mac80211: fix another race in aggregation start
[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/nsproxy.h>
34 #include <linux/rculist_nulls.h>
35
36 #include <net/netfilter/nf_conntrack.h>
37 #include <net/netfilter/nf_conntrack_l3proto.h>
38 #include <net/netfilter/nf_conntrack_l4proto.h>
39 #include <net/netfilter/nf_conntrack_expect.h>
40 #include <net/netfilter/nf_conntrack_helper.h>
41 #include <net/netfilter/nf_conntrack_core.h>
42 #include <net/netfilter/nf_conntrack_extend.h>
43 #include <net/netfilter/nf_conntrack_acct.h>
44 #include <net/netfilter/nf_conntrack_ecache.h>
45 #include <net/netfilter/nf_conntrack_zones.h>
46 #include <net/netfilter/nf_conntrack_timestamp.h>
47 #include <net/netfilter/nf_nat.h>
48 #include <net/netfilter/nf_nat_core.h>
49
50 #define NF_CONNTRACK_VERSION    "0.5.0"
51
52 int (*nfnetlink_parse_nat_setup_hook)(struct nf_conn *ct,
53                                       enum nf_nat_manip_type manip,
54                                       const struct nlattr *attr) __read_mostly;
55 EXPORT_SYMBOL_GPL(nfnetlink_parse_nat_setup_hook);
56
57 DEFINE_SPINLOCK(nf_conntrack_lock);
58 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
59
60 unsigned int nf_conntrack_htable_size __read_mostly;
61 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
62
63 unsigned int nf_conntrack_max __read_mostly;
64 EXPORT_SYMBOL_GPL(nf_conntrack_max);
65
66 DEFINE_PER_CPU(struct nf_conn, nf_conntrack_untracked);
67 EXPORT_PER_CPU_SYMBOL(nf_conntrack_untracked);
68
69 unsigned int nf_conntrack_hash_rnd __read_mostly;
70
71 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple, u16 zone)
72 {
73         unsigned int n;
74
75         /* The direction must be ignored, so we hash everything up to the
76          * destination ports (which is a multiple of 4) and treat the last
77          * three bytes manually.
78          */
79         n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
80         return jhash2((u32 *)tuple, n, zone ^ nf_conntrack_hash_rnd ^
81                       (((__force __u16)tuple->dst.u.all << 16) |
82                       tuple->dst.protonum));
83 }
84
85 static u32 __hash_bucket(u32 hash, unsigned int size)
86 {
87         return ((u64)hash * size) >> 32;
88 }
89
90 static u32 hash_bucket(u32 hash, const struct net *net)
91 {
92         return __hash_bucket(hash, net->ct.htable_size);
93 }
94
95 static u_int32_t __hash_conntrack(const struct nf_conntrack_tuple *tuple,
96                                   u16 zone, unsigned int size)
97 {
98         return __hash_bucket(hash_conntrack_raw(tuple, zone), size);
99 }
100
101 static inline u_int32_t hash_conntrack(const struct net *net, u16 zone,
102                                        const struct nf_conntrack_tuple *tuple)
103 {
104         return __hash_conntrack(tuple, zone, net->ct.htable_size);
105 }
106
107 bool
108 nf_ct_get_tuple(const struct sk_buff *skb,
109                 unsigned int nhoff,
110                 unsigned int dataoff,
111                 u_int16_t l3num,
112                 u_int8_t protonum,
113                 struct nf_conntrack_tuple *tuple,
114                 const struct nf_conntrack_l3proto *l3proto,
115                 const struct nf_conntrack_l4proto *l4proto)
116 {
117         memset(tuple, 0, sizeof(*tuple));
118
119         tuple->src.l3num = l3num;
120         if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
121                 return false;
122
123         tuple->dst.protonum = protonum;
124         tuple->dst.dir = IP_CT_DIR_ORIGINAL;
125
126         return l4proto->pkt_to_tuple(skb, dataoff, tuple);
127 }
128 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
129
130 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
131                        u_int16_t l3num, struct nf_conntrack_tuple *tuple)
132 {
133         struct nf_conntrack_l3proto *l3proto;
134         struct nf_conntrack_l4proto *l4proto;
135         unsigned int protoff;
136         u_int8_t protonum;
137         int ret;
138
139         rcu_read_lock();
140
141         l3proto = __nf_ct_l3proto_find(l3num);
142         ret = l3proto->get_l4proto(skb, nhoff, &protoff, &protonum);
143         if (ret != NF_ACCEPT) {
144                 rcu_read_unlock();
145                 return false;
146         }
147
148         l4proto = __nf_ct_l4proto_find(l3num, protonum);
149
150         ret = nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, tuple,
151                               l3proto, l4proto);
152
153         rcu_read_unlock();
154         return ret;
155 }
156 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
157
158 bool
159 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
160                    const struct nf_conntrack_tuple *orig,
161                    const struct nf_conntrack_l3proto *l3proto,
162                    const struct nf_conntrack_l4proto *l4proto)
163 {
164         memset(inverse, 0, sizeof(*inverse));
165
166         inverse->src.l3num = orig->src.l3num;
167         if (l3proto->invert_tuple(inverse, orig) == 0)
168                 return false;
169
170         inverse->dst.dir = !orig->dst.dir;
171
172         inverse->dst.protonum = orig->dst.protonum;
173         return l4proto->invert_tuple(inverse, orig);
174 }
175 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
176
177 static void
178 clean_from_lists(struct nf_conn *ct)
179 {
180         pr_debug("clean_from_lists(%p)\n", ct);
181         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
182         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
183
184         /* Destroy all pending expectations */
185         nf_ct_remove_expectations(ct);
186 }
187
188 static void
189 destroy_conntrack(struct nf_conntrack *nfct)
190 {
191         struct nf_conn *ct = (struct nf_conn *)nfct;
192         struct net *net = nf_ct_net(ct);
193         struct nf_conntrack_l4proto *l4proto;
194
195         pr_debug("destroy_conntrack(%p)\n", ct);
196         NF_CT_ASSERT(atomic_read(&nfct->use) == 0);
197         NF_CT_ASSERT(!timer_pending(&ct->timeout));
198
199         /* To make sure we don't get any weird locking issues here:
200          * destroy_conntrack() MUST NOT be called with a write lock
201          * to nf_conntrack_lock!!! -HW */
202         rcu_read_lock();
203         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
204         if (l4proto && l4proto->destroy)
205                 l4proto->destroy(ct);
206
207         rcu_read_unlock();
208
209         spin_lock_bh(&nf_conntrack_lock);
210         /* Expectations will have been removed in clean_from_lists,
211          * except TFTP can create an expectation on the first packet,
212          * before connection is in the list, so we need to clean here,
213          * too. */
214         nf_ct_remove_expectations(ct);
215
216         /* We overload first tuple to link into unconfirmed list. */
217         if (!nf_ct_is_confirmed(ct)) {
218                 BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
219                 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
220         }
221
222         NF_CT_STAT_INC(net, delete);
223         spin_unlock_bh(&nf_conntrack_lock);
224
225         if (ct->master)
226                 nf_ct_put(ct->master);
227
228         pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
229         nf_conntrack_free(ct);
230 }
231
232 void nf_ct_delete_from_lists(struct nf_conn *ct)
233 {
234         struct net *net = nf_ct_net(ct);
235
236         nf_ct_helper_destroy(ct);
237         spin_lock_bh(&nf_conntrack_lock);
238         /* Inside lock so preempt is disabled on module removal path.
