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 __inline__
void __ipq_unlink(struct ipq
*qp
)
113 hlist_del(&qp
->q
.list
);
114 list_del(&qp
->q
.lru_list
);
118 static __inline__
void ipq_unlink(struct ipq
*ipq
)
120 write_lock(&ip4_frags
.lock
);
122 write_unlock(&ip4_frags
.lock
);
125 static unsigned int ipqhashfn(__be16 id
, __be32 saddr
, __be32 daddr
, u8 prot
)
127 return jhash_3words((__force u32
)id
<< 16 | prot
,
128 (__force u32
)saddr
, (__force u32
)daddr
,
129 ip4_frags
.rnd
) & (INETFRAGS_HASHSZ
- 1);
132 static void ipfrag_secret_rebuild(unsigned long dummy
)
134 unsigned long now
= jiffies
;
137 write_lock(&ip4_frags
.lock
);
138 get_random_bytes(&ip4_frags
.rnd
, sizeof(u32
));
139 for (i
= 0; i
< INETFRAGS_HASHSZ
; i
++) {
141 struct hlist_node
*p
, *n
;
143 hlist_for_each_entry_safe(q
, p
, n
, &ip4_frags
.hash
[i
], q
.list
) {
144 unsigned int hval
= ipqhashfn(q
->id
, q
->saddr
,
145 q
->daddr
, q
->protocol
);
148 hlist_del(&q
->q
.list
);
150 /* Relink to new hash chain. */
151 hlist_add_head(&q
->q
.list
, &ip4_frags
.hash
[hval
]);
155 write_unlock(&ip4_frags
.lock
);
157 mod_timer(&ip4_frags
.secret_timer
, now
+ ip4_frags_ctl
.secret_interval
);
160 /* Memory Tracking Functions. */
161 static __inline__
void frag_kfree_skb(struct sk_buff
*skb
, int *work
)
164 *work
-= skb
->truesize
;
165 atomic_sub(skb
->truesize
, &ip4_frags
.mem
);
169 static __inline__
void frag_free_queue(struct ipq
*qp
, int *work
)
172 *work
-= sizeof(struct ipq
);
173 atomic_sub(sizeof(struct ipq
), &ip4_frags
.mem
);
177 static __inline__
struct ipq
*frag_alloc_queue(void)
179 struct ipq
*qp
= kmalloc(sizeof(struct ipq
), GFP_ATOMIC
);
183 atomic_add(sizeof(struct ipq
), &ip4_frags
.mem
);
188 /* Destruction primitives. */
190 /* Complete destruction of ipq. */
191 static void ip_frag_destroy(struct ipq
*qp
, int *work
)
195 BUG_TRAP(qp
->q
.last_in
&COMPLETE
);
196 BUG_TRAP(del_timer(&qp
->q
.timer
) == 0);
199 inet_putpeer(qp
->peer
);
201 /* Release all fragment data. */
202 fp
= qp
->q
.fragments
;
204 struct sk_buff
*xp
= fp
->next
;
206 frag_kfree_skb(fp
, work
);
210 /* Finally, release the queue descriptor itself. */
211 frag_free_queue(qp
, work
);
214 static __inline__
void ipq_put(struct ipq
*ipq
, int *work
)
216 if (atomic_dec_and_test(&ipq
->q
.refcnt
))
217 ip_frag_destroy(ipq
, work
);
220 /* Kill ipq entry. It is not destroyed immediately,
221 * because caller (and someone more) holds reference count.
223 static void ipq_kill(struct ipq
*ipq
)
225 if (del_timer(&ipq
->q
.timer
))
226 atomic_dec(&ipq
->q
.refcnt
);
228 if (!(ipq
->q
.last_in
& COMPLETE
)) {
230 atomic_dec(&ipq
->q
.refcnt
);
231 ipq
->q
.last_in
|= COMPLETE
;
235 /* Memory limiting on fragments. Evictor trashes the oldest
236 * fragment queue until we are back under the threshold.
238 static void ip_evictor(void)
241 struct list_head
*tmp
;
244 work
= atomic_read(&ip4_frags
.mem
) - ip4_frags_ctl
.low_thresh
;
249 read_lock(&ip4_frags
.lock
);
250 if (list_empty(&ip4_frags
.lru_list
)) {
251 read_unlock(&ip4_frags
.lock
);
254 tmp
= ip4_frags
.lru_list
.next
;
255 qp
= list_entry(tmp
, struct ipq
, q
.lru_list
);
256 atomic_inc(&qp
->q
.refcnt
);
257 read_unlock(&ip4_frags
.lock
);
259 spin_lock(&qp
->q
.lock
);
260 if (!(qp
->q
.last_in
&COMPLETE
))
262 spin_unlock(&qp
->q
.lock
);
265 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
270 * Oops, a fragment queue timed out. Kill it and send an ICMP reply.
