xref: /freebsd/sys/netinet6/frag6.c (revision 0bf56da32d83fbd3b5db8d6c72cd1e7cc26fbc66)
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 | CTLFLAG_NEEDGIANT,
201 	NULL, 0, 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 __unused)
222 {
223 	struct ip6_hdr *ip6;
224 
225 	KASSERT(m->m_len >= offset + sizeof(struct ip6_frag),
226 	    ("%s: ext headers not contigous in mbuf %p m_len %d >= "
227 	    "offset %d + %zu\n", __func__, m, m->m_len, offset,
228 	    sizeof(struct ip6_frag)));
229 
230 	/* Delete frag6 header. */
231 	ip6 = mtod(m, struct ip6_hdr *);
232 	bcopy(ip6, (char *)ip6 + sizeof(struct ip6_frag), offset);
233 	m->m_data += sizeof(struct ip6_frag);
234 	m->m_len -= sizeof(struct ip6_frag);
235 	m->m_flags |= M_FRAGMENTED;
236 
237 	return (0);
238 }
239 
240 /*
241  * Free a fragment reassembly header and all associated datagrams.
242  */
243 static void
244 frag6_freef(struct ip6q *q6, uint32_t bucket)
245 {
246 	struct ip6_hdr *ip6;
247 	struct ip6asfrag *af6;
248 	struct mbuf *m;
249 
250 	IP6QB_LOCK_ASSERT(bucket);
251 
252 	while ((af6 = TAILQ_FIRST(&q6->ip6q_frags)) != NULL) {
253 
254 		m = af6->ip6af_m;
255 		TAILQ_REMOVE(&q6->ip6q_frags, af6, ip6af_tq);
256 
257 		/*
258 		 * Return ICMP time exceeded error for the 1st fragment.
259 		 * Just free other fragments.
260 		 */
261 		if (af6->ip6af_off == 0 && m->m_pkthdr.rcvif != NULL) {
262 
263 			/* Adjust pointer. */
264 			ip6 = mtod(m, struct ip6_hdr *);
265 
266 			/* Restore source and destination addresses. */
267 			ip6->ip6_src = q6->ip6q_src;
268 			ip6->ip6_dst = q6->ip6q_dst;
269 
270 			icmp6_error(m, ICMP6_TIME_EXCEEDED,
271 			    ICMP6_TIME_EXCEED_REASSEMBLY, 0);
272 		} else
273 			m_freem(m);
274 
275 		free(af6, M_FRAG6);
276 	}
277 
278 	TAILQ_REMOVE(IP6QB_HEAD(bucket), q6, ip6q_tq);
279 	V_ip6qb[bucket].count--;
280 	atomic_subtract_int(&frag6_nfrags, q6->ip6q_nfrag);
281 #ifdef MAC
282 	mac_ip6q_destroy(q6);
283 #endif
284 	free(q6, M_FRAG6);
285 	atomic_subtract_int(&V_frag6_nfragpackets, 1);
286 }
287 
288 /*
289  * Drain off all datagram fragments belonging to
290  * the given network interface.
291  */
292 static void
293 frag6_cleanup(void *arg __unused, struct ifnet *ifp)
294 {
295 	struct ip6qhead *head;
296 	struct ip6q *q6;
297 	struct ip6asfrag *af6;
298 	uint32_t bucket;
299 
300 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
301 
302 	CURVNET_SET_QUIET(ifp->if_vnet);
303 #ifdef VIMAGE
304 	/*
305 	 * Skip processing if IPv6 reassembly is not initialised or
306 	 * torn down by frag6_destroy().
307 	 */
308 	if (!V_frag6_on) {
309 		CURVNET_RESTORE();
310 		return;
311 	}
312 #endif
313 
314 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
315 		IP6QB_LOCK(bucket);
316 		head = IP6QB_HEAD(bucket);
317 		/* Scan fragment list. */
318 		TAILQ_FOREACH(q6, head, ip6q_tq) {
319 			TAILQ_FOREACH(af6, &q6->ip6q_frags, ip6af_tq) {
320 
321 				/* Clear no longer valid rcvif pointer. */
322 				if (af6->ip6af_m->m_pkthdr.rcvif == ifp)
323 					af6->ip6af_m->m_pkthdr.rcvif = NULL;
324 			}
325 		}
326 		IP6QB_UNLOCK(bucket);
327 	}
328 	CURVNET_RESTORE();
329 }
330 EVENTHANDLER_DEFINE(ifnet_departure_event, frag6_cleanup, NULL, 0);
331 
332 /*
333  * Like in RFC2460, in RFC8200, fragment and reassembly rules do not agree with
334  * each other, in terms of next header field handling in fragment header.
