Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (c) 2013 Red Hat, Inc. and Parallels Inc. All rights reserved.
4 * Authors: David Chinner and Glauber Costa
5 *
6 * Generic LRU infrastructure
7 */
8#include <linux/kernel.h>
9#include <linux/module.h>
10#include <linux/mm.h>
11#include <linux/list_lru.h>
12#include <linux/slab.h>
13#include <linux/mutex.h>
14#include <linux/memcontrol.h>
15#include "slab.h"
16
17#ifdef CONFIG_MEMCG_KMEM
18static LIST_HEAD(list_lrus);
19static DEFINE_MUTEX(list_lrus_mutex);
20
21static void list_lru_register(struct list_lru *lru)
22{
23 mutex_lock(&list_lrus_mutex);
24 list_add(&lru->list, &list_lrus);
25 mutex_unlock(&list_lrus_mutex);
26}
27
28static void list_lru_unregister(struct list_lru *lru)
29{
30 mutex_lock(&list_lrus_mutex);
31 list_del(&lru->list);
32 mutex_unlock(&list_lrus_mutex);
33}
34
35static int lru_shrinker_id(struct list_lru *lru)
36{
37 return lru->shrinker_id;
38}
39
40static inline bool list_lru_memcg_aware(struct list_lru *lru)
41{
42 return lru->memcg_aware;
43}
44
45static inline struct list_lru_one *
46list_lru_from_memcg_idx(struct list_lru_node *nlru, int idx)
47{
48 struct list_lru_memcg *memcg_lrus;
49 /*
50 * Either lock or RCU protects the array of per cgroup lists
51 * from relocation (see memcg_update_list_lru_node).
52 */
53 memcg_lrus = rcu_dereference_check(nlru->memcg_lrus,
54 lockdep_is_held(&nlru->lock));
55 if (memcg_lrus && idx >= 0)
56 return memcg_lrus->lru[idx];
57 return &nlru->lru;
58}
59
60static inline struct list_lru_one *
61list_lru_from_kmem(struct list_lru_node *nlru, void *ptr,
62 struct mem_cgroup **memcg_ptr)
63{
64 struct list_lru_one *l = &nlru->lru;
65 struct mem_cgroup *memcg = NULL;
66
67 if (!nlru->memcg_lrus)
68 goto out;
69
70 memcg = mem_cgroup_from_obj(ptr);
71 if (!memcg)
72 goto out;
73
74 l = list_lru_from_memcg_idx(nlru, memcg_cache_id(memcg));
75out:
76 if (memcg_ptr)
77 *memcg_ptr = memcg;
78 return l;
79}
80#else
81static void list_lru_register(struct list_lru *lru)
82{
83}
84
85static void list_lru_unregister(struct list_lru *lru)
86{
87}
88
89static int lru_shrinker_id(struct list_lru *lru)
90{
91 return -1;
92}
93
94static inline bool list_lru_memcg_aware(struct list_lru *lru)
95{
96 return false;
97}
98
99static inline struct list_lru_one *
100list_lru_from_memcg_idx(struct list_lru_node *nlru, int idx)
101{
102 return &nlru->lru;
103}
104
105static inline struct list_lru_one *
106list_lru_from_kmem(struct list_lru_node *nlru, void *ptr,
107 struct mem_cgroup **memcg_ptr)
108{
109 if (memcg_ptr)
110 *memcg_ptr = NULL;
111 return &nlru->lru;
112}
113#endif /* CONFIG_MEMCG_KMEM */
114
115bool list_lru_add(struct list_lru *lru, struct list_head *item)
116{
117 int nid = page_to_nid(virt_to_page(item));
118 struct list_lru_node *nlru = &lru->node[nid];
119 struct mem_cgroup *memcg;
120 struct list_lru_one *l;
121
122 spin_lock(&nlru->lock);
123 if (list_empty(item)) {
124 l = list_lru_from_kmem(nlru, item, &memcg);
125 list_add_tail(item, &l->list);
126 /* Set shrinker bit if the first element was added */
127 if (!l->nr_items++)
128 memcg_set_shrinker_bit(memcg, nid,
129 lru_shrinker_id(lru));
130 nlru->nr_items++;
131 spin_unlock(&nlru->lock);
132 return true;
133 }
134 spin_unlock(&nlru->lock);
135 return false;
136}
137EXPORT_SYMBOL_GPL(list_lru_add);
138
139bool list_lru_del(struct list_lru *lru, struct list_head *item)
140{
141 int nid = page_to_nid(virt_to_page(item));
142 struct list_lru_node *nlru = &lru->node[nid];
143 struct list_lru_one *l;
144
145 spin_lock(&nlru->lock);
146 if (!list_empty(item)) {
147 l = list_lru_from_kmem(nlru, item, NULL);
148 list_del_init(item);
149 l->nr_items--;
150 nlru->nr_items--;
151 spin_unlock(&nlru->lock);
152 return true;
153 }
154 spin_unlock(&nlru->lock);
155 return false;
156}
157EXPORT_SYMBOL_GPL(list_lru_del);
158
159void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
160{
161 list_del_init(item);
162 list->nr_items--;
163}
164EXPORT_SYMBOL_GPL(list_lru_isolate);
165
166void list_lru_isolate_move(struct list_lru_one *list, struct list_head *item,
