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 __always_inline struct mem_cgroup *mem_cgroup_from_kmem(void *ptr)
61{
62 struct page *page;
63
64 if (!memcg_kmem_enabled())
65 return NULL;
66 page = virt_to_head_page(ptr);
67 return memcg_from_slab_page(page);
68}
69
70static inline struct list_lru_one *
71list_lru_from_kmem(struct list_lru_node *nlru, void *ptr,
72 struct mem_cgroup **memcg_ptr)
73{
74 struct list_lru_one *l = &nlru->lru;
75 struct mem_cgroup *memcg = NULL;
76
77 if (!nlru->memcg_lrus)
78 goto out;
79
80 memcg = mem_cgroup_from_kmem(ptr);
81 if (!memcg)
82 goto out;
83
84 l = list_lru_from_memcg_idx(nlru, memcg_cache_id(memcg));
85out:
86 if (memcg_ptr)
87 *memcg_ptr = memcg;
88 return l;
89}
90#else
91static void list_lru_register(struct list_lru *lru)
92{
93}
94
95static void list_lru_unregister(struct list_lru *lru)
96{
97}
98
99static int lru_shrinker_id(struct list_lru *lru)
100{
101 return -1;
102}
103
104static inline bool list_lru_memcg_aware(struct list_lru *lru)
105{
106 return false;
107}
108
109static inline struct list_lru_one *
110list_lru_from_memcg_idx(struct list_lru_node *nlru, int idx)
111{
112 return &nlru->lru;
113}
114
115static inline struct list_lru_one *
116list_lru_from_kmem(struct list_lru_node *nlru, void *ptr,
117 struct mem_cgroup **memcg_ptr)
118{
119 if (memcg_ptr)
120 *memcg_ptr = NULL;
121 return &nlru->lru;
122}
123#endif /* CONFIG_MEMCG_KMEM */
124
125bool list_lru_add(struct list_lru *lru, struct list_head *item)
126{
127 int nid = page_to_nid(virt_to_page(item));
128 struct list_lru_node *nlru = &lru->node[nid];
129 struct mem_cgroup *memcg;
130 struct list_lru_one *l;
131
132 spin_lock(&nlru->lock);
133 if (list_empty(item)) {
134 l = list_lru_from_kmem(nlru, item, &memcg);
135 list_add_tail(item, &l->list);
136 /* Set shrinker bit if the first element was added */
137 if (!l->nr_items++)
138 memcg_set_shrinker_bit(memcg, nid,
139 lru_shrinker_id(lru));
140 nlru->nr_items++;
141 spin_unlock(&nlru->lock);
142 return true;
143 }
144 spin_unlock(&nlru->lock);
145 return false;
146}
147EXPORT_SYMBOL_GPL(list_lru_add);
148
149bool list_lru_del(struct list_lru *lru, struct list_head *item)
150{
151 int nid = page_to_nid(virt_to_page(item));
152 struct list_lru_node *nlru = &lru->node[nid];
153 struct list_lru_one *l;
154
155 spin_lock(&nlru->lock);
156 if (!list_empty(item)) {
157 l = list_lru_from_kmem(nlru, item, NULL);
158 list_del_init(item);
159 l->nr_items--;
160 nlru->nr_items--;
161 spin_unlock(&nlru->lock);
162 return true;
163 }
164 spin_unlock(&nlru->lock);
165 return false;
166}
167EXPORT_SYMBOL_GPL(list_lru_del);
168
169void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
170{
171 list_del_init(item);
172 list->nr_items--;
173}
174EXPORT_SYMBOL_GPL(list_lru_isolate);
175
176void list_lru_isolate_move(struct list_lru_one *list, struct list_head *item,
177 struct list_head *head)
178{
179 list_move(item, head);
180 list->nr_items--;
181}
182EXPORT_SYMBOL_GPL(list_lru_isolate_move);
183
184unsigned long list_lru_count_one(struct list_lru *lru,
185 int nid, struct mem_cgroup *memcg)
186{
187 struct list_lru_node *nlru = &lru->node[nid];
188 struct list_lru_one *l;
189 unsigned long count;
190
191 rcu_read_lock();
192 l = list_lru_from_memcg_idx(nlru, memcg_cache_id(memcg));
193 count = l->nr_items;
194 rcu_read_unlock();
195
196 return count;
197}
198EXPORT_SYMBOL_GPL(list_lru_count_one);
199
200unsigned long list_lru_count_node(struct list_lru *lru, int nid)
201{
202 struct list_lru_node *nlru;
203
204 nlru = &lru->node[nid];
205 return nlru->nr_items;
206}
207EXPORT_SYMBOL_GPL(list_lru_count_node);
208
