blob: bc68eced0105716dca6f04abbdd491a1cfb9e908 [file] [log] [blame]
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -07001/*
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
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -070016#include <linux/module.h>
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -080017#include <linux/random.h>
Arnaldo Carvalho de Melof3f05f72005-08-09 20:08:09 -070018#include <linux/sched.h>
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070019#include <linux/slab.h>
Arnaldo Carvalho de Melof3f05f72005-08-09 20:08:09 -070020#include <linux/wait.h>
Eric Dumazet095dc8e2015-05-26 07:55:34 -070021#include <linux/vmalloc.h>
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070022
Craig Gallekc125e802016-02-10 11:50:40 -050023#include <net/addrconf.h>
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -070024#include <net/inet_connection_sock.h>
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070025#include <net/inet_hashtables.h>
David S. Miller6e5714e2011-08-03 20:50:44 -070026#include <net/secure_seq.h>
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -080027#include <net/ip.h>
Craig Gallekc125e802016-02-10 11:50:40 -050028#include <net/sock_reuseport.h>
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070029
Eric Dumazet6eada012015-03-18 14:05:33 -070030static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
31 const __u16 lport, const __be32 faddr,
32 const __be16 fport)
Hannes Frederic Sowa65cd8032013-10-19 21:48:51 +020033{
Hannes Frederic Sowa1bbdcee2013-10-19 21:48:57 +020034 static u32 inet_ehash_secret __read_mostly;
35
36 net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
37
Hannes Frederic Sowa65cd8032013-10-19 21:48:51 +020038 return __inet_ehashfn(laddr, lport, faddr, fport,
39 inet_ehash_secret + net_hash_mix(net));
40}
41
Eric Dumazetd1e559d2015-03-18 14:05:35 -070042/* This function handles inet_sock, but also timewait and request sockets
43 * for IPv4/IPv6.
44 */
Eric Dumazet5b441f72015-03-18 14:05:34 -070045u32 sk_ehashfn(const struct sock *sk)
Hannes Frederic Sowa65cd8032013-10-19 21:48:51 +020046{
Eric Dumazetd1e559d2015-03-18 14:05:35 -070047#if IS_ENABLED(CONFIG_IPV6)
48 if (sk->sk_family == AF_INET6 &&
49 !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
50 return inet6_ehashfn(sock_net(sk),
51 &sk->sk_v6_rcv_saddr, sk->sk_num,
52 &sk->sk_v6_daddr, sk->sk_dport);
53#endif
Eric Dumazet5b441f72015-03-18 14:05:34 -070054 return inet_ehashfn(sock_net(sk),
55 sk->sk_rcv_saddr, sk->sk_num,
56 sk->sk_daddr, sk->sk_dport);
Hannes Frederic Sowa65cd8032013-10-19 21:48:51 +020057}
58
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070059/*
60 * Allocate and initialize a new local port bind bucket.
61 * The bindhash mutex for snum's hash chain must be held here.
62 */
Christoph Lametere18b8902006-12-06 20:33:20 -080063struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
Pavel Emelyanov941b1d22008-01-31 05:05:50 -080064 struct net *net,
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070065 struct inet_bind_hashbucket *head,
66 const unsigned short snum)
67{
Christoph Lameter54e6ecb2006-12-06 20:33:16 -080068 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070069
Ian Morris00db4122015-04-03 09:17:27 +010070 if (tb) {
Eric W. Biedermanefd7ef12015-03-11 23:04:08 -050071 write_pnet(&tb->ib_net, net);
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070072 tb->port = snum;
73 tb->fastreuse = 0;
Tom Herbertda5e3632013-01-22 09:50:24 +000074 tb->fastreuseport = 0;
Evgeniy Polyakova9d8f912009-01-19 16:46:02 -080075 tb->num_owners = 0;
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070076 INIT_HLIST_HEAD(&tb->owners);
77 hlist_add_head(&tb->node, &head->chain);
78 }
79 return tb;
80}
81
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070082/*
83 * Caller must hold hashbucket lock for this tb with local BH disabled
84 */
Christoph Lametere18b8902006-12-06 20:33:20 -080085void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
Arnaldo Carvalho de Melo77d8bf92005-08-09 20:00:51 -070086{
87 if (hlist_empty(&tb->owners)) {
88 __hlist_del(&tb->node);
89 kmem_cache_free(cachep, tb);
90 }
91}
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -070092
93void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
94 const unsigned short snum)
95{
Eric Dumazetc720c7e2009-10-15 06:30:45 +000096 inet_sk(sk)->inet_num = snum;
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -070097 sk_add_bind_node(sk, &tb->owners);
Evgeniy Polyakova9d8f912009-01-19 16:46:02 -080098 tb->num_owners++;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -070099 inet_csk(sk)->icsk_bind_hash = tb;
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -0700100}
101
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -0700102/*
103 * Get rid of any references to a local port held by the given sock.
