xref: /freebsd-13-stable/sys/netinet/ip_reass.c (revision 3bc80996974a61a4223eae4c1ccd47b6ee32a48a)
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/rss_config.h>
53 #include <net/netisr.h>
54 #include <net/vnet.h>
55 
56 #include <netinet/in.h>
57 #include <netinet/ip.h>
58 #include <netinet/ip_var.h>
59 #include <netinet/in_rss.h>
60 #ifdef MAC
61 #include <security/mac/mac_framework.h>
62 #endif
63 
64 SYSCTL_DECL(_net_inet_ip);
65 
66 /*
67  * Reassembly headers are stored in hash buckets.
68  */
69 #define	IPREASS_NHASH_LOG2	10
70 #define	IPREASS_NHASH		(1 << IPREASS_NHASH_LOG2)
71 #define	IPREASS_HMASK		(IPREASS_NHASH - 1)
72 
73 struct ipqbucket {
74 	TAILQ_HEAD(ipqhead, ipq) head;
75 	struct mtx		 lock;
76 	int			 count;
77 };
78 
79 VNET_DEFINE_STATIC(struct ipqbucket, ipq[IPREASS_NHASH]);
80 #define	V_ipq		VNET(ipq)
81 VNET_DEFINE_STATIC(uint32_t, ipq_hashseed);
82 #define V_ipq_hashseed   VNET(ipq_hashseed)
83 
84 #define	IPQ_LOCK(i)	mtx_lock(&V_ipq[i].lock)
85 #define	IPQ_TRYLOCK(i)	mtx_trylock(&V_ipq[i].lock)
86 #define	IPQ_UNLOCK(i)	mtx_unlock(&V_ipq[i].lock)
87 #define	IPQ_LOCK_ASSERT(i)	mtx_assert(&V_ipq[i].lock, MA_OWNED)
88 
89 VNET_DEFINE_STATIC(int, ipreass_maxbucketsize);
90 #define	V_ipreass_maxbucketsize	VNET(ipreass_maxbucketsize)
91 
92 void		ipreass_init(void);
93 void		ipreass_drain(void);
94 void		ipreass_slowtimo(void);
95 #ifdef VIMAGE
96 void		ipreass_destroy(void);
97 #endif
98 static int	sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS);
99 static int	sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS);
100 static void	ipreass_zone_change(void *);
101 static void	ipreass_drain_tomax(void);
102 static void	ipq_free(struct ipqbucket *, struct ipq *);
103 static struct ipq * ipq_reuse(int);
104 
105 static inline void
ipq_timeout(struct ipqbucket * bucket,struct ipq * fp)106 ipq_timeout(struct ipqbucket *bucket, struct ipq *fp)
107 {
108 
109 	IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
110 	ipq_free(bucket, fp);
111 }
112 
113 static inline void
ipq_drop(struct ipqbucket * bucket,struct ipq * fp)114 ipq_drop(struct ipqbucket *bucket, struct ipq *fp)
115 {
116 
117 	IPSTAT_ADD(ips_fragdropped, fp->ipq_nfrags);
118 	ipq_free(bucket, fp);
119 }
120 
121 /*
122  * By default, limit the number of IP fragments across all reassembly
123  * queues to  1/32 of the total number of mbuf clusters.
124  *
125  * Limit the total number of reassembly queues per VNET to the
126  * IP fragment limit, but ensure the limit will not allow any bucket
127  * to grow above 100 items. (The bucket limit is
128  * IP_MAXFRAGPACKETS / (IPREASS_NHASH / 2), so the 50 is the correct
129  * multiplier to reach a 100-item limit.)
130  * The 100-item limit was chosen as brief testing seems to show that
131  * this produces "reasonable" performance on some subset of systems
132  * under DoS attack.
