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(bool, tcp_syncookies) = true;
106 #define V_tcp_syncookies VNET(tcp_syncookies)
107 SYSCTL_BOOL(_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(bool, tcp_syncookiesonly) = false;
112 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
113 SYSCTL_BOOL(_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 #ifdef TCP_OFFLOAD
118 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
119 #endif
120
121 static void syncache_drop(struct syncache *, struct syncache_head *);
122 static void syncache_free(struct syncache *);
123 static void syncache_insert(struct syncache *, struct syncache_head *);
124 static int syncache_respond(struct syncache *, const struct mbuf *, int);
125 static void syncache_send_challenge_ack(struct syncache *, struct mbuf *);
126 static struct socket *syncache_socket(struct syncache *, struct socket *,
127 struct mbuf *m);
128 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
129 int docallout);
130 static void syncache_timer(void *);
131
132 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
133 uint8_t *, uintptr_t);
134 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
135 static bool syncookie_expand(struct in_conninfo *,
136 const struct syncache_head *, struct syncache *,
137 struct tcphdr *, struct tcpopt *, struct socket *,
138 uint16_t);
139 static void syncache_pause(struct in_conninfo *);
140 static void syncache_unpause(void *);
141 static void syncookie_reseed(void *);
142 #ifdef INVARIANTS
143 static void syncookie_cmp(struct in_conninfo *,
144 const struct syncache_head *, struct syncache *,
145 struct tcphdr *, struct tcpopt *, struct socket *,
146 uint16_t);
147 #endif
148
149 /*
150 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
151 * 3 retransmits corresponds to a timeout with default values of
152 * tcp_rexmit_initial * ( 1 +
153 * tcp_backoff[1] +
154 * tcp_backoff[2] +
155 * tcp_backoff[3]) + 3 * tcp_rexmit_slop,
156 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms,
157 * the odds are that the user has given up attempting to connect by then.
158 */
159 #define SYNCACHE_MAXREXMTS 3
160
161 /* Arbitrary values */
162 #define TCP_SYNCACHE_HASHSIZE 512
163 #define TCP_SYNCACHE_BUCKETLIMIT 30
164
165 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
166 #define V_tcp_syncache VNET(tcp_syncache)
167
168 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
169 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
170 "TCP SYN cache");
171
172 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
173 &VNET_NAME(tcp_syncache.bucket_limit), 0,
174 "Per-bucket hash limit for syncache");
175
176 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
177 &VNET_NAME(tcp_syncache.cache_limit), 0,
178 "Overall entry limit for syncache");
179
180 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
181 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
182
183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
184 &VNET_NAME(tcp_syncache.hashsize), 0,
185 "Size of TCP syncache hashtable");
186
187 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
188 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
189 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
190 "and mac(4) checks");
191
192 static int
sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)193 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
194 {
195 int error;
196 u_int new;
197
198 new = V_tcp_syncache.rexmt_limit;
199 error = sysctl_handle_int(oidp, &new, 0, req);
200 if ((error == 0) && (req->newptr != NULL)) {
201 if (new > TCP_MAXRXTSHIFT)
202 error = EINVAL;
203 else
204 V_tcp_syncache.rexmt_limit = new;
205 }
206 return (error);
207 }
208
209 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
210 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
211 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
212 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "IU",
213 "Limit on SYN/ACK retransmissions");
214
215 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
216 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
217 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
218 "Send reset on socket allocation failure");
219
220 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
221
222 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
223 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
224 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
225
226 /*
227 * Requires the syncache entry to be already removed from the bucket list.
228 */
229 static void
syncache_free(struct syncache * sc)230 syncache_free(struct syncache *sc)
231 {
232
233 if (sc->sc_ipopts)
234 (void)m_free(sc->sc_ipopts);
235 if (sc->sc_cred)
236 crfree(sc->sc_cred);
237 #ifdef MAC
238 mac_syncache_destroy(&sc->sc_label);
239 #endif
240
241 uma_zfree(V_tcp_syncache.zone, sc);
242 }
243
244 void
syncache_init(void)245 syncache_init(void)
246 {
247 int i;
248
249 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
250 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
251 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
252 V_tcp_syncache.hash_secret = arc4random();
253
254 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
255 &V_tcp_syncache.hashsize);
256 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
257 &V_tcp_syncache.bucket_limit);
258 if (!powerof2(V_tcp_syncache.hashsize) ||
259 V_tcp_syncache.hashsize == 0) {
260 printf("WARNING: syncache hash size is not a power of 2.\n");
261 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
262 }
263 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
264
265 /* Set limits. */
266 V_tcp_syncache.cache_limit =
267 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
268 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
269 &V_tcp_syncache.cache_limit);
270
271 /* Allocate the hash table. */
272 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
273 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
274
275 #ifdef VIMAGE
276 V_tcp_syncache.vnet = curvnet;
277 #endif
278
279 /* Initialize the hash buckets. */
280 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
281 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
282 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
283 NULL, MTX_DEF);
284 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
285 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
286 V_tcp_syncache.hashbase[i].sch_length = 0;
287 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
288 V_tcp_syncache.hashbase[i].sch_last_overflow =
289 -(SYNCOOKIE_LIFETIME + 1);
290 }
291
292 /* Create the syncache entry zone. */
293 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
294 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
295 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
296 V_tcp_syncache.cache_limit);
297
298 /* Start the SYN cookie reseeder callout. */
299 callout_init(&V_tcp_syncache.secret.reseed, 1);
300 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
301 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
302 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
303 syncookie_reseed, &V_tcp_syncache);
304
305 /* Initialize the pause machinery. */
306 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
307 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
308 0);
309 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
310 V_tcp_syncache.pause_backoff = 0;
311 V_tcp_syncache.paused = false;
312 }
313
314 #ifdef VIMAGE
315 void
syncache_destroy(void)316 syncache_destroy(void)
317 {
318 struct syncache_head *sch;
319 struct syncache *sc, *nsc;
320 int i;
321
322 /*
323 * Stop the re-seed timer before freeing resources. No need to
324 * possibly schedule it another time.
325 */
326 callout_drain(&V_tcp_syncache.secret.reseed);
327
328 /* Stop the SYN cache pause callout. */
329 mtx_lock(&V_tcp_syncache.pause_mtx);
330 if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
331 mtx_unlock(&V_tcp_syncache.pause_mtx);
332 callout_drain(&V_tcp_syncache.pause_co);
333 } else
334 mtx_unlock(&V_tcp_syncache.pause_mtx);
335
336 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
337 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
338 sch = &V_tcp_syncache.hashbase[i];
339 callout_drain(&sch->sch_timer);
340
341 SCH_LOCK(sch);
342 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
343 syncache_drop(sc, sch);
344 SCH_UNLOCK(sch);
345 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
346 ("%s: sch->sch_bucket not empty", __func__));
347 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
348 __func__, sch->sch_length));
349 mtx_destroy(&sch->sch_mtx);
350 }
351
352 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
353 ("%s: cache_count not 0", __func__));
354
355 /* Free the allocated global resources. */
356 uma_zdestroy(V_tcp_syncache.zone);
357 free(V_tcp_syncache.hashbase, M_SYNCACHE);
358 mtx_destroy(&V_tcp_syncache.pause_mtx);
359 }
360 #endif
361
362 /*
363 * Inserts a syncache entry into the specified bucket row.
364 * Locks and unlocks the syncache_head autonomously.
365 */
366 static void
syncache_insert(struct syncache * sc,struct syncache_head * sch)367 syncache_insert(struct syncache *sc, struct syncache_head *sch)
368 {
369 struct syncache *sc2;
370
371 SCH_LOCK(sch);
372
373 /*
374 * Make sure that we don't overflow the per-bucket limit.
375 * If the bucket is full, toss the oldest element.
376 */
377 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
378 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
379 ("sch->sch_length incorrect"));
380 syncache_pause(&sc->sc_inc);
381 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
382 sch->sch_last_overflow = time_uptime;
383 syncache_drop(sc2, sch);
384 }
385
386 /* Put it into the bucket. */
387 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
388 sch->sch_length++;
389
390 #ifdef TCP_OFFLOAD
391 if (ADDED_BY_TOE(sc)) {
392 struct toedev *tod = sc->sc_tod;
393
394 tod->tod_syncache_added(tod, sc->sc_todctx);
395 }
396 #endif
397
398 /* Reinitialize the bucket row's timer. */
399 if (sch->sch_length == 1)
400 sch->sch_nextc = ticks + INT_MAX;
401 syncache_timeout(sc, sch, 1);
402
403 SCH_UNLOCK(sch);
404
405 TCPSTATES_INC(TCPS_SYN_RECEIVED);
406 TCPSTAT_INC(tcps_sc_added);
407 }
408
409 /*
410 * Remove and free entry from syncache bucket row.
411 * Expects locked syncache head.
412 */
413 static void
syncache_drop(struct syncache * sc,struct syncache_head * sch)414 syncache_drop(struct syncache *sc, struct syncache_head *sch)
415 {
416
417 SCH_LOCK_ASSERT(sch);
418
419 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
420 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
421 sch->sch_length--;
422
423 #ifdef TCP_OFFLOAD
424 if (ADDED_BY_TOE(sc)) {
425 struct toedev *tod = sc->sc_tod;
426
427 tod->tod_syncache_removed(tod, sc->sc_todctx);
428 }
429 #endif
430
431 syncache_free(sc);
432 }
433
434 /*
435 * Engage/reengage time on bucket row.
