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 6
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 static volatile u_int nfrags;
121 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfrags, CTLFLAG_RW,
123 "Maximum number of IPv4 fragments allowed across all reassembly queues");
124 SYSCTL_UINT(_net_inet_ip, OID_AUTO, curfrags, CTLFLAG_RD,
125 __DEVOLATILE(u_int *, &nfrags), 0,
126 "Current number of IPv4 fragments across all reassembly queues");
128 VNET_DEFINE_STATIC(uma_zone_t, ipq_zone);
129 #define V_ipq_zone VNET(ipq_zone)
130 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragpackets, CTLFLAG_VNET |
131 CTLTYPE_INT | CTLFLAG_RW, NULL, 0, sysctl_maxfragpackets, "I",
132 "Maximum number of IPv4 fragment reassembly queue entries");
133 SYSCTL_UMA_CUR(_net_inet_ip, OID_AUTO, fragpackets, CTLFLAG_VNET,
134 &VNET_NAME(ipq_zone),
135 "Current number of IPv4 fragment reassembly queue entries");
137 VNET_DEFINE_STATIC(int, noreass);
138 #define V_noreass VNET(noreass)
140 VNET_DEFINE_STATIC(int, maxfragsperpacket);
141 #define V_maxfragsperpacket VNET(maxfragsperpacket)
142 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfragsperpacket, CTLFLAG_VNET | CTLFLAG_RW,
143 &VNET_NAME(maxfragsperpacket), 0,
144 "Maximum number of IPv4 fragments allowed per packet");
145 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragbucketsize,
146 CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0,
147 sysctl_maxfragbucketsize, "I",
148 "Maximum number of IPv4 fragment reassembly queue entries per bucket");
151 * Take incoming datagram fragment and try to reassemble it into
152 * whole datagram. If the argument is the first fragment or one
153 * in between the function will return NULL and store the mbuf
154 * in the fragment chain. If the argument is the last fragment
155 * the packet will be reassembled and the pointer to the new
156 * mbuf returned for further processing. Only m_tags attached
157 * to the first packet/fragment are preserved.
158 * The IP header is *NOT* adjusted out of iplen.
160 #define M_IP_FRAG M_PROTO9
162 ip_reass(struct mbuf *m)
165 struct mbuf *p, *q, *nq, *t;
167 struct ipqhead *head;
168 int i, hlen, next, tmpmax;
170 uint32_t hash, hashkey[3];
172 uint32_t rss_hash, rss_type;
176 * If no reassembling or maxfragsperpacket are 0,
177 * never accept fragments.
178 * Also, drop packet if it would exceed the maximum
179 * number of fragments.
182 if (V_noreass == 1 || V_maxfragsperpacket == 0 ||
183 (tmpmax >= 0 && atomic_load_int(&nfrags) >= (u_int)tmpmax)) {
184 IPSTAT_INC(ips_fragments);
185 IPSTAT_INC(ips_fragdropped);
190 ip = mtod(m, struct ip *);
191 hlen = ip->ip_hl << 2;
194 * Adjust ip_len to not reflect header,
195 * convert offset of this to bytes.
197 ip->ip_len = htons(ntohs(ip->ip_len) - hlen);
198 if (ip->ip_off & htons(IP_MF)) {
200 * Make sure that fragments have a data length
201 * that's a non-zero multiple of 8 bytes.
203 if (ip->ip_len == htons(0) || (ntohs(ip->ip_len) & 0x7) != 0) {
204 IPSTAT_INC(ips_toosmall); /* XXX */
205 IPSTAT_INC(ips_fragdropped);
209 m->m_flags |= M_IP_FRAG;
211 m->m_flags &= ~M_IP_FRAG;
212 ip->ip_off = htons(ntohs(ip->ip_off) << 3);
215 * Attempt reassembly; if it succeeds, proceed.
216 * ip_reass() will return a different mbuf.
218 IPSTAT_INC(ips_fragments);
219 m->m_pkthdr.PH_loc.ptr = ip;
222 * Presence of header sizes in mbufs
223 * would confuse code below.
228 hashkey[0] = ip->ip_src.s_addr;
229 hashkey[1] = ip->ip_dst.s_addr;
230 hashkey[2] = (uint32_t)ip->ip_p << 16;
231 hashkey[2] += ip->ip_id;
232 hash = jenkins_hash32(hashkey, nitems(hashkey), V_ipq_hashseed);
233 hash &= IPREASS_HMASK;
234 head = &V_ipq[hash].head;
238 * Look for queue of fragments
241 TAILQ_FOREACH(fp, head, ipq_list)
242 if (ip->ip_id == fp->ipq_id &&
243 ip->ip_src.s_addr == fp->ipq_src.s_addr &&
244 ip->ip_dst.s_addr == fp->ipq_dst.s_addr &&
246 mac_ipq_match(m, fp) &&
248 ip->ip_p == fp->ipq_p)
251 * If first fragment to arrive, create a reassembly queue.
