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/config.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 <linux/tcp.h>
42 #include <linux/udp.h>
43 #include <linux/inet.h>
44 #include <linux/netfilter_ipv4.h>
46 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
47 * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
48 * as well. Or notify me, at least. --ANK
51 /* Fragment cache limits. We will commit 256K at one time. Should we
52 * cross that limit we will prune down to 192K. This should cope with
53 * even the most extreme cases without allowing an attacker to measurably
54 * harm machine performance.
56 int sysctl_ipfrag_high_thresh = 256*1024;
57 int sysctl_ipfrag_low_thresh = 192*1024;
59 /* Important NOTE! Fragment queue must be destroyed before MSL expires.
60 * RFC791 is wrong proposing to prolongate timer each fragment arrival by TTL.
62 int sysctl_ipfrag_time = IP_FRAG_TIME;
66 struct inet_skb_parm h;
70 #define FRAG_CB(skb) ((struct ipfrag_skb_cb*)((skb)->cb))
72 /* Describe an entry in the "incomplete datagrams" queue. */
74 struct hlist_node list;
75 struct list_head lru_list; /* lru list member */
86 struct sk_buff *fragments; /* linked list of received fragments */
87 int len; /* total length of original datagram */
91 struct timer_list timer; /* when will this queue expire? */
100 /* Per-bucket lock is easy to add now. */
101 static struct hlist_head ipq_hash[IPQ_HASHSZ];
102 static DEFINE_RWLOCK(ipfrag_lock);
103 static u32 ipfrag_hash_rnd;
104 static LIST_HEAD(ipq_lru_list);
105 int ip_frag_nqueues = 0;
107 static __inline__ void __ipq_unlink(struct ipq *qp)
109 hlist_del(&qp->list);
110 list_del(&qp->lru_list);
114 static __inline__ void ipq_unlink(struct ipq *ipq)
116 write_lock(&ipfrag_lock);
118 write_unlock(&ipfrag_lock);
121 static unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
123 return jhash_3words((u32)id << 16 | prot, saddr, daddr,
124 ipfrag_hash_rnd) & (IPQ_HASHSZ - 1);
127 static struct timer_list ipfrag_secret_timer;
128 int sysctl_ipfrag_secret_interval = 10 * 60 * HZ;
130 static void ipfrag_secret_rebuild(unsigned long dummy)
132 unsigned long now = jiffies;
135 write_lock(&ipfrag_lock);
136 get_random_bytes(&ipfrag_hash_rnd, sizeof(u32));
137 for (i = 0; i < IPQ_HASHSZ; i++) {
139 struct hlist_node *p, *n;
141 hlist_for_each_entry_safe(q, p, n, &ipq_hash[i], list) {
142 unsigned int hval = ipqhashfn(q->id, q->saddr,
143 q->daddr, q->protocol);
148 /* Relink to new hash chain. */
149 hlist_add_head(&q->list, &ipq_hash[hval]);
153 write_unlock(&ipfrag_lock);
155 mod_timer(&ipfrag_secret_timer, now + sysctl_ipfrag_secret_interval);
158 atomic_t ip_frag_mem = ATOMIC_INIT(0); /* Memory used for fragments */
160 /* Memory Tracking Functions. */
161 static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
164 *work -= skb->truesize;
165 atomic_sub(skb->truesize, &ip_frag_mem);
169 static __inline__ void frag_free_queue(struct ipq *qp, int *work)
172 *work -= sizeof(struct ipq);
173 atomic_sub(sizeof(struct ipq), &ip_frag_mem);
177 static __inline__ struct ipq *frag_alloc_queue(void)
179 struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
183 atomic_add(sizeof(struct ipq), &ip_frag_mem);
188 /* Destruction primitives. */
190 /* Complete destruction of ipq. */
191 static void ip_frag_destroy(struct ipq *qp, int *work)
195 BUG_TRAP(qp->last_in&COMPLETE);
196 BUG_TRAP(del_timer(&qp->timer) == 0);
198 /* Release all fragment data. */
201 struct sk_buff *xp = fp->next;
203 frag_kfree_skb(fp, work);
207 /* Finally, release the queue descriptor itself. */
208 frag_free_queue(qp, work);
211 static __inline__ void ipq_put(struct ipq *ipq, int *work)
213 if (atomic_dec_and_test(&ipq->refcnt))
214 ip_frag_destroy(ipq, work);
217 /* Kill ipq entry. It is not destroyed immediately,
218 * because caller (and someone more) holds reference count.
