]>
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> |
a04a480d | 28 | #include <net/tcp.h> |
c125e80b | 29 | #include <net/sock_reuseport.h> |
77d8bf9c | 30 | |
6eada011 ED |
31 | static u32 inet_ehashfn(const struct net *net, const __be32 laddr, |
32 | const __u16 lport, const __be32 faddr, | |
33 | const __be16 fport) | |
65cd8033 | 34 | { |
1bbdceef HFS |
35 | static u32 inet_ehash_secret __read_mostly; |
36 | ||
37 | net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret)); | |
38 | ||
65cd8033 HFS |
39 | return __inet_ehashfn(laddr, lport, faddr, fport, |
40 | inet_ehash_secret + net_hash_mix(net)); | |
41 | } | |
42 | ||
d1e559d0 ED |
43 | /* This function handles inet_sock, but also timewait and request sockets |
44 | * for IPv4/IPv6. | |
45 | */ | |
5b441f76 | 46 | u32 sk_ehashfn(const struct sock *sk) |
65cd8033 | 47 | { |
d1e559d0 ED |
48 | #if IS_ENABLED(CONFIG_IPV6) |
49 | if (sk->sk_family == AF_INET6 && | |
50 | !ipv6_addr_v4mapped(&sk->sk_v6_daddr)) | |
51 | return inet6_ehashfn(sock_net(sk), | |
52 | &sk->sk_v6_rcv_saddr, sk->sk_num, | |
53 | &sk->sk_v6_daddr, sk->sk_dport); | |
54 | #endif | |
5b441f76 ED |
55 | return inet_ehashfn(sock_net(sk), |
56 | sk->sk_rcv_saddr, sk->sk_num, | |
57 | sk->sk_daddr, sk->sk_dport); | |
65cd8033 HFS |
58 | } |
59 | ||
77d8bf9c ACM |
60 | /* |
61 | * Allocate and initialize a new local port bind bucket. | |
62 | * The bindhash mutex for snum's hash chain must be held here. | |
63 | */ | |
e18b890b | 64 | struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep, |
941b1d22 | 65 | struct net *net, |
77d8bf9c ACM |
66 | struct inet_bind_hashbucket *head, |
67 | const unsigned short snum) | |
68 | { | |
54e6ecb2 | 69 | struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC); |
77d8bf9c | 70 | |
00db4124 | 71 | if (tb) { |
efd7ef1c | 72 | write_pnet(&tb->ib_net, net); |
77d8bf9c ACM |
73 | tb->port = snum; |
74 | tb->fastreuse = 0; | |
da5e3630 | 75 | tb->fastreuseport = 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); |
463c84b9 | 98 | inet_csk(sk)->icsk_bind_hash = tb; |
2d8c4ce5 ACM |
99 | } |
100 | ||
2d8c4ce5 ACM |
101 | /* |
102 | * Get rid of any references to a local port held by the given sock. | |
103 | */ | |
ab1e0a13 | 104 | static void __inet_put_port(struct sock *sk) |
2d8c4ce5 | 105 | { |
39d8cda7 | 106 | struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; |
c720c7e8 | 107 | const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num, |
7f635ab7 | 108 | hashinfo->bhash_size); |
2d8c4ce5 ACM |
109 | struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash]; |
110 | struct inet_bind_bucket *tb; | |
111 | ||
112 | spin_lock(&head->lock); | |
463c84b9 | 113 | tb = inet_csk(sk)->icsk_bind_hash; |
2d8c4ce5 | 114 | __sk_del_bind_node(sk); |
463c84b9 | 115 | inet_csk(sk)->icsk_bind_hash = NULL; |
c720c7e8 | 116 | inet_sk(sk)->inet_num = 0; |
2d8c4ce5 ACM |
117 | inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb); |
118 | spin_unlock(&head->lock); | |
119 | } | |
120 | ||
ab1e0a13 | 121 | void inet_put_port(struct sock *sk) |
2d8c4ce5 ACM |
122 | { |
123 | local_bh_disable(); | |
ab1e0a13 | 124 | __inet_put_port(sk); |
2d8c4ce5 ACM |
125 | local_bh_enable(); |
