xref: /netbsd-src/sys/netinet/ip_reass.c (revision b7b7574d3bf8eeb51a1fa3977b59142ec6434a55) 
1 /*	$NetBSD: ip_reass.c,v 1.9 2014/02/25 18:30:12 pooka Exp $	*/
2 
3 /*
4  * Copyright (c) 1982, 1986, 1988, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	@(#)ip_input.c	8.2 (Berkeley) 1/4/94
32  */
33 
34 /*
35  * IP reassembly.
36  *
37  * Additive-Increase/Multiplicative-Decrease (AIMD) strategy for IP
38  * reassembly queue buffer managment.
39  *
40  * We keep a count of total IP fragments (NB: not fragmented packets),
41  * awaiting reassembly (ip_nfrags) and a limit (ip_maxfrags) on fragments.
42  * If ip_nfrags exceeds ip_maxfrags the limit, we drop half the total
43  * fragments in reassembly queues.  This AIMD policy avoids repeatedly
44  * deleting single packets under heavy fragmentation load (e.g., from lossy
45  * NFS peers).
46  */
47 
48 #include <sys/cdefs.h>
49 __KERNEL_RCSID(0, "$NetBSD: ip_reass.c,v 1.9 2014/02/25 18:30:12 pooka Exp $");
50 
51 #include <sys/param.h>
52 #include <sys/types.h>
53 
54 #include <sys/malloc.h>
55 #include <sys/mbuf.h>
56 #include <sys/mutex.h>
57 #include <sys/domain.h>
58 #include <sys/protosw.h>
59 #include <sys/pool.h>
60 #include <sys/queue.h>
61 #include <sys/sysctl.h>
62 #include <sys/systm.h>
63 
64 #include <net/if.h>
65 #include <net/route.h>
66 
67 #include <netinet/in.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/ip.h>
70 #include <netinet/in_pcb.h>
71 #include <netinet/ip_var.h>
72 #include <netinet/in_proto.h>
73 #include <netinet/ip_private.h>
74 #include <netinet/in_var.h>
75 
76 /*
77  * IP reassembly queue structures.  Each fragment being reassembled is
78  * attached to one of these structures.  They are timed out after TTL
79  * drops to 0, and may also be reclaimed if memory becomes tight.
80  */
81 
82 typedef struct ipfr_qent {
83 	TAILQ_ENTRY(ipfr_qent)	ipqe_q;
84 	struct ip *		ipqe_ip;
85 	struct mbuf *		ipqe_m;
86 	bool			ipqe_mff;
87 } ipfr_qent_t;
88 
89 TAILQ_HEAD(ipfr_qent_head, ipfr_qent);
90 
91 typedef struct ipfr_queue {
92 	LIST_ENTRY(ipfr_queue)	ipq_q;		/* to other reass headers */
93 	struct ipfr_qent_head	ipq_fragq;	/* queue of fragment entries */
94 	uint8_t			ipq_ttl;	/* time for reass q to live */
95 	uint8_t			ipq_p;		/* protocol of this fragment */
96 	uint16_t		ipq_id;		/* sequence id for reassembly */
97 	struct in_addr		ipq_src;
98 	struct in_addr		ipq_dst;
99 	uint16_t		ipq_nfrags;	/* frags in this queue entry */
100 	uint8_t 		ipq_tos;	/* TOS of this fragment */
101 } ipfr_queue_t;
102 
103 /*
104  * Hash table of IP reassembly queues.
105  */
106 #define	IPREASS_HASH_SHIFT	6
107 #define	IPREASS_HASH_SIZE	(1 << IPREASS_HASH_SHIFT)
108 #define	IPREASS_HASH_MASK	(IPREASS_HASH_SIZE - 1)
109 #define	IPREASS_HASH(x, y) \
110 	(((((x) & 0xf) | ((((x) >> 8) & 0xf) << 4)) ^ (y)) & IPREASS_HASH_MASK)
111 
112 static LIST_HEAD(, ipfr_queue)	ip_frags[IPREASS_HASH_SIZE];
113 static pool_cache_t	ipfren_cache;
114 static kmutex_t		ipfr_lock;
115 
116 /* Number of packets in reassembly queue and total number of fragments. */
117 static int		ip_nfragpackets;
118 static int		ip_nfrags;
119 
120 /* Limits on packet and fragments. */
121 static int		ip_maxfragpackets;
122 static int		ip_maxfrags;
123 
124 /*
125  * Cached copy of nmbclusters.  If nbclusters is different, recalculate
126  * IP parameters derived from nmbclusters.
