xref: /trueos/sys/netinet/tcp_syncache.c (revision 8943816bb4812ac55b5f3738b955ac07db05a3b2)
1 /*-
2  * Copyright (c) 2001 McAfee, Inc.
3  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
4  * All rights reserved.
5  *
6  * This software was developed for the FreeBSD Project by Jonathan Lemon
7  * and McAfee Research, the Security Research Division of McAfee, Inc. under
8  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
9  * DARPA CHATS research program. [2001 McAfee, Inc.]
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_pcbgroup.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/sysctl.h>
45 #include <sys/limits.h>
46 #include <sys/lock.h>
47 #include <sys/mutex.h>
48 #include <sys/malloc.h>
49 #include <sys/mbuf.h>
50 #include <sys/proc.h>		/* for proc0 declaration */
51 #include <sys/random.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syslog.h>
55 #include <sys/ucred.h>
56 
57 #include <sys/md5.h>
58 #include <crypto/siphash/siphash.h>
59 
60 #include <vm/uma.h>
61 
62 #include <net/if.h>
63 #include <net/route.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_systm.h>
68 #include <netinet/ip.h>
69 #include <netinet/in_var.h>
70 #include <netinet/in_pcb.h>
71 #include <netinet/ip_var.h>
72 #include <netinet/ip_options.h>
73 #ifdef INET6
74 #include <netinet/ip6.h>
75 #include <netinet/icmp6.h>
76 #include <netinet6/nd6.h>
77 #include <netinet6/ip6_var.h>
78 #include <netinet6/in6_pcb.h>
79 #endif
80 #include <netinet/tcp.h>
81 #include <netinet/tcp_fsm.h>
82 #include <netinet/tcp_seq.h>
83 #include <netinet/tcp_timer.h>
84 #include <netinet/tcp_var.h>
85 #include <netinet/tcp_syncache.h>
86 #ifdef INET6
87 #include <netinet6/tcp6_var.h>
88 #endif
89 #ifdef TCP_OFFLOAD
90 #include <netinet/toecore.h>
91 #endif
92 
93 #ifdef IPSEC
94 #include <netipsec/ipsec.h>
95 #ifdef INET6
96 #include <netipsec/ipsec6.h>
97 #endif
98 #include <netipsec/key.h>
99 #endif /*IPSEC*/
100 
101 #include <machine/in_cksum.h>
102 
103 #include <security/mac/mac_framework.h>
104 
105 static VNET_DEFINE(int, tcp_syncookies) = 1;
106 #define	V_tcp_syncookies		VNET(tcp_syncookies)
107 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
108     &VNET_NAME(tcp_syncookies), 0,
109     "Use TCP SYN cookies if the syncache overflows");
110 
111 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
112 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
113 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_RW,
114     &VNET_NAME(tcp_syncookiesonly), 0,
115     "Use only TCP SYN cookies");
116 
117 #ifdef TCP_OFFLOAD
118 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
119 #endif
120 
121 static void	 syncache_drop(struct syncache *, struct syncache_head *);
122 static void	 syncache_free(struct syncache *);
123 static void	 syncache_insert(struct syncache *, struct syncache_head *);
124 static int	 syncache_respond(struct syncache *);
125 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
126 		    struct mbuf *m);
127 static int	 syncache_sysctl_count(SYSCTL_HANDLER_ARGS);
128 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
129 		    int docallout);
130 static void	 syncache_timer(void *);
131 
132 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
133 		    uint8_t *, uintptr_t);
134 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
135 static struct syncache
136 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
137 		    struct syncache *, struct tcphdr *, struct tcpopt *,
138 		    struct socket *);
139 static void	 syncookie_reseed(void *);
140 #ifdef INVARIANTS
141 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
142 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
143 		    struct socket *lso);
144 #endif
145 
146 /*
147  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
148  * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
149  * the odds are that the user has given up attempting to connect by then.
150  */
151 #define SYNCACHE_MAXREXMTS		3
152 
153 /* Arbitrary values */
154 #define TCP_SYNCACHE_HASHSIZE		512
155 #define TCP_SYNCACHE_BUCKETLIMIT	30
156 
157 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
158 #define	V_tcp_syncache			VNET(tcp_syncache)
159 
160 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
161     "TCP SYN cache");
162 
163 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN,
164     &VNET_NAME(tcp_syncache.bucket_limit), 0,
165     "Per-bucket hash limit for syncache");
166 
167 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN,
168     &VNET_NAME(tcp_syncache.cache_limit), 0,
169     "Overall entry limit for syncache");
170 
171 SYSCTL_VNET_PROC(_net_inet_tcp_syncache, OID_AUTO, count, (CTLTYPE_UINT|CTLFLAG_RD),
172     NULL, 0, &syncache_sysctl_count, "IU",
173     "Current number of entries in syncache");
174 
175 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN,
176     &VNET_NAME(tcp_syncache.hashsize), 0,
177     "Size of TCP syncache hashtable");
178 
179 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
180     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
181     "Limit on SYN/ACK retransmissions");
182 
183 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
184 SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
185     CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
186     "Send reset on socket allocation failure");
187 
188 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
189 
190 #define SYNCACHE_HASH(inc, mask)					\
191 	((V_tcp_syncache.hash_secret ^					\
192 	  (inc)->inc_faddr.s_addr ^					\
193 	  ((inc)->inc_faddr.s_addr >> 16) ^				\
194 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
195 
196 #define SYNCACHE_HASH6(inc, mask)					\
197 	((V_tcp_syncache.hash_secret ^					\
198 	  (inc)->inc6_faddr.s6_addr32[0] ^				\
199 	  (inc)->inc6_faddr.s6_addr32[3] ^				\
200 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
201 
202 #define ENDPTS_EQ(a, b) (						\
203 	(a)->ie_fport == (b)->ie_fport &&				\
204 	(a)->ie_lport == (b)->ie_lport &&				\
205 	(a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr &&			\
206 	(a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr			\
207 )
208 
209 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
210 
211 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
212 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
213 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
214 
215 /*
216  * Requires the syncache entry to be already removed from the bucket list.
217  */
218 static void
syncache_free(struct syncache * sc)219 syncache_free(struct syncache *sc)
220 {
221 
222 	if (sc->sc_ipopts)
223 		(void) m_free(sc->sc_ipopts);
224 	if (sc->sc_cred)
225 		crfree(sc->sc_cred);
226 #ifdef MAC
227 	mac_syncache_destroy(&sc->sc_label);
228 #endif
229 
230 	uma_zfree(V_tcp_syncache.zone, sc);
231 }
232 
233 void
syncache_init(void)234 syncache_init(void)
235 {
236 	int i;
237 
238 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
239 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
240 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
241 	V_tcp_syncache.hash_secret = arc4random();
242 
243 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
244 	    &V_tcp_syncache.hashsize);
245 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
246 	    &V_tcp_syncache.bucket_limit);
247 	if (!powerof2(V_tcp_syncache.hashsize) ||
248 	    V_tcp_syncache.hashsize == 0) {
249 		printf("WARNING: syncache hash size is not a power of 2.\n");
250 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
251 	}
252 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
253 
254 	/* Set limits. */
255 	V_tcp_syncache.cache_limit =
256 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
257 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
258 	    &V_tcp_syncache.cache_limit);
259 
260 	/* Allocate the hash table. */
261 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
262 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
263 
264 #ifdef VIMAGE
265 	V_tcp_syncache.vnet = curvnet;
266 #endif
267 
268 	/* Initialize the hash buckets. */
269 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
270 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
271 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
272 			 NULL, MTX_DEF);
273 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
274 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
275 		V_tcp_syncache.hashbase[i].sch_length = 0;
276 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
277 	}
278 
279 	/* Create the syncache entry zone. */
280 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
281 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
282 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
283 	    V_tcp_syncache.cache_limit);
284 
285 	/* Start the SYN cookie reseeder callout. */
286 	callout_init(&V_tcp_syncache.secret.reseed, 1);
287 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
288 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
289 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
290 	    syncookie_reseed, &V_tcp_syncache);
291 }
292 
293 #ifdef VIMAGE
294 void
syncache_destroy(void)295 syncache_destroy(void)
296 {
297 	struct syncache_head *sch;
298 	struct syncache *sc, *nsc;
299 	int i;
300 
301 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
302 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
303 
304 		sch = &V_tcp_syncache.hashbase[i];
305 		callout_drain(&sch->sch_timer);
306 
307 		SCH_LOCK(sch);
308 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
309 			syncache_drop(sc, sch);
310 		SCH_UNLOCK(sch);
311 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
312 		    ("%s: sch->sch_bucket not empty", __func__));
313 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
314 		    __func__, sch->sch_length));
315 		mtx_destroy(&sch->sch_mtx);
316 	}
317 
318 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
319 	    ("%s: cache_count not 0", __func__));
320 
321 	/* Free the allocated global resources. */
322 	uma_zdestroy(V_tcp_syncache.zone);
323 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
324 
325 	callout_drain(&V_tcp_syncache.secret.reseed);
326 }
327 #endif
328 
329 static int
syncache_sysctl_count(SYSCTL_HANDLER_ARGS)330 syncache_sysctl_count(SYSCTL_HANDLER_ARGS)
331 {
332 	int count;
333 
334 	count = uma_zone_get_cur(V_tcp_syncache.zone);
335 	return (sysctl_handle_int(oidp, &count, 0, req));
336 }
337 
338 /*
339  * Inserts a syncache entry into the specified bucket row.
