xref: /freebsd/sys/netpfil/pf/pf_norm.c (revision a208223130fb12daac9b4f52be1b3477b03ac521)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright 2001 Niels Provos <provos@citi.umich.edu>
5  * Copyright 2011-2018 Alexander Bluhm <bluhm@openbsd.org>
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  *
28  *	$OpenBSD: pf_norm.c,v 1.114 2009/01/29 14:11:45 henning Exp $
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include "opt_inet.h"
35 #include "opt_inet6.h"
36 #include "opt_pf.h"
37 
38 #include <sys/param.h>
39 #include <sys/kernel.h>
40 #include <sys/lock.h>
41 #include <sys/mbuf.h>
42 #include <sys/mutex.h>
43 #include <sys/refcount.h>
44 #include <sys/socket.h>
45 
46 #include <net/if.h>
47 #include <net/vnet.h>
48 #include <net/pfvar.h>
49 #include <net/if_pflog.h>
50 
51 #include <netinet/in.h>
52 #include <netinet/ip.h>
53 #include <netinet/ip_var.h>
54 #include <netinet6/ip6_var.h>
55 #include <netinet/tcp.h>
56 #include <netinet/tcp_fsm.h>
57 #include <netinet/tcp_seq.h>
58 
59 #ifdef INET6
60 #include <netinet/ip6.h>
61 #endif /* INET6 */
62 
63 struct pf_frent {
64 	TAILQ_ENTRY(pf_frent)	fr_next;
65 	struct mbuf	*fe_m;
66 	uint16_t	fe_hdrlen;	/* ipv4 header length with ip options
67 					   ipv6, extension, fragment header */
68 	uint16_t	fe_extoff;	/* last extension header offset or 0 */
69 	uint16_t	fe_len;		/* fragment length */
70 	uint16_t	fe_off;		/* fragment offset */
71 	uint16_t	fe_mff;		/* more fragment flag */
72 };
73 
74 struct pf_fragment_cmp {
75 	struct pf_addr	frc_src;
76 	struct pf_addr	frc_dst;
77 	uint32_t	frc_id;
78 	sa_family_t	frc_af;
79 	uint8_t		frc_proto;
80 };
81 
82 struct pf_fragment {
83 	struct pf_fragment_cmp	fr_key;
84 #define fr_src	fr_key.frc_src
85 #define fr_dst	fr_key.frc_dst
86 #define fr_id	fr_key.frc_id
87 #define fr_af	fr_key.frc_af
88 #define fr_proto	fr_key.frc_proto
89 
90 	/* pointers to queue element */
91 	struct pf_frent	*fr_firstoff[PF_FRAG_ENTRY_POINTS];
92 	/* count entries between pointers */
93 	uint8_t	fr_entries[PF_FRAG_ENTRY_POINTS];
94 	RB_ENTRY(pf_fragment) fr_entry;
95 	TAILQ_ENTRY(pf_fragment) frag_next;
96 	uint32_t	fr_timeout;
97 	uint16_t	fr_maxlen;	/* maximum length of single fragment */
98 	u_int16_t	fr_holes;	/* number of holes in the queue */
99 	TAILQ_HEAD(pf_fragq, pf_frent) fr_queue;
100 };
101 
102 struct pf_fragment_tag {
103 	uint16_t	ft_hdrlen;	/* header length of reassembled pkt */
104 	uint16_t	ft_extoff;	/* last extension header offset or 0 */
105 	uint16_t	ft_maxlen;	/* maximum fragment payload length */
106 	uint32_t	ft_id;		/* fragment id */
107 };
108 
109 VNET_DEFINE_STATIC(struct mtx, pf_frag_mtx);
110 #define V_pf_frag_mtx		VNET(pf_frag_mtx)
111 #define PF_FRAG_LOCK()		mtx_lock(&V_pf_frag_mtx)
112 #define PF_FRAG_UNLOCK()	mtx_unlock(&V_pf_frag_mtx)
113 #define PF_FRAG_ASSERT()	mtx_assert(&V_pf_frag_mtx, MA_OWNED)
114 
115 VNET_DEFINE(uma_zone_t, pf_state_scrub_z);	/* XXX: shared with pfsync */
116 
117 VNET_DEFINE_STATIC(uma_zone_t, pf_frent_z);
118 #define	V_pf_frent_z	VNET(pf_frent_z)
119 VNET_DEFINE_STATIC(uma_zone_t, pf_frag_z);
120 #define	V_pf_frag_z	VNET(pf_frag_z)
121 
122 TAILQ_HEAD(pf_fragqueue, pf_fragment);
123 TAILQ_HEAD(pf_cachequeue, pf_fragment);
124 VNET_DEFINE_STATIC(struct pf_fragqueue,	pf_fragqueue);
125 #define	V_pf_fragqueue			VNET(pf_fragqueue)
126 RB_HEAD(pf_frag_tree, pf_fragment);
127 VNET_DEFINE_STATIC(struct pf_frag_tree,	pf_frag_tree);
128 #define	V_pf_frag_tree			VNET(pf_frag_tree)
129 static int		 pf_frag_compare(struct pf_fragment *,
130 			    struct pf_fragment *);
131 static RB_PROTOTYPE(pf_frag_tree, pf_fragment, fr_entry, pf_frag_compare);
132 static RB_GENERATE(pf_frag_tree, pf_fragment, fr_entry, pf_frag_compare);
133 
134 static void	pf_flush_fragments(void);
135 static void	pf_free_fragment(struct pf_fragment *);
136 static void	pf_remove_fragment(struct pf_fragment *);
137 static int	pf_normalize_tcpopt(struct pf_krule *, struct mbuf *,
138 		    struct tcphdr *, int, sa_family_t);
139 static struct pf_frent *pf_create_fragment(u_short *);
140 static int	pf_frent_holes(struct pf_frent *frent);
141 static struct pf_fragment *pf_find_fragment(struct pf_fragment_cmp *key,
142 		    struct pf_frag_tree *tree);
143 static inline int	pf_frent_index(struct pf_frent *);
144 static int	pf_frent_insert(struct pf_fragment *,
145 			    struct pf_frent *, struct pf_frent *);
146 void			pf_frent_remove(struct pf_fragment *,
147 			    struct pf_frent *);
148 struct pf_frent		*pf_frent_previous(struct pf_fragment *,
149 			    struct pf_frent *);
150 static struct pf_fragment *pf_fillup_fragment(struct pf_fragment_cmp *,
151 		    struct pf_frent *, u_short *);
152 static struct mbuf *pf_join_fragment(struct pf_fragment *);
153 #ifdef INET
154 static void	pf_scrub_ip(struct mbuf **, uint32_t, uint8_t, uint8_t);
155 static int	pf_reassemble(struct mbuf **, struct ip *, int, u_short *);
156 #endif	/* INET */
157 #ifdef INET6
158 static int	pf_reassemble6(struct mbuf **, struct ip6_hdr *,
159 		    struct ip6_frag *, uint16_t, uint16_t, u_short *);
160 static void	pf_scrub_ip6(struct mbuf **, uint8_t);
161 #endif	/* INET6 */
162 
163 #define	DPFPRINTF(x) do {				\
164 	if (V_pf_status.debug >= PF_DEBUG_MISC) {	\
165 		printf("%s: ", __func__);		\
166 		printf x ;				\
167 	}						\
168 } while(0)
169 
170 #ifdef INET
171 static void
172 pf_ip2key(struct ip *ip, int dir, struct pf_fragment_cmp *key)
173 {
174 
175 	key->frc_src.v4 = ip->ip_src;
176 	key->frc_dst.v4 = ip->ip_dst;
177 	key->frc_af = AF_INET;
178 	key->frc_proto = ip->ip_p;
179 	key->frc_id = ip->ip_id;
180 }
181 #endif	/* INET */
182 
183 void
184 pf_normalize_init(void)
185 {
186 
187 	V_pf_frag_z = uma_zcreate("pf frags", sizeof(struct pf_fragment),
188 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
189 	V_pf_frent_z = uma_zcreate("pf frag entries", sizeof(struct pf_frent),
190 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
191 	V_pf_state_scrub_z = uma_zcreate("pf state scrubs",
192 	    sizeof(struct pf_state_scrub),  NULL, NULL, NULL, NULL,
193 	    UMA_ALIGN_PTR, 0);
194 
195 	mtx_init(&V_pf_frag_mtx, "pf fragments", NULL, MTX_DEF);
196 
197 	V_pf_limits[PF_LIMIT_FRAGS].zone = V_pf_frent_z;
198 	V_pf_limits[PF_LIMIT_FRAGS].limit = PFFRAG_FRENT_HIWAT;
199 	uma_zone_set_max(V_pf_frent_z, PFFRAG_FRENT_HIWAT);
200 	uma_zone_set_warning(V_pf_frent_z, "PF frag entries limit reached");
201 
202 	TAILQ_INIT(&V_pf_fragqueue);
203 }
204 
205 void
206 pf_normalize_cleanup(void)
207 {
208 
209 	uma_zdestroy(V_pf_state_scrub_z);
210 	uma_zdestroy(V_pf_frent_z);
211 	uma_zdestroy(V_pf_frag_z);
212 
213 	mtx_destroy(&V_pf_frag_mtx);
214 }
215 
216 static int
217 pf_frag_compare(struct pf_fragment *a, struct pf_fragment *b)
218 {
219 	int	diff;
220 
221 	if ((diff = a->fr_id - b->fr_id) != 0)
222 		return (diff);
223 	if ((diff = a->fr_proto - b->fr_proto) != 0)
224 		return (diff);
225 	if ((diff = a->fr_af - b->fr_af) != 0)
226 		return (diff);
227 	if ((diff = pf_addr_cmp(&a->fr_src, &b->fr_src, a->fr_af)) != 0)
228 		return (diff);
229 	if ((diff = pf_addr_cmp(&a->fr_dst, &b->fr_dst, a->fr_af)) != 0)
230 		return (diff);
231 	return (0);
232 }
233 
234 void
235 pf_purge_expired_fragments(void)
236 {
237 	u_int32_t	expire = time_uptime -
238 			    V_pf_default_rule.timeout[PFTM_FRAG];
239 
240 	pf_purge_fragments(expire);
241 }
242 
243 void
244 pf_purge_fragments(uint32_t expire)
245 {
246 	struct pf_fragment	*frag;
247 
248 	PF_FRAG_LOCK();
249 	while ((frag = TAILQ_LAST(&V_pf_fragqueue, pf_fragqueue)) != NULL) {
250 		if (frag->fr_timeout > expire)
251 			break;
252 
253 		DPFPRINTF(("expiring %d(%p)\n", frag->fr_id, frag));
254 		pf_free_fragment(frag);
255 	}
256 
257 	PF_FRAG_UNLOCK();
258 }
259 
260 /*
261  * Try to flush old fragments to make space for new ones
262  */
263 static void
264 pf_flush_fragments(void)
265 {
266 	struct pf_fragment	*frag;
267 	int			 goal;
268 
269 	PF_FRAG_ASSERT();
270 
271 	goal = uma_zone_get_cur(V_pf_frent_z) * 9 / 10;
272 	DPFPRINTF(("trying to free %d frag entriess\n", goal));
273 	while (goal < uma_zone_get_cur(V_pf_frent_z)) {
274 		frag = TAILQ_LAST(&V_pf_fragqueue, pf_fragqueue);
275 		if (frag)
276 			pf_free_fragment(frag);
277 		else
278 			break;
279 	}
280 }
281 
282 /* Frees the fragments and all associated entries */
283 static void
284 pf_free_fragment(struct pf_fragment *frag)
285 {
286 	struct pf_frent		*frent;
287 
288 	PF_FRAG_ASSERT();
289 
290 	/* Free all fragments */
291 	for (frent = TAILQ_FIRST(&frag->fr_queue); frent;
292 	    frent = TAILQ_FIRST(&frag->fr_queue)) {
293 		TAILQ_REMOVE(&frag->fr_queue, frent, fr_next);
294 
295 		m_freem(frent->fe_m);
296 		uma_zfree(V_pf_frent_z, frent);
297 	}
298 
299 	pf_remove_fragment(frag);
300 }
301 
302 static struct pf_fragment *
303 pf_find_fragment(struct pf_fragment_cmp *key, struct pf_frag_tree *tree)
304 {
305 	struct pf_fragment	*frag;
306 
307 	PF_FRAG_ASSERT();
308 
309 	frag = RB_FIND(pf_frag_tree, tree, (struct pf_fragment *)key);
310 	if (frag != NULL) {
311 		/* XXX Are we sure we want to update the timeout? */
312 		frag->fr_timeout = time_uptime;
313 		TAILQ_REMOVE(&V_pf_fragqueue, frag, frag_next);
314 		TAILQ_INSERT_HEAD(&V_pf_fragqueue, frag, frag_next);
315 	}
316 
317 	return (frag);
318 }
319 
320 /* Removes a fragment from the fragment queue and frees the fragment */
321 static void
322 pf_remove_fragment(struct pf_fragment *frag)
323 {
324 
325 	PF_FRAG_ASSERT();
326 	KASSERT(frag, ("frag != NULL"));
327 
328 	RB_REMOVE(pf_frag_tree, &V_pf_frag_tree, frag);
329 	TAILQ_REMOVE(&V_pf_fragqueue, frag, frag_next);
330 	uma_zfree(V_pf_frag_z, frag);
331 }
332 
333 static struct pf_frent *
334 pf_create_fragment(u_short *reason)
335 {
336 	struct pf_frent *frent;
337 
338 	PF_FRAG_ASSERT();
339 
340 	frent = uma_zalloc(V_pf_frent_z, M_NOWAIT);
341 	if (frent == NULL) {
342 		pf_flush_fragments();
343 		frent = uma_zalloc(V_pf_frent_z, M_NOWAIT);
344 		if (frent == NULL) {
345 			REASON_SET(reason, PFRES_MEMORY);
346 			return (NULL);
347 		}
348 	}
349 
350 	return (frent);
351 }
352 
353 /*
354  * Calculate the additional holes that were created in the fragment
355  * queue by inserting this fragment.  A fragment in the middle
356  * creates one more hole by splitting.  For each connected side,
357  * it loses one hole.
