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 #if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
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 && !CONFIG_SLOB */
42 #if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
43 static inline bool list_lru_memcg_aware(struct list_lru
*lru
)
46 * This needs node 0 to be always present, even
47 * in the systems supporting sparse numa ids.
49 return !!lru
->node
[0].memcg_lrus
;
52 static inline struct list_lru_one
*
53 list_lru_from_memcg_idx(struct list_lru_node
*nlru
, int idx
)
55 struct list_lru_memcg
*memcg_lrus
;
57 * Either lock or RCU protects the array of per cgroup lists
58 * from relocation (see memcg_update_list_lru_node).
60 memcg_lrus
= rcu_dereference_check(nlru
->memcg_lrus
,
61 lockdep_is_held(&nlru
->lock
));
62 if (memcg_lrus
&& idx
>= 0)
63 return memcg_lrus
->lru
[idx
];
67 static __always_inline
struct mem_cgroup
*mem_cgroup_from_kmem(void *ptr
)
71 if (!memcg_kmem_enabled())
73 page
= virt_to_head_page(ptr
);
74 return page
->mem_cgroup
;
77 static inline struct list_lru_one
*
78 list_lru_from_kmem(struct list_lru_node
*nlru
, void *ptr
)
80 struct mem_cgroup
*memcg
;
82 if (!nlru
->memcg_lrus
)
85 memcg
= mem_cgroup_from_kmem(ptr
);
89 return list_lru_from_memcg_idx(nlru
, memcg_cache_id(memcg
));
92 static inline bool list_lru_memcg_aware(struct list_lru
*lru
)
97 static inline struct list_lru_one
*
98 list_lru_from_memcg_idx(struct list_lru_node
*nlru
, int idx
)
103 static inline struct list_lru_one
*
104 list_lru_from_kmem(struct list_lru_node
*nlru
, void *ptr
)
108 #endif /* CONFIG_MEMCG && !CONFIG_SLOB */
110 bool list_lru_add(struct list_lru
*lru
, struct list_head
*item
)
112 int nid
= page_to_nid(virt_to_page(item
));
113 struct list_lru_node
*nlru
= &lru
->node
[nid
];
114 struct list_lru_one
*l
;
116 spin_lock(&nlru
->lock
);
117 if (list_empty(item
)) {
118 l
= list_lru_from_kmem(nlru
, item
);
119 list_add_tail(item
, &l
->list
);
122 spin_unlock(&nlru
->lock
);
125 spin_unlock(&nlru
->lock
);
128 EXPORT_SYMBOL_GPL(list_lru_add
);
130 bool list_lru_del(struct list_lru
*lru
, struct list_head
*item
)
132 int nid
= page_to_nid(virt_to_page(item
));
133 struct list_lru_node
*nlru
= &lru
->node
[nid
];
134 struct list_lru_one
*l
;
136 spin_lock(&nlru
->lock
);
137 if (!list_empty(item
)) {
138 l
= list_lru_from_kmem(nlru
, item
);
142 spin_unlock(&nlru
->lock
);
145 spin_unlock(&nlru
->lock
);
148 EXPORT_SYMBOL_GPL(list_lru_del
);
150 void list_lru_isolate(struct list_lru_one
*list
, struct list_head
*item
)
155 EXPORT_SYMBOL_GPL(list_lru_isolate
);
157 void list_lru_isolate_move(struct list_lru_one
*list
, struct list_head
*item
,
158 struct list_head
*head
)
160 list_move(item
, head
);
163 EXPORT_SYMBOL_GPL(list_lru_isolate_move
);
165 static unsigned long __list_lru_count_one(struct list_lru
*lru
,
166 int nid
, int memcg_idx
)
168 struct list_lru_node
*nlru
= &lru
->node
[nid
];
169 struct list_lru_one
*l
;
173 l
= list_lru_from_memcg_idx(nlru
, memcg_idx
);
180 unsigned long list_lru_count_one(struct list_lru
*lru
,
181 int nid
, struct mem_cgroup
*memcg
)
183 return __list_lru_count_one(lru
, nid
, memcg_cache_id(memcg
));
185 EXPORT_SYMBOL_GPL(list_lru_count_one
);
187 unsigned long list_lru_count_node(struct list_lru
*lru
, int nid
)
189 struct list_lru_node
*nlru
;
191 nlru
= &lru
->node
[nid
];
192 return nlru
->nr_items
;
194 EXPORT_SYMBOL_GPL(list_lru_count_node
);
197 __list_lru_walk_one(struct list_lru
*lru
, int nid
, int memcg_idx
,
198 list_lru_walk_cb isolate
, void *cb_arg
,
199 unsigned long *nr_to_walk
)
202 struct list_lru_node
*nlru
= &lru
->node
[nid
];
203 struct list_lru_one
*l
;
204 struct list_head
*item
, *n
;
205 unsigned long isolated
= 0;
207 spin_lock(&nlru
->lock
);
208 l
= list_lru_from_memcg_idx(nlru
, memcg_idx
);
210 list_for_each_safe(item
, n
, &l
->list
) {
214 * decrement nr_to_walk first so that we don't livelock if we
215 * get stuck on large numbesr of LRU_RETRY items
221 ret
= isolate(item
, l
, &nlru
->lock
, cb_arg
);
223 case LRU_REMOVED_RETRY
:
224 assert_spin_locked(&nlru
->lock
);
230 * If the lru lock has been dropped, our list
231 * traversal is now invalid and so we have to
232 * restart from scratch.
