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 <net/inet_frag.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 int sysctl_ipfrag_max_dist __read_mostly
= 64;
57 struct inet_skb_parm h
;
61 #define FRAG_CB(skb) ((struct ipfrag_skb_cb*)((skb)->cb))
63 /* Describe an entry in the "incomplete datagrams" queue. */
65 struct inet_frag_queue q
;
74 struct inet_peer
*peer
;
77 struct inet_frags_ctl ip4_frags_ctl __read_mostly
= {
79 * Fragment cache limits. We will commit 256K at one time. Should we
80 * cross that limit we will prune down to 192K. This should cope with
81 * even the most extreme cases without allowing an attacker to
82 * measurably harm machine performance.
84 .high_thresh
= 256 * 1024,
85 .low_thresh
= 192 * 1024,
88 * Important NOTE! Fragment queue must be destroyed before MSL expires.
89 * RFC791 is wrong proposing to prolongate timer each fragment arrival
92 .timeout
= IP_FRAG_TIME
,
93 .secret_interval
= 10 * 60 * HZ
,
96 static struct inet_frags ip4_frags
;
98 int ip_frag_nqueues(void)
100 return ip4_frags
.nqueues
;
103 int ip_frag_mem(void)
105 return atomic_read(&ip4_frags
.mem
);
108 static int ip_frag_reasm(struct ipq
*qp
, struct sk_buff
*prev
,
109 struct net_device
*dev
);
111 static unsigned int ipqhashfn(__be16 id
, __be32 saddr
, __be32 daddr
, u8 prot
)
113 return jhash_3words((__force u32
)id
<< 16 | prot
,
114 (__force u32
)saddr
, (__force u32
)daddr
,
115 ip4_frags
.rnd
) & (INETFRAGS_HASHSZ
- 1);
118 static unsigned int ip4_hashfn(struct inet_frag_queue
*q
)
122 ipq
= container_of(q
, struct ipq
, q
);
123 return ipqhashfn(ipq
->id
, ipq
->saddr
, ipq
->daddr
, ipq
->protocol
);
126 /* Memory Tracking Functions. */
127 static __inline__
void frag_kfree_skb(struct sk_buff
*skb
, int *work
)
130 *work
-= skb
->truesize
;
131 atomic_sub(skb
->truesize
, &ip4_frags
.mem
);
135 static __inline__
void ip4_frag_free(struct inet_frag_queue
*q
)
139 qp
= container_of(q
, struct ipq
, q
);
141 inet_putpeer(qp
->peer
);
145 static __inline__
struct ipq
*frag_alloc_queue(void)
147 struct ipq
*qp
= kmalloc(sizeof(struct ipq
), GFP_ATOMIC
);
151 atomic_add(sizeof(struct ipq
), &ip4_frags
.mem
);
156 /* Destruction primitives. */
158 static __inline__
void ipq_put(struct ipq
*ipq
, int *work
)
160 if (atomic_dec_and_test(&ipq
->q
.refcnt
))
161 inet_frag_destroy(&ipq
->q
, &ip4_frags
, work
);
164 /* Kill ipq entry. It is not destroyed immediately,
165 * because caller (and someone more) holds reference count.
167 static void ipq_kill(struct ipq
*ipq
)
169 inet_frag_kill(&ipq
->q
, &ip4_frags
);
172 /* Memory limiting on fragments. Evictor trashes the oldest
173 * fragment queue until we are back under the threshold.
175 static void ip_evictor(void)
178 struct list_head
*tmp
;
181 work
= atomic_read(&ip4_frags
.mem
) - ip4_frags_ctl
.low_thresh
;
186 read_lock(&ip4_frags
.lock
);
187 if (list_empty(&ip4_frags
.lru_list
)) {
188 read_unlock(&ip4_frags
.lock
);
191 tmp
= ip4_frags
.lru_list
.next
;
192 qp
= list_entry(tmp
, struct ipq
, q
.lru_list
);
193 atomic_inc(&qp
->q
.refcnt
);
194 read_unlock(&ip4_frags
.lock
);
196 spin_lock(&qp
->q
.lock
);
197 if (!(qp
->q
.last_in
&COMPLETE
))
199 spin_unlock(&qp
->q
.lock
);
202 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
207 * Oops, a fragment queue timed out. Kill it and send an ICMP reply.