340  * Locks and unlocks the syncache_head autonomously.
341  */
342 static void
syncache_insert(struct syncache * sc,struct syncache_head * sch)343 syncache_insert(struct syncache *sc, struct syncache_head *sch)
344 {
345 	struct syncache *sc2;
346 
347 	SCH_LOCK(sch);
348 
349 	/*
350 	 * Make sure that we don't overflow the per-bucket limit.
351 	 * If the bucket is full, toss the oldest element.
352 	 */
353 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
354 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
355 			("sch->sch_length incorrect"));
356 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
357 		syncache_drop(sc2, sch);
358 		TCPSTAT_INC(tcps_sc_bucketoverflow);
359 	}
360 
361 	/* Put it into the bucket. */
362 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
363 	sch->sch_length++;
364 
365 #ifdef TCP_OFFLOAD
366 	if (ADDED_BY_TOE(sc)) {
367 		struct toedev *tod = sc->sc_tod;
368 
369 		tod->tod_syncache_added(tod, sc->sc_todctx);
370 	}
371 #endif
372 
373 	/* Reinitialize the bucket row's timer. */
374 	if (sch->sch_length == 1)
375 		sch->sch_nextc = ticks + INT_MAX;
376 	syncache_timeout(sc, sch, 1);
377 
378 	SCH_UNLOCK(sch);
379 
380 	TCPSTAT_INC(tcps_sc_added);
381 }
382 
383 /*
384  * Remove and free entry from syncache bucket row.
385  * Expects locked syncache head.
386  */
387 static void
syncache_drop(struct syncache * sc,struct syncache_head * sch)388 syncache_drop(struct syncache *sc, struct syncache_head *sch)
389 {
390 
391 	SCH_LOCK_ASSERT(sch);
392 
393 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
394 	sch->sch_length--;
395 
396 #ifdef TCP_OFFLOAD
397 	if (ADDED_BY_TOE(sc)) {
398 		struct toedev *tod = sc->sc_tod;
399 
400 		tod->tod_syncache_removed(tod, sc->sc_todctx);
401 	}
402 #endif
403 
404 	syncache_free(sc);
405 }
406 
407 /*
408  * Engage/reengage time on bucket row.
409  */
410 static void
syncache_timeout(struct syncache * sc,struct syncache_head * sch,int docallout)411 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
412 {
413 	sc->sc_rxttime = ticks +
414 		TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
415 	sc->sc_rxmits++;
416 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
417 		sch->sch_nextc = sc->sc_rxttime;
418 		if (docallout)
419 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
420 			    syncache_timer, (void *)sch);
421 	}
422 }
423 
424 /*
425  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
426  * If we have retransmitted an entry the maximum number of times, expire it.
427  * One separate timer for each bucket row.
428  */
429 static void
syncache_timer(void * xsch)430 syncache_timer(void *xsch)
431 {
432 	struct syncache_head *sch = (struct syncache_head *)xsch;
433 	struct syncache *sc, *nsc;
434 	int tick = ticks;
435 	char *s;
436 
437 	CURVNET_SET(sch->sch_sc->vnet);
438 
439 	/* NB: syncache_head has already been locked by the callout. */
440 	SCH_LOCK_ASSERT(sch);
441 
442 	/*
443 	 * In the following cycle we may remove some entries and/or
444 	 * advance some timeouts, so re-initialize the bucket timer.
445 	 */
446 	sch->sch_nextc = tick + INT_MAX;
447 
448 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
449 		/*
450 		 * We do not check if the listen socket still exists
451 		 * and accept the case where the listen socket may be
452 		 * gone by the time we resend the SYN/ACK.  We do
453 		 * not expect this to happens often. If it does,
454 		 * then the RST will be sent by the time the remote
455 		 * host does the SYN/ACK->ACK.
456 		 */
457 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
458 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
459 				sch->sch_nextc = sc->sc_rxttime;
460 			continue;
461 		}
462 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
463 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
464 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
465 				    "giving up and removing syncache entry\n",
466 				    s, __func__);
467 				free(s, M_TCPLOG);
468 			}
469 			syncache_drop(sc, sch);
470 			TCPSTAT_INC(tcps_sc_stale);
471 			continue;
472 		}
473 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
474 			log(LOG_DEBUG, "%s; %s: Response timeout, "
475 			    "retransmitting (%u) SYN|ACK\n",
476 			    s, __func__, sc->sc_rxmits);
477 			free(s, M_TCPLOG);
478 		}
479 
480 		(void) syncache_respond(sc);
481 		TCPSTAT_INC(tcps_sc_retransmitted);
482 		syncache_timeout(sc, sch, 0);
483 	}
484 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
485 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
486 			syncache_timer, (void *)(sch));
487 	CURVNET_RESTORE();
488 }
489 
490 /*
491  * Find an entry in the syncache.
492  * Returns always with locked syncache_head plus a matching entry or NULL.
493  */
494 static struct syncache *
syncache_lookup(struct in_conninfo * inc,struct syncache_head ** schp)495 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
496 {
497 	struct syncache *sc;
498 	struct syncache_head *sch;
499 
500 #ifdef INET6
501 	if (inc->inc_flags & INC_ISIPV6) {
502 		sch = &V_tcp_syncache.hashbase[
503 		    SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)];
504 		*schp = sch;
505 
506 		SCH_LOCK(sch);
507 
508 		/* Circle through bucket row to find matching entry. */
509 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
510 			if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
511 				return (sc);
512 		}
513 	} else
514 #endif
515 	{
516 		sch = &V_tcp_syncache.hashbase[
517 		    SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)];
518 		*schp = sch;
519 
520 		SCH_LOCK(sch);
521 
522 		/* Circle through bucket row to find matching entry. */
523 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
524 #ifdef INET6
525 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
526 				continue;
527 #endif
528 			if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
529 				return (sc);
530 		}
531 	}
532 	SCH_LOCK_ASSERT(*schp);
533 	return (NULL);			/* always returns with locked sch */
534 }
535 
536 /*
537  * This function is called when we get a RST for a
538  * non-existent connection, so that we can see if the
539  * connection is in the syn cache.  If it is, zap it.
540  */
541 void
syncache_chkrst(struct in_conninfo * inc,struct tcphdr * th)542 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
543 {
544 	struct syncache *sc;
545 	struct syncache_head *sch;
546 	char *s = NULL;
547 
548 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
549 	SCH_LOCK_ASSERT(sch);
550 
551 	/*
552 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
553 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
554 	 */
555 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
556 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
557 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
558 			    "FIN flag set, segment ignored\n", s, __func__);
559 		TCPSTAT_INC(tcps_badrst);
560 		goto done;
561 	}
562 
563 	/*
564 	 * No corresponding connection was found in syncache.
565 	 * If syncookies are enabled and possibly exclusively
566 	 * used, or we are under memory pressure, a valid RST
567 	 * may not find a syncache entry.  In that case we're
568 	 * done and no SYN|ACK retransmissions will happen.
569 	 * Otherwise the RST was misdirected or spoofed.
570 	 */
571 	if (sc == NULL) {
572 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
573 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
574 			    "syncache entry (possibly syncookie only), "
575 			    "segment ignored\n", s, __func__);
576 		TCPSTAT_INC(tcps_badrst);
577 		goto done;
578 	}
579 
580 	/*
581 	 * If the RST bit is set, check the sequence number to see
582 	 * if this is a valid reset segment.
