xref: /freebsd-14.2/sys/netinet/tcp_syncache.c (revision b785f83e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/hash.h>
43 #include <sys/refcount.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
46 #include <sys/limits.h>
47 #include <sys/lock.h>
48 #include <sys/mutex.h>
49 #include <sys/malloc.h>
50 #include <sys/mbuf.h>
51 #include <sys/proc.h>		/* for proc0 declaration */
52 #include <sys/random.h>
53 #include <sys/socket.h>
54 #include <sys/socketvar.h>
55 #include <sys/syslog.h>
56 #include <sys/ucred.h>
57 
58 #include <sys/md5.h>
59 #include <crypto/siphash/siphash.h>
60 
61 #include <vm/uma.h>
62 
63 #include <net/if.h>
64 #include <net/if_var.h>
65 #include <net/route.h>
66 #include <net/vnet.h>
67 
68 #include <netinet/in.h>
69 #include <netinet/in_kdtrace.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 #include <netinet/tcp_fastopen.h>
85 #include <netinet/tcp_fsm.h>
86 #include <netinet/tcp_seq.h>
87 #include <netinet/tcp_timer.h>
88 #include <netinet/tcp_var.h>
89 #include <netinet/tcp_syncache.h>
90 #include <netinet/tcp_ecn.h>
91 #ifdef TCP_BLACKBOX
92 #include <netinet/tcp_log_buf.h>
93 #endif
94 #ifdef TCP_OFFLOAD
95 #include <netinet/toecore.h>
96 #endif
97 #include <netinet/udp.h>
98 
99 #include <netipsec/ipsec_support.h>
100 
101 #include <machine/in_cksum.h>
102 
103 #include <security/mac/mac_framework.h>
104 
105 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
106 #define	V_tcp_syncookies		VNET(tcp_syncookies)
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
108     &VNET_NAME(tcp_syncookies), 0,
109     "Use TCP SYN cookies if the syncache overflows");
110 
111 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
112 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
114     &VNET_NAME(tcp_syncookiesonly), 0,
115     "Use only TCP SYN cookies");
116 
117 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
118 #define V_functions_inherit_listen_socket_stack \
119     VNET(functions_inherit_listen_socket_stack)
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
121     CTLFLAG_VNET | CTLFLAG_RW,
122     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
123     "Inherit listen socket's stack");
124 
125 #ifdef TCP_OFFLOAD
126 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
127 #endif
128 
129 static void	 syncache_drop(struct syncache *, struct syncache_head *);
130 static void	 syncache_free(struct syncache *);
131 static void	 syncache_insert(struct syncache *, struct syncache_head *);
132 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
133 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
134 		    struct mbuf *m);
135 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
136 		    int docallout);
137 static void	 syncache_timer(void *);
138 
139 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
140 		    uint8_t *, uintptr_t);
141 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
142 static struct syncache
143 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
144 		    struct syncache *, struct tcphdr *, struct tcpopt *,
145 		    struct socket *, uint16_t);
146 static void	syncache_pause(struct in_conninfo *);
147 static void	syncache_unpause(void *);
148 static void	 syncookie_reseed(void *);
149 #ifdef INVARIANTS
150 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
151 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
152 		    struct socket *lso, uint16_t port);
153 #endif
154 
155 /*
156  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
157  * 3 retransmits corresponds to a timeout with default values of
158  * tcp_rexmit_initial * (             1 +
159  *                       tcp_backoff[1] +
160  *                       tcp_backoff[2] +
161  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
162  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
163  * the odds are that the user has given up attempting to connect by then.
164  */
165 #define SYNCACHE_MAXREXMTS		3
166 
167 /* Arbitrary values */
168 #define TCP_SYNCACHE_HASHSIZE		512
169 #define TCP_SYNCACHE_BUCKETLIMIT	30
170 
171 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
172 #define	V_tcp_syncache			VNET(tcp_syncache)
173 
174 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
175     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
176     "TCP SYN cache");
177 
178 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
179     &VNET_NAME(tcp_syncache.bucket_limit), 0,
180     "Per-bucket hash limit for syncache");
181 
182 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
183     &VNET_NAME(tcp_syncache.cache_limit), 0,
184     "Overall entry limit for syncache");
185 
186 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
187     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
188 
189 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
190     &VNET_NAME(tcp_syncache.hashsize), 0,
191     "Size of TCP syncache hashtable");
192 
193 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
194     CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
195     "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
196     "and mac(4) checks");
197 
198 static int
sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)199 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
200 {
201 	int error;
202 	u_int new;
203 
204 	new = V_tcp_syncache.rexmt_limit;
205 	error = sysctl_handle_int(oidp, &new, 0, req);
206 	if ((error == 0) && (req->newptr != NULL)) {
207 		if (new > TCP_MAXRXTSHIFT)
208 			error = EINVAL;
209 		else
210 			V_tcp_syncache.rexmt_limit = new;
211 	}
212 	return (error);
213 }
214 
215 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
216     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
217     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
218     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "IU",
219     "Limit on SYN/ACK retransmissions");
220 
221 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
222 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
223     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
224     "Send reset on socket allocation failure");
225 
226 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
227 
228 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
229 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
230 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
231 
232 /*
233  * Requires the syncache entry to be already removed from the bucket list.
234  */
235 static void
syncache_free(struct syncache * sc)236 syncache_free(struct syncache *sc)
237 {
238 
239 	if (sc->sc_ipopts)
240 		(void)m_free(sc->sc_ipopts);
241 	if (sc->sc_cred)
242 		crfree(sc->sc_cred);
243 #ifdef MAC
244 	mac_syncache_destroy(&sc->sc_label);
245 #endif
246 
247 	uma_zfree(V_tcp_syncache.zone, sc);
248 }
249 
250 void
syncache_init(void)251 syncache_init(void)
252 {
253 	int i;
254 
255 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
256 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
257 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
258 	V_tcp_syncache.hash_secret = arc4random();
259 
260 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
261 	    &V_tcp_syncache.hashsize);
262 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
263 	    &V_tcp_syncache.bucket_limit);
264 	if (!powerof2(V_tcp_syncache.hashsize) ||
265 	    V_tcp_syncache.hashsize == 0) {
266 		printf("WARNING: syncache hash size is not a power of 2.\n");
267 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
268 	}
269 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
270 
271 	/* Set limits. */
272 	V_tcp_syncache.cache_limit =
273 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
274 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
275 	    &V_tcp_syncache.cache_limit);
276 
277 	/* Allocate the hash table. */
278 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
279 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
280 
281 #ifdef VIMAGE
282 	V_tcp_syncache.vnet = curvnet;
283 #endif
284 
285 	/* Initialize the hash buckets. */
286 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
287 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
288 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
289 			 NULL, MTX_DEF);
290 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
291 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
292 		V_tcp_syncache.hashbase[i].sch_length = 0;
293 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
294 		V_tcp_syncache.hashbase[i].sch_last_overflow =
295 		    -(SYNCOOKIE_LIFETIME + 1);
296 	}
297 
298 	/* Create the syncache entry zone. */
299 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
300 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
301 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
302 	    V_tcp_syncache.cache_limit);
303 
304 	/* Start the SYN cookie reseeder callout. */
305 	callout_init(&V_tcp_syncache.secret.reseed, 1);
306 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
307 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
308 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
309 	    syncookie_reseed, &V_tcp_syncache);
310 
311 	/* Initialize the pause machinery. */
312 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
313 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
314 	    0);
315 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
316 	V_tcp_syncache.pause_backoff = 0;
317 	V_tcp_syncache.paused = false;
318 }
319 
320 #ifdef VIMAGE
321 void
syncache_destroy(void)322 syncache_destroy(void)
323 {
324 	struct syncache_head *sch;
325 	struct syncache *sc, *nsc;
326 	int i;
327 
328 	/*
329 	 * Stop the re-seed timer before freeing resources.  No need to
330 	 * possibly schedule it another time.
331 	 */
332 	callout_drain(&V_tcp_syncache.secret.reseed);
333 
334 	/* Stop the SYN cache pause callout. */
335 	mtx_lock(&V_tcp_syncache.pause_mtx);
336 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
337 		mtx_unlock(&V_tcp_syncache.pause_mtx);
338 		callout_drain(&V_tcp_syncache.pause_co);
339 	} else
340 		mtx_unlock(&V_tcp_syncache.pause_mtx);
341 
342 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
343 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
344 		sch = &V_tcp_syncache.hashbase[i];
345 		callout_drain(&sch->sch_timer);
346 
347 		SCH_LOCK(sch);
348 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
349 			syncache_drop(sc, sch);
350 		SCH_UNLOCK(sch);
351 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
352 		    ("%s: sch->sch_bucket not empty", __func__));
353 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
354 		    __func__, sch->sch_length));
355 		mtx_destroy(&sch->sch_mtx);
356 	}
357 
358 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
359 	    ("%s: cache_count not 0", __func__));
360 
361 	/* Free the allocated global resources. */
362 	uma_zdestroy(V_tcp_syncache.zone);
363 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
364 	mtx_destroy(&V_tcp_syncache.pause_mtx);
365 }
366 #endif
367 
368 /*
369  * Inserts a syncache entry into the specified bucket row.
370  * Locks and unlocks the syncache_head autonomously.
371  */
372 static void
syncache_insert(struct syncache * sc,struct syncache_head * sch)373 syncache_insert(struct syncache *sc, struct syncache_head *sch)
374 {
375 	struct syncache *sc2;
376 
377 	SCH_LOCK(sch);
378 
379 	/*
380 	 * Make sure that we don't overflow the per-bucket limit.
381 	 * If the bucket is full, toss the oldest element.
382 	 */
383 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
384 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
385 			("sch->sch_length incorrect"));
386 		syncache_pause(&sc->sc_inc);
387 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
388 		sch->sch_last_overflow = time_uptime;
389 		syncache_drop(sc2, sch);
390 	}
391 
392 	/* Put it into the bucket. */
393 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
394 	sch->sch_length++;
395 
396 #ifdef TCP_OFFLOAD
397 	if (ADDED_BY_TOE(sc)) {
398 		struct toedev *tod = sc->sc_tod;
399 
400 		tod->tod_syncache_added(tod, sc->sc_todctx);
401 	}
402 #endif
403 
404 	/* Reinitialize the bucket row's timer. */
405 	if (sch->sch_length == 1)
406 		sch->sch_nextc = ticks + INT_MAX;
407 	syncache_timeout(sc, sch, 1);
408 
409 	SCH_UNLOCK(sch);
410 
411 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
412 	TCPSTAT_INC(tcps_sc_added);
413 }
414 
415 /*
416  * Remove and free entry from syncache bucket row.