239          * Otherwise we can get spurious warnings. */
240         NF_CT_STAT_INC(net, delete_list);
241         clean_from_lists(ct);
242         spin_unlock_bh(&nf_conntrack_lock);
243 }
244 EXPORT_SYMBOL_GPL(nf_ct_delete_from_lists);
245
246 static void death_by_event(unsigned long ul_conntrack)
247 {
248         struct nf_conn *ct = (void *)ul_conntrack;
249         struct net *net = nf_ct_net(ct);
250
251         if (nf_conntrack_event(IPCT_DESTROY, ct) < 0) {
252                 /* bad luck, let's retry again */
253                 ct->timeout.expires = jiffies +
254                         (random32() % net->ct.sysctl_events_retry_timeout);
255                 add_timer(&ct->timeout);
256                 return;
257         }
258         /* we've got the event delivered, now it's dying */
259         set_bit(IPS_DYING_BIT, &ct->status);
260         spin_lock(&nf_conntrack_lock);
261         hlist_nulls_del(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
262         spin_unlock(&nf_conntrack_lock);
263         nf_ct_put(ct);
264 }
265
266 void nf_ct_insert_dying_list(struct nf_conn *ct)
267 {
268         struct net *net = nf_ct_net(ct);
269
270         /* add this conntrack to the dying list */
271         spin_lock_bh(&nf_conntrack_lock);
272         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
273                              &net->ct.dying);
274         spin_unlock_bh(&nf_conntrack_lock);
275         /* set a new timer to retry event delivery */
276         setup_timer(&ct->timeout, death_by_event, (unsigned long)ct);
277         ct->timeout.expires = jiffies +
278                 (random32() % net->ct.sysctl_events_retry_timeout);
279         add_timer(&ct->timeout);
280 }
281 EXPORT_SYMBOL_GPL(nf_ct_insert_dying_list);
282
283 static void death_by_timeout(unsigned long ul_conntrack)
284 {
285         struct nf_conn *ct = (void *)ul_conntrack;
286         struct nf_conn_tstamp *tstamp;
287
288         tstamp = nf_conn_tstamp_find(ct);
289         if (tstamp && tstamp->stop == 0)
290                 tstamp->stop = ktime_to_ns(ktime_get_real());
291
292         if (!test_bit(IPS_DYING_BIT, &ct->status) &&
293             unlikely(nf_conntrack_event(IPCT_DESTROY, ct) < 0)) {
294                 /* destroy event was not delivered */
295                 nf_ct_delete_from_lists(ct);
296                 nf_ct_insert_dying_list(ct);
297                 return;
298         }
299         set_bit(IPS_DYING_BIT, &ct->status);
300         nf_ct_delete_from_lists(ct);
301         nf_ct_put(ct);
302 }
303
304 /*
305  * Warning :
306  * - Caller must take a reference on returned object
307  *   and recheck nf_ct_tuple_equal(tuple, &h->tuple)
308  * OR
309  * - Caller must lock nf_conntrack_lock before calling this function
310  */
311 static struct nf_conntrack_tuple_hash *
312 ____nf_conntrack_find(struct net *net, u16 zone,
313                       const struct nf_conntrack_tuple *tuple, u32 hash)
314 {
315         struct nf_conntrack_tuple_hash *h;
316         struct hlist_nulls_node *n;
317         unsigned int bucket = hash_bucket(hash, net);
318
319         /* Disable BHs the entire time since we normally need to disable them
320          * at least once for the stats anyway.
321          */
322         local_bh_disable();
323 begin:
324         hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[bucket], hnnode) {
325                 if (nf_ct_tuple_equal(tuple, &h->tuple) &&
326                     nf_ct_zone(nf_ct_tuplehash_to_ctrack(h)) == zone) {
327                         NF_CT_STAT_INC(net, found);
328                         local_bh_enable();
329                         return h;
330                 }
331                 NF_CT_STAT_INC(net, searched);
332         }
333         /*
334          * if the nulls value we got at the end of this lookup is
335          * not the expected one, we must restart lookup.
336          * We probably met an item that was moved to another chain.
337          */
338         if (get_nulls_value(n) != bucket) {
339                 NF_CT_STAT_INC(net, search_restart);
340                 goto begin;
341         }
342         local_bh_enable();
343
344         return NULL;
345 }
346
347 struct nf_conntrack_tuple_hash *
348 __nf_conntrack_find(struct net *net, u16 zone,
349                     const struct nf_conntrack_tuple *tuple)
350 {
351         return ____nf_conntrack_find(net, zone, tuple,
352                                      hash_conntrack_raw(tuple, zone));
353 }
354 EXPORT_SYMBOL_GPL(__nf_conntrack_find);
355
356 /* Find a connection corresponding to a tuple. */
357 static struct nf_conntrack_tuple_hash *
358 __nf_conntrack_find_get(struct net *net, u16 zone,
359                         const struct nf_conntrack_tuple *tuple, u32 hash)
360 {
361         struct nf_conntrack_tuple_hash *h;
362         struct nf_conn *ct;
363
364         rcu_read_lock();
365 begin:
366         h = ____nf_conntrack_find(net, zone, tuple, hash);
367         if (h) {
368                 ct = nf_ct_tuplehash_to_ctrack(h);
369                 if (unlikely(nf_ct_is_dying(ct) ||
370                              !atomic_inc_not_zero(&ct->ct_general.use)))
371                         h = NULL;
372                 else {
373                         if (unlikely(!nf_ct_tuple_equal(tuple, &h->tuple) ||
374                                      nf_ct_zone(ct) != zone)) {
375                                 nf_ct_put(ct);
376                                 goto begin;
377                         }
378                 }
379         }
380         rcu_read_unlock();
381
382         return h;
383 }
384
385 struct nf_conntrack_tuple_hash *
386 nf_conntrack_find_get(struct net *net, u16 zone,
387                       const struct nf_conntrack_tuple *tuple)
388 {
389         return __nf_conntrack_find_get(net, zone, tuple,
390                                        hash_conntrack_raw(tuple, zone));
391 }
392 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
393
394 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
395                                        unsigned int hash,
396                                        unsigned int repl_hash)
397 {
398         struct net *net = nf_ct_net(ct);
399
400         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
401                            &net->ct.hash[hash]);
402         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
403                            &net->ct.hash[repl_hash]);
404 }
405
406 void nf_conntrack_hash_insert(struct nf_conn *ct)
407 {
408         struct net *net = nf_ct_net(ct);
409         unsigned int hash, repl_hash;
410         u16 zone;
411
412         zone = nf_ct_zone(ct);
413         hash = hash_conntrack(net, zone, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
414         repl_hash = hash_conntrack(net, zone, &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
415
416         __nf_conntrack_hash_insert(ct, hash, repl_hash);
417 }
418 EXPORT_SYMBOL_GPL(nf_conntrack_hash_insert);
419
420 /* Confirm a connection given skb; places it in hash table */
421 int
422 __nf_conntrack_confirm(struct sk_buff *skb)
423 {
424         unsigned int hash, repl_hash;
425         struct nf_conntrack_tuple_hash *h;
426         struct nf_conn *ct;
427         struct nf_conn_help *help;
428         struct nf_conn_tstamp *tstamp;
429         struct hlist_nulls_node *n;
430         enum ip_conntrack_info ctinfo;
431         struct net *net;
432         u16 zone;
433
434         ct = nf_ct_get(skb, &ctinfo);
435         net = nf_ct_net(ct);
436
437         /* ipt_REJECT uses nf_conntrack_attach to attach related
