]>
Commit | Line | Data |
---|---|---|
77d8bf9c ACM |
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 | ||
2d8c4ce5 | 16 | #include <linux/module.h> |
a7f5e7f1 | 17 | #include <linux/random.h> |
f3f05f70 | 18 | #include <linux/sched.h> |
77d8bf9c | 19 | #include <linux/slab.h> |
f3f05f70 | 20 | #include <linux/wait.h> |
095dc8e0 | 21 | #include <linux/vmalloc.h> |
77d8bf9c | 22 | |
c125e80b | 23 | #include <net/addrconf.h> |
463c84b9 | 24 | #include <net/inet_connection_sock.h> |
77d8bf9c | 25 | #include <net/inet_hashtables.h> |
6e5714ea | 26 | #include <net/secure_seq.h> |
a7f5e7f1 | 27 | #include <net/ip.h> |
c125e80b | 28 | #include <net/sock_reuseport.h> |
77d8bf9c | 29 | |
6eada011 ED |
30 | static u32 inet_ehashfn(const struct net *net, const __be32 laddr, |
31 | const __u16 lport, const __be32 faddr, | |
32 | const __be16 fport) | |
65cd8033 | 33 | { |
1bbdceef HFS |
34 | static u32 inet_ehash_secret __read_mostly; |
35 | ||
36 | net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret)); | |
37 | ||
65cd8033 HFS |
38 | return __inet_ehashfn(laddr, lport, faddr, fport, |
39 | inet_ehash_secret + net_hash_mix(net)); | |
40 | } | |
41 | ||
d1e559d0 ED |
42 | /* This function handles inet_sock, but also timewait and request sockets |
43 | * for IPv4/IPv6. | |
44 | */ | |
5b441f76 | 45 | u32 sk_ehashfn(const struct sock *sk) |
65cd8033 | 46 | { |
d1e559d0 ED |
47 | #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 | |
5b441f76 ED |
54 | return inet_ehashfn(sock_net(sk), |
55 | sk->sk_rcv_saddr, sk->sk_num, | |
56 | sk->sk_daddr, sk->sk_dport); | |
65cd8033 HFS |
57 | } |
58 | ||
77d8bf9c ACM |
59 | /* |
60 | * Allocate and initialize a new local port bind bucket. | |
61 | * The bindhash mutex for snum's hash chain must be held here. | |
62 | */ | |
e18b890b | 63 | struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep, |
941b1d22 | 64 | struct net *net, |
77d8bf9c ACM |
65 | struct inet_bind_hashbucket *head, |
66 | const unsigned short snum) | |
67 | { | |
54e6ecb2 | 68 | struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC); |
77d8bf9c | 69 | |
00db4124 | 70 | if (tb) { |
efd7ef1c | 71 | write_pnet(&tb->ib_net, net); |
77d8bf9c ACM |
72 | tb->port = snum; |
73 | tb->fastreuse = 0; | |
da5e3630 | 74 | tb->fastreuseport = 0; |
a9d8f911 | 75 | tb->num_owners = 0; |
77d8bf9c ACM |
76 | INIT_HLIST_HEAD(&tb->owners); |
77 | hlist_add_head(&tb->node, &head->chain); | |
78 | } | |
79 | return tb; | |
80 | } | |
81 | ||
77d8bf9c ACM |
82 | /* |
83 | * Caller must hold hashbucket lock for this tb with local BH disabled | |
84 | */ | |
e18b890b | 85 | void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb) |
77d8bf9c ACM |
86 | { |
87 | if (hlist_empty(&tb->owners)) { | |
88 | __hlist_del(&tb->node); | |
89 | kmem_cache_free(cachep, tb); | |
90 | } | |
91 | } | |
2d8c4ce5 ACM |
92 | |
93 | void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb, | |
94 | const unsigned short snum) | |
95 | { | |
c720c7e8 | 96 | inet_sk(sk)->inet_num = snum; |
2d8c4ce5 | 97 | sk_add_bind_node(sk, &tb->owners); |
a9d8f911 | 98 | tb->num_owners++; |
463c84b9 | 99 | inet_csk(sk)->icsk_bind_hash = tb; |
2d8c4ce5 ACM |
100 | } |
101 | ||
2d8c4ce5 ACM |
102 | /* |
