]>
Commit | Line | Data |
---|---|---|
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> | |
b64e74e9 | 19 | #include "internal.h" |
a528910e JW |
20 | |
21 | /* | |
22 | * Double CLOCK lists | |
23 | * | |
1e6b1085 | 24 | * Per node, two clock lists are maintained for file pages: the |
a528910e JW |
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 | * | |
ed8f3f99 YY |
115 | * If we have swap we should consider about NR_inactive_anon and |
116 | * NR_active_anon, so for page cache and anonymous respectively: | |
a528910e | 117 | * |
ed8f3f99 YY |
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 | |
a528910e JW |
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 | * | |
1899ad18 | 137 | * Refaulting inactive pages |
a528910e JW |
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 | |
ed8f3f99 YY |
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 | |
a528910e JW |
147 | * used less frequently than the refaulting page - or even not used at |
148 | * all anymore. | |
149 | * | |
1899ad18 JW |
150 | * That means if inactive cache is refaulting with a suitable refault |
151 | * distance, we assume the cache workingset is transitioning and put | |
ed8f3f99 | 152 | * pressure on the current workingset. |
1899ad18 | 153 | * |
a528910e JW |
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 | * | |
1899ad18 JW |
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 | * | |
a528910e JW |
169 | * |
170 | * Implementation | |
171 | * | |
31d8fcac JW |
172 | * For each node's LRU lists, a counter for inactive evictions and |
173 | * activations is maintained (node->nonresident_age). | |
a528910e JW |
174 | * |
175 | * On eviction, a snapshot of this counter (along with some bits to | |
a97e7904 | 176 | * identify the node) is stored in the now empty page cache |
a528910e JW |
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 | ||
3ebc57f4 | 183 | #define WORKINGSET_SHIFT 1 |
3159f943 | 184 | #define EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \ |
3ebc57f4 ML |
185 | WORKINGSET_SHIFT + NODES_SHIFT + \ |
186 | MEM_CGROUP_ID_SHIFT) | |
689c94f0 JW |
187 | #define EVICTION_MASK (~0UL >> EVICTION_SHIFT) |
188 | ||
612e4493 JW |
189 | /* |
190 | * Eviction timestamps need to be able to cover the full range of | |
a97e7904 | 191 | * actionable refaults. However, bits are tight in the xarray |
612e4493 JW |
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 | ||
1899ad18 JW |
199 | static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction, |
200 | bool workingset) | |
a528910e | 201 | { |
3159f943 | 202 | eviction &= EVICTION_MASK; |
23047a96 | 203 | eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid; |
1e6b1085 | 204 | eviction = (eviction << NODES_SHIFT) | pgdat->node_id; |
3ebc57f4 | 205 | eviction = (eviction << WORKINGSET_SHIFT) | workingset; |
a528910e | 206 | |
3159f943 | 207 | return xa_mk_value(eviction); |
a528910e JW |
208 | } |
209 | ||
1e6b1085 | 210 | static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat, |
1899ad18 | 211 | unsigned long *evictionp, bool *workingsetp) |
a528910e | 212 | { |
3159f943 | 213 | unsigned long entry = xa_to_value(shadow); |
1e6b1085 | 214 | int memcgid, nid; |
1899ad18 | 215 | bool workingset; |
a528910e | 216 | |
3ebc57f4 ML |
217 | workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1); |
218 | entry >>= WORKINGSET_SHIFT; | |
a528910e JW |
219 | nid = entry & ((1UL << NODES_SHIFT) - 1); |
220 | entry >>= NODES_SHIFT; | |
23047a96 JW |
221 | memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1); |
222 | entry >>= MEM_CGROUP_ID_SHIFT; | |
a528910e | 223 | |
23047a96 | 224 | *memcgidp = memcgid; |
1e6b1085 | 225 | *pgdat = NODE_DATA(nid); |
ac35a490 | 226 | *evictionp = entry; |
1899ad18 | 227 | *workingsetp = workingset; |
