xref: /freebsd/sys/netinet/ip_reass.c (revision 5e3190f700637fcfc1a52daeaa4a031fdd2557c7)
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 #include "opt_rss.h"
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/hash.h>
42 #include <sys/mbuf.h>
43 #include <sys/malloc.h>
44 #include <sys/limits.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/sysctl.h>
48 #include <sys/socket.h>
49 
50 #include <net/if.h>
51 #include <net/if_var.h>
52 #include <net/if_private.h>
53 #include <net/rss_config.h>
54 #include <net/netisr.h>
55 #include <net/vnet.h>
56 
57 #include <netinet/in.h>
58 #include <netinet/ip.h>
59 #include <netinet/ip_var.h>
60 #include <netinet/in_rss.h>
61 #ifdef MAC
62 #include <security/mac/mac_framework.h>
63 #endif
64 
65 SYSCTL_DECL(_net_inet_ip);
66 
67 /*
68  * Reassembly headers are stored in hash buckets.
69  */
70 #define	IPREASS_NHASH_LOG2	10
71 #define	IPREASS_NHASH		(1 << IPREASS_NHASH_LOG2)
72 #define	IPREASS_HMASK		(V_ipq_hashsize - 1)
73 
74 struct ipqbucket {
75 	TAILQ_HEAD(ipqhead, ipq) head;
76 	struct mtx		 lock;
77 	struct callout		 timer;
78 #ifdef VIMAGE
79 	struct vnet		 *vnet;
80 #endif
81 	int			 count;
82 };
83 
84 VNET_DEFINE_STATIC(struct ipqbucket *, ipq);
85 #define	V_ipq		VNET(ipq)
86 VNET_DEFINE_STATIC(uint32_t, ipq_hashseed);
87 #define	V_ipq_hashseed	VNET(ipq_hashseed)
88 VNET_DEFINE_STATIC(uint32_t, ipq_hashsize);
89 #define	V_ipq_hashsize	VNET(ipq_hashsize)
90 
91 #define	IPQ_LOCK(i)	mtx_lock(&V_ipq[i].lock)
92 #define	IPQ_TRYLOCK(i)	mtx_trylock(&V_ipq[i].lock)
93 #define	IPQ_UNLOCK(i)	mtx_unlock(&V_ipq[i].lock)
94 #define	IPQ_LOCK_ASSERT(i)	mtx_assert(&V_ipq[i].lock, MA_OWNED)
95 #define	IPQ_BUCKET_LOCK_ASSERT(b)	mtx_assert(&(b)->lock, MA_OWNED)
96 
97 VNET_DEFINE_STATIC(int, ipreass_maxbucketsize);
98 #define	V_ipreass_maxbucketsize	VNET(ipreass_maxbucketsize)
99 
100 void		ipreass_init(void);
101 void		ipreass_vnet_init(void);
102 #ifdef VIMAGE
103 void		ipreass_destroy(void);
104 #endif
105 static int	sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS);
106 static int	sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS);
107 static int	sysctl_fragttl(SYSCTL_HANDLER_ARGS);
108 static void	ipreass_zone_change(void *);
109 static void	ipreass_drain_tomax(void);
110 static void	ipq_free(struct ipqbucket *, struct ipq *);
111 static struct ipq * ipq_reuse(int);
112 static void	ipreass_callout(void *);
113 static void	ipreass_reschedule(struct ipqbucket *);
114 
115 static inline void
116 ipq_timeout(struct ipqbucket *bucket, struct ipq *fp)
117 {
118 
119 	IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
120 	ipq_free(bucket, fp);
121 }
122 
123 static inline void
124 ipq_drop(struct ipqbucket *bucket, struct ipq *fp)
125 {
126 
127 	IPSTAT_ADD(ips_fragdropped, fp->ipq_nfrags);
128 	ipq_free(bucket, fp);
129 	ipreass_reschedule(bucket);
130 }
131 
132 /*
133  * By default, limit the number of IP fragments across all reassembly
134  * queues to  1/32 of the total number of mbuf clusters.
135  *
136  * Limit the total number of reassembly queues per VNET to the
137  * IP fragment limit, but ensure the limit will not allow any bucket
138  * to grow above 100 items. (The bucket limit is
139  * IP_MAXFRAGPACKETS / (V_ipq_hashsize / 2), so the 50 is the correct
140  * multiplier to reach a 100-item limit.)
141  * The 100-item limit was chosen as brief testing seems to show that
142  * this produces "reasonable" performance on some subset of systems
143  * under DoS attack.