583 	 * RFC 793 page 37:
584 	 *   In all states except SYN-SENT, all reset (RST) segments
585 	 *   are validated by checking their SEQ-fields.  A reset is
586 	 *   valid if its sequence number is in the window.
587 	 *
588 	 *   The sequence number in the reset segment is normally an
589 	 *   echo of our outgoing acknowlegement numbers, but some hosts
590 	 *   send a reset with the sequence number at the rightmost edge
591 	 *   of our receive window, and we have to handle this case.
592 	 */
593 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
594 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
595 		syncache_drop(sc, sch);
596 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
597 			log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
598 			    "connection attempt aborted by remote endpoint\n",
599 			    s, __func__);
600 		TCPSTAT_INC(tcps_sc_reset);
601 	} else {
602 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
603 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
604 			    "IRS %u (+WND %u), segment ignored\n",
605 			    s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
606 		TCPSTAT_INC(tcps_badrst);
607 	}
608 
609 done:
610 	if (s != NULL)
611 		free(s, M_TCPLOG);
612 	SCH_UNLOCK(sch);
613 }
614 
615 void
syncache_badack(struct in_conninfo * inc)616 syncache_badack(struct in_conninfo *inc)
617 {
618 	struct syncache *sc;
619 	struct syncache_head *sch;
620 
621 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
622 	SCH_LOCK_ASSERT(sch);
623 	if (sc != NULL) {
624 		syncache_drop(sc, sch);
625 		TCPSTAT_INC(tcps_sc_badack);
626 	}
627 	SCH_UNLOCK(sch);
628 }
629 
630 void
syncache_unreach(struct in_conninfo * inc,struct tcphdr * th)631 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
632 {
633 	struct syncache *sc;
634 	struct syncache_head *sch;
635 
636 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
637 	SCH_LOCK_ASSERT(sch);
638 	if (sc == NULL)
639 		goto done;
640 
641 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
642 	if (ntohl(th->th_seq) != sc->sc_iss)
643 		goto done;
644 
645 	/*
646 	 * If we've rertransmitted 3 times and this is our second error,
647 	 * we remove the entry.  Otherwise, we allow it to continue on.
648 	 * This prevents us from incorrectly nuking an entry during a
649 	 * spurious network outage.
650 	 *
651 	 * See tcp_notify().
652 	 */
653 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
654 		sc->sc_flags |= SCF_UNREACH;
655 		goto done;
656 	}
657 	syncache_drop(sc, sch);
658 	TCPSTAT_INC(tcps_sc_unreach);
659 done:
660 	SCH_UNLOCK(sch);
661 }
662 
663 /*
664  * Build a new TCP socket structure from a syncache entry.
665  */
666 static struct socket *
syncache_socket(struct syncache * sc,struct socket * lso,struct mbuf * m)667 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
668 {
669 	struct inpcb *inp = NULL;
670 	struct socket *so;
671 	struct tcpcb *tp;
672 	int error;
673 	char *s;
674 
675 	INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
676 
677 	/*
678 	 * Ok, create the full blown connection, and set things up
679 	 * as they would have been set up if we had created the
680 	 * connection when the SYN arrived.  If we can't create
681 	 * the connection, abort it.
682 	 */
683 	so = sonewconn(lso, SS_ISCONNECTED);
684 	if (so == NULL) {
685 		/*
686 		 * Drop the connection; we will either send a RST or
687 		 * have the peer retransmit its SYN again after its
688 		 * RTO and try again.
689 		 */
690 		TCPSTAT_INC(tcps_listendrop);
691 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
692 			log(LOG_DEBUG, "%s; %s: Socket create failed "
693 			    "due to limits or memory shortage\n",
694 			    s, __func__);
695 			free(s, M_TCPLOG);
696 		}
697 		goto abort2;
698 	}
699 #ifdef MAC
700 	mac_socketpeer_set_from_mbuf(m, so);
701 #endif
702 
703 	inp = sotoinpcb(so);
704 	inp->inp_inc.inc_fibnum = so->so_fibnum;
705 	INP_WLOCK(inp);
706 	INP_HASH_WLOCK(&V_tcbinfo);
707 
708 	/* Insert new socket into PCB hash list. */
709 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
710 #ifdef INET6
711 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
712 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
713 	} else {
714 		inp->inp_vflag &= ~INP_IPV6;
715 		inp->inp_vflag |= INP_IPV4;
716 #endif
717 		inp->inp_laddr = sc->sc_inc.inc_laddr;
718 #ifdef INET6
719 	}
720 #endif
721 
722 	/*
723 	 * If there's an mbuf and it has a flowid, then let's initialise the
724 	 * inp with that particular flowid.
725 	 */
726 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
727 		inp->inp_flowid = m->m_pkthdr.flowid;
728 	}
729 
730 	/*
731 	 * Install in the reservation hash table for now, but don't yet
732 	 * install a connection group since the full 4-tuple isn't yet
733 	 * configured.
734 	 */
735 	inp->inp_lport = sc->sc_inc.inc_lport;
736 	if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
737 		/*
738 		 * Undo the assignments above if we failed to
739 		 * put the PCB on the hash lists.
740 		 */
741 #ifdef INET6
742 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
743 			inp->in6p_laddr = in6addr_any;
744 		else
745 #endif
746 			inp->inp_laddr.s_addr = INADDR_ANY;
747 		inp->inp_lport = 0;
748 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
749 			log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
750 			    "with error %i\n",
751 			    s, __func__, error);
752 			free(s, M_TCPLOG);
753 		}
754 		INP_HASH_WUNLOCK(&V_tcbinfo);
755 		goto abort;
756 	}
757 #ifdef IPSEC
758 	/* Copy old policy into new socket's. */
759 	if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
760 		printf("syncache_socket: could not copy policy\n");
761 #endif
762 #ifdef INET6
763 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
764 		struct inpcb *oinp = sotoinpcb(lso);
765 		struct in6_addr laddr6;
766 		struct sockaddr_in6 sin6;
767 		/*
768 		 * Inherit socket options from the listening socket.
769 		 * Note that in6p_inputopts are not (and should not be)
770 		 * copied, since it stores previously received options and is
771 		 * used to detect if each new option is different than the
772 		 * previous one and hence should be passed to a user.
773 		 * If we copied in6p_inputopts, a user would not be able to
774 		 * receive options just after calling the accept system call.
775 		 */
776 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
777 		if (oinp->in6p_outputopts)
778 			inp->in6p_outputopts =
779 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
780 
781 		sin6.sin6_family = AF_INET6;
782 		sin6.sin6_len = sizeof(sin6);
783 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
784 		sin6.sin6_port = sc->sc_inc.inc_fport;
785 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
786 		laddr6 = inp->in6p_laddr;
787 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
788 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
789 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
790 		    thread0.td_ucred, m)) != 0) {
791 			inp->in6p_laddr = laddr6;
792 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
793 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
794 				    "with error %i\n",
795 				    s, __func__, error);
796 				free(s, M_TCPLOG);
797 			}
798 			INP_HASH_WUNLOCK(&V_tcbinfo);
799 			goto abort;
800 		}
801 		/* Override flowlabel from in6_pcbconnect. */
802 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
803 		inp->inp_flow |= sc->sc_flowlabel;
804 	}
805 #endif /* INET6 */
806 #if defined(INET) && defined(INET6)
807 	else
808 #endif
809 #ifdef INET
810 	{
811 		struct in_addr laddr;
812 		struct sockaddr_in sin;
813 
814 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
815 
816 		if (inp->inp_options == NULL) {
817 			inp->inp_options = sc->sc_ipopts;
818 			sc->sc_ipopts = NULL;
819 		}
820 
821 		sin.sin_family = AF_INET;
822 		sin.sin_len = sizeof(sin);
823 		sin.sin_addr = sc->sc_inc.inc_faddr;
824 		sin.sin_port = sc->sc_inc.inc_fport;
825 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
826 		laddr = inp->inp_laddr;
827 		if (inp->inp_laddr.s_addr == INADDR_ANY)
828 			inp->inp_laddr = sc->sc_inc.inc_laddr;
829 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
830 		    thread0.td_ucred, m)) != 0) {
831 			inp->inp_laddr = laddr;
832 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
833 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
834 				    "with error %i\n",
835 				    s, __func__, error);
836 				free(s, M_TCPLOG);
837 			}
838 			INP_HASH_WUNLOCK(&V_tcbinfo);
839 			goto abort;
840 		}
841 	}
842 #endif /* INET */
843 	INP_HASH_WUNLOCK(&V_tcbinfo);
844 	tp = intotcpcb(inp);
845 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
846 	tp->iss = sc->sc_iss;
847 	tp->irs = sc->sc_irs;
848 	tcp_rcvseqinit(tp);
849 	tcp_sendseqinit(tp);
850 	tp->snd_wl1 = sc->sc_irs;
851 	tp->snd_max = tp->iss + 1;
852 	tp->snd_nxt = tp->iss + 1;
853 	tp->rcv_up = sc->sc_irs + 1;
854 	tp->rcv_wnd = sc->sc_wnd;
855 	tp->rcv_adv += tp->rcv_wnd;
856 	tp->last_ack_sent = tp->rcv_nxt;
857 
858 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
859 	if (sc->sc_flags & SCF_NOOPT)
860 		tp->t_flags |= TF_NOOPT;
861 	else {
862 		if (sc->sc_flags & SCF_WINSCALE) {
863 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
864 			tp->snd_scale = sc->sc_requested_s_scale;
865 			tp->request_r_scale = sc->sc_requested_r_scale;
866 		}
867 		if (sc->sc_flags & SCF_TIMESTAMP) {
868 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
869 			tp->ts_recent = sc->sc_tsreflect;
870 			tp->ts_recent_age = tcp_ts_getticks();
871 			tp->ts_offset = sc->sc_tsoff;
872 		}
873 #ifdef TCP_SIGNATURE
874 		if (sc->sc_flags & SCF_SIGNATURE)
875 			tp->t_flags |= TF_SIGNATURE;
876 #endif
877 		if (sc->sc_flags & SCF_SACK)
878 			tp->t_flags |= TF_SACK_PERMIT;
879 	}
880 
881 	if (sc->sc_flags & SCF_ECN)
882 		tp->t_flags |= TF_ECN_PERMIT;
883 
884 	/*
885 	 * Set up MSS and get cached values from tcp_hostcache.