335  * While the sender will use the same value for all of the fragmented packets,
336  * receiver is suggested not to check for consistency.
337  *
338  * Fragment rules (p18,p19):
339  *	(2)  A Fragment header containing:
340  *	The Next Header value that identifies the first header
341  *	after the Per-Fragment headers of the original packet.
342  *		-> next header field is same for all fragments
343  *
344  * Reassembly rule (p20):
345  *	The Next Header field of the last header of the Per-Fragment
346  *	headers is obtained from the Next Header field of the first
347  *	fragment's Fragment header.
348  *		-> should grab it from the first fragment only
349  *
350  * The following note also contradicts with fragment rule - no one is going to
351  * send different fragment with different next header field.
352  *
353  * Additional note (p22) [not an error]:
354  *	The Next Header values in the Fragment headers of different
355  *	fragments of the same original packet may differ.  Only the value
356  *	from the Offset zero fragment packet is used for reassembly.
357  *		-> should grab it from the first fragment only
358  *
359  * There is no explicit reason given in the RFC.  Historical reason maybe?
360  */
361 /*
362  * Fragment input.
363  */
364 int
365 frag6_input(struct mbuf **mp, int *offp, int proto)
366 {
367 	struct mbuf *m, *t;
368 	struct ip6_hdr *ip6;
369 	struct ip6_frag *ip6f;
370 	struct ip6qhead *head;
371 	struct ip6q *q6;
372 	struct ip6asfrag *af6, *ip6af, *af6tmp;
373 	struct in6_ifaddr *ia6;
374 	struct ifnet *dstifp, *srcifp;
375 	uint32_t hashkey[(sizeof(struct in6_addr) * 2 +
376 		    sizeof(ip6f->ip6f_ident)) / sizeof(uint32_t)];
377 	uint32_t bucket, *hashkeyp;
378 	int fragoff, frgpartlen;	/* Must be larger than uint16_t. */
379 	int nxt, offset, plen;
380 	uint8_t ecn, ecn0;
381 	bool only_frag;
382 #ifdef RSS
383 	struct ip6_direct_ctx *ip6dc;
384 	struct m_tag *mtag;
385 #endif
386 
387 	m = *mp;
388 	offset = *offp;
389 
390 	M_ASSERTPKTHDR(m);
391 
392 	if (m->m_len < offset + sizeof(struct ip6_frag)) {
393 		m = m_pullup(m, offset + sizeof(struct ip6_frag));
394 		if (m == NULL) {
395 			IP6STAT_INC(ip6s_exthdrtoolong);
396 			*mp = NULL;
397 			return (IPPROTO_DONE);
398 		}
399 	}
400 	ip6 = mtod(m, struct ip6_hdr *);
401 
402 	dstifp = NULL;
403 	/* Find the destination interface of the packet. */
404 	ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */);
405 	if (ia6 != NULL) {
406 		dstifp = ia6->ia_ifp;
407 		ifa_free(&ia6->ia_ifa);
408 	}
409 
410 	/* Jumbo payload cannot contain a fragment header. */
411 	if (ip6->ip6_plen == 0) {
412 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER, offset);
413 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
414 		*mp = NULL;
415 		return (IPPROTO_DONE);
416 	}
417 
418 	/*
419 	 * Check whether fragment packet's fragment length is a
420 	 * multiple of 8 octets (unless it is the last one).
421 	 * sizeof(struct ip6_frag) == 8
422 	 * sizeof(struct ip6_hdr) = 40
423 	 */
424 	ip6f = (struct ip6_frag *)((caddr_t)ip6 + offset);
425 	if ((ip6f->ip6f_offlg & IP6F_MORE_FRAG) &&
426 	    (((ntohs(ip6->ip6_plen) - offset) & 0x7) != 0)) {
427 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
428 		    offsetof(struct ip6_hdr, ip6_plen));
429 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
430 		*mp = NULL;
431 		return (IPPROTO_DONE);
432 	}
433 
434 	IP6STAT_INC(ip6s_fragments);
435 	in6_ifstat_inc(dstifp, ifs6_reass_reqd);
436 
437 	/*
438 	 * Handle "atomic" fragments (offset and m bit set to 0) upfront,
439 	 * unrelated to any reassembly.  We need to remove the frag hdr
440 	 * which is ugly.
441 	 * See RFC 6946 and section 4.5 of RFC 8200.
442 	 */
443 	if ((ip6f->ip6f_offlg & ~IP6F_RESERVED_MASK) == 0) {
444 		IP6STAT_INC(ip6s_atomicfrags);
445 		nxt = ip6f->ip6f_nxt;
446 		/*
447 		 * Set nxt(-hdr field value) to the original value.
