Linux Audio

Check our new training course

In-person Linux kernel drivers training

Jun 16-20, 2025
Register
Loading...
Note: File does not exist in v3.1.
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Workingset detection
  4 *
  5 * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner
  6 */
  7
  8#include <linux/memcontrol.h>
  9#include <linux/mm_inline.h>
 10#include <linux/writeback.h>
 11#include <linux/shmem_fs.h>
 12#include <linux/pagemap.h>
 13#include <linux/atomic.h>
 14#include <linux/module.h>
 15#include <linux/swap.h>
 16#include <linux/dax.h>
 17#include <linux/fs.h>
 18#include <linux/mm.h>
 19
 20/*
 21 *		Double CLOCK lists
 22 *
 23 * Per node, two clock lists are maintained for file pages: the
 24 * inactive and the active list.  Freshly faulted pages start out at
 25 * the head of the inactive list and page reclaim scans pages from the
 26 * tail.  Pages that are accessed multiple times on the inactive list
 27 * are promoted to the active list, to protect them from reclaim,
 28 * whereas active pages are demoted to the inactive list when the
 29 * active list grows too big.
 30 *
 31 *   fault ------------------------+
 32 *                                 |
 33 *              +--------------+   |            +-------------+
 34 *   reclaim <- |   inactive   | <-+-- demotion |    active   | <--+
 35 *              +--------------+                +-------------+    |
 36 *                     |                                           |
 37 *                     +-------------- promotion ------------------+
 38 *
 39 *
 40 *		Access frequency and refault distance
 41 *
 42 * A workload is thrashing when its pages are frequently used but they
 43 * are evicted from the inactive list every time before another access
 44 * would have promoted them to the active list.
 45 *
 46 * In cases where the average access distance between thrashing pages
 47 * is bigger than the size of memory there is nothing that can be
 48 * done - the thrashing set could never fit into memory under any
 49 * circumstance.
 50 *
 51 * However, the average access distance could be bigger than the
 52 * inactive list, yet smaller than the size of memory.  In this case,
 53 * the set could fit into memory if it weren't for the currently
 54 * active pages - which may be used more, hopefully less frequently:
 55 *
 56 *      +-memory available to cache-+
 57 *      |                           |
 58 *      +-inactive------+-active----+
 59 *  a b | c d e f g h i | J K L M N |
 60 *      +---------------+-----------+
 61 *
 62 * It is prohibitively expensive to accurately track access frequency
 63 * of pages.  But a reasonable approximation can be made to measure
 64 * thrashing on the inactive list, after which refaulting pages can be
 65 * activated optimistically to compete with the existing active pages.
 66 *
 67 * Approximating inactive page access frequency - Observations:
 68 *
 69 * 1. When a page is accessed for the first time, it is added to the
 70 *    head of the inactive list, slides every existing inactive page
 71 *    towards the tail by one slot, and pushes the current tail page
 72 *    out of memory.
 73 *
 74 * 2. When a page is accessed for the second time, it is promoted to
 75 *    the active list, shrinking the inactive list by one slot.  This
 76 *    also slides all inactive pages that were faulted into the cache
 77 *    more recently than the activated page towards the tail of the
 78 *    inactive list.
 79 *
 80 * Thus:
 81 *
 82 * 1. The sum of evictions and activations between any two points in
 83 *    time indicate the minimum number of inactive pages accessed in
 84 *    between.
 85 *
 86 * 2. Moving one inactive page N page slots towards the tail of the
 87 *    list requires at least N inactive page accesses.
 88 *
 89 * Combining these:
 90 *
 91 * 1. When a page is finally evicted from memory, the number of
 92 *    inactive pages accessed while the page was in cache is at least
 93 *    the number of page slots on the inactive list.
 94 *
 95 * 2. In addition, measuring the sum of evictions and activations (E)
 96 *    at the time of a page's eviction, and comparing it to another
 97 *    reading (R) at the time the page faults back into memory tells
 98 *    the minimum number of accesses while the page was not cached.
