xref: /openbmc/linux/net/ipv4/inet_hashtables.c (revision ca48739e)
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 
31 static 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  */
46 static 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  */
64 struct 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  */
87 void 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 
95 void 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  */
106 static 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 
123 void inet_put_port(struct sock *sk)
124 {
125 	local_bh_disable();
126 	__inet_put_port(sk);
127 	local_bh_enable();
128 }
129 EXPORT_SYMBOL(inet_put_port);
130 
131 int __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 }
176 EXPORT_SYMBOL_GPL(__inet_inherit_port);
177 
178 static struct inet_listen_hashbucket *
179 inet_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 
196 static 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 
216 static 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 
232 static 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 
246 		score = sk->sk_family == PF_INET ? 2 : 1;
247 		if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
248 			score++;
249 	}
250 	return score;
251 }
252 
253 static inline struct sock *lookup_reuseport(struct net *net, struct sock *sk,
254 					    struct sk_buff *skb, int doff,
255 					    __be32 saddr, __be16 sport,
256 					    __be32 daddr, unsigned short hnum)
257 {
258 	struct sock *reuse_sk = NULL;
259 	u32 phash;
260 
261 	if (sk->sk_reuseport) {
262 		phash = inet_ehashfn(net, daddr, hnum, saddr, sport);
263 		reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
264 	}
265 	return reuse_sk;
266 }
267 
268 /*
269  * Here are some nice properties to exploit here. The BSD API
270  * does not allow a listening sock to specify the remote port nor the
271  * remote address for the connection. So always assume those are both
272  * wildcarded during the search since they can never be otherwise.
273  */
274 
275 /* called with rcu_read_lock() : No refcount taken on the socket */
276 static struct sock *inet_lhash2_lookup(struct net *net,
277 				struct inet_listen_hashbucket *ilb2,
278 				struct sk_buff *skb, int doff,
279 				const __be32 saddr, __be16 sport,
280 				const __be32 daddr, const unsigned short hnum,
281 				const int dif, const int sdif)
282 {
283 	struct inet_connection_sock *icsk;
284 	struct sock *sk, *result = NULL;
285 	int score, hiscore = 0;
286 
287 	inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
288 		sk = (struct sock *)icsk;
289 		score = compute_score(sk, net, hnum, daddr, dif, sdif);
290 		if (score > hiscore) {
291 			result = lookup_reuseport(net, sk, skb, doff,
292 						  saddr, sport, daddr, hnum);
293 			if (result)
294 				return result;
295 
296 			result = sk;
297 			hiscore = score;
298 		}
299 	}
300 
301 	return result;
302 }
303 
304 static inline struct sock *inet_lookup_run_bpf(struct net *net,
305 					       struct inet_hashinfo *hashinfo,
306 					       struct sk_buff *skb, int doff,
307 					       __be32 saddr, __be16 sport,
308 					       __be32 daddr, u16 hnum)
309 {
310 	struct sock *sk, *reuse_sk;
311 	bool no_reuseport;
312 
313 	if (hashinfo != &tcp_hashinfo)
314 		return NULL; /* only TCP is supported */
315 
316 	no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP,
317 					    saddr, sport, daddr, hnum, &sk);
318 	if (no_reuseport || IS_ERR_OR_NULL(sk))
319 		return sk;
320 
321 	reuse_sk = lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum);
322 	if (reuse_sk)
323 		sk = reuse_sk;
324 	return sk;
325 }
326 
327 struct sock *__inet_lookup_listener(struct net *net,
328 				    struct inet_hashinfo *hashinfo,
329 				    struct sk_buff *skb, int doff,
330 				    const __be32 saddr, __be16 sport,
331 				    const __be32 daddr, const unsigned short hnum,
332 				    const int dif, const int sdif)
333 {
334 	struct inet_listen_hashbucket *ilb2;
335 	struct sock *result = NULL;
336 	unsigned int hash2;
337 
338 	/* Lookup redirect from BPF */
339 	if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
340 		result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
341 					     saddr, sport, daddr, hnum);
342 		if (result)
343 			goto done;