103 | * Get rid of any references to a local port held by the given sock. | |
104 | */ | |
ab1e0a13 | 105 | static void __inet_put_port(struct sock *sk) |
2d8c4ce5 | 106 | { |
39d8cda7 | 107 | struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; |
c720c7e8 | 108 | const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num, |
7f635ab7 | 109 | hashinfo->bhash_size); |
2d8c4ce5 ACM |
110 | struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash]; |
111 | struct inet_bind_bucket *tb; | |
112 | ||
113 | spin_lock(&head->lock); | |
463c84b9 | 114 | tb = inet_csk(sk)->icsk_bind_hash; |
2d8c4ce5 | 115 | __sk_del_bind_node(sk); |
a9d8f911 | 116 | tb->num_owners--; |
463c84b9 | 117 | inet_csk(sk)->icsk_bind_hash = NULL; |
c720c7e8 | 118 | inet_sk(sk)->inet_num = 0; |
2d8c4ce5 ACM |
119 | inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb); |
120 | spin_unlock(&head->lock); | |
121 | } | |
122 | ||
ab1e0a13 | 123 | void inet_put_port(struct sock *sk) |
2d8c4ce5 ACM |
124 | { |
125 | local_bh_disable(); | |
ab1e0a13 | 126 | __inet_put_port(sk); |
2d8c4ce5 ACM |
127 | local_bh_enable(); |
128 | } | |
2d8c4ce5 | 129 | EXPORT_SYMBOL(inet_put_port); |
f3f05f70 | 130 | |
1ce31c9e | 131 | int __inet_inherit_port(const struct sock *sk, struct sock *child) |
53083773 PE |
132 | { |
133 | struct inet_hashinfo *table = sk->sk_prot->h.hashinfo; | |
093d2823 BS |
134 | unsigned short port = inet_sk(child)->inet_num; |
135 | const int bhash = inet_bhashfn(sock_net(sk), port, | |
7f635ab7 | 136 | table->bhash_size); |
53083773 PE |
137 | 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; | |
c2f34a65 ED |
142 | if (unlikely(!tb)) { |
143 | spin_unlock(&head->lock); | |
144 | return -ENOENT; | |
145 | } | |
093d2823 BS |
146 | 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. */ | |
b67bfe0d | 152 | inet_bind_bucket_for_each(tb, &head->chain) { |
093d2823 BS |
153 | if (net_eq(ib_net(tb), sock_net(sk)) && |
154 | tb->port == port) | |
155 | break; | |
156 | } | |
b67bfe0d | 157 | if (!tb) { |
093d2823 BS |
158 | 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 | } | |
b4ff3c90 | 166 | inet_bind_hash(child, tb, port); |
53083773 | 167 | spin_unlock(&head->lock); |
093d2823 BS |
168 | |
169 | return 0; | |
53083773 | 170 | } |
53083773 PE |
171 | EXPORT_SYMBOL_GPL(__inet_inherit_port); |
172 | ||
c25eb3bf ED |
173 | static 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 | ||
c720c7e8 | 180 | if (net_eq(sock_net(sk), net) && inet->inet_num == hnum && |
c25eb3bf | 181 | !ipv6_only_sock(sk)) { |
c720c7e8 | 182 | __be32 rcv_saddr = inet->inet_rcv_saddr; |
da5e3630 | 183 | score = sk->sk_family == PF_INET ? 2 : 1; |
c25eb3bf ED |
184 | if (rcv_saddr) { |
185 | if (rcv_saddr != daddr) | |
186 | return -1; | |
da5e3630 | 187 | score += 4; |
c25eb3bf ED |
188 | } |
189 | if (sk->sk_bound_dev_if) { | |
190 | if (sk->sk_bound_dev_if != dif) | |
191 | return -1; | |
da5e3630 | 192 | score += 4; |
c25eb3bf | 193 | } |
70da268b ED |
194 | if (sk->sk_incoming_cpu == raw_smp_processor_id()) |
195 | score++; | |
c25eb3bf ED |
196 | } |
197 | return score; | |
198 | } | |
199 | ||
33b62231 | 200 | /* |
3b24d854 ED |
201 | * Here are some nice properties to exploit here. The BSD API |
202 | * does not allow a listening sock to specify the remote port nor the | |