126 | } | |
2d8c4ce5 | 127 | EXPORT_SYMBOL(inet_put_port); |
f3f05f70 | 128 | |
1ce31c9e | 129 | int __inet_inherit_port(const struct sock *sk, struct sock *child) |
53083773 PE |
130 | { |
131 | struct inet_hashinfo *table = sk->sk_prot->h.hashinfo; | |
093d2823 BS |
132 | unsigned short port = inet_sk(child)->inet_num; |
133 | const int bhash = inet_bhashfn(sock_net(sk), port, | |
7f635ab7 | 134 | table->bhash_size); |
53083773 PE |
135 | struct inet_bind_hashbucket *head = &table->bhash[bhash]; |
136 | struct inet_bind_bucket *tb; | |
137 | ||
138 | spin_lock(&head->lock); | |
139 | tb = inet_csk(sk)->icsk_bind_hash; | |
c2f34a65 ED |
140 | if (unlikely(!tb)) { |
141 | spin_unlock(&head->lock); | |
142 | return -ENOENT; | |
143 | } | |
093d2823 BS |
144 | if (tb->port != port) { |
145 | /* NOTE: using tproxy and redirecting skbs to a proxy | |
146 | * on a different listener port breaks the assumption | |
147 | * that the listener socket's icsk_bind_hash is the same | |
148 | * as that of the child socket. We have to look up or | |
149 | * create a new bind bucket for the child here. */ | |
b67bfe0d | 150 | inet_bind_bucket_for_each(tb, &head->chain) { |
093d2823 BS |
151 | if (net_eq(ib_net(tb), sock_net(sk)) && |
152 | tb->port == port) | |
153 | break; | |
154 | } | |
b67bfe0d | 155 | if (!tb) { |
093d2823 BS |
156 | tb = inet_bind_bucket_create(table->bind_bucket_cachep, |
157 | sock_net(sk), head, port); | |
158 | if (!tb) { | |
159 | spin_unlock(&head->lock); | |
160 | return -ENOMEM; | |
161 | } | |
162 | } | |
163 | } | |
b4ff3c90 | 164 | inet_bind_hash(child, tb, port); |
53083773 | 165 | spin_unlock(&head->lock); |
093d2823 BS |
166 | |
167 | return 0; | |
53083773 | 168 | } |
53083773 PE |
169 | EXPORT_SYMBOL_GPL(__inet_inherit_port); |
170 | ||
c25eb3bf ED |
171 | static inline int compute_score(struct sock *sk, struct net *net, |
172 | const unsigned short hnum, const __be32 daddr, | |
a04a480d | 173 | const int dif, bool exact_dif) |
c25eb3bf ED |
174 | { |
175 | int score = -1; | |
176 | struct inet_sock *inet = inet_sk(sk); | |
177 | ||
c720c7e8 | 178 | if (net_eq(sock_net(sk), net) && inet->inet_num == hnum && |
c25eb3bf | 179 | !ipv6_only_sock(sk)) { |
c720c7e8 | 180 | __be32 rcv_saddr = inet->inet_rcv_saddr; |
da5e3630 | 181 | score = sk->sk_family == PF_INET ? 2 : 1; |
c25eb3bf ED |
182 | if (rcv_saddr) { |
183 | if (rcv_saddr != daddr) | |
184 | return -1; | |
da5e3630 | 185 | score += 4; |
c25eb3bf | 186 | } |
a04a480d | 187 | if (sk->sk_bound_dev_if || exact_dif) { |
c25eb3bf ED |
188 | if (sk->sk_bound_dev_if != dif) |
189 | return -1; | |
da5e3630 | 190 | score += 4; |
c25eb3bf | 191 | } |
70da268b ED |
192 | if (sk->sk_incoming_cpu == raw_smp_processor_id()) |
193 | score++; | |
c25eb3bf ED |
194 | } |
195 | return score; | |
196 | } | |
197 | ||
33b62231 | 198 | /* |
3b24d854 ED |
199 | * Here are some nice properties to exploit here. The BSD API |
200 | * does not allow a listening sock to specify the remote port nor the | |
33b62231 ACM |
201 | * remote address for the connection. So always assume those are both |
202 | * wildcarded during the search since they can never be otherwise. | |
203 | */ | |
e48c414e | 204 | |
