xref: /freebsd/sys/netinet6/frag6.c (revision edca4938f74db18d091868237592abbf7e718669)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
5  * 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 project 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 PROJECT 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 PROJECT 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  *	$KAME: frag6.c,v 1.33 2002/01/07 11:34:48 kjc Exp $
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include "opt_rss.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/domain.h>
42 #include <sys/eventhandler.h>
43 #include <sys/hash.h>
44 #include <sys/kernel.h>
45 #include <sys/malloc.h>
46 #include <sys/mbuf.h>
47 #include <sys/protosw.h>
48 #include <sys/socket.h>
49 #include <sys/sysctl.h>
50 #include <sys/syslog.h>
51 
52 #include <net/if.h>
53 #include <net/if_var.h>
54 #include <net/netisr.h>
55 #include <net/route.h>
56 #include <net/vnet.h>
57 
58 #include <netinet/in.h>
59 #include <netinet/in_var.h>
60 #include <netinet/ip6.h>
61 #include <netinet6/ip6_var.h>
62 #include <netinet/icmp6.h>
63 #include <netinet/in_systm.h>	/* For ECN definitions. */
64 #include <netinet/ip.h>		/* For ECN definitions. */
65 
66 #ifdef MAC
67 #include <security/mac/mac_framework.h>
68 #endif
69 
70 /* Reassembly headers are stored in hash buckets. */
71 #define	IP6REASS_NHASH_LOG2	10
72 #define	IP6REASS_NHASH		(1 << IP6REASS_NHASH_LOG2)
73 #define	IP6REASS_HMASK		(IP6REASS_NHASH - 1)
74 
75 static void frag6_enq(struct ip6asfrag *, struct ip6asfrag *,
76     uint32_t bucket __unused);
77 static void frag6_deq(struct ip6asfrag *, uint32_t bucket __unused);
78 static void frag6_insque_head(struct ip6q *, struct ip6q *,
79     uint32_t bucket);
80 static void frag6_remque(struct ip6q *, uint32_t bucket);
81 static void frag6_freef(struct ip6q *, uint32_t bucket);
82 
83 struct ip6qbucket {
84 	struct ip6q	ip6q;
85 	struct mtx	lock;
86 	int		count;
87 };
88 
89 struct	ip6asfrag {
90 	struct ip6asfrag *ip6af_down;
91 	struct ip6asfrag *ip6af_up;
92 	struct mbuf	*ip6af_m;
93 	int		ip6af_offset;	/* offset in ip6af_m to next header */
94 	int		ip6af_frglen;	/* fragmentable part length */
95 	int		ip6af_off;	/* fragment offset */
96 	u_int16_t	ip6af_mff;	/* more fragment bit in frag off */
97 };
98 
99 #define IP6_REASS_MBUF(ip6af) (*(struct mbuf **)&((ip6af)->ip6af_m))
100 
101 static MALLOC_DEFINE(M_FRAG6, "frag6", "IPv6 fragment reassembly header");
102 
103 /* System wide (global) maximum and count of packets in reassembly queues. */
104 static int ip6_maxfrags;
105 static volatile u_int frag6_nfrags = 0;
106 
107 /* Maximum and current packets in per-VNET reassembly queue. */
108 VNET_DEFINE_STATIC(int,			ip6_maxfragpackets);
109 VNET_DEFINE_STATIC(volatile u_int,	frag6_nfragpackets);
110 #define	V_ip6_maxfragpackets		VNET(ip6_maxfragpackets)
111 #define	V_frag6_nfragpackets		VNET(frag6_nfragpackets)
112 
113 /* Maximum per-VNET reassembly queues per bucket and fragments per packet. */
114 VNET_DEFINE_STATIC(int,			ip6_maxfragbucketsize);
115 VNET_DEFINE_STATIC(int,			ip6_maxfragsperpacket);
116 #define	V_ip6_maxfragbucketsize		VNET(ip6_maxfragbucketsize)
117 #define	V_ip6_maxfragsperpacket		VNET(ip6_maxfragsperpacket)
118 
119 /* Per-VNET reassembly queue buckets. */
120 VNET_DEFINE_STATIC(struct ip6qbucket,	ip6qb[IP6REASS_NHASH]);
121 VNET_DEFINE_STATIC(uint32_t,		ip6qb_hashseed);
122 #define	V_ip6qb				VNET(ip6qb)
123 #define	V_ip6qb_hashseed		VNET(ip6qb_hashseed)
124 
125 #define	IP6QB_LOCK(_b)		mtx_lock(&V_ip6qb[(_b)].lock)
126 #define	IP6QB_TRYLOCK(_b)	mtx_trylock(&V_ip6qb[(_b)].lock)
127 #define	IP6QB_LOCK_ASSERT(_b)	mtx_assert(&V_ip6qb[(_b)].lock, MA_OWNED)
128 #define	IP6QB_UNLOCK(_b)	mtx_unlock(&V_ip6qb[(_b)].lock)
129 #define	IP6QB_HEAD(_b)		(&V_ip6qb[(_b)].ip6q)
130 
131 /*
132  * By default, limit the number of IP6 fragments across all reassembly
133  * queues to  1/32 of the total number of mbuf clusters.
134  *
135  * Limit the total number of reassembly queues per VNET to the
136  * IP6 fragment limit, but ensure the limit will not allow any bucket
137  * to grow above 100 items. (The bucket limit is
138  * IP_MAXFRAGPACKETS / (IPREASS_NHASH / 2), so the 50 is the correct
139  * multiplier to reach a 100-item limit.)
