port more changes to make PCI work
[linux-2.4.git] / 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 1.58.2.1 2002/01/12 07:53:15 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/types.h>
27 #include <linux/mm.h>
28 #include <linux/sched.h>
29 #include <linux/skbuff.h>
30 #include <linux/list.h>
31 #include <linux/ip.h>
32 #include <linux/icmp.h>
33 #include <linux/netdevice.h>
34 #include <linux/jhash.h>
35 #include <linux/random.h>
36 #include <net/sock.h>
37 #include <net/ip.h>
38 #include <net/icmp.h>
39 #include <net/checksum.h>
40 #include <linux/tcp.h>
41 #include <linux/udp.h>
42 #include <linux/inet.h>
43 #include <linux/netfilter_ipv4.h>
44
45 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
46  * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
47  * as well. Or notify me, at least. --ANK
48  */
49
50 /* Fragment cache limits. We will commit 256K at one time. Should we
51  * cross that limit we will prune down to 192K. This should cope with
52  * even the most extreme cases without allowing an attacker to measurably
53  * harm machine performance.
54  */
55 int sysctl_ipfrag_high_thresh = 256*1024;
56 int sysctl_ipfrag_low_thresh = 192*1024;
57
58 /* Important NOTE! Fragment queue must be destroyed before MSL expires.
59  * RFC791 is wrong proposing to prolongate timer each fragment arrival by TTL.
60  */
61 int sysctl_ipfrag_time = IP_FRAG_TIME;
62
63 struct ipfrag_skb_cb
64 {
65         struct inet_skb_parm    h;
66         int                     offset;
67 };
68
69 #define FRAG_CB(skb)    ((struct ipfrag_skb_cb*)((skb)->cb))
70
71 /* Describe an entry in the "incomplete datagrams" queue. */
72 struct ipq {
73         struct ipq      *next;          /* linked list pointers                 */
74         struct list_head lru_list;      /* lru list member                      */
75         u32             user;
76         u32             saddr;
77         u32             daddr;
78         u16             id;
79         u8              protocol;
80         u8              last_in;
81 #define COMPLETE                4
82 #define FIRST_IN                2
83 #define LAST_IN                 1
84
85         struct sk_buff  *fragments;     /* linked list of received fragments    */
86         int             len;            /* total length of original datagram    */
87         int             meat;
88         spinlock_t      lock;
89         atomic_t        refcnt;
90         struct timer_list timer;        /* when will this queue expire?         */
91         struct ipq      **pprev;
92         int             iif;
93         struct timeval  stamp;
94 };
95
96 /* Hash table. */
97
98 #define IPQ_HASHSZ      64
99
100 /* Per-bucket lock is easy to add now. */
101 static struct ipq *ipq_hash[IPQ_HASHSZ];
102 static rwlock_t ipfrag_lock = RW_LOCK_UNLOCKED;
103 static u32 ipfrag_hash_rnd;
104 static LIST_HEAD(ipq_lru_list);
105 int ip_frag_nqueues = 0;
106
107 static __inline__ void __ipq_unlink(struct ipq *qp)
108 {
109         if(qp->next)
110                 qp->next->pprev = qp->pprev;
111         *qp->pprev = qp->next;
112         list_del(&qp->lru_list);
113         ip_frag_nqueues--;
114 }
115
116 static __inline__ void ipq_unlink(struct ipq *ipq)
117 {
118         write_lock(&ipfrag_lock);
119         __ipq_unlink(ipq);
120         write_unlock(&ipfrag_lock);
121 }
122
123 static unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
124 {
125         return jhash_3words((u32)id << 16 | prot, saddr, daddr,
126                             ipfrag_hash_rnd) & (IPQ_HASHSZ - 1);
127 }
128
129 static struct timer_list ipfrag_secret_timer;
130 int sysctl_ipfrag_secret_interval = 10 * 60 * HZ;
131
132 static void ipfrag_secret_rebuild(unsigned long dummy)
133 {
134         unsigned long now = jiffies;
135         int i;
136
137         write_lock(&ipfrag_lock);
138         get_random_bytes(&ipfrag_hash_rnd, sizeof(u32));
139         for (i = 0; i < IPQ_HASHSZ; i++) {
140                 struct ipq *q;
141
142                 q = ipq_hash[i];
143                 while (q) {
144                         struct ipq *next = q->next;
145                         unsigned int hval = ipqhashfn(q->id, q->saddr,
146                                                       q->daddr, q->protocol);