417  * Expects locked syncache head.
418  */
419 static void
syncache_drop(struct syncache * sc,struct syncache_head * sch)420 syncache_drop(struct syncache *sc, struct syncache_head *sch)
421 {
422 
423 	SCH_LOCK_ASSERT(sch);
424 
425 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
426 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
427 	sch->sch_length--;
428 
429 #ifdef TCP_OFFLOAD
430 	if (ADDED_BY_TOE(sc)) {
431 		struct toedev *tod = sc->sc_tod;
432 
433 		tod->tod_syncache_removed(tod, sc->sc_todctx);
434 	}
435 #endif
436 
437 	syncache_free(sc);
438 }
439 
440 /*
441  * Engage/reengage time on bucket row.
442  */
443 static void
syncache_timeout(struct syncache * sc,struct syncache_head * sch,int docallout)444 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
445 {
446 	int rexmt;
447 
448 	if (sc->sc_rxmits == 0)
449 		rexmt = tcp_rexmit_initial;
450 	else
451 		TCPT_RANGESET(rexmt,
452 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
453 		    tcp_rexmit_min, TCPTV_REXMTMAX);
454 	sc->sc_rxttime = ticks + rexmt;
455 	sc->sc_rxmits++;
456 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
457 		sch->sch_nextc = sc->sc_rxttime;
458 		if (docallout)
459 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
460 			    syncache_timer, (void *)sch);
461 	}
462 }
463 
464 /*
465  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
466  * If we have retransmitted an entry the maximum number of times, expire it.
467  * One separate timer for each bucket row.
468  */
469 static void
syncache_timer(void * xsch)470 syncache_timer(void *xsch)
471 {
472 	struct syncache_head *sch = (struct syncache_head *)xsch;
473 	struct syncache *sc, *nsc;
474 	struct epoch_tracker et;
475 	int tick = ticks;
476 	char *s;
477 	bool paused;
478 
479 	CURVNET_SET(sch->sch_sc->vnet);
480 
481 	/* NB: syncache_head has already been locked by the callout. */
482 	SCH_LOCK_ASSERT(sch);
483 
484 	/*
485 	 * In the following cycle we may remove some entries and/or
486 	 * advance some timeouts, so re-initialize the bucket timer.
487 	 */
488 	sch->sch_nextc = tick + INT_MAX;
489 
490 	/*
491 	 * If we have paused processing, unconditionally remove
492 	 * all syncache entries.
493 	 */
494 	mtx_lock(&V_tcp_syncache.pause_mtx);
495 	paused = V_tcp_syncache.paused;
496 	mtx_unlock(&V_tcp_syncache.pause_mtx);
497 
498 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
499 		if (paused) {
500 			syncache_drop(sc, sch);
501 			continue;
502 		}
503 		/*
504 		 * We do not check if the listen socket still exists
505 		 * and accept the case where the listen socket may be
506 		 * gone by the time we resend the SYN/ACK.  We do
507 		 * not expect this to happens often. If it does,
508 		 * then the RST will be sent by the time the remote
509 		 * host does the SYN/ACK->ACK.
510 		 */
511 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
512 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
513 				sch->sch_nextc = sc->sc_rxttime;
514 			continue;
515 		}
516 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
517 			sc->sc_flags &= ~SCF_ECN_MASK;
518 		}
519 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
520 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
521 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
522 				    "giving up and removing syncache entry\n",
523 				    s, __func__);
524 				free(s, M_TCPLOG);
525 			}
526 			syncache_drop(sc, sch);
527 			TCPSTAT_INC(tcps_sc_stale);
528 			continue;
529 		}
530 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
531 			log(LOG_DEBUG, "%s; %s: Response timeout, "
532 			    "retransmitting (%u) SYN|ACK\n",
533 			    s, __func__, sc->sc_rxmits);
534 			free(s, M_TCPLOG);
535 		}
536 
537 		NET_EPOCH_ENTER(et);
538 		if (syncache_respond(sc, NULL, TH_SYN|TH_ACK) == 0) {
539 			syncache_timeout(sc, sch, 0);
540 			TCPSTAT_INC(tcps_sndacks);
541 			TCPSTAT_INC(tcps_sndtotal);
542 			TCPSTAT_INC(tcps_sc_retransmitted);
543 		} else {
544 			syncache_drop(sc, sch);
545 			TCPSTAT_INC(tcps_sc_dropped);
546 		}
547 		NET_EPOCH_EXIT(et);
548 	}
549 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
550 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
551 			syncache_timer, (void *)(sch));
552 	CURVNET_RESTORE();
553 }
554 
555 /*
556  * Returns true if the system is only using cookies at the moment.
557  * This could be due to a sysadmin decision to only use cookies, or it
558  * could be due to the system detecting an attack.
559  */
560 static inline bool
syncache_cookiesonly(void)561 syncache_cookiesonly(void)
562 {
563 
564 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
565 	    V_tcp_syncookiesonly));
566 }
567 
568 /*
569  * Find the hash bucket for the given connection.
570  */
571 static struct syncache_head *
syncache_hashbucket(struct in_conninfo * inc)572 syncache_hashbucket(struct in_conninfo *inc)
573 {
574 	uint32_t hash;
575 
576 	/*
577 	 * The hash is built on foreign port + local port + foreign address.
578 	 * We rely on the fact that struct in_conninfo starts with 16 bits
579 	 * of foreign port, then 16 bits of local port then followed by 128
580 	 * bits of foreign address.  In case of IPv4 address, the first 3
581 	 * 32-bit words of the address always are zeroes.
582 	 */
583 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
584 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
585 
586 	return (&V_tcp_syncache.hashbase[hash]);
587 }
588 
589 /*
590  * Find an entry in the syncache.
591  * Returns always with locked syncache_head plus a matching entry or NULL.
592  */
593 static struct syncache *
syncache_lookup(struct in_conninfo * inc,struct syncache_head ** schp)594 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
595 {
596 	struct syncache *sc;
597 	struct syncache_head *sch;
598 
599 	*schp = sch = syncache_hashbucket(inc);
600 	SCH_LOCK(sch);
601 
602 	/* Circle through bucket row to find matching entry. */
603 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
604 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
605 		    sizeof(struct in_endpoints)) == 0)
606 			break;
607 
608 	return (sc);	/* Always returns with locked sch. */
609 }
610 
611 /*
612  * This function is called when we get a RST for a
613  * non-existent connection, so that we can see if the
614  * connection is in the syn cache.  If it is, zap it.
615  * If required send a challenge ACK.
616  */
617 void
syncache_chkrst(struct in_conninfo * inc,struct tcphdr * th,struct mbuf * m,uint16_t port)618 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
619     uint16_t port)
620 {
621 	struct syncache *sc;
622 	struct syncache_head *sch;
623 	char *s = NULL;
624 
625 	if (syncache_cookiesonly())
626 		return;
627 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
628 	SCH_LOCK_ASSERT(sch);
629 
630 	/*
631 	 * No corresponding connection was found in syncache.
632 	 * If syncookies are enabled and possibly exclusively
633 	 * used, or we are under memory pressure, a valid RST
634 	 * may not find a syncache entry.  In that case we're
635 	 * done and no SYN|ACK retransmissions will happen.
636 	 * Otherwise the RST was misdirected or spoofed.
637 	 */
638 	if (sc == NULL) {
639 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
640 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
641 			    "syncache entry (possibly syncookie only), "
642 			    "segment ignored\n", s, __func__);
643 		TCPSTAT_INC(tcps_badrst);
644 		goto done;
645 	}
646 
647 	/* The remote UDP encaps port does not match. */
648 	if (sc->sc_port != port) {
649 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
650 			log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
651 			    "syncache entry but non-matching UDP encaps port, "
652 			    "segment ignored\n", s, __func__);
653 		TCPSTAT_INC(tcps_badrst);
654 		goto done;
655 	}
656 
657 	/*
658 	 * If the RST bit is set, check the sequence number to see
659 	 * if this is a valid reset segment.
660 	 *
661 	 * RFC 793 page 37:
662 	 *   In all states except SYN-SENT, all reset (RST) segments
663 	 *   are validated by checking their SEQ-fields.  A reset is
664 	 *   valid if its sequence number is in the window.
665 	 *
666 	 * RFC 793 page 69:
667 	 *   There are four cases for the acceptability test for an incoming
668 	 *   segment:
669 	 *
670 	 * Segment Receive  Test
671 	 * Length  Window
672 	 * ------- -------  -------------------------------------------
673 	 *    0       0     SEG.SEQ = RCV.NXT
674 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
675 	 *   >0       0     not acceptable
676 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
677 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
678 	 *
679 	 * Note that when receiving a SYN segment in the LISTEN state,
680 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
681 	 * described in RFC 793, page 66.
682 	 */
683 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
684 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
685 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
686 		if (V_tcp_insecure_rst ||
687 		    th->th_seq == sc->sc_irs + 1) {
688 			syncache_drop(sc, sch);
689 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
690 				log(LOG_DEBUG,
691 				    "%s; %s: Our SYN|ACK was rejected, "
692 				    "connection attempt aborted by remote "
693 				    "endpoint\n",
694 				    s, __func__);
695 			TCPSTAT_INC(tcps_sc_reset);
696 		} else {
697 			TCPSTAT_INC(tcps_badrst);
698 			/* Send challenge ACK. */
699 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
700 				log(LOG_DEBUG, "%s; %s: RST with invalid "
701 				    " SEQ %u != NXT %u (+WND %u), "
702 				    "sending challenge ACK\n",
703 				    s, __func__,
704 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
705 			if (syncache_respond(sc, m, TH_ACK) == 0) {
706 				TCPSTAT_INC(tcps_sndacks);
707 				TCPSTAT_INC(tcps_sndtotal);
708 			} else {
709 				syncache_drop(sc, sch);
710 				TCPSTAT_INC(tcps_sc_dropped);
711 			}
712 		}
713 	} else {
714 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
715 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
716 			    "NXT %u (+WND %u), segment ignored\n",
717 			    s, __func__,
718 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
719 		TCPSTAT_INC(tcps_badrst);
720 	}
721 
722 done:
723 	if (s != NULL)
724 		free(s, M_TCPLOG);
725 	SCH_UNLOCK(sch);
726 }
727 
728 void
syncache_badack(struct in_conninfo * inc,uint16_t port)729 syncache_badack(struct in_conninfo *inc, uint16_t port)
730 {
731 	struct syncache *sc;
732 	struct syncache_head *sch;
733 
734 	if (syncache_cookiesonly())
735 		return;
736 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
737 	SCH_LOCK_ASSERT(sch);
738 	if ((sc != NULL) && (sc->sc_port == port)) {
739 		syncache_drop(sc, sch);
740 		TCPSTAT_INC(tcps_sc_badack);
741 	}
742 	SCH_UNLOCK(sch);
743 }
744 
745 void
syncache_unreach(struct in_conninfo * inc,tcp_seq th_seq,uint16_t port)746 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
747 {
748 	struct syncache *sc;
749 	struct syncache_head *sch;
750 
751 	if (syncache_cookiesonly())
752 		return;
753 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
754 	SCH_LOCK_ASSERT(sch);
755 	if (sc == NULL)
756 		goto done;
757 
758 	/* If the port != sc_port, then it's a bogus ICMP msg */
759 	if (port != sc->sc_port)
760 		goto done;
761 
762 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
763 	if (ntohl(th_seq) != sc->sc_iss)
764 		goto done;
765 
766 	/*
767 	 * If we've rertransmitted 3 times and this is our second error,
768 	 * we remove the entry.  Otherwise, we allow it to continue on.
