xref: /linux/include/net/inet_hashtables.h (revision 643d1f7fe3aa12c8bdea6fa5b4ba874ff6dd601d)
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  * Authors:	Lotsa people, from code originally in tcp
7  *
8  *	This program is free software; you can redistribute it and/or
9  *      modify it under the terms of the GNU General Public License
10  *      as published by the Free Software Foundation; either version
11  *      2 of the License, or (at your option) any later version.
12  */
13 
14 #ifndef _INET_HASHTABLES_H
15 #define _INET_HASHTABLES_H
16 
17 
18 #include <linux/interrupt.h>
19 #include <linux/ipv6.h>
20 #include <linux/list.h>
21 #include <linux/slab.h>
22 #include <linux/socket.h>
23 #include <linux/spinlock.h>
24 #include <linux/types.h>
25 #include <linux/wait.h>
26 #include <linux/vmalloc.h>
27 
28 #include <net/inet_connection_sock.h>
29 #include <net/inet_sock.h>
30 #include <net/sock.h>
31 #include <net/tcp_states.h>
32 
33 #include <asm/atomic.h>
34 #include <asm/byteorder.h>
35 
36 /* This is for all connections with a full identity, no wildcards.
37  * One chain is dedicated to TIME_WAIT sockets.
38  * I'll experiment with dynamic table growth later.
39  */
40 struct inet_ehash_bucket {
41 	struct hlist_head chain;
42 	struct hlist_head twchain;
43 };
44 
45 /* There are a few simple rules, which allow for local port reuse by
46  * an application.  In essence:
47  *
48  *	1) Sockets bound to different interfaces may share a local port.
49  *	   Failing that, goto test 2.
50  *	2) If all sockets have sk->sk_reuse set, and none of them are in
51  *	   TCP_LISTEN state, the port may be shared.
52  *	   Failing that, goto test 3.
53  *	3) If all sockets are bound to a specific inet_sk(sk)->rcv_saddr local
54  *	   address, and none of them are the same, the port may be
55  *	   shared.
56  *	   Failing this, the port cannot be shared.
57  *
58  * The interesting point, is test #2.  This is what an FTP server does
59  * all day.  To optimize this case we use a specific flag bit defined
60  * below.  As we add sockets to a bind bucket list, we perform a
61  * check of: (newsk->sk_reuse && (newsk->sk_state != TCP_LISTEN))
62  * As long as all sockets added to a bind bucket pass this test,
63  * the flag bit will be set.
64  * The resulting situation is that tcp_v[46]_verify_bind() can just check
65  * for this flag bit, if it is set and the socket trying to bind has
66  * sk->sk_reuse set, we don't even have to walk the owners list at all,
67  * we return that it is ok to bind this socket to the requested local port.
68  *
69  * Sounds like a lot of work, but it is worth it.  In a more naive
70  * implementation (ie. current FreeBSD etc.) the entire list of ports
71  * must be walked for each data port opened by an ftp server.  Needless
72  * to say, this does not scale at all.  With a couple thousand FTP
73  * users logged onto your box, isn't it nice to know that new data
74  * ports are created in O(1) time?  I thought so. ;-)	-DaveM
75  */
76 struct inet_bind_bucket {
77 	struct net		*ib_net;
78 	unsigned short		port;
79 	signed short		fastreuse;
80 	struct hlist_node	node;
81 	struct hlist_head	owners;
82 };
83 
84 #define inet_bind_bucket_for_each(tb, node, head) \
85 	hlist_for_each_entry(tb, node, head, node)
86 
87 struct inet_bind_hashbucket {
88 	spinlock_t		lock;
89 	struct hlist_head	chain;
90 };
91 
92 /* This is for listening sockets, thus all sockets which possess wildcards. */
93 #define INET_LHTABLE_SIZE	32	/* Yes, really, this is all you need. */
94 
95 struct inet_hashinfo {
96 	/* This is for sockets with full identity only.  Sockets here will
97 	 * always be without wildcards and will have the following invariant:
98 	 *
99 	 *          TCP_ESTABLISHED <= sk->sk_state < TCP_CLOSE
100 	 *
101 	 * TIME_WAIT sockets use a separate chain (twchain).
102 	 */
103 	struct inet_ehash_bucket	*ehash;
104 	rwlock_t			*ehash_locks;
105 	unsigned int			ehash_size;
106 	unsigned int			ehash_locks_mask;
107 
108 	/* Ok, let's try this, I give up, we do need a local binding
109 	 * TCP hash as well as the others for fast bind/connect.
110 	 */
111 	struct inet_bind_hashbucket	*bhash;
112 
113 	unsigned int			bhash_size;
114 	/* Note : 4 bytes padding on 64 bit arches */
115 
116 	/* All sockets in TCP_LISTEN state will be in here.  This is the only
117 	 * table where wildcard'd TCP sockets can exist.  Hash function here
118 	 * is just local port number.