438            ICMP/TCP RST packets in other direction.  Actual packet
439            which created connection will be IP_CT_NEW or for an
440            expected connection, IP_CT_RELATED. */
441         if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
442                 return NF_ACCEPT;
443
444         zone = nf_ct_zone(ct);
445         /* reuse the hash saved before */
446         hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
447         hash = hash_bucket(hash, net);
448         repl_hash = hash_conntrack(net, zone,
449                                    &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
450
451         /* We're not in hash table, and we refuse to set up related
452            connections for unconfirmed conns.  But packet copies and
453            REJECT will give spurious warnings here. */
454         /* NF_CT_ASSERT(atomic_read(&ct->ct_general.use) == 1); */
455
456         /* No external references means no one else could have
457            confirmed us. */
458         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
459         pr_debug("Confirming conntrack %p\n", ct);
460
461         spin_lock_bh(&nf_conntrack_lock);
462
463         /* We have to check the DYING flag inside the lock to prevent
464            a race against nf_ct_get_next_corpse() possibly called from
465            user context, else we insert an already 'dead' hash, blocking
466            further use of that particular connection -JM */
467
468         if (unlikely(nf_ct_is_dying(ct))) {
469                 spin_unlock_bh(&nf_conntrack_lock);
470                 return NF_ACCEPT;
471         }
472
473         /* See if there's one in the list already, including reverse:
474            NAT could have grabbed it without realizing, since we're
475            not in the hash.  If there is, we lost race. */
476         hlist_nulls_for_each_entry(h, n, &net->ct.hash[hash], hnnode)
477                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
478                                       &h->tuple) &&
479                     zone == nf_ct_zone(nf_ct_tuplehash_to_ctrack(h)))
480                         goto out;
481         hlist_nulls_for_each_entry(h, n, &net->ct.hash[repl_hash], hnnode)
482                 if (nf_ct_tuple_equal(&ct->tuplehash[IP_CT_DIR_REPLY].tuple,
483                                       &h->tuple) &&
484                     zone == nf_ct_zone(nf_ct_tuplehash_to_ctrack(h)))
485                         goto out;
486
487         /* Remove from unconfirmed list */
488         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
489
490         /* Timer relative to confirmation time, not original
491            setting time, otherwise we'd get timer wrap in
492            weird delay cases. */
493         ct->timeout.expires += jiffies;
494         add_timer(&ct->timeout);
495         atomic_inc(&ct->ct_general.use);
496         ct->status |= IPS_CONFIRMED;
497
498         /* set conntrack timestamp, if enabled. */
499         tstamp = nf_conn_tstamp_find(ct);
500         if (tstamp) {
501                 if (skb->tstamp.tv64 == 0)
502                         __net_timestamp((struct sk_buff *)skb);
503
504                 tstamp->start = ktime_to_ns(skb->tstamp);
505         }
506         /* Since the lookup is lockless, hash insertion must be done after
507          * starting the timer and setting the CONFIRMED bit. The RCU barriers
508          * guarantee that no other CPU can find the conntrack before the above
509          * stores are visible.
510          */
511         __nf_conntrack_hash_insert(ct, hash, repl_hash);
512         NF_CT_STAT_INC(net, insert);
513         spin_unlock_bh(&nf_conntrack_lock);
514
515         help = nfct_help(ct);
516         if (help && help->helper)
517                 nf_conntrack_event_cache(IPCT_HELPER, ct);
518
519         nf_conntrack_event_cache(master_ct(ct) ?
520                                  IPCT_RELATED : IPCT_NEW, ct);
521         return NF_ACCEPT;
522
523 out:
524         NF_CT_STAT_INC(net, insert_failed);
525         spin_unlock_bh(&nf_conntrack_lock);
526         return NF_DROP;
527 }
528 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
529
530 /* Returns true if a connection correspondings to the tuple (required
531    for NAT). */
532 int
533 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
534                          const struct nf_conn *ignored_conntrack)
535 {
536         struct net *net = nf_ct_net(ignored_conntrack);
537         struct nf_conntrack_tuple_hash *h;
538         struct hlist_nulls_node *n;
539         struct nf_conn *ct;
540         u16 zone = nf_ct_zone(ignored_conntrack);
541         unsigned int hash = hash_conntrack(net, zone, tuple);
542
543         /* Disable BHs the entire time since we need to disable them at
544          * least once for the stats anyway.
545          */
546         rcu_read_lock_bh();
547         hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash], hnnode) {
548                 ct = nf_ct_tuplehash_to_ctrack(h);
549                 if (ct != ignored_conntrack &&
550                     nf_ct_tuple_equal(tuple, &h->tuple) &&
551                     nf_ct_zone(ct) == zone) {
552                         NF_CT_STAT_INC(net, found);
553                         rcu_read_unlock_bh();
554                         return 1;
555                 }
556                 NF_CT_STAT_INC(net, searched);
557         }
558         rcu_read_unlock_bh();
559
560         return 0;
561 }
562 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
563
564 #define NF_CT_EVICTION_RANGE    8
565
566 /* There's a small race here where we may free a just-assured
567    connection.  Too bad: we're in trouble anyway. */
568 static noinline int early_drop(struct net *net, unsigned int hash)
569 {
570         /* Use oldest entry, which is roughly LRU */
571         struct nf_conntrack_tuple_hash *h;
572         struct nf_conn *ct = NULL, *tmp;
573         struct hlist_nulls_node *n;
574         unsigned int i, cnt = 0;
575         int dropped = 0;
576
577         rcu_read_lock();
578         for (i = 0; i < net->ct.htable_size; i++) {
579                 hlist_nulls_for_each_entry_rcu(h, n, &net->ct.hash[hash],
580                                          hnnode) {
581                         tmp = nf_ct_tuplehash_to_ctrack(h);
582                         if (!test_bit(IPS_ASSURED_BIT, &tmp->status))
583                                 ct = tmp;
584                         cnt++;
585                 }
586
587                 if (ct != NULL) {
588                         if (likely(!nf_ct_is_dying(ct) &&
589                                    atomic_inc_not_zero(&ct->ct_general.use)))
590                                 break;
591                         else
592                                 ct = NULL;
593                 }
594
595                 if (cnt >= NF_CT_EVICTION_RANGE)
596                         break;
597
598                 hash = (hash + 1) % net->ct.htable_size;
599         }
600         rcu_read_unlock();
601
602         if (!ct)
603                 return dropped;
604
605         if (del_timer(&ct->timeout)) {
606                 death_by_timeout((unsigned long)ct);
607                 dropped = 1;
608                 NF_CT_STAT_INC_ATOMIC(net, early_drop);
609         }
610         nf_ct_put(ct);
611         return dropped;
612 }
613
614 void init_nf_conntrack_hash_rnd(void)
615 {
616         unsigned int rand;
617
618         /*
619          * Why not initialize nf_conntrack_rnd in a "init()" function ?
620          * Because there isn't enough entropy when system initializing,
621          * and we initialize it as late as possible.
