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#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#else
63static void list_lru_register(struct list_lru *lru)
64{
65}
66
67static void list_lru_unregister(struct list_lru *lru)
68{
69}
70
71static int lru_shrinker_id(struct list_lru *lru)
72{
73 return -1;
74}
75
76static inline bool list_lru_memcg_aware(struct list_lru *lru)
77{
78 return false;
79}
80
81static inline struct list_lru_one *
82list_lru_from_memcg_idx(struct list_lru *lru, int nid, int idx)
83{
84 return &lru->node[nid].lru;
85}
86#endif /* CONFIG_MEMCG_KMEM */
87
88bool list_lru_add(struct list_lru *lru, struct list_head *item, int nid,
89 struct mem_cgroup *memcg)
90{
91 struct list_lru_node *nlru = &lru->node[nid];
92 struct list_lru_one *l;
93
94 spin_lock(&nlru->lock);
95 if (list_empty(item)) {
96 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
97 list_add_tail(item, &l->list);
98 /* Set shrinker bit if the first element was added */
99 if (!l->nr_items++)
100 set_shrinker_bit(memcg, nid, lru_shrinker_id(lru));
101 nlru->nr_items++;
102 spin_unlock(&nlru->lock);
103 return true;
104 }
105 spin_unlock(&nlru->lock);
106 return false;
107}
108EXPORT_SYMBOL_GPL(list_lru_add);
109
110bool list_lru_add_obj(struct list_lru *lru, struct list_head *item)
111{
112 int nid = page_to_nid(virt_to_page(item));
113 struct mem_cgroup *memcg = list_lru_memcg_aware(lru) ?
114 mem_cgroup_from_slab_obj(item) : NULL;
115
116 return list_lru_add(lru, item, nid, memcg);
117}
118EXPORT_SYMBOL_GPL(list_lru_add_obj);
119
120bool list_lru_del(struct list_lru *lru, struct list_head *item, int nid,
121 struct mem_cgroup *memcg)
122{
123 struct list_lru_node *nlru = &lru->node[nid];
124 struct list_lru_one *l;
125
126 spin_lock(&nlru->lock);
127 if (!list_empty(item)) {
128 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
129 list_del_init(item);
130 l->nr_items--;
131 nlru->nr_items--;
132 spin_unlock(&nlru->lock);
133 return true;
134 }
135 spin_unlock(&nlru->lock);
136 return false;
137}
138EXPORT_SYMBOL_GPL(list_lru_del);
139
140bool list_lru_del_obj(struct list_lru *lru, struct list_head *item)
141{
142 int nid = page_to_nid(virt_to_page(item));
143 struct mem_cgroup *memcg = list_lru_memcg_aware(lru) ?
144 mem_cgroup_from_slab_obj(item) : NULL;
145
146 return list_lru_del(lru, item, nid, memcg);
147}
148EXPORT_SYMBOL_GPL(list_lru_del_obj);
149
150void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
151{
152 list_del_init(item);
153 list->nr_items--;
154}
155EXPORT_SYMBOL_GPL(list_lru_isolate);
156
157void list_lru_isolate_move(struct list_lru_one *list, struct list_head *item,
158 struct list_head *head)
159{
160 list_move(item, head);
161 list->nr_items--;
162}
163EXPORT_SYMBOL_GPL(list_lru_isolate_move);
164
165void list_lru_putback(struct list_lru *lru, struct list_head *item, int nid,
166 struct mem_cgroup *memcg)
167{
168 struct list_lru_one *list =
169 list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
170
171 if (list_empty(item)) {
172 list_add_tail(item, &list->list);
173 if (!list->nr_items++)
174 set_shrinker_bit(memcg, nid, lru_shrinker_id(lru));
175 }
176}
177EXPORT_SYMBOL_GPL(list_lru_putback);
178
179unsigned long list_lru_count_one(struct list_lru *lru,
180 int nid, struct mem_cgroup *memcg)
181{
182 struct list_lru_one *l;
183 long count;
184
185 rcu_read_lock();
186 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
187 count = l ? READ_ONCE(l->nr_items) : 0;
188 rcu_read_unlock();
189