133  */
134 #define	IP_MAXFRAGS		(nmbclusters / 32)
135 #define	IP_MAXFRAGPACKETS	(imin(IP_MAXFRAGS, IPREASS_NHASH * 50))
136 
137 static int		maxfrags;
138 static u_int __exclusive_cache_line	nfrags;
139 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfrags, CTLFLAG_RW,
140     &maxfrags, 0,
141     "Maximum number of IPv4 fragments allowed across all reassembly queues");
142 SYSCTL_UINT(_net_inet_ip, OID_AUTO, curfrags, CTLFLAG_RD,
143     &nfrags, 0,
144     "Current number of IPv4 fragments across all reassembly queues");
145 
146 VNET_DEFINE_STATIC(uma_zone_t, ipq_zone);
147 #define	V_ipq_zone	VNET(ipq_zone)
148 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragpackets,
149     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
150     NULL, 0, sysctl_maxfragpackets, "I",
151     "Maximum number of IPv4 fragment reassembly queue entries");
152 SYSCTL_UMA_CUR(_net_inet_ip, OID_AUTO, fragpackets, CTLFLAG_VNET,
153     &VNET_NAME(ipq_zone),
154     "Current number of IPv4 fragment reassembly queue entries");
155 
156 VNET_DEFINE_STATIC(int, noreass);
157 #define	V_noreass	VNET(noreass)
158 
159 VNET_DEFINE_STATIC(int, maxfragsperpacket);
160 #define	V_maxfragsperpacket	VNET(maxfragsperpacket)
161 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfragsperpacket, CTLFLAG_VNET | CTLFLAG_RW,
162     &VNET_NAME(maxfragsperpacket), 0,
163     "Maximum number of IPv4 fragments allowed per packet");
164 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragbucketsize,
165     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0,
166     sysctl_maxfragbucketsize, "I",
167     "Maximum number of IPv4 fragment reassembly queue entries per bucket");
168 
169 /*
170  * Take incoming datagram fragment and try to reassemble it into
171  * whole datagram.  If the argument is the first fragment or one
172  * in between the function will return NULL and store the mbuf
173  * in the fragment chain.  If the argument is the last fragment
174  * the packet will be reassembled and the pointer to the new
175  * mbuf returned for further processing.  Only m_tags attached
176  * to the first packet/fragment are preserved.
177  * The IP header is *NOT* adjusted out of iplen.
178  */
179 #define	M_IP_FRAG	M_PROTO9
180 struct mbuf *
ip_reass(struct mbuf * m)181 ip_reass(struct mbuf *m)
182 {
183 	struct ip *ip;
184 	struct mbuf *p, *q, *nq, *t;
185 	struct ipq *fp;
186 	struct ifnet *srcifp;
187 	struct ipqhead *head;
188 	int i, hlen, next, tmpmax;
189 	u_int8_t ecn, ecn0;
190 	uint32_t hash, hashkey[3];
191 #ifdef	RSS
192 	uint32_t rss_hash, rss_type;
193 #endif
194 
195 	/*
196 	 * If no reassembling or maxfragsperpacket are 0,
197 	 * never accept fragments.
198 	 * Also, drop packet if it would exceed the maximum
199 	 * number of fragments.
200 	 */
201 	tmpmax = maxfrags;
202 	if (V_noreass == 1 || V_maxfragsperpacket == 0 ||
203 	    (tmpmax >= 0 && atomic_load_int(&nfrags) >= (u_int)tmpmax)) {
204 		IPSTAT_INC(ips_fragments);
205 		IPSTAT_INC(ips_fragdropped);
206 		m_freem(m);
207 		return (NULL);
208 	}
209 
210 	ip = mtod(m, struct ip *);
211 	hlen = ip->ip_hl << 2;
212 
213 	/*
214 	 * Adjust ip_len to not reflect header,
215 	 * convert offset of this to bytes.
216 	 */
217 	ip->ip_len = htons(ntohs(ip->ip_len) - hlen);
218 	/*
219 	 * Make sure that fragments have a data length
220 	 * that's a non-zero multiple of 8 bytes, unless
221 	 * this is the last fragment.
222 	 */
223 	if (ip->ip_len == htons(0) ||
224 	    ((ip->ip_off & htons(IP_MF)) && (ntohs(ip->ip_len) & 0x7) != 0)) {
225 		IPSTAT_INC(ips_toosmall); /* XXX */
226 		IPSTAT_INC(ips_fragdropped);
227 		m_freem(m);
228 		return (NULL);
229 	}
230 	if (ip->ip_off & htons(IP_MF))
231 		m->m_flags |= M_IP_FRAG;
232 	else
233 		m->m_flags &= ~M_IP_FRAG;
234 	ip->ip_off = htons(ntohs(ip->ip_off) << 3);
235 
236 	/*
237 	 * Make sure the fragment lies within a packet of valid size.