358  * Fragment entry must be in the queue when calling this function.
359  */
360 static int
361 pf_frent_holes(struct pf_frent *frent)
362 {
363 	struct pf_frent *prev = TAILQ_PREV(frent, pf_fragq, fr_next);
364 	struct pf_frent *next = TAILQ_NEXT(frent, fr_next);
365 	int holes = 1;
366 
367 	if (prev == NULL) {
368 		if (frent->fe_off == 0)
369 			holes--;
370 	} else {
371 		KASSERT(frent->fe_off != 0, ("frent->fe_off != 0"));
372 		if (frent->fe_off == prev->fe_off + prev->fe_len)
373 			holes--;
374 	}
375 	if (next == NULL) {
376 		if (!frent->fe_mff)
377 			holes--;
378 	} else {
379 		KASSERT(frent->fe_mff, ("frent->fe_mff"));
380 		if (next->fe_off == frent->fe_off + frent->fe_len)
381 			holes--;
382 	}
383 	return holes;
384 }
385 
386 static inline int
387 pf_frent_index(struct pf_frent *frent)
388 {
389 	/*
390 	 * We have an array of 16 entry points to the queue.  A full size
391 	 * 65535 octet IP packet can have 8192 fragments.  So the queue
392 	 * traversal length is at most 512 and at most 16 entry points are
393 	 * checked.  We need 128 additional bytes on a 64 bit architecture.
394 	 */
395 	CTASSERT(((u_int16_t)0xffff &~ 7) / (0x10000 / PF_FRAG_ENTRY_POINTS) ==
396 	    16 - 1);
397 	CTASSERT(((u_int16_t)0xffff >> 3) / PF_FRAG_ENTRY_POINTS == 512 - 1);
398 
399 	return frent->fe_off / (0x10000 / PF_FRAG_ENTRY_POINTS);
400 }
401 
402 static int
403 pf_frent_insert(struct pf_fragment *frag, struct pf_frent *frent,
404     struct pf_frent *prev)
405 {
406 	int index;
407 
408 	CTASSERT(PF_FRAG_ENTRY_LIMIT <= 0xff);
409 
410 	/*
411 	 * A packet has at most 65536 octets.  With 16 entry points, each one
412 	 * spawns 4096 octets.  We limit these to 64 fragments each, which
413 	 * means on average every fragment must have at least 64 octets.
414 	 */
415 	index = pf_frent_index(frent);
416 	if (frag->fr_entries[index] >= PF_FRAG_ENTRY_LIMIT)
417 		return ENOBUFS;
418 	frag->fr_entries[index]++;
419 
420 	if (prev == NULL) {
421 		TAILQ_INSERT_HEAD(&frag->fr_queue, frent, fr_next);
422 	} else {
423 		KASSERT(prev->fe_off + prev->fe_len <= frent->fe_off,
424 		    ("overlapping fragment"));
425 		TAILQ_INSERT_AFTER(&frag->fr_queue, prev, frent, fr_next);
426 	}
427 
428 	if (frag->fr_firstoff[index] == NULL) {
429 		KASSERT(prev == NULL || pf_frent_index(prev) < index,
430 		    ("prev == NULL || pf_frent_index(pref) < index"));
431 		frag->fr_firstoff[index] = frent;
432 	} else {
433 		if (frent->fe_off < frag->fr_firstoff[index]->fe_off) {
434 			KASSERT(prev == NULL || pf_frent_index(prev) < index,
435 			    ("prev == NULL || pf_frent_index(pref) < index"));
436 			frag->fr_firstoff[index] = frent;
437 		} else {
438 			KASSERT(prev != NULL, ("prev != NULL"));
439 			KASSERT(pf_frent_index(prev) == index,
440 			    ("pf_frent_index(prev) == index"));
441 		}
442 	}
443 
444 	frag->fr_holes += pf_frent_holes(frent);
445 
446 	return 0;
447 }
448 
449 void
450 pf_frent_remove(struct pf_fragment *frag, struct pf_frent *frent)
451 {
452 #ifdef INVARIANTS
453 	struct pf_frent *prev = TAILQ_PREV(frent, pf_fragq, fr_next);
454 #endif
455 	struct pf_frent *next = TAILQ_NEXT(frent, fr_next);
456 	int index;
457 
458 	frag->fr_holes -= pf_frent_holes(frent);
459 
460 	index = pf_frent_index(frent);
461 	KASSERT(frag->fr_firstoff[index] != NULL, ("frent not found"));
462 	if (frag->fr_firstoff[index]->fe_off == frent->fe_off) {
463 		if (next == NULL) {
464 			frag->fr_firstoff[index] = NULL;
465 		} else {
466 			KASSERT(frent->fe_off + frent->fe_len <= next->fe_off,
467 			    ("overlapping fragment"));
468 			if (pf_frent_index(next) == index) {
469 				frag->fr_firstoff[index] = next;
470 			} else {
471 				frag->fr_firstoff[index] = NULL;
472 			}
473 		}
474 	} else {
475 		KASSERT(frag->fr_firstoff[index]->fe_off < frent->fe_off,
476 		    ("frag->fr_firstoff[index]->fe_off < frent->fe_off"));
477 		KASSERT(prev != NULL, ("prev != NULL"));
478 		KASSERT(prev->fe_off + prev->fe_len <= frent->fe_off,
479 		    ("overlapping fragment"));
480 		KASSERT(pf_frent_index(prev) == index,
481 		    ("pf_frent_index(prev) == index"));
482 	}
483 
484 	TAILQ_REMOVE(&frag->fr_queue, frent, fr_next);
485 
486 	KASSERT(frag->fr_entries[index] > 0, ("No fragments remaining"));
487 	frag->fr_entries[index]--;
488 }
489 
490 struct pf_frent *
491 pf_frent_previous(struct pf_fragment *frag, struct pf_frent *frent)
492 {
493 	struct pf_frent *prev, *next;
494 	int index;
495 
496 	/*
497 	 * If there are no fragments after frag, take the final one.  Assume
498 	 * that the global queue is not empty.
499 	 */
500 	prev = TAILQ_LAST(&frag->fr_queue, pf_fragq);
501 	KASSERT(prev != NULL, ("prev != NULL"));
502 	if (prev->fe_off <= frent->fe_off)
503 		return prev;
504 	/*
505 	 * We want to find a fragment entry that is before frag, but still
506 	 * close to it.  Find the first fragment entry that is in the same
507 	 * entry point or in the first entry point after that.  As we have
508 	 * already checked that there are entries behind frag, this will
509 	 * succeed.
510 	 */
511 	for (index = pf_frent_index(frent); index < PF_FRAG_ENTRY_POINTS;
512 	    index++) {
513 		prev = frag->fr_firstoff[index];
514 		if (prev != NULL)
515 			break;
516 	}
517 	KASSERT(prev != NULL, ("prev != NULL"));
518 	/*
519 	 * In prev we may have a fragment from the same entry point that is
520 	 * before frent, or one that is just one position behind frent.
521 	 * In the latter case, we go back one step and have the predecessor.
522 	 * There may be none if the new fragment will be the first one.
523 	 */
524 	if (prev->fe_off > frent->fe_off) {
525 		prev = TAILQ_PREV(prev, pf_fragq, fr_next);
526 		if (prev == NULL)
527 			return NULL;
528 		KASSERT(prev->fe_off <= frent->fe_off,
529 		    ("prev->fe_off <= frent->fe_off"));
530 		return prev;
531 	}
532 	/*
533 	 * In prev is the first fragment of the entry point.  The offset
534 	 * of frag is behind it.  Find the closest previous fragment.