220 static void ipq_kill(struct ipq *ipq)
222 if (del_timer(&ipq->timer))
223 atomic_dec(&ipq->refcnt);
225 if (!(ipq->last_in & COMPLETE)) {
227 atomic_dec(&ipq->refcnt);
228 ipq->last_in |= COMPLETE;
232 /* Memory limiting on fragments. Evictor trashes the oldest
233 * fragment queue until we are back under the threshold.
235 static void ip_evictor(void)
238 struct list_head *tmp;
241 work = atomic_read(&ip_frag_mem) - sysctl_ipfrag_low_thresh;
246 read_lock(&ipfrag_lock);
247 if (list_empty(&ipq_lru_list)) {
248 read_unlock(&ipfrag_lock);
251 tmp = ipq_lru_list.next;
252 qp = list_entry(tmp, struct ipq, lru_list);
253 atomic_inc(&qp->refcnt);
254 read_unlock(&ipfrag_lock);
256 spin_lock(&qp->lock);
257 if (!(qp->last_in&COMPLETE))
259 spin_unlock(&qp->lock);
262 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
267 * Oops, a fragment queue timed out. Kill it and send an ICMP reply.
269 static void ip_expire(unsigned long arg)
271 struct ipq *qp = (struct ipq *) arg;
273 spin_lock(&qp->lock);
275 if (qp->last_in & COMPLETE)
280 IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT);
281 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
283 if ((qp->last_in&FIRST_IN) && qp->fragments != NULL) {
284 struct sk_buff *head = qp->fragments;
285 /* Send an ICMP "Fragment Reassembly Timeout" message. */
286 if ((head->dev = dev_get_by_index(qp->iif)) != NULL) {
287 icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
292 spin_unlock(&qp->lock);
296 /* Creation primitives. */
298 static struct ipq *ip_frag_intern(unsigned int hash, struct ipq *qp_in)
302 struct hlist_node *n;
304 write_lock(&ipfrag_lock);
306 /* With SMP race we have to recheck hash table, because
307 * such entry could be created on other cpu, while we
308 * promoted read lock to write lock.
310 hlist_for_each_entry(qp, n, &ipq_hash[hash], list) {
311 if(qp->id == qp_in->id &&
312 qp->saddr == qp_in->saddr &&
313 qp->daddr == qp_in->daddr &&
314 qp->protocol == qp_in->protocol &&
315 qp->user == qp_in->user) {
316 atomic_inc(&qp->refcnt);
317 write_unlock(&ipfrag_lock);
318 qp_in->last_in |= COMPLETE;
319 ipq_put(qp_in, NULL);
326 if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time))
327 atomic_inc(&qp->refcnt);
329 atomic_inc(&qp->refcnt);
330 hlist_add_head(&qp->list, &ipq_hash[hash]);
331 INIT_LIST_HEAD(&qp->lru_list);
332 list_add_tail(&qp->lru_list, &ipq_lru_list);
334 write_unlock(&ipfrag_lock);
338 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
339 static struct ipq *ip_frag_create(unsigned hash, struct iphdr *iph, u32 user)
343 if ((qp = frag_alloc_queue()) == NULL)
346 qp->protocol = iph->protocol;
349 qp->saddr = iph->saddr;
350 qp->daddr = iph->daddr;
354 qp->fragments = NULL;
357 /* Initialize a timer for this entry. */
358 init_timer(&qp->timer);
359 qp->timer.data = (unsigned long) qp; /* pointer to queue */
360 qp->timer.function = ip_expire; /* expire function */
361 spin_lock_init(&qp->lock);
362 atomic_set(&qp->refcnt, 1);
364 return ip_frag_intern(hash, qp);
367 LIMIT_NETDEBUG(KERN_ERR "ip_frag_create: no memory left !\n");
371 /* Find the correct entry in the "incomplete datagrams" queue for
372 * this IP datagram, and create new one, if nothing is found.
