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 #include "internal.h"
20 
21 /*
22  *		Double CLOCK lists
23  *
24  * Per node, two clock lists are maintained for file pages: the
25  * inactive and the active list.  Freshly faulted pages start out at
26  * the head of the inactive list and page reclaim scans pages from the
27  * tail.  Pages that are accessed multiple times on the inactive list
28  * are promoted to the active list, to protect them from reclaim,
29  * whereas active pages are demoted to the inactive list when the
30  * active list grows too big.
31  *
32  *   fault ------------------------+
33  *                                 |
34  *              +--------------+   |            +-------------+
35  *   reclaim <- |   inactive   | <-+-- demotion |    active   | <--+
36  *              +--------------+                +-------------+    |
37  *                     |                                           |
38  *                     +-------------- promotion ------------------+
39  *
40  *
41  *		Access frequency and refault distance
42  *
43  * A workload is thrashing when its pages are frequently used but they
44  * are evicted from the inactive list every time before another access
45  * would have promoted them to the active list.
46  *
47  * In cases where the average access distance between thrashing pages
48  * is bigger than the size of memory there is nothing that can be
49  * done - the thrashing set could never fit into memory under any
50  * circumstance.
51  *
52  * However, the average access distance could be bigger than the
53  * inactive list, yet smaller than the size of memory.  In this case,
54  * the set could fit into memory if it weren't for the currently
55  * active pages - which may be used more, hopefully less frequently:
56  *
57  *      +-memory available to cache-+
58  *      |                           |
59  *      +-inactive------+-active----+
60  *  a b | c d e f g h i | J K L M N |
61  *      +---------------+-----------+
62  *
63  * It is prohibitively expensive to accurately track access frequency
64  * of pages.  But a reasonable approximation can be made to measure
65  * thrashing on the inactive list, after which refaulting pages can be
66  * activated optimistically to compete with the existing active pages.
67  *
68  * Approximating inactive page access frequency - Observations:
69  *
70  * 1. When a page is accessed for the first time, it is added to the
71  *    head of the inactive list, slides every existing inactive page
72  *    towards the tail by one slot, and pushes the current tail page
73  *    out of memory.
74  *
75  * 2. When a page is accessed for the second time, it is promoted to
76  *    the active list, shrinking the inactive list by one slot.  This
77  *    also slides all inactive pages that were faulted into the cache
78  *    more recently than the activated page towards the tail of the
79  *    inactive list.
80  *
81  * Thus:
82  *
83  * 1. The sum of evictions and activations between any two points in
84  *    time indicate the minimum number of inactive pages accessed in
85  *    between.
86  *
87  * 2. Moving one inactive page N page slots towards the tail of the
88  *    list requires at least N inactive page accesses.
89  *
90  * Combining these:
91  *
92  * 1. When a page is finally evicted from memory, the number of
93  *    inactive pages accessed while the page was in cache is at least
94  *    the number of page slots on the inactive list.
95  *
96  * 2. In addition, measuring the sum of evictions and activations (E)
97  *    at the time of a page's eviction, and comparing it to another
98  *    reading (R) at the time the page faults back into memory tells
99  *    the minimum number of accesses while the page was not cached.
100  *    This is called the refault distance.
101  *
102  * Because the first access of the page was the fault and the second
103  * access the refault, we combine the in-cache distance with the
104  * out-of-cache distance to get the complete minimum access distance
105  * of this page:
106  *
107  *      NR_inactive + (R - E)
108  *
109  * And knowing the minimum access distance of a page, we can easily
110  * tell if the page would be able to stay in cache assuming all page
111  * slots in the cache were available:
112  *
113  *   NR_inactive + (R - E) <= NR_inactive + NR_active
114  *
115  * If we have swap we should consider about NR_inactive_anon and
116  * NR_active_anon, so for page cache and anonymous respectively:
117  *
118  *   NR_inactive_file + (R - E) <= NR_inactive_file + NR_active_file
119  *   + NR_inactive_anon + NR_active_anon
120  *
121  *   NR_inactive_anon + (R - E) <= NR_inactive_anon + NR_active_anon
122  *   + NR_inactive_file + NR_active_file
123  *
124  * Which can be further simplified to:
125  *
126  *   (R - E) <= NR_active_file + NR_inactive_anon + NR_active_anon
127  *
128  *   (R - E) <= NR_active_anon + NR_inactive_file + NR_active_file
129  *
130  * Put into words, the refault distance (out-of-cache) can be seen as
131  * a deficit in inactive list space (in-cache).  If the inactive list
132  * had (R - E) more page slots, the page would not have been evicted
133  * in between accesses, but activated instead.  And on a full system,
134  * the only thing eating into inactive list space is active pages.