254 if (V_ipq[hash].count < V_ipreass_maxbucketsize)
255 fp = uma_zalloc(V_ipq_zone, M_NOWAIT);
257 fp = ipq_reuse(hash);
261 if (mac_ipq_init(fp, M_NOWAIT) != 0) {
262 uma_zfree(V_ipq_zone, fp);
266 mac_ipq_create(m, fp);
268 TAILQ_INSERT_HEAD(head, fp, ipq_list);
271 atomic_add_int(&nfrags, 1);
272 fp->ipq_ttl = IPFRAGTTL;
273 fp->ipq_p = ip->ip_p;
274 fp->ipq_id = ip->ip_id;
275 fp->ipq_src = ip->ip_src;
276 fp->ipq_dst = ip->ip_dst;
282 atomic_add_int(&nfrags, 1);
284 mac_ipq_update(m, fp);
288 #define GETIP(m) ((struct ip*)((m)->m_pkthdr.PH_loc.ptr))
291 * Handle ECN by comparing this segment with the first one;
292 * if CE is set, do not lose CE.
293 * drop if CE and not-ECT are mixed for the same packet.
295 ecn = ip->ip_tos & IPTOS_ECN_MASK;
296 ecn0 = GETIP(fp->ipq_frags)->ip_tos & IPTOS_ECN_MASK;
297 if (ecn == IPTOS_ECN_CE) {
298 if (ecn0 == IPTOS_ECN_NOTECT)
300 if (ecn0 != IPTOS_ECN_CE)
301 GETIP(fp->ipq_frags)->ip_tos |= IPTOS_ECN_CE;
303 if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT)
307 * Find a segment which begins after this one does.
309 for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt)
310 if (ntohs(GETIP(q)->ip_off) > ntohs(ip->ip_off))
314 * If there is a preceding segment, it may provide some of
315 * our data already. If so, drop the data from the incoming
316 * segment. If it provides all of our data, drop us, otherwise
317 * stick new segment in the proper place.
319 * If some of the data is dropped from the preceding
320 * segment, then it's checksum is invalidated.
323 i = ntohs(GETIP(p)->ip_off) + ntohs(GETIP(p)->ip_len) -
326 if (i >= ntohs(ip->ip_len))
329 m->m_pkthdr.csum_flags = 0;
330 ip->ip_off = htons(ntohs(ip->ip_off) + i);
331 ip->ip_len = htons(ntohs(ip->ip_len) - i);
333 m->m_nextpkt = p->m_nextpkt;
336 m->m_nextpkt = fp->ipq_frags;
341 * While we overlap succeeding segments trim them or,
342 * if they are completely covered, dequeue them.
344 for (; q != NULL && ntohs(ip->ip_off) + ntohs(ip->ip_len) >
345 ntohs(GETIP(q)->ip_off); q = nq) {
346 i = (ntohs(ip->ip_off) + ntohs(ip->ip_len)) -
347 ntohs(GETIP(q)->ip_off);
348 if (i < ntohs(GETIP(q)->ip_len)) {
349 GETIP(q)->ip_len = htons(ntohs(GETIP(q)->ip_len) - i);
350 GETIP(q)->ip_off = htons(ntohs(GETIP(q)->ip_off) + i);
352 q->m_pkthdr.csum_flags = 0;
357 IPSTAT_INC(ips_fragdropped);
359 atomic_subtract_int(&nfrags, 1);
364 * Check for complete reassembly and perform frag per packet
367 * Frag limiting is performed here so that the nth frag has
368 * a chance to complete the packet before we drop the packet.
369 * As a result, n+1 frags are actually allowed per packet, but
370 * only n will ever be stored. (n = maxfragsperpacket.)
374 for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt) {
375 if (ntohs(GETIP(q)->ip_off) != next) {
376 if (fp->ipq_nfrags > V_maxfragsperpacket)
377 ipq_drop(&V_ipq[hash], fp);
380 next += ntohs(GETIP(q)->ip_len);
382 /* Make sure the last packet didn't have the IP_MF flag */
383 if (p->m_flags & M_IP_FRAG) {
384 if (fp->ipq_nfrags > V_maxfragsperpacket)
385 ipq_drop(&V_ipq[hash], fp);
390 * Reassembly is complete. Make sure the packet is a sane size.
