7e95b39de9fdd369d0ae0011a42c993c554f37f2
[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 <net/flow.h>
24 #include <asm/atomic.h>
25 #include <asm/semaphore.h>
26
27 struct flow_cache_entry {
28         struct flow_cache_entry *next;
29         u16                     family;
30         u8                      dir;
31         struct flowi            key;
32         u32                     genid;
33         void                    *object;
34         atomic_t                *object_ref;
35 };
36
37 atomic_t flow_cache_genid = ATOMIC_INIT(0);
38
39 static u32 flow_hash_shift;
40 #define flow_hash_size  (1 << flow_hash_shift)
41 static DEFINE_PER_CPU(struct flow_cache_entry **, flow_tables) = { NULL };
42
43 #define flow_table(cpu) (per_cpu(flow_tables, cpu))
44
45 static kmem_cache_t *flow_cachep __read_mostly;
46
47 static int flow_lwm, flow_hwm;
48
49 struct flow_percpu_info {
50         int hash_rnd_recalc;
51         u32 hash_rnd;
52         int count;
53 } ____cacheline_aligned;
54 static DEFINE_PER_CPU(struct flow_percpu_info, flow_hash_info) = { 0 };
55
56 #define flow_hash_rnd_recalc(cpu) \
57         (per_cpu(flow_hash_info, cpu).hash_rnd_recalc)
58 #define flow_hash_rnd(cpu) \
59         (per_cpu(flow_hash_info, cpu).hash_rnd)
60 #define flow_count(cpu) \
61         (per_cpu(flow_hash_info, cpu).count)
62
63 static struct timer_list flow_hash_rnd_timer;
64
65 #define FLOW_HASH_RND_PERIOD    (10 * 60 * HZ)
66
67 struct flow_flush_info {
68         atomic_t cpuleft;
69         struct completion completion;
70 };
71 static DEFINE_PER_CPU(struct tasklet_struct, flow_flush_tasklets) = { NULL };
72
73 #define flow_flush_tasklet(cpu) (&per_cpu(flow_flush_tasklets, cpu))
74
75 static void flow_cache_new_hashrnd(unsigned long arg)
76 {
77         int i;
78
79         for_each_cpu(i)
80                 flow_hash_rnd_recalc(i) = 1;
81
82         flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
83         add_timer(&flow_hash_rnd_timer);
84 }
85
86 static void __flow_cache_shrink(int cpu, int shrink_to)
87 {
88         struct flow_cache_entry *fle, **flp;
89         int i;
90
91         for (i = 0; i < flow_hash_size; i++) {
92                 int k = 0;
93
94                 flp = &flow_table(cpu)[i];
95                 while ((fle = *flp) != NULL && k < shrink_to) {
96                         k++;
97                         flp = &fle->next;
98                 }
99                 while ((fle = *flp) != NULL) {
100                         *flp = fle->next;
101                         if (fle->object)
102                                 atomic_dec(fle->object_ref);
103                         kmem_cache_free(flow_cachep, fle);
104                         flow_count(cpu)--;
105                 }
106         }
107 }
108
109 static void flow_cache_shrink(int cpu)
110 {
111         int shrink_to = flow_lwm / flow_hash_size;
112
113         __flow_cache_shrink(cpu, shrink_to);
114 }
115
116 static void flow_new_hash_rnd(int cpu)
117 {
118         get_random_bytes(&flow_hash_rnd(cpu), sizeof(u32));
119         flow_hash_rnd_recalc(cpu) = 0;
120
121         __flow_cache_shrink(cpu, 0);
122 }
123
124 static u32 flow_hash_code(struct flowi *key, int cpu)
125 {
126         u32 *k = (u32 *) key;
127
128         return (jhash2(k, (sizeof(*key) / sizeof(u32)), flow_hash_rnd(cpu)) &
129                 (flow_hash_size - 1));
130 }
131
132 #if (BITS_PER_LONG == 64)
133 typedef u64 flow_compare_t;
134 #else
135 typedef u32 flow_compare_t;
136 #endif
137
138 extern void flowi_is_missized(void);
139
140 /* I hear what you're saying, use memcmp.  But memcmp cannot make
141  * important assumptions that we can here, such as alignment and
142  * constant size.
