xref: /freebsd/sys/netinet6/frag6.c (revision d7d962ead0b6e5e8a39202d0590022082bf5bfb6)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
5  * All rights reserved.
6  * Copyright (c) 2019 Netflix, Inc.
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  * 3. Neither the name of the project nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	$KAME: frag6.c,v 1.33 2002/01/07 11:34:48 kjc Exp $
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "opt_rss.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #include <sys/eventhandler.h>
44 #include <sys/hash.h>
45 #include <sys/kernel.h>
46 #include <sys/malloc.h>
47 #include <sys/mbuf.h>
48 #include <sys/protosw.h>
49 #include <sys/queue.h>
50 #include <sys/socket.h>
51 #include <sys/sysctl.h>
52 #include <sys/syslog.h>
53 
54 #include <net/if.h>
55 #include <net/if_var.h>
56 #include <net/netisr.h>
57 #include <net/route.h>
58 #include <net/vnet.h>
59 
60 #include <netinet/in.h>
61 #include <netinet/in_var.h>
62 #include <netinet/ip6.h>
63 #include <netinet6/ip6_var.h>
64 #include <netinet/icmp6.h>
65 #include <netinet/in_systm.h>	/* For ECN definitions. */
66 #include <netinet/ip.h>		/* For ECN definitions. */
67 
68 #ifdef MAC
69 #include <security/mac/mac_framework.h>
70 #endif
71 
72 /*
73  * A "big picture" of how IPv6 fragment queues are all linked together.
74  *
75  * struct ip6qbucket ip6qb[...];			hashed buckets
76  * ||||||||
77  * |
78  * +--- TAILQ(struct ip6q, packets) *q6;		tailq entries holding
79  *      ||||||||					fragmented packets
80  *      |						(1 per original packet)
81  *      |
82  *      +--- TAILQ(struct ip6asfrag, ip6q_frags) *af6;	tailq entries of IPv6
83  *           |                                   *ip6af;fragment packets
84  *           |						for one original packet
85  *           + *mbuf
86  */
87 
88 /* Reassembly headers are stored in hash buckets. */
89 #define	IP6REASS_NHASH_LOG2	10
90 #define	IP6REASS_NHASH		(1 << IP6REASS_NHASH_LOG2)
91 #define	IP6REASS_HMASK		(IP6REASS_NHASH - 1)
92 
93 TAILQ_HEAD(ip6qhead, ip6q);
94 struct ip6qbucket {
95 	struct ip6qhead	packets;
96 	struct mtx	lock;
97 	int		count;
98 };
99 
100 struct ip6asfrag {
101 	TAILQ_ENTRY(ip6asfrag) ip6af_tq;
102 	struct mbuf	*ip6af_m;
103 	int		ip6af_offset;	/* Offset in ip6af_m to next header. */
104 	int		ip6af_frglen;	/* Fragmentable part length. */
105 	int		ip6af_off;	/* Fragment offset. */
106 	bool		ip6af_mff;	/* More fragment bit in frag off. */
107 };
108 
109 static MALLOC_DEFINE(M_FRAG6, "frag6", "IPv6 fragment reassembly header");
110 
111 #ifdef VIMAGE
112 /* A flag to indicate if IPv6 fragmentation is initialized. */
113 VNET_DEFINE_STATIC(bool,		frag6_on);
114 #define	V_frag6_on			VNET(frag6_on)
115 #endif
116 
117 /* System wide (global) maximum and count of packets in reassembly queues. */
118 static int ip6_maxfrags;
119 static u_int __exclusive_cache_line frag6_nfrags;
120 
121 /* Maximum and current packets in per-VNET reassembly queue. */
122 VNET_DEFINE_STATIC(int,			ip6_maxfragpackets);
123 VNET_DEFINE_STATIC(volatile u_int,	frag6_nfragpackets);
124 #define	V_ip6_maxfragpackets		VNET(ip6_maxfragpackets)
125 #define	V_frag6_nfragpackets		VNET(frag6_nfragpackets)
126 
127 /* Maximum per-VNET reassembly queues per bucket and fragments per packet. */
128 VNET_DEFINE_STATIC(int,			ip6_maxfragbucketsize);
129 VNET_DEFINE_STATIC(int,			ip6_maxfragsperpacket);
130 #define	V_ip6_maxfragbucketsize		VNET(ip6_maxfragbucketsize)
131 #define	V_ip6_maxfragsperpacket		VNET(ip6_maxfragsperpacket)
132 
133 /* Per-VNET reassembly queue buckets. */
134 VNET_DEFINE_STATIC(struct ip6qbucket,	ip6qb[IP6REASS_NHASH]);
135 VNET_DEFINE_STATIC(uint32_t,		ip6qb_hashseed);
136 #define	V_ip6qb				VNET(ip6qb)
137 #define	V_ip6qb_hashseed		VNET(ip6qb_hashseed)
138 
139 #define	IP6QB_LOCK(_b)		mtx_lock(&V_ip6qb[(_b)].lock)
140 #define	IP6QB_TRYLOCK(_b)	mtx_trylock(&V_ip6qb[(_b)].lock)
141 #define	IP6QB_LOCK_ASSERT(_b)	mtx_assert(&V_ip6qb[(_b)].lock, MA_OWNED)
142 #define	IP6QB_UNLOCK(_b)	mtx_unlock(&V_ip6qb[(_b)].lock)
143 #define	IP6QB_HEAD(_b)		(&V_ip6qb[(_b)].packets)
144 
145 /*
146  * By default, limit the number of IP6 fragments across all reassembly
147  * queues to  1/32 of the total number of mbuf clusters.
