xref: /freebsd/sys/netinet/ip_reass.c (revision f4f33ea0c752ff0f9bfad34991d5bbb54e71133d)
1 /*-
2  * Copyright (c) 2015 Gleb Smirnoff <glebius@FreeBSD.org>
3  * Copyright (c) 2015 Adrian Chadd <adrian@FreeBSD.org>
4  * Copyright (c) 1982, 1986, 1988, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	@(#)ip_input.c	8.2 (Berkeley) 1/4/94
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include "opt_rss.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/eventhandler.h>
42 #include <sys/hash.h>
43 #include <sys/mbuf.h>
44 #include <sys/malloc.h>
45 #include <sys/limits.h>
46 #include <sys/lock.h>
47 #include <sys/mutex.h>
48 #include <sys/sysctl.h>
49 
50 #include <net/rss_config.h>
51 #include <net/netisr.h>
52 #include <net/vnet.h>
53 
54 #include <netinet/in.h>
55 #include <netinet/ip.h>
56 #include <netinet/ip_var.h>
57 #include <netinet/in_rss.h>
58 #ifdef MAC
59 #include <security/mac/mac_framework.h>
60 #endif
61 
62 SYSCTL_DECL(_net_inet_ip);
63 
64 /*
65  * Reassembly headers are stored in hash buckets.
66  */
67 #define	IPREASS_NHASH_LOG2	10
68 #define	IPREASS_NHASH		(1 << IPREASS_NHASH_LOG2)
69 #define	IPREASS_HMASK		(IPREASS_NHASH - 1)
70 
71 struct ipqbucket {
72 	TAILQ_HEAD(ipqhead, ipq) head;
73 	struct mtx		 lock;
74 	int			 count;
75 };
76 
77 VNET_DEFINE_STATIC(struct ipqbucket, ipq[IPREASS_NHASH]);
78 #define	V_ipq		VNET(ipq)
79 VNET_DEFINE_STATIC(uint32_t, ipq_hashseed);
80 #define V_ipq_hashseed   VNET(ipq_hashseed)
81 
82 #define	IPQ_LOCK(i)	mtx_lock(&V_ipq[i].lock)
83 #define	IPQ_TRYLOCK(i)	mtx_trylock(&V_ipq[i].lock)
84 #define	IPQ_UNLOCK(i)	mtx_unlock(&V_ipq[i].lock)
85 #define	IPQ_LOCK_ASSERT(i)	mtx_assert(&V_ipq[i].lock, MA_OWNED)
86 
87 VNET_DEFINE_STATIC(int, ipreass_maxbucketsize);
88 #define	V_ipreass_maxbucketsize	VNET(ipreass_maxbucketsize)
89 
90 void		ipreass_init(void);
91 void		ipreass_drain(void);
92 void		ipreass_slowtimo(void);
93 #ifdef VIMAGE
94 void		ipreass_destroy(void);
95 #endif
96 static int	sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS);
97 static int	sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS);
98 static void	ipreass_zone_change(void *);
99 static void	ipreass_drain_tomax(void);
100 static void	ipq_free(struct ipqbucket *, struct ipq *);
101 static struct ipq * ipq_reuse(int);
102 
103 static inline void
104 ipq_timeout(struct ipqbucket *bucket, struct ipq *fp)
105 {
106 
107 	IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
108 	ipq_free(bucket, fp);
109 }
110 
111 static inline void
112 ipq_drop(struct ipqbucket *bucket, struct ipq *fp)
113 {
114 
115 	IPSTAT_ADD(ips_fragdropped, fp->ipq_nfrags);
116 	ipq_free(bucket, fp);
117 }
118 
119 /*
120  * By default, limit the number of IP fragments across all reassembly
121  * queues to  1/32 of the total number of mbuf clusters.
122  *
123  * Limit the total number of reassembly queues per VNET to the
124  * IP fragment limit, but ensure the limit will not allow any bucket
125  * to grow above 100 items. (The bucket limit is
126  * IP_MAXFRAGPACKETS / (IPREASS_NHASH / 2), so the 50 is the correct
127  * multiplier to reach a 100-item limit.)
128  * The 100-item limit was chosen as brief testing seems to show that
129  * this produces "reasonable" performance on some subset of systems
130  * under DoS attack.
