xref: /freebsd/sys/netinet/tcp_syncache.c (revision 8ef24a0d4b28fe230e20637f56869cc4148cd2ca)
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/hash.h>
44 #include <sys/refcount.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/limits.h>
48 #include <sys/lock.h>
49 #include <sys/mutex.h>
50 #include <sys/malloc.h>
51 #include <sys/mbuf.h>
52 #include <sys/proc.h>		/* for proc0 declaration */
53 #include <sys/random.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/syslog.h>
57 #include <sys/ucred.h>
58 
59 #include <sys/md5.h>
60 #include <crypto/siphash/siphash.h>
61 
62 #include <vm/uma.h>
63 
64 #include <net/if.h>
65 #include <net/if_var.h>
66 #include <net/route.h>
67 #include <net/vnet.h>
68 
69 #include <netinet/in.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/ip.h>
72 #include <netinet/in_var.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
76 #ifdef INET6
77 #include <netinet/ip6.h>
78 #include <netinet/icmp6.h>
79 #include <netinet6/nd6.h>
80 #include <netinet6/ip6_var.h>
81 #include <netinet6/in6_pcb.h>
82 #endif
83 #include <netinet/tcp.h>
84 #ifdef TCP_RFC7413
85 #include <netinet/tcp_fastopen.h>
86 #endif
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 #ifdef INET6
93 #include <netinet6/tcp6_var.h>
94 #endif
95 #ifdef TCP_OFFLOAD
96 #include <netinet/toecore.h>
97 #endif
98 
99 #ifdef IPSEC
100 #include <netipsec/ipsec.h>
101 #ifdef INET6
102 #include <netipsec/ipsec6.h>
103 #endif
104 #include <netipsec/key.h>
105 #endif /*IPSEC*/
106 
107 #include <machine/in_cksum.h>
108 
109 #include <security/mac/mac_framework.h>
110 
111 static VNET_DEFINE(int, tcp_syncookies) = 1;
112 #define	V_tcp_syncookies		VNET(tcp_syncookies)
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
114     &VNET_NAME(tcp_syncookies), 0,
115     "Use TCP SYN cookies if the syncache overflows");
116 
117 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
118 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
120     &VNET_NAME(tcp_syncookiesonly), 0,
121     "Use only TCP SYN cookies");
122 
123 #ifdef TCP_OFFLOAD
124 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
125 #endif
126 
127 static void	 syncache_drop(struct syncache *, struct syncache_head *);
128 static void	 syncache_free(struct syncache *);
129 static void	 syncache_insert(struct syncache *, struct syncache_head *);
130 static int	 syncache_respond(struct syncache *, struct syncache_head *, int,
131 		    const struct mbuf *);
132 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
133 		    struct mbuf *m);
134 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
135 		    int docallout);
136 static void	 syncache_timer(void *);
137 
138 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
139 		    uint8_t *, uintptr_t);
140 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
141 static struct syncache
142 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
143 		    struct syncache *, struct tcphdr *, struct tcpopt *,
144 		    struct socket *);
145 static void	 syncookie_reseed(void *);
146 #ifdef INVARIANTS
147 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
148 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
149 		    struct socket *lso);
150 #endif
151 
152 /*
153  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
154  * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
155  * the odds are that the user has given up attempting to connect by then.
156  */
157 #define SYNCACHE_MAXREXMTS		3
158 
159 /* Arbitrary values */
160 #define TCP_SYNCACHE_HASHSIZE		512
161 #define TCP_SYNCACHE_BUCKETLIMIT	30
162 
163 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
164 #define	V_tcp_syncache			VNET(tcp_syncache)
165 
166 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
167     "TCP SYN cache");
168 
169 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
170     &VNET_NAME(tcp_syncache.bucket_limit), 0,
171     "Per-bucket hash limit for syncache");
172 
173 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
174     &VNET_NAME(tcp_syncache.cache_limit), 0,
175     "Overall entry limit for syncache");
176 
177 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
178     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
179 
180 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
181     &VNET_NAME(tcp_syncache.hashsize), 0,
182     "Size of TCP syncache hashtable");
183 
184 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW,
185     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
186     "Limit on SYN/ACK retransmissions");
187 
188 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
189 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
190     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
191     "Send reset on socket allocation failure");
192 
193 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
194 
195 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
196 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
197 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
198 
199 /*
200  * Requires the syncache entry to be already removed from the bucket list.
201  */
202 static void
203 syncache_free(struct syncache *sc)
204 {
205 
206 	if (sc->sc_ipopts)
207 		(void) m_free(sc->sc_ipopts);
208 	if (sc->sc_cred)
209 		crfree(sc->sc_cred);
210 #ifdef MAC
211 	mac_syncache_destroy(&sc->sc_label);
212 #endif
213 
214 	uma_zfree(V_tcp_syncache.zone, sc);
215 }
216 
217 void
218 syncache_init(void)
219 {
220 	int i;
221 
222 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
223 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
224 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
225 	V_tcp_syncache.hash_secret = arc4random();
226 
227 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
228 	    &V_tcp_syncache.hashsize);
229 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
230 	    &V_tcp_syncache.bucket_limit);
231 	if (!powerof2(V_tcp_syncache.hashsize) ||
232 	    V_tcp_syncache.hashsize == 0) {
233 		printf("WARNING: syncache hash size is not a power of 2.\n");
234 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
235 	}
236 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
237 
238 	/* Set limits. */
239 	V_tcp_syncache.cache_limit =
240 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
241 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
242 	    &V_tcp_syncache.cache_limit);
243 
244 	/* Allocate the hash table. */
245 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
246 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
247 
248 #ifdef VIMAGE
249 	V_tcp_syncache.vnet = curvnet;
250 #endif
251 
252 	/* Initialize the hash buckets. */
253 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
254 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
255 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
256 			 NULL, MTX_DEF);
257 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
258 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
259 		V_tcp_syncache.hashbase[i].sch_length = 0;
260 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
261 	}
262 
263 	/* Create the syncache entry zone. */
264 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
265 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
266 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
267 	    V_tcp_syncache.cache_limit);
268 
269 	/* Start the SYN cookie reseeder callout. */
270 	callout_init(&V_tcp_syncache.secret.reseed, 1);
271 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
272 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
273 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
274 	    syncookie_reseed, &V_tcp_syncache);
275 }
276 
277 #ifdef VIMAGE
278 void
279 syncache_destroy(void)
280 {
281 	struct syncache_head *sch;
282 	struct syncache *sc, *nsc;
283 	int i;
284 
285 	/*
286 	 * Stop the re-seed timer before freeing resources.  No need to
287 	 * possibly schedule it another time.
288 	 */
289 	callout_drain(&V_tcp_syncache.secret.reseed);
290 
291 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
292 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
293 
294 		sch = &V_tcp_syncache.hashbase[i];
295 		callout_drain(&sch->sch_timer);
296 
297 		SCH_LOCK(sch);
298 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
299 			syncache_drop(sc, sch);
300 		SCH_UNLOCK(sch);
301 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
302 		    ("%s: sch->sch_bucket not empty", __func__));
303 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
304 		    __func__, sch->sch_length));
305 		mtx_destroy(&sch->sch_mtx);
306 	}
307 
308 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
309 	    ("%s: cache_count not 0", __func__));
310 
311 	/* Free the allocated global resources. */
312 	uma_zdestroy(V_tcp_syncache.zone);
313 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
314 }
315 #endif
316 
317 /*
318  * Inserts a syncache entry into the specified bucket row.
319  * Locks and unlocks the syncache_head autonomously.
320  */
321 static void
322 syncache_insert(struct syncache *sc, struct syncache_head *sch)
323 {
324 	struct syncache *sc2;
325 
326 	SCH_LOCK(sch);
327 
328 	/*
329 	 * Make sure that we don't overflow the per-bucket limit.
330 	 * If the bucket is full, toss the oldest element.
331 	 */
332 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
333 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
334 			("sch->sch_length incorrect"));
335 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
336 		syncache_drop(sc2, sch);
337 		TCPSTAT_INC(tcps_sc_bucketoverflow);
338 	}
339 
340 	/* Put it into the bucket. */
341 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
342 	sch->sch_length++;
343 
344 #ifdef TCP_OFFLOAD
345 	if (ADDED_BY_TOE(sc)) {
346 		struct toedev *tod = sc->sc_tod;
347 
348 		tod->tod_syncache_added(tod, sc->sc_todctx);
349 	}
350 #endif
351 
352 	/* Reinitialize the bucket row's timer. */
353 	if (sch->sch_length == 1)
354 		sch->sch_nextc = ticks + INT_MAX;
355 	syncache_timeout(sc, sch, 1);
356 
357 	SCH_UNLOCK(sch);
358 
359 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
360 	TCPSTAT_INC(tcps_sc_added);
361 }
362 
363 /*
364  * Remove and free entry from syncache bucket row.
