Loading...
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#if IS_ENABLED(CONFIG_IPV6)
24#include <net/inet6_hashtables.h>
25#endif
26#include <net/secure_seq.h>
27#include <net/hotdata.h>
28#include <net/ip.h>
29#include <net/tcp.h>
30#include <net/sock_reuseport.h>
31
32u32 inet_ehashfn(const struct net *net, const __be32 laddr,
33 const __u16 lport, const __be32 faddr,
34 const __be16 fport)
35{
36 net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
37
38 return __inet_ehashfn(laddr, lport, faddr, fport,
39 inet_ehash_secret + net_hash_mix(net));
40}
41EXPORT_SYMBOL_GPL(inet_ehashfn);
42
43/* This function handles inet_sock, but also timewait and request sockets
44 * for IPv4/IPv6.
45 */
46static u32 sk_ehashfn(const struct sock *sk)
47{
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
55 return inet_ehashfn(sock_net(sk),
56 sk->sk_rcv_saddr, sk->sk_num,
57 sk->sk_daddr, sk->sk_dport);
58}
59
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 */
64struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
65 struct net *net,
66 struct inet_bind_hashbucket *head,
67 const unsigned short snum,
68 int l3mdev)
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->l3mdev = l3mdev;
75 tb->port = snum;
76 tb->fastreuse = 0;
77 tb->fastreuseport = 0;
78 INIT_HLIST_HEAD(&tb->bhash2);
79 hlist_add_head(&tb->node, &head->chain);
80 }
81 return tb;
82}
83
84/*
85 * Caller must hold hashbucket lock for this tb with local BH disabled
86 */
87void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
88{
89 if (hlist_empty(&tb->bhash2)) {
90 __hlist_del(&tb->node);
91 kmem_cache_free(cachep, tb);
92 }
93}
94
95bool inet_bind_bucket_match(const struct inet_bind_bucket *tb, const struct net *net,
96 unsigned short port, int l3mdev)
97{
98 return net_eq(ib_net(tb), net) && tb->port == port &&
99 tb->l3mdev == l3mdev;
100}
101
102static void inet_bind2_bucket_init(struct inet_bind2_bucket *tb2,
103 struct net *net,
104 struct inet_bind_hashbucket *head,
105 struct inet_bind_bucket *tb,
106 const struct sock *sk)
107{
108 write_pnet(&tb2->ib_net, net);
109 tb2->l3mdev = tb->l3mdev;
110 tb2->port = tb->port;
111#if IS_ENABLED(CONFIG_IPV6)
112 BUILD_BUG_ON(USHRT_MAX < (IPV6_ADDR_ANY | IPV6_ADDR_MAPPED));
113 if (sk->sk_family == AF_INET6) {
114 tb2->addr_type = ipv6_addr_type(&sk->sk_v6_rcv_saddr);
115 tb2->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
116 } else {
117 tb2->addr_type = IPV6_ADDR_MAPPED;
118 ipv6_addr_set_v4mapped(sk->sk_rcv_saddr, &tb2->v6_rcv_saddr);
119 }
120#else
121 tb2->rcv_saddr = sk->sk_rcv_saddr;
122#endif
123 INIT_HLIST_HEAD(&tb2->owners);
124 hlist_add_head(&tb2->node, &head->chain);
125 hlist_add_head(&tb2->bhash_node, &tb->bhash2);
126}
127
128struct inet_bind2_bucket *inet_bind2_bucket_create(struct kmem_cache *cachep,
129 struct net *net,
130 struct inet_bind_hashbucket *head,
131 struct inet_bind_bucket *tb,
132 const struct sock *sk)
133{
134 struct inet_bind2_bucket *tb2 = kmem_cache_alloc(cachep, GFP_ATOMIC);
135
136 if (tb2)
137 inet_bind2_bucket_init(tb2, net, head, tb, sk);
138
139 return tb2;
140}
141
142/* Caller must hold hashbucket lock for this tb with local BH disabled */
143void inet_bind2_bucket_destroy(struct kmem_cache *cachep, struct inet_bind2_bucket *tb)
144{
145 if (hlist_empty(&tb->owners)) {
146 __hlist_del(&tb->node);
147 __hlist_del(&tb->bhash_node);
148 kmem_cache_free(cachep, tb);
149 }
150}
151
152static bool inet_bind2_bucket_addr_match(const struct inet_bind2_bucket *tb2,
153 const struct sock *sk)
154{
155#if IS_ENABLED(CONFIG_IPV6)
156 if (sk->sk_family == AF_INET6)
157 return ipv6_addr_equal(&tb2->v6_rcv_saddr, &sk->sk_v6_rcv_saddr);
158
159 if (tb2->addr_type != IPV6_ADDR_MAPPED)
160 return false;
161#endif
162 return tb2->rcv_saddr == sk->sk_rcv_saddr;
163}
164
165void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
166 struct inet_bind2_bucket *tb2, unsigned short port)
167{
168 inet_sk(sk)->inet_num = port;
169 inet_csk(sk)->icsk_bind_hash = tb;
170 inet_csk(sk)->icsk_bind2_hash = tb2;
171 sk_add_bind_node(sk, &tb2->owners);
172}
173
174/*
175 * Get rid of any references to a local port held by the given sock.
176 */
177static void __inet_put_port(struct sock *sk)
178{
179 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
180 struct inet_bind_hashbucket *head, *head2;
181 struct net *net = sock_net(sk);
182 struct inet_bind_bucket *tb;
183 int bhash;
184
185 bhash = inet_bhashfn(net, inet_sk(sk)->inet_num, hashinfo->bhash_size);
186 head = &hashinfo->bhash[bhash];
187 head2 = inet_bhashfn_portaddr(hashinfo, sk, net, inet_sk(sk)->inet_num);
188
189 spin_lock(&head->lock);
190 tb = inet_csk(sk)->icsk_bind_hash;
191 inet_csk(sk)->icsk_bind_hash = NULL;
192 inet_sk(sk)->inet_num = 0;
193
194 spin_lock(&head2->lock);
195 if (inet_csk(sk)->icsk_bind2_hash) {
196 struct inet_bind2_bucket *tb2 = inet_csk(sk)->icsk_bind2_hash;
197
198 __sk_del_bind_node(sk);
199 inet_csk(sk)->icsk_bind2_hash = NULL;
200 inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
201 }
202 spin_unlock(&head2->lock);
203
204 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
205 spin_unlock(&head->lock);
206}
207
208void inet_put_port(struct sock *sk)
209{
210 local_bh_disable();
211 __inet_put_port(sk);
212 local_bh_enable();
213}
214EXPORT_SYMBOL(inet_put_port);
215
216int __inet_inherit_port(const struct sock *sk, struct sock *child)
217{
218 struct inet_hashinfo *table = tcp_or_dccp_get_hashinfo(sk);
219 unsigned short port = inet_sk(child)->inet_num;
220 struct inet_bind_hashbucket *head, *head2;
221 bool created_inet_bind_bucket = false;
222 struct net *net = sock_net(sk);
223 bool update_fastreuse = false;
224 struct inet_bind2_bucket *tb2;
225 struct inet_bind_bucket *tb;
226 int bhash, l3mdev;
227
228 bhash = inet_bhashfn(net, port, table->bhash_size);
229 head = &table->bhash[bhash];
230 head2 = inet_bhashfn_portaddr(table, child, net, port);
231
232 spin_lock(&head->lock);
233 spin_lock(&head2->lock);
234 tb = inet_csk(sk)->icsk_bind_hash;
235 tb2 = inet_csk(sk)->icsk_bind2_hash;
236 if (unlikely(!tb || !tb2)) {
237 spin_unlock(&head2->lock);
238 spin_unlock(&head->lock);
239 return -ENOENT;
240 }
241 if (tb->port != port) {
242 l3mdev = inet_sk_bound_l3mdev(sk);
243
244 /* NOTE: using tproxy and redirecting skbs to a proxy
245 * on a different listener port breaks the assumption
246 * that the listener socket's icsk_bind_hash is the same
247 * as that of the child socket. We have to look up or
248 * create a new bind bucket for the child here. */
249 inet_bind_bucket_for_each(tb, &head->chain) {
250 if (inet_bind_bucket_match(tb, net, port, l3mdev))
251 break;
252 }
253 if (!tb) {
254 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
255 net, head, port, l3mdev);
256 if (!tb) {
257 spin_unlock(&head2->lock);
258 spin_unlock(&head->lock);
259 return -ENOMEM;
260 }
261 created_inet_bind_bucket = true;
262 }
263 update_fastreuse = true;
264
265 goto bhash2_find;
266 } else if (!inet_bind2_bucket_addr_match(tb2, child)) {
267 l3mdev = inet_sk_bound_l3mdev(sk);
268
269bhash2_find:
270 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, child);
271 if (!tb2) {
272 tb2 = inet_bind2_bucket_create(table->bind2_bucket_cachep,
273 net, head2, tb, child);
274 if (!tb2)
275 goto error;
276 }
277 }
278 if (update_fastreuse)
279 inet_csk_update_fastreuse(tb, child);
280 inet_bind_hash(child, tb, tb2, port);
281 spin_unlock(&head2->lock);
282 spin_unlock(&head->lock);
283
284 return 0;
285
286error:
287 if (created_inet_bind_bucket)
288 inet_bind_bucket_destroy(table->bind_bucket_cachep, tb);
289 spin_unlock(&head2->lock);
290 spin_unlock(&head->lock);
291 return -ENOMEM;
292}
293EXPORT_SYMBOL_GPL(__inet_inherit_port);
294
295static struct inet_listen_hashbucket *
296inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
297{
298 u32 hash;
299
300#if IS_ENABLED(CONFIG_IPV6)
301 if (sk->sk_family == AF_INET6)
302 hash = ipv6_portaddr_hash(sock_net(sk),
303 &sk->sk_v6_rcv_saddr,
304 inet_sk(sk)->inet_num);
305 else
306#endif
307 hash = ipv4_portaddr_hash(sock_net(sk),
308 inet_sk(sk)->inet_rcv_saddr,
309 inet_sk(sk)->inet_num);
310 return inet_lhash2_bucket(h, hash);
311}
312
313static inline int compute_score(struct sock *sk, const struct net *net,
314 const unsigned short hnum, const __be32 daddr,
315 const int dif, const int sdif)
316{
317 int score = -1;
318
319 if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
320 !ipv6_only_sock(sk)) {
321 if (sk->sk_rcv_saddr != daddr)
322 return -1;
323
324 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
325 return -1;
326 score = sk->sk_bound_dev_if ? 2 : 1;
327
328 if (sk->sk_family == PF_INET)
329 score++;
330 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
331 score++;
332 }
333 return score;
334}
335
336/**
337 * inet_lookup_reuseport() - execute reuseport logic on AF_INET socket if necessary.
