net: Convert TCP/DCCP listening hash tables to use RCU
[linux-2.6.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                 INIT_HLIST_HEAD(&tb->owners);
42                 hlist_add_head(&tb->node, &head->chain);
43         }
44         return tb;
45 }
46
47 /*
48  * Caller must hold hashbucket lock for this tb with local BH disabled
49  */
50 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
51 {
52         if (hlist_empty(&tb->owners)) {
53                 __hlist_del(&tb->node);
54                 release_net(ib_net(tb));
55                 kmem_cache_free(cachep, tb);
56         }
57 }
58
59 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
60                     const unsigned short snum)
61 {
62         inet_sk(sk)->num = snum;
63         sk_add_bind_node(sk, &tb->owners);
64         inet_csk(sk)->icsk_bind_hash = tb;
65 }
66
67 /*
68  * Get rid of any references to a local port held by the given sock.
69  */
70 static void __inet_put_port(struct sock *sk)
71 {
72         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
73         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->num,
74                         hashinfo->bhash_size);
75         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
76         struct inet_bind_bucket *tb;
77
78         spin_lock(&head->lock);
79         tb = inet_csk(sk)->icsk_bind_hash;
80         __sk_del_bind_node(sk);
81         inet_csk(sk)->icsk_bind_hash = NULL;
82         inet_sk(sk)->num = 0;
83         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
84         spin_unlock(&head->lock);
85 }
86
87 void inet_put_port(struct sock *sk)
88 {
89         local_bh_disable();
90         __inet_put_port(sk);
91         local_bh_enable();
92 }
93
94 EXPORT_SYMBOL(inet_put_port);
95
96 void __inet_inherit_port(struct sock *sk, struct sock *child)
97 {
98         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
99         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(child)->num,
100                         table->bhash_size);
101         struct inet_bind_hashbucket *head = &table->bhash[bhash];
102         struct inet_bind_bucket *tb;
103
104         spin_lock(&head->lock);
105         tb = inet_csk(sk)->icsk_bind_hash;
106         sk_add_bind_node(child, &tb->owners);
107         inet_csk(child)->icsk_bind_hash = tb;
108         spin_unlock(&head->lock);
109 }
110
111 EXPORT_SYMBOL_GPL(__inet_inherit_port);
112
113 static inline int compute_score(struct sock *sk, struct net *net,
114                                 const unsigned short hnum, const __be32 daddr,
115                                 const int dif)
116 {
117         int score = -1;
118         struct inet_sock *inet = inet_sk(sk);
119
120         if (net_eq(sock_net(sk), net) && inet->num == hnum &&
121                         !ipv6_only_sock(sk)) {
122                 __be32 rcv_saddr = inet->rcv_saddr;
123                 score = sk->sk_family == PF_INET ? 1 : 0;
124                 if (rcv_saddr) {
125                         if (rcv_saddr != daddr)
126                                 return -1;
127                         score += 2;
128                 }
129                 if (sk->sk_bound_dev_if) {
130                         if (sk->sk_bound_dev_if != dif)
131                                 return -1;
132                         score += 2;
133                 }
134         }
135         return score;
136 }
137
138 /*
139  * Don't inline this cruft. Here are some nice properties to exploit here. The
140  * BSD API does not allow a listening sock to specify the remote port nor the
141  * remote address for the connection. So always assume those are both
142  * wildcarded during the search since they can never be otherwise.
143  */
144
145
146 struct sock *__inet_lookup_listener(struct net *net,
147                                     struct inet_hashinfo *hashinfo,
148                                     const __be32 daddr, const unsigned short hnum,
149                                     const int dif)
150 {
151         struct sock *sk, *result;
152         struct hlist_nulls_node *node;
153         unsigned int hash = inet_lhashfn(net, hnum);
154         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
155         int score, hiscore;
156
157         rcu_read_lock();
158 begin:
159         result = NULL;
160         hiscore = -1;
161         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
162                 score = compute_score(sk, net, hnum, daddr, dif);
163                 if (score > hiscore) {
164                         result = sk;
165                         hiscore = score;
166                 }
167         }
168         /*
169          * if the nulls value we got at the end of this lookup is
170          * not the expected one, we must restart lookup.
