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.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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.
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
31 * @(#)ip_input.c 8.2 (Berkeley) 1/4/94
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/eventhandler.h>
44 #include <sys/malloc.h>
45 #include <sys/limits.h>
47 #include <sys/mutex.h>
48 #include <sys/sysctl.h>
50 #include <net/rss_config.h>
51 #include <net/netisr.h>
54 #include <netinet/in.h>
55 #include <netinet/ip.h>
56 #include <netinet/ip_var.h>
57 #include <netinet/in_rss.h>
59 #include <security/mac/mac_framework.h>
62 SYSCTL_DECL(_net_inet_ip);
65 * Reassembly headers are stored in hash buckets.
67 #define IPREASS_NHASH_LOG2 10
68 #define IPREASS_NHASH (1 << IPREASS_NHASH_LOG2)
69 #define IPREASS_HMASK (IPREASS_NHASH - 1)
72 TAILQ_HEAD(ipqhead, ipq) head;
77 VNET_DEFINE_STATIC(struct ipqbucket, ipq[IPREASS_NHASH]);
78 #define V_ipq VNET(ipq)
79 VNET_DEFINE_STATIC(uint32_t, ipq_hashseed);
80 #define V_ipq_hashseed VNET(ipq_hashseed)
82 #define IPQ_LOCK(i) mtx_lock(&V_ipq[i].lock)
83 #define IPQ_TRYLOCK(i) mtx_trylock(&V_ipq[i].lock)
84 #define IPQ_UNLOCK(i) mtx_unlock(&V_ipq[i].lock)
85 #define IPQ_LOCK_ASSERT(i) mtx_assert(&V_ipq[i].lock, MA_OWNED)
87 VNET_DEFINE_STATIC(int, ipreass_maxbucketsize);
88 #define V_ipreass_maxbucketsize VNET(ipreass_maxbucketsize)
90 void ipreass_init(void);
91 void ipreass_drain(void);
92 void ipreass_slowtimo(void);
94 void ipreass_destroy(void);
96 static int sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS);
97 static int sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS);
98 static void ipreass_zone_change(void *);
99 static void ipreass_drain_tomax(void);
100 static void ipq_free(struct ipqbucket *, struct ipq *);
101 static struct ipq * ipq_reuse(int);
104 ipq_timeout(struct ipqbucket *bucket, struct ipq *fp)
107 IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
108 ipq_free(bucket, fp);
112 ipq_drop(struct ipqbucket *bucket, struct ipq *fp)
115 IPSTAT_ADD(ips_fragdropped, fp->ipq_nfrags);
116 ipq_free(bucket, fp);
120 * By default, limit the number of IP fragments across all reassembly
121 * queues to 1/32 of the total number of mbuf clusters.
123 * Limit the total number of reassembly queues per VNET to the
124 * IP fragment limit, but ensure the limit will not allow any bucket
125 * to grow above 100 items. (The bucket limit is
126 * IP_MAXFRAGPACKETS / (IPREASS_NHASH / 2), so the 50 is the correct
127 * multiplier to reach a 100-item limit.)
128 * The 100-item limit was chosen as brief testing seems to show that
129 * this produces "reasonable" performance on some subset of systems
132 #define IP_MAXFRAGS (nmbclusters / 32)
133 #define IP_MAXFRAGPACKETS (imin(IP_MAXFRAGS, IPREASS_NHASH * 50))
136 static volatile u_int nfrags;
137 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfrags, CTLFLAG_RW,
139 "Maximum number of IPv4 fragments allowed across all reassembly queues");
140 SYSCTL_UINT(_net_inet_ip, OID_AUTO, curfrags, CTLFLAG_RD,
141 __DEVOLATILE(u_int *, &nfrags), 0,
142 "Current number of IPv4 fragments across all reassembly queues");
144 VNET_DEFINE_STATIC(uma_zone_t, ipq_zone);
145 #define V_ipq_zone VNET(ipq_zone)
146 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragpackets, CTLFLAG_VNET |
147 CTLTYPE_INT | CTLFLAG_RW, NULL, 0, sysctl_maxfragpackets, "I",
148 "Maximum number of IPv4 fragment reassembly queue entries");
149 SYSCTL_UMA_CUR(_net_inet_ip, OID_AUTO, fragpackets, CTLFLAG_VNET,
150 &VNET_NAME(ipq_zone),
151 "Current number of IPv4 fragment reassembly queue entries");
153 VNET_DEFINE_STATIC(int, noreass);
154 #define V_noreass VNET(noreass)
156 VNET_DEFINE_STATIC(int, maxfragsperpacket);
157 #define V_maxfragsperpacket VNET(maxfragsperpacket)
158 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfragsperpacket, CTLFLAG_VNET | CTLFLAG_RW,
159 &VNET_NAME(maxfragsperpacket), 0,
160 "Maximum number of IPv4 fragments allowed per packet");
161 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragbucketsize,
162 CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0,
163 sysctl_maxfragbucketsize, "I",
164 "Maximum number of IPv4 fragment reassembly queue entries per bucket");
167 * Take incoming datagram fragment and try to reassemble it into
168 * whole datagram. If the argument is the first fragment or one
169 * in between the function will return NULL and store the mbuf
170 * in the fragment chain. If the argument is the last fragment
171 * the packet will be reassembled and the pointer to the new
172 * mbuf returned for further processing. Only m_tags attached
173 * to the first packet/fragment are preserved.
