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->memcg_aware;
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 __always_inline struct mem_cgroup *mem_cgroup_from_kmem(void *ptr)
66 if (!memcg_kmem_enabled())
68 page = virt_to_head_page(ptr);
69 return page->mem_cgroup;
72 static inline struct list_lru_one *
73 list_lru_from_kmem(struct list_lru_node *nlru, void *ptr)
75 struct mem_cgroup *memcg;
77 if (!nlru->memcg_lrus)
80 memcg = mem_cgroup_from_kmem(ptr);
84 return list_lru_from_memcg_idx(nlru, memcg_cache_id(memcg));
87 static inline bool list_lru_memcg_aware(struct list_lru *lru)
92 static inline struct list_lru_one *
93 list_lru_from_memcg_idx(struct list_lru_node *nlru, int idx)
98 static inline struct list_lru_one *
99 list_lru_from_kmem(struct list_lru_node *nlru, void *ptr)
103 #endif /* CONFIG_MEMCG_KMEM */
105 bool list_lru_add(struct list_lru *lru, struct list_head *item)
107 int nid = page_to_nid(virt_to_page(item));
108 struct list_lru_node *nlru = &lru->node[nid];
109 struct list_lru_one *l;
111 spin_lock(&nlru->lock);
112 if (list_empty(item)) {
113 l = list_lru_from_kmem(nlru, item);
114 list_add_tail(item, &l->list);
117 spin_unlock(&nlru->lock);
120 spin_unlock(&nlru->lock);
123 EXPORT_SYMBOL_GPL(list_lru_add);
125 bool list_lru_del(struct list_lru *lru, struct list_head *item)
127 int nid = page_to_nid(virt_to_page(item));
128 struct list_lru_node *nlru = &lru->node[nid];
129 struct list_lru_one *l;
131 spin_lock(&nlru->lock);
132 if (!list_empty(item)) {
133 l = list_lru_from_kmem(nlru, item);
137 spin_unlock(&nlru->lock);
140 spin_unlock(&nlru->lock);
143 EXPORT_SYMBOL_GPL(list_lru_del);
145 void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
150 EXPORT_SYMBOL_GPL(list_lru_isolate);
152 void list_lru_isolate_move(struct list_lru_one *list, struct list_head *item,
153 struct list_head *head)
155 list_move(item, head);
158 EXPORT_SYMBOL_GPL(list_lru_isolate_move);
160 static unsigned long __list_lru_count_one(struct list_lru *lru,
161 int nid, int memcg_idx)
163 struct list_lru_node *nlru = &lru->node[nid];
164 struct list_lru_one *l;
167 spin_lock(&nlru->lock);
168 l = list_lru_from_memcg_idx(nlru, memcg_idx);
170 spin_unlock(&nlru->lock);
175 unsigned long list_lru_count_one(struct list_lru *lru,
176 int nid, struct mem_cgroup *memcg)
178 return __list_lru_count_one(lru, nid, memcg_cache_id(memcg));
180 EXPORT_SYMBOL_GPL(list_lru_count_one);
182 unsigned long list_lru_count_node(struct list_lru *lru, int nid)
184 struct list_lru_node *nlru;
186 nlru = &lru->node[nid];
187 return nlru->nr_items;
189 EXPORT_SYMBOL_GPL(list_lru_count_node);
192 __list_lru_walk_one(struct list_lru *lru, int nid, int memcg_idx,
193 list_lru_walk_cb isolate, void *cb_arg,
194 unsigned long *nr_to_walk)
197 struct list_lru_node *nlru = &lru->node[nid];
198 struct list_lru_one *l;
199 struct list_head *item, *n;
200 unsigned long isolated = 0;
202 spin_lock(&nlru->lock);
203 l = list_lru_from_memcg_idx(nlru, memcg_idx);
205 list_for_each_safe(item, n, &l->list) {
209 * decrement nr_to_walk first so that we don't livelock if we
210 * get stuck on large numbesr of LRU_RETRY items
216 ret = isolate(item, l, &nlru->lock, cb_arg);
218 case LRU_REMOVED_RETRY:
219 assert_spin_locked(&nlru->lock);
224 * If the lru lock has been dropped, our list
225 * traversal is now invalid and so we have to
226 * restart from scratch.