33b62231 ACM |
203 | * remote address for the connection. So always assume those are both |
204 | * wildcarded during the search since they can never be otherwise. | |
205 | */ | |
e48c414e | 206 | |
3b24d854 | 207 | /* called with rcu_read_lock() : No refcount taken on the socket */ |
c67499c0 PE |
208 | struct sock *__inet_lookup_listener(struct net *net, |
209 | struct inet_hashinfo *hashinfo, | |
a583636a | 210 | struct sk_buff *skb, int doff, |
da5e3630 | 211 | const __be32 saddr, __be16 sport, |
fb99c848 | 212 | const __be32 daddr, const unsigned short hnum, |
8f491069 | 213 | const int dif) |
99a92ff5 | 214 | { |
c25eb3bf ED |
215 | unsigned int hash = inet_lhashfn(net, hnum); |
216 | struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash]; | |
3b24d854 ED |
217 | int score, hiscore = 0, matches = 0, reuseport = 0; |
218 | struct sock *sk, *result = NULL; | |
da5e3630 | 219 | u32 phash = 0; |
99a92ff5 | 220 | |
3b24d854 | 221 | sk_for_each_rcu(sk, &ilb->head) { |
c25eb3bf ED |
222 | score = compute_score(sk, net, hnum, daddr, dif); |
223 | if (score > hiscore) { | |
da5e3630 TH |
224 | reuseport = sk->sk_reuseport; |
225 | if (reuseport) { | |
226 | phash = inet_ehashfn(net, daddr, hnum, | |
227 | saddr, sport); | |
3b24d854 ED |
228 | result = reuseport_select_sock(sk, phash, |
229 | skb, doff); | |
230 | if (result) | |
231 | return result; | |
da5e3630 TH |
232 | matches = 1; |
233 | } | |
3b24d854 ED |
234 | result = sk; |
235 | hiscore = score; | |
da5e3630 TH |
236 | } else if (score == hiscore && reuseport) { |
237 | matches++; | |
8fc54f68 | 238 | if (reciprocal_scale(phash, matches) == 0) |
da5e3630 TH |
239 | result = sk; |
240 | phash = next_pseudo_random32(phash); | |
c25eb3bf | 241 | } |
99a92ff5 | 242 | } |
c25eb3bf | 243 | return result; |
99a92ff5 | 244 | } |
8f491069 | 245 | EXPORT_SYMBOL_GPL(__inet_lookup_listener); |
a7f5e7f1 | 246 | |
05dbc7b5 ED |
247 | /* All sockets share common refcount, but have different destructors */ |
248 | void sock_gen_put(struct sock *sk) | |
249 | { | |
250 | if (!atomic_dec_and_test(&sk->sk_refcnt)) | |
251 | return; | |
252 | ||
253 | if (sk->sk_state == TCP_TIME_WAIT) | |
254 | inet_twsk_free(inet_twsk(sk)); | |
41b822c5 ED |
255 | else if (sk->sk_state == TCP_NEW_SYN_RECV) |
256 | reqsk_free(inet_reqsk(sk)); | |
05dbc7b5 ED |
257 | else |
258 | sk_free(sk); | |
259 | } | |
260 | EXPORT_SYMBOL_GPL(sock_gen_put); | |
261 | ||
2c13270b ED |
262 | void sock_edemux(struct sk_buff *skb) |
263 | { | |
264 | sock_gen_put(skb->sk); | |
265 | } | |
266 | EXPORT_SYMBOL(sock_edemux); | |
267 | ||
5e73ea1a | 268 | struct sock *__inet_lookup_established(struct net *net, |
c67499c0 | 269 | struct inet_hashinfo *hashinfo, |
77a5ba55 PE |
270 | const __be32 saddr, const __be16 sport, |
271 | const __be32 daddr, const u16 hnum, | |
272 | const int dif) | |
273 | { | |
c7228317 | 274 | INET_ADDR_COOKIE(acookie, saddr, daddr); |
77a5ba55 PE |
275 | const __portpair ports = INET_COMBINED_PORTS(sport, hnum); |
276 | struct sock *sk; | |
3ab5aee7 | 277 | const struct hlist_nulls_node *node; |
77a5ba55 PE |
278 | /* Optimize here for direct hit, only listening connections can |
279 | * have wildcards anyways. | |