238 	 */
239 	if (ntohs(ip->ip_len) + ntohs(ip->ip_off) > IP_MAXPACKET) {
240 		IPSTAT_INC(ips_toolong);
241 		IPSTAT_INC(ips_fragdropped);
242 		m_freem(m);
243 		return (NULL);
244 	}
245 
246 	/*
247 	 * Store receive network interface pointer for later.
248 	 */
249 	srcifp = m->m_pkthdr.rcvif;
250 
251 	/*
252 	 * Attempt reassembly; if it succeeds, proceed.
253 	 * ip_reass() will return a different mbuf.
254 	 */
255 	IPSTAT_INC(ips_fragments);
256 	m->m_pkthdr.PH_loc.ptr = ip;
257 
258 	/*
259 	 * Presence of header sizes in mbufs
260 	 * would confuse code below.
261 	 */
262 	m->m_data += hlen;
263 	m->m_len -= hlen;
264 
265 	hashkey[0] = ip->ip_src.s_addr;
266 	hashkey[1] = ip->ip_dst.s_addr;
267 	hashkey[2] = (uint32_t)ip->ip_p << 16;
268 	hashkey[2] += ip->ip_id;
269 	hash = jenkins_hash32(hashkey, nitems(hashkey), V_ipq_hashseed);
270 	hash &= IPREASS_HMASK;
271 	head = &V_ipq[hash].head;
272 	IPQ_LOCK(hash);
273 
274 	/*
275 	 * Look for queue of fragments
276 	 * of this datagram.
277 	 */
278 	TAILQ_FOREACH(fp, head, ipq_list)
279 		if (ip->ip_id == fp->ipq_id &&
280 		    ip->ip_src.s_addr == fp->ipq_src.s_addr &&
281 		    ip->ip_dst.s_addr == fp->ipq_dst.s_addr &&
282 #ifdef MAC
283 		    mac_ipq_match(m, fp) &&
284 #endif
285 		    ip->ip_p == fp->ipq_p)
286 			break;
287 	/*
288 	 * If first fragment to arrive, create a reassembly queue.
289 	 */
290 	if (fp == NULL) {
291 		if (V_ipq[hash].count < V_ipreass_maxbucketsize)
292 			fp = uma_zalloc(V_ipq_zone, M_NOWAIT);
293 		if (fp == NULL)
294 			fp = ipq_reuse(hash);
295 		if (fp == NULL)
296 			goto dropfrag;
297 #ifdef MAC
298 		if (mac_ipq_init(fp, M_NOWAIT) != 0) {
299 			uma_zfree(V_ipq_zone, fp);
300 			fp = NULL;
301 			goto dropfrag;
302 		}
303 		mac_ipq_create(m, fp);
304 #endif
305 		TAILQ_INSERT_HEAD(head, fp, ipq_list);
306 		V_ipq[hash].count++;
307 		fp->ipq_nfrags = 1;
308 		atomic_add_int(&nfrags, 1);
309 		fp->ipq_ttl = IPFRAGTTL;
310 		fp->ipq_p = ip->ip_p;
311 		fp->ipq_id = ip->ip_id;
312 		fp->ipq_src = ip->ip_src;
313 		fp->ipq_dst = ip->ip_dst;
314 		fp->ipq_frags = m;
315 		if (m->m_flags & M_IP_FRAG)
316 			fp->ipq_maxoff = -1;
317 		else
318 			fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len);
319 		m->m_nextpkt = NULL;
320 		goto done;
321 	} else {
322 		/*
323 		 * If we already saw the last fragment, make sure
324 		 * this fragment's offset looks sane. Otherwise, if
325 		 * this is the last fragment, record its endpoint.
