net: Batch inet_twsk_purge
[linux-2.6.git] / 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 <linux/kmemcheck.h>
13 #include <net/inet_hashtables.h>
14 #include <net/inet_timewait_sock.h>
15 #include <net/ip.h>
16
17 /* Must be called with locally disabled BHs. */
18 static void __inet_twsk_kill(struct inet_timewait_sock *tw,
19                              struct inet_hashinfo *hashinfo)
20 {
21         struct inet_bind_hashbucket *bhead;
22         struct inet_bind_bucket *tb;
23         /* Unlink from established hashes. */
24         spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
25
26         spin_lock(lock);
27         if (hlist_nulls_unhashed(&tw->tw_node)) {
28                 spin_unlock(lock);
29                 return;
30         }
31         hlist_nulls_del_rcu(&tw->tw_node);
32         sk_nulls_node_init(&tw->tw_node);
33         spin_unlock(lock);
34
35         /* Disassociate with bind bucket. */
36         bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), tw->tw_num,
37                         hashinfo->bhash_size)];
38         spin_lock(&bhead->lock);
39         tb = tw->tw_tb;
40         __hlist_del(&tw->tw_bind_node);
41         tw->tw_tb = NULL;
42         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
43         spin_unlock(&bhead->lock);
44 #ifdef SOCK_REFCNT_DEBUG
45         if (atomic_read(&tw->tw_refcnt) != 1) {
46                 printk(KERN_DEBUG "%s timewait_sock %p refcnt=%d\n",
47                        tw->tw_prot->name, tw, atomic_read(&tw->tw_refcnt));
48         }
49 #endif
50         inet_twsk_put(tw);
51 }
52
53 static noinline void inet_twsk_free(struct inet_timewait_sock *tw)
54 {
55         struct module *owner = tw->tw_prot->owner;
56         twsk_destructor((struct sock *)tw);
57 #ifdef SOCK_REFCNT_DEBUG
58         pr_debug("%s timewait_sock %p released\n", tw->tw_prot->name, tw);
59 #endif
60         release_net(twsk_net(tw));
61         kmem_cache_free(tw->tw_prot->twsk_prot->twsk_slab, tw);
62         module_put(owner);
63 }
64
65 void inet_twsk_put(struct inet_timewait_sock *tw)
66 {
67         if (atomic_dec_and_test(&tw->tw_refcnt))
68                 inet_twsk_free(tw);
69 }
70 EXPORT_SYMBOL_GPL(inet_twsk_put);
71
72 /*
73  * Enter the time wait state. This is called with locally disabled BH.
74  * Essentially we whip up a timewait bucket, copy the relevant info into it
75  * from the SK, and mess with hash chains and list linkage.
76  */
77 void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
78                            struct inet_hashinfo *hashinfo)
79 {
80         const struct inet_sock *inet = inet_sk(sk);
81         const struct inet_connection_sock *icsk = inet_csk(sk);
82         struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, sk->sk_hash);
83         spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
84         struct inet_bind_hashbucket *bhead;
85         /* Step 1: Put TW into bind hash. Original socket stays there too.
86            Note, that any socket with inet->num != 0 MUST be bound in
87            binding cache, even if it is closed.
88          */
89         bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), inet->inet_num,
90                         hashinfo->bhash_size)];
91         spin_lock(&bhead->lock);
92         tw->tw_tb = icsk->icsk_bind_hash;
93         WARN_ON(!icsk->icsk_bind_hash);
94         inet_twsk_add_bind_node(tw, &tw->tw_tb->owners);
95         spin_unlock(&bhead->lock);
96
97         spin_lock(lock);
98
99         /*
100          * Step 2: Hash TW into TIMEWAIT chain.
101          * Should be done before removing sk from established chain
102          * because readers are lockless and search established first.
