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 int sysctl_ipfrag_max_dist __read_mostly = 64;
57 struct inet_skb_parm h;
61 #define FRAG_CB(skb) ((struct ipfrag_skb_cb*)((skb)->cb))
63 /* Describe an entry in the "incomplete datagrams" queue. */
65 struct inet_frag_queue q;
74 struct inet_peer *peer;
77 struct inet_frags_ctl ip4_frags_ctl __read_mostly = {
79 * Fragment cache limits. We will commit 256K at one time. Should we
80 * cross that limit we will prune down to 192K. This should cope with
81 * even the most extreme cases without allowing an attacker to
82 * measurably harm machine performance.
84 .high_thresh = 256 * 1024,
85 .low_thresh = 192 * 1024,
88 * Important NOTE! Fragment queue must be destroyed before MSL expires.
89 * RFC791 is wrong proposing to prolongate timer each fragment arrival
92 .timeout = IP_FRAG_TIME,
93 .secret_interval = 10 * 60 * HZ,
96 static struct inet_frags ip4_frags;
98 int ip_frag_nqueues(void)
100 return ip4_frags.nqueues;
103 int ip_frag_mem(void)
105 return atomic_read(&ip4_frags.mem);
108 static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
109 struct net_device *dev);
111 static unsigned int ipqhashfn(__be16 id, __be32 saddr, __be32 daddr, u8 prot)
113 return jhash_3words((__force u32)id << 16 | prot,
114 (__force u32)saddr, (__force u32)daddr,
115 ip4_frags.rnd) & (INETFRAGS_HASHSZ - 1);
118 static unsigned int ip4_hashfn(struct inet_frag_queue *q)
122 ipq = container_of(q, struct ipq, q);
123 return ipqhashfn(ipq->id, ipq->saddr, ipq->daddr, ipq->protocol);
126 /* Memory Tracking Functions. */
127 static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
130 *work -= skb->truesize;
131 atomic_sub(skb->truesize, &ip4_frags.mem);
135 static __inline__ void ip4_frag_free(struct inet_frag_queue *q)
139 qp = container_of(q, struct ipq, q);
141 inet_putpeer(qp->peer);
145 static __inline__ struct ipq *frag_alloc_queue(void)
147 struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
151 atomic_add(sizeof(struct ipq), &ip4_frags.mem);
156 /* Destruction primitives. */
158 static __inline__ void ipq_put(struct ipq *ipq, int *work)
160 if (atomic_dec_and_test(&ipq->q.refcnt))
161 inet_frag_destroy(&ipq->q, &ip4_frags, work);
164 /* Kill ipq entry. It is not destroyed immediately,
165 * because caller (and someone more) holds reference count.
167 static void ipq_kill(struct ipq *ipq)
169 inet_frag_kill(&ipq->q, &ip4_frags);
172 /* Memory limiting on fragments. Evictor trashes the oldest
173 * fragment queue until we are back under the threshold.
175 static void ip_evictor(void)
179 evicted = inet_frag_evictor(&ip4_frags);
181 IP_ADD_STATS_BH(IPSTATS_MIB_REASMFAILS, evicted);
185 * Oops, a fragment queue timed out. Kill it and send an ICMP reply.
187 static void ip_expire(unsigned long arg)
189 struct ipq *qp = (struct ipq *) arg;
191 spin_lock(&qp->q.lock);
193 if (qp->q.last_in & COMPLETE)
198 IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT);
199 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
201 if ((qp->q.last_in&FIRST_IN) && qp->q.fragments != NULL) {
202 struct sk_buff *head = qp->q.fragments;
203 /* Send an ICMP "Fragment Reassembly Timeout" message. */
204 if ((head->dev = dev_get_by_index(&init_net, qp->iif)) != NULL) {
205 icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
210 spin_unlock(&qp->q.lock);
214 /* Creation primitives. */
216 static struct ipq *ip_frag_intern(struct ipq *qp_in)
220 struct hlist_node *n;
224 write_lock(&ip4_frags.lock);
225 hash = ipqhashfn(qp_in->id, qp_in->saddr, qp_in->daddr,
228 /* With SMP race we have to recheck hash table, because
229 * such entry could be created on other cpu, while we
230 * promoted read lock to write lock.