 99 *    This is called the refault distance.
100 *
101 * Because the first access of the page was the fault and the second
102 * access the refault, we combine the in-cache distance with the
103 * out-of-cache distance to get the complete minimum access distance
104 * of this page:
105 *
106 *      NR_inactive + (R - E)
107 *
108 * And knowing the minimum access distance of a page, we can easily
109 * tell if the page would be able to stay in cache assuming all page
110 * slots in the cache were available:
111 *
112 *   NR_inactive + (R - E) <= NR_inactive + NR_active
113 *
114 * which can be further simplified to
115 *
116 *   (R - E) <= NR_active
117 *
118 * Put into words, the refault distance (out-of-cache) can be seen as
119 * a deficit in inactive list space (in-cache).  If the inactive list
120 * had (R - E) more page slots, the page would not have been evicted
121 * in between accesses, but activated instead.  And on a full system,
122 * the only thing eating into inactive list space is active pages.
123 *
124 *
125 *		Refaulting inactive pages
126 *
127 * All that is known about the active list is that the pages have been
128 * accessed more than once in the past.  This means that at any given
129 * time there is actually a good chance that pages on the active list
130 * are no longer in active use.
131 *
132 * So when a refault distance of (R - E) is observed and there are at
133 * least (R - E) active pages, the refaulting page is activated
134 * optimistically in the hope that (R - E) active pages are actually
135 * used less frequently than the refaulting page - or even not used at
136 * all anymore.
137 *
138 * That means if inactive cache is refaulting with a suitable refault
139 * distance, we assume the cache workingset is transitioning and put
140 * pressure on the current active list.
141 *
142 * If this is wrong and demotion kicks in, the pages which are truly
143 * used more frequently will be reactivated while the less frequently
144 * used once will be evicted from memory.
145 *
146 * But if this is right, the stale pages will be pushed out of memory
147 * and the used pages get to stay in cache.
148 *
149 *		Refaulting active pages
150 *
151 * If on the other hand the refaulting pages have recently been
152 * deactivated, it means that the active list is no longer protecting
153 * actively used cache from reclaim. The cache is NOT transitioning to
154 * a different workingset; the existing workingset is thrashing in the
155 * space allocated to the page cache.
156 *
157 *
158 *		Implementation
159 *
160 * For each node's LRU lists, a counter for inactive evictions and
161 * activations is maintained (node->nonresident_age).
162 *
163 * On eviction, a snapshot of this counter (along with some bits to
164 * identify the node) is stored in the now empty page cache
165 * slot of the evicted page.  This is called a shadow entry.
166 *
167 * On cache misses for which there are shadow entries, an eligible
168 * refault distance will immediately activate the refaulting page.
169 */
170
171#define WORKINGSET_SHIFT 1
172#define EVICTION_SHIFT	((BITS_PER_LONG - BITS_PER_XA_VALUE) +	\
173			 WORKINGSET_SHIFT + NODES_SHIFT + \
174			 MEM_CGROUP_ID_SHIFT)
175#define EVICTION_MASK	(~0UL >> EVICTION_SHIFT)
176
177/*
178 * Eviction timestamps need to be able to cover the full range of
179 * actionable refaults. However, bits are tight in the xarray
180 * entry, and after storing the identifier for the lruvec there might
181 * not be enough left to represent every single actionable refault. In
182 * that case, we have to sacrifice granularity for distance, and group
183 * evictions into coarser buckets by shaving off lower timestamp bits.
