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 
flow_cache_new_hashrnd(unsigned long arg)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 
flow_entry_valid(struct flow_cache_entry * fle)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 
flow_entry_kill(struct flow_cache_entry * fle)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 
flow_cache_gc_task(struct work_struct * work)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 
flow_cache_queue_garbage(struct flow_cache_percpu * fcp,int deleted,struct list_head * gc_list)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 
__flow_cache_shrink(struct flow_cache * fc,struct flow_cache_percpu * fcp,int shrink_to)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 
flow_cache_shrink(struct flow_cache * fc,struct flow_cache_percpu * fcp)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 
flow_new_hash_rnd(struct flow_cache * fc,struct flow_cache_percpu * fcp)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 
flow_hash_code(struct flow_cache * fc,struct flow_cache_percpu * fcp,const struct flowi * key,size_t keysize)174 static u32 flow_hash_code(struct flow_cache *fc,
175 			  struct flow_cache_percpu *fcp,
176 			  const struct flowi *key,
177 			  size_t keysize)
178 {
179 	const u32 *k = (const u32 *) key;
180 	const u32 length = keysize * sizeof(flow_compare_t) / sizeof(u32);
181 
182 	return jhash2(k, length, fcp->hash_rnd)
183 		& (flow_cache_hash_size(fc) - 1);
184 }
185 
186 /* I hear what you're saying, use memcmp.  But memcmp cannot make
187  * important assumptions that we can here, such as alignment.
188  */
flow_key_compare(const struct flowi * key1,const struct flowi * key2,size_t keysize)189 static int flow_key_compare(const struct flowi *key1, const struct flowi *key2,
190 			    size_t keysize)
191 {
192 	const flow_compare_t *k1, *k1_lim, *k2;
193 
194 	k1 = (const flow_compare_t *) key1;
195 	k1_lim = k1 + keysize;
196 
197 	k2 = (const flow_compare_t *) key2;
198 
199 	do {
200 		if (*k1++ != *k2++)
201 			return 1;
202 	} while (k1 < k1_lim);
203 
204 	return 0;
205 }
206 
207 struct flow_cache_object *
flow_cache_lookup(struct net * net,const struct flowi * key,u16 family,u8 dir,flow_resolve_t resolver,void * ctx)208 flow_cache_lookup(struct net *net, const struct flowi *key, u16 family, u8 dir,
209 		  flow_resolve_t resolver, void *ctx)
210 {
211 	struct flow_cache *fc = &flow_cache_global;
212 	struct flow_cache_percpu *fcp;
213 	struct flow_cache_entry *fle, *tfle;
214 	struct hlist_node *entry;
215 	struct flow_cache_object *flo;
216 	size_t keysize;
217 	unsigned int hash;
218 
219 	local_bh_disable();
220 	fcp = this_cpu_ptr(fc->percpu);
221 
222 	fle = NULL;
223 	flo = NULL;
224 
225 	keysize = flow_key_size(family);
226 	if (!keysize)
227 		goto nocache;
228 
229 	/* Packet really early in init?  Making flow_cache_init a
230 	 * pre-smp initcall would solve this.  --RR */
231 	if (!fcp->hash_table)
232 		goto nocache;
233 
234 	if (fcp->hash_rnd_recalc)
235 		flow_new_hash_rnd(fc, fcp);
236 
237 	hash = flow_hash_code(fc, fcp, key, keysize);
238 	hlist_for_each_entry(tfle, entry, &fcp->hash_table[hash], u.hlist) {
239 		if (tfle->net == net &&
240 		    tfle->family == family &&
241 		    tfle->dir == dir &&