3b24d854 | 205 | /* called with rcu_read_lock() : No refcount taken on the socket */ |
c67499c0 PE |
206 | struct sock *__inet_lookup_listener(struct net *net, |
207 | struct inet_hashinfo *hashinfo, | |
a583636a | 208 | struct sk_buff *skb, int doff, |
da5e3630 | 209 | const __be32 saddr, __be16 sport, |
fb99c848 | 210 | const __be32 daddr, const unsigned short hnum, |
8f491069 | 211 | const int dif) |
99a92ff5 | 212 | { |
c25eb3bf ED |
213 | unsigned int hash = inet_lhashfn(net, hnum); |
214 | struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash]; | |
3b24d854 | 215 | int score, hiscore = 0, matches = 0, reuseport = 0; |
a04a480d | 216 | bool exact_dif = inet_exact_dif_match(net, skb); |
3b24d854 | 217 | struct sock *sk, *result = NULL; |
da5e3630 | 218 | u32 phash = 0; |
99a92ff5 | 219 | |
3b24d854 | 220 | sk_for_each_rcu(sk, &ilb->head) { |
a04a480d | 221 | score = compute_score(sk, net, hnum, daddr, dif, exact_dif); |
c25eb3bf | 222 | if (score > hiscore) { |
da5e3630 TH |
223 | reuseport = sk->sk_reuseport; |
224 | if (reuseport) { | |
225 | phash = inet_ehashfn(net, daddr, hnum, | |
226 | saddr, sport); | |
3b24d854 ED |
227 | result = reuseport_select_sock(sk, phash, |
228 | skb, doff); | |
229 | if (result) | |
230 | return result; | |
da5e3630 TH |
231 | matches = 1; |
232 | } | |
3b24d854 ED |
233 | result = sk; |
234 | hiscore = score; | |
da5e3630 TH |
235 | } else if (score == hiscore && reuseport) { |
236 | matches++; | |
8fc54f68 | 237 | if (reciprocal_scale(phash, matches) == 0) |
da5e3630 TH |
238 | result = sk; |
239 | phash = next_pseudo_random32(phash); | |
c25eb3bf | 240 | } |
99a92ff5 | 241 | } |
c25eb3bf | 242 | return result; |
99a92ff5 | 243 | } |
8f491069 | 244 | EXPORT_SYMBOL_GPL(__inet_lookup_listener); |
a7f5e7f1 | 245 | |
05dbc7b5 ED |
246 | /* All sockets share common refcount, but have different destructors */ |
247 | void sock_gen_put(struct sock *sk) | |
248 | { | |
249 | if (!atomic_dec_and_test(&sk->sk_refcnt)) | |
250 | return; | |
251 | ||
252 | if (sk->sk_state == TCP_TIME_WAIT) | |
253 | inet_twsk_free(inet_twsk(sk)); | |
41b822c5 ED |
254 | else if (sk->sk_state == TCP_NEW_SYN_RECV) |
255 | reqsk_free(inet_reqsk(sk)); | |
05dbc7b5 ED |
256 | else |
257 | sk_free(sk); | |
258 | } | |
259 | EXPORT_SYMBOL_GPL(sock_gen_put); | |
260 | ||
2c13270b ED |
261 | void sock_edemux(struct sk_buff *skb) |
262 | { | |
263 | sock_gen_put(skb->sk); | |
264 | } | |
265 | EXPORT_SYMBOL(sock_edemux); | |
266 | ||
5e73ea1a | 267 | struct sock *__inet_lookup_established(struct net *net, |
c67499c0 | 268 | struct inet_hashinfo *hashinfo, |
77a5ba55 PE |
269 | const __be32 saddr, const __be16 sport, |
270 | const __be32 daddr, const u16 hnum, | |
271 | const int dif) | |
272 | { | |
c7228317 | 273 | INET_ADDR_COOKIE(acookie, saddr, daddr); |
77a5ba55 PE |
274 | const __portpair ports = INET_COMBINED_PORTS(sport, hnum); |
275 | struct sock *sk; | |
3ab5aee7 | 276 | const struct hlist_nulls_node *node; |
77a5ba55 PE |
277 | /* Optimize here for direct hit, only listening connections can |
278 | * have wildcards anyways. | |
279 | */ | |
9f26b3ad | 280 | unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport); |
f373b53b | 281 | unsigned int slot = hash & hashinfo->ehash_mask; |
