]>
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> | |
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 | * | |
ed8f3f99 YY |
114 | * If we have swap we should consider about NR_inactive_anon and |
115 | * NR_active_anon, so for page cache and anonymous respectively: | |
a528910e | 116 | * |
ed8f3f99 YY |
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 | |
a528910e JW |
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 | * | |
1899ad18 | 136 | * Refaulting inactive pages |
a528910e JW |
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 | |
ed8f3f99 YY |
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 | |
a528910e JW |
146 | * used less frequently than the refaulting page - or even not used at |
147 | * all anymore. | |
148 | * | |
1899ad18 JW |
149 | * That means if inactive cache is refaulting with a suitable refault |
150 | * distance, we assume the cache workingset is transitioning and put | |
ed8f3f99 | 151 | * pressure on the current workingset. |
1899ad18 | 152 | * |
a528910e JW |
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 | * | |
1899ad18 JW |
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 | * | |
a528910e JW |
168 | * |
169 | * Implementation | |
170 | * | |
31d8fcac JW |
171 | * For each node's LRU lists, a counter for inactive evictions and |
172 | * activations is maintained (node->nonresident_age). | |
a528910e JW |
173 | * |
174 | * On eviction, a snapshot of this counter (along with some bits to | |
a97e7904 | 175 | * identify the node) is stored in the now empty page cache |
a528910e JW |
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 | ||
3ebc57f4 | 182 | #define WORKINGSET_SHIFT 1 |
3159f943 | 183 | #define EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \ |
3ebc57f4 ML |
184 | WORKINGSET_SHIFT + NODES_SHIFT + \ |
185 | MEM_CGROUP_ID_SHIFT) | |
689c94f0 JW |
186 | #define EVICTION_MASK (~0UL >> EVICTION_SHIFT) |
187 | ||
612e4493 JW |
188 | /* |
189 | * Eviction timestamps need to be able to cover the full range of | |
a97e7904 | 190 | * actionable refaults. However, bits are tight in the xarray |
612e4493 JW |
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 | */ | |
196 | static unsigned int bucket_order __read_mostly; | |
197 | ||
1899ad18 JW |
198 | static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction, |
199 | bool workingset) | |
a528910e | 200 | { |
3159f943 | 201 | eviction &= EVICTION_MASK; |
23047a96 | 202 | eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid; |
1e6b1085 | 203 | eviction = (eviction << NODES_SHIFT) | pgdat->node_id; |
3ebc57f4 | 204 | eviction = (eviction << WORKINGSET_SHIFT) | workingset; |
a528910e | 205 | |
3159f943 | 206 | return xa_mk_value(eviction); |
a528910e JW |
207 | } |
208 | ||
1e6b1085 | 209 | static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat, |
1899ad18 | 210 | unsigned long *evictionp, bool *workingsetp) |
a528910e | 211 | { |
3159f943 | 212 | unsigned long entry = xa_to_value(shadow); |
1e6b1085 | 213 | int memcgid, nid; |
1899ad18 | 214 | bool workingset; |
a528910e | 215 | |
3ebc57f4 ML |
216 | workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1); |
217 | entry >>= WORKINGSET_SHIFT; | |
a528910e JW |
218 | nid = entry & ((1UL << NODES_SHIFT) - 1); |
219 | entry >>= NODES_SHIFT; | |
23047a96 JW |
220 | memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1); |