622          */
623         do {
624                 get_random_bytes(&rand, sizeof(rand));
625         } while (!rand);
626         cmpxchg(&nf_conntrack_hash_rnd, 0, rand);
627 }
628
629 static struct nf_conn *
630 __nf_conntrack_alloc(struct net *net, u16 zone,
631                      const struct nf_conntrack_tuple *orig,
632                      const struct nf_conntrack_tuple *repl,
633                      gfp_t gfp, u32 hash)
634 {
635         struct nf_conn *ct;
636
637         if (unlikely(!nf_conntrack_hash_rnd)) {
638                 init_nf_conntrack_hash_rnd();
639                 /* recompute the hash as nf_conntrack_hash_rnd is initialized */
640                 hash = hash_conntrack_raw(orig, zone);
641         }
642
643         /* We don't want any race condition at early drop stage */
644         atomic_inc(&net->ct.count);
645
646         if (nf_conntrack_max &&
647             unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
648                 if (!early_drop(net, hash_bucket(hash, net))) {
649                         atomic_dec(&net->ct.count);
650                         if (net_ratelimit())
651                                 printk(KERN_WARNING
652                                        "nf_conntrack: table full, dropping"
653                                        " packet.\n");
654                         return ERR_PTR(-ENOMEM);
655                 }
656         }
657
658         /*
659          * Do not use kmem_cache_zalloc(), as this cache uses
660          * SLAB_DESTROY_BY_RCU.
661          */
662         ct = kmem_cache_alloc(net->ct.nf_conntrack_cachep, gfp);
663         if (ct == NULL) {
664                 pr_debug("nf_conntrack_alloc: Can't alloc conntrack.\n");
665                 atomic_dec(&net->ct.count);
666                 return ERR_PTR(-ENOMEM);
667         }
668         /*
669          * Let ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.next
670          * and ct->tuplehash[IP_CT_DIR_REPLY].hnnode.next unchanged.
671          */
672         memset(&ct->tuplehash[IP_CT_DIR_MAX], 0,
673                offsetof(struct nf_conn, proto) -
674                offsetof(struct nf_conn, tuplehash[IP_CT_DIR_MAX]));
675         spin_lock_init(&ct->lock);
676         ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
677         ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
678         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
679         /* save hash for reusing when confirming */
680         *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
681         /* Don't set timer yet: wait for confirmation */
682         setup_timer(&ct->timeout, death_by_timeout, (unsigned long)ct);
683         write_pnet(&ct->ct_net, net);
684 #ifdef CONFIG_NF_CONNTRACK_ZONES
685         if (zone) {
686                 struct nf_conntrack_zone *nf_ct_zone;
687
688                 nf_ct_zone = nf_ct_ext_add(ct, NF_CT_EXT_ZONE, GFP_ATOMIC);
689                 if (!nf_ct_zone)
690                         goto out_free;
691                 nf_ct_zone->id = zone;
692         }
693 #endif
694         /*
695          * changes to lookup keys must be done before setting refcnt to 1
696          */
697         smp_wmb();
698         atomic_set(&ct->ct_general.use, 1);
699         return ct;
700
701 #ifdef CONFIG_NF_CONNTRACK_ZONES
702 out_free:
703         kmem_cache_free(net->ct.nf_conntrack_cachep, ct);
704         return ERR_PTR(-ENOMEM);
705 #endif
706 }
707
708 struct nf_conn *nf_conntrack_alloc(struct net *net, u16 zone,
709                                    const struct nf_conntrack_tuple *orig,
710                                    const struct nf_conntrack_tuple *repl,
711                                    gfp_t gfp)
712 {
713         return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
714 }
715 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
716
717 void nf_conntrack_free(struct nf_conn *ct)
718 {
719         struct net *net = nf_ct_net(ct);
720
721         nf_ct_ext_destroy(ct);
722         atomic_dec(&net->ct.count);
723         nf_ct_ext_free(ct);
724         kmem_cache_free(net->ct.nf_conntrack_cachep, ct);
725 }
726 EXPORT_SYMBOL_GPL(nf_conntrack_free);
727
728 /* Allocate a new conntrack: we return -ENOMEM if classification
729    failed due to stress.  Otherwise it really is unclassifiable. */
730 static struct nf_conntrack_tuple_hash *
731 init_conntrack(struct net *net, struct nf_conn *tmpl,
732                const struct nf_conntrack_tuple *tuple,
733                struct nf_conntrack_l3proto *l3proto,
734                struct nf_conntrack_l4proto *l4proto,
735                struct sk_buff *skb,
736                unsigned int dataoff, u32 hash)
737 {
738         struct nf_conn *ct;
739         struct nf_conn_help *help;
740         struct nf_conntrack_tuple repl_tuple;
741         struct nf_conntrack_ecache *ecache;
742         struct nf_conntrack_expect *exp;
743         u16 zone = tmpl ? nf_ct_zone(tmpl) : NF_CT_DEFAULT_ZONE;
744
745         if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
746                 pr_debug("Can't invert tuple.\n");
747                 return NULL;
748         }
749
750         ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
751                                   hash);
752         if (IS_ERR(ct)) {
753                 pr_debug("Can't allocate conntrack.\n");
754                 return (struct nf_conntrack_tuple_hash *)ct;
755         }
756
757         if (!l4proto->new(ct, skb, dataoff)) {
758                 nf_conntrack_free(ct);
759                 pr_debug("init conntrack: can't track with proto module\n");
760                 return NULL;
761         }
762
763         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
764         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
765
766         ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
767         nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
768                                  ecache ? ecache->expmask : 0,
769                              GFP_ATOMIC);
770
771         spin_lock_bh(&nf_conntrack_lock);
772         exp = nf_ct_find_expectation(net, zone, tuple);
773         if (exp) {
774                 pr_debug("conntrack: expectation arrives ct=%p exp=%p\n",
775                          ct, exp);
776                 /* Welcome, Mr. Bond.  We've been expecting you... */
777                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
778                 ct->master = exp->master;
779                 if (exp->helper) {
780                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
781                         if (help)
782                                 rcu_assign_pointer(help->helper, exp->helper);
783                 }
784
785 #ifdef CONFIG_NF_CONNTRACK_MARK
786                 ct->mark = exp->master->mark;
787 #endif
788 #ifdef CONFIG_NF_CONNTRACK_SECMARK
789                 ct->secmark = exp->master->secmark;
790 #endif
791                 nf_conntrack_get(&ct->master->ct_general);
792                 NF_CT_STAT_INC(net, expect_new);
793         } else {
794                 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
795                 NF_CT_STAT_INC(net, new);
796         }
797
798         /* Overload tuple linked list to put us in unconfirmed list. */
799         hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
800                        &net->ct.unconfirmed);
801
802         spin_unlock_bh(&nf_conntrack_lock);
803
804         if (exp) {
