inet: Fix __inet_inherit_port() to correctly increment bsockets and num_owners
[linux-3.10.git] / net / ipv4 / inet_hashtables.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Generic INET transport hashtables
7  *
8  * Authors:     Lotsa people, from code originally in tcp
9  *
10  *      This program is free software; you can redistribute it and/or
11  *      modify it under the terms of the GNU General Public License
12  *      as published by the Free Software Foundation; either version
13  *      2 of the License, or (at your option) any later version.
14  */
15
16 #include <linux/module.h>
17 #include <linux/random.h>
18 #include <linux/sched.h>
19 #include <linux/slab.h>
20 #include <linux/wait.h>
21
22 #include <net/inet_connection_sock.h>
23 #include <net/inet_hashtables.h>
24 #include <net/ip.h>
25
26 /*
27  * Allocate and initialize a new local port bind bucket.
28  * The bindhash mutex for snum's hash chain must be held here.
29  */
30 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
31                                                  struct net *net,
32                                                  struct inet_bind_hashbucket *head,
33                                                  const unsigned short snum)
34 {
35         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
36
37         if (tb != NULL) {
38                 write_pnet(&tb->ib_net, hold_net(net));
39                 tb->port      = snum;
40                 tb->fastreuse = 0;
41                 tb->num_owners = 0;
42                 INIT_HLIST_HEAD(&tb->owners);
43                 hlist_add_head(&tb->node, &head->chain);
44         }
45         return tb;
46 }
47
48 /*
49  * Caller must hold hashbucket lock for this tb with local BH disabled
50  */
51 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
52 {
53         if (hlist_empty(&tb->owners)) {
54                 __hlist_del(&tb->node);
55                 release_net(ib_net(tb));
56                 kmem_cache_free(cachep, tb);
57         }
58 }
59
60 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
61                     const unsigned short snum)
62 {
63         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
64
65         atomic_inc(&hashinfo->bsockets);
66
67         inet_sk(sk)->inet_num = snum;
68         sk_add_bind_node(sk, &tb->owners);
69         tb->num_owners++;
70         inet_csk(sk)->icsk_bind_hash = tb;
71 }
72
73 /*
74  * Get rid of any references to a local port held by the given sock.
75  */
76 static void __inet_put_port(struct sock *sk)
77 {
78         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
79         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
80                         hashinfo->bhash_size);
81         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
82         struct inet_bind_bucket *tb;
83
84         atomic_dec(&hashinfo->bsockets);
85
86         spin_lock(&head->lock);
87         tb = inet_csk(sk)->icsk_bind_hash;
88         __sk_del_bind_node(sk);
89         tb->num_owners--;
90         inet_csk(sk)->icsk_bind_hash = NULL;
91         inet_sk(sk)->inet_num = 0;
92         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
93         spin_unlock(&head->lock);
94 }
95
96 void inet_put_port(struct sock *sk)
97 {
98         local_bh_disable();
99         __inet_put_port(sk);
100         local_bh_enable();
101 }
102 EXPORT_SYMBOL(inet_put_port);
103
104 int __inet_inherit_port(struct sock *sk, struct sock *child)
105 {
106         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
107         unsigned short port = inet_sk(child)->inet_num;
108         const int bhash = inet_bhashfn(sock_net(sk), port,
109                         table->bhash_size);
110         struct inet_bind_hashbucket *head = &table->bhash[bhash];
111         struct inet_bind_bucket *tb;
112
113         spin_lock(&head->lock);
114         tb = inet_csk(sk)->icsk_bind_hash;
115         if (tb->port != port) {
116                 /* NOTE: using tproxy and redirecting skbs to a proxy
117                  * on a different listener port breaks the assumption
118                  * that the listener socket's icsk_bind_hash is the same
119                  * as that of the child socket. We have to look up or
120                  * create a new bind bucket for the child here. */
121                 struct hlist_node *node;
122                 inet_bind_bucket_for_each(tb, node, &head->chain) {
123                         if (net_eq(ib_net(tb), sock_net(sk)) &&
124                             tb->port == port)
125                                 break;
126                 }
127                 if (!node) {
128                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
129                                                      sock_net(sk), head, port);
130                         if (!tb) {
131                                 spin_unlock(&head->lock);
132                                 return -ENOMEM;
133                         }
134                 }
135         }
136         inet_bind_hash(child, tb, port);
137         spin_unlock(&head->lock);
138
139         return 0;
140 }
141 EXPORT_SYMBOL_GPL(__inet_inherit_port);
142
143 static inline int compute_score(struct sock *sk, struct net *net,
144                                 const unsigned short hnum, const __be32 daddr,
145                                 const int dif)
146 {
147         int score = -1;
148         struct inet_sock *inet = inet_sk(sk);
149
150         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
151                         !ipv6_only_sock(sk)) {
152                 __be32 rcv_saddr = inet->inet_rcv_saddr;
153                 score = sk->sk_family == PF_INET ? 1 : 0;
154                 if (rcv_saddr) {
155                         if (rcv_saddr != daddr)
156                                 return -1;
157                         score += 2;
158                 }
159                 if (sk->sk_bound_dev_if) {
160                         if (sk->sk_bound_dev_if != dif)
161                                 return -1;
162                         score += 2;
163                 }
164         }
165         return score;
166 }
167
168 /*
169  * Don't inline this cruft. Here are some nice properties to exploit here. The
170  * BSD API does not allow a listening sock to specify the remote port nor the
171  * remote address for the connection. So always assume those are both
172  * wildcarded during the search since they can never be otherwise.