338 * @net: network namespace.
339 * @sk: AF_INET socket, must be in TCP_LISTEN state for TCP or TCP_CLOSE for UDP.
340 * @skb: context for a potential SK_REUSEPORT program.
341 * @doff: header offset.
342 * @saddr: source address.
343 * @sport: source port.
344 * @daddr: destination address.
345 * @hnum: destination port in host byte order.
346 * @ehashfn: hash function used to generate the fallback hash.
347 *
348 * Return: NULL if sk doesn't have SO_REUSEPORT set, otherwise a pointer to
349 * the selected sock or an error.
350 */
351struct sock *inet_lookup_reuseport(const struct net *net, struct sock *sk,
352 struct sk_buff *skb, int doff,
353 __be32 saddr, __be16 sport,
354 __be32 daddr, unsigned short hnum,
355 inet_ehashfn_t *ehashfn)
356{
357 struct sock *reuse_sk = NULL;
358 u32 phash;
359
360 if (sk->sk_reuseport) {
361 phash = INDIRECT_CALL_2(ehashfn, udp_ehashfn, inet_ehashfn,
362 net, daddr, hnum, saddr, sport);
363 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
364 }
365 return reuse_sk;
366}
367EXPORT_SYMBOL_GPL(inet_lookup_reuseport);
368
369/*
370 * Here are some nice properties to exploit here. The BSD API
371 * does not allow a listening sock to specify the remote port nor the
372 * remote address for the connection. So always assume those are both
373 * wildcarded during the search since they can never be otherwise.
374 */
375
376/* called with rcu_read_lock() : No refcount taken on the socket */
377static struct sock *inet_lhash2_lookup(const struct net *net,
378 struct inet_listen_hashbucket *ilb2,
379 struct sk_buff *skb, int doff,
380 const __be32 saddr, __be16 sport,
381 const __be32 daddr, const unsigned short hnum,
382 const int dif, const int sdif)
383{
384 struct sock *sk, *result = NULL;
385 struct hlist_nulls_node *node;
386 int score, hiscore = 0;
387
388 sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
389 score = compute_score(sk, net, hnum, daddr, dif, sdif);
390 if (score > hiscore) {
391 result = inet_lookup_reuseport(net, sk, skb, doff,
392 saddr, sport, daddr, hnum, inet_ehashfn);
393 if (result)
394 return result;
395
396 result = sk;
397 hiscore = score;
398 }
399 }
400
401 return result;
402}
403
404struct sock *inet_lookup_run_sk_lookup(const struct net *net,
405 int protocol,
406 struct sk_buff *skb, int doff,
407 __be32 saddr, __be16 sport,
408 __be32 daddr, u16 hnum, const int dif,
409 inet_ehashfn_t *ehashfn)
410{
411 struct sock *sk, *reuse_sk;
412 bool no_reuseport;
413
414 no_reuseport = bpf_sk_lookup_run_v4(net, protocol, saddr, sport,
415 daddr, hnum, dif, &sk);
416 if (no_reuseport || IS_ERR_OR_NULL(sk))
417 return sk;
418
419 reuse_sk = inet_lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum,
420 ehashfn);
421 if (reuse_sk)
422 sk = reuse_sk;
423 return sk;
424}
425
426struct sock *__inet_lookup_listener(const struct net *net,
427 struct inet_hashinfo *hashinfo,
428 struct sk_buff *skb, int doff,
429 const __be32 saddr, __be16 sport,
430 const __be32 daddr, const unsigned short hnum,
431 const int dif, const int sdif)
432{
433 struct inet_listen_hashbucket *ilb2;
434 struct sock *result = NULL;
435 unsigned int hash2;
436
437 /* Lookup redirect from BPF */
438 if (static_branch_unlikely(&bpf_sk_lookup_enabled) &&
439 hashinfo == net->ipv4.tcp_death_row.hashinfo) {
440 result = inet_lookup_run_sk_lookup(net, IPPROTO_TCP, skb, doff,
441 saddr, sport, daddr, hnum, dif,
442 inet_ehashfn);
443 if (result)
444 goto done;
445 }
446
447 hash2 = ipv4_portaddr_hash(net, daddr, hnum);
448 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
449
450 result = inet_lhash2_lookup(net, ilb2, skb, doff,
451 saddr, sport, daddr, hnum,
452 dif, sdif);
453 if (result)
454 goto done;
455
456 /* Lookup lhash2 with INADDR_ANY */
457 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
458 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
459
460 result = inet_lhash2_lookup(net, ilb2, skb, doff,
461 saddr, sport, htonl(INADDR_ANY), hnum,
462 dif, sdif);
463done:
464 if (IS_ERR(result))
465 return NULL;
466 return result;
467}
468EXPORT_SYMBOL_GPL(__inet_lookup_listener);
469
470/* All sockets share common refcount, but have different destructors */
471void sock_gen_put(struct sock *sk)
472{
473 if (!refcount_dec_and_test(&sk->sk_refcnt))
474 return;
475
476 if (sk->sk_state == TCP_TIME_WAIT)
477 inet_twsk_free(inet_twsk(sk));
478 else if (sk->sk_state == TCP_NEW_SYN_RECV)
479 reqsk_free(inet_reqsk(sk));
480 else
481 sk_free(sk);
482}
483EXPORT_SYMBOL_GPL(sock_gen_put);
484
485void sock_edemux(struct sk_buff *skb)
486{
487 sock_gen_put(skb->sk);
488}
489EXPORT_SYMBOL(sock_edemux);
490
491struct sock *__inet_lookup_established(const struct net *net,
492 struct inet_hashinfo *hashinfo,
493 const __be32 saddr, const __be16 sport,
494 const __be32 daddr, const u16 hnum,
495 const int dif, const int sdif)
496{
497 INET_ADDR_COOKIE(acookie, saddr, daddr);
498 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
499 struct sock *sk;
500 const struct hlist_nulls_node *node;
501 /* Optimize here for direct hit, only listening connections can
502 * have wildcards anyways.
503 */
504 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
505 unsigned int slot = hash & hashinfo->ehash_mask;
506 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
507
508begin:
509 sk_nulls_for_each_rcu(sk, node, &head->chain) {
510 if (sk->sk_hash != hash)
511 continue;
512 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
513 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
514 goto out;
515 if (unlikely(!inet_match(net, sk, acookie,
516 ports, dif, sdif))) {
517 sock_gen_put(sk);
518 goto begin;
519 }
520 goto found;
521 }
522 }
523 /*
524 * if the nulls value we got at the end of this lookup is
525 * not the expected one, we must restart lookup.
526 * We probably met an item that was moved to another chain.
