xref: /netbsd/sys/netinet/ip_reass.c (revision b8014fa8) 
1 /*	$NetBSD: ip_reass.c,v 1.10 2016/04/26 08:44:44 ozaki-r 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 1.10 2016/04/26 08:44:44 ozaki-r 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 
66 #include <netinet/in.h>
67 #include <netinet/in_systm.h>
68 #include <netinet/ip.h>
69 #include <netinet/in_pcb.h>
70 #include <netinet/ip_var.h>
71 #include <netinet/in_proto.h>
72 #include <netinet/ip_private.h>
73 #include <netinet/in_var.h>
74 
75 /*
76  * IP reassembly queue structures.  Each fragment being reassembled is
77  * attached to one of these structures.  They are timed out after TTL
78  * drops to 0, and may also be reclaimed if memory becomes tight.
79  */
80 
81 typedef struct ipfr_qent {
82 	TAILQ_ENTRY(ipfr_qent)	ipqe_q;
83 	struct ip *		ipqe_ip;
84 	struct mbuf *		ipqe_m;
85 	bool			ipqe_mff;
86 } ipfr_qent_t;
87 
88 TAILQ_HEAD(ipfr_qent_head, ipfr_qent);
89 
90 typedef struct ipfr_queue {
91 	LIST_ENTRY(ipfr_queue)	ipq_q;		/* to other reass headers */
92 	struct ipfr_qent_head	ipq_fragq;	/* queue of fragment entries */
93 	uint8_t			ipq_ttl;	/* time for reass q to live */
94 	uint8_t			ipq_p;		/* protocol of this fragment */
95 	uint16_t		ipq_id;		/* sequence id for reassembly */
96 	struct in_addr		ipq_src;
97 	struct in_addr		ipq_dst;
98 	uint16_t		ipq_nfrags;	/* frags in this queue entry */
99 	uint8_t 		ipq_tos;	/* TOS of this fragment */
100 } ipfr_queue_t;
101 
102 /*
103  * Hash table of IP reassembly queues.
104  */
105 #define	IPREASS_HASH_SHIFT	6
106 #define	IPREASS_HASH_SIZE	(1 << IPREASS_HASH_SHIFT)
107 #define	IPREASS_HASH_MASK	(IPREASS_HASH_SIZE - 1)
108 #define	IPREASS_HASH(x, y) \
109 	(((((x) & 0xf) | ((((x) >> 8) & 0xf) << 4)) ^ (y)) & IPREASS_HASH_MASK)
110 
111 static LIST_HEAD(, ipfr_queue)	ip_frags[IPREASS_HASH_SIZE];
112 static pool_cache_t	ipfren_cache;
113 static kmutex_t		ipfr_lock;
114 
115 /* Number of packets in reassembly queue and total number of fragments. */
116 static int		ip_nfragpackets;
117 static int		ip_nfrags;
118 
119 /* Limits on packet and fragments. */
120 static int		ip_maxfragpackets;
121 static int		ip_maxfrags;
122 
123 /*
124  * Cached copy of nmbclusters.  If nbclusters is different, recalculate
125  * IP parameters derived from nmbclusters.
126  */
127 static int		ip_nmbclusters;
128 
129 /*
130  * IP reassembly TTL machinery for multiplicative drop.
131  */
132 static u_int		fragttl_histo[IPFRAGTTL + 1];
133 
134 static struct sysctllog *ip_reass_sysctllog;
135 
136 void			sysctl_ip_reass_setup(void);
137 static void		ip_nmbclusters_changed(void);
138 
139 static struct mbuf *	ip_reass(ipfr_qent_t *, ipfr_queue_t *, u_int);
140 static u_int		ip_reass_ttl_decr(u_int ticks);
141 static void		ip_reass_drophalf(void);
142 static void		ip_freef(ipfr_queue_t *);
143 
144 /*
145  * ip_reass_init:
146  *
147  *	Initialization of IP reassembly mechanism.
