xref: /freebsd/sys/netinet/tcp_syncache.c (revision 81e0e7b9e36d6a25b3af6482811318e085537d2f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2001 McAfee, Inc.
5  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Jonathan Lemon
9  * and McAfee Research, the Security Research Division of McAfee, Inc. under
10  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11  * DARPA CHATS research program. [2001 McAfee, Inc.]
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "opt_inet.h"
39 #include "opt_inet6.h"
40 #include "opt_ipsec.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/hash.h>
45 #include <sys/refcount.h>
46 #include <sys/kernel.h>
47 #include <sys/sysctl.h>
48 #include <sys/limits.h>
49 #include <sys/lock.h>
50 #include <sys/mutex.h>
51 #include <sys/malloc.h>
52 #include <sys/mbuf.h>
53 #include <sys/proc.h>		/* for proc0 declaration */
54 #include <sys/random.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/syslog.h>
58 #include <sys/ucred.h>
59 
60 #include <sys/md5.h>
61 #include <crypto/siphash/siphash.h>
62 
63 #include <vm/uma.h>
64 
65 #include <net/if.h>
66 #include <net/if_var.h>
67 #include <net/route.h>
68 #include <net/vnet.h>
69 
70 #include <netinet/in.h>
71 #include <netinet/in_kdtrace.h>
72 #include <netinet/in_systm.h>
73 #include <netinet/ip.h>
74 #include <netinet/in_var.h>
75 #include <netinet/in_pcb.h>
76 #include <netinet/ip_var.h>
77 #include <netinet/ip_options.h>
78 #ifdef INET6
79 #include <netinet/ip6.h>
80 #include <netinet/icmp6.h>
81 #include <netinet6/nd6.h>
82 #include <netinet6/ip6_var.h>
83 #include <netinet6/in6_pcb.h>
84 #endif
85 #include <netinet/tcp.h>
86 #include <netinet/tcp_fastopen.h>
87 #include <netinet/tcp_fsm.h>
88 #include <netinet/tcp_seq.h>
89 #include <netinet/tcp_timer.h>
90 #include <netinet/tcp_var.h>
91 #include <netinet/tcp_syncache.h>
92 #include <netinet/tcp_ecn.h>
93 #ifdef INET6
94 #include <netinet6/tcp6_var.h>
95 #endif
96 #ifdef TCP_OFFLOAD
97 #include <netinet/toecore.h>
98 #endif
99 #include <netinet/udp.h>
100 #include <netinet/udp_var.h>
101 
102 #include <netipsec/ipsec_support.h>
103 
104 #include <machine/in_cksum.h>
105 
106 #include <security/mac/mac_framework.h>
107 
108 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
109 #define	V_tcp_syncookies		VNET(tcp_syncookies)
110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
111     &VNET_NAME(tcp_syncookies), 0,
112     "Use TCP SYN cookies if the syncache overflows");
113 
114 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
115 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
117     &VNET_NAME(tcp_syncookiesonly), 0,
118     "Use only TCP SYN cookies");
119 
120 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
121 #define V_functions_inherit_listen_socket_stack \
122     VNET(functions_inherit_listen_socket_stack)
123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
124     CTLFLAG_VNET | CTLFLAG_RW,
125     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
126     "Inherit listen socket's stack");
127 
128 #ifdef TCP_OFFLOAD
129 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
130 #endif
131 
132 static void	 syncache_drop(struct syncache *, struct syncache_head *);
133 static void	 syncache_free(struct syncache *);
134 static void	 syncache_insert(struct syncache *, struct syncache_head *);
135 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
136 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
137 		    struct mbuf *m);
138 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
139 		    int docallout);
140 static void	 syncache_timer(void *);
141 
142 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
143 		    uint8_t *, uintptr_t);
144 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
145 static struct syncache
146 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
147 		    struct syncache *, struct tcphdr *, struct tcpopt *,
148 		    struct socket *, uint16_t);
149 static void	syncache_pause(struct in_conninfo *);
150 static void	syncache_unpause(void *);
151 static void	 syncookie_reseed(void *);
152 #ifdef INVARIANTS
153 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
154 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
155 		    struct socket *lso, uint16_t port);
156 #endif
157 
158 /*
159  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
160  * 3 retransmits corresponds to a timeout with default values of
161  * tcp_rexmit_initial * (             1 +
162  *                       tcp_backoff[1] +
163  *                       tcp_backoff[2] +
164  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
165  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
166  * the odds are that the user has given up attempting to connect by then.
167  */
168 #define SYNCACHE_MAXREXMTS		3
169 
170 /* Arbitrary values */
171 #define TCP_SYNCACHE_HASHSIZE		512
172 #define TCP_SYNCACHE_BUCKETLIMIT	30
173 
174 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
175 #define	V_tcp_syncache			VNET(tcp_syncache)
176 
177 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
178     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
179     "TCP SYN cache");
180 
181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
182     &VNET_NAME(tcp_syncache.bucket_limit), 0,
183     "Per-bucket hash limit for syncache");
184 
185 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
186     &VNET_NAME(tcp_syncache.cache_limit), 0,
187     "Overall entry limit for syncache");
188 
189 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
190     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
191 
192 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
193     &VNET_NAME(tcp_syncache.hashsize), 0,
194     "Size of TCP syncache hashtable");
195 
196 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
197     CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
198     "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
199     "and mac(4) checks");
200 
201 static int
202 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
203 {
204 	int error;
205 	u_int new;
206 
207 	new = V_tcp_syncache.rexmt_limit;
208 	error = sysctl_handle_int(oidp, &new, 0, req);
209 	if ((error == 0) && (req->newptr != NULL)) {
210 		if (new > TCP_MAXRXTSHIFT)
211 			error = EINVAL;
212 		else
213 			V_tcp_syncache.rexmt_limit = new;
214 	}
215 	return (error);
216 }
217 
218 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
219     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
220     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
221     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
222     "Limit on SYN/ACK retransmissions");
223 
224 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
225 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
226     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
227     "Send reset on socket allocation failure");
228 
229 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
230 
231 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
232 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
233 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
234 
235 /*
236  * Requires the syncache entry to be already removed from the bucket list.
237  */
238 static void
239 syncache_free(struct syncache *sc)
240 {
241 
242 	if (sc->sc_ipopts)
243 		(void) m_free(sc->sc_ipopts);
244 	if (sc->sc_cred)
245 		crfree(sc->sc_cred);
246 #ifdef MAC
247 	mac_syncache_destroy(&sc->sc_label);
248 #endif
249 
250 	uma_zfree(V_tcp_syncache.zone, sc);
251 }
252 
253 void
254 syncache_init(void)
255 {
256 	int i;
257 
258 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
259 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
260 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
261 	V_tcp_syncache.hash_secret = arc4random();
262 
263 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
264 	    &V_tcp_syncache.hashsize);
265 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
266 	    &V_tcp_syncache.bucket_limit);
267 	if (!powerof2(V_tcp_syncache.hashsize) ||
268 	    V_tcp_syncache.hashsize == 0) {
269 		printf("WARNING: syncache hash size is not a power of 2.\n");
270 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
271 	}
272 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
273 
274 	/* Set limits. */
275 	V_tcp_syncache.cache_limit =
276 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
277 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
278 	    &V_tcp_syncache.cache_limit);
279 
280 	/* Allocate the hash table. */
281 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
282 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
283 
284 #ifdef VIMAGE
285 	V_tcp_syncache.vnet = curvnet;
286 #endif
287 
288 	/* Initialize the hash buckets. */
289 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
290 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
291 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
292 			 NULL, MTX_DEF);
293 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
294 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
295 		V_tcp_syncache.hashbase[i].sch_length = 0;
296 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
297 		V_tcp_syncache.hashbase[i].sch_last_overflow =
298 		    -(SYNCOOKIE_LIFETIME + 1);
299 	}
300 
301 	/* Create the syncache entry zone. */
302 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
303 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
304 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
305 	    V_tcp_syncache.cache_limit);
306 
307 	/* Start the SYN cookie reseeder callout. */
308 	callout_init(&V_tcp_syncache.secret.reseed, 1);
309 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
310 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
311 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
312 	    syncookie_reseed, &V_tcp_syncache);
313 
314 	/* Initialize the pause machinery. */
315 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
316 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
317 	    0);
318 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
319 	V_tcp_syncache.pause_backoff = 0;
320 	V_tcp_syncache.paused = false;
321 }
322 
323 #ifdef VIMAGE
324 void
325 syncache_destroy(void)
326 {
327 	struct syncache_head *sch;
328 	struct syncache *sc, *nsc;
329 	int i;
330 
331 	/*
332 	 * Stop the re-seed timer before freeing resources.  No need to
333 	 * possibly schedule it another time.
334 	 */
335 	callout_drain(&V_tcp_syncache.secret.reseed);
336 
337 	/* Stop the SYN cache pause callout. */
338 	mtx_lock(&V_tcp_syncache.pause_mtx);
339 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
340 		mtx_unlock(&V_tcp_syncache.pause_mtx);
341 		callout_drain(&V_tcp_syncache.pause_co);
342 	} else
343 		mtx_unlock(&V_tcp_syncache.pause_mtx);
344 
345 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
346 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
347 		sch = &V_tcp_syncache.hashbase[i];
348 		callout_drain(&sch->sch_timer);
349 
350 		SCH_LOCK(sch);
351 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
352 			syncache_drop(sc, sch);
353 		SCH_UNLOCK(sch);
354 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
355 		    ("%s: sch->sch_bucket not empty", __func__));
356 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
357 		    __func__, sch->sch_length));
358 		mtx_destroy(&sch->sch_mtx);
359 	}
360 
361 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
362 	    ("%s: cache_count not 0", __func__));
363 
364 	/* Free the allocated global resources. */
365 	uma_zdestroy(V_tcp_syncache.zone);
366 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
367 	mtx_destroy(&V_tcp_syncache.pause_mtx);
368 }
369 #endif
370 
371 /*
372  * Inserts a syncache entry into the specified bucket row.
373  * Locks and unlocks the syncache_head autonomously.
374  */
375 static void
376 syncache_insert(struct syncache *sc, struct syncache_head *sch)
377 {
378 	struct syncache *sc2;
379 
380 	SCH_LOCK(sch);
381 
382 	/*
383 	 * Make sure that we don't overflow the per-bucket limit.
384 	 * If the bucket is full, toss the oldest element.
