[NETFILTER] Fix conntrack event cache deadlock/oops
[linux-2.6] / net / ipv4 / ip_fragment.c
1 /*
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.
5  *
6  *              The IP fragmentation functionality.
7  *              
8  * Version:     $Id: ip_fragment.c,v 1.59 2002/01/12 07:54:56 davem Exp $
9  *
10  * Authors:     Fred N. van Kempen <waltje@uWalt.NL.Mugnet.ORG>
11  *              Alan Cox <Alan.Cox@linux.org>
12  *
13  * Fixes:
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.
23  */
24
25 #include <linux/config.h>
26 #include <linux/module.h>
27 #include <linux/types.h>
28 #include <linux/mm.h>
29 #include <linux/jiffies.h>
30 #include <linux/skbuff.h>
31 #include <linux/list.h>
32 #include <linux/ip.h>
33 #include <linux/icmp.h>
34 #include <linux/netdevice.h>
35 #include <linux/jhash.h>
36 #include <linux/random.h>
37 #include <net/sock.h>
38 #include <net/ip.h>
39 #include <net/icmp.h>
40 #include <net/checksum.h>
41 #include <linux/tcp.h>
42 #include <linux/udp.h>
43 #include <linux/inet.h>
44 #include <linux/netfilter_ipv4.h>
45
46 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
47  * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
48  * as well. Or notify me, at least. --ANK
49  */
50
51 /* Fragment cache limits. We will commit 256K at one time. Should we
52  * cross that limit we will prune down to 192K. This should cope with
53  * even the most extreme cases without allowing an attacker to measurably
54  * harm machine performance.
55  */
56 int sysctl_ipfrag_high_thresh = 256*1024;
57 int sysctl_ipfrag_low_thresh = 192*1024;
58
59 /* Important NOTE! Fragment queue must be destroyed before MSL expires.
60  * RFC791 is wrong proposing to prolongate timer each fragment arrival by TTL.
61  */
62 int sysctl_ipfrag_time = IP_FRAG_TIME;
63
64 struct ipfrag_skb_cb
65 {
66         struct inet_skb_parm    h;
67         int                     offset;
68 };
69
70 #define FRAG_CB(skb)    ((struct ipfrag_skb_cb*)((skb)->cb))
71
72 /* Describe an entry in the "incomplete datagrams" queue. */
73 struct ipq {
74         struct ipq      *next;          /* linked list pointers                 */
75         struct list_head lru_list;      /* lru list member                      */
76         u32             user;
77         u32             saddr;
78         u32             daddr;
79         u16             id;
80         u8              protocol;
81         u8              last_in;
82 #define COMPLETE                4
83 #define FIRST_IN                2
84 #define LAST_IN                 1
85
86         struct sk_buff  *fragments;     /* linked list of received fragments    */
87         int             len;            /* total length of original datagram    */
88         int             meat;
89         spinlock_t      lock;
90         atomic_t        refcnt;
91         struct timer_list timer;        /* when will this queue expire?         */
92         struct ipq      **pprev;
93         int             iif;
94         struct timeval  stamp;
95 };
96
97 /* Hash table. */
98
99 #define IPQ_HASHSZ      64
100
101 /* Per-bucket lock is easy to add now. */
102 static struct ipq *ipq_hash[IPQ_HASHSZ];
103 static DEFINE_RWLOCK(ipfrag_lock);
104 static u32 ipfrag_hash_rnd;
105 static LIST_HEAD(ipq_lru_list);
106 int ip_frag_nqueues = 0;
107
108 static __inline__ void __ipq_unlink(struct ipq *qp)
109 {
110         if(qp->next)
111                 qp->next->pprev = qp->pprev;
112         *qp->pprev = qp->next;
113         list_del(&qp->lru_list);
114         ip_frag_nqueues--;
115 }
116
117 static __inline__ void ipq_unlink(struct ipq *ipq)
118 {
119         write_lock(&ipfrag_lock);
120         __ipq_unlink(ipq);
121         write_unlock(&ipfrag_lock);
122 }
123
124 static unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
125 {
126         return jhash_3words((u32)id << 16 | prot, saddr, daddr,
127                             ipfrag_hash_rnd) & (IPQ_HASHSZ - 1);
128 }
129
130 static struct timer_list ipfrag_secret_timer;
131 int sysctl_ipfrag_secret_interval = 10 * 60 * HZ;
132
133 static void ipfrag_secret_rebuild(unsigned long dummy)
134 {
135         unsigned long now = jiffies;
136         int i;
137
138         write_lock(&ipfrag_lock);
139         get_random_bytes(&ipfrag_hash_rnd, sizeof(u32));
140         for (i = 0; i < IPQ_HASHSZ; i++) {
141                 struct ipq *q;
142
143                 q = ipq_hash[i];
144                 while (q) {