3ab5aee7 | 282 | struct inet_ehash_bucket *head = &hashinfo->ehash[slot]; |
77a5ba55 | 283 | |
3ab5aee7 ED |
284 | begin: |
285 | sk_nulls_for_each_rcu(sk, node, &head->chain) { | |
ce43b03e ED |
286 | if (sk->sk_hash != hash) |
287 | continue; | |
288 | if (likely(INET_MATCH(sk, net, acookie, | |
289 | saddr, daddr, ports, dif))) { | |
3ab5aee7 | 290 | if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt))) |
05dbc7b5 | 291 | goto out; |
ce43b03e ED |
292 | if (unlikely(!INET_MATCH(sk, net, acookie, |
293 | saddr, daddr, ports, dif))) { | |
05dbc7b5 | 294 | sock_gen_put(sk); |
3ab5aee7 ED |
295 | goto begin; |
296 | } | |
05dbc7b5 | 297 | goto found; |
3ab5aee7 | 298 | } |
77a5ba55 | 299 | } |
3ab5aee7 ED |
300 | /* |
301 | * if the nulls value we got at the end of this lookup is | |
302 | * not the expected one, we must restart lookup. | |
303 | * We probably met an item that was moved to another chain. | |
304 | */ | |
305 | if (get_nulls_value(node) != slot) | |
306 | goto begin; | |
77a5ba55 | 307 | out: |
05dbc7b5 ED |
308 | sk = NULL; |
309 | found: | |
77a5ba55 | 310 | return sk; |
77a5ba55 PE |
311 | } |
312 | EXPORT_SYMBOL_GPL(__inet_lookup_established); | |
313 | ||
a7f5e7f1 ACM |
314 | /* called with local bh disabled */ |
315 | static int __inet_check_established(struct inet_timewait_death_row *death_row, | |
316 | struct sock *sk, __u16 lport, | |
317 | struct inet_timewait_sock **twp) | |
318 | { | |
319 | struct inet_hashinfo *hinfo = death_row->hashinfo; | |
320 | struct inet_sock *inet = inet_sk(sk); | |
c720c7e8 ED |
321 | __be32 daddr = inet->inet_rcv_saddr; |
322 | __be32 saddr = inet->inet_daddr; | |
a7f5e7f1 | 323 | int dif = sk->sk_bound_dev_if; |
c7228317 | 324 | INET_ADDR_COOKIE(acookie, saddr, daddr); |
c720c7e8 | 325 | const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport); |
9f26b3ad | 326 | struct net *net = sock_net(sk); |
c720c7e8 ED |
327 | unsigned int hash = inet_ehashfn(net, daddr, lport, |
328 | saddr, inet->inet_dport); | |
a7f5e7f1 | 329 | struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash); |
9db66bdc | 330 | spinlock_t *lock = inet_ehash_lockp(hinfo, hash); |
a7f5e7f1 | 331 | struct sock *sk2; |
3ab5aee7 | 332 | const struct hlist_nulls_node *node; |
05dbc7b5 | 333 | struct inet_timewait_sock *tw = NULL; |
a7f5e7f1 | 334 | |
9db66bdc | 335 | spin_lock(lock); |
a7f5e7f1 | 336 | |
3ab5aee7 | 337 | sk_nulls_for_each(sk2, node, &head->chain) { |
ce43b03e ED |
338 | if (sk2->sk_hash != hash) |
339 | continue; | |
05dbc7b5 | 340 | |
ce43b03e | 341 | if (likely(INET_MATCH(sk2, net, acookie, |
05dbc7b5 ED |
342 | saddr, daddr, ports, dif))) { |
343 | if (sk2->sk_state == TCP_TIME_WAIT) { | |
344 | tw = inet_twsk(sk2); | |
345 | if (twsk_unique(sk, sk2, twp)) | |
346 | break; | |
347 | } | |
a7f5e7f1 | 348 | goto not_unique; |
05dbc7b5 | 349 | } |
a7f5e7f1 ACM |
350 | } |
351 | ||
a7f5e7f1 | 352 | /* Must record num and sport now. Otherwise we will see |
05dbc7b5 ED |
353 | * in hash table socket with a funny identity. |
354 | */ | |
c720c7e8 ED |
355 | inet->inet_num = lport; |
356 | inet->inet_sport = htons(lport); | |
a7f5e7f1 | 357 | sk->sk_hash = hash; |