394 if (next + (ip->ip_hl << 2) > IP_MAXPACKET) {
395 IPSTAT_INC(ips_toolong);
396 ipq_drop(&V_ipq[hash], fp);
401 * Concatenate fragments.
409 for (q = nq; q != NULL; q = nq) {
412 m->m_pkthdr.csum_flags &= q->m_pkthdr.csum_flags;
413 m->m_pkthdr.csum_data += q->m_pkthdr.csum_data;
418 * In order to do checksumming faster we do 'end-around carry' here
419 * (and not in for{} loop), though it implies we are not going to
420 * reassemble more than 64k fragments.
422 while (m->m_pkthdr.csum_data & 0xffff0000)
423 m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) +
424 (m->m_pkthdr.csum_data >> 16);
425 atomic_subtract_int(&nfrags, fp->ipq_nfrags);
427 mac_ipq_reassemble(fp, m);
432 * Create header for new ip packet by modifying header of first
433 * packet; dequeue and discard fragment reassembly header.
434 * Make header visible.
436 ip->ip_len = htons((ip->ip_hl << 2) + next);
437 ip->ip_src = fp->ipq_src;
438 ip->ip_dst = fp->ipq_dst;
439 TAILQ_REMOVE(head, fp, ipq_list);
441 uma_zfree(V_ipq_zone, fp);
442 m->m_len += (ip->ip_hl << 2);
443 m->m_data -= (ip->ip_hl << 2);
444 /* some debugging cruft by sklower, below, will go away soon */
445 if (m->m_flags & M_PKTHDR) /* XXX this should be done elsewhere */
447 IPSTAT_INC(ips_reassembled);
452 * Query the RSS layer for the flowid / flowtype for the
455 * For now, just assume we have to calculate a new one.
456 * Later on we should check to see if the assigned flowid matches
457 * what RSS wants for the given IP protocol and if so, just keep it.
459 * We then queue into the relevant netisr so it can be dispatched
460 * to the correct CPU.
462 * Note - this may return 1, which means the flowid in the mbuf
463 * is correct for the configured RSS hash types and can be used.
465 if (rss_mbuf_software_hash_v4(m, 0, &rss_hash, &rss_type) == 0) {
466 m->m_pkthdr.flowid = rss_hash;
467 M_HASHTYPE_SET(m, rss_type);
471 * Queue/dispatch for reprocessing.
473 * Note: this is much slower than just handling the frame in the
474 * current receive context. It's likely worth investigating
477 netisr_dispatch(NETISR_IP_DIRECT, m);
485 IPSTAT_INC(ips_fragdropped);
488 atomic_subtract_int(&nfrags, 1);
499 * Initialize IP reassembly structures.
506 for (int i = 0; i < IPREASS_NHASH; i++) {
507 TAILQ_INIT(&V_ipq[i].head);
508 mtx_init(&V_ipq[i].lock, "IP reassembly", NULL,
509 MTX_DEF | MTX_DUPOK);
512 V_ipq_hashseed = arc4random();
513 V_maxfragsperpacket = 16;
514 V_ipq_zone = uma_zcreate("ipq", sizeof(struct ipq), NULL, NULL, NULL,
515 NULL, UMA_ALIGN_PTR, 0);
516 max = nmbclusters / 32;
517 max = uma_zone_set_max(V_ipq_zone, max);
518 V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
520 if (IS_DEFAULT_VNET(curvnet)) {
521 maxfrags = nmbclusters / 32;
522 EVENTHANDLER_REGISTER(nmbclusters_change, ipreass_zone_change,
523 NULL, EVENTHANDLER_PRI_ANY);
528 * If a timer expires on a reassembly queue, discard it.
531 ipreass_slowtimo(void)
533 struct ipq *fp, *tmp;
535 for (int i = 0; i < IPREASS_NHASH; i++) {
537 TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, tmp)
538 if (--fp->ipq_ttl == 0)
539 ipq_timeout(&V_ipq[i], fp);
545 * Drain off all datagram fragments.
551 for (int i = 0; i < IPREASS_NHASH; i++) {
553 while(!TAILQ_EMPTY(&V_ipq[i].head))
554 ipq_drop(&V_ipq[i], TAILQ_FIRST(&V_ipq[i].head));
555 KASSERT(V_ipq[i].count == 0,
556 ("%s: V_ipq[%d] count %d (V_ipq=%p)", __func__, i,
557 V_ipq[i].count, V_ipq));
564 * Destroy IP reassembly structures.