104 */
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800105static void __inet_put_port(struct sock *sk)
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -0700106{
Pavel Emelyanov39d8cda2008-03-22 16:50:58 -0700107 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000108 const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
Pavel Emelyanov7f635ab2008-06-16 17:12:49 -0700109 hashinfo->bhash_size);
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -0700110 struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
111 struct inet_bind_bucket *tb;
112
113 spin_lock(&head->lock);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700114 tb = inet_csk(sk)->icsk_bind_hash;
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -0700115 __sk_del_bind_node(sk);
Evgeniy Polyakova9d8f912009-01-19 16:46:02 -0800116 tb->num_owners--;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700117 inet_csk(sk)->icsk_bind_hash = NULL;
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000118 inet_sk(sk)->inet_num = 0;
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -0700119 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
120 spin_unlock(&head->lock);
121}
122
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800123void inet_put_port(struct sock *sk)
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -0700124{
125 local_bh_disable();
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800126 __inet_put_port(sk);
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -0700127 local_bh_enable();
128}
Arnaldo Carvalho de Melo2d8c4ce2005-08-09 20:07:13 -0700129EXPORT_SYMBOL(inet_put_port);
Arnaldo Carvalho de Melof3f05f72005-08-09 20:08:09 -0700130
Eric Dumazet1ce31c92015-09-29 07:42:44 -0700131int __inet_inherit_port(const struct sock *sk, struct sock *child)
Pavel Emelyanov53083772008-04-17 23:18:15 -0700132{
133 struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
Balazs Scheidler093d2822010-10-21 13:06:43 +0200134 unsigned short port = inet_sk(child)->inet_num;
135 const int bhash = inet_bhashfn(sock_net(sk), port,
Pavel Emelyanov7f635ab2008-06-16 17:12:49 -0700136 table->bhash_size);
Pavel Emelyanov53083772008-04-17 23:18:15 -0700137 struct inet_bind_hashbucket *head = &table->bhash[bhash];
138 struct inet_bind_bucket *tb;
139
140 spin_lock(&head->lock);
141 tb = inet_csk(sk)->icsk_bind_hash;
Eric Dumazetc2f34a62015-10-14 05:58:38 -0700142 if (unlikely(!tb)) {
143 spin_unlock(&head->lock);
144 return -ENOENT;
145 }
Balazs Scheidler093d2822010-10-21 13:06:43 +0200146 if (tb->port != port) {
147 /* NOTE: using tproxy and redirecting skbs to a proxy
148 * on a different listener port breaks the assumption
149 * that the listener socket's icsk_bind_hash is the same
150 * as that of the child socket. We have to look up or
151 * create a new bind bucket for the child here. */
Sasha Levinb67bfe02013-02-27 17:06:00 -0800152 inet_bind_bucket_for_each(tb, &head->chain) {
Balazs Scheidler093d2822010-10-21 13:06:43 +0200153 if (net_eq(ib_net(tb), sock_net(sk)) &&
154 tb->port == port)
155 break;
156 }
Sasha Levinb67bfe02013-02-27 17:06:00 -0800157 if (!tb) {
Balazs Scheidler093d2822010-10-21 13:06:43 +0200158 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
159 sock_net(sk), head, port);
160 if (!tb) {
161 spin_unlock(&head->lock);
162 return -ENOMEM;
163 }
164 }
165 }
Nagendra Tomarb4ff3c92010-11-26 14:26:27 +0000166 inet_bind_hash(child, tb, port);
Pavel Emelyanov53083772008-04-17 23:18:15 -0700167 spin_unlock(&head->lock);
Balazs Scheidler093d2822010-10-21 13:06:43 +0200168
169 return 0;
Pavel Emelyanov53083772008-04-17 23:18:15 -0700170}
Pavel Emelyanov53083772008-04-17 23:18:15 -0700171EXPORT_SYMBOL_GPL(__inet_inherit_port);
172
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800173static inline int compute_score(struct sock *sk, struct net *net,
174 const unsigned short hnum, const __be32 daddr,
175 const int dif)
176{
177 int score = -1;
178 struct inet_sock *inet = inet_sk(sk);
179
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000180 if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800181 !ipv6_only_sock(sk)) {
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000182 __be32 rcv_saddr = inet->inet_rcv_saddr;
Tom Herbertda5e3632013-01-22 09:50:24 +0000183 score = sk->sk_family == PF_INET ? 2 : 1;
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800184 if (rcv_saddr) {
185 if (rcv_saddr != daddr)
186 return -1;
Tom Herbertda5e3632013-01-22 09:50:24 +0000187 score += 4;
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800188 }
189 if (sk->sk_bound_dev_if) {
190 if (sk->sk_bound_dev_if != dif)
191 return -1;
Tom Herbertda5e3632013-01-22 09:50:24 +0000192 score += 4;
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800193 }
Eric Dumazet70da2682015-10-08 19:33:21 -0700194 if (sk->sk_incoming_cpu == raw_smp_processor_id())
195 score++;
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800196 }
197 return score;
198}
199
Arnaldo Carvalho de Melof3f05f72005-08-09 20:08:09 -0700200/*
Arnaldo Carvalho de Melo33b62232005-08-09 20:09:06 -0700201 * Don't inline this cruft. Here are some nice properties to exploit here. The
202 * BSD API does not allow a listening sock to specify the remote port nor the
203 * remote address for the connection. So always assume those are both
204 * wildcarded during the search since they can never be otherwise.