448 		 * We cannot just set ip6->ip6_nxt as there might be
449 		 * an unfragmentable part with extension headers and
450 		 * we must update the last one.
451 		 */
452 		m_copyback(m, ip6_get_prevhdr(m, offset), sizeof(uint8_t),
453 		    (caddr_t)&nxt);
454 		ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) -
455 		    sizeof(struct ip6_frag));
456 		if (ip6_deletefraghdr(m, offset, M_NOWAIT) != 0)
457 			goto dropfrag2;
458 		m->m_pkthdr.len -= sizeof(struct ip6_frag);
459 		in6_ifstat_inc(dstifp, ifs6_reass_ok);
460 		*mp = m;
461 		return (nxt);
462 	}
463 
464 	/* Offset now points to data portion. */
465 	offset += sizeof(struct ip6_frag);
466 
467 	/* Get fragment length and discard 0-byte fragments. */
468 	frgpartlen = sizeof(struct ip6_hdr) + ntohs(ip6->ip6_plen) - offset;
469 	if (frgpartlen == 0) {
470 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
471 		    offsetof(struct ip6_hdr, ip6_plen));
472 		in6_ifstat_inc(dstifp, ifs6_reass_fail);
473 		IP6STAT_INC(ip6s_fragdropped);
474 		*mp = NULL;
475 		return (IPPROTO_DONE);
476 	}
477 
478 	/*
479 	 * Enforce upper bound on number of fragments for the entire system.
480 	 * If maxfrag is 0, never accept fragments.
481 	 * If maxfrag is -1, accept all fragments without limitation.
482 	 */
483 	if (ip6_maxfrags < 0)
484 		;
485 	else if (atomic_load_int(&frag6_nfrags) >= (u_int)ip6_maxfrags)
486 		goto dropfrag2;
487 
488 	/*
489 	 * Validate that a full header chain to the ULP is present in the
490 	 * packet containing the first fragment as per RFC RFC7112 and
491 	 * RFC 8200 pages 18,19:
492 	 * The first fragment packet is composed of:
493 	 * (3)  Extension headers, if any, and the Upper-Layer header.  These
494 	 *      headers must be in the first fragment.  ...
495 	 */
496 	fragoff = ntohs(ip6f->ip6f_offlg & IP6F_OFF_MASK);
497 	/* XXX TODO.  thj has D16851 open for this. */
498 	/* Send ICMPv6 4,3 in case of violation. */
499 
500 	/* Store receive network interface pointer for later. */
501 	srcifp = m->m_pkthdr.rcvif;
502 
503 	/* Generate a hash value for fragment bucket selection. */
504 	hashkeyp = hashkey;
505 	memcpy(hashkeyp, &ip6->ip6_src, sizeof(struct in6_addr));
506 	hashkeyp += sizeof(struct in6_addr) / sizeof(*hashkeyp);
507 	memcpy(hashkeyp, &ip6->ip6_dst, sizeof(struct in6_addr));
508 	hashkeyp += sizeof(struct in6_addr) / sizeof(*hashkeyp);
509 	*hashkeyp = ip6f->ip6f_ident;
510 	bucket = jenkins_hash32(hashkey, nitems(hashkey), V_ip6qb_hashseed);
511 	bucket &= IP6REASS_HMASK;
512 	IP6QB_LOCK(bucket);
513 	head = IP6QB_HEAD(bucket);
514 
515 	TAILQ_FOREACH(q6, head, ip6q_tq)
516 		if (ip6f->ip6f_ident == q6->ip6q_ident &&
517 		    IN6_ARE_ADDR_EQUAL(&ip6->ip6_src, &q6->ip6q_src) &&
518 		    IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst, &q6->ip6q_dst)
519 #ifdef MAC
520 		    && mac_ip6q_match(m, q6)
521 #endif
522 		    )
523 			break;
524 
525 	only_frag = false;
526 	if (q6 == NULL) {
527 
528 		/* A first fragment to arrive creates a reassembly queue. */
529 		only_frag = true;
530 
531 		/*
532 		 * Enforce upper bound on number of fragmented packets
533 		 * for which we attempt reassembly;
534 		 * If maxfragpackets is 0, never accept fragments.
535 		 * If maxfragpackets is -1, accept all fragments without
536 		 * limitation.
