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/config.h>
27 #include <linux/module.h>
28 #include <linux/types.h>
30 #include <linux/jiffies.h>
31 #include <linux/skbuff.h>
32 #include <linux/list.h>
34 #include <linux/icmp.h>
35 #include <linux/netdevice.h>
36 #include <linux/jhash.h>
37 #include <linux/random.h>
41 #include <net/checksum.h>
42 #include <net/inetpeer.h>
43 #include <linux/tcp.h>
44 #include <linux/udp.h>
45 #include <linux/inet.h>
46 #include <linux/netfilter_ipv4.h>
48 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
49 * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
50 * as well. Or notify me, at least. --ANK
53 /* Fragment cache limits. We will commit 256K at one time. Should we
54 * cross that limit we will prune down to 192K. This should cope with
55 * even the most extreme cases without allowing an attacker to measurably
56 * harm machine performance.
58 int sysctl_ipfrag_high_thresh
= 256*1024;
59 int sysctl_ipfrag_low_thresh
= 192*1024;
61 int sysctl_ipfrag_max_dist
= 64;
63 /* Important NOTE! Fragment queue must be destroyed before MSL expires.
64 * RFC791 is wrong proposing to prolongate timer each fragment arrival by TTL.
66 int sysctl_ipfrag_time
= IP_FRAG_TIME
;
70 struct inet_skb_parm h
;
74 #define FRAG_CB(skb) ((struct ipfrag_skb_cb*)((skb)->cb))
76 /* Describe an entry in the "incomplete datagrams" queue. */
78 struct hlist_node list
;
79 struct list_head lru_list
; /* lru list member */
90 struct sk_buff
*fragments
; /* linked list of received fragments */
91 int len
; /* total length of original datagram */
95 struct timer_list timer
; /* when will this queue expire? */
99 struct inet_peer
*peer
;
104 #define IPQ_HASHSZ 64
106 /* Per-bucket lock is easy to add now. */
107 static struct hlist_head ipq_hash
[IPQ_HASHSZ
];
108 static DEFINE_RWLOCK(ipfrag_lock
);
109 static u32 ipfrag_hash_rnd
;
110 static LIST_HEAD(ipq_lru_list
);
111 int ip_frag_nqueues
= 0;
113 static __inline__
void __ipq_unlink(struct ipq
*qp
)
115 hlist_del(&qp
->list
);
116 list_del(&qp
->lru_list
);
120 static __inline__
void ipq_unlink(struct ipq
*ipq
)
122 write_lock(&ipfrag_lock
);
124 write_unlock(&ipfrag_lock
);
127 static unsigned int ipqhashfn(u16 id
, u32 saddr
, u32 daddr
, u8 prot
)
129 return jhash_3words((u32
)id
<< 16 | prot
, saddr
, daddr
,
130 ipfrag_hash_rnd
) & (IPQ_HASHSZ
- 1);
133 static struct timer_list ipfrag_secret_timer
;
134 int sysctl_ipfrag_secret_interval
= 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(unsigned int hash
, struct ipq
*qp_in
)
311 struct hlist_node
*n
;
313 write_lock(&ipfrag_lock
);
315 /* With SMP race we have to recheck hash table, because
316 * such entry could be created on other cpu, while we
317 * promoted read lock to write lock.
319 hlist_for_each_entry(qp
, n
, &ipq_hash
[hash
], list
) {
320 if(qp
->id
== qp_in
->id
&&
321 qp
->saddr
== qp_in
->saddr
&&
322 qp
->daddr
== qp_in
->daddr
&&
323 qp
->protocol
== qp_in
->protocol
&&
324 qp
->user
== qp_in
->user
) {
325 atomic_inc(&qp
->refcnt
);
326 write_unlock(&ipfrag_lock
);
327 qp_in
->last_in
|= COMPLETE
;
328 ipq_put(qp_in
, NULL
);
335 if (!mod_timer(&qp
->timer
, jiffies
+ sysctl_ipfrag_time
))
336 atomic_inc(&qp
->refcnt
);
338 atomic_inc(&qp
->refcnt
);
339 hlist_add_head(&qp
->list
, &ipq_hash
[hash
]);
340 INIT_LIST_HEAD(&qp
->lru_list
);
341 list_add_tail(&qp
->lru_list
, &ipq_lru_list
);
343 write_unlock(&ipfrag_lock
);
347 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
348 static struct ipq
*ip_frag_create(unsigned hash
, struct iphdr
*iph
, u32 user
)
352 if ((qp
= frag_alloc_queue()) == NULL
)
355 qp
->protocol
= iph
->protocol
;
358 qp
->saddr
= iph
->saddr
;
359 qp
->daddr
= iph
->daddr
;
363 qp
->fragments
= NULL
;
365 qp
->peer
= sysctl_ipfrag_max_dist
? inet_getpeer(iph
->saddr
, 1) : NULL
;
367 /* Initialize a timer for this entry. */
368 init_timer(&qp
->timer
);
369 qp
->timer
.data
= (unsigned long) qp
; /* pointer to queue */
370 qp
->timer
.function
= ip_expire
; /* expire function */
371 spin_lock_init(&qp
->lock
);
372 atomic_set(&qp
->refcnt
, 1);
374 return ip_frag_intern(hash
, qp
);
377 LIMIT_NETDEBUG(KERN_ERR
"ip_frag_create: no memory left !\n");
381 /* Find the correct entry in the "incomplete datagrams" queue for
382 * this IP datagram, and create new one, if nothing is found.
