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 ipq *next; /* linked list pointers */
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? */
101 /* Per-bucket lock is easy to add now. */
102 static struct ipq *ipq_hash[IPQ_HASHSZ];
103 static DEFINE_RWLOCK(ipfrag_lock);
104 static u32 ipfrag_hash_rnd;
105 static LIST_HEAD(ipq_lru_list);
106 int ip_frag_nqueues = 0;
108 static __inline__ void __ipq_unlink(struct ipq *qp)
111 qp->next->pprev = qp->pprev;
112 *qp->pprev = qp->next;
113 list_del(&qp->lru_list);
117 static __inline__ void ipq_unlink(struct ipq *ipq)
119 write_lock(&ipfrag_lock);
121 write_unlock(&ipfrag_lock);
124 static unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
126 return jhash_3words((u32)id << 16 | prot, saddr, daddr,
127 ipfrag_hash_rnd) & (IPQ_HASHSZ - 1);
130 static struct timer_list ipfrag_secret_timer;
131 int sysctl_ipfrag_secret_interval = 10 * 60 * HZ;
133 static void ipfrag_secret_rebuild(unsigned long dummy)
135 unsigned long now = jiffies;
138 write_lock(&ipfrag_lock);
139 get_random_bytes(&ipfrag_hash_rnd, sizeof(u32));
140 for (i = 0; i < IPQ_HASHSZ; i++) {
145 struct ipq *next = q->next;
146 unsigned int hval = ipqhashfn(q->id, q->saddr,
147 q->daddr, q->protocol);
152 q->next->pprev = q->pprev;
155 /* Relink to new hash chain. */
156 if ((q->next = ipq_hash[hval]) != NULL)
157 q->next->pprev = &q->next;
159 q->pprev = &ipq_hash[hval];
165 write_unlock(&ipfrag_lock);
167 mod_timer(&ipfrag_secret_timer, now + sysctl_ipfrag_secret_interval);
170 atomic_t ip_frag_mem = ATOMIC_INIT(0); /* Memory used for fragments */
172 /* Memory Tracking Functions. */
173 static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
176 *work -= skb->truesize;
177 atomic_sub(skb->truesize, &ip_frag_mem);
181 static __inline__ void frag_free_queue(struct ipq *qp, int *work)
184 *work -= sizeof(struct ipq);
185 atomic_sub(sizeof(struct ipq), &ip_frag_mem);
189 static __inline__ struct ipq *frag_alloc_queue(void)
191 struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
195 atomic_add(sizeof(struct ipq), &ip_frag_mem);
200 /* Destruction primitives. */
202 /* Complete destruction of ipq. */
203 static void ip_frag_destroy(struct ipq *qp, int *work)
207 BUG_TRAP(qp->last_in&COMPLETE);
208 BUG_TRAP(del_timer(&qp->timer) == 0);
210 /* Release all fragment data. */
213 struct sk_buff *xp = fp->next;
215 frag_kfree_skb(fp, work);
219 /* Finally, release the queue descriptor itself. */
220 frag_free_queue(qp, work);
223 static __inline__ void ipq_put(struct ipq *ipq, int *work)
225 if (atomic_dec_and_test(&ipq->refcnt))
226 ip_frag_destroy(ipq, work);
229 /* Kill ipq entry. It is not destroyed immediately,
230 * because caller (and someone more) holds reference count.
232 static void ipq_kill(struct ipq *ipq)
234 if (del_timer(&ipq->timer))
235 atomic_dec(&ipq->refcnt);
237 if (!(ipq->last_in & COMPLETE)) {
239 atomic_dec(&ipq->refcnt);
240 ipq->last_in |= COMPLETE;
244 /* Memory limiting on fragments. Evictor trashes the oldest
245 * fragment queue until we are back under the threshold.
247 static void ip_evictor(void)
250 struct list_head *tmp;
253 work = atomic_read(&ip_frag_mem) - sysctl_ipfrag_low_thresh;
258 read_lock(&ipfrag_lock);
259 if (list_empty(&ipq_lru_list)) {
260 read_unlock(&ipfrag_lock);
263 tmp = ipq_lru_list.next;
264 qp = list_entry(tmp, struct ipq, lru_list);
265 atomic_inc(&qp->refcnt);
266 read_unlock(&ipfrag_lock);
268 spin_lock(&qp->lock);
269 if (!(qp->last_in&COMPLETE))
271 spin_unlock(&qp->lock);
274 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
279 * Oops, a fragment queue timed out. Kill it and send an ICMP reply.