527 */
528 if (get_nulls_value(node) != slot)
529 goto begin;
530out:
531 sk = NULL;
532found:
533 return sk;
534}
535EXPORT_SYMBOL_GPL(__inet_lookup_established);
536
537/* called with local bh disabled */
538static int __inet_check_established(struct inet_timewait_death_row *death_row,
539 struct sock *sk, __u16 lport,
540 struct inet_timewait_sock **twp)
541{
542 struct inet_hashinfo *hinfo = death_row->hashinfo;
543 struct inet_sock *inet = inet_sk(sk);
544 __be32 daddr = inet->inet_rcv_saddr;
545 __be32 saddr = inet->inet_daddr;
546 int dif = sk->sk_bound_dev_if;
547 struct net *net = sock_net(sk);
548 int sdif = l3mdev_master_ifindex_by_index(net, dif);
549 INET_ADDR_COOKIE(acookie, saddr, daddr);
550 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
551 unsigned int hash = inet_ehashfn(net, daddr, lport,
552 saddr, inet->inet_dport);
553 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
554 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
555 struct sock *sk2;
556 const struct hlist_nulls_node *node;
557 struct inet_timewait_sock *tw = NULL;
558
559 spin_lock(lock);
560
561 sk_nulls_for_each(sk2, node, &head->chain) {
562 if (sk2->sk_hash != hash)
563 continue;
564
565 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
566 if (sk2->sk_state == TCP_TIME_WAIT) {
567 tw = inet_twsk(sk2);
568 if (sk->sk_protocol == IPPROTO_TCP &&
569 tcp_twsk_unique(sk, sk2, twp))
570 break;
571 }
572 goto not_unique;
573 }
574 }
575
576 /* Must record num and sport now. Otherwise we will see
577 * in hash table socket with a funny identity.
578 */
579 inet->inet_num = lport;
580 inet->inet_sport = htons(lport);
581 sk->sk_hash = hash;
582 WARN_ON(!sk_unhashed(sk));
583 __sk_nulls_add_node_rcu(sk, &head->chain);
584 if (tw) {
585 sk_nulls_del_node_init_rcu((struct sock *)tw);
586 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
587 }
588 spin_unlock(lock);
589 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
590
591 if (twp) {
592 *twp = tw;
593 } else if (tw) {
594 /* Silly. Should hash-dance instead... */
595 inet_twsk_deschedule_put(tw);
596 }
597 return 0;
598
599not_unique:
600 spin_unlock(lock);
601 return -EADDRNOTAVAIL;
602}
603
604static u64 inet_sk_port_offset(const struct sock *sk)
605{
606 const struct inet_sock *inet = inet_sk(sk);
607
608 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
609 inet->inet_daddr,
610 inet->inet_dport);
611}
612
613/* Searches for an exsiting socket in the ehash bucket list.
614 * Returns true if found, false otherwise.
615 */
616static bool inet_ehash_lookup_by_sk(struct sock *sk,
617 struct hlist_nulls_head *list)
618{
619 const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
620 const int sdif = sk->sk_bound_dev_if;
621 const int dif = sk->sk_bound_dev_if;
622 const struct hlist_nulls_node *node;
623 struct net *net = sock_net(sk);
624 struct sock *esk;
625
626 INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
627
628 sk_nulls_for_each_rcu(esk, node, list) {
629 if (esk->sk_hash != sk->sk_hash)
630 continue;
631 if (sk->sk_family == AF_INET) {
632 if (unlikely(inet_match(net, esk, acookie,
633 ports, dif, sdif))) {
634 return true;
635 }
636 }
637#if IS_ENABLED(CONFIG_IPV6)
638 else if (sk->sk_family == AF_INET6) {
639 if (unlikely(inet6_match(net, esk,
640 &sk->sk_v6_daddr,
641 &sk->sk_v6_rcv_saddr,
642 ports, dif, sdif))) {
643 return true;
644 }
645 }
646#endif
647 }
648 return false;
649}
650
651/* Insert a socket into ehash, and eventually remove another one
652 * (The another one can be a SYN_RECV or TIMEWAIT)
653 * If an existing socket already exists, socket sk is not inserted,
654 * and sets found_dup_sk parameter to true.
655 */
656bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
657{
658 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
659 struct inet_ehash_bucket *head;
660 struct hlist_nulls_head *list;
661 spinlock_t *lock;
662 bool ret = true;
663
664 WARN_ON_ONCE(!sk_unhashed(sk));
665
666 sk->sk_hash = sk_ehashfn(sk);
667 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
668 list = &head->chain;
669 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
670
671 spin_lock(lock);
672 if (osk) {
673 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
674 ret = sk_nulls_del_node_init_rcu(osk);
675 } else if (found_dup_sk) {
676 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
677 if (*found_dup_sk)
678 ret = false;
679 }
680
681 if (ret)
682 __sk_nulls_add_node_rcu(sk, list);
683
684 spin_unlock(lock);
685
686 return ret;
687}
688
689bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
690{
691 bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
692
693 if (ok) {
694 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
695 } else {
696 this_cpu_inc(*sk->sk_prot->orphan_count);
697 inet_sk_set_state(sk, TCP_CLOSE);
698 sock_set_flag(sk, SOCK_DEAD);
699 inet_csk_destroy_sock(sk);
700 }
701 return ok;
702}
703EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
704
705static int inet_reuseport_add_sock(struct sock *sk,
706 struct inet_listen_hashbucket *ilb)
707{
708 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
709 const struct hlist_nulls_node *node;
710 struct sock *sk2;
711 kuid_t uid = sock_i_uid(sk);
712
713 sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
714 if (sk2 != sk &&
715 sk2->sk_family == sk->sk_family &&
716 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
717 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
718 inet_csk(sk2)->icsk_bind_hash == tb &&
719 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
720 inet_rcv_saddr_equal(sk, sk2, false))
721 return reuseport_add_sock(sk, sk2,
722 inet_rcv_saddr_any(sk));
723 }
724
725 return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
726}
727
728int __inet_hash(struct sock *sk, struct sock *osk)
729{
730 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
731 struct inet_listen_hashbucket *ilb2;
732 int err = 0;
733
734 if (sk->sk_state != TCP_LISTEN) {
735 local_bh_disable();
736 inet_ehash_nolisten(sk, osk, NULL);
737 local_bh_enable();
738 return 0;
739 }
740 WARN_ON(!sk_unhashed(sk));
741 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
742
743 spin_lock(&ilb2->lock);
744 if (sk->sk_reuseport) {
745 err = inet_reuseport_add_sock(sk, ilb2);
746 if (err)
747 goto unlock;
748 }
749 sock_set_flag(sk, SOCK_RCU_FREE);
750 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
751 sk->sk_family == AF_INET6)
752 __sk_nulls_add_node_tail_rcu(sk, &ilb2->nulls_head);
753 else
754 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
755 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
756unlock:
757 spin_unlock(&ilb2->lock);
758
759 return err;
760}
761EXPORT_SYMBOL(__inet_hash);
762
763int inet_hash(struct sock *sk)
764{
765 int err = 0;
766
767 if (sk->sk_state != TCP_CLOSE)
768 err = __inet_hash(sk, NULL);
769
770 return err;
771}
772EXPORT_SYMBOL_GPL(inet_hash);
773
774void inet_unhash(struct sock *sk)
775{
776 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
777
778 if (sk_unhashed(sk))
779 return;
780
781 if (sk->sk_state == TCP_LISTEN) {
782 struct inet_listen_hashbucket *ilb2;
783
784 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
785 /* Don't disable bottom halves while acquiring the lock to
786 * avoid circular locking dependency on PREEMPT_RT.
