xref: /freebsd/sys/netinet/ip_reass.c (revision 3d0d5b21c9b2cfd7e48209340a090f63ca4bce03)
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/if_private.h>
55 #include <net/rss_config.h>
56 #include <net/netisr.h>
57 #include <net/vnet.h>
58 
59 #include <netinet/in.h>
60 #include <netinet/ip.h>
61 #include <netinet/ip_var.h>
62 #include <netinet/in_rss.h>
63 #ifdef MAC
64 #include <security/mac/mac_framework.h>
65 #endif
66 
67 SYSCTL_DECL(_net_inet_ip);
68 
69 /*
70  * Reassembly headers are stored in hash buckets.
71  */
72 #define	IPREASS_NHASH_LOG2	10
73 #define	IPREASS_NHASH		(1 << IPREASS_NHASH_LOG2)
74 #define	IPREASS_HMASK		(V_ipq_hashsize - 1)
75 
76 struct ipqbucket {
77 	TAILQ_HEAD(ipqhead, ipq) head;
78 	struct mtx		 lock;
79 	struct callout		 timer;
80 #ifdef VIMAGE
81 	struct vnet		 *vnet;
82 #endif
83 	int			 count;
84 };
85 
86 VNET_DEFINE_STATIC(struct ipqbucket *, ipq);
87 #define	V_ipq		VNET(ipq)
88 VNET_DEFINE_STATIC(uint32_t, ipq_hashseed);
89 #define	V_ipq_hashseed	VNET(ipq_hashseed)
90 VNET_DEFINE_STATIC(uint32_t, ipq_hashsize);
91 #define	V_ipq_hashsize	VNET(ipq_hashsize)
92 
93 #define	IPQ_LOCK(i)	mtx_lock(&V_ipq[i].lock)
94 #define	IPQ_TRYLOCK(i)	mtx_trylock(&V_ipq[i].lock)
95 #define	IPQ_UNLOCK(i)	mtx_unlock(&V_ipq[i].lock)
96 #define	IPQ_LOCK_ASSERT(i)	mtx_assert(&V_ipq[i].lock, MA_OWNED)
97 #define	IPQ_BUCKET_LOCK_ASSERT(b)	mtx_assert(&(b)->lock, MA_OWNED)
98 
99 VNET_DEFINE_STATIC(int, ipreass_maxbucketsize);
100 #define	V_ipreass_maxbucketsize	VNET(ipreass_maxbucketsize)
101 
102 void		ipreass_init(void);
103 void		ipreass_vnet_init(void);
104 #ifdef VIMAGE
105 void		ipreass_destroy(void);
106 #endif
107 static int	sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS);
108 static int	sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS);
109 static int	sysctl_fragttl(SYSCTL_HANDLER_ARGS);
110 static void	ipreass_zone_change(void *);
111 static void	ipreass_drain_tomax(void);
112 static void	ipq_free(struct ipqbucket *, struct ipq *);
113 static struct ipq * ipq_reuse(int);
114 static void	ipreass_callout(void *);
115 static void	ipreass_reschedule(struct ipqbucket *);
116 
117 static inline void
118 ipq_timeout(struct ipqbucket *bucket, struct ipq *fp)
119 {
120 
121 	IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
122 	ipq_free(bucket, fp);
123 }
124 
125 static inline void
126 ipq_drop(struct ipqbucket *bucket, struct ipq *fp)
127 {
128 
129 	IPSTAT_ADD(ips_fragdropped, fp->ipq_nfrags);
130 	ipq_free(bucket, fp);
131 	ipreass_reschedule(bucket);
132 }
133 
134 /*
135  * By default, limit the number of IP fragments across all reassembly
136  * queues to  1/32 of the total number of mbuf clusters.
137  *
138  * Limit the total number of reassembly queues per VNET to the
139  * IP fragment limit, but ensure the limit will not allow any bucket
140  * to grow above 100 items. (The bucket limit is
141  * IP_MAXFRAGPACKETS / (V_ipq_hashsize / 2), so the 50 is the correct
142  * multiplier to reach a 100-item limit.)
143  * The 100-item limit was chosen as brief testing seems to show that
144  * this produces "reasonable" performance on some subset of systems
145  * under DoS attack.
