xref: /freebsd/sys/netinet/ip_reass.c (revision 4c1a82cea504df7a79f5bd8f7d0a41cacccff16e)
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 
51 #include <net/rss_config.h>
52 #include <net/netisr.h>
53 #include <net/vnet.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/ip_var.h>
58 #include <netinet/in_rss.h>
59 #ifdef MAC
60 #include <security/mac/mac_framework.h>
61 #endif
62 
63 SYSCTL_DECL(_net_inet_ip);
64 
65 /*
66  * Reassembly headers are stored in hash buckets.
67  */
68 #define	IPREASS_NHASH_LOG2	10
69 #define	IPREASS_NHASH		(1 << IPREASS_NHASH_LOG2)
70 #define	IPREASS_HMASK		(IPREASS_NHASH - 1)
71 
72 struct ipqbucket {
73 	TAILQ_HEAD(ipqhead, ipq) head;
74 	struct mtx		 lock;
75 	int			 count;
76 };
77 
78 VNET_DEFINE_STATIC(struct ipqbucket, ipq[IPREASS_NHASH]);
79 #define	V_ipq		VNET(ipq)
80 VNET_DEFINE_STATIC(uint32_t, ipq_hashseed);
81 #define V_ipq_hashseed   VNET(ipq_hashseed)
82 
83 #define	IPQ_LOCK(i)	mtx_lock(&V_ipq[i].lock)
84 #define	IPQ_TRYLOCK(i)	mtx_trylock(&V_ipq[i].lock)
85 #define	IPQ_UNLOCK(i)	mtx_unlock(&V_ipq[i].lock)
86 #define	IPQ_LOCK_ASSERT(i)	mtx_assert(&V_ipq[i].lock, MA_OWNED)
87 
88 VNET_DEFINE_STATIC(int, ipreass_maxbucketsize);
89 #define	V_ipreass_maxbucketsize	VNET(ipreass_maxbucketsize)
90 
91 void		ipreass_init(void);
92 void		ipreass_drain(void);
93 void		ipreass_slowtimo(void);
94 #ifdef VIMAGE
95 void		ipreass_destroy(void);
96 #endif
97 static int	sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS);
98 static int	sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS);
99 static void	ipreass_zone_change(void *);
100 static void	ipreass_drain_tomax(void);
101 static void	ipq_free(struct ipqbucket *, struct ipq *);
102 static struct ipq * ipq_reuse(int);
103 
104 static inline void
105 ipq_timeout(struct ipqbucket *bucket, struct ipq *fp)
106 {
107 
108 	IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
109 	ipq_free(bucket, fp);
110 }
111 
112 static inline void
113 ipq_drop(struct ipqbucket *bucket, struct ipq *fp)
114 {
115 
116 	IPSTAT_ADD(ips_fragdropped, fp->ipq_nfrags);
117 	ipq_free(bucket, fp);
118 }
119 
120 /*
121  * By default, limit the number of IP fragments across all reassembly
122  * queues to  1/32 of the total number of mbuf clusters.
123  *
124  * Limit the total number of reassembly queues per VNET to the
125  * IP fragment limit, but ensure the limit will not allow any bucket
126  * to grow above 100 items. (The bucket limit is
127  * IP_MAXFRAGPACKETS / (IPREASS_NHASH / 2), so the 50 is the correct
128  * multiplier to reach a 100-item limit.)
129  * The 100-item limit was chosen as brief testing seems to show that
130  * this produces "reasonable" performance on some subset of systems
131  * under DoS attack.