a528910e JW |
228 | } |
229 | ||
ac35a490 YZ |
230 | #ifdef CONFIG_LRU_GEN |
231 | ||
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; | |
391655fe | 238 | struct lru_gen_folio *lrugen; |
ac35a490 YZ |
239 | int type = folio_is_file_lru(folio); |
240 | int delta = folio_nr_pages(folio); | |
241 | int refs = folio_lru_refs(folio); | |
242 | int tier = lru_tier_from_refs(refs); | |
243 | struct mem_cgroup *memcg = folio_memcg(folio); | |
244 | struct pglist_data *pgdat = folio_pgdat(folio); | |
245 | ||
246 | BUILD_BUG_ON(LRU_GEN_WIDTH + LRU_REFS_WIDTH > BITS_PER_LONG - EVICTION_SHIFT); | |
247 | ||
248 | lruvec = mem_cgroup_lruvec(memcg, pgdat); | |
249 | lrugen = &lruvec->lrugen; | |
250 | min_seq = READ_ONCE(lrugen->min_seq[type]); | |
251 | token = (min_seq << LRU_REFS_WIDTH) | max(refs - 1, 0); | |
252 | ||
253 | hist = lru_hist_from_seq(min_seq); | |
254 | atomic_long_add(delta, &lrugen->evicted[hist][type][tier]); | |
255 | ||
256 | return pack_shadow(mem_cgroup_id(memcg), pgdat, token, refs); | |
257 | } | |
258 | ||
ffcb5f52 NP |
259 | /* |
260 | * Tests if the shadow entry is for a folio that was recently evicted. | |
d7f1afd0 | 261 | * Fills in @lruvec, @token, @workingset with the values unpacked from shadow. |
ffcb5f52 | 262 | */ |
d7f1afd0 A |
263 | static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec, |
264 | unsigned long *token, bool *workingset) | |
ffcb5f52 | 265 | { |
d7f1afd0 | 266 | int memcg_id; |
ffcb5f52 | 267 | unsigned long min_seq; |
d7f1afd0 A |
268 | struct mem_cgroup *memcg; |
269 | struct pglist_data *pgdat; | |
ffcb5f52 | 270 | |
d7f1afd0 | 271 | unpack_shadow(shadow, &memcg_id, &pgdat, token, workingset); |
ffcb5f52 | 272 | |
d7f1afd0 A |
273 | memcg = mem_cgroup_from_id(memcg_id); |
274 | *lruvec = mem_cgroup_lruvec(memcg, pgdat); | |
ffcb5f52 | 275 | |
d7f1afd0 | 276 | min_seq = READ_ONCE((*lruvec)->lrugen.min_seq[file]); |
ffcb5f52 NP |
277 | return (*token >> LRU_REFS_WIDTH) == (min_seq & (EVICTION_MASK >> LRU_REFS_WIDTH)); |
278 | } | |
279 | ||
ac35a490 YZ |
280 | static void lru_gen_refault(struct folio *folio, void *shadow) |
281 | { | |
3af0191a | 282 | bool recent; |
ac35a490 | 283 | int hist, tier, refs; |
ac35a490 YZ |
284 | bool workingset; |
285 | unsigned long token; | |
ac35a490 | 286 | struct lruvec *lruvec; |
391655fe | 287 | struct lru_gen_folio *lrugen; |
ac35a490 YZ |
288 | int type = folio_is_file_lru(folio); |
289 | int delta = folio_nr_pages(folio); | |
290 | ||
ac35a490 YZ |
291 | rcu_read_lock(); |
292 | ||
3af0191a KS |
293 | recent = lru_gen_test_recent(shadow, type, &lruvec, &token, &workingset); |
294 | if (lruvec != folio_lruvec(folio)) | |
ac35a490 YZ |
295 | goto unlock; |
296 | ||
3af0191a KS |
297 | mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + type, delta); |
298 | ||
299 | if (!recent) | |
ffcb5f52 NP |
300 | goto unlock; |
301 | ||
ac35a490 | 302 | lrugen = &lruvec->lrugen; |
ac35a490 | 303 | |
d7f1afd0 | 304 | hist = lru_hist_from_seq(READ_ONCE(lrugen->min_seq[type])); |
ac35a490 YZ |
305 | /* see the comment in folio_lru_refs() */ |
306 | refs = (token & (BIT(LRU_REFS_WIDTH) - 1)) + workingset; | |
307 | tier = lru_tier_from_refs(refs); | |
308 | ||
309 | atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]); | |
3af0191a | 310 | mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta); |
ac35a490 YZ |
311 | |
312 | /* | |
313 | * Count the following two cases as stalls: | |
314 | * 1. For pages accessed through page tables, hotter pages pushed out | |
315 | * hot pages which refaulted immediately. | |
316 | * 2. For pages accessed multiple times through file descriptors, | |
08148805 | 317 | * they would have been protected by sort_folio(). |
ac35a490 | 318 | */ |
08148805 YZ |