272 static void ip_expire(unsigned long arg
)
274 struct ipq
*qp
= (struct ipq
*) arg
;
276 spin_lock(&qp
->q
.lock
);
278 if (qp
->q
.last_in
& COMPLETE
)
283 IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT
);
284 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
286 if ((qp
->q
.last_in
&FIRST_IN
) && qp
->q
.fragments
!= NULL
) {
287 struct sk_buff
*head
= qp
->q
.fragments
;
288 /* Send an ICMP "Fragment Reassembly Timeout" message. */
289 if ((head
->dev
= dev_get_by_index(&init_net
, qp
->iif
)) != NULL
) {
290 icmp_send(head
, ICMP_TIME_EXCEEDED
, ICMP_EXC_FRAGTIME
, 0);
295 spin_unlock(&qp
->q
.lock
);
299 /* Creation primitives. */
301 static struct ipq
*ip_frag_intern(struct ipq
*qp_in
)
305 struct hlist_node
*n
;
309 write_lock(&ip4_frags
.lock
);
310 hash
= ipqhashfn(qp_in
->id
, qp_in
->saddr
, qp_in
->daddr
,
313 /* With SMP race we have to recheck hash table, because
314 * such entry could be created on other cpu, while we
315 * promoted read lock to write lock.
317 hlist_for_each_entry(qp
, n
, &ip4_frags
.hash
[hash
], q
.list
) {
318 if (qp
->id
== qp_in
->id
&&
319 qp
->saddr
== qp_in
->saddr
&&
320 qp
->daddr
== qp_in
->daddr
&&
321 qp
->protocol
== qp_in
->protocol
&&
322 qp
->user
== qp_in
->user
) {
323 atomic_inc(&qp
->q
.refcnt
);
324 write_unlock(&ip4_frags
.lock
);
325 qp_in
->q
.last_in
|= COMPLETE
;
326 ipq_put(qp_in
, NULL
);
333 if (!mod_timer(&qp
->q
.timer
, jiffies
+ ip4_frags_ctl
.timeout
))
334 atomic_inc(&qp
->q
.refcnt
);
336 atomic_inc(&qp
->q
.refcnt
);
337 hlist_add_head(&qp
->q
.list
, &ip4_frags
.hash
[hash
]);
338 INIT_LIST_HEAD(&qp
->q
.lru_list
);
339 list_add_tail(&qp
->q
.lru_list
, &ip4_frags
.lru_list
);
341 write_unlock(&ip4_frags
.lock
);
345 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
346 static struct ipq
*ip_frag_create(struct iphdr
*iph
, u32 user
)
350 if ((qp
= frag_alloc_queue()) == NULL
)
353 qp
->protocol
= iph
->protocol
;
356 qp
->saddr
= iph
->saddr
;
357 qp
->daddr
= iph
->daddr
;
361 qp
->q
.fragments
= NULL
;
363 qp
->peer
= sysctl_ipfrag_max_dist
? inet_getpeer(iph
->saddr
, 1) : NULL
;
365 /* Initialize a timer for this entry. */
366 init_timer(&qp
->q
.timer
);
367 qp
->q
.timer
.data
= (unsigned long) qp
; /* pointer to queue */
368 qp
->q
.timer
.function
= ip_expire
; /* expire function */
369 spin_lock_init(&qp
->q
.lock
);
370 atomic_set(&qp
->q
.refcnt
, 1);
372 return ip_frag_intern(qp
);
375 LIMIT_NETDEBUG(KERN_ERR
"ip_frag_create: no memory left !\n");
379 /* Find the correct entry in the "incomplete datagrams" queue for
380 * this IP datagram, and create new one, if nothing is found.