537 		 */
538 		if (V_ip6_maxfragpackets < 0)
539 			;
540 		else if (V_ip6qb[bucket].count >= V_ip6_maxfragbucketsize ||
541 		    atomic_load_int(&V_frag6_nfragpackets) >=
542 		    (u_int)V_ip6_maxfragpackets)
543 			goto dropfrag;
544 
545 		/* Allocate IPv6 fragement packet queue entry. */
546 		q6 = (struct ip6q *)malloc(sizeof(struct ip6q), M_FRAG6,
547 		    M_NOWAIT | M_ZERO);
548 		if (q6 == NULL)
549 			goto dropfrag;
550 #ifdef MAC
551 		if (mac_ip6q_init(q6, M_NOWAIT) != 0) {
552 			free(q6, M_FRAG6);
553 			goto dropfrag;
554 		}
555 		mac_ip6q_create(m, q6);
556 #endif
557 		atomic_add_int(&V_frag6_nfragpackets, 1);
558 
559 		/* ip6q_nxt will be filled afterwards, from 1st fragment. */
560 		TAILQ_INIT(&q6->ip6q_frags);
561 		q6->ip6q_ident	= ip6f->ip6f_ident;
562 		q6->ip6q_ttl	= IPV6_FRAGTTL;
563 		q6->ip6q_src	= ip6->ip6_src;
564 		q6->ip6q_dst	= ip6->ip6_dst;
565 		q6->ip6q_ecn	=
566 		    (ntohl(ip6->ip6_flow) >> 20) & IPTOS_ECN_MASK;
567 		q6->ip6q_unfrglen = -1;	/* The 1st fragment has not arrived. */
568 
569 		/* Add the fragemented packet to the bucket. */
570 		TAILQ_INSERT_HEAD(head, q6, ip6q_tq);
571 		V_ip6qb[bucket].count++;
572 	}
573 
574 	/*
575 	 * If it is the 1st fragment, record the length of the
576 	 * unfragmentable part and the next header of the fragment header.
577 	 * Assume the first 1st fragement to arrive will be correct.
578 	 * We do not have any duplicate checks here yet so another packet
579 	 * with fragoff == 0 could come and overwrite the ip6q_unfrglen
580 	 * and worse, the next header, at any time.
581 	 */
582 	if (fragoff == 0 && q6->ip6q_unfrglen == -1) {
583 		q6->ip6q_unfrglen = offset - sizeof(struct ip6_hdr) -
584 		    sizeof(struct ip6_frag);
585 		q6->ip6q_nxt = ip6f->ip6f_nxt;
586 		/* XXX ECN? */
587 	}
588 
589 	/*
590 	 * Check that the reassembled packet would not exceed 65535 bytes
591 	 * in size.
592 	 * If it would exceed, discard the fragment and return an ICMP error.
593 	 */
594 	if (q6->ip6q_unfrglen >= 0) {
595 		/* The 1st fragment has already arrived. */
596 		if (q6->ip6q_unfrglen + fragoff + frgpartlen > IPV6_MAXPACKET) {
597 			if (only_frag) {
598 				TAILQ_REMOVE(head, q6, ip6q_tq);
599 				V_ip6qb[bucket].count--;
600 				atomic_subtract_int(&V_frag6_nfragpackets, 1);
601 #ifdef MAC
602 				mac_ip6q_destroy(q6);
603 #endif
604 				free(q6, M_FRAG6);
605 			}
606 			IP6QB_UNLOCK(bucket);
607 			icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
608 			    offset - sizeof(struct ip6_frag) +
609 			    offsetof(struct ip6_frag, ip6f_offlg));
610 			*mp = NULL;
611 			return (IPPROTO_DONE);
612 		}
613 	} else if (fragoff + frgpartlen > IPV6_MAXPACKET) {
614 		if (only_frag) {
615 			TAILQ_REMOVE(head, q6, ip6q_tq);
616 			V_ip6qb[bucket].count--;
617 			atomic_subtract_int(&V_frag6_nfragpackets, 1);
618 #ifdef MAC
619 			mac_ip6q_destroy(q6);
620 #endif
621 			free(q6, M_FRAG6);
622 		}
623 		IP6QB_UNLOCK(bucket);
624 		icmp6_error(m, ICMP6_PARAM_PROB, ICMP6_PARAMPROB_HEADER,
625 		    offset - sizeof(struct ip6_frag) +
626 		    offsetof(struct ip6_frag, ip6f_offlg));
627 		*mp = NULL;
628 		return (IPPROTO_DONE);
629 	}
630 
631 	/*
632 	 * If it is the first fragment, do the above check for each
633 	 * fragment already stored in the reassembly queue.