148  *
149  * Limit the total number of reassembly queues per VNET to the
150  * IP6 fragment limit, but ensure the limit will not allow any bucket
151  * to grow above 100 items. (The bucket limit is
152  * IP_MAXFRAGPACKETS / (IPREASS_NHASH / 2), so the 50 is the correct
153  * multiplier to reach a 100-item limit.)
154  * The 100-item limit was chosen as brief testing seems to show that
155  * this produces "reasonable" performance on some subset of systems
156  * under DoS attack.
157  */
158 #define	IP6_MAXFRAGS		(nmbclusters / 32)
159 #define	IP6_MAXFRAGPACKETS	(imin(IP6_MAXFRAGS, IP6REASS_NHASH * 50))
160 
161 /*
162  * Sysctls and helper function.
163  */
164 SYSCTL_DECL(_net_inet6_ip6);
165 
166 SYSCTL_UINT(_net_inet6_ip6, OID_AUTO, frag6_nfrags,
167 	CTLFLAG_RD, &frag6_nfrags, 0,
168 	"Global number of IPv6 fragments across all reassembly queues.");
169 
170 static void
171 frag6_set_bucketsize(void)
172 {
173 	int i;
174 
175 	if ((i = V_ip6_maxfragpackets) > 0)
176 		V_ip6_maxfragbucketsize = imax(i / (IP6REASS_NHASH / 2), 1);
177 }
178 
179 SYSCTL_INT(_net_inet6_ip6, IPV6CTL_MAXFRAGS, maxfrags,
180 	CTLFLAG_RW, &ip6_maxfrags, 0,
181 	"Maximum allowed number of outstanding IPv6 packet fragments. "
182 	"A value of 0 means no fragmented packets will be accepted, while a "
183 	"a value of -1 means no limit");
184 
185 static int
186 sysctl_ip6_maxfragpackets(SYSCTL_HANDLER_ARGS)
187 {
188 	int error, val;
189 
190 	val = V_ip6_maxfragpackets;
191 	error = sysctl_handle_int(oidp, &val, 0, req);
192 	if (error != 0 || !req->newptr)
193 		return (error);
194 	V_ip6_maxfragpackets = val;
195 	frag6_set_bucketsize();
196 	return (0);
197 }
198 SYSCTL_PROC(_net_inet6_ip6, IPV6CTL_MAXFRAGPACKETS, maxfragpackets,
199 	CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
200 	NULL, 0, sysctl_ip6_maxfragpackets, "I",
201 	"Default maximum number of outstanding fragmented IPv6 packets. "
202 	"A value of 0 means no fragmented packets will be accepted, while a "
203 	"a value of -1 means no limit");
204 SYSCTL_UINT(_net_inet6_ip6, OID_AUTO, frag6_nfragpackets,
205 	CTLFLAG_VNET | CTLFLAG_RD,
206 	__DEVOLATILE(u_int *, &VNET_NAME(frag6_nfragpackets)), 0,
207 	"Per-VNET number of IPv6 fragments across all reassembly queues.");
208 SYSCTL_INT(_net_inet6_ip6, IPV6CTL_MAXFRAGSPERPACKET, maxfragsperpacket,
209 	CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ip6_maxfragsperpacket), 0,
210 	"Maximum allowed number of fragments per packet");
211 SYSCTL_INT(_net_inet6_ip6, IPV6CTL_MAXFRAGBUCKETSIZE, maxfragbucketsize,
212 	CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ip6_maxfragbucketsize), 0,
213 	"Maximum number of reassembly queues per hash bucket");
214 
215 /*
216  * Remove the IPv6 fragmentation header from the mbuf.
217  */
218 int
219 ip6_deletefraghdr(struct mbuf *m, int offset, int wait __unused)
220 {
221 	struct ip6_hdr *ip6;
222 
223 	KASSERT(m->m_len >= offset + sizeof(struct ip6_frag),
224 	    ("%s: ext headers not contigous in mbuf %p m_len %d >= "
225 	    "offset %d + %zu\n", __func__, m, m->m_len, offset,
226 	    sizeof(struct ip6_frag)));
227 
228 	/* Delete frag6 header. */
229 	ip6 = mtod(m, struct ip6_hdr *);
230 	bcopy(ip6, (char *)ip6 + sizeof(struct ip6_frag), offset);
231 	m->m_data += sizeof(struct ip6_frag);
232 	m->m_len -= sizeof(struct ip6_frag);
233 	m->m_flags |= M_FRAGMENTED;
234 
235 	return (0);
236 }
237 
238 /*
239  * Free a fragment reassembly header and all associated datagrams.
240  */
241 static void
242 frag6_freef(struct ip6q *q6, uint32_t bucket)
243 {
244 	struct ip6_hdr *ip6;
245 	struct ip6asfrag *af6;
246 	struct mbuf *m;
247 
248 	IP6QB_LOCK_ASSERT(bucket);
249 
250 	while ((af6 = TAILQ_FIRST(&q6->ip6q_frags)) != NULL) {
251 		m = af6->ip6af_m;
252 		TAILQ_REMOVE(&q6->ip6q_frags, af6, ip6af_tq);
253 
254 		/*
255 		 * Return ICMP time exceeded error for the 1st fragment.
256 		 * Just free other fragments.