184 */
185static unsigned int bucket_order __read_mostly;
186
187static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction,
188			 bool workingset)
189{
190	eviction &= EVICTION_MASK;
191	eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
192	eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
193	eviction = (eviction << WORKINGSET_SHIFT) | workingset;
194
195	return xa_mk_value(eviction);
196}
197
198static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
199			  unsigned long *evictionp, bool *workingsetp)
200{
201	unsigned long entry = xa_to_value(shadow);
202	int memcgid, nid;
203	bool workingset;
204
205	workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1);
206	entry >>= WORKINGSET_SHIFT;
207	nid = entry & ((1UL << NODES_SHIFT) - 1);
208	entry >>= NODES_SHIFT;
209	memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
210	entry >>= MEM_CGROUP_ID_SHIFT;
211
212	*memcgidp = memcgid;
213	*pgdat = NODE_DATA(nid);
214	*evictionp = entry;
215	*workingsetp = workingset;
216}
217
218#ifdef CONFIG_LRU_GEN
219
220static void *lru_gen_eviction(struct folio *folio)
221{
222	int hist;
223	unsigned long token;
224	unsigned long min_seq;
225	struct lruvec *lruvec;
226	struct lru_gen_struct *lrugen;
227	int type = folio_is_file_lru(folio);
228	int delta = folio_nr_pages(folio);
229	int refs = folio_lru_refs(folio);
230	int tier = lru_tier_from_refs(refs);
231	struct mem_cgroup *memcg = folio_memcg(folio);
232	struct pglist_data *pgdat = folio_pgdat(folio);
233
234	BUILD_BUG_ON(LRU_GEN_WIDTH + LRU_REFS_WIDTH > BITS_PER_LONG - EVICTION_SHIFT);
235
236	lruvec = mem_cgroup_lruvec(memcg, pgdat);
237	lrugen = &lruvec->lrugen;
238	min_seq = READ_ONCE(lrugen->min_seq[type]);
239	token = (min_seq << LRU_REFS_WIDTH) | max(refs - 1, 0);
240
241	hist = lru_hist_from_seq(min_seq);
242	atomic_long_add(delta, &lrugen->evicted[hist][type][tier]);
243
244	return pack_shadow(mem_cgroup_id(memcg), pgdat, token, refs);
245}
246
247static void lru_gen_refault(struct folio *folio, void *shadow)
248{
249	int hist, tier, refs;
250	int memcg_id;
251	bool workingset;
252	unsigned long token;
253	unsigned long min_seq;
254	struct lruvec *lruvec;
255	struct lru_gen_struct *lrugen;
256	struct mem_cgroup *memcg;
257	struct pglist_data *pgdat;
258	int type = folio_is_file_lru(folio);
259	int delta = folio_nr_pages(folio);
260
261	unpack_shadow(shadow, &memcg_id, &pgdat, &token, &workingset);
262
263	if (pgdat != folio_pgdat(folio))
264		return;
265
266	rcu_read_lock();
267
268	memcg = folio_memcg_rcu(folio);
269	if (memcg_id != mem_cgroup_id(memcg))
270		goto unlock;
271
272	lruvec = mem_cgroup_lruvec(memcg, pgdat);
273	lrugen = &lruvec->lrugen;
274
275	min_seq = READ_ONCE(lrugen->min_seq[type]);
276	if ((token >> LRU_REFS_WIDTH) != (min_seq & (EVICTION_MASK >> LRU_REFS_WIDTH)))
277		goto unlock;
278
279	hist = lru_hist_from_seq(min_seq);
280	/* see the comment in folio_lru_refs() */
281	refs = (token & (BIT(LRU_REFS_WIDTH) - 1)) + workingset;
282	tier = lru_tier_from_refs(refs);
283
284	atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]);
285	mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + type, delta);
286
287	/*
288	 * Count the following two cases as stalls:
289	 * 1. For pages accessed through page tables, hotter pages pushed out
290	 *    hot pages which refaulted immediately.
291	 * 2. For pages accessed multiple times through file descriptors,
292	 *    numbers of accesses might have been out of the range.
