2 * Copyright (c) 2013 Red Hat, Inc. and Parallels Inc. All rights reserved.
3 * Authors: David Chinner and Glauber Costa
5 * Generic LRU infrastructure
7 #include <linux/kernel.h>
8 #include <linux/module.h>
10 #include <linux/list_lru.h>
11 #include <linux/slab.h>
12 #include <linux/mutex.h>
13 #include <linux/memcontrol.h>
15 #ifdef CONFIG_MEMCG_KMEM
16 static LIST_HEAD(list_lrus);
17 static DEFINE_MUTEX(list_lrus_mutex);
19 static void list_lru_register(struct list_lru *lru)
21 mutex_lock(&list_lrus_mutex);
22 list_add(&lru->list, &list_lrus);
23 mutex_unlock(&list_lrus_mutex);
26 static void list_lru_unregister(struct list_lru *lru)
28 mutex_lock(&list_lrus_mutex);
30 mutex_unlock(&list_lrus_mutex);
33 static void list_lru_register(struct list_lru *lru)
37 static void list_lru_unregister(struct list_lru *lru)
40 #endif /* CONFIG_MEMCG_KMEM */
42 #ifdef CONFIG_MEMCG_KMEM
43 static inline bool list_lru_memcg_aware(struct list_lru *lru)
45 return !!lru->node[0].memcg_lrus;
48 static inline struct list_lru_one *
49 list_lru_from_memcg_idx(struct list_lru_node *nlru, int idx)
52 * The lock protects the array of per cgroup lists from relocation
53 * (see memcg_update_list_lru_node).
55 lockdep_assert_held(&nlru->lock);
56 if (nlru->memcg_lrus && idx >= 0)
57 return nlru->memcg_lrus->lru[idx];
62 static inline struct list_lru_one *
63 list_lru_from_kmem(struct list_lru_node *nlru, void *ptr)
65 struct mem_cgroup *memcg;
67 if (!nlru->memcg_lrus)
70 memcg = mem_cgroup_from_kmem(ptr);
74 return list_lru_from_memcg_idx(nlru, memcg_cache_id(memcg));
77 static inline bool list_lru_memcg_aware(struct list_lru *lru)
82 static inline struct list_lru_one *
83 list_lru_from_memcg_idx(struct list_lru_node *nlru, int idx)
88 static inline struct list_lru_one *
89 list_lru_from_kmem(struct list_lru_node *nlru, void *ptr)
93 #endif /* CONFIG_MEMCG_KMEM */
95 bool list_lru_add(struct list_lru *lru, struct list_head *item)
97 int nid = page_to_nid(virt_to_page(item));
98 struct list_lru_node *nlru = &lru->node[nid];
99 struct list_lru_one *l;
101 spin_lock(&nlru->lock);
102 l = list_lru_from_kmem(nlru, item);
103 WARN_ON_ONCE(l->nr_items < 0);
104 if (list_empty(item)) {
105 list_add_tail(item, &l->list);
107 spin_unlock(&nlru->lock);
110 spin_unlock(&nlru->lock);
113 EXPORT_SYMBOL_GPL(list_lru_add);
115 bool list_lru_del(struct list_lru *lru, struct list_head *item)
117 int nid = page_to_nid(virt_to_page(item));
118 struct list_lru_node *nlru = &lru->node[nid];
119 struct list_lru_one *l;
121 spin_lock(&nlru->lock);
122 l = list_lru_from_kmem(nlru, item);
123 if (!list_empty(item)) {
126 WARN_ON_ONCE(l->nr_items < 0);
127 spin_unlock(&nlru->lock);
130 spin_unlock(&nlru->lock);
133 EXPORT_SYMBOL_GPL(list_lru_del);
135 static unsigned long __list_lru_count_one(struct list_lru *lru,
136 int nid, int memcg_idx)
138 struct list_lru_node *nlru = &lru->node[nid];
139 struct list_lru_one *l;
142 spin_lock(&nlru->lock);
143 l = list_lru_from_memcg_idx(nlru, memcg_idx);
144 WARN_ON_ONCE(l->nr_items < 0);
146 spin_unlock(&nlru->lock);
151 unsigned long list_lru_count_one(struct list_lru *lru,
152 int nid, struct mem_cgroup *memcg)
154 return __list_lru_count_one(lru, nid, memcg_cache_id(memcg));
156 EXPORT_SYMBOL_GPL(list_lru_count_one);
158 unsigned long list_lru_count_node(struct list_lru *lru, int nid)
163 count += __list_lru_count_one(lru, nid, -1);
164 if (list_lru_memcg_aware(lru)) {