173  */
174
175
176 struct sock *__inet_lookup_listener(struct net *net,
177                                     struct inet_hashinfo *hashinfo,
178                                     const __be32 daddr, const unsigned short hnum,
179                                     const int dif)
180 {
181         struct sock *sk, *result;
182         struct hlist_nulls_node *node;
183         unsigned int hash = inet_lhashfn(net, hnum);
184         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
185         int score, hiscore;
186
187         rcu_read_lock();
188 begin:
189         result = NULL;
190         hiscore = -1;
191         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
192                 score = compute_score(sk, net, hnum, daddr, dif);
193                 if (score > hiscore) {
194                         result = sk;
195                         hiscore = score;
196                 }
197         }
198         /*
199          * if the nulls value we got at the end of this lookup is
200          * not the expected one, we must restart lookup.
201          * We probably met an item that was moved to another chain.
202          */
203         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
204                 goto begin;
205         if (result) {
206                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
207                         result = NULL;
208                 else if (unlikely(compute_score(result, net, hnum, daddr,
209                                   dif) < hiscore)) {
210                         sock_put(result);
211                         goto begin;
212                 }
213         }
214         rcu_read_unlock();
215         return result;
216 }
217 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
218
219 struct sock * __inet_lookup_established(struct net *net,
220                                   struct inet_hashinfo *hashinfo,
221                                   const __be32 saddr, const __be16 sport,
222                                   const __be32 daddr, const u16 hnum,
223                                   const int dif)
224 {
225         INET_ADDR_COOKIE(acookie, saddr, daddr)
226         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
227         struct sock *sk;
228         const struct hlist_nulls_node *node;
229         /* Optimize here for direct hit, only listening connections can
230          * have wildcards anyways.
231          */
232         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
233         unsigned int slot = hash & hashinfo->ehash_mask;
234         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
235
236         rcu_read_lock();
237 begin:
238         sk_nulls_for_each_rcu(sk, node, &head->chain) {
239                 if (INET_MATCH(sk, net, hash, acookie,
240                                         saddr, daddr, ports, dif)) {
241                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
242                                 goto begintw;
243                         if (unlikely(!INET_MATCH(sk, net, hash, acookie,
244                                 saddr, daddr, ports, dif))) {
245                                 sock_put(sk);
246                                 goto begin;
247                         }
248                         goto out;
249                 }
250         }
251         /*
252          * if the nulls value we got at the end of this lookup is
253          * not the expected one, we must restart lookup.
254          * We probably met an item that was moved to another chain.
255          */
256         if (get_nulls_value(node) != slot)
257                 goto begin;
258
259 begintw:
260         /* Must check for a TIME_WAIT'er before going to listener hash. */
261         sk_nulls_for_each_rcu(sk, node, &head->twchain) {
262                 if (INET_TW_MATCH(sk, net, hash, acookie,
263                                         saddr, daddr, ports, dif)) {
264                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt))) {
265                                 sk = NULL;
266                                 goto out;
267                         }
268                         if (unlikely(!INET_TW_MATCH(sk, net, hash, acookie,
269                                  saddr, daddr, ports, dif))) {
270                                 sock_put(sk);
271                                 goto begintw;
272                         }
273                         goto out;
274                 }
275         }
276         /*
277          * if the nulls value we got at the end of this lookup is
278          * not the expected one, we must restart lookup.
279          * We probably met an item that was moved to another chain.