385 	 */
386 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
387 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
388 			("sch->sch_length incorrect"));
389 		syncache_pause(&sc->sc_inc);
390 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
391 		sch->sch_last_overflow = time_uptime;
392 		syncache_drop(sc2, sch);
393 	}
394 
395 	/* Put it into the bucket. */
396 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
397 	sch->sch_length++;
398 
399 #ifdef TCP_OFFLOAD
400 	if (ADDED_BY_TOE(sc)) {
401 		struct toedev *tod = sc->sc_tod;
402 
403 		tod->tod_syncache_added(tod, sc->sc_todctx);
404 	}
405 #endif
406 
407 	/* Reinitialize the bucket row's timer. */
408 	if (sch->sch_length == 1)
409 		sch->sch_nextc = ticks + INT_MAX;
410 	syncache_timeout(sc, sch, 1);
411 
412 	SCH_UNLOCK(sch);
413 
414 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
415 	TCPSTAT_INC(tcps_sc_added);
416 }
417 
418 /*
419  * Remove and free entry from syncache bucket row.
420  * Expects locked syncache head.
421  */
422 static void
423 syncache_drop(struct syncache *sc, struct syncache_head *sch)
424 {
425 
426 	SCH_LOCK_ASSERT(sch);
427 
428 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
429 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
430 	sch->sch_length--;
431 
432 #ifdef TCP_OFFLOAD
433 	if (ADDED_BY_TOE(sc)) {
434 		struct toedev *tod = sc->sc_tod;
435 
436 		tod->tod_syncache_removed(tod, sc->sc_todctx);
437 	}
438 #endif
439 
440 	syncache_free(sc);
441 }
442 
443 /*
444  * Engage/reengage time on bucket row.
445  */
446 static void
447 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
448 {
449 	int rexmt;
450 
451 	if (sc->sc_rxmits == 0)
452 		rexmt = tcp_rexmit_initial;
453 	else
454 		TCPT_RANGESET(rexmt,
455 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
456 		    tcp_rexmit_min, TCPTV_REXMTMAX);
457 	sc->sc_rxttime = ticks + rexmt;
458 	sc->sc_rxmits++;
459 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
460 		sch->sch_nextc = sc->sc_rxttime;
461 		if (docallout)
462 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
463 			    syncache_timer, (void *)sch);
464 	}
465 }
466 
467 /*
468  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
469  * If we have retransmitted an entry the maximum number of times, expire it.
470  * One separate timer for each bucket row.
471  */
472 static void
473 syncache_timer(void *xsch)
474 {
475 	struct syncache_head *sch = (struct syncache_head *)xsch;
476 	struct syncache *sc, *nsc;
477 	struct epoch_tracker et;
478 	int tick = ticks;
479 	char *s;
480 	bool paused;
481 
482 	CURVNET_SET(sch->sch_sc->vnet);
483 
484 	/* NB: syncache_head has already been locked by the callout. */
485 	SCH_LOCK_ASSERT(sch);
486 
487 	/*
488 	 * In the following cycle we may remove some entries and/or
489 	 * advance some timeouts, so re-initialize the bucket timer.
490 	 */
491 	sch->sch_nextc = tick + INT_MAX;
492 
493 	/*
494 	 * If we have paused processing, unconditionally remove
495 	 * all syncache entries.
496 	 */
497 	mtx_lock(&V_tcp_syncache.pause_mtx);
498 	paused = V_tcp_syncache.paused;
499 	mtx_unlock(&V_tcp_syncache.pause_mtx);
500 
501 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
502 		if (paused) {
503 			syncache_drop(sc, sch);
504 			continue;
505 		}
506 		/*
507 		 * We do not check if the listen socket still exists
508 		 * and accept the case where the listen socket may be
509 		 * gone by the time we resend the SYN/ACK.  We do
510 		 * not expect this to happens often. If it does,
511 		 * then the RST will be sent by the time the remote
512 		 * host does the SYN/ACK->ACK.
513 		 */
514 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
515 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
516 				sch->sch_nextc = sc->sc_rxttime;
517 			continue;
518 		}
519 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
520 			sc->sc_flags &= ~SCF_ECN;
521 		}
522 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
523 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
524 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
525 				    "giving up and removing syncache entry\n",
526 				    s, __func__);
527 				free(s, M_TCPLOG);
528 			}
529 			syncache_drop(sc, sch);
530 			TCPSTAT_INC(tcps_sc_stale);
531 			continue;
532 		}
533 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
534 			log(LOG_DEBUG, "%s; %s: Response timeout, "
535 			    "retransmitting (%u) SYN|ACK\n",
536 			    s, __func__, sc->sc_rxmits);
537 			free(s, M_TCPLOG);
538 		}
539 
540 		NET_EPOCH_ENTER(et);
541 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
542 		NET_EPOCH_EXIT(et);
543 		TCPSTAT_INC(tcps_sc_retransmitted);
544 		syncache_timeout(sc, sch, 0);
545 	}
546 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
547 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
548 			syncache_timer, (void *)(sch));
549 	CURVNET_RESTORE();
550 }
551 
552 /*
553  * Returns true if the system is only using cookies at the moment.
554  * This could be due to a sysadmin decision to only use cookies, or it
555  * could be due to the system detecting an attack.
556  */
557 static inline bool
558 syncache_cookiesonly(void)
559 {
560 
561 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
562 	    V_tcp_syncookiesonly));
563 }
564 
565 /*
566  * Find the hash bucket for the given connection.
567  */
568 static struct syncache_head *
569 syncache_hashbucket(struct in_conninfo *inc)
570 {
571 	uint32_t hash;
572 
573 	/*
574 	 * The hash is built on foreign port + local port + foreign address.
575 	 * We rely on the fact that struct in_conninfo starts with 16 bits
576 	 * of foreign port, then 16 bits of local port then followed by 128
577 	 * bits of foreign address.  In case of IPv4 address, the first 3
578 	 * 32-bit words of the address always are zeroes.
579 	 */
580 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
581 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
582 
583 	return (&V_tcp_syncache.hashbase[hash]);
584 }
585 
586 /*
587  * Find an entry in the syncache.
588  * Returns always with locked syncache_head plus a matching entry or NULL.
589  */
590 static struct syncache *
591 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
592 {
593 	struct syncache *sc;
594 	struct syncache_head *sch;
595 
596 	*schp = sch = syncache_hashbucket(inc);
597 	SCH_LOCK(sch);
598 
599 	/* Circle through bucket row to find matching entry. */
600 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
601 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
602 		    sizeof(struct in_endpoints)) == 0)
603 			break;
604 
605 	return (sc);	/* Always returns with locked sch. */
606 }
607 
608 /*
609  * This function is called when we get a RST for a
610  * non-existent connection, so that we can see if the
611  * connection is in the syn cache.  If it is, zap it.
612  * If required send a challenge ACK.
613  */
614 void
615 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
616     uint16_t port)
617 {
618 	struct syncache *sc;
619 	struct syncache_head *sch;
620 	char *s = NULL;
621 
622 	if (syncache_cookiesonly())
623 		return;
624 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
625 	SCH_LOCK_ASSERT(sch);
626 
627 	/*
628 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
629 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
630 	 */
631 	if (tcp_get_flags(th) & (TH_ACK|TH_SYN|TH_FIN)) {
632 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
633 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
634 			    "FIN flag set, segment ignored\n", s, __func__);
635 		TCPSTAT_INC(tcps_badrst);
636 		goto done;
637 	}
638 
639 	/*
640 	 * No corresponding connection was found in syncache.
641 	 * If syncookies are enabled and possibly exclusively
642 	 * used, or we are under memory pressure, a valid RST
643 	 * may not find a syncache entry.  In that case we're
644 	 * done and no SYN|ACK retransmissions will happen.
645 	 * Otherwise the RST was misdirected or spoofed.
646 	 */
647 	if (sc == NULL) {
648 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
649 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
650 			    "syncache entry (possibly syncookie only), "
651 			    "segment ignored\n", s, __func__);
652 		TCPSTAT_INC(tcps_badrst);
653 		goto done;
654 	}
655 
656 	/* The remote UDP encaps port does not match. */
657 	if (sc->sc_port != port) {
658 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
659 			log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
660 			    "syncache entry but non-matching UDP encaps port, "
661 			    "segment ignored\n", s, __func__);
662 		TCPSTAT_INC(tcps_badrst);
663 		goto done;
664 	}
665 
666 	/*
667 	 * If the RST bit is set, check the sequence number to see
668 	 * if this is a valid reset segment.
669 	 *
670 	 * RFC 793 page 37:
671 	 *   In all states except SYN-SENT, all reset (RST) segments
672 	 *   are validated by checking their SEQ-fields.  A reset is
673 	 *   valid if its sequence number is in the window.
674 	 *
675 	 * RFC 793 page 69:
676 	 *   There are four cases for the acceptability test for an incoming
677 	 *   segment:
678 	 *
679 	 * Segment Receive  Test
680 	 * Length  Window
681 	 * ------- -------  -------------------------------------------
682 	 *    0       0     SEG.SEQ = RCV.NXT
683 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
684 	 *   >0       0     not acceptable
685 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
686 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
687 	 *
688 	 * Note that when receiving a SYN segment in the LISTEN state,
689 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
690 	 * described in RFC 793, page 66.
691 	 */
692 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
693 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
694 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
695 		if (V_tcp_insecure_rst ||
696 		    th->th_seq == sc->sc_irs + 1) {
697 			syncache_drop(sc, sch);
698 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
699 				log(LOG_DEBUG,
700 				    "%s; %s: Our SYN|ACK was rejected, "
701 				    "connection attempt aborted by remote "
702 				    "endpoint\n",
703 				    s, __func__);
704 			TCPSTAT_INC(tcps_sc_reset);
705 		} else {
706 			TCPSTAT_INC(tcps_badrst);
707 			/* Send challenge ACK. */
708 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
709 				log(LOG_DEBUG, "%s; %s: RST with invalid "
710 				    " SEQ %u != NXT %u (+WND %u), "
711 				    "sending challenge ACK\n",
712 				    s, __func__,
713 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
714 			syncache_respond(sc, m, TH_ACK);
715 		}
716 	} else {
717 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
718 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
719 			    "NXT %u (+WND %u), segment ignored\n",
720 			    s, __func__,
721 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
722 		TCPSTAT_INC(tcps_badrst);
723 	}
724 
725 done:
726 	if (s != NULL)
727 		free(s, M_TCPLOG);
728 	SCH_UNLOCK(sch);
729 }
730 
731 void
732 syncache_badack(struct in_conninfo *inc, uint16_t port)
733 {
734 	struct syncache *sc;
735 	struct syncache_head *sch;
736 
737 	if (syncache_cookiesonly())
738 		return;
739 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
740 	SCH_LOCK_ASSERT(sch);
741 	if ((sc != NULL) && (sc->sc_port == port)) {
742 		syncache_drop(sc, sch);
743 		TCPSTAT_INC(tcps_sc_badack);
744 	}
745 	SCH_UNLOCK(sch);
746 }
747 
748 void
749 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
750 {
751 	struct syncache *sc;
752 	struct syncache_head *sch;
753 
754 	if (syncache_cookiesonly())
755 		return;
756 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
757 	SCH_LOCK_ASSERT(sch);
758 	if (sc == NULL)
759 		goto done;
760 
761 	/* If the port != sc_port, then it's a bogus ICMP msg */
762 	if (port != sc->sc_port)
763 		goto done;
764 
765 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
766 	if (ntohl(th_seq) != sc->sc_iss)
767 		goto done;
768 
769 	/*
770 	 * If we've rertransmitted 3 times and this is our second error,
771 	 * we remove the entry.  Otherwise, we allow it to continue on.