209static unsigned long
210__list_lru_walk_one(struct list_lru_node *nlru, int memcg_idx,
211 list_lru_walk_cb isolate, void *cb_arg,
212 unsigned long *nr_to_walk)
213{
214
215 struct list_lru_one *l;
216 struct list_head *item, *n;
217 unsigned long isolated = 0;
218
219 l = list_lru_from_memcg_idx(nlru, memcg_idx);
220restart:
221 list_for_each_safe(item, n, &l->list) {
222 enum lru_status ret;
223
224 /*
225 * decrement nr_to_walk first so that we don't livelock if we
226 * get stuck on large numbesr of LRU_RETRY items
227 */
228 if (!*nr_to_walk)
229 break;
230 --*nr_to_walk;
231
232 ret = isolate(item, l, &nlru->lock, cb_arg);
233 switch (ret) {
234 case LRU_REMOVED_RETRY:
235 assert_spin_locked(&nlru->lock);
236 /* fall through */
237 case LRU_REMOVED:
238 isolated++;
239 nlru->nr_items--;
240 /*
241 * If the lru lock has been dropped, our list
242 * traversal is now invalid and so we have to
243 * restart from scratch.
244 */
245 if (ret == LRU_REMOVED_RETRY)
246 goto restart;
247 break;
248 case LRU_ROTATE:
249 list_move_tail(item, &l->list);
250 break;
251 case LRU_SKIP:
252 break;
253 case LRU_RETRY:
254 /*
255 * The lru lock has been dropped, our list traversal is
256 * now invalid and so we have to restart from scratch.
257 */
258 assert_spin_locked(&nlru->lock);
259 goto restart;
260 default:
261 BUG();
262 }
263 }
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(nlru, memcg_cache_id(memcg), isolate, cb_arg,
277 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(nlru, memcg_cache_id(memcg), isolate, cb_arg,
293 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 int memcg_idx;
304
305 isolated += list_lru_walk_one(lru, nid, NULL, isolate, cb_arg,
306 nr_to_walk);
307 if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
308 for_each_memcg_cache_index(memcg_idx) {
309 struct list_lru_node *nlru = &lru->node[nid];
310
311 spin_lock(&nlru->lock);
312 isolated += __list_lru_walk_one(nlru, memcg_idx,
313 isolate, cb_arg,
314 nr_to_walk);
315 spin_unlock(&nlru->lock);
316
317 if (*nr_to_walk <= 0)
318 break;
319 }
320 }
321 return isolated;
322}
323EXPORT_SYMBOL_GPL(list_lru_walk_node);
324
325static void init_one_lru(struct list_lru_one *l)
326{
327 INIT_LIST_HEAD(&l->list);
328 l->nr_items = 0;
329}
330
331#ifdef CONFIG_MEMCG_KMEM
332static void __memcg_destroy_list_lru_node(struct list_lru_memcg *memcg_lrus,
333 int begin, int end)
334{
335 int i;
336
337 for (i = begin; i < end; i++)
338 kfree(memcg_lrus->lru[i]);
339}
340
341static int __memcg_init_list_lru_node(struct list_lru_memcg *memcg_lrus,
342 int begin, int end)
343{
344 int i;
345
346 for (i = begin; i < end; i++) {
347 struct list_lru_one *l;
348
349 l = kmalloc(sizeof(struct list_lru_one), GFP_KERNEL);
350 if (!l)
351 goto fail;
352
353 init_one_lru(l);
354 memcg_lrus->lru[i] = l;
355 }
356 return 0;
357fail:
358 __memcg_destroy_list_lru_node(memcg_lrus, begin, i);
359 return -ENOMEM;
360}
361
362static int memcg_init_list_lru_node(struct list_lru_node *nlru)
363{
364 struct list_lru_memcg *memcg_lrus;
365 int size = memcg_nr_cache_ids;
366
367 memcg_lrus = kvmalloc(sizeof(*memcg_lrus) +
368 size * sizeof(void *), GFP_KERNEL);
369 if (!memcg_lrus)
370 return -ENOMEM;
371
372 if (__memcg_init_list_lru_node(memcg_lrus, 0, size)) {
373 kvfree(memcg_lrus);
374 return -ENOMEM;
375 }
376 RCU_INIT_POINTER(nlru->memcg_lrus, memcg_lrus);
377
378 return 0;
379}
380
381static void memcg_destroy_list_lru_node(struct list_lru_node *nlru)
382{
383 struct list_lru_memcg *memcg_lrus;
384 /*
385 * This is called when shrinker has already been unregistered,
386 * and nobody can use it. So, there is no need to use kvfree_rcu().