167 struct list_head *head)
168{
169 list_move(item, head);
170 list->nr_items--;
171}
172EXPORT_SYMBOL_GPL(list_lru_isolate_move);
173
174unsigned long list_lru_count_one(struct list_lru *lru,
175 int nid, struct mem_cgroup *memcg)
176{
177 struct list_lru_node *nlru = &lru->node[nid];
178 struct list_lru_one *l;
179 unsigned long count;
180
181 rcu_read_lock();
182 l = list_lru_from_memcg_idx(nlru, memcg_cache_id(memcg));
183 count = READ_ONCE(l->nr_items);
184 rcu_read_unlock();
185
186 return count;
187}
188EXPORT_SYMBOL_GPL(list_lru_count_one);
189
190unsigned long list_lru_count_node(struct list_lru *lru, int nid)
191{
192 struct list_lru_node *nlru;
193
194 nlru = &lru->node[nid];
195 return nlru->nr_items;
196}
197EXPORT_SYMBOL_GPL(list_lru_count_node);
198
199static unsigned long
200__list_lru_walk_one(struct list_lru_node *nlru, int memcg_idx,
201 list_lru_walk_cb isolate, void *cb_arg,
202 unsigned long *nr_to_walk)
203{
204
205 struct list_lru_one *l;
206 struct list_head *item, *n;
207 unsigned long isolated = 0;
208
209 l = list_lru_from_memcg_idx(nlru, memcg_idx);
210restart:
211 list_for_each_safe(item, n, &l->list) {
212 enum lru_status ret;
213
214 /*
215 * decrement nr_to_walk first so that we don't livelock if we
216 * get stuck on large numbers of LRU_RETRY items
217 */
218 if (!*nr_to_walk)
219 break;
220 --*nr_to_walk;
221
222 ret = isolate(item, l, &nlru->lock, cb_arg);
223 switch (ret) {
224 case LRU_REMOVED_RETRY:
225 assert_spin_locked(&nlru->lock);
226 fallthrough;
227 case LRU_REMOVED:
228 isolated++;
229 nlru->nr_items--;
230 /*
231 * If the lru lock has been dropped, our list
232 * traversal is now invalid and so we have to
233 * restart from scratch.
234 */
235 if (ret == LRU_REMOVED_RETRY)
236 goto restart;
237 break;
238 case LRU_ROTATE:
239 list_move_tail(item, &l->list);
240 break;
241 case LRU_SKIP:
242 break;
243 case LRU_RETRY:
244 /*
245 * The lru lock has been dropped, our list traversal is
246 * now invalid and so we have to restart from scratch.
247 */
248 assert_spin_locked(&nlru->lock);
249 goto restart;
250 default:
251 BUG();
252 }
253 }
254 return isolated;
255}
256
257unsigned long
258list_lru_walk_one(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
259 list_lru_walk_cb isolate, void *cb_arg,
260 unsigned long *nr_to_walk)
261{
262 struct list_lru_node *nlru = &lru->node[nid];
263 unsigned long ret;
264
265 spin_lock(&nlru->lock);
266 ret = __list_lru_walk_one(nlru, memcg_cache_id(memcg), isolate, cb_arg,
267 nr_to_walk);
268 spin_unlock(&nlru->lock);
269 return ret;
270}
271EXPORT_SYMBOL_GPL(list_lru_walk_one);
272
273unsigned long
274list_lru_walk_one_irq(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
275 list_lru_walk_cb isolate, void *cb_arg,
276 unsigned long *nr_to_walk)
277{
278 struct list_lru_node *nlru = &lru->node[nid];
279 unsigned long ret;
280
281 spin_lock_irq(&nlru->lock);
282 ret = __list_lru_walk_one(nlru, memcg_cache_id(memcg), isolate, cb_arg,
283 nr_to_walk);
284 spin_unlock_irq(&nlru->lock);
285 return ret;
286}
287
288unsigned long list_lru_walk_node(struct list_lru *lru, int nid,
289 list_lru_walk_cb isolate, void *cb_arg,
290 unsigned long *nr_to_walk)
291{
292 long isolated = 0;
293 int memcg_idx;
294
295 isolated += list_lru_walk_one(lru, nid, NULL, isolate, cb_arg,
296 nr_to_walk);
297 if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
298 for_each_memcg_cache_index(memcg_idx) {
299 struct list_lru_node *nlru = &lru->node[nid];
300
301 spin_lock(&nlru->lock);
302 isolated += __list_lru_walk_one(nlru, memcg_idx,
303 isolate, cb_arg,
304 nr_to_walk);
305 spin_unlock(&nlru->lock);
306
307 if (*nr_to_walk <= 0)
308 break;
309 }
310 }
311 return isolated;
312}
313EXPORT_SYMBOL_GPL(list_lru_walk_node);
314
315static void init_one_lru(struct list_lru_one *l)
316{
317 INIT_LIST_HEAD(&l->list);
318 l->nr_items = 0;
319}
320
321#ifdef CONFIG_MEMCG_KMEM
322static void __memcg_destroy_list_lru_node(struct list_lru_memcg *memcg_lrus,
323 int begin, int end)
324{
