Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6.git] / net / netfilter / nf_conntrack_core.c
1 /* Connection state tracking for netfilter.  This is separated from,
2    but required by, the NAT layer; it can also be used by an iptables
3    extension. */
4
5 /* (C) 1999-2001 Paul `Rusty' Russell
6  * (C) 2002-2006 Netfilter Core Team <coreteam@netfilter.org>
7  * (C) 2003,2004 USAGI/WIDE Project <http://www.linux-ipv6.org>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License version 2 as
11  * published by the Free Software Foundation.
12  */
13
14 #include <linux/types.h>
15 #include <linux/netfilter.h>
16 #include <linux/module.h>
17 #include <linux/sched.h>
18 #include <linux/skbuff.h>
19 #include <linux/proc_fs.h>
20 #include <linux/vmalloc.h>
21 #include <linux/stddef.h>
22 #include <linux/slab.h>
23 #include <linux/random.h>
24 #include <linux/jhash.h>
25 #include <linux/err.h>
26 #include <linux/percpu.h>
27 #include <linux/moduleparam.h>
28 #include <linux/notifier.h>
29 #include <linux/kernel.h>
30 #include <linux/netdevice.h>
31 #include <linux/socket.h>
32 #include <linux/mm.h>
33 #include <linux/rculist_nulls.h>
34
35 #include <net/netfilter/nf_conntrack.h>
36 #include <net/netfilter/nf_conntrack_l3proto.h>
37 #include <net/netfilter/nf_conntrack_l4proto.h>
38 #include <net/netfilter/nf_conntrack_expect.h>
39 #include <net/netfilter/nf_conntrack_helper.h>
40 #include <net/netfilter/nf_conntrack_core.h>
41 #include <net/netfilter/nf_conntrack_extend.h>
42 #include <net/netfilter/nf_conntrack_acct.h>
43 #include <net/netfilter/nf_conntrack_ecache.h>
44 #include <net/netfilter/nf_nat.h>
45 #include <net/netfilter/nf_nat_core.h>
46
47 #define NF_CONNTRACK_VERSION    "0.5.0"
48
49 int (*nfnetlink_parse_nat_setup_hook)(struct nf_conn *ct,
50                                       enum nf_nat_manip_type manip,
51                                       const struct nlattr *attr) __read_mostly;
52 EXPORT_SYMBOL_GPL(nfnetlink_parse_nat_setup_hook);
53
54 DEFINE_SPINLOCK(nf_conntrack_lock);
55 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
56
57 unsigned int nf_conntrack_htable_size __read_mostly;
58 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
59
60 unsigned int nf_conntrack_max __read_mostly;
61 EXPORT_SYMBOL_GPL(nf_conntrack_max);
62
63 struct nf_conn nf_conntrack_untracked __read_mostly;
64 EXPORT_SYMBOL_GPL(nf_conntrack_untracked);
65
66 static struct kmem_cache *nf_conntrack_cachep __read_mostly;
67
68 static int nf_conntrack_hash_rnd_initted;
69 static unsigned int nf_conntrack_hash_rnd;
70
71 static u_int32_t __hash_conntrack(const struct nf_conntrack_tuple *tuple,
72                                   unsigned int size, unsigned int rnd)
73 {
74         unsigned int n;
75         u_int32_t h;
76
77         /* The direction must be ignored, so we hash everything up to the
78          * destination ports (which is a multiple of 4) and treat the last
79          * three bytes manually.
80          */
81         n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
82         h = jhash2((u32 *)tuple, n,
83                    rnd ^ (((__force __u16)tuple->dst.u.all << 16) |
84                           tuple->dst.protonum));
85
86         return ((u64)h * size) >> 32;
87 }
88
89 static inline u_int32_t hash_conntrack(const struct nf_conntrack_tuple *tuple)
90 {
91         return __hash_conntrack(tuple, nf_conntrack_htable_size,
92                                 nf_conntrack_hash_rnd);
93 }
94
95 bool
96 nf_ct_get_tuple(const struct sk_buff *skb,
97                 unsigned int nhoff,
98                 unsigned int dataoff,
99                 u_int16_t l3num,
100                 u_int8_t protonum,
101                 struct nf_conntrack_tuple *tuple,
102                 const struct nf_conntrack_l3proto *l3proto,
103                 const struct nf_conntrack_l4proto *l4proto)
104 {
105         memset(tuple, 0, sizeof(*tuple));
106
107         tuple->src.l3num = l3num;
108         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
109                 return false;
110
111         tuple->dst.protonum = protonum;
112         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
113
114         return l4proto->pkt_to_tuple(skb, dataoff, tuple);
115 }
116 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
117
118 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
119                        u_int16_t l3num, struct nf_conntrack_tuple *tuple)
120 {
121         struct nf_conntrack_l3proto *l3proto;
122         struct nf_conntrack_l4proto *l4proto;
123         unsigned int protoff;
124         u_int8_t protonum;
125         int ret;
126
127         rcu_read_lock();
128
129         l3proto = __nf_ct_l3proto_find(l3num);
130         ret = l3proto->get_l4proto(skb, nhoff, &protoff, &protonum);
131         if (ret != NF_ACCEPT) {
132                 rcu_read_unlock();
133                 return false;
134         }
135
136         l4proto = __nf_ct_l4proto_find(l3num, protonum);
137
138         ret = nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, tuple,
139                               l3proto, l4proto);
140
141         rcu_read_unlock();
142         return ret;
143 }
144 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
145
146 bool
147 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
148                    const struct nf_conntrack_tuple *orig,
149                    const struct nf_conntrack_l3proto *l3proto,
150                    const struct nf_conntrack_l4proto *l4proto)
151 {
152         memset(inverse, 0, sizeof(*inverse));
153
154         inverse->src.l3num = orig->src.l3num;
155         if (l3proto->invert_tuple(inverse, orig) == 0)
156                 return false;
157
158         inverse->dst.dir = !orig->dst.dir;
159
160         inverse->dst.protonum = orig->dst.protonum;
161         return l4proto->invert_tuple(inverse, orig);
162 }
163 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
164
