Loading...
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
16#include <linux/module.h>
17#include <linux/random.h>
18#include <linux/sched.h>
19#include <linux/slab.h>
20#include <linux/wait.h>
21#include <linux/vmalloc.h>
22#include <linux/bootmem.h>
23
24#include <net/addrconf.h>
25#include <net/inet_connection_sock.h>
26#include <net/inet_hashtables.h>
27#include <net/secure_seq.h>
28#include <net/ip.h>
29#include <net/tcp.h>
30#include <net/sock_reuseport.h>
31
32static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
33 const __u16 lport, const __be32 faddr,
34 const __be16 fport)
35{
36 static u32 inet_ehash_secret __read_mostly;
37
38 net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
39
40 return __inet_ehashfn(laddr, lport, faddr, fport,
41 inet_ehash_secret + net_hash_mix(net));
42}
43
44/* This function handles inet_sock, but also timewait and request sockets
45 * for IPv4/IPv6.
46 */
47static u32 sk_ehashfn(const struct sock *sk)
48{
49#if IS_ENABLED(CONFIG_IPV6)
50 if (sk->sk_family == AF_INET6 &&
51 !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
52 return inet6_ehashfn(sock_net(sk),
53 &sk->sk_v6_rcv_saddr, sk->sk_num,
54 &sk->sk_v6_daddr, sk->sk_dport);
55#endif
56 return inet_ehashfn(sock_net(sk),
57 sk->sk_rcv_saddr, sk->sk_num,
58 sk->sk_daddr, sk->sk_dport);
59}
60
61/*
62 * Allocate and initialize a new local port bind bucket.
63 * The bindhash mutex for snum's hash chain must be held here.
64 */
65struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
66 struct net *net,
67 struct inet_bind_hashbucket *head,
68 const unsigned short snum)
69{
70 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
71
72 if (tb) {
73 write_pnet(&tb->ib_net, net);
74 tb->port = snum;
75 tb->fastreuse = 0;
76 tb->fastreuseport = 0;
77 INIT_HLIST_HEAD(&tb->owners);
78 hlist_add_head(&tb->node, &head->chain);
79 }
80 return tb;
81}
82
83/*
84 * Caller must hold hashbucket lock for this tb with local BH disabled
85 */
86void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
87{
88 if (hlist_empty(&tb->owners)) {
89 __hlist_del(&tb->node);
90 kmem_cache_free(cachep, tb);
91 }
92}
93
94void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
95 const unsigned short snum)
96{
97 inet_sk(sk)->inet_num = snum;
98 sk_add_bind_node(sk, &tb->owners);
99 inet_csk(sk)->icsk_bind_hash = tb;
100}
101
102/*
103 * Get rid of any references to a local port held by the given sock.
104 */
105static void __inet_put_port(struct sock *sk)
106{
107 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
108 const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
109 hashinfo->bhash_size);
110 struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
111 struct inet_bind_bucket *tb;
112
113 spin_lock(&head->lock);
114 tb = inet_csk(sk)->icsk_bind_hash;
115 __sk_del_bind_node(sk);
116 inet_csk(sk)->icsk_bind_hash = NULL;
117 inet_sk(sk)->inet_num = 0;
118 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
119 spin_unlock(&head->lock);
120}
121
122void inet_put_port(struct sock *sk)
123{
124 local_bh_disable();
125 __inet_put_port(sk);
126 local_bh_enable();
127}
128EXPORT_SYMBOL(inet_put_port);
129
130int __inet_inherit_port(const struct sock *sk, struct sock *child)
131{
132 struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
133 unsigned short port = inet_sk(child)->inet_num;
134 const int bhash = inet_bhashfn(sock_net(sk), port,
135 table->bhash_size);
136 struct inet_bind_hashbucket *head = &table->bhash[bhash];
137 struct inet_bind_bucket *tb;
138
139 spin_lock(&head->lock);
140 tb = inet_csk(sk)->icsk_bind_hash;
141 if (unlikely(!tb)) {
142 spin_unlock(&head->lock);
143 return -ENOENT;
144 }
145 if (tb->port != port) {
146 /* NOTE: using tproxy and redirecting skbs to a proxy
147 * on a different listener port breaks the assumption
148 * that the listener socket's icsk_bind_hash is the same
149 * as that of the child socket. We have to look up or
150 * create a new bind bucket for the child here. */
151 inet_bind_bucket_for_each(tb, &head->chain) {
152 if (net_eq(ib_net(tb), sock_net(sk)) &&
153 tb->port == port)
154 break;
155 }
156 if (!tb) {
157 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
158 sock_net(sk), head, port);
159 if (!tb) {
160 spin_unlock(&head->lock);
161 return -ENOMEM;
162 }
163 }
164 }
165 inet_bind_hash(child, tb, port);
166 spin_unlock(&head->lock);
167
168 return 0;
169}
170EXPORT_SYMBOL_GPL(__inet_inherit_port);
171
172static struct inet_listen_hashbucket *
173inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
174{
175 u32 hash;
176
177#if IS_ENABLED(CONFIG_IPV6)
178 if (sk->sk_family == AF_INET6)
179 hash = ipv6_portaddr_hash(sock_net(sk),
180 &sk->sk_v6_rcv_saddr,
181 inet_sk(sk)->inet_num);
182 else
183#endif
184 hash = ipv4_portaddr_hash(sock_net(sk),
185 inet_sk(sk)->inet_rcv_saddr,
186 inet_sk(sk)->inet_num);
187 return inet_lhash2_bucket(h, hash);
188}
189
190static void inet_hash2(struct inet_hashinfo *h, struct sock *sk)
191{
192 struct inet_listen_hashbucket *ilb2;
193
194 if (!h->lhash2)
195 return;
196
197 ilb2 = inet_lhash2_bucket_sk(h, sk);
198
199 spin_lock(&ilb2->lock);
200 if (sk->sk_reuseport && sk->sk_family == AF_INET6)
201 hlist_add_tail_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
202 &ilb2->head);
203 else
204 hlist_add_head_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
205 &ilb2->head);
206 ilb2->count++;
207 spin_unlock(&ilb2->lock);
208}
209
210static void inet_unhash2(struct inet_hashinfo *h, struct sock *sk)
211{
212 struct inet_listen_hashbucket *ilb2;
213
214 if (!h->lhash2 ||
215 WARN_ON_ONCE(hlist_unhashed(&inet_csk(sk)->icsk_listen_portaddr_node)))
216 return;
217
218 ilb2 = inet_lhash2_bucket_sk(h, sk);
219
220 spin_lock(&ilb2->lock);
221 hlist_del_init_rcu(&inet_csk(sk)->icsk_listen_portaddr_node);
222 ilb2->count--;
223 spin_unlock(&ilb2->lock);
224}
225
226static inline int compute_score(struct sock *sk, struct net *net,
227 const unsigned short hnum, const __be32 daddr,
228 const int dif, const int sdif, bool exact_dif)
229{
230 int score = -1;
231 struct inet_sock *inet = inet_sk(sk);
232
233 if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
234 !ipv6_only_sock(sk)) {
235 __be32 rcv_saddr = inet->inet_rcv_saddr;
236 score = sk->sk_family == PF_INET ? 2 : 1;
237 if (rcv_saddr) {
238 if (rcv_saddr != daddr)
239 return -1;
240 score += 4;
241 }
242 if (sk->sk_bound_dev_if || exact_dif) {
243 bool dev_match = (sk->sk_bound_dev_if == dif ||
244 sk->sk_bound_dev_if == sdif);
245
246 if (exact_dif && !dev_match)
247 return -1;
248 if (sk->sk_bound_dev_if && dev_match)
249 score += 4;
250 }
251 if (sk->sk_incoming_cpu == raw_smp_processor_id())
252 score++;
253 }
254 return score;
255}
256
257/*
258 * Here are some nice properties to exploit here. The BSD API
259 * does not allow a listening sock to specify the remote port nor the
260 * remote address for the connection. So always assume those are both
261 * wildcarded during the search since they can never be otherwise.
