xref: /linux/include/net/inet_hashtables.h (revision 367b8112fe2ea5c39a7bb4d263dcdd9b612fae18)
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/ip.h>
20 #include <linux/ipv6.h>
21 #include <linux/list.h>
22 #include <linux/slab.h>
23 #include <linux/socket.h>
24 #include <linux/spinlock.h>
25 #include <linux/types.h>
26 #include <linux/wait.h>
27 #include <linux/vmalloc.h>
28 
29 #include <net/inet_connection_sock.h>
30 #include <net/inet_sock.h>
31 #include <net/sock.h>
32 #include <net/route.h>
33 #include <net/tcp_states.h>
34 #include <net/netns/hash.h>
35 
36 #include <asm/atomic.h>
37 #include <asm/byteorder.h>
38 
39 /* This is for all connections with a full identity, no wildcards.
40  * One chain is dedicated to TIME_WAIT sockets.
41  * I'll experiment with dynamic table growth later.
42  */
43 struct inet_ehash_bucket {
44 	struct hlist_head chain;
45 	struct hlist_head twchain;
46 };
47 
48 /* There are a few simple rules, which allow for local port reuse by
49  * an application.  In essence:
50  *
51  *	1) Sockets bound to different interfaces may share a local port.
52  *	   Failing that, goto test 2.
53  *	2) If all sockets have sk->sk_reuse set, and none of them are in
54  *	   TCP_LISTEN state, the port may be shared.
55  *	   Failing that, goto test 3.
56  *	3) If all sockets are bound to a specific inet_sk(sk)->rcv_saddr local
57  *	   address, and none of them are the same, the port may be
58  *	   shared.
59  *	   Failing this, the port cannot be shared.
60  *
61  * The interesting point, is test #2.  This is what an FTP server does
62  * all day.  To optimize this case we use a specific flag bit defined
63  * below.  As we add sockets to a bind bucket list, we perform a
64  * check of: (newsk->sk_reuse && (newsk->sk_state != TCP_LISTEN))
65  * As long as all sockets added to a bind bucket pass this test,
66  * the flag bit will be set.
67  * The resulting situation is that tcp_v[46]_verify_bind() can just check
68  * for this flag bit, if it is set and the socket trying to bind has
69  * sk->sk_reuse set, we don't even have to walk the owners list at all,
70  * we return that it is ok to bind this socket to the requested local port.
71  *
72  * Sounds like a lot of work, but it is worth it.  In a more naive
73  * implementation (ie. current FreeBSD etc.) the entire list of ports
74  * must be walked for each data port opened by an ftp server.  Needless
75  * to say, this does not scale at all.  With a couple thousand FTP
76  * users logged onto your box, isn't it nice to know that new data
77  * ports are created in O(1) time?  I thought so. ;-)	-DaveM
78  */
79 struct inet_bind_bucket {
80 	struct net		*ib_net;
81 	unsigned short		port;
82 	signed short		fastreuse;
83 	struct hlist_node	node;
84 	struct hlist_head	owners;
85 };
86 
87 #define inet_bind_bucket_for_each(tb, node, head) \
88 	hlist_for_each_entry(tb, node, head, node)
89 
90 struct inet_bind_hashbucket {
91 	spinlock_t		lock;
92 	struct hlist_head	chain;
93 };
94 
95 /* This is for listening sockets, thus all sockets which possess wildcards. */
96 #define INET_LHTABLE_SIZE	32	/* Yes, really, this is all you need. */
97 
98 struct inet_hashinfo {
99 	/* This is for sockets with full identity only.  Sockets here will
100 	 * always be without wildcards and will have the following invariant:
101 	 *
102 	 *          TCP_ESTABLISHED <= sk->sk_state < TCP_CLOSE
103 	 *
104 	 * TIME_WAIT sockets use a separate chain (twchain).
105 	 */
106 	struct inet_ehash_bucket	*ehash;
107 	rwlock_t			*ehash_locks;
108 	unsigned int			ehash_size;
109 	unsigned int			ehash_locks_mask;
110 
111 	/* Ok, let's try this, I give up, we do need a local binding
112 	 * TCP hash as well as the others for fast bind/connect.
113 	 */
114 	struct inet_bind_hashbucket	*bhash;
115 
116 	unsigned int			bhash_size;
117 	/* Note : 4 bytes padding on 64 bit arches */
118 
119 	/* All sockets in TCP_LISTEN state will be in here.  This is the only
120 	 * table where wildcard'd TCP sockets can exist.  Hash function here
121 	 * is just local port number.