132  */
133 #define	IP_MAXFRAGS		(nmbclusters / 32)
134 #define	IP_MAXFRAGPACKETS	(imin(IP_MAXFRAGS, IPREASS_NHASH * 50))
135 
136 static int		maxfrags;
137 static volatile u_int	nfrags;
138 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfrags, CTLFLAG_RW,
139     &maxfrags, 0,
140     "Maximum number of IPv4 fragments allowed across all reassembly queues");
141 SYSCTL_UINT(_net_inet_ip, OID_AUTO, curfrags, CTLFLAG_RD,
142     __DEVOLATILE(u_int *, &nfrags), 0,
143     "Current number of IPv4 fragments across all reassembly queues");
144 
145 VNET_DEFINE_STATIC(uma_zone_t, ipq_zone);
146 #define	V_ipq_zone	VNET(ipq_zone)
147 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragpackets, CTLFLAG_VNET |
148     CTLTYPE_INT | CTLFLAG_RW, NULL, 0, sysctl_maxfragpackets, "I",
149     "Maximum number of IPv4 fragment reassembly queue entries");
150 SYSCTL_UMA_CUR(_net_inet_ip, OID_AUTO, fragpackets, CTLFLAG_VNET,
151     &VNET_NAME(ipq_zone),
152     "Current number of IPv4 fragment reassembly queue entries");
153 
154 VNET_DEFINE_STATIC(int, noreass);
155 #define	V_noreass	VNET(noreass)
156 
157 VNET_DEFINE_STATIC(int, maxfragsperpacket);
158 #define	V_maxfragsperpacket	VNET(maxfragsperpacket)
159 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfragsperpacket, CTLFLAG_VNET | CTLFLAG_RW,
160     &VNET_NAME(maxfragsperpacket), 0,
161     "Maximum number of IPv4 fragments allowed per packet");
162 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragbucketsize,
163     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0,
164     sysctl_maxfragbucketsize, "I",
165     "Maximum number of IPv4 fragment reassembly queue entries per bucket");
166 
167 /*
168  * Take incoming datagram fragment and try to reassemble it into
169  * whole datagram.  If the argument is the first fragment or one
170  * in between the function will return NULL and store the mbuf
171  * in the fragment chain.  If the argument is the last fragment
172  * the packet will be reassembled and the pointer to the new
173  * mbuf returned for further processing.  Only m_tags attached
174  * to the first packet/fragment are preserved.
175  * The IP header is *NOT* adjusted out of iplen.
176  */
177 #define	M_IP_FRAG	M_PROTO9
178 struct mbuf *
179 ip_reass(struct mbuf *m)
180 {
181 	struct ip *ip;
182 	struct mbuf *p, *q, *nq, *t;
183 	struct ipq *fp;
184 	struct ipqhead *head;
185 	int i, hlen, next, tmpmax;
186 	u_int8_t ecn, ecn0;
187 	uint32_t hash, hashkey[3];
188 #ifdef	RSS
189 	uint32_t rss_hash, rss_type;
190 #endif
191 
192 	/*
193 	 * If no reassembling or maxfragsperpacket are 0,
194 	 * never accept fragments.
195 	 * Also, drop packet if it would exceed the maximum
196 	 * number of fragments.
197 	 */
198 	tmpmax = maxfrags;
199 	if (V_noreass == 1 || V_maxfragsperpacket == 0 ||
200 	    (tmpmax >= 0 && atomic_load_int(&nfrags) >= (u_int)tmpmax)) {
201 		IPSTAT_INC(ips_fragments);
202 		IPSTAT_INC(ips_fragdropped);
203 		m_freem(m);
204 		return (NULL);
205 	}
206 
207 	ip = mtod(m, struct ip *);
208 	hlen = ip->ip_hl << 2;
209 
210 	/*
211 	 * Adjust ip_len to not reflect header,
212 	 * convert offset of this to bytes.
213 	 */
214 	ip->ip_len = htons(ntohs(ip->ip_len) - hlen);
215 	/*
216 	 * Make sure that fragments have a data length
217 	 * that's a non-zero multiple of 8 bytes, unless
218 	 * this is the last fragment.
219 	 */
220 	if (ip->ip_len == htons(0) ||
221 	    ((ip->ip_off & htons(IP_MF)) && (ntohs(ip->ip_len) & 0x7) != 0)) {
222 		IPSTAT_INC(ips_toosmall); /* XXX */
223 		IPSTAT_INC(ips_fragdropped);
224 		m_freem(m);
225 		return (NULL);
226 	}
227 	if (ip->ip_off & htons(IP_MF))
228 		m->m_flags |= M_IP_FRAG;
229 	else
230 		m->m_flags &= ~M_IP_FRAG;
231 	ip->ip_off = htons(ntohs(ip->ip_off) << 3);
232 
233 	/*
234 	 * Make sure the fragment lies within a packet of valid size.
235 	 */
236 	if (ntohs(ip->ip_len) + ntohs(ip->ip_off) > IP_MAXPACKET) {
237 		IPSTAT_INC(ips_toolong);
238 		IPSTAT_INC(ips_fragdropped);
239 		m_freem(m);
240 		return (NULL);
241 	}
242 
243 	/*
244 	 * Attempt reassembly; if it succeeds, proceed.