365  * Expects locked syncache head.
366  */
367 static void
368 syncache_drop(struct syncache *sc, struct syncache_head *sch)
369 {
370 
371 	SCH_LOCK_ASSERT(sch);
372 
373 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
374 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
375 	sch->sch_length--;
376 
377 #ifdef TCP_OFFLOAD
378 	if (ADDED_BY_TOE(sc)) {
379 		struct toedev *tod = sc->sc_tod;
380 
381 		tod->tod_syncache_removed(tod, sc->sc_todctx);
382 	}
383 #endif
384 
385 	syncache_free(sc);
386 }
387 
388 /*
389  * Engage/reengage time on bucket row.
390  */
391 static void
392 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
393 {
394 	sc->sc_rxttime = ticks +
395 		TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
396 	sc->sc_rxmits++;
397 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
398 		sch->sch_nextc = sc->sc_rxttime;
399 		if (docallout)
400 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
401 			    syncache_timer, (void *)sch);
402 	}
403 }
404 
405 /*
406  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
407  * If we have retransmitted an entry the maximum number of times, expire it.
408  * One separate timer for each bucket row.
409  */
410 static void
411 syncache_timer(void *xsch)
412 {
413 	struct syncache_head *sch = (struct syncache_head *)xsch;
414 	struct syncache *sc, *nsc;
415 	int tick = ticks;
416 	char *s;
417 
418 	CURVNET_SET(sch->sch_sc->vnet);
419 
420 	/* NB: syncache_head has already been locked by the callout. */
421 	SCH_LOCK_ASSERT(sch);
422 
423 	/*
424 	 * In the following cycle we may remove some entries and/or
425 	 * advance some timeouts, so re-initialize the bucket timer.
426 	 */
427 	sch->sch_nextc = tick + INT_MAX;
428 
429 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
430 		/*
431 		 * We do not check if the listen socket still exists
432 		 * and accept the case where the listen socket may be
433 		 * gone by the time we resend the SYN/ACK.  We do
434 		 * not expect this to happens often. If it does,
435 		 * then the RST will be sent by the time the remote
436 		 * host does the SYN/ACK->ACK.
437 		 */
438 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
439 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
440 				sch->sch_nextc = sc->sc_rxttime;
441 			continue;
442 		}
443 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
444 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
445 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
446 				    "giving up and removing syncache entry\n",
447 				    s, __func__);
448 				free(s, M_TCPLOG);
449 			}
450 			syncache_drop(sc, sch);
451 			TCPSTAT_INC(tcps_sc_stale);
452 			continue;
453 		}
454 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
455 			log(LOG_DEBUG, "%s; %s: Response timeout, "
456 			    "retransmitting (%u) SYN|ACK\n",
457 			    s, __func__, sc->sc_rxmits);
458 			free(s, M_TCPLOG);
459 		}
460 
461 		syncache_respond(sc, sch, 1, NULL);
462 		TCPSTAT_INC(tcps_sc_retransmitted);
463 		syncache_timeout(sc, sch, 0);
464 	}
465 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
466 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
467 			syncache_timer, (void *)(sch));
468 	CURVNET_RESTORE();
469 }
470 
471 /*
472  * Find an entry in the syncache.
473  * Returns always with locked syncache_head plus a matching entry or NULL.
474  */
475 static struct syncache *
476 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
477 {
478 	struct syncache *sc;
479 	struct syncache_head *sch;
480 	uint32_t hash;
481 
482 	/*
483 	 * The hash is built on foreign port + local port + foreign address.
484 	 * We rely on the fact that struct in_conninfo starts with 16 bits
485 	 * of foreign port, then 16 bits of local port then followed by 128
486 	 * bits of foreign address.  In case of IPv4 address, the first 3
487 	 * 32-bit words of the address always are zeroes.
488 	 */
489 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
490 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
491 
492 	sch = &V_tcp_syncache.hashbase[hash];
493 	*schp = sch;
494 	SCH_LOCK(sch);
495 
496 	/* Circle through bucket row to find matching entry. */
497 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
498 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
499 		    sizeof(struct in_endpoints)) == 0)
500 			break;
501 
502 	return (sc);	/* Always returns with locked sch. */
503 }
504 
505 /*
506  * This function is called when we get a RST for a
507  * non-existent connection, so that we can see if the
508  * connection is in the syn cache.  If it is, zap it.
509  */
510 void
511 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
512 {
513 	struct syncache *sc;
514 	struct syncache_head *sch;
515 	char *s = NULL;
516 
517 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
518 	SCH_LOCK_ASSERT(sch);
519 
520 	/*
521 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
522 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
523 	 */
524 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
525 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
526 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
527 			    "FIN flag set, segment ignored\n", s, __func__);
528 		TCPSTAT_INC(tcps_badrst);
529 		goto done;
530 	}
531 
532 	/*
533 	 * No corresponding connection was found in syncache.
534 	 * If syncookies are enabled and possibly exclusively
535 	 * used, or we are under memory pressure, a valid RST
536 	 * may not find a syncache entry.  In that case we're
537 	 * done and no SYN|ACK retransmissions will happen.
538 	 * Otherwise the RST was misdirected or spoofed.
539 	 */
540 	if (sc == NULL) {
541 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
542 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
543 			    "syncache entry (possibly syncookie only), "
544 			    "segment ignored\n", s, __func__);
545 		TCPSTAT_INC(tcps_badrst);
546 		goto done;
547 	}
548 
549 	/*
550 	 * If the RST bit is set, check the sequence number to see
551 	 * if this is a valid reset segment.
552 	 * RFC 793 page 37:
553 	 *   In all states except SYN-SENT, all reset (RST) segments
554 	 *   are validated by checking their SEQ-fields.  A reset is
555 	 *   valid if its sequence number is in the window.
556 	 *
557 	 *   The sequence number in the reset segment is normally an
558 	 *   echo of our outgoing acknowlegement numbers, but some hosts
559 	 *   send a reset with the sequence number at the rightmost edge
560 	 *   of our receive window, and we have to handle this case.
561 	 */
562 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
563 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
564 		syncache_drop(sc, sch);
565 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
566 			log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
567 			    "connection attempt aborted by remote endpoint\n",
568 			    s, __func__);
569 		TCPSTAT_INC(tcps_sc_reset);
570 	} else {
571 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
572 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
573 			    "IRS %u (+WND %u), segment ignored\n",
574 			    s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
575 		TCPSTAT_INC(tcps_badrst);
576 	}
577 
578 done:
579 	if (s != NULL)
580 		free(s, M_TCPLOG);
581 	SCH_UNLOCK(sch);
582 }
583 
584 void
585 syncache_badack(struct in_conninfo *inc)
586 {
587 	struct syncache *sc;
588 	struct syncache_head *sch;
589 
590 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
591 	SCH_LOCK_ASSERT(sch);
592 	if (sc != NULL) {
593 		syncache_drop(sc, sch);
594 		TCPSTAT_INC(tcps_sc_badack);
595 	}
596 	SCH_UNLOCK(sch);
597 }
598 
599 void
600 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
601 {
602 	struct syncache *sc;
603 	struct syncache_head *sch;
604 
605 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
606 	SCH_LOCK_ASSERT(sch);
607 	if (sc == NULL)
608 		goto done;
609 
610 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
611 	if (ntohl(th->th_seq) != sc->sc_iss)
612 		goto done;
613 
614 	/*
615 	 * If we've rertransmitted 3 times and this is our second error,
616 	 * we remove the entry.  Otherwise, we allow it to continue on.
617 	 * This prevents us from incorrectly nuking an entry during a
618 	 * spurious network outage.
619 	 *
620 	 * See tcp_notify().
621 	 */
622 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
623 		sc->sc_flags |= SCF_UNREACH;
624 		goto done;
625 	}
626 	syncache_drop(sc, sch);
627 	TCPSTAT_INC(tcps_sc_unreach);
628 done:
629 	SCH_UNLOCK(sch);
630 }
631 
632 /*
633  * Build a new TCP socket structure from a syncache entry.
634  *
635  * On success return the newly created socket with its underlying inp locked.
636  */
637 static struct socket *
638 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
639 {
640 	struct tcp_function_block *blk;
641 	struct inpcb *inp = NULL;
642 	struct socket *so;
643 	struct tcpcb *tp;
644 	int error;
645 	char *s;
646 
647 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
648 
649 	/*
650 	 * Ok, create the full blown connection, and set things up
651 	 * as they would have been set up if we had created the
652 	 * connection when the SYN arrived.  If we can't create
653 	 * the connection, abort it.