122 	 */
123 	struct hlist_head		listening_hash[INET_LHTABLE_SIZE];
124 
125 	/* All the above members are written once at bootup and
126 	 * never written again _or_ are predominantly read-access.
127 	 *
128 	 * Now align to a new cache line as all the following members
129 	 * are often dirty.
130 	 */
131 	rwlock_t			lhash_lock ____cacheline_aligned;
132 	atomic_t			lhash_users;
133 	wait_queue_head_t		lhash_wait;
134 	struct kmem_cache			*bind_bucket_cachep;
135 };
136 
137 static inline struct inet_ehash_bucket *inet_ehash_bucket(
138 	struct inet_hashinfo *hashinfo,
139 	unsigned int hash)
140 {
141 	return &hashinfo->ehash[hash & (hashinfo->ehash_size - 1)];
142 }
143 
144 static inline rwlock_t *inet_ehash_lockp(
145 	struct inet_hashinfo *hashinfo,
146 	unsigned int hash)
147 {
148 	return &hashinfo->ehash_locks[hash & hashinfo->ehash_locks_mask];
149 }
150 
151 static inline int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
152 {
153 	unsigned int i, size = 256;
154 #if defined(CONFIG_PROVE_LOCKING)
155 	unsigned int nr_pcpus = 2;
156 #else
157 	unsigned int nr_pcpus = num_possible_cpus();
158 #endif
159 	if (nr_pcpus >= 4)
160 		size = 512;
161 	if (nr_pcpus >= 8)
162 		size = 1024;
163 	if (nr_pcpus >= 16)
164 		size = 2048;
165 	if (nr_pcpus >= 32)
166 		size = 4096;
167 	if (sizeof(rwlock_t) != 0) {
168 #ifdef CONFIG_NUMA
169 		if (size * sizeof(rwlock_t) > PAGE_SIZE)
170 			hashinfo->ehash_locks = vmalloc(size * sizeof(rwlock_t));
171 		else
172 #endif
173 		hashinfo->ehash_locks =	kmalloc(size * sizeof(rwlock_t),
174 						GFP_KERNEL);
175 		if (!hashinfo->ehash_locks)
176 			return ENOMEM;
177 		for (i = 0; i < size; i++)
178 			rwlock_init(&hashinfo->ehash_locks[i]);
179 	}
180 	hashinfo->ehash_locks_mask = size - 1;
181 	return 0;
182 }
183 
184 static inline void inet_ehash_locks_free(struct inet_hashinfo *hashinfo)
185 {
186 	if (hashinfo->ehash_locks) {
187 #ifdef CONFIG_NUMA
188 		unsigned int size = (hashinfo->ehash_locks_mask + 1) *
189 							sizeof(rwlock_t);
190 		if (size > PAGE_SIZE)
191 			vfree(hashinfo->ehash_locks);
192 		else
193 #endif
194 		kfree(hashinfo->ehash_locks);
195 		hashinfo->ehash_locks = NULL;
196 	}
197 }
198 
199 extern struct inet_bind_bucket *
200 		    inet_bind_bucket_create(struct kmem_cache *cachep,
201 					    struct net *net,
202 					    struct inet_bind_hashbucket *head,
203 					    const unsigned short snum);
204 extern void inet_bind_bucket_destroy(struct kmem_cache *cachep,
205 				     struct inet_bind_bucket *tb);
206 
207 static inline int inet_bhashfn(struct net *net,
208 		const __u16 lport, const int bhash_size)
209 {
210 	return (lport + net_hash_mix(net)) & (bhash_size - 1);
211 }
212 
213 extern void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
214 			   const unsigned short snum);
215 
216 /* These can have wildcards, don't try too hard. */
217 static inline int inet_lhashfn(struct net *net, const unsigned short num)
218 {
219 	return (num + net_hash_mix(net)) & (INET_LHTABLE_SIZE - 1);
220 }
221 
222 static inline int inet_sk_listen_hashfn(const struct sock *sk)
223 {
224 	return inet_lhashfn(sock_net(sk), inet_sk(sk)->num);
225 }
226 
227 /* Caller must disable local BH processing. */
228 extern void __inet_inherit_port(struct sock *sk, struct sock *child);
229 
230 extern void inet_put_port(struct sock *sk);
231 
232 extern void inet_listen_wlock(struct inet_hashinfo *hashinfo);
233 
234 /*
235  * - We may sleep inside this lock.