325 int i;
326
327 for (i = begin; i < end; i++)
328 kfree(memcg_lrus->lru[i]);
329}
330
331static int __memcg_init_list_lru_node(struct list_lru_memcg *memcg_lrus,
332 int begin, int end)
333{
334 int i;
335
336 for (i = begin; i < end; i++) {
337 struct list_lru_one *l;
338
339 l = kmalloc(sizeof(struct list_lru_one), GFP_KERNEL);
340 if (!l)
341 goto fail;
342
343 init_one_lru(l);
344 memcg_lrus->lru[i] = l;
345 }
346 return 0;
347fail:
348 __memcg_destroy_list_lru_node(memcg_lrus, begin, i);
349 return -ENOMEM;
350}
351
352static int memcg_init_list_lru_node(struct list_lru_node *nlru)
353{
354 struct list_lru_memcg *memcg_lrus;
355 int size = memcg_nr_cache_ids;
356
357 memcg_lrus = kvmalloc(sizeof(*memcg_lrus) +
358 size * sizeof(void *), GFP_KERNEL);
359 if (!memcg_lrus)
360 return -ENOMEM;
361
362 if (__memcg_init_list_lru_node(memcg_lrus, 0, size)) {
363 kvfree(memcg_lrus);
364 return -ENOMEM;
365 }
366 RCU_INIT_POINTER(nlru->memcg_lrus, memcg_lrus);
367
368 return 0;
369}
370
371static void memcg_destroy_list_lru_node(struct list_lru_node *nlru)
372{
373 struct list_lru_memcg *memcg_lrus;
374 /*
375 * This is called when shrinker has already been unregistered,
376 * and nobody can use it. So, there is no need to use kvfree_rcu_local().
377 */
378 memcg_lrus = rcu_dereference_protected(nlru->memcg_lrus, true);
379 __memcg_destroy_list_lru_node(memcg_lrus, 0, memcg_nr_cache_ids);
380 kvfree(memcg_lrus);
381}
382
383static void kvfree_rcu_local(struct rcu_head *head)
384{
385 struct list_lru_memcg *mlru;
386
387 mlru = container_of(head, struct list_lru_memcg, rcu);
388 kvfree(mlru);
389}
390
391static int memcg_update_list_lru_node(struct list_lru_node *nlru,
392 int old_size, int new_size)
393{
394 struct list_lru_memcg *old, *new;
395
396 BUG_ON(old_size > new_size);
397
398 old = rcu_dereference_protected(nlru->memcg_lrus,
399 lockdep_is_held(&list_lrus_mutex));
400 new = kvmalloc(sizeof(*new) + new_size * sizeof(void *), GFP_KERNEL);
401 if (!new)
402 return -ENOMEM;
403
404 if (__memcg_init_list_lru_node(new, old_size, new_size)) {
405 kvfree(new);
406 return -ENOMEM;
407 }
408
409 memcpy(&new->lru, &old->lru, old_size * sizeof(void *));
410
411 /*
412 * The locking below allows readers that hold nlru->lock avoid taking
413 * rcu_read_lock (see list_lru_from_memcg_idx).
414 *
415 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
416 * we have to use IRQ-safe primitives here to avoid deadlock.
417 */
418 spin_lock_irq(&nlru->lock);
419 rcu_assign_pointer(nlru->memcg_lrus, new);
420 spin_unlock_irq(&nlru->lock);
421
422 call_rcu(&old->rcu, kvfree_rcu_local);
423 return 0;
424}
425
426static void memcg_cancel_update_list_lru_node(struct list_lru_node *nlru,
427 int old_size, int new_size)
428{
429 struct list_lru_memcg *memcg_lrus;
430
431 memcg_lrus = rcu_dereference_protected(nlru->memcg_lrus,
432 lockdep_is_held(&list_lrus_mutex));
433 /* do not bother shrinking the array back to the old size, because we
434 * cannot handle allocation failures here */
435 __memcg_destroy_list_lru_node(memcg_lrus, old_size, new_size);
436}
437
438static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
439{
440 int i;
441
442 lru->memcg_aware = memcg_aware;
443
444 if (!memcg_aware)
445 return 0;
446
447 for_each_node(i) {
448 if (memcg_init_list_lru_node(&lru->node[i]))
449 goto fail;
450 }
451 return 0;
452fail:
453 for (i = i - 1; i >= 0; i--) {
454 if (!lru->node[i].memcg_lrus)
455 continue;
456 memcg_destroy_list_lru_node(&lru->node[i]);
457 }
458 return -ENOMEM;
459}
460
461static void memcg_destroy_list_lru(struct list_lru *lru)
462{
463 int i;
464
465 if (!list_lru_memcg_aware(lru))
466 return;
467
468 for_each_node(i)
469 memcg_destroy_list_lru_node(&lru->node[i]);
470}
471
472static int memcg_update_list_lru(struct list_lru *lru,
473 int old_size, int new_size)
474{
475 int i;
476
477 if (!list_lru_memcg_aware(lru))
478 return 0;
479
480 for_each_node(i) {
481 if (memcg_update_list_lru_node(&lru->node[i],
482 old_size, new_size))
483 goto fail;
484 }
485 return 0;
