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
19 static LIST_HEAD(memcg_list_lrus);
20 static DEFINE_MUTEX(list_lrus_mutex);
21
22 static inline bool list_lru_memcg_aware(struct list_lru *lru)
23 {
24 return lru->memcg_aware;
25 }
26
27 static 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
37 static 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
47 static int lru_shrinker_id(struct list_lru *lru)
48 {
49 return lru->shrinker_id;
50 }
51
52 static inline struct list_lru_one *
53 list_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
63 static void list_lru_register(struct list_lru *lru)
64 {
65 }
66
67 static void list_lru_unregister(struct list_lru *lru)
68 {
69 }
70
71 static int lru_shrinker_id(struct list_lru *lru)
72 {
73 return -1;
74 }
75
76 static inline bool list_lru_memcg_aware(struct list_lru *lru)
77 {
78 return false;
79 }
80
81 static inline struct list_lru_one *
82 list_lru_from_memcg_idx(struct list_lru *lru, int nid, int idx)
83 {
84 return &lru->node[nid].lru;
85 }
86 #endif /* CONFIG_MEMCG */
87
88 /* The caller must ensure the memcg lifetime. */
89 bool list_lru_add(struct list_lru *lru, struct list_head *item, int nid,
90 struct mem_cgroup *memcg)
91 {
92 struct list_lru_node *nlru = &lru->node[nid];
93 struct list_lru_one *l;
94
95 spin_lock(&nlru->lock);
96 if (list_empty(item)) {
97 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
98 list_add_tail(item, &l->list);
99 /* Set shrinker bit if the first element was added */
100 if (!l->nr_items++)
101 set_shrinker_bit(memcg, nid, lru_shrinker_id(lru));
102 nlru->nr_items++;
103 spin_unlock(&nlru->lock);
104 return true;
105 }
106 spin_unlock(&nlru->lock);
107 return false;
108 }
109 EXPORT_SYMBOL_GPL(list_lru_add);
110
111 bool list_lru_add_obj(struct list_lru *lru, struct list_head *item)
112 {
113 bool ret;
114 int nid = page_to_nid(virt_to_page(item));
115
116 if (list_lru_memcg_aware(lru)) {
117 rcu_read_lock();
118 ret = list_lru_add(lru, item, nid, mem_cgroup_from_slab_obj(item));
119 rcu_read_unlock();
120 } else {
121 ret = list_lru_add(lru, item, nid, NULL);
122 }
123
124 return ret;
125 }
126 EXPORT_SYMBOL_GPL(list_lru_add_obj);
127
128 /* The caller must ensure the memcg lifetime. */
129 bool list_lru_del(struct list_lru *lru, struct list_head *item, int nid,
130 struct mem_cgroup *memcg)
131 {
132 struct list_lru_node *nlru = &lru->node[nid];
133 struct list_lru_one *l;
134
135 spin_lock(&nlru->lock);
136 if (!list_empty(item)) {
137 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
138 list_del_init(item);
139 l->nr_items--;
140 nlru->nr_items--;
141 spin_unlock(&nlru->lock);
142 return true;
143 }
144 spin_unlock(&nlru->lock);
145 return false;
146 }
147 EXPORT_SYMBOL_GPL(list_lru_del);
148
149 bool list_lru_del_obj(struct list_lru *lru, struct list_head *item)
150 {
151 bool ret;
152 int nid = page_to_nid(virt_to_page(item));
153
154 if (list_lru_memcg_aware(lru)) {
155 rcu_read_lock();
156 ret = list_lru_del(lru, item, nid, mem_cgroup_from_slab_obj(item));
157 rcu_read_unlock();
158 } else {
159 ret = list_lru_del(lru, item, nid, NULL);
160 }
161
162 return ret;
163 }
164 EXPORT_SYMBOL_GPL(list_lru_del_obj);
165
166 void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
167 {
168 list_del_init(item);
169 list->nr_items--;
170 }
171 EXPORT_SYMBOL_GPL(list_lru_isolate);
172
173 void list_lru_isolate_move(struct list_lru_one *list, struct list_head *item,
174 struct list_head *head)
175 {
176 list_move(item, head);
177 list->nr_items--;
178 }
179 EXPORT_SYMBOL_GPL(list_lru_isolate_move);
180
181 unsigned long list_lru_count_one(struct list_lru *lru,
182 int nid, struct mem_cgroup *memcg)
183 {
184 struct list_lru_one *l;
185 long count;
186
187 rcu_read_lock();
188 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
189 count = l ? READ_ONCE(l->nr_items) : 0;
190 rcu_read_unlock();
191
192 if (unlikely(count < 0))
193 count = 0;
194
195 return count;
196 }
197 EXPORT_SYMBOL_GPL(list_lru_count_one);
