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