344 	}
345 
346 	hash2 = ipv4_portaddr_hash(net, daddr, hnum);
347 	ilb2 = inet_lhash2_bucket(hashinfo, hash2);
348 
349 	result = inet_lhash2_lookup(net, ilb2, skb, doff,
350 				    saddr, sport, daddr, hnum,
351 				    dif, sdif);
352 	if (result)
353 		goto done;
354 
355 	/* Lookup lhash2 with INADDR_ANY */
356 	hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
357 	ilb2 = inet_lhash2_bucket(hashinfo, hash2);
358 
359 	result = inet_lhash2_lookup(net, ilb2, skb, doff,
360 				    saddr, sport, htonl(INADDR_ANY), hnum,
361 				    dif, sdif);
362 done:
363 	if (IS_ERR(result))
364 		return NULL;
365 	return result;
366 }
367 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
368 
369 /* All sockets share common refcount, but have different destructors */
370 void sock_gen_put(struct sock *sk)
371 {
372 	if (!refcount_dec_and_test(&sk->sk_refcnt))
373 		return;
374 
375 	if (sk->sk_state == TCP_TIME_WAIT)
376 		inet_twsk_free(inet_twsk(sk));
377 	else if (sk->sk_state == TCP_NEW_SYN_RECV)
378 		reqsk_free(inet_reqsk(sk));
379 	else
380 		sk_free(sk);
381 }
382 EXPORT_SYMBOL_GPL(sock_gen_put);
383 
384 void sock_edemux(struct sk_buff *skb)
385 {
386 	sock_gen_put(skb->sk);
387 }
388 EXPORT_SYMBOL(sock_edemux);
389 
390 struct sock *__inet_lookup_established(struct net *net,
391 				  struct inet_hashinfo *hashinfo,
392 				  const __be32 saddr, const __be16 sport,
393 				  const __be32 daddr, const u16 hnum,
394 				  const int dif, const int sdif)
395 {
396 	INET_ADDR_COOKIE(acookie, saddr, daddr);
397 	const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
398 	struct sock *sk;
399 	const struct hlist_nulls_node *node;
400 	/* Optimize here for direct hit, only listening connections can
401 	 * have wildcards anyways.
402 	 */
403 	unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
404 	unsigned int slot = hash & hashinfo->ehash_mask;
405 	struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
406 
407 begin:
408 	sk_nulls_for_each_rcu(sk, node, &head->chain) {
409 		if (sk->sk_hash != hash)
410 			continue;
411 		if (likely(INET_MATCH(sk, net, acookie,
412 				      saddr, daddr, ports, dif, sdif))) {
413 			if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
414 				goto out;
415 			if (unlikely(!INET_MATCH(sk, net, acookie,
416 						 saddr, daddr, ports,
417 						 dif, sdif))) {
418 				sock_gen_put(sk);
419 				goto begin;
420 			}
421 			goto found;
422 		}
423 	}
424 	/*
425 	 * if the nulls value we got at the end of this lookup is
426 	 * not the expected one, we must restart lookup.
427 	 * We probably met an item that was moved to another chain.
428 	 */
429 	if (get_nulls_value(node) != slot)
430 		goto begin;
431 out:
432 	sk = NULL;
433 found:
434 	return sk;
435 }
436 EXPORT_SYMBOL_GPL(__inet_lookup_established);
437 
438 /* called with local bh disabled */
439 static int __inet_check_established(struct inet_timewait_death_row *death_row,
440 				    struct sock *sk, __u16 lport,
441 				    struct inet_timewait_sock **twp)
442 {
443 	struct inet_hashinfo *hinfo = death_row->hashinfo;
444 	struct inet_sock *inet = inet_sk(sk);
445 	__be32 daddr = inet->inet_rcv_saddr;
446 	__be32 saddr = inet->inet_daddr;
447 	int dif = sk->sk_bound_dev_if;
448 	struct net *net = sock_net(sk);
449 	int sdif = l3mdev_master_ifindex_by_index(net, dif);
450 	INET_ADDR_COOKIE(acookie, saddr, daddr);
451 	const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
452 	unsigned int hash = inet_ehashfn(net, daddr, lport,
453 					 saddr, inet->inet_dport);
454 	struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
455 	spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
456 	struct sock *sk2;
457 	const struct hlist_nulls_node *node;
458 	struct inet_timewait_sock *tw = NULL;
459 
460 	spin_lock(lock);
461 
462 	sk_nulls_for_each(sk2, node, &head->chain) {
463 		if (sk2->sk_hash != hash)
464 			continue;
465 
466 		if (likely(INET_MATCH(sk2, net, acookie,
467 					 saddr, daddr, ports, dif, sdif))) {
468 			if (sk2->sk_state == TCP_TIME_WAIT) {
469 				tw = inet_twsk(sk2);
470 				if (twsk_unique(sk, sk2, twp))
471 					break;
472 			}
473 			goto not_unique;
474 		}
475 	}
476 
477 	/* Must record num and sport now. Otherwise we will see
478 	 * in hash table socket with a funny identity.