205 */
Arnaldo Carvalho de Melo33b62232005-08-09 20:09:06 -0700206
Arnaldo Carvalho de Melo33b62232005-08-09 20:09:06 -0700207
Pavel Emelyanovc67499c2008-01-31 05:06:40 -0800208struct sock *__inet_lookup_listener(struct net *net,
209 struct inet_hashinfo *hashinfo,
Craig Galleka5836362016-02-10 11:50:38 -0500210 struct sk_buff *skb, int doff,
Tom Herbertda5e3632013-01-22 09:50:24 +0000211 const __be32 saddr, __be16 sport,
Al Virofb99c842006-09-27 18:43:33 -0700212 const __be32 daddr, const unsigned short hnum,
Herbert Xu8f4910692006-08-09 15:47:12 -0700213 const int dif)
Herbert Xu99a92ff2006-08-08 02:18:10 -0700214{
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800215 struct sock *sk, *result;
216 struct hlist_nulls_node *node;
217 unsigned int hash = inet_lhashfn(net, hnum);
218 struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
Tom Herbertda5e3632013-01-22 09:50:24 +0000219 int score, hiscore, matches = 0, reuseport = 0;
Craig Gallekc125e802016-02-10 11:50:40 -0500220 bool select_ok = true;
Tom Herbertda5e3632013-01-22 09:50:24 +0000221 u32 phash = 0;
Herbert Xu99a92ff2006-08-08 02:18:10 -0700222
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800223 rcu_read_lock();
224begin:
225 result = NULL;
Tom Herbertda5e3632013-01-22 09:50:24 +0000226 hiscore = 0;
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800227 sk_nulls_for_each_rcu(sk, node, &ilb->head) {
228 score = compute_score(sk, net, hnum, daddr, dif);
229 if (score > hiscore) {
230 result = sk;
231 hiscore = score;
Tom Herbertda5e3632013-01-22 09:50:24 +0000232 reuseport = sk->sk_reuseport;
233 if (reuseport) {
234 phash = inet_ehashfn(net, daddr, hnum,
235 saddr, sport);
Craig Gallekc125e802016-02-10 11:50:40 -0500236 if (select_ok) {
237 struct sock *sk2;
238 sk2 = reuseport_select_sock(sk, phash,
239 skb, doff);
240 if (sk2) {
241 result = sk2;
242 goto found;
243 }
244 }
Tom Herbertda5e3632013-01-22 09:50:24 +0000245 matches = 1;
246 }
247 } else if (score == hiscore && reuseport) {
248 matches++;
Daniel Borkmann8fc54f62014-08-23 20:58:54 +0200249 if (reciprocal_scale(phash, matches) == 0)
Tom Herbertda5e3632013-01-22 09:50:24 +0000250 result = sk;
251 phash = next_pseudo_random32(phash);
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800252 }
Herbert Xu99a92ff2006-08-08 02:18:10 -0700253 }
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800254 /*
255 * if the nulls value we got at the end of this lookup is
256 * not the expected one, we must restart lookup.
257 * We probably met an item that was moved to another chain.