486fail:
487 for (i = i - 1; i >= 0; i--) {
488 if (!lru->node[i].memcg_lrus)
489 continue;
490
491 memcg_cancel_update_list_lru_node(&lru->node[i],
492 old_size, new_size);
493 }
494 return -ENOMEM;
495}
496
497static void memcg_cancel_update_list_lru(struct list_lru *lru,
498 int old_size, int new_size)
499{
500 int i;
501
502 if (!list_lru_memcg_aware(lru))
503 return;
504
505 for_each_node(i)
506 memcg_cancel_update_list_lru_node(&lru->node[i],
507 old_size, new_size);
508}
509
510int memcg_update_all_list_lrus(int new_size)
511{
512 int ret = 0;
513 struct list_lru *lru;
514 int old_size = memcg_nr_cache_ids;
515
516 mutex_lock(&list_lrus_mutex);
517 list_for_each_entry(lru, &list_lrus, list) {
518 ret = memcg_update_list_lru(lru, old_size, new_size);
519 if (ret)
520 goto fail;
521 }
522out:
523 mutex_unlock(&list_lrus_mutex);
524 return ret;
525fail:
526 list_for_each_entry_continue_reverse(lru, &list_lrus, list)
527 memcg_cancel_update_list_lru(lru, old_size, new_size);
528 goto out;
529}
530
531static void memcg_drain_list_lru_node(struct list_lru *lru, int nid,
532 int src_idx, struct mem_cgroup *dst_memcg)
533{
534 struct list_lru_node *nlru = &lru->node[nid];
535 int dst_idx = dst_memcg->kmemcg_id;
536 struct list_lru_one *src, *dst;
537 bool set;
538
539 /*
540 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
541 * we have to use IRQ-safe primitives here to avoid deadlock.
542 */
543 spin_lock_irq(&nlru->lock);
544
545 src = list_lru_from_memcg_idx(nlru, src_idx);
546 dst = list_lru_from_memcg_idx(nlru, dst_idx);
547
548 list_splice_init(&src->list, &dst->list);
549 set = (!dst->nr_items && src->nr_items);
550 dst->nr_items += src->nr_items;
551 if (set)
552 memcg_set_shrinker_bit(dst_memcg, nid, lru_shrinker_id(lru));
553 src->nr_items = 0;
554
555 spin_unlock_irq(&nlru->lock);
556}
557
558static void memcg_drain_list_lru(struct list_lru *lru,
559 int src_idx, struct mem_cgroup *dst_memcg)
560{
561 int i;
562
563 if (!list_lru_memcg_aware(lru))
564 return;
565
566 for_each_node(i)
567 memcg_drain_list_lru_node(lru, i, src_idx, dst_memcg);
568}
569
570void memcg_drain_all_list_lrus(int src_idx, struct mem_cgroup *dst_memcg)
571{
572 struct list_lru *lru;
573
574 mutex_lock(&list_lrus_mutex);
575 list_for_each_entry(lru, &list_lrus, list)
576 memcg_drain_list_lru(lru, src_idx, dst_memcg);
577 mutex_unlock(&list_lrus_mutex);
578}
579#else
580static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
581{
582 return 0;
583}
584
585static void memcg_destroy_list_lru(struct list_lru *lru)
586{
587}
588#endif /* CONFIG_MEMCG_KMEM */
589
590int __list_lru_init(struct list_lru *lru, bool memcg_aware,
591 struct lock_class_key *key, struct shrinker *shrinker)
592{
593 int i;
594 int err = -ENOMEM;
595
596#ifdef CONFIG_MEMCG_KMEM
597 if (shrinker)
598 lru->shrinker_id = shrinker->id;
599 else
600 lru->shrinker_id = -1;
601#endif
602 memcg_get_cache_ids();
603
604 lru->node = kcalloc(nr_node_ids, sizeof(*lru->node), GFP_KERNEL);
605 if (!lru->node)
606 goto out;
607
608 for_each_node(i) {
609 spin_lock_init(&lru->node[i].lock);
610 if (key)
611 lockdep_set_class(&lru->node[i].lock, key);
612 init_one_lru(&lru->node[i].lru);
613 }
614
615 err = memcg_init_list_lru(lru, memcg_aware);
616 if (err) {
617 kfree(lru->node);
618 /* Do this so a list_lru_destroy() doesn't crash: */
619 lru->node = NULL;
620 goto out;
621 }
622
623 list_lru_register(lru);
624out:
625 memcg_put_cache_ids();
626 return err;
627}
628EXPORT_SYMBOL_GPL(__list_lru_init);
629
630void list_lru_destroy(struct list_lru *lru)
631{
632 /* Already destroyed or not yet initialized? */
633 if (!lru->node)
634 return;
635
636 memcg_get_cache_ids();
637
638 list_lru_unregister(lru);
639
640 memcg_destroy_list_lru(lru);
641 kfree(lru->node);
642 lru->node = NULL;
643
644#ifdef CONFIG_MEMCG_KMEM
645 lru->shrinker_id = -1;
646#endif
647 memcg_put_cache_ids();
648}
649EXPORT_SYMBOL_GPL(list_lru_destroy);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (c) 2013 Red Hat, Inc. and Parallels Inc. All rights reserved.