805                 if (exp->expectfn)
806                         exp->expectfn(ct, exp);
807                 nf_ct_expect_put(exp);
808         }
809
810         return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
811 }
812
813 /* On success, returns conntrack ptr, sets skb->nfct and ctinfo */
814 static inline struct nf_conn *
815 resolve_normal_ct(struct net *net, struct nf_conn *tmpl,
816                   struct sk_buff *skb,
817                   unsigned int dataoff,
818                   u_int16_t l3num,
819                   u_int8_t protonum,
820                   struct nf_conntrack_l3proto *l3proto,
821                   struct nf_conntrack_l4proto *l4proto,
822                   int *set_reply,
823                   enum ip_conntrack_info *ctinfo)
824 {
825         struct nf_conntrack_tuple tuple;
826         struct nf_conntrack_tuple_hash *h;
827         struct nf_conn *ct;
828         u16 zone = tmpl ? nf_ct_zone(tmpl) : NF_CT_DEFAULT_ZONE;
829         u32 hash;
830
831         if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
832                              dataoff, l3num, protonum, &tuple, l3proto,
833                              l4proto)) {
834                 pr_debug("resolve_normal_ct: Can't get tuple\n");
835                 return NULL;
836         }
837
838         /* look for tuple match */
839         hash = hash_conntrack_raw(&tuple, zone);
840         h = __nf_conntrack_find_get(net, zone, &tuple, hash);
841         if (!h) {
842                 h = init_conntrack(net, tmpl, &tuple, l3proto, l4proto,
843                                    skb, dataoff, hash);
844                 if (!h)
845                         return NULL;
846                 if (IS_ERR(h))
847                         return (void *)h;
848         }
849         ct = nf_ct_tuplehash_to_ctrack(h);
850
851         /* It exists; we have (non-exclusive) reference. */
852         if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
853                 *ctinfo = IP_CT_ESTABLISHED_REPLY;
854                 /* Please set reply bit if this packet OK */
855                 *set_reply = 1;
856         } else {
857                 /* Once we've had two way comms, always ESTABLISHED. */
858                 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
859                         pr_debug("nf_conntrack_in: normal packet for %p\n", ct);
860                         *ctinfo = IP_CT_ESTABLISHED;
861                 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
862                         pr_debug("nf_conntrack_in: related packet for %p\n",
863                                  ct);
864                         *ctinfo = IP_CT_RELATED;
865                 } else {
866                         pr_debug("nf_conntrack_in: new packet for %p\n", ct);
867                         *ctinfo = IP_CT_NEW;
868                 }
869                 *set_reply = 0;
870         }
871         skb->nfct = &ct->ct_general;
872         skb->nfctinfo = *ctinfo;
873         return ct;
874 }
875
876 unsigned int
877 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
878                 struct sk_buff *skb)
879 {
880         struct nf_conn *ct, *tmpl = NULL;
881         enum ip_conntrack_info ctinfo;
882         struct nf_conntrack_l3proto *l3proto;
883         struct nf_conntrack_l4proto *l4proto;
884         unsigned int dataoff;
885         u_int8_t protonum;
886         int set_reply = 0;
887         int ret;
888
889         if (skb->nfct) {
890                 /* Previously seen (loopback or untracked)?  Ignore. */
891                 tmpl = (struct nf_conn *)skb->nfct;
892                 if (!nf_ct_is_template(tmpl)) {
893                         NF_CT_STAT_INC_ATOMIC(net, ignore);
894                         return NF_ACCEPT;
895                 }
896                 skb->nfct = NULL;
897         }
898
899         /* rcu_read_lock()ed by nf_hook_slow */
900         l3proto = __nf_ct_l3proto_find(pf);
901         ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
902                                    &dataoff, &protonum);
903         if (ret <= 0) {
904                 pr_debug("not prepared to track yet or error occurred\n");
905                 NF_CT_STAT_INC_ATOMIC(net, error);
906                 NF_CT_STAT_INC_ATOMIC(net, invalid);
907                 ret = -ret;
908                 goto out;
909         }
910
911         l4proto = __nf_ct_l4proto_find(pf, protonum);
912
913         /* It may be an special packet, error, unclean...
914          * inverse of the return code tells to the netfilter
915          * core what to do with the packet. */
916         if (l4proto->error != NULL) {
917                 ret = l4proto->error(net, tmpl, skb, dataoff, &ctinfo,
918                                      pf, hooknum);
919                 if (ret <= 0) {
920                         NF_CT_STAT_INC_ATOMIC(net, error);
921                         NF_CT_STAT_INC_ATOMIC(net, invalid);
922                         ret = -ret;
923                         goto out;
924                 }
925                 /* ICMP[v6] protocol trackers may assign one conntrack. */
926                 if (skb->nfct)
927                         goto out;
928         }
929
930         ct = resolve_normal_ct(net, tmpl, skb, dataoff, pf, protonum,
931                                l3proto, l4proto, &set_reply, &ctinfo);
932         if (!ct) {
933                 /* Not valid part of a connection */
934                 NF_CT_STAT_INC_ATOMIC(net, invalid);
935                 ret = NF_ACCEPT;
936                 goto out;
937         }
938
939         if (IS_ERR(ct)) {
940                 /* Too stressed to deal. */
941                 NF_CT_STAT_INC_ATOMIC(net, drop);
942                 ret = NF_DROP;
943                 goto out;
944         }
945
946         NF_CT_ASSERT(skb->nfct);
947
948         ret = l4proto->packet(ct, skb, dataoff, ctinfo, pf, hooknum);
949         if (ret <= 0) {
950                 /* Invalid: inverse of the return code tells
951                  * the netfilter core what to do */
952                 pr_debug("nf_conntrack_in: Can't track with proto module\n");
953                 nf_conntrack_put(skb->nfct);
954                 skb->nfct = NULL;
955                 NF_CT_STAT_INC_ATOMIC(net, invalid);
956                 if (ret == -NF_DROP)
957                         NF_CT_STAT_INC_ATOMIC(net, drop);
958                 ret = -ret;
959                 goto out;
960         }
961
962         if (set_reply && !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
963                 nf_conntrack_event_cache(IPCT_REPLY, ct);
964 out:
965         if (tmpl) {
966                 /* Special case: we have to repeat this hook, assign the
967                  * template again to this packet. We assume that this packet
968                  * has no conntrack assigned. This is used by nf_ct_tcp. */
969                 if (ret == NF_REPEAT)
970                         skb->nfct = (struct nf_conntrack *)tmpl;
971                 else
972                         nf_ct_put(tmpl);
973         }
974
975         return ret;
976 }
977 EXPORT_SYMBOL_GPL(nf_conntrack_in);
978
979 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
980                           const struct nf_conntrack_tuple *orig)
981 {
982         bool ret;
983
984         rcu_read_lock();
985         ret = nf_ct_invert_tuple(inverse, orig,
986                                  __nf_ct_l3proto_find(orig->src.l3num),
987                                  __nf_ct_l4proto_find(orig->src.l3num,