221 | entry >>= MEM_CGROUP_ID_SHIFT; | |
a528910e | 222 | |
23047a96 | 223 | *memcgidp = memcgid; |
1e6b1085 | 224 | *pgdat = NODE_DATA(nid); |
ac35a490 | 225 | *evictionp = entry; |
1899ad18 | 226 | *workingsetp = workingset; |
a528910e JW |
227 | } |
228 | ||
ac35a490 YZ |
229 | #ifdef CONFIG_LRU_GEN |
230 | ||
231 | static void *lru_gen_eviction(struct folio *folio) | |
232 | { | |
233 | int hist; | |
234 | unsigned long token; | |
235 | unsigned long min_seq; | |
236 | struct lruvec *lruvec; | |
391655fe | 237 | struct lru_gen_folio *lrugen; |
ac35a490 YZ |
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 | static void lru_gen_refault(struct folio *folio, void *shadow) | |
259 | { | |
260 | int hist, tier, refs; | |
261 | int memcg_id; | |
262 | bool workingset; | |
263 | unsigned long token; | |
264 | unsigned long min_seq; | |
265 | struct lruvec *lruvec; | |
391655fe | 266 | struct lru_gen_folio *lrugen; |
ac35a490 YZ |
267 | struct mem_cgroup *memcg; |
268 | struct pglist_data *pgdat; | |
269 | int type = folio_is_file_lru(folio); | |
270 | int delta = folio_nr_pages(folio); | |
271 | ||
272 | unpack_shadow(shadow, &memcg_id, &pgdat, &token, &workingset); | |
273 | ||
274 | if (pgdat != folio_pgdat(folio)) | |
275 | return; | |
276 | ||
277 | rcu_read_lock(); | |
278 | ||
279 | memcg = folio_memcg_rcu(folio); | |
280 | if (memcg_id != mem_cgroup_id(memcg)) | |
281 | goto unlock; | |
282 | ||
283 | lruvec = mem_cgroup_lruvec(memcg, pgdat); | |
284 | lrugen = &lruvec->lrugen; | |
285 | ||
286 | min_seq = READ_ONCE(lrugen->min_seq[type]); | |
287 | if ((token >> LRU_REFS_WIDTH) != (min_seq & (EVICTION_MASK >> LRU_REFS_WIDTH))) | |
288 | goto unlock; | |
289 | ||
290 | hist = lru_hist_from_seq(min_seq); | |
291 | /* see the comment in folio_lru_refs() */ | |
292 | refs = (token & (BIT(LRU_REFS_WIDTH) - 1)) + workingset; | |
293 | tier = lru_tier_from_refs(refs); | |
294 | ||
295 | atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]); | |
296 | mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + type, delta); | |
297 | ||
298 | /* | |
299 | * Count the following two cases as stalls: | |
300 | * 1. For pages accessed through page tables, hotter pages pushed out | |
301 | * hot pages which refaulted immediately. | |
302 | * 2. For pages accessed multiple times through file descriptors, | |
303 | * numbers of accesses might have been out of the range. | |
304 | */ | |
305 | if (lru_gen_in_fault() || refs == BIT(LRU_REFS_WIDTH)) { | |
306 | folio_set_workingset(folio); | |
307 | mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + type, delta); | |
308 | } | |
309 | unlock: | |
310 | rcu_read_unlock(); | |
311 | } | |
312 | ||
313 | #else /* !CONFIG_LRU_GEN */ | |
314 | ||
315 | static void *lru_gen_eviction(struct folio *folio) | |
316 | { | |
317 | return NULL; | |
318 | } | |
319 | ||
320 | static void lru_gen_refault(struct folio *folio, void *shadow) | |
321 | { | |
322 | } | |
323 | ||
324 | #endif /* CONFIG_LRU_GEN */ | |
325 | ||
31d8fcac JW |
326 | /** |
327 | * workingset_age_nonresident - age non-resident entries as LRU ages | |
e755f4af | 328 | * @lruvec: the lruvec that was aged |
31d8fcac JW |
329 | * @nr_pages: the number of pages to count |
330 | * | |
331 | * As in-memory pages are aged, non-resident pages need to be aged as | |
