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 "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_rss.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 *, int);
125 static void syncache_send_challenge_ack(struct syncache *);
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, 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,uint16_t port)615 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, uint16_t port)
616 {
617 struct syncache *sc;
618 struct syncache_head *sch;
619 char *s = NULL;
620
621 if (syncache_cookiesonly())
622 return;
623 sc = syncache_lookup(inc, &sch); /* returns locked sch */
624 SCH_LOCK_ASSERT(sch);
625
626 /*
627 * No corresponding connection was found in syncache.
628 * If syncookies are enabled and possibly exclusively
629 * used, or we are under memory pressure, a valid RST
630 * may not find a syncache entry. In that case we're
631 * done and no SYN|ACK retransmissions will happen.
632 * Otherwise the RST was misdirected or spoofed.
633 */
634 if (sc == NULL) {
635 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
636 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
637 "syncache entry (possibly syncookie only), "
638 "segment ignored\n", s, __func__);
639 TCPSTAT_INC(tcps_badrst);
640 goto done;
641 }
642
643 /* The remote UDP encaps port does not match. */
644 if (sc->sc_port != port) {
645 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
646 log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
647 "syncache entry but non-matching UDP encaps port, "
648 "segment ignored\n", s, __func__);
649 TCPSTAT_INC(tcps_badrst);
650 goto done;
651 }
652
653 /*
654 * If the RST bit is set, check the sequence number to see
655 * if this is a valid reset segment.
656 *
657 * RFC 793 page 37:
658 * In all states except SYN-SENT, all reset (RST) segments
659 * are validated by checking their SEQ-fields. A reset is
660 * valid if its sequence number is in the window.
661 *
662 * RFC 793 page 69:
663 * There are four cases for the acceptability test for an incoming
664 * segment:
665 *
666 * Segment Receive Test
667 * Length Window
668 * ------- ------- -------------------------------------------
669 * 0 0 SEG.SEQ = RCV.NXT
670 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
671 * >0 0 not acceptable
672 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
673 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
674 *
675 * Note that when receiving a SYN segment in the LISTEN state,
676 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
677 * described in RFC 793, page 66.
678 */
679 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
680 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
681 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
682 if (V_tcp_insecure_rst ||
683 th->th_seq == sc->sc_irs + 1) {
684 syncache_drop(sc, sch);
685 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
686 log(LOG_DEBUG,
687 "%s; %s: Our SYN|ACK was rejected, "
688 "connection attempt aborted by remote "
689 "endpoint\n",
690 s, __func__);
691 TCPSTAT_INC(tcps_sc_reset);
692 } else {
693 TCPSTAT_INC(tcps_badrst);
694 /* Send challenge ACK. */
695 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
696 log(LOG_DEBUG, "%s; %s: RST with invalid "
697 " SEQ %u != NXT %u (+WND %u), "
698 "sending challenge ACK\n",
699 s, __func__,
700 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
701 syncache_send_challenge_ack(sc);
702 }
703 } else {
704 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
705 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
706 "NXT %u (+WND %u), segment ignored\n",
707 s, __func__,
708 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
709 TCPSTAT_INC(tcps_badrst);
710 }
711
712 done:
713 if (s != NULL)
714 free(s, M_TCPLOG);
715 SCH_UNLOCK(sch);
716 }
717
718 void
syncache_unreach(struct in_conninfo * inc,tcp_seq th_seq,uint16_t port)719 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
720 {
721 struct syncache *sc;
722 struct syncache_head *sch;
723
724 if (syncache_cookiesonly())
725 return;
726 sc = syncache_lookup(inc, &sch); /* returns locked sch */
727 SCH_LOCK_ASSERT(sch);
728 if (sc == NULL)
729 goto done;
730
731 /* If the port != sc_port, then it's a bogus ICMP msg */
732 if (port != sc->sc_port)
733 goto done;
734
735 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
736 if (ntohl(th_seq) != sc->sc_iss)
737 goto done;
738
739 /*
740 * If we've retransmitted 3 times and this is our second error,
741 * we remove the entry. Otherwise, we allow it to continue on.
742 * This prevents us from incorrectly nuking an entry during a
743 * spurious network outage.
744 *
745 * See tcp_notify().
746 */
747 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
748 sc->sc_flags |= SCF_UNREACH;
749 goto done;
750 }
751 syncache_drop(sc, sch);
752 TCPSTAT_INC(tcps_sc_unreach);
753 done:
754 SCH_UNLOCK(sch);
755 }
756
757 /*
758 * Build a new TCP socket structure from a syncache entry.
759 *
760 * On success return the newly created socket with its underlying inp locked.
761 */
762 static struct socket *
syncache_socket(struct syncache * sc,struct socket * lso,struct mbuf * m)763 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
764 {
765 struct inpcb *inp = NULL;
766 struct socket *so;
767 struct tcpcb *tp;
768 int error;
769 char *s;
770
771 NET_EPOCH_ASSERT();
772
773 /*
774 * Ok, create the full blown connection, and set things up
775 * as they would have been set up if we had created the
776 * connection when the SYN arrived.
777 */
778 if ((so = solisten_clone(lso)) == NULL)
779 goto allocfail;
780 #ifdef MAC
781 mac_socketpeer_set_from_mbuf(m, so);
782 #endif
783 error = in_pcballoc(so, &V_tcbinfo);
784 if (error) {
785 sodealloc(so);
786 goto allocfail;
787 }
788 inp = sotoinpcb(so);
789 if ((tp = tcp_newtcpcb(inp, sototcpcb(lso))) == NULL) {
790 in_pcbfree(inp);
791 sodealloc(so);
792 goto allocfail;
793 }
794 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
795 #ifdef INET6
796 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
797 inp->inp_vflag &= ~INP_IPV4;
798 inp->inp_vflag |= INP_IPV6;
799 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
800 } else {
801 inp->inp_vflag &= ~INP_IPV6;
802 inp->inp_vflag |= INP_IPV4;
803 #endif
804 inp->inp_ip_ttl = sc->sc_ip_ttl;
805 inp->inp_ip_tos = sc->sc_ip_tos;
806 inp->inp_laddr = sc->sc_inc.inc_laddr;
807 #ifdef INET6
808 }
809 #endif
810 inp->inp_lport = sc->sc_inc.inc_lport;
811 #ifdef INET6
812 if (inp->inp_vflag & INP_IPV6PROTO) {
813 struct inpcb *oinp = sotoinpcb(lso);
814
815 /*
816 * Inherit socket options from the listening socket.
817 * Note that in6p_inputopts are not (and should not be)
818 * copied, since it stores previously received options and is
819 * used to detect if each new option is different than the
820 * previous one and hence should be passed to a user.
821 * If we copied in6p_inputopts, a user would not be able to
822 * receive options just after calling the accept system call.