280 | */ | |
9f26b3ad | 281 | unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport); |
f373b53b | 282 | unsigned int slot = hash & hashinfo->ehash_mask; |
3ab5aee7 | 283 | struct inet_ehash_bucket *head = &hashinfo->ehash[slot]; |
77a5ba55 | 284 | |
3ab5aee7 ED |
285 | begin: |
286 | sk_nulls_for_each_rcu(sk, node, &head->chain) { | |
ce43b03e ED |
287 | if (sk->sk_hash != hash) |
288 | continue; | |
289 | if (likely(INET_MATCH(sk, net, acookie, | |
290 | saddr, daddr, ports, dif))) { | |
3ab5aee7 | 291 | if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt))) |
05dbc7b5 | 292 | goto out; |
ce43b03e ED |
293 | if (unlikely(!INET_MATCH(sk, net, acookie, |
294 | saddr, daddr, ports, dif))) { | |
05dbc7b5 | 295 | sock_gen_put(sk); |
3ab5aee7 ED |
296 | goto begin; |
297 | } | |
05dbc7b5 | 298 | goto found; |
3ab5aee7 | 299 | } |
77a5ba55 | 300 | } |
3ab5aee7 ED |
301 | /* |
302 | * if the nulls value we got at the end of this lookup is | |
303 | * not the expected one, we must restart lookup. | |
304 | * We probably met an item that was moved to another chain. | |
305 | */ | |
306 | if (get_nulls_value(node) != slot) | |
307 | goto begin; | |
77a5ba55 | 308 | out: |
05dbc7b5 ED |
309 | sk = NULL; |
310 | found: | |
77a5ba55 | 311 | return sk; |
77a5ba55 PE |
312 | } |
313 | EXPORT_SYMBOL_GPL(__inet_lookup_established); | |
314 | ||
a7f5e7f1 ACM |
315 | /* called with local bh disabled */ |
316 | static int __inet_check_established(struct inet_timewait_death_row *death_row, | |
317 | struct sock *sk, __u16 lport, | |
318 | struct inet_timewait_sock **twp) | |
319 | { | |
320 | struct inet_hashinfo *hinfo = death_row->hashinfo; | |
321 | struct inet_sock *inet = inet_sk(sk); | |
c720c7e8 ED |
322 | __be32 daddr = inet->inet_rcv_saddr; |
323 | __be32 saddr = inet->inet_daddr; | |
a7f5e7f1 | 324 | int dif = sk->sk_bound_dev_if; |
c7228317 | 325 | INET_ADDR_COOKIE(acookie, saddr, daddr); |
c720c7e8 | 326 | const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport); |
9f26b3ad | 327 | struct net *net = sock_net(sk); |
c720c7e8 ED |
328 | unsigned int hash = inet_ehashfn(net, daddr, lport, |
329 | saddr, inet->inet_dport); | |
a7f5e7f1 | 330 | struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash); |
9db66bdc | 331 | spinlock_t *lock = inet_ehash_lockp(hinfo, hash); |
a7f5e7f1 | 332 | struct sock *sk2; |
3ab5aee7 | 333 | const struct hlist_nulls_node *node; |
05dbc7b5 | 334 | struct inet_timewait_sock *tw = NULL; |
a7f5e7f1 | 335 | |
9db66bdc | 336 | spin_lock(lock); |
a7f5e7f1 | 337 | |
3ab5aee7 | 338 | sk_nulls_for_each(sk2, node, &head->chain) { |
ce43b03e ED |
339 | if (sk2->sk_hash != hash) |
340 | continue; | |
05dbc7b5 | 341 | |
ce43b03e | 342 | if (likely(INET_MATCH(sk2, net, acookie, |
05dbc7b5 ED |
343 | saddr, daddr, ports, dif))) { |
344 | if (sk2->sk_state == TCP_TIME_WAIT) { | |
345 | tw = inet_twsk(sk2); | |
346 | if (twsk_unique(sk, sk2, twp)) | |
347 | break; | |
348 | } | |
a7f5e7f1 | 349 | goto not_unique; |
05dbc7b5 | 350 | } |
a7f5e7f1 ACM |
351 | } |
352 | ||
a7f5e7f1 | 353 | /* Must record num and sport now. Otherwise we will see |
05dbc7b5 ED |
354 | * in hash table socket with a funny identity. |
355 | */ | |
c720c7e8 ED |
356 | inet->inet_num = lport; |