165 for_each_memcg_cache_index(memcg_idx)
166 count += __list_lru_count_one(lru, nid, memcg_idx);
170 EXPORT_SYMBOL_GPL(list_lru_count_node);
173 __list_lru_walk_one(struct list_lru *lru, int nid, int memcg_idx,
174 list_lru_walk_cb isolate, void *cb_arg,
175 unsigned long *nr_to_walk)
178 struct list_lru_node *nlru = &lru->node[nid];
179 struct list_lru_one *l;
180 struct list_head *item, *n;
181 unsigned long isolated = 0;
183 spin_lock(&nlru->lock);
184 l = list_lru_from_memcg_idx(nlru, memcg_idx);
186 list_for_each_safe(item, n, &l->list) {
190 * decrement nr_to_walk first so that we don't livelock if we
191 * get stuck on large numbesr of LRU_RETRY items
197 ret = isolate(item, &nlru->lock, cb_arg);
199 case LRU_REMOVED_RETRY:
200 assert_spin_locked(&nlru->lock);
203 WARN_ON_ONCE(l->nr_items < 0);
206 * If the lru lock has been dropped, our list
207 * traversal is now invalid and so we have to
208 * restart from scratch.
210 if (ret == LRU_REMOVED_RETRY)
214 list_move_tail(item, &l->list);
220 * The lru lock has been dropped, our list traversal is
221 * now invalid and so we have to restart from scratch.
223 assert_spin_locked(&nlru->lock);
230 spin_unlock(&nlru->lock);
235 list_lru_walk_one(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
236 list_lru_walk_cb isolate, void *cb_arg,
237 unsigned long *nr_to_walk)
239 return __list_lru_walk_one(lru, nid, memcg_cache_id(memcg),
240 isolate, cb_arg, nr_to_walk);
242 EXPORT_SYMBOL_GPL(list_lru_walk_one);
244 unsigned long list_lru_walk_node(struct list_lru *lru, int nid,
245 list_lru_walk_cb isolate, void *cb_arg,
246 unsigned long *nr_to_walk)
251 isolated += __list_lru_walk_one(lru, nid, -1, isolate, cb_arg,
253 if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
254 for_each_memcg_cache_index(memcg_idx) {
255 isolated += __list_lru_walk_one(lru, nid, memcg_idx,
256 isolate, cb_arg, nr_to_walk);
257 if (*nr_to_walk <= 0)
263 EXPORT_SYMBOL_GPL(list_lru_walk_node);
265 static void init_one_lru(struct list_lru_one *l)
267 INIT_LIST_HEAD(&l->list);
271 #ifdef CONFIG_MEMCG_KMEM
272 static void __memcg_destroy_list_lru_node(struct list_lru_memcg *memcg_lrus,
277 for (i = begin; i < end; i++)
278 kfree(memcg_lrus->lru[i]);
281 static int __memcg_init_list_lru_node(struct list_lru_memcg *memcg_lrus,
286 for (i = begin; i < end; i++) {
287 struct list_lru_one *l;
289 l = kmalloc(sizeof(struct list_lru_one), GFP_KERNEL);
294 memcg_lrus->lru[i] = l;
298 __memcg_destroy_list_lru_node(memcg_lrus, begin, i - 1);
302 static int memcg_init_list_lru_node(struct list_lru_node *nlru)
304 int size = memcg_nr_cache_ids;
306 nlru->memcg_lrus = kmalloc(size * sizeof(void *), GFP_KERNEL);
307 if (!nlru->memcg_lrus)
310 if (__memcg_init_list_lru_node(nlru->memcg_lrus, 0, size)) {
311 kfree(nlru->memcg_lrus);
318 static void memcg_destroy_list_lru_node(struct list_lru_node *nlru)
320 __memcg_destroy_list_lru_node(nlru->memcg_lrus, 0, memcg_nr_cache_ids);
321 kfree(nlru->memcg_lrus);
324 static int memcg_update_list_lru_node(struct list_lru_node *nlru,
325 int old_size, int new_size)
327 struct list_lru_memcg *old, *new;
329 BUG_ON(old_size > new_size);
331 old = nlru->memcg_lrus;
332 new = kmalloc(new_size * sizeof(void *), GFP_KERNEL);
336 if (__memcg_init_list_lru_node(new, old_size, new_size)) {
341 memcpy(new, old, old_size * sizeof(void *));
344 * The lock guarantees that we won't race with a reader
345 * (see list_lru_from_memcg_idx).
