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