332 | * well, in order for the refault distances later on to be comparable | |
333 | * to the in-memory dimensions. This function allows reclaim and LRU | |
334 | * operations to drive the non-resident aging along in parallel. | |
335 | */ | |
336 | void workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages) | |
b910718a JW |
337 | { |
338 | /* | |
339 | * Reclaiming a cgroup means reclaiming all its children in a | |
340 | * round-robin fashion. That means that each cgroup has an LRU | |
341 | * order that is composed of the LRU orders of its child | |
342 | * cgroups; and every page has an LRU position not just in the | |
343 | * cgroup that owns it, but in all of that group's ancestors. | |
344 | * | |
345 | * So when the physical inactive list of a leaf cgroup ages, | |
346 | * the virtual inactive lists of all its parents, including | |
347 | * the root cgroup's, age as well. | |
348 | */ | |
349 | do { | |
31d8fcac JW |
350 | atomic_long_add(nr_pages, &lruvec->nonresident_age); |
351 | } while ((lruvec = parent_lruvec(lruvec))); | |
b910718a JW |
352 | } |
353 | ||
a528910e | 354 | /** |
8927f647 | 355 | * workingset_eviction - note the eviction of a folio from memory |
b910718a | 356 | * @target_memcg: the cgroup that is causing the reclaim |
8927f647 | 357 | * @folio: the folio being evicted |
a528910e | 358 | * |
8927f647 MWO |
359 | * Return: a shadow entry to be stored in @folio->mapping->i_pages in place |
360 | * of the evicted @folio so that a later refault can be detected. | |
a528910e | 361 | */ |
8927f647 | 362 | void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg) |
a528910e | 363 | { |
8927f647 | 364 | struct pglist_data *pgdat = folio_pgdat(folio); |
a528910e | 365 | unsigned long eviction; |
23047a96 | 366 | struct lruvec *lruvec; |
b910718a | 367 | int memcgid; |
a528910e | 368 | |
8927f647 MWO |
369 | /* Folio is fully exclusive and pins folio's memory cgroup pointer */ |
370 | VM_BUG_ON_FOLIO(folio_test_lru(folio), folio); | |
371 | VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); | |
372 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); | |
23047a96 | 373 | |
ac35a490 YZ |
374 | if (lru_gen_enabled()) |
375 | return lru_gen_eviction(folio); | |
376 | ||
b910718a JW |
377 | lruvec = mem_cgroup_lruvec(target_memcg, pgdat); |
378 | /* XXX: target_memcg can be NULL, go through lruvec */ | |
379 | memcgid = mem_cgroup_id(lruvec_memcg(lruvec)); | |
31d8fcac | 380 | eviction = atomic_long_read(&lruvec->nonresident_age); |
ac35a490 | 381 | eviction >>= bucket_order; |
8927f647 MWO |
382 | workingset_age_nonresident(lruvec, folio_nr_pages(folio)); |
383 | return pack_shadow(memcgid, pgdat, eviction, | |
384 | folio_test_workingset(folio)); | |
a528910e JW |
385 | } |
386 | ||
387 | /** | |
0995d7e5 MWO |
388 | * workingset_refault - Evaluate the refault of a previously evicted folio. |
389 | * @folio: The freshly allocated replacement folio. | |
390 | * @shadow: Shadow entry of the evicted folio. | |
a528910e JW |
391 | * |
392 | * Calculates and evaluates the refault distance of the previously | |
0995d7e5 | 393 | * evicted folio in the context of the node and the memcg whose memory |
b910718a | 394 | * pressure caused the eviction. |
a528910e | 395 | */ |
0995d7e5 | 396 | void workingset_refault(struct folio *folio, void *shadow) |
a528910e | 397 | { |
0995d7e5 | 398 | bool file = folio_is_file_lru(folio); |