319 | if (lru_gen_in_fault() || refs >= BIT(LRU_REFS_WIDTH) - 1) { |
320 | set_mask_bits(&folio->flags, 0, LRU_REFS_MASK | BIT(PG_workingset)); | |
ac35a490 YZ |
321 | mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + type, delta); |
322 | } | |
323 | unlock: | |
324 | rcu_read_unlock(); | |
325 | } | |
326 | ||
327 | #else /* !CONFIG_LRU_GEN */ | |
328 | ||
329 | static void *lru_gen_eviction(struct folio *folio) | |
330 | { | |
331 | return NULL; | |
332 | } | |
333 | ||
d7f1afd0 A |
334 | static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec, |
335 | unsigned long *token, bool *workingset) | |
ffcb5f52 NP |
336 | { |
337 | return false; | |
338 | } | |
339 | ||
ac35a490 YZ |
340 | static void lru_gen_refault(struct folio *folio, void *shadow) |
341 | { | |
342 | } | |
343 | ||
344 | #endif /* CONFIG_LRU_GEN */ | |
345 | ||
31d8fcac JW |
346 | /** |
347 | * workingset_age_nonresident - age non-resident entries as LRU ages | |
e755f4af | 348 | * @lruvec: the lruvec that was aged |
31d8fcac JW |
349 | * @nr_pages: the number of pages to count |
350 | * | |
351 | * As in-memory pages are aged, non-resident pages need to be aged as | |
352 | * well, in order for the refault distances later on to be comparable | |
353 | * to the in-memory dimensions. This function allows reclaim and LRU | |
354 | * operations to drive the non-resident aging along in parallel. | |
355 | */ | |
356 | void workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages) | |
b910718a JW |
357 | { |
358 | /* | |
359 | * Reclaiming a cgroup means reclaiming all its children in a | |
360 | * round-robin fashion. That means that each cgroup has an LRU | |
361 | * order that is composed of the LRU orders of its child | |
362 | * cgroups; and every page has an LRU position not just in the | |
363 | * cgroup that owns it, but in all of that group's ancestors. | |
364 | * | |
365 | * So when the physical inactive list of a leaf cgroup ages, | |
366 | * the virtual inactive lists of all its parents, including | |
367 | * the root cgroup's, age as well. | |
368 | */ | |
369 | do { | |
31d8fcac JW |
370 | atomic_long_add(nr_pages, &lruvec->nonresident_age); |
371 | } while ((lruvec = parent_lruvec(lruvec))); | |
b910718a JW |
372 | } |
373 | ||
a528910e | 374 | /** |
8927f647 | 375 | * workingset_eviction - note the eviction of a folio from memory |
b910718a | 376 | * @target_memcg: the cgroup that is causing the reclaim |
8927f647 | 377 | * @folio: the folio being evicted |
a528910e | 378 | * |
8927f647 MWO |
379 | * Return: a shadow entry to be stored in @folio->mapping->i_pages in place |
380 | * of the evicted @folio so that a later refault can be detected. | |
a528910e | 381 | */ |
8927f647 | 382 | void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg) |
a528910e | 383 | { |
8927f647 | 384 | struct pglist_data *pgdat = folio_pgdat(folio); |
a528910e | 385 | unsigned long eviction; |
23047a96 | 386 | struct lruvec *lruvec; |
b910718a | 387 | int memcgid; |
a528910e | 388 | |
8927f647 MWO |
389 | /* Folio is fully exclusive and pins folio's memory cgroup pointer */ |
390 | VM_BUG_ON_FOLIO(folio_test_lru(folio), folio); | |
391 | VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); | |
392 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); | |
23047a96 | 393 | |
ac35a490 YZ |
394 | if (lru_gen_enabled()) |
395 | return lru_gen_eviction(folio); | |
396 | ||
b910718a JW |
397 | lruvec = mem_cgroup_lruvec(target_memcg, pgdat); |
398 | /* XXX: target_memcg can be NULL, go through lruvec */ | |
399 | memcgid = mem_cgroup_id(lruvec_memcg(lruvec)); | |
31d8fcac | 400 | eviction = atomic_long_read(&lruvec->nonresident_age); |
ac35a490 | 401 | eviction >>= bucket_order; |
8927f647 MWO |
402 | workingset_age_nonresident(lruvec, folio_nr_pages(folio)); |
