2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * The IP fragmentation functionality.
8 * Version: $Id: ip_fragment.c,v 1.59 2002/01/12 07:54:56 davem Exp $
10 * Authors: Fred N. van Kempen <waltje@uWalt.NL.Mugnet.ORG>
11 * Alan Cox <Alan.Cox@linux.org>
14 * Alan Cox : Split from ip.c , see ip_input.c for history.
15 * David S. Miller : Begin massive cleanup...
16 * Andi Kleen : Add sysctls.
17 * xxxx : Overlapfrag bug.
18 * Ultima : ip_expire() kernel panic.
19 * Bill Hawes : Frag accounting and evictor fixes.
20 * John McDonald : 0 length frag bug.
21 * Alexey Kuznetsov: SMP races, threading, cleanup.
22 * Patrick McHardy : LRU queue of frag heads for evictor.
25 #include <linux/compiler.h>
26 #include <linux/module.h>
27 #include <linux/types.h>
29 #include <linux/jiffies.h>
30 #include <linux/skbuff.h>
31 #include <linux/list.h>
33 #include <linux/icmp.h>
34 #include <linux/netdevice.h>
35 #include <linux/jhash.h>
36 #include <linux/random.h>
40 #include <net/checksum.h>
41 #include <net/inetpeer.h>
42 #include <net/inet_frag.h>
43 #include <linux/tcp.h>
44 #include <linux/udp.h>
45 #include <linux/inet.h>
46 #include <linux/netfilter_ipv4.h>
48 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
49 * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
50 * as well. Or notify me, at least. --ANK
53 /* Fragment cache limits. We will commit 256K at one time. Should we
54 * cross that limit we will prune down to 192K. This should cope with
55 * even the most extreme cases without allowing an attacker to measurably
56 * harm machine performance.
58 int sysctl_ipfrag_high_thresh __read_mostly = 256*1024;
59 int sysctl_ipfrag_low_thresh __read_mostly = 192*1024;
61 int sysctl_ipfrag_max_dist __read_mostly = 64;
63 /* Important NOTE! Fragment queue must be destroyed before MSL expires.
64 * RFC791 is wrong proposing to prolongate timer each fragment arrival by TTL.
66 int sysctl_ipfrag_time __read_mostly = IP_FRAG_TIME;
70 struct inet_skb_parm h;
74 #define FRAG_CB(skb) ((struct ipfrag_skb_cb*)((skb)->cb))
76 /* Describe an entry in the "incomplete datagrams" queue. */
78 struct inet_frag_queue q;
87 struct inet_peer *peer;
90 static struct inet_frags ip4_frags;
92 int ip_frag_nqueues(void)
94 return ip4_frags.nqueues;
99 return atomic_read(&ip4_frags.mem);
102 static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
103 struct net_device *dev);
105 static __inline__ void __ipq_unlink(struct ipq *qp)
107 hlist_del(&qp->q.list);
108 list_del(&qp->q.lru_list);
112 static __inline__ void ipq_unlink(struct ipq *ipq)
114 write_lock(&ip4_frags.lock);
116 write_unlock(&ip4_frags.lock);
119 static unsigned int ipqhashfn(__be16 id, __be32 saddr, __be32 daddr, u8 prot)
121 return jhash_3words((__force u32)id << 16 | prot,
122 (__force u32)saddr, (__force u32)daddr,
123 ip4_frags.rnd) & (INETFRAGS_HASHSZ - 1);
126 int sysctl_ipfrag_secret_interval __read_mostly = 10 * 60 * HZ;
128 static void ipfrag_secret_rebuild(unsigned long dummy)
130 unsigned long now = jiffies;
133 write_lock(&ip4_frags.lock);
134 get_random_bytes(&ip4_frags.rnd, sizeof(u32));
135 for (i = 0; i < INETFRAGS_HASHSZ; i++) {
137 struct hlist_node *p, *n;
139 hlist_for_each_entry_safe(q, p, n, &ip4_frags.hash[i], q.list) {
140 unsigned int hval = ipqhashfn(q->id, q->saddr,
141 q->daddr, q->protocol);
144 hlist_del(&q->q.list);
146 /* Relink to new hash chain. */
147 hlist_add_head(&q->q.list, &ip4_frags.hash[hval]);
151 write_unlock(&ip4_frags.lock);
153 mod_timer(&ip4_frags.secret_timer, now + sysctl_ipfrag_secret_interval);
156 /* Memory Tracking Functions. */
157 static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
160 *work -= skb->truesize;
161 atomic_sub(skb->truesize, &ip4_frags.mem);
165 static __inline__ void frag_free_queue(struct ipq *qp, int *work)
168 *work -= sizeof(struct ipq);
169 atomic_sub(sizeof(struct ipq), &ip4_frags.mem);
173 static __inline__ struct ipq *frag_alloc_queue(void)
175 struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
179 atomic_add(sizeof(struct ipq), &ip4_frags.mem);
184 /* Destruction primitives. */
186 /* Complete destruction of ipq. */
187 static void ip_frag_destroy(struct ipq *qp, int *work)
191 BUG_TRAP(qp->q.last_in&COMPLETE);
192 BUG_TRAP(del_timer(&qp->q.timer) == 0);
195 inet_putpeer(qp->peer);
197 /* Release all fragment data. */
198 fp = qp->q.fragments;
200 struct sk_buff *xp = fp->next;
202 frag_kfree_skb(fp, work);
206 /* Finally, release the queue descriptor itself. */
207 frag_free_queue(qp, work);
210 static __inline__ void ipq_put(struct ipq *ipq, int *work)
212 if (atomic_dec_and_test(&ipq->q.refcnt))
213 ip_frag_destroy(ipq, work);
216 /* Kill ipq entry. It is not destroyed immediately,
217 * because caller (and someone more) holds reference count.