257 		 */
258 		if (af6->ip6af_off == 0 && m->m_pkthdr.rcvif != NULL) {
259 			/* Adjust pointer. */
260 			ip6 = mtod(m, struct ip6_hdr *);
261 
262 			/* Restore source and destination addresses. */
263 			ip6->ip6_src = q6->ip6q_src;
264 			ip6->ip6_dst = q6->ip6q_dst;
265 
266 			icmp6_error(m, ICMP6_TIME_EXCEEDED,
267 			    ICMP6_TIME_EXCEED_REASSEMBLY, 0);
268 		} else
269 			m_freem(m);
270 
271 		free(af6, M_FRAG6);
272 	}
273 
274 	TAILQ_REMOVE(IP6QB_HEAD(bucket), q6, ip6q_tq);
275 	V_ip6qb[bucket].count--;
276 	atomic_subtract_int(&frag6_nfrags, q6->ip6q_nfrag);
277 #ifdef MAC
278 	mac_ip6q_destroy(q6);
279 #endif
280 	free(q6, M_FRAG6);
281 	atomic_subtract_int(&V_frag6_nfragpackets, 1);
282 }
283 
284 /*
285  * Drain off all datagram fragments belonging to
286  * the given network interface.
287  */
288 static void
289 frag6_cleanup(void *arg __unused, struct ifnet *ifp)
290 {
291 	struct ip6qhead *head;
292 	struct ip6q *q6;
293 	struct ip6asfrag *af6;
294 	uint32_t bucket;
295 
296 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
297 
298 	CURVNET_SET_QUIET(ifp->if_vnet);
299 #ifdef VIMAGE
300 	/*
301 	 * Skip processing if IPv6 reassembly is not initialised or
302 	 * torn down by frag6_destroy().
303 	 */
304 	if (!V_frag6_on) {
305 		CURVNET_RESTORE();
306 		return;
307 	}
308 #endif
309 
310 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
311 		IP6QB_LOCK(bucket);
312 		head = IP6QB_HEAD(bucket);
313 		/* Scan fragment list. */
314 		TAILQ_FOREACH(q6, head, ip6q_tq) {
315 			TAILQ_FOREACH(af6, &q6->ip6q_frags, ip6af_tq) {
316 				/* Clear no longer valid rcvif pointer. */
317 				if (af6->ip6af_m->m_pkthdr.rcvif == ifp)
318 					af6->ip6af_m->m_pkthdr.rcvif = NULL;
319 			}
320 		}
321 		IP6QB_UNLOCK(bucket);
322 	}
323 	CURVNET_RESTORE();
324 }
325 EVENTHANDLER_DEFINE(ifnet_departure_event, frag6_cleanup, NULL, 0);
326 
327 /*
328  * Like in RFC2460, in RFC8200, fragment and reassembly rules do not agree with
329  * each other, in terms of next header field handling in fragment header.
330  * While the sender will use the same value for all of the fragmented packets,
331  * receiver is suggested not to check for consistency.
332  *
333  * Fragment rules (p18,p19):
334  *	(2)  A Fragment header containing:
335  *	The Next Header value that identifies the first header
336  *	after the Per-Fragment headers of the original packet.
337  *		-> next header field is same for all fragments
338  *
339  * Reassembly rule (p20):
340  *	The Next Header field of the last header of the Per-Fragment
341  *	headers is obtained from the Next Header field of the first
342  *	fragment's Fragment header.
343  *		-> should grab it from the first fragment only
344  *
345  * The following note also contradicts with fragment rule - no one is going to
346  * send different fragment with different next header field.
347  *
348  * Additional note (p22) [not an error]:
349  *	The Next Header values in the Fragment headers of different
350  *	fragments of the same original packet may differ.  Only the value
351  *	from the Offset zero fragment packet is used for reassembly.
352  *		-> should grab it from the first fragment only
353  *
354  * There is no explicit reason given in the RFC.  Historical reason maybe?
355  */
356 /*
357  * Fragment input.
358  */
359 int
360 frag6_input(struct mbuf **mp, int *offp, int proto)
361 {
362 	struct mbuf *m, *t;
363 	struct ip6_hdr *ip6;
364 	struct ip6_frag *ip6f;
365 	struct ip6qhead *head;
366 	struct ip6q *q6;
367 	struct ip6asfrag *af6, *ip6af, *af6tmp;
368 	struct in6_ifaddr *ia6;
369 	struct ifnet *dstifp, *srcifp;
370 	uint32_t hashkey[(sizeof(struct in6_addr) * 2 +
371 		    sizeof(ip6f->ip6f_ident)) / sizeof(uint32_t)];
372 	uint32_t bucket, *hashkeyp;
373 	int fragoff, frgpartlen;	/* Must be larger than uint16_t. */
374 	int nxt, offset, plen;
375 	uint8_t ecn, ecn0;
376 	bool only_frag;
377 #ifdef RSS
378 	struct ip6_direct_ctx *ip6dc;
379 	struct m_tag *mtag;
380 #endif
381 
382 	m = *mp;
383 	offset = *offp;
384 
385 	M_ASSERTPKTHDR(m);
386 
387 	if (m->m_len < offset + sizeof(struct ip6_frag)) {
388 		m = m_pullup(m, offset + sizeof(struct ip6_frag));
389 		if (m == NULL) {
390 			IP6STAT_INC(ip6s_exthdrtoolong);
391 			*mp = NULL;
392 			return (IPPROTO_DONE);
393 		}
394 	}
395 	ip6 = mtod(m, struct ip6_hdr *);
396 
397 	dstifp = NULL;
398 	/* Find the destination interface of the packet. */
399 	ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */, false);
400 	if (ia6 != NULL)
401 		dstifp = ia6->ia_ifp;
402 
403 	/* Jumbo payload cannot contain a fragment header. */
404 	if (ip6->ip6_plen == 0) {
405 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER, offset);
406 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
407 		*mp = NULL;
408 		return (IPPROTO_DONE);
409 	}
410 
411 	/*
412 	 * Check whether fragment packet's fragment length is a
413 	 * multiple of 8 octets (unless it is the last one).