232 hlist_for_each_entry(qp, n, &ip4_frags.hash[hash], q.list) {
233 if (qp->id == qp_in->id &&
234 qp->saddr == qp_in->saddr &&
235 qp->daddr == qp_in->daddr &&
236 qp->protocol == qp_in->protocol &&
237 qp->user == qp_in->user) {
238 atomic_inc(&qp->q.refcnt);
239 write_unlock(&ip4_frags.lock);
240 qp_in->q.last_in |= COMPLETE;
241 ipq_put(qp_in, NULL);
248 if (!mod_timer(&qp->q.timer, jiffies + ip4_frags_ctl.timeout))
249 atomic_inc(&qp->q.refcnt);
251 atomic_inc(&qp->q.refcnt);
252 hlist_add_head(&qp->q.list, &ip4_frags.hash[hash]);
253 INIT_LIST_HEAD(&qp->q.lru_list);
254 list_add_tail(&qp->q.lru_list, &ip4_frags.lru_list);
256 write_unlock(&ip4_frags.lock);
260 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
261 static struct ipq *ip_frag_create(struct iphdr *iph, u32 user)
265 if ((qp = frag_alloc_queue()) == NULL)
268 qp->protocol = iph->protocol;
271 qp->saddr = iph->saddr;
272 qp->daddr = iph->daddr;
276 qp->q.fragments = NULL;
278 qp->peer = sysctl_ipfrag_max_dist ? inet_getpeer(iph->saddr, 1) : NULL;
280 /* Initialize a timer for this entry. */
281 init_timer(&qp->q.timer);
282 qp->q.timer.data = (unsigned long) qp; /* pointer to queue */
283 qp->q.timer.function = ip_expire; /* expire function */
284 spin_lock_init(&qp->q.lock);
285 atomic_set(&qp->q.refcnt, 1);
287 return ip_frag_intern(qp);
290 LIMIT_NETDEBUG(KERN_ERR "ip_frag_create: no memory left !\n");
294 /* Find the correct entry in the "incomplete datagrams" queue for
295 * this IP datagram, and create new one, if nothing is found.
297 static inline struct ipq *ip_find(struct iphdr *iph, u32 user)
300 __be32 saddr = iph->saddr;
301 __be32 daddr = iph->daddr;
302 __u8 protocol = iph->protocol;
305 struct hlist_node *n;
307 read_lock(&ip4_frags.lock);
308 hash = ipqhashfn(id, saddr, daddr, protocol);
309 hlist_for_each_entry(qp, n, &ip4_frags.hash[hash], q.list) {
311 qp->saddr == saddr &&
312 qp->daddr == daddr &&
313 qp->protocol == protocol &&
315 atomic_inc(&qp->q.refcnt);
316 read_unlock(&ip4_frags.lock);
320 read_unlock(&ip4_frags.lock);
322 return ip_frag_create(iph, user);
325 /* Is the fragment too far ahead to be part of ipq? */
326 static inline int ip_frag_too_far(struct ipq *qp)
328 struct inet_peer *peer = qp->peer;
329 unsigned int max = sysctl_ipfrag_max_dist;
330 unsigned int start, end;
338 end = atomic_inc_return(&peer->rid);
341 rc = qp->q.fragments && (end - start) > max;
344 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
350 static int ip_frag_reinit(struct ipq *qp)
354 if (!mod_timer(&qp->q.timer, jiffies + ip4_frags_ctl.timeout)) {
355 atomic_inc(&qp->q.refcnt);
359 fp = qp->q.fragments;
361 struct sk_buff *xp = fp->next;
362 frag_kfree_skb(fp, NULL);
369 qp->q.fragments = NULL;
375 /* Add new segment to existing queue. */
376 static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
378 struct sk_buff *prev, *next;
379 struct net_device *dev;
384 if (qp->q.last_in & COMPLETE)
387 if (!(IPCB(skb)->flags & IPSKB_FRAG_COMPLETE) &&
388 unlikely(ip_frag_too_far(qp)) &&
389 unlikely(err = ip_frag_reinit(qp))) {
394 offset = ntohs(ip_hdr(skb)->frag_off);
395 flags = offset & ~IP_OFFSET;
397 offset <<= 3; /* offset is in 8-byte chunks */
398 ihl = ip_hdrlen(skb);
400 /* Determine the position of this fragment. */
401 end = offset + skb->len - ihl;
404 /* Is this the final fragment? */
405 if ((flags & IP_MF) == 0) {
406 /* If we already have some bits beyond end
407 * or have different end, the segment is corrrupted.
409 if (end < qp->q.len ||
410 ((qp->q.last_in & LAST_IN) && end != qp->q.len))
412 qp->q.last_in |= LAST_IN;
417 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
418 skb->ip_summed = CHECKSUM_NONE;
420 if (end > qp->q.len) {
421 /* Some bits beyond end -> corruption. */
422 if (qp->q.last_in & LAST_IN)
431 if (pskb_pull(skb, ihl) == NULL)
434 err = pskb_trim_rcsum(skb, end - offset);
438 /* Find out which fragments are in front and at the back of us
439 * in the chain of fragments so far. We must know where to put
440 * this fragment, right?
443 for (next = qp->q.fragments; next != NULL; next = next->next) {
444 if (FRAG_CB(next)->offset >= offset)
449 /* We found where to put this one. Check for overlap with
450 * preceding fragment, and, if needed, align things so that
451 * any overlaps are eliminated.
454 int i = (FRAG_CB(prev)->offset + prev->len) - offset;
462 if (!pskb_pull(skb, i))
464 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
465 skb->ip_summed = CHECKSUM_NONE;
471 while (next && FRAG_CB(next)->offset < end) {
472 int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
475 /* Eat head of the next overlapped fragment
476 * and leave the loop. The next ones cannot overlap.