567 ipreass_destroy(void)
571 uma_zdestroy(V_ipq_zone);
572 for (int i = 0; i < IPREASS_NHASH; i++)
573 mtx_destroy(&V_ipq[i].lock);
578 * After maxnipq has been updated, propagate the change to UMA. The UMA zone
579 * max has slightly different semantics than the sysctl, for historical
583 ipreass_drain_tomax(void)
589 * Make sure each bucket is under the new limit. If
590 * necessary, drop enough of the oldest elements from
591 * each bucket to get under the new limit.
593 for (int i = 0; i < IPREASS_NHASH; i++) {
595 while (V_ipq[i].count > V_ipreass_maxbucketsize &&
596 (fp = TAILQ_LAST(&V_ipq[i].head, ipqhead)) != NULL)
597 ipq_timeout(&V_ipq[i], fp);
602 * If we are over the maximum number of fragments,
603 * drain off enough to get down to the new limit,
604 * stripping off last elements on queues. Every
605 * run we strip the oldest element from each bucket.
607 target = uma_zone_get_max(V_ipq_zone);
608 while (uma_zone_get_cur(V_ipq_zone) > target) {
609 for (int i = 0; i < IPREASS_NHASH; i++) {
611 fp = TAILQ_LAST(&V_ipq[i].head, ipqhead);
613 ipq_timeout(&V_ipq[i], fp);
620 ipreass_zone_change(void *tag)
622 VNET_ITERATOR_DECL(vnet_iter);
625 maxfrags = nmbclusters / 32;
626 max = nmbclusters / 32;
627 VNET_LIST_RLOCK_NOSLEEP();
628 VNET_FOREACH(vnet_iter) {
629 CURVNET_SET(vnet_iter);
630 max = uma_zone_set_max(V_ipq_zone, max);
631 V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
632 ipreass_drain_tomax();
635 VNET_LIST_RUNLOCK_NOSLEEP();
639 * Change the limit on the UMA zone, or disable the fragment allocation
640 * at all. Since 0 and -1 is a special values here, we need our own handler,
641 * instead of sysctl_handle_uma_zone_max().
644 sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS)
648 if (V_noreass == 0) {
649 max = uma_zone_get_max(V_ipq_zone);
654 error = sysctl_handle_int(oidp, &max, 0, req);
655 if (error || !req->newptr)
659 * XXXRW: Might be a good idea to sanity check the argument
660 * and place an extreme upper bound.
662 max = uma_zone_set_max(V_ipq_zone, max);
663 V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1);
664 ipreass_drain_tomax();
666 } else if (max == 0) {
669 } else if (max == -1) {
671 uma_zone_set_max(V_ipq_zone, 0);
672 V_ipreass_maxbucketsize = INT_MAX;
679 * Seek for old fragment queue header that can be reused. Try to
680 * reuse a header from currently locked hash bucket.
688 IPQ_LOCK_ASSERT(start);
690 for (i = 0; i < IPREASS_NHASH; i++) {
691 bucket = (start + i) % IPREASS_NHASH;
692 if (bucket != start && IPQ_TRYLOCK(bucket) == 0)
694 fp = TAILQ_LAST(&V_ipq[bucket].head, ipqhead);
698 IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags);
699 atomic_subtract_int(&nfrags, fp->ipq_nfrags);
700 while (fp->ipq_frags) {
702 fp->ipq_frags = m->m_nextpkt;
705 TAILQ_REMOVE(&V_ipq[bucket].head, fp, ipq_list);
706 V_ipq[bucket].count--;
714 IPQ_LOCK_ASSERT(start);
719 * Free a fragment reassembly header and all associated datagrams.
722 ipq_free(struct ipqbucket *bucket, struct ipq *fp)
726 atomic_subtract_int(&nfrags, fp->ipq_nfrags);
727 while (fp->ipq_frags) {
729 fp->ipq_frags = q->m_nextpkt;
732 TAILQ_REMOVE(&bucket->head, fp, ipq_list);
734 uma_zfree(V_ipq_zone, fp);
738 * Get or set the maximum number of reassembly queues per bucket.
741 sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS)
745 max = V_ipreass_maxbucketsize;
746 error = sysctl_handle_int(oidp, &max, 0, req);
747 if (error || !req->newptr)
751 V_ipreass_maxbucketsize = max;
752 ipreass_drain_tomax();