144  */
145 #define	IP_MAXFRAGS		(nmbclusters / 32)
146 #define	IP_MAXFRAGPACKETS	(imin(IP_MAXFRAGS, V_ipq_hashsize * 50))
147 
148 static int		maxfrags;
149 static u_int __exclusive_cache_line	nfrags;
150 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfrags, CTLFLAG_RW,
151     &maxfrags, 0,
152     "Maximum number of IPv4 fragments allowed across all reassembly queues");
153 SYSCTL_UINT(_net_inet_ip, OID_AUTO, curfrags, CTLFLAG_RD,
154     &nfrags, 0,
155     "Current number of IPv4 fragments across all reassembly queues");
156 
157 VNET_DEFINE_STATIC(uma_zone_t, ipq_zone);
158 #define	V_ipq_zone	VNET(ipq_zone)
159 
160 SYSCTL_UINT(_net_inet_ip, OID_AUTO, reass_hashsize,
161     CTLFLAG_VNET | CTLFLAG_RDTUN, &VNET_NAME(ipq_hashsize), 0,
162     "Size of IP fragment reassembly hashtable");
163 
164 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragpackets,
165     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
166     NULL, 0, sysctl_maxfragpackets, "I",
167     "Maximum number of IPv4 fragment reassembly queue entries");
168 SYSCTL_UMA_CUR(_net_inet_ip, OID_AUTO, fragpackets, CTLFLAG_VNET,
169     &VNET_NAME(ipq_zone),
170     "Current number of IPv4 fragment reassembly queue entries");
171 
172 VNET_DEFINE_STATIC(int, noreass);
173 #define	V_noreass	VNET(noreass)
174 
175 VNET_DEFINE_STATIC(int, maxfragsperpacket);
176 #define	V_maxfragsperpacket	VNET(maxfragsperpacket)
177 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfragsperpacket, CTLFLAG_VNET | CTLFLAG_RW,
178     &VNET_NAME(maxfragsperpacket), 0,
179     "Maximum number of IPv4 fragments allowed per packet");
180 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragbucketsize,
181     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0,
182     sysctl_maxfragbucketsize, "I",
183     "Maximum number of IPv4 fragment reassembly queue entries per bucket");
184 
185 VNET_DEFINE_STATIC(u_int, ipfragttl) = 30;
186 #define	V_ipfragttl	VNET(ipfragttl)
187 SYSCTL_PROC(_net_inet_ip, OID_AUTO, fragttl, CTLTYPE_INT | CTLFLAG_RW |
188     CTLFLAG_MPSAFE | CTLFLAG_VNET, NULL, 0, sysctl_fragttl, "IU",
189     "IP fragment life time on reassembly queue (seconds)");
190 
191 /*
192  * Take incoming datagram fragment and try to reassemble it into
193  * whole datagram.  If the argument is the first fragment or one
194  * in between the function will return NULL and store the mbuf
195  * in the fragment chain.  If the argument is the last fragment
196  * the packet will be reassembled and the pointer to the new
197  * mbuf returned for further processing.  Only m_tags attached
198  * to the first packet/fragment are preserved.
199  * The IP header is *NOT* adjusted out of iplen.
200  */
201 #define	M_IP_FRAG	M_PROTO9
202 struct mbuf *
203 ip_reass(struct mbuf *m)
204 {
205 	struct ip *ip;
206 	struct mbuf *p, *q, *nq, *t;
207 	struct ipq *fp;
208 	struct ifnet *srcifp;
209 	struct ipqhead *head;
210 	int i, hlen, next, tmpmax;
211 	u_int8_t ecn, ecn0;
212 	uint32_t hash, hashkey[3];
213 #ifdef	RSS
214 	uint32_t rss_hash, rss_type;
215 #endif
216 
217 	/*
218 	 * If no reassembling or maxfragsperpacket are 0,
219 	 * never accept fragments.
220 	 * Also, drop packet if it would exceed the maximum
221 	 * number of fragments.
222 	 */
223 	tmpmax = maxfrags;
224 	if (V_noreass == 1 || V_maxfragsperpacket == 0 ||
225 	    (tmpmax >= 0 && atomic_load_int(&nfrags) >= (u_int)tmpmax)) {
226 		IPSTAT_INC(ips_fragments);
227 		IPSTAT_INC(ips_fragdropped);
228 		m_freem(m);
229 		return (NULL);
230 	}
231 
232 	ip = mtod(m, struct ip *);
233 	hlen = ip->ip_hl << 2;
234 
235 	/*
236 	 * Adjust ip_len to not reflect header,
237 	 * convert offset of this to bytes.
238 	 */
239 	ip->ip_len = htons(ntohs(ip->ip_len) - hlen);
240 	/*
241 	 * Make sure that fragments have a data length
242 	 * that's a non-zero multiple of 8 bytes, unless
243 	 * this is the last fragment.