414 	 * sizeof(struct ip6_frag) == 8
415 	 * sizeof(struct ip6_hdr) = 40
416 	 */
417 	ip6f = (struct ip6_frag *)((caddr_t)ip6 + offset);
418 	if ((ip6f->ip6f_offlg & IP6F_MORE_FRAG) &&
419 	    (((ntohs(ip6->ip6_plen) - offset) & 0x7) != 0)) {
420 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
421 		    offsetof(struct ip6_hdr, ip6_plen));
422 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
423 		*mp = NULL;
424 		return (IPPROTO_DONE);
425 	}
426 
427 	IP6STAT_INC(ip6s_fragments);
428 	in6_ifstat_inc(dstifp, ifs6_reass_reqd);
429 
430 	/*
431 	 * Handle "atomic" fragments (offset and m bit set to 0) upfront,
432 	 * unrelated to any reassembly.  We need to remove the frag hdr
433 	 * which is ugly.
434 	 * See RFC 6946 and section 4.5 of RFC 8200.
435 	 */
436 	if ((ip6f->ip6f_offlg & ~IP6F_RESERVED_MASK) == 0) {
437 		IP6STAT_INC(ip6s_atomicfrags);
438 		nxt = ip6f->ip6f_nxt;
439 		/*
440 		 * Set nxt(-hdr field value) to the original value.
441 		 * We cannot just set ip6->ip6_nxt as there might be
442 		 * an unfragmentable part with extension headers and
443 		 * we must update the last one.
444 		 */
445 		m_copyback(m, ip6_get_prevhdr(m, offset), sizeof(uint8_t),
446 		    (caddr_t)&nxt);
447 		ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) -
448 		    sizeof(struct ip6_frag));
449 		if (ip6_deletefraghdr(m, offset, M_NOWAIT) != 0)
450 			goto dropfrag2;
451 		m->m_pkthdr.len -= sizeof(struct ip6_frag);
452 		in6_ifstat_inc(dstifp, ifs6_reass_ok);
453 		*mp = m;
454 		return (nxt);
455 	}
456 
457 	/* Offset now points to data portion. */
458 	offset += sizeof(struct ip6_frag);
459 
460 	/* Get fragment length and discard 0-byte fragments. */
461 	frgpartlen = sizeof(struct ip6_hdr) + ntohs(ip6->ip6_plen) - offset;
462 	if (frgpartlen == 0) {
463 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
464 		    offsetof(struct ip6_hdr, ip6_plen));
465 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
466 		IP6STAT_INC(ip6s_fragdropped);
467 		*mp = NULL;
468 		return (IPPROTO_DONE);
469 	}
470 
471 	/*
472 	 * Enforce upper bound on number of fragments for the entire system.
473 	 * If maxfrag is 0, never accept fragments.
474 	 * If maxfrag is -1, accept all fragments without limitation.
475 	 */
476 	if (ip6_maxfrags < 0)
477 		;
478 	else if (atomic_load_int(&frag6_nfrags) >= (u_int)ip6_maxfrags)
479 		goto dropfrag2;
480 
481 	/*
482 	 * Validate that a full header chain to the ULP is present in the
483 	 * packet containing the first fragment as per RFC RFC7112 and
484 	 * RFC 8200 pages 18,19:
485 	 * The first fragment packet is composed of:
486 	 * (3)  Extension headers, if any, and the Upper-Layer header.  These
487 	 *      headers must be in the first fragment.  ...
488 	 */
489 	fragoff = ntohs(ip6f->ip6f_offlg & IP6F_OFF_MASK);
490 	/* XXX TODO.  thj has D16851 open for this. */
491 	/* Send ICMPv6 4,3 in case of violation. */
492 
493 	/* Store receive network interface pointer for later. */
494 	srcifp = m->m_pkthdr.rcvif;
495 
496 	/* Generate a hash value for fragment bucket selection. */
497 	hashkeyp = hashkey;
498 	memcpy(hashkeyp, &ip6->ip6_src, sizeof(struct in6_addr));
499 	hashkeyp += sizeof(struct in6_addr) / sizeof(*hashkeyp);
500 	memcpy(hashkeyp, &ip6->ip6_dst, sizeof(struct in6_addr));
501 	hashkeyp += sizeof(struct in6_addr) / sizeof(*hashkeyp);
502 	*hashkeyp = ip6f->ip6f_ident;
503 	bucket = jenkins_hash32(hashkey, nitems(hashkey), V_ip6qb_hashseed);
504 	bucket &= IP6REASS_HMASK;
505 	IP6QB_LOCK(bucket);
506 	head = IP6QB_HEAD(bucket);
507 
508 	TAILQ_FOREACH(q6, head, ip6q_tq)
509 		if (ip6f->ip6f_ident == q6->ip6q_ident &&
510 		    IN6_ARE_ADDR_EQUAL(&ip6->ip6_src, &q6->ip6q_src) &&
511 		    IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst, &q6->ip6q_dst)
512 #ifdef MAC
513 		    && mac_ip6q_match(m, q6)
514 #endif
515 		    )
516 			break;
517 
518 	only_frag = false;
519 	if (q6 == NULL) {
520 		/* A first fragment to arrive creates a reassembly queue. */
521 		only_frag = true;
522 
523 		/*
524 		 * Enforce upper bound on number of fragmented packets
525 		 * for which we attempt reassembly;
526 		 * If maxfragpackets is 0, never accept fragments.
527 		 * If maxfragpackets is -1, accept all fragments without
528 		 * limitation.