374 static inline struct ipq *ip_find(struct iphdr *iph, u32 user)
377 __u32 saddr = iph->saddr;
378 __u32 daddr = iph->daddr;
379 __u8 protocol = iph->protocol;
380 unsigned int hash = ipqhashfn(id, saddr, daddr, protocol);
382 struct hlist_node *n;
384 read_lock(&ipfrag_lock);
385 hlist_for_each_entry(qp, n, &ipq_hash[hash], list) {
387 qp->saddr == saddr &&
388 qp->daddr == daddr &&
389 qp->protocol == protocol &&
391 atomic_inc(&qp->refcnt);
392 read_unlock(&ipfrag_lock);
396 read_unlock(&ipfrag_lock);
398 return ip_frag_create(hash, iph, user);
401 /* Add new segment to existing queue. */
402 static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
404 struct sk_buff *prev, *next;
408 if (qp->last_in & COMPLETE)
411 offset = ntohs(skb->nh.iph->frag_off);
412 flags = offset & ~IP_OFFSET;
414 offset <<= 3; /* offset is in 8-byte chunks */
415 ihl = skb->nh.iph->ihl * 4;
417 /* Determine the position of this fragment. */
418 end = offset + skb->len - ihl;
420 /* Is this the final fragment? */
421 if ((flags & IP_MF) == 0) {
422 /* If we already have some bits beyond end
423 * or have different end, the segment is corrrupted.
426 ((qp->last_in & LAST_IN) && end != qp->len))
428 qp->last_in |= LAST_IN;
433 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
434 skb->ip_summed = CHECKSUM_NONE;
437 /* Some bits beyond end -> corruption. */
438 if (qp->last_in & LAST_IN)
446 if (pskb_pull(skb, ihl) == NULL)
448 if (pskb_trim_rcsum(skb, end-offset))
451 /* Find out which fragments are in front and at the back of us
452 * in the chain of fragments so far. We must know where to put
453 * this fragment, right?
456 for(next = qp->fragments; next != NULL; next = next->next) {
457 if (FRAG_CB(next)->offset >= offset)
462 /* We found where to put this one. Check for overlap with
463 * preceding fragment, and, if needed, align things so that
464 * any overlaps are eliminated.
467 int i = (FRAG_CB(prev)->offset + prev->len) - offset;
473 if (!pskb_pull(skb, i))
475 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
476 skb->ip_summed = CHECKSUM_NONE;
480 while (next && FRAG_CB(next)->offset < end) {
481 int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
484 /* Eat head of the next overlapped fragment
485 * and leave the loop. The next ones cannot overlap.
