1 /*-
2 * Copyright (c) 2001 McAfee, Inc.
3 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
4 * All rights reserved.
5 *
6 * This software was developed for the FreeBSD Project by Jonathan Lemon
7 * and McAfee Research, the Security Research Division of McAfee, Inc. under
8 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
9 * DARPA CHATS research program. [2001 McAfee, Inc.]
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 */
32
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_pcbgroup.h"
40
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/hash.h>
44 #include <sys/refcount.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/limits.h>
48 #include <sys/lock.h>
49 #include <sys/mutex.h>
50 #include <sys/malloc.h>
51 #include <sys/mbuf.h>
52 #include <sys/proc.h> /* for proc0 declaration */
53 #include <sys/random.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/syslog.h>
57 #include <sys/ucred.h>
58
59 #include <sys/md5.h>
60 #include <crypto/siphash/siphash.h>
61
62 #include <vm/uma.h>
63
64 #include <net/if.h>
65 #include <net/if_var.h>
66 #include <net/route.h>
67 #include <net/vnet.h>
68
69 #include <netinet/in.h>
70 #include <netinet/in_kdtrace.h>
71 #include <netinet/in_systm.h>
72 #include <netinet/ip.h>
73 #include <netinet/in_var.h>
74 #include <netinet/in_pcb.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/ip_options.h>
77 #ifdef INET6
78 #include <netinet/ip6.h>
79 #include <netinet/icmp6.h>
80 #include <netinet6/nd6.h>
81 #include <netinet6/ip6_var.h>
82 #include <netinet6/in6_pcb.h>
83 #endif
84 #include <netinet/tcp.h>
85 #ifdef TCP_RFC7413
86 #include <netinet/tcp_fastopen.h>
87 #endif
88 #include <netinet/tcp_fsm.h>
89 #include <netinet/tcp_seq.h>
90 #include <netinet/tcp_timer.h>
91 #include <netinet/tcp_var.h>
92 #include <netinet/tcp_syncache.h>
93 #ifdef INET6
94 #include <netinet6/tcp6_var.h>
95 #endif
96 #ifdef TCP_OFFLOAD
97 #include <netinet/toecore.h>
98 #endif
99
100 #include <netipsec/ipsec_support.h>
101
102 #include <machine/in_cksum.h>
103
104 #include <security/mac/mac_framework.h>
105
106 static VNET_DEFINE(int, tcp_syncookies) = 1;
107 #define V_tcp_syncookies VNET(tcp_syncookies)
108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
109 &VNET_NAME(tcp_syncookies), 0,
110 "Use TCP SYN cookies if the syncache overflows");
111
112 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
115 &VNET_NAME(tcp_syncookiesonly), 0,
116 "Use only TCP SYN cookies");
117
118 #ifdef TCP_OFFLOAD
119 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
120 #endif
121
122 static void syncache_drop(struct syncache *, struct syncache_head *);
123 static void syncache_free(struct syncache *);
124 static void syncache_insert(struct syncache *, struct syncache_head *);
125 static int syncache_respond(struct syncache *, struct syncache_head *, int,
126 const struct mbuf *);
127 static struct socket *syncache_socket(struct syncache *, struct socket *,
128 struct mbuf *m);
129 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
130 int docallout);
131 static void syncache_timer(void *);
132
133 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
134 uint8_t *, uintptr_t);
135 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
136 static struct syncache
137 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
138 struct syncache *, struct tcphdr *, struct tcpopt *,
139 struct socket *);
140 static void syncookie_reseed(void *);
141 #ifdef INVARIANTS
142 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
143 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
144 struct socket *lso);
145 #endif
146
147 /*
148 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
149 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
150 * the odds are that the user has given up attempting to connect by then.
151 */
152 #define SYNCACHE_MAXREXMTS 3
153
154 /* Arbitrary values */
155 #define TCP_SYNCACHE_HASHSIZE 512
156 #define TCP_SYNCACHE_BUCKETLIMIT 30
157
158 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
159 #define V_tcp_syncache VNET(tcp_syncache)
160
161 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
162 "TCP SYN cache");
163
164 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
165 &VNET_NAME(tcp_syncache.bucket_limit), 0,
166 "Per-bucket hash limit for syncache");
167
168 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
169 &VNET_NAME(tcp_syncache.cache_limit), 0,
170 "Overall entry limit for syncache");
171
172 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
173 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
174
175 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
176 &VNET_NAME(tcp_syncache.hashsize), 0,
177 "Size of TCP syncache hashtable");
178
179 static int
180 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
181 {
182 int error;
183 u_int new;
184
185 new = V_tcp_syncache.rexmt_limit;
186 error = sysctl_handle_int(oidp, &new, 0, req);
187 if ((error == 0) && (req->newptr != NULL)) {
188 if (new > TCP_MAXRXTSHIFT)
189 error = EINVAL;
190 else
191 V_tcp_syncache.rexmt_limit = new;
192 }
193 return (error);
194 }
195
196 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
197 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
198 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
199 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
200 "Limit on SYN/ACK retransmissions");
201
202 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
203 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
204 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
205 "Send reset on socket allocation failure");
206
207 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
208
209 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
210 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
211 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
212
213 /*
214 * Requires the syncache entry to be already removed from the bucket list.
215 */
216 static void
217 syncache_free(struct syncache *sc)
218 {
219
220 if (sc->sc_ipopts)
221 (void) m_free(sc->sc_ipopts);
222 if (sc->sc_cred)
223 crfree(sc->sc_cred);
224 #ifdef MAC
225 mac_syncache_destroy(&sc->sc_label);
226 #endif
227
228 uma_zfree(V_tcp_syncache.zone, sc);
229 }
230
231 void
232 syncache_init(void)
233 {
234 int i;
235
236 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
237 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
238 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
239 V_tcp_syncache.hash_secret = arc4random();
240
241 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
242 &V_tcp_syncache.hashsize);
243 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
244 &V_tcp_syncache.bucket_limit);
245 if (!powerof2(V_tcp_syncache.hashsize) ||
246 V_tcp_syncache.hashsize == 0) {
247 printf("WARNING: syncache hash size is not a power of 2.\n");
248 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
249 }
250 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
251
252 /* Set limits. */
253 V_tcp_syncache.cache_limit =
254 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
255 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
256 &V_tcp_syncache.cache_limit);
257
258 /* Allocate the hash table. */
259 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
260 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
261
262 #ifdef VIMAGE
263 V_tcp_syncache.vnet = curvnet;
264 #endif
265
266 /* Initialize the hash buckets. */
267 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
268 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
269 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
270 NULL, MTX_DEF);
271 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
272 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
273 V_tcp_syncache.hashbase[i].sch_length = 0;
274 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
275 V_tcp_syncache.hashbase[i].sch_last_overflow =
276 -(SYNCOOKIE_LIFETIME + 1);
277 }
278
279 /* Create the syncache entry zone. */
280 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
281 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
282 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
283 V_tcp_syncache.cache_limit);
284
285 /* Start the SYN cookie reseeder callout. */
286 callout_init(&V_tcp_syncache.secret.reseed, 1);
287 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
288 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
289 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
290 syncookie_reseed, &V_tcp_syncache);
291 }
292
293 #ifdef VIMAGE
294 void
295 syncache_destroy(void)
296 {
297 struct syncache_head *sch;
298 struct syncache *sc, *nsc;
299 int i;
300
301 /*
302 * Stop the re-seed timer before freeing resources. No need to
303 * possibly schedule it another time.
304 */
305 callout_drain(&V_tcp_syncache.secret.reseed);
306
307 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
308 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
309
310 sch = &V_tcp_syncache.hashbase[i];
311 callout_drain(&sch->sch_timer);
312
313 SCH_LOCK(sch);
314 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
315 syncache_drop(sc, sch);
316 SCH_UNLOCK(sch);
317 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
318 ("%s: sch->sch_bucket not empty", __func__));
319 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
320 __func__, sch->sch_length));
321 mtx_destroy(&sch->sch_mtx);
322 }
323
324 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
325 ("%s: cache_count not 0", __func__));
326
327 /* Free the allocated global resources. */
328 uma_zdestroy(V_tcp_syncache.zone);
329 free(V_tcp_syncache.hashbase, M_SYNCACHE);
330 }
331 #endif
332
333 /*
334 * Inserts a syncache entry into the specified bucket row.
335 * Locks and unlocks the syncache_head autonomously.
336 */
337 static void
338 syncache_insert(struct syncache *sc, struct syncache_head *sch)
339 {
340 struct syncache *sc2;
341
342 SCH_LOCK(sch);
343
344 /*
345 * Make sure that we don't overflow the per-bucket limit.
346 * If the bucket is full, toss the oldest element.