b910718a JW |
399 | struct mem_cgroup *eviction_memcg; |
400 | struct lruvec *eviction_lruvec; | |
a528910e | 401 | unsigned long refault_distance; |
34e58cac | 402 | unsigned long workingset_size; |
1899ad18 | 403 | struct pglist_data *pgdat; |
23047a96 | 404 | struct mem_cgroup *memcg; |
162453bf | 405 | unsigned long eviction; |
23047a96 | 406 | struct lruvec *lruvec; |
162453bf | 407 | unsigned long refault; |
1899ad18 | 408 | bool workingset; |
23047a96 | 409 | int memcgid; |
0995d7e5 | 410 | long nr; |
a528910e | 411 | |
ac35a490 YZ |
412 | if (lru_gen_enabled()) { |
413 | lru_gen_refault(folio, shadow); | |
414 | return; | |
415 | } | |
416 | ||
1899ad18 | 417 | unpack_shadow(shadow, &memcgid, &pgdat, &eviction, &workingset); |
ac35a490 | 418 | eviction <<= bucket_order; |
162453bf | 419 | |
4009b2f1 YA |
420 | /* Flush stats (and potentially sleep) before holding RCU read lock */ |
421 | mem_cgroup_flush_stats_ratelimited(); | |
422 | ||
23047a96 JW |
423 | rcu_read_lock(); |
424 | /* | |
425 | * Look up the memcg associated with the stored ID. It might | |
0995d7e5 | 426 | * have been deleted since the folio's eviction. |
23047a96 JW |
427 | * |
428 | * Note that in rare events the ID could have been recycled | |
0995d7e5 | 429 | * for a new cgroup that refaults a shared folio. This is |
23047a96 JW |
430 | * impossible to tell from the available data. However, this |
431 | * should be a rare and limited disturbance, and activations | |
432 | * are always speculative anyway. Ultimately, it's the aging | |
433 | * algorithm's job to shake out the minimum access frequency | |
434 | * for the active cache. | |
435 | * | |
436 | * XXX: On !CONFIG_MEMCG, this will always return NULL; it | |
437 | * would be better if the root_mem_cgroup existed in all | |
438 | * configurations instead. | |
439 | */ | |
b910718a JW |
440 | eviction_memcg = mem_cgroup_from_id(memcgid); |
441 | if (!mem_cgroup_disabled() && !eviction_memcg) | |
1899ad18 | 442 | goto out; |
b910718a | 443 | eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat); |
31d8fcac | 444 | refault = atomic_long_read(&eviction_lruvec->nonresident_age); |
162453bf JW |
445 | |
446 | /* | |
1899ad18 | 447 | * Calculate the refault distance |
162453bf | 448 | * |
1899ad18 | 449 | * The unsigned subtraction here gives an accurate distance |
31d8fcac | 450 | * across nonresident_age overflows in most cases. There is a |
1899ad18 JW |
451 | * special case: usually, shadow entries have a short lifetime |
452 | * and are either refaulted or reclaimed along with the inode | |
453 | * before they get too old. But it is not impossible for the | |
31d8fcac JW |
454 | * nonresident_age to lap a shadow entry in the field, which |
455 | * can then result in a false small refault distance, leading | |
456 | * to a false activation should this old entry actually | |
457 | * refault again. However, earlier kernels used to deactivate | |
1899ad18 JW |
458 | * unconditionally with *every* reclaim invocation for the |
459 | * longest time, so the occasional inappropriate activation | |
460 | * leading to pressure on the active list is not a problem. | |
162453bf JW |
461 | */ |
462 | refault_distance = (refault - eviction) & EVICTION_MASK; | |
463 | ||
b910718a | 464 | /* |
0995d7e5 | 465 | * The activation decision for this folio is made at the level |