384 static inline struct ipq
*ip_find(struct iphdr
*iph
, u32 user
)
387 __u32 saddr
= iph
->saddr
;
388 __u32 daddr
= iph
->daddr
;
389 __u8 protocol
= iph
->protocol
;
390 unsigned int hash
= ipqhashfn(id
, saddr
, daddr
, protocol
);
392 struct hlist_node
*n
;
394 read_lock(&ipfrag_lock
);
395 hlist_for_each_entry(qp
, n
, &ipq_hash
[hash
], list
) {
397 qp
->saddr
== saddr
&&
398 qp
->daddr
== daddr
&&
399 qp
->protocol
== protocol
&&
401 atomic_inc(&qp
->refcnt
);
402 read_unlock(&ipfrag_lock
);
406 read_unlock(&ipfrag_lock
);
408 return ip_frag_create(hash
, iph
, user
);
411 /* Is the fragment too far ahead to be part of ipq? */
412 static inline int ip_frag_too_far(struct ipq
*qp
)
414 struct inet_peer
*peer
= qp
->peer
;
415 unsigned int max
= sysctl_ipfrag_max_dist
;
416 unsigned int start
, end
;
424 end
= atomic_inc_return(&peer
->rid
);
427 rc
= qp
->fragments
&& (end
- start
) > max
;
430 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
436 static int ip_frag_reinit(struct ipq
*qp
)
440 if (!mod_timer(&qp
->timer
, jiffies
+ sysctl_ipfrag_time
)) {
441 atomic_inc(&qp
->refcnt
);
447 struct sk_buff
*xp
= fp
->next
;
448 frag_kfree_skb(fp
, NULL
);
455 qp
->fragments
= NULL
;
461 /* Add new segment to existing queue. */
462 static void ip_frag_queue(struct ipq
*qp
, struct sk_buff
*skb
)
464 struct sk_buff
*prev
, *next
;
468 if (qp
->last_in
& COMPLETE
)
471 if (!(IPCB(skb
)->flags
& IPSKB_FRAG_COMPLETE
) &&
472 unlikely(ip_frag_too_far(qp
)) && unlikely(ip_frag_reinit(qp
))) {
477 offset
= ntohs(skb
->nh
.iph
->frag_off
);
478 flags
= offset
& ~IP_OFFSET
;
480 offset
<<= 3; /* offset is in 8-byte chunks */
481 ihl
= skb
->nh
.iph
->ihl
* 4;
483 /* Determine the position of this fragment. */
484 end
= offset
+ skb
->len
- ihl
;
486 /* Is this the final fragment? */
487 if ((flags
& IP_MF
) == 0) {
488 /* If we already have some bits beyond end
489 * or have different end, the segment is corrrupted.
492 ((qp
->last_in
& LAST_IN
) && end
!= qp
->len
))
494 qp
->last_in
|= LAST_IN
;
499 if (skb
->ip_summed
!= CHECKSUM_UNNECESSARY
)
500 skb
->ip_summed
= CHECKSUM_NONE
;
503 /* Some bits beyond end -> corruption. */
504 if (qp
->last_in
& LAST_IN
)
512 if (pskb_pull(skb
, ihl
) == NULL
)
514 if (pskb_trim_rcsum(skb
, end
-offset
))
517 /* Find out which fragments are in front and at the back of us
518 * in the chain of fragments so far. We must know where to put
519 * this fragment, right?
522 for(next
= qp
->fragments
; next
!= NULL
; next
= next
->next
) {
523 if (FRAG_CB(next
)->offset
>= offset
)
528 /* We found where to put this one. Check for overlap with
529 * preceding fragment, and, if needed, align things so that
530 * any overlaps are eliminated.
533 int i
= (FRAG_CB(prev
)->offset
+ prev
->len
) - offset
;
539 if (!pskb_pull(skb
, i
))
541 if (skb
->ip_summed
!= CHECKSUM_UNNECESSARY
)
542 skb
->ip_summed
= CHECKSUM_NONE
;
546 while (next
&& FRAG_CB(next
)->offset
< end
) {
547 int i
= end
- FRAG_CB(next
)->offset
; /* overlap is 'i' bytes */
550 /* Eat head of the next overlapped fragment
551 * and leave the loop. The next ones cannot overlap.