436 */
437 static void
syncache_timeout(struct syncache * sc,struct syncache_head * sch,int docallout)438 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
439 {
440 int rexmt;
441
442 if (sc->sc_rxmits == 0)
443 rexmt = tcp_rexmit_initial;
444 else
445 TCPT_RANGESET(rexmt,
446 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
447 tcp_rexmit_min, tcp_rexmit_max);
448 sc->sc_rxttime = ticks + rexmt;
449 sc->sc_rxmits++;
450 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
451 sch->sch_nextc = sc->sc_rxttime;
452 if (docallout)
453 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
454 syncache_timer, (void *)sch);
455 }
456 }
457
458 /*
459 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
460 * If we have retransmitted an entry the maximum number of times, expire it.
461 * One separate timer for each bucket row.
462 */
463 static void
syncache_timer(void * xsch)464 syncache_timer(void *xsch)
465 {
466 struct syncache_head *sch = (struct syncache_head *)xsch;
467 struct syncache *sc, *nsc;
468 struct epoch_tracker et;
469 int tick = ticks;
470 char *s;
471 bool paused;
472
473 CURVNET_SET(sch->sch_sc->vnet);
474
475 /* NB: syncache_head has already been locked by the callout. */
476 SCH_LOCK_ASSERT(sch);
477
478 /*
479 * In the following cycle we may remove some entries and/or
480 * advance some timeouts, so re-initialize the bucket timer.
481 */
482 sch->sch_nextc = tick + INT_MAX;
483
484 /*
485 * If we have paused processing, unconditionally remove
486 * all syncache entries.
487 */
488 mtx_lock(&V_tcp_syncache.pause_mtx);
489 paused = V_tcp_syncache.paused;
490 mtx_unlock(&V_tcp_syncache.pause_mtx);
491
492 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
493 if (paused) {
494 syncache_drop(sc, sch);
495 continue;
496 }
497 /*
498 * We do not check if the listen socket still exists
499 * and accept the case where the listen socket may be
500 * gone by the time we resend the SYN/ACK. We do
501 * not expect this to happens often. If it does,
502 * then the RST will be sent by the time the remote
503 * host does the SYN/ACK->ACK.
504 */
505 if (TSTMP_GT(sc->sc_rxttime, tick)) {
506 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
507 sch->sch_nextc = sc->sc_rxttime;
508 continue;
509 }
510 if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
511 sc->sc_flags &= ~SCF_ECN_MASK;
512 }
513 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
514 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
515 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
516 "giving up and removing syncache entry\n",
517 s, __func__);
518 free(s, M_TCPLOG);
519 }
520 syncache_drop(sc, sch);
521 TCPSTAT_INC(tcps_sc_stale);
522 continue;
523 }
524 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
525 log(LOG_DEBUG, "%s; %s: Response timeout, "
526 "retransmitting (%u) SYN|ACK\n",
527 s, __func__, sc->sc_rxmits);
528 free(s, M_TCPLOG);
529 }
530
531 NET_EPOCH_ENTER(et);
532 if (syncache_respond(sc, NULL, TH_SYN|TH_ACK) == 0) {
533 syncache_timeout(sc, sch, 0);
534 TCPSTAT_INC(tcps_sndacks);
535 TCPSTAT_INC(tcps_sndtotal);
536 TCPSTAT_INC(tcps_sc_retransmitted);
537 } else {
538 /*
539 * Most likely we are memory constrained, so free
540 * resources.
541 */
542 syncache_drop(sc, sch);
543 TCPSTAT_INC(tcps_sc_dropped);
544 }
545 NET_EPOCH_EXIT(et);
546 }
547 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
548 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
549 syncache_timer, (void *)(sch));
550 CURVNET_RESTORE();
551 }
552
553 /*
554 * Returns true if the system is only using cookies at the moment.
555 * This could be due to a sysadmin decision to only use cookies, or it
556 * could be due to the system detecting an attack.
557 */
558 static inline bool
syncache_cookiesonly(void)559 syncache_cookiesonly(void)
560 {
561 return ((V_tcp_syncookies && V_tcp_syncache.paused) ||
562 V_tcp_syncookiesonly);
563 }
564
565 /*
566 * Find the hash bucket for the given connection.
567 */
568 static struct syncache_head *
syncache_hashbucket(struct in_conninfo * inc)569 syncache_hashbucket(struct in_conninfo *inc)
570 {
571 uint32_t hash;
572
573 /*
574 * The hash is built on foreign port + local port + foreign address.
575 * We rely on the fact that struct in_conninfo starts with 16 bits
576 * of foreign port, then 16 bits of local port then followed by 128
577 * bits of foreign address. In case of IPv4 address, the first 3
578 * 32-bit words of the address always are zeroes.
579 */
580 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
581 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
582
583 return (&V_tcp_syncache.hashbase[hash]);
584 }
585
586 /*
587 * Find an entry in the syncache.
588 * Returns always with locked syncache_head plus a matching entry or NULL.
589 */
590 static struct syncache *
syncache_lookup(struct in_conninfo * inc,struct syncache_head ** schp)591 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
592 {
593 struct syncache *sc;
594 struct syncache_head *sch;
595
596 *schp = sch = syncache_hashbucket(inc);
597 SCH_LOCK(sch);
598
599 /* Circle through bucket row to find matching entry. */
600 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
601 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
602 sizeof(struct in_endpoints)) == 0)
603 break;
604
605 return (sc); /* Always returns with locked sch. */
606 }
607
608 /*
609 * This function is called when we get a RST for a
610 * non-existent connection, so that we can see if the
611 * connection is in the syn cache. If it is, zap it.
612 * If required send a challenge ACK.
613 */
614 void
syncache_chkrst(struct in_conninfo * inc,struct tcphdr * th,struct mbuf * m,uint16_t port)615 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
616 uint16_t port)
617 {
618 struct syncache *sc;
619 struct syncache_head *sch;
620 char *s = NULL;
621
622 if (syncache_cookiesonly())
623 return;
624 sc = syncache_lookup(inc, &sch); /* returns locked sch */
625 SCH_LOCK_ASSERT(sch);
626
627 /*
628 * No corresponding connection was found in syncache.
629 * If syncookies are enabled and possibly exclusively
630 * used, or we are under memory pressure, a valid RST
631 * may not find a syncache entry. In that case we're
632 * done and no SYN|ACK retransmissions will happen.
633 * Otherwise the RST was misdirected or spoofed.
634 */
635 if (sc == NULL) {
636 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
637 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
638 "syncache entry (possibly syncookie only), "
639 "segment ignored\n", s, __func__);
640 TCPSTAT_INC(tcps_badrst);
641 goto done;
642 }
643
644 /* The remote UDP encaps port does not match. */
645 if (sc->sc_port != port) {
646 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
647 log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
648 "syncache entry but non-matching UDP encaps port, "
649 "segment ignored\n", s, __func__);
650 TCPSTAT_INC(tcps_badrst);
651 goto done;
652 }
653
654 /*
655 * If the RST bit is set, check the sequence number to see
656 * if this is a valid reset segment.
657 *
658 * RFC 793 page 37:
659 * In all states except SYN-SENT, all reset (RST) segments
660 * are validated by checking their SEQ-fields. A reset is
661 * valid if its sequence number is in the window.
662 *
663 * RFC 793 page 69:
664 * There are four cases for the acceptability test for an incoming
665 * segment:
666 *
667 * Segment Receive Test
668 * Length Window
669 * ------- ------- -------------------------------------------
670 * 0 0 SEG.SEQ = RCV.NXT
671 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
672 * >0 0 not acceptable
673 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
674 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
675 *
676 * Note that when receiving a SYN segment in the LISTEN state,
677 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
678 * described in RFC 793, page 66.
679 */
680 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
681 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
682 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
683 if (V_tcp_insecure_rst ||
684 th->th_seq == sc->sc_irs + 1) {
685 syncache_drop(sc, sch);
686 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
687 log(LOG_DEBUG,
688 "%s; %s: Our SYN|ACK was rejected, "
689 "connection attempt aborted by remote "
690 "endpoint\n",
691 s, __func__);
692 TCPSTAT_INC(tcps_sc_reset);
693 } else {
694 TCPSTAT_INC(tcps_badrst);
695 /* Send challenge ACK. */
696 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
697 log(LOG_DEBUG, "%s; %s: RST with invalid "
698 " SEQ %u != NXT %u (+WND %u), "
699 "sending challenge ACK\n",
700 s, __func__,
701 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
702 syncache_send_challenge_ack(sc, m);
703 }
704 } else {
705 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
706 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
707 "NXT %u (+WND %u), segment ignored\n",
708 s, __func__,
709 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
710 TCPSTAT_INC(tcps_badrst);
711 }
712
713 done:
714 if (s != NULL)
715 free(s, M_TCPLOG);
716 SCH_UNLOCK(sch);
717 }
718
719 void
syncache_unreach(struct in_conninfo * inc,tcp_seq th_seq,uint16_t port)720 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
721 {
722 struct syncache *sc;
723 struct syncache_head *sch;
724
725 if (syncache_cookiesonly())
726 return;
727 sc = syncache_lookup(inc, &sch); /* returns locked sch */
728 SCH_LOCK_ASSERT(sch);
729 if (sc == NULL)
730 goto done;
731
732 /* If the port != sc_port, then it's a bogus ICMP msg */
733 if (port != sc->sc_port)
734 goto done;
735
736 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
737 if (ntohl(th_seq) != sc->sc_iss)
738 goto done;
739
740 /*
741 * If we've retransmitted 3 times and this is our second error,
742 * we remove the entry. Otherwise, we allow it to continue on.
743 * This prevents us from incorrectly nuking an entry during a
744 * spurious network outage.
745 *
746 * See tcp_notify().