131  */
132 #define	IP_MAXFRAGS		(nmbclusters / 32)
133 #define	IP_MAXFRAGPACKETS	(imin(IP_MAXFRAGS, IPREASS_NHASH * 50))
134 
135 static int		maxfrags;
136 static volatile u_int	nfrags;
137 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfrags, CTLFLAG_RW,
138     &maxfrags, 0,
139     "Maximum number of IPv4 fragments allowed across all reassembly queues");
140 SYSCTL_UINT(_net_inet_ip, OID_AUTO, curfrags, CTLFLAG_RD,
141     __DEVOLATILE(u_int *, &nfrags), 0,
142     "Current number of IPv4 fragments across all reassembly queues");
143 
144 VNET_DEFINE_STATIC(uma_zone_t, ipq_zone);
145 #define	V_ipq_zone	VNET(ipq_zone)
146 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragpackets, CTLFLAG_VNET |
147     CTLTYPE_INT | CTLFLAG_RW, NULL, 0, sysctl_maxfragpackets, "I",
148     "Maximum number of IPv4 fragment reassembly queue entries");
149 SYSCTL_UMA_CUR(_net_inet_ip, OID_AUTO, fragpackets, CTLFLAG_VNET,
150     &VNET_NAME(ipq_zone),
151     "Current number of IPv4 fragment reassembly queue entries");
152 
153 VNET_DEFINE_STATIC(int, noreass);
154 #define	V_noreass	VNET(noreass)
155 
156 VNET_DEFINE_STATIC(int, maxfragsperpacket);
157 #define	V_maxfragsperpacket	VNET(maxfragsperpacket)
158 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfragsperpacket, CTLFLAG_VNET | CTLFLAG_RW,
159     &VNET_NAME(maxfragsperpacket), 0,
160     "Maximum number of IPv4 fragments allowed per packet");
161 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragbucketsize,
162     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0,
163     sysctl_maxfragbucketsize, "I",
164     "Maximum number of IPv4 fragment reassembly queue entries per bucket");
165 
166 /*
167  * Take incoming datagram fragment and try to reassemble it into
168  * whole datagram.  If the argument is the first fragment or one
169  * in between the function will return NULL and store the mbuf
170  * in the fragment chain.  If the argument is the last fragment
171  * the packet will be reassembled and the pointer to the new
172  * mbuf returned for further processing.  Only m_tags attached
173  * to the first packet/fragment are preserved.
174  * The IP header is *NOT* adjusted out of iplen.
175  */
176 #define	M_IP_FRAG	M_PROTO9
177 struct mbuf *
178 ip_reass(struct mbuf *m)
179 {
180 	struct ip *ip;
181 	struct mbuf *p, *q, *nq, *t;
182 	struct ipq *fp;
183 	struct ipqhead *head;
184 	int i, hlen, next, tmpmax;
185 	u_int8_t ecn, ecn0;
186 	uint32_t hash, hashkey[3];
187 #ifdef	RSS
188 	uint32_t rss_hash, rss_type;
189 #endif
190 
191 	/*
192 	 * If no reassembling or maxfragsperpacket are 0,
193 	 * never accept fragments.
194 	 * Also, drop packet if it would exceed the maximum
195 	 * number of fragments.
196 	 */
197 	tmpmax = maxfrags;
198 	if (V_noreass == 1 || V_maxfragsperpacket == 0 ||
199 	    (tmpmax >= 0 && atomic_load_int(&nfrags) >= (u_int)tmpmax)) {
200 		IPSTAT_INC(ips_fragments);
201 		IPSTAT_INC(ips_fragdropped);
202 		m_freem(m);
203 		return (NULL);
204 	}
205 
206 	ip = mtod(m, struct ip *);
207 	hlen = ip->ip_hl << 2;
208 
209 	/*
210 	 * Adjust ip_len to not reflect header,
211 	 * convert offset of this to bytes.
212 	 */
213 	ip->ip_len = htons(ntohs(ip->ip_len) - hlen);
214 	if (ip->ip_off & htons(IP_MF)) {
215 		/*
216 		 * Make sure that fragments have a data length
217 		 * that's a non-zero multiple of 8 bytes.