654 	 */
655 	so = sonewconn(lso, 0);
656 	if (so == NULL) {
657 		/*
658 		 * Drop the connection; we will either send a RST or
659 		 * have the peer retransmit its SYN again after its
660 		 * RTO and try again.
661 		 */
662 		TCPSTAT_INC(tcps_listendrop);
663 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
664 			log(LOG_DEBUG, "%s; %s: Socket create failed "
665 			    "due to limits or memory shortage\n",
666 			    s, __func__);
667 			free(s, M_TCPLOG);
668 		}
669 		goto abort2;
670 	}
671 #ifdef MAC
672 	mac_socketpeer_set_from_mbuf(m, so);
673 #endif
674 
675 	inp = sotoinpcb(so);
676 	inp->inp_inc.inc_fibnum = so->so_fibnum;
677 	INP_WLOCK(inp);
678 	/*
679 	 * Exclusive pcbinfo lock is not required in syncache socket case even
680 	 * if two inpcb locks can be acquired simultaneously:
681 	 *  - the inpcb in LISTEN state,
682 	 *  - the newly created inp.
683 	 *
684 	 * In this case, an inp cannot be at same time in LISTEN state and
685 	 * just created by an accept() call.
686 	 */
687 	INP_HASH_WLOCK(&V_tcbinfo);
688 
689 	/* Insert new socket into PCB hash list. */
690 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
691 #ifdef INET6
692 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
693 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
694 	} else {
695 		inp->inp_vflag &= ~INP_IPV6;
696 		inp->inp_vflag |= INP_IPV4;
697 #endif
698 		inp->inp_laddr = sc->sc_inc.inc_laddr;
699 #ifdef INET6
700 	}
701 #endif
702 
703 	/*
704 	 * If there's an mbuf and it has a flowid, then let's initialise the
705 	 * inp with that particular flowid.
706 	 */
707 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
708 		inp->inp_flowid = m->m_pkthdr.flowid;
709 		inp->inp_flowtype = M_HASHTYPE_GET(m);
710 	}
711 
712 	/*
713 	 * Install in the reservation hash table for now, but don't yet
714 	 * install a connection group since the full 4-tuple isn't yet
715 	 * configured.
716 	 */
717 	inp->inp_lport = sc->sc_inc.inc_lport;
718 	if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
719 		/*
720 		 * Undo the assignments above if we failed to
721 		 * put the PCB on the hash lists.
722 		 */
723 #ifdef INET6
724 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
725 			inp->in6p_laddr = in6addr_any;
726 		else
727 #endif
728 			inp->inp_laddr.s_addr = INADDR_ANY;
729 		inp->inp_lport = 0;
730 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
731 			log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
732 			    "with error %i\n",
733 			    s, __func__, error);
734 			free(s, M_TCPLOG);
735 		}
736 		INP_HASH_WUNLOCK(&V_tcbinfo);
737 		goto abort;
738 	}
739 #ifdef IPSEC
740 	/* Copy old policy into new socket's. */
741 	if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
742 		printf("syncache_socket: could not copy policy\n");
743 #endif
744 #ifdef INET6
745 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
746 		struct inpcb *oinp = sotoinpcb(lso);
747 		struct in6_addr laddr6;
748 		struct sockaddr_in6 sin6;
749 		/*
750 		 * Inherit socket options from the listening socket.
751 		 * Note that in6p_inputopts are not (and should not be)
752 		 * copied, since it stores previously received options and is
753 		 * used to detect if each new option is different than the
754 		 * previous one and hence should be passed to a user.
755 		 * If we copied in6p_inputopts, a user would not be able to
756 		 * receive options just after calling the accept system call.
757 		 */
758 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
759 		if (oinp->in6p_outputopts)
760 			inp->in6p_outputopts =
761 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
762 
763 		sin6.sin6_family = AF_INET6;
764 		sin6.sin6_len = sizeof(sin6);
765 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
766 		sin6.sin6_port = sc->sc_inc.inc_fport;
767 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
768 		laddr6 = inp->in6p_laddr;
769 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
770 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
771 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
772 		    thread0.td_ucred, m)) != 0) {
773 			inp->in6p_laddr = laddr6;
774 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
775 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
776 				    "with error %i\n",
777 				    s, __func__, error);
778 				free(s, M_TCPLOG);
779 			}
780 			INP_HASH_WUNLOCK(&V_tcbinfo);
781 			goto abort;
782 		}
783 		/* Override flowlabel from in6_pcbconnect. */
784 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
785 		inp->inp_flow |= sc->sc_flowlabel;
786 	}
787 #endif /* INET6 */
788 #if defined(INET) && defined(INET6)
789 	else
790 #endif
791 #ifdef INET
792 	{
793 		struct in_addr laddr;
794 		struct sockaddr_in sin;
795 
796 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
797 
798 		if (inp->inp_options == NULL) {
799 			inp->inp_options = sc->sc_ipopts;
800 			sc->sc_ipopts = NULL;
801 		}
802 
803 		sin.sin_family = AF_INET;
804 		sin.sin_len = sizeof(sin);
805 		sin.sin_addr = sc->sc_inc.inc_faddr;
806 		sin.sin_port = sc->sc_inc.inc_fport;
807 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
808 		laddr = inp->inp_laddr;
809 		if (inp->inp_laddr.s_addr == INADDR_ANY)
810 			inp->inp_laddr = sc->sc_inc.inc_laddr;
811 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
812 		    thread0.td_ucred, m)) != 0) {
813 			inp->inp_laddr = laddr;
814 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
815 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
816 				    "with error %i\n",
817 				    s, __func__, error);
818 				free(s, M_TCPLOG);
819 			}
820 			INP_HASH_WUNLOCK(&V_tcbinfo);
821 			goto abort;
822 		}
823 	}
824 #endif /* INET */
825 	INP_HASH_WUNLOCK(&V_tcbinfo);
826 	tp = intotcpcb(inp);
827 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
828 	tp->iss = sc->sc_iss;
829 	tp->irs = sc->sc_irs;
830 	tcp_rcvseqinit(tp);
831 	tcp_sendseqinit(tp);
832 	blk = sototcpcb(lso)->t_fb;
833 	if (blk != tp->t_fb) {
834 		/*
835 		 * Our parents t_fb was not the default,
836 		 * we need to release our ref on tp->t_fb and
837 		 * pickup one on the new entry.
838 		 */
839 		struct tcp_function_block *rblk;
840 
841 		rblk = find_and_ref_tcp_fb(blk);
842 		KASSERT(rblk != NULL,
843 		    ("cannot find blk %p out of syncache?", blk));
844 		if (tp->t_fb->tfb_tcp_fb_fini)
845 			(*tp->t_fb->tfb_tcp_fb_fini)(tp);
846 		refcount_release(&tp->t_fb->tfb_refcnt);
847 		tp->t_fb = rblk;
848 		if (tp->t_fb->tfb_tcp_fb_init) {
849 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
850 		}
851 	}
852 	tp->snd_wl1 = sc->sc_irs;
853 	tp->snd_max = tp->iss + 1;
854 	tp->snd_nxt = tp->iss + 1;
855 	tp->rcv_up = sc->sc_irs + 1;
856 	tp->rcv_wnd = sc->sc_wnd;
857 	tp->rcv_adv += tp->rcv_wnd;
858 	tp->last_ack_sent = tp->rcv_nxt;
859 
860 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
861 	if (sc->sc_flags & SCF_NOOPT)
862 		tp->t_flags |= TF_NOOPT;
863 	else {
864 		if (sc->sc_flags & SCF_WINSCALE) {
865 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
866 			tp->snd_scale = sc->sc_requested_s_scale;
867 			tp->request_r_scale = sc->sc_requested_r_scale;
868 		}
869 		if (sc->sc_flags & SCF_TIMESTAMP) {
870 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
871 			tp->ts_recent = sc->sc_tsreflect;
872 			tp->ts_recent_age = tcp_ts_getticks();
873 			tp->ts_offset = sc->sc_tsoff;
874 		}
875 #ifdef TCP_SIGNATURE
876 		if (sc->sc_flags & SCF_SIGNATURE)
877 			tp->t_flags |= TF_SIGNATURE;
878 #endif
879 		if (sc->sc_flags & SCF_SACK)
880 			tp->t_flags |= TF_SACK_PERMIT;
881 	}
882 
883 	if (sc->sc_flags & SCF_ECN)
884 		tp->t_flags |= TF_ECN_PERMIT;
885 
886 	/*
887 	 * Set up MSS and get cached values from tcp_hostcache.
888 	 * This might overwrite some of the defaults we just set.
889 	 */
890 	tcp_mss(tp, sc->sc_peer_mss);
891 
892 	/*
893 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
894 	 * limited to one segment in cc_conn_init().
