]> git.ipfire.org Git - thirdparty/linux.git/blame - mm/workingset.c
mm/hugetlb: convert get_hwpoison_huge_page() to folios
[thirdparty/linux.git] / mm / workingset.c
CommitLineData
b2441318 1// SPDX-License-Identifier: GPL-2.0
a528910e
JW
2/*
3 * Workingset detection
4 *
5 * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner
6 */
7
8#include <linux/memcontrol.h>
170b04b7 9#include <linux/mm_inline.h>
a528910e 10#include <linux/writeback.h>
3a4f8a0b 11#include <linux/shmem_fs.h>
a528910e
JW
12#include <linux/pagemap.h>
13#include <linux/atomic.h>
14#include <linux/module.h>
15#include <linux/swap.h>
14b46879 16#include <linux/dax.h>
a528910e
JW
17#include <linux/fs.h>
18#include <linux/mm.h>
19
20/*
21 * Double CLOCK lists
22 *
1e6b1085 23 * Per node, two clock lists are maintained for file pages: the
a528910e
JW
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 *
1899ad18 125 * Refaulting inactive pages
a528910e
JW
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 *
1899ad18
JW
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 *
a528910e
JW
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 *
1899ad18
JW
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 *
a528910e
JW
157 *
158 * Implementation
159 *
31d8fcac
JW
160 * For each node's LRU lists, a counter for inactive evictions and
161 * activations is maintained (node->nonresident_age).
a528910e
JW
162 *
163 * On eviction, a snapshot of this counter (along with some bits to
a97e7904 164 * identify the node) is stored in the now empty page cache
a528910e
JW
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
3ebc57f4 171#define WORKINGSET_SHIFT 1
3159f943 172#define EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \
3ebc57f4
ML
173 WORKINGSET_SHIFT + NODES_SHIFT + \
174 MEM_CGROUP_ID_SHIFT)
689c94f0
JW
175#define EVICTION_MASK (~0UL >> EVICTION_SHIFT)
176
612e4493
JW
177/*
178 * Eviction timestamps need to be able to cover the full range of
a97e7904 179 * actionable refaults. However, bits are tight in the xarray
612e4493
JW
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
1899ad18
JW
187static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction,
188 bool workingset)
a528910e 189{
3159f943 190 eviction &= EVICTION_MASK;
23047a96 191 eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
1e6b1085 192 eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
3ebc57f4 193 eviction = (eviction << WORKINGSET_SHIFT) | workingset;
a528910e 194
3159f943 195 return xa_mk_value(eviction);
a528910e
JW
196}
197
1e6b1085 198static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
1899ad18 199 unsigned long *evictionp, bool *workingsetp)
a528910e 200{
3159f943 201 unsigned long entry = xa_to_value(shadow);
1e6b1085 202 int memcgid, nid;
1899ad18 203 bool workingset;
a528910e 204
3ebc57f4
ML
205 workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1);
206 entry >>= WORKINGSET_SHIFT;
a528910e
JW
207 nid = entry & ((1UL << NODES_SHIFT) - 1);
208 entry >>= NODES_SHIFT;
23047a96
JW
209 memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
210 entry >>= MEM_CGROUP_ID_SHIFT;
a528910e 211
23047a96 212 *memcgidp = memcgid;
1e6b1085 213 *pgdat = NODE_DATA(nid);
ac35a490 214 *evictionp = entry;
1899ad18 215 *workingsetp = workingset;
a528910e
JW
216}
217
ac35a490
YZ
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;
391655fe 226 struct lru_gen_folio *lrugen;
ac35a490
YZ
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;
391655fe 255 struct lru_gen_folio *lrugen;
ac35a490
YZ
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
31d8fcac
JW
315/**
316 * workingset_age_nonresident - age non-resident entries as LRU ages
e755f4af 317 * @lruvec: the lruvec that was aged
31d8fcac
JW
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)
b910718a
JW
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 {
31d8fcac
JW
339 atomic_long_add(nr_pages, &lruvec->nonresident_age);
340 } while ((lruvec = parent_lruvec(lruvec)));
b910718a
JW
341}
342
a528910e 343/**
8927f647 344 * workingset_eviction - note the eviction of a folio from memory
b910718a 345 * @target_memcg: the cgroup that is causing the reclaim
8927f647 346 * @folio: the folio being evicted
a528910e 347 *
8927f647
MWO
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.