171          * We probably met an item that was moved to another chain.
172          */
173         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
174                 goto begin;
175         if (result) {
176                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
177                         result = NULL;
178                 else if (unlikely(compute_score(result, net, hnum, daddr,
179                                   dif) < hiscore)) {
180                         sock_put(result);
181                         goto begin;
182                 }
183         }
184         rcu_read_unlock();
185         return result;
186 }
187 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
188
189 struct sock * __inet_lookup_established(struct net *net,
190                                   struct inet_hashinfo *hashinfo,
191                                   const __be32 saddr, const __be16 sport,
192                                   const __be32 daddr, const u16 hnum,
193                                   const int dif)
194 {
195         INET_ADDR_COOKIE(acookie, saddr, daddr)
196         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
197         struct sock *sk;
198         const struct hlist_nulls_node *node;
199         /* Optimize here for direct hit, only listening connections can
200          * have wildcards anyways.
201          */
202         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
203         unsigned int slot = hash & (hashinfo->ehash_size - 1);
204         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
205
206         rcu_read_lock();
207 begin:
208         sk_nulls_for_each_rcu(sk, node, &head->chain) {
209                 if (INET_MATCH(sk, net, hash, acookie,
210                                         saddr, daddr, ports, dif)) {
211                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
212                                 goto begintw;
213                         if (unlikely(!INET_MATCH(sk, net, hash, acookie,
214                                 saddr, daddr, ports, dif))) {
215                                 sock_put(sk);
216                                 goto begin;
217                         }
218                         goto out;
219                 }
220         }
221         /*
222          * if the nulls value we got at the end of this lookup is
223          * not the expected one, we must restart lookup.
224          * We probably met an item that was moved to another chain.
225          */
226         if (get_nulls_value(node) != slot)
227                 goto begin;
228
229 begintw:
230         /* Must check for a TIME_WAIT'er before going to listener hash. */
231         sk_nulls_for_each_rcu(sk, node, &head->twchain) {
232                 if (INET_TW_MATCH(sk, net, hash, acookie,
233                                         saddr, daddr, ports, dif)) {
234                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt))) {
235                                 sk = NULL;
236                                 goto out;
237                         }
238                         if (unlikely(!INET_TW_MATCH(sk, net, hash, acookie,
239                                  saddr, daddr, ports, dif))) {
240                                 sock_put(sk);
241                                 goto begintw;
242                         }
243                         goto out;
244                 }
245         }
246         /*
247          * if the nulls value we got at the end of this lookup is
248          * not the expected one, we must restart lookup.
249          * We probably met an item that was moved to another chain.