895 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
896 	 */
897 	if (sc->sc_rxmits > 1)
898 		tp->snd_cwnd = 1;
899 
900 #ifdef TCP_OFFLOAD
901 	/*
902 	 * Allow a TOE driver to install its hooks.  Note that we hold the
903 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
904 	 * new connection before the TOE driver has done its thing.
905 	 */
906 	if (ADDED_BY_TOE(sc)) {
907 		struct toedev *tod = sc->sc_tod;
908 
909 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
910 	}
911 #endif
912 	/*
913 	 * Copy and activate timers.
914 	 */
915 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
916 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
917 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
918 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
919 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
920 
921 	soisconnected(so);
922 
923 	TCPSTAT_INC(tcps_accepts);
924 	return (so);
925 
926 abort:
927 	INP_WUNLOCK(inp);
928 abort2:
929 	if (so != NULL)
930 		soabort(so);
931 	return (NULL);
932 }
933 
934 /*
935  * This function gets called when we receive an ACK for a
936  * socket in the LISTEN state.  We look up the connection
937  * in the syncache, and if its there, we pull it out of
938  * the cache and turn it into a full-blown connection in
939  * the SYN-RECEIVED state.
940  *
941  * On syncache_socket() success the newly created socket
942  * has its underlying inp locked.
943  */
944 int
945 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
946     struct socket **lsop, struct mbuf *m)
947 {
948 	struct syncache *sc;
949 	struct syncache_head *sch;
950 	struct syncache scs;
951 	char *s;
952 
953 	/*
954 	 * Global TCP locks are held because we manipulate the PCB lists
955 	 * and create a new socket.
956 	 */
957 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
958 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
959 	    ("%s: can handle only ACK", __func__));
960 
961 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
962 	SCH_LOCK_ASSERT(sch);
963 
964 #ifdef INVARIANTS
965 	/*
966 	 * Test code for syncookies comparing the syncache stored
967 	 * values with the reconstructed values from the cookie.
968 	 */
969 	if (sc != NULL)
970 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
971 #endif
972 
973 	if (sc == NULL) {
974 		/*
975 		 * There is no syncache entry, so see if this ACK is
976 		 * a returning syncookie.  To do this, first:
977 		 *  A. See if this socket has had a syncache entry dropped in
978 		 *     the past.  We don't want to accept a bogus syncookie
979 		 *     if we've never received a SYN.
980 		 *  B. check that the syncookie is valid.  If it is, then
981 		 *     cobble up a fake syncache entry, and return.
982 		 */
983 		if (!V_tcp_syncookies) {
984 			SCH_UNLOCK(sch);
985 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
986 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
987 				    "segment rejected (syncookies disabled)\n",
988 				    s, __func__);
989 			goto failed;
990 		}
991 		bzero(&scs, sizeof(scs));
992 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
993 		SCH_UNLOCK(sch);
994 		if (sc == NULL) {
995 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
996 				log(LOG_DEBUG, "%s; %s: Segment failed "
997 				    "SYNCOOKIE authentication, segment rejected "
998 				    "(probably spoofed)\n", s, __func__);
999 			goto failed;
1000 		}
1001 	} else {
1002 		/*
1003 		 * Pull out the entry to unlock the bucket row.
1004 		 *
1005 		 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1006 		 * tcp_state_change().  The tcpcb is not existent at this
1007 		 * moment.  A new one will be allocated via syncache_socket->
1008 		 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1009 		 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1010 		 */
1011 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1012 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1013 		sch->sch_length--;
1014 #ifdef TCP_OFFLOAD
1015 		if (ADDED_BY_TOE(sc)) {
1016 			struct toedev *tod = sc->sc_tod;
1017 
1018 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1019 		}
1020 #endif
1021 		SCH_UNLOCK(sch);
1022 	}
1023 
1024 	/*
1025 	 * Segment validation:
1026 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1027 	 */
1028 	if (th->th_ack != sc->sc_iss + 1) {
1029 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1030 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1031 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1032 		goto failed;
1033 	}
1034 
1035 	/*
1036 	 * The SEQ must fall in the window starting at the received
1037 	 * initial receive sequence number + 1 (the SYN).
1038 	 */
1039 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1040 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1041 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1042 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1043 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1044 		goto failed;
1045 	}
1046 
1047 	/*
1048 	 * If timestamps were not negotiated during SYN/ACK they
1049 	 * must not appear on any segment during this session.
1050 	 */
1051 	if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1052 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1053 			log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1054 			    "segment rejected\n", s, __func__);
1055 		goto failed;
1056 	}
1057 
1058 	/*
1059 	 * If timestamps were negotiated during SYN/ACK they should
1060 	 * appear on every segment during this session.
1061 	 * XXXAO: This is only informal as there have been unverified
1062 	 * reports of non-compliants stacks.
1063 	 */
1064 	if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1065 		if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1066 			log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1067 			    "no action\n", s, __func__);
1068 			free(s, M_TCPLOG);
1069 			s = NULL;
1070 		}
1071 	}
1072 
1073 	/*
1074 	 * If timestamps were negotiated the reflected timestamp
1075 	 * must be equal to what we actually sent in the SYN|ACK.
1076 	 */
1077 	if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) {
1078 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1079 			log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1080 			    "segment rejected\n",
1081 			    s, __func__, to->to_tsecr, sc->sc_ts);
1082 		goto failed;
1083 	}
1084 
1085 	*lsop = syncache_socket(sc, *lsop, m);
1086 
1087 	if (*lsop == NULL)
1088 		TCPSTAT_INC(tcps_sc_aborted);
1089 	else
1090 		TCPSTAT_INC(tcps_sc_completed);
1091 
1092 /* how do we find the inp for the new socket? */
1093 	if (sc != &scs)
1094 		syncache_free(sc);
1095 	return (1);
1096 failed:
1097 	if (sc != NULL && sc != &scs)
1098 		syncache_free(sc);
1099 	if (s != NULL)
1100 		free(s, M_TCPLOG);
1101 	*lsop = NULL;
1102 	return (0);
1103 }
1104 
1105 #ifdef TCP_RFC7413
1106 static void
1107 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1108     uint64_t response_cookie)
1109 {
1110 	struct inpcb *inp;
1111 	struct tcpcb *tp;
1112 	unsigned int *pending_counter;
1113 
1114 	/*
1115 	 * Global TCP locks are held because we manipulate the PCB lists
1116 	 * and create a new socket.
1117 	 */
1118 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1119 
1120 	pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1121 	*lsop = syncache_socket(sc, *lsop, m);
1122 	if (*lsop == NULL) {
1123 		TCPSTAT_INC(tcps_sc_aborted);
1124 		atomic_subtract_int(pending_counter, 1);
1125 	} else {
1126 		inp = sotoinpcb(*lsop);
1127 		tp = intotcpcb(inp);
1128 		tp->t_flags |= TF_FASTOPEN;
1129 		tp->t_tfo_cookie = response_cookie;
1130 		tp->snd_max = tp->iss;
1131 		tp->snd_nxt = tp->iss;
1132 		tp->t_tfo_pending = pending_counter;
1133 		TCPSTAT_INC(tcps_sc_completed);
1134 	}
1135 }
1136 #endif /* TCP_RFC7413 */
1137 
1138 /*
1139  * Given a LISTEN socket and an inbound SYN request, add
1140  * this to the syn cache, and send back a segment:
1141  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1142  * to the source.
1143  *
1144  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1145  * Doing so would require that we hold onto the data and deliver it
1146  * to the application.  However, if we are the target of a SYN-flood
1147  * DoS attack, an attacker could send data which would eventually
1148  * consume all available buffer space if it were ACKed.  By not ACKing
1149  * the data, we avoid this DoS scenario.
1150  *
1151  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1152  * cookie is processed, V_tcp_fastopen_enabled set to true, and the
1153  * TCP_FASTOPEN socket option is set.  In this case, a new socket is created
1154  * and returned via lsop, the mbuf is not freed so that tcp_input() can
1155  * queue its data to the socket, and 1 is returned to indicate the
1156  * TFO-socket-creation path was taken.
1157  */
1158 int
1159 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1160     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1161     void *todctx)
1162 {
1163 	struct tcpcb *tp;
1164 	struct socket *so;
1165 	struct syncache *sc = NULL;
1166 	struct syncache_head *sch;
1167 	struct mbuf *ipopts = NULL;
1168 	u_int ltflags;
1169 	int win, sb_hiwat, ip_ttl, ip_tos;
1170 	char *s;
1171 	int rv = 0;
1172 #ifdef INET6
1173 	int autoflowlabel = 0;
1174 #endif
1175 #ifdef MAC
1176 	struct label *maclabel;
1177 #endif
1178 	struct syncache scs;
1179 	struct ucred *cred;
1180 #ifdef TCP_RFC7413
1181 	uint64_t tfo_response_cookie;
1182 	int tfo_cookie_valid = 0;
1183 	int tfo_response_cookie_valid = 0;
1184 #endif
1185 
1186 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1187 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1188 	    ("%s: unexpected tcp flags", __func__));
1189 
1190 	/*
1191 	 * Combine all so/tp operations very early to drop the INP lock as
1192 	 * soon as possible.