145                         struct ipq *next = q->next;
146                         unsigned int hval = ipqhashfn(q->id, q->saddr,
147                                                       q->daddr, q->protocol);
148
149                         if (hval != i) {
150                                 /* Unlink. */
151                                 if (q->next)
152                                         q->next->pprev = q->pprev;
153                                 *q->pprev = q->next;
154
155                                 /* Relink to new hash chain. */
156                                 if ((q->next = ipq_hash[hval]) != NULL)
157                                         q->next->pprev = &q->next;
158                                 ipq_hash[hval] = q;
159                                 q->pprev = &ipq_hash[hval];
160                         }
161
162                         q = next;
163                 }
164         }
165         write_unlock(&ipfrag_lock);
166
167         mod_timer(&ipfrag_secret_timer, now + sysctl_ipfrag_secret_interval);
168 }
169
170 atomic_t ip_frag_mem = ATOMIC_INIT(0);  /* Memory used for fragments */
171
172 /* Memory Tracking Functions. */
173 static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
174 {
175         if (work)
176                 *work -= skb->truesize;
177         atomic_sub(skb->truesize, &ip_frag_mem);
178         kfree_skb(skb);
179 }
180
181 static __inline__ void frag_free_queue(struct ipq *qp, int *work)
182 {
183         if (work)
184                 *work -= sizeof(struct ipq);
185         atomic_sub(sizeof(struct ipq), &ip_frag_mem);
186         kfree(qp);
187 }
188
189 static __inline__ struct ipq *frag_alloc_queue(void)
190 {
191         struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
192
193         if(!qp)
194                 return NULL;
195         atomic_add(sizeof(struct ipq), &ip_frag_mem);
196         return qp;
197 }
198
199
200 /* Destruction primitives. */
201
202 /* Complete destruction of ipq. */
203 static void ip_frag_destroy(struct ipq *qp, int *work)
204 {
205         struct sk_buff *fp;
206
207         BUG_TRAP(qp->last_in&COMPLETE);
208         BUG_TRAP(del_timer(&qp->timer) == 0);
209
210         /* Release all fragment data. */
211         fp = qp->fragments;
212         while (fp) {
213                 struct sk_buff *xp = fp->next;
214
215                 frag_kfree_skb(fp, work);
216                 fp = xp;
217         }
218
219         /* Finally, release the queue descriptor itself. */
220         frag_free_queue(qp, work);
221 }
222
223 static __inline__ void ipq_put(struct ipq *ipq, int *work)
224 {
225         if (atomic_dec_and_test(&ipq->refcnt))
226                 ip_frag_destroy(ipq, work);
227 }
228
229 /* Kill ipq entry. It is not destroyed immediately,
230  * because caller (and someone more) holds reference count.
231  */
232 static void ipq_kill(struct ipq *ipq)
233 {
234         if (del_timer(&ipq->timer))
235                 atomic_dec(&ipq->refcnt);
236
237         if (!(ipq->last_in & COMPLETE)) {
238                 ipq_unlink(ipq);
239                 atomic_dec(&ipq->refcnt);
240                 ipq->last_in |= COMPLETE;
241         }
242 }
243
244 /* Memory limiting on fragments.  Evictor trashes the oldest 
245  * fragment queue until we are back under the threshold.
246  */
247 static void ip_evictor(void)
248 {
249         struct ipq *qp;
250         struct list_head *tmp;
251         int work;
252
253         work = atomic_read(&ip_frag_mem) - sysctl_ipfrag_low_thresh;
254         if (work <= 0)
255                 return;
256
257         while (work > 0) {
258                 read_lock(&ipfrag_lock);
259                 if (list_empty(&ipq_lru_list)) {
260                         read_unlock(&ipfrag_lock);
261                         return;
262                 }
263                 tmp = ipq_lru_list.next;
264                 qp = list_entry(tmp, struct ipq, lru_list);
265                 atomic_inc(&qp->refcnt);
266                 read_unlock(&ipfrag_lock);
267
268                 spin_lock(&qp->lock);
269                 if (!(qp->last_in&COMPLETE))
270                         ipq_kill(qp);
271                 spin_unlock(&qp->lock);
272
273                 ipq_put(qp, &work);
274                 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
275         }
276 }
277
278 /*
279  * Oops, a fragment queue timed out.  Kill it and send an ICMP reply.
