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 <linux/tcp.h>
43 #include <linux/udp.h>
44 #include <linux/inet.h>
45 #include <linux/netfilter_ipv4.h>
47 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
48 * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
49 * as well. Or notify me, at least. --ANK
52 /* Fragment cache limits. We will commit 256K at one time. Should we
53 * cross that limit we will prune down to 192K. This should cope with
54 * even the most extreme cases without allowing an attacker to measurably
55 * harm machine performance.
57 int sysctl_ipfrag_high_thresh = 256*1024;
58 int sysctl_ipfrag_low_thresh = 192*1024;
60 int sysctl_ipfrag_max_dist = 64;
62 /* Important NOTE! Fragment queue must be destroyed before MSL expires.
63 * RFC791 is wrong proposing to prolongate timer each fragment arrival by TTL.
65 int sysctl_ipfrag_time = IP_FRAG_TIME;
69 struct inet_skb_parm h;
73 #define FRAG_CB(skb) ((struct ipfrag_skb_cb*)((skb)->cb))
75 /* Describe an entry in the "incomplete datagrams" queue. */
77 struct hlist_node list;
78 struct list_head lru_list; /* lru list member */
89 struct sk_buff *fragments; /* linked list of received fragments */
90 int len; /* total length of original datagram */
94 struct timer_list timer; /* when will this queue expire? */
98 struct inet_peer *peer;
103 #define IPQ_HASHSZ 64
105 /* Per-bucket lock is easy to add now. */
106 static struct hlist_head ipq_hash[IPQ_HASHSZ];
107 static DEFINE_RWLOCK(ipfrag_lock);
108 static u32 ipfrag_hash_rnd;
109 static LIST_HEAD(ipq_lru_list);
110 int ip_frag_nqueues = 0;
112 static __inline__ void __ipq_unlink(struct ipq *qp)
114 hlist_del(&qp->list);
115 list_del(&qp->lru_list);
119 static __inline__ void ipq_unlink(struct ipq *ipq)
121 write_lock(&ipfrag_lock);
123 write_unlock(&ipfrag_lock);
126 static unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
128 return jhash_3words((u32)id << 16 | prot, saddr, daddr,
129 ipfrag_hash_rnd) & (IPQ_HASHSZ - 1);
132 static struct timer_list ipfrag_secret_timer;
133 int sysctl_ipfrag_secret_interval = 10 * 60 * HZ;
135 static void ipfrag_secret_rebuild(unsigned long dummy)
137 unsigned long now = jiffies;
140 write_lock(&ipfrag_lock);
141 get_random_bytes(&ipfrag_hash_rnd, sizeof(u32));
142 for (i = 0; i < IPQ_HASHSZ; i++) {
144 struct hlist_node *p, *n;
146 hlist_for_each_entry_safe(q, p, n, &ipq_hash[i], list) {
147 unsigned int hval = ipqhashfn(q->id, q->saddr,
148 q->daddr, q->protocol);
153 /* Relink to new hash chain. */
154 hlist_add_head(&q->list, &ipq_hash[hval]);
158 write_unlock(&ipfrag_lock);
160 mod_timer(&ipfrag_secret_timer, now + sysctl_ipfrag_secret_interval);
163 atomic_t ip_frag_mem = ATOMIC_INIT(0); /* Memory used for fragments */
165 /* Memory Tracking Functions. */
166 static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
169 *work -= skb->truesize;
170 atomic_sub(skb->truesize, &ip_frag_mem);
174 static __inline__ void frag_free_queue(struct ipq *qp, int *work)
177 *work -= sizeof(struct ipq);
178 atomic_sub(sizeof(struct ipq), &ip_frag_mem);
182 static __inline__ struct ipq *frag_alloc_queue(void)
184 struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
188 atomic_add(sizeof(struct ipq), &ip_frag_mem);
193 /* Destruction primitives. */
195 /* Complete destruction of ipq. */
196 static void ip_frag_destroy(struct ipq *qp, int *work)
200 BUG_TRAP(qp->last_in&COMPLETE);
201 BUG_TRAP(del_timer(&qp->timer) == 0);
204 inet_putpeer(qp->peer);
206 /* Release all fragment data. */
209 struct sk_buff *xp = fp->next;
211 frag_kfree_skb(fp, work);
215 /* Finally, release the queue descriptor itself. */
216 frag_free_queue(qp, work);
219 static __inline__ void ipq_put(struct ipq *ipq, int *work)
221 if (atomic_dec_and_test(&ipq->refcnt))
222 ip_frag_destroy(ipq, work);
225 /* Kill ipq entry. It is not destroyed immediately,
226 * because caller (and someone more) holds reference count.