293	 */
294	if (lru_gen_in_fault() || refs == BIT(LRU_REFS_WIDTH)) {
295		folio_set_workingset(folio);
296		mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + type, delta);
297	}
298unlock:
299	rcu_read_unlock();
300}
301
302#else /* !CONFIG_LRU_GEN */
303
304static void *lru_gen_eviction(struct folio *folio)
305{
306	return NULL;
307}
308
309static void lru_gen_refault(struct folio *folio, void *shadow)
310{
311}
312
313#endif /* CONFIG_LRU_GEN */
314
315/**
316 * workingset_age_nonresident - age non-resident entries as LRU ages
317 * @lruvec: the lruvec that was aged
318 * @nr_pages: the number of pages to count
319 *
320 * As in-memory pages are aged, non-resident pages need to be aged as
321 * well, in order for the refault distances later on to be comparable
322 * to the in-memory dimensions. This function allows reclaim and LRU
323 * operations to drive the non-resident aging along in parallel.
324 */
325void workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages)
326{
327	/*
328	 * Reclaiming a cgroup means reclaiming all its children in a
329	 * round-robin fashion. That means that each cgroup has an LRU
330	 * order that is composed of the LRU orders of its child
331	 * cgroups; and every page has an LRU position not just in the
332	 * cgroup that owns it, but in all of that group's ancestors.
333	 *
334	 * So when the physical inactive list of a leaf cgroup ages,
335	 * the virtual inactive lists of all its parents, including
336	 * the root cgroup's, age as well.
337	 */
338	do {
339		atomic_long_add(nr_pages, &lruvec->nonresident_age);
340	} while ((lruvec = parent_lruvec(lruvec)));
341}
342
343/**
344 * workingset_eviction - note the eviction of a folio from memory
345 * @target_memcg: the cgroup that is causing the reclaim
346 * @folio: the folio being evicted
347 *
348 * Return: a shadow entry to be stored in @folio->mapping->i_pages in place
349 * of the evicted @folio so that a later refault can be detected.
350 */
351void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg)
352{
353	struct pglist_data *pgdat = folio_pgdat(folio);
354	unsigned long eviction;
355	struct lruvec *lruvec;
356	int memcgid;
357
358	/* Folio is fully exclusive and pins folio's memory cgroup pointer */
359	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
360	VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
361	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
362
363	if (lru_gen_enabled())
364		return lru_gen_eviction(folio);
365
366	lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
367	/* XXX: target_memcg can be NULL, go through lruvec */
368	memcgid = mem_cgroup_id(lruvec_memcg(lruvec));
369	eviction = atomic_long_read(&lruvec->nonresident_age);
370	eviction >>= bucket_order;
371	workingset_age_nonresident(lruvec, folio_nr_pages(folio));
372	return pack_shadow(memcgid, pgdat, eviction,
373				folio_test_workingset(folio));
374}
375
376/**
377 * workingset_refault - Evaluate the refault of a previously evicted folio.
378 * @folio: The freshly allocated replacement folio.
379 * @shadow: Shadow entry of the evicted folio.
380 *
381 * Calculates and evaluates the refault distance of the previously
382 * evicted folio in the context of the node and the memcg whose memory
383 * pressure caused the eviction.
384 */
385void workingset_refault(struct folio *folio, void *shadow)
386{
387	bool file = folio_is_file_lru(folio);
388	struct mem_cgroup *eviction_memcg;
389	struct lruvec *eviction_lruvec;
390	unsigned long refault_distance;
391	unsigned long workingset_size;
392	struct pglist_data *pgdat;
393	struct mem_cgroup *memcg;
394	unsigned long eviction;
395	struct lruvec *lruvec;
396	unsigned long refault;
397	bool workingset;
398	int memcgid;
399	long nr;
400
401	if (lru_gen_enabled()) {
402		lru_gen_refault(folio, shadow);
403		return;
404	}
405
406	unpack_shadow(shadow, &memcgid, &pgdat, &eviction, &workingset);
407	eviction <<= bucket_order;
408
409	rcu_read_lock();
410	/*
411	 * Look up the memcg associated with the stored ID. It might
412	 * have been deleted since the folio's eviction.