148  */
149 void
150 ip_reass_init(void)
151 {
152 	int i;
153 
154 	ipfren_cache = pool_cache_init(sizeof(ipfr_qent_t), coherency_unit,
155 	    0, 0, "ipfrenpl", NULL, IPL_NET, NULL, NULL, NULL);
156 	mutex_init(&ipfr_lock, MUTEX_DEFAULT, IPL_VM);
157 
158 	for (i = 0; i < IPREASS_HASH_SIZE; i++) {
159 		LIST_INIT(&ip_frags[i]);
160 	}
161 	ip_maxfragpackets = 200;
162 	ip_maxfrags = 0;
163 	ip_nmbclusters_changed();
164 
165 	sysctl_ip_reass_setup();
166 }
167 
168 void
169 sysctl_ip_reass_setup(void)
170 {
171 
172 	sysctl_createv(&ip_reass_sysctllog, 0, NULL, NULL,
173 		CTLFLAG_PERMANENT,
174 		CTLTYPE_NODE, "inet",
175 		SYSCTL_DESCR("PF_INET related settings"),
176 		NULL, 0, NULL, 0,
177 		CTL_NET, PF_INET, CTL_EOL);
178 	sysctl_createv(&ip_reass_sysctllog, 0, NULL, NULL,
179 		CTLFLAG_PERMANENT,
180 		CTLTYPE_NODE, "ip",
181 		SYSCTL_DESCR("IPv4 related settings"),
182 		NULL, 0, NULL, 0,
183 		CTL_NET, PF_INET, IPPROTO_IP, CTL_EOL);
184 
185 	sysctl_createv(&ip_reass_sysctllog, 0, NULL, NULL,
186 		CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
187 		CTLTYPE_INT, "maxfragpackets",
188 		SYSCTL_DESCR("Maximum number of fragments to retain for "
189 			     "possible reassembly"),
190 		NULL, 0, &ip_maxfragpackets, 0,
191 		CTL_NET, PF_INET, IPPROTO_IP, IPCTL_MAXFRAGPACKETS, CTL_EOL);
192 }
193 
194 #define CHECK_NMBCLUSTER_PARAMS()				\
195 do {								\
196 	if (__predict_false(ip_nmbclusters != nmbclusters))	\
197 		ip_nmbclusters_changed();			\
198 } while (/*CONSTCOND*/0)
199 
200 /*
201  * Compute IP limits derived from the value of nmbclusters.
202  */
203 static void
204 ip_nmbclusters_changed(void)
205 {
206 	ip_maxfrags = nmbclusters / 4;
207 	ip_nmbclusters = nmbclusters;
208 }
209 
210 /*
211  * ip_reass:
212  *
213  *	Take incoming datagram fragment and try to reassemble it into whole
214  *	datagram.  If a chain for reassembly of this datagram already exists,
215  *	then it is given as 'fp'; otherwise have to make a chain.
216  */
217 struct mbuf *
218 ip_reass(ipfr_qent_t *ipqe, ipfr_queue_t *fp, const u_int hash)
219 {
220 	struct ip *ip = ipqe->ipqe_ip, *qip;
221 	const int hlen = ip->ip_hl << 2;
222 	struct mbuf *m = ipqe->ipqe_m, *t;
223 	ipfr_qent_t *nq, *p, *q;
224 	int i, next;
225 
226 	KASSERT(mutex_owned(&ipfr_lock));
227 
228 	/*
229 	 * Presence of header sizes in mbufs would confuse code below.
230 	 */
231 	m->m_data += hlen;
232 	m->m_len -= hlen;
233 
234 #ifdef	notyet
235 	/* Make sure fragment limit is up-to-date. */
236 	CHECK_NMBCLUSTER_PARAMS();
237 
238 	/* If we have too many fragments, drop the older half. */
239 	if (ip_nfrags >= ip_maxfrags) {
240 		ip_reass_drophalf(void);
241 	}
242 #endif
243 
244 	/*
245 	 * We are about to add a fragment; increment frag count.
246 	 */
247 	ip_nfrags++;
248 
249 	/*
250 	 * If first fragment to arrive, create a reassembly queue.
251 	 */
252 	if (fp == NULL) {
253 		/*
254 		 * Enforce upper bound on number of fragmented packets
255 		 * for which we attempt reassembly:  a) if maxfrag is 0,
256 		 * never accept fragments  b) if maxfrag is -1, accept
257 		 * all fragments without limitation.
