xref: /freebsd/sys/netinet6/frag6.c (revision f5ef5f675d9d9eb6e35ed9142d70ecf774456ad4)
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 volatile u_int frag6_nfrags = 0;
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 /*
163  * Sysctls and helper function.
164  */
165 SYSCTL_DECL(_net_inet6_ip6);
166 
167 SYSCTL_UINT(_net_inet6_ip6, OID_AUTO, frag6_nfrags,
168 	CTLFLAG_RD, __DEVOLATILE(u_int *, &frag6_nfrags), 0,
169 	"Global number of IPv6 fragments across all reassembly queues.");
170 
171 static void
172 frag6_set_bucketsize(void)
173 {
174 	int i;
175 
176 	if ((i = V_ip6_maxfragpackets) > 0)
177 		V_ip6_maxfragbucketsize = imax(i / (IP6REASS_NHASH / 2), 1);
178 }
179 
180 SYSCTL_INT(_net_inet6_ip6, IPV6CTL_MAXFRAGS, maxfrags,
181 	CTLFLAG_RW, &ip6_maxfrags, 0,
182 	"Maximum allowed number of outstanding IPv6 packet fragments. "
183 	"A value of 0 means no fragmented packets will be accepted, while a "
184 	"a value of -1 means no limit");
185 
186 static int
187 sysctl_ip6_maxfragpackets(SYSCTL_HANDLER_ARGS)
188 {
189 	int error, val;
190 
191 	val = V_ip6_maxfragpackets;
192 	error = sysctl_handle_int(oidp, &val, 0, req);
193 	if (error != 0 || !req->newptr)
194 		return (error);
195 	V_ip6_maxfragpackets = val;
196 	frag6_set_bucketsize();
197 	return (0);
198 }
199 SYSCTL_PROC(_net_inet6_ip6, IPV6CTL_MAXFRAGPACKETS, maxfragpackets,
200 	CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW, NULL, 0,
201 	sysctl_ip6_maxfragpackets, "I",
202 	"Default maximum number of outstanding fragmented IPv6 packets. "
203 	"A value of 0 means no fragmented packets will be accepted, while a "
204 	"a value of -1 means no limit");
205 SYSCTL_UINT(_net_inet6_ip6, OID_AUTO, frag6_nfragpackets,
206 	CTLFLAG_VNET | CTLFLAG_RD,
207 	__DEVOLATILE(u_int *, &VNET_NAME(frag6_nfragpackets)), 0,
208 	"Per-VNET number of IPv6 fragments across all reassembly queues.");
209 SYSCTL_INT(_net_inet6_ip6, IPV6CTL_MAXFRAGSPERPACKET, maxfragsperpacket,
210 	CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ip6_maxfragsperpacket), 0,
211 	"Maximum allowed number of fragments per packet");
212 SYSCTL_INT(_net_inet6_ip6, IPV6CTL_MAXFRAGBUCKETSIZE, maxfragbucketsize,
213 	CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ip6_maxfragbucketsize), 0,
214 	"Maximum number of reassembly queues per hash bucket");
215 
216 
217 /*
218  * Remove the IPv6 fragmentation header from the mbuf.
219  */
220 int
221 ip6_deletefraghdr(struct mbuf *m, int offset, int wait)
222 {
223 	struct ip6_hdr *ip6;
224 	struct mbuf *t;
225 
226 	/* Delete frag6 header. */
227 	if (m->m_len >= offset + sizeof(struct ip6_frag)) {
228 
229 		/* This is the only possible case with !PULLDOWN_TEST. */
230 		ip6 = mtod(m, struct ip6_hdr *);
231 		bcopy(ip6, (char *)ip6 + sizeof(struct ip6_frag),
232 		    offset);
233 		m->m_data += sizeof(struct ip6_frag);
234 		m->m_len -= sizeof(struct ip6_frag);
235 	} else {
236 
237 		/* This comes with no copy if the boundary is on cluster. */
238 		if ((t = m_split(m, offset, wait)) == NULL)
239 			return (ENOMEM);
240 		m_adj(t, sizeof(struct ip6_frag));
241 		m_cat(m, t);
242 	}
243 
244 	m->m_flags |= M_FRAGMENTED;
245 	return (0);
246 }
247 
248 /*
249  * Free a fragment reassembly header and all associated datagrams.
250  */
251 static void
252 frag6_freef(struct ip6q *q6, uint32_t bucket)
253 {
254 	struct ip6_hdr *ip6;
255 	struct ip6asfrag *af6;
256 	struct mbuf *m;
257 
258 	IP6QB_LOCK_ASSERT(bucket);
259 
260 	while ((af6 = TAILQ_FIRST(&q6->ip6q_frags)) != NULL) {
261 
262 		m = af6->ip6af_m;
263 		TAILQ_REMOVE(&q6->ip6q_frags, af6, ip6af_tq);
264 
265 		/*
266 		 * Return ICMP time exceeded error for the 1st fragment.