127  */
128 static int		ip_nmbclusters;
129 
130 /*
131  * IP reassembly TTL machinery for multiplicative drop.
132  */
133 static u_int		fragttl_histo[IPFRAGTTL + 1];
134 
135 static struct sysctllog *ip_reass_sysctllog;
136 
137 void			sysctl_ip_reass_setup(void);
138 static void		ip_nmbclusters_changed(void);
139 
140 static struct mbuf *	ip_reass(ipfr_qent_t *, ipfr_queue_t *, u_int);
141 static u_int		ip_reass_ttl_decr(u_int ticks);
142 static void		ip_reass_drophalf(void);
143 static void		ip_freef(ipfr_queue_t *);
144 
145 /*
146  * ip_reass_init:
147  *
148  *	Initialization of IP reassembly mechanism.
149  */
150 void
151 ip_reass_init(void)
152 {
153 	int i;
154 
155 	ipfren_cache = pool_cache_init(sizeof(ipfr_qent_t), coherency_unit,
156 	    0, 0, "ipfrenpl", NULL, IPL_NET, NULL, NULL, NULL);
157 	mutex_init(&ipfr_lock, MUTEX_DEFAULT, IPL_VM);
158 
159 	for (i = 0; i < IPREASS_HASH_SIZE; i++) {
160 		LIST_INIT(&ip_frags[i]);
161 	}
162 	ip_maxfragpackets = 200;
163 	ip_maxfrags = 0;
164 	ip_nmbclusters_changed();
165 
166 	sysctl_ip_reass_setup();
167 }
168 
169 void
170 sysctl_ip_reass_setup(void)
171 {
172 
173 	sysctl_createv(&ip_reass_sysctllog, 0, NULL, NULL,
174 		CTLFLAG_PERMANENT,
175 		CTLTYPE_NODE, "inet",
176 		SYSCTL_DESCR("PF_INET related settings"),
177 		NULL, 0, NULL, 0,
178 		CTL_NET, PF_INET, CTL_EOL);
179 	sysctl_createv(&ip_reass_sysctllog, 0, NULL, NULL,
180 		CTLFLAG_PERMANENT,
181 		CTLTYPE_NODE, "ip",
182 		SYSCTL_DESCR("IPv4 related settings"),
183 		NULL, 0, NULL, 0,
184 		CTL_NET, PF_INET, IPPROTO_IP, CTL_EOL);
185 
186 	sysctl_createv(&ip_reass_sysctllog, 0, NULL, NULL,
187 		CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
188 		CTLTYPE_INT, "maxfragpackets",
189 		SYSCTL_DESCR("Maximum number of fragments to retain for "
190 			     "possible reassembly"),
191 		NULL, 0, &ip_maxfragpackets, 0,
192 		CTL_NET, PF_INET, IPPROTO_IP, IPCTL_MAXFRAGPACKETS, CTL_EOL);
193 }
194 
195 #define CHECK_NMBCLUSTER_PARAMS()				\
196 do {								\
197 	if (__predict_false(ip_nmbclusters != nmbclusters))	\
198 		ip_nmbclusters_changed();			\
199 } while (/*CONSTCOND*/0)
200 
201 /*
202  * Compute IP limits derived from the value of nmbclusters.
203  */
204 static void
205 ip_nmbclusters_changed(void)
206 {
207 	ip_maxfrags = nmbclusters / 4;
208 	ip_nmbclusters = nmbclusters;
209 }
210 
211 /*
212  * ip_reass:
213  *
214  *	Take incoming datagram fragment and try to reassemble it into whole
215  *	datagram.  If a chain for reassembly of this datagram already exists,
216  *	then it is given as 'fp'; otherwise have to make a chain.
217  */
218 struct mbuf *
219 ip_reass(ipfr_qent_t *ipqe, ipfr_queue_t *fp, const u_int hash)
220 {
221 	struct ip *ip = ipqe->ipqe_ip, *qip;
222 	const int hlen = ip->ip_hl << 2;
223 	struct mbuf *m = ipqe->ipqe_m, *t;
224 	ipfr_qent_t *nq, *p, *q;
225 	int i, next;
226 
227 	KASSERT(mutex_owned(&ipfr_lock));
228 
229 	/*
230 	 * Presence of header sizes in mbufs would confuse code below.