236  * - If sleeping is not required (or called from BH),
237  *   use plain read_(un)lock(&inet_hashinfo.lhash_lock).
238  */
239 static inline void inet_listen_lock(struct inet_hashinfo *hashinfo)
240 {
241 	/* read_lock synchronizes to candidates to writers */
242 	read_lock(&hashinfo->lhash_lock);
243 	atomic_inc(&hashinfo->lhash_users);
244 	read_unlock(&hashinfo->lhash_lock);
245 }
246 
247 static inline void inet_listen_unlock(struct inet_hashinfo *hashinfo)
248 {
249 	if (atomic_dec_and_test(&hashinfo->lhash_users))
250 		wake_up(&hashinfo->lhash_wait);
251 }
252 
253 extern void __inet_hash_nolisten(struct sock *sk);
254 extern void inet_hash(struct sock *sk);
255 extern void inet_unhash(struct sock *sk);
256 
257 extern struct sock *__inet_lookup_listener(struct net *net,
258 					   struct inet_hashinfo *hashinfo,
259 					   const __be32 daddr,
260 					   const unsigned short hnum,
261 					   const int dif);
262 
263 static inline struct sock *inet_lookup_listener(struct net *net,
264 		struct inet_hashinfo *hashinfo,
265 		__be32 daddr, __be16 dport, int dif)
266 {
267 	return __inet_lookup_listener(net, hashinfo, daddr, ntohs(dport), dif);
268 }
269 
270 /* Socket demux engine toys. */
271 /* What happens here is ugly; there's a pair of adjacent fields in
272    struct inet_sock; __be16 dport followed by __u16 num.  We want to
273    search by pair, so we combine the keys into a single 32bit value
274    and compare with 32bit value read from &...->dport.  Let's at least
275    make sure that it's not mixed with anything else...
276    On 64bit targets we combine comparisons with pair of adjacent __be32
277    fields in the same way.
278 */
279 typedef __u32 __bitwise __portpair;
280 #ifdef __BIG_ENDIAN
281 #define INET_COMBINED_PORTS(__sport, __dport) \
282 	((__force __portpair)(((__force __u32)(__be16)(__sport) << 16) | (__u32)(__dport)))
283 #else /* __LITTLE_ENDIAN */
284 #define INET_COMBINED_PORTS(__sport, __dport) \
285 	((__force __portpair)(((__u32)(__dport) << 16) | (__force __u32)(__be16)(__sport)))
286 #endif
287 
288 #if (BITS_PER_LONG == 64)
289 typedef __u64 __bitwise __addrpair;
290 #ifdef __BIG_ENDIAN
291 #define INET_ADDR_COOKIE(__name, __saddr, __daddr) \
292 	const __addrpair __name = (__force __addrpair) ( \
293 				   (((__force __u64)(__be32)(__saddr)) << 32) | \
294 				   ((__force __u64)(__be32)(__daddr)));
295 #else /* __LITTLE_ENDIAN */
296 #define INET_ADDR_COOKIE(__name, __saddr, __daddr) \
297 	const __addrpair __name = (__force __addrpair) ( \
298 				   (((__force __u64)(__be32)(__daddr)) << 32) | \
299 				   ((__force __u64)(__be32)(__saddr)));
300 #endif /* __BIG_ENDIAN */
301 #define INET_MATCH(__sk, __net, __hash, __cookie, __saddr, __daddr, __ports, __dif)\
302 	(((__sk)->sk_hash == (__hash)) && sock_net((__sk)) == (__net)	&&	\
303 	 ((*((__addrpair *)&(inet_sk(__sk)->daddr))) == (__cookie))	&&	\
304 	 ((*((__portpair *)&(inet_sk(__sk)->dport))) == (__ports))	&&	\
305 	 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
306 #define INET_TW_MATCH(__sk, __net, __hash, __cookie, __saddr, __daddr, __ports, __dif)\
307 	(((__sk)->sk_hash == (__hash)) && sock_net((__sk)) == (__net)	&&	\
308 	 ((*((__addrpair *)&(inet_twsk(__sk)->tw_daddr))) == (__cookie)) &&	\
309 	 ((*((__portpair *)&(inet_twsk(__sk)->tw_dport))) == (__ports)) &&	\
310 	 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
311 #else /* 32-bit arch */