119 	 */
120 	struct hlist_head		listening_hash[INET_LHTABLE_SIZE];
121 
122 	/* All the above members are written once at bootup and
123 	 * never written again _or_ are predominantly read-access.
124 	 *
125 	 * Now align to a new cache line as all the following members
126 	 * are often dirty.
127 	 */
128 	rwlock_t			lhash_lock ____cacheline_aligned;
129 	atomic_t			lhash_users;
130 	wait_queue_head_t		lhash_wait;
131 	struct kmem_cache			*bind_bucket_cachep;
132 };
133 
134 static inline struct inet_ehash_bucket *inet_ehash_bucket(
135 	struct inet_hashinfo *hashinfo,
136 	unsigned int hash)
137 {
138 	return &hashinfo->ehash[hash & (hashinfo->ehash_size - 1)];
139 }
140 
141 static inline rwlock_t *inet_ehash_lockp(
142 	struct inet_hashinfo *hashinfo,
143 	unsigned int hash)
144 {
145 	return &hashinfo->ehash_locks[hash & hashinfo->ehash_locks_mask];
146 }
147 
148 static inline int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
149 {
150 	unsigned int i, size = 256;
151 #if defined(CONFIG_PROVE_LOCKING)
152 	unsigned int nr_pcpus = 2;
153 #else
154 	unsigned int nr_pcpus = num_possible_cpus();
155 #endif
156 	if (nr_pcpus >= 4)
157 		size = 512;
158 	if (nr_pcpus >= 8)
159 		size = 1024;
160 	if (nr_pcpus >= 16)
161 		size = 2048;
162 	if (nr_pcpus >= 32)
163 		size = 4096;
164 	if (sizeof(rwlock_t) != 0) {
165 #ifdef CONFIG_NUMA
166 		if (size * sizeof(rwlock_t) > PAGE_SIZE)
167 			hashinfo->ehash_locks = vmalloc(size * sizeof(rwlock_t));
168 		else
169 #endif
170 		hashinfo->ehash_locks =	kmalloc(size * sizeof(rwlock_t),
171 						GFP_KERNEL);
172 		if (!hashinfo->ehash_locks)
173 			return ENOMEM;
174 		for (i = 0; i < size; i++)
175 			rwlock_init(&hashinfo->ehash_locks[i]);
176 	}
177 	hashinfo->ehash_locks_mask = size - 1;
178 	return 0;
179 }
180 
181 static inline void inet_ehash_locks_free(struct inet_hashinfo *hashinfo)
182 {
183 	if (hashinfo->ehash_locks) {
184 #ifdef CONFIG_NUMA
185 		unsigned int size = (hashinfo->ehash_locks_mask + 1) *
186 							sizeof(rwlock_t);
187 		if (size > PAGE_SIZE)
188 			vfree(hashinfo->ehash_locks);
189 		else
190 #endif
191 		kfree(hashinfo->ehash_locks);
192 		hashinfo->ehash_locks = NULL;
193 	}
194 }
195 
196 extern struct inet_bind_bucket *
197 		    inet_bind_bucket_create(struct kmem_cache *cachep,
198 					    struct net *net,
199 					    struct inet_bind_hashbucket *head,
200 					    const unsigned short snum);
201 extern void inet_bind_bucket_destroy(struct kmem_cache *cachep,
202 				     struct inet_bind_bucket *tb);
203 
204 static inline int inet_bhashfn(const __u16 lport, const int bhash_size)
205 {
206 	return lport & (bhash_size - 1);
207 }
208 
209 extern void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
210 			   const unsigned short snum);
211 
212 /* These can have wildcards, don't try too hard. */
213 static inline int inet_lhashfn(const unsigned short num)
214 {
215 	return num & (INET_LHTABLE_SIZE - 1);
216 }
217 
218 static inline int inet_sk_listen_hashfn(const struct sock *sk)
219 {
220 	return inet_lhashfn(inet_sk(sk)->num);
221 }
222 
223 /* Caller must disable local BH processing. */
224 static inline void __inet_inherit_port(struct inet_hashinfo *table,
225 				       struct sock *sk, struct sock *child)
226 {
227 	const int bhash = inet_bhashfn(inet_sk(child)->num, table->bhash_size);
228 	struct inet_bind_hashbucket *head = &table->bhash[bhash];
229 	struct inet_bind_bucket *tb;
230 
231 	spin_lock(&head->lock);
232 	tb = inet_csk(sk)->icsk_bind_hash;
233 	sk_add_bind_node(child, &tb->owners);
234 	inet_csk(child)->icsk_bind_hash = tb;
235 	spin_unlock(&head->lock);
236 }
237 
238 static inline void inet_inherit_port(struct inet_hashinfo *table,
239 				     struct sock *sk, struct sock *child)
240 {
241 	local_bh_disable();
242 	__inet_inherit_port(table, sk, child);
243 	local_bh_enable();
244 }
245 
246 extern void inet_put_port(struct inet_hashinfo *table, struct sock *sk);
247 
248 extern void inet_listen_wlock(struct inet_hashinfo *hashinfo);
249 
250 /*
251  * - We may sleep inside this lock.