4 * Authors: David Chinner and Glauber Costa
5 *
6 * Generic LRU infrastructure
7 */
8#include <linux/kernel.h>
9#include <linux/module.h>
10#include <linux/mm.h>
11#include <linux/list_lru.h>
12#include <linux/slab.h>
13#include <linux/mutex.h>
14#include <linux/memcontrol.h>
15#include "slab.h"
16#include "internal.h"
17
18#ifdef CONFIG_MEMCG_KMEM
19static LIST_HEAD(memcg_list_lrus);
20static DEFINE_MUTEX(list_lrus_mutex);
21
22static inline bool list_lru_memcg_aware(struct list_lru *lru)
23{
24 return lru->memcg_aware;
25}
26
27static void list_lru_register(struct list_lru *lru)
28{
29 if (!list_lru_memcg_aware(lru))
30 return;
31
32 mutex_lock(&list_lrus_mutex);
33 list_add(&lru->list, &memcg_list_lrus);
34 mutex_unlock(&list_lrus_mutex);
35}
36
37static void list_lru_unregister(struct list_lru *lru)
38{
39 if (!list_lru_memcg_aware(lru))
40 return;
41
42 mutex_lock(&list_lrus_mutex);
43 list_del(&lru->list);
44 mutex_unlock(&list_lrus_mutex);
45}
46
47static int lru_shrinker_id(struct list_lru *lru)
48{
49 return lru->shrinker_id;
50}
51
52static inline struct list_lru_one *
53list_lru_from_memcg_idx(struct list_lru *lru, int nid, int idx)
54{
55 if (list_lru_memcg_aware(lru) && idx >= 0) {
56 struct list_lru_memcg *mlru = xa_load(&lru->xa, idx);
57
58 return mlru ? &mlru->node[nid] : NULL;
59 }
60 return &lru->node[nid].lru;
61}
62
63static inline struct list_lru_one *
64list_lru_from_kmem(struct list_lru *lru, int nid, void *ptr,
65 struct mem_cgroup **memcg_ptr)
66{
67 struct list_lru_node *nlru = &lru->node[nid];
68 struct list_lru_one *l = &nlru->lru;
69 struct mem_cgroup *memcg = NULL;
70
71 if (!list_lru_memcg_aware(lru))
72 goto out;
73
74 memcg = mem_cgroup_from_slab_obj(ptr);
75 if (!memcg)
76 goto out;
77
78 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
79out:
80 if (memcg_ptr)
81 *memcg_ptr = memcg;
82 return l;
83}
84#else
85static void list_lru_register(struct list_lru *lru)
86{
87}
88
89static void list_lru_unregister(struct list_lru *lru)
90{
91}
92
93static int lru_shrinker_id(struct list_lru *lru)
94{
95 return -1;
96}
97
98static inline bool list_lru_memcg_aware(struct list_lru *lru)
99{
100 return false;
101}
102
103static inline struct list_lru_one *
104list_lru_from_memcg_idx(struct list_lru *lru, int nid, int idx)
105{
106 return &lru->node[nid].lru;
107}
108
109static inline struct list_lru_one *
110list_lru_from_kmem(struct list_lru *lru, int nid, void *ptr,
111 struct mem_cgroup **memcg_ptr)
112{
113 if (memcg_ptr)
114 *memcg_ptr = NULL;
115 return &lru->node[nid].lru;
116}
117#endif /* CONFIG_MEMCG_KMEM */
118
119bool list_lru_add(struct list_lru *lru, struct list_head *item)
120{
121 int nid = page_to_nid(virt_to_page(item));
122 struct list_lru_node *nlru = &lru->node[nid];
123 struct mem_cgroup *memcg;
124 struct list_lru_one *l;
125
126 spin_lock(&nlru->lock);
127 if (list_empty(item)) {
128 l = list_lru_from_kmem(lru, nid, item, &memcg);
129 list_add_tail(item, &l->list);
130 /* Set shrinker bit if the first element was added */
131 if (!l->nr_items++)
132 set_shrinker_bit(memcg, nid,
133 lru_shrinker_id(lru));
134 nlru->nr_items++;
135 spin_unlock(&nlru->lock);
136 return true;
137 }
138 spin_unlock(&nlru->lock);
139 return false;
140}
141EXPORT_SYMBOL_GPL(list_lru_add);
142
143bool list_lru_del(struct list_lru *lru, struct list_head *item)
144{
145 int nid = page_to_nid(virt_to_page(item));
146 struct list_lru_node *nlru = &lru->node[nid];
147 struct list_lru_one *l;
148
149 spin_lock(&nlru->lock);
150 if (!list_empty(item)) {
151 l = list_lru_from_kmem(lru, nid, item, NULL);
152 list_del_init(item);
153 l->nr_items--;
154 nlru->nr_items--;
155 spin_unlock(&nlru->lock);
156 return true;
157 }
158 spin_unlock(&nlru->lock);