347 */
348 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
349 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
350 ("sch->sch_length incorrect"));
351 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
352 sch->sch_last_overflow = time_uptime;
353 syncache_drop(sc2, sch);
354 TCPSTAT_INC(tcps_sc_bucketoverflow);
355 }
356
357 /* Put it into the bucket. */
358 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
359 sch->sch_length++;
360
361 #ifdef TCP_OFFLOAD
362 if (ADDED_BY_TOE(sc)) {
363 struct toedev *tod = sc->sc_tod;
364
365 tod->tod_syncache_added(tod, sc->sc_todctx);
366 }
367 #endif
368
369 /* Reinitialize the bucket row's timer. */
370 if (sch->sch_length == 1)
371 sch->sch_nextc = ticks + INT_MAX;
372 syncache_timeout(sc, sch, 1);
373
374 SCH_UNLOCK(sch);
375
376 TCPSTATES_INC(TCPS_SYN_RECEIVED);
377 TCPSTAT_INC(tcps_sc_added);
378 }
379
380 /*
381 * Remove and free entry from syncache bucket row.
382 * Expects locked syncache head.
383 */
384 static void
385 syncache_drop(struct syncache *sc, struct syncache_head *sch)
386 {
387
388 SCH_LOCK_ASSERT(sch);
389
390 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
391 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
392 sch->sch_length--;
393
394 #ifdef TCP_OFFLOAD
395 if (ADDED_BY_TOE(sc)) {
396 struct toedev *tod = sc->sc_tod;
397
398 tod->tod_syncache_removed(tod, sc->sc_todctx);
399 }
400 #endif
401
402 syncache_free(sc);
403 }
404
405 /*
406 * Engage/reengage time on bucket row.
407 */
408 static void
409 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
410 {
411 int rexmt;
412
413 if (sc->sc_rxmits == 0)
414 rexmt = TCPTV_RTOBASE;
415 else
416 TCPT_RANGESET(rexmt, TCPTV_RTOBASE * tcp_syn_backoff[sc->sc_rxmits],
417 tcp_rexmit_min, TCPTV_REXMTMAX);
418 sc->sc_rxttime = ticks + rexmt;
419 sc->sc_rxmits++;
420 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
421 sch->sch_nextc = sc->sc_rxttime;
422 if (docallout)
423 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
424 syncache_timer, (void *)sch);
425 }
426 }
427
428 /*
429 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
430 * If we have retransmitted an entry the maximum number of times, expire it.
431 * One separate timer for each bucket row.
432 */
433 static void
434 syncache_timer(void *xsch)
435 {
436 struct syncache_head *sch = (struct syncache_head *)xsch;
437 struct syncache *sc, *nsc;
438 int tick = ticks;
439 char *s;
440
441 CURVNET_SET(sch->sch_sc->vnet);
442
443 /* NB: syncache_head has already been locked by the callout. */
444 SCH_LOCK_ASSERT(sch);
445
446 /*
447 * In the following cycle we may remove some entries and/or
448 * advance some timeouts, so re-initialize the bucket timer.
449 */
450 sch->sch_nextc = tick + INT_MAX;
451
452 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
453 /*
454 * We do not check if the listen socket still exists
455 * and accept the case where the listen socket may be
456 * gone by the time we resend the SYN/ACK. We do
457 * not expect this to happens often. If it does,
458 * then the RST will be sent by the time the remote
459 * host does the SYN/ACK->ACK.
460 */
461 if (TSTMP_GT(sc->sc_rxttime, tick)) {
462 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
463 sch->sch_nextc = sc->sc_rxttime;
464 continue;
465 }
466 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
467 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
468 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
469 "giving up and removing syncache entry\n",
470 s, __func__);
471 free(s, M_TCPLOG);
472 }
473 syncache_drop(sc, sch);
474 TCPSTAT_INC(tcps_sc_stale);
475 continue;
476 }
477 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
478 log(LOG_DEBUG, "%s; %s: Response timeout, "
479 "retransmitting (%u) SYN|ACK\n",
480 s, __func__, sc->sc_rxmits);
481 free(s, M_TCPLOG);
482 }
483
484 syncache_respond(sc, sch, 1, NULL);
485 TCPSTAT_INC(tcps_sc_retransmitted);
486 syncache_timeout(sc, sch, 0);
487 }
488 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
489 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
490 syncache_timer, (void *)(sch));
491 CURVNET_RESTORE();
492 }
493
494 /*
495 * Find an entry in the syncache.
496 * Returns always with locked syncache_head plus a matching entry or NULL.
497 */
498 static struct syncache *
499 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
500 {
501 struct syncache *sc;
502 struct syncache_head *sch;
503 uint32_t hash;
504
505 /*
506 * The hash is built on foreign port + local port + foreign address.
507 * We rely on the fact that struct in_conninfo starts with 16 bits
508 * of foreign port, then 16 bits of local port then followed by 128
509 * bits of foreign address. In case of IPv4 address, the first 3
510 * 32-bit words of the address always are zeroes.
511 */
512 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
513 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
514
515 sch = &V_tcp_syncache.hashbase[hash];
516 *schp = sch;
517 SCH_LOCK(sch);
518
519 /* Circle through bucket row to find matching entry. */
520 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
521 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
522 sizeof(struct in_endpoints)) == 0)
523 break;
524
525 return (sc); /* Always returns with locked sch. */
526 }
527
528 /*
529 * This function is called when we get a RST for a
530 * non-existent connection, so that we can see if the
531 * connection is in the syn cache. If it is, zap it.
532 */
533 void
534 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
535 {
536 struct syncache *sc;
537 struct syncache_head *sch;
538 char *s = NULL;
539
540 sc = syncache_lookup(inc, &sch); /* returns locked sch */
541 SCH_LOCK_ASSERT(sch);
542
543 /*
544 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
545 * See RFC 793 page 65, section SEGMENT ARRIVES.
546 */
547 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
548 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
549 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
550 "FIN flag set, segment ignored\n", s, __func__);
551 TCPSTAT_INC(tcps_badrst);
552 goto done;
553 }
554
555 /*
556 * No corresponding connection was found in syncache.
557 * If syncookies are enabled and possibly exclusively
558 * used, or we are under memory pressure, a valid RST
559 * may not find a syncache entry. In that case we're
560 * done and no SYN|ACK retransmissions will happen.
561 * Otherwise the RST was misdirected or spoofed.
562 */
563 if (sc == NULL) {
564 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
565 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
566 "syncache entry (possibly syncookie only), "
567 "segment ignored\n", s, __func__);
568 TCPSTAT_INC(tcps_badrst);
569 goto done;
570 }
571
572 /*
573 * If the RST bit is set, check the sequence number to see
574 * if this is a valid reset segment.
575 * RFC 793 page 37:
576 * In all states except SYN-SENT, all reset (RST) segments
577 * are validated by checking their SEQ-fields. A reset is
578 * valid if its sequence number is in the window.
579 *
580 * The sequence number in the reset segment is normally an
581 * echo of our outgoing acknowlegement numbers, but some hosts
582 * send a reset with the sequence number at the rightmost edge
583 * of our receive window, and we have to handle this case.
584 */
585 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
586 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
587 syncache_drop(sc, sch);
588 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
589 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
590 "connection attempt aborted by remote endpoint\n",
591 s, __func__);
592 TCPSTAT_INC(tcps_sc_reset);
593 } else {
594 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
595 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
596 "IRS %u (+WND %u), segment ignored\n",
597 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
598 TCPSTAT_INC(tcps_badrst);
599 }
600
601 done:
602 if (s != NULL)
603 free(s, M_TCPLOG);
604 SCH_UNLOCK(sch);
605 }
606
607 void
608 syncache_badack(struct in_conninfo *inc)
609 {
610 struct syncache *sc;
611 struct syncache_head *sch;
612
613 sc = syncache_lookup(inc, &sch); /* returns locked sch */
614 SCH_LOCK_ASSERT(sch);
615 if (sc != NULL) {
616 syncache_drop(sc, sch);
617 TCPSTAT_INC(tcps_sc_badack);
618 }
619 SCH_UNLOCK(sch);
620 }
621
622 void
623 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
624 {
625 struct syncache *sc;
626 struct syncache_head *sch;
627
628 sc = syncache_lookup(inc, &sch); /* returns locked sch */
629 SCH_LOCK_ASSERT(sch);
630 if (sc == NULL)
631 goto done;
632
633 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
634 if (ntohl(th_seq) != sc->sc_iss)
635 goto done;
636
637 /*
638 * If we've rertransmitted 3 times and this is our second error,
639 * we remove the entry. Otherwise, we allow it to continue on.
640 * This prevents us from incorrectly nuking an entry during a
641 * spurious network outage.
642 *
643 * See tcp_notify().
644 */
645 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
646 sc->sc_flags |= SCF_UNREACH;
647 goto done;
648 }
649 syncache_drop(sc, sch);
650 TCPSTAT_INC(tcps_sc_unreach);
651 done:
652 SCH_UNLOCK(sch);
653 }
654
655 /*
656 * Build a new TCP socket structure from a syncache entry.
