tcp: Fix a connect() race with timewait sockets
[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                 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
103 EXPORT_SYMBOL(inet_put_port);
104
105 void __inet_inherit_port(struct sock *sk, struct sock *child)
106 {
107         struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
108         const int bhash = inet_bhashfn(sock_net(sk), inet_sk(child)->inet_num,
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         sk_add_bind_node(child, &tb->owners);
116         inet_csk(child)->icsk_bind_hash = tb;
117         spin_unlock(&head->lock);
118 }
119
120 EXPORT_SYMBOL_GPL(__inet_inherit_port);
121
122 static inline int compute_score(struct sock *sk, struct net *net,
123                                 const unsigned short hnum, const __be32 daddr,
124                                 const int dif)
125 {
126         int score = -1;
127         struct inet_sock *inet = inet_sk(sk);
128
129         if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
130                         !ipv6_only_sock(sk)) {
131                 __be32 rcv_saddr = inet->inet_rcv_saddr;
132                 score = sk->sk_family == PF_INET ? 1 : 0;
133                 if (rcv_saddr) {
134                         if (rcv_saddr != daddr)
135                                 return -1;
136                         score += 2;
137                 }
138                 if (sk->sk_bound_dev_if) {
139                         if (sk->sk_bound_dev_if != dif)
140                                 return -1;
141                         score += 2;
142                 }
143         }
144         return score;
145 }
146
147 /*
148  * Don't inline this cruft. Here are some nice properties to exploit here. The
149  * BSD API does not allow a listening sock to specify the remote port nor the
150  * remote address for the connection. So always assume those are both
151  * wildcarded during the search since they can never be otherwise.
152  */
153
154
155 struct sock *__inet_lookup_listener(struct net *net,
156                                     struct inet_hashinfo *hashinfo,
157                                     const __be32 daddr, const unsigned short hnum,
158                                     const int dif)
159 {
160         struct sock *sk, *result;
161         struct hlist_nulls_node *node;
162         unsigned int hash = inet_lhashfn(net, hnum);
163         struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
164         int score, hiscore;
165
166         rcu_read_lock();
167 begin:
168         result = NULL;
169         hiscore = -1;
170         sk_nulls_for_each_rcu(sk, node, &ilb->head) {
171                 score = compute_score(sk, net, hnum, daddr, dif);
172                 if (score > hiscore) {
173                         result = sk;
174                         hiscore = score;
175                 }
176         }
177         /*
178          * if the nulls value we got at the end of this lookup is
179          * not the expected one, we must restart lookup.
180          * We probably met an item that was moved to another chain.
181          */
182         if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
183                 goto begin;
184         if (result) {
185                 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
186                         result = NULL;
187                 else if (unlikely(compute_score(result, net, hnum, daddr,
188                                   dif) < hiscore)) {
189                         sock_put(result);
190                         goto begin;
191                 }
192         }
193         rcu_read_unlock();
194         return result;
195 }
196 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
197
198 struct sock * __inet_lookup_established(struct net *net,
199                                   struct inet_hashinfo *hashinfo,
200                                   const __be32 saddr, const __be16 sport,
201                                   const __be32 daddr, const u16 hnum,
202                                   const int dif)
203 {
204         INET_ADDR_COOKIE(acookie, saddr, daddr)
205         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
206         struct sock *sk;
207         const struct hlist_nulls_node *node;
208         /* Optimize here for direct hit, only listening connections can
209          * have wildcards anyways.
210          */
211         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
212         unsigned int slot = hash & hashinfo->ehash_mask;
213         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
214
215         rcu_read_lock();
216 begin:
217         sk_nulls_for_each_rcu(sk, node, &head->chain) {
218                 if (INET_MATCH(sk, net, hash, acookie,
219                                         saddr, daddr, ports, dif)) {
220                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
221                                 goto begintw;
222                         if (unlikely(!INET_MATCH(sk, net, hash, acookie,
223                                 saddr, daddr, ports, dif))) {
224                                 sock_put(sk);
225                                 goto begin;
226                         }
227                         goto out;
228                 }
229         }
230         /*
231          * if the nulls value we got at the end of this lookup is
232          * not the expected one, we must restart lookup.
233          * We probably met an item that was moved to another chain.
234          */
235         if (get_nulls_value(node) != slot)
236                 goto begin;
237
238 begintw:
239         /* Must check for a TIME_WAIT'er before going to listener hash. */
240         sk_nulls_for_each_rcu(sk, node, &head->twchain) {
241                 if (INET_TW_MATCH(sk, net, hash, acookie,
242                                         saddr, daddr, ports, dif)) {
243                         if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt))) {
244                                 sk = NULL;
245                                 goto out;
246                         }
247                         if (unlikely(!INET_TW_MATCH(sk, net, hash, acookie,
248                                  saddr, daddr, ports, dif))) {
249                                 sock_put(sk);
250                                 goto begintw;
251                         }
252                         goto out;
253                 }
254         }
255         /*
256          * if the nulls value we got at the end of this lookup is
257          * not the expected one, we must restart lookup.