747 */
748 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
749 sc->sc_flags |= SCF_UNREACH;
750 goto done;
751 }
752 syncache_drop(sc, sch);
753 TCPSTAT_INC(tcps_sc_unreach);
754 done:
755 SCH_UNLOCK(sch);
756 }
757
758 /*
759 * Build a new TCP socket structure from a syncache entry.
760 *
761 * On success return the newly created socket with its underlying inp locked.
762 */
763 static struct socket *
syncache_socket(struct syncache * sc,struct socket * lso,struct mbuf * m)764 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
765 {
766 struct inpcb *inp = NULL;
767 struct socket *so;
768 struct tcpcb *tp;
769 int error;
770 char *s;
771
772 NET_EPOCH_ASSERT();
773
774 /*
775 * Ok, create the full blown connection, and set things up
776 * as they would have been set up if we had created the
777 * connection when the SYN arrived.
778 */
779 if ((so = solisten_clone(lso)) == NULL)
780 goto allocfail;
781 #ifdef MAC
782 mac_socketpeer_set_from_mbuf(m, so);
783 #endif
784 error = in_pcballoc(so, &V_tcbinfo);
785 if (error) {
786 sodealloc(so);
787 goto allocfail;
788 }
789 inp = sotoinpcb(so);
790 if ((tp = tcp_newtcpcb(inp, sototcpcb(lso))) == NULL) {
791 in_pcbfree(inp);
792 sodealloc(so);
793 goto allocfail;
794 }
795 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
796 #ifdef INET6
797 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
798 inp->inp_vflag &= ~INP_IPV4;
799 inp->inp_vflag |= INP_IPV6;
800 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
801 } else {
802 inp->inp_vflag &= ~INP_IPV6;
803 inp->inp_vflag |= INP_IPV4;
804 #endif
805 inp->inp_ip_ttl = sc->sc_ip_ttl;
806 inp->inp_ip_tos = sc->sc_ip_tos;
807 inp->inp_laddr = sc->sc_inc.inc_laddr;
808 #ifdef INET6
809 }
810 #endif
811
812 /*
813 * If there's an mbuf and it has a flowid, then let's initialise the
814 * inp with that particular flowid.
815 */
816 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
817 inp->inp_flowid = m->m_pkthdr.flowid;
818 inp->inp_flowtype = M_HASHTYPE_GET(m);
819 #ifdef NUMA
820 inp->inp_numa_domain = m->m_pkthdr.numa_domain;
821 #endif
822 }
823
824 inp->inp_lport = sc->sc_inc.inc_lport;
825 #ifdef INET6
826 if (inp->inp_vflag & INP_IPV6PROTO) {
827 struct inpcb *oinp = sotoinpcb(lso);
828
829 /*
830 * Inherit socket options from the listening socket.
831 * Note that in6p_inputopts are not (and should not be)
832 * copied, since it stores previously received options and is
833 * used to detect if each new option is different than the
834 * previous one and hence should be passed to a user.
835 * If we copied in6p_inputopts, a user would not be able to
836 * receive options just after calling the accept system call.
837 */
838 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
839 if (oinp->in6p_outputopts)
840 inp->in6p_outputopts =
841 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
842 inp->in6p_hops = oinp->in6p_hops;
843 }
844
845 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
846 struct sockaddr_in6 sin6;
847
848 sin6.sin6_family = AF_INET6;
849 sin6.sin6_len = sizeof(sin6);
850 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
851 sin6.sin6_port = sc->sc_inc.inc_fport;
852 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
853 INP_HASH_WLOCK(&V_tcbinfo);
854 error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
855 INP_HASH_WUNLOCK(&V_tcbinfo);
856 if (error != 0)
857 goto abort;
858 /* Override flowlabel from in6_pcbconnect. */
859 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
860 inp->inp_flow |= sc->sc_flowlabel;
861 }
862 #endif /* INET6 */
863 #if defined(INET) && defined(INET6)
864 else
865 #endif
866 #ifdef INET
867 {
868 struct sockaddr_in sin;
869
870 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
871
872 if (inp->inp_options == NULL) {
873 inp->inp_options = sc->sc_ipopts;
874 sc->sc_ipopts = NULL;
875 }
876
877 sin.sin_family = AF_INET;
878 sin.sin_len = sizeof(sin);
879 sin.sin_addr = sc->sc_inc.inc_faddr;
880 sin.sin_port = sc->sc_inc.inc_fport;
881 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
882 INP_HASH_WLOCK(&V_tcbinfo);
883 error = in_pcbconnect(inp, &sin, thread0.td_ucred);
884 INP_HASH_WUNLOCK(&V_tcbinfo);
885 if (error != 0)
886 goto abort;
887 }
888 #endif /* INET */
889 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
890 /* Copy old policy into new socket's. */
891 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
892 printf("syncache_socket: could not copy policy\n");
893 #endif
894 tp->t_state = TCPS_SYN_RECEIVED;
895 tp->iss = sc->sc_iss;
896 tp->irs = sc->sc_irs;
897 tp->t_port = sc->sc_port;
898 tcp_rcvseqinit(tp);
899 tcp_sendseqinit(tp);
900 tp->snd_wl1 = sc->sc_irs;
901 tp->snd_max = tp->iss + 1;
902 tp->snd_nxt = tp->iss + 1;
903 tp->rcv_up = sc->sc_irs + 1;
904 tp->rcv_wnd = sc->sc_wnd;
905 tp->rcv_adv += tp->rcv_wnd;
906 tp->last_ack_sent = tp->rcv_nxt;
907
908 tp->t_flags = sototcpcb(lso)->t_flags &
909 (TF_LRD|TF_NOPUSH|TF_NODELAY);
910 if (sc->sc_flags & SCF_NOOPT)
911 tp->t_flags |= TF_NOOPT;
912 else {
913 if (sc->sc_flags & SCF_WINSCALE) {
914 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
915 tp->snd_scale = sc->sc_requested_s_scale;
916 tp->request_r_scale = sc->sc_requested_r_scale;
917 }
918 if (sc->sc_flags & SCF_TIMESTAMP) {
919 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
920 tp->ts_recent = sc->sc_tsreflect;
921 tp->ts_recent_age = tcp_ts_getticks();
922 tp->ts_offset = sc->sc_tsoff;
923 }
924 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
925 if (sc->sc_flags & SCF_SIGNATURE)
926 tp->t_flags |= TF_SIGNATURE;
927 #endif
928 if (sc->sc_flags & SCF_SACK)
929 tp->t_flags |= TF_SACK_PERMIT;
930 }
931
932 tcp_ecn_syncache_socket(tp, sc);
933
934 /*
935 * Set up MSS and get cached values from tcp_hostcache.
936 * This might overwrite some of the defaults we just set.
937 */
938 tcp_mss(tp, sc->sc_peer_mss);
939
940 /*
941 * If the SYN,ACK was retransmitted, indicate that CWND to be
942 * limited to one segment in cc_conn_init().
943 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
944 */
945 if (sc->sc_rxmits > 1)
946 tp->snd_cwnd = 1;
947
948 /* Copy over the challenge ACK state. */
949 tp->t_challenge_ack_end = sc->sc_challenge_ack_end;
950 tp->t_challenge_ack_cnt = sc->sc_challenge_ack_cnt;
951
952 #ifdef TCP_OFFLOAD
953 /*
954 * Allow a TOE driver to install its hooks. Note that we hold the
955 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
956 * new connection before the TOE driver has done its thing.
957 */
958 if (ADDED_BY_TOE(sc)) {
959 struct toedev *tod = sc->sc_tod;
960
961 tod->tod_offload_socket(tod, sc->sc_todctx, so);
962 }
963 #endif
964 #ifdef TCP_BLACKBOX
965 /*
966 * Inherit the log state from the listening socket, if
967 * - the log state of the listening socket is not off and
968 * - the listening socket was not auto selected from all sessions and
969 * - a log id is not set on the listening socket.
970 * This avoids inheriting a log state which was automatically set.
971 */
972 if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
973 ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
974 (sototcpcb(lso)->t_lib == NULL)) {
975 tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
976 }
977 #endif
978 /*
979 * Copy and activate timers.
980 */
981 tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
982 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
983 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
984 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
985 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
986 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
987
988 TCPSTAT_INC(tcps_accepts);
989 TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
990
991 if (!solisten_enqueue(so, SS_ISCONNECTED))
992 tp->t_flags |= TF_SONOTCONN;
993 /* Can we inherit anything from the listener? */
994 if (tp->t_fb->tfb_inherit != NULL) {
995 (*tp->t_fb->tfb_inherit)(tp, sotoinpcb(lso));
996 }
997 return (so);
998
999 allocfail:
1000 /*
1001 * Drop the connection; we will either send a RST or have the peer
1002 * retransmit its SYN again after its RTO and try again.
1003 */
1004 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1005 log(LOG_DEBUG, "%s; %s: Socket create failed "
1006 "due to limits or memory shortage\n",
1007 s, __func__);
1008 free(s, M_TCPLOG);
1009 }
1010 TCPSTAT_INC(tcps_listendrop);
1011 return (NULL);
1012
1013 abort:
1014 tcp_discardcb(tp);
1015 in_pcbfree(inp);
1016 sodealloc(so);
1017 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1018 log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1019 s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1020 error);
1021 free(s, M_TCPLOG);
1022 }
1023 TCPSTAT_INC(tcps_listendrop);
1024 return (NULL);
1025 }
1026
1027 /*
1028 * This function gets called when we receive an ACK for a
1029 * socket in the LISTEN state. We look up the connection
1030 * in the syncache, and if its there, we pull it out of
1031 * the cache and turn it into a full-blown connection in
1032 * the SYN-RECEIVED state.
