ip_fragment.c revision e7c8a41e817f381ac5c2a59ecc81b483bd68a7df 
1/*
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.
5 *
6 *		The IP fragmentation functionality.
7 *
8 * Version:	$Id: ip_fragment.c,v 1.59 2002/01/12 07:54:56 davem Exp $
9 *
10 * Authors:	Fred N. van Kempen <waltje@uWalt.NL.Mugnet.ORG>
11 *		Alan Cox <Alan.Cox@linux.org>
12 *
13 * Fixes:
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.
23 */
24
25#include <linux/config.h>
26#include <linux/module.h>
27#include <linux/types.h>
28#include <linux/mm.h>
29#include <linux/jiffies.h>
30#include <linux/skbuff.h>
31#include <linux/list.h>
32#include <linux/ip.h>
33#include <linux/icmp.h>
34#include <linux/netdevice.h>
35#include <linux/jhash.h>
36#include <linux/random.h>
37#include <net/sock.h>
38#include <net/ip.h>
39#include <net/icmp.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>
45
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
49 */
50
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.
55 */
56int sysctl_ipfrag_high_thresh = 256*1024;
57int sysctl_ipfrag_low_thresh = 192*1024;
58
59/* Important NOTE! Fragment queue must be destroyed before MSL expires.
60 * RFC791 is wrong proposing to prolongate timer each fragment arrival by TTL.
61 */
62int sysctl_ipfrag_time = IP_FRAG_TIME;
63
64struct ipfrag_skb_cb
65{
66	struct inet_skb_parm	h;
67	int			offset;
68};
69
70#define FRAG_CB(skb)	((struct ipfrag_skb_cb*)((skb)->cb))
71
72/* Describe an entry in the "incomplete datagrams" queue. */
73struct ipq {
74	struct hlist_node list;
75	struct list_head lru_list;	/* lru list member 			*/
76	u32		user;
77	u32		saddr;
78	u32		daddr;
79	u16		id;
80	u8		protocol;
81	u8		last_in;
82#define COMPLETE		4
83#define FIRST_IN		2
84#define LAST_IN			1
85
86	struct sk_buff	*fragments;	/* linked list of received fragments	*/
87	int		len;		/* total length of original datagram	*/
88	int		meat;
89	spinlock_t	lock;
90	atomic_t	refcnt;
91	struct timer_list timer;	/* when will this queue expire?		*/
92	int		iif;
93	struct timeval	stamp;
94};
95
96/* Hash table. */
97
98#define IPQ_HASHSZ	64
99
100/* Per-bucket lock is easy to add now. */
101static struct hlist_head ipq_hash[IPQ_HASHSZ];
102static DEFINE_RWLOCK(ipfrag_lock);
103static u32 ipfrag_hash_rnd;
104static LIST_HEAD(ipq_lru_list);
105int ip_frag_nqueues = 0;
106
107static __inline__ void __ipq_unlink(struct ipq *qp)
108{
109	hlist_del(&qp->list);
110	list_del(&qp->lru_list);
111	ip_frag_nqueues--;
112}
113
114static __inline__ void ipq_unlink(struct ipq *ipq)
115{
116	write_lock(&ipfrag_lock);
117	__ipq_unlink(ipq);
118	write_unlock(&ipfrag_lock);
119}
120
121static unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
122{
123	return jhash_3words((u32)id << 16 | prot, saddr, daddr,
124			    ipfrag_hash_rnd) & (IPQ_HASHSZ - 1);
125}
126
127static struct timer_list ipfrag_secret_timer;
128int sysctl_ipfrag_secret_interval = 10 * 60 * HZ;
129
130static void ipfrag_secret_rebuild(unsigned long dummy)
131{
132	unsigned long now = jiffies;
133	int i;
134
135	write_lock(&ipfrag_lock);
136	get_random_bytes(&ipfrag_hash_rnd, sizeof(u32));
137	for (i = 0; i < IPQ_HASHSZ; i++) {
138		struct ipq *q;
139		struct hlist_node *p, *n;
140
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);
144
145			if (hval != i) {
146				hlist_del(&q->list);
147
148				/* Relink to new hash chain. */
149				hlist_add_head(&q->list, &ipq_hash[hval]);
150			}
151		}
152	}
153	write_unlock(&ipfrag_lock);
154