357 | inet->inet_sport = htons(lport); | |
a7f5e7f1 | 358 | sk->sk_hash = hash; |
547b792c | 359 | WARN_ON(!sk_unhashed(sk)); |
3ab5aee7 | 360 | __sk_nulls_add_node_rcu(sk, &head->chain); |
13475a30 | 361 | if (tw) { |
fc01538f | 362 | sk_nulls_del_node_init_rcu((struct sock *)tw); |
13475a30 ED |
363 | NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED); |
364 | } | |
9db66bdc | 365 | spin_unlock(lock); |
c29a0bc4 | 366 | sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); |
a7f5e7f1 ACM |
367 | |
368 | if (twp) { | |
369 | *twp = tw; | |
a7f5e7f1 ACM |
370 | } else if (tw) { |
371 | /* Silly. Should hash-dance instead... */ | |
dbe7faa4 | 372 | inet_twsk_deschedule_put(tw); |
a7f5e7f1 | 373 | } |
a7f5e7f1 ACM |
374 | return 0; |
375 | ||
376 | not_unique: | |
9db66bdc | 377 | spin_unlock(lock); |
a7f5e7f1 ACM |
378 | return -EADDRNOTAVAIL; |
379 | } | |
380 | ||
e2baad9e | 381 | static u32 inet_sk_port_offset(const struct sock *sk) |
a7f5e7f1 ACM |
382 | { |
383 | const struct inet_sock *inet = inet_sk(sk); | |
e2baad9e | 384 | |
c720c7e8 ED |
385 | return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr, |
386 | inet->inet_daddr, | |
387 | inet->inet_dport); | |
a7f5e7f1 ACM |
388 | } |
389 | ||
079096f1 ED |
390 | /* insert a socket into ehash, and eventually remove another one |
391 | * (The another one can be a SYN_RECV or TIMEWAIT | |
392 | */ | |
5e0724d0 | 393 | bool inet_ehash_insert(struct sock *sk, struct sock *osk) |
152da81d | 394 | { |
39d8cda7 | 395 | struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; |
3ab5aee7 | 396 | struct hlist_nulls_head *list; |
152da81d | 397 | struct inet_ehash_bucket *head; |
5b441f76 | 398 | spinlock_t *lock; |
5e0724d0 | 399 | bool ret = true; |
152da81d | 400 | |
079096f1 | 401 | WARN_ON_ONCE(!sk_unhashed(sk)); |
152da81d | 402 | |
5b441f76 | 403 | sk->sk_hash = sk_ehashfn(sk); |
152da81d PE |
404 | head = inet_ehash_bucket(hashinfo, sk->sk_hash); |
405 | list = &head->chain; | |
406 | lock = inet_ehash_lockp(hashinfo, sk->sk_hash); | |
407 | ||
9db66bdc | 408 | spin_lock(lock); |
fc01538f | 409 | if (osk) { |
5e0724d0 ED |
410 | WARN_ON_ONCE(sk->sk_hash != osk->sk_hash); |
411 | ret = sk_nulls_del_node_init_rcu(osk); | |
9327f705 | 412 | } |
5e0724d0 ED |
413 | if (ret) |
414 | __sk_nulls_add_node_rcu(sk, list); | |
9db66bdc | 415 | spin_unlock(lock); |
079096f1 ED |
416 | return ret; |
417 | } | |
418 | ||
5e0724d0 | 419 | bool inet_ehash_nolisten(struct sock *sk, struct sock *osk) |
079096f1 | 420 | { |
5e0724d0 ED |
421 | bool ok = inet_ehash_insert(sk, osk); |
422 | ||
423 | if (ok) { | |
424 | sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); | |
425 | } else { | |
426 | percpu_counter_inc(sk->sk_prot->orphan_count); | |
427 | sk->sk_state = TCP_CLOSE; | |
428 | sock_set_flag(sk, SOCK_DEAD); | |
429 | inet_csk_destroy_sock(sk); | |
430 | } | |
431 | return ok; | |
152da81d | 432 | } |
5e0724d0 | 433 | EXPORT_SYMBOL_GPL(inet_ehash_nolisten); |
152da81d | 434 | |
c125e80b CG |
435 | static int inet_reuseport_add_sock(struct sock *sk, |
436 | struct inet_listen_hashbucket *ilb, | |
437 | int (*saddr_same)(const struct sock *sk1, | |
438 | const struct sock *sk2, | |
439 | bool match_wildcard)) | |
440 | { | |
