);
853 splx(s);
854 tcpstat.tcps_sndacks++;
855 tcpstat.tcps_sndtotal++;
856 }
857 *sop = NULL;
858 return (1);
859 }
860
861 sc = zalloc(tcp_syncache.zone);
862 if (sc == NULL) {
863 /*
864 * The zone allocator couldn't provide more entries.
865 * Treat this as if the cache was full; drop the oldest
866 * entry and insert the new one.
867 */
868 s = splnet();
869 for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
870 sc = TAILQ_FIRST(&tcp_syncache.timerq[i]);
871 if (sc != NULL)
872 break;
873 }
874 sc->sc_tp->ts_recent = ticks;
875 syncache_drop(sc, NULL);
876 splx(s);
877 tcpstat.tcps_sc_zonefail++;
878 sc = zalloc(tcp_syncache.zone);
879 if (sc == NULL) {
880 if (ipopts)
881 (void) m_free(ipopts);
882 return (0);
883 }
884 }
885
886 /*
887 * Fill in the syncache values.
888 */
889 bzero(sc, sizeof(*sc));
890 sc->sc_tp = tp;
891 sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
892 sc->sc_ipopts = ipopts;
893 sc->sc_inc.inc_fport = inc->inc_fport;
894 sc->sc_inc.inc_lport = inc->inc_lport;
895#ifdef INET6
896 sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
897 if (inc->inc_isipv6) {
898 sc->sc_inc.inc6_faddr = inc->inc6_faddr;
899 sc->sc_inc.inc6_laddr = inc->inc6_laddr;
900 sc->sc_route6.ro_rt = NULL;
901 } else
902#endif
903 {
904 sc->sc_inc.inc_faddr = inc->inc_faddr;
905 sc->sc_inc.inc_laddr = inc->inc_laddr;
906 sc->sc_route.ro_rt = NULL;
907 }
908 sc->sc_irs = th->th_seq;
909 if (tcp_syncookies)
910 sc->sc_iss = syncookie_generate(sc);
911 else
912 sc->sc_iss = arc4random();
913
914 /* Initial receive window: clip sbspace to [0 .. TCP_MAXWIN] */
915 win = sbspace(&so->so_rcv);
916 win = imax(win, 0);
917 win = imin(win, TCP_MAXWIN);
918 sc->sc_wnd = win;
919
920 sc->sc_flags = 0;
921 sc->sc_peer_mss = to->to_flags & TOF_MSS ? to->to_mss : 0;
922 if (tcp_do_rfc1323) {
923 /*
924 * A timestamp received in a SYN makes
925 * it ok to send timestamp requests and replies.
926 */
927 if (to->to_flags & TOF_TS) {
928 sc->sc_tsrecent = to->to_tsval;
929 sc->sc_flags |= SCF_TIMESTAMP;
930 }
931 if (to->to_flags & TOF_SCALE) {
932 int wscale = 0;
933
934 /* Compute proper scaling value from buffer space */
935 while (wscale < TCP_MAX_WINSHIFT &&
936 (TCP_MAXWIN << wscale) < so->so_rcv.sb_hiwat)
937 wscale++;
938 sc->sc_request_r_scale = wscale;
939 sc->sc_requested_s_scale = to->to_requested_s_scale;
940 sc->sc_flags |= SCF_WINSCALE;
941 }
942 }
943 if (tcp_do_rfc1644) {
944 /*
945 * A CC or CC.new option received in a SYN makes
946 * it ok to send CC in subsequent segments.
947 */
948 if (to->to_flags & (TOF_CC|TOF_CCNEW)) {
949 sc->sc_cc_recv = to->to_cc;
950 sc->sc_cc_send = CC_INC(tcp_ccgen);
951 sc->sc_flags |= SCF_CC;
952 }
953 }
954 if (tp->t_flags & TF_NOOPT)
955 sc->sc_flags = SCF_NOOPT;
956
957 /*
958 * XXX
959 * We have the option here of not doing TAO (even if the segment
960 * qualifies) and instead fall back to a normal 3WHS via the syncache.
961 * This allows us to apply synflood protection to TAO-qualifying SYNs
962 * also. However, there should be a hueristic to determine when to
963 * do this, and is not present at the moment.
964 */
965
966 /*
967 * Perform TAO test on incoming CC (SEG.CC) option, if any.
968 * - compare SEG.CC against cached CC from the same host, if any.
969 * - if SEG.CC > chached value, SYN must be new and is accepted
970 * immediately: save new CC in the cache, mark the socket
971 * connected, enter ESTABLISHED state, turn on flag to
972 * send a SYN in the next segment.
973 * A virtual advertised window is set in rcv_adv to
974 * initialize SWS prevention. Then enter normal segment
975 * processing: drop SYN, process data and FIN.
976 * - otherwise do a normal 3-way handshake.
977 */
978 taop = tcp_gettaocache(&sc->sc_inc);
979 if ((to->to_flags & TOF_CC) != 0) {
980 if (((tp->t_flags & TF_NOPUSH) != 0) &&
981 sc->sc_flags & SCF_CC &&
982 taop != NULL && taop->tao_cc != 0 &&
983 CC_GT(to->to_cc, taop->tao_cc)) {
984 sc->sc_rxtslot = 0;
985 so = syncache_socket(sc, *sop);
986 if (so != NULL) {
987 sc->sc_flags |= SCF_KEEPROUTE;
988 taop->tao_cc = to->to_cc;
989 *sop = so;
990 }
991 syncache_free(sc);
992 return (so != NULL);
993 }
994 } else {
995 /*
996 * No CC option, but maybe CC.NEW: invalidate cached value.
997 */
998 if (taop != NULL)
999 taop->tao_cc = 0;
1000 }
1001 /*
1002 * TAO test failed or there was no CC option,
1003 * do a standard 3-way handshake.
1004 */
1005 if (syncache_respond(sc, m) == 0) {
1006 syncache_insert(sc, sch);
1007 tcpstat.tcps_sndacks++;
1008 tcpstat.tcps_sndtotal++;
1009 } else {
1010 syncache_free(sc);
1011 tcpstat.tcps_sc_dropped++;
1012 }
1013 *sop = NULL;
1014 return (1);
1015}
1016
1017static int
1018syncache_respond(sc, m)
1019 struct syncache *sc;
1020 struct mbuf *m;
1021{
1022 u_int8_t *optp;
1023 int optlen, error;
1024 u_int16_t tlen, hlen, mssopt;
1025 struct ip *ip = NULL;
1026 struct rtentry *rt;
1027 struct tcphdr *th;
1028#ifdef INET6
1029 struct ip6_hdr *ip6 = NULL;
1030#endif
1031
1032#ifdef INET6
1033 if (sc->sc_inc.inc_isipv6) {
1034 rt = tcp_rtlookup6(&sc->sc_inc);
1035 if (rt != NULL)
1036 mssopt = rt->rt_ifp->if_mtu -
1037 (sizeof(struct ip6_hdr) + sizeof(struct tcphdr));
1038 else
1039 mssopt = tcp_v6mssdflt;
1040 hlen = sizeof(struct ip6_hdr);
1041 } else
1042#endif
1043 {
1044 rt = tcp_rtlookup(&sc->sc_inc);
1045 if (rt != NULL)
1046 mssopt = rt->rt_ifp->if_mtu -
1047 (sizeof(struct ip) + sizeof(struct tcphdr));
1048 else
1049 mssopt = tcp_mssdflt;
1050 hlen = sizeof(struct ip);
1051 }
1052
1053 /* Compute the size of the TCP options. */
1054 if (sc->sc_flags & SCF_NOOPT) {
1055 optlen = 0;
1056 } else {
1057 optlen = TCPOLEN_MAXSEG +
1058 ((sc->sc_flags & SCF_WINSCALE) ? 4 : 0) +
1059 ((sc->sc_flags & SCF_TIMESTAMP) ? TCPOLEN_TSTAMP_APPA : 0) +
1060 ((sc->sc_flags & SCF_CC) ? TCPOLEN_CC_APPA * 2 : 0);
1061 }
1062 tlen = hlen + sizeof(struct tcphdr) + optlen;
1063
1064 /*
1065 * XXX
1066 * assume that the entire packet will fit in a header mbuf
1067 */
1068 KASSERT(max_linkhdr + tlen <= MHLEN, ("syncache: mbuf too small"));
1069
1070 /*
1071 * XXX shouldn't this reuse the mbuf if possible ?
