Merge branch 'for-linus' of git://oss.sgi.com:8090/xfs/xfs-2.6
[linux-2.6] / 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 inet_bind_hashbucket *head,
32                                                  const unsigned short snum)
33 {
34         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
35
36         if (tb != NULL) {
37                 tb->port      = snum;
38                 tb->fastreuse = 0;
39                 INIT_HLIST_HEAD(&tb->owners);
40                 hlist_add_head(&tb->node, &head->chain);
41         }
42         return tb;
43 }
44
45 /*
46  * Caller must hold hashbucket lock for this tb with local BH disabled
47  */
48 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
49 {
50         if (hlist_empty(&tb->owners)) {
51                 __hlist_del(&tb->node);
52                 kmem_cache_free(cachep, tb);
53         }
54 }
55
56 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
57                     const unsigned short snum)
58 {
59         inet_sk(sk)->num = snum;
60         sk_add_bind_node(sk, &tb->owners);
61         inet_csk(sk)->icsk_bind_hash = tb;
62 }
63
64 /*
65  * Get rid of any references to a local port held by the given sock.
66  */
67 static void __inet_put_port(struct inet_hashinfo *hashinfo, struct sock *sk)
68 {
69         const int bhash = inet_bhashfn(inet_sk(sk)->num, hashinfo->bhash_size);
70         struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
71         struct inet_bind_bucket *tb;
72
73         spin_lock(&head->lock);
74         tb = inet_csk(sk)->icsk_bind_hash;
75         __sk_del_bind_node(sk);
76         inet_csk(sk)->icsk_bind_hash = NULL;
77         inet_sk(sk)->num = 0;
78         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
79         spin_unlock(&head->lock);
80 }
81
82 void inet_put_port(struct inet_hashinfo *hashinfo, struct sock *sk)
83 {
84         local_bh_disable();
85         __inet_put_port(hashinfo, sk);
86         local_bh_enable();
87 }
88
89 EXPORT_SYMBOL(inet_put_port);
90
91 /*
92  * This lock without WQ_FLAG_EXCLUSIVE is good on UP and it can be very bad on SMP.
93  * Look, when several writers sleep and reader wakes them up, all but one
94  * immediately hit write lock and grab all the cpus. Exclusive sleep solves
95  * this, _but_ remember, it adds useless work on UP machines (wake up each
96  * exclusive lock release). It should be ifdefed really.
97  */
98 void inet_listen_wlock(struct inet_hashinfo *hashinfo)
99 {
100         write_lock(&hashinfo->lhash_lock);
101
102         if (atomic_read(&hashinfo->lhash_users)) {
103                 DEFINE_WAIT(wait);
104
105                 for (;;) {
106                         prepare_to_wait_exclusive(&hashinfo->lhash_wait,
107                                                   &wait, TASK_UNINTERRUPTIBLE);
108                         if (!atomic_read(&hashinfo->lhash_users))
109                                 break;
110                         write_unlock_bh(&hashinfo->lhash_lock);
111                         schedule();
112                         write_lock_bh(&hashinfo->lhash_lock);
113                 }
114
115                 finish_wait(&hashinfo->lhash_wait, &wait);
116         }
117 }
118
119 EXPORT_SYMBOL(inet_listen_wlock);
120
121 /*
122  * Don't inline this cruft. Here are some nice properties to exploit here. The
123  * BSD API does not allow a listening sock to specify the remote port nor the
124  * remote address for the connection. So always assume those are both
125  * wildcarded during the search since they can never be otherwise.