529 		 */
530 		if (V_ip6_maxfragpackets < 0)
531 			;
532 		else if (V_ip6qb[bucket].count >= V_ip6_maxfragbucketsize ||
533 		    atomic_load_int(&V_frag6_nfragpackets) >=
534 		    (u_int)V_ip6_maxfragpackets)
535 			goto dropfrag;
536 
537 		/* Allocate IPv6 fragement packet queue entry. */
538 		q6 = (struct ip6q *)malloc(sizeof(struct ip6q), M_FRAG6,
539 		    M_NOWAIT | M_ZERO);
540 		if (q6 == NULL)
541 			goto dropfrag;
542 #ifdef MAC
543 		if (mac_ip6q_init(q6, M_NOWAIT) != 0) {
544 			free(q6, M_FRAG6);
545 			goto dropfrag;
546 		}
547 		mac_ip6q_create(m, q6);
548 #endif
549 		atomic_add_int(&V_frag6_nfragpackets, 1);
550 
551 		/* ip6q_nxt will be filled afterwards, from 1st fragment. */
552 		TAILQ_INIT(&q6->ip6q_frags);
553 		q6->ip6q_ident	= ip6f->ip6f_ident;
554 		q6->ip6q_ttl	= IPV6_FRAGTTL;
555 		q6->ip6q_src	= ip6->ip6_src;
556 		q6->ip6q_dst	= ip6->ip6_dst;
557 		q6->ip6q_ecn	= IPV6_ECN(ip6);
558 		q6->ip6q_unfrglen = -1;	/* The 1st fragment has not arrived. */
559 
560 		/* Add the fragemented packet to the bucket. */
561 		TAILQ_INSERT_HEAD(head, q6, ip6q_tq);
562 		V_ip6qb[bucket].count++;
563 	}
564 
565 	/*
566 	 * If it is the 1st fragment, record the length of the
567 	 * unfragmentable part and the next header of the fragment header.
568 	 * Assume the first 1st fragement to arrive will be correct.
569 	 * We do not have any duplicate checks here yet so another packet
570 	 * with fragoff == 0 could come and overwrite the ip6q_unfrglen
571 	 * and worse, the next header, at any time.
572 	 */
573 	if (fragoff == 0 && q6->ip6q_unfrglen == -1) {
574 		q6->ip6q_unfrglen = offset - sizeof(struct ip6_hdr) -
575 		    sizeof(struct ip6_frag);
576 		q6->ip6q_nxt = ip6f->ip6f_nxt;
577 		/* XXX ECN? */
578 	}
579 
580 	/*
581 	 * Check that the reassembled packet would not exceed 65535 bytes
582 	 * in size.
583 	 * If it would exceed, discard the fragment and return an ICMP error.
584 	 */
585 	if (q6->ip6q_unfrglen >= 0) {
586 		/* The 1st fragment has already arrived. */
587 		if (q6->ip6q_unfrglen + fragoff + frgpartlen > IPV6_MAXPACKET) {
588 			if (only_frag) {
589 				TAILQ_REMOVE(head, q6, ip6q_tq);
590 				V_ip6qb[bucket].count--;
591 				atomic_subtract_int(&V_frag6_nfragpackets, 1);
592 #ifdef MAC
593 				mac_ip6q_destroy(q6);
594 #endif
595 				free(q6, M_FRAG6);
596 			}
597 			IP6QB_UNLOCK(bucket);
598 			icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
599 			    offset - sizeof(struct ip6_frag) +
600 			    offsetof(struct ip6_frag, ip6f_offlg));
601 			*mp = NULL;
602 			return (IPPROTO_DONE);
603 		}
604 	} else if (fragoff + frgpartlen > IPV6_MAXPACKET) {
605 		if (only_frag) {
606 			TAILQ_REMOVE(head, q6, ip6q_tq);
607 			V_ip6qb[bucket].count--;
608 			atomic_subtract_int(&V_frag6_nfragpackets, 1);
609 #ifdef MAC
610 			mac_ip6q_destroy(q6);
611 #endif
612 			free(q6, M_FRAG6);
613 		}
614 		IP6QB_UNLOCK(bucket);
615 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
616 		    offset - sizeof(struct ip6_frag) +
617 		    offsetof(struct ip6_frag, ip6f_offlg));
618 		*mp = NULL;
619 		return (IPPROTO_DONE);
620 	}
621 
622 	/*
623 	 * If it is the first fragment, do the above check for each
624 	 * fragment already stored in the reassembly queue.
625 	 */
626 	if (fragoff == 0 && !only_frag) {
627 		TAILQ_FOREACH_SAFE(af6, &q6->ip6q_frags, ip6af_tq, af6tmp) {
628 			if (q6->ip6q_unfrglen + af6->ip6af_off +
629 			    af6->ip6af_frglen > IPV6_MAXPACKET) {
630 				struct ip6_hdr *ip6err;
631 				struct mbuf *merr;
632 				int erroff;
633 
634 				merr = af6->ip6af_m;
635 				erroff = af6->ip6af_offset;
636 
637 				/* Dequeue the fragment. */
638 				TAILQ_REMOVE(&q6->ip6q_frags, af6, ip6af_tq);
639 				q6->ip6q_nfrag--;
640 				atomic_subtract_int(&frag6_nfrags, 1);
641 				free(af6, M_FRAG6);
642 
643 				/* Set a valid receive interface pointer. */
644 				merr->m_pkthdr.rcvif = srcifp;
645 
646 				/* Adjust pointer. */
647 				ip6err = mtod(merr, struct ip6_hdr *);
648 
649 				/*
650 				 * Restore source and destination addresses
651 				 * in the erroneous IPv6 header.