403 | return pack_shadow(memcgid, pgdat, eviction, | |
404 | folio_test_workingset(folio)); | |
a528910e JW |
405 | } |
406 | ||
407 | /** | |
ffcb5f52 NP |
408 | * workingset_test_recent - tests if the shadow entry is for a folio that was |
409 | * recently evicted. Also fills in @workingset with the value unpacked from | |
410 | * shadow. | |
411 | * @shadow: the shadow entry to be tested. | |
412 | * @file: whether the corresponding folio is from the file lru. | |
413 | * @workingset: where the workingset value unpacked from shadow should | |
414 | * be stored. | |
415 | * | |
416 | * Return: true if the shadow is for a recently evicted folio; false otherwise. | |
a528910e | 417 | */ |
ffcb5f52 | 418 | bool workingset_test_recent(void *shadow, bool file, bool *workingset) |
a528910e | 419 | { |
b910718a JW |
420 | struct mem_cgroup *eviction_memcg; |
421 | struct lruvec *eviction_lruvec; | |
a528910e | 422 | unsigned long refault_distance; |
34e58cac | 423 | unsigned long workingset_size; |
162453bf | 424 | unsigned long refault; |
23047a96 | 425 | int memcgid; |
ffcb5f52 NP |
426 | struct pglist_data *pgdat; |
427 | unsigned long eviction; | |
a528910e | 428 | |
b0068472 YA |
429 | rcu_read_lock(); |
430 | ||
431 | if (lru_gen_enabled()) { | |
432 | bool recent = lru_gen_test_recent(shadow, file, | |
433 | &eviction_lruvec, &eviction, workingset); | |
434 | ||
435 | rcu_read_unlock(); | |
436 | return recent; | |
437 | } | |
438 | ||
ac35a490 | 439 | |
ffcb5f52 | 440 | unpack_shadow(shadow, &memcgid, &pgdat, &eviction, workingset); |
ac35a490 | 441 | eviction <<= bucket_order; |
162453bf | 442 | |
23047a96 JW |
443 | /* |
444 | * Look up the memcg associated with the stored ID. It might | |
0995d7e5 | 445 | * have been deleted since the folio's eviction. |
23047a96 JW |
446 | * |
447 | * Note that in rare events the ID could have been recycled | |
0995d7e5 | 448 | * for a new cgroup that refaults a shared folio. This is |
23047a96 JW |
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 | */ | |
b910718a | 459 | eviction_memcg = mem_cgroup_from_id(memcgid); |
b0068472 YA |
460 | if (!mem_cgroup_disabled() && |
461 | (!eviction_memcg || !mem_cgroup_tryget(eviction_memcg))) { | |
462 | rcu_read_unlock(); | |
ffcb5f52 | 463 | return false; |
b0068472 YA |
464 | } |
465 | ||
466 | rcu_read_unlock(); | |
467 | ||
7d7ef0a4 YA |
468 | /* |
469 | * Flush stats (and potentially sleep) outside the RCU read section. | |
470 | * XXX: With per-memcg flushing and thresholding, is ratelimiting | |
471 | * still needed here? | |
472 | */ | |
473 | mem_cgroup_flush_stats_ratelimited(eviction_memcg); | |
ffcb5f52 | 474 | |
b910718a | 475 | eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat); |
31d8fcac | 476 | refault = atomic_long_read(&eviction_lruvec->nonresident_age); |
162453bf JW |
477 | |
478 | /* | |
1899ad18 | 479 | * Calculate the refault distance |
162453bf | 480 | * |
1899ad18 | 481 | * The unsigned subtraction here gives an accurate distance |
31d8fcac | 482 | * across nonresident_age overflows in most cases. There is a |
1899ad18 JW |
483 | * special case: usually, shadow entries have a short lifetime |
484 | * and are either refaulted or reclaimed along with the inode | |
485 | * before they get too old. But it is not impossible for the | |
31d8fcac JW |
486 | * nonresident_age to lap a shadow entry in the field, which |
487 | * can then result in a false small refault distance, leading | |
488 | * to a false activation should this old entry actually | |
489 | * refault again. However, earlier kernels used to deactivate | |
1899ad18 JW |
490 | * unconditionally with *every* reclaim invocation for the |
491 | * longest time, so the occasional inappropriate activation | |