244 	 */
245 	if (ip->ip_len == htons(0) ||
246 	    ((ip->ip_off & htons(IP_MF)) && (ntohs(ip->ip_len) & 0x7) != 0)) {
247 		IPSTAT_INC(ips_toosmall); /* XXX */
248 		IPSTAT_INC(ips_fragdropped);
249 		m_freem(m);
250 		return (NULL);
251 	}
252 	if (ip->ip_off & htons(IP_MF))
253 		m->m_flags |= M_IP_FRAG;
254 	else
255 		m->m_flags &= ~M_IP_FRAG;
256 	ip->ip_off = htons(ntohs(ip->ip_off) << 3);
257 
258 	/*
259 	 * Make sure the fragment lies within a packet of valid size.
260 	 */
261 	if (ntohs(ip->ip_len) + ntohs(ip->ip_off) > IP_MAXPACKET) {
262 		IPSTAT_INC(ips_toolong);
263 		IPSTAT_INC(ips_fragdropped);
264 		m_freem(m);
265 		return (NULL);
266 	}
267 
268 	/*
269 	 * Store receive network interface pointer for later.
270 	 */
271 	srcifp = m->m_pkthdr.rcvif;
272 
273 	/*
274 	 * Attempt reassembly; if it succeeds, proceed.
275 	 * ip_reass() will return a different mbuf.
276 	 */
277 	IPSTAT_INC(ips_fragments);
278 	m->m_pkthdr.PH_loc.ptr = ip;
279 
280 	/*
281 	 * Presence of header sizes in mbufs
282 	 * would confuse code below.
283 	 */
284 	m->m_data += hlen;
285 	m->m_len -= hlen;
286 
287 	hashkey[0] = ip->ip_src.s_addr;
288 	hashkey[1] = ip->ip_dst.s_addr;
289 	hashkey[2] = (uint32_t)ip->ip_p << 16;
290 	hashkey[2] += ip->ip_id;
291 	hash = jenkins_hash32(hashkey, nitems(hashkey), V_ipq_hashseed);
292 	hash &= IPREASS_HMASK;
293 	head = &V_ipq[hash].head;
294 	IPQ_LOCK(hash);
295 
296 	/*
297 	 * Look for queue of fragments
298 	 * of this datagram.
299 	 */
300 	TAILQ_FOREACH(fp, head, ipq_list)
301 		if (ip->ip_id == fp->ipq_id &&
302 		    ip->ip_src.s_addr == fp->ipq_src.s_addr &&
303 		    ip->ip_dst.s_addr == fp->ipq_dst.s_addr &&
304 #ifdef MAC
305 		    mac_ipq_match(m, fp) &&
306 #endif
307 		    ip->ip_p == fp->ipq_p)
308 			break;
309 	/*
310 	 * If first fragment to arrive, create a reassembly queue.
311 	 */
312 	if (fp == NULL) {
313 		if (V_ipq[hash].count < V_ipreass_maxbucketsize)
314 			fp = uma_zalloc(V_ipq_zone, M_NOWAIT);
315 		if (fp == NULL)
316 			fp = ipq_reuse(hash);
317 		if (fp == NULL)
318 			goto dropfrag;
319 #ifdef MAC
320 		if (mac_ipq_init(fp, M_NOWAIT) != 0) {
321 			uma_zfree(V_ipq_zone, fp);
322 			fp = NULL;
323 			goto dropfrag;
324 		}
325 		mac_ipq_create(m, fp);
326 #endif
327 		TAILQ_INSERT_HEAD(head, fp, ipq_list);
328 		V_ipq[hash].count++;
329 		fp->ipq_nfrags = 1;
330 		atomic_add_int(&nfrags, 1);
331 		fp->ipq_expire = time_uptime + V_ipfragttl;
332 		fp->ipq_p = ip->ip_p;
333 		fp->ipq_id = ip->ip_id;
334 		fp->ipq_src = ip->ip_src;
335 		fp->ipq_dst = ip->ip_dst;
336 		fp->ipq_frags = m;
337 		if (m->m_flags & M_IP_FRAG)
338 			fp->ipq_maxoff = -1;
339 		else
340 			fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len);
341 		m->m_nextpkt = NULL;
342 		if (fp == TAILQ_LAST(head, ipqhead))
343 			callout_reset_sbt(&V_ipq[hash].timer,
344 			    SBT_1S * V_ipfragttl, SBT_1S, ipreass_callout,
345 			    &V_ipq[hash], 0);
346 		else
347 			MPASS(callout_active(&V_ipq[hash].timer));
348 		goto done;
349 	} else {
350 		/*
351 		 * If we already saw the last fragment, make sure
352 		 * this fragment's offset looks sane. Otherwise, if
353 		 * this is the last fragment, record its endpoint.