823 */
824 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
825 if (oinp->in6p_outputopts)
826 inp->in6p_outputopts =
827 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
828 inp->in6p_hops = oinp->in6p_hops;
829 }
830
831 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
832 struct sockaddr_in6 sin6;
833
834 sin6.sin6_family = AF_INET6;
835 sin6.sin6_len = sizeof(sin6);
836 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
837 sin6.sin6_port = sc->sc_inc.inc_fport;
838 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
839 INP_HASH_WLOCK(&V_tcbinfo);
840 error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
841 INP_HASH_WUNLOCK(&V_tcbinfo);
842 if (error != 0)
843 goto abort;
844 /* Override flowlabel from in6_pcbconnect. */
845 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
846 inp->inp_flow |= sc->sc_flowlabel;
847 }
848 #endif /* INET6 */
849 #if defined(INET) && defined(INET6)
850 else
851 #endif
852 #ifdef INET
853 {
854 struct sockaddr_in sin;
855
856 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
857
858 if (inp->inp_options == NULL) {
859 inp->inp_options = sc->sc_ipopts;
860 sc->sc_ipopts = NULL;
861 }
862
863 sin.sin_family = AF_INET;
864 sin.sin_len = sizeof(sin);
865 sin.sin_addr = sc->sc_inc.inc_faddr;
866 sin.sin_port = sc->sc_inc.inc_fport;
867 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
868 INP_HASH_WLOCK(&V_tcbinfo);
869 error = in_pcbconnect(inp, &sin, thread0.td_ucred);
870 INP_HASH_WUNLOCK(&V_tcbinfo);
871 if (error != 0)
872 goto abort;
873 }
874 #endif /* INET */
875 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
876 /* Copy old policy into new socket's. */
877 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
878 printf("syncache_socket: could not copy policy\n");
879 #endif
880 if (sc->sc_flowtype != M_HASHTYPE_NONE) {
881 inp->inp_flowid = sc->sc_flowid;
882 inp->inp_flowtype = sc->sc_flowtype;
883 #ifdef RSS
884 } else {
885 /* assign flowid by software RSS hash */
886 #ifdef INET6
887 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
888 rss_proto_software_hash_v6(&inp->in6p_faddr,
889 &inp->in6p_laddr,
890 inp->inp_fport,
891 inp->inp_lport,
892 IPPROTO_TCP,
893 &inp->inp_flowid,
894 &inp->inp_flowtype);
895 } else
896 #endif /* INET6 */
897 {
898 rss_proto_software_hash_v4(inp->inp_faddr,
899 inp->inp_laddr,
900 inp->inp_fport,
901 inp->inp_lport,
902 IPPROTO_TCP,
903 &inp->inp_flowid,
904 &inp->inp_flowtype);
905 }
906 #endif /* RSS */
907 }
908 #ifdef NUMA
909 inp->inp_numa_domain = sc->sc_numa_domain;
910 #endif
911
912 tp->t_state = TCPS_SYN_RECEIVED;
913 tp->iss = sc->sc_iss;
914 tp->irs = sc->sc_irs;
915 tp->t_port = sc->sc_port;
916 tcp_rcvseqinit(tp);
917 tcp_sendseqinit(tp);
918 tp->snd_wl1 = sc->sc_irs;
919 tp->snd_max = tp->iss + 1;
920 tp->snd_nxt = tp->iss + 1;
921 tp->rcv_up = sc->sc_irs + 1;
922 tp->rcv_wnd = sc->sc_wnd;
923 tp->rcv_adv += tp->rcv_wnd;
924 tp->last_ack_sent = tp->rcv_nxt;
925
926 tp->t_flags = sototcpcb(lso)->t_flags &
927 (TF_LRD|TF_NOPUSH|TF_NODELAY);
928 if (sc->sc_flags & SCF_NOOPT)
929 tp->t_flags |= TF_NOOPT;
930 else {
931 if (sc->sc_flags & SCF_WINSCALE) {
932 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
933 tp->snd_scale = sc->sc_requested_s_scale;
934 tp->request_r_scale = sc->sc_requested_r_scale;
935 }
936 if (sc->sc_flags & SCF_TIMESTAMP) {
937 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
938 tp->ts_recent = sc->sc_tsreflect;
939 tp->ts_recent_age = tcp_ts_getticks();
940 tp->ts_offset = sc->sc_tsoff;
941 }
942 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
943 if (sc->sc_flags & SCF_SIGNATURE)
944 tp->t_flags |= TF_SIGNATURE;
945 #endif
946 if (sc->sc_flags & SCF_SACK)
947 tp->t_flags |= TF_SACK_PERMIT;
948 }
949
950 tcp_ecn_syncache_socket(tp, sc);
951
952 /*
953 * Set up MSS and get cached values from tcp_hostcache.
954 * This might overwrite some of the defaults we just set.
955 */
956 tcp_mss(tp, sc->sc_peer_mss);
957
958 /*
959 * If the SYN,ACK was retransmitted, indicate that CWND to be
960 * limited to one segment in cc_conn_init().
961 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
962 */
963 if (sc->sc_rxmits > 1)
964 tp->snd_cwnd = 1;
965
966 /* Copy over the challenge ACK state. */
967 tp->t_challenge_ack_end = sc->sc_challenge_ack_end;
968 tp->t_challenge_ack_cnt = sc->sc_challenge_ack_cnt;
969
970 #ifdef TCP_OFFLOAD
971 /*
972 * Allow a TOE driver to install its hooks. Note that we hold the
973 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
974 * new connection before the TOE driver has done its thing.
975 */
976 if (ADDED_BY_TOE(sc)) {
977 struct toedev *tod = sc->sc_tod;
978
979 tod->tod_offload_socket(tod, sc->sc_todctx, so);
980 }
981 #endif
982 #ifdef TCP_BLACKBOX
983 /*
984 * Inherit the log state from the listening socket, if
985 * - the log state of the listening socket is not off and
986 * - the listening socket was not auto selected from all sessions and
987 * - a log id is not set on the listening socket.
988 * This avoids inheriting a log state which was automatically set.
989 */
990 if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
991 ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
992 (sototcpcb(lso)->t_lib == NULL)) {
993 tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
994 }
995 #endif
996 /*
997 * Copy and activate timers.
998 */
999 tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
1000 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1001 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1002 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1003 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1004 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1005
1006 TCPSTAT_INC(tcps_accepts);
1007 TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1008
1009 if (!solisten_enqueue(so, SS_ISCONNECTED))
1010 tp->t_flags |= TF_SONOTCONN;
1011 /* Can we inherit anything from the listener? */
1012 if (tp->t_fb->tfb_inherit != NULL) {
1013 (*tp->t_fb->tfb_inherit)(tp, sotoinpcb(lso));
1014 }
1015 return (so);
1016
1017 allocfail:
1018 /*
1019 * Drop the connection; we will either send a RST or have the peer
1020 * retransmit its SYN again after its RTO and try again.
1021 */
1022 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1023 log(LOG_DEBUG, "%s; %s: Socket create failed "
1024 "due to limits or memory shortage\n",
1025 s, __func__);
1026 free(s, M_TCPLOG);
1027 }
1028 TCPSTAT_INC(tcps_listendrop);
1029 return (NULL);
1030
1031 abort:
1032 tcp_discardcb(tp);
1033 in_pcbfree(inp);
1034 sodealloc(so);
1035 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1036 log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1037 s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1038 error);
1039 free(s, M_TCPLOG);
1040 }
1041 TCPSTAT_INC(tcps_listendrop);
1042 return (NULL);
1043 }
1044
1045 /*
1046 * This function gets called when we receive an ACK for a
1047 * socket in the LISTEN state. We look up the connection
1048 * in the syncache, and if its there, we pull it out of
1049 * the cache and turn it into a full-blown connection in
1050 * the SYN-RECEIVED state.
1051 *
1052 * On syncache_socket() success the newly created socket
1053 * has its underlying inp locked.
1054 *
1055 * *lsop is updated, if and only if 1 is returned.
1056 */
1057 int
syncache_expand(struct in_conninfo * inc,struct tcpopt * to,struct tcphdr * th,struct socket ** lsop,struct mbuf * m,uint16_t port)1058 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1059 struct socket **lsop, struct mbuf *m, uint16_t port)
1060 {
1061 struct syncache *sc;
1062 struct syncache_head *sch;
1063 struct syncache scs;
1064 char *s;
1065 bool locked;
1066
1067 NET_EPOCH_ASSERT();
1068 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1069 ("%s: can handle only ACK", __func__));
1070
1071 if (syncache_cookiesonly()) {
1072 sc = NULL;
1073 sch = syncache_hashbucket(inc);
1074 locked = false;
1075 } else {
1076 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1077 locked = true;
1078 SCH_LOCK_ASSERT(sch);
1079 }
1080
1081 #ifdef INVARIANTS
1082 /*
1083 * Test code for syncookies comparing the syncache stored
1084 * values with the reconstructed values from the cookie.
1085 */
1086 if (sc != NULL)
1087 syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1088 #endif
1089
1090 if (sc == NULL) {
1091 if (locked) {
1092 /*
1093 * The syncache is currently in use (neither disabled,
1094 * nor paused), but no entry was found.
1095 */
1096 if (!V_tcp_syncookies) {
1097 /*
1098 * Since no syncookies are used in case of
1099 * a bucket overflow, don't even check for
1100 * a valid syncookie.
1101 */
1102 SCH_UNLOCK(sch);
1103 TCPSTAT_INC(tcps_sc_spurcookie);
1104 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1105 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1106 "segment rejected "
1107 "(syncookies disabled)\n",
1108 s, __func__);
1109 free(s, M_TCPLOG);
1110 }
1111 return (0);
1112 }
1113 if (sch->sch_last_overflow <
1114 time_uptime - SYNCOOKIE_LIFETIME) {
1115 /*
1116 * Since the bucket did not overflow recently,
1117 * don't even check for a valid syncookie.
1118 */
1119 SCH_UNLOCK(sch);
1120 TCPSTAT_INC(tcps_sc_spurcookie);
1121 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1122 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1123 "segment rejected "
1124 "(no syncache entry)\n",
1125 s, __func__);
1126 free(s, M_TCPLOG);
1127 }
1128 return (0);
1129 }
1130 SCH_UNLOCK(sch);
1131 }
1132 bzero(&scs, sizeof(scs));
1133 /*
1134 * Now check, if the syncookie is valid. If it is, create an on
1135 * stack syncache entry.