146  */
147 #define	IP_MAXFRAGS		(nmbclusters / 32)
148 #define	IP_MAXFRAGPACKETS	(imin(IP_MAXFRAGS, V_ipq_hashsize * 50))
149 
150 static int		maxfrags;
151 static u_int __exclusive_cache_line	nfrags;
152 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfrags, CTLFLAG_RW,
153     &maxfrags, 0,
154     "Maximum number of IPv4 fragments allowed across all reassembly queues");
155 SYSCTL_UINT(_net_inet_ip, OID_AUTO, curfrags, CTLFLAG_RD,
156     &nfrags, 0,
157     "Current number of IPv4 fragments across all reassembly queues");
158 
159 VNET_DEFINE_STATIC(uma_zone_t, ipq_zone);
160 #define	V_ipq_zone	VNET(ipq_zone)
161 
162 SYSCTL_UINT(_net_inet_ip, OID_AUTO, reass_hashsize,
163     CTLFLAG_VNET | CTLFLAG_RDTUN, &VNET_NAME(ipq_hashsize), 0,
164     "Size of IP fragment reassembly hashtable");
165 
166 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragpackets,
167     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
168     NULL, 0, sysctl_maxfragpackets, "I",
169     "Maximum number of IPv4 fragment reassembly queue entries");
170 SYSCTL_UMA_CUR(_net_inet_ip, OID_AUTO, fragpackets, CTLFLAG_VNET,
171     &VNET_NAME(ipq_zone),
172     "Current number of IPv4 fragment reassembly queue entries");
173 
174 VNET_DEFINE_STATIC(int, noreass);
175 #define	V_noreass	VNET(noreass)
176 
177 VNET_DEFINE_STATIC(int, maxfragsperpacket);
178 #define	V_maxfragsperpacket	VNET(maxfragsperpacket)
179 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfragsperpacket, CTLFLAG_VNET | CTLFLAG_RW,
180     &VNET_NAME(maxfragsperpacket), 0,
181     "Maximum number of IPv4 fragments allowed per packet");
182 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragbucketsize,
183     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0,
184     sysctl_maxfragbucketsize, "I",
185     "Maximum number of IPv4 fragment reassembly queue entries per bucket");
186 
187 VNET_DEFINE_STATIC(u_int, ipfragttl) = 30;
188 #define	V_ipfragttl	VNET(ipfragttl)
189 SYSCTL_PROC(_net_inet_ip, OID_AUTO, fragttl, CTLTYPE_INT | CTLFLAG_RW |
190     CTLFLAG_MPSAFE | CTLFLAG_VNET, NULL, 0, sysctl_fragttl, "IU",
191     "IP fragment life time on reassembly queue (seconds)");
192 
193 /*
194  * Take incoming datagram fragment and try to reassemble it into
195  * whole datagram.  If the argument is the first fragment or one
196  * in between the function will return NULL and store the mbuf
197  * in the fragment chain.  If the argument is the last fragment
198  * the packet will be reassembled and the pointer to the new
199  * mbuf returned for further processing.  Only m_tags attached
200  * to the first packet/fragment are preserved.
201  * The IP header is *NOT* adjusted out of iplen.
202  */
203 #define	M_IP_FRAG	M_PROTO9
204 struct mbuf *
205 ip_reass(struct mbuf *m)
206 {
207 	struct ip *ip;
208 	struct mbuf *p, *q, *nq, *t;
209 	struct ipq *fp;
210 	struct ifnet *srcifp;
211 	struct ipqhead *head;
212 	int i, hlen, next, tmpmax;
213 	u_int8_t ecn, ecn0;
214 	uint32_t hash, hashkey[3];
215 #ifdef	RSS
216 	uint32_t rss_hash, rss_type;
217 #endif
218 
219 	/*
220 	 * If no reassembling or maxfragsperpacket are 0,
221 	 * never accept fragments.
222 	 * Also, drop packet if it would exceed the maximum
223 	 * number of fragments.
224 	 */
225 	tmpmax = maxfrags;
226 	if (V_noreass == 1 || V_maxfragsperpacket == 0 ||
227 	    (tmpmax >= 0 && atomic_load_int(&nfrags) >= (u_int)tmpmax)) {
228 		IPSTAT_INC(ips_fragments);
229 		IPSTAT_INC(ips_fragdropped);
230 		m_freem(m);
231 		return (NULL);
232 	}
233 
234 	ip = mtod(m, struct ip *);
235 	hlen = ip->ip_hl << 2;
236 
237 	/*
238 	 * Adjust ip_len to not reflect header,
239 	 * convert offset of this to bytes.
240 	 */
241 	ip->ip_len = htons(ntohs(ip->ip_len) - hlen);
242 	/*
243 	 * Make sure that fragments have a data length
244 	 * that's a non-zero multiple of 8 bytes, unless
245 	 * this is the last fragment.
246 	 */
247 	if (ip->ip_len == htons(0) ||
248 	    ((ip->ip_off & htons(IP_MF)) && (ntohs(ip->ip_len) & 0x7) != 0)) {
249 		IPSTAT_INC(ips_toosmall); /* XXX */
250 		IPSTAT_INC(ips_fragdropped);
251 		m_freem(m);
252 		return (NULL);
253 	}
254 	if (ip->ip_off & htons(IP_MF))
255 		m->m_flags |= M_IP_FRAG;
256 	else
257 		m->m_flags &= ~M_IP_FRAG;
258 	ip->ip_off = htons(ntohs(ip->ip_off) << 3);
259 
260 	/*
261 	 * Make sure the fragment lies within a packet of valid size.