147
148                         if (hval != i) {
149                                 /* Unlink. */
150                                 if (q->next)
151                                         q->next->pprev = q->pprev;
152                                 *q->pprev = q->next;
153
154                                 /* Relink to new hash chain. */
155                                 if ((q->next = ipq_hash[hval]) != NULL)
156                                         q->next->pprev = &q->next;
157                                 ipq_hash[hval] = q;
158                                 q->pprev = &ipq_hash[hval];
159                         }
160
161                         q = next;
162                 }
163         }
164         write_unlock(&ipfrag_lock);
165
166         mod_timer(&ipfrag_secret_timer, now + sysctl_ipfrag_secret_interval);
167 }
168
169 atomic_t ip_frag_mem = ATOMIC_INIT(0);  /* Memory used for fragments */
170
171 /* Memory Tracking Functions. */
172 static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
173 {
174         if (work)
175                 *work -= skb->truesize;
176         atomic_sub(skb->truesize, &ip_frag_mem);
177         kfree_skb(skb);
178 }
179
180 static __inline__ void frag_free_queue(struct ipq *qp, int *work)
181 {
182         if (work)
183                 *work -= sizeof(struct ipq);
184         atomic_sub(sizeof(struct ipq), &ip_frag_mem);
185         kfree(qp);
186 }
187
188 static __inline__ struct ipq *frag_alloc_queue(void)
189 {
190         struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
191
192         if(!qp)
193                 return NULL;
194         atomic_add(sizeof(struct ipq), &ip_frag_mem);
195         return qp;
196 }
197
198
199 /* Destruction primitives. */
200
201 /* Complete destruction of ipq. */
202 static void ip_frag_destroy(struct ipq *qp, int *work)
203 {
204         struct sk_buff *fp;
205
206         BUG_TRAP(qp->last_in&COMPLETE);
207         BUG_TRAP(del_timer(&qp->timer) == 0);
208
209         /* Release all fragment data. */
210         fp = qp->fragments;
211         while (fp) {
212                 struct sk_buff *xp = fp->next;
213
214                 frag_kfree_skb(fp, work);
215                 fp = xp;
216         }
217
218         /* Finally, release the queue descriptor itself. */
219         frag_free_queue(qp, work);
220 }
221
222 static __inline__ void ipq_put(struct ipq *ipq, int *work)
223 {
224         if (atomic_dec_and_test(&ipq->refcnt))
225                 ip_frag_destroy(ipq, work);
226 }
227
228 /* Kill ipq entry. It is not destroyed immediately,
229  * because caller (and someone more) holds reference count.
230  */
231 static __inline__ void ipq_kill(struct ipq *ipq)
232 {
233         if (del_timer(&ipq->timer))
234                 atomic_dec(&ipq->refcnt);
235
236         if (!(ipq->last_in & COMPLETE)) {
237                 ipq_unlink(ipq);
238                 atomic_dec(&ipq->refcnt);
239                 ipq->last_in |= COMPLETE;
240         }
241 }
242
243 /* Memory limiting on fragments.  Evictor trashes the oldest 
244  * fragment queue until we are back under the threshold.
245  */
246 static void ip_evictor(void)
247 {
248         struct ipq *qp;
249         struct list_head *tmp;
250         int work;
251
252         work = atomic_read(&ip_frag_mem) - sysctl_ipfrag_low_thresh;
253         if (work <= 0)
254                 return;
255
256         while (work > 0) {
257                 read_lock(&ipfrag_lock);
258                 if (list_empty(&ipq_lru_list)) {
259                         read_unlock(&ipfrag_lock);
260                         return;
261                 }
262                 tmp = ipq_lru_list.next;
263                 qp = list_entry(tmp, struct ipq, lru_list);
264                 atomic_inc(&qp->refcnt);
265                 read_unlock(&ipfrag_lock);
266
267                 spin_lock(&qp->lock);
268                 if (!(qp->last_in&COMPLETE))
269                         ipq_kill(qp);
270                 spin_unlock(&qp->lock);
271
272                 ipq_put(qp, &work);
273                 IP_INC_STATS_BH(IpReasmFails);
274         }
275 }
276
277 /*
278  * Oops, a fragment queue timed out.  Kill it and send an ICMP reply.