155	mod_timer(&ipfrag_secret_timer, now + sysctl_ipfrag_secret_interval);
156}
157
158atomic_t ip_frag_mem = ATOMIC_INIT(0);	/* Memory used for fragments */
159
160/* Memory Tracking Functions. */
161static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
162{
163	if (work)
164		*work -= skb->truesize;
165	atomic_sub(skb->truesize, &ip_frag_mem);
166	kfree_skb(skb);
167}
168
169static __inline__ void frag_free_queue(struct ipq *qp, int *work)
170{
171	if (work)
172		*work -= sizeof(struct ipq);
173	atomic_sub(sizeof(struct ipq), &ip_frag_mem);
174	kfree(qp);
175}
176
177static __inline__ struct ipq *frag_alloc_queue(void)
178{
179	struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
180
181	if(!qp)
182		return NULL;
183	atomic_add(sizeof(struct ipq), &ip_frag_mem);
184	return qp;
185}
186
187
188/* Destruction primitives. */
189
190/* Complete destruction of ipq. */
191static void ip_frag_destroy(struct ipq *qp, int *work)
192{
193	struct sk_buff *fp;
194
195	BUG_TRAP(qp->last_in&COMPLETE);
196	BUG_TRAP(del_timer(&qp->timer) == 0);
197
198	/* Release all fragment data. */
199	fp = qp->fragments;
200	while (fp) {
201		struct sk_buff *xp = fp->next;
202
203		frag_kfree_skb(fp, work);
204		fp = xp;
205	}
206
207	/* Finally, release the queue descriptor itself. */
208	frag_free_queue(qp, work);
209}
210
211static __inline__ void ipq_put(struct ipq *ipq, int *work)
212{
213	if (atomic_dec_and_test(&ipq->refcnt))
214		ip_frag_destroy(ipq, work);
215}
216
217/* Kill ipq entry. It is not destroyed immediately,
218 * because caller (and someone more) holds reference count.
219 */
220static void ipq_kill(struct ipq *ipq)
221{
222	if (del_timer(&ipq->timer))
223		atomic_dec(&ipq->refcnt);
224
225	if (!(ipq->last_in & COMPLETE)) {
226		ipq_unlink(ipq);
227		atomic_dec(&ipq->refcnt);
228		ipq->last_in |= COMPLETE;
229	}
230}
231
232/* Memory limiting on fragments.  Evictor trashes the oldest
233 * fragment queue until we are back under the threshold.
234 */
235static void ip_evictor(void)
236{
237	struct ipq *qp;
238	struct list_head *tmp;
239	int work;
240
241	work = atomic_read(&ip_frag_mem) - sysctl_ipfrag_low_thresh;
242	if (work <= 0)
243		return;
244
245	while (work > 0) {
246		read_lock(&ipfrag_lock);
247		if (list_empty(&ipq_lru_list)) {
248			read_unlock(&ipfrag_lock);
249			return;
250		}
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);
255
256		spin_lock(&qp->lock);
257		if (!(qp->last_in&COMPLETE))
258			ipq_kill(qp);
259		spin_unlock(&qp->lock);
260
261		ipq_put(qp, &work);
262		IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
263	}
264}
265
266/*
267 * Oops, a fragment queue timed out.  Kill it and send an ICMP reply.
268 */
269static void ip_expire(unsigned long arg)
270{
271	struct ipq *qp = (struct ipq *) arg;
272
273	spin_lock(&qp->lock);
274
275	if (qp->last_in & COMPLETE)
276		goto out;
277
278	ipq_kill(qp);
279
280	IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT);
281	IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
282
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);
288			dev_put(head->dev);
289		}
290	}
291out:
292	spin_unlock(&qp->lock);
293	ipq_put(qp, NULL);
294}
295
296/* Creation primitives. */
297
298static struct ipq *ip_frag_intern(unsigned int hash, struct ipq *qp_in)
299{
300	struct ipq *qp;
301#ifdef CONFIG_SMP
302	struct hlist_node *n;
303#endif
304	write_lock(&ipfrag_lock);
305#ifdef CONFIG_SMP
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.