262 	 */
263 	if (ntohs(ip->ip_len) + ntohs(ip->ip_off) > IP_MAXPACKET) {
264 		IPSTAT_INC(ips_toolong);
265 		IPSTAT_INC(ips_fragdropped);
266 		m_freem(m);
267 		return (NULL);
268 	}
269 
270 	/*
271 	 * Store receive network interface pointer for later.
272 	 */
273 	srcifp = m->m_pkthdr.rcvif;
274 
275 	/*
276 	 * Attempt reassembly; if it succeeds, proceed.
277 	 * ip_reass() will return a different mbuf.
278 	 */
279 	IPSTAT_INC(ips_fragments);
280 	m->m_pkthdr.PH_loc.ptr = ip;
281 
282 	/*
283 	 * Presence of header sizes in mbufs
284 	 * would confuse code below.
285 	 */
286 	m->m_data += hlen;
287 	m->m_len -= hlen;
288 
289 	hashkey[0] = ip->ip_src.s_addr;
290 	hashkey[1] = ip->ip_dst.s_addr;
291 	hashkey[2] = (uint32_t)ip->ip_p << 16;
292 	hashkey[2] += ip->ip_id;
293 	hash = jenkins_hash32(hashkey, nitems(hashkey), V_ipq_hashseed);
294 	hash &= IPREASS_HMASK;
295 	head = &V_ipq[hash].head;
296 	IPQ_LOCK(hash);
297 
298 	/*
299 	 * Look for queue of fragments
300 	 * of this datagram.
301 	 */
302 	TAILQ_FOREACH(fp, head, ipq_list)
303 		if (ip->ip_id == fp->ipq_id &&
304 		    ip->ip_src.s_addr == fp->ipq_src.s_addr &&
305 		    ip->ip_dst.s_addr == fp->ipq_dst.s_addr &&
306 #ifdef MAC
307 		    mac_ipq_match(m, fp) &&
308 #endif
309 		    ip->ip_p == fp->ipq_p)
310 			break;
311 	/*
312 	 * If first fragment to arrive, create a reassembly queue.
313 	 */
314 	if (fp == NULL) {
315 		if (V_ipq[hash].count < V_ipreass_maxbucketsize)
316 			fp = uma_zalloc(V_ipq_zone, M_NOWAIT);
317 		if (fp == NULL)
318 			fp = ipq_reuse(hash);
319 		if (fp == NULL)
320 			goto dropfrag;
321 #ifdef MAC
322 		if (mac_ipq_init(fp, M_NOWAIT) != 0) {
323 			uma_zfree(V_ipq_zone, fp);
324 			fp = NULL;
325 			goto dropfrag;
326 		}
327 		mac_ipq_create(m, fp);
328 #endif
329 		TAILQ_INSERT_HEAD(head, fp, ipq_list);
330 		V_ipq[hash].count++;
331 		fp->ipq_nfrags = 1;
332 		atomic_add_int(&nfrags, 1);
333 		fp->ipq_expire = time_uptime + V_ipfragttl;
334 		fp->ipq_p = ip->ip_p;
335 		fp->ipq_id = ip->ip_id;
336 		fp->ipq_src = ip->ip_src;
337 		fp->ipq_dst = ip->ip_dst;
338 		fp->ipq_frags = m;
339 		if (m->m_flags & M_IP_FRAG)
340 			fp->ipq_maxoff = -1;
341 		else
342 			fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len);
343 		m->m_nextpkt = NULL;
344 		if (fp == TAILQ_LAST(head, ipqhead))
345 			callout_reset_sbt(&V_ipq[hash].timer,
346 			    SBT_1S * V_ipfragttl, SBT_1S, ipreass_callout,
347 			    &V_ipq[hash], 0);
348 		else
349 			MPASS(callout_active(&V_ipq[hash].timer));
350 		goto done;
351 	} else {
352 		/*
353 		 * If we already saw the last fragment, make sure
354 		 * this fragment's offset looks sane. Otherwise, if
355 		 * this is the last fragment, record its endpoint.
356 		 */
357 		if (fp->ipq_maxoff > 0) {
358 			i = ntohs(ip->ip_off) + ntohs(ip->ip_len);
359 			if (((m->m_flags & M_IP_FRAG) && i >= fp->ipq_maxoff) ||
360 			    ((m->m_flags & M_IP_FRAG) == 0 &&
361 			    i != fp->ipq_maxoff)) {
362 				fp = NULL;
363 				goto dropfrag;
364 			}
365 		} else if ((m->m_flags & M_IP_FRAG) == 0)
366 			fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len);
367 		fp->ipq_nfrags++;
368 		atomic_add_int(&nfrags, 1);
369 #ifdef MAC
370 		mac_ipq_update(m, fp);
371 #endif
372 	}
373 
374 #define GETIP(m)	((struct ip*)((m)->m_pkthdr.PH_loc.ptr))
375 
376 	/*
377 	 * Handle ECN by comparing this segment with the first one;
378 	 * if CE is set, do not lose CE.