228 if (ret == LRU_REMOVED_RETRY)
232 list_move_tail(item, &l->list);
238 * The lru lock has been dropped, our list traversal is
239 * now invalid and so we have to restart from scratch.
241 assert_spin_locked(&nlru->lock);
248 spin_unlock(&nlru->lock);
253 list_lru_walk_one(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
254 list_lru_walk_cb isolate, void *cb_arg,
255 unsigned long *nr_to_walk)
257 return __list_lru_walk_one(lru, nid, memcg_cache_id(memcg),
258 isolate, cb_arg, nr_to_walk);
260 EXPORT_SYMBOL_GPL(list_lru_walk_one);
262 unsigned long list_lru_walk_node(struct list_lru *lru, int nid,
263 list_lru_walk_cb isolate, void *cb_arg,
264 unsigned long *nr_to_walk)
269 isolated += __list_lru_walk_one(lru, nid, -1, isolate, cb_arg,
271 if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
272 for_each_memcg_cache_index(memcg_idx) {
273 isolated += __list_lru_walk_one(lru, nid, memcg_idx,
274 isolate, cb_arg, nr_to_walk);
275 if (*nr_to_walk <= 0)
281 EXPORT_SYMBOL_GPL(list_lru_walk_node);
283 static void init_one_lru(struct list_lru_one *l)
285 INIT_LIST_HEAD(&l->list);
289 #ifdef CONFIG_MEMCG_KMEM
290 static void __memcg_destroy_list_lru_node(struct list_lru_memcg *memcg_lrus,
295 for (i = begin; i < end; i++)
296 kfree(memcg_lrus->lru[i]);
299 static int __memcg_init_list_lru_node(struct list_lru_memcg *memcg_lrus,
304 for (i = begin; i < end; i++) {
305 struct list_lru_one *l;
307 l = kmalloc(sizeof(struct list_lru_one), GFP_KERNEL);
312 memcg_lrus->lru[i] = l;
316 __memcg_destroy_list_lru_node(memcg_lrus, begin, i - 1);
320 static int memcg_init_list_lru_node(struct list_lru_node *nlru)
322 int size = memcg_nr_cache_ids;
324 nlru->memcg_lrus = kmalloc(size * sizeof(void *), GFP_KERNEL);
325 if (!nlru->memcg_lrus)
328 if (__memcg_init_list_lru_node(nlru->memcg_lrus, 0, size)) {
329 kfree(nlru->memcg_lrus);
336 static void memcg_destroy_list_lru_node(struct list_lru_node *nlru)
338 __memcg_destroy_list_lru_node(nlru->memcg_lrus, 0, memcg_nr_cache_ids);
339 kfree(nlru->memcg_lrus);
342 static int memcg_update_list_lru_node(struct list_lru_node *nlru,
343 int old_size, int new_size)
345 struct list_lru_memcg *old, *new;
347 BUG_ON(old_size > new_size);
349 old = nlru->memcg_lrus;
350 new = kmalloc(new_size * sizeof(void *), GFP_KERNEL);
354 if (__memcg_init_list_lru_node(new, old_size, new_size)) {
359 memcpy(new, old, old_size * sizeof(void *));
362 * The lock guarantees that we won't race with a reader
363 * (see list_lru_from_memcg_idx).
365 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
366 * we have to use IRQ-safe primitives here to avoid deadlock.