772 	 * This prevents us from incorrectly nuking an entry during a
773 	 * spurious network outage.
774 	 *
775 	 * See tcp_notify().
776 	 */
777 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
778 		sc->sc_flags |= SCF_UNREACH;
779 		goto done;
780 	}
781 	syncache_drop(sc, sch);
782 	TCPSTAT_INC(tcps_sc_unreach);
783 done:
784 	SCH_UNLOCK(sch);
785 }
786 
787 /*
788  * Build a new TCP socket structure from a syncache entry.
789  *
790  * On success return the newly created socket with its underlying inp locked.
791  */
792 static struct socket *
793 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
794 {
795 	struct tcp_function_block *blk;
796 	struct inpcb *inp = NULL;
797 	struct socket *so;
798 	struct tcpcb *tp;
799 	int error;
800 	char *s;
801 
802 	NET_EPOCH_ASSERT();
803 
804 	/*
805 	 * Ok, create the full blown connection, and set things up
806 	 * as they would have been set up if we had created the
807 	 * connection when the SYN arrived.  If we can't create
808 	 * the connection, abort it.
809 	 */
810 	so = sonewconn(lso, 0);
811 	if (so == NULL) {
812 		/*
813 		 * Drop the connection; we will either send a RST or
814 		 * have the peer retransmit its SYN again after its
815 		 * RTO and try again.
816 		 */
817 		TCPSTAT_INC(tcps_listendrop);
818 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
819 			log(LOG_DEBUG, "%s; %s: Socket create failed "
820 			    "due to limits or memory shortage\n",
821 			    s, __func__);
822 			free(s, M_TCPLOG);
823 		}
824 		goto abort2;
825 	}
826 #ifdef MAC
827 	mac_socketpeer_set_from_mbuf(m, so);
828 #endif
829 
830 	inp = sotoinpcb(so);
831 	inp->inp_inc.inc_fibnum = so->so_fibnum;
832 	INP_WLOCK(inp);
833 	/*
834 	 * Exclusive pcbinfo lock is not required in syncache socket case even
835 	 * if two inpcb locks can be acquired simultaneously:
836 	 *  - the inpcb in LISTEN state,
837 	 *  - the newly created inp.
838 	 *
839 	 * In this case, an inp cannot be at same time in LISTEN state and
840 	 * just created by an accept() call.
841 	 */
842 	INP_HASH_WLOCK(&V_tcbinfo);
843 
844 	/* Insert new socket into PCB hash list. */
845 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
846 #ifdef INET6
847 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
848 		inp->inp_vflag &= ~INP_IPV4;
849 		inp->inp_vflag |= INP_IPV6;
850 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
851 	} else {
852 		inp->inp_vflag &= ~INP_IPV6;
853 		inp->inp_vflag |= INP_IPV4;
854 #endif
855 		inp->inp_ip_ttl = sc->sc_ip_ttl;
856 		inp->inp_ip_tos = sc->sc_ip_tos;
857 		inp->inp_laddr = sc->sc_inc.inc_laddr;
858 #ifdef INET6
859 	}
860 #endif
861 
862 	/*
863 	 * If there's an mbuf and it has a flowid, then let's initialise the
864 	 * inp with that particular flowid.
865 	 */
866 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
867 		inp->inp_flowid = m->m_pkthdr.flowid;
868 		inp->inp_flowtype = M_HASHTYPE_GET(m);
869 #ifdef NUMA
870 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
871 #endif
872 	}
873 
874 	inp->inp_lport = sc->sc_inc.inc_lport;
875 #ifdef INET6
876 	if (inp->inp_vflag & INP_IPV6PROTO) {
877 		struct inpcb *oinp = sotoinpcb(lso);
878 
879 		/*
880 		 * Inherit socket options from the listening socket.
881 		 * Note that in6p_inputopts are not (and should not be)
882 		 * copied, since it stores previously received options and is
883 		 * used to detect if each new option is different than the
884 		 * previous one and hence should be passed to a user.
885 		 * If we copied in6p_inputopts, a user would not be able to
886 		 * receive options just after calling the accept system call.
887 		 */
888 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
889 		if (oinp->in6p_outputopts)
890 			inp->in6p_outputopts =
891 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
892 		inp->in6p_hops = oinp->in6p_hops;
893 	}
894 
895 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
896 		struct in6_addr laddr6;
897 		struct sockaddr_in6 sin6;
898 
899 		sin6.sin6_family = AF_INET6;
900 		sin6.sin6_len = sizeof(sin6);
901 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
902 		sin6.sin6_port = sc->sc_inc.inc_fport;
903 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
904 		laddr6 = inp->in6p_laddr;
905 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
906 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
907 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
908 		    thread0.td_ucred, m, false)) != 0) {
909 			inp->in6p_laddr = laddr6;
910 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
911 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
912 				    "with error %i\n",
913 				    s, __func__, error);
914 				free(s, M_TCPLOG);
915 			}
916 			INP_HASH_WUNLOCK(&V_tcbinfo);
917 			goto abort;
918 		}
919 		/* Override flowlabel from in6_pcbconnect. */
920 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
921 		inp->inp_flow |= sc->sc_flowlabel;
922 	}
923 #endif /* INET6 */
924 #if defined(INET) && defined(INET6)
925 	else
926 #endif
927 #ifdef INET
928 	{
929 		struct in_addr laddr;
930 		struct sockaddr_in sin;
931 
932 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
933 
934 		if (inp->inp_options == NULL) {
935 			inp->inp_options = sc->sc_ipopts;
936 			sc->sc_ipopts = NULL;
937 		}
938 
939 		sin.sin_family = AF_INET;
940 		sin.sin_len = sizeof(sin);
941 		sin.sin_addr = sc->sc_inc.inc_faddr;
942 		sin.sin_port = sc->sc_inc.inc_fport;
943 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
944 		laddr = inp->inp_laddr;
945 		if (inp->inp_laddr.s_addr == INADDR_ANY)
946 			inp->inp_laddr = sc->sc_inc.inc_laddr;
947 		if ((error = in_pcbconnect(inp, (struct sockaddr *)&sin,
948 		    thread0.td_ucred, false)) != 0) {
949 			inp->inp_laddr = laddr;
950 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
951 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
952 				    "with error %i\n",
953 				    s, __func__, error);
954 				free(s, M_TCPLOG);
955 			}
956 			INP_HASH_WUNLOCK(&V_tcbinfo);
957 			goto abort;
958 		}
959 	}
960 #endif /* INET */
961 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
962 	/* Copy old policy into new socket's. */
963 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
964 		printf("syncache_socket: could not copy policy\n");
965 #endif
966 	INP_HASH_WUNLOCK(&V_tcbinfo);
967 	tp = intotcpcb(inp);
968 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
969 	tp->iss = sc->sc_iss;
970 	tp->irs = sc->sc_irs;
971 	tp->t_port = sc->sc_port;
972 	tcp_rcvseqinit(tp);
973 	tcp_sendseqinit(tp);
974 	blk = sototcpcb(lso)->t_fb;
975 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
976 		/*
977 		 * Our parents t_fb was not the default,
978 		 * we need to release our ref on tp->t_fb and
979 		 * pickup one on the new entry.
980 		 */
981 		struct tcp_function_block *rblk;
982 
983 		rblk = find_and_ref_tcp_fb(blk);
984 		KASSERT(rblk != NULL,
985 		    ("cannot find blk %p out of syncache?", blk));
986 		if (tp->t_fb->tfb_tcp_fb_fini)
987 			(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
988 		refcount_release(&tp->t_fb->tfb_refcnt);
989 		tp->t_fb = rblk;
990 		/*
991 		 * XXXrrs this is quite dangerous, it is possible
992 		 * for the new function to fail to init. We also
993 		 * are not asking if the handoff_is_ok though at
994 		 * the very start thats probalbly ok.
995 		 */
996 		if (tp->t_fb->tfb_tcp_fb_init) {
997 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
998 		}
999 	}
1000 	tp->snd_wl1 = sc->sc_irs;
1001 	tp->snd_max = tp->iss + 1;
1002 	tp->snd_nxt = tp->iss + 1;
1003 	tp->rcv_up = sc->sc_irs + 1;
1004 	tp->rcv_wnd = sc->sc_wnd;
1005 	tp->rcv_adv += tp->rcv_wnd;
1006 	tp->last_ack_sent = tp->rcv_nxt;
1007 
1008 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
1009 	if (sc->sc_flags & SCF_NOOPT)
1010 		tp->t_flags |= TF_NOOPT;
1011 	else {
1012 		if (sc->sc_flags & SCF_WINSCALE) {
1013 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
1014 			tp->snd_scale = sc->sc_requested_s_scale;
1015 			tp->request_r_scale = sc->sc_requested_r_scale;
1016 		}
1017 		if (sc->sc_flags & SCF_TIMESTAMP) {
1018 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
1019 			tp->ts_recent = sc->sc_tsreflect;
1020 			tp->ts_recent_age = tcp_ts_getticks();
1021 			tp->ts_offset = sc->sc_tsoff;
1022 		}
1023 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1024 		if (sc->sc_flags & SCF_SIGNATURE)
1025 			tp->t_flags |= TF_SIGNATURE;
1026 #endif
1027 		if (sc->sc_flags & SCF_SACK)
1028 			tp->t_flags |= TF_SACK_PERMIT;
1029 	}
1030 
1031 	tcp_ecn_syncache_socket(tp, sc);
1032 
1033 	/*
1034 	 * Set up MSS and get cached values from tcp_hostcache.