267 		 * Just free other fragments.
268 		 */
269 		if (af6->ip6af_off == 0 && m->m_pkthdr.rcvif != NULL) {
270 
271 			/* Adjust pointer. */
272 			ip6 = mtod(m, struct ip6_hdr *);
273 
274 			/* Restore source and destination addresses. */
275 			ip6->ip6_src = q6->ip6q_src;
276 			ip6->ip6_dst = q6->ip6q_dst;
277 
278 			icmp6_error(m, ICMP6_TIME_EXCEEDED,
279 			    ICMP6_TIME_EXCEED_REASSEMBLY, 0);
280 		} else
281 			m_freem(m);
282 
283 		free(af6, M_FRAG6);
284 	}
285 
286 	TAILQ_REMOVE(IP6QB_HEAD(bucket), q6, ip6q_tq);
287 	V_ip6qb[bucket].count--;
288 	atomic_subtract_int(&frag6_nfrags, q6->ip6q_nfrag);
289 #ifdef MAC
290 	mac_ip6q_destroy(q6);
291 #endif
292 	free(q6, M_FRAG6);
293 	atomic_subtract_int(&V_frag6_nfragpackets, 1);
294 }
295 
296 /*
297  * Drain off all datagram fragments belonging to
298  * the given network interface.
299  */
300 static void
301 frag6_cleanup(void *arg __unused, struct ifnet *ifp)
302 {
303 	struct ip6qhead *head;
304 	struct ip6q *q6;
305 	struct ip6asfrag *af6;
306 	uint32_t bucket;
307 
308 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
309 
310 	CURVNET_SET_QUIET(ifp->if_vnet);
311 #ifdef VIMAGE
312 	/*
313 	 * Skip processing if IPv6 reassembly is not initialised or
314 	 * torn down by frag6_destroy().
315 	 */
316 	if (!V_frag6_on) {
317 		CURVNET_RESTORE();
318 		return;
319 	}
320 #endif
321 
322 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
323 		IP6QB_LOCK(bucket);
324 		head = IP6QB_HEAD(bucket);
325 		/* Scan fragment list. */
326 		TAILQ_FOREACH(q6, head, ip6q_tq) {
327 			TAILQ_FOREACH(af6, &q6->ip6q_frags, ip6af_tq) {
328 
329 				/* Clear no longer valid rcvif pointer. */
330 				if (af6->ip6af_m->m_pkthdr.rcvif == ifp)
331 					af6->ip6af_m->m_pkthdr.rcvif = NULL;
332 			}
333 		}
334 		IP6QB_UNLOCK(bucket);
335 	}
336 	CURVNET_RESTORE();
337 }
338 EVENTHANDLER_DEFINE(ifnet_departure_event, frag6_cleanup, NULL, 0);
339 
340 /*
341  * Like in RFC2460, in RFC8200, fragment and reassembly rules do not agree with
342  * each other, in terms of next header field handling in fragment header.
343  * While the sender will use the same value for all of the fragmented packets,
344  * receiver is suggested not to check for consistency.
345  *
346  * Fragment rules (p18,p19):
347  *	(2)  A Fragment header containing:
348  *	The Next Header value that identifies the first header
349  *	after the Per-Fragment headers of the original packet.
350  *		-> next header field is same for all fragments
351  *
352  * Reassembly rule (p20):
353  *	The Next Header field of the last header of the Per-Fragment
354  *	headers is obtained from the Next Header field of the first
355  *	fragment's Fragment header.
356  *		-> should grab it from the first fragment only
357  *
358  * The following note also contradicts with fragment rule - no one is going to
359  * send different fragment with different next header field.
360  *
361  * Additional note (p22) [not an error]:
362  *	The Next Header values in the Fragment headers of different
363  *	fragments of the same original packet may differ.  Only the value
364  *	from the Offset zero fragment packet is used for reassembly.
365  *		-> should grab it from the first fragment only
366  *
367  * There is no explicit reason given in the RFC.  Historical reason maybe?
368  */
369 /*
370  * Fragment input.