387 */
388 memcg_lrus = rcu_dereference_protected(nlru->memcg_lrus, true);
389 __memcg_destroy_list_lru_node(memcg_lrus, 0, memcg_nr_cache_ids);
390 kvfree(memcg_lrus);
391}
392
393static void kvfree_rcu(struct rcu_head *head)
394{
395 struct list_lru_memcg *mlru;
396
397 mlru = container_of(head, struct list_lru_memcg, rcu);
398 kvfree(mlru);
399}
400
401static int memcg_update_list_lru_node(struct list_lru_node *nlru,
402 int old_size, int new_size)
403{
404 struct list_lru_memcg *old, *new;
405
406 BUG_ON(old_size > new_size);
407
408 old = rcu_dereference_protected(nlru->memcg_lrus,
409 lockdep_is_held(&list_lrus_mutex));
410 new = kvmalloc(sizeof(*new) + new_size * sizeof(void *), GFP_KERNEL);
411 if (!new)
412 return -ENOMEM;
413
414 if (__memcg_init_list_lru_node(new, old_size, new_size)) {
415 kvfree(new);
416 return -ENOMEM;
417 }
418
419 memcpy(&new->lru, &old->lru, old_size * sizeof(void *));
420
421 /*
422 * The locking below allows readers that hold nlru->lock avoid taking
423 * rcu_read_lock (see list_lru_from_memcg_idx).
424 *
425 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
426 * we have to use IRQ-safe primitives here to avoid deadlock.
427 */
428 spin_lock_irq(&nlru->lock);
429 rcu_assign_pointer(nlru->memcg_lrus, new);
430 spin_unlock_irq(&nlru->lock);
431
432 call_rcu(&old->rcu, kvfree_rcu);
433 return 0;
434}
435
436static void memcg_cancel_update_list_lru_node(struct list_lru_node *nlru,
437 int old_size, int new_size)
438{
439 struct list_lru_memcg *memcg_lrus;
440
441 memcg_lrus = rcu_dereference_protected(nlru->memcg_lrus,
442 lockdep_is_held(&list_lrus_mutex));
443 /* do not bother shrinking the array back to the old size, because we
444 * cannot handle allocation failures here */
445 __memcg_destroy_list_lru_node(memcg_lrus, old_size, new_size);
446}
447
448static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
449{
450 int i;
451
452 lru->memcg_aware = memcg_aware;
453
454 if (!memcg_aware)
455 return 0;
456
457 for_each_node(i) {
458 if (memcg_init_list_lru_node(&lru->node[i]))
459 goto fail;
460 }
461 return 0;
462fail:
463 for (i = i - 1; i >= 0; i--) {
464 if (!lru->node[i].memcg_lrus)
465 continue;
466 memcg_destroy_list_lru_node(&lru->node[i]);
467 }
468 return -ENOMEM;
469}
470
471static void memcg_destroy_list_lru(struct list_lru *lru)
472{
473 int i;
474
475 if (!list_lru_memcg_aware(lru))
476 return;
477
478 for_each_node(i)
479 memcg_destroy_list_lru_node(&lru->node[i]);
480}
481
482static int memcg_update_list_lru(struct list_lru *lru,
483 int old_size, int new_size)
484{
485 int i;
486
487 if (!list_lru_memcg_aware(lru))
488 return 0;
489
490 for_each_node(i) {
491 if (memcg_update_list_lru_node(&lru->node[i],
492 old_size, new_size))
493 goto fail;
494 }
495 return 0;
496fail:
497 for (i = i - 1; i >= 0; i--) {
498 if (!lru->node[i].memcg_lrus)
499 continue;
500
501 memcg_cancel_update_list_lru_node(&lru->node[i],
502 old_size, new_size);
503 }
504 return -ENOMEM;
505}
506
507static void memcg_cancel_update_list_lru(struct list_lru *lru,
508 int old_size, int new_size)
509{
510 int i;
511
512 if (!list_lru_memcg_aware(lru))
513 return;
514
515 for_each_node(i)
516 memcg_cancel_update_list_lru_node(&lru->node[i],
517 old_size, new_size);
518}
519
520int memcg_update_all_list_lrus(int new_size)
521{
522 int ret = 0;
523 struct list_lru *lru;
524 int old_size = memcg_nr_cache_ids;
525
526 mutex_lock(&list_lrus_mutex);
527 list_for_each_entry(lru, &list_lrus, list) {
528 ret = memcg_update_list_lru(lru, old_size, new_size);
529 if (ret)
530 goto fail;
531 }
532out:
533 mutex_unlock(&list_lrus_mutex);
534 return ret;
535fail:
536 list_for_each_entry_continue_reverse(lru, &list_lrus, list)
537 memcg_cancel_update_list_lru(lru, old_size, new_size);
538 goto out;
539}
540
541static void memcg_drain_list_lru_node(struct list_lru *lru, int nid,
542 int src_idx, struct mem_cgroup *dst_memcg)
543{
544 struct list_lru_node *nlru = &lru->node[nid];
545 int dst_idx = dst_memcg->kmemcg_id;
546 struct list_lru_one *src, *dst;
547 bool set;
548
549 /*
550 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
551 * we have to use IRQ-safe primitives here to avoid deadlock.