140  * The 100-item limit was chosen as brief testing seems to show that
141  * this produces "reasonable" performance on some subset of systems
142  * under DoS attack.
143  */
144 #define	IP6_MAXFRAGS		(nmbclusters / 32)
145 #define	IP6_MAXFRAGPACKETS	(imin(IP6_MAXFRAGS, IP6REASS_NHASH * 50))
146 
147 
148 /*
149  * Sysctls and helper function.
150  */
151 SYSCTL_DECL(_net_inet6_ip6);
152 
153 static void
154 frag6_set_bucketsize(void)
155 {
156 	int i;
157 
158 	if ((i = V_ip6_maxfragpackets) > 0)
159 		V_ip6_maxfragbucketsize = imax(i / (IP6REASS_NHASH / 2), 1);
160 }
161 
162 SYSCTL_INT(_net_inet6_ip6, IPV6CTL_MAXFRAGS, maxfrags,
163 	CTLFLAG_RW, &ip6_maxfrags, 0,
164 	"Maximum allowed number of outstanding IPv6 packet fragments. "
165 	"A value of 0 means no fragmented packets will be accepted, while a "
166 	"a value of -1 means no limit");
167 
168 static int
169 sysctl_ip6_maxfragpackets(SYSCTL_HANDLER_ARGS)
170 {
171 	int error, val;
172 
173 	val = V_ip6_maxfragpackets;
174 	error = sysctl_handle_int(oidp, &val, 0, req);
175 	if (error != 0 || !req->newptr)
176 		return (error);
177 	V_ip6_maxfragpackets = val;
178 	frag6_set_bucketsize();
179 	return (0);
180 }
181 SYSCTL_PROC(_net_inet6_ip6, IPV6CTL_MAXFRAGPACKETS, maxfragpackets,
182 	CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW, NULL, 0,
183 	sysctl_ip6_maxfragpackets, "I",
184 	"Default maximum number of outstanding fragmented IPv6 packets. "
185 	"A value of 0 means no fragmented packets will be accepted, while a "
186 	"a value of -1 means no limit");
187 SYSCTL_INT(_net_inet6_ip6, IPV6CTL_MAXFRAGSPERPACKET, maxfragsperpacket,
188 	CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ip6_maxfragsperpacket), 0,
189 	"Maximum allowed number of fragments per packet");
190 SYSCTL_INT(_net_inet6_ip6, IPV6CTL_MAXFRAGBUCKETSIZE, maxfragbucketsize,
191 	CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ip6_maxfragbucketsize), 0,
192 	"Maximum number of reassembly queues per hash bucket");
193 
194 
195 /*
196  * Remove the IPv6 fragmentation header from the mbuf.
197  */
198 int
199 ip6_deletefraghdr(struct mbuf *m, int offset, int wait)
200 {
201 	struct ip6_hdr *ip6;
202 	struct mbuf *t;
203 
204 	/* Delete frag6 header. */
205 	if (m->m_len >= offset + sizeof(struct ip6_frag)) {
206 
207 		/* This is the only possible case with !PULLDOWN_TEST. */
208 		ip6  = mtod(m, struct ip6_hdr *);
209 		bcopy(ip6, (char *)ip6 + sizeof(struct ip6_frag),
210 		    offset);
211 		m->m_data += sizeof(struct ip6_frag);
212 		m->m_len -= sizeof(struct ip6_frag);
213 	} else {
214 
215 		/* This comes with no copy if the boundary is on cluster. */
216 		if ((t = m_split(m, offset, wait)) == NULL)
217 			return (ENOMEM);
218 		m_adj(t, sizeof(struct ip6_frag));
219 		m_cat(m, t);
220 	}
221 
222 	m->m_flags |= M_FRAGMENTED;
223 	return (0);
224 }
225 
226 /*
227  * Free a fragment reassembly header and all associated datagrams.
228  */
229 static void
230 frag6_freef(struct ip6q *q6, uint32_t bucket)
231 {
232 	struct ip6_hdr *ip6;
233 	struct ip6asfrag *af6, *down6;
234 	struct mbuf *m;
235 
236 	IP6QB_LOCK_ASSERT(bucket);
237 
238 	for (af6 = q6->ip6q_down; af6 != (struct ip6asfrag *)q6;
239 	     af6 = down6) {
240 
241 		m = IP6_REASS_MBUF(af6);
242 		down6 = af6->ip6af_down;
243 		frag6_deq(af6, bucket);
244 
245 		/*
246 		 * Return ICMP time exceeded error for the 1st fragment.
247 		 * Just free other fragments.
248 		 */
249 		if (af6->ip6af_off == 0 && m->m_pkthdr.rcvif != NULL) {
250 
251 			/* Adjust pointer. */
252 			ip6 = mtod(m, struct ip6_hdr *);
253 
254 			/* Restore source and destination addresses. */
255 			ip6->ip6_src = q6->ip6q_src;
256 			ip6->ip6_dst = q6->ip6q_dst;
257 
258 			icmp6_error(m, ICMP6_TIME_EXCEEDED,
259 			    ICMP6_TIME_EXCEED_REASSEMBLY, 0);
260 		} else
261 			m_freem(m);
262 
263 		free(af6, M_FRAG6);
264 	}
265 	frag6_remque(q6, bucket);
266 	atomic_subtract_int(&frag6_nfrags, q6->ip6q_nfrag);
267 #ifdef MAC
268 	mac_ip6q_destroy(q6);
269 #endif
270 	free(q6, M_FRAG6);
271 	atomic_subtract_int(&V_frag6_nfragpackets, 1);
272 }
273 
274 /*
275  * Drain off all datagram fragments belonging to
276  * the given network interface.