103          */
104         atomic_inc(&tw->tw_refcnt);
105         inet_twsk_add_node_rcu(tw, &ehead->twchain);
106
107         /* Step 3: Remove SK from established hash. */
108         if (__sk_nulls_del_node_init_rcu(sk))
109                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
110
111         spin_unlock(lock);
112 }
113
114 EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
115
116 struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk, const int state)
117 {
118         struct inet_timewait_sock *tw =
119                 kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
120                                  GFP_ATOMIC);
121         if (tw != NULL) {
122                 const struct inet_sock *inet = inet_sk(sk);
123
124                 kmemcheck_annotate_bitfield(tw, flags);
125
126                 /* Give us an identity. */
127                 tw->tw_daddr        = inet->inet_daddr;
128                 tw->tw_rcv_saddr    = inet->inet_rcv_saddr;
129                 tw->tw_bound_dev_if = sk->sk_bound_dev_if;
130                 tw->tw_num          = inet->inet_num;
131                 tw->tw_state        = TCP_TIME_WAIT;
132                 tw->tw_substate     = state;
133                 tw->tw_sport        = inet->inet_sport;
134                 tw->tw_dport        = inet->inet_dport;
135                 tw->tw_family       = sk->sk_family;
136                 tw->tw_reuse        = sk->sk_reuse;
137                 tw->tw_hash         = sk->sk_hash;
138                 tw->tw_ipv6only     = 0;
139                 tw->tw_transparent  = inet->transparent;
140                 tw->tw_prot         = sk->sk_prot_creator;
141                 twsk_net_set(tw, hold_net(sock_net(sk)));
142                 atomic_set(&tw->tw_refcnt, 1);
143                 inet_twsk_dead_node_init(tw);
144                 __module_get(tw->tw_prot->owner);
145         }
146
147         return tw;
148 }
149
150 EXPORT_SYMBOL_GPL(inet_twsk_alloc);
151
152 /* Returns non-zero if quota exceeded.  */
153 static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
154                                     const int slot)
155 {
156         struct inet_timewait_sock *tw;
157         struct hlist_node *node;
158         unsigned int killed;
159         int ret;
160
161         /* NOTE: compare this to previous version where lock
162          * was released after detaching chain. It was racy,
163          * because tw buckets are scheduled in not serialized context
164          * in 2.3 (with netfilter), and with softnet it is common, because
165          * soft irqs are not sequenced.
166          */
167         killed = 0;
168         ret = 0;
169 rescan:
170         inet_twsk_for_each_inmate(tw, node, &twdr->cells[slot]) {
171                 __inet_twsk_del_dead_node(tw);
172                 spin_unlock(&twdr->death_lock);
173                 __inet_twsk_kill(tw, twdr->hashinfo);
174 #ifdef CONFIG_NET_NS
175                 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITED);
176 #endif
177                 inet_twsk_put(tw);
178                 killed++;
179                 spin_lock(&twdr->death_lock);
180                 if (killed > INET_TWDR_TWKILL_QUOTA) {
181                         ret = 1;
182                         break;
183                 }
184
185                 /* While we dropped twdr->death_lock, another cpu may have
186                  * killed off the next TW bucket in the list, therefore
187                  * do a fresh re-read of the hlist head node with the
188                  * lock reacquired.  We still use the hlist traversal
189                  * macro in order to get the prefetches.
190                  */
191                 goto rescan;
192         }
193
194         twdr->tw_count -= killed;
195 #ifndef CONFIG_NET_NS
196         NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITED, killed);
197 #endif
198         return ret;
199 }
200
201 void inet_twdr_hangman(unsigned long data)
202 {
203         struct inet_timewait_death_row *twdr;
204         int unsigned need_timer;
205
206         twdr = (struct inet_timewait_death_row *)data;
207         spin_lock(&twdr->death_lock);
208
209         if (twdr->tw_count == 0)
210                 goto out;
211
212         need_timer = 0;
213         if (inet_twdr_do_twkill_work(twdr, twdr->slot)) {
214                 twdr->thread_slots |= (1 << twdr->slot);
215                 schedule_work(&twdr->twkill_work);
216                 need_timer = 1;
217         } else {
218                 /* We purged the entire slot, anything left?  */
219                 if (twdr->tw_count)
220                         need_timer = 1;
221                 twdr->slot = ((twdr->slot + 1) & (INET_TWDR_TWKILL_SLOTS - 1));
222         }
223         if (need_timer)
224                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
225 out:
226         spin_unlock(&twdr->death_lock);
227 }
228
229 EXPORT_SYMBOL_GPL(inet_twdr_hangman);
230
231 void inet_twdr_twkill_work(struct work_struct *work)
232 {
233         struct inet_timewait_death_row *twdr =
234                 container_of(work, struct inet_timewait_death_row, twkill_work);
235         int i;
236
237         BUILD_BUG_ON((INET_TWDR_TWKILL_SLOTS - 1) >
238                         (sizeof(twdr->thread_slots) * 8));
239
240         while (twdr->thread_slots) {
241                 spin_lock_bh(&twdr->death_lock);
242                 for (i = 0; i < INET_TWDR_TWKILL_SLOTS; i++) {
243                         if (!(twdr->thread_slots & (1 << i)))
244                                 continue;
245
246                         while (inet_twdr_do_twkill_work(twdr, i) != 0) {
247                                 if (need_resched()) {
248                                         spin_unlock_bh(&twdr->death_lock);
249                                         schedule();
250                                         spin_lock_bh(&twdr->death_lock);
251                                 }
252                         }
253
254                         twdr->thread_slots &= ~(1 << i);
255                 }
256                 spin_unlock_bh(&twdr->death_lock);
257         }
258 }
259
260 EXPORT_SYMBOL_GPL(inet_twdr_twkill_work);
261
262 /* These are always called from BH context.  See callers in
263  * tcp_input.c to verify this.