492 | * leading to pressure on the active list is not a problem. | |
162453bf JW |
493 | */ |
494 | refault_distance = (refault - eviction) & EVICTION_MASK; | |
495 | ||
1899ad18 JW |
496 | /* |
497 | * Compare the distance to the existing workingset size. We | |
34e58cac | 498 | * don't activate pages that couldn't stay resident even if |
aae466b0 JK |
499 | * all the memory was available to the workingset. Whether |
500 | * workingset competition needs to consider anon or not depends | |
ed8f3f99 | 501 | * on having free swap space. |
1899ad18 | 502 | */ |
34e58cac | 503 | workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE); |
aae466b0 | 504 | if (!file) { |
34e58cac | 505 | workingset_size += lruvec_page_state(eviction_lruvec, |
aae466b0 JK |
506 | NR_INACTIVE_FILE); |
507 | } | |
f78dfc7b | 508 | if (mem_cgroup_get_nr_swap_pages(eviction_memcg) > 0) { |
34e58cac JW |
509 | workingset_size += lruvec_page_state(eviction_lruvec, |
510 | NR_ACTIVE_ANON); | |
aae466b0 JK |
511 | if (file) { |
512 | workingset_size += lruvec_page_state(eviction_lruvec, | |
513 | NR_INACTIVE_ANON); | |
514 | } | |
34e58cac | 515 | } |
ffcb5f52 | 516 | |
b0068472 | 517 | mem_cgroup_put(eviction_memcg); |
ffcb5f52 NP |
518 | return refault_distance <= workingset_size; |
519 | } | |
520 | ||
521 | /** | |
522 | * workingset_refault - Evaluate the refault of a previously evicted folio. | |
523 | * @folio: The freshly allocated replacement folio. | |
524 | * @shadow: Shadow entry of the evicted folio. | |
525 | * | |
526 | * Calculates and evaluates the refault distance of the previously | |
527 | * evicted folio in the context of the node and the memcg whose memory | |
528 | * pressure caused the eviction. | |
529 | */ | |
530 | void workingset_refault(struct folio *folio, void *shadow) | |
531 | { | |
532 | bool file = folio_is_file_lru(folio); | |
533 | struct pglist_data *pgdat; | |
534 | struct mem_cgroup *memcg; | |
535 | struct lruvec *lruvec; | |
536 | bool workingset; | |
537 | long nr; | |
538 | ||
539 | if (lru_gen_enabled()) { | |
540 | lru_gen_refault(folio, shadow); | |
541 | return; | |
542 | } | |
543 | ||
ffcb5f52 NP |
544 | /* |
545 | * The activation decision for this folio is made at the level | |
546 | * where the eviction occurred, as that is where the LRU order | |
547 | * during folio reclaim is being determined. | |
548 | * | |
549 | * However, the cgroup that will own the folio is the one that | |
b0068472 YA |
550 | * is actually experiencing the refault event. Make sure the folio is |
551 | * locked to guarantee folio_memcg() stability throughout. | |
ffcb5f52 | 552 | */ |
b0068472 | 553 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
ffcb5f52 NP |
554 | nr = folio_nr_pages(folio); |
555 | memcg = folio_memcg(folio); | |
556 | pgdat = folio_pgdat(folio); | |
557 | lruvec = mem_cgroup_lruvec(memcg, pgdat); | |
558 | ||
559 | mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr); | |
560 | ||
561 | if (!workingset_test_recent(shadow, file, &workingset)) | |
b0068472 | 562 | return; |
1899ad18 | 563 | |
0995d7e5 MWO |
564 | folio_set_active(folio); |
565 | workingset_age_nonresident(lruvec, nr); | |
566 | mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr); | |
1899ad18 | 567 | |
0995d7e5 | 568 | /* Folio was active prior to eviction */ |
1899ad18 | 569 | if (workingset) { |
0995d7e5 | 570 | folio_set_workingset(folio); |
6e1ca48d VMO |
571 | /* |
572 | * XXX: Move to folio_add_lru() when it supports new vs | |
573 | * putback | |
574 | */ | |
0538a82c | 575 | lru_note_cost_refault(folio); |
0995d7e5 | 576 | mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr); |
a528910e | 577 | } |
a528910e JW |
578 | } |
579 | ||
580 | /** | |
581 | * workingset_activation - note a page activation | |