165 static void
166 clean_from_lists(struct nf_conn *ct)
167 {
168         pr_debug("clean_from_lists(%p)\n", ct);
169         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
170         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
171
172         /* Destroy all pending expectations */
173         nf_ct_remove_expectations(ct);
174 }
175
176 static void
177 destroy_conntrack(struct nf_conntrack *nfct)
178 {
179         struct nf_conn *ct = (struct nf_conn *)nfct;
180         struct net *net = nf_ct_net(ct);
181         struct nf_conntrack_l4proto *l4proto;
182
183         pr_debug("destroy_conntrack(%p)\n", ct);
184         NF_CT_ASSERT(atomic_read(&nfct->use) == 0);
185         NF_CT_ASSERT(!timer_pending(&ct->timeout));
186
187         /* To make sure we don't get any weird locking issues here:
188          * destroy_conntrack() MUST NOT be called with a write lock
189          * to nf_conntrack_lock!!! -HW */
190         rcu_read_lock();
191         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
192         if (l4proto && l4proto->destroy)
193                 l4proto->destroy(ct);
194
195         rcu_read_unlock();
196
197         spin_lock_bh(&nf_conntrack_lock);
198         /* Expectations will have been removed in clean_from_lists,
199          * except TFTP can create an expectation on the first packet,
200          * before connection is in the list, so we need to clean here,
201          * too. */
202         nf_ct_remove_expectations(ct);
203
204         /* We overload first tuple to link into unconfirmed list. */
205         if (!nf_ct_is_confirmed(ct)) {
206                 BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
207                 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
208         }
209
210         NF_CT_STAT_INC(net, delete);
211         spin_unlock_bh(&nf_conntrack_lock);
212
213         if (ct->master)
214                 nf_ct_put(ct->master);
215
216         pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
217         nf_conntrack_free(ct);
218 }
219
220 void nf_ct_delete_from_lists(struct nf_conn *ct)
221 {
222         struct net *net = nf_ct_net(ct);
223
224         nf_ct_helper_destroy(ct);
225         spin_lock_bh(&nf_conntrack_lock);
226         /* Inside lock so preempt is disabled on module removal path.
227          * Otherwise we can get spurious warnings. */
228         NF_CT_STAT_INC(net, delete_list);
229         clean_from_lists(ct);
230         spin_unlock_bh(&nf_conntrack_lock);
231 }
232 EXPORT_SYMBOL_GPL(nf_ct_delete_from_lists);
233
234 static void death_by_event(unsigned long ul_conntrack)
235 {
236         struct nf_conn *ct = (void *)ul_conntrack;
237         struct net *net = nf_ct_net(ct);
238
239         if (nf_conntrack_event(IPCT_DESTROY, ct) < 0) {
240                 /* bad luck, let's retry again */
241                 ct->timeout.expires = jiffies +
242                         (random32() % net->ct.sysctl_events_retry_timeout);
243                 add_timer(&ct->timeout);
244                 return;
245         }
246         /* we've got the event delivered, now it's dying */
247         set_bit(IPS_DYING_BIT, &ct->status);
248         spin_lock(&nf_conntrack_lock);
249         hlist_nulls_del(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
250         spin_unlock(&nf_conntrack_lock);
251         nf_ct_put(ct);
252 }
253
254 void nf_ct_insert_dying_list(struct nf_conn *ct)
255 {
256         struct net *net = nf_ct_net(ct);
257
258         /* add this conntrack to the dying list */
259         spin_lock_bh(&nf_conntrack_lock);
260         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
261                              &net->ct.dying);
262         spin_unlock_bh(&nf_conntrack_lock);
263         /* set a new timer to retry event delivery */
264         setup_timer(&ct->timeout, death_by_event, (unsigned long)ct);
265         ct->timeout.expires = jiffies +
266                 (random32() % net->ct.sysctl_events_retry_timeout);
267         add_timer(&ct->timeout);
268 }
269 EXPORT_SYMBOL_GPL(nf_ct_insert_dying_list);
270
271 static void death_by_timeout(unsigned long ul_conntrack)
272 {
273         struct nf_conn *ct = (void *)ul_conntrack;
274
275         if (!test_bit(IPS_DYING_BIT, &ct->status) &&
276             unlikely(nf_conntrack_event(IPCT_DESTROY, ct) < 0)) {
277                 /* destroy event was not delivered */
278                 nf_ct_delete_from_lists(ct);
279                 nf_ct_insert_dying_list(ct);
280                 return;
281         }
282         set_bit(IPS_DYING_BIT, &ct->status);
283         nf_ct_delete_from_lists(ct);
284         nf_ct_put(ct);
285 }
286
287 /*
288  * Warning :
289  * - Caller must take a reference on returned object
290  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
291  * OR
292  * - Caller must lock nf_conntrack_lock before calling this function
293  */
294 struct nf_conntrack_tuple_hash *
295 __nf_conntrack_find(struct net *net, const struct nf_conntrack_tuple *tuple)
296 {
297         struct nf_conntrack_tuple_hash *h;
298         struct hlist_nulls_node *n;
299         unsigned int hash = hash_conntrack(tuple);
300
301         /* Disable BHs the entire time since we normally need to disable them
302          * at least once for the stats anyway.
303          */
304         local_bh_disable();
305 begin:
306         hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash], hnnode) {
307                 if (nf_ct_tuple_equal(tuple, &h->tuple)) {
308                         NF_CT_STAT_INC(net, found);
309                         local_bh_enable();
310                         return h;
311                 }
312                 NF_CT_STAT_INC(net, searched);
313         }
314         /*
315          * if the nulls value we got at the end of this lookup is
316          * not the expected one, we must restart lookup.
317          * We probably met an item that was moved to another chain.