135  *
136  *
137  *		Refaulting inactive pages
138  *
139  * All that is known about the active list is that the pages have been
140  * accessed more than once in the past.  This means that at any given
141  * time there is actually a good chance that pages on the active list
142  * are no longer in active use.
143  *
144  * So when a refault distance of (R - E) is observed and there are at
145  * least (R - E) pages in the userspace workingset, the refaulting page
146  * is activated optimistically in the hope that (R - E) pages are actually
147  * used less frequently than the refaulting page - or even not used at
148  * all anymore.
149  *
150  * That means if inactive cache is refaulting with a suitable refault
151  * distance, we assume the cache workingset is transitioning and put
152  * pressure on the current workingset.
153  *
154  * If this is wrong and demotion kicks in, the pages which are truly
155  * used more frequently will be reactivated while the less frequently
156  * used once will be evicted from memory.
157  *
158  * But if this is right, the stale pages will be pushed out of memory
159  * and the used pages get to stay in cache.
160  *
161  *		Refaulting active pages
162  *
163  * If on the other hand the refaulting pages have recently been
164  * deactivated, it means that the active list is no longer protecting
165  * actively used cache from reclaim. The cache is NOT transitioning to
166  * a different workingset; the existing workingset is thrashing in the
167  * space allocated to the page cache.
168  *
169  *
170  *		Implementation
171  *
172  * For each node's LRU lists, a counter for inactive evictions and
173  * activations is maintained (node->nonresident_age).
174  *
175  * On eviction, a snapshot of this counter (along with some bits to
176  * identify the node) is stored in the now empty page cache
177  * slot of the evicted page.  This is called a shadow entry.
178  *
179  * On cache misses for which there are shadow entries, an eligible
180  * refault distance will immediately activate the refaulting page.
181  */
182 
183 #define WORKINGSET_SHIFT 1
184 #define EVICTION_SHIFT	((BITS_PER_LONG - BITS_PER_XA_VALUE) +	\
185 			 WORKINGSET_SHIFT + NODES_SHIFT + \
186 			 MEM_CGROUP_ID_SHIFT)
187 #define EVICTION_MASK	(~0UL >> EVICTION_SHIFT)
188 
189 /*
190  * Eviction timestamps need to be able to cover the full range of
191  * actionable refaults. However, bits are tight in the xarray
192  * entry, and after storing the identifier for the lruvec there might
193  * not be enough left to represent every single actionable refault. In
194  * that case, we have to sacrifice granularity for distance, and group
195  * evictions into coarser buckets by shaving off lower timestamp bits.