159 return false;
160}
161EXPORT_SYMBOL_GPL(list_lru_del);
162
163void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
164{
165 list_del_init(item);
166 list->nr_items--;
167}
168EXPORT_SYMBOL_GPL(list_lru_isolate);
169
170void list_lru_isolate_move(struct list_lru_one *list, struct list_head *item,
171 struct list_head *head)
172{
173 list_move(item, head);
174 list->nr_items--;
175}
176EXPORT_SYMBOL_GPL(list_lru_isolate_move);
177
178unsigned long list_lru_count_one(struct list_lru *lru,
179 int nid, struct mem_cgroup *memcg)
180{
181 struct list_lru_one *l;
182 long count;
183
184 rcu_read_lock();
185 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
186 count = l ? READ_ONCE(l->nr_items) : 0;
187 rcu_read_unlock();
188
189 if (unlikely(count < 0))
190 count = 0;
191
192 return count;
193}
194EXPORT_SYMBOL_GPL(list_lru_count_one);
195
196unsigned long list_lru_count_node(struct list_lru *lru, int nid)
197{
198 struct list_lru_node *nlru;
199
200 nlru = &lru->node[nid];
201 return nlru->nr_items;
202}
203EXPORT_SYMBOL_GPL(list_lru_count_node);
204
205static unsigned long
206__list_lru_walk_one(struct list_lru *lru, int nid, int memcg_idx,
207 list_lru_walk_cb isolate, void *cb_arg,
208 unsigned long *nr_to_walk)
209{
210 struct list_lru_node *nlru = &lru->node[nid];
211 struct list_lru_one *l;
212 struct list_head *item, *n;
213 unsigned long isolated = 0;
214
215restart:
216 l = list_lru_from_memcg_idx(lru, nid, memcg_idx);
217 if (!l)
218 goto out;
219
220 list_for_each_safe(item, n, &l->list) {
221 enum lru_status ret;
222
223 /*
224 * decrement nr_to_walk first so that we don't livelock if we
225 * get stuck on large numbers of LRU_RETRY items
226 */
227 if (!*nr_to_walk)
228 break;
229 --*nr_to_walk;
230
231 ret = isolate(item, l, &nlru->lock, cb_arg);
232 switch (ret) {
233 case LRU_REMOVED_RETRY:
234 assert_spin_locked(&nlru->lock);
235 fallthrough;
236 case LRU_REMOVED:
237 isolated++;
238 nlru->nr_items--;
239 /*
240 * If the lru lock has been dropped, our list
241 * traversal is now invalid and so we have to
242 * restart from scratch.
243 */
244 if (ret == LRU_REMOVED_RETRY)
245 goto restart;
246 break;
247 case LRU_ROTATE:
248 list_move_tail(item, &l->list);
249 break;
250 case LRU_SKIP:
251 break;
252 case LRU_RETRY:
253 /*
254 * The lru lock has been dropped, our list traversal is
255 * now invalid and so we have to restart from scratch.
256 */
257 assert_spin_locked(&nlru->lock);
258 goto restart;
259 default:
260 BUG();
261 }
262 }
263out:
264 return isolated;
265}
266
267unsigned long
268list_lru_walk_one(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
269 list_lru_walk_cb isolate, void *cb_arg,
270 unsigned long *nr_to_walk)
271{
272 struct list_lru_node *nlru = &lru->node[nid];
273 unsigned long ret;
274
275 spin_lock(&nlru->lock);
276 ret = __list_lru_walk_one(lru, nid, memcg_kmem_id(memcg), isolate,
277 cb_arg, nr_to_walk);
278 spin_unlock(&nlru->lock);
279 return ret;
280}
281EXPORT_SYMBOL_GPL(list_lru_walk_one);
282
283unsigned long
284list_lru_walk_one_irq(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
285 list_lru_walk_cb isolate, void *cb_arg,
286 unsigned long *nr_to_walk)
287{
288 struct list_lru_node *nlru = &lru->node[nid];
289 unsigned long ret;
290
291 spin_lock_irq(&nlru->lock);
292 ret = __list_lru_walk_one(lru, nid, memcg_kmem_id(memcg), isolate,
293 cb_arg, nr_to_walk);
294 spin_unlock_irq(&nlru->lock);
295 return ret;
296}
297
298unsigned long list_lru_walk_node(struct list_lru *lru, int nid,
299 list_lru_walk_cb isolate, void *cb_arg,
300 unsigned long *nr_to_walk)
301{
302 long isolated = 0;
303
304 isolated += list_lru_walk_one(lru, nid, NULL, isolate, cb_arg,
305 nr_to_walk);
306
307#ifdef CONFIG_MEMCG_KMEM