413	 *
414	 * Note that in rare events the ID could have been recycled
415	 * for a new cgroup that refaults a shared folio. This is
416	 * impossible to tell from the available data. However, this
417	 * should be a rare and limited disturbance, and activations
418	 * are always speculative anyway. Ultimately, it's the aging
419	 * algorithm's job to shake out the minimum access frequency
420	 * for the active cache.
421	 *
422	 * XXX: On !CONFIG_MEMCG, this will always return NULL; it
423	 * would be better if the root_mem_cgroup existed in all
424	 * configurations instead.
425	 */
426	eviction_memcg = mem_cgroup_from_id(memcgid);
427	if (!mem_cgroup_disabled() && !eviction_memcg)
428		goto out;
429	eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
430	refault = atomic_long_read(&eviction_lruvec->nonresident_age);
431
432	/*
433	 * Calculate the refault distance
434	 *
435	 * The unsigned subtraction here gives an accurate distance
436	 * across nonresident_age overflows in most cases. There is a
437	 * special case: usually, shadow entries have a short lifetime
438	 * and are either refaulted or reclaimed along with the inode
439	 * before they get too old.  But it is not impossible for the
440	 * nonresident_age to lap a shadow entry in the field, which
441	 * can then result in a false small refault distance, leading
442	 * to a false activation should this old entry actually
443	 * refault again.  However, earlier kernels used to deactivate
444	 * unconditionally with *every* reclaim invocation for the
445	 * longest time, so the occasional inappropriate activation
446	 * leading to pressure on the active list is not a problem.
447	 */
448	refault_distance = (refault - eviction) & EVICTION_MASK;
449
450	/*
451	 * The activation decision for this folio is made at the level
452	 * where the eviction occurred, as that is where the LRU order
453	 * during folio reclaim is being determined.
454	 *
455	 * However, the cgroup that will own the folio is the one that
456	 * is actually experiencing the refault event.
457	 */
458	nr = folio_nr_pages(folio);
459	memcg = folio_memcg(folio);
460	lruvec = mem_cgroup_lruvec(memcg, pgdat);
461
462	mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr);
463
464	mem_cgroup_flush_stats_delayed();
465	/*
466	 * Compare the distance to the existing workingset size. We
467	 * don't activate pages that couldn't stay resident even if
468	 * all the memory was available to the workingset. Whether
469	 * workingset competition needs to consider anon or not depends
470	 * on having swap.
471	 */
472	workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE);
473	if (!file) {
474		workingset_size += lruvec_page_state(eviction_lruvec,
475						     NR_INACTIVE_FILE);
476	}
477	if (mem_cgroup_get_nr_swap_pages(memcg) > 0) {
478		workingset_size += lruvec_page_state(eviction_lruvec,
479						     NR_ACTIVE_ANON);
480		if (file) {
481			workingset_size += lruvec_page_state(eviction_lruvec,
482						     NR_INACTIVE_ANON);
483		}
484	}
485	if (refault_distance > workingset_size)
486		goto out;
487
488	folio_set_active(folio);
489	workingset_age_nonresident(lruvec, nr);
490	mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr);
491
492	/* Folio was active prior to eviction */
493	if (workingset) {
494		folio_set_workingset(folio);
495		/*
496		 * XXX: Move to folio_add_lru() when it supports new vs
497		 * putback
498		 */
499		lru_note_cost_refault(folio);
500		mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr);
501	}
502out:
503	rcu_read_unlock();
504}
505
506/**
507 * workingset_activation - note a page activation
508 * @folio: Folio that is being activated.
509 */
510void workingset_activation(struct folio *folio)
511{
512	struct mem_cgroup *memcg;
513
514	rcu_read_lock();
515	/*
516	 * Filter non-memcg pages here, e.g. unmap can call
517	 * mark_page_accessed() on VDSO pages.
518	 *
519	 * XXX: See workingset_refault() - this should return
520	 * root_mem_cgroup even for !CONFIG_MEMCG.