1072 * Create the IP+TCP header from scratch.
1073 */
1074 if (m)
1075 m_freem(m);
1076
1077 m = m_gethdr(M_DONTWAIT, MT_HEADER);
1078 if (m == NULL)
1079 return (ENOBUFS);
1080 m->m_data += max_linkhdr;
1081 m->m_len = tlen;
1082 m->m_pkthdr.len = tlen;
1083 m->m_pkthdr.rcvif = NULL;
1084
1085#ifdef IPSEC
1086 /* use IPsec policy on listening socket to send SYN,ACK */
1087 if (ipsec_setsocket(m, sc->sc_tp->t_inpcb->inp_socket) != 0) {
1088 m_freem(m);
1089 return (ENOBUFS);
1090 }
1091#endif
1092
1093#ifdef INET6
1094 if (sc->sc_inc.inc_isipv6) {
1095 ip6 = mtod(m, struct ip6_hdr *);
1096 ip6->ip6_vfc = IPV6_VERSION;
1097 ip6->ip6_nxt = IPPROTO_TCP;
1098 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1099 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1100 ip6->ip6_plen = htons(tlen - hlen);
1101 /* ip6_hlim is set after checksum */
1102 /* ip6_flow = ??? */
1103
1104 th = (struct tcphdr *)(ip6 + 1);
1105 } else
1106#endif
1107 {
1108 ip = mtod(m, struct ip *);
1109 ip->ip_v = IPVERSION;
1110 ip->ip_hl = sizeof(struct ip) >> 2;
1111 ip->ip_tos = 0;
1112 ip->ip_len = tlen;
1113 ip->ip_id = 0;
1114 ip->ip_off = 0;
1115 ip->ip_ttl = ip_defttl;
1116 ip->ip_sum = 0;
1117 ip->ip_p = IPPROTO_TCP;
1118 ip->ip_src = sc->sc_inc.inc_laddr;
1119 ip->ip_dst = sc->sc_inc.inc_faddr;
1120
1121 th = (struct tcphdr *)(ip + 1);
1122 }
1123 th->th_sport = sc->sc_inc.inc_lport;
1124 th->th_dport = sc->sc_inc.inc_fport;
1125
1126 th->th_seq = htonl(sc->sc_iss);
1127 th->th_ack = htonl(sc->sc_irs + 1);
1128 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1129 th->th_x2 = 0;
1130 th->th_flags = TH_SYN|TH_ACK;
1131 th->th_win = htons(sc->sc_wnd);
1132 th->th_urp = 0;
1133
1134 /* Tack on the TCP options. */
1135 if (optlen == 0)
1136 goto no_options;
1137 optp = (u_int8_t *)(th + 1);
1138 *optp++ = TCPOPT_MAXSEG;
1139 *optp++ = TCPOLEN_MAXSEG;
1140 *optp++ = (mssopt >> 8) & 0xff;
1141 *optp++ = mssopt & 0xff;
1142
1143 if (sc->sc_flags & SCF_WINSCALE) {
1144 *((u_int32_t *)optp) = htonl(TCPOPT_NOP << 24 |
1145 TCPOPT_WINDOW << 16 | TCPOLEN_WINDOW << 8 |
1146 sc->sc_request_r_scale);
1147 optp += 4;
1148 }
1149
1150 if (sc->sc_flags & SCF_TIMESTAMP) {
1151 u_int32_t *lp = (u_int32_t *)(optp);
1152
1153 /* Form timestamp option as shown in appendix A of RFC 1323. */
1154 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
1155 *lp++ = htonl(ticks);
1156 *lp = htonl(sc->sc_tsrecent);
1157 optp += TCPOLEN_TSTAMP_APPA;
1158 }
1159
1160 /*
1161 * Send CC and CC.echo if we received CC from our peer.
1162 */
1163 if (sc->sc_flags & SCF_CC) {
1164 u_int32_t *lp = (u_int32_t *)(optp);
1165
1166 *lp++ = htonl(TCPOPT_CC_HDR(TCPOPT_CC));
1167 *lp++ = htonl(sc->sc_cc_send);
1168 *lp++ = htonl(TCPOPT_CC_HDR(TCPOPT_CCECHO));
1169 *lp = htonl(sc->sc_cc_recv);
1170 optp += TCPOLEN_CC_APPA * 2;
1171 }
1172no_options:
1173
1174#ifdef INET6
1175 if (sc->sc_inc.inc_isipv6) {
1176 struct route_in6 *ro6 = &sc->sc_route6;
1177
1178 th->th_sum = 0;
1179 th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, tlen - hlen);
1180 ip6->ip6_hlim = in6_selecthlim(NULL,
1181 ro6->ro_rt ? ro6->ro_rt->rt_ifp : NULL);
1182 error = ip6_output(m, NULL, ro6, 0, NULL, NULL);
1183 } else
1184#endif
1185 {
1186 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1187 htons(tlen - hlen + IPPROTO_TCP));
1188 m->m_pkthdr.csum_flags = CSUM_TCP;
1189 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1190 error = ip_output(m, sc->sc_ipopts, &sc->sc_route, 0, NULL);
1191 }
1192 return (error);
1193}
1194
1195/*
1196 * cookie layers:
1197 *
1198 * |. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .|
1199 * | peer iss |
1200 * | MD5(laddr,faddr,lport,fport,secret) |. . . . . . .|
1201 * | 0 |(A)| |
1202 * (A): peer mss index
1203 */
1204
1205/*
1206 * The values below are chosen to minimize the size of the tcp_secret
1207 * table, as well as providing roughly a 4 second lifetime for the cookie.
1208 */
1209
1210#define SYNCOOKIE_HASHSHIFT 2 /* log2(# of 32bit words from hash) */
1211#define SYNCOOKIE_WNDBITS 7 /* exposed bits for window indexing */
1212#define SYNCOOKIE_TIMESHIFT 5 /* scale ticks to window time units */
1213
1214#define SYNCOOKIE_HASHMASK ((1 << SYNCOOKIE_HASHSHIFT) - 1)
1215#define SYNCOOKIE_WNDMASK ((1 << SYNCOOKIE_WNDBITS) - 1)
1216#define SYNCOOKIE_NSECRETS (1 << (SYNCOOKIE_WNDBITS - SYNCOOKIE_HASHSHIFT))
1217#define SYNCOOKIE_TIMEOUT \
1218 (hz * (1 << SYNCOOKIE_WNDBITS) / (1 << SYNCOOKIE_TIMESHIFT))
1219#define SYNCOOKIE_DATAMASK ((3 << SYNCOOKIE_WNDBITS) | SYNCOOKIE_WNDMASK)
1220
1221static struct {
1222 u_int32_t ts_secbits;
1223 u_int ts_expire;
1224} tcp_secret[SYNCOOKIE_NSECRETS];
1225
1226static int tcp_msstab[] = { 0, 536, 1460, 8960 };
1227
1228static MD5_CTX syn_ctx;
1229
1230#define MD5Add(v) MD5Update(&syn_ctx, (u_char *)&v, sizeof(v))
1231
1232/*
1233 * Consider the problem of a recreated (and retransmitted) cookie. If the
1234 * original SYN was accepted, the connection is established. The second
1235 * SYN is inflight, and if it arrives with an ISN that falls within the
1236 * receive window, the connection is killed.
1237 *
1238 * However, since cookies have other problems, this may not be worth
1239 * worrying about.