196  */
197 static unsigned int bucket_order __read_mostly;
198 
pack_shadow(int memcgid,pg_data_t * pgdat,unsigned long eviction,bool workingset)199 static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction,
200 			 bool workingset)
201 {
202 	eviction &= EVICTION_MASK;
203 	eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
204 	eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
205 	eviction = (eviction << WORKINGSET_SHIFT) | workingset;
206 
207 	return xa_mk_value(eviction);
208 }
209 
unpack_shadow(void * shadow,int * memcgidp,pg_data_t ** pgdat,unsigned long * evictionp,bool * workingsetp)210 static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
211 			  unsigned long *evictionp, bool *workingsetp)
212 {
213 	unsigned long entry = xa_to_value(shadow);
214 	int memcgid, nid;
215 	bool workingset;
216 
217 	workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1);
218 	entry >>= WORKINGSET_SHIFT;
219 	nid = entry & ((1UL << NODES_SHIFT) - 1);
220 	entry >>= NODES_SHIFT;
221 	memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
222 	entry >>= MEM_CGROUP_ID_SHIFT;
223 
224 	*memcgidp = memcgid;
225 	*pgdat = NODE_DATA(nid);
226 	*evictionp = entry;
227 	*workingsetp = workingset;
228 }
229 
230 #ifdef CONFIG_LRU_GEN
231 
lru_gen_eviction(struct folio * folio)232 static void *lru_gen_eviction(struct folio *folio)
233 {
234 	int hist;
235 	unsigned long token;
236 	unsigned long min_seq;
237 	struct lruvec *lruvec;
238 	struct lru_gen_folio *lrugen;
239 	int type = folio_is_file_lru(folio);
240 	int delta = folio_nr_pages(folio);
241 	int refs = folio_lru_refs(folio);
242 	bool workingset = folio_test_workingset(folio);
243 	int tier = lru_tier_from_refs(refs, workingset);
244 	struct mem_cgroup *memcg = folio_memcg(folio);
245 	struct pglist_data *pgdat = folio_pgdat(folio);
246 
247 	BUILD_BUG_ON(LRU_GEN_WIDTH + LRU_REFS_WIDTH > BITS_PER_LONG - EVICTION_SHIFT);
248 
249 	lruvec = mem_cgroup_lruvec(memcg, pgdat);
250 	lrugen = &lruvec->lrugen;
251 	min_seq = READ_ONCE(lrugen->min_seq[type]);
252 	token = (min_seq << LRU_REFS_WIDTH) | max(refs - 1, 0);
253 
254 	hist = lru_hist_from_seq(min_seq);
255 	atomic_long_add(delta, &lrugen->evicted[hist][type][tier]);
256 
257 	return pack_shadow(mem_cgroup_id(memcg), pgdat, token, workingset);
258 }
259 
260 /*
261  * Tests if the shadow entry is for a folio that was recently evicted.
262  * Fills in @lruvec, @token, @workingset with the values unpacked from shadow.
263  */
lru_gen_test_recent(void * shadow,struct lruvec ** lruvec,unsigned long * token,bool * workingset)264 static bool lru_gen_test_recent(void *shadow, struct lruvec **lruvec,
265 				unsigned long *token, bool *workingset)
266 {
267 	int memcg_id;
268 	unsigned long max_seq;
269 	struct mem_cgroup *memcg;
270 	struct pglist_data *pgdat;
271 
272 	unpack_shadow(shadow, &memcg_id, &pgdat, token, workingset);
273 
274 	memcg = mem_cgroup_from_id(memcg_id);
275 	*lruvec = mem_cgroup_lruvec(memcg, pgdat);
276 
277 	max_seq = READ_ONCE((*lruvec)->lrugen.max_seq);
278 	max_seq &= EVICTION_MASK >> LRU_REFS_WIDTH;
279 
280 	return abs_diff(max_seq, *token >> LRU_REFS_WIDTH) < MAX_NR_GENS;
281 }
282 
lru_gen_refault(struct folio * folio,void * shadow)283 static void lru_gen_refault(struct folio *folio, void *shadow)
284 {
285 	bool recent;
286 	int hist, tier, refs;
287 	bool workingset;
288 	unsigned long token;
289 	struct lruvec *lruvec;
290 	struct lru_gen_folio *lrugen;
291 	int type = folio_is_file_lru(folio);
292 	int delta = folio_nr_pages(folio);
293 
294 	rcu_read_lock();
295 
296 	recent = lru_gen_test_recent(shadow, &lruvec, &token, &workingset);
297 	if (lruvec != folio_lruvec(folio))
298 		goto unlock;
299 
300 	mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + type, delta);
301 
302 	if (!recent)
303 		goto unlock;
304 
305 	lrugen = &lruvec->lrugen;
306 
307 	hist = lru_hist_from_seq(READ_ONCE(lrugen->min_seq[type]));
308 	refs = (token & (BIT(LRU_REFS_WIDTH) - 1)) + 1;
309 	tier = lru_tier_from_refs(refs, workingset);