277  */
278 static void
279 frag6_cleanup(void *arg __unused, struct ifnet *ifp)
280 {
281 	struct ip6q *q6, *q6n, *head;
282 	struct ip6asfrag *af6;
283 	struct mbuf *m;
284 	int i;
285 
286 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
287 
288 	CURVNET_SET_QUIET(ifp->if_vnet);
289 	for (i = 0; i < IP6REASS_NHASH; i++) {
290 		IP6QB_LOCK(i);
291 		head = IP6QB_HEAD(i);
292 		/* Scan fragment list. */
293 		for (q6 = head->ip6q_next; q6 != head; q6 = q6n) {
294 			q6n = q6->ip6q_next;
295 
296 			for (af6 = q6->ip6q_down; af6 != (struct ip6asfrag *)q6;
297 			     af6 = af6->ip6af_down) {
298 				m = IP6_REASS_MBUF(af6);
299 
300 				/* clear no longer valid rcvif pointer */
301 				if (m->m_pkthdr.rcvif == ifp)
302 					m->m_pkthdr.rcvif = NULL;
303 			}
304 		}
305 		IP6QB_UNLOCK(i);
306 	}
307 	CURVNET_RESTORE();
308 }
309 EVENTHANDLER_DEFINE(ifnet_departure_event, frag6_cleanup, NULL, 0);
310 
311 /*
312  * Like in RFC2460, in RFC8200, fragment and reassembly rules do not agree with
313  * each other, in terms of next header field handling in fragment header.
314  * While the sender will use the same value for all of the fragmented packets,
315  * receiver is suggested not to check for consistency.
316  *
317  * Fragment rules (p18,p19):
318  *	(2)  A Fragment header containing:
319  *	The Next Header value that identifies the first header
320  *	after the Per-Fragment headers of the original packet.
321  *		-> next header field is same for all fragments
322  *
323  * Reassembly rule (p20):
324  *	The Next Header field of the last header of the Per-Fragment
325  *	headers is obtained from the Next Header field of the first
326  *	fragment's Fragment header.
327  *		-> should grab it from the first fragment only
328  *
329  * The following note also contradicts with fragment rule - no one is going to
330  * send different fragment with different next header field.
331  *
332  * Additional note (p22) [not an error]:
333  *	The Next Header values in the Fragment headers of different
334  *	fragments of the same original packet may differ.  Only the value
335  *	from the Offset zero fragment packet is used for reassembly.
336  *		-> should grab it from the first fragment only
337  *
338  * There is no explicit reason given in the RFC.  Historical reason maybe?
339  */
340 /*
341  * Fragment input.
342  */
343 int
344 frag6_input(struct mbuf **mp, int *offp, int proto)
345 {
346 	struct ifnet *dstifp;
347 	struct ifnet *srcifp;
348 	struct in6_ifaddr *ia6;
349 	struct ip6_hdr *ip6;
350 	struct ip6_frag *ip6f;
351 	struct ip6q *head, *q6;
352 	struct ip6asfrag *af6, *af6dwn, *ip6af;
353 	struct mbuf *m, *t;
354 	uint32_t hashkey[(sizeof(struct in6_addr) * 2 +
355 		    sizeof(ip6f->ip6f_ident)) / sizeof(uint32_t)];
356 	uint32_t bucket, *hashkeyp;
357 	int fragoff, frgpartlen;	/* Must be larger than uint16_t. */
358 	int nxt, offset, plen;
359 	uint8_t ecn, ecn0;
360 	bool only_frag;
361 #ifdef RSS
362 	struct ip6_direct_ctx *ip6dc;
363 	struct m_tag *mtag;
364 #endif
365 
366 	m = *mp;
367 	offset = *offp;
368 
369 	ip6 = mtod(m, struct ip6_hdr *);
370 #ifndef PULLDOWN_TEST
371 	IP6_EXTHDR_CHECK(m, offset, sizeof(struct ip6_frag), IPPROTO_DONE);
372 	ip6f = (struct ip6_frag *)((caddr_t)ip6 + offset);
373 #else
374 	IP6_EXTHDR_GET(ip6f, struct ip6_frag *, m, offset, sizeof(*ip6f));
375 	if (ip6f == NULL)
376 		return (IPPROTO_DONE);
377 #endif
378 
379 	/*
380 	 * Store receive network interface pointer for later.
381 	 */
382 	srcifp = m->m_pkthdr.rcvif;
383 
384 	dstifp = NULL;
385 	/* Find the destination interface of the packet. */
386 	ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */);
387 	if (ia6 != NULL) {
388 		dstifp = ia6->ia_ifp;
389 		ifa_free(&ia6->ia_ifa);
390 	}
391 
392 	/* Jumbo payload cannot contain a fragment header. */
393 	if (ip6->ip6_plen == 0) {
394 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER, offset);
395 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
396 		return (IPPROTO_DONE);
397 	}
398 
399 	/*
400 	 * Check whether fragment packet's fragment length is a
401 	 * multiple of 8 octets (unless it is the last one).