1240 */
1241
1242static u_int32_t
1243syncookie_generate(struct syncache *sc)
1244{
1245 u_int32_t md5_buffer[4];
1246 u_int32_t data;
1247 int wnd, idx;
1248
1249 wnd = ((ticks << SYNCOOKIE_TIMESHIFT) / hz) & SYNCOOKIE_WNDMASK;
1250 idx = wnd >> SYNCOOKIE_HASHSHIFT;
1251 if (tcp_secret[idx].ts_expire < ticks) {
1252 tcp_secret[idx].ts_secbits = arc4random();
1253 tcp_secret[idx].ts_expire = ticks + SYNCOOKIE_TIMEOUT;
1254 }
1255 for (data = sizeof(tcp_msstab) / sizeof(int) - 1; data > 0; data--)
1256 if (tcp_msstab[data] <= sc->sc_peer_mss)
1257 break;
1258 data = (data << SYNCOOKIE_WNDBITS) | wnd;
1259 data ^= sc->sc_irs; /* peer's iss */
1260 MD5Init(&syn_ctx);
1261#ifdef INET6
1262 if (sc->sc_inc.inc_isipv6) {
1263 MD5Add(sc->sc_inc.inc6_laddr);
1264 MD5Add(sc->sc_inc.inc6_faddr);
1265 } else
1266#endif
1267 {
1268 MD5Add(sc->sc_inc.inc_laddr);
1269 MD5Add(sc->sc_inc.inc_faddr);
1270 }
1271 MD5Add(sc->sc_inc.inc_lport);
1272 MD5Add(sc->sc_inc.inc_fport);
1273 MD5Add(tcp_secret[idx].ts_secbits);
1274 MD5Final((u_char *)&md5_buffer, &syn_ctx);
1275 data ^= (md5_buffer[wnd & SYNCOOKIE_HASHMASK] & ~SYNCOOKIE_WNDMASK);
1276 return (data);
1277}
1278
1279static struct syncache *
1280syncookie_lookup(inc, th, so)
1281 struct in_conninfo *inc;
1282 struct tcphdr *th;
1283 struct socket *so;
1284{
1285 u_int32_t md5_buffer[4];
1286 struct syncache *sc;
1287 u_int32_t data;
1288 int wnd, idx;
1289
1290 data = (th->th_ack - 1) ^ (th->th_seq - 1); /* remove ISS */
1291 wnd = data & SYNCOOKIE_WNDMASK;
1292 idx = wnd >> SYNCOOKIE_HASHSHIFT;
1293 if (tcp_secret[idx].ts_expire < ticks ||
1294 sototcpcb(so)->ts_recent + SYNCOOKIE_TIMEOUT < ticks)
1295 return (NULL);
1296 MD5Init(&syn_ctx);
1297#ifdef INET6
1298 if (inc->inc_isipv6) {
1299 MD5Add(inc->inc6_laddr);
1300 MD5Add(inc->inc6_faddr);
1301 } else
1302#endif
1303 {
1304 MD5Add(inc->inc_laddr);
1305 MD5Add(inc->inc_faddr);
1306 }
1307 MD5Add(inc->inc_lport);
1308 MD5Add(inc->inc_fport);
1309 MD5Add(tcp_secret[idx].ts_secbits);
1310 MD5Final((u_char *)&md5_buffer, &syn_ctx);
1311 data ^= md5_buffer[wnd & SYNCOOKIE_HASHMASK];
1312 if ((data & ~SYNCOOKIE_DATAMASK) != 0)
1313 return (NULL);
1314 data = data >> SYNCOOKIE_WNDBITS;
1315
1316 sc = zalloc(tcp_syncache.zone);
1317 if (sc == NULL)
1318 return (NULL);
1319 /*
1320 * Fill in the syncache values.
1321 * XXX duplicate code from syncache_add
1322 */
1323 sc->sc_ipopts = NULL;
1324 sc->sc_inc.inc_fport = inc->inc_fport;
1325 sc->sc_inc.inc_lport = inc->inc_lport;
1326#ifdef INET6
1327 sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1328 if (inc->inc_isipv6) {
1329 sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1330 sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1331 sc->sc_route6.ro_rt = NULL;
1332 } else
1333#endif
1334 {
1335 sc->sc_inc.inc_faddr = inc->inc_faddr;
1336 sc->sc_inc.inc_laddr = inc->inc_laddr;
1337 sc->sc_route.ro_rt = NULL;
1338 }
1339 sc->sc_irs = th->th_seq - 1;
1340 sc->sc_iss = th->th_ack - 1;
1341 wnd = sbspace(&so->so_rcv);
1342 wnd = imax(wnd, 0);
1343 wnd = imin(wnd, TCP_MAXWIN);
1344 sc->sc_wnd = wnd;
1345 sc->sc_flags = 0;
1346 sc->sc_rxtslot = 0;
1347 sc->sc_peer_mss = tcp_msstab[data];
1348 return (sc);
1349}
| 118 u_int hashsize;
119 u_int hashmask;
120 u_int bucket_limit;
121 u_int cache_count;
122 u_int cache_limit;
123 u_int rexmt_limit;
124 u_int hash_secret;
125 u_int next_reseed;
126 TAILQ_HEAD(, syncache) timerq[SYNCACHE_MAXREXMTS + 1];
127 struct callout tt_timerq[SYNCACHE_MAXREXMTS + 1];
128};
129static struct tcp_syncache tcp_syncache;
130
131SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache");
132
133SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RD,
134 &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache");
135
136SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RD,
137 &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache");
138
139SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD,
140 &tcp_syncache.cache_count, 0, "Current number of entries in syncache");
141
142SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RD,
143 &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable");
144
145SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
146 &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions");
147
148static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
149
150#define SYNCACHE_HASH(inc, mask) \
151 ((tcp_syncache.hash_secret ^ \
152 (inc)->inc_faddr.s_addr ^ \
153 ((inc)->inc_faddr.s_addr >> 16) ^ \
154 (inc)->inc_fport ^ (inc)->inc_lport) & mask)
155
156#define SYNCACHE_HASH6(inc, mask) \
157 ((tcp_syncache.hash_secret ^ \
158 (inc)->inc6_faddr.s6_addr32[0] ^ \
159 (inc)->inc6_faddr.s6_addr32[3] ^ \
160 (inc)->inc_fport ^ (inc)->inc_lport) & mask)
161
162#define ENDPTS_EQ(a, b) ( \
163 (a)->ie_fport == (b)->ie_fport && \
164 (a)->ie_lport == (b)->ie_lport && \
165 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \
166 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \
167)
168
169#define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
170
171#define SYNCACHE_TIMEOUT(sc, slot) do { \
172 sc->sc_rxtslot = slot; \
173 sc->sc_rxttime = ticks + TCPTV_RTOBASE * tcp_backoff[slot]; \
174 TAILQ_INSERT_TAIL(&tcp_syncache.timerq[slot], sc, sc_timerq); \
175 if (!callout_active(&tcp_syncache.tt_timerq[slot])) \
176 callout_reset(&tcp_syncache.tt_timerq[slot], \
177 TCPTV_RTOBASE * tcp_backoff[slot], \
178 syncache_timer, (void *)((intptr_t)slot)); \
179} while (0)
180
181static void
182syncache_free(struct syncache *sc)
183{
184 struct rtentry *rt;
185
186 if (sc->sc_ipopts)
187 (void) m_free(sc->sc_ipopts);
188#ifdef INET6
189 if (sc->sc_inc.inc_isipv6)
190 rt = sc->sc_route6.ro_rt;
191 else
192#endif
193 rt = sc->sc_route.ro_rt;
194 if (rt != NULL) {
195 /*
196 * If this is the only reference to a protocol cloned
197 * route, remove it immediately.
198 */
199 if (rt->rt_flags & RTF_WASCLONED &&
200 (sc->sc_flags & SCF_KEEPROUTE) == 0 &&
201 rt->rt_refcnt == 1)
202 rtrequest(RTM_DELETE, rt_key(rt),
203 rt->rt_gateway, rt_mask(rt),
204 rt->rt_flags, NULL);
205 RTFREE(rt);
206 }
207 zfree(tcp_syncache.zone, sc);
208}
209
210void
211syncache_init(void)
212{
213 int i;
214
215 tcp_syncache.cache_count = 0;
216 tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
217 tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
218 tcp_syncache.cache_limit =
219 tcp_syncache.hashsize * tcp_syncache.bucket_limit;
220 tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
221 tcp_syncache.next_reseed = 0;
222 tcp_syncache.hash_secret = arc4random();
223
224 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
225 &tcp_syncache.hashsize);
226 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
227 &tcp_syncache.cache_limit);
228 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
229 &tcp_syncache.bucket_limit);
230 if (!powerof2(tcp_syncache.hashsize)) {
231 printf("WARNING: syncache hash size is not a power of 2.\n");
232 tcp_syncache.hashsize = 512; /* safe default */
233 }
234 tcp_syncache.hashmask = tcp_syncache.hashsize - 1;
235
236 /* Allocate the hash table. */
237 MALLOC(tcp_syncache.hashbase, struct syncache_head *,
238 tcp_syncache.hashsize * sizeof(struct syncache_head),
239 M_SYNCACHE, M_WAITOK | M_ZERO);
240
241 /* Initialize the hash buckets. */
242 for (i = 0; i < tcp_syncache.hashsize; i++) {
243 TAILQ_INIT(&tcp_syncache.hashbase[i].sch_bucket);
244 tcp_syncache.hashbase[i].sch_length = 0;
245 }
246
247 /* Initialize the timer queues. */
248 for (i = 0; i <= SYNCACHE_MAXREXMTS; i++) {
249 TAILQ_INIT(&tcp_syncache.timerq[i]);
250 callout_init(&tcp_syncache.tt_timerq[i], 0);
251 }
252
253 /*
254 * Allocate the syncache entries. Allow the zone to allocate one
255 * more entry than cache limit, so a new entry can bump out an
256 * older one.