382 static inline struct ipq
*ip_find(struct iphdr
*iph
, u32 user
)
385 __be32 saddr
= iph
->saddr
;
386 __be32 daddr
= iph
->daddr
;
387 __u8 protocol
= iph
->protocol
;
390 struct hlist_node
*n
;
392 read_lock(&ip4_frags
.lock
);
393 hash
= ipqhashfn(id
, saddr
, daddr
, protocol
);
394 hlist_for_each_entry(qp
, n
, &ip4_frags
.hash
[hash
], q
.list
) {
396 qp
->saddr
== saddr
&&
397 qp
->daddr
== daddr
&&
398 qp
->protocol
== protocol
&&
400 atomic_inc(&qp
->q
.refcnt
);
401 read_unlock(&ip4_frags
.lock
);
405 read_unlock(&ip4_frags
.lock
);
407 return ip_frag_create(iph
, user
);
410 /* Is the fragment too far ahead to be part of ipq? */
411 static inline int ip_frag_too_far(struct ipq
*qp
)
413 struct inet_peer
*peer
= qp
->peer
;
414 unsigned int max
= sysctl_ipfrag_max_dist
;
415 unsigned int start
, end
;
423 end
= atomic_inc_return(&peer
->rid
);
426 rc
= qp
->q
.fragments
&& (end
- start
) > max
;
429 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
435 static int ip_frag_reinit(struct ipq
*qp
)
439 if (!mod_timer(&qp
->q
.timer
, jiffies
+ ip4_frags_ctl
.timeout
)) {
440 atomic_inc(&qp
->q
.refcnt
);
444 fp
= qp
->q
.fragments
;
446 struct sk_buff
*xp
= fp
->next
;
447 frag_kfree_skb(fp
, NULL
);
454 qp
->q
.fragments
= NULL
;
460 /* Add new segment to existing queue. */
461 static int ip_frag_queue(struct ipq
*qp
, struct sk_buff
*skb
)
463 struct sk_buff
*prev
, *next
;
464 struct net_device
*dev
;
469 if (qp
->q
.last_in
& COMPLETE
)
472 if (!(IPCB(skb
)->flags
& IPSKB_FRAG_COMPLETE
) &&
473 unlikely(ip_frag_too_far(qp
)) &&
474 unlikely(err
= ip_frag_reinit(qp
))) {
479 offset
= ntohs(ip_hdr(skb
)->frag_off
);
480 flags
= offset
& ~IP_OFFSET
;
482 offset
<<= 3; /* offset is in 8-byte chunks */
483 ihl
= ip_hdrlen(skb
);
485 /* Determine the position of this fragment. */
486 end
= offset
+ skb
->len
- ihl
;
489 /* Is this the final fragment? */
490 if ((flags
& IP_MF
) == 0) {
491 /* If we already have some bits beyond end
492 * or have different end, the segment is corrrupted.
494 if (end
< qp
->q
.len
||
495 ((qp
->q
.last_in
& LAST_IN
) && end
!= qp
->q
.len
))
497 qp
->q
.last_in
|= LAST_IN
;
502 if (skb
->ip_summed
!= CHECKSUM_UNNECESSARY
)
503 skb
->ip_summed
= CHECKSUM_NONE
;
505 if (end
> qp
->q
.len
) {
506 /* Some bits beyond end -> corruption. */
507 if (qp
->q
.last_in
& LAST_IN
)
516 if (pskb_pull(skb
, ihl
) == NULL
)
519 err
= pskb_trim_rcsum(skb
, end
- offset
);
523 /* Find out which fragments are in front and at the back of us
524 * in the chain of fragments so far. We must know where to put
525 * this fragment, right?
528 for (next
= qp
->q
.fragments
; next
!= NULL
; next
= next
->next
) {
529 if (FRAG_CB(next
)->offset
>= offset
)
534 /* We found where to put this one. Check for overlap with
535 * preceding fragment, and, if needed, align things so that
536 * any overlaps are eliminated.
539 int i
= (FRAG_CB(prev
)->offset
+ prev
->len
) - offset
;
547 if (!pskb_pull(skb
, i
))
549 if (skb
->ip_summed
!= CHECKSUM_UNNECESSARY
)
550 skb
->ip_summed
= CHECKSUM_NONE
;
556 while (next
&& FRAG_CB(next
)->offset
< end
) {
557 int i
= end
- FRAG_CB(next
)->offset
; /* overlap is 'i' bytes */
560 /* Eat head of the next overlapped fragment
561 * and leave the loop. The next ones cannot overlap.