326 		 */
327 		if (fp->ipq_maxoff > 0) {
328 			i = ntohs(ip->ip_off) + ntohs(ip->ip_len);
329 			if (((m->m_flags & M_IP_FRAG) && i >= fp->ipq_maxoff) ||
330 			    ((m->m_flags & M_IP_FRAG) == 0 &&
331 			    i != fp->ipq_maxoff)) {
332 				fp = NULL;
333 				goto dropfrag;
334 			}
335 		} else if ((m->m_flags & M_IP_FRAG) == 0)
336 			fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len);
337 		fp->ipq_nfrags++;
338 		atomic_add_int(&nfrags, 1);
339 #ifdef MAC
340 		mac_ipq_update(m, fp);
341 #endif
342 	}
343 
344 #define GETIP(m)	((struct ip*)((m)->m_pkthdr.PH_loc.ptr))
345 
346 	/*
347 	 * Handle ECN by comparing this segment with the first one;
348 	 * if CE is set, do not lose CE.
349 	 * drop if CE and not-ECT are mixed for the same packet.
350 	 */
351 	ecn = ip->ip_tos & IPTOS_ECN_MASK;
352 	ecn0 = GETIP(fp->ipq_frags)->ip_tos & IPTOS_ECN_MASK;
353 	if (ecn == IPTOS_ECN_CE) {
354 		if (ecn0 == IPTOS_ECN_NOTECT)
355 			goto dropfrag;
356 		if (ecn0 != IPTOS_ECN_CE)
357 			GETIP(fp->ipq_frags)->ip_tos |= IPTOS_ECN_CE;
358 	}
359 	if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT)
360 		goto dropfrag;
361 
362 	/*
363 	 * Find a segment which begins after this one does.
364 	 */
365 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt)
366 		if (ntohs(GETIP(q)->ip_off) > ntohs(ip->ip_off))
367 			break;
368 
369 	/*
370 	 * If there is a preceding segment, it may provide some of
371 	 * our data already.  If so, drop the data from the incoming
372 	 * segment.  If it provides all of our data, drop us, otherwise
373 	 * stick new segment in the proper place.
374 	 *
375 	 * If some of the data is dropped from the preceding
376 	 * segment, then it's checksum is invalidated.
377 	 */
378 	if (p) {
379 		i = ntohs(GETIP(p)->ip_off) + ntohs(GETIP(p)->ip_len) -
380 		    ntohs(ip->ip_off);
381 		if (i > 0) {
382 			if (i >= ntohs(ip->ip_len))
383 				goto dropfrag;
384 			m_adj(m, i);
385 			m->m_pkthdr.csum_flags = 0;
386 			ip->ip_off = htons(ntohs(ip->ip_off) + i);
387 			ip->ip_len = htons(ntohs(ip->ip_len) - i);
388 		}
389 		m->m_nextpkt = p->m_nextpkt;
390 		p->m_nextpkt = m;
391 	} else {
392 		m->m_nextpkt = fp->ipq_frags;
393 		fp->ipq_frags = m;
394 	}
395 
396 	/*
397 	 * While we overlap succeeding segments trim them or,
398 	 * if they are completely covered, dequeue them.
399 	 */
400 	for (; q != NULL && ntohs(ip->ip_off) + ntohs(ip->ip_len) >
401 	    ntohs(GETIP(q)->ip_off); q = nq) {
402 		i = (ntohs(ip->ip_off) + ntohs(ip->ip_len)) -
403 		    ntohs(GETIP(q)->ip_off);
404 		if (i < ntohs(GETIP(q)->ip_len)) {
405 			GETIP(q)->ip_len = htons(ntohs(GETIP(q)->ip_len) - i);
406 			GETIP(q)->ip_off = htons(ntohs(GETIP(q)->ip_off) + i);
407 			m_adj(q, i);
408 			q->m_pkthdr.csum_flags = 0;
409 			break;
410 		}
411 		nq = q->m_nextpkt;
412 		m->m_nextpkt = nq;
413 		IPSTAT_INC(ips_fragdropped);
414 		fp->ipq_nfrags--;
415 		atomic_subtract_int(&nfrags, 1);
416 		m_freem(q);
417 	}
418 
419 	/*
420 	 * Check for complete reassembly and perform frag per packet
421 	 * limiting.