190 if (unlikely(count < 0))
191 count = 0;
192
193 return count;
194}
195EXPORT_SYMBOL_GPL(list_lru_count_one);
196
197unsigned long list_lru_count_node(struct list_lru *lru, int nid)
198{
199 struct list_lru_node *nlru;
200
201 nlru = &lru->node[nid];
202 return nlru->nr_items;
203}
204EXPORT_SYMBOL_GPL(list_lru_count_node);
205
206static unsigned long
207__list_lru_walk_one(struct list_lru *lru, int nid, int memcg_idx,
208 list_lru_walk_cb isolate, void *cb_arg,
209 unsigned long *nr_to_walk)
210{
211 struct list_lru_node *nlru = &lru->node[nid];
212 struct list_lru_one *l;
213 struct list_head *item, *n;
214 unsigned long isolated = 0;
215
216restart:
217 l = list_lru_from_memcg_idx(lru, nid, memcg_idx);
218 if (!l)
219 goto out;
220
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 numbers 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 fallthrough;
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 }
264out:
265 return isolated;
266}
267
268unsigned long
269list_lru_walk_one(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
270 list_lru_walk_cb isolate, void *cb_arg,
271 unsigned long *nr_to_walk)
272{
273 struct list_lru_node *nlru = &lru->node[nid];
274 unsigned long ret;
275
276 spin_lock(&nlru->lock);
277 ret = __list_lru_walk_one(lru, nid, memcg_kmem_id(memcg), isolate,
278 cb_arg, nr_to_walk);
279 spin_unlock(&nlru->lock);
280 return ret;
281}
282EXPORT_SYMBOL_GPL(list_lru_walk_one);
283
284unsigned long
285list_lru_walk_one_irq(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
286 list_lru_walk_cb isolate, void *cb_arg,
287 unsigned long *nr_to_walk)
288{
289 struct list_lru_node *nlru = &lru->node[nid];
290 unsigned long ret;
291
292 spin_lock_irq(&nlru->lock);
293 ret = __list_lru_walk_one(lru, nid, memcg_kmem_id(memcg), isolate,
294 cb_arg, nr_to_walk);
295 spin_unlock_irq(&nlru->lock);
296 return ret;
297}
298
299unsigned long list_lru_walk_node(struct list_lru *lru, int nid,
300 list_lru_walk_cb isolate, void *cb_arg,
301 unsigned long *nr_to_walk)
302{
303 long isolated = 0;
304
305 isolated += list_lru_walk_one(lru, nid, NULL, isolate, cb_arg,
306 nr_to_walk);
307
308#ifdef CONFIG_MEMCG_KMEM
309 if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
310 struct list_lru_memcg *mlru;
311 unsigned long index;
312
313 xa_for_each(&lru->xa, index, mlru) {
314 struct list_lru_node *nlru = &lru->node[nid];
315
316 spin_lock(&nlru->lock);
317 isolated += __list_lru_walk_one(lru, nid, index,
318 isolate, cb_arg,
319 nr_to_walk);
320 spin_unlock(&nlru->lock);
321
322 if (*nr_to_walk <= 0)
323 break;
324 }
325 }
326#endif
327
328 return isolated;
329}
330EXPORT_SYMBOL_GPL(list_lru_walk_node);
331
332static void init_one_lru(struct list_lru_one *l)
333{
334 INIT_LIST_HEAD(&l->list);
335 l->nr_items = 0;
336}
337
338#ifdef CONFIG_MEMCG_KMEM
339static struct list_lru_memcg *memcg_init_list_lru_one(gfp_t gfp)
340{
341 int nid;
342 struct list_lru_memcg *mlru;
343
344 mlru = kmalloc(struct_size(mlru, node, nr_node_ids), gfp);
345 if (!mlru)
346 return NULL;
347
348 for_each_node(nid)
349 init_one_lru(&mlru->node[nid]);
350
351 return mlru;
352}
353
354static void memcg_list_lru_free(struct list_lru *lru, int src_idx)
355{
356 struct list_lru_memcg *mlru = xa_erase_irq(&lru->xa, src_idx);
357
358 /*
359 * The __list_lru_walk_one() can walk the list of this node.
360 * We need kvfree_rcu() here. And the walking of the list
361 * is under lru->node[nid]->lock, which can serve as a RCU
362 * read-side critical section.