218 		 */
219 		if (ip->ip_len == htons(0) || (ntohs(ip->ip_len) & 0x7) != 0) {
220 			IPSTAT_INC(ips_toosmall); /* XXX */
221 			IPSTAT_INC(ips_fragdropped);
222 			m_freem(m);
223 			return (NULL);
224 		}
225 		m->m_flags |= M_IP_FRAG;
226 	} else
227 		m->m_flags &= ~M_IP_FRAG;
228 	ip->ip_off = htons(ntohs(ip->ip_off) << 3);
229 
230 	/*
231 	 * Attempt reassembly; if it succeeds, proceed.
232 	 * ip_reass() will return a different mbuf.
233 	 */
234 	IPSTAT_INC(ips_fragments);
235 	m->m_pkthdr.PH_loc.ptr = ip;
236 
237 	/*
238 	 * Presence of header sizes in mbufs
239 	 * would confuse code below.
240 	 */
241 	m->m_data += hlen;
242 	m->m_len -= hlen;
243 
244 	hashkey[0] = ip->ip_src.s_addr;
245 	hashkey[1] = ip->ip_dst.s_addr;
246 	hashkey[2] = (uint32_t)ip->ip_p << 16;
247 	hashkey[2] += ip->ip_id;
248 	hash = jenkins_hash32(hashkey, nitems(hashkey), V_ipq_hashseed);
249 	hash &= IPREASS_HMASK;
250 	head = &V_ipq[hash].head;
251 	IPQ_LOCK(hash);
252 
253 	/*
254 	 * Look for queue of fragments
255 	 * of this datagram.
256 	 */
257 	TAILQ_FOREACH(fp, head, ipq_list)
258 		if (ip->ip_id == fp->ipq_id &&
259 		    ip->ip_src.s_addr == fp->ipq_src.s_addr &&
260 		    ip->ip_dst.s_addr == fp->ipq_dst.s_addr &&
261 #ifdef MAC
262 		    mac_ipq_match(m, fp) &&
263 #endif
264 		    ip->ip_p == fp->ipq_p)
265 			break;
266 	/*
267 	 * If first fragment to arrive, create a reassembly queue.
268 	 */
269 	if (fp == NULL) {
270 		if (V_ipq[hash].count < V_ipreass_maxbucketsize)
271 			fp = uma_zalloc(V_ipq_zone, M_NOWAIT);
272 		if (fp == NULL)
273 			fp = ipq_reuse(hash);
274 		if (fp == NULL)
275 			goto dropfrag;
276 #ifdef MAC
277 		if (mac_ipq_init(fp, M_NOWAIT) != 0) {
278 			uma_zfree(V_ipq_zone, fp);
279 			fp = NULL;
280 			goto dropfrag;
281 		}
282 		mac_ipq_create(m, fp);
283 #endif
284 		TAILQ_INSERT_HEAD(head, fp, ipq_list);
285 		V_ipq[hash].count++;
286 		fp->ipq_nfrags = 1;
287 		atomic_add_int(&nfrags, 1);
288 		fp->ipq_ttl = IPFRAGTTL;
289 		fp->ipq_p = ip->ip_p;
290 		fp->ipq_id = ip->ip_id;
291 		fp->ipq_src = ip->ip_src;
292 		fp->ipq_dst = ip->ip_dst;
293 		fp->ipq_frags = m;
294 		m->m_nextpkt = NULL;
295 		goto done;
296 	} else {
297 		fp->ipq_nfrags++;
298 		atomic_add_int(&nfrags, 1);
299 #ifdef MAC
300 		mac_ipq_update(m, fp);
301 #endif
302 	}
303 
304 #define GETIP(m)	((struct ip*)((m)->m_pkthdr.PH_loc.ptr))
305 
306 	/*
307 	 * Handle ECN by comparing this segment with the first one;
308 	 * if CE is set, do not lose CE.
309 	 * drop if CE and not-ECT are mixed for the same packet.
310 	 */
311 	ecn = ip->ip_tos & IPTOS_ECN_MASK;
312 	ecn0 = GETIP(fp->ipq_frags)->ip_tos & IPTOS_ECN_MASK;
313 	if (ecn == IPTOS_ECN_CE) {
314 		if (ecn0 == IPTOS_ECN_NOTECT)
315 			goto dropfrag;
316 		if (ecn0 != IPTOS_ECN_CE)
317 			GETIP(fp->ipq_frags)->ip_tos |= IPTOS_ECN_CE;
318 	}
319 	if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT)
320 		goto dropfrag;
321 
322 	/*
323 	 * Find a segment which begins after this one does.