281 static void ip_expire(unsigned long arg)
283 struct ipq *qp = (struct ipq *) arg;
285 spin_lock(&qp->lock);
287 if (qp->last_in & COMPLETE)
292 IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT);
293 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
295 if ((qp->last_in&FIRST_IN) && qp->fragments != NULL) {
296 struct sk_buff *head = qp->fragments;
297 /* Send an ICMP "Fragment Reassembly Timeout" message. */
298 if ((head->dev = dev_get_by_index(qp->iif)) != NULL) {
299 icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
304 spin_unlock(&qp->lock);
308 /* Creation primitives. */
310 static struct ipq *ip_frag_intern(unsigned int hash, struct ipq *qp_in)
314 write_lock(&ipfrag_lock);
316 /* With SMP race we have to recheck hash table, because
317 * such entry could be created on other cpu, while we
318 * promoted read lock to write lock.
320 for(qp = ipq_hash[hash]; qp; qp = qp->next) {
321 if(qp->id == qp_in->id &&
322 qp->saddr == qp_in->saddr &&
323 qp->daddr == qp_in->daddr &&
324 qp->protocol == qp_in->protocol &&
325 qp->user == qp_in->user) {
326 atomic_inc(&qp->refcnt);
327 write_unlock(&ipfrag_lock);
328 qp_in->last_in |= COMPLETE;
329 ipq_put(qp_in, NULL);
336 if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time))
337 atomic_inc(&qp->refcnt);
339 atomic_inc(&qp->refcnt);
340 if((qp->next = ipq_hash[hash]) != NULL)
341 qp->next->pprev = &qp->next;
343 qp->pprev = &ipq_hash[hash];
344 INIT_LIST_HEAD(&qp->lru_list);
345 list_add_tail(&qp->lru_list, &ipq_lru_list);
347 write_unlock(&ipfrag_lock);
351 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
352 static struct ipq *ip_frag_create(unsigned hash, struct iphdr *iph, u32 user)
356 if ((qp = frag_alloc_queue()) == NULL)
359 qp->protocol = iph->protocol;
362 qp->saddr = iph->saddr;
363 qp->daddr = iph->daddr;
367 qp->fragments = NULL;
370 /* Initialize a timer for this entry. */
371 init_timer(&qp->timer);
372 qp->timer.data = (unsigned long) qp; /* pointer to queue */
373 qp->timer.function = ip_expire; /* expire function */
374 spin_lock_init(&qp->lock);
375 atomic_set(&qp->refcnt, 1);
377 return ip_frag_intern(hash, qp);
380 LIMIT_NETDEBUG(printk(KERN_ERR "ip_frag_create: no memory left !\n"));
384 /* Find the correct entry in the "incomplete datagrams" queue for
385 * this IP datagram, and create new one, if nothing is found.
387 static inline struct ipq *ip_find(struct iphdr *iph, u32 user)
390 __u32 saddr = iph->saddr;
391 __u32 daddr = iph->daddr;
392 __u8 protocol = iph->protocol;
393 unsigned int hash = ipqhashfn(id, saddr, daddr, protocol);
396 read_lock(&ipfrag_lock);
397 for(qp = ipq_hash[hash]; qp; qp = qp->next) {
399 qp->saddr == saddr &&
400 qp->daddr == daddr &&
401 qp->protocol == protocol &&
403 atomic_inc(&qp->refcnt);
404 read_unlock(&ipfrag_lock);
408 read_unlock(&ipfrag_lock);
410 return ip_frag_create(hash, iph, user);
413 /* Add new segment to existing queue. */
414 static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
416 struct sk_buff *prev, *next;
420 if (qp->last_in & COMPLETE)
423 offset = ntohs(skb->nh.iph->frag_off);
424 flags = offset & ~IP_OFFSET;
426 offset <<= 3; /* offset is in 8-byte chunks */
427 ihl = skb->nh.iph->ihl * 4;
429 /* Determine the position of this fragment. */
430 end = offset + skb->len - ihl;
432 /* Is this the final fragment? */
433 if ((flags & IP_MF) == 0) {
434 /* If we already have some bits beyond end
435 * or have different end, the segment is corrrupted.
438 ((qp->last_in & LAST_IN) && end != qp->len))
440 qp->last_in |= LAST_IN;
445 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
446 skb->ip_summed = CHECKSUM_NONE;
449 /* Some bits beyond end -> corruption. */
450 if (qp->last_in & LAST_IN)
458 if (pskb_pull(skb, ihl) == NULL)
460 if (pskb_trim(skb, end-offset))
463 /* Find out which fragments are in front and at the back of us
464 * in the chain of fragments so far. We must know where to put
465 * this fragment, right?
468 for(next = qp->fragments; next != NULL; next = next->next) {
469 if (FRAG_CB(next)->offset >= offset)
474 /* We found where to put this one. Check for overlap with
475 * preceding fragment, and, if needed, align things so that
476 * any overlaps are eliminated.