371  */
372 int
373 frag6_input(struct mbuf **mp, int *offp, int proto)
374 {
375 	struct mbuf *m, *t;
376 	struct ip6_hdr *ip6;
377 	struct ip6_frag *ip6f;
378 	struct ip6qhead *head;
379 	struct ip6q *q6;
380 	struct ip6asfrag *af6, *ip6af, *af6tmp;
381 	struct in6_ifaddr *ia6;
382 	struct ifnet *dstifp, *srcifp;
383 	uint32_t hashkey[(sizeof(struct in6_addr) * 2 +
384 		    sizeof(ip6f->ip6f_ident)) / sizeof(uint32_t)];
385 	uint32_t bucket, *hashkeyp;
386 	int fragoff, frgpartlen;	/* Must be larger than uint16_t. */
387 	int nxt, offset, plen;
388 	uint8_t ecn, ecn0;
389 	bool only_frag;
390 #ifdef RSS
391 	struct ip6_direct_ctx *ip6dc;
392 	struct m_tag *mtag;
393 #endif
394 
395 	m = *mp;
396 	offset = *offp;
397 
398 	ip6 = mtod(m, struct ip6_hdr *);
399 #ifndef PULLDOWN_TEST
400 	IP6_EXTHDR_CHECK(m, offset, sizeof(struct ip6_frag), IPPROTO_DONE);
401 	ip6f = (struct ip6_frag *)((caddr_t)ip6 + offset);
402 #else
403 	IP6_EXTHDR_GET(ip6f, struct ip6_frag *, m, offset, sizeof(*ip6f));
404 	if (ip6f == NULL)
405 		return (IPPROTO_DONE);
406 #endif
407 
408 	dstifp = NULL;
409 	/* Find the destination interface of the packet. */
410 	ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */);
411 	if (ia6 != NULL) {
412 		dstifp = ia6->ia_ifp;
413 		ifa_free(&ia6->ia_ifa);
414 	}
415 
416 	/* Jumbo payload cannot contain a fragment header. */
417 	if (ip6->ip6_plen == 0) {
418 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER, offset);
419 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
420 		return (IPPROTO_DONE);
421 	}
422 
423 	/*
424 	 * Check whether fragment packet's fragment length is a
425 	 * multiple of 8 octets (unless it is the last one).
426 	 * sizeof(struct ip6_frag) == 8
427 	 * sizeof(struct ip6_hdr) = 40
428 	 */
429 	if ((ip6f->ip6f_offlg & IP6F_MORE_FRAG) &&
430 	    (((ntohs(ip6->ip6_plen) - offset) & 0x7) != 0)) {
431 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
432 		    offsetof(struct ip6_hdr, ip6_plen));
433 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
434 		return (IPPROTO_DONE);
435 	}
436 
437 	IP6STAT_INC(ip6s_fragments);
438 	in6_ifstat_inc(dstifp, ifs6_reass_reqd);
439 
440 	/* Offset now points to data portion. */
441 	offset += sizeof(struct ip6_frag);
442 
443 	/*
444 	 * Handle "atomic" fragments (offset and m bit set to 0) upfront,
445 	 * unrelated to any reassembly.  Still need to remove the frag hdr.
446 	 * See RFC 6946 and section 4.5 of RFC 8200.
447 	 */
448 	if ((ip6f->ip6f_offlg & ~IP6F_RESERVED_MASK) == 0) {
449 		IP6STAT_INC(ip6s_atomicfrags);
450 		/* XXX-BZ handle correctly. */
451 		in6_ifstat_inc(dstifp, ifs6_reass_ok);
452 		*offp = offset;
453 		m->m_flags |= M_FRAGMENTED;
454 		return (ip6f->ip6f_nxt);
455 	}
456 
457 	/* Get fragment length and discard 0-byte fragments. */
458 	frgpartlen = sizeof(struct ip6_hdr) + ntohs(ip6->ip6_plen) - offset;
459 	if (frgpartlen == 0) {
460 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
461 		    offsetof(struct ip6_hdr, ip6_plen));
462 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
463 		IP6STAT_INC(ip6s_fragdropped);
464 		return (IPPROTO_DONE);
465 	}
466 
467 	/*
468 	 * Enforce upper bound on number of fragments for the entire system.
469 	 * If maxfrag is 0, never accept fragments.
470 	 * If maxfrag is -1, accept all fragments without limitation.
471 	 */
472 	if (ip6_maxfrags < 0)
473 		;
474 	else if (atomic_load_int(&frag6_nfrags) >= (u_int)ip6_maxfrags)
475 		goto dropfrag2;
476 
477 	/*
478 	 * Validate that a full header chain to the ULP is present in the
479 	 * packet containing the first fragment as per RFC RFC7112 and
480 	 * RFC 8200 pages 18,19:
481 	 * The first fragment packet is composed of:
482 	 * (3)  Extension headers, if any, and the Upper-Layer header.  These
483 	 *      headers must be in the first fragment.  ...