479 	 */
480 	inet->inet_num = lport;
481 	inet->inet_sport = htons(lport);
482 	sk->sk_hash = hash;
483 	WARN_ON(!sk_unhashed(sk));
484 	__sk_nulls_add_node_rcu(sk, &head->chain);
485 	if (tw) {
486 		sk_nulls_del_node_init_rcu((struct sock *)tw);
487 		__NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
488 	}
489 	spin_unlock(lock);
490 	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
491 
492 	if (twp) {
493 		*twp = tw;
494 	} else if (tw) {
495 		/* Silly. Should hash-dance instead... */
496 		inet_twsk_deschedule_put(tw);
497 	}
498 	return 0;
499 
500 not_unique:
501 	spin_unlock(lock);
502 	return -EADDRNOTAVAIL;
503 }
504 
505 static u32 inet_sk_port_offset(const struct sock *sk)
506 {
507 	const struct inet_sock *inet = inet_sk(sk);
508 
509 	return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
510 					  inet->inet_daddr,
511 					  inet->inet_dport);
512 }
513 
514 /* Searches for an exsiting socket in the ehash bucket list.
515  * Returns true if found, false otherwise.
516  */
517 static bool inet_ehash_lookup_by_sk(struct sock *sk,
518 				    struct hlist_nulls_head *list)
519 {
520 	const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
521 	const int sdif = sk->sk_bound_dev_if;
522 	const int dif = sk->sk_bound_dev_if;
523 	const struct hlist_nulls_node *node;
524 	struct net *net = sock_net(sk);
525 	struct sock *esk;
526 
527 	INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
528 
529 	sk_nulls_for_each_rcu(esk, node, list) {
530 		if (esk->sk_hash != sk->sk_hash)
531 			continue;
532 		if (sk->sk_family == AF_INET) {
533 			if (unlikely(INET_MATCH(esk, net, acookie,
534 						sk->sk_daddr,
535 						sk->sk_rcv_saddr,
536 						ports, dif, sdif))) {
537 				return true;
538 			}
539 		}
540 #if IS_ENABLED(CONFIG_IPV6)
541 		else if (sk->sk_family == AF_INET6) {
542 			if (unlikely(INET6_MATCH(esk, net,
543 						 &sk->sk_v6_daddr,
544 						 &sk->sk_v6_rcv_saddr,
545 						 ports, dif, sdif))) {
546 				return true;
547 			}
548 		}
549 #endif
550 	}
551 	return false;
552 }
553 
554 /* Insert a socket into ehash, and eventually remove another one
555  * (The another one can be a SYN_RECV or TIMEWAIT)
556  * If an existing socket already exists, socket sk is not inserted,
557  * and sets found_dup_sk parameter to true.
