[INET]: speedup inet (tcp/dccp) lookups
[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 = kmem_cache_alloc(sk->sk_prot_creator->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         }
115
116         return tw;
117 }
118
119 EXPORT_SYMBOL_GPL(inet_twsk_alloc);
120
121 /* Returns non-zero if quota exceeded.  */
122 static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
123                                     const int slot)
124 {
125         struct inet_timewait_sock *tw;
126         struct hlist_node *node;
127         unsigned int killed;
128         int ret;
129
130         /* NOTE: compare this to previous version where lock
131          * was released after detaching chain. It was racy,
132          * because tw buckets are scheduled in not serialized context
133          * in 2.3 (with netfilter), and with softnet it is common, because
134          * soft irqs are not sequenced.
135          */
136         killed = 0;
137         ret = 0;
138 rescan:
139         inet_twsk_for_each_inmate(tw, node, &twdr->cells[slot]) {
140                 __inet_twsk_del_dead_node(tw);
141                 spin_unlock(&twdr->death_lock);
142                 __inet_twsk_kill(tw, twdr->hashinfo);
143                 inet_twsk_put(tw);
144                 killed++;
145                 spin_lock(&twdr->death_lock);
146                 if (killed > INET_TWDR_TWKILL_QUOTA) {
147                         ret = 1;
148                         break;
149                 }
150
151                 /* While we dropped twdr->death_lock, another cpu may have
152                  * killed off the next TW bucket in the list, therefore
153                  * do a fresh re-read of the hlist head node with the
154                  * lock reacquired.  We still use the hlist traversal
155                  * macro in order to get the prefetches.
156                  */
157                 goto rescan;
158         }
159
160         twdr->tw_count -= killed;
161         NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITED, killed);
162
163         return ret;
164 }
165
166 void inet_twdr_hangman(unsigned long data)
167 {
168         struct inet_timewait_death_row *twdr;
169         int unsigned need_timer;
170
171         twdr = (struct inet_timewait_death_row *)data;
172         spin_lock(&twdr->death_lock);
173
174         if (twdr->tw_count == 0)
175                 goto out;
176
177         need_timer = 0;
178         if (inet_twdr_do_twkill_work(twdr, twdr->slot)) {
179                 twdr->thread_slots |= (1 << twdr->slot);
180                 mb();
181                 schedule_work(&twdr->twkill_work);
182                 need_timer = 1;
183         } else {
184                 /* We purged the entire slot, anything left?  */
185                 if (twdr->tw_count)
186                         need_timer = 1;
187         }
188         twdr->slot = ((twdr->slot + 1) & (INET_TWDR_TWKILL_SLOTS - 1));
189         if (need_timer)
190                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
191 out:
192         spin_unlock(&twdr->death_lock);
193 }
194
195 EXPORT_SYMBOL_GPL(inet_twdr_hangman);
196
197 extern void twkill_slots_invalid(void);
198
199 void inet_twdr_twkill_work(void *data)
200 {
201         struct inet_timewait_death_row *twdr = data;
202         int i;
203
204         if ((INET_TWDR_TWKILL_SLOTS - 1) > (sizeof(twdr->thread_slots) * 8))
205                 twkill_slots_invalid();
206
207         while (twdr->thread_slots) {
208                 spin_lock_bh(&twdr->death_lock);
209                 for (i = 0; i < INET_TWDR_TWKILL_SLOTS; i++) {
210                         if (!(twdr->thread_slots & (1 << i)))
211                                 continue;
212
213                         while (inet_twdr_do_twkill_work(twdr, i) != 0) {
214                                 if (need_resched()) {
215                                         spin_unlock_bh(&twdr->death_lock);
216                                         schedule();
217                                         spin_lock_bh(&twdr->death_lock);
218                                 }
219                         }
220
221                         twdr->thread_slots &= ~(1 << i);
222                 }
223                 spin_unlock_bh(&twdr->death_lock);
224         }
225 }
226
227 EXPORT_SYMBOL_GPL(inet_twdr_twkill_work);
228
229 /* These are always called from BH context.  See callers in
230  * tcp_input.c to verify this.
231  */
232
233 /* This is for handling early-kills of TIME_WAIT sockets. */
234 void inet_twsk_deschedule(struct inet_timewait_sock *tw,
235                           struct inet_timewait_death_row *twdr)
236 {
237         spin_lock(&twdr->death_lock);
238         if (inet_twsk_del_dead_node(tw)) {
239                 inet_twsk_put(tw);
240                 if (--twdr->tw_count == 0)
241                         del_timer(&twdr->tw_timer);
242         }
243         spin_unlock(&twdr->death_lock);
244         __inet_twsk_kill(tw, twdr->hashinfo);
245 }
246
247 EXPORT_SYMBOL(inet_twsk_deschedule);
248
249 void inet_twsk_schedule(struct inet_timewait_sock *tw,
250                        struct inet_timewait_death_row *twdr,
251                        const int timeo, const int timewait_len)
252 {
253         struct hlist_head *list;
254         int slot;
255
256         /* timeout := RTO * 3.5
257          *
258          * 3.5 = 1+2+0.5 to wait for two retransmits.