219 static void ipq_kill(struct ipq *ipq)
221 if (del_timer(&ipq->q.timer))
222 atomic_dec(&ipq->q.refcnt);
224 if (!(ipq->q.last_in & COMPLETE)) {
226 atomic_dec(&ipq->q.refcnt);
227 ipq->q.last_in |= COMPLETE;
231 /* Memory limiting on fragments. Evictor trashes the oldest
232 * fragment queue until we are back under the threshold.
234 static void ip_evictor(void)
237 struct list_head *tmp;
240 work = atomic_read(&ip4_frags.mem) - sysctl_ipfrag_low_thresh;
245 read_lock(&ip4_frags.lock);
246 if (list_empty(&ip4_frags.lru_list)) {
247 read_unlock(&ip4_frags.lock);
250 tmp = ip4_frags.lru_list.next;
251 qp = list_entry(tmp, struct ipq, q.lru_list);
252 atomic_inc(&qp->q.refcnt);
253 read_unlock(&ip4_frags.lock);
255 spin_lock(&qp->q.lock);
256 if (!(qp->q.last_in&COMPLETE))
258 spin_unlock(&qp->q.lock);
261 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
266 * Oops, a fragment queue timed out. Kill it and send an ICMP reply.
268 static void ip_expire(unsigned long arg)
270 struct ipq *qp = (struct ipq *) arg;
272 spin_lock(&qp->q.lock);
274 if (qp->q.last_in & COMPLETE)
279 IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT);
280 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
282 if ((qp->q.last_in&FIRST_IN) && qp->q.fragments != NULL) {
283 struct sk_buff *head = qp->q.fragments;
284 /* Send an ICMP "Fragment Reassembly Timeout" message. */
285 if ((head->dev = dev_get_by_index(&init_net, qp->iif)) != NULL) {
286 icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
291 spin_unlock(&qp->q.lock);
295 /* Creation primitives. */
297 static struct ipq *ip_frag_intern(struct ipq *qp_in)
301 struct hlist_node *n;
305 write_lock(&ip4_frags.lock);
306 hash = ipqhashfn(qp_in->id, qp_in->saddr, qp_in->daddr,
309 /* With SMP race we have to recheck hash table, because
310 * such entry could be created on other cpu, while we
311 * promoted read lock to write lock.
313 hlist_for_each_entry(qp, n, &ip4_frags.hash[hash], q.list) {
314 if (qp->id == qp_in->id &&
315 qp->saddr == qp_in->saddr &&
316 qp->daddr == qp_in->daddr &&
317 qp->protocol == qp_in->protocol &&
318 qp->user == qp_in->user) {
319 atomic_inc(&qp->q.refcnt);
320 write_unlock(&ip4_frags.lock);
321 qp_in->q.last_in |= COMPLETE;
322 ipq_put(qp_in, NULL);
329 if (!mod_timer(&qp->q.timer, jiffies + sysctl_ipfrag_time))
330 atomic_inc(&qp->q.refcnt);
332 atomic_inc(&qp->q.refcnt);
333 hlist_add_head(&qp->q.list, &ip4_frags.hash[hash]);
334 INIT_LIST_HEAD(&qp->q.lru_list);
335 list_add_tail(&qp->q.lru_list, &ip4_frags.lru_list);
337 write_unlock(&ip4_frags.lock);
341 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
342 static struct ipq *ip_frag_create(struct iphdr *iph, u32 user)
346 if ((qp = frag_alloc_queue()) == NULL)
349 qp->protocol = iph->protocol;
352 qp->saddr = iph->saddr;
353 qp->daddr = iph->daddr;
357 qp->q.fragments = NULL;
359 qp->peer = sysctl_ipfrag_max_dist ? inet_getpeer(iph->saddr, 1) : NULL;
361 /* Initialize a timer for this entry. */
362 init_timer(&qp->q.timer);
363 qp->q.timer.data = (unsigned long) qp; /* pointer to queue */
364 qp->q.timer.function = ip_expire; /* expire function */
365 spin_lock_init(&qp->q.lock);
366 atomic_set(&qp->q.refcnt, 1);
368 return ip_frag_intern(qp);
371 LIMIT_NETDEBUG(KERN_ERR "ip_frag_create: no memory left !\n");
375 /* Find the correct entry in the "incomplete datagrams" queue for
376 * this IP datagram, and create new one, if nothing is found.
