inet_timewait_sock.c revision 9db66bdcc83749affe61c61eb8ff3cf08f42afec 
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 TIME_WAIT sockets functions
7 *
8 *		From code orinally in TCP
9 */
10
11#include <linux/kernel.h>
12#include <net/inet_hashtables.h>
13#include <net/inet_timewait_sock.h>
14#include <net/ip.h>
15
16/* Must be called with locally disabled BHs. */
17static void __inet_twsk_kill(struct inet_timewait_sock *tw,
18			     struct inet_hashinfo *hashinfo)
19{
20	struct inet_bind_hashbucket *bhead;
21	struct inet_bind_bucket *tb;
22	/* Unlink from established hashes. */
23	spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
24
25	spin_lock(lock);
26	if (hlist_nulls_unhashed(&tw->tw_node)) {
27		spin_unlock(lock);
28		return;
29	}
30	hlist_nulls_del_rcu(&tw->tw_node);
31	sk_nulls_node_init(&tw->tw_node);
32	spin_unlock(lock);
33
34	/* Disassociate with bind bucket. */
35	bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), tw->tw_num,
36			hashinfo->bhash_size)];
37	spin_lock(&bhead->lock);
38	tb = tw->tw_tb;
39	__hlist_del(&tw->tw_bind_node);
40	tw->tw_tb = NULL;
41	inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
42	spin_unlock(&bhead->lock);
43#ifdef SOCK_REFCNT_DEBUG
44	if (atomic_read(&tw->tw_refcnt) != 1) {
45		printk(KERN_DEBUG "%s timewait_sock %p refcnt=%d\n",
46		       tw->tw_prot->name, tw, atomic_read(&tw->tw_refcnt));
47	}
48#endif
49	inet_twsk_put(tw);
50}
51
52void inet_twsk_put(struct inet_timewait_sock *tw)
53{
54	if (atomic_dec_and_test(&tw->tw_refcnt)) {
55		struct module *owner = tw->tw_prot->owner;
56		twsk_destructor((struct sock *)tw);
57#ifdef SOCK_REFCNT_DEBUG
58		printk(KERN_DEBUG "%s timewait_sock %p released\n",
59		       tw->tw_prot->name, tw);
60#endif
61		release_net(twsk_net(tw));
62		kmem_cache_free(tw->tw_prot->twsk_prot->twsk_slab, tw);
63		module_put(owner);
64	}
65}
66EXPORT_SYMBOL_GPL(inet_twsk_put);
67
68/*
69 * Enter the time wait state. This is called with locally disabled BH.
70 * Essentially we whip up a timewait bucket, copy the relevant info into it
71 * from the SK, and mess with hash chains and list linkage.
72 */
73void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
74			   struct inet_hashinfo *hashinfo)
75{
76	const struct inet_sock *inet = inet_sk(sk);
77	const struct inet_connection_sock *icsk = inet_csk(sk);
78	struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, sk->sk_hash);
79	spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
80	struct inet_bind_hashbucket *bhead;
81	/* Step 1: Put TW into bind hash. Original socket stays there too.
82	   Note, that any socket with inet->num != 0 MUST be bound in
83	   binding cache, even if it is closed.
84	 */
85	bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), inet->num,
86			hashinfo->bhash_size)];
87	spin_lock(&bhead->lock);
88	tw->tw_tb = icsk->icsk_bind_hash;
89	WARN_ON(!icsk->icsk_bind_hash);
90	inet_twsk_add_bind_node(tw, &tw->tw_tb->owners);
91	spin_unlock(&bhead->lock);
92
93	spin_lock(lock);
94
95	/*
96	 * Step 2: Hash TW into TIMEWAIT chain.
97	 * Should be done before removing sk from established chain
98	 * because readers are lockless and search established first.