547b792c | 358 | WARN_ON(!sk_unhashed(sk)); |
3ab5aee7 | 359 | __sk_nulls_add_node_rcu(sk, &head->chain); |
13475a30 | 360 | if (tw) { |
fc01538f | 361 | sk_nulls_del_node_init_rcu((struct sock *)tw); |
02a1d6e7 | 362 | __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED); |
13475a30 | 363 | } |
9db66bdc | 364 | spin_unlock(lock); |
c29a0bc4 | 365 | sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); |
a7f5e7f1 ACM |
366 | |
367 | if (twp) { | |
368 | *twp = tw; | |
a7f5e7f1 ACM |
369 | } else if (tw) { |
370 | /* Silly. Should hash-dance instead... */ | |
dbe7faa4 | 371 | inet_twsk_deschedule_put(tw); |
a7f5e7f1 | 372 | } |
a7f5e7f1 ACM |
373 | return 0; |
374 | ||
375 | not_unique: | |
9db66bdc | 376 | spin_unlock(lock); |
a7f5e7f1 ACM |
377 | return -EADDRNOTAVAIL; |
378 | } | |
379 | ||
e2baad9e | 380 | static u32 inet_sk_port_offset(const struct sock *sk) |
a7f5e7f1 ACM |
381 | { |
382 | const struct inet_sock *inet = inet_sk(sk); | |
e2baad9e | 383 | |
c720c7e8 ED |
384 | return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr, |
385 | inet->inet_daddr, | |
386 | inet->inet_dport); | |
a7f5e7f1 ACM |
387 | } |
388 | ||
079096f1 ED |
389 | /* insert a socket into ehash, and eventually remove another one |
390 | * (The another one can be a SYN_RECV or TIMEWAIT | |
391 | */ | |
5e0724d0 | 392 | bool inet_ehash_insert(struct sock *sk, struct sock *osk) |
152da81d | 393 | { |
39d8cda7 | 394 | struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; |
3ab5aee7 | 395 | struct hlist_nulls_head *list; |
152da81d | 396 | struct inet_ehash_bucket *head; |
5b441f76 | 397 | spinlock_t *lock; |
5e0724d0 | 398 | bool ret = true; |
152da81d | 399 | |
079096f1 | 400 | WARN_ON_ONCE(!sk_unhashed(sk)); |
152da81d | 401 | |
5b441f76 | 402 | sk->sk_hash = sk_ehashfn(sk); |
152da81d PE |
403 | head = inet_ehash_bucket(hashinfo, sk->sk_hash); |
404 | list = &head->chain; | |
405 | lock = inet_ehash_lockp(hashinfo, sk->sk_hash); | |
406 | ||
9db66bdc | 407 | spin_lock(lock); |
fc01538f | 408 | if (osk) { |
5e0724d0 ED |
409 | WARN_ON_ONCE(sk->sk_hash != osk->sk_hash); |
410 | ret = sk_nulls_del_node_init_rcu(osk); | |
9327f705 | 411 | } |
5e0724d0 ED |
412 | if (ret) |
413 | __sk_nulls_add_node_rcu(sk, list); | |
9db66bdc | 414 | spin_unlock(lock); |
079096f1 ED |
415 | return ret; |
416 | } | |
417 | ||
5e0724d0 | 418 | bool inet_ehash_nolisten(struct sock *sk, struct sock *osk) |
079096f1 | 419 | { |
5e0724d0 ED |
420 | bool ok = inet_ehash_insert(sk, osk); |
421 | ||
422 | if (ok) { | |
423 | sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); | |
424 | } else { | |
425 | percpu_counter_inc(sk->sk_prot->orphan_count); | |
426 | sk->sk_state = TCP_CLOSE; | |
427 | sock_set_flag(sk, SOCK_DEAD); | |
428 | inet_csk_destroy_sock(sk); | |
429 | } | |
430 | return ok; | |
152da81d | 431 | } |
5e0724d0 | 432 | EXPORT_SYMBOL_GPL(inet_ehash_nolisten); |
152da81d | 433 | |
c125e80b | 434 | static int inet_reuseport_add_sock(struct sock *sk, |
fe38d2a1 | 435 | struct inet_listen_hashbucket *ilb) |
c125e80b | 436 | { |
90e5d0db | 437 | struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash; |
c125e80b | 438 | struct sock *sk2; |
c125e80b CG |
439 | kuid_t uid = sock_i_uid(sk); |
440 | ||
