xref: /linux/net/ipv4/inet_hashtables.c (revision f3d9478b2ce468c3115b02ecae7e975990697f15)
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/config.h>
17 #include <linux/module.h>
18 #include <linux/random.h>
19 #include <linux/sched.h>
20 #include <linux/slab.h>
21 #include <linux/wait.h>
22 
23 #include <net/inet_connection_sock.h>
24 #include <net/inet_hashtables.h>
25 #include <net/ip.h>
26 
27 /*
28  * Allocate and initialize a new local port bind bucket.
29  * The bindhash mutex for snum's hash chain must be held here.
30  */
31 struct inet_bind_bucket *inet_bind_bucket_create(kmem_cache_t *cachep,
32 						 struct inet_bind_hashbucket *head,
33 						 const unsigned short snum)
34 {
35 	struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, SLAB_ATOMIC);
36 
37 	if (tb != NULL) {
38 		tb->port      = snum;
39 		tb->fastreuse = 0;
40 		INIT_HLIST_HEAD(&tb->owners);
41 		hlist_add_head(&tb->node, &head->chain);
42 	}
43 	return tb;
44 }
45 
46 /*
47  * Caller must hold hashbucket lock for this tb with local BH disabled
48  */
49 void inet_bind_bucket_destroy(kmem_cache_t *cachep, struct inet_bind_bucket *tb)
50 {
51 	if (hlist_empty(&tb->owners)) {
52 		__hlist_del(&tb->node);
53 		kmem_cache_free(cachep, tb);
54 	}
55 }
56 
57 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
58 		    const unsigned short snum)
59 {
60 	inet_sk(sk)->num = snum;
61 	sk_add_bind_node(sk, &tb->owners);
62 	inet_csk(sk)->icsk_bind_hash = tb;
63 }
64 
65 /*
66  * Get rid of any references to a local port held by the given sock.
67  */
68 static void __inet_put_port(struct inet_hashinfo *hashinfo, struct sock *sk)
69 {
70 	const int bhash = inet_bhashfn(inet_sk(sk)->num, hashinfo->bhash_size);
71 	struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
72 	struct inet_bind_bucket *tb;
73 
74 	spin_lock(&head->lock);
75 	tb = inet_csk(sk)->icsk_bind_hash;
76 	__sk_del_bind_node(sk);
77 	inet_csk(sk)->icsk_bind_hash = NULL;
78 	inet_sk(sk)->num = 0;
79 	inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
80 	spin_unlock(&head->lock);
81 }
82 
83 void inet_put_port(struct inet_hashinfo *hashinfo, struct sock *sk)
84 {
85 	local_bh_disable();
86 	__inet_put_port(hashinfo, sk);
87 	local_bh_enable();
88 }
89 
90 EXPORT_SYMBOL(inet_put_port);
91 
92 /*
93  * This lock without WQ_FLAG_EXCLUSIVE is good on UP and it can be very bad on SMP.
94  * Look, when several writers sleep and reader wakes them up, all but one
95  * immediately hit write lock and grab all the cpus. Exclusive sleep solves
96  * this, _but_ remember, it adds useless work on UP machines (wake up each
97  * exclusive lock release). It should be ifdefed really.
98  */
99 void inet_listen_wlock(struct inet_hashinfo *hashinfo)
100 {
101 	write_lock(&hashinfo->lhash_lock);
102 
103 	if (atomic_read(&hashinfo->lhash_users)) {
104 		DEFINE_WAIT(wait);
105 
106 		for (;;) {
107 			prepare_to_wait_exclusive(&hashinfo->lhash_wait,
108 						  &wait, TASK_UNINTERRUPTIBLE);
109 			if (!atomic_read(&hashinfo->lhash_users))
110 				break;
111 			write_unlock_bh(&hashinfo->lhash_lock);
112 			schedule();
113 			write_lock_bh(&hashinfo->lhash_lock);
114 		}
115 
116 		finish_wait(&hashinfo->lhash_wait, &wait);
117 	}
118 }
119 
120 EXPORT_SYMBOL(inet_listen_wlock);
121 
122 /*
123  * Don't inline this cruft. Here are some nice properties to exploit here. The
124  * BSD API does not allow a listening sock to specify the remote port nor the
125  * remote address for the connection. So always assume those are both
126  * wildcarded during the search since they can never be otherwise.
