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
= kzalloc(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
)
160 inet_frag_put(&ipq
->q
, &ip4_frags
);
163 /* Kill ipq entry. It is not destroyed immediately,
164 * because caller (and someone more) holds reference count.
166 static void ipq_kill(struct ipq
*ipq
)
168 inet_frag_kill(&ipq
->q
, &ip4_frags
);
171 /* Memory limiting on fragments. Evictor trashes the oldest
172 * fragment queue until we are back under the threshold.
174 static void ip_evictor(void)
178 evicted
= inet_frag_evictor(&ip4_frags
);
180 IP_ADD_STATS_BH(IPSTATS_MIB_REASMFAILS
, evicted
);
184 * Oops, a fragment queue timed out. Kill it and send an ICMP reply.
186 static void ip_expire(unsigned long arg
)
188 struct ipq
*qp
= (struct ipq
*) arg
;
190 spin_lock(&qp
->q
.lock
);
192 if (qp
->q
.last_in
& COMPLETE
)
197 IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT
);
198 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
200 if ((qp
->q
.last_in
&FIRST_IN
) && qp
->q
.fragments
!= NULL
) {
201 struct sk_buff
*head
= qp
->q
.fragments
;
202 /* Send an ICMP "Fragment Reassembly Timeout" message. */
203 if ((head
->dev
= dev_get_by_index(&init_net
, qp
->iif
)) != NULL
) {
204 icmp_send(head
, ICMP_TIME_EXCEEDED
, ICMP_EXC_FRAGTIME
, 0);
209 spin_unlock(&qp
->q
.lock
);
213 /* Creation primitives. */
215 static struct ipq
*ip_frag_intern(struct ipq
*qp_in
, unsigned int hash
)
219 struct hlist_node
*n
;
222 write_lock(&ip4_frags
.lock
);
224 /* With SMP race we have to recheck hash table, because
225 * such entry could be created on other cpu, while we
226 * promoted read lock to write lock.
228 hlist_for_each_entry(qp
, n
, &ip4_frags
.hash
[hash
], q
.list
) {
229 if (qp
->id
== qp_in
->id
&&
230 qp
->saddr
== qp_in
->saddr
&&
231 qp
->daddr
== qp_in
->daddr
&&
232 qp
->protocol
== qp_in
->protocol
&&
233 qp
->user
== qp_in
->user
) {
234 atomic_inc(&qp
->q
.refcnt
);
235 write_unlock(&ip4_frags
.lock
);
236 qp_in
->q
.last_in
|= COMPLETE
;
244 if (!mod_timer(&qp
->q
.timer
, jiffies
+ ip4_frags_ctl
.timeout
))
245 atomic_inc(&qp
->q
.refcnt
);
247 atomic_inc(&qp
->q
.refcnt
);
248 hlist_add_head(&qp
->q
.list
, &ip4_frags
.hash
[hash
]);
249 INIT_LIST_HEAD(&qp
->q
.lru_list
);
250 list_add_tail(&qp
->q
.lru_list
, &ip4_frags
.lru_list
);
252 write_unlock(&ip4_frags
.lock
);
256 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
257 static struct ipq
*ip_frag_create(struct iphdr
*iph
, u32 user
, unsigned int h
)
261 if ((qp
= frag_alloc_queue()) == NULL
)
264 qp
->protocol
= iph
->protocol
;
266 qp
->saddr
= iph
->saddr
;
267 qp
->daddr
= iph
->daddr
;
269 qp
->peer
= sysctl_ipfrag_max_dist
? inet_getpeer(iph
->saddr
, 1) : NULL
;
271 /* Initialize a timer for this entry. */
272 init_timer(&qp
->q
.timer
);
273 qp
->q
.timer
.data
= (unsigned long) qp
; /* pointer to queue */
274 qp
->q
.timer
.function
= ip_expire
; /* expire function */
275 spin_lock_init(&qp
->q
.lock
);
276 atomic_set(&qp
->q
.refcnt
, 1);
278 return ip_frag_intern(qp
, h
);
281 LIMIT_NETDEBUG(KERN_ERR
"ip_frag_create: no memory left !\n");
285 /* Find the correct entry in the "incomplete datagrams" queue for
286 * this IP datagram, and create new one, if nothing is found.