769 	 * This prevents us from incorrectly nuking an entry during a
770 	 * spurious network outage.
771 	 *
772 	 * See tcp_notify().
773 	 */
774 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
775 		sc->sc_flags |= SCF_UNREACH;
776 		goto done;
777 	}
778 	syncache_drop(sc, sch);
779 	TCPSTAT_INC(tcps_sc_unreach);
780 done:
781 	SCH_UNLOCK(sch);
782 }
783 
784 /*
785  * Build a new TCP socket structure from a syncache entry.
786  *
787  * On success return the newly created socket with its underlying inp locked.
788  */
789 static struct socket *
syncache_socket(struct syncache * sc,struct socket * lso,struct mbuf * m)790 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
791 {
792 	struct tcpcb *listening_tcb;
793 	struct inpcb *inp = NULL;
794 	struct socket *so;
795 	struct tcpcb *tp;
796 	int error;
797 	char *s;
798 
799 	NET_EPOCH_ASSERT();
800 
801 	/*
802 	 * Ok, create the full blown connection, and set things up
803 	 * as they would have been set up if we had created the
804 	 * connection when the SYN arrived.
805 	 */
806 	if ((so = solisten_clone(lso)) == NULL)
807 		goto allocfail;
808 #ifdef MAC
809 	mac_socketpeer_set_from_mbuf(m, so);
810 #endif
811 	error = in_pcballoc(so, &V_tcbinfo);
812 	if (error) {
813 		sodealloc(so);
814 		goto allocfail;
815 	}
816 	inp = sotoinpcb(so);
817 	if (V_functions_inherit_listen_socket_stack)
818 		listening_tcb = sototcpcb(lso);
819 	else
820 		listening_tcb = NULL;
821 	if ((tp = tcp_newtcpcb(inp, listening_tcb)) == NULL) {
822 		in_pcbfree(inp);
823 		sodealloc(so);
824 		goto allocfail;
825 	}
826 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
827 #ifdef INET6
828 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
829 		inp->inp_vflag &= ~INP_IPV4;
830 		inp->inp_vflag |= INP_IPV6;
831 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
832 	} else {
833 		inp->inp_vflag &= ~INP_IPV6;
834 		inp->inp_vflag |= INP_IPV4;
835 #endif
836 		inp->inp_ip_ttl = sc->sc_ip_ttl;
837 		inp->inp_ip_tos = sc->sc_ip_tos;
838 		inp->inp_laddr = sc->sc_inc.inc_laddr;
839 #ifdef INET6
840 	}
841 #endif
842 
843 	/*
844 	 * If there's an mbuf and it has a flowid, then let's initialise the
845 	 * inp with that particular flowid.
846 	 */
847 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
848 		inp->inp_flowid = m->m_pkthdr.flowid;
849 		inp->inp_flowtype = M_HASHTYPE_GET(m);
850 #ifdef NUMA
851 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
852 #endif
853 	}
854 
855 	inp->inp_lport = sc->sc_inc.inc_lport;
856 #ifdef INET6
857 	if (inp->inp_vflag & INP_IPV6PROTO) {
858 		struct inpcb *oinp = sotoinpcb(lso);
859 
860 		/*
861 		 * Inherit socket options from the listening socket.
862 		 * Note that in6p_inputopts are not (and should not be)
863 		 * copied, since it stores previously received options and is
864 		 * used to detect if each new option is different than the
865 		 * previous one and hence should be passed to a user.
866 		 * If we copied in6p_inputopts, a user would not be able to
867 		 * receive options just after calling the accept system call.
868 		 */
869 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
870 		if (oinp->in6p_outputopts)
871 			inp->in6p_outputopts =
872 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
873 		inp->in6p_hops = oinp->in6p_hops;
874 	}
875 
876 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
877 		struct sockaddr_in6 sin6;
878 
879 		sin6.sin6_family = AF_INET6;
880 		sin6.sin6_len = sizeof(sin6);
881 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
882 		sin6.sin6_port = sc->sc_inc.inc_fport;
883 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
884 		INP_HASH_WLOCK(&V_tcbinfo);
885 		error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
886 		INP_HASH_WUNLOCK(&V_tcbinfo);
887 		if (error != 0)
888 			goto abort;
889 		/* Override flowlabel from in6_pcbconnect. */
890 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
891 		inp->inp_flow |= sc->sc_flowlabel;
892 	}
893 #endif /* INET6 */
894 #if defined(INET) && defined(INET6)
895 	else
896 #endif
897 #ifdef INET
898 	{
899 		struct sockaddr_in sin;
900 
901 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
902 
903 		if (inp->inp_options == NULL) {
904 			inp->inp_options = sc->sc_ipopts;
905 			sc->sc_ipopts = NULL;
906 		}
907 
908 		sin.sin_family = AF_INET;
909 		sin.sin_len = sizeof(sin);
910 		sin.sin_addr = sc->sc_inc.inc_faddr;
911 		sin.sin_port = sc->sc_inc.inc_fport;
912 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
913 		INP_HASH_WLOCK(&V_tcbinfo);
914 		error = in_pcbconnect(inp, &sin, thread0.td_ucred, false);
915 		INP_HASH_WUNLOCK(&V_tcbinfo);
916 		if (error != 0)
917 			goto abort;
918 	}
919 #endif /* INET */
920 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
921 	/* Copy old policy into new socket's. */
922 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
923 		printf("syncache_socket: could not copy policy\n");
924 #endif
925 	tp->t_state = TCPS_SYN_RECEIVED;
926 	tp->iss = sc->sc_iss;
927 	tp->irs = sc->sc_irs;
928 	tp->t_port = sc->sc_port;
929 	tcp_rcvseqinit(tp);
930 	tcp_sendseqinit(tp);
931 	tp->snd_wl1 = sc->sc_irs;
932 	tp->snd_max = tp->iss + 1;
933 	tp->snd_nxt = tp->iss + 1;
934 	tp->rcv_up = sc->sc_irs + 1;
935 	tp->rcv_wnd = sc->sc_wnd;
936 	tp->rcv_adv += tp->rcv_wnd;
937 	tp->last_ack_sent = tp->rcv_nxt;
938 
939 	tp->t_flags = sototcpcb(lso)->t_flags &
940 	    (TF_LRD|TF_NOPUSH|TF_NODELAY);
941 	if (sc->sc_flags & SCF_NOOPT)
942 		tp->t_flags |= TF_NOOPT;
943 	else {
944 		if (sc->sc_flags & SCF_WINSCALE) {
945 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
946 			tp->snd_scale = sc->sc_requested_s_scale;
947 			tp->request_r_scale = sc->sc_requested_r_scale;
948 		}
949 		if (sc->sc_flags & SCF_TIMESTAMP) {
950 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
951 			tp->ts_recent = sc->sc_tsreflect;
952 			tp->ts_recent_age = tcp_ts_getticks();
953 			tp->ts_offset = sc->sc_tsoff;
954 		}
955 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
956 		if (sc->sc_flags & SCF_SIGNATURE)
957 			tp->t_flags |= TF_SIGNATURE;
958 #endif
959 		if (sc->sc_flags & SCF_SACK)
960 			tp->t_flags |= TF_SACK_PERMIT;
961 	}
962 
963 	tcp_ecn_syncache_socket(tp, sc);
964 
965 	/*
966 	 * Set up MSS and get cached values from tcp_hostcache.
967 	 * This might overwrite some of the defaults we just set.
968 	 */
969 	tcp_mss(tp, sc->sc_peer_mss);
970 
971 	/*
972 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
973 	 * limited to one segment in cc_conn_init().
974 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
975 	 */
976 	if (sc->sc_rxmits > 1)
977 		tp->snd_cwnd = 1;
978 
979 #ifdef TCP_OFFLOAD
980 	/*
981 	 * Allow a TOE driver to install its hooks.  Note that we hold the
982 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
983 	 * new connection before the TOE driver has done its thing.
984 	 */
985 	if (ADDED_BY_TOE(sc)) {
986 		struct toedev *tod = sc->sc_tod;
987 
988 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
989 	}
990 #endif
991 #ifdef TCP_BLACKBOX
992 	/*
993 	 * Inherit the log state from the listening socket, if
994 	 * - the log state of the listening socket is not off and
995 	 * - the listening socket was not auto selected from all sessions and
996 	 * - a log id is not set on the listening socket.
997 	 * This avoids inheriting a log state which was automatically set.
998 	 */
999 	if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
1000 	    ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
1001 	    (sototcpcb(lso)->t_lib == NULL)) {
1002 		tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
1003 	}
1004 #endif
1005 	/*
1006 	 * Copy and activate timers.
1007 	 */
1008 	tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
1009 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1010 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1011 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1012 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1013 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1014 
1015 	TCPSTAT_INC(tcps_accepts);
1016 	TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1017 
1018 	if (!solisten_enqueue(so, SS_ISCONNECTED))
1019 		tp->t_flags |= TF_SONOTCONN;
1020 
1021 	return (so);
1022 
1023 allocfail:
1024 	/*
1025 	 * Drop the connection; we will either send a RST or have the peer
1026 	 * retransmit its SYN again after its RTO and try again.
