47b6d26c2afb395ff2ecc19232af7679eafb2a20
[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 <linux/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         struct net                      *net;
34         u16                             family;
35         u8                              dir;
36         u32                             genid;
37         struct flowi                    key;
38         struct flow_cache_object        *object;
39 };
40
41 struct flow_cache_percpu {
42         struct hlist_head               *hash_table;
43         int                             hash_count;
44         u32                             hash_rnd;
45         int                             hash_rnd_recalc;
46         struct tasklet_struct           flush_tasklet;
47 };
48
49 struct flow_flush_info {
50         struct flow_cache               *cache;
51         atomic_t                        cpuleft;
52         struct completion               completion;
53 };
54
55 struct flow_cache {
56         u32                             hash_shift;
57         struct flow_cache_percpu __percpu *percpu;
58         struct notifier_block           hotcpu_notifier;
59         int                             low_watermark;
60         int                             high_watermark;
61         struct timer_list               rnd_timer;
62 };
63
64 atomic_t flow_cache_genid = ATOMIC_INIT(0);
65 EXPORT_SYMBOL(flow_cache_genid);
66 static struct flow_cache flow_cache_global;
67 static struct kmem_cache *flow_cachep __read_mostly;
68
69 static DEFINE_SPINLOCK(flow_cache_gc_lock);
70 static LIST_HEAD(flow_cache_gc_list);
71
72 #define flow_cache_hash_size(cache)     (1 << (cache)->hash_shift)
73 #define FLOW_HASH_RND_PERIOD            (10 * 60 * HZ)
74
75 static void flow_cache_new_hashrnd(unsigned long arg)
76 {
77         struct flow_cache *fc = (void *) arg;
78         int i;
79
80         for_each_possible_cpu(i)
81                 per_cpu_ptr(fc->percpu, i)->hash_rnd_recalc = 1;
82
83         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
84         add_timer(&fc->rnd_timer);
85 }
86
87 static int flow_entry_valid(struct flow_cache_entry *fle)
88 {
89         if (atomic_read(&flow_cache_genid) != fle->genid)
90                 return 0;
91         if (fle->object && !fle->object->ops->check(fle->object))
92                 return 0;
93         return 1;
94 }
95
96 static void flow_entry_kill(struct flow_cache_entry *fle)
97 {
98         if (fle->object)
99                 fle->object->ops->delete(fle->object);
100         kmem_cache_free(flow_cachep, fle);
101 }
102
103 static void flow_cache_gc_task(struct work_struct *work)
104 {
105         struct list_head gc_list;
106         struct flow_cache_entry *fce, *n;
107
108         INIT_LIST_HEAD(&gc_list);
109         spin_lock_bh(&flow_cache_gc_lock);
110         list_splice_tail_init(&flow_cache_gc_list, &gc_list);
111         spin_unlock_bh(&flow_cache_gc_lock);
112
113         list_for_each_entry_safe(fce, n, &gc_list, u.gc_list)
114                 flow_entry_kill(fce);
115 }
116 static DECLARE_WORK(flow_cache_gc_work, flow_cache_gc_task);
117
118 static void flow_cache_queue_garbage(struct flow_cache_percpu *fcp,
119                                      int deleted, struct list_head *gc_list)
120 {
121         if (deleted) {
122                 fcp->hash_count -= deleted;
123                 spin_lock_bh(&flow_cache_gc_lock);
124                 list_splice_tail(gc_list, &flow_cache_gc_list);
125                 spin_unlock_bh(&flow_cache_gc_lock);
126                 schedule_work(&flow_cache_gc_work);
127         }
128 }
129
130 static void __flow_cache_shrink(struct flow_cache *fc,
131                                 struct flow_cache_percpu *fcp,
132                                 int shrink_to)
133 {
134         struct flow_cache_entry *fle;
135         struct hlist_node *entry, *tmp;
136         LIST_HEAD(gc_list);
137         int i, deleted = 0;
138
139         for (i = 0; i < flow_cache_hash_size(fc); i++) {
140                 int saved = 0;
141
142                 hlist_for_each_entry_safe(fle, entry, tmp,
143                                           &fcp->hash_table[i], u.hlist) {
144                         if (saved < shrink_to &&
145                             flow_entry_valid(fle)) {
146                                 saved++;
147                         } else {
148                                 deleted++;
149                                 hlist_del(&fle->u.hlist);
150                                 list_add_tail(&fle->u.gc_list, &gc_list);
151                         }
152                 }
153         }
154
155         flow_cache_queue_garbage(fcp, deleted, &gc_list);
156 }
157
158 static void flow_cache_shrink(struct flow_cache *fc,
159                               struct flow_cache_percpu *fcp)
160 {
161         int shrink_to = fc->low_watermark / flow_cache_hash_size(fc);
162
163         __flow_cache_shrink(fc, fcp, shrink_to);
164 }
165
166 static void flow_new_hash_rnd(struct flow_cache *fc,
167                               struct flow_cache_percpu *fcp)
168 {
169         get_random_bytes(&fcp->hash_rnd, sizeof(u32));
170         fcp->hash_rnd_recalc = 0;
171         __flow_cache_shrink(fc, fcp, 0);
172 }
173
174 static u32 flow_hash_code(struct flow_cache *fc,
175                           struct flow_cache_percpu *fcp,
176                           const struct flowi *key)
177 {
178         const u32 *k = (const u32 *) key;
179
180         return jhash2(k, (sizeof(*key) / sizeof(u32)), fcp->hash_rnd)
181                 & (flow_cache_hash_size(fc) - 1);
182 }
183
184 typedef unsigned long flow_compare_t;
185
186 /* I hear what you're saying, use memcmp.  But memcmp cannot make
187  * important assumptions that we can here, such as alignment and
188  * constant size.