886 	 * This might overwrite some of the defaults we just set.
887 	 */
888 	tcp_mss(tp, sc->sc_peer_mss);
889 
890 	/*
891 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
892 	 * limited to one segment in cc_conn_init().
893 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
894 	 */
895 	if (sc->sc_rxmits > 1)
896 		tp->snd_cwnd = 1;
897 
898 #ifdef TCP_OFFLOAD
899 	/*
900 	 * Allow a TOE driver to install its hooks.  Note that we hold the
901 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
902 	 * new connection before the TOE driver has done its thing.
903 	 */
904 	if (ADDED_BY_TOE(sc)) {
905 		struct toedev *tod = sc->sc_tod;
906 
907 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
908 	}
909 #endif
910 	/*
911 	 * Copy and activate timers.
912 	 */
913 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
914 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
915 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
916 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
917 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
918 
919 	INP_WUNLOCK(inp);
920 
921 	TCPSTAT_INC(tcps_accepts);
922 	return (so);
923 
924 abort:
925 	INP_WUNLOCK(inp);
926 abort2:
927 	if (so != NULL)
928 		soabort(so);
929 	return (NULL);
930 }
931 
932 /*
933  * This function gets called when we receive an ACK for a
934  * socket in the LISTEN state.  We look up the connection
935  * in the syncache, and if its there, we pull it out of
936  * the cache and turn it into a full-blown connection in
937  * the SYN-RECEIVED state.
938  */
939 int
syncache_expand(struct in_conninfo * inc,struct tcpopt * to,struct tcphdr * th,struct socket ** lsop,struct mbuf * m)940 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
941     struct socket **lsop, struct mbuf *m)
942 {
943 	struct syncache *sc;
944 	struct syncache_head *sch;
945 	struct syncache scs;
946 	char *s;
947 
948 	/*
949 	 * Global TCP locks are held because we manipulate the PCB lists
950 	 * and create a new socket.
951 	 */
952 	INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
953 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
954 	    ("%s: can handle only ACK", __func__));
955 
956 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
957 	SCH_LOCK_ASSERT(sch);
958 
959 #ifdef INVARIANTS
960 	/*
961 	 * Test code for syncookies comparing the syncache stored
962 	 * values with the reconstructed values from the cookie.
963 	 */
964 	if (sc != NULL)
965 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
966 #endif
967 
968 	if (sc == NULL) {
969 		/*
970 		 * There is no syncache entry, so see if this ACK is
971 		 * a returning syncookie.  To do this, first:
972 		 *  A. See if this socket has had a syncache entry dropped in
973 		 *     the past.  We don't want to accept a bogus syncookie
974 		 *     if we've never received a SYN.
975 		 *  B. check that the syncookie is valid.  If it is, then
976 		 *     cobble up a fake syncache entry, and return.
977 		 */
978 		if (!V_tcp_syncookies) {
979 			SCH_UNLOCK(sch);
980 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
981 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
982 				    "segment rejected (syncookies disabled)\n",
983 				    s, __func__);
984 			goto failed;
985 		}
986 		bzero(&scs, sizeof(scs));
987 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
988 		SCH_UNLOCK(sch);
989 		if (sc == NULL) {
990 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
991 				log(LOG_DEBUG, "%s; %s: Segment failed "
992 				    "SYNCOOKIE authentication, segment rejected "
993 				    "(probably spoofed)\n", s, __func__);
994 			goto failed;
995 		}
996 	} else {
997 		/* Pull out the entry to unlock the bucket row. */
998 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
999 		sch->sch_length--;
1000 #ifdef TCP_OFFLOAD
1001 		if (ADDED_BY_TOE(sc)) {
1002 			struct toedev *tod = sc->sc_tod;
1003 
1004 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1005 		}
1006 #endif
1007 		SCH_UNLOCK(sch);
1008 	}
1009 
1010 	/*
1011 	 * Segment validation:
1012 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1013 	 */
1014 	if (th->th_ack != sc->sc_iss + 1) {
1015 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1016 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1017 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1018 		goto failed;
1019 	}
1020 
1021 	/*
1022 	 * The SEQ must fall in the window starting at the received
1023 	 * initial receive sequence number + 1 (the SYN).
1024 	 */
1025 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1026 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1027 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1028 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1029 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1030 		goto failed;
1031 	}
1032 
1033 	/*
1034 	 * If timestamps were not negotiated during SYN/ACK they
1035 	 * must not appear on any segment during this session.
1036 	 */
1037 	if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1038 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1039 			log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1040 			    "segment rejected\n", s, __func__);
1041 		goto failed;
1042 	}
1043 
1044 	/*
1045 	 * If timestamps were negotiated during SYN/ACK they should
1046 	 * appear on every segment during this session.
1047 	 * XXXAO: This is only informal as there have been unverified
1048 	 * reports of non-compliants stacks.
1049 	 */
1050 	if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1051 		if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1052 			log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1053 			    "no action\n", s, __func__);
1054 			free(s, M_TCPLOG);
1055 			s = NULL;
1056 		}
1057 	}
1058 
1059 	/*
1060 	 * If timestamps were negotiated the reflected timestamp
1061 	 * must be equal to what we actually sent in the SYN|ACK.
1062 	 */
1063 	if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) {
1064 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1065 			log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1066 			    "segment rejected\n",
1067 			    s, __func__, to->to_tsecr, sc->sc_ts);
1068 		goto failed;
1069 	}
1070 
1071 	*lsop = syncache_socket(sc, *lsop, m);
1072 
1073 	if (*lsop == NULL)
1074 		TCPSTAT_INC(tcps_sc_aborted);
1075 	else
1076 		TCPSTAT_INC(tcps_sc_completed);
1077 
1078 /* how do we find the inp for the new socket? */
1079 	if (sc != &scs)
1080 		syncache_free(sc);
1081 	return (1);
1082 failed:
1083 	if (sc != NULL && sc != &scs)
1084 		syncache_free(sc);
1085 	if (s != NULL)
1086 		free(s, M_TCPLOG);
1087 	*lsop = NULL;
1088 	return (0);
1089 }
1090 
1091 /*
1092  * Given a LISTEN socket and an inbound SYN request, add
1093  * this to the syn cache, and send back a segment:
1094  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1095  * to the source.
1096  *
1097  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1098  * Doing so would require that we hold onto the data and deliver it
1099  * to the application.  However, if we are the target of a SYN-flood
1100  * DoS attack, an attacker could send data which would eventually
1101  * consume all available buffer space if it were ACKed.  By not ACKing
1102  * the data, we avoid this DoS scenario.