354 		 */
355 		if (fp->ipq_maxoff > 0) {
356 			i = ntohs(ip->ip_off) + ntohs(ip->ip_len);
357 			if (((m->m_flags & M_IP_FRAG) && i >= fp->ipq_maxoff) ||
358 			    ((m->m_flags & M_IP_FRAG) == 0 &&
359 			    i != fp->ipq_maxoff)) {
360 				fp = NULL;
361 				goto dropfrag;
362 			}
363 		} else if ((m->m_flags & M_IP_FRAG) == 0)
364 			fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len);
365 		fp->ipq_nfrags++;
366 		atomic_add_int(&nfrags, 1);
367 #ifdef MAC
368 		mac_ipq_update(m, fp);
369 #endif
370 	}
371 
372 #define GETIP(m)	((struct ip*)((m)->m_pkthdr.PH_loc.ptr))
373 
374 	/*
375 	 * Handle ECN by comparing this segment with the first one;
376 	 * if CE is set, do not lose CE.
377 	 * drop if CE and not-ECT are mixed for the same packet.
378 	 */
379 	ecn = ip->ip_tos & IPTOS_ECN_MASK;
380 	ecn0 = GETIP(fp->ipq_frags)->ip_tos & IPTOS_ECN_MASK;
381 	if (ecn == IPTOS_ECN_CE) {
382 		if (ecn0 == IPTOS_ECN_NOTECT)
383 			goto dropfrag;
384 		if (ecn0 != IPTOS_ECN_CE)
385 			GETIP(fp->ipq_frags)->ip_tos |= IPTOS_ECN_CE;
386 	}
387 	if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT)
388 		goto dropfrag;
389 
390 	/*
391 	 * Find a segment which begins after this one does.
392 	 */
393 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt)
394 		if (ntohs(GETIP(q)->ip_off) > ntohs(ip->ip_off))
395 			break;
396 
397 	/*
398 	 * If there is a preceding segment, it may provide some of
399 	 * our data already.  If so, drop the data from the incoming
400 	 * segment.  If it provides all of our data, drop us, otherwise
401 	 * stick new segment in the proper place.
402 	 *
403 	 * If some of the data is dropped from the preceding
404 	 * segment, then it's checksum is invalidated.
405 	 */
406 	if (p) {
407 		i = ntohs(GETIP(p)->ip_off) + ntohs(GETIP(p)->ip_len) -
408 		    ntohs(ip->ip_off);
409 		if (i > 0) {
410 			if (i >= ntohs(ip->ip_len))
411 				goto dropfrag;
412 			m_adj(m, i);
413 			m->m_pkthdr.csum_flags = 0;
414 			ip->ip_off = htons(ntohs(ip->ip_off) + i);
415 			ip->ip_len = htons(ntohs(ip->ip_len) - i);
416 		}
417 		m->m_nextpkt = p->m_nextpkt;
418 		p->m_nextpkt = m;
419 	} else {
420 		m->m_nextpkt = fp->ipq_frags;
421 		fp->ipq_frags = m;
422 	}
423 
424 	/*
425 	 * While we overlap succeeding segments trim them or,
426 	 * if they are completely covered, dequeue them.
427 	 */
428 	for (; q != NULL && ntohs(ip->ip_off) + ntohs(ip->ip_len) >
429 	    ntohs(GETIP(q)->ip_off); q = nq) {
430 		i = (ntohs(ip->ip_off) + ntohs(ip->ip_len)) -
431 		    ntohs(GETIP(q)->ip_off);
432 		if (i < ntohs(GETIP(q)->ip_len)) {
433 			GETIP(q)->ip_len = htons(ntohs(GETIP(q)->ip_len) - i);
434 			GETIP(q)->ip_off = htons(ntohs(GETIP(q)->ip_off) + i);
435 			m_adj(q, i);
436 			q->m_pkthdr.csum_flags = 0;
437 			break;
438 		}
439 		nq = q->m_nextpkt;
440 		m->m_nextpkt = nq;
441 		IPSTAT_INC(ips_fragdropped);
442 		fp->ipq_nfrags--;
443 		atomic_subtract_int(&nfrags, 1);
444 		m_freem(q);
445 	}
446 
447 	/*
448 	 * Check for complete reassembly and perform frag per packet
449 	 * limiting.
450 	 *
451 	 * Frag limiting is performed here so that the nth frag has
452 	 * a chance to complete the packet before we drop the packet.
453 	 * As a result, n+1 frags are actually allowed per packet, but
454 	 * only n will ever be stored. (n = maxfragsperpacket.)
455 	 *
456 	 */
457 	next = 0;
458 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt) {
459 		if (ntohs(GETIP(q)->ip_off) != next) {
460 			if (fp->ipq_nfrags > V_maxfragsperpacket)
461 				ipq_drop(&V_ipq[hash], fp);
462 			goto done;
463 		}
464 		next += ntohs(GETIP(q)->ip_len);
465 	}
466 	/* Make sure the last packet didn't have the IP_MF flag */
467 	if (p->m_flags & M_IP_FRAG) {
468 		if (fp->ipq_nfrags > V_maxfragsperpacket)
469 			ipq_drop(&V_ipq[hash], fp);
470 		goto done;
471 	}
472 
473 	/*
474 	 * Reassembly is complete.  Make sure the packet is a sane size.