264  */
265
266 /* This is for handling early-kills of TIME_WAIT sockets. */
267 void inet_twsk_deschedule(struct inet_timewait_sock *tw,
268                           struct inet_timewait_death_row *twdr)
269 {
270         spin_lock(&twdr->death_lock);
271         if (inet_twsk_del_dead_node(tw)) {
272                 inet_twsk_put(tw);
273                 if (--twdr->tw_count == 0)
274                         del_timer(&twdr->tw_timer);
275         }
276         spin_unlock(&twdr->death_lock);
277         __inet_twsk_kill(tw, twdr->hashinfo);
278 }
279
280 EXPORT_SYMBOL(inet_twsk_deschedule);
281
282 void inet_twsk_schedule(struct inet_timewait_sock *tw,
283                        struct inet_timewait_death_row *twdr,
284                        const int timeo, const int timewait_len)
285 {
286         struct hlist_head *list;
287         int slot;
288
289         /* timeout := RTO * 3.5
290          *
291          * 3.5 = 1+2+0.5 to wait for two retransmits.
292          *
293          * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
294          * our ACK acking that FIN can be lost. If N subsequent retransmitted
295          * FINs (or previous seqments) are lost (probability of such event
296          * is p^(N+1), where p is probability to lose single packet and
297          * time to detect the loss is about RTO*(2^N - 1) with exponential
298          * backoff). Normal timewait length is calculated so, that we
299          * waited at least for one retransmitted FIN (maximal RTO is 120sec).
300          * [ BTW Linux. following BSD, violates this requirement waiting
301          *   only for 60sec, we should wait at least for 240 secs.
302          *   Well, 240 consumes too much of resources 8)
303          * ]
304          * This interval is not reduced to catch old duplicate and
305          * responces to our wandering segments living for two MSLs.
306          * However, if we use PAWS to detect
307          * old duplicates, we can reduce the interval to bounds required
308          * by RTO, rather than MSL. So, if peer understands PAWS, we
309          * kill tw bucket after 3.5*RTO (it is important that this number
310          * is greater than TS tick!) and detect old duplicates with help
311          * of PAWS.
312          */
313         slot = (timeo + (1 << INET_TWDR_RECYCLE_TICK) - 1) >> INET_TWDR_RECYCLE_TICK;
314
315         spin_lock(&twdr->death_lock);
316
317         /* Unlink it, if it was scheduled */
318         if (inet_twsk_del_dead_node(tw))
319                 twdr->tw_count--;
320         else
321                 atomic_inc(&tw->tw_refcnt);
322
323         if (slot >= INET_TWDR_RECYCLE_SLOTS) {
324                 /* Schedule to slow timer */
325                 if (timeo >= timewait_len) {
326                         slot = INET_TWDR_TWKILL_SLOTS - 1;
327                 } else {
328                         slot = DIV_ROUND_UP(timeo, twdr->period);
329                         if (slot >= INET_TWDR_TWKILL_SLOTS)
330                                 slot = INET_TWDR_TWKILL_SLOTS - 1;
331                 }
332                 tw->tw_ttd = jiffies + timeo;
333                 slot = (twdr->slot + slot) & (INET_TWDR_TWKILL_SLOTS - 1);
334                 list = &twdr->cells[slot];
335         } else {
336                 tw->tw_ttd = jiffies + (slot << INET_TWDR_RECYCLE_TICK);
337
338                 if (twdr->twcal_hand < 0) {
339                         twdr->twcal_hand = 0;
340                         twdr->twcal_jiffie = jiffies;
341                         twdr->twcal_timer.expires = twdr->twcal_jiffie +
342                                               (slot << INET_TWDR_RECYCLE_TICK);
343                         add_timer(&twdr->twcal_timer);
344                 } else {
345                         if (time_after(twdr->twcal_timer.expires,
346                                        jiffies + (slot << INET_TWDR_RECYCLE_TICK)))
347                                 mod_timer(&twdr->twcal_timer,
348                                           jiffies + (slot << INET_TWDR_RECYCLE_TICK));
349                         slot = (twdr->twcal_hand + slot) & (INET_TWDR_RECYCLE_SLOTS - 1);
350                 }
351                 list = &twdr->twcal_row[slot];
352         }
353
354         hlist_add_head(&tw->tw_death_node, list);
355
356         if (twdr->tw_count++ == 0)
