Linux Audio

Check our new training course

Loading...
v4.6
  1/*
  2 * INET		An implementation of the TCP/IP protocol suite for the LINUX
  3 *		operating system.  INET is implemented using the BSD Socket
  4 *		interface as the means of communication with the user level.
  5 *
  6 *		Generic INET transport hashtables
  7 *
  8 * Authors:	Lotsa people, from code originally in tcp
  9 *
 10 *	This program is free software; you can redistribute it and/or
 11 *      modify it under the terms of the GNU General Public License
 12 *      as published by the Free Software Foundation; either version
 13 *      2 of the License, or (at your option) any later version.
 14 */
 15
 16#include <linux/module.h>
 17#include <linux/random.h>
 18#include <linux/sched.h>
 19#include <linux/slab.h>
 20#include <linux/wait.h>
 21#include <linux/vmalloc.h>
 22
 23#include <net/addrconf.h>
 24#include <net/inet_connection_sock.h>
 25#include <net/inet_hashtables.h>
 26#include <net/secure_seq.h>
 27#include <net/ip.h>
 
 28#include <net/sock_reuseport.h>
 29
 30static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
 31			const __u16 lport, const __be32 faddr,
 32			const __be16 fport)
 33{
 34	static u32 inet_ehash_secret __read_mostly;
 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}
 41
 42/* This function handles inet_sock, but also timewait and request sockets
 43 * for IPv4/IPv6.
 44 */
 45u32 sk_ehashfn(const struct sock *sk)
 46{
 47#if IS_ENABLED(CONFIG_IPV6)
 48	if (sk->sk_family == AF_INET6 &&
 49	    !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
 50		return inet6_ehashfn(sock_net(sk),
 51				     &sk->sk_v6_rcv_saddr, sk->sk_num,
 52				     &sk->sk_v6_daddr, sk->sk_dport);
 53#endif
 54	return inet_ehashfn(sock_net(sk),
 55			    sk->sk_rcv_saddr, sk->sk_num,
 56			    sk->sk_daddr, sk->sk_dport);
 57}
 58
 59/*
 60 * Allocate and initialize a new local port bind bucket.
 61 * The bindhash mutex for snum's hash chain must be held here.
 62 */
 63struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
 64						 struct net *net,
 65						 struct inet_bind_hashbucket *head,
 66						 const unsigned short snum)
 67{
 68	struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
 69
 70	if (tb) {
 71		write_pnet(&tb->ib_net, net);
 72		tb->port      = snum;
 73		tb->fastreuse = 0;
 74		tb->fastreuseport = 0;
 75		tb->num_owners = 0;
 76		INIT_HLIST_HEAD(&tb->owners);
 77		hlist_add_head(&tb->node, &head->chain);
 78	}
 79	return tb;
 80}
 81
 82/*
 83 * Caller must hold hashbucket lock for this tb with local BH disabled
 84 */
 85void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
 86{
 87	if (hlist_empty(&tb->owners)) {
 88		__hlist_del(&tb->node);
 89		kmem_cache_free(cachep, tb);
 90	}
 91}
 92
 93void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 94		    const unsigned short snum)
 95{
 96	inet_sk(sk)->inet_num = snum;
 97	sk_add_bind_node(sk, &tb->owners);
 98	tb->num_owners++;
 99	inet_csk(sk)->icsk_bind_hash = tb;
100}
101
102/*
103 * Get rid of any references to a local port held by the given sock.
104 */
105static void __inet_put_port(struct sock *sk)
106{
107	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
108	const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
109			hashinfo->bhash_size);
110	struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
111	struct inet_bind_bucket *tb;
112
113	spin_lock(&head->lock);
114	tb = inet_csk(sk)->icsk_bind_hash;
115	__sk_del_bind_node(sk);
116	tb->num_owners--;
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
140	spin_lock(&head->lock);
141	tb = inet_csk(sk)->icsk_bind_hash;
142	if (unlikely(!tb)) {
143		spin_unlock(&head->lock);
144		return -ENOENT;
145	}
146	if (tb->port != port) {
147		/* NOTE: using tproxy and redirecting skbs to a proxy
148		 * on a different listener port breaks the assumption
149		 * that the listener socket's icsk_bind_hash is the same
150		 * as that of the child socket. We have to look up or
151		 * create a new bind bucket for the child here. */
152		inet_bind_bucket_for_each(tb, &head->chain) {
153			if (net_eq(ib_net(tb), sock_net(sk)) &&
154			    tb->port == port)
155				break;
156		}
157		if (!tb) {
158			tb = inet_bind_bucket_create(table->bind_bucket_cachep,
159						     sock_net(sk), head, port);
160			if (!tb) {
161				spin_unlock(&head->lock);
162				return -ENOMEM;
163			}
164		}
165	}
166	inet_bind_hash(child, tb, port);
167	spin_unlock(&head->lock);
168
169	return 0;
170}
171EXPORT_SYMBOL_GPL(__inet_inherit_port);
172
173static inline int compute_score(struct sock *sk, struct net *net,
174				const unsigned short hnum, const __be32 daddr,
175				const int dif)
176{
177	int score = -1;
178	struct inet_sock *inet = inet_sk(sk);
179
180	if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
181			!ipv6_only_sock(sk)) {
182		__be32 rcv_saddr = inet->inet_rcv_saddr;
183		score = sk->sk_family == PF_INET ? 2 : 1;
184		if (rcv_saddr) {
185			if (rcv_saddr != daddr)
186				return -1;
187			score += 4;
188		}
189		if (sk->sk_bound_dev_if) {
190			if (sk->sk_bound_dev_if != dif)
191				return -1;
192			score += 4;
193		}
194		if (sk->sk_incoming_cpu == raw_smp_processor_id())
195			score++;
196	}
197	return score;
198}
199
200/*
201 * Don't inline this cruft. Here are some nice properties to exploit here. The
202 * BSD API does not allow a listening sock to specify the remote port nor the
203 * remote address for the connection. So always assume those are both
204 * wildcarded during the search since they can never be otherwise.
205 */
206
207
208struct sock *__inet_lookup_listener(struct net *net,
209				    struct inet_hashinfo *hashinfo,
210				    struct sk_buff *skb, int doff,
211				    const __be32 saddr, __be16 sport,
212				    const __be32 daddr, const unsigned short hnum,
213				    const int dif)
214{
215	struct sock *sk, *result;
216	struct hlist_nulls_node *node;
217	unsigned int hash = inet_lhashfn(net, hnum);
218	struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
219	int score, hiscore, matches = 0, reuseport = 0;
220	bool select_ok = true;
 