245 	 * ip_reass() will return a different mbuf.
246 	 */
247 	IPSTAT_INC(ips_fragments);
248 	m->m_pkthdr.PH_loc.ptr = ip;
249 
250 	/*
251 	 * Presence of header sizes in mbufs
252 	 * would confuse code below.
253 	 */
254 	m->m_data += hlen;
255 	m->m_len -= hlen;
256 
257 	hashkey[0] = ip->ip_src.s_addr;
258 	hashkey[1] = ip->ip_dst.s_addr;
259 	hashkey[2] = (uint32_t)ip->ip_p << 16;
260 	hashkey[2] += ip->ip_id;
261 	hash = jenkins_hash32(hashkey, nitems(hashkey), V_ipq_hashseed);
262 	hash &= IPREASS_HMASK;
263 	head = &V_ipq[hash].head;
264 	IPQ_LOCK(hash);
265 
266 	/*
267 	 * Look for queue of fragments
268 	 * of this datagram.
269 	 */
270 	TAILQ_FOREACH(fp, head, ipq_list)
271 		if (ip->ip_id == fp->ipq_id &&
272 		    ip->ip_src.s_addr == fp->ipq_src.s_addr &&
273 		    ip->ip_dst.s_addr == fp->ipq_dst.s_addr &&
274 #ifdef MAC
275 		    mac_ipq_match(m, fp) &&
276 #endif
277 		    ip->ip_p == fp->ipq_p)
278 			break;
279 	/*
280 	 * If first fragment to arrive, create a reassembly queue.
281 	 */
282 	if (fp == NULL) {
283 		if (V_ipq[hash].count < V_ipreass_maxbucketsize)
284 			fp = uma_zalloc(V_ipq_zone, M_NOWAIT);
285 		if (fp == NULL)
286 			fp = ipq_reuse(hash);
287 		if (fp == NULL)
288 			goto dropfrag;
289 #ifdef MAC
290 		if (mac_ipq_init(fp, M_NOWAIT) != 0) {
291 			uma_zfree(V_ipq_zone, fp);
292 			fp = NULL;
293 			goto dropfrag;
294 		}
295 		mac_ipq_create(m, fp);
296 #endif
297 		TAILQ_INSERT_HEAD(head, fp, ipq_list);
298 		V_ipq[hash].count++;
299 		fp->ipq_nfrags = 1;
300 		atomic_add_int(&nfrags, 1);
301 		fp->ipq_ttl = IPFRAGTTL;
302 		fp->ipq_p = ip->ip_p;
303 		fp->ipq_id = ip->ip_id;
304 		fp->ipq_src = ip->ip_src;
305 		fp->ipq_dst = ip->ip_dst;
306 		fp->ipq_frags = m;
307 		if (m->m_flags & M_IP_FRAG)
308 			fp->ipq_maxoff = -1;
309 		else
310 			fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len);
311 		m->m_nextpkt = NULL;
312 		goto done;
313 	} else {
314 		/*
315 		 * If we already saw the last fragment, make sure
316 		 * this fragment's offset looks sane. Otherwise, if
317 		 * this is the last fragment, record its endpoint.
318 		 */
319 		if (fp->ipq_maxoff > 0) {
320 			i = ntohs(ip->ip_off) + ntohs(ip->ip_len);
321 			if (((m->m_flags & M_IP_FRAG) && i >= fp->ipq_maxoff) ||
322 			    ((m->m_flags & M_IP_FRAG) == 0 &&
323 			    i != fp->ipq_maxoff)) {
324 				fp = NULL;
325 				goto dropfrag;
326 			}
327 		} else if ((m->m_flags & M_IP_FRAG) == 0)
328 			fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len);
329 		fp->ipq_nfrags++;
330 		atomic_add_int(&nfrags, 1);
331 #ifdef MAC
332 		mac_ipq_update(m, fp);
333 #endif
334 	}
335 
336 #define GETIP(m)	((struct ip*)((m)->m_pkthdr.PH_loc.ptr))
337 
338 	/*
339 	 * Handle ECN by comparing this segment with the first one;
340 	 * if CE is set, do not lose CE.
341 	 * drop if CE and not-ECT are mixed for the same packet.
