b143b86b1f2a433e8a69573bf1887b986d970e8a
[linux-3.10.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 <linux/security.h>
27
28 struct flow_cache_entry {
29         union {
30                 struct hlist_node       hlist;
31                 struct list_head        gc_list;
32         } u;
33         u16                             family;
34         u8                              dir;
35         u32                             genid;
36         struct flowi                    key;
37         struct flow_cache_object        *object;
38 };
39
40 struct flow_cache_percpu {
41         struct hlist_head               *hash_table;
42         int                             hash_count;
43         u32                             hash_rnd;
44         int                             hash_rnd_recalc;
45         struct tasklet_struct           flush_tasklet;
46 };
47
48 struct flow_flush_info {
49         struct flow_cache               *cache;
50         atomic_t                        cpuleft;
51         struct completion               completion;
52 };
53
54 struct flow_cache {
55         u32                             hash_shift;
56         struct flow_cache_percpu __percpu *percpu;
57         struct notifier_block           hotcpu_notifier;
58         int                             low_watermark;
59         int                             high_watermark;
60         struct timer_list               rnd_timer;
61 };
62
63 atomic_t flow_cache_genid = ATOMIC_INIT(0);
64 EXPORT_SYMBOL(flow_cache_genid);
65 static struct flow_cache flow_cache_global;
66 static struct kmem_cache *flow_cachep __read_mostly;
67
68 static DEFINE_SPINLOCK(flow_cache_gc_lock);
69 static LIST_HEAD(flow_cache_gc_list);
70
71 #define flow_cache_hash_size(cache)     (1 << (cache)->hash_shift)
72 #define FLOW_HASH_RND_PERIOD            (10 * 60 * HZ)
73
74 static void flow_cache_new_hashrnd(unsigned long arg)
75 {
76         struct flow_cache *fc = (void *) arg;
77         int i;
78
79         for_each_possible_cpu(i)
80                 per_cpu_ptr(fc->percpu, i)->hash_rnd_recalc = 1;
81
82         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
83         add_timer(&fc->rnd_timer);
84 }
85
86 static int flow_entry_valid(struct flow_cache_entry *fle)
87 {
88         if (atomic_read(&flow_cache_genid) != fle->genid)
89                 return 0;
90         if (fle->object && !fle->object->ops->check(fle->object))
91                 return 0;
92         return 1;
93 }
94
95 static void flow_entry_kill(struct flow_cache_entry *fle)
96 {
97         if (fle->object)
98                 fle->object->ops->delete(fle->object);
99         kmem_cache_free(flow_cachep, fle);
100 }
101
102 static void flow_cache_gc_task(struct work_struct *work)
103 {
104         struct list_head gc_list;
105         struct flow_cache_entry *fce, *n;
106
107         INIT_LIST_HEAD(&gc_list);
108         spin_lock_bh(&flow_cache_gc_lock);
109         list_splice_tail_init(&flow_cache_gc_list, &gc_list);
110         spin_unlock_bh(&flow_cache_gc_lock);
111
112         list_for_each_entry_safe(fce, n, &gc_list, u.gc_list)
113                 flow_entry_kill(fce);
114 }
115 static DECLARE_WORK(flow_cache_gc_work, flow_cache_gc_task);
116
117 static void flow_cache_queue_garbage(struct flow_cache_percpu *fcp,
118                                      int deleted, struct list_head *gc_list)
119 {
120         if (deleted) {
121                 fcp->hash_count -= deleted;
122                 spin_lock_bh(&flow_cache_gc_lock);
123                 list_splice_tail(gc_list, &flow_cache_gc_list);
124                 spin_unlock_bh(&flow_cache_gc_lock);
125                 schedule_work(&flow_cache_gc_work);
126         }
127 }
128
129 static void __flow_cache_shrink(struct flow_cache *fc,
130                                 struct flow_cache_percpu *fcp,
131                                 int shrink_to)
132 {
133         struct flow_cache_entry *fle;
134         struct hlist_node *entry, *tmp;
135         LIST_HEAD(gc_list);
136         int i, deleted = 0;
137
138         for (i = 0; i < flow_cache_hash_size(fc); i++) {
139                 int saved = 0;
140
141                 hlist_for_each_entry_safe(fle, entry, tmp,
142                                           &fcp->hash_table[i], u.hlist) {
143                         if (saved < shrink_to &&
144                             flow_entry_valid(fle)) {
145                                 saved++;
146                         } else {
147                                 deleted++;
148                                 hlist_del(&fle->u.hlist);
149                                 list_add_tail(&fle->u.gc_list, &gc_list);
150                         }
151                 }
152         }
153
154         flow_cache_queue_garbage(fcp, deleted, &gc_list);
155 }
156
157 static void flow_cache_shrink(struct flow_cache *fc,
158                               struct flow_cache_percpu *fcp)
159 {
160         int shrink_to = fc->low_watermark / flow_cache_hash_size(fc);
161
162         __flow_cache_shrink(fc, fcp, shrink_to);
163 }
164
165 static void flow_new_hash_rnd(struct flow_cache *fc,
166                               struct flow_cache_percpu *fcp)
167 {
168         get_random_bytes(&fcp->hash_rnd, sizeof(u32));
169         fcp->hash_rnd_recalc = 0;
170         __flow_cache_shrink(fc, fcp, 0);
171 }
172
173 static u32 flow_hash_code(struct flow_cache *fc,
174                           struct flow_cache_percpu *fcp,
175                           struct flowi *key)
176 {
177         u32 *k = (u32 *) key;
178
179         return (jhash2(k, (sizeof(*key) / sizeof(u32)), fcp->hash_rnd)
180                 & (flow_cache_hash_size(fc) - 1));
181 }
182
183 typedef unsigned long flow_compare_t;
184
185 /* I hear what you're saying, use memcmp.  But memcmp cannot make
186  * important assumptions that we can here, such as alignment and
187  * constant size.