475 	 */
476 	q = fp->ipq_frags;
477 	ip = GETIP(q);
478 	if (next + (ip->ip_hl << 2) > IP_MAXPACKET) {
479 		IPSTAT_INC(ips_toolong);
480 		ipq_drop(&V_ipq[hash], fp);
481 		goto done;
482 	}
483 
484 	/*
485 	 * Concatenate fragments.
486 	 */
487 	m = q;
488 	t = m->m_next;
489 	m->m_next = NULL;
490 	m_cat(m, t);
491 	nq = q->m_nextpkt;
492 	q->m_nextpkt = NULL;
493 	for (q = nq; q != NULL; q = nq) {
494 		nq = q->m_nextpkt;
495 		q->m_nextpkt = NULL;
496 		m->m_pkthdr.csum_flags &= q->m_pkthdr.csum_flags;
497 		m->m_pkthdr.csum_data += q->m_pkthdr.csum_data;
498 		m_demote_pkthdr(q);
499 		m_cat(m, q);
500 	}
501 	/*
502 	 * In order to do checksumming faster we do 'end-around carry' here
503 	 * (and not in for{} loop), though it implies we are not going to
504 	 * reassemble more than 64k fragments.
505 	 */
506 	while (m->m_pkthdr.csum_data & 0xffff0000)
507 		m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
508 		    (m->m_pkthdr.csum_data >> 16);
509 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
510 #ifdef MAC
511 	mac_ipq_reassemble(fp, m);
512 	mac_ipq_destroy(fp);
513 #endif
514 
515 	/*
516 	 * Create header for new ip packet by modifying header of first
517 	 * packet;  dequeue and discard fragment reassembly header.
518 	 * Make header visible.
519 	 */
520 	ip->ip_len = htons((ip->ip_hl << 2) + next);
521 	ip->ip_src = fp->ipq_src;
522 	ip->ip_dst = fp->ipq_dst;
523 	TAILQ_REMOVE(head, fp, ipq_list);
524 	V_ipq[hash].count--;
525 	uma_zfree(V_ipq_zone, fp);
526 	m->m_len += (ip->ip_hl << 2);
527 	m->m_data -= (ip->ip_hl << 2);
528 	/* some debugging cruft by sklower, below, will go away soon */
529 	if (m->m_flags & M_PKTHDR) {	/* XXX this should be done elsewhere */
530 		m_fixhdr(m);
531 		/* set valid receive interface pointer */
532 		m->m_pkthdr.rcvif = srcifp;
533 	}
534 	IPSTAT_INC(ips_reassembled);
535 	ipreass_reschedule(&V_ipq[hash]);
536 	IPQ_UNLOCK(hash);
537 
538 #ifdef	RSS
539 	/*
540 	 * Query the RSS layer for the flowid / flowtype for the
541 	 * mbuf payload.
542 	 *
543 	 * For now, just assume we have to calculate a new one.
544 	 * Later on we should check to see if the assigned flowid matches
545 	 * what RSS wants for the given IP protocol and if so, just keep it.
546 	 *
547 	 * We then queue into the relevant netisr so it can be dispatched
548 	 * to the correct CPU.
549 	 *
550 	 * Note - this may return 1, which means the flowid in the mbuf
551 	 * is correct for the configured RSS hash types and can be used.
552 	 */
553 	if (rss_mbuf_software_hash_v4(m, 0, &rss_hash, &rss_type) == 0) {
554 		m->m_pkthdr.flowid = rss_hash;
555 		M_HASHTYPE_SET(m, rss_type);
556 	}
557 
558 	/*
559 	 * Queue/dispatch for reprocessing.
560 	 *
561 	 * Note: this is much slower than just handling the frame in the
562 	 * current receive context.  It's likely worth investigating
563 	 * why this is.
564 	 */
565 	netisr_dispatch(NETISR_IP_DIRECT, m);
566 	return (NULL);
567 #endif
568 
569 	/* Handle in-line */
570 	return (m);
571 
572 dropfrag:
573 	IPSTAT_INC(ips_fragdropped);
574 	if (fp != NULL) {
575 		fp->ipq_nfrags--;
576 		atomic_subtract_int(&nfrags, 1);
577 	}
578 	m_freem(m);
579 done:
580 	IPQ_UNLOCK(hash);
581 	return (NULL);
582 
583 #undef GETIP
584 }
585 
586 /*
587  * Timer expired on a bucket.
588  * There should be at least one ipq to be timed out.