174 * The IP header is *NOT* adjusted out of iplen.
176 #define M_IP_FRAG M_PROTO9
178 ip_reass(struct mbuf *m)
181 struct mbuf *p, *q, *nq, *t;
183 struct ipqhead *head;
184 int i, hlen, next, tmpmax;
186 uint32_t hash, hashkey[3];
188 uint32_t rss_hash, rss_type;
192 * If no reassembling or maxfragsperpacket are 0,
193 * never accept fragments.
194 * Also, drop packet if it would exceed the maximum
195 * number of fragments.
198 if (V_noreass == 1 || V_maxfragsperpacket == 0 ||
199 (tmpmax >= 0 && atomic_load_int(&nfrags) >= (u_int)tmpmax)) {
200 IPSTAT_INC(ips_fragments);
201 IPSTAT_INC(ips_fragdropped);
206 ip = mtod(m, struct ip *);
207 hlen = ip->ip_hl << 2;
210 * Adjust ip_len to not reflect header,
211 * convert offset of this to bytes.
213 ip->ip_len = htons(ntohs(ip->ip_len) - hlen);
214 if (ip->ip_off & htons(IP_MF)) {
216 * Make sure that fragments have a data length
217 * that's a non-zero multiple of 8 bytes.
219 if (ip->ip_len == htons(0) || (ntohs(ip->ip_len) & 0x7) != 0) {
220 IPSTAT_INC(ips_toosmall); /* XXX */
221 IPSTAT_INC(ips_fragdropped);
225 m->m_flags |= M_IP_FRAG;
227 m->m_flags &= ~M_IP_FRAG;
228 ip->ip_off = htons(ntohs(ip->ip_off) << 3);
231 * Make sure the fragment lies within a packet of valid size.
233 if (ntohs(ip->ip_len) + ntohs(ip->ip_off) > IP_MAXPACKET) {
234 IPSTAT_INC(ips_toolong);
235 IPSTAT_INC(ips_fragdropped);
241 * Attempt reassembly; if it succeeds, proceed.
242 * ip_reass() will return a different mbuf.
244 IPSTAT_INC(ips_fragments);
245 m->m_pkthdr.PH_loc.ptr = ip;
248 * Presence of header sizes in mbufs
249 * would confuse code below.
254 hashkey[0] = ip->ip_src.s_addr;
255 hashkey[1] = ip->ip_dst.s_addr;
256 hashkey[2] = (uint32_t)ip->ip_p << 16;
257 hashkey[2] += ip->ip_id;
258 hash = jenkins_hash32(hashkey, nitems(hashkey), V_ipq_hashseed);
259 hash &= IPREASS_HMASK;
260 head = &V_ipq[hash].head;
264 * Look for queue of fragments
267 TAILQ_FOREACH(fp, head, ipq_list)
268 if (ip->ip_id == fp->ipq_id &&
269 ip->ip_src.s_addr == fp->ipq_src.s_addr &&
270 ip->ip_dst.s_addr == fp->ipq_dst.s_addr &&
272 mac_ipq_match(m, fp) &&
274 ip->ip_p == fp->ipq_p)
277 * If first fragment to arrive, create a reassembly queue.