787 */
788 spin_lock(&ilb2->lock);
789 if (sk_unhashed(sk)) {
790 spin_unlock(&ilb2->lock);
791 return;
792 }
793
794 if (rcu_access_pointer(sk->sk_reuseport_cb))
795 reuseport_stop_listen_sock(sk);
796
797 __sk_nulls_del_node_init_rcu(sk);
798 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
799 spin_unlock(&ilb2->lock);
800 } else {
801 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
802
803 spin_lock_bh(lock);
804 if (sk_unhashed(sk)) {
805 spin_unlock_bh(lock);
806 return;
807 }
808 __sk_nulls_del_node_init_rcu(sk);
809 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
810 spin_unlock_bh(lock);
811 }
812}
813EXPORT_SYMBOL_GPL(inet_unhash);
814
815static bool inet_bind2_bucket_match(const struct inet_bind2_bucket *tb,
816 const struct net *net, unsigned short port,
817 int l3mdev, const struct sock *sk)
818{
819 if (!net_eq(ib2_net(tb), net) || tb->port != port ||
820 tb->l3mdev != l3mdev)
821 return false;
822
823 return inet_bind2_bucket_addr_match(tb, sk);
824}
825
826bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const struct net *net,
827 unsigned short port, int l3mdev, const struct sock *sk)
828{
829 if (!net_eq(ib2_net(tb), net) || tb->port != port ||
830 tb->l3mdev != l3mdev)
831 return false;
832
833#if IS_ENABLED(CONFIG_IPV6)
834 if (tb->addr_type == IPV6_ADDR_ANY)
835 return true;
836
837 if (tb->addr_type != IPV6_ADDR_MAPPED)
838 return false;
839
840 if (sk->sk_family == AF_INET6 &&
841 !ipv6_addr_v4mapped(&sk->sk_v6_rcv_saddr))
842 return false;
843#endif
844 return tb->rcv_saddr == 0;
845}
846
847/* The socket's bhash2 hashbucket spinlock must be held when this is called */
848struct inet_bind2_bucket *
849inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct net *net,
850 unsigned short port, int l3mdev, const struct sock *sk)
851{
852 struct inet_bind2_bucket *bhash2 = NULL;
853
854 inet_bind_bucket_for_each(bhash2, &head->chain)
855 if (inet_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
856 break;
857
858 return bhash2;
859}
860
861struct inet_bind_hashbucket *
862inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port)
863{
864 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
865 u32 hash;
866
867#if IS_ENABLED(CONFIG_IPV6)
868 if (sk->sk_family == AF_INET6)
869 hash = ipv6_portaddr_hash(net, &in6addr_any, port);
870 else
871#endif
872 hash = ipv4_portaddr_hash(net, 0, port);
873
874 return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
875}
876
877static void inet_update_saddr(struct sock *sk, void *saddr, int family)
878{
879 if (family == AF_INET) {
880 inet_sk(sk)->inet_saddr = *(__be32 *)saddr;
881 sk_rcv_saddr_set(sk, inet_sk(sk)->inet_saddr);
882 }
883#if IS_ENABLED(CONFIG_IPV6)
884 else {
885 sk->sk_v6_rcv_saddr = *(struct in6_addr *)saddr;
886 }
887#endif
888}
889
890static int __inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family, bool reset)
891{
892 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
893 struct inet_bind_hashbucket *head, *head2;
894 struct inet_bind2_bucket *tb2, *new_tb2;
895 int l3mdev = inet_sk_bound_l3mdev(sk);
896 int port = inet_sk(sk)->inet_num;
897 struct net *net = sock_net(sk);
898 int bhash;
899
900 if (!inet_csk(sk)->icsk_bind2_hash) {
901 /* Not bind()ed before. */
902 if (reset)
903 inet_reset_saddr(sk);
904 else
905 inet_update_saddr(sk, saddr, family);
906
907 return 0;
908 }
909
910 /* Allocate a bind2 bucket ahead of time to avoid permanently putting
911 * the bhash2 table in an inconsistent state if a new tb2 bucket
912 * allocation fails.
913 */
914 new_tb2 = kmem_cache_alloc(hinfo->bind2_bucket_cachep, GFP_ATOMIC);
915 if (!new_tb2) {
916 if (reset) {
917 /* The (INADDR_ANY, port) bucket might have already
918 * been freed, then we cannot fixup icsk_bind2_hash,
919 * so we give up and unlink sk from bhash/bhash2 not
920 * to leave inconsistency in bhash2.
921 */
922 inet_put_port(sk);
923 inet_reset_saddr(sk);
924 }
925
926 return -ENOMEM;
927 }
928
929 bhash = inet_bhashfn(net, port, hinfo->bhash_size);
930 head = &hinfo->bhash[bhash];
931 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
932
933 /* If we change saddr locklessly, another thread
934 * iterating over bhash might see corrupted address.
935 */
936 spin_lock_bh(&head->lock);
937
938 spin_lock(&head2->lock);
939 __sk_del_bind_node(sk);
940 inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, inet_csk(sk)->icsk_bind2_hash);
941 spin_unlock(&head2->lock);
942
943 if (reset)
944 inet_reset_saddr(sk);
945 else
946 inet_update_saddr(sk, saddr, family);
947
948 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
949
950 spin_lock(&head2->lock);
951 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
952 if (!tb2) {
953 tb2 = new_tb2;
954 inet_bind2_bucket_init(tb2, net, head2, inet_csk(sk)->icsk_bind_hash, sk);
955 }
956 inet_csk(sk)->icsk_bind2_hash = tb2;
957 sk_add_bind_node(sk, &tb2->owners);
958 spin_unlock(&head2->lock);
959
960 spin_unlock_bh(&head->lock);
961
962 if (tb2 != new_tb2)
963 kmem_cache_free(hinfo->bind2_bucket_cachep, new_tb2);
964
965 return 0;
966}
967
968int inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family)
969{
970 return __inet_bhash2_update_saddr(sk, saddr, family, false);
971}
972EXPORT_SYMBOL_GPL(inet_bhash2_update_saddr);
973
974void inet_bhash2_reset_saddr(struct sock *sk)
975{
976 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
977 __inet_bhash2_update_saddr(sk, NULL, 0, true);
978}
979EXPORT_SYMBOL_GPL(inet_bhash2_reset_saddr);
980
981/* RFC 6056 3.3.4. Algorithm 4: Double-Hash Port Selection Algorithm
982 * Note that we use 32bit integers (vs RFC 'short integers')
983 * because 2^16 is not a multiple of num_ephemeral and this
984 * property might be used by clever attacker.
985 *
986 * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
987 * attacks were since demonstrated, thus we use 65536 by default instead
988 * to really give more isolation and privacy, at the expense of 256kB
989 * of kernel memory.
990 */
991#define INET_TABLE_PERTURB_SIZE (1 << CONFIG_INET_TABLE_PERTURB_ORDER)
992static u32 *table_perturb;
993
994int __inet_hash_connect(struct inet_timewait_death_row *death_row,
995 struct sock *sk, u64 port_offset,
996 int (*check_established)(struct inet_timewait_death_row *,
997 struct sock *, __u16, struct inet_timewait_sock **))
998{
999 struct inet_hashinfo *hinfo = death_row->hashinfo;
1000 struct inet_bind_hashbucket *head, *head2;
1001 struct inet_timewait_sock *tw = NULL;
1002 int port = inet_sk(sk)->inet_num;
1003 struct net *net = sock_net(sk);
1004 struct inet_bind2_bucket *tb2;
1005 struct inet_bind_bucket *tb;
1006 bool tb_created = false;
1007 u32 remaining, offset;
1008 int ret, i, low, high;
1009 bool local_ports;
1010 int step, l3mdev;
1011 u32 index;
1012
1013 if (port) {
1014 local_bh_disable();
1015 ret = check_established(death_row, sk, port, NULL);
1016 local_bh_enable();
1017 return ret;
1018 }
1019
1020 l3mdev = inet_sk_bound_l3mdev(sk);
1021
1022 local_ports = inet_sk_get_local_port_range(sk, &low, &high);
1023 step = local_ports ? 1 : 2;
1024
1025 high++; /* [32768, 60999] -> [32768, 61000[ */
1026 remaining = high - low;
1027 if (!local_ports && remaining > 1)
1028 remaining &= ~1U;
1029
1030 get_random_sleepable_once(table_perturb,
1031 INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
1032 index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
1033
1034 offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
1035 offset %= remaining;
1036
1037 /* In first pass we try ports of @low parity.
1038 * inet_csk_get_port() does the opposite choice.
1039 */
1040 if (!local_ports)
1041 offset &= ~1U;
1042other_parity_scan:
1043 port = low + offset;
1044 for (i = 0; i < remaining; i += step, port += step) {
1045 if (unlikely(port >= high))
1046 port -= remaining;
1047 if (inet_is_local_reserved_port(net, port))
1048 continue;
1049 head = &hinfo->bhash[inet_bhashfn(net, port,
1050 hinfo->bhash_size)];
1051 spin_lock_bh(&head->lock);
1052
1053 /* Does not bother with rcv_saddr checks, because
1054 * the established check is already unique enough.