422 	 *
423 	 * Frag limiting is performed here so that the nth frag has
424 	 * a chance to complete the packet before we drop the packet.
425 	 * As a result, n+1 frags are actually allowed per packet, but
426 	 * only n will ever be stored. (n = maxfragsperpacket.)
427 	 *
428 	 */
429 	next = 0;
430 	for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt) {
431 		if (ntohs(GETIP(q)->ip_off) != next) {
432 			if (fp->ipq_nfrags > V_maxfragsperpacket)
433 				ipq_drop(&V_ipq[hash], fp);
434 			goto done;
435 		}
436 		next += ntohs(GETIP(q)->ip_len);
437 	}
438 	/* Make sure the last packet didn't have the IP_MF flag */
439 	if (p->m_flags & M_IP_FRAG) {
440 		if (fp->ipq_nfrags > V_maxfragsperpacket)
441 			ipq_drop(&V_ipq[hash], fp);
442 		goto done;
443 	}
444 
445 	/*
446 	 * Reassembly is complete.  Make sure the packet is a sane size.
447 	 */
448 	q = fp->ipq_frags;
449 	ip = GETIP(q);
450 	if (next + (ip->ip_hl << 2) > IP_MAXPACKET) {
451 		IPSTAT_INC(ips_toolong);
452 		ipq_drop(&V_ipq[hash], fp);
453 		goto done;
454 	}
455 
456 	/*
457 	 * Concatenate fragments.
458 	 */
459 	m = q;
460 	t = m->m_next;
461 	m->m_next = NULL;
462 	m_cat(m, t);
463 	nq = q->m_nextpkt;
464 	q->m_nextpkt = NULL;
465 	for (q = nq; q != NULL; q = nq) {
466 		nq = q->m_nextpkt;
467 		q->m_nextpkt = NULL;
468 		m->m_pkthdr.csum_flags &= q->m_pkthdr.csum_flags;
469 		m->m_pkthdr.csum_data += q->m_pkthdr.csum_data;
470 		m_demote_pkthdr(q);
471 		m_cat(m, q);
472 	}
473 	/*
474 	 * In order to do checksumming faster we do 'end-around carry' here
475 	 * (and not in for{} loop), though it implies we are not going to
476 	 * reassemble more than 64k fragments.
477 	 */
478 	while (m->m_pkthdr.csum_data & 0xffff0000)
479 		m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
480 		    (m->m_pkthdr.csum_data >> 16);
481 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
482 #ifdef MAC
483 	mac_ipq_reassemble(fp, m);
484 	mac_ipq_destroy(fp);
485 #endif
486 
487 	/*
488 	 * Create header for new ip packet by modifying header of first
489 	 * packet;  dequeue and discard fragment reassembly header.
490 	 * Make header visible.
491 	 */
492 	ip->ip_len = htons((ip->ip_hl << 2) + next);
493 	ip->ip_src = fp->ipq_src;
494 	ip->ip_dst = fp->ipq_dst;
495 	TAILQ_REMOVE(head, fp, ipq_list);
496 	V_ipq[hash].count--;
497 	uma_zfree(V_ipq_zone, fp);
498 	m->m_len += (ip->ip_hl << 2);
499 	m->m_data -= (ip->ip_hl << 2);
500 	/* some debugging cruft by sklower, below, will go away soon */
501 	if (m->m_flags & M_PKTHDR) {	/* XXX this should be done elsewhere */
502 		m_fixhdr(m);
503 		/* set valid receive interface pointer */
504 		m->m_pkthdr.rcvif = srcifp;
505 	}
506 	IPSTAT_INC(ips_reassembled);
507 	IPQ_UNLOCK(hash);
508 
509 #ifdef	RSS
510 	/*
511 	 * Query the RSS layer for the flowid / flowtype for the
512 	 * mbuf payload.
513 	 *
514 	 * For now, just assume we have to calculate a new one.
515 	 * Later on we should check to see if the assigned flowid matches
516 	 * what RSS wants for the given IP protocol and if so, just keep it.
517 	 *
518 	 * We then queue into the relevant netisr so it can be dispatched
519 	 * to the correct CPU.
