[NET]: Fix bug in sk_filter race cures.
[linux-2.6.git] / net / core / flow.c
1 /* flow.c: Generic flow cache.
2  *
3  * Copyright (C) 2003 Alexey N. Kuznetsov (kuznet@ms2.inr.ac.ru)
4  * Copyright (C) 2003 David S. Miller (davem@redhat.com)
5  */
6
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/list.h>
10 #include <linux/jhash.h>
11 #include <linux/interrupt.h>
12 #include <linux/mm.h>
13 #include <linux/random.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/smp.h>
17 #include <linux/completion.h>
18 #include <linux/percpu.h>
19 #include <linux/bitops.h>
20 #include <linux/notifier.h>
21 #include <linux/cpu.h>
22 #include <linux/cpumask.h>
23 #include <linux/mutex.h>
24 #include <net/flow.h>
25 #include <asm/atomic.h>
26 #include <asm/semaphore.h>
27 #include <linux/security.h>
28
29 struct flow_cache_entry {
30         struct flow_cache_entry *next;
31         u16                     family;
32         u8                      dir;
33         struct flowi            key;
34         u32                     genid;
35         void                    *object;
36         atomic_t                *object_ref;
37 };
38
39 atomic_t flow_cache_genid = ATOMIC_INIT(0);
40
41 static u32 flow_hash_shift;
42 #define flow_hash_size  (1 << flow_hash_shift)
43 static DEFINE_PER_CPU(struct flow_cache_entry **, flow_tables) = { NULL };
44
45 #define flow_table(cpu) (per_cpu(flow_tables, cpu))
46
47 static struct kmem_cache *flow_cachep __read_mostly;
48
49 static int flow_lwm, flow_hwm;
50
51 struct flow_percpu_info {
52         int hash_rnd_recalc;
53         u32 hash_rnd;
54         int count;
55 } ____cacheline_aligned;
56 static DEFINE_PER_CPU(struct flow_percpu_info, flow_hash_info) = { 0 };
57
58 #define flow_hash_rnd_recalc(cpu) \
59         (per_cpu(flow_hash_info, cpu).hash_rnd_recalc)
60 #define flow_hash_rnd(cpu) \
61         (per_cpu(flow_hash_info, cpu).hash_rnd)
62 #define flow_count(cpu) \
63         (per_cpu(flow_hash_info, cpu).count)
64
65 static struct timer_list flow_hash_rnd_timer;
66
67 #define FLOW_HASH_RND_PERIOD    (10 * 60 * HZ)
68
69 struct flow_flush_info {
70         atomic_t cpuleft;
71         struct completion completion;
72 };
73 static DEFINE_PER_CPU(struct tasklet_struct, flow_flush_tasklets) = { NULL };
74
75 #define flow_flush_tasklet(cpu) (&per_cpu(flow_flush_tasklets, cpu))
76
77 static void flow_cache_new_hashrnd(unsigned long arg)
78 {
79         int i;
80
81         for_each_possible_cpu(i)
82                 flow_hash_rnd_recalc(i) = 1;
83
84         flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
85         add_timer(&flow_hash_rnd_timer);
86 }
87
88 static void flow_entry_kill(int cpu, struct flow_cache_entry *fle)
89 {
90         if (fle->object)
91                 atomic_dec(fle->object_ref);
92         kmem_cache_free(flow_cachep, fle);
93         flow_count(cpu)--;
94 }
95
96 static void __flow_cache_shrink(int cpu, int shrink_to)
97 {
98         struct flow_cache_entry *fle, **flp;
99         int i;
100
101         for (i = 0; i < flow_hash_size; i++) {
102                 int k = 0;
103
104                 flp = &flow_table(cpu)[i];
105                 while ((fle = *flp) != NULL && k < shrink_to) {
106                         k++;
107                         flp = &fle->next;
108                 }
109                 while ((fle = *flp) != NULL) {
110                         *flp = fle->next;
111                         flow_entry_kill(cpu, fle);
112                 }
113         }
114 }
115
116 static void flow_cache_shrink(int cpu)
117 {
118         int shrink_to = flow_lwm / flow_hash_size;
119
120         __flow_cache_shrink(cpu, shrink_to);
121 }
122
123 static void flow_new_hash_rnd(int cpu)
124 {
125         get_random_bytes(&flow_hash_rnd(cpu), sizeof(u32));
126         flow_hash_rnd_recalc(cpu) = 0;
127
128         __flow_cache_shrink(cpu, 0);
129 }
130
131 static u32 flow_hash_code(struct flowi *key, int cpu)
132 {
133         u32 *k = (u32 *) key;
134
135         return (jhash2(k, (sizeof(*key) / sizeof(u32)), flow_hash_rnd(cpu)) &
136                 (flow_hash_size - 1));
137 }
138
139 #if (BITS_PER_LONG == 64)
140 typedef u64 flow_compare_t;
141 #else
142 typedef u32 flow_compare_t;
143 #endif
144
145 extern void flowi_is_missized(void);
146
147 /* I hear what you're saying, use memcmp.  But memcmp cannot make
148  * important assumptions that we can here, such as alignment and
149  * constant size.