126  */
127 static struct sock *inet_lookup_listener_slow(const struct hlist_head *head,
128                                               const __be32 daddr,
129                                               const unsigned short hnum,
130                                               const int dif)
131 {
132         struct sock *result = NULL, *sk;
133         const struct hlist_node *node;
134         int hiscore = -1;
135
136         sk_for_each(sk, node, head) {
137                 const struct inet_sock *inet = inet_sk(sk);
138
139                 if (inet->num == hnum && !ipv6_only_sock(sk)) {
140                         const __be32 rcv_saddr = inet->rcv_saddr;
141                         int score = sk->sk_family == PF_INET ? 1 : 0;
142
143                         if (rcv_saddr) {
144                                 if (rcv_saddr != daddr)
145                                         continue;
146                                 score += 2;
147                         }
148                         if (sk->sk_bound_dev_if) {
149                                 if (sk->sk_bound_dev_if != dif)
150                                         continue;
151                                 score += 2;
152                         }
153                         if (score == 5)
154                                 return sk;
155                         if (score > hiscore) {
156                                 hiscore = score;
157                                 result  = sk;
158                         }
159                 }
160         }
161         return result;
162 }
163
164 /* Optimize the common listener case. */
165 struct sock *__inet_lookup_listener(struct inet_hashinfo *hashinfo,
166                                     const __be32 daddr, const unsigned short hnum,
167                                     const int dif)
168 {
169         struct sock *sk = NULL;
170         const struct hlist_head *head;
171
172         read_lock(&hashinfo->lhash_lock);
173         head = &hashinfo->listening_hash[inet_lhashfn(hnum)];
174         if (!hlist_empty(head)) {
175                 const struct inet_sock *inet = inet_sk((sk = __sk_head(head)));
176
177                 if (inet->num == hnum && !sk->sk_node.next &&
178                     (!inet->rcv_saddr || inet->rcv_saddr == daddr) &&
179                     (sk->sk_family == PF_INET || !ipv6_only_sock(sk)) &&
180                     !sk->sk_bound_dev_if)
181                         goto sherry_cache;
182                 sk = inet_lookup_listener_slow(head, daddr, hnum, dif);
183         }
184         if (sk) {
185 sherry_cache:
186                 sock_hold(sk);
187         }
188         read_unlock(&hashinfo->lhash_lock);
189         return sk;
190 }
191 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
192
193 /* called with local bh disabled */
194 static int __inet_check_established(struct inet_timewait_death_row *death_row,
195                                     struct sock *sk, __u16 lport,
196                                     struct inet_timewait_sock **twp)
197 {
198         struct inet_hashinfo *hinfo = death_row->hashinfo;
199         struct inet_sock *inet = inet_sk(sk);
200         __be32 daddr = inet->rcv_saddr;
201         __be32 saddr = inet->daddr;
202         int dif = sk->sk_bound_dev_if;
203         INET_ADDR_COOKIE(acookie, saddr, daddr)
204         const __portpair ports = INET_COMBINED_PORTS(inet->dport, lport);
205         unsigned int hash = inet_ehashfn(daddr, lport, saddr, inet->dport);
206         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
207         struct sock *sk2;
208         const struct hlist_node *node;
209         struct inet_timewait_sock *tw;
210
211         prefetch(head->chain.first);
212         write_lock(&head->lock);
213
214         /* Check TIME-WAIT sockets first. */
215         sk_for_each(sk2, node, &head->twchain) {
216                 tw = inet_twsk(sk2);
217
218                 if (INET_TW_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif)) {
219                         if (twsk_unique(sk, sk2, twp))
220                                 goto unique;
221                         else
222                                 goto not_unique;
223                 }
224         }
225         tw = NULL;
226
227         /* And established part... */
228         sk_for_each(sk2, node, &head->chain) {
229                 if (INET_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif))
230                         goto not_unique;
231         }
232
233 unique:
234         /* Must record num and sport now. Otherwise we will see
235          * in hash table socket with a funny identity. */
236         inet->num = lport;
237         inet->sport = htons(lport);
238         sk->sk_hash = hash;
239         BUG_TRAP(sk_unhashed(sk));
240         __sk_add_node(sk, &head->chain);
241         sock_prot_inc_use(sk->sk_prot);
242         write_unlock(&head->lock);
243
244         if (twp) {
245                 *twp = tw;
246                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
247         } else if (tw) {
248                 /* Silly. Should hash-dance instead... */
249                 inet_twsk_deschedule(tw, death_row);
250                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
251
252                 inet_twsk_put(tw);
253         }
254
255         return 0;
256
257 not_unique:
258         write_unlock(&head->lock);
259         return -EADDRNOTAVAIL;
260 }
261
262 static inline u32 inet_sk_port_offset(const struct sock *sk)
263 {
264         const struct inet_sock *inet = inet_sk(sk);
265         return secure_ipv4_port_ephemeral(inet->rcv_saddr, inet->daddr,
266                                           inet->dport);
267 }
268
269 /*
270  * Bind a port for a connect operation and hash it.