312 #define INET_ADDR_COOKIE(__name, __saddr, __daddr)
313 #define INET_MATCH(__sk, __net, __hash, __cookie, __saddr, __daddr, __ports, __dif)	\
314 	(((__sk)->sk_hash == (__hash)) && sock_net((__sk)) == (__net)	&&	\
315 	 (inet_sk(__sk)->daddr		== (__saddr))		&&	\
316 	 (inet_sk(__sk)->rcv_saddr	== (__daddr))		&&	\
317 	 ((*((__portpair *)&(inet_sk(__sk)->dport))) == (__ports))	&&	\
318 	 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
319 #define INET_TW_MATCH(__sk, __net, __hash,__cookie, __saddr, __daddr, __ports, __dif)	\
320 	(((__sk)->sk_hash == (__hash)) && sock_net((__sk)) == (__net)	&&	\
321 	 (inet_twsk(__sk)->tw_daddr	== (__saddr))		&&	\
322 	 (inet_twsk(__sk)->tw_rcv_saddr	== (__daddr))		&&	\
323 	 ((*((__portpair *)&(inet_twsk(__sk)->tw_dport))) == (__ports)) &&	\
324 	 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
325 #endif /* 64-bit arch */
326 
327 /*
328  * Sockets in TCP_CLOSE state are _always_ taken out of the hash, so we need
329  * not check it for lookups anymore, thanks Alexey. -DaveM
330  *
331  * Local BH must be disabled here.
332  */
333 extern struct sock * __inet_lookup_established(struct net *net,
334 		struct inet_hashinfo *hashinfo,
335 		const __be32 saddr, const __be16 sport,
336 		const __be32 daddr, const u16 hnum, const int dif);
337 
338 static inline struct sock *
339 	inet_lookup_established(struct net *net, struct inet_hashinfo *hashinfo,
340 				const __be32 saddr, const __be16 sport,
341 				const __be32 daddr, const __be16 dport,
342 				const int dif)
343 {
344 	return __inet_lookup_established(net, hashinfo, saddr, sport, daddr,
345 					 ntohs(dport), dif);
346 }
347 
348 static inline struct sock *__inet_lookup(struct net *net,
349 					 struct inet_hashinfo *hashinfo,
350 					 const __be32 saddr, const __be16 sport,
351 					 const __be32 daddr, const __be16 dport,
352 					 const int dif)
353 {
354 	u16 hnum = ntohs(dport);
355 	struct sock *sk = __inet_lookup_established(net, hashinfo,
356 				saddr, sport, daddr, hnum, dif);
357 
358 	return sk ? : __inet_lookup_listener(net, hashinfo, daddr, hnum, dif);
359 }
360 
361 static inline struct sock *inet_lookup(struct net *net,
362 				       struct inet_hashinfo *hashinfo,
363 				       const __be32 saddr, const __be16 sport,
364 				       const __be32 daddr, const __be16 dport,
365 				       const int dif)
366 {
367 	struct sock *sk;
368 
369 	local_bh_disable();
370 	sk = __inet_lookup(net, hashinfo, saddr, sport, daddr, dport, dif);
371 	local_bh_enable();
372 
373 	return sk;
374 }
375 
376 static inline struct sock *__inet_lookup_skb(struct inet_hashinfo *hashinfo,
377 					     struct sk_buff *skb,
378 					     const __be16 sport,
379 					     const __be16 dport)
380 {
381 	struct sock *sk;
382 	const struct iphdr *iph = ip_hdr(skb);
383 
384 	if (unlikely(sk = skb_steal_sock(skb)))
385 		return sk;
386 	else
387 		return __inet_lookup(dev_net(skb->dst->dev), hashinfo,
388 				     iph->saddr, sport,
389 				     iph->daddr, dport, inet_iif(skb));
390 }
391 
392 extern int __inet_hash_connect(struct inet_timewait_death_row *death_row,
393 		struct sock *sk, u32 port_offset,
394 		int (*check_established)(struct inet_timewait_death_row *,
395 			struct sock *, __u16, struct inet_timewait_sock **),
396 			       void (*hash)(struct sock *sk));
397 extern int inet_hash_connect(struct inet_timewait_death_row *death_row,
398 			     struct sock *sk);
399 #endif /* _INET_HASHTABLES_H */
400