535 	 */
536 	for (next = TAILQ_NEXT(prev, fr_next); next != NULL;
537 	    next = TAILQ_NEXT(next, fr_next)) {
538 		if (next->fe_off > frent->fe_off)
539 			break;
540 		prev = next;
541 	}
542 	return prev;
543 }
544 
545 static struct pf_fragment *
546 pf_fillup_fragment(struct pf_fragment_cmp *key, struct pf_frent *frent,
547     u_short *reason)
548 {
549 	struct pf_frent		*after, *next, *prev;
550 	struct pf_fragment	*frag;
551 	uint16_t		total;
552 	int			old_index, new_index;
553 
554 	PF_FRAG_ASSERT();
555 
556 	/* No empty fragments. */
557 	if (frent->fe_len == 0) {
558 		DPFPRINTF(("bad fragment: len 0\n"));
559 		goto bad_fragment;
560 	}
561 
562 	/* All fragments are 8 byte aligned. */
563 	if (frent->fe_mff && (frent->fe_len & 0x7)) {
564 		DPFPRINTF(("bad fragment: mff and len %d\n", frent->fe_len));
565 		goto bad_fragment;
566 	}
567 
568 	/* Respect maximum length, IP_MAXPACKET == IPV6_MAXPACKET. */
569 	if (frent->fe_off + frent->fe_len > IP_MAXPACKET) {
570 		DPFPRINTF(("bad fragment: max packet %d\n",
571 		    frent->fe_off + frent->fe_len));
572 		goto bad_fragment;
573 	}
574 
575 	DPFPRINTF((key->frc_af == AF_INET ?
576 	    "reass frag %d @ %d-%d\n" : "reass frag %#08x @ %d-%d\n",
577 	    key->frc_id, frent->fe_off, frent->fe_off + frent->fe_len));
578 
579 	/* Fully buffer all of the fragments in this fragment queue. */
580 	frag = pf_find_fragment(key, &V_pf_frag_tree);
581 
582 	/* Create a new reassembly queue for this packet. */
583 	if (frag == NULL) {
584 		frag = uma_zalloc(V_pf_frag_z, M_NOWAIT);
585 		if (frag == NULL) {
586 			pf_flush_fragments();
587 			frag = uma_zalloc(V_pf_frag_z, M_NOWAIT);
588 			if (frag == NULL) {
589 				REASON_SET(reason, PFRES_MEMORY);
590 				goto drop_fragment;
591 			}
592 		}
593 
594 		*(struct pf_fragment_cmp *)frag = *key;
595 		memset(frag->fr_firstoff, 0, sizeof(frag->fr_firstoff));
596 		memset(frag->fr_entries, 0, sizeof(frag->fr_entries));
597 		frag->fr_timeout = time_uptime;
598 		frag->fr_maxlen = frent->fe_len;
599 		frag->fr_holes = 1;
600 		TAILQ_INIT(&frag->fr_queue);
601 
602 		RB_INSERT(pf_frag_tree, &V_pf_frag_tree, frag);
603 		TAILQ_INSERT_HEAD(&V_pf_fragqueue, frag, frag_next);
604 
605 		/* We do not have a previous fragment, cannot fail. */
606 		pf_frent_insert(frag, frent, NULL);
607 
608 		return (frag);
609 	}
610 
611 	KASSERT(!TAILQ_EMPTY(&frag->fr_queue), ("!TAILQ_EMPTY()->fr_queue"));
612 
613 	/* Remember maximum fragment len for refragmentation. */
614 	if (frent->fe_len > frag->fr_maxlen)
615 		frag->fr_maxlen = frent->fe_len;
616 
617 	/* Maximum data we have seen already. */
618 	total = TAILQ_LAST(&frag->fr_queue, pf_fragq)->fe_off +
619 		TAILQ_LAST(&frag->fr_queue, pf_fragq)->fe_len;
620 
621 	/* Non terminal fragments must have more fragments flag. */
622 	if (frent->fe_off + frent->fe_len < total && !frent->fe_mff)
623 		goto bad_fragment;
624 
625 	/* Check if we saw the last fragment already. */
626 	if (!TAILQ_LAST(&frag->fr_queue, pf_fragq)->fe_mff) {
627 		if (frent->fe_off + frent->fe_len > total ||
628 		    (frent->fe_off + frent->fe_len == total && frent->fe_mff))
629 			goto bad_fragment;
630 	} else {
631 		if (frent->fe_off + frent->fe_len == total && !frent->fe_mff)
632 			goto bad_fragment;
633 	}
634 
635 	/* Find neighbors for newly inserted fragment */
636 	prev = pf_frent_previous(frag, frent);
637 	if (prev == NULL) {
638 		after = TAILQ_FIRST(&frag->fr_queue);
639 		KASSERT(after != NULL, ("after != NULL"));
640 	} else {
641 		after = TAILQ_NEXT(prev, fr_next);
642 	}
643 
644 	if (prev != NULL && prev->fe_off + prev->fe_len > frent->fe_off) {
645 		uint16_t precut;
646 
647 		precut = prev->fe_off + prev->fe_len - frent->fe_off;
648 		if (precut >= frent->fe_len)
649 			goto bad_fragment;
650 		DPFPRINTF(("overlap -%d\n", precut));
651 		m_adj(frent->fe_m, precut);
652 		frent->fe_off += precut;
653 		frent->fe_len -= precut;
654 	}
655 
656 	for (; after != NULL && frent->fe_off + frent->fe_len > after->fe_off;
657 	    after = next) {
658 		uint16_t aftercut;
659 
660 		aftercut = frent->fe_off + frent->fe_len - after->fe_off;
661 		DPFPRINTF(("adjust overlap %d\n", aftercut));
662 		if (aftercut < after->fe_len) {
663 			m_adj(after->fe_m, aftercut);
664 			old_index = pf_frent_index(after);
665 			after->fe_off += aftercut;
666 			after->fe_len -= aftercut;
667 			new_index = pf_frent_index(after);
668 			if (old_index != new_index) {
669 				DPFPRINTF(("frag index %d, new %d",
670 				    old_index, new_index));
671 				/* Fragment switched queue as fe_off changed */
672 				after->fe_off -= aftercut;
673 				after->fe_len += aftercut;
674 				/* Remove restored fragment from old queue */
675 				pf_frent_remove(frag, after);
676 				after->fe_off += aftercut;
677 				after->fe_len -= aftercut;
678 				/* Insert into correct queue */
679 				if (pf_frent_insert(frag, after, prev)) {
680 					DPFPRINTF(
681 					    ("fragment requeue limit exceeded"));
682 					m_freem(after->fe_m);
683 					uma_zfree(V_pf_frent_z, after);
684 					/* There is not way to recover */
685 					goto bad_fragment;
686 				}
687 			}
688 			break;
689 		}
690 
691 		/* This fragment is completely overlapped, lose it. */
692 		next = TAILQ_NEXT(after, fr_next);
693 		pf_frent_remove(frag, after);
694 		m_freem(after->fe_m);
695 		uma_zfree(V_pf_frent_z, after);
696 	}
697 
698 	/* If part of the queue gets too long, there is not way to recover. */
699 	if (pf_frent_insert(frag, frent, prev)) {
700 		DPFPRINTF(("fragment queue limit exceeded\n"));
701 		goto bad_fragment;
702 	}
703 
704 	return (frag);
705 
706 bad_fragment:
707 	REASON_SET(reason, PFRES_FRAG);
708 drop_fragment:
709 	uma_zfree(V_pf_frent_z, frent);
710 	return (NULL);
711 }
712 
713 static struct mbuf *
714 pf_join_fragment(struct pf_fragment *frag)
715 {
716 	struct mbuf *m, *m2;
717 	struct pf_frent	*frent, *next;
718 
719 	frent = TAILQ_FIRST(&frag->fr_queue);
720 	next = TAILQ_NEXT(frent, fr_next);
721 
722 	m = frent->fe_m;
723 	m_adj(m, (frent->fe_hdrlen + frent->fe_len) - m->m_pkthdr.len);
724 	uma_zfree(V_pf_frent_z, frent);
725 	for (frent = next; frent != NULL; frent = next) {
726 		next = TAILQ_NEXT(frent, fr_next);
727 
728 		m2 = frent->fe_m;
729 		/* Strip off ip header. */
730 		m_adj(m2, frent->fe_hdrlen);
731 		/* Strip off any trailing bytes. */
732 		m_adj(m2, frent->fe_len - m2->m_pkthdr.len);
733 
734 		uma_zfree(V_pf_frent_z, frent);
735 		m_cat(m, m2);
736 	}
737 
738 	/* Remove from fragment queue. */
739 	pf_remove_fragment(frag);
740 
741 	return (m);
742 }
743 
744 #ifdef INET
745 static int
746 pf_reassemble(struct mbuf **m0, struct ip *ip, int dir, u_short *reason)
747 {
748 	struct mbuf		*m = *m0;
749 	struct pf_frent		*frent;
750 	struct pf_fragment	*frag;
751 	struct pf_fragment_cmp	key;
752 	uint16_t		total, hdrlen;
753 
754 	/* Get an entry for the fragment queue */
755 	if ((frent = pf_create_fragment(reason)) == NULL)
756 		return (PF_DROP);
757 
758 	frent->fe_m = m;
759 	frent->fe_hdrlen = ip->ip_hl << 2;
760 	frent->fe_extoff = 0;
761 	frent->fe_len = ntohs(ip->ip_len) - (ip->ip_hl << 2);
762 	frent->fe_off = (ntohs(ip->ip_off) & IP_OFFMASK) << 3;
763 	frent->fe_mff = ntohs(ip->ip_off) & IP_MF;
764 
765 	pf_ip2key(ip, dir, &key);
766 
767 	if ((frag = pf_fillup_fragment(&key, frent, reason)) == NULL)
768 		return (PF_DROP);
769 
770 	/* The mbuf is part of the fragment entry, no direct free or access */
771 	m = *m0 = NULL;
772 
773 	if (frag->fr_holes) {
774 		DPFPRINTF(("frag %d, holes %d\n", frag->fr_id, frag->fr_holes));
775 		return (PF_PASS);  /* drop because *m0 is NULL, no error */
776 	}
777 
778 	/* We have all the data */
779 	frent = TAILQ_FIRST(&frag->fr_queue);
780 	KASSERT(frent != NULL, ("frent != NULL"));
781 	total = TAILQ_LAST(&frag->fr_queue, pf_fragq)->fe_off +
782 		TAILQ_LAST(&frag->fr_queue, pf_fragq)->fe_len;
783 	hdrlen = frent->fe_hdrlen;
784 
785 	m = *m0 = pf_join_fragment(frag);
786 	frag = NULL;
787 
788 	if (m->m_flags & M_PKTHDR) {
789 		int plen = 0;
790 		for (m = *m0; m; m = m->m_next)
791 			plen += m->m_len;
792 		m = *m0;
793 		m->m_pkthdr.len = plen;
794 	}
795 
796 	ip = mtod(m, struct ip *);
797 	ip->ip_sum = pf_cksum_fixup(ip->ip_sum, ip->ip_len,
798 	    htons(hdrlen + total), 0);
799 	ip->ip_len = htons(hdrlen + total);
800 	ip->ip_sum = pf_cksum_fixup(ip->ip_sum, ip->ip_off,
801 	    ip->ip_off & ~(IP_MF|IP_OFFMASK), 0);
802 	ip->ip_off &= ~(IP_MF|IP_OFFMASK);
803 
804 	if (hdrlen + total > IP_MAXPACKET) {
805 		DPFPRINTF(("drop: too big: %d\n", total));
806 		ip->ip_len = 0;
807 		REASON_SET(reason, PFRES_SHORT);
808 		/* PF_DROP requires a valid mbuf *m0 in pf_test() */
809 		return (PF_DROP);
810 	}
811 
812 	DPFPRINTF(("complete: %p(%d)\n", m, ntohs(ip->ip_len)));
813 	return (PF_PASS);
814 }
815 #endif	/* INET */
816 
817 #ifdef INET6
818 static int