324 	 */
325 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt)
326 		if (ntohs(GETIP(q)->ip_off) > ntohs(ip->ip_off))
327 			break;
328 
329 	/*
330 	 * If there is a preceding segment, it may provide some of
331 	 * our data already.  If so, drop the data from the incoming
332 	 * segment.  If it provides all of our data, drop us, otherwise
333 	 * stick new segment in the proper place.
334 	 *
335 	 * If some of the data is dropped from the preceding
336 	 * segment, then it's checksum is invalidated.
337 	 */
338 	if (p) {
339 		i = ntohs(GETIP(p)->ip_off) + ntohs(GETIP(p)->ip_len) -
340 		    ntohs(ip->ip_off);
341 		if (i > 0) {
342 			if (i >= ntohs(ip->ip_len))
343 				goto dropfrag;
344 			m_adj(m, i);
345 			m->m_pkthdr.csum_flags = 0;
346 			ip->ip_off = htons(ntohs(ip->ip_off) + i);
347 			ip->ip_len = htons(ntohs(ip->ip_len) - i);
348 		}
349 		m->m_nextpkt = p->m_nextpkt;
350 		p->m_nextpkt = m;
351 	} else {
352 		m->m_nextpkt = fp->ipq_frags;
353 		fp->ipq_frags = m;
354 	}
355 
356 	/*
357 	 * While we overlap succeeding segments trim them or,
358 	 * if they are completely covered, dequeue them.
359 	 */
360 	for (; q != NULL && ntohs(ip->ip_off) + ntohs(ip->ip_len) >
361 	    ntohs(GETIP(q)->ip_off); q = nq) {
362 		i = (ntohs(ip->ip_off) + ntohs(ip->ip_len)) -
363 		    ntohs(GETIP(q)->ip_off);
364 		if (i < ntohs(GETIP(q)->ip_len)) {
365 			GETIP(q)->ip_len = htons(ntohs(GETIP(q)->ip_len) - i);
366 			GETIP(q)->ip_off = htons(ntohs(GETIP(q)->ip_off) + i);
367 			m_adj(q, i);
368 			q->m_pkthdr.csum_flags = 0;
369 			break;
370 		}
371 		nq = q->m_nextpkt;
372 		m->m_nextpkt = nq;
373 		IPSTAT_INC(ips_fragdropped);
374 		fp->ipq_nfrags--;
375 		atomic_subtract_int(&nfrags, 1);
376 		m_freem(q);
377 	}
378 
379 	/*
380 	 * Check for complete reassembly and perform frag per packet
381 	 * limiting.
382 	 *
383 	 * Frag limiting is performed here so that the nth frag has
384 	 * a chance to complete the packet before we drop the packet.
385 	 * As a result, n+1 frags are actually allowed per packet, but
386 	 * only n will ever be stored. (n = maxfragsperpacket.)
387 	 *
388 	 */
389 	next = 0;
390 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt) {
391 		if (ntohs(GETIP(q)->ip_off) != next) {
392 			if (fp->ipq_nfrags > V_maxfragsperpacket)
393 				ipq_drop(&V_ipq[hash], fp);
394 			goto done;
395 		}
396 		next += ntohs(GETIP(q)->ip_len);
397 	}
398 	/* Make sure the last packet didn't have the IP_MF flag */
399 	if (p->m_flags & M_IP_FRAG) {
400 		if (fp->ipq_nfrags > V_maxfragsperpacket)
401 			ipq_drop(&V_ipq[hash], fp);
402 		goto done;
403 	}
404 
405 	/*
406 	 * Reassembly is complete.  Make sure the packet is a sane size.
407 	 */
408 	q = fp->ipq_frags;
409 	ip = GETIP(q);
410 	if (next + (ip->ip_hl << 2) > IP_MAXPACKET) {
411 		IPSTAT_INC(ips_toolong);
412 		ipq_drop(&V_ipq[hash], fp);
413 		goto done;
414 	}
415 
416 	/*
417 	 * Concatenate fragments.