553 if (!pskb_pull(next
, i
))
555 FRAG_CB(next
)->offset
+= i
;
557 if (next
->ip_summed
!= CHECKSUM_UNNECESSARY
)
558 next
->ip_summed
= CHECKSUM_NONE
;
561 struct sk_buff
*free_it
= next
;
563 /* Old fragmnet is completely overridden with
571 qp
->fragments
= next
;
573 qp
->meat
-= free_it
->len
;
574 frag_kfree_skb(free_it
, NULL
);
578 FRAG_CB(skb
)->offset
= offset
;
580 /* Insert this fragment in the chain of fragments. */
588 qp
->iif
= skb
->dev
->ifindex
;
590 skb_get_timestamp(skb
, &qp
->stamp
);
591 qp
->meat
+= skb
->len
;
592 atomic_add(skb
->truesize
, &ip_frag_mem
);
594 qp
->last_in
|= FIRST_IN
;
596 write_lock(&ipfrag_lock
);
597 list_move_tail(&qp
->lru_list
, &ipq_lru_list
);
598 write_unlock(&ipfrag_lock
);
607 /* Build a new IP datagram from all its fragments. */
609 static struct sk_buff
*ip_frag_reasm(struct ipq
*qp
, struct net_device
*dev
)
612 struct sk_buff
*fp
, *head
= qp
->fragments
;
618 BUG_TRAP(head
!= NULL
);
619 BUG_TRAP(FRAG_CB(head
)->offset
== 0);
621 /* Allocate a new buffer for the datagram. */
622 ihlen
= head
->nh
.iph
->ihl
*4;
623 len
= ihlen
+ qp
->len
;
628 /* Head of list must not be cloned. */
629 if (skb_cloned(head
) && pskb_expand_head(head
, 0, 0, GFP_ATOMIC
))
632 /* If the first fragment is fragmented itself, we split
633 * it to two chunks: the first with data and paged part
634 * and the second, holding only fragments. */
635 if (skb_shinfo(head
)->frag_list
) {
636 struct sk_buff
*clone
;
639 if ((clone
= alloc_skb(0, GFP_ATOMIC
)) == NULL
)
641 clone
->next
= head
->next
;
643 skb_shinfo(clone
)->frag_list
= skb_shinfo(head
)->frag_list
;
644 skb_shinfo(head
)->frag_list
= NULL
;
645 for (i
=0; i
<skb_shinfo(head
)->nr_frags
; i
++)
646 plen
+= skb_shinfo(head
)->frags
[i
].size
;
647 clone
->len
= clone
->data_len
= head
->data_len
- plen
;
648 head
->data_len
-= clone
->len
;
649 head
->len
-= clone
->len
;
651 clone
->ip_summed
= head
->ip_summed
;
652 atomic_add(clone
->truesize
, &ip_frag_mem
);
655 skb_shinfo(head
)->frag_list
= head
->next
;
656 skb_push(head
, head
->data
- head
->nh
.raw
);
657 atomic_sub(head
->truesize
, &ip_frag_mem
);
659 for (fp
=head
->next
; fp
; fp
= fp
->next
) {
660 head
->data_len
+= fp
->len
;
661 head
->len
+= fp
->len
;
662 if (head
->ip_summed
!= fp
->ip_summed
)
663 head
->ip_summed
= CHECKSUM_NONE
;
664 else if (head
->ip_summed
== CHECKSUM_HW
)
665 head
->csum
= csum_add(head
->csum
, fp
->csum
);
666 head
->truesize
+= fp
->truesize
;
667 atomic_sub(fp
->truesize
, &ip_frag_mem
);
672 skb_set_timestamp(head
, &qp
->stamp
);
676 iph
->tot_len
= htons(len
);
677 IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS
);
678 qp
->fragments
= NULL
;
682 LIMIT_NETDEBUG(KERN_ERR
"IP: queue_glue: no memory for gluing "
688 "Oversized IP packet from %d.%d.%d.%d.\n",
691 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
695 /* Process an incoming IP datagram fragment. */
696 struct sk_buff
*ip_defrag(struct sk_buff
*skb
, u32 user
)
698 struct iphdr
*iph
= skb
->nh
.iph
;
700 struct net_device
*dev
;
702 IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS
);
704 /* Start by cleaning up the memory. */
705 if (atomic_read(&ip_frag_mem
) > sysctl_ipfrag_high_thresh
)
710 /* Lookup (or create) queue header */
711 if ((qp
= ip_find(iph
, user
)) != NULL
) {
712 struct sk_buff
*ret
= NULL
;
714 spin_lock(&qp
->lock
);
716 ip_frag_queue(qp
, skb
);
718 if (qp
->last_in
== (FIRST_IN
|LAST_IN
) &&
720 ret
= ip_frag_reasm(qp
, dev
);
722 spin_unlock(&qp
->lock
);
727 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
732 void ipfrag_init(void)
734 ipfrag_hash_rnd
= (u32
) ((num_physpages
^ (num_physpages
>>7)) ^
735 (jiffies
^ (jiffies
>> 6)));
737 init_timer(&ipfrag_secret_timer
);
738 ipfrag_secret_timer
.function
= ipfrag_secret_rebuild
;
739 ipfrag_secret_timer
.expires
= jiffies
+ sysctl_ipfrag_secret_interval
;
740 add_timer(&ipfrag_secret_timer
);
743 EXPORT_SYMBOL(ip_defrag
);