242 		    flow_key_compare(key, &tfle->key, keysize) == 0) {
243 			fle = tfle;
244 			break;
245 		}
246 	}
247 
248 	if (unlikely(!fle)) {
249 		if (fcp->hash_count > fc->high_watermark)
250 			flow_cache_shrink(fc, fcp);
251 
252 		fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
253 		if (fle) {
254 			fle->net = net;
255 			fle->family = family;
256 			fle->dir = dir;
257 			memcpy(&fle->key, key, keysize * sizeof(flow_compare_t));
258 			fle->object = NULL;
259 			hlist_add_head(&fle->u.hlist, &fcp->hash_table[hash]);
260 			fcp->hash_count++;
261 		}
262 	} else if (likely(fle->genid == atomic_read(&flow_cache_genid))) {
263 		flo = fle->object;
264 		if (!flo)
265 			goto ret_object;
266 		flo = flo->ops->get(flo);
267 		if (flo)
268 			goto ret_object;
269 	} else if (fle->object) {
270 	        flo = fle->object;
271 	        flo->ops->delete(flo);
272 	        fle->object = NULL;
273 	}
274 
275 nocache:
276 	flo = NULL;
277 	if (fle) {
278 		flo = fle->object;
279 		fle->object = NULL;
280 	}
281 	flo = resolver(net, key, family, dir, flo, ctx);
282 	if (fle) {
283 		fle->genid = atomic_read(&flow_cache_genid);
284 		if (!IS_ERR(flo))
285 			fle->object = flo;
286 		else
287 			fle->genid--;
288 	} else {
289 		if (flo && !IS_ERR(flo))
290 			flo->ops->delete(flo);
291 	}
292 ret_object:
293 	local_bh_enable();
294 	return flo;
295 }
296 EXPORT_SYMBOL(flow_cache_lookup);
297 
flow_cache_flush_tasklet(unsigned long data)298 static void flow_cache_flush_tasklet(unsigned long data)
299 {
300 	struct flow_flush_info *info = (void *)data;
301 	struct flow_cache *fc = info->cache;
302 	struct flow_cache_percpu *fcp;
303 	struct flow_cache_entry *fle;
304 	struct hlist_node *entry, *tmp;
305 	LIST_HEAD(gc_list);
306 	int i, deleted = 0;
307 
308 	fcp = this_cpu_ptr(fc->percpu);
309 	for (i = 0; i < flow_cache_hash_size(fc); i++) {
310 		hlist_for_each_entry_safe(fle, entry, tmp,
311 					  &fcp->hash_table[i], u.hlist) {
312 			if (flow_entry_valid(fle))
313 				continue;
314 
315 			deleted++;
316 			hlist_del(&fle->u.hlist);
317 			list_add_tail(&fle->u.gc_list, &gc_list);
318 		}
319 	}
320 
321 	flow_cache_queue_garbage(fcp, deleted, &gc_list);
322 
323 	if (atomic_dec_and_test(&info->cpuleft))
324 		complete(&info->completion);
325 }
326 
flow_cache_flush_per_cpu(void * data)327 static void flow_cache_flush_per_cpu(void *data)
328 {
329 	struct flow_flush_info *info = data;
330 	int cpu;
331 	struct tasklet_struct *tasklet;
332 
333 	cpu = smp_processor_id();
334 	tasklet = &per_cpu_ptr(info->cache->percpu, cpu)->flush_tasklet;
335 	tasklet->data = (unsigned long)info;
336 	tasklet_schedule(tasklet);
337 }
338 
flow_cache_flush(void)339 void flow_cache_flush(void)
340 {
341 	struct flow_flush_info info;
342 	static DEFINE_MUTEX(flow_flush_sem);
343 
344 	/* Don't want cpus going down or up during this. */
345 	get_online_cpus();
346 	mutex_lock(&flow_flush_sem);
347 	info.cache = &flow_cache_global;
348 	atomic_set(&info.cpuleft, num_online_cpus());
349 	init_completion(&info.completion);
350 
351 	local_bh_disable();
352 	smp_call_function(flow_cache_flush_per_cpu, &info, 0);
353 	flow_cache_flush_tasklet((unsigned long)&info);
354 	local_bh_enable();
355 
356 	wait_for_completion(&info.completion);
