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 * If we have swap we should consider about NR_inactive_anon and
115 * NR_active_anon, so for page cache and anonymous respectively:
116 *
117 * NR_inactive_file + (R - E) <= NR_inactive_file + NR_active_file
118 * + NR_inactive_anon + NR_active_anon
119 *
120 * NR_inactive_anon + (R - E) <= NR_inactive_anon + NR_active_anon
121 * + NR_inactive_file + NR_active_file
122 *
123 * Which can be further simplified to:
124 *
125 * (R - E) <= NR_active_file + NR_inactive_anon + NR_active_anon
126 *
127 * (R - E) <= NR_active_anon + NR_inactive_file + NR_active_file
128 *
129 * Put into words, the refault distance (out-of-cache) can be seen as
130 * a deficit in inactive list space (in-cache). If the inactive list
131 * had (R - E) more page slots, the page would not have been evicted
132 * in between accesses, but activated instead. And on a full system,
133 * the only thing eating into inactive list space is active pages.
134 *
135 *
136 * Refaulting inactive pages
137 *
138 * All that is known about the active list is that the pages have been
139 * accessed more than once in the past. This means that at any given
140 * time there is actually a good chance that pages on the active list
141 * are no longer in active use.
142 *
143 * So when a refault distance of (R - E) is observed and there are at
144 * least (R - E) pages in the userspace workingset, the refaulting page
145 * is activated optimistically in the hope that (R - E) pages are actually
146 * used less frequently than the refaulting page - or even not used at
147 * all anymore.
148 *
149 * That means if inactive cache is refaulting with a suitable refault
150 * distance, we assume the cache workingset is transitioning and put
151 * pressure on the current workingset.
152 *
153 * If this is wrong and demotion kicks in, the pages which are truly
154 * used more frequently will be reactivated while the less frequently
155 * used once will be evicted from memory.
156 *
157 * But if this is right, the stale pages will be pushed out of memory
158 * and the used pages get to stay in cache.
159 *
160 * Refaulting active pages
161 *
162 * If on the other hand the refaulting pages have recently been
163 * deactivated, it means that the active list is no longer protecting
164 * actively used cache from reclaim. The cache is NOT transitioning to
165 * a different workingset; the existing workingset is thrashing in the
166 * space allocated to the page cache.
167 *
168 *
169 * Implementation
170 *
171 * For each node's LRU lists, a counter for inactive evictions and
172 * activations is maintained (node->nonresident_age).
173 *
174 * On eviction, a snapshot of this counter (along with some bits to
175 * identify the node) is stored in the now empty page cache
176 * slot of the evicted page. This is called a shadow entry.
177 *
178 * On cache misses for which there are shadow entries, an eligible
179 * refault distance will immediately activate the refaulting page.
180 */
181
182#define WORKINGSET_SHIFT 1
183#define EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \
184 WORKINGSET_SHIFT + NODES_SHIFT + \
185 MEM_CGROUP_ID_SHIFT)
186#define EVICTION_MASK (~0UL >> EVICTION_SHIFT)
187
188/*
189 * Eviction timestamps need to be able to cover the full range of
190 * actionable refaults. However, bits are tight in the xarray
191 * entry, and after storing the identifier for the lruvec there might
192 * not be enough left to represent every single actionable refault. In
193 * that case, we have to sacrifice granularity for distance, and group
194 * evictions into coarser buckets by shaving off lower timestamp bits.