99	 */
100	atomic_inc(&tw->tw_refcnt);
101	inet_twsk_add_node_rcu(tw, &ehead->twchain);
102
103	/* Step 3: Remove SK from established hash. */
104	if (__sk_nulls_del_node_init_rcu(sk))
105		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
106
107	spin_unlock(lock);
108}
109
110EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
111
112struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk, const int state)
113{
114	struct inet_timewait_sock *tw =
115		kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
116				 GFP_ATOMIC);
117	if (tw != NULL) {
118		const struct inet_sock *inet = inet_sk(sk);
119
120		/* Give us an identity. */
121		tw->tw_daddr	    = inet->daddr;
122		tw->tw_rcv_saddr    = inet->rcv_saddr;
123		tw->tw_bound_dev_if = sk->sk_bound_dev_if;
124		tw->tw_num	    = inet->num;
125		tw->tw_state	    = TCP_TIME_WAIT;
126		tw->tw_substate	    = state;
127		tw->tw_sport	    = inet->sport;
128		tw->tw_dport	    = inet->dport;
129		tw->tw_family	    = sk->sk_family;
130		tw->tw_reuse	    = sk->sk_reuse;
131		tw->tw_hash	    = sk->sk_hash;
132		tw->tw_ipv6only	    = 0;
133		tw->tw_transparent  = inet->transparent;
134		tw->tw_prot	    = sk->sk_prot_creator;
135		twsk_net_set(tw, hold_net(sock_net(sk)));
136		atomic_set(&tw->tw_refcnt, 1);
137		inet_twsk_dead_node_init(tw);
138		__module_get(tw->tw_prot->owner);
139	}
140
141	return tw;
142}
143
144EXPORT_SYMBOL_GPL(inet_twsk_alloc);
145
146/* Returns non-zero if quota exceeded.  */
147static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
148				    const int slot)
149{
150	struct inet_timewait_sock *tw;
151	struct hlist_node *node;
152	unsigned int killed;
153	int ret;
154
155	/* NOTE: compare this to previous version where lock
156	 * was released after detaching chain. It was racy,
157	 * because tw buckets are scheduled in not serialized context
158	 * in 2.3 (with netfilter), and with softnet it is common, because
159	 * soft irqs are not sequenced.
160	 */
161	killed = 0;
162	ret = 0;
163rescan:
164	inet_twsk_for_each_inmate(tw, node, &twdr->cells[slot]) {
165		__inet_twsk_del_dead_node(tw);
166		spin_unlock(&twdr->death_lock);
167		__inet_twsk_kill(tw, twdr->hashinfo);
168#ifdef CONFIG_NET_NS
169		NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITED);
170#endif
171		inet_twsk_put(tw);
172		killed++;
173		spin_lock(&twdr->death_lock);
174		if (killed > INET_TWDR_TWKILL_QUOTA) {
175			ret = 1;
176			break;
177		}
178
179		/* While we dropped twdr->death_lock, another cpu may have
180		 * killed off the next TW bucket in the list, therefore
181		 * do a fresh re-read of the hlist head node with the
182		 * lock reacquired.  We still use the hlist traversal
183		 * macro in order to get the prefetches.
184		 */
185		goto rescan;
186	}
187
188	twdr->tw_count -= killed;
189#ifndef CONFIG_NET_NS
190	NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITED, killed);
191#endif
192	return ret;
193}
194
195void inet_twdr_hangman(unsigned long data)
196{
197	struct inet_timewait_death_row *twdr;
198	int unsigned need_timer;
199
200	twdr = (struct inet_timewait_death_row *)data;
201	spin_lock(&twdr->death_lock);
202
203	if (twdr->tw_count == 0)
204		goto out;
205
206	need_timer = 0;
207	if (inet_twdr_do_twkill_work(twdr, twdr->slot)) {
208		twdr->thread_slots |= (1 << twdr->slot);
209		schedule_work(&twdr->twkill_work);
210		need_timer = 1;
211	} else {
212		/* We purged the entire slot, anything left?  */
213		if (twdr->tw_count)
214			need_timer = 1;
215	}
216	twdr->slot = ((twdr->slot + 1) & (INET_TWDR_TWKILL_SLOTS - 1));
217	if (need_timer)
218		mod_timer(&twdr->tw_timer, jiffies + twdr->period);
219out:
220	spin_unlock(&twdr->death_lock);
221}
222
223EXPORT_SYMBOL_GPL(inet_twdr_hangman);
224
225void inet_twdr_twkill_work(struct work_struct *work)
226{
227	struct inet_timewait_death_row *twdr =
228		container_of(work, struct inet_timewait_death_row, twkill_work);
229	int i;
230
231	BUILD_BUG_ON((INET_TWDR_TWKILL_SLOTS - 1) >
232			(sizeof(twdr->thread_slots) * 8));
233
234	while (twdr->thread_slots) {
235		spin_lock_bh(&twdr->death_lock);
236		for (i = 0; i < INET_TWDR_TWKILL_SLOTS; i++) {
237			if (!(twdr->thread_slots & (1 << i)))
238				continue;
239
240			while (inet_twdr_do_twkill_work(twdr, i) != 0) {
241				if (need_resched()) {
242					spin_unlock_bh(&twdr->death_lock);
243					schedule();
244					spin_lock_bh(&twdr->death_lock);
245				}
246			}
247
248			twdr->thread_slots &= ~(1 << i);
249		}
250		spin_unlock_bh(&twdr->death_lock);
251	}
252}
253
254EXPORT_SYMBOL_GPL(inet_twdr_twkill_work);
255
256/* These are always called from BH context.  See callers in
257 * tcp_input.c to verify this.