484 	 */
485 	fragoff = ntohs(ip6f->ip6f_offlg & IP6F_OFF_MASK);
486 	/* XXX TODO.  thj has D16851 open for this. */
487 	/* Send ICMPv6 4,3 in case of violation. */
488 
489 	/* Store receive network interface pointer for later. */
490 	srcifp = m->m_pkthdr.rcvif;
491 
492 	/* Generate a hash value for fragment bucket selection. */
493 	hashkeyp = hashkey;
494 	memcpy(hashkeyp, &ip6->ip6_src, sizeof(struct in6_addr));
495 	hashkeyp += sizeof(struct in6_addr) / sizeof(*hashkeyp);
496 	memcpy(hashkeyp, &ip6->ip6_dst, sizeof(struct in6_addr));
497 	hashkeyp += sizeof(struct in6_addr) / sizeof(*hashkeyp);
498 	*hashkeyp = ip6f->ip6f_ident;
499 	bucket = jenkins_hash32(hashkey, nitems(hashkey), V_ip6qb_hashseed);
500 	bucket &= IP6REASS_HMASK;
501 	IP6QB_LOCK(bucket);
502 	head = IP6QB_HEAD(bucket);
503 
504 	TAILQ_FOREACH(q6, head, ip6q_tq)
505 		if (ip6f->ip6f_ident == q6->ip6q_ident &&
506 		    IN6_ARE_ADDR_EQUAL(&ip6->ip6_src, &q6->ip6q_src) &&
507 		    IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst, &q6->ip6q_dst)
508 #ifdef MAC
509 		    && mac_ip6q_match(m, q6)
510 #endif
511 		    )
512 			break;
513 
514 	only_frag = false;
515 	if (q6 == NULL) {
516 
517 		/* A first fragment to arrive creates a reassembly queue. */
518 		only_frag = true;
519 
520 		/*
521 		 * Enforce upper bound on number of fragmented packets
522 		 * for which we attempt reassembly;
523 		 * If maxfragpackets is 0, never accept fragments.
524 		 * If maxfragpackets is -1, accept all fragments without
525 		 * limitation.
526 		 */
527 		if (V_ip6_maxfragpackets < 0)
528 			;
529 		else if (V_ip6qb[bucket].count >= V_ip6_maxfragbucketsize ||
530 		    atomic_load_int(&V_frag6_nfragpackets) >=
531 		    (u_int)V_ip6_maxfragpackets)
532 			goto dropfrag;
533 
534 		/* Allocate IPv6 fragement packet queue entry. */
535 		q6 = (struct ip6q *)malloc(sizeof(struct ip6q), M_FRAG6,
536 		    M_NOWAIT | M_ZERO);
537 		if (q6 == NULL)
538 			goto dropfrag;
539 #ifdef MAC
540 		if (mac_ip6q_init(q6, M_NOWAIT) != 0) {
541 			free(q6, M_FRAG6);
542 			goto dropfrag;
543 		}
544 		mac_ip6q_create(m, q6);
545 #endif
546 		atomic_add_int(&V_frag6_nfragpackets, 1);
547 
548 		/* ip6q_nxt will be filled afterwards, from 1st fragment. */
549 		TAILQ_INIT(&q6->ip6q_frags);
550 		q6->ip6q_ident	= ip6f->ip6f_ident;
551 		q6->ip6q_ttl	= IPV6_FRAGTTL;
552 		q6->ip6q_src	= ip6->ip6_src;
553 		q6->ip6q_dst	= ip6->ip6_dst;
554 		q6->ip6q_ecn	=
555 		    (ntohl(ip6->ip6_flow) >> 20) & IPTOS_ECN_MASK;
556 		q6->ip6q_unfrglen = -1;	/* The 1st fragment has not arrived. */
557 
558 		/* Add the fragemented packet to the bucket. */
559 		TAILQ_INSERT_HEAD(head, q6, ip6q_tq);
560 		V_ip6qb[bucket].count++;
561 	}
562 
563 	/*
564 	 * If it is the 1st fragment, record the length of the
565 	 * unfragmentable part and the next header of the fragment header.
566 	 * Assume the first 1st fragement to arrive will be correct.
567 	 * We do not have any duplicate checks here yet so another packet
568 	 * with fragoff == 0 could come and overwrite the ip6q_unfrglen
569 	 * and worse, the next header, at any time.
570 	 */
571 	if (fragoff == 0 && q6->ip6q_unfrglen == -1) {
572 		q6->ip6q_unfrglen = offset - sizeof(struct ip6_hdr) -
573 		    sizeof(struct ip6_frag);
574 		q6->ip6q_nxt = ip6f->ip6f_nxt;
575 		/* XXX ECN? */
576 	}
577 
578 	/*
579 	 * Check that the reassembled packet would not exceed 65535 bytes
580 	 * in size.
581 	 * If it would exceed, discard the fragment and return an ICMP error.
