xref: /freebsd/sys/netinet/ip_reass.c (revision fe6060f10f634930ff71b7c50291ddc610da2475)
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/kernel.h>
43 #include <sys/hash.h>
44 #include <sys/mbuf.h>
45 #include <sys/malloc.h>
46 #include <sys/limits.h>
47 #include <sys/lock.h>
48 #include <sys/mutex.h>
49 #include <sys/sysctl.h>
50 #include <sys/socket.h>
51 
52 #include <net/if.h>
53 #include <net/if_var.h>
54 #include <net/rss_config.h>
55 #include <net/netisr.h>
56 #include <net/vnet.h>
57 
58 #include <netinet/in.h>
59 #include <netinet/ip.h>
60 #include <netinet/ip_var.h>
61 #include <netinet/in_rss.h>
62 #ifdef MAC
63 #include <security/mac/mac_framework.h>
64 #endif
65 
66 SYSCTL_DECL(_net_inet_ip);
67 
68 /*
69  * Reassembly headers are stored in hash buckets.
70  */
71 #define	IPREASS_NHASH_LOG2	10
72 #define	IPREASS_NHASH		(1 << IPREASS_NHASH_LOG2)
73 #define	IPREASS_HMASK		(IPREASS_NHASH - 1)
74 
75 struct ipqbucket {
76 	TAILQ_HEAD(ipqhead, ipq) head;
77 	struct mtx		 lock;
78 	int			 count;
79 };
80 
81 VNET_DEFINE_STATIC(struct ipqbucket, ipq[IPREASS_NHASH]);
82 #define	V_ipq		VNET(ipq)
83 VNET_DEFINE_STATIC(uint32_t, ipq_hashseed);
84 #define V_ipq_hashseed   VNET(ipq_hashseed)
85 
86 #define	IPQ_LOCK(i)	mtx_lock(&V_ipq[i].lock)
87 #define	IPQ_TRYLOCK(i)	mtx_trylock(&V_ipq[i].lock)
88 #define	IPQ_UNLOCK(i)	mtx_unlock(&V_ipq[i].lock)
89 #define	IPQ_LOCK_ASSERT(i)	mtx_assert(&V_ipq[i].lock, MA_OWNED)
90 
91 VNET_DEFINE_STATIC(int, ipreass_maxbucketsize);
92 #define	V_ipreass_maxbucketsize	VNET(ipreass_maxbucketsize)
93 
94 void		ipreass_init(void);
95 void		ipreass_drain(void);
96 void		ipreass_slowtimo(void);
97 #ifdef VIMAGE
98 void		ipreass_destroy(void);
99 #endif
100 static int	sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS);
101 static int	sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS);
102 static void	ipreass_zone_change(void *);
103 static void	ipreass_drain_tomax(void);
104 static void	ipq_free(struct ipqbucket *, struct ipq *);
105 static struct ipq * ipq_reuse(int);
106 
107 static inline void
108 ipq_timeout(struct ipqbucket *bucket, struct ipq *fp)
109 {
110 
111 	IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
112 	ipq_free(bucket, fp);
113 }
114 
115 static inline void
116 ipq_drop(struct ipqbucket *bucket, struct ipq *fp)
117 {
118 
119 	IPSTAT_ADD(ips_fragdropped, fp->ipq_nfrags);
120 	ipq_free(bucket, fp);
121 }
122 
123 /*
124  * By default, limit the number of IP fragments across all reassembly
125  * queues to  1/32 of the total number of mbuf clusters.
126  *
127  * Limit the total number of reassembly queues per VNET to the
128  * IP fragment limit, but ensure the limit will not allow any bucket
129  * to grow above 100 items. (The bucket limit is
130  * IP_MAXFRAGPACKETS / (IPREASS_NHASH / 2), so the 50 is the correct
131  * multiplier to reach a 100-item limit.)
132  * The 100-item limit was chosen as brief testing seems to show that
133  * this produces "reasonable" performance on some subset of systems
134  * under DoS attack.