280 if (V_ipq[hash].count < V_ipreass_maxbucketsize)
281 fp = uma_zalloc(V_ipq_zone, M_NOWAIT);
283 fp = ipq_reuse(hash);
287 if (mac_ipq_init(fp, M_NOWAIT) != 0) {
288 uma_zfree(V_ipq_zone, fp);
292 mac_ipq_create(m, fp);
294 TAILQ_INSERT_HEAD(head, fp, ipq_list);
297 atomic_add_int(&nfrags, 1);
298 fp->ipq_ttl = IPFRAGTTL;
299 fp->ipq_p = ip->ip_p;
300 fp->ipq_id = ip->ip_id;
301 fp->ipq_src = ip->ip_src;
302 fp->ipq_dst = ip->ip_dst;
308 atomic_add_int(&nfrags, 1);
310 mac_ipq_update(m, fp);
314 #define GETIP(m) ((struct ip*)((m)->m_pkthdr.PH_loc.ptr))
317 * Handle ECN by comparing this segment with the first one;
318 * if CE is set, do not lose CE.
319 * drop if CE and not-ECT are mixed for the same packet.
321 ecn = ip->ip_tos & IPTOS_ECN_MASK;
322 ecn0 = GETIP(fp->ipq_frags)->ip_tos & IPTOS_ECN_MASK;
323 if (ecn == IPTOS_ECN_CE) {
324 if (ecn0 == IPTOS_ECN_NOTECT)
326 if (ecn0 != IPTOS_ECN_CE)
327 GETIP(fp->ipq_frags)->ip_tos |= IPTOS_ECN_CE;
329 if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT)
333 * Find a segment which begins after this one does.
335 for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt)
336 if (ntohs(GETIP(q)->ip_off) > ntohs(ip->ip_off))
340 * If there is a preceding segment, it may provide some of
341 * our data already. If so, drop the data from the incoming
342 * segment. If it provides all of our data, drop us, otherwise
343 * stick new segment in the proper place.
345 * If some of the data is dropped from the preceding
346 * segment, then it's checksum is invalidated.
349 i = ntohs(GETIP(p)->ip_off) + ntohs(GETIP(p)->ip_len) -
352 if (i >= ntohs(ip->ip_len))
355 m->m_pkthdr.csum_flags = 0;
356 ip->ip_off = htons(ntohs(ip->ip_off) + i);
357 ip->ip_len = htons(ntohs(ip->ip_len) - i);
359 m->m_nextpkt = p->m_nextpkt;
362 m->m_nextpkt = fp->ipq_frags;
367 * While we overlap succeeding segments trim them or,
368 * if they are completely covered, dequeue them.
370 for (; q != NULL && ntohs(ip->ip_off) + ntohs(ip->ip_len) >
371 ntohs(GETIP(q)->ip_off); q = nq) {
372 i = (ntohs(ip->ip_off) + ntohs(ip->ip_len)) -
373 ntohs(GETIP(q)->ip_off);
374 if (i < ntohs(GETIP(q)->ip_len)) {
375 GETIP(q)->ip_len = htons(ntohs(GETIP(q)->ip_len) - i);
376 GETIP(q)->ip_off = htons(ntohs(GETIP(q)->ip_off) + i);
378 q->m_pkthdr.csum_flags = 0;
383 IPSTAT_INC(ips_fragdropped);
385 atomic_subtract_int(&nfrags, 1);
390 * Check for complete reassembly and perform frag per packet
393 * Frag limiting is performed here so that the nth frag has
394 * a chance to complete the packet before we drop the packet.
395 * As a result, n+1 frags are actually allowed per packet, but
396 * only n will ever be stored. (n = maxfragsperpacket.)
400 for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt) {
401 if (ntohs(GETIP(q)->ip_off) != next) {
402 if (fp->ipq_nfrags > V_maxfragsperpacket)
403 ipq_drop(&V_ipq[hash], fp);
406 next += ntohs(GETIP(q)->ip_len);
408 /* Make sure the last packet didn't have the IP_MF flag */
409 if (p->m_flags & M_IP_FRAG) {
410 if (fp->ipq_nfrags > V_maxfragsperpacket)
411 ipq_drop(&V_ipq[hash], fp);
416 * Reassembly is complete. Make sure the packet is a sane size.
420 if (next + (ip->ip_hl << 2) > IP_MAXPACKET) {
421 IPSTAT_INC(ips_toolong);
422 ipq_drop(&V_ipq[hash], fp);
427 * Concatenate fragments.