143  */
144 static int flow_key_compare(struct flowi *key1, struct flowi *key2)
145 {
146         flow_compare_t *k1, *k1_lim, *k2;
147         const int n_elem = sizeof(struct flowi) / sizeof(flow_compare_t);
148
149         if (sizeof(struct flowi) % sizeof(flow_compare_t))
150                 flowi_is_missized();
151
152         k1 = (flow_compare_t *) key1;
153         k1_lim = k1 + n_elem;
154
155         k2 = (flow_compare_t *) key2;
156
157         do {
158                 if (*k1++ != *k2++)
159                         return 1;
160         } while (k1 < k1_lim);
161
162         return 0;
163 }
164
165 void *flow_cache_lookup(struct flowi *key, u16 family, u8 dir,
166                         flow_resolve_t resolver)
167 {
168         struct flow_cache_entry *fle, **head;
169         unsigned int hash;
170         int cpu;
171
172         local_bh_disable();
173         cpu = smp_processor_id();
174
175         fle = NULL;
176         /* Packet really early in init?  Making flow_cache_init a
177          * pre-smp initcall would solve this.  --RR */
178         if (!flow_table(cpu))
179                 goto nocache;
180
181         if (flow_hash_rnd_recalc(cpu))
182                 flow_new_hash_rnd(cpu);
183         hash = flow_hash_code(key, cpu);
184
185         head = &flow_table(cpu)[hash];
186         for (fle = *head; fle; fle = fle->next) {
187                 if (fle->family == family &&
188                     fle->dir == dir &&
189                     flow_key_compare(key, &fle->key) == 0) {
190                         if (fle->genid == atomic_read(&flow_cache_genid)) {
191                                 void *ret = fle->object;
192
193                                 if (ret)
194                                         atomic_inc(fle->object_ref);
195                                 local_bh_enable();
196
197                                 return ret;
198                         }
199                         break;
200                 }
201         }
202
203         if (!fle) {
204                 if (flow_count(cpu) > flow_hwm)
205                         flow_cache_shrink(cpu);
206
207                 fle = kmem_cache_alloc(flow_cachep, SLAB_ATOMIC);
208                 if (fle) {
209                         fle->next = *head;
210                         *head = fle;
211                         fle->family = family;
212                         fle->dir = dir;
213                         memcpy(&fle->key, key, sizeof(*key));
214                         fle->object = NULL;
215                         flow_count(cpu)++;
216                 }
217         }
218
219 nocache:
220         {
221                 void *obj;
222                 atomic_t *obj_ref;
223
224                 resolver(key, family, dir, &obj, &obj_ref);
225
226                 if (fle) {
227                         fle->genid = atomic_read(&flow_cache_genid);
228
229                         if (fle->object)
230                                 atomic_dec(fle->object_ref);
231
232                         fle->object = obj;
233                         fle->object_ref = obj_ref;
234                         if (obj)
235                                 atomic_inc(fle->object_ref);
236                 }
237                 local_bh_enable();
238
239                 return obj;
240         }
241 }
242
243 static void flow_cache_flush_tasklet(unsigned long data)
244 {
245         struct flow_flush_info *info = (void *)data;
246         int i;
247         int cpu;
248
249         cpu = smp_processor_id();
250         for (i = 0; i < flow_hash_size; i++) {
251                 struct flow_cache_entry *fle;
252
253                 fle = flow_table(cpu)[i];
254                 for (; fle; fle = fle->next) {
255                         unsigned genid = atomic_read(&flow_cache_genid);