1055 */
1056 inet_bind_bucket_for_each(tb, &head->chain) {
1057 if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
1058 if (tb->fastreuse >= 0 ||
1059 tb->fastreuseport >= 0)
1060 goto next_port;
1061 WARN_ON(hlist_empty(&tb->bhash2));
1062 if (!check_established(death_row, sk,
1063 port, &tw))
1064 goto ok;
1065 goto next_port;
1066 }
1067 }
1068
1069 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
1070 net, head, port, l3mdev);
1071 if (!tb) {
1072 spin_unlock_bh(&head->lock);
1073 return -ENOMEM;
1074 }
1075 tb_created = true;
1076 tb->fastreuse = -1;
1077 tb->fastreuseport = -1;
1078 goto ok;
1079next_port:
1080 spin_unlock_bh(&head->lock);
1081 cond_resched();
1082 }
1083
1084 if (!local_ports) {
1085 offset++;
1086 if ((offset & 1) && remaining > 1)
1087 goto other_parity_scan;
1088 }
1089 return -EADDRNOTAVAIL;
1090
1091ok:
1092 /* Find the corresponding tb2 bucket since we need to
1093 * add the socket to the bhash2 table as well
1094 */
1095 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
1096 spin_lock(&head2->lock);
1097
1098 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
1099 if (!tb2) {
1100 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
1101 head2, tb, sk);
1102 if (!tb2)
1103 goto error;
1104 }
1105
1106 /* Here we want to add a little bit of randomness to the next source
1107 * port that will be chosen. We use a max() with a random here so that
1108 * on low contention the randomness is maximal and on high contention
1109 * it may be inexistent.
1110 */
1111 i = max_t(int, i, get_random_u32_below(8) * step);
1112 WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + step);
1113
1114 /* Head lock still held and bh's disabled */
1115 inet_bind_hash(sk, tb, tb2, port);
1116
1117 if (sk_unhashed(sk)) {
1118 inet_sk(sk)->inet_sport = htons(port);
1119 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
1120 }
1121 if (tw)
1122 inet_twsk_bind_unhash(tw, hinfo);
1123
1124 spin_unlock(&head2->lock);
1125 spin_unlock(&head->lock);
1126
1127 if (tw)
1128 inet_twsk_deschedule_put(tw);
1129 local_bh_enable();
1130 return 0;
1131
1132error:
1133 if (sk_hashed(sk)) {
1134 spinlock_t *lock = inet_ehash_lockp(hinfo, sk->sk_hash);
1135
1136 sock_prot_inuse_add(net, sk->sk_prot, -1);
1137
1138 spin_lock(lock);
1139 __sk_nulls_del_node_init_rcu(sk);
1140 spin_unlock(lock);
1141
1142 sk->sk_hash = 0;
1143 inet_sk(sk)->inet_sport = 0;
1144 inet_sk(sk)->inet_num = 0;
1145
1146 if (tw)
1147 inet_twsk_bind_unhash(tw, hinfo);
1148 }
1149
1150 spin_unlock(&head2->lock);
1151 if (tb_created)
1152 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
1153 spin_unlock(&head->lock);
1154
1155 if (tw)
1156 inet_twsk_deschedule_put(tw);
1157
1158 local_bh_enable();
1159
1160 return -ENOMEM;
1161}
1162
1163/*
1164 * Bind a port for a connect operation and hash it.
1165 */
1166int inet_hash_connect(struct inet_timewait_death_row *death_row,
1167 struct sock *sk)
1168{
1169 u64 port_offset = 0;
1170
1171 if (!inet_sk(sk)->inet_num)
1172 port_offset = inet_sk_port_offset(sk);
1173 return __inet_hash_connect(death_row, sk, port_offset,
1174 __inet_check_established);
1175}
1176EXPORT_SYMBOL_GPL(inet_hash_connect);
1177
1178static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1179{
1180 int i;
1181
1182 for (i = 0; i <= h->lhash2_mask; i++) {
1183 spin_lock_init(&h->lhash2[i].lock);
1184 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1185 i + LISTENING_NULLS_BASE);
1186 }
1187}
1188
1189void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1190 unsigned long numentries, int scale,
1191 unsigned long low_limit,
1192 unsigned long high_limit)
1193{
1194 h->lhash2 = alloc_large_system_hash(name,
1195 sizeof(*h->lhash2),
1196 numentries,
1197 scale,
1198 0,
1199 NULL,
1200 &h->lhash2_mask,
1201 low_limit,
1202 high_limit);
1203 init_hashinfo_lhash2(h);
1204
1205 /* this one is used for source ports of outgoing connections */
1206 table_perturb = alloc_large_system_hash("Table-perturb",
1207 sizeof(*table_perturb),
1208 INET_TABLE_PERTURB_SIZE,
1209 0, 0, NULL, NULL,
1210 INET_TABLE_PERTURB_SIZE,
1211 INET_TABLE_PERTURB_SIZE);
1212}
1213
1214int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1215{
1216 h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1217 if (!h->lhash2)
1218 return -ENOMEM;
1219
1220 h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1221 /* INET_LHTABLE_SIZE must be a power of 2 */
1222 BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1223
1224 init_hashinfo_lhash2(h);
1225 return 0;
1226}
1227EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
1228
1229int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1230{
1231 unsigned int locksz = sizeof(spinlock_t);
1232 unsigned int i, nblocks = 1;
1233
1234 if (locksz != 0) {
1235 /* allocate 2 cache lines or at least one spinlock per cpu */
1236 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
1237 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
1238
1239 /* no more locks than number of hash buckets */
1240 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1241
1242 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1243 if (!hashinfo->ehash_locks)
1244 return -ENOMEM;
1245
1246 for (i = 0; i < nblocks; i++)
1247 spin_lock_init(&hashinfo->ehash_locks[i]);
1248 }
1249 hashinfo->ehash_locks_mask = nblocks - 1;
1250 return 0;
1251}
1252EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
1253
1254struct inet_hashinfo *inet_pernet_hashinfo_alloc(struct inet_hashinfo *hashinfo,
1255 unsigned int ehash_entries)
1256{
1257 struct inet_hashinfo *new_hashinfo;
1258 int i;
1259
1260 new_hashinfo = kmemdup(hashinfo, sizeof(*hashinfo), GFP_KERNEL);
1261 if (!new_hashinfo)
1262 goto err;
1263
1264 new_hashinfo->ehash = vmalloc_huge(ehash_entries * sizeof(struct inet_ehash_bucket),
1265 GFP_KERNEL_ACCOUNT);
1266 if (!new_hashinfo->ehash)
1267 goto free_hashinfo;
1268
1269 new_hashinfo->ehash_mask = ehash_entries - 1;
1270
1271 if (inet_ehash_locks_alloc(new_hashinfo))
1272 goto free_ehash;
1273
1274 for (i = 0; i < ehash_entries; i++)
1275 INIT_HLIST_NULLS_HEAD(&new_hashinfo->ehash[i].chain, i);
1276
1277 new_hashinfo->pernet = true;
1278
1279 return new_hashinfo;
1280
1281free_ehash:
1282 vfree(new_hashinfo->ehash);
1283free_hashinfo:
1284 kfree(new_hashinfo);
1285err:
1286 return NULL;
1287}
1288EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_alloc);
1289
1290void inet_pernet_hashinfo_free(struct inet_hashinfo *hashinfo)
1291{
1292 if (!hashinfo->pernet)
1293 return;
1294
1295 inet_ehash_locks_free(hashinfo);
1296 vfree(hashinfo->ehash);
1297 kfree(hashinfo);
1298}
1299EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_free);
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#if IS_ENABLED(CONFIG_IPV6)
24#include <net/inet6_hashtables.h>
25#endif
26#include <net/secure_seq.h>
27#include <net/ip.h>
28#include <net/tcp.h>
29#include <net/sock_reuseport.h>
30
31static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
32 const __u16 lport, const __be32 faddr,
33 const __be16 fport)
34{
35 static u32 inet_ehash_secret __read_mostly;
36
37 net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
38
39 return __inet_ehashfn(laddr, lport, faddr, fport,
40 inet_ehash_secret + net_hash_mix(net));
41}
42
43/* This function handles inet_sock, but also timewait and request sockets
44 * for IPv4/IPv6.
45 */
46static u32 sk_ehashfn(const struct sock *sk)
47{
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
55 return inet_ehashfn(sock_net(sk),
56 sk->sk_rcv_saddr, sk->sk_num,
57 sk->sk_daddr, sk->sk_dport);
58}
59
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 */
64struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
65 struct net *net,
66 struct inet_bind_hashbucket *head,
67 const unsigned short snum,
68 int l3mdev)
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->l3mdev = l3mdev;
75 tb->port = snum;
76 tb->fastreuse = 0;
77 tb->fastreuseport = 0;
78 INIT_HLIST_HEAD(&tb->owners);
79 hlist_add_head(&tb->node, &head->chain);
80 }
81 return tb;
82}
83
84/*
85 * Caller must hold hashbucket lock for this tb with local BH disabled
86 */
87void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
88{
89 if (hlist_empty(&tb->owners)) {
90 __hlist_del(&tb->node);
91 kmem_cache_free(cachep, tb);
92 }
93}
94
95void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
96 const unsigned short snum)
97{
98 inet_sk(sk)->inet_num = snum;
99 sk_add_bind_node(sk, &tb->owners);
100 inet_csk(sk)->icsk_bind_hash = tb;
101}
102
103/*
104 * Get rid of any references to a local port held by the given sock.