262 */
263
264/* called with rcu_read_lock() : No refcount taken on the socket */
265static struct sock *inet_lhash2_lookup(struct net *net,
266 struct inet_listen_hashbucket *ilb2,
267 struct sk_buff *skb, int doff,
268 const __be32 saddr, __be16 sport,
269 const __be32 daddr, const unsigned short hnum,
270 const int dif, const int sdif)
271{
272 bool exact_dif = inet_exact_dif_match(net, skb);
273 struct inet_connection_sock *icsk;
274 struct sock *sk, *result = NULL;
275 int score, hiscore = 0;
276 u32 phash = 0;
277
278 inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
279 sk = (struct sock *)icsk;
280 score = compute_score(sk, net, hnum, daddr,
281 dif, sdif, exact_dif);
282 if (score > hiscore) {
283 if (sk->sk_reuseport) {
284 phash = inet_ehashfn(net, daddr, hnum,
285 saddr, sport);
286 result = reuseport_select_sock(sk, phash,
287 skb, doff);
288 if (result)
289 return result;
290 }
291 result = sk;
292 hiscore = score;
293 }
294 }
295
296 return result;
297}
298
299struct sock *__inet_lookup_listener(struct net *net,
300 struct inet_hashinfo *hashinfo,
301 struct sk_buff *skb, int doff,
302 const __be32 saddr, __be16 sport,
303 const __be32 daddr, const unsigned short hnum,
304 const int dif, const int sdif)
305{
306 unsigned int hash = inet_lhashfn(net, hnum);
307 struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
308 bool exact_dif = inet_exact_dif_match(net, skb);
309 struct inet_listen_hashbucket *ilb2;
310 struct sock *sk, *result = NULL;
311 int score, hiscore = 0;
312 unsigned int hash2;
313 u32 phash = 0;
314
315 if (ilb->count <= 10 || !hashinfo->lhash2)
316 goto port_lookup;
317
318 /* Too many sk in the ilb bucket (which is hashed by port alone).
319 * Try lhash2 (which is hashed by port and addr) instead.
320 */
321
322 hash2 = ipv4_portaddr_hash(net, daddr, hnum);
323 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
324 if (ilb2->count > ilb->count)
325 goto port_lookup;
326
327 result = inet_lhash2_lookup(net, ilb2, skb, doff,
328 saddr, sport, daddr, hnum,
329 dif, sdif);
330 if (result)
331 return result;
332
333 /* Lookup lhash2 with INADDR_ANY */
334
335 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
336 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
337 if (ilb2->count > ilb->count)
338 goto port_lookup;
339
340 return inet_lhash2_lookup(net, ilb2, skb, doff,
341 saddr, sport, daddr, hnum,
342 dif, sdif);
343
344port_lookup:
345 sk_for_each_rcu(sk, &ilb->head) {
346 score = compute_score(sk, net, hnum, daddr,
347 dif, sdif, exact_dif);
348 if (score > hiscore) {
349 if (sk->sk_reuseport) {
350 phash = inet_ehashfn(net, daddr, hnum,
351 saddr, sport);
352 result = reuseport_select_sock(sk, phash,
353 skb, doff);
354 if (result)
355 return result;
356 }
357 result = sk;
358 hiscore = score;
359 }
360 }
361 return result;
362}
363EXPORT_SYMBOL_GPL(__inet_lookup_listener);
364
365/* All sockets share common refcount, but have different destructors */
366void sock_gen_put(struct sock *sk)
367{
368 if (!refcount_dec_and_test(&sk->sk_refcnt))
369 return;
370
371 if (sk->sk_state == TCP_TIME_WAIT)
372 inet_twsk_free(inet_twsk(sk));
373 else if (sk->sk_state == TCP_NEW_SYN_RECV)
374 reqsk_free(inet_reqsk(sk));
375 else
376 sk_free(sk);
377}
378EXPORT_SYMBOL_GPL(sock_gen_put);
379
380void sock_edemux(struct sk_buff *skb)
381{
382 sock_gen_put(skb->sk);
383}
384EXPORT_SYMBOL(sock_edemux);
385
386struct sock *__inet_lookup_established(struct net *net,
387 struct inet_hashinfo *hashinfo,
388 const __be32 saddr, const __be16 sport,
389 const __be32 daddr, const u16 hnum,
390 const int dif, const int sdif)
391{
392 INET_ADDR_COOKIE(acookie, saddr, daddr);
393 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
394 struct sock *sk;
395 const struct hlist_nulls_node *node;
396 /* Optimize here for direct hit, only listening connections can
397 * have wildcards anyways.