520 	 *
521 	 * Note - this may return 1, which means the flowid in the mbuf
522 	 * is correct for the configured RSS hash types and can be used.
523 	 */
524 	if (rss_mbuf_software_hash_v4(m, 0, &rss_hash, &rss_type) == 0) {
525 		m->m_pkthdr.flowid = rss_hash;
526 		M_HASHTYPE_SET(m, rss_type);
527 	}
528 
529 	/*
530 	 * Queue/dispatch for reprocessing.
531 	 *
532 	 * Note: this is much slower than just handling the frame in the
533 	 * current receive context.  It's likely worth investigating
534 	 * why this is.
535 	 */
536 	netisr_dispatch(NETISR_IP_DIRECT, m);
537 	return (NULL);
538 #endif
539 
540 	/* Handle in-line */
541 	return (m);
542 
543 dropfrag:
544 	IPSTAT_INC(ips_fragdropped);
545 	if (fp != NULL) {
546 		fp->ipq_nfrags--;
547 		atomic_subtract_int(&nfrags, 1);
548 	}
549 	m_freem(m);
550 done:
551 	IPQ_UNLOCK(hash);
552 	return (NULL);
553 
554 #undef GETIP
555 }
556 
557 /*
558  * Initialize IP reassembly structures.
559  */
560 void
ipreass_init(void)561 ipreass_init(void)
562 {
563 	int max;
564 
565 	for (int i = 0; i < IPREASS_NHASH; i++) {
566 		TAILQ_INIT(&V_ipq[i].head);
567 		mtx_init(&V_ipq[i].lock, "IP reassembly", NULL,
568 		    MTX_DEF | MTX_DUPOK);
569 		V_ipq[i].count = 0;
570 	}
571 	V_ipq_hashseed = arc4random();
572 	V_maxfragsperpacket = 16;
573 	V_ipq_zone = uma_zcreate("ipq", sizeof(struct ipq), NULL, NULL, NULL,
574 	    NULL, UMA_ALIGN_PTR, 0);
575 	max = IP_MAXFRAGPACKETS;
576 	max = uma_zone_set_max(V_ipq_zone, max);
577 	V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
578 
579 	if (IS_DEFAULT_VNET(curvnet)) {
580 		maxfrags = IP_MAXFRAGS;
581 		EVENTHANDLER_REGISTER(nmbclusters_change, ipreass_zone_change,
582 		    NULL, EVENTHANDLER_PRI_ANY);
583 	}
584 }
585 
586 /*
587  * If a timer expires on a reassembly queue, discard it.
588  */
589 void
ipreass_slowtimo(void)590 ipreass_slowtimo(void)
591 {
592 	struct ipq *fp, *tmp;
593 
594 	if (atomic_load_int(&nfrags) == 0)
595 		return;
596 
597 	for (int i = 0; i < IPREASS_NHASH; i++) {
598 		if (TAILQ_EMPTY(&V_ipq[i].head))
599 			continue;
600 		IPQ_LOCK(i);
601 		TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, tmp)
602 		if (--fp->ipq_ttl == 0)
603 			ipq_timeout(&V_ipq[i], fp);
604 		IPQ_UNLOCK(i);
605 	}
606 }
607 
608 /*
609  * Drain off all datagram fragments.
610  */
611 void
ipreass_drain(void)612 ipreass_drain(void)
613 {
614 
615 	for (int i = 0; i < IPREASS_NHASH; i++) {
616 		IPQ_LOCK(i);
617 		while(!TAILQ_EMPTY(&V_ipq[i].head))
618 			ipq_drop(&V_ipq[i], TAILQ_FIRST(&V_ipq[i].head));
619 		KASSERT(V_ipq[i].count == 0,
620 		    ("%s: V_ipq[%d] count %d (V_ipq=%p)", __func__, i,
621 		    V_ipq[i].count, V_ipq));
622 		IPQ_UNLOCK(i);
623 	}
624 }
625 
626 /*
627  * Drain off all datagram fragments belonging to
628  * the given network interface.
