1b344f30b463fab9ed70a8f19a19d348d7c626f7
[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         sk_add_bind_node(child, &tb->owners);
137         inet_csk(child)->icsk_bind_hash = tb;
138         spin_unlock(&head->lock);
139
140         return 0;
141 }
142 EXPORT_SYMBOL_GPL(__inet_inherit_port);
143
144 static inline int compute_score(struct sock *sk, struct net *net,
145                                 const unsigned short hnum, const __be32 daddr,
146                                 const int dif)
147 {
148         int score = -1;
149         struct inet_sock *inet = inet_sk(sk);
150
151         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
152                         !ipv6_only_sock(sk)) {
153                 __be32 rcv_saddr = inet->inet_rcv_saddr;
154                 score = sk->sk_family == PF_INET ? 1 : 0;
155                 if (rcv_saddr) {
156                         if (rcv_saddr != daddr)
157                                 return -1;
158                         score += 2;
159                 }
160                 if (sk->sk_bound_dev_if) {
161                         if (sk->sk_bound_dev_if != dif)
162                                 return -1;
163                         score += 2;
164                 }
165         }
166         return score;
167 }
168
169 /*
170  * Don't inline this cruft. Here are some nice properties to exploit here. The
171  * BSD API does not allow a listening sock to specify the remote port nor the
172  * remote address for the connection. So always assume those are both
173  * wildcarded during the search since they can never be otherwise.
174  */
175
176
177 struct sock *__inet_lookup_listener(struct net *net,
178                                     struct inet_hashinfo *hashinfo,
179                                     const __be32 daddr, const unsigned short hnum,
180                                     const int dif)
181 {
182         struct sock *sk, *result;
183         struct hlist_nulls_node *node;
184         unsigned int hash = inet_lhashfn(net, hnum);
185         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
186         int score, hiscore;
187
188         rcu_read_lock();
189 begin:
190         result = NULL;
191         hiscore = -1;
192         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
193                 score = compute_score(sk, net, hnum, daddr, dif);
194                 if (score > hiscore) {
195                         result = sk;
196                         hiscore = score;
197                 }
198         }
199         /*
200          * if the nulls value we got at the end of this lookup is
201          * not the expected one, we must restart lookup.
202          * We probably met an item that was moved to another chain.
203          */
204         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
205                 goto begin;
206         if (result) {
207                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
208                         result = NULL;
209                 else if (unlikely(compute_score(result, net, hnum, daddr,
210                                   dif) < hiscore)) {
211                         sock_put(result);
212                         goto begin;
213                 }
214         }
215         rcu_read_unlock();
216         return result;
217 }
218 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
219
220 struct sock * __inet_lookup_established(struct net *net,
221                                   struct inet_hashinfo *hashinfo,
222                                   const __be32 saddr, const __be16 sport,
223                                   const __be32 daddr, const u16 hnum,
224                                   const int dif)
225 {
226         INET_ADDR_COOKIE(acookie, saddr, daddr)
227         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
228         struct sock *sk;
229         const struct hlist_nulls_node *node;
230         /* Optimize here for direct hit, only listening connections can
231          * have wildcards anyways.
232          */
233         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
234         unsigned int slot = hash & hashinfo->ehash_mask;
235         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
236
237         rcu_read_lock();
238 begin:
239         sk_nulls_for_each_rcu(sk, node, &head->chain) {
240                 if (INET_MATCH(sk, net, hash, acookie,
241                                         saddr, daddr, ports, dif)) {
242                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
243                                 goto begintw;
244                         if (unlikely(!INET_MATCH(sk, net, hash, acookie,
245                                 saddr, daddr, ports, dif))) {
246                                 sock_put(sk);
247                                 goto begin;
248                         }
249                         goto out;
250                 }
251         }
252         /*
253          * if the nulls value we got at the end of this lookup is
254          * not the expected one, we must restart lookup.
255          * We probably met an item that was moved to another chain.
256          */
257         if (get_nulls_value(node) != slot)
258                 goto begin;
259
260 begintw:
261         /* Must check for a TIME_WAIT'er before going to listener hash. */
262         sk_nulls_for_each_rcu(sk, node, &head->twchain) {
263                 if (INET_TW_MATCH(sk, net, hash, acookie,
264                                         saddr, daddr, ports, dif)) {
265                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt))) {
266                                 sk = NULL;
267                                 goto out;
268                         }
269                         if (unlikely(!INET_TW_MATCH(sk, net, hash, acookie,
270                                  saddr, daddr, ports, dif))) {
271                                 sock_put(sk);
272                                 goto begintw;
273                         }
274                         goto out;
275                 }
276         }
277         /*
278          * if the nulls value we got at the end of this lookup is
279          * not the expected one, we must restart lookup.