582 	 */
583 	if (q6->ip6q_unfrglen >= 0) {
584 		/* The 1st fragment has already arrived. */
585 		if (q6->ip6q_unfrglen + fragoff + frgpartlen > IPV6_MAXPACKET) {
586 			if (only_frag) {
587 				TAILQ_REMOVE(head, q6, ip6q_tq);
588 				V_ip6qb[bucket].count--;
589 				atomic_subtract_int(&V_frag6_nfragpackets, 1);
590 #ifdef MAC
591 				mac_ip6q_destroy(q6);
592 #endif
593 				free(q6, M_FRAG6);
594 			}
595 			IP6QB_UNLOCK(bucket);
596 			icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
597 			    offset - sizeof(struct ip6_frag) +
598 			    offsetof(struct ip6_frag, ip6f_offlg));
599 			return (IPPROTO_DONE);
600 		}
601 	} else if (fragoff + frgpartlen > IPV6_MAXPACKET) {
602 		if (only_frag) {
603 			TAILQ_REMOVE(head, q6, ip6q_tq);
604 			V_ip6qb[bucket].count--;
605 			atomic_subtract_int(&V_frag6_nfragpackets, 1);
606 #ifdef MAC
607 			mac_ip6q_destroy(q6);
608 #endif
609 			free(q6, M_FRAG6);
610 		}
611 		IP6QB_UNLOCK(bucket);
612 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
613 		    offset - sizeof(struct ip6_frag) +
614 		    offsetof(struct ip6_frag, ip6f_offlg));
615 		return (IPPROTO_DONE);
616 	}
617 
618 	/*
619 	 * If it is the first fragment, do the above check for each
620 	 * fragment already stored in the reassembly queue.
621 	 */
622 	if (fragoff == 0 && !only_frag) {
623 		TAILQ_FOREACH_SAFE(af6, &q6->ip6q_frags, ip6af_tq, af6tmp) {
624 
625 			if (q6->ip6q_unfrglen + af6->ip6af_off +
626 			    af6->ip6af_frglen > IPV6_MAXPACKET) {
627 				struct ip6_hdr *ip6err;
628 				struct mbuf *merr;
629 				int erroff;
630 
631 				merr = af6->ip6af_m;
632 				erroff = af6->ip6af_offset;
633 
634 				/* Dequeue the fragment. */
635 				TAILQ_REMOVE(&q6->ip6q_frags, af6, ip6af_tq);
636 				q6->ip6q_nfrag--;
637 				atomic_subtract_int(&frag6_nfrags, 1);
638 				free(af6, M_FRAG6);
639 
640 				/* Set a valid receive interface pointer. */
641 				merr->m_pkthdr.rcvif = srcifp;
642 
643 				/* Adjust pointer. */
644 				ip6err = mtod(merr, struct ip6_hdr *);
645 
646 				/*
647 				 * Restore source and destination addresses
648 				 * in the erroneous IPv6 header.
649 				 */
650 				ip6err->ip6_src = q6->ip6q_src;
651 				ip6err->ip6_dst = q6->ip6q_dst;
652 
653 				icmp6_error(merr, ICMP6_PARAM_PROB,
654 				    ICMP6_PARAMPROB_HEADER,
655 				    erroff - sizeof(struct ip6_frag) +
656 				    offsetof(struct ip6_frag, ip6f_offlg));
657 			}
658 		}
659 	}
660 
661 	/* Allocate an IPv6 fragement queue entry for this fragmented part. */
662 	ip6af = (struct ip6asfrag *)malloc(sizeof(struct ip6asfrag), M_FRAG6,
663 	    M_NOWAIT | M_ZERO);
664 	if (ip6af == NULL)
665 		goto dropfrag;
666 	ip6af->ip6af_mff = (ip6f->ip6f_offlg & IP6F_MORE_FRAG) ? true : false;
667 	ip6af->ip6af_off = fragoff;
668 	ip6af->ip6af_frglen = frgpartlen;
669 	ip6af->ip6af_offset = offset;
670 	ip6af->ip6af_m = m;
671 
672 	if (only_frag) {
673 		/*
674 		 * Do a manual insert rather than a hard-to-understand cast
675 		 * to a different type relying on data structure order to work.
676 		 */
677 		TAILQ_INSERT_HEAD(&q6->ip6q_frags, ip6af, ip6af_tq);
678 		goto postinsert;
679 	}
680 
681 	/* Do duplicate, condition, and boundry checks. */
682 	/*
683 	 * Handle ECN by comparing this segment with the first one;
684 	 * if CE is set, do not lose CE.
685 	 * Drop if CE and not-ECT are mixed for the same packet.