819 pf_reassemble6(struct mbuf **m0, struct ip6_hdr *ip6, struct ip6_frag *fraghdr,
820     uint16_t hdrlen, uint16_t extoff, u_short *reason)
821 {
822 	struct mbuf		*m = *m0;
823 	struct pf_frent		*frent;
824 	struct pf_fragment	*frag;
825 	struct pf_fragment_cmp	 key;
826 	struct m_tag		*mtag;
827 	struct pf_fragment_tag	*ftag;
828 	int			 off;
829 	uint32_t		 frag_id;
830 	uint16_t		 total, maxlen;
831 	uint8_t			 proto;
832 
833 	PF_FRAG_LOCK();
834 
835 	/* Get an entry for the fragment queue. */
836 	if ((frent = pf_create_fragment(reason)) == NULL) {
837 		PF_FRAG_UNLOCK();
838 		return (PF_DROP);
839 	}
840 
841 	frent->fe_m = m;
842 	frent->fe_hdrlen = hdrlen;
843 	frent->fe_extoff = extoff;
844 	frent->fe_len = sizeof(struct ip6_hdr) + ntohs(ip6->ip6_plen) - hdrlen;
845 	frent->fe_off = ntohs(fraghdr->ip6f_offlg & IP6F_OFF_MASK);
846 	frent->fe_mff = fraghdr->ip6f_offlg & IP6F_MORE_FRAG;
847 
848 	key.frc_src.v6 = ip6->ip6_src;
849 	key.frc_dst.v6 = ip6->ip6_dst;
850 	key.frc_af = AF_INET6;
851 	/* Only the first fragment's protocol is relevant. */
852 	key.frc_proto = 0;
853 	key.frc_id = fraghdr->ip6f_ident;
854 
855 	if ((frag = pf_fillup_fragment(&key, frent, reason)) == NULL) {
856 		PF_FRAG_UNLOCK();
857 		return (PF_DROP);
858 	}
859 
860 	/* The mbuf is part of the fragment entry, no direct free or access. */
861 	m = *m0 = NULL;
862 
863 	if (frag->fr_holes) {
864 		DPFPRINTF(("frag %d, holes %d\n", frag->fr_id,
865 		    frag->fr_holes));
866 		PF_FRAG_UNLOCK();
867 		return (PF_PASS);  /* Drop because *m0 is NULL, no error. */
868 	}
869 
870 	/* We have all the data. */
871 	frent = TAILQ_FIRST(&frag->fr_queue);
872 	KASSERT(frent != NULL, ("frent != NULL"));
873 	extoff = frent->fe_extoff;
874 	maxlen = frag->fr_maxlen;
875 	frag_id = frag->fr_id;
876 	total = TAILQ_LAST(&frag->fr_queue, pf_fragq)->fe_off +
877 		TAILQ_LAST(&frag->fr_queue, pf_fragq)->fe_len;
878 	hdrlen = frent->fe_hdrlen - sizeof(struct ip6_frag);
879 
880 	m = *m0 = pf_join_fragment(frag);
881 	frag = NULL;
882 
883 	PF_FRAG_UNLOCK();
884 
885 	/* Take protocol from first fragment header. */
886 	m = m_getptr(m, hdrlen + offsetof(struct ip6_frag, ip6f_nxt), &off);
887 	KASSERT(m, ("%s: short mbuf chain", __func__));
888 	proto = *(mtod(m, caddr_t) + off);
889 	m = *m0;
890 
891 	/* Delete frag6 header */
892 	if (ip6_deletefraghdr(m, hdrlen, M_NOWAIT) != 0)
893 		goto fail;
894 
895 	if (m->m_flags & M_PKTHDR) {
896 		int plen = 0;
897 		for (m = *m0; m; m = m->m_next)
898 			plen += m->m_len;
899 		m = *m0;
900 		m->m_pkthdr.len = plen;
901 	}
902 
903 	if ((mtag = m_tag_get(PF_REASSEMBLED, sizeof(struct pf_fragment_tag),
904 	    M_NOWAIT)) == NULL)
905 		goto fail;
906 	ftag = (struct pf_fragment_tag *)(mtag + 1);
907 	ftag->ft_hdrlen = hdrlen;
908 	ftag->ft_extoff = extoff;
909 	ftag->ft_maxlen = maxlen;
910 	ftag->ft_id = frag_id;
911 	m_tag_prepend(m, mtag);
912 
913 	ip6 = mtod(m, struct ip6_hdr *);
914 	ip6->ip6_plen = htons(hdrlen - sizeof(struct ip6_hdr) + total);
915 	if (extoff) {
916 		/* Write protocol into next field of last extension header. */
917 		m = m_getptr(m, extoff + offsetof(struct ip6_ext, ip6e_nxt),
918 		    &off);
919 		KASSERT(m, ("%s: short mbuf chain", __func__));
920 		*(mtod(m, char *) + off) = proto;
921 		m = *m0;
922 	} else
923 		ip6->ip6_nxt = proto;
924 
925 	if (hdrlen - sizeof(struct ip6_hdr) + total > IPV6_MAXPACKET) {
926 		DPFPRINTF(("drop: too big: %d\n", total));
927 		ip6->ip6_plen = 0;
928 		REASON_SET(reason, PFRES_SHORT);
929 		/* PF_DROP requires a valid mbuf *m0 in pf_test6(). */
930 		return (PF_DROP);
931 	}
932 
933 	DPFPRINTF(("complete: %p(%d)\n", m, ntohs(ip6->ip6_plen)));
934 	return (PF_PASS);
935 
936 fail:
937 	REASON_SET(reason, PFRES_MEMORY);
938 	/* PF_DROP requires a valid mbuf *m0 in pf_test6(), will free later. */
939 	return (PF_DROP);
940 }
941 #endif	/* INET6 */
942 
943 #ifdef INET6
944 int
945 pf_refragment6(struct ifnet *ifp, struct mbuf **m0, struct m_tag *mtag)
946 {
947 	struct mbuf		*m = *m0, *t;
948 	struct pf_fragment_tag	*ftag = (struct pf_fragment_tag *)(mtag + 1);
949 	struct pf_pdesc		 pd;
950 	uint32_t		 frag_id;
951 	uint16_t		 hdrlen, extoff, maxlen;
952 	uint8_t			 proto;
953 	int			 error, action;
954 
955 	hdrlen = ftag->ft_hdrlen;
956 	extoff = ftag->ft_extoff;
957 	maxlen = ftag->ft_maxlen;
958 	frag_id = ftag->ft_id;
959 	m_tag_delete(m, mtag);
960 	mtag = NULL;
961 	ftag = NULL;
962 
963 	if (extoff) {
964 		int off;
965 
966 		/* Use protocol from next field of last extension header */
967 		m = m_getptr(m, extoff + offsetof(struct ip6_ext, ip6e_nxt),
968 		    &off);
969 		KASSERT((m != NULL), ("pf_refragment6: short mbuf chain"));
970 		proto = *(mtod(m, caddr_t) + off);
971 		*(mtod(m, char *) + off) = IPPROTO_FRAGMENT;
972 		m = *m0;
973 	} else {
974 		struct ip6_hdr *hdr;
975 
976 		hdr = mtod(m, struct ip6_hdr *);
977 		proto = hdr->ip6_nxt;
978 		hdr->ip6_nxt = IPPROTO_FRAGMENT;
979 	}
980 
981 	/* The MTU must be a multiple of 8 bytes, or we risk doing the
982 	 * fragmentation wrong. */
983 	maxlen = maxlen & ~7;
984 
985 	/*
986 	 * Maxlen may be less than 8 if there was only a single
987 	 * fragment.  As it was fragmented before, add a fragment
988 	 * header also for a single fragment.  If total or maxlen
989 	 * is less than 8, ip6_fragment() will return EMSGSIZE and
990 	 * we drop the packet.
991 	 */
992 	error = ip6_fragment(ifp, m, hdrlen, proto, maxlen, frag_id);
993 	m = (*m0)->m_nextpkt;
994 	(*m0)->m_nextpkt = NULL;
995 	if (error == 0) {
996 		/* The first mbuf contains the unfragmented packet. */
997 		m_freem(*m0);
998 		*m0 = NULL;
999 		action = PF_PASS;
1000 	} else {
1001 		/* Drop expects an mbuf to free. */
1002 		DPFPRINTF(("refragment error %d\n", error));
1003 		action = PF_DROP;
1004 	}
1005 	for (t = m; m; m = t) {
1006 		t = m->m_nextpkt;
1007 		m->m_nextpkt = NULL;
1008 		m->m_flags |= M_SKIP_FIREWALL;
1009 		memset(&pd, 0, sizeof(pd));
1010 		pd.pf_mtag = pf_find_mtag(m);
1011 		if (error == 0)
1012 			ip6_forward(m, 0);
1013 		else
1014 			m_freem(m);
1015 	}
1016 
1017 	return (action);
1018 }
1019 #endif /* INET6 */
1020 
1021 #ifdef INET
1022 int
1023 pf_normalize_ip(struct mbuf **m0, int dir, struct pfi_kkif *kif, u_short *reason,
1024     struct pf_pdesc *pd)
1025 {
1026 	struct mbuf		*m = *m0;
1027 	struct pf_krule		*r;
1028 	struct ip		*h = mtod(m, struct ip *);
1029 	int			 mff = (ntohs(h->ip_off) & IP_MF);
1030 	int			 hlen = h->ip_hl << 2;
1031 	u_int16_t		 fragoff = (ntohs(h->ip_off) & IP_OFFMASK) << 3;
1032 	u_int16_t		 max;
1033 	int			 ip_len;
1034 	int			 ip_off;
1035 	int			 tag = -1;
1036 	int			 verdict;
1037 
1038 	PF_RULES_RASSERT();
1039 
1040 	r = TAILQ_FIRST(pf_main_ruleset.rules[PF_RULESET_SCRUB].active.ptr);
1041 	while (r != NULL) {
1042 		pf_counter_u64_add(&r->evaluations, 1);
1043 		if (pfi_kkif_match(r->kif, kif) == r->ifnot)
1044 			r = r->skip[PF_SKIP_IFP].ptr;
1045 		else if (r->direction && r->direction != dir)
1046 			r = r->skip[PF_SKIP_DIR].ptr;
1047 		else if (r->af && r->af != AF_INET)
1048 			r = r->skip[PF_SKIP_AF].ptr;
1049 		else if (r->proto && r->proto != h->ip_p)
1050 			r = r->skip[PF_SKIP_PROTO].ptr;
1051 		else if (PF_MISMATCHAW(&r->src.addr,
1052 		    (struct pf_addr *)&h->ip_src.s_addr, AF_INET,
1053 		    r->src.neg, kif, M_GETFIB(m)))
1054 			r = r->skip[PF_SKIP_SRC_ADDR].ptr;
1055 		else if (PF_MISMATCHAW(&r->dst.addr,
1056 		    (struct pf_addr *)&h->ip_dst.s_addr, AF_INET,
1057 		    r->dst.neg, NULL, M_GETFIB(m)))
1058 			r = r->skip[PF_SKIP_DST_ADDR].ptr;
1059 		else if (r->match_tag && !pf_match_tag(m, r, &tag,
1060 		    pd->pf_mtag ? pd->pf_mtag->tag : 0))
1061 			r = TAILQ_NEXT(r, entries);
1062 		else
1063 			break;
1064 	}
1065 
1066 	if (r == NULL || r->action == PF_NOSCRUB)
1067 		return (PF_PASS);
1068 
1069 	pf_counter_u64_critical_enter();
1070 	pf_counter_u64_add_protected(&r->packets[dir == PF_OUT], 1);
1071 	pf_counter_u64_add_protected(&r->bytes[dir == PF_OUT], pd->tot_len);
1072 	pf_counter_u64_critical_exit();
1073 
1074 	/* Check for illegal packets */
1075 	if (hlen < (int)sizeof(struct ip)) {
1076 		REASON_SET(reason, PFRES_NORM);
1077 		goto drop;
1078 	}
1079 
1080 	if (hlen > ntohs(h->ip_len)) {
1081 		REASON_SET(reason, PFRES_NORM);
1082 		goto drop;
1083 	}
1084 
1085 	/* Clear IP_DF if the rule uses the no-df option */
1086 	if (r->rule_flag & PFRULE_NODF && h->ip_off & htons(IP_DF)) {
1087 		u_int16_t ip_off = h->ip_off;
1088 
1089 		h->ip_off &= htons(~IP_DF);
1090 		h->ip_sum = pf_cksum_fixup(h->ip_sum, ip_off, h->ip_off, 0);
1091 	}
1092 
1093 	/* We will need other tests here */
1094 	if (!fragoff && !mff)
1095 		goto no_fragment;
1096 
1097 	/* We're dealing with a fragment now. Don't allow fragments
1098 	 * with IP_DF to enter the cache. If the flag was cleared by
1099 	 * no-df above, fine. Otherwise drop it.