318          */
319         if (get_nulls_value(n) != hash)
320                 goto begin;
321         local_bh_enable();
322
323         return NULL;
324 }
325 EXPORT_SYMBOL_GPL(__nf_conntrack_find);
326
327 /* Find a connection corresponding to a tuple. */
328 struct nf_conntrack_tuple_hash *
329 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_tuple *tuple)
330 {
331         struct nf_conntrack_tuple_hash *h;
332         struct nf_conn *ct;
333
334         rcu_read_lock();
335 begin:
336         h = __nf_conntrack_find(net, tuple);
337         if (h) {
338                 ct = nf_ct_tuplehash_to_ctrack(h);
339                 if (unlikely(nf_ct_is_dying(ct) ||
340                              !atomic_inc_not_zero(&ct->ct_general.use)))
341                         h = NULL;
342                 else {
343                         if (unlikely(!nf_ct_tuple_equal(tuple, &h->tuple))) {
344                                 nf_ct_put(ct);
345                                 goto begin;
346                         }
347                 }
348         }
349         rcu_read_unlock();
350
351         return h;
352 }
353 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
354
355 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
356                                        unsigned int hash,
357                                        unsigned int repl_hash)
358 {
359         struct net *net = nf_ct_net(ct);
360
361         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
362                            &net->ct.hash[hash]);
363         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
364                            &net->ct.hash[repl_hash]);
365 }
366
367 void nf_conntrack_hash_insert(struct nf_conn *ct)
368 {
369         unsigned int hash, repl_hash;
370
371         hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
372         repl_hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_REPLY].tuple);
373
374         __nf_conntrack_hash_insert(ct, hash, repl_hash);
375 }
376 EXPORT_SYMBOL_GPL(nf_conntrack_hash_insert);
377
378 /* Confirm a connection given skb; places it in hash table */
379 int
380 __nf_conntrack_confirm(struct sk_buff *skb)
381 {
382         unsigned int hash, repl_hash;
383         struct nf_conntrack_tuple_hash *h;
384         struct nf_conn *ct;
385         struct nf_conn_help *help;
386         struct hlist_nulls_node *n;
387         enum ip_conntrack_info ctinfo;
388         struct net *net;
389
390         ct = nf_ct_get(skb, &ctinfo);
391         net = nf_ct_net(ct);
392
393         /* ipt_REJECT uses nf_conntrack_attach to attach related
394            ICMP/TCP RST packets in other direction.  Actual packet
395            which created connection will be IP_CT_NEW or for an
396            expected connection, IP_CT_RELATED. */
397         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
398                 return NF_ACCEPT;
399
400         hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
401         repl_hash = hash_conntrack(&ct->tuplehash[IP_CT_DIR_REPLY].tuple);
402
403         /* We're not in hash table, and we refuse to set up related
404            connections for unconfirmed conns.  But packet copies and
405            REJECT will give spurious warnings here. */
406         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
407
408         /* No external references means noone else could have
409            confirmed us. */
410         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
411         pr_debug("Confirming conntrack %p\n", ct);
412
413         spin_lock_bh(&nf_conntrack_lock);
414
415         /* See if there's one in the list already, including reverse:
416            NAT could have grabbed it without realizing, since we're
417            not in the hash.  If there is, we lost race. */
418         hlist_nulls_for_each_entry(h, n, &net->ct.hash[hash], hnnode)
419                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
420                                       &h->tuple))
421                         goto out;
422         hlist_nulls_for_each_entry(h, n, &net->ct.hash[repl_hash], hnnode)
423                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_REPLY].tuple,
424                                       &h->tuple))
425                         goto out;
426
427         /* Remove from unconfirmed list */
428         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
429
430         /* Timer relative to confirmation time, not original
431            setting time, otherwise we'd get timer wrap in
432            weird delay cases. */
433         ct->timeout.expires += jiffies;
434         add_timer(&ct->timeout);
435         atomic_inc(&ct->ct_general.use);
436         set_bit(IPS_CONFIRMED_BIT, &ct->status);
437
438         /* Since the lookup is lockless, hash insertion must be done after
439          * starting the timer and setting the CONFIRMED bit. The RCU barriers
440          * guarantee that no other CPU can find the conntrack before the above
441          * stores are visible.
442          */
443         __nf_conntrack_hash_insert(ct, hash, repl_hash);
444         NF_CT_STAT_INC(net, insert);
445         spin_unlock_bh(&nf_conntrack_lock);
446
447         help = nfct_help(ct);
448         if (help && help->helper)
449                 nf_conntrack_event_cache(IPCT_HELPER, ct);
450
451         nf_conntrack_event_cache(master_ct(ct) ?
452                                  IPCT_RELATED : IPCT_NEW, ct);
453         return NF_ACCEPT;
454
455 out:
456         NF_CT_STAT_INC(net, insert_failed);
457         spin_unlock_bh(&nf_conntrack_lock);
458         return NF_DROP;
459 }
460 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
461
462 /* Returns true if a connection correspondings to the tuple (required
463    for NAT). */
464 int
465 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
466                          const struct nf_conn *ignored_conntrack)
467 {
468         struct net *net = nf_ct_net(ignored_conntrack);
469         struct nf_conntrack_tuple_hash *h;
470         struct hlist_nulls_node *n;
471         unsigned int hash = hash_conntrack(tuple);
472
473         /* Disable BHs the entire time since we need to disable them at
474          * least once for the stats anyway.