310 
311 	atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]);
312 
313 	/* see folio_add_lru() where folio_set_active() will be called */
314 	if (lru_gen_in_fault())
315 		mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
316 
317 	if (workingset) {
318 		folio_set_workingset(folio);
319 		mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + type, delta);
320 	} else
321 		set_mask_bits(&folio->flags, LRU_REFS_MASK, (refs - 1UL) << LRU_REFS_PGOFF);
322 unlock:
323 	rcu_read_unlock();
324 }
325 
326 #else /* !CONFIG_LRU_GEN */
327 
lru_gen_eviction(struct folio * folio)328 static void *lru_gen_eviction(struct folio *folio)
329 {
330 	return NULL;
331 }
332 
lru_gen_test_recent(void * shadow,struct lruvec ** lruvec,unsigned long * token,bool * workingset)333 static bool lru_gen_test_recent(void *shadow, struct lruvec **lruvec,
334 				unsigned long *token, bool *workingset)
335 {
336 	return false;
337 }
338 
lru_gen_refault(struct folio * folio,void * shadow)339 static void lru_gen_refault(struct folio *folio, void *shadow)
340 {
341 }
342 
343 #endif /* CONFIG_LRU_GEN */
344 
345 /**
346  * workingset_age_nonresident - age non-resident entries as LRU ages
347  * @lruvec: the lruvec that was aged
348  * @nr_pages: the number of pages to count
349  *
350  * As in-memory pages are aged, non-resident pages need to be aged as
351  * well, in order for the refault distances later on to be comparable
352  * to the in-memory dimensions. This function allows reclaim and LRU
353  * operations to drive the non-resident aging along in parallel.
354  */
workingset_age_nonresident(struct lruvec * lruvec,unsigned long nr_pages)355 void workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages)
356 {
357 	/*
358 	 * Reclaiming a cgroup means reclaiming all its children in a
359 	 * round-robin fashion. That means that each cgroup has an LRU
360 	 * order that is composed of the LRU orders of its child
361 	 * cgroups; and every page has an LRU position not just in the
362 	 * cgroup that owns it, but in all of that group's ancestors.
363 	 *
364 	 * So when the physical inactive list of a leaf cgroup ages,
365 	 * the virtual inactive lists of all its parents, including
366 	 * the root cgroup's, age as well.
367 	 */
368 	do {
369 		atomic_long_add(nr_pages, &lruvec->nonresident_age);
370 	} while ((lruvec = parent_lruvec(lruvec)));
371 }
372 
373 /**
374  * workingset_eviction - note the eviction of a folio from memory
375  * @target_memcg: the cgroup that is causing the reclaim
376  * @folio: the folio being evicted
377  *
378  * Return: a shadow entry to be stored in @folio->mapping->i_pages in place
379  * of the evicted @folio so that a later refault can be detected.
380  */
workingset_eviction(struct folio * folio,struct mem_cgroup * target_memcg)381 void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg)
382 {
383 	struct pglist_data *pgdat = folio_pgdat(folio);
384 	unsigned long eviction;
385 	struct lruvec *lruvec;
386 	int memcgid;
387 
388 	/* Folio is fully exclusive and pins folio's memory cgroup pointer */
389 	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
390 	VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
391 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
392 
393 	if (lru_gen_enabled())
394 		return lru_gen_eviction(folio);
395 
396 	lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
397 	/* XXX: target_memcg can be NULL, go through lruvec */
398 	memcgid = mem_cgroup_id(lruvec_memcg(lruvec));
399 	eviction = atomic_long_read(&lruvec->nonresident_age);
400 	eviction >>= bucket_order;
401 	workingset_age_nonresident(lruvec, folio_nr_pages(folio));
402 	return pack_shadow(memcgid, pgdat, eviction,
403 				folio_test_workingset(folio));
404 }
405 
406 /**
407  * workingset_test_recent - tests if the shadow entry is for a folio that was
408  * recently evicted. Also fills in @workingset with the value unpacked from
409  * shadow.
410  * @shadow: the shadow entry to be tested.