487 if (!pskb_pull(next, i))
489 FRAG_CB(next)->offset += i;
491 if (next->ip_summed != CHECKSUM_UNNECESSARY)
492 next->ip_summed = CHECKSUM_NONE;
495 struct sk_buff *free_it = next;
497 /* Old fragmnet is completely overridden with
505 qp->fragments = next;
507 qp->meat -= free_it->len;
508 frag_kfree_skb(free_it, NULL);
512 FRAG_CB(skb)->offset = offset;
514 /* Insert this fragment in the chain of fragments. */
522 qp->iif = skb->dev->ifindex;
524 skb_get_timestamp(skb, &qp->stamp);
525 qp->meat += skb->len;
526 atomic_add(skb->truesize, &ip_frag_mem);
528 qp->last_in |= FIRST_IN;
530 write_lock(&ipfrag_lock);
531 list_move_tail(&qp->lru_list, &ipq_lru_list);
532 write_unlock(&ipfrag_lock);
541 /* Build a new IP datagram from all its fragments. */
543 static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
546 struct sk_buff *fp, *head = qp->fragments;
552 BUG_TRAP(head != NULL);
553 BUG_TRAP(FRAG_CB(head)->offset == 0);
555 /* Allocate a new buffer for the datagram. */
556 ihlen = head->nh.iph->ihl*4;
557 len = ihlen + qp->len;
562 /* Head of list must not be cloned. */
563 if (skb_cloned(head) && pskb_expand_head(head, 0, 0, GFP_ATOMIC))
566 /* If the first fragment is fragmented itself, we split
567 * it to two chunks: the first with data and paged part
568 * and the second, holding only fragments. */
569 if (skb_shinfo(head)->frag_list) {
570 struct sk_buff *clone;
573 if ((clone = alloc_skb(0, GFP_ATOMIC)) == NULL)
575 clone->next = head->next;
577 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
578 skb_shinfo(head)->frag_list = NULL;
579 for (i=0; i<skb_shinfo(head)->nr_frags; i++)
580 plen += skb_shinfo(head)->frags[i].size;
581 clone->len = clone->data_len = head->data_len - plen;
582 head->data_len -= clone->len;
583 head->len -= clone->len;
585 clone->ip_summed = head->ip_summed;
586 atomic_add(clone->truesize, &ip_frag_mem);
589 skb_shinfo(head)->frag_list = head->next;
590 skb_push(head, head->data - head->nh.raw);
591 atomic_sub(head->truesize, &ip_frag_mem);
593 for (fp=head->next; fp; fp = fp->next) {
594 head->data_len += fp->len;
595 head->len += fp->len;
596 if (head->ip_summed != fp->ip_summed)
597 head->ip_summed = CHECKSUM_NONE;
598 else if (head->ip_summed == CHECKSUM_HW)
599 head->csum = csum_add(head->csum, fp->csum);
600 head->truesize += fp->truesize;
601 atomic_sub(fp->truesize, &ip_frag_mem);
606 skb_set_timestamp(head, &qp->stamp);
610 iph->tot_len = htons(len);
611 IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS);
612 qp->fragments = NULL;
616 LIMIT_NETDEBUG(KERN_ERR "IP: queue_glue: no memory for gluing "
622 "Oversized IP packet from %d.%d.%d.%d.\n",
625 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
629 /* Process an incoming IP datagram fragment. */
630 struct sk_buff *ip_defrag(struct sk_buff *skb, u32 user)
632 struct iphdr *iph = skb->nh.iph;
634 struct net_device *dev;
636 IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS);
638 /* Start by cleaning up the memory. */
639 if (atomic_read(&ip_frag_mem) > sysctl_ipfrag_high_thresh)
644 /* Lookup (or create) queue header */
645 if ((qp = ip_find(iph, user)) != NULL) {
646 struct sk_buff *ret = NULL;
648 spin_lock(&qp->lock);
650 ip_frag_queue(qp, skb);
652 if (qp->last_in == (FIRST_IN|LAST_IN) &&
654 ret = ip_frag_reasm(qp, dev);
656 spin_unlock(&qp->lock);
661 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
666 void ipfrag_init(void)
668 ipfrag_hash_rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
669 (jiffies ^ (jiffies >> 6)));
671 init_timer(&ipfrag_secret_timer);
672 ipfrag_secret_timer.function = ipfrag_secret_rebuild;
673 ipfrag_secret_timer.expires = jiffies + sysctl_ipfrag_secret_interval;
674 add_timer(&ipfrag_secret_timer);
677 EXPORT_SYMBOL(ip_defrag);