Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound-2.6
[linux-2.6] / net / ipv4 / inet_timewait_sock.c
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. */
17 static 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         rwlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
24
25         write_lock(lock);
26         if (hlist_unhashed(&tw->tw_node)) {
27                 write_unlock(lock);
28                 return;
29         }
30         __hlist_del(&tw->tw_node);
31         sk_node_init(&tw->tw_node);
32         write_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
52 void 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 }
66 EXPORT_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  */
73 void __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         rwlock_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         write_lock(lock);
94
95         /* Step 2: Remove SK from established hash. */
96         if (__sk_del_node_init(sk))
97                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
98
99         /* Step 3: Hash TW into TIMEWAIT chain. */
100         inet_twsk_add_node(tw, &ehead->twchain);
101         atomic_inc(&tw->tw_refcnt);
102
103         write_unlock(lock);
104 }
105
106 EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
107
108 struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk, const int state)
109 {
110         struct inet_timewait_sock *tw =
111                 kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
112                                  GFP_ATOMIC);
113         if (tw != NULL) {
114                 const struct inet_sock *inet = inet_sk(sk);
115
116                 /* Give us an identity. */
117                 tw->tw_daddr        = inet->daddr;
118                 tw->tw_rcv_saddr    = inet->rcv_saddr;
119                 tw->tw_bound_dev_if = sk->sk_bound_dev_if;
120                 tw->tw_num          = inet->num;
121                 tw->tw_state        = TCP_TIME_WAIT;
122                 tw->tw_substate     = state;
123                 tw->tw_sport        = inet->sport;
124                 tw->tw_dport        = inet->dport;
125                 tw->tw_family       = sk->sk_family;
126                 tw->tw_reuse        = sk->sk_reuse;
127                 tw->tw_hash         = sk->sk_hash;
128                 tw->tw_ipv6only     = 0;
129                 tw->tw_prot         = sk->sk_prot_creator;
130                 twsk_net_set(tw, hold_net(sock_net(sk)));
131                 atomic_set(&tw->tw_refcnt, 1);
132                 inet_twsk_dead_node_init(tw);
133                 __module_get(tw->tw_prot->owner);
134         }
135
136         return tw;
137 }
138
139 EXPORT_SYMBOL_GPL(inet_twsk_alloc);
140
141 /* Returns non-zero if quota exceeded.  */
142 static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
143                                     const int slot)
144 {
145         struct inet_timewait_sock *tw;
146         struct hlist_node *node;
147         unsigned int killed;
148         int ret;
149
150         /* NOTE: compare this to previous version where lock
151          * was released after detaching chain. It was racy,
152          * because tw buckets are scheduled in not serialized context
153          * in 2.3 (with netfilter), and with softnet it is common, because
154          * soft irqs are not sequenced.
155          */
156         killed = 0;
157         ret = 0;
158 rescan:
159         inet_twsk_for_each_inmate(tw, node, &twdr->cells[slot]) {
160                 __inet_twsk_del_dead_node(tw);
161                 spin_unlock(&twdr->death_lock);
162                 __inet_twsk_kill(tw, twdr->hashinfo);
163 #ifdef CONFIG_NET_NS
164                 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITED);
165 #endif
166                 inet_twsk_put(tw);
167                 killed++;
168                 spin_lock(&twdr->death_lock);
169                 if (killed > INET_TWDR_TWKILL_QUOTA) {
170                         ret = 1;
171                         break;
172                 }
173
174                 /* While we dropped twdr->death_lock, another cpu may have
175                  * killed off the next TW bucket in the list, therefore
176                  * do a fresh re-read of the hlist head node with the
177                  * lock reacquired.  We still use the hlist traversal
178                  * macro in order to get the prefetches.