411  * @file: whether the corresponding folio is from the file lru.
412  * @workingset: where the workingset value unpacked from shadow should
413  * be stored.
414  * @flush: whether to flush cgroup rstat.
415  *
416  * Return: true if the shadow is for a recently evicted folio; false otherwise.
417  */
workingset_test_recent(void * shadow,bool file,bool * workingset,bool flush)418 bool workingset_test_recent(void *shadow, bool file, bool *workingset,
419 				bool flush)
420 {
421 	struct mem_cgroup *eviction_memcg;
422 	struct lruvec *eviction_lruvec;
423 	unsigned long refault_distance;
424 	unsigned long workingset_size;
425 	unsigned long refault;
426 	int memcgid;
427 	struct pglist_data *pgdat;
428 	unsigned long eviction;
429 
430 	if (lru_gen_enabled()) {
431 		bool recent;
432 
433 		rcu_read_lock();
434 		recent = lru_gen_test_recent(shadow, &eviction_lruvec, &eviction, workingset);
435 		rcu_read_unlock();
436 		return recent;
437 	}
438 
439 	rcu_read_lock();
440 	unpack_shadow(shadow, &memcgid, &pgdat, &eviction, workingset);
441 	eviction <<= bucket_order;
442 
443 	/*
444 	 * Look up the memcg associated with the stored ID. It might
445 	 * have been deleted since the folio's eviction.
446 	 *
447 	 * Note that in rare events the ID could have been recycled
448 	 * for a new cgroup that refaults a shared folio. This is
449 	 * impossible to tell from the available data. However, this
450 	 * should be a rare and limited disturbance, and activations
451 	 * are always speculative anyway. Ultimately, it's the aging
452 	 * algorithm's job to shake out the minimum access frequency
453 	 * for the active cache.
454 	 *
455 	 * XXX: On !CONFIG_MEMCG, this will always return NULL; it
456 	 * would be better if the root_mem_cgroup existed in all
457 	 * configurations instead.
458 	 */
459 	eviction_memcg = mem_cgroup_from_id(memcgid);
460 	if (!mem_cgroup_tryget(eviction_memcg))
461 		eviction_memcg = NULL;
462 	rcu_read_unlock();
463 
464 	if (!mem_cgroup_disabled() && !eviction_memcg)
465 		return false;
466 	/*
467 	 * Flush stats (and potentially sleep) outside the RCU read section.
468 	 *
469 	 * Note that workingset_test_recent() itself might be called in RCU read
470 	 * section (for e.g, in cachestat) - these callers need to skip flushing
471 	 * stats (via the flush argument).
472 	 *
473 	 * XXX: With per-memcg flushing and thresholding, is ratelimiting
474 	 * still needed here?
475 	 */
476 	if (flush)
477 		mem_cgroup_flush_stats_ratelimited(eviction_memcg);
478 
479 	eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
480 	refault = atomic_long_read(&eviction_lruvec->nonresident_age);
481 
482 	/*
483 	 * Calculate the refault distance
484 	 *
485 	 * The unsigned subtraction here gives an accurate distance
486 	 * across nonresident_age overflows in most cases. There is a
487 	 * special case: usually, shadow entries have a short lifetime
488 	 * and are either refaulted or reclaimed along with the inode
489 	 * before they get too old.  But it is not impossible for the
490 	 * nonresident_age to lap a shadow entry in the field, which
491 	 * can then result in a false small refault distance, leading
492 	 * to a false activation should this old entry actually
493 	 * refault again.  However, earlier kernels used to deactivate
494 	 * unconditionally with *every* reclaim invocation for the
495 	 * longest time, so the occasional inappropriate activation
496 	 * leading to pressure on the active list is not a problem.
497 	 */
498 	refault_distance = (refault - eviction) & EVICTION_MASK;
499 
500 	/*
501 	 * Compare the distance to the existing workingset size. We
502 	 * don't activate pages that couldn't stay resident even if
503 	 * all the memory was available to the workingset. Whether
504 	 * workingset competition needs to consider anon or not depends
505 	 * on having free swap space.