589  */
590 static void
591 ipreass_callout(void *arg)
592 {
593 	struct ipqbucket *bucket = arg;
594 	struct ipq *fp;
595 
596 	IPQ_BUCKET_LOCK_ASSERT(bucket);
597 	MPASS(atomic_load_int(&nfrags) > 0);
598 
599 	CURVNET_SET(bucket->vnet);
600 	fp = TAILQ_LAST(&bucket->head, ipqhead);
601 	KASSERT(fp != NULL && fp->ipq_expire <= time_uptime,
602 	    ("%s: stray callout on bucket %p, %ju < %ju", __func__, bucket,
603 	    fp ? (uintmax_t)fp->ipq_expire : 0, (uintmax_t)time_uptime));
604 
605 	while (fp != NULL && fp->ipq_expire <= time_uptime) {
606 		ipq_timeout(bucket, fp);
607 		fp = TAILQ_LAST(&bucket->head, ipqhead);
608 	}
609 	ipreass_reschedule(bucket);
610 	CURVNET_RESTORE();
611 }
612 
613 static void
614 ipreass_reschedule(struct ipqbucket *bucket)
615 {
616 	struct ipq *fp;
617 
618 	IPQ_BUCKET_LOCK_ASSERT(bucket);
619 
620 	if ((fp = TAILQ_LAST(&bucket->head, ipqhead)) != NULL) {
621 		time_t t;
622 
623 		/* Protect against time_uptime tick. */
624 		t = fp->ipq_expire - time_uptime;
625 		t = (t > 0) ? t : 1;
626 		callout_reset_sbt(&bucket->timer, SBT_1S * t, SBT_1S,
627 		    ipreass_callout, bucket, 0);
628 	} else
629 		callout_stop(&bucket->timer);
630 }
631 
632 static void
633 ipreass_drain_vnet(void)
634 {
635 	u_int dropped = 0;
636 
637 	for (int i = 0; i < V_ipq_hashsize; i++) {
638 		bool resched;
639 
640 		IPQ_LOCK(i);
641 		resched = !TAILQ_EMPTY(&V_ipq[i].head);
642 		while(!TAILQ_EMPTY(&V_ipq[i].head)) {
643 			struct ipq *fp = TAILQ_FIRST(&V_ipq[i].head);
644 
645 			dropped += fp->ipq_nfrags;
646 			ipq_free(&V_ipq[i], fp);
647 		}
648 		if (resched)
649 			ipreass_reschedule(&V_ipq[i]);
650 		KASSERT(V_ipq[i].count == 0,
651 		    ("%s: V_ipq[%d] count %d (V_ipq=%p)", __func__, i,
652 		    V_ipq[i].count, V_ipq));
653 		IPQ_UNLOCK(i);
654 	}
655 	IPSTAT_ADD(ips_fragdropped, dropped);
656 }
657 
658 /*
659  * Drain off all datagram fragments.
660  */
661 static void
662 ipreass_drain(void)
663 {
664 	VNET_ITERATOR_DECL(vnet_iter);
665 
666 	VNET_FOREACH(vnet_iter) {
667 		CURVNET_SET(vnet_iter);
668 		ipreass_drain_vnet();
669 		CURVNET_RESTORE();
670 	}
671 }
672 
673 
674 /*
675  * Initialize IP reassembly structures.
676  */
677 MALLOC_DEFINE(M_IPREASS_HASH, "IP reass", "IP packet reassembly hash headers");
678 void
679 ipreass_vnet_init(void)
680 {
681 	int max;
682 
683 	V_ipq_hashsize = IPREASS_NHASH;
684 	TUNABLE_INT_FETCH("net.inet.ip.reass_hashsize", &V_ipq_hashsize);
685 	V_ipq = malloc(sizeof(struct ipqbucket) * V_ipq_hashsize,
686 	    M_IPREASS_HASH, M_WAITOK);
687 
688 	for (int i = 0; i < V_ipq_hashsize; i++) {
689 		TAILQ_INIT(&V_ipq[i].head);
690 		mtx_init(&V_ipq[i].lock, "IP reassembly", NULL,
691 		    MTX_DEF | MTX_DUPOK | MTX_NEW);
692 		callout_init_mtx(&V_ipq[i].timer, &V_ipq[i].lock, 0);
693 		V_ipq[i].count = 0;
694 #ifdef VIMAGE
695 		V_ipq[i].vnet = curvnet;
696 #endif
697 	}
698 	V_ipq_hashseed = arc4random();
699 	V_maxfragsperpacket = 16;
700 	V_ipq_zone = uma_zcreate("ipq", sizeof(struct ipq), NULL, NULL, NULL,
701 	    NULL, UMA_ALIGN_PTR, 0);
702 	max = IP_MAXFRAGPACKETS;
703 	max = uma_zone_set_max(V_ipq_zone, max);
704 	V_ipreass_maxbucketsize = imax(max / (V_ipq_hashsize / 2), 1);
705 }
706 
707 void
708 ipreass_init(void)
709 {
710 
711 	maxfrags = IP_MAXFRAGS;
712 	EVENTHANDLER_REGISTER(nmbclusters_change, ipreass_zone_change,
713 	    NULL, EVENTHANDLER_PRI_ANY);
714 	EVENTHANDLER_REGISTER(vm_lowmem, ipreass_drain, NULL,
715 	    LOWMEM_PRI_DEFAULT);
716 	EVENTHANDLER_REGISTER(mbuf_lowmem, ipreass_drain, NULL,
717 		LOWMEM_PRI_DEFAULT);
718 }
719 
720 /*
721  * Drain off all datagram fragments belonging to
722  * the given network interface.