198
199 unsigned long list_lru_count_node(struct list_lru *lru, int nid)
200 {
201 struct list_lru_node *nlru;
202
203 nlru = &lru->node[nid];
204 return nlru->nr_items;
205 }
206 EXPORT_SYMBOL_GPL(list_lru_count_node);
207
208 static unsigned long
209 __list_lru_walk_one(struct list_lru *lru, int nid, int memcg_idx,
210 list_lru_walk_cb isolate, void *cb_arg,
211 unsigned long *nr_to_walk)
212 {
213 struct list_lru_node *nlru = &lru->node[nid];
214 struct list_lru_one *l;
215 struct list_head *item, *n;
216 unsigned long isolated = 0;
217
218 restart:
219 l = list_lru_from_memcg_idx(lru, nid, memcg_idx);
220 if (!l)
221 goto out;
222
223 list_for_each_safe(item, n, &l->list) {
224 enum lru_status ret;
225
226 /*
227 * decrement nr_to_walk first so that we don't livelock if we
228 * get stuck on large numbers of LRU_RETRY items
229 */
230 if (!*nr_to_walk)
231 break;
232 --*nr_to_walk;
233
234 ret = isolate(item, l, &nlru->lock, cb_arg);
235 switch (ret) {
236 case LRU_REMOVED_RETRY:
237 assert_spin_locked(&nlru->lock);
238 fallthrough;
239 case LRU_REMOVED:
240 isolated++;
241 nlru->nr_items--;
242 /*
243 * If the lru lock has been dropped, our list
244 * traversal is now invalid and so we have to
245 * restart from scratch.
246 */
247 if (ret == LRU_REMOVED_RETRY)
248 goto restart;
249 break;
250 case LRU_ROTATE:
251 list_move_tail(item, &l->list);
252 break;
253 case LRU_SKIP:
254 break;
255 case LRU_RETRY:
256 /*
257 * The lru lock has been dropped, our list traversal is
258 * now invalid and so we have to restart from scratch.
259 */
260 assert_spin_locked(&nlru->lock);
261 goto restart;
262 case LRU_STOP:
263 assert_spin_locked(&nlru->lock);
264 goto out;
265 default:
266 BUG();
267 }
268 }
269 out:
270 return isolated;
271 }
272
273 unsigned long
274 list_lru_walk_one(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(&nlru->lock);
282 ret = __list_lru_walk_one(lru, nid, memcg_kmem_id(memcg), isolate,
283 cb_arg, nr_to_walk);
284 spin_unlock(&nlru->lock);
285 return ret;
286 }
287 EXPORT_SYMBOL_GPL(list_lru_walk_one);
288
289 unsigned long
290 list_lru_walk_one_irq(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
291 list_lru_walk_cb isolate, void *cb_arg,
292 unsigned long *nr_to_walk)
293 {
294 struct list_lru_node *nlru = &lru->node[nid];
295 unsigned long ret;
296
297 spin_lock_irq(&nlru->lock);
298 ret = __list_lru_walk_one(lru, nid, memcg_kmem_id(memcg), isolate,
299 cb_arg, nr_to_walk);
300 spin_unlock_irq(&nlru->lock);
301 return ret;
302 }
303
304 unsigned long list_lru_walk_node(struct list_lru *lru, int nid,
305 list_lru_walk_cb isolate, void *cb_arg,
306 unsigned long *nr_to_walk)
307 {
308 long isolated = 0;
309
310 isolated += list_lru_walk_one(lru, nid, NULL, isolate, cb_arg,
311 nr_to_walk);
312
313 #ifdef CONFIG_MEMCG
314 if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
315 struct list_lru_memcg *mlru;
316 unsigned long index;
317
318 xa_for_each(&lru->xa, index, mlru) {
319 struct list_lru_node *nlru = &lru->node[nid];
320
321 spin_lock(&nlru->lock);
322 isolated += __list_lru_walk_one(lru, nid, index,
323 isolate, cb_arg,
324 nr_to_walk);
325 spin_unlock(&nlru->lock);
326
327 if (*nr_to_walk <= 0)
328 break;
329 }
330 }
331 #endif
332
333 return isolated;
334 }
335 EXPORT_SYMBOL_GPL(list_lru_walk_node);
336
337 static void init_one_lru(struct list_lru_one *l)
338 {
339 INIT_LIST_HEAD(&l->list);
340 l->nr_items = 0;
341 }
342
343 #ifdef CONFIG_MEMCG
344 static struct list_lru_memcg *memcg_init_list_lru_one(gfp_t gfp)
345 {
346 int nid;
347 struct list_lru_memcg *mlru;
348
349 mlru = kmalloc(struct_size(mlru, node, nr_node_ids), gfp);
350 if (!mlru)
351 return NULL;
352
353 for_each_node(nid)
354 init_one_lru(&mlru->node[nid]);
355
356 return mlru;
357 }
358
359 static void memcg_list_lru_free(struct list_lru *lru, int src_idx)
360 {
361 struct list_lru_memcg *mlru = xa_erase_irq(&lru->xa, src_idx);
362
363 /*
364 * The __list_lru_walk_one() can walk the list of this node.