398 */
399 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
400 unsigned int slot = hash & hashinfo->ehash_mask;
401 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
402
403begin:
404 sk_nulls_for_each_rcu(sk, node, &head->chain) {
405 if (sk->sk_hash != hash)
406 continue;
407 if (likely(INET_MATCH(sk, net, acookie,
408 saddr, daddr, ports, dif, sdif))) {
409 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
410 goto out;
411 if (unlikely(!INET_MATCH(sk, net, acookie,
412 saddr, daddr, ports,
413 dif, sdif))) {
414 sock_gen_put(sk);
415 goto begin;
416 }
417 goto found;
418 }
419 }
420 /*
421 * if the nulls value we got at the end of this lookup is
422 * not the expected one, we must restart lookup.
423 * We probably met an item that was moved to another chain.
424 */
425 if (get_nulls_value(node) != slot)
426 goto begin;
427out:
428 sk = NULL;
429found:
430 return sk;
431}
432EXPORT_SYMBOL_GPL(__inet_lookup_established);
433
434/* called with local bh disabled */
435static int __inet_check_established(struct inet_timewait_death_row *death_row,
436 struct sock *sk, __u16 lport,
437 struct inet_timewait_sock **twp)
438{
439 struct inet_hashinfo *hinfo = death_row->hashinfo;
440 struct inet_sock *inet = inet_sk(sk);
441 __be32 daddr = inet->inet_rcv_saddr;
442 __be32 saddr = inet->inet_daddr;
443 int dif = sk->sk_bound_dev_if;
444 struct net *net = sock_net(sk);
445 int sdif = l3mdev_master_ifindex_by_index(net, dif);
446 INET_ADDR_COOKIE(acookie, saddr, daddr);
447 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
448 unsigned int hash = inet_ehashfn(net, daddr, lport,
449 saddr, inet->inet_dport);
450 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
451 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
452 struct sock *sk2;
453 const struct hlist_nulls_node *node;
454 struct inet_timewait_sock *tw = NULL;
455
456 spin_lock(lock);
457
458 sk_nulls_for_each(sk2, node, &head->chain) {
459 if (sk2->sk_hash != hash)
460 continue;
461
462 if (likely(INET_MATCH(sk2, net, acookie,
463 saddr, daddr, ports, dif, sdif))) {
464 if (sk2->sk_state == TCP_TIME_WAIT) {
465 tw = inet_twsk(sk2);
466 if (twsk_unique(sk, sk2, twp))
467 break;
468 }
469 goto not_unique;
470 }
471 }
472
473 /* Must record num and sport now. Otherwise we will see
474 * in hash table socket with a funny identity.
475 */
476 inet->inet_num = lport;
477 inet->inet_sport = htons(lport);
478 sk->sk_hash = hash;
479 WARN_ON(!sk_unhashed(sk));
480 __sk_nulls_add_node_rcu(sk, &head->chain);
481 if (tw) {
482 sk_nulls_del_node_init_rcu((struct sock *)tw);
483 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
484 }
485 spin_unlock(lock);
486 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
487
488 if (twp) {
489 *twp = tw;
490 } else if (tw) {
491 /* Silly. Should hash-dance instead... */
492 inet_twsk_deschedule_put(tw);
493 }
494 return 0;
495
496not_unique:
497 spin_unlock(lock);
498 return -EADDRNOTAVAIL;
499}
500
501static u32 inet_sk_port_offset(const struct sock *sk)
502{
503 const struct inet_sock *inet = inet_sk(sk);
504
505 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
506 inet->inet_daddr,
507 inet->inet_dport);
508}
509
510/* insert a socket into ehash, and eventually remove another one
511 * (The another one can be a SYN_RECV or TIMEWAIT
512 */
513bool inet_ehash_insert(struct sock *sk, struct sock *osk)
514{
515 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
516 struct hlist_nulls_head *list;
517 struct inet_ehash_bucket *head;
518 spinlock_t *lock;
519 bool ret = true;
520
521 WARN_ON_ONCE(!sk_unhashed(sk));
522
523 sk->sk_hash = sk_ehashfn(sk);
524 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
525 list = &head->chain;
526 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
527
528 spin_lock(lock);
529 if (osk) {
530 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
531 ret = sk_nulls_del_node_init_rcu(osk);
532 }
533 if (ret)
534 __sk_nulls_add_node_rcu(sk, list);
535 spin_unlock(lock);
536 return ret;
537}
538
539bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
540{
541 bool ok = inet_ehash_insert(sk, osk);
542
543 if (ok) {
544 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
545 } else {
546 percpu_counter_inc(sk->sk_prot->orphan_count);
547 inet_sk_set_state(sk, TCP_CLOSE);
548 sock_set_flag(sk, SOCK_DEAD);
549 inet_csk_destroy_sock(sk);
550 }
551 return ok;
552}
553EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
554
555static int inet_reuseport_add_sock(struct sock *sk,
556 struct inet_listen_hashbucket *ilb)
557{
558 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
559 struct sock *sk2;
560 kuid_t uid = sock_i_uid(sk);
561
562 sk_for_each_rcu(sk2, &ilb->head) {
563 if (sk2 != sk &&
564 sk2->sk_family == sk->sk_family &&
565 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
566 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
567 inet_csk(sk2)->icsk_bind_hash == tb &&
568 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
569 inet_rcv_saddr_equal(sk, sk2, false))
570 return reuseport_add_sock(sk, sk2);
571 }
572
573 return reuseport_alloc(sk);
574}
575
576int __inet_hash(struct sock *sk, struct sock *osk)
577{
578 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
579 struct inet_listen_hashbucket *ilb;
580 int err = 0;
581
582 if (sk->sk_state != TCP_LISTEN) {
583 inet_ehash_nolisten(sk, osk);
584 return 0;
585 }
586 WARN_ON(!sk_unhashed(sk));
587 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
588
589 spin_lock(&ilb->lock);
590 if (sk->sk_reuseport) {
591 err = inet_reuseport_add_sock(sk, ilb);
592 if (err)
593 goto unlock;
594 }
595 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
596 sk->sk_family == AF_INET6)
597 hlist_add_tail_rcu(&sk->sk_node, &ilb->head);
598 else
599 hlist_add_head_rcu(&sk->sk_node, &ilb->head);
600 inet_hash2(hashinfo, sk);
601 ilb->count++;
602 sock_set_flag(sk, SOCK_RCU_FREE);
603 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
604unlock:
605 spin_unlock(&ilb->lock);
606
607 return err;
608}
609EXPORT_SYMBOL(__inet_hash);
610
611int inet_hash(struct sock *sk)
612{
613 int err = 0;
614
615 if (sk->sk_state != TCP_CLOSE) {
616 local_bh_disable();
617 err = __inet_hash(sk, NULL);
618 local_bh_enable();
619 }
620
621 return err;
622}
623EXPORT_SYMBOL_GPL(inet_hash);
624
625void inet_unhash(struct sock *sk)
626{
627 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
628 struct inet_listen_hashbucket *ilb = NULL;
629 spinlock_t *lock;
630
631 if (sk_unhashed(sk))
632 return;
633
634 if (sk->sk_state == TCP_LISTEN) {
635 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