252  * - If sleeping is not required (or called from BH),
253  *   use plain read_(un)lock(&inet_hashinfo.lhash_lock).
254  */
255 static inline void inet_listen_lock(struct inet_hashinfo *hashinfo)
256 {
257 	/* read_lock synchronizes to candidates to writers */
258 	read_lock(&hashinfo->lhash_lock);
259 	atomic_inc(&hashinfo->lhash_users);
260 	read_unlock(&hashinfo->lhash_lock);
261 }
262 
263 static inline void inet_listen_unlock(struct inet_hashinfo *hashinfo)
264 {
265 	if (atomic_dec_and_test(&hashinfo->lhash_users))
266 		wake_up(&hashinfo->lhash_wait);
267 }
268 
269 extern void __inet_hash(struct inet_hashinfo *hashinfo, struct sock *sk);
270 extern void __inet_hash_nolisten(struct inet_hashinfo *hinfo, struct sock *sk);
271 
272 static inline void inet_hash(struct inet_hashinfo *hashinfo, struct sock *sk)
273 {
274 	if (sk->sk_state != TCP_CLOSE) {
275 		local_bh_disable();
276 		__inet_hash(hashinfo, sk);
277 		local_bh_enable();
278 	}
279 }
280 
281 static inline void inet_unhash(struct inet_hashinfo *hashinfo, struct sock *sk)
282 {
283 	rwlock_t *lock;
284 
285 	if (sk_unhashed(sk))
286 		goto out;
287 
288 	if (sk->sk_state == TCP_LISTEN) {
289 		local_bh_disable();
290 		inet_listen_wlock(hashinfo);
291 		lock = &hashinfo->lhash_lock;
292 	} else {
293 		lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
294 		write_lock_bh(lock);
295 	}
296 
297 	if (__sk_del_node_init(sk))
298 		sock_prot_inuse_add(sk->sk_prot, -1);
299 	write_unlock_bh(lock);
300 out:
301 	if (sk->sk_state == TCP_LISTEN)
302 		wake_up(&hashinfo->lhash_wait);
303 }
304 
305 extern struct sock *__inet_lookup_listener(struct net *net,
306 					   struct inet_hashinfo *hashinfo,
307 					   const __be32 daddr,
308 					   const unsigned short hnum,
309 					   const int dif);
310 
311 static inline struct sock *inet_lookup_listener(struct net *net,
312 		struct inet_hashinfo *hashinfo,
313 		__be32 daddr, __be16 dport, int dif)
314 {
315 	return __inet_lookup_listener(net, hashinfo, daddr, ntohs(dport), dif);
316 }
317 
318 /* Socket demux engine toys. */
319 /* What happens here is ugly; there's a pair of adjacent fields in
320    struct inet_sock; __be16 dport followed by __u16 num.  We want to
321    search by pair, so we combine the keys into a single 32bit value
322    and compare with 32bit value read from &...->dport.  Let's at least
323    make sure that it's not mixed with anything else...
324    On 64bit targets we combine comparisons with pair of adjacent __be32
325    fields in the same way.