1103  */
1104 void
syncache_add(struct in_conninfo * inc,struct tcpopt * to,struct tcphdr * th,struct inpcb * inp,struct socket ** lsop,struct mbuf * m,void * tod,void * todctx)1105 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1106     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1107     void *todctx)
1108 {
1109 	struct tcpcb *tp;
1110 	struct socket *so;
1111 	struct syncache *sc = NULL;
1112 	struct syncache_head *sch;
1113 	struct mbuf *ipopts = NULL;
1114 	u_int ltflags;
1115 	int win, sb_hiwat, ip_ttl, ip_tos;
1116 	char *s;
1117 #ifdef INET6
1118 	int autoflowlabel = 0;
1119 #endif
1120 #ifdef MAC
1121 	struct label *maclabel;
1122 #endif
1123 	struct syncache scs;
1124 	struct ucred *cred;
1125 
1126 	INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
1127 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1128 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1129 	    ("%s: unexpected tcp flags", __func__));
1130 
1131 	/*
1132 	 * Combine all so/tp operations very early to drop the INP lock as
1133 	 * soon as possible.
1134 	 */
1135 	so = *lsop;
1136 	tp = sototcpcb(so);
1137 	cred = crhold(so->so_cred);
1138 
1139 #ifdef INET6
1140 	if ((inc->inc_flags & INC_ISIPV6) &&
1141 	    (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1142 		autoflowlabel = 1;
1143 #endif
1144 	ip_ttl = inp->inp_ip_ttl;
1145 	ip_tos = inp->inp_ip_tos;
1146 	win = sbspace(&so->so_rcv);
1147 	sb_hiwat = so->so_rcv.sb_hiwat;
1148 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1149 
1150 	/* By the time we drop the lock these should no longer be used. */
1151 	so = NULL;
1152 	tp = NULL;
1153 
1154 #ifdef MAC
1155 	if (mac_syncache_init(&maclabel) != 0) {
1156 		INP_WUNLOCK(inp);
1157 		INP_INFO_WUNLOCK(&V_tcbinfo);
1158 		goto done;
1159 	} else
1160 		mac_syncache_create(maclabel, inp);
1161 #endif
1162 	INP_WUNLOCK(inp);
1163 	INP_INFO_WUNLOCK(&V_tcbinfo);
1164 
1165 	/*
1166 	 * Remember the IP options, if any.
1167 	 */
1168 #ifdef INET6
1169 	if (!(inc->inc_flags & INC_ISIPV6))
1170 #endif
1171 #ifdef INET
1172 		ipopts = (m) ? ip_srcroute(m) : NULL;
1173 #else
1174 		ipopts = NULL;
1175 #endif
1176 
1177 	/*
1178 	 * See if we already have an entry for this connection.
1179 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1180 	 *
1181 	 * XXX: should the syncache be re-initialized with the contents
1182 	 * of the new SYN here (which may have different options?)
1183 	 *
1184 	 * XXX: We do not check the sequence number to see if this is a
1185 	 * real retransmit or a new connection attempt.  The question is
1186 	 * how to handle such a case; either ignore it as spoofed, or
1187 	 * drop the current entry and create a new one?
1188 	 */
1189 	sc = syncache_lookup(inc, &sch);	/* returns locked entry */
1190 	SCH_LOCK_ASSERT(sch);
1191 	if (sc != NULL) {
1192 		TCPSTAT_INC(tcps_sc_dupsyn);
1193 		if (ipopts) {
1194 			/*
1195 			 * If we were remembering a previous source route,
1196 			 * forget it and use the new one we've been given.
1197 			 */
1198 			if (sc->sc_ipopts)
1199 				(void) m_free(sc->sc_ipopts);
1200 			sc->sc_ipopts = ipopts;
1201 		}
1202 		/*
1203 		 * Update timestamp if present.
1204 		 */
1205 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1206 			sc->sc_tsreflect = to->to_tsval;
1207 		else
1208 			sc->sc_flags &= ~SCF_TIMESTAMP;
1209 #ifdef MAC
1210 		/*
1211 		 * Since we have already unconditionally allocated label
1212 		 * storage, free it up.  The syncache entry will already
1213 		 * have an initialized label we can use.
1214 		 */
1215 		mac_syncache_destroy(&maclabel);
1216 #endif
1217 		/* Retransmit SYN|ACK and reset retransmit count. */
1218 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1219 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1220 			    "resetting timer and retransmitting SYN|ACK\n",
1221 			    s, __func__);
1222 			free(s, M_TCPLOG);
1223 		}
1224 		if (syncache_respond(sc) == 0) {
1225 			sc->sc_rxmits = 0;
1226 			syncache_timeout(sc, sch, 1);
1227 			TCPSTAT_INC(tcps_sndacks);
1228 			TCPSTAT_INC(tcps_sndtotal);
1229 		}
1230 		SCH_UNLOCK(sch);
1231 		goto done;
1232 	}
1233 
1234 	sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1235 	if (sc == NULL) {
1236 		/*
1237 		 * The zone allocator couldn't provide more entries.
1238 		 * Treat this as if the cache was full; drop the oldest
1239 		 * entry and insert the new one.
1240 		 */
1241 		TCPSTAT_INC(tcps_sc_zonefail);
1242 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL)
1243 			syncache_drop(sc, sch);
1244 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1245 		if (sc == NULL) {
1246 			if (V_tcp_syncookies) {
1247 				bzero(&scs, sizeof(scs));
1248 				sc = &scs;
1249 			} else {
1250 				SCH_UNLOCK(sch);
1251 				if (ipopts)
1252 					(void) m_free(ipopts);
1253 				goto done;
1254 			}
1255 		}
1256 	}
1257 
1258 	/*
1259 	 * Fill in the syncache values.
1260 	 */
1261 #ifdef MAC
1262 	sc->sc_label = maclabel;
1263 #endif
1264 	sc->sc_cred = cred;
1265 	cred = NULL;
1266 	sc->sc_ipopts = ipopts;
1267 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1268 #ifdef INET6
1269 	if (!(inc->inc_flags & INC_ISIPV6))
1270 #endif
1271 	{
1272 		sc->sc_ip_tos = ip_tos;
1273 		sc->sc_ip_ttl = ip_ttl;
1274 	}
1275 #ifdef TCP_OFFLOAD
1276 	sc->sc_tod = tod;
1277 	sc->sc_todctx = todctx;
1278 #endif
1279 	sc->sc_irs = th->th_seq;
1280 	sc->sc_iss = arc4random();
1281 	sc->sc_flags = 0;
1282 	sc->sc_flowlabel = 0;
1283 
1284 	/*
1285 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1286 	 * win was derived from socket earlier in the function.
1287 	 */
1288 	win = imax(win, 0);
1289 	win = imin(win, TCP_MAXWIN);
1290 	sc->sc_wnd = win;
1291 
1292 	if (V_tcp_do_rfc1323) {
1293 		/*
1294 		 * A timestamp received in a SYN makes
1295 		 * it ok to send timestamp requests and replies.
1296 		 */
1297 		if (to->to_flags & TOF_TS) {
1298 			sc->sc_tsreflect = to->to_tsval;
1299 			sc->sc_ts = tcp_ts_getticks();
1300 			sc->sc_flags |= SCF_TIMESTAMP;
1301 		}
1302 		if (to->to_flags & TOF_SCALE) {
1303 			int wscale = 0;
1304 
1305 			/*
1306 			 * Pick the smallest possible scaling factor that
1307 			 * will still allow us to scale up to sb_max, aka
1308 			 * kern.ipc.maxsockbuf.
1309 			 *
1310 			 * We do this because there are broken firewalls that
1311 			 * will corrupt the window scale option, leading to
1312 			 * the other endpoint believing that our advertised
1313 			 * window is unscaled.  At scale factors larger than
1314 			 * 5 the unscaled window will drop below 1500 bytes,
1315 			 * leading to serious problems when traversing these
1316 			 * broken firewalls.
1317 			 *
1318 			 * With the default maxsockbuf of 256K, a scale factor
1319 			 * of 3 will be chosen by this algorithm.  Those who
1320 			 * choose a larger maxsockbuf should watch out
1321 			 * for the compatiblity problems mentioned above.
1322 			 *
1323 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1324 			 * or <SYN,ACK>) segment itself is never scaled.