259          *
260          * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
261          * our ACK acking that FIN can be lost. If N subsequent retransmitted
262          * FINs (or previous seqments) are lost (probability of such event
263          * is p^(N+1), where p is probability to lose single packet and
264          * time to detect the loss is about RTO*(2^N - 1) with exponential
265          * backoff). Normal timewait length is calculated so, that we
266          * waited at least for one retransmitted FIN (maximal RTO is 120sec).
267          * [ BTW Linux. following BSD, violates this requirement waiting
268          *   only for 60sec, we should wait at least for 240 secs.
269          *   Well, 240 consumes too much of resources 8)
270          * ]
271          * This interval is not reduced to catch old duplicate and
272          * responces to our wandering segments living for two MSLs.
273          * However, if we use PAWS to detect
274          * old duplicates, we can reduce the interval to bounds required
275          * by RTO, rather than MSL. So, if peer understands PAWS, we
276          * kill tw bucket after 3.5*RTO (it is important that this number
277          * is greater than TS tick!) and detect old duplicates with help
278          * of PAWS.
279          */
280         slot = (timeo + (1 << INET_TWDR_RECYCLE_TICK) - 1) >> INET_TWDR_RECYCLE_TICK;
281
282         spin_lock(&twdr->death_lock);
283
284         /* Unlink it, if it was scheduled */
285         if (inet_twsk_del_dead_node(tw))
286                 twdr->tw_count--;
287         else
288                 atomic_inc(&tw->tw_refcnt);
289
290         if (slot >= INET_TWDR_RECYCLE_SLOTS) {
291                 /* Schedule to slow timer */
292                 if (timeo >= timewait_len) {
293                         slot = INET_TWDR_TWKILL_SLOTS - 1;
294                 } else {
295                         slot = (timeo + twdr->period - 1) / twdr->period;
296                         if (slot >= INET_TWDR_TWKILL_SLOTS)
297                                 slot = INET_TWDR_TWKILL_SLOTS - 1;
298                 }
299                 tw->tw_ttd = jiffies + timeo;
300                 slot = (twdr->slot + slot) & (INET_TWDR_TWKILL_SLOTS - 1);
301                 list = &twdr->cells[slot];
302         } else {
303                 tw->tw_ttd = jiffies + (slot << INET_TWDR_RECYCLE_TICK);
304
305                 if (twdr->twcal_hand < 0) {
306                         twdr->twcal_hand = 0;
307                         twdr->twcal_jiffie = jiffies;
308                         twdr->twcal_timer.expires = twdr->twcal_jiffie +
309                                               (slot << INET_TWDR_RECYCLE_TICK);
310                         add_timer(&twdr->twcal_timer);
311                 } else {
312                         if (time_after(twdr->twcal_timer.expires,
313                                        jiffies + (slot << INET_TWDR_RECYCLE_TICK)))
314                                 mod_timer(&twdr->twcal_timer,
315                                           jiffies + (slot << INET_TWDR_RECYCLE_TICK));
316                         slot = (twdr->twcal_hand + slot) & (INET_TWDR_RECYCLE_SLOTS - 1);
317                 }
318                 list = &twdr->twcal_row[slot];
319         }
320
321         hlist_add_head(&tw->tw_death_node, list);
322
323         if (twdr->tw_count++ == 0)
324                 mod_timer(&twdr->tw_timer, jiffies + twdr->period);
325         spin_unlock(&twdr->death_lock);
326 }
327
328 EXPORT_SYMBOL_GPL(inet_twsk_schedule);
329
330 void inet_twdr_twcal_tick(unsigned long data)
331 {
332         struct inet_timewait_death_row *twdr;
333         int n, slot;
334         unsigned long j;
335         unsigned long now = jiffies;
336         int killed = 0;
337         int adv = 0;
338
339         twdr = (struct inet_timewait_death_row *)data;
340
341         spin_lock(&twdr->death_lock);
342         if (twdr->twcal_hand < 0)
343                 goto out;
344
345         slot = twdr->twcal_hand;
346         j = twdr->twcal_jiffie;
347
348         for (n = 0; n < INET_TWDR_RECYCLE_SLOTS; n++) {
349                 if (time_before_eq(j, now)) {
350                         struct hlist_node *node, *safe;
351                         struct inet_timewait_sock *tw;
352
353                         inet_twsk_for_each_inmate_safe(tw, node, safe,
354                                                        &twdr->twcal_row[slot]) {
355                                 __inet_twsk_del_dead_node(tw);
356                                 __inet_twsk_kill(tw, twdr->hashinfo);
357                                 inet_twsk_put(tw);
358                                 killed++;
359                         }
360                 } else {
361                         if (!adv) {
362                                 adv = 1;
363                                 twdr->twcal_jiffie = j;
364                                 twdr->twcal_hand = slot;
365                         }
366
367                         if (!hlist_empty(&twdr->twcal_row[slot])) {
368                                 mod_timer(&twdr->twcal_timer, j);
369                                 goto out;
370                         }
371                 }
372                 j += 1 << INET_TWDR_RECYCLE_TICK;
373                 slot = (slot + 1) & (INET_TWDR_RECYCLE_SLOTS - 1);
374         }
375         twdr->twcal_hand = -1;
376
377 out:
378         if ((twdr->tw_count -= killed) == 0)
379                 del_timer(&twdr->tw_timer);
380         NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITKILLED, killed);
381         spin_unlock(&twdr->death_lock);
382 }
383
384 EXPORT_SYMBOL_GPL(inet_twdr_twcal_tick);