b910718a | 466 | * where the eviction occurred, as that is where the LRU order |
0995d7e5 | 467 | * during folio reclaim is being determined. |
b910718a | 468 | * |
0995d7e5 | 469 | * However, the cgroup that will own the folio is the one that |
b910718a JW |
470 | * is actually experiencing the refault event. |
471 | */ | |
0995d7e5 MWO |
472 | nr = folio_nr_pages(folio); |
473 | memcg = folio_memcg(folio); | |
f78dfc7b | 474 | pgdat = folio_pgdat(folio); |
b910718a JW |
475 | lruvec = mem_cgroup_lruvec(memcg, pgdat); |
476 | ||
0995d7e5 | 477 | mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr); |
1899ad18 JW |
478 | /* |
479 | * Compare the distance to the existing workingset size. We | |
34e58cac | 480 | * don't activate pages that couldn't stay resident even if |
aae466b0 JK |
481 | * all the memory was available to the workingset. Whether |
482 | * workingset competition needs to consider anon or not depends | |
ed8f3f99 | 483 | * on having free swap space. |
1899ad18 | 484 | */ |
34e58cac | 485 | workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE); |
aae466b0 | 486 | if (!file) { |
34e58cac | 487 | workingset_size += lruvec_page_state(eviction_lruvec, |
aae466b0 JK |
488 | NR_INACTIVE_FILE); |
489 | } | |
f78dfc7b | 490 | if (mem_cgroup_get_nr_swap_pages(eviction_memcg) > 0) { |
34e58cac JW |
491 | workingset_size += lruvec_page_state(eviction_lruvec, |
492 | NR_ACTIVE_ANON); | |
aae466b0 JK |
493 | if (file) { |
494 | workingset_size += lruvec_page_state(eviction_lruvec, | |
495 | NR_INACTIVE_ANON); | |
496 | } | |
34e58cac JW |
497 | } |
498 | if (refault_distance > workingset_size) | |
1899ad18 JW |
499 | goto out; |
500 | ||
0995d7e5 MWO |
501 | folio_set_active(folio); |
502 | workingset_age_nonresident(lruvec, nr); | |
503 | mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr); | |
1899ad18 | 504 | |
0995d7e5 | 505 | /* Folio was active prior to eviction */ |
1899ad18 | 506 | if (workingset) { |
0995d7e5 | 507 | folio_set_workingset(folio); |
6e1ca48d VMO |
508 | /* |
509 | * XXX: Move to folio_add_lru() when it supports new vs | |
510 | * putback | |
511 | */ | |
0538a82c | 512 | lru_note_cost_refault(folio); |
0995d7e5 | 513 | mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr); |
a528910e | 514 | } |
1899ad18 | 515 | out: |
2a2e4885 | 516 | rcu_read_unlock(); |
a528910e JW |
517 | } |
518 | ||
519 | /** | |
520 | * workingset_activation - note a page activation | |
c5ce619a | 521 | * @folio: Folio that is being activated. |
a528910e | 522 | */ |
c5ce619a | 523 | void workingset_activation(struct folio *folio) |
a528910e | 524 | { |
55779ec7 | 525 | struct mem_cgroup *memcg; |
23047a96 | 526 | |
55779ec7 | 527 | rcu_read_lock(); |
23047a96 JW |
528 | /* |
529 | * Filter non-memcg pages here, e.g. unmap can call | |
530 | * mark_page_accessed() on VDSO pages. | |
531 | * | |
532 | * XXX: See workingset_refault() - this should return | |
533 | * root_mem_cgroup even for !CONFIG_MEMCG. | |
534 | */ | |
c5ce619a | 535 | memcg = folio_memcg_rcu(folio); |
55779ec7 | 536 | if (!mem_cgroup_disabled() && !memcg) |
23047a96 | 537 | goto out; |
c5ce619a | 538 | workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio)); |
23047a96 | 539 | out: |
55779ec7 | 540 | rcu_read_unlock(); |
a528910e | 541 | } |
449dd698 JW |
542 | |
543 | /* | |
544 | * Shadow entries reflect the share of the working set that does not | |