135  */
136 #define	IP_MAXFRAGS		(nmbclusters / 32)
137 #define	IP_MAXFRAGPACKETS	(imin(IP_MAXFRAGS, IPREASS_NHASH * 50))
138 
139 static int		maxfrags;
140 static u_int __exclusive_cache_line	nfrags;
141 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfrags, CTLFLAG_RW,
142     &maxfrags, 0,
143     "Maximum number of IPv4 fragments allowed across all reassembly queues");
144 SYSCTL_UINT(_net_inet_ip, OID_AUTO, curfrags, CTLFLAG_RD,
145     &nfrags, 0,
146     "Current number of IPv4 fragments across all reassembly queues");
147 
148 VNET_DEFINE_STATIC(uma_zone_t, ipq_zone);
149 #define	V_ipq_zone	VNET(ipq_zone)
150 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragpackets,
151     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
152     NULL, 0, sysctl_maxfragpackets, "I",
153     "Maximum number of IPv4 fragment reassembly queue entries");
154 SYSCTL_UMA_CUR(_net_inet_ip, OID_AUTO, fragpackets, CTLFLAG_VNET,
155     &VNET_NAME(ipq_zone),
156     "Current number of IPv4 fragment reassembly queue entries");
157 
158 VNET_DEFINE_STATIC(int, noreass);
159 #define	V_noreass	VNET(noreass)
160 
161 VNET_DEFINE_STATIC(int, maxfragsperpacket);
162 #define	V_maxfragsperpacket	VNET(maxfragsperpacket)
163 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfragsperpacket, CTLFLAG_VNET | CTLFLAG_RW,
164     &VNET_NAME(maxfragsperpacket), 0,
165     "Maximum number of IPv4 fragments allowed per packet");
166 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragbucketsize,
167     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0,
168     sysctl_maxfragbucketsize, "I",
169     "Maximum number of IPv4 fragment reassembly queue entries per bucket");
170 
171 /*
172  * Take incoming datagram fragment and try to reassemble it into
173  * whole datagram.  If the argument is the first fragment or one
174  * in between the function will return NULL and store the mbuf
175  * in the fragment chain.  If the argument is the last fragment
176  * the packet will be reassembled and the pointer to the new
177  * mbuf returned for further processing.  Only m_tags attached
178  * to the first packet/fragment are preserved.
179  * The IP header is *NOT* adjusted out of iplen.
180  */
181 #define	M_IP_FRAG	M_PROTO9
182 struct mbuf *
183 ip_reass(struct mbuf *m)
184 {
185 	struct ip *ip;
186 	struct mbuf *p, *q, *nq, *t;
187 	struct ipq *fp;
188 	struct ifnet *srcifp;
189 	struct ipqhead *head;
190 	int i, hlen, next, tmpmax;
191 	u_int8_t ecn, ecn0;
192 	uint32_t hash, hashkey[3];
193 #ifdef	RSS
194 	uint32_t rss_hash, rss_type;
195 #endif
196 
197 	/*
198 	 * If no reassembling or maxfragsperpacket are 0,
199 	 * never accept fragments.
200 	 * Also, drop packet if it would exceed the maximum
201 	 * number of fragments.
202 	 */
203 	tmpmax = maxfrags;
204 	if (V_noreass == 1 || V_maxfragsperpacket == 0 ||
205 	    (tmpmax >= 0 && atomic_load_int(&nfrags) >= (u_int)tmpmax)) {
206 		IPSTAT_INC(ips_fragments);
207 		IPSTAT_INC(ips_fragdropped);
208 		m_freem(m);
209 		return (NULL);
210 	}
211 
212 	ip = mtod(m, struct ip *);
213 	hlen = ip->ip_hl << 2;
214 
215 	/*
216 	 * Adjust ip_len to not reflect header,
217 	 * convert offset of this to bytes.
218 	 */
219 	ip->ip_len = htons(ntohs(ip->ip_len) - hlen);
220 	/*
221 	 * Make sure that fragments have a data length
222 	 * that's a non-zero multiple of 8 bytes, unless
223 	 * this is the last fragment.