636 lock = &ilb->lock;
637 } else {
638 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
639 }
640 spin_lock_bh(lock);
641 if (sk_unhashed(sk))
642 goto unlock;
643
644 if (rcu_access_pointer(sk->sk_reuseport_cb))
645 reuseport_detach_sock(sk);
646 if (ilb) {
647 inet_unhash2(hashinfo, sk);
648 __sk_del_node_init(sk);
649 ilb->count--;
650 } else {
651 __sk_nulls_del_node_init_rcu(sk);
652 }
653 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
654unlock:
655 spin_unlock_bh(lock);
656}
657EXPORT_SYMBOL_GPL(inet_unhash);
658
659int __inet_hash_connect(struct inet_timewait_death_row *death_row,
660 struct sock *sk, u32 port_offset,
661 int (*check_established)(struct inet_timewait_death_row *,
662 struct sock *, __u16, struct inet_timewait_sock **))
663{
664 struct inet_hashinfo *hinfo = death_row->hashinfo;
665 struct inet_timewait_sock *tw = NULL;
666 struct inet_bind_hashbucket *head;
667 int port = inet_sk(sk)->inet_num;
668 struct net *net = sock_net(sk);
669 struct inet_bind_bucket *tb;
670 u32 remaining, offset;
671 int ret, i, low, high;
672 static u32 hint;
673
674 if (port) {
675 head = &hinfo->bhash[inet_bhashfn(net, port,
676 hinfo->bhash_size)];
677 tb = inet_csk(sk)->icsk_bind_hash;
678 spin_lock_bh(&head->lock);
679 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
680 inet_ehash_nolisten(sk, NULL);
681 spin_unlock_bh(&head->lock);
682 return 0;
683 }
684 spin_unlock(&head->lock);
685 /* No definite answer... Walk to established hash table */
686 ret = check_established(death_row, sk, port, NULL);
687 local_bh_enable();
688 return ret;
689 }
690
691 inet_get_local_port_range(net, &low, &high);
692 high++; /* [32768, 60999] -> [32768, 61000[ */
693 remaining = high - low;
694 if (likely(remaining > 1))
695 remaining &= ~1U;
696
697 offset = (hint + port_offset) % remaining;
698 /* In first pass we try ports of @low parity.
699 * inet_csk_get_port() does the opposite choice.
700 */
701 offset &= ~1U;
702other_parity_scan:
703 port = low + offset;
704 for (i = 0; i < remaining; i += 2, port += 2) {
705 if (unlikely(port >= high))
706 port -= remaining;
707 if (inet_is_local_reserved_port(net, port))
708 continue;
709 head = &hinfo->bhash[inet_bhashfn(net, port,
710 hinfo->bhash_size)];
711 spin_lock_bh(&head->lock);
712
713 /* Does not bother with rcv_saddr checks, because
714 * the established check is already unique enough.
715 */
716 inet_bind_bucket_for_each(tb, &head->chain) {
717 if (net_eq(ib_net(tb), net) && tb->port == port) {
718 if (tb->fastreuse >= 0 ||
719 tb->fastreuseport >= 0)
720 goto next_port;
721 WARN_ON(hlist_empty(&tb->owners));
722 if (!check_established(death_row, sk,
723 port, &tw))
724 goto ok;
725 goto next_port;
726 }
727 }
728
729 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
730 net, head, port);
731 if (!tb) {
732 spin_unlock_bh(&head->lock);
733 return -ENOMEM;
734 }
735 tb->fastreuse = -1;
736 tb->fastreuseport = -1;
737 goto ok;
738next_port:
739 spin_unlock_bh(&head->lock);
740 cond_resched();
741 }
742
743 offset++;
744 if ((offset & 1) && remaining > 1)
745 goto other_parity_scan;
746
747 return -EADDRNOTAVAIL;
748
749ok:
750 hint += i + 2;
751
752 /* Head lock still held and bh's disabled */
753 inet_bind_hash(sk, tb, port);
754 if (sk_unhashed(sk)) {
755 inet_sk(sk)->inet_sport = htons(port);
756 inet_ehash_nolisten(sk, (struct sock *)tw);
757 }
758 if (tw)
759 inet_twsk_bind_unhash(tw, hinfo);
760 spin_unlock(&head->lock);
761 if (tw)
762 inet_twsk_deschedule_put(tw);
763 local_bh_enable();
764 return 0;
765}
766
767/*
768 * Bind a port for a connect operation and hash it.
769 */
770int inet_hash_connect(struct inet_timewait_death_row *death_row,
771 struct sock *sk)
772{
773 u32 port_offset = 0;
774
775 if (!inet_sk(sk)->inet_num)
776 port_offset = inet_sk_port_offset(sk);
777 return __inet_hash_connect(death_row, sk, port_offset,
778 __inet_check_established);
779}
780EXPORT_SYMBOL_GPL(inet_hash_connect);
781
782void inet_hashinfo_init(struct inet_hashinfo *h)
783{
784 int i;
785
786 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
787 spin_lock_init(&h->listening_hash[i].lock);
788 INIT_HLIST_HEAD(&h->listening_hash[i].head);
789 h->listening_hash[i].count = 0;
790 }
791
792 h->lhash2 = NULL;
793}
794EXPORT_SYMBOL_GPL(inet_hashinfo_init);
795
796void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
797 unsigned long numentries, int scale,
798 unsigned long low_limit,
799 unsigned long high_limit)
800{
801 unsigned int i;
802
803 h->lhash2 = alloc_large_system_hash(name,
804 sizeof(*h->lhash2),
805 numentries,
806 scale,
807 0,
808 NULL,
809 &h->lhash2_mask,
810 low_limit,
811 high_limit);
812
813 for (i = 0; i <= h->lhash2_mask; i++) {
814 spin_lock_init(&h->lhash2[i].lock);
815 INIT_HLIST_HEAD(&h->lhash2[i].head);
816 h->lhash2[i].count = 0;
817 }
818}
819
820int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
821{
822 unsigned int locksz = sizeof(spinlock_t);
823 unsigned int i, nblocks = 1;
824
825 if (locksz != 0) {
826 /* allocate 2 cache lines or at least one spinlock per cpu */
827 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
828 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
829
830 /* no more locks than number of hash buckets */
831 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
832
833 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
834 if (!hashinfo->ehash_locks)
835 return -ENOMEM;
836
837 for (i = 0; i < nblocks; i++)
838 spin_lock_init(&hashinfo->ehash_locks[i]);
839 }
840 hashinfo->ehash_locks_mask = nblocks - 1;
841 return 0;
842}
843EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Generic INET transport hashtables
8 *
9 * Authors: Lotsa people, from code originally in tcp
10 */
11
12#include <linux/module.h>
13#include <linux/random.h>
14#include <linux/sched.h>
15#include <linux/slab.h>
16#include <linux/wait.h>
17#include <linux/vmalloc.h>
18#include <linux/memblock.h>
19
20#include <net/addrconf.h>
21#include <net/inet_connection_sock.h>
22#include <net/inet_hashtables.h>
23#include <net/secure_seq.h>
24#include <net/ip.h>
25#include <net/tcp.h>
26#include <net/sock_reuseport.h>
27
28static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
29 const __u16 lport, const __be32 faddr,
30 const __be16 fport)
31{
32 static u32 inet_ehash_secret __read_mostly;
33
34 net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
35
36 return __inet_ehashfn(laddr, lport, faddr, fport,
37 inet_ehash_secret + net_hash_mix(net));
38}
39
40/* This function handles inet_sock, but also timewait and request sockets
41 * for IPv4/IPv6.