c5ce619a | 582 | * @folio: Folio that is being activated. |
a528910e | 583 | */ |
c5ce619a | 584 | void workingset_activation(struct folio *folio) |
a528910e | 585 | { |
55779ec7 | 586 | struct mem_cgroup *memcg; |
23047a96 | 587 | |
55779ec7 | 588 | rcu_read_lock(); |
23047a96 JW |
589 | /* |
590 | * Filter non-memcg pages here, e.g. unmap can call | |
591 | * mark_page_accessed() on VDSO pages. | |
592 | * | |
593 | * XXX: See workingset_refault() - this should return | |
594 | * root_mem_cgroup even for !CONFIG_MEMCG. | |
595 | */ | |
c5ce619a | 596 | memcg = folio_memcg_rcu(folio); |
55779ec7 | 597 | if (!mem_cgroup_disabled() && !memcg) |
23047a96 | 598 | goto out; |
c5ce619a | 599 | workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio)); |
23047a96 | 600 | out: |
55779ec7 | 601 | rcu_read_unlock(); |
a528910e | 602 | } |
449dd698 JW |
603 | |
604 | /* | |
605 | * Shadow entries reflect the share of the working set that does not | |
606 | * fit into memory, so their number depends on the access pattern of | |
607 | * the workload. In most cases, they will refault or get reclaimed | |
608 | * along with the inode, but a (malicious) workload that streams | |
609 | * through files with a total size several times that of available | |
610 | * memory, while preventing the inodes from being reclaimed, can | |
611 | * create excessive amounts of shadow nodes. To keep a lid on this, | |
612 | * track shadow nodes and reclaim them when they grow way past the | |
613 | * point where they would still be useful. | |
614 | */ | |
615 | ||
9bbdc0f3 | 616 | struct list_lru shadow_nodes; |
14b46879 | 617 | |
a97e7904 | 618 | void workingset_update_node(struct xa_node *node) |
14b46879 | 619 | { |
2386eef2 | 620 | struct address_space *mapping; |
4715c6a7 | 621 | struct page *page = virt_to_page(node); |
2386eef2 | 622 | |
14b46879 JW |
623 | /* |
624 | * Track non-empty nodes that contain only shadow entries; | |
625 | * unlink those that contain pages or are being freed. | |
626 | * | |
627 | * Avoid acquiring the list_lru lock when the nodes are | |
628 | * already where they should be. The list_empty() test is safe | |
b93b0163 | 629 | * as node->private_list is protected by the i_pages lock. |
14b46879 | 630 | */ |
2386eef2 SAS |
631 | mapping = container_of(node->array, struct address_space, i_pages); |
632 | lockdep_assert_held(&mapping->i_pages.xa_lock); | |
68d48e6a | 633 | |
01959dfe | 634 | if (node->count && node->count == node->nr_values) { |
68d48e6a | 635 | if (list_empty(&node->private_list)) { |
0a97c01c | 636 | list_lru_add_obj(&shadow_nodes, &node->private_list); |
4715c6a7 | 637 | __inc_node_page_state(page, WORKINGSET_NODES); |
68d48e6a | 638 | } |
14b46879 | 639 | } else { |
68d48e6a | 640 | if (!list_empty(&node->private_list)) { |
0a97c01c | 641 | list_lru_del_obj(&shadow_nodes, &node->private_list); |
4715c6a7 | 642 | __dec_node_page_state(page, WORKINGSET_NODES); |
68d48e6a | 643 | } |
14b46879 JW |
644 | } |
645 | } | |
449dd698 JW |
646 | |
647 | static unsigned long count_shadow_nodes(struct shrinker *shrinker, | |
648 | struct shrink_control *sc) | |
649 | { | |
449dd698 | 650 | unsigned long max_nodes; |
14b46879 | 651 | unsigned long nodes; |
95f9ab2d | 652 | unsigned long pages; |
449dd698 | 653 | |
14b46879 | 654 | nodes = list_lru_shrink_count(&shadow_nodes, sc); |
725cac1c ML |
655 | if (!nodes) |
656 | return SHRINK_EMPTY; | |
449dd698 | 657 | |
449dd698 | 658 | /* |
a97e7904 | 659 | * Approximate a reasonable limit for the nodes |
b5388998 JW |
660 | * containing shadow entries. We don't need to keep more |
661 | * shadow entries than possible pages on the active list, | |
662 | * since refault distances bigger than that are dismissed. | |
663 | * | |