250          */
251         if (get_nulls_value(node) != slot)
252                 goto begintw;
253         sk = NULL;
254 out:
255         rcu_read_unlock();
256         return sk;
257 }
258 EXPORT_SYMBOL_GPL(__inet_lookup_established);
259
260 /* called with local bh disabled */
261 static int __inet_check_established(struct inet_timewait_death_row *death_row,
262                                     struct sock *sk, __u16 lport,
263                                     struct inet_timewait_sock **twp)
264 {
265         struct inet_hashinfo *hinfo = death_row->hashinfo;
266         struct inet_sock *inet = inet_sk(sk);
267         __be32 daddr = inet->rcv_saddr;
268         __be32 saddr = inet->daddr;
269         int dif = sk->sk_bound_dev_if;
270         INET_ADDR_COOKIE(acookie, saddr, daddr)
271         const __portpair ports = INET_COMBINED_PORTS(inet->dport, lport);
272         struct net *net = sock_net(sk);
273         unsigned int hash = inet_ehashfn(net, daddr, lport, saddr, inet->dport);
274         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
275         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
276         struct sock *sk2;
277         const struct hlist_nulls_node *node;
278         struct inet_timewait_sock *tw;
279
280         spin_lock(lock);
281
282         /* Check TIME-WAIT sockets first. */
283         sk_nulls_for_each(sk2, node, &head->twchain) {
284                 tw = inet_twsk(sk2);
285
286                 if (INET_TW_MATCH(sk2, net, hash, acookie,
287                                         saddr, daddr, ports, dif)) {
288                         if (twsk_unique(sk, sk2, twp))
289                                 goto unique;
290                         else
291                                 goto not_unique;
292                 }
293         }
294         tw = NULL;
295
296         /* And established part... */
297         sk_nulls_for_each(sk2, node, &head->chain) {
298                 if (INET_MATCH(sk2, net, hash, acookie,
299                                         saddr, daddr, ports, dif))
300                         goto not_unique;
301         }
302
303 unique:
304         /* Must record num and sport now. Otherwise we will see
305          * in hash table socket with a funny identity. */
306         inet->num = lport;
307         inet->sport = htons(lport);
308         sk->sk_hash = hash;
309         WARN_ON(!sk_unhashed(sk));
310         __sk_nulls_add_node_rcu(sk, &head->chain);
311         spin_unlock(lock);
312         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
313
314         if (twp) {
315                 *twp = tw;
316                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
317         } else if (tw) {
318                 /* Silly. Should hash-dance instead... */
319                 inet_twsk_deschedule(tw, death_row);
320                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
321
322                 inet_twsk_put(tw);
323         }
324
325         return 0;
326
327 not_unique:
328         spin_unlock(lock);
329         return -EADDRNOTAVAIL;
330 }
331
332 static inline u32 inet_sk_port_offset(const struct sock *sk)
333 {
334         const struct inet_sock *inet = inet_sk(sk);
335         return secure_ipv4_port_ephemeral(inet->rcv_saddr, inet->daddr,
336                                           inet->dport);
337 }
338
339 void __inet_hash_nolisten(struct sock *sk)
340 {
341         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
342         struct hlist_nulls_head *list;
343         spinlock_t *lock;
344         struct inet_ehash_bucket *head;
345
346         WARN_ON(!sk_unhashed(sk));
347
348         sk->sk_hash = inet_sk_ehashfn(sk);
349         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
350         list = &head->chain;
351         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
352
353         spin_lock(lock);
354         __sk_nulls_add_node_rcu(sk, list);
355         spin_unlock(lock);
356         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
357 }
358 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
359
360 static void __inet_hash(struct sock *sk)
361 {
362         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
363         struct inet_listen_hashbucket *ilb;
364
365         if (sk->sk_state != TCP_LISTEN) {
366                 __inet_hash_nolisten(sk);
367                 return;
368         }
369
370         WARN_ON(!sk_unhashed(sk));
371         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
372
373         spin_lock(&ilb->lock);
374         __sk_nulls_add_node_rcu(sk, &ilb->head);
375         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
376         spin_unlock(&ilb->lock);
377 }
378
379 void inet_hash(struct sock *sk)
380 {
381         if (sk->sk_state != TCP_CLOSE) {
382                 local_bh_disable();
383                 __inet_hash(sk);
384                 local_bh_enable();
385         }
386 }
387 EXPORT_SYMBOL_GPL(inet_hash);
388
389 void inet_unhash(struct sock *sk)
390 {
391         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
392         spinlock_t *lock;
393         int done;
394
395         if (sk_unhashed(sk))
396                 return;
397
398         if (sk->sk_state == TCP_LISTEN)
399                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