347 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
348 * we have to use IRQ-safe primitives here to avoid deadlock.
350 spin_lock_irq(&nlru->lock);
351 nlru->memcg_lrus = new;
352 spin_unlock_irq(&nlru->lock);
358 static void memcg_cancel_update_list_lru_node(struct list_lru_node *nlru,
359 int old_size, int new_size)
361 /* do not bother shrinking the array back to the old size, because we
362 * cannot handle allocation failures here */
363 __memcg_destroy_list_lru_node(nlru->memcg_lrus, old_size, new_size);
366 static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
370 for (i = 0; i < nr_node_ids; i++) {
372 lru->node[i].memcg_lrus = NULL;
373 else if (memcg_init_list_lru_node(&lru->node[i]))
378 for (i = i - 1; i >= 0; i--)
379 memcg_destroy_list_lru_node(&lru->node[i]);
383 static void memcg_destroy_list_lru(struct list_lru *lru)
387 if (!list_lru_memcg_aware(lru))
390 for (i = 0; i < nr_node_ids; i++)
391 memcg_destroy_list_lru_node(&lru->node[i]);
394 static int memcg_update_list_lru(struct list_lru *lru,
395 int old_size, int new_size)
399 if (!list_lru_memcg_aware(lru))
402 for (i = 0; i < nr_node_ids; i++) {
403 if (memcg_update_list_lru_node(&lru->node[i],
409 for (i = i - 1; i >= 0; i--)
410 memcg_cancel_update_list_lru_node(&lru->node[i],
415 static void memcg_cancel_update_list_lru(struct list_lru *lru,
416 int old_size, int new_size)
420 if (!list_lru_memcg_aware(lru))
423 for (i = 0; i < nr_node_ids; i++)
424 memcg_cancel_update_list_lru_node(&lru->node[i],
428 int memcg_update_all_list_lrus(int new_size)
431 struct list_lru *lru;
432 int old_size = memcg_nr_cache_ids;
434 mutex_lock(&list_lrus_mutex);
435 list_for_each_entry(lru, &list_lrus, list) {
436 ret = memcg_update_list_lru(lru, old_size, new_size);
441 mutex_unlock(&list_lrus_mutex);
444 list_for_each_entry_continue_reverse(lru, &list_lrus, list)
445 memcg_cancel_update_list_lru(lru, old_size, new_size);
449 static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
454 static void memcg_destroy_list_lru(struct list_lru *lru)
457 #endif /* CONFIG_MEMCG_KMEM */
459 int __list_lru_init(struct list_lru *lru, bool memcg_aware,
460 struct lock_class_key *key)
463 size_t size = sizeof(*lru->node) * nr_node_ids;
466 memcg_get_cache_ids();
468 lru->node = kzalloc(size, GFP_KERNEL);
472 for (i = 0; i < nr_node_ids; i++) {
473 spin_lock_init(&lru->node[i].lock);
475 lockdep_set_class(&lru->node[i].lock, key);
476 init_one_lru(&lru->node[i].lru);
479 err = memcg_init_list_lru(lru, memcg_aware);
485 list_lru_register(lru);
487 memcg_put_cache_ids();
490 EXPORT_SYMBOL_GPL(__list_lru_init);
492 void list_lru_destroy(struct list_lru *lru)
494 /* Already destroyed or not yet initialized? */
498 memcg_get_cache_ids();
500 list_lru_unregister(lru);
502 memcg_destroy_list_lru(lru);
506 memcg_put_cache_ids();
508 EXPORT_SYMBOL_GPL(list_lru_destroy);