a528910e 350 */
8927f647 351void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg)
a528910e 352{
8927f647 353 struct pglist_data *pgdat = folio_pgdat(folio);
a528910e 354 unsigned long eviction;
23047a96 355 struct lruvec *lruvec;
b910718a 356 int memcgid;
a528910e 357
8927f647
MWO
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);
23047a96 362
ac35a490
YZ
363 if (lru_gen_enabled())
364 return lru_gen_eviction(folio);
365
b910718a
JW
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));
31d8fcac 369 eviction = atomic_long_read(&lruvec->nonresident_age);
ac35a490 370 eviction >>= bucket_order;
8927f647
MWO
371 workingset_age_nonresident(lruvec, folio_nr_pages(folio));
372 return pack_shadow(memcgid, pgdat, eviction,
373 folio_test_workingset(folio));
a528910e
JW
374}
375
376/**
0995d7e5
MWO
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.
a528910e
JW
380 *
381 * Calculates and evaluates the refault distance of the previously
0995d7e5 382 * evicted folio in the context of the node and the memcg whose memory
b910718a 383 * pressure caused the eviction.
a528910e 384 */
0995d7e5 385void workingset_refault(struct folio *folio, void *shadow)
a528910e 386{
0995d7e5 387 bool file = folio_is_file_lru(folio);
b910718a
JW
388 struct mem_cgroup *eviction_memcg;
389 struct lruvec *eviction_lruvec;
a528910e 390 unsigned long refault_distance;
34e58cac 391 unsigned long workingset_size;
1899ad18 392 struct pglist_data *pgdat;
23047a96 393 struct mem_cgroup *memcg;
162453bf 394 unsigned long eviction;
23047a96 395 struct lruvec *lruvec;
162453bf 396 unsigned long refault;
1899ad18 397 bool workingset;
23047a96 398 int memcgid;
0995d7e5 399 long nr;
a528910e 400
ac35a490
YZ
401 if (lru_gen_enabled()) {
402 lru_gen_refault(folio, shadow);
403 return;
404 }
405
1899ad18 406 unpack_shadow(shadow, &memcgid, &pgdat, &eviction, &workingset);
ac35a490 407 eviction <<= bucket_order;
162453bf 408
23047a96
JW
409 rcu_read_lock();
410 /*
411 * Look up the memcg associated with the stored ID. It might
0995d7e5 412 * have been deleted since the folio's eviction.
23047a96
JW
413 *
414 * Note that in rare events the ID could have been recycled
0995d7e5 415 * for a new cgroup that refaults a shared folio. This is
23047a96
JW
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 */
b910718a
JW
426 eviction_memcg = mem_cgroup_from_id(memcgid);
427 if (!mem_cgroup_disabled() && !eviction_memcg)
1899ad18 428 goto out;
b910718a 429 eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
31d8fcac 430 refault = atomic_long_read(&eviction_lruvec->nonresident_age);
162453bf
JW
431
432 /*
1899ad18 433 * Calculate the refault distance
162453bf 434 *
1899ad18 435 * The unsigned subtraction here gives an accurate distance
31d8fcac 436 * across nonresident_age overflows in most cases. There is a
1899ad18
JW
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
31d8fcac
JW
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
1899ad18
JW
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.
162453bf
JW
447 */
448 refault_distance = (refault - eviction) & EVICTION_MASK;
449
b910718a 450 /*
0995d7e5 451 * The activation decision for this folio is made at the level
b910718a 452 * where the eviction occurred, as that is where the LRU order
0995d7e5 453 * during folio reclaim is being determined.
b910718a 454 *
0995d7e5 455 * However, the cgroup that will own the folio is the one that
b910718a
JW
456 * is actually experiencing the refault event.