402 	 * sizeof(struct ip6_frag) == 8
403 	 * sizeof(struct ip6_hdr) = 40
404 	 */
405 	if ((ip6f->ip6f_offlg & IP6F_MORE_FRAG) &&
406 	    (((ntohs(ip6->ip6_plen) - offset) & 0x7) != 0)) {
407 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
408 		    offsetof(struct ip6_hdr, ip6_plen));
409 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
410 		return (IPPROTO_DONE);
411 	}
412 
413 	IP6STAT_INC(ip6s_fragments);
414 	in6_ifstat_inc(dstifp, ifs6_reass_reqd);
415 
416 	/* Offset now points to data portion. */
417 	offset += sizeof(struct ip6_frag);
418 
419 	/*
420 	 * Handle "atomic" fragments (offset and m bit set to 0) upfront,
421 	 * unrelated to any reassembly.  Still need to remove the frag hdr.
422 	 * See RFC 6946 and section 4.5 of RFC 8200.
423 	 */
424 	if ((ip6f->ip6f_offlg & ~IP6F_RESERVED_MASK) == 0) {
425 		IP6STAT_INC(ip6s_atomicfrags);
426 		/* XXX-BZ handle correctly. */
427 		in6_ifstat_inc(dstifp, ifs6_reass_ok);
428 		*offp = offset;
429 		m->m_flags |= M_FRAGMENTED;
430 		return (ip6f->ip6f_nxt);
431 	}
432 
433 	/* Get fragment length and discard 0-byte fragments. */
434 	frgpartlen = sizeof(struct ip6_hdr) + ntohs(ip6->ip6_plen) - offset;
435 	if (frgpartlen == 0) {
436 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
437 		    offsetof(struct ip6_hdr, ip6_plen));
438 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
439 		IP6STAT_INC(ip6s_fragdropped);
440 		return (IPPROTO_DONE);
441 	}
442 
443 	/* Generate a hash value for fragment bucket selection. */
444 	hashkeyp = hashkey;
445 	memcpy(hashkeyp, &ip6->ip6_src, sizeof(struct in6_addr));
446 	hashkeyp += sizeof(struct in6_addr) / sizeof(*hashkeyp);
447 	memcpy(hashkeyp, &ip6->ip6_dst, sizeof(struct in6_addr));
448 	hashkeyp += sizeof(struct in6_addr) / sizeof(*hashkeyp);
449 	*hashkeyp = ip6f->ip6f_ident;
450 	bucket = jenkins_hash32(hashkey, nitems(hashkey), V_ip6qb_hashseed);
451 	bucket &= IP6REASS_HMASK;
452 	head = IP6QB_HEAD(bucket);
453 	IP6QB_LOCK(bucket);
454 
455 	/*
456 	 * Enforce upper bound on number of fragments for the entire system.
457 	 * If maxfrag is 0, never accept fragments.
458 	 * If maxfrag is -1, accept all fragments without limitation.
459 	 */
460 	if (ip6_maxfrags < 0)
461 		;
462 	else if (atomic_load_int(&frag6_nfrags) >= (u_int)ip6_maxfrags)
463 		goto dropfrag;
464 
465 	for (q6 = head->ip6q_next; q6 != head; q6 = q6->ip6q_next)
466 		if (ip6f->ip6f_ident == q6->ip6q_ident &&
467 		    IN6_ARE_ADDR_EQUAL(&ip6->ip6_src, &q6->ip6q_src) &&
468 		    IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst, &q6->ip6q_dst)
469 #ifdef MAC
470 		    && mac_ip6q_match(m, q6)
471 #endif
472 		    )
473 			break;
474 
475 	only_frag = false;
476 	if (q6 == head) {
477 
478 		/* A first fragment to arrive creates a reassembly queue. */
479 		only_frag = true;
480 
481 		/*
482 		 * Enforce upper bound on number of fragmented packets
483 		 * for which we attempt reassembly;
484 		 * If maxfragpackets is 0, never accept fragments.
485 		 * If maxfragpackets is -1, accept all fragments without
486 		 * limitation.
487 		 */
488 		if (V_ip6_maxfragpackets < 0)
489 			;
490 		else if (V_ip6qb[bucket].count >= V_ip6_maxfragbucketsize ||
491 		    atomic_load_int(&V_frag6_nfragpackets) >=
492 		    (u_int)V_ip6_maxfragpackets)
493 			goto dropfrag;
494 		atomic_add_int(&V_frag6_nfragpackets, 1);
495 
496 		/* Allocate IPv6 fragement packet queue entry. */
497 		q6 = (struct ip6q *)malloc(sizeof(struct ip6q), M_FRAG6,
498 		    M_NOWAIT | M_ZERO);
499 		if (q6 == NULL)
500 			goto dropfrag;
501 #ifdef MAC
502 		if (mac_ip6q_init(q6, M_NOWAIT) != 0) {
503 			free(q6, M_FRAG6);
504 			goto dropfrag;
505 		}
506 		mac_ip6q_create(m, q6);
507 #endif
508 		frag6_insque_head(q6, head, bucket);
509 
510 		/* ip6q_nxt will be filled afterwards, from 1st fragment. */
511 		q6->ip6q_down	= q6->ip6q_up = (struct ip6asfrag *)q6;
512 #ifdef notyet
513 		q6->ip6q_nxtp	= (u_char *)nxtp;
514 #endif
515 		q6->ip6q_ident	= ip6f->ip6f_ident;
516 		q6->ip6q_ttl	= IPV6_FRAGTTL;
517 		q6->ip6q_src	= ip6->ip6_src;
518 		q6->ip6q_dst	= ip6->ip6_dst;
519 		q6->ip6q_ecn	=
520 		    (ntohl(ip6->ip6_flow) >> 20) & IPTOS_ECN_MASK;
521 		q6->ip6q_unfrglen = -1;	/* The 1st fragment has not arrived. */
522 
523 		q6->ip6q_nfrag = 0;
524 	}
525 
526 	/*
527 	 * If it is the 1st fragment, record the length of the
528 	 * unfragmentable part and the next header of the fragment header.