657 *
658 * On success return the newly created socket with its underlying inp locked.
659 */
660 static struct socket *
661 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
662 {
663 struct tcp_function_block *blk;
664 struct inpcb *inp = NULL;
665 struct socket *so;
666 struct tcpcb *tp;
667 int error;
668 char *s;
669
670 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
671
672 /*
673 * Ok, create the full blown connection, and set things up
674 * as they would have been set up if we had created the
675 * connection when the SYN arrived. If we can't create
676 * the connection, abort it.
677 */
678 so = sonewconn(lso, 0);
679 if (so == NULL) {
680 /*
681 * Drop the connection; we will either send a RST or
682 * have the peer retransmit its SYN again after its
683 * RTO and try again.
684 */
685 TCPSTAT_INC(tcps_listendrop);
686 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
687 log(LOG_DEBUG, "%s; %s: Socket create failed "
688 "due to limits or memory shortage\n",
689 s, __func__);
690 free(s, M_TCPLOG);
691 }
692 goto abort2;
693 }
694 #ifdef MAC
695 mac_socketpeer_set_from_mbuf(m, so);
696 #endif
697
698 inp = sotoinpcb(so);
699 inp->inp_inc.inc_fibnum = so->so_fibnum;
700 INP_WLOCK(inp);
701 /*
702 * Exclusive pcbinfo lock is not required in syncache socket case even
703 * if two inpcb locks can be acquired simultaneously:
704 * - the inpcb in LISTEN state,
705 * - the newly created inp.
706 *
707 * In this case, an inp cannot be at same time in LISTEN state and
708 * just created by an accept() call.
709 */
710 INP_HASH_WLOCK(&V_tcbinfo);
711
712 /* Insert new socket into PCB hash list. */
713 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
714 #ifdef INET6
715 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
716 inp->inp_vflag &= ~INP_IPV4;
717 inp->inp_vflag |= INP_IPV6;
718 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
719 } else {
720 inp->inp_vflag &= ~INP_IPV6;
721 inp->inp_vflag |= INP_IPV4;
722 #endif
723 inp->inp_laddr = sc->sc_inc.inc_laddr;
724 #ifdef INET6
725 }
726 #endif
727
728 /*
729 * If there's an mbuf and it has a flowid, then let's initialise the
730 * inp with that particular flowid.
731 */
732 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
733 inp->inp_flowid = m->m_pkthdr.flowid;
734 inp->inp_flowtype = M_HASHTYPE_GET(m);
735 }
736
737 /*
738 * Install in the reservation hash table for now, but don't yet
739 * install a connection group since the full 4-tuple isn't yet
740 * configured.
741 */
742 inp->inp_lport = sc->sc_inc.inc_lport;
743 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
744 /*
745 * Undo the assignments above if we failed to
746 * put the PCB on the hash lists.
747 */
748 #ifdef INET6
749 if (sc->sc_inc.inc_flags & INC_ISIPV6)
750 inp->in6p_laddr = in6addr_any;
751 else
752 #endif
753 inp->inp_laddr.s_addr = INADDR_ANY;
754 inp->inp_lport = 0;
755 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
756 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
757 "with error %i\n",
758 s, __func__, error);
759 free(s, M_TCPLOG);
760 }
761 INP_HASH_WUNLOCK(&V_tcbinfo);
762 goto abort;
763 }
764 #ifdef INET6
765 if (inp->inp_vflag & INP_IPV6PROTO) {
766 struct inpcb *oinp = sotoinpcb(lso);
767
768 /*
769 * Inherit socket options from the listening socket.
770 * Note that in6p_inputopts are not (and should not be)
771 * copied, since it stores previously received options and is
772 * used to detect if each new option is different than the
773 * previous one and hence should be passed to a user.
774 * If we copied in6p_inputopts, a user would not be able to
775 * receive options just after calling the accept system call.
776 */
777 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
778 if (oinp->in6p_outputopts)
779 inp->in6p_outputopts =
780 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
781 }
782
783 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
784 struct in6_addr laddr6;
785 struct sockaddr_in6 sin6;
786
787 sin6.sin6_family = AF_INET6;
788 sin6.sin6_len = sizeof(sin6);
789 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
790 sin6.sin6_port = sc->sc_inc.inc_fport;
791 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
792 laddr6 = inp->in6p_laddr;
793 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
794 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
795 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
796 thread0.td_ucred, m)) != 0) {
797 inp->in6p_laddr = laddr6;
798 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
799 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
800 "with error %i\n",
801 s, __func__, error);
802 free(s, M_TCPLOG);
803 }
804 INP_HASH_WUNLOCK(&V_tcbinfo);
805 goto abort;
806 }
807 /* Override flowlabel from in6_pcbconnect. */
808 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
809 inp->inp_flow |= sc->sc_flowlabel;
810 }
811 #endif /* INET6 */
812 #if defined(INET) && defined(INET6)
813 else
814 #endif
815 #ifdef INET
816 {
817 struct in_addr laddr;
818 struct sockaddr_in sin;
819
820 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
821
822 if (inp->inp_options == NULL) {
823 inp->inp_options = sc->sc_ipopts;
824 sc->sc_ipopts = NULL;
825 }
826
827 sin.sin_family = AF_INET;
828 sin.sin_len = sizeof(sin);
829 sin.sin_addr = sc->sc_inc.inc_faddr;
830 sin.sin_port = sc->sc_inc.inc_fport;
831 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
832 laddr = inp->inp_laddr;
833 if (inp->inp_laddr.s_addr == INADDR_ANY)
834 inp->inp_laddr = sc->sc_inc.inc_laddr;
835 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
836 thread0.td_ucred, m)) != 0) {
837 inp->inp_laddr = laddr;
838 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
839 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
840 "with error %i\n",
841 s, __func__, error);
842 free(s, M_TCPLOG);
843 }
844 INP_HASH_WUNLOCK(&V_tcbinfo);
845 goto abort;
846 }
847 }
848 #endif /* INET */
849 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
850 /* Copy old policy into new socket's. */
851 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
852 printf("syncache_socket: could not copy policy\n");
853 #endif
854 INP_HASH_WUNLOCK(&V_tcbinfo);
855 tp = intotcpcb(inp);
856 tcp_state_change(tp, TCPS_SYN_RECEIVED);
857 tp->iss = sc->sc_iss;
858 tp->irs = sc->sc_irs;
859 tcp_rcvseqinit(tp);
860 tcp_sendseqinit(tp);
861 blk = sototcpcb(lso)->t_fb;
862 if (blk != tp->t_fb) {
863 /*
864 * Our parents t_fb was not the default,
865 * we need to release our ref on tp->t_fb and
866 * pickup one on the new entry.
867 */
868 struct tcp_function_block *rblk;
869
870 rblk = find_and_ref_tcp_fb(blk);
871 KASSERT(rblk != NULL,
872 ("cannot find blk %p out of syncache?", blk));
873 if (tp->t_fb->tfb_tcp_fb_fini)
874 (*tp->t_fb->tfb_tcp_fb_fini)(tp);
875 refcount_release(&tp->t_fb->tfb_refcnt);
876 tp->t_fb = rblk;
877 if (tp->t_fb->tfb_tcp_fb_init) {
878 (*tp->t_fb->tfb_tcp_fb_init)(tp);
879 }
880 }
881 tp->snd_wl1 = sc->sc_irs;
882 tp->snd_max = tp->iss + 1;
883 tp->snd_nxt = tp->iss + 1;
884 tp->rcv_up = sc->sc_irs + 1;
885 tp->rcv_wnd = sc->sc_wnd;
886 tp->rcv_adv += tp->rcv_wnd;
887 tp->last_ack_sent = tp->rcv_nxt;
888
889 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
890 if (sc->sc_flags & SCF_NOOPT)
891 tp->t_flags |= TF_NOOPT;
892 else {
893 if (sc->sc_flags & SCF_WINSCALE) {
894 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
895 tp->snd_scale = sc->sc_requested_s_scale;
896 tp->request_r_scale = sc->sc_requested_r_scale;
897 }
898 if (sc->sc_flags & SCF_TIMESTAMP) {
899 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
900 tp->ts_recent = sc->sc_tsreflect;
901 tp->ts_recent_age = tcp_ts_getticks();
902 tp->ts_offset = sc->sc_tsoff;
903 }
904 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
905 if (sc->sc_flags & SCF_SIGNATURE)
906 tp->t_flags |= TF_SIGNATURE;
907 #endif
908 if (sc->sc_flags & SCF_SACK)
909 tp->t_flags |= TF_SACK_PERMIT;
910 }
911
912 if (sc->sc_flags & SCF_ECN)
913 tp->t_flags |= TF_ECN_PERMIT;
914
915 /*
916 * Set up MSS and get cached values from tcp_hostcache.
917 * This might overwrite some of the defaults we just set.