179                  */
180                 goto rescan;
181         }
182
183         twdr->tw_count -= killed;
184 #ifndef CONFIG_NET_NS
185         NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITED, killed);
186 #endif
187         return ret;
188 }
189
190 void inet_twdr_hangman(unsigned long data)
191 {
192         struct inet_timewait_death_row *twdr;
193         int unsigned need_timer;
194
195         twdr = (struct inet_timewait_death_row *)data;
196         spin_lock(&twdr->death_lock);
197
198         if (twdr->tw_count == 0)
199                 goto out;
200
201         need_timer = 0;
202         if (inet_twdr_do_twkill_work(twdr, twdr->slot)) {
203                 twdr->thread_slots |= (1 << twdr->slot);
204                 schedule_work(&twdr->twkill_work);
205                 need_timer = 1;
206         } else {
207                 /* We purged the entire slot, anything left?  */
208                 if (twdr->tw_count)
209                         need_timer = 1;
210         }
211         twdr->slot = ((twdr->slot + 1) & (INET_TWDR_TWKILL_SLOTS - 1));
212         if (need_timer)
213                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
214 out:
215         spin_unlock(&twdr->death_lock);
216 }
217
218 EXPORT_SYMBOL_GPL(inet_twdr_hangman);
219
220 void inet_twdr_twkill_work(struct work_struct *work)
221 {
222         struct inet_timewait_death_row *twdr =
223                 container_of(work, struct inet_timewait_death_row, twkill_work);
224         int i;
225
226         BUILD_BUG_ON((INET_TWDR_TWKILL_SLOTS - 1) >
227                         (sizeof(twdr->thread_slots) * 8));
228
229         while (twdr->thread_slots) {
230                 spin_lock_bh(&twdr->death_lock);
231                 for (i = 0; i < INET_TWDR_TWKILL_SLOTS; i++) {
232                         if (!(twdr->thread_slots & (1 << i)))
233                                 continue;
234
235                         while (inet_twdr_do_twkill_work(twdr, i) != 0) {
236                                 if (need_resched()) {
237                                         spin_unlock_bh(&twdr->death_lock);
238                                         schedule();
239                                         spin_lock_bh(&twdr->death_lock);
240                                 }
241                         }
242
243                         twdr->thread_slots &= ~(1 << i);
244                 }
245                 spin_unlock_bh(&twdr->death_lock);
246         }
247 }
248
249 EXPORT_SYMBOL_GPL(inet_twdr_twkill_work);
250
251 /* These are always called from BH context.  See callers in
252  * tcp_input.c to verify this.
253  */
254
255 /* This is for handling early-kills of TIME_WAIT sockets. */
256 void inet_twsk_deschedule(struct inet_timewait_sock *tw,
257                           struct inet_timewait_death_row *twdr)
258 {
259         spin_lock(&twdr->death_lock);
260         if (inet_twsk_del_dead_node(tw)) {
261                 inet_twsk_put(tw);
262                 if (--twdr->tw_count == 0)
263                         del_timer(&twdr->tw_timer);
264         }
265         spin_unlock(&twdr->death_lock);
266         __inet_twsk_kill(tw, twdr->hashinfo);
267 }
268
269 EXPORT_SYMBOL(inet_twsk_deschedule);
270
271 void inet_twsk_schedule(struct inet_timewait_sock *tw,
272                        struct inet_timewait_death_row *twdr,
273                        const int timeo, const int timewait_len)
274 {
275         struct hlist_head *list;
276         int slot;
277
278         /* timeout := RTO * 3.5
279          *
280          * 3.5 = 1+2+0.5 to wait for two retransmits.
281          *
282          * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
283          * our ACK acking that FIN can be lost. If N subsequent retransmitted
284          * FINs (or previous seqments) are lost (probability of such event
285          * is p^(N+1), where p is probability to lose single packet and
286          * time to detect the loss is about RTO*(2^N - 1) with exponential
287          * backoff). Normal timewait length is calculated so, that we
288          * waited at least for one retransmitted FIN (maximal RTO is 120sec).
289          * [ BTW Linux. following BSD, violates this requirement waiting
290          *   only for 60sec, we should wait at least for 240 secs.
291          *   Well, 240 consumes too much of resources 8)
292          * ]
293          * This interval is not reduced to catch old duplicate and
294          * responces to our wandering segments living for two MSLs.
295          * However, if we use PAWS to detect
296          * old duplicates, we can reduce the interval to bounds required
297          * by RTO, rather than MSL. So, if peer understands PAWS, we
298          * kill tw bucket after 3.5*RTO (it is important that this number
299          * is greater than TS tick!) and detect old duplicates with help
300          * of PAWS.