280          * We probably met an item that was moved to another chain.
281          */
282         if (get_nulls_value(node) != slot)
283                 goto begintw;
284         sk = NULL;
285 out:
286         rcu_read_unlock();
287         return sk;
288 }
289 EXPORT_SYMBOL_GPL(__inet_lookup_established);
290
291 /* called with local bh disabled */
292 static int __inet_check_established(struct inet_timewait_death_row *death_row,
293                                     struct sock *sk, __u16 lport,
294                                     struct inet_timewait_sock **twp)
295 {
296         struct inet_hashinfo *hinfo = death_row->hashinfo;
297         struct inet_sock *inet = inet_sk(sk);
298         __be32 daddr = inet->inet_rcv_saddr;
299         __be32 saddr = inet->inet_daddr;
300         int dif = sk->sk_bound_dev_if;
301         INET_ADDR_COOKIE(acookie, saddr, daddr)
302         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
303         struct net *net = sock_net(sk);
304         unsigned int hash = inet_ehashfn(net, daddr, lport,
305                                          saddr, inet->inet_dport);
306         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
307         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
308         struct sock *sk2;
309         const struct hlist_nulls_node *node;
310         struct inet_timewait_sock *tw;
311         int twrefcnt = 0;
312
313         spin_lock(lock);
314
315         /* Check TIME-WAIT sockets first. */
316         sk_nulls_for_each(sk2, node, &head->twchain) {
317                 tw = inet_twsk(sk2);
318
319                 if (INET_TW_MATCH(sk2, net, hash, acookie,
320                                         saddr, daddr, ports, dif)) {
321                         if (twsk_unique(sk, sk2, twp))
322                                 goto unique;
323                         else
324                                 goto not_unique;
325                 }
326         }
327         tw = NULL;
328
329         /* And established part... */
330         sk_nulls_for_each(sk2, node, &head->chain) {
331                 if (INET_MATCH(sk2, net, hash, acookie,
332                                         saddr, daddr, ports, dif))
333                         goto not_unique;
334         }
335
336 unique:
337         /* Must record num and sport now. Otherwise we will see
338          * in hash table socket with a funny identity. */
339         inet->inet_num = lport;
340         inet->inet_sport = htons(lport);
341         sk->sk_hash = hash;
342         WARN_ON(!sk_unhashed(sk));
343         __sk_nulls_add_node_rcu(sk, &head->chain);
344         if (tw) {
345                 twrefcnt = inet_twsk_unhash(tw);
346                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
347         }
348         spin_unlock(lock);
349         if (twrefcnt)
350                 inet_twsk_put(tw);
351         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
352
353         if (twp) {
354                 *twp = tw;
355         } else if (tw) {
356                 /* Silly. Should hash-dance instead... */
357                 inet_twsk_deschedule(tw, death_row);
358
359                 inet_twsk_put(tw);
360         }
361         return 0;
362
363 not_unique:
364         spin_unlock(lock);
365         return -EADDRNOTAVAIL;
366 }
367
368 static inline u32 inet_sk_port_offset(const struct sock *sk)
369 {
370         const struct inet_sock *inet = inet_sk(sk);
371         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
372                                           inet->inet_daddr,
373                                           inet->inet_dport);
374 }
375
376 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
377 {
378         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
379         struct hlist_nulls_head *list;
380         spinlock_t *lock;
381         struct inet_ehash_bucket *head;
382         int twrefcnt = 0;
383
384         WARN_ON(!sk_unhashed(sk));
385
386         sk->sk_hash = inet_sk_ehashfn(sk);
387         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
388         list = &head->chain;
389         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
390
391         spin_lock(lock);
392         __sk_nulls_add_node_rcu(sk, list);
393         if (tw) {
394                 WARN_ON(sk->sk_hash != tw->tw_hash);
395                 twrefcnt = inet_twsk_unhash(tw);
396         }
397         spin_unlock(lock);
398         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
399         return twrefcnt;
400 }
401 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
402
403 static void __inet_hash(struct sock *sk)
404 {
405         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
406         struct inet_listen_hashbucket *ilb;
407
408         if (sk->sk_state != TCP_LISTEN) {
409                 __inet_hash_nolisten(sk, NULL);
410                 return;
411         }
412
413         WARN_ON(!sk_unhashed(sk));
414         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
415
416         spin_lock(&ilb->lock);
417         __sk_nulls_add_node_rcu(sk, &ilb->head);
418         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
419         spin_unlock(&ilb->lock);
420 }
421
422 void inet_hash(struct sock *sk)
423 {
424         if (sk->sk_state != TCP_CLOSE) {
425                 local_bh_disable();
426                 __inet_hash(sk);
427                 local_bh_enable();
428         }
429 }
430 EXPORT_SYMBOL_GPL(inet_hash);
431
432 void inet_unhash(struct sock *sk)
433 {
434         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
435         spinlock_t *lock;
436         int done;
437
438         if (sk_unhashed(sk))
439                 return;
440
441         if (sk->sk_state == TCP_LISTEN)
442                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