195 */
196static unsigned int bucket_order __read_mostly;
197
198static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction,
199 bool workingset)
200{
201 eviction &= EVICTION_MASK;
202 eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
203 eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
204 eviction = (eviction << WORKINGSET_SHIFT) | workingset;
205
206 return xa_mk_value(eviction);
207}
208
209static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
210 unsigned long *evictionp, bool *workingsetp)
211{
212 unsigned long entry = xa_to_value(shadow);
213 int memcgid, nid;
214 bool workingset;
215
216 workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1);
217 entry >>= WORKINGSET_SHIFT;
218 nid = entry & ((1UL << NODES_SHIFT) - 1);
219 entry >>= NODES_SHIFT;
220 memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
221 entry >>= MEM_CGROUP_ID_SHIFT;
222
223 *memcgidp = memcgid;
224 *pgdat = NODE_DATA(nid);
225 *evictionp = entry;
226 *workingsetp = workingset;
227}
228
229#ifdef CONFIG_LRU_GEN
230
231static void *lru_gen_eviction(struct folio *folio)
232{
233 int hist;
234 unsigned long token;
235 unsigned long min_seq;
236 struct lruvec *lruvec;
237 struct lru_gen_folio *lrugen;
238 int type = folio_is_file_lru(folio);
239 int delta = folio_nr_pages(folio);
240 int refs = folio_lru_refs(folio);
241 int tier = lru_tier_from_refs(refs);
242 struct mem_cgroup *memcg = folio_memcg(folio);
243 struct pglist_data *pgdat = folio_pgdat(folio);
244
245 BUILD_BUG_ON(LRU_GEN_WIDTH + LRU_REFS_WIDTH > BITS_PER_LONG - EVICTION_SHIFT);
246
247 lruvec = mem_cgroup_lruvec(memcg, pgdat);
248 lrugen = &lruvec->lrugen;
249 min_seq = READ_ONCE(lrugen->min_seq[type]);
250 token = (min_seq << LRU_REFS_WIDTH) | max(refs - 1, 0);
251
252 hist = lru_hist_from_seq(min_seq);
253 atomic_long_add(delta, &lrugen->evicted[hist][type][tier]);
254
255 return pack_shadow(mem_cgroup_id(memcg), pgdat, token, refs);
256}
257
258/*
259 * Tests if the shadow entry is for a folio that was recently evicted.
260 * Fills in @lruvec, @token, @workingset with the values unpacked from shadow.
261 */
262static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec,
263 unsigned long *token, bool *workingset)
264{
265 int memcg_id;
266 unsigned long min_seq;
267 struct mem_cgroup *memcg;
268 struct pglist_data *pgdat;
269
270 unpack_shadow(shadow, &memcg_id, &pgdat, token, workingset);
271
272 memcg = mem_cgroup_from_id(memcg_id);
273 *lruvec = mem_cgroup_lruvec(memcg, pgdat);
274
275 min_seq = READ_ONCE((*lruvec)->lrugen.min_seq[file]);
276 return (*token >> LRU_REFS_WIDTH) == (min_seq & (EVICTION_MASK >> LRU_REFS_WIDTH));
277}
278
279static void lru_gen_refault(struct folio *folio, void *shadow)
280{
281 bool recent;
282 int hist, tier, refs;
283 bool workingset;
284 unsigned long token;
285 struct lruvec *lruvec;
286 struct lru_gen_folio *lrugen;
287 int type = folio_is_file_lru(folio);
288 int delta = folio_nr_pages(folio);
289
290 rcu_read_lock();
291
292 recent = lru_gen_test_recent(shadow, type, &lruvec, &token, &workingset);
293 if (lruvec != folio_lruvec(folio))
294 goto unlock;
295
296 mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + type, delta);
297
298 if (!recent)
299 goto unlock;
300
301 lrugen = &lruvec->lrugen;
302
303 hist = lru_hist_from_seq(READ_ONCE(lrugen->min_seq[type]));
304 /* see the comment in folio_lru_refs() */
305 refs = (token & (BIT(LRU_REFS_WIDTH) - 1)) + workingset;
306 tier = lru_tier_from_refs(refs);
307
308 atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]);
309 mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
310
311 /*
312 * Count the following two cases as stalls:
313 * 1. For pages accessed through page tables, hotter pages pushed out
314 * hot pages which refaulted immediately.
315 * 2. For pages accessed multiple times through file descriptors,
316 * they would have been protected by sort_folio().
317 */
318 if (lru_gen_in_fault() || refs >= BIT(LRU_REFS_WIDTH) - 1) {
319 set_mask_bits(&folio->flags, 0, LRU_REFS_MASK | BIT(PG_workingset));
320 mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + type, delta);
321 }
322unlock:
323 rcu_read_unlock();
324}
325
326#else /* !CONFIG_LRU_GEN */
327
328static void *lru_gen_eviction(struct folio *folio)
329{
330 return NULL;
331}
332
333static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec,
334 unsigned long *token, bool *workingset)
335{
336 return false;
337}
338
339static 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 */
355void 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 */
381void *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 *
415 * Return: true if the shadow is for a recently evicted folio; false otherwise.
416 */
417bool workingset_test_recent(void *shadow, bool file, bool *workingset)
418{
419 struct mem_cgroup *eviction_memcg;
420 struct lruvec *eviction_lruvec;
421 unsigned long refault_distance;
422 unsigned long workingset_size;
423 unsigned long refault;
424 int memcgid;
425 struct pglist_data *pgdat;
426 unsigned long eviction;
427
428 rcu_read_lock();
429
430 if (lru_gen_enabled()) {
431 bool recent = lru_gen_test_recent(shadow, file,
432 &eviction_lruvec, &eviction, workingset);
433
434 rcu_read_unlock();
435 return recent;
436 }
437
438
439 unpack_shadow(shadow, &memcgid, &pgdat, &eviction, workingset);
440 eviction <<= bucket_order;
441
442 /*
443 * Look up the memcg associated with the stored ID. It might
444 * have been deleted since the folio's eviction.