224 	 */
225 	if (ip->ip_len == htons(0) ||
226 	    ((ip->ip_off & htons(IP_MF)) && (ntohs(ip->ip_len) & 0x7) != 0)) {
227 		IPSTAT_INC(ips_toosmall); /* XXX */
228 		IPSTAT_INC(ips_fragdropped);
229 		m_freem(m);
230 		return (NULL);
231 	}
232 	if (ip->ip_off & htons(IP_MF))
233 		m->m_flags |= M_IP_FRAG;
234 	else
235 		m->m_flags &= ~M_IP_FRAG;
236 	ip->ip_off = htons(ntohs(ip->ip_off) << 3);
237 
238 	/*
239 	 * Make sure the fragment lies within a packet of valid size.
240 	 */
241 	if (ntohs(ip->ip_len) + ntohs(ip->ip_off) > IP_MAXPACKET) {
242 		IPSTAT_INC(ips_toolong);
243 		IPSTAT_INC(ips_fragdropped);
244 		m_freem(m);
245 		return (NULL);
246 	}
247 
248 	/*
249 	 * Store receive network interface pointer for later.
250 	 */
251 	srcifp = m->m_pkthdr.rcvif;
252 
253 	/*
254 	 * Attempt reassembly; if it succeeds, proceed.
255 	 * ip_reass() will return a different mbuf.
256 	 */
257 	IPSTAT_INC(ips_fragments);
258 	m->m_pkthdr.PH_loc.ptr = ip;
259 
260 	/*
261 	 * Presence of header sizes in mbufs
262 	 * would confuse code below.
263 	 */
264 	m->m_data += hlen;
265 	m->m_len -= hlen;
266 
267 	hashkey[0] = ip->ip_src.s_addr;
268 	hashkey[1] = ip->ip_dst.s_addr;
269 	hashkey[2] = (uint32_t)ip->ip_p << 16;
270 	hashkey[2] += ip->ip_id;
271 	hash = jenkins_hash32(hashkey, nitems(hashkey), V_ipq_hashseed);
272 	hash &= IPREASS_HMASK;
273 	head = &V_ipq[hash].head;
274 	IPQ_LOCK(hash);
275 
276 	/*
277 	 * Look for queue of fragments
278 	 * of this datagram.
279 	 */
280 	TAILQ_FOREACH(fp, head, ipq_list)
281 		if (ip->ip_id == fp->ipq_id &&
282 		    ip->ip_src.s_addr == fp->ipq_src.s_addr &&
283 		    ip->ip_dst.s_addr == fp->ipq_dst.s_addr &&
284 #ifdef MAC
285 		    mac_ipq_match(m, fp) &&
286 #endif
287 		    ip->ip_p == fp->ipq_p)
288 			break;
289 	/*
290 	 * If first fragment to arrive, create a reassembly queue.
291 	 */
292 	if (fp == NULL) {
293 		if (V_ipq[hash].count < V_ipreass_maxbucketsize)
294 			fp = uma_zalloc(V_ipq_zone, M_NOWAIT);
295 		if (fp == NULL)
296 			fp = ipq_reuse(hash);
297 		if (fp == NULL)
298 			goto dropfrag;
299 #ifdef MAC
300 		if (mac_ipq_init(fp, M_NOWAIT) != 0) {
301 			uma_zfree(V_ipq_zone, fp);
302 			fp = NULL;
303 			goto dropfrag;
304 		}
305 		mac_ipq_create(m, fp);
306 #endif
307 		TAILQ_INSERT_HEAD(head, fp, ipq_list);
308 		V_ipq[hash].count++;
309 		fp->ipq_nfrags = 1;
310 		atomic_add_int(&nfrags, 1);
311 		fp->ipq_ttl = IPFRAGTTL;
312 		fp->ipq_p = ip->ip_p;
313 		fp->ipq_id = ip->ip_id;
314 		fp->ipq_src = ip->ip_src;
315 		fp->ipq_dst = ip->ip_dst;
316 		fp->ipq_frags = m;
317 		if (m->m_flags & M_IP_FRAG)
318 			fp->ipq_maxoff = -1;
319 		else
320 			fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len);
321 		m->m_nextpkt = NULL;
322 		goto done;
323 	} else {
324 		/*
325 		 * If we already saw the last fragment, make sure
326 		 * this fragment's offset looks sane. Otherwise, if
327 		 * this is the last fragment, record its endpoint.