258 		 */
259 		if (ip_maxfragpackets < 0)
260 			;
261 		else if (ip_nfragpackets >= ip_maxfragpackets) {
262 			goto dropfrag;
263 		}
264 		fp = malloc(sizeof(ipfr_queue_t), M_FTABLE, M_NOWAIT);
265 		if (fp == NULL) {
266 			goto dropfrag;
267 		}
268 		ip_nfragpackets++;
269 		TAILQ_INIT(&fp->ipq_fragq);
270 		fp->ipq_nfrags = 1;
271 		fp->ipq_ttl = IPFRAGTTL;
272 		fp->ipq_p = ip->ip_p;
273 		fp->ipq_id = ip->ip_id;
274 		fp->ipq_tos = ip->ip_tos;
275 		fp->ipq_src = ip->ip_src;
276 		fp->ipq_dst = ip->ip_dst;
277 		LIST_INSERT_HEAD(&ip_frags[hash], fp, ipq_q);
278 		p = NULL;
279 		goto insert;
280 	} else {
281 		fp->ipq_nfrags++;
282 	}
283 
284 	/*
285 	 * Find a segment which begins after this one does.
286 	 */
287 	TAILQ_FOREACH(q, &fp->ipq_fragq, ipqe_q) {
288 		if (ntohs(q->ipqe_ip->ip_off) > ntohs(ip->ip_off))
289 			break;
290 	}
291 	if (q != NULL) {
292 		p = TAILQ_PREV(q, ipfr_qent_head, ipqe_q);
293 	} else {
294 		p = TAILQ_LAST(&fp->ipq_fragq, ipfr_qent_head);
295 	}
296 
297 	/*
298 	 * If there is a preceding segment, it may provide some of our
299 	 * data already.  If so, drop the data from the incoming segment.
300 	 * If it provides all of our data, drop us.
301 	 */
302 	if (p != NULL) {
303 		i = ntohs(p->ipqe_ip->ip_off) + ntohs(p->ipqe_ip->ip_len) -
304 		    ntohs(ip->ip_off);
305 		if (i > 0) {
306 			if (i >= ntohs(ip->ip_len)) {
307 				goto dropfrag;
308 			}
309 			m_adj(ipqe->ipqe_m, i);
310 			ip->ip_off = htons(ntohs(ip->ip_off) + i);
311 			ip->ip_len = htons(ntohs(ip->ip_len) - i);
312 		}
313 	}
314 
315 	/*
316 	 * While we overlap succeeding segments trim them or, if they are
317 	 * completely covered, dequeue them.
318 	 */
319 	while (q != NULL) {
320 		size_t end;
321 
322 		qip = q->ipqe_ip;
323 		end = ntohs(ip->ip_off) + ntohs(ip->ip_len);
324 		if (end <= ntohs(qip->ip_off)) {
325 			break;
326 		}
327 		i = end - ntohs(qip->ip_off);
328 		if (i < ntohs(qip->ip_len)) {
329 			qip->ip_len = htons(ntohs(qip->ip_len) - i);
330 			qip->ip_off = htons(ntohs(qip->ip_off) + i);
331 			m_adj(q->ipqe_m, i);
332 			break;
333 		}
334 		nq = TAILQ_NEXT(q, ipqe_q);
335 		m_freem(q->ipqe_m);
336 		TAILQ_REMOVE(&fp->ipq_fragq, q, ipqe_q);
337 		pool_cache_put(ipfren_cache, q);
338 		fp->ipq_nfrags--;
339 		ip_nfrags--;
340 		q = nq;
341 	}
342 
343 insert:
344 	/*
345 	 * Stick new segment in its place; check for complete reassembly.
346 	 */
347 	if (p == NULL) {
348 		TAILQ_INSERT_HEAD(&fp->ipq_fragq, ipqe, ipqe_q);
349 	} else {
350 		TAILQ_INSERT_AFTER(&fp->ipq_fragq, p, ipqe, ipqe_q);
351 	}
352 	next = 0;
353 	TAILQ_FOREACH(q, &fp->ipq_fragq, ipqe_q) {
354 		qip = q->ipqe_ip;
355 		if (ntohs(qip->ip_off) != next) {
356 			mutex_exit(&ipfr_lock);
357 			return NULL;
358 		}
359 		next += ntohs(qip->ip_len);
360 	}
361 	p = TAILQ_LAST(&fp->ipq_fragq, ipfr_qent_head);
362 	if (p->ipqe_mff) {
363 		mutex_exit(&ipfr_lock);
364 		return NULL;
365 	}
366 
367 	/*
368 	 * Reassembly is complete.  Check for a bogus message size.