1027 	 */
1028 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1029 		log(LOG_DEBUG, "%s; %s: Socket create failed "
1030 		    "due to limits or memory shortage\n",
1031 		    s, __func__);
1032 		free(s, M_TCPLOG);
1033 	}
1034 	TCPSTAT_INC(tcps_listendrop);
1035 	return (NULL);
1036 
1037 abort:
1038 	tcp_discardcb(tp);
1039 	in_pcbfree(inp);
1040 	sodealloc(so);
1041 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1042 		log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1043 		    s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1044 		    error);
1045 		free(s, M_TCPLOG);
1046 	}
1047 	TCPSTAT_INC(tcps_listendrop);
1048 	return (NULL);
1049 }
1050 
1051 /*
1052  * This function gets called when we receive an ACK for a
1053  * socket in the LISTEN state.  We look up the connection
1054  * in the syncache, and if its there, we pull it out of
1055  * the cache and turn it into a full-blown connection in
1056  * the SYN-RECEIVED state.
1057  *
1058  * On syncache_socket() success the newly created socket
1059  * has its underlying inp locked.
1060  */
1061 int
syncache_expand(struct in_conninfo * inc,struct tcpopt * to,struct tcphdr * th,struct socket ** lsop,struct mbuf * m,uint16_t port)1062 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1063     struct socket **lsop, struct mbuf *m, uint16_t port)
1064 {
1065 	struct syncache *sc;
1066 	struct syncache_head *sch;
1067 	struct syncache scs;
1068 	char *s;
1069 	bool locked;
1070 
1071 	NET_EPOCH_ASSERT();
1072 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1073 	    ("%s: can handle only ACK", __func__));
1074 
1075 	if (syncache_cookiesonly()) {
1076 		sc = NULL;
1077 		sch = syncache_hashbucket(inc);
1078 		locked = false;
1079 	} else {
1080 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1081 		locked = true;
1082 		SCH_LOCK_ASSERT(sch);
1083 	}
1084 
1085 #ifdef INVARIANTS
1086 	/*
1087 	 * Test code for syncookies comparing the syncache stored
1088 	 * values with the reconstructed values from the cookie.
1089 	 */
1090 	if (sc != NULL)
1091 		syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1092 #endif
1093 
1094 	if (sc == NULL) {
1095 		/*
1096 		 * There is no syncache entry, so see if this ACK is
1097 		 * a returning syncookie.  To do this, first:
1098 		 *  A. Check if syncookies are used in case of syncache
1099 		 *     overflows
1100 		 *  B. See if this socket has had a syncache entry dropped in
1101 		 *     the recent past. We don't want to accept a bogus
1102 		 *     syncookie if we've never received a SYN or accept it
1103 		 *     twice.
1104 		 *  C. check that the syncookie is valid.  If it is, then
1105 		 *     cobble up a fake syncache entry, and return.
1106 		 */
1107 		if (locked && !V_tcp_syncookies) {
1108 			SCH_UNLOCK(sch);
1109 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1110 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1111 				    "segment rejected (syncookies disabled)\n",
1112 				    s, __func__);
1113 			goto failed;
1114 		}
1115 		if (locked && !V_tcp_syncookiesonly &&
1116 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1117 			SCH_UNLOCK(sch);
1118 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1119 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1120 				    "segment rejected (no syncache entry)\n",
1121 				    s, __func__);
1122 			goto failed;
1123 		}
1124 		bzero(&scs, sizeof(scs));
1125 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1126 		if (locked)
1127 			SCH_UNLOCK(sch);
1128 		if (sc == NULL) {
1129 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1130 				log(LOG_DEBUG, "%s; %s: Segment failed "
1131 				    "SYNCOOKIE authentication, segment rejected "
1132 				    "(probably spoofed)\n", s, __func__);
1133 			goto failed;
1134 		}
1135 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1136 		/* If received ACK has MD5 signature, check it. */
1137 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1138 		    (!TCPMD5_ENABLED() ||
1139 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1140 			/* Drop the ACK. */
1141 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1142 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1143 				    "MD5 signature doesn't match.\n",
1144 				    s, __func__);
1145 				free(s, M_TCPLOG);
1146 			}
1147 			TCPSTAT_INC(tcps_sig_err_sigopt);
1148 			return (-1); /* Do not send RST */
1149 		}
1150 #endif /* TCP_SIGNATURE */
1151 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1152 	} else {
1153 		if (sc->sc_port != port) {
1154 			SCH_UNLOCK(sch);
1155 			return (0);
1156 		}
1157 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1158 		/*
1159 		 * If listening socket requested TCP digests, check that
1160 		 * received ACK has signature and it is correct.
1161 		 * If not, drop the ACK and leave sc entry in th cache,
1162 		 * because SYN was received with correct signature.
1163 		 */
1164 		if (sc->sc_flags & SCF_SIGNATURE) {
1165 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1166 				/* No signature */
1167 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1168 				SCH_UNLOCK(sch);
1169 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1170 					log(LOG_DEBUG, "%s; %s: Segment "
1171 					    "rejected, MD5 signature wasn't "
1172 					    "provided.\n", s, __func__);
1173 					free(s, M_TCPLOG);
1174 				}
1175 				return (-1); /* Do not send RST */
1176 			}
1177 			if (!TCPMD5_ENABLED() ||
1178 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1179 				/* Doesn't match or no SA */
1180 				SCH_UNLOCK(sch);
1181 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1182 					log(LOG_DEBUG, "%s; %s: Segment "
1183 					    "rejected, MD5 signature doesn't "
1184 					    "match.\n", s, __func__);
1185 					free(s, M_TCPLOG);
1186 				}
1187 				return (-1); /* Do not send RST */
1188 			}
1189 		}
1190 #endif /* TCP_SIGNATURE */
1191 
1192 		/*
1193 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1194 		 * it's less than ts_recent, drop it.
1195 		 * XXXMT: RFC 7323 also requires to send an ACK.
1196 		 *        In tcp_input.c this is only done for TCP segments
1197 		 *        with user data, so be consistent here and just drop
1198 		 *        the segment.
1199 		 */
1200 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1201 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1202 			SCH_UNLOCK(sch);
1203 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1204 				log(LOG_DEBUG,
1205 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1206 				    "segment dropped\n", s, __func__,
1207 				    to->to_tsval, sc->sc_tsreflect);
1208 				free(s, M_TCPLOG);
1209 			}
1210 			return (-1);  /* Do not send RST */
1211 		}
1212 
1213 		/*
1214 		 * If timestamps were not negotiated during SYN/ACK and a
1215 		 * segment with a timestamp is received, ignore the
1216 		 * timestamp and process the packet normally.
1217 		 * See section 3.2 of RFC 7323.
1218 		 */
1219 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1220 		    (to->to_flags & TOF_TS)) {
1221 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1222 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1223 				    "expected, segment processed normally\n",
1224 				    s, __func__);
1225 				free(s, M_TCPLOG);
1226 				s = NULL;
1227 			}
1228 		}
1229 
1230 		/*
1231 		 * If timestamps were negotiated during SYN/ACK and a
1232 		 * segment without a timestamp is received, silently drop
1233 		 * the segment, unless the missing timestamps are tolerated.
1234 		 * See section 3.2 of RFC 7323.
1235 		 */
1236 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1237 		    !(to->to_flags & TOF_TS)) {
1238 			if (V_tcp_tolerate_missing_ts) {
1239 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1240 					log(LOG_DEBUG,
1241 					    "%s; %s: Timestamp missing, "
1242 					    "segment processed normally\n",
1243 					    s, __func__);
1244 					free(s, M_TCPLOG);
1245 				}
1246 			} else {
1247 				SCH_UNLOCK(sch);
1248 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1249 					log(LOG_DEBUG,
1250 					    "%s; %s: Timestamp missing, "
1251 					    "segment silently dropped\n",
1252 					    s, __func__);
1253 					free(s, M_TCPLOG);
1254 				}
1255 				return (-1);  /* Do not send RST */
1256 			}
1257 		}
1258 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1259 		sch->sch_length--;
1260 #ifdef TCP_OFFLOAD
1261 		if (ADDED_BY_TOE(sc)) {
1262 			struct toedev *tod = sc->sc_tod;
1263 
1264 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1265 		}
1266 #endif
1267 		SCH_UNLOCK(sch);
1268 	}
1269 
1270 	/*
1271 	 * Segment validation:
1272 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1273 	 */
1274 	if (th->th_ack != sc->sc_iss + 1) {
1275 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1276 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1277 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1278 		goto failed;
1279 	}
1280 
1281 	/*
1282 	 * The SEQ must fall in the window starting at the received
1283 	 * initial receive sequence number + 1 (the SYN).
1284 	 */
1285 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1286 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1287 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1288 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1289 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1290 		goto failed;
1291 	}
1292 
1293 	*lsop = syncache_socket(sc, *lsop, m);
1294 
1295 	if (__predict_false(*lsop == NULL)) {
1296 		TCPSTAT_INC(tcps_sc_aborted);
1297 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1298 	} else
1299 		TCPSTAT_INC(tcps_sc_completed);
1300 
1301 /* how do we find the inp for the new socket? */
1302 	if (sc != &scs)
1303 		syncache_free(sc);
1304 	return (1);
1305 failed:
1306 	if (sc != NULL) {
1307 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1308 		if (sc != &scs)
1309 			syncache_free(sc);
1310 	}
1311 	if (s != NULL)
1312 		free(s, M_TCPLOG);
1313 	*lsop = NULL;
1314 	return (0);
1315 }
1316 
1317 static struct socket *
syncache_tfo_expand(struct syncache * sc,struct socket * lso,struct mbuf * m,uint64_t response_cookie)1318 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1319     uint64_t response_cookie)
1320 {
1321 	struct inpcb *inp;
1322 	struct tcpcb *tp;
1323 	unsigned int *pending_counter;
1324 	struct socket *so;
1325 
1326 	NET_EPOCH_ASSERT();
1327 
1328 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1329 	so = syncache_socket(sc, lso, m);
1330 	if (so == NULL) {
1331 		TCPSTAT_INC(tcps_sc_aborted);
1332 		atomic_subtract_int(pending_counter, 1);
1333 	} else {
1334 		soisconnected(so);
1335 		inp = sotoinpcb(so);
1336 		tp = intotcpcb(inp);
1337 		tp->t_flags |= TF_FASTOPEN;
1338 		tp->t_tfo_cookie.server = response_cookie;
1339 		tp->snd_max = tp->iss;
1340 		tp->snd_nxt = tp->iss;
1341 		tp->t_tfo_pending = pending_counter;
1342 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1343 		TCPSTAT_INC(tcps_sc_completed);
1344 	}
1345 
1346 	return (so);
1347 }
1348 
1349 /*
1350  * Given a LISTEN socket and an inbound SYN request, add
1351  * this to the syn cache, and send back a segment:
1352  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1353  * to the source.