664 | * The size of the active list converges toward 100% of | |
665 | * overall page cache as memory grows, with only a tiny | |
666 | * inactive list. Assume the total cache size for that. | |
667 | * | |
668 | * Nodes might be sparsely populated, with only one shadow | |
669 | * entry in the extreme case. Obviously, we cannot keep one | |
670 | * node for every eligible shadow entry, so compromise on a | |
671 | * worst-case density of 1/8th. Below that, not all eligible | |
672 | * refaults can be detected anymore. | |
449dd698 | 673 | * |
a97e7904 | 674 | * On 64-bit with 7 xa_nodes per page and 64 slots |
449dd698 | 675 | * each, this will reclaim shadow entries when they consume |
b5388998 | 676 | * ~1.8% of available memory: |
449dd698 | 677 | * |
a97e7904 | 678 | * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE |
449dd698 | 679 | */ |
95f9ab2d | 680 | #ifdef CONFIG_MEMCG |
b5388998 | 681 | if (sc->memcg) { |
95f9ab2d | 682 | struct lruvec *lruvec; |
2b487e59 | 683 | int i; |
95f9ab2d | 684 | |
d4a5b369 | 685 | mem_cgroup_flush_stats_ratelimited(sc->memcg); |
867e5e1d | 686 | lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid)); |
2b487e59 | 687 | for (pages = 0, i = 0; i < NR_LRU_LISTS; i++) |
205b20cc JW |
688 | pages += lruvec_page_state_local(lruvec, |
689 | NR_LRU_BASE + i); | |
d42f3245 RG |
690 | pages += lruvec_page_state_local( |
691 | lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT; | |
692 | pages += lruvec_page_state_local( | |
693 | lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT; | |
95f9ab2d JW |
694 | } else |
695 | #endif | |
696 | pages = node_present_pages(sc->nid); | |
697 | ||
dad4f140 | 698 | max_nodes = pages >> (XA_CHUNK_SHIFT - 3); |
449dd698 | 699 | |
14b46879 | 700 | if (nodes <= max_nodes) |
449dd698 | 701 | return 0; |
14b46879 | 702 | return nodes - max_nodes; |
449dd698 JW |
703 | } |
704 | ||
705 | static enum lru_status shadow_lru_isolate(struct list_head *item, | |
3f97b163 | 706 | struct list_lru_one *lru, |
449dd698 | 707 | spinlock_t *lru_lock, |
a97e7904 | 708 | void *arg) __must_hold(lru_lock) |
449dd698 | 709 | { |
a97e7904 | 710 | struct xa_node *node = container_of(item, struct xa_node, private_list); |
449dd698 | 711 | struct address_space *mapping; |
449dd698 JW |
712 | int ret; |
713 | ||
714 | /* | |
f82cd2f0 | 715 | * Page cache insertions and deletions synchronously maintain |
b93b0163 | 716 | * the shadow node LRU under the i_pages lock and the |
449dd698 JW |
717 | * lru_lock. Because the page cache tree is emptied before |
718 | * the inode can be destroyed, holding the lru_lock pins any | |
a97e7904 | 719 | * address_space that has nodes on the LRU. |
449dd698 | 720 | * |
b93b0163 | 721 | * We can then safely transition to the i_pages lock to |
449dd698 JW |
722 | * pin only the address_space of the particular node we want |
723 | * to reclaim, take the node off-LRU, and drop the lru_lock. | |
724 | */ | |
725 | ||
01959dfe | 726 | mapping = container_of(node->array, struct address_space, i_pages); |
449dd698 JW |
727 | |
728 | /* Coming from the list, invert the lock order */ | |
b93b0163 | 729 | if (!xa_trylock(&mapping->i_pages)) { |
6ca342d0 | 730 | spin_unlock_irq(lru_lock); |
449dd698 JW |
731 | ret = LRU_RETRY; |
732 | goto out; | |
733 | } | |
734 | ||
5649d113 YY |
735 | /* For page cache we need to hold i_lock */ |
736 | if (mapping->host != NULL) { | |
737 | if (!spin_trylock(&mapping->host->i_lock)) { | |
738 | xa_unlock(&mapping->i_pages); | |
739 | spin_unlock_irq(lru_lock); | |
740 | ret = LRU_RETRY; | |
741 | goto out; | |
742 | } | |
51b8c1fe JW |
743 | } |
744 | ||
3f97b163 | 745 | list_lru_isolate(lru, item); |
4715c6a7 | 746 | __dec_node_page_state(virt_to_page(node), WORKINGSET_NODES); |