529 	 */
530 	fragoff = ntohs(ip6f->ip6f_offlg & IP6F_OFF_MASK);
531 	if (fragoff == 0) {
532 		q6->ip6q_unfrglen = offset - sizeof(struct ip6_hdr) -
533 		    sizeof(struct ip6_frag);
534 		q6->ip6q_nxt = ip6f->ip6f_nxt;
535 	}
536 
537 	/*
538 	 * Check that the reassembled packet would not exceed 65535 bytes
539 	 * in size.
540 	 * If it would exceed, discard the fragment and return an ICMP error.
541 	 */
542 	if (q6->ip6q_unfrglen >= 0) {
543 		/* The 1st fragment has already arrived. */
544 		if (q6->ip6q_unfrglen + fragoff + frgpartlen > IPV6_MAXPACKET) {
545 			icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
546 			    offset - sizeof(struct ip6_frag) +
547 			    offsetof(struct ip6_frag, ip6f_offlg));
548 			IP6QB_UNLOCK(bucket);
549 			return (IPPROTO_DONE);
550 		}
551 	} else if (fragoff + frgpartlen > IPV6_MAXPACKET) {
552 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
553 		    offset - sizeof(struct ip6_frag) +
554 		    offsetof(struct ip6_frag, ip6f_offlg));
555 		IP6QB_UNLOCK(bucket);
556 		return (IPPROTO_DONE);
557 	}
558 	/*
559 	 * If it is the first fragment, do the above check for each
560 	 * fragment already stored in the reassembly queue.
561 	 */
562 	if (fragoff == 0) {
563 		for (af6 = q6->ip6q_down; af6 != (struct ip6asfrag *)q6;
564 		     af6 = af6dwn) {
565 			af6dwn = af6->ip6af_down;
566 
567 			if (q6->ip6q_unfrglen + af6->ip6af_off + af6->ip6af_frglen >
568 			    IPV6_MAXPACKET) {
569 				struct ip6_hdr *ip6err;
570 				struct mbuf *merr;
571 				int erroff;
572 
573 				merr = IP6_REASS_MBUF(af6);
574 				erroff = af6->ip6af_offset;
575 
576 				/* Dequeue the fragment. */
577 				frag6_deq(af6, bucket);
578 				free(af6, M_FRAG6);
579 
580 				/* Set a valid receive interface pointer. */
581 				merr->m_pkthdr.rcvif = srcifp;
582 
583 				/* Adjust pointer. */
584 				ip6err = mtod(merr, struct ip6_hdr *);
585 
586 				/*
587 				 * Restore source and destination addresses
588 				 * in the erroneous IPv6 header.
589 				 */
590 				ip6err->ip6_src = q6->ip6q_src;
591 				ip6err->ip6_dst = q6->ip6q_dst;
592 
593 				icmp6_error(merr, ICMP6_PARAM_PROB,
594 				    ICMP6_PARAMPROB_HEADER,
595 				    erroff - sizeof(struct ip6_frag) +
596 				    offsetof(struct ip6_frag, ip6f_offlg));
597 			}
598 		}
599 	}
600 
601 	/* Allocate an IPv6 fragement queue entry for this fragmented part. */
602 	ip6af = (struct ip6asfrag *)malloc(sizeof(struct ip6asfrag), M_FRAG6,
603 	    M_NOWAIT | M_ZERO);
604 	if (ip6af == NULL)
605 		goto dropfrag;
606 	ip6af->ip6af_mff = ip6f->ip6f_offlg & IP6F_MORE_FRAG;
607 	ip6af->ip6af_off = fragoff;
608 	ip6af->ip6af_frglen = frgpartlen;
609 	ip6af->ip6af_offset = offset;
610 	IP6_REASS_MBUF(ip6af) = m;
611 
612 	if (only_frag) {
613 		af6 = (struct ip6asfrag *)q6;
614 		goto insert;
615 	}
616 
617 	/* Do duplicate, condition, and boundry checks. */
618 	/*
619 	 * Handle ECN by comparing this segment with the first one;
620 	 * if CE is set, do not lose CE.
621 	 * Drop if CE and not-ECT are mixed for the same packet.