457 */
0995d7e5
MWO
458 nr = folio_nr_pages(folio);
459 memcg = folio_memcg(folio);
f78dfc7b 460 pgdat = folio_pgdat(folio);
b910718a
JW
461 lruvec = mem_cgroup_lruvec(memcg, pgdat);
462
0995d7e5 463 mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr);
a528910e 464
9b301615 465 mem_cgroup_flush_stats_delayed();
1899ad18
JW
466 /*
467 * Compare the distance to the existing workingset size. We
34e58cac 468 * don't activate pages that couldn't stay resident even if
aae466b0
JK
469 * all the memory was available to the workingset. Whether
470 * workingset competition needs to consider anon or not depends
471 * on having swap.
1899ad18 472 */
34e58cac 473 workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE);
aae466b0 474 if (!file) {
34e58cac 475 workingset_size += lruvec_page_state(eviction_lruvec,
aae466b0
JK
476 NR_INACTIVE_FILE);
477 }
f78dfc7b 478 if (mem_cgroup_get_nr_swap_pages(eviction_memcg) > 0) {
34e58cac
JW
479 workingset_size += lruvec_page_state(eviction_lruvec,
480 NR_ACTIVE_ANON);
aae466b0
JK
481 if (file) {
482 workingset_size += lruvec_page_state(eviction_lruvec,
483 NR_INACTIVE_ANON);
484 }
34e58cac
JW
485 }
486 if (refault_distance > workingset_size)
1899ad18
JW
487 goto out;
488
0995d7e5
MWO
489 folio_set_active(folio);
490 workingset_age_nonresident(lruvec, nr);
491 mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr);
1899ad18 492
0995d7e5 493 /* Folio was active prior to eviction */
1899ad18 494 if (workingset) {
0995d7e5 495 folio_set_workingset(folio);
6e1ca48d
VMO
496 /*
497 * XXX: Move to folio_add_lru() when it supports new vs
498 * putback
499 */
0538a82c 500 lru_note_cost_refault(folio);
0995d7e5 501 mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr);
a528910e 502 }
1899ad18 503out:
2a2e4885 504 rcu_read_unlock();
a528910e
JW
505}
506
507/**
508 * workingset_activation - note a page activation
c5ce619a 509 * @folio: Folio that is being activated.
a528910e 510 */
c5ce619a 511void workingset_activation(struct folio *folio)
a528910e 512{
55779ec7 513 struct mem_cgroup *memcg;
23047a96 514
55779ec7 515 rcu_read_lock();
23047a96
JW
516 /*
517 * Filter non-memcg pages here, e.g. unmap can call
518 * mark_page_accessed() on VDSO pages.
519 *
520 * XXX: See workingset_refault() - this should return
521 * root_mem_cgroup even for !CONFIG_MEMCG.
522 */
c5ce619a 523 memcg = folio_memcg_rcu(folio);
55779ec7 524 if (!mem_cgroup_disabled() && !memcg)
23047a96 525 goto out;
c5ce619a 526 workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio));
23047a96 527out:
55779ec7 528 rcu_read_unlock();
a528910e 529}
449dd698
JW
530
531/*
532 * Shadow entries reflect the share of the working set that does not
533 * fit into memory, so their number depends on the access pattern of
534 * the workload. In most cases, they will refault or get reclaimed
535 * along with the inode, but a (malicious) workload that streams
536 * through files with a total size several times that of available
537 * memory, while preventing the inodes from being reclaimed, can
538 * create excessive amounts of shadow nodes. To keep a lid on this,
539 * track shadow nodes and reclaim them when they grow way past the
540 * point where they would still be useful.
541 */
542
9bbdc0f3 543struct list_lru shadow_nodes;
14b46879 544
a97e7904 545void workingset_update_node(struct xa_node *node)
14b46879 546{
2386eef2
SAS
547 struct address_space *mapping;
548
14b46879
JW
549 /*
550 * Track non-empty nodes that contain only shadow entries;
551 * unlink those that contain pages or are being freed.
552 *
553 * Avoid acquiring the list_lru lock when the nodes are
554 * already where they should be. The list_empty() test is safe
b93b0163 555 * as node->private_list is protected by the i_pages lock.