379 	 * drop if CE and not-ECT are mixed for the same packet.
380 	 */
381 	ecn = ip->ip_tos & IPTOS_ECN_MASK;
382 	ecn0 = GETIP(fp->ipq_frags)->ip_tos & IPTOS_ECN_MASK;
383 	if (ecn == IPTOS_ECN_CE) {
384 		if (ecn0 == IPTOS_ECN_NOTECT)
385 			goto dropfrag;
386 		if (ecn0 != IPTOS_ECN_CE)
387 			GETIP(fp->ipq_frags)->ip_tos |= IPTOS_ECN_CE;
388 	}
389 	if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT)
390 		goto dropfrag;
391 
392 	/*
393 	 * Find a segment which begins after this one does.
394 	 */
395 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt)
396 		if (ntohs(GETIP(q)->ip_off) > ntohs(ip->ip_off))
397 			break;
398 
399 	/*
400 	 * If there is a preceding segment, it may provide some of
401 	 * our data already.  If so, drop the data from the incoming
402 	 * segment.  If it provides all of our data, drop us, otherwise
403 	 * stick new segment in the proper place.
404 	 *
405 	 * If some of the data is dropped from the preceding
406 	 * segment, then it's checksum is invalidated.
407 	 */
408 	if (p) {
409 		i = ntohs(GETIP(p)->ip_off) + ntohs(GETIP(p)->ip_len) -
410 		    ntohs(ip->ip_off);
411 		if (i > 0) {
412 			if (i >= ntohs(ip->ip_len))
413 				goto dropfrag;
414 			m_adj(m, i);
415 			m->m_pkthdr.csum_flags = 0;
416 			ip->ip_off = htons(ntohs(ip->ip_off) + i);
417 			ip->ip_len = htons(ntohs(ip->ip_len) - i);
418 		}
419 		m->m_nextpkt = p->m_nextpkt;
420 		p->m_nextpkt = m;
421 	} else {
422 		m->m_nextpkt = fp->ipq_frags;
423 		fp->ipq_frags = m;
424 	}
425 
426 	/*
427 	 * While we overlap succeeding segments trim them or,
428 	 * if they are completely covered, dequeue them.
429 	 */
430 	for (; q != NULL && ntohs(ip->ip_off) + ntohs(ip->ip_len) >
431 	    ntohs(GETIP(q)->ip_off); q = nq) {
432 		i = (ntohs(ip->ip_off) + ntohs(ip->ip_len)) -
433 		    ntohs(GETIP(q)->ip_off);
434 		if (i < ntohs(GETIP(q)->ip_len)) {
435 			GETIP(q)->ip_len = htons(ntohs(GETIP(q)->ip_len) - i);
436 			GETIP(q)->ip_off = htons(ntohs(GETIP(q)->ip_off) + i);
437 			m_adj(q, i);
438 			q->m_pkthdr.csum_flags = 0;
439 			break;
440 		}
441 		nq = q->m_nextpkt;
442 		m->m_nextpkt = nq;
443 		IPSTAT_INC(ips_fragdropped);
444 		fp->ipq_nfrags--;
445 		atomic_subtract_int(&nfrags, 1);
446 		m_freem(q);
447 	}
448 
449 	/*
450 	 * Check for complete reassembly and perform frag per packet
451 	 * limiting.
452 	 *
453 	 * Frag limiting is performed here so that the nth frag has
454 	 * a chance to complete the packet before we drop the packet.
455 	 * As a result, n+1 frags are actually allowed per packet, but
456 	 * only n will ever be stored. (n = maxfragsperpacket.)
457 	 *
458 	 */
459 	next = 0;
460 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt) {
461 		if (ntohs(GETIP(q)->ip_off) != next) {
462 			if (fp->ipq_nfrags > V_maxfragsperpacket)
463 				ipq_drop(&V_ipq[hash], fp);
464 			goto done;
465 		}
466 		next += ntohs(GETIP(q)->ip_len);
467 	}
468 	/* Make sure the last packet didn't have the IP_MF flag */
469 	if (p->m_flags & M_IP_FRAG) {
470 		if (fp->ipq_nfrags > V_maxfragsperpacket)
471 			ipq_drop(&V_ipq[hash], fp);
472 		goto done;
473 	}
474 
475 	/*
476 	 * Reassembly is complete.  Make sure the packet is a sane size.