506 	 */
507 	workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE);
508 	if (!file) {
509 		workingset_size += lruvec_page_state(eviction_lruvec,
510 						     NR_INACTIVE_FILE);
511 	}
512 	if (mem_cgroup_get_nr_swap_pages(eviction_memcg) > 0) {
513 		workingset_size += lruvec_page_state(eviction_lruvec,
514 						     NR_ACTIVE_ANON);
515 		if (file) {
516 			workingset_size += lruvec_page_state(eviction_lruvec,
517 						     NR_INACTIVE_ANON);
518 		}
519 	}
520 
521 	mem_cgroup_put(eviction_memcg);
522 	return refault_distance <= workingset_size;
523 }
524 
525 /**
526  * workingset_refault - Evaluate the refault of a previously evicted folio.
527  * @folio: The freshly allocated replacement folio.
528  * @shadow: Shadow entry of the evicted folio.
529  *
530  * Calculates and evaluates the refault distance of the previously
531  * evicted folio in the context of the node and the memcg whose memory
532  * pressure caused the eviction.
533  */
workingset_refault(struct folio * folio,void * shadow)534 void workingset_refault(struct folio *folio, void *shadow)
535 {
536 	bool file = folio_is_file_lru(folio);
537 	struct pglist_data *pgdat;
538 	struct mem_cgroup *memcg;
539 	struct lruvec *lruvec;
540 	bool workingset;
541 	long nr;
542 
543 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
544 
545 	if (lru_gen_enabled()) {
546 		lru_gen_refault(folio, shadow);
547 		return;
548 	}
549 
550 	/*
551 	 * The activation decision for this folio is made at the level
552 	 * where the eviction occurred, as that is where the LRU order
553 	 * during folio reclaim is being determined.
554 	 *
555 	 * However, the cgroup that will own the folio is the one that
556 	 * is actually experiencing the refault event. Make sure the folio is
557 	 * locked to guarantee folio_memcg() stability throughout.
558 	 */
559 	nr = folio_nr_pages(folio);
560 	memcg = folio_memcg(folio);
561 	pgdat = folio_pgdat(folio);
562 	lruvec = mem_cgroup_lruvec(memcg, pgdat);
563 
564 	mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr);
565 
566 	if (!workingset_test_recent(shadow, file, &workingset, true))
567 		return;
568 
569 	folio_set_active(folio);
570 	workingset_age_nonresident(lruvec, nr);
571 	mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr);
572 
573 	/* Folio was active prior to eviction */
574 	if (workingset) {
575 		folio_set_workingset(folio);
576 		/*
577 		 * XXX: Move to folio_add_lru() when it supports new vs
578 		 * putback
579 		 */
580 		lru_note_cost_refault(folio);
581 		mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr);
582 	}
583 }
584 
585 /**
586  * workingset_activation - note a page activation
587  * @folio: Folio that is being activated.
588  */
workingset_activation(struct folio * folio)589 void workingset_activation(struct folio *folio)
590 {
591 	/*
592 	 * Filter non-memcg pages here, e.g. unmap can call
593 	 * mark_page_accessed() on VDSO pages.
594 	 */
595 	if (mem_cgroup_disabled() || folio_memcg_charged(folio))
596 		workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio));
597 }
598 
599 /*
600  * Shadow entries reflect the share of the working set that does not
601  * fit into memory, so their number depends on the access pattern of
602  * the workload.  In most cases, they will refault or get reclaimed
603  * along with the inode, but a (malicious) workload that streams
604  * through files with a total size several times that of available
605  * memory, while preventing the inodes from being reclaimed, can
606  * create excessive amounts of shadow nodes.  To keep a lid on this,
607  * track shadow nodes and reclaim them when they grow way past the
608  * point where they would still be useful.
609  */
610 
611 struct list_lru shadow_nodes;
612 
workingset_update_node(struct xa_node * node)613 void workingset_update_node(struct xa_node *node)
614 {
615 	struct address_space *mapping;
616 	struct page *page = virt_to_page(node);
617 
618 	/*
619 	 * Track non-empty nodes that contain only shadow entries;
620 	 * unlink those that contain pages or are being freed.