918 */
919 tcp_mss(tp, sc->sc_peer_mss);
920
921 /*
922 * If the SYN,ACK was retransmitted, indicate that CWND to be
923 * limited to one segment in cc_conn_init().
924 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
925 */
926 if (sc->sc_rxmits > 1)
927 tp->snd_cwnd = 1;
928
929 #ifdef TCP_OFFLOAD
930 /*
931 * Allow a TOE driver to install its hooks. Note that we hold the
932 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
933 * new connection before the TOE driver has done its thing.
934 */
935 if (ADDED_BY_TOE(sc)) {
936 struct toedev *tod = sc->sc_tod;
937
938 tod->tod_offload_socket(tod, sc->sc_todctx, so);
939 }
940 #endif
941 /*
942 * Copy and activate timers.
943 */
944 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
945 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
946 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
947 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
948 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
949
950 TCPSTAT_INC(tcps_accepts);
951 return (so);
952
953 abort:
954 INP_WUNLOCK(inp);
955 abort2:
956 if (so != NULL)
957 soabort(so);
958 return (NULL);
959 }
960
961 /*
962 * This function gets called when we receive an ACK for a
963 * socket in the LISTEN state. We look up the connection
964 * in the syncache, and if its there, we pull it out of
965 * the cache and turn it into a full-blown connection in
966 * the SYN-RECEIVED state.
967 *
968 * On syncache_socket() success the newly created socket
969 * has its underlying inp locked.
970 */
971 int
972 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
973 struct socket **lsop, struct mbuf *m)
974 {
975 struct syncache *sc;
976 struct syncache_head *sch;
977 struct syncache scs;
978 char *s;
979
980 /*
981 * Global TCP locks are held because we manipulate the PCB lists
982 * and create a new socket.
983 */
984 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
985 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
986 ("%s: can handle only ACK", __func__));
987
988 sc = syncache_lookup(inc, &sch); /* returns locked sch */
989 SCH_LOCK_ASSERT(sch);
990
991 #ifdef INVARIANTS
992 /*
993 * Test code for syncookies comparing the syncache stored
994 * values with the reconstructed values from the cookie.
995 */
996 if (sc != NULL)
997 syncookie_cmp(inc, sch, sc, th, to, *lsop);
998 #endif
999
1000 if (sc == NULL) {
1001 /*
1002 * There is no syncache entry, so see if this ACK is
1003 * a returning syncookie. To do this, first:
1004 * A. Check if syncookies are used in case of syncache
1005 * overflows
1006 * B. See if this socket has had a syncache entry dropped in
1007 * the recent past. We don't want to accept a bogus
1008 * syncookie if we've never received a SYN or accept it
1009 * twice.
1010 * C. check that the syncookie is valid. If it is, then
1011 * cobble up a fake syncache entry, and return.
1012 */
1013 if (!V_tcp_syncookies) {
1014 SCH_UNLOCK(sch);
1015 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1016 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1017 "segment rejected (syncookies disabled)\n",
1018 s, __func__);
1019 goto failed;
1020 }
1021 if (!V_tcp_syncookiesonly &&
1022 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1023 SCH_UNLOCK(sch);
1024 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1025 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1026 "segment rejected (no syncache entry)\n",
1027 s, __func__);
1028 goto failed;
1029 }
1030 bzero(&scs, sizeof(scs));
1031 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1032 SCH_UNLOCK(sch);
1033 if (sc == NULL) {
1034 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1035 log(LOG_DEBUG, "%s; %s: Segment failed "
1036 "SYNCOOKIE authentication, segment rejected "
1037 "(probably spoofed)\n", s, __func__);
1038 goto failed;
1039 }
1040 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1041 /* If received ACK has MD5 signature, check it. */
1042 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1043 (!TCPMD5_ENABLED() ||
1044 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1045 /* Drop the ACK. */
1046 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1047 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1048 "MD5 signature doesn't match.\n",
1049 s, __func__);
1050 free(s, M_TCPLOG);
1051 }
1052 TCPSTAT_INC(tcps_sig_err_sigopt);
1053 return (-1); /* Do not send RST */
1054 }
1055 #endif /* TCP_SIGNATURE */
1056 } else {
1057 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1058 /*
1059 * If listening socket requested TCP digests, check that
1060 * received ACK has signature and it is correct.
1061 * If not, drop the ACK and leave sc entry in th cache,
1062 * because SYN was received with correct signature.
1063 */
1064 if (sc->sc_flags & SCF_SIGNATURE) {
1065 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1066 /* No signature */
1067 TCPSTAT_INC(tcps_sig_err_nosigopt);
1068 SCH_UNLOCK(sch);
1069 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1070 log(LOG_DEBUG, "%s; %s: Segment "
1071 "rejected, MD5 signature wasn't "
1072 "provided.\n", s, __func__);
1073 free(s, M_TCPLOG);
1074 }
1075 return (-1); /* Do not send RST */
1076 }
1077 if (!TCPMD5_ENABLED() ||
1078 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1079 /* Doesn't match or no SA */
1080 SCH_UNLOCK(sch);
1081 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1082 log(LOG_DEBUG, "%s; %s: Segment "
1083 "rejected, MD5 signature doesn't "
1084 "match.\n", s, __func__);
1085 free(s, M_TCPLOG);
1086 }
1087 return (-1); /* Do not send RST */
1088 }
1089 }
1090 #endif /* TCP_SIGNATURE */
1091
1092 /*
1093 * RFC 7323 PAWS: If we have a timestamp on this segment and
1094 * it's less than ts_recent, drop it.
1095 * XXXMT: RFC 7323 also requires to send an ACK.
1096 * In tcp_input.c this is only done for TCP segments
1097 * with user data, so be consistent here and just drop
1098 * the segment.
1099 */
1100 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1101 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1102 SCH_UNLOCK(sch);
1103 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1104 log(LOG_DEBUG,
1105 "%s; %s: SEG.TSval %u < TS.Recent %u, "
1106 "segment dropped\n", s, __func__,
1107 to->to_tsval, sc->sc_tsreflect);
1108 free(s, M_TCPLOG);
1109 }
1110 return (-1); /* Do not send RST */
1111 }
1112
1113 /*
1114 * If timestamps were not negotiated during SYN/ACK and a
1115 * segment with a timestamp is received, ignore the
1116 * timestamp and process the packet normally.
1117 * See section 3.2 of RFC 7323.
1118 */
1119 if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1120 (to->to_flags & TOF_TS)) {
1121 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1122 log(LOG_DEBUG, "%s; %s: Timestamp not "
1123 "expected, segment processed normally\n",
1124 s, __func__);
1125 free(s, M_TCPLOG);
1126 s = NULL;
1127 }
1128 }
1129
1130 /*
1131 * If timestamps were negotiated during SYN/ACK and a
1132 * segment without a timestamp is received, silently drop
1133 * the segment.
1134 * See section 3.2 of RFC 7323.
1135 */
1136 if ((sc->sc_flags & SCF_TIMESTAMP) &&
1137 !(to->to_flags & TOF_TS)) {
1138 SCH_UNLOCK(sch);
1139 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1140 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1141 "segment silently dropped\n", s, __func__);
1142 free(s, M_TCPLOG);
1143 }
1144 return (-1); /* Do not send RST */
1145 }
1146
1147 /*
1148 * Pull out the entry to unlock the bucket row.
1149 *
1150 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1151 * tcp_state_change(). The tcpcb is not existent at this
1152 * moment. A new one will be allocated via syncache_socket->
1153 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1154 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1155 */
1156 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1157 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1158 sch->sch_length--;
1159 #ifdef TCP_OFFLOAD
1160 if (ADDED_BY_TOE(sc)) {
1161 struct toedev *tod = sc->sc_tod;
1162
1163 tod->tod_syncache_removed(tod, sc->sc_todctx);
1164 }
1165 #endif
1166 SCH_UNLOCK(sch);
1167 }
1168
1169 /*
1170 * Segment validation:
1171 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1172 */
1173 if (th->th_ack != sc->sc_iss + 1) {
1174 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1175 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1176 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1177 goto failed;
1178 }
1179
1180 /*
1181 * The SEQ must fall in the window starting at the received
1182 * initial receive sequence number + 1 (the SYN).
1183 */
1184 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1185 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1186 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1187 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1188 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1189 goto failed;
1190 }
1191
1192 *lsop = syncache_socket(sc, *lsop, m);
1193
1194 if (*lsop == NULL)
1195 TCPSTAT_INC(tcps_sc_aborted);
1196 else
1197 TCPSTAT_INC(tcps_sc_completed);
1198
1199 /* how do we find the inp for the new socket? */
1200 if (sc != &scs)
1201 syncache_free(sc);
1202 return (1);
1203 failed:
1204 if (sc != NULL && sc != &scs)
1205 syncache_free(sc);
1206 if (s != NULL)
1207 free(s, M_TCPLOG);
1208 *lsop = NULL;
1209 return (0);
1210 }
1211
1212 #ifdef TCP_RFC7413
1213 static void
1214 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1215 uint64_t response_cookie)
1216 {
1217 struct inpcb *inp;
1218 struct tcpcb *tp;
1219 unsigned int *pending_counter;
1220
1221 /*
1222 * Global TCP locks are held because we manipulate the PCB lists
1223 * and create a new socket.