378 static inline struct ipq *ip_find(struct iphdr *iph, u32 user)
381 __be32 saddr = iph->saddr;
382 __be32 daddr = iph->daddr;
383 __u8 protocol = iph->protocol;
386 struct hlist_node *n;
388 read_lock(&ip4_frags.lock);
389 hash = ipqhashfn(id, saddr, daddr, protocol);
390 hlist_for_each_entry(qp, n, &ip4_frags.hash[hash], q.list) {
392 qp->saddr == saddr &&
393 qp->daddr == daddr &&
394 qp->protocol == protocol &&
396 atomic_inc(&qp->q.refcnt);
397 read_unlock(&ip4_frags.lock);
401 read_unlock(&ip4_frags.lock);
403 return ip_frag_create(iph, user);
406 /* Is the fragment too far ahead to be part of ipq? */
407 static inline int ip_frag_too_far(struct ipq *qp)
409 struct inet_peer *peer = qp->peer;
410 unsigned int max = sysctl_ipfrag_max_dist;
411 unsigned int start, end;
419 end = atomic_inc_return(&peer->rid);
422 rc = qp->q.fragments && (end - start) > max;
425 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
431 static int ip_frag_reinit(struct ipq *qp)
435 if (!mod_timer(&qp->q.timer, jiffies + sysctl_ipfrag_time)) {
436 atomic_inc(&qp->q.refcnt);
440 fp = qp->q.fragments;
442 struct sk_buff *xp = fp->next;
443 frag_kfree_skb(fp, NULL);
450 qp->q.fragments = NULL;
456 /* Add new segment to existing queue. */
457 static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
459 struct sk_buff *prev, *next;
460 struct net_device *dev;
465 if (qp->q.last_in & COMPLETE)
468 if (!(IPCB(skb)->flags & IPSKB_FRAG_COMPLETE) &&
469 unlikely(ip_frag_too_far(qp)) &&
470 unlikely(err = ip_frag_reinit(qp))) {
475 offset = ntohs(ip_hdr(skb)->frag_off);
476 flags = offset & ~IP_OFFSET;
478 offset <<= 3; /* offset is in 8-byte chunks */
479 ihl = ip_hdrlen(skb);
481 /* Determine the position of this fragment. */
482 end = offset + skb->len - ihl;
485 /* Is this the final fragment? */
486 if ((flags & IP_MF) == 0) {
487 /* If we already have some bits beyond end
488 * or have different end, the segment is corrrupted.
490 if (end < qp->q.len ||
491 ((qp->q.last_in & LAST_IN) && end != qp->q.len))
493 qp->q.last_in |= LAST_IN;
498 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
499 skb->ip_summed = CHECKSUM_NONE;
501 if (end > qp->q.len) {
502 /* Some bits beyond end -> corruption. */
503 if (qp->q.last_in & LAST_IN)
512 if (pskb_pull(skb, ihl) == NULL)
515 err = pskb_trim_rcsum(skb, end - offset);
519 /* Find out which fragments are in front and at the back of us
520 * in the chain of fragments so far. We must know where to put
521 * this fragment, right?
524 for (next = qp->q.fragments; next != NULL; next = next->next) {
525 if (FRAG_CB(next)->offset >= offset)
530 /* We found where to put this one. Check for overlap with
531 * preceding fragment, and, if needed, align things so that
532 * any overlaps are eliminated.
535 int i = (FRAG_CB(prev)->offset + prev->len) - offset;
543 if (!pskb_pull(skb, i))
545 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
546 skb->ip_summed = CHECKSUM_NONE;
552 while (next && FRAG_CB(next)->offset < end) {
553 int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
556 /* Eat head of the next overlapped fragment
557 * and leave the loop. The next ones cannot overlap.
