xref: /openbmc/linux/mm/list_lru.c (revision 2f164822)
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
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 
63 static inline struct list_lru_one *
64 list_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));
79 out:
80 	if (memcg_ptr)
81 		*memcg_ptr = memcg;
82 	return l;
83 }
84 #else
85 static void list_lru_register(struct list_lru *lru)
86 {
87 }
88 
89 static void list_lru_unregister(struct list_lru *lru)
90 {
91 }
92 
93 static int lru_shrinker_id(struct list_lru *lru)
94 {
95 	return -1;
96 }
97 
98 static inline bool list_lru_memcg_aware(struct list_lru *lru)
99 {
100 	return false;
101 }
102 
103 static inline struct list_lru_one *
104 list_lru_from_memcg_idx(struct list_lru *lru, int nid, int idx)
105 {
106 	return &lru->node[nid].lru;
107 }
108 
109 static inline struct list_lru_one *
110 list_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 
119 bool 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 }
141 EXPORT_SYMBOL_GPL(list_lru_add);
142 
143 bool 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 }
161 EXPORT_SYMBOL_GPL(list_lru_del);
162 
163 void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
164 {
165 	list_del_init(item);
166 	list->nr_items--;
167 }
168 EXPORT_SYMBOL_GPL(list_lru_isolate);
169 
170 void 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 }
176 EXPORT_SYMBOL_GPL(list_lru_isolate_move);
177 
178 unsigned 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 }
194 EXPORT_SYMBOL_GPL(list_lru_count_one);
195 
196 unsigned 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 }
203 EXPORT_SYMBOL_GPL(list_lru_count_node);
204 
205 static 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 
215 restart:
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 	}
263 out:
264 	return isolated;
265 }
266 
267 unsigned long
268 list_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 }
281 EXPORT_SYMBOL_GPL(list_lru_walk_one);
282 
283 unsigned long
284 list_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 
298 unsigned 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 }
329 EXPORT_SYMBOL_GPL(list_lru_walk_node);
330 
331 static 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
338 static 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 
353 static 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 
367 static 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 
374 static 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 
390 static 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 	}
415 out:
416 	spin_unlock_irq(&nlru->lock);
417 }
418 
419 static 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 
430 void 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 
464 static 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 
472 int 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);
516 retry:
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
550 static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
551 {
552 }
553 
554 static void memcg_destroy_list_lru(struct list_lru *lru)
555 {
556 }
557 #endif /* CONFIG_MEMCG_KMEM */
558 
559 int __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 }
587 EXPORT_SYMBOL_GPL(__list_lru_init);
588 
589 void 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 }
605 EXPORT_SYMBOL_GPL(list_lru_destroy);
606