328 		 */
329 		if (fp->ipq_maxoff > 0) {
330 			i = ntohs(ip->ip_off) + ntohs(ip->ip_len);
331 			if (((m->m_flags & M_IP_FRAG) && i >= fp->ipq_maxoff) ||
332 			    ((m->m_flags & M_IP_FRAG) == 0 &&
333 			    i != fp->ipq_maxoff)) {
334 				fp = NULL;
335 				goto dropfrag;
336 			}
337 		} else if ((m->m_flags & M_IP_FRAG) == 0)
338 			fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len);
339 		fp->ipq_nfrags++;
340 		atomic_add_int(&nfrags, 1);
341 #ifdef MAC
342 		mac_ipq_update(m, fp);
343 #endif
344 	}
345 
346 #define GETIP(m)	((struct ip*)((m)->m_pkthdr.PH_loc.ptr))
347 
348 	/*
349 	 * Handle ECN by comparing this segment with the first one;
350 	 * if CE is set, do not lose CE.
351 	 * drop if CE and not-ECT are mixed for the same packet.
352 	 */
353 	ecn = ip->ip_tos & IPTOS_ECN_MASK;
354 	ecn0 = GETIP(fp->ipq_frags)->ip_tos & IPTOS_ECN_MASK;
355 	if (ecn == IPTOS_ECN_CE) {
356 		if (ecn0 == IPTOS_ECN_NOTECT)
357 			goto dropfrag;
358 		if (ecn0 != IPTOS_ECN_CE)
359 			GETIP(fp->ipq_frags)->ip_tos |= IPTOS_ECN_CE;
360 	}
361 	if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT)
362 		goto dropfrag;
363 
364 	/*
365 	 * Find a segment which begins after this one does.
366 	 */
367 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt)
368 		if (ntohs(GETIP(q)->ip_off) > ntohs(ip->ip_off))
369 			break;
370 
371 	/*
372 	 * If there is a preceding segment, it may provide some of
373 	 * our data already.  If so, drop the data from the incoming
374 	 * segment.  If it provides all of our data, drop us, otherwise
375 	 * stick new segment in the proper place.
376 	 *
377 	 * If some of the data is dropped from the preceding
378 	 * segment, then it's checksum is invalidated.
379 	 */
380 	if (p) {
381 		i = ntohs(GETIP(p)->ip_off) + ntohs(GETIP(p)->ip_len) -
382 		    ntohs(ip->ip_off);
383 		if (i > 0) {
384 			if (i >= ntohs(ip->ip_len))
385 				goto dropfrag;
386 			m_adj(m, i);
387 			m->m_pkthdr.csum_flags = 0;
388 			ip->ip_off = htons(ntohs(ip->ip_off) + i);
389 			ip->ip_len = htons(ntohs(ip->ip_len) - i);
390 		}
391 		m->m_nextpkt = p->m_nextpkt;
392 		p->m_nextpkt = m;
393 	} else {
394 		m->m_nextpkt = fp->ipq_frags;
395 		fp->ipq_frags = m;
396 	}
397 
398 	/*
399 	 * While we overlap succeeding segments trim them or,
400 	 * if they are completely covered, dequeue them.
401 	 */
402 	for (; q != NULL && ntohs(ip->ip_off) + ntohs(ip->ip_len) >
403 	    ntohs(GETIP(q)->ip_off); q = nq) {
404 		i = (ntohs(ip->ip_off) + ntohs(ip->ip_len)) -
405 		    ntohs(GETIP(q)->ip_off);
406 		if (i < ntohs(GETIP(q)->ip_len)) {
407 			GETIP(q)->ip_len = htons(ntohs(GETIP(q)->ip_len) - i);
408 			GETIP(q)->ip_off = htons(ntohs(GETIP(q)->ip_off) + i);
409 			m_adj(q, i);
410 			q->m_pkthdr.csum_flags = 0;
411 			break;
412 		}
413 		nq = q->m_nextpkt;
414 		m->m_nextpkt = nq;
415 		IPSTAT_INC(ips_fragdropped);
416 		fp->ipq_nfrags--;
417 		atomic_subtract_int(&nfrags, 1);
418 		m_freem(q);
419 	}
420 
421 	/*
422 	 * Check for complete reassembly and perform frag per packet
423 	 * limiting.