279  */
280 static void ip_expire(unsigned long arg)
281 {
282         struct ipq *qp = (struct ipq *) arg;
283
284         spin_lock(&qp->lock);
285
286         if (qp->last_in & COMPLETE)
287                 goto out;
288
289         ipq_kill(qp);
290
291         IP_INC_STATS_BH(IpReasmTimeout);
292         IP_INC_STATS_BH(IpReasmFails);
293
294         if ((qp->last_in&FIRST_IN) && qp->fragments != NULL) {
295                 struct sk_buff *head = qp->fragments;
296                 /* Send an ICMP "Fragment Reassembly Timeout" message. */
297                 if ((head->dev = dev_get_by_index(qp->iif)) != NULL) {
298                         icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
299                         dev_put(head->dev);
300                 }
301         }
302 out:
303         spin_unlock(&qp->lock);
304         ipq_put(qp, NULL);
305 }
306
307 /* Creation primitives. */
308
309 static struct ipq *ip_frag_intern(unsigned int hash, struct ipq *qp_in)
310 {
311         struct ipq *qp;
312
313         write_lock(&ipfrag_lock);
314 #ifdef CONFIG_SMP
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.
318          */
319         for(qp = ipq_hash[hash]; qp; qp = qp->next) {
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);
329                         return qp;
330                 }
331         }
332 #endif
333         qp = qp_in;
334
335         if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time))
336                 atomic_inc(&qp->refcnt);
337
338         atomic_inc(&qp->refcnt);
339         if((qp->next = ipq_hash[hash]) != NULL)
340                 qp->next->pprev = &qp->next;
341         ipq_hash[hash] = qp;
342         qp->pprev = &ipq_hash[hash];
343         INIT_LIST_HEAD(&qp->lru_list);
344         list_add_tail(&qp->lru_list, &ipq_lru_list);
345         ip_frag_nqueues++;
346         write_unlock(&ipfrag_lock);
347         return qp;
348 }
349
350 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
351 static struct ipq *ip_frag_create(unsigned hash, struct iphdr *iph, u32 user)
352 {
353         struct ipq *qp;
354
355         if ((qp = frag_alloc_queue()) == NULL)
356                 goto out_nomem;
357
358         qp->protocol = iph->protocol;
359         qp->last_in = 0;
360         qp->id = iph->id;
361         qp->saddr = iph->saddr;
362         qp->daddr = iph->daddr;
363         qp->user = user;
364         qp->len = 0;
365         qp->meat = 0;
366         qp->fragments = NULL;
367         qp->iif = 0;
368
369         /* Initialize a timer for this entry. */
370         init_timer(&qp->timer);
371         qp->timer.data = (unsigned long) qp;    /* pointer to queue     */
372         qp->timer.function = ip_expire;         /* expire function      */
373         qp->lock = SPIN_LOCK_UNLOCKED;
374         atomic_set(&qp->refcnt, 1);
375
376         return ip_frag_intern(hash, qp);
377
378 out_nomem:
379         NETDEBUG(if (net_ratelimit()) printk(KERN_ERR "ip_frag_create: no memory left !\n"));
380         return NULL;
381 }
382
383 /* Find the correct entry in the "incomplete datagrams" queue for
384  * this IP datagram, and create new one, if nothing is found.
385  */
386 static inline struct ipq *ip_find(struct iphdr *iph, u32 user)
387 {
388         __u16 id = iph->id;
389         __u32 saddr = iph->saddr;
390         __u32 daddr = iph->daddr;
391         __u8 protocol = iph->protocol;
392         unsigned int hash = ipqhashfn(id, saddr, daddr, protocol);
393         struct ipq *qp;
394
395         read_lock(&ipfrag_lock);
396         for(qp = ipq_hash[hash]; qp; qp = qp->next) {
397                 if(qp->id == id         &&
398                    qp->saddr == saddr   &&
399                    qp->daddr == daddr   &&
400                    qp->protocol == protocol &&
401                    qp->user == user) {
402                         atomic_inc(&qp->refcnt);
403                         read_unlock(&ipfrag_lock);
404                         return qp;
405                 }
406         }
407         read_unlock(&ipfrag_lock);
408
409         return ip_frag_create(hash, iph, user);
410 }
411
412 /* Add new segment to existing queue. */
413 static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
414 {
415         struct sk_buff *prev, *next;
416         int flags, offset;
417         int ihl, end;
418
419         if (qp->last_in & COMPLETE)
420                 goto err;
421
422         offset = ntohs(skb->nh.iph->frag_off);
423         flags = offset & ~IP_OFFSET;
424         offset &= IP_OFFSET;
425         offset <<= 3;           /* offset is in 8-byte chunks */
426         ihl = skb->nh.iph->ihl * 4;
427
428         /* Determine the position of this fragment. */
429         end = offset + skb->len - ihl;
430
431         /* Is this the final fragment? */
432         if ((flags & IP_MF) == 0) {
433                 /* If we already have some bits beyond end
434                  * or have different end, the segment is corrrupted.