477 	 */
478 	q = fp->ipq_frags;
479 	ip = GETIP(q);
480 	if (next + (ip->ip_hl << 2) > IP_MAXPACKET) {
481 		IPSTAT_INC(ips_toolong);
482 		ipq_drop(&V_ipq[hash], fp);
483 		goto done;
484 	}
485 
486 	/*
487 	 * Concatenate fragments.
488 	 */
489 	m = q;
490 	t = m->m_next;
491 	m->m_next = NULL;
492 	m_cat(m, t);
493 	nq = q->m_nextpkt;
494 	q->m_nextpkt = NULL;
495 	for (q = nq; q != NULL; q = nq) {
496 		nq = q->m_nextpkt;
497 		q->m_nextpkt = NULL;
498 		m->m_pkthdr.csum_flags &= q->m_pkthdr.csum_flags;
499 		m->m_pkthdr.csum_data += q->m_pkthdr.csum_data;
500 		m_demote_pkthdr(q);
501 		m_cat(m, q);
502 	}
503 	/*
504 	 * In order to do checksumming faster we do 'end-around carry' here
505 	 * (and not in for{} loop), though it implies we are not going to
506 	 * reassemble more than 64k fragments.
507 	 */
508 	while (m->m_pkthdr.csum_data & 0xffff0000)
509 		m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
510 		    (m->m_pkthdr.csum_data >> 16);
511 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
512 #ifdef MAC
513 	mac_ipq_reassemble(fp, m);
514 	mac_ipq_destroy(fp);
515 #endif
516 
517 	/*
518 	 * Create header for new ip packet by modifying header of first
519 	 * packet;  dequeue and discard fragment reassembly header.
520 	 * Make header visible.
521 	 */
522 	ip->ip_len = htons((ip->ip_hl << 2) + next);
523 	ip->ip_src = fp->ipq_src;
524 	ip->ip_dst = fp->ipq_dst;
525 	TAILQ_REMOVE(head, fp, ipq_list);
526 	V_ipq[hash].count--;
527 	uma_zfree(V_ipq_zone, fp);
528 	m->m_len += (ip->ip_hl << 2);
529 	m->m_data -= (ip->ip_hl << 2);
530 	/* some debugging cruft by sklower, below, will go away soon */
531 	if (m->m_flags & M_PKTHDR) {	/* XXX this should be done elsewhere */
532 		m_fixhdr(m);
533 		/* set valid receive interface pointer */
534 		m->m_pkthdr.rcvif = srcifp;
535 	}
536 	IPSTAT_INC(ips_reassembled);
537 	ipreass_reschedule(&V_ipq[hash]);
538 	IPQ_UNLOCK(hash);
539 
540 #ifdef	RSS
541 	/*
542 	 * Query the RSS layer for the flowid / flowtype for the
543 	 * mbuf payload.
544 	 *
545 	 * For now, just assume we have to calculate a new one.
546 	 * Later on we should check to see if the assigned flowid matches
547 	 * what RSS wants for the given IP protocol and if so, just keep it.
548 	 *
549 	 * We then queue into the relevant netisr so it can be dispatched
550 	 * to the correct CPU.
551 	 *
552 	 * Note - this may return 1, which means the flowid in the mbuf
553 	 * is correct for the configured RSS hash types and can be used.
554 	 */
555 	if (rss_mbuf_software_hash_v4(m, 0, &rss_hash, &rss_type) == 0) {
556 		m->m_pkthdr.flowid = rss_hash;
557 		M_HASHTYPE_SET(m, rss_type);
558 	}
559 
560 	/*
561 	 * Queue/dispatch for reprocessing.
562 	 *
563 	 * Note: this is much slower than just handling the frame in the
564 	 * current receive context.  It's likely worth investigating
565 	 * why this is.
566 	 */
567 	netisr_dispatch(NETISR_IP_DIRECT, m);
568 	return (NULL);
569 #endif
570 
571 	/* Handle in-line */
572 	return (m);
573 
574 dropfrag:
575 	IPSTAT_INC(ips_fragdropped);
576 	if (fp != NULL) {
577 		fp->ipq_nfrags--;
578 		atomic_subtract_int(&nfrags, 1);
579 	}
580 	m_freem(m);
581 done:
582 	IPQ_UNLOCK(hash);
583 	return (NULL);
584 
585 #undef GETIP
586 }
587 
588 /*
589  * Timer expired on a bucket.
590  * There should be at least one ipq to be timed out.