363 */
364 if (mlru)
365 kvfree_rcu(mlru, rcu);
366}
367
368static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
369{
370 if (memcg_aware)
371 xa_init_flags(&lru->xa, XA_FLAGS_LOCK_IRQ);
372 lru->memcg_aware = memcg_aware;
373}
374
375static void memcg_destroy_list_lru(struct list_lru *lru)
376{
377 XA_STATE(xas, &lru->xa, 0);
378 struct list_lru_memcg *mlru;
379
380 if (!list_lru_memcg_aware(lru))
381 return;
382
383 xas_lock_irq(&xas);
384 xas_for_each(&xas, mlru, ULONG_MAX) {
385 kfree(mlru);
386 xas_store(&xas, NULL);
387 }
388 xas_unlock_irq(&xas);
389}
390
391static void memcg_reparent_list_lru_node(struct list_lru *lru, int nid,
392 int src_idx, struct mem_cgroup *dst_memcg)
393{
394 struct list_lru_node *nlru = &lru->node[nid];
395 int dst_idx = dst_memcg->kmemcg_id;
396 struct list_lru_one *src, *dst;
397
398 /*
399 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
400 * we have to use IRQ-safe primitives here to avoid deadlock.
401 */
402 spin_lock_irq(&nlru->lock);
403
404 src = list_lru_from_memcg_idx(lru, nid, src_idx);
405 if (!src)
406 goto out;
407 dst = list_lru_from_memcg_idx(lru, nid, dst_idx);
408
409 list_splice_init(&src->list, &dst->list);
410
411 if (src->nr_items) {
412 dst->nr_items += src->nr_items;
413 set_shrinker_bit(dst_memcg, nid, lru_shrinker_id(lru));
414 src->nr_items = 0;
415 }
416out:
417 spin_unlock_irq(&nlru->lock);
418}
419
420static void memcg_reparent_list_lru(struct list_lru *lru,
421 int src_idx, struct mem_cgroup *dst_memcg)
422{
423 int i;
424
425 for_each_node(i)
426 memcg_reparent_list_lru_node(lru, i, src_idx, dst_memcg);
427
428 memcg_list_lru_free(lru, src_idx);
429}
430
431void memcg_reparent_list_lrus(struct mem_cgroup *memcg, struct mem_cgroup *parent)
432{
433 struct cgroup_subsys_state *css;
434 struct list_lru *lru;
435 int src_idx = memcg->kmemcg_id;
436
437 /*
438 * Change kmemcg_id of this cgroup and all its descendants to the
439 * parent's id, and then move all entries from this cgroup's list_lrus
440 * to ones of the parent.
441 *
442 * After we have finished, all list_lrus corresponding to this cgroup
443 * are guaranteed to remain empty. So we can safely free this cgroup's
444 * list lrus in memcg_list_lru_free().
445 *
446 * Changing ->kmemcg_id to the parent can prevent memcg_list_lru_alloc()
447 * from allocating list lrus for this cgroup after memcg_list_lru_free()
448 * call.
449 */
450 rcu_read_lock();
451 css_for_each_descendant_pre(css, &memcg->css) {
452 struct mem_cgroup *child;
453
454 child = mem_cgroup_from_css(css);
455 WRITE_ONCE(child->kmemcg_id, parent->kmemcg_id);
456 }
457 rcu_read_unlock();
458
459 mutex_lock(&list_lrus_mutex);
460 list_for_each_entry(lru, &memcg_list_lrus, list)
461 memcg_reparent_list_lru(lru, src_idx, parent);
462 mutex_unlock(&list_lrus_mutex);
463}
464
465static inline bool memcg_list_lru_allocated(struct mem_cgroup *memcg,
466 struct list_lru *lru)
467{
468 int idx = memcg->kmemcg_id;
469
470 return idx < 0 || xa_load(&lru->xa, idx);
471}
472
473int memcg_list_lru_alloc(struct mem_cgroup *memcg, struct list_lru *lru,
474 gfp_t gfp)
475{
476 int i;
477 unsigned long flags;
478 struct list_lru_memcg_table {
479 struct list_lru_memcg *mlru;
480 struct mem_cgroup *memcg;
481 } *table;
482 XA_STATE(xas, &lru->xa, 0);
483
484 if (!list_lru_memcg_aware(lru) || memcg_list_lru_allocated(memcg, lru))
485 return 0;
486
487 gfp &= GFP_RECLAIM_MASK;
488 table = kmalloc_array(memcg->css.cgroup->level, sizeof(*table), gfp);
489 if (!table)
490 return -ENOMEM;
491
492 /*
493 * Because the list_lru can be reparented to the parent cgroup's
494 * list_lru, we should make sure that this cgroup and all its
495 * ancestors have allocated list_lru_memcg.