475          */
476         rcu_read_lock_bh();
477         hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash], hnnode) {
478                 if (nf_ct_tuplehash_to_ctrack(h) != ignored_conntrack &&
479                     nf_ct_tuple_equal(tuple, &h->tuple)) {
480                         NF_CT_STAT_INC(net, found);
481                         rcu_read_unlock_bh();
482                         return 1;
483                 }
484                 NF_CT_STAT_INC(net, searched);
485         }
486         rcu_read_unlock_bh();
487
488         return 0;
489 }
490 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
491
492 #define NF_CT_EVICTION_RANGE    8
493
494 /* There's a small race here where we may free a just-assured
495    connection.  Too bad: we're in trouble anyway. */
496 static noinline int early_drop(struct net *net, unsigned int hash)
497 {
498         /* Use oldest entry, which is roughly LRU */
499         struct nf_conntrack_tuple_hash *h;
500         struct nf_conn *ct = NULL, *tmp;
501         struct hlist_nulls_node *n;
502         unsigned int i, cnt = 0;
503         int dropped = 0;
504
505         rcu_read_lock();
506         for (i = 0; i < nf_conntrack_htable_size; i++) {
507                 hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash],
508                                          hnnode) {
509                         tmp = nf_ct_tuplehash_to_ctrack(h);
510                         if (!test_bit(IPS_ASSURED_BIT, &tmp->status))
511                                 ct = tmp;
512                         cnt++;
513                 }
514
515                 if (ct && unlikely(nf_ct_is_dying(ct) ||
516                                    !atomic_inc_not_zero(&ct->ct_general.use)))
517                         ct = NULL;
518                 if (ct || cnt >= NF_CT_EVICTION_RANGE)
519                         break;
520                 hash = (hash + 1) % nf_conntrack_htable_size;
521         }
522         rcu_read_unlock();
523
524         if (!ct)
525                 return dropped;
526
527         if (del_timer(&ct->timeout)) {
528                 death_by_timeout((unsigned long)ct);
529                 dropped = 1;
530                 NF_CT_STAT_INC_ATOMIC(net, early_drop);
531         }
532         nf_ct_put(ct);
533         return dropped;
534 }
535
536 struct nf_conn *nf_conntrack_alloc(struct net *net,
537                                    const struct nf_conntrack_tuple *orig,
538                                    const struct nf_conntrack_tuple *repl,
539                                    gfp_t gfp)
540 {
541         struct nf_conn *ct;
542
543         if (unlikely(!nf_conntrack_hash_rnd_initted)) {
544                 get_random_bytes(&nf_conntrack_hash_rnd,
545                                 sizeof(nf_conntrack_hash_rnd));
546                 nf_conntrack_hash_rnd_initted = 1;
547         }
548
549         /* We don't want any race condition at early drop stage */
550         atomic_inc(&net->ct.count);
551
552         if (nf_conntrack_max &&
553             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
554                 unsigned int hash = hash_conntrack(orig);
555                 if (!early_drop(net, hash)) {
556                         atomic_dec(&net->ct.count);
557                         if (net_ratelimit())
558                                 printk(KERN_WARNING
559                                        "nf_conntrack: table full, dropping"
560                                        " packet.\n");
561                         return ERR_PTR(-ENOMEM);
562                 }
563         }
564
565         /*
566          * Do not use kmem_cache_zalloc(), as this cache uses
567          * SLAB_DESTROY_BY_RCU.
568          */
569         ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
570         if (ct == NULL) {
571                 pr_debug("nf_conntrack_alloc: Can't alloc conntrack.\n");
572                 atomic_dec(&net->ct.count);
573                 return ERR_PTR(-ENOMEM);
574         }
575         /*
576          * Let ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.next
577          * and ct->tuplehash[IP_CT_DIR_REPLY].hnnode.next unchanged.
578          */
579         memset(&ct->tuplehash[IP_CT_DIR_MAX], 0,
580                sizeof(*ct) - offsetof(struct nf_conn, tuplehash[IP_CT_DIR_MAX]));
581         spin_lock_init(&ct->lock);
582         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
583         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
584         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
585         ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev = NULL;
586         /* Don't set timer yet: wait for confirmation */
587         setup_timer(&ct->timeout, death_by_timeout, (unsigned long)ct);
588 #ifdef CONFIG_NET_NS
589         ct->ct_net = net;
590 #endif
591
592         /*
593          * changes to lookup keys must be done before setting refcnt to 1
594          */
595         smp_wmb();
596         atomic_set(&ct->ct_general.use, 1);
597         return ct;
598 }
599 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
600
601 void nf_conntrack_free(struct nf_conn *ct)
602 {
603         struct net *net = nf_ct_net(ct);
604
605         nf_ct_ext_destroy(ct);
606         atomic_dec(&net->ct.count);
607         nf_ct_ext_free(ct);
608         kmem_cache_free(nf_conntrack_cachep, ct);
609 }
610 EXPORT_SYMBOL_GPL(nf_conntrack_free);
611
612 /* Allocate a new conntrack: we return -ENOMEM if classification
613    failed due to stress.  Otherwise it really is unclassifiable. */
614 static struct nf_conntrack_tuple_hash *
615 init_conntrack(struct net *net,
616                const struct nf_conntrack_tuple *tuple,
617                struct nf_conntrack_l3proto *l3proto,
618                struct nf_conntrack_l4proto *l4proto,
619                struct sk_buff *skb,
620                unsigned int dataoff)
621 {
622         struct nf_conn *ct;
623         struct nf_conn_help *help;
624         struct nf_conntrack_tuple repl_tuple;
625         struct nf_conntrack_expect *exp;
626
627         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
628                 pr_debug("Can't invert tuple.\n");
629                 return NULL;
630         }
631
632         ct = nf_conntrack_alloc(net, tuple, &repl_tuple, GFP_ATOMIC);
633         if (IS_ERR(ct)) {
634                 pr_debug("Can't allocate conntrack.\n");
635                 return (struct nf_conntrack_tuple_hash *)ct;
636         }
637
638         if (!l4proto->new(ct, skb, dataoff)) {
639                 nf_conntrack_free(ct);
640                 pr_debug("init conntrack: can't track with proto module\n");
641                 return NULL;
642         }
643
644         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
645         nf_ct_ecache_ext_add(ct, GFP_ATOMIC);
646
647         spin_lock_bh(&nf_conntrack_lock);
648         exp = nf_ct_find_expectation(net, tuple);
649         if (exp) {
650                 pr_debug("conntrack: expectation arrives ct=%p exp=%p\n",
651                          ct, exp);
652                 /* Welcome, Mr. Bond.  We've been expecting you... */
653                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