209 static void ip_expire(unsigned long arg
)
211 struct ipq
*qp
= (struct ipq
*) arg
;
213 spin_lock(&qp
->q
.lock
);
215 if (qp
->q
.last_in
& COMPLETE
)
220 IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT
);
221 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
223 if ((qp
->q
.last_in
&FIRST_IN
) && qp
->q
.fragments
!= NULL
) {
224 struct sk_buff
*head
= qp
->q
.fragments
;
225 /* Send an ICMP "Fragment Reassembly Timeout" message. */
226 if ((head
->dev
= dev_get_by_index(&init_net
, qp
->iif
)) != NULL
) {
227 icmp_send(head
, ICMP_TIME_EXCEEDED
, ICMP_EXC_FRAGTIME
, 0);
232 spin_unlock(&qp
->q
.lock
);
236 /* Creation primitives. */
238 static struct ipq
*ip_frag_intern(struct ipq
*qp_in
)
242 struct hlist_node
*n
;
246 write_lock(&ip4_frags
.lock
);
247 hash
= ipqhashfn(qp_in
->id
, qp_in
->saddr
, qp_in
->daddr
,
250 /* With SMP race we have to recheck hash table, because
251 * such entry could be created on other cpu, while we
252 * promoted read lock to write lock.
254 hlist_for_each_entry(qp
, n
, &ip4_frags
.hash
[hash
], q
.list
) {
255 if (qp
->id
== qp_in
->id
&&
256 qp
->saddr
== qp_in
->saddr
&&
257 qp
->daddr
== qp_in
->daddr
&&
258 qp
->protocol
== qp_in
->protocol
&&
259 qp
->user
== qp_in
->user
) {
260 atomic_inc(&qp
->q
.refcnt
);
261 write_unlock(&ip4_frags
.lock
);
262 qp_in
->q
.last_in
|= COMPLETE
;
263 ipq_put(qp_in
, NULL
);
270 if (!mod_timer(&qp
->q
.timer
, jiffies
+ ip4_frags_ctl
.timeout
))
271 atomic_inc(&qp
->q
.refcnt
);
273 atomic_inc(&qp
->q
.refcnt
);
274 hlist_add_head(&qp
->q
.list
, &ip4_frags
.hash
[hash
]);
275 INIT_LIST_HEAD(&qp
->q
.lru_list
);
276 list_add_tail(&qp
->q
.lru_list
, &ip4_frags
.lru_list
);
278 write_unlock(&ip4_frags
.lock
);
282 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
283 static struct ipq
*ip_frag_create(struct iphdr
*iph
, u32 user
)
287 if ((qp
= frag_alloc_queue()) == NULL
)
290 qp
->protocol
= iph
->protocol
;
293 qp
->saddr
= iph
->saddr
;
294 qp
->daddr
= iph
->daddr
;
298 qp
->q
.fragments
= NULL
;
300 qp
->peer
= sysctl_ipfrag_max_dist
? inet_getpeer(iph
->saddr
, 1) : NULL
;
302 /* Initialize a timer for this entry. */
303 init_timer(&qp
->q
.timer
);
304 qp
->q
.timer
.data
= (unsigned long) qp
; /* pointer to queue */
305 qp
->q
.timer
.function
= ip_expire
; /* expire function */
306 spin_lock_init(&qp
->q
.lock
);
307 atomic_set(&qp
->q
.refcnt
, 1);
309 return ip_frag_intern(qp
);
312 LIMIT_NETDEBUG(KERN_ERR
"ip_frag_create: no memory left !\n");
316 /* Find the correct entry in the "incomplete datagrams" queue for
317 * this IP datagram, and create new one, if nothing is found.