445 *
446 * Note that in rare events the ID could have been recycled
447 * for a new cgroup that refaults a shared folio. This is
448 * impossible to tell from the available data. However, this
449 * should be a rare and limited disturbance, and activations
450 * are always speculative anyway. Ultimately, it's the aging
451 * algorithm's job to shake out the minimum access frequency
452 * for the active cache.
453 *
454 * XXX: On !CONFIG_MEMCG, this will always return NULL; it
455 * would be better if the root_mem_cgroup existed in all
456 * configurations instead.
457 */
458 eviction_memcg = mem_cgroup_from_id(memcgid);
459 if (!mem_cgroup_disabled() &&
460 (!eviction_memcg || !mem_cgroup_tryget(eviction_memcg))) {
461 rcu_read_unlock();
462 return false;
463 }
464
465 rcu_read_unlock();
466
467 /*
468 * Flush stats (and potentially sleep) outside the RCU read section.
469 * XXX: With per-memcg flushing and thresholding, is ratelimiting
470 * still needed here?
471 */
472 mem_cgroup_flush_stats_ratelimited(eviction_memcg);
473
474 eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
475 refault = atomic_long_read(&eviction_lruvec->nonresident_age);
476
477 /*
478 * Calculate the refault distance
479 *
480 * The unsigned subtraction here gives an accurate distance
481 * across nonresident_age overflows in most cases. There is a
482 * special case: usually, shadow entries have a short lifetime
483 * and are either refaulted or reclaimed along with the inode
484 * before they get too old. But it is not impossible for the
485 * nonresident_age to lap a shadow entry in the field, which
486 * can then result in a false small refault distance, leading
487 * to a false activation should this old entry actually
488 * refault again. However, earlier kernels used to deactivate
489 * unconditionally with *every* reclaim invocation for the
490 * longest time, so the occasional inappropriate activation
491 * leading to pressure on the active list is not a problem.
492 */
493 refault_distance = (refault - eviction) & EVICTION_MASK;
494
495 /*
496 * Compare the distance to the existing workingset size. We
497 * don't activate pages that couldn't stay resident even if
498 * all the memory was available to the workingset. Whether
499 * workingset competition needs to consider anon or not depends
500 * on having free swap space.
501 */
502 workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE);
503 if (!file) {
504 workingset_size += lruvec_page_state(eviction_lruvec,
505 NR_INACTIVE_FILE);
506 }
507 if (mem_cgroup_get_nr_swap_pages(eviction_memcg) > 0) {
508 workingset_size += lruvec_page_state(eviction_lruvec,
509 NR_ACTIVE_ANON);
510 if (file) {
511 workingset_size += lruvec_page_state(eviction_lruvec,
512 NR_INACTIVE_ANON);
513 }
514 }
515
516 mem_cgroup_put(eviction_memcg);
517 return refault_distance <= workingset_size;
518}
519
520/**
521 * workingset_refault - Evaluate the refault of a previously evicted folio.
522 * @folio: The freshly allocated replacement folio.
523 * @shadow: Shadow entry of the evicted folio.
524 *
525 * Calculates and evaluates the refault distance of the previously
526 * evicted folio in the context of the node and the memcg whose memory
527 * pressure caused the eviction.
528 */
529void workingset_refault(struct folio *folio, void *shadow)
530{
531 bool file = folio_is_file_lru(folio);
532 struct pglist_data *pgdat;
533 struct mem_cgroup *memcg;
534 struct lruvec *lruvec;
535 bool workingset;
536 long nr;
537
538 if (lru_gen_enabled()) {
539 lru_gen_refault(folio, shadow);
540 return;
541 }
542
543 /*
544 * The activation decision for this folio is made at the level
545 * where the eviction occurred, as that is where the LRU order
546 * during folio reclaim is being determined.
547 *
548 * However, the cgroup that will own the folio is the one that
549 * is actually experiencing the refault event. Make sure the folio is
550 * locked to guarantee folio_memcg() stability throughout.