1224 */
1225 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1226
1227 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1228 *lsop = syncache_socket(sc, *lsop, m);
1229 if (*lsop == NULL) {
1230 TCPSTAT_INC(tcps_sc_aborted);
1231 atomic_subtract_int(pending_counter, 1);
1232 } else {
1233 inp = sotoinpcb(*lsop);
1234 tp = intotcpcb(inp);
1235 tp->t_flags |= TF_FASTOPEN;
1236 tp->t_tfo_cookie = response_cookie;
1237 tp->snd_max = tp->iss;
1238 tp->snd_nxt = tp->iss;
1239 tp->t_tfo_pending = pending_counter;
1240 TCPSTAT_INC(tcps_sc_completed);
1241 }
1242 }
1243 #endif /* TCP_RFC7413 */
1244
1245 /*
1246 * Given a LISTEN socket and an inbound SYN request, add
1247 * this to the syn cache, and send back a segment:
1248 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1249 * to the source.
1250 *
1251 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1252 * Doing so would require that we hold onto the data and deliver it
1253 * to the application. However, if we are the target of a SYN-flood
1254 * DoS attack, an attacker could send data which would eventually
1255 * consume all available buffer space if it were ACKed. By not ACKing
1256 * the data, we avoid this DoS scenario.
1257 *
1258 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1259 * cookie is processed, V_tcp_fastopen_enabled set to true, and the
1260 * TCP_FASTOPEN socket option is set. In this case, a new socket is created
1261 * and returned via lsop, the mbuf is not freed so that tcp_input() can
1262 * queue its data to the socket, and 1 is returned to indicate the
1263 * TFO-socket-creation path was taken.
1264 */
1265 int
1266 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1267 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1268 void *todctx)
1269 {
1270 struct tcpcb *tp;
1271 struct socket *so;
1272 struct syncache *sc = NULL;
1273 struct syncache_head *sch;
1274 struct mbuf *ipopts = NULL;
1275 u_int ltflags;
1276 int win, sb_hiwat, ip_ttl, ip_tos;
1277 char *s;
1278 int rv = 0;
1279 #ifdef INET6
1280 int autoflowlabel = 0;
1281 #endif
1282 #ifdef MAC
1283 struct label *maclabel;
1284 #endif
1285 struct syncache scs;
1286 struct ucred *cred;
1287 #ifdef TCP_RFC7413
1288 uint64_t tfo_response_cookie;
1289 int tfo_cookie_valid = 0;
1290 int tfo_response_cookie_valid = 0;
1291 #endif
1292
1293 INP_WLOCK_ASSERT(inp); /* listen socket */
1294 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1295 ("%s: unexpected tcp flags", __func__));
1296
1297 /*
1298 * Combine all so/tp operations very early to drop the INP lock as
1299 * soon as possible.
1300 */
1301 so = *lsop;
1302 tp = sototcpcb(so);
1303 cred = crhold(so->so_cred);
1304
1305 #ifdef INET6
1306 if ((inc->inc_flags & INC_ISIPV6) &&
1307 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1308 autoflowlabel = 1;
1309 #endif
1310 ip_ttl = inp->inp_ip_ttl;
1311 ip_tos = inp->inp_ip_tos;
1312 win = sbspace(&so->so_rcv);
1313 sb_hiwat = so->so_rcv.sb_hiwat;
1314 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1315
1316 #ifdef TCP_RFC7413
1317 if (V_tcp_fastopen_enabled && (tp->t_flags & TF_FASTOPEN) &&
1318 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) {
1319 /*
1320 * Limit the number of pending TFO connections to
1321 * approximately half of the queue limit. This prevents TFO
1322 * SYN floods from starving the service by filling the
1323 * listen queue with bogus TFO connections.
1324 */
1325 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1326 (so->so_qlimit / 2)) {
1327 int result;
1328
1329 result = tcp_fastopen_check_cookie(inc,
1330 to->to_tfo_cookie, to->to_tfo_len,
1331 &tfo_response_cookie);
1332 tfo_cookie_valid = (result > 0);
1333 tfo_response_cookie_valid = (result >= 0);
1334 } else
1335 atomic_subtract_int(tp->t_tfo_pending, 1);
1336 }
1337 #endif
1338
1339 /* By the time we drop the lock these should no longer be used. */
1340 so = NULL;
1341 tp = NULL;
1342
1343 #ifdef MAC
1344 if (mac_syncache_init(&maclabel) != 0) {
1345 INP_WUNLOCK(inp);
1346 goto done;
1347 } else
1348 mac_syncache_create(maclabel, inp);
1349 #endif
1350 #ifdef TCP_RFC7413
1351 if (!tfo_cookie_valid)
1352 #endif
1353 INP_WUNLOCK(inp);
1354
1355 /*
1356 * Remember the IP options, if any.
1357 */
1358 #ifdef INET6
1359 if (!(inc->inc_flags & INC_ISIPV6))
1360 #endif
1361 #ifdef INET
1362 ipopts = (m) ? ip_srcroute(m) : NULL;
1363 #else
1364 ipopts = NULL;
1365 #endif
1366
1367 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1368 /*
1369 * If listening socket requested TCP digests, check that received
1370 * SYN has signature and it is correct. If signature doesn't match
1371 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1372 */
1373 if (ltflags & TF_SIGNATURE) {
1374 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1375 TCPSTAT_INC(tcps_sig_err_nosigopt);
1376 goto done;
1377 }
1378 if (!TCPMD5_ENABLED() ||
1379 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1380 goto done;
1381 }
1382 #endif /* TCP_SIGNATURE */
1383 /*
1384 * See if we already have an entry for this connection.
1385 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1386 *
1387 * XXX: should the syncache be re-initialized with the contents
1388 * of the new SYN here (which may have different options?)
1389 *
1390 * XXX: We do not check the sequence number to see if this is a
1391 * real retransmit or a new connection attempt. The question is
1392 * how to handle such a case; either ignore it as spoofed, or
1393 * drop the current entry and create a new one?
1394 */
1395 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1396 SCH_LOCK_ASSERT(sch);
1397 if (sc != NULL) {
1398 #ifdef TCP_RFC7413
1399 if (tfo_cookie_valid)
1400 INP_WUNLOCK(inp);
1401 #endif
1402 TCPSTAT_INC(tcps_sc_dupsyn);
1403 if (ipopts) {
1404 /*
1405 * If we were remembering a previous source route,
1406 * forget it and use the new one we've been given.
1407 */
1408 if (sc->sc_ipopts)
1409 (void) m_free(sc->sc_ipopts);
1410 sc->sc_ipopts = ipopts;
1411 }
1412 /*
1413 * Update timestamp if present.
1414 */
1415 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1416 sc->sc_tsreflect = to->to_tsval;
1417 else
1418 sc->sc_flags &= ~SCF_TIMESTAMP;
1419 #ifdef MAC
1420 /*
1421 * Since we have already unconditionally allocated label
1422 * storage, free it up. The syncache entry will already
1423 * have an initialized label we can use.
1424 */
1425 mac_syncache_destroy(&maclabel);
1426 #endif
1427 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1428 /* Retransmit SYN|ACK and reset retransmit count. */
1429 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1430 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1431 "resetting timer and retransmitting SYN|ACK\n",
1432 s, __func__);
1433 free(s, M_TCPLOG);
1434 }
1435 if (syncache_respond(sc, sch, 1, m) == 0) {
1436 sc->sc_rxmits = 0;
1437 syncache_timeout(sc, sch, 1);
1438 TCPSTAT_INC(tcps_sndacks);
1439 TCPSTAT_INC(tcps_sndtotal);
1440 }
1441 SCH_UNLOCK(sch);
1442 goto donenoprobe;
1443 }
1444
1445 #ifdef TCP_RFC7413
1446 if (tfo_cookie_valid) {
1447 bzero(&scs, sizeof(scs));
1448 sc = &scs;
1449 goto skip_alloc;
1450 }
1451 #endif
1452
1453 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1454 if (sc == NULL) {
1455 /*
1456 * The zone allocator couldn't provide more entries.
1457 * Treat this as if the cache was full; drop the oldest
1458 * entry and insert the new one.
1459 */
1460 TCPSTAT_INC(tcps_sc_zonefail);
1461 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1462 sch->sch_last_overflow = time_uptime;
1463 syncache_drop(sc, sch);
1464 }
1465 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1466 if (sc == NULL) {
1467 if (V_tcp_syncookies) {
1468 bzero(&scs, sizeof(scs));
1469 sc = &scs;
1470 } else {
1471 SCH_UNLOCK(sch);
1472 if (ipopts)
1473 (void) m_free(ipopts);
1474 goto done;
1475 }
1476 }
1477 }
1478
1479 #ifdef TCP_RFC7413
1480 skip_alloc:
1481 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1482 sc->sc_tfo_cookie = &tfo_response_cookie;
1483 #endif
1484
1485 /*
1486 * Fill in the syncache values.