1354  *
1355  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1356  * Doing so would require that we hold onto the data and deliver it
1357  * to the application.  However, if we are the target of a SYN-flood
1358  * DoS attack, an attacker could send data which would eventually
1359  * consume all available buffer space if it were ACKed.  By not ACKing
1360  * the data, we avoid this DoS scenario.
1361  *
1362  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1363  * cookie is processed and a new socket is created.  In this case, any data
1364  * accompanying the SYN will be queued to the socket by tcp_input() and will
1365  * be ACKed either when the application sends response data or the delayed
1366  * ACK timer expires, whichever comes first.
1367  */
1368 struct socket *
syncache_add(struct in_conninfo * inc,struct tcpopt * to,struct tcphdr * th,struct inpcb * inp,struct socket * so,struct mbuf * m,void * tod,void * todctx,uint8_t iptos,uint16_t port)1369 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1370     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1371     void *todctx, uint8_t iptos, uint16_t port)
1372 {
1373 	struct tcpcb *tp;
1374 	struct socket *rv = NULL;
1375 	struct syncache *sc = NULL;
1376 	struct syncache_head *sch;
1377 	struct mbuf *ipopts = NULL;
1378 	u_int ltflags;
1379 	int win, ip_ttl, ip_tos;
1380 	char *s;
1381 #ifdef INET6
1382 	int autoflowlabel = 0;
1383 #endif
1384 #ifdef MAC
1385 	struct label *maclabel = NULL;
1386 #endif
1387 	struct syncache scs;
1388 	uint64_t tfo_response_cookie;
1389 	unsigned int *tfo_pending = NULL;
1390 	int tfo_cookie_valid = 0;
1391 	int tfo_response_cookie_valid = 0;
1392 	bool locked;
1393 
1394 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1395 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1396 	    ("%s: unexpected tcp flags", __func__));
1397 
1398 	/*
1399 	 * Combine all so/tp operations very early to drop the INP lock as
1400 	 * soon as possible.
1401 	 */
1402 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1403 	tp = sototcpcb(so);
1404 
1405 #ifdef INET6
1406 	if (inc->inc_flags & INC_ISIPV6) {
1407 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1408 			autoflowlabel = 1;
1409 		}
1410 		ip_ttl = in6_selecthlim(inp, NULL);
1411 		if ((inp->in6p_outputopts == NULL) ||
1412 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1413 			ip_tos = 0;
1414 		} else {
1415 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1416 		}
1417 	}
1418 #endif
1419 #if defined(INET6) && defined(INET)
1420 	else
1421 #endif
1422 #ifdef INET
1423 	{
1424 		ip_ttl = inp->inp_ip_ttl;
1425 		ip_tos = inp->inp_ip_tos;
1426 	}
1427 #endif
1428 	win = so->sol_sbrcv_hiwat;
1429 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1430 
1431 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1432 	    (tp->t_tfo_pending != NULL) &&
1433 	    (to->to_flags & TOF_FASTOPEN)) {
1434 		/*
1435 		 * Limit the number of pending TFO connections to
1436 		 * approximately half of the queue limit.  This prevents TFO
1437 		 * SYN floods from starving the service by filling the
1438 		 * listen queue with bogus TFO connections.
1439 		 */
1440 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1441 		    (so->sol_qlimit / 2)) {
1442 			int result;
1443 
1444 			result = tcp_fastopen_check_cookie(inc,
1445 			    to->to_tfo_cookie, to->to_tfo_len,
1446 			    &tfo_response_cookie);
1447 			tfo_cookie_valid = (result > 0);
1448 			tfo_response_cookie_valid = (result >= 0);
1449 		}
1450 
1451 		/*
1452 		 * Remember the TFO pending counter as it will have to be
1453 		 * decremented below if we don't make it to syncache_tfo_expand().
1454 		 */
1455 		tfo_pending = tp->t_tfo_pending;
1456 	}
1457 
1458 #ifdef MAC
1459 	if (mac_syncache_init(&maclabel) != 0) {
1460 		INP_RUNLOCK(inp);
1461 		goto done;
1462 	} else
1463 		mac_syncache_create(maclabel, inp);
1464 #endif
1465 	if (!tfo_cookie_valid)
1466 		INP_RUNLOCK(inp);
1467 
1468 	/*
1469 	 * Remember the IP options, if any.
1470 	 */
1471 #ifdef INET6
1472 	if (!(inc->inc_flags & INC_ISIPV6))
1473 #endif
1474 #ifdef INET
1475 		ipopts = (m) ? ip_srcroute(m) : NULL;
1476 #else
1477 		ipopts = NULL;
1478 #endif
1479 
1480 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1481 	/*
1482 	 * When the socket is TCP-MD5 enabled check that,
1483 	 *  - a signed packet is valid
1484 	 *  - a non-signed packet does not have a security association
1485 	 *
1486 	 *  If a signed packet fails validation or a non-signed packet has a
1487 	 *  security association, the packet will be dropped.
1488 	 */
1489 	if (ltflags & TF_SIGNATURE) {
1490 		if (to->to_flags & TOF_SIGNATURE) {
1491 			if (!TCPMD5_ENABLED() ||
1492 			    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1493 				goto done;
1494 		} else {
1495 			if (TCPMD5_ENABLED() &&
1496 			    TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1497 				goto done;
1498 		}
1499 	} else if (to->to_flags & TOF_SIGNATURE)
1500 		goto done;
1501 #endif	/* TCP_SIGNATURE */
1502 	/*
1503 	 * See if we already have an entry for this connection.
1504 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1505 	 *
1506 	 * XXX: should the syncache be re-initialized with the contents
1507 	 * of the new SYN here (which may have different options?)
1508 	 *
1509 	 * XXX: We do not check the sequence number to see if this is a
1510 	 * real retransmit or a new connection attempt.  The question is
1511 	 * how to handle such a case; either ignore it as spoofed, or
1512 	 * drop the current entry and create a new one?
1513 	 */
1514 	if (syncache_cookiesonly()) {
1515 		sc = NULL;
1516 		sch = syncache_hashbucket(inc);
1517 		locked = false;
1518 	} else {
1519 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1520 		locked = true;
1521 		SCH_LOCK_ASSERT(sch);
1522 	}
1523 	if (sc != NULL) {
1524 		if (tfo_cookie_valid)
1525 			INP_RUNLOCK(inp);
1526 		TCPSTAT_INC(tcps_sc_dupsyn);
1527 		if (ipopts) {
1528 			/*
1529 			 * If we were remembering a previous source route,
1530 			 * forget it and use the new one we've been given.
1531 			 */
1532 			if (sc->sc_ipopts)
1533 				(void)m_free(sc->sc_ipopts);
1534 			sc->sc_ipopts = ipopts;
1535 		}
1536 		/*
1537 		 * Update timestamp if present.
1538 		 */
1539 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1540 			sc->sc_tsreflect = to->to_tsval;
1541 		else
1542 			sc->sc_flags &= ~SCF_TIMESTAMP;
1543 		/*
1544 		 * Adjust ECN response if needed, e.g. different
1545 		 * IP ECN field, or a fallback by the remote host.
1546 		 */
1547 		if (sc->sc_flags & SCF_ECN_MASK) {
1548 			sc->sc_flags &= ~SCF_ECN_MASK;
1549 			sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1550 		}
1551 #ifdef MAC
1552 		/*
1553 		 * Since we have already unconditionally allocated label
1554 		 * storage, free it up.  The syncache entry will already
1555 		 * have an initialized label we can use.
1556 		 */
1557 		mac_syncache_destroy(&maclabel);
1558 #endif
1559 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1560 		/* Retransmit SYN|ACK and reset retransmit count. */
1561 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1562 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1563 			    "resetting timer and retransmitting SYN|ACK\n",
1564 			    s, __func__);
1565 			free(s, M_TCPLOG);
1566 		}
1567 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1568 			sc->sc_rxmits = 0;
1569 			syncache_timeout(sc, sch, 1);
1570 			TCPSTAT_INC(tcps_sndacks);
1571 			TCPSTAT_INC(tcps_sndtotal);
1572 		} else {
1573 			syncache_drop(sc, sch);
1574 			TCPSTAT_INC(tcps_sc_dropped);
1575 		}
1576 		SCH_UNLOCK(sch);
1577 		goto donenoprobe;
1578 	}
1579 
1580 	KASSERT(sc == NULL, ("sc(%p) != NULL", sc));
1581 	/*
1582 	 * Skip allocating a syncache entry if we are just going to discard
1583 	 * it later.
1584 	 */
1585 	if (!locked || tfo_cookie_valid) {
1586 		bzero(&scs, sizeof(scs));
1587 		sc = &scs;
1588 	} else {
1589 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1590 		if (sc == NULL) {
1591 			/*
1592 			 * The zone allocator couldn't provide more entries.
1593 			 * Treat this as if the cache was full; drop the oldest
1594 			 * entry and insert the new one.
1595 			 */
1596 			TCPSTAT_INC(tcps_sc_zonefail);
1597 			sc = TAILQ_LAST(&sch->sch_bucket, sch_head);
1598 			if (sc != NULL) {
1599 				sch->sch_last_overflow = time_uptime;
1600 				syncache_drop(sc, sch);
1601 				syncache_pause(inc);
1602 			}
1603 			sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1604 			if (sc == NULL) {
1605 				if (V_tcp_syncookies) {
1606 					bzero(&scs, sizeof(scs));
1607 					sc = &scs;
1608 				} else {
1609 					KASSERT(locked,
1610 					    ("%s: bucket unexpectedly unlocked",
1611 					    __func__));
1612 					SCH_UNLOCK(sch);
1613 					goto done;
1614 				}
1615 			}
1616 		}
1617 	}
1618 
1619 	KASSERT(sc != NULL, ("sc == NULL"));
1620 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1621 		sc->sc_tfo_cookie = &tfo_response_cookie;
1622 
1623 	/*
1624 	 * Fill in the syncache values.
1625 	 */
1626 #ifdef MAC
1627 	sc->sc_label = maclabel;
1628 #endif
1629 	/*
1630 	 * sc_cred is only used in syncache_pcblist() to list TCP endpoints in
1631 	 * TCPS_SYN_RECEIVED state when V_tcp_syncache.see_other is false.
1632 	 * Therefore, store the credentials and take a reference count only
1633 	 * when needed:
1634 	 * - sc is allocated from the zone and not using the on stack instance.
1635 	 * - the sysctl variable net.inet.tcp.syncache.see_other is false.
1636 	 * The reference count is decremented when a zone allocated sc is
1637 	 * freed in syncache_free().