221	u32 phash = 0;
222
223	rcu_read_lock();
224begin:
225	result = NULL;
226	hiscore = 0;
227	sk_nulls_for_each_rcu(sk, node, &ilb->head) {
228		score = compute_score(sk, net, hnum, daddr, dif);
229		if (score > hiscore) {
230			result = sk;
231			hiscore = score;
232			reuseport = sk->sk_reuseport;
233			if (reuseport) {
234				phash = inet_ehashfn(net, daddr, hnum,
235						     saddr, sport);
236				if (select_ok) {
237					struct sock *sk2;
238					sk2 = reuseport_select_sock(sk, phash,
239								    skb, doff);
240					if (sk2) {
241						result = sk2;
242						goto found;
243					}
244				}
245				matches = 1;
246			}
 
 
247		} else if (score == hiscore && reuseport) {
248			matches++;
249			if (reciprocal_scale(phash, matches) == 0)
250				result = sk;
251			phash = next_pseudo_random32(phash);
252		}
253	}
254	/*
255	 * if the nulls value we got at the end of this lookup is
256	 * not the expected one, we must restart lookup.
257	 * We probably met an item that was moved to another chain.
258	 */
259	if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
260		goto begin;
261	if (result) {
262found:
263		if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
264			result = NULL;
265		else if (unlikely(compute_score(result, net, hnum, daddr,
266				  dif) < hiscore)) {
267			sock_put(result);
268			select_ok = false;
269			goto begin;
270		}
271	}
272	rcu_read_unlock();
273	return result;
274}
275EXPORT_SYMBOL_GPL(__inet_lookup_listener);
276
277/* All sockets share common refcount, but have different destructors */
278void sock_gen_put(struct sock *sk)
279{
280	if (!atomic_dec_and_test(&sk->sk_refcnt))
281		return;
282
283	if (sk->sk_state == TCP_TIME_WAIT)
284		inet_twsk_free(inet_twsk(sk));
285	else if (sk->sk_state == TCP_NEW_SYN_RECV)
286		reqsk_free(inet_reqsk(sk));
287	else
288		sk_free(sk);
289}
290EXPORT_SYMBOL_GPL(sock_gen_put);
291
292void sock_edemux(struct sk_buff *skb)
293{
294	sock_gen_put(skb->sk);
295}
296EXPORT_SYMBOL(sock_edemux);
297
298struct sock *__inet_lookup_established(struct net *net,
299				  struct inet_hashinfo *hashinfo,
300				  const __be32 saddr, const __be16 sport,
301				  const __be32 daddr, const u16 hnum,
302				  const int dif)
303{
304	INET_ADDR_COOKIE(acookie, saddr, daddr);
305	const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
306	struct sock *sk;
307	const struct hlist_nulls_node *node;
308	/* Optimize here for direct hit, only listening connections can
309	 * have wildcards anyways.
310	 */
311	unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
312	unsigned int slot = hash & hashinfo->ehash_mask;
313	struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
314
315	rcu_read_lock();
316begin:
317	sk_nulls_for_each_rcu(sk, node, &head->chain) {
318		if (sk->sk_hash != hash)
319			continue;
320		if (likely(INET_MATCH(sk, net, acookie,
321				      saddr, daddr, ports, dif))) {
322			if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
323				goto out;
324			if (unlikely(!INET_MATCH(sk, net, acookie,
325						 saddr, daddr, ports, dif))) {
326				sock_gen_put(sk);
327				goto begin;
328			}
329			goto found;
330		}
331	}
332	/*
333	 * if the nulls value we got at the end of this lookup is
334	 * not the expected one, we must restart lookup.
335	 * We probably met an item that was moved to another chain.
336	 */
337	if (get_nulls_value(node) != slot)
338		goto begin;
339out:
340	sk = NULL;
341found:
342	rcu_read_unlock();
343	return sk;
344}
345EXPORT_SYMBOL_GPL(__inet_lookup_established);
346
347/* called with local bh disabled */
348static int __inet_check_established(struct inet_timewait_death_row *death_row,
349				    struct sock *sk, __u16 lport,
350				    struct inet_timewait_sock **twp)
351{
352	struct inet_hashinfo *hinfo = death_row->hashinfo;
353	struct inet_sock *inet = inet_sk(sk);
354	__be32 daddr = inet->inet_rcv_saddr;
355	__be32 saddr = inet->inet_daddr;
356	int dif = sk->sk_bound_dev_if;
357	INET_ADDR_COOKIE(acookie, saddr, daddr);
358	const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
359	struct net *net = sock_net(sk);
360	unsigned int hash = inet_ehashfn(net, daddr, lport,
361					 saddr, inet->inet_dport);
362	struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
363	spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
364	struct sock *sk2;
365	const struct hlist_nulls_node *node;
366	struct inet_timewait_sock *tw = NULL;
367
368	spin_lock(lock);
369
370	sk_nulls_for_each(sk2, node, &head->chain) {
371		if (sk2->sk_hash != hash)
372			continue;
373
374		if (likely(INET_MATCH(sk2, net, acookie,
375					 saddr, daddr, ports, dif))) {