552 */
553 spin_lock_irq(&nlru->lock);
554
555 src = list_lru_from_memcg_idx(nlru, src_idx);
556 dst = list_lru_from_memcg_idx(nlru, dst_idx);
557
558 list_splice_init(&src->list, &dst->list);
559 set = (!dst->nr_items && src->nr_items);
560 dst->nr_items += src->nr_items;
561 if (set)
562 memcg_set_shrinker_bit(dst_memcg, nid, lru_shrinker_id(lru));
563 src->nr_items = 0;
564
565 spin_unlock_irq(&nlru->lock);
566}
567
568static void memcg_drain_list_lru(struct list_lru *lru,
569 int src_idx, struct mem_cgroup *dst_memcg)
570{
571 int i;
572
573 if (!list_lru_memcg_aware(lru))
574 return;
575
576 for_each_node(i)
577 memcg_drain_list_lru_node(lru, i, src_idx, dst_memcg);
578}
579
580void memcg_drain_all_list_lrus(int src_idx, struct mem_cgroup *dst_memcg)
581{
582 struct list_lru *lru;
583
584 mutex_lock(&list_lrus_mutex);
585 list_for_each_entry(lru, &list_lrus, list)
586 memcg_drain_list_lru(lru, src_idx, dst_memcg);
587 mutex_unlock(&list_lrus_mutex);
588}
589#else
590static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
591{
592 return 0;
593}
594
595static void memcg_destroy_list_lru(struct list_lru *lru)
596{
597}
598#endif /* CONFIG_MEMCG_KMEM */
599
600int __list_lru_init(struct list_lru *lru, bool memcg_aware,
601 struct lock_class_key *key, struct shrinker *shrinker)
602{
603 int i;
604 int err = -ENOMEM;
605
606#ifdef CONFIG_MEMCG_KMEM
607 if (shrinker)
608 lru->shrinker_id = shrinker->id;
609 else
610 lru->shrinker_id = -1;
611#endif
612 memcg_get_cache_ids();
613
614 lru->node = kcalloc(nr_node_ids, sizeof(*lru->node), GFP_KERNEL);
615 if (!lru->node)
616 goto out;
617
618 for_each_node(i) {
619 spin_lock_init(&lru->node[i].lock);
620 if (key)
621 lockdep_set_class(&lru->node[i].lock, key);
622 init_one_lru(&lru->node[i].lru);
623 }
624
625 err = memcg_init_list_lru(lru, memcg_aware);
626 if (err) {
627 kfree(lru->node);
628 /* Do this so a list_lru_destroy() doesn't crash: */
629 lru->node = NULL;
630 goto out;
631 }
632
633 list_lru_register(lru);
634out:
635 memcg_put_cache_ids();
636 return err;
637}
638EXPORT_SYMBOL_GPL(__list_lru_init);
639
640void list_lru_destroy(struct list_lru *lru)
641{
642 /* Already destroyed or not yet initialized? */
643 if (!lru->node)
644 return;
645
646 memcg_get_cache_ids();
647
648 list_lru_unregister(lru);
649
650 memcg_destroy_list_lru(lru);
651 kfree(lru->node);
652 lru->node = NULL;
653
654#ifdef CONFIG_MEMCG_KMEM
655 lru->shrinker_id = -1;
656#endif
657 memcg_put_cache_ids();
658}
659EXPORT_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
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 *
64lock_list_lru_of_memcg(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
65 bool irq, bool skip_empty)
66{
67 struct list_lru_one *l;
68 long nr_items;
69
70 rcu_read_lock();
71again:
72 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
73 if (likely(l)) {
74 if (irq)
75 spin_lock_irq(&l->lock);
76 else
77 spin_lock(&l->lock);
78 nr_items = READ_ONCE(l->nr_items);
79 if (likely(nr_items != LONG_MIN)) {
80 rcu_read_unlock();
81 return l;
82 }
83 if (irq)
84 spin_unlock_irq(&l->lock);
85 else
86 spin_unlock(&l->lock);
87 }
88 /*
89 * Caller may simply bail out if raced with reparenting or
90 * may iterate through the list_lru and expect empty slots.