441 | struct sock *sk2; | |
442 | struct hlist_nulls_node *node; | |
443 | kuid_t uid = sock_i_uid(sk); | |
444 | ||
445 | sk_nulls_for_each_rcu(sk2, node, &ilb->head) { | |
446 | if (sk2 != sk && | |
447 | sk2->sk_family == sk->sk_family && | |
448 | ipv6_only_sock(sk2) == ipv6_only_sock(sk) && | |
449 | sk2->sk_bound_dev_if == sk->sk_bound_dev_if && | |
450 | sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) && | |
451 | saddr_same(sk, sk2, false)) | |
452 | return reuseport_add_sock(sk, sk2); | |
453 | } | |
454 | ||
455 | /* Initial allocation may have already happened via setsockopt */ | |
456 | if (!rcu_access_pointer(sk->sk_reuseport_cb)) | |
457 | return reuseport_alloc(sk); | |
458 | return 0; | |
459 | } | |
460 | ||
461 | int __inet_hash(struct sock *sk, struct sock *osk, | |
462 | int (*saddr_same)(const struct sock *sk1, | |
463 | const struct sock *sk2, | |
464 | bool match_wildcard)) | |
152da81d | 465 | { |
39d8cda7 | 466 | struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; |
5caea4ea | 467 | struct inet_listen_hashbucket *ilb; |
c125e80b | 468 | int err = 0; |
152da81d | 469 | |
5e0724d0 ED |
470 | if (sk->sk_state != TCP_LISTEN) { |
471 | inet_ehash_nolisten(sk, osk); | |
c125e80b | 472 | return 0; |
5e0724d0 | 473 | } |
547b792c | 474 | WARN_ON(!sk_unhashed(sk)); |
5caea4ea | 475 | ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)]; |
152da81d | 476 | |
5caea4ea | 477 | spin_lock(&ilb->lock); |
c125e80b CG |
478 | if (sk->sk_reuseport) { |
479 | err = inet_reuseport_add_sock(sk, ilb, saddr_same); | |
480 | if (err) | |
481 | goto unlock; | |
482 | } | |
3b24d854 ED |
483 | hlist_add_head_rcu(&sk->sk_node, &ilb->head); |
484 | sock_set_flag(sk, SOCK_RCU_FREE); | |
c29a0bc4 | 485 | sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); |
c125e80b | 486 | unlock: |
5caea4ea | 487 | spin_unlock(&ilb->lock); |
c125e80b CG |
488 | |
489 | return err; | |
152da81d | 490 | } |
77a6a471 | 491 | EXPORT_SYMBOL(__inet_hash); |
ab1e0a13 | 492 | |
086c653f | 493 | int inet_hash(struct sock *sk) |
ab1e0a13 | 494 | { |
c125e80b CG |
495 | int err = 0; |
496 | ||
ab1e0a13 ACM |
497 | if (sk->sk_state != TCP_CLOSE) { |
498 | local_bh_disable(); | |
c125e80b | 499 | err = __inet_hash(sk, NULL, ipv4_rcv_saddr_equal); |
ab1e0a13 ACM |
500 | local_bh_enable(); |
501 | } | |
086c653f | 502 | |
c125e80b | 503 | return err; |
ab1e0a13 ACM |
504 | } |
505 | EXPORT_SYMBOL_GPL(inet_hash); | |
506 | ||
507 | void inet_unhash(struct sock *sk) | |
508 | { | |
39d8cda7 | 509 | struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; |
c25eb3bf | 510 | spinlock_t *lock; |
3b24d854 | 511 | bool listener = false; |
c25eb3bf | 512 | int done; |
ab1e0a13 ACM |
513 | |
514 | if (sk_unhashed(sk)) | |
5caea4ea | 515 | return; |
ab1e0a13 | 516 | |
3b24d854 | 517 | if (sk->sk_state == TCP_LISTEN) { |
c25eb3bf | 518 | lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock; |
3b24d854 ED |
519 | listener = true; |
520 | } else { | |
c25eb3bf | 521 | lock = inet_ehash_lockp(hashinfo, sk->sk_hash); |
3b24d854 | 522 | } |
c25eb3bf | 523 | spin_lock_bh(lock); |
c125e80b CG |
524 | if (rcu_access_pointer(sk->sk_reuseport_cb)) |
525 | reuseport_detach_sock(sk); | |
3b24d854 ED |
526 | if (listener) |