1136 */
1137 if (syncookie_expand(inc, sch, &scs, th, to, *lsop, port)) {
1138 sc = &scs;
1139 TCPSTAT_INC(tcps_sc_recvcookie);
1140 } else {
1141 TCPSTAT_INC(tcps_sc_failcookie);
1142 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1143 log(LOG_DEBUG, "%s; %s: Segment failed "
1144 "SYNCOOKIE authentication, segment rejected "
1145 "(probably spoofed)\n", s, __func__);
1146 free(s, M_TCPLOG);
1147 }
1148 return (0);
1149 }
1150 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1151 /* If received ACK has MD5 signature, check it. */
1152 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1153 (!TCPMD5_ENABLED() ||
1154 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1155 /* Drop the ACK. */
1156 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1157 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1158 "MD5 signature doesn't match.\n",
1159 s, __func__);
1160 free(s, M_TCPLOG);
1161 }
1162 TCPSTAT_INC(tcps_sig_err_sigopt);
1163 return (-1); /* Do not send RST */
1164 }
1165 #endif /* TCP_SIGNATURE */
1166 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
1167 sc->sc_flowid = m->m_pkthdr.flowid;
1168 sc->sc_flowtype = M_HASHTYPE_GET(m);
1169 }
1170 #ifdef NUMA
1171 sc->sc_numa_domain = m ? m->m_pkthdr.numa_domain : M_NODOM;
1172 #endif
1173 TCPSTATES_INC(TCPS_SYN_RECEIVED);
1174 } else {
1175 if (sc->sc_port != port) {
1176 SCH_UNLOCK(sch);
1177 return (0);
1178 }
1179 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1180 /*
1181 * If listening socket requested TCP digests, check that
1182 * received ACK has signature and it is correct.
1183 * If not, drop the ACK and leave sc entry in the cache,
1184 * because SYN was received with correct signature.
1185 */
1186 if (sc->sc_flags & SCF_SIGNATURE) {
1187 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1188 /* No signature */
1189 TCPSTAT_INC(tcps_sig_err_nosigopt);
1190 SCH_UNLOCK(sch);
1191 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1192 log(LOG_DEBUG, "%s; %s: Segment "
1193 "rejected, MD5 signature wasn't "
1194 "provided.\n", s, __func__);
1195 free(s, M_TCPLOG);
1196 }
1197 return (-1); /* Do not send RST */
1198 }
1199 if (!TCPMD5_ENABLED() ||
1200 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1201 /* Doesn't match or no SA */
1202 SCH_UNLOCK(sch);
1203 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1204 log(LOG_DEBUG, "%s; %s: Segment "
1205 "rejected, MD5 signature doesn't "
1206 "match.\n", s, __func__);
1207 free(s, M_TCPLOG);
1208 }
1209 return (-1); /* Do not send RST */
1210 }
1211 }
1212 #endif /* TCP_SIGNATURE */
1213
1214 /*
1215 * RFC 7323 PAWS: If we have a timestamp on this segment and
1216 * it's less than ts_recent, drop it.
1217 * XXXMT: RFC 7323 also requires to send an ACK.
1218 * In tcp_input.c this is only done for TCP segments
1219 * with user data, so be consistent here and just drop
1220 * the segment.
1221 */
1222 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1223 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1224 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1225 log(LOG_DEBUG,
1226 "%s; %s: SEG.TSval %u < TS.Recent %u, "
1227 "segment dropped\n", s, __func__,
1228 to->to_tsval, sc->sc_tsreflect);
1229 }
1230 SCH_UNLOCK(sch);
1231 free(s, M_TCPLOG);
1232 return (-1); /* Do not send RST */
1233 }
1234
1235 /*
1236 * If timestamps were not negotiated during SYN/ACK and a
1237 * segment with a timestamp is received, ignore the
1238 * timestamp and process the packet normally.
1239 * See section 3.2 of RFC 7323.
1240 */
1241 if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1242 (to->to_flags & TOF_TS)) {
1243 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1244 log(LOG_DEBUG, "%s; %s: Timestamp not "
1245 "expected, segment processed normally\n",
1246 s, __func__);
1247 free(s, M_TCPLOG);
1248 }
1249 }
1250
1251 /*
1252 * If timestamps were negotiated during SYN/ACK and a
1253 * segment without a timestamp is received, silently drop
1254 * the segment, unless the missing timestamps are tolerated.
1255 * See section 3.2 of RFC 7323.
1256 */
1257 if ((sc->sc_flags & SCF_TIMESTAMP) &&
1258 !(to->to_flags & TOF_TS)) {
1259 if (V_tcp_tolerate_missing_ts) {
1260 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1261 log(LOG_DEBUG,
1262 "%s; %s: Timestamp missing, "
1263 "segment processed normally\n",
1264 s, __func__);
1265 free(s, M_TCPLOG);
1266 }
1267 } else {
1268 SCH_UNLOCK(sch);
1269 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1270 log(LOG_DEBUG,
1271 "%s; %s: Timestamp missing, "
1272 "segment silently dropped\n",
1273 s, __func__);
1274 free(s, M_TCPLOG);
1275 }
1276 return (-1); /* Do not send RST */
1277 }
1278 }
1279
1280 /*
1281 * SEG.SEQ validation:
1282 * The SEG.SEQ must be in the window starting at our
1283 * initial receive sequence number + 1.
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, "
1289 "sending challenge ACK\n",
1290 s, __func__, th->th_seq, sc->sc_irs + 1);
1291 syncache_send_challenge_ack(sc);
1292 SCH_UNLOCK(sch);
1293 free(s, M_TCPLOG);
1294 return (-1); /* Do not send RST */
1295 }
1296
1297 /*
1298 * SEG.ACK validation:
1299 * SEG.ACK must match our initial send sequence number + 1.
1300 */
1301 if (th->th_ack != sc->sc_iss + 1) {
1302 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1303 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, "
1304 "segment rejected\n",
1305 s, __func__, th->th_ack, sc->sc_iss + 1);
1306 SCH_UNLOCK(sch);
1307 free(s, M_TCPLOG);
1308 return (0); /* Do send RST, do not free sc. */
1309 }
1310
1311 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1312 sch->sch_length--;
1313 #ifdef TCP_OFFLOAD
1314 if (ADDED_BY_TOE(sc)) {
1315 struct toedev *tod = sc->sc_tod;
1316
1317 tod->tod_syncache_removed(tod, sc->sc_todctx);
1318 }
1319 #endif
1320 SCH_UNLOCK(sch);
1321 }
1322
1323 *lsop = syncache_socket(sc, *lsop, m);
1324
1325 if (__predict_false(*lsop == NULL)) {
1326 TCPSTAT_INC(tcps_sc_aborted);
1327 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1328 } else if (sc != &scs)
1329 TCPSTAT_INC(tcps_sc_completed);
1330
1331 if (sc != &scs)
1332 syncache_free(sc);
1333 return (1);
1334 }
1335
1336 static struct socket *
syncache_tfo_expand(struct syncache * sc,struct socket * lso,struct mbuf * m,uint64_t response_cookie)1337 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1338 uint64_t response_cookie)
1339 {
1340 struct inpcb *inp;
1341 struct tcpcb *tp;
1342 unsigned int *pending_counter;
1343 struct socket *so;
1344
1345 NET_EPOCH_ASSERT();
1346
1347 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1348 so = syncache_socket(sc, lso, m);
1349 if (so == NULL) {
1350 TCPSTAT_INC(tcps_sc_aborted);
1351 atomic_subtract_int(pending_counter, 1);
1352 } else {
1353 soisconnected(so);
1354 inp = sotoinpcb(so);
1355 tp = intotcpcb(inp);
1356 tp->t_flags |= TF_FASTOPEN;
1357 tp->t_tfo_cookie.server = response_cookie;
1358 tp->snd_max = tp->iss;
1359 tp->snd_nxt = tp->iss;
1360 tp->t_tfo_pending = pending_counter;
1361 TCPSTATES_INC(TCPS_SYN_RECEIVED);
1362 TCPSTAT_INC(tcps_sc_completed);
1363 }
1364
1365 return (so);
1366 }
1367
1368 /*
1369 * Given a LISTEN socket and an inbound SYN request, add
1370 * this to the syn cache, and send back a segment:
1371 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1372 * to the source.
1373 *
1374 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1375 * Doing so would require that we hold onto the data and deliver it
1376 * to the application. However, if we are the target of a SYN-flood
1377 * DoS attack, an attacker could send data which would eventually
1378 * consume all available buffer space if it were ACKed. By not ACKing
1379 * the data, we avoid this DoS scenario.
1380 *
1381 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1382 * cookie is processed and a new socket is created. In this case, any data
1383 * accompanying the SYN will be queued to the socket by tcp_input() and will
1384 * be ACKed either when the application sends response data or the delayed
1385 * ACK timer expires, whichever comes first.