258 */
259 if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
260 goto begin;
261 if (result) {
Craig Gallekc125e802016-02-10 11:50:40 -0500262found:
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800263 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
264 result = NULL;
265 else if (unlikely(compute_score(result, net, hnum, daddr,
266 dif) < hiscore)) {
267 sock_put(result);
Craig Gallekc125e802016-02-10 11:50:40 -0500268 select_ok = false;
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800269 goto begin;
270 }
Herbert Xu99a92ff2006-08-08 02:18:10 -0700271 }
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800272 rcu_read_unlock();
273 return result;
Herbert Xu99a92ff2006-08-08 02:18:10 -0700274}
Herbert Xu8f4910692006-08-09 15:47:12 -0700275EXPORT_SYMBOL_GPL(__inet_lookup_listener);
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800276
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700277/* All sockets share common refcount, but have different destructors */
278void sock_gen_put(struct sock *sk)
279{
280 if (!atomic_dec_and_test(&sk->sk_refcnt))
281 return;
282
283 if (sk->sk_state == TCP_TIME_WAIT)
284 inet_twsk_free(inet_twsk(sk));
Eric Dumazet41b822c2015-03-12 16:44:08 -0700285 else if (sk->sk_state == TCP_NEW_SYN_RECV)
286 reqsk_free(inet_reqsk(sk));
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700287 else
288 sk_free(sk);
289}
290EXPORT_SYMBOL_GPL(sock_gen_put);
291
Eric Dumazet2c132702015-03-15 21:12:15 -0700292void sock_edemux(struct sk_buff *skb)
293{
294 sock_gen_put(skb->sk);
295}
296EXPORT_SYMBOL(sock_edemux);
297
Daniel Baluta5e73ea12012-04-15 01:34:41 +0000298struct sock *__inet_lookup_established(struct net *net,
Pavel Emelyanovc67499c2008-01-31 05:06:40 -0800299 struct inet_hashinfo *hashinfo,
Pavel Emelyanov77a5ba52007-12-20 15:32:17 -0800300 const __be32 saddr, const __be16 sport,
301 const __be32 daddr, const u16 hnum,
302 const int dif)
303{
Joe Perchesc7228312014-05-13 20:30:07 -0700304 INET_ADDR_COOKIE(acookie, saddr, daddr);
Pavel Emelyanov77a5ba52007-12-20 15:32:17 -0800305 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
306 struct sock *sk;
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800307 const struct hlist_nulls_node *node;
Pavel Emelyanov77a5ba52007-12-20 15:32:17 -0800308 /* Optimize here for direct hit, only listening connections can
309 * have wildcards anyways.
310 */
Pavel Emelyanov9f26b3a2008-06-16 17:13:27 -0700311 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
Eric Dumazetf373b532009-10-09 00:16:19 +0000312 unsigned int slot = hash & hashinfo->ehash_mask;
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800313 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
Pavel Emelyanov77a5ba52007-12-20 15:32:17 -0800314
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800315 rcu_read_lock();
316begin:
317 sk_nulls_for_each_rcu(sk, node, &head->chain) {
Eric Dumazetce43b032012-11-30 09:49:27 +0000318 if (sk->sk_hash != hash)
319 continue;
320 if (likely(INET_MATCH(sk, net, acookie,
321 saddr, daddr, ports, dif))) {
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800322 if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700323 goto out;
Eric Dumazetce43b032012-11-30 09:49:27 +0000324 if (unlikely(!INET_MATCH(sk, net, acookie,
325 saddr, daddr, ports, dif))) {
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700326 sock_gen_put(sk);
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800327 goto begin;
328 }
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700329 goto found;
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800330 }
Pavel Emelyanov77a5ba52007-12-20 15:32:17 -0800331 }
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800332 /*
333 * if the nulls value we got at the end of this lookup is
334 * not the expected one, we must restart lookup.
335 * We probably met an item that was moved to another chain.