376			if (sk2->sk_state == TCP_TIME_WAIT) {
377				tw = inet_twsk(sk2);
378				if (twsk_unique(sk, sk2, twp))
379					break;
380			}
381			goto not_unique;
382		}
383	}
384
385	/* Must record num and sport now. Otherwise we will see
386	 * in hash table socket with a funny identity.
387	 */
388	inet->inet_num = lport;
389	inet->inet_sport = htons(lport);
390	sk->sk_hash = hash;
391	WARN_ON(!sk_unhashed(sk));
392	__sk_nulls_add_node_rcu(sk, &head->chain);
393	if (tw) {
394		sk_nulls_del_node_init_rcu((struct sock *)tw);
395		NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
396	}
397	spin_unlock(lock);
398	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
399
400	if (twp) {
401		*twp = tw;
402	} else if (tw) {
403		/* Silly. Should hash-dance instead... */
404		inet_twsk_deschedule_put(tw);
405	}
406	return 0;
407
408not_unique:
409	spin_unlock(lock);
410	return -EADDRNOTAVAIL;
411}
412
413static u32 inet_sk_port_offset(const struct sock *sk)
414{
415	const struct inet_sock *inet = inet_sk(sk);
416
417	return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
418					  inet->inet_daddr,
419					  inet->inet_dport);
420}
421
422/* insert a socket into ehash, and eventually remove another one
423 * (The another one can be a SYN_RECV or TIMEWAIT
424 */
425bool inet_ehash_insert(struct sock *sk, struct sock *osk)
426{
427	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
428	struct hlist_nulls_head *list;
429	struct inet_ehash_bucket *head;
430	spinlock_t *lock;
431	bool ret = true;
432
433	WARN_ON_ONCE(!sk_unhashed(sk));
434
435	sk->sk_hash = sk_ehashfn(sk);
436	head = inet_ehash_bucket(hashinfo, sk->sk_hash);
437	list = &head->chain;
438	lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
439
440	spin_lock(lock);
441	if (osk) {
442		WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
443		ret = sk_nulls_del_node_init_rcu(osk);
444	}
445	if (ret)
446		__sk_nulls_add_node_rcu(sk, list);
447	spin_unlock(lock);
448	return ret;
449}
450
451bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
452{
453	bool ok = inet_ehash_insert(sk, osk);
454
455	if (ok) {
456		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
457	} else {
458		percpu_counter_inc(sk->sk_prot->orphan_count);
459		sk->sk_state = TCP_CLOSE;
460		sock_set_flag(sk, SOCK_DEAD);
461		inet_csk_destroy_sock(sk);
462	}
463	return ok;
464}
465EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
466
467static int inet_reuseport_add_sock(struct sock *sk,
468				   struct inet_listen_hashbucket *ilb,
469				   int (*saddr_same)(const struct sock *sk1,
470						     const struct sock *sk2,
471						     bool match_wildcard))
472{
473	struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
474	struct sock *sk2;
475	struct hlist_nulls_node *node;
476	kuid_t uid = sock_i_uid(sk);
477
478	sk_nulls_for_each_rcu(sk2, node, &ilb->head) {
479		if (sk2 != sk &&
480		    sk2->sk_family == sk->sk_family &&
481		    ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
482		    sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
483		    inet_csk(sk2)->icsk_bind_hash == tb &&
484		    sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
485		    saddr_same(sk, sk2, false))
486			return reuseport_add_sock(sk, sk2);
487	}
488
489	/* Initial allocation may have already happened via setsockopt */
490	if (!rcu_access_pointer(sk->sk_reuseport_cb))
491		return reuseport_alloc(sk);
492	return 0;
493}
494
495int __inet_hash(struct sock *sk, struct sock *osk,
496		 int (*saddr_same)(const struct sock *sk1,
497				   const struct sock *sk2,
498				   bool match_wildcard))
499{
500	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
501	struct inet_listen_hashbucket *ilb;
502	int err = 0;
503
504	if (sk->sk_state != TCP_LISTEN) {
505		inet_ehash_nolisten(sk, osk);
506		return 0;
507	}
508	WARN_ON(!sk_unhashed(sk));
509	ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
510
511	spin_lock(&ilb->lock);
512	if (sk->sk_reuseport) {
513		err = inet_reuseport_add_sock(sk, ilb, saddr_same);
514		if (err)
515			goto unlock;
516	}
517	__sk_nulls_add_node_rcu(sk, &ilb->head);
 
 
 
 
 
518	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
519unlock:
520	spin_unlock(&ilb->lock);
521
522	return err;
523}
524EXPORT_SYMBOL(__inet_hash);
525
526int inet_hash(struct sock *sk)
527{
528	int err = 0;
529
530	if (sk->sk_state != TCP_CLOSE) {
531		local_bh_disable();
532		err = __inet_hash(sk, NULL, ipv4_rcv_saddr_equal);
533		local_bh_enable();
534	}
535
536	return err;
537}
538EXPORT_SYMBOL_GPL(inet_hash);
539
540void inet_unhash(struct sock *sk)
541{
542	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
543	spinlock_t *lock;
 