1035 	 * This might overwrite some of the defaults we just set.
1036 	 */
1037 	tcp_mss(tp, sc->sc_peer_mss);
1038 
1039 	/*
1040 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
1041 	 * limited to one segment in cc_conn_init().
1042 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1043 	 */
1044 	if (sc->sc_rxmits > 1)
1045 		tp->snd_cwnd = 1;
1046 
1047 #ifdef TCP_OFFLOAD
1048 	/*
1049 	 * Allow a TOE driver to install its hooks.  Note that we hold the
1050 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1051 	 * new connection before the TOE driver has done its thing.
1052 	 */
1053 	if (ADDED_BY_TOE(sc)) {
1054 		struct toedev *tod = sc->sc_tod;
1055 
1056 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
1057 	}
1058 #endif
1059 	/*
1060 	 * Copy and activate timers.
1061 	 */
1062 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1063 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1064 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1065 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1066 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1067 
1068 	TCPSTAT_INC(tcps_accepts);
1069 	return (so);
1070 
1071 abort:
1072 	INP_WUNLOCK(inp);
1073 abort2:
1074 	if (so != NULL)
1075 		soabort(so);
1076 	return (NULL);
1077 }
1078 
1079 /*
1080  * This function gets called when we receive an ACK for a
1081  * socket in the LISTEN state.  We look up the connection
1082  * in the syncache, and if its there, we pull it out of
1083  * the cache and turn it into a full-blown connection in
1084  * the SYN-RECEIVED state.
1085  *
1086  * On syncache_socket() success the newly created socket
1087  * has its underlying inp locked.
1088  */
1089 int
1090 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1091     struct socket **lsop, struct mbuf *m, uint16_t port)
1092 {
1093 	struct syncache *sc;
1094 	struct syncache_head *sch;
1095 	struct syncache scs;
1096 	char *s;
1097 	bool locked;
1098 
1099 	NET_EPOCH_ASSERT();
1100 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1101 	    ("%s: can handle only ACK", __func__));
1102 
1103 	if (syncache_cookiesonly()) {
1104 		sc = NULL;
1105 		sch = syncache_hashbucket(inc);
1106 		locked = false;
1107 	} else {
1108 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1109 		locked = true;
1110 		SCH_LOCK_ASSERT(sch);
1111 	}
1112 
1113 #ifdef INVARIANTS
1114 	/*
1115 	 * Test code for syncookies comparing the syncache stored
1116 	 * values with the reconstructed values from the cookie.
1117 	 */
1118 	if (sc != NULL)
1119 		syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1120 #endif
1121 
1122 	if (sc == NULL) {
1123 		/*
1124 		 * There is no syncache entry, so see if this ACK is
1125 		 * a returning syncookie.  To do this, first:
1126 		 *  A. Check if syncookies are used in case of syncache
1127 		 *     overflows
1128 		 *  B. See if this socket has had a syncache entry dropped in
1129 		 *     the recent past. We don't want to accept a bogus
1130 		 *     syncookie if we've never received a SYN or accept it
1131 		 *     twice.
1132 		 *  C. check that the syncookie is valid.  If it is, then
1133 		 *     cobble up a fake syncache entry, and return.
1134 		 */
1135 		if (locked && !V_tcp_syncookies) {
1136 			SCH_UNLOCK(sch);
1137 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1138 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1139 				    "segment rejected (syncookies disabled)\n",
1140 				    s, __func__);
1141 			goto failed;
1142 		}
1143 		if (locked && !V_tcp_syncookiesonly &&
1144 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1145 			SCH_UNLOCK(sch);
1146 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1147 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1148 				    "segment rejected (no syncache entry)\n",
1149 				    s, __func__);
1150 			goto failed;
1151 		}
1152 		bzero(&scs, sizeof(scs));
1153 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1154 		if (locked)
1155 			SCH_UNLOCK(sch);
1156 		if (sc == NULL) {
1157 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1158 				log(LOG_DEBUG, "%s; %s: Segment failed "
1159 				    "SYNCOOKIE authentication, segment rejected "
1160 				    "(probably spoofed)\n", s, __func__);
1161 			goto failed;
1162 		}
1163 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1164 		/* If received ACK has MD5 signature, check it. */
1165 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1166 		    (!TCPMD5_ENABLED() ||
1167 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1168 			/* Drop the ACK. */
1169 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1170 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1171 				    "MD5 signature doesn't match.\n",
1172 				    s, __func__);
1173 				free(s, M_TCPLOG);
1174 			}
1175 			TCPSTAT_INC(tcps_sig_err_sigopt);
1176 			return (-1); /* Do not send RST */
1177 		}
1178 #endif /* TCP_SIGNATURE */
1179 	} else {
1180 		if (sc->sc_port != port) {
1181 			SCH_UNLOCK(sch);
1182 			return (0);
1183 		}
1184 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1185 		/*
1186 		 * If listening socket requested TCP digests, check that
1187 		 * received ACK has signature and it is correct.
1188 		 * If not, drop the ACK and leave sc entry in th cache,
1189 		 * because SYN was received with correct signature.
1190 		 */
1191 		if (sc->sc_flags & SCF_SIGNATURE) {
1192 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1193 				/* No signature */
1194 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1195 				SCH_UNLOCK(sch);
1196 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1197 					log(LOG_DEBUG, "%s; %s: Segment "
1198 					    "rejected, MD5 signature wasn't "
1199 					    "provided.\n", s, __func__);
1200 					free(s, M_TCPLOG);
1201 				}
1202 				return (-1); /* Do not send RST */
1203 			}
1204 			if (!TCPMD5_ENABLED() ||
1205 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1206 				/* Doesn't match or no SA */
1207 				SCH_UNLOCK(sch);
1208 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1209 					log(LOG_DEBUG, "%s; %s: Segment "
1210 					    "rejected, MD5 signature doesn't "
1211 					    "match.\n", s, __func__);
1212 					free(s, M_TCPLOG);
1213 				}
1214 				return (-1); /* Do not send RST */
1215 			}
1216 		}
1217 #endif /* TCP_SIGNATURE */
1218 
1219 		/*
1220 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1221 		 * it's less than ts_recent, drop it.
1222 		 * XXXMT: RFC 7323 also requires to send an ACK.
1223 		 *        In tcp_input.c this is only done for TCP segments
1224 		 *        with user data, so be consistent here and just drop
1225 		 *        the segment.
1226 		 */
1227 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1228 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1229 			SCH_UNLOCK(sch);
1230 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1231 				log(LOG_DEBUG,
1232 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1233 				    "segment dropped\n", s, __func__,
1234 				    to->to_tsval, sc->sc_tsreflect);
1235 				free(s, M_TCPLOG);
1236 			}
1237 			return (-1);  /* Do not send RST */
1238 		}
1239 
1240 		/*
1241 		 * If timestamps were not negotiated during SYN/ACK and a
1242 		 * segment with a timestamp is received, ignore the
1243 		 * timestamp and process the packet normally.
1244 		 * See section 3.2 of RFC 7323.
1245 		 */
1246 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1247 		    (to->to_flags & TOF_TS)) {
1248 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1249 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1250 				    "expected, segment processed normally\n",
1251 				    s, __func__);
1252 				free(s, M_TCPLOG);
1253 				s = NULL;
1254 			}
1255 		}
1256 
1257 		/*
1258 		 * If timestamps were negotiated during SYN/ACK and a
1259 		 * segment without a timestamp is received, silently drop
1260 		 * the segment, unless the missing timestamps are tolerated.
1261 		 * See section 3.2 of RFC 7323.
1262 		 */
1263 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1264 		    !(to->to_flags & TOF_TS)) {
1265 			if (V_tcp_tolerate_missing_ts) {
1266 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1267 					log(LOG_DEBUG,
1268 					    "%s; %s: Timestamp missing, "
1269 					    "segment processed normally\n",
1270 					    s, __func__);
1271 					free(s, M_TCPLOG);
1272 				}
1273 			} else {
1274 				SCH_UNLOCK(sch);
1275 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1276 					log(LOG_DEBUG,
1277 					    "%s; %s: Timestamp missing, "
1278 					    "segment silently dropped\n",
1279 					    s, __func__);
1280 					free(s, M_TCPLOG);
1281 				}
1282 				return (-1);  /* Do not send RST */
1283 			}
1284 		}
1285 
1286 		/*
1287 		 * Pull out the entry to unlock the bucket row.
1288 		 *
1289 		 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1290 		 * tcp_state_change().  The tcpcb is not existent at this
1291 		 * moment.  A new one will be allocated via syncache_socket->
1292 		 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1293 		 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1294 		 */
1295 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1296 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1297 		sch->sch_length--;
1298 #ifdef TCP_OFFLOAD
1299 		if (ADDED_BY_TOE(sc)) {
1300 			struct toedev *tod = sc->sc_tod;
1301 
1302 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1303 		}
1304 #endif
1305 		SCH_UNLOCK(sch);
1306 	}
1307 
1308 	/*
1309 	 * Segment validation:
1310 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1311 	 */
1312 	if (th->th_ack != sc->sc_iss + 1) {
1313 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1314 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1315 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1316 		goto failed;
1317 	}
1318 
1319 	/*
1320 	 * The SEQ must fall in the window starting at the received
1321 	 * initial receive sequence number + 1 (the SYN).