686 	 */
687 	ecn = (ntohl(ip6->ip6_flow) >> 20) & IPTOS_ECN_MASK;
688 	ecn0 = q6->ip6q_ecn;
689 	if (ecn == IPTOS_ECN_CE) {
690 		if (ecn0 == IPTOS_ECN_NOTECT) {
691 			free(ip6af, M_FRAG6);
692 			goto dropfrag;
693 		}
694 		if (ecn0 != IPTOS_ECN_CE)
695 			q6->ip6q_ecn = IPTOS_ECN_CE;
696 	}
697 	if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT) {
698 		free(ip6af, M_FRAG6);
699 		goto dropfrag;
700 	}
701 
702 	/* Find a fragmented part which begins after this one does. */
703 	TAILQ_FOREACH(af6, &q6->ip6q_frags, ip6af_tq)
704 		if (af6->ip6af_off > ip6af->ip6af_off)
705 			break;
706 
707 	/*
708 	 * If the incoming framgent overlaps some existing fragments in
709 	 * the reassembly queue, drop both the new fragment and the
710 	 * entire reassembly queue.  However, if the new fragment
711 	 * is an exact duplicate of an existing fragment, only silently
712 	 * drop the existing fragment and leave the fragmentation queue
713 	 * unchanged, as allowed by the RFC.  (RFC 8200, 4.5)
714 	 */
715 	if (af6 != NULL)
716 		af6tmp = TAILQ_PREV(af6, ip6fraghead, ip6af_tq);
717 	else
718 		af6tmp = TAILQ_LAST(&q6->ip6q_frags, ip6fraghead);
719 	if (af6tmp != NULL) {
720 		if (af6tmp->ip6af_off + af6tmp->ip6af_frglen -
721 		    ip6af->ip6af_off > 0) {
722 			if (af6tmp->ip6af_off != ip6af->ip6af_off ||
723 			    af6tmp->ip6af_frglen != ip6af->ip6af_frglen)
724 				frag6_freef(q6, bucket);
725 			free(ip6af, M_FRAG6);
726 			goto dropfrag;
727 		}
728 	}
729 	if (af6 != NULL) {
730 		if (ip6af->ip6af_off + ip6af->ip6af_frglen -
731 		    af6->ip6af_off > 0) {
732 			if (af6->ip6af_off != ip6af->ip6af_off ||
733 			    af6->ip6af_frglen != ip6af->ip6af_frglen)
734 				frag6_freef(q6, bucket);
735 			free(ip6af, M_FRAG6);
736 			goto dropfrag;
737 		}
738 	}
739 
740 #ifdef MAC
741 	mac_ip6q_update(m, q6);
742 #endif
743 
744 	/*
745 	 * Stick new segment in its place; check for complete reassembly.
746 	 * If not complete, check fragment limit.  Move to front of packet
747 	 * queue, as we are the most recently active fragmented packet.
748 	 */
749 	if (af6 != NULL)
750 		TAILQ_INSERT_BEFORE(af6, ip6af, ip6af_tq);
751 	else
752 		TAILQ_INSERT_TAIL(&q6->ip6q_frags, ip6af, ip6af_tq);
753 postinsert:
754 	atomic_add_int(&frag6_nfrags, 1);
755 	q6->ip6q_nfrag++;
756 
757 	plen = 0;
758 	TAILQ_FOREACH(af6, &q6->ip6q_frags, ip6af_tq) {
759 		if (af6->ip6af_off != plen) {
760 			if (q6->ip6q_nfrag > V_ip6_maxfragsperpacket) {
761 				IP6STAT_ADD(ip6s_fragdropped, q6->ip6q_nfrag);
762 				frag6_freef(q6, bucket);
763 			}
764 			IP6QB_UNLOCK(bucket);
765 			return (IPPROTO_DONE);
766 		}
767 		plen += af6->ip6af_frglen;
768 	}
769 	af6 = TAILQ_LAST(&q6->ip6q_frags, ip6fraghead);
770 	if (af6->ip6af_mff) {
771 		if (q6->ip6q_nfrag > V_ip6_maxfragsperpacket) {
772 			IP6STAT_ADD(ip6s_fragdropped, q6->ip6q_nfrag);
773 			frag6_freef(q6, bucket);
774 		}
775 		IP6QB_UNLOCK(bucket);
776 		return (IPPROTO_DONE);
777 	}
778 
779 	/* Reassembly is complete; concatenate fragments. */
780 	ip6af = TAILQ_FIRST(&q6->ip6q_frags);
781 	t = m = ip6af->ip6af_m;
782 	TAILQ_REMOVE(&q6->ip6q_frags, ip6af, ip6af_tq);
783 	while ((af6 = TAILQ_FIRST(&q6->ip6q_frags)) != NULL) {
784 		m->m_pkthdr.csum_flags &=
785 		    af6->ip6af_m->m_pkthdr.csum_flags;
786 		m->m_pkthdr.csum_data +=
787 		    af6->ip6af_m->m_pkthdr.csum_data;
788 
789 		TAILQ_REMOVE(&q6->ip6q_frags, af6, ip6af_tq);
790 		t = m_last(t);
791 		m_adj(af6->ip6af_m, af6->ip6af_offset);
792 		m_demote_pkthdr(af6->ip6af_m);
793 		m_cat(t, af6->ip6af_m);