1100 	 */
1101 	if (h->ip_off & htons(IP_DF)) {
1102 		DPFPRINTF(("IP_DF\n"));
1103 		goto bad;
1104 	}
1105 
1106 	ip_len = ntohs(h->ip_len) - hlen;
1107 	ip_off = (ntohs(h->ip_off) & IP_OFFMASK) << 3;
1108 
1109 	/* All fragments are 8 byte aligned */
1110 	if (mff && (ip_len & 0x7)) {
1111 		DPFPRINTF(("mff and %d\n", ip_len));
1112 		goto bad;
1113 	}
1114 
1115 	/* Respect maximum length */
1116 	if (fragoff + ip_len > IP_MAXPACKET) {
1117 		DPFPRINTF(("max packet %d\n", fragoff + ip_len));
1118 		goto bad;
1119 	}
1120 	max = fragoff + ip_len;
1121 
1122 	/* Fully buffer all of the fragments
1123 	 * Might return a completely reassembled mbuf, or NULL */
1124 	PF_FRAG_LOCK();
1125 	DPFPRINTF(("reass frag %d @ %d-%d\n", h->ip_id, fragoff, max));
1126 	verdict = pf_reassemble(m0, h, dir, reason);
1127 	PF_FRAG_UNLOCK();
1128 
1129 	if (verdict != PF_PASS)
1130 		return (PF_DROP);
1131 
1132 	m = *m0;
1133 	if (m == NULL)
1134 		return (PF_DROP);
1135 
1136 	h = mtod(m, struct ip *);
1137 
1138  no_fragment:
1139 	/* At this point, only IP_DF is allowed in ip_off */
1140 	if (h->ip_off & ~htons(IP_DF)) {
1141 		u_int16_t ip_off = h->ip_off;
1142 
1143 		h->ip_off &= htons(IP_DF);
1144 		h->ip_sum = pf_cksum_fixup(h->ip_sum, ip_off, h->ip_off, 0);
1145 	}
1146 
1147 	pf_scrub_ip(&m, r->rule_flag, r->min_ttl, r->set_tos);
1148 
1149 	return (PF_PASS);
1150 
1151  bad:
1152 	DPFPRINTF(("dropping bad fragment\n"));
1153 	REASON_SET(reason, PFRES_FRAG);
1154  drop:
1155 	if (r != NULL && r->log)
1156 		PFLOG_PACKET(kif, m, AF_INET, dir, *reason, r, NULL, NULL, pd,
1157 		    1);
1158 
1159 	return (PF_DROP);
1160 }
1161 #endif
1162 
1163 #ifdef INET6
1164 int
1165 pf_normalize_ip6(struct mbuf **m0, int dir, struct pfi_kkif *kif,
1166     u_short *reason, struct pf_pdesc *pd)
1167 {
1168 	struct mbuf		*m = *m0;
1169 	struct pf_krule		*r;
1170 	struct ip6_hdr		*h = mtod(m, struct ip6_hdr *);
1171 	int			 extoff;
1172 	int			 off;
1173 	struct ip6_ext		 ext;
1174 	struct ip6_opt		 opt;
1175 	struct ip6_frag		 frag;
1176 	u_int32_t		 plen;
1177 	int			 optend;
1178 	int			 ooff;
1179 	u_int8_t		 proto;
1180 	int			 terminal;
1181 
1182 	PF_RULES_RASSERT();
1183 
1184 	r = TAILQ_FIRST(pf_main_ruleset.rules[PF_RULESET_SCRUB].active.ptr);
1185 	while (r != NULL) {
1186 		pf_counter_u64_add(&r->evaluations, 1);
1187 		if (pfi_kkif_match(r->kif, kif) == r->ifnot)
1188 			r = r->skip[PF_SKIP_IFP].ptr;
1189 		else if (r->direction && r->direction != dir)
1190 			r = r->skip[PF_SKIP_DIR].ptr;
1191 		else if (r->af && r->af != AF_INET6)
1192 			r = r->skip[PF_SKIP_AF].ptr;
1193 #if 0 /* header chain! */
1194 		else if (r->proto && r->proto != h->ip6_nxt)
1195 			r = r->skip[PF_SKIP_PROTO].ptr;
1196 #endif
1197 		else if (PF_MISMATCHAW(&r->src.addr,
1198 		    (struct pf_addr *)&h->ip6_src, AF_INET6,
1199 		    r->src.neg, kif, M_GETFIB(m)))
1200 			r = r->skip[PF_SKIP_SRC_ADDR].ptr;
1201 		else if (PF_MISMATCHAW(&r->dst.addr,
1202 		    (struct pf_addr *)&h->ip6_dst, AF_INET6,
1203 		    r->dst.neg, NULL, M_GETFIB(m)))
1204 			r = r->skip[PF_SKIP_DST_ADDR].ptr;
1205 		else
1206 			break;
1207 	}
1208 
1209 	if (r == NULL || r->action == PF_NOSCRUB)
1210 		return (PF_PASS);
1211 
1212 	pf_counter_u64_critical_enter();
1213 	pf_counter_u64_add_protected(&r->packets[dir == PF_OUT], 1);
1214 	pf_counter_u64_add_protected(&r->bytes[dir == PF_OUT], pd->tot_len);
1215 	pf_counter_u64_critical_exit();
1216 
1217 	/* Check for illegal packets */
1218 	if (sizeof(struct ip6_hdr) + IPV6_MAXPACKET < m->m_pkthdr.len)
1219 		goto drop;
1220 
1221 	plen = ntohs(h->ip6_plen);
1222 	/* jumbo payload option not supported */
1223 	if (plen == 0)
1224 		goto drop;
1225 
1226 	extoff = 0;
1227 	off = sizeof(struct ip6_hdr);
1228 	proto = h->ip6_nxt;
1229 	terminal = 0;
1230 	do {
1231 		switch (proto) {
1232 		case IPPROTO_FRAGMENT:
1233 			goto fragment;
1234 			break;
1235 		case IPPROTO_AH:
1236 		case IPPROTO_ROUTING:
1237 		case IPPROTO_DSTOPTS:
1238 			if (!pf_pull_hdr(m, off, &ext, sizeof(ext), NULL,
1239 			    NULL, AF_INET6))
1240 				goto shortpkt;
1241 			extoff = off;
1242 			if (proto == IPPROTO_AH)
1243 				off += (ext.ip6e_len + 2) * 4;
1244 			else
1245 				off += (ext.ip6e_len + 1) * 8;
1246 			proto = ext.ip6e_nxt;
1247 			break;
1248 		case IPPROTO_HOPOPTS:
1249 			if (!pf_pull_hdr(m, off, &ext, sizeof(ext), NULL,
1250 			    NULL, AF_INET6))
1251 				goto shortpkt;
1252 			extoff = off;
1253 			optend = off + (ext.ip6e_len + 1) * 8;
1254 			ooff = off + sizeof(ext);
1255 			do {
1256 				if (!pf_pull_hdr(m, ooff, &opt.ip6o_type,
1257 				    sizeof(opt.ip6o_type), NULL, NULL,
1258 				    AF_INET6))
1259 					goto shortpkt;
1260 				if (opt.ip6o_type == IP6OPT_PAD1) {
1261 					ooff++;
1262 					continue;
1263 				}
1264 				if (!pf_pull_hdr(m, ooff, &opt, sizeof(opt),
1265 				    NULL, NULL, AF_INET6))
1266 					goto shortpkt;
1267 				if (ooff + sizeof(opt) + opt.ip6o_len > optend)
1268 					goto drop;
1269 				if (opt.ip6o_type == IP6OPT_JUMBO)
1270 					goto drop;
1271 				ooff += sizeof(opt) + opt.ip6o_len;
1272 			} while (ooff < optend);
1273 
1274 			off = optend;
1275 			proto = ext.ip6e_nxt;
1276 			break;
1277 		default:
1278 			terminal = 1;
1279 			break;
1280 		}
1281 	} while (!terminal);
1282 
1283 	if (sizeof(struct ip6_hdr) + plen > m->m_pkthdr.len)
1284 		goto shortpkt;
1285 
1286 	pf_scrub_ip6(&m, r->min_ttl);
1287 
1288 	return (PF_PASS);
1289 
1290  fragment:
1291 	if (sizeof(struct ip6_hdr) + plen > m->m_pkthdr.len)
1292 		goto shortpkt;
1293 
1294 	if (!pf_pull_hdr(m, off, &frag, sizeof(frag), NULL, NULL, AF_INET6))
1295 		goto shortpkt;
1296 
1297 	/* Offset now points to data portion. */
1298 	off += sizeof(frag);
1299 
1300 	/* Returns PF_DROP or *m0 is NULL or completely reassembled mbuf. */
1301 	if (pf_reassemble6(m0, h, &frag, off, extoff, reason) != PF_PASS)
1302 		return (PF_DROP);
1303 	m = *m0;
1304 	if (m == NULL)
1305 		return (PF_DROP);
1306 
1307 	pd->flags |= PFDESC_IP_REAS;
1308 	return (PF_PASS);
1309 
1310  shortpkt:
1311 	REASON_SET(reason, PFRES_SHORT);
1312 	if (r != NULL && r->log)
1313 		PFLOG_PACKET(kif, m, AF_INET6, dir, *reason, r, NULL, NULL, pd,
1314 		    1);
1315 	return (PF_DROP);
1316 
1317  drop:
1318 	REASON_SET(reason, PFRES_NORM);
1319 	if (r != NULL && r->log)
1320 		PFLOG_PACKET(kif, m, AF_INET6, dir, *reason, r, NULL, NULL, pd,
1321 		    1);
1322 	return (PF_DROP);
1323 }
1324 #endif /* INET6 */
1325 
1326 int
1327 pf_normalize_tcp(int dir, struct pfi_kkif *kif, struct mbuf *m, int ipoff,
1328     int off, void *h, struct pf_pdesc *pd)
1329 {
1330 	struct pf_krule	*r, *rm = NULL;
1331 	struct tcphdr	*th = &pd->hdr.tcp;
1332 	int		 rewrite = 0;
1333 	u_short		 reason;
1334 	u_int8_t	 flags;
1335 	sa_family_t	 af = pd->af;
1336 
1337 	PF_RULES_RASSERT();
1338 
1339 	r = TAILQ_FIRST(pf_main_ruleset.rules[PF_RULESET_SCRUB].active.ptr);
1340 	while (r != NULL) {
1341 		pf_counter_u64_add(&r->evaluations, 1);
1342 		if (pfi_kkif_match(r->kif, kif) == r->ifnot)
1343 			r = r->skip[PF_SKIP_IFP].ptr;
1344 		else if (r->direction && r->direction != dir)
1345 			r = r->skip[PF_SKIP_DIR].ptr;
1346 		else if (r->af && r->af != af)
1347 			r = r->skip[PF_SKIP_AF].ptr;
1348 		else if (r->proto && r->proto != pd->proto)
1349 			r = r->skip[PF_SKIP_PROTO].ptr;
1350 		else if (PF_MISMATCHAW(&r->src.addr, pd->src, af,
1351 		    r->src.neg, kif, M_GETFIB(m)))
1352 			r = r->skip[PF_SKIP_SRC_ADDR].ptr;
1353 		else if (r->src.port_op && !pf_match_port(r->src.port_op,
1354 			    r->src.port[0], r->src.port[1], th->th_sport))
1355 			r = r->skip[PF_SKIP_SRC_PORT].ptr;