622 	 */
623 	ecn = (ntohl(ip6->ip6_flow) >> 20) & IPTOS_ECN_MASK;
624 	ecn0 = q6->ip6q_ecn;
625 	if (ecn == IPTOS_ECN_CE) {
626 		if (ecn0 == IPTOS_ECN_NOTECT) {
627 			free(ip6af, M_FRAG6);
628 			goto dropfrag;
629 		}
630 		if (ecn0 != IPTOS_ECN_CE)
631 			q6->ip6q_ecn = IPTOS_ECN_CE;
632 	}
633 	if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT) {
634 		free(ip6af, M_FRAG6);
635 		goto dropfrag;
636 	}
637 
638 	/* Find a fragmented part which begins after this one does. */
639 	for (af6 = q6->ip6q_down; af6 != (struct ip6asfrag *)q6;
640 	     af6 = af6->ip6af_down)
641 		if (af6->ip6af_off > ip6af->ip6af_off)
642 			break;
643 
644 	/*
645 	 * If the incoming framgent overlaps some existing fragments in
646 	 * the reassembly queue, drop both the new fragment and the
647 	 * entire reassembly queue.  However, if the new fragment
648 	 * is an exact duplicate of an existing fragment, only silently
649 	 * drop the existing fragment and leave the fragmentation queue
650 	 * unchanged, as allowed by the RFC.  (RFC 8200, 4.5)
651 	 */
652 	if (af6->ip6af_up != (struct ip6asfrag *)q6) {
653 		if (af6->ip6af_up->ip6af_off + af6->ip6af_up->ip6af_frglen -
654 		    ip6af->ip6af_off > 0) {
655 			free(ip6af, M_FRAG6);
656 			goto dropfrag;
657 		}
658 	}
659 	if (af6 != (struct ip6asfrag *)q6) {
660 		if (ip6af->ip6af_off + ip6af->ip6af_frglen -
661 		    af6->ip6af_off > 0) {
662 			free(ip6af, M_FRAG6);
663 			goto dropfrag;
664 		}
665 	}
666 
667 insert:
668 #ifdef MAC
669 	if (!only_frag)
670 		mac_ip6q_update(m, q6);
671 #endif
672 
673 	/*
674 	 * Stick new segment in its place; check for complete reassembly.
675 	 * If not complete, check fragment limit.  Move to front of packet
676 	 * queue, as we are the most recently active fragmented packet.
677 	 */
678 	frag6_enq(ip6af, af6->ip6af_up, bucket);
679 	atomic_add_int(&frag6_nfrags, 1);
680 	q6->ip6q_nfrag++;
681 	plen = 0;
682 	for (af6 = q6->ip6q_down; af6 != (struct ip6asfrag *)q6;
683 	     af6 = af6->ip6af_down) {
684 		if (af6->ip6af_off != plen) {
685 			if (q6->ip6q_nfrag > V_ip6_maxfragsperpacket) {
686 				IP6STAT_ADD(ip6s_fragdropped, q6->ip6q_nfrag);
687 				frag6_freef(q6, bucket);
688 			}
689 			IP6QB_UNLOCK(bucket);
690 			return (IPPROTO_DONE);
691 		}
692 		plen += af6->ip6af_frglen;
693 	}
694 	if (af6->ip6af_up->ip6af_mff) {
695 		if (q6->ip6q_nfrag > V_ip6_maxfragsperpacket) {
696 			IP6STAT_ADD(ip6s_fragdropped, q6->ip6q_nfrag);
697 			frag6_freef(q6, bucket);
698 		}
699 		IP6QB_UNLOCK(bucket);
700 		return (IPPROTO_DONE);
701 	}
702 
703 	/* Reassembly is complete; concatenate fragments. */
704 	ip6af = q6->ip6q_down;
705 	t = m = IP6_REASS_MBUF(ip6af);
706 	af6 = ip6af->ip6af_down;
707 	frag6_deq(ip6af, bucket);
708 	while (af6 != (struct ip6asfrag *)q6) {
709 		m->m_pkthdr.csum_flags &=
710 		    IP6_REASS_MBUF(af6)->m_pkthdr.csum_flags;
711 		m->m_pkthdr.csum_data +=
712 		    IP6_REASS_MBUF(af6)->m_pkthdr.csum_data;
713 
714 		af6dwn = af6->ip6af_down;
715 		frag6_deq(af6, bucket);
716 		while (t->m_next)
717 			t = t->m_next;
718 		m_adj(IP6_REASS_MBUF(af6), af6->ip6af_offset);
719 		m_demote_pkthdr(IP6_REASS_MBUF(af6));
720 		m_cat(t, IP6_REASS_MBUF(af6));
721 		free(af6, M_FRAG6);
722 		af6 = af6dwn;
723 	}
724 
725 	while (m->m_pkthdr.csum_data & 0xffff0000)
726 		m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
727 		    (m->m_pkthdr.csum_data >> 16);
728 
729 	/* Adjust offset to point where the original next header starts. */
730 	offset = ip6af->ip6af_offset - sizeof(struct ip6_frag);
731 	free(ip6af, M_FRAG6);
732 	ip6 = mtod(m, struct ip6_hdr *);
733 	ip6->ip6_plen = htons((u_short)plen + offset - sizeof(struct ip6_hdr));
734 	if (q6->ip6q_ecn == IPTOS_ECN_CE)
735 		ip6->ip6_flow |= htonl(IPTOS_ECN_CE << 20);
736 	nxt = q6->ip6q_nxt;
737 