1322 	 */
1323 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1324 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1325 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1326 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1327 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1328 		goto failed;
1329 	}
1330 
1331 	*lsop = syncache_socket(sc, *lsop, m);
1332 
1333 	if (*lsop == NULL)
1334 		TCPSTAT_INC(tcps_sc_aborted);
1335 	else
1336 		TCPSTAT_INC(tcps_sc_completed);
1337 
1338 /* how do we find the inp for the new socket? */
1339 	if (sc != &scs)
1340 		syncache_free(sc);
1341 	return (1);
1342 failed:
1343 	if (sc != NULL && sc != &scs)
1344 		syncache_free(sc);
1345 	if (s != NULL)
1346 		free(s, M_TCPLOG);
1347 	*lsop = NULL;
1348 	return (0);
1349 }
1350 
1351 static struct socket *
1352 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1353     uint64_t response_cookie)
1354 {
1355 	struct inpcb *inp;
1356 	struct tcpcb *tp;
1357 	unsigned int *pending_counter;
1358 	struct socket *so;
1359 
1360 	NET_EPOCH_ASSERT();
1361 
1362 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1363 	so = syncache_socket(sc, lso, m);
1364 	if (so == NULL) {
1365 		TCPSTAT_INC(tcps_sc_aborted);
1366 		atomic_subtract_int(pending_counter, 1);
1367 	} else {
1368 		soisconnected(so);
1369 		inp = sotoinpcb(so);
1370 		tp = intotcpcb(inp);
1371 		tp->t_flags |= TF_FASTOPEN;
1372 		tp->t_tfo_cookie.server = response_cookie;
1373 		tp->snd_max = tp->iss;
1374 		tp->snd_nxt = tp->iss;
1375 		tp->t_tfo_pending = pending_counter;
1376 		TCPSTAT_INC(tcps_sc_completed);
1377 	}
1378 
1379 	return (so);
1380 }
1381 
1382 /*
1383  * Given a LISTEN socket and an inbound SYN request, add
1384  * this to the syn cache, and send back a segment:
1385  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1386  * to the source.
1387  *
1388  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1389  * Doing so would require that we hold onto the data and deliver it
1390  * to the application.  However, if we are the target of a SYN-flood
1391  * DoS attack, an attacker could send data which would eventually
1392  * consume all available buffer space if it were ACKed.  By not ACKing
1393  * the data, we avoid this DoS scenario.
1394  *
1395  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1396  * cookie is processed and a new socket is created.  In this case, any data
1397  * accompanying the SYN will be queued to the socket by tcp_input() and will
1398  * be ACKed either when the application sends response data or the delayed
1399  * ACK timer expires, whichever comes first.
1400  */
1401 struct socket *
1402 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1403     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1404     void *todctx, uint8_t iptos, uint16_t port)
1405 {
1406 	struct tcpcb *tp;
1407 	struct socket *rv = NULL;
1408 	struct syncache *sc = NULL;
1409 	struct syncache_head *sch;
1410 	struct mbuf *ipopts = NULL;
1411 	u_int ltflags;
1412 	int win, ip_ttl, ip_tos;
1413 	char *s;
1414 #ifdef INET6
1415 	int autoflowlabel = 0;
1416 #endif
1417 #ifdef MAC
1418 	struct label *maclabel;
1419 #endif
1420 	struct syncache scs;
1421 	struct ucred *cred;
1422 	uint64_t tfo_response_cookie;
1423 	unsigned int *tfo_pending = NULL;
1424 	int tfo_cookie_valid = 0;
1425 	int tfo_response_cookie_valid = 0;
1426 	bool locked;
1427 
1428 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1429 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1430 	    ("%s: unexpected tcp flags", __func__));
1431 
1432 	/*
1433 	 * Combine all so/tp operations very early to drop the INP lock as
1434 	 * soon as possible.
1435 	 */
1436 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1437 	tp = sototcpcb(so);
1438 	cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1439 
1440 #ifdef INET6
1441 	if (inc->inc_flags & INC_ISIPV6) {
1442 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1443 			autoflowlabel = 1;
1444 		}
1445 		ip_ttl = in6_selecthlim(inp, NULL);
1446 		if ((inp->in6p_outputopts == NULL) ||
1447 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1448 			ip_tos = 0;
1449 		} else {
1450 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1451 		}
1452 	}
1453 #endif
1454 #if defined(INET6) && defined(INET)
1455 	else
1456 #endif
1457 #ifdef INET
1458 	{
1459 		ip_ttl = inp->inp_ip_ttl;
1460 		ip_tos = inp->inp_ip_tos;
1461 	}
1462 #endif
1463 	win = so->sol_sbrcv_hiwat;
1464 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1465 
1466 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1467 	    (tp->t_tfo_pending != NULL) &&
1468 	    (to->to_flags & TOF_FASTOPEN)) {
1469 		/*
1470 		 * Limit the number of pending TFO connections to
1471 		 * approximately half of the queue limit.  This prevents TFO
1472 		 * SYN floods from starving the service by filling the
1473 		 * listen queue with bogus TFO connections.
1474 		 */
1475 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1476 		    (so->sol_qlimit / 2)) {
1477 			int result;
1478 
1479 			result = tcp_fastopen_check_cookie(inc,
1480 			    to->to_tfo_cookie, to->to_tfo_len,
1481 			    &tfo_response_cookie);
1482 			tfo_cookie_valid = (result > 0);
1483 			tfo_response_cookie_valid = (result >= 0);
1484 		}
1485 
1486 		/*
1487 		 * Remember the TFO pending counter as it will have to be
1488 		 * decremented below if we don't make it to syncache_tfo_expand().
1489 		 */
1490 		tfo_pending = tp->t_tfo_pending;
1491 	}
1492 
1493 #ifdef MAC
1494 	if (mac_syncache_init(&maclabel) != 0) {
1495 		INP_RUNLOCK(inp);
1496 		goto done;
1497 	} else
1498 		mac_syncache_create(maclabel, inp);
1499 #endif
1500 	if (!tfo_cookie_valid)
1501 		INP_RUNLOCK(inp);
1502 
1503 	/*
1504 	 * Remember the IP options, if any.
1505 	 */
1506 #ifdef INET6
1507 	if (!(inc->inc_flags & INC_ISIPV6))
1508 #endif
1509 #ifdef INET
1510 		ipopts = (m) ? ip_srcroute(m) : NULL;
1511 #else
1512 		ipopts = NULL;
1513 #endif
1514 
1515 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1516 	/*
1517 	 * When the socket is TCP-MD5 enabled check that,
1518 	 *  - a signed packet is valid
1519 	 *  - a non-signed packet does not have a security association
1520 	 *
1521 	 *  If a signed packet fails validation or a non-signed packet has a
1522 	 *  security association, the packet will be dropped.
1523 	 */
1524 	if (ltflags & TF_SIGNATURE) {
1525 		if (to->to_flags & TOF_SIGNATURE) {
1526 			if (!TCPMD5_ENABLED() ||
1527 			    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1528 				goto done;
1529 		} else {
1530 			if (TCPMD5_ENABLED() &&
1531 			    TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1532 				goto done;
1533 		}
1534 	} else if (to->to_flags & TOF_SIGNATURE)
1535 		goto done;
1536 #endif	/* TCP_SIGNATURE */
1537 	/*
1538 	 * See if we already have an entry for this connection.
1539 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1540 	 *
1541 	 * XXX: should the syncache be re-initialized with the contents
1542 	 * of the new SYN here (which may have different options?)
1543 	 *
1544 	 * XXX: We do not check the sequence number to see if this is a
1545 	 * real retransmit or a new connection attempt.  The question is
1546 	 * how to handle such a case; either ignore it as spoofed, or
1547 	 * drop the current entry and create a new one?
1548 	 */
1549 	if (syncache_cookiesonly()) {
1550 		sc = NULL;
1551 		sch = syncache_hashbucket(inc);
1552 		locked = false;
1553 	} else {
1554 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1555 		locked = true;
1556 		SCH_LOCK_ASSERT(sch);
1557 	}
1558 	if (sc != NULL) {
1559 		if (tfo_cookie_valid)
1560 			INP_RUNLOCK(inp);
1561 		TCPSTAT_INC(tcps_sc_dupsyn);
1562 		if (ipopts) {
1563 			/*
1564 			 * If we were remembering a previous source route,
1565 			 * forget it and use the new one we've been given.
1566 			 */
1567 			if (sc->sc_ipopts)
1568 				(void) m_free(sc->sc_ipopts);
1569 			sc->sc_ipopts = ipopts;
1570 		}
1571 		/*
1572 		 * Update timestamp if present.
1573 		 */
1574 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1575 			sc->sc_tsreflect = to->to_tsval;
1576 		else
1577 			sc->sc_flags &= ~SCF_TIMESTAMP;
1578 		/*
1579 		 * Disable ECN if needed.
1580 		 */
1581 		if ((sc->sc_flags & SCF_ECN) &&
1582 		    ((tcp_get_flags(th) & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) {
1583 			sc->sc_flags &= ~SCF_ECN;
1584 		}
1585 #ifdef MAC
1586 		/*
1587 		 * Since we have already unconditionally allocated label
1588 		 * storage, free it up.  The syncache entry will already
1589 		 * have an initialized label we can use.
1590 		 */
1591 		mac_syncache_destroy(&maclabel);
1592 #endif
1593 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1594 		/* Retransmit SYN|ACK and reset retransmit count. */
1595 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1596 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1597 			    "resetting timer and retransmitting SYN|ACK\n",
1598 			    s, __func__);
1599 			free(s, M_TCPLOG);
1600 		}
1601 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1602 			sc->sc_rxmits = 0;
1603 			syncache_timeout(sc, sch, 1);
1604 			TCPSTAT_INC(tcps_sndacks);
1605 			TCPSTAT_INC(tcps_sndtotal);
1606 		}
1607 		SCH_UNLOCK(sch);
1608 		goto donenoprobe;
1609 	}
1610 
1611 	if (tfo_cookie_valid) {
1612 		bzero(&scs, sizeof(scs));
1613 		sc = &scs;
1614 		goto skip_alloc;
1615 	}
1616 
1617 	/*
1618 	 * Skip allocating a syncache entry if we are just going to discard
1619 	 * it later.
1620 	 */
1621 	if (!locked) {
1622 		bzero(&scs, sizeof(scs));
1623 		sc = &scs;
1624 	} else
1625 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1626 	if (sc == NULL) {
1627 		/*
1628 		 * The zone allocator couldn't provide more entries.
1629 		 * Treat this as if the cache was full; drop the oldest
1630 		 * entry and insert the new one.
1631 		 */
1632 		TCPSTAT_INC(tcps_sc_zonefail);
1633 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1634 			sch->sch_last_overflow = time_uptime;
1635 			syncache_drop(sc, sch);
1636 			syncache_pause(inc);
1637 		}
1638 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1639 		if (sc == NULL) {
1640 			if (V_tcp_syncookies) {
1641 				bzero(&scs, sizeof(scs));
1642 				sc = &scs;
1643 			} else {
1644 				KASSERT(locked,
1645 				    ("%s: bucket unexpectedly unlocked",
1646 				    __func__));
1647 				SCH_UNLOCK(sch);
1648 				if (ipopts)
1649 					(void) m_free(ipopts);
1650 				goto done;
1651 			}
1652 		}
1653 	}
1654 
1655 skip_alloc:
1656 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1657 		sc->sc_tfo_cookie = &tfo_response_cookie;
1658 
1659 	/*
1660 	 * Fill in the syncache values.