280  */
281 static void ip_expire(unsigned long arg)
282 {
283         struct ipq *qp = (struct ipq *) arg;
284
285         spin_lock(&qp->lock);
286
287         if (qp->last_in & COMPLETE)
288                 goto out;
289
290         ipq_kill(qp);
291
292         IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT);
293         IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
294
295         if ((qp->last_in&FIRST_IN) && qp->fragments != NULL) {
296                 struct sk_buff *head = qp->fragments;
297                 /* Send an ICMP "Fragment Reassembly Timeout" message. */
298                 if ((head->dev = dev_get_by_index(qp->iif)) != NULL) {
299                         icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
300                         dev_put(head->dev);
301                 }
302         }
303 out:
304         spin_unlock(&qp->lock);
305         ipq_put(qp, NULL);
306 }
307
308 /* Creation primitives. */
309
310 static struct ipq *ip_frag_intern(unsigned int hash, struct ipq *qp_in)
311 {
312         struct ipq *qp;
313
314         write_lock(&ipfrag_lock);
315 #ifdef CONFIG_SMP
316         /* With SMP race we have to recheck hash table, because
317          * such entry could be created on other cpu, while we
318          * promoted read lock to write lock.
319          */
320         for(qp = ipq_hash[hash]; qp; qp = qp->next) {
321                 if(qp->id == qp_in->id          &&
322                    qp->saddr == qp_in->saddr    &&
323                    qp->daddr == qp_in->daddr    &&
324                    qp->protocol == qp_in->protocol &&
325                    qp->user == qp_in->user) {
326                         atomic_inc(&qp->refcnt);
327                         write_unlock(&ipfrag_lock);
328                         qp_in->last_in |= COMPLETE;
329                         ipq_put(qp_in, NULL);
330                         return qp;
331                 }
332         }
333 #endif
334         qp = qp_in;
335
336         if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time))
337                 atomic_inc(&qp->refcnt);
338
339         atomic_inc(&qp->refcnt);
340         if((qp->next = ipq_hash[hash]) != NULL)
341                 qp->next->pprev = &qp->next;
342         ipq_hash[hash] = qp;
343         qp->pprev = &ipq_hash[hash];
344         INIT_LIST_HEAD(&qp->lru_list);
345         list_add_tail(&qp->lru_list, &ipq_lru_list);
346         ip_frag_nqueues++;
347         write_unlock(&ipfrag_lock);
348         return qp;
349 }
350
351 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
352 static struct ipq *ip_frag_create(unsigned hash, struct iphdr *iph, u32 user)
353 {
354         struct ipq *qp;
355
356         if ((qp = frag_alloc_queue()) == NULL)
357                 goto out_nomem;
358
359         qp->protocol = iph->protocol;
360         qp->last_in = 0;
361         qp->id = iph->id;
362         qp->saddr = iph->saddr;
363         qp->daddr = iph->daddr;
364         qp->user = user;
365         qp->len = 0;
366         qp->meat = 0;
367         qp->fragments = NULL;
368         qp->iif = 0;
369
370         /* Initialize a timer for this entry. */
371         init_timer(&qp->timer);
372         qp->timer.data = (unsigned long) qp;    /* pointer to queue     */
373         qp->timer.function = ip_expire;         /* expire function      */
374         spin_lock_init(&qp->lock);
375         atomic_set(&qp->refcnt, 1);
376
377         return ip_frag_intern(hash, qp);
378
379 out_nomem:
380         LIMIT_NETDEBUG(KERN_ERR "ip_frag_create: no memory left !\n");
381         return NULL;
382 }
383
384 /* Find the correct entry in the "incomplete datagrams" queue for
385  * this IP datagram, and create new one, if nothing is found.
386  */
387 static inline struct ipq *ip_find(struct iphdr *iph, u32 user)
388 {
389         __u16 id = iph->id;
390         __u32 saddr = iph->saddr;
391         __u32 daddr = iph->daddr;
392         __u8 protocol = iph->protocol;
393         unsigned int hash = ipqhashfn(id, saddr, daddr, protocol);
394         struct ipq *qp;
395
396         read_lock(&ipfrag_lock);
397         for(qp = ipq_hash[hash]; qp; qp = qp->next) {
398                 if(qp->id == id         &&
399                    qp->saddr == saddr   &&
400                    qp->daddr == daddr   &&
401                    qp->protocol == protocol &&
402                    qp->user == user) {
403                         atomic_inc(&qp->refcnt);
404                         read_unlock(&ipfrag_lock);
405                         return qp;
406                 }
407         }
408         read_unlock(&ipfrag_lock);
409
410         return ip_frag_create(hash, iph, user);
411 }
412
413 /* Add new segment to existing queue. */
414 static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
415 {
416         struct sk_buff *prev, *next;
417         int flags, offset;
418         int ihl, end;
419
420         if (qp->last_in & COMPLETE)
421                 goto err;
422
423         offset = ntohs(skb->nh.iph->frag_off);
424         flags = offset & ~IP_OFFSET;
425         offset &= IP_OFFSET;
426         offset <<= 3;           /* offset is in 8-byte chunks */
427         ihl = skb->nh.iph->ihl * 4;
428
429         /* Determine the position of this fragment. */
430         end = offset + skb->len - ihl;
431
432         /* Is this the final fragment? */
433         if ((flags & IP_MF) == 0) {
434                 /* If we already have some bits beyond end
435                  * or have different end, the segment is corrrupted.