91 */
92 if (skip_empty) {
93 rcu_read_unlock();
94 return NULL;
95 }
96 VM_WARN_ON(!css_is_dying(&memcg->css));
97 memcg = parent_mem_cgroup(memcg);
98 goto again;
99}
100
101static inline void unlock_list_lru(struct list_lru_one *l, bool irq_off)
102{
103 if (irq_off)
104 spin_unlock_irq(&l->lock);
105 else
106 spin_unlock(&l->lock);
107}
108#else
109static void list_lru_register(struct list_lru *lru)
110{
111}
112
113static void list_lru_unregister(struct list_lru *lru)
114{
115}
116
117static int lru_shrinker_id(struct list_lru *lru)
118{
119 return -1;
120}
121
122static inline bool list_lru_memcg_aware(struct list_lru *lru)
123{
124 return false;
125}
126
127static inline struct list_lru_one *
128list_lru_from_memcg_idx(struct list_lru *lru, int nid, int idx)
129{
130 return &lru->node[nid].lru;
131}
132
133static inline struct list_lru_one *
134lock_list_lru_of_memcg(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
135 bool irq, bool skip_empty)
136{
137 struct list_lru_one *l = &lru->node[nid].lru;
138
139 if (irq)
140 spin_lock_irq(&l->lock);
141 else
142 spin_lock(&l->lock);
143
144 return l;
145}
146
147static inline void unlock_list_lru(struct list_lru_one *l, bool irq_off)
148{
149 if (irq_off)
150 spin_unlock_irq(&l->lock);
151 else
152 spin_unlock(&l->lock);
153}
154#endif /* CONFIG_MEMCG */
155
156/* The caller must ensure the memcg lifetime. */
157bool list_lru_add(struct list_lru *lru, struct list_head *item, int nid,
158 struct mem_cgroup *memcg)
159{
160 struct list_lru_node *nlru = &lru->node[nid];
161 struct list_lru_one *l;
162
163 l = lock_list_lru_of_memcg(lru, nid, memcg, false, false);
164 if (!l)
165 return false;
166 if (list_empty(item)) {
167 list_add_tail(item, &l->list);
168 /* Set shrinker bit if the first element was added */
169 if (!l->nr_items++)
170 set_shrinker_bit(memcg, nid, lru_shrinker_id(lru));
171 unlock_list_lru(l, false);
172 atomic_long_inc(&nlru->nr_items);
173 return true;
174 }
175 unlock_list_lru(l, false);
176 return false;
177}
178
179bool list_lru_add_obj(struct list_lru *lru, struct list_head *item)
180{
181 bool ret;
182 int nid = page_to_nid(virt_to_page(item));
183
184 if (list_lru_memcg_aware(lru)) {
185 rcu_read_lock();
186 ret = list_lru_add(lru, item, nid, mem_cgroup_from_slab_obj(item));
187 rcu_read_unlock();
188 } else {
189 ret = list_lru_add(lru, item, nid, NULL);
190 }
191
192 return ret;
193}
194EXPORT_SYMBOL_GPL(list_lru_add_obj);
195
196/* The caller must ensure the memcg lifetime. */
197bool list_lru_del(struct list_lru *lru, struct list_head *item, int nid,
198 struct mem_cgroup *memcg)
199{
200 struct list_lru_node *nlru = &lru->node[nid];
201 struct list_lru_one *l;