336 */
337 if (get_nulls_value(node) != slot)
338 goto begin;
Pavel Emelyanov77a5ba52007-12-20 15:32:17 -0800339out:
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700340 sk = NULL;
341found:
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800342 rcu_read_unlock();
Pavel Emelyanov77a5ba52007-12-20 15:32:17 -0800343 return sk;
Pavel Emelyanov77a5ba52007-12-20 15:32:17 -0800344}
345EXPORT_SYMBOL_GPL(__inet_lookup_established);
346
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800347/* called with local bh disabled */
348static int __inet_check_established(struct inet_timewait_death_row *death_row,
349 struct sock *sk, __u16 lport,
350 struct inet_timewait_sock **twp)
351{
352 struct inet_hashinfo *hinfo = death_row->hashinfo;
353 struct inet_sock *inet = inet_sk(sk);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000354 __be32 daddr = inet->inet_rcv_saddr;
355 __be32 saddr = inet->inet_daddr;
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800356 int dif = sk->sk_bound_dev_if;
Joe Perchesc7228312014-05-13 20:30:07 -0700357 INET_ADDR_COOKIE(acookie, saddr, daddr);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000358 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
Pavel Emelyanov9f26b3a2008-06-16 17:13:27 -0700359 struct net *net = sock_net(sk);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000360 unsigned int hash = inet_ehashfn(net, daddr, lport,
361 saddr, inet->inet_dport);
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800362 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
Eric Dumazet9db66bd2008-11-20 20:39:09 -0800363 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800364 struct sock *sk2;
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800365 const struct hlist_nulls_node *node;
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700366 struct inet_timewait_sock *tw = NULL;
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800367
Eric Dumazet9db66bd2008-11-20 20:39:09 -0800368 spin_lock(lock);
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800369
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800370 sk_nulls_for_each(sk2, node, &head->chain) {
Eric Dumazetce43b032012-11-30 09:49:27 +0000371 if (sk2->sk_hash != hash)
372 continue;
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700373
Eric Dumazetce43b032012-11-30 09:49:27 +0000374 if (likely(INET_MATCH(sk2, net, acookie,
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700375 saddr, daddr, ports, dif))) {
376 if (sk2->sk_state == TCP_TIME_WAIT) {
377 tw = inet_twsk(sk2);
378 if (twsk_unique(sk, sk2, twp))
379 break;
380 }
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800381 goto not_unique;
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700382 }
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800383 }
384
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800385 /* Must record num and sport now. Otherwise we will see
Eric Dumazet05dbc7b2013-10-03 00:22:02 -0700386 * in hash table socket with a funny identity.
387 */
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000388 inet->inet_num = lport;
389 inet->inet_sport = htons(lport);
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800390 sk->sk_hash = hash;
Ilpo Järvinen547b7922008-07-25 21:43:18 -0700391 WARN_ON(!sk_unhashed(sk));
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800392 __sk_nulls_add_node_rcu(sk, &head->chain);
Eric Dumazet13475a32009-12-02 22:31:19 +0000393 if (tw) {
Eric Dumazetfc01538f2015-07-08 14:28:29 -0700394 sk_nulls_del_node_init_rcu((struct sock *)tw);
Eric Dumazet13475a32009-12-02 22:31:19 +0000395 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
396 }
Eric Dumazet9db66bd2008-11-20 20:39:09 -0800397 spin_unlock(lock);
Pavel Emelyanovc29a0bc2008-03-31 19:41:46 -0700398 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800399
400 if (twp) {
401 *twp = tw;
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800402 } else if (tw) {
403 /* Silly. Should hash-dance instead... */
Eric Dumazetdbe7faa2015-07-08 14:28:30 -0700404 inet_twsk_deschedule_put(tw);
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800405 }
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800406 return 0;
407
408not_unique:
Eric Dumazet9db66bd2008-11-20 20:39:09 -0800409 spin_unlock(lock);
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800410 return -EADDRNOTAVAIL;
411}
412
Eric Dumazete2baad92015-05-27 10:46:02 -0700413static u32 inet_sk_port_offset(const struct sock *sk)
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800414{
415 const struct inet_sock *inet = inet_sk(sk);
Eric Dumazete2baad92015-05-27 10:46:02 -0700416
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000417 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
418 inet->inet_daddr,
419 inet->inet_dport);
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800420}
421