436                  */
437                 if (end < qp->len ||
438                     ((qp->last_in & LAST_IN) && end != qp->len))
439                         goto err;
440                 qp->last_in |= LAST_IN;
441                 qp->len = end;
442         } else {
443                 if (end&7) {
444                         end &= ~7;
445                         if (skb->ip_summed != CHECKSUM_UNNECESSARY)
446                                 skb->ip_summed = CHECKSUM_NONE;
447                 }
448                 if (end > qp->len) {
449                         /* Some bits beyond end -> corruption. */
450                         if (qp->last_in & LAST_IN)
451                                 goto err;
452                         qp->len = end;
453                 }
454         }
455         if (end == offset)
456                 goto err;
457
458         if (pskb_pull(skb, ihl) == NULL)
459                 goto err;
460         if (pskb_trim_rcsum(skb, end-offset))
461                 goto err;
462
463         /* Find out which fragments are in front and at the back of us
464          * in the chain of fragments so far.  We must know where to put
465          * this fragment, right?
466          */
467         prev = NULL;
468         for(next = qp->fragments; next != NULL; next = next->next) {
469                 if (FRAG_CB(next)->offset >= offset)
470                         break;  /* bingo! */
471                 prev = next;
472         }
473
474         /* We found where to put this one.  Check for overlap with
475          * preceding fragment, and, if needed, align things so that
476          * any overlaps are eliminated.
477          */
478         if (prev) {
479                 int i = (FRAG_CB(prev)->offset + prev->len) - offset;
480
481                 if (i > 0) {
482                         offset += i;
483                         if (end <= offset)
484                                 goto err;
485                         if (!pskb_pull(skb, i))
486                                 goto err;
487                         if (skb->ip_summed != CHECKSUM_UNNECESSARY)
488                                 skb->ip_summed = CHECKSUM_NONE;
489                 }
490         }
491
492         while (next && FRAG_CB(next)->offset < end) {
493                 int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
494
495                 if (i < next->len) {
496                         /* Eat head of the next overlapped fragment
497                          * and leave the loop. The next ones cannot overlap.
498                          */
499                         if (!pskb_pull(next, i))
500                                 goto err;
501                         FRAG_CB(next)->offset += i;
502                         qp->meat -= i;
503                         if (next->ip_summed != CHECKSUM_UNNECESSARY)
504                                 next->ip_summed = CHECKSUM_NONE;
505                         break;
506                 } else {
507                         struct sk_buff *free_it = next;
508
509                         /* Old fragmnet is completely overridden with
510                          * new one drop it.