634 	 */
635 	if (fragoff == 0 && !only_frag) {
636 		TAILQ_FOREACH_SAFE(af6, &q6->ip6q_frags, ip6af_tq, af6tmp) {
637 
638 			if (q6->ip6q_unfrglen + af6->ip6af_off +
639 			    af6->ip6af_frglen > IPV6_MAXPACKET) {
640 				struct ip6_hdr *ip6err;
641 				struct mbuf *merr;
642 				int erroff;
643 
644 				merr = af6->ip6af_m;
645 				erroff = af6->ip6af_offset;
646 
647 				/* Dequeue the fragment. */
648 				TAILQ_REMOVE(&q6->ip6q_frags, af6, ip6af_tq);
649 				q6->ip6q_nfrag--;
650 				atomic_subtract_int(&frag6_nfrags, 1);
651 				free(af6, M_FRAG6);
652 
653 				/* Set a valid receive interface pointer. */
654 				merr->m_pkthdr.rcvif = srcifp;
655 
656 				/* Adjust pointer. */
657 				ip6err = mtod(merr, struct ip6_hdr *);
658 
659 				/*
660 				 * Restore source and destination addresses
661 				 * in the erroneous IPv6 header.
662 				 */
663 				ip6err->ip6_src = q6->ip6q_src;
664 				ip6err->ip6_dst = q6->ip6q_dst;
665 
666 				icmp6_error(merr, ICMP6_PARAM_PROB,
667 				    ICMP6_PARAMPROB_HEADER,
668 				    erroff - sizeof(struct ip6_frag) +
669 				    offsetof(struct ip6_frag, ip6f_offlg));
670 			}
671 		}
672 	}
673 
674 	/* Allocate an IPv6 fragement queue entry for this fragmented part. */
675 	ip6af = (struct ip6asfrag *)malloc(sizeof(struct ip6asfrag), M_FRAG6,
676 	    M_NOWAIT | M_ZERO);
677 	if (ip6af == NULL)
678 		goto dropfrag;
679 	ip6af->ip6af_mff = (ip6f->ip6f_offlg & IP6F_MORE_FRAG) ? true : false;
680 	ip6af->ip6af_off = fragoff;
681 	ip6af->ip6af_frglen = frgpartlen;
682 	ip6af->ip6af_offset = offset;
683 	ip6af->ip6af_m = m;
684 
685 	if (only_frag) {
686 		/*
687 		 * Do a manual insert rather than a hard-to-understand cast
688 		 * to a different type relying on data structure order to work.
689 		 */
690 		TAILQ_INSERT_HEAD(&q6->ip6q_frags, ip6af, ip6af_tq);
691 		goto postinsert;
692 	}
693 
694 	/* Do duplicate, condition, and boundry checks. */
695 	/*
696 	 * Handle ECN by comparing this segment with the first one;
697 	 * if CE is set, do not lose CE.
698 	 * Drop if CE and not-ECT are mixed for the same packet.
699 	 */
700 	ecn = (ntohl(ip6->ip6_flow) >> 20) & IPTOS_ECN_MASK;
701 	ecn0 = q6->ip6q_ecn;
702 	if (ecn == IPTOS_ECN_CE) {
703 		if (ecn0 == IPTOS_ECN_NOTECT) {
704 			free(ip6af, M_FRAG6);
705 			goto dropfrag;
706 		}
707 		if (ecn0 != IPTOS_ECN_CE)
708 			q6->ip6q_ecn = IPTOS_ECN_CE;
709 	}
710 	if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT) {
711 		free(ip6af, M_FRAG6);
712 		goto dropfrag;
713 	}
714 
715 	/* Find a fragmented part which begins after this one does. */
716 	TAILQ_FOREACH(af6, &q6->ip6q_frags, ip6af_tq)
717 		if (af6->ip6af_off > ip6af->ip6af_off)
718 			break;
719 
720 	/*
721 	 * If the incoming framgent overlaps some existing fragments in
722 	 * the reassembly queue, drop both the new fragment and the
723 	 * entire reassembly queue.  However, if the new fragment
724 	 * is an exact duplicate of an existing fragment, only silently
725 	 * drop the existing fragment and leave the fragmentation queue
726 	 * unchanged, as allowed by the RFC.  (RFC 8200, 4.5)
727 	 */
728 	if (af6 != NULL)
729 		af6tmp = TAILQ_PREV(af6, ip6fraghead, ip6af_tq);
730 	else
731 		af6tmp = TAILQ_LAST(&q6->ip6q_frags, ip6fraghead);
732 	if (af6tmp != NULL) {
733 		if (af6tmp->ip6af_off + af6tmp->ip6af_frglen -
734 		    ip6af->ip6af_off > 0) {
735 			if (af6tmp->ip6af_off != ip6af->ip6af_off ||
736 			    af6tmp->ip6af_frglen != ip6af->ip6af_frglen)
737 				frag6_freef(q6, bucket);
738 			free(ip6af, M_FRAG6);
739 			goto dropfrag;
740 		}
741 	}
742 	if (af6 != NULL) {
743 		if (ip6af->ip6af_off + ip6af->ip6af_frglen -
744 		    af6->ip6af_off > 0) {
745 			if (af6->ip6af_off != ip6af->ip6af_off ||
746 			    af6->ip6af_frglen != ip6af->ip6af_frglen)
747 				frag6_freef(q6, bucket);
748 			free(ip6af, M_FRAG6);
749 			goto dropfrag;
750 		}
751 	}
752 
753 #ifdef MAC
754 	mac_ip6q_update(m, q6);
755 #endif
756 
757 	/*
758 	 * Stick new segment in its place; check for complete reassembly.