591  */
592 static void
593 ipreass_callout(void *arg)
594 {
595 	struct ipqbucket *bucket = arg;
596 	struct ipq *fp;
597 
598 	IPQ_BUCKET_LOCK_ASSERT(bucket);
599 	MPASS(atomic_load_int(&nfrags) > 0);
600 
601 	CURVNET_SET(bucket->vnet);
602 	fp = TAILQ_LAST(&bucket->head, ipqhead);
603 	KASSERT(fp != NULL && fp->ipq_expire <= time_uptime,
604 	    ("%s: stray callout on bucket %p, %ju < %ju", __func__, bucket,
605 	    fp ? (uintmax_t)fp->ipq_expire : 0, (uintmax_t)time_uptime));
606 
607 	while (fp != NULL && fp->ipq_expire <= time_uptime) {
608 		ipq_timeout(bucket, fp);
609 		fp = TAILQ_LAST(&bucket->head, ipqhead);
610 	}
611 	ipreass_reschedule(bucket);
612 	CURVNET_RESTORE();
613 }
614 
615 static void
616 ipreass_reschedule(struct ipqbucket *bucket)
617 {
618 	struct ipq *fp;
619 
620 	IPQ_BUCKET_LOCK_ASSERT(bucket);
621 
622 	if ((fp = TAILQ_LAST(&bucket->head, ipqhead)) != NULL) {
623 		time_t t;
624 
625 		/* Protect against time_uptime tick. */
626 		t = fp->ipq_expire - time_uptime;
627 		t = (t > 0) ? t : 1;
628 		callout_reset_sbt(&bucket->timer, SBT_1S * t, SBT_1S,
629 		    ipreass_callout, bucket, 0);
630 	} else
631 		callout_stop(&bucket->timer);
632 }
633 
634 static void
635 ipreass_drain_vnet(void)
636 {
637 	u_int dropped = 0;
638 
639 	for (int i = 0; i < V_ipq_hashsize; i++) {
640 		bool resched;
641 
642 		IPQ_LOCK(i);
643 		resched = !TAILQ_EMPTY(&V_ipq[i].head);
644 		while(!TAILQ_EMPTY(&V_ipq[i].head)) {
645 			struct ipq *fp = TAILQ_FIRST(&V_ipq[i].head);
646 
647 			dropped += fp->ipq_nfrags;
648 			ipq_free(&V_ipq[i], fp);
649 		}
650 		if (resched)
651 			ipreass_reschedule(&V_ipq[i]);
652 		KASSERT(V_ipq[i].count == 0,
653 		    ("%s: V_ipq[%d] count %d (V_ipq=%p)", __func__, i,
654 		    V_ipq[i].count, V_ipq));
655 		IPQ_UNLOCK(i);
656 	}
657 	IPSTAT_ADD(ips_fragdropped, dropped);
658 }
659 
660 /*
661  * Drain off all datagram fragments.
662  */
663 static void
664 ipreass_drain(void)
665 {
666 	VNET_ITERATOR_DECL(vnet_iter);
667 
668 	VNET_FOREACH(vnet_iter) {
669 		CURVNET_SET(vnet_iter);
670 		ipreass_drain_vnet();
671 		CURVNET_RESTORE();
672 	}
673 }
674 
675 
676 /*
677  * Initialize IP reassembly structures.
678  */
679 MALLOC_DEFINE(M_IPREASS_HASH, "IP reass", "IP packet reassembly hash headers");
680 void
681 ipreass_vnet_init(void)
682 {
683 	int max;
684 
685 	V_ipq_hashsize = IPREASS_NHASH;
686 	TUNABLE_INT_FETCH("net.inet.ip.reass_hashsize", &V_ipq_hashsize);
687 	V_ipq = malloc(sizeof(struct ipqbucket) * V_ipq_hashsize,
688 	    M_IPREASS_HASH, M_WAITOK);
689 
690 	for (int i = 0; i < V_ipq_hashsize; i++) {
691 		TAILQ_INIT(&V_ipq[i].head);
692 		mtx_init(&V_ipq[i].lock, "IP reassembly", NULL,
693 		    MTX_DEF | MTX_DUPOK | MTX_NEW);
694 		callout_init_mtx(&V_ipq[i].timer, &V_ipq[i].lock, 0);
695 		V_ipq[i].count = 0;
696 #ifdef VIMAGE
697 		V_ipq[i].vnet = curvnet;
698 #endif
699 	}
700 	V_ipq_hashseed = arc4random();
701 	V_maxfragsperpacket = 16;
702 	V_ipq_zone = uma_zcreate("ipq", sizeof(struct ipq), NULL, NULL, NULL,
703 	    NULL, UMA_ALIGN_PTR, 0);
704 	max = IP_MAXFRAGPACKETS;
705 	max = uma_zone_set_max(V_ipq_zone, max);
706 	V_ipreass_maxbucketsize = imax(max / (V_ipq_hashsize / 2), 1);
707 }
708 
709 void
710 ipreass_init(void)
711 {
712 
713 	maxfrags = IP_MAXFRAGS;
714 	EVENTHANDLER_REGISTER(nmbclusters_change, ipreass_zone_change,
715 	    NULL, EVENTHANDLER_PRI_ANY);
716 	EVENTHANDLER_REGISTER(vm_lowmem, ipreass_drain, NULL,
717 	    LOWMEM_PRI_DEFAULT);
718 	EVENTHANDLER_REGISTER(mbuf_lowmem, ipreass_drain, NULL,
719 		LOWMEM_PRI_DEFAULT);
720 }
721 
722 /*
723  * Drain off all datagram fragments belonging to
724  * the given network interface.