1356 		else if (PF_MISMATCHAW(&r->dst.addr, pd->dst, af,
1357 		    r->dst.neg, NULL, M_GETFIB(m)))
1358 			r = r->skip[PF_SKIP_DST_ADDR].ptr;
1359 		else if (r->dst.port_op && !pf_match_port(r->dst.port_op,
1360 			    r->dst.port[0], r->dst.port[1], th->th_dport))
1361 			r = r->skip[PF_SKIP_DST_PORT].ptr;
1362 		else if (r->os_fingerprint != PF_OSFP_ANY && !pf_osfp_match(
1363 			    pf_osfp_fingerprint(pd, m, off, th),
1364 			    r->os_fingerprint))
1365 			r = TAILQ_NEXT(r, entries);
1366 		else {
1367 			rm = r;
1368 			break;
1369 		}
1370 	}
1371 
1372 	if (rm == NULL || rm->action == PF_NOSCRUB)
1373 		return (PF_PASS);
1374 
1375 	pf_counter_u64_critical_enter();
1376 	pf_counter_u64_add_protected(&r->packets[dir == PF_OUT], 1);
1377 	pf_counter_u64_add_protected(&r->bytes[dir == PF_OUT], pd->tot_len);
1378 	pf_counter_u64_critical_exit();
1379 
1380 	if (rm->rule_flag & PFRULE_REASSEMBLE_TCP)
1381 		pd->flags |= PFDESC_TCP_NORM;
1382 
1383 	flags = th->th_flags;
1384 	if (flags & TH_SYN) {
1385 		/* Illegal packet */
1386 		if (flags & TH_RST)
1387 			goto tcp_drop;
1388 
1389 		if (flags & TH_FIN)
1390 			goto tcp_drop;
1391 	} else {
1392 		/* Illegal packet */
1393 		if (!(flags & (TH_ACK|TH_RST)))
1394 			goto tcp_drop;
1395 	}
1396 
1397 	if (!(flags & TH_ACK)) {
1398 		/* These flags are only valid if ACK is set */
1399 		if ((flags & TH_FIN) || (flags & TH_PUSH) || (flags & TH_URG))
1400 			goto tcp_drop;
1401 	}
1402 
1403 	/* Check for illegal header length */
1404 	if (th->th_off < (sizeof(struct tcphdr) >> 2))
1405 		goto tcp_drop;
1406 
1407 	/* If flags changed, or reserved data set, then adjust */
1408 	if (flags != th->th_flags || th->th_x2 != 0) {
1409 		u_int16_t	ov, nv;
1410 
1411 		ov = *(u_int16_t *)(&th->th_ack + 1);
1412 		th->th_flags = flags;
1413 		th->th_x2 = 0;
1414 		nv = *(u_int16_t *)(&th->th_ack + 1);
1415 
1416 		th->th_sum = pf_proto_cksum_fixup(m, th->th_sum, ov, nv, 0);
1417 		rewrite = 1;
1418 	}
1419 
1420 	/* Remove urgent pointer, if TH_URG is not set */
1421 	if (!(flags & TH_URG) && th->th_urp) {
1422 		th->th_sum = pf_proto_cksum_fixup(m, th->th_sum, th->th_urp,
1423 		    0, 0);
1424 		th->th_urp = 0;
1425 		rewrite = 1;
1426 	}
1427 
1428 	/* Process options */
1429 	if (r->max_mss && pf_normalize_tcpopt(r, m, th, off, pd->af))
1430 		rewrite = 1;
1431 
1432 	/* copy back packet headers if we sanitized */
1433 	if (rewrite)
1434 		m_copyback(m, off, sizeof(*th), (caddr_t)th);
1435 
1436 	return (PF_PASS);
1437 
1438  tcp_drop:
1439 	REASON_SET(&reason, PFRES_NORM);
1440 	if (rm != NULL && r->log)
1441 		PFLOG_PACKET(kif, m, AF_INET, dir, reason, r, NULL, NULL, pd,
1442 		    1);
1443 	return (PF_DROP);
1444 }
1445 
1446 int
1447 pf_normalize_tcp_init(struct mbuf *m, int off, struct pf_pdesc *pd,
1448     struct tcphdr *th, struct pf_state_peer *src, struct pf_state_peer *dst)
1449 {
1450 	u_int32_t tsval, tsecr;
1451 	u_int8_t hdr[60];
1452 	u_int8_t *opt;
1453 
1454 	KASSERT((src->scrub == NULL),
1455 	    ("pf_normalize_tcp_init: src->scrub != NULL"));
1456 
1457 	src->scrub = uma_zalloc(V_pf_state_scrub_z, M_ZERO | M_NOWAIT);
1458 	if (src->scrub == NULL)
1459 		return (1);
1460 
1461 	switch (pd->af) {
1462 #ifdef INET
1463 	case AF_INET: {
1464 		struct ip *h = mtod(m, struct ip *);
1465 		src->scrub->pfss_ttl = h->ip_ttl;
1466 		break;
1467 	}
1468 #endif /* INET */
1469 #ifdef INET6
1470 	case AF_INET6: {
1471 		struct ip6_hdr *h = mtod(m, struct ip6_hdr *);
1472 		src->scrub->pfss_ttl = h->ip6_hlim;
1473 		break;
1474 	}
1475 #endif /* INET6 */
1476 	}
1477 
1478 	/*
1479 	 * All normalizations below are only begun if we see the start of
1480 	 * the connections.  They must all set an enabled bit in pfss_flags
1481 	 */
1482 	if ((th->th_flags & TH_SYN) == 0)
1483 		return (0);
1484 
1485 	if (th->th_off > (sizeof(struct tcphdr) >> 2) && src->scrub &&
1486 	    pf_pull_hdr(m, off, hdr, th->th_off << 2, NULL, NULL, pd->af)) {
1487 		/* Diddle with TCP options */
1488 		int hlen;
1489 		opt = hdr + sizeof(struct tcphdr);
1490 		hlen = (th->th_off << 2) - sizeof(struct tcphdr);
1491 		while (hlen >= TCPOLEN_TIMESTAMP) {
1492 			switch (*opt) {
1493 			case TCPOPT_EOL:	/* FALLTHROUGH */
1494 			case TCPOPT_NOP:
1495 				opt++;
1496 				hlen--;
1497 				break;
1498 			case TCPOPT_TIMESTAMP:
1499 				if (opt[1] >= TCPOLEN_TIMESTAMP) {
1500 					src->scrub->pfss_flags |=
1501 					    PFSS_TIMESTAMP;
1502 					src->scrub->pfss_ts_mod =
1503 					    htonl(arc4random());
1504 
1505 					/* note PFSS_PAWS not set yet */
1506 					memcpy(&tsval, &opt[2],
1507 					    sizeof(u_int32_t));
1508 					memcpy(&tsecr, &opt[6],
1509 					    sizeof(u_int32_t));
1510 					src->scrub->pfss_tsval0 = ntohl(tsval);
1511 					src->scrub->pfss_tsval = ntohl(tsval);
1512 					src->scrub->pfss_tsecr = ntohl(tsecr);
1513 					getmicrouptime(&src->scrub->pfss_last);
1514 				}
1515 				/* FALLTHROUGH */
1516 			default:
1517 				hlen -= MAX(opt[1], 2);
1518 				opt += MAX(opt[1], 2);
1519 				break;
1520 			}
1521 		}
1522 	}
1523 
1524 	return (0);
1525 }
1526 
1527 void
1528 pf_normalize_tcp_cleanup(struct pf_kstate *state)
1529 {
1530 	if (state->src.scrub)
1531 		uma_zfree(V_pf_state_scrub_z, state->src.scrub);
1532 	if (state->dst.scrub)
1533 		uma_zfree(V_pf_state_scrub_z, state->dst.scrub);
1534 
1535 	/* Someday... flush the TCP segment reassembly descriptors. */
1536 }
1537 
1538 int
1539 pf_normalize_tcp_stateful(struct mbuf *m, int off, struct pf_pdesc *pd,
1540     u_short *reason, struct tcphdr *th, struct pf_kstate *state,
1541     struct pf_state_peer *src, struct pf_state_peer *dst, int *writeback)
1542 {
1543 	struct timeval uptime;
1544 	u_int32_t tsval, tsecr;
1545 	u_int tsval_from_last;
1546 	u_int8_t hdr[60];
1547 	u_int8_t *opt;
1548 	int copyback = 0;
1549 	int got_ts = 0;
1550 	size_t startoff;
1551 
1552 	KASSERT((src->scrub || dst->scrub),
1553 	    ("%s: src->scrub && dst->scrub!", __func__));
1554 
1555 	/*
1556 	 * Enforce the minimum TTL seen for this connection.  Negate a common
1557 	 * technique to evade an intrusion detection system and confuse
1558 	 * firewall state code.
1559 	 */
1560 	switch (pd->af) {
1561 #ifdef INET
1562 	case AF_INET: {
1563 		if (src->scrub) {
1564 			struct ip *h = mtod(m, struct ip *);
1565 			if (h->ip_ttl > src->scrub->pfss_ttl)
1566 				src->scrub->pfss_ttl = h->ip_ttl;
1567 			h->ip_ttl = src->scrub->pfss_ttl;
1568 		}
1569 		break;
1570 	}
1571 #endif /* INET */
1572 #ifdef INET6
1573 	case AF_INET6: {
1574 		if (src->scrub) {
1575 			struct ip6_hdr *h = mtod(m, struct ip6_hdr *);
1576 			if (h->ip6_hlim > src->scrub->pfss_ttl)
1577 				src->scrub->pfss_ttl = h->ip6_hlim;
1578 			h->ip6_hlim = src->scrub->pfss_ttl;
1579 		}
1580 		break;
1581 	}
1582 #endif /* INET6 */
1583 	}
1584 
1585 	if (th->th_off > (sizeof(struct tcphdr) >> 2) &&
1586 	    ((src->scrub && (src->scrub->pfss_flags & PFSS_TIMESTAMP)) ||
1587 	    (dst->scrub && (dst->scrub->pfss_flags & PFSS_TIMESTAMP))) &&
1588 	    pf_pull_hdr(m, off, hdr, th->th_off << 2, NULL, NULL, pd->af)) {
1589 		/* Diddle with TCP options */
1590 		int hlen;
1591 		opt = hdr + sizeof(struct tcphdr);
1592 		hlen = (th->th_off << 2) - sizeof(struct tcphdr);
1593 		while (hlen >= TCPOLEN_TIMESTAMP) {
1594 			startoff = opt - (hdr + sizeof(struct tcphdr));
1595 			switch (*opt) {
1596 			case TCPOPT_EOL:	/* FALLTHROUGH */
1597 			case TCPOPT_NOP:
1598 				opt++;
1599 				hlen--;
1600 				break;
1601 			case TCPOPT_TIMESTAMP:
1602 				/* Modulate the timestamps.  Can be used for
1603 				 * NAT detection, OS uptime determination or
1604 				 * reboot detection.