Eric Dumazet079096f2015-10-02 11:43:32 -0700422/* insert a socket into ehash, and eventually remove another one
423 * (The another one can be a SYN_RECV or TIMEWAIT
424 */
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700425bool inet_ehash_insert(struct sock *sk, struct sock *osk)
Pavel Emelyanov152da812007-12-20 15:31:33 -0800426{
Pavel Emelyanov39d8cda2008-03-22 16:50:58 -0700427 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
Eric Dumazet3ab5aee2008-11-16 19:40:17 -0800428 struct hlist_nulls_head *list;
Pavel Emelyanov152da812007-12-20 15:31:33 -0800429 struct inet_ehash_bucket *head;
Eric Dumazet5b441f72015-03-18 14:05:34 -0700430 spinlock_t *lock;
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700431 bool ret = true;
Pavel Emelyanov152da812007-12-20 15:31:33 -0800432
Eric Dumazet079096f2015-10-02 11:43:32 -0700433 WARN_ON_ONCE(!sk_unhashed(sk));
Pavel Emelyanov152da812007-12-20 15:31:33 -0800434
Eric Dumazet5b441f72015-03-18 14:05:34 -0700435 sk->sk_hash = sk_ehashfn(sk);
Pavel Emelyanov152da812007-12-20 15:31:33 -0800436 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
437 list = &head->chain;
438 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
439
Eric Dumazet9db66bd2008-11-20 20:39:09 -0800440 spin_lock(lock);
Eric Dumazetfc01538f2015-07-08 14:28:29 -0700441 if (osk) {
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700442 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
443 ret = sk_nulls_del_node_init_rcu(osk);
Eric Dumazet9327f702009-12-04 03:46:54 +0000444 }
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700445 if (ret)
446 __sk_nulls_add_node_rcu(sk, list);
Eric Dumazet9db66bd2008-11-20 20:39:09 -0800447 spin_unlock(lock);
Eric Dumazet079096f2015-10-02 11:43:32 -0700448 return ret;
449}
450
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700451bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
Eric Dumazet079096f2015-10-02 11:43:32 -0700452{
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700453 bool ok = inet_ehash_insert(sk, osk);
454
455 if (ok) {
456 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
457 } else {
458 percpu_counter_inc(sk->sk_prot->orphan_count);
459 sk->sk_state = TCP_CLOSE;
460 sock_set_flag(sk, SOCK_DEAD);
461 inet_csk_destroy_sock(sk);
462 }
463 return ok;
Pavel Emelyanov152da812007-12-20 15:31:33 -0800464}
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700465EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
Pavel Emelyanov152da812007-12-20 15:31:33 -0800466
Craig Gallekc125e802016-02-10 11:50:40 -0500467static int inet_reuseport_add_sock(struct sock *sk,
468 struct inet_listen_hashbucket *ilb,
469 int (*saddr_same)(const struct sock *sk1,
470 const struct sock *sk2,
471 bool match_wildcard))
472{
473 struct sock *sk2;
474 struct hlist_nulls_node *node;
475 kuid_t uid = sock_i_uid(sk);
476
477 sk_nulls_for_each_rcu(sk2, node, &ilb->head) {
478 if (sk2 != sk &&
479 sk2->sk_family == sk->sk_family &&
480 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
481 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
482 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
483 saddr_same(sk, sk2, false))
484 return reuseport_add_sock(sk, sk2);
485 }
486
487 /* Initial allocation may have already happened via setsockopt */
488 if (!rcu_access_pointer(sk->sk_reuseport_cb))
489 return reuseport_alloc(sk);
490 return 0;
491}
492
493int __inet_hash(struct sock *sk, struct sock *osk,
494 int (*saddr_same)(const struct sock *sk1,
495 const struct sock *sk2,
496 bool match_wildcard))
Pavel Emelyanov152da812007-12-20 15:31:33 -0800497{
Pavel Emelyanov39d8cda2008-03-22 16:50:58 -0700498 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
Eric Dumazet5caea4e2008-11-20 00:40:07 -0800499 struct inet_listen_hashbucket *ilb;
Craig Gallekc125e802016-02-10 11:50:40 -0500500 int err = 0;
Pavel Emelyanov152da812007-12-20 15:31:33 -0800501
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700502 if (sk->sk_state != TCP_LISTEN) {
503 inet_ehash_nolisten(sk, osk);
Craig Gallekc125e802016-02-10 11:50:40 -0500504 return 0;
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700505 }
Ilpo Järvinen547b7922008-07-25 21:43:18 -0700506 WARN_ON(!sk_unhashed(sk));
Eric Dumazet5caea4e2008-11-20 00:40:07 -0800507 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
Pavel Emelyanov152da812007-12-20 15:31:33 -0800508
Eric Dumazet5caea4e2008-11-20 00:40:07 -0800509 spin_lock(&ilb->lock);
Craig Gallekc125e802016-02-10 11:50:40 -0500510 if (sk->sk_reuseport) {
511 err = inet_reuseport_add_sock(sk, ilb, saddr_same);
512 if (err)
513 goto unlock;
514 }
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800515 __sk_nulls_add_node_rcu(sk, &ilb->head);
Pavel Emelyanovc29a0bc2008-03-31 19:41:46 -0700516 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
Craig Gallekc125e802016-02-10 11:50:40 -0500517unlock:
Eric Dumazet5caea4e2008-11-20 00:40:07 -0800518 spin_unlock(&ilb->lock);