738 	if (ip6_deletefraghdr(m, offset, M_NOWAIT) != 0) {
739 		frag6_remque(q6, bucket);
740 		atomic_subtract_int(&frag6_nfrags, q6->ip6q_nfrag);
741 #ifdef MAC
742 		mac_ip6q_destroy(q6);
743 #endif
744 		free(q6, M_FRAG6);
745 		atomic_subtract_int(&V_frag6_nfragpackets, 1);
746 
747 		goto dropfrag;
748 	}
749 
750 	/* Set nxt(-hdr field value) to the original value. */
751 	m_copyback(m, ip6_get_prevhdr(m, offset), sizeof(uint8_t),
752 	    (caddr_t)&nxt);
753 
754 	frag6_remque(q6, bucket);
755 	atomic_subtract_int(&frag6_nfrags, q6->ip6q_nfrag);
756 #ifdef MAC
757 	mac_ip6q_reassemble(q6, m);
758 	mac_ip6q_destroy(q6);
759 #endif
760 	free(q6, M_FRAG6);
761 	atomic_subtract_int(&V_frag6_nfragpackets, 1);
762 
763 	if (m->m_flags & M_PKTHDR) { /* Isn't it always true? */
764 
765 		plen = 0;
766 		for (t = m; t; t = t->m_next)
767 			plen += t->m_len;
768 		m->m_pkthdr.len = plen;
769 		/* Set a valid receive interface pointer. */
770 		m->m_pkthdr.rcvif = srcifp;
771 	}
772 
773 #ifdef RSS
774 	mtag = m_tag_alloc(MTAG_ABI_IPV6, IPV6_TAG_DIRECT, sizeof(*ip6dc),
775 	    M_NOWAIT);
776 	if (mtag == NULL)
777 		goto dropfrag;
778 
779 	ip6dc = (struct ip6_direct_ctx *)(mtag + 1);
780 	ip6dc->ip6dc_nxt = nxt;
781 	ip6dc->ip6dc_off = offset;
782 
783 	m_tag_prepend(m, mtag);
784 #endif
785 
786 	IP6QB_UNLOCK(bucket);
787 	IP6STAT_INC(ip6s_reassembled);
788 	in6_ifstat_inc(dstifp, ifs6_reass_ok);
789 
790 #ifdef RSS
791 	/* Queue/dispatch for reprocessing. */
792 	netisr_dispatch(NETISR_IPV6_DIRECT, m);
793 	return (IPPROTO_DONE);
794 #endif
795 
796 	/* Tell launch routine the next header. */
797 	*mp = m;
798 	*offp = offset;
799 
800 	return (nxt);
801 
802 dropfrag:
803 	IP6QB_UNLOCK(bucket);
804 	in6_ifstat_inc(dstifp, ifs6_reass_fail);
805 	IP6STAT_INC(ip6s_fragdropped);
806 	m_freem(m);
807 	return (IPPROTO_DONE);
808 }
809 
810 /*
811  * IPv6 reassembling timer processing;
812  * if a timer expires on a reassembly queue, discard it.
813  */
814 void
815 frag6_slowtimo(void)
816 {
817 	VNET_ITERATOR_DECL(vnet_iter);
818 	struct ip6q *head, *q6;
819 	uint32_t bucket;
820 
821 	VNET_LIST_RLOCK_NOSLEEP();
822 	VNET_FOREACH(vnet_iter) {
823 		CURVNET_SET(vnet_iter);
824 		for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
825 			IP6QB_LOCK(bucket);
826 			head = IP6QB_HEAD(bucket);
827 			q6 = head->ip6q_next;
828 			if (q6 == NULL) {
829 				/*
830 				 * XXXJTL: This should never happen. This
831 				 * should turn into an assertion.
832 				 */
833 				IP6QB_UNLOCK(bucket);
834 				continue;
835 			}
836 			while (q6 != head) {
837 				--q6->ip6q_ttl;
838 				q6 = q6->ip6q_next;
839 				if (q6->ip6q_prev->ip6q_ttl == 0) {
840 					IP6STAT_ADD(ip6s_fragtimeout,
841 						q6->ip6q_prev->ip6q_nfrag);
842 					/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
843 					frag6_freef(q6->ip6q_prev, bucket);
844 				}
845 			}
846 			/*
847 			 * If we are over the maximum number of fragments
848 			 * (due to the limit being lowered), drain off
849 			 * enough to get down to the new limit.
850 			 * Note that we drain all reassembly queues if
851 			 * maxfragpackets is 0 (fragmentation is disabled),
852 			 * and do not enforce a limit when maxfragpackets
853 			 * is negative.
854 			 */
855 			while ((V_ip6_maxfragpackets == 0 ||
856 			    (V_ip6_maxfragpackets > 0 &&
857 			    V_ip6qb[bucket].count > V_ip6_maxfragbucketsize)) &&
858 			    head->ip6q_prev != head) {
859 				IP6STAT_ADD(ip6s_fragoverflow,
860 					q6->ip6q_prev->ip6q_nfrag);
861 				/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
862 				frag6_freef(head->ip6q_prev, bucket);
863 			}
864 			IP6QB_UNLOCK(bucket);
865 		}
866 		/*
867 		 * If we are still over the maximum number of fragmented
868 		 * packets, drain off enough to get down to the new limit.