127  */
128 struct sock *__inet_lookup_listener(const struct hlist_head *head, const u32 daddr,
129 				    const unsigned short hnum, const int dif)
130 {
131 	struct sock *result = NULL, *sk;
132 	const struct hlist_node *node;
133 	int hiscore = -1;
134 
135 	sk_for_each(sk, node, head) {
136 		const struct inet_sock *inet = inet_sk(sk);
137 
138 		if (inet->num == hnum && !ipv6_only_sock(sk)) {
139 			const __u32 rcv_saddr = inet->rcv_saddr;
140 			int score = sk->sk_family == PF_INET ? 1 : 0;
141 
142 			if (rcv_saddr) {
143 				if (rcv_saddr != daddr)
144 					continue;
145 				score += 2;
146 			}
147 			if (sk->sk_bound_dev_if) {
148 				if (sk->sk_bound_dev_if != dif)
149 					continue;
150 				score += 2;
151 			}
152 			if (score == 5)
153 				return sk;
154 			if (score > hiscore) {
155 				hiscore	= score;
156 				result	= sk;
157 			}
158 		}
159 	}
160 	return result;
161 }
162 
163 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
164 
165 /* called with local bh disabled */
166 static int __inet_check_established(struct inet_timewait_death_row *death_row,
167 				    struct sock *sk, __u16 lport,
168 				    struct inet_timewait_sock **twp)
169 {
170 	struct inet_hashinfo *hinfo = death_row->hashinfo;
171 	struct inet_sock *inet = inet_sk(sk);
172 	u32 daddr = inet->rcv_saddr;
173 	u32 saddr = inet->daddr;
174 	int dif = sk->sk_bound_dev_if;
175 	INET_ADDR_COOKIE(acookie, saddr, daddr)
176 	const __u32 ports = INET_COMBINED_PORTS(inet->dport, lport);
177 	unsigned int hash = inet_ehashfn(daddr, lport, saddr, inet->dport);
178 	struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
179 	struct sock *sk2;
180 	const struct hlist_node *node;
181 	struct inet_timewait_sock *tw;
182 
183 	prefetch(head->chain.first);
184 	write_lock(&head->lock);
185 
186 	/* Check TIME-WAIT sockets first. */
187 	sk_for_each(sk2, node, &(head + hinfo->ehash_size)->chain) {
188 		tw = inet_twsk(sk2);
189 
190 		if (INET_TW_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif)) {
191 			if (twsk_unique(sk, sk2, twp))
192 				goto unique;
193 			else
194 				goto not_unique;
195 		}
196 	}
197 	tw = NULL;
198 
199 	/* And established part... */
200 	sk_for_each(sk2, node, &head->chain) {
201 		if (INET_MATCH(sk2, hash, acookie, saddr, daddr, ports, dif))
202 			goto not_unique;
203 	}
204 
205 unique:
206 	/* Must record num and sport now. Otherwise we will see
207 	 * in hash table socket with a funny identity. */
208 	inet->num = lport;
209 	inet->sport = htons(lport);
210 	sk->sk_hash = hash;
211 	BUG_TRAP(sk_unhashed(sk));
212 	__sk_add_node(sk, &head->chain);
213 	sock_prot_inc_use(sk->sk_prot);
214 	write_unlock(&head->lock);
215 
216 	if (twp) {
217 		*twp = tw;
218 		NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
219 	} else if (tw) {
220 		/* Silly. Should hash-dance instead... */
221 		inet_twsk_deschedule(tw, death_row);
222 		NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
223 
224 		inet_twsk_put(tw);
225 	}
226 
227 	return 0;
228 
229 not_unique:
230 	write_unlock(&head->lock);
231 	return -EADDRNOTAVAIL;
232 }
233 
234 static inline u32 inet_sk_port_offset(const struct sock *sk)
235 {
236 	const struct inet_sock *inet = inet_sk(sk);
237 	return secure_ipv4_port_ephemeral(inet->rcv_saddr, inet->daddr,
238 					  inet->dport);
239 }
240 
241 /*
242  * Bind a port for a connect operation and hash it.