1638 	 */
1639 	if (sc != &scs && !V_tcp_syncache.see_other)
1640 		sc->sc_cred = crhold(so->so_cred);
1641 	else
1642 		sc->sc_cred = NULL;
1643 	sc->sc_port = port;
1644 	sc->sc_ipopts = ipopts;
1645 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1646 	sc->sc_ip_tos = ip_tos;
1647 	sc->sc_ip_ttl = ip_ttl;
1648 #ifdef TCP_OFFLOAD
1649 	sc->sc_tod = tod;
1650 	sc->sc_todctx = todctx;
1651 #endif
1652 	sc->sc_irs = th->th_seq;
1653 	sc->sc_flags = 0;
1654 	sc->sc_flowlabel = 0;
1655 
1656 	/*
1657 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1658 	 * win was derived from socket earlier in the function.
1659 	 */
1660 	win = imax(win, 0);
1661 	win = imin(win, TCP_MAXWIN);
1662 	sc->sc_wnd = win;
1663 
1664 	if (V_tcp_do_rfc1323 &&
1665 	    !(ltflags & TF_NOOPT)) {
1666 		/*
1667 		 * A timestamp received in a SYN makes
1668 		 * it ok to send timestamp requests and replies.
1669 		 */
1670 		if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1671 			sc->sc_tsreflect = to->to_tsval;
1672 			sc->sc_flags |= SCF_TIMESTAMP;
1673 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1674 		}
1675 		if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1676 			int wscale = 0;
1677 
1678 			/*
1679 			 * Pick the smallest possible scaling factor that
1680 			 * will still allow us to scale up to sb_max, aka
1681 			 * kern.ipc.maxsockbuf.
1682 			 *
1683 			 * We do this because there are broken firewalls that
1684 			 * will corrupt the window scale option, leading to
1685 			 * the other endpoint believing that our advertised
1686 			 * window is unscaled.  At scale factors larger than
1687 			 * 5 the unscaled window will drop below 1500 bytes,
1688 			 * leading to serious problems when traversing these
1689 			 * broken firewalls.
1690 			 *
1691 			 * With the default maxsockbuf of 256K, a scale factor
1692 			 * of 3 will be chosen by this algorithm.  Those who
1693 			 * choose a larger maxsockbuf should watch out
1694 			 * for the compatibility problems mentioned above.
1695 			 *
1696 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1697 			 * or <SYN,ACK>) segment itself is never scaled.
1698 			 */
1699 			while (wscale < TCP_MAX_WINSHIFT &&
1700 			    (TCP_MAXWIN << wscale) < sb_max)
1701 				wscale++;
1702 			sc->sc_requested_r_scale = wscale;
1703 			sc->sc_requested_s_scale = to->to_wscale;
1704 			sc->sc_flags |= SCF_WINSCALE;
1705 		}
1706 	}
1707 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1708 	/*
1709 	 * If incoming packet has an MD5 signature, flag this in the
1710 	 * syncache so that syncache_respond() will do the right thing
1711 	 * with the SYN+ACK.
1712 	 */
1713 	if (to->to_flags & TOF_SIGNATURE)
1714 		sc->sc_flags |= SCF_SIGNATURE;
1715 #endif	/* TCP_SIGNATURE */
1716 	if (to->to_flags & TOF_SACKPERM)
1717 		sc->sc_flags |= SCF_SACK;
1718 	if (to->to_flags & TOF_MSS)
1719 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1720 	if (ltflags & TF_NOOPT)
1721 		sc->sc_flags |= SCF_NOOPT;
1722 	/* ECN Handshake */
1723 	if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0)
1724 		sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1725 
1726 	if (V_tcp_syncookies)
1727 		sc->sc_iss = syncookie_generate(sch, sc);
1728 	else
1729 		sc->sc_iss = arc4random();
1730 #ifdef INET6
1731 	if (autoflowlabel) {
1732 		if (V_tcp_syncookies)
1733 			sc->sc_flowlabel = sc->sc_iss;
1734 		else
1735 			sc->sc_flowlabel = ip6_randomflowlabel();
1736 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1737 	}
1738 #endif
1739 	if (locked)
1740 		SCH_UNLOCK(sch);
1741 
1742 	if (tfo_cookie_valid) {
1743 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1744 		/* INP_RUNLOCK(inp) will be performed by the caller */
1745 		goto tfo_expanded;
1746 	}
1747 
1748 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1749 	/*
1750 	 * Do a standard 3-way handshake.
1751 	 */
1752 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1753 		if (sc != &scs)
1754 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1755 		TCPSTAT_INC(tcps_sndacks);
1756 		TCPSTAT_INC(tcps_sndtotal);
1757 	} else {
1758 		if (sc != &scs)
1759 			syncache_free(sc);
1760 		TCPSTAT_INC(tcps_sc_dropped);
1761 	}
1762 	goto donenoprobe;
1763 
1764 done:
1765 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1766 donenoprobe:
1767 	if (m)
1768 		m_freem(m);
1769 	/*
1770 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1771 	 * result in a new socket was processed and the associated pending
1772 	 * counter has not yet been decremented.  All such TFO processing paths
1773 	 * transit this point.
1774 	 */
1775 	if (tfo_pending != NULL)
1776 		tcp_fastopen_decrement_counter(tfo_pending);
1777 
1778 tfo_expanded:
1779 	if (sc == NULL || sc == &scs) {
1780 #ifdef MAC
1781 		mac_syncache_destroy(&maclabel);
1782 #endif
1783 		if (ipopts)
1784 			(void)m_free(ipopts);
1785 	}
1786 	return (rv);
1787 }
1788 
1789 /*
1790  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1791  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1792  */
1793 static int
syncache_respond(struct syncache * sc,const struct mbuf * m0,int flags)1794 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1795 {
1796 	struct ip *ip = NULL;
1797 	struct mbuf *m;
1798 	struct tcphdr *th = NULL;
1799 	struct udphdr *udp = NULL;
1800 	int optlen, error = 0;	/* Make compiler happy */
1801 	u_int16_t hlen, tlen, mssopt, ulen;
1802 	struct tcpopt to;
1803 #ifdef INET6
1804 	struct ip6_hdr *ip6 = NULL;
1805 #endif
1806 
1807 	NET_EPOCH_ASSERT();
1808 
1809 	hlen =
1810 #ifdef INET6
1811 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1812 #endif
1813 		sizeof(struct ip);
1814 	tlen = hlen + sizeof(struct tcphdr);
1815 	if (sc->sc_port) {
1816 		tlen += sizeof(struct udphdr);
1817 	}
1818 	/* Determine MSS we advertize to other end of connection. */
1819 	mssopt = tcp_mssopt(&sc->sc_inc);
1820 	if (sc->sc_port)
1821 		mssopt -= V_tcp_udp_tunneling_overhead;
1822 	mssopt = max(mssopt, V_tcp_minmss);
1823 
1824 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1825 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1826 	    ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1827 	    "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1828 	    max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1829 
1830 	/* Create the IP+TCP header from scratch. */
1831 	m = m_gethdr(M_NOWAIT, MT_DATA);
1832 	if (m == NULL)
1833 		return (ENOBUFS);
1834 #ifdef MAC
1835 	mac_syncache_create_mbuf(sc->sc_label, m);
1836 #endif
1837 	m->m_data += max_linkhdr;
1838 	m->m_len = tlen;
1839 	m->m_pkthdr.len = tlen;
1840 	m->m_pkthdr.rcvif = NULL;
1841 
1842 #ifdef INET6
1843 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1844 		ip6 = mtod(m, struct ip6_hdr *);
1845 		ip6->ip6_vfc = IPV6_VERSION;
1846 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1847 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1848 		ip6->ip6_plen = htons(tlen - hlen);
1849 		/* ip6_hlim is set after checksum */
1850 		/* Zero out traffic class and flow label. */
1851 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1852 		ip6->ip6_flow |= sc->sc_flowlabel;
1853 		if (sc->sc_port != 0) {
1854 			ip6->ip6_nxt = IPPROTO_UDP;
1855 			udp = (struct udphdr *)(ip6 + 1);
1856 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1857 			udp->uh_dport = sc->sc_port;
1858 			ulen = (tlen - sizeof(struct ip6_hdr));
1859 			th = (struct tcphdr *)(udp + 1);
1860 		} else {
1861 			ip6->ip6_nxt = IPPROTO_TCP;
1862 			th = (struct tcphdr *)(ip6 + 1);
1863 		}
1864 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1865 	}
1866 #endif
1867 #if defined(INET6) && defined(INET)
1868 	else
1869 #endif
1870 #ifdef INET
1871 	{
1872 		ip = mtod(m, struct ip *);
1873 		ip->ip_v = IPVERSION;
1874 		ip->ip_hl = sizeof(struct ip) >> 2;
1875 		ip->ip_len = htons(tlen);
1876 		ip->ip_id = 0;
1877 		ip->ip_off = 0;
1878 		ip->ip_sum = 0;
1879 		ip->ip_src = sc->sc_inc.inc_laddr;
1880 		ip->ip_dst = sc->sc_inc.inc_faddr;
1881 		ip->ip_ttl = sc->sc_ip_ttl;
1882 		ip->ip_tos = sc->sc_ip_tos;
1883 
1884 		/*
1885 		 * See if we should do MTU discovery.  Route lookups are
1886 		 * expensive, so we will only unset the DF bit if:
1887 		 *
1888 		 *	1) path_mtu_discovery is disabled
1889 		 *	2) the SCF_UNREACH flag has been set
1890 		 */
1891 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1892 		       ip->ip_off |= htons(IP_DF);
1893 		if (sc->sc_port == 0) {
1894 			ip->ip_p = IPPROTO_TCP;
1895 			th = (struct tcphdr *)(ip + 1);
1896 		} else {
1897 			ip->ip_p = IPPROTO_UDP;
1898 			udp = (struct udphdr *)(ip + 1);
1899 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1900 			udp->uh_dport = sc->sc_port;
1901 			ulen = (tlen - sizeof(struct ip));
1902 			th = (struct tcphdr *)(udp + 1);
1903 		}
1904 	}
1905 #endif /* INET */
1906 	th->th_sport = sc->sc_inc.inc_lport;
1907 	th->th_dport = sc->sc_inc.inc_fport;
1908 
1909 	if (flags & TH_SYN)
1910 		th->th_seq = htonl(sc->sc_iss);
1911 	else
1912 		th->th_seq = htonl(sc->sc_iss + 1);
1913 	th->th_ack = htonl(sc->sc_irs + 1);
1914 	th->th_off = sizeof(struct tcphdr) >> 2;
1915 	th->th_win = htons(sc->sc_wnd);
1916 	th->th_urp = 0;
1917 
1918 	flags = tcp_ecn_syncache_respond(flags, sc);
1919 	tcp_set_flags(th, flags);
1920 
1921 	/* Tack on the TCP options. */
1922 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1923 		to.to_flags = 0;
1924 
1925 		if (flags & TH_SYN) {
1926 			to.to_mss = mssopt;
1927 			to.to_flags = TOF_MSS;
1928 			if (sc->sc_flags & SCF_WINSCALE) {
1929 				to.to_wscale = sc->sc_requested_r_scale;
1930 				to.to_flags |= TOF_SCALE;
1931 			}
1932 			if (sc->sc_flags & SCF_SACK)
1933 				to.to_flags |= TOF_SACKPERM;
1934 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1935 			if (sc->sc_flags & SCF_SIGNATURE)
1936 				to.to_flags |= TOF_SIGNATURE;
1937 #endif
1938 			if (sc->sc_tfo_cookie) {
1939 				to.to_flags |= TOF_FASTOPEN;
1940 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1941 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1942 				/* don't send cookie again when retransmitting response */
1943 				sc->sc_tfo_cookie = NULL;
1944 			}
1945 		}
1946 		if (sc->sc_flags & SCF_TIMESTAMP) {
1947 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1948 			to.to_tsecr = sc->sc_tsreflect;
1949 			to.to_flags |= TOF_TS;
1950 		}
1951 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1952 
1953 		/* Adjust headers by option size. */
1954 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1955 		m->m_len += optlen;
1956 		m->m_pkthdr.len += optlen;
1957 #ifdef INET6
1958 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1959 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1960 		else
1961 #endif
1962 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1963 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1964 		if (sc->sc_flags & SCF_SIGNATURE) {
1965 			KASSERT(to.to_flags & TOF_SIGNATURE,
1966 			    ("tcp_addoptions() didn't set tcp_signature"));
1967 
1968 			/* NOTE: to.to_signature is inside of mbuf */
1969 			if (!TCPMD5_ENABLED() ||
1970 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1971 				m_freem(m);
1972 				return (EACCES);
1973 			}
1974 		}
1975 #endif
1976 	} else
1977 		optlen = 0;
1978 
1979 	if (udp) {
1980 		ulen += optlen;
1981 		udp->uh_ulen = htons(ulen);
1982 	}
1983 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1984 	/*
1985 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1986 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1987 	 * to SYN|ACK due to lack of inp here.