357                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
358         spin_unlock(&twdr->death_lock);
359 }
360
361 EXPORT_SYMBOL_GPL(inet_twsk_schedule);
362
363 void inet_twdr_twcal_tick(unsigned long data)
364 {
365         struct inet_timewait_death_row *twdr;
366         int n, slot;
367         unsigned long j;
368         unsigned long now = jiffies;
369         int killed = 0;
370         int adv = 0;
371
372         twdr = (struct inet_timewait_death_row *)data;
373
374         spin_lock(&twdr->death_lock);
375         if (twdr->twcal_hand < 0)
376                 goto out;
377
378         slot = twdr->twcal_hand;
379         j = twdr->twcal_jiffie;
380
381         for (n = 0; n < INET_TWDR_RECYCLE_SLOTS; n++) {
382                 if (time_before_eq(j, now)) {
383                         struct hlist_node *node, *safe;
384                         struct inet_timewait_sock *tw;
385
386                         inet_twsk_for_each_inmate_safe(tw, node, safe,
387                                                        &twdr->twcal_row[slot]) {
388                                 __inet_twsk_del_dead_node(tw);
389                                 __inet_twsk_kill(tw, twdr->hashinfo);
390 #ifdef CONFIG_NET_NS
391                                 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITKILLED);
392 #endif
393                                 inet_twsk_put(tw);
394                                 killed++;
395                         }
396                 } else {
397                         if (!adv) {
398                                 adv = 1;
399                                 twdr->twcal_jiffie = j;
400                                 twdr->twcal_hand = slot;
401                         }
402
403                         if (!hlist_empty(&twdr->twcal_row[slot])) {
404                                 mod_timer(&twdr->twcal_timer, j);
405                                 goto out;
406                         }
407                 }
408                 j += 1 << INET_TWDR_RECYCLE_TICK;
409                 slot = (slot + 1) & (INET_TWDR_RECYCLE_SLOTS - 1);
410         }
411         twdr->twcal_hand = -1;
412
413 out:
414         if ((twdr->tw_count -= killed) == 0)
415                 del_timer(&twdr->tw_timer);
416 #ifndef CONFIG_NET_NS
417         NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITKILLED, killed);
418 #endif
419         spin_unlock(&twdr->death_lock);
420 }
421
422 EXPORT_SYMBOL_GPL(inet_twdr_twcal_tick);
423
424 void inet_twsk_purge(struct inet_hashinfo *hashinfo,
425                      struct inet_timewait_death_row *twdr, int family)
426 {
427         struct inet_timewait_sock *tw;
428         struct sock *sk;
429         struct hlist_nulls_node *node;
430         unsigned int slot;
431
432         for (slot = 0; slot <= hashinfo->ehash_mask; slot++) {
433                 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
434 restart_rcu:
435                 rcu_read_lock();
436 restart:
437                 sk_nulls_for_each_rcu(sk, node, &head->twchain) {
438                         tw = inet_twsk(sk);
439                         if ((tw->tw_family != family) ||
440                                 atomic_read(&twsk_net(tw)->count))
441                                 continue;
442
443                         if (unlikely(!atomic_inc_not_zero(&tw->tw_refcnt)))
444                                 continue;
445
446                         if (unlikely((tw->tw_family != family) ||
447                                      atomic_read(&twsk_net(tw)->count))) {
448                                 inet_twsk_put(tw);
449                                 goto restart;
450                         }
451
452                         rcu_read_unlock();
453                         inet_twsk_deschedule(tw, twdr);
454                         inet_twsk_put(tw);
455                         goto restart_rcu;
456                 }
457                 /* If the nulls value we got at the end of this lookup is
458                  * not the expected one, we must restart lookup.
459                  * We probably met an item that was moved to another chain.
460                  */
461                 if (get_nulls_value(node) != slot)
462                         goto restart;
463                 rcu_read_unlock();
464         }
465 }
466 EXPORT_SYMBOL_GPL(inet_twsk_purge);