629  */
630 static void
ipreass_cleanup(void * arg __unused,struct ifnet * ifp)631 ipreass_cleanup(void *arg __unused, struct ifnet *ifp)
632 {
633 	struct ipq *fp, *temp;
634 	struct mbuf *m;
635 	int i;
636 
637 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
638 
639 	CURVNET_SET_QUIET(ifp->if_vnet);
640 
641 	/*
642 	 * Skip processing if IPv4 reassembly is not initialised or
643 	 * torn down by ipreass_destroy().
644 	 */
645 	if (V_ipq_zone == NULL) {
646 		CURVNET_RESTORE();
647 		return;
648 	}
649 
650 	for (i = 0; i < IPREASS_NHASH; i++) {
651 		IPQ_LOCK(i);
652 		/* Scan fragment list. */
653 		TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, temp) {
654 			for (m = fp->ipq_frags; m != NULL; m = m->m_nextpkt) {
655 				/* clear no longer valid rcvif pointer */
656 				if (m->m_pkthdr.rcvif == ifp)
657 					m->m_pkthdr.rcvif = NULL;
658 			}
659 		}
660 		IPQ_UNLOCK(i);
661 	}
662 	CURVNET_RESTORE();
663 }
664 EVENTHANDLER_DEFINE(ifnet_departure_event, ipreass_cleanup, NULL, 0);
665 
666 #ifdef VIMAGE
667 /*
668  * Destroy IP reassembly structures.
669  */
670 void
ipreass_destroy(void)671 ipreass_destroy(void)
672 {
673 
674 	ipreass_drain();
675 	uma_zdestroy(V_ipq_zone);
676 	V_ipq_zone = NULL;
677 	for (int i = 0; i < IPREASS_NHASH; i++)
678 		mtx_destroy(&V_ipq[i].lock);
679 }
680 #endif
681 
682 /*
683  * After maxnipq has been updated, propagate the change to UMA.  The UMA zone
684  * max has slightly different semantics than the sysctl, for historical
685  * reasons.
686  */
687 static void
ipreass_drain_tomax(void)688 ipreass_drain_tomax(void)
689 {
690 	struct ipq *fp;
691 	int target;
692 
693 	/*
694 	 * Make sure each bucket is under the new limit. If
695 	 * necessary, drop enough of the oldest elements from
696 	 * each bucket to get under the new limit.
697 	 */
698 	for (int i = 0; i < IPREASS_NHASH; i++) {
699 		IPQ_LOCK(i);
700 		while (V_ipq[i].count > V_ipreass_maxbucketsize &&
701 		    (fp = TAILQ_LAST(&V_ipq[i].head, ipqhead)) != NULL)
702 			ipq_timeout(&V_ipq[i], fp);
703 		IPQ_UNLOCK(i);
704 	}
705 
706 	/*
707 	 * If we are over the maximum number of fragments,
708 	 * drain off enough to get down to the new limit,
709 	 * stripping off last elements on queues.  Every
710 	 * run we strip the oldest element from each bucket.
711 	 */
712 	target = uma_zone_get_max(V_ipq_zone);
713 	while (uma_zone_get_cur(V_ipq_zone) > target) {
714 		for (int i = 0; i < IPREASS_NHASH; i++) {
715 			IPQ_LOCK(i);
716 			fp = TAILQ_LAST(&V_ipq[i].head, ipqhead);
717 			if (fp != NULL)
718 				ipq_timeout(&V_ipq[i], fp);
719 			IPQ_UNLOCK(i);
720 		}
721 	}
722 }
723 
724 static void
ipreass_zone_change(void * tag)725 ipreass_zone_change(void *tag)
726 {
727 	VNET_ITERATOR_DECL(vnet_iter);
728 	int max;
729 
730 	maxfrags = IP_MAXFRAGS;
731 	max = IP_MAXFRAGPACKETS;
732 	VNET_LIST_RLOCK_NOSLEEP();
733 	VNET_FOREACH(vnet_iter) {
734 		CURVNET_SET(vnet_iter);
735 		max = uma_zone_set_max(V_ipq_zone, max);
736 		V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
737 		ipreass_drain_tomax();
738 		CURVNET_RESTORE();
739 	}
740 	VNET_LIST_RUNLOCK_NOSLEEP();
741 }
742 
743 /*
744  * Change the limit on the UMA zone, or disable the fragment allocation
745  * at all.  Since 0 and -1 is a special values here, we need our own handler,
746  * instead of sysctl_handle_uma_zone_max().