400         else
401                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
402
403         spin_lock_bh(lock);
404         done =__sk_nulls_del_node_init_rcu(sk);
405         spin_unlock_bh(lock);
406         if (done)
407                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
408 }
409 EXPORT_SYMBOL_GPL(inet_unhash);
410
411 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
412                 struct sock *sk, u32 port_offset,
413                 int (*check_established)(struct inet_timewait_death_row *,
414                         struct sock *, __u16, struct inet_timewait_sock **),
415                 void (*hash)(struct sock *sk))
416 {
417         struct inet_hashinfo *hinfo = death_row->hashinfo;
418         const unsigned short snum = inet_sk(sk)->num;
419         struct inet_bind_hashbucket *head;
420         struct inet_bind_bucket *tb;
421         int ret;
422         struct net *net = sock_net(sk);
423
424         if (!snum) {
425                 int i, remaining, low, high, port;
426                 static u32 hint;
427                 u32 offset = hint + port_offset;
428                 struct hlist_node *node;
429                 struct inet_timewait_sock *tw = NULL;
430
431                 inet_get_local_port_range(&low, &high);
432                 remaining = (high - low) + 1;
433
434                 local_bh_disable();
435                 for (i = 1; i <= remaining; i++) {
436                         port = low + (i + offset) % remaining;
437                         head = &hinfo->bhash[inet_bhashfn(net, port,
438                                         hinfo->bhash_size)];
439                         spin_lock(&head->lock);
440
441                         /* Does not bother with rcv_saddr checks,
442                          * because the established check is already
443                          * unique enough.
444                          */
445                         inet_bind_bucket_for_each(tb, node, &head->chain) {
446                                 if (ib_net(tb) == net && tb->port == port) {
447                                         WARN_ON(hlist_empty(&tb->owners));
448                                         if (tb->fastreuse >= 0)
449                                                 goto next_port;
450                                         if (!check_established(death_row, sk,
451                                                                 port, &tw))
452                                                 goto ok;
453                                         goto next_port;
454                                 }
455                         }
456
457                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
458                                         net, head, port);
459                         if (!tb) {
460                                 spin_unlock(&head->lock);
461                                 break;
462                         }
463                         tb->fastreuse = -1;
464                         goto ok;
465
466                 next_port:
467                         spin_unlock(&head->lock);
468                 }
469                 local_bh_enable();
470
471                 return -EADDRNOTAVAIL;
472
473 ok:
474                 hint += i;
475
476                 /* Head lock still held and bh's disabled */
477                 inet_bind_hash(sk, tb, port);
478                 if (sk_unhashed(sk)) {
479                         inet_sk(sk)->sport = htons(port);
480                         hash(sk);
481                 }
482                 spin_unlock(&head->lock);
483
484                 if (tw) {
485                         inet_twsk_deschedule(tw, death_row);
486                         inet_twsk_put(tw);
487                 }
488
489                 ret = 0;
490                 goto out;
491         }
492
493         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
494         tb  = inet_csk(sk)->icsk_bind_hash;
495         spin_lock_bh(&head->lock);
496         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
497                 hash(sk);
498                 spin_unlock_bh(&head->lock);
499                 return 0;
500         } else {
501                 spin_unlock(&head->lock);
502                 /* No definite answer... Walk to established hash table */
503                 ret = check_established(death_row, sk, snum, NULL);
504 out:
505                 local_bh_enable();
506                 return ret;
507         }
508 }
509
510 /*
511  * Bind a port for a connect operation and hash it.
512  */
513 int inet_hash_connect(struct inet_timewait_death_row *death_row,
514                       struct sock *sk)
515 {
516         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
517                         __inet_check_established, __inet_hash_nolisten);
518 }
519
520 EXPORT_SYMBOL_GPL(inet_hash_connect);
521
522 void inet_hashinfo_init(struct inet_hashinfo *h)
523 {
524         int i;
525
526         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
527                 spin_lock_init(&h->listening_hash[i].lock);
528                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
529                                       i + LISTENING_NULLS_BASE);
530                 }
531 }
532
533 EXPORT_SYMBOL_GPL(inet_hashinfo_init);