42 */
43static u32 sk_ehashfn(const struct sock *sk)
44{
45#if IS_ENABLED(CONFIG_IPV6)
46 if (sk->sk_family == AF_INET6 &&
47 !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
48 return inet6_ehashfn(sock_net(sk),
49 &sk->sk_v6_rcv_saddr, sk->sk_num,
50 &sk->sk_v6_daddr, sk->sk_dport);
51#endif
52 return inet_ehashfn(sock_net(sk),
53 sk->sk_rcv_saddr, sk->sk_num,
54 sk->sk_daddr, sk->sk_dport);
55}
56
57/*
58 * Allocate and initialize a new local port bind bucket.
59 * The bindhash mutex for snum's hash chain must be held here.
60 */
61struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
62 struct net *net,
63 struct inet_bind_hashbucket *head,
64 const unsigned short snum,
65 int l3mdev)
66{
67 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
68
69 if (tb) {
70 write_pnet(&tb->ib_net, net);
71 tb->l3mdev = l3mdev;
72 tb->port = snum;
73 tb->fastreuse = 0;
74 tb->fastreuseport = 0;
75 INIT_HLIST_HEAD(&tb->owners);
76 hlist_add_head(&tb->node, &head->chain);
77 }
78 return tb;
79}
80
81/*
82 * Caller must hold hashbucket lock for this tb with local BH disabled
83 */
84void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
85{
86 if (hlist_empty(&tb->owners)) {
87 __hlist_del(&tb->node);
88 kmem_cache_free(cachep, tb);
89 }
90}
91
92void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
93 const unsigned short snum)
94{
95 inet_sk(sk)->inet_num = snum;
96 sk_add_bind_node(sk, &tb->owners);
97 inet_csk(sk)->icsk_bind_hash = tb;
98}
99
100/*
101 * Get rid of any references to a local port held by the given sock.
102 */
103static void __inet_put_port(struct sock *sk)
104{
105 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
106 const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
107 hashinfo->bhash_size);
108 struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
109 struct inet_bind_bucket *tb;
110
111 spin_lock(&head->lock);
112 tb = inet_csk(sk)->icsk_bind_hash;
113 __sk_del_bind_node(sk);
114 inet_csk(sk)->icsk_bind_hash = NULL;
115 inet_sk(sk)->inet_num = 0;
116 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
117 spin_unlock(&head->lock);
118}
119
120void inet_put_port(struct sock *sk)
121{
122 local_bh_disable();
123 __inet_put_port(sk);
124 local_bh_enable();
125}
126EXPORT_SYMBOL(inet_put_port);
127
128int __inet_inherit_port(const struct sock *sk, struct sock *child)
129{
130 struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
131 unsigned short port = inet_sk(child)->inet_num;
132 const int bhash = inet_bhashfn(sock_net(sk), port,
133 table->bhash_size);
134 struct inet_bind_hashbucket *head = &table->bhash[bhash];
135 struct inet_bind_bucket *tb;
136 int l3mdev;
137
138 spin_lock(&head->lock);
139 tb = inet_csk(sk)->icsk_bind_hash;
140 if (unlikely(!tb)) {
141 spin_unlock(&head->lock);
142 return -ENOENT;
143 }
144 if (tb->port != port) {
145 l3mdev = inet_sk_bound_l3mdev(sk);
146
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. */
152 inet_bind_bucket_for_each(tb, &head->chain) {
153 if (net_eq(ib_net(tb), sock_net(sk)) &&
154 tb->l3mdev == l3mdev && tb->port == port)
155 break;
156 }
157 if (!tb) {
158 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
159 sock_net(sk), head, port,
160 l3mdev);
161 if (!tb) {
162 spin_unlock(&head->lock);
163 return -ENOMEM;
164 }
165 }
166 }
167 inet_bind_hash(child, tb, port);
168 spin_unlock(&head->lock);
169
170 return 0;
171}
172EXPORT_SYMBOL_GPL(__inet_inherit_port);
173
174static struct inet_listen_hashbucket *
175inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
176{
177 u32 hash;
178
179#if IS_ENABLED(CONFIG_IPV6)
180 if (sk->sk_family == AF_INET6)
181 hash = ipv6_portaddr_hash(sock_net(sk),
182 &sk->sk_v6_rcv_saddr,
183 inet_sk(sk)->inet_num);
184 else
185#endif
186 hash = ipv4_portaddr_hash(sock_net(sk),
187 inet_sk(sk)->inet_rcv_saddr,
188 inet_sk(sk)->inet_num);
189 return inet_lhash2_bucket(h, hash);
190}
191
192static void inet_hash2(struct inet_hashinfo *h, struct sock *sk)
193{
194 struct inet_listen_hashbucket *ilb2;
195
196 if (!h->lhash2)
197 return;
198
199 ilb2 = inet_lhash2_bucket_sk(h, sk);
200
201 spin_lock(&ilb2->lock);
202 if (sk->sk_reuseport && sk->sk_family == AF_INET6)
203 hlist_add_tail_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
204 &ilb2->head);
205 else
206 hlist_add_head_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
207 &ilb2->head);
208 ilb2->count++;
209 spin_unlock(&ilb2->lock);
210}
211
212static void inet_unhash2(struct inet_hashinfo *h, struct sock *sk)
213{
214 struct inet_listen_hashbucket *ilb2;
215
216 if (!h->lhash2 ||
217 WARN_ON_ONCE(hlist_unhashed(&inet_csk(sk)->icsk_listen_portaddr_node)))
218 return;
219
220 ilb2 = inet_lhash2_bucket_sk(h, sk);
221
222 spin_lock(&ilb2->lock);
223 hlist_del_init_rcu(&inet_csk(sk)->icsk_listen_portaddr_node);
224 ilb2->count--;
225 spin_unlock(&ilb2->lock);
226}
227
228static inline int compute_score(struct sock *sk, struct net *net,
229 const unsigned short hnum, const __be32 daddr,
230 const int dif, const int sdif, bool exact_dif)
231{
232 int score = -1;
233
234 if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
235 !ipv6_only_sock(sk)) {
236 if (sk->sk_rcv_saddr != daddr)
237 return -1;
238
239 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
240 return -1;
241
242 score = sk->sk_family == PF_INET ? 2 : 1;
243 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
244 score++;
245 }
246 return score;
247}
248
249/*
250 * Here are some nice properties to exploit here. The BSD API
251 * does not allow a listening sock to specify the remote port nor the
252 * remote address for the connection. So always assume those are both
253 * wildcarded during the search since they can never be otherwise.