1386 */
1387 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)1388 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1389 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1390 void *todctx, uint8_t iptos, uint16_t port)
1391 {
1392 struct tcpcb *tp;
1393 struct socket *rv = NULL;
1394 struct syncache *sc = NULL;
1395 struct ucred *cred;
1396 struct syncache_head *sch;
1397 struct mbuf *ipopts = NULL;
1398 u_int ltflags;
1399 int win, ip_ttl, ip_tos;
1400 char *s;
1401 #ifdef INET6
1402 int autoflowlabel = 0;
1403 #endif
1404 #ifdef MAC
1405 struct label *maclabel = NULL;
1406 #endif
1407 struct syncache scs;
1408 uint64_t tfo_response_cookie;
1409 unsigned int *tfo_pending = NULL;
1410 int tfo_cookie_valid = 0;
1411 int tfo_response_cookie_valid = 0;
1412 bool locked;
1413
1414 INP_RLOCK_ASSERT(inp); /* listen socket */
1415 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1416 ("%s: unexpected tcp flags", __func__));
1417
1418 /*
1419 * Combine all so/tp operations very early to drop the INP lock as
1420 * soon as possible.
1421 */
1422 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1423 tp = sototcpcb(so);
1424 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1425
1426 #ifdef INET6
1427 if (inc->inc_flags & INC_ISIPV6) {
1428 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1429 autoflowlabel = 1;
1430 }
1431 ip_ttl = in6_selecthlim(inp, NULL);
1432 if ((inp->in6p_outputopts == NULL) ||
1433 (inp->in6p_outputopts->ip6po_tclass == -1)) {
1434 ip_tos = 0;
1435 } else {
1436 ip_tos = inp->in6p_outputopts->ip6po_tclass;
1437 }
1438 }
1439 #endif
1440 #if defined(INET6) && defined(INET)
1441 else
1442 #endif
1443 #ifdef INET
1444 {
1445 ip_ttl = inp->inp_ip_ttl;
1446 ip_tos = inp->inp_ip_tos;
1447 }
1448 #endif
1449 win = so->sol_sbrcv_hiwat;
1450 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1451
1452 if (V_tcp_fastopen_server_enable && (tp->t_flags & TF_FASTOPEN) &&
1453 (tp->t_tfo_pending != NULL) &&
1454 (to->to_flags & TOF_FASTOPEN)) {
1455 /*
1456 * Limit the number of pending TFO connections to
1457 * approximately half of the queue limit. This prevents TFO
1458 * SYN floods from starving the service by filling the
1459 * listen queue with bogus TFO connections.
1460 */
1461 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1462 (so->sol_qlimit / 2)) {
1463 int result;
1464
1465 result = tcp_fastopen_check_cookie(inc,
1466 to->to_tfo_cookie, to->to_tfo_len,
1467 &tfo_response_cookie);
1468 tfo_cookie_valid = (result > 0);
1469 tfo_response_cookie_valid = (result >= 0);
1470 }
1471
1472 /*
1473 * Remember the TFO pending counter as it will have to be
1474 * decremented below if we don't make it to syncache_tfo_expand().
1475 */
1476 tfo_pending = tp->t_tfo_pending;
1477 }
1478
1479 #ifdef MAC
1480 if (mac_syncache_init(&maclabel) != 0) {
1481 INP_RUNLOCK(inp);
1482 goto done;
1483 } else
1484 mac_syncache_create(maclabel, inp);
1485 #endif
1486 if (!tfo_cookie_valid)
1487 INP_RUNLOCK(inp);
1488
1489 /*
1490 * Remember the IP options, if any.
1491 */
1492 #ifdef INET6
1493 if (!(inc->inc_flags & INC_ISIPV6))
1494 #endif
1495 #ifdef INET
1496 ipopts = (m) ? ip_srcroute(m) : NULL;
1497 #else
1498 ipopts = NULL;
1499 #endif
1500
1501 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1502 /*
1503 * When the socket is TCP-MD5 enabled check that,
1504 * - a signed packet is valid
1505 * - a non-signed packet does not have a security association
1506 *
1507 * If a signed packet fails validation or a non-signed packet has a
1508 * security association, the packet will be dropped.
1509 */
1510 if (ltflags & TF_SIGNATURE) {
1511 if (to->to_flags & TOF_SIGNATURE) {
1512 if (!TCPMD5_ENABLED() ||
1513 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1514 goto done;
1515 } else {
1516 if (TCPMD5_ENABLED() &&
1517 TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1518 goto done;
1519 }
1520 } else if (to->to_flags & TOF_SIGNATURE)
1521 goto done;
1522 #endif /* TCP_SIGNATURE */
1523 /*
1524 * See if we already have an entry for this connection.
1525 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1526 *
1527 * XXX: should the syncache be re-initialized with the contents
1528 * of the new SYN here (which may have different options?)
1529 *
1530 * XXX: We do not check the sequence number to see if this is a
1531 * real retransmit or a new connection attempt. The question is
1532 * how to handle such a case; either ignore it as spoofed, or
1533 * drop the current entry and create a new one?
1534 */
1535 if (syncache_cookiesonly()) {
1536 sc = NULL;
1537 sch = syncache_hashbucket(inc);
1538 locked = false;
1539 } else {
1540 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1541 locked = true;
1542 SCH_LOCK_ASSERT(sch);
1543 }
1544 if (sc != NULL) {
1545 if (tfo_cookie_valid)
1546 INP_RUNLOCK(inp);
1547 TCPSTAT_INC(tcps_sc_dupsyn);
1548 if (ipopts) {
1549 /*
1550 * If we were remembering a previous source route,
1551 * forget it and use the new one we've been given.
1552 */
1553 if (sc->sc_ipopts)
1554 (void)m_free(sc->sc_ipopts);
1555 sc->sc_ipopts = ipopts;
1556 }
1557 /*
1558 * Update timestamp if present.
1559 */
1560 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1561 sc->sc_tsreflect = to->to_tsval;
1562 else
1563 sc->sc_flags &= ~SCF_TIMESTAMP;
1564 /*
1565 * Adjust ECN response if needed, e.g. different
1566 * IP ECN field, or a fallback by the remote host.
1567 */
1568 if (sc->sc_flags & SCF_ECN_MASK) {
1569 sc->sc_flags &= ~SCF_ECN_MASK;
1570 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1571 }
1572 #ifdef MAC
1573 /*
1574 * Since we have already unconditionally allocated label
1575 * storage, free it up. The syncache entry will already
1576 * have an initialized label we can use.
1577 */
1578 mac_syncache_destroy(&maclabel);
1579 #endif
1580 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1581 /* Retransmit SYN|ACK and reset retransmit count. */
1582 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1583 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1584 "resetting timer and retransmitting SYN|ACK\n",
1585 s, __func__);
1586 free(s, M_TCPLOG);
1587 }
1588 if (syncache_respond(sc, TH_SYN|TH_ACK) == 0) {
1589 sc->sc_rxmits = 0;
1590 syncache_timeout(sc, sch, 1);
1591 TCPSTAT_INC(tcps_sndacks);
1592 TCPSTAT_INC(tcps_sndtotal);
1593 } else {
1594 /*
1595 * Most likely we are memory constrained, so free
1596 * resources.
1597 */
1598 syncache_drop(sc, sch);
1599 TCPSTAT_INC(tcps_sc_dropped);
1600 }
1601 SCH_UNLOCK(sch);
1602 goto donenoprobe;
1603 }
1604
1605 KASSERT(sc == NULL, ("sc(%p) != NULL", sc));
1606 /*
1607 * Skip allocating a syncache entry if we are just going to discard
1608 * it later.
1609 */
1610 if (!locked || tfo_cookie_valid) {
1611 bzero(&scs, sizeof(scs));
1612 sc = &scs;
1613 } else {
1614 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1615 if (sc == NULL) {
1616 /*
1617 * The zone allocator couldn't provide more entries.
1618 * Treat this as if the cache was full; drop the oldest
1619 * entry and insert the new one.
1620 */
1621 TCPSTAT_INC(tcps_sc_zonefail);
1622 sc = TAILQ_LAST(&sch->sch_bucket, sch_head);
1623 if (sc != NULL) {
1624 sch->sch_last_overflow = time_uptime;
1625 syncache_drop(sc, sch);
1626 syncache_pause(inc);
1627 }
1628 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1629 if (sc == NULL) {
1630 if (V_tcp_syncookies) {
1631 bzero(&scs, sizeof(scs));
1632 sc = &scs;
1633 } else {
1634 KASSERT(locked,
1635 ("%s: bucket unexpectedly unlocked",
1636 __func__));
1637 SCH_UNLOCK(sch);
1638 goto done;
1639 }
1640 }
1641 }
1642 }
1643
1644 KASSERT(sc != NULL, ("sc == NULL"));
1645 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1646 sc->sc_tfo_cookie = &tfo_response_cookie;
1647
1648 /*
1649 * Fill in the syncache values.