258          * We probably met an item that was moved to another chain.
259          */
260         if (get_nulls_value(node) != slot)
261                 goto begintw;
262         sk = NULL;
263 out:
264         rcu_read_unlock();
265         return sk;
266 }
267 EXPORT_SYMBOL_GPL(__inet_lookup_established);
268
269 /* called with local bh disabled */
270 static int __inet_check_established(struct inet_timewait_death_row *death_row,
271                                     struct sock *sk, __u16 lport,
272                                     struct inet_timewait_sock **twp)
273 {
274         struct inet_hashinfo *hinfo = death_row->hashinfo;
275         struct inet_sock *inet = inet_sk(sk);
276         __be32 daddr = inet->inet_rcv_saddr;
277         __be32 saddr = inet->inet_daddr;
278         int dif = sk->sk_bound_dev_if;
279         INET_ADDR_COOKIE(acookie, saddr, daddr)
280         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
281         struct net *net = sock_net(sk);
282         unsigned int hash = inet_ehashfn(net, daddr, lport,
283                                          saddr, inet->inet_dport);
284         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
285         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
286         struct sock *sk2;
287         const struct hlist_nulls_node *node;
288         struct inet_timewait_sock *tw;
289         int twrefcnt = 0;
290
291         spin_lock(lock);
292
293         /* Check TIME-WAIT sockets first. */
294         sk_nulls_for_each(sk2, node, &head->twchain) {
295                 tw = inet_twsk(sk2);
296
297                 if (INET_TW_MATCH(sk2, net, hash, acookie,
298                                         saddr, daddr, ports, dif)) {
299                         if (twsk_unique(sk, sk2, twp))
300                                 goto unique;
301                         else
302                                 goto not_unique;
303                 }
304         }
305         tw = NULL;
306
307         /* And established part... */
308         sk_nulls_for_each(sk2, node, &head->chain) {
309                 if (INET_MATCH(sk2, net, hash, acookie,
310                                         saddr, daddr, ports, dif))
311                         goto not_unique;
312         }
313
314 unique:
315         /* Must record num and sport now. Otherwise we will see
316          * in hash table socket with a funny identity. */
317         inet->inet_num = lport;
318         inet->inet_sport = htons(lport);
319         sk->sk_hash = hash;
320         WARN_ON(!sk_unhashed(sk));
321         __sk_nulls_add_node_rcu(sk, &head->chain);
322         if (tw) {
323                 twrefcnt = inet_twsk_unhash(tw);
324                 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
325         }
326         spin_unlock(lock);
327         if (twrefcnt)
328                 inet_twsk_put(tw);
329         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
330
331         if (twp) {
332                 *twp = tw;
333         } else if (tw) {
334                 /* Silly. Should hash-dance instead... */
335                 inet_twsk_deschedule(tw, death_row);
336
337                 inet_twsk_put(tw);
338         }
339         return 0;
340
341 not_unique:
342         spin_unlock(lock);
343         return -EADDRNOTAVAIL;
344 }
345
346 static inline u32 inet_sk_port_offset(const struct sock *sk)
347 {
348         const struct inet_sock *inet = inet_sk(sk);
349         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
350                                           inet->inet_daddr,
351                                           inet->inet_dport);
352 }
353
354 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
355 {
356         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
357         struct hlist_nulls_head *list;
358         spinlock_t *lock;
359         struct inet_ehash_bucket *head;
360         int twrefcnt = 0;
361
362         WARN_ON(!sk_unhashed(sk));
363
364         sk->sk_hash = inet_sk_ehashfn(sk);
365         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
366         list = &head->chain;
367         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
368
369         spin_lock(lock);
370         __sk_nulls_add_node_rcu(sk, list);
371         if (tw) {
372                 WARN_ON(sk->sk_hash != tw->tw_hash);
373                 twrefcnt = inet_twsk_unhash(tw);
374         }
375         spin_unlock(lock);
376         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
377         return twrefcnt;
378 }
379 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
380
381 static void __inet_hash(struct sock *sk)
382 {
383         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
384         struct inet_listen_hashbucket *ilb;
385
386         if (sk->sk_state != TCP_LISTEN) {
387                 __inet_hash_nolisten(sk, NULL);
388                 return;
389         }
390
391         WARN_ON(!sk_unhashed(sk));
392         ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
393
394         spin_lock(&ilb->lock);
395         __sk_nulls_add_node_rcu(sk, &ilb->head);
396         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
397         spin_unlock(&ilb->lock);
398 }
399
400 void inet_hash(struct sock *sk)
401 {
402         if (sk->sk_state != TCP_CLOSE) {
403                 local_bh_disable();
404                 __inet_hash(sk);
405                 local_bh_enable();
406         }
407 }
408 EXPORT_SYMBOL_GPL(inet_hash);
409
410 void inet_unhash(struct sock *sk)
411 {
412         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
413         spinlock_t *lock;
414         int done;
415
416         if (sk_unhashed(sk))
417                 return;
418