1033 *
1034 * On syncache_socket() success the newly created socket
1035 * has its underlying inp locked.
1036 *
1037 * *lsop is updated, if and only if 1 is returned.
1038 */
1039 int
syncache_expand(struct in_conninfo * inc,struct tcpopt * to,struct tcphdr * th,struct socket ** lsop,struct mbuf * m,uint16_t port)1040 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1041 struct socket **lsop, struct mbuf *m, uint16_t port)
1042 {
1043 struct syncache *sc;
1044 struct syncache_head *sch;
1045 struct syncache scs;
1046 char *s;
1047 bool locked;
1048
1049 NET_EPOCH_ASSERT();
1050 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1051 ("%s: can handle only ACK", __func__));
1052
1053 if (syncache_cookiesonly()) {
1054 sc = NULL;
1055 sch = syncache_hashbucket(inc);
1056 locked = false;
1057 } else {
1058 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1059 locked = true;
1060 SCH_LOCK_ASSERT(sch);
1061 }
1062
1063 #ifdef INVARIANTS
1064 /*
1065 * Test code for syncookies comparing the syncache stored
1066 * values with the reconstructed values from the cookie.
1067 */
1068 if (sc != NULL)
1069 syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1070 #endif
1071
1072 if (sc == NULL) {
1073 if (locked) {
1074 /*
1075 * The syncache is currently in use (neither disabled,
1076 * nor paused), but no entry was found.
1077 */
1078 if (!V_tcp_syncookies) {
1079 /*
1080 * Since no syncookies are used in case of
1081 * a bucket overflow, don't even check for
1082 * a valid syncookie.
1083 */
1084 SCH_UNLOCK(sch);
1085 TCPSTAT_INC(tcps_sc_spurcookie);
1086 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1087 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1088 "segment rejected "
1089 "(syncookies disabled)\n",
1090 s, __func__);
1091 free(s, M_TCPLOG);
1092 }
1093 return (0);
1094 }
1095 if (sch->sch_last_overflow <
1096 time_uptime - SYNCOOKIE_LIFETIME) {
1097 /*
1098 * Since the bucket did not overflow recently,
1099 * don't even check for a valid syncookie.
1100 */
1101 SCH_UNLOCK(sch);
1102 TCPSTAT_INC(tcps_sc_spurcookie);
1103 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1104 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1105 "segment rejected "
1106 "(no syncache entry)\n",
1107 s, __func__);
1108 free(s, M_TCPLOG);
1109 }
1110 return (0);
1111 }
1112 SCH_UNLOCK(sch);
1113 }
1114 bzero(&scs, sizeof(scs));
1115 /*
1116 * Now check, if the syncookie is valid. If it is, create an on
1117 * stack syncache entry.
1118 */
1119 if (syncookie_expand(inc, sch, &scs, th, to, *lsop, port)) {
1120 sc = &scs;
1121 TCPSTAT_INC(tcps_sc_recvcookie);
1122 } else {
1123 TCPSTAT_INC(tcps_sc_failcookie);
1124 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1125 log(LOG_DEBUG, "%s; %s: Segment failed "
1126 "SYNCOOKIE authentication, segment rejected "
1127 "(probably spoofed)\n", s, __func__);
1128 free(s, M_TCPLOG);
1129 }
1130 return (0);
1131 }
1132 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1133 /* If received ACK has MD5 signature, check it. */
1134 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1135 (!TCPMD5_ENABLED() ||
1136 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1137 /* Drop the ACK. */
1138 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1139 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1140 "MD5 signature doesn't match.\n",
1141 s, __func__);
1142 free(s, M_TCPLOG);
1143 }
1144 TCPSTAT_INC(tcps_sig_err_sigopt);
1145 return (-1); /* Do not send RST */
1146 }
1147 #endif /* TCP_SIGNATURE */
1148 TCPSTATES_INC(TCPS_SYN_RECEIVED);
1149 } else {
1150 if (sc->sc_port != port) {
1151 SCH_UNLOCK(sch);
1152 return (0);
1153 }
1154 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1155 /*
1156 * If listening socket requested TCP digests, check that
1157 * received ACK has signature and it is correct.
1158 * If not, drop the ACK and leave sc entry in the cache,
1159 * because SYN was received with correct signature.
1160 */
1161 if (sc->sc_flags & SCF_SIGNATURE) {
1162 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1163 /* No signature */
1164 TCPSTAT_INC(tcps_sig_err_nosigopt);
1165 SCH_UNLOCK(sch);
1166 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1167 log(LOG_DEBUG, "%s; %s: Segment "
1168 "rejected, MD5 signature wasn't "
1169 "provided.\n", s, __func__);
1170 free(s, M_TCPLOG);
1171 }
1172 return (-1); /* Do not send RST */
1173 }
1174 if (!TCPMD5_ENABLED() ||
1175 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1176 /* Doesn't match or no SA */
1177 SCH_UNLOCK(sch);
1178 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1179 log(LOG_DEBUG, "%s; %s: Segment "
1180 "rejected, MD5 signature doesn't "
1181 "match.\n", s, __func__);
1182 free(s, M_TCPLOG);
1183 }
1184 return (-1); /* Do not send RST */
1185 }
1186 }
1187 #endif /* TCP_SIGNATURE */
1188
1189 /*
1190 * RFC 7323 PAWS: If we have a timestamp on this segment and
1191 * it's less than ts_recent, drop it.
1192 * XXXMT: RFC 7323 also requires to send an ACK.
1193 * In tcp_input.c this is only done for TCP segments
1194 * with user data, so be consistent here and just drop
1195 * the segment.
1196 */
1197 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1198 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1199 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1200 log(LOG_DEBUG,
1201 "%s; %s: SEG.TSval %u < TS.Recent %u, "
1202 "segment dropped\n", s, __func__,
1203 to->to_tsval, sc->sc_tsreflect);
1204 }
1205 SCH_UNLOCK(sch);
1206 free(s, M_TCPLOG);
1207 return (-1); /* Do not send RST */
1208 }
1209
1210 /*
1211 * If timestamps were not negotiated during SYN/ACK and a
1212 * segment with a timestamp is received, ignore the
1213 * timestamp and process the packet normally.
1214 * See section 3.2 of RFC 7323.
1215 */
1216 if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1217 (to->to_flags & TOF_TS)) {
1218 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1219 log(LOG_DEBUG, "%s; %s: Timestamp not "
1220 "expected, segment processed normally\n",
1221 s, __func__);
1222 free(s, M_TCPLOG);
1223 }
1224 }
1225
1226 /*
1227 * If timestamps were negotiated during SYN/ACK and a
1228 * segment without a timestamp is received, silently drop
1229 * the segment, unless the missing timestamps are tolerated.
1230 * See section 3.2 of RFC 7323.
1231 */
1232 if ((sc->sc_flags & SCF_TIMESTAMP) &&
1233 !(to->to_flags & TOF_TS)) {
1234 if (V_tcp_tolerate_missing_ts) {
1235 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1236 log(LOG_DEBUG,
1237 "%s; %s: Timestamp missing, "
1238 "segment processed normally\n",
1239 s, __func__);
1240 free(s, M_TCPLOG);
1241 }
1242 } else {
1243 SCH_UNLOCK(sch);
1244 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1245 log(LOG_DEBUG,
1246 "%s; %s: Timestamp missing, "
1247 "segment silently dropped\n",
1248 s, __func__);
1249 free(s, M_TCPLOG);
1250 }
1251 return (-1); /* Do not send RST */
1252 }
1253 }
1254
1255 /*
1256 * SEG.SEQ validation:
1257 * The SEG.SEQ must be in the window starting at our
1258 * initial receive sequence number + 1.
1259 */
1260 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1261 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1262 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1263 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, "
1264 "sending challenge ACK\n",
1265 s, __func__, th->th_seq, sc->sc_irs + 1);
1266 syncache_send_challenge_ack(sc, m);
1267 SCH_UNLOCK(sch);
1268 free(s, M_TCPLOG);
1269 return (-1); /* Do not send RST */
1270 }
1271
1272 /*
1273 * SEG.ACK validation:
1274 * SEG.ACK must match our initial send sequence number + 1.
1275 */
1276 if (th->th_ack != sc->sc_iss + 1) {
1277 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1278 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, "
1279 "segment rejected\n",
1280 s, __func__, th->th_ack, sc->sc_iss + 1);
1281 SCH_UNLOCK(sch);
1282 free(s, M_TCPLOG);
1283 return (0); /* Do send RST, do not free sc. */
1284 }
1285
1286 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1287 sch->sch_length--;
1288 #ifdef TCP_OFFLOAD
1289 if (ADDED_BY_TOE(sc)) {
1290 struct toedev *tod = sc->sc_tod;
1291
1292 tod->tod_syncache_removed(tod, sc->sc_todctx);
1293 }
1294 #endif
1295 SCH_UNLOCK(sch);
1296 }
1297
1298 *lsop = syncache_socket(sc, *lsop, m);
1299
1300 if (__predict_false(*lsop == NULL)) {
1301 TCPSTAT_INC(tcps_sc_aborted);
1302 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1303 } else if (sc != &scs)
1304 TCPSTAT_INC(tcps_sc_completed);
1305
1306 if (sc != &scs)
1307 syncache_free(sc);
1308 return (1);
1309 }
1310
1311 static struct socket *
syncache_tfo_expand(struct syncache * sc,struct socket * lso,struct mbuf * m,uint64_t response_cookie)1312 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1313 uint64_t response_cookie)
1314 {
1315 struct inpcb *inp;
1316 struct tcpcb *tp;
1317 unsigned int *pending_counter;
1318 struct socket *so;
1319
1320 NET_EPOCH_ASSERT();
1321
1322 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1323 so = syncache_socket(sc, lso, m);
1324 if (so == NULL) {
1325 TCPSTAT_INC(tcps_sc_aborted);
1326 atomic_subtract_int(pending_counter, 1);
1327 } else {
1328 soisconnected(so);
1329 inp = sotoinpcb(so);
1330 tp = intotcpcb(inp);
1331 tp->t_flags |= TF_FASTOPEN;
1332 tp->t_tfo_cookie.server = response_cookie;
1333 tp->snd_max = tp->iss;
1334 tp->snd_nxt = tp->iss;
1335 tp->t_tfo_pending = pending_counter;
1336 TCPSTATES_INC(TCPS_SYN_RECEIVED);
1337 TCPSTAT_INC(tcps_sc_completed);
1338 }
1339
1340 return (so);
1341 }
1342
1343 /*
1344 * Given a LISTEN socket and an inbound SYN request, add
1345 * this to the syn cache, and send back a segment:
1346 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1347 * to the source.