1650 */
1651 #ifdef MAC
1652 sc->sc_label = maclabel;
1653 #endif
1654 /*
1655 * sc_cred is only used in syncache_pcblist() to list TCP endpoints in
1656 * TCPS_SYN_RECEIVED state when V_tcp_syncache.see_other is false.
1657 * Therefore, store the credentials only when needed:
1658 * - sc is allocated from the zone and not using the on stack instance.
1659 * - the sysctl variable net.inet.tcp.syncache.see_other is false.
1660 * The reference count is decremented when a zone allocated sc is
1661 * freed in syncache_free().
1662 */
1663 if (sc != &scs && !V_tcp_syncache.see_other) {
1664 sc->sc_cred = cred;
1665 cred = NULL;
1666 } else
1667 sc->sc_cred = NULL;
1668 sc->sc_port = port;
1669 sc->sc_ipopts = ipopts;
1670 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1671 sc->sc_ip_tos = ip_tos;
1672 sc->sc_ip_ttl = ip_ttl;
1673 #ifdef TCP_OFFLOAD
1674 sc->sc_tod = tod;
1675 sc->sc_todctx = todctx;
1676 #endif
1677 sc->sc_irs = th->th_seq;
1678 sc->sc_flags = 0;
1679 sc->sc_flowlabel = 0;
1680
1681 /*
1682 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1683 * win was derived from socket earlier in the function.
1684 */
1685 win = imax(win, 0);
1686 win = imin(win, TCP_MAXWIN);
1687 sc->sc_wnd = win;
1688
1689 if (V_tcp_do_rfc1323 &&
1690 !(ltflags & TF_NOOPT)) {
1691 /*
1692 * A timestamp received in a SYN makes
1693 * it ok to send timestamp requests and replies.
1694 */
1695 if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1696 sc->sc_tsreflect = to->to_tsval;
1697 sc->sc_flags |= SCF_TIMESTAMP;
1698 sc->sc_tsoff = tcp_new_ts_offset(inc);
1699 }
1700 if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1701 u_int wscale = 0;
1702
1703 /*
1704 * Pick the smallest possible scaling factor that
1705 * will still allow us to scale up to sb_max, aka
1706 * kern.ipc.maxsockbuf.
1707 *
1708 * We do this because there are broken firewalls that
1709 * will corrupt the window scale option, leading to
1710 * the other endpoint believing that our advertised
1711 * window is unscaled. At scale factors larger than
1712 * 5 the unscaled window will drop below 1500 bytes,
1713 * leading to serious problems when traversing these
1714 * broken firewalls.
1715 *
1716 * With the default maxsockbuf of 256K, a scale factor
1717 * of 3 will be chosen by this algorithm. Those who
1718 * choose a larger maxsockbuf should watch out
1719 * for the compatibility problems mentioned above.
1720 *
1721 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1722 * or <SYN,ACK>) segment itself is never scaled.
1723 */
1724 while (wscale < TCP_MAX_WINSHIFT &&
1725 (TCP_MAXWIN << wscale) < sb_max)
1726 wscale++;
1727 sc->sc_requested_r_scale = wscale;
1728 sc->sc_requested_s_scale = to->to_wscale;
1729 sc->sc_flags |= SCF_WINSCALE;
1730 }
1731 }
1732 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1733 /*
1734 * If incoming packet has an MD5 signature, flag this in the
1735 * syncache so that syncache_respond() will do the right thing
1736 * with the SYN+ACK.
1737 */
1738 if (to->to_flags & TOF_SIGNATURE)
1739 sc->sc_flags |= SCF_SIGNATURE;
1740 #endif /* TCP_SIGNATURE */
1741 if (to->to_flags & TOF_SACKPERM)
1742 sc->sc_flags |= SCF_SACK;
1743 if (to->to_flags & TOF_MSS)
1744 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1745 if (ltflags & TF_NOOPT)
1746 sc->sc_flags |= SCF_NOOPT;
1747 /* ECN Handshake */
1748 if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0)
1749 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1750
1751 if (V_tcp_syncookies || V_tcp_syncookiesonly)
1752 sc->sc_iss = syncookie_generate(sch, sc);
1753 else
1754 sc->sc_iss = arc4random();
1755 #ifdef INET6
1756 if (autoflowlabel) {
1757 if (V_tcp_syncookies || V_tcp_syncookiesonly)
1758 sc->sc_flowlabel = sc->sc_iss;
1759 else
1760 sc->sc_flowlabel = ip6_randomflowlabel();
1761 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1762 }
1763 #endif
1764 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
1765 sc->sc_flowid = m->m_pkthdr.flowid;
1766 sc->sc_flowtype = M_HASHTYPE_GET(m);
1767 }
1768 #ifdef NUMA
1769 sc->sc_numa_domain = m ? m->m_pkthdr.numa_domain : M_NODOM;
1770 #endif
1771 if (locked)
1772 SCH_UNLOCK(sch);
1773
1774 if (tfo_cookie_valid) {
1775 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1776 /* INP_RUNLOCK(inp) will be performed by the caller */
1777 goto tfo_expanded;
1778 }
1779
1780 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1781 /*
1782 * Do a standard 3-way handshake.
1783 */
1784 if (syncache_respond(sc, TH_SYN|TH_ACK) == 0) {
1785 if (sc != &scs)
1786 syncache_insert(sc, sch); /* locks and unlocks sch */
1787 TCPSTAT_INC(tcps_sndacks);
1788 TCPSTAT_INC(tcps_sndtotal);
1789 } else {
1790 /*
1791 * Most likely we are memory constrained, so free resources.
1792 */
1793 if (sc != &scs)
1794 syncache_free(sc);
1795 TCPSTAT_INC(tcps_sc_dropped);
1796 }
1797 goto donenoprobe;
1798
1799 done:
1800 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1801 donenoprobe:
1802 if (m)
1803 m_freem(m);
1804 /*
1805 * If tfo_pending is not NULL here, then a TFO SYN that did not
1806 * result in a new socket was processed and the associated pending
1807 * counter has not yet been decremented. All such TFO processing paths
1808 * transit this point.
1809 */
1810 if (tfo_pending != NULL)
1811 tcp_fastopen_decrement_counter(tfo_pending);
1812
1813 tfo_expanded:
1814 if (cred != NULL)
1815 crfree(cred);
1816 if (sc == NULL || sc == &scs) {
1817 #ifdef MAC
1818 mac_syncache_destroy(&maclabel);
1819 #endif
1820 if (ipopts)
1821 (void)m_free(ipopts);
1822 }
1823 return (rv);
1824 }
1825
1826 /*
1827 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment
1828 * or upon 3WHS ACK timeout.