368 spin_lock_irq(&nlru->lock);
369 nlru->memcg_lrus = new;
370 spin_unlock_irq(&nlru->lock);
376 static void memcg_cancel_update_list_lru_node(struct list_lru_node *nlru,
377 int old_size, int new_size)
379 /* do not bother shrinking the array back to the old size, because we
380 * cannot handle allocation failures here */
381 __memcg_destroy_list_lru_node(nlru->memcg_lrus, old_size, new_size);
384 static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
388 lru->memcg_aware = memcg_aware;
394 if (memcg_init_list_lru_node(&lru->node[i]))
399 for (i = i - 1; i >= 0; i--) {
400 if (!lru->node[i].memcg_lrus)
402 memcg_destroy_list_lru_node(&lru->node[i]);
407 static void memcg_destroy_list_lru(struct list_lru *lru)
411 if (!list_lru_memcg_aware(lru))
415 memcg_destroy_list_lru_node(&lru->node[i]);
418 static int memcg_update_list_lru(struct list_lru *lru,
419 int old_size, int new_size)
423 if (!list_lru_memcg_aware(lru))
427 if (memcg_update_list_lru_node(&lru->node[i],
433 for (i = i - 1; i >= 0; i--) {
434 if (!lru->node[i].memcg_lrus)
437 memcg_cancel_update_list_lru_node(&lru->node[i],
443 static void memcg_cancel_update_list_lru(struct list_lru *lru,
444 int old_size, int new_size)
448 if (!list_lru_memcg_aware(lru))
452 memcg_cancel_update_list_lru_node(&lru->node[i],
456 int memcg_update_all_list_lrus(int new_size)
459 struct list_lru *lru;
460 int old_size = memcg_nr_cache_ids;
462 mutex_lock(&list_lrus_mutex);
463 list_for_each_entry(lru, &list_lrus, list) {
464 ret = memcg_update_list_lru(lru, old_size, new_size);
469 mutex_unlock(&list_lrus_mutex);
472 list_for_each_entry_continue_reverse(lru, &list_lrus, list)
473 memcg_cancel_update_list_lru(lru, old_size, new_size);
477 static void memcg_drain_list_lru_node(struct list_lru_node *nlru,
478 int src_idx, int dst_idx)
480 struct list_lru_one *src, *dst;
483 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
484 * we have to use IRQ-safe primitives here to avoid deadlock.
486 spin_lock_irq(&nlru->lock);
488 src = list_lru_from_memcg_idx(nlru, src_idx);
489 dst = list_lru_from_memcg_idx(nlru, dst_idx);
491 list_splice_init(&src->list, &dst->list);
492 dst->nr_items += src->nr_items;
495 spin_unlock_irq(&nlru->lock);
498 static void memcg_drain_list_lru(struct list_lru *lru,
499 int src_idx, int dst_idx)
503 if (!list_lru_memcg_aware(lru))
507 memcg_drain_list_lru_node(&lru->node[i], src_idx, dst_idx);
510 void memcg_drain_all_list_lrus(int src_idx, int dst_idx)
512 struct list_lru *lru;
514 mutex_lock(&list_lrus_mutex);
515 list_for_each_entry(lru, &list_lrus, list)
516 memcg_drain_list_lru(lru, src_idx, dst_idx);
517 mutex_unlock(&list_lrus_mutex);
520 static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
525 static void memcg_destroy_list_lru(struct list_lru *lru)
528 #endif /* CONFIG_MEMCG_KMEM */
530 int __list_lru_init(struct list_lru *lru, bool memcg_aware,
531 struct lock_class_key *key)
534 size_t size = sizeof(*lru->node) * nr_node_ids;
537 memcg_get_cache_ids();
539 lru->node = kzalloc(size, GFP_KERNEL);
544 spin_lock_init(&lru->node[i].lock);
546 lockdep_set_class(&lru->node[i].lock, key);
547 init_one_lru(&lru->node[i].lru);
550 err = memcg_init_list_lru(lru, memcg_aware);
553 /* Do this so a list_lru_destroy() doesn't crash: */
558 list_lru_register(lru);
560 memcg_put_cache_ids();
563 EXPORT_SYMBOL_GPL(__list_lru_init);
565 void list_lru_destroy(struct list_lru *lru)
567 /* Already destroyed or not yet initialized? */
571 memcg_get_cache_ids();
573 list_lru_unregister(lru);
575 memcg_destroy_list_lru(lru);
579 memcg_put_cache_ids();
581 EXPORT_SYMBOL_GPL(list_lru_destroy);