189  */
190 static int flow_key_compare(const struct flowi *key1, const struct flowi *key2)
191 {
192         const flow_compare_t *k1, *k1_lim, *k2;
193         const int n_elem = sizeof(struct flowi) / sizeof(flow_compare_t);
194
195         BUILD_BUG_ON(sizeof(struct flowi) % sizeof(flow_compare_t));
196
197         k1 = (const flow_compare_t *) key1;
198         k1_lim = k1 + n_elem;
199
200         k2 = (const flow_compare_t *) key2;
201
202         do {
203                 if (*k1++ != *k2++)
204                         return 1;
205         } while (k1 < k1_lim);
206
207         return 0;
208 }
209
210 struct flow_cache_object *
211 flow_cache_lookup(struct net *net, const struct flowi *key, u16 family, u8 dir,
212                   flow_resolve_t resolver, void *ctx)
213 {
214         struct flow_cache *fc = &flow_cache_global;
215         struct flow_cache_percpu *fcp;
216         struct flow_cache_entry *fle, *tfle;
217         struct hlist_node *entry;
218         struct flow_cache_object *flo;
219         unsigned int hash;
220
221         local_bh_disable();
222         fcp = this_cpu_ptr(fc->percpu);
223
224         fle = NULL;
225         flo = NULL;
226         /* Packet really early in init?  Making flow_cache_init a
227          * pre-smp initcall would solve this.  --RR */
228         if (!fcp->hash_table)
229                 goto nocache;
230
231         if (fcp->hash_rnd_recalc)
232                 flow_new_hash_rnd(fc, fcp);
233
234         hash = flow_hash_code(fc, fcp, key);
235         hlist_for_each_entry(tfle, entry, &fcp->hash_table[hash], u.hlist) {
236                 if (tfle->net == net &&
237                     tfle->family == family &&
238                     tfle->dir == dir &&
239                     flow_key_compare(key, &tfle->key) == 0) {
240                         fle = tfle;
241                         break;
242                 }
243         }
244
245         if (unlikely(!fle)) {
246                 if (fcp->hash_count > fc->high_watermark)
247                         flow_cache_shrink(fc, fcp);
248
249                 fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
250                 if (fle) {
251                         fle->net = net;
252                         fle->family = family;
253                         fle->dir = dir;
254                         memcpy(&fle->key, key, sizeof(*key));
255                         fle->object = NULL;
256                         hlist_add_head(&fle->u.hlist, &fcp->hash_table[hash]);
257                         fcp->hash_count++;
258                 }
259         } else if (likely(fle->genid == atomic_read(&flow_cache_genid))) {
260                 flo = fle->object;
261                 if (!flo)
262                         goto ret_object;
263                 flo = flo->ops->get(flo);
264                 if (flo)
265                         goto ret_object;
266         } else if (fle->object) {
267                 flo = fle->object;
268                 flo->ops->delete(flo);
269                 fle->object = NULL;
270         }
271
272 nocache:
273         flo = NULL;
274         if (fle) {
275                 flo = fle->object;
276                 fle->object = NULL;
277         }
278         flo = resolver(net, key, family, dir, flo, ctx);
279         if (fle) {
280                 fle->genid = atomic_read(&flow_cache_genid);
281                 if (!IS_ERR(flo))
282                         fle->object = flo;
283                 else
284                         fle->genid--;
285         } else {
286                 if (flo && !IS_ERR(flo))
287                         flo->ops->delete(flo);
288         }
289 ret_object:
290         local_bh_enable();
291         return flo;
292 }
293 EXPORT_SYMBOL(flow_cache_lookup);
294
295 static void flow_cache_flush_tasklet(unsigned long data)
296 {
297         struct flow_flush_info *info = (void *)data;
298         struct flow_cache *fc = info->cache;
299         struct flow_cache_percpu *fcp;
300         struct flow_cache_entry *fle;
301         struct hlist_node *entry, *tmp;
302         LIST_HEAD(gc_list);
303         int i, deleted = 0;
304
305         fcp = this_cpu_ptr(fc->percpu);
306         for (i = 0; i < flow_cache_hash_size(fc); i++) {
307                 hlist_for_each_entry_safe(fle, entry, tmp,
308                                           &fcp->hash_table[i], u.hlist) {
309                         if (flow_entry_valid(fle))
310                                 continue;