342 	 */
343 	ecn = ip->ip_tos & IPTOS_ECN_MASK;
344 	ecn0 = GETIP(fp->ipq_frags)->ip_tos & IPTOS_ECN_MASK;
345 	if (ecn == IPTOS_ECN_CE) {
346 		if (ecn0 == IPTOS_ECN_NOTECT)
347 			goto dropfrag;
348 		if (ecn0 != IPTOS_ECN_CE)
349 			GETIP(fp->ipq_frags)->ip_tos |= IPTOS_ECN_CE;
350 	}
351 	if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT)
352 		goto dropfrag;
353 
354 	/*
355 	 * Find a segment which begins after this one does.
356 	 */
357 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt)
358 		if (ntohs(GETIP(q)->ip_off) > ntohs(ip->ip_off))
359 			break;
360 
361 	/*
362 	 * If there is a preceding segment, it may provide some of
363 	 * our data already.  If so, drop the data from the incoming
364 	 * segment.  If it provides all of our data, drop us, otherwise
365 	 * stick new segment in the proper place.
366 	 *
367 	 * If some of the data is dropped from the preceding
368 	 * segment, then it's checksum is invalidated.
369 	 */
370 	if (p) {
371 		i = ntohs(GETIP(p)->ip_off) + ntohs(GETIP(p)->ip_len) -
372 		    ntohs(ip->ip_off);
373 		if (i > 0) {
374 			if (i >= ntohs(ip->ip_len))
375 				goto dropfrag;
376 			m_adj(m, i);
377 			m->m_pkthdr.csum_flags = 0;
378 			ip->ip_off = htons(ntohs(ip->ip_off) + i);
379 			ip->ip_len = htons(ntohs(ip->ip_len) - i);
380 		}
381 		m->m_nextpkt = p->m_nextpkt;
382 		p->m_nextpkt = m;
383 	} else {
384 		m->m_nextpkt = fp->ipq_frags;
385 		fp->ipq_frags = m;
386 	}
387 
388 	/*
389 	 * While we overlap succeeding segments trim them or,
390 	 * if they are completely covered, dequeue them.
391 	 */
392 	for (; q != NULL && ntohs(ip->ip_off) + ntohs(ip->ip_len) >
393 	    ntohs(GETIP(q)->ip_off); q = nq) {
394 		i = (ntohs(ip->ip_off) + ntohs(ip->ip_len)) -
395 		    ntohs(GETIP(q)->ip_off);
396 		if (i < ntohs(GETIP(q)->ip_len)) {
397 			GETIP(q)->ip_len = htons(ntohs(GETIP(q)->ip_len) - i);
398 			GETIP(q)->ip_off = htons(ntohs(GETIP(q)->ip_off) + i);
399 			m_adj(q, i);
400 			q->m_pkthdr.csum_flags = 0;
401 			break;
402 		}
403 		nq = q->m_nextpkt;
404 		m->m_nextpkt = nq;
405 		IPSTAT_INC(ips_fragdropped);
406 		fp->ipq_nfrags--;
407 		atomic_subtract_int(&nfrags, 1);
408 		m_freem(q);
409 	}
410 
411 	/*
412 	 * Check for complete reassembly and perform frag per packet
413 	 * limiting.
414 	 *
415 	 * Frag limiting is performed here so that the nth frag has
416 	 * a chance to complete the packet before we drop the packet.
417 	 * As a result, n+1 frags are actually allowed per packet, but
418 	 * only n will ever be stored. (n = maxfragsperpacket.)
419 	 *
420 	 */
421 	next = 0;
422 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt) {
423 		if (ntohs(GETIP(q)->ip_off) != next) {
424 			if (fp->ipq_nfrags > V_maxfragsperpacket)
425 				ipq_drop(&V_ipq[hash], fp);
426 			goto done;
427 		}
428 		next += ntohs(GETIP(q)->ip_len);
429 	}
430 	/* Make sure the last packet didn't have the IP_MF flag */
431 	if (p->m_flags & M_IP_FRAG) {
432 		if (fp->ipq_nfrags > V_maxfragsperpacket)
433 			ipq_drop(&V_ipq[hash], fp);
434 		goto done;
435 	}
436 
437 	/*
438 	 * Reassembly is complete.  Make sure the packet is a sane size.