1348 *
1349 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1350 * Doing so would require that we hold onto the data and deliver it
1351 * to the application. However, if we are the target of a SYN-flood
1352 * DoS attack, an attacker could send data which would eventually
1353 * consume all available buffer space if it were ACKed. By not ACKing
1354 * the data, we avoid this DoS scenario.
1355 *
1356 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1357 * cookie is processed and a new socket is created. In this case, any data
1358 * accompanying the SYN will be queued to the socket by tcp_input() and will
1359 * be ACKed either when the application sends response data or the delayed
1360 * ACK timer expires, whichever comes first.
1361 */
1362 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)1363 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1364 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1365 void *todctx, uint8_t iptos, uint16_t port)
1366 {
1367 struct tcpcb *tp;
1368 struct socket *rv = NULL;
1369 struct syncache *sc = NULL;
1370 struct ucred *cred;
1371 struct syncache_head *sch;
1372 struct mbuf *ipopts = NULL;
1373 u_int ltflags;
1374 int win, ip_ttl, ip_tos;
1375 char *s;
1376 #ifdef INET6
1377 int autoflowlabel = 0;
1378 #endif
1379 #ifdef MAC
1380 struct label *maclabel = NULL;
1381 #endif
1382 struct syncache scs;
1383 uint64_t tfo_response_cookie;
1384 unsigned int *tfo_pending = NULL;
1385 int tfo_cookie_valid = 0;
1386 int tfo_response_cookie_valid = 0;
1387 bool locked;
1388
1389 INP_RLOCK_ASSERT(inp); /* listen socket */
1390 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1391 ("%s: unexpected tcp flags", __func__));
1392
1393 /*
1394 * Combine all so/tp operations very early to drop the INP lock as
1395 * soon as possible.
1396 */
1397 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1398 tp = sototcpcb(so);
1399 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1400
1401 #ifdef INET6
1402 if (inc->inc_flags & INC_ISIPV6) {
1403 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1404 autoflowlabel = 1;
1405 }
1406 ip_ttl = in6_selecthlim(inp, NULL);
1407 if ((inp->in6p_outputopts == NULL) ||
1408 (inp->in6p_outputopts->ip6po_tclass == -1)) {
1409 ip_tos = 0;
1410 } else {
1411 ip_tos = inp->in6p_outputopts->ip6po_tclass;
1412 }
1413 }
1414 #endif
1415 #if defined(INET6) && defined(INET)
1416 else
1417 #endif
1418 #ifdef INET
1419 {
1420 ip_ttl = inp->inp_ip_ttl;
1421 ip_tos = inp->inp_ip_tos;
1422 }
1423 #endif
1424 win = so->sol_sbrcv_hiwat;
1425 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1426
1427 if (V_tcp_fastopen_server_enable && (tp->t_flags & TF_FASTOPEN) &&
1428 (tp->t_tfo_pending != NULL) &&
1429 (to->to_flags & TOF_FASTOPEN)) {
1430 /*
1431 * Limit the number of pending TFO connections to
1432 * approximately half of the queue limit. This prevents TFO
1433 * SYN floods from starving the service by filling the
1434 * listen queue with bogus TFO connections.
1435 */
1436 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1437 (so->sol_qlimit / 2)) {
1438 int result;
1439
1440 result = tcp_fastopen_check_cookie(inc,
1441 to->to_tfo_cookie, to->to_tfo_len,
1442 &tfo_response_cookie);
1443 tfo_cookie_valid = (result > 0);
1444 tfo_response_cookie_valid = (result >= 0);
1445 }
1446
1447 /*
1448 * Remember the TFO pending counter as it will have to be
1449 * decremented below if we don't make it to syncache_tfo_expand().
1450 */
1451 tfo_pending = tp->t_tfo_pending;
1452 }
1453
1454 #ifdef MAC
1455 if (mac_syncache_init(&maclabel) != 0) {
1456 INP_RUNLOCK(inp);
1457 goto done;
1458 } else
1459 mac_syncache_create(maclabel, inp);
1460 #endif
1461 if (!tfo_cookie_valid)
1462 INP_RUNLOCK(inp);
1463
1464 /*
1465 * Remember the IP options, if any.
1466 */
1467 #ifdef INET6
1468 if (!(inc->inc_flags & INC_ISIPV6))
1469 #endif
1470 #ifdef INET
1471 ipopts = (m) ? ip_srcroute(m) : NULL;
1472 #else
1473 ipopts = NULL;
1474 #endif
1475
1476 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1477 /*
1478 * When the socket is TCP-MD5 enabled check that,
1479 * - a signed packet is valid
1480 * - a non-signed packet does not have a security association
1481 *
1482 * If a signed packet fails validation or a non-signed packet has a
1483 * security association, the packet will be dropped.
1484 */
1485 if (ltflags & TF_SIGNATURE) {
1486 if (to->to_flags & TOF_SIGNATURE) {
1487 if (!TCPMD5_ENABLED() ||
1488 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1489 goto done;
1490 } else {
1491 if (TCPMD5_ENABLED() &&
1492 TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1493 goto done;
1494 }
1495 } else if (to->to_flags & TOF_SIGNATURE)
1496 goto done;
1497 #endif /* TCP_SIGNATURE */
1498 /*
1499 * See if we already have an entry for this connection.
1500 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1501 *
1502 * XXX: should the syncache be re-initialized with the contents
1503 * of the new SYN here (which may have different options?)
1504 *
1505 * XXX: We do not check the sequence number to see if this is a
1506 * real retransmit or a new connection attempt. The question is
1507 * how to handle such a case; either ignore it as spoofed, or
1508 * drop the current entry and create a new one?
1509 */
1510 if (syncache_cookiesonly()) {
1511 sc = NULL;
1512 sch = syncache_hashbucket(inc);
1513 locked = false;
1514 } else {
1515 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1516 locked = true;
1517 SCH_LOCK_ASSERT(sch);
1518 }
1519 if (sc != NULL) {
1520 if (tfo_cookie_valid)
1521 INP_RUNLOCK(inp);
1522 TCPSTAT_INC(tcps_sc_dupsyn);
1523 if (ipopts) {
1524 /*
1525 * If we were remembering a previous source route,
1526 * forget it and use the new one we've been given.
1527 */
1528 if (sc->sc_ipopts)
1529 (void)m_free(sc->sc_ipopts);
1530 sc->sc_ipopts = ipopts;
1531 }
1532 /*
1533 * Update timestamp if present.
1534 */
1535 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1536 sc->sc_tsreflect = to->to_tsval;
1537 else
1538 sc->sc_flags &= ~SCF_TIMESTAMP;
1539 /*
1540 * Adjust ECN response if needed, e.g. different
1541 * IP ECN field, or a fallback by the remote host.
1542 */
1543 if (sc->sc_flags & SCF_ECN_MASK) {
1544 sc->sc_flags &= ~SCF_ECN_MASK;
1545 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1546 }
1547 #ifdef MAC
1548 /*
1549 * Since we have already unconditionally allocated label
1550 * storage, free it up. The syncache entry will already
1551 * have an initialized label we can use.
1552 */
1553 mac_syncache_destroy(&maclabel);
1554 #endif
1555 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1556 /* Retransmit SYN|ACK and reset retransmit count. */
1557 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1558 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1559 "resetting timer and retransmitting SYN|ACK\n",
1560 s, __func__);
1561 free(s, M_TCPLOG);
1562 }
1563 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1564 sc->sc_rxmits = 0;
1565 syncache_timeout(sc, sch, 1);
1566 TCPSTAT_INC(tcps_sndacks);
1567 TCPSTAT_INC(tcps_sndtotal);
1568 } else {
1569 /*
1570 * Most likely we are memory constrained, so free
1571 * resources.
1572 */
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 only when needed:
1633 * - sc is allocated from the zone and not using the on stack instance.
1634 * - the sysctl variable net.inet.tcp.syncache.see_other is false.
1635 * The reference count is decremented when a zone allocated sc is
1636 * freed in syncache_free().
1637 */
1638 if (sc != &scs && !V_tcp_syncache.see_other) {
1639 sc->sc_cred = cred;
1640 cred = NULL;
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 u_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 || V_tcp_syncookiesonly)
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 || V_tcp_syncookiesonly)
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 /*
1759 * Most likely we are memory constrained, so free resources.