150  */
151 static int flow_key_compare(struct flowi *key1, struct flowi *key2)
152 {
153         flow_compare_t *k1, *k1_lim, *k2;
154         const int n_elem = sizeof(struct flowi) / sizeof(flow_compare_t);
155
156         if (sizeof(struct flowi) % sizeof(flow_compare_t))
157                 flowi_is_missized();
158
159         k1 = (flow_compare_t *) key1;
160         k1_lim = k1 + n_elem;
161
162         k2 = (flow_compare_t *) key2;
163
164         do {
165                 if (*k1++ != *k2++)
166                         return 1;
167         } while (k1 < k1_lim);
168
169         return 0;
170 }
171
172 void *flow_cache_lookup(struct flowi *key, u16 family, u8 dir,
173                         flow_resolve_t resolver)
174 {
175         struct flow_cache_entry *fle, **head;
176         unsigned int hash;
177         int cpu;
178
179         local_bh_disable();
180         cpu = smp_processor_id();
181
182         fle = NULL;
183         /* Packet really early in init?  Making flow_cache_init a
184          * pre-smp initcall would solve this.  --RR */
185         if (!flow_table(cpu))
186                 goto nocache;
187
188         if (flow_hash_rnd_recalc(cpu))
189                 flow_new_hash_rnd(cpu);
190         hash = flow_hash_code(key, cpu);
191
192         head = &flow_table(cpu)[hash];
193         for (fle = *head; fle; fle = fle->next) {
194                 if (fle->family == family &&
195                     fle->dir == dir &&
196                     flow_key_compare(key, &fle->key) == 0) {
197                         if (fle->genid == atomic_read(&flow_cache_genid)) {
198                                 void *ret = fle->object;
199
200                                 if (ret)
201                                         atomic_inc(fle->object_ref);
202                                 local_bh_enable();
203
204                                 return ret;
205                         }
206                         break;
207                 }
208         }
209
210         if (!fle) {
211                 if (flow_count(cpu) > flow_hwm)
212                         flow_cache_shrink(cpu);
213
214                 fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
215                 if (fle) {
216                         fle->next = *head;
217                         *head = fle;
218                         fle->family = family;
219                         fle->dir = dir;
220                         memcpy(&fle->key, key, sizeof(*key));
221                         fle->object = NULL;
222                         flow_count(cpu)++;
223                 }
224         }
225
226 nocache:
227         {
228                 int err;
229                 void *obj;
230                 atomic_t *obj_ref;
231
232                 err = resolver(key, family, dir, &obj, &obj_ref);
233
234                 if (fle && !err) {
235                         fle->genid = atomic_read(&flow_cache_genid);
236
237                         if (fle->object)
238                                 atomic_dec(fle->object_ref);
239
240                         fle->object = obj;
241                         fle->object_ref = obj_ref;
242                         if (obj)
243                                 atomic_inc(fle->object_ref);
244                 }
245                 local_bh_enable();
246
247                 if (err)
248                         obj = ERR_PTR(err);
249                 return obj;
250         }
251 }
252
253 static void flow_cache_flush_tasklet(unsigned long data)
254 {
255         struct flow_flush_info *info = (void *)data;
256         int i;
257         int cpu;
258
259         cpu = smp_processor_id();
260         for (i = 0; i < flow_hash_size; i++) {
261                 struct flow_cache_entry *fle;
262