435                  */
436                 if (end < qp->len ||
437                     ((qp->last_in & LAST_IN) && end != qp->len))
438                         goto err;
439                 qp->last_in |= LAST_IN;
440                 qp->len = end;
441         } else {
442                 if (end&7) {
443                         end &= ~7;
444                         if (skb->ip_summed != CHECKSUM_UNNECESSARY)
445                                 skb->ip_summed = CHECKSUM_NONE;
446                 }
447                 if (end > qp->len) {
448                         /* Some bits beyond end -> corruption. */
449                         if (qp->last_in & LAST_IN)
450                                 goto err;
451                         qp->len = end;
452                 }
453         }
454         if (end == offset)
455                 goto err;
456
457         if (pskb_pull(skb, ihl) == NULL)
458                 goto err;
459         if (pskb_trim(skb, end-offset))
460                 goto err;
461
462         /* Find out which fragments are in front and at the back of us
463          * in the chain of fragments so far.  We must know where to put
464          * this fragment, right?
465          */
466         prev = NULL;
467         for(next = qp->fragments; next != NULL; next = next->next) {
468                 if (FRAG_CB(next)->offset >= offset)
469                         break;  /* bingo! */
470                 prev = next;
471         }
472
473         /* We found where to put this one.  Check for overlap with
474          * preceding fragment, and, if needed, align things so that
475          * any overlaps are eliminated.
476          */
477         if (prev) {
478                 int i = (FRAG_CB(prev)->offset + prev->len) - offset;
479
480                 if (i > 0) {
481                         offset += i;
482                         if (end <= offset)
483                                 goto err;
484                         if (!pskb_pull(skb, i))
485                                 goto err;
486                         if (skb->ip_summed != CHECKSUM_UNNECESSARY)
487                                 skb->ip_summed = CHECKSUM_NONE;
488                 }
489         }
490
491         while (next && FRAG_CB(next)->offset < end) {
492                 int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
493
494                 if (i < next->len) {
495                         /* Eat head of the next overlapped fragment
496                          * and leave the loop. The next ones cannot overlap.
497                          */
498                         if (!pskb_pull(next, i))
499                                 goto err;
500                         FRAG_CB(next)->offset += i;
501                         qp->meat -= i;
502                         if (next->ip_summed != CHECKSUM_UNNECESSARY)
503                                 next->ip_summed = CHECKSUM_NONE;
504                         break;
505                 } else {
506                         struct sk_buff *free_it = next;
507
508                         /* Old fragmnet is completely overridden with
509                          * new one drop it.
510                          */
511                         next = next->next;
512
513                         if (prev)
514                                 prev->next = next;
515                         else
516                                 qp->fragments = next;
517
518                         qp->meat -= free_it->len;
519                         frag_kfree_skb(free_it, NULL);
520                 }
521         }
522
523         FRAG_CB(skb)->offset = offset;
524
525         /* Insert this fragment in the chain of fragments. */
526         skb->next = next;
527         if (prev)
528                 prev->next = skb;
529         else
530                 qp->fragments = skb;
531
532         if (skb->dev)
533                 qp->iif = skb->dev->ifindex;
534         skb->dev = NULL;
535         qp->stamp = skb->stamp;
536         qp->meat += skb->len;
537         atomic_add(skb->truesize, &ip_frag_mem);
538         if (offset == 0)
539                 qp->last_in |= FIRST_IN;
540
541         write_lock(&ipfrag_lock);
542         list_move_tail(&qp->lru_list, &ipq_lru_list);
543         write_unlock(&ipfrag_lock);
544
545         return;
546
547 err:
548         kfree_skb(skb);
549 }
550
551
552 /* Build a new IP datagram from all its fragments. */
553
554 static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
555 {
556         struct iphdr *iph;
557         struct sk_buff *fp, *head = qp->fragments;
558         int len;
559         int ihlen;
560
561         ipq_kill(qp);
562