14b46879 556 */
2386eef2
SAS
557 mapping = container_of(node->array, struct address_space, i_pages);
558 lockdep_assert_held(&mapping->i_pages.xa_lock);
68d48e6a 559
01959dfe 560 if (node->count && node->count == node->nr_values) {
68d48e6a 561 if (list_empty(&node->private_list)) {
14b46879 562 list_lru_add(&shadow_nodes, &node->private_list);
da3ceeff 563 __inc_lruvec_kmem_state(node, WORKINGSET_NODES);
68d48e6a 564 }
14b46879 565 } else {
68d48e6a 566 if (!list_empty(&node->private_list)) {
14b46879 567 list_lru_del(&shadow_nodes, &node->private_list);
da3ceeff 568 __dec_lruvec_kmem_state(node, WORKINGSET_NODES);
68d48e6a 569 }
14b46879
JW
570 }
571}
449dd698
JW
572
573static unsigned long count_shadow_nodes(struct shrinker *shrinker,
574 struct shrink_control *sc)
575{
449dd698 576 unsigned long max_nodes;
14b46879 577 unsigned long nodes;
95f9ab2d 578 unsigned long pages;
449dd698 579
14b46879 580 nodes = list_lru_shrink_count(&shadow_nodes, sc);
725cac1c
ML
581 if (!nodes)
582 return SHRINK_EMPTY;
449dd698 583
449dd698 584 /*
a97e7904 585 * Approximate a reasonable limit for the nodes
b5388998
JW
586 * containing shadow entries. We don't need to keep more
587 * shadow entries than possible pages on the active list,
588 * since refault distances bigger than that are dismissed.
589 *
590 * The size of the active list converges toward 100% of
591 * overall page cache as memory grows, with only a tiny
592 * inactive list. Assume the total cache size for that.
593 *
594 * Nodes might be sparsely populated, with only one shadow
595 * entry in the extreme case. Obviously, we cannot keep one
596 * node for every eligible shadow entry, so compromise on a
597 * worst-case density of 1/8th. Below that, not all eligible
598 * refaults can be detected anymore.
449dd698 599 *
a97e7904 600 * On 64-bit with 7 xa_nodes per page and 64 slots
449dd698 601 * each, this will reclaim shadow entries when they consume
b5388998 602 * ~1.8% of available memory:
449dd698 603 *
a97e7904 604 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
449dd698 605 */
95f9ab2d 606#ifdef CONFIG_MEMCG
b5388998 607 if (sc->memcg) {
95f9ab2d 608 struct lruvec *lruvec;
2b487e59 609 int i;
95f9ab2d 610
867e5e1d 611 lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
2b487e59 612 for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
205b20cc
JW
613 pages += lruvec_page_state_local(lruvec,
614 NR_LRU_BASE + i);
d42f3245
RG
615 pages += lruvec_page_state_local(
616 lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT;
617 pages += lruvec_page_state_local(
618 lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT;
95f9ab2d
JW
619 } else
620#endif
621 pages = node_present_pages(sc->nid);
622
dad4f140 623 max_nodes = pages >> (XA_CHUNK_SHIFT - 3);
449dd698 624
14b46879 625 if (nodes <= max_nodes)
449dd698 626 return 0;
14b46879 627 return nodes - max_nodes;
449dd698
JW
628}
629
630static enum lru_status shadow_lru_isolate(struct list_head *item,
3f97b163 631 struct list_lru_one *lru,
449dd698 632 spinlock_t *lru_lock,
a97e7904 633 void *arg) __must_hold(lru_lock)
449dd698 634{
a97e7904 635 struct xa_node *node = container_of(item, struct xa_node, private_list);
449dd698 636 struct address_space *mapping;
449dd698
JW
637 int ret;
638
639 /*
f82cd2f0 640 * Page cache insertions and deletions synchronously maintain
b93b0163 641 * the shadow node LRU under the i_pages lock and the
449dd698
JW
642 * lru_lock. Because the page cache tree is emptied before
643 * the inode can be destroyed, holding the lru_lock pins any
a97e7904 644 * address_space that has nodes on the LRU.
449dd698 645 *
b93b0163 646 * We can then safely transition to the i_pages lock to
449dd698
JW
647 * pin only the address_space of the particular node we want
648 * to reclaim, take the node off-LRU, and drop the lru_lock.