85017869 | 441 | sk_for_each_rcu(sk2, &ilb->head) { |
c125e80b CG |
442 | if (sk2 != sk && |
443 | sk2->sk_family == sk->sk_family && | |
444 | ipv6_only_sock(sk2) == ipv6_only_sock(sk) && | |
445 | sk2->sk_bound_dev_if == sk->sk_bound_dev_if && | |
90e5d0db | 446 | inet_csk(sk2)->icsk_bind_hash == tb && |
c125e80b | 447 | sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) && |
fe38d2a1 | 448 | inet_rcv_saddr_equal(sk, sk2, false)) |
c125e80b CG |
449 | return reuseport_add_sock(sk, sk2); |
450 | } | |
451 | ||
452 | /* Initial allocation may have already happened via setsockopt */ | |
453 | if (!rcu_access_pointer(sk->sk_reuseport_cb)) | |
454 | return reuseport_alloc(sk); | |
455 | return 0; | |
456 | } | |
457 | ||
fe38d2a1 | 458 | int __inet_hash(struct sock *sk, struct sock *osk) |
152da81d | 459 | { |
39d8cda7 | 460 | struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; |
5caea4ea | 461 | struct inet_listen_hashbucket *ilb; |
c125e80b | 462 | int err = 0; |
152da81d | 463 | |
5e0724d0 ED |
464 | if (sk->sk_state != TCP_LISTEN) { |
465 | inet_ehash_nolisten(sk, osk); | |
c125e80b | 466 | return 0; |
5e0724d0 | 467 | } |
547b792c | 468 | WARN_ON(!sk_unhashed(sk)); |
5caea4ea | 469 | ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)]; |
152da81d | 470 | |
5caea4ea | 471 | spin_lock(&ilb->lock); |
c125e80b | 472 | if (sk->sk_reuseport) { |
fe38d2a1 | 473 | err = inet_reuseport_add_sock(sk, ilb); |
c125e80b CG |
474 | if (err) |
475 | goto unlock; | |
476 | } | |
d296ba60 CG |
477 | if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport && |
478 | sk->sk_family == AF_INET6) | |
479 | hlist_add_tail_rcu(&sk->sk_node, &ilb->head); | |
480 | else | |
481 | hlist_add_head_rcu(&sk->sk_node, &ilb->head); | |
3b24d854 | 482 | sock_set_flag(sk, SOCK_RCU_FREE); |
c29a0bc4 | 483 | sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); |
c125e80b | 484 | unlock: |
5caea4ea | 485 | spin_unlock(&ilb->lock); |
c125e80b CG |
486 | |
487 | return err; | |
152da81d | 488 | } |
77a6a471 | 489 | EXPORT_SYMBOL(__inet_hash); |
ab1e0a13 | 490 | |
086c653f | 491 | int inet_hash(struct sock *sk) |
ab1e0a13 | 492 | { |
c125e80b CG |
493 | int err = 0; |
494 | ||
ab1e0a13 ACM |
495 | if (sk->sk_state != TCP_CLOSE) { |
496 | local_bh_disable(); | |
fe38d2a1 | 497 | err = __inet_hash(sk, NULL); |
ab1e0a13 ACM |
498 | local_bh_enable(); |
499 | } | |
086c653f | 500 | |
c125e80b | 501 | return err; |
ab1e0a13 ACM |
502 | } |
503 | EXPORT_SYMBOL_GPL(inet_hash); | |
504 | ||
505 | void inet_unhash(struct sock *sk) | |
506 | { | |
39d8cda7 | 507 | struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; |
c25eb3bf | 508 | spinlock_t *lock; |
3b24d854 | 509 | bool listener = false; |
c25eb3bf | 510 | int done; |
ab1e0a13 ACM |
511 | |
512 | if (sk_unhashed(sk)) | |
5caea4ea | 513 | return; |
ab1e0a13 | 514 | |
3b24d854 | 515 | if (sk->sk_state == TCP_LISTEN) { |
c25eb3bf | 516 | lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock; |
3b24d854 ED |
517 | listener = true; |
518 | } else { | |
c25eb3bf | 519 | lock = inet_ehash_lockp(hashinfo, sk->sk_hash); |
3b24d854 | 520 | } |
c25eb3bf | 521 | spin_lock_bh(lock); |
c125e80b CG |
522 | if (rcu_access_pointer(sk->sk_reuseport_cb)) |