369 	 */
370 	q = TAILQ_FIRST(&fp->ipq_fragq);
371 	ip = q->ipqe_ip;
372 	if ((next + (ip->ip_hl << 2)) > IP_MAXPACKET) {
373 		IP_STATINC(IP_STAT_TOOLONG);
374 		ip_freef(fp);
375 		mutex_exit(&ipfr_lock);
376 		return NULL;
377 	}
378 	LIST_REMOVE(fp, ipq_q);
379 	ip_nfrags -= fp->ipq_nfrags;
380 	ip_nfragpackets--;
381 	mutex_exit(&ipfr_lock);
382 
383 	/* Concatenate all fragments. */
384 	m = q->ipqe_m;
385 	t = m->m_next;
386 	m->m_next = NULL;
387 	m_cat(m, t);
388 	nq = TAILQ_NEXT(q, ipqe_q);
389 	pool_cache_put(ipfren_cache, q);
390 
391 	for (q = nq; q != NULL; q = nq) {
392 		t = q->ipqe_m;
393 		nq = TAILQ_NEXT(q, ipqe_q);
394 		pool_cache_put(ipfren_cache, q);
395 		m_cat(m, t);
396 	}
397 
398 	/*
399 	 * Create header for new packet by modifying header of first
400 	 * packet.  Dequeue and discard fragment reassembly header.  Make
401 	 * header visible.
402 	 */
403 	ip->ip_len = htons((ip->ip_hl << 2) + next);
404 	ip->ip_src = fp->ipq_src;
405 	ip->ip_dst = fp->ipq_dst;
406 	free(fp, M_FTABLE);
407 
408 	m->m_len += (ip->ip_hl << 2);
409 	m->m_data -= (ip->ip_hl << 2);
410 
411 	/* Fix up mbuf.  XXX This should be done elsewhere. */
412 	if (m->m_flags & M_PKTHDR) {
413 		int plen = 0;
414 		for (t = m; t; t = t->m_next) {
415 			plen += t->m_len;
416 		}
417 		m->m_pkthdr.len = plen;
418 		m->m_pkthdr.csum_flags = 0;
419 	}
420 	return m;
421 
422 dropfrag:
423 	if (fp != NULL) {
424 		fp->ipq_nfrags--;
425 	}
426 	ip_nfrags--;
427 	IP_STATINC(IP_STAT_FRAGDROPPED);
428 	mutex_exit(&ipfr_lock);
429 
430 	pool_cache_put(ipfren_cache, ipqe);
431 	m_freem(m);
432 	return NULL;
433 }
434 
435 /*
436  * ip_freef:
437  *
438  *	Free a fragment reassembly header and all associated datagrams.
439  */
440 static void
441 ip_freef(ipfr_queue_t *fp)
442 {
443 	ipfr_qent_t *q;
444 
445 	KASSERT(mutex_owned(&ipfr_lock));
446 
447 	LIST_REMOVE(fp, ipq_q);
448 	ip_nfrags -= fp->ipq_nfrags;
449 	ip_nfragpackets--;
450 
451 	while ((q = TAILQ_FIRST(&fp->ipq_fragq)) != NULL) {
452 		TAILQ_REMOVE(&fp->ipq_fragq, q, ipqe_q);
453 		m_freem(q->ipqe_m);
454 		pool_cache_put(ipfren_cache, q);
455 	}
456 	free(fp, M_FTABLE);
457 }
458 
459 /*
460  * ip_reass_ttl_decr:
461  *
462  *	Decrement TTL of all reasembly queue entries by `ticks'.  Count
463  *	number of distinct fragments (as opposed to partial, fragmented
464  *	datagrams) inthe reassembly queue.  While we  traverse the entire
465  *	reassembly queue, compute and return the median TTL over all
466  *	fragments.