202 l = lock_list_lru_of_memcg(lru, nid, memcg, false, false);
203 if (!l)
204 return false;
205 if (!list_empty(item)) {
206 list_del_init(item);
207 l->nr_items--;
208 unlock_list_lru(l, false);
209 atomic_long_dec(&nlru->nr_items);
210 return true;
211 }
212 unlock_list_lru(l, false);
213 return false;
214}
215
216bool list_lru_del_obj(struct list_lru *lru, struct list_head *item)
217{
218 bool ret;
219 int nid = page_to_nid(virt_to_page(item));
220
221 if (list_lru_memcg_aware(lru)) {
222 rcu_read_lock();
223 ret = list_lru_del(lru, item, nid, mem_cgroup_from_slab_obj(item));
224 rcu_read_unlock();
225 } else {
226 ret = list_lru_del(lru, item, nid, NULL);
227 }
228
229 return ret;
230}
231EXPORT_SYMBOL_GPL(list_lru_del_obj);
232
233void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
234{
235 list_del_init(item);
236 list->nr_items--;
237}
238EXPORT_SYMBOL_GPL(list_lru_isolate);
239
240void list_lru_isolate_move(struct list_lru_one *list, struct list_head *item,
241 struct list_head *head)
242{
243 list_move(item, head);
244 list->nr_items--;
245}
246EXPORT_SYMBOL_GPL(list_lru_isolate_move);
247
248unsigned long list_lru_count_one(struct list_lru *lru,
249 int nid, struct mem_cgroup *memcg)
250{
251 struct list_lru_one *l;
252 long count;
253
254 rcu_read_lock();
255 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
256 count = l ? READ_ONCE(l->nr_items) : 0;
257 rcu_read_unlock();
258
259 if (unlikely(count < 0))
260 count = 0;
261
262 return count;
263}
264EXPORT_SYMBOL_GPL(list_lru_count_one);
265
266unsigned long list_lru_count_node(struct list_lru *lru, int nid)
267{
268 struct list_lru_node *nlru;
269
270 nlru = &lru->node[nid];
271 return atomic_long_read(&nlru->nr_items);
272}
273EXPORT_SYMBOL_GPL(list_lru_count_node);
274
275static unsigned long
276__list_lru_walk_one(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
277 list_lru_walk_cb isolate, void *cb_arg,
278 unsigned long *nr_to_walk, bool irq_off)
279{
280 struct list_lru_node *nlru = &lru->node[nid];
281 struct list_lru_one *l = NULL;
282 struct list_head *item, *n;
283 unsigned long isolated = 0;
284
285restart:
286 l = lock_list_lru_of_memcg(lru, nid, memcg, irq_off, true);
287 if (!l)
288 return isolated;
289 list_for_each_safe(item, n, &l->list) {
290 enum lru_status ret;
291
292 /*
293 * decrement nr_to_walk first so that we don't livelock if we
294 * get stuck on large numbers of LRU_RETRY items
295 */
296 if (!*nr_to_walk)
297 break;
298 --*nr_to_walk;
299
300 ret = isolate(item, l, cb_arg);
301 switch (ret) {
302 /*
303 * LRU_RETRY, LRU_REMOVED_RETRY and LRU_STOP will drop the lru
304 * lock. List traversal will have to restart from scratch.