511                          */
512                         next = next->next;
513
514                         if (prev)
515                                 prev->next = next;
516                         else
517                                 qp->fragments = next;
518
519                         qp->meat -= free_it->len;
520                         frag_kfree_skb(free_it, NULL);
521                 }
522         }
523
524         FRAG_CB(skb)->offset = offset;
525
526         /* Insert this fragment in the chain of fragments. */
527         skb->next = next;
528         if (prev)
529                 prev->next = skb;
530         else
531                 qp->fragments = skb;
532
533         if (skb->dev)
534                 qp->iif = skb->dev->ifindex;
535         skb->dev = NULL;
536         skb_get_timestamp(skb, &qp->stamp);
537         qp->meat += skb->len;
538         atomic_add(skb->truesize, &ip_frag_mem);
539         if (offset == 0)
540                 qp->last_in |= FIRST_IN;
541
542         write_lock(&ipfrag_lock);
543         list_move_tail(&qp->lru_list, &ipq_lru_list);
544         write_unlock(&ipfrag_lock);
545
546         return;
547
548 err:
549         kfree_skb(skb);
550 }
551
552
553 /* Build a new IP datagram from all its fragments. */
554
555 static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
556 {
557         struct iphdr *iph;
558         struct sk_buff *fp, *head = qp->fragments;
559         int len;
560         int ihlen;
561
562         ipq_kill(qp);
563
564         BUG_TRAP(head != NULL);
565         BUG_TRAP(FRAG_CB(head)->offset == 0);
566
567         /* Allocate a new buffer for the datagram. */
568         ihlen = head->nh.iph->ihl*4;
569         len = ihlen + qp->len;
570
571         if(len > 65535)
572                 goto out_oversize;
573
574         /* Head of list must not be cloned. */
575         if (skb_cloned(head) && pskb_expand_head(head, 0, 0, GFP_ATOMIC))
576                 goto out_nomem;
577
578         /* If the first fragment is fragmented itself, we split
579          * it to two chunks: the first with data and paged part
580          * and the second, holding only fragments. */
581         if (skb_shinfo(head)->frag_list) {
582                 struct sk_buff *clone;
583                 int i, plen = 0;
584
585                 if ((clone = alloc_skb(0, GFP_ATOMIC)) == NULL)
586                         goto out_nomem;
587                 clone->next = head->next;
588                 head->next = clone;
589                 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
590                 skb_shinfo(head)->frag_list = NULL;
591                 for (i=0; i<skb_shinfo(head)->nr_frags; i++)
592                         plen += skb_shinfo(head)->frags[i].size;
593                 clone->len = clone->data_len = head->data_len - plen;
594                 head->data_len -= clone->len;
595                 head->len -= clone->len;
596                 clone->csum = 0;
597                 clone->ip_summed = head->ip_summed;
598                 atomic_add(clone->truesize, &ip_frag_mem);
599         }
600
601         skb_shinfo(head)->frag_list = head->next;
602         skb_push(head, head->data - head->nh.raw);
603         atomic_sub(head->truesize, &ip_frag_mem);
604
605         for (fp=head->next; fp; fp = fp->next) {
606                 head->data_len += fp->len;
607                 head->len += fp->len;
608                 if (head->ip_summed != fp->ip_summed)
609                         head->ip_summed = CHECKSUM_NONE;
610                 else if (head->ip_summed == CHECKSUM_HW)
611                         head->csum = csum_add(head->csum, fp->csum);
612                 head->truesize += fp->truesize;
613                 atomic_sub(fp->truesize, &ip_frag_mem);
614         }
615
616         head->next = NULL;
617         head->dev = dev;
618         skb_set_timestamp(head, &qp->stamp);
619
620         iph = head->nh.iph;
621         iph->frag_off = 0;
622         iph->tot_len = htons(len);
623         IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS);
624         qp->fragments = NULL;
625         return head;
626
627 out_nomem:
628         LIMIT_NETDEBUG(KERN_ERR "IP: queue_glue: no memory for gluing "
629                               "queue %p\n", qp);
630         goto out_fail;
631 out_oversize:
632         if (net_ratelimit())
633                 printk(KERN_INFO
634                         "Oversized IP packet from %d.%d.%d.%d.\n",
635                         NIPQUAD(qp->saddr));
636 out_fail:
637         IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
638         return NULL;
639 }
640
641 /* Process an incoming IP datagram fragment. */
642 struct sk_buff *ip_defrag(struct sk_buff *skb, u32 user)
643 {
644         struct iphdr *iph = skb->nh.iph;
645         struct ipq *qp;
646         struct net_device *dev;
647         
648         IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS);
649
650         /* Start by cleaning up the memory. */
651         if (atomic_read(&ip_frag_mem) > sysctl_ipfrag_high_thresh)
652                 ip_evictor();
653
654         dev = skb->dev;
655
656         /* Lookup (or create) queue header */
657         if ((qp = ip_find(iph, user)) != NULL) {
658                 struct sk_buff *ret = NULL;
659
660                 spin_lock(&qp->lock);
661
662                 ip_frag_queue(qp, skb);
663
664                 if (qp->last_in == (FIRST_IN|LAST_IN) &&
665                     qp->meat == qp->len)
666                         ret = ip_frag_reasm(qp, dev);
667
668                 spin_unlock(&qp->lock);
669                 ipq_put(qp, NULL);
670                 return ret;
671         }
672
673         IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
674         kfree_skb(skb);
675         return NULL;
676 }
677
678 void ipfrag_init(void)
679 {
680         ipfrag_hash_rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
681                                  (jiffies ^ (jiffies >> 6)));
682
683         init_timer(&ipfrag_secret_timer);
684         ipfrag_secret_timer.function = ipfrag_secret_rebuild;
685         ipfrag_secret_timer.expires = jiffies + sysctl_ipfrag_secret_interval;
686         add_timer(&ipfrag_secret_timer);
687 }
688
689 EXPORT_SYMBOL(ip_defrag);