551 */
552 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
553 nr = folio_nr_pages(folio);
554 memcg = folio_memcg(folio);
555 pgdat = folio_pgdat(folio);
556 lruvec = mem_cgroup_lruvec(memcg, pgdat);
557
558 mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr);
559
560 if (!workingset_test_recent(shadow, file, &workingset))
561 return;
562
563 folio_set_active(folio);
564 workingset_age_nonresident(lruvec, nr);
565 mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr);
566
567 /* Folio was active prior to eviction */
568 if (workingset) {
569 folio_set_workingset(folio);
570 /*
571 * XXX: Move to folio_add_lru() when it supports new vs
572 * putback
573 */
574 lru_note_cost_refault(folio);
575 mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr);
576 }
577}
578
579/**
580 * workingset_activation - note a page activation
581 * @folio: Folio that is being activated.
582 */
583void workingset_activation(struct folio *folio)
584{
585 struct mem_cgroup *memcg;
586
587 rcu_read_lock();
588 /*
589 * Filter non-memcg pages here, e.g. unmap can call
590 * mark_page_accessed() on VDSO pages.
591 *
592 * XXX: See workingset_refault() - this should return
593 * root_mem_cgroup even for !CONFIG_MEMCG.
594 */
595 memcg = folio_memcg_rcu(folio);
596 if (!mem_cgroup_disabled() && !memcg)
597 goto out;
598 workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio));
599out:
600 rcu_read_unlock();
601}
602
603/*
604 * Shadow entries reflect the share of the working set that does not
605 * fit into memory, so their number depends on the access pattern of
606 * the workload. In most cases, they will refault or get reclaimed
607 * along with the inode, but a (malicious) workload that streams
608 * through files with a total size several times that of available
609 * memory, while preventing the inodes from being reclaimed, can
610 * create excessive amounts of shadow nodes. To keep a lid on this,
611 * track shadow nodes and reclaim them when they grow way past the
612 * point where they would still be useful.
613 */
614
615struct list_lru shadow_nodes;
616
617void workingset_update_node(struct xa_node *node)
618{
619 struct address_space *mapping;
620
621 /*
622 * Track non-empty nodes that contain only shadow entries;
623 * unlink those that contain pages or are being freed.
624 *
625 * Avoid acquiring the list_lru lock when the nodes are
626 * already where they should be. The list_empty() test is safe
627 * as node->private_list is protected by the i_pages lock.
628 */
629 mapping = container_of(node->array, struct address_space, i_pages);
630 lockdep_assert_held(&mapping->i_pages.xa_lock);
631
632 if (node->count && node->count == node->nr_values) {
633 if (list_empty(&node->private_list)) {
634 list_lru_add_obj(&shadow_nodes, &node->private_list);
635 __inc_lruvec_kmem_state(node, WORKINGSET_NODES);
636 }
637 } else {
638 if (!list_empty(&node->private_list)) {
639 list_lru_del_obj(&shadow_nodes, &node->private_list);
640 __dec_lruvec_kmem_state(node, WORKINGSET_NODES);
641 }
642 }
643}
644
645static unsigned long count_shadow_nodes(struct shrinker *shrinker,
646 struct shrink_control *sc)
647{
648 unsigned long max_nodes;
649 unsigned long nodes;
650 unsigned long pages;
651
652 nodes = list_lru_shrink_count(&shadow_nodes, sc);
653 if (!nodes)
654 return SHRINK_EMPTY;
655
656 /*
657 * Approximate a reasonable limit for the nodes
658 * containing shadow entries. We don't need to keep more
659 * shadow entries than possible pages on the active list,
660 * since refault distances bigger than that are dismissed.
661 *
662 * The size of the active list converges toward 100% of
663 * overall page cache as memory grows, with only a tiny
664 * inactive list. Assume the total cache size for that.
665 *
666 * Nodes might be sparsely populated, with only one shadow
667 * entry in the extreme case. Obviously, we cannot keep one
668 * node for every eligible shadow entry, so compromise on a
669 * worst-case density of 1/8th. Below that, not all eligible
670 * refaults can be detected anymore.