1661 	 */
1662 #ifdef MAC
1663 	sc->sc_label = maclabel;
1664 #endif
1665 	sc->sc_cred = cred;
1666 	sc->sc_port = port;
1667 	cred = NULL;
1668 	sc->sc_ipopts = ipopts;
1669 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1670 	sc->sc_ip_tos = ip_tos;
1671 	sc->sc_ip_ttl = ip_ttl;
1672 #ifdef TCP_OFFLOAD
1673 	sc->sc_tod = tod;
1674 	sc->sc_todctx = todctx;
1675 #endif
1676 	sc->sc_irs = th->th_seq;
1677 	sc->sc_flags = 0;
1678 	sc->sc_flowlabel = 0;
1679 
1680 	/*
1681 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1682 	 * win was derived from socket earlier in the function.
1683 	 */
1684 	win = imax(win, 0);
1685 	win = imin(win, TCP_MAXWIN);
1686 	sc->sc_wnd = win;
1687 
1688 	if (V_tcp_do_rfc1323 &&
1689 	    !(ltflags & TF_NOOPT)) {
1690 		/*
1691 		 * A timestamp received in a SYN makes
1692 		 * it ok to send timestamp requests and replies.
1693 		 */
1694 		if (to->to_flags & TOF_TS) {
1695 			sc->sc_tsreflect = to->to_tsval;
1696 			sc->sc_flags |= SCF_TIMESTAMP;
1697 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1698 		}
1699 		if (to->to_flags & TOF_SCALE) {
1700 			int wscale = 0;
1701 
1702 			/*
1703 			 * Pick the smallest possible scaling factor that
1704 			 * will still allow us to scale up to sb_max, aka
1705 			 * kern.ipc.maxsockbuf.
1706 			 *
1707 			 * We do this because there are broken firewalls that
1708 			 * will corrupt the window scale option, leading to
1709 			 * the other endpoint believing that our advertised
1710 			 * window is unscaled.  At scale factors larger than
1711 			 * 5 the unscaled window will drop below 1500 bytes,
1712 			 * leading to serious problems when traversing these
1713 			 * broken firewalls.
1714 			 *
1715 			 * With the default maxsockbuf of 256K, a scale factor
1716 			 * of 3 will be chosen by this algorithm.  Those who
1717 			 * choose a larger maxsockbuf should watch out
1718 			 * for the compatibility problems mentioned above.
1719 			 *
1720 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1721 			 * or <SYN,ACK>) segment itself is never scaled.
1722 			 */
1723 			while (wscale < TCP_MAX_WINSHIFT &&
1724 			    (TCP_MAXWIN << wscale) < sb_max)
1725 				wscale++;
1726 			sc->sc_requested_r_scale = wscale;
1727 			sc->sc_requested_s_scale = to->to_wscale;
1728 			sc->sc_flags |= SCF_WINSCALE;
1729 		}
1730 	}
1731 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1732 	/*
1733 	 * If incoming packet has an MD5 signature, flag this in the
1734 	 * syncache so that syncache_respond() will do the right thing
1735 	 * with the SYN+ACK.
1736 	 */
1737 	if (to->to_flags & TOF_SIGNATURE)
1738 		sc->sc_flags |= SCF_SIGNATURE;
1739 #endif	/* TCP_SIGNATURE */
1740 	if (to->to_flags & TOF_SACKPERM)
1741 		sc->sc_flags |= SCF_SACK;
1742 	if (to->to_flags & TOF_MSS)
1743 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1744 	if (ltflags & TF_NOOPT)
1745 		sc->sc_flags |= SCF_NOOPT;
1746 	/* ECN Handshake */
1747 	if (V_tcp_do_ecn)
1748 		sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1749 
1750 	if (V_tcp_syncookies)
1751 		sc->sc_iss = syncookie_generate(sch, sc);
1752 	else
1753 		sc->sc_iss = arc4random();
1754 #ifdef INET6
1755 	if (autoflowlabel) {
1756 		if (V_tcp_syncookies)
1757 			sc->sc_flowlabel = sc->sc_iss;
1758 		else
1759 			sc->sc_flowlabel = ip6_randomflowlabel();
1760 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1761 	}
1762 #endif
1763 	if (locked)
1764 		SCH_UNLOCK(sch);
1765 
1766 	if (tfo_cookie_valid) {
1767 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1768 		/* INP_RUNLOCK(inp) will be performed by the caller */
1769 		goto tfo_expanded;
1770 	}
1771 
1772 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1773 	/*
1774 	 * Do a standard 3-way handshake.
1775 	 */
1776 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1777 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1778 			syncache_free(sc);
1779 		else if (sc != &scs)
1780 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1781 		TCPSTAT_INC(tcps_sndacks);
1782 		TCPSTAT_INC(tcps_sndtotal);
1783 	} else {
1784 		if (sc != &scs)
1785 			syncache_free(sc);
1786 		TCPSTAT_INC(tcps_sc_dropped);
1787 	}
1788 	goto donenoprobe;
1789 
1790 done:
1791 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1792 donenoprobe:
1793 	if (m)
1794 		m_freem(m);
1795 	/*
1796 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1797 	 * result in a new socket was processed and the associated pending
1798 	 * counter has not yet been decremented.  All such TFO processing paths
1799 	 * transit this point.
1800 	 */
1801 	if (tfo_pending != NULL)
1802 		tcp_fastopen_decrement_counter(tfo_pending);
1803 
1804 tfo_expanded:
1805 	if (cred != NULL)
1806 		crfree(cred);
1807 #ifdef MAC
1808 	if (sc == &scs)
1809 		mac_syncache_destroy(&maclabel);
1810 #endif
1811 	return (rv);
1812 }
1813 
1814 /*
1815  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1816  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1817  */
1818 static int
1819 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1820 {
1821 	struct ip *ip = NULL;
1822 	struct mbuf *m;
1823 	struct tcphdr *th = NULL;
1824 	struct udphdr *udp = NULL;
1825 	int optlen, error = 0;	/* Make compiler happy */
1826 	u_int16_t hlen, tlen, mssopt, ulen;
1827 	struct tcpopt to;
1828 #ifdef INET6
1829 	struct ip6_hdr *ip6 = NULL;
1830 #endif
1831 
1832 	NET_EPOCH_ASSERT();
1833 
1834 	hlen =
1835 #ifdef INET6
1836 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1837 #endif
1838 		sizeof(struct ip);
1839 	tlen = hlen + sizeof(struct tcphdr);
1840 	if (sc->sc_port) {
1841 		tlen += sizeof(struct udphdr);
1842 	}
1843 	/* Determine MSS we advertize to other end of connection. */
1844 	mssopt = tcp_mssopt(&sc->sc_inc);
1845 	if (sc->sc_port)
1846 		mssopt -= V_tcp_udp_tunneling_overhead;
1847 	mssopt = max(mssopt, V_tcp_minmss);
1848 
1849 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1850 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1851 	    ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1852 	    "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1853 	    max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1854 
1855 	/* Create the IP+TCP header from scratch. */
1856 	m = m_gethdr(M_NOWAIT, MT_DATA);
1857 	if (m == NULL)
1858 		return (ENOBUFS);
1859 #ifdef MAC
1860 	mac_syncache_create_mbuf(sc->sc_label, m);
1861 #endif
1862 	m->m_data += max_linkhdr;
1863 	m->m_len = tlen;
1864 	m->m_pkthdr.len = tlen;
1865 	m->m_pkthdr.rcvif = NULL;
1866 
1867 #ifdef INET6
1868 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1869 		ip6 = mtod(m, struct ip6_hdr *);
1870 		ip6->ip6_vfc = IPV6_VERSION;
1871 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1872 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1873 		ip6->ip6_plen = htons(tlen - hlen);
1874 		/* ip6_hlim is set after checksum */
1875 		/* Zero out traffic class and flow label. */
1876 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1877 		ip6->ip6_flow |= sc->sc_flowlabel;
1878 		if (sc->sc_port != 0) {
1879 			ip6->ip6_nxt = IPPROTO_UDP;
1880 			udp = (struct udphdr *)(ip6 + 1);
1881 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1882 			udp->uh_dport = sc->sc_port;
1883 			ulen = (tlen - sizeof(struct ip6_hdr));
1884 			th = (struct tcphdr *)(udp + 1);
1885 		} else {
1886 			ip6->ip6_nxt = IPPROTO_TCP;
1887 			th = (struct tcphdr *)(ip6 + 1);
1888 		}
1889 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20);
1890 	}
1891 #endif
1892 #if defined(INET6) && defined(INET)
1893 	else
1894 #endif
1895 #ifdef INET
1896 	{
1897 		ip = mtod(m, struct ip *);
1898 		ip->ip_v = IPVERSION;
1899 		ip->ip_hl = sizeof(struct ip) >> 2;
1900 		ip->ip_len = htons(tlen);
1901 		ip->ip_id = 0;
1902 		ip->ip_off = 0;
1903 		ip->ip_sum = 0;
1904 		ip->ip_src = sc->sc_inc.inc_laddr;
1905 		ip->ip_dst = sc->sc_inc.inc_faddr;
1906 		ip->ip_ttl = sc->sc_ip_ttl;
1907 		ip->ip_tos = sc->sc_ip_tos;
1908 
1909 		/*
1910 		 * See if we should do MTU discovery.  Route lookups are
1911 		 * expensive, so we will only unset the DF bit if:
1912 		 *
1913 		 *	1) path_mtu_discovery is disabled
1914 		 *	2) the SCF_UNREACH flag has been set
1915 		 */
1916 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1917 		       ip->ip_off |= htons(IP_DF);
1918 		if (sc->sc_port == 0) {
1919 			ip->ip_p = IPPROTO_TCP;
1920 			th = (struct tcphdr *)(ip + 1);
1921 		} else {
1922 			ip->ip_p = IPPROTO_UDP;
1923 			udp = (struct udphdr *)(ip + 1);
1924 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1925 			udp->uh_dport = sc->sc_port;
1926 			ulen = (tlen - sizeof(struct ip));
1927 			th = (struct tcphdr *)(udp + 1);
1928 		}
1929 	}
1930 #endif /* INET */
1931 	th->th_sport = sc->sc_inc.inc_lport;
1932 	th->th_dport = sc->sc_inc.inc_fport;
1933 
1934 	if (flags & TH_SYN)
1935 		th->th_seq = htonl(sc->sc_iss);
1936 	else
1937 		th->th_seq = htonl(sc->sc_iss + 1);
1938 	th->th_ack = htonl(sc->sc_irs + 1);
1939 	th->th_off = sizeof(struct tcphdr) >> 2;
1940 	th->th_win = htons(sc->sc_wnd);
1941 	th->th_urp = 0;
1942 
1943 	flags = tcp_ecn_syncache_respond(flags, sc);
1944 	tcp_set_flags(th, flags);
1945 
1946 	/* Tack on the TCP options. */
1947 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1948 		to.to_flags = 0;
1949 
1950 		if (flags & TH_SYN) {
1951 			to.to_mss = mssopt;
1952 			to.to_flags = TOF_MSS;
1953 			if (sc->sc_flags & SCF_WINSCALE) {
1954 				to.to_wscale = sc->sc_requested_r_scale;
1955 				to.to_flags |= TOF_SCALE;
1956 			}
1957 			if (sc->sc_flags & SCF_SACK)
1958 				to.to_flags |= TOF_SACKPERM;
1959 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1960 			if (sc->sc_flags & SCF_SIGNATURE)
1961 				to.to_flags |= TOF_SIGNATURE;
1962 #endif
1963 			if (sc->sc_tfo_cookie) {
1964 				to.to_flags |= TOF_FASTOPEN;
1965 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1966 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1967 				/* don't send cookie again when retransmitting response */
1968 				sc->sc_tfo_cookie = NULL;
1969 			}
1970 		}
1971 		if (sc->sc_flags & SCF_TIMESTAMP) {
1972 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1973 			to.to_tsecr = sc->sc_tsreflect;
1974 			to.to_flags |= TOF_TS;
1975 		}
1976 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1977 
1978 		/* Adjust headers by option size. */
1979 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1980 		m->m_len += optlen;
1981 		m->m_pkthdr.len += optlen;
1982 #ifdef INET6
1983 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1984 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1985 		else
1986 #endif
1987 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1988 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1989 		if (sc->sc_flags & SCF_SIGNATURE) {
1990 			KASSERT(to.to_flags & TOF_SIGNATURE,
1991 			    ("tcp_addoptions() didn't set tcp_signature"));
1992 
1993 			/* NOTE: to.to_signature is inside of mbuf */
1994 			if (!TCPMD5_ENABLED() ||
1995 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1996 				m_freem(m);
1997 				return (EACCES);
1998 			}
1999 		}
2000 #endif
2001 	} else
2002 		optlen = 0;
2003 
2004 	if (udp) {
2005 		ulen += optlen;
2006 		udp->uh_ulen = htons(ulen);
2007 	}
2008 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
2009 	/*
2010 	 * If we have peer's SYN and it has a flowid, then let's assign it to
2011 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
2012 	 * to SYN|ACK due to lack of inp here.