424 	 *
425 	 * Frag limiting is performed here so that the nth frag has
426 	 * a chance to complete the packet before we drop the packet.
427 	 * As a result, n+1 frags are actually allowed per packet, but
428 	 * only n will ever be stored. (n = maxfragsperpacket.)
429 	 *
430 	 */
431 	next = 0;
432 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt) {
433 		if (ntohs(GETIP(q)->ip_off) != next) {
434 			if (fp->ipq_nfrags > V_maxfragsperpacket)
435 				ipq_drop(&V_ipq[hash], fp);
436 			goto done;
437 		}
438 		next += ntohs(GETIP(q)->ip_len);
439 	}
440 	/* Make sure the last packet didn't have the IP_MF flag */
441 	if (p->m_flags & M_IP_FRAG) {
442 		if (fp->ipq_nfrags > V_maxfragsperpacket)
443 			ipq_drop(&V_ipq[hash], fp);
444 		goto done;
445 	}
446 
447 	/*
448 	 * Reassembly is complete.  Make sure the packet is a sane size.
449 	 */
450 	q = fp->ipq_frags;
451 	ip = GETIP(q);
452 	if (next + (ip->ip_hl << 2) > IP_MAXPACKET) {
453 		IPSTAT_INC(ips_toolong);
454 		ipq_drop(&V_ipq[hash], fp);
455 		goto done;
456 	}
457 
458 	/*
459 	 * Concatenate fragments.
460 	 */
461 	m = q;
462 	t = m->m_next;
463 	m->m_next = NULL;
464 	m_cat(m, t);
465 	nq = q->m_nextpkt;
466 	q->m_nextpkt = NULL;
467 	for (q = nq; q != NULL; q = nq) {
468 		nq = q->m_nextpkt;
469 		q->m_nextpkt = NULL;
470 		m->m_pkthdr.csum_flags &= q->m_pkthdr.csum_flags;
471 		m->m_pkthdr.csum_data += q->m_pkthdr.csum_data;
472 		m_demote_pkthdr(q);
473 		m_cat(m, q);
474 	}
475 	/*
476 	 * In order to do checksumming faster we do 'end-around carry' here
477 	 * (and not in for{} loop), though it implies we are not going to
478 	 * reassemble more than 64k fragments.
479 	 */
480 	while (m->m_pkthdr.csum_data & 0xffff0000)
481 		m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
482 		    (m->m_pkthdr.csum_data >> 16);
483 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
484 #ifdef MAC
485 	mac_ipq_reassemble(fp, m);
486 	mac_ipq_destroy(fp);
487 #endif
488 
489 	/*
490 	 * Create header for new ip packet by modifying header of first
491 	 * packet;  dequeue and discard fragment reassembly header.
492 	 * Make header visible.
493 	 */
494 	ip->ip_len = htons((ip->ip_hl << 2) + next);
495 	ip->ip_src = fp->ipq_src;
496 	ip->ip_dst = fp->ipq_dst;
497 	TAILQ_REMOVE(head, fp, ipq_list);
498 	V_ipq[hash].count--;
499 	uma_zfree(V_ipq_zone, fp);
500 	m->m_len += (ip->ip_hl << 2);
501 	m->m_data -= (ip->ip_hl << 2);
502 	/* some debugging cruft by sklower, below, will go away soon */
503 	if (m->m_flags & M_PKTHDR) {	/* XXX this should be done elsewhere */
504 		m_fixhdr(m);
505 		/* set valid receive interface pointer */
506 		m->m_pkthdr.rcvif = srcifp;
507 	}
508 	IPSTAT_INC(ips_reassembled);
509 	IPQ_UNLOCK(hash);
510 
511 #ifdef	RSS
512 	/*
513 	 * Query the RSS layer for the flowid / flowtype for the
514 	 * mbuf payload.