258 */
259
260/* This is for handling early-kills of TIME_WAIT sockets. */
261void inet_twsk_deschedule(struct inet_timewait_sock *tw,
262			  struct inet_timewait_death_row *twdr)
263{
264	spin_lock(&twdr->death_lock);
265	if (inet_twsk_del_dead_node(tw)) {
266		inet_twsk_put(tw);
267		if (--twdr->tw_count == 0)
268			del_timer(&twdr->tw_timer);
269	}
270	spin_unlock(&twdr->death_lock);
271	__inet_twsk_kill(tw, twdr->hashinfo);
272}
273
274EXPORT_SYMBOL(inet_twsk_deschedule);
275
276void inet_twsk_schedule(struct inet_timewait_sock *tw,
277		       struct inet_timewait_death_row *twdr,
278		       const int timeo, const int timewait_len)
279{
280	struct hlist_head *list;
281	int slot;
282
283	/* timeout := RTO * 3.5
284	 *
285	 * 3.5 = 1+2+0.5 to wait for two retransmits.
286	 *
287	 * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
288	 * our ACK acking that FIN can be lost. If N subsequent retransmitted
289	 * FINs (or previous seqments) are lost (probability of such event
290	 * is p^(N+1), where p is probability to lose single packet and
291	 * time to detect the loss is about RTO*(2^N - 1) with exponential
292	 * backoff). Normal timewait length is calculated so, that we
293	 * waited at least for one retransmitted FIN (maximal RTO is 120sec).
294	 * [ BTW Linux. following BSD, violates this requirement waiting
295	 *   only for 60sec, we should wait at least for 240 secs.
296	 *   Well, 240 consumes too much of resources 8)
297	 * ]
298	 * This interval is not reduced to catch old duplicate and
299	 * responces to our wandering segments living for two MSLs.
300	 * However, if we use PAWS to detect
301	 * old duplicates, we can reduce the interval to bounds required
302	 * by RTO, rather than MSL. So, if peer understands PAWS, we
303	 * kill tw bucket after 3.5*RTO (it is important that this number
304	 * is greater than TS tick!) and detect old duplicates with help
305	 * of PAWS.
306	 */
307	slot = (timeo + (1 << INET_TWDR_RECYCLE_TICK) - 1) >> INET_TWDR_RECYCLE_TICK;
308
309	spin_lock(&twdr->death_lock);
310
311	/* Unlink it, if it was scheduled */
312	if (inet_twsk_del_dead_node(tw))
313		twdr->tw_count--;
314	else
315		atomic_inc(&tw->tw_refcnt);
316
317	if (slot >= INET_TWDR_RECYCLE_SLOTS) {
318		/* Schedule to slow timer */
319		if (timeo >= timewait_len) {
320			slot = INET_TWDR_TWKILL_SLOTS - 1;
321		} else {
322			slot = DIV_ROUND_UP(timeo, twdr->period);
323			if (slot >= INET_TWDR_TWKILL_SLOTS)
324				slot = INET_TWDR_TWKILL_SLOTS - 1;
325		}
326		tw->tw_ttd = jiffies + timeo;
327		slot = (twdr->slot + slot) & (INET_TWDR_TWKILL_SLOTS - 1);
328		list = &twdr->cells[slot];
329	} else {
330		tw->tw_ttd = jiffies + (slot << INET_TWDR_RECYCLE_TICK);
331
332		if (twdr->twcal_hand < 0) {
333			twdr->twcal_hand = 0;
334			twdr->twcal_jiffie = jiffies;
335			twdr->twcal_timer.expires = twdr->twcal_jiffie +
336					      (slot << INET_TWDR_RECYCLE_TICK);
337			add_timer(&twdr->twcal_timer);
338		} else {
339			if (time_after(twdr->twcal_timer.expires,