2013 	 */
2014 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
2015 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
2016 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
2017 	}
2018 #ifdef INET6
2019 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2020 		if (sc->sc_port) {
2021 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2022 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2023 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2024 			      IPPROTO_UDP, 0);
2025 			th->th_sum = htons(0);
2026 		} else {
2027 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2028 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2029 			th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2030 			    IPPROTO_TCP, 0);
2031 		}
2032 		ip6->ip6_hlim = sc->sc_ip_ttl;
2033 #ifdef TCP_OFFLOAD
2034 		if (ADDED_BY_TOE(sc)) {
2035 			struct toedev *tod = sc->sc_tod;
2036 
2037 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2038 
2039 			return (error);
2040 		}
2041 #endif
2042 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2043 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2044 	}
2045 #endif
2046 #if defined(INET6) && defined(INET)
2047 	else
2048 #endif
2049 #ifdef INET
2050 	{
2051 		if (sc->sc_port) {
2052 			m->m_pkthdr.csum_flags = CSUM_UDP;
2053 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2054 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2055 			      ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2056 			th->th_sum = htons(0);
2057 		} else {
2058 			m->m_pkthdr.csum_flags = CSUM_TCP;
2059 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2060 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2061 			    htons(tlen + optlen - hlen + IPPROTO_TCP));
2062 		}
2063 #ifdef TCP_OFFLOAD
2064 		if (ADDED_BY_TOE(sc)) {
2065 			struct toedev *tod = sc->sc_tod;
2066 
2067 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2068 
2069 			return (error);
2070 		}
2071 #endif
2072 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2073 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2074 	}
2075 #endif
2076 	return (error);
2077 }
2078 
2079 /*
2080  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2081  * that exceed the capacity of the syncache by avoiding the storage of any
2082  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2083  * attacks where the attacker does not have access to our responses.
2084  *
2085  * Syncookies encode and include all necessary information about the
2086  * connection setup within the SYN|ACK that we send back.  That way we
2087  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2088  * (if ever).  Normally the syncache and syncookies are running in parallel
2089  * with the latter taking over when the former is exhausted.  When matching
2090  * syncache entry is found the syncookie is ignored.
2091  *
2092  * The only reliable information persisting the 3WHS is our initial sequence
2093  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2094  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2095  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2096  * returns and signifies a legitimate connection if it matches the ACK.
2097  *
2098  * The available space of 32 bits to store the hash and to encode the SYN
2099  * option information is very tight and we should have at least 24 bits for
2100  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2101  *
2102  * SYN option information we have to encode to fully restore a connection:
2103  * MSS: is imporant to chose an optimal segment size to avoid IP level
2104  *   fragmentation along the path.  The common MSS values can be encoded
2105  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2106  *   in the table leading to a slight increase in packetization overhead.
2107  * WSCALE: is necessary to allow large windows to be used for high delay-
2108  *   bandwidth product links.  Not scaling the window when it was initially
2109  *   negotiated is bad for performance as lack of scaling further decreases
2110  *   the apparent available send window.  We only need to encode the WSCALE
2111  *   we received from the remote end.  Our end can be recalculated at any
2112  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2113  *   Uncommon values are captured by the next lower value in the table
2114  *   making us under-estimate the available window size halving our
2115  *   theoretically possible maximum throughput for that connection.
2116  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2117  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2118  *   that are included in all segments on a connection.  We enable them when
2119  *   the ACK has them.
2120  *
2121  * Security of syncookies and attack vectors:
2122  *
2123  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2124  * together with the gloabl secret to make it unique per connection attempt.
2125  * Thus any change of any of those parameters results in a different MAC output
2126  * in an unpredictable way unless a collision is encountered.  24 bits of the
2127  * MAC are embedded into the ISS.
2128  *
2129  * To prevent replay attacks two rotating global secrets are updated with a
2130  * new random value every 15 seconds.  The life-time of a syncookie is thus
2131  * 15-30 seconds.
2132  *
2133  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2134  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2135  * text attack by varying and testing the all parameters under his control.
2136  * The strength depends on the size and randomness of the secret, and the
2137  * cryptographic security of the MAC function.  Due to the constant updating
2138  * of the secret the attacker has at most 29.999 seconds to find the secret
2139  * and launch spoofed connections.  After that he has to start all over again.
2140  *
2141  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2142  * size an average of 4,823 attempts are required for a 50% chance of success
2143  * to spoof a single syncookie (birthday collision paradox).  However the
2144  * attacker is blind and doesn't know if one of his attempts succeeded unless
2145  * he has a side channel to interfere success from.  A single connection setup
2146  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2147  * This many attempts are required for each one blind spoofed connection.  For
2148  * every additional spoofed connection he has to launch another N attempts.
2149  * Thus for a sustained rate 100 spoofed connections per second approximately
2150  * 1,800,000 packets per second would have to be sent.
2151  *
2152  * NB: The MAC function should be fast so that it doesn't become a CPU
2153  * exhaustion attack vector itself.
2154  *
2155  * References:
2156  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2157  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2158  *   http://cr.yp.to/syncookies.html    (overview)
2159  *   http://cr.yp.to/syncookies/archive (details)
2160  *
2161  *
2162  * Schematic construction of a syncookie enabled Initial Sequence Number:
2163  *  0        1         2         3
2164  *  12345678901234567890123456789012
2165  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2166  *
2167  *  x 24 MAC (truncated)
2168  *  W  3 Send Window Scale index
2169  *  M  3 MSS index
2170  *  S  1 SACK permitted
2171  *  P  1 Odd/even secret
2172  */
2173 
2174 /*
2175  * Distribution and probability of certain MSS values.  Those in between are
2176  * rounded down to the next lower one.
2177  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2178  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2179  */
2180 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2181 
2182 /*
2183  * Distribution and probability of certain WSCALE values.  We have to map the
2184  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2185  * bits based on prevalence of certain values.  Where we don't have an exact
2186  * match for are rounded down to the next lower one letting us under-estimate
2187  * the true available window.  At the moment this would happen only for the
2188  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2189  * and window size).  The absence of the WSCALE option (no scaling in either
2190  * direction) is encoded with index zero.
2191  * [WSCALE values histograms, Allman, 2012]
2192  *                            X 10 10 35  5  6 14 10%   by host
2193  *                            X 11  4  5  5 18 49  3%   by connections
2194  */
2195 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2196 
2197 /*
2198  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2199  * and good cryptographic properties.
2200  */
2201 static uint32_t
2202 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2203     uint8_t *secbits, uintptr_t secmod)
2204 {
2205 	SIPHASH_CTX ctx;
2206 	uint32_t siphash[2];
2207 
2208 	SipHash24_Init(&ctx);
2209 	SipHash_SetKey(&ctx, secbits);
2210 	switch (inc->inc_flags & INC_ISIPV6) {
2211 #ifdef INET
2212 	case 0:
2213 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2214 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2215 		break;
2216 #endif
2217 #ifdef INET6
2218 	case INC_ISIPV6:
2219 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2220 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2221 		break;
2222 #endif
2223 	}
2224 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2225 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2226 	SipHash_Update(&ctx, &irs, sizeof(irs));
2227 	SipHash_Update(&ctx, &flags, sizeof(flags));
2228 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2229 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2230 
2231 	return (siphash[0] ^ siphash[1]);
2232 }
2233 
2234 static tcp_seq
2235 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2236 {
2237 	u_int i, secbit, wscale;
2238 	uint32_t iss, hash;
2239 	uint8_t *secbits;
2240 	union syncookie cookie;
2241 
2242 	cookie.cookie = 0;
2243 
2244 	/* Map our computed MSS into the 3-bit index. */
2245 	for (i = nitems(tcp_sc_msstab) - 1;
2246 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2247 	     i--)
2248 		;
2249 	cookie.flags.mss_idx = i;
2250 
2251 	/*
2252 	 * Map the send window scale into the 3-bit index but only if
2253 	 * the wscale option was received.