309	 */
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);
320			return qp;
321		}
322	}
323#endif
324	qp = qp_in;
325
326	if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time))
327		atomic_inc(&qp->refcnt);
328
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);
333	ip_frag_nqueues++;
334	write_unlock(&ipfrag_lock);
335	return qp;
336}
337
338/* Add an entry to the 'ipq' queue for a newly received IP datagram. */
339static struct ipq *ip_frag_create(unsigned hash, struct iphdr *iph, u32 user)
340{
341	struct ipq *qp;
342
343	if ((qp = frag_alloc_queue()) == NULL)
344		goto out_nomem;
345
346	qp->protocol = iph->protocol;
347	qp->last_in = 0;
348	qp->id = iph->id;
349	qp->saddr = iph->saddr;
350	qp->daddr = iph->daddr;
351	qp->user = user;
352	qp->len = 0;
353	qp->meat = 0;
354	qp->fragments = NULL;
355	qp->iif = 0;
356
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);
363
364	return ip_frag_intern(hash, qp);
365
366out_nomem:
367	LIMIT_NETDEBUG(KERN_ERR "ip_frag_create: no memory left !\n");
368	return NULL;
369}
370
371/* Find the correct entry in the "incomplete datagrams" queue for
372 * this IP datagram, and create new one, if nothing is found.
373 */
374static inline struct ipq *ip_find(struct iphdr *iph, u32 user)
375{
376	__u16 id = iph->id;
377	__u32 saddr = iph->saddr;
378	__u32 daddr = iph->daddr;
379	__u8 protocol = iph->protocol;
380	unsigned int hash = ipqhashfn(id, saddr, daddr, protocol);
381	struct ipq *qp;
382	struct hlist_node *n;
383
384	read_lock(&ipfrag_lock);
385	hlist_for_each_entry(qp, n, &ipq_hash[hash], list) {
386		if(qp->id == id		&&
387		   qp->saddr == saddr	&&
388		   qp->daddr == daddr	&&
389		   qp->protocol == protocol &&
390		   qp->user == user) {
391			atomic_inc(&qp->refcnt);
392			read_unlock(&ipfrag_lock);
393			return qp;
394		}
395	}
396	read_unlock(&ipfrag_lock);
397
398	return ip_frag_create(hash, iph, user);
399}
400
401/* Add new segment to existing queue. */
402static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
403{
404	struct sk_buff *prev, *next;
405	int flags, offset;
406	int ihl, end;
407
408	if (qp->last_in & COMPLETE)
409		goto err;
410
411 	offset = ntohs(skb->nh.iph->frag_off);
412	flags = offset & ~IP_OFFSET;
413	offset &= IP_OFFSET;
414	offset <<= 3;		/* offset is in 8-byte chunks */
415 	ihl = skb->nh.iph->ihl * 4;
416
417	/* Determine the position of this fragment. */
418 	end = offset + skb->len - ihl;
419
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.
424		 */
425		if (end < qp->len ||
426		    ((qp->last_in & LAST_IN) && end != qp->len))
427			goto err;
428		qp->last_in |= LAST_IN;
429		qp->len = end;
430	} else {
431		if (end&7) {
432			end &= ~7;
433			if (skb->ip_summed != CHECKSUM_UNNECESSARY)
434				skb->ip_summed = CHECKSUM_NONE;
435		}
436		if (end > qp->len) {
437			/* Some bits beyond end -> corruption. */
438			if (qp->last_in & LAST_IN)
439				goto err;
440			qp->len = end;
441		}
442	}
443	if (end == offset)
444		goto err;
445
446	if (pskb_pull(skb, ihl) == NULL)
447		goto err;
448	if (pskb_trim_rcsum(skb, end-offset))
449		goto err;
450
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?
454	 */
455	prev = NULL;
456	for(next = qp->fragments; next != NULL; next = next->next) {
457		if (FRAG_CB(next)->offset >= offset)
458			break;	/* bingo! */
459		prev = next;
460	}
461
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.
465	 */
466	if (prev) {
467		int i = (FRAG_CB(prev)->offset + prev->len) - offset;
468
469		if (i > 0) {
470			offset += i;
471			if (end <= offset)
472				goto err;
473			if (!pskb_pull(skb, i))
474				goto err;
475			if (skb->ip_summed != CHECKSUM_UNNECESSARY)
476				skb->ip_summed = CHECKSUM_NONE;
477		}
478	}
479
480	while (next && FRAG_CB(next)->offset < end) {
481		int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
482
483		if (i < next->len) {
484			/* Eat head of the next overlapped fragment
485			 * and leave the loop. The next ones cannot overlap.
486			 */
487			if (!pskb_pull(next, i))
488				goto err;
489			FRAG_CB(next)->offset += i;
490			qp->meat -= i;
491			if (next->ip_summed != CHECKSUM_UNNECESSARY)
492				next->ip_summed = CHECKSUM_NONE;
493			break;
494		} else {
495			struct sk_buff *free_it = next;
496
497			/* Old fragmnet is completely overridden with
498			 * new one drop it.