340				       jiffies + (slot << INET_TWDR_RECYCLE_TICK)))
341				mod_timer(&twdr->twcal_timer,
342					  jiffies + (slot << INET_TWDR_RECYCLE_TICK));
343			slot = (twdr->twcal_hand + slot) & (INET_TWDR_RECYCLE_SLOTS - 1);
344		}
345		list = &twdr->twcal_row[slot];
346	}
347
348	hlist_add_head(&tw->tw_death_node, list);
349
350	if (twdr->tw_count++ == 0)
351		mod_timer(&twdr->tw_timer, jiffies + twdr->period);
352	spin_unlock(&twdr->death_lock);
353}
354
355EXPORT_SYMBOL_GPL(inet_twsk_schedule);
356
357void inet_twdr_twcal_tick(unsigned long data)
358{
359	struct inet_timewait_death_row *twdr;
360	int n, slot;
361	unsigned long j;
362	unsigned long now = jiffies;
363	int killed = 0;
364	int adv = 0;
365
366	twdr = (struct inet_timewait_death_row *)data;
367
368	spin_lock(&twdr->death_lock);
369	if (twdr->twcal_hand < 0)
370		goto out;
371
372	slot = twdr->twcal_hand;
373	j = twdr->twcal_jiffie;
374
375	for (n = 0; n < INET_TWDR_RECYCLE_SLOTS; n++) {
376		if (time_before_eq(j, now)) {
377			struct hlist_node *node, *safe;
378			struct inet_timewait_sock *tw;
379
380			inet_twsk_for_each_inmate_safe(tw, node, safe,
381						       &twdr->twcal_row[slot]) {
382				__inet_twsk_del_dead_node(tw);
383				__inet_twsk_kill(tw, twdr->hashinfo);
384#ifdef CONFIG_NET_NS
385				NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITKILLED);
386#endif
387				inet_twsk_put(tw);
388				killed++;
389			}
390		} else {
391			if (!adv) {
392				adv = 1;
393				twdr->twcal_jiffie = j;
394				twdr->twcal_hand = slot;
395			}
396
397			if (!hlist_empty(&twdr->twcal_row[slot])) {
398				mod_timer(&twdr->twcal_timer, j);
399				goto out;
400			}
401		}
402		j += 1 << INET_TWDR_RECYCLE_TICK;
403		slot = (slot + 1) & (INET_TWDR_RECYCLE_SLOTS - 1);
404	}
405	twdr->twcal_hand = -1;
406
407out:
408	if ((twdr->tw_count -= killed) == 0)
409		del_timer(&twdr->tw_timer);
410#ifndef CONFIG_NET_NS
411	NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITKILLED, killed);
412#endif
413	spin_unlock(&twdr->death_lock);
414}
415
416EXPORT_SYMBOL_GPL(inet_twdr_twcal_tick);
417
418void inet_twsk_purge(struct net *net, struct inet_hashinfo *hashinfo,
419		     struct inet_timewait_death_row *twdr, int family)
420{
421	struct inet_timewait_sock *tw;
422	struct sock *sk;
423	struct hlist_nulls_node *node;
424	int h;
425
426	local_bh_disable();
427	for (h = 0; h < (hashinfo->ehash_size); h++) {
428		struct inet_ehash_bucket *head =
429			inet_ehash_bucket(hashinfo, h);
430		spinlock_t *lock = inet_ehash_lockp(hashinfo, h);
431restart:
432		spin_lock(lock);
433		sk_nulls_for_each(sk, node, &head->twchain) {
434
435			tw = inet_twsk(sk);
436			if (!net_eq(twsk_net(tw), net) ||
437			    tw->tw_family != family)
438				continue;
439
440			atomic_inc(&tw->tw_refcnt);
441			spin_unlock(lock);
442			inet_twsk_deschedule(tw, twdr);
443			inet_twsk_put(tw);
444
445			goto restart;
446		}
447		spin_unlock(lock);
448	}
449	local_bh_enable();
450}
451EXPORT_SYMBOL_GPL(inet_twsk_purge);
452