257 */
258 tcp_syncache.cache_limit -= 1;
259 tcp_syncache.zone = zinit("syncache", sizeof(struct syncache),
260 tcp_syncache.cache_limit, ZONE_INTERRUPT, 0);
261}
262
263static void
264syncache_insert(sc, sch)
265 struct syncache *sc;
266 struct syncache_head *sch;
267{
268 struct syncache *sc2;
269 int s, i;
270
271 /*
272 * Make sure that we don't overflow the per-bucket
273 * limit or the total cache size limit.
274 */
275 s = splnet();
276 if (sch->sch_length >= tcp_syncache.bucket_limit) {
277 /*
278 * The bucket is full, toss the oldest element.
279 */
280 sc2 = TAILQ_FIRST(&sch->sch_bucket);
281 sc2->sc_tp->ts_recent = ticks;
282 syncache_drop(sc2, sch);
283 tcpstat.tcps_sc_bucketoverflow++;
284 } else if (tcp_syncache.cache_count >= tcp_syncache.cache_limit) {
285 /*
286 * The cache is full. Toss the oldest entry in the
287 * entire cache. This is the front entry in the
288 * first non-empty timer queue with the largest
289 * timeout value.
290 */
291 for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
292 sc2 = TAILQ_FIRST(&tcp_syncache.timerq[i]);
293 if (sc2 != NULL)
294 break;
295 }
296 sc2->sc_tp->ts_recent = ticks;
297 syncache_drop(sc2, NULL);
298 tcpstat.tcps_sc_cacheoverflow++;
299 }
300
301 /* Initialize the entry's timer. */
302 SYNCACHE_TIMEOUT(sc, 0);
303
304 /* Put it into the bucket. */
305 TAILQ_INSERT_TAIL(&sch->sch_bucket, sc, sc_hash);
306 sch->sch_length++;
307 tcp_syncache.cache_count++;
308 tcpstat.tcps_sc_added++;
309 splx(s);
310}
311
312static void
313syncache_drop(sc, sch)
314 struct syncache *sc;
315 struct syncache_head *sch;
316{
317 int s;
318
319 if (sch == NULL) {
320#ifdef INET6
321 if (sc->sc_inc.inc_isipv6) {
322 sch = &tcp_syncache.hashbase[
323 SYNCACHE_HASH6(&sc->sc_inc, tcp_syncache.hashmask)];
324 } else
325#endif
326 {
327 sch = &tcp_syncache.hashbase[
328 SYNCACHE_HASH(&sc->sc_inc, tcp_syncache.hashmask)];
329 }
330 }
331
332 s = splnet();
333
334 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
335 sch->sch_length--;
336 tcp_syncache.cache_count--;
337
338 TAILQ_REMOVE(&tcp_syncache.timerq[sc->sc_rxtslot], sc, sc_timerq);
339 if (TAILQ_EMPTY(&tcp_syncache.timerq[sc->sc_rxtslot]))
340 callout_stop(&tcp_syncache.tt_timerq[sc->sc_rxtslot]);
341 splx(s);
342
343 syncache_free(sc);
344}
345
346/*
347 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
348 * If we have retransmitted an entry the maximum number of times, expire it.
349 */
350static void
351syncache_timer(xslot)
352 void *xslot;
353{
354 intptr_t slot = (intptr_t)xslot;
355 struct syncache *sc, *nsc;
356 struct inpcb *inp;
357 int s;
358
359 s = splnet();
360 if (callout_pending(&tcp_syncache.tt_timerq[slot]) ||
361 !callout_active(&tcp_syncache.tt_timerq[slot])) {
362 splx(s);
363 return;
364 }
365 callout_deactivate(&tcp_syncache.tt_timerq[slot]);
366
367 nsc = TAILQ_FIRST(&tcp_syncache.timerq[slot]);
368 while (nsc != NULL) {
369 if (ticks < nsc->sc_rxttime)
370 break;
371 sc = nsc;
372 nsc = TAILQ_NEXT(sc, sc_timerq);
373 inp = sc->sc_tp->t_inpcb;
374 if (slot == SYNCACHE_MAXREXMTS ||
375 slot >= tcp_syncache.rexmt_limit ||
376 inp->inp_gencnt != sc->sc_inp_gencnt) {
377 syncache_drop(sc, NULL);
378 tcpstat.tcps_sc_stale++;
379 continue;
380 }
381 (void) syncache_respond(sc, NULL);
382 tcpstat.tcps_sc_retransmitted++;
383 TAILQ_REMOVE(&tcp_syncache.timerq[slot], sc, sc_timerq);
384 SYNCACHE_TIMEOUT(sc, slot + 1);
385 }
386 if (nsc != NULL)
387 callout_reset(&tcp_syncache.tt_timerq[slot],
388 nsc->sc_rxttime - ticks, syncache_timer, (void *)(slot));
389 splx(s);
390}
391
392/*
393 * Find an entry in the syncache.
394 */
395struct syncache *
396syncache_lookup(inc, schp)
397 struct in_conninfo *inc;
398 struct syncache_head **schp;
399{
400 struct syncache *sc;
401 struct syncache_head *sch;
402 int s;
403
404#ifdef INET6
405 if (inc->inc_isipv6) {
406 sch = &tcp_syncache.hashbase[
407 SYNCACHE_HASH6(inc, tcp_syncache.hashmask)];
408 *schp = sch;
409 s = splnet();
410 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
411 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) {
412 splx(s);
413 return (sc);
414 }
415 }
416 splx(s);
417 } else
418#endif
419 {
420 sch = &tcp_syncache.hashbase[
421 SYNCACHE_HASH(inc, tcp_syncache.hashmask)];
422 *schp = sch;
423 s = splnet();
424 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
425#ifdef INET6
426 if (sc->sc_inc.inc_isipv6)
427 continue;
428#endif
429 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) {
430 splx(s);
431 return (sc);
432 }
433 }
434 splx(s);
435 }
436 return (NULL);
437}
438
439/*
440 * This function is called when we get a RST for a
441 * non-existent connection, so that we can see if the
442 * connection is in the syn cache. If it is, zap it.
443 */
444void
445syncache_chkrst(inc, th)
446 struct in_conninfo *inc;
447 struct tcphdr *th;
448{
449 struct syncache *sc;
450 struct syncache_head *sch;
451
452 sc = syncache_lookup(inc, &sch);
453 if (sc == NULL)
454 return;
455 /*
456 * If the RST bit is set, check the sequence number to see
457 * if this is a valid reset segment.
458 * RFC 793 page 37:
459 * In all states except SYN-SENT, all reset (RST) segments
460 * are validated by checking their SEQ-fields. A reset is
461 * valid if its sequence number is in the window.
462 *
463 * The sequence number in the reset segment is normally an
464 * echo of our outgoing acknowlegement numbers, but some hosts
465 * send a reset with the sequence number at the rightmost edge
466 * of our receive window, and we have to handle this case.