759 	 * If not complete, check fragment limit.  Move to front of packet
760 	 * queue, as we are the most recently active fragmented packet.
761 	 */
762 	if (af6 != NULL)
763 		TAILQ_INSERT_BEFORE(af6, ip6af, ip6af_tq);
764 	else
765 		TAILQ_INSERT_TAIL(&q6->ip6q_frags, ip6af, ip6af_tq);
766 postinsert:
767 	atomic_add_int(&frag6_nfrags, 1);
768 	q6->ip6q_nfrag++;
769 
770 	plen = 0;
771 	TAILQ_FOREACH(af6, &q6->ip6q_frags, ip6af_tq) {
772 		if (af6->ip6af_off != plen) {
773 			if (q6->ip6q_nfrag > V_ip6_maxfragsperpacket) {
774 				IP6STAT_ADD(ip6s_fragdropped, q6->ip6q_nfrag);
775 				frag6_freef(q6, bucket);
776 			}
777 			IP6QB_UNLOCK(bucket);
778 			*mp = NULL;
779 			return (IPPROTO_DONE);
780 		}
781 		plen += af6->ip6af_frglen;
782 	}
783 	af6 = TAILQ_LAST(&q6->ip6q_frags, ip6fraghead);
784 	if (af6->ip6af_mff) {
785 		if (q6->ip6q_nfrag > V_ip6_maxfragsperpacket) {
786 			IP6STAT_ADD(ip6s_fragdropped, q6->ip6q_nfrag);
787 			frag6_freef(q6, bucket);
788 		}
789 		IP6QB_UNLOCK(bucket);
790 		*mp = NULL;
791 		return (IPPROTO_DONE);
792 	}
793 
794 	/* Reassembly is complete; concatenate fragments. */
795 	ip6af = TAILQ_FIRST(&q6->ip6q_frags);
796 	t = m = ip6af->ip6af_m;
797 	TAILQ_REMOVE(&q6->ip6q_frags, ip6af, ip6af_tq);
798 	while ((af6 = TAILQ_FIRST(&q6->ip6q_frags)) != NULL) {
799 		m->m_pkthdr.csum_flags &=
800 		    af6->ip6af_m->m_pkthdr.csum_flags;
801 		m->m_pkthdr.csum_data +=
802 		    af6->ip6af_m->m_pkthdr.csum_data;
803 
804 		TAILQ_REMOVE(&q6->ip6q_frags, af6, ip6af_tq);
805 		t = m_last(t);
806 		m_adj(af6->ip6af_m, af6->ip6af_offset);
807 		m_demote_pkthdr(af6->ip6af_m);
808 		m_cat(t, af6->ip6af_m);
809 		free(af6, M_FRAG6);
810 	}
811 
812 	while (m->m_pkthdr.csum_data & 0xffff0000)
813 		m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
814 		    (m->m_pkthdr.csum_data >> 16);
815 
816 	/* Adjust offset to point where the original next header starts. */
817 	offset = ip6af->ip6af_offset - sizeof(struct ip6_frag);
818 	free(ip6af, M_FRAG6);
819 	ip6 = mtod(m, struct ip6_hdr *);
820 	ip6->ip6_plen = htons((u_short)plen + offset - sizeof(struct ip6_hdr));
821 	if (q6->ip6q_ecn == IPTOS_ECN_CE)
822 		ip6->ip6_flow |= htonl(IPTOS_ECN_CE << 20);
823 	nxt = q6->ip6q_nxt;
824 
825 	TAILQ_REMOVE(head, q6, ip6q_tq);
826 	V_ip6qb[bucket].count--;
827 	atomic_subtract_int(&frag6_nfrags, q6->ip6q_nfrag);
828 
829 	ip6_deletefraghdr(m, offset, M_NOWAIT);
830 
831 	/* Set nxt(-hdr field value) to the original value. */
832 	m_copyback(m, ip6_get_prevhdr(m, offset), sizeof(uint8_t),
833 	    (caddr_t)&nxt);
834 
835 #ifdef MAC
836 	mac_ip6q_reassemble(q6, m);
837 	mac_ip6q_destroy(q6);
838 #endif
839 	free(q6, M_FRAG6);
840 	atomic_subtract_int(&V_frag6_nfragpackets, 1);
841 
842 	if (m->m_flags & M_PKTHDR) { /* Isn't it always true? */