725  */
726 static void
727 ipreass_cleanup(void *arg __unused, struct ifnet *ifp)
728 {
729 	struct ipq *fp, *temp;
730 	struct mbuf *m;
731 	int i;
732 
733 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
734 
735 	CURVNET_SET_QUIET(ifp->if_vnet);
736 
737 	/*
738 	 * Skip processing if IPv4 reassembly is not initialised or
739 	 * torn down by ipreass_destroy().
740 	 */
741 	if (V_ipq_zone == NULL) {
742 		CURVNET_RESTORE();
743 		return;
744 	}
745 
746 	for (i = 0; i < V_ipq_hashsize; i++) {
747 		IPQ_LOCK(i);
748 		/* Scan fragment list. */
749 		TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, temp) {
750 			for (m = fp->ipq_frags; m != NULL; m = m->m_nextpkt) {
751 				/* clear no longer valid rcvif pointer */
752 				if (m->m_pkthdr.rcvif == ifp)
753 					m->m_pkthdr.rcvif = NULL;
754 			}
755 		}
756 		IPQ_UNLOCK(i);
757 	}
758 	CURVNET_RESTORE();
759 }
760 EVENTHANDLER_DEFINE(ifnet_departure_event, ipreass_cleanup, NULL, 0);
761 
762 #ifdef VIMAGE
763 /*
764  * Destroy IP reassembly structures.
765  */
766 void
767 ipreass_destroy(void)
768 {
769 
770 	ipreass_drain_vnet();
771 	uma_zdestroy(V_ipq_zone);
772 	V_ipq_zone = NULL;
773 	for (int i = 0; i < V_ipq_hashsize; i++)
774 		mtx_destroy(&V_ipq[i].lock);
775 	free(V_ipq, M_IPREASS_HASH);
776 }
777 #endif
778 
779 /*
780  * After maxnipq has been updated, propagate the change to UMA.  The UMA zone
781  * max has slightly different semantics than the sysctl, for historical
782  * reasons.
783  */
784 static void
785 ipreass_drain_tomax(void)
786 {
787 	struct ipq *fp;
788 	int target;
789 
790 	/*
791 	 * Make sure each bucket is under the new limit. If
792 	 * necessary, drop enough of the oldest elements from
793 	 * each bucket to get under the new limit.
794 	 */
795 	for (int i = 0; i < V_ipq_hashsize; i++) {
796 		IPQ_LOCK(i);
797 		while (V_ipq[i].count > V_ipreass_maxbucketsize &&
798 		    (fp = TAILQ_LAST(&V_ipq[i].head, ipqhead)) != NULL)
799 			ipq_timeout(&V_ipq[i], fp);
800 		ipreass_reschedule(&V_ipq[i]);
801 		IPQ_UNLOCK(i);
802 	}
803 
804 	/*
805 	 * If we are over the maximum number of fragments,
806 	 * drain off enough to get down to the new limit,
807 	 * stripping off last elements on queues.  Every
808 	 * run we strip the oldest element from each bucket.
809 	 */
810 	target = uma_zone_get_max(V_ipq_zone);
811 	while (uma_zone_get_cur(V_ipq_zone) > target) {
812 		for (int i = 0; i < V_ipq_hashsize; i++) {
813 			IPQ_LOCK(i);
814 			fp = TAILQ_LAST(&V_ipq[i].head, ipqhead);
815 			if (fp != NULL) {
816 				ipq_timeout(&V_ipq[i], fp);
817 				ipreass_reschedule(&V_ipq[i]);
818 			}
819 			IPQ_UNLOCK(i);
820 		}
821 	}
822 }
823 
824 static void
825 ipreass_zone_change(void *tag)
826 {
827 	VNET_ITERATOR_DECL(vnet_iter);
828 	int max;
829 
830 	maxfrags = IP_MAXFRAGS;
831 	max = IP_MAXFRAGPACKETS;
832 	VNET_LIST_RLOCK_NOSLEEP();
833 	VNET_FOREACH(vnet_iter) {
834 		CURVNET_SET(vnet_iter);
835 		max = uma_zone_set_max(V_ipq_zone, max);
836 		V_ipreass_maxbucketsize = imax(max / (V_ipq_hashsize / 2), 1);
837 		ipreass_drain_tomax();
838 		CURVNET_RESTORE();
839 	}
840 	VNET_LIST_RUNLOCK_NOSLEEP();
841 }
842 
843 /*
844  * Change the limit on the UMA zone, or disable the fragment allocation
845  * at all.  Since 0 and -1 is a special values here, we need our own handler,
846  * instead of sysctl_handle_uma_zone_max().