649 */
650
01959dfe 651 mapping = container_of(node->array, struct address_space, i_pages);
449dd698
JW
652
653 /* Coming from the list, invert the lock order */
b93b0163 654 if (!xa_trylock(&mapping->i_pages)) {
6ca342d0 655 spin_unlock_irq(lru_lock);
449dd698
JW
656 ret = LRU_RETRY;
657 goto out;
658 }
659
51b8c1fe
JW
660 if (!spin_trylock(&mapping->host->i_lock)) {
661 xa_unlock(&mapping->i_pages);
662 spin_unlock_irq(lru_lock);
663 ret = LRU_RETRY;
664 goto out;
665 }
666
3f97b163 667 list_lru_isolate(lru, item);
da3ceeff 668 __dec_lruvec_kmem_state(node, WORKINGSET_NODES);
68d48e6a 669
449dd698
JW
670 spin_unlock(lru_lock);
671
672 /*
673 * The nodes should only contain one or more shadow entries,
674 * no pages, so we expect to be able to remove them all and
675 * delete and free the empty node afterwards.
676 */
01959dfe 677 if (WARN_ON_ONCE(!node->nr_values))
b936887e 678 goto out_invalid;
01959dfe 679 if (WARN_ON_ONCE(node->count != node->nr_values))
b936887e 680 goto out_invalid;
f82cd2f0 681 xa_delete_node(node, workingset_update_node);
da3ceeff 682 __inc_lruvec_kmem_state(node, WORKINGSET_NODERECLAIM);
449dd698 683
b936887e 684out_invalid:
6ca342d0 685 xa_unlock_irq(&mapping->i_pages);
51b8c1fe
JW
686 if (mapping_shrinkable(mapping))
687 inode_add_lru(mapping->host);
688 spin_unlock(&mapping->host->i_lock);
449dd698
JW
689 ret = LRU_REMOVED_RETRY;
690out:
449dd698 691 cond_resched();
6ca342d0 692 spin_lock_irq(lru_lock);
449dd698
JW
693 return ret;
694}
695
696static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
697 struct shrink_control *sc)
698{
b93b0163 699 /* list_lru lock nests inside the IRQ-safe i_pages lock */
6b51e881
SAS
700 return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate,
701 NULL);
449dd698
JW
702}
703
704static struct shrinker workingset_shadow_shrinker = {
705 .count_objects = count_shadow_nodes,
706 .scan_objects = scan_shadow_nodes,
4b85afbd 707 .seeks = 0, /* ->count reports only fully expendable nodes */
0a6b76dd 708 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
449dd698
JW
709};
710
711/*
712 * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
b93b0163 713 * i_pages lock.
449dd698
JW
714 */
715static struct lock_class_key shadow_nodes_key;
716
717static int __init workingset_init(void)
718{
612e4493
JW
719 unsigned int timestamp_bits;
720 unsigned int max_order;
449dd698
JW
721 int ret;
722
612e4493
JW
723 BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
724 /*
725 * Calculate the eviction bucket size to cover the longest
726 * actionable refault distance, which is currently half of
727 * memory (totalram_pages/2). However, memory hotplug may add
728 * some more pages at runtime, so keep working with up to
729 * double the initial memory by using totalram_pages as-is.
730 */
731 timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
ca79b0c2 732 max_order = fls_long(totalram_pages() - 1);
612e4493
JW
733 if (max_order > timestamp_bits)
734 bucket_order = max_order - timestamp_bits;
d3d36c4b 735 pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
612e4493
JW
736 timestamp_bits, max_order, bucket_order);
737
e33c267a 738 ret = prealloc_shrinker(&workingset_shadow_shrinker, "mm-shadow");
449dd698
JW
739 if (ret)
740 goto err;
c92e8e10
KT
741 ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key,
742 &workingset_shadow_shrinker);
449dd698
JW
743 if (ret)
744 goto err_list_lru;
39887653 745 register_shrinker_prepared(&workingset_shadow_shrinker);
449dd698
JW
746 return 0;
747err_list_lru:
39887653 748 free_prealloced_shrinker(&workingset_shadow_shrinker);
449dd698
JW
749err:
750 return ret;
751}
752module_init(workingset_init);