435 for (q = nq; q != NULL; q = nq) {
438 m->m_pkthdr.csum_flags &= q->m_pkthdr.csum_flags;
439 m->m_pkthdr.csum_data += q->m_pkthdr.csum_data;
444 * In order to do checksumming faster we do 'end-around carry' here
445 * (and not in for{} loop), though it implies we are not going to
446 * reassemble more than 64k fragments.
448 while (m->m_pkthdr.csum_data & 0xffff0000)
449 m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
450 (m->m_pkthdr.csum_data >> 16);
451 atomic_subtract_int(&nfrags, fp->ipq_nfrags);
453 mac_ipq_reassemble(fp, m);
458 * Create header for new ip packet by modifying header of first
459 * packet; dequeue and discard fragment reassembly header.
460 * Make header visible.
462 ip->ip_len = htons((ip->ip_hl << 2) + next);
463 ip->ip_src = fp->ipq_src;
464 ip->ip_dst = fp->ipq_dst;
465 TAILQ_REMOVE(head, fp, ipq_list);
467 uma_zfree(V_ipq_zone, fp);
468 m->m_len += (ip->ip_hl << 2);
469 m->m_data -= (ip->ip_hl << 2);
470 /* some debugging cruft by sklower, below, will go away soon */
471 if (m->m_flags & M_PKTHDR) /* XXX this should be done elsewhere */
473 IPSTAT_INC(ips_reassembled);
478 * Query the RSS layer for the flowid / flowtype for the
481 * For now, just assume we have to calculate a new one.
482 * Later on we should check to see if the assigned flowid matches
483 * what RSS wants for the given IP protocol and if so, just keep it.
485 * We then queue into the relevant netisr so it can be dispatched
486 * to the correct CPU.
488 * Note - this may return 1, which means the flowid in the mbuf
489 * is correct for the configured RSS hash types and can be used.
491 if (rss_mbuf_software_hash_v4(m, 0, &rss_hash, &rss_type) == 0) {
492 m->m_pkthdr.flowid = rss_hash;
493 M_HASHTYPE_SET(m, rss_type);
497 * Queue/dispatch for reprocessing.
499 * Note: this is much slower than just handling the frame in the
500 * current receive context. It's likely worth investigating
503 netisr_dispatch(NETISR_IP_DIRECT, m);
511 IPSTAT_INC(ips_fragdropped);
514 atomic_subtract_int(&nfrags, 1);
525 * Initialize IP reassembly structures.
532 for (int i = 0; i < IPREASS_NHASH; i++) {
533 TAILQ_INIT(&V_ipq[i].head);
534 mtx_init(&V_ipq[i].lock, "IP reassembly", NULL,
535 MTX_DEF | MTX_DUPOK);
538 V_ipq_hashseed = arc4random();
539 V_maxfragsperpacket = 16;
540 V_ipq_zone = uma_zcreate("ipq", sizeof(struct ipq), NULL, NULL, NULL,
541 NULL, UMA_ALIGN_PTR, 0);
542 max = IP_MAXFRAGPACKETS;
543 max = uma_zone_set_max(V_ipq_zone, max);
544 V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
546 if (IS_DEFAULT_VNET(curvnet)) {
547 maxfrags = IP_MAXFRAGS;
548 EVENTHANDLER_REGISTER(nmbclusters_change, ipreass_zone_change,
549 NULL, EVENTHANDLER_PRI_ANY);
554 * If a timer expires on a reassembly queue, discard it.
557 ipreass_slowtimo(void)
559 struct ipq *fp, *tmp;
561 for (int i = 0; i < IPREASS_NHASH; i++) {
563 TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, tmp)
564 if (--fp->ipq_ttl == 0)
565 ipq_timeout(&V_ipq[i], fp);
571 * Drain off all datagram fragments.
577 for (int i = 0; i < IPREASS_NHASH; i++) {
579 while(!TAILQ_EMPTY(&V_ipq[i].head))
580 ipq_drop(&V_ipq[i], TAILQ_FIRST(&V_ipq[i].head));
581 KASSERT(V_ipq[i].count == 0,
582 ("%s: V_ipq[%d] count %d (V_ipq=%p)", __func__, i,
583 V_ipq[i].count, V_ipq));
590 * Destroy IP reassembly structures.