231 	 */
232 	m->m_data += hlen;
233 	m->m_len -= hlen;
234 
235 #ifdef	notyet
236 	/* Make sure fragment limit is up-to-date. */
237 	CHECK_NMBCLUSTER_PARAMS();
238 
239 	/* If we have too many fragments, drop the older half. */
240 	if (ip_nfrags >= ip_maxfrags) {
241 		ip_reass_drophalf(void);
242 	}
243 #endif
244 
245 	/*
246 	 * We are about to add a fragment; increment frag count.
247 	 */
248 	ip_nfrags++;
249 
250 	/*
251 	 * If first fragment to arrive, create a reassembly queue.
252 	 */
253 	if (fp == NULL) {
254 		/*
255 		 * Enforce upper bound on number of fragmented packets
256 		 * for which we attempt reassembly:  a) if maxfrag is 0,
257 		 * never accept fragments  b) if maxfrag is -1, accept
258 		 * all fragments without limitation.
259 		 */
260 		if (ip_maxfragpackets < 0)
261 			;
262 		else if (ip_nfragpackets >= ip_maxfragpackets) {
263 			goto dropfrag;
264 		}
265 		fp = malloc(sizeof(ipfr_queue_t), M_FTABLE, M_NOWAIT);
266 		if (fp == NULL) {
267 			goto dropfrag;
268 		}
269 		ip_nfragpackets++;
270 		TAILQ_INIT(&fp->ipq_fragq);
271 		fp->ipq_nfrags = 1;
272 		fp->ipq_ttl = IPFRAGTTL;
273 		fp->ipq_p = ip->ip_p;
274 		fp->ipq_id = ip->ip_id;
275 		fp->ipq_tos = ip->ip_tos;
276 		fp->ipq_src = ip->ip_src;
277 		fp->ipq_dst = ip->ip_dst;
278 		LIST_INSERT_HEAD(&ip_frags[hash], fp, ipq_q);
279 		p = NULL;
280 		goto insert;
281 	} else {
282 		fp->ipq_nfrags++;
283 	}
284 
285 	/*
286 	 * Find a segment which begins after this one does.
287 	 */
288 	TAILQ_FOREACH(q, &fp->ipq_fragq, ipqe_q) {
289 		if (ntohs(q->ipqe_ip->ip_off) > ntohs(ip->ip_off))
290 			break;
291 	}
292 	if (q != NULL) {
293 		p = TAILQ_PREV(q, ipfr_qent_head, ipqe_q);
294 	} else {
295 		p = TAILQ_LAST(&fp->ipq_fragq, ipfr_qent_head);
296 	}
297 
298 	/*
299 	 * If there is a preceding segment, it may provide some of our
300 	 * data already.  If so, drop the data from the incoming segment.
301 	 * If it provides all of our data, drop us.
302 	 */
303 	if (p != NULL) {
304 		i = ntohs(p->ipqe_ip->ip_off) + ntohs(p->ipqe_ip->ip_len) -
305 		    ntohs(ip->ip_off);
306 		if (i > 0) {
307 			if (i >= ntohs(ip->ip_len)) {
308 				goto dropfrag;
309 			}
310 			m_adj(ipqe->ipqe_m, i);
311 			ip->ip_off = htons(ntohs(ip->ip_off) + i);
312 			ip->ip_len = htons(ntohs(ip->ip_len) - i);
313 		}
314 	}
315 
316 	/*
317 	 * While we overlap succeeding segments trim them or, if they are
318 	 * completely covered, dequeue them.
319 	 */
320 	while (q != NULL) {
321 		size_t end;
322 
323 		qip = q->ipqe_ip;
324 		end = ntohs(ip->ip_off) + ntohs(ip->ip_len);
325 		if (end <= ntohs(qip->ip_off)) {
326 			break;
327 		}
328 		i = end - ntohs(qip->ip_off);
329 		if (i < ntohs(qip->ip_len)) {
330 			qip->ip_len = htons(ntohs(qip->ip_len) - i);
331 			qip->ip_off = htons(ntohs(qip->ip_off) + i);
332 			m_adj(q->ipqe_m, i);
333 			break;
334 		}
335 		nq = TAILQ_NEXT(q, ipqe_q);
336 		m_freem(q->ipqe_m);
337 		TAILQ_REMOVE(&fp->ipq_fragq, q, ipqe_q);
338 		pool_cache_put(ipfren_cache, q);
339 		fp->ipq_nfrags--;
340 		ip_nfrags--;
341 		q = nq;
342 	}
343 
344 insert:
345 	/*
346 	 * Stick new segment in its place; check for complete reassembly.