467  */
468 static u_int
469 ip_reass_ttl_decr(u_int ticks)
470 {
471 	u_int nfrags, median, dropfraction, keepfraction;
472 	ipfr_queue_t *fp, *nfp;
473 	int i;
474 
475 	nfrags = 0;
476 	memset(fragttl_histo, 0, sizeof(fragttl_histo));
477 
478 	for (i = 0; i < IPREASS_HASH_SIZE; i++) {
479 		for (fp = LIST_FIRST(&ip_frags[i]); fp != NULL; fp = nfp) {
480 			fp->ipq_ttl = ((fp->ipq_ttl <= ticks) ?
481 			    0 : fp->ipq_ttl - ticks);
482 			nfp = LIST_NEXT(fp, ipq_q);
483 			if (fp->ipq_ttl == 0) {
484 				IP_STATINC(IP_STAT_FRAGTIMEOUT);
485 				ip_freef(fp);
486 			} else {
487 				nfrags += fp->ipq_nfrags;
488 				fragttl_histo[fp->ipq_ttl] += fp->ipq_nfrags;
489 			}
490 		}
491 	}
492 
493 	KASSERT(ip_nfrags == nfrags);
494 
495 	/* Find median (or other drop fraction) in histogram. */
496 	dropfraction = (ip_nfrags / 2);
497 	keepfraction = ip_nfrags - dropfraction;
498 	for (i = IPFRAGTTL, median = 0; i >= 0; i--) {
499 		median += fragttl_histo[i];
500 		if (median >= keepfraction)
501 			break;
502 	}
503 
504 	/* Return TTL of median (or other fraction). */
505 	return (u_int)i;
506 }
507 
508 static void
509 ip_reass_drophalf(void)
510 {
511 	u_int median_ticks;
512 
513 	KASSERT(mutex_owned(&ipfr_lock));
514 
515 	/*
516 	 * Compute median TTL of all fragments, and count frags
517 	 * with that TTL or lower (roughly half of all fragments).
518 	 */
519 	median_ticks = ip_reass_ttl_decr(0);
520 
521 	/* Drop half. */
522 	median_ticks = ip_reass_ttl_decr(median_ticks);
523 }
524 
525 /*
526  * ip_reass_drain: drain off all datagram fragments.  Do not acquire
527  * softnet_lock as can be called from hardware interrupt context.
528  */
529 void
530 ip_reass_drain(void)
531 {
532 
533 	/*
534 	 * We may be called from a device's interrupt context.  If
535 	 * the ipq is already busy, just bail out now.
536 	 */
537 	if (mutex_tryenter(&ipfr_lock)) {
538 		/*
539 		 * Drop half the total fragments now. If more mbufs are
540 		 * needed, we will be called again soon.
541 		 */
542 		ip_reass_drophalf();
543 		mutex_exit(&ipfr_lock);
544 	}
545 }
546 
547 /*
548  * ip_reass_slowtimo:
549  *
550  *	If a timer expires on a reassembly queue, discard it.
551  */
552 void
553 ip_reass_slowtimo(void)
554 {
555 	static u_int dropscanidx = 0;
556 	u_int i, median_ttl;
557 
558 	mutex_enter(&ipfr_lock);
559 
560 	/* Age TTL of all fragments by 1 tick .*/
561 	median_ttl = ip_reass_ttl_decr(1);
562 
563 	/* Make sure fragment limit is up-to-date. */
564 	CHECK_NMBCLUSTER_PARAMS();
565 
566 	/* If we have too many fragments, drop the older half. */
567 	if (ip_nfrags > ip_maxfrags) {
568 		ip_reass_ttl_decr(median_ttl);
569 	}
570 
571 	/*
572 	 * If we are over the maximum number of fragmented packets (due to
573 	 * the limit being lowered), drain off enough to get down to the
574 	 * new limit.  Start draining from the reassembly hashqueue most
575 	 * recently drained.
576 	 */
577 	if (ip_maxfragpackets < 0)
578 		;
579 	else {
580 		int wrapped = 0;
581 
582 		i = dropscanidx;
583 		while (ip_nfragpackets > ip_maxfragpackets && wrapped == 0) {
584 			while (LIST_FIRST(&ip_frags[i]) != NULL) {
585 				ip_freef(LIST_FIRST(&ip_frags[i]));
586 			}
587 			if (++i >= IPREASS_HASH_SIZE) {
588 				i = 0;
589 			}
590 			/*
591 			 * Do not scan forever even if fragment counters are
592 			 * wrong: stop after scanning entire reassembly queue.