357 	mutex_unlock(&flow_flush_sem);
358 	put_online_cpus();
359 }
360 
flow_cache_flush_task(struct work_struct * work)361 static void flow_cache_flush_task(struct work_struct *work)
362 {
363 	flow_cache_flush();
364 }
365 
366 static DECLARE_WORK(flow_cache_flush_work, flow_cache_flush_task);
367 
flow_cache_flush_deferred(void)368 void flow_cache_flush_deferred(void)
369 {
370 	schedule_work(&flow_cache_flush_work);
371 }
372 
flow_cache_cpu_prepare(struct flow_cache * fc,int cpu)373 static int __cpuinit flow_cache_cpu_prepare(struct flow_cache *fc, int cpu)
374 {
375 	struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
376 	size_t sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
377 
378 	if (!fcp->hash_table) {
379 		fcp->hash_table = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
380 		if (!fcp->hash_table) {
381 			pr_err("NET: failed to allocate flow cache sz %zu\n", sz);
382 			return -ENOMEM;
383 		}
384 		fcp->hash_rnd_recalc = 1;
385 		fcp->hash_count = 0;
386 		tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
387 	}
388 	return 0;
389 }
390 
flow_cache_cpu(struct notifier_block * nfb,unsigned long action,void * hcpu)391 static int __cpuinit flow_cache_cpu(struct notifier_block *nfb,
392 			  unsigned long action,
393 			  void *hcpu)
394 {
395 	struct flow_cache *fc = container_of(nfb, struct flow_cache, hotcpu_notifier);
396 	int res, cpu = (unsigned long) hcpu;
397 	struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
398 
399 	switch (action) {
400 	case CPU_UP_PREPARE:
401 	case CPU_UP_PREPARE_FROZEN:
402 		res = flow_cache_cpu_prepare(fc, cpu);
403 		if (res)
404 			return notifier_from_errno(res);
405 		break;
406 	case CPU_DEAD:
407 	case CPU_DEAD_FROZEN:
408 		__flow_cache_shrink(fc, fcp, 0);
409 		break;
410 	}
411 	return NOTIFY_OK;
412 }
413 
flow_cache_init(struct flow_cache * fc)414 static int __init flow_cache_init(struct flow_cache *fc)
415 {
416 	int i;
417 
418 	fc->hash_shift = 10;
419 	fc->low_watermark = 2 * flow_cache_hash_size(fc);
420 	fc->high_watermark = 4 * flow_cache_hash_size(fc);
421 
422 	fc->percpu = alloc_percpu(struct flow_cache_percpu);
423 	if (!fc->percpu)
424 		return -ENOMEM;
425 
426 	for_each_online_cpu(i) {
427 		if (flow_cache_cpu_prepare(fc, i))
428 			goto err;
429 	}
430 	fc->hotcpu_notifier = (struct notifier_block){
431 		.notifier_call = flow_cache_cpu,
432 	};
433 	register_hotcpu_notifier(&fc->hotcpu_notifier);
434 
435 	setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
436 		    (unsigned long) fc);
437 	fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
438 	add_timer(&fc->rnd_timer);
439 
440 	return 0;
441 
442 err:
443 	for_each_possible_cpu(i) {
444 		struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
445 		kfree(fcp->hash_table);
446 		fcp->hash_table = NULL;
447 	}
448 
449 	free_percpu(fc->percpu);
450 	fc->percpu = NULL;
451 
452 	return -ENOMEM;
453 }
454 
flow_cache_init_global(void)455 static int __init flow_cache_init_global(void)
456 {
457 	flow_cachep = kmem_cache_create("flow_cache",
458 					sizeof(struct flow_cache_entry),
459 					0, SLAB_PANIC, NULL);
460 
461 	return flow_cache_init(&flow_cache_global);
462 }
463 
464 module_init(flow_cache_init_global);
465