563 if (!pskb_pull(next
, i
))
565 FRAG_CB(next
)->offset
+= i
;
567 if (next
->ip_summed
!= CHECKSUM_UNNECESSARY
)
568 next
->ip_summed
= CHECKSUM_NONE
;
571 struct sk_buff
*free_it
= next
;
573 /* Old fragment is completely overridden with
581 qp
->q
.fragments
= next
;
583 qp
->q
.meat
-= free_it
->len
;
584 frag_kfree_skb(free_it
, NULL
);
588 FRAG_CB(skb
)->offset
= offset
;
590 /* Insert this fragment in the chain of fragments. */
595 qp
->q
.fragments
= skb
;
599 qp
->iif
= dev
->ifindex
;
602 qp
->q
.stamp
= skb
->tstamp
;
603 qp
->q
.meat
+= skb
->len
;
604 atomic_add(skb
->truesize
, &ip4_frags
.mem
);
606 qp
->q
.last_in
|= FIRST_IN
;
608 if (qp
->q
.last_in
== (FIRST_IN
| LAST_IN
) && qp
->q
.meat
== qp
->q
.len
)
609 return ip_frag_reasm(qp
, prev
, dev
);
611 write_lock(&ip4_frags
.lock
);
612 list_move_tail(&qp
->q
.lru_list
, &ip4_frags
.lru_list
);
613 write_unlock(&ip4_frags
.lock
);
622 /* Build a new IP datagram from all its fragments. */
624 static int ip_frag_reasm(struct ipq
*qp
, struct sk_buff
*prev
,
625 struct net_device
*dev
)
628 struct sk_buff
*fp
, *head
= qp
->q
.fragments
;
635 /* Make the one we just received the head. */
638 fp
= skb_clone(head
, GFP_ATOMIC
);
643 fp
->next
= head
->next
;
646 skb_morph(head
, qp
->q
.fragments
);
647 head
->next
= qp
->q
.fragments
->next
;
649 kfree_skb(qp
->q
.fragments
);
650 qp
->q
.fragments
= head
;
653 BUG_TRAP(head
!= NULL
);
654 BUG_TRAP(FRAG_CB(head
)->offset
== 0);
656 /* Allocate a new buffer for the datagram. */
657 ihlen
= ip_hdrlen(head
);
658 len
= ihlen
+ qp
->q
.len
;
664 /* Head of list must not be cloned. */
666 if (skb_cloned(head
) && pskb_expand_head(head
, 0, 0, GFP_ATOMIC
))
669 /* If the first fragment is fragmented itself, we split
670 * it to two chunks: the first with data and paged part
671 * and the second, holding only fragments. */
672 if (skb_shinfo(head
)->frag_list
) {
673 struct sk_buff
*clone
;
676 if ((clone
= alloc_skb(0, GFP_ATOMIC
)) == NULL
)
678 clone
->next
= head
->next
;
680 skb_shinfo(clone
)->frag_list
= skb_shinfo(head
)->frag_list
;
681 skb_shinfo(head
)->frag_list
= NULL
;
682 for (i
=0; i
<skb_shinfo(head
)->nr_frags
; i
++)
683 plen
+= skb_shinfo(head
)->frags
[i
].size
;
684 clone
->len
= clone
->data_len
= head
->data_len
- plen
;
685 head
->data_len
-= clone
->len
;
686 head
->len
-= clone
->len
;
688 clone
->ip_summed
= head
->ip_summed
;
689 atomic_add(clone
->truesize
, &ip4_frags
.mem
);
692 skb_shinfo(head
)->frag_list
= head
->next
;
693 skb_push(head
, head
->data
- skb_network_header(head
));
694 atomic_sub(head
->truesize
, &ip4_frags
.mem
);
696 for (fp
=head
->next
; fp
; fp
= fp
->next
) {
697 head
->data_len
+= fp
->len
;
698 head
->len
+= fp
->len
;
699 if (head
->ip_summed
!= fp
->ip_summed
)
700 head
->ip_summed
= CHECKSUM_NONE
;
701 else if (head
->ip_summed
== CHECKSUM_COMPLETE
)
702 head
->csum
= csum_add(head
->csum
, fp
->csum
);
703 head
->truesize
+= fp
->truesize
;
704 atomic_sub(fp
->truesize
, &ip4_frags
.mem
);
709 head
->tstamp
= qp
->q
.stamp
;
713 iph
->tot_len
= htons(len
);
714 IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS
);
715 qp
->q
.fragments
= NULL
;
719 LIMIT_NETDEBUG(KERN_ERR
"IP: queue_glue: no memory for gluing "
725 "Oversized IP packet from %d.%d.%d.%d.\n",
728 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
732 /* Process an incoming IP datagram fragment. */
733 int ip_defrag(struct sk_buff
*skb
, u32 user
)
737 IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS
);
739 /* Start by cleaning up the memory. */
740 if (atomic_read(&ip4_frags
.mem
) > ip4_frags_ctl
.high_thresh
)
743 /* Lookup (or create) queue header */
744 if ((qp
= ip_find(ip_hdr(skb
), user
)) != NULL
) {
747 spin_lock(&qp
->q
.lock
);
749 ret
= ip_frag_queue(qp
, skb
);
751 spin_unlock(&qp
->q
.lock
);
756 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS
);
761 void __init
ipfrag_init(void)
763 init_timer(&ip4_frags
.secret_timer
);
764 ip4_frags
.secret_timer
.function
= ipfrag_secret_rebuild
;
765 ip4_frags
.secret_timer
.expires
= jiffies
+ ip4_frags_ctl
.secret_interval
;
766 add_timer(&ip4_frags
.secret_timer
);
768 ip4_frags
.ctl
= &ip4_frags_ctl
;
769 inet_frags_init(&ip4_frags
);
772 EXPORT_SYMBOL(ip_defrag
);