105 */
106static void __inet_put_port(struct sock *sk)
107{
108 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
109 const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
110 hashinfo->bhash_size);
111 struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
112 struct inet_bind_bucket *tb;
113
114 spin_lock(&head->lock);
115 tb = inet_csk(sk)->icsk_bind_hash;
116 __sk_del_bind_node(sk);
117 inet_csk(sk)->icsk_bind_hash = NULL;
118 inet_sk(sk)->inet_num = 0;
119 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
120 spin_unlock(&head->lock);
121}
122
123void inet_put_port(struct sock *sk)
124{
125 local_bh_disable();
126 __inet_put_port(sk);
127 local_bh_enable();
128}
129EXPORT_SYMBOL(inet_put_port);
130
131int __inet_inherit_port(const struct sock *sk, struct sock *child)
132{
133 struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
134 unsigned short port = inet_sk(child)->inet_num;
135 const int bhash = inet_bhashfn(sock_net(sk), port,
136 table->bhash_size);
137 struct inet_bind_hashbucket *head = &table->bhash[bhash];
138 struct inet_bind_bucket *tb;
139 int l3mdev;
140
141 spin_lock(&head->lock);
142 tb = inet_csk(sk)->icsk_bind_hash;
143 if (unlikely(!tb)) {
144 spin_unlock(&head->lock);
145 return -ENOENT;
146 }
147 if (tb->port != port) {
148 l3mdev = inet_sk_bound_l3mdev(sk);
149
150 /* NOTE: using tproxy and redirecting skbs to a proxy
151 * on a different listener port breaks the assumption
152 * that the listener socket's icsk_bind_hash is the same
153 * as that of the child socket. We have to look up or
154 * create a new bind bucket for the child here. */
155 inet_bind_bucket_for_each(tb, &head->chain) {
156 if (net_eq(ib_net(tb), sock_net(sk)) &&
157 tb->l3mdev == l3mdev && tb->port == port)
158 break;
159 }
160 if (!tb) {
161 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
162 sock_net(sk), head, port,
163 l3mdev);
164 if (!tb) {
165 spin_unlock(&head->lock);
166 return -ENOMEM;
167 }
168 }
169 inet_csk_update_fastreuse(tb, child);
170 }
171 inet_bind_hash(child, tb, port);
172 spin_unlock(&head->lock);
173
174 return 0;
175}
176EXPORT_SYMBOL_GPL(__inet_inherit_port);
177
178static struct inet_listen_hashbucket *
179inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
180{
181 u32 hash;
182
183#if IS_ENABLED(CONFIG_IPV6)
184 if (sk->sk_family == AF_INET6)
185 hash = ipv6_portaddr_hash(sock_net(sk),
186 &sk->sk_v6_rcv_saddr,
187 inet_sk(sk)->inet_num);
188 else
189#endif
190 hash = ipv4_portaddr_hash(sock_net(sk),
191 inet_sk(sk)->inet_rcv_saddr,
192 inet_sk(sk)->inet_num);
193 return inet_lhash2_bucket(h, hash);
194}
195
196static void inet_hash2(struct inet_hashinfo *h, struct sock *sk)
197{
198 struct inet_listen_hashbucket *ilb2;
199
200 if (!h->lhash2)
201 return;
202
203 ilb2 = inet_lhash2_bucket_sk(h, sk);
204
205 spin_lock(&ilb2->lock);
206 if (sk->sk_reuseport && sk->sk_family == AF_INET6)
207 hlist_add_tail_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
208 &ilb2->head);
209 else
210 hlist_add_head_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
211 &ilb2->head);
212 ilb2->count++;
213 spin_unlock(&ilb2->lock);
214}
215
216static void inet_unhash2(struct inet_hashinfo *h, struct sock *sk)
217{
218 struct inet_listen_hashbucket *ilb2;
219
220 if (!h->lhash2 ||
221 WARN_ON_ONCE(hlist_unhashed(&inet_csk(sk)->icsk_listen_portaddr_node)))
222 return;
223
224 ilb2 = inet_lhash2_bucket_sk(h, sk);
225
226 spin_lock(&ilb2->lock);
227 hlist_del_init_rcu(&inet_csk(sk)->icsk_listen_portaddr_node);
228 ilb2->count--;
229 spin_unlock(&ilb2->lock);
230}
231
232static inline int compute_score(struct sock *sk, struct net *net,
233 const unsigned short hnum, const __be32 daddr,
234 const int dif, const int sdif)
235{
236 int score = -1;
237
238 if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
239 !ipv6_only_sock(sk)) {
240 if (sk->sk_rcv_saddr != daddr)
241 return -1;
242
243 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
244 return -1;
245 score = sk->sk_bound_dev_if ? 2 : 1;
246
247 if (sk->sk_family == PF_INET)
248 score++;
249 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
250 score++;
251 }
252 return score;
253}
254
255static inline struct sock *lookup_reuseport(struct net *net, struct sock *sk,
256 struct sk_buff *skb, int doff,
257 __be32 saddr, __be16 sport,
258 __be32 daddr, unsigned short hnum)
259{
260 struct sock *reuse_sk = NULL;
261 u32 phash;
262
263 if (sk->sk_reuseport) {
264 phash = inet_ehashfn(net, daddr, hnum, saddr, sport);
265 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
266 }
267 return reuse_sk;
268}
269
270/*
271 * Here are some nice properties to exploit here. The BSD API
272 * does not allow a listening sock to specify the remote port nor the
273 * remote address for the connection. So always assume those are both
274 * wildcarded during the search since they can never be otherwise.
275 */
276
277/* called with rcu_read_lock() : No refcount taken on the socket */
278static struct sock *inet_lhash2_lookup(struct net *net,
279 struct inet_listen_hashbucket *ilb2,
280 struct sk_buff *skb, int doff,
281 const __be32 saddr, __be16 sport,
282 const __be32 daddr, const unsigned short hnum,
283 const int dif, const int sdif)
284{
285 struct inet_connection_sock *icsk;
286 struct sock *sk, *result = NULL;
287 int score, hiscore = 0;
288
289 inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
290 sk = (struct sock *)icsk;
291 score = compute_score(sk, net, hnum, daddr, dif, sdif);
292 if (score > hiscore) {
293 result = lookup_reuseport(net, sk, skb, doff,
294 saddr, sport, daddr, hnum);
295 if (result)
296 return result;
297
298 result = sk;
299 hiscore = score;
300 }
301 }
302
303 return result;
304}
305
306static inline struct sock *inet_lookup_run_bpf(struct net *net,
307 struct inet_hashinfo *hashinfo,
308 struct sk_buff *skb, int doff,
309 __be32 saddr, __be16 sport,
310 __be32 daddr, u16 hnum)
311{
312 struct sock *sk, *reuse_sk;
313 bool no_reuseport;
314
315 if (hashinfo != &tcp_hashinfo)
316 return NULL; /* only TCP is supported */
317
318 no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP,
319 saddr, sport, daddr, hnum, &sk);
320 if (no_reuseport || IS_ERR_OR_NULL(sk))
321 return sk;
322
323 reuse_sk = lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum);
324 if (reuse_sk)
325 sk = reuse_sk;
326 return sk;
327}
328
329struct sock *__inet_lookup_listener(struct net *net,
330 struct inet_hashinfo *hashinfo,
331 struct sk_buff *skb, int doff,
332 const __be32 saddr, __be16 sport,
333 const __be32 daddr, const unsigned short hnum,
334 const int dif, const int sdif)
335{
336 struct inet_listen_hashbucket *ilb2;
337 struct sock *result = NULL;
338 unsigned int hash2;
339
340 /* Lookup redirect from BPF */
341 if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
342 result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
343 saddr, sport, daddr, hnum);
344 if (result)
345 goto done;
346 }
347
348 hash2 = ipv4_portaddr_hash(net, daddr, hnum);
349 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
350
351 result = inet_lhash2_lookup(net, ilb2, skb, doff,
352 saddr, sport, daddr, hnum,
353 dif, sdif);
354 if (result)
355 goto done;
356
357 /* Lookup lhash2 with INADDR_ANY */
358 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
359 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
360
361 result = inet_lhash2_lookup(net, ilb2, skb, doff,
362 saddr, sport, htonl(INADDR_ANY), hnum,
363 dif, sdif);
364done:
365 if (IS_ERR(result))
366 return NULL;
367 return result;
368}
369EXPORT_SYMBOL_GPL(__inet_lookup_listener);
370
371/* All sockets share common refcount, but have different destructors */
372void sock_gen_put(struct sock *sk)
373{
374 if (!refcount_dec_and_test(&sk->sk_refcnt))
375 return;
376
377 if (sk->sk_state == TCP_TIME_WAIT)
378 inet_twsk_free(inet_twsk(sk));
379 else if (sk->sk_state == TCP_NEW_SYN_RECV)
380 reqsk_free(inet_reqsk(sk));
381 else
382 sk_free(sk);
383}
384EXPORT_SYMBOL_GPL(sock_gen_put);
385
386void sock_edemux(struct sk_buff *skb)
387{
388 sock_gen_put(skb->sk);
389}
390EXPORT_SYMBOL(sock_edemux);
391
392struct sock *__inet_lookup_established(struct net *net,
393 struct inet_hashinfo *hashinfo,
394 const __be32 saddr, const __be16 sport,
395 const __be32 daddr, const u16 hnum,
396 const int dif, const int sdif)
397{
398 INET_ADDR_COOKIE(acookie, saddr, daddr);
399 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
400 struct sock *sk;
401 const struct hlist_nulls_node *node;
402 /* Optimize here for direct hit, only listening connections can
403 * have wildcards anyways.