308 if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
309 struct list_lru_memcg *mlru;
310 unsigned long index;
311
312 xa_for_each(&lru->xa, index, mlru) {
313 struct list_lru_node *nlru = &lru->node[nid];
314
315 spin_lock(&nlru->lock);
316 isolated += __list_lru_walk_one(lru, nid, index,
317 isolate, cb_arg,
318 nr_to_walk);
319 spin_unlock(&nlru->lock);
320
321 if (*nr_to_walk <= 0)
322 break;
323 }
324 }
325#endif
326
327 return isolated;
328}
329EXPORT_SYMBOL_GPL(list_lru_walk_node);
330
331static void init_one_lru(struct list_lru_one *l)
332{
333 INIT_LIST_HEAD(&l->list);
334 l->nr_items = 0;
335}
336
337#ifdef CONFIG_MEMCG_KMEM
338static struct list_lru_memcg *memcg_init_list_lru_one(gfp_t gfp)
339{
340 int nid;
341 struct list_lru_memcg *mlru;
342
343 mlru = kmalloc(struct_size(mlru, node, nr_node_ids), gfp);
344 if (!mlru)
345 return NULL;
346
347 for_each_node(nid)
348 init_one_lru(&mlru->node[nid]);
349
350 return mlru;
351}
352
353static void memcg_list_lru_free(struct list_lru *lru, int src_idx)
354{
355 struct list_lru_memcg *mlru = xa_erase_irq(&lru->xa, src_idx);
356
357 /*
358 * The __list_lru_walk_one() can walk the list of this node.
359 * We need kvfree_rcu() here. And the walking of the list
360 * is under lru->node[nid]->lock, which can serve as a RCU
361 * read-side critical section.
362 */
363 if (mlru)
364 kvfree_rcu(mlru, rcu);
365}
366
367static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
368{
369 if (memcg_aware)
370 xa_init_flags(&lru->xa, XA_FLAGS_LOCK_IRQ);
371 lru->memcg_aware = memcg_aware;
372}
373
374static void memcg_destroy_list_lru(struct list_lru *lru)
375{
376 XA_STATE(xas, &lru->xa, 0);
377 struct list_lru_memcg *mlru;
378
379 if (!list_lru_memcg_aware(lru))
380 return;
381
382 xas_lock_irq(&xas);
383 xas_for_each(&xas, mlru, ULONG_MAX) {
384 kfree(mlru);
385 xas_store(&xas, NULL);
386 }
387 xas_unlock_irq(&xas);
388}
389
390static void memcg_reparent_list_lru_node(struct list_lru *lru, int nid,
391 int src_idx, struct mem_cgroup *dst_memcg)
392{
393 struct list_lru_node *nlru = &lru->node[nid];
394 int dst_idx = dst_memcg->kmemcg_id;
395 struct list_lru_one *src, *dst;
396
397 /*
398 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
399 * we have to use IRQ-safe primitives here to avoid deadlock.
400 */
401 spin_lock_irq(&nlru->lock);
402
403 src = list_lru_from_memcg_idx(lru, nid, src_idx);
404 if (!src)
405 goto out;
406 dst = list_lru_from_memcg_idx(lru, nid, dst_idx);
407
408 list_splice_init(&src->list, &dst->list);
409
410 if (src->nr_items) {
411 dst->nr_items += src->nr_items;
412 set_shrinker_bit(dst_memcg, nid, lru_shrinker_id(lru));
413 src->nr_items = 0;
414 }
415out:
416 spin_unlock_irq(&nlru->lock);
417}
418
419static void memcg_reparent_list_lru(struct list_lru *lru,
420 int src_idx, struct mem_cgroup *dst_memcg)
421{
422 int i;
423
424 for_each_node(i)
425 memcg_reparent_list_lru_node(lru, i, src_idx, dst_memcg);
426
427 memcg_list_lru_free(lru, src_idx);
428}
429
430void memcg_reparent_list_lrus(struct mem_cgroup *memcg, struct mem_cgroup *parent)
431{
432 struct cgroup_subsys_state *css;
433 struct list_lru *lru;
434 int src_idx = memcg->kmemcg_id;
435
436 /*
437 * Change kmemcg_id of this cgroup and all its descendants to the
438 * parent's id, and then move all entries from this cgroup's list_lrus
439 * to ones of the parent.
440 *
441 * After we have finished, all list_lrus corresponding to this cgroup
442 * are guaranteed to remain empty. So we can safely free this cgroup's
443 * list lrus in memcg_list_lru_free().
444 *
445 * Changing ->kmemcg_id to the parent can prevent memcg_list_lru_alloc()
446 * from allocating list lrus for this cgroup after memcg_list_lru_free()
447 * call.