558  */
559 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
560 {
561 	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
562 	struct hlist_nulls_head *list;
563 	struct inet_ehash_bucket *head;
564 	spinlock_t *lock;
565 	bool ret = true;
566 
567 	WARN_ON_ONCE(!sk_unhashed(sk));
568 
569 	sk->sk_hash = sk_ehashfn(sk);
570 	head = inet_ehash_bucket(hashinfo, sk->sk_hash);
571 	list = &head->chain;
572 	lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
573 
574 	spin_lock(lock);
575 	if (osk) {
576 		WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
577 		ret = sk_nulls_del_node_init_rcu(osk);
578 	} else if (found_dup_sk) {
579 		*found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
580 		if (*found_dup_sk)
581 			ret = false;
582 	}
583 
584 	if (ret)
585 		__sk_nulls_add_node_rcu(sk, list);
586 
587 	spin_unlock(lock);
588 
589 	return ret;
590 }
591 
592 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
593 {
594 	bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
595 
596 	if (ok) {
597 		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
598 	} else {
599 		percpu_counter_inc(sk->sk_prot->orphan_count);
600 		inet_sk_set_state(sk, TCP_CLOSE);
601 		sock_set_flag(sk, SOCK_DEAD);
602 		inet_csk_destroy_sock(sk);
603 	}
604 	return ok;
605 }
606 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
607 
608 static int inet_reuseport_add_sock(struct sock *sk,
609 				   struct inet_listen_hashbucket *ilb)
610 {
611 	struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
612 	const struct hlist_nulls_node *node;
613 	struct sock *sk2;
614 	kuid_t uid = sock_i_uid(sk);
615 
616 	sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
617 		if (sk2 != sk &&
618 		    sk2->sk_family == sk->sk_family &&
619 		    ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
620 		    sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
621 		    inet_csk(sk2)->icsk_bind_hash == tb &&
622 		    sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
623 		    inet_rcv_saddr_equal(sk, sk2, false))
624 			return reuseport_add_sock(sk, sk2,
625 						  inet_rcv_saddr_any(sk));
626 	}
627 
628 	return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
629 }
630 
631 int __inet_hash(struct sock *sk, struct sock *osk)
632 {
633 	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
634 	struct inet_listen_hashbucket *ilb;
635 	int err = 0;
636 
637 	if (sk->sk_state != TCP_LISTEN) {
638 		inet_ehash_nolisten(sk, osk, NULL);
639 		return 0;
640 	}
641 	WARN_ON(!sk_unhashed(sk));
642 	ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
643 
644 	spin_lock(&ilb->lock);
645 	if (sk->sk_reuseport) {
646 		err = inet_reuseport_add_sock(sk, ilb);
647 		if (err)
648 			goto unlock;
649 	}
650 	if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
651 		sk->sk_family == AF_INET6)
652 		__sk_nulls_add_node_tail_rcu(sk, &ilb->nulls_head);
653 	else
654 		__sk_nulls_add_node_rcu(sk, &ilb->nulls_head);
655 	inet_hash2(hashinfo, sk);
656 	ilb->count++;
657 	sock_set_flag(sk, SOCK_RCU_FREE);
658 	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
659 unlock:
660 	spin_unlock(&ilb->lock);
661 
662 	return err;
663 }
664 EXPORT_SYMBOL(__inet_hash);
665 
666 int inet_hash(struct sock *sk)
667 {
668 	int err = 0;
669 
670 	if (sk->sk_state != TCP_CLOSE) {
671 		local_bh_disable();
672 		err = __inet_hash(sk, NULL);
673 		local_bh_enable();
674 	}
675 
676 	return err;
677 }
678 EXPORT_SYMBOL_GPL(inet_hash);
679 
680 void inet_unhash(struct sock *sk)
681 {
682 	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
683 	struct inet_listen_hashbucket *ilb = NULL;
684 	spinlock_t *lock;
685 
686 	if (sk_unhashed(sk))
687 		return;
688 
689 	if (sk->sk_state == TCP_LISTEN) {
690 		ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
691 		lock = &ilb->lock;
692 	} else {
693 		lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
694 	}
695 	spin_lock_bh(lock);
696 	if (sk_unhashed(sk))
697 		goto unlock;
698 
699 	if (rcu_access_pointer(sk->sk_reuseport_cb))
700 		reuseport_detach_sock(sk);
701 	if (ilb) {
702 		inet_unhash2(hashinfo, sk);
703 		ilb->count--;
704 	}
705 	__sk_nulls_del_node_init_rcu(sk);
706 	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
707 unlock:
708 	spin_unlock_bh(lock);
709 }
710 EXPORT_SYMBOL_GPL(inet_unhash);
711 
712 /* RFC 6056 3.3.4.  Algorithm 4: Double-Hash Port Selection Algorithm
713  * Note that we use 32bit integers (vs RFC 'short integers')
714  * because 2^16 is not a multiple of num_ephemeral and this
715  * property might be used by clever attacker.