1829 */
1830 static int
syncache_respond(struct syncache * sc,int flags)1831 syncache_respond(struct syncache *sc, int flags)
1832 {
1833 struct ip *ip = NULL;
1834 struct mbuf *m;
1835 struct tcphdr *th = NULL;
1836 struct udphdr *udp = NULL;
1837 int optlen, error = 0; /* Make compiler happy */
1838 u_int16_t hlen, tlen, mssopt, ulen;
1839 struct tcpopt to;
1840 #ifdef INET6
1841 struct ip6_hdr *ip6 = NULL;
1842 #endif
1843
1844 NET_EPOCH_ASSERT();
1845
1846 hlen =
1847 #ifdef INET6
1848 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1849 #endif
1850 sizeof(struct ip);
1851 tlen = hlen + sizeof(struct tcphdr);
1852 if (sc->sc_port) {
1853 tlen += sizeof(struct udphdr);
1854 }
1855 /* Determine MSS we advertize to other end of connection. */
1856 mssopt = tcp_mssopt(&sc->sc_inc);
1857 if (sc->sc_port)
1858 mssopt -= V_tcp_udp_tunneling_overhead;
1859 mssopt = max(mssopt, V_tcp_minmss);
1860
1861 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1862 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1863 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1864 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1865 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1866
1867 /* Create the IP+TCP header from scratch. */
1868 m = m_gethdr(M_NOWAIT, MT_DATA);
1869 if (m == NULL)
1870 return (ENOBUFS);
1871 #ifdef MAC
1872 mac_syncache_create_mbuf(sc->sc_label, m);
1873 #endif
1874 m->m_data += max_linkhdr;
1875 m->m_len = tlen;
1876 m->m_pkthdr.len = tlen;
1877 m->m_pkthdr.rcvif = NULL;
1878
1879 #ifdef INET6
1880 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1881 ip6 = mtod(m, struct ip6_hdr *);
1882 ip6->ip6_vfc = IPV6_VERSION;
1883 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1884 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1885 ip6->ip6_plen = htons(tlen - hlen);
1886 /* ip6_hlim is set after checksum */
1887 /* Zero out traffic class and flow label. */
1888 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1889 ip6->ip6_flow |= sc->sc_flowlabel;
1890 if (sc->sc_port != 0) {
1891 ip6->ip6_nxt = IPPROTO_UDP;
1892 udp = (struct udphdr *)(ip6 + 1);
1893 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1894 udp->uh_dport = sc->sc_port;
1895 ulen = (tlen - sizeof(struct ip6_hdr));
1896 th = (struct tcphdr *)(udp + 1);
1897 } else {
1898 ip6->ip6_nxt = IPPROTO_TCP;
1899 th = (struct tcphdr *)(ip6 + 1);
1900 }
1901 ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1902 }
1903 #endif
1904 #if defined(INET6) && defined(INET)
1905 else
1906 #endif
1907 #ifdef INET
1908 {
1909 ip = mtod(m, struct ip *);
1910 ip->ip_v = IPVERSION;
1911 ip->ip_hl = sizeof(struct ip) >> 2;
1912 ip->ip_len = htons(tlen);
1913 ip->ip_id = 0;
1914 ip->ip_off = 0;
1915 ip->ip_sum = 0;
1916 ip->ip_src = sc->sc_inc.inc_laddr;
1917 ip->ip_dst = sc->sc_inc.inc_faddr;
1918 ip->ip_ttl = sc->sc_ip_ttl;
1919 ip->ip_tos = sc->sc_ip_tos;
1920
1921 /*
1922 * See if we should do MTU discovery. Route lookups are
1923 * expensive, so we will only unset the DF bit if:
1924 *
1925 * 1) path_mtu_discovery is disabled
1926 * 2) the SCF_UNREACH flag has been set
1927 */
1928 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1929 ip->ip_off |= htons(IP_DF);
1930 if (sc->sc_port == 0) {
1931 ip->ip_p = IPPROTO_TCP;
1932 th = (struct tcphdr *)(ip + 1);
1933 } else {
1934 ip->ip_p = IPPROTO_UDP;
1935 udp = (struct udphdr *)(ip + 1);
1936 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1937 udp->uh_dport = sc->sc_port;
1938 ulen = (tlen - sizeof(struct ip));
1939 th = (struct tcphdr *)(udp + 1);
1940 }
1941 }
1942 #endif /* INET */
1943 th->th_sport = sc->sc_inc.inc_lport;
1944 th->th_dport = sc->sc_inc.inc_fport;
1945
1946 if (flags & TH_SYN)
1947 th->th_seq = htonl(sc->sc_iss);
1948 else
1949 th->th_seq = htonl(sc->sc_iss + 1);
1950 th->th_ack = htonl(sc->sc_irs + 1);
1951 th->th_off = sizeof(struct tcphdr) >> 2;
1952 th->th_win = htons(sc->sc_wnd);
1953 th->th_urp = 0;
1954
1955 flags = tcp_ecn_syncache_respond(flags, sc);
1956 tcp_set_flags(th, flags);
1957
1958 /* Tack on the TCP options. */
1959 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1960 to.to_flags = 0;
1961
1962 if (flags & TH_SYN) {
1963 to.to_mss = mssopt;
1964 to.to_flags = TOF_MSS;
1965 if (sc->sc_flags & SCF_WINSCALE) {
1966 to.to_wscale = sc->sc_requested_r_scale;
1967 to.to_flags |= TOF_SCALE;
1968 }
1969 if (sc->sc_flags & SCF_SACK)
1970 to.to_flags |= TOF_SACKPERM;
1971 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1972 if (sc->sc_flags & SCF_SIGNATURE)
1973 to.to_flags |= TOF_SIGNATURE;
1974 #endif
1975 if (sc->sc_tfo_cookie) {
1976 to.to_flags |= TOF_FASTOPEN;
1977 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1978 to.to_tfo_cookie = sc->sc_tfo_cookie;
1979 /* don't send cookie again when retransmitting response */
1980 sc->sc_tfo_cookie = NULL;
1981 }
1982 }
1983 if (sc->sc_flags & SCF_TIMESTAMP) {
1984 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1985 to.to_tsecr = sc->sc_tsreflect;
1986 to.to_flags |= TOF_TS;
1987 }
1988 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1989
1990 /* Adjust headers by option size. */
1991 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1992 m->m_len += optlen;
1993 m->m_pkthdr.len += optlen;
1994 #ifdef INET6
1995 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1996 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1997 else
1998 #endif
1999 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
2000 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
2001 if (sc->sc_flags & SCF_SIGNATURE) {
2002 KASSERT(to.to_flags & TOF_SIGNATURE,
2003 ("tcp_addoptions() didn't set tcp_signature"));
2004
2005 /* NOTE: to.to_signature is inside of mbuf */
2006 if (!TCPMD5_ENABLED() ||
2007 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
2008 m_freem(m);
2009 return (EACCES);
2010 }
2011 }
2012 #endif
2013 } else
2014 optlen = 0;
2015
2016 if (udp) {
2017 ulen += optlen;
2018 udp->uh_ulen = htons(ulen);
2019 }
2020 M_SETFIB(m, sc->sc_inc.inc_fibnum);
2021 m->m_pkthdr.flowid = sc->sc_flowid;
2022 M_HASHTYPE_SET(m, sc->sc_flowtype);
2023 #ifdef NUMA
2024 m->m_pkthdr.numa_domain = sc->sc_numa_domain;
2025 #endif
2026 #ifdef INET6
2027 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2028 if (sc->sc_port) {
2029 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2030 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2031 udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2032 IPPROTO_UDP, 0);
2033 th->th_sum = htons(0);
2034 } else {
2035 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2036 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2037 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2038 IPPROTO_TCP, 0);
2039 }
2040 ip6->ip6_hlim = sc->sc_ip_ttl;
2041 #ifdef TCP_OFFLOAD
2042 if (ADDED_BY_TOE(sc)) {
2043 struct toedev *tod = sc->sc_tod;
2044
2045 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2046
2047 return (error);
2048 }
2049 #endif
2050 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2051 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2052 }
2053 #endif
2054 #if defined(INET6) && defined(INET)
2055 else
2056 #endif
2057 #ifdef INET
2058 {
2059 if (sc->sc_port) {
2060 m->m_pkthdr.csum_flags = CSUM_UDP;
2061 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2062 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2063 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2064 th->th_sum = htons(0);
2065 } else {
2066 m->m_pkthdr.csum_flags = CSUM_TCP;
2067 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2068 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2069 htons(tlen + optlen - hlen + IPPROTO_TCP));
2070 }
2071 #ifdef TCP_OFFLOAD
2072 if (ADDED_BY_TOE(sc)) {
2073 struct toedev *tod = sc->sc_tod;
2074
2075 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2076
2077 return (error);
2078 }
2079 #endif
2080 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2081 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2082 }
2083 #endif
2084 return (error);
2085 }
2086
2087 static void
syncache_send_challenge_ack(struct syncache * sc)2088 syncache_send_challenge_ack(struct syncache *sc)
2089 {
2090 if (tcp_challenge_ack_check(&sc->sc_challenge_ack_end,
2091 &sc->sc_challenge_ack_cnt)) {
2092 if (syncache_respond(sc, TH_ACK) == 0) {
2093 TCPSTAT_INC(tcps_sndacks);
2094 TCPSTAT_INC(tcps_sndtotal);
2095 }
2096 }
2097 }
2098
2099 /*
2100 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2101 * that exceed the capacity of the syncache by avoiding the storage of any
2102 * of the SYNs we receive. Syncookies defend against blind SYN flooding
2103 * attacks where the attacker does not have access to our responses.
2104 *
2105 * Syncookies encode and include all necessary information about the
2106 * connection setup within the SYN|ACK that we send back. That way we
2107 * can avoid keeping any local state until the ACK to our SYN|ACK returns
2108 * (if ever). Normally the syncache and syncookies are running in parallel
2109 * with the latter taking over when the former is exhausted. When matching
2110 * syncache entry is found the syncookie is ignored.
2111 *
2112 * The only reliable information persisting the 3WHS is our initial sequence
2113 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
2114 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2115 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
2116 * returns and signifies a legitimate connection if it matches the ACK.
2117 *
2118 * The available space of 32 bits to store the hash and to encode the SYN
2119 * option information is very tight and we should have at least 24 bits for
2120 * the MAC to keep the number of guesses by blind spoofing reasonably high.