254 */
255
256/* called with rcu_read_lock() : No refcount taken on the socket */
257static struct sock *inet_lhash2_lookup(struct net *net,
258 struct inet_listen_hashbucket *ilb2,
259 struct sk_buff *skb, int doff,
260 const __be32 saddr, __be16 sport,
261 const __be32 daddr, const unsigned short hnum,
262 const int dif, const int sdif)
263{
264 bool exact_dif = inet_exact_dif_match(net, skb);
265 struct inet_connection_sock *icsk;
266 struct sock *sk, *result = NULL;
267 int score, hiscore = 0;
268 u32 phash = 0;
269
270 inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
271 sk = (struct sock *)icsk;
272 score = compute_score(sk, net, hnum, daddr,
273 dif, sdif, exact_dif);
274 if (score > hiscore) {
275 if (sk->sk_reuseport) {
276 phash = inet_ehashfn(net, daddr, hnum,
277 saddr, sport);
278 result = reuseport_select_sock(sk, phash,
279 skb, doff);
280 if (result)
281 return result;
282 }
283 result = sk;
284 hiscore = score;
285 }
286 }
287
288 return result;
289}
290
291struct sock *__inet_lookup_listener(struct net *net,
292 struct inet_hashinfo *hashinfo,
293 struct sk_buff *skb, int doff,
294 const __be32 saddr, __be16 sport,
295 const __be32 daddr, const unsigned short hnum,
296 const int dif, const int sdif)
297{
298 struct inet_listen_hashbucket *ilb2;
299 struct sock *result = NULL;
300 unsigned int hash2;
301
302 hash2 = ipv4_portaddr_hash(net, daddr, hnum);
303 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
304
305 result = inet_lhash2_lookup(net, ilb2, skb, doff,
306 saddr, sport, daddr, hnum,
307 dif, sdif);
308 if (result)
309 goto done;
310
311 /* Lookup lhash2 with INADDR_ANY */
312 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
313 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
314
315 result = inet_lhash2_lookup(net, ilb2, skb, doff,
316 saddr, sport, htonl(INADDR_ANY), hnum,
317 dif, sdif);
318done:
319 if (IS_ERR(result))
320 return NULL;
321 return result;
322}
323EXPORT_SYMBOL_GPL(__inet_lookup_listener);
324
325/* All sockets share common refcount, but have different destructors */
326void sock_gen_put(struct sock *sk)
327{
328 if (!refcount_dec_and_test(&sk->sk_refcnt))
329 return;
330
331 if (sk->sk_state == TCP_TIME_WAIT)
332 inet_twsk_free(inet_twsk(sk));
333 else if (sk->sk_state == TCP_NEW_SYN_RECV)
334 reqsk_free(inet_reqsk(sk));
335 else
336 sk_free(sk);
337}
338EXPORT_SYMBOL_GPL(sock_gen_put);
339
340void sock_edemux(struct sk_buff *skb)
341{
342 sock_gen_put(skb->sk);
343}
344EXPORT_SYMBOL(sock_edemux);
345
346struct sock *__inet_lookup_established(struct net *net,
347 struct inet_hashinfo *hashinfo,
348 const __be32 saddr, const __be16 sport,
349 const __be32 daddr, const u16 hnum,
350 const int dif, const int sdif)
351{
352 INET_ADDR_COOKIE(acookie, saddr, daddr);
353 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
354 struct sock *sk;
355 const struct hlist_nulls_node *node;
356 /* Optimize here for direct hit, only listening connections can
357 * have wildcards anyways.
358 */
359 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
360 unsigned int slot = hash & hashinfo->ehash_mask;
361 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
362
363begin:
364 sk_nulls_for_each_rcu(sk, node, &head->chain) {
365 if (sk->sk_hash != hash)
366 continue;
367 if (likely(INET_MATCH(sk, net, acookie,
368 saddr, daddr, ports, dif, sdif))) {
369 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
370 goto out;
371 if (unlikely(!INET_MATCH(sk, net, acookie,
372 saddr, daddr, ports,
373 dif, sdif))) {
374 sock_gen_put(sk);
375 goto begin;
376 }
377 goto found;
378 }
379 }
380 /*
381 * if the nulls value we got at the end of this lookup is
382 * not the expected one, we must restart lookup.
383 * We probably met an item that was moved to another chain.