747  */
748 static int
sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS)749 sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS)
750 {
751 	int error, max;
752 
753 	if (V_noreass == 0) {
754 		max = uma_zone_get_max(V_ipq_zone);
755 		if (max == 0)
756 			max = -1;
757 	} else
758 		max = 0;
759 	error = sysctl_handle_int(oidp, &max, 0, req);
760 	if (error || !req->newptr)
761 		return (error);
762 	if (max > 0) {
763 		/*
764 		 * XXXRW: Might be a good idea to sanity check the argument
765 		 * and place an extreme upper bound.
766 		 */
767 		max = uma_zone_set_max(V_ipq_zone, max);
768 		V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
769 		ipreass_drain_tomax();
770 		V_noreass = 0;
771 	} else if (max == 0) {
772 		V_noreass = 1;
773 		ipreass_drain();
774 	} else if (max == -1) {
775 		V_noreass = 0;
776 		uma_zone_set_max(V_ipq_zone, 0);
777 		V_ipreass_maxbucketsize = INT_MAX;
778 	} else
779 		return (EINVAL);
780 	return (0);
781 }
782 
783 /*
784  * Seek for old fragment queue header that can be reused.  Try to
785  * reuse a header from currently locked hash bucket.
786  */
787 static struct ipq *
ipq_reuse(int start)788 ipq_reuse(int start)
789 {
790 	struct ipq *fp;
791 	int bucket, i;
792 
793 	IPQ_LOCK_ASSERT(start);
794 
795 	for (i = 0; i < IPREASS_NHASH; i++) {
796 		bucket = (start + i) % IPREASS_NHASH;
797 		if (bucket != start && IPQ_TRYLOCK(bucket) == 0)
798 			continue;
799 		fp = TAILQ_LAST(&V_ipq[bucket].head, ipqhead);
800 		if (fp) {
801 			struct mbuf *m;
802 
803 			IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
804 			atomic_subtract_int(&nfrags, fp->ipq_nfrags);
805 			while (fp->ipq_frags) {
806 				m = fp->ipq_frags;
807 				fp->ipq_frags = m->m_nextpkt;
808 				m_freem(m);
809 			}
810 			TAILQ_REMOVE(&V_ipq[bucket].head, fp, ipq_list);
811 			V_ipq[bucket].count--;
812 			if (bucket != start)
813 				IPQ_UNLOCK(bucket);
814 			break;
815 		}
816 		if (bucket != start)
817 			IPQ_UNLOCK(bucket);
818 	}
819 	IPQ_LOCK_ASSERT(start);
820 	return (fp);
821 }
822 
823 /*
824  * Free a fragment reassembly header and all associated datagrams.
825  */
826 static void
ipq_free(struct ipqbucket * bucket,struct ipq * fp)827 ipq_free(struct ipqbucket *bucket, struct ipq *fp)
828 {
829 	struct mbuf *q;
830 
831 	atomic_subtract_int(&nfrags, fp->ipq_nfrags);
832 	while (fp->ipq_frags) {
833 		q = fp->ipq_frags;
834 		fp->ipq_frags = q->m_nextpkt;
835 		m_freem(q);
836 	}
837 	TAILQ_REMOVE(&bucket->head, fp, ipq_list);
838 	bucket->count--;
839 	uma_zfree(V_ipq_zone, fp);
840 }
841 
842 /*
843  * Get or set the maximum number of reassembly queues per bucket.
844  */
845 static int
sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS)846 sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS)
847 {
848 	int error, max;
849 
850 	max = V_ipreass_maxbucketsize;
851 	error = sysctl_handle_int(oidp, &max, 0, req);
852 	if (error || !req->newptr)
853 		return (error);
854 	if (max <= 0)
855 		return (EINVAL);
856 	V_ipreass_maxbucketsize = max;
857 	ipreass_drain_tomax();
858 	return (0);
859 }
860