621 	 *
622 	 * Avoid acquiring the list_lru lock when the nodes are
623 	 * already where they should be. The list_empty() test is safe
624 	 * as node->private_list is protected by the i_pages lock.
625 	 */
626 	mapping = container_of(node->array, struct address_space, i_pages);
627 	lockdep_assert_held(&mapping->i_pages.xa_lock);
628 
629 	if (node->count && node->count == node->nr_values) {
630 		if (list_empty(&node->private_list)) {
631 			list_lru_add_obj(&shadow_nodes, &node->private_list);
632 			__inc_node_page_state(page, WORKINGSET_NODES);
633 		}
634 	} else {
635 		if (!list_empty(&node->private_list)) {
636 			list_lru_del_obj(&shadow_nodes, &node->private_list);
637 			__dec_node_page_state(page, WORKINGSET_NODES);
638 		}
639 	}
640 }
641 
count_shadow_nodes(struct shrinker * shrinker,struct shrink_control * sc)642 static unsigned long count_shadow_nodes(struct shrinker *shrinker,
643 					struct shrink_control *sc)
644 {
645 	unsigned long max_nodes;
646 	unsigned long nodes;
647 	unsigned long pages;
648 
649 	nodes = list_lru_shrink_count(&shadow_nodes, sc);
650 	if (!nodes)
651 		return SHRINK_EMPTY;
652 
653 	/*
654 	 * Approximate a reasonable limit for the nodes
655 	 * containing shadow entries. We don't need to keep more
656 	 * shadow entries than possible pages on the active list,
657 	 * since refault distances bigger than that are dismissed.
658 	 *
659 	 * The size of the active list converges toward 100% of
660 	 * overall page cache as memory grows, with only a tiny
661 	 * inactive list. Assume the total cache size for that.
662 	 *
663 	 * Nodes might be sparsely populated, with only one shadow
664 	 * entry in the extreme case. Obviously, we cannot keep one
665 	 * node for every eligible shadow entry, so compromise on a
666 	 * worst-case density of 1/8th. Below that, not all eligible
667 	 * refaults can be detected anymore.
668 	 *
669 	 * On 64-bit with 7 xa_nodes per page and 64 slots
670 	 * each, this will reclaim shadow entries when they consume
671 	 * ~1.8% of available memory:
672 	 *
673 	 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
674 	 */
675 #ifdef CONFIG_MEMCG
676 	if (sc->memcg) {
677 		struct lruvec *lruvec;
678 		int i;
679 
680 		mem_cgroup_flush_stats_ratelimited(sc->memcg);
681 		lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
682 		for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
683 			pages += lruvec_page_state_local(lruvec,
684 							 NR_LRU_BASE + i);
685 		pages += lruvec_page_state_local(
686 			lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT;
687 		pages += lruvec_page_state_local(
688 			lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT;
689 	} else
690 #endif
691 		pages = node_present_pages(sc->nid);
692 
693 	max_nodes = pages >> (XA_CHUNK_SHIFT - 3);
694 
695 	if (nodes <= max_nodes)
696 		return 0;
697 	return nodes - max_nodes;
698 }
699 
shadow_lru_isolate(struct list_head * item,struct list_lru_one * lru,void * arg)700 static enum lru_status shadow_lru_isolate(struct list_head *item,
701 					  struct list_lru_one *lru,
702 					  void *arg) __must_hold(lru->lock)
703 {
704 	struct xa_node *node = container_of(item, struct xa_node, private_list);
705 	struct address_space *mapping;
706 	int ret;
707 
708 	/*
709 	 * Page cache insertions and deletions synchronously maintain
710 	 * the shadow node LRU under the i_pages lock and the
711 	 * &lru->lock. Because the page cache tree is emptied before
712 	 * the inode can be destroyed, holding the &lru->lock pins any
713 	 * address_space that has nodes on the LRU.
714 	 *
715 	 * We can then safely transition to the i_pages lock to
716 	 * pin only the address_space of the particular node we want
717 	 * to reclaim, take the node off-LRU, and drop the &lru->lock.