1325 			 */
1326 			while (wscale < TCP_MAX_WINSHIFT &&
1327 			    (TCP_MAXWIN << wscale) < sb_max)
1328 				wscale++;
1329 			sc->sc_requested_r_scale = wscale;
1330 			sc->sc_requested_s_scale = to->to_wscale;
1331 			sc->sc_flags |= SCF_WINSCALE;
1332 		}
1333 	}
1334 #ifdef TCP_SIGNATURE
1335 	/*
1336 	 * If listening socket requested TCP digests, and received SYN
1337 	 * contains the option, flag this in the syncache so that
1338 	 * syncache_respond() will do the right thing with the SYN+ACK.
1339 	 * XXX: Currently we always record the option by default and will
1340 	 * attempt to use it in syncache_respond().
1341 	 */
1342 	if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE)
1343 		sc->sc_flags |= SCF_SIGNATURE;
1344 #endif
1345 	if (to->to_flags & TOF_SACKPERM)
1346 		sc->sc_flags |= SCF_SACK;
1347 	if (to->to_flags & TOF_MSS)
1348 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1349 	if (ltflags & TF_NOOPT)
1350 		sc->sc_flags |= SCF_NOOPT;
1351 	if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1352 		sc->sc_flags |= SCF_ECN;
1353 
1354 	if (V_tcp_syncookies)
1355 		sc->sc_iss = syncookie_generate(sch, sc);
1356 #ifdef INET6
1357 	if (autoflowlabel) {
1358 		if (V_tcp_syncookies)
1359 			sc->sc_flowlabel = sc->sc_iss;
1360 		else
1361 			sc->sc_flowlabel = ip6_randomflowlabel();
1362 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1363 	}
1364 #endif
1365 	SCH_UNLOCK(sch);
1366 
1367 	/*
1368 	 * Do a standard 3-way handshake.
1369 	 */
1370 	if (syncache_respond(sc) == 0) {
1371 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1372 			syncache_free(sc);
1373 		else if (sc != &scs)
1374 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1375 		TCPSTAT_INC(tcps_sndacks);
1376 		TCPSTAT_INC(tcps_sndtotal);
1377 	} else {
1378 		if (sc != &scs)
1379 			syncache_free(sc);
1380 		TCPSTAT_INC(tcps_sc_dropped);
1381 	}
1382 
1383 done:
1384 	if (cred != NULL)
1385 		crfree(cred);
1386 #ifdef MAC
1387 	if (sc == &scs)
1388 		mac_syncache_destroy(&maclabel);
1389 #endif
1390 	if (m) {
1391 
1392 		*lsop = NULL;
1393 		m_freem(m);
1394 	}
1395 }
1396 
1397 static int
syncache_respond(struct syncache * sc)1398 syncache_respond(struct syncache *sc)
1399 {
1400 	struct ip *ip = NULL;
1401 	struct mbuf *m;
1402 	struct tcphdr *th = NULL;
1403 	int optlen, error = 0;	/* Make compiler happy */
1404 	u_int16_t hlen, tlen, mssopt;
1405 	struct tcpopt to;
1406 #ifdef INET6
1407 	struct ip6_hdr *ip6 = NULL;
1408 #endif
1409 
1410 	hlen =
1411 #ifdef INET6
1412 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1413 #endif
1414 		sizeof(struct ip);
1415 	tlen = hlen + sizeof(struct tcphdr);
1416 
1417 	/* Determine MSS we advertize to other end of connection. */
1418 	mssopt = tcp_mssopt(&sc->sc_inc);
1419 	if (sc->sc_peer_mss)
1420 		mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1421 
1422 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1423 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1424 	    ("syncache: mbuf too small"));
1425 
1426 	/* Create the IP+TCP header from scratch. */
1427 	m = m_gethdr(M_NOWAIT, MT_DATA);
1428 	if (m == NULL)
1429 		return (ENOBUFS);
1430 #ifdef MAC
1431 	mac_syncache_create_mbuf(sc->sc_label, m);
1432 #endif
1433 	m->m_data += max_linkhdr;
1434 	m->m_len = tlen;
1435 	m->m_pkthdr.len = tlen;
1436 	m->m_pkthdr.rcvif = NULL;
1437 
1438 #ifdef INET6
1439 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1440 		ip6 = mtod(m, struct ip6_hdr *);
1441 		ip6->ip6_vfc = IPV6_VERSION;
1442 		ip6->ip6_nxt = IPPROTO_TCP;
1443 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1444 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1445 		ip6->ip6_plen = htons(tlen - hlen);
1446 		/* ip6_hlim is set after checksum */
1447 		ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1448 		ip6->ip6_flow |= sc->sc_flowlabel;
1449 
1450 		th = (struct tcphdr *)(ip6 + 1);
1451 	}
1452 #endif
1453 #if defined(INET6) && defined(INET)
1454 	else
1455 #endif
1456 #ifdef INET
1457 	{
1458 		ip = mtod(m, struct ip *);
1459 		ip->ip_v = IPVERSION;
1460 		ip->ip_hl = sizeof(struct ip) >> 2;
1461 		ip->ip_len = htons(tlen);
1462 		ip->ip_id = 0;
1463 		ip->ip_off = 0;
1464 		ip->ip_sum = 0;
1465 		ip->ip_p = IPPROTO_TCP;
1466 		ip->ip_src = sc->sc_inc.inc_laddr;
1467 		ip->ip_dst = sc->sc_inc.inc_faddr;
1468 		ip->ip_ttl = sc->sc_ip_ttl;
1469 		ip->ip_tos = sc->sc_ip_tos;
1470 
1471 		/*
1472 		 * See if we should do MTU discovery.  Route lookups are
1473 		 * expensive, so we will only unset the DF bit if:
1474 		 *
1475 		 *	1) path_mtu_discovery is disabled
1476 		 *	2) the SCF_UNREACH flag has been set
1477 		 */
1478 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1479 		       ip->ip_off |= htons(IP_DF);
1480 
1481 		th = (struct tcphdr *)(ip + 1);
1482 	}
1483 #endif /* INET */
1484 	th->th_sport = sc->sc_inc.inc_lport;
1485 	th->th_dport = sc->sc_inc.inc_fport;
1486 
1487 	th->th_seq = htonl(sc->sc_iss);
1488 	th->th_ack = htonl(sc->sc_irs + 1);
1489 	th->th_off = sizeof(struct tcphdr) >> 2;
1490 	th->th_x2 = 0;
1491 	th->th_flags = TH_SYN|TH_ACK;
1492 	th->th_win = htons(sc->sc_wnd);
1493 	th->th_urp = 0;
1494 
1495 	if (sc->sc_flags & SCF_ECN) {
1496 		th->th_flags |= TH_ECE;
1497 		TCPSTAT_INC(tcps_ecn_shs);
1498 	}
1499 
1500 	/* Tack on the TCP options. */
1501 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1502 		to.to_flags = 0;
1503 
1504 		to.to_mss = mssopt;
1505 		to.to_flags = TOF_MSS;
1506 		if (sc->sc_flags & SCF_WINSCALE) {
1507 			to.to_wscale = sc->sc_requested_r_scale;
1508 			to.to_flags |= TOF_SCALE;
1509 		}
1510 		if (sc->sc_flags & SCF_TIMESTAMP) {
1511 			/* Virgin timestamp or TCP cookie enhanced one. */
1512 			to.to_tsval = sc->sc_ts;
1513 			to.to_tsecr = sc->sc_tsreflect;
1514 			to.to_flags |= TOF_TS;
1515 		}
1516 		if (sc->sc_flags & SCF_SACK)
1517 			to.to_flags |= TOF_SACKPERM;
1518 #ifdef TCP_SIGNATURE
1519 		if (sc->sc_flags & SCF_SIGNATURE)
1520 			to.to_flags |= TOF_SIGNATURE;
1521 #endif
1522 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1523 
1524 		/* Adjust headers by option size. */
1525 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1526 		m->m_len += optlen;
1527 		m->m_pkthdr.len += optlen;
1528 
1529 #ifdef TCP_SIGNATURE
1530 		if (sc->sc_flags & SCF_SIGNATURE)
1531 			tcp_signature_compute(m, 0, 0, optlen,
1532 			    to.to_signature, IPSEC_DIR_OUTBOUND);
1533 #endif
1534 #ifdef INET6
1535 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1536 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1537 		else
1538 #endif
1539 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1540 	} else
1541 		optlen = 0;
1542 
1543 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1544 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1545 #ifdef INET6
1546 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1547 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1548 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1549 		    IPPROTO_TCP, 0);
1550 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1551 #ifdef TCP_OFFLOAD
1552 		if (ADDED_BY_TOE(sc)) {
1553 			struct toedev *tod = sc->sc_tod;
1554 
1555 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1556 
1557 			return (error);
1558 		}
1559 #endif
1560 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1561 	}
1562 #endif
1563 #if defined(INET6) && defined(INET)
1564 	else
1565 #endif
1566 #ifdef INET
1567 	{
1568 		m->m_pkthdr.csum_flags = CSUM_TCP;
1569 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1570 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1571 #ifdef TCP_OFFLOAD
1572 		if (ADDED_BY_TOE(sc)) {
1573 			struct toedev *tod = sc->sc_tod;
1574 
1575 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1576 
1577 			return (error);
1578 		}
1579 #endif
1580 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1581 	}
1582 #endif
1583 	return (error);
1584 }
1585 
1586 /*
1587  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1588  * that exceed the capacity of the syncache by avoiding the storage of any
1589  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
1590  * attacks where the attacker does not have access to our responses.