1988 	 */
1989 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1990 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1991 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1992 	}
1993 #ifdef INET6
1994 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1995 		if (sc->sc_port) {
1996 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1997 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1998 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
1999 			      IPPROTO_UDP, 0);
2000 			th->th_sum = htons(0);
2001 		} else {
2002 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2003 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2004 			th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2005 			    IPPROTO_TCP, 0);
2006 		}
2007 		ip6->ip6_hlim = sc->sc_ip_ttl;
2008 #ifdef TCP_OFFLOAD
2009 		if (ADDED_BY_TOE(sc)) {
2010 			struct toedev *tod = sc->sc_tod;
2011 
2012 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2013 
2014 			return (error);
2015 		}
2016 #endif
2017 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2018 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2019 	}
2020 #endif
2021 #if defined(INET6) && defined(INET)
2022 	else
2023 #endif
2024 #ifdef INET
2025 	{
2026 		if (sc->sc_port) {
2027 			m->m_pkthdr.csum_flags = CSUM_UDP;
2028 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2029 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2030 			      ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2031 			th->th_sum = htons(0);
2032 		} else {
2033 			m->m_pkthdr.csum_flags = CSUM_TCP;
2034 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2035 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2036 			    htons(tlen + optlen - hlen + IPPROTO_TCP));
2037 		}
2038 #ifdef TCP_OFFLOAD
2039 		if (ADDED_BY_TOE(sc)) {
2040 			struct toedev *tod = sc->sc_tod;
2041 
2042 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2043 
2044 			return (error);
2045 		}
2046 #endif
2047 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2048 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2049 	}
2050 #endif
2051 	return (error);
2052 }
2053 
2054 /*
2055  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2056  * that exceed the capacity of the syncache by avoiding the storage of any
2057  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2058  * attacks where the attacker does not have access to our responses.
2059  *
2060  * Syncookies encode and include all necessary information about the
2061  * connection setup within the SYN|ACK that we send back.  That way we
2062  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2063  * (if ever).  Normally the syncache and syncookies are running in parallel
2064  * with the latter taking over when the former is exhausted.  When matching
2065  * syncache entry is found the syncookie is ignored.
2066  *
2067  * The only reliable information persisting the 3WHS is our initial sequence
2068  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2069  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2070  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2071  * returns and signifies a legitimate connection if it matches the ACK.
2072  *
2073  * The available space of 32 bits to store the hash and to encode the SYN
2074  * option information is very tight and we should have at least 24 bits for
2075  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2076  *
2077  * SYN option information we have to encode to fully restore a connection:
2078  * MSS: is imporant to chose an optimal segment size to avoid IP level
2079  *   fragmentation along the path.  The common MSS values can be encoded
2080  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2081  *   in the table leading to a slight increase in packetization overhead.
2082  * WSCALE: is necessary to allow large windows to be used for high delay-
2083  *   bandwidth product links.  Not scaling the window when it was initially
2084  *   negotiated is bad for performance as lack of scaling further decreases
2085  *   the apparent available send window.  We only need to encode the WSCALE
2086  *   we received from the remote end.  Our end can be recalculated at any
2087  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2088  *   Uncommon values are captured by the next lower value in the table
2089  *   making us under-estimate the available window size halving our
2090  *   theoretically possible maximum throughput for that connection.
2091  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2092  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2093  *   that are included in all segments on a connection.  We enable them when
2094  *   the ACK has them.
2095  *
2096  * Security of syncookies and attack vectors:
2097  *
2098  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2099  * together with the gloabl secret to make it unique per connection attempt.
2100  * Thus any change of any of those parameters results in a different MAC output
2101  * in an unpredictable way unless a collision is encountered.  24 bits of the
2102  * MAC are embedded into the ISS.
2103  *
2104  * To prevent replay attacks two rotating global secrets are updated with a
2105  * new random value every 15 seconds.  The life-time of a syncookie is thus
2106  * 15-30 seconds.
2107  *
2108  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2109  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2110  * text attack by varying and testing the all parameters under his control.
2111  * The strength depends on the size and randomness of the secret, and the
2112  * cryptographic security of the MAC function.  Due to the constant updating
2113  * of the secret the attacker has at most 29.999 seconds to find the secret
2114  * and launch spoofed connections.  After that he has to start all over again.
2115  *
2116  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2117  * size an average of 4,823 attempts are required for a 50% chance of success
2118  * to spoof a single syncookie (birthday collision paradox).  However the
2119  * attacker is blind and doesn't know if one of his attempts succeeded unless
2120  * he has a side channel to interfere success from.  A single connection setup
2121  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2122  * This many attempts are required for each one blind spoofed connection.  For
2123  * every additional spoofed connection he has to launch another N attempts.
2124  * Thus for a sustained rate 100 spoofed connections per second approximately
2125  * 1,800,000 packets per second would have to be sent.
2126  *
2127  * NB: The MAC function should be fast so that it doesn't become a CPU
2128  * exhaustion attack vector itself.
2129  *
2130  * References:
2131  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2132  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2133  *   http://cr.yp.to/syncookies.html    (overview)
2134  *   http://cr.yp.to/syncookies/archive (details)
2135  *
2136  *
2137  * Schematic construction of a syncookie enabled Initial Sequence Number:
2138  *  0        1         2         3
2139  *  12345678901234567890123456789012
2140  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2141  *
2142  *  x 24 MAC (truncated)
2143  *  W  3 Send Window Scale index
2144  *  M  3 MSS index
2145  *  S  1 SACK permitted
2146  *  P  1 Odd/even secret
2147  */
2148 
2149 /*
2150  * Distribution and probability of certain MSS values.  Those in between are
2151  * rounded down to the next lower one.
2152  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2153  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2154  */
2155 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2156 
2157 /*
2158  * Distribution and probability of certain WSCALE values.  We have to map the
2159  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2160  * bits based on prevalence of certain values.  Where we don't have an exact
2161  * match for are rounded down to the next lower one letting us under-estimate
2162  * the true available window.  At the moment this would happen only for the
2163  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2164  * and window size).  The absence of the WSCALE option (no scaling in either
2165  * direction) is encoded with index zero.
2166  * [WSCALE values histograms, Allman, 2012]
2167  *                            X 10 10 35  5  6 14 10%   by host
2168  *                            X 11  4  5  5 18 49  3%   by connections
2169  */
2170 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2171 
2172 /*
2173  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2174  * and good cryptographic properties.
2175  */
2176 static uint32_t
syncookie_mac(struct in_conninfo * inc,tcp_seq irs,uint8_t flags,uint8_t * secbits,uintptr_t secmod)2177 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2178     uint8_t *secbits, uintptr_t secmod)
2179 {
2180 	SIPHASH_CTX ctx;
2181 	uint32_t siphash[2];
2182 
2183 	SipHash24_Init(&ctx);
2184 	SipHash_SetKey(&ctx, secbits);
2185 	switch (inc->inc_flags & INC_ISIPV6) {
2186 #ifdef INET
2187 	case 0:
2188 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2189 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2190 		break;
2191 #endif
2192 #ifdef INET6
2193 	case INC_ISIPV6:
2194 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2195 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2196 		break;
2197 #endif
2198 	}
2199 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2200 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2201 	SipHash_Update(&ctx, &irs, sizeof(irs));
2202 	SipHash_Update(&ctx, &flags, sizeof(flags));
2203 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2204 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2205 
2206 	return (siphash[0] ^ siphash[1]);
2207 }
2208 
2209 static tcp_seq
syncookie_generate(struct syncache_head * sch,struct syncache * sc)2210 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2211 {
2212 	u_int i, secbit, wscale;
2213 	uint32_t iss, hash;
2214 	uint8_t *secbits;
2215 	union syncookie cookie;
2216 
2217 	cookie.cookie = 0;
2218 
2219 	/* Map our computed MSS into the 3-bit index. */
2220 	for (i = nitems(tcp_sc_msstab) - 1;
2221 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2222 	     i--)
2223 		;
2224 	cookie.flags.mss_idx = i;
2225 
2226 	/*
2227 	 * Map the send window scale into the 3-bit index but only if
2228 	 * the wscale option was received.