467 */
468 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
469 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
470 syncache_drop(sc, sch);
471 tcpstat.tcps_sc_reset++;
472 }
473}
474
475void
476syncache_badack(inc)
477 struct in_conninfo *inc;
478{
479 struct syncache *sc;
480 struct syncache_head *sch;
481
482 sc = syncache_lookup(inc, &sch);
483 if (sc != NULL) {
484 syncache_drop(sc, sch);
485 tcpstat.tcps_sc_badack++;
486 }
487}
488
489void
490syncache_unreach(inc, th)
491 struct in_conninfo *inc;
492 struct tcphdr *th;
493{
494 struct syncache *sc;
495 struct syncache_head *sch;
496
497 /* we are called at splnet() here */
498 sc = syncache_lookup(inc, &sch);
499 if (sc == NULL)
500 return;
501
502 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
503 if (ntohl(th->th_seq) != sc->sc_iss)
504 return;
505
506 /*
507 * If we've rertransmitted 3 times and this is our second error,
508 * we remove the entry. Otherwise, we allow it to continue on.
509 * This prevents us from incorrectly nuking an entry during a
510 * spurious network outage.
511 *
512 * See tcp_notify().
513 */
514 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxtslot < 3) {
515 sc->sc_flags |= SCF_UNREACH;
516 return;
517 }
518 syncache_drop(sc, sch);
519 tcpstat.tcps_sc_unreach++;
520}
521
522/*
523 * Build a new TCP socket structure from a syncache entry.
524 */
525static struct socket *
526syncache_socket(sc, lso)
527 struct syncache *sc;
528 struct socket *lso;
529{
530 struct inpcb *inp = NULL;
531 struct socket *so;
532 struct tcpcb *tp;
533
534 /*
535 * Ok, create the full blown connection, and set things up
536 * as they would have been set up if we had created the
537 * connection when the SYN arrived. If we can't create
538 * the connection, abort it.
539 */
540 so = sonewconn(lso, SS_ISCONNECTED);
541 if (so == NULL) {
542 /*
543 * Drop the connection; we will send a RST if the peer
544 * retransmits the ACK,
545 */
546 tcpstat.tcps_listendrop++;
547 goto abort;
548 }
549
550 inp = sotoinpcb(so);
551
552 /*
553 * Insert new socket into hash list.
554 */
555 inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6;
556#ifdef INET6
557 if (sc->sc_inc.inc_isipv6) {
558 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
559 } else {
560 inp->inp_vflag &= ~INP_IPV6;
561 inp->inp_vflag |= INP_IPV4;
562#endif
563 inp->inp_laddr = sc->sc_inc.inc_laddr;
564#ifdef INET6
565 }
566#endif
567 inp->inp_lport = sc->sc_inc.inc_lport;
568 if (in_pcbinshash(inp) != 0) {
569 /*
570 * Undo the assignments above if we failed to
571 * put the PCB on the hash lists.
572 */
573#ifdef INET6
574 if (sc->sc_inc.inc_isipv6)
575 inp->in6p_laddr = in6addr_any;
576 else
577#endif
578 inp->inp_laddr.s_addr = INADDR_ANY;
579 inp->inp_lport = 0;
580 goto abort;
581 }
582#ifdef IPSEC
583 /* copy old policy into new socket's */
584 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
585 printf("syncache_expand: could not copy policy\n");
586#endif
587#ifdef INET6
588 if (sc->sc_inc.inc_isipv6) {
589 struct inpcb *oinp = sotoinpcb(lso);
590 struct in6_addr laddr6;
591 struct sockaddr_in6 *sin6;
592 /*
593 * Inherit socket options from the listening socket.
594 * Note that in6p_inputopts are not (and should not be)
595 * copied, since it stores previously received options and is
596 * used to detect if each new option is different than the
597 * previous one and hence should be passed to a user.
598 * If we copied in6p_inputopts, a user would not be able to
599 * receive options just after calling the accept system call.
600 */
601 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
602 if (oinp->in6p_outputopts)
603 inp->in6p_outputopts =
604 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
605 inp->in6p_route = sc->sc_route6;
606 sc->sc_route6.ro_rt = NULL;
607
608 MALLOC(sin6, struct sockaddr_in6 *, sizeof *sin6,
609 M_SONAME, M_NOWAIT | M_ZERO);
610 if (sin6 == NULL)
611 goto abort;
612 sin6->sin6_family = AF_INET6;
613 sin6->sin6_len = sizeof(*sin6);
614 sin6->sin6_addr = sc->sc_inc.inc6_faddr;
615 sin6->sin6_port = sc->sc_inc.inc_fport;
616 laddr6 = inp->in6p_laddr;
617 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
618 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
619 if (in6_pcbconnect(inp, (struct sockaddr *)sin6, &thread0)) {
620 inp->in6p_laddr = laddr6;
621 FREE(sin6, M_SONAME);
622 goto abort;
623 }
624 FREE(sin6, M_SONAME);
625 } else
626#endif
627 {
628 struct in_addr laddr;
629 struct sockaddr_in *sin;
630
631 inp->inp_options = ip_srcroute();
632 if (inp->inp_options == NULL) {
633 inp->inp_options = sc->sc_ipopts;
634 sc->sc_ipopts = NULL;
635 }
636 inp->inp_route = sc->sc_route;
637 sc->sc_route.ro_rt = NULL;
638
639 MALLOC(sin, struct sockaddr_in *, sizeof *sin,
640 M_SONAME, M_NOWAIT | M_ZERO);
641 if (sin == NULL)
642 goto abort;
643 sin->sin_family = AF_INET;
644 sin->sin_len = sizeof(*sin);
645 sin->sin_addr = sc->sc_inc.inc_faddr;
646 sin->sin_port = sc->sc_inc.inc_fport;
647 bzero((caddr_t)sin->sin_zero, sizeof(sin->sin_zero));
648 laddr = inp->inp_laddr;
649 if (inp->inp_laddr.s_addr == INADDR_ANY)
650 inp->inp_laddr = sc->sc_inc.inc_laddr;
651 if (in_pcbconnect(inp, (struct sockaddr *)sin, &thread0)) {
652 inp->inp_laddr = laddr;
653 FREE(sin, M_SONAME);
654 goto abort;
655 }
656 FREE(sin, M_SONAME);
657 }
658
659 tp = intotcpcb(inp);
660 tp->t_state = TCPS_SYN_RECEIVED;
661 tp->iss = sc->sc_iss;
662 tp->irs = sc->sc_irs;
663 tcp_rcvseqinit(tp);
664 tcp_sendseqinit(tp);
665 tp->snd_wl1 = sc->sc_irs;
666 tp->rcv_up = sc->sc_irs + 1;
667 tp->rcv_wnd = sc->sc_wnd;
668 tp->rcv_adv += tp->rcv_wnd;
669
670 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
671 if (sc->sc_flags & SCF_NOOPT)
672 tp->t_flags |= TF_NOOPT;
673 if (sc->sc_flags & SCF_WINSCALE) {
674 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
675 tp->requested_s_scale = sc->sc_requested_s_scale;
676 tp->request_r_scale = sc->sc_request_r_scale;
677 }
678 if (sc->sc_flags & SCF_TIMESTAMP) {
679 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
680 tp->ts_recent = sc->sc_tsrecent;
681 tp->ts_recent_age = ticks;
682 }
683 if (sc->sc_flags & SCF_CC) {
684 /*
685 * Initialization of the tcpcb for transaction;
686 * set SND.WND = SEG.WND,
687 * initialize CCsend and CCrecv.
688 */
689 tp->t_flags |= TF_REQ_CC|TF_RCVD_CC;
690 tp->cc_send = sc->sc_cc_send;
691 tp->cc_recv = sc->sc_cc_recv;
692 }
693
694 tcp_mss(tp, sc->sc_peer_mss);
695
696 /*
697 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment.
698 */
699 if (sc->sc_rxtslot != 0)
700 tp->snd_cwnd = tp->t_maxseg;
701 callout_reset(tp->tt_keep, tcp_keepinit, tcp_timer_keep, tp);
702
703 tcpstat.tcps_accepts++;
704 return (so);
705
706abort:
707 if (so != NULL)
708 (void) soabort(so);
709 return (NULL);
710}
711
712/*
713 * This function gets called when we receive an ACK for a
714 * socket in the LISTEN state. We look up the connection
715 * in the syncache, and if its there, we pull it out of
716 * the cache and turn it into a full-blown connection in
717 * the SYN-RECEIVED state.