319 static inline struct ipq
*ip_find(struct iphdr
*iph
, u32 user
)
322 __be32 saddr
= iph
->saddr
;
323 __be32 daddr
= iph
->daddr
;
324 __u8 protocol
= iph
->protocol
;
327 struct hlist_node
*n
;
329 read_lock(&ip4_frags
.lock
);
330 hash
= ipqhashfn(id
, saddr
, daddr
, protocol
);
331 hlist_for_each_entry(qp
, n
, &ip4_frags
.hash
[hash
], q
.list
) {
333 qp
->saddr
== saddr
&&
334 qp
->daddr
== daddr
&&
335 qp
->protocol
== protocol
&&
337 atomic_inc(&qp
->q
.refcnt
);
338 read_unlock(&ip4_frags
.lock
);
342 read_unlock(&ip4_frags
.lock
);
344 return ip_frag_create(iph
, user
);
347 /* Is the fragment too far ahead to be part of ipq? */
348 static inline int ip_frag_too_far(struct ipq
*qp
)
350 struct inet_peer
*peer
= qp
->peer
;
351 unsigned int max
= sysctl_ipfrag_max_dist
;
352 unsigned int start
, end
;
360 end
= atomic_inc_return(&peer
->rid
);
363 rc
= qp
->q
.fragments
&& (end
- start
) > max
;
366 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
372 static int ip_frag_reinit(struct ipq
*qp
)
376 if (!mod_timer(&qp
->q
.timer
, jiffies
+ ip4_frags_ctl
.timeout
)) {
377 atomic_inc(&qp
->q
.refcnt
);
381 fp
= qp
->q
.fragments
;
383 struct sk_buff
*xp
= fp
->next
;
384 frag_kfree_skb(fp
, NULL
);
391 qp
->q
.fragments
= NULL
;
397 /* Add new segment to existing queue. */
398 static int ip_frag_queue(struct ipq
*qp
, struct sk_buff
*skb
)
400 struct sk_buff
*prev
, *next
;
401 struct net_device
*dev
;
406 if (qp
->q
.last_in
& COMPLETE
)
409 if (!(IPCB(skb
)->flags
& IPSKB_FRAG_COMPLETE
) &&
410 unlikely(ip_frag_too_far(qp
)) &&
411 unlikely(err
= ip_frag_reinit(qp
))) {
416 offset
= ntohs(ip_hdr(skb
)->frag_off
);
417 flags
= offset
& ~IP_OFFSET
;
419 offset
<<= 3; /* offset is in 8-byte chunks */
420 ihl
= ip_hdrlen(skb
);
422 /* Determine the position of this fragment. */
423 end
= offset
+ skb
->len
- ihl
;
426 /* Is this the final fragment? */
427 if ((flags
& IP_MF
) == 0) {
428 /* If we already have some bits beyond end
429 * or have different end, the segment is corrrupted.
431 if (end
< qp
->q
.len
||
432 ((qp
->q
.last_in
& LAST_IN
) && end
!= qp
->q
.len
))
434 qp
->q
.last_in
|= LAST_IN
;
439 if (skb
->ip_summed
!= CHECKSUM_UNNECESSARY
)
440 skb
->ip_summed
= CHECKSUM_NONE
;
442 if (end
> qp
->q
.len
) {
443 /* Some bits beyond end -> corruption. */
444 if (qp
->q
.last_in
& LAST_IN
)
453 if (pskb_pull(skb
, ihl
) == NULL
)
456 err
= pskb_trim_rcsum(skb
, end
- offset
);
460 /* Find out which fragments are in front and at the back of us
461 * in the chain of fragments so far. We must know where to put
462 * this fragment, right?
465 for (next
= qp
->q
.fragments
; next
!= NULL
; next
= next
->next
) {
466 if (FRAG_CB(next
)->offset
>= offset
)
471 /* We found where to put this one. Check for overlap with
472 * preceding fragment, and, if needed, align things so that
473 * any overlaps are eliminated.
476 int i
= (FRAG_CB(prev
)->offset
+ prev
->len
) - offset
;
484 if (!pskb_pull(skb
, i
))
486 if (skb
->ip_summed
!= CHECKSUM_UNNECESSARY
)
487 skb
->ip_summed
= CHECKSUM_NONE
;
493 while (next
&& FRAG_CB(next
)->offset
< end
) {
494 int i
= end
- FRAG_CB(next
)->offset
; /* overlap is 'i' bytes */
497 /* Eat head of the next overlapped fragment
498 * and leave the loop. The next ones cannot overlap.