365 * We need kvfree_rcu() here. And the walking of the list
366 * is under lru->node[nid]->lock, which can serve as a RCU
367 * read-side critical section.
368 */
369 if (mlru)
370 kvfree_rcu(mlru, rcu);
371 }
372
373 static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
374 {
375 if (memcg_aware)
376 xa_init_flags(&lru->xa, XA_FLAGS_LOCK_IRQ);
377 lru->memcg_aware = memcg_aware;
378 }
379
380 static void memcg_destroy_list_lru(struct list_lru *lru)
381 {
382 XA_STATE(xas, &lru->xa, 0);
383 struct list_lru_memcg *mlru;
384
385 if (!list_lru_memcg_aware(lru))
386 return;
387
388 xas_lock_irq(&xas);
389 xas_for_each(&xas, mlru, ULONG_MAX) {
390 kfree(mlru);
391 xas_store(&xas, NULL);
392 }
393 xas_unlock_irq(&xas);
394 }
395
396 static void memcg_reparent_list_lru_node(struct list_lru *lru, int nid,
397 int src_idx, struct mem_cgroup *dst_memcg)
398 {
399 struct list_lru_node *nlru = &lru->node[nid];
400 int dst_idx = dst_memcg->kmemcg_id;
401 struct list_lru_one *src, *dst;
402
403 /*
404 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
405 * we have to use IRQ-safe primitives here to avoid deadlock.
406 */
407 spin_lock_irq(&nlru->lock);
408
409 src = list_lru_from_memcg_idx(lru, nid, src_idx);
410 if (!src)
411 goto out;
412 dst = list_lru_from_memcg_idx(lru, nid, dst_idx);
413
414 list_splice_init(&src->list, &dst->list);
415
416 if (src->nr_items) {
417 dst->nr_items += src->nr_items;
418 set_shrinker_bit(dst_memcg, nid, lru_shrinker_id(lru));
419 src->nr_items = 0;
420 }
421 out:
422 spin_unlock_irq(&nlru->lock);
423 }
424
425 static void memcg_reparent_list_lru(struct list_lru *lru,
426 int src_idx, struct mem_cgroup *dst_memcg)
427 {
428 int i;
429
430 for_each_node(i)
431 memcg_reparent_list_lru_node(lru, i, src_idx, dst_memcg);
432
433 memcg_list_lru_free(lru, src_idx);
434 }
435
436 void memcg_reparent_list_lrus(struct mem_cgroup *memcg, struct mem_cgroup *parent)
437 {
438 struct cgroup_subsys_state *css;
439 struct list_lru *lru;
440 int src_idx = memcg->kmemcg_id;
441
442 /*
443 * Change kmemcg_id of this cgroup and all its descendants to the
444 * parent's id, and then move all entries from this cgroup's list_lrus
445 * to ones of the parent.
446 *
447 * After we have finished, all list_lrus corresponding to this cgroup
448 * are guaranteed to remain empty. So we can safely free this cgroup's
449 * list lrus in memcg_list_lru_free().
450 *
451 * Changing ->kmemcg_id to the parent can prevent memcg_list_lru_alloc()
452 * from allocating list lrus for this cgroup after memcg_list_lru_free()
453 * call.