988                                                       orig->dst.protonum));
989         rcu_read_unlock();
990         return ret;
991 }
992 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
993
994 /* Alter reply tuple (maybe alter helper).  This is for NAT, and is
995    implicitly racy: see __nf_conntrack_confirm */
996 void nf_conntrack_alter_reply(struct nf_conn *ct,
997                               const struct nf_conntrack_tuple *newreply)
998 {
999         struct nf_conn_help *help = nfct_help(ct);
1000
1001         /* Should be unconfirmed, so not in hash table yet */
1002         NF_CT_ASSERT(!nf_ct_is_confirmed(ct));
1003
1004         pr_debug("Altering reply tuple of %p to ", ct);
1005         nf_ct_dump_tuple(newreply);
1006
1007         ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1008         if (ct->master || (help && !hlist_empty(&help->expectations)))
1009                 return;
1010
1011         rcu_read_lock();
1012         __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1013         rcu_read_unlock();
1014 }
1015 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1016
1017 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1018 void __nf_ct_refresh_acct(struct nf_conn *ct,
1019                           enum ip_conntrack_info ctinfo,
1020                           const struct sk_buff *skb,
1021                           unsigned long extra_jiffies,
1022                           int do_acct)
1023 {
1024         NF_CT_ASSERT(ct->timeout.data == (unsigned long)ct);
1025         NF_CT_ASSERT(skb);
1026
1027         /* Only update if this is not a fixed timeout */
1028         if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1029                 goto acct;
1030
1031         /* If not in hash table, timer will not be active yet */
1032         if (!nf_ct_is_confirmed(ct)) {
1033                 ct->timeout.expires = extra_jiffies;
1034         } else {
1035                 unsigned long newtime = jiffies + extra_jiffies;
1036
1037                 /* Only update the timeout if the new timeout is at least
1038                    HZ jiffies from the old timeout. Need del_timer for race
1039                    avoidance (may already be dying). */
1040                 if (newtime - ct->timeout.expires >= HZ)
1041                         mod_timer_pending(&ct->timeout, newtime);
1042         }
1043
1044 acct:
1045         if (do_acct) {
1046                 struct nf_conn_counter *acct;
1047
1048                 acct = nf_conn_acct_find(ct);
1049                 if (acct) {
1050                         spin_lock_bh(&ct->lock);
1051                         acct[CTINFO2DIR(ctinfo)].packets++;
1052                         acct[CTINFO2DIR(ctinfo)].bytes += skb->len;
1053                         spin_unlock_bh(&ct->lock);
1054                 }
1055         }
1056 }
1057 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1058
1059 bool __nf_ct_kill_acct(struct nf_conn *ct,
1060                        enum ip_conntrack_info ctinfo,
1061                        const struct sk_buff *skb,
1062                        int do_acct)
1063 {
1064         if (do_acct) {
1065                 struct nf_conn_counter *acct;
1066
1067                 acct = nf_conn_acct_find(ct);
1068                 if (acct) {
1069                         spin_lock_bh(&ct->lock);
1070                         acct[CTINFO2DIR(ctinfo)].packets++;
1071                         acct[CTINFO2DIR(ctinfo)].bytes +=
1072                                 skb->len - skb_network_offset(skb);
1073                         spin_unlock_bh(&ct->lock);
1074                 }
1075         }
1076
1077         if (del_timer(&ct->timeout)) {
1078                 ct->timeout.function((unsigned long)ct);
1079                 return true;
1080         }
1081         return false;
1082 }
1083 EXPORT_SYMBOL_GPL(__nf_ct_kill_acct);
1084
1085 #ifdef CONFIG_NF_CONNTRACK_ZONES
1086 static struct nf_ct_ext_type nf_ct_zone_extend __read_mostly = {
1087         .len    = sizeof(struct nf_conntrack_zone),
1088         .align  = __alignof__(struct nf_conntrack_zone),
1089         .id     = NF_CT_EXT_ZONE,
1090 };
1091 #endif
1092
1093 #if defined(CONFIG_NF_CT_NETLINK) || defined(CONFIG_NF_CT_NETLINK_MODULE)
1094
1095 #include <linux/netfilter/nfnetlink.h>
1096 #include <linux/netfilter/nfnetlink_conntrack.h>
1097 #include <linux/mutex.h>
1098
1099 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
1100  * in ip_conntrack_core, since we don't want the protocols to autoload
1101  * or depend on ctnetlink */
1102 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
1103                                const struct nf_conntrack_tuple *tuple)
1104 {
1105         NLA_PUT_BE16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port);
1106         NLA_PUT_BE16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port);
1107         return 0;
1108
1109 nla_put_failure:
1110         return -1;
1111 }
1112 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
1113
1114 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
1115         [CTA_PROTO_SRC_PORT]  = { .type = NLA_U16 },
1116         [CTA_PROTO_DST_PORT]  = { .type = NLA_U16 },
1117 };
1118 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
1119
1120 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
1121                                struct nf_conntrack_tuple *t)
1122 {
1123         if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
1124                 return -EINVAL;
1125
1126         t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
1127         t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
1128
1129         return 0;
1130 }
1131 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
1132
1133 int nf_ct_port_nlattr_tuple_size(void)
1134 {
1135         return nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
1136 }
1137 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
1138 #endif
1139
1140 /* Used by ipt_REJECT and ip6t_REJECT. */
1141 static void nf_conntrack_attach(struct sk_buff *nskb, struct sk_buff *skb)
1142 {
1143         struct nf_conn *ct;
1144         enum ip_conntrack_info ctinfo;
1145
1146         /* This ICMP is in reverse direction to the packet which caused it */
1147         ct = nf_ct_get(skb, &ctinfo);
1148         if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
1149                 ctinfo = IP_CT_RELATED_REPLY;
1150         else
1151                 ctinfo = IP_CT_RELATED;
1152
1153         /* Attach to new skbuff, and increment count */
1154         nskb->nfct = &ct->ct_general;
1155         nskb->nfctinfo = ctinfo;
1156         nf_conntrack_get(nskb->nfct);
1157 }
1158
1159 /* Bring out ya dead! */
1160 static struct nf_conn *
1161 get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
1162                 void *data, unsigned int *bucket)
1163 {
1164         struct nf_conntrack_tuple_hash *h;
1165         struct nf_conn *ct;
1166         struct hlist_nulls_node *n;
1167
1168         spin_lock_bh(&nf_conntrack_lock);
1169         for (; *bucket < net->ct.htable_size; (*bucket)++) {
1170                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[*bucket], hnnode) {
1171                         ct = nf_ct_tuplehash_to_ctrack(h);
1172                         if (iter(ct, data))
1173                                 goto found;
1174                 }
1175         }
1176         hlist_nulls_for_each_entry(h, n, &net->ct.unconfirmed, hnnode) {
1177                 ct = nf_ct_tuplehash_to_ctrack(h);