188  */
189 static int flow_key_compare(struct flowi *key1, struct flowi *key2)
190 {
191         flow_compare_t *k1, *k1_lim, *k2;
192         const int n_elem = sizeof(struct flowi) / sizeof(flow_compare_t);
193
194         BUILD_BUG_ON(sizeof(struct flowi) % sizeof(flow_compare_t));
195
196         k1 = (flow_compare_t *) key1;
197         k1_lim = k1 + n_elem;
198
199         k2 = (flow_compare_t *) key2;
200
201         do {
202                 if (*k1++ != *k2++)
203                         return 1;
204         } while (k1 < k1_lim);
205
206         return 0;
207 }
208
209 struct flow_cache_object *
210 flow_cache_lookup(struct net *net, struct flowi *key, u16 family, u8 dir,
211                   flow_resolve_t resolver, void *ctx)
212 {
213         struct flow_cache *fc = &flow_cache_global;
214         struct flow_cache_percpu *fcp;
215         struct flow_cache_entry *fle, *tfle;
216         struct hlist_node *entry;
217         struct flow_cache_object *flo;
218         unsigned int hash;
219
220         local_bh_disable();
221         fcp = this_cpu_ptr(fc->percpu);
222
223         fle = NULL;
224         flo = NULL;
225         /* Packet really early in init?  Making flow_cache_init a
226          * pre-smp initcall would solve this.  --RR */
227         if (!fcp->hash_table)
228                 goto nocache;
229
230         if (fcp->hash_rnd_recalc)
231                 flow_new_hash_rnd(fc, fcp);
232
233         hash = flow_hash_code(fc, fcp, key);
234         hlist_for_each_entry(tfle, entry, &fcp->hash_table[hash], u.hlist) {
235                 if (tfle->family == family &&
236                     tfle->dir == dir &&
237                     flow_key_compare(key, &tfle->key) == 0) {
238                         fle = tfle;
239                         break;
240                 }
241         }
242
243         if (unlikely(!fle)) {
244                 if (fcp->hash_count > fc->high_watermark)
245                         flow_cache_shrink(fc, fcp);
246
247                 fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
248                 if (fle) {
249                         fle->family = family;
250                         fle->dir = dir;
251                         memcpy(&fle->key, key, sizeof(*key));
252                         fle->object = NULL;
253                         hlist_add_head(&fle->u.hlist, &fcp->hash_table[hash]);
254                         fcp->hash_count++;
255                 }
256         } else if (likely(fle->genid == atomic_read(&flow_cache_genid))) {
257                 flo = fle->object;
258                 if (!flo)
259                         goto ret_object;
260                 flo = flo->ops->get(flo);
261                 if (flo)
262                         goto ret_object;
263         } else if (fle->object) {
264                 flo = fle->object;
265                 flo->ops->delete(flo);
266                 fle->object = NULL;
267         }
268
269 nocache:
270         flo = NULL;
271         if (fle) {
272                 flo = fle->object;
273                 fle->object = NULL;
274         }
275         flo = resolver(net, key, family, dir, flo, ctx);
276         if (fle) {
277                 fle->genid = atomic_read(&flow_cache_genid);
278                 if (!IS_ERR(flo))
279                         fle->object = flo;
280                 else
281                         fle->genid--;
282         } else {
283                 if (flo && !IS_ERR(flo))
284                         flo->ops->delete(flo);
285         }
286 ret_object:
287         local_bh_enable();
288         return flo;
289 }
290 EXPORT_SYMBOL(flow_cache_lookup);
291
292 static void flow_cache_flush_tasklet(unsigned long data)
293 {
294         struct flow_flush_info *info = (void *)data;
295         struct flow_cache *fc = info->cache;
296         struct flow_cache_percpu *fcp;
297         struct flow_cache_entry *fle;
298         struct hlist_node *entry, *tmp;
299         LIST_HEAD(gc_list);
300         int i, deleted = 0;
301
302         fcp = this_cpu_ptr(fc->percpu);
303         for (i = 0; i < flow_cache_hash_size(fc); i++) {
304                 hlist_for_each_entry_safe(fle, entry, tmp,
305                                           &fcp->hash_table[i], u.hlist) {
306                         if (flow_entry_valid(fle))
307                                 continue;
308
309                         deleted++;