652 				 */
653 				ip6err->ip6_src = q6->ip6q_src;
654 				ip6err->ip6_dst = q6->ip6q_dst;
655 
656 				icmp6_error(merr, ICMP6_PARAM_PROB,
657 				    ICMP6_PARAMPROB_HEADER,
658 				    erroff - sizeof(struct ip6_frag) +
659 				    offsetof(struct ip6_frag, ip6f_offlg));
660 			}
661 		}
662 	}
663 
664 	/* Allocate an IPv6 fragement queue entry for this fragmented part. */
665 	ip6af = (struct ip6asfrag *)malloc(sizeof(struct ip6asfrag), M_FRAG6,
666 	    M_NOWAIT | M_ZERO);
667 	if (ip6af == NULL)
668 		goto dropfrag;
669 	ip6af->ip6af_mff = (ip6f->ip6f_offlg & IP6F_MORE_FRAG) ? true : false;
670 	ip6af->ip6af_off = fragoff;
671 	ip6af->ip6af_frglen = frgpartlen;
672 	ip6af->ip6af_offset = offset;
673 	ip6af->ip6af_m = m;
674 
675 	if (only_frag) {
676 		/*
677 		 * Do a manual insert rather than a hard-to-understand cast
678 		 * to a different type relying on data structure order to work.
679 		 */
680 		TAILQ_INSERT_HEAD(&q6->ip6q_frags, ip6af, ip6af_tq);
681 		goto postinsert;
682 	}
683 
684 	/* Do duplicate, condition, and boundry checks. */
685 	/*
686 	 * Handle ECN by comparing this segment with the first one;
687 	 * if CE is set, do not lose CE.
688 	 * Drop if CE and not-ECT are mixed for the same packet.
689 	 */
690 	ecn = IPV6_ECN(ip6);
691 	ecn0 = q6->ip6q_ecn;
692 	if (ecn == IPTOS_ECN_CE) {
693 		if (ecn0 == IPTOS_ECN_NOTECT) {
694 			free(ip6af, M_FRAG6);
695 			goto dropfrag;
696 		}
697 		if (ecn0 != IPTOS_ECN_CE)
698 			q6->ip6q_ecn = IPTOS_ECN_CE;
699 	}
700 	if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT) {
701 		free(ip6af, M_FRAG6);
702 		goto dropfrag;
703 	}
704 
705 	/* Find a fragmented part which begins after this one does. */
706 	TAILQ_FOREACH(af6, &q6->ip6q_frags, ip6af_tq)
707 		if (af6->ip6af_off > ip6af->ip6af_off)
708 			break;
709 
710 	/*
711 	 * If the incoming framgent overlaps some existing fragments in
712 	 * the reassembly queue, drop both the new fragment and the
713 	 * entire reassembly queue.  However, if the new fragment
714 	 * is an exact duplicate of an existing fragment, only silently
715 	 * drop the existing fragment and leave the fragmentation queue
716 	 * unchanged, as allowed by the RFC.  (RFC 8200, 4.5)
717 	 */
718 	if (af6 != NULL)
719 		af6tmp = TAILQ_PREV(af6, ip6fraghead, ip6af_tq);
720 	else
721 		af6tmp = TAILQ_LAST(&q6->ip6q_frags, ip6fraghead);
722 	if (af6tmp != NULL) {
723 		if (af6tmp->ip6af_off + af6tmp->ip6af_frglen -
724 		    ip6af->ip6af_off > 0) {
725 			if (af6tmp->ip6af_off != ip6af->ip6af_off ||
726 			    af6tmp->ip6af_frglen != ip6af->ip6af_frglen)
727 				frag6_freef(q6, bucket);
728 			free(ip6af, M_FRAG6);
729 			goto dropfrag;
730 		}
731 	}
732 	if (af6 != NULL) {
733 		if (ip6af->ip6af_off + ip6af->ip6af_frglen -
734 		    af6->ip6af_off > 0) {
735 			if (af6->ip6af_off != ip6af->ip6af_off ||
736 			    af6->ip6af_frglen != ip6af->ip6af_frglen)
737 				frag6_freef(q6, bucket);
738 			free(ip6af, M_FRAG6);
739 			goto dropfrag;
740 		}
741 	}
742 
743 #ifdef MAC
744 	mac_ip6q_update(m, q6);
745 #endif
746 
747 	/*
748 	 * Stick new segment in its place; check for complete reassembly.
749 	 * If not complete, check fragment limit.  Move to front of packet
750 	 * queue, as we are the most recently active fragmented packet.