404 */
405 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
406 unsigned int slot = hash & hashinfo->ehash_mask;
407 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
408
409begin:
410 sk_nulls_for_each_rcu(sk, node, &head->chain) {
411 if (sk->sk_hash != hash)
412 continue;
413 if (likely(INET_MATCH(sk, net, acookie,
414 saddr, daddr, ports, dif, sdif))) {
415 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
416 goto out;
417 if (unlikely(!INET_MATCH(sk, net, acookie,
418 saddr, daddr, ports,
419 dif, sdif))) {
420 sock_gen_put(sk);
421 goto begin;
422 }
423 goto found;
424 }
425 }
426 /*
427 * if the nulls value we got at the end of this lookup is
428 * not the expected one, we must restart lookup.
429 * We probably met an item that was moved to another chain.
430 */
431 if (get_nulls_value(node) != slot)
432 goto begin;
433out:
434 sk = NULL;
435found:
436 return sk;
437}
438EXPORT_SYMBOL_GPL(__inet_lookup_established);
439
440/* called with local bh disabled */
441static int __inet_check_established(struct inet_timewait_death_row *death_row,
442 struct sock *sk, __u16 lport,
443 struct inet_timewait_sock **twp)
444{
445 struct inet_hashinfo *hinfo = death_row->hashinfo;
446 struct inet_sock *inet = inet_sk(sk);
447 __be32 daddr = inet->inet_rcv_saddr;
448 __be32 saddr = inet->inet_daddr;
449 int dif = sk->sk_bound_dev_if;
450 struct net *net = sock_net(sk);
451 int sdif = l3mdev_master_ifindex_by_index(net, dif);
452 INET_ADDR_COOKIE(acookie, saddr, daddr);
453 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
454 unsigned int hash = inet_ehashfn(net, daddr, lport,
455 saddr, inet->inet_dport);
456 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
457 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
458 struct sock *sk2;
459 const struct hlist_nulls_node *node;
460 struct inet_timewait_sock *tw = NULL;
461
462 spin_lock(lock);
463
464 sk_nulls_for_each(sk2, node, &head->chain) {
465 if (sk2->sk_hash != hash)
466 continue;
467
468 if (likely(INET_MATCH(sk2, net, acookie,
469 saddr, daddr, ports, dif, sdif))) {
470 if (sk2->sk_state == TCP_TIME_WAIT) {
471 tw = inet_twsk(sk2);
472 if (twsk_unique(sk, sk2, twp))
473 break;
474 }
475 goto not_unique;
476 }
477 }
478
479 /* Must record num and sport now. Otherwise we will see
480 * in hash table socket with a funny identity.
481 */
482 inet->inet_num = lport;
483 inet->inet_sport = htons(lport);
484 sk->sk_hash = hash;
485 WARN_ON(!sk_unhashed(sk));
486 __sk_nulls_add_node_rcu(sk, &head->chain);
487 if (tw) {
488 sk_nulls_del_node_init_rcu((struct sock *)tw);
489 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
490 }
491 spin_unlock(lock);
492 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
493
494 if (twp) {
495 *twp = tw;
496 } else if (tw) {
497 /* Silly. Should hash-dance instead... */
498 inet_twsk_deschedule_put(tw);
499 }
500 return 0;
501
502not_unique:
503 spin_unlock(lock);
504 return -EADDRNOTAVAIL;
505}
506
507static u32 inet_sk_port_offset(const struct sock *sk)
508{
509 const struct inet_sock *inet = inet_sk(sk);
510
511 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
512 inet->inet_daddr,
513 inet->inet_dport);
514}
515
516/* Searches for an exsiting socket in the ehash bucket list.
517 * Returns true if found, false otherwise.
518 */
519static bool inet_ehash_lookup_by_sk(struct sock *sk,
520 struct hlist_nulls_head *list)
521{
522 const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
523 const int sdif = sk->sk_bound_dev_if;
524 const int dif = sk->sk_bound_dev_if;
525 const struct hlist_nulls_node *node;
526 struct net *net = sock_net(sk);
527 struct sock *esk;
528
529 INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
530
531 sk_nulls_for_each_rcu(esk, node, list) {
532 if (esk->sk_hash != sk->sk_hash)
533 continue;
534 if (sk->sk_family == AF_INET) {
535 if (unlikely(INET_MATCH(esk, net, acookie,
536 sk->sk_daddr,
537 sk->sk_rcv_saddr,
538 ports, dif, sdif))) {
539 return true;
540 }
541 }
542#if IS_ENABLED(CONFIG_IPV6)
543 else if (sk->sk_family == AF_INET6) {
544 if (unlikely(INET6_MATCH(esk, net,
545 &sk->sk_v6_daddr,
546 &sk->sk_v6_rcv_saddr,
547 ports, dif, sdif))) {
548 return true;
549 }
550 }
551#endif
552 }
553 return false;
554}
555
556/* Insert a socket into ehash, and eventually remove another one
557 * (The another one can be a SYN_RECV or TIMEWAIT)
558 * If an existing socket already exists, socket sk is not inserted,
559 * and sets found_dup_sk parameter to true.
560 */
561bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
562{
563 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
564 struct hlist_nulls_head *list;
565 struct inet_ehash_bucket *head;
566 spinlock_t *lock;
567 bool ret = true;
568
569 WARN_ON_ONCE(!sk_unhashed(sk));
570
571 sk->sk_hash = sk_ehashfn(sk);
572 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
573 list = &head->chain;
574 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
575
576 spin_lock(lock);
577 if (osk) {
578 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
579 ret = sk_nulls_del_node_init_rcu(osk);
580 } else if (found_dup_sk) {
581 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
582 if (*found_dup_sk)
583 ret = false;
584 }
585
586 if (ret)
587 __sk_nulls_add_node_rcu(sk, list);
588
589 spin_unlock(lock);
590
591 return ret;
592}
593
594bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
595{
596 bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
597
598 if (ok) {
599 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
600 } else {
601 percpu_counter_inc(sk->sk_prot->orphan_count);
602 inet_sk_set_state(sk, TCP_CLOSE);
603 sock_set_flag(sk, SOCK_DEAD);
604 inet_csk_destroy_sock(sk);
605 }
606 return ok;
607}
608EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
609
610static int inet_reuseport_add_sock(struct sock *sk,
611 struct inet_listen_hashbucket *ilb)
612{
613 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
614 const struct hlist_nulls_node *node;
615 struct sock *sk2;
616 kuid_t uid = sock_i_uid(sk);
617
618 sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
619 if (sk2 != sk &&
620 sk2->sk_family == sk->sk_family &&
621 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
622 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
623 inet_csk(sk2)->icsk_bind_hash == tb &&
624 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
625 inet_rcv_saddr_equal(sk, sk2, false))
626 return reuseport_add_sock(sk, sk2,
627 inet_rcv_saddr_any(sk));
628 }
629
630 return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
631}
632
633int __inet_hash(struct sock *sk, struct sock *osk)
634{
635 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
636 struct inet_listen_hashbucket *ilb;
637 int err = 0;
638
639 if (sk->sk_state != TCP_LISTEN) {
640 inet_ehash_nolisten(sk, osk, NULL);
641 return 0;
642 }
643 WARN_ON(!sk_unhashed(sk));
644 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
645
646 spin_lock(&ilb->lock);
647 if (sk->sk_reuseport) {
648 err = inet_reuseport_add_sock(sk, ilb);
649 if (err)
650 goto unlock;
651 }
652 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
653 sk->sk_family == AF_INET6)
654 __sk_nulls_add_node_tail_rcu(sk, &ilb->nulls_head);
655 else
656 __sk_nulls_add_node_rcu(sk, &ilb->nulls_head);
657 inet_hash2(hashinfo, sk);
658 ilb->count++;
659 sock_set_flag(sk, SOCK_RCU_FREE);
660 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
661unlock:
662 spin_unlock(&ilb->lock);
663
664 return err;
665}
666EXPORT_SYMBOL(__inet_hash);
667
668int inet_hash(struct sock *sk)
669{
670 int err = 0;
671
672 if (sk->sk_state != TCP_CLOSE) {
673 local_bh_disable();
674 err = __inet_hash(sk, NULL);
675 local_bh_enable();
676 }
677
678 return err;
679}
680EXPORT_SYMBOL_GPL(inet_hash);
681
682void inet_unhash(struct sock *sk)
683{
684 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
685 struct inet_listen_hashbucket *ilb = NULL;
686 spinlock_t *lock;
687
688 if (sk_unhashed(sk))
689 return;
690
691 if (sk->sk_state == TCP_LISTEN) {
692 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
693 lock = &ilb->lock;
694 } else {
695 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
696 }
697 spin_lock_bh(lock);
698 if (sk_unhashed(sk))
699 goto unlock;
700
701 if (rcu_access_pointer(sk->sk_reuseport_cb))
702 reuseport_stop_listen_sock(sk);
703 if (ilb) {
704 inet_unhash2(hashinfo, sk);
705 ilb->count--;
706 }
707 __sk_nulls_del_node_init_rcu(sk);
708 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
709unlock:
710 spin_unlock_bh(lock);
711}
712EXPORT_SYMBOL_GPL(inet_unhash);
713
714/* RFC 6056 3.3.4. Algorithm 4: Double-Hash Port Selection Algorithm
715 * Note that we use 32bit integers (vs RFC 'short integers')
716 * because 2^16 is not a multiple of num_ephemeral and this
717 * property might be used by clever attacker.