418 	 */
419 	m = q;
420 	t = m->m_next;
421 	m->m_next = NULL;
422 	m_cat(m, t);
423 	nq = q->m_nextpkt;
424 	q->m_nextpkt = NULL;
425 	for (q = nq; q != NULL; q = nq) {
426 		nq = q->m_nextpkt;
427 		q->m_nextpkt = NULL;
428 		m->m_pkthdr.csum_flags &= q->m_pkthdr.csum_flags;
429 		m->m_pkthdr.csum_data += q->m_pkthdr.csum_data;
430 		m_demote_pkthdr(q);
431 		m_cat(m, q);
432 	}
433 	/*
434 	 * In order to do checksumming faster we do 'end-around carry' here
435 	 * (and not in for{} loop), though it implies we are not going to
436 	 * reassemble more than 64k fragments.
437 	 */
438 	while (m->m_pkthdr.csum_data & 0xffff0000)
439 		m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
440 		    (m->m_pkthdr.csum_data >> 16);
441 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
442 #ifdef MAC
443 	mac_ipq_reassemble(fp, m);
444 	mac_ipq_destroy(fp);
445 #endif
446 
447 	/*
448 	 * Create header for new ip packet by modifying header of first
449 	 * packet;  dequeue and discard fragment reassembly header.
450 	 * Make header visible.
451 	 */
452 	ip->ip_len = htons((ip->ip_hl << 2) + next);
453 	ip->ip_src = fp->ipq_src;
454 	ip->ip_dst = fp->ipq_dst;
455 	TAILQ_REMOVE(head, fp, ipq_list);
456 	V_ipq[hash].count--;
457 	uma_zfree(V_ipq_zone, fp);
458 	m->m_len += (ip->ip_hl << 2);
459 	m->m_data -= (ip->ip_hl << 2);
460 	/* some debugging cruft by sklower, below, will go away soon */
461 	if (m->m_flags & M_PKTHDR)	/* XXX this should be done elsewhere */
462 		m_fixhdr(m);
463 	IPSTAT_INC(ips_reassembled);
464 	IPQ_UNLOCK(hash);
465 
466 #ifdef	RSS
467 	/*
468 	 * Query the RSS layer for the flowid / flowtype for the
469 	 * mbuf payload.
470 	 *
471 	 * For now, just assume we have to calculate a new one.
472 	 * Later on we should check to see if the assigned flowid matches
473 	 * what RSS wants for the given IP protocol and if so, just keep it.
474 	 *
475 	 * We then queue into the relevant netisr so it can be dispatched
476 	 * to the correct CPU.
477 	 *
478 	 * Note - this may return 1, which means the flowid in the mbuf
479 	 * is correct for the configured RSS hash types and can be used.
480 	 */
481 	if (rss_mbuf_software_hash_v4(m, 0, &rss_hash, &rss_type) == 0) {
482 		m->m_pkthdr.flowid = rss_hash;
483 		M_HASHTYPE_SET(m, rss_type);
484 	}
485 
486 	/*
487 	 * Queue/dispatch for reprocessing.
488 	 *
489 	 * Note: this is much slower than just handling the frame in the
490 	 * current receive context.  It's likely worth investigating
491 	 * why this is.
492 	 */
493 	netisr_dispatch(NETISR_IP_DIRECT, m);
494 	return (NULL);
495 #endif
496 
497 	/* Handle in-line */
498 	return (m);
499 
500 dropfrag:
501 	IPSTAT_INC(ips_fragdropped);
502 	if (fp != NULL) {
503 		fp->ipq_nfrags--;
504 		atomic_subtract_int(&nfrags, 1);
505 	}
506 	m_freem(m);
507 done:
508 	IPQ_UNLOCK(hash);
509 	return (NULL);
510 
511 #undef GETIP
512 }
513 
514 /*
515  * Initialize IP reassembly structures.