Craig Gallekc125e802016-02-10 11:50:40 -0500519
520 return err;
Pavel Emelyanov152da812007-12-20 15:31:33 -0800521}
Eric Dumazet77a6a472015-03-18 14:05:36 -0700522EXPORT_SYMBOL(__inet_hash);
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800523
Craig Gallek086c6532016-02-10 11:50:35 -0500524int inet_hash(struct sock *sk)
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800525{
Craig Gallekc125e802016-02-10 11:50:40 -0500526 int err = 0;
527
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800528 if (sk->sk_state != TCP_CLOSE) {
529 local_bh_disable();
Craig Gallekc125e802016-02-10 11:50:40 -0500530 err = __inet_hash(sk, NULL, ipv4_rcv_saddr_equal);
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800531 local_bh_enable();
532 }
Craig Gallek086c6532016-02-10 11:50:35 -0500533
Craig Gallekc125e802016-02-10 11:50:40 -0500534 return err;
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800535}
536EXPORT_SYMBOL_GPL(inet_hash);
537
538void inet_unhash(struct sock *sk)
539{
Pavel Emelyanov39d8cda2008-03-22 16:50:58 -0700540 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800541 spinlock_t *lock;
542 int done;
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800543
544 if (sk_unhashed(sk))
Eric Dumazet5caea4e2008-11-20 00:40:07 -0800545 return;
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800546
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800547 if (sk->sk_state == TCP_LISTEN)
548 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
549 else
550 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
Eric Dumazet5caea4e2008-11-20 00:40:07 -0800551
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800552 spin_lock_bh(lock);
Craig Gallekc125e802016-02-10 11:50:40 -0500553 if (rcu_access_pointer(sk->sk_reuseport_cb))
554 reuseport_detach_sock(sk);
Camelia Groza3b8ccd42013-07-11 09:55:51 +0300555 done = __sk_nulls_del_node_init_rcu(sk);
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800556 if (done)
557 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
Eric Dumazet920de802008-11-24 00:09:29 -0800558 spin_unlock_bh(lock);
Arnaldo Carvalho de Meloab1e0a12008-02-03 04:06:04 -0800559}
560EXPORT_SYMBOL_GPL(inet_unhash);
Pavel Emelyanov152da812007-12-20 15:31:33 -0800561
Pavel Emelyanov5ee31fc2008-01-31 05:04:45 -0800562int __inet_hash_connect(struct inet_timewait_death_row *death_row,
Pavel Emelyanov5d8c0aa2008-02-05 03:14:44 -0800563 struct sock *sk, u32 port_offset,
Pavel Emelyanov5ee31fc2008-01-31 05:04:45 -0800564 int (*check_established)(struct inet_timewait_death_row *,
Eric Dumazetb4d64442015-03-18 14:05:37 -0700565 struct sock *, __u16, struct inet_timewait_sock **))
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800566{
567 struct inet_hashinfo *hinfo = death_row->hashinfo;
Eric Dumazet1580ab62016-02-11 16:28:49 -0800568 struct inet_timewait_sock *tw = NULL;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900569 struct inet_bind_hashbucket *head;
Eric Dumazet1580ab62016-02-11 16:28:49 -0800570 int port = inet_sk(sk)->inet_num;
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900571 struct net *net = sock_net(sk);
Eric Dumazet1580ab62016-02-11 16:28:49 -0800572 struct inet_bind_bucket *tb;
573 u32 remaining, offset;
574 int ret, i, low, high;
575 static u32 hint;
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800576
Eric Dumazet1580ab62016-02-11 16:28:49 -0800577 if (port) {
578 head = &hinfo->bhash[inet_bhashfn(net, port,
579 hinfo->bhash_size)];
580 tb = inet_csk(sk)->icsk_bind_hash;
581 spin_lock_bh(&head->lock);
582 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
583 inet_ehash_nolisten(sk, NULL);
584 spin_unlock_bh(&head->lock);
585 return 0;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900586 }
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800587 spin_unlock(&head->lock);
588 /* No definite answer... Walk to established hash table */
Eric Dumazet1580ab62016-02-11 16:28:49 -0800589 ret = check_established(death_row, sk, port, NULL);
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800590 local_bh_enable();
591 return ret;
592 }
Eric Dumazet1580ab62016-02-11 16:28:49 -0800593
594 inet_get_local_port_range(net, &low, &high);
595 high++; /* [32768, 60999] -> [32768, 61000[ */
596 remaining = high - low;
597 if (likely(remaining > 1))
598 remaining &= ~1U;
599
600 offset = (hint + port_offset) % remaining;
601 /* In first pass we try ports of @low parity.
602 * inet_csk_get_port() does the opposite choice.
603 */
604 offset &= ~1U;
605other_parity_scan:
606 port = low + offset;
607 for (i = 0; i < remaining; i += 2, port += 2) {
608 if (unlikely(port >= high))
609 port -= remaining;
610 if (inet_is_local_reserved_port(net, port))
611 continue;
612 head = &hinfo->bhash[inet_bhashfn(net, port,
613 hinfo->bhash_size)];
614 spin_lock_bh(&head->lock);
615
616 /* Does not bother with rcv_saddr checks, because
617 * the established check is already unique enough.