545 | * fit into memory, so their number depends on the access pattern of | |
546 | * the workload. In most cases, they will refault or get reclaimed | |
547 | * along with the inode, but a (malicious) workload that streams | |
548 | * through files with a total size several times that of available | |
549 | * memory, while preventing the inodes from being reclaimed, can | |
550 | * create excessive amounts of shadow nodes. To keep a lid on this, | |
551 | * track shadow nodes and reclaim them when they grow way past the | |
552 | * point where they would still be useful. | |
553 | */ | |
554 | ||
9bbdc0f3 | 555 | struct list_lru shadow_nodes; |
14b46879 | 556 | |
a97e7904 | 557 | void workingset_update_node(struct xa_node *node) |
14b46879 | 558 | { |
2386eef2 SAS |
559 | struct address_space *mapping; |
560 | ||
14b46879 JW |
561 | /* |
562 | * Track non-empty nodes that contain only shadow entries; | |
563 | * unlink those that contain pages or are being freed. | |
564 | * | |
565 | * Avoid acquiring the list_lru lock when the nodes are | |
566 | * already where they should be. The list_empty() test is safe | |
b93b0163 | 567 | * as node->private_list is protected by the i_pages lock. |
14b46879 | 568 | */ |
2386eef2 SAS |
569 | mapping = container_of(node->array, struct address_space, i_pages); |
570 | lockdep_assert_held(&mapping->i_pages.xa_lock); | |
68d48e6a | 571 | |
01959dfe | 572 | if (node->count && node->count == node->nr_values) { |
68d48e6a | 573 | if (list_empty(&node->private_list)) { |
14b46879 | 574 | list_lru_add(&shadow_nodes, &node->private_list); |
da3ceeff | 575 | __inc_lruvec_kmem_state(node, WORKINGSET_NODES); |
68d48e6a | 576 | } |
14b46879 | 577 | } else { |
68d48e6a | 578 | if (!list_empty(&node->private_list)) { |
14b46879 | 579 | list_lru_del(&shadow_nodes, &node->private_list); |
da3ceeff | 580 | __dec_lruvec_kmem_state(node, WORKINGSET_NODES); |
68d48e6a | 581 | } |
14b46879 JW |
582 | } |
583 | } | |
449dd698 JW |
584 | |
585 | static unsigned long count_shadow_nodes(struct shrinker *shrinker, | |
586 | struct shrink_control *sc) | |
587 | { | |
449dd698 | 588 | unsigned long max_nodes; |
14b46879 | 589 | unsigned long nodes; |
95f9ab2d | 590 | unsigned long pages; |
449dd698 | 591 | |
14b46879 | 592 | nodes = list_lru_shrink_count(&shadow_nodes, sc); |
725cac1c ML |
593 | if (!nodes) |
594 | return SHRINK_EMPTY; | |
449dd698 | 595 | |
449dd698 | 596 | /* |
a97e7904 | 597 | * Approximate a reasonable limit for the nodes |
b5388998 JW |
598 | * containing shadow entries. We don't need to keep more |
599 | * shadow entries than possible pages on the active list, | |
600 | * since refault distances bigger than that are dismissed. | |
601 | * | |
602 | * The size of the active list converges toward 100% of | |
603 | * overall page cache as memory grows, with only a tiny | |
604 | * inactive list. Assume the total cache size for that. | |
605 | * | |
606 | * Nodes might be sparsely populated, with only one shadow | |
607 | * entry in the extreme case. Obviously, we cannot keep one | |
608 | * node for every eligible shadow entry, so compromise on a | |
609 | * worst-case density of 1/8th. Below that, not all eligible | |
610 | * refaults can be detected anymore. | |
449dd698 | 611 | * |
a97e7904 | 612 | * On 64-bit with 7 xa_nodes per page and 64 slots |