794 		free(af6, M_FRAG6);
795 	}
796 
797 	while (m->m_pkthdr.csum_data & 0xffff0000)
798 		m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
799 		    (m->m_pkthdr.csum_data >> 16);
800 
801 	/* Adjust offset to point where the original next header starts. */
802 	offset = ip6af->ip6af_offset - sizeof(struct ip6_frag);
803 	free(ip6af, M_FRAG6);
804 	ip6 = mtod(m, struct ip6_hdr *);
805 	ip6->ip6_plen = htons((u_short)plen + offset - sizeof(struct ip6_hdr));
806 	if (q6->ip6q_ecn == IPTOS_ECN_CE)
807 		ip6->ip6_flow |= htonl(IPTOS_ECN_CE << 20);
808 	nxt = q6->ip6q_nxt;
809 
810 	TAILQ_REMOVE(head, q6, ip6q_tq);
811 	V_ip6qb[bucket].count--;
812 	atomic_subtract_int(&frag6_nfrags, q6->ip6q_nfrag);
813 
814 	if (ip6_deletefraghdr(m, offset, M_NOWAIT) != 0) {
815 #ifdef MAC
816 		mac_ip6q_destroy(q6);
817 #endif
818 		free(q6, M_FRAG6);
819 		atomic_subtract_int(&V_frag6_nfragpackets, 1);
820 
821 		goto dropfrag;
822 	}
823 
824 	/* Set nxt(-hdr field value) to the original value. */
825 	m_copyback(m, ip6_get_prevhdr(m, offset), sizeof(uint8_t),
826 	    (caddr_t)&nxt);
827 
828 #ifdef MAC
829 	mac_ip6q_reassemble(q6, m);
830 	mac_ip6q_destroy(q6);
831 #endif
832 	free(q6, M_FRAG6);
833 	atomic_subtract_int(&V_frag6_nfragpackets, 1);
834 
835 	if (m->m_flags & M_PKTHDR) { /* Isn't it always true? */
836 
837 		plen = 0;
838 		for (t = m; t; t = t->m_next)
839 			plen += t->m_len;
840 		m->m_pkthdr.len = plen;
841 		/* Set a valid receive interface pointer. */
842 		m->m_pkthdr.rcvif = srcifp;
843 	}
844 
845 #ifdef RSS
846 	mtag = m_tag_alloc(MTAG_ABI_IPV6, IPV6_TAG_DIRECT, sizeof(*ip6dc),
847 	    M_NOWAIT);
848 	if (mtag == NULL)
849 		goto dropfrag;
850 
851 	ip6dc = (struct ip6_direct_ctx *)(mtag + 1);
852 	ip6dc->ip6dc_nxt = nxt;
853 	ip6dc->ip6dc_off = offset;
854 
855 	m_tag_prepend(m, mtag);
856 #endif
857 
858 	IP6QB_UNLOCK(bucket);
859 	IP6STAT_INC(ip6s_reassembled);
860 	in6_ifstat_inc(dstifp, ifs6_reass_ok);
861 
862 #ifdef RSS
863 	/* Queue/dispatch for reprocessing. */
864 	netisr_dispatch(NETISR_IPV6_DIRECT, m);
865 	return (IPPROTO_DONE);
866 #endif
867 
868 	/* Tell launch routine the next header. */
869 	*mp = m;
870 	*offp = offset;
871 
872 	return (nxt);
873 
874 dropfrag:
875 	IP6QB_UNLOCK(bucket);
876 dropfrag2:
877 	in6_ifstat_inc(dstifp, ifs6_reass_fail);
878 	IP6STAT_INC(ip6s_fragdropped);
879 	m_freem(m);
880 	return (IPPROTO_DONE);
881 }
882 
883 /*
884  * IPv6 reassembling timer processing;
885  * if a timer expires on a reassembly queue, discard it.
886  */
887 void
888 frag6_slowtimo(void)
889 {
890 	VNET_ITERATOR_DECL(vnet_iter);
891 	struct ip6qhead *head;
892 	struct ip6q *q6, *q6tmp;
893 	uint32_t bucket;
894 
895 	VNET_LIST_RLOCK_NOSLEEP();
896 	VNET_FOREACH(vnet_iter) {
897 		CURVNET_SET(vnet_iter);
898 		for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
899 			IP6QB_LOCK(bucket);
900 			head = IP6QB_HEAD(bucket);
901 			TAILQ_FOREACH_SAFE(q6, head, ip6q_tq, q6tmp)
902 				if (--q6->ip6q_ttl == 0) {
903 					IP6STAT_ADD(ip6s_fragtimeout,
904 						q6->ip6q_nfrag);
905 					/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
906 					frag6_freef(q6, bucket);
907 				}
908 			/*
909 			 * If we are over the maximum number of fragments
910 			 * (due to the limit being lowered), drain off
911 			 * enough to get down to the new limit.
912 			 * Note that we drain all reassembly queues if
913 			 * maxfragpackets is 0 (fragmentation is disabled),
914 			 * and do not enforce a limit when maxfragpackets
915 			 * is negative.