671 *
672 * On 64-bit with 7 xa_nodes per page and 64 slots
673 * each, this will reclaim shadow entries when they consume
674 * ~1.8% of available memory:
675 *
676 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
677 */
678#ifdef CONFIG_MEMCG
679 if (sc->memcg) {
680 struct lruvec *lruvec;
681 int i;
682
683 mem_cgroup_flush_stats_ratelimited(sc->memcg);
684 lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
685 for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
686 pages += lruvec_page_state_local(lruvec,
687 NR_LRU_BASE + i);
688 pages += lruvec_page_state_local(
689 lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT;
690 pages += lruvec_page_state_local(
691 lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT;
692 } else
693#endif
694 pages = node_present_pages(sc->nid);
695
696 max_nodes = pages >> (XA_CHUNK_SHIFT - 3);
697
698 if (nodes <= max_nodes)
699 return 0;
700 return nodes - max_nodes;
701}
702
703static enum lru_status shadow_lru_isolate(struct list_head *item,
704 struct list_lru_one *lru,
705 spinlock_t *lru_lock,
706 void *arg) __must_hold(lru_lock)
707{
708 struct xa_node *node = container_of(item, struct xa_node, private_list);
709 struct address_space *mapping;
710 int ret;
711
712 /*
713 * Page cache insertions and deletions synchronously maintain
714 * the shadow node LRU under the i_pages lock and the
715 * lru_lock. Because the page cache tree is emptied before
716 * the inode can be destroyed, holding the lru_lock pins any
717 * address_space that has nodes on the LRU.
718 *
719 * We can then safely transition to the i_pages lock to
720 * pin only the address_space of the particular node we want
721 * to reclaim, take the node off-LRU, and drop the lru_lock.
722 */
723
724 mapping = container_of(node->array, struct address_space, i_pages);
725
726 /* Coming from the list, invert the lock order */
727 if (!xa_trylock(&mapping->i_pages)) {
728 spin_unlock_irq(lru_lock);
729 ret = LRU_RETRY;
730 goto out;
731 }
732
733 /* For page cache we need to hold i_lock */
734 if (mapping->host != NULL) {
735 if (!spin_trylock(&mapping->host->i_lock)) {
736 xa_unlock(&mapping->i_pages);
737 spin_unlock_irq(lru_lock);
738 ret = LRU_RETRY;
739 goto out;
740 }
741 }
742
743 list_lru_isolate(lru, item);
744 __dec_lruvec_kmem_state(node, WORKINGSET_NODES);
745
746 spin_unlock(lru_lock);
747
748 /*
749 * The nodes should only contain one or more shadow entries,
750 * no pages, so we expect to be able to remove them all and
751 * delete and free the empty node afterwards.
752 */
753 if (WARN_ON_ONCE(!node->nr_values))
754 goto out_invalid;
755 if (WARN_ON_ONCE(node->count != node->nr_values))
756 goto out_invalid;
757 xa_delete_node(node, workingset_update_node);
758 __inc_lruvec_kmem_state(node, WORKINGSET_NODERECLAIM);
759
760out_invalid:
761 xa_unlock_irq(&mapping->i_pages);
762 if (mapping->host != NULL) {
763 if (mapping_shrinkable(mapping))
764 inode_add_lru(mapping->host);
765 spin_unlock(&mapping->host->i_lock);
766 }
767 ret = LRU_REMOVED_RETRY;
768out:
769 cond_resched();
770 spin_lock_irq(lru_lock);
771 return ret;
772}
773
774static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
775 struct shrink_control *sc)
776{
777 /* list_lru lock nests inside the IRQ-safe i_pages lock */
778 return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate,
779 NULL);
780}
781
782/*
783 * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
784 * i_pages lock.
785 */
786static struct lock_class_key shadow_nodes_key;
787
788static int __init workingset_init(void)
789{
790 struct shrinker *workingset_shadow_shrinker;
791 unsigned int timestamp_bits;
792 unsigned int max_order;
793 int ret = -ENOMEM;
794
795 BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
796 /*
797 * Calculate the eviction bucket size to cover the longest
798 * actionable refault distance, which is currently half of
799 * memory (totalram_pages/2). However, memory hotplug may add
800 * some more pages at runtime, so keep working with up to
801 * double the initial memory by using totalram_pages as-is.
802 */
803 timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
804 max_order = fls_long(totalram_pages() - 1);
805 if (max_order > timestamp_bits)
806 bucket_order = max_order - timestamp_bits;
807 pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
808 timestamp_bits, max_order, bucket_order);
809
810 workingset_shadow_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE |
811 SHRINKER_MEMCG_AWARE,
812 "mm-shadow");
813 if (!workingset_shadow_shrinker)
814 goto err;
815
816 ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key,
817 workingset_shadow_shrinker);
818 if (ret)
819 goto err_list_lru;
820
821 workingset_shadow_shrinker->count_objects = count_shadow_nodes;
822 workingset_shadow_shrinker->scan_objects = scan_shadow_nodes;
823 /* ->count reports only fully expendable nodes */
824 workingset_shadow_shrinker->seeks = 0;
825
826 shrinker_register(workingset_shadow_shrinker);
827 return 0;
828err_list_lru:
829 shrinker_free(workingset_shadow_shrinker);
830err:
831 return ret;
832}
833module_init(workingset_init);