593 ipreass_destroy(void)
597 uma_zdestroy(V_ipq_zone);
598 for (int i = 0; i < IPREASS_NHASH; i++)
599 mtx_destroy(&V_ipq[i].lock);
604 * After maxnipq has been updated, propagate the change to UMA. The UMA zone
605 * max has slightly different semantics than the sysctl, for historical
609 ipreass_drain_tomax(void)
615 * Make sure each bucket is under the new limit. If
616 * necessary, drop enough of the oldest elements from
617 * each bucket to get under the new limit.
619 for (int i = 0; i < IPREASS_NHASH; i++) {
621 while (V_ipq[i].count > V_ipreass_maxbucketsize &&
622 (fp = TAILQ_LAST(&V_ipq[i].head, ipqhead)) != NULL)
623 ipq_timeout(&V_ipq[i], fp);
628 * If we are over the maximum number of fragments,
629 * drain off enough to get down to the new limit,
630 * stripping off last elements on queues. Every
631 * run we strip the oldest element from each bucket.
633 target = uma_zone_get_max(V_ipq_zone);
634 while (uma_zone_get_cur(V_ipq_zone) > target) {
635 for (int i = 0; i < IPREASS_NHASH; i++) {
637 fp = TAILQ_LAST(&V_ipq[i].head, ipqhead);
639 ipq_timeout(&V_ipq[i], fp);
646 ipreass_zone_change(void *tag)
648 VNET_ITERATOR_DECL(vnet_iter);
651 maxfrags = IP_MAXFRAGS;
652 max = IP_MAXFRAGPACKETS;
653 VNET_LIST_RLOCK_NOSLEEP();
654 VNET_FOREACH(vnet_iter) {
655 CURVNET_SET(vnet_iter);
656 max = uma_zone_set_max(V_ipq_zone, max);
657 V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
658 ipreass_drain_tomax();
661 VNET_LIST_RUNLOCK_NOSLEEP();
665 * Change the limit on the UMA zone, or disable the fragment allocation
666 * at all. Since 0 and -1 is a special values here, we need our own handler,
667 * instead of sysctl_handle_uma_zone_max().
670 sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS)
674 if (V_noreass == 0) {
675 max = uma_zone_get_max(V_ipq_zone);
680 error = sysctl_handle_int(oidp, &max, 0, req);
681 if (error || !req->newptr)
685 * XXXRW: Might be a good idea to sanity check the argument
686 * and place an extreme upper bound.
688 max = uma_zone_set_max(V_ipq_zone, max);
689 V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
690 ipreass_drain_tomax();
692 } else if (max == 0) {
695 } else if (max == -1) {
697 uma_zone_set_max(V_ipq_zone, 0);
698 V_ipreass_maxbucketsize = INT_MAX;
705 * Seek for old fragment queue header that can be reused. Try to
706 * reuse a header from currently locked hash bucket.
714 IPQ_LOCK_ASSERT(start);
716 for (i = 0; i < IPREASS_NHASH; i++) {
717 bucket = (start + i) % IPREASS_NHASH;
718 if (bucket != start && IPQ_TRYLOCK(bucket) == 0)
720 fp = TAILQ_LAST(&V_ipq[bucket].head, ipqhead);
724 IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
725 atomic_subtract_int(&nfrags, fp->ipq_nfrags);
726 while (fp->ipq_frags) {
728 fp->ipq_frags = m->m_nextpkt;
731 TAILQ_REMOVE(&V_ipq[bucket].head, fp, ipq_list);
732 V_ipq[bucket].count--;
740 IPQ_LOCK_ASSERT(start);
745 * Free a fragment reassembly header and all associated datagrams.
748 ipq_free(struct ipqbucket *bucket, struct ipq *fp)
752 atomic_subtract_int(&nfrags, fp->ipq_nfrags);
753 while (fp->ipq_frags) {
755 fp->ipq_frags = q->m_nextpkt;
758 TAILQ_REMOVE(&bucket->head, fp, ipq_list);
760 uma_zfree(V_ipq_zone, fp);
764 * Get or set the maximum number of reassembly queues per bucket.
767 sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS)
771 max = V_ipreass_maxbucketsize;
772 error = sysctl_handle_int(oidp, &max, 0, req);
773 if (error || !req->newptr)
777 V_ipreass_maxbucketsize = max;
778 ipreass_drain_tomax();