439 	 */
440 	q = fp->ipq_frags;
441 	ip = GETIP(q);
442 	if (next + (ip->ip_hl << 2) > IP_MAXPACKET) {
443 		IPSTAT_INC(ips_toolong);
444 		ipq_drop(&V_ipq[hash], fp);
445 		goto done;
446 	}
447 
448 	/*
449 	 * Concatenate fragments.
450 	 */
451 	m = q;
452 	t = m->m_next;
453 	m->m_next = NULL;
454 	m_cat(m, t);
455 	nq = q->m_nextpkt;
456 	q->m_nextpkt = NULL;
457 	for (q = nq; q != NULL; q = nq) {
458 		nq = q->m_nextpkt;
459 		q->m_nextpkt = NULL;
460 		m->m_pkthdr.csum_flags &= q->m_pkthdr.csum_flags;
461 		m->m_pkthdr.csum_data += q->m_pkthdr.csum_data;
462 		m_demote_pkthdr(q);
463 		m_cat(m, q);
464 	}
465 	/*
466 	 * In order to do checksumming faster we do 'end-around carry' here
467 	 * (and not in for{} loop), though it implies we are not going to
468 	 * reassemble more than 64k fragments.
469 	 */
470 	while (m->m_pkthdr.csum_data & 0xffff0000)
471 		m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
472 		    (m->m_pkthdr.csum_data >> 16);
473 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
474 #ifdef MAC
475 	mac_ipq_reassemble(fp, m);
476 	mac_ipq_destroy(fp);
477 #endif
478 
479 	/*
480 	 * Create header for new ip packet by modifying header of first
481 	 * packet;  dequeue and discard fragment reassembly header.
482 	 * Make header visible.
483 	 */
484 	ip->ip_len = htons((ip->ip_hl << 2) + next);
485 	ip->ip_src = fp->ipq_src;
486 	ip->ip_dst = fp->ipq_dst;
487 	TAILQ_REMOVE(head, fp, ipq_list);
488 	V_ipq[hash].count--;
489 	uma_zfree(V_ipq_zone, fp);
490 	m->m_len += (ip->ip_hl << 2);
491 	m->m_data -= (ip->ip_hl << 2);
492 	/* some debugging cruft by sklower, below, will go away soon */
493 	if (m->m_flags & M_PKTHDR)	/* XXX this should be done elsewhere */
494 		m_fixhdr(m);
495 	IPSTAT_INC(ips_reassembled);
496 	IPQ_UNLOCK(hash);
497 
498 #ifdef	RSS
499 	/*
500 	 * Query the RSS layer for the flowid / flowtype for the
501 	 * mbuf payload.
502 	 *
503 	 * For now, just assume we have to calculate a new one.
504 	 * Later on we should check to see if the assigned flowid matches
505 	 * what RSS wants for the given IP protocol and if so, just keep it.
506 	 *
507 	 * We then queue into the relevant netisr so it can be dispatched
508 	 * to the correct CPU.
509 	 *
510 	 * Note - this may return 1, which means the flowid in the mbuf
511 	 * is correct for the configured RSS hash types and can be used.
512 	 */
513 	if (rss_mbuf_software_hash_v4(m, 0, &rss_hash, &rss_type) == 0) {
514 		m->m_pkthdr.flowid = rss_hash;
515 		M_HASHTYPE_SET(m, rss_type);
516 	}
517 
518 	/*
519 	 * Queue/dispatch for reprocessing.
520 	 *
521 	 * Note: this is much slower than just handling the frame in the
522 	 * current receive context.  It's likely worth investigating
523 	 * why this is.
524 	 */
525 	netisr_dispatch(NETISR_IP_DIRECT, m);
526 	return (NULL);
527 #endif
528 
529 	/* Handle in-line */
530 	return (m);
531 
532 dropfrag:
533 	IPSTAT_INC(ips_fragdropped);
534 	if (fp != NULL) {
535 		fp->ipq_nfrags--;
536 		atomic_subtract_int(&nfrags, 1);
537 	}
538 	m_freem(m);
539 done:
540 	IPQ_UNLOCK(hash);
541 	return (NULL);
542 
543 #undef GETIP
544 }
545 
546 /*
547  * Initialize IP reassembly structures.