1193 	 */
1194 	so = *lsop;
1195 	tp = sototcpcb(so);
1196 	cred = crhold(so->so_cred);
1197 
1198 #ifdef INET6
1199 	if ((inc->inc_flags & INC_ISIPV6) &&
1200 	    (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1201 		autoflowlabel = 1;
1202 #endif
1203 	ip_ttl = inp->inp_ip_ttl;
1204 	ip_tos = inp->inp_ip_tos;
1205 	win = sbspace(&so->so_rcv);
1206 	sb_hiwat = so->so_rcv.sb_hiwat;
1207 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1208 
1209 #ifdef TCP_RFC7413
1210 	if (V_tcp_fastopen_enabled && (tp->t_flags & TF_FASTOPEN) &&
1211 	    (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) {
1212 		/*
1213 		 * Limit the number of pending TFO connections to
1214 		 * approximately half of the queue limit.  This prevents TFO
1215 		 * SYN floods from starving the service by filling the
1216 		 * listen queue with bogus TFO connections.
1217 		 */
1218 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1219 		    (so->so_qlimit / 2)) {
1220 			int result;
1221 
1222 			result = tcp_fastopen_check_cookie(inc,
1223 			    to->to_tfo_cookie, to->to_tfo_len,
1224 			    &tfo_response_cookie);
1225 			tfo_cookie_valid = (result > 0);
1226 			tfo_response_cookie_valid = (result >= 0);
1227 		} else
1228 			atomic_subtract_int(tp->t_tfo_pending, 1);
1229 	}
1230 #endif
1231 
1232 	/* By the time we drop the lock these should no longer be used. */
1233 	so = NULL;
1234 	tp = NULL;
1235 
1236 #ifdef MAC
1237 	if (mac_syncache_init(&maclabel) != 0) {
1238 		INP_WUNLOCK(inp);
1239 		goto done;
1240 	} else
1241 		mac_syncache_create(maclabel, inp);
1242 #endif
1243 #ifdef TCP_RFC7413
1244 	if (!tfo_cookie_valid)
1245 #endif
1246 		INP_WUNLOCK(inp);
1247 
1248 	/*
1249 	 * Remember the IP options, if any.
1250 	 */
1251 #ifdef INET6
1252 	if (!(inc->inc_flags & INC_ISIPV6))
1253 #endif
1254 #ifdef INET
1255 		ipopts = (m) ? ip_srcroute(m) : NULL;
1256 #else
1257 		ipopts = NULL;
1258 #endif
1259 
1260 	/*
1261 	 * See if we already have an entry for this connection.
1262 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1263 	 *
1264 	 * XXX: should the syncache be re-initialized with the contents
1265 	 * of the new SYN here (which may have different options?)
1266 	 *
1267 	 * XXX: We do not check the sequence number to see if this is a
1268 	 * real retransmit or a new connection attempt.  The question is
1269 	 * how to handle such a case; either ignore it as spoofed, or
1270 	 * drop the current entry and create a new one?
1271 	 */
1272 	sc = syncache_lookup(inc, &sch);	/* returns locked entry */
1273 	SCH_LOCK_ASSERT(sch);
1274 	if (sc != NULL) {
1275 #ifdef TCP_RFC7413
1276 		if (tfo_cookie_valid)
1277 			INP_WUNLOCK(inp);
1278 #endif
1279 		TCPSTAT_INC(tcps_sc_dupsyn);
1280 		if (ipopts) {
1281 			/*
1282 			 * If we were remembering a previous source route,
1283 			 * forget it and use the new one we've been given.
1284 			 */
1285 			if (sc->sc_ipopts)
1286 				(void) m_free(sc->sc_ipopts);
1287 			sc->sc_ipopts = ipopts;
1288 		}
1289 		/*
1290 		 * Update timestamp if present.
1291 		 */
1292 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1293 			sc->sc_tsreflect = to->to_tsval;
1294 		else
1295 			sc->sc_flags &= ~SCF_TIMESTAMP;
1296 #ifdef MAC
1297 		/*
1298 		 * Since we have already unconditionally allocated label
1299 		 * storage, free it up.  The syncache entry will already
1300 		 * have an initialized label we can use.
1301 		 */
1302 		mac_syncache_destroy(&maclabel);
1303 #endif
1304 		/* Retransmit SYN|ACK and reset retransmit count. */
1305 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1306 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1307 			    "resetting timer and retransmitting SYN|ACK\n",
1308 			    s, __func__);
1309 			free(s, M_TCPLOG);
1310 		}
1311 		if (syncache_respond(sc, sch, 1, m) == 0) {
1312 			sc->sc_rxmits = 0;
1313 			syncache_timeout(sc, sch, 1);
1314 			TCPSTAT_INC(tcps_sndacks);
1315 			TCPSTAT_INC(tcps_sndtotal);
1316 		}
1317 		SCH_UNLOCK(sch);
1318 		goto done;
1319 	}
1320 
1321 #ifdef TCP_RFC7413
1322 	if (tfo_cookie_valid) {
1323 		bzero(&scs, sizeof(scs));
1324 		sc = &scs;
1325 		goto skip_alloc;
1326 	}
1327 #endif
1328 
1329 	sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1330 	if (sc == NULL) {
1331 		/*
1332 		 * The zone allocator couldn't provide more entries.
1333 		 * Treat this as if the cache was full; drop the oldest
1334 		 * entry and insert the new one.
1335 		 */
1336 		TCPSTAT_INC(tcps_sc_zonefail);
1337 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL)
1338 			syncache_drop(sc, sch);
1339 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1340 		if (sc == NULL) {
1341 			if (V_tcp_syncookies) {
1342 				bzero(&scs, sizeof(scs));
1343 				sc = &scs;
1344 			} else {
1345 				SCH_UNLOCK(sch);
1346 				if (ipopts)
1347 					(void) m_free(ipopts);
1348 				goto done;
1349 			}
1350 		}
1351 	}
1352 
1353 #ifdef TCP_RFC7413
1354 skip_alloc:
1355 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1356 		sc->sc_tfo_cookie = &tfo_response_cookie;
1357 #endif
1358 
1359 	/*
1360 	 * Fill in the syncache values.
1361 	 */
1362 #ifdef MAC
1363 	sc->sc_label = maclabel;
1364 #endif
1365 	sc->sc_cred = cred;
1366 	cred = NULL;
1367 	sc->sc_ipopts = ipopts;
1368 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1369 #ifdef INET6
1370 	if (!(inc->inc_flags & INC_ISIPV6))
1371 #endif
1372 	{
1373 		sc->sc_ip_tos = ip_tos;
1374 		sc->sc_ip_ttl = ip_ttl;
1375 	}
1376 #ifdef TCP_OFFLOAD
1377 	sc->sc_tod = tod;
1378 	sc->sc_todctx = todctx;
1379 #endif
1380 	sc->sc_irs = th->th_seq;
1381 	sc->sc_iss = arc4random();
1382 	sc->sc_flags = 0;
1383 	sc->sc_flowlabel = 0;
1384 
1385 	/*
1386 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1387 	 * win was derived from socket earlier in the function.
1388 	 */
1389 	win = imax(win, 0);
1390 	win = imin(win, TCP_MAXWIN);
1391 	sc->sc_wnd = win;
1392 
1393 	if (V_tcp_do_rfc1323) {
1394 		/*
1395 		 * A timestamp received in a SYN makes
1396 		 * it ok to send timestamp requests and replies.
1397 		 */
1398 		if (to->to_flags & TOF_TS) {
1399 			sc->sc_tsreflect = to->to_tsval;
1400 			sc->sc_ts = tcp_ts_getticks();
1401 			sc->sc_flags |= SCF_TIMESTAMP;
1402 		}
1403 		if (to->to_flags & TOF_SCALE) {
1404 			int wscale = 0;
1405 
1406 			/*
1407 			 * Pick the smallest possible scaling factor that
1408 			 * will still allow us to scale up to sb_max, aka
1409 			 * kern.ipc.maxsockbuf.
1410 			 *
1411 			 * We do this because there are broken firewalls that
1412 			 * will corrupt the window scale option, leading to
1413 			 * the other endpoint believing that our advertised
1414 			 * window is unscaled.  At scale factors larger than
1415 			 * 5 the unscaled window will drop below 1500 bytes,
1416 			 * leading to serious problems when traversing these
1417 			 * broken firewalls.