751 	 */
752 	if (af6 != NULL)
753 		TAILQ_INSERT_BEFORE(af6, ip6af, ip6af_tq);
754 	else
755 		TAILQ_INSERT_TAIL(&q6->ip6q_frags, ip6af, ip6af_tq);
756 postinsert:
757 	atomic_add_int(&frag6_nfrags, 1);
758 	q6->ip6q_nfrag++;
759 
760 	plen = 0;
761 	TAILQ_FOREACH(af6, &q6->ip6q_frags, ip6af_tq) {
762 		if (af6->ip6af_off != plen) {
763 			if (q6->ip6q_nfrag > V_ip6_maxfragsperpacket) {
764 				IP6STAT_ADD(ip6s_fragdropped, q6->ip6q_nfrag);
765 				frag6_freef(q6, bucket);
766 			}
767 			IP6QB_UNLOCK(bucket);
768 			*mp = NULL;
769 			return (IPPROTO_DONE);
770 		}
771 		plen += af6->ip6af_frglen;
772 	}
773 	af6 = TAILQ_LAST(&q6->ip6q_frags, ip6fraghead);
774 	if (af6->ip6af_mff) {
775 		if (q6->ip6q_nfrag > V_ip6_maxfragsperpacket) {
776 			IP6STAT_ADD(ip6s_fragdropped, q6->ip6q_nfrag);
777 			frag6_freef(q6, bucket);
778 		}
779 		IP6QB_UNLOCK(bucket);
780 		*mp = NULL;
781 		return (IPPROTO_DONE);
782 	}
783 
784 	/* Reassembly is complete; concatenate fragments. */
785 	ip6af = TAILQ_FIRST(&q6->ip6q_frags);
786 	t = m = ip6af->ip6af_m;
787 	TAILQ_REMOVE(&q6->ip6q_frags, ip6af, ip6af_tq);
788 	while ((af6 = TAILQ_FIRST(&q6->ip6q_frags)) != NULL) {
789 		m->m_pkthdr.csum_flags &=
790 		    af6->ip6af_m->m_pkthdr.csum_flags;
791 		m->m_pkthdr.csum_data +=
792 		    af6->ip6af_m->m_pkthdr.csum_data;
793 
794 		TAILQ_REMOVE(&q6->ip6q_frags, af6, ip6af_tq);
795 		t = m_last(t);
796 		m_adj(af6->ip6af_m, af6->ip6af_offset);
797 		m_demote_pkthdr(af6->ip6af_m);
798 		m_cat(t, af6->ip6af_m);
799 		free(af6, M_FRAG6);
800 	}
801 
802 	while (m->m_pkthdr.csum_data & 0xffff0000)
803 		m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
804 		    (m->m_pkthdr.csum_data >> 16);
805 
806 	/* Adjust offset to point where the original next header starts. */
807 	offset = ip6af->ip6af_offset - sizeof(struct ip6_frag);
808 	free(ip6af, M_FRAG6);
809 	ip6 = mtod(m, struct ip6_hdr *);
810 	ip6->ip6_plen = htons((u_short)plen + offset - sizeof(struct ip6_hdr));
811 	if (q6->ip6q_ecn == IPTOS_ECN_CE)
812 		ip6->ip6_flow |= htonl(IPTOS_ECN_CE << 20);
813 	nxt = q6->ip6q_nxt;
814 
815 	TAILQ_REMOVE(head, q6, ip6q_tq);
816 	V_ip6qb[bucket].count--;
817 	atomic_subtract_int(&frag6_nfrags, q6->ip6q_nfrag);
818 
819 	ip6_deletefraghdr(m, offset, M_NOWAIT);
820 
821 	/* Set nxt(-hdr field value) to the original value. */
822 	m_copyback(m, ip6_get_prevhdr(m, offset), sizeof(uint8_t),
823 	    (caddr_t)&nxt);
824 
825 #ifdef MAC
826 	mac_ip6q_reassemble(q6, m);
827 	mac_ip6q_destroy(q6);
828 #endif
829 	free(q6, M_FRAG6);
830 	atomic_subtract_int(&V_frag6_nfragpackets, 1);
831 
832 	if (m->m_flags & M_PKTHDR) { /* Isn't it always true? */
833 
834 		plen = 0;
835 		for (t = m; t; t = t->m_next)
836 			plen += t->m_len;
837 		m->m_pkthdr.len = plen;
838 		/* Set a valid receive interface pointer. */
839 		m->m_pkthdr.rcvif = srcifp;
840 	}
841 
842 #ifdef RSS
843 	mtag = m_tag_alloc(MTAG_ABI_IPV6, IPV6_TAG_DIRECT, sizeof(*ip6dc),
844 	    M_NOWAIT);
845 	if (mtag == NULL)
846 		goto dropfrag;
847 
848 	ip6dc = (struct ip6_direct_ctx *)(mtag + 1);
849 	ip6dc->ip6dc_nxt = nxt;
850 	ip6dc->ip6dc_off = offset;
851 
852 	m_tag_prepend(m, mtag);
853 #endif
854 
855 	IP6QB_UNLOCK(bucket);
856 	IP6STAT_INC(ip6s_reassembled);
857 	in6_ifstat_inc(dstifp, ifs6_reass_ok);
858 
859 #ifdef RSS
860 	/* Queue/dispatch for reprocessing. */
861 	netisr_dispatch(NETISR_IPV6_DIRECT, m);
862 	*mp = NULL;
863 	return (IPPROTO_DONE);
864 #endif
865 
866 	/* Tell launch routine the next header. */
867 	*mp = m;
868 	*offp = offset;
869 
870 	return (nxt);
871 
872 dropfrag:
873 	IP6QB_UNLOCK(bucket);
874 dropfrag2:
875 	in6_ifstat_inc(dstifp, ifs6_reass_fail);
876 	IP6STAT_INC(ip6s_fragdropped);
877 	m_freem(m);
878 	*mp = NULL;
879 	return (IPPROTO_DONE);
880 }
881 
882 /*
883  * IPv6 reassembling timer processing;
884  * if a timer expires on a reassembly queue, discard it.
885  */
886 void
887 frag6_slowtimo(void)
888 {
889 	VNET_ITERATOR_DECL(vnet_iter);
890 	struct ip6qhead *head;
891 	struct ip6q *q6, *q6tmp;
892 	uint32_t bucket;
893 
894 	if (atomic_load_int(&frag6_nfrags) == 0)
895 		return;
896 
897 	VNET_LIST_RLOCK_NOSLEEP();
898 	VNET_FOREACH(vnet_iter) {
899 		CURVNET_SET(vnet_iter);
900 		for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
901 			if (V_ip6qb[bucket].count == 0)
902 				continue;
903 			IP6QB_LOCK(bucket);
904 			head = IP6QB_HEAD(bucket);
905 			TAILQ_FOREACH_SAFE(q6, head, ip6q_tq, q6tmp)
906 				if (--q6->ip6q_ttl == 0) {
907 					IP6STAT_ADD(ip6s_fragtimeout,
908 						q6->ip6q_nfrag);
909 					/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
910 					frag6_freef(q6, bucket);
911 				}
912 			/*
913 			 * If we are over the maximum number of fragments
914 			 * (due to the limit being lowered), drain off
915 			 * enough to get down to the new limit.