311
312                         deleted++;
313                         hlist_del(&fle->u.hlist);
314                         list_add_tail(&fle->u.gc_list, &gc_list);
315                 }
316         }
317
318         flow_cache_queue_garbage(fcp, deleted, &gc_list);
319
320         if (atomic_dec_and_test(&info->cpuleft))
321                 complete(&info->completion);
322 }
323
324 static void flow_cache_flush_per_cpu(void *data)
325 {
326         struct flow_flush_info *info = data;
327         int cpu;
328         struct tasklet_struct *tasklet;
329
330         cpu = smp_processor_id();
331         tasklet = &per_cpu_ptr(info->cache->percpu, cpu)->flush_tasklet;
332         tasklet->data = (unsigned long)info;
333         tasklet_schedule(tasklet);
334 }
335
336 void flow_cache_flush(void)
337 {
338         struct flow_flush_info info;
339         static DEFINE_MUTEX(flow_flush_sem);
340
341         /* Don't want cpus going down or up during this. */
342         get_online_cpus();
343         mutex_lock(&flow_flush_sem);
344         info.cache = &flow_cache_global;
345         atomic_set(&info.cpuleft, num_online_cpus());
346         init_completion(&info.completion);
347
348         local_bh_disable();
349         smp_call_function(flow_cache_flush_per_cpu, &info, 0);
350         flow_cache_flush_tasklet((unsigned long)&info);
351         local_bh_enable();
352
353         wait_for_completion(&info.completion);
354         mutex_unlock(&flow_flush_sem);
355         put_online_cpus();
356 }
357
358 static int __cpuinit flow_cache_cpu_prepare(struct flow_cache *fc, int cpu)
359 {
360         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
361         size_t sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
362
363         if (!fcp->hash_table) {
364                 fcp->hash_table = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
365                 if (!fcp->hash_table) {
366                         pr_err("NET: failed to allocate flow cache sz %zu\n", sz);
367                         return -ENOMEM;
368                 }
369                 fcp->hash_rnd_recalc = 1;
370                 fcp->hash_count = 0;
371                 tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
372         }
373         return 0;
374 }
375
376 static int __cpuinit flow_cache_cpu(struct notifier_block *nfb,
377                           unsigned long action,
378                           void *hcpu)
379 {
380         struct flow_cache *fc = container_of(nfb, struct flow_cache, hotcpu_notifier);
381         int res, cpu = (unsigned long) hcpu;
382         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
383
384         switch (action) {
385         case CPU_UP_PREPARE:
386         case CPU_UP_PREPARE_FROZEN:
387                 res = flow_cache_cpu_prepare(fc, cpu);
388                 if (res)
389                         return notifier_from_errno(res);
390                 break;
391         case CPU_DEAD:
392         case CPU_DEAD_FROZEN:
393                 __flow_cache_shrink(fc, fcp, 0);
394                 break;
395         }
396         return NOTIFY_OK;
397 }
398
399 static int __init flow_cache_init(struct flow_cache *fc)
400 {
401         int i;
402
403         fc->hash_shift = 10;
404         fc->low_watermark = 2 * flow_cache_hash_size(fc);
405         fc->high_watermark = 4 * flow_cache_hash_size(fc);
406
407         fc->percpu = alloc_percpu(struct flow_cache_percpu);
408         if (!fc->percpu)
409                 return -ENOMEM;
410
411         for_each_online_cpu(i) {
412                 if (flow_cache_cpu_prepare(fc, i))
413                         return -ENOMEM;
414         }
415         fc->hotcpu_notifier = (struct notifier_block){
416                 .notifier_call = flow_cache_cpu,
417         };
418         register_hotcpu_notifier(&fc->hotcpu_notifier);
419
420         setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
421                     (unsigned long) fc);
422         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
423         add_timer(&fc->rnd_timer);
424
425         return 0;
426 }
427
428 static int __init flow_cache_init_global(void)
429 {
430         flow_cachep = kmem_cache_create("flow_cache",
431                                         sizeof(struct flow_cache_entry),
432                                         0, SLAB_PANIC, NULL);
433
434         return flow_cache_init(&flow_cache_global);
435 }
436
437 module_init(flow_cache_init_global);