234 if (ret
== LRU_REMOVED_RETRY
)
238 list_move_tail(item
, &l
->list
);
244 * The lru lock has been dropped, our list traversal is
245 * now invalid and so we have to restart from scratch.
247 assert_spin_locked(&nlru
->lock
);
254 spin_unlock(&nlru
->lock
);
259 list_lru_walk_one(struct list_lru
*lru
, int nid
, struct mem_cgroup
*memcg
,
260 list_lru_walk_cb isolate
, void *cb_arg
,
261 unsigned long *nr_to_walk
)
263 return __list_lru_walk_one(lru
, nid
, memcg_cache_id(memcg
),
264 isolate
, cb_arg
, nr_to_walk
);
266 EXPORT_SYMBOL_GPL(list_lru_walk_one
);
268 unsigned long list_lru_walk_node(struct list_lru
*lru
, int nid
,
269 list_lru_walk_cb isolate
, void *cb_arg
,
270 unsigned long *nr_to_walk
)
275 isolated
+= __list_lru_walk_one(lru
, nid
, -1, isolate
, cb_arg
,
277 if (*nr_to_walk
> 0 && list_lru_memcg_aware(lru
)) {
278 for_each_memcg_cache_index(memcg_idx
) {
279 isolated
+= __list_lru_walk_one(lru
, nid
, memcg_idx
,
280 isolate
, cb_arg
, nr_to_walk
);
281 if (*nr_to_walk
<= 0)
287 EXPORT_SYMBOL_GPL(list_lru_walk_node
);
289 static void init_one_lru(struct list_lru_one
*l
)
291 INIT_LIST_HEAD(&l
->list
);
295 #if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
296 static void __memcg_destroy_list_lru_node(struct list_lru_memcg
*memcg_lrus
,
301 for (i
= begin
; i
< end
; i
++)
302 kfree(memcg_lrus
->lru
[i
]);
305 static int __memcg_init_list_lru_node(struct list_lru_memcg
*memcg_lrus
,
310 for (i
= begin
; i
< end
; i
++) {
311 struct list_lru_one
*l
;
313 l
= kmalloc(sizeof(struct list_lru_one
), GFP_KERNEL
);
318 memcg_lrus
->lru
[i
] = l
;
322 __memcg_destroy_list_lru_node(memcg_lrus
, begin
, i
- 1);
326 static int memcg_init_list_lru_node(struct list_lru_node
*nlru
)
328 struct list_lru_memcg
*memcg_lrus
;
329 int size
= memcg_nr_cache_ids
;
331 memcg_lrus
= kvmalloc(sizeof(*memcg_lrus
) +
332 size
* sizeof(void *), GFP_KERNEL
);
336 if (__memcg_init_list_lru_node(memcg_lrus
, 0, size
)) {
340 RCU_INIT_POINTER(nlru
->memcg_lrus
, memcg_lrus
);
345 static void memcg_destroy_list_lru_node(struct list_lru_node
*nlru
)
347 struct list_lru_memcg
*memcg_lrus
;
349 * This is called when shrinker has already been unregistered,
350 * and nobody can use it. So, there is no need to use kvfree_rcu().
352 memcg_lrus
= rcu_dereference_protected(nlru
->memcg_lrus
, true);
353 __memcg_destroy_list_lru_node(memcg_lrus
, 0, memcg_nr_cache_ids
);
357 static void kvfree_rcu(struct rcu_head
*head
)
359 struct list_lru_memcg
*mlru
;
361 mlru
= container_of(head
, struct list_lru_memcg
, rcu
);
365 static int memcg_update_list_lru_node(struct list_lru_node
*nlru
,
366 int old_size
, int new_size
)
368 struct list_lru_memcg
*old
, *new;
370 BUG_ON(old_size
> new_size
);
372 old
= rcu_dereference_protected(nlru
->memcg_lrus
,
373 lockdep_is_held(&list_lrus_mutex
));
374 new = kvmalloc(sizeof(*new) + new_size
* sizeof(void *), GFP_KERNEL
);
378 if (__memcg_init_list_lru_node(new, old_size
, new_size
)) {
383 memcpy(&new->lru
, &old
->lru
, old_size
* sizeof(void *));
386 * The locking below allows readers that hold nlru->lock avoid taking
387 * rcu_read_lock (see list_lru_from_memcg_idx).
389 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
390 * we have to use IRQ-safe primitives here to avoid deadlock.