479 int i = (FRAG_CB(prev)->offset + prev->len) - offset;
485 if (!pskb_pull(skb, i))
487 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
488 skb->ip_summed = CHECKSUM_NONE;
492 while (next && FRAG_CB(next)->offset < end) {
493 int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
496 /* Eat head of the next overlapped fragment
497 * and leave the loop. The next ones cannot overlap.
499 if (!pskb_pull(next, i))
501 FRAG_CB(next)->offset += i;
503 if (next->ip_summed != CHECKSUM_UNNECESSARY)
504 next->ip_summed = CHECKSUM_NONE;
507 struct sk_buff *free_it = next;
509 /* Old fragmnet is completely overridden with
517 qp->fragments = next;
519 qp->meat -= free_it->len;
520 frag_kfree_skb(free_it, NULL);
524 FRAG_CB(skb)->offset = offset;
526 /* Insert this fragment in the chain of fragments. */
534 qp->iif = skb->dev->ifindex;
536 qp->stamp = skb->stamp;
537 qp->meat += skb->len;
538 atomic_add(skb->truesize, &ip_frag_mem);
540 qp->last_in |= FIRST_IN;
542 write_lock(&ipfrag_lock);
543 list_move_tail(&qp->lru_list, &ipq_lru_list);
544 write_unlock(&ipfrag_lock);
553 /* Build a new IP datagram from all its fragments. */
555 static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
558 struct sk_buff *fp, *head = qp->fragments;
564 BUG_TRAP(head != NULL);
565 BUG_TRAP(FRAG_CB(head)->offset == 0);
567 /* Allocate a new buffer for the datagram. */
568 ihlen = head->nh.iph->ihl*4;
569 len = ihlen + qp->len;
574 /* Head of list must not be cloned. */
575 if (skb_cloned(head) && pskb_expand_head(head, 0, 0, GFP_ATOMIC))
578 /* If the first fragment is fragmented itself, we split
579 * it to two chunks: the first with data and paged part
580 * and the second, holding only fragments. */
581 if (skb_shinfo(head)->frag_list) {
582 struct sk_buff *clone;
585 if ((clone = alloc_skb(0, GFP_ATOMIC)) == NULL)
587 clone->next = head->next;
589 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
590 skb_shinfo(head)->frag_list = NULL;
591 for (i=0; i<skb_shinfo(head)->nr_frags; i++)
592 plen += skb_shinfo(head)->frags[i].size;
593 clone->len = clone->data_len = head->data_len - plen;
594 head->data_len -= clone->len;
595 head->len -= clone->len;
597 clone->ip_summed = head->ip_summed;
598 atomic_add(clone->truesize, &ip_frag_mem);
601 skb_shinfo(head)->frag_list = head->next;
602 skb_push(head, head->data - head->nh.raw);
603 atomic_sub(head->truesize, &ip_frag_mem);
605 for (fp=head->next; fp; fp = fp->next) {
606 head->data_len += fp->len;
607 head->len += fp->len;
608 if (head->ip_summed != fp->ip_summed)
609 head->ip_summed = CHECKSUM_NONE;
610 else if (head->ip_summed == CHECKSUM_HW)
611 head->csum = csum_add(head->csum, fp->csum);
612 head->truesize += fp->truesize;
613 atomic_sub(fp->truesize, &ip_frag_mem);
618 head->stamp = qp->stamp;
622 iph->tot_len = htons(len);
623 IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS);
624 qp->fragments = NULL;
628 LIMIT_NETDEBUG(printk(KERN_ERR "IP: queue_glue: no memory for gluing "
634 "Oversized IP packet from %d.%d.%d.%d.\n",
637 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
641 /* Process an incoming IP datagram fragment. */
642 struct sk_buff *ip_defrag(struct sk_buff *skb, u32 user)
644 struct iphdr *iph = skb->nh.iph;
646 struct net_device *dev;
648 IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS);
650 /* Start by cleaning up the memory. */
651 if (atomic_read(&ip_frag_mem) > sysctl_ipfrag_high_thresh)
656 /* Lookup (or create) queue header */
657 if ((qp = ip_find(iph, user)) != NULL) {
658 struct sk_buff *ret = NULL;
660 spin_lock(&qp->lock);
662 ip_frag_queue(qp, skb);
664 if (qp->last_in == (FIRST_IN|LAST_IN) &&
666 ret = ip_frag_reasm(qp, dev);
668 spin_unlock(&qp->lock);
673 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
678 void ipfrag_init(void)
680 ipfrag_hash_rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
681 (jiffies ^ (jiffies >> 6)));
683 init_timer(&ipfrag_secret_timer);
684 ipfrag_secret_timer.function = ipfrag_secret_rebuild;
685 ipfrag_secret_timer.expires = jiffies + sysctl_ipfrag_secret_interval;
686 add_timer(&ipfrag_secret_timer);
689 EXPORT_SYMBOL(ip_defrag);