347 	 */
348 	if (p == NULL) {
349 		TAILQ_INSERT_HEAD(&fp->ipq_fragq, ipqe, ipqe_q);
350 	} else {
351 		TAILQ_INSERT_AFTER(&fp->ipq_fragq, p, ipqe, ipqe_q);
352 	}
353 	next = 0;
354 	TAILQ_FOREACH(q, &fp->ipq_fragq, ipqe_q) {
355 		qip = q->ipqe_ip;
356 		if (ntohs(qip->ip_off) != next) {
357 			mutex_exit(&ipfr_lock);
358 			return NULL;
359 		}
360 		next += ntohs(qip->ip_len);
361 	}
362 	p = TAILQ_LAST(&fp->ipq_fragq, ipfr_qent_head);
363 	if (p->ipqe_mff) {
364 		mutex_exit(&ipfr_lock);
365 		return NULL;
366 	}
367 
368 	/*
369 	 * Reassembly is complete.  Check for a bogus message size.
370 	 */
371 	q = TAILQ_FIRST(&fp->ipq_fragq);
372 	ip = q->ipqe_ip;
373 	if ((next + (ip->ip_hl << 2)) > IP_MAXPACKET) {
374 		IP_STATINC(IP_STAT_TOOLONG);
375 		ip_freef(fp);
376 		mutex_exit(&ipfr_lock);
377 		return NULL;
378 	}
379 	LIST_REMOVE(fp, ipq_q);
380 	ip_nfrags -= fp->ipq_nfrags;
381 	ip_nfragpackets--;
382 	mutex_exit(&ipfr_lock);
383 
384 	/* Concatenate all fragments. */
385 	m = q->ipqe_m;
386 	t = m->m_next;
387 	m->m_next = NULL;
388 	m_cat(m, t);
389 	nq = TAILQ_NEXT(q, ipqe_q);
390 	pool_cache_put(ipfren_cache, q);
391 
392 	for (q = nq; q != NULL; q = nq) {
393 		t = q->ipqe_m;
394 		nq = TAILQ_NEXT(q, ipqe_q);
395 		pool_cache_put(ipfren_cache, q);
396 		m_cat(m, t);
397 	}
398 
399 	/*
400 	 * Create header for new packet by modifying header of first
401 	 * packet.  Dequeue and discard fragment reassembly header.  Make
402 	 * header visible.
403 	 */
404 	ip->ip_len = htons((ip->ip_hl << 2) + next);
405 	ip->ip_src = fp->ipq_src;
406 	ip->ip_dst = fp->ipq_dst;
407 	free(fp, M_FTABLE);
408 
409 	m->m_len += (ip->ip_hl << 2);
410 	m->m_data -= (ip->ip_hl << 2);
411 
412 	/* Fix up mbuf.  XXX This should be done elsewhere. */
413 	if (m->m_flags & M_PKTHDR) {
414 		int plen = 0;
415 		for (t = m; t; t = t->m_next) {
416 			plen += t->m_len;
417 		}
418 		m->m_pkthdr.len = plen;
419 		m->m_pkthdr.csum_flags = 0;
420 	}
421 	return m;
422 
423 dropfrag:
424 	if (fp != NULL) {
425 		fp->ipq_nfrags--;
426 	}
427 	ip_nfrags--;
428 	IP_STATINC(IP_STAT_FRAGDROPPED);
429 	mutex_exit(&ipfr_lock);
430 
431 	pool_cache_put(ipfren_cache, ipqe);
432 	m_freem(m);
433 	return NULL;
434 }
435 
436 /*
437  * ip_freef:
438  *
439  *	Free a fragment reassembly header and all associated datagrams.
440  */
441 static void
442 ip_freef(ipfr_queue_t *fp)
443 {
444 	ipfr_qent_t *q;
445 
446 	KASSERT(mutex_owned(&ipfr_lock));
447 
448 	LIST_REMOVE(fp, ipq_q);
449 	ip_nfrags -= fp->ipq_nfrags;
450 	ip_nfragpackets--;
451 
452 	while ((q = TAILQ_FIRST(&fp->ipq_fragq)) != NULL) {
453 		TAILQ_REMOVE(&fp->ipq_fragq, q, ipqe_q);
454 		m_freem(q->ipqe_m);
455 		pool_cache_put(ipfren_cache, q);
456 	}
457 	free(fp, M_FTABLE);
458 }
459 
460 /*
461  * ip_reass_ttl_decr:
462  *
463  *	Decrement TTL of all reasembly queue entries by `ticks'.  Count
464  *	number of distinct fragments (as opposed to partial, fragmented
465  *	datagrams) inthe reassembly queue.  While we  traverse the entire
466  *	reassembly queue, compute and return the median TTL over all
467  *	fragments.