499			 */
500			next = next->next;
501
502			if (prev)
503				prev->next = next;
504			else
505				qp->fragments = next;
506
507			qp->meat -= free_it->len;
508			frag_kfree_skb(free_it, NULL);
509		}
510	}
511
512	FRAG_CB(skb)->offset = offset;
513
514	/* Insert this fragment in the chain of fragments. */
515	skb->next = next;
516	if (prev)
517		prev->next = skb;
518	else
519		qp->fragments = skb;
520
521 	if (skb->dev)
522 		qp->iif = skb->dev->ifindex;
523	skb->dev = NULL;
524	skb_get_timestamp(skb, &qp->stamp);
525	qp->meat += skb->len;
526	atomic_add(skb->truesize, &ip_frag_mem);
527	if (offset == 0)
528		qp->last_in |= FIRST_IN;
529
530	write_lock(&ipfrag_lock);
531	list_move_tail(&qp->lru_list, &ipq_lru_list);
532	write_unlock(&ipfrag_lock);
533
534	return;
535
536err:
537	kfree_skb(skb);
538}
539
540
541/* Build a new IP datagram from all its fragments. */
542
543static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
544{
545	struct iphdr *iph;
546	struct sk_buff *fp, *head = qp->fragments;
547	int len;
548	int ihlen;
549
550	ipq_kill(qp);
551
552	BUG_TRAP(head != NULL);
553	BUG_TRAP(FRAG_CB(head)->offset == 0);
554
555	/* Allocate a new buffer for the datagram. */
556	ihlen = head->nh.iph->ihl*4;
557	len = ihlen + qp->len;
558
559	if(len > 65535)
560		goto out_oversize;
561
562	/* Head of list must not be cloned. */
563	if (skb_cloned(head) && pskb_expand_head(head, 0, 0, GFP_ATOMIC))
564		goto out_nomem;
565
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;
571		int i, plen = 0;
572
573		if ((clone = alloc_skb(0, GFP_ATOMIC)) == NULL)
574			goto out_nomem;
575		clone->next = head->next;
576		head->next = clone;
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;
584		clone->csum = 0;
585		clone->ip_summed = head->ip_summed;
586		atomic_add(clone->truesize, &ip_frag_mem);
587	}
588
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);
592
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);
602	}
603
604	head->next = NULL;
605	head->dev = dev;
606	skb_set_timestamp(head, &qp->stamp);
607
608	iph = head->nh.iph;
609	iph->frag_off = 0;
610	iph->tot_len = htons(len);
611	IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS);
612	qp->fragments = NULL;
613	return head;
614
615out_nomem:
616 	LIMIT_NETDEBUG(KERN_ERR "IP: queue_glue: no memory for gluing "
617			      "queue %p\n", qp);
618	goto out_fail;
619out_oversize:
620	if (net_ratelimit())
621		printk(KERN_INFO
622			"Oversized IP packet from %d.%d.%d.%d.\n",
623			NIPQUAD(qp->saddr));
624out_fail:
625	IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
626	return NULL;
627}
628
629/* Process an incoming IP datagram fragment. */
630struct sk_buff *ip_defrag(struct sk_buff *skb, u32 user)
631{
632	struct iphdr *iph = skb->nh.iph;
633	struct ipq *qp;
634	struct net_device *dev;
635
636	IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS);
637
638	/* Start by cleaning up the memory. */
639	if (atomic_read(&ip_frag_mem) > sysctl_ipfrag_high_thresh)
640		ip_evictor();
641
642	dev = skb->dev;
643
644	/* Lookup (or create) queue header */
645	if ((qp = ip_find(iph, user)) != NULL) {
646		struct sk_buff *ret = NULL;
647
648		spin_lock(&qp->lock);
649
650		ip_frag_queue(qp, skb);
651
652		if (qp->last_in == (FIRST_IN|LAST_IN) &&
653		    qp->meat == qp->len)
654			ret = ip_frag_reasm(qp, dev);
655
656		spin_unlock(&qp->lock);
657		ipq_put(qp, NULL);
658		return ret;
659	}
660
661	IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
662	kfree_skb(skb);
663	return NULL;
664}
665
666void ipfrag_init(void)
667{
668	ipfrag_hash_rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
669				 (jiffies ^ (jiffies >> 6)));
670
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);
675}
676
677EXPORT_SYMBOL(ip_defrag);
678