2229 	 */
2230 	if (sc->sc_flags & SCF_WINSCALE) {
2231 		wscale = sc->sc_requested_s_scale;
2232 		for (i = nitems(tcp_sc_wstab) - 1;
2233 		    tcp_sc_wstab[i] > wscale && i > 0;
2234 		     i--)
2235 			;
2236 		cookie.flags.wscale_idx = i;
2237 	}
2238 
2239 	/* Can we do SACK? */
2240 	if (sc->sc_flags & SCF_SACK)
2241 		cookie.flags.sack_ok = 1;
2242 
2243 	/* Which of the two secrets to use. */
2244 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2245 	cookie.flags.odd_even = secbit;
2246 
2247 	secbits = V_tcp_syncache.secret.key[secbit];
2248 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2249 	    (uintptr_t)sch);
2250 
2251 	/*
2252 	 * Put the flags into the hash and XOR them to get better ISS number
2253 	 * variance.  This doesn't enhance the cryptographic strength and is
2254 	 * done to prevent the 8 cookie bits from showing up directly on the
2255 	 * wire.
2256 	 */
2257 	iss = hash & ~0xff;
2258 	iss |= cookie.cookie ^ (hash >> 24);
2259 
2260 	TCPSTAT_INC(tcps_sc_sendcookie);
2261 	return (iss);
2262 }
2263 
2264 static struct syncache *
syncookie_lookup(struct in_conninfo * inc,struct syncache_head * sch,struct syncache * sc,struct tcphdr * th,struct tcpopt * to,struct socket * lso,uint16_t port)2265 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2266     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2267     struct socket *lso, uint16_t port)
2268 {
2269 	uint32_t hash;
2270 	uint8_t *secbits;
2271 	tcp_seq ack, seq;
2272 	int wnd, wscale = 0;
2273 	union syncookie cookie;
2274 
2275 	/*
2276 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2277 	 * advances.
2278 	 */
2279 	ack = th->th_ack - 1;
2280 	seq = th->th_seq - 1;
2281 
2282 	/*
2283 	 * Unpack the flags containing enough information to restore the
2284 	 * connection.
2285 	 */
2286 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2287 
2288 	/* Which of the two secrets to use. */
2289 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2290 
2291 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2292 
2293 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2294 	if ((ack & ~0xff) != (hash & ~0xff))
2295 		return (NULL);
2296 
2297 	/* Fill in the syncache values. */
2298 	sc->sc_flags = 0;
2299 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2300 	sc->sc_ipopts = NULL;
2301 
2302 	sc->sc_irs = seq;
2303 	sc->sc_iss = ack;
2304 
2305 	switch (inc->inc_flags & INC_ISIPV6) {
2306 #ifdef INET
2307 	case 0:
2308 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2309 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2310 		break;
2311 #endif
2312 #ifdef INET6
2313 	case INC_ISIPV6:
2314 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2315 			sc->sc_flowlabel =
2316 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2317 		break;
2318 #endif
2319 	}
2320 
2321 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2322 
2323 	/* We can simply recompute receive window scale we sent earlier. */
2324 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2325 		wscale++;
2326 
2327 	/* Only use wscale if it was enabled in the orignal SYN. */
2328 	if (cookie.flags.wscale_idx > 0) {
2329 		sc->sc_requested_r_scale = wscale;
2330 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2331 		sc->sc_flags |= SCF_WINSCALE;
2332 	}
2333 
2334 	wnd = lso->sol_sbrcv_hiwat;
2335 	wnd = imax(wnd, 0);
2336 	wnd = imin(wnd, TCP_MAXWIN);
2337 	sc->sc_wnd = wnd;
2338 
2339 	if (cookie.flags.sack_ok)
2340 		sc->sc_flags |= SCF_SACK;
2341 
2342 	if (to->to_flags & TOF_TS) {
2343 		sc->sc_flags |= SCF_TIMESTAMP;
2344 		sc->sc_tsreflect = to->to_tsval;
2345 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2346 	}
2347 
2348 	if (to->to_flags & TOF_SIGNATURE)
2349 		sc->sc_flags |= SCF_SIGNATURE;
2350 
2351 	sc->sc_rxmits = 0;
2352 
2353 	sc->sc_port = port;
2354 
2355 	TCPSTAT_INC(tcps_sc_recvcookie);
2356 	return (sc);
2357 }
2358 
2359 #ifdef INVARIANTS
2360 static int
syncookie_cmp(struct in_conninfo * inc,struct syncache_head * sch,struct syncache * sc,struct tcphdr * th,struct tcpopt * to,struct socket * lso,uint16_t port)2361 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2362     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2363     struct socket *lso, uint16_t port)
2364 {
2365 	struct syncache scs, *scx;
2366 	char *s;
2367 
2368 	bzero(&scs, sizeof(scs));
2369 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2370 
2371 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2372 		return (0);
2373 
2374 	if (scx != NULL) {
2375 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2376 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2377 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2378 
2379 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2380 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2381 			    s, __func__, sc->sc_requested_r_scale,
2382 			    scx->sc_requested_r_scale);
2383 
2384 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2385 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2386 			    s, __func__, sc->sc_requested_s_scale,
2387 			    scx->sc_requested_s_scale);
2388 
2389 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2390 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2391 	}
2392 
2393 	if (s != NULL)
2394 		free(s, M_TCPLOG);
2395 	return (0);
2396 }
2397 #endif /* INVARIANTS */
2398 
2399 static void
syncookie_reseed(void * arg)2400 syncookie_reseed(void *arg)
2401 {
2402 	struct tcp_syncache *sc = arg;
2403 	uint8_t *secbits;
2404 	int secbit;
2405 
2406 	/*
2407 	 * Reseeding the secret doesn't have to be protected by a lock.
2408 	 * It only must be ensured that the new random values are visible
2409 	 * to all CPUs in a SMP environment.  The atomic with release
2410 	 * semantics ensures that.
2411 	 */
2412 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2413 	secbits = sc->secret.key[secbit];
2414 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2415 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2416 
2417 	/* Reschedule ourself. */
2418 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2419 }
2420 
2421 /*
2422  * We have overflowed a bucket. Let's pause dealing with the syncache.
2423  * This function will increment the bucketoverflow statistics appropriately
2424  * (once per pause when pausing is enabled; otherwise, once per overflow).
2425  */
2426 static void
syncache_pause(struct in_conninfo * inc)2427 syncache_pause(struct in_conninfo *inc)
2428 {
2429 	time_t delta;
2430 	const char *s;
2431 
2432 	/* XXX:
2433 	 * 2. Add sysctl read here so we don't get the benefit of this
2434 	 * change without the new sysctl.
2435 	 */
2436 
2437 	/*
2438 	 * Try an unlocked read. If we already know that another thread
2439 	 * has activated the feature, there is no need to proceed.
2440 	 */
2441 	if (V_tcp_syncache.paused)
2442 		return;
2443 
2444 	/* Are cookied enabled? If not, we can't pause. */
2445 	if (!V_tcp_syncookies) {
2446 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2447 		return;
2448 	}
2449 
2450 	/*
2451 	 * We may be the first thread to find an overflow. Get the lock
2452 	 * and evaluate if we need to take action.
2453 	 */
2454 	mtx_lock(&V_tcp_syncache.pause_mtx);
2455 	if (V_tcp_syncache.paused) {
2456 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2457 		return;
2458 	}
2459 
2460 	/* Activate protection. */
2461 	V_tcp_syncache.paused = true;
2462 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2463 
2464 	/*
2465 	 * Determine the last backoff time. If we are seeing a re-newed
2466 	 * attack within that same time after last reactivating the syncache,
2467 	 * consider it an extension of the same attack.
2468 	 */
2469 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2470 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2471 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2472 			delta <<= 1;
2473 			V_tcp_syncache.pause_backoff++;
2474 		}
2475 	} else {
2476 		delta = TCP_SYNCACHE_PAUSE_TIME;
2477 		V_tcp_syncache.pause_backoff = 0;
2478 	}
2479 
2480 	/* Log a warning, including IP addresses, if able. */
2481 	if (inc != NULL)
2482 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2483 	else
2484 		s = (const char *)NULL;
2485 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2486 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2487 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2488 	    (s != NULL) ? ")" : "");
2489 	free(__DECONST(void *, s), M_TCPLOG);
2490 
2491 	/* Use the calculated delta to set a new pause time. */
2492 	V_tcp_syncache.pause_until = time_uptime + delta;
2493 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2494 	    &V_tcp_syncache);
2495 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2496 }
2497 
2498 /* Evaluate whether we need to unpause. */
2499 static void
syncache_unpause(void * arg)2500 syncache_unpause(void *arg)
2501 {
2502 	struct tcp_syncache *sc;
2503 	time_t delta;
2504 
2505 	sc = arg;
2506 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2507 	callout_deactivate(&sc->pause_co);
2508 
2509 	/*
2510 	 * Check to make sure we are not running early. If the pause
2511 	 * time has expired, then deactivate the protection.
2512 	 */
2513 	if ((delta = sc->pause_until - time_uptime) > 0)
2514 		callout_schedule(&sc->pause_co, delta * hz);
2515 	else
2516 		sc->paused = false;
2517 }
2518 
2519 /*
2520  * Exports the syncache entries to userland so that netstat can display
2521  * them alongside the other sockets.  This function is intended to be
2522  * called only from tcp_pcblist.
2523  *
2524  * Due to concurrency on an active system, the number of pcbs exported
2525  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2526  * amount of space the caller allocated for this function to use.
2527  */
2528 int
syncache_pcblist(struct sysctl_req * req)2529 syncache_pcblist(struct sysctl_req *req)
2530 {
2531 	struct xtcpcb xt;
2532 	struct syncache *sc;
2533 	struct syncache_head *sch;
2534 	int error, i;
2535 
2536 	bzero(&xt, sizeof(xt));
2537 	xt.xt_len = sizeof(xt);
2538 	xt.t_state = TCPS_SYN_RECEIVED;
2539 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2540 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2541 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2542 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2543 
2544 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2545 		sch = &V_tcp_syncache.hashbase[i];
2546 		SCH_LOCK(sch);
2547 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2548 			if (sc->sc_cred != NULL &&
2549 			    cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2550 				continue;
2551 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2552 				xt.xt_inp.inp_vflag = INP_IPV6;
2553 			else
2554 				xt.xt_inp.inp_vflag = INP_IPV4;
2555 			xt.xt_encaps_port = sc->sc_port;
2556 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2557 			    sizeof (struct in_conninfo));
2558 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2559 			if (error) {
2560 				SCH_UNLOCK(sch);
2561 				return (0);
2562 			}
2563 		}
2564 		SCH_UNLOCK(sch);
2565 	}
2566 
2567 	return (0);
2568 }
2569