516  */
517 void
518 ipreass_init(void)
519 {
520 	int max;
521 
522 	for (int i = 0; i < IPREASS_NHASH; i++) {
523 		TAILQ_INIT(&V_ipq[i].head);
524 		mtx_init(&V_ipq[i].lock, "IP reassembly", NULL,
525 		    MTX_DEF | MTX_DUPOK);
526 		V_ipq[i].count = 0;
527 	}
528 	V_ipq_hashseed = arc4random();
529 	V_maxfragsperpacket = 16;
530 	V_ipq_zone = uma_zcreate("ipq", sizeof(struct ipq), NULL, NULL, NULL,
531 	    NULL, UMA_ALIGN_PTR, 0);
532 	max = IP_MAXFRAGPACKETS;
533 	max = uma_zone_set_max(V_ipq_zone, max);
534 	V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
535 
536 	if (IS_DEFAULT_VNET(curvnet)) {
537 		maxfrags = IP_MAXFRAGS;
538 		EVENTHANDLER_REGISTER(nmbclusters_change, ipreass_zone_change,
539 		    NULL, EVENTHANDLER_PRI_ANY);
540 	}
541 }
542 
543 /*
544  * If a timer expires on a reassembly queue, discard it.
545  */
546 void
547 ipreass_slowtimo(void)
548 {
549 	struct ipq *fp, *tmp;
550 
551 	for (int i = 0; i < IPREASS_NHASH; i++) {
552 		IPQ_LOCK(i);
553 		TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, tmp)
554 		if (--fp->ipq_ttl == 0)
555 				ipq_timeout(&V_ipq[i], fp);
556 		IPQ_UNLOCK(i);
557 	}
558 }
559 
560 /*
561  * Drain off all datagram fragments.
562  */
563 void
564 ipreass_drain(void)
565 {
566 
567 	for (int i = 0; i < IPREASS_NHASH; i++) {
568 		IPQ_LOCK(i);
569 		while(!TAILQ_EMPTY(&V_ipq[i].head))
570 			ipq_drop(&V_ipq[i], TAILQ_FIRST(&V_ipq[i].head));
571 		KASSERT(V_ipq[i].count == 0,
572 		    ("%s: V_ipq[%d] count %d (V_ipq=%p)", __func__, i,
573 		    V_ipq[i].count, V_ipq));
574 		IPQ_UNLOCK(i);
575 	}
576 }
577 
578 #ifdef VIMAGE
579 /*
580  * Destroy IP reassembly structures.
581  */
582 void
583 ipreass_destroy(void)
584 {
585 
586 	ipreass_drain();
587 	uma_zdestroy(V_ipq_zone);
588 	for (int i = 0; i < IPREASS_NHASH; i++)
589 		mtx_destroy(&V_ipq[i].lock);
590 }
591 #endif
592 
593 /*
594  * After maxnipq has been updated, propagate the change to UMA.  The UMA zone
595  * max has slightly different semantics than the sysctl, for historical
596  * reasons.
597  */
598 static void
599 ipreass_drain_tomax(void)
600 {
601 	struct ipq *fp;
602 	int target;
603 
604 	/*
605 	 * Make sure each bucket is under the new limit. If
606 	 * necessary, drop enough of the oldest elements from
607 	 * each bucket to get under the new limit.
608 	 */
609 	for (int i = 0; i < IPREASS_NHASH; i++) {
610 		IPQ_LOCK(i);
611 		while (V_ipq[i].count > V_ipreass_maxbucketsize &&
612 		    (fp = TAILQ_LAST(&V_ipq[i].head, ipqhead)) != NULL)
613 			ipq_timeout(&V_ipq[i], fp);
614 		IPQ_UNLOCK(i);
615 	}
616 
617 	/*
618 	 * If we are over the maximum number of fragments,
619 	 * drain off enough to get down to the new limit,
620 	 * stripping off last elements on queues.  Every
621 	 * run we strip the oldest element from each bucket.
622 	 */
623 	target = uma_zone_get_max(V_ipq_zone);
624 	while (uma_zone_get_cur(V_ipq_zone) > target) {
625 		for (int i = 0; i < IPREASS_NHASH; i++) {
626 			IPQ_LOCK(i);
627 			fp = TAILQ_LAST(&V_ipq[i].head, ipqhead);
628 			if (fp != NULL)
629 				ipq_timeout(&V_ipq[i], fp);
630 			IPQ_UNLOCK(i);
631 		}
632 	}
633 }
634 
635 static void
636 ipreass_zone_change(void *tag)
637 {
638 	VNET_ITERATOR_DECL(vnet_iter);
639 	int max;
640 
641 	maxfrags = IP_MAXFRAGS;
642 	max = IP_MAXFRAGPACKETS;
643 	VNET_LIST_RLOCK_NOSLEEP();
644 	VNET_FOREACH(vnet_iter) {
645 		CURVNET_SET(vnet_iter);
646 		max = uma_zone_set_max(V_ipq_zone, max);
647 		V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
648 		ipreass_drain_tomax();
649 		CURVNET_RESTORE();
650 	}
651 	VNET_LIST_RUNLOCK_NOSLEEP();
652 }
653 
654 /*
655  * Change the limit on the UMA zone, or disable the fragment allocation
656  * at all.  Since 0 and -1 is a special values here, we need our own handler,
657  * instead of sysctl_handle_uma_zone_max().