515 	 *
516 	 * For now, just assume we have to calculate a new one.
517 	 * Later on we should check to see if the assigned flowid matches
518 	 * what RSS wants for the given IP protocol and if so, just keep it.
519 	 *
520 	 * We then queue into the relevant netisr so it can be dispatched
521 	 * to the correct CPU.
522 	 *
523 	 * Note - this may return 1, which means the flowid in the mbuf
524 	 * is correct for the configured RSS hash types and can be used.
525 	 */
526 	if (rss_mbuf_software_hash_v4(m, 0, &rss_hash, &rss_type) == 0) {
527 		m->m_pkthdr.flowid = rss_hash;
528 		M_HASHTYPE_SET(m, rss_type);
529 	}
530 
531 	/*
532 	 * Queue/dispatch for reprocessing.
533 	 *
534 	 * Note: this is much slower than just handling the frame in the
535 	 * current receive context.  It's likely worth investigating
536 	 * why this is.
537 	 */
538 	netisr_dispatch(NETISR_IP_DIRECT, m);
539 	return (NULL);
540 #endif
541 
542 	/* Handle in-line */
543 	return (m);
544 
545 dropfrag:
546 	IPSTAT_INC(ips_fragdropped);
547 	if (fp != NULL) {
548 		fp->ipq_nfrags--;
549 		atomic_subtract_int(&nfrags, 1);
550 	}
551 	m_freem(m);
552 done:
553 	IPQ_UNLOCK(hash);
554 	return (NULL);
555 
556 #undef GETIP
557 }
558 
559 /*
560  * Initialize IP reassembly structures.
561  */
562 void
563 ipreass_init(void)
564 {
565 	int max;
566 
567 	for (int i = 0; i < IPREASS_NHASH; i++) {
568 		TAILQ_INIT(&V_ipq[i].head);
569 		mtx_init(&V_ipq[i].lock, "IP reassembly", NULL,
570 		    MTX_DEF | MTX_DUPOK);
571 		V_ipq[i].count = 0;
572 	}
573 	V_ipq_hashseed = arc4random();
574 	V_maxfragsperpacket = 16;
575 	V_ipq_zone = uma_zcreate("ipq", sizeof(struct ipq), NULL, NULL, NULL,
576 	    NULL, UMA_ALIGN_PTR, 0);
577 	max = IP_MAXFRAGPACKETS;
578 	max = uma_zone_set_max(V_ipq_zone, max);
579 	V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
580 
581 	if (IS_DEFAULT_VNET(curvnet)) {
582 		maxfrags = IP_MAXFRAGS;
583 		EVENTHANDLER_REGISTER(nmbclusters_change, ipreass_zone_change,
584 		    NULL, EVENTHANDLER_PRI_ANY);
585 	}
586 }
587 
588 /*
589  * If a timer expires on a reassembly queue, discard it.
590  */
591 void
592 ipreass_slowtimo(void)
593 {
594 	struct ipq *fp, *tmp;
595 
596 	if (atomic_load_int(&nfrags) == 0)
597 		return;
598 
599 	for (int i = 0; i < IPREASS_NHASH; i++) {
600 		if (TAILQ_EMPTY(&V_ipq[i].head))
601 			continue;
602 		IPQ_LOCK(i);
603 		TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, tmp)
604 		if (--fp->ipq_ttl == 0)
605 			ipq_timeout(&V_ipq[i], fp);
606 		IPQ_UNLOCK(i);
607 	}
608 }
609 
610 /*
611  * Drain off all datagram fragments.
612  */
613 void
614 ipreass_drain(void)
615 {
616 
617 	for (int i = 0; i < IPREASS_NHASH; i++) {
618 		IPQ_LOCK(i);
619 		while(!TAILQ_EMPTY(&V_ipq[i].head))
620 			ipq_drop(&V_ipq[i], TAILQ_FIRST(&V_ipq[i].head));
621 		KASSERT(V_ipq[i].count == 0,
622 		    ("%s: V_ipq[%d] count %d (V_ipq=%p)", __func__, i,
623 		    V_ipq[i].count, V_ipq));
624 		IPQ_UNLOCK(i);
625 	}
626 }
627 
628 /*
629  * Drain off all datagram fragments belonging to
630  * the given network interface.