305 */
306 case LRU_RETRY:
307 goto restart;
308 case LRU_REMOVED_RETRY:
309 fallthrough;
310 case LRU_REMOVED:
311 isolated++;
312 atomic_long_dec(&nlru->nr_items);
313 if (ret == LRU_REMOVED_RETRY)
314 goto restart;
315 break;
316 case LRU_ROTATE:
317 list_move_tail(item, &l->list);
318 break;
319 case LRU_SKIP:
320 break;
321 case LRU_STOP:
322 goto out;
323 default:
324 BUG();
325 }
326 }
327 unlock_list_lru(l, irq_off);
328out:
329 return isolated;
330}
331
332unsigned long
333list_lru_walk_one(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
334 list_lru_walk_cb isolate, void *cb_arg,
335 unsigned long *nr_to_walk)
336{
337 return __list_lru_walk_one(lru, nid, memcg, isolate,
338 cb_arg, nr_to_walk, false);
339}
340EXPORT_SYMBOL_GPL(list_lru_walk_one);
341
342unsigned long
343list_lru_walk_one_irq(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
344 list_lru_walk_cb isolate, void *cb_arg,
345 unsigned long *nr_to_walk)
346{
347 return __list_lru_walk_one(lru, nid, memcg, isolate,
348 cb_arg, nr_to_walk, true);
349}
350
351unsigned long list_lru_walk_node(struct list_lru *lru, int nid,
352 list_lru_walk_cb isolate, void *cb_arg,
353 unsigned long *nr_to_walk)
354{
355 long isolated = 0;
356
357 isolated += list_lru_walk_one(lru, nid, NULL, isolate, cb_arg,
358 nr_to_walk);
359
360#ifdef CONFIG_MEMCG
361 if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
362 struct list_lru_memcg *mlru;
363 struct mem_cgroup *memcg;
364 unsigned long index;
365
366 xa_for_each(&lru->xa, index, mlru) {
367 rcu_read_lock();
368 memcg = mem_cgroup_from_id(index);
369 if (!mem_cgroup_tryget(memcg)) {
370 rcu_read_unlock();
371 continue;
372 }
373 rcu_read_unlock();
374 isolated += __list_lru_walk_one(lru, nid, memcg,
375 isolate, cb_arg,
376 nr_to_walk, false);
377 mem_cgroup_put(memcg);
378
379 if (*nr_to_walk <= 0)
380 break;
381 }
382 }
383#endif
384
385 return isolated;
386}
387EXPORT_SYMBOL_GPL(list_lru_walk_node);
388
389static void init_one_lru(struct list_lru *lru, struct list_lru_one *l)
390{
391 INIT_LIST_HEAD(&l->list);
392 spin_lock_init(&l->lock);
393 l->nr_items = 0;
394#ifdef CONFIG_LOCKDEP
395 if (lru->key)
396 lockdep_set_class(&l->lock, lru->key);
397#endif
398}
399
400#ifdef CONFIG_MEMCG
401static struct list_lru_memcg *memcg_init_list_lru_one(struct list_lru *lru, gfp_t gfp)
402{
403 int nid;
404 struct list_lru_memcg *mlru;
405
406 mlru = kmalloc(struct_size(mlru, node, nr_node_ids), gfp);
407 if (!mlru)
408 return NULL;
409
410 for_each_node(nid)
411 init_one_lru(lru, &mlru->node[nid]);
412
413 return mlru;
414}
415
416static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
417{
418 if (memcg_aware)
419 xa_init_flags(&lru->xa, XA_FLAGS_LOCK_IRQ);
420 lru->memcg_aware = memcg_aware;
421}
422
423static void memcg_destroy_list_lru(struct list_lru *lru)
424{
425 XA_STATE(xas, &lru->xa, 0);
426 struct list_lru_memcg *mlru;
427
428 if (!list_lru_memcg_aware(lru))
429 return;
430
431 xas_lock_irq(&xas);
432 xas_for_each(&xas, mlru, ULONG_MAX) {
433 kfree(mlru);
434 xas_store(&xas, NULL);
435 }
436 xas_unlock_irq(&xas);
437}
438
439static void memcg_reparent_list_lru_one(struct list_lru *lru, int nid,
440 struct list_lru_one *src,
441 struct mem_cgroup *dst_memcg)
442{
443 int dst_idx = dst_memcg->kmemcg_id;
444 struct list_lru_one *dst;
445
446 spin_lock_irq(&src->lock);
447 dst = list_lru_from_memcg_idx(lru, nid, dst_idx);
448 spin_lock_nested(&dst->lock, SINGLE_DEPTH_NESTING);
449
450 list_splice_init(&src->list, &dst->list);
451 if (src->nr_items) {
452 WARN_ON(src->nr_items < 0);
453 dst->nr_items += src->nr_items;
454 set_shrinker_bit(dst_memcg, nid, lru_shrinker_id(lru));
455 }
456 /* Mark the list_lru_one dead */
457 src->nr_items = LONG_MIN;
458
459 spin_unlock(&dst->lock);
460 spin_unlock_irq(&src->lock);
461}
462
463void memcg_reparent_list_lrus(struct mem_cgroup *memcg, struct mem_cgroup *parent)
464{
465 struct list_lru *lru;
466 int i;
467
468 mutex_lock(&list_lrus_mutex);
469 list_for_each_entry(lru, &memcg_list_lrus, list) {
470 struct list_lru_memcg *mlru;
471 XA_STATE(xas, &lru->xa, memcg->kmemcg_id);
472
473 /*
474 * Lock the Xarray to ensure no on going list_lru_memcg
475 * allocation and further allocation will see css_is_dying().