478 if (!pskb_pull(next, i))
480 FRAG_CB(next)->offset += i;
482 if (next->ip_summed != CHECKSUM_UNNECESSARY)
483 next->ip_summed = CHECKSUM_NONE;
486 struct sk_buff *free_it = next;
488 /* Old fragment is completely overridden with
496 qp->q.fragments = next;
498 qp->q.meat -= free_it->len;
499 frag_kfree_skb(free_it, NULL);
503 FRAG_CB(skb)->offset = offset;
505 /* Insert this fragment in the chain of fragments. */
510 qp->q.fragments = skb;
514 qp->iif = dev->ifindex;
517 qp->q.stamp = skb->tstamp;
518 qp->q.meat += skb->len;
519 atomic_add(skb->truesize, &ip4_frags.mem);
521 qp->q.last_in |= FIRST_IN;
523 if (qp->q.last_in == (FIRST_IN | LAST_IN) && qp->q.meat == qp->q.len)
524 return ip_frag_reasm(qp, prev, dev);
526 write_lock(&ip4_frags.lock);
527 list_move_tail(&qp->q.lru_list, &ip4_frags.lru_list);
528 write_unlock(&ip4_frags.lock);
537 /* Build a new IP datagram from all its fragments. */
539 static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
540 struct net_device *dev)
543 struct sk_buff *fp, *head = qp->q.fragments;
550 /* Make the one we just received the head. */
553 fp = skb_clone(head, GFP_ATOMIC);
558 fp->next = head->next;
561 skb_morph(head, qp->q.fragments);
562 head->next = qp->q.fragments->next;
564 kfree_skb(qp->q.fragments);
565 qp->q.fragments = head;
568 BUG_TRAP(head != NULL);
569 BUG_TRAP(FRAG_CB(head)->offset == 0);
571 /* Allocate a new buffer for the datagram. */
572 ihlen = ip_hdrlen(head);
573 len = ihlen + qp->q.len;
579 /* Head of list must not be cloned. */
581 if (skb_cloned(head) && pskb_expand_head(head, 0, 0, GFP_ATOMIC))
584 /* If the first fragment is fragmented itself, we split
585 * it to two chunks: the first with data and paged part
586 * and the second, holding only fragments. */
587 if (skb_shinfo(head)->frag_list) {
588 struct sk_buff *clone;
591 if ((clone = alloc_skb(0, GFP_ATOMIC)) == NULL)
593 clone->next = head->next;
595 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
596 skb_shinfo(head)->frag_list = NULL;
597 for (i=0; i<skb_shinfo(head)->nr_frags; i++)
598 plen += skb_shinfo(head)->frags[i].size;
599 clone->len = clone->data_len = head->data_len - plen;
600 head->data_len -= clone->len;
601 head->len -= clone->len;
603 clone->ip_summed = head->ip_summed;
604 atomic_add(clone->truesize, &ip4_frags.mem);
607 skb_shinfo(head)->frag_list = head->next;
608 skb_push(head, head->data - skb_network_header(head));
609 atomic_sub(head->truesize, &ip4_frags.mem);
611 for (fp=head->next; fp; fp = fp->next) {
612 head->data_len += fp->len;
613 head->len += fp->len;
614 if (head->ip_summed != fp->ip_summed)
615 head->ip_summed = CHECKSUM_NONE;
616 else if (head->ip_summed == CHECKSUM_COMPLETE)
617 head->csum = csum_add(head->csum, fp->csum);
618 head->truesize += fp->truesize;
619 atomic_sub(fp->truesize, &ip4_frags.mem);
624 head->tstamp = qp->q.stamp;
628 iph->tot_len = htons(len);
629 IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS);
630 qp->q.fragments = NULL;
634 LIMIT_NETDEBUG(KERN_ERR "IP: queue_glue: no memory for gluing "
640 "Oversized IP packet from %d.%d.%d.%d.\n",
643 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
647 /* Process an incoming IP datagram fragment. */
648 int ip_defrag(struct sk_buff *skb, u32 user)
652 IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS);
654 /* Start by cleaning up the memory. */
655 if (atomic_read(&ip4_frags.mem) > ip4_frags_ctl.high_thresh)
658 /* Lookup (or create) queue header */
659 if ((qp = ip_find(ip_hdr(skb), user)) != NULL) {
662 spin_lock(&qp->q.lock);
664 ret = ip_frag_queue(qp, skb);
666 spin_unlock(&qp->q.lock);
671 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
676 void __init ipfrag_init(void)
678 ip4_frags.ctl = &ip4_frags_ctl;
679 ip4_frags.hashfn = ip4_hashfn;
680 ip4_frags.destructor = ip4_frag_free;
681 ip4_frags.skb_free = NULL;
682 ip4_frags.qsize = sizeof(struct ipq);
683 inet_frags_init(&ip4_frags);
686 EXPORT_SYMBOL(ip_defrag);