243  */
244 int inet_hash_connect(struct inet_timewait_death_row *death_row,
245 		      struct sock *sk)
246 {
247 	struct inet_hashinfo *hinfo = death_row->hashinfo;
248 	const unsigned short snum = inet_sk(sk)->num;
249  	struct inet_bind_hashbucket *head;
250  	struct inet_bind_bucket *tb;
251 	int ret;
252 
253  	if (!snum) {
254  		int low = sysctl_local_port_range[0];
255  		int high = sysctl_local_port_range[1];
256 		int range = high - low;
257  		int i;
258 		int port;
259 		static u32 hint;
260 		u32 offset = hint + inet_sk_port_offset(sk);
261 		struct hlist_node *node;
262  		struct inet_timewait_sock *tw = NULL;
263 
264  		local_bh_disable();
265 		for (i = 1; i <= range; i++) {
266 			port = low + (i + offset) % range;
267  			head = &hinfo->bhash[inet_bhashfn(port, hinfo->bhash_size)];
268  			spin_lock(&head->lock);
269 
270  			/* Does not bother with rcv_saddr checks,
271  			 * because the established check is already
272  			 * unique enough.
273  			 */
274 			inet_bind_bucket_for_each(tb, node, &head->chain) {
275  				if (tb->port == port) {
276  					BUG_TRAP(!hlist_empty(&tb->owners));
277  					if (tb->fastreuse >= 0)
278  						goto next_port;
279  					if (!__inet_check_established(death_row,
280 								      sk, port,
281 								      &tw))
282  						goto ok;
283  					goto next_port;
284  				}
285  			}
286 
287  			tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, head, port);
288  			if (!tb) {
289  				spin_unlock(&head->lock);
290  				break;
291  			}
292  			tb->fastreuse = -1;
293  			goto ok;
294 
295  		next_port:
296  			spin_unlock(&head->lock);
297  		}
298  		local_bh_enable();
299 
300  		return -EADDRNOTAVAIL;
301 
302 ok:
303 		hint += i;
304 
305  		/* Head lock still held and bh's disabled */
306  		inet_bind_hash(sk, tb, port);
307 		if (sk_unhashed(sk)) {
308  			inet_sk(sk)->sport = htons(port);
309  			__inet_hash(hinfo, sk, 0);
310  		}
311  		spin_unlock(&head->lock);
312 
313  		if (tw) {
314  			inet_twsk_deschedule(tw, death_row);
315  			inet_twsk_put(tw);
316  		}
317 
318 		ret = 0;
319 		goto out;
320  	}
321 
322  	head = &hinfo->bhash[inet_bhashfn(snum, hinfo->bhash_size)];
323  	tb  = inet_csk(sk)->icsk_bind_hash;
324 	spin_lock_bh(&head->lock);
325 	if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
326 		__inet_hash(hinfo, sk, 0);
327 		spin_unlock_bh(&head->lock);
328 		return 0;
329 	} else {
330 		spin_unlock(&head->lock);
331 		/* No definite answer... Walk to established hash table */
332 		ret = __inet_check_established(death_row, sk, snum, NULL);
333 out:
334 		local_bh_enable();
335 		return ret;
336 	}
337 }
338 
339 EXPORT_SYMBOL_GPL(inet_hash_connect);
340