384 */
385 if (get_nulls_value(node) != slot)
386 goto begin;
387out:
388 sk = NULL;
389found:
390 return sk;
391}
392EXPORT_SYMBOL_GPL(__inet_lookup_established);
393
394/* called with local bh disabled */
395static int __inet_check_established(struct inet_timewait_death_row *death_row,
396 struct sock *sk, __u16 lport,
397 struct inet_timewait_sock **twp)
398{
399 struct inet_hashinfo *hinfo = death_row->hashinfo;
400 struct inet_sock *inet = inet_sk(sk);
401 __be32 daddr = inet->inet_rcv_saddr;
402 __be32 saddr = inet->inet_daddr;
403 int dif = sk->sk_bound_dev_if;
404 struct net *net = sock_net(sk);
405 int sdif = l3mdev_master_ifindex_by_index(net, dif);
406 INET_ADDR_COOKIE(acookie, saddr, daddr);
407 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
408 unsigned int hash = inet_ehashfn(net, daddr, lport,
409 saddr, inet->inet_dport);
410 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
411 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
412 struct sock *sk2;
413 const struct hlist_nulls_node *node;
414 struct inet_timewait_sock *tw = NULL;
415
416 spin_lock(lock);
417
418 sk_nulls_for_each(sk2, node, &head->chain) {
419 if (sk2->sk_hash != hash)
420 continue;
421
422 if (likely(INET_MATCH(sk2, net, acookie,
423 saddr, daddr, ports, dif, sdif))) {
424 if (sk2->sk_state == TCP_TIME_WAIT) {
425 tw = inet_twsk(sk2);
426 if (twsk_unique(sk, sk2, twp))
427 break;
428 }
429 goto not_unique;
430 }
431 }
432
433 /* Must record num and sport now. Otherwise we will see
434 * in hash table socket with a funny identity.
435 */
436 inet->inet_num = lport;
437 inet->inet_sport = htons(lport);
438 sk->sk_hash = hash;
439 WARN_ON(!sk_unhashed(sk));
440 __sk_nulls_add_node_rcu(sk, &head->chain);
441 if (tw) {
442 sk_nulls_del_node_init_rcu((struct sock *)tw);
443 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
444 }
445 spin_unlock(lock);
446 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
447
448 if (twp) {
449 *twp = tw;
450 } else if (tw) {
451 /* Silly. Should hash-dance instead... */
452 inet_twsk_deschedule_put(tw);
453 }
454 return 0;
455
456not_unique:
457 spin_unlock(lock);
458 return -EADDRNOTAVAIL;
459}
460
461static u32 inet_sk_port_offset(const struct sock *sk)
462{
463 const struct inet_sock *inet = inet_sk(sk);
464
465 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
466 inet->inet_daddr,
467 inet->inet_dport);
468}
469
470/* insert a socket into ehash, and eventually remove another one
471 * (The another one can be a SYN_RECV or TIMEWAIT
472 */
473bool inet_ehash_insert(struct sock *sk, struct sock *osk)
474{
475 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
476 struct hlist_nulls_head *list;
477 struct inet_ehash_bucket *head;
478 spinlock_t *lock;
479 bool ret = true;
480
481 WARN_ON_ONCE(!sk_unhashed(sk));
482
483 sk->sk_hash = sk_ehashfn(sk);
484 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
485 list = &head->chain;
486 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
487
488 spin_lock(lock);
489 if (osk) {
490 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
491 ret = sk_nulls_del_node_init_rcu(osk);
492 }
493 if (ret)
494 __sk_nulls_add_node_rcu(sk, list);
495 spin_unlock(lock);
496 return ret;
497}
498
499bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
500{
501 bool ok = inet_ehash_insert(sk, osk);
502
503 if (ok) {
504 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
505 } else {
506 percpu_counter_inc(sk->sk_prot->orphan_count);
507 inet_sk_set_state(sk, TCP_CLOSE);
508 sock_set_flag(sk, SOCK_DEAD);
509 inet_csk_destroy_sock(sk);
510 }
511 return ok;
512}
513EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
514
515static int inet_reuseport_add_sock(struct sock *sk,
516 struct inet_listen_hashbucket *ilb)
517{
518 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
519 struct sock *sk2;
520 kuid_t uid = sock_i_uid(sk);
521
522 sk_for_each_rcu(sk2, &ilb->head) {
523 if (sk2 != sk &&
524 sk2->sk_family == sk->sk_family &&
525 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
526 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
527 inet_csk(sk2)->icsk_bind_hash == tb &&
528 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
529 inet_rcv_saddr_equal(sk, sk2, false))
530 return reuseport_add_sock(sk, sk2,
531 inet_rcv_saddr_any(sk));
532 }
533
534 return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
535}
536
537int __inet_hash(struct sock *sk, struct sock *osk)
538{
539 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
540 struct inet_listen_hashbucket *ilb;
541 int err = 0;
542
543 if (sk->sk_state != TCP_LISTEN) {
544 inet_ehash_nolisten(sk, osk);
545 return 0;
546 }
547 WARN_ON(!sk_unhashed(sk));
548 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
549
550 spin_lock(&ilb->lock);
551 if (sk->sk_reuseport) {
552 err = inet_reuseport_add_sock(sk, ilb);
553 if (err)
554 goto unlock;
555 }
556 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
557 sk->sk_family == AF_INET6)
558 hlist_add_tail_rcu(&sk->sk_node, &ilb->head);
559 else
560 hlist_add_head_rcu(&sk->sk_node, &ilb->head);
561 inet_hash2(hashinfo, sk);
562 ilb->count++;
563 sock_set_flag(sk, SOCK_RCU_FREE);
564 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
565unlock:
566 spin_unlock(&ilb->lock);
567
568 return err;
569}
570EXPORT_SYMBOL(__inet_hash);
571
572int inet_hash(struct sock *sk)
573{
574 int err = 0;
575
576 if (sk->sk_state != TCP_CLOSE) {
577 local_bh_disable();
578 err = __inet_hash(sk, NULL);
579 local_bh_enable();
580 }
581
582 return err;
583}
584EXPORT_SYMBOL_GPL(inet_hash);
585
586void inet_unhash(struct sock *sk)
587{
588 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
589 struct inet_listen_hashbucket *ilb = NULL;
590 spinlock_t *lock;
591
592 if (sk_unhashed(sk))
593 return;
594
595 if (sk->sk_state == TCP_LISTEN) {
596 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
597 lock = &ilb->lock;
598 } else {
599 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
600 }
601 spin_lock_bh(lock);
602 if (sk_unhashed(sk))
603 goto unlock;
604
605 if (rcu_access_pointer(sk->sk_reuseport_cb))
606 reuseport_detach_sock(sk);
607 if (ilb) {
608 inet_unhash2(hashinfo, sk);
609 __sk_del_node_init(sk);
610 ilb->count--;
611 } else {
612 __sk_nulls_del_node_init_rcu(sk);
613 }
614 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
615unlock:
616 spin_unlock_bh(lock);
617}
618EXPORT_SYMBOL_GPL(inet_unhash);
619
620int __inet_hash_connect(struct inet_timewait_death_row *death_row,
621 struct sock *sk, u32 port_offset,
622 int (*check_established)(struct inet_timewait_death_row *,
623 struct sock *, __u16, struct inet_timewait_sock **))
624{
625 struct inet_hashinfo *hinfo = death_row->hashinfo;
626 struct inet_timewait_sock *tw = NULL;
627 struct inet_bind_hashbucket *head;
628 int port = inet_sk(sk)->inet_num;
629 struct net *net = sock_net(sk);
630 struct inet_bind_bucket *tb;
631 u32 remaining, offset;
632 int ret, i, low, high;
633 static u32 hint;
634 int l3mdev;
635
636 if (port) {
637 head = &hinfo->bhash[inet_bhashfn(net, port,
638 hinfo->bhash_size)];
639 tb = inet_csk(sk)->icsk_bind_hash;
640 spin_lock_bh(&head->lock);
641 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
642 inet_ehash_nolisten(sk, NULL);
643 spin_unlock_bh(&head->lock);
644 return 0;
645 }
646 spin_unlock(&head->lock);
647 /* No definite answer... Walk to established hash table */
648 ret = check_established(death_row, sk, port, NULL);
649 local_bh_enable();
650 return ret;
651 }
652
653 l3mdev = inet_sk_bound_l3mdev(sk);
654
655 inet_get_local_port_range(net, &low, &high);
656 high++; /* [32768, 60999] -> [32768, 61000[ */
657 remaining = high - low;
658 if (likely(remaining > 1))
659 remaining &= ~1U;
660
661 offset = (hint + port_offset) % remaining;
662 /* In first pass we try ports of @low parity.