496 */
497 for (i = 0; memcg; memcg = parent_mem_cgroup(memcg), i++) {
498 if (memcg_list_lru_allocated(memcg, lru))
499 break;
500
501 table[i].memcg = memcg;
502 table[i].mlru = memcg_init_list_lru_one(gfp);
503 if (!table[i].mlru) {
504 while (i--)
505 kfree(table[i].mlru);
506 kfree(table);
507 return -ENOMEM;
508 }
509 }
510
511 xas_lock_irqsave(&xas, flags);
512 while (i--) {
513 int index = READ_ONCE(table[i].memcg->kmemcg_id);
514 struct list_lru_memcg *mlru = table[i].mlru;
515
516 xas_set(&xas, index);
517retry:
518 if (unlikely(index < 0 || xas_error(&xas) || xas_load(&xas))) {
519 kfree(mlru);
520 } else {
521 xas_store(&xas, mlru);
522 if (xas_error(&xas) == -ENOMEM) {
523 xas_unlock_irqrestore(&xas, flags);
524 if (xas_nomem(&xas, gfp))
525 xas_set_err(&xas, 0);
526 xas_lock_irqsave(&xas, flags);
527 /*
528 * The xas lock has been released, this memcg
529 * can be reparented before us. So reload
530 * memcg id. More details see the comments
531 * in memcg_reparent_list_lrus().
532 */
533 index = READ_ONCE(table[i].memcg->kmemcg_id);
534 if (index < 0)
535 xas_set_err(&xas, 0);
536 else if (!xas_error(&xas) && index != xas.xa_index)
537 xas_set(&xas, index);
538 goto retry;
539 }
540 }
541 }
542 /* xas_nomem() is used to free memory instead of memory allocation. */
543 if (xas.xa_alloc)
544 xas_nomem(&xas, gfp);
545 xas_unlock_irqrestore(&xas, flags);
546 kfree(table);
547
548 return xas_error(&xas);
549}
550#else
551static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
552{
553}
554
555static void memcg_destroy_list_lru(struct list_lru *lru)
556{
557}
558#endif /* CONFIG_MEMCG_KMEM */
559
560int __list_lru_init(struct list_lru *lru, bool memcg_aware,
561 struct lock_class_key *key, struct shrinker *shrinker)
562{
563 int i;
564
565#ifdef CONFIG_MEMCG_KMEM
566 if (shrinker)
567 lru->shrinker_id = shrinker->id;
568 else
569 lru->shrinker_id = -1;
570#endif
571
572 lru->node = kcalloc(nr_node_ids, sizeof(*lru->node), GFP_KERNEL);
573 if (!lru->node)
574 return -ENOMEM;
575
576 for_each_node(i) {
577 spin_lock_init(&lru->node[i].lock);
578 if (key)
579 lockdep_set_class(&lru->node[i].lock, key);
580 init_one_lru(&lru->node[i].lru);
581 }
582
583 memcg_init_list_lru(lru, memcg_aware);
584 list_lru_register(lru);
585
586 return 0;
587}
588EXPORT_SYMBOL_GPL(__list_lru_init);
589
590void list_lru_destroy(struct list_lru *lru)
591{
592 /* Already destroyed or not yet initialized? */
593 if (!lru->node)
594 return;
595
596 list_lru_unregister(lru);
597
598 memcg_destroy_list_lru(lru);
599 kfree(lru->node);
600 lru->node = NULL;
601
602#ifdef CONFIG_MEMCG_KMEM
603 lru->shrinker_id = -1;
604#endif
605}
606EXPORT_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
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);