521	 */
522	memcg = folio_memcg_rcu(folio);
523	if (!mem_cgroup_disabled() && !memcg)
524		goto out;
525	workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio));
526out:
527	rcu_read_unlock();
528}
529
530/*
531 * Shadow entries reflect the share of the working set that does not
532 * fit into memory, so their number depends on the access pattern of
533 * the workload.  In most cases, they will refault or get reclaimed
534 * along with the inode, but a (malicious) workload that streams
535 * through files with a total size several times that of available
536 * memory, while preventing the inodes from being reclaimed, can
537 * create excessive amounts of shadow nodes.  To keep a lid on this,
538 * track shadow nodes and reclaim them when they grow way past the
539 * point where they would still be useful.
540 */
541
542struct list_lru shadow_nodes;
543
544void workingset_update_node(struct xa_node *node)
545{
546	struct address_space *mapping;
547
548	/*
549	 * Track non-empty nodes that contain only shadow entries;
550	 * unlink those that contain pages or are being freed.
551	 *
552	 * Avoid acquiring the list_lru lock when the nodes are
553	 * already where they should be. The list_empty() test is safe
554	 * as node->private_list is protected by the i_pages lock.
555	 */
556	mapping = container_of(node->array, struct address_space, i_pages);
557	lockdep_assert_held(&mapping->i_pages.xa_lock);
558
559	if (node->count && node->count == node->nr_values) {
560		if (list_empty(&node->private_list)) {
561			list_lru_add(&shadow_nodes, &node->private_list);
562			__inc_lruvec_kmem_state(node, WORKINGSET_NODES);
563		}
564	} else {
565		if (!list_empty(&node->private_list)) {
566			list_lru_del(&shadow_nodes, &node->private_list);
567			__dec_lruvec_kmem_state(node, WORKINGSET_NODES);
568		}
569	}
570}
571
572static unsigned long count_shadow_nodes(struct shrinker *shrinker,
573					struct shrink_control *sc)
574{
575	unsigned long max_nodes;
576	unsigned long nodes;
577	unsigned long pages;
578
579	nodes = list_lru_shrink_count(&shadow_nodes, sc);
580	if (!nodes)
581		return SHRINK_EMPTY;
582
583	/*
584	 * Approximate a reasonable limit for the nodes
585	 * containing shadow entries. We don't need to keep more
586	 * shadow entries than possible pages on the active list,
587	 * since refault distances bigger than that are dismissed.
588	 *
589	 * The size of the active list converges toward 100% of
590	 * overall page cache as memory grows, with only a tiny
591	 * inactive list. Assume the total cache size for that.
592	 *
593	 * Nodes might be sparsely populated, with only one shadow
594	 * entry in the extreme case. Obviously, we cannot keep one
595	 * node for every eligible shadow entry, so compromise on a
596	 * worst-case density of 1/8th. Below that, not all eligible
597	 * refaults can be detected anymore.
598	 *
599	 * On 64-bit with 7 xa_nodes per page and 64 slots
600	 * each, this will reclaim shadow entries when they consume
601	 * ~1.8% of available memory:
602	 *
603	 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
604	 */
605#ifdef CONFIG_MEMCG
606	if (sc->memcg) {
607		struct lruvec *lruvec;
608		int i;
609
610		lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
611		for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
612			pages += lruvec_page_state_local(lruvec,
613							 NR_LRU_BASE + i);
614		pages += lruvec_page_state_local(
615			lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT;
616		pages += lruvec_page_state_local(
617			lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT;
618	} else
619#endif
620		pages = node_present_pages(sc->nid);
621
622	max_nodes = pages >> (XA_CHUNK_SHIFT - 3);
623
624	if (nodes <= max_nodes)
625		return 0;
626	return nodes - max_nodes;
627}
628
629static enum lru_status shadow_lru_isolate(struct list_head *item,
630					  struct list_lru_one *lru,
631					  spinlock_t *lru_lock,
632					  void *arg) __must_hold(lru_lock)
633{
634	struct xa_node *node = container_of(item, struct xa_node, private_list);
635	struct address_space *mapping;
636	int ret;
637
638	/*
639	 * Page cache insertions and deletions synchronously maintain
640	 * the shadow node LRU under the i_pages lock and the
641	 * lru_lock.  Because the page cache tree is emptied before
642	 * the inode can be destroyed, holding the lru_lock pins any
643	 * address_space that has nodes on the LRU.