1487 */
1488 #ifdef MAC
1489 sc->sc_label = maclabel;
1490 #endif
1491 sc->sc_cred = cred;
1492 cred = NULL;
1493 sc->sc_ipopts = ipopts;
1494 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1495 #ifdef INET6
1496 if (!(inc->inc_flags & INC_ISIPV6))
1497 #endif
1498 {
1499 sc->sc_ip_tos = ip_tos;
1500 sc->sc_ip_ttl = ip_ttl;
1501 }
1502 #ifdef TCP_OFFLOAD
1503 sc->sc_tod = tod;
1504 sc->sc_todctx = todctx;
1505 #endif
1506 sc->sc_irs = th->th_seq;
1507 sc->sc_iss = arc4random();
1508 sc->sc_flags = 0;
1509 sc->sc_flowlabel = 0;
1510
1511 /*
1512 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1513 * win was derived from socket earlier in the function.
1514 */
1515 win = imax(win, 0);
1516 win = imin(win, TCP_MAXWIN);
1517 sc->sc_wnd = win;
1518
1519 if (V_tcp_do_rfc1323) {
1520 /*
1521 * A timestamp received in a SYN makes
1522 * it ok to send timestamp requests and replies.
1523 */
1524 if (to->to_flags & TOF_TS) {
1525 sc->sc_tsreflect = to->to_tsval;
1526 sc->sc_flags |= SCF_TIMESTAMP;
1527 sc->sc_tsoff = tcp_new_ts_offset(inc);
1528 }
1529 if (to->to_flags & TOF_SCALE) {
1530 int wscale = 0;
1531
1532 /*
1533 * Pick the smallest possible scaling factor that
1534 * will still allow us to scale up to sb_max, aka
1535 * kern.ipc.maxsockbuf.
1536 *
1537 * We do this because there are broken firewalls that
1538 * will corrupt the window scale option, leading to
1539 * the other endpoint believing that our advertised
1540 * window is unscaled. At scale factors larger than
1541 * 5 the unscaled window will drop below 1500 bytes,
1542 * leading to serious problems when traversing these
1543 * broken firewalls.
1544 *
1545 * With the default maxsockbuf of 256K, a scale factor
1546 * of 3 will be chosen by this algorithm. Those who
1547 * choose a larger maxsockbuf should watch out
1548 * for the compatibility problems mentioned above.
1549 *
1550 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1551 * or <SYN,ACK>) segment itself is never scaled.
1552 */
1553 while (wscale < TCP_MAX_WINSHIFT &&
1554 (TCP_MAXWIN << wscale) < sb_max)
1555 wscale++;
1556 sc->sc_requested_r_scale = wscale;
1557 sc->sc_requested_s_scale = to->to_wscale;
1558 sc->sc_flags |= SCF_WINSCALE;
1559 }
1560 }
1561 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1562 /*
1563 * If listening socket requested TCP digests, flag this in the
1564 * syncache so that syncache_respond() will do the right thing
1565 * with the SYN+ACK.
1566 */
1567 if (ltflags & TF_SIGNATURE)
1568 sc->sc_flags |= SCF_SIGNATURE;
1569 #endif /* TCP_SIGNATURE */
1570 if (to->to_flags & TOF_SACKPERM)
1571 sc->sc_flags |= SCF_SACK;
1572 if (to->to_flags & TOF_MSS)
1573 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1574 if (ltflags & TF_NOOPT)
1575 sc->sc_flags |= SCF_NOOPT;
1576 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1577 sc->sc_flags |= SCF_ECN;
1578
1579 if (V_tcp_syncookies)
1580 sc->sc_iss = syncookie_generate(sch, sc);
1581 #ifdef INET6
1582 if (autoflowlabel) {
1583 if (V_tcp_syncookies)
1584 sc->sc_flowlabel = sc->sc_iss;
1585 else
1586 sc->sc_flowlabel = ip6_randomflowlabel();
1587 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1588 }
1589 #endif
1590 SCH_UNLOCK(sch);
1591
1592 #ifdef TCP_RFC7413
1593 if (tfo_cookie_valid) {
1594 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1595 /* INP_WUNLOCK(inp) will be performed by the called */
1596 rv = 1;
1597 goto tfo_done;
1598 }
1599 #endif
1600
1601 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1602 /*
1603 * Do a standard 3-way handshake.
1604 */
1605 if (syncache_respond(sc, sch, 0, m) == 0) {
1606 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1607 syncache_free(sc);
1608 else if (sc != &scs)
1609 syncache_insert(sc, sch); /* locks and unlocks sch */
1610 TCPSTAT_INC(tcps_sndacks);
1611 TCPSTAT_INC(tcps_sndtotal);
1612 } else {
1613 if (sc != &scs)
1614 syncache_free(sc);
1615 TCPSTAT_INC(tcps_sc_dropped);
1616 }
1617 goto donenoprobe;
1618
1619 done:
1620 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1621 donenoprobe:
1622 if (m) {
1623 *lsop = NULL;
1624 m_freem(m);
1625 }
1626 #ifdef TCP_RFC7413
1627 tfo_done:
1628 #endif
1629 if (cred != NULL)
1630 crfree(cred);
1631 #ifdef MAC
1632 if (sc == &scs)
1633 mac_syncache_destroy(&maclabel);
1634 #endif
1635 return (rv);
1636 }
1637
1638 /*
1639 * Send SYN|ACK to the peer. Either in response to the peer's SYN,
1640 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1641 */
1642 static int
1643 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked,
1644 const struct mbuf *m0)
1645 {
1646 struct ip *ip = NULL;
1647 struct mbuf *m;
1648 struct tcphdr *th = NULL;
1649 int optlen, error = 0; /* Make compiler happy */
1650 u_int16_t hlen, tlen, mssopt;
1651 struct tcpopt to;
1652 #ifdef INET6
1653 struct ip6_hdr *ip6 = NULL;
1654 #endif
1655 hlen =
1656 #ifdef INET6
1657 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1658 #endif
1659 sizeof(struct ip);
1660 tlen = hlen + sizeof(struct tcphdr);
1661
1662 /* Determine MSS we advertize to other end of connection. */
1663 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1664
1665 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1666 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1667 ("syncache: mbuf too small"));
1668
1669 /* Create the IP+TCP header from scratch. */
1670 m = m_gethdr(M_NOWAIT, MT_DATA);
1671 if (m == NULL)
1672 return (ENOBUFS);
1673 #ifdef MAC
1674 mac_syncache_create_mbuf(sc->sc_label, m);
1675 #endif
1676 m->m_data += max_linkhdr;
1677 m->m_len = tlen;
1678 m->m_pkthdr.len = tlen;
1679 m->m_pkthdr.rcvif = NULL;
1680
1681 #ifdef INET6
1682 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1683 ip6 = mtod(m, struct ip6_hdr *);
1684 ip6->ip6_vfc = IPV6_VERSION;
1685 ip6->ip6_nxt = IPPROTO_TCP;
1686 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1687 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1688 ip6->ip6_plen = htons(tlen - hlen);
1689 /* ip6_hlim is set after checksum */
1690 /* Zero out traffic class and flow label. */
1691 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1692 ip6->ip6_flow |= sc->sc_flowlabel;
1693
1694 th = (struct tcphdr *)(ip6 + 1);
1695 }
1696 #endif
1697 #if defined(INET6) && defined(INET)
1698 else
1699 #endif
1700 #ifdef INET
1701 {
1702 ip = mtod(m, struct ip *);
1703 ip->ip_v = IPVERSION;
1704 ip->ip_hl = sizeof(struct ip) >> 2;
1705 ip->ip_len = htons(tlen);
1706 ip->ip_id = 0;
1707 ip->ip_off = 0;
1708 ip->ip_sum = 0;
1709 ip->ip_p = IPPROTO_TCP;
1710 ip->ip_src = sc->sc_inc.inc_laddr;
1711 ip->ip_dst = sc->sc_inc.inc_faddr;
1712 ip->ip_ttl = sc->sc_ip_ttl;
1713 ip->ip_tos = sc->sc_ip_tos;
1714
1715 /*
1716 * See if we should do MTU discovery. Route lookups are
1717 * expensive, so we will only unset the DF bit if:
1718 *
1719 * 1) path_mtu_discovery is disabled
1720 * 2) the SCF_UNREACH flag has been set
1721 */
1722 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1723 ip->ip_off |= htons(IP_DF);
1724
1725 th = (struct tcphdr *)(ip + 1);
1726 }
1727 #endif /* INET */
1728 th->th_sport = sc->sc_inc.inc_lport;
1729 th->th_dport = sc->sc_inc.inc_fport;
1730
1731 th->th_seq = htonl(sc->sc_iss);
1732 th->th_ack = htonl(sc->sc_irs + 1);
1733 th->th_off = sizeof(struct tcphdr) >> 2;
1734 th->th_x2 = 0;
1735 th->th_flags = TH_SYN|TH_ACK;
1736 th->th_win = htons(sc->sc_wnd);
1737 th->th_urp = 0;
1738
1739 if (sc->sc_flags & SCF_ECN) {
1740 th->th_flags |= TH_ECE;
1741 TCPSTAT_INC(tcps_ecn_shs);
1742 }
1743
1744 /* Tack on the TCP options. */
1745 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1746 to.to_flags = 0;
1747
1748 to.to_mss = mssopt;
1749 to.to_flags = TOF_MSS;
1750 if (sc->sc_flags & SCF_WINSCALE) {
1751 to.to_wscale = sc->sc_requested_r_scale;
1752 to.to_flags |= TOF_SCALE;
1753 }
1754 if (sc->sc_flags & SCF_TIMESTAMP) {
1755 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1756 to.to_tsecr = sc->sc_tsreflect;
1757 to.to_flags |= TOF_TS;
1758 }
1759 if (sc->sc_flags & SCF_SACK)
1760 to.to_flags |= TOF_SACKPERM;
1761 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1762 if (sc->sc_flags & SCF_SIGNATURE)
1763 to.to_flags |= TOF_SIGNATURE;
1764 #endif
1765 #ifdef TCP_RFC7413
1766 if (sc->sc_tfo_cookie) {
1767 to.to_flags |= TOF_FASTOPEN;
1768 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1769 to.to_tfo_cookie = sc->sc_tfo_cookie;
1770 /* don't send cookie again when retransmitting response */
1771 sc->sc_tfo_cookie = NULL;
1772 }
1773 #endif
1774 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1775
1776 /* Adjust headers by option size. */
1777 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1778 m->m_len += optlen;
1779 m->m_pkthdr.len += optlen;