544	int done;
545
546	if (sk_unhashed(sk))
547		return;
548
549	if (sk->sk_state == TCP_LISTEN)
550		lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
551	else
 
552		lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
553
554	spin_lock_bh(lock);
555	if (rcu_access_pointer(sk->sk_reuseport_cb))
556		reuseport_detach_sock(sk);
557	done = __sk_nulls_del_node_init_rcu(sk);
 
 
 
558	if (done)
559		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
560	spin_unlock_bh(lock);
561}
562EXPORT_SYMBOL_GPL(inet_unhash);
563
564int __inet_hash_connect(struct inet_timewait_death_row *death_row,
565		struct sock *sk, u32 port_offset,
566		int (*check_established)(struct inet_timewait_death_row *,
567			struct sock *, __u16, struct inet_timewait_sock **))
568{
569	struct inet_hashinfo *hinfo = death_row->hashinfo;
570	struct inet_timewait_sock *tw = NULL;
571	struct inet_bind_hashbucket *head;
572	int port = inet_sk(sk)->inet_num;
573	struct net *net = sock_net(sk);
574	struct inet_bind_bucket *tb;
575	u32 remaining, offset;
576	int ret, i, low, high;
577	static u32 hint;
578
579	if (port) {
580		head = &hinfo->bhash[inet_bhashfn(net, port,
581						  hinfo->bhash_size)];
582		tb = inet_csk(sk)->icsk_bind_hash;
583		spin_lock_bh(&head->lock);
584		if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
585			inet_ehash_nolisten(sk, NULL);
586			spin_unlock_bh(&head->lock);
587			return 0;
588		}
589		spin_unlock(&head->lock);
590		/* No definite answer... Walk to established hash table */
591		ret = check_established(death_row, sk, port, NULL);
592		local_bh_enable();
593		return ret;
594	}
595
596	inet_get_local_port_range(net, &low, &high);
597	high++; /* [32768, 60999] -> [32768, 61000[ */
598	remaining = high - low;
599	if (likely(remaining > 1))
600		remaining &= ~1U;
601
602	offset = (hint + port_offset) % remaining;
603	/* In first pass we try ports of @low parity.
604	 * inet_csk_get_port() does the opposite choice.
605	 */
606	offset &= ~1U;
607other_parity_scan:
608	port = low + offset;
609	for (i = 0; i < remaining; i += 2, port += 2) {
610		if (unlikely(port >= high))
611			port -= remaining;
612		if (inet_is_local_reserved_port(net, port))
613			continue;
614		head = &hinfo->bhash[inet_bhashfn(net, port,
615						  hinfo->bhash_size)];
616		spin_lock_bh(&head->lock);
617
618		/* Does not bother with rcv_saddr checks, because
619		 * the established check is already unique enough.
620		 */
621		inet_bind_bucket_for_each(tb, &head->chain) {
622			if (net_eq(ib_net(tb), net) && tb->port == port) {
623				if (tb->fastreuse >= 0 ||
624				    tb->fastreuseport >= 0)
625					goto next_port;
626				WARN_ON(hlist_empty(&tb->owners));
627				if (!check_established(death_row, sk,
628						       port, &tw))
629					goto ok;
630				goto next_port;
631			}
632		}
633
634		tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
635					     net, head, port);
636		if (!tb) {
637			spin_unlock_bh(&head->lock);
638			return -ENOMEM;
639		}
640		tb->fastreuse = -1;
641		tb->fastreuseport = -1;
642		goto ok;
643next_port:
644		spin_unlock_bh(&head->lock);
645		cond_resched();
646	}
647
648	offset++;
649	if ((offset & 1) && remaining > 1)
650		goto other_parity_scan;
651
652	return -EADDRNOTAVAIL;
653
654ok:
655	hint += i + 2;
656
657	/* Head lock still held and bh's disabled */
658	inet_bind_hash(sk, tb, port);
659	if (sk_unhashed(sk)) {
660		inet_sk(sk)->inet_sport = htons(port);
661		inet_ehash_nolisten(sk, (struct sock *)tw);
662	}
663	if (tw)
664		inet_twsk_bind_unhash(tw, hinfo);
665	spin_unlock(&head->lock);
666	if (tw)
667		inet_twsk_deschedule_put(tw);
668	local_bh_enable();
669	return 0;
670}
671
672/*
673 * Bind a port for a connect operation and hash it.
674 */
675int inet_hash_connect(struct inet_timewait_death_row *death_row,
676		      struct sock *sk)
677{
678	u32 port_offset = 0;
679
680	if (!inet_sk(sk)->inet_num)
681		port_offset = inet_sk_port_offset(sk);
682	return __inet_hash_connect(death_row, sk, port_offset,
683				   __inet_check_established);
684}
685EXPORT_SYMBOL_GPL(inet_hash_connect);
686
687void inet_hashinfo_init(struct inet_hashinfo *h)
688{
689	int i;
690
691	for (i = 0; i < INET_LHTABLE_SIZE; i++) {
692		spin_lock_init(&h->listening_hash[i].lock);
693		INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
694				      i + LISTENING_NULLS_BASE);
695		}
696}
697EXPORT_SYMBOL_GPL(inet_hashinfo_init);
698
699int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
700{
701	unsigned int locksz = sizeof(spinlock_t);
702	unsigned int i, nblocks = 1;
703
704	if (locksz != 0) {
705		/* allocate 2 cache lines or at least one spinlock per cpu */
706		nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
707		nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
708
709		/* no more locks than number of hash buckets */
710		nblocks = min(nblocks, hashinfo->ehash_mask + 1);
711
712		hashinfo->ehash_locks =	kmalloc_array(nblocks, locksz,
713						      GFP_KERNEL | __GFP_NOWARN);
714		if (!hashinfo->ehash_locks)
715			hashinfo->ehash_locks = vmalloc(nblocks * locksz);