593 			 */
594 			if (i == dropscanidx) {
595 				wrapped = 1;
596 			}
597 		}
598 		dropscanidx = i;
599 	}
600 	mutex_exit(&ipfr_lock);
601 }
602 
603 /*
604  * ip_reass_packet: generic routine to perform IP reassembly.
605  *
606  * => Passed fragment should have IP_MF flag and/or offset set.
607  * => Fragment should not have other than IP_MF flags set.
608  *
609  * => Returns 0 on success or error otherwise.
610  * => On complete, m0 represents a constructed final packet.
611  */
612 int
613 ip_reass_packet(struct mbuf **m0, struct ip *ip)
614 {
615 	const int hlen = ip->ip_hl << 2;
616 	const int len = ntohs(ip->ip_len);
617 	struct mbuf *m = *m0;
618 	ipfr_queue_t *fp;
619 	ipfr_qent_t *ipqe;
620 	u_int hash, off, flen;
621 	bool mff;
622 
623 	/*
624 	 * Prevent TCP blind data attacks by not allowing non-initial
625 	 * fragments to start at less than 68 bytes (minimal fragment
626 	 * size) and making sure the first fragment is at least 68
627 	 * bytes.
628 	 */
629 	off = (ntohs(ip->ip_off) & IP_OFFMASK) << 3;
630 	if ((off > 0 ? off + hlen : len) < IP_MINFRAGSIZE - 1) {
631 		IP_STATINC(IP_STAT_BADFRAGS);
632 		return EINVAL;
633 	}
634 
635 	/*
636 	 * Fragment length and MF flag.  Make sure that fragments have
637 	 * a data length which is non-zero and multiple of 8 bytes.
638 	 */
639 	flen = ntohs(ip->ip_len) - hlen;
640 	mff = (ip->ip_off & htons(IP_MF)) != 0;
641 	if (mff && (flen == 0 || (flen & 0x7) != 0)) {
642 		IP_STATINC(IP_STAT_BADFRAGS);
643 		return EINVAL;
644 	}
645 
646 	/*
647 	 * Adjust total IP length to not reflect header and convert
648 	 * offset of this to bytes.  XXX: clobbers struct ip.
649 	 */
650 	ip->ip_len = htons(flen);
651 	ip->ip_off = htons(off);
652 
653 	/* Look for queue of fragments of this datagram. */
654 	mutex_enter(&ipfr_lock);
655 	hash = IPREASS_HASH(ip->ip_src.s_addr, ip->ip_id);
656 	LIST_FOREACH(fp, &ip_frags[hash], ipq_q) {
657 		if (ip->ip_id != fp->ipq_id)
658 			continue;
659 		if (!in_hosteq(ip->ip_src, fp->ipq_src))
660 			continue;
661 		if (!in_hosteq(ip->ip_dst, fp->ipq_dst))
662 			continue;
663 		if (ip->ip_p != fp->ipq_p)
664 			continue;
665 		break;
666 	}
667 
668 	/* Make sure that TOS matches previous fragments. */
669 	if (fp && fp->ipq_tos != ip->ip_tos) {
670 		IP_STATINC(IP_STAT_BADFRAGS);
671 		mutex_exit(&ipfr_lock);
672 		return EINVAL;
673 	}
674 
675 	/*
676 	 * Create new entry and attempt to reassembly.
677 	 */
678 	IP_STATINC(IP_STAT_FRAGMENTS);
679 	ipqe = pool_cache_get(ipfren_cache, PR_NOWAIT);
680 	if (ipqe == NULL) {
681 		IP_STATINC(IP_STAT_RCVMEMDROP);
682 		mutex_exit(&ipfr_lock);
683 		return ENOMEM;
684 	}
685 	ipqe->ipqe_mff = mff;
686 	ipqe->ipqe_m = m;
687 	ipqe->ipqe_ip = ip;
688 
689 	*m0 = ip_reass(ipqe, fp, hash);
690 	if (*m0) {
691 		/* Note that finally reassembled. */
692 		IP_STATINC(IP_STAT_REASSEMBLED);
693 	}
694 	return 0;
695 }
696