847  */
848 static int
849 sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS)
850 {
851 	int error, max;
852 
853 	if (V_noreass == 0) {
854 		max = uma_zone_get_max(V_ipq_zone);
855 		if (max == 0)
856 			max = -1;
857 	} else
858 		max = 0;
859 	error = sysctl_handle_int(oidp, &max, 0, req);
860 	if (error || !req->newptr)
861 		return (error);
862 	if (max > 0) {
863 		/*
864 		 * XXXRW: Might be a good idea to sanity check the argument
865 		 * and place an extreme upper bound.
866 		 */
867 		max = uma_zone_set_max(V_ipq_zone, max);
868 		V_ipreass_maxbucketsize = imax(max / (V_ipq_hashsize / 2), 1);
869 		ipreass_drain_tomax();
870 		V_noreass = 0;
871 	} else if (max == 0) {
872 		V_noreass = 1;
873 		ipreass_drain();
874 	} else if (max == -1) {
875 		V_noreass = 0;
876 		uma_zone_set_max(V_ipq_zone, 0);
877 		V_ipreass_maxbucketsize = INT_MAX;
878 	} else
879 		return (EINVAL);
880 	return (0);
881 }
882 
883 /*
884  * Seek for old fragment queue header that can be reused.  Try to
885  * reuse a header from currently locked hash bucket.
886  */
887 static struct ipq *
888 ipq_reuse(int start)
889 {
890 	struct ipq *fp;
891 	int bucket, i;
892 
893 	IPQ_LOCK_ASSERT(start);
894 
895 	for (i = 0; i < V_ipq_hashsize; i++) {
896 		bucket = (start + i) % V_ipq_hashsize;
897 		if (bucket != start && IPQ_TRYLOCK(bucket) == 0)
898 			continue;
899 		fp = TAILQ_LAST(&V_ipq[bucket].head, ipqhead);
900 		if (fp) {
901 			struct mbuf *m;
902 
903 			IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
904 			atomic_subtract_int(&nfrags, fp->ipq_nfrags);
905 			while (fp->ipq_frags) {
906 				m = fp->ipq_frags;
907 				fp->ipq_frags = m->m_nextpkt;
908 				m_freem(m);
909 			}
910 			TAILQ_REMOVE(&V_ipq[bucket].head, fp, ipq_list);
911 			V_ipq[bucket].count--;
912 			ipreass_reschedule(&V_ipq[bucket]);
913 			if (bucket != start)
914 				IPQ_UNLOCK(bucket);
915 			break;
916 		}
917 		if (bucket != start)
918 			IPQ_UNLOCK(bucket);
919 	}
920 	IPQ_LOCK_ASSERT(start);
921 	return (fp);
922 }
923 
924 /*
925  * Free a fragment reassembly header and all associated datagrams.
926  */
927 static void
928 ipq_free(struct ipqbucket *bucket, struct ipq *fp)
929 {
930 	struct mbuf *q;
931 
932 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
933 	while (fp->ipq_frags) {
934 		q = fp->ipq_frags;
935 		fp->ipq_frags = q->m_nextpkt;
936 		m_freem(q);
937 	}
938 	TAILQ_REMOVE(&bucket->head, fp, ipq_list);
939 	bucket->count--;
940 	uma_zfree(V_ipq_zone, fp);
941 }
942 
943 /*
944  * Get or set the maximum number of reassembly queues per bucket.
945  */
946 static int
947 sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS)
948 {
949 	int error, max;
950 
951 	max = V_ipreass_maxbucketsize;
952 	error = sysctl_handle_int(oidp, &max, 0, req);
953 	if (error || !req->newptr)
954 		return (error);
955 	if (max <= 0)
956 		return (EINVAL);
957 	V_ipreass_maxbucketsize = max;
958 	ipreass_drain_tomax();
959 	return (0);
960 }
961 
962 /*
963  * Get or set the IP fragment time to live.
964  */
965 static int
966 sysctl_fragttl(SYSCTL_HANDLER_ARGS)
967 {
968 	u_int ttl;
969 	int error;
970 
971 	ttl = V_ipfragttl;
972 	error = sysctl_handle_int(oidp, &ttl, 0, req);
973 	if (error || !req->newptr)
974 		return (error);
975 
976 	if (ttl < 1 || ttl > MAXTTL)
977 		return (EINVAL);
978 
979 	atomic_store_int(&V_ipfragttl, ttl);
980 	return (0);
981 }
982