723  */
724 static void
725 ipreass_cleanup(void *arg __unused, struct ifnet *ifp)
726 {
727 	struct ipq *fp, *temp;
728 	struct mbuf *m;
729 	int i;
730 
731 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
732 
733 	CURVNET_SET_QUIET(ifp->if_vnet);
734 
735 	/*
736 	 * Skip processing if IPv4 reassembly is not initialised or
737 	 * torn down by ipreass_destroy().
738 	 */
739 	if (V_ipq_zone == NULL) {
740 		CURVNET_RESTORE();
741 		return;
742 	}
743 
744 	for (i = 0; i < V_ipq_hashsize; i++) {
745 		IPQ_LOCK(i);
746 		/* Scan fragment list. */
747 		TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, temp) {
748 			for (m = fp->ipq_frags; m != NULL; m = m->m_nextpkt) {
749 				/* clear no longer valid rcvif pointer */
750 				if (m->m_pkthdr.rcvif == ifp)
751 					m->m_pkthdr.rcvif = NULL;
752 			}
753 		}
754 		IPQ_UNLOCK(i);
755 	}
756 	CURVNET_RESTORE();
757 }
758 EVENTHANDLER_DEFINE(ifnet_departure_event, ipreass_cleanup, NULL, 0);
759 
760 #ifdef VIMAGE
761 /*
762  * Destroy IP reassembly structures.
763  */
764 void
765 ipreass_destroy(void)
766 {
767 
768 	ipreass_drain_vnet();
769 	uma_zdestroy(V_ipq_zone);
770 	V_ipq_zone = NULL;
771 	for (int i = 0; i < V_ipq_hashsize; i++)
772 		mtx_destroy(&V_ipq[i].lock);
773 	free(V_ipq, M_IPREASS_HASH);
774 }
775 #endif
776 
777 /*
778  * After maxnipq has been updated, propagate the change to UMA.  The UMA zone
779  * max has slightly different semantics than the sysctl, for historical
780  * reasons.
781  */
782 static void
783 ipreass_drain_tomax(void)
784 {
785 	struct ipq *fp;
786 	int target;
787 
788 	/*
789 	 * Make sure each bucket is under the new limit. If
790 	 * necessary, drop enough of the oldest elements from
791 	 * each bucket to get under the new limit.
792 	 */
793 	for (int i = 0; i < V_ipq_hashsize; i++) {
794 		IPQ_LOCK(i);
795 		while (V_ipq[i].count > V_ipreass_maxbucketsize &&
796 		    (fp = TAILQ_LAST(&V_ipq[i].head, ipqhead)) != NULL)
797 			ipq_timeout(&V_ipq[i], fp);
798 		ipreass_reschedule(&V_ipq[i]);
799 		IPQ_UNLOCK(i);
800 	}
801 
802 	/*
803 	 * If we are over the maximum number of fragments,
804 	 * drain off enough to get down to the new limit,
805 	 * stripping off last elements on queues.  Every
806 	 * run we strip the oldest element from each bucket.
807 	 */
808 	target = uma_zone_get_max(V_ipq_zone);
809 	while (uma_zone_get_cur(V_ipq_zone) > target) {
810 		for (int i = 0; i < V_ipq_hashsize; i++) {
811 			IPQ_LOCK(i);
812 			fp = TAILQ_LAST(&V_ipq[i].head, ipqhead);
813 			if (fp != NULL) {
814 				ipq_timeout(&V_ipq[i], fp);
815 				ipreass_reschedule(&V_ipq[i]);
816 			}
817 			IPQ_UNLOCK(i);
818 		}
819 	}
820 }
821 
822 static void
823 ipreass_zone_change(void *tag)
824 {
825 	VNET_ITERATOR_DECL(vnet_iter);
826 	int max;
827 
828 	maxfrags = IP_MAXFRAGS;
829 	max = IP_MAXFRAGPACKETS;
830 	VNET_LIST_RLOCK_NOSLEEP();
831 	VNET_FOREACH(vnet_iter) {
832 		CURVNET_SET(vnet_iter);
833 		max = uma_zone_set_max(V_ipq_zone, max);
834 		V_ipreass_maxbucketsize = imax(max / (V_ipq_hashsize / 2), 1);
835 		ipreass_drain_tomax();
836 		CURVNET_RESTORE();
837 	}
838 	VNET_LIST_RUNLOCK_NOSLEEP();
839 }
840 
841 /*
842  * Change the limit on the UMA zone, or disable the fragment allocation
843  * at all.  Since 0 and -1 is a special values here, we need our own handler,
844  * instead of sysctl_handle_uma_zone_max().