1760 */
1761 if (sc != &scs)
1762 syncache_free(sc);
1763 TCPSTAT_INC(tcps_sc_dropped);
1764 }
1765 goto donenoprobe;
1766
1767 done:
1768 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1769 donenoprobe:
1770 if (m)
1771 m_freem(m);
1772 /*
1773 * If tfo_pending is not NULL here, then a TFO SYN that did not
1774 * result in a new socket was processed and the associated pending
1775 * counter has not yet been decremented. All such TFO processing paths
1776 * transit this point.
1777 */
1778 if (tfo_pending != NULL)
1779 tcp_fastopen_decrement_counter(tfo_pending);
1780
1781 tfo_expanded:
1782 if (cred != NULL)
1783 crfree(cred);
1784 if (sc == NULL || sc == &scs) {
1785 #ifdef MAC
1786 mac_syncache_destroy(&maclabel);
1787 #endif
1788 if (ipopts)
1789 (void)m_free(ipopts);
1790 }
1791 return (rv);
1792 }
1793
1794 /*
1795 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment,
1796 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1797 */
1798 static int
syncache_respond(struct syncache * sc,const struct mbuf * m0,int flags)1799 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1800 {
1801 struct ip *ip = NULL;
1802 struct mbuf *m;
1803 struct tcphdr *th = NULL;
1804 struct udphdr *udp = NULL;
1805 int optlen, error = 0; /* Make compiler happy */
1806 u_int16_t hlen, tlen, mssopt, ulen;
1807 struct tcpopt to;
1808 #ifdef INET6
1809 struct ip6_hdr *ip6 = NULL;
1810 #endif
1811
1812 NET_EPOCH_ASSERT();
1813
1814 hlen =
1815 #ifdef INET6
1816 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1817 #endif
1818 sizeof(struct ip);
1819 tlen = hlen + sizeof(struct tcphdr);
1820 if (sc->sc_port) {
1821 tlen += sizeof(struct udphdr);
1822 }
1823 /* Determine MSS we advertize to other end of connection. */
1824 mssopt = tcp_mssopt(&sc->sc_inc);
1825 if (sc->sc_port)
1826 mssopt -= V_tcp_udp_tunneling_overhead;
1827 mssopt = max(mssopt, V_tcp_minmss);
1828
1829 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1830 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1831 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1832 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1833 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1834
1835 /* Create the IP+TCP header from scratch. */
1836 m = m_gethdr(M_NOWAIT, MT_DATA);
1837 if (m == NULL)
1838 return (ENOBUFS);
1839 #ifdef MAC
1840 mac_syncache_create_mbuf(sc->sc_label, m);
1841 #endif
1842 m->m_data += max_linkhdr;
1843 m->m_len = tlen;
1844 m->m_pkthdr.len = tlen;
1845 m->m_pkthdr.rcvif = NULL;
1846
1847 #ifdef INET6
1848 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1849 ip6 = mtod(m, struct ip6_hdr *);
1850 ip6->ip6_vfc = IPV6_VERSION;
1851 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1852 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1853 ip6->ip6_plen = htons(tlen - hlen);
1854 /* ip6_hlim is set after checksum */
1855 /* Zero out traffic class and flow label. */
1856 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1857 ip6->ip6_flow |= sc->sc_flowlabel;
1858 if (sc->sc_port != 0) {
1859 ip6->ip6_nxt = IPPROTO_UDP;
1860 udp = (struct udphdr *)(ip6 + 1);
1861 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1862 udp->uh_dport = sc->sc_port;
1863 ulen = (tlen - sizeof(struct ip6_hdr));
1864 th = (struct tcphdr *)(udp + 1);
1865 } else {
1866 ip6->ip6_nxt = IPPROTO_TCP;
1867 th = (struct tcphdr *)(ip6 + 1);
1868 }
1869 ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1870 }
1871 #endif
1872 #if defined(INET6) && defined(INET)
1873 else
1874 #endif
1875 #ifdef INET
1876 {
1877 ip = mtod(m, struct ip *);
1878 ip->ip_v = IPVERSION;
1879 ip->ip_hl = sizeof(struct ip) >> 2;
1880 ip->ip_len = htons(tlen);
1881 ip->ip_id = 0;
1882 ip->ip_off = 0;
1883 ip->ip_sum = 0;
1884 ip->ip_src = sc->sc_inc.inc_laddr;
1885 ip->ip_dst = sc->sc_inc.inc_faddr;
1886 ip->ip_ttl = sc->sc_ip_ttl;
1887 ip->ip_tos = sc->sc_ip_tos;
1888
1889 /*
1890 * See if we should do MTU discovery. Route lookups are
1891 * expensive, so we will only unset the DF bit if:
1892 *
1893 * 1) path_mtu_discovery is disabled
1894 * 2) the SCF_UNREACH flag has been set
1895 */
1896 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1897 ip->ip_off |= htons(IP_DF);
1898 if (sc->sc_port == 0) {
1899 ip->ip_p = IPPROTO_TCP;
1900 th = (struct tcphdr *)(ip + 1);
1901 } else {
1902 ip->ip_p = IPPROTO_UDP;
1903 udp = (struct udphdr *)(ip + 1);
1904 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1905 udp->uh_dport = sc->sc_port;
1906 ulen = (tlen - sizeof(struct ip));
1907 th = (struct tcphdr *)(udp + 1);
1908 }
1909 }
1910 #endif /* INET */
1911 th->th_sport = sc->sc_inc.inc_lport;
1912 th->th_dport = sc->sc_inc.inc_fport;
1913
1914 if (flags & TH_SYN)
1915 th->th_seq = htonl(sc->sc_iss);
1916 else
1917 th->th_seq = htonl(sc->sc_iss + 1);
1918 th->th_ack = htonl(sc->sc_irs + 1);
1919 th->th_off = sizeof(struct tcphdr) >> 2;
1920 th->th_win = htons(sc->sc_wnd);
1921 th->th_urp = 0;
1922
1923 flags = tcp_ecn_syncache_respond(flags, sc);
1924 tcp_set_flags(th, flags);
1925
1926 /* Tack on the TCP options. */
1927 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1928 to.to_flags = 0;
1929
1930 if (flags & TH_SYN) {
1931 to.to_mss = mssopt;
1932 to.to_flags = TOF_MSS;
1933 if (sc->sc_flags & SCF_WINSCALE) {
1934 to.to_wscale = sc->sc_requested_r_scale;
1935 to.to_flags |= TOF_SCALE;
1936 }
1937 if (sc->sc_flags & SCF_SACK)
1938 to.to_flags |= TOF_SACKPERM;
1939 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1940 if (sc->sc_flags & SCF_SIGNATURE)
1941 to.to_flags |= TOF_SIGNATURE;
1942 #endif
1943 if (sc->sc_tfo_cookie) {
1944 to.to_flags |= TOF_FASTOPEN;
1945 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1946 to.to_tfo_cookie = sc->sc_tfo_cookie;
1947 /* don't send cookie again when retransmitting response */
1948 sc->sc_tfo_cookie = NULL;
1949 }
1950 }
1951 if (sc->sc_flags & SCF_TIMESTAMP) {
1952 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1953 to.to_tsecr = sc->sc_tsreflect;
1954 to.to_flags |= TOF_TS;
1955 }
1956 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1957
1958 /* Adjust headers by option size. */
1959 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1960 m->m_len += optlen;
1961 m->m_pkthdr.len += optlen;
1962 #ifdef INET6
1963 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1964 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1965 else
1966 #endif
1967 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1968 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1969 if (sc->sc_flags & SCF_SIGNATURE) {
1970 KASSERT(to.to_flags & TOF_SIGNATURE,
1971 ("tcp_addoptions() didn't set tcp_signature"));
1972
1973 /* NOTE: to.to_signature is inside of mbuf */
1974 if (!TCPMD5_ENABLED() ||
1975 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1976 m_freem(m);
1977 return (EACCES);
1978 }
1979 }
1980 #endif
1981 } else
1982 optlen = 0;
1983
1984 if (udp) {
1985 ulen += optlen;
1986 udp->uh_ulen = htons(ulen);
1987 }
1988 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1989 /*
1990 * If we have peer's SYN and it has a flowid, then let's assign it to
1991 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1992 * to SYN|ACK due to lack of inp here.
1993 */
1994 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1995 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1996 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1997 }
1998 #ifdef INET6
1999 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2000 if (sc->sc_port) {
2001 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2002 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2003 udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2004 IPPROTO_UDP, 0);
2005 th->th_sum = htons(0);
2006 } else {
2007 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2008 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2009 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2010 IPPROTO_TCP, 0);
2011 }
2012 ip6->ip6_hlim = sc->sc_ip_ttl;
2013 #ifdef TCP_OFFLOAD
2014 if (ADDED_BY_TOE(sc)) {
2015 struct toedev *tod = sc->sc_tod;
2016
2017 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2018
2019 return (error);
2020 }
2021 #endif
2022 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2023 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2024 }
2025 #endif
2026 #if defined(INET6) && defined(INET)
2027 else
2028 #endif
2029 #ifdef INET
2030 {
2031 if (sc->sc_port) {
2032 m->m_pkthdr.csum_flags = CSUM_UDP;
2033 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2034 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2035 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2036 th->th_sum = htons(0);
2037 } else {
2038 m->m_pkthdr.csum_flags = CSUM_TCP;
2039 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2040 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2041 htons(tlen + optlen - hlen + IPPROTO_TCP));
2042 }
2043 #ifdef TCP_OFFLOAD
2044 if (ADDED_BY_TOE(sc)) {
2045 struct toedev *tod = sc->sc_tod;
2046
2047 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2048
2049 return (error);
2050 }
2051 #endif
2052 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2053 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2054 }
2055 #endif
2056 return (error);
2057 }
2058
2059 static void
syncache_send_challenge_ack(struct syncache * sc,struct mbuf * m)2060 syncache_send_challenge_ack(struct syncache *sc, struct mbuf *m)
2061 {
2062 if (tcp_challenge_ack_check(&sc->sc_challenge_ack_end,
2063 &sc->sc_challenge_ack_cnt)) {
2064 if (syncache_respond(sc, m, TH_ACK) == 0) {
2065 TCPSTAT_INC(tcps_sndacks);
2066 TCPSTAT_INC(tcps_sndtotal);
2067 }
2068 }
2069 }
2070
2071 /*
2072 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2073 * that exceed the capacity of the syncache by avoiding the storage of any
2074 * of the SYNs we receive. Syncookies defend against blind SYN flooding
2075 * attacks where the attacker does not have access to our responses.