527 | done = __sk_del_node_init(sk); | |
528 | else | |
529 | done = __sk_nulls_del_node_init_rcu(sk); | |
c25eb3bf ED |
530 | if (done) |
531 | sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1); | |
920de804 | 532 | spin_unlock_bh(lock); |
ab1e0a13 ACM |
533 | } |
534 | EXPORT_SYMBOL_GPL(inet_unhash); | |
152da81d | 535 | |
5ee31fc1 | 536 | int __inet_hash_connect(struct inet_timewait_death_row *death_row, |
5d8c0aa9 | 537 | struct sock *sk, u32 port_offset, |
5ee31fc1 | 538 | int (*check_established)(struct inet_timewait_death_row *, |
b4d6444e | 539 | struct sock *, __u16, struct inet_timewait_sock **)) |
a7f5e7f1 ACM |
540 | { |
541 | struct inet_hashinfo *hinfo = death_row->hashinfo; | |
1580ab63 | 542 | struct inet_timewait_sock *tw = NULL; |
e905a9ed | 543 | struct inet_bind_hashbucket *head; |
1580ab63 | 544 | int port = inet_sk(sk)->inet_num; |
3b1e0a65 | 545 | struct net *net = sock_net(sk); |
1580ab63 ED |
546 | struct inet_bind_bucket *tb; |
547 | u32 remaining, offset; | |
548 | int ret, i, low, high; | |
549 | static u32 hint; | |
550 | ||
551 | if (port) { | |
552 | head = &hinfo->bhash[inet_bhashfn(net, port, | |
553 | hinfo->bhash_size)]; | |
554 | tb = inet_csk(sk)->icsk_bind_hash; | |
555 | spin_lock_bh(&head->lock); | |
556 | if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) { | |
557 | inet_ehash_nolisten(sk, NULL); | |
558 | spin_unlock_bh(&head->lock); | |
559 | return 0; | |
560 | } | |
561 | spin_unlock(&head->lock); | |
562 | /* No definite answer... Walk to established hash table */ | |
563 | ret = check_established(death_row, sk, port, NULL); | |
564 | local_bh_enable(); | |
565 | return ret; | |
566 | } | |
a7f5e7f1 | 567 | |
1580ab63 ED |
568 | inet_get_local_port_range(net, &low, &high); |
569 | high++; /* [32768, 60999] -> [32768, 61000[ */ | |
570 | remaining = high - low; | |
571 | if (likely(remaining > 1)) | |
572 | remaining &= ~1U; | |
a7f5e7f1 | 573 | |
1580ab63 ED |
574 | offset = (hint + port_offset) % remaining; |
575 | /* In first pass we try ports of @low parity. | |
576 | * inet_csk_get_port() does the opposite choice. | |
577 | */ | |
578 | offset &= ~1U; | |
579 | other_parity_scan: | |
580 | port = low + offset; | |
581 | for (i = 0; i < remaining; i += 2, port += 2) { | |
582 | if (unlikely(port >= high)) | |
583 | port -= remaining; | |
584 | if (inet_is_local_reserved_port(net, port)) | |
585 | continue; | |
586 | head = &hinfo->bhash[inet_bhashfn(net, port, | |
587 | hinfo->bhash_size)]; | |
588 | spin_lock_bh(&head->lock); | |
227b60f5 | 589 | |
1580ab63 ED |
590 | /* Does not bother with rcv_saddr checks, because |
591 | * the established check is already unique enough. | |
07f4c900 | 592 | */ |
1580ab63 ED |
593 | inet_bind_bucket_for_each(tb, &head->chain) { |
594 | if (net_eq(ib_net(tb), net) && tb->port == port) { | |
595 | if (tb->fastreuse >= 0 || | |
596 | tb->fastreuseport >= 0) | |
e905a9ed | 597 | goto next_port; |
1580ab63 ED |
598 | WARN_ON(hlist_empty(&tb->owners)); |
599 | if (!check_established(death_row, sk, | |
600 | port, &tw)) | |
601 | goto ok; | |
602 | goto next_port; | |
e905a9ed | 603 | } |
e905a9ed | 604 | } |
a7f5e7f1 | 605 | |
1580ab63 ED |
606 | tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, |