654                 ct->master = exp->master;
655                 if (exp->helper) {
656                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
657                         if (help)
658                                 rcu_assign_pointer(help->helper, exp->helper);
659                 }
660
661 #ifdef CONFIG_NF_CONNTRACK_MARK
662                 ct->mark = exp->master->mark;
663 #endif
664 #ifdef CONFIG_NF_CONNTRACK_SECMARK
665                 ct->secmark = exp->master->secmark;
666 #endif
667                 nf_conntrack_get(&ct->master->ct_general);
668                 NF_CT_STAT_INC(net, expect_new);
669         } else {
670                 __nf_ct_try_assign_helper(ct, GFP_ATOMIC);
671                 NF_CT_STAT_INC(net, new);
672         }
673
674         /* Overload tuple linked list to put us in unconfirmed list. */
675         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
676                        &net->ct.unconfirmed);
677
678         spin_unlock_bh(&nf_conntrack_lock);
679
680         if (exp) {
681                 if (exp->expectfn)
682                         exp->expectfn(ct, exp);
683                 nf_ct_expect_put(exp);
684         }
685
686         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
687 }
688
689 /* On success, returns conntrack ptr, sets skb->nfct and ctinfo */
690 static inline struct nf_conn *
691 resolve_normal_ct(struct net *net,
692                   struct sk_buff *skb,
693                   unsigned int dataoff,
694                   u_int16_t l3num,
695                   u_int8_t protonum,
696                   struct nf_conntrack_l3proto *l3proto,
697                   struct nf_conntrack_l4proto *l4proto,
698                   int *set_reply,
699                   enum ip_conntrack_info *ctinfo)
700 {
701         struct nf_conntrack_tuple tuple;
702         struct nf_conntrack_tuple_hash *h;
703         struct nf_conn *ct;
704
705         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
706                              dataoff, l3num, protonum, &tuple, l3proto,
707                              l4proto)) {
708                 pr_debug("resolve_normal_ct: Can't get tuple\n");
709                 return NULL;
710         }
711
712         /* look for tuple match */
713         h = nf_conntrack_find_get(net, &tuple);
714         if (!h) {
715                 h = init_conntrack(net, &tuple, l3proto, l4proto, skb, dataoff);
716                 if (!h)
717                         return NULL;
718                 if (IS_ERR(h))
719                         return (void *)h;
720         }
721         ct = nf_ct_tuplehash_to_ctrack(h);
722
723         /* It exists; we have (non-exclusive) reference. */
724         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
725                 *ctinfo = IP_CT_ESTABLISHED + IP_CT_IS_REPLY;
726                 /* Please set reply bit if this packet OK */
727                 *set_reply = 1;
728         } else {
729                 /* Once we've had two way comms, always ESTABLISHED. */
730                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
731                         pr_debug("nf_conntrack_in: normal packet for %p\n", ct);
732                         *ctinfo = IP_CT_ESTABLISHED;
733                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
734                         pr_debug("nf_conntrack_in: related packet for %p\n",
735                                  ct);
736                         *ctinfo = IP_CT_RELATED;
737                 } else {
738                         pr_debug("nf_conntrack_in: new packet for %p\n", ct);
739                         *ctinfo = IP_CT_NEW;
740                 }
741                 *set_reply = 0;
742         }
743         skb->nfct = &ct->ct_general;
744         skb->nfctinfo = *ctinfo;
745         return ct;
746 }
747
748 unsigned int
749 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
750                 struct sk_buff *skb)
751 {
752         struct nf_conn *ct;
753         enum ip_conntrack_info ctinfo;
754         struct nf_conntrack_l3proto *l3proto;
755         struct nf_conntrack_l4proto *l4proto;
756         unsigned int dataoff;
757         u_int8_t protonum;
758         int set_reply = 0;
759         int ret;
760
761         /* Previously seen (loopback or untracked)?  Ignore. */
762         if (skb->nfct) {
763                 NF_CT_STAT_INC_ATOMIC(net, ignore);
764                 return NF_ACCEPT;
765         }
766
767         /* rcu_read_lock()ed by nf_hook_slow */
768         l3proto = __nf_ct_l3proto_find(pf);
769         ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
770                                    &dataoff, &protonum);
771         if (ret <= 0) {
772                 pr_debug("not prepared to track yet or error occured\n");
773                 NF_CT_STAT_INC_ATOMIC(net, error);
774                 NF_CT_STAT_INC_ATOMIC(net, invalid);
775                 return -ret;
776         }
777
778         l4proto = __nf_ct_l4proto_find(pf, protonum);
779
780         /* It may be an special packet, error, unclean...
781          * inverse of the return code tells to the netfilter
782          * core what to do with the packet. */
783         if (l4proto->error != NULL) {
784                 ret = l4proto->error(net, skb, dataoff, &ctinfo, pf, hooknum);
785                 if (ret <= 0) {
786                         NF_CT_STAT_INC_ATOMIC(net, error);
787                         NF_CT_STAT_INC_ATOMIC(net, invalid);
788                         return -ret;
789                 }
790         }
791
792         ct = resolve_normal_ct(net, skb, dataoff, pf, protonum,
793                                l3proto, l4proto, &set_reply, &ctinfo);
794         if (!ct) {
795                 /* Not valid part of a connection */
796                 NF_CT_STAT_INC_ATOMIC(net, invalid);
797                 return NF_ACCEPT;
798         }
799
800         if (IS_ERR(ct)) {
801                 /* Too stressed to deal. */
802                 NF_CT_STAT_INC_ATOMIC(net, drop);
803                 return NF_DROP;
804         }
805
806         NF_CT_ASSERT(skb->nfct);
807
808         ret = l4proto->packet(ct, skb, dataoff, ctinfo, pf, hooknum);
809         if (ret <= 0) {
810                 /* Invalid: inverse of the return code tells
811                  * the netfilter core what to do */
812                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
813                 nf_conntrack_put(skb->nfct);
814                 skb->nfct = NULL;
815                 NF_CT_STAT_INC_ATOMIC(net, invalid);
816                 if (ret == -NF_DROP)
817                         NF_CT_STAT_INC_ATOMIC(net, drop);
818                 return -ret;
819         }
820
821         if (set_reply && !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
822                 nf_conntrack_event_cache(IPCT_STATUS, ct);
823
824         return ret;
825 }
826 EXPORT_SYMBOL_GPL(nf_conntrack_in);
827
828 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
829                           const struct nf_conntrack_tuple *orig)
830 {
831         bool ret;
832
833         rcu_read_lock();
834         ret = nf_ct_invert_tuple(inverse, orig,