631  */
632 static void
633 ipreass_cleanup(void *arg __unused, struct ifnet *ifp)
634 {
635 	struct ipq *fp, *temp;
636 	struct mbuf *m;
637 	int i;
638 
639 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
640 
641 	CURVNET_SET_QUIET(ifp->if_vnet);
642 
643 	/*
644 	 * Skip processing if IPv4 reassembly is not initialised or
645 	 * torn down by ipreass_destroy().
646 	 */
647 	if (V_ipq_zone == NULL) {
648 		CURVNET_RESTORE();
649 		return;
650 	}
651 
652 	for (i = 0; i < IPREASS_NHASH; i++) {
653 		IPQ_LOCK(i);
654 		/* Scan fragment list. */
655 		TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, temp) {
656 			for (m = fp->ipq_frags; m != NULL; m = m->m_nextpkt) {
657 				/* clear no longer valid rcvif pointer */
658 				if (m->m_pkthdr.rcvif == ifp)
659 					m->m_pkthdr.rcvif = NULL;
660 			}
661 		}
662 		IPQ_UNLOCK(i);
663 	}
664 	CURVNET_RESTORE();
665 }
666 EVENTHANDLER_DEFINE(ifnet_departure_event, ipreass_cleanup, NULL, 0);
667 
668 #ifdef VIMAGE
669 /*
670  * Destroy IP reassembly structures.
671  */
672 void
673 ipreass_destroy(void)
674 {
675 
676 	ipreass_drain();
677 	uma_zdestroy(V_ipq_zone);
678 	V_ipq_zone = NULL;
679 	for (int i = 0; i < IPREASS_NHASH; i++)
680 		mtx_destroy(&V_ipq[i].lock);
681 }
682 #endif
683 
684 /*
685  * After maxnipq has been updated, propagate the change to UMA.  The UMA zone
686  * max has slightly different semantics than the sysctl, for historical
687  * reasons.
688  */
689 static void
690 ipreass_drain_tomax(void)
691 {
692 	struct ipq *fp;
693 	int target;
694 
695 	/*
696 	 * Make sure each bucket is under the new limit. If
697 	 * necessary, drop enough of the oldest elements from
698 	 * each bucket to get under the new limit.
699 	 */
700 	for (int i = 0; i < IPREASS_NHASH; i++) {
701 		IPQ_LOCK(i);
702 		while (V_ipq[i].count > V_ipreass_maxbucketsize &&
703 		    (fp = TAILQ_LAST(&V_ipq[i].head, ipqhead)) != NULL)
704 			ipq_timeout(&V_ipq[i], fp);
705 		IPQ_UNLOCK(i);
706 	}
707 
708 	/*
709 	 * If we are over the maximum number of fragments,
710 	 * drain off enough to get down to the new limit,
711 	 * stripping off last elements on queues.  Every
712 	 * run we strip the oldest element from each bucket.
713 	 */
714 	target = uma_zone_get_max(V_ipq_zone);
715 	while (uma_zone_get_cur(V_ipq_zone) > target) {
716 		for (int i = 0; i < IPREASS_NHASH; i++) {
717 			IPQ_LOCK(i);
718 			fp = TAILQ_LAST(&V_ipq[i].head, ipqhead);
719 			if (fp != NULL)
720 				ipq_timeout(&V_ipq[i], fp);
721 			IPQ_UNLOCK(i);
722 		}
723 	}
724 }
725 
726 static void
727 ipreass_zone_change(void *tag)
728 {
729 	VNET_ITERATOR_DECL(vnet_iter);
730 	int max;
731 
732 	maxfrags = IP_MAXFRAGS;
733 	max = IP_MAXFRAGPACKETS;
734 	VNET_LIST_RLOCK_NOSLEEP();
735 	VNET_FOREACH(vnet_iter) {
736 		CURVNET_SET(vnet_iter);
737 		max = uma_zone_set_max(V_ipq_zone, max);
738 		V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
739 		ipreass_drain_tomax();
740 		CURVNET_RESTORE();
741 	}
742 	VNET_LIST_RUNLOCK_NOSLEEP();
743 }
744 
745 /*
746  * Change the limit on the UMA zone, or disable the fragment allocation
747  * at all.  Since 0 and -1 is a special values here, we need our own handler,
748  * instead of sysctl_handle_uma_zone_max().