263                 fle = flow_table(cpu)[i];
264                 for (; fle; fle = fle->next) {
265                         unsigned genid = atomic_read(&flow_cache_genid);
266
267                         if (!fle->object || fle->genid == genid)
268                                 continue;
269
270                         fle->object = NULL;
271                         atomic_dec(fle->object_ref);
272                 }
273         }
274
275         if (atomic_dec_and_test(&info->cpuleft))
276                 complete(&info->completion);
277 }
278
279 static void flow_cache_flush_per_cpu(void *) __attribute__((__unused__));
280 static void flow_cache_flush_per_cpu(void *data)
281 {
282         struct flow_flush_info *info = data;
283         int cpu;
284         struct tasklet_struct *tasklet;
285
286         cpu = smp_processor_id();
287
288         tasklet = flow_flush_tasklet(cpu);
289         tasklet->data = (unsigned long)info;
290         tasklet_schedule(tasklet);
291 }
292
293 void flow_cache_flush(void)
294 {
295         struct flow_flush_info info;
296         static DEFINE_MUTEX(flow_flush_sem);
297
298         /* Don't want cpus going down or up during this. */
299         lock_cpu_hotplug();
300         mutex_lock(&flow_flush_sem);
301         atomic_set(&info.cpuleft, num_online_cpus());
302         init_completion(&info.completion);
303
304         local_bh_disable();
305         smp_call_function(flow_cache_flush_per_cpu, &info, 1, 0);
306         flow_cache_flush_tasklet((unsigned long)&info);
307         local_bh_enable();
308
309         wait_for_completion(&info.completion);
310         mutex_unlock(&flow_flush_sem);
311         unlock_cpu_hotplug();
312 }
313
314 static void __devinit flow_cache_cpu_prepare(int cpu)
315 {
316         struct tasklet_struct *tasklet;
317         unsigned long order;
318
319         for (order = 0;
320              (PAGE_SIZE << order) <
321                      (sizeof(struct flow_cache_entry *)*flow_hash_size);
322              order++)
323                 /* NOTHING */;
324
325         flow_table(cpu) = (struct flow_cache_entry **)
326                 __get_free_pages(GFP_KERNEL|__GFP_ZERO, order);
327         if (!flow_table(cpu))
328                 panic("NET: failed to allocate flow cache order %lu\n", order);
329
330         flow_hash_rnd_recalc(cpu) = 1;
331         flow_count(cpu) = 0;
332
333         tasklet = flow_flush_tasklet(cpu);
334         tasklet_init(tasklet, flow_cache_flush_tasklet, 0);
335 }
336
337 static int flow_cache_cpu(struct notifier_block *nfb,
338                           unsigned long action,
339                           void *hcpu)
340 {
341         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
342                 __flow_cache_shrink((unsigned long)hcpu, 0);
343         return NOTIFY_OK;
344 }
345
346 static int __init flow_cache_init(void)
347 {
348         int i;
349
350         flow_cachep = kmem_cache_create("flow_cache",
351                                         sizeof(struct flow_cache_entry),
352                                         0, SLAB_HWCACHE_ALIGN|SLAB_PANIC,
353                                         NULL);
354         flow_hash_shift = 10;
355         flow_lwm = 2 * flow_hash_size;
356         flow_hwm = 4 * flow_hash_size;
357
358         init_timer(&flow_hash_rnd_timer);
359         flow_hash_rnd_timer.function = flow_cache_new_hashrnd;
360         flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
361         add_timer(&flow_hash_rnd_timer);
362
363         for_each_possible_cpu(i)
364                 flow_cache_cpu_prepare(i);
365
366         hotcpu_notifier(flow_cache_cpu, 0);
367         return 0;
368 }
369
370 module_init(flow_cache_init);
371
372 EXPORT_SYMBOL(flow_cache_genid);
373 EXPORT_SYMBOL(flow_cache_lookup);