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 __read_mostly = 256*1024;
58 int sysctl_ipfrag_low_thresh __read_mostly = 192*1024;
60 int sysctl_ipfrag_max_dist __read_mostly = 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 __read_mostly = 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(__be16 id, __be32 saddr, __be32 daddr, u8 prot)
128 return jhash_3words((__force u32)id << 16 | prot,
129 (__force u32)saddr, (__force u32)daddr,
130 ipfrag_hash_rnd) & (IPQ_HASHSZ - 1);
133 static struct timer_list ipfrag_secret_timer;
134 int sysctl_ipfrag_secret_interval __read_mostly = 10 * 60 * HZ;
136 static void ipfrag_secret_rebuild(unsigned long dummy)
138 unsigned long now = jiffies;
141 write_lock(&ipfrag_lock);
142 get_random_bytes(&ipfrag_hash_rnd, sizeof(u32));
143 for (i = 0; i < IPQ_HASHSZ; i++) {
145 struct hlist_node *p, *n;
147 hlist_for_each_entry_safe(q, p, n, &ipq_hash[i], list) {
148 unsigned int hval = ipqhashfn(q->id, q->saddr,
149 q->daddr, q->protocol);
154 /* Relink to new hash chain. */
155 hlist_add_head(&q->list, &ipq_hash[hval]);
159 write_unlock(&ipfrag_lock);
161 mod_timer(&ipfrag_secret_timer, now + sysctl_ipfrag_secret_interval);
164 atomic_t ip_frag_mem = ATOMIC_INIT(0); /* Memory used for fragments */
166 /* Memory Tracking Functions. */
167 static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
170 *work -= skb->truesize;
171 atomic_sub(skb->truesize, &ip_frag_mem);
175 static __inline__ void frag_free_queue(struct ipq *qp, int *work)
178 *work -= sizeof(struct ipq);
179 atomic_sub(sizeof(struct ipq), &ip_frag_mem);
183 static __inline__ struct ipq *frag_alloc_queue(void)
185 struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
189 atomic_add(sizeof(struct ipq), &ip_frag_mem);
194 /* Destruction primitives. */
196 /* Complete destruction of ipq. */
197 static void ip_frag_destroy(struct ipq *qp, int *work)
201 BUG_TRAP(qp->last_in&COMPLETE);
202 BUG_TRAP(del_timer(&qp->timer) == 0);
205 inet_putpeer(qp->peer);
207 /* Release all fragment data. */
210 struct sk_buff *xp = fp->next;
212 frag_kfree_skb(fp, work);
216 /* Finally, release the queue descriptor itself. */
217 frag_free_queue(qp, work);
220 static __inline__ void ipq_put(struct ipq *ipq, int *work)
222 if (atomic_dec_and_test(&ipq->refcnt))
223 ip_frag_destroy(ipq, work);
226 /* Kill ipq entry. It is not destroyed immediately,
227 * because caller (and someone more) holds reference count.
229 static void ipq_kill(struct ipq *ipq)
231 if (del_timer(&ipq->timer))
232 atomic_dec(&ipq->refcnt);
234 if (!(ipq->last_in & COMPLETE)) {
236 atomic_dec(&ipq->refcnt);
237 ipq->last_in |= COMPLETE;
241 /* Memory limiting on fragments. Evictor trashes the oldest
242 * fragment queue until we are back under the threshold.
244 static void ip_evictor(void)
247 struct list_head *tmp;
250 work = atomic_read(&ip_frag_mem) - sysctl_ipfrag_low_thresh;
255 read_lock(&ipfrag_lock);
256 if (list_empty(&ipq_lru_list)) {
257 read_unlock(&ipfrag_lock);
260 tmp = ipq_lru_list.next;
261 qp = list_entry(tmp, struct ipq, lru_list);
262 atomic_inc(&qp->refcnt);
263 read_unlock(&ipfrag_lock);
265 spin_lock(&qp->lock);
266 if (!(qp->last_in&COMPLETE))
268 spin_unlock(&qp->lock);
271 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
276 * Oops, a fragment queue timed out. Kill it and send an ICMP reply.