916 			 */
917 			while ((V_ip6_maxfragpackets == 0 ||
918 			    (V_ip6_maxfragpackets > 0 &&
919 			    V_ip6qb[bucket].count > V_ip6_maxfragbucketsize)) &&
920 			    (q6 = TAILQ_LAST(head, ip6qhead)) != NULL) {
921 				IP6STAT_ADD(ip6s_fragoverflow, q6->ip6q_nfrag);
922 				/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
923 				frag6_freef(q6, bucket);
924 			}
925 			IP6QB_UNLOCK(bucket);
926 		}
927 		/*
928 		 * If we are still over the maximum number of fragmented
929 		 * packets, drain off enough to get down to the new limit.
930 		 */
931 		bucket = 0;
932 		while (V_ip6_maxfragpackets >= 0 &&
933 		    atomic_load_int(&V_frag6_nfragpackets) >
934 		    (u_int)V_ip6_maxfragpackets) {
935 			IP6QB_LOCK(bucket);
936 			q6 = TAILQ_LAST(IP6QB_HEAD(bucket), ip6qhead);
937 			if (q6 != NULL) {
938 				IP6STAT_ADD(ip6s_fragoverflow, q6->ip6q_nfrag);
939 				/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
940 				frag6_freef(q6, bucket);
941 			}
942 			IP6QB_UNLOCK(bucket);
943 			bucket = (bucket + 1) % IP6REASS_NHASH;
944 		}
945 		CURVNET_RESTORE();
946 	}
947 	VNET_LIST_RUNLOCK_NOSLEEP();
948 }
949 
950 /*
951  * Eventhandler to adjust limits in case nmbclusters change.
952  */
953 static void
954 frag6_change(void *tag)
955 {
956 	VNET_ITERATOR_DECL(vnet_iter);
957 
958 	ip6_maxfrags = IP6_MAXFRAGS;
959 	VNET_LIST_RLOCK_NOSLEEP();
960 	VNET_FOREACH(vnet_iter) {
961 		CURVNET_SET(vnet_iter);
962 		V_ip6_maxfragpackets = IP6_MAXFRAGPACKETS;
963 		frag6_set_bucketsize();
964 		CURVNET_RESTORE();
965 	}
966 	VNET_LIST_RUNLOCK_NOSLEEP();
967 }
968 
969 /*
970  * Initialise reassembly queue and fragment identifier.
971  */
972 void
973 frag6_init(void)
974 {
975 	uint32_t bucket;
976 
977 	V_ip6_maxfragpackets = IP6_MAXFRAGPACKETS;
978 	frag6_set_bucketsize();
979 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
980 		TAILQ_INIT(IP6QB_HEAD(bucket));
981 		mtx_init(&V_ip6qb[bucket].lock, "ip6qb", NULL, MTX_DEF);
982 		V_ip6qb[bucket].count = 0;
983 	}
984 	V_ip6qb_hashseed = arc4random();
985 	V_ip6_maxfragsperpacket = 64;
986 #ifdef VIMAGE
987 	V_frag6_on = true;
988 #endif
989 	if (!IS_DEFAULT_VNET(curvnet))
990 		return;
991 
992 	ip6_maxfrags = IP6_MAXFRAGS;
993 	EVENTHANDLER_REGISTER(nmbclusters_change,
994 	    frag6_change, NULL, EVENTHANDLER_PRI_ANY);
995 }
996 
997 /*
998  * Drain off all datagram fragments.
999  */
1000 static void
1001 frag6_drain_one(void)
1002 {
1003 	struct ip6q *q6;
1004 	uint32_t bucket;
1005 
1006 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
1007 		IP6QB_LOCK(bucket);
1008 		while ((q6 = TAILQ_FIRST(IP6QB_HEAD(bucket))) != NULL) {
1009 			IP6STAT_INC(ip6s_fragdropped);
1010 			/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
1011 			frag6_freef(q6, bucket);
1012 		}
1013 		IP6QB_UNLOCK(bucket);
1014 	}
1015 }
1016 
1017 void
1018 frag6_drain(void)
1019 {
1020 	VNET_ITERATOR_DECL(vnet_iter);
1021 
1022 	VNET_LIST_RLOCK_NOSLEEP();
1023 	VNET_FOREACH(vnet_iter) {
1024 		CURVNET_SET(vnet_iter);
1025 		frag6_drain_one();
1026 		CURVNET_RESTORE();
1027 	}
1028 	VNET_LIST_RUNLOCK_NOSLEEP();
1029 }
1030 
1031 #ifdef VIMAGE
1032 /*
1033  * Clear up IPv6 reassembly structures.
1034  */
1035 void
1036 frag6_destroy(void)
1037 {
1038 	uint32_t bucket;
1039 
1040 	frag6_drain_one();
1041 	V_frag6_on = false;
1042 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
1043 		KASSERT(V_ip6qb[bucket].count == 0,
1044 		    ("%s: V_ip6qb[%d] (%p) count not 0 (%d)", __func__,
1045 		    bucket, &V_ip6qb[bucket], V_ip6qb[bucket].count));
1046 		mtx_destroy(&V_ip6qb[bucket].lock);
1047 	}
1048 }
1049 #endif
1050