1418 			 *
1419 			 * With the default maxsockbuf of 256K, a scale factor
1420 			 * of 3 will be chosen by this algorithm.  Those who
1421 			 * choose a larger maxsockbuf should watch out
1422 			 * for the compatibility problems mentioned above.
1423 			 *
1424 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1425 			 * or <SYN,ACK>) segment itself is never scaled.
1426 			 */
1427 			while (wscale < TCP_MAX_WINSHIFT &&
1428 			    (TCP_MAXWIN << wscale) < sb_max)
1429 				wscale++;
1430 			sc->sc_requested_r_scale = wscale;
1431 			sc->sc_requested_s_scale = to->to_wscale;
1432 			sc->sc_flags |= SCF_WINSCALE;
1433 		}
1434 	}
1435 #ifdef TCP_SIGNATURE
1436 	/*
1437 	 * If listening socket requested TCP digests, OR received SYN
1438 	 * contains the option, flag this in the syncache so that
1439 	 * syncache_respond() will do the right thing with the SYN+ACK.
1440 	 */
1441 	if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE)
1442 		sc->sc_flags |= SCF_SIGNATURE;
1443 #endif
1444 	if (to->to_flags & TOF_SACKPERM)
1445 		sc->sc_flags |= SCF_SACK;
1446 	if (to->to_flags & TOF_MSS)
1447 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1448 	if (ltflags & TF_NOOPT)
1449 		sc->sc_flags |= SCF_NOOPT;
1450 	if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1451 		sc->sc_flags |= SCF_ECN;
1452 
1453 	if (V_tcp_syncookies)
1454 		sc->sc_iss = syncookie_generate(sch, sc);
1455 #ifdef INET6
1456 	if (autoflowlabel) {
1457 		if (V_tcp_syncookies)
1458 			sc->sc_flowlabel = sc->sc_iss;
1459 		else
1460 			sc->sc_flowlabel = ip6_randomflowlabel();
1461 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1462 	}
1463 #endif
1464 	SCH_UNLOCK(sch);
1465 
1466 #ifdef TCP_RFC7413
1467 	if (tfo_cookie_valid) {
1468 		syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1469 		/* INP_WUNLOCK(inp) will be performed by the called */
1470 		rv = 1;
1471 		goto tfo_done;
1472 	}
1473 #endif
1474 
1475 	/*
1476 	 * Do a standard 3-way handshake.
1477 	 */
1478 	if (syncache_respond(sc, sch, 0, m) == 0) {
1479 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1480 			syncache_free(sc);
1481 		else if (sc != &scs)
1482 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1483 		TCPSTAT_INC(tcps_sndacks);
1484 		TCPSTAT_INC(tcps_sndtotal);
1485 	} else {
1486 		if (sc != &scs)
1487 			syncache_free(sc);
1488 		TCPSTAT_INC(tcps_sc_dropped);
1489 	}
1490 
1491 done:
1492 	if (m) {
1493 		*lsop = NULL;
1494 		m_freem(m);
1495 	}
1496 #ifdef TCP_RFC7413
1497 tfo_done:
1498 #endif
1499 	if (cred != NULL)
1500 		crfree(cred);
1501 #ifdef MAC
1502 	if (sc == &scs)
1503 		mac_syncache_destroy(&maclabel);
1504 #endif
1505 	return (rv);
1506 }
1507 
1508 /*
1509  * Send SYN|ACK to the peer.  Either in response to the peer's SYN,
1510  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1511  */
1512 static int
1513 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked,
1514     const struct mbuf *m0)
1515 {
1516 	struct ip *ip = NULL;
1517 	struct mbuf *m;
1518 	struct tcphdr *th = NULL;
1519 	int optlen, error = 0;	/* Make compiler happy */
1520 	u_int16_t hlen, tlen, mssopt;
1521 	struct tcpopt to;
1522 #ifdef INET6
1523 	struct ip6_hdr *ip6 = NULL;
1524 #endif
1525 #ifdef TCP_SIGNATURE
1526 	struct secasvar *sav;
1527 #endif
1528 
1529 	hlen =
1530 #ifdef INET6
1531 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1532 #endif
1533 		sizeof(struct ip);
1534 	tlen = hlen + sizeof(struct tcphdr);
1535 
1536 	/* Determine MSS we advertize to other end of connection. */
1537 	mssopt = tcp_mssopt(&sc->sc_inc);
1538 	if (sc->sc_peer_mss)
1539 		mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1540 
1541 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1542 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1543 	    ("syncache: mbuf too small"));
1544 
1545 	/* Create the IP+TCP header from scratch. */
1546 	m = m_gethdr(M_NOWAIT, MT_DATA);
1547 	if (m == NULL)
1548 		return (ENOBUFS);
1549 #ifdef MAC
1550 	mac_syncache_create_mbuf(sc->sc_label, m);
1551 #endif
1552 	m->m_data += max_linkhdr;
1553 	m->m_len = tlen;
1554 	m->m_pkthdr.len = tlen;
1555 	m->m_pkthdr.rcvif = NULL;
1556 
1557 #ifdef INET6
1558 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1559 		ip6 = mtod(m, struct ip6_hdr *);
1560 		ip6->ip6_vfc = IPV6_VERSION;
1561 		ip6->ip6_nxt = IPPROTO_TCP;
1562 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1563 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1564 		ip6->ip6_plen = htons(tlen - hlen);
1565 		/* ip6_hlim is set after checksum */
1566 		ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1567 		ip6->ip6_flow |= sc->sc_flowlabel;
1568 
1569 		th = (struct tcphdr *)(ip6 + 1);
1570 	}
1571 #endif
1572 #if defined(INET6) && defined(INET)
1573 	else
1574 #endif
1575 #ifdef INET
1576 	{
1577 		ip = mtod(m, struct ip *);
1578 		ip->ip_v = IPVERSION;
1579 		ip->ip_hl = sizeof(struct ip) >> 2;
1580 		ip->ip_len = htons(tlen);
1581 		ip->ip_id = 0;
1582 		ip->ip_off = 0;
1583 		ip->ip_sum = 0;
1584 		ip->ip_p = IPPROTO_TCP;
1585 		ip->ip_src = sc->sc_inc.inc_laddr;
1586 		ip->ip_dst = sc->sc_inc.inc_faddr;
1587 		ip->ip_ttl = sc->sc_ip_ttl;
1588 		ip->ip_tos = sc->sc_ip_tos;
1589 
1590 		/*
1591 		 * See if we should do MTU discovery.  Route lookups are
1592 		 * expensive, so we will only unset the DF bit if:
1593 		 *
1594 		 *	1) path_mtu_discovery is disabled
1595 		 *	2) the SCF_UNREACH flag has been set
1596 		 */
1597 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1598 		       ip->ip_off |= htons(IP_DF);
1599 
1600 		th = (struct tcphdr *)(ip + 1);
1601 	}
1602 #endif /* INET */
1603 	th->th_sport = sc->sc_inc.inc_lport;
1604 	th->th_dport = sc->sc_inc.inc_fport;
1605 
1606 	th->th_seq = htonl(sc->sc_iss);
1607 	th->th_ack = htonl(sc->sc_irs + 1);
1608 	th->th_off = sizeof(struct tcphdr) >> 2;
1609 	th->th_x2 = 0;
1610 	th->th_flags = TH_SYN|TH_ACK;
1611 	th->th_win = htons(sc->sc_wnd);
1612 	th->th_urp = 0;
1613 
1614 	if (sc->sc_flags & SCF_ECN) {
1615 		th->th_flags |= TH_ECE;
1616 		TCPSTAT_INC(tcps_ecn_shs);
1617 	}
1618 
1619 	/* Tack on the TCP options. */
1620 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1621 		to.to_flags = 0;
1622 
1623 		to.to_mss = mssopt;
1624 		to.to_flags = TOF_MSS;
1625 		if (sc->sc_flags & SCF_WINSCALE) {
1626 			to.to_wscale = sc->sc_requested_r_scale;
1627 			to.to_flags |= TOF_SCALE;
1628 		}
1629 		if (sc->sc_flags & SCF_TIMESTAMP) {
1630 			/* Virgin timestamp or TCP cookie enhanced one. */
1631 			to.to_tsval = sc->sc_ts;
1632 			to.to_tsecr = sc->sc_tsreflect;
1633 			to.to_flags |= TOF_TS;
1634 		}
1635 		if (sc->sc_flags & SCF_SACK)
1636 			to.to_flags |= TOF_SACKPERM;
1637 #ifdef TCP_SIGNATURE
1638 		sav = NULL;
1639 		if (sc->sc_flags & SCF_SIGNATURE) {
1640 			sav = tcp_get_sav(m, IPSEC_DIR_OUTBOUND);
1641 			if (sav != NULL)
1642 				to.to_flags |= TOF_SIGNATURE;
1643 			else {
1644 
1645 				/*
1646 				 * We've got SCF_SIGNATURE flag
1647 				 * inherited from listening socket,
1648 				 * but no SADB key for given source
1649 				 * address. Assume signature is not
1650 				 * required and remove signature flag
1651 				 * instead of silently dropping
1652 				 * connection.