68d48e6a | 747 | |
449dd698 JW |
748 | spin_unlock(lru_lock); |
749 | ||
750 | /* | |
751 | * The nodes should only contain one or more shadow entries, | |
752 | * no pages, so we expect to be able to remove them all and | |
753 | * delete and free the empty node afterwards. | |
754 | */ | |
01959dfe | 755 | if (WARN_ON_ONCE(!node->nr_values)) |
b936887e | 756 | goto out_invalid; |
01959dfe | 757 | if (WARN_ON_ONCE(node->count != node->nr_values)) |
b936887e | 758 | goto out_invalid; |
f82cd2f0 | 759 | xa_delete_node(node, workingset_update_node); |
da3ceeff | 760 | __inc_lruvec_kmem_state(node, WORKINGSET_NODERECLAIM); |
449dd698 | 761 | |
b936887e | 762 | out_invalid: |
6ca342d0 | 763 | xa_unlock_irq(&mapping->i_pages); |
5649d113 YY |
764 | if (mapping->host != NULL) { |
765 | if (mapping_shrinkable(mapping)) | |
766 | inode_add_lru(mapping->host); | |
767 | spin_unlock(&mapping->host->i_lock); | |
768 | } | |
449dd698 JW |
769 | ret = LRU_REMOVED_RETRY; |
770 | out: | |
449dd698 | 771 | cond_resched(); |
6ca342d0 | 772 | spin_lock_irq(lru_lock); |
449dd698 JW |
773 | return ret; |
774 | } | |
775 | ||
776 | static unsigned long scan_shadow_nodes(struct shrinker *shrinker, | |
777 | struct shrink_control *sc) | |
778 | { | |
b93b0163 | 779 | /* list_lru lock nests inside the IRQ-safe i_pages lock */ |
6b51e881 SAS |
780 | return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate, |
781 | NULL); | |
449dd698 JW |
782 | } |
783 | ||
449dd698 JW |
784 | /* |
785 | * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe | |
b93b0163 | 786 | * i_pages lock. |
449dd698 JW |
787 | */ |
788 | static struct lock_class_key shadow_nodes_key; | |
789 | ||
790 | static int __init workingset_init(void) | |
791 | { | |
219c666e | 792 | struct shrinker *workingset_shadow_shrinker; |
612e4493 JW |
793 | unsigned int timestamp_bits; |
794 | unsigned int max_order; | |
219c666e | 795 | int ret = -ENOMEM; |
449dd698 | 796 | |
612e4493 JW |
797 | BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT); |
798 | /* | |
799 | * Calculate the eviction bucket size to cover the longest | |
800 | * actionable refault distance, which is currently half of | |
801 | * memory (totalram_pages/2). However, memory hotplug may add | |
802 | * some more pages at runtime, so keep working with up to | |
803 | * double the initial memory by using totalram_pages as-is. | |
804 | */ | |
805 | timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT; | |
ca79b0c2 | 806 | max_order = fls_long(totalram_pages() - 1); |
612e4493 JW |
807 | if (max_order > timestamp_bits) |
808 | bucket_order = max_order - timestamp_bits; | |
d3d36c4b | 809 | pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n", |
612e4493 JW |
810 | timestamp_bits, max_order, bucket_order); |
811 | ||
219c666e QZ |
812 | workingset_shadow_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE | |
813 | SHRINKER_MEMCG_AWARE, | |
814 | "mm-shadow"); | |
815 | if (!workingset_shadow_shrinker) | |
449dd698 | 816 | goto err; |
219c666e | 817 | |
c92e8e10 | 818 | ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key, |
219c666e | 819 | workingset_shadow_shrinker); |
449dd698 JW |
820 | if (ret) |
821 | goto err_list_lru; | |
219c666e QZ |
822 | |
823 | workingset_shadow_shrinker->count_objects = count_shadow_nodes; | |
824 | workingset_shadow_shrinker->scan_objects = scan_shadow_nodes; | |
825 | /* ->count reports only fully expendable nodes */ | |
826 | workingset_shadow_shrinker->seeks = 0; | |
827 | ||
828 | shrinker_register(workingset_shadow_shrinker); | |
449dd698 JW |
829 | return 0; |
830 | err_list_lru: | |
219c666e | 831 | shrinker_free(workingset_shadow_shrinker); |
449dd698 JW |
832 | err: |
833 | return ret; | |
834 | } | |
835 | module_init(workingset_init); |