716
717		if (!hashinfo->ehash_locks)
718			return -ENOMEM;
719
720		for (i = 0; i < nblocks; i++)
721			spin_lock_init(&hashinfo->ehash_locks[i]);
722	}
723	hashinfo->ehash_locks_mask = nblocks - 1;
724	return 0;
725}
726EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
v4.10.11
  1/*
  2 * INET		An implementation of the TCP/IP protocol suite for the LINUX
  3 *		operating system.  INET is implemented using the BSD Socket
  4 *		interface as the means of communication with the user level.
  5 *
  6 *		Generic INET transport hashtables
  7 *
  8 * Authors:	Lotsa people, from code originally in tcp
  9 *
 10 *	This program is free software; you can redistribute it and/or
 11 *      modify it under the terms of the GNU General Public License
 12 *      as published by the Free Software Foundation; either version
 13 *      2 of the License, or (at your option) any later version.
 14 */
 15
 16#include <linux/module.h>
 17#include <linux/random.h>
 18#include <linux/sched.h>
 19#include <linux/slab.h>
 20#include <linux/wait.h>
 21#include <linux/vmalloc.h>
 22
 23#include <net/addrconf.h>
 24#include <net/inet_connection_sock.h>
 25#include <net/inet_hashtables.h>
 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 */
 46u32 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{
 69	struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
 70
 71	if (tb) {
 72		write_pnet(&tb->ib_net, net);
 73		tb->port      = snum;
 74		tb->fastreuse = 0;
 75		tb->fastreuseport = 0;
 76		tb->num_owners = 0;
 77		INIT_HLIST_HEAD(&tb->owners);
 78		hlist_add_head(&tb->node, &head->chain);
 79	}
 80	return tb;
 81}
 82
 83/*
 84 * Caller must hold hashbucket lock for this tb with local BH disabled
 85 */
 86void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
 87{
 88	if (hlist_empty(&tb->owners)) {
 89		__hlist_del(&tb->node);
 90		kmem_cache_free(cachep, tb);
 91	}
 92}
 93
 94void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 95		    const unsigned short snum)
 96{
 97	inet_sk(sk)->inet_num = snum;
 98	sk_add_bind_node(sk, &tb->owners);
 99	tb->num_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	tb->num_owners--;
118	inet_csk(sk)->icsk_bind_hash = NULL;
119	inet_sk(sk)->inet_num = 0;
120	inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
121	spin_unlock(&head->lock);
122}
123
124void inet_put_port(struct sock *sk)
125{
126	local_bh_disable();
127	__inet_put_port(sk);
128	local_bh_enable();
129}
130EXPORT_SYMBOL(inet_put_port);
131
132int __inet_inherit_port(const struct sock *sk, struct sock *child)
133{
134	struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
135	unsigned short port = inet_sk(child)->inet_num;
136	const int bhash = inet_bhashfn(sock_net(sk), port,
137			table->bhash_size);
138	struct inet_bind_hashbucket *head = &table->bhash[bhash];
139	struct inet_bind_bucket *tb;
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		/* NOTE: using tproxy and redirecting skbs to a proxy
149		 * on a different listener port breaks the assumption
150		 * that the listener socket's icsk_bind_hash is the same
151		 * as that of the child socket. We have to look up or
152		 * create a new bind bucket for the child here. */
153		inet_bind_bucket_for_each(tb, &head->chain) {
154			if (net_eq(ib_net(tb), sock_net(sk)) &&
155			    tb->port == port)
156				break;
157		}
158		if (!tb) {
159			tb = inet_bind_bucket_create(table->bind_bucket_cachep,
160						     sock_net(sk), head, port);
161			if (!tb) {
162				spin_unlock(&head->lock);
163				return -ENOMEM;
164			}
165		}
166	}
167	inet_bind_hash(child, tb, port);
168	spin_unlock(&head->lock);
169
170	return 0;
171}
172EXPORT_SYMBOL_GPL(__inet_inherit_port);
173
174static inline int compute_score(struct sock *sk, struct net *net,
175				const unsigned short hnum, const __be32 daddr,
176				const int dif, bool exact_dif)
177{
178	int score = -1;
179	struct inet_sock *inet = inet_sk(sk);
180
181	if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
182			!ipv6_only_sock(sk)) {
183		__be32 rcv_saddr = inet->inet_rcv_saddr;
184		score = sk->sk_family == PF_INET ? 2 : 1;
185		if (rcv_saddr) {
186			if (rcv_saddr != daddr)
187				return -1;
188			score += 4;
189		}
190		if (sk->sk_bound_dev_if || exact_dif) {
191			if (sk->sk_bound_dev_if != dif)
192				return -1;
193			score += 4;
194		}
195		if (sk->sk_incoming_cpu == raw_smp_processor_id())
196			score++;
197	}
198	return score;
199}
200
201/*
202 * Here are some nice properties to exploit here. The BSD API
203 * does not allow a listening sock to specify the remote port nor the
204 * remote address for the connection. So always assume those are both
205 * wildcarded during the search since they can never be otherwise.
206 */
207
208/* called with rcu_read_lock() : No refcount taken on the socket */
209struct sock *__inet_lookup_listener(struct net *net,
210				    struct inet_hashinfo *hashinfo,
211				    struct sk_buff *skb, int doff,
212				    const __be32 saddr, __be16 sport,
213				    const __be32 daddr, const unsigned short hnum,
214				    const int dif)
215{
 