2121 *
2122 * SYN option information we have to encode to fully restore a connection:
2123 * MSS: is imporant to chose an optimal segment size to avoid IP level
2124 * fragmentation along the path. The common MSS values can be encoded
2125 * in a 3-bit table. Uncommon values are captured by the next lower value
2126 * in the table leading to a slight increase in packetization overhead.
2127 * WSCALE: is necessary to allow large windows to be used for high delay-
2128 * bandwidth product links. Not scaling the window when it was initially
2129 * negotiated is bad for performance as lack of scaling further decreases
2130 * the apparent available send window. We only need to encode the WSCALE
2131 * we received from the remote end. Our end can be recalculated at any
2132 * time. The common WSCALE values can be encoded in a 3-bit table.
2133 * Uncommon values are captured by the next lower value in the table
2134 * making us under-estimate the available window size halving our
2135 * theoretically possible maximum throughput for that connection.
2136 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2137 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2138 * that are included in all segments on a connection. We enable them when
2139 * the ACK has them.
2140 *
2141 * Security of syncookies and attack vectors:
2142 *
2143 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2144 * together with the gloabl secret to make it unique per connection attempt.
2145 * Thus any change of any of those parameters results in a different MAC output
2146 * in an unpredictable way unless a collision is encountered. 24 bits of the
2147 * MAC are embedded into the ISS.
2148 *
2149 * To prevent replay attacks two rotating global secrets are updated with a
2150 * new random value every 15 seconds. The life-time of a syncookie is thus
2151 * 15-30 seconds.
2152 *
2153 * Vector 1: Attacking the secret. This requires finding a weakness in the
2154 * MAC itself or the way it is used here. The attacker can do a chosen plain
2155 * text attack by varying and testing the all parameters under his control.
2156 * The strength depends on the size and randomness of the secret, and the
2157 * cryptographic security of the MAC function. Due to the constant updating
2158 * of the secret the attacker has at most 29.999 seconds to find the secret
2159 * and launch spoofed connections. After that he has to start all over again.
2160 *
2161 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
2162 * size an average of 4,823 attempts are required for a 50% chance of success
2163 * to spoof a single syncookie (birthday collision paradox). However the
2164 * attacker is blind and doesn't know if one of his attempts succeeded unless
2165 * he has a side channel to interfere success from. A single connection setup
2166 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2167 * This many attempts are required for each one blind spoofed connection. For
2168 * every additional spoofed connection he has to launch another N attempts.
2169 * Thus for a sustained rate 100 spoofed connections per second approximately
2170 * 1,800,000 packets per second would have to be sent.
2171 *
2172 * NB: The MAC function should be fast so that it doesn't become a CPU
2173 * exhaustion attack vector itself.
2174 *
2175 * References:
2176 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2177 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2178 * http://cr.yp.to/syncookies.html (overview)
2179 * http://cr.yp.to/syncookies/archive (details)
2180 *
2181 *
2182 * Schematic construction of a syncookie enabled Initial Sequence Number:
2183 * 0 1 2 3
2184 * 12345678901234567890123456789012
2185 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2186 *
2187 * x 24 MAC (truncated)
2188 * W 3 Send Window Scale index
2189 * M 3 MSS index
2190 * S 1 SACK permitted
2191 * P 1 Odd/even secret
2192 */
2193
2194 /*
2195 * Distribution and probability of certain MSS values. Those in between are
2196 * rounded down to the next lower one.
2197 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2198 * .2% .3% 5% 7% 7% 20% 15% 45%
2199 */
2200 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2201
2202 /*
2203 * Distribution and probability of certain WSCALE values. We have to map the
2204 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2205 * bits based on prevalence of certain values. Where we don't have an exact
2206 * match for are rounded down to the next lower one letting us under-estimate
2207 * the true available window. At the moment this would happen only for the
2208 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2209 * and window size). The absence of the WSCALE option (no scaling in either
2210 * direction) is encoded with index zero.
2211 * [WSCALE values histograms, Allman, 2012]
2212 * X 10 10 35 5 6 14 10% by host
2213 * X 11 4 5 5 18 49 3% by connections
2214 */
2215 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2216
2217 /*
2218 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
2219 * and good cryptographic properties.
2220 */
2221 static uint32_t
syncookie_mac(struct in_conninfo * inc,tcp_seq irs,uint8_t flags,uint8_t * secbits,uintptr_t secmod)2222 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2223 uint8_t *secbits, uintptr_t secmod)
2224 {
2225 SIPHASH_CTX ctx;
2226 uint32_t siphash[2];
2227
2228 SipHash24_Init(&ctx);
2229 SipHash_SetKey(&ctx, secbits);
2230 switch (inc->inc_flags & INC_ISIPV6) {
2231 #ifdef INET
2232 case 0:
2233 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2234 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2235 break;
2236 #endif
2237 #ifdef INET6
2238 case INC_ISIPV6:
2239 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2240 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2241 break;
2242 #endif
2243 }
2244 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2245 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2246 SipHash_Update(&ctx, &irs, sizeof(irs));
2247 SipHash_Update(&ctx, &flags, sizeof(flags));
2248 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2249 SipHash_Final((u_int8_t *)&siphash, &ctx);
2250
2251 return (siphash[0] ^ siphash[1]);
2252 }
2253
2254 static tcp_seq
syncookie_generate(struct syncache_head * sch,struct syncache * sc)2255 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2256 {
2257 u_int i, secbit, wscale;
2258 uint32_t iss, hash;
2259 uint8_t *secbits;
2260 union syncookie cookie;
2261
2262 cookie.cookie = 0;
2263
2264 /* Map our computed MSS into the 3-bit index. */
2265 for (i = nitems(tcp_sc_msstab) - 1;
2266 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2267 i--)
2268 ;
2269 cookie.flags.mss_idx = i;
2270
2271 /*
2272 * Map the send window scale into the 3-bit index but only if
2273 * the wscale option was received.
2274 */
2275 if (sc->sc_flags & SCF_WINSCALE) {
2276 wscale = sc->sc_requested_s_scale;
2277 for (i = nitems(tcp_sc_wstab) - 1;
2278 tcp_sc_wstab[i] > wscale && i > 0;
2279 i--)
2280 ;
2281 cookie.flags.wscale_idx = i;
2282 }
2283
2284 /* Can we do SACK? */
2285 if (sc->sc_flags & SCF_SACK)
2286 cookie.flags.sack_ok = 1;
2287
2288 /* Which of the two secrets to use. */
2289 secbit = V_tcp_syncache.secret.oddeven & 0x1;
2290 cookie.flags.odd_even = secbit;
2291
2292 secbits = V_tcp_syncache.secret.key[secbit];
2293 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2294 (uintptr_t)sch);
2295
2296 /*
2297 * Put the flags into the hash and XOR them to get better ISS number
2298 * variance. This doesn't enhance the cryptographic strength and is
2299 * done to prevent the 8 cookie bits from showing up directly on the
2300 * wire.
2301 */
2302 iss = hash & ~0xff;
2303 iss |= cookie.cookie ^ (hash >> 24);
2304
2305 TCPSTAT_INC(tcps_sc_sendcookie);
2306 return (iss);
2307 }
2308
2309 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)2310 syncookie_expand(struct in_conninfo *inc, const struct syncache_head *sch,
2311 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2312 struct socket *lso, uint16_t port)
2313 {
2314 uint32_t hash;
2315 uint8_t *secbits;
2316 tcp_seq ack, seq;
2317 int wnd;
2318 union syncookie cookie;
2319
2320 /*
2321 * Pull information out of SYN-ACK/ACK and revert sequence number
2322 * advances.
2323 */
2324 ack = th->th_ack - 1;
2325 seq = th->th_seq - 1;
2326
2327 /*
2328 * Unpack the flags containing enough information to restore the
2329 * connection.