718 */
719int
720syncache_expand(inc, th, sop, m)
721 struct in_conninfo *inc;
722 struct tcphdr *th;
723 struct socket **sop;
724 struct mbuf *m;
725{
726 struct syncache *sc;
727 struct syncache_head *sch;
728 struct socket *so;
729
730 sc = syncache_lookup(inc, &sch);
731 if (sc == NULL) {
732 /*
733 * There is no syncache entry, so see if this ACK is
734 * a returning syncookie. To do this, first:
735 * A. See if this socket has had a syncache entry dropped in
736 * the past. We don't want to accept a bogus syncookie
737 * if we've never received a SYN.
738 * B. check that the syncookie is valid. If it is, then
739 * cobble up a fake syncache entry, and return.
740 */
741 if (!tcp_syncookies)
742 return (0);
743 sc = syncookie_lookup(inc, th, *sop);
744 if (sc == NULL)
745 return (0);
746 sch = NULL;
747 tcpstat.tcps_sc_recvcookie++;
748 }
749
750 /*
751 * If seg contains an ACK, but not for our SYN/ACK, send a RST.
752 */
753 if (th->th_ack != sc->sc_iss + 1)
754 return (0);
755
756 so = syncache_socket(sc, *sop);
757 if (so == NULL) {
758#if 0
759resetandabort:
760 /* XXXjlemon check this - is this correct? */
761 (void) tcp_respond(NULL, m, m, th,
762 th->th_seq + tlen, (tcp_seq)0, TH_RST|TH_ACK);
763#endif
764 m_freem(m); /* XXX only needed for above */
765 tcpstat.tcps_sc_aborted++;
766 } else {
767 sc->sc_flags |= SCF_KEEPROUTE;
768 tcpstat.tcps_sc_completed++;
769 }
770 if (sch == NULL)
771 syncache_free(sc);
772 else
773 syncache_drop(sc, sch);
774 *sop = so;
775 return (1);
776}
777
778/*
779 * Given a LISTEN socket and an inbound SYN request, add
780 * this to the syn cache, and send back a segment:
781 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
782 * to the source.
783 *
784 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
785 * Doing so would require that we hold onto the data and deliver it
786 * to the application. However, if we are the target of a SYN-flood
787 * DoS attack, an attacker could send data which would eventually
788 * consume all available buffer space if it were ACKed. By not ACKing
789 * the data, we avoid this DoS scenario.
790 */
791int
792syncache_add(inc, to, th, sop, m)
793 struct in_conninfo *inc;
794 struct tcpopt *to;
795 struct tcphdr *th;
796 struct socket **sop;
797 struct mbuf *m;
798{
799 struct tcpcb *tp;
800 struct socket *so;
801 struct syncache *sc = NULL;
802 struct syncache_head *sch;
803 struct mbuf *ipopts = NULL;
804 struct rmxp_tao *taop;
805 int i, s, win;
806
807 so = *sop;
808 tp = sototcpcb(so);
809
810 /*
811 * Remember the IP options, if any.
812 */
813#ifdef INET6
814 if (!inc->inc_isipv6)
815#endif
816 ipopts = ip_srcroute();
817
818 /*
819 * See if we already have an entry for this connection.
820 * If we do, resend the SYN,ACK, and reset the retransmit timer.
821 *
822 * XXX
823 * should the syncache be re-initialized with the contents
824 * of the new SYN here (which may have different options?)
825 */
826 sc = syncache_lookup(inc, &sch);
827 if (sc != NULL) {
828 tcpstat.tcps_sc_dupsyn++;
829 if (ipopts) {
830 /*
831 * If we were remembering a previous source route,
832 * forget it and use the new one we've been given.
833 */
834 if (sc->sc_ipopts)
835 (void) m_free(sc->sc_ipopts);
836 sc->sc_ipopts = ipopts;
837 }
838 /*
839 * Update timestamp if present.
840 */
841 if (sc->sc_flags & SCF_TIMESTAMP)
842 sc->sc_tsrecent = to->to_tsval;
843 /*
844 * PCB may have changed, pick up new values.
845 */
846 sc->sc_tp = tp;
847 sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
848 if (syncache_respond(sc, m) == 0) {
849 s = splnet();
850 TAILQ_REMOVE(&tcp_syncache.timerq[sc->sc_rxtslot],
851 sc, sc_timerq);
852 SYNCACHE_TIMEOUT(sc, sc->sc_rxtslot);
853 splx(s);
854 tcpstat.tcps_sndacks++;
855 tcpstat.tcps_sndtotal++;
856 }
857 *sop = NULL;
858 return (1);
859 }
860
861 sc = zalloc(tcp_syncache.zone);
862 if (sc == NULL) {
863 /*
864 * The zone allocator couldn't provide more entries.
865 * Treat this as if the cache was full; drop the oldest
866 * entry and insert the new one.
867 */
868 s = splnet();
869 for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
870 sc = TAILQ_FIRST(&tcp_syncache.timerq[i]);
871 if (sc != NULL)
872 break;
873 }
874 sc->sc_tp->ts_recent = ticks;
875 syncache_drop(sc, NULL);
876 splx(s);
877 tcpstat.tcps_sc_zonefail++;
878 sc = zalloc(tcp_syncache.zone);
879 if (sc == NULL) {
880 if (ipopts)
881 (void) m_free(ipopts);
882 return (0);
883 }
884 }
885
886 /*
887 * Fill in the syncache values.
888 */
889 bzero(sc, sizeof(*sc));
890 sc->sc_tp = tp;
891 sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
892 sc->sc_ipopts = ipopts;
893 sc->sc_inc.inc_fport = inc->inc_fport;
894 sc->sc_inc.inc_lport = inc->inc_lport;
895#ifdef INET6
896 sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
897 if (inc->inc_isipv6) {
898 sc->sc_inc.inc6_faddr = inc->inc6_faddr;
899 sc->sc_inc.inc6_laddr = inc->inc6_laddr;
900 sc->sc_route6.ro_rt = NULL;
901 } else
902#endif
903 {
904 sc->sc_inc.inc_faddr = inc->inc_faddr;
905 sc->sc_inc.inc_laddr = inc->inc_laddr;
906 sc->sc_route.ro_rt = NULL;
907 }
908 sc->sc_irs = th->th_seq;
909 if (tcp_syncookies)
910 sc->sc_iss = syncookie_generate(sc);
911 else
912 sc->sc_iss = arc4random();
913
914 /* Initial receive window: clip sbspace to [0 .. TCP_MAXWIN] */
915 win = sbspace(&so->so_rcv);
916 win = imax(win, 0);
917 win = imin(win, TCP_MAXWIN);
918 sc->sc_wnd = win;
919
920 sc->sc_flags = 0;
921 sc->sc_peer_mss = to->to_flags & TOF_MSS ? to->to_mss : 0;
922 if (tcp_do_rfc1323) {
923 /*
924 * A timestamp received in a SYN makes
925 * it ok to send timestamp requests and replies.
926 */
927 if (to->to_flags & TOF_TS) {
928 sc->sc_tsrecent = to->to_tsval;
929 sc->sc_flags |= SCF_TIMESTAMP;
930 }
931 if (to->to_flags & TOF_SCALE) {
932 int wscale = 0;
933
934 /* Compute proper scaling value from buffer space */
935 while (wscale < TCP_MAX_WINSHIFT &&
936 (TCP_MAXWIN << wscale) < so->so_rcv.sb_hiwat)
937 wscale++;
938 sc->sc_request_r_scale = wscale;
939 sc->sc_requested_s_scale = to->to_requested_s_scale;
940 sc->sc_flags |= SCF_WINSCALE;
941 }
942 }
943 if (tcp_do_rfc1644) {
944 /*
945 * A CC or CC.new option received in a SYN makes
946 * it ok to send CC in subsequent segments.
947 */
948 if (to->to_flags & (TOF_CC|TOF_CCNEW)) {
949 sc->sc_cc_recv = to->to_cc;
950 sc->sc_cc_send = CC_INC(tcp_ccgen);
951 sc->sc_flags |= SCF_CC;
952 }
953 }
954 if (tp->t_flags & TF_NOOPT)
955 sc->sc_flags = SCF_NOOPT;
956
957 /*
958 * XXX
959 * We have the option here of not doing TAO (even if the segment
960 * qualifies) and instead fall back to a normal 3WHS via the syncache.
961 * This allows us to apply synflood protection to TAO-qualifying SYNs
962 * also. However, there should be a hueristic to determine when to
963 * do this, and is not present at the moment.