1591  *
1592  * Syncookies encode and include all necessary information about the
1593  * connection setup within the SYN|ACK that we send back.  That way we
1594  * can avoid keeping any local state until the ACK to our SYN|ACK returns
1595  * (if ever).  Normally the syncache and syncookies are running in parallel
1596  * with the latter taking over when the former is exhausted.  When matching
1597  * syncache entry is found the syncookie is ignored.
1598  *
1599  * The only reliable information persisting the 3WHS is our inital sequence
1600  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
1601  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1602  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
1603  * returns and signifies a legitimate connection if it matches the ACK.
1604  *
1605  * The available space of 32 bits to store the hash and to encode the SYN
1606  * option information is very tight and we should have at least 24 bits for
1607  * the MAC to keep the number of guesses by blind spoofing reasonably high.
1608  *
1609  * SYN option information we have to encode to fully restore a connection:
1610  * MSS: is imporant to chose an optimal segment size to avoid IP level
1611  *   fragmentation along the path.  The common MSS values can be encoded
1612  *   in a 3-bit table.  Uncommon values are captured by the next lower value
1613  *   in the table leading to a slight increase in packetization overhead.
1614  * WSCALE: is necessary to allow large windows to be used for high delay-
1615  *   bandwidth product links.  Not scaling the window when it was initially
1616  *   negotiated is bad for performance as lack of scaling further decreases
1617  *   the apparent available send window.  We only need to encode the WSCALE
1618  *   we received from the remote end.  Our end can be recalculated at any
1619  *   time.  The common WSCALE values can be encoded in a 3-bit table.
1620  *   Uncommon values are captured by the next lower value in the table
1621  *   making us under-estimate the available window size halving our
1622  *   theoretically possible maximum throughput for that connection.
1623  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1624  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1625  *   that are included in all segments on a connection.  We enable them when
1626  *   the ACK has them.
1627  *
1628  * Security of syncookies and attack vectors:
1629  *
1630  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1631  * together with the gloabl secret to make it unique per connection attempt.
1632  * Thus any change of any of those parameters results in a different MAC output
1633  * in an unpredictable way unless a collision is encountered.  24 bits of the
1634  * MAC are embedded into the ISS.
1635  *
1636  * To prevent replay attacks two rotating global secrets are updated with a
1637  * new random value every 15 seconds.  The life-time of a syncookie is thus
1638  * 15-30 seconds.
1639  *
1640  * Vector 1: Attacking the secret.  This requires finding a weakness in the
1641  * MAC itself or the way it is used here.  The attacker can do a chosen plain
1642  * text attack by varying and testing the all parameters under his control.
1643  * The strength depends on the size and randomness of the secret, and the
1644  * cryptographic security of the MAC function.  Due to the constant updating
1645  * of the secret the attacker has at most 29.999 seconds to find the secret
1646  * and launch spoofed connections.  After that he has to start all over again.
1647  *
1648  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
1649  * size an average of 4,823 attempts are required for a 50% chance of success
1650  * to spoof a single syncookie (birthday collision paradox).  However the
1651  * attacker is blind and doesn't know if one of his attempts succeeded unless
1652  * he has a side channel to interfere success from.  A single connection setup
1653  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1654  * This many attempts are required for each one blind spoofed connection.  For
1655  * every additional spoofed connection he has to launch another N attempts.
1656  * Thus for a sustained rate 100 spoofed connections per second approximately
1657  * 1,800,000 packets per second would have to be sent.
1658  *
1659  * NB: The MAC function should be fast so that it doesn't become a CPU
1660  * exhaustion attack vector itself.
1661  *
1662  * References:
1663  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1664  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1665  *   http://cr.yp.to/syncookies.html    (overview)
1666  *   http://cr.yp.to/syncookies/archive (details)
1667  *
1668  *
1669  * Schematic construction of a syncookie enabled Initial Sequence Number:
1670  *  0        1         2         3
1671  *  12345678901234567890123456789012
1672  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1673  *
1674  *  x 24 MAC (truncated)
1675  *  W  3 Send Window Scale index
1676  *  M  3 MSS index
1677  *  S  1 SACK permitted
1678  *  P  1 Odd/even secret
1679  */
1680 
1681 /*
1682  * Distribution and probability of certain MSS values.  Those in between are
1683  * rounded down to the next lower one.
1684  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1685  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
1686  */
1687 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1688 
1689 /*
1690  * Distribution and probability of certain WSCALE values.  We have to map the
1691  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1692  * bits based on prevalence of certain values.  Where we don't have an exact
1693  * match for are rounded down to the next lower one letting us under-estimate
1694  * the true available window.  At the moment this would happen only for the
1695  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1696  * and window size).  The absence of the WSCALE option (no scaling in either
1697  * direction) is encoded with index zero.
1698  * [WSCALE values histograms, Allman, 2012]
1699  *                            X 10 10 35  5  6 14 10%   by host
1700  *                            X 11  4  5  5 18 49  3%   by connections
1701  */
1702 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1703 
1704 /*
1705  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
1706  * and good cryptographic properties.
1707  */
1708 static uint32_t
syncookie_mac(struct in_conninfo * inc,tcp_seq irs,uint8_t flags,uint8_t * secbits,uintptr_t secmod)1709 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1710     uint8_t *secbits, uintptr_t secmod)
1711 {
1712 	SIPHASH_CTX ctx;
1713 	uint32_t siphash[2];
1714 
1715 	SipHash24_Init(&ctx);
1716 	SipHash_SetKey(&ctx, secbits);
1717 	switch (inc->inc_flags & INC_ISIPV6) {
1718 #ifdef INET
1719 	case 0:
1720 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1721 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1722 		break;
1723 #endif
1724 #ifdef INET6
1725 	case INC_ISIPV6:
1726 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1727 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1728 		break;
1729 #endif
1730 	}
1731 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1732 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1733 	SipHash_Update(&ctx, &flags, sizeof(flags));
1734 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
1735 	SipHash_Final((u_int8_t *)&siphash, &ctx);
1736 
1737 	return (siphash[0] ^ siphash[1]);
1738 }
1739 
1740 static tcp_seq
syncookie_generate(struct syncache_head * sch,struct syncache * sc)1741 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1742 {
1743 	u_int i, mss, secbit, wscale;
1744 	uint32_t iss, hash;
1745 	uint8_t *secbits;
1746 	union syncookie cookie;
1747 
1748 	SCH_LOCK_ASSERT(sch);
1749 
1750 	cookie.cookie = 0;
1751 
1752 	/* Map our computed MSS into the 3-bit index. */
1753 	mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1754 	for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1;
1755 	     tcp_sc_msstab[i] > mss && i > 0;
1756 	     i--)
1757 		;
1758 	cookie.flags.mss_idx = i;
1759 
1760 	/*
1761 	 * Map the send window scale into the 3-bit index but only if
1762 	 * the wscale option was received.
1763 	 */
1764 	if (sc->sc_flags & SCF_WINSCALE) {
1765 		wscale = sc->sc_requested_s_scale;
1766 		for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1;
1767 		     tcp_sc_wstab[i] > wscale && i > 0;
1768 		     i--)
1769 			;
1770 		cookie.flags.wscale_idx = i;
1771 	}
1772 
1773 	/* Can we do SACK? */
1774 	if (sc->sc_flags & SCF_SACK)
1775 		cookie.flags.sack_ok = 1;
1776 
1777 	/* Which of the two secrets to use. */
1778 	secbit = sch->sch_sc->secret.oddeven & 0x1;
1779 	cookie.flags.odd_even = secbit;
1780 
1781 	secbits = sch->sch_sc->secret.key[secbit];
1782 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
1783 	    (uintptr_t)sch);
1784 
1785 	/*
1786 	 * Put the flags into the hash and XOR them to get better ISS number
1787 	 * variance.  This doesn't enhance the cryptographic strength and is
1788 	 * done to prevent the 8 cookie bits from showing up directly on the
1789 	 * wire.