228 static void ipq_kill(struct ipq *ipq)
230 if (del_timer(&ipq->timer))
231 atomic_dec(&ipq->refcnt);
233 if (!(ipq->last_in & COMPLETE)) {
235 atomic_dec(&ipq->refcnt);
236 ipq->last_in |= COMPLETE;
240 /* Memory limiting on fragments. Evictor trashes the oldest
241 * fragment queue until we are back under the threshold.
243 static void ip_evictor(void)
246 struct list_head *tmp;
249 work = atomic_read(&ip_frag_mem) - sysctl_ipfrag_low_thresh;
254 read_lock(&ipfrag_lock);
255 if (list_empty(&ipq_lru_list)) {
256 read_unlock(&ipfrag_lock);
259 tmp = ipq_lru_list.next;
260 qp = list_entry(tmp, struct ipq, lru_list);
261 atomic_inc(&qp->refcnt);
262 read_unlock(&ipfrag_lock);
264 spin_lock(&qp->lock);
265 if (!(qp->last_in&COMPLETE))
267 spin_unlock(&qp->lock);
270 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
275 * Oops, a fragment queue timed out. Kill it and send an ICMP reply.
277 static void ip_expire(unsigned long arg)
279 struct ipq *qp = (struct ipq *) arg;
281 spin_lock(&qp->lock);
283 if (qp->last_in & COMPLETE)
288 IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT);
289 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
291 if ((qp->last_in&FIRST_IN) && qp->fragments != NULL) {
292 struct sk_buff *head = qp->fragments;
293 /* Send an ICMP "Fragment Reassembly Timeout" message. */
294 if ((head->dev = dev_get_by_index(qp->iif)) != NULL) {
295 icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
300 spin_unlock(&qp->lock);
304 /* Creation primitives. */
306 static struct ipq *ip_frag_intern(struct ipq *qp_in)
310 struct hlist_node *n;
314 write_lock(&ipfrag_lock);
315 hash = ipqhashfn(qp_in->id, qp_in->saddr, qp_in->daddr,
318 /* With SMP race we have to recheck hash table, because
319 * such entry could be created on other cpu, while we
320 * promoted read lock to write lock.