326 */
327 typedef __u32 __bitwise __portpair;
328 #ifdef __BIG_ENDIAN
329 #define INET_COMBINED_PORTS(__sport, __dport) \
330 	((__force __portpair)(((__force __u32)(__be16)(__sport) << 16) | (__u32)(__dport)))
331 #else /* __LITTLE_ENDIAN */
332 #define INET_COMBINED_PORTS(__sport, __dport) \
333 	((__force __portpair)(((__u32)(__dport) << 16) | (__force __u32)(__be16)(__sport)))
334 #endif
335 
336 #if (BITS_PER_LONG == 64)
337 typedef __u64 __bitwise __addrpair;
338 #ifdef __BIG_ENDIAN
339 #define INET_ADDR_COOKIE(__name, __saddr, __daddr) \
340 	const __addrpair __name = (__force __addrpair) ( \
341 				   (((__force __u64)(__be32)(__saddr)) << 32) | \
342 				   ((__force __u64)(__be32)(__daddr)));
343 #else /* __LITTLE_ENDIAN */
344 #define INET_ADDR_COOKIE(__name, __saddr, __daddr) \
345 	const __addrpair __name = (__force __addrpair) ( \
346 				   (((__force __u64)(__be32)(__daddr)) << 32) | \
347 				   ((__force __u64)(__be32)(__saddr)));
348 #endif /* __BIG_ENDIAN */
349 #define INET_MATCH(__sk, __net, __hash, __cookie, __saddr, __daddr, __ports, __dif)\
350 	(((__sk)->sk_hash == (__hash)) && ((__sk)->sk_net == (__net))	&&	\
351 	 ((*((__addrpair *)&(inet_sk(__sk)->daddr))) == (__cookie))	&&	\
352 	 ((*((__portpair *)&(inet_sk(__sk)->dport))) == (__ports))	&&	\
353 	 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
354 #define INET_TW_MATCH(__sk, __net, __hash, __cookie, __saddr, __daddr, __ports, __dif)\
355 	(((__sk)->sk_hash == (__hash)) && ((__sk)->sk_net == (__net))	&&	\
356 	 ((*((__addrpair *)&(inet_twsk(__sk)->tw_daddr))) == (__cookie)) &&	\
357 	 ((*((__portpair *)&(inet_twsk(__sk)->tw_dport))) == (__ports)) &&	\
358 	 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
359 #else /* 32-bit arch */
360 #define INET_ADDR_COOKIE(__name, __saddr, __daddr)
361 #define INET_MATCH(__sk, __net, __hash, __cookie, __saddr, __daddr, __ports, __dif)	\
362 	(((__sk)->sk_hash == (__hash)) && ((__sk)->sk_net == (__net))	&&	\
363 	 (inet_sk(__sk)->daddr		== (__saddr))		&&	\
364 	 (inet_sk(__sk)->rcv_saddr	== (__daddr))		&&	\
365 	 ((*((__portpair *)&(inet_sk(__sk)->dport))) == (__ports))	&&	\
366 	 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
367 #define INET_TW_MATCH(__sk, __net, __hash,__cookie, __saddr, __daddr, __ports, __dif)	\
368 	(((__sk)->sk_hash == (__hash)) && ((__sk)->sk_net == (__net))	&&	\
369 	 (inet_twsk(__sk)->tw_daddr	== (__saddr))		&&	\
370 	 (inet_twsk(__sk)->tw_rcv_saddr	== (__daddr))		&&	\
371 	 ((*((__portpair *)&(inet_twsk(__sk)->tw_dport))) == (__ports)) &&	\
372 	 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
373 #endif /* 64-bit arch */
374 
375 /*
376  * Sockets in TCP_CLOSE state are _always_ taken out of the hash, so we need
377  * not check it for lookups anymore, thanks Alexey. -DaveM
378  *
379  * Local BH must be disabled here.
380  */
381 extern struct sock * __inet_lookup_established(struct net *net,
382 		struct inet_hashinfo *hashinfo,
383 		const __be32 saddr, const __be16 sport,
384 		const __be32 daddr, const u16 hnum, const int dif);
385 
386 static inline struct sock *
387 	inet_lookup_established(struct net *net, struct inet_hashinfo *hashinfo,
388 				const __be32 saddr, const __be16 sport,
389 				const __be32 daddr, const __be16 dport,
390 				const int dif)
391 {
392 	return __inet_lookup_established(net, hashinfo, saddr, sport, daddr,
393 					 ntohs(dport), dif);
394 }
395 
396 static inline struct sock *__inet_lookup(struct net *net,
397 					 struct inet_hashinfo *hashinfo,
398 					 const __be32 saddr, const __be16 sport,
399 					 const __be32 daddr, const __be16 dport,
400 					 const int dif)
401 {
402 	u16 hnum = ntohs(dport);
403 	struct sock *sk = __inet_lookup_established(net, hashinfo,
404 				saddr, sport, daddr, hnum, dif);
405 
406 	return sk ? : __inet_lookup_listener(net, hashinfo, daddr, hnum, dif);
407 }
408 
409 static inline struct sock *inet_lookup(struct net *net,
410 				       struct inet_hashinfo *hashinfo,
411 				       const __be32 saddr, const __be16 sport,
412 				       const __be32 daddr, const __be16 dport,
413 				       const int dif)
414 {
415 	struct sock *sk;
416 
417 	local_bh_disable();
418 	sk = __inet_lookup(net, hashinfo, saddr, sport, daddr, dport, dif);
419 	local_bh_enable();
420 
421 	return sk;
422 }
423 
424 extern int __inet_hash_connect(struct inet_timewait_death_row *death_row,
425 		struct sock *sk,
426 		int (*check_established)(struct inet_timewait_death_row *,
427 			struct sock *, __u16, struct inet_timewait_sock **),
428 		void (*hash)(struct inet_hashinfo *, struct sock *));
429 extern int inet_hash_connect(struct inet_timewait_death_row *death_row,
430 			     struct sock *sk);
431 #endif /* _INET_HASHTABLES_H */
432