256
257                         if (!fle->object || fle->genid == genid)
258                                 continue;
259
260                         fle->object = NULL;
261                         atomic_dec(fle->object_ref);
262                 }
263         }
264
265         if (atomic_dec_and_test(&info->cpuleft))
266                 complete(&info->completion);
267 }
268
269 static void flow_cache_flush_per_cpu(void *) __attribute__((__unused__));
270 static void flow_cache_flush_per_cpu(void *data)
271 {
272         struct flow_flush_info *info = data;
273         int cpu;
274         struct tasklet_struct *tasklet;
275
276         cpu = smp_processor_id();
277
278         tasklet = flow_flush_tasklet(cpu);
279         tasklet->data = (unsigned long)info;
280         tasklet_schedule(tasklet);
281 }
282
283 void flow_cache_flush(void)
284 {
285         struct flow_flush_info info;
286         static DECLARE_MUTEX(flow_flush_sem);
287
288         /* Don't want cpus going down or up during this. */
289         lock_cpu_hotplug();
290         down(&flow_flush_sem);
291         atomic_set(&info.cpuleft, num_online_cpus());
292         init_completion(&info.completion);
293
294         local_bh_disable();
295         smp_call_function(flow_cache_flush_per_cpu, &info, 1, 0);
296         flow_cache_flush_tasklet((unsigned long)&info);
297         local_bh_enable();
298
299         wait_for_completion(&info.completion);
300         up(&flow_flush_sem);
301         unlock_cpu_hotplug();
302 }
303
304 static void __devinit flow_cache_cpu_prepare(int cpu)
305 {
306         struct tasklet_struct *tasklet;
307         unsigned long order;
308
309         for (order = 0;
310              (PAGE_SIZE << order) <
311                      (sizeof(struct flow_cache_entry *)*flow_hash_size);
312              order++)
313                 /* NOTHING */;
314
315         flow_table(cpu) = (struct flow_cache_entry **)
316                 __get_free_pages(GFP_KERNEL, order);
317         if (!flow_table(cpu))
318                 panic("NET: failed to allocate flow cache order %lu\n", order);
319
320         memset(flow_table(cpu), 0, PAGE_SIZE << order);
321
322         flow_hash_rnd_recalc(cpu) = 1;
323         flow_count(cpu) = 0;
324
325         tasklet = flow_flush_tasklet(cpu);
326         tasklet_init(tasklet, flow_cache_flush_tasklet, 0);
327 }
328
329 #ifdef CONFIG_HOTPLUG_CPU
330 static int flow_cache_cpu(struct notifier_block *nfb,
331                           unsigned long action,
332                           void *hcpu)
333 {
334         if (action == CPU_DEAD)
335                 __flow_cache_shrink((unsigned long)hcpu, 0);
336         return NOTIFY_OK;
337 }
338 #endif /* CONFIG_HOTPLUG_CPU */
339
340 static int __init flow_cache_init(void)
341 {
342         int i;
343
344         flow_cachep = kmem_cache_create("flow_cache",
345                                         sizeof(struct flow_cache_entry),
346                                         0, SLAB_HWCACHE_ALIGN,
347                                         NULL, NULL);
348
349         if (!flow_cachep)
350                 panic("NET: failed to allocate flow cache slab\n");
351
352         flow_hash_shift = 10;
353         flow_lwm = 2 * flow_hash_size;
354         flow_hwm = 4 * flow_hash_size;
355
356         init_timer(&flow_hash_rnd_timer);
357         flow_hash_rnd_timer.function = flow_cache_new_hashrnd;
358         flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
359         add_timer(&flow_hash_rnd_timer);
360
361         for_each_cpu(i)
362                 flow_cache_cpu_prepare(i);
363
364         hotcpu_notifier(flow_cache_cpu, 0);
365         return 0;
366 }
367
368 module_init(flow_cache_init);
369
370 EXPORT_SYMBOL(flow_cache_genid);
371 EXPORT_SYMBOL(flow_cache_lookup);