454 */
455 rcu_read_lock();
456 css_for_each_descendant_pre(css, &memcg->css) {
457 struct mem_cgroup *child;
458
459 child = mem_cgroup_from_css(css);
460 WRITE_ONCE(child->kmemcg_id, parent->kmemcg_id);
461 }
462 rcu_read_unlock();
463
464 mutex_lock(&list_lrus_mutex);
465 list_for_each_entry(lru, &memcg_list_lrus, list)
466 memcg_reparent_list_lru(lru, src_idx, parent);
467 mutex_unlock(&list_lrus_mutex);
468 }
469
470 static inline bool memcg_list_lru_allocated(struct mem_cgroup *memcg,
471 struct list_lru *lru)
472 {
473 int idx = memcg->kmemcg_id;
474
475 return idx < 0 || xa_load(&lru->xa, idx);
476 }
477
478 int memcg_list_lru_alloc(struct mem_cgroup *memcg, struct list_lru *lru,
479 gfp_t gfp)
480 {
481 int i;
482 unsigned long flags;
483 struct list_lru_memcg_table {
484 struct list_lru_memcg *mlru;
485 struct mem_cgroup *memcg;
486 } *table;
487 XA_STATE(xas, &lru->xa, 0);
488
489 if (!list_lru_memcg_aware(lru) || memcg_list_lru_allocated(memcg, lru))
490 return 0;
491
492 gfp &= GFP_RECLAIM_MASK;
493 table = kmalloc_array(memcg->css.cgroup->level, sizeof(*table), gfp);
494 if (!table)
495 return -ENOMEM;
496
497 /*
498 * Because the list_lru can be reparented to the parent cgroup's
499 * list_lru, we should make sure that this cgroup and all its
500 * ancestors have allocated list_lru_memcg.
501 */
502 for (i = 0; memcg; memcg = parent_mem_cgroup(memcg), i++) {
503 if (memcg_list_lru_allocated(memcg, lru))
504 break;
505
506 table[i].memcg = memcg;
507 table[i].mlru = memcg_init_list_lru_one(gfp);
508 if (!table[i].mlru) {
509 while (i--)
510 kfree(table[i].mlru);
511 kfree(table);
512 return -ENOMEM;
513 }
514 }
515
516 xas_lock_irqsave(&xas, flags);
517 while (i--) {
518 int index = READ_ONCE(table[i].memcg->kmemcg_id);
519 struct list_lru_memcg *mlru = table[i].mlru;
520
521 xas_set(&xas, index);
522 retry:
523 if (unlikely(index < 0 || xas_error(&xas) || xas_load(&xas))) {
524 kfree(mlru);
525 } else {
526 xas_store(&xas, mlru);
527 if (xas_error(&xas) == -ENOMEM) {
528 xas_unlock_irqrestore(&xas, flags);
529 if (xas_nomem(&xas, gfp))
530 xas_set_err(&xas, 0);
531 xas_lock_irqsave(&xas, flags);
532 /*
533 * The xas lock has been released, this memcg
534 * can be reparented before us. So reload
535 * memcg id. More details see the comments
536 * in memcg_reparent_list_lrus().
537 */
538 index = READ_ONCE(table[i].memcg->kmemcg_id);
539 if (index < 0)
540 xas_set_err(&xas, 0);
541 else if (!xas_error(&xas) && index != xas.xa_index)
542 xas_set(&xas, index);
543 goto retry;
544 }
545 }
546 }
547 /* xas_nomem() is used to free memory instead of memory allocation. */
548 if (xas.xa_alloc)
549 xas_nomem(&xas, gfp);
550 xas_unlock_irqrestore(&xas, flags);
551 kfree(table);
552
553 return xas_error(&xas);
554 }
555 #else
556 static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
557 {
558 }
559
560 static void memcg_destroy_list_lru(struct list_lru *lru)
561 {
562 }
563 #endif /* CONFIG_MEMCG */
564
565 int __list_lru_init(struct list_lru *lru, bool memcg_aware,
566 struct lock_class_key *key, struct shrinker *shrinker)
567 {
568 int i;
569
570 #ifdef CONFIG_MEMCG
571 if (shrinker)
572 lru->shrinker_id = shrinker->id;
573 else
574 lru->shrinker_id = -1;
575
576 if (mem_cgroup_kmem_disabled())
577 memcg_aware = false;
578 #endif
579
580 lru->node = kcalloc(nr_node_ids, sizeof(*lru->node), GFP_KERNEL);
581 if (!lru->node)
582 return -ENOMEM;
583
584 for_each_node(i) {
585 spin_lock_init(&lru->node[i].lock);
586 if (key)
587 lockdep_set_class(&lru->node[i].lock, key);
588 init_one_lru(&lru->node[i].lru);
589 }
590
591 memcg_init_list_lru(lru, memcg_aware);
592 list_lru_register(lru);
593
594 return 0;
595 }
596 EXPORT_SYMBOL_GPL(__list_lru_init);
597
598 void list_lru_destroy(struct list_lru *lru)
599 {
600 /* Already destroyed or not yet initialized? */
601 if (!lru->node)
602 return;
603
604 list_lru_unregister(lru);
605
606 memcg_destroy_list_lru(lru);
607 kfree(lru->node);
608 lru->node = NULL;
609
610 #ifdef CONFIG_MEMCG
611 lru->shrinker_id = -1;
612 #endif
613 }
614 EXPORT_SYMBOL_GPL(list_lru_destroy);
615
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.