559 if (!pskb_pull(next, i))
561 FRAG_CB(next)->offset += i;
563 if (next->ip_summed != CHECKSUM_UNNECESSARY)
564 next->ip_summed = CHECKSUM_NONE;
567 struct sk_buff *free_it = next;
569 /* Old fragment is completely overridden with
577 qp->q.fragments = next;
579 qp->q.meat -= free_it->len;
580 frag_kfree_skb(free_it, NULL);
584 FRAG_CB(skb)->offset = offset;
586 /* Insert this fragment in the chain of fragments. */
591 qp->q.fragments = skb;
595 qp->iif = dev->ifindex;
598 qp->q.stamp = skb->tstamp;
599 qp->q.meat += skb->len;
600 atomic_add(skb->truesize, &ip4_frags.mem);
602 qp->q.last_in |= FIRST_IN;
604 if (qp->q.last_in == (FIRST_IN | LAST_IN) && qp->q.meat == qp->q.len)
605 return ip_frag_reasm(qp, prev, dev);
607 write_lock(&ip4_frags.lock);
608 list_move_tail(&qp->q.lru_list, &ip4_frags.lru_list);
609 write_unlock(&ip4_frags.lock);
618 /* Build a new IP datagram from all its fragments. */
620 static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
621 struct net_device *dev)
624 struct sk_buff *fp, *head = qp->q.fragments;
631 /* Make the one we just received the head. */
634 fp = skb_clone(head, GFP_ATOMIC);
639 fp->next = head->next;
642 skb_morph(head, qp->q.fragments);
643 head->next = qp->q.fragments->next;
645 kfree_skb(qp->q.fragments);
646 qp->q.fragments = head;
649 BUG_TRAP(head != NULL);
650 BUG_TRAP(FRAG_CB(head)->offset == 0);
652 /* Allocate a new buffer for the datagram. */
653 ihlen = ip_hdrlen(head);
654 len = ihlen + qp->q.len;
660 /* Head of list must not be cloned. */
662 if (skb_cloned(head) && pskb_expand_head(head, 0, 0, GFP_ATOMIC))
665 /* If the first fragment is fragmented itself, we split
666 * it to two chunks: the first with data and paged part
667 * and the second, holding only fragments. */
668 if (skb_shinfo(head)->frag_list) {
669 struct sk_buff *clone;
672 if ((clone = alloc_skb(0, GFP_ATOMIC)) == NULL)
674 clone->next = head->next;
676 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
677 skb_shinfo(head)->frag_list = NULL;
678 for (i=0; i<skb_shinfo(head)->nr_frags; i++)
679 plen += skb_shinfo(head)->frags[i].size;
680 clone->len = clone->data_len = head->data_len - plen;
681 head->data_len -= clone->len;
682 head->len -= clone->len;
684 clone->ip_summed = head->ip_summed;
685 atomic_add(clone->truesize, &ip4_frags.mem);
688 skb_shinfo(head)->frag_list = head->next;
689 skb_push(head, head->data - skb_network_header(head));
690 atomic_sub(head->truesize, &ip4_frags.mem);
692 for (fp=head->next; fp; fp = fp->next) {
693 head->data_len += fp->len;
694 head->len += fp->len;
695 if (head->ip_summed != fp->ip_summed)
696 head->ip_summed = CHECKSUM_NONE;
697 else if (head->ip_summed == CHECKSUM_COMPLETE)
698 head->csum = csum_add(head->csum, fp->csum);
699 head->truesize += fp->truesize;
700 atomic_sub(fp->truesize, &ip4_frags.mem);
705 head->tstamp = qp->q.stamp;
709 iph->tot_len = htons(len);
710 IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS);
711 qp->q.fragments = NULL;
715 LIMIT_NETDEBUG(KERN_ERR "IP: queue_glue: no memory for gluing "
721 "Oversized IP packet from %d.%d.%d.%d.\n",
724 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
728 /* Process an incoming IP datagram fragment. */
729 int ip_defrag(struct sk_buff *skb, u32 user)
733 IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS);
735 /* Start by cleaning up the memory. */
736 if (atomic_read(&ip4_frags.mem) > sysctl_ipfrag_high_thresh)
739 /* Lookup (or create) queue header */
740 if ((qp = ip_find(ip_hdr(skb), user)) != NULL) {
743 spin_lock(&qp->q.lock);
745 ret = ip_frag_queue(qp, skb);
747 spin_unlock(&qp->q.lock);
752 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
757 void __init ipfrag_init(void)
759 init_timer(&ip4_frags.secret_timer);
760 ip4_frags.secret_timer.function = ipfrag_secret_rebuild;
761 ip4_frags.secret_timer.expires = jiffies + sysctl_ipfrag_secret_interval;
762 add_timer(&ip4_frags.secret_timer);
764 inet_frags_init(&ip4_frags);
767 EXPORT_SYMBOL(ip_defrag);