663 * inet_csk_get_port() does the opposite choice.
664 */
665 offset &= ~1U;
666other_parity_scan:
667 port = low + offset;
668 for (i = 0; i < remaining; i += 2, port += 2) {
669 if (unlikely(port >= high))
670 port -= remaining;
671 if (inet_is_local_reserved_port(net, port))
672 continue;
673 head = &hinfo->bhash[inet_bhashfn(net, port,
674 hinfo->bhash_size)];
675 spin_lock_bh(&head->lock);
676
677 /* Does not bother with rcv_saddr checks, because
678 * the established check is already unique enough.
679 */
680 inet_bind_bucket_for_each(tb, &head->chain) {
681 if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
682 tb->port == port) {
683 if (tb->fastreuse >= 0 ||
684 tb->fastreuseport >= 0)
685 goto next_port;
686 WARN_ON(hlist_empty(&tb->owners));
687 if (!check_established(death_row, sk,
688 port, &tw))
689 goto ok;
690 goto next_port;
691 }
692 }
693
694 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
695 net, head, port, l3mdev);
696 if (!tb) {
697 spin_unlock_bh(&head->lock);
698 return -ENOMEM;
699 }
700 tb->fastreuse = -1;
701 tb->fastreuseport = -1;
702 goto ok;
703next_port:
704 spin_unlock_bh(&head->lock);
705 cond_resched();
706 }
707
708 offset++;
709 if ((offset & 1) && remaining > 1)
710 goto other_parity_scan;
711
712 return -EADDRNOTAVAIL;
713
714ok:
715 hint += i + 2;
716
717 /* Head lock still held and bh's disabled */
718 inet_bind_hash(sk, tb, port);
719 if (sk_unhashed(sk)) {
720 inet_sk(sk)->inet_sport = htons(port);
721 inet_ehash_nolisten(sk, (struct sock *)tw);
722 }
723 if (tw)
724 inet_twsk_bind_unhash(tw, hinfo);
725 spin_unlock(&head->lock);
726 if (tw)
727 inet_twsk_deschedule_put(tw);
728 local_bh_enable();
729 return 0;
730}
731
732/*
733 * Bind a port for a connect operation and hash it.
734 */
735int inet_hash_connect(struct inet_timewait_death_row *death_row,
736 struct sock *sk)
737{
738 u32 port_offset = 0;
739
740 if (!inet_sk(sk)->inet_num)
741 port_offset = inet_sk_port_offset(sk);
742 return __inet_hash_connect(death_row, sk, port_offset,
743 __inet_check_established);
744}
745EXPORT_SYMBOL_GPL(inet_hash_connect);
746
747void inet_hashinfo_init(struct inet_hashinfo *h)
748{
749 int i;
750
751 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
752 spin_lock_init(&h->listening_hash[i].lock);
753 INIT_HLIST_HEAD(&h->listening_hash[i].head);
754 h->listening_hash[i].count = 0;
755 }
756
757 h->lhash2 = NULL;
758}
759EXPORT_SYMBOL_GPL(inet_hashinfo_init);
760
761static void init_hashinfo_lhash2(struct inet_hashinfo *h)
762{
763 int i;
764
765 for (i = 0; i <= h->lhash2_mask; i++) {
766 spin_lock_init(&h->lhash2[i].lock);
767 INIT_HLIST_HEAD(&h->lhash2[i].head);
768 h->lhash2[i].count = 0;
769 }
770}
771
772void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
773 unsigned long numentries, int scale,
774 unsigned long low_limit,
775 unsigned long high_limit)
776{
777 h->lhash2 = alloc_large_system_hash(name,
778 sizeof(*h->lhash2),
779 numentries,
780 scale,
781 0,
782 NULL,
783 &h->lhash2_mask,
784 low_limit,
785 high_limit);
786 init_hashinfo_lhash2(h);
787}
788
789int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
790{
791 h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
792 if (!h->lhash2)
793 return -ENOMEM;
794
795 h->lhash2_mask = INET_LHTABLE_SIZE - 1;
796 /* INET_LHTABLE_SIZE must be a power of 2 */
797 BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
798
799 init_hashinfo_lhash2(h);
800 return 0;
801}
802EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
803
804int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
805{
806 unsigned int locksz = sizeof(spinlock_t);
807 unsigned int i, nblocks = 1;
808
809 if (locksz != 0) {
810 /* allocate 2 cache lines or at least one spinlock per cpu */
811 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
812 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
813
814 /* no more locks than number of hash buckets */
815 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
816
817 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
818 if (!hashinfo->ehash_locks)
819 return -ENOMEM;
820
821 for (i = 0; i < nblocks; i++)
822 spin_lock_init(&hashinfo->ehash_locks[i]);
823 }
824 hashinfo->ehash_locks_mask = nblocks - 1;
825 return 0;
826}
827EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);