523 | reuseport_detach_sock(sk); | |
3b24d854 ED |
524 | if (listener) |
525 | done = __sk_del_node_init(sk); | |
526 | else | |
527 | done = __sk_nulls_del_node_init_rcu(sk); | |
c25eb3bf ED |
528 | if (done) |
529 | sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1); | |
920de804 | 530 | spin_unlock_bh(lock); |
ab1e0a13 ACM |
531 | } |
532 | EXPORT_SYMBOL_GPL(inet_unhash); | |
152da81d | 533 | |
5ee31fc1 | 534 | int __inet_hash_connect(struct inet_timewait_death_row *death_row, |
5d8c0aa9 | 535 | struct sock *sk, u32 port_offset, |
5ee31fc1 | 536 | int (*check_established)(struct inet_timewait_death_row *, |
b4d6444e | 537 | struct sock *, __u16, struct inet_timewait_sock **)) |
a7f5e7f1 ACM |
538 | { |
539 | struct inet_hashinfo *hinfo = death_row->hashinfo; | |
1580ab63 | 540 | struct inet_timewait_sock *tw = NULL; |
e905a9ed | 541 | struct inet_bind_hashbucket *head; |
1580ab63 | 542 | int port = inet_sk(sk)->inet_num; |
3b1e0a65 | 543 | struct net *net = sock_net(sk); |
1580ab63 ED |
544 | struct inet_bind_bucket *tb; |
545 | u32 remaining, offset; | |
546 | int ret, i, low, high; | |
547 | static u32 hint; | |
548 | ||
549 | if (port) { | |
550 | head = &hinfo->bhash[inet_bhashfn(net, port, | |
551 | hinfo->bhash_size)]; | |
552 | tb = inet_csk(sk)->icsk_bind_hash; | |
553 | spin_lock_bh(&head->lock); | |
554 | if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) { | |
555 | inet_ehash_nolisten(sk, NULL); | |
556 | spin_unlock_bh(&head->lock); | |
557 | return 0; | |
558 | } | |
559 | spin_unlock(&head->lock); | |
560 | /* No definite answer... Walk to established hash table */ | |
561 | ret = check_established(death_row, sk, port, NULL); | |
562 | local_bh_enable(); | |
563 | return ret; | |
564 | } | |
a7f5e7f1 | 565 | |
1580ab63 ED |
566 | inet_get_local_port_range(net, &low, &high); |
567 | high++; /* [32768, 60999] -> [32768, 61000[ */ | |
568 | remaining = high - low; | |
569 | if (likely(remaining > 1)) | |
570 | remaining &= ~1U; | |
a7f5e7f1 | 571 | |
1580ab63 ED |
572 | offset = (hint + port_offset) % remaining; |
573 | /* In first pass we try ports of @low parity. | |
574 | * inet_csk_get_port() does the opposite choice. | |
575 | */ | |
576 | offset &= ~1U; | |
577 | other_parity_scan: | |
578 | port = low + offset; | |
579 | for (i = 0; i < remaining; i += 2, port += 2) { | |
580 | if (unlikely(port >= high)) | |
581 | port -= remaining; | |
582 | if (inet_is_local_reserved_port(net, port)) | |
583 | continue; | |
584 | head = &hinfo->bhash[inet_bhashfn(net, port, | |
585 | hinfo->bhash_size)]; | |
586 | spin_lock_bh(&head->lock); | |
227b60f5 | 587 | |
1580ab63 ED |
588 | /* Does not bother with rcv_saddr checks, because |
589 | * the established check is already unique enough. | |
07f4c900 | 590 | */ |
1580ab63 ED |
591 | inet_bind_bucket_for_each(tb, &head->chain) { |
592 | if (net_eq(ib_net(tb), net) && tb->port == port) { | |
593 | if (tb->fastreuse >= 0 || | |
594 | tb->fastreuseport >= 0) | |
e905a9ed | 595 | goto next_port; |
1580ab63 ED |
596 | WARN_ON(hlist_empty(&tb->owners)); |
597 | if (!check_established(death_row, sk, | |
598 | port, &tw)) | |
599 | goto ok; | |
600 | goto next_port; | |
e905a9ed | 601 | } |