718 * RFC claims using TABLE_LENGTH=10 buckets gives an improvement,
719 * we use 256 instead to really give more isolation and
720 * privacy, this only consumes 1 KB of kernel memory.
721 */
722#define INET_TABLE_PERTURB_SHIFT 8
723static u32 table_perturb[1 << INET_TABLE_PERTURB_SHIFT];
724
725int __inet_hash_connect(struct inet_timewait_death_row *death_row,
726 struct sock *sk, u32 port_offset,
727 int (*check_established)(struct inet_timewait_death_row *,
728 struct sock *, __u16, struct inet_timewait_sock **))
729{
730 struct inet_hashinfo *hinfo = death_row->hashinfo;
731 struct inet_timewait_sock *tw = NULL;
732 struct inet_bind_hashbucket *head;
733 int port = inet_sk(sk)->inet_num;
734 struct net *net = sock_net(sk);
735 struct inet_bind_bucket *tb;
736 u32 remaining, offset;
737 int ret, i, low, high;
738 int l3mdev;
739 u32 index;
740
741 if (port) {
742 head = &hinfo->bhash[inet_bhashfn(net, port,
743 hinfo->bhash_size)];
744 tb = inet_csk(sk)->icsk_bind_hash;
745 spin_lock_bh(&head->lock);
746 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
747 inet_ehash_nolisten(sk, NULL, NULL);
748 spin_unlock_bh(&head->lock);
749 return 0;
750 }
751 spin_unlock(&head->lock);
752 /* No definite answer... Walk to established hash table */
753 ret = check_established(death_row, sk, port, NULL);
754 local_bh_enable();
755 return ret;
756 }
757
758 l3mdev = inet_sk_bound_l3mdev(sk);
759
760 inet_get_local_port_range(net, &low, &high);
761 high++; /* [32768, 60999] -> [32768, 61000[ */
762 remaining = high - low;
763 if (likely(remaining > 1))
764 remaining &= ~1U;
765
766 net_get_random_once(table_perturb, sizeof(table_perturb));
767 index = hash_32(port_offset, INET_TABLE_PERTURB_SHIFT);
768
769 offset = (READ_ONCE(table_perturb[index]) + port_offset) % remaining;
770 /* In first pass we try ports of @low parity.
771 * inet_csk_get_port() does the opposite choice.
772 */
773 offset &= ~1U;
774other_parity_scan:
775 port = low + offset;
776 for (i = 0; i < remaining; i += 2, port += 2) {
777 if (unlikely(port >= high))
778 port -= remaining;
779 if (inet_is_local_reserved_port(net, port))
780 continue;
781 head = &hinfo->bhash[inet_bhashfn(net, port,
782 hinfo->bhash_size)];
783 spin_lock_bh(&head->lock);
784
785 /* Does not bother with rcv_saddr checks, because
786 * the established check is already unique enough.
787 */
788 inet_bind_bucket_for_each(tb, &head->chain) {
789 if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
790 tb->port == port) {
791 if (tb->fastreuse >= 0 ||
792 tb->fastreuseport >= 0)
793 goto next_port;
794 WARN_ON(hlist_empty(&tb->owners));
795 if (!check_established(death_row, sk,
796 port, &tw))
797 goto ok;
798 goto next_port;
799 }
800 }
801
802 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
803 net, head, port, l3mdev);
804 if (!tb) {
805 spin_unlock_bh(&head->lock);
806 return -ENOMEM;
807 }
808 tb->fastreuse = -1;
809 tb->fastreuseport = -1;
810 goto ok;
811next_port:
812 spin_unlock_bh(&head->lock);
813 cond_resched();
814 }
815
816 offset++;
817 if ((offset & 1) && remaining > 1)
818 goto other_parity_scan;
819
820 return -EADDRNOTAVAIL;
821
822ok:
823 /* If our first attempt found a candidate, skip next candidate
824 * in 1/16 of cases to add some noise.
825 */
826 if (!i && !(prandom_u32() % 16))
827 i = 2;
828 WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
829
830 /* Head lock still held and bh's disabled */
831 inet_bind_hash(sk, tb, port);
832 if (sk_unhashed(sk)) {
833 inet_sk(sk)->inet_sport = htons(port);
834 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
835 }
836 if (tw)
837 inet_twsk_bind_unhash(tw, hinfo);
838 spin_unlock(&head->lock);
839 if (tw)
840 inet_twsk_deschedule_put(tw);
841 local_bh_enable();
842 return 0;
843}
844
845/*
846 * Bind a port for a connect operation and hash it.
847 */
848int inet_hash_connect(struct inet_timewait_death_row *death_row,
849 struct sock *sk)
850{
851 u32 port_offset = 0;
852
853 if (!inet_sk(sk)->inet_num)
854 port_offset = inet_sk_port_offset(sk);
855 return __inet_hash_connect(death_row, sk, port_offset,
856 __inet_check_established);
857}
858EXPORT_SYMBOL_GPL(inet_hash_connect);
859
860void inet_hashinfo_init(struct inet_hashinfo *h)
861{
862 int i;
863
864 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
865 spin_lock_init(&h->listening_hash[i].lock);
866 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].nulls_head,
867 i + LISTENING_NULLS_BASE);
868 h->listening_hash[i].count = 0;
869 }
870
871 h->lhash2 = NULL;
872}
873EXPORT_SYMBOL_GPL(inet_hashinfo_init);
874
875static void init_hashinfo_lhash2(struct inet_hashinfo *h)
876{
877 int i;
878
879 for (i = 0; i <= h->lhash2_mask; i++) {
880 spin_lock_init(&h->lhash2[i].lock);
881 INIT_HLIST_HEAD(&h->lhash2[i].head);
882 h->lhash2[i].count = 0;
883 }
884}
885
886void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
887 unsigned long numentries, int scale,
888 unsigned long low_limit,
889 unsigned long high_limit)
890{
891 h->lhash2 = alloc_large_system_hash(name,
892 sizeof(*h->lhash2),
893 numentries,
894 scale,
895 0,
896 NULL,
897 &h->lhash2_mask,
898 low_limit,
899 high_limit);
900 init_hashinfo_lhash2(h);
901}
902
903int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
904{
905 h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
906 if (!h->lhash2)
907 return -ENOMEM;
908
909 h->lhash2_mask = INET_LHTABLE_SIZE - 1;
910 /* INET_LHTABLE_SIZE must be a power of 2 */
911 BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
912
913 init_hashinfo_lhash2(h);
914 return 0;
915}
916EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
917
918int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
919{
920 unsigned int locksz = sizeof(spinlock_t);
921 unsigned int i, nblocks = 1;
922
923 if (locksz != 0) {
924 /* allocate 2 cache lines or at least one spinlock per cpu */
925 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
926 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
927
928 /* no more locks than number of hash buckets */
929 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
930
931 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
932 if (!hashinfo->ehash_locks)
933 return -ENOMEM;
934
935 for (i = 0; i < nblocks; i++)
936 spin_lock_init(&hashinfo->ehash_locks[i]);
937 }
938 hashinfo->ehash_locks_mask = nblocks - 1;
939 return 0;
940}
941EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);