449dd698 | 613 | * each, this will reclaim shadow entries when they consume |
b5388998 | 614 | * ~1.8% of available memory: |
449dd698 | 615 | * |
a97e7904 | 616 | * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE |
449dd698 | 617 | */ |
95f9ab2d | 618 | #ifdef CONFIG_MEMCG |
b5388998 | 619 | if (sc->memcg) { |
95f9ab2d | 620 | struct lruvec *lruvec; |
2b487e59 | 621 | int i; |
95f9ab2d | 622 | |
867e5e1d | 623 | lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid)); |
2b487e59 | 624 | for (pages = 0, i = 0; i < NR_LRU_LISTS; i++) |
205b20cc JW |
625 | pages += lruvec_page_state_local(lruvec, |
626 | NR_LRU_BASE + i); | |
d42f3245 RG |
627 | pages += lruvec_page_state_local( |
628 | lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT; | |
629 | pages += lruvec_page_state_local( | |
630 | lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT; | |
95f9ab2d JW |
631 | } else |
632 | #endif | |
633 | pages = node_present_pages(sc->nid); | |
634 | ||
dad4f140 | 635 | max_nodes = pages >> (XA_CHUNK_SHIFT - 3); |
449dd698 | 636 | |
14b46879 | 637 | if (nodes <= max_nodes) |
449dd698 | 638 | return 0; |
14b46879 | 639 | return nodes - max_nodes; |
449dd698 JW |
640 | } |
641 | ||
642 | static enum lru_status shadow_lru_isolate(struct list_head *item, | |
3f97b163 | 643 | struct list_lru_one *lru, |
449dd698 | 644 | spinlock_t *lru_lock, |
a97e7904 | 645 | void *arg) __must_hold(lru_lock) |
449dd698 | 646 | { |
a97e7904 | 647 | struct xa_node *node = container_of(item, struct xa_node, private_list); |
449dd698 | 648 | struct address_space *mapping; |
449dd698 JW |
649 | int ret; |
650 | ||
651 | /* | |
f82cd2f0 | 652 | * Page cache insertions and deletions synchronously maintain |
b93b0163 | 653 | * the shadow node LRU under the i_pages lock and the |
449dd698 JW |
654 | * lru_lock. Because the page cache tree is emptied before |
655 | * the inode can be destroyed, holding the lru_lock pins any | |
a97e7904 | 656 | * address_space that has nodes on the LRU. |
449dd698 | 657 | * |
b93b0163 | 658 | * We can then safely transition to the i_pages lock to |
449dd698 JW |
659 | * pin only the address_space of the particular node we want |
660 | * to reclaim, take the node off-LRU, and drop the lru_lock. | |
661 | */ | |
662 | ||
01959dfe | 663 | mapping = container_of(node->array, struct address_space, i_pages); |
449dd698 JW |
664 | |
665 | /* Coming from the list, invert the lock order */ | |
b93b0163 | 666 | if (!xa_trylock(&mapping->i_pages)) { |
6ca342d0 | 667 | spin_unlock_irq(lru_lock); |
449dd698 JW |
668 | ret = LRU_RETRY; |
669 | goto out; | |
670 | } | |
671 | ||
5649d113 YY |
672 | /* For page cache we need to hold i_lock */ |
673 | if (mapping->host != NULL) { | |
674 | if (!spin_trylock(&mapping->host->i_lock)) { | |
675 | xa_unlock(&mapping->i_pages); | |
676 | spin_unlock_irq(lru_lock); | |
677 | ret = LRU_RETRY; | |
678 | goto out; | |
679 | } | |
51b8c1fe JW |
680 | } |
681 | ||
3f97b163 | 682 | list_lru_isolate(lru, item); |
da3ceeff | 683 | __dec_lruvec_kmem_state(node, WORKINGSET_NODES); |
68d48e6a | 684 | |
449dd698 JW |
685 | spin_unlock(lru_lock); |
686 | ||
687 | /* | |
688 | * The nodes should only contain one or more shadow entries, | |
689 | * no pages, so we expect to be able to remove them all and | |
690 | * delete and free the empty node afterwards. | |
691 | */ | |