2254 	 */
2255 	if (sc->sc_flags & SCF_WINSCALE) {
2256 		wscale = sc->sc_requested_s_scale;
2257 		for (i = nitems(tcp_sc_wstab) - 1;
2258 		    tcp_sc_wstab[i] > wscale && i > 0;
2259 		     i--)
2260 			;
2261 		cookie.flags.wscale_idx = i;
2262 	}
2263 
2264 	/* Can we do SACK? */
2265 	if (sc->sc_flags & SCF_SACK)
2266 		cookie.flags.sack_ok = 1;
2267 
2268 	/* Which of the two secrets to use. */
2269 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2270 	cookie.flags.odd_even = secbit;
2271 
2272 	secbits = V_tcp_syncache.secret.key[secbit];
2273 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2274 	    (uintptr_t)sch);
2275 
2276 	/*
2277 	 * Put the flags into the hash and XOR them to get better ISS number
2278 	 * variance.  This doesn't enhance the cryptographic strength and is
2279 	 * done to prevent the 8 cookie bits from showing up directly on the
2280 	 * wire.
2281 	 */
2282 	iss = hash & ~0xff;
2283 	iss |= cookie.cookie ^ (hash >> 24);
2284 
2285 	TCPSTAT_INC(tcps_sc_sendcookie);
2286 	return (iss);
2287 }
2288 
2289 static struct syncache *
2290 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2291     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2292     struct socket *lso, uint16_t port)
2293 {
2294 	uint32_t hash;
2295 	uint8_t *secbits;
2296 	tcp_seq ack, seq;
2297 	int wnd, wscale = 0;
2298 	union syncookie cookie;
2299 
2300 	/*
2301 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2302 	 * advances.
2303 	 */
2304 	ack = th->th_ack - 1;
2305 	seq = th->th_seq - 1;
2306 
2307 	/*
2308 	 * Unpack the flags containing enough information to restore the
2309 	 * connection.
2310 	 */
2311 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2312 
2313 	/* Which of the two secrets to use. */
2314 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2315 
2316 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2317 
2318 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2319 	if ((ack & ~0xff) != (hash & ~0xff))
2320 		return (NULL);
2321 
2322 	/* Fill in the syncache values. */
2323 	sc->sc_flags = 0;
2324 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2325 	sc->sc_ipopts = NULL;
2326 
2327 	sc->sc_irs = seq;
2328 	sc->sc_iss = ack;
2329 
2330 	switch (inc->inc_flags & INC_ISIPV6) {
2331 #ifdef INET
2332 	case 0:
2333 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2334 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2335 		break;
2336 #endif
2337 #ifdef INET6
2338 	case INC_ISIPV6:
2339 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2340 			sc->sc_flowlabel =
2341 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2342 		break;
2343 #endif
2344 	}
2345 
2346 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2347 
2348 	/* We can simply recompute receive window scale we sent earlier. */
2349 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2350 		wscale++;
2351 
2352 	/* Only use wscale if it was enabled in the orignal SYN. */
2353 	if (cookie.flags.wscale_idx > 0) {
2354 		sc->sc_requested_r_scale = wscale;
2355 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2356 		sc->sc_flags |= SCF_WINSCALE;
2357 	}
2358 
2359 	wnd = lso->sol_sbrcv_hiwat;
2360 	wnd = imax(wnd, 0);
2361 	wnd = imin(wnd, TCP_MAXWIN);
2362 	sc->sc_wnd = wnd;
2363 
2364 	if (cookie.flags.sack_ok)
2365 		sc->sc_flags |= SCF_SACK;
2366 
2367 	if (to->to_flags & TOF_TS) {
2368 		sc->sc_flags |= SCF_TIMESTAMP;
2369 		sc->sc_tsreflect = to->to_tsval;
2370 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2371 	}
2372 
2373 	if (to->to_flags & TOF_SIGNATURE)
2374 		sc->sc_flags |= SCF_SIGNATURE;
2375 
2376 	sc->sc_rxmits = 0;
2377 
2378 	sc->sc_port = port;
2379 
2380 	TCPSTAT_INC(tcps_sc_recvcookie);
2381 	return (sc);
2382 }
2383 
2384 #ifdef INVARIANTS
2385 static int
2386 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2387     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2388     struct socket *lso, uint16_t port)
2389 {
2390 	struct syncache scs, *scx;
2391 	char *s;
2392 
2393 	bzero(&scs, sizeof(scs));
2394 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2395 
2396 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2397 		return (0);
2398 
2399 	if (scx != NULL) {
2400 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2401 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2402 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2403 
2404 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2405 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2406 			    s, __func__, sc->sc_requested_r_scale,
2407 			    scx->sc_requested_r_scale);
2408 
2409 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2410 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2411 			    s, __func__, sc->sc_requested_s_scale,
2412 			    scx->sc_requested_s_scale);
2413 
2414 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2415 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2416 	}
2417 
2418 	if (s != NULL)
2419 		free(s, M_TCPLOG);
2420 	return (0);
2421 }
2422 #endif /* INVARIANTS */
2423 
2424 static void
2425 syncookie_reseed(void *arg)
2426 {
2427 	struct tcp_syncache *sc = arg;
2428 	uint8_t *secbits;
2429 	int secbit;
2430 
2431 	/*
2432 	 * Reseeding the secret doesn't have to be protected by a lock.
2433 	 * It only must be ensured that the new random values are visible
2434 	 * to all CPUs in a SMP environment.  The atomic with release
2435 	 * semantics ensures that.
2436 	 */
2437 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2438 	secbits = sc->secret.key[secbit];
2439 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2440 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2441 
2442 	/* Reschedule ourself. */
2443 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2444 }
2445 
2446 /*
2447  * We have overflowed a bucket. Let's pause dealing with the syncache.
2448  * This function will increment the bucketoverflow statistics appropriately
2449  * (once per pause when pausing is enabled; otherwise, once per overflow).
2450  */
2451 static void
2452 syncache_pause(struct in_conninfo *inc)
2453 {
2454 	time_t delta;
2455 	const char *s;
2456 
2457 	/* XXX:
2458 	 * 2. Add sysctl read here so we don't get the benefit of this
2459 	 * change without the new sysctl.
2460 	 */
2461 
2462 	/*
2463 	 * Try an unlocked read. If we already know that another thread
2464 	 * has activated the feature, there is no need to proceed.
2465 	 */
2466 	if (V_tcp_syncache.paused)
2467 		return;
2468 
2469 	/* Are cookied enabled? If not, we can't pause. */
2470 	if (!V_tcp_syncookies) {
2471 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2472 		return;
2473 	}
2474 
2475 	/*
2476 	 * We may be the first thread to find an overflow. Get the lock
2477 	 * and evaluate if we need to take action.
2478 	 */
2479 	mtx_lock(&V_tcp_syncache.pause_mtx);
2480 	if (V_tcp_syncache.paused) {
2481 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2482 		return;
2483 	}
2484 
2485 	/* Activate protection. */
2486 	V_tcp_syncache.paused = true;
2487 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2488 
2489 	/*
2490 	 * Determine the last backoff time. If we are seeing a re-newed
2491 	 * attack within that same time after last reactivating the syncache,
2492 	 * consider it an extension of the same attack.
2493 	 */
2494 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2495 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2496 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2497 			delta <<= 1;
2498 			V_tcp_syncache.pause_backoff++;
2499 		}
2500 	} else {
2501 		delta = TCP_SYNCACHE_PAUSE_TIME;
2502 		V_tcp_syncache.pause_backoff = 0;
2503 	}
2504 
2505 	/* Log a warning, including IP addresses, if able. */
2506 	if (inc != NULL)
2507 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2508 	else
2509 		s = (const char *)NULL;
2510 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2511 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2512 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2513 	    (s != NULL) ? ")" : "");
2514 	free(__DECONST(void *, s), M_TCPLOG);
2515 
2516 	/* Use the calculated delta to set a new pause time. */
2517 	V_tcp_syncache.pause_until = time_uptime + delta;
2518 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2519 	    &V_tcp_syncache);
2520 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2521 }
2522 
2523 /* Evaluate whether we need to unpause. */
2524 static void
2525 syncache_unpause(void *arg)
2526 {
2527 	struct tcp_syncache *sc;
2528 	time_t delta;
2529 
2530 	sc = arg;
2531 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2532 	callout_deactivate(&sc->pause_co);
2533 
2534 	/*
2535 	 * Check to make sure we are not running early. If the pause
2536 	 * time has expired, then deactivate the protection.
2537 	 */
2538 	if ((delta = sc->pause_until - time_uptime) > 0)
2539 		callout_schedule(&sc->pause_co, delta * hz);
2540 	else
2541 		sc->paused = false;
2542 }
2543 
2544 /*
2545  * Exports the syncache entries to userland so that netstat can display
2546  * them alongside the other sockets.  This function is intended to be
2547  * called only from tcp_pcblist.
2548  *
2549  * Due to concurrency on an active system, the number of pcbs exported
2550  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2551  * amount of space the caller allocated for this function to use.
2552  */
2553 int
2554 syncache_pcblist(struct sysctl_req *req)
2555 {
2556 	struct xtcpcb xt;
2557 	struct syncache *sc;
2558 	struct syncache_head *sch;
2559 	int error, i;
2560 
2561 	bzero(&xt, sizeof(xt));
2562 	xt.xt_len = sizeof(xt);
2563 	xt.t_state = TCPS_SYN_RECEIVED;
2564 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2565 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2566 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2567 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2568 
2569 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2570 		sch = &V_tcp_syncache.hashbase[i];
2571 		SCH_LOCK(sch);
2572 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2573 			if (sc->sc_cred != NULL &&
2574 			    cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2575 				continue;
2576 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2577 				xt.xt_inp.inp_vflag = INP_IPV6;
2578 			else
2579 				xt.xt_inp.inp_vflag = INP_IPV4;
2580 			xt.xt_encaps_port = sc->sc_port;
2581 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2582 			    sizeof (struct in_conninfo));
2583 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2584 			if (error) {
2585 				SCH_UNLOCK(sch);
2586 				return (0);
2587 			}
2588 		}
2589 		SCH_UNLOCK(sch);
2590 	}
2591 
2592 	return (0);
2593 }
2594