e905a9ed | 602 | } |
a7f5e7f1 | 603 | |
1580ab63 ED |
604 | tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, |
605 | net, head, port); | |
606 | if (!tb) { | |
607 | spin_unlock_bh(&head->lock); | |
608 | return -ENOMEM; | |
e905a9ed | 609 | } |
1580ab63 ED |
610 | tb->fastreuse = -1; |
611 | tb->fastreuseport = -1; | |
612 | goto ok; | |
613 | next_port: | |
614 | spin_unlock_bh(&head->lock); | |
615 | cond_resched(); | |
616 | } | |
a7f5e7f1 | 617 | |
1580ab63 ED |
618 | offset++; |
619 | if ((offset & 1) && remaining > 1) | |
620 | goto other_parity_scan; | |
a7f5e7f1 | 621 | |
1580ab63 | 622 | return -EADDRNOTAVAIL; |
a7f5e7f1 | 623 | |
1580ab63 ED |
624 | ok: |
625 | hint += i + 2; | |
626 | ||
627 | /* Head lock still held and bh's disabled */ | |
628 | inet_bind_hash(sk, tb, port); | |
629 | if (sk_unhashed(sk)) { | |
630 | inet_sk(sk)->inet_sport = htons(port); | |
631 | inet_ehash_nolisten(sk, (struct sock *)tw); | |
a7f5e7f1 | 632 | } |
1580ab63 ED |
633 | if (tw) |
634 | inet_twsk_bind_unhash(tw, hinfo); | |
635 | spin_unlock(&head->lock); | |
636 | if (tw) | |
637 | inet_twsk_deschedule_put(tw); | |
638 | local_bh_enable(); | |
639 | return 0; | |
a7f5e7f1 | 640 | } |
5ee31fc1 PE |
641 | |
642 | /* | |
643 | * Bind a port for a connect operation and hash it. | |
644 | */ | |
645 | int inet_hash_connect(struct inet_timewait_death_row *death_row, | |
646 | struct sock *sk) | |
647 | { | |
e2baad9e ED |
648 | u32 port_offset = 0; |
649 | ||
650 | if (!inet_sk(sk)->inet_num) | |
651 | port_offset = inet_sk_port_offset(sk); | |
652 | return __inet_hash_connect(death_row, sk, port_offset, | |
b4d6444e | 653 | __inet_check_established); |
5ee31fc1 | 654 | } |
a7f5e7f1 | 655 | EXPORT_SYMBOL_GPL(inet_hash_connect); |
5caea4ea ED |
656 | |
657 | void inet_hashinfo_init(struct inet_hashinfo *h) | |
658 | { | |
659 | int i; | |
660 | ||
c25eb3bf | 661 | for (i = 0; i < INET_LHTABLE_SIZE; i++) { |
5caea4ea | 662 | spin_lock_init(&h->listening_hash[i].lock); |
3b24d854 ED |
663 | INIT_HLIST_HEAD(&h->listening_hash[i].head); |
664 | } | |
5caea4ea | 665 | } |
5caea4ea | 666 | EXPORT_SYMBOL_GPL(inet_hashinfo_init); |
095dc8e0 ED |
667 | |
668 | int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo) | |
669 | { | |
89e478a2 | 670 | unsigned int locksz = sizeof(spinlock_t); |
095dc8e0 ED |
671 | unsigned int i, nblocks = 1; |
672 | ||
89e478a2 | 673 | if (locksz != 0) { |
095dc8e0 | 674 | /* allocate 2 cache lines or at least one spinlock per cpu */ |
89e478a2 | 675 | nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U); |
095dc8e0 ED |
676 | nblocks = roundup_pow_of_two(nblocks * num_possible_cpus()); |
677 | ||
678 | /* no more locks than number of hash buckets */ | |
679 | nblocks = min(nblocks, hashinfo->ehash_mask + 1); | |
680 | ||
752ade68 | 681 | hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL); |
095dc8e0 ED |
682 | if (!hashinfo->ehash_locks) |
683 | return -ENOMEM; | |
684 | ||
685 | for (i = 0; i < nblocks; i++) | |
686 | spin_lock_init(&hashinfo->ehash_locks[i]); | |
687 | } | |
688 | hashinfo->ehash_locks_mask = nblocks - 1; | |
689 | return 0; | |
690 | } | |
691 | EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc); |