716  * RFC claims using TABLE_LENGTH=10 buckets gives an improvement,
717  * we use 256 instead to really give more isolation and
718  * privacy, this only consumes 1 KB of kernel memory.
719  */
720 #define INET_TABLE_PERTURB_SHIFT 8
721 static u32 table_perturb[1 << INET_TABLE_PERTURB_SHIFT];
722 
723 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
724 		struct sock *sk, u32 port_offset,
725 		int (*check_established)(struct inet_timewait_death_row *,
726 			struct sock *, __u16, struct inet_timewait_sock **))
727 {
728 	struct inet_hashinfo *hinfo = death_row->hashinfo;
729 	struct inet_timewait_sock *tw = NULL;
730 	struct inet_bind_hashbucket *head;
731 	int port = inet_sk(sk)->inet_num;
732 	struct net *net = sock_net(sk);
733 	struct inet_bind_bucket *tb;
734 	u32 remaining, offset;
735 	int ret, i, low, high;
736 	int l3mdev;
737 	u32 index;
738 
739 	if (port) {
740 		head = &hinfo->bhash[inet_bhashfn(net, port,
741 						  hinfo->bhash_size)];
742 		tb = inet_csk(sk)->icsk_bind_hash;
743 		spin_lock_bh(&head->lock);
744 		if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
745 			inet_ehash_nolisten(sk, NULL, NULL);
746 			spin_unlock_bh(&head->lock);
747 			return 0;
748 		}
749 		spin_unlock(&head->lock);
750 		/* No definite answer... Walk to established hash table */
751 		ret = check_established(death_row, sk, port, NULL);
752 		local_bh_enable();
753 		return ret;
754 	}
755 
756 	l3mdev = inet_sk_bound_l3mdev(sk);
757 
758 	inet_get_local_port_range(net, &low, &high);
759 	high++; /* [32768, 60999] -> [32768, 61000[ */
760 	remaining = high - low;
761 	if (likely(remaining > 1))
762 		remaining &= ~1U;
763 
764 	net_get_random_once(table_perturb, sizeof(table_perturb));
765 	index = hash_32(port_offset, INET_TABLE_PERTURB_SHIFT);
766 
767 	offset = (READ_ONCE(table_perturb[index]) + port_offset) % remaining;
768 	/* In first pass we try ports of @low parity.
769 	 * inet_csk_get_port() does the opposite choice.
770 	 */
771 	offset &= ~1U;
772 other_parity_scan:
773 	port = low + offset;
774 	for (i = 0; i < remaining; i += 2, port += 2) {
775 		if (unlikely(port >= high))
776 			port -= remaining;
777 		if (inet_is_local_reserved_port(net, port))
778 			continue;
779 		head = &hinfo->bhash[inet_bhashfn(net, port,
780 						  hinfo->bhash_size)];
781 		spin_lock_bh(&head->lock);
782 
783 		/* Does not bother with rcv_saddr checks, because
784 		 * the established check is already unique enough.
785 		 */
786 		inet_bind_bucket_for_each(tb, &head->chain) {
787 			if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
788 			    tb->port == port) {
789 				if (tb->fastreuse >= 0 ||
790 				    tb->fastreuseport >= 0)
791 					goto next_port;
792 				WARN_ON(hlist_empty(&tb->owners));
793 				if (!check_established(death_row, sk,
794 						       port, &tw))
795 					goto ok;
796 				goto next_port;
797 			}
798 		}
799 
800 		tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
801 					     net, head, port, l3mdev);
802 		if (!tb) {
803 			spin_unlock_bh(&head->lock);
804 			return -ENOMEM;
805 		}
806 		tb->fastreuse = -1;
807 		tb->fastreuseport = -1;
808 		goto ok;
809 next_port:
810 		spin_unlock_bh(&head->lock);
811 		cond_resched();
812 	}
813 
814 	offset++;
815 	if ((offset & 1) && remaining > 1)
816 		goto other_parity_scan;
817 
818 	return -EADDRNOTAVAIL;
819 
820 ok:
821 	/* If our first attempt found a candidate, skip next candidate
822 	 * in 1/16 of cases to add some noise.