1780 #ifdef INET6
1781 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1782 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1783 else
1784 #endif
1785 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1786 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1787 if (sc->sc_flags & SCF_SIGNATURE) {
1788 KASSERT(to.to_flags & TOF_SIGNATURE,
1789 ("tcp_addoptions() didn't set tcp_signature"));
1790
1791 /* NOTE: to.to_signature is inside of mbuf */
1792 if (!TCPMD5_ENABLED() ||
1793 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1794 m_freem(m);
1795 return (EACCES);
1796 }
1797 }
1798 #endif
1799 } else
1800 optlen = 0;
1801
1802 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1803 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1804 /*
1805 * If we have peer's SYN and it has a flowid, then let's assign it to
1806 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1807 * to SYN|ACK due to lack of inp here.
1808 */
1809 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1810 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1811 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1812 }
1813 #ifdef INET6
1814 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1815 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1816 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1817 IPPROTO_TCP, 0);
1818 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1819 #ifdef TCP_OFFLOAD
1820 if (ADDED_BY_TOE(sc)) {
1821 struct toedev *tod = sc->sc_tod;
1822
1823 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1824
1825 return (error);
1826 }
1827 #endif
1828 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1829 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1830 }
1831 #endif
1832 #if defined(INET6) && defined(INET)
1833 else
1834 #endif
1835 #ifdef INET
1836 {
1837 m->m_pkthdr.csum_flags = CSUM_TCP;
1838 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1839 htons(tlen + optlen - hlen + IPPROTO_TCP));
1840 #ifdef TCP_OFFLOAD
1841 if (ADDED_BY_TOE(sc)) {
1842 struct toedev *tod = sc->sc_tod;
1843
1844 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1845
1846 return (error);
1847 }
1848 #endif
1849 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
1850 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1851 }
1852 #endif
1853 return (error);
1854 }
1855
1856 /*
1857 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1858 * that exceed the capacity of the syncache by avoiding the storage of any
1859 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1860 * attacks where the attacker does not have access to our responses.
1861 *
1862 * Syncookies encode and include all necessary information about the
1863 * connection setup within the SYN|ACK that we send back. That way we
1864 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1865 * (if ever). Normally the syncache and syncookies are running in parallel
1866 * with the latter taking over when the former is exhausted. When matching
1867 * syncache entry is found the syncookie is ignored.
1868 *
1869 * The only reliable information persisting the 3WHS is our initial sequence
1870 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1871 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1872 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1873 * returns and signifies a legitimate connection if it matches the ACK.
1874 *
1875 * The available space of 32 bits to store the hash and to encode the SYN
1876 * option information is very tight and we should have at least 24 bits for
1877 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1878 *
1879 * SYN option information we have to encode to fully restore a connection:
1880 * MSS: is imporant to chose an optimal segment size to avoid IP level
1881 * fragmentation along the path. The common MSS values can be encoded
1882 * in a 3-bit table. Uncommon values are captured by the next lower value
1883 * in the table leading to a slight increase in packetization overhead.
1884 * WSCALE: is necessary to allow large windows to be used for high delay-
1885 * bandwidth product links. Not scaling the window when it was initially
1886 * negotiated is bad for performance as lack of scaling further decreases
1887 * the apparent available send window. We only need to encode the WSCALE
1888 * we received from the remote end. Our end can be recalculated at any
1889 * time. The common WSCALE values can be encoded in a 3-bit table.
1890 * Uncommon values are captured by the next lower value in the table
1891 * making us under-estimate the available window size halving our
1892 * theoretically possible maximum throughput for that connection.
1893 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1894 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1895 * that are included in all segments on a connection. We enable them when
1896 * the ACK has them.
1897 *
1898 * Security of syncookies and attack vectors:
1899 *
1900 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1901 * together with the gloabl secret to make it unique per connection attempt.
1902 * Thus any change of any of those parameters results in a different MAC output
1903 * in an unpredictable way unless a collision is encountered. 24 bits of the
1904 * MAC are embedded into the ISS.
1905 *
1906 * To prevent replay attacks two rotating global secrets are updated with a
1907 * new random value every 15 seconds. The life-time of a syncookie is thus
1908 * 15-30 seconds.
1909 *
1910 * Vector 1: Attacking the secret. This requires finding a weakness in the
1911 * MAC itself or the way it is used here. The attacker can do a chosen plain
1912 * text attack by varying and testing the all parameters under his control.
1913 * The strength depends on the size and randomness of the secret, and the
1914 * cryptographic security of the MAC function. Due to the constant updating
1915 * of the secret the attacker has at most 29.999 seconds to find the secret
1916 * and launch spoofed connections. After that he has to start all over again.
1917 *
1918 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1919 * size an average of 4,823 attempts are required for a 50% chance of success
1920 * to spoof a single syncookie (birthday collision paradox). However the
1921 * attacker is blind and doesn't know if one of his attempts succeeded unless
1922 * he has a side channel to interfere success from. A single connection setup
1923 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1924 * This many attempts are required for each one blind spoofed connection. For
1925 * every additional spoofed connection he has to launch another N attempts.
1926 * Thus for a sustained rate 100 spoofed connections per second approximately
1927 * 1,800,000 packets per second would have to be sent.
1928 *
1929 * NB: The MAC function should be fast so that it doesn't become a CPU
1930 * exhaustion attack vector itself.
1931 *
1932 * References:
1933 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1934 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1935 * http://cr.yp.to/syncookies.html (overview)
1936 * http://cr.yp.to/syncookies/archive (details)
1937 *
1938 *
1939 * Schematic construction of a syncookie enabled Initial Sequence Number:
1940 * 0 1 2 3
1941 * 12345678901234567890123456789012
1942 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1943 *
1944 * x 24 MAC (truncated)
1945 * W 3 Send Window Scale index
1946 * M 3 MSS index
1947 * S 1 SACK permitted
1948 * P 1 Odd/even secret
1949 */
1950
1951 /*
1952 * Distribution and probability of certain MSS values. Those in between are
1953 * rounded down to the next lower one.
1954 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1955 * .2% .3% 5% 7% 7% 20% 15% 45%
1956 */
1957 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1958
1959 /*
1960 * Distribution and probability of certain WSCALE values. We have to map the
1961 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1962 * bits based on prevalence of certain values. Where we don't have an exact
1963 * match for are rounded down to the next lower one letting us under-estimate
1964 * the true available window. At the moment this would happen only for the
1965 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1966 * and window size). The absence of the WSCALE option (no scaling in either
1967 * direction) is encoded with index zero.
1968 * [WSCALE values histograms, Allman, 2012]
1969 * X 10 10 35 5 6 14 10% by host
1970 * X 11 4 5 5 18 49 3% by connections
1971 */
1972 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1973
1974 /*
1975 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
1976 * and good cryptographic properties.