468  */
469 static u_int
470 ip_reass_ttl_decr(u_int ticks)
471 {
472 	u_int nfrags, median, dropfraction, keepfraction;
473 	ipfr_queue_t *fp, *nfp;
474 	int i;
475 
476 	nfrags = 0;
477 	memset(fragttl_histo, 0, sizeof(fragttl_histo));
478 
479 	for (i = 0; i < IPREASS_HASH_SIZE; i++) {
480 		for (fp = LIST_FIRST(&ip_frags[i]); fp != NULL; fp = nfp) {
481 			fp->ipq_ttl = ((fp->ipq_ttl <= ticks) ?
482 			    0 : fp->ipq_ttl - ticks);
483 			nfp = LIST_NEXT(fp, ipq_q);
484 			if (fp->ipq_ttl == 0) {
485 				IP_STATINC(IP_STAT_FRAGTIMEOUT);
486 				ip_freef(fp);
487 			} else {
488 				nfrags += fp->ipq_nfrags;
489 				fragttl_histo[fp->ipq_ttl] += fp->ipq_nfrags;
490 			}
491 		}
492 	}
493 
494 	KASSERT(ip_nfrags == nfrags);
495 
496 	/* Find median (or other drop fraction) in histogram. */
497 	dropfraction = (ip_nfrags / 2);
498 	keepfraction = ip_nfrags - dropfraction;
499 	for (i = IPFRAGTTL, median = 0; i >= 0; i--) {
500 		median += fragttl_histo[i];
501 		if (median >= keepfraction)
502 			break;
503 	}
504 
505 	/* Return TTL of median (or other fraction). */
506 	return (u_int)i;
507 }
508 
509 static void
510 ip_reass_drophalf(void)
511 {
512 	u_int median_ticks;
513 
514 	KASSERT(mutex_owned(&ipfr_lock));
515 
516 	/*
517 	 * Compute median TTL of all fragments, and count frags
518 	 * with that TTL or lower (roughly half of all fragments).
519 	 */
520 	median_ticks = ip_reass_ttl_decr(0);
521 
522 	/* Drop half. */
523 	median_ticks = ip_reass_ttl_decr(median_ticks);
524 }
525 
526 /*
527  * ip_reass_drain: drain off all datagram fragments.  Do not acquire
528  * softnet_lock as can be called from hardware interrupt context.
529  */
530 void
531 ip_reass_drain(void)
532 {
533 
534 	/*
535 	 * We may be called from a device's interrupt context.  If
536 	 * the ipq is already busy, just bail out now.
537 	 */
538 	if (mutex_tryenter(&ipfr_lock)) {
539 		/*
540 		 * Drop half the total fragments now. If more mbufs are
541 		 * needed, we will be called again soon.
542 		 */
543 		ip_reass_drophalf();
544 		mutex_exit(&ipfr_lock);
545 	}
546 }
547 
548 /*
549  * ip_reass_slowtimo:
550  *
551  *	If a timer expires on a reassembly queue, discard it.
552  */
553 void
554 ip_reass_slowtimo(void)
555 {
556 	static u_int dropscanidx = 0;
557 	u_int i, median_ttl;
558 
559 	mutex_enter(&ipfr_lock);
560 
561 	/* Age TTL of all fragments by 1 tick .*/
562 	median_ttl = ip_reass_ttl_decr(1);
563 
564 	/* Make sure fragment limit is up-to-date. */
565 	CHECK_NMBCLUSTER_PARAMS();
566 
567 	/* If we have too many fragments, drop the older half. */
568 	if (ip_nfrags > ip_maxfrags) {
569 		ip_reass_ttl_decr(median_ttl);
570 	}
571 
572 	/*
573 	 * If we are over the maximum number of fragmented packets (due to
574 	 * the limit being lowered), drain off enough to get down to the
575 	 * new limit.  Start draining from the reassembly hashqueue most
576 	 * recently drained.