718 	 */
719 
720 	mapping = container_of(node->array, struct address_space, i_pages);
721 
722 	/* Coming from the list, invert the lock order */
723 	if (!xa_trylock(&mapping->i_pages)) {
724 		spin_unlock_irq(&lru->lock);
725 		ret = LRU_RETRY;
726 		goto out;
727 	}
728 
729 	/* For page cache we need to hold i_lock */
730 	if (mapping->host != NULL) {
731 		if (!spin_trylock(&mapping->host->i_lock)) {
732 			xa_unlock(&mapping->i_pages);
733 			spin_unlock_irq(&lru->lock);
734 			ret = LRU_RETRY;
735 			goto out;
736 		}
737 	}
738 
739 	list_lru_isolate(lru, item);
740 	__dec_node_page_state(virt_to_page(node), WORKINGSET_NODES);
741 
742 	spin_unlock(&lru->lock);
743 
744 	/*
745 	 * The nodes should only contain one or more shadow entries,
746 	 * no pages, so we expect to be able to remove them all and
747 	 * delete and free the empty node afterwards.
748 	 */
749 	if (WARN_ON_ONCE(!node->nr_values))
750 		goto out_invalid;
751 	if (WARN_ON_ONCE(node->count != node->nr_values))
752 		goto out_invalid;
753 	xa_delete_node(node, workingset_update_node);
754 	__inc_lruvec_kmem_state(node, WORKINGSET_NODERECLAIM);
755 
756 out_invalid:
757 	xa_unlock_irq(&mapping->i_pages);
758 	if (mapping->host != NULL) {
759 		if (mapping_shrinkable(mapping))
760 			inode_add_lru(mapping->host);
761 		spin_unlock(&mapping->host->i_lock);
762 	}
763 	ret = LRU_REMOVED_RETRY;
764 out:
765 	cond_resched();
766 	return ret;
767 }
768 
scan_shadow_nodes(struct shrinker * shrinker,struct shrink_control * sc)769 static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
770 				       struct shrink_control *sc)
771 {
772 	/* list_lru lock nests inside the IRQ-safe i_pages lock */
773 	return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate,
774 					NULL);
775 }
776 
777 /*
778  * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
779  * i_pages lock.
780  */
781 static struct lock_class_key shadow_nodes_key;
782 
workingset_init(void)783 static int __init workingset_init(void)
784 {
785 	struct shrinker *workingset_shadow_shrinker;
786 	unsigned int timestamp_bits;
787 	unsigned int max_order;
788 	int ret = -ENOMEM;
789 
790 	BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
791 	/*
792 	 * Calculate the eviction bucket size to cover the longest
793 	 * actionable refault distance, which is currently half of
794 	 * memory (totalram_pages/2). However, memory hotplug may add
795 	 * some more pages at runtime, so keep working with up to
796 	 * double the initial memory by using totalram_pages as-is.
797 	 */
798 	timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
799 	max_order = fls_long(totalram_pages() - 1);
800 	if (max_order > timestamp_bits)
801 		bucket_order = max_order - timestamp_bits;
802 	pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
803 	       timestamp_bits, max_order, bucket_order);
804 
805 	workingset_shadow_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE |
806 						    SHRINKER_MEMCG_AWARE,
807 						    "mm-shadow");
808 	if (!workingset_shadow_shrinker)
809 		goto err;
810 
811 	ret = list_lru_init_memcg_key(&shadow_nodes, workingset_shadow_shrinker,
812 				      &shadow_nodes_key);
813 	if (ret)
814 		goto err_list_lru;
815 
816 	workingset_shadow_shrinker->count_objects = count_shadow_nodes;
817 	workingset_shadow_shrinker->scan_objects = scan_shadow_nodes;
818 	/* ->count reports only fully expendable nodes */
819 	workingset_shadow_shrinker->seeks = 0;
820 
821 	shrinker_register(workingset_shadow_shrinker);
822 	return 0;
823 err_list_lru:
824 	shrinker_free(workingset_shadow_shrinker);
825 err:
826 	return ret;
827 }
828 module_init(workingset_init);
829