1178                 if (iter(ct, data))
1179                         set_bit(IPS_DYING_BIT, &ct->status);
1180         }
1181         spin_unlock_bh(&nf_conntrack_lock);
1182         return NULL;
1183 found:
1184         atomic_inc(&ct->ct_general.use);
1185         spin_unlock_bh(&nf_conntrack_lock);
1186         return ct;
1187 }
1188
1189 void nf_ct_iterate_cleanup(struct net *net,
1190                            int (*iter)(struct nf_conn *i, void *data),
1191                            void *data)
1192 {
1193         struct nf_conn *ct;
1194         unsigned int bucket = 0;
1195
1196         while ((ct = get_next_corpse(net, iter, data, &bucket)) != NULL) {
1197                 /* Time to push up daises... */
1198                 if (del_timer(&ct->timeout))
1199                         death_by_timeout((unsigned long)ct);
1200                 /* ... else the timer will get him soon. */
1201
1202                 nf_ct_put(ct);
1203         }
1204 }
1205 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);
1206
1207 struct __nf_ct_flush_report {
1208         u32 pid;
1209         int report;
1210 };
1211
1212 static int kill_report(struct nf_conn *i, void *data)
1213 {
1214         struct __nf_ct_flush_report *fr = (struct __nf_ct_flush_report *)data;
1215         struct nf_conn_tstamp *tstamp;
1216
1217         tstamp = nf_conn_tstamp_find(i);
1218         if (tstamp && tstamp->stop == 0)
1219                 tstamp->stop = ktime_to_ns(ktime_get_real());
1220
1221         /* If we fail to deliver the event, death_by_timeout() will retry */
1222         if (nf_conntrack_event_report(IPCT_DESTROY, i,
1223                                       fr->pid, fr->report) < 0)
1224                 return 1;
1225
1226         /* Avoid the delivery of the destroy event in death_by_timeout(). */
1227         set_bit(IPS_DYING_BIT, &i->status);
1228         return 1;
1229 }
1230
1231 static int kill_all(struct nf_conn *i, void *data)
1232 {
1233         return 1;
1234 }
1235
1236 void nf_ct_free_hashtable(void *hash, unsigned int size)
1237 {
1238         if (is_vmalloc_addr(hash))
1239                 vfree(hash);
1240         else
1241                 free_pages((unsigned long)hash,
1242                            get_order(sizeof(struct hlist_head) * size));
1243 }
1244 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1245
1246 void nf_conntrack_flush_report(struct net *net, u32 pid, int report)
1247 {
1248         struct __nf_ct_flush_report fr = {
1249                 .pid    = pid,
1250                 .report = report,
1251         };
1252         nf_ct_iterate_cleanup(net, kill_report, &fr);
1253 }
1254 EXPORT_SYMBOL_GPL(nf_conntrack_flush_report);
1255
1256 static void nf_ct_release_dying_list(struct net *net)
1257 {
1258         struct nf_conntrack_tuple_hash *h;
1259         struct nf_conn *ct;
1260         struct hlist_nulls_node *n;
1261
1262         spin_lock_bh(&nf_conntrack_lock);
1263         hlist_nulls_for_each_entry(h, n, &net->ct.dying, hnnode) {
1264                 ct = nf_ct_tuplehash_to_ctrack(h);
1265                 /* never fails to remove them, no listeners at this point */
1266                 nf_ct_kill(ct);
1267         }
1268         spin_unlock_bh(&nf_conntrack_lock);
1269 }
1270
1271 static int untrack_refs(void)
1272 {
1273         int cnt = 0, cpu;
1274
1275         for_each_possible_cpu(cpu) {
1276                 struct nf_conn *ct = &per_cpu(nf_conntrack_untracked, cpu);
1277
1278                 cnt += atomic_read(&ct->ct_general.use) - 1;
1279         }
1280         return cnt;
1281 }
1282
1283 static void nf_conntrack_cleanup_init_net(void)
1284 {
1285         while (untrack_refs() > 0)
1286                 schedule();
1287
1288         nf_conntrack_helper_fini();
1289         nf_conntrack_proto_fini();
1290 #ifdef CONFIG_NF_CONNTRACK_ZONES
1291         nf_ct_extend_unregister(&nf_ct_zone_extend);
1292 #endif
1293 }
1294
1295 static void nf_conntrack_cleanup_net(struct net *net)
1296 {
1297  i_see_dead_people:
1298         nf_ct_iterate_cleanup(net, kill_all, NULL);
1299         nf_ct_release_dying_list(net);
1300         if (atomic_read(&net->ct.count) != 0) {
1301                 schedule();
1302                 goto i_see_dead_people;
1303         }
1304
1305         nf_ct_free_hashtable(net->ct.hash, net->ct.htable_size);
1306         nf_conntrack_ecache_fini(net);
1307         nf_conntrack_tstamp_fini(net);
1308         nf_conntrack_acct_fini(net);
1309         nf_conntrack_expect_fini(net);
1310         kmem_cache_destroy(net->ct.nf_conntrack_cachep);
1311         kfree(net->ct.slabname);
1312         free_percpu(net->ct.stat);
1313 }
1314
1315 /* Mishearing the voices in his head, our hero wonders how he's
1316    supposed to kill the mall. */
1317 void nf_conntrack_cleanup(struct net *net)
1318 {
1319         if (net_eq(net, &init_net))
1320                 rcu_assign_pointer(ip_ct_attach, NULL);
1321
1322         /* This makes sure all current packets have passed through
1323            netfilter framework.  Roll on, two-stage module
1324            delete... */
1325         synchronize_net();
1326
1327         nf_conntrack_cleanup_net(net);
1328
1329         if (net_eq(net, &init_net)) {
1330                 rcu_assign_pointer(nf_ct_destroy, NULL);
1331                 nf_conntrack_cleanup_init_net();
1332         }
1333 }
1334
1335 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
1336 {
1337         struct hlist_nulls_head *hash;
1338         unsigned int nr_slots, i;
1339         size_t sz;
1340
1341         BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1342         nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1343         sz = nr_slots * sizeof(struct hlist_nulls_head);
1344         hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1345                                         get_order(sz));
1346         if (!hash) {
1347                 printk(KERN_WARNING "nf_conntrack: falling back to vmalloc.\n");
1348                 hash = __vmalloc(sz, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1349                                  PAGE_KERNEL);
1350         }
1351
1352         if (hash && nulls)
1353                 for (i = 0; i < nr_slots; i++)
1354                         INIT_HLIST_NULLS_HEAD(&hash[i], i);
1355
1356         return hash;
1357 }
1358 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1359
1360 int nf_conntrack_set_hashsize(const char *val, struct kernel_param *kp)
1361 {
1362         int i, bucket;
1363         unsigned int hashsize, old_size;
1364         struct hlist_nulls_head *hash, *old_hash;
1365         struct nf_conntrack_tuple_hash *h;
1366         struct nf_conn *ct;
1367
1368         if (current->nsproxy->net_ns != &init_net)
1369                 return -EOPNOTSUPP;
1370
1371         /* On boot, we can set this without any fancy locking. */
1372         if (!nf_conntrack_htable_size)
1373                 return param_set_uint(val, kp);
1374
1375         hashsize = simple_strtoul(val, NULL, 0);
1376         if (!hashsize)
1377                 return -EINVAL;
1378
1379         hash = nf_ct_alloc_hashtable(&hashsize, 1);
1380         if (!hash)
1381                 return -ENOMEM;
1382
1383         /* Lookups in the old hash might happen in parallel, which means we
1384          * might get false negatives during connection lookup. New connections
1385          * created because of a false negative won't make it into the hash
1386          * though since that required taking the lock.