823 	 */
824 	if (!i && !(prandom_u32() % 16))
825 		i = 2;
826 	WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
827 
828 	/* Head lock still held and bh's disabled */
829 	inet_bind_hash(sk, tb, port);
830 	if (sk_unhashed(sk)) {
831 		inet_sk(sk)->inet_sport = htons(port);
832 		inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
833 	}
834 	if (tw)
835 		inet_twsk_bind_unhash(tw, hinfo);
836 	spin_unlock(&head->lock);
837 	if (tw)
838 		inet_twsk_deschedule_put(tw);
839 	local_bh_enable();
840 	return 0;
841 }
842 
843 /*
844  * Bind a port for a connect operation and hash it.
845  */
846 int inet_hash_connect(struct inet_timewait_death_row *death_row,
847 		      struct sock *sk)
848 {
849 	u32 port_offset = 0;
850 
851 	if (!inet_sk(sk)->inet_num)
852 		port_offset = inet_sk_port_offset(sk);
853 	return __inet_hash_connect(death_row, sk, port_offset,
854 				   __inet_check_established);
855 }
856 EXPORT_SYMBOL_GPL(inet_hash_connect);
857 
858 void inet_hashinfo_init(struct inet_hashinfo *h)
859 {
860 	int i;
861 
862 	for (i = 0; i < INET_LHTABLE_SIZE; i++) {
863 		spin_lock_init(&h->listening_hash[i].lock);
864 		INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].nulls_head,
865 				      i + LISTENING_NULLS_BASE);
866 		h->listening_hash[i].count = 0;
867 	}
868 
869 	h->lhash2 = NULL;
870 }
871 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
872 
873 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
874 {
875 	int i;
876 
877 	for (i = 0; i <= h->lhash2_mask; i++) {
878 		spin_lock_init(&h->lhash2[i].lock);
879 		INIT_HLIST_HEAD(&h->lhash2[i].head);
880 		h->lhash2[i].count = 0;
881 	}
882 }
883 
884 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
885 				unsigned long numentries, int scale,
886 				unsigned long low_limit,
887 				unsigned long high_limit)
888 {
889 	h->lhash2 = alloc_large_system_hash(name,
890 					    sizeof(*h->lhash2),
891 					    numentries,
892 					    scale,
893 					    0,
894 					    NULL,
895 					    &h->lhash2_mask,
896 					    low_limit,
897 					    high_limit);
898 	init_hashinfo_lhash2(h);
899 }
900 
901 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
902 {
903 	h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
904 	if (!h->lhash2)
905 		return -ENOMEM;
906 
907 	h->lhash2_mask = INET_LHTABLE_SIZE - 1;
908 	/* INET_LHTABLE_SIZE must be a power of 2 */
909 	BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
910 
911 	init_hashinfo_lhash2(h);
912 	return 0;
913 }
914 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
915 
916 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
917 {
918 	unsigned int locksz = sizeof(spinlock_t);
919 	unsigned int i, nblocks = 1;
920 
921 	if (locksz != 0) {
922 		/* allocate 2 cache lines or at least one spinlock per cpu */
923 		nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
924 		nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
925 
926 		/* no more locks than number of hash buckets */
927 		nblocks = min(nblocks, hashinfo->ehash_mask + 1);
928 
929 		hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
930 		if (!hashinfo->ehash_locks)
931 			return -ENOMEM;
932 
933 		for (i = 0; i < nblocks; i++)
934 			spin_lock_init(&hashinfo->ehash_locks[i]);
935 	}
936 	hashinfo->ehash_locks_mask = nblocks - 1;
937 	return 0;
938 }
939 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
940