2076 *
2077 * Syncookies encode and include all necessary information about the
2078 * connection setup within the SYN|ACK that we send back. That way we
2079 * can avoid keeping any local state until the ACK to our SYN|ACK returns
2080 * (if ever). Normally the syncache and syncookies are running in parallel
2081 * with the latter taking over when the former is exhausted. When matching
2082 * syncache entry is found the syncookie is ignored.
2083 *
2084 * The only reliable information persisting the 3WHS is our initial sequence
2085 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
2086 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2087 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
2088 * returns and signifies a legitimate connection if it matches the ACK.
2089 *
2090 * The available space of 32 bits to store the hash and to encode the SYN
2091 * option information is very tight and we should have at least 24 bits for
2092 * the MAC to keep the number of guesses by blind spoofing reasonably high.
2093 *
2094 * SYN option information we have to encode to fully restore a connection:
2095 * MSS: is imporant to chose an optimal segment size to avoid IP level
2096 * fragmentation along the path. The common MSS values can be encoded
2097 * in a 3-bit table. Uncommon values are captured by the next lower value
2098 * in the table leading to a slight increase in packetization overhead.
2099 * WSCALE: is necessary to allow large windows to be used for high delay-
2100 * bandwidth product links. Not scaling the window when it was initially
2101 * negotiated is bad for performance as lack of scaling further decreases
2102 * the apparent available send window. We only need to encode the WSCALE
2103 * we received from the remote end. Our end can be recalculated at any
2104 * time. The common WSCALE values can be encoded in a 3-bit table.
2105 * Uncommon values are captured by the next lower value in the table
2106 * making us under-estimate the available window size halving our
2107 * theoretically possible maximum throughput for that connection.
2108 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2109 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2110 * that are included in all segments on a connection. We enable them when
2111 * the ACK has them.
2112 *
2113 * Security of syncookies and attack vectors:
2114 *
2115 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2116 * together with the gloabl secret to make it unique per connection attempt.
2117 * Thus any change of any of those parameters results in a different MAC output
2118 * in an unpredictable way unless a collision is encountered. 24 bits of the
2119 * MAC are embedded into the ISS.
2120 *
2121 * To prevent replay attacks two rotating global secrets are updated with a
2122 * new random value every 15 seconds. The life-time of a syncookie is thus
2123 * 15-30 seconds.
2124 *
2125 * Vector 1: Attacking the secret. This requires finding a weakness in the
2126 * MAC itself or the way it is used here. The attacker can do a chosen plain
2127 * text attack by varying and testing the all parameters under his control.
2128 * The strength depends on the size and randomness of the secret, and the
2129 * cryptographic security of the MAC function. Due to the constant updating
2130 * of the secret the attacker has at most 29.999 seconds to find the secret
2131 * and launch spoofed connections. After that he has to start all over again.
2132 *
2133 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
2134 * size an average of 4,823 attempts are required for a 50% chance of success
2135 * to spoof a single syncookie (birthday collision paradox). However the
2136 * attacker is blind and doesn't know if one of his attempts succeeded unless
2137 * he has a side channel to interfere success from. A single connection setup
2138 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2139 * This many attempts are required for each one blind spoofed connection. For
2140 * every additional spoofed connection he has to launch another N attempts.
2141 * Thus for a sustained rate 100 spoofed connections per second approximately
2142 * 1,800,000 packets per second would have to be sent.
2143 *
2144 * NB: The MAC function should be fast so that it doesn't become a CPU
2145 * exhaustion attack vector itself.
2146 *
2147 * References:
2148 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2149 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2150 * http://cr.yp.to/syncookies.html (overview)
2151 * http://cr.yp.to/syncookies/archive (details)
2152 *
2153 *
2154 * Schematic construction of a syncookie enabled Initial Sequence Number:
2155 * 0 1 2 3
2156 * 12345678901234567890123456789012
2157 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2158 *
2159 * x 24 MAC (truncated)
2160 * W 3 Send Window Scale index
2161 * M 3 MSS index
2162 * S 1 SACK permitted
2163 * P 1 Odd/even secret
2164 */
2165
2166 /*
2167 * Distribution and probability of certain MSS values. Those in between are
2168 * rounded down to the next lower one.
2169 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2170 * .2% .3% 5% 7% 7% 20% 15% 45%
2171 */
2172 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2173
2174 /*
2175 * Distribution and probability of certain WSCALE values. We have to map the
2176 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2177 * bits based on prevalence of certain values. Where we don't have an exact
2178 * match for are rounded down to the next lower one letting us under-estimate
2179 * the true available window. At the moment this would happen only for the
2180 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2181 * and window size). The absence of the WSCALE option (no scaling in either
2182 * direction) is encoded with index zero.
2183 * [WSCALE values histograms, Allman, 2012]
2184 * X 10 10 35 5 6 14 10% by host
2185 * X 11 4 5 5 18 49 3% by connections
2186 */
2187 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2188
2189 /*
2190 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
2191 * and good cryptographic properties.
2192 */
2193 static uint32_t
syncookie_mac(struct in_conninfo * inc,tcp_seq irs,uint8_t flags,uint8_t * secbits,uintptr_t secmod)2194 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2195 uint8_t *secbits, uintptr_t secmod)
2196 {
2197 SIPHASH_CTX ctx;
2198 uint32_t siphash[2];
2199
2200 SipHash24_Init(&ctx);
2201 SipHash_SetKey(&ctx, secbits);
2202 switch (inc->inc_flags & INC_ISIPV6) {
2203 #ifdef INET
2204 case 0:
2205 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2206 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2207 break;
2208 #endif
2209 #ifdef INET6
2210 case INC_ISIPV6:
2211 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2212 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2213 break;
2214 #endif
2215 }
2216 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2217 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2218 SipHash_Update(&ctx, &irs, sizeof(irs));
2219 SipHash_Update(&ctx, &flags, sizeof(flags));
2220 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2221 SipHash_Final((u_int8_t *)&siphash, &ctx);
2222
2223 return (siphash[0] ^ siphash[1]);
2224 }
2225
2226 static tcp_seq
syncookie_generate(struct syncache_head * sch,struct syncache * sc)2227 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2228 {
2229 u_int i, secbit, wscale;
2230 uint32_t iss, hash;
2231 uint8_t *secbits;
2232 union syncookie cookie;
2233
2234 cookie.cookie = 0;
2235
2236 /* Map our computed MSS into the 3-bit index. */
2237 for (i = nitems(tcp_sc_msstab) - 1;
2238 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2239 i--)
2240 ;
2241 cookie.flags.mss_idx = i;
2242
2243 /*
2244 * Map the send window scale into the 3-bit index but only if
2245 * the wscale option was received.
2246 */
2247 if (sc->sc_flags & SCF_WINSCALE) {
2248 wscale = sc->sc_requested_s_scale;
2249 for (i = nitems(tcp_sc_wstab) - 1;
2250 tcp_sc_wstab[i] > wscale && i > 0;
2251 i--)
2252 ;
2253 cookie.flags.wscale_idx = i;
2254 }
2255
2256 /* Can we do SACK? */
2257 if (sc->sc_flags & SCF_SACK)
2258 cookie.flags.sack_ok = 1;
2259
2260 /* Which of the two secrets to use. */
2261 secbit = V_tcp_syncache.secret.oddeven & 0x1;
2262 cookie.flags.odd_even = secbit;
2263
2264 secbits = V_tcp_syncache.secret.key[secbit];
2265 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2266 (uintptr_t)sch);
2267
2268 /*
2269 * Put the flags into the hash and XOR them to get better ISS number
2270 * variance. This doesn't enhance the cryptographic strength and is
2271 * done to prevent the 8 cookie bits from showing up directly on the
2272 * wire.
2273 */
2274 iss = hash & ~0xff;
2275 iss |= cookie.cookie ^ (hash >> 24);
2276
2277 TCPSTAT_INC(tcps_sc_sendcookie);
2278 return (iss);
2279 }
2280
2281 static bool
syncookie_expand(struct in_conninfo * inc,const struct syncache_head * sch,struct syncache * sc,struct tcphdr * th,struct tcpopt * to,struct socket * lso,uint16_t port)2282 syncookie_expand(struct in_conninfo *inc, const struct syncache_head *sch,
2283 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2284 struct socket *lso, uint16_t port)
2285 {
2286 uint32_t hash;
2287 uint8_t *secbits;
2288 tcp_seq ack, seq;
2289 int wnd;
2290 union syncookie cookie;
2291
2292 /*
2293 * Pull information out of SYN-ACK/ACK and revert sequence number
2294 * advances.
2295 */
2296 ack = th->th_ack - 1;
2297 seq = th->th_seq - 1;
2298
2299 /*
2300 * Unpack the flags containing enough information to restore the
2301 * connection.