280          */
281         if (get_nulls_value(node) != slot)
282                 goto begintw;
283         sk = NULL;
284 out:
285         rcu_read_unlock();
286         return sk;
287 }
288 EXPORT_SYMBOL_GPL(__inet_lookup_established);
289
290 /* called with local bh disabled */
291 static int __inet_check_established(struct inet_timewait_death_row *death_row,
292                                     struct sock *sk, __u16 lport,
293                                     struct inet_timewait_sock **twp)
294 {
295         struct inet_hashinfo *hinfo = death_row->hashinfo;
296         struct inet_sock *inet = inet_sk(sk);
297         __be32 daddr = inet->inet_rcv_saddr;
298         __be32 saddr = inet->inet_daddr;
299         int dif = sk->sk_bound_dev_if;
300         INET_ADDR_COOKIE(acookie, saddr, daddr)
301         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
302         struct net *net = sock_net(sk);
303         unsigned int hash = inet_ehashfn(net, daddr, lport,
304                                          saddr, inet->inet_dport);
305         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
306         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
307         struct sock *sk2;
308         const struct hlist_nulls_node *node;
309         struct inet_timewait_sock *tw;
310         int twrefcnt = 0;
311
312         spin_lock(lock);
313
314         /* Check TIME-WAIT sockets first. */
315         sk_nulls_for_each(sk2, node, &head->twchain) {
316                 tw = inet_twsk(sk2);
317
318                 if (INET_TW_MATCH(sk2, net, hash, acookie,
319                                         saddr, daddr, ports, dif)) {
320                         if (twsk_unique(sk, sk2, twp))
321                                 goto unique;
322                         else
323                                 goto not_unique;
324                 }
325         }
326         tw = NULL;
327
328         /* And established part... */
329         sk_nulls_for_each(sk2, node, &head->chain) {
330                 if (INET_MATCH(sk2, net, hash, acookie,
331                                         saddr, daddr, ports, dif))
332                         goto not_unique;
333         }
334
335 unique:
336         /* Must record num and sport now. Otherwise we will see
337          * in hash table socket with a funny identity. */
338         inet->inet_num = lport;
339         inet->inet_sport = htons(lport);
340         sk->sk_hash = hash;
341         WARN_ON(!sk_unhashed(sk));
342         __sk_nulls_add_node_rcu(sk, &head->chain);
343         if (tw) {
344                 twrefcnt = inet_twsk_unhash(tw);
345                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
346         }
347         spin_unlock(lock);
348         if (twrefcnt)
349                 inet_twsk_put(tw);
350         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
351
352         if (twp) {
353                 *twp = tw;
354         } else if (tw) {
355                 /* Silly. Should hash-dance instead... */
356                 inet_twsk_deschedule(tw, death_row);
357
358                 inet_twsk_put(tw);
359         }
360         return 0;
361
362 not_unique:
363         spin_unlock(lock);
364         return -EADDRNOTAVAIL;
365 }
366
367 static inline u32 inet_sk_port_offset(const struct sock *sk)
368 {
369         const struct inet_sock *inet = inet_sk(sk);
370         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
371                                           inet->inet_daddr,
372                                           inet->inet_dport);
373 }
374
375 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
376 {
377         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
378         struct hlist_nulls_head *list;
379         spinlock_t *lock;
380         struct inet_ehash_bucket *head;
381         int twrefcnt = 0;
382
383         WARN_ON(!sk_unhashed(sk));
384
385         sk->sk_hash = inet_sk_ehashfn(sk);
386         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
387         list = &head->chain;
388         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
389
390         spin_lock(lock);
391         __sk_nulls_add_node_rcu(sk, list);
392         if (tw) {
393                 WARN_ON(sk->sk_hash != tw->tw_hash);
394                 twrefcnt = inet_twsk_unhash(tw);
395         }
396         spin_unlock(lock);
397         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
398         return twrefcnt;
399 }
400 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
401
402 static void __inet_hash(struct sock *sk)
403 {
404         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
405         struct inet_listen_hashbucket *ilb;
406
407         if (sk->sk_state != TCP_LISTEN) {
408                 __inet_hash_nolisten(sk, NULL);
409                 return;
410         }
411
412         WARN_ON(!sk_unhashed(sk));
413         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
414
415         spin_lock(&ilb->lock);
416         __sk_nulls_add_node_rcu(sk, &ilb->head);
417         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
418         spin_unlock(&ilb->lock);
419 }
420
421 void inet_hash(struct sock *sk)
422 {
423         if (sk->sk_state != TCP_CLOSE) {
424                 local_bh_disable();
425                 __inet_hash(sk);
426                 local_bh_enable();
427         }
428 }
429 EXPORT_SYMBOL_GPL(inet_hash);
430
431 void inet_unhash(struct sock *sk)
432 {
433         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
434         spinlock_t *lock;
435         int done;
436
437         if (sk_unhashed(sk))
438                 return;
439
440         if (sk->sk_state == TCP_LISTEN)
441                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
442         else