869 		 */
870 		bucket = 0;
871 		while (V_ip6_maxfragpackets >= 0 &&
872 		    atomic_load_int(&V_frag6_nfragpackets) >
873 		    (u_int)V_ip6_maxfragpackets) {
874 			IP6QB_LOCK(bucket);
875 			head = IP6QB_HEAD(bucket);
876 			if (head->ip6q_prev != head) {
877 				IP6STAT_ADD(ip6s_fragoverflow,
878 					q6->ip6q_prev->ip6q_nfrag);
879 				/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
880 				frag6_freef(head->ip6q_prev, bucket);
881 			}
882 			IP6QB_UNLOCK(bucket);
883 			bucket = (bucket + 1) % IP6REASS_NHASH;
884 		}
885 		CURVNET_RESTORE();
886 	}
887 	VNET_LIST_RUNLOCK_NOSLEEP();
888 }
889 
890 /*
891  * Eventhandler to adjust limits in case nmbclusters change.
892  */
893 static void
894 frag6_change(void *tag)
895 {
896 	VNET_ITERATOR_DECL(vnet_iter);
897 
898 	ip6_maxfrags = IP6_MAXFRAGS;
899 	VNET_LIST_RLOCK_NOSLEEP();
900 	VNET_FOREACH(vnet_iter) {
901 		CURVNET_SET(vnet_iter);
902 		V_ip6_maxfragpackets = IP6_MAXFRAGPACKETS;
903 		frag6_set_bucketsize();
904 		CURVNET_RESTORE();
905 	}
906 	VNET_LIST_RUNLOCK_NOSLEEP();
907 }
908 
909 /*
910  * Initialise reassembly queue and fragment identifier.
911  */
912 void
913 frag6_init(void)
914 {
915 	struct ip6q *q6;
916 	uint32_t bucket;
917 
918 	V_ip6_maxfragpackets = IP6_MAXFRAGPACKETS;
919 	frag6_set_bucketsize();
920 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
921 		q6 = IP6QB_HEAD(bucket);
922 		q6->ip6q_next = q6->ip6q_prev = q6;
923 		mtx_init(&V_ip6qb[bucket].lock, "ip6qlock", NULL, MTX_DEF);
924 		V_ip6qb[bucket].count = 0;
925 	}
926 	V_ip6qb_hashseed = arc4random();
927 	V_ip6_maxfragsperpacket = 64;
928 	if (!IS_DEFAULT_VNET(curvnet))
929 		return;
930 
931 	ip6_maxfrags = IP6_MAXFRAGS;
932 	EVENTHANDLER_REGISTER(nmbclusters_change,
933 	    frag6_change, NULL, EVENTHANDLER_PRI_ANY);
934 }
935 
936 /*
937  * Drain off all datagram fragments.
938  */
939 void
940 frag6_drain(void)
941 {
942 	VNET_ITERATOR_DECL(vnet_iter);
943 	struct ip6q *head;
944 	uint32_t bucket;
945 
946 	VNET_LIST_RLOCK_NOSLEEP();
947 	VNET_FOREACH(vnet_iter) {
948 		CURVNET_SET(vnet_iter);
949 		for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
950 			if (IP6QB_TRYLOCK(bucket) == 0)
951 				continue;
952 			head = IP6QB_HEAD(bucket);
953 			while (head->ip6q_next != head) {
954 				IP6STAT_INC(ip6s_fragdropped);
955 				/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
956 				frag6_freef(head->ip6q_next, bucket);
957 			}
958 			IP6QB_UNLOCK(bucket);
959 		}
960 		CURVNET_RESTORE();
961 	}
962 	VNET_LIST_RUNLOCK_NOSLEEP();
963 }
964 
965 /*
966  * Put an ip fragment on a reassembly chain.
967  * Like insque, but pointers in middle of structure.
968  */
969 static void
970 frag6_enq(struct ip6asfrag *af6, struct ip6asfrag *up6,
971     uint32_t bucket __unused)
972 {
973 
974 	IP6QB_LOCK_ASSERT(bucket);
975 
976 	af6->ip6af_up = up6;
977 	af6->ip6af_down = up6->ip6af_down;
978 	up6->ip6af_down->ip6af_up = af6;
979 	up6->ip6af_down = af6;
980 }
981 
982 /*
983  * To frag6_enq as remque is to insque.
984  */
985 static void
986 frag6_deq(struct ip6asfrag *af6, uint32_t bucket __unused)
987 {
988 
989 	IP6QB_LOCK_ASSERT(bucket);
990 
991 	af6->ip6af_up->ip6af_down = af6->ip6af_down;
992 	af6->ip6af_down->ip6af_up = af6->ip6af_up;
993 }
994 
995 static void
996 frag6_insque_head(struct ip6q *new, struct ip6q *old, uint32_t bucket)
997 {
998 
999 	IP6QB_LOCK_ASSERT(bucket);
1000 	KASSERT(IP6QB_HEAD(bucket) == old,
1001 	    ("%s: attempt to insert at head of wrong bucket"
1002 	    " (bucket=%u, old=%p)", __func__, bucket, old));
1003 
1004 	new->ip6q_prev = old;
1005 	new->ip6q_next = old->ip6q_next;
1006 	old->ip6q_next->ip6q_prev= new;
1007 	old->ip6q_next = new;
1008 	V_ip6qb[bucket].count++;
1009 }
1010 
1011 static void
1012 frag6_remque(struct ip6q *p6, uint32_t bucket)
1013 {
1014 
1015 	IP6QB_LOCK_ASSERT(bucket);
1016 
1017 	p6->ip6q_prev->ip6q_next = p6->ip6q_next;
1018 	p6->ip6q_next->ip6q_prev = p6->ip6q_prev;
1019 	V_ip6qb[bucket].count--;
1020 }
1021