843 
844 		plen = 0;
845 		for (t = m; t; t = t->m_next)
846 			plen += t->m_len;
847 		m->m_pkthdr.len = plen;
848 		/* Set a valid receive interface pointer. */
849 		m->m_pkthdr.rcvif = srcifp;
850 	}
851 
852 #ifdef RSS
853 	mtag = m_tag_alloc(MTAG_ABI_IPV6, IPV6_TAG_DIRECT, sizeof(*ip6dc),
854 	    M_NOWAIT);
855 	if (mtag == NULL)
856 		goto dropfrag;
857 
858 	ip6dc = (struct ip6_direct_ctx *)(mtag + 1);
859 	ip6dc->ip6dc_nxt = nxt;
860 	ip6dc->ip6dc_off = offset;
861 
862 	m_tag_prepend(m, mtag);
863 #endif
864 
865 	IP6QB_UNLOCK(bucket);
866 	IP6STAT_INC(ip6s_reassembled);
867 	in6_ifstat_inc(dstifp, ifs6_reass_ok);
868 
869 #ifdef RSS
870 	/* Queue/dispatch for reprocessing. */
871 	netisr_dispatch(NETISR_IPV6_DIRECT, m);
872 	*mp = NULL;
873 	return (IPPROTO_DONE);
874 #endif
875 
876 	/* Tell launch routine the next header. */
877 	*mp = m;
878 	*offp = offset;
879 
880 	return (nxt);
881 
882 dropfrag:
883 	IP6QB_UNLOCK(bucket);
884 dropfrag2:
885 	in6_ifstat_inc(dstifp, ifs6_reass_fail);
886 	IP6STAT_INC(ip6s_fragdropped);
887 	m_freem(m);
888 	*mp = NULL;
889 	return (IPPROTO_DONE);
890 }
891 
892 /*
893  * IPv6 reassembling timer processing;
894  * if a timer expires on a reassembly queue, discard it.
895  */
896 void
897 frag6_slowtimo(void)
898 {
899 	VNET_ITERATOR_DECL(vnet_iter);
900 	struct ip6qhead *head;
901 	struct ip6q *q6, *q6tmp;
902 	uint32_t bucket;
903 
904 	VNET_LIST_RLOCK_NOSLEEP();
905 	VNET_FOREACH(vnet_iter) {
906 		CURVNET_SET(vnet_iter);
907 		for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
908 			IP6QB_LOCK(bucket);
909 			head = IP6QB_HEAD(bucket);
910 			TAILQ_FOREACH_SAFE(q6, head, ip6q_tq, q6tmp)
911 				if (--q6->ip6q_ttl == 0) {
912 					IP6STAT_ADD(ip6s_fragtimeout,
913 						q6->ip6q_nfrag);
914 					/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
915 					frag6_freef(q6, bucket);
916 				}
917 			/*
918 			 * If we are over the maximum number of fragments
919 			 * (due to the limit being lowered), drain off
920 			 * enough to get down to the new limit.
921 			 * Note that we drain all reassembly queues if
922 			 * maxfragpackets is 0 (fragmentation is disabled),
923 			 * and do not enforce a limit when maxfragpackets
924 			 * is negative.
925 			 */
926 			while ((V_ip6_maxfragpackets == 0 ||
927 			    (V_ip6_maxfragpackets > 0 &&
928 			    V_ip6qb[bucket].count > V_ip6_maxfragbucketsize)) &&
929 			    (q6 = TAILQ_LAST(head, ip6qhead)) != NULL) {
930 				IP6STAT_ADD(ip6s_fragoverflow, q6->ip6q_nfrag);
931 				/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
932 				frag6_freef(q6, bucket);
933 			}
934 			IP6QB_UNLOCK(bucket);
935 		}
936 		/*
937 		 * If we are still over the maximum number of fragmented
938 		 * packets, drain off enough to get down to the new limit.