2302 */
2303 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2304
2305 /* Which of the two secrets to use. */
2306 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2307
2308 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2309
2310 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2311 if ((ack & ~0xff) != (hash & ~0xff))
2312 return (false);
2313
2314 /* Fill in the syncache values. */
2315 sc->sc_flags = 0;
2316 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2317 sc->sc_ipopts = NULL;
2318
2319 sc->sc_irs = seq;
2320 sc->sc_iss = ack;
2321
2322 switch (inc->inc_flags & INC_ISIPV6) {
2323 #ifdef INET
2324 case 0:
2325 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2326 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2327 break;
2328 #endif
2329 #ifdef INET6
2330 case INC_ISIPV6:
2331 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2332 sc->sc_flowlabel =
2333 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2334 break;
2335 #endif
2336 }
2337
2338 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2339
2340 /* Only use wscale if it was enabled in the orignal SYN. */
2341 if (cookie.flags.wscale_idx > 0) {
2342 u_int wscale = 0;
2343
2344 /* Recompute the receive window scale that was sent earlier. */
2345 while (wscale < TCP_MAX_WINSHIFT &&
2346 (TCP_MAXWIN << wscale) < sb_max)
2347 wscale++;
2348 sc->sc_requested_r_scale = wscale;
2349 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2350 sc->sc_flags |= SCF_WINSCALE;
2351 }
2352
2353 wnd = lso->sol_sbrcv_hiwat;
2354 wnd = imax(wnd, 0);
2355 wnd = imin(wnd, TCP_MAXWIN);
2356 sc->sc_wnd = wnd;
2357
2358 if (cookie.flags.sack_ok)
2359 sc->sc_flags |= SCF_SACK;
2360
2361 if (to->to_flags & TOF_TS) {
2362 sc->sc_flags |= SCF_TIMESTAMP;
2363 sc->sc_tsreflect = to->to_tsval;
2364 sc->sc_tsoff = tcp_new_ts_offset(inc);
2365 }
2366
2367 if (to->to_flags & TOF_SIGNATURE)
2368 sc->sc_flags |= SCF_SIGNATURE;
2369
2370 sc->sc_rxmits = 0;
2371
2372 sc->sc_port = port;
2373
2374 return (true);
2375 }
2376
2377 #ifdef INVARIANTS
2378 static void
syncookie_cmp(struct in_conninfo * inc,const struct syncache_head * sch,struct syncache * sc,struct tcphdr * th,struct tcpopt * to,struct socket * lso,uint16_t port)2379 syncookie_cmp(struct in_conninfo *inc, const struct syncache_head *sch,
2380 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2381 struct socket *lso, uint16_t port)
2382 {
2383 struct syncache scs;
2384 char *s;
2385
2386 bzero(&scs, sizeof(scs));
2387 if (syncookie_expand(inc, sch, &scs, th, to, lso, port) &&
2388 (sc->sc_peer_mss != scs.sc_peer_mss ||
2389 sc->sc_requested_r_scale != scs.sc_requested_r_scale ||
2390 sc->sc_requested_s_scale != scs.sc_requested_s_scale ||
2391 (sc->sc_flags & SCF_SACK) != (scs.sc_flags & SCF_SACK))) {
2392
2393 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2394 return;
2395
2396 if (sc->sc_peer_mss != scs.sc_peer_mss)
2397 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2398 s, __func__, sc->sc_peer_mss, scs.sc_peer_mss);
2399
2400 if (sc->sc_requested_r_scale != scs.sc_requested_r_scale)
2401 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2402 s, __func__, sc->sc_requested_r_scale,
2403 scs.sc_requested_r_scale);
2404
2405 if (sc->sc_requested_s_scale != scs.sc_requested_s_scale)
2406 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2407 s, __func__, sc->sc_requested_s_scale,
2408 scs.sc_requested_s_scale);
2409
2410 if ((sc->sc_flags & SCF_SACK) != (scs.sc_flags & SCF_SACK))
2411 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2412
2413 free(s, M_TCPLOG);
2414 }
2415 }
2416 #endif /* INVARIANTS */
2417
2418 static void
syncookie_reseed(void * arg)2419 syncookie_reseed(void *arg)
2420 {
2421 struct tcp_syncache *sc = arg;
2422 uint8_t *secbits;
2423 int secbit;
2424
2425 /*
2426 * Reseeding the secret doesn't have to be protected by a lock.
2427 * It only must be ensured that the new random values are visible
2428 * to all CPUs in a SMP environment. The atomic with release
2429 * semantics ensures that.
2430 */
2431 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2432 secbits = sc->secret.key[secbit];
2433 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2434 atomic_add_rel_int(&sc->secret.oddeven, 1);
2435
2436 /* Reschedule ourself. */
2437 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2438 }
2439
2440 /*
2441 * We have overflowed a bucket. Let's pause dealing with the syncache.
2442 * This function will increment the bucketoverflow statistics appropriately
2443 * (once per pause when pausing is enabled; otherwise, once per overflow).
2444 */
2445 static void
syncache_pause(struct in_conninfo * inc)2446 syncache_pause(struct in_conninfo *inc)
2447 {
2448 time_t delta;
2449 const char *s;
2450
2451 /* XXX:
2452 * 2. Add sysctl read here so we don't get the benefit of this
2453 * change without the new sysctl.
2454 */
2455
2456 /*
2457 * Try an unlocked read. If we already know that another thread
2458 * has activated the feature, there is no need to proceed.
2459 */
2460 if (V_tcp_syncache.paused)
2461 return;
2462
2463 /* Are cookied enabled? If not, we can't pause. */
2464 if (!V_tcp_syncookies) {
2465 TCPSTAT_INC(tcps_sc_bucketoverflow);
2466 return;
2467 }
2468
2469 /*
2470 * We may be the first thread to find an overflow. Get the lock
2471 * and evaluate if we need to take action.
2472 */
2473 mtx_lock(&V_tcp_syncache.pause_mtx);
2474 if (V_tcp_syncache.paused) {
2475 mtx_unlock(&V_tcp_syncache.pause_mtx);
2476 return;
2477 }
2478
2479 /* Activate protection. */
2480 V_tcp_syncache.paused = true;
2481 TCPSTAT_INC(tcps_sc_bucketoverflow);
2482
2483 /*
2484 * Determine the last backoff time. If we are seeing a re-newed
2485 * attack within that same time after last reactivating the syncache,
2486 * consider it an extension of the same attack.
2487 */
2488 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2489 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2490 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2491 delta <<= 1;
2492 V_tcp_syncache.pause_backoff++;
2493 }
2494 } else {
2495 delta = TCP_SYNCACHE_PAUSE_TIME;
2496 V_tcp_syncache.pause_backoff = 0;
2497 }
2498
2499 /* Log a warning, including IP addresses, if able. */
2500 if (inc != NULL)
2501 s = tcp_log_addrs(inc, NULL, NULL, NULL);
2502 else
2503 s = (const char *)NULL;
2504 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2505 "the next %lld seconds%s%s%s\n", (long long)delta,
2506 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2507 (s != NULL) ? ")" : "");
2508 free(__DECONST(void *, s), M_TCPLOG);
2509
2510 /* Use the calculated delta to set a new pause time. */
2511 V_tcp_syncache.pause_until = time_uptime + delta;
2512 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2513 &V_tcp_syncache);
2514 mtx_unlock(&V_tcp_syncache.pause_mtx);
2515 }
2516
2517 /* Evaluate whether we need to unpause. */
2518 static void
syncache_unpause(void * arg)2519 syncache_unpause(void *arg)
2520 {
2521 struct tcp_syncache *sc;
2522 time_t delta;
2523
2524 sc = arg;
2525 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2526 callout_deactivate(&sc->pause_co);
2527
2528 /*
2529 * Check to make sure we are not running early. If the pause
2530 * time has expired, then deactivate the protection.
2531 */
2532 if ((delta = sc->pause_until - time_uptime) > 0)
2533 callout_schedule(&sc->pause_co, delta * hz);
2534 else
2535 sc->paused = false;
2536 }
2537
2538 /*
2539 * Exports the syncache entries to userland so that netstat can display
2540 * them alongside the other sockets. This function is intended to be
2541 * called only from tcp_pcblist.
2542 *
2543 * Due to concurrency on an active system, the number of pcbs exported
2544 * may have no relation to max_pcbs. max_pcbs merely indicates the
2545 * amount of space the caller allocated for this function to use.
2546 */
2547 int
syncache_pcblist(struct sysctl_req * req)2548 syncache_pcblist(struct sysctl_req *req)
2549 {
2550 struct xtcpcb xt;
2551 struct syncache *sc;
2552 struct syncache_head *sch;
2553 int error, i;
2554
2555 bzero(&xt, sizeof(xt));
2556 xt.xt_len = sizeof(xt);
2557 xt.t_state = TCPS_SYN_RECEIVED;
2558 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2559 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2560 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2561 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2562
2563 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2564 sch = &V_tcp_syncache.hashbase[i];
2565 SCH_LOCK(sch);
2566 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2567 if (sc->sc_cred != NULL &&
2568 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2569 continue;
2570 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2571 xt.xt_inp.inp_vflag = INP_IPV6;
2572 else
2573 xt.xt_inp.inp_vflag = INP_IPV4;
2574 xt.xt_encaps_port = sc->sc_port;
2575 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2576 sizeof (struct in_conninfo));
2577 error = SYSCTL_OUT(req, &xt, sizeof xt);
2578 if (error) {
2579 SCH_UNLOCK(sch);
2580 return (0);
2581 }
2582 }
2583 SCH_UNLOCK(sch);
2584 }
2585
2586 return (0);
2587 }
2588