288 static inline struct ipq
*ip_find(struct iphdr
*iph
, u32 user
)
291 __be32 saddr
= iph
->saddr
;
292 __be32 daddr
= iph
->daddr
;
293 __u8 protocol
= iph
->protocol
;
296 struct hlist_node
*n
;
298 read_lock(&ip4_frags
.lock
);
299 hash
= ipqhashfn(id
, saddr
, daddr
, protocol
);
300 hlist_for_each_entry(qp
, n
, &ip4_frags
.hash
[hash
], q
.list
) {
302 qp
->saddr
== saddr
&&
303 qp
->daddr
== daddr
&&
304 qp
->protocol
== protocol
&&
306 atomic_inc(&qp
->q
.refcnt
);
307 read_unlock(&ip4_frags
.lock
);
311 read_unlock(&ip4_frags
.lock
);
313 return ip_frag_create(iph
, user
, hash
);
316 /* Is the fragment too far ahead to be part of ipq? */
317 static inline int ip_frag_too_far(struct ipq
*qp
)
319 struct inet_peer
*peer
= qp
->peer
;
320 unsigned int max
= sysctl_ipfrag_max_dist
;
321 unsigned int start
, end
;
329 end
= atomic_inc_return(&peer
->rid
);
332 rc
= qp
->q
.fragments
&& (end
- start
) > max
;
335 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
341 static int ip_frag_reinit(struct ipq
*qp
)
345 if (!mod_timer(&qp
->q
.timer
, jiffies
+ ip4_frags_ctl
.timeout
)) {
346 atomic_inc(&qp
->q
.refcnt
);
350 fp
= qp
->q
.fragments
;
352 struct sk_buff
*xp
= fp
->next
;
353 frag_kfree_skb(fp
, NULL
);
360 qp
->q
.fragments
= NULL
;
366 /* Add new segment to existing queue. */
367 static int ip_frag_queue(struct ipq
*qp
, struct sk_buff
*skb
)
369 struct sk_buff
*prev
, *next
;
370 struct net_device
*dev
;
375 if (qp
->q
.last_in
& COMPLETE
)
378 if (!(IPCB(skb
)->flags
& IPSKB_FRAG_COMPLETE
) &&
379 unlikely(ip_frag_too_far(qp
)) &&
380 unlikely(err
= ip_frag_reinit(qp
))) {
385 offset
= ntohs(ip_hdr(skb
)->frag_off
);
386 flags
= offset
& ~IP_OFFSET
;
388 offset
<<= 3; /* offset is in 8-byte chunks */
389 ihl
= ip_hdrlen(skb
);
391 /* Determine the position of this fragment. */
392 end
= offset
+ skb
->len
- ihl
;
395 /* Is this the final fragment? */
396 if ((flags
& IP_MF
) == 0) {
397 /* If we already have some bits beyond end
398 * or have different end, the segment is corrrupted.
400 if (end
< qp
->q
.len
||
401 ((qp
->q
.last_in
& LAST_IN
) && end
!= qp
->q
.len
))
403 qp
->q
.last_in
|= LAST_IN
;
408 if (skb
->ip_summed
!= CHECKSUM_UNNECESSARY
)
409 skb
->ip_summed
= CHECKSUM_NONE
;
411 if (end
> qp
->q
.len
) {
412 /* Some bits beyond end -> corruption. */
413 if (qp
->q
.last_in
& LAST_IN
)
422 if (pskb_pull(skb
, ihl
) == NULL
)
425 err
= pskb_trim_rcsum(skb
, end
- offset
);
429 /* Find out which fragments are in front and at the back of us
430 * in the chain of fragments so far. We must know where to put
431 * this fragment, right?
434 for (next
= qp
->q
.fragments
; next
!= NULL
; next
= next
->next
) {
435 if (FRAG_CB(next
)->offset
>= offset
)
440 /* We found where to put this one. Check for overlap with
441 * preceding fragment, and, if needed, align things so that
442 * any overlaps are eliminated.
445 int i
= (FRAG_CB(prev
)->offset
+ prev
->len
) - offset
;
453 if (!pskb_pull(skb
, i
))
455 if (skb
->ip_summed
!= CHECKSUM_UNNECESSARY
)
456 skb
->ip_summed
= CHECKSUM_NONE
;
462 while (next
&& FRAG_CB(next
)->offset
< end
) {
463 int i
= end
- FRAG_CB(next
)->offset
; /* overlap is 'i' bytes */
466 /* Eat head of the next overlapped fragment
467 * and leave the loop. The next ones cannot overlap.