1387          */
1388         spin_lock_bh(&nf_conntrack_lock);
1389         for (i = 0; i < init_net.ct.htable_size; i++) {
1390                 while (!hlist_nulls_empty(&init_net.ct.hash[i])) {
1391                         h = hlist_nulls_entry(init_net.ct.hash[i].first,
1392                                         struct nf_conntrack_tuple_hash, hnnode);
1393                         ct = nf_ct_tuplehash_to_ctrack(h);
1394                         hlist_nulls_del_rcu(&h->hnnode);
1395                         bucket = __hash_conntrack(&h->tuple, nf_ct_zone(ct),
1396                                                   hashsize);
1397                         hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
1398                 }
1399         }
1400         old_size = init_net.ct.htable_size;
1401         old_hash = init_net.ct.hash;
1402
1403         init_net.ct.htable_size = nf_conntrack_htable_size = hashsize;
1404         init_net.ct.hash = hash;
1405         spin_unlock_bh(&nf_conntrack_lock);
1406
1407         nf_ct_free_hashtable(old_hash, old_size);
1408         return 0;
1409 }
1410 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
1411
1412 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
1413                   &nf_conntrack_htable_size, 0600);
1414
1415 void nf_ct_untracked_status_or(unsigned long bits)
1416 {
1417         int cpu;
1418
1419         for_each_possible_cpu(cpu)
1420                 per_cpu(nf_conntrack_untracked, cpu).status |= bits;
1421 }
1422 EXPORT_SYMBOL_GPL(nf_ct_untracked_status_or);
1423
1424 static int nf_conntrack_init_init_net(void)
1425 {
1426         int max_factor = 8;
1427         int ret, cpu;
1428
1429         /* Idea from tcp.c: use 1/16384 of memory.  On i386: 32MB
1430          * machine has 512 buckets. >= 1GB machines have 16384 buckets. */
1431         if (!nf_conntrack_htable_size) {
1432                 nf_conntrack_htable_size
1433                         = (((totalram_pages << PAGE_SHIFT) / 16384)
1434                            / sizeof(struct hlist_head));
1435                 if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
1436                         nf_conntrack_htable_size = 16384;
1437                 if (nf_conntrack_htable_size < 32)
1438                         nf_conntrack_htable_size = 32;
1439
1440                 /* Use a max. factor of four by default to get the same max as
1441                  * with the old struct list_heads. When a table size is given
1442                  * we use the old value of 8 to avoid reducing the max.
1443                  * entries. */
1444                 max_factor = 4;
1445         }
1446         nf_conntrack_max = max_factor * nf_conntrack_htable_size;
1447
1448         printk(KERN_INFO "nf_conntrack version %s (%u buckets, %d max)\n",
1449                NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
1450                nf_conntrack_max);
1451
1452         ret = nf_conntrack_proto_init();
1453         if (ret < 0)
1454                 goto err_proto;
1455
1456         ret = nf_conntrack_helper_init();
1457         if (ret < 0)
1458                 goto err_helper;
1459
1460 #ifdef CONFIG_NF_CONNTRACK_ZONES
1461         ret = nf_ct_extend_register(&nf_ct_zone_extend);
1462         if (ret < 0)
1463                 goto err_extend;
1464 #endif
1465         /* Set up fake conntrack: to never be deleted, not in any hashes */
1466         for_each_possible_cpu(cpu) {
1467                 struct nf_conn *ct = &per_cpu(nf_conntrack_untracked, cpu);
1468                 write_pnet(&ct->ct_net, &init_net);
1469                 atomic_set(&ct->ct_general.use, 1);
1470         }
1471         /*  - and look it like as a confirmed connection */
1472         nf_ct_untracked_status_or(IPS_CONFIRMED | IPS_UNTRACKED);
1473         return 0;
1474
1475 #ifdef CONFIG_NF_CONNTRACK_ZONES
1476 err_extend:
1477         nf_conntrack_helper_fini();
1478 #endif
1479 err_helper:
1480         nf_conntrack_proto_fini();
1481 err_proto:
1482         return ret;
1483 }
1484
1485 /*
1486  * We need to use special "null" values, not used in hash table
1487  */
1488 #define UNCONFIRMED_NULLS_VAL   ((1<<30)+0)
1489 #define DYING_NULLS_VAL         ((1<<30)+1)
1490
1491 static int nf_conntrack_init_net(struct net *net)
1492 {
1493         int ret;
1494
1495         atomic_set(&net->ct.count, 0);
1496         INIT_HLIST_NULLS_HEAD(&net->ct.unconfirmed, UNCONFIRMED_NULLS_VAL);
1497         INIT_HLIST_NULLS_HEAD(&net->ct.dying, DYING_NULLS_VAL);
1498         net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
1499         if (!net->ct.stat) {
1500                 ret = -ENOMEM;
1501                 goto err_stat;
1502         }
1503
1504         net->ct.slabname = kasprintf(GFP_KERNEL, "nf_conntrack_%p", net);
1505         if (!net->ct.slabname) {
1506                 ret = -ENOMEM;
1507                 goto err_slabname;
1508         }
1509
1510         net->ct.nf_conntrack_cachep = kmem_cache_create(net->ct.slabname,
1511                                                         sizeof(struct nf_conn), 0,
1512                                                         SLAB_DESTROY_BY_RCU, NULL);
1513         if (!net->ct.nf_conntrack_cachep) {
1514                 printk(KERN_ERR "Unable to create nf_conn slab cache\n");
1515                 ret = -ENOMEM;
1516                 goto err_cache;
1517         }
1518
1519         net->ct.htable_size = nf_conntrack_htable_size;
1520         net->ct.hash = nf_ct_alloc_hashtable(&net->ct.htable_size, 1);
1521         if (!net->ct.hash) {
1522                 ret = -ENOMEM;
1523                 printk(KERN_ERR "Unable to create nf_conntrack_hash\n");
1524                 goto err_hash;
1525         }
1526         ret = nf_conntrack_expect_init(net);
1527         if (ret < 0)
1528                 goto err_expect;
1529         ret = nf_conntrack_acct_init(net);
1530         if (ret < 0)
1531                 goto err_acct;
1532         ret = nf_conntrack_tstamp_init(net);
1533         if (ret < 0)
1534                 goto err_tstamp;
1535         ret = nf_conntrack_ecache_init(net);
1536         if (ret < 0)
1537                 goto err_ecache;
1538
1539         return 0;
1540
1541 err_ecache:
1542         nf_conntrack_tstamp_fini(net);
1543 err_tstamp:
1544         nf_conntrack_acct_fini(net);
1545 err_acct:
1546         nf_conntrack_expect_fini(net);
1547 err_expect:
1548         nf_ct_free_hashtable(net->ct.hash, net->ct.htable_size);
1549 err_hash:
1550         kmem_cache_destroy(net->ct.nf_conntrack_cachep);
1551 err_cache:
1552         kfree(net->ct.slabname);
1553 err_slabname:
1554         free_percpu(net->ct.stat);
1555 err_stat:
1556         return ret;
1557 }
1558
1559 s16 (*nf_ct_nat_offset)(const struct nf_conn *ct,
1560                         enum ip_conntrack_dir dir,
1561                         u32 seq);
1562 EXPORT_SYMBOL_GPL(nf_ct_nat_offset);
1563
1564 int nf_conntrack_init(struct net *net)
1565 {
1566         int ret;
1567
1568         if (net_eq(net, &init_net)) {
1569                 ret = nf_conntrack_init_init_net();
1570                 if (ret < 0)
1571                         goto out_init_net;
1572         }
1573         ret = nf_conntrack_init_net(net);
1574         if (ret < 0)
1575                 goto out_net;
1576
1577         if (net_eq(net, &init_net)) {
1578                 /* For use by REJECT target */
1579                 rcu_assign_pointer(ip_ct_attach, nf_conntrack_attach);
1580                 rcu_assign_pointer(nf_ct_destroy, destroy_conntrack);
1581
1582                 /* Howto get NAT offsets */
1583                 rcu_assign_pointer(nf_ct_nat_offset, NULL);
1584         }
1585         return 0;
1586
1587 out_net:
1588         if (net_eq(net, &init_net))
1589                 nf_conntrack_cleanup_init_net();
1590 out_init_net:
1591         return ret;
1592 }