548  */
549 void
550 ipreass_init(void)
551 {
552 	int max;
553 
554 	for (int i = 0; i < IPREASS_NHASH; i++) {
555 		TAILQ_INIT(&V_ipq[i].head);
556 		mtx_init(&V_ipq[i].lock, "IP reassembly", NULL,
557 		    MTX_DEF | MTX_DUPOK);
558 		V_ipq[i].count = 0;
559 	}
560 	V_ipq_hashseed = arc4random();
561 	V_maxfragsperpacket = 16;
562 	V_ipq_zone = uma_zcreate("ipq", sizeof(struct ipq), NULL, NULL, NULL,
563 	    NULL, UMA_ALIGN_PTR, 0);
564 	max = IP_MAXFRAGPACKETS;
565 	max = uma_zone_set_max(V_ipq_zone, max);
566 	V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
567 
568 	if (IS_DEFAULT_VNET(curvnet)) {
569 		maxfrags = IP_MAXFRAGS;
570 		EVENTHANDLER_REGISTER(nmbclusters_change, ipreass_zone_change,
571 		    NULL, EVENTHANDLER_PRI_ANY);
572 	}
573 }
574 
575 /*
576  * If a timer expires on a reassembly queue, discard it.
577  */
578 void
579 ipreass_slowtimo(void)
580 {
581 	struct ipq *fp, *tmp;
582 
583 	for (int i = 0; i < IPREASS_NHASH; i++) {
584 		IPQ_LOCK(i);
585 		TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, tmp)
586 		if (--fp->ipq_ttl == 0)
587 				ipq_timeout(&V_ipq[i], fp);
588 		IPQ_UNLOCK(i);
589 	}
590 }
591 
592 /*
593  * Drain off all datagram fragments.
594  */
595 void
596 ipreass_drain(void)
597 {
598 
599 	for (int i = 0; i < IPREASS_NHASH; i++) {
600 		IPQ_LOCK(i);
601 		while(!TAILQ_EMPTY(&V_ipq[i].head))
602 			ipq_drop(&V_ipq[i], TAILQ_FIRST(&V_ipq[i].head));
603 		KASSERT(V_ipq[i].count == 0,
604 		    ("%s: V_ipq[%d] count %d (V_ipq=%p)", __func__, i,
605 		    V_ipq[i].count, V_ipq));
606 		IPQ_UNLOCK(i);
607 	}
608 }
609 
610 #ifdef VIMAGE
611 /*
612  * Destroy IP reassembly structures.
613  */
614 void
615 ipreass_destroy(void)
616 {
617 
618 	ipreass_drain();
619 	uma_zdestroy(V_ipq_zone);
620 	for (int i = 0; i < IPREASS_NHASH; i++)
621 		mtx_destroy(&V_ipq[i].lock);
622 }
623 #endif
624 
625 /*
626  * After maxnipq has been updated, propagate the change to UMA.  The UMA zone
627  * max has slightly different semantics than the sysctl, for historical
628  * reasons.
629  */
630 static void
631 ipreass_drain_tomax(void)
632 {
633 	struct ipq *fp;
634 	int target;
635 
636 	/*
637 	 * Make sure each bucket is under the new limit. If
638 	 * necessary, drop enough of the oldest elements from
639 	 * each bucket to get under the new limit.
640 	 */
641 	for (int i = 0; i < IPREASS_NHASH; i++) {
642 		IPQ_LOCK(i);
643 		while (V_ipq[i].count > V_ipreass_maxbucketsize &&
644 		    (fp = TAILQ_LAST(&V_ipq[i].head, ipqhead)) != NULL)
645 			ipq_timeout(&V_ipq[i], fp);
646 		IPQ_UNLOCK(i);
647 	}
648 
649 	/*
650 	 * If we are over the maximum number of fragments,
651 	 * drain off enough to get down to the new limit,
652 	 * stripping off last elements on queues.  Every
653 	 * run we strip the oldest element from each bucket.