 
216	unsigned int hash = inet_lhashfn(net, hnum);
217	struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
218	int score, hiscore = 0, matches = 0, reuseport = 0;
219	bool exact_dif = inet_exact_dif_match(net, skb);
220	struct sock *sk, *result = NULL;
221	u32 phash = 0;
222
223	sk_for_each_rcu(sk, &ilb->head) {
224		score = compute_score(sk, net, hnum, daddr, dif, exact_dif);
 
 
 
 
225		if (score > hiscore) {
 
 
226			reuseport = sk->sk_reuseport;
227			if (reuseport) {
228				phash = inet_ehashfn(net, daddr, hnum,
229						     saddr, sport);
230				result = reuseport_select_sock(sk, phash,
231							       skb, doff);
232				if (result)
233					return result;
 
 
 
 
 
234				matches = 1;
235			}
236			result = sk;
237			hiscore = score;
238		} else if (score == hiscore && reuseport) {
239			matches++;
240			if (reciprocal_scale(phash, matches) == 0)
241				result = sk;
242			phash = next_pseudo_random32(phash);
243		}
244	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
245	return result;
246}
247EXPORT_SYMBOL_GPL(__inet_lookup_listener);
248
249/* All sockets share common refcount, but have different destructors */
250void sock_gen_put(struct sock *sk)
251{
252	if (!atomic_dec_and_test(&sk->sk_refcnt))
253		return;
254
255	if (sk->sk_state == TCP_TIME_WAIT)
256		inet_twsk_free(inet_twsk(sk));
257	else if (sk->sk_state == TCP_NEW_SYN_RECV)
258		reqsk_free(inet_reqsk(sk));
259	else
260		sk_free(sk);
261}
262EXPORT_SYMBOL_GPL(sock_gen_put);
263
264void sock_edemux(struct sk_buff *skb)
265{
266	sock_gen_put(skb->sk);
267}
268EXPORT_SYMBOL(sock_edemux);
269
270struct sock *__inet_lookup_established(struct net *net,
271				  struct inet_hashinfo *hashinfo,
272				  const __be32 saddr, const __be16 sport,
273				  const __be32 daddr, const u16 hnum,
274				  const int dif)
275{
276	INET_ADDR_COOKIE(acookie, saddr, daddr);
277	const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
278	struct sock *sk;
279	const struct hlist_nulls_node *node;
280	/* Optimize here for direct hit, only listening connections can
281	 * have wildcards anyways.
282	 */
283	unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
284	unsigned int slot = hash & hashinfo->ehash_mask;
285	struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
286
 
287begin:
288	sk_nulls_for_each_rcu(sk, node, &head->chain) {
289		if (sk->sk_hash != hash)
290			continue;
291		if (likely(INET_MATCH(sk, net, acookie,
292				      saddr, daddr, ports, dif))) {
293			if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
294				goto out;
295			if (unlikely(!INET_MATCH(sk, net, acookie,
296						 saddr, daddr, ports, dif))) {
297				sock_gen_put(sk);
298				goto begin;
299			}
300			goto found;
301		}
302	}
303	/*
304	 * if the nulls value we got at the end of this lookup is
305	 * not the expected one, we must restart lookup.
306	 * We probably met an item that was moved to another chain.
307	 */
308	if (get_nulls_value(node) != slot)
309		goto begin;
310out:
311	sk = NULL;
312found:
 