2330 */
2331 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2332
2333 /* Which of the two secrets to use. */
2334 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2335
2336 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2337
2338 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2339 if ((ack & ~0xff) != (hash & ~0xff))
2340 return (false);
2341
2342 /* Fill in the syncache values. */
2343 sc->sc_flags = 0;
2344 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2345 sc->sc_ipopts = NULL;
2346
2347 sc->sc_irs = seq;
2348 sc->sc_iss = ack;
2349
2350 switch (inc->inc_flags & INC_ISIPV6) {
2351 #ifdef INET
2352 case 0:
2353 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2354 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2355 break;
2356 #endif
2357 #ifdef INET6
2358 case INC_ISIPV6:
2359 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2360 sc->sc_flowlabel =
2361 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2362 break;
2363 #endif
2364 }
2365
2366 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2367
2368 /* Only use wscale if it was enabled in the orignal SYN. */
2369 if (cookie.flags.wscale_idx > 0) {
2370 u_int wscale = 0;
2371
2372 /* Recompute the receive window scale that was sent earlier. */
2373 while (wscale < TCP_MAX_WINSHIFT &&
2374 (TCP_MAXWIN << wscale) < sb_max)
2375 wscale++;
2376 sc->sc_requested_r_scale = wscale;
2377 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2378 sc->sc_flags |= SCF_WINSCALE;
2379 }
2380
2381 wnd = lso->sol_sbrcv_hiwat;
2382 wnd = imax(wnd, 0);
2383 wnd = imin(wnd, TCP_MAXWIN);
2384 sc->sc_wnd = wnd;
2385
2386 if (cookie.flags.sack_ok)
2387 sc->sc_flags |= SCF_SACK;
2388
2389 if (to->to_flags & TOF_TS) {
2390 sc->sc_flags |= SCF_TIMESTAMP;
2391 sc->sc_tsreflect = to->to_tsval;
2392 sc->sc_tsoff = tcp_new_ts_offset(inc);
2393 }
2394
2395 if (to->to_flags & TOF_SIGNATURE)
2396 sc->sc_flags |= SCF_SIGNATURE;
2397
2398 sc->sc_rxmits = 0;
2399
2400 sc->sc_port = port;
2401
2402 return (true);
2403 }
2404
2405 #ifdef INVARIANTS
2406 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)2407 syncookie_cmp(struct in_conninfo *inc, const struct syncache_head *sch,
2408 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2409 struct socket *lso, uint16_t port)
2410 {
2411 struct syncache scs;
2412 char *s;
2413
2414 bzero(&scs, sizeof(scs));
2415 if (syncookie_expand(inc, sch, &scs, th, to, lso, port) &&
2416 (sc->sc_peer_mss != scs.sc_peer_mss ||
2417 sc->sc_requested_r_scale != scs.sc_requested_r_scale ||
2418 sc->sc_requested_s_scale != scs.sc_requested_s_scale ||
2419 (sc->sc_flags & SCF_SACK) != (scs.sc_flags & SCF_SACK))) {
2420
2421 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2422 return;
2423
2424 if (sc->sc_peer_mss != scs.sc_peer_mss)
2425 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2426 s, __func__, sc->sc_peer_mss, scs.sc_peer_mss);
2427
2428 if (sc->sc_requested_r_scale != scs.sc_requested_r_scale)
2429 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2430 s, __func__, sc->sc_requested_r_scale,
2431 scs.sc_requested_r_scale);
2432
2433 if (sc->sc_requested_s_scale != scs.sc_requested_s_scale)
2434 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2435 s, __func__, sc->sc_requested_s_scale,
2436 scs.sc_requested_s_scale);
2437
2438 if ((sc->sc_flags & SCF_SACK) != (scs.sc_flags & SCF_SACK))
2439 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2440
2441 free(s, M_TCPLOG);
2442 }
2443 }
2444 #endif /* INVARIANTS */
2445
2446 static void
syncookie_reseed(void * arg)2447 syncookie_reseed(void *arg)
2448 {
2449 struct tcp_syncache *sc = arg;
2450 uint8_t *secbits;
2451 int secbit;
2452
2453 /*
2454 * Reseeding the secret doesn't have to be protected by a lock.
2455 * It only must be ensured that the new random values are visible
2456 * to all CPUs in a SMP environment. The atomic with release
2457 * semantics ensures that.
2458 */
2459 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2460 secbits = sc->secret.key[secbit];
2461 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2462 atomic_add_rel_int(&sc->secret.oddeven, 1);
2463
2464 /* Reschedule ourself. */
2465 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2466 }
2467
2468 /*
2469 * We have overflowed a bucket. Let's pause dealing with the syncache.
2470 * This function will increment the bucketoverflow statistics appropriately
2471 * (once per pause when pausing is enabled; otherwise, once per overflow).
2472 */
2473 static void
syncache_pause(struct in_conninfo * inc)2474 syncache_pause(struct in_conninfo *inc)
2475 {
2476 time_t delta;
2477 const char *s;
2478
2479 /* XXX:
2480 * 2. Add sysctl read here so we don't get the benefit of this
2481 * change without the new sysctl.
2482 */
2483
2484 /*
2485 * Try an unlocked read. If we already know that another thread
2486 * has activated the feature, there is no need to proceed.
2487 */
2488 if (V_tcp_syncache.paused)
2489 return;
2490
2491 /* Are cookied enabled? If not, we can't pause. */
2492 if (!V_tcp_syncookies) {
2493 TCPSTAT_INC(tcps_sc_bucketoverflow);
2494 return;
2495 }
2496
2497 /*
2498 * We may be the first thread to find an overflow. Get the lock
2499 * and evaluate if we need to take action.
2500 */
2501 mtx_lock(&V_tcp_syncache.pause_mtx);
2502 if (V_tcp_syncache.paused) {
2503 mtx_unlock(&V_tcp_syncache.pause_mtx);
2504 return;
2505 }
2506
2507 /* Activate protection. */
2508 V_tcp_syncache.paused = true;
2509 TCPSTAT_INC(tcps_sc_bucketoverflow);
2510
2511 /*
2512 * Determine the last backoff time. If we are seeing a re-newed
2513 * attack within that same time after last reactivating the syncache,
2514 * consider it an extension of the same attack.
2515 */
2516 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2517 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2518 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2519 delta <<= 1;
2520 V_tcp_syncache.pause_backoff++;
2521 }
2522 } else {
2523 delta = TCP_SYNCACHE_PAUSE_TIME;
2524 V_tcp_syncache.pause_backoff = 0;
2525 }
2526
2527 /* Log a warning, including IP addresses, if able. */
2528 if (inc != NULL)
2529 s = tcp_log_addrs(inc, NULL, NULL, NULL);
2530 else
2531 s = (const char *)NULL;
2532 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2533 "the next %lld seconds%s%s%s\n", (long long)delta,
2534 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2535 (s != NULL) ? ")" : "");
2536 free(__DECONST(void *, s), M_TCPLOG);
2537
2538 /* Use the calculated delta to set a new pause time. */
2539 V_tcp_syncache.pause_until = time_uptime + delta;
2540 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2541 &V_tcp_syncache);
2542 mtx_unlock(&V_tcp_syncache.pause_mtx);
2543 }
2544
2545 /* Evaluate whether we need to unpause. */
2546 static void
syncache_unpause(void * arg)2547 syncache_unpause(void *arg)
2548 {
2549 struct tcp_syncache *sc;
2550 time_t delta;
2551
2552 sc = arg;
2553 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2554 callout_deactivate(&sc->pause_co);
2555
2556 /*
2557 * Check to make sure we are not running early. If the pause
2558 * time has expired, then deactivate the protection.
2559 */
2560 if ((delta = sc->pause_until - time_uptime) > 0)
2561 callout_schedule(&sc->pause_co, delta * hz);
2562 else
2563 sc->paused = false;
2564 }
2565
2566 /*
2567 * Exports the syncache entries to userland so that netstat can display
2568 * them alongside the other sockets. This function is intended to be
2569 * called only from tcp_pcblist.
2570 *
2571 * Due to concurrency on an active system, the number of pcbs exported
2572 * may have no relation to max_pcbs. max_pcbs merely indicates the
2573 * amount of space the caller allocated for this function to use.
2574 */
2575 int
syncache_pcblist(struct sysctl_req * req)2576 syncache_pcblist(struct sysctl_req *req)
2577 {
2578 struct xtcpcb xt;
2579 struct syncache *sc;
2580 struct syncache_head *sch;
2581 int error, i;
2582
2583 bzero(&xt, sizeof(xt));
2584 xt.xt_len = sizeof(xt);
2585 xt.t_state = TCPS_SYN_RECEIVED;
2586 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2587 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2588 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2589 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2590
2591 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2592 sch = &V_tcp_syncache.hashbase[i];
2593 SCH_LOCK(sch);
2594 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2595 if (sc->sc_cred != NULL &&
2596 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2597 continue;
2598 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2599 xt.xt_inp.inp_vflag = INP_IPV6;
2600 else
2601 xt.xt_inp.inp_vflag = INP_IPV4;
2602 xt.xt_encaps_port = sc->sc_port;
2603 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2604 sizeof (struct in_conninfo));
2605 error = SYSCTL_OUT(req, &xt, sizeof xt);
2606 if (error) {
2607 SCH_UNLOCK(sch);
2608 return (0);
2609 }
2610 }
2611 SCH_UNLOCK(sch);
2612 }
2613
2614 return (0);
2615 }
2616