278 static void ip_expire(unsigned long arg)
280 struct ipq *qp = (struct ipq *) arg;
282 spin_lock(&qp->lock);
284 if (qp->last_in & COMPLETE)
289 IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT);
290 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
292 if ((qp->last_in&FIRST_IN) && qp->fragments != NULL) {
293 struct sk_buff *head = qp->fragments;
294 /* Send an ICMP "Fragment Reassembly Timeout" message. */
295 if ((head->dev = dev_get_by_index(qp->iif)) != NULL) {
296 icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
301 spin_unlock(&qp->lock);
305 /* Creation primitives. */
307 static struct ipq *ip_frag_intern(struct ipq *qp_in)
311 struct hlist_node *n;
315 write_lock(&ipfrag_lock);
316 hash = ipqhashfn(qp_in->id, qp_in->saddr, qp_in->daddr,
319 /* With SMP race we have to recheck hash table, because
320 * such entry could be created on other cpu, while we
321 * promoted read lock to write lock.
323 hlist_for_each_entry(qp, n, &ipq_hash[hash], list) {
324 if (qp->id == qp_in->id &&
325 qp->saddr == qp_in->saddr &&
326 qp->daddr == qp_in->daddr &&
327 qp->protocol == qp_in->protocol &&
328 qp->user == qp_in->user) {
329 atomic_inc(&qp->refcnt);
330 write_unlock(&ipfrag_lock);
331 qp_in->last_in |= COMPLETE;
332 ipq_put(qp_in, NULL);
339 if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time))
340 atomic_inc(&qp->refcnt);
342 atomic_inc(&qp->refcnt);
343 hlist_add_head(&qp->list, &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(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;
369 qp->peer = sysctl_ipfrag_max_dist ? inet_getpeer(iph->saddr, 1) : NULL;
371 /* Initialize a timer for this entry. */
372 init_timer(&qp->timer);
373 qp->timer.data = (unsigned long) qp; /* pointer to queue */
374 qp->timer.function = ip_expire; /* expire function */
375 spin_lock_init(&qp->lock);
376 atomic_set(&qp->refcnt, 1);
378 return ip_frag_intern(qp);
381 LIMIT_NETDEBUG(KERN_ERR "ip_frag_create: no memory left !\n");
385 /* Find the correct entry in the "incomplete datagrams" queue for
386 * this IP datagram, and create new one, if nothing is found.
388 static inline struct ipq *ip_find(struct iphdr *iph, u32 user)
391 __be32 saddr = iph->saddr;
392 __be32 daddr = iph->daddr;
393 __u8 protocol = iph->protocol;
396 struct hlist_node *n;
398 read_lock(&ipfrag_lock);
399 hash = ipqhashfn(id, saddr, daddr, protocol);
400 hlist_for_each_entry(qp, n, &ipq_hash[hash], list) {
402 qp->saddr == saddr &&
403 qp->daddr == daddr &&
404 qp->protocol == protocol &&
406 atomic_inc(&qp->refcnt);
407 read_unlock(&ipfrag_lock);
411 read_unlock(&ipfrag_lock);
413 return ip_frag_create(iph, user);
416 /* Is the fragment too far ahead to be part of ipq? */
417 static inline int ip_frag_too_far(struct ipq *qp)
419 struct inet_peer *peer = qp->peer;
420 unsigned int max = sysctl_ipfrag_max_dist;
421 unsigned int start, end;
429 end = atomic_inc_return(&peer->rid);
432 rc = qp->fragments && (end - start) > max;
435 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
441 static int ip_frag_reinit(struct ipq *qp)
445 if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time)) {
446 atomic_inc(&qp->refcnt);
452 struct sk_buff *xp = fp->next;
453 frag_kfree_skb(fp, NULL);
460 qp->fragments = NULL;
466 /* Add new segment to existing queue. */
467 static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
469 struct sk_buff *prev, *next;
473 if (qp->last_in & COMPLETE)
476 if (!(IPCB(skb)->flags & IPSKB_FRAG_COMPLETE) &&
477 unlikely(ip_frag_too_far(qp)) && unlikely(ip_frag_reinit(qp))) {
482 offset = ntohs(ip_hdr(skb)->frag_off);
483 flags = offset & ~IP_OFFSET;
485 offset <<= 3; /* offset is in 8-byte chunks */
486 ihl = ip_hdrlen(skb);
488 /* Determine the position of this fragment. */
489 end = offset + skb->len - ihl;
491 /* Is this the final fragment? */
492 if ((flags & IP_MF) == 0) {
493 /* If we already have some bits beyond end
494 * or have different end, the segment is corrrupted.