1977 */
1978 static uint32_t
1979 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1980 uint8_t *secbits, uintptr_t secmod)
1981 {
1982 SIPHASH_CTX ctx;
1983 uint32_t siphash[2];
1984
1985 SipHash24_Init(&ctx);
1986 SipHash_SetKey(&ctx, secbits);
1987 switch (inc->inc_flags & INC_ISIPV6) {
1988 #ifdef INET
1989 case 0:
1990 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1991 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1992 break;
1993 #endif
1994 #ifdef INET6
1995 case INC_ISIPV6:
1996 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1997 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1998 break;
1999 #endif
2000 }
2001 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2002 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2003 SipHash_Update(&ctx, &irs, sizeof(irs));
2004 SipHash_Update(&ctx, &flags, sizeof(flags));
2005 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2006 SipHash_Final((u_int8_t *)&siphash, &ctx);
2007
2008 return (siphash[0] ^ siphash[1]);
2009 }
2010
2011 static tcp_seq
2012 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2013 {
2014 u_int i, secbit, wscale;
2015 uint32_t iss, hash;
2016 uint8_t *secbits;
2017 union syncookie cookie;
2018
2019 SCH_LOCK_ASSERT(sch);
2020
2021 cookie.cookie = 0;
2022
2023 /* Map our computed MSS into the 3-bit index. */
2024 for (i = nitems(tcp_sc_msstab) - 1;
2025 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2026 i--)
2027 ;
2028 cookie.flags.mss_idx = i;
2029
2030 /*
2031 * Map the send window scale into the 3-bit index but only if
2032 * the wscale option was received.
2033 */
2034 if (sc->sc_flags & SCF_WINSCALE) {
2035 wscale = sc->sc_requested_s_scale;
2036 for (i = nitems(tcp_sc_wstab) - 1;
2037 tcp_sc_wstab[i] > wscale && i > 0;
2038 i--)
2039 ;
2040 cookie.flags.wscale_idx = i;
2041 }
2042
2043 /* Can we do SACK? */
2044 if (sc->sc_flags & SCF_SACK)
2045 cookie.flags.sack_ok = 1;
2046
2047 /* Which of the two secrets to use. */
2048 secbit = sch->sch_sc->secret.oddeven & 0x1;
2049 cookie.flags.odd_even = secbit;
2050
2051 secbits = sch->sch_sc->secret.key[secbit];
2052 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2053 (uintptr_t)sch);
2054
2055 /*
2056 * Put the flags into the hash and XOR them to get better ISS number
2057 * variance. This doesn't enhance the cryptographic strength and is
2058 * done to prevent the 8 cookie bits from showing up directly on the
2059 * wire.
2060 */
2061 iss = hash & ~0xff;
2062 iss |= cookie.cookie ^ (hash >> 24);
2063
2064 TCPSTAT_INC(tcps_sc_sendcookie);
2065 return (iss);
2066 }
2067
2068 static struct syncache *
2069 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2070 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2071 struct socket *lso)
2072 {
2073 uint32_t hash;
2074 uint8_t *secbits;
2075 tcp_seq ack, seq;
2076 int wnd, wscale = 0;
2077 union syncookie cookie;
2078
2079 SCH_LOCK_ASSERT(sch);
2080
2081 /*
2082 * Pull information out of SYN-ACK/ACK and revert sequence number
2083 * advances.
2084 */
2085 ack = th->th_ack - 1;
2086 seq = th->th_seq - 1;
2087
2088 /*
2089 * Unpack the flags containing enough information to restore the
2090 * connection.
2091 */
2092 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2093
2094 /* Which of the two secrets to use. */
2095 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2096
2097 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2098
2099 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2100 if ((ack & ~0xff) != (hash & ~0xff))
2101 return (NULL);
2102
2103 /* Fill in the syncache values. */
2104 sc->sc_flags = 0;
2105 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2106 sc->sc_ipopts = NULL;
2107
2108 sc->sc_irs = seq;
2109 sc->sc_iss = ack;
2110
2111 switch (inc->inc_flags & INC_ISIPV6) {
2112 #ifdef INET
2113 case 0:
2114 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2115 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2116 break;
2117 #endif
2118 #ifdef INET6
2119 case INC_ISIPV6:
2120 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2121 sc->sc_flowlabel =
2122 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2123 break;
2124 #endif
2125 }
2126
2127 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2128
2129 /* We can simply recompute receive window scale we sent earlier. */
2130 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2131 wscale++;
2132
2133 /* Only use wscale if it was enabled in the orignal SYN. */
2134 if (cookie.flags.wscale_idx > 0) {
2135 sc->sc_requested_r_scale = wscale;
2136 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2137 sc->sc_flags |= SCF_WINSCALE;
2138 }
2139
2140 wnd = sbspace(&lso->so_rcv);
2141 wnd = imax(wnd, 0);
2142 wnd = imin(wnd, TCP_MAXWIN);
2143 sc->sc_wnd = wnd;
2144
2145 if (cookie.flags.sack_ok)
2146 sc->sc_flags |= SCF_SACK;
2147
2148 if (to->to_flags & TOF_TS) {
2149 sc->sc_flags |= SCF_TIMESTAMP;
2150 sc->sc_tsreflect = to->to_tsval;
2151 sc->sc_tsoff = tcp_new_ts_offset(inc);
2152 }
2153
2154 if (to->to_flags & TOF_SIGNATURE)
2155 sc->sc_flags |= SCF_SIGNATURE;
2156
2157 sc->sc_rxmits = 0;
2158
2159 TCPSTAT_INC(tcps_sc_recvcookie);
2160 return (sc);
2161 }
2162
2163 #ifdef INVARIANTS
2164 static int
2165 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2166 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2167 struct socket *lso)
2168 {
2169 struct syncache scs, *scx;
2170 char *s;
2171
2172 bzero(&scs, sizeof(scs));
2173 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2174
2175 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2176 return (0);
2177
2178 if (scx != NULL) {
2179 if (sc->sc_peer_mss != scx->sc_peer_mss)
2180 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2181 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2182
2183 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2184 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2185 s, __func__, sc->sc_requested_r_scale,
2186 scx->sc_requested_r_scale);
2187
2188 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2189 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2190 s, __func__, sc->sc_requested_s_scale,
2191 scx->sc_requested_s_scale);
2192
2193 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2194 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2195 }
2196
2197 if (s != NULL)
2198 free(s, M_TCPLOG);
2199 return (0);
2200 }
2201 #endif /* INVARIANTS */
2202
2203 static void
2204 syncookie_reseed(void *arg)
2205 {
2206 struct tcp_syncache *sc = arg;
2207 uint8_t *secbits;
2208 int secbit;
2209
2210 /*
2211 * Reseeding the secret doesn't have to be protected by a lock.
2212 * It only must be ensured that the new random values are visible
2213 * to all CPUs in a SMP environment. The atomic with release
2214 * semantics ensures that.
2215 */
2216 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2217 secbits = sc->secret.key[secbit];
2218 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2219 atomic_add_rel_int(&sc->secret.oddeven, 1);
2220
2221 /* Reschedule ourself. */
2222 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2223 }
2224
2225 /*
2226 * Exports the syncache entries to userland so that netstat can display
2227 * them alongside the other sockets. This function is intended to be
2228 * called only from tcp_pcblist.
2229 *
2230 * Due to concurrency on an active system, the number of pcbs exported
2231 * may have no relation to max_pcbs. max_pcbs merely indicates the
2232 * amount of space the caller allocated for this function to use.
2233 */
2234 int
2235 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2236 {
2237 struct xtcpcb xt;
2238 struct syncache *sc;
2239 struct syncache_head *sch;
2240 int count, error, i;
2241
2242 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2243 sch = &V_tcp_syncache.hashbase[i];
2244 SCH_LOCK(sch);
2245 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2246 if (count >= max_pcbs) {
2247 SCH_UNLOCK(sch);
2248 goto exit;
2249 }
2250 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2251 continue;
2252 bzero(&xt, sizeof(xt));
2253 xt.xt_len = sizeof(xt);
2254 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2255 xt.xt_inp.inp_vflag = INP_IPV6;
2256 else
2257 xt.xt_inp.inp_vflag = INP_IPV4;
2258 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2259 xt.xt_tp.t_inpcb = &xt.xt_inp;
2260 xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2261 xt.xt_socket.xso_protocol = IPPROTO_TCP;
2262 xt.xt_socket.xso_len = sizeof (struct xsocket);
2263 xt.xt_socket.so_type = SOCK_STREAM;
2264 xt.xt_socket.so_state = SS_ISCONNECTING;
2265 error = SYSCTL_OUT(req, &xt, sizeof xt);
2266 if (error) {
2267 SCH_UNLOCK(sch);
2268 goto exit;
2269 }
2270 count++;
2271 }
2272 SCH_UNLOCK(sch);
2273 }
2274 exit:
2275 *pcbs_exported = count;
2276 return error;
2277 }
Cache object: 6895f90818cbe2771a5458cfcc463978
|