322 hlist_for_each_entry(qp, n, &ipq_hash[hash], list) {
323 if(qp->id == qp_in->id &&
324 qp->saddr == qp_in->saddr &&
325 qp->daddr == qp_in->daddr &&
326 qp->protocol == qp_in->protocol &&
327 qp->user == qp_in->user) {
328 atomic_inc(&qp->refcnt);
329 write_unlock(&ipfrag_lock);
330 qp_in->last_in |= COMPLETE;
331 ipq_put(qp_in, NULL);
338 if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time))
339 atomic_inc(&qp->refcnt);
341 atomic_inc(&qp->refcnt);
342 hlist_add_head(&qp->list, &ipq_hash[hash]);
343 INIT_LIST_HEAD(&qp->lru_list);
344 list_add_tail(&qp->lru_list, &ipq_lru_list);
346 write_unlock(&ipfrag_lock);
350 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
351 static struct ipq *ip_frag_create(struct iphdr *iph, u32 user)
355 if ((qp = frag_alloc_queue()) == NULL)
358 qp->protocol = iph->protocol;
361 qp->saddr = iph->saddr;
362 qp->daddr = iph->daddr;
366 qp->fragments = NULL;
368 qp->peer = sysctl_ipfrag_max_dist ? inet_getpeer(iph->saddr, 1) : 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(qp);
380 LIMIT_NETDEBUG(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;
395 struct hlist_node *n;
397 read_lock(&ipfrag_lock);
398 hash = ipqhashfn(id, saddr, daddr, protocol);
399 hlist_for_each_entry(qp, n, &ipq_hash[hash], list) {
401 qp->saddr == saddr &&
402 qp->daddr == daddr &&
403 qp->protocol == protocol &&
405 atomic_inc(&qp->refcnt);
406 read_unlock(&ipfrag_lock);
410 read_unlock(&ipfrag_lock);
412 return ip_frag_create(iph, user);
415 /* Is the fragment too far ahead to be part of ipq? */
416 static inline int ip_frag_too_far(struct ipq *qp)
418 struct inet_peer *peer = qp->peer;
419 unsigned int max = sysctl_ipfrag_max_dist;
420 unsigned int start, end;
428 end = atomic_inc_return(&peer->rid);
431 rc = qp->fragments && (end - start) > max;
434 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
440 static int ip_frag_reinit(struct ipq *qp)
444 if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time)) {
445 atomic_inc(&qp->refcnt);
451 struct sk_buff *xp = fp->next;
452 frag_kfree_skb(fp, NULL);
459 qp->fragments = NULL;
465 /* Add new segment to existing queue. */
466 static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
468 struct sk_buff *prev, *next;
472 if (qp->last_in & COMPLETE)
475 if (!(IPCB(skb)->flags & IPSKB_FRAG_COMPLETE) &&
476 unlikely(ip_frag_too_far(qp)) && unlikely(ip_frag_reinit(qp))) {
481 offset = ntohs(skb->nh.iph->frag_off);
482 flags = offset & ~IP_OFFSET;
484 offset <<= 3; /* offset is in 8-byte chunks */
485 ihl = skb->nh.iph->ihl * 4;
487 /* Determine the position of this fragment. */
488 end = offset + skb->len - ihl;
490 /* Is this the final fragment? */
491 if ((flags & IP_MF) == 0) {
492 /* If we already have some bits beyond end
493 * or have different end, the segment is corrrupted.
496 ((qp->last_in & LAST_IN) && end != qp->len))
498 qp->last_in |= LAST_IN;
503 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
504 skb->ip_summed = CHECKSUM_NONE;
507 /* Some bits beyond end -> corruption. */
508 if (qp->last_in & LAST_IN)
516 if (pskb_pull(skb, ihl) == NULL)
518 if (pskb_trim_rcsum(skb, end-offset))
521 /* Find out which fragments are in front and at the back of us
522 * in the chain of fragments so far. We must know where to put
523 * this fragment, right?
526 for(next = qp->fragments; next != NULL; next = next->next) {
527 if (FRAG_CB(next)->offset >= offset)
532 /* We found where to put this one. Check for overlap with
533 * preceding fragment, and, if needed, align things so that
534 * any overlaps are eliminated.
537 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;
550 while (next && FRAG_CB(next)->offset < end) {
551 int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
554 /* Eat head of the next overlapped fragment
555 * and leave the loop. The next ones cannot overlap.