476 */
477 xas_lock_irq(&xas);
478 mlru = xas_store(&xas, NULL);
479 xas_unlock_irq(&xas);
480 if (!mlru)
481 continue;
482
483 /*
484 * With Xarray value set to NULL, holding the lru lock below
485 * prevents list_lru_{add,del,isolate} from touching the lru,
486 * safe to reparent.
487 */
488 for_each_node(i)
489 memcg_reparent_list_lru_one(lru, i, &mlru->node[i], parent);
490
491 /*
492 * Here all list_lrus corresponding to the cgroup are guaranteed
493 * to remain empty, we can safely free this lru, any further
494 * memcg_list_lru_alloc() call will simply bail out.
495 */
496 kvfree_rcu(mlru, rcu);
497 }
498 mutex_unlock(&list_lrus_mutex);
499}
500
501static inline bool memcg_list_lru_allocated(struct mem_cgroup *memcg,
502 struct list_lru *lru)
503{
504 int idx = memcg->kmemcg_id;
505
506 return idx < 0 || xa_load(&lru->xa, idx);
507}
508
509int memcg_list_lru_alloc(struct mem_cgroup *memcg, struct list_lru *lru,
510 gfp_t gfp)
511{
512 unsigned long flags;
513 struct list_lru_memcg *mlru;
514 struct mem_cgroup *pos, *parent;
515 XA_STATE(xas, &lru->xa, 0);
516
517 if (!list_lru_memcg_aware(lru) || memcg_list_lru_allocated(memcg, lru))
518 return 0;
519
520 gfp &= GFP_RECLAIM_MASK;
521 /*
522 * Because the list_lru can be reparented to the parent cgroup's
523 * list_lru, we should make sure that this cgroup and all its
524 * ancestors have allocated list_lru_memcg.
525 */
526 do {
527 /*
528 * Keep finding the farest parent that wasn't populated
529 * until found memcg itself.
530 */
531 pos = memcg;
532 parent = parent_mem_cgroup(pos);
533 while (!memcg_list_lru_allocated(parent, lru)) {
534 pos = parent;
535 parent = parent_mem_cgroup(pos);
536 }
537
538 mlru = memcg_init_list_lru_one(lru, gfp);
539 if (!mlru)
540 return -ENOMEM;
541 xas_set(&xas, pos->kmemcg_id);
542 do {
543 xas_lock_irqsave(&xas, flags);
544 if (!xas_load(&xas) && !css_is_dying(&pos->css)) {
545 xas_store(&xas, mlru);
546 if (!xas_error(&xas))
547 mlru = NULL;
548 }
549 xas_unlock_irqrestore(&xas, flags);
550 } while (xas_nomem(&xas, gfp));
551 if (mlru)
552 kfree(mlru);
553 } while (pos != memcg && !css_is_dying(&pos->css));
554
555 return xas_error(&xas);
556}
557#else
558static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
559{
560}
561
562static void memcg_destroy_list_lru(struct list_lru *lru)
563{
564}
565#endif /* CONFIG_MEMCG */
566
567int __list_lru_init(struct list_lru *lru, bool memcg_aware, struct shrinker *shrinker)
568{
569 int i;
570
571#ifdef CONFIG_MEMCG
572 if (shrinker)
573 lru->shrinker_id = shrinker->id;
574 else
575 lru->shrinker_id = -1;
576
577 if (mem_cgroup_kmem_disabled())
578 memcg_aware = false;
579#endif
580
581 lru->node = kcalloc(nr_node_ids, sizeof(*lru->node), GFP_KERNEL);
582 if (!lru->node)
583 return -ENOMEM;
584
585 for_each_node(i)
586 init_one_lru(lru, &lru->node[i].lru);
587
588 memcg_init_list_lru(lru, memcg_aware);
589 list_lru_register(lru);
590
591 return 0;
592}
593EXPORT_SYMBOL_GPL(__list_lru_init);
594
595void list_lru_destroy(struct list_lru *lru)
596{
597 /* Already destroyed or not yet initialized? */
598 if (!lru->node)
599 return;
600
601 list_lru_unregister(lru);
602
603 memcg_destroy_list_lru(lru);
604 kfree(lru->node);
605 lru->node = NULL;
606
607#ifdef CONFIG_MEMCG
608 lru->shrinker_id = -1;
609#endif
610}
611EXPORT_SYMBOL_GPL(list_lru_destroy);