443                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
444
445         spin_lock_bh(lock);
446         done =__sk_nulls_del_node_init_rcu(sk);
447         if (done)
448                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
449         spin_unlock_bh(lock);
450 }
451 EXPORT_SYMBOL_GPL(inet_unhash);
452
453 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
454                 struct sock *sk, u32 port_offset,
455                 int (*check_established)(struct inet_timewait_death_row *,
456                         struct sock *, __u16, struct inet_timewait_sock **),
457                 int (*hash)(struct sock *sk, struct inet_timewait_sock *twp))
458 {
459         struct inet_hashinfo *hinfo = death_row->hashinfo;
460         const unsigned short snum = inet_sk(sk)->inet_num;
461         struct inet_bind_hashbucket *head;
462         struct inet_bind_bucket *tb;
463         int ret;
464         struct net *net = sock_net(sk);
465         int twrefcnt = 1;
466
467         if (!snum) {
468                 int i, remaining, low, high, port;
469                 static u32 hint;
470                 u32 offset = hint + port_offset;
471                 struct hlist_node *node;
472                 struct inet_timewait_sock *tw = NULL;
473
474                 inet_get_local_port_range(&low, &high);
475                 remaining = (high - low) + 1;
476
477                 local_bh_disable();
478                 for (i = 1; i <= remaining; i++) {
479                         port = low + (i + offset) % remaining;
480                         if (inet_is_reserved_local_port(port))
481                                 continue;
482                         head = &hinfo->bhash[inet_bhashfn(net, port,
483                                         hinfo->bhash_size)];
484                         spin_lock(&head->lock);
485
486                         /* Does not bother with rcv_saddr checks,
487                          * because the established check is already
488                          * unique enough.
489                          */
490                         inet_bind_bucket_for_each(tb, node, &head->chain) {
491                                 if (net_eq(ib_net(tb), net) &&
492                                     tb->port == port) {
493                                         if (tb->fastreuse >= 0)
494                                                 goto next_port;
495                                         WARN_ON(hlist_empty(&tb->owners));
496                                         if (!check_established(death_row, sk,
497                                                                 port, &tw))
498                                                 goto ok;
499                                         goto next_port;
500                                 }
501                         }
502
503                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
504                                         net, head, port);
505                         if (!tb) {
506                                 spin_unlock(&head->lock);
507                                 break;
508                         }
509                         tb->fastreuse = -1;
510                         goto ok;
511
512                 next_port:
513                         spin_unlock(&head->lock);
514                 }
515                 local_bh_enable();
516
517                 return -EADDRNOTAVAIL;
518
519 ok:
520                 hint += i;
521
522                 /* Head lock still held and bh's disabled */
523                 inet_bind_hash(sk, tb, port);
524                 if (sk_unhashed(sk)) {
525                         inet_sk(sk)->inet_sport = htons(port);
526                         twrefcnt += hash(sk, tw);
527                 }
528                 if (tw)
529                         twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
530                 spin_unlock(&head->lock);
531
532                 if (tw) {
533                         inet_twsk_deschedule(tw, death_row);
534                         while (twrefcnt) {
535                                 twrefcnt--;
536                                 inet_twsk_put(tw);
537                         }
538                 }
539
540                 ret = 0;
541                 goto out;
542         }
543
544         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
545         tb  = inet_csk(sk)->icsk_bind_hash;
546         spin_lock_bh(&head->lock);
547         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
548                 hash(sk, NULL);
549                 spin_unlock_bh(&head->lock);
550                 return 0;
551         } else {
552                 spin_unlock(&head->lock);
553                 /* No definite answer... Walk to established hash table */
554                 ret = check_established(death_row, sk, snum, NULL);
555 out:
556                 local_bh_enable();
557                 return ret;
558         }
559 }
560
561 /*
562  * Bind a port for a connect operation and hash it.
563  */
564 int inet_hash_connect(struct inet_timewait_death_row *death_row,
565                       struct sock *sk)
566 {
567         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
568                         __inet_check_established, __inet_hash_nolisten);
569 }
570 EXPORT_SYMBOL_GPL(inet_hash_connect);
571
572 void inet_hashinfo_init(struct inet_hashinfo *h)
573 {
574         int i;
575
576         atomic_set(&h->bsockets, 0);
577         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
578                 spin_lock_init(&h->listening_hash[i].lock);
579                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
580                                       i + LISTENING_NULLS_BASE);
581                 }
582 }
583 EXPORT_SYMBOL_GPL(inet_hashinfo_init);