563         BUG_TRAP(head != NULL);
564         BUG_TRAP(FRAG_CB(head)->offset == 0);
565
566         /* Allocate a new buffer for the datagram. */
567         ihlen = head->nh.iph->ihl*4;
568         len = ihlen + qp->len;
569
570         if(len > 65535)
571                 goto out_oversize;
572
573         /* Head of list must not be cloned. */
574         if (skb_cloned(head) && pskb_expand_head(head, 0, 0, GFP_ATOMIC))
575                 goto out_nomem;
576
577         /* If the first fragment is fragmented itself, we split
578          * it to two chunks: the first with data and paged part
579          * and the second, holding only fragments. */
580         if (skb_shinfo(head)->frag_list) {
581                 struct sk_buff *clone;
582                 int i, plen = 0;
583
584                 if ((clone = alloc_skb(0, GFP_ATOMIC)) == NULL)
585                         goto out_nomem;
586                 clone->next = head->next;
587                 head->next = clone;
588                 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
589                 skb_shinfo(head)->frag_list = NULL;
590                 for (i=0; i<skb_shinfo(head)->nr_frags; i++)
591                         plen += skb_shinfo(head)->frags[i].size;
592                 clone->len = clone->data_len = head->data_len - plen;
593                 head->data_len -= clone->len;
594                 head->len -= clone->len;
595                 clone->csum = 0;
596                 clone->ip_summed = head->ip_summed;
597                 atomic_add(clone->truesize, &ip_frag_mem);
598         }
599
600         skb_shinfo(head)->frag_list = head->next;
601         skb_push(head, head->data - head->nh.raw);
602         atomic_sub(head->truesize, &ip_frag_mem);
603
604         for (fp=head->next; fp; fp = fp->next) {
605                 head->data_len += fp->len;
606                 head->len += fp->len;
607                 if (head->ip_summed != fp->ip_summed)
608                         head->ip_summed = CHECKSUM_NONE;
609                 else if (head->ip_summed == CHECKSUM_HW)
610                         head->csum = csum_add(head->csum, fp->csum);
611                 head->truesize += fp->truesize;
612                 atomic_sub(fp->truesize, &ip_frag_mem);
613         }
614
615         head->next = NULL;
616         head->dev = dev;
617         head->stamp = qp->stamp;
618
619         iph = head->nh.iph;
620         iph->frag_off = 0;
621         iph->tot_len = htons(len);
622         IP_INC_STATS_BH(IpReasmOKs);
623         qp->fragments = NULL;
624         return head;
625
626 out_nomem:
627         NETDEBUG(if (net_ratelimit())
628                  printk(KERN_ERR 
629                         "IP: queue_glue: no memory for gluing queue %p\n",
630                         qp));
631         goto out_fail;
632 out_oversize:
633         if (net_ratelimit())
634                 printk(KERN_INFO
635                         "Oversized IP packet from %d.%d.%d.%d.\n",
636                         NIPQUAD(qp->saddr));
637 out_fail:
638         IP_INC_STATS_BH(IpReasmFails);
639         return NULL;
640 }
641
642 /* Process an incoming IP datagram fragment. */
643 struct sk_buff *ip_defrag(struct sk_buff *skb, u32 user)
644 {
645         struct iphdr *iph = skb->nh.iph;
646         struct ipq *qp;
647         struct net_device *dev;
648         
649         IP_INC_STATS_BH(IpReasmReqds);
650
651         /* Start by cleaning up the memory. */
652         if (atomic_read(&ip_frag_mem) > sysctl_ipfrag_high_thresh)
653                 ip_evictor();
654
655         dev = skb->dev;
656
657         /* Lookup (or create) queue header */
658         if ((qp = ip_find(iph, user)) != NULL) {
659                 struct sk_buff *ret = NULL;
660
661                 spin_lock(&qp->lock);
662
663                 ip_frag_queue(qp, skb);
664
665                 if (qp->last_in == (FIRST_IN|LAST_IN) &&
666                     qp->meat == qp->len)
667                         ret = ip_frag_reasm(qp, dev);
668
669                 spin_unlock(&qp->lock);
670                 ipq_put(qp, NULL);
671                 return ret;
672         }
673
674         IP_INC_STATS_BH(IpReasmFails);
675         kfree_skb(skb);
676         return NULL;
677 }
678
679 void ipfrag_init(void)
680 {
681         ipfrag_hash_rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
682                                  (jiffies ^ (jiffies >> 6)));
683
684         init_timer(&ipfrag_secret_timer);
685         ipfrag_secret_timer.function = ipfrag_secret_rebuild;
686         ipfrag_secret_timer.expires = jiffies + sysctl_ipfrag_secret_interval;
687         add_timer(&ipfrag_secret_timer);
688 }