916 			 * Note that we drain all reassembly queues if
917 			 * maxfragpackets is 0 (fragmentation is disabled),
918 			 * and do not enforce a limit when maxfragpackets
919 			 * is negative.
920 			 */
921 			while ((V_ip6_maxfragpackets == 0 ||
922 			    (V_ip6_maxfragpackets > 0 &&
923 			    V_ip6qb[bucket].count > V_ip6_maxfragbucketsize)) &&
924 			    (q6 = TAILQ_LAST(head, ip6qhead)) != NULL) {
925 				IP6STAT_ADD(ip6s_fragoverflow, q6->ip6q_nfrag);
926 				/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
927 				frag6_freef(q6, bucket);
928 			}
929 			IP6QB_UNLOCK(bucket);
930 		}
931 		/*
932 		 * If we are still over the maximum number of fragmented
933 		 * packets, drain off enough to get down to the new limit.
934 		 */
935 		bucket = 0;
936 		while (V_ip6_maxfragpackets >= 0 &&
937 		    atomic_load_int(&V_frag6_nfragpackets) >
938 		    (u_int)V_ip6_maxfragpackets) {
939 			IP6QB_LOCK(bucket);
940 			q6 = TAILQ_LAST(IP6QB_HEAD(bucket), ip6qhead);
941 			if (q6 != NULL) {
942 				IP6STAT_ADD(ip6s_fragoverflow, q6->ip6q_nfrag);
943 				/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
944 				frag6_freef(q6, bucket);
945 			}
946 			IP6QB_UNLOCK(bucket);
947 			bucket = (bucket + 1) % IP6REASS_NHASH;
948 		}
949 		CURVNET_RESTORE();
950 	}
951 	VNET_LIST_RUNLOCK_NOSLEEP();
952 }
953 
954 /*
955  * Eventhandler to adjust limits in case nmbclusters change.
956  */
957 static void
958 frag6_change(void *tag)
959 {
960 	VNET_ITERATOR_DECL(vnet_iter);
961 
962 	ip6_maxfrags = IP6_MAXFRAGS;
963 	VNET_LIST_RLOCK_NOSLEEP();
964 	VNET_FOREACH(vnet_iter) {
965 		CURVNET_SET(vnet_iter);
966 		V_ip6_maxfragpackets = IP6_MAXFRAGPACKETS;
967 		frag6_set_bucketsize();
968 		CURVNET_RESTORE();
969 	}
970 	VNET_LIST_RUNLOCK_NOSLEEP();
971 }
972 
973 /*
974  * Initialise reassembly queue and fragment identifier.
975  */
976 void
977 frag6_init(void)
978 {
979 	uint32_t bucket;
980 
981 	V_ip6_maxfragpackets = IP6_MAXFRAGPACKETS;
982 	frag6_set_bucketsize();
983 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
984 		TAILQ_INIT(IP6QB_HEAD(bucket));
985 		mtx_init(&V_ip6qb[bucket].lock, "ip6qb", NULL, MTX_DEF);
986 		V_ip6qb[bucket].count = 0;
987 	}
988 	V_ip6qb_hashseed = arc4random();
989 	V_ip6_maxfragsperpacket = 64;
990 #ifdef VIMAGE
991 	V_frag6_on = true;
992 #endif
993 	if (!IS_DEFAULT_VNET(curvnet))
994 		return;
995 
996 	ip6_maxfrags = IP6_MAXFRAGS;
997 	EVENTHANDLER_REGISTER(nmbclusters_change,
998 	    frag6_change, NULL, EVENTHANDLER_PRI_ANY);
999 }
1000 
1001 /*
1002  * Drain off all datagram fragments.
1003  */
1004 static void
1005 frag6_drain_one(void)
1006 {
1007 	struct ip6q *q6;
1008 	uint32_t bucket;
1009 
1010 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
1011 		IP6QB_LOCK(bucket);
1012 		while ((q6 = TAILQ_FIRST(IP6QB_HEAD(bucket))) != NULL) {
1013 			IP6STAT_INC(ip6s_fragdropped);
1014 			/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
1015 			frag6_freef(q6, bucket);
1016 		}
1017 		IP6QB_UNLOCK(bucket);
1018 	}
1019 }
1020 
1021 void
1022 frag6_drain(void)
1023 {
1024 	VNET_ITERATOR_DECL(vnet_iter);
1025 
1026 	VNET_LIST_RLOCK_NOSLEEP();
1027 	VNET_FOREACH(vnet_iter) {
1028 		CURVNET_SET(vnet_iter);
1029 		frag6_drain_one();
1030 		CURVNET_RESTORE();
1031 	}
1032 	VNET_LIST_RUNLOCK_NOSLEEP();
1033 }
1034 
1035 #ifdef VIMAGE
1036 /*
1037  * Clear up IPv6 reassembly structures.
1038  */
1039 void
1040 frag6_destroy(void)
1041 {
1042 	uint32_t bucket;
1043 
1044 	frag6_drain_one();
1045 	V_frag6_on = false;
1046 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
1047 		KASSERT(V_ip6qb[bucket].count == 0,
1048 		    ("%s: V_ip6qb[%d] (%p) count not 0 (%d)", __func__,
1049 		    bucket, &V_ip6qb[bucket], V_ip6qb[bucket].count));
1050 		mtx_destroy(&V_ip6qb[bucket].lock);
1051 	}
1052 }
1053 #endif
1054