644	 *
645	 * We can then safely transition to the i_pages lock to
646	 * pin only the address_space of the particular node we want
647	 * to reclaim, take the node off-LRU, and drop the lru_lock.
648	 */
649
650	mapping = container_of(node->array, struct address_space, i_pages);
651
652	/* Coming from the list, invert the lock order */
653	if (!xa_trylock(&mapping->i_pages)) {
654		spin_unlock_irq(lru_lock);
655		ret = LRU_RETRY;
656		goto out;
657	}
658
659	if (!spin_trylock(&mapping->host->i_lock)) {
660		xa_unlock(&mapping->i_pages);
661		spin_unlock_irq(lru_lock);
662		ret = LRU_RETRY;
663		goto out;
664	}
665
666	list_lru_isolate(lru, item);
667	__dec_lruvec_kmem_state(node, WORKINGSET_NODES);
668
669	spin_unlock(lru_lock);
670
671	/*
672	 * The nodes should only contain one or more shadow entries,
673	 * no pages, so we expect to be able to remove them all and
674	 * delete and free the empty node afterwards.
675	 */
676	if (WARN_ON_ONCE(!node->nr_values))
677		goto out_invalid;
678	if (WARN_ON_ONCE(node->count != node->nr_values))
679		goto out_invalid;
680	xa_delete_node(node, workingset_update_node);
681	__inc_lruvec_kmem_state(node, WORKINGSET_NODERECLAIM);
682
683out_invalid:
684	xa_unlock_irq(&mapping->i_pages);
685	if (mapping_shrinkable(mapping))
686		inode_add_lru(mapping->host);
687	spin_unlock(&mapping->host->i_lock);
688	ret = LRU_REMOVED_RETRY;
689out:
690	cond_resched();
691	spin_lock_irq(lru_lock);
692	return ret;
693}
694
695static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
696				       struct shrink_control *sc)
697{
698	/* list_lru lock nests inside the IRQ-safe i_pages lock */
699	return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate,
700					NULL);
701}
702
703static struct shrinker workingset_shadow_shrinker = {
704	.count_objects = count_shadow_nodes,
705	.scan_objects = scan_shadow_nodes,
706	.seeks = 0, /* ->count reports only fully expendable nodes */
707	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
708};
709
710/*
711 * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
712 * i_pages lock.
713 */
714static struct lock_class_key shadow_nodes_key;
715
716static int __init workingset_init(void)
717{
718	unsigned int timestamp_bits;
719	unsigned int max_order;
720	int ret;
721
722	BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
723	/*
724	 * Calculate the eviction bucket size to cover the longest
725	 * actionable refault distance, which is currently half of
726	 * memory (totalram_pages/2). However, memory hotplug may add
727	 * some more pages at runtime, so keep working with up to
728	 * double the initial memory by using totalram_pages as-is.
729	 */
730	timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
731	max_order = fls_long(totalram_pages() - 1);
732	if (max_order > timestamp_bits)
733		bucket_order = max_order - timestamp_bits;
734	pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
735	       timestamp_bits, max_order, bucket_order);
736
737	ret = prealloc_shrinker(&workingset_shadow_shrinker, "mm-shadow");
738	if (ret)
739		goto err;
740	ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key,
741			      &workingset_shadow_shrinker);
742	if (ret)
743		goto err_list_lru;
744	register_shrinker_prepared(&workingset_shadow_shrinker);
745	return 0;
746err_list_lru:
747	free_prealloced_shrinker(&workingset_shadow_shrinker);
748err:
749	return ret;
750}
751module_init(workingset_init);