607 | net, head, port); | |
608 | if (!tb) { | |
609 | spin_unlock_bh(&head->lock); | |
610 | return -ENOMEM; | |
e905a9ed | 611 | } |
1580ab63 ED |
612 | tb->fastreuse = -1; |
613 | tb->fastreuseport = -1; | |
614 | goto ok; | |
615 | next_port: | |
616 | spin_unlock_bh(&head->lock); | |
617 | cond_resched(); | |
618 | } | |
a7f5e7f1 | 619 | |
1580ab63 ED |
620 | offset++; |
621 | if ((offset & 1) && remaining > 1) | |
622 | goto other_parity_scan; | |
a7f5e7f1 | 623 | |
1580ab63 | 624 | return -EADDRNOTAVAIL; |
a7f5e7f1 | 625 | |
1580ab63 ED |
626 | ok: |
627 | hint += i + 2; | |
628 | ||
629 | /* Head lock still held and bh's disabled */ | |
630 | inet_bind_hash(sk, tb, port); | |
631 | if (sk_unhashed(sk)) { | |
632 | inet_sk(sk)->inet_sport = htons(port); | |
633 | inet_ehash_nolisten(sk, (struct sock *)tw); | |
a7f5e7f1 | 634 | } |
1580ab63 ED |
635 | if (tw) |
636 | inet_twsk_bind_unhash(tw, hinfo); | |
637 | spin_unlock(&head->lock); | |
638 | if (tw) | |
639 | inet_twsk_deschedule_put(tw); | |
640 | local_bh_enable(); | |
641 | return 0; | |
a7f5e7f1 | 642 | } |
5ee31fc1 PE |
643 | |
644 | /* | |
645 | * Bind a port for a connect operation and hash it. | |
646 | */ | |
647 | int inet_hash_connect(struct inet_timewait_death_row *death_row, | |
648 | struct sock *sk) | |
649 | { | |
e2baad9e ED |
650 | u32 port_offset = 0; |
651 | ||
652 | if (!inet_sk(sk)->inet_num) | |
653 | port_offset = inet_sk_port_offset(sk); | |
654 | return __inet_hash_connect(death_row, sk, port_offset, | |
b4d6444e | 655 | __inet_check_established); |
5ee31fc1 | 656 | } |
a7f5e7f1 | 657 | EXPORT_SYMBOL_GPL(inet_hash_connect); |
5caea4ea ED |
658 | |
659 | void inet_hashinfo_init(struct inet_hashinfo *h) | |
660 | { | |
661 | int i; | |
662 | ||
c25eb3bf | 663 | for (i = 0; i < INET_LHTABLE_SIZE; i++) { |
5caea4ea | 664 | spin_lock_init(&h->listening_hash[i].lock); |
3b24d854 ED |
665 | INIT_HLIST_HEAD(&h->listening_hash[i].head); |
666 | } | |
5caea4ea | 667 | } |
5caea4ea | 668 | EXPORT_SYMBOL_GPL(inet_hashinfo_init); |
095dc8e0 ED |
669 | |
670 | int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo) | |
671 | { | |
89e478a2 | 672 | unsigned int locksz = sizeof(spinlock_t); |
095dc8e0 ED |
673 | unsigned int i, nblocks = 1; |
674 | ||
89e478a2 | 675 | if (locksz != 0) { |
095dc8e0 | 676 | /* allocate 2 cache lines or at least one spinlock per cpu */ |
89e478a2 | 677 | nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U); |
095dc8e0 ED |
678 | nblocks = roundup_pow_of_two(nblocks * num_possible_cpus()); |
679 | ||
680 | /* no more locks than number of hash buckets */ | |
681 | nblocks = min(nblocks, hashinfo->ehash_mask + 1); | |
682 | ||
89e478a2 | 683 | hashinfo->ehash_locks = kmalloc_array(nblocks, locksz, |
095dc8e0 ED |
684 | GFP_KERNEL | __GFP_NOWARN); |
685 | if (!hashinfo->ehash_locks) | |
89e478a2 | 686 | hashinfo->ehash_locks = vmalloc(nblocks * locksz); |
095dc8e0 ED |
687 | |
688 | if (!hashinfo->ehash_locks) | |
689 | return -ENOMEM; | |
690 | ||
691 | for (i = 0; i < nblocks; i++) | |
692 | spin_lock_init(&hashinfo->ehash_locks[i]); | |
693 | } | |
694 | hashinfo->ehash_locks_mask = nblocks - 1; | |
695 | return 0; | |
696 | } | |
697 | EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc); |