835                                  __nf_ct_l3proto_find(orig->src.l3num),
836                                  __nf_ct_l4proto_find(orig->src.l3num,
837                                                       orig->dst.protonum));
838         rcu_read_unlock();
839         return ret;
840 }
841 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
842
843 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
844    implicitly racy: see __nf_conntrack_confirm */
845 void nf_conntrack_alter_reply(struct nf_conn *ct,
846                               const struct nf_conntrack_tuple *newreply)
847 {
848         struct nf_conn_help *help = nfct_help(ct);
849
850         /* Should be unconfirmed, so not in hash table yet */
851         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
852
853         pr_debug("Altering reply tuple of %p to ", ct);
854         nf_ct_dump_tuple(newreply);
855
856         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
857         if (ct->master || (help && !hlist_empty(&help->expectations)))
858                 return;
859
860         rcu_read_lock();
861         __nf_ct_try_assign_helper(ct, GFP_ATOMIC);
862         rcu_read_unlock();
863 }
864 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
865
866 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
867 void __nf_ct_refresh_acct(struct nf_conn *ct,
868                           enum ip_conntrack_info ctinfo,
869                           const struct sk_buff *skb,
870                           unsigned long extra_jiffies,
871                           int do_acct)
872 {
873         NF_CT_ASSERT(ct->timeout.data == (unsigned long)ct);
874         NF_CT_ASSERT(skb);
875
876         /* Only update if this is not a fixed timeout */
877         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
878                 goto acct;
879
880         /* If not in hash table, timer will not be active yet */
881         if (!nf_ct_is_confirmed(ct)) {
882                 ct->timeout.expires = extra_jiffies;
883         } else {
884                 unsigned long newtime = jiffies + extra_jiffies;
885
886                 /* Only update the timeout if the new timeout is at least
887                    HZ jiffies from the old timeout. Need del_timer for race
888                    avoidance (may already be dying). */
889                 if (newtime - ct->timeout.expires >= HZ)
890                         mod_timer_pending(&ct->timeout, newtime);
891         }
892
893 acct:
894         if (do_acct) {
895                 struct nf_conn_counter *acct;
896
897                 acct = nf_conn_acct_find(ct);
898                 if (acct) {
899                         spin_lock_bh(&ct->lock);
900                         acct[CTINFO2DIR(ctinfo)].packets++;
901                         acct[CTINFO2DIR(ctinfo)].bytes +=
902                                 skb->len - skb_network_offset(skb);
903                         spin_unlock_bh(&ct->lock);
904                 }
905         }
906 }
907 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
908
909 bool __nf_ct_kill_acct(struct nf_conn *ct,
910                        enum ip_conntrack_info ctinfo,
911                        const struct sk_buff *skb,
912                        int do_acct)
913 {
914         if (do_acct) {
915                 struct nf_conn_counter *acct;
916
917                 acct = nf_conn_acct_find(ct);
918                 if (acct) {
919                         spin_lock_bh(&ct->lock);
920                         acct[CTINFO2DIR(ctinfo)].packets++;
921                         acct[CTINFO2DIR(ctinfo)].bytes +=
922                                 skb->len - skb_network_offset(skb);
923                         spin_unlock_bh(&ct->lock);
924                 }
925         }
926
927         if (del_timer(&ct->timeout)) {
928                 ct->timeout.function((unsigned long)ct);
929                 return true;
930         }
931         return false;
932 }
933 EXPORT_SYMBOL_GPL(__nf_ct_kill_acct);
934
935 #if defined(CONFIG_NF_CT_NETLINK) || defined(CONFIG_NF_CT_NETLINK_MODULE)
936
937 #include <linux/netfilter/nfnetlink.h>
938 #include <linux/netfilter/nfnetlink_conntrack.h>
939 #include <linux/mutex.h>
940
941 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
942  * in ip_conntrack_core, since we don't want the protocols to autoload
943  * or depend on ctnetlink */
944 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
945                                const struct nf_conntrack_tuple *tuple)
946 {
947         NLA_PUT_BE16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port);
948         NLA_PUT_BE16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port);
949         return 0;
950
951 nla_put_failure:
952         return -1;
953 }
954 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
955
956 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
957         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
958         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
959 };
960 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
961
962 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
963                                struct nf_conntrack_tuple *t)
964 {
965         if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
966                 return -EINVAL;
967
968         t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
969         t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
970
971         return 0;
972 }
973 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
974
975 int nf_ct_port_nlattr_tuple_size(void)
976 {
977         return nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
978 }
979 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
980 #endif
981
982 /* Used by ipt_REJECT and ip6t_REJECT. */
983 static void nf_conntrack_attach(struct sk_buff *nskb, struct sk_buff *skb)
984 {
985         struct nf_conn *ct;
986         enum ip_conntrack_info ctinfo;
987
988         /* This ICMP is in reverse direction to the packet which caused it */
989         ct = nf_ct_get(skb, &ctinfo);
990         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
991                 ctinfo = IP_CT_RELATED + IP_CT_IS_REPLY;
992         else
993                 ctinfo = IP_CT_RELATED;
994
995         /* Attach to new skbuff, and increment count */
996         nskb->nfct = &ct->ct_general;
997         nskb->nfctinfo = ctinfo;
998         nf_conntrack_get(nskb->nfct);
999 }
1000
1001 /* Bring out ya dead! */
1002 static struct nf_conn *
1003 get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
1004                 void *data, unsigned int *bucket)
1005 {
1006         struct nf_conntrack_tuple_hash *h;
1007         struct nf_conn *ct;
1008         struct hlist_nulls_node *n;
1009
1010         spin_lock_bh(&nf_conntrack_lock);
1011         for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
1012                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[*bucket], hnnode) {
1013                         ct = nf_ct_tuplehash_to_ctrack(h);
1014                         if (iter(ct, data))
1015                                 goto found;
1016                 }
1017         }