500 if (!pskb_pull(next
, i
))
502 FRAG_CB(next
)->offset
+= i
;
504 if (next
->ip_summed
!= CHECKSUM_UNNECESSARY
)
505 next
->ip_summed
= CHECKSUM_NONE
;
508 struct sk_buff
*free_it
= next
;
510 /* Old fragment is completely overridden with
518 qp
->q
.fragments
= next
;
520 qp
->q
.meat
-= free_it
->len
;
521 frag_kfree_skb(free_it
, NULL
);
525 FRAG_CB(skb
)->offset
= offset
;
527 /* Insert this fragment in the chain of fragments. */
532 qp
->q
.fragments
= skb
;
536 qp
->iif
= dev
->ifindex
;
539 qp
->q
.stamp
= skb
->tstamp
;
540 qp
->q
.meat
+= skb
->len
;
541 atomic_add(skb
->truesize
, &ip4_frags
.mem
);
543 qp
->q
.last_in
|= FIRST_IN
;
545 if (qp
->q
.last_in
== (FIRST_IN
| LAST_IN
) && qp
->q
.meat
== qp
->q
.len
)
546 return ip_frag_reasm(qp
, prev
, dev
);
548 write_lock(&ip4_frags
.lock
);
549 list_move_tail(&qp
->q
.lru_list
, &ip4_frags
.lru_list
);
550 write_unlock(&ip4_frags
.lock
);
559 /* Build a new IP datagram from all its fragments. */
561 static int ip_frag_reasm(struct ipq
*qp
, struct sk_buff
*prev
,
562 struct net_device
*dev
)
565 struct sk_buff
*fp
, *head
= qp
->q
.fragments
;
572 /* Make the one we just received the head. */
575 fp
= skb_clone(head
, GFP_ATOMIC
);
580 fp
->next
= head
->next
;
583 skb_morph(head
, qp
->q
.fragments
);
584 head
->next
= qp
->q
.fragments
->next
;
586 kfree_skb(qp
->q
.fragments
);
587 qp
->q
.fragments
= head
;
590 BUG_TRAP(head
!= NULL
);
591 BUG_TRAP(FRAG_CB(head
)->offset
== 0);
593 /* Allocate a new buffer for the datagram. */
594 ihlen
= ip_hdrlen(head
);
595 len
= ihlen
+ qp
->q
.len
;
601 /* Head of list must not be cloned. */
603 if (skb_cloned(head
) && pskb_expand_head(head
, 0, 0, GFP_ATOMIC
))
606 /* If the first fragment is fragmented itself, we split
607 * it to two chunks: the first with data and paged part
608 * and the second, holding only fragments. */
609 if (skb_shinfo(head
)->frag_list
) {
610 struct sk_buff
*clone
;
613 if ((clone
= alloc_skb(0, GFP_ATOMIC
)) == NULL
)
615 clone
->next
= head
->next
;
617 skb_shinfo(clone
)->frag_list
= skb_shinfo(head
)->frag_list
;
618 skb_shinfo(head
)->frag_list
= NULL
;
619 for (i
=0; i
<skb_shinfo(head
)->nr_frags
; i
++)
620 plen
+= skb_shinfo(head
)->frags
[i
].size
;
621 clone
->len
= clone
->data_len
= head
->data_len
- plen
;
622 head
->data_len
-= clone
->len
;
623 head
->len
-= clone
->len
;
625 clone
->ip_summed
= head
->ip_summed
;
626 atomic_add(clone
->truesize
, &ip4_frags
.mem
);
629 skb_shinfo(head
)->frag_list
= head
->next
;
630 skb_push(head
, head
->data
- skb_network_header(head
));
631 atomic_sub(head
->truesize
, &ip4_frags
.mem
);
633 for (fp
=head
->next
; fp
; fp
= fp
->next
) {
634 head
->data_len
+= fp
->len
;
635 head
->len
+= fp
->len
;
636 if (head
->ip_summed
!= fp
->ip_summed
)
637 head
->ip_summed
= CHECKSUM_NONE
;
638 else if (head
->ip_summed
== CHECKSUM_COMPLETE
)
639 head
->csum
= csum_add(head
->csum
, fp
->csum
);
640 head
->truesize
+= fp
->truesize
;
641 atomic_sub(fp
->truesize
, &ip4_frags
.mem
);
646 head
->tstamp
= qp
->q
.stamp
;
650 iph
->tot_len
= htons(len
);
651 IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS
);
652 qp
->q
.fragments
= NULL
;
656 LIMIT_NETDEBUG(KERN_ERR
"IP: queue_glue: no memory for gluing "
662 "Oversized IP packet from %d.%d.%d.%d.\n",
665 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
669 /* Process an incoming IP datagram fragment. */
670 int ip_defrag(struct sk_buff
*skb
, u32 user
)
674 IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS
);
676 /* Start by cleaning up the memory. */
677 if (atomic_read(&ip4_frags
.mem
) > ip4_frags_ctl
.high_thresh
)
680 /* Lookup (or create) queue header */
681 if ((qp
= ip_find(ip_hdr(skb
), user
)) != NULL
) {
684 spin_lock(&qp
->q
.lock
);
686 ret
= ip_frag_queue(qp
, skb
);
688 spin_unlock(&qp
->q
.lock
);
693 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
698 void __init
ipfrag_init(void)
700 ip4_frags
.ctl
= &ip4_frags_ctl
;
701 ip4_frags
.hashfn
= ip4_hashfn
;
702 ip4_frags
.destructor
= ip4_frag_free
;
703 ip4_frags
.skb_free
= NULL
;
704 ip4_frags
.qsize
= sizeof(struct ipq
);
705 inet_frags_init(&ip4_frags
);
708 EXPORT_SYMBOL(ip_defrag
);