497 ((qp->last_in & LAST_IN) && end != qp->len))
499 qp->last_in |= LAST_IN;
504 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
505 skb->ip_summed = CHECKSUM_NONE;
508 /* Some bits beyond end -> corruption. */
509 if (qp->last_in & LAST_IN)
517 if (pskb_pull(skb, ihl) == NULL)
519 if (pskb_trim_rcsum(skb, end-offset))
522 /* Find out which fragments are in front and at the back of us
523 * in the chain of fragments so far. We must know where to put
524 * this fragment, right?
527 for (next = qp->fragments; next != NULL; next = next->next) {
528 if (FRAG_CB(next)->offset >= offset)
533 /* We found where to put this one. Check for overlap with
534 * preceding fragment, and, if needed, align things so that
535 * any overlaps are eliminated.
538 int i = (FRAG_CB(prev)->offset + prev->len) - offset;
544 if (!pskb_pull(skb, i))
546 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
547 skb->ip_summed = CHECKSUM_NONE;
551 while (next && FRAG_CB(next)->offset < end) {
552 int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
555 /* Eat head of the next overlapped fragment
556 * and leave the loop. The next ones cannot overlap.
558 if (!pskb_pull(next, i))
560 FRAG_CB(next)->offset += i;
562 if (next->ip_summed != CHECKSUM_UNNECESSARY)
563 next->ip_summed = CHECKSUM_NONE;
566 struct sk_buff *free_it = next;
568 /* Old fragment is completely overridden with
576 qp->fragments = next;
578 qp->meat -= free_it->len;
579 frag_kfree_skb(free_it, NULL);
583 FRAG_CB(skb)->offset = offset;
585 /* Insert this fragment in the chain of fragments. */
593 qp->iif = skb->dev->ifindex;
595 qp->stamp = skb->tstamp;
596 qp->meat += skb->len;
597 atomic_add(skb->truesize, &ip_frag_mem);
599 qp->last_in |= FIRST_IN;
601 write_lock(&ipfrag_lock);
602 list_move_tail(&qp->lru_list, &ipq_lru_list);
603 write_unlock(&ipfrag_lock);
612 /* Build a new IP datagram from all its fragments. */
614 static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
617 struct sk_buff *fp, *head = qp->fragments;
623 BUG_TRAP(head != NULL);
624 BUG_TRAP(FRAG_CB(head)->offset == 0);
626 /* Allocate a new buffer for the datagram. */
627 ihlen = ip_hdrlen(head);
628 len = ihlen + qp->len;
633 /* Head of list must not be cloned. */
634 if (skb_cloned(head) && pskb_expand_head(head, 0, 0, GFP_ATOMIC))
637 /* If the first fragment is fragmented itself, we split
638 * it to two chunks: the first with data and paged part
639 * and the second, holding only fragments. */
640 if (skb_shinfo(head)->frag_list) {
641 struct sk_buff *clone;
644 if ((clone = alloc_skb(0, GFP_ATOMIC)) == NULL)
646 clone->next = head->next;
648 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
649 skb_shinfo(head)->frag_list = NULL;
650 for (i=0; i<skb_shinfo(head)->nr_frags; i++)
651 plen += skb_shinfo(head)->frags[i].size;
652 clone->len = clone->data_len = head->data_len - plen;
653 head->data_len -= clone->len;
654 head->len -= clone->len;
656 clone->ip_summed = head->ip_summed;
657 atomic_add(clone->truesize, &ip_frag_mem);
660 skb_shinfo(head)->frag_list = head->next;
661 skb_push(head, head->data - skb_network_header(head));
662 atomic_sub(head->truesize, &ip_frag_mem);
664 for (fp=head->next; fp; fp = fp->next) {
665 head->data_len += fp->len;
666 head->len += fp->len;
667 if (head->ip_summed != fp->ip_summed)
668 head->ip_summed = CHECKSUM_NONE;
669 else if (head->ip_summed == CHECKSUM_COMPLETE)
670 head->csum = csum_add(head->csum, fp->csum);
671 head->truesize += fp->truesize;
672 atomic_sub(fp->truesize, &ip_frag_mem);
677 head->tstamp = qp->stamp;
681 iph->tot_len = htons(len);
682 IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS);
683 qp->fragments = NULL;
687 LIMIT_NETDEBUG(KERN_ERR "IP: queue_glue: no memory for gluing "
693 "Oversized IP packet from %d.%d.%d.%d.\n",
696 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
700 /* Process an incoming IP datagram fragment. */
701 struct sk_buff *ip_defrag(struct sk_buff *skb, u32 user)
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(ip_hdr(skb), 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 __init 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);