1018         hlist_nulls_for_each_entry(h, n, &net->ct.unconfirmed, hnnode) {
1019                 ct = nf_ct_tuplehash_to_ctrack(h);
1020                 if (iter(ct, data))
1021                         set_bit(IPS_DYING_BIT, &ct->status);
1022         }
1023         spin_unlock_bh(&nf_conntrack_lock);
1024         return NULL;
1025 found:
1026         atomic_inc(&ct->ct_general.use);
1027         spin_unlock_bh(&nf_conntrack_lock);
1028         return ct;
1029 }
1030
1031 void nf_ct_iterate_cleanup(struct net *net,
1032                            int (*iter)(struct nf_conn *i, void *data),
1033                            void *data)
1034 {
1035         struct nf_conn *ct;
1036         unsigned int bucket = 0;
1037
1038         while ((ct = get_next_corpse(net, iter, data, &bucket)) != NULL) {
1039                 /* Time to push up daises... */
1040                 if (del_timer(&ct->timeout))
1041                         death_by_timeout((unsigned long)ct);
1042                 /* ... else the timer will get him soon. */
1043
1044                 nf_ct_put(ct);
1045         }
1046 }
1047 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);
1048
1049 struct __nf_ct_flush_report {
1050         u32 pid;
1051         int report;
1052 };
1053
1054 static int kill_report(struct nf_conn *i, void *data)
1055 {
1056         struct __nf_ct_flush_report *fr = (struct __nf_ct_flush_report *)data;
1057
1058         /* If we fail to deliver the event, death_by_timeout() will retry */
1059         if (nf_conntrack_event_report(IPCT_DESTROY, i,
1060                                       fr->pid, fr->report) < 0)
1061                 return 1;
1062
1063         /* Avoid the delivery of the destroy event in death_by_timeout(). */
1064         set_bit(IPS_DYING_BIT, &i->status);
1065         return 1;
1066 }
1067
1068 static int kill_all(struct nf_conn *i, void *data)
1069 {
1070         return 1;
1071 }
1072
1073 void nf_ct_free_hashtable(void *hash, int vmalloced, unsigned int size)
1074 {
1075         if (vmalloced)
1076                 vfree(hash);
1077         else
1078                 free_pages((unsigned long)hash,
1079                            get_order(sizeof(struct hlist_head) * size));
1080 }
1081 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1082
1083 void nf_conntrack_flush_report(struct net *net, u32 pid, int report)
1084 {
1085         struct __nf_ct_flush_report fr = {
1086                 .pid    = pid,
1087                 .report = report,
1088         };
1089         nf_ct_iterate_cleanup(net, kill_report, &fr);
1090 }
1091 EXPORT_SYMBOL_GPL(nf_conntrack_flush_report);
1092
1093 static void nf_ct_release_dying_list(struct net *net)
1094 {
1095         struct nf_conntrack_tuple_hash *h;
1096         struct nf_conn *ct;
1097         struct hlist_nulls_node *n;
1098
1099         spin_lock_bh(&nf_conntrack_lock);
1100         hlist_nulls_for_each_entry(h, n, &net->ct.dying, hnnode) {
1101                 ct = nf_ct_tuplehash_to_ctrack(h);
1102                 /* never fails to remove them, no listeners at this point */
1103                 nf_ct_kill(ct);
1104         }
1105         spin_unlock_bh(&nf_conntrack_lock);
1106 }
1107
1108 static void nf_conntrack_cleanup_init_net(void)
1109 {
1110         nf_conntrack_helper_fini();
1111         nf_conntrack_proto_fini();
1112         kmem_cache_destroy(nf_conntrack_cachep);
1113 }
1114
1115 static void nf_conntrack_cleanup_net(struct net *net)
1116 {
1117  i_see_dead_people:
1118         nf_ct_iterate_cleanup(net, kill_all, NULL);
1119         nf_ct_release_dying_list(net);
1120         if (atomic_read(&net->ct.count) != 0) {
1121                 schedule();
1122                 goto i_see_dead_people;
1123         }
1124         /* wait until all references to nf_conntrack_untracked are dropped */
1125         while (atomic_read(&nf_conntrack_untracked.ct_general.use) > 1)
1126                 schedule();
1127
1128         nf_ct_free_hashtable(net->ct.hash, net->ct.hash_vmalloc,
1129                              nf_conntrack_htable_size);
1130         nf_conntrack_ecache_fini(net);
1131         nf_conntrack_acct_fini(net);
1132         nf_conntrack_expect_fini(net);
1133         free_percpu(net->ct.stat);
1134 }
1135
1136 /* Mishearing the voices in his head, our hero wonders how he's
1137    supposed to kill the mall. */
1138 void nf_conntrack_cleanup(struct net *net)
1139 {
1140         if (net_eq(net, &init_net))
1141                 rcu_assign_pointer(ip_ct_attach, NULL);
1142
1143         /* This makes sure all current packets have passed through
1144            netfilter framework.  Roll on, two-stage module
1145            delete... */
1146         synchronize_net();
1147
1148         nf_conntrack_cleanup_net(net);
1149
1150         if (net_eq(net, &init_net)) {
1151                 rcu_assign_pointer(nf_ct_destroy, NULL);
1152                 nf_conntrack_cleanup_init_net();
1153         }
1154 }
1155
1156 void *nf_ct_alloc_hashtable(unsigned int *sizep, int *vmalloced, int nulls)
1157 {
1158         struct hlist_nulls_head *hash;
1159         unsigned int nr_slots, i;
1160         size_t sz;
1161
1162         *vmalloced = 0;
1163
1164         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1165         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1166         sz = nr_slots * sizeof(struct hlist_nulls_head);
1167         hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1168                                         get_order(sz));
1169         if (!hash) {
1170                 *vmalloced = 1;
1171                 printk(KERN_WARNING "nf_conntrack: falling back to vmalloc.\n");
1172                 hash = __vmalloc(sz, GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL);
1173         }
1174
1175         if (hash && nulls)
1176                 for (i = 0; i < nr_slots; i++)
1177                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
1178
1179         return hash;
1180 }
1181 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1182
1183 int nf_conntrack_set_hashsize(const char *val, struct kernel_param *kp)
1184 {
1185         int i, bucket, vmalloced, old_vmalloced;
1186         unsigned int hashsize, old_size;
1187         int rnd;
1188         struct hlist_nulls_head *hash, *old_hash;
1189         struct nf_conntrack_tuple_hash *h;
1190
1191         /* On boot, we can set this without any fancy locking. */
1192         if (!nf_conntrack_htable_size)
1193                 return param_set_uint(val, kp);
1194
1195         hashsize = simple_strtoul(val, NULL, 0);
1196         if (!hashsize)
1197                 return -EINVAL;
1198
1199         hash = nf_ct_alloc_hashtable(&hashsize, &vmalloced, 1);
1200         if (!hash)
1201                 return -ENOMEM;
1202
1203         /* We have to rehahs for the new table anyway, so we also can
1204          * use a newrandom seed */
1205         get_random_bytes(&rnd, sizeof(rnd));
1206
1207         /* Lookups in the old hash might happen in parallel, which means we
1208          * might get false negatives during connection lookup. New connections
1209          * created because of a false negative won't make it into the hash
1210          * though since that required taking the lock.