301          */
302         slot = (timeo + (1 << INET_TWDR_RECYCLE_TICK) - 1) >> INET_TWDR_RECYCLE_TICK;
303
304         spin_lock(&twdr->death_lock);
305
306         /* Unlink it, if it was scheduled */
307         if (inet_twsk_del_dead_node(tw))
308                 twdr->tw_count--;
309         else
310                 atomic_inc(&tw->tw_refcnt);
311
312         if (slot >= INET_TWDR_RECYCLE_SLOTS) {
313                 /* Schedule to slow timer */
314                 if (timeo >= timewait_len) {
315                         slot = INET_TWDR_TWKILL_SLOTS - 1;
316                 } else {
317                         slot = DIV_ROUND_UP(timeo, twdr->period);
318                         if (slot >= INET_TWDR_TWKILL_SLOTS)
319                                 slot = INET_TWDR_TWKILL_SLOTS - 1;
320                 }
321                 tw->tw_ttd = jiffies + timeo;
322                 slot = (twdr->slot + slot) & (INET_TWDR_TWKILL_SLOTS - 1);
323                 list = &twdr->cells[slot];
324         } else {
325                 tw->tw_ttd = jiffies + (slot << INET_TWDR_RECYCLE_TICK);
326
327                 if (twdr->twcal_hand < 0) {
328                         twdr->twcal_hand = 0;
329                         twdr->twcal_jiffie = jiffies;
330                         twdr->twcal_timer.expires = twdr->twcal_jiffie +
331                                               (slot << INET_TWDR_RECYCLE_TICK);
332                         add_timer(&twdr->twcal_timer);
333                 } else {
334                         if (time_after(twdr->twcal_timer.expires,
335                                        jiffies + (slot << INET_TWDR_RECYCLE_TICK)))
336                                 mod_timer(&twdr->twcal_timer,
337                                           jiffies + (slot << INET_TWDR_RECYCLE_TICK));
338                         slot = (twdr->twcal_hand + slot) & (INET_TWDR_RECYCLE_SLOTS - 1);
339                 }
340                 list = &twdr->twcal_row[slot];
341         }
342
343         hlist_add_head(&tw->tw_death_node, list);
344
345         if (twdr->tw_count++ == 0)
346                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
347         spin_unlock(&twdr->death_lock);
348 }
349
350 EXPORT_SYMBOL_GPL(inet_twsk_schedule);
351
352 void inet_twdr_twcal_tick(unsigned long data)
353 {
354         struct inet_timewait_death_row *twdr;
355         int n, slot;
356         unsigned long j;
357         unsigned long now = jiffies;
358         int killed = 0;
359         int adv = 0;
360
361         twdr = (struct inet_timewait_death_row *)data;
362
363         spin_lock(&twdr->death_lock);
364         if (twdr->twcal_hand < 0)
365                 goto out;
366
367         slot = twdr->twcal_hand;
368         j = twdr->twcal_jiffie;
369
370         for (n = 0; n < INET_TWDR_RECYCLE_SLOTS; n++) {
371                 if (time_before_eq(j, now)) {
372                         struct hlist_node *node, *safe;
373                         struct inet_timewait_sock *tw;
374
375                         inet_twsk_for_each_inmate_safe(tw, node, safe,
376                                                        &twdr->twcal_row[slot]) {
377                                 __inet_twsk_del_dead_node(tw);
378                                 __inet_twsk_kill(tw, twdr->hashinfo);
379 #ifdef CONFIG_NET_NS
380                                 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITKILLED);
381 #endif
382                                 inet_twsk_put(tw);
383                                 killed++;
384                         }
385                 } else {
386                         if (!adv) {
387                                 adv = 1;
388                                 twdr->twcal_jiffie = j;
389                                 twdr->twcal_hand = slot;
390                         }
391
392                         if (!hlist_empty(&twdr->twcal_row[slot])) {
393                                 mod_timer(&twdr->twcal_timer, j);
394                                 goto out;
395                         }
396                 }
397                 j += 1 << INET_TWDR_RECYCLE_TICK;
398                 slot = (slot + 1) & (INET_TWDR_RECYCLE_SLOTS - 1);
399         }
400         twdr->twcal_hand = -1;
401
402 out:
403         if ((twdr->tw_count -= killed) == 0)
404                 del_timer(&twdr->tw_timer);
405 #ifndef CONFIG_NET_NS
406         NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITKILLED, killed);
407 #endif
408         spin_unlock(&twdr->death_lock);
409 }
410
411 EXPORT_SYMBOL_GPL(inet_twdr_twcal_tick);
412
413 void inet_twsk_purge(struct net *net, struct inet_hashinfo *hashinfo,
414                      struct inet_timewait_death_row *twdr, int family)
415 {
416         struct inet_timewait_sock *tw;
417         struct sock *sk;
418         struct hlist_node *node;
419         int h;
420
421         local_bh_disable();
422         for (h = 0; h < (hashinfo->ehash_size); h++) {
423                 struct inet_ehash_bucket *head =
424                         inet_ehash_bucket(hashinfo, h);
425                 rwlock_t *lock = inet_ehash_lockp(hashinfo, h);
426 restart:
427                 write_lock(lock);
428                 sk_for_each(sk, node, &head->twchain) {
429
430                         tw = inet_twsk(sk);
431                         if (!net_eq(twsk_net(tw), net) ||
432                             tw->tw_family != family)
433                                 continue;
434
435                         atomic_inc(&tw->tw_refcnt);
436                         write_unlock(lock);
437                         inet_twsk_deschedule(tw, twdr);
438                         inet_twsk_put(tw);
439
440                         goto restart;
441                 }
442                 write_unlock(lock);
443         }
444         local_bh_enable();
445 }
446 EXPORT_SYMBOL_GPL(inet_twsk_purge);