443         else
444                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
445
446         spin_lock_bh(lock);
447         done =__sk_nulls_del_node_init_rcu(sk);
448         if (done)
449                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
450         spin_unlock_bh(lock);
451 }
452 EXPORT_SYMBOL_GPL(inet_unhash);
453
454 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
455                 struct sock *sk, u32 port_offset,
456                 int (*check_established)(struct inet_timewait_death_row *,
457                         struct sock *, __u16, struct inet_timewait_sock **),
458                 int (*hash)(struct sock *sk, struct inet_timewait_sock *twp))
459 {
460         struct inet_hashinfo *hinfo = death_row->hashinfo;
461         const unsigned short snum = inet_sk(sk)->inet_num;
462         struct inet_bind_hashbucket *head;
463         struct inet_bind_bucket *tb;
464         int ret;
465         struct net *net = sock_net(sk);
466         int twrefcnt = 1;
467
468         if (!snum) {
469                 int i, remaining, low, high, port;
470                 static u32 hint;
471                 u32 offset = hint + port_offset;
472                 struct hlist_node *node;
473                 struct inet_timewait_sock *tw = NULL;
474
475                 inet_get_local_port_range(&low, &high);
476                 remaining = (high - low) + 1;
477
478                 local_bh_disable();
479                 for (i = 1; i <= remaining; i++) {
480                         port = low + (i + offset) % remaining;
481                         if (inet_is_reserved_local_port(port))
482                                 continue;
483                         head = &hinfo->bhash[inet_bhashfn(net, port,
484                                         hinfo->bhash_size)];
485                         spin_lock(&head->lock);
486
487                         /* Does not bother with rcv_saddr checks,
488                          * because the established check is already
489                          * unique enough.
490                          */
491                         inet_bind_bucket_for_each(tb, node, &head->chain) {
492                                 if (net_eq(ib_net(tb), net) &&
493                                     tb->port == port) {
494                                         if (tb->fastreuse >= 0)
495                                                 goto next_port;
496                                         WARN_ON(hlist_empty(&tb->owners));
497                                         if (!check_established(death_row, sk,
498                                                                 port, &tw))
499                                                 goto ok;
500                                         goto next_port;
501                                 }
502                         }
503
504                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
505                                         net, head, port);
506                         if (!tb) {
507                                 spin_unlock(&head->lock);
508                                 break;
509                         }
510                         tb->fastreuse = -1;
511                         goto ok;
512
513                 next_port:
514                         spin_unlock(&head->lock);
515                 }
516                 local_bh_enable();
517
518                 return -EADDRNOTAVAIL;
519
520 ok:
521                 hint += i;
522
523                 /* Head lock still held and bh's disabled */
524                 inet_bind_hash(sk, tb, port);
525                 if (sk_unhashed(sk)) {
526                         inet_sk(sk)->inet_sport = htons(port);
527                         twrefcnt += hash(sk, tw);
528                 }
529                 if (tw)
530                         twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
531                 spin_unlock(&head->lock);
532
533                 if (tw) {
534                         inet_twsk_deschedule(tw, death_row);
535                         while (twrefcnt) {
536                                 twrefcnt--;
537                                 inet_twsk_put(tw);
538                         }
539                 }
540
541                 ret = 0;
542                 goto out;
543         }
544
545         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
546         tb  = inet_csk(sk)->icsk_bind_hash;
547         spin_lock_bh(&head->lock);
548         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
549                 hash(sk, NULL);
550                 spin_unlock_bh(&head->lock);
551                 return 0;
552         } else {
553                 spin_unlock(&head->lock);
554                 /* No definite answer... Walk to established hash table */
555                 ret = check_established(death_row, sk, snum, NULL);
556 out:
557                 local_bh_enable();
558                 return ret;
559         }
560 }
561
562 /*
563  * Bind a port for a connect operation and hash it.
564  */
565 int inet_hash_connect(struct inet_timewait_death_row *death_row,
566                       struct sock *sk)
567 {
568         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
569                         __inet_check_established, __inet_hash_nolisten);
570 }
571 EXPORT_SYMBOL_GPL(inet_hash_connect);
572
573 void inet_hashinfo_init(struct inet_hashinfo *h)
574 {
575         int i;
576
577         atomic_set(&h->bsockets, 0);
578         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
579                 spin_lock_init(&h->listening_hash[i].lock);
580                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
581                                       i + LISTENING_NULLS_BASE);
582                 }
583 }
584 EXPORT_SYMBOL_GPL(inet_hashinfo_init);