1211          */
1212         spin_lock_bh(&nf_conntrack_lock);
1213         for (i = 0; i < nf_conntrack_htable_size; i++) {
1214                 while (!hlist_nulls_empty(&init_net.ct.hash[i])) {
1215                         h = hlist_nulls_entry(init_net.ct.hash[i].first,
1216                                         struct nf_conntrack_tuple_hash, hnnode);
1217                         hlist_nulls_del_rcu(&h->hnnode);
1218                         bucket = __hash_conntrack(&h->tuple, hashsize, rnd);
1219                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
1220                 }
1221         }
1222         old_size = nf_conntrack_htable_size;
1223         old_vmalloced = init_net.ct.hash_vmalloc;
1224         old_hash = init_net.ct.hash;
1225
1226         nf_conntrack_htable_size = hashsize;
1227         init_net.ct.hash_vmalloc = vmalloced;
1228         init_net.ct.hash = hash;
1229         nf_conntrack_hash_rnd = rnd;
1230         spin_unlock_bh(&nf_conntrack_lock);
1231
1232         nf_ct_free_hashtable(old_hash, old_vmalloced, old_size);
1233         return 0;
1234 }
1235 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
1236
1237 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
1238                   &nf_conntrack_htable_size, 0600);
1239
1240 static int nf_conntrack_init_init_net(void)
1241 {
1242         int max_factor = 8;
1243         int ret;
1244
1245         /* Idea from tcp.c: use 1/16384 of memory.  On i386: 32MB
1246          * machine has 512 buckets. >= 1GB machines have 16384 buckets. */
1247         if (!nf_conntrack_htable_size) {
1248                 nf_conntrack_htable_size
1249                         = (((totalram_pages << PAGE_SHIFT) / 16384)
1250                            / sizeof(struct hlist_head));
1251                 if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
1252                         nf_conntrack_htable_size = 16384;
1253                 if (nf_conntrack_htable_size < 32)
1254                         nf_conntrack_htable_size = 32;
1255
1256                 /* Use a max. factor of four by default to get the same max as
1257                  * with the old struct list_heads. When a table size is given
1258                  * we use the old value of 8 to avoid reducing the max.
1259                  * entries. */
1260                 max_factor = 4;
1261         }
1262         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
1263
1264         printk("nf_conntrack version %s (%u buckets, %d max)\n",
1265                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
1266                nf_conntrack_max);
1267
1268         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
1269                                                 sizeof(struct nf_conn),
1270                                                 0, SLAB_DESTROY_BY_RCU, NULL);
1271         if (!nf_conntrack_cachep) {
1272                 printk(KERN_ERR "Unable to create nf_conn slab cache\n");
1273                 ret = -ENOMEM;
1274                 goto err_cache;
1275         }
1276
1277         ret = nf_conntrack_proto_init();
1278         if (ret < 0)
1279                 goto err_proto;
1280
1281         ret = nf_conntrack_helper_init();
1282         if (ret < 0)
1283                 goto err_helper;
1284
1285         return 0;
1286
1287 err_helper:
1288         nf_conntrack_proto_fini();
1289 err_proto:
1290         kmem_cache_destroy(nf_conntrack_cachep);
1291 err_cache:
1292         return ret;
1293 }
1294
1295 /*
1296  * We need to use special "null" values, not used in hash table
1297  */
1298 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
1299 #define DYING_NULLS_VAL         ((1<<30)+1)
1300
1301 static int nf_conntrack_init_net(struct net *net)
1302 {
1303         int ret;
1304
1305         atomic_set(&net->ct.count, 0);
1306         INIT_HLIST_NULLS_HEAD(&net->ct.unconfirmed, UNCONFIRMED_NULLS_VAL);
1307         INIT_HLIST_NULLS_HEAD(&net->ct.dying, DYING_NULLS_VAL);
1308         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
1309         if (!net->ct.stat) {
1310                 ret = -ENOMEM;
1311                 goto err_stat;
1312         }
1313         net->ct.hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size,
1314                                              &net->ct.hash_vmalloc, 1);
1315         if (!net->ct.hash) {
1316                 ret = -ENOMEM;
1317                 printk(KERN_ERR "Unable to create nf_conntrack_hash\n");
1318                 goto err_hash;
1319         }
1320         ret = nf_conntrack_expect_init(net);
1321         if (ret < 0)
1322                 goto err_expect;
1323         ret = nf_conntrack_acct_init(net);
1324         if (ret < 0)
1325                 goto err_acct;
1326         ret = nf_conntrack_ecache_init(net);
1327         if (ret < 0)
1328                 goto err_ecache;
1329
1330         /* Set up fake conntrack:
1331             - to never be deleted, not in any hashes */
1332 #ifdef CONFIG_NET_NS
1333         nf_conntrack_untracked.ct_net = &init_net;
1334 #endif
1335         atomic_set(&nf_conntrack_untracked.ct_general.use, 1);
1336         /*  - and look it like as a confirmed connection */
1337         set_bit(IPS_CONFIRMED_BIT, &nf_conntrack_untracked.status);
1338
1339         return 0;
1340
1341 err_ecache:
1342         nf_conntrack_acct_fini(net);
1343 err_acct:
1344         nf_conntrack_expect_fini(net);
1345 err_expect:
1346         nf_ct_free_hashtable(net->ct.hash, net->ct.hash_vmalloc,
1347                              nf_conntrack_htable_size);
1348 err_hash:
1349         free_percpu(net->ct.stat);
1350 err_stat:
1351         return ret;
1352 }
1353
1354 s16 (*nf_ct_nat_offset)(const struct nf_conn *ct,
1355                         enum ip_conntrack_dir dir,
1356                         u32 seq);
1357 EXPORT_SYMBOL_GPL(nf_ct_nat_offset);
1358
1359 int nf_conntrack_init(struct net *net)
1360 {
1361         int ret;
1362
1363         if (net_eq(net, &init_net)) {
1364                 ret = nf_conntrack_init_init_net();
1365                 if (ret < 0)
1366                         goto out_init_net;
1367         }
1368         ret = nf_conntrack_init_net(net);
1369         if (ret < 0)
1370                 goto out_net;
1371
1372         if (net_eq(net, &init_net)) {
1373                 /* For use by REJECT target */
1374                 rcu_assign_pointer(ip_ct_attach, nf_conntrack_attach);
1375                 rcu_assign_pointer(nf_ct_destroy, destroy_conntrack);
1376
1377                 /* Howto get NAT offsets */
1378                 rcu_assign_pointer(nf_ct_nat_offset, NULL);
1379         }
1380         return 0;
1381
1382 out_net:
1383         if (net_eq(net, &init_net))
1384                 nf_conntrack_cleanup_init_net();
1385 out_init_net:
1386         return ret;
1387 }