939 		 */
940 		bucket = 0;
941 		while (V_ip6_maxfragpackets >= 0 &&
942 		    atomic_load_int(&V_frag6_nfragpackets) >
943 		    (u_int)V_ip6_maxfragpackets) {
944 			IP6QB_LOCK(bucket);
945 			q6 = TAILQ_LAST(IP6QB_HEAD(bucket), ip6qhead);
946 			if (q6 != NULL) {
947 				IP6STAT_ADD(ip6s_fragoverflow, q6->ip6q_nfrag);
948 				/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
949 				frag6_freef(q6, bucket);
950 			}
951 			IP6QB_UNLOCK(bucket);
952 			bucket = (bucket + 1) % IP6REASS_NHASH;
953 		}
954 		CURVNET_RESTORE();
955 	}
956 	VNET_LIST_RUNLOCK_NOSLEEP();
957 }
958 
959 /*
960  * Eventhandler to adjust limits in case nmbclusters change.
961  */
962 static void
963 frag6_change(void *tag)
964 {
965 	VNET_ITERATOR_DECL(vnet_iter);
966 
967 	ip6_maxfrags = IP6_MAXFRAGS;
968 	VNET_LIST_RLOCK_NOSLEEP();
969 	VNET_FOREACH(vnet_iter) {
970 		CURVNET_SET(vnet_iter);
971 		V_ip6_maxfragpackets = IP6_MAXFRAGPACKETS;
972 		frag6_set_bucketsize();
973 		CURVNET_RESTORE();
974 	}
975 	VNET_LIST_RUNLOCK_NOSLEEP();
976 }
977 
978 /*
979  * Initialise reassembly queue and fragment identifier.
980  */
981 void
982 frag6_init(void)
983 {
984 	uint32_t bucket;
985 
986 	V_ip6_maxfragpackets = IP6_MAXFRAGPACKETS;
987 	frag6_set_bucketsize();
988 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
989 		TAILQ_INIT(IP6QB_HEAD(bucket));
990 		mtx_init(&V_ip6qb[bucket].lock, "ip6qb", NULL, MTX_DEF);
991 		V_ip6qb[bucket].count = 0;
992 	}
993 	V_ip6qb_hashseed = arc4random();
994 	V_ip6_maxfragsperpacket = 64;
995 #ifdef VIMAGE
996 	V_frag6_on = true;
997 #endif
998 	if (!IS_DEFAULT_VNET(curvnet))
999 		return;
1000 
1001 	ip6_maxfrags = IP6_MAXFRAGS;
1002 	EVENTHANDLER_REGISTER(nmbclusters_change,
1003 	    frag6_change, NULL, EVENTHANDLER_PRI_ANY);
1004 }
1005 
1006 /*
1007  * Drain off all datagram fragments.
1008  */
1009 static void
1010 frag6_drain_one(void)
1011 {
1012 	struct ip6q *q6;
1013 	uint32_t bucket;
1014 
1015 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
1016 		IP6QB_LOCK(bucket);
1017 		while ((q6 = TAILQ_FIRST(IP6QB_HEAD(bucket))) != NULL) {
1018 			IP6STAT_INC(ip6s_fragdropped);
1019 			/* XXX in6_ifstat_inc(ifp, ifs6_reass_fail) */
1020 			frag6_freef(q6, bucket);
1021 		}
1022 		IP6QB_UNLOCK(bucket);
1023 	}
1024 }
1025 
1026 void
1027 frag6_drain(void)
1028 {
1029 	VNET_ITERATOR_DECL(vnet_iter);
1030 
1031 	VNET_LIST_RLOCK_NOSLEEP();
1032 	VNET_FOREACH(vnet_iter) {
1033 		CURVNET_SET(vnet_iter);
1034 		frag6_drain_one();
1035 		CURVNET_RESTORE();
1036 	}
1037 	VNET_LIST_RUNLOCK_NOSLEEP();
1038 }
1039 
1040 #ifdef VIMAGE
1041 /*
1042  * Clear up IPv6 reassembly structures.
1043  */
1044 void
1045 frag6_destroy(void)
1046 {
1047 	uint32_t bucket;
1048 
1049 	frag6_drain_one();
1050 	V_frag6_on = false;
1051 	for (bucket = 0; bucket < IP6REASS_NHASH; bucket++) {
1052 		KASSERT(V_ip6qb[bucket].count == 0,
1053 		    ("%s: V_ip6qb[%d] (%p) count not 0 (%d)", __func__,
1054 		    bucket, &V_ip6qb[bucket], V_ip6qb[bucket].count));
1055 		mtx_destroy(&V_ip6qb[bucket].lock);
1056 	}
1057 }
1058 #endif
1059