1653 				 */
1654 				if (locked == 0)
1655 					SCH_LOCK(sch);
1656 				sc->sc_flags &= ~SCF_SIGNATURE;
1657 				if (locked == 0)
1658 					SCH_UNLOCK(sch);
1659 			}
1660 		}
1661 #endif
1662 
1663 #ifdef TCP_RFC7413
1664 		if (sc->sc_tfo_cookie) {
1665 			to.to_flags |= TOF_FASTOPEN;
1666 			to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1667 			to.to_tfo_cookie = sc->sc_tfo_cookie;
1668 			/* don't send cookie again when retransmitting response */
1669 			sc->sc_tfo_cookie = NULL;
1670 		}
1671 #endif
1672 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1673 
1674 		/* Adjust headers by option size. */
1675 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1676 		m->m_len += optlen;
1677 		m->m_pkthdr.len += optlen;
1678 
1679 #ifdef TCP_SIGNATURE
1680 		if (sc->sc_flags & SCF_SIGNATURE)
1681 			tcp_signature_do_compute(m, 0, optlen,
1682 			    to.to_signature, sav);
1683 #endif
1684 #ifdef INET6
1685 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1686 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1687 		else
1688 #endif
1689 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1690 	} else
1691 		optlen = 0;
1692 
1693 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1694 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1695 	/*
1696 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1697 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1698 	 * to SYN|ACK due to lack of inp here.
1699 	 */
1700 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1701 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1702 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1703 	}
1704 #ifdef INET6
1705 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1706 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1707 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1708 		    IPPROTO_TCP, 0);
1709 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1710 #ifdef TCP_OFFLOAD
1711 		if (ADDED_BY_TOE(sc)) {
1712 			struct toedev *tod = sc->sc_tod;
1713 
1714 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1715 
1716 			return (error);
1717 		}
1718 #endif
1719 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1720 	}
1721 #endif
1722 #if defined(INET6) && defined(INET)
1723 	else
1724 #endif
1725 #ifdef INET
1726 	{
1727 		m->m_pkthdr.csum_flags = CSUM_TCP;
1728 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1729 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1730 #ifdef TCP_OFFLOAD
1731 		if (ADDED_BY_TOE(sc)) {
1732 			struct toedev *tod = sc->sc_tod;
1733 
1734 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1735 
1736 			return (error);
1737 		}
1738 #endif
1739 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1740 	}
1741 #endif
1742 	return (error);
1743 }
1744 
1745 /*
1746  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1747  * that exceed the capacity of the syncache by avoiding the storage of any
1748  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
1749  * attacks where the attacker does not have access to our responses.
1750  *
1751  * Syncookies encode and include all necessary information about the
1752  * connection setup within the SYN|ACK that we send back.  That way we
1753  * can avoid keeping any local state until the ACK to our SYN|ACK returns
1754  * (if ever).  Normally the syncache and syncookies are running in parallel
1755  * with the latter taking over when the former is exhausted.  When matching
1756  * syncache entry is found the syncookie is ignored.
1757  *
1758  * The only reliable information persisting the 3WHS is our initial sequence
1759  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
1760  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1761  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
1762  * returns and signifies a legitimate connection if it matches the ACK.
1763  *
1764  * The available space of 32 bits to store the hash and to encode the SYN
1765  * option information is very tight and we should have at least 24 bits for
1766  * the MAC to keep the number of guesses by blind spoofing reasonably high.
1767  *
1768  * SYN option information we have to encode to fully restore a connection:
1769  * MSS: is imporant to chose an optimal segment size to avoid IP level
1770  *   fragmentation along the path.  The common MSS values can be encoded
1771  *   in a 3-bit table.  Uncommon values are captured by the next lower value
1772  *   in the table leading to a slight increase in packetization overhead.
1773  * WSCALE: is necessary to allow large windows to be used for high delay-
1774  *   bandwidth product links.  Not scaling the window when it was initially
1775  *   negotiated is bad for performance as lack of scaling further decreases
1776  *   the apparent available send window.  We only need to encode the WSCALE
1777  *   we received from the remote end.  Our end can be recalculated at any
1778  *   time.  The common WSCALE values can be encoded in a 3-bit table.
1779  *   Uncommon values are captured by the next lower value in the table
1780  *   making us under-estimate the available window size halving our
1781  *   theoretically possible maximum throughput for that connection.
1782  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1783  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1784  *   that are included in all segments on a connection.  We enable them when
1785  *   the ACK has them.
1786  *
1787  * Security of syncookies and attack vectors:
1788  *
1789  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1790  * together with the gloabl secret to make it unique per connection attempt.
1791  * Thus any change of any of those parameters results in a different MAC output
1792  * in an unpredictable way unless a collision is encountered.  24 bits of the
1793  * MAC are embedded into the ISS.
1794  *
1795  * To prevent replay attacks two rotating global secrets are updated with a
1796  * new random value every 15 seconds.  The life-time of a syncookie is thus
1797  * 15-30 seconds.
1798  *
1799  * Vector 1: Attacking the secret.  This requires finding a weakness in the
1800  * MAC itself or the way it is used here.  The attacker can do a chosen plain
1801  * text attack by varying and testing the all parameters under his control.
1802  * The strength depends on the size and randomness of the secret, and the
1803  * cryptographic security of the MAC function.  Due to the constant updating
1804  * of the secret the attacker has at most 29.999 seconds to find the secret
1805  * and launch spoofed connections.  After that he has to start all over again.
1806  *
1807  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
1808  * size an average of 4,823 attempts are required for a 50% chance of success
1809  * to spoof a single syncookie (birthday collision paradox).  However the
1810  * attacker is blind and doesn't know if one of his attempts succeeded unless
1811  * he has a side channel to interfere success from.  A single connection setup
1812  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1813  * This many attempts are required for each one blind spoofed connection.  For
1814  * every additional spoofed connection he has to launch another N attempts.
1815  * Thus for a sustained rate 100 spoofed connections per second approximately
1816  * 1,800,000 packets per second would have to be sent.
1817  *
1818  * NB: The MAC function should be fast so that it doesn't become a CPU
1819  * exhaustion attack vector itself.
1820  *
1821  * References:
1822  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1823  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1824  *   http://cr.yp.to/syncookies.html    (overview)
1825  *   http://cr.yp.to/syncookies/archive (details)
1826  *
1827  *
1828  * Schematic construction of a syncookie enabled Initial Sequence Number:
1829  *  0        1         2         3
1830  *  12345678901234567890123456789012
1831  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1832  *
1833  *  x 24 MAC (truncated)
1834  *  W  3 Send Window Scale index
1835  *  M  3 MSS index
1836  *  S  1 SACK permitted
1837  *  P  1 Odd/even secret
1838  */
1839 
1840 /*
1841  * Distribution and probability of certain MSS values.  Those in between are
1842  * rounded down to the next lower one.
1843  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1844  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
1845  */
1846 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1847 
1848 /*
1849  * Distribution and probability of certain WSCALE values.  We have to map the
1850  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1851  * bits based on prevalence of certain values.  Where we don't have an exact
1852  * match for are rounded down to the next lower one letting us under-estimate
1853  * the true available window.  At the moment this would happen only for the
1854  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1855  * and window size).  The absence of the WSCALE option (no scaling in either
1856  * direction) is encoded with index zero.
1857  * [WSCALE values histograms, Allman, 2012]
1858  *                            X 10 10 35  5  6 14 10%   by host
1859  *                            X 11  4  5  5 18 49  3%   by connections
1860  */
1861 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1862 
1863 /*
1864  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
1865  * and good cryptographic properties.