618 */
619 inet_bind_bucket_for_each(tb, &head->chain) {
620 if (net_eq(ib_net(tb), net) && tb->port == port) {
621 if (tb->fastreuse >= 0 ||
622 tb->fastreuseport >= 0)
623 goto next_port;
624 WARN_ON(hlist_empty(&tb->owners));
625 if (!check_established(death_row, sk,
626 port, &tw))
627 goto ok;
628 goto next_port;
629 }
630 }
631
632 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
633 net, head, port);
634 if (!tb) {
635 spin_unlock_bh(&head->lock);
636 return -ENOMEM;
637 }
638 tb->fastreuse = -1;
639 tb->fastreuseport = -1;
640 goto ok;
641next_port:
642 spin_unlock_bh(&head->lock);
643 cond_resched();
644 }
645
646 offset++;
647 if ((offset & 1) && remaining > 1)
648 goto other_parity_scan;
649
650 return -EADDRNOTAVAIL;
651
652ok:
653 hint += i + 2;
654
655 /* Head lock still held and bh's disabled */
656 inet_bind_hash(sk, tb, port);
657 if (sk_unhashed(sk)) {
658 inet_sk(sk)->inet_sport = htons(port);
659 inet_ehash_nolisten(sk, (struct sock *)tw);
660 }
661 if (tw)
662 inet_twsk_bind_unhash(tw, hinfo);
663 spin_unlock(&head->lock);
664 if (tw)
665 inet_twsk_deschedule_put(tw);
666 local_bh_enable();
667 return 0;
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800668}
Pavel Emelyanov5ee31fc2008-01-31 05:04:45 -0800669
670/*
671 * Bind a port for a connect operation and hash it.
672 */
673int inet_hash_connect(struct inet_timewait_death_row *death_row,
674 struct sock *sk)
675{
Eric Dumazete2baad92015-05-27 10:46:02 -0700676 u32 port_offset = 0;
677
678 if (!inet_sk(sk)->inet_num)
679 port_offset = inet_sk_port_offset(sk);
680 return __inet_hash_connect(death_row, sk, port_offset,
Eric Dumazetb4d64442015-03-18 14:05:37 -0700681 __inet_check_established);
Pavel Emelyanov5ee31fc2008-01-31 05:04:45 -0800682}
Arnaldo Carvalho de Meloa7f5e7f2005-12-13 23:25:31 -0800683EXPORT_SYMBOL_GPL(inet_hash_connect);
Eric Dumazet5caea4e2008-11-20 00:40:07 -0800684
685void inet_hashinfo_init(struct inet_hashinfo *h)
686{
687 int i;
688
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800689 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
Eric Dumazet5caea4e2008-11-20 00:40:07 -0800690 spin_lock_init(&h->listening_hash[i].lock);
Eric Dumazetc25eb3b2008-11-23 17:22:55 -0800691 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
692 i + LISTENING_NULLS_BASE);
693 }
Eric Dumazet5caea4e2008-11-20 00:40:07 -0800694}
Eric Dumazet5caea4e2008-11-20 00:40:07 -0800695EXPORT_SYMBOL_GPL(inet_hashinfo_init);
Eric Dumazet095dc8e2015-05-26 07:55:34 -0700696
697int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
698{
Eric Dumazet89e478a2015-07-22 07:02:00 +0200699 unsigned int locksz = sizeof(spinlock_t);
Eric Dumazet095dc8e2015-05-26 07:55:34 -0700700 unsigned int i, nblocks = 1;
701
Eric Dumazet89e478a2015-07-22 07:02:00 +0200702 if (locksz != 0) {
Eric Dumazet095dc8e2015-05-26 07:55:34 -0700703 /* allocate 2 cache lines or at least one spinlock per cpu */
Eric Dumazet89e478a2015-07-22 07:02:00 +0200704 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
Eric Dumazet095dc8e2015-05-26 07:55:34 -0700705 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
706
707 /* no more locks than number of hash buckets */
708 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
709
Eric Dumazet89e478a2015-07-22 07:02:00 +0200710 hashinfo->ehash_locks = kmalloc_array(nblocks, locksz,
Eric Dumazet095dc8e2015-05-26 07:55:34 -0700711 GFP_KERNEL | __GFP_NOWARN);
712 if (!hashinfo->ehash_locks)
Eric Dumazet89e478a2015-07-22 07:02:00 +0200713 hashinfo->ehash_locks = vmalloc(nblocks * locksz);
Eric Dumazet095dc8e2015-05-26 07:55:34 -0700714
715 if (!hashinfo->ehash_locks)
716 return -ENOMEM;
717
718 for (i = 0; i < nblocks; i++)
719 spin_lock_init(&hashinfo->ehash_locks[i]);
720 }
721 hashinfo->ehash_locks_mask = nblocks - 1;
722 return 0;
723}
724EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);