313	return sk;
314}
315EXPORT_SYMBOL_GPL(__inet_lookup_established);
316
317/* called with local bh disabled */
318static int __inet_check_established(struct inet_timewait_death_row *death_row,
319				    struct sock *sk, __u16 lport,
320				    struct inet_timewait_sock **twp)
321{
322	struct inet_hashinfo *hinfo = death_row->hashinfo;
323	struct inet_sock *inet = inet_sk(sk);
324	__be32 daddr = inet->inet_rcv_saddr;
325	__be32 saddr = inet->inet_daddr;
326	int dif = sk->sk_bound_dev_if;
327	INET_ADDR_COOKIE(acookie, saddr, daddr);
328	const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
329	struct net *net = sock_net(sk);
330	unsigned int hash = inet_ehashfn(net, daddr, lport,
331					 saddr, inet->inet_dport);
332	struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
333	spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
334	struct sock *sk2;
335	const struct hlist_nulls_node *node;
336	struct inet_timewait_sock *tw = NULL;
337
338	spin_lock(lock);
339
340	sk_nulls_for_each(sk2, node, &head->chain) {
341		if (sk2->sk_hash != hash)
342			continue;
343
344		if (likely(INET_MATCH(sk2, net, acookie,
345					 saddr, daddr, ports, dif))) {
346			if (sk2->sk_state == TCP_TIME_WAIT) {
347				tw = inet_twsk(sk2);
348				if (twsk_unique(sk, sk2, twp))
349					break;
350			}
351			goto not_unique;
352		}
353	}
354
355	/* Must record num and sport now. Otherwise we will see
356	 * in hash table socket with a funny identity.
357	 */
358	inet->inet_num = lport;
359	inet->inet_sport = htons(lport);
360	sk->sk_hash = hash;
361	WARN_ON(!sk_unhashed(sk));
362	__sk_nulls_add_node_rcu(sk, &head->chain);
363	if (tw) {
364		sk_nulls_del_node_init_rcu((struct sock *)tw);
365		__NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
366	}
367	spin_unlock(lock);
368	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
369
370	if (twp) {
371		*twp = tw;
372	} else if (tw) {
373		/* Silly. Should hash-dance instead... */
374		inet_twsk_deschedule_put(tw);
375	}
376	return 0;
377
378not_unique:
379	spin_unlock(lock);
380	return -EADDRNOTAVAIL;
381}
382
383static u32 inet_sk_port_offset(const struct sock *sk)
384{
385	const struct inet_sock *inet = inet_sk(sk);
386
387	return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
388					  inet->inet_daddr,
389					  inet->inet_dport);
390}
391
392/* insert a socket into ehash, and eventually remove another one
393 * (The another one can be a SYN_RECV or TIMEWAIT
394 */
395bool inet_ehash_insert(struct sock *sk, struct sock *osk)
396{
397	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
398	struct hlist_nulls_head *list;
399	struct inet_ehash_bucket *head;
400	spinlock_t *lock;
401	bool ret = true;
402
403	WARN_ON_ONCE(!sk_unhashed(sk));
404
405	sk->sk_hash = sk_ehashfn(sk);
406	head = inet_ehash_bucket(hashinfo, sk->sk_hash);
407	list = &head->chain;
408	lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
409
410	spin_lock(lock);
411	if (osk) {
412		WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
413		ret = sk_nulls_del_node_init_rcu(osk);
414	}
415	if (ret)
416		__sk_nulls_add_node_rcu(sk, list);
417	spin_unlock(lock);
418	return ret;
419}
420
421bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
422{
423	bool ok = inet_ehash_insert(sk, osk);
424
425	if (ok) {
426		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
427	} else {
428		percpu_counter_inc(sk->sk_prot->orphan_count);
429		sk->sk_state = TCP_CLOSE;
430		sock_set_flag(sk, SOCK_DEAD);
431		inet_csk_destroy_sock(sk);
432	}
433	return ok;
434}
435EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
436
437static int inet_reuseport_add_sock(struct sock *sk,
438				   struct inet_listen_hashbucket *ilb,
439				   int (*saddr_same)(const struct sock *sk1,
440						     const struct sock *sk2,
441						     bool match_wildcard))
442{
443	struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
444	struct sock *sk2;
 