1866  */
1867 static uint32_t
1868 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1869     uint8_t *secbits, uintptr_t secmod)
1870 {
1871 	SIPHASH_CTX ctx;
1872 	uint32_t siphash[2];
1873 
1874 	SipHash24_Init(&ctx);
1875 	SipHash_SetKey(&ctx, secbits);
1876 	switch (inc->inc_flags & INC_ISIPV6) {
1877 #ifdef INET
1878 	case 0:
1879 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1880 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1881 		break;
1882 #endif
1883 #ifdef INET6
1884 	case INC_ISIPV6:
1885 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1886 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1887 		break;
1888 #endif
1889 	}
1890 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1891 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1892 	SipHash_Update(&ctx, &irs, sizeof(irs));
1893 	SipHash_Update(&ctx, &flags, sizeof(flags));
1894 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
1895 	SipHash_Final((u_int8_t *)&siphash, &ctx);
1896 
1897 	return (siphash[0] ^ siphash[1]);
1898 }
1899 
1900 static tcp_seq
1901 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1902 {
1903 	u_int i, mss, secbit, wscale;
1904 	uint32_t iss, hash;
1905 	uint8_t *secbits;
1906 	union syncookie cookie;
1907 
1908 	SCH_LOCK_ASSERT(sch);
1909 
1910 	cookie.cookie = 0;
1911 
1912 	/* Map our computed MSS into the 3-bit index. */
1913 	mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1914 	for (i = nitems(tcp_sc_msstab) - 1; tcp_sc_msstab[i] > mss && i > 0;
1915 	     i--)
1916 		;
1917 	cookie.flags.mss_idx = i;
1918 
1919 	/*
1920 	 * Map the send window scale into the 3-bit index but only if
1921 	 * the wscale option was received.
1922 	 */
1923 	if (sc->sc_flags & SCF_WINSCALE) {
1924 		wscale = sc->sc_requested_s_scale;
1925 		for (i = nitems(tcp_sc_wstab) - 1;
1926 		    tcp_sc_wstab[i] > wscale && i > 0;
1927 		     i--)
1928 			;
1929 		cookie.flags.wscale_idx = i;
1930 	}
1931 
1932 	/* Can we do SACK? */
1933 	if (sc->sc_flags & SCF_SACK)
1934 		cookie.flags.sack_ok = 1;
1935 
1936 	/* Which of the two secrets to use. */
1937 	secbit = sch->sch_sc->secret.oddeven & 0x1;
1938 	cookie.flags.odd_even = secbit;
1939 
1940 	secbits = sch->sch_sc->secret.key[secbit];
1941 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
1942 	    (uintptr_t)sch);
1943 
1944 	/*
1945 	 * Put the flags into the hash and XOR them to get better ISS number
1946 	 * variance.  This doesn't enhance the cryptographic strength and is
1947 	 * done to prevent the 8 cookie bits from showing up directly on the
1948 	 * wire.
1949 	 */
1950 	iss = hash & ~0xff;
1951 	iss |= cookie.cookie ^ (hash >> 24);
1952 
1953 	/* Randomize the timestamp. */
1954 	if (sc->sc_flags & SCF_TIMESTAMP) {
1955 		sc->sc_ts = arc4random();
1956 		sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
1957 	}
1958 
1959 	TCPSTAT_INC(tcps_sc_sendcookie);
1960 	return (iss);
1961 }
1962 
1963 static struct syncache *
1964 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1965     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1966     struct socket *lso)
1967 {
1968 	uint32_t hash;
1969 	uint8_t *secbits;
1970 	tcp_seq ack, seq;
1971 	int wnd, wscale = 0;
1972 	union syncookie cookie;
1973 
1974 	SCH_LOCK_ASSERT(sch);
1975 
1976 	/*
1977 	 * Pull information out of SYN-ACK/ACK and revert sequence number
1978 	 * advances.
1979 	 */
1980 	ack = th->th_ack - 1;
1981 	seq = th->th_seq - 1;
1982 
1983 	/*
1984 	 * Unpack the flags containing enough information to restore the
1985 	 * connection.
1986 	 */
1987 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
1988 
1989 	/* Which of the two secrets to use. */
1990 	secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
1991 
1992 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
1993 
1994 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
1995 	if ((ack & ~0xff) != (hash & ~0xff))
1996 		return (NULL);
1997 
1998 	/* Fill in the syncache values. */
1999 	sc->sc_flags = 0;
2000 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2001 	sc->sc_ipopts = NULL;
2002 
2003 	sc->sc_irs = seq;
2004 	sc->sc_iss = ack;
2005 
2006 	switch (inc->inc_flags & INC_ISIPV6) {
2007 #ifdef INET
2008 	case 0:
2009 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2010 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2011 		break;
2012 #endif
2013 #ifdef INET6
2014 	case INC_ISIPV6:
2015 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2016 			sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2017 		break;
2018 #endif
2019 	}
2020 
2021 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2022 
2023 	/* We can simply recompute receive window scale we sent earlier. */
2024 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2025 		wscale++;
2026 
2027 	/* Only use wscale if it was enabled in the orignal SYN. */
2028 	if (cookie.flags.wscale_idx > 0) {
2029 		sc->sc_requested_r_scale = wscale;
2030 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2031 		sc->sc_flags |= SCF_WINSCALE;
2032 	}
2033 
2034 	wnd = sbspace(&lso->so_rcv);
2035 	wnd = imax(wnd, 0);
2036 	wnd = imin(wnd, TCP_MAXWIN);
2037 	sc->sc_wnd = wnd;
2038 
2039 	if (cookie.flags.sack_ok)
2040 		sc->sc_flags |= SCF_SACK;
2041 
2042 	if (to->to_flags & TOF_TS) {
2043 		sc->sc_flags |= SCF_TIMESTAMP;
2044 		sc->sc_tsreflect = to->to_tsval;
2045 		sc->sc_ts = to->to_tsecr;
2046 		sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
2047 	}
2048 
2049 	if (to->to_flags & TOF_SIGNATURE)
2050 		sc->sc_flags |= SCF_SIGNATURE;
2051 
2052 	sc->sc_rxmits = 0;
2053 
2054 	TCPSTAT_INC(tcps_sc_recvcookie);
2055 	return (sc);
2056 }
2057 
2058 #ifdef INVARIANTS
2059 static int
2060 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2061     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2062     struct socket *lso)
2063 {
2064 	struct syncache scs, *scx;
2065 	char *s;
2066 
2067 	bzero(&scs, sizeof(scs));
2068 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2069 
2070 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2071 		return (0);
2072 
2073 	if (scx != NULL) {
2074 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2075 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2076 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2077 
2078 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2079 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2080 			    s, __func__, sc->sc_requested_r_scale,
2081 			    scx->sc_requested_r_scale);
2082 
2083 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2084 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2085 			    s, __func__, sc->sc_requested_s_scale,
2086 			    scx->sc_requested_s_scale);
2087 
2088 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2089 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2090 	}
2091 
2092 	if (s != NULL)
2093 		free(s, M_TCPLOG);
2094 	return (0);
2095 }
2096 #endif /* INVARIANTS */
2097 
2098 static void
2099 syncookie_reseed(void *arg)
2100 {
2101 	struct tcp_syncache *sc = arg;
2102 	uint8_t *secbits;
2103 	int secbit;
2104 
2105 	/*
2106 	 * Reseeding the secret doesn't have to be protected by a lock.
2107 	 * It only must be ensured that the new random values are visible
2108 	 * to all CPUs in a SMP environment.  The atomic with release
2109 	 * semantics ensures that.
2110 	 */
2111 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2112 	secbits = sc->secret.key[secbit];
2113 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2114 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2115 
2116 	/* Reschedule ourself. */
2117 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2118 }
2119 
2120 /*
2121  * Exports the syncache entries to userland so that netstat can display
2122  * them alongside the other sockets.  This function is intended to be
2123  * called only from tcp_pcblist.
2124  *
2125  * Due to concurrency on an active system, the number of pcbs exported
2126  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2127  * amount of space the caller allocated for this function to use.
2128  */
2129 int
2130 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2131 {
2132 	struct xtcpcb xt;
2133 	struct syncache *sc;
2134 	struct syncache_head *sch;
2135 	int count, error, i;
2136 
2137 	for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2138 		sch = &V_tcp_syncache.hashbase[i];
2139 		SCH_LOCK(sch);
2140 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2141 			if (count >= max_pcbs) {
2142 				SCH_UNLOCK(sch);
2143 				goto exit;
2144 			}
2145 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2146 				continue;
2147 			bzero(&xt, sizeof(xt));
2148 			xt.xt_len = sizeof(xt);
2149 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2150 				xt.xt_inp.inp_vflag = INP_IPV6;
2151 			else
2152 				xt.xt_inp.inp_vflag = INP_IPV4;
2153 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2154 			xt.xt_tp.t_inpcb = &xt.xt_inp;
2155 			xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2156 			xt.xt_socket.xso_protocol = IPPROTO_TCP;
2157 			xt.xt_socket.xso_len = sizeof (struct xsocket);
2158 			xt.xt_socket.so_type = SOCK_STREAM;
2159 			xt.xt_socket.so_state = SS_ISCONNECTING;
2160 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2161 			if (error) {
2162 				SCH_UNLOCK(sch);
2163 				goto exit;
2164 			}
2165 			count++;
2166 		}
2167 		SCH_UNLOCK(sch);
2168 	}
2169 exit:
2170 	*pcbs_exported = count;
2171 	return error;
2172 }
2173