419         if (sk->sk_state == TCP_LISTEN)
420                 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
421         else
422                 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
423
424         spin_lock_bh(lock);
425         done =__sk_nulls_del_node_init_rcu(sk);
426         if (done)
427                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
428         spin_unlock_bh(lock);
429 }
430 EXPORT_SYMBOL_GPL(inet_unhash);
431
432 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
433                 struct sock *sk, u32 port_offset,
434                 int (*check_established)(struct inet_timewait_death_row *,
435                         struct sock *, __u16, struct inet_timewait_sock **),
436                 int (*hash)(struct sock *sk, struct inet_timewait_sock *twp))
437 {
438         struct inet_hashinfo *hinfo = death_row->hashinfo;
439         const unsigned short snum = inet_sk(sk)->inet_num;
440         struct inet_bind_hashbucket *head;
441         struct inet_bind_bucket *tb;
442         int ret;
443         struct net *net = sock_net(sk);
444         int twrefcnt = 1;
445
446         if (!snum) {
447                 int i, remaining, low, high, port;
448                 static u32 hint;
449                 u32 offset = hint + port_offset;
450                 struct hlist_node *node;
451                 struct inet_timewait_sock *tw = NULL;
452
453                 inet_get_local_port_range(&low, &high);
454                 remaining = (high - low) + 1;
455
456                 local_bh_disable();
457                 for (i = 1; i <= remaining; i++) {
458                         port = low + (i + offset) % remaining;
459                         head = &hinfo->bhash[inet_bhashfn(net, port,
460                                         hinfo->bhash_size)];
461                         spin_lock(&head->lock);
462
463                         /* Does not bother with rcv_saddr checks,
464                          * because the established check is already
465                          * unique enough.
466                          */
467                         inet_bind_bucket_for_each(tb, node, &head->chain) {
468                                 if (net_eq(ib_net(tb), net) &&
469                                     tb->port == port) {
470                                         if (tb->fastreuse >= 0)
471                                                 goto next_port;
472                                         WARN_ON(hlist_empty(&tb->owners));
473                                         if (!check_established(death_row, sk,
474                                                                 port, &tw))
475                                                 goto ok;
476                                         goto next_port;
477                                 }
478                         }
479
480                         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
481                                         net, head, port);
482                         if (!tb) {
483                                 spin_unlock(&head->lock);
484                                 break;
485                         }
486                         tb->fastreuse = -1;
487                         goto ok;
488
489                 next_port:
490                         spin_unlock(&head->lock);
491                 }
492                 local_bh_enable();
493
494                 return -EADDRNOTAVAIL;
495
496 ok:
497                 hint += i;
498
499                 /* Head lock still held and bh's disabled */
500                 inet_bind_hash(sk, tb, port);
501                 if (sk_unhashed(sk)) {
502                         inet_sk(sk)->inet_sport = htons(port);
503                         twrefcnt += hash(sk, tw);
504                 }
505                 spin_unlock(&head->lock);
506
507                 if (tw) {
508                         inet_twsk_deschedule(tw, death_row);
509                         while (twrefcnt) {
510                                 twrefcnt--;
511                                 inet_twsk_put(tw);
512                         }
513                 }
514
515                 ret = 0;
516                 goto out;
517         }
518
519         head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
520         tb  = inet_csk(sk)->icsk_bind_hash;
521         spin_lock_bh(&head->lock);
522         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
523                 hash(sk, NULL);
524                 spin_unlock_bh(&head->lock);
525                 return 0;
526         } else {
527                 spin_unlock(&head->lock);
528                 /* No definite answer... Walk to established hash table */
529                 ret = check_established(death_row, sk, snum, NULL);
530 out:
531                 local_bh_enable();
532                 return ret;
533         }
534 }
535
536 /*
537  * Bind a port for a connect operation and hash it.
538  */
539 int inet_hash_connect(struct inet_timewait_death_row *death_row,
540                       struct sock *sk)
541 {
542         return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
543                         __inet_check_established, __inet_hash_nolisten);
544 }
545
546 EXPORT_SYMBOL_GPL(inet_hash_connect);
547
548 void inet_hashinfo_init(struct inet_hashinfo *h)
549 {
550         int i;
551
552         atomic_set(&h->bsockets, 0);
553         for (i = 0; i < INET_LHTABLE_SIZE; i++) {
554                 spin_lock_init(&h->listening_hash[i].lock);
555                 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
556                                       i + LISTENING_NULLS_BASE);
557                 }
558 }
559
560 EXPORT_SYMBOL_GPL(inet_hashinfo_init);