310                         hlist_del(&fle->u.hlist);
311                         list_add_tail(&fle->u.gc_list, &gc_list);
312                 }
313         }
314
315         flow_cache_queue_garbage(fcp, deleted, &gc_list);
316
317         if (atomic_dec_and_test(&info->cpuleft))
318                 complete(&info->completion);
319 }
320
321 static void flow_cache_flush_per_cpu(void *data)
322 {
323         struct flow_flush_info *info = data;
324         int cpu;
325         struct tasklet_struct *tasklet;
326
327         cpu = smp_processor_id();
328         tasklet = &per_cpu_ptr(info->cache->percpu, cpu)->flush_tasklet;
329         tasklet->data = (unsigned long)info;
330         tasklet_schedule(tasklet);
331 }
332
333 void flow_cache_flush(void)
334 {
335         struct flow_flush_info info;
336         static DEFINE_MUTEX(flow_flush_sem);
337
338         /* Don't want cpus going down or up during this. */
339         get_online_cpus();
340         mutex_lock(&flow_flush_sem);
341         info.cache = &flow_cache_global;
342         atomic_set(&info.cpuleft, num_online_cpus());
343         init_completion(&info.completion);
344
345         local_bh_disable();
346         smp_call_function(flow_cache_flush_per_cpu, &info, 0);
347         flow_cache_flush_tasklet((unsigned long)&info);
348         local_bh_enable();
349
350         wait_for_completion(&info.completion);
351         mutex_unlock(&flow_flush_sem);
352         put_online_cpus();
353 }
354
355 static int __cpuinit flow_cache_cpu_prepare(struct flow_cache *fc, int cpu)
356 {
357         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
358         size_t sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
359
360         if (!fcp->hash_table) {
361                 fcp->hash_table = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
362                 if (!fcp->hash_table) {
363                         pr_err("NET: failed to allocate flow cache sz %zu\n", sz);
364                         return -ENOMEM;
365                 }
366                 fcp->hash_rnd_recalc = 1;
367                 fcp->hash_count = 0;
368                 tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
369         }
370         return 0;
371 }
372
373 static int __cpuinit flow_cache_cpu(struct notifier_block *nfb,
374                           unsigned long action,
375                           void *hcpu)
376 {
377         struct flow_cache *fc = container_of(nfb, struct flow_cache, hotcpu_notifier);
378         int res, cpu = (unsigned long) hcpu;
379         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
380
381         switch (action) {
382         case CPU_UP_PREPARE:
383         case CPU_UP_PREPARE_FROZEN:
384                 res = flow_cache_cpu_prepare(fc, cpu);
385                 if (res)
386                         return notifier_from_errno(res);
387                 break;
388         case CPU_DEAD:
389         case CPU_DEAD_FROZEN:
390                 __flow_cache_shrink(fc, fcp, 0);
391                 break;
392         }
393         return NOTIFY_OK;
394 }
395
396 static int __init flow_cache_init(struct flow_cache *fc)
397 {
398         int i;
399
400         fc->hash_shift = 10;
401         fc->low_watermark = 2 * flow_cache_hash_size(fc);
402         fc->high_watermark = 4 * flow_cache_hash_size(fc);
403
404         fc->percpu = alloc_percpu(struct flow_cache_percpu);
405         if (!fc->percpu)
406                 return -ENOMEM;
407
408         for_each_online_cpu(i) {
409                 if (flow_cache_cpu_prepare(fc, i))
410                         return -ENOMEM;
411         }
412         fc->hotcpu_notifier = (struct notifier_block){
413                 .notifier_call = flow_cache_cpu,
414         };
415         register_hotcpu_notifier(&fc->hotcpu_notifier);
416
417         setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
418                     (unsigned long) fc);
419         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
420         add_timer(&fc->rnd_timer);
421
422         return 0;
423 }
424
425 static int __init flow_cache_init_global(void)
426 {
427         flow_cachep = kmem_cache_create("flow_cache",
428                                         sizeof(struct flow_cache_entry),
429                                         0, SLAB_PANIC, NULL);
430
431         return flow_cache_init(&flow_cache_global);
432 }
433
434 module_init(flow_cache_init_global);