01959dfe | 692 | if (WARN_ON_ONCE(!node->nr_values)) |
b936887e | 693 | goto out_invalid; |
01959dfe | 694 | if (WARN_ON_ONCE(node->count != node->nr_values)) |
b936887e | 695 | goto out_invalid; |
f82cd2f0 | 696 | xa_delete_node(node, workingset_update_node); |
da3ceeff | 697 | __inc_lruvec_kmem_state(node, WORKINGSET_NODERECLAIM); |
449dd698 | 698 | |
b936887e | 699 | out_invalid: |
6ca342d0 | 700 | xa_unlock_irq(&mapping->i_pages); |
5649d113 YY |
701 | if (mapping->host != NULL) { |
702 | if (mapping_shrinkable(mapping)) | |
703 | inode_add_lru(mapping->host); | |
704 | spin_unlock(&mapping->host->i_lock); | |
705 | } | |
449dd698 JW |
706 | ret = LRU_REMOVED_RETRY; |
707 | out: | |
449dd698 | 708 | cond_resched(); |
6ca342d0 | 709 | spin_lock_irq(lru_lock); |
449dd698 JW |
710 | return ret; |
711 | } | |
712 | ||
713 | static unsigned long scan_shadow_nodes(struct shrinker *shrinker, | |
714 | struct shrink_control *sc) | |
715 | { | |
b93b0163 | 716 | /* list_lru lock nests inside the IRQ-safe i_pages lock */ |
6b51e881 SAS |
717 | return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate, |
718 | NULL); | |
449dd698 JW |
719 | } |
720 | ||
721 | static struct shrinker workingset_shadow_shrinker = { | |
722 | .count_objects = count_shadow_nodes, | |
723 | .scan_objects = scan_shadow_nodes, | |
4b85afbd | 724 | .seeks = 0, /* ->count reports only fully expendable nodes */ |
0a6b76dd | 725 | .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, |
449dd698 JW |
726 | }; |
727 | ||
728 | /* | |
729 | * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe | |
b93b0163 | 730 | * i_pages lock. |
449dd698 JW |
731 | */ |
732 | static struct lock_class_key shadow_nodes_key; | |
733 | ||
734 | static int __init workingset_init(void) | |
735 | { | |
612e4493 JW |
736 | unsigned int timestamp_bits; |
737 | unsigned int max_order; | |
449dd698 JW |
738 | int ret; |
739 | ||
612e4493 JW |
740 | BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT); |
741 | /* | |
742 | * Calculate the eviction bucket size to cover the longest | |
743 | * actionable refault distance, which is currently half of | |
744 | * memory (totalram_pages/2). However, memory hotplug may add | |
745 | * some more pages at runtime, so keep working with up to | |
746 | * double the initial memory by using totalram_pages as-is. | |
747 | */ | |
748 | timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT; | |
ca79b0c2 | 749 | max_order = fls_long(totalram_pages() - 1); |
612e4493 JW |
750 | if (max_order > timestamp_bits) |
751 | bucket_order = max_order - timestamp_bits; | |
d3d36c4b | 752 | pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n", |
612e4493 JW |
753 | timestamp_bits, max_order, bucket_order); |
754 | ||
e33c267a | 755 | ret = prealloc_shrinker(&workingset_shadow_shrinker, "mm-shadow"); |
449dd698 JW |
756 | if (ret) |
757 | goto err; | |
c92e8e10 KT |
758 | ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key, |
759 | &workingset_shadow_shrinker); | |
449dd698 JW |
760 | if (ret) |
761 | goto err_list_lru; | |
39887653 | 762 | register_shrinker_prepared(&workingset_shadow_shrinker); |
449dd698 JW |
763 | return 0; |
764 | err_list_lru: | |
39887653 | 765 | free_prealloced_shrinker(&workingset_shadow_shrinker); |
449dd698 JW |
766 | err: |
767 | return ret; | |
768 | } | |
769 | module_init(workingset_init); |