1605 				 */
1606 
1607 				if (got_ts) {
1608 					/* Huh?  Multiple timestamps!? */
1609 					if (V_pf_status.debug >= PF_DEBUG_MISC) {
1610 						DPFPRINTF(("multiple TS??\n"));
1611 						pf_print_state(state);
1612 						printf("\n");
1613 					}
1614 					REASON_SET(reason, PFRES_TS);
1615 					return (PF_DROP);
1616 				}
1617 				if (opt[1] >= TCPOLEN_TIMESTAMP) {
1618 					memcpy(&tsval, &opt[2],
1619 					    sizeof(u_int32_t));
1620 					if (tsval && src->scrub &&
1621 					    (src->scrub->pfss_flags &
1622 					    PFSS_TIMESTAMP)) {
1623 						tsval = ntohl(tsval);
1624 						pf_patch_32_unaligned(m,
1625 						    &th->th_sum,
1626 						    &opt[2],
1627 						    htonl(tsval +
1628 						    src->scrub->pfss_ts_mod),
1629 						    PF_ALGNMNT(startoff),
1630 						    0);
1631 						copyback = 1;
1632 					}
1633 
1634 					/* Modulate TS reply iff valid (!0) */
1635 					memcpy(&tsecr, &opt[6],
1636 					    sizeof(u_int32_t));
1637 					if (tsecr && dst->scrub &&
1638 					    (dst->scrub->pfss_flags &
1639 					    PFSS_TIMESTAMP)) {
1640 						tsecr = ntohl(tsecr)
1641 						    - dst->scrub->pfss_ts_mod;
1642 						pf_patch_32_unaligned(m,
1643 						    &th->th_sum,
1644 						    &opt[6],
1645 						    htonl(tsecr),
1646 						    PF_ALGNMNT(startoff),
1647 						    0);
1648 						copyback = 1;
1649 					}
1650 					got_ts = 1;
1651 				}
1652 				/* FALLTHROUGH */
1653 			default:
1654 				hlen -= MAX(opt[1], 2);
1655 				opt += MAX(opt[1], 2);
1656 				break;
1657 			}
1658 		}
1659 		if (copyback) {
1660 			/* Copyback the options, caller copys back header */
1661 			*writeback = 1;
1662 			m_copyback(m, off + sizeof(struct tcphdr),
1663 			    (th->th_off << 2) - sizeof(struct tcphdr), hdr +
1664 			    sizeof(struct tcphdr));
1665 		}
1666 	}
1667 
1668 	/*
1669 	 * Must invalidate PAWS checks on connections idle for too long.
1670 	 * The fastest allowed timestamp clock is 1ms.  That turns out to
1671 	 * be about 24 days before it wraps.  XXX Right now our lowerbound
1672 	 * TS echo check only works for the first 12 days of a connection
1673 	 * when the TS has exhausted half its 32bit space
1674 	 */
1675 #define TS_MAX_IDLE	(24*24*60*60)
1676 #define TS_MAX_CONN	(12*24*60*60)	/* XXX remove when better tsecr check */
1677 
1678 	getmicrouptime(&uptime);
1679 	if (src->scrub && (src->scrub->pfss_flags & PFSS_PAWS) &&
1680 	    (uptime.tv_sec - src->scrub->pfss_last.tv_sec > TS_MAX_IDLE ||
1681 	    time_uptime - state->creation > TS_MAX_CONN))  {
1682 		if (V_pf_status.debug >= PF_DEBUG_MISC) {
1683 			DPFPRINTF(("src idled out of PAWS\n"));
1684 			pf_print_state(state);
1685 			printf("\n");
1686 		}
1687 		src->scrub->pfss_flags = (src->scrub->pfss_flags & ~PFSS_PAWS)
1688 		    | PFSS_PAWS_IDLED;
1689 	}
1690 	if (dst->scrub && (dst->scrub->pfss_flags & PFSS_PAWS) &&
1691 	    uptime.tv_sec - dst->scrub->pfss_last.tv_sec > TS_MAX_IDLE) {
1692 		if (V_pf_status.debug >= PF_DEBUG_MISC) {
1693 			DPFPRINTF(("dst idled out of PAWS\n"));
1694 			pf_print_state(state);
1695 			printf("\n");
1696 		}
1697 		dst->scrub->pfss_flags = (dst->scrub->pfss_flags & ~PFSS_PAWS)
1698 		    | PFSS_PAWS_IDLED;
1699 	}
1700 
1701 	if (got_ts && src->scrub && dst->scrub &&
1702 	    (src->scrub->pfss_flags & PFSS_PAWS) &&
1703 	    (dst->scrub->pfss_flags & PFSS_PAWS)) {
1704 		/* Validate that the timestamps are "in-window".
1705 		 * RFC1323 describes TCP Timestamp options that allow
1706 		 * measurement of RTT (round trip time) and PAWS
1707 		 * (protection against wrapped sequence numbers).  PAWS
1708 		 * gives us a set of rules for rejecting packets on
1709 		 * long fat pipes (packets that were somehow delayed
1710 		 * in transit longer than the time it took to send the
1711 		 * full TCP sequence space of 4Gb).  We can use these
1712 		 * rules and infer a few others that will let us treat
1713 		 * the 32bit timestamp and the 32bit echoed timestamp
1714 		 * as sequence numbers to prevent a blind attacker from
1715 		 * inserting packets into a connection.
1716 		 *
1717 		 * RFC1323 tells us:
1718 		 *  - The timestamp on this packet must be greater than
1719 		 *    or equal to the last value echoed by the other
1720 		 *    endpoint.  The RFC says those will be discarded
1721 		 *    since it is a dup that has already been acked.
1722 		 *    This gives us a lowerbound on the timestamp.
1723 		 *        timestamp >= other last echoed timestamp
1724 		 *  - The timestamp will be less than or equal to
1725 		 *    the last timestamp plus the time between the
1726 		 *    last packet and now.  The RFC defines the max
1727 		 *    clock rate as 1ms.  We will allow clocks to be
1728 		 *    up to 10% fast and will allow a total difference
1729 		 *    or 30 seconds due to a route change.  And this
1730 		 *    gives us an upperbound on the timestamp.
1731 		 *        timestamp <= last timestamp + max ticks
1732 		 *    We have to be careful here.  Windows will send an
1733 		 *    initial timestamp of zero and then initialize it
1734 		 *    to a random value after the 3whs; presumably to
1735 		 *    avoid a DoS by having to call an expensive RNG
1736 		 *    during a SYN flood.  Proof MS has at least one
1737 		 *    good security geek.
1738 		 *
1739 		 *  - The TCP timestamp option must also echo the other
1740 		 *    endpoints timestamp.  The timestamp echoed is the
1741 		 *    one carried on the earliest unacknowledged segment
1742 		 *    on the left edge of the sequence window.  The RFC
1743 		 *    states that the host will reject any echoed
1744 		 *    timestamps that were larger than any ever sent.
1745 		 *    This gives us an upperbound on the TS echo.
1746 		 *        tescr <= largest_tsval
1747 		 *  - The lowerbound on the TS echo is a little more
1748 		 *    tricky to determine.  The other endpoint's echoed
1749 		 *    values will not decrease.  But there may be
1750 		 *    network conditions that re-order packets and
1751 		 *    cause our view of them to decrease.  For now the
1752 		 *    only lowerbound we can safely determine is that
1753 		 *    the TS echo will never be less than the original
1754 		 *    TS.  XXX There is probably a better lowerbound.
1755 		 *    Remove TS_MAX_CONN with better lowerbound check.
1756 		 *        tescr >= other original TS
1757 		 *
1758 		 * It is also important to note that the fastest
1759 		 * timestamp clock of 1ms will wrap its 32bit space in
1760 		 * 24 days.  So we just disable TS checking after 24
1761 		 * days of idle time.  We actually must use a 12d
1762 		 * connection limit until we can come up with a better
1763 		 * lowerbound to the TS echo check.
1764 		 */
1765 		struct timeval delta_ts;
1766 		int ts_fudge;
1767 
1768 		/*
1769 		 * PFTM_TS_DIFF is how many seconds of leeway to allow
1770 		 * a host's timestamp.  This can happen if the previous
1771 		 * packet got delayed in transit for much longer than
1772 		 * this packet.
1773 		 */
1774 		if ((ts_fudge = state->rule.ptr->timeout[PFTM_TS_DIFF]) == 0)
1775 			ts_fudge = V_pf_default_rule.timeout[PFTM_TS_DIFF];
1776 
1777 		/* Calculate max ticks since the last timestamp */
1778 #define TS_MAXFREQ	1100		/* RFC max TS freq of 1Khz + 10% skew */
1779 #define TS_MICROSECS	1000000		/* microseconds per second */
1780 		delta_ts = uptime;
1781 		timevalsub(&delta_ts, &src->scrub->pfss_last);
1782 		tsval_from_last = (delta_ts.tv_sec + ts_fudge) * TS_MAXFREQ;
1783 		tsval_from_last += delta_ts.tv_usec / (TS_MICROSECS/TS_MAXFREQ);
1784 
1785 		if ((src->state >= TCPS_ESTABLISHED &&
1786 		    dst->state >= TCPS_ESTABLISHED) &&
1787 		    (SEQ_LT(tsval, dst->scrub->pfss_tsecr) ||
1788 		    SEQ_GT(tsval, src->scrub->pfss_tsval + tsval_from_last) ||
1789 		    (tsecr && (SEQ_GT(tsecr, dst->scrub->pfss_tsval) ||
1790 		    SEQ_LT(tsecr, dst->scrub->pfss_tsval0))))) {
1791 			/* Bad RFC1323 implementation or an insertion attack.
1792 			 *
1793 			 * - Solaris 2.6 and 2.7 are known to send another ACK
1794 			 *   after the FIN,FIN|ACK,ACK closing that carries
1795 			 *   an old timestamp.