964 */
965
966 /*
967 * Perform TAO test on incoming CC (SEG.CC) option, if any.
968 * - compare SEG.CC against cached CC from the same host, if any.
969 * - if SEG.CC > chached value, SYN must be new and is accepted
970 * immediately: save new CC in the cache, mark the socket
971 * connected, enter ESTABLISHED state, turn on flag to
972 * send a SYN in the next segment.
973 * A virtual advertised window is set in rcv_adv to
974 * initialize SWS prevention. Then enter normal segment
975 * processing: drop SYN, process data and FIN.
976 * - otherwise do a normal 3-way handshake.
977 */
978 taop = tcp_gettaocache(&sc->sc_inc);
979 if ((to->to_flags & TOF_CC) != 0) {
980 if (((tp->t_flags & TF_NOPUSH) != 0) &&
981 sc->sc_flags & SCF_CC &&
982 taop != NULL && taop->tao_cc != 0 &&
983 CC_GT(to->to_cc, taop->tao_cc)) {
984 sc->sc_rxtslot = 0;
985 so = syncache_socket(sc, *sop);
986 if (so != NULL) {
987 sc->sc_flags |= SCF_KEEPROUTE;
988 taop->tao_cc = to->to_cc;
989 *sop = so;
990 }
991 syncache_free(sc);
992 return (so != NULL);
993 }
994 } else {
995 /*
996 * No CC option, but maybe CC.NEW: invalidate cached value.
997 */
998 if (taop != NULL)
999 taop->tao_cc = 0;
1000 }
1001 /*
1002 * TAO test failed or there was no CC option,
1003 * do a standard 3-way handshake.
1004 */
1005 if (syncache_respond(sc, m) == 0) {
1006 syncache_insert(sc, sch);
1007 tcpstat.tcps_sndacks++;
1008 tcpstat.tcps_sndtotal++;
1009 } else {
1010 syncache_free(sc);
1011 tcpstat.tcps_sc_dropped++;
1012 }
1013 *sop = NULL;
1014 return (1);
1015}
1016
1017static int
1018syncache_respond(sc, m)
1019 struct syncache *sc;
1020 struct mbuf *m;
1021{
1022 u_int8_t *optp;
1023 int optlen, error;
1024 u_int16_t tlen, hlen, mssopt;
1025 struct ip *ip = NULL;
1026 struct rtentry *rt;
1027 struct tcphdr *th;
1028#ifdef INET6
1029 struct ip6_hdr *ip6 = NULL;
1030#endif
1031
1032#ifdef INET6
1033 if (sc->sc_inc.inc_isipv6) {
1034 rt = tcp_rtlookup6(&sc->sc_inc);
1035 if (rt != NULL)
1036 mssopt = rt->rt_ifp->if_mtu -
1037 (sizeof(struct ip6_hdr) + sizeof(struct tcphdr));
1038 else
1039 mssopt = tcp_v6mssdflt;
1040 hlen = sizeof(struct ip6_hdr);
1041 } else
1042#endif
1043 {
1044 rt = tcp_rtlookup(&sc->sc_inc);
1045 if (rt != NULL)
1046 mssopt = rt->rt_ifp->if_mtu -
1047 (sizeof(struct ip) + sizeof(struct tcphdr));
1048 else
1049 mssopt = tcp_mssdflt;
1050 hlen = sizeof(struct ip);
1051 }
1052
1053 /* Compute the size of the TCP options. */
1054 if (sc->sc_flags & SCF_NOOPT) {
1055 optlen = 0;
1056 } else {
1057 optlen = TCPOLEN_MAXSEG +
1058 ((sc->sc_flags & SCF_WINSCALE) ? 4 : 0) +
1059 ((sc->sc_flags & SCF_TIMESTAMP) ? TCPOLEN_TSTAMP_APPA : 0) +
1060 ((sc->sc_flags & SCF_CC) ? TCPOLEN_CC_APPA * 2 : 0);
1061 }
1062 tlen = hlen + sizeof(struct tcphdr) + optlen;
1063
1064 /*
1065 * XXX
1066 * assume that the entire packet will fit in a header mbuf
1067 */
1068 KASSERT(max_linkhdr + tlen <= MHLEN, ("syncache: mbuf too small"));
1069
1070 /*
1071 * XXX shouldn't this reuse the mbuf if possible ?
1072 * Create the IP+TCP header from scratch.
1073 */
1074 if (m)
1075 m_freem(m);
1076
1077 m = m_gethdr(M_DONTWAIT, MT_HEADER);
1078 if (m == NULL)
1079 return (ENOBUFS);
1080 m->m_data += max_linkhdr;
1081 m->m_len = tlen;
1082 m->m_pkthdr.len = tlen;
1083 m->m_pkthdr.rcvif = NULL;
1084
1085#ifdef IPSEC
1086 /* use IPsec policy on listening socket to send SYN,ACK */
1087 if (ipsec_setsocket(m, sc->sc_tp->t_inpcb->inp_socket) != 0) {
1088 m_freem(m);
1089 return (ENOBUFS);
1090 }
1091#endif
1092
1093#ifdef INET6
1094 if (sc->sc_inc.inc_isipv6) {
1095 ip6 = mtod(m, struct ip6_hdr *);
1096 ip6->ip6_vfc = IPV6_VERSION;
1097 ip6->ip6_nxt = IPPROTO_TCP;
1098 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1099 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1100 ip6->ip6_plen = htons(tlen - hlen);
1101 /* ip6_hlim is set after checksum */
1102 /* ip6_flow = ??? */
1103
1104 th = (struct tcphdr *)(ip6 + 1);
1105 } else
1106#endif
1107 {
1108 ip = mtod(m, struct ip *);
1109 ip->ip_v = IPVERSION;
1110 ip->ip_hl = sizeof(struct ip) >> 2;
1111 ip->ip_tos = 0;
1112 ip->ip_len = tlen;
1113 ip->ip_id = 0;
1114 ip->ip_off = 0;
1115 ip->ip_ttl = ip_defttl;
1116 ip->ip_sum = 0;
1117 ip->ip_p = IPPROTO_TCP;
1118 ip->ip_src = sc->sc_inc.inc_laddr;
1119 ip->ip_dst = sc->sc_inc.inc_faddr;
1120
1121 th = (struct tcphdr *)(ip + 1);
1122 }
1123 th->th_sport = sc->sc_inc.inc_lport;
1124 th->th_dport = sc->sc_inc.inc_fport;
1125
1126 th->th_seq = htonl(sc->sc_iss);
1127 th->th_ack = htonl(sc->sc_irs + 1);
1128 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1129 th->th_x2 = 0;
1130 th->th_flags = TH_SYN|TH_ACK;
1131 th->th_win = htons(sc->sc_wnd);
1132 th->th_urp = 0;
1133
1134 /* Tack on the TCP options. */
1135 if (optlen == 0)
1136 goto no_options;
1137 optp = (u_int8_t *)(th + 1);
1138 *optp++ = TCPOPT_MAXSEG;
1139 *optp++ = TCPOLEN_MAXSEG;
1140 *optp++ = (mssopt >> 8) & 0xff;
1141 *optp++ = mssopt & 0xff;
1142
1143 if (sc->sc_flags & SCF_WINSCALE) {
1144 *((u_int32_t *)optp) = htonl(TCPOPT_NOP << 24 |
1145 TCPOPT_WINDOW << 16 | TCPOLEN_WINDOW << 8 |
1146 sc->sc_request_r_scale);
1147 optp += 4;
1148 }
1149
1150 if (sc->sc_flags & SCF_TIMESTAMP) {
1151 u_int32_t *lp = (u_int32_t *)(optp);
1152
1153 /* Form timestamp option as shown in appendix A of RFC 1323. */
1154 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
1155 *lp++ = htonl(ticks);
1156 *lp = htonl(sc->sc_tsrecent);
1157 optp += TCPOLEN_TSTAMP_APPA;
1158 }
1159
1160 /*
1161 * Send CC and CC.echo if we received CC from our peer.