469 if (!pskb_pull(next
, i
))
471 FRAG_CB(next
)->offset
+= i
;
473 if (next
->ip_summed
!= CHECKSUM_UNNECESSARY
)
474 next
->ip_summed
= CHECKSUM_NONE
;
477 struct sk_buff
*free_it
= next
;
479 /* Old fragment is completely overridden with
487 qp
->q
.fragments
= next
;
489 qp
->q
.meat
-= free_it
->len
;
490 frag_kfree_skb(free_it
, NULL
);
494 FRAG_CB(skb
)->offset
= offset
;
496 /* Insert this fragment in the chain of fragments. */
501 qp
->q
.fragments
= skb
;
505 qp
->iif
= dev
->ifindex
;
508 qp
->q
.stamp
= skb
->tstamp
;
509 qp
->q
.meat
+= skb
->len
;
510 atomic_add(skb
->truesize
, &ip4_frags
.mem
);
512 qp
->q
.last_in
|= FIRST_IN
;
514 if (qp
->q
.last_in
== (FIRST_IN
| LAST_IN
) && qp
->q
.meat
== qp
->q
.len
)
515 return ip_frag_reasm(qp
, prev
, dev
);
517 write_lock(&ip4_frags
.lock
);
518 list_move_tail(&qp
->q
.lru_list
, &ip4_frags
.lru_list
);
519 write_unlock(&ip4_frags
.lock
);
528 /* Build a new IP datagram from all its fragments. */
530 static int ip_frag_reasm(struct ipq
*qp
, struct sk_buff
*prev
,
531 struct net_device
*dev
)
534 struct sk_buff
*fp
, *head
= qp
->q
.fragments
;
541 /* Make the one we just received the head. */
544 fp
= skb_clone(head
, GFP_ATOMIC
);
549 fp
->next
= head
->next
;
552 skb_morph(head
, qp
->q
.fragments
);
553 head
->next
= qp
->q
.fragments
->next
;
555 kfree_skb(qp
->q
.fragments
);
556 qp
->q
.fragments
= head
;
559 BUG_TRAP(head
!= NULL
);
560 BUG_TRAP(FRAG_CB(head
)->offset
== 0);
562 /* Allocate a new buffer for the datagram. */
563 ihlen
= ip_hdrlen(head
);
564 len
= ihlen
+ qp
->q
.len
;
570 /* Head of list must not be cloned. */
572 if (skb_cloned(head
) && pskb_expand_head(head
, 0, 0, GFP_ATOMIC
))
575 /* If the first fragment is fragmented itself, we split
576 * it to two chunks: the first with data and paged part
577 * and the second, holding only fragments. */
578 if (skb_shinfo(head
)->frag_list
) {
579 struct sk_buff
*clone
;
582 if ((clone
= alloc_skb(0, GFP_ATOMIC
)) == NULL
)
584 clone
->next
= head
->next
;
586 skb_shinfo(clone
)->frag_list
= skb_shinfo(head
)->frag_list
;
587 skb_shinfo(head
)->frag_list
= NULL
;
588 for (i
=0; i
<skb_shinfo(head
)->nr_frags
; i
++)
589 plen
+= skb_shinfo(head
)->frags
[i
].size
;
590 clone
->len
= clone
->data_len
= head
->data_len
- plen
;
591 head
->data_len
-= clone
->len
;
592 head
->len
-= clone
->len
;
594 clone
->ip_summed
= head
->ip_summed
;
595 atomic_add(clone
->truesize
, &ip4_frags
.mem
);
598 skb_shinfo(head
)->frag_list
= head
->next
;
599 skb_push(head
, head
->data
- skb_network_header(head
));
600 atomic_sub(head
->truesize
, &ip4_frags
.mem
);
602 for (fp
=head
->next
; fp
; fp
= fp
->next
) {
603 head
->data_len
+= fp
->len
;
604 head
->len
+= fp
->len
;
605 if (head
->ip_summed
!= fp
->ip_summed
)
606 head
->ip_summed
= CHECKSUM_NONE
;
607 else if (head
->ip_summed
== CHECKSUM_COMPLETE
)
608 head
->csum
= csum_add(head
->csum
, fp
->csum
);
609 head
->truesize
+= fp
->truesize
;
610 atomic_sub(fp
->truesize
, &ip4_frags
.mem
);
615 head
->tstamp
= qp
->q
.stamp
;
619 iph
->tot_len
= htons(len
);
620 IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS
);
621 qp
->q
.fragments
= NULL
;
625 LIMIT_NETDEBUG(KERN_ERR
"IP: queue_glue: no memory for gluing "
631 "Oversized IP packet from %d.%d.%d.%d.\n",
634 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
638 /* Process an incoming IP datagram fragment. */
639 int ip_defrag(struct sk_buff
*skb
, u32 user
)
643 IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS
);
645 /* Start by cleaning up the memory. */
646 if (atomic_read(&ip4_frags
.mem
) > ip4_frags_ctl
.high_thresh
)
649 /* Lookup (or create) queue header */
650 if ((qp
= ip_find(ip_hdr(skb
), user
)) != NULL
) {
653 spin_lock(&qp
->q
.lock
);
655 ret
= ip_frag_queue(qp
, skb
);
657 spin_unlock(&qp
->q
.lock
);
662 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
667 void __init
ipfrag_init(void)
669 ip4_frags
.ctl
= &ip4_frags_ctl
;
670 ip4_frags
.hashfn
= ip4_hashfn
;
671 ip4_frags
.destructor
= ip4_frag_free
;
672 ip4_frags
.skb_free
= NULL
;
673 ip4_frags
.qsize
= sizeof(struct ipq
);
674 inet_frags_init(&ip4_frags
);
677 EXPORT_SYMBOL(ip_defrag
);