1790 	 */
1791 	iss = hash & ~0xff;
1792 	iss |= cookie.cookie ^ (hash >> 24);
1793 
1794 	/* Randomize the timestamp. */
1795 	if (sc->sc_flags & SCF_TIMESTAMP) {
1796 		sc->sc_ts = arc4random();
1797 		sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
1798 	}
1799 
1800 	TCPSTAT_INC(tcps_sc_sendcookie);
1801 	return (iss);
1802 }
1803 
1804 static struct syncache *
syncookie_lookup(struct in_conninfo * inc,struct syncache_head * sch,struct syncache * sc,struct tcphdr * th,struct tcpopt * to,struct socket * lso)1805 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1806     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1807     struct socket *lso)
1808 {
1809 	uint32_t hash;
1810 	uint8_t *secbits;
1811 	tcp_seq ack, seq;
1812 	int wnd, wscale = 0;
1813 	union syncookie cookie;
1814 
1815 	SCH_LOCK_ASSERT(sch);
1816 
1817 	/*
1818 	 * Pull information out of SYN-ACK/ACK and revert sequence number
1819 	 * advances.
1820 	 */
1821 	ack = th->th_ack - 1;
1822 	seq = th->th_seq - 1;
1823 
1824 	/*
1825 	 * Unpack the flags containing enough information to restore the
1826 	 * connection.
1827 	 */
1828 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
1829 
1830 	/* Which of the two secrets to use. */
1831 	secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
1832 
1833 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
1834 
1835 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
1836 	if ((ack & ~0xff) != (hash & ~0xff))
1837 		return (NULL);
1838 
1839 	/* Fill in the syncache values. */
1840 	sc->sc_flags = 0;
1841 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1842 	sc->sc_ipopts = NULL;
1843 
1844 	sc->sc_irs = seq;
1845 	sc->sc_iss = ack;
1846 
1847 	switch (inc->inc_flags & INC_ISIPV6) {
1848 #ifdef INET
1849 	case 0:
1850 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
1851 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
1852 		break;
1853 #endif
1854 #ifdef INET6
1855 	case INC_ISIPV6:
1856 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
1857 			sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
1858 		break;
1859 #endif
1860 	}
1861 
1862 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
1863 
1864 	/* We can simply recompute receive window scale we sent earlier. */
1865 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
1866 		wscale++;
1867 
1868 	/* Only use wscale if it was enabled in the orignal SYN. */
1869 	if (cookie.flags.wscale_idx > 0) {
1870 		sc->sc_requested_r_scale = wscale;
1871 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
1872 		sc->sc_flags |= SCF_WINSCALE;
1873 	}
1874 
1875 	wnd = sbspace(&lso->so_rcv);
1876 	wnd = imax(wnd, 0);
1877 	wnd = imin(wnd, TCP_MAXWIN);
1878 	sc->sc_wnd = wnd;
1879 
1880 	if (cookie.flags.sack_ok)
1881 		sc->sc_flags |= SCF_SACK;
1882 
1883 	if (to->to_flags & TOF_TS) {
1884 		sc->sc_flags |= SCF_TIMESTAMP;
1885 		sc->sc_tsreflect = to->to_tsval;
1886 		sc->sc_ts = to->to_tsecr;
1887 		sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
1888 	}
1889 
1890 	if (to->to_flags & TOF_SIGNATURE)
1891 		sc->sc_flags |= SCF_SIGNATURE;
1892 
1893 	sc->sc_rxmits = 0;
1894 
1895 	TCPSTAT_INC(tcps_sc_recvcookie);
1896 	return (sc);
1897 }
1898 
1899 #ifdef INVARIANTS
1900 static int
syncookie_cmp(struct in_conninfo * inc,struct syncache_head * sch,struct syncache * sc,struct tcphdr * th,struct tcpopt * to,struct socket * lso)1901 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
1902     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1903     struct socket *lso)
1904 {
1905 	struct syncache scs, *scx;
1906 	char *s;
1907 
1908 	bzero(&scs, sizeof(scs));
1909 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
1910 
1911 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
1912 		return (0);
1913 
1914 	if (scx != NULL) {
1915 		if (sc->sc_peer_mss != scx->sc_peer_mss)
1916 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
1917 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
1918 
1919 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
1920 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
1921 			    s, __func__, sc->sc_requested_r_scale,
1922 			    scx->sc_requested_r_scale);
1923 
1924 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
1925 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
1926 			    s, __func__, sc->sc_requested_s_scale,
1927 			    scx->sc_requested_s_scale);
1928 
1929 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
1930 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
1931 	}
1932 
1933 	if (s != NULL)
1934 		free(s, M_TCPLOG);
1935 	return (0);
1936 }
1937 #endif /* INVARIANTS */
1938 
1939 static void
syncookie_reseed(void * arg)1940 syncookie_reseed(void *arg)
1941 {
1942 	struct tcp_syncache *sc = arg;
1943 	uint8_t *secbits;
1944 	int secbit;
1945 
1946 	/*
1947 	 * Reseeding the secret doesn't have to be protected by a lock.
1948 	 * It only must be ensured that the new random values are visible
1949 	 * to all CPUs in a SMP environment.  The atomic with release
1950 	 * semantics ensures that.
1951 	 */
1952 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
1953 	secbits = sc->secret.key[secbit];
1954 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
1955 	atomic_add_rel_int(&sc->secret.oddeven, 1);
1956 
1957 	/* Reschedule ourself. */
1958 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
1959 }
1960 
1961 /*
1962  * Returns the current number of syncache entries.  This number
1963  * will probably change before you get around to calling
1964  * syncache_pcblist.
1965  */
1966 int
syncache_pcbcount(void)1967 syncache_pcbcount(void)
1968 {
1969 	struct syncache_head *sch;
1970 	int count, i;
1971 
1972 	for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
1973 		/* No need to lock for a read. */
1974 		sch = &V_tcp_syncache.hashbase[i];
1975 		count += sch->sch_length;
1976 	}
1977 	return count;
1978 }
1979 
1980 /*
1981  * Exports the syncache entries to userland so that netstat can display
1982  * them alongside the other sockets.  This function is intended to be
1983  * called only from tcp_pcblist.
1984  *
1985  * Due to concurrency on an active system, the number of pcbs exported
1986  * may have no relation to max_pcbs.  max_pcbs merely indicates the
1987  * amount of space the caller allocated for this function to use.
1988  */
1989 int
syncache_pcblist(struct sysctl_req * req,int max_pcbs,int * pcbs_exported)1990 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
1991 {
1992 	struct xtcpcb xt;
1993 	struct syncache *sc;
1994 	struct syncache_head *sch;
1995 	int count, error, i;
1996 
1997 	for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
1998 		sch = &V_tcp_syncache.hashbase[i];
1999 		SCH_LOCK(sch);
2000 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2001 			if (count >= max_pcbs) {
2002 				SCH_UNLOCK(sch);
2003 				goto exit;
2004 			}
2005 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2006 				continue;
2007 			bzero(&xt, sizeof(xt));
2008 			xt.xt_len = sizeof(xt);
2009 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2010 				xt.xt_inp.inp_vflag = INP_IPV6;
2011 			else
2012 				xt.xt_inp.inp_vflag = INP_IPV4;
2013 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2014 			xt.xt_tp.t_inpcb = &xt.xt_inp;
2015 			xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2016 			xt.xt_socket.xso_protocol = IPPROTO_TCP;
2017 			xt.xt_socket.xso_len = sizeof (struct xsocket);
2018 			xt.xt_socket.so_type = SOCK_STREAM;
2019 			xt.xt_socket.so_state = SS_ISCONNECTING;
2020 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2021 			if (error) {
2022 				SCH_UNLOCK(sch);
2023 				goto exit;
2024 			}
2025 			count++;
2026 		}
2027 		SCH_UNLOCK(sch);
2028 	}
2029 exit:
2030 	*pcbs_exported = count;
2031 	return error;
2032 }
2033