654 	 */
655 	target = uma_zone_get_max(V_ipq_zone);
656 	while (uma_zone_get_cur(V_ipq_zone) > target) {
657 		for (int i = 0; i < IPREASS_NHASH; i++) {
658 			IPQ_LOCK(i);
659 			fp = TAILQ_LAST(&V_ipq[i].head, ipqhead);
660 			if (fp != NULL)
661 				ipq_timeout(&V_ipq[i], fp);
662 			IPQ_UNLOCK(i);
663 		}
664 	}
665 }
666 
667 static void
668 ipreass_zone_change(void *tag)
669 {
670 	VNET_ITERATOR_DECL(vnet_iter);
671 	int max;
672 
673 	maxfrags = IP_MAXFRAGS;
674 	max = IP_MAXFRAGPACKETS;
675 	VNET_LIST_RLOCK_NOSLEEP();
676 	VNET_FOREACH(vnet_iter) {
677 		CURVNET_SET(vnet_iter);
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 		CURVNET_RESTORE();
682 	}
683 	VNET_LIST_RUNLOCK_NOSLEEP();
684 }
685 
686 /*
687  * Change the limit on the UMA zone, or disable the fragment allocation
688  * at all.  Since 0 and -1 is a special values here, we need our own handler,
689  * instead of sysctl_handle_uma_zone_max().
690  */
691 static int
692 sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS)
693 {
694 	int error, max;
695 
696 	if (V_noreass == 0) {
697 		max = uma_zone_get_max(V_ipq_zone);
698 		if (max == 0)
699 			max = -1;
700 	} else
701 		max = 0;
702 	error = sysctl_handle_int(oidp, &max, 0, req);
703 	if (error || !req->newptr)
704 		return (error);
705 	if (max > 0) {
706 		/*
707 		 * XXXRW: Might be a good idea to sanity check the argument
708 		 * and place an extreme upper bound.
709 		 */
710 		max = uma_zone_set_max(V_ipq_zone, max);
711 		V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
712 		ipreass_drain_tomax();
713 		V_noreass = 0;
714 	} else if (max == 0) {
715 		V_noreass = 1;
716 		ipreass_drain();
717 	} else if (max == -1) {
718 		V_noreass = 0;
719 		uma_zone_set_max(V_ipq_zone, 0);
720 		V_ipreass_maxbucketsize = INT_MAX;
721 	} else
722 		return (EINVAL);
723 	return (0);
724 }
725 
726 /*
727  * Seek for old fragment queue header that can be reused.  Try to
728  * reuse a header from currently locked hash bucket.
729  */
730 static struct ipq *
731 ipq_reuse(int start)
732 {
733 	struct ipq *fp;
734 	int bucket, i;
735 
736 	IPQ_LOCK_ASSERT(start);
737 
738 	for (i = 0; i < IPREASS_NHASH; i++) {
739 		bucket = (start + i) % IPREASS_NHASH;
740 		if (bucket != start && IPQ_TRYLOCK(bucket) == 0)
741 			continue;
742 		fp = TAILQ_LAST(&V_ipq[bucket].head, ipqhead);
743 		if (fp) {
744 			struct mbuf *m;
745 
746 			IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
747 			atomic_subtract_int(&nfrags, fp->ipq_nfrags);
748 			while (fp->ipq_frags) {
749 				m = fp->ipq_frags;
750 				fp->ipq_frags = m->m_nextpkt;
751 				m_freem(m);
752 			}
753 			TAILQ_REMOVE(&V_ipq[bucket].head, fp, ipq_list);
754 			V_ipq[bucket].count--;
755 			if (bucket != start)
756 				IPQ_UNLOCK(bucket);
757 			break;
758 		}
759 		if (bucket != start)
760 			IPQ_UNLOCK(bucket);
761 	}
762 	IPQ_LOCK_ASSERT(start);
763 	return (fp);
764 }
765 
766 /*
767  * Free a fragment reassembly header and all associated datagrams.
768  */
769 static void
770 ipq_free(struct ipqbucket *bucket, struct ipq *fp)
771 {
772 	struct mbuf *q;
773 
774 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
775 	while (fp->ipq_frags) {
776 		q = fp->ipq_frags;
777 		fp->ipq_frags = q->m_nextpkt;
778 		m_freem(q);
779 	}
780 	TAILQ_REMOVE(&bucket->head, fp, ipq_list);
781 	bucket->count--;
782 	uma_zfree(V_ipq_zone, fp);
783 }
784 
785 /*
786  * Get or set the maximum number of reassembly queues per bucket.
787  */
788 static int
789 sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS)
790 {
791 	int error, max;
792 
793 	max = V_ipreass_maxbucketsize;
794 	error = sysctl_handle_int(oidp, &max, 0, req);
795 	if (error || !req->newptr)
796 		return (error);
797 	if (max <= 0)
798 		return (EINVAL);
799 	V_ipreass_maxbucketsize = max;
800 	ipreass_drain_tomax();
801 	return (0);
802 }
803