658  */
659 static int
660 sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS)
661 {
662 	int error, max;
663 
664 	if (V_noreass == 0) {
665 		max = uma_zone_get_max(V_ipq_zone);
666 		if (max == 0)
667 			max = -1;
668 	} else
669 		max = 0;
670 	error = sysctl_handle_int(oidp, &max, 0, req);
671 	if (error || !req->newptr)
672 		return (error);
673 	if (max > 0) {
674 		/*
675 		 * XXXRW: Might be a good idea to sanity check the argument
676 		 * and place an extreme upper bound.
677 		 */
678 		max = uma_zone_set_max(V_ipq_zone, max);
679 		V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
680 		ipreass_drain_tomax();
681 		V_noreass = 0;
682 	} else if (max == 0) {
683 		V_noreass = 1;
684 		ipreass_drain();
685 	} else if (max == -1) {
686 		V_noreass = 0;
687 		uma_zone_set_max(V_ipq_zone, 0);
688 		V_ipreass_maxbucketsize = INT_MAX;
689 	} else
690 		return (EINVAL);
691 	return (0);
692 }
693 
694 /*
695  * Seek for old fragment queue header that can be reused.  Try to
696  * reuse a header from currently locked hash bucket.
697  */
698 static struct ipq *
699 ipq_reuse(int start)
700 {
701 	struct ipq *fp;
702 	int bucket, i;
703 
704 	IPQ_LOCK_ASSERT(start);
705 
706 	for (i = 0; i < IPREASS_NHASH; i++) {
707 		bucket = (start + i) % IPREASS_NHASH;
708 		if (bucket != start && IPQ_TRYLOCK(bucket) == 0)
709 			continue;
710 		fp = TAILQ_LAST(&V_ipq[bucket].head, ipqhead);
711 		if (fp) {
712 			struct mbuf *m;
713 
714 			IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
715 			atomic_subtract_int(&nfrags, fp->ipq_nfrags);
716 			while (fp->ipq_frags) {
717 				m = fp->ipq_frags;
718 				fp->ipq_frags = m->m_nextpkt;
719 				m_freem(m);
720 			}
721 			TAILQ_REMOVE(&V_ipq[bucket].head, fp, ipq_list);
722 			V_ipq[bucket].count--;
723 			if (bucket != start)
724 				IPQ_UNLOCK(bucket);
725 			break;
726 		}
727 		if (bucket != start)
728 			IPQ_UNLOCK(bucket);
729 	}
730 	IPQ_LOCK_ASSERT(start);
731 	return (fp);
732 }
733 
734 /*
735  * Free a fragment reassembly header and all associated datagrams.
736  */
737 static void
738 ipq_free(struct ipqbucket *bucket, struct ipq *fp)
739 {
740 	struct mbuf *q;
741 
742 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
743 	while (fp->ipq_frags) {
744 		q = fp->ipq_frags;
745 		fp->ipq_frags = q->m_nextpkt;
746 		m_freem(q);
747 	}
748 	TAILQ_REMOVE(&bucket->head, fp, ipq_list);
749 	bucket->count--;
750 	uma_zfree(V_ipq_zone, fp);
751 }
752 
753 /*
754  * Get or set the maximum number of reassembly queues per bucket.
755  */
756 static int
757 sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS)
758 {
759 	int error, max;
760 
761 	max = V_ipreass_maxbucketsize;
762 	error = sysctl_handle_int(oidp, &max, 0, req);
763 	if (error || !req->newptr)
764 		return (error);
765 	if (max <= 0)
766 		return (EINVAL);
767 	V_ipreass_maxbucketsize = max;
768 	ipreass_drain_tomax();
769 	return (0);
770 }
771