845  */
846 static int
847 sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS)
848 {
849 	int error, max;
850 
851 	if (V_noreass == 0) {
852 		max = uma_zone_get_max(V_ipq_zone);
853 		if (max == 0)
854 			max = -1;
855 	} else
856 		max = 0;
857 	error = sysctl_handle_int(oidp, &max, 0, req);
858 	if (error || !req->newptr)
859 		return (error);
860 	if (max > 0) {
861 		/*
862 		 * XXXRW: Might be a good idea to sanity check the argument
863 		 * and place an extreme upper bound.
864 		 */
865 		max = uma_zone_set_max(V_ipq_zone, max);
866 		V_ipreass_maxbucketsize = imax(max / (V_ipq_hashsize / 2), 1);
867 		ipreass_drain_tomax();
868 		V_noreass = 0;
869 	} else if (max == 0) {
870 		V_noreass = 1;
871 		ipreass_drain();
872 	} else if (max == -1) {
873 		V_noreass = 0;
874 		uma_zone_set_max(V_ipq_zone, 0);
875 		V_ipreass_maxbucketsize = INT_MAX;
876 	} else
877 		return (EINVAL);
878 	return (0);
879 }
880 
881 /*
882  * Seek for old fragment queue header that can be reused.  Try to
883  * reuse a header from currently locked hash bucket.
884  */
885 static struct ipq *
886 ipq_reuse(int start)
887 {
888 	struct ipq *fp;
889 	int bucket, i;
890 
891 	IPQ_LOCK_ASSERT(start);
892 
893 	for (i = 0; i < V_ipq_hashsize; i++) {
894 		bucket = (start + i) % V_ipq_hashsize;
895 		if (bucket != start && IPQ_TRYLOCK(bucket) == 0)
896 			continue;
897 		fp = TAILQ_LAST(&V_ipq[bucket].head, ipqhead);
898 		if (fp) {
899 			struct mbuf *m;
900 
901 			IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
902 			atomic_subtract_int(&nfrags, fp->ipq_nfrags);
903 			while (fp->ipq_frags) {
904 				m = fp->ipq_frags;
905 				fp->ipq_frags = m->m_nextpkt;
906 				m_freem(m);
907 			}
908 			TAILQ_REMOVE(&V_ipq[bucket].head, fp, ipq_list);
909 			V_ipq[bucket].count--;
910 			ipreass_reschedule(&V_ipq[bucket]);
911 			if (bucket != start)
912 				IPQ_UNLOCK(bucket);
913 			break;
914 		}
915 		if (bucket != start)
916 			IPQ_UNLOCK(bucket);
917 	}
918 	IPQ_LOCK_ASSERT(start);
919 	return (fp);
920 }
921 
922 /*
923  * Free a fragment reassembly header and all associated datagrams.
924  */
925 static void
926 ipq_free(struct ipqbucket *bucket, struct ipq *fp)
927 {
928 	struct mbuf *q;
929 
930 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
931 	while (fp->ipq_frags) {
932 		q = fp->ipq_frags;
933 		fp->ipq_frags = q->m_nextpkt;
934 		m_freem(q);
935 	}
936 	TAILQ_REMOVE(&bucket->head, fp, ipq_list);
937 	bucket->count--;
938 	uma_zfree(V_ipq_zone, fp);
939 }
940 
941 /*
942  * Get or set the maximum number of reassembly queues per bucket.
943  */
944 static int
945 sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS)
946 {
947 	int error, max;
948 
949 	max = V_ipreass_maxbucketsize;
950 	error = sysctl_handle_int(oidp, &max, 0, req);
951 	if (error || !req->newptr)
952 		return (error);
953 	if (max <= 0)
954 		return (EINVAL);
955 	V_ipreass_maxbucketsize = max;
956 	ipreass_drain_tomax();
957 	return (0);
958 }
959 
960 /*
961  * Get or set the IP fragment time to live.
962  */
963 static int
964 sysctl_fragttl(SYSCTL_HANDLER_ARGS)
965 {
966 	u_int ttl;
967 	int error;
968 
969 	ttl = V_ipfragttl;
970 	error = sysctl_handle_int(oidp, &ttl, 0, req);
971 	if (error || !req->newptr)
972 		return (error);
973 
974 	if (ttl < 1 || ttl > MAXTTL)
975 		return (EINVAL);
976 
977 	atomic_store_int(&V_ipfragttl, ttl);
978 	return (0);
979 }
980