1796 			 */
1797 
1798 			DPFPRINTF(("Timestamp failed %c%c%c%c\n",
1799 			    SEQ_LT(tsval, dst->scrub->pfss_tsecr) ? '0' : ' ',
1800 			    SEQ_GT(tsval, src->scrub->pfss_tsval +
1801 			    tsval_from_last) ? '1' : ' ',
1802 			    SEQ_GT(tsecr, dst->scrub->pfss_tsval) ? '2' : ' ',
1803 			    SEQ_LT(tsecr, dst->scrub->pfss_tsval0)? '3' : ' '));
1804 			DPFPRINTF((" tsval: %u  tsecr: %u  +ticks: %u  "
1805 			    "idle: %jus %lums\n",
1806 			    tsval, tsecr, tsval_from_last,
1807 			    (uintmax_t)delta_ts.tv_sec,
1808 			    delta_ts.tv_usec / 1000));
1809 			DPFPRINTF((" src->tsval: %u  tsecr: %u\n",
1810 			    src->scrub->pfss_tsval, src->scrub->pfss_tsecr));
1811 			DPFPRINTF((" dst->tsval: %u  tsecr: %u  tsval0: %u"
1812 			    "\n", dst->scrub->pfss_tsval,
1813 			    dst->scrub->pfss_tsecr, dst->scrub->pfss_tsval0));
1814 			if (V_pf_status.debug >= PF_DEBUG_MISC) {
1815 				pf_print_state(state);
1816 				pf_print_flags(th->th_flags);
1817 				printf("\n");
1818 			}
1819 			REASON_SET(reason, PFRES_TS);
1820 			return (PF_DROP);
1821 		}
1822 
1823 		/* XXX I'd really like to require tsecr but it's optional */
1824 
1825 	} else if (!got_ts && (th->th_flags & TH_RST) == 0 &&
1826 	    ((src->state == TCPS_ESTABLISHED && dst->state == TCPS_ESTABLISHED)
1827 	    || pd->p_len > 0 || (th->th_flags & TH_SYN)) &&
1828 	    src->scrub && dst->scrub &&
1829 	    (src->scrub->pfss_flags & PFSS_PAWS) &&
1830 	    (dst->scrub->pfss_flags & PFSS_PAWS)) {
1831 		/* Didn't send a timestamp.  Timestamps aren't really useful
1832 		 * when:
1833 		 *  - connection opening or closing (often not even sent).
1834 		 *    but we must not let an attacker to put a FIN on a
1835 		 *    data packet to sneak it through our ESTABLISHED check.
1836 		 *  - on a TCP reset.  RFC suggests not even looking at TS.
1837 		 *  - on an empty ACK.  The TS will not be echoed so it will
1838 		 *    probably not help keep the RTT calculation in sync and
1839 		 *    there isn't as much danger when the sequence numbers
1840 		 *    got wrapped.  So some stacks don't include TS on empty
1841 		 *    ACKs :-(
1842 		 *
1843 		 * To minimize the disruption to mostly RFC1323 conformant
1844 		 * stacks, we will only require timestamps on data packets.
1845 		 *
1846 		 * And what do ya know, we cannot require timestamps on data
1847 		 * packets.  There appear to be devices that do legitimate
1848 		 * TCP connection hijacking.  There are HTTP devices that allow
1849 		 * a 3whs (with timestamps) and then buffer the HTTP request.
1850 		 * If the intermediate device has the HTTP response cache, it
1851 		 * will spoof the response but not bother timestamping its
1852 		 * packets.  So we can look for the presence of a timestamp in
1853 		 * the first data packet and if there, require it in all future
1854 		 * packets.
1855 		 */
1856 
1857 		if (pd->p_len > 0 && (src->scrub->pfss_flags & PFSS_DATA_TS)) {
1858 			/*
1859 			 * Hey!  Someone tried to sneak a packet in.  Or the
1860 			 * stack changed its RFC1323 behavior?!?!
1861 			 */
1862 			if (V_pf_status.debug >= PF_DEBUG_MISC) {
1863 				DPFPRINTF(("Did not receive expected RFC1323 "
1864 				    "timestamp\n"));
1865 				pf_print_state(state);
1866 				pf_print_flags(th->th_flags);
1867 				printf("\n");
1868 			}
1869 			REASON_SET(reason, PFRES_TS);
1870 			return (PF_DROP);
1871 		}
1872 	}
1873 
1874 	/*
1875 	 * We will note if a host sends his data packets with or without
1876 	 * timestamps.  And require all data packets to contain a timestamp
1877 	 * if the first does.  PAWS implicitly requires that all data packets be
1878 	 * timestamped.  But I think there are middle-man devices that hijack
1879 	 * TCP streams immediately after the 3whs and don't timestamp their
1880 	 * packets (seen in a WWW accelerator or cache).
1881 	 */
1882 	if (pd->p_len > 0 && src->scrub && (src->scrub->pfss_flags &
1883 	    (PFSS_TIMESTAMP|PFSS_DATA_TS|PFSS_DATA_NOTS)) == PFSS_TIMESTAMP) {
1884 		if (got_ts)
1885 			src->scrub->pfss_flags |= PFSS_DATA_TS;
1886 		else {
1887 			src->scrub->pfss_flags |= PFSS_DATA_NOTS;
1888 			if (V_pf_status.debug >= PF_DEBUG_MISC && dst->scrub &&
1889 			    (dst->scrub->pfss_flags & PFSS_TIMESTAMP)) {
1890 				/* Don't warn if other host rejected RFC1323 */
1891 				DPFPRINTF(("Broken RFC1323 stack did not "
1892 				    "timestamp data packet. Disabled PAWS "
1893 				    "security.\n"));
1894 				pf_print_state(state);
1895 				pf_print_flags(th->th_flags);
1896 				printf("\n");
1897 			}
1898 		}
1899 	}
1900 
1901 	/*
1902 	 * Update PAWS values
1903 	 */
1904 	if (got_ts && src->scrub && PFSS_TIMESTAMP == (src->scrub->pfss_flags &
1905 	    (PFSS_PAWS_IDLED|PFSS_TIMESTAMP))) {
1906 		getmicrouptime(&src->scrub->pfss_last);
1907 		if (SEQ_GEQ(tsval, src->scrub->pfss_tsval) ||
1908 		    (src->scrub->pfss_flags & PFSS_PAWS) == 0)
1909 			src->scrub->pfss_tsval = tsval;
1910 
1911 		if (tsecr) {
1912 			if (SEQ_GEQ(tsecr, src->scrub->pfss_tsecr) ||
1913 			    (src->scrub->pfss_flags & PFSS_PAWS) == 0)
1914 				src->scrub->pfss_tsecr = tsecr;
1915 
1916 			if ((src->scrub->pfss_flags & PFSS_PAWS) == 0 &&
1917 			    (SEQ_LT(tsval, src->scrub->pfss_tsval0) ||
1918 			    src->scrub->pfss_tsval0 == 0)) {
1919 				/* tsval0 MUST be the lowest timestamp */
1920 				src->scrub->pfss_tsval0 = tsval;
1921 			}
1922 
1923 			/* Only fully initialized after a TS gets echoed */
1924 			if ((src->scrub->pfss_flags & PFSS_PAWS) == 0)
1925 				src->scrub->pfss_flags |= PFSS_PAWS;
1926 		}
1927 	}
1928 
1929 	/* I have a dream....  TCP segment reassembly.... */
1930 	return (0);
1931 }
1932 
1933 static int
1934 pf_normalize_tcpopt(struct pf_krule *r, struct mbuf *m, struct tcphdr *th,
1935     int off, sa_family_t af)
1936 {
1937 	u_int16_t	*mss;
1938 	int		 thoff;
1939 	int		 opt, cnt, optlen = 0;
1940 	int		 rewrite = 0;
1941 	u_char		 opts[TCP_MAXOLEN];
1942 	u_char		*optp = opts;
1943 	size_t		 startoff;
1944 
1945 	thoff = th->th_off << 2;
1946 	cnt = thoff - sizeof(struct tcphdr);
1947 
1948 	if (cnt > 0 && !pf_pull_hdr(m, off + sizeof(*th), opts, cnt,
1949 	    NULL, NULL, af))
1950 		return (rewrite);
1951 
1952 	for (; cnt > 0; cnt -= optlen, optp += optlen) {
1953 		startoff = optp - opts;
1954 		opt = optp[0];
1955 		if (opt == TCPOPT_EOL)
1956 			break;
1957 		if (opt == TCPOPT_NOP)
1958 			optlen = 1;
1959 		else {
1960 			if (cnt < 2)
1961 				break;
1962 			optlen = optp[1];
1963 			if (optlen < 2 || optlen > cnt)
1964 				break;
1965 		}
1966 		switch (opt) {
1967 		case TCPOPT_MAXSEG:
1968 			mss = (u_int16_t *)(optp + 2);
1969 			if ((ntohs(*mss)) > r->max_mss) {
1970 				pf_patch_16_unaligned(m,
1971 				    &th->th_sum,
1972 				    mss, htons(r->max_mss),
1973 				    PF_ALGNMNT(startoff),
1974 				    0);
1975 				rewrite = 1;
1976 			}
1977 			break;
1978 		default:
1979 			break;
1980 		}
1981 	}
1982 
1983 	if (rewrite)
1984 		m_copyback(m, off + sizeof(*th), thoff - sizeof(*th), opts);
1985 
1986 	return (rewrite);
1987 }
1988 
1989 #ifdef INET
1990 static void
1991 pf_scrub_ip(struct mbuf **m0, u_int32_t flags, u_int8_t min_ttl, u_int8_t tos)
1992 {
1993 	struct mbuf		*m = *m0;
1994 	struct ip		*h = mtod(m, struct ip *);
1995 
1996 	/* Clear IP_DF if no-df was requested */
1997 	if (flags & PFRULE_NODF && h->ip_off & htons(IP_DF)) {
1998 		u_int16_t ip_off = h->ip_off;
1999 
2000 		h->ip_off &= htons(~IP_DF);
2001 		h->ip_sum = pf_cksum_fixup(h->ip_sum, ip_off, h->ip_off, 0);
2002 	}
2003 
2004 	/* Enforce a minimum ttl, may cause endless packet loops */
2005 	if (min_ttl && h->ip_ttl < min_ttl) {
2006 		u_int16_t ip_ttl = h->ip_ttl;
2007 
2008 		h->ip_ttl = min_ttl;
2009 		h->ip_sum = pf_cksum_fixup(h->ip_sum, ip_ttl, h->ip_ttl, 0);
2010 	}
2011 
2012 	/* Enforce tos */
2013 	if (flags & PFRULE_SET_TOS) {
2014 		u_int16_t	ov, nv;
2015 
2016 		ov = *(u_int16_t *)h;
2017 		h->ip_tos = tos | (h->ip_tos & IPTOS_ECN_MASK);
2018 		nv = *(u_int16_t *)h;
2019 
2020 		h->ip_sum = pf_cksum_fixup(h->ip_sum, ov, nv, 0);
2021 	}
2022 
2023 	/* random-id, but not for fragments */
2024 	if (flags & PFRULE_RANDOMID && !(h->ip_off & ~htons(IP_DF))) {
2025 		uint16_t ip_id = h->ip_id;
2026 
2027 		ip_fillid(h);
2028 		h->ip_sum = pf_cksum_fixup(h->ip_sum, ip_id, h->ip_id, 0);
2029 	}
2030 }
2031 #endif /* INET */
2032 
2033 #ifdef INET6
2034 static void
2035 pf_scrub_ip6(struct mbuf **m0, u_int8_t min_ttl)
2036 {
2037 	struct mbuf		*m = *m0;
2038 	struct ip6_hdr		*h = mtod(m, struct ip6_hdr *);
2039 
2040 	/* Enforce a minimum ttl, may cause endless packet loops */
2041 	if (min_ttl && h->ip6_hlim < min_ttl)
2042 		h->ip6_hlim = min_ttl;
2043 }
2044 #endif
2045