577 	 */
578 	if (ip_maxfragpackets < 0)
579 		;
580 	else {
581 		int wrapped = 0;
582 
583 		i = dropscanidx;
584 		while (ip_nfragpackets > ip_maxfragpackets && wrapped == 0) {
585 			while (LIST_FIRST(&ip_frags[i]) != NULL) {
586 				ip_freef(LIST_FIRST(&ip_frags[i]));
587 			}
588 			if (++i >= IPREASS_HASH_SIZE) {
589 				i = 0;
590 			}
591 			/*
592 			 * Do not scan forever even if fragment counters are
593 			 * wrong: stop after scanning entire reassembly queue.
594 			 */
595 			if (i == dropscanidx) {
596 				wrapped = 1;
597 			}
598 		}
599 		dropscanidx = i;
600 	}
601 	mutex_exit(&ipfr_lock);
602 }
603 
604 /*
605  * ip_reass_packet: generic routine to perform IP reassembly.
606  *
607  * => Passed fragment should have IP_MF flag and/or offset set.
608  * => Fragment should not have other than IP_MF flags set.
609  *
610  * => Returns 0 on success or error otherwise.
611  * => On complete, m0 represents a constructed final packet.
612  */
613 int
614 ip_reass_packet(struct mbuf **m0, struct ip *ip)
615 {
616 	const int hlen = ip->ip_hl << 2;
617 	const int len = ntohs(ip->ip_len);
618 	struct mbuf *m = *m0;
619 	ipfr_queue_t *fp;
620 	ipfr_qent_t *ipqe;
621 	u_int hash, off, flen;
622 	bool mff;
623 
624 	/*
625 	 * Prevent TCP blind data attacks by not allowing non-initial
626 	 * fragments to start at less than 68 bytes (minimal fragment
627 	 * size) and making sure the first fragment is at least 68
628 	 * bytes.
629 	 */
630 	off = (ntohs(ip->ip_off) & IP_OFFMASK) << 3;
631 	if ((off > 0 ? off + hlen : len) < IP_MINFRAGSIZE - 1) {
632 		IP_STATINC(IP_STAT_BADFRAGS);
633 		return EINVAL;
634 	}
635 
636 	/*
637 	 * Fragment length and MF flag.  Make sure that fragments have
638 	 * a data length which is non-zero and multiple of 8 bytes.
639 	 */
640 	flen = ntohs(ip->ip_len) - hlen;
641 	mff = (ip->ip_off & htons(IP_MF)) != 0;
642 	if (mff && (flen == 0 || (flen & 0x7) != 0)) {
643 		IP_STATINC(IP_STAT_BADFRAGS);
644 		return EINVAL;
645 	}
646 
647 	/*
648 	 * Adjust total IP length to not reflect header and convert
649 	 * offset of this to bytes.  XXX: clobbers struct ip.
650 	 */
651 	ip->ip_len = htons(flen);
652 	ip->ip_off = htons(off);
653 
654 	/* Look for queue of fragments of this datagram. */
655 	mutex_enter(&ipfr_lock);
656 	hash = IPREASS_HASH(ip->ip_src.s_addr, ip->ip_id);
657 	LIST_FOREACH(fp, &ip_frags[hash], ipq_q) {
658 		if (ip->ip_id != fp->ipq_id)
659 			continue;
660 		if (!in_hosteq(ip->ip_src, fp->ipq_src))
661 			continue;
662 		if (!in_hosteq(ip->ip_dst, fp->ipq_dst))
663 			continue;
664 		if (ip->ip_p != fp->ipq_p)
665 			continue;
666 		break;
667 	}
668 
669 	/* Make sure that TOS matches previous fragments. */
670 	if (fp && fp->ipq_tos != ip->ip_tos) {
671 		IP_STATINC(IP_STAT_BADFRAGS);
672 		mutex_exit(&ipfr_lock);
673 		return EINVAL;
674 	}
675 
676 	/*
677 	 * Create new entry and attempt to reassembly.
678 	 */
679 	IP_STATINC(IP_STAT_FRAGMENTS);
680 	ipqe = pool_cache_get(ipfren_cache, PR_NOWAIT);
681 	if (ipqe == NULL) {
682 		IP_STATINC(IP_STAT_RCVMEMDROP);
683 		mutex_exit(&ipfr_lock);
684 		return ENOMEM;
685 	}
686 	ipqe->ipqe_mff = mff;
687 	ipqe->ipqe_m = m;
688 	ipqe->ipqe_ip = ip;
689 
690 	*m0 = ip_reass(ipqe, fp, hash);
691 	if (*m0) {
692 		/* Note that finally reassembled. */
693 		IP_STATINC(IP_STAT_REASSEMBLED);
694 	}
695 	return 0;
696 }
697