1162 */
1163 if (sc->sc_flags & SCF_CC) {
1164 u_int32_t *lp = (u_int32_t *)(optp);
1165
1166 *lp++ = htonl(TCPOPT_CC_HDR(TCPOPT_CC));
1167 *lp++ = htonl(sc->sc_cc_send);
1168 *lp++ = htonl(TCPOPT_CC_HDR(TCPOPT_CCECHO));
1169 *lp = htonl(sc->sc_cc_recv);
1170 optp += TCPOLEN_CC_APPA * 2;
1171 }
1172no_options:
1173
1174#ifdef INET6
1175 if (sc->sc_inc.inc_isipv6) {
1176 struct route_in6 *ro6 = &sc->sc_route6;
1177
1178 th->th_sum = 0;
1179 th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, tlen - hlen);
1180 ip6->ip6_hlim = in6_selecthlim(NULL,
1181 ro6->ro_rt ? ro6->ro_rt->rt_ifp : NULL);
1182 error = ip6_output(m, NULL, ro6, 0, NULL, NULL);
1183 } else
1184#endif
1185 {
1186 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1187 htons(tlen - hlen + IPPROTO_TCP));
1188 m->m_pkthdr.csum_flags = CSUM_TCP;
1189 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1190 error = ip_output(m, sc->sc_ipopts, &sc->sc_route, 0, NULL);
1191 }
1192 return (error);
1193}
1194
1195/*
1196 * cookie layers:
1197 *
1198 * |. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .|
1199 * | peer iss |
1200 * | MD5(laddr,faddr,lport,fport,secret) |. . . . . . .|
1201 * | 0 |(A)| |
1202 * (A): peer mss index
1203 */
1204
1205/*
1206 * The values below are chosen to minimize the size of the tcp_secret
1207 * table, as well as providing roughly a 4 second lifetime for the cookie.
1208 */
1209
1210#define SYNCOOKIE_HASHSHIFT 2 /* log2(# of 32bit words from hash) */
1211#define SYNCOOKIE_WNDBITS 7 /* exposed bits for window indexing */
1212#define SYNCOOKIE_TIMESHIFT 5 /* scale ticks to window time units */
1213
1214#define SYNCOOKIE_HASHMASK ((1 << SYNCOOKIE_HASHSHIFT) - 1)
1215#define SYNCOOKIE_WNDMASK ((1 << SYNCOOKIE_WNDBITS) - 1)
1216#define SYNCOOKIE_NSECRETS (1 << (SYNCOOKIE_WNDBITS - SYNCOOKIE_HASHSHIFT))
1217#define SYNCOOKIE_TIMEOUT \
1218 (hz * (1 << SYNCOOKIE_WNDBITS) / (1 << SYNCOOKIE_TIMESHIFT))
1219#define SYNCOOKIE_DATAMASK ((3 << SYNCOOKIE_WNDBITS) | SYNCOOKIE_WNDMASK)
1220
1221static struct {
1222 u_int32_t ts_secbits;
1223 u_int ts_expire;
1224} tcp_secret[SYNCOOKIE_NSECRETS];
1225
1226static int tcp_msstab[] = { 0, 536, 1460, 8960 };
1227
1228static MD5_CTX syn_ctx;
1229
1230#define MD5Add(v) MD5Update(&syn_ctx, (u_char *)&v, sizeof(v))
1231
1232/*
1233 * Consider the problem of a recreated (and retransmitted) cookie. If the
1234 * original SYN was accepted, the connection is established. The second
1235 * SYN is inflight, and if it arrives with an ISN that falls within the
1236 * receive window, the connection is killed.
1237 *
1238 * However, since cookies have other problems, this may not be worth
1239 * worrying about.
1240 */
1241
1242static u_int32_t
1243syncookie_generate(struct syncache *sc)
1244{
1245 u_int32_t md5_buffer[4];
1246 u_int32_t data;
1247 int wnd, idx;
1248
1249 wnd = ((ticks << SYNCOOKIE_TIMESHIFT) / hz) & SYNCOOKIE_WNDMASK;
1250 idx = wnd >> SYNCOOKIE_HASHSHIFT;
1251 if (tcp_secret[idx].ts_expire < ticks) {
1252 tcp_secret[idx].ts_secbits = arc4random();
1253 tcp_secret[idx].ts_expire = ticks + SYNCOOKIE_TIMEOUT;
1254 }
1255 for (data = sizeof(tcp_msstab) / sizeof(int) - 1; data > 0; data--)
1256 if (tcp_msstab[data] <= sc->sc_peer_mss)
1257 break;
1258 data = (data << SYNCOOKIE_WNDBITS) | wnd;
1259 data ^= sc->sc_irs; /* peer's iss */
1260 MD5Init(&syn_ctx);
1261#ifdef INET6
1262 if (sc->sc_inc.inc_isipv6) {
1263 MD5Add(sc->sc_inc.inc6_laddr);
1264 MD5Add(sc->sc_inc.inc6_faddr);
1265 } else
1266#endif
1267 {
1268 MD5Add(sc->sc_inc.inc_laddr);
1269 MD5Add(sc->sc_inc.inc_faddr);
1270 }
1271 MD5Add(sc->sc_inc.inc_lport);
1272 MD5Add(sc->sc_inc.inc_fport);
1273 MD5Add(tcp_secret[idx].ts_secbits);
1274 MD5Final((u_char *)&md5_buffer, &syn_ctx);
1275 data ^= (md5_buffer[wnd & SYNCOOKIE_HASHMASK] & ~SYNCOOKIE_WNDMASK);
1276 return (data);
1277}
1278
1279static struct syncache *
1280syncookie_lookup(inc, th, so)
1281 struct in_conninfo *inc;
1282 struct tcphdr *th;
1283 struct socket *so;
1284{
1285 u_int32_t md5_buffer[4];
1286 struct syncache *sc;
1287 u_int32_t data;
1288 int wnd, idx;
1289
1290 data = (th->th_ack - 1) ^ (th->th_seq - 1); /* remove ISS */
1291 wnd = data & SYNCOOKIE_WNDMASK;
1292 idx = wnd >> SYNCOOKIE_HASHSHIFT;
1293 if (tcp_secret[idx].ts_expire < ticks ||
1294 sototcpcb(so)->ts_recent + SYNCOOKIE_TIMEOUT < ticks)
1295 return (NULL);
1296 MD5Init(&syn_ctx);
1297#ifdef INET6
1298 if (inc->inc_isipv6) {
1299 MD5Add(inc->inc6_laddr);
1300 MD5Add(inc->inc6_faddr);
1301 } else
1302#endif
1303 {
1304 MD5Add(inc->inc_laddr);
1305 MD5Add(inc->inc_faddr);
1306 }
1307 MD5Add(inc->inc_lport);
1308 MD5Add(inc->inc_fport);
1309 MD5Add(tcp_secret[idx].ts_secbits);
1310 MD5Final((u_char *)&md5_buffer, &syn_ctx);
1311 data ^= md5_buffer[wnd & SYNCOOKIE_HASHMASK];
1312 if ((data & ~SYNCOOKIE_DATAMASK) != 0)
1313 return (NULL);
1314 data = data >> SYNCOOKIE_WNDBITS;
1315
1316 sc = zalloc(tcp_syncache.zone);
1317 if (sc == NULL)
1318 return (NULL);
1319 /*
1320 * Fill in the syncache values.
1321 * XXX duplicate code from syncache_add
1322 */
1323 sc->sc_ipopts = NULL;
1324 sc->sc_inc.inc_fport = inc->inc_fport;
1325 sc->sc_inc.inc_lport = inc->inc_lport;
1326#ifdef INET6
1327 sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1328 if (inc->inc_isipv6) {
1329 sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1330 sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1331 sc->sc_route6.ro_rt = NULL;
1332 } else
1333#endif
1334 {
1335 sc->sc_inc.inc_faddr = inc->inc_faddr;
1336 sc->sc_inc.inc_laddr = inc->inc_laddr;
1337 sc->sc_route.ro_rt = NULL;
1338 }
1339 sc->sc_irs = th->th_seq - 1;
1340 sc->sc_iss = th->th_ack - 1;
1341 wnd = sbspace(&so->so_rcv);
1342 wnd = imax(wnd, 0);
1343 wnd = imin(wnd, TCP_MAXWIN);
1344 sc->sc_wnd = wnd;
1345 sc->sc_flags = 0;
1346 sc->sc_rxtslot = 0;
1347 sc->sc_peer_mss = tcp_msstab[data];
1348 return (sc);
1349}
|