749  */
750 static int
751 sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS)
752 {
753 	int error, max;
754 
755 	if (V_noreass == 0) {
756 		max = uma_zone_get_max(V_ipq_zone);
757 		if (max == 0)
758 			max = -1;
759 	} else
760 		max = 0;
761 	error = sysctl_handle_int(oidp, &max, 0, req);
762 	if (error || !req->newptr)
763 		return (error);
764 	if (max > 0) {
765 		/*
766 		 * XXXRW: Might be a good idea to sanity check the argument
767 		 * and place an extreme upper bound.
768 		 */
769 		max = uma_zone_set_max(V_ipq_zone, max);
770 		V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
771 		ipreass_drain_tomax();
772 		V_noreass = 0;
773 	} else if (max == 0) {
774 		V_noreass = 1;
775 		ipreass_drain();
776 	} else if (max == -1) {
777 		V_noreass = 0;
778 		uma_zone_set_max(V_ipq_zone, 0);
779 		V_ipreass_maxbucketsize = INT_MAX;
780 	} else
781 		return (EINVAL);
782 	return (0);
783 }
784 
785 /*
786  * Seek for old fragment queue header that can be reused.  Try to
787  * reuse a header from currently locked hash bucket.
788  */
789 static struct ipq *
790 ipq_reuse(int start)
791 {
792 	struct ipq *fp;
793 	int bucket, i;
794 
795 	IPQ_LOCK_ASSERT(start);
796 
797 	for (i = 0; i < IPREASS_NHASH; i++) {
798 		bucket = (start + i) % IPREASS_NHASH;
799 		if (bucket != start && IPQ_TRYLOCK(bucket) == 0)
800 			continue;
801 		fp = TAILQ_LAST(&V_ipq[bucket].head, ipqhead);
802 		if (fp) {
803 			struct mbuf *m;
804 
805 			IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
806 			atomic_subtract_int(&nfrags, fp->ipq_nfrags);
807 			while (fp->ipq_frags) {
808 				m = fp->ipq_frags;
809 				fp->ipq_frags = m->m_nextpkt;
810 				m_freem(m);
811 			}
812 			TAILQ_REMOVE(&V_ipq[bucket].head, fp, ipq_list);
813 			V_ipq[bucket].count--;
814 			if (bucket != start)
815 				IPQ_UNLOCK(bucket);
816 			break;
817 		}
818 		if (bucket != start)
819 			IPQ_UNLOCK(bucket);
820 	}
821 	IPQ_LOCK_ASSERT(start);
822 	return (fp);
823 }
824 
825 /*
826  * Free a fragment reassembly header and all associated datagrams.
827  */
828 static void
829 ipq_free(struct ipqbucket *bucket, struct ipq *fp)
830 {
831 	struct mbuf *q;
832 
833 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
834 	while (fp->ipq_frags) {
835 		q = fp->ipq_frags;
836 		fp->ipq_frags = q->m_nextpkt;
837 		m_freem(q);
838 	}
839 	TAILQ_REMOVE(&bucket->head, fp, ipq_list);
840 	bucket->count--;
841 	uma_zfree(V_ipq_zone, fp);
842 }
843 
844 /*
845  * Get or set the maximum number of reassembly queues per bucket.
846  */
847 static int
848 sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS)
849 {
850 	int error, max;
851 
852 	max = V_ipreass_maxbucketsize;
853 	error = sysctl_handle_int(oidp, &max, 0, req);
854 	if (error || !req->newptr)
855 		return (error);
856 	if (max <= 0)
857 		return (EINVAL);
858 	V_ipreass_maxbucketsize = max;
859 	ipreass_drain_tomax();
860 	return (0);
861 }
862