557 if (!pskb_pull(next, i))
559 FRAG_CB(next)->offset += i;
561 if (next->ip_summed != CHECKSUM_UNNECESSARY)
562 next->ip_summed = CHECKSUM_NONE;
565 struct sk_buff *free_it = next;
567 /* Old fragmnet is completely overridden with
575 qp->fragments = next;
577 qp->meat -= free_it->len;
578 frag_kfree_skb(free_it, NULL);
582 FRAG_CB(skb)->offset = offset;
584 /* Insert this fragment in the chain of fragments. */
592 qp->iif = skb->dev->ifindex;
594 skb_get_timestamp(skb, &qp->stamp);
595 qp->meat += skb->len;
596 atomic_add(skb->truesize, &ip_frag_mem);
598 qp->last_in |= FIRST_IN;
600 write_lock(&ipfrag_lock);
601 list_move_tail(&qp->lru_list, &ipq_lru_list);
602 write_unlock(&ipfrag_lock);
611 /* Build a new IP datagram from all its fragments. */
613 static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
616 struct sk_buff *fp, *head = qp->fragments;
622 BUG_TRAP(head != NULL);
623 BUG_TRAP(FRAG_CB(head)->offset == 0);
625 /* Allocate a new buffer for the datagram. */
626 ihlen = head->nh.iph->ihl*4;
627 len = ihlen + qp->len;
632 /* Head of list must not be cloned. */
633 if (skb_cloned(head) && pskb_expand_head(head, 0, 0, GFP_ATOMIC))
636 /* If the first fragment is fragmented itself, we split
637 * it to two chunks: the first with data and paged part
638 * and the second, holding only fragments. */
639 if (skb_shinfo(head)->frag_list) {
640 struct sk_buff *clone;
643 if ((clone = alloc_skb(0, GFP_ATOMIC)) == NULL)
645 clone->next = head->next;
647 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
648 skb_shinfo(head)->frag_list = NULL;
649 for (i=0; i<skb_shinfo(head)->nr_frags; i++)
650 plen += skb_shinfo(head)->frags[i].size;
651 clone->len = clone->data_len = head->data_len - plen;
652 head->data_len -= clone->len;
653 head->len -= clone->len;
655 clone->ip_summed = head->ip_summed;
656 atomic_add(clone->truesize, &ip_frag_mem);
659 skb_shinfo(head)->frag_list = head->next;
660 skb_push(head, head->data - head->nh.raw);
661 atomic_sub(head->truesize, &ip_frag_mem);
663 for (fp=head->next; fp; fp = fp->next) {
664 head->data_len += fp->len;
665 head->len += fp->len;
666 if (head->ip_summed != fp->ip_summed)
667 head->ip_summed = CHECKSUM_NONE;
668 else if (head->ip_summed == CHECKSUM_HW)
669 head->csum = csum_add(head->csum, fp->csum);
670 head->truesize += fp->truesize;
671 atomic_sub(fp->truesize, &ip_frag_mem);
676 skb_set_timestamp(head, &qp->stamp);
680 iph->tot_len = htons(len);
681 IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS);
682 qp->fragments = NULL;
686 LIMIT_NETDEBUG(KERN_ERR "IP: queue_glue: no memory for gluing "
692 "Oversized IP packet from %d.%d.%d.%d.\n",
695 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
699 /* Process an incoming IP datagram fragment. */
700 struct sk_buff *ip_defrag(struct sk_buff *skb, u32 user)
702 struct iphdr *iph = skb->nh.iph;
704 struct net_device *dev;
706 IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS);
708 /* Start by cleaning up the memory. */
709 if (atomic_read(&ip_frag_mem) > sysctl_ipfrag_high_thresh)
714 /* Lookup (or create) queue header */
715 if ((qp = ip_find(iph, user)) != NULL) {
716 struct sk_buff *ret = NULL;
718 spin_lock(&qp->lock);
720 ip_frag_queue(qp, skb);
722 if (qp->last_in == (FIRST_IN|LAST_IN) &&
724 ret = ip_frag_reasm(qp, dev);
726 spin_unlock(&qp->lock);
731 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
736 void ipfrag_init(void)
738 ipfrag_hash_rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
739 (jiffies ^ (jiffies >> 6)));
741 init_timer(&ipfrag_secret_timer);
742 ipfrag_secret_timer.function = ipfrag_secret_rebuild;
743 ipfrag_secret_timer.expires = jiffies + sysctl_ipfrag_secret_interval;
744 add_timer(&ipfrag_secret_timer);
747 EXPORT_SYMBOL(ip_defrag);