445	kuid_t uid = sock_i_uid(sk);
446
447	sk_for_each_rcu(sk2, &ilb->head) {
448		if (sk2 != sk &&
449		    sk2->sk_family == sk->sk_family &&
450		    ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
451		    sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
452		    inet_csk(sk2)->icsk_bind_hash == tb &&
453		    sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
454		    saddr_same(sk, sk2, false))
455			return reuseport_add_sock(sk, sk2);
456	}
457
458	/* Initial allocation may have already happened via setsockopt */
459	if (!rcu_access_pointer(sk->sk_reuseport_cb))
460		return reuseport_alloc(sk);
461	return 0;
462}
463
464int __inet_hash(struct sock *sk, struct sock *osk,
465		 int (*saddr_same)(const struct sock *sk1,
466				   const struct sock *sk2,
467				   bool match_wildcard))
468{
469	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
470	struct inet_listen_hashbucket *ilb;
471	int err = 0;
472
473	if (sk->sk_state != TCP_LISTEN) {
474		inet_ehash_nolisten(sk, osk);
475		return 0;
476	}
477	WARN_ON(!sk_unhashed(sk));
478	ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
479
480	spin_lock(&ilb->lock);
481	if (sk->sk_reuseport) {
482		err = inet_reuseport_add_sock(sk, ilb, saddr_same);
483		if (err)
484			goto unlock;
485	}
486	if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
487		sk->sk_family == AF_INET6)
488		hlist_add_tail_rcu(&sk->sk_node, &ilb->head);
489	else
490		hlist_add_head_rcu(&sk->sk_node, &ilb->head);
491	sock_set_flag(sk, SOCK_RCU_FREE);
492	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
493unlock:
494	spin_unlock(&ilb->lock);
495
496	return err;
497}
498EXPORT_SYMBOL(__inet_hash);
499
500int inet_hash(struct sock *sk)
501{
502	int err = 0;
503
504	if (sk->sk_state != TCP_CLOSE) {
505		local_bh_disable();
506		err = __inet_hash(sk, NULL, ipv4_rcv_saddr_equal);
507		local_bh_enable();
508	}
509
510	return err;
511}
512EXPORT_SYMBOL_GPL(inet_hash);
513
514void inet_unhash(struct sock *sk)
515{
516	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
517	spinlock_t *lock;
518	bool listener = false;
519	int done;
520
521	if (sk_unhashed(sk))
522		return;
523
524	if (sk->sk_state == TCP_LISTEN) {
525		lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
526		listener = true;
527	} else {
528		lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
529	}
530	spin_lock_bh(lock);
531	if (rcu_access_pointer(sk->sk_reuseport_cb))
532		reuseport_detach_sock(sk);
533	if (listener)
534		done = __sk_del_node_init(sk);
535	else
536		done = __sk_nulls_del_node_init_rcu(sk);
537	if (done)
538		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
539	spin_unlock_bh(lock);
540}
541EXPORT_SYMBOL_GPL(inet_unhash);
542
543int __inet_hash_connect(struct inet_timewait_death_row *death_row,
544		struct sock *sk, u32 port_offset,
545		int (*check_established)(struct inet_timewait_death_row *,
546			struct sock *, __u16, struct inet_timewait_sock **))
547{
548	struct inet_hashinfo *hinfo = death_row->hashinfo;
549	struct inet_timewait_sock *tw = NULL;
550	struct inet_bind_hashbucket *head;
551	int port = inet_sk(sk)->inet_num;
552	struct net *net = sock_net(sk);
553	struct inet_bind_bucket *tb;
554	u32 remaining, offset;
555	int ret, i, low, high;
556	static u32 hint;
557
558	if (port) {
559		head = &hinfo->bhash[inet_bhashfn(net, port,
560						  hinfo->bhash_size)];
561		tb = inet_csk(sk)->icsk_bind_hash;
562		spin_lock_bh(&head->lock);
563		if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
564			inet_ehash_nolisten(sk, NULL);
565			spin_unlock_bh(&head->lock);
566			return 0;
567		}
568		spin_unlock(&head->lock);
569		/* No definite answer... Walk to established hash table */
570		ret = check_established(death_row, sk, port, NULL);
571		local_bh_enable();
572		return ret;
573	}
574
575	inet_get_local_port_range(net, &low, &high);
576	high++; /* [32768, 60999] -> [32768, 61000[ */
577	remaining = high - low;
578	if (likely(remaining > 1))
579		remaining &= ~1U;
580
581	offset = (hint + port_offset) % remaining;
582	/* In first pass we try ports of @low parity.
583	 * inet_csk_get_port() does the opposite choice.
584	 */
585	offset &= ~1U;
586other_parity_scan:
587	port = low + offset;
588	for (i = 0; i < remaining; i += 2, port += 2) {
589		if (unlikely(port >= high))
590			port -= remaining;
591		if (inet_is_local_reserved_port(net, port))
592			continue;
593		head = &hinfo->bhash[inet_bhashfn(net, port,
594						  hinfo->bhash_size)];
595		spin_lock_bh(&head->lock);
596
597		/* Does not bother with rcv_saddr checks, because
598		 * the established check is already unique enough.
599		 */
600		inet_bind_bucket_for_each(tb, &head->chain) {
601			if (net_eq(ib_net(tb), net) && tb->port == port) {
602				if (tb->fastreuse >= 0 ||
603				    tb->fastreuseport >= 0)
604					goto next_port;
605				WARN_ON(hlist_empty(&tb->owners));
606				if (!check_established(death_row, sk,
607						       port, &tw))
608					goto ok;
609				goto next_port;
610			}
611		}
612
613		tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
614					     net, head, port);
615		if (!tb) {
616			spin_unlock_bh(&head->lock);
617			return -ENOMEM;
618		}
619		tb->fastreuse = -1;
620		tb->fastreuseport = -1;
621		goto ok;
622next_port:
623		spin_unlock_bh(&head->lock);
624		cond_resched();
625	}
626
627	offset++;
628	if ((offset & 1) && remaining > 1)
629		goto other_parity_scan;
630
631	return -EADDRNOTAVAIL;
632
633ok:
634	hint += i + 2;
635
636	/* Head lock still held and bh's disabled */
637	inet_bind_hash(sk, tb, port);
638	if (sk_unhashed(sk)) {
639		inet_sk(sk)->inet_sport = htons(port);
640		inet_ehash_nolisten(sk, (struct sock *)tw);
641	}
642	if (tw)
643		inet_twsk_bind_unhash(tw, hinfo);
644	spin_unlock(&head->lock);
645	if (tw)
646		inet_twsk_deschedule_put(tw);
647	local_bh_enable();
648	return 0;
649}
650
651/*
652 * Bind a port for a connect operation and hash it.
653 */
654int inet_hash_connect(struct inet_timewait_death_row *death_row,
655		      struct sock *sk)
656{
657	u32 port_offset = 0;
658
659	if (!inet_sk(sk)->inet_num)
660		port_offset = inet_sk_port_offset(sk);
661	return __inet_hash_connect(death_row, sk, port_offset,
662				   __inet_check_established);
663}
664EXPORT_SYMBOL_GPL(inet_hash_connect);
665
666void inet_hashinfo_init(struct inet_hashinfo *h)
667{
668	int i;
669
670	for (i = 0; i < INET_LHTABLE_SIZE; i++) {
671		spin_lock_init(&h->listening_hash[i].lock);
672		INIT_HLIST_HEAD(&h->listening_hash[i].head);
673	}
 
674}
675EXPORT_SYMBOL_GPL(inet_hashinfo_init);
676
677int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
678{
679	unsigned int locksz = sizeof(spinlock_t);
680	unsigned int i, nblocks = 1;
681
682	if (locksz != 0) {
683		/* allocate 2 cache lines or at least one spinlock per cpu */
684		nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
685		nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
686
687		/* no more locks than number of hash buckets */
688		nblocks = min(nblocks, hashinfo->ehash_mask + 1);
689
690		hashinfo->ehash_locks =	kmalloc_array(nblocks, locksz,
691						      GFP_KERNEL | __GFP_NOWARN);
692		if (!hashinfo->ehash_locks)
693			hashinfo->ehash_locks = vmalloc(nblocks * locksz);
694
695		if (!hashinfo->ehash_locks)
696			return -ENOMEM;
697
698		for (i = 0; i < nblocks; i++)
699			spin_lock_init(&hashinfo->ehash_locks[i]);
700	}
701	hashinfo->ehash_locks_mask = nblocks - 1;
702	return 0;
703}
704EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);