392 spin_lock_irq(&nlru
->lock
);
393 rcu_assign_pointer(nlru
->memcg_lrus
, new);
394 spin_unlock_irq(&nlru
->lock
);
396 call_rcu(&old
->rcu
, kvfree_rcu
);
400 static void memcg_cancel_update_list_lru_node(struct list_lru_node
*nlru
,
401 int old_size
, int new_size
)
403 struct list_lru_memcg
*memcg_lrus
;
405 memcg_lrus
= rcu_dereference_protected(nlru
->memcg_lrus
,
406 lockdep_is_held(&list_lrus_mutex
));
407 /* do not bother shrinking the array back to the old size, because we
408 * cannot handle allocation failures here */
409 __memcg_destroy_list_lru_node(memcg_lrus
, old_size
, new_size
);
412 static int memcg_init_list_lru(struct list_lru
*lru
, bool memcg_aware
)
420 if (memcg_init_list_lru_node(&lru
->node
[i
]))
425 for (i
= i
- 1; i
>= 0; i
--) {
426 if (!lru
->node
[i
].memcg_lrus
)
428 memcg_destroy_list_lru_node(&lru
->node
[i
]);
433 static void memcg_destroy_list_lru(struct list_lru
*lru
)
437 if (!list_lru_memcg_aware(lru
))
441 memcg_destroy_list_lru_node(&lru
->node
[i
]);
444 static int memcg_update_list_lru(struct list_lru
*lru
,
445 int old_size
, int new_size
)
449 if (!list_lru_memcg_aware(lru
))
453 if (memcg_update_list_lru_node(&lru
->node
[i
],
459 for (i
= i
- 1; i
>= 0; i
--) {
460 if (!lru
->node
[i
].memcg_lrus
)
463 memcg_cancel_update_list_lru_node(&lru
->node
[i
],
469 static void memcg_cancel_update_list_lru(struct list_lru
*lru
,
470 int old_size
, int new_size
)
474 if (!list_lru_memcg_aware(lru
))
478 memcg_cancel_update_list_lru_node(&lru
->node
[i
],
482 int memcg_update_all_list_lrus(int new_size
)
485 struct list_lru
*lru
;
486 int old_size
= memcg_nr_cache_ids
;
488 mutex_lock(&list_lrus_mutex
);
489 list_for_each_entry(lru
, &list_lrus
, list
) {
490 ret
= memcg_update_list_lru(lru
, old_size
, new_size
);
495 mutex_unlock(&list_lrus_mutex
);
498 list_for_each_entry_continue_reverse(lru
, &list_lrus
, list
)
499 memcg_cancel_update_list_lru(lru
, old_size
, new_size
);
503 static void memcg_drain_list_lru_node(struct list_lru_node
*nlru
,
504 int src_idx
, int dst_idx
)
506 struct list_lru_one
*src
, *dst
;
509 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
510 * we have to use IRQ-safe primitives here to avoid deadlock.
512 spin_lock_irq(&nlru
->lock
);
514 src
= list_lru_from_memcg_idx(nlru
, src_idx
);
515 dst
= list_lru_from_memcg_idx(nlru
, dst_idx
);
517 list_splice_init(&src
->list
, &dst
->list
);
518 dst
->nr_items
+= src
->nr_items
;
521 spin_unlock_irq(&nlru
->lock
);
524 static void memcg_drain_list_lru(struct list_lru
*lru
,
525 int src_idx
, int dst_idx
)
529 if (!list_lru_memcg_aware(lru
))
533 memcg_drain_list_lru_node(&lru
->node
[i
], src_idx
, dst_idx
);
536 void memcg_drain_all_list_lrus(int src_idx
, int dst_idx
)
538 struct list_lru
*lru
;
540 mutex_lock(&list_lrus_mutex
);
541 list_for_each_entry(lru
, &list_lrus
, list
)
542 memcg_drain_list_lru(lru
, src_idx
, dst_idx
);
543 mutex_unlock(&list_lrus_mutex
);
546 static int memcg_init_list_lru(struct list_lru
*lru
, bool memcg_aware
)
551 static void memcg_destroy_list_lru(struct list_lru
*lru
)
554 #endif /* CONFIG_MEMCG && !CONFIG_SLOB */
556 int __list_lru_init(struct list_lru
*lru
, bool memcg_aware
,
557 struct lock_class_key
*key
)
560 size_t size
= sizeof(*lru
->node
) * nr_node_ids
;
563 memcg_get_cache_ids();
565 lru
->node
= kzalloc(size
, GFP_KERNEL
);
570 spin_lock_init(&lru
->node
[i
].lock
);
572 lockdep_set_class(&lru
->node
[i
].lock
, key
);
573 init_one_lru(&lru
->node
[i
].lru
);
576 err
= memcg_init_list_lru(lru
, memcg_aware
);
579 /* Do this so a list_lru_destroy() doesn't crash: */
584 list_lru_register(lru
);
586 memcg_put_cache_ids();
589 EXPORT_SYMBOL_GPL(__list_lru_init
);
591 void list_lru_destroy(struct list_lru
*lru
)
593 /* Already destroyed or not yet initialized? */
597 memcg_get_cache_ids();
599 list_lru_unregister(lru
);
601 memcg_destroy_list_lru(lru
);
605 memcg_put_cache_ids();
607 EXPORT_SYMBOL_GPL(list_lru_destroy
);