271  */
272 int inet_hash_connect(struct inet_timewait_death_row *death_row,
273                       struct sock *sk)
274 {
275         struct inet_hashinfo *hinfo = death_row->hashinfo;
276         const unsigned short snum = inet_sk(sk)->num;
277         struct inet_bind_hashbucket *head;
278         struct inet_bind_bucket *tb;
279         int ret;
280
281         if (!snum) {
282                 int i, remaining, low, high, port;
283                 static u32 hint;
284                 u32 offset = hint + inet_sk_port_offset(sk);
285                 struct hlist_node *node;
286                 struct inet_timewait_sock *tw = NULL;
287
288                 inet_get_local_port_range(&low, &high);
289                 remaining = (high - low) + 1;
290
291                 local_bh_disable();
292                 for (i = 1; i <= remaining; i++) {
293                         port = low + (i + offset) % remaining;
294                         head = &hinfo->bhash[inet_bhashfn(port, hinfo->bhash_size)];
295                         spin_lock(&head->lock);
296
297                         /* Does not bother with rcv_saddr checks,
298                          * because the established check is already
299                          * unique enough.
300                          */
301                         inet_bind_bucket_for_each(tb, node, &head->chain) {
302                                 if (tb->port == port) {
303                                         BUG_TRAP(!hlist_empty(&tb->owners));
304                                         if (tb->fastreuse >= 0)
305                                                 goto next_port;
306                                         if (!__inet_check_established(death_row,
307                                                                       sk, port,
308                                                                       &tw))
309                                                 goto ok;
310                                         goto next_port;
311                                 }
312                         }
313
314                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, head, port);
315                         if (!tb) {
316                                 spin_unlock(&head->lock);
317                                 break;
318                         }
319                         tb->fastreuse = -1;
320                         goto ok;
321
322                 next_port:
323                         spin_unlock(&head->lock);
324                 }
325                 local_bh_enable();
326
327                 return -EADDRNOTAVAIL;
328
329 ok:
330                 hint += i;
331
332                 /* Head lock still held and bh's disabled */
333                 inet_bind_hash(sk, tb, port);
334                 if (sk_unhashed(sk)) {
335                         inet_sk(sk)->sport = htons(port);
336                         __inet_hash(hinfo, sk, 0);
337                 }
338                 spin_unlock(&head->lock);
339
340                 if (tw) {
341                         inet_twsk_deschedule(tw, death_row);
342                         inet_twsk_put(tw);
343                 }
344
345                 ret = 0;
346                 goto out;
347         }
348
349         head = &hinfo->bhash[inet_bhashfn(snum, hinfo->bhash_size)];
350         tb  = inet_csk(sk)->icsk_bind_hash;
351         spin_lock_bh(&head->lock);
352         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
353                 __inet_hash(hinfo, sk, 0);
354                 spin_unlock_bh(&head->lock);
355                 return 0;
356         } else {
357                 spin_unlock(&head->lock);
358                 /* No definite answer... Walk to established hash table */
359                 ret = __inet_check_established(death_row, sk, snum, NULL);
360 out:
361                 local_bh_enable();
362                 return ret;
363         }
364 }
365
366 EXPORT_SYMBOL_GPL(inet_hash_connect);