448 */
449 rcu_read_lock();
450 css_for_each_descendant_pre(css, &memcg->css) {
451 struct mem_cgroup *child;
452
453 child = mem_cgroup_from_css(css);
454 WRITE_ONCE(child->kmemcg_id, parent->kmemcg_id);
455 }
456 rcu_read_unlock();
457
458 mutex_lock(&list_lrus_mutex);
459 list_for_each_entry(lru, &memcg_list_lrus, list)
460 memcg_reparent_list_lru(lru, src_idx, parent);
461 mutex_unlock(&list_lrus_mutex);
462}
463
464static inline bool memcg_list_lru_allocated(struct mem_cgroup *memcg,
465 struct list_lru *lru)
466{
467 int idx = memcg->kmemcg_id;
468
469 return idx < 0 || xa_load(&lru->xa, idx);
470}
471
472int memcg_list_lru_alloc(struct mem_cgroup *memcg, struct list_lru *lru,
473 gfp_t gfp)
474{
475 int i;
476 unsigned long flags;
477 struct list_lru_memcg_table {
478 struct list_lru_memcg *mlru;
479 struct mem_cgroup *memcg;
480 } *table;
481 XA_STATE(xas, &lru->xa, 0);
482
483 if (!list_lru_memcg_aware(lru) || memcg_list_lru_allocated(memcg, lru))
484 return 0;
485
486 gfp &= GFP_RECLAIM_MASK;
487 table = kmalloc_array(memcg->css.cgroup->level, sizeof(*table), gfp);
488 if (!table)
489 return -ENOMEM;
490
491 /*
492 * Because the list_lru can be reparented to the parent cgroup's
493 * list_lru, we should make sure that this cgroup and all its
494 * ancestors have allocated list_lru_memcg.
495 */
496 for (i = 0; memcg; memcg = parent_mem_cgroup(memcg), i++) {
497 if (memcg_list_lru_allocated(memcg, lru))
498 break;
499
500 table[i].memcg = memcg;
501 table[i].mlru = memcg_init_list_lru_one(gfp);
502 if (!table[i].mlru) {
503 while (i--)
504 kfree(table[i].mlru);
505 kfree(table);
506 return -ENOMEM;
507 }
508 }
509
510 xas_lock_irqsave(&xas, flags);
511 while (i--) {
512 int index = READ_ONCE(table[i].memcg->kmemcg_id);
513 struct list_lru_memcg *mlru = table[i].mlru;
514
515 xas_set(&xas, index);
516retry:
517 if (unlikely(index < 0 || xas_error(&xas) || xas_load(&xas))) {
518 kfree(mlru);
519 } else {
520 xas_store(&xas, mlru);
521 if (xas_error(&xas) == -ENOMEM) {
522 xas_unlock_irqrestore(&xas, flags);
523 if (xas_nomem(&xas, gfp))
524 xas_set_err(&xas, 0);
525 xas_lock_irqsave(&xas, flags);
526 /*
527 * The xas lock has been released, this memcg
528 * can be reparented before us. So reload
529 * memcg id. More details see the comments
530 * in memcg_reparent_list_lrus().
531 */
532 index = READ_ONCE(table[i].memcg->kmemcg_id);
533 if (index < 0)
534 xas_set_err(&xas, 0);
535 else if (!xas_error(&xas) && index != xas.xa_index)
536 xas_set(&xas, index);
537 goto retry;
538 }
539 }
540 }
541 /* xas_nomem() is used to free memory instead of memory allocation. */
542 if (xas.xa_alloc)
543 xas_nomem(&xas, gfp);
544 xas_unlock_irqrestore(&xas, flags);
545 kfree(table);
546
547 return xas_error(&xas);
548}
549#else
550static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
551{
552}
553
554static void memcg_destroy_list_lru(struct list_lru *lru)
555{
556}
557#endif /* CONFIG_MEMCG_KMEM */
558
559int __list_lru_init(struct list_lru *lru, bool memcg_aware,
560 struct lock_class_key *key, struct shrinker *shrinker)
561{
562 int i;
563
564#ifdef CONFIG_MEMCG_KMEM
565 if (shrinker)
566 lru->shrinker_id = shrinker->id;
567 else
568 lru->shrinker_id = -1;
569#endif
570
571 lru->node = kcalloc(nr_node_ids, sizeof(*lru->node), GFP_KERNEL);
572 if (!lru->node)
573 return -ENOMEM;
574
575 for_each_node(i) {
576 spin_lock_init(&lru->node[i].lock);
577 if (key)
578 lockdep_set_class(&lru->node[i].lock, key);
579 init_one_lru(&lru->node[i].lru);
580 }
581
582 memcg_init_list_lru(lru, memcg_aware);
583 list_lru_register(lru);
584
585 return 0;
586}
587EXPORT_SYMBOL_GPL(__list_lru_init);
588
589void list_lru_destroy(struct list_lru *lru)
590{
591 /* Already destroyed or not yet initialized? */
592 if (!lru->node)
593 return;
594
595 list_lru_unregister(lru);
596
597 memcg_destroy_list_lru(lru);
598 kfree(lru->node);
599 lru->node = NULL;
600
601#ifdef CONFIG_MEMCG_KMEM
602 lru->shrinker_id = -1;
603#endif
604}
605EXPORT_SYMBOL_GPL(list_lru_destroy);